[deliverable/binutils-gdb.git] / zlib / contrib / masmx86 / inffas32.asm

;/* inffas32.asm is a hand tuned assembler version of inffast.c -- fast decoding\r
; *\r
; * inffas32.asm is derivated from inffas86.c, with translation of assembly code\r
; *\r
; * Copyright (C) 1995-2003 Mark Adler\r
; * For conditions of distribution and use, see copyright notice in zlib.h\r
; *\r
; * Copyright (C) 2003 Chris Anderson <christop@charm.net>\r
; * Please use the copyright conditions above.\r
; *\r
; * Mar-13-2003 -- Most of this is derived from inffast.S which is derived from\r
; * the gcc -S output of zlib-1.2.0/inffast.c.  Zlib-1.2.0 is in beta release at\r
; * the moment.  I have successfully compiled and tested this code with gcc2.96,\r
; * gcc3.2, icc5.0, msvc6.0.  It is very close to the speed of inffast.S\r
; * compiled with gcc -DNO_MMX, but inffast.S is still faster on the P3 with MMX\r
; * enabled.  I will attempt to merge the MMX code into this version.  Newer\r
; * versions of this and inffast.S can be found at\r
; * http://www.eetbeetee.com/zlib/ and http://www.charm.net/~christop/zlib/\r
; *\r
; * 2005 : modification by Gilles Vollant\r
; */\r
; For Visual C++ 4.x and higher and ML 6.x and higher\r
;   ml.exe is in directory \MASM611C of Win95 DDK\r
;   ml.exe is also distributed in http://www.masm32.com/masmdl.htm\r
;    and in VC++2003 toolkit at http://msdn.microsoft.com/visualc/vctoolkit2003/\r
;\r
;\r
;   compile with command line option\r
;   ml  /coff /Zi /c /Flinffas32.lst inffas32.asm\r
\r
;   if you define NO_GZIP (see inflate.h), compile with\r
;   ml  /coff /Zi /c /Flinffas32.lst /DNO_GUNZIP inffas32.asm\r
\r
\r
; zlib122sup is 0 fort zlib 1.2.2.1 and lower\r
; zlib122sup is 8 fort zlib 1.2.2.2 and more (with addition of dmax and head\r
;        in inflate_state in inflate.h)\r
zlib1222sup      equ    8\r
\r
\r
IFDEF GUNZIP\r
  INFLATE_MODE_TYPE    equ 11\r
  INFLATE_MODE_BAD     equ 26\r
ELSE\r
  IFNDEF NO_GUNZIP\r
    INFLATE_MODE_TYPE    equ 11\r
    INFLATE_MODE_BAD     equ 26\r
  ELSE\r
    INFLATE_MODE_TYPE    equ 3\r
    INFLATE_MODE_BAD     equ 17\r
  ENDIF\r
ENDIF\r
\r
\r
; 75 "inffast.S"\r
;FILE "inffast.S"\r
\r
;;;GLOBAL _inflate_fast\r
\r
;;;SECTION .text\r
\r
\r
\r
	.586p\r
	.mmx\r
\r
	name	inflate_fast_x86\r
	.MODEL	FLAT\r
\r
_DATA			segment\r
inflate_fast_use_mmx:\r
	dd	1\r
\r
\r
_TEXT			segment\r
\r
\r
\r
ALIGN 4\r
	db	'Fast decoding Code from Chris Anderson'\r
	db	0\r
\r
ALIGN 4\r
invalid_literal_length_code_msg:\r
	db	'invalid literal/length code'\r
	db	0\r
\r
ALIGN 4\r
invalid_distance_code_msg:\r
	db	'invalid distance code'\r
	db	0\r
\r
ALIGN 4\r
invalid_distance_too_far_msg:\r
	db	'invalid distance too far back'\r
	db	0\r
\r
\r
ALIGN 4\r
inflate_fast_mask:\r
dd	0\r
dd	1\r
dd	3\r
dd	7\r
dd	15\r
dd	31\r
dd	63\r
dd	127\r
dd	255\r
dd	511\r
dd	1023\r
dd	2047\r
dd	4095\r
dd	8191\r
dd	16383\r
dd	32767\r
dd	65535\r
dd	131071\r
dd	262143\r
dd	524287\r
dd	1048575\r
dd	2097151\r
dd	4194303\r
dd	8388607\r
dd	16777215\r
dd	33554431\r
dd	67108863\r
dd	134217727\r
dd	268435455\r
dd	536870911\r
dd	1073741823\r
dd	2147483647\r
dd	4294967295\r
\r
\r
mode_state	 equ	0	;/* state->mode	*/\r
wsize_state	 equ	(32+zlib1222sup)	;/* state->wsize */\r
write_state	 equ	(36+4+zlib1222sup)	;/* state->write */\r
window_state	 equ	(40+4+zlib1222sup)	;/* state->window */\r
hold_state	 equ	(44+4+zlib1222sup)	;/* state->hold	*/\r
bits_state	 equ	(48+4+zlib1222sup)	;/* state->bits	*/\r
lencode_state	 equ	(64+4+zlib1222sup)	;/* state->lencode */\r
distcode_state	 equ	(68+4+zlib1222sup)	;/* state->distcode */\r
lenbits_state	 equ	(72+4+zlib1222sup)	;/* state->lenbits */\r
distbits_state	 equ	(76+4+zlib1222sup)	;/* state->distbits */\r
\r
\r
;;SECTION .text\r
; 205 "inffast.S"\r
;GLOBAL	inflate_fast_use_mmx\r
\r
;SECTION .data\r
\r
\r
; GLOBAL inflate_fast_use_mmx:object\r
;.size inflate_fast_use_mmx, 4\r
; 226 "inffast.S"\r
;SECTION .text\r
\r
ALIGN 4\r
_inflate_fast proc near\r
.FPO (16, 4, 0, 0, 1, 0)\r
	push  edi\r
	push  esi\r
	push  ebp\r
	push  ebx\r
	pushfd\r
	sub  esp,64\r
	cld\r
\r
\r
\r
\r
	mov  esi, [esp+88]\r
	mov  edi, [esi+28]\r
\r
\r
\r
\r
\r
\r
\r
	mov  edx, [esi+4]\r
	mov  eax, [esi+0]\r
\r
	add  edx,eax\r
	sub  edx,11\r
\r
	mov  [esp+44],eax\r
	mov  [esp+20],edx\r
\r
	mov  ebp, [esp+92]\r
	mov  ecx, [esi+16]\r
	mov  ebx, [esi+12]\r
\r
	sub  ebp,ecx\r
	neg  ebp\r
	add  ebp,ebx\r
\r
	sub  ecx,257\r
	add  ecx,ebx\r
\r
	mov  [esp+60],ebx\r
	mov  [esp+40],ebp\r
	mov  [esp+16],ecx\r
; 285 "inffast.S"\r
	mov  eax, [edi+lencode_state]\r
	mov  ecx, [edi+distcode_state]\r
\r
	mov  [esp+8],eax\r
	mov  [esp+12],ecx\r
\r
	mov  eax,1\r
	mov  ecx, [edi+lenbits_state]\r
	shl  eax,cl\r
	dec  eax\r
	mov  [esp+0],eax\r
\r
	mov  eax,1\r
	mov  ecx, [edi+distbits_state]\r
	shl  eax,cl\r
