From 33701d3a84daf96b187b36f45480f61b5eafc169 Mon Sep 17 00:00:00 2001 From: magnum Date: Fri, 11 Nov 2011 01:54:38 +0100 Subject: [PATCH 7/7] j8: Further revisions of pre-built intrinsics Unified sse-intrinsics-win32.S and sse-intrinsics-32.S Added correct PARA defines to SHA1 too in arch.h Added -msse2 to linux-x86-sse2i target Changed icc options to produce code that runs on any x86-64 but has alternate (optimised) code paths --- src/Makefile | 36 ++++---- src/sse-intrinsics-32.S | 208 +++------------------------------------------- src/x86-64.h | 5 +- src/x86-ssei.h | 4 +- 4 files changed, 38 insertions(+), 215 deletions(-) diff --git a/src/Makefile b/src/Makefile index 236c2e7..c2ab226 100644 --- a/src/Makefile +++ b/src/Makefile @@ -136,18 +136,18 @@ default: @echo "To build John the Ripper, type:" @echo " make clean SYSTEM" @echo "where SYSTEM can be one of the following:" - @echo "linux-x86-64i Linux, x86-64 with pre-built (icc) intrinsics" - @echo "linux-x86-64 Linux, x86-64 with SSE2 (best tested)" + @echo "linux-x86-64i Linux, x86-64 with pre-built intrinsics (best)" + @echo "linux-x86-64 Linux, x86-64 with SSE2" @echo "linux-x86-64-avx Linux, x86-64 with AVX (experimental)" @echo "linux-x86-64-xop Linux, x86-64 with AVX and XOP (experimental)" - @echo "linux-x86-64-icc Linux, x86-64 compiled with icc (best)" - @echo "linux-x86-64-clang Linux, x86-64 compiled with clang (good)" + @echo "linux-x86-64-icc Linux, x86-64 compiled with icc" + @echo "linux-x86-64-clang Linux, x86-64 compiled with clang" +# @echo "linux-x86-64-32-sse2i Linux, x86-64, 32-bit with pre-built intrinsics (for regression tests)" # @echo "linux-x86-64-32-sse2 Linux, x86-64, 32-bit with SSE2 (for regression tests)" -# @echo "linux-x86-64-32-sse2i Linux, x86-64, 32-bit with SSE2 intrinsics (for regression tests)" # @echo "linux-x86-64-32-mmx Linux, x86-64, 32-bit with MMX (for regression tests)" # @echo "linux-x86-64-32-any Linux, x86-64, 32-bit (for regression tests)" - @echo "linux-x86-sse2 Linux, x86 32-bit with SSE2 (best tested if 32-bit)" - @echo "linux-x86-sse2i Linux, x86 32-bit with SSE2 (32-bit, intrinsic)" + @echo "linux-x86-sse2i Linux, x86 32-bit with pre-built intrinsics (best)" + @echo "linux-x86-sse2 Linux, x86 32-bit with SSE2" @echo "linux-x86-mmx Linux, x86 32-bit with MMX (for old computers)" @echo "linux-x86-any Linux, x86 32-bit (for truly ancient computers)" @echo "linux-x86-avx Linux, x86 32-bit with AVX (experimental)" @@ -208,11 +208,11 @@ default: @echo "irix-mips32 IRIX, MIPS 32-bit" @echo "dos-djgpp-x86-mmx DOS, DJGPP, x86 with MMX" @echo "dos-djgpp-x86-any