/* p_lx_elf.cpp --

   This file is part of the UPX executable compressor.

   Copyright (C) 1996-2025 Markus Franz Xaver Johannes Oberhumer
   Copyright (C) 1996-2025 Laszlo Molnar
   Copyright (C) 2000-2025 John F. Reiser
   All Rights Reserved.

   UPX and the UCL library are free software; you can redistribute them
   and/or modify them under the terms of the GNU General Public License as
   published by the Free Software Foundation; either version 2 of
   the License, or (at your option) any later version.

   This program is distributed in the hope that it will be useful,
   but WITHOUT ANY WARRANTY; without even the implied warranty of
   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
   GNU General Public License for more details.

   You should have received a copy of the GNU General Public License
   along with this program; see the file COPYING.
   If not, write to the Free Software Foundation, Inc.,
   59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.

   Markus F.X.J. Oberhumer              Laszlo Molnar
   <markus@oberhumer.com>               <ezerotven+github@gmail.com>

   John F. Reiser
   <jreiser@users.sourceforge.net>
 */


#define ALLOW_INT_PLUS_MEMBUFFER 1
#include "conf.h"

#include "file.h"
#include "filter.h"
#include "linker.h"
#include "packer.h"
#include "p_elf.h"
#include "p_unix.h"
#include "p_lx_exc.h"
#include "p_lx_elf.h"
#include "ui.h"

// NOLINTBEGIN(clang-analyzer-core.CallAndMessage)
// NOLINTBEGIN(clang-analyzer-deadcode.DeadStores)

#if defined(__CHERI__) && defined(__CHERI_PURE_CAPABILITY__)
#  pragma clang diagnostic ignored "-Wcheri-capability-misuse" // TODO later
#  pragma clang diagnostic ignored "-Wcheri-provenance" // TODO later
#endif

using upx::umin;

#define PT_LOAD Elf32_Phdr::PT_LOAD  /* 64-bit PT_LOAD is the same */
#define PT_NOTE32   Elf32_Phdr::PT_NOTE
#define PT_NOTE64   Elf64_Phdr::PT_NOTE
#define PT_GNU_STACK32  Elf32_Phdr::PT_GNU_STACK
#define PT_GNU_STACK64  Elf64_Phdr::PT_GNU_STACK
#define PT_GNU_RELRO32  Elf32_Phdr::PT_GNU_RELRO
#define PT_GNU_RELRO64  Elf64_Phdr::PT_GNU_RELRO

// also see stub/src/MAX_ELF_HDR.[Sc]
static constexpr unsigned MAX_ELF_HDR_32 = 512;
static constexpr unsigned MAX_ELF_HDR_64 = 1024;

//static unsigned const EF_ARM_HASENTRY = 0x02;
static unsigned const EF_ARM_EABI_VER4 = 0x04000000;
static unsigned const EF_ARM_EABI_VER5 = 0x05000000;

/*static*/ const unsigned char PackLinuxElf::o_shstrtab[] = {  \
/*start*/       '\0',
/*offset  1*/   '.','n','o','t','e','.','g','n','u','.','b','u','i','l','d','-','i','d','\0',
/*offset 20*/   '.','s','h','s','t','r','t','a','b','\0'
};

static unsigned
up4(unsigned x)
{
    return ~3u & (3+ x);
}

static unsigned
up8(unsigned x)
{
    return ~7u & (7+ x);
}

static off_t
fpadN(OutputFile *fo, unsigned len)
{
    if (len) {
        MemBuffer pad(len); pad.clear();
        fo->write(pad, len);
    }
    return fo->st_size();
}

static off_t
fpad4(OutputFile *fo, unsigned pos)
{
    (void)pos;  // debug: compare 'pos' with "shell grep pos /proc/PID/fdinfo/FD"
    if (!fo) { // --test, --list
        return 0;
    }
    off_t len = fo->st_size();
    unsigned d = 3u & (0 - len);
    if (d) {
        unsigned zero = 0;
        fo->write(&zero, d);
    }
    return d + len;
}

static off_t
fpad8(OutputFile *fo, unsigned pos)
{
    (void)pos;  // debug: compare 'pos' with "shell grep pos /proc/PID/fdinfo/FD"
    if (!fo) { // --test, --list
        return 0;
    }
    off_t len = fo->st_size();
    unsigned d = 7u & (0 - len);
    if (d) {
        upx_uint64_t zero = 0;
        fo->write(&zero, d);
    }
    return d + len;
}

static unsigned
funpad4(InputFile *fi)
{
    unsigned d = 3u & (0 - fi->tell());
    if (d)
        fi->seek(d, SEEK_CUR);
    return d;
}

static void alloc_file_image(MemBuffer &mb, off_t size)
{
    assert(mem_size_valid_bytes(size));
    if (mb.getVoidPtr() == nullptr) {
        mb.alloc(size);
    } else {
        assert((u32_t)size <= mb.getSize());
    }
}

int
PackLinuxElf32::checkEhdr(Elf32_Ehdr const *ehdr) const
{
    const unsigned char * const buf = ehdr->e_ident;

    if (0!=memcmp(buf, "\x7f\x45\x4c\x46", 4)  // "\177ELF"
    ||  buf[Elf32_Ehdr::EI_CLASS]!=ei_class
    ||  buf[Elf32_Ehdr::EI_DATA] !=ei_data
    ) {
        return -1;
    }
    if (!memcmp(buf+8, "FreeBSD", 7))                   // branded
        return 1;

    int const type = get_te16(&ehdr->e_type);
    if (type != Elf32_Ehdr::ET_EXEC && type != Elf32_Ehdr::ET_DYN)
        return 2;
    if (get_te16(&ehdr->e_machine) != (unsigned) e_machine)
        return 3;
    if (get_te32(&ehdr->e_version) != Elf32_Ehdr::EV_CURRENT)
        return 4;
    if (e_phnum < 1)
        return 5;
    if (get_te16(&ehdr->e_phentsize) != sizeof(Elf32_Phdr))
        return 6;

    if (type == Elf32_Ehdr::ET_EXEC) {
        // check for Linux kernels
        unsigned const entry = get_te32(&ehdr->e_entry);
        if (entry == 0xC0100000)    // uncompressed vmlinux
            return 1000;
        if (entry == 0x00001000)    // compressed vmlinux
            return 1001;
        if (entry == 0x00100000)    // compressed bvmlinux
            return 1002;
    }

    // FIXME: add more checks for kernels

    // FIXME: add special checks for other ELF i386 formats, like
    //        NetBSD, OpenBSD, Solaris, ....

    // success
    return 0;
}

int
PackLinuxElf64::checkEhdr(Elf64_Ehdr const *ehdr) const
{
    const unsigned char * const buf = ehdr->e_ident;
    unsigned char osabi0 = buf[Elf32_Ehdr::EI_OSABI];
    if (0==osabi0) {
        osabi0 = opt->o_unix.osabi0;
    }

    if (0!=memcmp(buf, "\x7f\x45\x4c\x46", 4)  // "\177ELF"
    ||  buf[Elf64_Ehdr::EI_CLASS]!=ei_class
    ||  buf[Elf64_Ehdr::EI_DATA] !=ei_data
    ||                     osabi0!=ei_osabi
    ) {
        return -1;
    }
    if (!memcmp(buf+8, "FreeBSD", 7))                   // branded
        return 1;

    int const type = get_te16(&ehdr->e_type);
    if (type != Elf64_Ehdr::ET_EXEC && type != Elf64_Ehdr::ET_DYN)
        return 2;
    if (get_te16(&ehdr->e_machine) != (unsigned) e_machine)
        return 3;
    if (get_te32(&ehdr->e_version) != Elf64_Ehdr::EV_CURRENT)
        return 4;
    if (e_phnum < 1)
        return 5;
    if (get_te16(&ehdr->e_phentsize) != sizeof(Elf64_Phdr))
        return 6;

    if (type == Elf64_Ehdr::ET_EXEC) {
        // check for Linux kernels
        upx_uint64_t const entry = get_te64(&ehdr->e_entry);
        if (entry == 0xC0100000)    // uncompressed vmlinux
            return 1000;
        if (entry == 0x00001000)    // compressed vmlinux
            return 1001;
        if (entry == 0x00100000)    // compressed bvmlinux
            return 1002;
    }

    // FIXME: add more checks for kernels

    // FIXME: add special checks for other ELF i386 formats, like
    //        NetBSD, OpenBSD, Solaris, ....

    // success
    return 0;
}

PackLinuxElf::PackLinuxElf(InputFile *f)
    : super(f), e_phnum(0), dynstr(nullptr),
    sz_phdrs(0), sz_elf_hdrs(0), sz_pack2(0), sz_pack2a(0),
    lg2_page(12), page_size(1u<<lg2_page), is_pie(0), is_asl(0),
    xct_off(0), o_binfo(0), so_slide(0), xct_va(0), jni_onload_va(0),
    user_init_va(0), user_init_off(0),
    e_machine(0), ei_class(0), ei_data(0), ei_osabi(0), osabi_note(nullptr),
    shstrtab(nullptr),
    o_elf_shnum(0)
{
    memset(dt_table, 0, sizeof(dt_table));
    symnum_max = 0;
    user_init_rp = nullptr;
}

PackLinuxElf::~PackLinuxElf()
{
}

int PackLinuxElf32::is_LOAD(Elf32_Phdr const *phdr) const
{
    // (1+ PT_LOPROC) can confuse!
    return PT_LOAD == get_te32(&phdr->p_type);
}

int PackLinuxElf64::is_LOAD(Elf64_Phdr const *phdr) const
{
    // (1+ PT_LOPROC) can confuse!
    return PT_LOAD == get_te32(&phdr->p_type);
}

void
PackLinuxElf32::PackLinuxElf32help1(InputFile *f)
{
    e_type  = get_te16(&ehdri.e_type);
    e_phnum = get_te16(&ehdri.e_phnum);
    e_shnum = get_te16(&ehdri.e_shnum);
    unsigned const e_phentsize = get_te16(&ehdri.e_phentsize);
    if (memcmp((char const *)&ehdri, "\x7f\x45\x4c\x46", 4)  // "\177ELF"
    || ehdri.e_ident[Elf32_Ehdr::EI_CLASS]!=Elf32_Ehdr::ELFCLASS32
    || sizeof(Elf32_Phdr) != e_phentsize
    || (Elf32_Ehdr::ELFDATA2MSB == ehdri.e_ident[Elf32_Ehdr::EI_DATA]
            && &N_BELE_RTP::be_policy != bele)
    || (Elf32_Ehdr::ELFDATA2LSB == ehdri.e_ident[Elf32_Ehdr::EI_DATA]
            && &N_BELE_RTP::le_policy != bele)) {
        e_phoff = 0;
        e_shoff = 0;
        sz_phdrs = 0;
        return;
    }
    if (0==e_phnum) throwCantUnpack("0==e_phnum");
    e_phoff = get_te32(&ehdri.e_phoff);
    unsigned const last_Phdr = e_phoff + e_phnum * usizeof(Elf32_Phdr);
    if (last_Phdr < e_phoff  // wrap-around
    ||  e_phoff != sizeof(Elf32_Ehdr)  // must be contiguous
    ||  (unsigned long)file_size < last_Phdr) {
        throwCantUnpack("bad e_phoff %#x", e_phoff);
    }
    e_shoff = get_te32(&ehdri.e_shoff);
    e_shstrndx = get_te16(&ehdri.e_shstrndx);
    unsigned const last_Shdr = e_shoff + e_shnum * usizeof(Elf32_Shdr);
    if (last_Shdr < e_shoff  // wrap-around
    ||  (e_shnum && e_shoff < last_Phdr)
    ||  (unsigned long)file_size < last_Shdr) {
        if (opt->cmd == CMD_COMPRESS) {
            throwCantUnpack("bad e_shoff %#x", e_shoff);
        }
    }
    sz_phdrs = e_phnum * e_phentsize;
    sz_elf_hdrs = sz_phdrs + sizeof(Elf32_Ehdr) +
        n_phdrx * sizeof(Elf32_Phdr);  // phdrx bodies later: ::generateElfHdr

    if (f && Elf32_Ehdr::ET_DYN!=e_type) {
        unsigned const len = file_size;  // (sz_phdrs + e_phoff) except --preserve-build-id
        alloc_file_image(file_image, len);
        f->seek(0, SEEK_SET);
        f->readx(file_image, len);
        phdri= (Elf32_Phdr       *)(e_phoff + file_image);  // do not free() !!
    }
    if (f && Elf32_Ehdr::ET_DYN==e_type) {
        // The DT_SYMTAB has no designated length.  Read the whole file.
        alloc_file_image(file_image, file_size);
        f->seek(0, SEEK_SET);
        f->readx(file_image, file_size);
        phdri= (Elf32_Phdr *)(e_phoff + file_image);  // do not free() !!
        if (opt->cmd != CMD_COMPRESS || !e_shoff ||  file_size < e_shoff) {
            shdri = nullptr;
        }
        else {  // FIXME: copy from file_image ?
            fi->seek(e_shoff, SEEK_SET);
            if (mb_shdr.getSize() != sizeof(Elf32_Shdr) * e_shnum) {
                mb_shdr.alloc(   sizeof(Elf32_Shdr) * e_shnum);
            }
            shdri = (Elf32_Shdr *)mb_shdr.getVoidPtr();
            fi->readx(shdri, sizeof(Elf32_Shdr) * e_shnum);
        }
        sec_dynsym = elf_find_section_type(Elf32_Shdr::SHT_DYNSYM);
        if (sec_dynsym) {
            unsigned t = get_te32(&sec_dynsym->sh_link);
            if (e_shnum <= t)
                throwCantPack("bad dynsym->sh_link");
            sec_dynstr = &shdri[t];
        }

        Elf32_Phdr const *phdr= phdri;
        for (int j = e_phnum; --j>=0; ++phdr)
        if (Elf32_Phdr::PT_DYNAMIC==get_te32(&phdr->p_type)) {
            unsigned offset = check_pt_dynamic(phdr);
            dynseg= (Elf32_Dyn *)(offset + file_image);
            invert_pt_dynamic(dynseg,
                umin(get_te32(&phdr->p_filesz), (unsigned)(file_size_u - offset)));
        }
        else if (is_LOAD(phdr)) {
            check_pt_load(phdr);
        }
        // elf_find_dynamic() returns 0 if 0==dynseg.
        dynstr =          (char const *)elf_find_dynamic(Elf32_Dyn::DT_STRTAB);
        dynsym = (Elf32_Sym /*const*/ *)elf_find_dynamic(Elf32_Dyn::DT_SYMTAB);
        gashtab =     (unsigned const *)elf_find_dynamic(Elf32_Dyn::DT_GNU_HASH);
        hashtab =     (unsigned const *)elf_find_dynamic(Elf32_Dyn::DT_HASH);
        if (3& ((upx_uintptr_t)dynsym | (upx_uintptr_t)gashtab | (upx_uintptr_t)hashtab)) {
            throwCantPack("unaligned DT_SYMTAB, DT_GNU_HASH, or DT_HASH/n");
        }
        jni_onload_sym = elf_lookup("JNI_OnLoad");
        if (jni_onload_sym) {
            jni_onload_va = get_te32(&jni_onload_sym->st_value);
            jni_onload_va = 0;  // FIXME not understood; need example
        }
    }
}

#define WANT_EHDR_ENUM
#define WANT_REL_ENUM
#include "p_elf_enum.h"
#undef WANT_REL_ENUM
#undef WANT_EHDR_ENUM

off_t PackLinuxElf::pack3(OutputFile *fo, Filter &ft) // return length of output
{
    if (!fo) {
        return 0;
    }
    unsigned disp;
    unsigned len = sz_pack2a;  // after headers and all PT_LOAD

    unsigned const t = (4 & len) ^ ((!!xct_off)<<2);  // 0 or 4
    if (t) {
        if (fo) {
            unsigned const zero = 0;
            fo->write(&zero, t);
        }
        len += t;  // force sz_pack2 (0 mod 8)  [see below]
    }

    set_te32(&disp, sz_elf_hdrs + usizeof(p_info) + usizeof(l_info) +
        (!!xct_off & !!is_asl));  // |1 iff android shlib
    fo->write(&disp, sizeof(disp));  // offset(b_info)
        // FIXME: If is_shlib then that is useful only for the is_asl bit.
        // Better info is the word below with (overlay_offset - sizeof(linfo)).

    len += sizeof(disp);
    set_te32(&disp, len);  // distance back to beginning (detect dynamic reloc)
    fo->write(&disp, sizeof(disp));
    len += sizeof(disp);

    if (xct_off) {  // is_shlib
        upx_uint64_t const firstpc_va = (jni_onload_va
            ? jni_onload_va
            : user_init_va);
        set_te32(&disp, firstpc_va - load_va);
        fo->write(&disp, sizeof(disp));  // DT_INIT.d_val or DT_INIT_ARRAY[0]
        len += sizeof(disp);

        set_te32(&disp, xct_off);
        fo->write(&disp, sizeof(disp));  // offset(lowest_executable_instr)
        len += sizeof(disp);

        if (is_asl) {
            xct_off += asl_delta;  // the extra page
        }
        set_te32(&disp, overlay_offset - sizeof(linfo));
        fo->write(&disp, sizeof(disp));  // &{l_info; p_info; b_info}
        len += sizeof(disp);
    }
    total_out += len - sz_pack2a;
    sz_pack2 = len;  // 0 mod 8  [see above]

    // FIXME: debugging aid: entry to decompressor
    if (lowmem.getSize()) {
        Elf32_Ehdr *const ehdr = (Elf32_Ehdr *)&lowmem[0];
        set_te32(&ehdr->e_entry, sz_pack2);  // hint for decompressor
    }
    // end debugging aid

    super::pack3(fo, ft);  // append the decompressor
    set_te16(&linfo.l_lsize, up4(  // MATCH03: up4
    get_te16(&linfo.l_lsize) + len - sz_pack2a));
    total_out = fpad4(fo, total_out);  // MATCH03
    return total_out;
}

static unsigned
is_pow2(u64_t x)
{
    return !((-1+ x) & x);
}

Elf32_Phdr const *
PackLinuxElf32::elf_find_Phdr_for_va(upx_uint32_t addr, Elf32_Phdr const *phdr, unsigned phnum)
{
    for (unsigned j = 0; j < phnum; ++j, ++phdr) {
        u32_t align = get_te32(&phdr->p_align);
        if (is_LOAD(phdr)
        && is_pow2(align) && !((-1+ align) &
            (get_te32(&phdr->p_vaddr) ^ get_te32(&phdr->p_offset)))
        &&  (addr - get_te32(&phdr->p_vaddr)) < get_te32(&phdr->p_filesz)) {
            return phdr;
        }
    }
    return nullptr;
}

Elf64_Phdr const *
PackLinuxElf64::elf_find_Phdr_for_va(upx_uint64_t addr, Elf64_Phdr const *phdr, unsigned phnum)
{
    for (unsigned j = 0; j < phnum; ++j, ++phdr) {
        u64_t align = get_te64(&phdr->p_align);
        if (is_LOAD(phdr)
        && is_pow2(align) && !((-1+ align) &
            (get_te64(&phdr->p_vaddr) ^ get_te64(&phdr->p_offset)))
        &&  (addr - get_te64(&phdr->p_vaddr)) < get_te64(&phdr->p_filesz)) {
            return phdr;
        }
    }
    return nullptr;
}

unsigned
PackLinuxElf32::slide_sh_offset(Elf32_Shdr *shdr)
{
    unsigned sh_offset = get_te32(&shdr->sh_offset);
    unsigned sh_addr   = get_te32(&shdr->sh_addr);
    if (Elf32_Shdr::SHF_WRITE & get_te32(&shdr->sh_flags)
      || (sh_offset && !sh_addr))
    {
        unsigned newoff = so_slide + sh_offset + (is_asl ? asl_delta : 0);
        if ((unsigned)this->file_size < newoff) {
            throwInternalError("bad slide %p %#x", shdr, (unsigned)so_slide);
        }
        set_te32(&shdr->sh_offset, newoff);
        if (sh_addr) // change only if non-zero
            set_te32(&shdr->sh_addr,
                so_slide + sh_addr + (is_asl ? asl_delta : 0));
        return newoff;
    }
    return sh_offset;
}

void
PackLinuxElf32::asl_slide_Shdrs()
{
    Elf32_Shdr *shdr = shdro;
    for (unsigned j = 0; j < e_shnum; ++shdr, ++j) {
        unsigned sh_offset = get_te32(&shdr->sh_offset);
        if (xct_off < sh_offset) {
            slide_sh_offset(shdr);
        }
    }
}

unsigned
PackLinuxElf64::slide_sh_offset(Elf64_Shdr *shdr)
{
    unsigned sh_offset = get_te64(&shdr->sh_offset);
    unsigned sh_addr   = get_te64(&shdr->sh_addr);
    if (Elf64_Shdr::SHF_WRITE & get_te64(&shdr->sh_flags)
      || (sh_offset && !sh_addr))
    {
        unsigned newoff = so_slide + sh_offset + (is_asl ? asl_delta : 0);
        if ((unsigned)this->file_size < newoff) {
            throwInternalError("bad slide %p %#x", shdr, (unsigned)so_slide);
        }
        set_te64(&shdr->sh_offset, newoff);
        if (sh_addr) // change only if non-zero
            set_te64(&shdr->sh_addr,
                so_slide + sh_addr + (is_asl ? asl_delta : 0));
        return newoff;
    }
    return sh_offset;
}

void
PackLinuxElf64::asl_slide_Shdrs()
{
    Elf64_Shdr *shdr = shdro;
    for (unsigned j = 0; j < e_shnum; ++shdr, ++j) {
        unsigned sh_offset = get_te64_32(&shdr->sh_offset);
        if (xct_off < sh_offset) {
            slide_sh_offset(shdr);
        }
    }
}

// C_BASE covers the convex hull of the PT_LOAD of the uncompressed module.
// It has (PF_W & .p_flags), and is ".bss": empty (0==.p_filesz, except a bug
// in Linux kernel forces 0x1000==.p_filesz) with .p_memsz equal to the brk(0).
// It is first in order to reserve all // pages, in particular so that if
// (64K == .p_align) but at runtime (4K == PAGE_SIZE) then the Linux kernel
// does not put [vdso] and [vvar] into alignment holes that the UPX runtime stub
// will overwrite.
//
// Note that C_TEXT[.p_vaddr, +.p_memsz) is a subset of C_BASE.
// This requires that the kernel process the ELFxx_Phdr in ascending order,
// and does not mind the overlap.  The UPX runtime stub will "re-program"
// the memory regions anyway.
enum { // ordinals in ELFxx_Phdr[] of compressed output
      C_BASE = 0  // reserve address space
    , C_TEXT = 1  // compressed data and stub
    , C_NOTE = 2  // PT_NOTE copied from input
    , C_GSTK = 3  // PT_GNU_STACK; will be 2 if no PT_NOTE
};

off_t PackLinuxElf32::pack3(OutputFile *fo, Filter &ft)
{
    if (!overlay_offset) {
        overlay_offset = sizeof(linfo) + (xct_off ? xct_off : sz_elf_hdrs);
    }

    total_out = super::pack3(fo, ft);  // loader follows compressed PT_LOADs
    if (fo && xct_off  && !is_asl && Elf32_Dyn::DT_INIT != upx_dt_init) {
        // Patch user_init_rp.  is_asl has done this already in asl_pack2_Shdrs()
        // Section .init_array[0] must have R_$(ARCH)_RELATIVE relocation.
        fo->seek((char *)user_init_rp - (char *)&file_image[0], SEEK_SET);
        Elf32_Rel rel(*(Elf32_Rel const *)user_init_rp);
        u32_t r_info = get_te32(&((Elf32_Rel const *)user_init_rp)->r_info);
        u32_t r_type = (Elf32_Ehdr::EM_386  == e_machine) ? R_386_RELATIVE
                     : (Elf32_Ehdr::EM_ARM  == e_machine) ? R_ARM_RELATIVE
                     : (Elf32_Ehdr::EM_PPC  == e_machine) ? R_PPC_RELATIVE
                     : (Elf32_Ehdr::EM_MIPS == e_machine) ? R_MIPS_32
                     : 0;
        set_te32(&rel.r_info, ELF32_R_INFO(ELF32_R_SYM(r_info), r_type));
        if (is_asl) {
            u32_t r_offset = get_te32(&((Elf32_Rel const *)user_init_rp)->r_offset);
            set_te32(&rel.r_offset, asl_delta + r_offset);
        }
        fo->rewrite(&rel, sizeof(rel));
        fo->seek(0, SEEK_END);

        // Value of init_array[0] will be changed later.
        // See write() of 'cpr_entry' below.
    }
    // NOTE: PackLinuxElf::pack3  adjusted xct_off for the extra page

    // Then compressed gaps (including debuginfo.)
    for (unsigned k = 0; k < e_phnum; ++k) {
        Extent x;
        x.size = find_LOAD_gap(phdri, k, e_phnum);
        if (x.size) {
            x.offset = get_te32(&phdri[k].p_offset) +
                       get_te32(&phdri[k].p_filesz);
            packExtent(x, nullptr, fo);
        }
    }
    // write block end marker (uncompressed size 0)
    b_info hdr; memset(&hdr, 0, sizeof(hdr));
    set_le32(&hdr.sz_cpr, UPX_MAGIC_LE32);
    fo->write(&hdr, sizeof(hdr));
    total_out = fpad4(fo, total_out);

    if (0==xct_off) { // not shared library
        set_te32(&elfout.phdr[C_BASE].p_align, 0u - page_mask);
        elfout.phdr[C_BASE].p_paddr = elfout.phdr[C_BASE].p_vaddr;
        elfout.phdr[C_BASE].p_offset = 0;
        unsigned abrk = getbrk(phdri, e_phnum);
        // vbase handles ET_EXEC.  FIXME: pre-linking?
        unsigned vbase = get_te32(&elfout.phdr[C_BASE].p_vaddr);
        set_te32(&elfout.phdr[C_BASE].p_filesz, 0x1000);  // Linux kernel SIGSEGV if (0==.p_filesz)
        set_te32(&elfout.phdr[C_BASE].p_memsz, abrk - vbase);
        set_te32(&elfout.phdr[C_BASE].p_flags, Elf32_Phdr::PF_W|Elf32_Phdr::PF_R);
        set_te32(&elfout.phdr[C_TEXT].p_filesz, sz_pack2 + lsize);
        set_te32(&elfout.phdr[C_TEXT].p_memsz,  sz_pack2 + lsize);
        set_te32(&elfout.phdr[C_TEXT].p_vaddr, abrk= (page_mask & (~page_mask + abrk)));
        elfout.phdr[C_TEXT].p_paddr = elfout.phdr[C_TEXT].p_vaddr;
        set_te32(&elfout.ehdr.e_entry, abrk + get_te32(&elfout.ehdr.e_entry) - vbase);
    }
    if (0!=xct_off) { // shared library
        unsigned const cpr_entry = load_va + sz_pack2;
        set_te32(&file_image[user_init_off], cpr_entry);  // set the hook

        if (user_init_rp) { // decompressor needs hint for DT_INIT_ARRAY
            Elf32_Dyn *dynp = (Elf32_Dyn *)elf_find_dynptr(Elf32_Dyn::DT_NULL);
            set_te32(&dynp->d_val, (char *)user_init_rp - (char *)&file_image[0]);
        }

        Elf32_Phdr *const phdr0 = (Elf32_Phdr *)lowmem.subref(
                "bad e_phoff", e_phoff, e_phnum * sizeof(Elf32_Phdr));
        Elf32_Phdr *phdr = phdr0;
        upx_off_t off = fo->st_size();  // 64 bits
        so_slide = 0;
        for (unsigned j = 0; j < e_phnum; ++j, ++phdr) {
            // p_vaddr and p_paddr do not change!
            unsigned const len  = get_te32(&phdr->p_filesz);
            unsigned const ioff = get_te32(&phdri[j].p_offset);  // without asl_delta
            unsigned       align= get_te32(&phdr->p_align);
            unsigned const type = get_te32(&phdr->p_type);
            if (Elf32_Phdr::PT_INTERP==type) {
                // Rotate to highest position, so it can be lopped
                // by decrementing e_phnum.
                memcpy((unsigned char *)ibuf, phdr, sizeof(*phdr));  // extract
                memmove(phdr, 1+phdr, (e_phnum - (1+ j))*sizeof(*phdr));  // overlapping
                memcpy(&phdr[e_phnum - (1+ j)], (unsigned char *)ibuf, sizeof(*phdr));  // to top
                --phdr; --e_phnum;
                set_te16(&ehdri.e_phnum, e_phnum);
                set_te16(&((Elf32_Ehdr *)(unsigned char *)lowmem)->e_phnum, e_phnum);
                continue;
            }
            if (PT_LOAD == type) {
                if (!ioff) { // first PT_LOAD must contain everything written so far
                    set_te32(&phdr->p_filesz, sz_pack2 + lsize);  // is this correct?
                    set_te32(&phdr->p_memsz,  sz_pack2 + lsize);
                    // so_entry._start is in 1st PT_LOAD, and must be eXecutable
                    unsigned p_flags = get_te32(&phdr->p_flags);
                    set_te32(&phdr->p_flags, Elf32_Phdr::PF_X | p_flags);
                }
                else if ((xct_off - ioff) < len) { // Change length of compressed PT_LOAD.
                    set_te32(&phdr->p_filesz, total_out - ioff);  // FIXME  (sz_pack2 + lsize - ioff) ?
                    set_te32(&phdr->p_memsz,  total_out - ioff);
                    if (user_init_off < xct_off) { // MIPS puts PT_DYNAMIC here
                        // Allow for DT_INIT in a new [stolen] slot
                        unsigned off2 = user_init_off - sizeof(unsigned);
                        fo->seek(off2, SEEK_SET);
                        fo->rewrite(&file_image[off2], 2*sizeof(unsigned));
                    }
                }
                else if (xct_off < ioff) { // Slide subsequent PT_LOAD.
                    if ((1u<<12) < align
                    &&  (  Elf32_Ehdr::EM_386 == e_machine
                        || Elf32_Ehdr::EM_ARM == e_machine)  // FIXME: other $ARCH ?
                    ) {
                        align = 1u<<12;
                        set_te32(&phdr->p_align, align);
                    }
                    off = fpadN(fo, (-1 + align) & (ioff - off));
                    if (!so_slide) { // only once
                        so_slide = off - ((is_asl ? asl_delta : 0) + ioff);
                        //fprintf(stderr, "\nso_slide = %#x\n", (unsigned)so_slide);
                        //asl_slide_Shdrs();
                    }
                    set_te32(&phdr->p_offset, off);
                    fo->seek(off, SEEK_SET);
                    fo->write(&file_image[ioff], len);
                    off += len;
                    total_out = off;

                    if ((user_init_off - ioff) < len) {
                        fo->seek(user_init_off + so_slide, SEEK_SET);
                        unsigned word = cpr_entry;
                        set_te32(&word, cpr_entry);
                        fo->rewrite(&word, sizeof(word));
                        fo->seek(0, SEEK_END);
                    }
                }
                continue;  // all done with this PT_LOAD
            }
            if (xct_off < ioff) {
                set_te32(&phdr->p_offset, so_slide + (is_asl ? asl_delta : 0) + ioff);
            }
        }  // end each Phdr

        if (sec_arm_attr || is_asl) { // must update Shdr.sh_offset for so_slide
            Elf32_Shdr *shdr = shdri;
            for (unsigned j = 0; j < e_shnum; ++shdr, ++j) {
                unsigned sh_type = get_te32(&shdr->sh_type);
                if (Elf32_Shdr::SHT_REL == sh_type
                &&  n_jmp_slot  // FIXME  who sets this?
                &&  !strcmp(".rel.plt", get_te32(&shdr->sh_name) + shstrtab)) {
                    unsigned va = elf_unsigned_dynamic(Elf32_Dyn::DT_PLTGOT)
                        - (is_asl ? asl_delta : 0);
                    // Now use the old Phdrs (phdri)
                    Elf32_Phdr const *phva;
                    phva = elf_find_Phdr_for_va(va, phdri, e_phnum);
                    unsigned old_off = (va - get_te32(&phva->p_vaddr))
                        + get_te32(&phva->p_offset);

                    // Now use the new Phdrs (phdr0)
                    va += (is_asl ? asl_delta : 0);
                    phva = elf_find_Phdr_for_va(va, phdr0, e_phnum);
                    unsigned new_off = (va - get_te32(&phva->p_vaddr))
                        + get_te32(&phva->p_offset);

                    if (fo && n_jmp_slot) {
                        fo->seek(new_off, SEEK_SET);
                        fo->rewrite(&file_image[old_off], n_jmp_slot * 4);
                    }
                }
                if (j && shdr->sh_addr == 0
                &&  get_te32(&shdr->sh_offset) < xct_off) {
                    // Try to be nice by sliding; but still fails if compressed.
                    // So don't do it unless appending plain text of shstrtab.
                    unsigned sh_off = get_te32(&shdr->sh_offset);
                    if (xct_off < sh_off) {
                        slide_sh_offset(shdr);
                    }
                }
            }
            // Maybe: append plain text of shstrtab strings?
            fo->seek(total_out, SEEK_SET);
            if (xct_off < e_shoff) {
                set_te32(&((Elf32_Ehdr *)lowmem.getVoidPtr())->e_shoff, total_out);
                if (fo) {
                    fo->write(shdri, e_shnum * sizeof(*shdr));
                    total_out += e_shnum * sizeof(*shdr);
                }
            }
        }
        else { // output has no Shdr
            ehdri.e_shnum = 0;
            ehdri.e_shoff = 0;
            ehdri.e_shstrndx = 0;
        }
    }
    return total_out;
}

off_t PackLinuxElf64::pack3(OutputFile *fo, Filter &ft)
{
    if (!overlay_offset) {
        overlay_offset = sizeof(linfo) + (xct_off ? xct_off : sz_elf_hdrs);
    }

    total_out = super::pack3(fo, ft);  // loader follows compressed PT_LOADs
    if (fo && xct_off && Elf64_Dyn::DT_INIT != upx_dt_init) { // patch user_init_rp
        fo->seek((char *)user_init_rp - (char *)&file_image[0], SEEK_SET);
        Elf64_Rela rela(*(Elf64_Rela const *)user_init_rp);
        //u64_t r_info = get_te64(&((Elf64_Rela const *)user_init_rp)->r_info);
        u32_t r_type = (Elf64_Ehdr::EM_AARCH64 == e_machine) ? R_AARCH64_RELATIVE
                     : (Elf64_Ehdr::EM_X86_64  == e_machine) ? R_X86_64_RELATIVE
                     : (Elf64_Ehdr::EM_PPC64   == e_machine) ? R_PPC64_RELATIVE
                     : 0;
        set_te64(&rela.r_info, ELF64_R_INFO(0 /*ELF64_R_SYM(r_info)*/, r_type));
        set_te64(&rela.r_addend, sz_pack2);  // entry to decompressor
        fo->rewrite(&rela, sizeof(rela));

        fo->seek(0, SEEK_END);
    }
    // NOTE: PackLinuxElf::pack3  adjusted xct_off for the extra page

    // Then compressed gaps (including debuginfo.)
    for (unsigned k = 0; k < e_phnum; ++k) {
        Extent x;
        x.size = find_LOAD_gap(phdri, k, e_phnum);
        if (x.size) {
            x.offset = get_te64(&phdri[k].p_offset) +
                       get_te64(&phdri[k].p_filesz);
            packExtent(x, nullptr, fo);
        }
    }
    // write block end marker (uncompressed size 0)
    b_info hdr; memset(&hdr, 0, sizeof(hdr));
    set_le32(&hdr.sz_cpr, UPX_MAGIC_LE32);
    fo->write(&hdr, sizeof(hdr));
    total_out = fpad4(fo, total_out);

    if (0==xct_off) { // not shared library
        set_te64(&elfout.phdr[C_BASE].p_align, ((u64_t)0) - page_mask);
        elfout.phdr[C_BASE].p_paddr = elfout.phdr[C_BASE].p_vaddr;
        elfout.phdr[C_BASE].p_offset = 0;
        u64_t abrk = getbrk(phdri, e_phnum);
        // vbase handles ET_EXEC.  FIXME: pre-linking?
        u64_t const vbase = get_te64(&elfout.phdr[C_BASE].p_vaddr);
        set_te64(&elfout.phdr[C_BASE].p_filesz, 0x1000);  // Linux kernel SIGSEGV if (0==.p_filesz)
        set_te64(&elfout.phdr[C_BASE].p_memsz, abrk - vbase);
        set_te32(&elfout.phdr[C_BASE].p_flags, Elf64_Phdr::PF_W|Elf64_Phdr::PF_R);
        set_te64(&elfout.phdr[C_TEXT].p_filesz, sz_pack2 + lsize);
        set_te64(&elfout.phdr[C_TEXT].p_memsz,  sz_pack2 + lsize);
        set_te64(&elfout.phdr[C_TEXT].p_vaddr, abrk= (page_mask & (~page_mask + abrk)));
        elfout.phdr[C_TEXT].p_paddr = elfout.phdr[C_TEXT].p_vaddr;
        set_te64(&elfout.ehdr.e_entry, abrk + get_te64(&elfout.ehdr.e_entry) - vbase);
    }
    if (0!=xct_off) { // shared library
        u64_t const cpr_entry = load_va + sz_pack2;
        set_te64(&file_image[user_init_off], cpr_entry);  // set the hook

        if (user_init_rp) { // decompressor needs hint for DT_INIT_ARRAY
            Elf64_Dyn *dynp = (Elf64_Dyn *)elf_find_dynptr(Elf64_Dyn::DT_NULL);
            set_te64(&dynp->d_val, (char *)user_init_rp - (char *)&file_image[0]);
        }

        Elf64_Phdr *const phdr0 = (Elf64_Phdr *)lowmem.subref(
                "bad e_phoff", e_phoff, e_phnum * sizeof(Elf64_Phdr));
        Elf64_Phdr *phdr = phdr0;
        upx_off_t off = fo->st_size();  // 64 bits
        so_slide = 0;
        for (unsigned j = 0; j < e_phnum; ++j, ++phdr) {
            // p_vaddr and p_paddr do not change!
            u64_t const len  = get_te64(&phdr->p_filesz);
            u64_t const ioff = get_te64(&phdri[j].p_offset);  // without asl_delta
            u64_t       align= get_te64(&phdr->p_align);
            unsigned const type = get_te32(&phdr->p_type);
            if (Elf64_Phdr::PT_INTERP==type) {
                // Rotate to highest position, so it can be lopped
                // by decrementing e_phnum.
                memcpy((unsigned char *)ibuf, phdr, sizeof(*phdr));  // extract
                memmove(phdr, 1+phdr, (e_phnum - (1+ j))*sizeof(*phdr));  // overlapping
                memcpy(&phdr[e_phnum - (1+ j)], (unsigned char *)ibuf, sizeof(*phdr));  // to top
                --phdr; --e_phnum;
                set_te16(&ehdri.e_phnum, e_phnum);
                set_te16(&((Elf64_Ehdr *)(unsigned char *)lowmem)->e_phnum, e_phnum);
                continue;
            }
            if (PT_LOAD == type) {
                if (!ioff) { // first PT_LOAD must contain everything written so far
                    set_te64(&phdr->p_filesz, sz_pack2 + lsize);  // is this correct?
                    set_te64(&phdr->p_memsz,  sz_pack2 + lsize);
                }
                else if ((xct_off - ioff) < len) { // Change length of compressed PT_LOAD.
                    set_te64(&phdr->p_filesz, total_out - ioff);  // FIXME  (sz_pack2 + lsize - ioff) ?
                    set_te64(&phdr->p_memsz,  total_out - ioff);
                    if (user_init_off < xct_off) { // MIPS puts PT_DYNAMIC here
                        // Allow for DT_INIT in a new [stolen] slot
                        unsigned off2 = user_init_off - sizeof(u64_t);
                        fo->seek(off2, SEEK_SET);
                        fo->rewrite(&file_image[off2], 2*sizeof(u64_t));
                    }
                }
                else if (xct_off < ioff) { // Slide subsequent PT_LOAD.
                    if ((1u<<12) < align
                    &&  Elf64_Ehdr::EM_X86_64 == e_machine  // FIXME: other $ARCH ?
                    ) {
                        align = 1u<<12;
                        set_te64(&phdr->p_align, align);
                    }
                    off = fpadN(fo, (-1 + align) & (ioff - off));
                    if (!so_slide) { // only once
                        so_slide = off - ((is_asl ? asl_delta : 0) + ioff);
                        //fprintf(stderr, "\nso_slide = %#x\n", (unsigned)so_slide);
                        //asl_slide_Shdrs();
                    }
                    set_te64(&phdr->p_offset, off);
                    fo->seek(off, SEEK_SET);
                    fo->write(&file_image[ioff], len);
                    off += len;
                    total_out = off;

                    if ((user_init_off - ioff) < len) {
                        fo->seek(user_init_off + so_slide, SEEK_SET);
                        u64_t word = cpr_entry;
                        set_te64(&word, cpr_entry);
                        fo->rewrite(&word, sizeof(word));
                        fo->seek(0, SEEK_END);
                    }
                }
                continue;  // all done with this PT_LOAD
            }
            if (xct_off < ioff) {
                set_te64(&phdr->p_offset, so_slide + (is_asl ? asl_delta : 0) + ioff);
            }
        }  // end each Phdr

        if (sec_arm_attr || is_asl) { // must update Shdr.sh_offset for so_slide
            // Update {DYNAMIC}.sh_offset by so_slide.
            Elf64_Shdr *shdr = (Elf64_Shdr *)lowmem.subref(  // FIXME: use shdri ?
                    "bad e_shoff", xct_off - (is_asl ? asl_delta : 0), e_shnum * sizeof(Elf64_Shdr));
            for (unsigned j = 0; j < e_shnum; ++shdr, ++j) {
                unsigned sh_type = get_te32(&shdr->sh_type);
                unsigned sh_flags = get_te64(&shdr->sh_flags);  // all SHF_ are 32-bit anyway
                unsigned sh_offset = get_te64(&shdr->sh_offset); // already asl_delta
                if (Elf64_Shdr::SHF_ALLOC & sh_flags
                &&  Elf64_Shdr::SHF_WRITE & sh_flags
                &&  xct_off < sh_offset) {
                    set_te64(&shdr->sh_offset, so_slide + sh_offset);
                }
                if (Elf64_Shdr::SHT_RELA == sh_type
                &&  n_jmp_slot  // FIXME: does this apply to SHT_RELA ?
                &&  !strcmp(".rel.plt", get_te32(&shdr->sh_name) + shstrtab)) {
                    u64_t va = elf_unsigned_dynamic(Elf64_Dyn::DT_PLTGOT) - (is_asl ? asl_delta : 0);
                    // Now use the old Phdrs (phdri)
                    Elf64_Phdr const *phva;
                    phva = elf_find_Phdr_for_va(va, phdri, e_phnum);
                    u64_t old_off = (va - get_te64(&phva->p_vaddr))
                        + get_te64(&phva->p_offset);

                    // Now use the new Phdrs (phdr0)
                    va += (is_asl ? asl_delta : 0);
                    phva = elf_find_Phdr_for_va(va, phdr0, e_phnum);
                    u64_t new_off = (va - get_te64(&phva->p_vaddr))
                        + get_te64(&phva->p_offset);

                    if (fo && n_jmp_slot) {
                        fo->seek(new_off, SEEK_SET);
                        fo->rewrite(&file_image[old_off], n_jmp_slot * 8);
                    }
                }
                if (j && shdr->sh_addr == 0
                &&  get_te64_32(&shdr->sh_offset) < xct_off) { // UPX_RSIZE_MAX_MEM protects us
                    // Try to be nice by sliding; but still fails if compressed.
                    // So don't do it unless appending plain text of shstrtab.
                    unsigned sh_off = get_te64_32(&shdr->sh_offset);  // UPX_RSIZE_MAX_MEM protects us
                    if (xct_off < sh_off) {
                        set_te64(&shdr->sh_offset, sh_off + so_slide);
                    }
                }
            }
            // Maybe: append plain text of shstrtab strings?
            fo->seek(total_out, SEEK_SET);
            if (xct_off < e_shoff) {
                set_te32(&((Elf32_Ehdr *)lowmem.getVoidPtr())->e_shoff, total_out);
                if (fo) {
                    fo->write(shdri, e_shnum * sizeof(*shdr));
                    total_out += e_shnum * sizeof(*shdr);
                }
            }
        }
        else { // output has no Shdr
            ehdri.e_shnum = 0;
            ehdri.e_shoff = 0;
            ehdri.e_shstrndx = 0;
        }
    }
    return total_out;
}

void
PackLinuxElf::addStubEntrySections(Filter const *, unsigned m_decompr)
{
    (void)m_decompr;  // FIXME
    if (hasLoaderSection("ELFMAINX")) {
        addLoader("ELFMAINX", nullptr);
    }
    if (hasLoaderSection("ELFMAINXu")) {
            // brk() trouble if static
        addLoader("ELFMAINXu", nullptr);
    }
    addLoader(
        ( M_IS_NRV2E(ph_forced_method(ph.method)) ? "NRV_HEAD,NRV2E,NRV_TAIL"
        : M_IS_NRV2D(ph_forced_method(ph.method)) ? "NRV_HEAD,NRV2D,NRV_TAIL"
        : M_IS_NRV2B(ph_forced_method(ph.method)) ? "NRV_HEAD,NRV2B,NRV_TAIL"
        : M_IS_LZMA(ph_forced_method(ph.method))  ? "LZMA_ELF00,LZMA_DEC20,LZMA_DEC30"
        : nullptr), nullptr);
    if (hasLoaderSection("CFLUSH"))
        addLoader("CFLUSH");
    addLoader("ELFMAINY,IDENTSTR", nullptr);
    if (hasLoaderSection("ELFMAINZe")) { // ppc64 big-endian only
        addLoader("ELFMAINZe", nullptr);
    }
    addLoader("+40,ELFMAINZ", nullptr);
    if (hasLoaderSection("ANDMAJNZ")) { // Android trouble with args to DT_INIT
        if (is_asl) {
            addLoader("ANDMAJNZ", nullptr);  // constant PAGE_SIZE
        }
        else {
            addLoader("ELFMAJNZ", nullptr);  // PAGE_SIZE from AT_PAGESZ
        }
        addLoader("ELFMAKNZ", nullptr);
    }
    if (hasLoaderSection("ELFMAINZu")) {
        addLoader("ELFMAINZu", nullptr);
    }
    addLoader("FOLDEXEC", nullptr);
}


void PackLinuxElf::defineSymbols(Filter const *)
{
    linker->defineSymbol("O_BINFO", o_binfo
       | ((!!opt->o_unix.is_ptinterp)     << 0)
       | ((!!opt->o_unix.unmap_all_pages) << 1) );
}

void PackLinuxElf32::defineSymbols(Filter const *ft)
{
    PackLinuxElf::defineSymbols(ft);
}

void PackLinuxElf64::defineSymbols(Filter const *ft)
{
    PackLinuxElf::defineSymbols(ft);
}

PackLinuxElf32::PackLinuxElf32(InputFile *f)
    : super(f), phdri(nullptr), shdri(nullptr),
    n_phdrx(0), sz_phdrx(0),
    page_mask(~0u<<lg2_page),
    dynseg(nullptr), hashtab(nullptr), hashend(nullptr),
                     gashtab(nullptr), gashend(nullptr), dynsym(nullptr),
    jni_onload_sym(nullptr),
    sec_strndx(nullptr), sec_dynsym(nullptr), sec_dynstr(nullptr)
    , sec_arm_attr(nullptr)
{
    memset(&ehdri, 0, sizeof(ehdri));
    n_jmp_slot = 0;
    if (f) {
        f->seek(0, SEEK_SET);
        f->readx(&ehdri, sizeof(ehdri));
    }
}

PackLinuxElf32::~PackLinuxElf32()
{
}

void PackLinuxElf32::add_phdrx(Elf32_Phdr const *phdr)
{
    if (END_PHDRX <= n_phdrx) {
        throwCantPack("too many Phdr %u", (unsigned)(phdr - phdri));
    }
    phdrx[n_phdrx++] = phdr;
}

PackLinuxElf64::PackLinuxElf64(InputFile *f)
    : super(f), phdri(nullptr), shdri(nullptr),
    n_phdrx(0), sz_phdrx(0),
    page_mask(~0ull<<lg2_page),
    dynseg(nullptr), hashtab(nullptr), hashend(nullptr),
                     gashtab(nullptr), gashend(nullptr), dynsym(nullptr),
    jni_onload_sym(nullptr),
    sec_strndx(nullptr), sec_dynsym(nullptr), sec_dynstr(nullptr)
    , sec_arm_attr(nullptr)
{
    memset(&ehdri, 0, sizeof(ehdri));
    n_jmp_slot = 0;
    if (f) {
        f->seek(0, SEEK_SET);
        f->readx(&ehdri, sizeof(ehdri));
    }
}

PackLinuxElf64::~PackLinuxElf64()
{
}

void PackLinuxElf64::add_phdrx(Elf64_Phdr const *phdr)
{
    if (END_PHDRX <= n_phdrx) {
        throwCantPack("too many Phdr %u", (unsigned)(phdr - phdri));
    }
    phdrx[n_phdrx++] = phdr;
}

// FIXME: should be templated with PackLinuxElf32help1
void
PackLinuxElf64::PackLinuxElf64help1(InputFile *f)
{
    e_type  = get_te16(&ehdri.e_type);
    e_phnum = get_te16(&ehdri.e_phnum);
    e_shnum = get_te16(&ehdri.e_shnum);
    unsigned const e_phentsize = get_te16(&ehdri.e_phentsize);
    if (memcmp((char const *)&ehdri, "\x7f\x45\x4c\x46", 4)  // "\177ELF"
    || ehdri.e_ident[Elf64_Ehdr::EI_CLASS]!=Elf64_Ehdr::ELFCLASS64
    || sizeof(Elf64_Phdr) != e_phentsize
    || (Elf64_Ehdr::ELFDATA2MSB == ehdri.e_ident[Elf64_Ehdr::EI_DATA]
            && &N_BELE_RTP::be_policy != bele)
    || (Elf64_Ehdr::ELFDATA2LSB == ehdri.e_ident[Elf64_Ehdr::EI_DATA]
            && &N_BELE_RTP::le_policy != bele)) {
        e_phoff = 0;
        e_shoff = 0;
        sz_phdrs = 0;
        return;
    }
    if (0==e_phnum) throwCantUnpack("0==e_phnum");
    e_phoff = get_te64(&ehdri.e_phoff);
    upx_uint64_t const last_Phdr = e_phoff + e_phnum * sizeof(Elf64_Phdr);
    if (last_Phdr < e_phoff  // wrap-around
    ||  e_phoff != sizeof(Elf64_Ehdr)  // must be contiguous
    ||  (unsigned long)file_size < last_Phdr) {
        throwCantUnpack("bad e_phoff %p", (void *)e_phoff);
    }
    e_shoff = get_te64(&ehdri.e_shoff);
    upx_uint64_t const last_Shdr = e_shoff + e_shnum * sizeof(Elf64_Shdr);
    if (last_Shdr < e_shoff  // wrap-around
    ||  (e_shnum && e_shoff < last_Phdr)
    ||  (unsigned long)file_size < last_Shdr) {
        if (opt->cmd == CMD_COMPRESS) {
            throwCantUnpack("bad e_shoff %p", (void *)e_shoff);
        }
    }
    sz_phdrs = e_phnum * e_phentsize;
    sz_elf_hdrs = sz_phdrs + sizeof(Elf64_Ehdr) +
        n_phdrx * sizeof(Elf32_Phdr);  // phdrx bodies later: ::generateElfHdr

    if (f && Elf64_Ehdr::ET_DYN!=e_type) {
        unsigned const len = file_size;  // (sz_phdrs + e_phoff) except --preserve-build-id
        alloc_file_image(file_image, len);
        f->seek(0, SEEK_SET);
        f->readx(file_image, len);
        phdri= (Elf64_Phdr       *)(e_phoff + file_image);  // do not free() !!
    }
    if (f && Elf64_Ehdr::ET_DYN==e_type) {
        // The DT_SYMTAB has no designated length.  Read the whole file.
        alloc_file_image(file_image, file_size);
        f->seek(0, SEEK_SET);
        f->readx(file_image, file_size);
        phdri= (file_size <= (unsigned)e_phoff) ? nullptr : (Elf64_Phdr *)(e_phoff + file_image);  // do not free() !!
        if (!(opt->cmd == CMD_COMPRESS && e_shoff < (upx_uint64_t)file_size && mb_shdr.getSize() == 0)) {
            shdri = nullptr;
        }
        else if (e_shnum && e_shoff && ehdri.e_shentsize) {
            fi->seek(e_shoff, SEEK_SET);
            mb_shdr.alloc(   sizeof(Elf64_Shdr) * e_shnum);
            shdri = (Elf64_Shdr *)mb_shdr.getVoidPtr();
            fi->readx(shdri, sizeof(Elf64_Shdr) * e_shnum);
        }
        else {
            shdri = nullptr;
        }
        sec_dynsym = elf_find_section_type(Elf64_Shdr::SHT_DYNSYM);
        if (sec_dynsym) {
            unsigned t = get_te32(&sec_dynsym->sh_link);
            if (e_shnum <= t)
                throwCantPack("bad dynsym->sh_link");
            sec_dynstr = &shdri[t];
        }