	dec  eax\r
	mov  [esp+4],eax\r
\r
	mov  eax, [edi+wsize_state]\r
	mov  ecx, [edi+write_state]\r
	mov  edx, [edi+window_state]\r
\r
	mov  [esp+52],eax\r
	mov  [esp+48],ecx\r
	mov  [esp+56],edx\r
\r
	mov  ebp, [edi+hold_state]\r
	mov  ebx, [edi+bits_state]\r
; 321 "inffast.S"\r
	mov  esi, [esp+44]\r
	mov  ecx, [esp+20]\r
	cmp  ecx,esi\r
	ja   L_align_long\r
\r
	add  ecx,11\r
	sub  ecx,esi\r
	mov  eax,12\r
	sub  eax,ecx\r
	lea  edi, [esp+28]\r
	rep movsb\r
	mov  ecx,eax\r
	xor  eax,eax\r
	rep stosb\r
	lea  esi, [esp+28]\r
	mov  [esp+20],esi\r
	jmp  L_is_aligned\r
\r
\r
L_align_long:\r
	test  esi,3\r
	jz   L_is_aligned\r
	xor  eax,eax\r
	mov  al, [esi]\r
	inc  esi\r
	mov  ecx,ebx\r
	add  ebx,8\r
	shl  eax,cl\r
	or  ebp,eax\r
	jmp L_align_long\r
\r
L_is_aligned:\r
	mov  edi, [esp+60]\r
; 366 "inffast.S"\r
L_check_mmx:\r
	cmp  dword ptr [inflate_fast_use_mmx],2\r
	je   L_init_mmx\r
	ja   L_do_loop\r
\r
	push  eax\r
	push  ebx\r
	push  ecx\r
	push  edx\r
	pushfd\r
	mov  eax, [esp]\r
	xor  dword ptr [esp],0200000h\r
\r
\r
\r
\r
	popfd\r
	pushfd\r
	pop  edx\r
	xor  edx,eax\r
	jz   L_dont_use_mmx\r
	xor  eax,eax\r
	cpuid\r
	cmp  ebx,0756e6547h\r
	jne  L_dont_use_mmx\r
	cmp  ecx,06c65746eh\r
	jne  L_dont_use_mmx\r
	cmp  edx,049656e69h\r
	jne  L_dont_use_mmx\r
	mov  eax,1\r
	cpuid\r
	shr  eax,8\r
	and  eax,15\r
	cmp  eax,6\r
	jne  L_dont_use_mmx\r
	test  edx,0800000h\r
	jnz  L_use_mmx\r
	jmp  L_dont_use_mmx\r
L_use_mmx:\r
	mov  dword ptr [inflate_fast_use_mmx],2\r
	jmp  L_check_mmx_pop\r
L_dont_use_mmx:\r
	mov  dword ptr [inflate_fast_use_mmx],3\r
L_check_mmx_pop:\r
	pop  edx\r
	pop  ecx\r
	pop  ebx\r
	pop  eax\r
	jmp  L_check_mmx\r
; 426 "inffast.S"\r
ALIGN 4\r
L_do_loop:\r
; 437 "inffast.S"\r
	cmp  bl,15\r
	ja   L_get_length_code\r
\r
	xor  eax,eax\r
	lodsw\r
	mov  cl,bl\r
	add  bl,16\r
	shl  eax,cl\r
	or  ebp,eax\r
\r
L_get_length_code:\r
	mov  edx, [esp+0]\r
	mov  ecx, [esp+8]\r
	and  edx,ebp\r
	mov  eax, [ecx+edx*4]\r
\r
L_dolen:\r
\r
\r
\r
\r
\r
\r
	mov  cl,ah\r
	sub  bl,ah\r
	shr  ebp,cl\r
\r
\r
\r
\r
\r
\r
	test  al,al\r
	jnz   L_test_for_length_base\r
\r
	shr  eax,16\r
	stosb\r
\r
L_while_test:\r
\r
\r
	cmp  [esp+16],edi\r
	jbe  L_break_loop\r
\r
	cmp  [esp+20],esi\r
	ja   L_do_loop\r
	jmp  L_break_loop\r
\r
L_test_for_length_base:\r
; 502 "inffast.S"\r
	mov  edx,eax\r
	shr  edx,16\r
	mov  cl,al\r
\r
	test  al,16\r
	jz   L_test_for_second_level_length\r
	and  cl,15\r
	jz   L_save_len\r
	cmp  bl,cl\r
	jae  L_add_bits_to_len\r
\r
	mov  ch,cl\r
	xor  eax,eax\r
	lodsw\r
	mov  cl,bl\r
	add  bl,16\r
	shl  eax,cl\r
	or  ebp,eax\r
	mov  cl,ch\r
\r
L_add_bits_to_len:\r
	mov  eax,1\r
	shl  eax,cl\r
	dec  eax\r
	sub  bl,cl\r
	and  eax,ebp\r
	shr  ebp,cl\r
	add  edx,eax\r
\r
L_save_len:\r
	mov  [esp+24],edx\r
\r
\r
L_decode_distance:\r
; 549 "inffast.S"\r
	cmp  bl,15\r
	ja   L_get_distance_code\r
\r
	xor  eax,eax\r
	lodsw\r
	mov  cl,bl\r
	add  bl,16\r
	shl  eax,cl\r
	or  ebp,eax\r
\r
L_get_distance_code:\r
	mov  edx, [esp+4]\r
	mov  ecx, [esp+12]\r
	and  edx,ebp\r
	mov  eax, [ecx+edx*4]\r
\r
\r
L_dodist:\r
	mov  edx,eax\r
	shr  edx,16\r
	mov  cl,ah\r
	sub  bl,ah\r
	shr  ebp,cl\r
; 584 "inffast.S"\r
	mov  cl,al\r
\r
	test  al,16\r
	jz  L_test_for_second_level_dist\r
	and  cl,15\r
	jz  L_check_dist_one\r
	cmp  bl,cl\r
	jae  L_add_bits_to_dist\r
\r
	mov  ch,cl\r
	xor  eax,eax\r
	lodsw\r
	mov  cl,bl\r
	add  bl,16\r
	shl  eax,cl\r
	or  ebp,eax\r
	mov  cl,ch\r
\r
L_add_bits_to_dist:\r
	mov  eax,1\r
	shl  eax,cl\r
	dec  eax\r
	sub  bl,cl\r
	and  eax,ebp\r
	shr  ebp,cl\r
	add  edx,eax\r
	jmp  L_check_window\r
\r
L_check_window:\r
; 625 "inffast.S"\r
	mov  [esp+44],esi\r
	mov  eax,edi\r
	sub  eax, [esp+40]\r
\r
	cmp  eax,edx\r
	jb   L_clip_window\r
\r
	mov  ecx, [esp+24]\r
	mov  esi,edi\r
	sub  esi,edx\r
\r
	sub  ecx,3\r
	mov  al, [esi]\r
	mov  [edi],al\r
	mov  al, [esi+1]\r
	mov  dl, [esi+2]\r
	add  esi,3\r
	mov  [edi+1],al\r
	mov  [edi+2],dl\r
	add  edi,3\r
	rep movsb\r
\r
	mov  esi, [esp+44]\r
	jmp  L_while_test\r
\r
ALIGN 4\r
L_check_dist_one:\r
	cmp  edx,1\r
	jne  L_check_window\r
	cmp  [esp+40],edi\r
	je  L_check_window\r
\r
	dec  edi\r
	mov  ecx, [esp+24]\r
	mov  al, [edi]\r
	sub  ecx,3\r
\r
	mov  [edi+1],al\r
	mov  [edi+2],al\r
	mov  [edi+3],al\r
	add  edi,4\r
	rep stosb\r
\r
	jmp  L_while_test\r
\r
ALIGN 4\r
L_test_for_second_level_length:\r
\r
\r
\r
\r
	test  al,64\r
	jnz   L_test_for_end_of_block\r
\r
	mov  eax,1\r
	shl  eax,cl\r
	dec  eax\r
	and  eax,ebp\r
	add  eax,edx\r
	mov  edx, [esp+8]\r
	mov  eax, [edx+eax*4]\r
	jmp  L_dolen\r
\r
ALIGN 4\r
L_test_for_second_level_dist:\r
\r
\r
\r
\r
	test  al,64\r
	jnz   L_invalid_distance_code\r
\r
	mov  eax,1\r
	shl  eax,cl\r
	dec  eax\r
	and  eax,ebp\r
	add  eax,edx\r
	mov  edx, [esp+12]\r
	mov  eax, [edx+eax*4]\r
	jmp  L_dodist\r
\r
ALIGN 4\r
L_clip_window:\r
