tuning

amd_bxx.S amd_d_nrv2b.S amd_d_nrv2e.S l_lx_elf64amd.S

committer: jreiser <jreiser> 1131939243 +0000

src/stub/amd_d_nrv2e.S

@@ -29,62 +29,6 @@
    <jreiser@users.sourceforge.net>
 */
 
-#include "amd_regs.h"
-
-M_NRV2B_LE32=2  # ../conf.h
-M_NRV2E_LE32=8
-
-#define ALIGN(n)  .align n
-
-/* Returns 0 on success; non-zero on failure. */
-decompress:  # (uchar const *src, size_t lsrc, uchar *dst, size_t &ldst, uint method)
-
-/* Arguments according to calling convention */
-#define src  %arg1
-#define lsrc %arg2
-#define dst  %arg3
-#define ldst %arg4  /* Out: actually a reference: &len_dst */
-#define meth %arg5l
-
-/* Working registers */
-#define off  %eax  /* XXX: 2GB */
-#define len  %ecx  /* XXX: 2GB */
-#define lenq %rcx
-#define bits %ebx
-#define disp %rbp
-
-	push %rbp; push %rbx
-	push ldst
-	push dst
-        addq src,lsrc; push lsrc  # &input_eof
-
-	movq src,%rsi  # hardware src for movsb, lodsb
-	movq dst,%rdi  # hardware dst for movsb
-	xorl bits,bits  # empty; force refill
-	xorl len,len  # create loop invariant
-	orq $~0,disp  # -1: initial displacement
-	call setup_rdx
-ra_setup_rdx:
-
-/* jump on next bit {0,1} with prediction {y==>likely, n==>unlikely} */
-/* Prediction omitted for now. */
-#define jnextb0n jnextb0y
-#define jnextb0y addl bits,bits; jnz 0f; call refill; 0: jnc
-#define jnextb1n jnextb1y
-#define jnextb1y addl bits,bits; jnz 0f; call refill; 0: jc
-
-/* rotate next bit into bottom bit of reg */
-#define getnextb(reg) addl bits,bits; jnz 0f; call refill; 0: adcl reg,reg
-
-	ALIGN(1<<3)
-getbit:
-	addl bits,bits; jz refill  # Carry= next bit
-	rep; ret
-refill:
-	movl (%rsi),bits; subq $-4,%rsi  # next 32 bits; set Carry
-	adcl bits,bits  # LSB= 1 (CarryIn); CarryOut= next bit
-	rep; ret
-
 	ALIGN(1<<3)
 lit_n2e:
 	movb (%rsi),%al; addq $1,%rsi
@@ -101,11 +45,11 @@ getoff_n2e:
         getnextb(off)
         jnextb0n off_n2e
 
-	/*xorl len,len  # len= 0*/
-	subl $3,off; jc offprev_n2e
-	shll $8,off
-	lodsb  # requires off===%eax
-	xorl $~0,off; jz eof_n2e
+	movzbl (%rsi),%edx  #; xorl len,len  # len= 0
+	subl $ 3,off; jc offprev_n2e
+	shll $ 8,off
+	orl %edx,off; incq %rsi
+	xorl $~0,off; jz eof
 	sarl off  # Carry= original low bit
 	movslq off,disp  # XXX: 2GB
 	jc lenlast_n2e
@@ -127,62 +71,6 @@ lenlast_n2e:
 gotlen_n2e:
 	cmpq $-0x500,disp
         adcl $2,len  # len += 2+ (disp < -0x500);
-	push %rsi
-	leaq (%rdi,disp),%rsi
-	rep; movsb  # len
-	pop %rsi
+	call copy
 bot_n2e:  # In: 0==len
-	prefetch 0x7f(%rdi)  # em64t has no prefetchw
         jmp top_n2e
-
-
-	ALIGN(1<<3)
-lit_n2b:
-	movb (%rsi),%al; addq $1,%rsi
-	movb %al,(%rdi); addq $1,%rdi
-top_n2b:
-	jnextb1y lit_n2b
-	lea 1(lenq),off  # [len= 0] off= 1
-offmore_n2b:
-	getnextb(off)
-	jnextb0n offmore_n2b
-
-	subl $3,off; jc len_n2b  # use previous offset
-	shll $8,off
-	lodsb  # requires off===%eax
-	xorl $~0,off; jz eof_n2b
-	movslq off,disp  # XXX: 2GB
-len_n2b:
-	lea 1(lenq),off  # [len= 0] off= 1
-	getnextb(len); getnextb(len)  # two bits; cc set on result
-	jnz gotlen_n2b  # raw 1,2,3 ==> 2,3,4
-	movl off,len  # len= 1, the msb
-	addl $3-1,off  # raw 2.. ==> 5..
-lenmore_n2b:
-	getnextb(len)
-	jnextb0n lenmore_n2b
-gotlen_n2b:
-	cmpq $-0xd00,disp
-	adcl off,len  # len += off + (disp < -0xd00)
-	push %rsi
-	leaq (%rdi,disp),%rsi
-	rep; movsb  # len
-	pop %rsi
-bot_n2b:  # In: 0==len
-	prefetch 0x7f(%rdi)  # em64t has no prefetchw
-        jmp top_n2b
-
-setup_rdx:
-	cld
-	pop %rdx; addq $ getbit - ra_setup_rdx,%rdx  # %rdx= &getbit
-	cmpl $ M_NRV2E_LE32,meth; je bot_n2e
-	cmpl $ M_NRV2B_LE32,meth; je bot_n2b
-eof_n2e:
-eof_n2b:
-	pop %rcx  # &input_eof
-	movq %rsi,%rax; subq %rcx,%rax  # src -= eof;  // return 0: good; else: bad
-	pop %rdx;       subq %rdx,%rdi  # dst -= original dst
-	pop %rcx;            movl %edi,(%rcx)  # actual length used at dst  XXX: 4GB
-	pop %rbx; pop %rbp
-	ret
-