        Elf64_Phdr const *phdr= phdri;
        for (int j = e_phnum; --j>=0; ++phdr)
        if (Elf64_Phdr::PT_DYNAMIC==get_te32(&phdr->p_type)) {
            upx_uint64_t offset = check_pt_dynamic(phdr);
            dynseg= (Elf64_Dyn *)(offset + file_image);
            invert_pt_dynamic(dynseg,
                umin(get_te64(&phdr->p_filesz), file_size_u - offset));
        }
        else if (is_LOAD(phdr)) {
            check_pt_load(phdr);
        }
        // elf_find_dynamic() returns 0 if 0==dynseg.
        dynstr =          (char const *)elf_find_dynamic(Elf64_Dyn::DT_STRTAB);
        dynsym = (Elf64_Sym /*const*/ *)elf_find_dynamic(Elf64_Dyn::DT_SYMTAB);
        gashtab =     (unsigned const *)elf_find_dynamic(Elf64_Dyn::DT_GNU_HASH);
        hashtab =     (unsigned const *)elf_find_dynamic(Elf64_Dyn::DT_HASH);
        if (3& ((upx_uintptr_t)dynsym | (upx_uintptr_t)gashtab | (upx_uintptr_t)hashtab)) {
            throwCantPack("unaligned DT_SYMTAB, DT_GNU_HASH, or DT_HASH/n");
        }
        jni_onload_sym = elf_lookup("JNI_OnLoad");
        if (jni_onload_sym) {
            jni_onload_va = get_te64(&jni_onload_sym->st_value);
            jni_onload_va = 0;  // FIXME not understood; need example
        }
    }
}

Linker* PackLinuxElf64amd::newLinker() const
{
    return new ElfLinkerAMD64;
}

Linker* PackLinuxElf64arm::newLinker() const
{
    return new ElfLinkerArm64LE;
}

int const *
PackLinuxElf::getCompressionMethods(int method, int level) const
{
    // No real dependency on LE32.
    return Packer::getDefaultCompressionMethods_le32(method, level);
}

int const *
PackLinuxElf32armLe::getCompressionMethods(int method, int level) const
{
    return Packer::getDefaultCompressionMethods_8(method, level);
}

int const *
PackLinuxElf32armBe::getCompressionMethods(int method, int level) const
{
    return Packer::getDefaultCompressionMethods_8(method, level);
}

int const *
PackLinuxElf32ppc::getFilters() const
{
    static const int filters[] = {
        0xd0,
    FT_END };
    return filters;
}

int const *
PackLinuxElf64ppcle::getFilters() const
{
    static const int filters[] = {
        0xd0,
    FT_END };
    return filters;
}

int const *
PackLinuxElf64ppc::getFilters() const
{
    static const int filters[] = {
        0xd0,
    FT_END };
    return filters;
}

int const *
PackLinuxElf64amd::getFilters() const
{
    static const int filters[] = {
        0x49,
    FT_END };
    return filters;
}

int const *
PackLinuxElf64arm::getFilters() const
{
    static const int filters[] = {
        0x52,
    FT_END };
    return filters;
}

void PackLinuxElf32::patchLoader()
{
}

void PackLinuxElf64::patchLoader()
{
}

void PackLinuxElf32::ARM_updateLoader(OutputFile * /*fo*/)
{
    set_te32(&elfout.ehdr.e_entry, sz_pack2 +
        linker->getSymbolOffset("_start") +
        get_te32(&elfout.phdr[C_TEXT].p_vaddr));
}

void PackLinuxElf32armLe::updateLoader(OutputFile *fo)
{
    ARM_updateLoader(fo);
}

void PackLinuxElf32armBe::updateLoader(OutputFile *fo)
{
    ARM_updateLoader(fo);
}

void PackLinuxElf32mipsel::updateLoader(OutputFile *fo)
{
    ARM_updateLoader(fo);  // not ARM specific; (no 32-bit immediates)
}

void PackLinuxElf32mipseb::updateLoader(OutputFile *fo)
{
    ARM_updateLoader(fo);  // not ARM specific; (no 32-bit immediates)
}

void PackLinuxElf32::updateLoader(OutputFile * /*fo*/)
{
    unsigned start = linker->getSymbolOffset("_start");
    unsigned vbase = get_te32(&elfout.phdr[C_TEXT].p_vaddr);
    set_te32(&elfout.ehdr.e_entry, start + sz_pack2 + vbase);
}

void PackLinuxElf64::updateLoader(OutputFile * /*fo*/)
{
    if (xct_off) {
        return;  // FIXME elfout has no values at all
    }
    upx_uint64_t const vbase = get_te64(&elfout.phdr[C_TEXT].p_vaddr);
    unsigned start = linker->getSymbolOffset("_start");

    if (get_te16(&elfout.ehdr.e_machine)==Elf64_Ehdr::EM_PPC64
    &&  elfout.ehdr.e_ident[Elf64_Ehdr::EI_DATA]==Elf64_Ehdr::ELFDATA2MSB) {
        unsigned descr = linker->getSymbolOffset("entry_descr");

        // External relocation of PPC64 function descriptor.
        upx_uint64_t dot_entry = start + sz_pack2 + vbase;
        upx_byte *p = getLoader();

        set_te64(&p[descr], dot_entry);
        // Kernel 3.16.0 (2017-09-19) uses start, not descr
        set_te64(&elfout.ehdr.e_entry, start + sz_pack2 + vbase);
    }
    else {
        set_te64(&elfout.ehdr.e_entry, start + sz_pack2 + vbase);
    }
}

PackLinuxElf32ppc::PackLinuxElf32ppc(InputFile *f)
    : super(f)
{
    e_machine = Elf32_Ehdr::EM_PPC;
    ei_class  = Elf32_Ehdr::ELFCLASS32;
    ei_data   = Elf32_Ehdr::ELFDATA2MSB;
    ei_osabi  = Elf32_Ehdr::ELFOSABI_LINUX;
}

PackLinuxElf32ppc::~PackLinuxElf32ppc()
{
}

Linker* PackLinuxElf32ppc::newLinker() const
{
    return new ElfLinkerPpc32;
}

PackLinuxElf64ppcle::PackLinuxElf64ppcle(InputFile *f)
    : super(f), lg2_page(16), page_size(1u<<lg2_page)
{
    e_machine = Elf64_Ehdr::EM_PPC64;
    ei_class  = Elf64_Ehdr::ELFCLASS64;
    ei_data   = Elf64_Ehdr::ELFDATA2LSB;
    ei_osabi  = Elf64_Ehdr::ELFOSABI_LINUX;
}

PackLinuxElf64ppc::PackLinuxElf64ppc(InputFile *f)
    : super(f), lg2_page(16), page_size(1u<<lg2_page)
{
    e_machine = Elf64_Ehdr::EM_PPC64;
    ei_class  = Elf64_Ehdr::ELFCLASS64;
    ei_data   = Elf64_Ehdr::ELFDATA2MSB;
    ei_osabi  = Elf32_Ehdr::ELFOSABI_LINUX;
}

PackLinuxElf64ppcle::~PackLinuxElf64ppcle()
{
}

PackLinuxElf64ppc::~PackLinuxElf64ppc()
{
}

Linker* PackLinuxElf64ppcle::newLinker() const
{
    return new ElfLinkerPpc64le;
}

Linker* PackLinuxElf64ppc::newLinker() const
{
    return new ElfLinkerPpc64;
}

PackLinuxElf64amd::PackLinuxElf64amd(InputFile *f)
    : super(f)
{
    // Why did PackLinuxElf64Le set lg2_page = 16 ?
    // It causes trouble for check_pt_dynamic() from canPack().
    lg2_page = 12;  page_size = 1u<<lg2_page;
    e_machine = Elf64_Ehdr::EM_X86_64;
    ei_class = Elf64_Ehdr::ELFCLASS64;
    ei_data = Elf64_Ehdr::ELFDATA2LSB;
    ei_osabi  = Elf32_Ehdr::ELFOSABI_LINUX;
}

PackLinuxElf64arm::PackLinuxElf64arm(InputFile *f)
    : super(f)
{
    e_machine = Elf64_Ehdr::EM_AARCH64;
    ei_class = Elf64_Ehdr::ELFCLASS64;
    ei_data = Elf64_Ehdr::ELFDATA2LSB;
    ei_osabi  = Elf32_Ehdr::ELFOSABI_LINUX;
}

PackLinuxElf64amd::~PackLinuxElf64amd()
{
}

PackLinuxElf64arm::~PackLinuxElf64arm()
{
}

void PackLinuxElf32x86::addStubEntrySections(Filter const *ft, unsigned m_decompr)
{
    (void)m_decompr;  // FIXME
    int const n_mru = ft->n_mru;  // FIXME: belongs to filter? packerf?

// Rely on "+80CXXXX" [etc] in getDecompressorSections() packer_c.cpp */
//    // Here is a quick summary of the format of the output file:
//    linker->setLoaderAlignOffset(
//            // Elf32_Ehdr
//        sizeof(elfout.ehdr) +
//            // Elf32_Phdr: 1 for exec86, 2 for sh86, 3 for elf86
//        (get_te16(&elfout.ehdr.e_phentsize) * get_te16(&elfout.ehdr.e_phnum)) +
//            // checksum UPX! lsize version format
//        sizeof(l_info) +
//            // PT_DYNAMIC with DT_NEEDED "forwarded" from original file
//        ((get_te16(&elfout.ehdr.e_phnum)==3)
//            ? (unsigned) get_te32(&elfout.phdr[C_NOTE].p_memsz)
//            : 0) +
//            // p_progid, p_filesize, p_blocksize
//        sizeof(p_info) +
//            // compressed data
//        b_len + ph.c_len );

            // entry to stub
    addLoader("LEXEC000", nullptr);

    if (ft->id) {
        { // decompr, unfilter are separate
            addLoader("LXUNF000", nullptr);
            addLoader("LXUNF002", nullptr);
                if (0x80==(ft->id & 0xF0)) {
                    if (256==n_mru) {
                        addLoader("MRUBYTE0", nullptr);
                    }
                    else if (n_mru) {
                        addLoader("LXMRU005", nullptr);
                    }
                    if (n_mru) {
                        addLoader("LXMRU006", nullptr);
                    }
                    else {
                        addLoader("LXMRU007", nullptr);
                    }
            }
            else if (0x40==(ft->id & 0xF0)) {
                addLoader("LXUNF008", nullptr);
            }
            addLoader("LXUNF010", nullptr);
        }
        if (n_mru) {
            addLoader("LEXEC009", nullptr);
        }
    }
    addLoader("LEXEC010", nullptr);
    addLoader(getDecompressorSections(), nullptr);
    addLoader("LEXEC015", nullptr);
    if (ft->id) {
        {  // decompr, unfilter are separate
            if (0x80!=(ft->id & 0xF0)) {
                addLoader("LXUNF042", nullptr);
            }
        }
        addFilter32(ft->id);
        { // decompr, unfilter are separate
            if (0x80==(ft->id & 0xF0)) {
                if (0==n_mru) {
                    addLoader("LXMRU058", nullptr);
                }
            }
            addLoader("LXUNF035", nullptr);
        }
    }
    else {
        addLoader("LEXEC017", nullptr);
    }

    addLoader("IDENTSTR", nullptr);
    addLoader("+40,LEXEC020", nullptr);
    addLoader("FOLDEXEC", nullptr);
}

void PackLinuxElf32x86::defineSymbols(Filter const *const ft)
{
    PackLinuxElf32::defineSymbols(ft);

    if (0x80==(ft->id & 0xF0)) {
        int const mru = ft->n_mru ? 1+ ft->n_mru : 0;
        if (mru && mru!=256) {
            unsigned const is_pwr2 = (0==((mru -1) & mru));
            linker->defineSymbol("NMRU", mru - is_pwr2);
        }
    }
}

void
PackLinuxElf32::buildLinuxLoader(
    upx_byte const *const proto,
    unsigned        const szproto,
    upx_byte const *const fold,
    unsigned        const szfold,
    Filter const *ft
)
{
    MemBuffer mb_cprLoader;
    unsigned sz_cpr = 0;
    unsigned sz_unc = 0;
    unsigned method = 0;
    upx_byte const *uncLoader = nullptr;

  if (0 < szfold) {
    if (xct_off // shlib
      && (  this->e_machine==Elf32_Ehdr::EM_ARM
         || this->e_machine==Elf32_Ehdr::EM_386)
    ) {
        initLoader(fold, szfold);
// Typical layout of 'sections' in compressed stub code for shared library:
//   SO_HEAD
//   ptr_NEXT
//   EXP_HEAD  NRV getbit(), copy
//   NRV2B etc: daisy chain of de-compressor for each method used
//   EXP_TAIL  FIXME: unfilter
//   SO_TAIL
//   SO_MAIN  C-language supervision based on PT_LOADs
        char sec[200]; memset(sec, 0, sizeof(sec));  // debug convenience
        int len = 0;
        unsigned m_decompr = methods_used | (1u << (0xFF & ph_forced_method(ph.method)));
        len += snprintf(sec, sizeof(sec), "%s", "SO_HEAD,ptr_NEXT,EXP_HEAD");

        // Start of dasiy-chain fall-through.
        if (((1u<<M_NRV2B_LE32)|(1u<<M_NRV2B_8)|(1u<<M_NRV2B_LE16)) & m_decompr) {
            len += snprintf(&sec[len], sizeof(sec) - len, ",%s", "NRV2B");
        }
        if (((1u<<M_NRV2D_LE32)|(1u<<M_NRV2D_8)|(1u<<M_NRV2D_LE16)) & m_decompr) {
            len += snprintf(&sec[len], sizeof(sec) - len, ",%s", "NRV2D");
        }
        if (((1u<<M_NRV2E_LE32)|(1u<<M_NRV2E_8)|(1u<<M_NRV2E_LE16)) & m_decompr) {
            len += snprintf(&sec[len], sizeof(sec) - len, ",%s", "NRV2E");
        }
        if (((1u<<M_LZMA)) & m_decompr) {
            len += snprintf(&sec[len], sizeof(sec) - len, ",%s",
                "LZMA_DAISY,LZMA_ELF00,LZMA_DEC20,LZMA_DEC30");
        }
        len += snprintf(&sec[len], sizeof(sec) - len, ",%s", "EXP_TAIL");
        // End of daisy-chain fall-through.

        // MIPS directly calls memfd_create
        if (this->e_machine != Elf32_Ehdr::EM_MIPS) {
            len += snprintf(&sec[len], sizeof(sec) - len, ",%s",
                (sec_arm_attr || is_asl)
                    ? "HUMF_A,UMF_ANDROID"
                    : "HUMF_L,UMF_LINUX");
        }
        if (hasLoaderSection("STRCON")) {
            len += snprintf(&sec[len], sizeof(sec) - len, ",%s", "STRCON");
        }
        len += snprintf(&sec[len], sizeof(sec) - len, ",%s", "SO_TAIL,SO_MAIN");
        (void)len;  // Pacify the anal-retentive static analyzer which hates a good idiom.
        NO_printf("\n%s\n", sec);
        addLoader(sec, nullptr);
        relocateLoader();
        {
            int sz_unc_int;
            uncLoader = linker->getLoader(&sz_unc_int);
            sz_unc = sz_unc_int;
        }
        method = M_NRV2B_LE32;  // requires unaligned fetch
        if (this->e_machine==Elf32_Ehdr::EM_ARM)
            method = M_NRV2B_8;  //only ARM v6 and above has unaligned fetch
    } // end shlib (folded portion)
    else if (this->e_machine==Elf32_Ehdr::EM_386
         ||  this->e_machine==Elf32_Ehdr::EM_ARM
         ||  this->e_machine==Elf32_Ehdr::EM_PPC
         ||  this->e_machine==Elf32_Ehdr::EM_MIPS
         ) { // main program with ELF2 de-compressor (folded portion)
        initLoader(fold, szfold);
        char sec[200]; memset(sec, 0, sizeof(sec));  // debug convenience
        int len = 0;
        unsigned m_decompr = methods_used | (1u << (0xFF & ph_forced_method(ph.method)));
        len += snprintf(sec, sizeof(sec), "%s", ".text,EXP_HEAD");
        if (((1u<<M_NRV2B_LE32)|(1u<<M_NRV2B_8)|(1u<<M_NRV2B_LE16)) & m_decompr) {
            len += snprintf(&sec[len], sizeof(sec) - len, ",%s", "NRV2B");
        }
        if (((1u<<M_NRV2D_LE32)|(1u<<M_NRV2D_8)|(1u<<M_NRV2D_LE16)) & m_decompr) {
            len += snprintf(&sec[len], sizeof(sec) - len, ",%s", "NRV2D");
        }
        if (((1u<<M_NRV2E_LE32)|(1u<<M_NRV2E_8)|(1u<<M_NRV2E_LE16)) & m_decompr) {
            len += snprintf(&sec[len], sizeof(sec) - len, ",%s", "NRV2E");
        }
        if (((1u<<M_LZMA)) & m_decompr) {
            len += snprintf(&sec[len], sizeof(sec) - len, ",%s",
                "LZMA_DAISY,LZMA_ELF00,LZMA_DEC20,LZMA_DEC30");
        }
        len += snprintf(&sec[len], sizeof(sec) - len, ",%s", "EXP_TAIL");

        // $ARCH-linux.elf-main2.c calls upx_mmap_and_fd, not direct memfd_create
        len += snprintf(&sec[len], sizeof(sec) - len, ",%s",
            (sec_arm_attr || is_asl)
                ? "HUMF_A,UMF_ANDROID"
                : "HUMF_L,UMF_LINUX");
        if (hasLoaderSection("SYSCALLS")) {
            len += snprintf(&sec[len], sizeof(sec) - len, ",%s", "SYSCALLS");
        }
        (void)len;  // Pacify the anal-retentive static analyzer which hates a good idiom.
        NO_printf("\n%s\n", sec);
        addLoader(sec, nullptr);
        relocateLoader();
        {
            int sz_unc_int;
            uncLoader = linker->getLoader(&sz_unc_int);
            sz_unc = sz_unc_int;
        }
        method = M_NRV2B_LE32;  // requires unaligned fetch
        if (this->e_machine==Elf32_Ehdr::EM_ARM)
            method = M_NRV2B_8;  //only ARM v6 and above has unaligned fetch
    }
    else { // main program with ELF1 de-compressor (folded portion)
        cprElfHdr1 const *const hf = (cprElfHdr1 const *)fold;
        unsigned fold_hdrlen = usizeof(hf->ehdr) +
            get_te16(&hf->ehdr.e_phentsize) * get_te16(&hf->ehdr.e_phnum);
        if (this->e_machine==Elf32_Ehdr::EM_MIPS) {
            fold_hdrlen = upx::umax(fold_hdrlen, (unsigned)0x80);
        }
        uncLoader = fold_hdrlen + fold;
        sz_unc = ((szfold < fold_hdrlen) ? 0 : (szfold - fold_hdrlen));
        method = ph.method;
    }

    struct b_info h; memset(&h, 0, sizeof(h));
    h.b_method = method;
    // _Ehdr and _Phdr are NOT filtered, so Leave h.b_ftid and h.b_cto8 as zero.

    mb_cprLoader.allocForCompression(sizeof(h) + sz_unc);
    unsigned char *const cprLoader = (unsigned char *)mb_cprLoader;  // less typing

    h.sz_unc = sz_unc;
    h.sz_cpr = mb_cprLoader.getSize();  // max that upx_compress may use
    {
        int r = upx_compress(uncLoader, sz_unc, sizeof(h) + cprLoader, &sz_cpr,
            nullptr, ph_forced_method(method), 10, nullptr, nullptr );
        h.sz_cpr = sz_cpr;  // actual length used
        if (r != UPX_E_OK || h.sz_cpr >= h.sz_unc)
            throwInternalError("loader compression failed");
    }
    set_te32(&h.sz_cpr, h.sz_cpr);
    set_te32(&h.sz_unc, h.sz_unc);
    memcpy(cprLoader, &h, sizeof(h)); // cprLoader will become FOLDEXEC
  }  // end (0 < szfold)

    initLoader(proto, szproto, -1, sz_cpr);
    NO_printf("FOLDEXEC unc=%#x  cpr=%#x\n", sz_unc, sz_cpr);
    linker->addSection("FOLDEXEC", mb_cprLoader, sizeof(b_info) + sz_cpr, 0);
    if (xct_off  // shlib
       && (this->e_machine==Elf32_Ehdr::EM_NONE
          || this->e_machine==Elf32_Ehdr::EM_386
          || this->e_machine==Elf32_Ehdr::EM_ARM
          || this->e_machine==Elf32_Ehdr::EM_PPC
          || this->e_machine==Elf32_Ehdr::EM_MIPS
          )
    ) { // shlib with ELF2 de-compressor
        addLoader("ELFMAINX");
        addLoader((sec_arm_attr || is_asl)
            ? "HUMF_A,UMF_ANDROID"
            : "HUMF_L,UMF_LINUX");
        addLoader("ELFMAINZ,FOLDEXEC,IDENTSTR");
    }
    else if (this->e_machine==Elf32_Ehdr::EM_NONE
          || this->e_machine==Elf32_Ehdr::EM_386
          || this->e_machine==Elf32_Ehdr::EM_ARM
          || this->e_machine==Elf32_Ehdr::EM_PPC
          || this->e_machine==Elf32_Ehdr::EM_MIPS
      ) { // main program with ELF2 de-compressor
        addLoader("ELFMAINX");
        if (this->e_machine==Elf32_Ehdr::EM_ARM) { // hardware is more problematic
            if (opt->o_unix.catch_sigsegv && hasLoaderSection("ELFSIGSEGV"))
                addLoader("ELFSIGSEGV");
            addLoader("ELFMAINX2");
        }
        // Only if $ARCH-linux.elf-entry.S calls upx_mmap_and_fd instead of memfd_create
        if (this->e_machine != Elf32_Ehdr::EM_PPC
        &&  this->e_machine != Elf32_Ehdr::EM_MIPS)
            addLoader((sec_arm_attr || is_asl)
                ? "HUMF_A,UMF_ANDROID"
                : "HUMF_L,UMF_LINUX");
        addLoader("ELFMAINZ,FOLDEXEC,IDENTSTR");
            defineSymbols(ft);
    }
    else { // main program with ELF1 de-compressor
        addStubEntrySections(ft, methods_used | (1u << (0xFF & ph_forced_method(ph.method))) );
        if (!xct_off) { // main program
            defineSymbols(ft);
        }
    }
    relocateLoader();
}

void
PackLinuxElf64::buildLinuxLoader(
    upx_byte const *const proto,
    unsigned        const szproto,
    upx_byte const *const fold,
    unsigned        const szfold,
    Filter const *ft
)
{
    MemBuffer mb_cprLoader;
    unsigned sz_cpr = 0;
    unsigned sz_unc = 0;
    unsigned method = 0;
    upx_byte const *uncLoader = nullptr;

  if (0 < szfold) {
    if (xct_off // shlib
      && (  this->e_machine==Elf64_Ehdr::EM_X86_64
         || this->e_machine==Elf64_Ehdr::EM_AARCH64)
    ) {
        initLoader(fold, szfold);
// Typical layout of 'sections' in compressed stub code for shared library:
//   SO_HEAD
//   ptr_NEXT
//   EXP_HEAD  NRV getbit(), copy
//   NRV2B etc: daisy chain of de-compressor for each method used
//   EXP_TAIL  FIXME: unfilter
//   SO_TAIL
//   SO_MAIN  C-language supervision based on PT_LOADs
        char sec[200]; memset(sec, 0, sizeof(sec));  // debug convenience
        int len = 0;
        unsigned m_decompr = methods_used | (1u << (0xFF & ph_forced_method(ph.method)));
        len += snprintf(sec, sizeof(sec), "%s", "SO_HEAD,ptr_NEXT,EXP_HEAD");

        // Start of dasiy-chain fall-through.
        if (((1u<<M_NRV2B_LE32)|(1u<<M_NRV2B_8)|(1u<<M_NRV2B_LE16)) & m_decompr) {
            len += snprintf(&sec[len], sizeof(sec) - len, ",%s", "NRV2B");
        }
        if (((1u<<M_NRV2D_LE32)|(1u<<M_NRV2D_8)|(1u<<M_NRV2D_LE16)) & m_decompr) {
            len += snprintf(&sec[len], sizeof(sec) - len, ",%s", "NRV2D");
        }
        if (((1u<<M_NRV2E_LE32)|(1u<<M_NRV2E_8)|(1u<<M_NRV2E_LE16)) & m_decompr) {
            len += snprintf(&sec[len], sizeof(sec) - len, ",%s", "NRV2E");
        }
        if (((1u<<M_LZMA)) & m_decompr) {
            len += snprintf(&sec[len], sizeof(sec) - len, ",%s",
                "LZMA_DAISY,LZMA_ELF00,LZMA_DEC20,LZMA_DEC30");
        }
        len += snprintf(&sec[len], sizeof(sec) - len, ",%s", "EXP_TAIL");
        // End of daisy-chain fall-through.

        // Android on EM_AARCH64 has memfd_create(), so UMF_ANDROID not needed.
        len += snprintf(&sec[len], sizeof(sec) - len, ",%s", "HUMF_L,UMF_LINUX");
        if (hasLoaderSection("STRCON")) {
            len += snprintf(&sec[len], sizeof(sec) - len, ",%s", "STRCON");
        }
        len += snprintf(&sec[len], sizeof(sec) - len, ",%s", "SO_TAIL,SO_MAIN");
        (void)len;  // Pacify the anal-retentive static analyzer which hates a good idiom.
        NO_printf("\n%s\n", sec);
        addLoader(sec, nullptr);
        relocateLoader();
        {
            int sz_unc_int;
            uncLoader = linker->getLoader(&sz_unc_int);
            sz_unc = sz_unc_int;
        }
        method = M_NRV2B_LE32;  // requires unaligned fetch
    }
    else if (this->e_machine==Elf32_Ehdr::EM_NONE
         ||  this->e_machine==Elf64_Ehdr::EM_X86_64
         ||  this->e_machine==Elf64_Ehdr::EM_AARCH64
         ||  this->e_machine==Elf64_Ehdr::EM_PPC64
         ) { // main program with ELF2 de-compressor (folded portion)
        initLoader(fold, szfold);
        char sec[200]; memset(sec, 0, sizeof(sec));  // debug convenience
        int len = 0;
        unsigned m_decompr = methods_used | (1u << (0xFF & ph_forced_method(ph.method)));
        len += snprintf(sec, sizeof(sec), "%s", ".text,EXP_HEAD");
        if (((1u<<M_NRV2B_LE32)|(1u<<M_NRV2B_8)|(1u<<M_NRV2B_LE16)) & m_decompr) {
            len += snprintf(&sec[len], sizeof(sec) - len, ",%s", "NRV2B");
        }
        if (((1u<<M_NRV2D_LE32)|(1u<<M_NRV2D_8)|(1u<<M_NRV2D_LE16)) & m_decompr) {
            len += snprintf(&sec[len], sizeof(sec) - len, ",%s", "NRV2D");
        }
        if (((1u<<M_NRV2E_LE32)|(1u<<M_NRV2E_8)|(1u<<M_NRV2E_LE16)) & m_decompr) {
            len += snprintf(&sec[len], sizeof(sec) - len, ",%s", "NRV2E");
        }
        if (((1u<<M_LZMA)) & m_decompr) {
            len += snprintf(&sec[len], sizeof(sec) - len, ",%s",
                "LZMA_DAISY,LZMA_ELF00,LZMA_DEC20,LZMA_DEC30");
        }
        len += snprintf(&sec[len], sizeof(sec) - len, ",%s", "EXP_TAIL");
        if (hasLoaderSection("SYSCALLS")) {
            len += snprintf(&sec[len], sizeof(sec) - len, ",%s", "SYSCALLS");
        }
        if (hasLoaderSection("STRCON")) {
            len += snprintf(&sec[len], sizeof(sec) - len, ",%s", "STRCON");
        }
        (void)len;
        NO_printf("\n%s\n", sec);
        addLoader(sec, nullptr);
        relocateLoader();
        {
            int sz_unc_int;
            uncLoader = linker->getLoader(&sz_unc_int);
            sz_unc = sz_unc_int;
        }
        method = M_NRV2B_LE32;  // requires unaligned fetch
    }
    else { // not shlib: main program with ELF1 de-compressor
        cprElfHdr1 const *hf = (cprElfHdr1 const *)fold;
        e_type = get_te16(&hf->ehdr.e_type);
        if (ET_REL == e_type) {
            initLoader(fold, szfold);
            addLoader(".text", nullptr);
            relocateLoader();
            int sz_unc_int(0);
            uncLoader = linker->getLoader(&sz_unc_int);
            sz_unc = sz_unc_int;
        }
        else if (ET_EXEC == e_type) {
            hf = (cprElfHdr1 const *)fold;
            unsigned fold_hdrlen = usizeof(hf->ehdr) +
                get_te16(&hf->ehdr.e_phentsize) * get_te16(&hf->ehdr.e_phnum);
            if (this->e_machine==Elf64_Ehdr::EM_X86_64) {
                fold_hdrlen = get_te64(&hf->ehdr.e_entry);
            }
            uncLoader = &fold[fold_hdrlen];
            sz_unc = ((szfold < fold_hdrlen) ? 0 : (szfold - fold_hdrlen));
        }
        method = ph.method;
    }

    struct b_info h; memset(&h, 0, sizeof(h));
    h.b_method = method;
    // _Ehdr and _Phdr are NOT filtered, so Leave h.b_ftid and h.b_cto8 as zero.

    mb_cprLoader.allocForCompression(sizeof(h) + sz_unc);
    unsigned char *const cprLoader = (unsigned char *)mb_cprLoader;  // less typing

    h.sz_unc = sz_unc;
    h.sz_cpr = mb_cprLoader.getSize();  // max that upx_compress may use
    {
        int r = upx_compress(uncLoader, sz_unc, sizeof(h) + cprLoader, &sz_cpr,
            nullptr, ph_forced_method(method), 10, nullptr, nullptr );
        h.sz_cpr = sz_cpr;  // actual length used
        if (r != UPX_E_OK || h.sz_cpr >= h.sz_unc)
            throwInternalError("loader compression failed");
    }
    set_te32(&h.sz_cpr, h.sz_cpr);
    set_te32(&h.sz_unc, h.sz_unc);
    memcpy(cprLoader, &h, sizeof(h)); // cprLoader will become FOLDEXEC
  }  // end (0 < szfold)

    initLoader(proto, szproto, -1, sz_cpr);
    NO_printf("FOLDEXEC unc=%#x  cpr=%#x\n", sz_unc, sz_cpr);
    linker->addSection("FOLDEXEC", mb_cprLoader, sizeof(b_info) + sz_cpr, 0);
    if (xct_off
       && (this->e_machine==Elf64_Ehdr::EM_NONE
          || this->e_machine==Elf64_Ehdr::EM_X86_64
          || this->e_machine==Elf64_Ehdr::EM_AARCH64
          || this->e_machine==Elf64_Ehdr::EM_PPC64
          )
    ) {
        addLoader("ELFMAINX,ELFMAINZ,FOLDEXEC,IDENTSTR");
    } // shlib
    else if (this->e_machine==Elf32_Ehdr::EM_NONE
         ||  this->e_machine==Elf64_Ehdr::EM_X86_64
         ||  this->e_machine==Elf64_Ehdr::EM_AARCH64
         ||  this->e_machine==Elf64_Ehdr::EM_PPC64
        ) { // main program with ELF2 de-compressor
        addLoader("ELFMAINX");
        // NYI for PPC64 {
        if (opt->o_unix.catch_sigsegv && hasLoaderSection("ELFSIGSEGV"))
            addLoader("ELFSIGSEGV");
        if (this->e_machine!=Elf64_Ehdr::EM_PPC64)
            addLoader("ELFMAINX2");
        // } end NYI for PPC64
        addLoader("ELFMAINZ,FOLDEXEC,IDENTSTR");
        if (this->e_machine==Elf64_Ehdr::EM_PPC64
        &&  ehdri.e_ident[Elf64_Ehdr::EI_DATA]==Elf64_Ehdr::ELFDATA2MSB) {
            addLoader("ELFMAINZe");
        }
        if (!xct_off) { // main program
            defineSymbols(ft);
        }
    }
    else { // main program with ELF1 de-compressor
        addStubEntrySections(ft, methods_used | (1u << (0xFF & ph_forced_method(ph.method))) );
        if (!xct_off) { // main program
            defineSymbols(ft);
        }
    }
    relocateLoader();
}

void
PackLinuxElf64amd::defineSymbols(Filter const *ft)
{
    PackLinuxElf64::defineSymbols(ft);
}

static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/i386-linux.elf-entry.h"
static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/i386-linux.elf-so_entry.h"
static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/i386-linux.elf-fold.h"
static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/i386-linux.elf-so_fold.h"
static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/i386-linux.shlib-init.h"

void
PackLinuxElf32x86::buildLoader(const Filter *ft)
{
    if (0!=xct_off) {  // shared library
        buildLinuxLoader(
            stub_i386_linux_elf_so_entry, sizeof(stub_i386_linux_elf_so_entry),
            stub_i386_linux_elf_so_fold,  sizeof(stub_i386_linux_elf_so_fold), ft);
        return;
    }
    unsigned char tmp[sizeof(stub_i386_linux_elf_fold)];
    memcpy(tmp, stub_i386_linux_elf_fold, sizeof(stub_i386_linux_elf_fold));
    checkPatch(nullptr, 0, 0, 0);  // reset
    if (opt->o_unix.is_ptinterp) {
        unsigned j;
        for (j = 0; j < sizeof(stub_i386_linux_elf_fold)-1; ++j) {
            if (0x60==tmp[  j]
            &&  0x47==tmp[1+j] ) {
                /* put INC EDI before PUSHA: inhibits auxv_up for PT_INTERP */
                tmp[  j] = 0x47;
                tmp[1+j] = 0x60;
                break;
            }
        }
    }
    buildLinuxLoader(
        stub_i386_linux_elf_entry, sizeof(stub_i386_linux_elf_entry),
        tmp,                       sizeof(stub_i386_linux_elf_fold),  ft );
}

static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/i386-bsd.elf-entry.h"
static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/i386-bsd.elf-fold.h"

void
PackBSDElf32x86::buildLoader(const Filter *ft)
{
    unsigned char tmp[sizeof(stub_i386_bsd_elf_fold)];
    memcpy(tmp, stub_i386_bsd_elf_fold, sizeof(stub_i386_bsd_elf_fold));
    checkPatch(nullptr, 0, 0, 0);  // reset
    if (opt->o_unix.is_ptinterp) {
        unsigned j;
        for (j = 0; j < sizeof(stub_i386_bsd_elf_fold)-1; ++j) {
            if (0x60==tmp[  j]
            &&  0x47==tmp[1+j] ) {
                /* put INC EDI before PUSHA: inhibits auxv_up for PT_INTERP */
                tmp[  j] = 0x47;
                tmp[1+j] = 0x60;
                break;
            }
        }
    }
    buildLinuxLoader(
        stub_i386_bsd_elf_entry, sizeof(stub_i386_bsd_elf_entry),
        tmp,                     sizeof(stub_i386_bsd_elf_fold), ft);
}

static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/i386-netbsd.elf-entry.h"

static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/i386-netbsd.elf-fold.h"

#define WANT_NHDR_ENUM
#include "p_elf_enum.h"
#undef WANT_NHDR_ENUM

void
PackNetBSDElf32x86::buildLoader(const Filter *ft)
{
    unsigned char tmp[sizeof(stub_i386_netbsd_elf_fold)];
    memcpy(tmp, stub_i386_netbsd_elf_fold, sizeof(stub_i386_netbsd_elf_fold));
    checkPatch(nullptr, 0, 0, 0);  // reset
    if (opt->o_unix.is_ptinterp) {
        unsigned j;
        for (j = 0; j < sizeof(stub_i386_netbsd_elf_fold)-1; ++j) {
            if (0x60==tmp[  j]
            &&  0x47==tmp[1+j] ) {
                /* put INC EDI before PUSHA: inhibits auxv_up for PT_INTERP */
                tmp[  j] = 0x47;
                tmp[1+j] = 0x60;
                break;
            }
        }
    }
    buildLinuxLoader(
        stub_i386_netbsd_elf_entry, sizeof(stub_i386_netbsd_elf_entry),
        tmp,                        sizeof(stub_i386_netbsd_elf_fold), ft);
}

static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/i386-openbsd.elf-fold.h"

void
PackOpenBSDElf32x86::buildLoader(const Filter *ft)
{
    unsigned char tmp[sizeof(stub_i386_openbsd_elf_fold)];
    memcpy(tmp, stub_i386_openbsd_elf_fold, sizeof(stub_i386_openbsd_elf_fold));
    checkPatch(nullptr, 0, 0, 0);  // reset
    if (opt->o_unix.is_ptinterp) {
        unsigned j;
        for (j = 0; j < sizeof(stub_i386_openbsd_elf_fold)-1; ++j) {
            if (0x60==tmp[  j]
            &&  0x47==tmp[1+j] ) {
                /* put INC EDI before PUSHA: inhibits auxv_up for PT_INTERP */
                tmp[  j] = 0x47;
                tmp[1+j] = 0x60;
                break;
            }
        }
    }
    buildLinuxLoader(
        stub_i386_bsd_elf_entry, sizeof(stub_i386_bsd_elf_entry),
        tmp,                     sizeof(stub_i386_openbsd_elf_fold), ft);
}

static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/arm.v5a-linux.elf-entry.h"
static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/arm.v5a-linux.elf-so_entry.h"
static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/arm.v5a-linux.elf-fold.h"
static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/arm.v5a-linux.elf-so_fold.h"
static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/arm.v5t-linux.shlib-init.h"

static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/arm.v4a-linux.elf-entry.h"
static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/arm.v4a-linux.elf-so_entry.h"
static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/arm.v4a-linux.elf-fold.h"
static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/arm.v4a-linux.elf-so_fold.h"
#if 0
static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/arm.v4a-linux.shlib-init.h"
#endif

static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/armeb.v4a-linux.elf-entry.h"
static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/armeb.v4a-linux.elf-fold.h"

void
PackLinuxElf32armBe::buildLoader(Filter const *ft)
{
    buildLinuxLoader(
        stub_armeb_v4a_linux_elf_entry, sizeof(stub_armeb_v4a_linux_elf_entry),
        stub_armeb_v4a_linux_elf_fold,  sizeof(stub_armeb_v4a_linux_elf_fold), ft);
}

void
PackLinuxElf32armLe::buildLoader(Filter const *ft)
{
    if (Elf32_Ehdr::ELFOSABI_LINUX==ei_osabi) {
        if (0!=xct_off) {  // shared library
            buildLinuxLoader( // FIXME: 4 vs 5 ?
                stub_arm_v5a_linux_elf_so_entry, sizeof(stub_arm_v5a_linux_elf_so_entry),
                stub_arm_v5a_linux_elf_so_fold,  sizeof(stub_arm_v5a_linux_elf_so_fold), ft);
            return;
        }
        buildLinuxLoader(
            stub_arm_v5a_linux_elf_entry, sizeof(stub_arm_v5a_linux_elf_entry),
            stub_arm_v5a_linux_elf_fold,  sizeof(stub_arm_v5a_linux_elf_fold), ft);
    }
    else {
        buildLinuxLoader(
            stub_arm_v4a_linux_elf_entry, sizeof(stub_arm_v4a_linux_elf_entry),
            stub_arm_v4a_linux_elf_fold,  sizeof(stub_arm_v4a_linux_elf_fold), ft);
    }
}

static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/mipsel.r3000-linux.elf-entry.h"
static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/mipsel.r3000-linux.elf-fold.h"
static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/mipsel.r3000-linux.shlib-init.h"

void
PackLinuxElf32mipsel::buildLoader(Filter const *ft)
{
    if (0!=xct_off) {  // shared library
        buildLinuxLoader(
            stub_mipsel_r3000_linux_shlib_init, sizeof(stub_mipsel_r3000_linux_shlib_init),
            nullptr,                        0,                                 ft );
        return;
    }
    buildLinuxLoader(
        stub_mipsel_r3000_linux_elf_entry, sizeof(stub_mipsel_r3000_linux_elf_entry),
        stub_mipsel_r3000_linux_elf_fold,  sizeof(stub_mipsel_r3000_linux_elf_fold), ft);
}

static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/mips.r3000-linux.elf-entry.h"
static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/mips.r3000-linux.elf-fold.h"
static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/mips.r3000-linux.shlib-init.h"

void
PackLinuxElf32mipseb::buildLoader(Filter const *ft)
{
    if (0!=xct_off) {  // shared library
        buildLinuxLoader(
            stub_mips_r3000_linux_shlib_init, sizeof(stub_mips_r3000_linux_shlib_init),
            nullptr,                        0,                                 ft );
        return;
    }
    buildLinuxLoader(
        stub_mips_r3000_linux_elf_entry, sizeof(stub_mips_r3000_linux_elf_entry),
        stub_mips_r3000_linux_elf_fold,  sizeof(stub_mips_r3000_linux_elf_fold), ft);
}

static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/powerpc-linux.elf-entry.h"
static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/powerpc-linux.elf-fold.h"

void
PackLinuxElf32ppc::buildLoader(const Filter *ft)
{
    buildLinuxLoader(
        stub_powerpc_linux_elf_entry, sizeof(stub_powerpc_linux_elf_entry),
        stub_powerpc_linux_elf_fold,  sizeof(stub_powerpc_linux_elf_fold), ft);
}

static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/powerpc64le-linux.elf-entry.h"
static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/powerpc64le-linux.elf-fold.h"

void
PackLinuxElf64ppcle::buildLoader(const Filter *ft)
{
    buildLinuxLoader(
        stub_powerpc64le_linux_elf_entry, sizeof(stub_powerpc64le_linux_elf_entry),
        stub_powerpc64le_linux_elf_fold,  sizeof(stub_powerpc64le_linux_elf_fold), ft);
}

static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/powerpc64-linux.elf-entry.h"
static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/powerpc64-linux.elf-fold.h"

void
PackLinuxElf64ppc::buildLoader(const Filter *ft)
{
    buildLinuxLoader(
        stub_powerpc64_linux_elf_entry, sizeof(stub_powerpc64_linux_elf_entry),
        stub_powerpc64_linux_elf_fold,  sizeof(stub_powerpc64_linux_elf_fold), ft);
}

static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/amd64-linux.elf-entry.h"
static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/amd64-linux.elf-fold.h"
static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/amd64-linux.elf-so_entry.h"
static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/amd64-linux.elf-so_fold.h"
static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/amd64-linux.shlib-init.h"

void
PackLinuxElf64amd::buildLoader(const Filter *ft)
{
    if (0!=xct_off) {  // shared library
        buildLinuxLoader(
            stub_amd64_linux_elf_so_entry, sizeof(stub_amd64_linux_elf_so_entry),
            stub_amd64_linux_elf_so_fold,  sizeof(stub_amd64_linux_elf_so_fold), ft);
        return;
    }
    buildLinuxLoader(
        stub_amd64_linux_elf_entry, sizeof(stub_amd64_linux_elf_entry),
        stub_amd64_linux_elf_fold,  sizeof(stub_amd64_linux_elf_fold), ft);
}

static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/arm64-linux.elf-entry.h"
static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/arm64-linux.elf-so_entry.h"
static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/arm64-linux.elf-fold.h"
static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/arm64-linux.elf-so_fold.h"
static const CLANG_FORMAT_DUMMY_STATEMENT
#include "stub/arm64-linux.shlib-init.h"

void
PackLinuxElf64arm::buildLoader(const Filter *ft)
{
    if (0!=xct_off) {  // shared library
        buildLinuxLoader(
            stub_arm64_linux_elf_so_entry, sizeof(stub_arm64_linux_elf_so_entry),
            stub_arm64_linux_elf_so_fold,  sizeof(stub_arm64_linux_elf_so_fold), ft);
        return;
    }
    buildLinuxLoader(
        stub_arm64_linux_elf_entry, sizeof(stub_arm64_linux_elf_entry),
        stub_arm64_linux_elf_fold,  sizeof(stub_arm64_linux_elf_fold), ft);
}

    // DT_HASH, DT_GNU_HASH have no explicit length (except in ElfXX_Shdr),
    // so it is hard to detect when the index of a hash chain is out-of-bounds.
    // Workaround: Assume no overlap of DT_* tables (and often contiguous.)
    // Then any given table ends at least as early as when another table begins.
    // So find the tables, and sort the offsets.
    // The 32-bit DT_xxxxx keys have the same values as 64-bit DT_xxxxx keys.
static unsigned const dt_keys[] = {
        Elf64_Dyn::DT_SYMTAB,
        Elf64_Dyn::DT_VERSYM,  // not small integer
        Elf64_Dyn::DT_VERNEED,  // not small integer
        Elf64_Dyn::DT_HASH,
        Elf64_Dyn::DT_GNU_HASH,  // not small integer
        Elf64_Dyn::DT_STRTAB,
        Elf64_Dyn::DT_VERDEF,  // not small integer
        Elf64_Dyn::DT_REL,
        Elf64_Dyn::DT_RELA,
        Elf64_Dyn::DT_FINI_ARRAY,
        Elf64_Dyn::DT_INIT_ARRAY,
        Elf64_Dyn::DT_PREINIT_ARRAY,
        0,
};

static int __acc_cdecl_qsort
qcmp_unsigned(void const *const aa, void const *const bb)
{
    unsigned a = *(unsigned const *)aa;
    unsigned b = *(unsigned const *)bb;
    if (a < b) return -1;
    if (a > b) return  1;
    return  0;
}

void
PackLinuxElf32::sort_DT32_offsets(Elf32_Dyn const *const dynp0)
{
    mb_dt_offsets.alloc(sizeof(unsigned) * sizeof(dt_keys)/sizeof(dt_keys[0]));
    mb_dt_offsets.clear();
    dt_offsets = (unsigned *)mb_dt_offsets.getVoidPtr();
    unsigned n_off = 0, k;
    for (unsigned j=0; ((k = dt_keys[j]),  k); ++j) {
        dt_offsets[n_off] = 0;  // default to "not found"
        u32_t rva = 0;
        if (k < DT_NUM) { // in range of easy table
            if (!dt_table[k]) {
                continue;  // not present in input
            }
            rva = get_te32(&dynp0[-1+ dt_table[k]].d_val);
        }
        else if (file_image) { // why is this guard necessary?
            rva = elf_unsigned_dynamic(k);  // zero if not found
        }
        if (!rva) {
            continue;  // not present in input
        }
        Elf32_Phdr const *phdr = elf_find_Phdr_for_va(rva, phdri, e_phnum);
        if (!phdr) {
            char msg[60]; snprintf(msg, sizeof(msg), "bad  DT_{%#x} = %#x (no Phdr)",
                k, rva);
            throwCantPack(msg);
        }
        dt_offsets[n_off] = (rva - get_te32(&phdr->p_vaddr)) + get_te32(&phdr->p_offset);

        if (file_size <= dt_offsets[n_off]) {
            char msg[60]; snprintf(msg, sizeof(msg), "bad DT_{%#x} = %#x (beyond EOF)",
                k, dt_offsets[n_off]);
                throwCantPack(msg);
        }
        n_off += !!dt_offsets[n_off];
    }
    dt_offsets[n_off++] = file_size;  // sentinel
    upx_qsort(dt_offsets, n_off, sizeof(dt_offsets[0]), qcmp_unsigned);
}

unsigned PackLinuxElf32::find_dt_ndx(unsigned rva)
{
    unsigned *const dto = (unsigned *)mb_dt_offsets.getVoidPtr();
    unsigned const dto_size = mb_dt_offsets.getSize() / sizeof(*dto);
    for (unsigned j = 0; j < dto_size && dto[j]; ++j) { // linear search of short table
        if (rva == dto[j]) {
            return j;
        }
    }
    return ~0u;
}

unsigned PackLinuxElf32::elf_find_table_size(unsigned dt_type, unsigned sh_type)
{
    Elf32_Shdr const *sec = elf_find_section_type(sh_type);
    if (sec) { // Cheat the easy way: use _Shdr.  (No _Shdr anyway for de-compression)
        return get_te32(&sec->sh_size);
    }
    // Honest hard work: use _Phdr
    unsigned x_rva;
    if (dt_type < DT_NUM) {
        unsigned const x_ndx = dt_table[dt_type];
        if (!x_ndx) { // no such entry
            return 0;
        }
        x_rva = get_te32(&dynseg[-1+ x_ndx].d_val);
    }
    else {
        x_rva = elf_unsigned_dynamic(dt_type);
    }
    Elf32_Phdr const *const x_phdr = elf_find_Phdr_for_va(x_rva, phdri, e_phnum);
    if (!x_phdr)
        return ~0u;  // corrupted Phdrs?
    unsigned const           d_off =             x_rva - get_te32(&x_phdr->p_vaddr);
    unsigned const           y_ndx = find_dt_ndx(d_off + get_te32(&x_phdr->p_offset));
    if (~0u != y_ndx) {
        return dt_offsets[1+ y_ndx] - dt_offsets[y_ndx];
    }
    return ~0u;
}

void
PackLinuxElf32::invert_pt_dynamic(Elf32_Dyn const *dynp, u32_t headway)
{
    if (dt_table[Elf32_Dyn::DT_NULL]) {
        return;  // not 1st time; do not change upx_dt_init
    }
    Elf32_Dyn const *const dynp0 = dynp;
    unsigned ndx = 0;
    unsigned const limit = headway / sizeof(*dynp);
    if (dynp)
    for (; ; ++ndx, ++dynp) {
        if (limit <= ndx) {
            throwCantPack("DT_NULL not found");
        }
        u32_t const d_tag = get_te32(&dynp->d_tag);
        if (d_tag < DT_NUM) {
            if (Elf32_Dyn::DT_NEEDED != d_tag
            &&  dt_table[d_tag]
            &&    get_te32(&dynp->d_val)
               != get_te32(&dynp0[-1+ dt_table[d_tag]].d_val)) {
                throwCantPack("duplicate DT_%#x: [%#x] [%#x]",
                    (unsigned)d_tag, -1+ dt_table[d_tag], ndx);
            }
            dt_table[d_tag] = 1+ ndx;
        }
        if (Elf32_Dyn::DT_NULL == d_tag) {
            break;  // check here so that dt_table[DT_NULL] is set
        }
    }
    sort_DT32_offsets(dynp0);

    upx_dt_init = 0;
         if (dt_table[Elf32_Dyn::DT_INIT])          upx_dt_init = Elf32_Dyn::DT_INIT;
    else if (dt_table[Elf32_Dyn::DT_PREINIT_ARRAY]) upx_dt_init = Elf32_Dyn::DT_PREINIT_ARRAY;
    else if (dt_table[Elf32_Dyn::DT_INIT_ARRAY])    upx_dt_init = Elf32_Dyn::DT_INIT_ARRAY;

    unsigned const z_str = dt_table[Elf32_Dyn::DT_STRSZ];
    strtab_max = !z_str ? 0 : get_te32(&dynp0[-1+ z_str].d_val);
    unsigned const z_tab = dt_table[Elf32_Dyn::DT_STRTAB];
    unsigned const tmp1 = !z_tab ? 0 : get_te32(&dynp0[-1+ z_tab].d_val);
    if (tmp1 < sz_elf_hdrs) {
        throwCantPack("bad DT_STRTAB %#x", tmp1);
    }
    unsigned const strtab_beg = !z_tab ? 0 : elf_get_offset_from_address(tmp1);

    if (!z_str || !z_tab || !(strtab_max + strtab_beg)
    || (this->file_size - strtab_beg) < strtab_max  // strtab overlaps EOF
        // last string in table must have terminating NUL
    ||  '\0' != ((char *)file_image.getVoidPtr())[-1+ strtab_max + strtab_beg]
    ) {
        throwCantPack("bad DT_STRSZ %#x", strtab_max);
    }