; 721 "inffast.S"\r
	mov  ecx,eax\r
	mov  eax, [esp+52]\r
	neg  ecx\r
	mov  esi, [esp+56]\r
\r
	cmp  eax,edx\r
	jb   L_invalid_distance_too_far\r
\r
	add  ecx,edx\r
	cmp  dword ptr [esp+48],0\r
	jne  L_wrap_around_window\r
\r
	sub  eax,ecx\r
	add  esi,eax\r
; 749 "inffast.S"\r
	mov  eax, [esp+24]\r
	cmp  eax,ecx\r
	jbe  L_do_copy1\r
\r
	sub  eax,ecx\r
	rep movsb\r
	mov  esi,edi\r
	sub  esi,edx\r
	jmp  L_do_copy1\r
\r
	cmp  eax,ecx\r
	jbe  L_do_copy1\r
\r
	sub  eax,ecx\r
	rep movsb\r
	mov  esi,edi\r
	sub  esi,edx\r
	jmp  L_do_copy1\r
\r
L_wrap_around_window:\r
; 793 "inffast.S"\r
	mov  eax, [esp+48]\r
	cmp  ecx,eax\r
	jbe  L_contiguous_in_window\r
\r
	add  esi, [esp+52]\r
	add  esi,eax\r
	sub  esi,ecx\r
	sub  ecx,eax\r
\r
\r
	mov  eax, [esp+24]\r
	cmp  eax,ecx\r
	jbe  L_do_copy1\r
\r
	sub  eax,ecx\r
	rep movsb\r
	mov  esi, [esp+56]\r
	mov  ecx, [esp+48]\r
	cmp  eax,ecx\r
	jbe  L_do_copy1\r
\r
	sub  eax,ecx\r
	rep movsb\r
	mov  esi,edi\r
	sub  esi,edx\r
	jmp  L_do_copy1\r
\r
L_contiguous_in_window:\r
; 836 "inffast.S"\r
	add  esi,eax\r
	sub  esi,ecx\r
\r
\r
	mov  eax, [esp+24]\r
	cmp  eax,ecx\r
	jbe  L_do_copy1\r
\r
	sub  eax,ecx\r
	rep movsb\r
	mov  esi,edi\r
	sub  esi,edx\r
\r
L_do_copy1:\r
; 862 "inffast.S"\r
	mov  ecx,eax\r
	rep movsb\r
\r
	mov  esi, [esp+44]\r
	jmp  L_while_test\r
; 878 "inffast.S"\r
ALIGN 4\r
L_init_mmx:\r
	emms\r
\r
\r
\r
\r
\r
	movd mm0,ebp\r
	mov  ebp,ebx\r
; 896 "inffast.S"\r
	movd mm4,dword ptr [esp+0]\r
	movq mm3,mm4\r
	movd mm5,dword ptr [esp+4]\r
	movq mm2,mm5\r
	pxor mm1,mm1\r
	mov  ebx, [esp+8]\r
	jmp  L_do_loop_mmx\r
\r
ALIGN 4\r
L_do_loop_mmx:\r
	psrlq mm0,mm1\r
\r
	cmp  ebp,32\r
	ja  L_get_length_code_mmx\r
\r
	movd mm6,ebp\r
	movd mm7,dword ptr [esi]\r
	add  esi,4\r
	psllq mm7,mm6\r
	add  ebp,32\r
	por mm0,mm7\r
\r
L_get_length_code_mmx:\r
	pand mm4,mm0\r
	movd eax,mm4\r
	movq mm4,mm3\r
	mov  eax, [ebx+eax*4]\r
\r
L_dolen_mmx:\r
	movzx  ecx,ah\r
	movd mm1,ecx\r
	sub  ebp,ecx\r
\r
	test  al,al\r
	jnz L_test_for_length_base_mmx\r
\r
	shr  eax,16\r
	stosb\r
\r
L_while_test_mmx:\r
\r
\r
	cmp  [esp+16],edi\r
	jbe L_break_loop\r
\r
	cmp  [esp+20],esi\r
	ja L_do_loop_mmx\r
	jmp L_break_loop\r
\r
L_test_for_length_base_mmx:\r
\r
	mov  edx,eax\r
	shr  edx,16\r
\r
	test  al,16\r
	jz  L_test_for_second_level_length_mmx\r
	and  eax,15\r
	jz L_decode_distance_mmx\r
\r
	psrlq mm0,mm1\r
	movd mm1,eax\r
	movd ecx,mm0\r
	sub  ebp,eax\r
	and  ecx, [inflate_fast_mask+eax*4]\r
	add  edx,ecx\r
\r
L_decode_distance_mmx:\r
	psrlq mm0,mm1\r
\r
	cmp  ebp,32\r
	ja L_get_dist_code_mmx\r
\r
	movd mm6,ebp\r
	movd mm7,dword ptr [esi]\r
	add  esi,4\r
	psllq mm7,mm6\r
	add  ebp,32\r
	por mm0,mm7\r
\r
L_get_dist_code_mmx:\r
	mov  ebx, [esp+12]\r
	pand mm5,mm0\r
	movd eax,mm5\r
	movq mm5,mm2\r
	mov  eax, [ebx+eax*4]\r
\r
L_dodist_mmx:\r
\r
	movzx  ecx,ah\r
	mov  ebx,eax\r
	shr  ebx,16\r
	sub  ebp,ecx\r
	movd mm1,ecx\r
\r
	test  al,16\r
	jz L_test_for_second_level_dist_mmx\r
	and  eax,15\r
	jz L_check_dist_one_mmx\r
\r
L_add_bits_to_dist_mmx:\r
	psrlq mm0,mm1\r
	movd mm1,eax\r
	movd ecx,mm0\r
	sub  ebp,eax\r
	and  ecx, [inflate_fast_mask+eax*4]\r
	add  ebx,ecx\r
\r
L_check_window_mmx:\r
	mov  [esp+44],esi\r
	mov  eax,edi\r
	sub  eax, [esp+40]\r
\r
	cmp  eax,ebx\r
	jb L_clip_window_mmx\r
\r
	mov  ecx,edx\r
	mov  esi,edi\r
	sub  esi,ebx\r
\r
	sub  ecx,3\r
	mov  al, [esi]\r
	mov  [edi],al\r
	mov  al, [esi+1]\r
	mov  dl, [esi+2]\r
	add  esi,3\r
	mov  [edi+1],al\r
	mov  [edi+2],dl\r
	add  edi,3\r
	rep movsb\r
\r
	mov  esi, [esp+44]\r
	mov  ebx, [esp+8]\r
	jmp  L_while_test_mmx\r
\r
ALIGN 4\r
L_check_dist_one_mmx:\r
	cmp  ebx,1\r
	jne  L_check_window_mmx\r
	cmp  [esp+40],edi\r
	je   L_check_window_mmx\r
\r
	dec  edi\r
	mov  ecx,edx\r
	mov  al, [edi]\r
	sub  ecx,3\r
\r
	mov  [edi+1],al\r
	mov  [edi+2],al\r
	mov  [edi+3],al\r
	add  edi,4\r
	rep stosb\r
\r
	mov  ebx, [esp+8]\r
	jmp  L_while_test_mmx\r
\r
ALIGN 4\r
L_test_for_second_level_length_mmx:\r
	test  al,64\r
	jnz L_test_for_end_of_block\r
\r
	and  eax,15\r
	psrlq mm0,mm1\r
	movd ecx,mm0\r
	and  ecx, [inflate_fast_mask+eax*4]\r
	add  ecx,edx\r
	mov  eax, [ebx+ecx*4]\r
	jmp L_dolen_mmx\r
\r
ALIGN 4\r
L_test_for_second_level_dist_mmx:\r
	test  al,64\r
	jnz L_invalid_distance_code\r
\r
	and  eax,15\r
	psrlq mm0,mm1\r
	movd ecx,mm0\r
	and  ecx, [inflate_fast_mask+eax*4]\r
	mov  eax, [esp+12]\r
	add  ecx,ebx\r
	mov  eax, [eax+ecx*4]\r
	jmp  L_dodist_mmx\r
\r
ALIGN 4\r
L_clip_window_mmx:\r
\r
	mov  ecx,eax\r
	mov  eax, [esp+52]\r
	neg  ecx\r
	mov  esi, [esp+56]\r
\r
	cmp  eax,ebx\r
	jb  L_invalid_distance_too_far\r
\r
	add  ecx,ebx\r
	cmp  dword ptr [esp+48],0\r
	jne  L_wrap_around_window_mmx\r
\r
	sub  eax,ecx\r
	add  esi,eax\r
\r
	cmp  edx,ecx\r
	jbe  L_do_copy1_mmx\r
\r
	sub  edx,ecx\r
	rep movsb\r
	mov  esi,edi\r
	sub  esi,ebx\r