DOS, DJGPP, x86" - @echo "win32-cygwin-x86-sse2 Win32, Cygwin, x86 with SSE2 (best)" + @echo "win32-cygwin-x86-sse2i Win32, Cygwin, x86 with pre-built intrinsics (best)" @echo "win32-cygwin-x86-mmx Win32, Cygwin, x86 with MMX" @echo "win32-cygwin-x86-any Win32, Cygwin, x86" - @echo "win32-mingw-x86-sse2 Win32, MinGW, x86 with SSE2 (best)" - @echo "win32-mingw-x86-sse2i Win32, MinGW, x86 with SSE2 (intrinsic)" + @echo "win32-mingw-x86-sse2i Win32, MinGW, x86 with pre-built intrinsics (best)" + @echo "win32-mingw-x86-sse2 Win32, MinGW, x86 with SSE2" @echo "win32-mingw-x86-mmx Win32, MinGW, x86 with MMX" @echo "win32-mingw-x86-any Win32, MinGW, x86" @echo "beos-x86-sse2 BeOS, x86 with SSE2 (best)" @@ -266,8 +266,8 @@ linux-x86-64-icc: @echo "#define JOHN_BLD" '"'$@'"' > john_build_rule.h $(MAKE) $(PROJ) \ JOHN_OBJS="$(JOHN_OBJS) c3_fmt.o x86-64.o sse-intrinsics.o" \ - CFLAGS="-c -fast -I/usr/include -static-intel -DHAVE_CRYPT -DHAVE_DL $(ICCOMPFLAGS)" \ - LDFLAGS="-lm -lssl -lcrypto -ipo -no-prec-div -xHost -static-intel -lcrypt -ldl $(ICCOMPFLAGS) -s " \ + CFLAGS="-c -msse2 -axCORE-AVX2 -no-prec-div -ipo -I/usr/include -static-intel -DHAVE_CRYPT -DHAVE_DL $(ICCOMPFLAGS)" \ + LDFLAGS="-lm -lssl -lcrypto -ipo -static-intel -lcrypt -ldl $(ICCOMPFLAGS) -s " \ CPP="icc" CC="icc" AS="icc" LD="icc" linux-x86-64-32-sse2: @@ -319,7 +319,8 @@ linux-x86-sse2i: @echo "#define JOHN_BLD" '"'$@'"' > john_build_rule.h $(MAKE) $(PROJ) \ JOHN_OBJS="$(JOHN_OBJS) c3_fmt.o x86.o x86-sse.o sha1-mmx.o md5-mmx.o sse-intrinsics-32.S" \ - CFLAGS="$(CFLAGS) -DHAVE_CRYPT -DHAVE_DL -DUSING_ICC_S_FILE" \ + CFLAGS="$(CFLAGS) -msse2 -DHAVE_CRYPT -DHAVE_DL -DUSING_ICC_S_FILE" \ + ASFLAGS="$(ASFLAGS) -msse2" \ LDFLAGS="$(LDFLAGS) -lcrypt -ldl" linux-x86-mmx: @@ -975,9 +976,8 @@ dos-djgpp-x86-any: win32-cygwin-x86-sse2i: $(CP) x86-ssei.h arch.h @echo "#define JOHN_BLD" '"'$@'"' > john_build_rule.h - perl ../run/sse2i_winfix.pl $(MAKE) $(PROJ_WIN32) \ - JOHN_OBJS="sse-intrinsics-win32.o $(JOHN_OBJS) x86.o x86-sse.o " \ + JOHN_OBJS="sse-intrinsics-32.S $(JOHN_OBJS) x86.o x86-sse.o " \ CFLAGS="$(CFLAGS) -Wall -mpreferred-stack-boundary=4 -msse2 -m32 -DUSING_ICC_S_FILE" \ ASFLAGS="$(ASFLAGS) -DUNDERSCORES" @@ -1008,9 +1008,8 @@ win32-cygwin-x86-any: win32-mingw-x86-sse2i: $(CP) x86-ssei.h arch.h @echo "#define JOHN_BLD" '"'$@'"' > john_build_rule.h - perl ../run/sse2i_winfix.pl $(MAKE) $(PROJ_WIN32_MINGW) \ - JOHN_OBJS="sse-intrinsics-win32.o $(JOHN_OBJS) x86.o x86-sse.o " \ + JOHN_OBJS="sse-intrinsics-32.S $(JOHN_OBJS) x86.o x86-sse.o " \ CFLAGS="$(CFLAGS) -Wall -mpreferred-stack-boundary=4 -msse2 -m32 -DUSING_ICC_S_FILE" \ ASFLAGS="$(ASFLAGS) -DUNDERSCORES" @@ -1276,7 +1275,6 @@ clean: $(RM) ../run/john.exe john-macosx-* *.o *.bak core $(RM) detect bench generic.h arch.h tmp.s $(RM) fmt_registers.h fmt_externs.h john_build_rule.h - $(RM) sse-intrinsics-win32.S $(CP) $(NULL) Makefile.dep # For the time being, icc builds a better sse-intrinsics.S but this may @@ -1294,5 +1292,7 @@ sse-intrinsics.S: sse-intrinsics-32.S: $(LN) x86-ssei.h arch.h icc -m32 -S -o sse-intrinsics-32.S sse-intrinsics.c -O3 -fno-builtin -fno-verbose-asm -DUSING_ICC_S_FILE -Wall + perl ../run/sse2i_winfix.pl + mv sse-intrinsics-win32.S sse-intrinsics-32.S include Makefile.dep diff --git a/src/sse-intrinsics-32.S b/src/sse-intrinsics-32.S index c55c11f..a439352 100644 --- a/src/sse-intrinsics-32.S +++ b/src/sse-intrinsics-32.S @@ -1,4 +1,18 @@ .file "sse-intrinsics.c" + +#ifdef UNDERSCORES +#define memcpy _memcpy +#define memset _memset +#define strlen _strlen +#define MD5_Init _MD5_Init +#define MD5_Update _MD5_Update +#define MD5_Final _MD5_Final +#define SSEmd5body _SSEmd5body +#define SSESHA1body _SSESHA1body +#define SSEmd4body _SSEmd4body +#define md5cryptsse _md5cryptsse +#endif + .text ..TXTST0: # -- Begin sse_debug @@ -9,8 +23,6 @@ sse_debug: movl $1, debug ret .align 16,0x90 - .type sse_debug,@function - .size sse_debug,.-sse_debug .data # -- End sse_debug .text @@ -107,8 +119,6 @@ mmxput: movl $1, %edx jmp ..B2.6 .align 16,0x90 - .type mmxput,@function - .size mmxput,.-mmxput .data # -- End mmxput .text @@ -156,8 +166,6 @@ mmxput2: popl %esi ret .align 16,0x90 - .type mmxput2,@function - .size mmxput2,.-mmxput2 .data # -- End mmxput2 .text @@ -379,8 +387,6 @@ dispatch: popl %esi ret .align 16,0x90 - .type dispatch,@function - .size dispatch,.-dispatch .section .rodata, "a" .align 32 .align 32 @@ -793,8 +799,6 @@ mmxput3.: popl %esi ret .align 16,0x90 - .type mmxput3,@function - .size mmxput3,.-mmxput3 .data # -- End mmxput3 .text @@ -3816,8 +3820,6 @@ SSEmd5body.: addl $1564, %esp ret .align 16,0x90 - .type SSEmd5body,@function - .size SSEmd5body,.-SSEmd5body .data # -- End SSEmd5body .text @@ -5285,8 +5287,6 @@ md5cryptsse: testl %eax, %eax jmp ..B7.47 .align 16,0x90 - .type md5cryptsse,@function - .size md5cryptsse,.-md5cryptsse .section .rodata, "a" .space 28, 0x00 # pad .align 32 @@ -7525,8 +7525,6 @@ SSEmd4body: addl $1644, %esp ret .align 16,0x90 - .type SSEmd4body,@function - .size SSEmd4body,.-SSEmd4body .data # -- End SSEmd4body .text @@ -10854,8 +10852,6 @@ SSESHA1body: movdqa %xmm2, 112(%esp) jmp ..B9.5 .align 16,0x90 - .type SSESHA1body,@function - .size SSESHA1body,.