stub: fix i386 syntax for current gas versions, and add some explicit

alignments at the end of the file so that we do not depend on gas defaults.

src/stub/src/arch/i386/macros.S

@@ -472,7 +472,7 @@ decompr0:
 
 section LXJCC010
 lxunf2:  // have seen 0x80..0x8f of possible recoded 6-byte Jcc <d32>
-        movzwl ebp, [edi]  // 2 bytes, zero-extended
+        movzx ebp, word ptr [edi]  // 2 bytes, zero-extended
 
 section LXMRU045 // 0!=n_mru
         sub ebp, __cto8_0f
@@ -644,7 +644,7 @@ section LXMRU040 // 0!=n_mru
 
 section LXUNF030
 lxctloop:
-        movzxw eax, [edi]  // 2 bytes, zero extended
+        movzx eax, word ptr [edi]  // 2 bytes, zero extended
         add edi, 1
 section LXJCC000
         cmp al, 0x80  // lo of Jcc <d32>

src/stub/src/i086-dos16.exe.S

@@ -112,7 +112,7 @@ addaxds:
                 movsw
                 cld
 section         LONGSUB
-                subb    [cs:si + addaxds + 4], 0x10
+                subb    cs:[si + addaxds + 4], 0x10
 section         SHORTSUB
                 .byte   0x2e, 0x80, 0x6c, addaxds + 4, 0x10
 section         JNCDOCOPY
@@ -177,7 +177,7 @@ reloc_0:
                 xor     ax, ax
 reloc_1:
                 add     di, ax
-                add     [es:di], bp
+                add     es:[di], bp
 reloc_2:
                 lodsb
                 dec     ax
@@ -188,9 +188,9 @@ section         EXEREL9A
                 inc     di
 reloc_4:
                 inc     di
-                cmpb    [es:di], 0x9a
+                cmpb    es:[di], 0x9a
                 jne     reloc_4
-                cmp     [es:di+3], dx
+                cmp     es:[di+3], dx
                 ja      reloc_4
                 mov     al, 3
                 jmps    reloc_1

src/stub/src/i386-bsd.elf-fold.S

@@ -202,5 +202,6 @@ L32:
         pop edi  // &auxtab
         ret
 
+.balign 4,0
 // vi:ts=8:et:nowrap
 

src/stub/src/i386-bsd.elf.execve-fold.S

@@ -55,4 +55,6 @@ fold_begin:  // enter: %ebx= &Elf32_Ehdr of this program
 .extern upx_main
                 call    upx_main            // Call the UPX main function
                 hlt                         // Crash if somehow upx_main does return
+
+.balign 4,0
 // vi:ts=8:et:nowrap

src/stub/src/i386-darwin.dylib-entry.S

@@ -209,8 +209,8 @@ dy_uncpr:
         add esp, (5+1)*4  // (5+1) args to decompress
 
         pop edx; pop eax  # edx= old dst; eax= old &b_info
-        movzbl ecx,[1+ b_method + eax]; jecxz dy_uncpr; push ecx  # ftid
+        movzx ecx, byte ptr [1+ b_method + eax]; jecxz dy_uncpr; push ecx  # ftid
-        movzbl ecx,[2+ b_method + eax];                 push ecx  # cto8
+        movzx ecx, byte ptr [2+ b_method + eax];                 push ecx  # cto8
         push [sz_unc + eax]
         push edx    # dst
         lea eax,[2+ ebp]; call eax  # f_unfilter(dst, dstlen, cto8, ftid)

src/stub/src/i386-darwin.macho-fold.S

@@ -135,6 +135,7 @@ read: .globl read
 0:
         ret
 
+.balign 2,144
 /*
 vi:ts=8:et:nowrap
 */

src/stub/src/i386-linux.elf-fold.S

@@ -257,5 +257,6 @@ mmap: .globl mmap
         pop ebx
         ret
 
+.balign 4,0
 // vi:ts=8:et:nowrap
 

src/stub/src/i386-linux.elf.execve-fold.S

@@ -68,4 +68,5 @@ mmap: .globl mmap
         pop ebx
         ret
 
+.balign 4,0
 // vi:ts=8:et:nowrap

src/stub/src/i386-linux.elf.interp-fold.S

@@ -125,5 +125,6 @@ mmap: .globl mmap
         pop ebx
         ret
 
+.balign 4,0
 // vi:ts=8:et:nowrap
 

src/stub/src/i386-linux.elf.shell-fold.S

@@ -172,5 +172,6 @@ mmap: .globl mmap
         ret
 
 
+.balign 4,0
 // vi:ts=8:et:nowrap
 

src/stub/src/i386-linux.kernel.vmlinuz.S

@@ -259,7 +259,7 @@ cmd_exp = 80  // allow kernel to expand command line "in place" by this much
 cl_nbp202:  // use < 2.02 method.  in: ebp= real_mode_ptr
                 mov esi,ebp  // absolute worst case default
                 cmpw [ebp+ BP_cmd_line_magic],0xA33F; jne cl_move  // nothing at all
-                movzwl esi,[ebp+ BP_cmd_line_offset]
+                movzx esi, word ptr [ebp+ BP_cmd_line_offset]
                 add esi,ebp
                 jmp cl_move
 cl_bp202:  // try >= 2.02 method

src/stub/src/i386-linux.shlib-init.S

@@ -201,8 +201,8 @@ p_mprot= -5*4
         sub eax,ebx  // restore
         add edi,ebx
 
-        movzbl ecx,[esi+b_method-4+1]; push ecx  // ftid
+        movzx ecx, byte ptr [esi+b_method-4+1]; push ecx  // ftid
-        movzbl ecx,[esi+b_method-4+2]; push ecx  // cto8
+        movzx ecx, byte ptr [esi+b_method-4+2]; push ecx  // cto8
         push eax; mov ecx,esp  // dstlen  also for unfilter  step 7
         push edi  // dst                 param for unfilter  step 7
 p_unflt= -9*4
@@ -248,12 +248,12 @@ supervise:
         pop ecx  // &dstlen
         pop edx  // discard method,filter,cto,junk
         add eax,[ecx]  // dst += dstlen
-        mov [dword ptr eax],0xc36180cd  // "int 0x80; popa; ret"
+        mov dword ptr [eax],0xc36180cd  // "int 0x80; popa; ret"
         mov [esp + p_hatch - o_uncpr],eax  // hatch at end of .text
 //o_uncpr
         pop eax  // &decompress
 //p_unflt
-        cmp [word ptr esp+3*4],0; je 0f  // 0==ftid ==> no filter
+        cmp word ptr [esp+3*4],0; je 0f  // 0==ftid ==> no filter
         add eax,2; call eax  // unfilter {i386 f_unf==(2+f_unc)}
 0:
         add esp,4*4

src/stub/src/i386-openbsd.elf-fold.S

@@ -202,5 +202,6 @@ L32:
         pop edi  // &auxtab
         ret
 
+.balign 4,0
 // vi:ts=8:et:nowrap
 

src/stub/tmp/i386-netbsd.elf-fold.map

@@ -11,8 +11,7 @@ TARGET(elf32-i386)
 
 .text           0x0000000000c01080      0x68c
  *(.text)
- .text          0x0000000000c01080       0xc2 tmp/i386-bsd.elf-fold.o
+ .text          0x0000000000c01080       0xc4 tmp/i386-bsd.elf-fold.o
- *fill*         0x0000000000c01142        0x2 00
  .text          0x0000000000c01144      0x558 tmp/i386-bsd.elf-main.o
                 0x0000000000c01564                upx_main
  .text          0x0000000000c0169c       0x6d tmp/i386-bsd.syscall.o