	jmp  L_do_copy1_mmx\r
\r
	cmp  edx,ecx\r
	jbe  L_do_copy1_mmx\r
\r
	sub  edx,ecx\r
	rep movsb\r
	mov  esi,edi\r
	sub  esi,ebx\r
	jmp  L_do_copy1_mmx\r
\r
L_wrap_around_window_mmx:\r
\r
	mov  eax, [esp+48]\r
	cmp  ecx,eax\r
	jbe  L_contiguous_in_window_mmx\r
\r
	add  esi, [esp+52]\r
	add  esi,eax\r
	sub  esi,ecx\r
	sub  ecx,eax\r
\r
\r
	cmp  edx,ecx\r
	jbe  L_do_copy1_mmx\r
\r
	sub  edx,ecx\r
	rep movsb\r
	mov  esi, [esp+56]\r
	mov  ecx, [esp+48]\r
	cmp  edx,ecx\r
	jbe  L_do_copy1_mmx\r
\r
	sub  edx,ecx\r
	rep movsb\r
	mov  esi,edi\r
	sub  esi,ebx\r
	jmp  L_do_copy1_mmx\r
\r
L_contiguous_in_window_mmx:\r
\r
	add  esi,eax\r
	sub  esi,ecx\r
\r
\r
	cmp  edx,ecx\r
	jbe  L_do_copy1_mmx\r
\r
	sub  edx,ecx\r
	rep movsb\r
	mov  esi,edi\r
	sub  esi,ebx\r
\r
L_do_copy1_mmx:\r
\r
\r
	mov  ecx,edx\r
	rep movsb\r
\r
	mov  esi, [esp+44]\r
	mov  ebx, [esp+8]\r
	jmp  L_while_test_mmx\r
; 1174 "inffast.S"\r
L_invalid_distance_code:\r
\r
\r
\r
\r
\r
	mov  ecx, invalid_distance_code_msg\r
	mov  edx,INFLATE_MODE_BAD\r
	jmp  L_update_stream_state\r
\r
L_test_for_end_of_block:\r
\r
\r
\r
\r
\r
	test  al,32\r
	jz  L_invalid_literal_length_code\r
\r
	mov  ecx,0\r
	mov  edx,INFLATE_MODE_TYPE\r
	jmp  L_update_stream_state\r
\r
L_invalid_literal_length_code:\r
\r
\r
\r
\r
\r
	mov  ecx, invalid_literal_length_code_msg\r
	mov  edx,INFLATE_MODE_BAD\r
	jmp  L_update_stream_state\r
\r
L_invalid_distance_too_far:\r
\r
\r
\r
	mov  esi, [esp+44]\r
	mov  ecx, invalid_distance_too_far_msg\r
	mov  edx,INFLATE_MODE_BAD\r
	jmp  L_update_stream_state\r
\r
L_update_stream_state:\r
\r
	mov  eax, [esp+88]\r
	test  ecx,ecx\r
	jz  L_skip_msg\r
	mov  [eax+24],ecx\r
L_skip_msg:\r
	mov  eax, [eax+28]\r
	mov  [eax+mode_state],edx\r
	jmp  L_break_loop\r
\r
ALIGN 4\r
L_break_loop:\r
; 1243 "inffast.S"\r
	cmp  dword ptr [inflate_fast_use_mmx],2\r
	jne  L_update_next_in\r
\r
\r
\r
	mov  ebx,ebp\r
\r
L_update_next_in:\r
; 1266 "inffast.S"\r
	mov  eax, [esp+88]\r
	mov  ecx,ebx\r
	mov  edx, [eax+28]\r
	shr  ecx,3\r
	sub  esi,ecx\r
	shl  ecx,3\r
	sub  ebx,ecx\r
	mov  [eax+12],edi\r
	mov  [edx+bits_state],ebx\r
	mov  ecx,ebx\r
\r
	lea  ebx, [esp+28]\r
	cmp  [esp+20],ebx\r
	jne  L_buf_not_used\r
\r
	sub  esi,ebx\r
	mov  ebx, [eax+0]\r
	mov  [esp+20],ebx\r
	add  esi,ebx\r
	mov  ebx, [eax+4]\r
	sub  ebx,11\r
	add  [esp+20],ebx\r
\r
L_buf_not_used:\r
	mov  [eax+0],esi\r
\r
	mov  ebx,1\r
	shl  ebx,cl\r
	dec  ebx\r
\r
\r
\r
\r
\r
	cmp  dword ptr [inflate_fast_use_mmx],2\r
	jne  L_update_hold\r
\r
\r
\r
	psrlq mm0,mm1\r
	movd ebp,mm0\r
\r
	emms\r
\r
L_update_hold:\r
\r
\r
\r
	and  ebp,ebx\r
	mov  [edx+hold_state],ebp\r
\r
\r
\r
\r
	mov  ebx, [esp+20]\r
	cmp  ebx,esi\r
	jbe  L_last_is_smaller\r
\r
	sub  ebx,esi\r
	add  ebx,11\r
	mov  [eax+4],ebx\r
	jmp  L_fixup_out\r
L_last_is_smaller:\r
	sub  esi,ebx\r
	neg  esi\r
	add  esi,11\r
	mov  [eax+4],esi\r
\r
\r
\r
\r
L_fixup_out:\r
\r
	mov  ebx, [esp+16]\r
	cmp  ebx,edi\r
	jbe  L_end_is_smaller\r
\r
	sub  ebx,edi\r
	add  ebx,257\r
	mov  [eax+16],ebx\r
	jmp  L_done\r
L_end_is_smaller:\r
	sub  edi,ebx\r
	neg  edi\r
	add  edi,257\r
	mov  [eax+16],edi\r
\r
\r
\r
\r
\r
L_done:\r
	add  esp,64\r
	popfd\r
	pop  ebx\r
	pop  ebp\r
	pop  esi\r
	pop  edi\r
	ret\r
_inflate_fast endp\r
\r
_TEXT	ends\r
end\r
Commit	Line	Data
5ca28f79 L	1	;/* inffas32.asm is a hand tuned assembler version of inffast.c -- fast decoding\r
	2	; *\r
	3	; * inffas32.asm is derivated from inffas86.c, with translation of assembly code\r
	4	; *\r
	5	; * Copyright (C) 1995-2003 Mark Adler\r
	6	; * For conditions of distribution and use, see copyright notice in zlib.h\r
	7	; *\r
	8	; * Copyright (C) 2003 Chris Anderson <christop@charm.net>\r
	9	; * Please use the copyright conditions above.\r
	10	; *\r
	11	; * Mar-13-2003 -- Most of this is derived from inffast.S which is derived from\r
	12	; * the gcc -S output of zlib-1.2.0/inffast.c. Zlib-1.2.0 is in beta release at\r
	13	; * the moment. I have successfully compiled and tested this code with gcc2.96,\r
	14	; * gcc3.2, icc5.0, msvc6.0. It is very close to the speed of inffast.S\r
	15	; * compiled with gcc -DNO_MMX, but inffast.S is still faster on the P3 with MMX\r
	16	; * enabled. I will attempt to merge the MMX code into this version. Newer\r
	17	; * versions of this and inffast.S can be found at\r
	18	; * http://www.eetbeetee.com/zlib/ and http://www.charm.net/~christop/zlib/\r
	19	; *\r
	20	; * 2005 : modification by Gilles Vollant\r
	21	; */\r
	22	; For Visual C++ 4.x and higher and ML 6.x and higher\r
	23	; ml.exe is in directory \MASM611C of Win95 DDK\r
	24	; ml.exe is also distributed in http://www.masm32.com/masmdl.htm\r
	25	; and in VC++2003 toolkit at http://msdn.microsoft.com/visualc/vctoolkit2003/\r
	26	;\r
	27	;\r
	28	; compile with command line option\r
	29	; ml /coff /Zi /c /Flinffas32.lst inffas32.asm\r