    { // Find end of DT_SYMTAB
        unsigned const tmp2 = elf_find_table_size(Elf32_Dyn::DT_SYMTAB,
            Elf32_Shdr::SHT_DYNSYM);
        symnum_max = (~0u == tmp2) ? 0 : tmp2 / sizeof(Elf32_Sym);
    }

    unsigned v_sym = dt_table[Elf32_Dyn::DT_SYMTAB];
    if (v_sym) {
        v_sym = elf_get_offset_from_address(get_te32(&dynp0[-1+ v_sym].d_val));
    }
    unsigned v_hsh = dt_table[Elf32_Dyn::DT_HASH];
    if (v_hsh) {
        v_hsh = elf_get_offset_from_address(get_te32(&dynp0[-1+ v_hsh].d_val));
    }
    if (v_hsh && file_image) {
        hashtab = (unsigned const *)elf_find_dynamic(Elf32_Dyn::DT_HASH);
        if (!hashtab) {
            throwCantPack("bad DT_HASH %#x", v_hsh);
        }
        // Find end of DT_HASH
        hashend = (unsigned const *)(void const *)(elf_find_table_size(
            Elf32_Dyn::DT_HASH, Elf32_Shdr::SHT_HASH) + (char const *)hashtab);
        if (!hashtab || (char const *)hashend <= (char const *)&hashtab[2]
        ||  file_image.getSizeInBytes()
            < (unsigned)((char const *)&hashtab[2] - (char *)&file_image[0]) )
        {
            throwCantPack("bad DT_HASH %#x", v_hsh);
        }

        unsigned const nbucket = get_te32(&hashtab[0]);
        unsigned const *const buckets = &hashtab[2];
        unsigned const *const chains = &buckets[nbucket]; (void)chains;
        if ((unsigned)(file_size - ((char const *)buckets - (char const *)(void const *)file_image))
                <= sizeof(unsigned)*nbucket ) {
            throwCantPack("bad nbucket %#x\n", nbucket);
        }

        if ((unsigned)(hashend - buckets) < nbucket
        || !v_sym || (unsigned)file_size <= v_sym
        || ((v_hsh < v_sym) && (v_sym - v_hsh) < sizeof(*buckets)*(2+ nbucket))
        ) {
            throwCantPack("bad DT_HASH nbucket=%#x  len=%#x",
                nbucket, (v_sym - v_hsh));
        }
        unsigned chmax = 0;
        for (unsigned j= 0; j < nbucket; ++j) {
            unsigned x = get_te32(&buckets[j]);
            if (chmax < x) {
                chmax = x;
            }
        }
        if ((v_hsh < v_sym) && (v_sym - v_hsh) <
                (sizeof(*buckets)*(2+ nbucket) + sizeof(*chains)*(1+ chmax))) {
            throwCantPack("bad DT_HASH nbucket=%#x  len=%#x",
                nbucket, (v_sym - v_hsh));
        }
    }
    unsigned const v_gsh = elf_unsigned_dynamic(Elf32_Dyn::DT_GNU_HASH);
    if (v_gsh && file_image) {
        // Not similar to DT_HASH because DT_GNU_HASH is not small (0x6ffffef5).
        gashtab = (unsigned const *)elf_find_dynamic(Elf32_Dyn::DT_GNU_HASH);
        gashend = (unsigned const *)(void const *)(elf_find_table_size(
            Elf32_Dyn::DT_GNU_HASH, Elf32_Shdr::SHT_GNU_HASH) + (char const *)gashtab);
        if (!gashtab || (char const *)gashend <= (char const *)&gashtab[4]
        ||  file_image.getSizeInBytes()
            < (unsigned)((char const *)&gashtab[4] - (char *)&file_image[0]) )
        {
            throwCantPack("bad DT_GNU_HASH %#x", v_gsh);
        }

        unsigned const n_bucket = get_te32(&gashtab[0]);
        unsigned const symbias  = get_te32(&gashtab[1]);
        unsigned const n_bitmask = get_te32(&gashtab[2]);
        unsigned const gnu_shift = get_te32(&gashtab[3]);
        u32_t const *const bitmask = (u32_t const *)(void const *)&gashtab[4];
        unsigned     const *const buckets = (unsigned const *)&bitmask[n_bitmask];
        unsigned     const *const hasharr = &buckets[n_bucket]; (void)hasharr;
        if (!n_bucket || (1u<<31) <= n_bucket  /* fie on fuzzers */
        || (unsigned)(gashend - buckets) < n_bucket
        || (file_size + file_image) <= (void const *)hasharr) {
            throwCantPack("bad n_bucket %#x\n", n_bucket);
        }
        // It would be better to detect zeroes shifted into low 5 bits of:
        //    (037 & (hash_32 >> gnu_shift))
        // but compilers can be stupid.
        if (31 < gnu_shift) {
            throwCantPack("bad gnu_shift %#x", gnu_shift);
        }
        // unsigned const *const gashend = &hasharr[n_bucket];
        // minimum, except:
        // Rust and Android trim unused zeroes from high end of hasharr[]
        unsigned bmax = 0;
        for (unsigned j= 0; j < n_bucket; ++j) {
            unsigned bj = get_te32(&buckets[j]);
            if (bj) {
                if (bj < symbias) {
                    throwCantPack("bad DT_GNU_HASH bucket[%d] < symbias{%#x}\n",
                        bj, symbias);
                }
                if (bmax < bj) {
                    bmax = bj;
                }
            }
        }
        if (1==n_bucket  && 0==buckets[0]
        &&  1==n_bitmask && 0==bitmask[0]) {
            // 2021-09-11 Rust on RaspberryPi apparently uses this to minimize space.
            // But then the DT_GNU_HASH symbol lookup algorithm always fails?
            // https://github.com/upx/upx/issues/525
        } else
        if (bmax) {
            if ((1+ bmax) < symbias) {
                throwCantPack("bad DT_GNU_HASH (1+ max_bucket)=%#x < symbias=%#x",
                    1+ bmax, symbias);
            }
            bmax -= symbias;
        }

        unsigned r = 0;
        if (!n_bucket || !n_bitmask || !v_sym
        || (r=1, ((-1+ n_bitmask) & n_bitmask))  // not a power of 2
        || (r=2, (8*sizeof(u32_t) <= gnu_shift))  // shifted result always == 0
        || (r=3, (n_bucket>>30))  // fie on fuzzers
        || (r=4, (n_bitmask>>30))
        || (r=5, ((file_size/sizeof(unsigned))
                <= ((sizeof(*bitmask)/sizeof(unsigned))*n_bitmask + 2*n_bucket)))  // FIXME: weak
        || (r=6, ((v_gsh < v_sym) && (v_sym - v_gsh) < (sizeof(unsigned)*4  // headers
                + sizeof(*bitmask)*n_bitmask  // bitmask
                + sizeof(*buckets)*n_bucket  // buckets
                + sizeof(*hasharr)*(!bmax ? 0 : (1+ bmax))  // hasharr
            )) )
        ) {
            char msg[90]; snprintf(msg, sizeof(msg),
                "bad DT_GNU_HASH n_bucket=%#x  n_bitmask=%#x  len=%#lx  r=%d",
                n_bucket, n_bitmask, (long unsigned)(v_sym - v_gsh), r);
            throwCantPack(msg);
        }
    }
    e_shstrndx = get_te16(&ehdri.e_shstrndx);  // who omitted this?
    if (e_shnum <= e_shstrndx
    &&  !(0==e_shnum && 0==e_shstrndx) ) {
        char msg[40]; snprintf(msg, sizeof(msg),
            "bad .e_shstrndx %d >= .e_shnum %d", e_shstrndx, e_shnum);
        throwCantPack(msg);
    }
}

Elf32_Phdr const *
PackLinuxElf32::elf_find_ptype(unsigned type, Elf32_Phdr const *phdr, unsigned phnum)
{
    for (unsigned j = 0; j < phnum; ++j, ++phdr) {
        if (type == get_te32(&phdr->p_type)) {
            return phdr;
        }
    }
    return nullptr;
}

Elf64_Phdr const *
PackLinuxElf64::elf_find_ptype(unsigned type, Elf64_Phdr const *phdr, unsigned phnum)
{
    for (unsigned j = 0; j < phnum; ++j, ++phdr) {
        if (type == get_te32(&phdr->p_type)) {
            return phdr;
        }
    }
    return nullptr;
}

Elf32_Shdr const *PackLinuxElf32::elf_find_section_name(
    char const *const name
) const
{
    Elf32_Shdr const *shdr = shdri;
    if (!shdr) {
        return nullptr;
    }
    int j = e_shnum;
    for (; 0 <=--j; ++shdr) {
        unsigned const sh_name = get_te32(&shdr->sh_name);
        if ((u32_t)file_size <= sh_name) {  // FIXME: weak
            char msg[50]; snprintf(msg, sizeof(msg),
                "bad Elf32_Shdr[%d].sh_name %#x",
                -1+ e_shnum -j, sh_name);
            throwCantPack(msg);
        }
        if (0==strcmp(name, &shstrtab[sh_name])) {
            return shdr;
        }
    }
    return nullptr;
}

Elf64_Shdr const *PackLinuxElf64::elf_find_section_name(
    char const *const name
) const
{
    Elf64_Shdr const *shdr = shdri;
    if (!shdr) {
        return nullptr;
    }
    int j = e_shnum;
    for (; 0 <=--j; ++shdr) {
        unsigned const sh_name = get_te32(&shdr->sh_name);
        if ((u32_t)file_size <= sh_name) {  // FIXME: weak
            char msg[50]; snprintf(msg, sizeof(msg),
                "bad Elf64_Shdr[%d].sh_name %#x",
                -1+ e_shnum -j, sh_name);
            throwCantPack(msg);
        }
        if (0==strcmp(name, &shstrtab[sh_name])) {
            return shdr;
        }
    }
    return nullptr;
}

Elf32_Shdr *PackLinuxElf32::elf_find_section_type(
    unsigned const type
) const
{
    Elf32_Shdr *shdr = shdri;
    if (!shdr) {
        return nullptr;
    }
    int j = e_shnum;
    for (; 0 <=--j; ++shdr) {
        if (type==get_te32(&shdr->sh_type)) {
            return shdr;
        }
    }
    return nullptr;
}

Elf64_Shdr *PackLinuxElf64::elf_find_section_type(
    unsigned const type
) const
{
    Elf64_Shdr *shdr = shdri;
    if (!shdr) {
        return nullptr;
    }
    int j = e_shnum;
    for (; 0 <=--j; ++shdr) {
        if (type==get_te32(&shdr->sh_type)) {
            return shdr;
        }
    }
    return nullptr;
}

char const *PackLinuxElf64::get_str_name(unsigned st_name, unsigned symnum) const
{
    if (strtab_max <= st_name) {
        char msg[70]; snprintf(msg, sizeof(msg),
            "bad .st_name %#x in DT_SYMTAB[%d]", st_name, symnum);
        throwCantPack(msg);
    }
    return &dynstr[st_name];
}

char const *PackLinuxElf64::get_dynsym_name(unsigned symnum, unsigned relnum) const
{
    if (symnum_max <= symnum) {
        (void)relnum;
        return nullptr;
    }
    return get_str_name(get_te32(&dynsym[symnum].st_name), symnum);
}

bool PackLinuxElf64::calls_crt1(Elf64_Rela const *rela, int sz)
{
    if (!dynsym || !dynstr || !rela) {
        return false;
    }
    for (unsigned relnum= 0; 0 < sz; (sz -= sizeof(Elf64_Rela)), ++rela, ++relnum) {
        unsigned const symnum = get_te64(&rela->r_info) >> 32;
        char const *const symnam = get_dynsym_name(symnum, relnum);
        if (0==strcmp(symnam, "__libc_start_main")  // glibc
        ||  0==strcmp(symnam, "__libc_init")  // Android
        ||  0==strcmp(symnam, "__uClibc_main")
        ||  0==strcmp(symnam, "__uClibc_start_main"))
            return true;
    }
    return false;
}

char const *PackLinuxElf32::get_str_name(unsigned st_name, unsigned symnum) const
{
    if (strtab_max <= st_name) {
        char msg[70]; snprintf(msg, sizeof(msg),
            "bad .st_name %#x in DT_SYMTAB[%d]\n", st_name, symnum);
        throwCantPack(msg);
    }
    return &dynstr[st_name];
}

char const *PackLinuxElf32::get_dynsym_name(unsigned symnum, unsigned relnum) const
{
    if (symnum_max <= symnum) {
        (void)relnum;
        return nullptr;
    }
    return get_str_name(get_te32(&dynsym[symnum].st_name), symnum);
}

bool PackLinuxElf32::calls_crt1(Elf32_Rel const *rel, int sz)
{
    if (!dynsym || !dynstr || !rel) {
        return false;
    }
    for (unsigned relnum= 0; 0 < sz; (sz -= sizeof(Elf32_Rel)), ++rel, ++relnum) {
        unsigned const symnum = get_te32(&rel->r_info) >> 8;
        char const *const symnam = get_dynsym_name(symnum, relnum);
        if (0==strcmp(symnam, "__libc_start_main")  // glibc
        ||  0==strcmp(symnam, "__libc_init")  // Android
        ||  0==strcmp(symnam, "__uClibc_main")
        ||  0==strcmp(symnam, "__uClibc_start_main"))
            return true;
    }
    return false;
}

tribool PackLinuxElf32::canUnpack() // bool, except -1: format known, but not packed
{
    if (checkEhdr(&ehdri)) {
        return false;
    }
    if (get_te16(&ehdri.e_phnum) < 2) {
        throwCantUnpack("e_phnum must be >= 2");
    }
    if (Elf32_Ehdr::ET_DYN==get_te16(&ehdri.e_type)) {
        PackLinuxElf32help1(fi);
    }
    if (super::canUnpack()) {
        return true;
    }
    return false;
}

bool  // false [often throwCantPack]: some defect;  true: good so far
PackLinuxElf32::canPackOSABI(Elf32_Ehdr const *ehdr)
{
    unsigned char osabi0 = ehdr->e_ident[Elf32_Ehdr::EI_OSABI];
    // The first PT_LOAD must cover the beginning of the file (0==p_offset).
    Elf32_Phdr const *phdr = phdri;
    note_size = 0;
    for (unsigned j=0; j < e_phnum; ++phdr, ++j) {
        if (j > ((MAX_ELF_HDR_32 - sizeof(Elf32_Ehdr)) / sizeof(Elf32_Phdr))) {
            throwCantPack("too many ElfXX_Phdr; try '--force-execve'");
            return false;
        }
        unsigned const p_type = get_te32(&phdr->p_type);
        unsigned const p_offset = get_te32(&phdr->p_offset);
        if (1!=exetype && PT_LOAD == p_type) { // 1st PT_LOAD
            exetype = 1;
            load_va = get_te32(&phdr->p_vaddr);  // class data member

            // Cast on next line is to avoid a compiler bug (incorrect complaint) in
            // Microsoft (R) C/C++ Optimizing Compiler Version 19.00.24215.1 for x64
            // error C4319: '~': zero extending 'unsigned int' to 'upx_uint64_t' of greater size
            unsigned const off = ~page_mask & (unsigned)load_va;

            if (off && off == p_offset) { // specific hint
                throwCantPack("Go-language PT_LOAD: try hemfix.c, or try '--force-execve'");
                // Fixing it inside upx fails because packExtent() reads original file.
                return false;
            }
            if (0 != p_offset) { // 1st PT_LOAD must cover Ehdr and Phdr
                throwCantPack("first PT_LOAD.p_offset != 0; try '--force-execve'");
                return false;
            }
            hatch_off = ~3u & (3+ get_te32(&phdr->p_memsz));
        }
        if (EM_MIPS == e_machine
        &&  (Elf32_Phdr::PT_MIPS_ABIFLAGS == p_type
          || Elf32_Phdr::PT_MIPS_REGINFO  == p_type)
        ) {
            add_phdrx(phdr);  // PT_MIPS_*
            unsigned mask = -1+ get_te32(&phdr->p_align);
            sz_phdrx = ~mask & (mask + sz_phdrx);
            sz_phdrx += get_te32(&phdr->p_memsz);

        }
        if (PT_NOTE32 == p_type) {
            if (osabi_note && Elf32_Ehdr::ELFOSABI_NONE==osabi0) { // Still seems to be generic.
                struct {
                    struct Elf32_Nhdr nhdr;
                    char name[8];
                    unsigned body;
                } note;
                memset(&note, 0, sizeof(note));
                fi->seek(p_offset, SEEK_SET);
                fi->readx(&note, sizeof(note));
                fi->seek(0, SEEK_SET);
                if (4==get_te32(&note.nhdr.descsz)
                &&  1==get_te32(&note.nhdr.type)
                // &&  0==note.end
                &&  (1+ strlen(osabi_note))==get_te32(&note.nhdr.namesz)
                &&  0==strcmp(osabi_note, (char const *)&note.name[0])
                ) {
                    osabi0 = ei_osabi;  // Specified by PT_NOTE.
                }
            }
        }
    }
    if (Elf32_Ehdr::ELFOSABI_NONE ==osabi0
    ||  Elf32_Ehdr::ELFOSABI_LINUX==osabi0) { // No EI_OSBAI, no PT_NOTE.
        unsigned const arm_eabi = 0xff000000u & get_te32(&ehdr->e_flags);
        if (Elf32_Ehdr::EM_ARM==e_machine
        &&   (EF_ARM_EABI_VER5==arm_eabi
          ||  EF_ARM_EABI_VER4==arm_eabi ) ) {
            // armel-eabi armeb-eabi ARM Linux EABI version 4 is a mess.
            ei_osabi = osabi0 = Elf32_Ehdr::ELFOSABI_LINUX;
        }
        else {
            osabi0 = opt->o_unix.osabi0;  // Possibly specified by command-line.
        }
    }
    if (osabi0!=ei_osabi) {
        return false;
    }
    return true;  // good so far
}

#define WANT_SHDR_ENUM
#include "p_elf_enum.h"
#undef  WANT_SHDR_ENUM

upx_uint64_t PackLinuxElf32::canPack_Shdr(Elf32_Phdr const *pload_x0)
{
    Elf32_Shdr const *shdr_xva = nullptr;
    Elf32_Shdr const *shdr = shdri;
  for (int j= e_shnum; --j>=0; ++shdr) {
    unsigned const sh_type = get_te32(&shdr->sh_type);
    if (!shdr_xva && Elf32_Shdr::SHF_EXECINSTR & get_te32(&shdr->sh_flags)) {
        shdr_xva = shdr;
        xct_va = get_te32(&shdr_xva->sh_addr);
    }
    // Hook the first slot of DT_PREINIT_ARRAY or DT_INIT_ARRAY.
    if (!user_init_rp && (
        (     Elf32_Dyn::DT_PREINIT_ARRAY==upx_dt_init
        &&  Elf32_Shdr::SHT_PREINIT_ARRAY==sh_type)
    ||  (     Elf32_Dyn::DT_INIT_ARRAY   ==upx_dt_init
        &&  Elf32_Shdr::SHT_INIT_ARRAY   ==sh_type) )) {
        unsigned user_init_ava = get_te32(&shdr->sh_addr);
        user_init_off = get_te32(&shdr->sh_offset);
        if ((u32_t)file_size <= user_init_off) {
            char msg[70]; snprintf(msg, sizeof(msg),
                "bad Elf32_Shdr[%d].sh_offset %#x",
                -1+ e_shnum - j, user_init_off);
            throwCantPack(msg);
        }
        // Check that &file_image[user_init_off] has
        // *_RELATIVE or *_ABS* relocation, and fetch user_init_va.
        // If Elf32_Rela then the actual value is in Rela.r_addend.
        int z_rel = dt_table[Elf32_Dyn::DT_REL];
        int z_rsz = dt_table[Elf32_Dyn::DT_RELSZ];
        if (z_rel && z_rsz) {
            unsigned rel_off = get_te32(&dynseg[-1+ z_rel].d_val);
            if ((unsigned)file_size <= rel_off) {
                char msg[70]; snprintf(msg, sizeof(msg),
                     "bad Elf32_Dynamic[DT_REL] %#x\n",
                     rel_off);
                throwCantPack(msg);
            }
            Elf32_Rel *rp = (Elf32_Rel *)&file_image[rel_off];
            unsigned relsz   = get_te32(&dynseg[-1+ z_rsz].d_val);
            if ((unsigned)file_size <= relsz) {
                char msg[70]; snprintf(msg, sizeof(msg),
                     "bad Elf32_Dynamic[DT_RELSZ] %#x\n",
                     relsz);
                throwCantPack(msg);
            }
            Elf32_Rel *last = (Elf32_Rel *)(relsz + (char *)rp);
            for (; rp < last; ++rp) {
                unsigned r_va = get_te32(&rp->r_offset);
                if (r_va == user_init_ava) { // found the Elf32_Rel
                    user_init_rp = rp;
                    unsigned r_info = get_te32(&rp->r_info);
                    unsigned r_type = ELF32_R_TYPE(r_info);
                    set_te32(&dynsym[0].st_name, r_va);  // for decompressor
                    set_te32(&dynsym[0].st_value, r_info);
                    if (Elf32_Ehdr::EM_ARM == e_machine) {
                        if (R_ARM_RELATIVE == r_type) {
                            user_init_va = get_te32(&file_image[user_init_off]);
                        }
                        else if (R_ARM_ABS32 == r_type) {
                            unsigned symj = ELF32_R_SYM(r_info);
                            user_init_va = get_te32(&dynsym[symj].st_value);
                            set_te32(&rp->r_info, ELF32_R_INFO(0, R_ARM_RELATIVE));
                            // pack3() will set &file_image[user_init_off]
                        }
                        else {
                            goto bad;
                        }
                    }
                    else if (Elf32_Ehdr::EM_386 == e_machine) {
                        if (R_386_RELATIVE == r_type) {
                            user_init_va = get_te32(&file_image[user_init_off]);
                        }
                        else if (R_386_32 == r_type) {
                            unsigned symj = ELF32_R_SYM(r_info);
                            user_init_va = get_te32(&dynsym[symj].st_value);
                            set_te32(&rp->r_info, ELF32_R_INFO(0, R_386_RELATIVE));
                            // pack3() will set &file_image[user_init_off]
                        }
                        else {
                            goto bad;
                        }
                    }
                    else {
bad:
                        char msg[50]; snprintf(msg, sizeof(msg),
                            "bad relocation %#x DT_INIT_ARRAY[0]",
                            r_info);
                        throwCantPack(msg);
                    }
                    break;
                }
            }
        }
        unsigned const p_filesz = get_te32(&pload_x0->p_filesz);
        if (!((user_init_va - xct_va) < p_filesz)) {
            // Not in executable portion of first executable PT_LOAD.
            if (0==user_init_va && is_asl) {
                // Android allows (0 ==> skip) ?
                upx_dt_init = 0;  // force steal of 'extra' DT_NULL
                // XXX: FIXME: depends on SHT_DYNAMIC coming later
            }
            else {
                char msg[70]; snprintf(msg, sizeof(msg),
                    "bad init address %#x in Elf32_Shdr[%d].%#x\n",
                    (unsigned)user_init_va, -1+ e_shnum - j, user_init_off);
                throwCantPack(msg);
            }
        }
    }
    // By default /usr/bin/ld leaves 4 extra DT_NULL to support pre-linking.
    // Take one as a last resort.
    if ((Elf32_Dyn::DT_INIT==upx_dt_init || !upx_dt_init)
    &&  Elf32_Shdr::SHT_DYNAMIC == sh_type) {
        unsigned sh_offset = get_te32(&shdr->sh_offset);
        unsigned sh_size = get_te32(&shdr->sh_size);
        if ((unsigned)file_size < sh_size
        ||  (unsigned)file_size < sh_offset
        || ((unsigned)file_size - sh_offset) < sh_size) {
            throwCantPack("bad SHT_DYNAMIC");
        }
        unsigned const n = get_te32(&shdr->sh_size) / sizeof(Elf32_Dyn);
        Elf32_Dyn *dynp = (Elf32_Dyn *)&file_image[get_te32(&shdr->sh_offset)];
        for (; Elf32_Dyn::DT_NULL != dynp->d_tag; ++dynp) {
            if (upx_dt_init == get_te32(&dynp->d_tag)) {
                break;  // re-found DT_INIT
            }
        }
        if ((1+ dynp) < (n+ dynseg)) { // not the terminator, so take it
            user_init_va = get_te32(&dynp->d_val);  // 0 if (0==upx_dt_init)
            set_te32(&dynp->d_tag, upx_dt_init = Elf32_Dyn::DT_INIT);
            user_init_off = (char const *)&dynp->d_val - (char const *)&file_image[0];
        }
    }
  }
    return xct_va;
}

upx_uint64_t PackLinuxElf64::canPack_Shdr(Elf64_Phdr const *pload_x0)
{
    Elf64_Shdr const *shdr_xva = nullptr;
    Elf64_Shdr const *shdr = shdri;
  for (int j= e_shnum; --j>=0; ++shdr) {
    unsigned const sh_type = get_te32(&shdr->sh_type);
    if (!shdr_xva && Elf64_Shdr::SHF_EXECINSTR & get_te64(&shdr->sh_flags)) {
        shdr_xva = shdr;
        xct_va = get_te64_32(&shdr_xva->sh_addr);
    }
    // Hook the first slot of DT_PREINIT_ARRAY or DT_INIT_ARRAY.
    if (!user_init_rp && (
           (     Elf64_Dyn::DT_PREINIT_ARRAY==upx_dt_init
           &&  Elf64_Shdr::SHT_PREINIT_ARRAY==sh_type)
        || (     Elf64_Dyn::DT_INIT_ARRAY   ==upx_dt_init
           &&  Elf64_Shdr::SHT_INIT_ARRAY   ==sh_type) )) {
        unsigned user_init_ava = get_te64(&shdr->sh_addr);
        user_init_off = get_te64(&shdr->sh_offset);
        if ((u64_t)file_size <= user_init_off) {
            char msg[70]; snprintf(msg, sizeof(msg),
                "bad Elf64_Shdr[%d].sh_offset %#x",
                -1+ e_shnum - j, user_init_off);
            throwCantPack(msg);
        }
        // Check that &file_image[user_init_off] has
        // *_RELATIVE or *_ABS* relocation, and fetch user_init_va.
        // If Elf_Rela then the actual value is in Rela.r_addend.
        int z_rel = dt_table[Elf64_Dyn::DT_RELA];
        int z_rsz = dt_table[Elf64_Dyn::DT_RELASZ];
        if (z_rel && z_rsz) {
            upx_uint64_t rel_off = get_te64(&dynseg[-1+ z_rel].d_val);
            if ((u64_t)file_size <= rel_off) {
                char msg[70]; snprintf(msg, sizeof(msg),
                     "bad Elf64_Dynamic[DT_RELA] %#llx\n",
                     rel_off);
                throwCantPack(msg);
            }
            Elf64_Rela *rp = (Elf64_Rela *)&file_image[rel_off];
            upx_uint64_t relsz   = get_te64(&dynseg[-1+ z_rsz].d_val);
            if ((u64_t)file_size <= relsz) {
                char msg[70]; snprintf(msg, sizeof(msg),
                     "bad Elf64_Dynamic[DT_RELASZ] %#llx\n",
                     relsz);
                throwCantPack(msg);
            }
            Elf64_Rela *last = (Elf64_Rela *)(relsz + (char *)rp);
            for (; rp < last; ++rp) {
                upx_uint64_t r_va = get_te64(&rp->r_offset);
                if (r_va == user_init_ava) { // found the Elf64_Rela
                    user_init_rp = rp;
                    upx_uint64_t r_info = get_te64(&rp->r_info);
                    unsigned r_type = ELF64_R_TYPE(r_info);
                    set_te32(&dynsym[0].st_name, r_va);  // for decompressor
                    set_te64(&dynsym[0].st_value, r_info);
                    set_te64(&dynsym[0].st_size, get_te64(&rp->r_addend));
                    if (Elf64_Ehdr::EM_AARCH64 == e_machine) {
                        if (R_AARCH64_RELATIVE == r_type) {
                            user_init_va = get_te64(&rp->r_addend);
                        }
                        else if (R_AARCH64_ABS64 == r_type) {
                            user_init_va = get_te64(&dynsym[ELF64_R_SYM(r_info)].st_value);
                        }
                        else {
                            char msg[50]; snprintf(msg, sizeof(msg),
                                "bad relocation %#llx DT_INIT_ARRAY[0]",
                                r_info);
                            throwCantPack(msg);
                        }
                    }
                    else if (Elf64_Ehdr::EM_X86_64 == e_machine) {
                        if (R_X86_64_RELATIVE == r_type) {
                            user_init_va = get_te64(&rp->r_addend);
                        }
                        else if (R_X86_64_64 == r_type) {
                            user_init_va = get_te64(&dynsym[ELF64_R_SYM(r_info)].st_value);
                        }
                        else {
                            char msg[50]; snprintf(msg, sizeof(msg),
                                "bad relocation %#llx DT_INIT_ARRAY[0]",
                                r_info);
                            throwCantPack(msg);
                        }
                    }
                    break;
                }
            }
        }
        unsigned const p_filesz = get_te64(&pload_x0->p_filesz);
        if (!((user_init_va - xct_va) < p_filesz)) {
            // Not in executable portion of first executable PT_LOAD.
            if (0==user_init_va && is_asl) {
                // Android allows (0 ==> skip) ?
                upx_dt_init = 0;  // force steal of 'extra' DT_NULL
                // XXX: FIXME: depends on SHT_DYNAMIC coming later
            }
            else {
                char msg[70]; snprintf(msg, sizeof(msg),
                    "bad init address %#x in Elf64_Shdr[%d].%#x\n",
                    (unsigned)user_init_va, -1+ e_shnum - j, user_init_off);
                throwCantPack(msg);
            }
        }
    }
    // By default /usr/bin/ld leaves 4 extra DT_NULL to support pre-linking.
    // Take one as a last resort.
    if ((Elf64_Dyn::DT_INIT==upx_dt_init || !upx_dt_init)
    &&  Elf64_Shdr::SHT_DYNAMIC == sh_type) {
        upx_uint64_t sh_offset = get_te64(&shdr->sh_offset);
        upx_uint64_t sh_size = get_te64(&shdr->sh_size);
        if ((upx_uint64_t)file_size < sh_size
        ||  (upx_uint64_t)file_size < sh_offset
        || ((upx_uint64_t)file_size - sh_offset) < sh_size) {
            throwCantPack("bad SHT_DYNAMIC");
        }
        unsigned const n = sh_size / sizeof(Elf64_Dyn);
        Elf64_Dyn *dynp = (Elf64_Dyn *)&file_image[sh_offset];
        for (; Elf64_Dyn::DT_NULL != dynp->d_tag; ++dynp) {
            if (upx_dt_init == get_te64(&dynp->d_tag)) {
                break;  // re-found DT_INIT
            }
        }
        if ((1+ dynp) < (n+ dynseg)) { // not the terminator, so take it
            user_init_va = get_te64(&dynp->d_val);  // 0 if (0==upx_dt_init)
            set_te64(&dynp->d_tag, upx_dt_init = Elf64_Dyn::DT_INIT);
            user_init_off = (char const *)&dynp->d_val - (char const *)&file_image[0];
        }
    }
  }
    return xct_va;
}

tribool PackLinuxElf32::canPack()
{
    union {
        unsigned char buf[MAX_ELF_HDR_32];
        //struct { Elf32_Ehdr ehdr; Elf32_Phdr phdr; } e;
    } u;
    COMPILE_TIME_ASSERT(sizeof(u.buf) <= (2*512))

// My earlier design with "extra" Shdrs in output at xct_off
// DOES NOT WORK because code for EM_ARM has embedded relocations
// that are not made visible, such as:
//    ----- glibc-2.31/sysdeps/arm/crti.S
//            .type call_weak_fn, %function
//    call_weak_fn:
//            ldr r3, .LGOT
//            ldr r2, .LGOT+4
//    .LPIC:
//            add r3, pc, r3
//            ldr r2, [r3, r2]
//            cmp r2, #0
//            bxeq lr
//            b PREINIT_FUNCTION
//            .p2align 2
//    .LGOT:
//            .word _GLOBAL_OFFSET_TABLE_-(.LPIC+8)  // unseen reloc!
//            .word PREINIT_FUNCTION(GOT)
//    -----
// So, PackUnix::PackUnix() disables (but silently accepts) --android-shlib,
// and see if appending ARM_ATTRIBUTES Shdr is good enough.

    fi->seek(0, SEEK_SET);
    fi->readx(u.buf, sizeof(u.buf));
    fi->seek(0, SEEK_SET);
    Elf32_Ehdr const *const ehdr = (Elf32_Ehdr *) u.buf;

    // now check the ELF header
    if (checkEhdr(ehdr) != 0)
        return false;

    // additional requirements for linux/elf386
    if (get_te16(&ehdr->e_ehsize) != sizeof(*ehdr)) {
        throwCantPack("invalid Ehdr e_ehsize; try '--force-execve'");
        return false;
    }
    if (e_phoff != sizeof(*ehdr)) {// Phdrs not contiguous with Ehdr
        throwCantPack("non-contiguous Ehdr/Phdr; try '--force-execve'");
        return false;
    }

    if (!canPackOSABI((Elf32_Ehdr *)u.buf)) {
        return false;
    }
    upx_uint32_t max_LOADsz = 0, max_offset = 0;
    Elf32_Phdr *phdr = phdri;
    for (unsigned j=0; j < e_phnum; ++phdr, ++j) {
        if (j > ((MAX_ELF_HDR_32 - sizeof(Elf32_Ehdr)) / sizeof(Elf32_Phdr))) {
            throwCantPack("too many ElfXX_Phdr; try '--force-execve'");
            return false;
        }
        if (is_LOAD(phdr)) {
            // The first PT_LOAD must cover the beginning of the file (0==p_offset).
            upx_uint32_t const p_offset = get_te32(&phdr->p_offset);
            if (1!= exetype) {
                exetype = 1;
                load_va = get_te32(&phdr->p_vaddr);  // class data member
                upx_uint32_t const off = ~page_mask & (upx_uint32_t)load_va;
                if (off && off == p_offset) { // specific hint
                    throwCantPack("Go-language PT_LOAD: try hemfix.c, or try '--force-execve'");
                    // Fixing it inside upx fails because packExtent() reads original file.
                    return false;
                }
                if (0 != p_offset) { // 1st PT_LOAD must cover Ehdr and Phdr
                    throwCantPack("first PT_LOAD.p_offset != 0; try '--force-execve'");
                    return false;
                }
                // FIXME: bad for shlib!
                hatch_off = ~3ul & (3+ get_te32(&phdr->p_memsz));
            }
            max_LOADsz = UPX_MAX(max_LOADsz, UPX_MIN(0x200000u, get_te32(&phdr->p_align)));
            unsigned filesz = get_te32(&phdr->p_filesz);
            max_LOADsz = UPX_MAX(max_LOADsz, filesz);
            max_offset = UPX_MAX(max_offset, filesz + p_offset);
        }
    }
    if (canUnpack()) {
        throwAlreadyPacked();
    }
    // Heuristic for lopped trailing PackHeader (packed and "hacked"!)
    if (3 == e_phnum  // not shlib: PT_LOAD.C_BASE, PT_LOAD.C_TEXT, PT_GNU_STACK
    && UPX_MAGIC_LE32 == get_le32(&((l_info *)&phdri[e_phnum])->l_magic)) {
        throwAlreadyPacked();
    }
    // We want to compress position-independent executable (gcc -pie)
    // main programs, but compressing a shared library must be avoided
    // because the result is no longer usable.  In theory, there is no way
    // to tell them apart: both are just ET_DYN.  Also in theory,
    // neither the presence nor the absence of any particular symbol name
    // can be used to tell them apart; there are counterexamples.
    // However, we will use the following heuristic suggested by
    // Peter S. Mazinger <ps.m@gmx.net> September 2005:
    // If a ET_DYN has __libc_start_main as a global undefined symbol,
    // then the file is a position-independent executable main program
    // (that depends on libc.so.6) and is eligible to be compressed.
    // Otherwise (no __libc_start_main as global undefined): skip it.
    // Also allow  __uClibc_main  and  __uClibc_start_main .

    if (Elf32_Ehdr::ET_DYN==get_te16(&ehdr->e_type)) {
        // The DT_SYMTAB has no designated length.  Read the whole file.
        alloc_file_image(file_image, file_size);
        fi->seek(0, SEEK_SET);
        fi->readx(file_image, file_size);
        memcpy(&ehdri, ehdr, sizeof(Elf32_Ehdr));
        phdri= (Elf32_Phdr *)((size_t)e_phoff + file_image);  // do not free() !!
        shdri= (Elf32_Shdr *)((size_t)e_shoff + file_image);  // do not free() !!

        sec_strndx = nullptr;
        shstrtab = nullptr;
        if (e_shnum) {
            e_shstrndx = get_te16(&ehdr->e_shstrndx);
            if (e_shstrndx) {
                if (e_shnum <= e_shstrndx) {
                    char msg[40]; snprintf(msg, sizeof(msg),
                        "bad e_shstrndx %#x >= e_shnum %d", e_shstrndx, e_shnum);
                    throwCantPack(msg);
                }
                sec_strndx = &shdri[e_shstrndx];
                unsigned const sh_offset = get_te32(&sec_strndx->sh_offset);
                if ((u32_t)file_size <= sh_offset) {
                    char msg[50]; snprintf(msg, sizeof(msg),
                        "bad .e_shstrndx->sh_offset %#x", sh_offset);
                    throwCantPack(msg);
                }
                shstrtab = (char const *)(sh_offset + file_image);
            }
            sec_dynsym = elf_find_section_type(Elf32_Shdr::SHT_DYNSYM);
            if (sec_dynsym) {
                unsigned const sh_link = get_te32(&sec_dynsym->sh_link);
                if (e_shnum <= sh_link) {
                    char msg[50]; snprintf(msg, sizeof(msg),
                        "bad SHT_DYNSYM.sh_link %#x", sh_link);
                }
                sec_dynstr = &shdri[sh_link];
            }

            if (sec_strndx) {
                unsigned const sh_name = get_te32(&sec_strndx->sh_name);
                if (Elf32_Shdr::SHT_STRTAB != get_te32(&sec_strndx->sh_type)
                || (u32_t)file_size <= (sizeof(".shstrtab")
                    + sh_name + (shstrtab - (const char *)&file_image[0]))
                || (sh_name
                  && 0!=strcmp((char const *)".shstrtab", &shstrtab[sh_name]))
                ) {
                    throwCantPack("bad e_shstrtab");
                }
            }
        }

        Elf32_Phdr const *pload_x0(nullptr);  // first eXecutable PT_LOAD
        phdr= phdri;
        for (int j= e_phnum; --j>=0; ++phdr)
        if (Elf32_Phdr::PT_DYNAMIC==get_te32(&phdr->p_type)) {
            unsigned offset = check_pt_dynamic(phdr);
            dynseg= (Elf32_Dyn *)(offset + file_image);
            invert_pt_dynamic(dynseg,
                umin(get_te32(&phdr->p_filesz), (unsigned)(file_size_u - offset)));
        }
        else if (is_LOAD(phdr)) {
            if (!pload_x0
            &&  Elf32_Phdr::PF_X & get_te32(&phdr->p_flags)
            ) {
                pload_x0 = phdr;
            }
            check_pt_load(phdr);
        }
        if (!pload_x0) {
            throwCantPack("No PT_LOAD has (p_flags & PF_X)");
        }
        // elf_find_dynamic() returns 0 if 0==dynseg.
        dynstr=          (char const *)elf_find_dynamic(Elf32_Dyn::DT_STRTAB);
        dynsym= (Elf32_Sym /*const*/ *)elf_find_dynamic(Elf32_Dyn::DT_SYMTAB);

        if (opt->o_unix.force_pie
        ||      Elf32_Dyn::DF_1_PIE & elf_unsigned_dynamic(Elf32_Dyn::DT_FLAGS_1)
        ||  calls_crt1((Elf32_Rel const *)elf_find_dynamic(Elf32_Dyn::DT_REL),
                                 (int)elf_unsigned_dynamic(Elf32_Dyn::DT_RELSZ))
        ||  calls_crt1((Elf32_Rel const *)elf_find_dynamic(Elf32_Dyn::DT_JMPREL),
                                 (int)elf_unsigned_dynamic(Elf32_Dyn::DT_PLTRELSZ))) {
            is_pie = true;  // UNUSED except to ignore is_shlib and xct_off
            goto proceed;  // calls C library init for main program
        }

        if (Elf32_Ehdr::EM_ARM==get_te16(&ehdri.e_machine)) {
            sec_arm_attr = elf_find_section_type(Elf32_Shdr::SHT_ARM_ATTRIBUTES);
            if (Elf32_Ehdr::ET_DYN == e_type) {
                // See earlier comment in this function.  is_asl does not work.
                //is_asl = (!!saved_opt_android_shlib) << 1;  // bit 1; see is_shlib
                is_asl = 0;
            }
        }

        // Heuristic HACK for shared libraries (compare Darwin (MacOS) Dylib.)
        // If there is an existing DT_INIT, and if everything that the dynamic
        // linker ld-linux needs to perform relocations before calling DT_INIT
        // resides below the first SHT_EXECINSTR Section in one PT_LOAD, then
        // compress from the first executable Section to the end of that PT_LOAD.
        // We must not alter anything that ld-linux might touch before it calls
        // the DT_INIT function.
        //
        // Obviously this hack requires that the linker script put pieces
        // into good positions when building the original shared library,
        // and also requires ld-linux to behave.

        if (/*jni_onload_sym ||*/ elf_find_dynamic(upx_dt_init)) {
            if (this->e_machine!=Elf32_Ehdr::EM_386
            &&  this->e_machine!=Elf32_Ehdr::EM_MIPS
            &&  this->e_machine!=Elf32_Ehdr::EM_ARM)
                goto abandon;  // need stub: EM_PPC
            if (elf_has_dynamic(Elf32_Dyn::DT_TEXTREL)) {
                throwCantPack("DT_TEXTREL found; re-compile with -fPIC");
                goto abandon;
            }
            if (!(Elf32_Dyn::DF_1_PIE & elf_unsigned_dynamic(Elf32_Dyn::DT_FLAGS_1))) {
                // not explicitly PIE main program
                if (Elf32_Ehdr::EM_ARM == e_machine  // Android is common
                &&  0 && !is_asl  // but not explicit
                ) {
                    opt->info_mode++;
                    info("note: use --android-shlib if appropriate");
                    opt->info_mode--;
                }
            }
            if (Elf32_Ehdr::EM_MIPS == get_te16(&ehdr->e_machine)
            ||  Elf32_Ehdr::EM_PPC  == get_te16(&ehdr->e_machine)) {
                throwCantPack("This test UPX cannot pack .so for MIPS or PowerPC; coming soon.");
            }
            xct_va = ~(upx_uint64_t)0;
            if (e_shnum) {
                xct_va = canPack_Shdr(pload_x0);
            }
            else { // no Sections; use heuristics
                unsigned const strsz  = elf_unsigned_dynamic(Elf32_Dyn::DT_STRSZ);
                unsigned const strtab = elf_unsigned_dynamic(Elf32_Dyn::DT_STRTAB);
                unsigned const relsz  = elf_unsigned_dynamic(Elf32_Dyn::DT_RELSZ);
                unsigned const rel    = elf_unsigned_dynamic(Elf32_Dyn::DT_REL);
                unsigned const init   = elf_unsigned_dynamic(upx_dt_init);
                if ((init == (relsz + rel   ) && rel    == (strsz + strtab))
                ||  (init == (strsz + strtab) && strtab == (relsz + rel   ))
                ) {
                    xct_va = init;
                    user_init_va = init;
                    user_init_off = elf_get_offset_from_address(init);
                }
            }
            // Rely on 0==elf_unsigned_dynamic(tag) if no such tag.
            unsigned const va_gash = elf_unsigned_dynamic(Elf32_Dyn::DT_GNU_HASH);
            unsigned const va_hash = elf_unsigned_dynamic(Elf32_Dyn::DT_HASH);
            unsigned y = 0;
            if ((y=1, xct_va < va_gash)  ||  (y=2, (0==va_gash && xct_va < va_hash))
            ||  (y=3, xct_va < elf_unsigned_dynamic(Elf32_Dyn::DT_STRTAB))
            ||  (y=4, xct_va < elf_unsigned_dynamic(Elf32_Dyn::DT_SYMTAB))
            ||  (y=5, xct_va < elf_unsigned_dynamic(Elf32_Dyn::DT_REL))
            ||  (y=6, xct_va < elf_unsigned_dynamic(Elf32_Dyn::DT_RELA))
            ||  (y=7, xct_va < elf_unsigned_dynamic(Elf32_Dyn::DT_JMPREL))
            ||  (y=8, xct_va < elf_unsigned_dynamic(Elf32_Dyn::DT_VERDEF))
            ||  (y=9, xct_va < elf_unsigned_dynamic(Elf32_Dyn::DT_VERSYM))
            ||  (y=10, xct_va < elf_unsigned_dynamic(Elf32_Dyn::DT_VERNEED)) ) {
                static char const *which[] = {
                    "unknown",
                    "DT_GNU_HASH",
                    "DT_HASH",
                    "DT_STRTAB",
                    "DT_SYMTAB",
                    "DT_REL",
                    "DT_RELA",
                    "DT_JMPREL",
                    "DT_VERDEF",
                    "DT_VERSYM",
                    "DT_VERNEED",
                };
                char buf[30]; snprintf(buf, sizeof(buf), "%s above stub", which[y]);
                throwCantPack(buf);
                goto abandon;
            }
            xct_off = elf_get_offset_from_address(xct_va);
            if (opt->debug.debug_level) {
                fprintf(stderr, "shlib canPack: xct_va=%#lx  xct_off=%#lx\n",
                    (long)xct_va, (long)xct_off);
            }
            goto proceed;  // But proper packing depends on checking xct_va.
        }
        else {
            throwCantPack("need DT_INIT; try \"void _init(void){}\"");
        }
abandon:
        return false;
proceed: ;
    }
    // XXX Theoretically the following test should be first,
    // but PackUnix::canPack() wants 0!=exetype ?
    if (!super::canPack())
        return false;
    assert(exetype == 1);
    exetype = 0;

    // set options
    // this->blocksize: avoid over-allocating.
    // (file_size - max_offset): debug info, non-globl symbols, etc.
    opt->o_unix.blocksize = blocksize = UPX_MAX(max_LOADsz, (unsigned)(file_size - max_offset));
    return true;
}

tribool PackLinuxElf64::canUnpack() // bool, except -1: format known, but not packed
{
    if (checkEhdr(&ehdri)) {
        return false;
    }
    if (get_te16(&ehdri.e_phnum) < 2) {
        throwCantUnpack("e_phnum must be >= 2");
    }
    if (Elf64_Ehdr::ET_DYN==get_te16(&ehdri.e_type)) {
        PackLinuxElf64help1(fi);
    }
    if (super::canUnpack()) {
        return true;
    }
    return false;
}

tribool PackLinuxElf64::canPack()
{
    union {
        unsigned char buf[MAX_ELF_HDR_64];
        //struct { Elf64_Ehdr ehdr; Elf64_Phdr phdr; } e;
    } u;
    COMPILE_TIME_ASSERT(sizeof(u) <= (2*1024))

    fi->readx(u.buf, sizeof(u.buf));
    fi->seek(0, SEEK_SET);
    Elf64_Ehdr const *const ehdr = (Elf64_Ehdr *) u.buf;

    // now check the ELF header
    if (checkEhdr(ehdr) != 0)
        return false;

    // additional requirements for linux/elf386
    if (get_te16(&ehdr->e_ehsize) != sizeof(*ehdr)) {
        throwCantPack("invalid Ehdr e_ehsize; try '--force-execve'");
        return false;
    }
    if (e_phoff != sizeof(*ehdr)) {// Phdrs not contiguous with Ehdr
        throwCantPack("non-contiguous Ehdr/Phdr; try '--force-execve'");
        return false;
    }

    upx_uint64_t max_LOADsz = 0, max_offset = 0;
    Elf64_Phdr const *phdr = phdri;
    for (unsigned j=0; j < e_phnum; ++phdr, ++j) {
        if (j > ((MAX_ELF_HDR_64 - sizeof(Elf64_Ehdr)) / sizeof(Elf64_Phdr))) {
            throwCantPack("too many ElfXX_Phdr; try '--force-execve'");
            return false;
        }
        if (is_LOAD(phdr)) {
            // The first PT_LOAD must cover the beginning of the file (0==p_offset).
            upx_uint64_t const p_offset = get_te64(&phdr->p_offset);
            if (1!= exetype) {
                exetype = 1;
                load_va = get_te64(&phdr->p_vaddr);  // class data member
                upx_uint64_t const off = ~page_mask & load_va;
                if (off && off == p_offset) { // specific hint
                    throwCantPack("Go-language PT_LOAD: try hemfix.c, or try '--force-execve'");
                    // Fixing it inside upx fails because packExtent() reads original file.
                    return false;
                }
                if (0 != p_offset) { // 1st PT_LOAD must cover Ehdr and Phdr
                    throwCantPack("first PT_LOAD.p_offset != 0; try '--force-execve'");
                    return false;
                }
                // FIXME: bad for shlib!
                hatch_off = ~3ul & (3+ get_te64(&phdr->p_memsz));
            }
            upx_uint64_t align = get_te64(&phdr->p_align);
            if (0x200000ul < align) align = 0x200000ul;  // UPX_MIN type problem
            max_LOADsz = UPX_MAX(max_LOADsz, align);
            upx_uint64_t filesz = get_te64(&phdr->p_filesz);
            max_LOADsz = UPX_MAX(max_LOADsz, filesz);
            max_offset = UPX_MAX(max_offset, filesz + p_offset);
        }
    }
    if (canUnpack()) {
        throwAlreadyPacked();
    }
    // Heuristic for lopped trailing PackHeader (packed and "hacked"!)
    if (3 == e_phnum  // not shlib: PT_LOAD.C_BASE, PT_LOAD.C_TEXT, PT_GNU_STACK
    && UPX_MAGIC_LE32 == get_le32(&((l_info *)&phdri[e_phnum])->l_magic)) {
        throwAlreadyPacked();
    }
    // We want to compress position-independent executable (gcc -pie)
    // main programs, but compressing a shared library must be avoided
    // because the result is no longer usable.  In theory, there is no way
    // to tell them apart: both are just ET_DYN.  Also in theory,
    // neither the presence nor the absence of any particular symbol name
    // can be used to tell them apart; there are counterexamples.
    // However, we will use the following heuristic suggested by
    // Peter S. Mazinger <ps.m@gmx.net> September 2005:
    // If a ET_DYN has __libc_start_main as a global undefined symbol,
    // then the file is a position-independent executable main program
    // (that depends on libc.so.6) and is eligible to be compressed.
    // Otherwise (no __libc_start_main as global undefined): skip it.
    // Also allow  __uClibc_main  and  __uClibc_start_main .

    if (Elf64_Ehdr::ET_DYN==get_te16(&ehdr->e_type)) {
        // The DT_SYMTAB has no designated length.  Read the whole file.
        alloc_file_image(file_image, file_size);
        fi->seek(0, SEEK_SET);
        fi->readx(file_image, file_size);
        memcpy(&ehdri, ehdr, sizeof(Elf64_Ehdr));
        phdri= (Elf64_Phdr *)((size_t)e_phoff + file_image);  // do not free() !!
        shdri= (Elf64_Shdr *)((size_t)e_shoff + file_image);  // do not free() !!