	30	\r
	31	; if you define NO_GZIP (see inflate.h), compile with\r
	32	; ml /coff /Zi /c /Flinffas32.lst /DNO_GUNZIP inffas32.asm\r
	33	\r
	34	\r
	35	; zlib122sup is 0 fort zlib 1.2.2.1 and lower\r
	36	; zlib122sup is 8 fort zlib 1.2.2.2 and more (with addition of dmax and head\r
	37	; in inflate_state in inflate.h)\r
	38	zlib1222sup equ 8\r
	39	\r
	40	\r
	41	IFDEF GUNZIP\r
	42	INFLATE_MODE_TYPE equ 11\r
	43	INFLATE_MODE_BAD equ 26\r
	44	ELSE\r
	45	IFNDEF NO_GUNZIP\r
	46	INFLATE_MODE_TYPE equ 11\r
	47	INFLATE_MODE_BAD equ 26\r
	48	ELSE\r
	49	INFLATE_MODE_TYPE equ 3\r
	50	INFLATE_MODE_BAD equ 17\r
	51	ENDIF\r
	52	ENDIF\r
	53	\r
	54	\r
	55	; 75 "inffast.S"\r
	56	;FILE "inffast.S"\r
	57	\r
	58	;;;GLOBAL _inflate_fast\r
	59	\r
	60	;;;SECTION .text\r
	61	\r
	62	\r
	63	\r
	64	.586p\r
65	.mmx\r
66	\r
67	name inflate_fast_x86\r
68	.MODEL FLAT\r
69	\r
70	_DATA segment\r
71	inflate_fast_use_mmx:\r
72	dd 1\r
73	\r
74	\r
75	_TEXT segment\r
76	\r
77	\r
78	\r
79	ALIGN 4\r
80	db 'Fast decoding Code from Chris Anderson'\r
81	db 0\r
82	\r
83	ALIGN 4\r
84	invalid_literal_length_code_msg:\r
85	db 'invalid literal/length code'\r
86	db 0\r
87	\r
88	ALIGN 4\r
89	invalid_distance_code_msg:\r
90	db 'invalid distance code'\r
91	db 0\r
92	\r
93	ALIGN 4\r
94	invalid_distance_too_far_msg:\r
95	db 'invalid distance too far back'\r
96	db 0\r
97	\r
98	\r
99	ALIGN 4\r
100	inflate_fast_mask:\r
101	dd 0\r
102	dd 1\r
103	dd 3\r
104	dd 7\r
105	dd 15\r
106	dd 31\r
107	dd 63\r
108	dd 127\r
109	dd 255\r
110	dd 511\r
111	dd 1023\r
112	dd 2047\r
113	dd 4095\r
114	dd 8191\r
115	dd 16383\r
116	dd 32767\r
117	dd 65535\r
118	dd 131071\r
119	dd 262143\r
120	dd 524287\r
121	dd 1048575\r
122	dd 2097151\r
123	dd 4194303\r
124	dd 8388607\r
125	dd 16777215\r
126	dd 33554431\r
127	dd 67108863\r
128	dd 134217727\r
129	dd 268435455\r
130	dd 536870911\r
131	dd 1073741823\r
132	dd 2147483647\r
133	dd 4294967295\r
134	\r
135	\r
136	mode_state equ 0 ;/* state->mode */\r
137	wsize_state equ (32+zlib1222sup) ;/* state->wsize */\r
138	write_state equ (36+4+zlib1222sup) ;/* state->write */\r
139	window_state equ (40+4+zlib1222sup) ;/* state->window */\r
140	hold_state equ (44+4+zlib1222sup) ;/* state->hold */\r
141	bits_state equ (48+4+zlib1222sup) ;/* state->bits */\r
142	lencode_state equ (64+4+zlib1222sup) ;/* state->lencode */\r
143	distcode_state equ (68+4+zlib1222sup) ;/* state->distcode */\r
144	lenbits_state equ (72+4+zlib1222sup) ;/* state->lenbits */\r
145	distbits_state equ (76+4+zlib1222sup) ;/* state->distbits */\r
146	\r
147	\r
148	;;SECTION .text\r
149	; 205 "inffast.S"\r
150	;GLOBAL inflate_fast_use_mmx\r
151	\r
152	;SECTION .data\r
153	\r
154	\r
155	; GLOBAL inflate_fast_use_mmx:object\r
156	;.size inflate_fast_use_mmx, 4\r
157	; 226 "inffast.S"\r
158	;SECTION .text\r
159	\r
160	ALIGN 4\r
161	_inflate_fast proc near\r
162	.FPO (16, 4, 0, 0, 1, 0)\r
163	push edi\r
164	push esi\r
165	push ebp\r
166	push ebx\r
167	pushfd\r
168	sub esp,64\r
169	cld\r
170	\r
171	\r
172	\r
173	\r
174	mov esi, [esp+88]\r
175	mov edi, [esi+28]\r
176	\r
177	\r
178	\r
179	\r
180	\r
181	\r
182	\r
183	mov edx, [esi+4]\r
184	mov eax, [esi+0]\r
185	\r
186	add edx,eax\r
187	sub edx,11\r
188	\r
189	mov [esp+44],eax\r
190	mov [esp+20],edx\r
191	\r
192	mov ebp, [esp+92]\r
193	mov ecx, [esi+16]\r
194	mov ebx, [esi+12]\r
195	\r
196	sub ebp,ecx\r
197	neg ebp\r
198	add ebp,ebx\r
199	\r
200	sub ecx,257\r
201	add ecx,ebx\r
202	\r
203	mov [esp+60],ebx\r
204	mov [esp+40],ebp\r
205	mov [esp+16],ecx\r
206	; 285 "inffast.S"\r
207	mov eax, [edi+lencode_state]\r
208	mov ecx, [edi+distcode_state]\r
209	\r
210	mov [esp+8],eax\r
211	mov [esp+12],ecx\r
212	\r
213	mov eax,1\r
214	mov ecx, [edi+lenbits_state]\r
215	shl eax,cl\r
216	dec eax\r
217	mov [esp+0],eax\r
218	\r
219	mov eax,1\r
220	mov ecx, [edi+distbits_state]\r
221	shl eax,cl\r
222	dec eax\r
223	mov [esp+4],eax\r
224	\r
225	mov eax, [edi+wsize_state]\r
226	mov ecx, [edi+write_state]\r
227	mov edx, [edi+window_state]\r
228	\r
229	mov [esp+52],eax\r
230	mov [esp+48],ecx\r
231	mov [esp+56],edx\r
232	\r
233	mov ebp, [edi+hold_state]\r
234	mov ebx, [edi+bits_state]\r
235	; 321 "inffast.S"\r
236	mov esi, [esp+44]\r
237	mov ecx, [esp+20]\r
238	cmp ecx,esi\r
239	ja L_align_long\r
240	\r
241	add ecx,11\r
242	sub ecx,esi\r
243	mov eax,12\r
244	sub eax,ecx\r
245	lea edi, [esp+28]\r
246	rep movsb\r
247	mov ecx,eax\r
248	xor eax,eax\r
249	rep stosb\r
250	lea esi, [esp+28]\r
251	mov [esp+20],esi\r
252	jmp L_is_aligned\r
253	\r
254	\r
255	L_align_long:\r
256	test esi,3\r
257	jz L_is_aligned\r
258	xor eax,eax\r
259	mov al, [esi]\r
260	inc esi\r
261	mov ecx,ebx\r
262	add ebx,8\r
263	shl eax,cl\r
264	or ebp,eax\r
265	jmp L_align_long\r
266	\r
267	L_is_aligned:\r
268	mov edi, [esp+60]\r
269	; 366 "inffast.S"\r
270	L_check_mmx:\r
271	cmp dword ptr [inflate_fast_use_mmx],2\r
272	je L_init_mmx\r
273	ja L_do_loop\r
274	\r
275	push eax\r
276	push ebx\r
277	push ecx\r
278	push edx\r
279	pushfd\r
280	mov eax, [esp]\r
281	xor dword ptr [esp],0200000h\r
282	\r
283	\r
284	\r
285	\r
286	popfd\r
287	pushfd\r
288	pop edx\r
289	xor edx,eax\r
290	jz L_dont_use_mmx\r
291	xor eax,eax\r
292	cpuid\r
293	cmp ebx,0756e6547h\r
294	jne L_dont_use_mmx\r
295	cmp ecx,06c65746eh\r
296	jne L_dont_use_mmx\r
297	cmp edx,049656e69h\r
298	jne L_dont_use_mmx\r
299	mov eax,1\r
300	cpuid\r
301	shr eax,8\r
302	and eax,15\r
303	cmp eax,6\r
304	jne L_dont_use_mmx\r
305	test edx,0800000h\r
306	jnz L_use_mmx\r
307	jmp L_dont_use_mmx\r
308	L_use_mmx:\r
309	mov dword ptr [inflate_fast_use_mmx],2\r
310	jmp L_check_mmx_pop\r
311	L_dont_use_mmx:\r
312	mov dword ptr [inflate_fast_use_mmx],3\r
313	L_check_mmx_pop:\r
314	pop edx\r
315	pop ecx\r
316	pop ebx\r
317	pop eax\r
318	jmp L_check_mmx\r
319	; 426 "inffast.S"\r
320	ALIGN 4\r
321	L_do_loop:\r
322	; 437 "inffast.S"\r
323	cmp bl,15\r
324	ja L_get_length_code\r
325	\r
326	xor eax,eax\r
327	lodsw\r
328	mov cl,bl\r
329	add bl,16\r
330	shl eax,cl\r
331	or ebp,eax\r
332	\r
333	L_get_length_code:\r
334	mov edx, [esp+0]\r
335	mov ecx, [esp+8]\r
336	and edx,ebp\r
337	mov eax, [ecx+edx*4]\r
338	\r
339	L_dolen:\r
340	\r
341	\r
342	\r
343	\r
344	\r
345	\r
346	mov cl,ah\r
347	sub bl,ah\r
348	shr ebp,cl\r
349	\r
350	\r
351	\r
352	\r
353	\r
354	\r
355	test al,al\r
356	jnz L_test_for_length_base\r
357	\r
358	shr eax,16\r
359	stosb\r
360	\r
361	L_while_test:\r
362	\r
363	\r
364	cmp [esp+16],edi\r
365	jbe L_break_loop\r
366	\r
367	cmp [esp+20],esi\r
368	ja L_do_loop\r
369	jmp L_break_loop\r
370	\r
371	L_test_for_length_base:\r
372	; 502 "inffast.S"\r
373	mov edx,eax\r
374	shr edx,16\r
375	mov cl,al\r
376	\r
377	test al,16\r
378	jz L_test_for_second_level_length\r
379	and cl,15\r
380	jz L_save_len\r
381	cmp bl,cl\r
382	jae L_add_bits_to_len\r
383	\r
384	mov ch,cl\r
385	xor eax,eax\r
386	lodsw\r
387	mov cl,bl\r
388	add bl,16\r
389	shl eax,cl\r
390	or ebp,eax\r
391	mov cl,ch\r
392	\r
393	L_add_bits_to_len:\r
394	mov eax,1\r
395	shl eax,cl\r
396	dec eax\r
397	sub bl,cl\r
398	and eax,ebp\r
399	shr ebp,cl\r
400	add edx,eax\r
401	\r
402	L_save_len:\r
403	mov [esp+24],edx\r
404	\r
405	\r
406	L_decode_distance:\r
407	; 549 "inffast.S"\r
408	cmp bl,15\r
409	ja L_get_distance_code\r
410	\r
411	xor eax,eax\r
412	lodsw\r
413	mov cl,bl\r
414	add bl,16\r
415	shl eax,cl\r
416	or ebp,eax\r
417	\r
418	L_get_distance_code:\r
419	mov edx, [esp+4]\r
420	mov ecx, [esp+12]\r
421	and edx,ebp\r
422	mov eax, [ecx+edx*4]\r
423	\r
424	\r
425	L_dodist:\r
426	mov edx,eax\r
427	shr edx,16\r
428	mov cl,ah\r
429	sub bl,ah\r
430	shr ebp,cl\r
431	; 584 "inffast.S"\r
432	mov cl,al\r
433	\r
434	test al,16\r
435	jz L_test_for_second_level_dist\r
436	and cl,15\r
437	jz L_check_dist_one\r
438	cmp bl,cl\r
439	jae L_add_bits_to_dist\r
440	\r
441	mov ch,cl\r
442	xor eax,eax\r
443	lodsw\r
444	mov cl,bl\r
445	add bl,16\r
446	shl eax,cl\r
447	or ebp,eax\r
448	mov cl,ch\r
449	\r
450	L_add_bits_to_dist:\r
451	mov eax,1\r
452	shl eax,cl\r
453	dec eax\r
454	sub bl,cl\r
455	and eax,ebp\r
456	shr ebp,cl\r
457	add edx,eax\r
458	jmp L_check_window\r
459	\r
460	L_check_window:\r
461	; 625 "inffast.S"\r
462	mov [esp+44],esi\r
463	mov eax,edi\r
464	sub eax, [esp+40]\r
465	\r
466	cmp eax,edx\r
467	jb L_clip_window\r
468	\r
469	mov ecx, [esp+24]\r
470	mov esi,edi\r
471	sub esi,edx\r
472	\r
473	sub ecx,3\r
474	mov al, [esi]\r
475	mov [edi],al\r
476	mov al, [esi+1]\r
477	mov dl, [esi+2]\r
478	add esi,3\r
479	mov [edi+1],al\r
480	mov [edi+2],dl\r
481	add edi,3\r
482	rep movsb\r
483	\r
484	mov esi, [esp+44]\r
485	jmp L_while_test\r
486	\r
487	ALIGN 4\r
488	L_check_dist_one:\r
489	cmp edx,1\r
490	jne L_check_window\r
491	cmp [esp+40],edi\r
492	je L_check_window\r
493	\r
494	dec edi\r
495	mov ecx, [esp+24]\r
496	mov al, [edi]\r
497	sub ecx,3\r
498	\r
499	mov [edi+1],al\r
500	mov [edi+2],al\r
501	mov [edi+3],al\r
502	add edi,4\r
503	rep stosb\r
504	\r
505	jmp L_while_test\r
506	\r
507	ALIGN 4\r
508	L_test_for_second_level_length:\r
509	\r
510	\r
511	\r
512	\r
513	test al,64\r
514	jnz L_test_for_end_of_block\r
515	\r
516	mov eax,1\r
517	shl eax,cl\r
518	dec eax\r
519	and eax,ebp\r
520	add eax,edx\r
521	mov edx, [esp+8]\r
522	mov eax, [edx+eax*4]\r
523	jmp L_dolen\r
524	\r
525	ALIGN 4\r
526	L_test_for_second_level_dist:\r
527	\r
528	\r
529	\r
530	\r
531	test al,64\r
532	jnz L_invalid_distance_code\r
533	\r
534	mov eax,1\r
535	shl eax,cl\r
536	dec eax\r
537	and eax,ebp\r
538	add eax,edx\r
539	mov edx, [esp+12]\r
540	mov eax, [edx+eax*4]\r
541	jmp L_dodist\r
542	\r
543	ALIGN 4\r
544	L_clip_window:\r
545	; 721 "inffast.S"\r
546	mov ecx,eax\r
547	mov eax, [esp+52]\r