        sec_strndx = nullptr;
        shstrtab = nullptr;
        if (e_shnum) {
            e_shstrndx = get_te16(&ehdr->e_shstrndx);
            if (e_shstrndx) {
                if (e_shnum <= e_shstrndx) {
                    char msg[40]; snprintf(msg, sizeof(msg),
                        "bad e_shstrndx %#x >= e_shnum %d", e_shstrndx, e_shnum);
                    throwCantPack(msg);
                }
                sec_strndx = &shdri[e_shstrndx];
                upx_uint64_t sh_offset = get_te64(&sec_strndx->sh_offset);
                if ((u64_t)file_size <= sh_offset) {
                    char msg[50]; snprintf(msg, sizeof(msg),
                        "bad .e_shstrndx->sh_offset %#lx", (long unsigned)sh_offset);
                    throwCantPack(msg);
                }
                shstrtab = (char const *)(sh_offset + file_image);
            }
            sec_dynsym = elf_find_section_type(Elf64_Shdr::SHT_DYNSYM);
            if (sec_dynsym) {
                unsigned const sh_link = get_te32(&sec_dynsym->sh_link);
                if (e_shnum <= sh_link) {
                    char msg[50]; snprintf(msg, sizeof(msg),
                        "bad SHT_DYNSYM.sh_link %#x", sh_link);
                }
                sec_dynstr = &shdri[sh_link];
            }

            if (sec_strndx) {
                unsigned const sh_name = get_te32(&sec_strndx->sh_name);
                if (Elf64_Shdr::SHT_STRTAB != get_te32(&sec_strndx->sh_type)
                || (u32_t)file_size <= (sizeof(".shstrtab")
                    + sh_name + (shstrtab - (const char *)&file_image[0]))
                || (sh_name
                  && 0!=strcmp((char const *)".shstrtab", &shstrtab[sh_name]))
                ) {
                    throwCantPack("bad e_shstrtab");
                }
            }
        }

        Elf64_Phdr const *pload_x0(nullptr);  // first eXecutable PT_LOAD
        phdr= phdri;
        for (int j= e_phnum; --j>=0; ++phdr)
        if (Elf64_Phdr::PT_DYNAMIC==get_te32(&phdr->p_type)) {
            upx_uint64_t offset = check_pt_dynamic(phdr);
            dynseg= (Elf64_Dyn *)(offset + file_image);
            invert_pt_dynamic(dynseg,
                umin(get_te64(&phdr->p_filesz), file_size - offset));
        }
        else if (is_LOAD(phdr)) {
            if (!pload_x0
            &&  Elf64_Phdr::PF_X & get_te32(&phdr->p_flags)
            ) {
                pload_x0 = phdr;
            }
            check_pt_load(phdr);
        }
        if (!pload_x0) {
            throwCantPack("No PT_LOAD has (p_flags & PF_X)");
        }
        // elf_find_dynamic() returns 0 if 0==dynseg.
        dynstr=          (char const *)elf_find_dynamic(Elf64_Dyn::DT_STRTAB);
        dynsym= (Elf64_Sym /*const*/ *)elf_find_dynamic(Elf64_Dyn::DT_SYMTAB);

        if (opt->o_unix.force_pie
        ||       Elf64_Dyn::DF_1_PIE & elf_unsigned_dynamic(Elf64_Dyn::DT_FLAGS_1)
        ||  calls_crt1((Elf64_Rela const *)elf_find_dynamic(Elf64_Dyn::DT_RELA),
                                  (int)elf_unsigned_dynamic(Elf64_Dyn::DT_RELASZ))
        ||  calls_crt1((Elf64_Rela const *)elf_find_dynamic(Elf64_Dyn::DT_JMPREL),
                                  (int)elf_unsigned_dynamic(Elf64_Dyn::DT_PLTRELSZ))) {
            is_pie = true;  // UNUSED except to ignore is_shlib and xct_off
            goto proceed;  // calls C library init for main program
        }

        if (Elf64_Ehdr::EM_ARM==get_te16(&ehdri.e_machine)) {
            sec_arm_attr = elf_find_section_type(Elf64_Shdr::SHT_ARM_ATTRIBUTES);
            if (Elf64_Ehdr::ET_DYN == e_type) {
                // See comment in Elf32_Linux::canPack().  is_asl does not work.
                // 64-bit ARM (aarch64) also has no ARM_ATTRIBUTES.
                //is_asl = (!!saved_opt_android_shlib) << 1;  // bit 1; see is_shlib
                is_asl = 0;
            }
        }

        // Heuristic HACK for shared libraries (compare Darwin (MacOS) Dylib.)
        // If there is an existing DT_INIT, and if everything that the dynamic
        // linker ld-linux needs to perform relocations before calling DT_INIT
        // resides below the first SHT_EXECINSTR Section in one PT_LOAD, then
        // compress from the first executable Section to the end of that PT_LOAD.
        // We must not alter anything that ld-linux might touch before it calls
        // the DT_INIT function.
        //
        // Obviously this hack requires that the linker script put pieces
        // into good positions when building the original shared library,
        // and also requires ld-linux to behave.

        if (/*jni_onload_sym ||*/ elf_find_dynamic(upx_dt_init)) {
            if (elf_has_dynamic(Elf64_Dyn::DT_TEXTREL)) {
                throwCantPack("DT_TEXTREL found; re-compile with -fPIC");
                goto abandon;
            }
            if (!(Elf64_Dyn::DF_1_PIE & elf_unsigned_dynamic(Elf64_Dyn::DT_FLAGS_1))) {
                // not explicitly PIE main program
                if (Elf64_Ehdr::EM_AARCH64 == e_machine  // Android is common
                &&  !is_asl  // but not explicit
                ) {
                    opt->info_mode++;
                    info("note: use --android-shlib if appropriate");
                    opt->info_mode--;
                }
            }
            if (Elf64_Ehdr::EM_PPC64 == get_te16(&ehdr->e_machine)) {
                throwCantPack("This test UPX cannot pack .so for PowerPC64; coming soon.");
            }
            xct_va = ~(upx_uint64_t)0;
            if (e_shnum) {
                xct_va = canPack_Shdr(pload_x0);
            }
            else { // no Sections; use heuristics
                upx_uint64_t const strsz  = elf_unsigned_dynamic(Elf64_Dyn::DT_STRSZ);
                upx_uint64_t const strtab = elf_unsigned_dynamic(Elf64_Dyn::DT_STRTAB);
                upx_uint64_t const relsz  = elf_unsigned_dynamic(Elf64_Dyn::DT_RELSZ);
                upx_uint64_t const rel    = elf_unsigned_dynamic(Elf64_Dyn::DT_REL);
                upx_uint64_t const init   = elf_unsigned_dynamic(upx_dt_init);
                if ((init == (relsz + rel   ) && rel    == (strsz + strtab))
                ||  (init == (strsz + strtab) && strtab == (relsz + rel   ))
                ) {
                    xct_va = init;
                    user_init_va = init;
                    user_init_off = elf_get_offset_from_address(init);
                }
            }
            // Rely on 0==elf_unsigned_dynamic(tag) if no such tag.
            upx_uint64_t const va_gash = elf_unsigned_dynamic(Elf64_Dyn::DT_GNU_HASH);
            upx_uint64_t const va_hash = elf_unsigned_dynamic(Elf64_Dyn::DT_HASH);
            unsigned y = 0;
            if ((y=1, xct_va < va_gash)  ||  (y=2, (0==va_gash && xct_va < va_hash))
            ||  (y=3, xct_va < elf_unsigned_dynamic(Elf64_Dyn::DT_STRTAB))
            ||  (y=4, xct_va < elf_unsigned_dynamic(Elf64_Dyn::DT_SYMTAB))
            ||  (y=5, xct_va < elf_unsigned_dynamic(Elf64_Dyn::DT_REL))
            ||  (y=6, xct_va < elf_unsigned_dynamic(Elf64_Dyn::DT_RELA))
            ||  (y=7, xct_va < elf_unsigned_dynamic(Elf64_Dyn::DT_JMPREL))
            ||  (y=8, xct_va < elf_unsigned_dynamic(Elf64_Dyn::DT_VERDEF))
            ||  (y=9, xct_va < elf_unsigned_dynamic(Elf64_Dyn::DT_VERSYM))
            ||  (y=10, xct_va < elf_unsigned_dynamic(Elf64_Dyn::DT_VERNEED)) ) {
                static char const *which[] = {
                    "unknown",
                    "DT_GNU_HASH",
                    "DT_HASH",
                    "DT_STRTAB",
                    "DT_SYMTAB",
                    "DT_REL",
                    "DT_RELA",
                    "DT_JMPREL",
                    "DT_VERDEF",
                    "DT_VERSYM",
                    "DT_VERNEED",
                };
                char buf[30]; snprintf(buf, sizeof(buf), "%s above stub", which[y]);
                throwCantPack(buf);
                goto abandon;
            }
            xct_off = elf_get_offset_from_address(xct_va);
            if (opt->debug.debug_level) {
                fprintf(stderr, "shlib canPack: xct_va=%#lx  xct_off=%#lx\n",
                    (long)xct_va, (long)xct_off);
            }
            goto proceed;  // But proper packing depends on checking xct_va.
        }
        else {
            throwCantPack("need DT_INIT; try \"void _init(void){}\"");
        }
abandon:
        return false;
proceed: ;
    }
    // XXX Theoretically the following test should be first,
    // but PackUnix::canPack() wants 0!=exetype ?
    if (!super::canPack())
        return false;
    assert(exetype == 1);
    exetype = 0;

    // set options
    // this->blocksize: avoid over-allocating.
    // (file_size - max_offset): debug info, non-globl symbols, etc.
    opt->o_unix.blocksize = blocksize = UPX_MAX(max_LOADsz, file_size - max_offset);
    return true;
}

off_t
PackLinuxElf32::getbrk(Elf32_Phdr const *phdr, int nph) const
{
    off_t brka = 0;
    for (int j = 0; j < nph; ++phdr, ++j) {
        if (is_LOAD(phdr)) {
            off_t b = get_te32(&phdr->p_vaddr) + get_te32(&phdr->p_memsz);
            if (b > brka)
                brka = b;
        }
    }
    return brka;
}

off_t
PackLinuxElf32::getbase(const Elf32_Phdr *phdr, int nph) const
{
    off_t base = ~0u;
    for (int j = 0; j < nph; ++phdr, ++j) {
        if (is_LOAD(phdr)) {
            unsigned const vaddr = get_te32(&phdr->p_vaddr);
            if (vaddr < (unsigned) base)
                base = vaddr;
        }
    }
    if (0!=base) {
        return base;
    }
    return 0x12000;
}

off_t
PackLinuxElf64::getbrk(const Elf64_Phdr *phdr, int nph) const
{
    off_t brka = 0;
    for (int j = 0; j < nph; ++phdr, ++j) {
        if (is_LOAD(phdr)) {
            off_t b = get_te64(&phdr->p_vaddr) + get_te64(&phdr->p_memsz);
            if (b > brka)
                brka = b;
        }
    }
    return brka;
}

static unsigned
is_pow2(unsigned x)
{
    return !((-1 + x) & x);
}

void
PackLinuxElf32::generateElfHdr(
    OutputFile *fo,
    void const *proto,
    unsigned const brka
)
{
    cprElfHdr2 *const h2 = (cprElfHdr2 *)(void *)&elfout;
    cprElfHdr3 *const h3 = (cprElfHdr3 *)(void *)&elfout;
    h3->ehdr =         ((cprElfHdr3 const *)proto)->ehdr;
    e_type = get_te16(&h3->ehdr.e_type);
    if (!memcmp("\x7f\x45\x4c\x46", proto, 4) && Elf32_Ehdr::ET_EXEC == e_type
    &&  2 <= get_te16(&h3->ehdr.e_phnum)) {
        h3->phdr[C_BASE] = ((cprElfHdr3 const *)proto)->phdr[1];  // .data; .p_align
        h3->phdr[C_TEXT] = ((cprElfHdr3 const *)proto)->phdr[0];  // .text
    }
    else  {
        memcpy(&h3->ehdr.e_ident[0], "\x7f\x45\x4c\x46", 4);
        h3->ehdr.e_ident[EI_CLASS] = ei_class;
        h3->ehdr.e_ident[EI_DATA] = ei_data;
        h3->ehdr.e_ident[EI_VERSION] = EV_CURRENT;
        h3->ehdr.e_ident[EI_OSABI] = ei_osabi;

        h3->ehdr.e_ident[EI_ABIVERSION] = ehdri.e_ident[EI_ABIVERSION];  // ei_abiversion ?
        // Perhaps NT_GNU_ABI_TAG from Shdr .note.ABI-tag  or PT_NOTE ?
        // But that is much too recent (and un-standardized) for other software.

        set_te32(&h3->ehdr.e_phoff, sizeof(Elf32_Ehdr));
        set_te16(&h3->ehdr.e_ehsize,sizeof(Elf32_Ehdr));
        set_te16(&h3->ehdr.e_phentsize, sizeof(Elf32_Phdr));
        set_te16(&h3->ehdr.e_phnum, 2);
        set_te16(&h3->ehdr.e_machine, e_machine);
        set_te16(&h3->ehdr.e_shstrndx, 0);
        memset(&h3->phdr[C_BASE], 0, sizeof(h3->phdr[C_BASE]));
        memset(&h3->phdr[C_TEXT], 0, sizeof(h3->phdr[C_TEXT]));
        memset(&h3->phdr[2     ], 0, sizeof(h3->phdr[2     ]));
        set_te32(&h3->phdr[C_BASE].p_flags, 0);
        set_te32(&h3->phdr[C_TEXT].p_flags, Elf32_Phdr::PF_X| Elf32_Phdr::PF_R);
        if (!memcmp("\x7f\x45\x4c\x46", proto, 4) && Elf32_Ehdr::ET_REL == e_type) {
        }
        else {
            throwCantPack("unknown e_type %#x", e_type);
        }
    }
    h3->ehdr.e_type = ehdri.e_type;  // ET_EXEC vs ET_DYN (gcc -pie -fPIC)
    memset(&h3->linfo, 0, sizeof(h3->linfo));

    h3->ehdr.e_ident[Elf32_Ehdr::EI_OSABI] = ei_osabi;
    if (Elf32_Ehdr::EM_MIPS==e_machine) { // MIPS R3000  FIXME
        h3->ehdr.e_ident[Elf32_Ehdr::EI_OSABI] = Elf32_Ehdr::ELFOSABI_NONE;
        h3->ehdr.e_flags = ehdri.e_flags;
    }

    if (ph.format != getFormat()) {
        assert(false);  // unknown ph.format, PackLinuxElf32
    }
    assert(get_te32(&h2->ehdr.e_phoff)     == sizeof(Elf32_Ehdr));
    assert(get_te16(&h2->ehdr.e_ehsize)    == sizeof(Elf32_Ehdr));
    assert(get_te16(&h2->ehdr.e_phentsize) == sizeof(Elf32_Phdr));

    h2->ehdr.e_shoff = 0;
    set_te16(&h2->ehdr.e_shentsize, sizeof(Elf32_Shdr));  // libbfd-2.41-38.fc40
    if (o_elf_shnum) {
        h2->ehdr.e_shnum = o_elf_shnum;
        h2->ehdr.e_shstrndx = o_elf_shnum - 1;
    }
    else {
        // https://bugzilla.redhat.com/show_bug.cgi?id=2131609
        // 0==.e_shnum is a special case for libbfd
        // that requires 0==.e_shentsize in order to force "no Shdrs"
        // But qemu 8.2.9 with libbfd-2.41-38.fc40  says EXEC format error
        // and uses 0==.e_shoff instead.
        // h2->ehdr.e_shentsize = 0;
        h2->ehdr.e_shnum = 0;
        h2->ehdr.e_shstrndx = 0;
    }

    unsigned phnum_o = 2 + n_phdrx;  // C_BASE, C_TEXT
    set_te16(&h2->ehdr.e_phnum, phnum_o);

    // Info for OS kernel to set the brk()
    if (brka) {
        // linux-2.6.14 binfmt_elf.c: SIGKILL if (0==.p_memsz) on a page boundary
        upx_uint32_t lo_va_user = ~0u;  // infinity
        for (int j= e_phnum; --j>=0; ) {
            if (is_LOAD(&phdri[j])) {
                upx_uint32_t const vaddr = get_te32(&phdri[j].p_vaddr);
                lo_va_user = umin(lo_va_user, vaddr);
            }
        }
        set_te32(                 &h2->phdr[C_BASE].p_vaddr, lo_va_user);
        h2->phdr[C_BASE].p_paddr = h2->phdr[C_BASE].p_vaddr;
        h2->phdr[C_TEXT].p_vaddr = h2->phdr[C_BASE].p_vaddr;
        h2->phdr[C_TEXT].p_paddr = h2->phdr[C_BASE].p_vaddr;
        set_te32(&h2->phdr[C_BASE].p_type, PT_LOAD);  // be sure
        set_te32(&h2->phdr[C_TEXT].p_type, PT_LOAD);  // be sure
        h2->phdr[C_BASE].p_offset = 0;
        h2->phdr[C_BASE].p_filesz = 0;
        // .p_memsz = brka;  temporary until sz_pack2
        set_te32(&h2->phdr[C_BASE].p_memsz, brka - lo_va_user);
        set_te32(&h2->phdr[C_BASE].p_flags, Elf32_Phdr::PF_R | Elf32_Phdr::PF_W);
    }
    set_te32(&h3->phdr[C_BASE].p_align, page_size);
    set_te32(&h3->phdr[C_TEXT].p_align, page_size);
    set_te32(&h2->phdr[C_TEXT].p_flags, ~Elf32_Phdr::PF_W & get_te32(&h2->phdr[C_TEXT].p_flags));
    fo->write(h2, sizeof(Elf32_Ehdr) + 2* sizeof(Elf32_Phdr));  // C_BASE, C_TEXT

    u64_t const zero_pad = 0;
    unsigned uva0 = get_te32(&h2->phdr[C_TEXT].p_vaddr);
    unsigned off_o = 0;
    for (unsigned phase = 0; phase < 2; ++phase) { // 0: Phdrs;  1: bodies
        off_o = sizeof(Elf32_Ehdr) + phnum_o * sizeof(Elf32_Phdr);
        for (unsigned j = 0; j < n_phdrx; ++j) {
            Elf32_Phdr phtmp = *phdrx[j];
            phtmp.p_vaddr = phtmp.p_paddr = 0;
            unsigned const off_i = get_te32(&phtmp.p_offset);
            unsigned const align = get_te32(&phtmp.p_align);
            unsigned const filesz = get_te32(&phtmp.p_filesz);
            unsigned pad = 0;
            if (filesz && align && is_pow2(align)) {
                unsigned alm1 = -1 + umin(8u, align);
                pad = alm1 & (0u - off_o);
                off_o += pad;
                set_te32(&phtmp.p_vaddr, off_o + uva0); phtmp.p_paddr = phtmp.p_vaddr;
            }
            if (0==phase) {
                set_te32(&phtmp.p_offset, (filesz ? off_o : 0));
                fo->write(&phtmp, sizeof(phtmp));  // Phdr
            }
            if (1==phase) {
                if (pad) fo->write(&zero_pad, pad);
                if (filesz) fo->write(&file_image[off_i], filesz);  // body
            }
            off_o += filesz;
        }
    }
    off_o = fpad4(fo, off_o);
    sz_phdrx = off_o - (sizeof(Elf32_Ehdr) + phnum_o * sizeof(Elf32_Phdr));
    set_te32(&h2->phdr[C_BASE].p_filesz, off_o);

    sz_elf_hdrs = sizeof(Elf32_Ehdr) + phnum_o * sizeof(Elf32_Phdr) + sz_phdrx;
    overlay_offset = sz_elf_hdrs + sizeof(l_info);
    o_binfo        = sz_elf_hdrs + sizeof(l_info) + sizeof(p_info);

    l_info linfo2; memset(&linfo2, 0, sizeof(linfo2));
    fo->write(&linfo2, sizeof(linfo2));
}

void
PackNetBSDElf32x86::generateElfHdr(
    OutputFile *fo,
    void const *proto,
    unsigned const brka
)
{
    cprElfHdr2 *const h2 = (cprElfHdr2 *)(void *)&elfout;
    super::generateElfHdr(fo, proto, brka);

    sz_elf_hdrs = sizeof(*h2) - sizeof(linfo);
    unsigned note_offset = sz_elf_hdrs;

    // Find the NetBSD PT_NOTE and the PaX PT_NOTE.
    Elf32_Nhdr const *np_NetBSD = nullptr;  unsigned sz_NetBSD = 0;
    Elf32_Nhdr const *np_PaX    = nullptr;  unsigned sz_PaX    = 0;
    unsigned char *cp = (unsigned char *)note_body;
    unsigned j;
    for (j=0; j < note_size; ) {
        Elf32_Nhdr const *const np = (Elf32_Nhdr const *)(void *)cp;
        int k = sizeof(*np) + up4(get_te32(&np->namesz))
            + up4(get_te32(&np->descsz));

        if (NHDR_NETBSD_TAG == np->type && 7== np->namesz
        &&  NETBSD_DESCSZ == np->descsz
        &&  0==strcmp(ELF_NOTE_NETBSD_NAME,
                /* &np->body */ (char const *)(1+ np))) {
            np_NetBSD = np;
            sz_NetBSD = k;
        }
        if (NHDR_PAX_TAG == np->type && 4== np->namesz
        &&  PAX_DESCSZ==np->descsz
        &&  0==strcmp(ELF_NOTE_PAX_NAME,
                /* &np->body */ (char const *)(1+ np))) {
            np_PaX = np;
            sz_PaX = k;
        }
        cp += k;
        j += k;
    }

    // Add PT_NOTE for the NetBSD note and PaX note, if any.
    note_offset += (np_NetBSD ? sizeof(Elf32_Phdr) : 0);
    note_offset += (np_PaX    ? sizeof(Elf32_Phdr) : 0);
    Elf32_Phdr *phdr = &elfout.phdr[C_NOTE];
    if (np_NetBSD) {
        set_te32(&phdr->p_type, PT_NOTE32);
        set_te32(&phdr->p_offset, note_offset);
        set_te32(&phdr->p_vaddr, note_offset);
        set_te32(&phdr->p_paddr, note_offset);
        set_te32(&phdr->p_filesz, sz_NetBSD);
        set_te32(&phdr->p_memsz,  sz_NetBSD);
        set_te32(&phdr->p_flags, Elf32_Phdr::PF_R);
        set_te32(&phdr->p_align, 4);

        sz_elf_hdrs += sz_NetBSD + sizeof(*phdr);
        note_offset += sz_NetBSD;
        ++phdr;
    }
    if (np_PaX) {
        set_te32(&phdr->p_type, PT_NOTE32);
        set_te32(&phdr->p_offset, note_offset);
        set_te32(&phdr->p_vaddr, note_offset);
        set_te32(&phdr->p_paddr, note_offset);
        set_te32(&phdr->p_filesz, sz_PaX);
        set_te32(&phdr->p_memsz,  sz_PaX);
        set_te32(&phdr->p_flags, Elf32_Phdr::PF_R);
        set_te32(&phdr->p_align, 4);

        /* &np_PaX->body[4] */
        const unsigned char *p4 =  &(ACC_CCAST(const unsigned char *, (1+ np_PaX)))[4];
        unsigned bits = get_te32(p4);
        bits &= ~PAX_MPROTECT;
        bits |=  PAX_NOMPROTECT;
        set_te32(ACC_UNCONST_CAST(unsigned char *, p4), bits);

        sz_elf_hdrs += sz_PaX + sizeof(*phdr);
        note_offset += sz_PaX;
        ++phdr;
    }
    set_te32(&h2->phdr[C_TEXT].p_filesz, note_offset);
              h2->phdr[C_TEXT].p_memsz = h2->phdr[C_TEXT].p_filesz;

    if (ph.format==getFormat()) {
        set_te16(&h2->ehdr.e_phnum, !!sz_NetBSD + !!sz_PaX +
        get_te16(&h2->ehdr.e_phnum));
        fo->seek(0, SEEK_SET);
        fo->rewrite(h2, sizeof(*h2) - sizeof(h2->linfo));

        // The 'if' guards on these two calls to memcpy are required
        // because the C Standard Committee did not debug the Standard
        // before publishing.  An empty region (0==size) must nevertheless
        // have a valid (non-nullptr) pointer.
        if (sz_NetBSD) memcpy(&((char *)phdr)[0],         np_NetBSD, sz_NetBSD);
        if (sz_PaX)    memcpy(&((char *)phdr)[sz_NetBSD], np_PaX,    sz_PaX);

        fo->write(&elfout.phdr[C_NOTE],
            &((char *)phdr)[sz_PaX + sz_NetBSD] - (char *)&elfout.phdr[C_NOTE]);

        l_info foo; memset(&foo, 0, sizeof(foo));
        fo->rewrite(&foo, sizeof(foo));
    }
    else {
        assert(false);  // unknown ph.format, PackLinuxElf32
    }
}

void
PackOpenBSDElf32x86::generateElfHdr(
    OutputFile *fo,
    void const *proto,
    unsigned const brka
)
{
    cprElfHdr3 *const h3 = (cprElfHdr3 *)(void *)&elfout;
    memcpy(h3, proto, sizeof(*h3));  // reads beyond, but OK
    h3->ehdr.e_ident[Elf32_Ehdr::EI_OSABI] = ei_osabi;
    assert(2==get_te16(&h3->ehdr.e_phnum));
    set_te16(&h3->ehdr.e_phnum, 3);

    assert(get_te32(&h3->ehdr.e_phoff)     == sizeof(Elf32_Ehdr));
                         h3->ehdr.e_shoff = 0;
    assert(get_te16(&h3->ehdr.e_ehsize)    == sizeof(Elf32_Ehdr));
    assert(get_te16(&h3->ehdr.e_phentsize) == sizeof(Elf32_Phdr));
    set_te16(&h3->ehdr.e_shentsize, sizeof(Elf32_Shdr));  // libbfd-2.41-38.fc40
    h3->ehdr.e_shnum = 0;
    h3->ehdr.e_shstrndx = 0;

    struct {
        Elf32_Nhdr nhdr;
        char name[8];
        unsigned body;
    } elfnote;

    unsigned const note_offset = sizeof(*h3) - sizeof(linfo);
    sz_elf_hdrs = sizeof(elfnote) + note_offset;

    set_te32(&h3->phdr[C_NOTE].p_type, PT_NOTE32);
    set_te32(&h3->phdr[C_NOTE].p_offset, note_offset);
    set_te32(&h3->phdr[C_NOTE].p_vaddr, note_offset);
    set_te32(&h3->phdr[C_NOTE].p_paddr, note_offset);
    set_te32(&h3->phdr[C_NOTE].p_filesz, sizeof(elfnote));
    set_te32(&h3->phdr[C_NOTE].p_memsz,  sizeof(elfnote));
    set_te32(&h3->phdr[C_NOTE].p_flags, Elf32_Phdr::PF_R);
    set_te32(&h3->phdr[C_NOTE].p_align, 4);

    // Q: Same as this->note_body[0 .. this->note_size-1] ?
    set_te32(&elfnote.nhdr.namesz, 8);
    set_te32(&elfnote.nhdr.descsz, OPENBSD_DESCSZ);
    set_te32(&elfnote.nhdr.type,   NHDR_OPENBSD_TAG);
    memcpy(elfnote.name, "OpenBSD", sizeof(elfnote.name));
    elfnote.body = 0;

    set_te32(&h3->phdr[C_TEXT].p_filesz, sz_elf_hdrs);
              h3->phdr[C_TEXT].p_memsz = h3->phdr[C_TEXT].p_filesz;

    unsigned const brkb = brka | ((0==(~page_mask & brka)) ? 0x20 : 0);
    set_te32(&h3->phdr[C_BASE].p_type, PT_LOAD);  // be sure
    set_te32(&h3->phdr[C_BASE].p_offset, ~page_mask & brkb);
    set_te32(&h3->phdr[C_BASE].p_vaddr, brkb);
    set_te32(&h3->phdr[C_BASE].p_paddr, brkb);
    h3->phdr[C_BASE].p_filesz = 0;
    // Too many kernels have bugs when 0==.p_memsz
    set_te32(&h3->phdr[C_BASE].p_memsz, 1);
    set_te32(&h3->phdr[C_BASE].p_flags, Elf32_Phdr::PF_R | Elf32_Phdr::PF_W);

    if (ph.format==getFormat()) {
        memset(&h3->linfo, 0, sizeof(h3->linfo));
        fo->write(h3, sizeof(*h3) - sizeof(h3->linfo));
        fo->write(&elfnote, sizeof(elfnote));
        fo->write(&h3->linfo, sizeof(h3->linfo));
    }
    else {
        assert(false);  // unknown ph.format, PackLinuxElf32
    }
}

void
PackLinuxElf64::generateElfHdr(
    OutputFile *fo,
    void const *proto,
    unsigned const brka
)
{
    cprElfHdr2 *const h2 = (cprElfHdr2 *)(void *)&elfout;
    cprElfHdr3 *const h3 = (cprElfHdr3 *)(void *)&elfout;
    h3->ehdr =         ((cprElfHdr3 const *)proto)->ehdr;
    if (Elf64_Ehdr::ET_REL == get_te16(&h3->ehdr.e_type)) {
        set_te64(&h3->ehdr.e_phoff, sizeof(Elf64_Ehdr));
        set_te16(&h3->ehdr.e_ehsize,sizeof(Elf64_Ehdr));
        set_te16(&h3->ehdr.e_phentsize, sizeof(Elf64_Phdr));
        set_te16(&h3->ehdr.e_phnum, 2);
        memset(&h3->phdr[C_BASE], 0, sizeof(h3->phdr[C_BASE]));
        memset(&h3->phdr[C_TEXT], 0, sizeof(h3->phdr[C_TEXT]));
        memset(&h3->phdr[2     ], 0, sizeof(h3->phdr[2     ]));
        set_te32(&h3->phdr[C_BASE].p_flags, 0);
        set_te32(&h3->phdr[C_TEXT].p_flags, Elf64_Phdr::PF_X| Elf64_Phdr::PF_R);
    }
    else {
        h3->phdr[C_BASE] = ((cprElfHdr3 const *)proto)->phdr[1];  // .data; .p_align
        h3->phdr[C_TEXT] = ((cprElfHdr3 const *)proto)->phdr[0];  // .text
    }
    memset(&h3->linfo, 0, sizeof(h3->linfo));

    h3->ehdr.e_type = ehdri.e_type;  // ET_EXEC vs ET_DYN (gcc -pie -fPIC)
    h3->ehdr.e_ident[Elf64_Ehdr::EI_OSABI] = ei_osabi;
    if (Elf64_Ehdr::ELFOSABI_LINUX == ei_osabi  // proper
    &&  Elf64_Ehdr::ELFOSABI_NONE  == ehdri.e_ident[Elf64_Ehdr::EI_OSABI]  // sloppy
    ) { // propagate sloppiness so that decompression does not complain
        h3->ehdr.e_ident[Elf64_Ehdr::EI_OSABI] = ehdri.e_ident[Elf64_Ehdr::EI_OSABI];
    }
    if (Elf64_Ehdr::EM_PPC64 == get_te16(&ehdri.e_machine)) {
        h3->ehdr.e_flags = ehdri.e_flags;  // "0x1, abiv1" vs "0x2, abiv2"
    }

    assert(get_te64(&h2->ehdr.e_phoff)     == sizeof(Elf64_Ehdr));
    assert(get_te16(&h2->ehdr.e_ehsize)    == sizeof(Elf64_Ehdr));
    assert(get_te16(&h2->ehdr.e_phentsize) == sizeof(Elf64_Phdr));

    h2->ehdr.e_shoff = 0;
    set_te16(&h2->ehdr.e_shentsize, sizeof(Elf64_Shdr));  // libbfd-2.41-38.fc40
    if (o_elf_shnum) {
        h2->ehdr.e_shnum = o_elf_shnum;
        h2->ehdr.e_shstrndx = o_elf_shnum - 1;
    }
    else {
        // https://bugzilla.redhat.com/show_bug.cgi?id=2131609
        // 0==.e_shnum is a special case for libbfd
        // that requires 0==.e_shentsize in order to force "no Shdrs"
        // But qemu 8.2.9 with libbfd-2.41-38.fc40  says EXEC format error
        // and uses 0==.e_shoff instead.
        // h2->ehdr.e_shentsize = 0;
        h2->ehdr.e_shnum = 0;
        h2->ehdr.e_shstrndx = 0;
    }

    unsigned const phnum_i = get_te16(&h2->ehdr.e_phnum);
    unsigned       phnum_o = 2 + n_phdrx;  // C_BASE, C_TEXT
    set_te16(&h2->ehdr.e_phnum, phnum_o);
    o_binfo =  sizeof(Elf64_Ehdr) + sizeof(Elf64_Phdr)*phnum_o + sizeof(l_info) + sizeof(p_info);
    set_te64(&h2->phdr[C_TEXT].p_filesz, sizeof(*h2));  // + identsize;
              h2->phdr[C_TEXT].p_memsz = h2->phdr[C_TEXT].p_filesz;
    set_te32(&h2->phdr[C_TEXT].p_type, PT_LOAD);  // be sure

    for (unsigned j=0; j < phnum_i; ++j) {
        if (is_LOAD(&h3->phdr[j])) {
            set_te64( &h3->phdr[j].p_align, page_size);
        }
    }

    // Info for OS kernel to set the brk()
    if (brka) {
        // linux-2.6.14 binfmt_elf.c: SIGKILL if (0==.p_memsz) on a page boundary
        upx_uint64_t lo_va_user(~(upx_uint64_t)0);  // infinity
        for (int j= e_phnum; --j>=0; ) {
            if (is_LOAD(&phdri[j])) {
                upx_uint64_t const vaddr = get_te64(&phdri[j].p_vaddr);
                lo_va_user = umin(lo_va_user, vaddr);
            }
        }
        set_te64(                 &h2->phdr[C_BASE].p_vaddr, lo_va_user);
        h2->phdr[C_BASE].p_paddr = h2->phdr[C_BASE].p_vaddr;
        h2->phdr[C_TEXT].p_vaddr = h2->phdr[C_BASE].p_vaddr;
        h2->phdr[C_TEXT].p_paddr = h2->phdr[C_BASE].p_vaddr;
        set_te32(&h2->phdr[C_BASE].p_type, PT_LOAD);  // be sure
        h2->phdr[C_BASE].p_offset = 0;
        h2->phdr[C_BASE].p_filesz = 0;
        // .p_memsz = brka;  temporary until sz_pack2
        set_te64(&h2->phdr[C_BASE].p_memsz, brka - lo_va_user);
        set_te32(&h2->phdr[C_BASE].p_flags, Elf64_Phdr::PF_R | Elf64_Phdr::PF_W);
    }
    set_te64(&h3->phdr[C_BASE].p_align, page_size);
    set_te64(&h3->phdr[C_TEXT].p_align, page_size);
    set_te32(&h2->phdr[C_TEXT].p_flags, ~Elf64_Phdr::PF_W & get_te32(&h2->phdr[C_TEXT].p_flags));
    fo->write(h2, sizeof(Elf64_Ehdr) + 2* sizeof(Elf64_Phdr));  // C_BASE, C_TEXT

    u64_t const zero_pad = 0;
    unsigned uva0 = get_te64(&h2->phdr[C_TEXT].p_vaddr);
    unsigned off_o = 0;
    for (unsigned phase = 0; phase < 2; ++phase) { // 0: Phdrs;  1: bodies
        off_o = sizeof(Elf64_Ehdr) + phnum_o * sizeof(Elf64_Phdr);
        for (unsigned j = 0; j < n_phdrx; ++j) {
            Elf64_Phdr phtmp = *phdrx[j];
            phtmp.p_vaddr = phtmp.p_paddr = 0;
            u64_t const off_i = get_te64(&phtmp.p_offset);
            u64_t const align = get_te64(&phtmp.p_align);
            u64_t const filesz = get_te64(&phtmp.p_filesz);
            unsigned pad = 0;
            if (filesz && align && is_pow2(align)) {
                unsigned alm1 = -1 + umin(8u, (unsigned)align);
                pad = alm1 & (0u - off_o);
                off_o += pad;
                set_te64(&phtmp.p_vaddr, off_o + uva0); phtmp.p_paddr = phtmp.p_vaddr;
            }
            if (0==phase) {
                set_te64(&phtmp.p_offset, (filesz ? off_o : 0));
                fo->write(&phtmp, sizeof(phtmp));  // Phdr
            }
            if (1==phase) {
                if (pad) fo->write(&zero_pad, pad);
                if (filesz) fo->write(&file_image[off_i], filesz);  // body
            }
            off_o += filesz;
        }
    }
    off_o = fpad4(fo, off_o);
    sz_phdrx = off_o - (sizeof(Elf64_Ehdr) + phnum_o * sizeof(Elf64_Phdr));
    set_te64(&h2->phdr[C_BASE].p_filesz, off_o);

    sz_elf_hdrs = sizeof(Elf64_Ehdr) + phnum_o * sizeof(Elf64_Phdr) + sz_phdrx;
    overlay_offset = sz_elf_hdrs + sizeof(l_info);
    o_binfo        = sz_elf_hdrs + sizeof(l_info) + sizeof(p_info);

    l_info linfo2; memset(&linfo2, 0, sizeof(linfo2));
    fo->write(&linfo2, sizeof(linfo2));
}

// Android shlib has ABS symbols that actually are relative.
static char const abs_symbol_names[][14] = {
      "__bss_end__"
    ,  "_bss_end__"
    , "__bss_start"
    , "__bss_start__"
    ,  "_edata"
    ,  "_end"
    , "__end__"
    , ""
};

int
PackLinuxElf32::adjABS(Elf32_Sym *sym, unsigned delta)
{
    unsigned st_name = get_te32(&sym->st_name);
    for (int j = 0; abs_symbol_names[j][0]; ++j) {
        if (!strcmp(abs_symbol_names[j], get_str_name(st_name, (unsigned)-1))) {
            sym->st_value += delta;
            return 1;
        }
    }
    return 0;
}

int
PackLinuxElf64::adjABS(Elf64_Sym *sym, unsigned long delta)
{
    unsigned st_name = get_te32(&sym->st_name);
    for (int j = 0; abs_symbol_names[j][0]; ++j) {
        if (!strcmp(abs_symbol_names[j], get_str_name(st_name, (unsigned)-1))) {
            sym->st_value += delta;
            return 1;
        }
    }
    return 0;
}

void PackLinuxElf32::pack1(OutputFile * /*fo*/, Filter &ft)
{
    fi->seek(0, SEEK_SET);
    fi->readx(&ehdri, sizeof(ehdri));
    assert(e_phoff == sizeof(Elf32_Ehdr));  // checked by canPack()
    sz_phdrs = e_phnum * get_te16(&ehdri.e_phentsize);

// We compress separate pieces (usually each PT_LOAD, plus the gaps in the file
// that are not covered by any PT_LOAD), but currently at run time there can be
// only one decompressor method.
// Therefore we must plan ahead because Packer::compressWithFilters tries
// to find the smallest result among the available methods, for one piece only.
// In the future we may allow more than one decompression method at run time.
// For now we must choose only one, and force PackUnix::packExtent
// (==> compressWithFilters) to use it.
    int nfilters = 0;
    {
        int const *fp = getFilters();
        while (FT_END != *fp++) {
            ++nfilters;
        }
    }
    {
        int npieces = 1;  // tail after highest PT_LOAD
        Elf32_Phdr *phdr = phdri;
        for (unsigned j=0; j < e_phnum; ++phdr, ++j) {
            if (is_LOAD(phdr)) {
                unsigned const  flags = get_te32(&phdr->p_flags);
                unsigned       offset = get_te32(&phdr->p_offset);
                if (!xct_off  // not shlib
                  // new-style shlib: PT_LOAD[0] has symbol table
                  // which must not be compressed, but also lacks PF_X
                ||    (Elf32_Phdr::PF_X & flags)
                  // Read-only, non-first PT_LOAD is _assumed_ to be compressible
                ||  (!(Elf32_Phdr::PF_W & flags) && 0!=offset))
                {
                    ++npieces;  // will attempt compression of this PT_LOAD
                }
            }
        }
        uip->ui_total_passes += npieces;
    }
    int methods[256];
    unsigned nmethods = prepareMethods(methods, ph.method, getCompressionMethods(M_ALL, ph.level));
    if (1 < nmethods) { // Many are available, but we must choose only one
        uip->ui_total_passes += 1;  // the batch for output
        uip->ui_total_passes *= nmethods * (1+ nfilters);  // finding smallest total
        PackHeader orig_ph = ph;
        Filter orig_ft = ft;
        unsigned max_offset = 0;
        unsigned sz_best= ~0u;
        int method_best = 0;
        for (unsigned k = 0; k < nmethods; ++k) { // FIXME: parallelize; cost: working space
            unsigned sz_this = 0;
            Elf32_Phdr *phdr = phdri;
            for (unsigned j=0; j < e_phnum; ++phdr, ++j) {
                if (is_LOAD(phdr)) {
                    unsigned const  flags = get_te32(&phdr->p_flags);
                    unsigned       offset = get_te32(&phdr->p_offset);
                    unsigned       filesz = get_te32(&phdr->p_filesz);
                    max_offset = UPX_MAX(max_offset, filesz + offset);
                    if (!xct_off  // not shlib
                      // new-style shlib: PT_LOAD[0] has symbol table
                      // which must not be compressed, but also lacks PF_X
                    ||    (Elf32_Phdr::PF_X & flags)
                      // Read-only, non-first PT_LOAD is _assumed_ to be compressible
                    ||  (!(Elf32_Phdr::PF_W & flags) && 0!=offset))
                    {
                        if (xct_off && 0==offset) { // old-style shlib
                            offset  = xct_off;
                            filesz -= xct_off;
                        }
                        fi->seek(offset, SEEK_SET);
                        fi->readx(ibuf, filesz);
                        ft = orig_ft;
                        ph = orig_ph;
                        ph.set_method(ph_force_method(methods[k]), offset);
                        ph.u_len = filesz;
                        compressWithFilters(&ft, OVERHEAD, NULL_cconf, 10, true);
                        sz_this += ph.c_len;
                    }
                }
            }
            unsigned const sz_tail = file_size - max_offset;  // debuginfo, etc.
            if (sz_tail) {
                fi->seek(max_offset, SEEK_SET);
                fi->readx(ibuf, sz_tail);
                ft = orig_ft;
                ph = orig_ph;
                ph.set_method(ph_force_method(methods[k]));
                ph.u_len = sz_tail;
                compressWithFilters(&ft, OVERHEAD, NULL_cconf, 10, true);
                sz_this += ph.c_len;
            }
            // FIXME: loader size also depends on method
            if (sz_best > sz_this) {
                sz_best = sz_this;
                method_best = methods[k];
            }
        }
        ft = orig_ft;
        ph = orig_ph;
        ph.set_method(ph_force_method(method_best));
    }

    Elf32_Phdr *phdr = phdri;
    for (unsigned j=0; j < e_phnum; ++phdr, ++j) {
        unsigned const type = get_te32(&phdr->p_type);
        if (PT_GNU_STACK32 == type || PT_NOTE32 == type) {
            add_phdrx(phdr);  // PT_GNU_STACK32, PT_NOTE32
        }
        if (PT_LOAD == type) {
            unsigned x = get_te32(&phdr->p_align) >> lg2_page;
            while (x>>=1) {
                ++lg2_page;
            }
        }
        if (PT_GNU_RELRO32 == type
        &&  16 < lg2_page) {
            // x86* allows 4K, 2M, 1G
            // arm32 and arm64 both allow 4K, 16K, 64K; Apple Silicon uses 16K
            // Oracle 5.4.17-2136.314.6.2.el8uek.aarch64 demands 64K
            // PowerPC and PowerPC64 has 4K, 64K, 1M (?), 16M (?)
            // MIPS  allows any power of 2 from 4K through 64K

            // If 64K < .p_align then we assume that 4K is also accepted.
            // .p_align can be like 2M, which is a huge over-estimate.
            // RELRO ends on a page boundary: usually close to actual page_size
            unsigned offset = get_te32(&phdr->p_offset);
            unsigned filesz = get_te32(&phdr->p_filesz);
            if (!(0xfff & (filesz + offset))) { // a 4KiB boundary
                unsigned b = 12;
                while (!(~(~0u << b) & (filesz + offset))) {
                    ++b;
                }
                lg2_page = umin(lg2_page, -1+ b);
            }
        }
    }
    page_size =  1u  <<lg2_page;
    page_mask = ~0ull<<lg2_page;

    progid = 0;  // getRandomId();  not useful, so do not clutter
    sz_elf_hdrs = sizeof(ehdri) + sz_phdrs;

    // only execute if option present
    if (opt->o_unix.preserve_build_id) {
        // set this so we can use elf_find_section_name
        e_shnum = get_te16(&ehdri.e_shnum);
        if (!shdri) {
            mb_shdr.alloc(e_shnum * sizeof(Elf32_Shdr));
            shdri = (Elf32_Shdr *)mb_shdr.getVoidPtr();
            e_shoff = get_te32(&ehdri.e_shoff);
            fi->seek(e_shoff, SEEK_SET);
            fi->readx(shdri, e_shnum * sizeof(Elf32_Shdr));
        }
        //set the shstrtab
        sec_strndx = &shdri[get_te16(&ehdri.e_shstrndx)];

        upx_uint32_t sh_size = get_te32(&sec_strndx->sh_size);
        mb_shstrtab.alloc(sh_size); shstrtab = (char *)mb_shstrtab.getVoidPtr();
        fi->seek(0,SEEK_SET);
        fi->seek(sec_strndx->sh_offset,SEEK_SET);
        fi->readx(mb_shstrtab, sh_size);

        Elf32_Shdr const *buildid = elf_find_section_name(".note.gnu.build-id");
        if (buildid) {
            unsigned bid_sh_size = get_te32(&buildid->sh_size);
            buildid_data.alloc(bid_sh_size);
            buildid_data.clear();
            fi->seek(0,SEEK_SET);
            fi->seek(buildid->sh_offset,SEEK_SET);
            fi->readx((void *)buildid_data, bid_sh_size);

            o_elf_shnum = 3;
            memset(&shdrout,0,sizeof(shdrout));

            //setup the build-id
            memcpy(&shdrout.shdr[1], buildid, sizeof(shdrout.shdr[1]));
            set_te32(&shdrout.shdr[1].sh_name, 1);

            //setup the shstrtab
            memcpy(&shdrout.shdr[2], sec_strndx, sizeof(shdrout.shdr[2]));
            set_te32(&shdrout.shdr[2].sh_name, 20);
            set_te32(&shdrout.shdr[2].sh_size, 29); //size of our static shstrtab
        }
    }
}

void PackLinuxElf32x86::pack1(OutputFile *fo, Filter &ft)
{
    super::pack1(fo, ft);
    if (0!=xct_off)  // shared library
        return;
    generateElfHdr(fo, stub_i386_linux_elf_fold, getbrk(phdri, e_phnum) );
}

void PackBSDElf32x86::pack1(OutputFile *fo, Filter &ft)
{
    super::pack1(fo, ft);
    if (0!=xct_off) // shared library
        return;
    generateElfHdr(fo, stub_i386_bsd_elf_fold, getbrk(phdri, e_phnum) );
}

void PackLinuxElf32armLe::pack1(OutputFile *fo, Filter &ft)
{
    super::pack1(fo, ft);
    if (0!=xct_off)  // shared library
        return;
    unsigned const e_flags = get_te32(&ehdri.e_flags);
    cprElfHdr3 h3;
    if (Elf32_Ehdr::ELFOSABI_LINUX==ei_osabi) {
        memcpy(&h3, stub_arm_v5a_linux_elf_fold, sizeof(Elf32_Ehdr) + 2*sizeof(Elf32_Phdr));

        h3.ehdr.e_ident[Elf32_Ehdr::EI_ABIVERSION] = e_flags>>24;
    }
    else {
        memcpy(&h3, stub_arm_v4a_linux_elf_fold,        sizeof(Elf32_Ehdr) + 2*sizeof(Elf32_Phdr));
    }
    // Fighting over .e_ident[EI_ABIVERSION]: Debian armhf is latest culprit.
    // So copy from input to output; but see PackLinuxElf32::generateElfHdr
    memcpy(&h3.ehdr.e_ident[0], &ehdri.e_ident[0], sizeof(ehdri.e_ident));
    set_te32(&h3.ehdr.e_flags, e_flags);
    generateElfHdr(fo, &h3, getbrk(phdri, e_phnum) );
}

void PackLinuxElf32armBe::pack1(OutputFile *fo, Filter &ft)
{
    super::pack1(fo, ft);
    if (0!=xct_off)  // shared library
        return;
    unsigned const e_flags = get_te32(&ehdri.e_flags);
    cprElfHdr3 h3;
    memcpy(&h3, stub_armeb_v4a_linux_elf_fold, sizeof(Elf32_Ehdr) + 2*sizeof(Elf32_Phdr));
    set_te32(&h3.ehdr.e_flags, e_flags);
    generateElfHdr(fo, &h3, getbrk(phdri, e_phnum) );
}

void PackLinuxElf32mipsel::pack1(OutputFile *fo, Filter &ft)
{
    super::pack1(fo, ft);
    if (0!=xct_off)  // shared library
        return;
    cprElfHdr3 h3;
    memcpy(&h3, stub_mipsel_r3000_linux_elf_fold, sizeof(Elf32_Ehdr) + 2*sizeof(Elf32_Phdr));
    generateElfHdr(fo, &h3, getbrk(phdri, e_phnum) );
}

void PackLinuxElf32mipseb::pack1(OutputFile *fo, Filter &ft)
{
    super::pack1(fo, ft);
    if (0!=xct_off)  // shared library
        return;
    cprElfHdr3 h3;
    memcpy(&h3, stub_mips_r3000_linux_elf_fold, sizeof(Elf32_Ehdr) + 2*sizeof(Elf32_Phdr));
    generateElfHdr(fo, &h3, getbrk(phdri, e_phnum) );
}

void PackLinuxElf32ppc::pack1(OutputFile *fo, Filter &ft)
{
    super::pack1(fo, ft);
    if (0!=xct_off)  // shared library
        return;
    generateElfHdr(fo, stub_powerpc_linux_elf_fold, getbrk(phdri, e_phnum) );
}

void PackLinuxElf64ppcle::pack1(OutputFile *fo, Filter &ft)
{
    super::pack1(fo, ft);
    if (0!=xct_off)  // shared library
        return;
    generateElfHdr(fo, stub_powerpc64le_linux_elf_fold, getbrk(phdri, e_phnum) );
}

void PackLinuxElf64ppc::pack1(OutputFile *fo, Filter &ft)
{
    super::pack1(fo, ft);
    if (0!=xct_off)  // shared library
        return;
    generateElfHdr(fo, stub_powerpc64_linux_elf_fold, getbrk(phdri, e_phnum) );
}

void PackLinuxElf64::asl_pack2_Shdrs(OutputFile *fo, unsigned pre_xct_top)
{
    if (!fo) {
        return;
    }
    // In order to pacify the runtime linker on Android "O" ("Oreo"),
    // we will splice-in a 4KiB page that contains an "extra" copy
    // of the Shdr, any PT_NOTE above xct_off, and shstrtab.
    // File order: Ehdr, Phdr[], section contents below xct_off,
    //    Shdr_copy[], PT_NOTEs.hi, shstrtab.
    if (is_asl)
        xct_va  += asl_delta;
    //xct_off += asl_delta;  // not until ::pack3()

    total_in = pre_xct_top;

    // Relocate PT_DYNAMIC (in PT_LOAD with PF_W)
    Elf64_Dyn *dyn = const_cast<Elf64_Dyn *>(dynseg);
    for (; dyn->d_tag; ++dyn) {
        upx_uint64_t d_tag = get_te64(&dyn->d_tag);
        if (Elf64_Dyn::DT_FINI       == d_tag
        ||  Elf64_Dyn::DT_FINI_ARRAY == d_tag
        ||  Elf64_Dyn::DT_INIT_ARRAY == d_tag
        ||  Elf64_Dyn::DT_PREINIT_ARRAY == d_tag
        ||  Elf64_Dyn::DT_PLTGOT      == d_tag) {
            upx_uint64_t d_val = get_te64(&dyn->d_val);
            set_te64(&dyn->d_val, asl_delta + d_val);
        }
    }
    // Updated dynseg (.dynamic, in PT_DYNAMIC (PT_LOAD{PF_W})) has not been written.
    // dynseg is in file_image[] but not in low_mem[].