548	neg ecx\r
549	mov esi, [esp+56]\r
550	\r
551	cmp eax,edx\r
552	jb L_invalid_distance_too_far\r
553	\r
554	add ecx,edx\r
555	cmp dword ptr [esp+48],0\r
556	jne L_wrap_around_window\r
557	\r
558	sub eax,ecx\r
559	add esi,eax\r
560	; 749 "inffast.S"\r
561	mov eax, [esp+24]\r
562	cmp eax,ecx\r
563	jbe L_do_copy1\r
564	\r
565	sub eax,ecx\r
566	rep movsb\r
567	mov esi,edi\r
568	sub esi,edx\r
569	jmp L_do_copy1\r
570	\r
571	cmp eax,ecx\r
572	jbe L_do_copy1\r
573	\r
574	sub eax,ecx\r
575	rep movsb\r
576	mov esi,edi\r
577	sub esi,edx\r
578	jmp L_do_copy1\r
579	\r
580	L_wrap_around_window:\r
581	; 793 "inffast.S"\r
582	mov eax, [esp+48]\r
583	cmp ecx,eax\r
584	jbe L_contiguous_in_window\r
585	\r
586	add esi, [esp+52]\r
587	add esi,eax\r
588	sub esi,ecx\r
589	sub ecx,eax\r
590	\r
591	\r
592	mov eax, [esp+24]\r
593	cmp eax,ecx\r
594	jbe L_do_copy1\r
595	\r
596	sub eax,ecx\r
597	rep movsb\r
598	mov esi, [esp+56]\r
599	mov ecx, [esp+48]\r
600	cmp eax,ecx\r
601	jbe L_do_copy1\r
602	\r
603	sub eax,ecx\r
604	rep movsb\r
605	mov esi,edi\r
606	sub esi,edx\r
607	jmp L_do_copy1\r
608	\r
609	L_contiguous_in_window:\r
610	; 836 "inffast.S"\r
611	add esi,eax\r
612	sub esi,ecx\r
613	\r
614	\r
615	mov eax, [esp+24]\r
616	cmp eax,ecx\r
617	jbe L_do_copy1\r
618	\r
619	sub eax,ecx\r
620	rep movsb\r
621	mov esi,edi\r
622	sub esi,edx\r
623	\r
624	L_do_copy1:\r
625	; 862 "inffast.S"\r
626	mov ecx,eax\r
627	rep movsb\r
628	\r
629	mov esi, [esp+44]\r
630	jmp L_while_test\r
631	; 878 "inffast.S"\r
632	ALIGN 4\r
633	L_init_mmx:\r
634	emms\r
635	\r
636	\r
637	\r
638	\r
639	\r
640	movd mm0,ebp\r
641	mov ebp,ebx\r
642	; 896 "inffast.S"\r
643	movd mm4,dword ptr [esp+0]\r
644	movq mm3,mm4\r
645	movd mm5,dword ptr [esp+4]\r
646	movq mm2,mm5\r
647	pxor mm1,mm1\r
648	mov ebx, [esp+8]\r
649	jmp L_do_loop_mmx\r
650	\r
651	ALIGN 4\r
652	L_do_loop_mmx:\r
653	psrlq mm0,mm1\r
654	\r
655	cmp ebp,32\r
656	ja L_get_length_code_mmx\r
657	\r
658	movd mm6,ebp\r
659	movd mm7,dword ptr [esi]\r
660	add esi,4\r
661	psllq mm7,mm6\r
662	add ebp,32\r
663	por mm0,mm7\r
664	\r
665	L_get_length_code_mmx:\r
666	pand mm4,mm0\r
667	movd eax,mm4\r
668	movq mm4,mm3\r
669	mov eax, [ebx+eax*4]\r
670	\r
671	L_dolen_mmx:\r
672	movzx ecx,ah\r
673	movd mm1,ecx\r
674	sub ebp,ecx\r
675	\r
676	test al,al\r
677	jnz L_test_for_length_base_mmx\r
678	\r
679	shr eax,16\r
680	stosb\r
681	\r
682	L_while_test_mmx:\r
683	\r
684	\r
685	cmp [esp+16],edi\r
686	jbe L_break_loop\r
687	\r
688	cmp [esp+20],esi\r
689	ja L_do_loop_mmx\r
690	jmp L_break_loop\r
691	\r
692	L_test_for_length_base_mmx:\r
693	\r
694	mov edx,eax\r
695	shr edx,16\r
696	\r
697	test al,16\r
698	jz L_test_for_second_level_length_mmx\r
699	and eax,15\r
700	jz L_decode_distance_mmx\r
701	\r
702	psrlq mm0,mm1\r
703	movd mm1,eax\r
704	movd ecx,mm0\r
705	sub ebp,eax\r
706	and ecx, [inflate_fast_mask+eax*4]\r
707	add edx,ecx\r
708	\r
709	L_decode_distance_mmx:\r
710	psrlq mm0,mm1\r
711	\r
712	cmp ebp,32\r
713	ja L_get_dist_code_mmx\r
714	\r
715	movd mm6,ebp\r
716	movd mm7,dword ptr [esi]\r
717	add esi,4\r
718	psllq mm7,mm6\r
719	add ebp,32\r
720	por mm0,mm7\r
721	\r
722	L_get_dist_code_mmx:\r
723	mov ebx, [esp+12]\r
724	pand mm5,mm0\r
725	movd eax,mm5\r
726	movq mm5,mm2\r
727	mov eax, [ebx+eax*4]\r
728	\r
729	L_dodist_mmx:\r
730	\r
731	movzx ecx,ah\r
732	mov ebx,eax\r
733	shr ebx,16\r
734	sub ebp,ecx\r
735	movd mm1,ecx\r
736	\r
737	test al,16\r
738	jz L_test_for_second_level_dist_mmx\r
739	and eax,15\r
740	jz L_check_dist_one_mmx\r
741	\r
742	L_add_bits_to_dist_mmx:\r
743	psrlq mm0,mm1\r
744	movd mm1,eax\r
745	movd ecx,mm0\r
746	sub ebp,eax\r
747	and ecx, [inflate_fast_mask+eax*4]\r
748	add ebx,ecx\r
749	\r
750	L_check_window_mmx:\r
751	mov [esp+44],esi\r
752	mov eax,edi\r
753	sub eax, [esp+40]\r
754	\r
755	cmp eax,ebx\r
756	jb L_clip_window_mmx\r
757	\r
758	mov ecx,edx\r
759	mov esi,edi\r
760	sub esi,ebx\r
761	\r
762	sub ecx,3\r
763	mov al, [esi]\r
764	mov [edi],al\r
765	mov al, [esi+1]\r
766	mov dl, [esi+2]\r
767	add esi,3\r
768	mov [edi+1],al\r
769	mov [edi+2],dl\r
770	add edi,3\r
771	rep movsb\r
772	\r
773	mov esi, [esp+44]\r
774	mov ebx, [esp+8]\r
775	jmp L_while_test_mmx\r
776	\r
777	ALIGN 4\r
778	L_check_dist_one_mmx:\r
779	cmp ebx,1\r
780	jne L_check_window_mmx\r
781	cmp [esp+40],edi\r
782	je L_check_window_mmx\r
783	\r
784	dec edi\r
785	mov ecx,edx\r
786	mov al, [edi]\r
787	sub ecx,3\r
788	\r
789	mov [edi+1],al\r
790	mov [edi+2],al\r
791	mov [edi+3],al\r
792	add edi,4\r
793	rep stosb\r
794	\r
795	mov ebx, [esp+8]\r
796	jmp L_while_test_mmx\r
797	\r
798	ALIGN 4\r
799	L_test_for_second_level_length_mmx:\r
800	test al,64\r
801	jnz L_test_for_end_of_block\r
802	\r
803	and eax,15\r
804	psrlq mm0,mm1\r
805	movd ecx,mm0\r
806	and ecx, [inflate_fast_mask+eax*4]\r
807	add ecx,edx\r
808	mov eax, [ebx+ecx*4]\r