    // Relocate dynsym (DT_SYMTAB) which is below xct_va
    upx_uint64_t const off_dynsym = get_te64(&sec_dynsym->sh_offset);
    upx_uint64_t const sz_dynsym  = get_te64(&sec_dynsym->sh_size);
    if ((upx_uint64_t)file_size < sz_dynsym
    ||  (upx_uint64_t)file_size < off_dynsym
    || ((upx_uint64_t)file_size - off_dynsym) < sz_dynsym) {
        throwCantPack("bad DT_SYMTAB");
    }
    Elf64_Sym *dyntym = (Elf64_Sym *)lowmem.subref(
        "bad dynsym", off_dynsym, sz_dynsym);
    Elf64_Sym *sym = dyntym;
    for (int j = sz_dynsym / sizeof(Elf64_Sym); --j>=0; ++sym) {
        upx_uint64_t symval = get_te64(&sym->st_value);
        unsigned symsec = get_te16(&sym->st_shndx);
        if (Elf64_Sym::SHN_UNDEF != symsec
        &&  Elf64_Sym::SHN_ABS   != symsec
        &&  xct_off <= symval) {
            set_te64(&sym->st_value, asl_delta + symval);
        }
        if (Elf64_Sym::SHN_ABS == symsec && xct_off <= symval) {
            adjABS(sym, asl_delta);
        }
    }

    // Relocate Phdr virtual addresses, but not physical offsets and sizes
    unsigned char buf_notes[512]; memset(buf_notes, 0, sizeof(buf_notes));
    unsigned len_notes = 0;
    Elf64_Phdr *phdr = (Elf64_Phdr *)lowmem.subref(
        "bad e_phoff", e_phoff, e_phnum * sizeof(Elf64_Phdr));
    for (unsigned j = 0; j < e_phnum; ++j, ++phdr) {
        upx_uint64_t offset = get_te64(&phdr->p_offset);
        if (xct_off <= offset) { // above the extra page
            {
                //set_te64(&phdr->p_offset, asl_delta + offset);  // physical
                upx_uint64_t addr = get_te64(&phdr->p_paddr);
                set_te64(&phdr->p_paddr, asl_delta + addr);
                             addr = get_te64(&phdr->p_vaddr);
                set_te64(&phdr->p_vaddr, asl_delta + addr);
            }
        }
        // .p_filesz,.p_memsz are updated in ::pack3
    }

    Elf64_Ehdr *const ehdr = (Elf64_Ehdr *)&lowmem[0];
    upx_uint64_t e_entry = get_te64(&ehdr->e_entry);
    if (xct_off < e_entry) {
        set_te64(&ehdr->e_entry, asl_delta + e_entry);
    }
    // Relocate Shdr; and Rela, Rel (below xct_off)
    unsigned const pal_xct_top = up8(pre_xct_top);
    set_te64(&ehdr->e_shoff, up8(pal_xct_top));  // Shdr alignment
    memcpy(&lowmem[pal_xct_top], shdri, e_shnum * sizeof(Elf64_Shdr));
    shdro = (Elf64_Shdr *)&lowmem[pal_xct_top];
    Elf64_Shdr *shdr = shdro;
    upx_uint64_t sz_shstrtab  = get_te64(&sec_strndx->sh_size);
    for (unsigned j = 0; j < e_shnum; ++j, ++shdr) {
        unsigned sh_type = get_te32(&shdr->sh_type);
        upx_uint64_t sh_size = get_te64(&shdr->sh_size);
        upx_uint64_t sh_offset = get_te64(&shdr->sh_offset);
        upx_uint64_t sh_entsize = get_te64(&shdr->sh_entsize);
        if ((upx_uint64_t)file_size < sh_size
        ||  (upx_uint64_t)file_size < sh_offset
        || (Elf64_Shdr::SHT_NOBITS != sh_type
           && ((upx_uint64_t)file_size - sh_offset) < sh_size) ) {
            throwCantPack("bad SHT_STRNDX");
        }

        if (xct_off <= sh_offset) {
            upx_uint64_t addr = get_te64(&shdr->sh_addr);
            set_te64(&shdr->sh_addr, asl_delta + addr);
            set_te64(&shdr->sh_offset, asl_delta + sh_offset);
        }
        switch (sh_type) {
        default: break;
        case Elf64_Shdr::SHT_RELA: {
            if (sizeof(Elf64_Rela) != sh_entsize) {
                char msg[50];
                snprintf(msg, sizeof(msg), "bad Rela.sh_entsize %lu", (long)sh_entsize);
                throwCantPack(msg);
            }
            plt_va = ~0ull;
            Elf64_Rela *const relb = (Elf64_Rela *)lowmem.subref(
                 "bad Rela offset", sh_offset, sh_size);
            Elf64_Rela *rela = relb;
            for (int k = sh_size / sh_entsize; --k >= 0; ++rela) {
                upx_uint64_t r_addend = get_te64(&rela->r_addend);
                upx_uint64_t r_offset = get_te64(&rela->r_offset);
                upx_uint64_t r_info   = get_te64(&rela->r_info);
                unsigned r_type = ELF64_R_TYPE(r_info);
                if (xct_off <= r_offset) {
                    set_te64(&rela->r_offset, asl_delta + r_offset);
                }
                if (Elf64_Ehdr::EM_AARCH64 == e_machine) switch (r_type) {
                    default: {
                        char msg[90]; snprintf(msg, sizeof(msg),
                            "unexpected relocation %#x [%#x]",
                            r_type, -1 + (unsigned)(sh_size / sh_entsize) - k);
                        throwCantPack(msg);
                    } break;
                    case R_AARCH64_ABS64: // FALL THROUGH
                    case R_AARCH64_GLOB_DAT: // FALL THROUGH
                    case R_AARCH64_RELATIVE: {
                        if (xct_off <= r_addend) {
                            set_te64(&rela->r_addend, asl_delta + r_addend);
                        }
                    } break;
                    case R_AARCH64_JUMP_SLOT: {
                        // .rela.plt contains offset of the "first time" target
                        if (plt_va > r_offset) {
                            plt_va = r_offset;
                        }
                        upx_uint64_t d = elf_get_offset_from_address(r_offset);
                        upx_uint64_t w = get_te64(&file_image[d]);
                        if (xct_off <= w) {
                            set_te64(&file_image[d], asl_delta + w);
                        }
                        ++n_jmp_slot;
                    } break;
                }
            }
        }; break;
        case Elf64_Shdr::SHT_REL: {
            if (sizeof(Elf64_Rel) != sh_entsize) {
                char msg[50];
                snprintf(msg, sizeof(msg), "bad Rel.sh_entsize %lu", (long)sh_entsize);
                throwCantPack(msg);
            }
            Elf64_Rel *rel = (Elf64_Rel *)lowmem.subref(
                    "bad Rel sh_offset", sh_offset, sh_size);
            for (int k = sh_size / sh_entsize; --k >= 0; ++rel) {
                upx_uint64_t r_offset = get_te64(&rel->r_offset);
                if (xct_off <= r_offset) {
                    set_te64(&rel->r_offset, asl_delta + r_offset);
                }
                // r_offset must be in 2nd PT_LOAD; .p_vaddr was already relocated
                upx_uint64_t d = elf_get_offset_from_address(asl_delta + r_offset);
                upx_uint64_t w = get_te64(&file_image[d]);
                upx_uint64_t r_info = get_te64(&rel->r_info);
                unsigned r_type = ELF64_R_TYPE(r_info);
                if (xct_off <= w
                &&  Elf64_Ehdr::EM_AARCH64 == e_machine
                &&  (  R_AARCH64_RELATIVE  == r_type
                    || R_AARCH64_JUMP_SLOT == r_type)) {
                    set_te64(&file_image[d], asl_delta + w);
                }
            }
        }; break;
        case Elf64_Shdr::SHT_NOTE: {
            if (!(Elf64_Shdr::SHF_ALLOC & get_te64(&shdr->sh_flags))) {
                // example: version number of 'gold' linker (static binder)
                if (sizeof(buf_notes) < (sh_size + len_notes)) {
                    throwCantPack("SHT_NOTEs too big");
                }
                set_te64(&shdro[j].sh_offset,
                    len_notes + (e_shnum * sizeof(Elf64_Shdr)) + xct_off);
                memcpy(&buf_notes[len_notes], &file_image[sh_offset], sh_size);
                len_notes += sh_size;
            }
            else { // SHF_ALLOC: in PT_LOAD; but move sh_addr and sh_offset
                // Not sure why we need this conditional.
                // Anyway, some Android have multiple SHT_NOTE sections.
                if (xct_off <= sh_offset) {
                    upx_uint64_t pos = xct_off + e_shnum * sizeof(Elf64_Shdr);
                    set_te64(&shdr->sh_addr,   pos);
                    set_te64(&shdr->sh_offset, pos);
                }
            }
        }; break;
        } // end switch (sh_type)
    }
    // shstrndx will move
    set_te64(&shdro[get_te16(&ehdri.e_shstrndx)].sh_offset,
        len_notes + e_shnum * sizeof(Elf64_Shdr) + pal_xct_top);

    // ("Re-")write all changes below pal_xct_top
    fo->seek(0, SEEK_SET);
    fo->write(lowmem, pal_xct_top);
    total_in = pal_xct_top;

    // New copy of Shdr
    Elf64_Shdr blank; memset(&blank, 0, sizeof(blank));
    set_te64(&blank.sh_offset, xct_off);  // hint for "upx -d"
    fpad8(fo, total_out);  // Shdr alignment
    fo->write(&blank, sizeof(blank));
    fo->write(&shdro[1], (-1+ e_shnum) * sizeof(Elf64_Shdr));

    if (len_notes) {
        fo->write(buf_notes, len_notes);
    }

    // New copy of Shdr[.e_shstrndx].[ sh_offset, +.sh_size )
    fo->write(shstrtab,  sz_shstrtab);

    sz_elf_hdrs = fpad8(fo, total_out);
    total_out = sz_elf_hdrs;
    //xct_off += asl_delta;  // wait until ::pack3
    unsigned d = asl_delta + pal_xct_top - sz_elf_hdrs;
    fo->seek(d, SEEK_CUR);
    total_out += d;
}

// asl (ANdroid Shared Library) creates a big mess!
void PackLinuxElf32::asl_pack2_Shdrs(OutputFile *fo, unsigned pre_xct_top)
{
    if (!fo) {
        return;
    }
    // In order to pacify the runtime linker on Android "O" ("Oreo"),
    // we will splice-in a 4KiB page that contains an "extra" copy
    // of the Shdr, any PT_NOTE above xct_off, and shstrtab.
    // File order: Ehdr, Phdr[], section contents below xct_off,
    //    Shdr_copy[], PT_NOTEs.hi, shstrtab.
    if (is_asl)
        xct_va  += asl_delta;
    //xct_off += asl_delta;  // not until ::pack3()

    total_in = pre_xct_top;

    // Relocate PT_DYNAMIC (in PT_LOAD with PF_W)
    Elf32_Dyn *dyn = const_cast<Elf32_Dyn *>(dynseg);
    for (; dyn->d_tag; ++dyn) {
        upx_uint32_t d_tag = get_te32(&dyn->d_tag);
        if (Elf32_Dyn::DT_FINI       == d_tag
        ||  Elf32_Dyn::DT_FINI_ARRAY == d_tag
        ||  Elf32_Dyn::DT_INIT_ARRAY == d_tag
        ||  Elf32_Dyn::DT_PREINIT_ARRAY == d_tag
        ||  Elf32_Dyn::DT_PLTGOT      == d_tag) {
            upx_uint32_t d_val = get_te32(&dyn->d_val);
            set_te32(&dyn->d_val, asl_delta + d_val);
        }
    }
    // Updated dynseg (.dynamic, in PT_DYNAMIC (PT_LOAD{PF_W})) has not been written.
    // dynseg is in file_image[] but not in low_mem[].

    // Relocate dynsym (DT_SYMTAB) which is below xct_va
    upx_uint32_t const off_dynsym = get_te32(&sec_dynsym->sh_offset);
    upx_uint32_t const sz_dynsym  = get_te32(&sec_dynsym->sh_size);
    if ((upx_uint32_t)file_size < sz_dynsym
    ||  (upx_uint32_t)file_size < off_dynsym
    || ((upx_uint32_t)file_size - off_dynsym) < sz_dynsym) {
        throwCantPack("bad DT_SYMTAB");
    }
    Elf32_Sym *dyntym = (Elf32_Sym *)lowmem.subref(
        "bad dynsym", off_dynsym, sz_dynsym);
    Elf32_Sym *sym = dyntym;
    for (int j = sz_dynsym / sizeof(Elf32_Sym); --j>=0; ++sym) {
        upx_uint32_t symval = get_te32(&sym->st_value);
        unsigned symsec = get_te16(&sym->st_shndx);
        if (Elf32_Sym::SHN_UNDEF != symsec
        &&  Elf32_Sym::SHN_ABS   != symsec
        &&  xct_off <= symval) {
            set_te32(&sym->st_value, asl_delta + symval);
        }
        if (Elf32_Sym::SHN_ABS == symsec && xct_off <= symval) {
            adjABS(sym, asl_delta);
        }
    }

    // Relocate Phdr virtual addresses, but not physical offsets and sizes
    unsigned char buf_notes[512]; memset(buf_notes, 0, sizeof(buf_notes));
    unsigned len_notes = 0;
    Elf32_Phdr *phdr = (Elf32_Phdr *)lowmem.subref(
        "bad e_phoff", e_phoff, e_phnum * sizeof(Elf32_Phdr));
    for (unsigned j = 0; j < e_phnum; ++j, ++phdr) {
        upx_uint32_t offset = get_te32(&phdr->p_offset);
        if (xct_off <= offset) { // above the extra page
            {
                //set_te32(&phdr->p_offset, asl_delta + offset);  // physical
                upx_uint32_t v_addr = get_te32(&phdr->p_vaddr);
                                      set_te32(&phdr->p_vaddr, asl_delta + v_addr);
                upx_uint32_t p_addr = get_te32(&phdr->p_paddr);
                                      set_te32(&phdr->p_paddr, asl_delta + p_addr);
            }
        }
        // .p_filesz,.p_memsz are updated in ::pack3
    }

    Elf32_Ehdr *const ehdr = (Elf32_Ehdr *)&lowmem[0];
    upx_uint32_t e_entry = get_te32(&ehdr->e_entry);
    if (xct_off <= e_entry) { // FIXME: --android-shlib is different
        set_te32(&ehdr->e_entry, asl_delta + e_entry);
    }
    // Relocate Shdr; and Rela, Rel (below xct_off)
    unsigned const pal_xct_top = up4(pre_xct_top);
    set_te32(&ehdr->e_shoff, pal_xct_top);  // Shdr alignment
    memcpy(&lowmem[pal_xct_top], shdri, e_shnum * sizeof(Elf32_Shdr));
    shdro = (Elf32_Shdr *)&lowmem[pal_xct_top];
    Elf32_Shdr *shdr = shdro;
    upx_uint32_t sz_shstrtab  = get_te32(&sec_strndx->sh_size);
    for (unsigned j = 0; j < e_shnum; ++j, ++shdr) {
        unsigned sh_type = get_te32(&shdr->sh_type);
        unsigned sh_flags = get_te32(&shdr->sh_flags);
        upx_uint32_t sh_size = get_te32(&shdr->sh_size);
        upx_uint32_t sh_offset = get_te32(&shdr->sh_offset);
        upx_uint32_t sh_entsize = get_te32(&shdr->sh_entsize);
        if ((upx_uint32_t)file_size < sh_size
        ||  (upx_uint32_t)file_size < sh_offset
        || (Elf32_Shdr::SHT_NOBITS != sh_type
           && ((upx_uint32_t)file_size - sh_offset) < sh_size) ) {
            throwCantPack("bad SHT_STRNDX");
        }

        if (xct_off <= sh_offset && Elf32_Shdr::SHF_ALLOC & sh_flags) {
            upx_uint32_t addr = get_te32(&shdr->sh_addr);
            set_te32(&shdr->sh_addr, asl_delta + addr);
            set_te32(&shdr->sh_offset, asl_delta + sh_offset);
        }
        switch (sh_type) {
        default: break;
        case Elf32_Shdr::SHT_RELA: { // 32-bit Elf_Rela is unused (by convention)
            if (sizeof(Elf32_Rela) != sh_entsize) {
                char msg[50];
                snprintf(msg, sizeof(msg), "bad Rela.sh_entsize %lu", (long)sh_entsize);
                throwCantPack(msg);
            }
            plt_va = ~0ull;
            Elf32_Rela *const relb = (Elf32_Rela *)lowmem.subref(
                 "bad Rela offset", sh_offset, sh_size);
            Elf32_Rela *rela = relb;
            for (int k = sh_size / sh_entsize; --k >= 0; ++rela) {
                upx_uint32_t r_addend = get_te32(&rela->r_addend);
                upx_uint32_t r_offset = get_te32(&rela->r_offset);
                upx_uint32_t r_info   = get_te32(&rela->r_info);
                unsigned r_type = ELF32_R_TYPE(r_info);
                if (xct_off <= r_offset) {
                    set_te32(&rela->r_offset, asl_delta + r_offset);
                }
                if (Elf32_Ehdr::EM_386 == e_machine) switch (r_type) {
                    default: {
                        char msg[90]; snprintf(msg, sizeof(msg),
                            "unexpected relocation %#x [%#x]",
                            r_type, -1 + (unsigned)(sh_size / sh_entsize) - k);
                        throwCantPack(msg);
                    } break;
                    case R_386_32: // FALL THROUGH
                    case R_386_GLOB_DAT: // FALL THROUGH
                    case R_386_RELATIVE: {
                        if (xct_off <= r_addend) {
                            set_te32(&rela->r_addend, asl_delta + r_addend);
                        }
                    } break;
                    case R_386_JMP_SLOT: {
                        // .rela.plt contains offset of the "first time" target
                        if (plt_va > r_offset) {
                            plt_va = r_offset;
                        }
                        upx_uint32_t d = elf_get_offset_from_address(r_offset);
                        upx_uint32_t w = get_te32(&file_image[d]);
                        if (xct_off <= w) {
                            set_te32(&file_image[d], asl_delta + w);
                        }
                        ++n_jmp_slot;
                    } break;
                } // end EM_386 r_type
                else if (Elf32_Ehdr::EM_ARM == e_machine) switch (r_type) {
                    default: {
                        char msg[90]; snprintf(msg, sizeof(msg),
                            "unexpected relocation %#x [%#x]",
                            r_type, -1 + (unsigned)(sh_size / sh_entsize) - k);
                        throwCantPack(msg);
                    } break;
                    case R_ARM_ABS32: // FALL THROUGH
                    case R_ARM_GLOB_DAT: // FALL THROUGH
                    case R_ARM_RELATIVE: {
                        if (xct_off <= r_addend) {
                            set_te32(&rela->r_addend, asl_delta + r_addend);
                        }
                    } break;
                    case R_ARM_JUMP_SLOT: {
                        // .rela.plt contains offset of the "first time" target
                        if (plt_va > r_offset) {
                            plt_va = r_offset;
                        }
                        upx_uint32_t d = elf_get_offset_from_address(r_offset);
                        upx_uint32_t w = get_te32(&file_image[d]);
                        if (xct_off <= w) {
                            set_te32(&file_image[d], asl_delta + w);
                        }
                        ++n_jmp_slot;
                    } break;
                }  // end EM_ARM r_type
                else {
                    char msg[40]; snprintf(msg, sizeof(msg),
                        "Unknown architecture %d", this->e_machine);
                    throwCantPack(msg);
                }  // end e_machine
            }
        }; break;  // end Elf32_Shdr::SHT_RELA
        case Elf32_Shdr::SHT_REL: {
            if (sizeof(Elf32_Rel) != sh_entsize) {
                char msg[50];
                snprintf(msg, sizeof(msg), "bad Rel.sh_entsize %lu", (long)sh_entsize);
                throwCantPack(msg);
            }
            Elf32_Rel *rel = (Elf32_Rel *)lowmem.subref(
                    "bad Rel sh_offset", sh_offset, sh_size);
            for (int k = sh_size / sh_entsize; --k >= 0; ++rel) {
                upx_uint32_t r_offset = get_te32(&rel->r_offset);
                if (xct_off <= r_offset) {
                    set_te32(&rel->r_offset, asl_delta + r_offset);
                }
                // r_offset must be in 2nd PT_LOAD; .p_vaddr was already relocated
                upx_uint32_t d = elf_get_offset_from_address(r_offset);
                upx_uint32_t w = get_te32(&file_image[d]);
                upx_uint32_t r_info = get_te32(&rel->r_info);
                unsigned r_type = ELF32_R_TYPE(r_info);
                //printf("d=%#x  w=%#x  r_info=%#x\n", d, w, r_info);
                // FIXME: why change file_image[d] instead of lowmem[d] ?
                if (Elf32_Ehdr::EM_386 == e_machine) switch (r_type) {
                    default: {
                        char msg[90]; snprintf(msg, sizeof(msg),
                            "unexpected relocation %#x [%#x]",
                            r_type, -1 + (unsigned)(sh_size / sh_entsize) - k);
                        throwCantPack(msg);
                    } break;
                    case R_386_32: // FALL THROUGH
                    case R_386_GLOB_DAT: // FALL THROUGH
                    case R_386_RELATIVE: {
                        if (xct_off <= w) {
                            set_te32(&file_image[d], asl_delta + w);
                        }
                    } break;
                    case R_386_JMP_SLOT: {
                        // .rela.plt contains offset of the "first time" target
                        if (plt_va > r_offset) {
                            plt_va = r_offset;
                        }
                        if (xct_off <= w) {
                            set_te32(&file_image[d], asl_delta + w);
                        }
                        ++n_jmp_slot;
                    } break;
                } // end EM_386 r_type
                else if (Elf32_Ehdr::EM_ARM == e_machine) switch (r_type) {
                    default: {
                        char msg[90]; snprintf(msg, sizeof(msg),
                            "unexpected relocation %#x [%#x]",
                            r_type, -1 + (unsigned)(sh_size / sh_entsize) - k);
                        throwCantPack(msg);
                    } break;
                    case R_ARM_ABS32: // FALL THROUGH
                    case R_ARM_GLOB_DAT: // FALL THROUGH
                    case R_ARM_RELATIVE: {
                        if (xct_off <= w) {
                            set_te32(&file_image[d], asl_delta + w);
                        }
                    } break;
                    case R_ARM_JUMP_SLOT: {
                        // .rela.plt contains offset of the "first time" target
                        if (plt_va > r_offset) {
                            plt_va = r_offset;
                        }
                        if (xct_off <= w) {
                            set_te32(&file_image[d], asl_delta + w);
                        }
                        ++n_jmp_slot;
                    } break;
                }  // end EM_ARM r_type
                else {
                    char msg[40]; snprintf(msg, sizeof(msg),
                        "Unknown architecture %d", this->e_machine);
                    throwCantPack(msg);
                }  // end e_machine
            }  // end rel
        }; break;  // end Elf32_Shdr::SHT_REL
        case Elf32_Shdr::SHT_NOTE: {
            if (!(Elf32_Shdr::SHF_ALLOC & sh_flags)) {
                // example: version number of 'gold' linker (static binder)
                if (sizeof(buf_notes) < (sh_size + len_notes)) {
                    throwCantPack("SHT_NOTEs too big");
                }
                set_te32(&shdro[j].sh_offset,
                    len_notes + (e_shnum * sizeof(Elf32_Shdr)) + xct_off);
                memcpy(&buf_notes[len_notes], &file_image[sh_offset], sh_size);
                len_notes += sh_size;
            }
            else { // SHF_ALLOC: in PT_LOAD; but move sh_addr and sh_offset
                // Not sure why we need this conditional.
                // Anyway, some Android have multiple SHT_NOTE sections.
                if (xct_off <= sh_offset) {
                    upx_uint32_t pos = xct_off + e_shnum * sizeof(Elf32_Shdr);
                    //set_te32(&shdr->sh_addr,   pos);
                    set_te32(&shdr->sh_offset, pos);
                }
            }
        }; break;  // end Elf32_Shdr::SHT_NOTE
        case Elf32_Shdr::SHT_ARM_ATTRIBUTES: {
            sec_arm_attr = shdr;
        }; break;
        } // end switch (sh_type)
    }
    // shstrndx will move
    set_te32(&shdro[get_te16(&ehdri.e_shstrndx)].sh_offset,
        len_notes + e_shnum * sizeof(Elf32_Shdr) + up8(pal_xct_top));

    // Write all changes below pal_xct_top
    // FIXME: why is this any more than Ehdr + Phdrs?
    if (fo) {
        fo->seek(0, SEEK_SET);
        fo->write(lowmem, pal_xct_top);
    }
    total_out = pal_xct_top;
    total_in  = pal_xct_top;

    // New copy of Shdr
    Elf32_Shdr blank; memset(&blank, 0, sizeof(blank));
    set_te32(&blank.sh_offset, xct_off);  // hint for "upx -d"
    set_te32(&shdro->sh_offset, xct_off);  // hint for "upx -d"
    total_out = fpad8(fo, total_out);  // Shdr alignment
    unsigned arm_attr_off = 0;
    if (sec_arm_attr) {
        arm_attr_off = get_te32(&sec_arm_attr->sh_offset);
        set_te32(&sec_arm_attr->sh_offset,
            total_out + e_shnum*sizeof(Elf32_Shdr)
            + len_notes + sz_shstrtab);
    }
    if (fo) {
        fo->write(&blank, sizeof(blank));
        fo->write(&shdro[1], (-1+ e_shnum) * sizeof(Elf32_Shdr));
        if (len_notes) {
            fo->write(buf_notes, len_notes);
        }
        // New copy of Shdr[.e_shstrndx].[ sh_offset, +.sh_size )
        fo->write(shstrtab,  sz_shstrtab);

        if (sec_arm_attr) {
            fo->write(&file_image[arm_attr_off],
                get_te32(&sec_arm_attr->sh_size));
        }
    }

    sz_elf_hdrs = fpad8(fo, total_out);
    total_out = sz_elf_hdrs;
    //xct_off += asl_delta;  // wait until ::pack3
    total_out = fpadN(fo, asl_delta - (sz_elf_hdrs - pal_xct_top));
}

void PackLinuxElf64::pack1(OutputFile * /*fo*/, Filter &ft)
{
    fi->seek(0, SEEK_SET);
    fi->readx(&ehdri, sizeof(ehdri));
    assert(e_phoff == sizeof(Elf64_Ehdr));  // checked by canPack()
    sz_phdrs = e_phnum * get_te16(&ehdri.e_phentsize);

// We compress separate pieces (usually each PT_LOAD, plus the gaps in the file
// that are not covered by any PT_LOAD), but currently at run time there can be
// only one decompressor method.
// Therefore we must plan ahead because Packer::compressWithFilters tries
// to find the smallest result among the available methods, for one piece only.
// In the future we may allow more than one decompression method at run time.
// For now we must choose only one, and force PackUnix::packExtent
// (==> compressWithFilters) to use it.
    int nfilters = 0;
    {
        int const *fp = getFilters();
        while (FT_END != *fp++) {
            ++nfilters;
        }
    }
    {
        int npieces = 1;  // tail after highest PT_LOAD
        Elf64_Phdr *phdr = phdri;
        for (unsigned j=0; j < e_phnum; ++phdr, ++j) {
            if (is_LOAD(phdr)) {
                unsigned const  flags = get_te32(&phdr->p_flags);
                unsigned       offset = get_te64(&phdr->p_offset);
                if (!xct_off  // not shlib
                  // new-style shlib: PT_LOAD[0] has symbol table
                  // which must not be compressed, but also lacks PF_X
                ||    (Elf64_Phdr::PF_X & flags)
                  // Read-only, non-first PT_LOAD is _assumed_ to be compressible
                ||  (!(Elf64_Phdr::PF_W & flags) && 0!=offset))
                {
                    ++npieces;  // will attempt compression of this PT_LOAD
                }
            }
        }
        uip->ui_total_passes += npieces;
    }
    int methods[256];
    unsigned nmethods = prepareMethods(methods, ph.method, getCompressionMethods(M_ALL, ph.level));
    if (1 < nmethods) { // Many are available, but we must choose only one
        uip->ui_total_passes += 1;  // the batch for output
        uip->ui_total_passes *= nmethods * (1+ nfilters);  // finding smallest total
        PackHeader orig_ph = ph;
        Filter orig_ft = ft;
        unsigned max_offset = 0;
        unsigned sz_best= ~0u;
        int method_best = 0;
        for (unsigned k = 0; k < nmethods; ++k) { // FIXME: parallelize; cost: working space
            unsigned sz_this = 0;
            Elf64_Phdr *phdr = phdri;
            for (unsigned j=0; j < e_phnum; ++phdr, ++j) {
                if (is_LOAD(phdr)) {
                    unsigned const  flags = get_te32(&phdr->p_flags);
                    unsigned       offset = get_te64(&phdr->p_offset);
                    unsigned       filesz = get_te64(&phdr->p_filesz);
                    max_offset = UPX_MAX(max_offset, filesz + offset);
                    if (!xct_off  // not shlib
                      // new-style shlib: PT_LOAD[0] has symbol table
                      // which must not be compressed, but also lacks PF_X
                    ||    (Elf64_Phdr::PF_X & flags)
                      // Read-only, non-first PT_LOAD is _assumed_ to be compressible
                    ||  (!(Elf64_Phdr::PF_W & flags) && 0!=offset))
                    {
                        if (xct_off && 0==offset) { // old-style shlib
                            offset  = xct_off;
                            filesz -= xct_off;
                        }
                        fi->seek(offset, SEEK_SET);
                        fi->readx(ibuf, filesz);
                        ft = orig_ft;
                        ph = orig_ph;
                        ph.set_method(ph_force_method(methods[k]));
                        ph.u_len = filesz;
                        compressWithFilters(&ft, OVERHEAD, NULL_cconf, 10, true);
                        sz_this += ph.c_len;
                    }
                }
            }
            unsigned const sz_tail = file_size - max_offset;  // debuginfo, etc.
            if (sz_tail) {
                fi->seek(max_offset, SEEK_SET);
                fi->readx(ibuf, sz_tail);
                ft = orig_ft;
                ph = orig_ph;
                ph.set_method(ph_force_method(methods[k]));
                ph.u_len = sz_tail;
                compressWithFilters(&ft, OVERHEAD, NULL_cconf, 10, true);
                sz_this += ph.c_len;
            }
            // FIXME: loader size also depends on method
            if (sz_best > sz_this) {
                sz_best = sz_this;
                method_best = methods[k];
            }
        }
        ft = orig_ft;
        ph = orig_ph;
        ph.set_method(ph_force_method(method_best));
    }

    Elf64_Phdr *phdr = phdri;
    for (unsigned j=0; j < e_phnum; ++phdr, ++j) {
        unsigned const type = get_te32(&phdr->p_type);
        if (PT_GNU_STACK64 == type || PT_NOTE64 == type) {
            add_phdrx(phdr);  // PT_GNU_STACK64, PT_NOTE64
        }
        if (PT_LOAD == type) {
            unsigned x = get_te64(&phdr->p_align) >> lg2_page;
            while (x>>=1) {
                ++lg2_page;
            }
        }
        if (PT_GNU_RELRO64 == type
        &&  16 < lg2_page) {
            // x86* allows 4K, 2M, 1G
            // arm32 and arm64 both allow 4K, 16K, 64K; Apple Silicon uses 16K
            // Oracle 5.4.17-2136.314.6.2.el8uek.aarch64 demands 64K
            // PowerPC and PowerPC64 has 4K, 64K, 1M (?), 16M (?)
            // MIPS  allows any power of 2 from 4K through 64K

            // If 64K < .p_align then we assume that 4K is also accepted.
            // .p_align can be like 2M, which is a huge over-estimate.
            // RELRO ends on a page boundary: usually close to actual page_size
            unsigned offset = get_te64(&phdr->p_offset);
            unsigned filesz = get_te64(&phdr->p_filesz);
            if (!(0xfff & (filesz + offset))) { // a 4KiB boundary
                unsigned b = 12;
                while (!(~(~0u << b) & (filesz + offset))) {
                    ++b;
                }
                lg2_page = umin(lg2_page, -1+ b);
            }
        }
    }
    page_size =  1u  <<lg2_page;
    page_mask = ~0ull<<lg2_page;

    progid = 0;  // getRandomId();  not useful, so do not clutter
    sz_elf_hdrs = sizeof(ehdri) + sz_phdrs;

    // only execute if option present
    if (opt->o_unix.preserve_build_id) {
        // set this so we can use elf_find_section_name
        e_shnum = get_te16(&ehdri.e_shnum);
        if (!shdri) {
            mb_shdr.alloc(e_shnum * sizeof(Elf64_Shdr));
            shdri = (Elf64_Shdr *)mb_shdr.getVoidPtr();
            e_shoff = get_te64(&ehdri.e_shoff);
            fi->seek(e_shoff, SEEK_SET);
            fi->readx(shdri, e_shnum * sizeof(Elf64_Shdr));
        }
        //set the shstrtab
        sec_strndx = &shdri[get_te16(&ehdri.e_shstrndx)];

        upx_uint64_t sh_size = get_te64(&sec_strndx->sh_size);
        mb_shstrtab.alloc(sh_size); shstrtab = (char *)mb_shstrtab.getVoidPtr();
        fi->seek(0,SEEK_SET);
        fi->seek(sec_strndx->sh_offset,SEEK_SET);
        fi->readx(mb_shstrtab, sh_size);

        Elf64_Shdr const *buildid = elf_find_section_name(".note.gnu.build-id");
        if (buildid) {
            unsigned bid_sh_size = get_te64(&buildid->sh_size);  // UPX_RSIZE_MAX_MEM protects us
            buildid_data.alloc(bid_sh_size);
            buildid_data.clear();
            fi->seek(0,SEEK_SET);
            fi->seek(buildid->sh_offset,SEEK_SET);
            fi->readx((void *)buildid_data, bid_sh_size);

            o_elf_shnum = 3;
            memset(&shdrout,0,sizeof(shdrout));

            //setup the build-id
            memcpy(&shdrout.shdr[1], buildid, sizeof(shdrout.shdr[1]));
            set_te32(&shdrout.shdr[1].sh_name, 1);

            //setup the shstrtab
            memcpy(&shdrout.shdr[2], sec_strndx, sizeof(shdrout.shdr[2]));
            set_te32(&shdrout.shdr[2].sh_name, 20);
            set_te64(&shdrout.shdr[2].sh_size, 29); //size of our static shstrtab; UPX_RSIZE_MAX_MEM
        }
    }
}

void PackLinuxElf64amd::pack1(OutputFile *fo, Filter &ft)
{
    super::pack1(fo, ft);
    if (0!=xct_off)  // shared library
        return;
    generateElfHdr(fo, stub_amd64_linux_elf_fold, getbrk(phdri, e_phnum) );
}

void PackLinuxElf64arm::pack1(OutputFile *fo, Filter &ft)
{
    super::pack1(fo, ft);
    if (0!=xct_off)  // shared library
        return;
    generateElfHdr(fo, stub_arm64_linux_elf_fold, getbrk(phdri, e_phnum) );
}

// Determine length of gap between PT_LOAD phdr[k] and closest PT_LOAD
// which follows in the file (or end-of-file).  Optimize for common case
// where the PT_LOAD are adjacent ascending by .p_offset.  Assume no overlap.

unsigned PackLinuxElf32::find_LOAD_gap(
    Elf32_Phdr const *const phdr,
    unsigned const k,
    unsigned const nph
)
{
    if (!is_LOAD(&phdr[k])) {
        return 0;
    }
    unsigned const hi = get_te32(&phdr[k].p_offset) +
                        get_te32(&phdr[k].p_filesz);
    unsigned lo = ph.u_file_size;
    if (lo < hi)
        throwCantPack("bad input: PT_LOAD beyond end-of-file");
    unsigned j = k;
    for (;;) { // circular search, optimize for adjacent ascending
        ++j;
        if (nph==j) {
            j = 0;
        }
        if (k==j) {
            break;
        }
        if (is_LOAD(&phdr[j])) {
            unsigned const t = get_te32(&phdr[j].p_offset);
            if ((t - hi) < (lo - hi)) {
                lo = t;
                if (hi==lo) {
                    break;
                }
            }
        }
    }
    return lo - hi;
}

unsigned PackLinuxElf::pack2_shlib_overlay_init(OutputFile *fo)
{
    linfo.l_checksum = 0;  // preliminary
    linfo.l_magic = UPX_MAGIC_LE32;

    set_le16(&linfo.l_lsize, lsize);  // preliminary (0)
    linfo.l_version = (unsigned char)ph.version;
    linfo.l_format =  (unsigned char)ph.format;
    linfo_off = total_out;
    fo->write(&linfo, sizeof(linfo));  total_out += sizeof(linfo);

    overlay_offset = total_out;

    p_info hbuf;
    set_te32(&hbuf.p_progid, 0);
    set_te32(&hbuf.p_filesize, file_size);
    set_te32(&hbuf.p_blocksize, blocksize);
    fo->write(&hbuf, sizeof(hbuf));  total_out += sizeof(hbuf);
    return total_out;
}

unsigned PackLinuxElf::pack2_shlib_overlay_write(OutputFile *fo, MemBuffer &mb,
        unsigned u_len, unsigned c_len)
{
    // Write mb with b_info header.
    b_info tmp;
    memset(&tmp, 0, sizeof(tmp));
    set_te32(&tmp.sz_unc, u_len);
    set_te32(&tmp.sz_cpr, c_len);
    tmp.b_method = (EM_ARM == e_machine) ? M_NRV2B_8 : M_NRV2B_LE32;
    tmp.b_extra = 0;
    fo->write(&tmp, sizeof(tmp));   total_out += sizeof(tmp);
    b_len += sizeof(b_info);
    fo->write(mb, c_len); total_out += c_len;
    return total_out;
}

// Returns compressed size
unsigned PackLinuxElf::pack2_shlib_overlay_compress(
    MemBuffer &bufo,
    upx_byte const *inp,
    unsigned u_len
)
{
    ph.saved_u_adler = ph.u_adler;
    ph.u_adler = upx_adler32(inp, u_len, ph.saved_u_adler);
    ph.u_len += u_len;

    unsigned const method = (EM_ARM == e_machine) ? M_NRV2B_8 : M_NRV2B_LE32;
    methods_used |= 1 << method;
    unsigned c_len = 0;
    int r = upx_compress(inp, u_len, bufo, &c_len,
        /* &progress callback */ nullptr,
        method, 10,
        /* &config_t */ nullptr, /* &result_t */ nullptr);
    if (r != UPX_E_OK)
        throwInternalError("header compression failed");
    if (c_len >= u_len)
        throwInternalError("header compression size increase");

    ph.saved_c_adler = ph.c_adler;
    ph.c_adler = upx_adler32(bufo,  c_len, ph.saved_c_adler);
    ph.c_len += c_len;
    return c_len;
}

// Layout for compressed shared library:
// 1. If first PT_LOAD is totally below xct_off
//      (namely, (.p_memsz + .p_offset) <= xct_off )
//    then it must be Ehdr+Phdrs followed only by PT_DYNAMIC (Dynamic
//    section) and loader tables (DT_* data), and the order of output is:
//      A. original first PT_LOAD (Ehdr+Phdrs will be overwritten later)
//      B. compressed Ehdr+Phdrs (for de-compressor to restore); M_NRV2B_LE32, no filter
//      C. compressed remaining PT_LOAD segments (except PT_WRITE),
//         regardless of xct_off.
//         Until xct_off is covered by some PT_LOAD, then there is a logical
//         two-pass problem.  The best compression algorithm is not chosen
//         until the PT_LOAD which contains xct_off, but before that
//         compression is desired anyway for any preceding PT_LOAD.
//         So use NRV2B for the "early" PT_LOADs, and do not use packExtent
//         which calls compressWithFilters and also creates a loader.
//         Instead, do the compression "by hand" explicitly using upx_compress
//         and write().  This may require ELF2 multi-decompressor (NRV2B
//         on the early PT_LOADs, then something better).
//
//         Example input Program Headers:
//         Type   Offset   VirtAddr   PhysAddr   FileSiz MemSiz  Flg Align
//         LOAD   0x000000 0x00000000 0x00000000 0x003a4 0x003a4 R   0x1000  Elf hdrs, loader tables
//         LOAD   0x001000 0x00001000 0x00001000 0x001d4 0x001d4 R E 0x1000  small code
//         LOAD   0x002000 0x00002000 0x00002000 0x09d40 0x09d40 R   0x1000  large app consts
//         LOAD   0x00bef0 0x0000cef0 0x0000cef0 0x00118 0x0011c RW  0x1000  writeable
//
//         xct_off will have been increased artifically to point to
//         the large compressable PT_LOAD (0x9d40==MemSiz), in order to avoid
//         NotCompressable because the earlier PT_LOADs were too short (< 4KiB).
// 2. If the first PT_LOAD covers xct_off, then the order of output is:
//      A. original Ehdr+Phdrs (overwritten later)
//      B. other original content below xct_off [loader tables]
//      C. compressed Ehdr+Phdrs (for de-compressor to restore); M_NRV2B_LE32, no filter.
//      D. compressed content above xct_off in first PT_LOAD;
//         use filter for executable code.
//      E. compressed remaining PT_LOAD (except PT_WRITE); no filter.
//

int PackLinuxElf32::pack2_shlib(OutputFile *fo, Filter &ft, unsigned pre_xct_top)
{
    // prepare to alter Phdrs and Shdrs
    lowmem.alloc(up8(xct_off + (!is_asl
        ? 0
        : e_shnum * sizeof(Elf32_Shdr))));
    memcpy(lowmem, file_image, xct_off);  // android omits Shdr from copy now

    MemBuffer elf_buf;
    elf_buf.allocForCompression(sz_elf_hdrs);
    unsigned u_len = sz_elf_hdrs;
    unsigned c_len = pack2_shlib_overlay_compress(elf_buf, lowmem, u_len);

    int n_ptload = 0;
    Elf32_Phdr *phdr = &phdri[0];
    for (unsigned k = 0; k < e_phnum; ++phdr, ++k)
    if (is_LOAD(phdr)) {
        unsigned p_offset = get_te32(&phdr->p_offset);
        unsigned p_filesz = get_te32(&phdr->p_filesz);
        if (!n_ptload) { // first PT_LOAD
            if (is_asl) { // Copy up to xct_off, then add 2nd copy of Shdrs
                asl_pack2_Shdrs(fo, pre_xct_top);
            }
            else {
                fo->write(&lowmem[p_offset], sz_elf_hdrs);  total_out += sz_elf_hdrs;
                total_in  += sz_elf_hdrs;
                fo->seek(sz_phdrx, SEEK_CUR);  total_out += sz_phdrx;  // leave space

                // Compare PackUnix::packExtent, especially "if (u_len)" .
                //

                unsigned const hi_offset = umin(xct_off, p_filesz + p_offset);
                if (sz_elf_hdrs < hi_offset) {
                    // Loader tables in first PT_LOAD, and below xct_off.
                    //
                    fo->write(&lowmem[sz_elf_hdrs], hi_offset - sz_elf_hdrs);  total_out += hi_offset - sz_elf_hdrs;
                    total_in  += hi_offset - sz_elf_hdrs;
                    Elf32_Phdr *lo_phdr = k + (Elf32_Phdr *)(1+ (Elf32_Ehdr *)&lowmem[0]);
                    set_te32(&lo_phdr->p_flags, Elf32_Phdr::PF_X | get_te32(&lo_phdr->p_flags));
                }
            }
            pack2_shlib_overlay_init(fo); // set overlay_offset
            // Compressed ELF headers go first.
            pack2_shlib_overlay_write(fo, elf_buf, u_len, c_len);

            // The (compressible) remainder above xct_off in first PT_LOAD
            if (         p_filesz > (xct_off - p_offset)) {
                Extent x;
                x.size = p_filesz - (xct_off - p_offset);
                x.offset = xct_off;
                packExtent(x, &ft, fo, 0, 0, true);
            }
        }  // end first PT_LOAD
        else if ((p_filesz + p_offset) <= xct_off) {
            // Not first PT_LOAD, but below xct_offset.  Thus "interloper" PT_LOAD,
            // below xct_off that was increased to a PT_LOAD large enough to compress.
            // "Manually" compress it, do not use PackExtent.
            MemBuffer buf2; buf2.allocForCompression(p_filesz);
            c_len = pack2_shlib_overlay_compress(buf2, &lowmem[p_offset], p_filesz); total_in += p_filesz;
            pack2_shlib_overlay_write(fo, buf2, p_filesz, c_len);
            Elf32_Phdr *lo_phdr = k + (Elf32_Phdr *)(1+ (Elf32_Ehdr *)&lowmem[0]);
            set_te32(&lo_phdr->p_type, Elf32_Phdr::PT_NULL);
        }
        else if ((xct_off - p_offset) < p_filesz) {
            unsigned const len = xct_off - p_offset;
            // Loader tables in non-first PT_LOAD.  (xct_off was increased.)
            // Bytes below xct_off belong to rtld, so cannot be compressed.
            // Note that asl_pack2_Shdrs() copies up to xct_off, then adds extra info.
            // Copy up to xct_off
            if (len) {
                fo->write(&lowmem[p_offset], len);  total_out += len;
                total_in += len;

                // Compressed ELF headers  FIXME: 1st PT_LOAD also did this!
                pack2_shlib_overlay_write(fo, elf_buf, u_len, c_len);  // set overlay_offset
            }
            // The rest, above xct_off.
            Extent x;
            x.offset = xct_off;
            x.size = p_filesz - len;
            packExtent(x, &ft, fo, 0, 0, true);
            Elf32_Phdr *lo_phdr = k + (Elf32_Phdr *)(1+ (Elf32_Ehdr *)&lowmem[0]);
            set_te32(&lo_phdr->p_type, Elf32_Phdr::PT_NULL);
        }
        else { // definitely compressible unless writeable
            if (!(Elf32_Phdr::PF_W & get_te32(&phdr->p_flags))) {
                // Read-only PT_LOAD, assume not written by relocations.
                // Also assume not the source for R_*_COPY relocation,
                // therefore compress it.
                Extent x;
                x.offset = p_offset;
                x.size = p_filesz;
                packExtent(x, &ft, fo, 0, 0, true);  // choose filter and method, then do it
                // De-compressing will re-create it, but otherwise ignore it.
                Elf32_Phdr *phdro = (Elf32_Phdr *)(1+ (Elf32_Ehdr *)&lowmem[0]);
                set_te32(&phdro[k].p_type, Elf32_Phdr::PT_NULL);
            }
            else {
                // Read-write PT_LOAD.
                // rtld might relocate, so we cannot compress.
                // (Could compress if not relocated; complicates run-time.)
                // Postpone writing until "slide", but account for its size.
                total_in +=  p_filesz;
            }
        }
        ++n_ptload;
    }
    return 0;  // FIXME
}

int PackLinuxElf32::pack2(OutputFile *fo, Filter &ft)
{
    Extent x;
    unsigned k;
    unsigned const is_shlib = (0!=xct_off) | is_asl;
    unsigned pre_xct_top = 0;  // offset of end of PT_LOAD _before_ xct_off

    // count passes, set ptload vars
    uip->ui_total_passes = 0;
    for (k = 0; k < e_phnum; ++k) {
        if (is_LOAD(&phdri[k])) {
            if (!is_shlib) {
                uip->ui_total_passes++;
            }
            else {
                unsigned p_flags = get_te32(&phdri[k].p_flags);
                unsigned p_offset = get_te32(&phdri[k].p_offset);
                unsigned p_filesz = get_te32(&phdri[k].p_filesz);
                if ((xct_off - p_offset) < p_filesz) { // PT_LOAD covers xct_off
                    if (!pre_xct_top && xct_off != p_offset) {
                        pre_xct_top = xct_off;
                    }
                }
                else if (p_offset < xct_off) { // candidate for pre_xct_top
                    unsigned top = p_filesz + p_offset;
                    if (pre_xct_top < top) {
                        pre_xct_top = top;
                    }
                }
                if (Elf32_Phdr::PF_W & p_flags) {
                    // rtld might write, so cannot compress
                }
                else {
                    // First PT_LOAD (partial) only if has instructions
                    if (k || xct_off < p_filesz) {
                        uip->ui_total_passes++;
                    }
                }
            }
            if (find_LOAD_gap(phdri, k, e_phnum)) {
                uip->ui_total_passes++;
            }
        }
    }

    // compress extents
    unsigned hdr_u_len = sizeof(Elf32_Ehdr) + sz_phdrs;

    total_in =  0;
    total_out = 0;
    uip->ui_pass = 0;
    ft.addvalue = 0;

    unsigned nk_f = 0; upx_uint32_t xsz_f = 0;
    for (k = 0; k < e_phnum; ++k)
    if (is_LOAD(&phdri[k])
    &&  Elf32_Phdr::PF_X & get_te32(&phdri[k].p_flags)) {
        upx_uint32_t xsz = get_te32(&phdri[k].p_filesz);
        if (xsz_f < xsz) {
            xsz_f = xsz;
            nk_f = k;
        }
    }
    if (is_shlib) {
        pack2_shlib(fo, ft, pre_xct_top);
    }
    else { // main program
        int n_ptload = 0;
        for (k = 0; k < e_phnum; ++k)
        if (is_LOAD(&phdri[k])) {
            if (ft.id < 0x40) {
                // FIXME: ??    ft.addvalue = phdri[k].p_vaddr;
            }
            unsigned p_offset = get_te32(&phdri[k].p_offset);
            unsigned p_filesz = get_te32(&phdri[k].p_filesz);
            x.offset = p_offset;
            x.size   = p_filesz;
            if ((u32_t)x.size < hdr_u_len) { // FIXME: main program, but how?
                total_in += x.size;
            }
            else {  // main program, not shared library
                if (0 == n_ptload) { // 1st PT_LOAD must cover Ehdr at 0==p_offset
                    unsigned const delta = hdr_u_len;
                    if (ft.id < 0x40) {
                        // FIXME: ??     ft.addvalue += asl_delta;
                    }
                    if ((off_t)delta == x.size) { // PT_LOAD[0] with ElfXX.Ehdr only
                        // QBE backend - http://c9x.me/compile/
                        hdr_u_len = 0;  // no fiddling necessary!
                        // &ft arg to packExtent will be zero because (k != nk_f)
                    }
                    else {
                        total_in += delta - hdr_u_len;
                        x.offset += delta;
                        x.size   -= delta;
                    }
                }
                // compressWithFilters() always assumes a "loader", so would
                // throw NotCompressible for small .data Extents, which PowerPC
                // sometimes marks as PF_X anyway.  So filter only first segment.
                packExtent(x,
                    (k==nk_f ? &ft : nullptr ), fo, hdr_u_len, 0, true);
                hdr_u_len = 0;
            }
            ++n_ptload;
        }
    }
    sz_pack2a = fpad4(fo, total_out);  // MATCH01
    total_out = up4(total_out);

    // Accounting only; ::pack3 will do the compression and output
    for (k = 0; k < e_phnum; ++k) {
        total_in += find_LOAD_gap(phdri, k, e_phnum);
    }

    if (total_in != (u32_t)file_size)
        throwEOFException();

    return 0;  // omit end-of-compression bhdr for now
}

// Determine length of gap between PT_LOAD phdr[k] and closest PT_LOAD
// which follows in the file (or end-of-file).  Optimize for common case
// where the PT_LOAD are adjacent ascending by .p_offset.  Assume no overlap.

unsigned PackLinuxElf64::find_LOAD_gap(
    Elf64_Phdr const *const phdr,
    unsigned const k,
    unsigned const nph
)
{
    if (!is_LOAD(&phdr[k])) {
        return 0;
    }
    unsigned const hi = get_te64(&phdr[k].p_offset) +
                        get_te64(&phdr[k].p_filesz);
    unsigned lo = ph.u_file_size;
    if (lo < hi)
        throwCantPack("bad input: PT_LOAD beyond end-of-file");
    unsigned j = k;
    for (;;) { // circular search, optimize for adjacent ascending
        ++j;
        if (nph==j) {
            j = 0;
        }
        if (k==j) {
            break;
        }
        if (is_LOAD(&phdr[j])) {
            unsigned const t = get_te64(&phdr[j].p_offset);
            if ((t - hi) < (lo - hi)) {
                lo = t;
                if (hi==lo) {
                    break;
                }
            }
        }
    }
    return lo - hi;
}

int PackLinuxElf64::pack2_shlib(OutputFile *fo, Filter &ft, unsigned pre_xct_top)
{
    // prepare to alter Phdrs and Shdrs
    lowmem.alloc(up8(xct_off + (!is_asl
        ? 0
        : e_shnum * sizeof(Elf64_Shdr))));
    memcpy(lowmem, file_image, xct_off);  // android omits Shdr from copy now

    MemBuffer elf_buf;
    elf_buf.allocForCompression(sz_elf_hdrs);
    unsigned u_len = sz_elf_hdrs;
    unsigned c_len = pack2_shlib_overlay_compress(elf_buf, lowmem, u_len);

    int n_ptload = 0;
    Elf64_Phdr *phdr = &phdri[0];
    for (unsigned k = 0; k < e_phnum; ++phdr, ++k)
    if (is_LOAD(phdr)) {
        unsigned p_offset = get_te64_32(&phdr->p_offset);
        unsigned p_filesz = get_te64_32(&phdr->p_filesz);
        if (!n_ptload) { // first PT_LOAD
            if (is_asl) { // Copy up to xct_off, then add 2nd copy of Shdrs
                asl_pack2_Shdrs(fo, pre_xct_top);
            }
            else {
                fo->write(&lowmem[p_offset], sz_elf_hdrs);  total_out += sz_elf_hdrs;
                total_in  += sz_elf_hdrs;
                fo->seek(sz_phdrx, SEEK_CUR);  total_out += sz_phdrx;  // leave space

                // Compare PackUnix::packExtent, especially "if (u_len)" .
                //

                unsigned const hi_offset = umin(xct_off, (unsigned)(p_filesz + p_offset));
                if (sz_elf_hdrs < hi_offset) {
                    // Loader tables in first PT_LOAD, and below xct_off.
                    //
                    fo->write(&lowmem[sz_elf_hdrs], hi_offset - sz_elf_hdrs);  total_out += hi_offset - sz_elf_hdrs;
                    total_in  += hi_offset - sz_elf_hdrs;
                    Elf64_Phdr *lo_phdr = k + (Elf64_Phdr *)(1+ (Elf64_Ehdr *)&lowmem[0]);
                    set_te32(&lo_phdr->p_flags, Elf64_Phdr::PF_X | get_te32(&lo_phdr->p_flags));
                }
            }
            pack2_shlib_overlay_init(fo); // set overlay_offset
            // Compressed ELF headers go first.
            pack2_shlib_overlay_write(fo, elf_buf, u_len, c_len);

            // The (compressible) remainder above xct_off in first PT_LOAD
            if (         p_filesz > (xct_off - p_offset)) {
                Extent x;
                x.size = p_filesz - (xct_off - p_offset);
                x.offset = xct_off;
                packExtent(x, &ft, fo, 0, 0, true);
            }
        }  // end first PT_LOAD
        else if ((p_filesz + p_offset) <= xct_off) {
            // Not first PT_LOAD, but below xct_offset.  Thus "interloper" PT_LOAD,
            // below xct_off that was increased to a PT_LOAD large enough to compress.
            // "Manually" compress it, do not use PackExtent.
            MemBuffer buf2; buf2.allocForCompression(p_filesz);
            c_len = pack2_shlib_overlay_compress(buf2, &lowmem[p_offset], p_filesz); total_in += p_filesz;
            pack2_shlib_overlay_write(fo, buf2, p_filesz, c_len);
            Elf64_Phdr *lo_phdr = k + (Elf64_Phdr *)(1+ (Elf64_Ehdr *)&lowmem[0]);
            set_te32(&lo_phdr->p_type, Elf64_Phdr::PT_NULL);
        }
        else if ((xct_off - p_offset) < p_filesz) {
            unsigned const len = xct_off - p_offset;
            // Loader tables in non-first PT_LOAD.  (xct_off was increased.)
            // Bytes below xct_off belong to rtld, so cannot be compressed.
            // Note that asl_pack2_Shdrs() copies up to xct_off, then adds extra info.
            // Copy up to xct_off
            if (len) {
                fo->write(&lowmem[p_offset], len);  total_out += len;
                total_in += len;

                // Compressed ELF headers  FIXME: 1st PT_LOAD also did this!
                pack2_shlib_overlay_write(fo, elf_buf, u_len, c_len);  // set overlay_offset
            }
            // The rest, above xct_off.
            Extent x;
            x.offset = xct_off;
            x.size = p_filesz - len;
            packExtent(x, &ft, fo, 0, 0, true);
            Elf64_Phdr *lo_phdr = k + (Elf64_Phdr *)(1+ (Elf64_Ehdr *)&lowmem[0]);
            set_te32(&lo_phdr->p_type, Elf64_Phdr::PT_NULL);
        }
        else { // definitely compressible unless writeable
            if (!(Elf64_Phdr::PF_W & get_te32(&phdr->p_flags))) {
                // Read-only PT_LOAD, assume not written by relocations.
                // Also assume not the source for R_*_COPY relocation,
                // therefore compress it.
                Extent x;
                x.offset = p_offset;
                x.size = p_filesz;
                packExtent(x, &ft, fo, 0, 0, true);  // choose filter and method, then do it
                // De-compressing will re-create it, but otherwise ignore it.
                Elf64_Phdr *phdro = (Elf64_Phdr *)(1+ (Elf64_Ehdr *)&lowmem[0]);
                set_te32(&phdro[k].p_type, Elf32_Phdr::PT_NULL);
            }
            else {
                // Read-write PT_LOAD.
                // rtld might relocate, so we cannot compress.
                // (Could compress if not relocated; complicates run-time.)
                // Postpone writing until "slide", but account for its size.
                total_in +=  p_filesz;
            }
        }
        ++n_ptload;
    }
    return 0;  // FIXME
}

int PackLinuxElf64::pack2(OutputFile *fo, Filter &ft)
{
    Extent x;
    unsigned k;
    // See comment in Elf32_Linux::canPack().  is_asl does not work.
    // 64-bit ARM (aarch64) also has no ARM_ATTRIBUTES.
    //is_asl = (!!saved_opt_android_shlib) << 1;  // bit 1; see is_shlib
    is_asl = 0;
    unsigned const is_shlib = (0!=xct_off) | is_asl;
    unsigned pre_xct_top = 0;  // offset of end of PT_LOAD _before_ xct_off

    if (Elf64_Ehdr::EM_ARM==get_te16(&ehdri.e_machine)) {
        sec_arm_attr = elf_find_section_type(Elf64_Shdr::SHT_ARM_ATTRIBUTES);
    }
    // count passes, set ptload vars
    uip->ui_total_passes = 0;
    for (k = 0; k < e_phnum; ++k) {
        if (is_LOAD(&phdri[k])) {
            if (!is_shlib) {
                uip->ui_total_passes++;
            }
            else {
                unsigned p_flags = get_te32(&phdri[k].p_flags);
                unsigned p_offset = get_te64_32(&phdri[k].p_offset);
                unsigned p_filesz = get_te64_32(&phdri[k].p_filesz);
                if ((xct_off - p_offset) < p_filesz) { // PT_LOAD covers xct_off
                    if (!pre_xct_top && xct_off != p_offset) {
                        pre_xct_top = xct_off;
                    }
                }
                else if (p_offset < xct_off) { // candidate for pre_xct_top
                    unsigned top = p_filesz + p_offset;
                    if (pre_xct_top < top) {
                        pre_xct_top = top;
                    }
                }
                if (Elf64_Phdr::PF_W & p_flags) {
                    // rtld might write, so cannot compress
                }
                else {
                    // First PT_LOAD (partial) only if has instructions
                    if (k || xct_off < p_filesz) {
                        uip->ui_total_passes++;
                    }
                }
            }
            if (find_LOAD_gap(phdri, k, e_phnum)) {
                uip->ui_total_passes++;
            }
        }
    }

    // compress extents
    unsigned hdr_u_len = sizeof(Elf64_Ehdr) + sz_phdrs;

    total_in =  0;
    total_out = 0;
    uip->ui_pass = 0;
    ft.addvalue = 0;

    unsigned nk_f = 0; upx_uint32_t xsz_f = 0;
    for (k = 0; k < e_phnum; ++k)
    if (is_LOAD(&phdri[k])
    &&  Elf32_Phdr::PF_X & get_te32(&phdri[k].p_flags)) {
        unsigned xsz = get_te64_32(&phdri[k].p_filesz);
        if (xsz_f < xsz) {
            xsz_f = xsz;
            nk_f = k;
        }
    }
    if (is_shlib) {
        pack2_shlib(fo, ft, pre_xct_top);
    }
    else { // main program
        int n_ptload = 0;
        for (k = 0; k < e_phnum; ++k)
        if (is_LOAD(&phdri[k])) {
            if (ft.id < 0x40) {
                // FIXME: ??    ft.addvalue = phdri[k].p_vaddr;
            }
            unsigned p_offset = get_te64_32(&phdri[k].p_offset);
            unsigned p_filesz = get_te64_32(&phdri[k].p_filesz);
            x.offset = p_offset;
            x.size   = p_filesz;
            if ((u32_t)x.size < hdr_u_len) { // FIXME: main program, but how?
                total_in += x.size;
            }
            else {  // main program, not shared library
                if (0 == n_ptload) { // 1st PT_LOAD must cover Ehdr at 0==p_offset
                    unsigned const delta = hdr_u_len;
                    if (ft.id < 0x40) {
                        // FIXME: ??     ft.addvalue += asl_delta;
                    }
                    if ((off_t)delta == x.size) { // PT_LOAD[0] with ElfXX.Ehdr only
                        // QBE backend - http://c9x.me/compile/
                        hdr_u_len = 0;  // no fiddling necessary!
                        // &ft arg to packExtent will be zero because (k != nk_f)
                    }
                    else {
                        total_in += delta - hdr_u_len;
                        x.offset += delta;
                        x.size   -= delta;
                    }
                }
                // compressWithFilters() always assumes a "loader", so would
                // throw NotCompressible for small .data Extents, which PowerPC
                // sometimes marks as PF_X anyway.  So filter only first segment.
                packExtent(x,
                    (k==nk_f ? &ft : nullptr ), fo, hdr_u_len, 0, true);
                hdr_u_len = 0;
            }
            ++n_ptload;
        }
    }
    sz_pack2a = fpad4(fo, total_out);  // MATCH01
    total_out = up4(total_out);

    // Accounting only; ::pack3 will do the compression and output
    for (k = 0; k < e_phnum; ++k) {
        total_in += find_LOAD_gap(phdri, k, e_phnum);
    }

    if (total_in != (u32_t)file_size)
        throwEOFException();

    return 0;  // omit end-of-compression bhdr for now
}

// Filter 0x50, 0x51 assume HostPolicy::isLE
static const int *
ARM_getFilters(bool const isBE)
{
    static const int f50[] = { 0x50, FT_END };
    static const int f51[] = { 0x51, FT_END };
    if (isBE)
        return f51;
    return f50;
}

const int *
PackLinuxElf32armBe::getFilters() const
{
    return ARM_getFilters(true);
}

const int *
PackLinuxElf32armLe::getFilters() const
{
    return ARM_getFilters(false);
}

const int *
PackLinuxElf32mipseb::getFilters() const
{
    static const int f_none[] = { FT_END };
    return f_none;
}

const int *
PackLinuxElf32mipsel::getFilters() const
{
    static const int f_none[] = { FT_END };
    return f_none;
}

// October 2011: QNX 6.3.0 has no unique signature?
int PackLinuxElf32::ARM_is_QNX(void)
{
    if (Elf32_Ehdr::EM_ARM==get_te16(&ehdri.e_machine)
    &&  Elf32_Ehdr::ELFDATA2MSB== ehdri.e_ident[Elf32_Ehdr::EI_DATA]
    &&  Elf32_Ehdr::ELFOSABI_ARM==ehdri.e_ident[Elf32_Ehdr::EI_OSABI]
    &&  0x100000==(page_mask & get_te32(&phdri[0].p_vaddr))) {
        Elf32_Phdr const *phdr = phdri;
        for (int j = get_te16(&ehdri.e_phnum); --j>=0; ++phdr) {
            if (Elf32_Phdr::PT_INTERP==get_te32(&phdr->p_type)) {
                char interp[64];
                unsigned const sz_interp = get_te32(&phdr->p_filesz);
                unsigned const pos_interp = get_te32(&phdr->p_offset);
                if (sz_interp <= sizeof(interp)
                &&  (sz_interp + pos_interp) <= (unsigned)file_size) {
                    fi->seek(pos_interp, SEEK_SET);
                    fi->readx(interp, sz_interp);
                    for (int k = sz_interp - 5; k>=0; --k) {
                        if (0==memcmp("ldqnx", &interp[k], 5))
                            return 1;
                    }
                }
            }
        }
    }
    return 0;
}

void PackLinuxElf32::ARM_defineSymbols(Filter const *ft)
{
    PackLinuxElf32::defineSymbols(ft);

#define MAP_PRIVATE      2     /* UNIX standard */
#define MAP_FIXED     0x10     /* UNIX standard */
#define MAP_ANONYMOUS 0x20     /* UNIX standard */
#define MAP_PRIVANON     3     /* QNX anonymous private memory */
    unsigned mflg = MAP_PRIVATE | MAP_ANONYMOUS;
    if (ARM_is_QNX())
        mflg = MAP_PRIVANON;
    linker->defineSymbol("MFLG", mflg);
}

void PackLinuxElf32armLe::defineSymbols(Filter const *ft)
{
    ARM_defineSymbols(ft);
}

void PackLinuxElf32armBe::defineSymbols(Filter const *ft)
{
    ARM_defineSymbols(ft);
}

void PackLinuxElf64arm::defineSymbols(Filter const *ft)
{
    PackLinuxElf64::defineSymbols(ft);

#define MAP_PRIVATE      2     /* UNIX standard */
#define MAP_FIXED     0x10     /* UNIX standard */
#define MAP_ANONYMOUS 0x20     /* UNIX standard */
#define MAP_PRIVANON     3     /* QNX anonymous private memory */
    unsigned mflg = MAP_PRIVATE | MAP_ANONYMOUS;
    //if (ARM_is_QNX())
    //    mflg = MAP_PRIVANON;
    linker->defineSymbol("MFLG", mflg);
}

void PackLinuxElf32mipseb::defineSymbols(Filter const *ft)
{
    PackLinuxElf32::defineSymbols(ft);
}

void PackLinuxElf32mipsel::defineSymbols(Filter const *ft)
{
    PackLinuxElf32::defineSymbols(ft);
}

// ::forward_Shdrs adds info for the benefit of gdb and Android dlopen().
// De-compression (runtime and offline) ignores the added information
// because it uses the de-compressed Ehdr etc.
// All the added space is redundant; libbfd should take a hint:
// if no Shdrs, then use PT_DYNAMIC instead.
// (.ARM_attributes (ARM_ATTRIBUTES) is not redundant.)
//
// want_types_mask: SHT_PROGBITS is needed else gdb complains:
//    /build/gdb-MVZsgD/gdb-10.1/gdb/symfile.c:878: internal-error: sect_index_text not initialized
// and Continuing is not reasonable.
// However, SHT_PROGBITS with compression gives:
//    BFD: warning: ./libmain.so.upx has a section extending past end of file
//    BFD: warning: ./libmain.so.upx has a section extending past end of file
//    BFD: warning: ./libmain.so.upx has a section extending past end of file
//    warning: Loadable section ".text" outside of ELF segments
//    warning: Loadable section ".plt" outside of ELF segments
// because compression gives smaller extents, with no reasonable _Shdr fields.
// At least gdb can continue.

unsigned PackLinuxElf32::forward_Shdrs(OutputFile *fo, Elf32_Ehdr *const eho)
{
    if (!fo) {
        return 0;
    }
    unsigned penalty = total_out;
    if (sec_arm_attr) { // Forward select _Shdr.  EM_ARM (and Android ?) only.
        // Keep _Shdr for rtld data (below xct_off).
        // Discard _Shdr for compressed regions, except ".text" for gdb.
        // Keep _Shdr for SHF_WRITE.
        // Discard _Shdr with (0==sh_addr), except _Shdr[0]
        // Keep ARM_ATTRIBUTES
        unsigned const want_types_mask = 0
            | 1u<<SHT_PROGBITS  // see comment above, and special code below
            | 1u<<SHT_HASH
            | 1u<<SHT_DYNAMIC
            | 1u<<SHT_NOTE
            | 1u<<SHT_REL
            | 1u<<SHT_RELA
            | 1u<<SHT_RELR
            | 1u<<SHT_DYNSYM  // but not SHT_SYMTAB because compression confuses gdb
            | 1u<<SHT_STRTAB  // .shstrtab and DYNSYM.sh_link; not SYMTAB.sh_link
            | 1u<<SHT_INIT_ARRAY
            | 1u<<SHT_FINI_ARRAY
            | 1u<<SHT_PREINIT_ARRAY
            | 1u<<(0x1f & SHT_GNU_versym)
            | 1u<<(0x1f & SHT_GNU_verneed)
            | 1u<<(0x1f & SHT_GNU_verdef)
            | 1u<<(0x1f & SHT_GNU_HASH);

        u32_t xct_off_hi = 0;
        Elf32_Phdr *ptr = phdri, *ptr_end = &phdri[e_phnum];
        for (; ptr < ptr_end; ++ptr) {
            if (is_LOAD(ptr)) {
                u32_t hi = get_te32(&ptr->p_filesz)
                    + get_te32(&ptr->p_offset);
                if (xct_off < hi) {
                    xct_off_hi = hi;
                    break;
                }
            }
        }

        MemBuffer mb_ask_for(e_shnum * sizeof(eho->e_shnum));
        memset(mb_ask_for, 0, mb_ask_for.getSize());
        unsigned short *const ask_for = (unsigned short *)mb_ask_for.getVoidPtr();

        MemBuffer mb_shdro(e_shnum * sizeof(*shdri));
        Elf32_Shdr *sh_out0 = (Elf32_Shdr *)mb_shdro.getVoidPtr();
        Elf32_Shdr *sh_out = sh_out0;
        Elf32_Shdr *sh_in = shdri;
        Elf32_Shdr *n_shstrsec = nullptr;

        // Some binutils does tail merging on section names; we don't.
        // ".plt" == (4+ ".rel.plt"); ".hash" == (4+ ".gnu.hash")
        MemBuffer mb_shstrings(100 + 2*get_te32(&sh_in[e_shstrndx].sh_size));
        char *ptr_shstrings = (char *)&mb_shstrings[0];
        *ptr_shstrings++ = '\0';

        memset(sh_out, 0, sizeof(*sh_out));  // blank sh_out[0]
        ++sh_in; ++sh_out; unsigned n_sh_out = 1;

        for (unsigned j = 1; j < e_shnum; ++j, ++sh_in) {
            unsigned sh_name   = get_te32(&sh_in->sh_name);
            unsigned sh_type   = get_te32(&sh_in->sh_type);
            unsigned sh_flags  = get_te32(&sh_in->sh_flags);
            //unsigned sh_addr   = get_te32(&sh_in->sh_addr);
            unsigned sh_offset = get_te32(&sh_in->sh_offset);
            unsigned sh_size   = get_te32(&sh_in->sh_size);
            unsigned sh_info   = get_te32(&sh_in->sh_info);
            char const *name = &shstrtab[sh_name];
            if (ask_for[j]) { // Some previous _Shdr requested  me
                // Tell them my new index
                set_te32(&sh_out0[ask_for[j]].sh_info, n_sh_out);  // sh_info vs st_shndx
            }
            if (sh_info < e_shnum) { // wild sh_info abounds!
                ask_for[sh_info] = j;  // Enter my request, if any
            }
            if (   (sh_offset && sh_offset < xct_off)
                || (j == e_shstrndx)
                || (Elf32_Shdr::SHF_WRITE & sh_flags)
                || (sh_type < Elf32_Shdr::SHT_LOPROC
                    && want_types_mask & (1<<(0x1f & sh_type)))
                || (Elf32_Shdr::SHT_ARM_ATTRIBUTES == sh_type)
            ) {
                *sh_out = *sh_in;  // *sh_in is a candidate for fowarding
                if (sh_offset > xct_off) { // may slide down: earlier compression
                    if (sh_offset >= xct_off_hi) { // easy: so_slide down
                        if (Elf32_Shdr::SHT_ARM_ATTRIBUTES != sh_type) {
                            slide_sh_offset(sh_out);
                        }
                    }
                    else { // somewhere in compressed; try proportional (aligned)
                        // But note that PROGBITS without SHF_ALLOC
                        // will be dropped below.
                        u32_t const slice = xct_off + (~0xFu & (unsigned)(
                             (sh_offset - xct_off) *
                            ((sh_offset - xct_off) / (float)(xct_off_hi - xct_off))));
                        //set_te32(&sh_out->sh_addr,   slice);
                        set_te32(&sh_out->sh_offset, slice);
                    }
                    u32_t const max_sz = total_out - get_te32(&sh_out->sh_offset);
                    if (sh_size > max_sz) { // avoid complaint "extends beyond EOF"
                        set_te32(&sh_out->sh_size, max_sz);
                    }
                }
                if (j == e_shstrndx) { // changes Elf32_Ehdr itself
                    set_te16(&eho->e_shstrndx, sh_out -
                        (Elf32_Shdr *)mb_shdro.getVoidPtr());
                }
                if (Elf32_Shdr::SHT_ARM_ATTRIBUTES == sh_type
                ||  (SHT_NOTE == sh_type && xct_off < sh_offset)
                ) { // append a copy
                    set_te32(&sh_out->sh_offset, total_out);
                    fi->seek((upx_off_t)sh_offset, SEEK_SET);
                    fi->read(ibuf,  sh_size);
                    fo->write(ibuf, sh_size);
                    total_out +=    sh_size;
                } else
                if (SHT_PROGBITS == sh_type) {
                    if (!(Elf32_Shdr::SHF_ALLOC & sh_flags)) {
                        // .debug_*, .gnu_debuglink etc.  Typically compressed
                        // but not in RAM, and gdb (BFD) gets confused.
                        continue;  // OMIT the commit: do not forward
                    } else
                    if (sh_offset <= xct_off
                    &&  0 == strcmp(".text", name) ) {
                        // .text was compressed (but perhaps omitting some leading
                        // portion, if less than 4 PT_LOAD)
                        set_te32(&sh_out->sh_size, so_slide + sh_size);
                        // FIXME: so_slide is negative; avoid negative result
                    }
                } else
                if (SHT_STRTAB == sh_type) {
                    if (j == e_shstrndx) {
                        n_shstrsec = sh_out;
                    } else
                    if (strcmp(".dynstr",    name)) {
                        continue;  // OMIT the commit of non-global symbol names
                    }
                }
                set_te32(&sh_out->sh_name, ptr_shstrings - (char *)mb_shstrings.getVoidPtr());
                do { // stupid MSVC lacks stpcpy()
                    *ptr_shstrings++ = *name;
                } while (*name++);
                ++sh_out; ++n_sh_out;  // actually commit the fowarding
            }
        }
        unsigned len = ptr_shstrings - (char *)mb_shstrings.getVoidPtr();
        set_te32(&n_shstrsec->sh_offset, total_out);
        set_te32(&n_shstrsec->sh_size, len);
        fo->write(mb_shstrings, len);
        total_out += len;

        total_out = fpad4(fo, total_out);  // align _Shdr[]
        set_te32(&eho->e_shoff, total_out);
        len = (char *)sh_out - (char *)mb_shdro.getVoidPtr();
        set_te16(&eho->e_shnum, len / sizeof(*sh_out));
        set_te16(&eho->e_shentsize, sizeof(Elf32_Shdr));
        fo->write(mb_shdro, len);
        total_out += len;

        // Try to pacify gdb (before DT_INIT) by making it look like
        // the compressed PT_LOAD extends all the way to the next PT_LOAD,
        // with no gap in address space.  Thus gdb should not complain about
        // "Loadable section "..." [Shdr] outside of ELF segments [PT_LOAD]".
        // gdb still "warning: section ... not found in .gnu_debugdata"
        // because .gdb_debugdata is not present (or gets removed),
        // but that is separate and "just" a warning.
        ptr = (Elf32_Phdr *)(1+ eho);
        if (0)  // FIXME:  This is not required?
        for (ptr_end = &ptr[e_phnum]; ptr < ptr_end; ++ptr) {
            if (is_LOAD(ptr)) {
                Elf32_Phdr *ptr2 = 1+ ptr;
                for (; ptr2 < ptr_end; ++ptr2) {
                    if (is_LOAD(ptr2)) {
                        unsigned pmask = 0u - get_te32(&ptr2->p_align);
                        set_te32(&ptr->p_memsz,
                            (pmask & get_te32(&ptr2->p_vaddr)) -
                            (pmask & get_te32(&ptr ->p_vaddr)) );
                        ptr  = ptr_end;  // force end of outer loop
                        ptr2 = ptr_end;  // force end of inner loop
                    }
                }
            }
        }

        fo->seek(0, SEEK_SET);
        fo->rewrite(eho, sizeof(*eho));
        fo->seek(0, SEEK_END);
    }
    penalty = total_out - penalty;
    info("Android penalty = %d bytes", penalty);
    return penalty;
}

unsigned PackLinuxElf64::forward_Shdrs(OutputFile *fo, Elf64_Ehdr *const eho)
{
    if (!fo) {
        return 0;
    }
    unsigned penalty = total_out;
    if (Elf64_Ehdr::EM_AARCH64 == e_machine
    &&  saved_opt_android_shlib) { // Forward select _Shdr
        // Keep _Shdr for rtld data (below xct_off).
        // Discard _Shdr for compressed regions, except ".text" for gdb.
        // Keep _Shdr with SHF_WRITE.
        // Keep ARM_ATTRIBUTES
        // Discard _Shdr with (0==sh_addr), except _Shdr[0]
        unsigned const want_types_mask =
              1u<<SHT_SYMTAB
            | 1u<<SHT_RELA
            | 1u<<SHT_PROGBITS  // see comment above
            | 1u<<SHT_HASH
            | 1u<<SHT_DYNAMIC
            | 1u<<SHT_NOTE
            | 1u<<SHT_REL
            | 1u<<SHT_RELR
            | 1u<<SHT_DYNSYM
            | 1u<<SHT_STRTAB  // .shstrtab and .dynstr
            | 1u<<SHT_INIT_ARRAY
            | 1u<<SHT_FINI_ARRAY
            | 1u<<SHT_PREINIT_ARRAY
            | 1u<<(0x1f & SHT_GNU_versym)
            | 1u<<(0x1f & SHT_GNU_verneed)
            | 1u<<(0x1f & SHT_GNU_verdef)
            | 1u<<(0x1f & SHT_GNU_HASH);

        upx_uint64_t xct_off_hi = 0;
        Elf64_Phdr const *ptr = phdri, *ptr_end = &phdri[e_phnum];
        for (; ptr < ptr_end; ++ptr) {
            if (is_LOAD(ptr)) {
                upx_uint64_t hi = get_te64(&ptr->p_filesz)
                    + get_te64(&ptr->p_offset);
                if (xct_off < hi) {
                    xct_off_hi = hi;
                    break;
                }
            }
        }

        MemBuffer mb_ask_for(e_shnum * sizeof(eho->e_shnum));
        memset(mb_ask_for, 0, mb_ask_for.getSize());
        unsigned short *const ask_for = (unsigned short *)mb_ask_for.getVoidPtr();

        MemBuffer mb_shdro(e_shnum * sizeof(*shdri));
        Elf64_Shdr *sh_out0 = (Elf64_Shdr *)mb_shdro.getVoidPtr();
        Elf64_Shdr *sh_out = sh_out0;
        Elf64_Shdr *sh_in = shdri;

        memset(sh_out, 0, sizeof(*sh_out));  // blank sh_out[0]
        ++sh_in; ++sh_out; unsigned n_sh_out = 1;

        for (unsigned j = 1; j < e_shnum; ++j, ++sh_in) {
            char const *sh_name = &shstrtab[get_te32(&sh_in->sh_name)];
            (void)sh_name;  // debugging
            unsigned sh_type = get_te32(&sh_in->sh_type);
            u64_t sh_flags   = get_te64(&sh_in->sh_flags);
            u64_t sh_addr    = get_te64(&sh_in->sh_addr);
            u64_t sh_offset  = get_te64(&sh_in->sh_offset);
            u64_t sh_size    = get_te64(&sh_in->sh_size);
            unsigned sh_info = get_te32(&sh_in->sh_info);
            if (ask_for[j]) { // Some previous _Shdr requested  me
                // Tell them my new index
                set_te32(&sh_out0[ask_for[j]].sh_info, n_sh_out);  // sh_info vs st_shndx
            }
            if (sh_info < e_shnum) { // wild sh_info abounds!
                ask_for[sh_info] = j;  // Enter my request, if any
            }
            if (   (sh_offset && sh_offset < xct_off)
                || (Elf64_Shdr::SHF_WRITE & sh_flags)
                || (j == e_shstrndx)
                || (sec_arm_attr == sh_in)
                || (want_types_mask & (1<<(0x1f & sh_type)))
                || (Elf32_Shdr::SHT_ARM_ATTRIBUTES == sh_type)
            ) {
                *sh_out = *sh_in;
                if (sh_offset > xct_off) { // may slide down: earlier compression
                    if (sh_offset >= xct_off_hi) { // easy: so_slide down
                        if (sh_out->sh_addr) // change only if non-zero
                        set_te64(&sh_out->sh_addr,   so_slide + sh_addr +
                            (is_asl ? asl_delta : 0));
                        set_te64(&sh_out->sh_offset, so_slide + sh_offset);
                    }
                    else { // somewhere in compressed; try proportional (aligned)
                        u64_t const slice = xct_off + (~0xFu & (unsigned)(
                             (sh_offset - xct_off) *
                            ((sh_offset - xct_off) / (float)(xct_off_hi - xct_off))));
                        set_te64(&sh_out->sh_addr,   slice);
                        set_te64(&sh_out->sh_offset, slice);
                    }
                    u64_t const max_sz = total_out - get_te64(&sh_out->sh_offset);
                    if (sh_size > max_sz) { // avoid complaint "extends beyond EOF"
                        set_te64(&sh_out->sh_size, max_sz);
                    }
                }
                if (j == e_shstrndx) { // changes Elf64_Ehdr itself
                    set_te16(&eho->e_shstrndx, sh_out -
                        (Elf64_Shdr *)mb_shdro.getVoidPtr());
                }
                if (j == e_shstrndx
                ||  sec_arm_attr == sh_in
                ||  (SHT_NOTE == sh_type && xct_off < sh_offset)
                ) { // append a copy
                    set_te64(&sh_out->sh_offset, total_out);
                    fi->seek((upx_off_t)sh_offset, SEEK_SET);
                    fi->read(ibuf,  sh_size);
                    fo->write(ibuf, sh_size);
                    total_out +=    sh_size;
                } else
                if (SHT_PROGBITS == sh_type) {
                    if (sh_offset <= xct_off
                    &&  0 == strcmp(".text", shstrtab + get_te32(&sh_in->sh_name)) ) {
                        // .text was compressed (but perhaps omitting some leading
                        // portion, if less than 4 PT_LOAD)
                        set_te64(&sh_out->sh_size, so_slide + sh_size);
                    } else
                    if (0 == sh_in->sh_addr) { // .gnu_debuglink etc
                        set_te64(&sh_out->sh_offset, so_slide + sh_offset);
                    }
                }
                // Exploration for updating Shdrs from Dynamic, to pacify Android.
                // Motivated by --force-pie with an input shared library;
                // see https://github.com/upx/upx/issues/694
                // But then realized that the pointed-to regions typically were
                // just above Elfhdrs (overlay_offset), so the data would be
                // incorrect because occupied by compressed PT_LOADS.
                // But don't lose the code, so comment-out via "if (0)".
                if (0) switch(sh_type) { // start {Shdr, Dynamic} pairs
                case SHT_DYNSYM: {
                    set_te64(&sh_out->sh_addr, elf_unsigned_dynamic(Elf64_Dyn::DT_SYMTAB));
                    sh_out->sh_offset = sh_out->sh_addr;
                } break;
                case SHT_STRTAB: {
                    if (e_shstrndx != j) {
                        set_te64(&sh_out->sh_addr, elf_unsigned_dynamic(Elf64_Dyn::DT_STRTAB));
                        sh_out->sh_offset = sh_out->sh_addr;
                    }
                } break;
                case SHT_GNU_versym: {
                    set_te64(&sh_out->sh_addr, elf_unsigned_dynamic(Elf64_Dyn::DT_VERSYM));
                    sh_out->sh_offset = sh_out->sh_addr;
                } break;
                case SHT_GNU_verneed: {
                    set_te64(&sh_out->sh_addr, elf_unsigned_dynamic(Elf64_Dyn::DT_VERNEED));
                    sh_out->sh_offset = sh_out->sh_addr;
                } break;
                case SHT_GNU_HASH: {
                    set_te64(&sh_out->sh_addr, elf_unsigned_dynamic(Elf64_Dyn::DT_GNU_HASH));
                    sh_out->sh_offset = sh_out->sh_addr;
                } break;
                case SHT_HASH: {
                    set_te64(&sh_out->sh_addr, elf_unsigned_dynamic(Elf64_Dyn::DT_HASH));
                    sh_out->sh_offset = sh_out->sh_addr;
                } break;
                case SHT_RELA: {
                    if (0==strcmp(".rela.dyn", sh_name)) {
                        set_te64(&sh_out->sh_addr, elf_unsigned_dynamic(Elf64_Dyn::DT_RELA));
                        sh_out->sh_offset = sh_out->sh_addr;
                    }
                    if (0==strcmp(".rela.plt", sh_name)) {
                        set_te64(&sh_out->sh_addr, elf_unsigned_dynamic(Elf64_Dyn::DT_JMPREL));
                        sh_out->sh_offset = sh_out->sh_addr;
                    }
                } break;
                } // end {Shdr, Dynamic} pairs
                ++sh_out; ++n_sh_out;
            }
        }
        total_out = fpad8(fo, total_out);
        set_te64(&eho->e_shoff, total_out);
        unsigned len = (char *)sh_out - (char *)mb_shdro.getVoidPtr();
        set_te16(&eho->e_shnum, len / sizeof(*sh_out));
        set_te16(&eho->e_shentsize, sizeof(Elf64_Shdr));
        fo->write(mb_shdro, len);
        total_out += len;
        fo->seek(0, SEEK_SET);
        fo->rewrite(eho, sizeof(*eho));
        fo->seek(0, SEEK_END);
    }
    penalty = total_out - penalty;
    info("Android penalty = %d bytes", penalty);
    return penalty;
}

void PackLinuxElf32::pack4(OutputFile *fo, Filter &ft)
{
    if (!xct_off) {
        overlay_offset = sz_elf_hdrs + sizeof(linfo);
    }

    cprElfHdr4 *eho = !xct_off
            ? (cprElfHdr4 *)(void *)&elfout  // not shlib  FIXME: ugly casting
            : (cprElfHdr4 *)lowmem.getVoidPtr();  // shlib
    unsigned penalty = forward_Shdrs(fo, &eho->ehdr); (void)penalty;

    if (opt->o_unix.preserve_build_id) { // FIXME: co-ordinate with forward_Shdrs
        // calc e_shoff here and write shdrout, then o_shstrtab
        //NOTE: these are pushed last to ensure nothing is stepped on
        //for the UPX structure.
        total_out = fpad4(fo, total_out);
        set_te32(&eho->ehdr.e_shoff, total_out);

        unsigned const ssize = sizeof(shdrout);
        unsigned const ssize1 = get_te32(&shdrout.shdr[1].sh_size);
        unsigned const ssize2 = get_te32(&shdrout.shdr[2].sh_size);

        set_te32(&shdrout.shdr[2].sh_offset,          ssize + total_out);
        set_te32(&shdrout.shdr[1].sh_offset, ssize2 + ssize + total_out);

        fo->write(&shdrout, ssize); total_out += ssize;

        fo->write(o_shstrtab, ssize2); total_out += ssize2;
        fo->write(buildid_data, ssize1); total_out += ssize1;
    }

    // ph.u_len and ph.c_len are leftover from earliest days when there was
    // only one compressed extent.  Use a good analogy for multiple extents.
    ph.u_len = file_size;
    ph.c_len = total_out;
    super::pack4(fo, ft);  // write PackHeader and overlay_offset

    fo->seek(0, SEEK_SET);
    if (0!=xct_off) {  // shared library
        { // Shouldn't this special case be handled earlier?
            if (overlay_offset < xct_off) {
                Elf32_Phdr *phdro = (Elf32_Phdr *)&eho->phdr;
                set_te32(&phdro->p_flags, Elf32_Phdr::PF_X | get_te32(&phdro->p_flags));
            }
        }
        if (!sec_arm_attr && !saved_opt_android_shlib) {
            // Make it abundantly clear that there are no Elf32_Shdr in this shlib
            eho->ehdr.e_shoff = 0;
            set_te16(&eho->ehdr.e_shentsize, sizeof(Elf32_Shdr));  // Android bug: cannot use 0
            eho->ehdr.e_shnum = 0;
            eho->ehdr.e_shstrndx = 0;
        }
        fo->rewrite(eho, sizeof(ehdri) + e_phnum * sizeof(*phdri));
        fo->seek(linfo_off, SEEK_SET);
        fo->rewrite(&linfo, sizeof(linfo));  // new info: l_checksum, l_size

        if (jni_onload_va) {
            unsigned tmp = sz_pack2 + get_te32(&eho->phdr[C_TEXT].p_vaddr);
            tmp |= (Elf32_Ehdr::EM_ARM==e_machine);  // THUMB mode
            set_te32(&tmp, tmp);
            fo->seek(ptr_udiff_bytes(&jni_onload_sym->st_value, file_image), SEEK_SET);
            fo->rewrite(&tmp, sizeof(tmp));
        }
    }
    else { // not shlib
        // Cannot pre-round .p_memsz.  If .p_filesz < .p_memsz, then kernel
        // tries to make .bss, which requires PF_W.
        // But strict SELinux (or PaX, grSecurity) disallows PF_W with PF_X.
        set_te32(&eho->phdr[C_TEXT].p_filesz, sz_pack2 + lsize);
                  eho->phdr[C_TEXT].p_memsz = eho->phdr[C_TEXT].p_filesz;

        fo->seek(0, SEEK_SET);
        fo->rewrite(&eho->ehdr, sizeof(Elf32_Ehdr) + 2* sizeof(Elf32_Phdr));  // C_BASE, C_TEXT
        fo->seek(overlay_offset - sizeof(l_info), SEEK_SET);
        fo->rewrite(&linfo, sizeof(linfo));
    }
}

void PackLinuxElf64::pack4(OutputFile *fo, Filter &ft)
{
    if (!xct_off) {
        overlay_offset = sz_elf_hdrs + sizeof(linfo);
    }

    cprElfHdr4 *eho = !xct_off
            ? &elfout  // not shlib
            : (cprElfHdr4 *)lowmem.getVoidPtr();  // shlib
    unsigned penalty = forward_Shdrs(fo, &eho->ehdr); (void)penalty;

    if (opt->o_unix.preserve_build_id) { // FIXME: co-ordinate with forward_Shdrs
        // calc e_shoff here and write shdrout, then o_shstrtab
        //NOTE: these are pushed last to ensure nothing is stepped on
        //for the UPX structure.
        total_out = fpad4(fo, total_out);
        set_te64(&eho->ehdr.e_shoff, total_out);

        unsigned const ssize = sizeof(shdrout);
        unsigned const ssize1 = get_te64(&shdrout.shdr[1].sh_size);
        unsigned const ssize2 = get_te64(&shdrout.shdr[2].sh_size);

        set_te64(&shdrout.shdr[2].sh_offset,          ssize + total_out);
        set_te64(&shdrout.shdr[1].sh_offset, ssize2 + ssize + total_out);

        fo->write(&shdrout, ssize); total_out += ssize;

        fo->write(o_shstrtab, ssize2); total_out += ssize2;
        fo->write(buildid_data, ssize1); total_out += ssize1;
    }

    // ph.u_len and ph.c_len are leftover from earliest days when there was
    // only one compressed extent.  Use a good analogy for multiple extents.
    ph.u_len = file_size;
    ph.c_len = total_out;
    super::pack4(fo, ft);  // write PackHeader and overlay_offset

    fo->seek(0, SEEK_SET);
    if (0!=xct_off) {  // shared library
        { // Shouldn't this special case be handled earlier?
            if (overlay_offset < xct_off) {
                Elf64_Phdr *phdro = (Elf64_Phdr *)(&eho->phdr);
                set_te32(&phdro->p_flags, Elf64_Phdr::PF_X | get_te32(&phdro->p_flags));
            }
        }
        if (!sec_arm_attr && !saved_opt_android_shlib) {
            // Make it abundantly clear that there are no Elf64_Shdr in this shlib
            eho->ehdr.e_shoff = 0;
            set_te16(&eho->ehdr.e_shentsize, sizeof(Elf64_Shdr));  // Android bug: cannot use 0
            eho->ehdr.e_shnum = 0;
            eho->ehdr.e_shstrndx = 0;
        }

        fo->rewrite(eho, sizeof(ehdri) + e_phnum * sizeof(*phdri));
        fo->seek(linfo_off, SEEK_SET);
        fo->rewrite(&linfo, sizeof(linfo));  // new info: l_checksum, l_size

        if (jni_onload_va) { // FIXME Does this apply to 64-bit, too?
            upx_uint64_t tmp = sz_pack2 + get_te64(&eho->phdr[C_TEXT].p_vaddr);
            tmp |= (Elf64_Ehdr::EM_ARM==e_machine);  // THUMB mode; no-op for 64-bit
            set_te64(&tmp, tmp);
            fo->seek(ptr_udiff_bytes(&jni_onload_sym->st_value, file_image), SEEK_SET);
            fo->rewrite(&tmp, sizeof(tmp));
        }
    }
    else { // not shlib
        // Cannot pre-round .p_memsz.  If .p_filesz < .p_memsz, then kernel
        // tries to make .bss, which requires PF_W.
        // But strict SELinux (or PaX, grSecurity) disallows PF_W with PF_X.
        set_te64(&eho->phdr[C_TEXT].p_filesz, sz_pack2 + lsize);
                  eho->phdr[C_TEXT].p_memsz = eho->phdr[C_TEXT].p_filesz;

        fo->seek(0, SEEK_SET);
        fo->rewrite(&eho->ehdr, sizeof(Elf64_Ehdr) + 2* sizeof(Elf64_Phdr));  // C_BASE, C_TEXT
        fo->seek(overlay_offset - sizeof(l_info), SEEK_SET);
        fo->rewrite(&linfo, sizeof(linfo));
    }
}

void
PackLinuxElf32::unRel32(
    unsigned dt_rel,
    Elf32_Rel *rel0,
    unsigned relsz,
    MemBuffer &ptload1,
    unsigned const load_off,
    OutputFile *fo
)
{
    Elf32_Rel *rel = rel0;
    for (int k = relsz / sizeof(Elf32_Rel); --k >= 0; ++rel) {
        unsigned r_offset = get_te32(&rel->r_offset);
        unsigned r_info   = get_te32(&rel->r_info);
        unsigned r_type = ELF32_R_TYPE(r_info);
        if (xct_off <= r_offset) {
            set_te32(&rel->r_offset, r_offset - asl_delta);
        }
        if (Elf32_Ehdr::EM_ARM == e_machine) {
            if (R_ARM_RELATIVE == r_type) {
                unsigned d = r_offset - load_off - asl_delta;
                unsigned w = get_te32(&ptload1[d]);
                if (xct_off <= w) {
                    set_te32(&ptload1[d], w - asl_delta);
                }
            }
            if (R_ARM_JUMP_SLOT == r_type) {
                ++n_jmp_slot;
                // .rel.plt contains offset of the "first time" target
                unsigned d = r_offset - load_off - asl_delta;
                if (plt_va > d) {
                    plt_va = d;
                }
                unsigned w = get_te32(&ptload1[d]);
                if (xct_off <= w) {
                    set_te32(&ptload1[d], w - asl_delta);
                }
            }
        }
        if (Elf32_Ehdr::EM_386 == e_machine) {
            if (R_386_RELATIVE == r_type) {
                unsigned d = r_offset - load_off - asl_delta;
                unsigned w = get_te32(&ptload1[d]);
                if (xct_off <= w) {
                    set_te32(&ptload1[d], w - asl_delta);
                }
            }
            if (R_386_JMP_SLOT == r_type) {
                ++n_jmp_slot;
                // .rel.plt contains offset of the "first time" target
                unsigned d = r_offset - load_off - asl_delta;
                if (plt_va > d) {
                    plt_va = d;
                }
                unsigned w = get_te32(&ptload1[d]);
                if (xct_off <= w) {
                    set_te32(&ptload1[d], w - asl_delta);
                }
            }
        }
    }
    fo->seek(dt_rel, SEEK_SET);
    fo->rewrite(rel0, relsz);
}

void
PackLinuxElf64::unRela64(
    upx_uint64_t const dt_rela,
    Elf64_Rela *const rela0,
    unsigned const relasz,
    upx_uint64_t const old_dtinit,
    OutputFile *const fo
)
{
    Elf64_Rela *rela = rela0;
    for (int k = relasz / sizeof(Elf64_Rela); --k >= 0; ++rela) {
        upx_uint64_t r_addend = get_te64(&rela->r_addend);
        if (xct_off <= r_addend) {
            r_addend -= asl_delta;
            set_te64(&rela->r_addend, r_addend);
        }

        upx_uint64_t r_offset = get_te64(&rela->r_offset);
        if (xct_off <= r_offset) {
            //r_offset -= asl_delta;  // keep compressed value vs plt_va
            set_te64(&rela->r_offset, r_offset - asl_delta);  // uncompressed value
        }

        // ElfXX_Rela (used only on 64-bit) ignores the contents of memory
        // at the target designated by r_offset.  The target is completely
        // overwritten by (r_addend + f_reloc(r_info)).
        //
        // Nevertheless, the existing targets of .rela.plt in the .got
        // seem to have values that matter to somebody. So restore original
        // values when is_asl.
        upx_uint64_t r_info   = get_te64(&rela->r_info);
        unsigned r_type = ELF64_R_TYPE(r_info);
        if (is_asl && Elf64_Ehdr::EM_AARCH64 == e_machine) {
            if (R_AARCH64_RELATIVE == r_type) {
#if 0  //{ FIXME
                if (old_dtinit == r_addend) {
                    set_te64(&ptload1[r_offset - plt_va], r_addend);
                }
#endif  //}
            }
            if (R_AARCH64_JUMP_SLOT == r_type) {
                ++n_jmp_slot;
                // .rela.plt contains offset of the "first time" target
                if (jump_slots.getSize() < (r_offset - plt_va)) {
                    throwInternalError("bad r_offset for jump_slots");
                }
                // really upx_uint64_t *, but clang makes it hard to say that
                unsigned char *slot = r_offset - plt_va
                    + (unsigned char *)jump_slots.getVoidPtr();
                upx_uint64_t w = get_te64(slot);
                if (xct_off <= w) {
                    set_te64(slot, w - asl_delta);
                }
            }
        }
        // FIXME: but what about old_dtinit?
        (void)old_dtinit;

    }  // end each RELA
    if (fo) {
        fo->seek(dt_rela, SEEK_SET);
        fo->rewrite(rela0, relasz);
    }
}

void
PackLinuxElf64::un_asl_dynsym( // ibuf has the input
    unsigned orig_file_size,
    OutputFile *fo  // else just leave in ibuf
)
{
    // un-Relocate dynsym (DT_SYMTAB) which is below xct_off
    dynstr = (char const *)elf_find_dynamic(Elf64_Dyn::DT_STRTAB);
    sec_dynsym = elf_find_section_type(Elf64_Shdr::SHT_DYNSYM);
    if (dynstr && sec_dynsym) {
        upx_uint64_t const off_dynsym = get_te64(&sec_dynsym->sh_offset);
        upx_uint64_t const sz_dynsym  = get_te64(&sec_dynsym->sh_size);
        if (orig_file_size < sz_dynsym
        ||  orig_file_size < off_dynsym
        || (orig_file_size - off_dynsym) < sz_dynsym) {
            throwCantUnpack("bad SHT_DYNSYM");
        }
        Elf64_Sym *const sym0 = (Elf64_Sym *)ibuf.subref(
            "bad dynsym", off_dynsym, sz_dynsym);
        Elf64_Sym *sym = sym0;
        for (int j = sz_dynsym / sizeof(Elf64_Sym); --j>=0; ++sym) {
            upx_uint64_t symval = get_te64(&sym->st_value);
            unsigned symsec = get_te16(&sym->st_shndx);
            if (Elf64_Sym::SHN_UNDEF != symsec
            &&  Elf64_Sym::SHN_ABS   != symsec
            &&  xct_off <= symval) {
                set_te64(&sym->st_value, symval - asl_delta);
            }
            if (Elf64_Sym::SHN_ABS == symsec && xct_off <= symval) {
                adjABS(sym, 0ul - (unsigned long)asl_delta);
            }
        }
        if (fo) {
            unsigned pos = fo->tell();
            fo->seek(off_dynsym, SEEK_SET);
            fo->rewrite(sym0, sz_dynsym);
            fo->seek(pos, SEEK_SET);
        }
    }
}

void
PackLinuxElf32::un_asl_dynsym( // ibuf has the input
    unsigned orig_file_size,
    OutputFile *fo  // else just leave in ibuf
)
{
    // un-Relocate dynsym (DT_SYMTAB) which is below xct_off
    dynstr = (char const *)elf_find_dynamic(Elf32_Dyn::DT_STRTAB);
    sec_dynsym = elf_find_section_type(Elf32_Shdr::SHT_DYNSYM);
    if (dynstr && sec_dynsym) {
        upx_uint32_t const off_dynsym = get_te32(&sec_dynsym->sh_offset);
        upx_uint32_t const sz_dynsym  = get_te32(&sec_dynsym->sh_size);
        if (orig_file_size < sz_dynsym
        ||  orig_file_size < off_dynsym
        || (orig_file_size - off_dynsym) < sz_dynsym) {
            throwCantUnpack("bad SHT_DYNSYM");
        }
        Elf32_Sym *const sym0 = (Elf32_Sym *)ibuf.subref(
            "bad dynsym", off_dynsym, sz_dynsym);
        Elf32_Sym *sym = sym0;
        for (int j = sz_dynsym / sizeof(Elf32_Sym); --j>=0; ++sym) {
            upx_uint32_t symval = get_te32(&sym->st_value);
            unsigned symsec = get_te16(&sym->st_shndx);
            if (Elf32_Sym::SHN_UNDEF != symsec
            &&  Elf32_Sym::SHN_ABS   != symsec
            &&  xct_off <= symval) {
                set_te32(&sym->st_value, symval - asl_delta);
            }
            if (Elf32_Sym::SHN_ABS == symsec && xct_off <= symval) {
                adjABS(sym, 0u - (unsigned)asl_delta);
            }
        }
        if (fo) {
            unsigned pos = fo->tell();
            fo->seek(off_dynsym, SEEK_SET);
            fo->rewrite(sym0, sz_dynsym);
            fo->seek(pos, SEEK_SET);
        }
    }
}

// File layout of compressed .so (new-style: 3 or 4 PT_LOAD) shared library:
// 1. new Elf headers: Ehdr, PT_LOAD (r-x), PT_LOAD (rw-, if any), non-PT_LOAD Phdrs
// 2. Space for (original - 2) PT_LOAD Phdr
// 3. Remaining original contents of file below xct_off
// xct_off: (&lowest eXecutable Shdr section; in original PT_LOAD[0] or [1])
// 3a. If --android-shlib, then 4KiB page of Shdr copy, etc.  (asl_pack2_Shdrs)
//    And xct_off gets incremented by 4KiB at the right time.
// 4. l_info (12 bytes)
// overlay_offset:
// 5. p_info (12 bytes)
// 6. compressed original Elf headers (prefixed by b_info as usual)
// 6a. un-compressed copy of input after Elf headers until xct_off.
//    *user_init_rp has been modified if no DT_INIT
// 7. compressed remainder of PT_LOAD above xct_off
// 8. compressed read-only PT_LOAD above xct_off (if any)  // FIXME: check decompressor
// 9. uncompressed Read-Write PT_LOAD (slide down N pages)
// 10. int[6] tables for UPX runtime de-compressor
// (new) DT_INIT:
// 11. UPX runtime de-compressing loader
// 12. compressed gaps between PT_LOADs (and EOF) above xct_off
// 13. 32-byte pack header
// 14. 4-byte overlay_offset

void PackLinuxElf64::un_shlib_1(
    OutputFile *const fo,
    MemBuffer &o_elfhdrs,
    unsigned &c_adler,
    unsigned &u_adler,
    unsigned const orig_file_size
)
{
    // xct_off [input side] was set by ::unpack when is_shlib
    // yct_off [output side] set here unless is_asl in next 'if' block
    unsigned yct_off = xct_off;

    // Below xct_off is not compressed (for benefit of rtld.)
    fi->seek(0, SEEK_SET);
    fi->readx(ibuf, umin(blocksize, file_size_u32));

    // Determine if the extra page with copy of _Shdrs was spliced in.
    // This used to be the result of --android-shlib.
    // But in 2023-02 the forwarding of ARM_ATTRIBUTES (by appending)
    // takes care of this, so the 5th word before e_entry does not
    // have the low bit 1, so is_asl should not be set.
    // However, .so that were compressed before 2023-03
    // may be marked.
    e_shoff = get_te64(&ehdri.e_shoff);
    if (e_shoff && e_shnum
            // +36: (sizeof(PackHeader) + sizeof(overlay_offset))
            //    after Shdrs for ARM_ATTRIBUTES
    &&  (((e_shoff + sizeof(Elf64_Shdr) * e_shnum) + 36) < file_size_u)
    ) { // possible --android-shlib
        unsigned x = get_te32(&file_image[get_te64(&ehdri.e_entry) - (1+ 4)*sizeof(int)]);
        if (1 & x) { // the clincher
            is_asl = 1;
            fi->seek(e_shoff, SEEK_SET);
            mb_shdr.alloc(   sizeof(Elf64_Shdr) * e_shnum);
            shdri = (Elf64_Shdr *)mb_shdr.getVoidPtr();
            fi->readx(shdri, sizeof(Elf64_Shdr) * e_shnum);
            yct_off = get_te64(&shdri->sh_offset);  // for the output file (de-compressed)
            xct_off = asl_delta + yct_off;  // for the input file (compressed)
        }
    }

    // Decompress first Extent.  Old style covers [0, xct_off)
    // which includes rtld constant data and eXecutable app code below DT_INIT.
    // In old style, the first compressed Extent is redundant
    // except for the compressed original Elf headers.
    // New style covers just Elf headers: the rest below xct_off is
    // rtld constant data: DT_*HASH, DT_SYMTAB, DT_STRTAB, etc.
    // New style puts eXecutable app code in second PT_LOAD
    // in order to mark Elf headers and rtld data as non-eXecutable.
    fi->seek(overlay_offset - sizeof(l_info), SEEK_SET);
    struct {
        struct l_info l;
        struct p_info p;
        struct b_info b;
    } hdr;
    fi->readx(&hdr, sizeof(hdr));
    if (hdr.l.l_magic != UPX_MAGIC_LE32
    ||  get_te16(&hdr.l.l_lsize) != (unsigned)lsize
    ||  get_te32(&hdr.p.p_filesize) != ph.u_file_size
    ||  get_te32(&hdr.b.sz_unc) < sz_elf_hdrs  // peek: 1st b_info covers Elf headers
    ) {
        throwCantUnpack("corrupt l_info/p_info/b_info");
    }
    fi->seek(-(off_t)sizeof(struct b_info), SEEK_CUR); // hdr.b_info was a peek

// The default layout for a shared library created by binutils-2.29
// (Fedora 28; 2018) has two PT_LOAD: permissions r-x and rw-.
// xct_off (the lowest address of executable instructions;
// the highest address of read-only data used by rtld (ld-linux))
// will be somewhere in the first PT_LOAD.
//
// The default layout for a shared library created by binutils-2.31
// (Fedora 29; 2018) has four PT_LOAD: permissions r--, r-x, r--, rw-.
// xct_off will be the base of the second [r-x] PT_LOAD.
//
// Bytes below xct_off cannot be compressed because they are used
// by rtld *before* the UPX run-time de-compression stub gets control
// via DT_INIT. Bytes in a Writeable PT_LOAD cannot be compressed
// because they may be relocated by rtld, again before stub execution.
//
// We need to know which layout of PT_LOAD. It seems risky to steal
// bits in the input ElfXX_Ehdr or ElfXX_Phdr, so we decompress
// the first compressed block.  For an old-style shared library
// the first compressed block covers [0, xct_off) which is redundant
// with the interval [sz_elf_hdrs, xct_off) because those bytes
// must be present for use by rtl  (So that is a large inefficiency.)
// Fortunately p_info.p_blocksize fits in ibuf, and unpackExtent
// will just decompress it all.  For new style, the first compressed
// block covers [0, sz_elf_hdrs).