809	jmp L_dolen_mmx\r
810	\r
811	ALIGN 4\r
812	L_test_for_second_level_dist_mmx:\r
813	test al,64\r
814	jnz L_invalid_distance_code\r
815	\r
816	and eax,15\r
817	psrlq mm0,mm1\r
818	movd ecx,mm0\r
819	and ecx, [inflate_fast_mask+eax*4]\r
820	mov eax, [esp+12]\r
821	add ecx,ebx\r
822	mov eax, [eax+ecx*4]\r
823	jmp L_dodist_mmx\r
824	\r
825	ALIGN 4\r
826	L_clip_window_mmx:\r
827	\r
828	mov ecx,eax\r
829	mov eax, [esp+52]\r
830	neg ecx\r
831	mov esi, [esp+56]\r
832	\r
833	cmp eax,ebx\r
834	jb L_invalid_distance_too_far\r
835	\r
836	add ecx,ebx\r
837	cmp dword ptr [esp+48],0\r
838	jne L_wrap_around_window_mmx\r
839	\r
840	sub eax,ecx\r
841	add esi,eax\r
842	\r
843	cmp edx,ecx\r
844	jbe L_do_copy1_mmx\r
845	\r
846	sub edx,ecx\r
847	rep movsb\r
848	mov esi,edi\r
849	sub esi,ebx\r
850	jmp L_do_copy1_mmx\r
851	\r
852	cmp edx,ecx\r
853	jbe L_do_copy1_mmx\r
854	\r
855	sub edx,ecx\r
856	rep movsb\r
857	mov esi,edi\r
858	sub esi,ebx\r
859	jmp L_do_copy1_mmx\r
860	\r
861	L_wrap_around_window_mmx:\r
862	\r
863	mov eax, [esp+48]\r
864	cmp ecx,eax\r
865	jbe L_contiguous_in_window_mmx\r
866	\r
867	add esi, [esp+52]\r
868	add esi,eax\r
869	sub esi,ecx\r
870	sub ecx,eax\r
871	\r
872	\r
873	cmp edx,ecx\r
874	jbe L_do_copy1_mmx\r
875	\r
876	sub edx,ecx\r
877	rep movsb\r
878	mov esi, [esp+56]\r
879	mov ecx, [esp+48]\r
880	cmp edx,ecx\r
881	jbe L_do_copy1_mmx\r
882	\r
883	sub edx,ecx\r
884	rep movsb\r
885	mov esi,edi\r
886	sub esi,ebx\r
887	jmp L_do_copy1_mmx\r
888	\r
889	L_contiguous_in_window_mmx:\r
890	\r
891	add esi,eax\r
892	sub esi,ecx\r
893	\r
894	\r
895	cmp edx,ecx\r
896	jbe L_do_copy1_mmx\r
897	\r
898	sub edx,ecx\r
899	rep movsb\r
900	mov esi,edi\r
901	sub esi,ebx\r
902	\r
903	L_do_copy1_mmx:\r
904	\r
905	\r
906	mov ecx,edx\r
907	rep movsb\r
908	\r
909	mov esi, [esp+44]\r
910	mov ebx, [esp+8]\r
911	jmp L_while_test_mmx\r
912	; 1174 "inffast.S"\r
913	L_invalid_distance_code:\r
914	\r
915	\r
916	\r
917	\r
918	\r
919	mov ecx, invalid_distance_code_msg\r
920	mov edx,INFLATE_MODE_BAD\r
921	jmp L_update_stream_state\r
922	\r
923	L_test_for_end_of_block:\r
924	\r
925	\r
926	\r
927	\r
928	\r
929	test al,32\r
930	jz L_invalid_literal_length_code\r
931	\r
932	mov ecx,0\r
933	mov edx,INFLATE_MODE_TYPE\r
934	jmp L_update_stream_state\r
935	\r
936	L_invalid_literal_length_code:\r
937	\r
938	\r
939	\r
940	\r
941	\r
942	mov ecx, invalid_literal_length_code_msg\r
943	mov edx,INFLATE_MODE_BAD\r
944	jmp L_update_stream_state\r
945	\r
946	L_invalid_distance_too_far:\r
947	\r
948	\r
949	\r
950	mov esi, [esp+44]\r
951	mov ecx, invalid_distance_too_far_msg\r
952	mov edx,INFLATE_MODE_BAD\r
953	jmp L_update_stream_state\r
954	\r
955	L_update_stream_state:\r
956	\r
957	mov eax, [esp+88]\r
958	test ecx,ecx\r
959	jz L_skip_msg\r
960	mov [eax+24],ecx\r
961	L_skip_msg:\r
962	mov eax, [eax+28]\r
963	mov [eax+mode_state],edx\r
964	jmp L_break_loop\r
965	\r
966	ALIGN 4\r
967	L_break_loop:\r
968	; 1243 "inffast.S"\r
969	cmp dword ptr [inflate_fast_use_mmx],2\r
970	jne L_update_next_in\r
971	\r
972	\r
973	\r
974	mov ebx,ebp\r
975	\r
976	L_update_next_in:\r
977	; 1266 "inffast.S"\r
978	mov eax, [esp+88]\r
979	mov ecx,ebx\r
980	mov edx, [eax+28]\r
981	shr ecx,3\r
982	sub esi,ecx\r
983	shl ecx,3\r
984	sub ebx,ecx\r
985	mov [eax+12],edi\r
986	mov [edx+bits_state],ebx\r
987	mov ecx,ebx\r
988	\r
989	lea ebx, [esp+28]\r
990	cmp [esp+20],ebx\r
991	jne L_buf_not_used\r
992	\r
993	sub esi,ebx\r
994	mov ebx, [eax+0]\r
995	mov [esp+20],ebx\r
996	add esi,ebx\r
997	mov ebx, [eax+4]\r
998	sub ebx,11\r
999	add [esp+20],ebx\r
1000	\r
1001	L_buf_not_used:\r
1002	mov [eax+0],esi\r
1003	\r
1004	mov ebx,1\r
1005	shl ebx,cl\r
1006	dec ebx\r
1007	\r
1008	\r
1009	\r
1010	\r
1011	\r
1012	cmp dword ptr [inflate_fast_use_mmx],2\r
1013	jne L_update_hold\r
1014	\r
1015	\r
1016	\r
1017	psrlq mm0,mm1\r
1018	movd ebp,mm0\r
1019	\r
1020	emms\r
1021	\r
1022	L_update_hold:\r
1023	\r
1024	\r
1025	\r
1026	and ebp,ebx\r
1027	mov [edx+hold_state],ebp\r
1028	\r
1029	\r
1030	\r
1031	\r
1032	mov ebx, [esp+20]\r
1033	cmp ebx,esi\r
1034	jbe L_last_is_smaller\r
1035	\r
1036	sub ebx,esi\r
1037	add ebx,11\r
1038	mov [eax+4],ebx\r
1039	jmp L_fixup_out\r
1040	L_last_is_smaller:\r
1041	sub esi,ebx\r
1042	neg esi\r
1043	add esi,11\r
1044	mov [eax+4],esi\r
1045	\r
1046	\r
1047	\r
1048	\r
1049	L_fixup_out:\r
1050	\r
1051	mov ebx, [esp+16]\r
1052	cmp ebx,edi\r
1053	jbe L_end_is_smaller\r
1054	\r
1055	sub ebx,edi\r
1056	add ebx,257\r
1057	mov [eax+16],ebx\r
1058	jmp L_done\r
1059	L_end_is_smaller:\r
1060	sub edi,ebx\r
1061	neg edi\r
1062	add edi,257\r
1063	mov [eax+16],edi\r
1064	\r
1065	\r
1066	\r
1067	\r
1068	\r
1069	L_done:\r
1070	add esp,64\r
1071	popfd\r
1072	pop ebx\r
1073	pop ebp\r
1074	pop esi\r
1075	pop edi\r
1076	ret\r
1077	_inflate_fast endp\r
1078	\r
1079	_TEXT ends\r
1080	end\r