    // Peek: unpack into ibuf, but do not write
    unsigned const sz_block1 = unpackExtent(sz_elf_hdrs, nullptr,
        c_adler, u_adler, false, -1);
    if (sz_block1 < sz_elf_hdrs) {
        throwCantUnpack("corrupt b_info");
    }
    memcpy(o_elfhdrs, ibuf, sz_elf_hdrs); // save de-compressed Elf headers
    Elf64_Ehdr const *const ehdro = (Elf64_Ehdr const *)(void const *)o_elfhdrs;
    if (ehdro->e_type   !=ehdri.e_type
    ||  ehdro->e_machine!=ehdri.e_machine
    ||  ehdro->e_version!=ehdri.e_version
        // less strict for EM_PPC64 to workaround earlier bug
    ||  !( ehdro->e_flags==ehdri.e_flags
        || Elf64_Ehdr::EM_PPC64 == get_te16(&ehdri.e_machine))
    ||  ehdro->e_ehsize !=ehdri.e_ehsize
        // check EI_MAG[0-3], EI_CLASS, EI_DATA, EI_VERSION
    ||  memcmp(ehdro->e_ident, ehdri.e_ident, Elf64_Ehdr::EI_OSABI)) {
        throwCantUnpack("ElfXX_Ehdr corrupted");
    }
    if (fo) {
        fo->write(ibuf, sz_block1);
        total_out = sz_block1;
    }
    Elf64_Phdr const *o_phdr = (Elf64_Phdr const *)(1+ ehdro);
    // Handle compressed PT_LOADs (must not have PF_W)
    unsigned not_first_LOAD = 0;
    for (unsigned j = 0; j < e_phnum; ++j, ++o_phdr) {
        unsigned type = get_te32(&o_phdr->p_type);
        unsigned flags = get_te32(&o_phdr->p_flags);
        if (PT_LOAD != type || Elf64_Phdr::PF_W & flags) {
            continue;
        }
        unsigned p_offset = get_te64(&o_phdr->p_offset);
        unsigned p_filesz = get_te64(&o_phdr->p_filesz);
        unsigned wanted = p_filesz;
        if (!not_first_LOAD++) { // first PT_LOAD
            wanted -= sz_block1;
            if (sz_block1 >  sz_elf_hdrs) { // old style
                if (is_asl) {
                    un_asl_dynsym(orig_file_size, fo);
                }
                p_offset += sz_block1;
            }
            if (sz_block1 == sz_elf_hdrs) { // new style
                unsigned const len = (yct_off ? yct_off : xct_off) - sz_elf_hdrs;
                unsigned const ipos = fi->tell();
                fi->seek(sz_elf_hdrs, SEEK_SET);
                fi->readx(&ibuf[sz_elf_hdrs], len);
                if (is_asl) {
                    un_asl_dynsym(orig_file_size, nullptr);
                }
                if (fo) {
                    fo->write(&ibuf[sz_elf_hdrs], len);
                }
                total_out += len;

// github-issue629: (overlay_offset = 0xa500), so initially (xct_off = 0xa494).
// But "yct_off = get_te64(&shdri->sh_offset)" so if _Shdrs are aligned (??)
// then (0x10500 == (xct_off = asl_delta + yct_off)), and we read+write
// more than we need.
// So assume the excess just lives there, or is overwritten later by seek+write.
                if (wanted < len) { // FIXME: why does this happen?
                    wanted = 0;
                }
                else {
                    wanted -= len;
                }
                fi->seek(ipos, SEEK_SET);
                if (total_out == p_filesz) {
                    continue;   // already entirely re-generated
                }
                p_offset = total_out;
            }
        }
        if (fo) {
            fo->seek(p_offset, SEEK_SET);
        }
        unpackExtent(wanted, fo, c_adler, u_adler, false);
    }
    funpad4(fi);
    loader_offset = fi->tell();

    // Handle PT_LOAD with PF_W: writeable, so not compressed.  "Slide"
    o_phdr = (Elf64_Phdr const *)(1+ ehdro);
    Elf64_Phdr const *i_phdr = phdri;
    for (unsigned j = 0; j < e_phnum; ++j, ++o_phdr, ++i_phdr) {
        unsigned type = get_te32(&o_phdr->p_type);
        unsigned flags = get_te32(&o_phdr->p_flags);
        if (PT_LOAD != type || !(Elf64_Phdr::PF_W & flags)) {
            continue;
        }
        unsigned filesz = get_te64(&o_phdr->p_filesz);
        unsigned o_offset = get_te64(&o_phdr->p_offset);
        unsigned i_offset = get_te64(&i_phdr->p_offset);
        fi->seek(i_offset, SEEK_SET);
        fi->readx(ibuf, filesz);
        total_in += filesz;
        if (fo) {
            fo->seek(o_offset, SEEK_SET);
            fo->write(ibuf, filesz);
        }
        total_out = filesz + o_offset;  // high-water mark
    }

    // Gaps between PT_LOAD will be handled by ::unpack()

    // position fi at loader offset
    fi->seek(loader_offset, SEEK_SET);
}

void PackLinuxElf32::un_shlib_1(
    OutputFile *const fo,
    MemBuffer &o_elfhdrs,
    unsigned &c_adler,
    unsigned &u_adler,
    unsigned const orig_file_size
)
{
    // xct_off [input side] was set by ::unpack when is_shlib
    // yct_off [output side] set here unless is_asl in next 'if' block
    unsigned yct_off = xct_off;

    // Below xct_off is not compressed (for benefit of rtld.)
    fi->seek(0, SEEK_SET);
    fi->readx(ibuf, umin(blocksize, file_size_u32));

    // Determine if the extra page with copy of _Shdrs was spliced in.
    // This used to be the result of --android-shlib.
    // But in 2023-02 the forwarding of ARM_ATTRIBUTES (by appending)
    // takes care of this, so the 5th word before e_entry does not
    // have the low bit 1, so is_asl should not be set.
    // However, .so that were compressed before 2023-03
    // may be marked.
    e_shoff = get_te32(&ehdri.e_shoff);
    if (e_shoff && e_shnum
            // +36: (sizeof(PackHeader) + sizeof(overlay_offset))
            //    after Shdrs for ARM_ATTRIBUTES
    &&  (((e_shoff + sizeof(Elf32_Shdr) * e_shnum) + 36) < file_size_u32)
    ) { // possible --android-shlib
        unsigned x = get_te32(&file_image[get_te32(&ehdri.e_entry) - (1+ 4)*sizeof(int)]);
        if (1 & x) { // the clincher
            is_asl = 1;
            fi->seek(e_shoff, SEEK_SET);
            mb_shdr.alloc(   sizeof(Elf32_Shdr) * e_shnum);
            shdri = (Elf32_Shdr *)mb_shdr.getVoidPtr();
            fi->readx(shdri, sizeof(Elf32_Shdr) * e_shnum);
            yct_off = get_te32(&shdri->sh_offset);  // for the output file (de-compressed)
            xct_off = asl_delta + yct_off;  // for the input file (compressed)
        }
    }

    // Decompress first Extent.  Old style covers [0, xct_off)
    // which includes rtld constant data and eXecutable app code below DT_INIT.
    // In old style, the first compressed Extent is redundant
    // except for the compressed original Elf headers.
    // New style covers just Elf headers: the rest below xct_off is
    // rtld constant data: DT_*HASH, DT_SYMTAB, DT_STRTAB, etc.
    // New style puts eXecutable app code in second PT_LOAD
    // in order to mark Elf headers and rtld data as non-eXecutable.
    fi->seek(overlay_offset - sizeof(l_info), SEEK_SET);
    struct {
        struct l_info l;
        struct p_info p;
        struct b_info b;
    } hdr;
    fi->readx(&hdr, sizeof(hdr));
    if (hdr.l.l_magic != UPX_MAGIC_LE32
    ||  get_te16(&hdr.l.l_lsize) != (unsigned)lsize
    ||  get_te32(&hdr.p.p_filesize) != ph.u_file_size
    ||  get_te32(&hdr.b.sz_unc) < sz_elf_hdrs  // peek: 1st b_info covers Elf headers
    ) {
        throwCantUnpack("corrupt l_info/p_info/b_info");
    }
    fi->seek(-(off_t)sizeof(struct b_info), SEEK_CUR); // hdr.b_info was a peek

// The default layout for a shared library created by binutils-2.29
// (Fedora 28; 2018) has two PT_LOAD: permissions r-x and rw-.
// xct_off (the lowest address of executable instructions;
// the highest address of read-only data used by rtld (ld-linux))
// will be somewhere in the first PT_LOAD.
//
// The default layout for a shared library created by binutils-2.31
// (Fedora 29; 2018) has four PT_LOAD: permissions r--, r-x, r--, rw-.
// xct_off will be the base of the second [r-x] PT_LOAD.
//
// Bytes below xct_off cannot be compressed because they are used
// by rtld *before* the UPX run-time de-compression stub gets control
// via DT_INIT. Bytes in a Writeable PT_LOAD cannot be compressed
// because they may be relocated by rtld, again before stub execution.
//
// We need to know which layout of PT_LOAD. It seems risky to steal
// bits in the input ElfXX_Ehdr or ElfXX_Phdr, so we decompress
// the first compressed block.  For an old-style shared library
// the first compressed block covers [0, xct_off) which is redundant
// with the interval [sz_elf_hdrs, xct_off) because those bytes
// must be present for use by rtl  (So that is a large inefficiency.)
// Fortunately p_info.p_blocksize fits in ibuf, and unpackExtent
// will just decompress it all.  For new style, the first compressed
// block covers [0, sz_elf_hdrs).

    // Peek: unpack into ibuf, but do not write
    unsigned const sz_block1 = unpackExtent(sz_elf_hdrs, nullptr,
        c_adler, u_adler, false, -1);
    if (sz_block1 < sz_elf_hdrs) {
        throwCantUnpack("corrupt b_info");
    }
    memcpy(o_elfhdrs, ibuf, sz_elf_hdrs); // save de-compressed Elf headers
    Elf32_Ehdr const *const ehdro = (Elf32_Ehdr const *)(void const *)o_elfhdrs;
    if (ehdro->e_type   !=ehdri.e_type
    ||  ehdro->e_machine!=ehdri.e_machine
    ||  ehdro->e_version!=ehdri.e_version
        // less strict for EM_PPC to workaround earlier bug
    ||  !( ehdro->e_flags==ehdri.e_flags
        || Elf32_Ehdr::EM_PPC == get_te16(&ehdri.e_machine))
    ||  ehdro->e_ehsize !=ehdri.e_ehsize
        // check EI_MAG[0-3], EI_CLASS, EI_DATA, EI_VERSION
    ||  memcmp(ehdro->e_ident, ehdri.e_ident, Elf32_Ehdr::EI_OSABI)) {
        throwCantUnpack("ElfXX_Ehdr corrupted");
    }
    if (fo) {
        fo->write(ibuf, sz_block1);
        total_out = sz_block1;
    }
    Elf32_Phdr const *o_phdr = (Elf32_Phdr const *)(1+ ehdro);
    // Handle compressed PT_LOADs (must not have PF_W)
    unsigned not_first_LOAD = 0;
    for (unsigned j = 0; j < e_phnum; ++j, ++o_phdr) {
        unsigned type = get_te32(&o_phdr->p_type);
        unsigned flags = get_te32(&o_phdr->p_flags);
        if (PT_LOAD != type || Elf32_Phdr::PF_W & flags) {
            continue;
        }
        unsigned p_offset = get_te32(&o_phdr->p_offset);
        unsigned p_filesz = get_te32(&o_phdr->p_filesz);
        unsigned wanted = p_filesz;
        if (!not_first_LOAD++) { // first PT_LOAD
            wanted -= sz_block1;
            if (sz_block1 >  sz_elf_hdrs) { // old style
                if (is_asl) {
                    un_asl_dynsym(orig_file_size, fo);
                }
                p_offset += sz_block1;
            }
            if (sz_block1 == sz_elf_hdrs) { // new style
                unsigned const len = (yct_off ? yct_off : xct_off) - sz_elf_hdrs;
                unsigned const ipos = fi->tell();
                fi->seek(sz_elf_hdrs, SEEK_SET);
                fi->readx(&ibuf[sz_elf_hdrs], len);
                if (is_asl) {
                    un_asl_dynsym(orig_file_size, nullptr);
                }
                if (fo) {
                    fo->write(&ibuf[sz_elf_hdrs], len);
                }
                total_out += len;

// github-issue629: (overlay_offset = 0xa500), so initially (xct_off = 0xa494).
// But "yct_off = get_te32(&shdri->sh_offset)" so if _Shdrs are aligned (??)
// then (0x10500 == (xct_off = asl_delta + yct_off)), and we read+write
// more than we need.
// So assume the excess just lives there, or is overwritten later by seek+write.
                if (wanted < len) { // FIXME: why does this happen?
                    wanted = 0;
                }
                else {
                    wanted -= len;
                }
                fi->seek(ipos, SEEK_SET);
                if (total_out == p_filesz) {
                    continue;   // already entirely re-generated
                }
                p_offset = total_out;
            }
        }
        if (fo) {
            fo->seek(p_offset, SEEK_SET);
        }
        unpackExtent(wanted, fo, c_adler, u_adler, false);
    }
    funpad4(fi);
    loader_offset = fi->tell();

    // Handle PT_LOAD with PF_W: writeable, so not compressed.  "Slide"
    o_phdr = (Elf32_Phdr const *)(1+ ehdro);
    Elf32_Phdr const *i_phdr = phdri;
    for (unsigned j = 0; j < e_phnum; ++j, ++o_phdr, ++i_phdr) {
        unsigned type = get_te32(&o_phdr->p_type);
        unsigned flags = get_te32(&o_phdr->p_flags);
        if (PT_LOAD != type || !(Elf32_Phdr::PF_W & flags)) {
            continue;
        }
        unsigned filesz = get_te32(&o_phdr->p_filesz);
        unsigned o_offset = get_te32(&o_phdr->p_offset);
        unsigned i_offset = get_te32(&i_phdr->p_offset);
        fi->seek(i_offset, SEEK_SET);
        fi->readx(ibuf, filesz);
        total_in += filesz;
        if (fo) {
            fo->seek(o_offset, SEEK_SET);
            fo->write(ibuf, filesz);
        }
        total_out = filesz + o_offset;  // high-water mark
    }

    // Gaps between PT_LOAD will be handled by ::unpack()

    // position fi at loader offset
    fi->seek(loader_offset, SEEK_SET);
}

void PackLinuxElf32::un_DT_INIT(
    unsigned old_dtinit,
    Elf32_Phdr const *const phdro,
    Elf32_Phdr const *const dynhdr,  // in phdri
    OutputFile *fo
)
{
    // DT_INIT must be restored.
    // If android_shlib, then the asl_delta relocations must be un-done.
    unsigned n_plt = 0;
    upx_uint32_t dt_pltrelsz(0), dt_jmprel(0), dt_pltgot(0);
    upx_uint32_t dt_relsz(0), dt_rel(0);
    upx_uint32_t const dyn_len = get_te32(&dynhdr->p_filesz);
    upx_uint32_t const dyn_off = get_te32(&dynhdr->p_offset);
    if (file_size_u32 < (dyn_len + dyn_off)) {
        char msg[50]; snprintf(msg, sizeof(msg),
                "bad PT_DYNAMIC .p_filesz %#lx", (long unsigned)dyn_len);
        throwCantUnpack(msg);
    }
    fi->seek(dyn_off, SEEK_SET);
    fi->readx(ibuf, dyn_len);
    Elf32_Dyn *dyn = (Elf32_Dyn *)(void *)ibuf;
    dynseg = dyn; invert_pt_dynamic(dynseg,
        umin(dyn_len, file_size_u32 - dyn_off));
    for (unsigned j2= 0; j2 < dyn_len; ++dyn, j2 += sizeof(*dyn)) {
        upx_uint32_t const tag = get_te32(&dyn->d_tag);
        upx_uint32_t       val = get_te32(&dyn->d_val);
        if (is_asl) switch (tag) {
        case Elf32_Dyn::DT_RELASZ:   { dt_relsz   = val; } break;
        case Elf32_Dyn::DT_RELA:     { dt_rel     = val; } break;
        case Elf32_Dyn::DT_JMPREL:   { dt_jmprel   = val; } break;
        case Elf32_Dyn::DT_PLTRELSZ: { dt_pltrelsz = val;
            n_plt = dt_pltrelsz / sizeof(Elf32_Rel);
            if (is_asl) {
                n_plt += 3;  // FIXME
            }
        };  break;

        case Elf32_Dyn::DT_PLTGOT:   { plt_va = dt_pltgot = val; (void)dt_pltgot; }
        // FALL THROUGH
        case Elf32_Dyn::DT_PREINIT_ARRAY:
        case Elf32_Dyn::DT_INIT_ARRAY:
        case Elf32_Dyn::DT_FINI_ARRAY:
        case Elf32_Dyn::DT_FINI: if (is_asl) {
            set_te32(&dyn->d_val, val - asl_delta);
        }; break;
        } // end switch() on tag when is_asl
        if (upx_dt_init == tag) {
            if (Elf32_Dyn::DT_INIT == tag) { // the easy case
                set_te32(&dyn->d_val, old_dtinit);
                if (!old_dtinit) { // compressor took the slot
                    dyn->d_tag = Elf32_Dyn::DT_NULL;
                    dyn->d_val = 0;
                }
            }
            // Apparently the hard case is common for some Android IDEs.
            else if (Elf32_Dyn::DT_INIT_ARRAY    == tag
            ||       Elf32_Dyn::DT_PREINIT_ARRAY == tag) {
                // 'val' is the RVA of the first slot, which is the slot that
                // the compressor changed to be the entry to the run-time stub.
                Elf32_Dyn *dyn_null = elf_find_dynptr(Elf32_Dyn::DT_NULL);
                if (!dyn_null)
                    throwCantUnpack("bad PT_DYNAMIC .end");
                Elf32_Rel *rp = (Elf32_Rel *)elf_find_dynamic(dyn_null->d_val);
                dyn_null->d_val = 0;
                if (rp) {
                    // Compressor saved the original *rp in dynsym[0]
                    Elf32_Rel *rp_unc = (Elf32_Rel *)&dynsym[0];  // pointer
                    rp->r_info = rp_unc->r_info;  // restore original r_info; r_offset not touched

                    unsigned e_entry = get_te32(&ehdri.e_entry);
                    unsigned init_rva = get_te32(&file_image[e_entry - 3*sizeof(unsigned)]);
                    unsigned arr_rva = get_te32(&rp_unc->r_offset);
                    Elf32_Phdr const *phdr = elf_find_Phdr_for_va(arr_rva, phdro, e_phnum);
                    unsigned arr_off = (arr_rva - get_te32(&phdr->p_vaddr)) + get_te32(&phdr->p_offset);

                    rp_unc->r_offset = 0;    rp_unc->r_info = 0;
                    if (fo)  {
                        fo->seek(elf_unsigned_dynamic(Elf32_Dyn::DT_SYMTAB), SEEK_SET);
                        fo->rewrite(rp_unc, sizeof(Elf32_Rel));  // clear dynsym[0]

                        fo->seek((char *)rp - (char *)&file_image[0], SEEK_SET);
                        fo->rewrite(rp, sizeof(*rp));  // restore original *rp
                    }

                    // Set arr[0] to the first user init routine.
                    unsigned r_info = get_te32(&rp->r_info);
                    unsigned r_type = ELF32_R_TYPE(r_info);
                    unsigned word;
                    if (Elf32_Ehdr::EM_ARM == e_machine) {
                        if (R_ARM_RELATIVE == r_type) {
                            set_te32(&word, init_rva);
                        }
                        else if (R_ARM_ABS32 == r_type) {
                            word = 0;
                        }
                        else {
                            char msg[40]; snprintf(msg, sizeof(msg), "unknown relocation: %#x",
                                r_type);
                            throwCantUnpack(msg);
                        }
                    }
                    else if (Elf32_Ehdr::EM_386 == e_machine) {
                        if (R_386_RELATIVE == r_type) {
                        }
                        else if (R_386_32 == r_type) {
                        }
                        if (R_386_RELATIVE == r_type) {
                            set_te32(&word, init_rva);
                        }
                        else if (R_386_32 == r_type) {
                            word = 0;
                        }
                        else {
                            char msg[40]; snprintf(msg, sizeof(msg), "unknown relocation: %#x",
                                r_type);
                            throwCantUnpack(msg);
                        }
                    }
                    if (fo) {
                        fo->seek(arr_off, SEEK_SET);
                        fo->rewrite(&word, sizeof(unsigned));
                        fo->seek(0, SEEK_END);
                    }
                }
            }
        }
    }
    if (fo) { // Write updated dt_*.val
        upx_uint32_t dyn_offo = get_te32(&phdro[dynhdr - phdri].p_offset);
        fo->seek(dyn_offo, SEEK_SET);
        fo->rewrite(ibuf, dyn_len);
    }
    if (is_asl) {
        MemBuffer ptload1;  // FIXME.  file_image has the whole file; ibuf is available
        lowmem.alloc(xct_off);
        fi->seek(0, SEEK_SET);
        fi->read(lowmem, xct_off);  // contains relocation tables
        if (dt_relsz && dt_rel) {
            Elf32_Rel *const rel0 = (Elf32_Rel *)lowmem.subref(
                "bad Rel offset", dt_rel, dt_relsz);
            unRel32(dt_rel, rel0, dt_relsz, ptload1, old_dtinit, fo);
        }
        if (dt_pltrelsz && dt_jmprel) { // FIXME:  overlap w/ DT_REL ?
            Elf32_Rel *const jmp0 = (Elf32_Rel *)lowmem.subref(
                "bad Jmprel offset", dt_jmprel, dt_pltrelsz);
            jump_slots.alloc(n_plt * sizeof(upx_uint32_t));
            Elf32_Phdr const *phdr = phdri;
            for (unsigned j = 0; j < e_phnum; ++j, ++phdr) if (is_LOAD(phdr)) {
                upx_uint32_t vaddr = get_te32(&phdr->p_vaddr);
                upx_uint32_t filesz = get_te32(&phdr->p_filesz);
                upx_uint32_t d = plt_va - vaddr;
                if (d < filesz) {
                    upx_uint32_t offset = get_te32(&phdr->p_offset);
                    fi->seek(d + offset, SEEK_SET);
                    fi->readx(jump_slots, n_plt * sizeof(upx_uint32_t));
                    break;
                }
            }
            unRel32(dt_jmprel, jmp0, dt_pltrelsz, ptload1, old_dtinit, fo);

            Elf32_Ehdr const *const o_ehdr = (Elf32_Ehdr const *)(void *)lowmem;
            unsigned const o_phnum = o_ehdr->e_phnum;
            if (((1<<16) - sizeof(Elf32_Ehdr)) / sizeof(Elf32_Phdr) < o_phnum)
                throwCantUnpack("bad Ehdr.e_phnum %#x", o_phnum);
            phdr = phdro;
            for (unsigned j = 0; j < o_phnum; ++j, ++phdr) if (is_LOAD(phdr)) {
                upx_uint32_t vaddr = get_te32(&phdr->p_vaddr);
                upx_uint32_t filesz = get_te32(&phdr->p_filesz);
                upx_uint32_t d = plt_va - vaddr - asl_delta;
                if (d < filesz) {
                    upx_uint32_t offset = get_te32(&phdr->p_offset);
                    if ((upx_uint32_t)file_size <= offset)
                        throwCantUnpack("bad phdr[%d].p_offset %#zx", j, (size_t)offset);
                    if (fo) {
                        fo->seek(d + offset, SEEK_SET);
                        fo->rewrite(jump_slots, n_plt * sizeof(upx_uint32_t));
                    }
                    break;
                }
            }
        }
        // Modified relocation tables are re-written by unRel32
    }
}

void PackLinuxElf64::un_DT_INIT(
    unsigned old_dtinit,
    Elf64_Phdr const *const phdro,
    Elf64_Phdr const *const dynhdr,  // in phdri
    OutputFile *fo
)
{
    // DT_INIT must be restored.
    // If android_shlib, then the asl_delta relocations must be un-done.
    unsigned n_plt = 0;
    upx_uint64_t dt_pltrelsz(0), dt_jmprel(0), dt_pltgot(0);
    upx_uint64_t dt_relasz(0), dt_rela(0);
    upx_uint64_t const dyn_len = get_te64(&dynhdr->p_filesz);
    upx_uint64_t const dyn_off = get_te64(&dynhdr->p_offset);
    if (file_size_u < (dyn_len + dyn_off)) {
        char msg[50]; snprintf(msg, sizeof(msg),
                "bad PT_DYNAMIC .p_filesz %#lx", (long unsigned)dyn_len);
        throwCantUnpack(msg);
    }
    fi->seek(dyn_off, SEEK_SET);
    fi->readx(ibuf, dyn_len);
    Elf64_Dyn *dyn = (Elf64_Dyn *)(void *)ibuf;
    dynseg = dyn; invert_pt_dynamic(dynseg,
        umin(dyn_len, file_size_u - dyn_off));
    for (unsigned j2= 0; j2 < dyn_len; ++dyn, j2 += sizeof(*dyn)) {
        upx_uint64_t const tag = get_te64(&dyn->d_tag);
        upx_uint64_t       val = get_te64(&dyn->d_val);
        if (is_asl) switch (tag) {
        case Elf64_Dyn::DT_RELASZ:   { dt_relasz   = val; } break;
        case Elf64_Dyn::DT_RELA:     { dt_rela     = val; } break;
        case Elf64_Dyn::DT_JMPREL:   { dt_jmprel   = val; } break;
        case Elf64_Dyn::DT_PLTRELSZ: { dt_pltrelsz = val;
            n_plt = dt_pltrelsz / sizeof(Elf32_Rel);
            if (is_asl) {
                n_plt += 3;  // FIXME
            }
        };  break;

        case Elf64_Dyn::DT_PLTGOT:   { plt_va = dt_pltgot = val; (void)dt_pltgot;}
        // FALL THROUGH
        case Elf64_Dyn::DT_PREINIT_ARRAY:
        case Elf64_Dyn::DT_INIT_ARRAY:
        case Elf64_Dyn::DT_FINI_ARRAY:
        case Elf64_Dyn::DT_FINI: if (is_asl) {
            set_te64(&dyn->d_val, val - asl_delta);
        }; break;
        } // end switch() on tag when is_asl
        if (upx_dt_init == tag) { // the easy case
            if (Elf64_Dyn::DT_INIT == tag) {
                set_te64(&dyn->d_val, old_dtinit);
                if (!old_dtinit) { // compressor took the slot
                    dyn->d_tag = Elf64_Dyn::DT_NULL;
                    dyn->d_val = 0;
                }
            }
            // Apparently the hard case is common for some Android IDEs.
            // No DT_INIT; only DT_INIT_ARRAY.
            else if (Elf64_Dyn::DT_INIT_ARRAY    == tag
            ||       Elf64_Dyn::DT_PREINIT_ARRAY == tag) {
                // 'val' is the RVA of the first slot, which is the slot that
                // the compressor changed to be the entry to the run-time stub.
                Elf64_Dyn *dyn_null = elf_find_dynptr(Elf64_Dyn::DT_NULL);
                if (!dyn_null)
                    throwCantUnpack("bad PT_DYNAMIC .end");
                Elf64_Rela *rp = (Elf64_Rela *)elf_find_dynamic(dyn_null->d_val);
                dyn_null->d_val = 0;
                if (rp) {
                    // Compressor saved the original *rp in dynsym[0]
                    Elf64_Rela *rp_unc = (Elf64_Rela *)&dynsym[0];  // pointer
                    rp->r_info = rp_unc->r_info;  // restore original r_info; r_offset not touched
                    rp->r_addend = rp_unc->r_addend;

                    unsigned arr_rva = get_te64(&rp_unc->r_offset);
                    Elf64_Phdr const *phdr = elf_find_Phdr_for_va(arr_rva, phdro, e_phnum);
                    unsigned arr_off = (arr_rva - get_te64(&phdr->p_vaddr)) + get_te64(&phdr->p_offset);

                    if (fo)  {
                        memset(rp_unc, 0, sizeof(*rp_unc));
                        fo->seek(elf_unsigned_dynamic(Elf64_Dyn::DT_SYMTAB), SEEK_SET);
                        fo->rewrite(rp_unc, sizeof(Elf64_Rela));  // clear dynsym[0]

                        fo->seek((char *)rp - (char *)&file_image[0], SEEK_SET);
                        fo->rewrite(rp, sizeof(*rp));  // restore original *rp

                        // Elf64_Rela overwrites; but put back original.
                        fo->seek(arr_off, SEEK_SET);
                        fo->rewrite(rp_unc, sizeof(u64_t));

                        fo->seek(0, SEEK_END);
                    }
                }
            }
        }
    }
    if (fo) { // Write updated dt_*.val
        upx_uint64_t dyn_offo = get_te64(&phdro[dynhdr - phdri].p_offset);
        fo->seek(dyn_offo, SEEK_SET);
        fo->rewrite(ibuf, dyn_len);
    }
    if (is_asl) {
        lowmem.alloc(xct_off);
        fi->seek(0, SEEK_SET);
        fi->read(lowmem, xct_off);  // contains relocation tables
        if (dt_relasz && dt_rela) {
            Elf64_Rela *const rela0 = (Elf64_Rela *)lowmem.subref(
                "bad Rela offset", dt_rela, dt_relasz);
            unRela64(dt_rela, rela0, dt_relasz, old_dtinit, fo);
        }
        if (dt_pltrelsz && dt_jmprel) { // FIXME:  overlap w/ DT_REL ?
            Elf64_Rela *const jmp0 = (Elf64_Rela *)lowmem.subref(
                "bad Jmprel offset", dt_jmprel, dt_pltrelsz);
            jump_slots.alloc(n_plt * sizeof(upx_uint64_t));
            Elf64_Phdr const *phdr = phdri;
            for (unsigned j = 0; j < e_phnum; ++j, ++phdr) if (is_LOAD(phdr)) {
                upx_uint64_t vaddr = get_te64(&phdr->p_vaddr);
                upx_uint64_t filesz = get_te64(&phdr->p_filesz);
                upx_uint64_t d = plt_va - vaddr;
                if (d < filesz) {
                    upx_uint64_t offset = get_te64(&phdr->p_offset);
                    fi->seek(d + offset, SEEK_SET);
                    fi->readx(jump_slots, n_plt * sizeof(upx_uint64_t));
                    break;
                }
            }
            unRela64(dt_jmprel, jmp0, dt_pltrelsz, old_dtinit, fo);

            Elf64_Ehdr const *const o_ehdr = (Elf64_Ehdr const *)(void *)lowmem;
            unsigned const o_phnum = o_ehdr->e_phnum;
            if (((1<<16) - sizeof(Elf64_Ehdr)) / sizeof(Elf64_Phdr) < o_phnum)
                throwCantUnpack("bad Ehdr.e_phnum %#x", o_phnum);
            phdr = phdro;
            for (unsigned j = 0; j < o_phnum; ++j, ++phdr) if (is_LOAD(phdr)) {
                upx_uint64_t vaddr = get_te64(&phdr->p_vaddr);
                upx_uint64_t filesz = get_te64(&phdr->p_filesz);
                upx_uint64_t d = plt_va - vaddr - asl_delta;
                if (d < filesz) {
                    upx_uint64_t offset = get_te64(&phdr->p_offset);
                    if ((upx_uint64_t)file_size <= offset)
                        throwCantUnpack("bad phdr[%d].p_offset %#zx", j, (size_t)offset);
                    if (fo) {
                        fo->seek(d + offset, SEEK_SET);
                        fo->rewrite(jump_slots, n_plt * sizeof(upx_uint64_t));
                    }
                    break;
                }
            }
        }
        // Modified relocation tables are re-written by unRela64
    }
}

void PackLinuxElf64::unpack(OutputFile *fo)
{
    if (e_phoff != sizeof(Elf64_Ehdr)) {// Phdrs not contiguous with Ehdr
        throwCantUnpack("bad e_phoff");
    }
    unsigned const c_phnum = get_te16(&ehdri.e_phnum);
    unsigned u_phnum = 0;
    upx_uint64_t old_dtinit = 0;

    if (Elf64_Ehdr::ET_EXEC == get_te16(&ehdri.e_type)) {
// 40fddf17153ee3db73a04ff1bf288b91676138d6 2001-02-01 ph.version 11; b_info 12 bytes
// df9db96bd1c013c07da1d7ec740021d588ab2815 2001-01-17 ph.version 11; no b_info (==> 8 bytes)
        if (ph.version <= 11
        &&  get_te64(&ehdri.e_entry) < 0x401180
        &&  get_te16(&ehdri.e_machine)==Elf64_Ehdr::EM_X86_64) {
            // old style, 8-byte b_info:
            // sizeof(b_info.sz_unc) + sizeof(b_info.sz_cpr);
            szb_info = 2*sizeof(unsigned);
        }
    }

    fi->seek(overlay_offset - sizeof(l_info), SEEK_SET);
    fi->readx(&linfo, sizeof(linfo));
    if (UPX_MAGIC_LE32 != get_le32(&linfo.l_magic)) {
        NE32 const *const lp = (NE32 const *)(void const *)&linfo;
        // Workaround for bug of extra linfo by some asl_pack2_Shdrs().
        if (0==lp[0] && 0==lp[1] && 0==lp[2]) { // looks like blank extra
            fi->readx(&linfo, sizeof(linfo));
            if (UPX_MAGIC_LE32 == get_le32(&linfo.l_magic)) {
                overlay_offset += sizeof(linfo);
            }
            else {
                throwCantUnpack("l_info corrupted");
            }
        }
        else {
            throwCantUnpack("l_info corrupted");
        }
    }
    lsize = get_te16(&linfo.l_lsize);
    p_info hbuf;  fi->readx(&hbuf, sizeof(hbuf));
    unsigned orig_file_size = get_te32(&hbuf.p_filesize);
    blocksize = get_te32(&hbuf.p_blocksize);
    if ((u32_t)file_size > orig_file_size || blocksize > orig_file_size
        || (orig_file_size >> 8) > (u32_t)file_size  // heuristic anti-fuzz
        ||      (blocksize >> 8) > (u32_t)file_size
        || !mem_size_valid(1, blocksize, OVERHEAD))
        throwCantUnpack("p_info corrupted");

    ibuf.alloc(blocksize + OVERHEAD);
    b_info bhdr; memset(&bhdr, 0, sizeof(bhdr));
    fi->readx(&bhdr, szb_info);
    ph.u_len = get_te32(&bhdr.sz_unc);
    ph.c_len = get_te32(&bhdr.sz_cpr);
    ph.set_method(bhdr.b_method, overlay_offset + sizeof(p_info));
    if (ph.c_len > file_size_u || ph.c_len == 0 || ph.u_len == 0
    ||  ph.u_len > orig_file_size)
        throwCantUnpack("b_info corrupted");
    ph.filter_cto = bhdr.b_cto8;
    prev_method = bhdr.b_method;  // FIXME if multiple de-compressors

    MemBuffer u(ph.u_len);
    Elf64_Ehdr *const ehdr = (Elf64_Ehdr *)&u[0];
    Elf64_Phdr const *phdr = nullptr;
    total_in = 0;
    total_out = 0;
    unsigned c_adler = upx_adler32(nullptr, 0);
    unsigned u_adler = upx_adler32(nullptr, 0);

    unsigned is_shlib = 0;
    loader_offset = 0;
    MemBuffer o_elfhdrs;
    Elf64_Phdr const *const dynhdr = elf_find_ptype(Elf64_Phdr::PT_DYNAMIC, phdri, c_phnum);
    // dynseg was set by PackLinuxElf64help1
    if (dynhdr && !(Elf64_Dyn::DF_1_PIE & elf_unsigned_dynamic(Elf64_Dyn::DT_FLAGS_1))) {
        // Packed shlib? (ET_DYN without -fPIE)
        is_shlib = 1;
        xct_off = overlay_offset - sizeof(l_info);
        u_phnum = get_te16(&ehdri.e_phnum);
        o_elfhdrs.alloc(sz_elf_hdrs);
        un_shlib_1(fo, o_elfhdrs, c_adler, u_adler, orig_file_size);
        *ehdr = ehdri;
    }
    else { // main executable
        // Uncompress Ehdr and Phdrs: info for control of unpacking
        if (ibuf.getSize() < ph.c_len)
            throwCompressedDataViolation();

        fi->readx(ibuf, ph.c_len);
        // "clickhouse" ET_EXEC for amd64 has 0x200000 <= .e_entry
        // instead of 0x400000 that we checked earlier.
        if (8 == szb_info
        &&  Elf64_Ehdr::EM_X86_64 == e_machine
        &&  Elf64_Ehdr::ET_EXEC   == e_type
        &&  ph.u_len <= MAX_ELF_HDR_64
        ) {
            unsigned b_method = ibuf[0];
            unsigned b_extra  = ibuf[3];
            if (M_ZSTD >= b_method && 0 == b_extra) {
                unsigned where = fi->seek( -(upx_off_t)(ph.c_len + szb_info), SEEK_CUR);
                szb_info = 12;
                fi->readx(&bhdr, szb_info);
                ph.filter_cto = bhdr.b_cto8;
                ph.set_method(bhdr.b_method, where);
                prev_method = bhdr.b_method;  // FIXME if multiple de-compressors
                fi->readx(ibuf, ph.c_len);
            }
        }
        if (ph.u_len < sizeof(*ehdr))
            throwCantUnpack("ElfXX_Ehdr corrupted");
        decompress(ibuf, (upx_byte *)ehdr, false);
        if (ehdr->e_type   !=ehdri.e_type
        ||  ehdr->e_machine!=ehdri.e_machine
        ||  ehdr->e_version!=ehdri.e_version
            // less strict for EM_PPC64 to workaround earlier bug
        ||  !( ehdr->e_flags==ehdri.e_flags
            || Elf64_Ehdr::EM_PPC64 == get_te16(&ehdri.e_machine))
        ||  ehdr->e_ehsize !=ehdri.e_ehsize
            // check EI_MAG[0-3], EI_CLASS, EI_DATA, EI_VERSION
        ||  memcmp(ehdr->e_ident, ehdri.e_ident, Elf64_Ehdr::EI_OSABI)) {
            throwCantUnpack("ElfXX_Ehdr corrupted");
        }
        // Rewind: prepare for data phase
        fi->seek(- (off_t) (szb_info + ph.c_len), SEEK_CUR);

        u_phnum = get_te16(&ehdr->e_phnum);
        if ((umin(MAX_ELF_HDR_64, ph.u_len) - sizeof(Elf64_Ehdr))/sizeof(Elf64_Phdr) < u_phnum) {
            throwCantUnpack("bad compressed e_phnum");
        }
        o_elfhdrs.alloc(sizeof(Elf64_Ehdr) + u_phnum * sizeof(Elf64_Phdr));
        memcpy(o_elfhdrs, ehdr, o_elfhdrs.getSize());

        // Decompress each PT_LOAD.
        bool first_PF_X = true;
        phdr = (Elf64_Phdr *) (void *) (1+ ehdr);  // uncompressed
        for (unsigned j=0; j < u_phnum; ++phdr, ++j) {
            if (is_LOAD(phdr)) {
                unsigned const filesz = get_te64(&phdr->p_filesz);
                unsigned const offset = get_te64(&phdr->p_offset);
                if (fo) {
                    fo->seek(offset, SEEK_SET);
                    if (total_out < offset) {
                        total_out = offset;  // FIXME: can it be re-write?
                    }
                }
                if (Elf64_Phdr::PF_X & get_te32(&phdr->p_flags)) {
                    unpackExtent(filesz, fo,
                        c_adler, u_adler, first_PF_X);
                    first_PF_X = false;
                }
                else {
                    unpackExtent(filesz, fo,
                        c_adler, u_adler, false);
                }
            }
        }
    }

    upx_uint64_t const e_entry = get_te64(&ehdri.e_entry);
    unsigned off_entry = 0;
    phdr = phdri;
    load_va = 0;
    for (unsigned j=0; j < c_phnum; ++j, ++phdr) {
        if (is_LOAD(phdr)) {
            upx_uint64_t offset = get_te64(&phdr->p_offset);
            upx_uint64_t vaddr  = get_te64(&phdr->p_vaddr);
            upx_uint64_t filesz = get_te64(&phdr->p_filesz);
            if (!load_va) {
                load_va = vaddr;
            }
            if ((e_entry - vaddr) < filesz) {
                off_entry = (e_entry - vaddr) + offset;
                break;
            }
        }
    }
    unsigned d_info[6];
    unsigned sz_d_info = sizeof(d_info);
    if (!is_shlib) {
        if (get_te32(&phdri[0].p_flags) & Elf64_Phdr::PF_X) {
            // Old style, such as upx-3.91 thru upx-3.95
            switch (this->e_machine) {
                default: {
                    char msg[40]; snprintf(msg, sizeof(msg),
                        "Unknown architecture %d", this->e_machine);
                    throwCantUnpack(msg);
                }; break;
                case Elf64_Ehdr::EM_AARCH64: sz_d_info = 4 * sizeof(unsigned); break;
                case Elf64_Ehdr::EM_PPC64:   sz_d_info = 3 * sizeof(unsigned); break;
                case Elf64_Ehdr::EM_X86_64:  sz_d_info = 2 * sizeof(unsigned); break;
            }
        }
        loader_offset = off_entry - sz_d_info;
    }

    if (0x1000==get_te64(&phdri[0].p_filesz)  // detect C_BASE style
    &&  0==get_te64(&phdri[1].p_offset)
    &&  0==get_te64(&phdri[0].p_offset)
    &&     get_te64(&phdri[1].p_filesz) == get_te64(&phdri[1].p_memsz)) {
        fi->seek(up4(get_te64(&phdri[1].p_memsz)), SEEK_SET);  // past the loader
    }
    else if (is_shlib
    ||  (off_entry + up4(lsize) + ph.getPackHeaderSize() + sizeof(overlay_offset))
            < up4(file_size)) {
        // Loader is not at end; skip past it.
        if (loader_offset) {
            fi->seek(loader_offset, SEEK_SET);
        }
        else {
            funpad4(fi);  // MATCH01
        }
        fi->readx(d_info, sz_d_info);
        if (is_shlib && 0==old_dtinit) {
            old_dtinit = get_te32(&d_info[2 + (0==d_info[0])]);
            is_asl = 1u& get_te32(&d_info[0 + (0==d_info[0])]);
        }
        fi->seek(lsize - sz_d_info, SEEK_CUR);
    }

    // The gaps between PT_LOAD and after last PT_LOAD
    phdr = (Elf64_Phdr const *)(1+ (Elf64_Ehdr const *)(void const *)o_elfhdrs);
    upx_uint64_t hi_offset(0);
    for (unsigned j = 0; j < u_phnum; ++j) {
        upx_uint64_t const offset = get_te64(&phdr[j].p_offset);
        if (is_LOAD(&phdr[j])
        &&  hi_offset < offset) {
            hi_offset = offset;
        }
    }
    for (unsigned j = 0; j < u_phnum; ++j) {
        unsigned const size = find_LOAD_gap(phdr, j, u_phnum);
        if (size) {
            unsigned const where = get_te64(&phdr[j].p_offset) +
                                   get_te64(&phdr[j].p_filesz);
            if (fo)
                fo->seek(where, SEEK_SET);
            { // Recover from some piracy [also serves as error tolerance :-) ]
              // Getting past the loader is problematic, due to unintended
              // variances between released versions:
              //   l_info.l_lsize might be rounded up by 8 instead of by 4, and
              //   sz_d_info might have changed.
                b_info b_peek, *bp = &b_peek;
                fi->readx(bp, sizeof(b_peek));
                upx_off_t pos = fi->seek(-(off_t)sizeof(b_peek), SEEK_CUR);
                unsigned sz_unc = get_te32(&bp->sz_unc);
                unsigned sz_cpr = get_te32(&bp->sz_cpr);
                unsigned word3  = get_te32(&bp->b_method);
                unsigned method = bp->b_method;
                unsigned ftid = bp->b_ftid;
                unsigned cto8 = bp->b_cto8;
                if (!( ((sz_cpr == sz_unc) && (0 == word3) && (size == sz_unc)) // incompressible literal
                    || ((sz_cpr <  sz_unc)
                        && (method == prev_method || M_NRV2B_LE32 == prev_method)
                        && (0 == ftid) && (0 == cto8)) )
                ) {
                    opt->info_mode++;
                    infoWarning("bad b_info at %#zx", (size_t)pos);
                    unsigned const N_PEEK(16 * sizeof(int)), H_PEEK(N_PEEK >> 1);
                    unsigned char peek_arr[N_PEEK];
                    fi->seek(pos - H_PEEK, SEEK_SET);
                    fi->readx(peek_arr, sizeof(peek_arr));
                    fi->seek(pos, SEEK_SET);
                    bool const is_be = ELFDATA2MSB == ehdri.e_ident[EI_DATA];
                    if (is_be) {
                        // Does the right thing for sz_unc and sz_cpr,
                        // but swaps b_method and b_extra.  Need find_be32() :-)
                        for (unsigned k = 0; k < N_PEEK; k += sizeof(int)) {
                            set_le32(&peek_arr[k], get_be32(&peek_arr[k]));
                        }
                    }
                    int boff = find_le32(peek_arr, sizeof(peek_arr), size);
                    if (boff < 0
                    || sizeof(peek_arr) < (boff + sizeof(b_info))) {
                        throwCantUnpack("b_info corrupted");
                    }
                    bp = (b_info *)(void *)&peek_arr[boff];

                    sz_unc = get_le32(&bp->sz_unc);
                    sz_cpr = get_le32(&bp->sz_cpr);
                    word3  = get_le32(&bp->b_method);
                    ftid = bp->b_ftid;
                    cto8 = bp->b_cto8;
                    if (0 <= boff  // found
                    && ( ((sz_cpr == sz_unc) && (0 == word3) && (size == sz_unc)) // incompressible literal
                      || ((sz_cpr <  sz_unc) && (0 == ftid) && (0 == cto8)
                          && ((is_be ? bp->b_extra : bp->b_method) == prev_method)) )
                    ) {
                        pos -= H_PEEK;
                        pos += boff;
                        infoWarning("... recovery at %#zx", (size_t)pos);
                        fi->seek(pos, SEEK_SET);
                    }
                    opt->info_mode--;
                }
            }
            unpackExtent(size, fo,
                c_adler, u_adler, false,
                (hi_offset != get_te64(&phdr[j].p_offset)));
        }
    }

    // check for end-of-file
    fi->readx(&bhdr, szb_info);
    ph.set_method(bhdr.b_method, ~0u);
    unsigned const sz_unc = ph.u_len = get_te32(&bhdr.sz_unc);

    if (sz_unc == 0) { // uncompressed size 0 -> EOF
        // note: magic is always stored le32
        unsigned const sz_cpr = get_le32(&bhdr.sz_cpr);
        if (sz_cpr != UPX_MAGIC_LE32)  // sz_cpr must be h->magic
            throwCompressedDataViolation();
    }
    else { // extra bytes after end?
        throwCompressedDataViolation();
    }

    if (is_shlib) {
        un_DT_INIT(old_dtinit, (Elf64_Phdr *)(1+ (Elf64_Ehdr *)(void *)o_elfhdrs), dynhdr, fo);
    }

    // update header with totals
    ph.c_len = total_in;
    ph.u_len = total_out;

    // all bytes must be written
    if (fo && total_out != orig_file_size)
        throwEOFException();

    // finally test the checksums
    if (ph.c_adler != c_adler || ph.u_adler != u_adler)
        throwChecksumError();
}


/*************************************************************************
//
**************************************************************************/

PackLinuxElf32x86::PackLinuxElf32x86(InputFile *f) : super(f)
{
    e_machine = Elf32_Ehdr::EM_386;
    ei_class  = Elf32_Ehdr::ELFCLASS32;
    ei_data   = Elf32_Ehdr::ELFDATA2LSB;
    ei_osabi  = Elf32_Ehdr::ELFOSABI_LINUX;
}

PackLinuxElf32x86::~PackLinuxElf32x86()
{
}

tribool PackLinuxElf32x86::canUnpack() // bool, except -1: format known, but not packed
{
    if (super::canUnpack()) {
        return true;
    }
    return false;
}

Linker* PackLinuxElf32x86::newLinker() const
{
    return new ElfLinkerX86;
}

PackBSDElf32x86::PackBSDElf32x86(InputFile *f) : super(f)
{
    e_machine = Elf32_Ehdr::EM_386;
    ei_class  = Elf32_Ehdr::ELFCLASS32;
    ei_data   = Elf32_Ehdr::ELFDATA2LSB;
}

PackBSDElf32x86::~PackBSDElf32x86()
{
}

PackFreeBSDElf32x86::PackFreeBSDElf32x86(InputFile *f) : super(f)
{
    ei_osabi  = Elf32_Ehdr::ELFOSABI_FREEBSD;
}

PackFreeBSDElf32x86::~PackFreeBSDElf32x86()
{
}

PackNetBSDElf32x86::PackNetBSDElf32x86(InputFile *f) : super(f)
{
    ei_osabi  = Elf32_Ehdr::ELFOSABI_NETBSD;
    osabi_note = "NetBSD";
}

PackNetBSDElf32x86::~PackNetBSDElf32x86()
{
}

PackOpenBSDElf32x86::PackOpenBSDElf32x86(InputFile *f) : super(f)
{
    ei_osabi  = Elf32_Ehdr::ELFOSABI_OPENBSD;
    osabi_note = "OpenBSD";
}

PackOpenBSDElf32x86::~PackOpenBSDElf32x86()
{
}

int const *
PackLinuxElf32x86::getFilters() const
{
    static const int filters[] = {
        // 0x49 is 5-byte CALL or JMP, and 6-byte Jxx
        // 0x46 is 5-byte CALL or JMP
        0x49, 0x46,
// FIXME 2002-11-11: We use stub/fold_elf86.asm, which calls the
// decompressor multiple times, and unfilter is independent of decompress.
// Currently only filters 0x49, 0x46, 0x80..0x87 can handle this;
// and 0x80..0x87 are regarded as "untested".
#if 0
        0x26, 0x24, 0x11, 0x14, 0x13, 0x16, 0x25, 0x15, 0x12,
#endif
#if 0
        0x83, 0x36, 0x26,
              0x86, 0x80,
        0x84, 0x87, 0x81,
        0x82, 0x85,
        0x24, 0x16, 0x13, 0x14, 0x11, 0x25, 0x15, 0x12,
#endif
    FT_END };
    return filters;
}

PackLinuxElf32armLe::PackLinuxElf32armLe(InputFile *f) : super(f)
{
    e_machine = Elf32_Ehdr::EM_ARM;
    ei_class  = Elf32_Ehdr::ELFCLASS32;
    ei_data   = Elf32_Ehdr::ELFDATA2LSB;
    ei_osabi  = Elf32_Ehdr::ELFOSABI_ARM;
}

PackLinuxElf32armLe::~PackLinuxElf32armLe()
{
}

PackLinuxElf32mipseb::PackLinuxElf32mipseb(InputFile *f) : super(f)
{
    e_machine = Elf32_Ehdr::EM_MIPS;
    ei_class  = Elf32_Ehdr::ELFCLASS32;
    ei_data   = Elf32_Ehdr::ELFDATA2MSB;
    ei_osabi  = Elf32_Ehdr::ELFOSABI_LINUX;
}

PackLinuxElf32mipseb::~PackLinuxElf32mipseb()
{
}

PackLinuxElf32mipsel::PackLinuxElf32mipsel(InputFile *f) : super(f)
{
    e_machine = Elf32_Ehdr::EM_MIPS;
    ei_class  = Elf32_Ehdr::ELFCLASS32;
    ei_data   = Elf32_Ehdr::ELFDATA2LSB;
    ei_osabi  = Elf32_Ehdr::ELFOSABI_LINUX;
}

PackLinuxElf32mipsel::~PackLinuxElf32mipsel()
{
}

Linker* PackLinuxElf32armLe::newLinker() const
{
    return new ElfLinkerArmLE();
}

Linker* PackLinuxElf32mipseb::newLinker() const
{
    return new ElfLinkerMipsBE();
}

Linker* PackLinuxElf32mipsel::newLinker() const
{
    return new ElfLinkerMipsLE();
}

PackLinuxElf32armBe::PackLinuxElf32armBe(InputFile *f) : super(f)
{
    e_machine = Elf32_Ehdr::EM_ARM;
    ei_class  = Elf32_Ehdr::ELFCLASS32;
    ei_data   = Elf32_Ehdr::ELFDATA2MSB;
    ei_osabi  = Elf32_Ehdr::ELFOSABI_ARM;
}

PackLinuxElf32armBe::~PackLinuxElf32armBe()
{
}

Linker* PackLinuxElf32armBe::newLinker() const
{
    return new ElfLinkerArmBE();
}

unsigned
PackLinuxElf32::elf_get_offset_from_address(unsigned addr) const
{
    return elf_get_offset_from_Phdrs(addr, phdri);
}

unsigned
PackLinuxElf32::elf_get_offset_from_Phdrs(unsigned addr, Elf32_Phdr const *phdr0) const
{
    Elf32_Phdr const *phdr = phdr0;
    int j = e_phnum;
    for (; --j>=0; ++phdr) if (is_LOAD(phdr)) {
        unsigned const t = addr - get_te32(&phdr->p_vaddr);
        if (t < get_te32(&phdr->p_filesz)) {
            unsigned const p_offset = get_te32(&phdr->p_offset);
            if (file_size_u <= p_offset) { // FIXME: weak
                char msg[40]; snprintf(msg, sizeof(msg),
                    "bad Elf32_Phdr[%d].p_offset %x",
                    -1+ e_phnum - j, p_offset);
                throwCantPack(msg);
            }
            return t + p_offset;
        }
    }
    return 0;
}

u32_t  // returns .p_offset
PackLinuxElf32::check_pt_load(Elf32_Phdr const *const phdr)
{
    u32_t filesz = get_te32(&phdr->p_filesz);
    u32_t offset = get_te32(&phdr->p_offset), offend = filesz + offset;
    u32_t vaddr  = get_te32(&phdr->p_vaddr);
    u32_t paddr  = get_te32(&phdr->p_paddr);
    u32_t align  = get_te32(&phdr->p_align);

    if ((-1+ align) & (paddr ^ vaddr)
    ||  file_size_u32 <= (u32_t)offset
    ||  file_size_u32 <  (u32_t)offend
    ||  file_size_u32 <  (u32_t)filesz) {
        char msg[50]; snprintf(msg, sizeof(msg), "bad PT_LOAD phdr[%u]",
            (unsigned)(phdr - phdri));
        throwCantPack(msg);
    }
    return offset;
}

Elf32_Dyn const *
PackLinuxElf32::elf_has_dynamic(unsigned int key) const
{
    Elf32_Dyn const *dynp= dynseg;
    if (dynp)
    for (; Elf32_Dyn::DT_NULL!=dynp->d_tag; ++dynp) if (get_te32(&dynp->d_tag)==key) {
        return dynp;
    }
    return nullptr;
}

unsigned  // checked .p_offset; sz_dynseg set
PackLinuxElf32::check_pt_dynamic(Elf32_Phdr const *const phdr)
{
    unsigned t = get_te32(&phdr->p_offset), s = sizeof(Elf32_Dyn) + t;
    unsigned vaddr = get_te32(&phdr->p_vaddr);
    unsigned filesz = get_te32(&phdr->p_filesz), memsz = get_te32(&phdr->p_memsz);
    unsigned align = get_te32(&phdr->p_align);
    if (file_size_u < t || s < t
    ||  file_size_u < filesz
    ||  file_size_u < (filesz + t)
    ||  t < (e_phnum*sizeof(Elf32_Phdr) + sizeof(Elf32_Ehdr))
    ||  (3 & t) || (7 & (filesz | memsz))  // .balign 4; 8==sizeof(Elf32_Dyn)
    ||  (-1+ align) & (t ^ vaddr)
    ||  file_size_u <= memsz
    ||  filesz < sizeof(Elf32_Dyn)
    ||  memsz  < sizeof(Elf32_Dyn)
    ||  filesz < memsz) {
        char msg[50]; snprintf(msg, sizeof(msg), "bad PT_DYNAMIC phdr[%u]",
            (unsigned)(phdr - phdri));
        throwCantPack(msg);
    }
    sz_dynseg = memsz;
    return t;
}

Elf32_Dyn *PackLinuxElf32::elf_find_dynptr(unsigned int key) const
{
    Elf32_Dyn *dynp= dynseg;
    if (dynp) {
        Elf32_Dyn *const last = (Elf32_Dyn *)(sz_dynseg + (char *)dynseg);
        for (; dynp < last; ++dynp) {
            if (get_te32(&dynp->d_tag)==key) {
                return dynp;
            }
            if (Elf32_Dyn::DT_NULL == dynp->d_tag) {
                return nullptr;
            }
        }
    }
    return nullptr;
}

Elf64_Dyn *PackLinuxElf64::elf_find_dynptr(unsigned int key) const
{
    Elf64_Dyn *dynp= dynseg;
    if (dynp) {
        Elf64_Dyn *const last = (Elf64_Dyn *)(sz_dynseg + (char *)dynseg);
        for (; dynp < last; ++dynp) {
            if (get_te64(&dynp->d_tag)==key) {
                return dynp;
            }
            if (Elf64_Dyn::DT_NULL == dynp->d_tag) {
                return nullptr;
            }
        }
    }
    return nullptr;
}

void *
PackLinuxElf32::elf_find_dynamic(unsigned int key) const
{
    Elf32_Dyn const *dynp= elf_find_dynptr(key);
    if (dynp) {
        unsigned const t= elf_get_offset_from_address(get_te32(&dynp->d_val));
        if (t && t < file_size_u) {
            return t + file_image;
        }
    }
    return nullptr;
}

void *
PackLinuxElf64::elf_find_dynamic(unsigned int key) const
{
    Elf64_Dyn const *dynp= elf_find_dynptr(key);
    if (dynp) {
        upx_uint64_t const t= elf_get_offset_from_address(get_te64(&dynp->d_val));
        if (t && t < file_size_u) {
            return t + file_image;
        }
    }
    return nullptr;
}

upx_uint64_t
PackLinuxElf64::elf_unsigned_dynamic(unsigned int key) const
{
    Elf64_Dyn const *dynp= elf_find_dynptr(key);
    if (dynp) {
        return get_te64(&dynp->d_val);
    }
    return 0;
}

upx_uint64_t
PackLinuxElf32::elf_unsigned_dynamic(unsigned int key) const
{
    Elf32_Dyn const *dynp= elf_find_dynptr(key);
    if (dynp) {
        return get_te32(&dynp->d_val);
    }
    return 0;
}

upx_uint64_t
PackLinuxElf64::elf_get_offset_from_address(upx_uint64_t addr) const
{
    Elf64_Phdr const *phdr = phdri;
    int j = e_phnum;
    for (; --j>=0; ++phdr) if (is_LOAD(phdr)) {
        upx_uint64_t const t = addr - get_te64(&phdr->p_vaddr);
        if (t < get_te64(&phdr->p_filesz)) {
            upx_uint64_t const p_offset = get_te64(&phdr->p_offset);
            if (file_size_u <= p_offset) { // FIXME: weak
                char msg[40]; snprintf(msg, sizeof(msg),
                    "bad Elf64_Phdr[%d].p_offset %#lx",
                    -1+ e_phnum - j, (long unsigned)p_offset);
                throwCantPack(msg);
            }
            return t + p_offset;
        }
    }
    return 0;
}

u64_t  // returns .p_offset
PackLinuxElf64::check_pt_load(Elf64_Phdr const *const phdr)
{
    u64_t filesz = get_te64(&phdr->p_filesz);
    u64_t offset = get_te64(&phdr->p_offset), offend = filesz + offset;
    u64_t vaddr  = get_te64(&phdr->p_vaddr);
    u64_t paddr  = get_te64(&phdr->p_paddr);
    u64_t align  = get_te64(&phdr->p_align);

    if ((-1+ align) & (paddr ^ vaddr)
    ||  file_size_u <= (u64_t)offset
    ||  file_size_u <  (u64_t)offend
    ||  file_size_u <  (u64_t)filesz) {
        char msg[50]; snprintf(msg, sizeof(msg), "bad PT_LOAD phdr[%u]",
            (unsigned)(phdr - phdri));
        throwCantPack(msg);
    }
    return offset;
}

Elf64_Dyn const *
PackLinuxElf64::elf_has_dynamic(unsigned int key) const
{
    Elf64_Dyn const *dynp= dynseg;
    if (dynp)
    for (; Elf64_Dyn::DT_NULL!=dynp->d_tag; ++dynp) if (get_te64(&dynp->d_tag)==key) {
        return dynp;
    }
    return nullptr;
}

upx_uint64_t  // checked .p_offset; sz_dynseg set
PackLinuxElf64::check_pt_dynamic(Elf64_Phdr const *const phdr)
{
    upx_uint64_t t = get_te64(&phdr->p_offset), s = sizeof(Elf64_Dyn) + t;
    upx_uint64_t vaddr = get_te64(&phdr->p_vaddr);
    upx_uint64_t filesz = get_te64(&phdr->p_filesz), memsz = get_te64(&phdr->p_memsz);
    upx_uint64_t align = get_te64(&phdr->p_align);
    if (file_size_u < t || s < t
    ||  file_size_u < filesz
    ||  file_size_u < (filesz + t)
    ||  t < (e_phnum*sizeof(Elf64_Phdr) + sizeof(Elf64_Ehdr))
    ||  (7 & t) || (0xf & (filesz | memsz))  // .balign 8; 16==sizeof(Elf64_Dyn)
    ||  (-1+ align) & (t ^ vaddr)
    ||  file_size_u <= memsz
    ||  filesz < sizeof(Elf64_Dyn)
    ||  memsz  < sizeof(Elf64_Dyn)
    ||  filesz < memsz) {
        char msg[50]; snprintf(msg, sizeof(msg), "bad PT_DYNAMIC phdr[%u]",
            (unsigned)(phdr - phdri));
        throwCantPack(msg);
    }
    sz_dynseg = memsz;
    return t;
}

void
PackLinuxElf64::sort_DT64_offsets(Elf64_Dyn const *const dynp0)
{
    mb_dt_offsets.alloc(sizeof(unsigned) * sizeof(dt_keys)/sizeof(dt_keys[0]));
    mb_dt_offsets.clear();
    dt_offsets = (unsigned *)mb_dt_offsets.getVoidPtr();
    unsigned n_off = 0, k;
    for (unsigned j=0; ((k = dt_keys[j]),  k); ++j) {
        dt_offsets[n_off] = 0;  // default to "not found"
        u64_t rva = 0;
        if (k < DT_NUM) { // in range of easy table
            if (!dt_table[k]) {
                continue;
            }
            rva = get_te64(&dynp0[-1+ dt_table[k]].d_val);
        }
        else if (file_image) { // why is this guard necessary?
            rva = elf_unsigned_dynamic(k);  // zero if not found
        }
        if (!rva) { // not present in input
            continue;
        }
        Elf64_Phdr const *const phdr = elf_find_Phdr_for_va(rva, phdri, e_phnum);
        if (!phdr) {
            char msg[60]; snprintf(msg, sizeof(msg), "bad DT_{%#x} = %#llx (no Phdr)",
                k, rva);
            throwCantPack(msg);
        }
        dt_offsets[n_off] = (rva - get_te64(&phdr->p_vaddr)) + get_te64(&phdr->p_offset);

        if (file_size <= dt_offsets[n_off]) {
            char msg[60]; snprintf(msg, sizeof(msg), "bad DT_{%#x} = %#x (beyond EOF)",
                k, dt_offsets[n_off]);
                throwCantPack(msg);
        }
        n_off += !!dt_offsets[n_off];
    }
    dt_offsets[n_off++] = file_size;  // sentinel
    upx_qsort(dt_offsets, n_off, sizeof(dt_offsets[0]), qcmp_unsigned);
}

unsigned PackLinuxElf64::find_dt_ndx(u64_t rva)
{
    unsigned *const dto = (unsigned *)mb_dt_offsets.getVoidPtr();
    unsigned const dto_size = mb_dt_offsets.getSize() / sizeof(*dto);
    for (unsigned j = 0; j < dto_size && dto[j]; ++j) { // linear search of short table
        if (rva == dto[j]) {
            return j;
        }
    }
    return ~0u;
}

unsigned PackLinuxElf64::elf_find_table_size(unsigned dt_type, unsigned sh_type)
{
    Elf64_Shdr const *sec = elf_find_section_type(sh_type);
    if (sec) { // Cheat the easy way: use _Shdr.  (No _Shdr anyway for de-compression)
        return get_te64(&sec->sh_size);
    }
    // Honest hard work: use _Phdr
    unsigned x_rva;
    if (dt_type < DT_NUM) {
        unsigned const x_ndx = dt_table[dt_type];
        if (!x_ndx) { // no such entry
            return 0;
        }
        x_rva = get_te64(&dynseg[-1+ x_ndx].d_val);
    }
    else {
        x_rva = elf_unsigned_dynamic(dt_type);
    }
    Elf64_Phdr const *const x_phdr = elf_find_Phdr_for_va(x_rva, phdri, e_phnum);
    if (!x_phdr)
        return ~0u;  // corrupted Phdrs?
    unsigned const           d_off =             x_rva - get_te64(&x_phdr->p_vaddr);
    unsigned const           y_ndx = find_dt_ndx(d_off + get_te64(&x_phdr->p_offset));
    if (~0u != y_ndx) {
        return dt_offsets[1+ y_ndx] - dt_offsets[y_ndx];
    }
    return ~0u;
}

void
PackLinuxElf64::invert_pt_dynamic(Elf64_Dyn const *dynp, upx_uint64_t headway)
{
    if (dt_table[Elf64_Dyn::DT_NULL]) {
        return;  // not 1st time; do not change upx_dt_init
    }
    Elf64_Dyn const *const dynp0 = dynp;
    unsigned ndx = 0;
    unsigned const limit = headway / sizeof(*dynp);
    if (dynp)
    for (; ; ++ndx, ++dynp) {
        if (limit <= ndx) {
            throwCantPack("DT_NULL not found");
        }
        upx_uint64_t const d_tag = get_te64(&dynp->d_tag);
        if (d_tag>>32) { // outrageous
            throwCantPack("bad Elf64_Dyn[%d].d_tag %#lx",
                ndx, (long unsigned)d_tag);
        }
        if (d_tag < DT_NUM) {
            if (Elf64_Dyn::DT_NEEDED != d_tag
            &&  dt_table[d_tag]
            &&    get_te64(&dynp->d_val)
               != get_te64(&dynp0[-1+ dt_table[d_tag]].d_val)) {
                throwCantPack("duplicate DT_%#x: [%#x] [%#x]",
                    (unsigned)d_tag, -1+ dt_table[d_tag], ndx);
            }
            dt_table[d_tag] = 1+ ndx;
        }
        if (Elf64_Dyn::DT_NULL == d_tag) {
            break;  // check here so that dt_table[DT_NULL] is set
        }
    }
    sort_DT64_offsets(dynp0);

    upx_dt_init = 0;
         if (dt_table[Elf64_Dyn::DT_INIT])          upx_dt_init = Elf64_Dyn::DT_INIT;
    else if (dt_table[Elf64_Dyn::DT_PREINIT_ARRAY]) upx_dt_init = Elf64_Dyn::DT_PREINIT_ARRAY;
    else if (dt_table[Elf64_Dyn::DT_INIT_ARRAY])    upx_dt_init = Elf64_Dyn::DT_INIT_ARRAY;

    unsigned const z_str = dt_table[Elf64_Dyn::DT_STRSZ];
    strtab_max = !z_str ? 0 : get_te64(&dynp0[-1+ z_str].d_val);
    unsigned const z_tab = dt_table[Elf64_Dyn::DT_STRTAB];
    unsigned const tmp1 = !z_tab ? 0 : get_te64(&dynp0[-1+ z_tab].d_val);
    if (tmp1 < sz_elf_hdrs) {
        throwCantPack("bad DT_STRTAB %#x", tmp1);
    }
    unsigned const strtab_beg = !z_tab ? 0 : elf_get_offset_from_address(tmp1);

    if (!z_str || !z_tab || !(strtab_max + strtab_beg)
    || (this->file_size - strtab_beg) < strtab_max  // strtab overlaps EOF
        // last string in table must have terminating NUL
    ||  '\0' != ((char *)file_image.getVoidPtr())[-1+ strtab_max + strtab_beg]
    ) {
        throwCantPack("bad DT_STRSZ %#x", strtab_max);
    }

    { // Find end of DT_SYMTAB
        unsigned const tmp2 = elf_find_table_size(Elf64_Dyn::DT_SYMTAB,
            Elf64_Shdr::SHT_DYNSYM);
        symnum_max = (~0u == tmp2) ? 0 : tmp2 / sizeof(Elf64_Sym);
    }

    unsigned v_sym = dt_table[Elf64_Dyn::DT_SYMTAB];
    if (v_sym) {
        v_sym = elf_get_offset_from_address(get_te64(&dynp0[-1+ v_sym].d_val));
    }

    unsigned v_hsh = dt_table[Elf64_Dyn::DT_HASH];
    if (v_hsh) {
        v_hsh = elf_get_offset_from_address(get_te64(&dynp0[-1+ v_hsh].d_val));
    }
    if (v_hsh && file_image) {
        hashtab = (unsigned const *)elf_find_dynamic(Elf64_Dyn::DT_HASH);
        if (!hashtab) {
            throwCantPack("bad DT_HASH %#x", v_hsh);
        }
        // Find end of DT_HASH
        hashend = (unsigned const *)(void const *)(elf_find_table_size(
            Elf64_Dyn::DT_HASH, Elf64_Shdr::SHT_HASH) + (char const *)hashtab);
        if (!hashtab || (char const *)hashend <= (char const *)&hashtab[2]
        ||  file_image.getSizeInBytes()
            < (unsigned)((char const *)&hashtab[2] - (char *)&file_image[0]) )
        {
            throwCantPack("bad DT_HASH %#x", v_hsh);
        }

        unsigned const nbucket = get_te32(&hashtab[0]);
        unsigned const *const buckets = &hashtab[2];
        unsigned const *const chains = &buckets[nbucket]; (void)chains;
        if ((unsigned)(file_size - ((char const *)buckets - (char const *)(void const *)file_image))
                <= sizeof(unsigned)*nbucket ) {
            throwCantPack("bad nbucket %#x\n", nbucket);
        }

        if ((unsigned)(hashend - buckets) < nbucket
        || !v_sym || file_size_u <= v_sym
        || ((v_hsh < v_sym) && (v_sym - v_hsh) < sizeof(*buckets)*(2+ nbucket))
        ) {
            throwCantPack("bad DT_HASH nbucket=%#x  len=%#x",
                nbucket, (v_sym - v_hsh));
        }
        unsigned chmax = 0;
        for (unsigned j= 0; j < nbucket; ++j) {
            unsigned x = get_te32(&buckets[j]);
            if (chmax < x) {
                chmax = x;
            }
        }
        if ((v_hsh < v_sym) && (v_sym - v_hsh) <
                (sizeof(*buckets)*(2+ nbucket) + sizeof(*chains)*(1+ chmax))) {
            throwCantPack("bad DT_HASH nbucket=%#x  len=%#x",
                nbucket, (v_sym - v_hsh));
        }
    }
    unsigned const v_gsh = elf_unsigned_dynamic(Elf64_Dyn::DT_GNU_HASH);
    if (v_gsh && file_image) {
        // Not similar to DT_HASH because DT_GNU_HASH is not small (0x6ffffef5).
        gashtab = (unsigned const *)elf_find_dynamic(Elf64_Dyn::DT_GNU_HASH);
        gashend = (unsigned const *)(void const *)(elf_find_table_size(
            Elf64_Dyn::DT_GNU_HASH, Elf64_Shdr::SHT_GNU_HASH) + (char const *)gashtab);
        if (!gashtab || (char const *)gashend <= (char const *)&gashtab[4]
        ||  file_image.getSizeInBytes()
            < (unsigned)((char const *)&gashtab[4] - (char *)&file_image[0]) )
        {
            throwCantPack("bad DT_GNU_HASH %#x", v_gsh);
        }

        unsigned const n_bucket = get_te32(&gashtab[0]);
        unsigned const symbias  = get_te32(&gashtab[1]);
        unsigned const n_bitmask = get_te32(&gashtab[2]);
        unsigned const gnu_shift = get_te32(&gashtab[3]);
        upx_uint64_t const *const bitmask = (upx_uint64_t const *)(void const *)&gashtab[4];
        unsigned     const *const buckets = (unsigned const *)&bitmask[n_bitmask];
        unsigned     const *const hasharr = &buckets[n_bucket]; (void)hasharr;
        if (!n_bucket || (1u<<31) <= n_bucket  /* fie on fuzzers */
        || (unsigned)(gashend - buckets) < n_bucket
        || (file_size + file_image) <= (void const *)hasharr) {
            throwCantPack("bad n_bucket %#x\n", n_bucket);
        }
        // It would be better to detect zeroes shifted into low 6 bits of:
        //    (077 & (hash_32 >> gnu_shift))
        // but compilers can be stupid.
        if (31 < gnu_shift) {
            throwCantPack("bad gnu_shift %#x", gnu_shift);
        }
        // unsigned const *const gashend = &hasharr[n_bucket];
        // minimum, except:
        // Rust and Android trim unused zeroes from high end of hasharr[]
        unsigned bmax = 0;
        for (unsigned j= 0; j < n_bucket; ++j) {
            unsigned bj = get_te32(&buckets[j]);
            if (bj) {
                if (bj < symbias) {
                    throwCantPack("bad DT_GNU_HASH bucket[%d] < symbias{%#x}\n",
                            bj, symbias);
                }
                if (bmax < bj) {
                    bmax = bj;
                }
            }
        }
        if (1==n_bucket  && 0==buckets[0]
        &&  1==n_bitmask && 0==bitmask[0]) {
            // 2021-09-11 Rust on RaspberryPi apparently uses this to minimize space.
            // But then the DT_GNU_HASH symbol lookup algorithm always fails?
            // https://github.com/upx/upx/issues/525
        } else
        if (bmax) {
            if ((1+ bmax) < symbias) {
                throwCantPack("bad DT_GNU_HASH (1+ max_bucket)=%#x < symbias=%#x",
                    1+ bmax, symbias);
            }
            bmax -= symbias;
        }

        unsigned r = 0;
        if (!n_bucket || !n_bitmask || !v_sym
        || (r=1, ((-1+ n_bitmask) & n_bitmask))  // not a power of 2
        || (r=2, (8*sizeof(upx_uint64_t) <= gnu_shift))  // shifted result always == 0
        || (r=3, (n_bucket>>30))  // fie on fuzzers
        || (r=4, (n_bitmask>>30))
        || (r=5, ((file_size/sizeof(unsigned))
                <= ((sizeof(*bitmask)/sizeof(unsigned))*n_bitmask + 2*n_bucket)))  // FIXME: weak
        || (r=6, ((v_gsh < v_sym) && (v_sym - v_gsh) < (sizeof(unsigned)*4  // headers
                + sizeof(*bitmask)*n_bitmask  // bitmask
                + sizeof(*buckets)*n_bucket  // buckets
                + sizeof(*hasharr)*(!bmax ? 0 : (1+ bmax))  // hasharr
            )) )
        ) {
            throwCantPack("bad DT_GNU_HASH n_bucket=%#x  n_bitmask=%#x  len=%#lx  r=%d",
                n_bucket, n_bitmask, (long unsigned)(v_sym - v_gsh), r);
        }
    }
    e_shstrndx = get_te16(&ehdri.e_shstrndx);  // who omitted this?
    if (e_shnum <= e_shstrndx
    &&  !(0==e_shnum && 0==e_shstrndx) ) {
        throwCantPack("bad .e_shstrndx %d >= .e_shnum %d", e_shstrndx, e_shnum);
    }
}

unsigned PackLinuxElf::gnu_hash(char const *q)
{
    unsigned char const *p = (unsigned char const *)q;
    unsigned h;

    for (h= 5381; 0!=*p; ++p) {
        h += *p + (h << 5);
    }
    return h;
}

unsigned PackLinuxElf::elf_hash(char const *p)
{
    unsigned h;
    for (h= 0; 0!=*p; ++p) {
        h = *p + (h<<4);
        {
            unsigned const t = 0xf0000000u & h;
            h &= ~t;
            h ^= t>>24;
        }
    }
    return h;
}

Elf32_Sym const *PackLinuxElf32::elf_lookup(char const *name) const
{
    if (hashtab && dynsym && dynstr) {
        unsigned const n_bucket = get_te32(&hashtab[0]);
        unsigned const *const buckets = &hashtab[2];
        unsigned const *const chains = &buckets[n_bucket];
        // Find the end of DT_HASH and DT_DYNSYM. Perhaps elf_find_table_size()
        // depends on too many valid input values?
        void const *l_hash = nullptr, *l_sym = nullptr;
        for (unsigned j = 0; j < DT_NUM; ++j) {
            void const *const ptr = dt_offsets[j] + (char const *)file_image.getVoidPtr();
            if (!l_hash && hashtab == ptr) {
                l_hash = dt_offsets[1+ j] + (char const *)file_image.getVoidPtr();
            }
            if (!l_sym && dynsym == ptr) {
                l_sym = dt_offsets[1+ j] + (char const *)file_image.getVoidPtr();
            }
            if (l_sym && l_hash)
                break;  // found both
            if (this->file_size == (off_t)dt_offsets[j])
                break;  // end sentinel
        }
        if (n_bucket) {
            void const *EOM = file_size + (char const *)file_image.getVoidPtr();
            unsigned const m = elf_hash(name) % n_bucket;
            unsigned n_visit = 0;
            unsigned si;
            if (l_hash <= &buckets[m])
                throwCantPack("bad DT_HASH %u", m);
            for (si= get_te32(&buckets[m]); si; si = get_te32(&chains[si])) {
                if (l_sym <= &dynsym[si])
                    throwCantPack("bad DT_HASH chain %d\n", si);
                char const *const p= get_dynsym_name(si, (unsigned)-1);
                if (p && 0==strcmp(name, p))
                    return &dynsym[si];
                if (l_sym <= &dynsym[n_visit++])
                    throwCantPack("circular DT_HASH chain %d\n", si);
                // Detect next si out-of-bounds
                if ((unsigned)((unsigned int const *)EOM - chains) <= si)
                    throwCantPack("bad DT_HASH chain %d\n", si);
            }
        }
    }
    if (gashtab && dynsym && dynstr) {
        unsigned const n_bucket = get_te32(&gashtab[0]);
        unsigned const symbias  = get_te32(&gashtab[1]);
        unsigned const n_bitmask = get_te32(&gashtab[2]);
        unsigned const gnu_shift = get_te32(&gashtab[3]);
        unsigned const *const bitmask = &gashtab[4];
        unsigned const *const buckets = &bitmask[n_bitmask];
        unsigned const *const hasharr = &buckets[n_bucket];
        if (31 < gnu_shift) {
            throwCantPack("bad gnu_shift %#x", gnu_shift);
        }
        if ((file_size + file_image) <= (void const *)hasharr) {
            throwCantPack("bad n_bucket %#x\n", n_bucket);
        }
        if (!n_bitmask
        || (unsigned)(file_size - ((char const *)bitmask - (char const *)(void const *)file_image))
                <= sizeof(unsigned)*n_bitmask ) {
            throwCantPack("bad n_bitmask %#x\n", n_bitmask);
        }
        if (n_bucket) { // -rust-musl can have "empty" hashtab
            unsigned const h = gnu_hash(name);
            unsigned const hbit1 = 037& h;
            unsigned const hbit2 = 037& (h>>gnu_shift);
            unsigned const w = get_te32(&bitmask[(n_bitmask -1) & (h>>5)]);

            if (1& (w>>hbit1) & (w>>hbit2)) {
                unsigned const hhead = get_te32(&buckets[h % n_bucket]);
                if (symnum_max <= hhead || (hhead && hhead < symbias)) {
                    throwCantPack("bad DT_GNU_HASH symnum_max{%#x} <= buckets[%d]{%#x} < symbias{%#x}\n",
                            symnum_max, h % n_bucket, hhead, symbias);
                }
                if (hhead) {
                    Elf32_Sym const *dsp = &dynsym[hhead];
                    unsigned const *hp = &hasharr[hhead - symbias];
                    unsigned k;
                    do {
                        if (gashend <= hp) {
                            throwCantPack("bad DT_GNU_HASH[%#x]  head=%u",
                                (unsigned)(hp - hasharr), hhead);
                        }
                        k = get_te32(hp);
                        if (0==((h ^ k)>>1)) {
                            unsigned const st_name = get_te32(&dsp->st_name);
                            char const *const p = get_str_name(st_name, (unsigned)-1);
                            if (0==strcmp(name, p)) {
                                return dsp;
                            }
                        }
                    } while (++dsp, ++hp, 0==(1u& k));
                }
            }
        }
    }
    // 2021-12-25  FIXME: Some Rust programs use
    //    (1==n_bucket && 0==buckets[0] && 1==n_bitmask && 0==bitmask[0])
    // to minimize space in DT_GNU_HASH. This causes the fancy lookup to fail.
    // Is a fallback to linear search assumed?
    // 2022-03-12  Some Rust programs have 0==n_bucket.
    return nullptr;

}

Elf64_Sym const *PackLinuxElf64::elf_lookup(char const *name) const
{
    if (hashtab && dynsym && dynstr) {
        unsigned const n_bucket = get_te32(&hashtab[0]);
        unsigned const *const buckets = &hashtab[2];
        unsigned const *const chains = &buckets[n_bucket];
        // Find the end of DT_HASH and DT_DYNSYM. Perhaps elf_find_table_size()
        // depends on too many valid input values?
        void const *l_hash = nullptr, *l_sym = nullptr;
        for (unsigned j = 0; j < DT_NUM; ++j) {
            void const *const ptr = dt_offsets[j] + (char const *)file_image.getVoidPtr();
            if (!l_hash && hashtab == ptr) {
                l_hash = dt_offsets[1+ j] + (char const *)file_image.getVoidPtr();
            }
            if (!l_sym && dynsym == ptr) {
                l_sym = dt_offsets[1+ j] + (char const *)file_image.getVoidPtr();
            }
            if (l_sym && l_hash)
                break;  // found both
            if (this->file_size == (off_t)dt_offsets[j])
                break;  //end
        }
        if (n_bucket) { // -rust-musl can have "empty" hashtab
            void const *const EOM = file_size + (char const *)file_image.getVoidPtr();
            unsigned const m = elf_hash(name) % n_bucket;
            unsigned n_visit = 0;
            unsigned si;
            if (l_hash <= &buckets[m])
                throwCantPack("bad DT_HASH %u", m);
            for (si= get_te32(&buckets[m]); si; si = get_te32(&chains[si])) {
                if (l_sym <= &dynsym[si])
                    throwCantPack("bad DT_HASH chain %d\n", si);
                char const *const p= get_dynsym_name(si, (unsigned)-1);
                if (p && 0==strcmp(name, p))
                    return &dynsym[si];
                if (l_sym <= &dynsym[n_visit++])
                    throwCantPack("circular DT_HASH chain %d\n", si);
                // Detect next si out-of-bounds
                if ((unsigned)((unsigned int const *)EOM - chains) <= si)
                    throwCantPack("bad DT_HASH chain %d\n", si);
            }
        }
    }
    if (gashtab && dynsym && dynstr) {
        unsigned const n_bucket = get_te32(&gashtab[0]);
        unsigned const symbias  = get_te32(&gashtab[1]);
        unsigned const n_bitmask = get_te32(&gashtab[2]);
        unsigned const gnu_shift = get_te32(&gashtab[3]);
        upx_uint64_t const *const bitmask = (upx_uint64_t const *)(void const *)&gashtab[4];
        unsigned     const *const buckets = (unsigned const *)&bitmask[n_bitmask];
        unsigned     const *const hasharr = &buckets[n_bucket];

        if (31 < gnu_shift) {
            throwCantPack("bad gnu_shift %#x", gnu_shift);
        }
        if ((file_size + file_image) <= (void const *)hasharr) {
            throwCantPack("bad n_bucket %#x\n", n_bucket);
        }
        if (!n_bitmask
        || (unsigned)(file_size - ((char const *)bitmask - (char const *)(void const *)file_image))
                <= sizeof(unsigned)*n_bitmask ) {
            throwCantPack("bad n_bitmask %#x\n", n_bitmask);
        }
        if (n_bucket) { // -rust-musl can have "empty" gashtab
            unsigned const h = gnu_hash(name);
            unsigned const hbit1 = 077& h;
            unsigned const hbit2 = 077& (h>>gnu_shift);
            upx_uint64_t const w = get_te64(&bitmask[(n_bitmask -1) & (h>>6)]);
            if (1& (w>>hbit1) & (w>>hbit2)) {
                unsigned hhead = get_te32(&buckets[h % n_bucket]);
                if (symnum_max <= hhead || (hhead && hhead < symbias)) {
                    throwCantPack("bad DT_GNU_HASH symnum_max{%#x} <= buckets[%d]{%#x} < symbias{%#x}\n",
                            symnum_max, h % n_bucket, hhead, symbias);
                }
                if (hhead) {
                    Elf64_Sym const *dsp = &dynsym[hhead];
                    unsigned const *hp = &hasharr[hhead - symbias];
                    unsigned k;
                    do {
                        if (gashend <= hp) {
                            throwCantPack("bad gnu_hash[%#tx]  head=%u",
                                hp - hasharr, hhead);
                        }
                        k = get_te32(hp);
                        if (0==((h ^ k)>>1)) {
                            unsigned const st_name = get_te32(&dsp->st_name);
                            char const *const p = get_str_name(st_name, (unsigned)-1);
                            if (0==strcmp(name, p)) {
                                return dsp;
                            }
                        }
                    } while (++dsp, ++hp, 0==(1u& k));
                }
            }
        }
    }
    // 2021-12-25  FIXME: Some Rust programs use
    //    (1==n_bucket && 0==buckets[0] && 1==n_bitmask && 0==bitmask[0])
    // to minimize space in DT_GNU_HASH. This causes the fancy lookup to fail.
    // Is a fallback to linear search assumed?
    // 2022-03-12  Some Rust programs have 0==n_bucket.
    return nullptr;

}

void PackLinuxElf32::unpack(OutputFile *fo)
{
    if (e_phoff != sizeof(Elf32_Ehdr)) {// Phdrs not contiguous with Ehdr
        throwCantUnpack("bad e_phoff");
    }
    unsigned const c_phnum = get_te16(&ehdri.e_phnum);
    unsigned u_phnum = 0;
    upx_uint32_t old_dtinit = 0;

    if (Elf32_Ehdr::ET_EXEC == get_te16(&ehdri.e_type)) {
// 40fddf17153ee3db73a04ff1bf288b91676138d6 2001-02-01 ph.version 11; b_info 12 bytes
// df9db96bd1c013c07da1d7ec740021d588ab2815 2001-01-17 ph.version 11; no b_info (==> 8 bytes)
        if (ph.version <= 11
        &&  get_te32(&ehdri.e_entry) < 0x401180
        &&  get_te16(&ehdri.e_machine)==Elf32_Ehdr::EM_386) {
            // old style, 8-byte b_info:
            // sizeof(b_info.sz_unc) + sizeof(b_info.sz_cpr);
            szb_info = 2*sizeof(unsigned);
        }
    }

    fi->seek(overlay_offset - sizeof(l_info), SEEK_SET);
    fi->readx(&linfo, sizeof(linfo));
    if (UPX_MAGIC_LE32 != get_le32(&linfo.l_magic)) {
        NE32 const *const lp = (NE32 const *)(void const *)&linfo;
        // Workaround for bug of extra linfo by some asl_pack2_Shdrs().
        if (0==lp[0] && 0==lp[1] && 0==lp[2]) { // looks like blank extra
            fi->readx(&linfo, sizeof(linfo));
            if (UPX_MAGIC_LE32 == get_le32(&linfo.l_magic)) {
                overlay_offset += sizeof(linfo);
            }
            else {
                throwCantUnpack("l_info corrupted");
            }
        }
        else {
            throwCantUnpack("l_info corrupted");
        }
    }
    lsize = get_te16(&linfo.l_lsize);
    p_info hbuf;  fi->readx(&hbuf, sizeof(hbuf));
    unsigned orig_file_size = get_te32(&hbuf.p_filesize);
    blocksize = get_te32(&hbuf.p_blocksize);
    if ((u32_t)file_size > orig_file_size || blocksize > orig_file_size
        || (orig_file_size >> 8) > (u32_t)file_size  // heuristic anti-fuzz
        ||      (blocksize >> 8) > (u32_t)file_size
        || !mem_size_valid(1, blocksize, OVERHEAD))
        throwCantUnpack("p_info corrupted");

    ibuf.alloc(blocksize + OVERHEAD);
    b_info bhdr; memset(&bhdr, 0, sizeof(bhdr));
    fi->readx(&bhdr, szb_info);
    ph.u_len = get_te32(&bhdr.sz_unc);
    ph.c_len = get_te32(&bhdr.sz_cpr);
    ph.set_method(bhdr.b_method, overlay_offset + sizeof(p_info));
    if (ph.c_len > (unsigned)file_size || ph.c_len == 0 || ph.u_len == 0
    ||  ph.u_len > orig_file_size)
        throwCantUnpack("b_info corrupted");
    ph.filter_cto = bhdr.b_cto8;
    prev_method = bhdr.b_method;  // FIXME if multiple de-compressors

    MemBuffer u(ph.u_len);
    Elf32_Ehdr *const ehdr = (Elf32_Ehdr *)&u[0];
    Elf32_Phdr const *phdr = nullptr;
    total_in = 0;
    total_out = 0;
    unsigned c_adler = upx_adler32(nullptr, 0);
    unsigned u_adler = upx_adler32(nullptr, 0);

    unsigned is_shlib = 0;
    loader_offset = 0;
    MemBuffer o_elfhdrs;
    Elf32_Phdr const *const dynhdr = elf_find_ptype(Elf32_Phdr::PT_DYNAMIC, phdri, c_phnum);
    // dynseg was set by PackLinuxElf32help1
    if (dynhdr && !(Elf32_Dyn::DF_1_PIE & elf_unsigned_dynamic(Elf32_Dyn::DT_FLAGS_1))) {
        // Packed shlib? (ET_DYN without -fPIE)
        is_shlib = 1;
        xct_off = overlay_offset - sizeof(l_info);
        u_phnum = get_te16(&ehdri.e_phnum);
        o_elfhdrs.alloc(sz_elf_hdrs);
        un_shlib_1(fo, o_elfhdrs, c_adler, u_adler, orig_file_size);
        *ehdr = ehdri;
    }
    else { // main executable
        // Uncompress Ehdr and Phdrs: info for control of unpacking
        if (ibuf.getSize() < ph.c_len)
            throwCompressedDataViolation();
        fi->readx(ibuf, ph.c_len);
        if (ph.u_len < sizeof(*ehdr))
            throwCantUnpack("ElfXX_Ehdr corrupted");
        decompress(ibuf, (upx_byte *)ehdr, false);
        if (ehdr->e_type   !=ehdri.e_type
        ||  ehdr->e_machine!=ehdri.e_machine
        ||  ehdr->e_version!=ehdri.e_version
            // less strict for EM_PPC to workaround earlier bug
        ||  !( ehdr->e_flags==ehdri.e_flags
            || Elf32_Ehdr::EM_PPC == get_te16(&ehdri.e_machine))
        ||  ehdr->e_ehsize !=ehdri.e_ehsize
            // check EI_MAG[0-3], EI_CLASS, EI_DATA, EI_VERSION
        ||  memcmp(ehdr->e_ident, ehdri.e_ident, Elf32_Ehdr::EI_OSABI)) {
            throwCantUnpack("ElfXX_Ehdr corrupted");
        }
        // Rewind: prepare for data phase
        fi->seek(- (off_t) (szb_info + ph.c_len), SEEK_CUR);

        u_phnum = get_te16(&ehdr->e_phnum);
        if ((umin(MAX_ELF_HDR_32, ph.u_len) - sizeof(Elf32_Ehdr))/sizeof(Elf32_Phdr) < u_phnum) {
            throwCantUnpack("bad compressed e_phnum");
        }
        o_elfhdrs.alloc(sizeof(Elf32_Ehdr) + u_phnum * sizeof(Elf32_Phdr));
        memcpy(o_elfhdrs, ehdr, o_elfhdrs.getSize());

        // Decompress each PT_LOAD.
        bool first_PF_X = true;
        phdr = (Elf32_Phdr *) (void *) (1+ ehdr);  // uncompressed
        for (unsigned j=0; j < u_phnum; ++phdr, ++j) {
            if (is_LOAD(phdr)) {
                unsigned const filesz = get_te32(&phdr->p_filesz);
                unsigned const offset = get_te32(&phdr->p_offset);
                if (fo) {
                    fo->seek(offset, SEEK_SET);
                    if (total_out < offset) {
                        total_out = offset;  // FIXME: can it be re-write?
                    }
                }
                if (Elf32_Phdr::PF_X & get_te32(&phdr->p_flags)) {
                    unpackExtent(filesz, fo,
                        c_adler, u_adler, first_PF_X);
                    first_PF_X = false;
                }
                else {
                    unpackExtent(filesz, fo,
                        c_adler, u_adler, false);
                }
            }
        }
    }

    upx_uint32_t const e_entry = get_te32(&ehdri.e_entry);
    unsigned off_entry = 0;
    phdr = phdri;
    load_va = 0;
    for (unsigned j=0; j < c_phnum; ++j, ++phdr) {
        if (is_LOAD(phdr)) {
            upx_uint32_t offset = get_te32(&phdr->p_offset);
            upx_uint32_t vaddr  = get_te32(&phdr->p_vaddr);
            upx_uint32_t filesz = get_te32(&phdr->p_filesz);
            if (!load_va) {
                load_va = vaddr;
            }
            if ((e_entry - vaddr) < filesz) {
                off_entry = (e_entry - vaddr) + offset;
                break;
            }
        }
    }
    unsigned d_info[6];
    unsigned sz_d_info = sizeof(d_info);
    if (!is_shlib) {
        if (get_te32(&phdri[0].p_flags) & Elf32_Phdr::PF_X) {
            // Old style, such as upx-3.91 thru upx-3.95
            switch (this->e_machine) {
                default: {
                    char msg[40]; snprintf(msg, sizeof(msg),
                        "Unknown architecture %d", this->e_machine);
                    throwCantUnpack(msg);
                }; break;
                case Elf32_Ehdr::EM_MIPS:sz_d_info = 1 * sizeof(unsigned); break;
                case Elf32_Ehdr::EM_ARM: sz_d_info = 4 * sizeof(unsigned); break;
                case Elf32_Ehdr::EM_PPC: sz_d_info = 3 * sizeof(unsigned); break;
                case Elf32_Ehdr::EM_386: sz_d_info = 2 * sizeof(unsigned); break;
            }
        }
        loader_offset = off_entry - sz_d_info;
    }

    if (0x1000==get_te32(&phdri[0].p_filesz)  // detect C_BASE style
    &&  0==get_te32(&phdri[1].p_offset)
    &&  0==get_te32(&phdri[0].p_offset)
    &&     get_te32(&phdri[1].p_filesz) == get_te32(&phdri[1].p_memsz)) {
        fi->seek(up4(get_te32(&phdri[1].p_memsz)), SEEK_SET);  // past the loader
    }
    else if (is_shlib
    ||  (off_entry + up4(lsize) + ph.getPackHeaderSize() + sizeof(overlay_offset))
            < up4(file_size)) {
        // Loader is not at end; skip past it.
        if (loader_offset) {
            fi->seek(loader_offset, SEEK_SET);
        }
        else {
            funpad4(fi);  // MATCH01
        }
        fi->readx(d_info, sz_d_info);
        if (is_shlib && 0==old_dtinit) {
            old_dtinit = get_te32(&d_info[2 + (0==d_info[0])]);
            is_asl = 1u& get_te32(&d_info[0 + (0==d_info[0])]);
        }
        fi->seek(lsize - sz_d_info, SEEK_CUR);
    }

    // The gaps between PT_LOAD and after last PT_LOAD
    phdr = (Elf32_Phdr const *)(1+ (Elf32_Ehdr const *)(void const *)o_elfhdrs);
    upx_uint32_t hi_offset(0);
    for (unsigned j = 0; j < u_phnum; ++j) {
        if (is_LOAD(&phdr[j])
        &&  hi_offset < get_te32(&phdr[j].p_offset))
            hi_offset = get_te32(&phdr[j].p_offset);
    }
    for (unsigned j = 0; j < u_phnum; ++j) {
        unsigned const size = find_LOAD_gap(phdr, j, u_phnum);
        if (size) {
            unsigned const where = get_te32(&phdr[j].p_offset) +
                                   get_te32(&phdr[j].p_filesz);
            if (fo)
                fo->seek(where, SEEK_SET);
            { // Recover from some piracy [also serves as error tolerance :-) ]
              // Getting past the loader is problematic, due to unintended
              // variances between released versions:
              //   l_info.l_lsize might be rounded up by 8 instead of by 4, and
              //   sz_d_info might have changed.
                b_info b_peek, *bp = &b_peek;
                fi->readx(bp, sizeof(b_peek));
                upx_off_t pos = fi->seek(-(off_t)sizeof(b_peek), SEEK_CUR);
                unsigned sz_unc = get_te32(&bp->sz_unc);
                unsigned sz_cpr = get_te32(&bp->sz_cpr);
                unsigned word3  = get_te32(&bp->b_method);
                unsigned method = bp->b_method;
                unsigned ftid = bp->b_ftid;
                unsigned cto8 = bp->b_cto8;
                if (!( ((sz_cpr == sz_unc) && (0 == word3) && (size == sz_unc)) // incompressible literal
                    || ((sz_cpr <  sz_unc)
                        && (method == prev_method || M_NRV2B_LE32 == prev_method)
                        && (0 == ftid) && (0 == cto8)) )
                ) {
                    opt->info_mode++;
                    infoWarning("bad b_info at %#zx", (size_t)pos);
                    unsigned const N_PEEK(16 * sizeof(int)), H_PEEK(N_PEEK >> 1);
                    unsigned char peek_arr[N_PEEK];
                    fi->seek(pos - H_PEEK, SEEK_SET);
                    fi->readx(peek_arr, sizeof(peek_arr));
                    fi->seek(pos, SEEK_SET);
                    bool const is_be = ELFDATA2MSB == ehdri.e_ident[EI_DATA];
                    if (is_be) {
                        // Does the right thing for sz_unc and sz_cpr,
                        // but swaps b_method and b_extra.  Need find_be32() :-)
                        for (unsigned k = 0; k < N_PEEK; k += sizeof(int)) {
                            set_le32(&peek_arr[k], get_be32(&peek_arr[k]));
                        }
                    }
                    int boff = find_le32(peek_arr, sizeof(peek_arr), size);
                    if (boff < 0 || sizeof(peek_arr) < (sizeof(*bp) + boff)) {
                        throwCantUnpack("b_info corrupted");
                    }
                    bp = (b_info *)(void *)&peek_arr[boff];

                    sz_unc = get_le32(&bp->sz_unc);
                    sz_cpr = get_le32(&bp->sz_cpr);
                    word3  = get_le32(&bp->b_method);
                    ftid = bp->b_ftid;
                    cto8 = bp->b_cto8;
                    if (0 <= boff  // found
                    && ( ((sz_cpr == sz_unc) && (0 == word3) && (size == sz_unc)) // incompressible literal
                      || ((sz_cpr <  sz_unc) && (0 == ftid) && (0 == cto8)
                          && ((is_be ? bp->b_extra : bp->b_method) == prev_method)) )
                    ) {
                        pos -= H_PEEK;
                        pos += boff;
                        infoWarning("... recovery at %#zx", (size_t)pos);
                        fi->seek(pos, SEEK_SET);
                    }
                    opt->info_mode--;
                }
            }
            unpackExtent(size, fo,
                c_adler, u_adler, false,
                (hi_offset != get_te32(&phdr[j].p_offset)));
        }
    }

    // check for end-of-file
    fi->readx(&bhdr, szb_info);
    ph.set_method(bhdr.b_method, ~0u);
    unsigned const sz_unc = ph.u_len = get_te32(&bhdr.sz_unc);

    if (sz_unc == 0) { // uncompressed size 0 -> EOF
        // note: magic is always stored le32
        unsigned const sz_cpr = get_le32(&bhdr.sz_cpr);
        if (sz_cpr != UPX_MAGIC_LE32)  // sz_cpr must be h->magic
            throwCompressedDataViolation();
    }
    else { // extra bytes after end?
        throwCompressedDataViolation();
    }

    if (is_shlib) {
        un_DT_INIT(old_dtinit, (Elf32_Phdr *)(1+ (Elf32_Ehdr *)(void *)o_elfhdrs), dynhdr, fo);
    }

    // update header with totals
    ph.c_len = total_in;
    ph.u_len = total_out;

    // all bytes must be written
    if (fo && total_out != orig_file_size)
        throwEOFException();

    // finally test the checksums
    if (ph.c_adler != c_adler || ph.u_adler != u_adler)
        throwChecksumError();
}

void PackLinuxElf::unpack(OutputFile * /*fo*/)
{
    throwCantUnpack("internal error");
}

// NOLINTEND(clang-analyzer-deadcode.DeadStores)
// NOLINTEND(clang-analyzer-core.CallAndMessage)

/* vim:set ts=4 sw=4 et: */