-SSESHA1body .data # -- End SSESHA1body .bss @@ -10863,433 +10859,262 @@ SSESHA1body: .align 4 .globl debug debug: - .type debug,@object - .size debug,4 .space 4 # pad .section .rodata, "a" .space 12, 0x00 # pad .align 16 .L_2il0floatpacket.61: .long 0x67452301,0x67452301,0x67452301,0x67452301 - .type .L_2il0floatpacket.61,@object - .size .L_2il0floatpacket.61,16 .align 16 .L_2il0floatpacket.62: .long 0xefcdab89,0xefcdab89,0xefcdab89,0xefcdab89 - .type .L_2il0floatpacket.62,@object - .size .L_2il0floatpacket.62,16 .align 16 .L_2il0floatpacket.63: .long 0x98badcfe,0x98badcfe,0x98badcfe,0x98badcfe - .type .L_2il0floatpacket.63,@object - .size .L_2il0floatpacket.63,16 .align 16 .L_2il0floatpacket.64: .long 0x10325476,0x10325476,0x10325476,0x10325476 - .type .L_2il0floatpacket.64,@object - .size .L_2il0floatpacket.64,16 .align 16 .L_2il0floatpacket.65: .long 0xd76aa478,0xd76aa478,0xd76aa478,0xd76aa478 - .type .L_2il0floatpacket.65,@object - .size .L_2il0floatpacket.65,16 .align 16 .L_2il0floatpacket.66: .long 0xe8c7b756,0xe8c7b756,0xe8c7b756,0xe8c7b756 - .type .L_2il0floatpacket.66,@object - .size .L_2il0floatpacket.66,16 .align 16 .L_2il0floatpacket.67: .long 0x242070db,0x242070db,0x242070db,0x242070db - .type .L_2il0floatpacket.67,@object - .size .L_2il0floatpacket.67,16 .align 16 .L_2il0floatpacket.68: .long 0xc1bdceee,0xc1bdceee,0xc1bdceee,0xc1bdceee - .type .L_2il0floatpacket.68,@object - .size .L_2il0floatpacket.68,16 .align 16 .L_2il0floatpacket.69: .long 0xf57c0faf,0xf57c0faf,0xf57c0faf,0xf57c0faf - .type .L_2il0floatpacket.69,@object - .size .L_2il0floatpacket.69,16 .align 16 .L_2il0floatpacket.70: .long 0x4787c62a,0x4787c62a,0x4787c62a,0x4787c62a - .type .L_2il0floatpacket.70,@object - .size .L_2il0floatpacket.70,16 .align 16 .L_2il0floatpacket.71: .long 0xa8304613,0xa8304613,0xa8304613,0xa8304613 - .type .L_2il0floatpacket.71,@object - .size .L_2il0floatpacket.71,16 .align 16 .L_2il0floatpacket.72: .long 0xfd469501,0xfd469501,0xfd469501,0xfd469501 - .type .L_2il0floatpacket.72,@object - .size .L_2il0floatpacket.72,16 .align 16 .L_2il0floatpacket.73: .long 0x698098d8,0x698098d8,0x698098d8,0x698098d8 - .type .L_2il0floatpacket.73,@object - .size .L_2il0floatpacket.73,16 .align 16 .L_2il0floatpacket.74: .long 0x8b44f7af,0x8b44f7af,0x8b44f7af,0x8b44f7af - .type .L_2il0floatpacket.74,@object - .size .L_2il0floatpacket.74,16 .align 16 .L_2il0floatpacket.75: .long 0xffff5bb1,0xffff5bb1,0xffff5bb1,0xffff5bb1 - .type .L_2il0floatpacket.75,@object - .size .L_2il0floatpacket.75,16 .align 16 .L_2il0floatpacket.76: .long 0x895cd7be,0x895cd7be,0x895cd7be,0x895cd7be - .type .L_2il0floatpacket.76,@object - .size .L_2il0floatpacket.76,16 .align 16 .L_2il0floatpacket.77: .long 0x6b901122,0x6b901122,0x6b901122,0x6b901122 - .type .L_2il0floatpacket.77,@object - .size .L_2il0floatpacket.77,16 .align 16 .L_2il0floatpacket.78: .long 0xfd987193,0xfd987193,0xfd987193,0xfd987193 - .type .L_2il0floatpacket.78,@object - .size .L_2il0floatpacket.78,16 .align 16 .L_2il0floatpacket.79: .long 0xa679438e,0xa679438e,0xa679438e,0xa679438e - .type .L_2il0floatpacket.79,@object - .size .L_2il0floatpacket.79,16 .align 16 .L_2il0floatpacket.80: .long 0x49b40821,0x49b40821,0x49b40821,0x49b40821 - .type .L_2il0floatpacket.80,@object - .size .L_2il0floatpacket.80,16 .align 16 .L_2il0floatpacket.81: .long 0xf61e2562,0xf61e2562,0xf61e2562,0xf61e2562 - .type .L_2il0floatpacket.81,@object - .size .L_2il0floatpacket.81,16 .align 16 .L_2il0floatpacket.82: .long 0xc040b340,0xc040b340,0xc040b340,0xc040b340 - .type .L_2il0floatpacket.82,@object - .size .L_2il0floatpacket.82,16 .align 16 .L_2il0floatpacket.83: .long 0x265e5a51,0x265e5a51,0x265e5a51,0x265e5a51 - .type .L_2il0floatpacket.83,@object - .size .L_2il0floatpacket.83,16 .align 16 .L_2il0floatpacket.84: .long 0xe9b6c7aa,0xe9b6c7aa,0xe9b6c7aa,0xe9b6c7aa - .type .L_2il0floatpacket.84,@object - .size .L_2il0floatpacket.84,16 .align 16 .L_2il0floatpacket.85: .long 0xd62f105d,0xd62f105d,0xd62f105d,0xd62f105d - .type .L_2il0floatpacket.85,@object - .size .L_2il0floatpacket.85,16 .align 16 .L_2il0floatpacket.86: .long 0x02441453,0x02441453,0x02441453,0x02441453 - .type .L_2il0floatpacket.86,@object - .size .L_2il0floatpacket.86,16 .align 16 .L_2il0floatpacket.87: .long 0xd8a1e681,0xd8a1e681,0xd8a1e681,0xd8a1e681 - .type .L_2il0floatpacket.87,@object - .size .L_2il0floatpacket.87,16 .align 16 .L_2il0floatpacket.88: .long 0xe7d3fbc8,0xe7d3fbc8,0xe7d3fbc8,0xe7d3fbc8 - .type .L_2il0floatpacket.88,@object - .size .L_2il0floatpacket.88,16 .align 16 .L_2il0floatpacket.89: .long 0x21e1cde6,0x21e1cde6,0x21e1cde6,0x21e1cde6 - .type .L_2il0floatpacket.89,@object - .size .L_2il0floatpacket.89,16 .align 16 .L_2il0floatpacket.90: .long 0xc33707d6,0xc33707d6,0xc33707d6,0xc33707d6 - .type .L_2il0floatpacket.90,@object - .size .L_2il0floatpacket.90,16 .align 16 .L_2il0floatpacket.91: .long 0xf4d50d87,0xf4d50d87,0xf4d50d87,0xf4d50d87 - .type .L_2il0floatpacket.91,@object - .size .L_2il0floatpacket.91,16 .align 16 .L_2il0floatpacket.92: .long 0x455a14ed,0x455a14ed,0x455a14ed,0x455a14ed - .type .L_2il0floatpacket.92,@object - .size .L_2il0floatpacket.92,16 .align 16 .L_2il0floatpacket.93: .long 0xa9e3e905,0xa9e3e905,0xa9e3e905,0xa9e3e905 - .type .L_2il0floatpacket.93,@object - .size .L_2il0floatpacket.93,16 .align 16 .L_2il0floatpacket.94: .long 0xfcefa3f8,0xfcefa3f8,0xfcefa3f8,0xfcefa3f8 - .type .L_2il0floatpacket.94,@object - .size .L_2il0floatpacket.94,16 .align 16 .L_2il0floatpacket.95: .long 0x676f02d9,0x676f02d9,0x676f02d9,0x676f02d9 - .type .L_2il0floatpacket.95,@object - .size .L_2il0floatpacket.95,16 .align 16 .L_2il0floatpacket.96: .long 0x8d2a4c8a,0x8d2a4c8a,0x8d2a4c8a,0x8d2a4c8a - .type .L_2il0floatpacket.96,@object - .size .L_2il0floatpacket.96,16 .align 16 .L_2il0floatpacket.97: .long 0xfffa3942,0xfffa3942,0xfffa3942,0xfffa3942 - .type .L_2il0floatpacket.97,@object - .size .L_2il0floatpacket.97,16 .align 16 .L_2il0floatpacket.98: .long 0x8771f681,0x8771f681,0x8771f681,0x8771f681 - .type .L_2il0floatpacket.98,@object - .size .L_2il0floatpacket.98,16 .align 16 .L_2il0floatpacket.99: .long 0x6d9d6122,0x6d9d6122,0x6d9d6122,0x6d9d6122 - .type .L_2il0floatpacket.99,@object - .size .L_2il0floatpacket.99,16 .align 16 .L_2il0floatpacket.100: .long 0xfde5380c,0xfde5380c,0xfde5380c,0xfde5380c - .type .L_2il0floatpacket.100,@object - .size .L_2il0floatpacket.100,16 .align 16 .L_2il0floatpacket.101: .long 0xa4beea44,0xa4beea44,0xa4beea44,0xa4beea44 - .type .L_2il0floatpacket.101,@object - .size .L_2il0floatpacket.101,16 .align 16 .L_2il0floatpacket.102: .long 0x4bdecfa9,0x4bdecfa9,0x4bdecfa9,0x4bdecfa9 - .type .L_2il0floatpacket.102,@object - .size .L_2il0floatpacket.102,16 .align 16 .L_2il0floatpacket.103: .long 0xf6bb4b60,0xf6bb4b60,0xf6bb4b60,0xf6bb4b60 - .type .L_2il0floatpacket.103,@object - .size .L_2il0floatpacket.103,16 .align 16 .L_2il0floatpacket.104: .long 0xbebfbc70,0xbebfbc70,0xbebfbc70,0xbebfbc70 - .type .L_2il0floatpacket.104,@object - .size .L_2il0floatpacket.104,16 .align 16 .L_2il0floatpacket.105: .long 0x289b7ec6,0x289b7ec6,0x289b7ec6,0x289b7ec6 - .type .L_2il0floatpacket.105,@object - .size .L_2il0floatpacket.105,16 .align 16 .L_2il0floatpacket.106: .long 0xeaa127fa,0xeaa127fa,0xeaa127fa,0xeaa127fa - .type .L_2il0floatpacket.106,@object - .size .L_2il0floatpacket.106,16 .align 16 .L_2il0floatpacket.107: .long 0xd4ef3085,0xd4ef3085,0xd4ef3085,0xd4ef3085 - .type .L_2il0floatpacket.107,@object - .size .L_2il0floatpacket.107,16 .align 16 .L_2il0floatpacket.108: .long 0x04881d05,0x04881d05,0x04881d05,0x04881d05 - .type .L_2il0floatpacket.108,@object - .size .L_2il0floatpacket.108,16 .align 16 .L_2il0floatpacket.109: .long 0xd9d4d039,0xd9d4d039,0xd9d4d039,0xd9d4d039 - .type .L_2il0floatpacket.109,@object - .size .L_2il0floatpacket.109,16 .align 16 .L_2il0floatpacket.110: .long 0xe6db99e5,0xe6db99e5,0xe6db99e5,0xe6db99e5 - .type .L_2il0floatpacket.110,@object - .size .L_2il0floatpacket.110,16 .align 16 .L_2il0floatpacket.111: .long 0x1fa27cf8,0x1fa27cf8,0x1fa27cf8,0x1fa27cf8 - .type .L_2il0floatpacket.111,@object - .size .L_2il0floatpacket.111,16 .align 16 .L_2il0floatpacket.112: .long 0xc4ac5665,0xc4ac5665,0xc4ac5665,0xc4ac5665 - .type .L_2il0floatpacket.112,@object - .size .L_2il0floatpacket.112,16 .align 16 .L_2il0floatpacket.113: .long 0xf4292244,0xf4292244,0xf4292244,0xf4292244 - .type .L_2il0floatpacket.113,@object - .size .L_2il0floatpacket.113,16 .align 16 .L_2il0floatpacket.114: .long 0x432aff97,0x432aff97,0x432aff97,0x432aff97 - .type .L_2il0floatpacket.114,@object - .size .L_2il0floatpacket.114,16 .align 16 .L_2il0floatpacket.115: .long 0xab9423a7,0xab9423a7,0xab9423a7,0xab9423a7 - .type .L_2il0floatpacket.115,@object - .size .L_2il0floatpacket.115,16 .align 16 .L_2il0floatpacket.116: .long 0xfc93a039,0xfc93a039,0xfc93a039,0xfc93a039 - .type .L_2il0floatpacket.116,@object - .size .L_2il0floatpacket.116,16 .align 16 .L_2il0floatpacket.117: .long 0x655b59c3,0x655b59c3,0x655b59c3,0x655b59c3 - .type .L_2il0floatpacket.117,@object - .size .L_2il0floatpacket.117,16 .align 16 .L_2il0floatpacket.118: .long 0x8f0ccc92,0x8f0ccc92,0x8f0ccc92,0x8f0ccc92 - .type .L_2il0floatpacket.118,@object - .size .L_2il0floatpacket.118,16 .align 16 .L_2il0floatpacket.119: .long 0xffeff47d,0xffeff47d,0xffeff47d,0xffeff47d - .type .L_2il0floatpacket.119,@object - .size .L_2il0floatpacket.119,16 .align 16 .L_2il0floatpacket.120: .long 0x85845dd1,0x85845dd1,0x85845dd1,0x85845dd1 - .type .L_2il0floatpacket.120,@object - .size .L_2il0floatpacket.120,16 .align 16 .L_2il0floatpacket.121: .long 0x6fa87e4f,0x6fa87e4f,0x6fa87e4f,0x6fa87e4f - .type .L_2il0floatpacket.121,@object - .size .L_2il0floatpacket.121,16 .align 16 .L_2il0floatpacket.122: .long 0xfe2ce6e0,0xfe2ce6e0,0xfe2ce6e0,0xfe2ce6e0 - .type .L_2il0floatpacket.122,@object - .size .L_2il0floatpacket.122,16 .align 16 .L_2il0floatpacket.123: .long 0xa3014314,0xa3014314,0xa3014314,0xa3014314 - .type .L_2il0floatpacket.123,@object - .size .L_2il0floatpacket.123,16 .align 16 .L_2il0floatpacket.124: .long 0x4e0811a1,0x4e0811a1,0x4e0811a1,0x4e0811a1 - .type .L_2il0floatpacket.124,@object - .size .L_2il0floatpacket.124,16 .align 16 .L_2il0floatpacket.125: .long 0xf7537e82,0xf7537e82,0xf7537e82,0xf7537e82 - .type .L_2il0floatpacket.125,@object - .size .L_2il0floatpacket.125,16 .align 16 .L_2il0floatpacket.126: .long 0xbd3af235,0xbd3af235,0xbd3af235,0xbd3af235 - .type .L_2il0floatpacket.126,@object - .size .L_2il0floatpacket.126,16 .align 16 .L_2il0floatpacket.127: .long 0x2ad7d2bb,0x2ad7d2bb,0x2ad7d2bb,0x2ad7d2bb - .type .L_2il0floatpacket.127,@object - .size .L_2il0floatpacket.127,16 .align 16 .L_2il0floatpacket.128: .long 0xeb86d391,0xeb86d391,0xeb86d391,0xeb86d391 - .type .L_2il0floatpacket.128,@object - .size .L_2il0floatpacket.128,16 .align 16 .L_2il0floatpacket.474: .long 0x67452301,0x67452301,0x67452301,0x67452301 - .type .L_2il0floatpacket.474,@object - .size .L_2il0floatpacket.474,16 .align 16 .L_2il0floatpacket.475: .long 0xefcdab89,0xefcdab89,0xefcdab89,0xefcdab89 - .type .L_2il0floatpacket.475,@object - .size .L_2il0floatpacket.475,16 .align 16 .L_2il0floatpacket.476: .long 0x98badcfe,0x98badcfe,0x98badcfe,0x98badcfe - .type .L_2il0floatpacket.476,@object - .size .L_2il0floatpacket.476,16 .align 16 .L_2il0floatpacket.477: .long 0x10325476,0x10325476,0x10325476,0x10325476 - .type .L_2il0floatpacket.477,@object - .size .L_2il0floatpacket.477,16 .align 16 .L_2il0floatpacket.478: .long 0x5a827999,0x5a827999,0x5a827999,0x5a827999 - .type .L_2il0floatpacket.478,@object - .size .L_2il0floatpacket.478,16 .align 16 .L_2il0floatpacket.479: .long 0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1 - .type .L_2il0floatpacket.479,@object - .size .L_2il0floatpacket.479,16 .align 16 .L_2il0floatpacket.498: .long 0x67452301,0x67452301,0x67452301,0x67452301 - .type .L_2il0floatpacket.498,@object - .size .L_2il0floatpacket.498,16 .align 16 .L_2il0floatpacket.499: .long 0xefcdab89,0xefcdab89,0xefcdab89,0xefcdab89 - .type .L_2il0floatpacket.499,@object - .size .L_2il0floatpacket.499,16 .align 16 .L_2il0floatpacket.500: .long 0x98badcfe,0x98badcfe,0x98badcfe,0x98badcfe - .type .L_2il0floatpacket.500,@object - .size .L_2il0floatpacket.500,16 .align 16 .L_2il0floatpacket.501: .long 0x10325476,0x10325476,0x10325476,0x10325476 - .type .L_2il0floatpacket.501,@object - .size .L_2il0floatpacket.501,16 .align 16 .L_2il0floatpacket.502: .long 0xc3d2e1f0,0xc3d2e1f0,0xc3d2e1f0,0xc3d2e1f0 - .type .L_2il0floatpacket.502,@object - .size .L_2il0floatpacket.502,16 .align 16 .L_2il0floatpacket.503: .long 0x5a827999,0x5a827999,0x5a827999,0x5a827999 - .type .L_2il0floatpacket.503,@object - .size .L_2il0floatpacket.503,16 .align 16 .L_2il0floatpacket.504: .long 0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1 - .type .L_2il0floatpacket.504,@object - .size .L_2il0floatpacket.504,16 .align 16 .L_2il0floatpacket.505: .long 0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc - .type .L_2il0floatpacket.505,@object - .size .L_2il0floatpacket.505,16 .align 16 .L_2il0floatpacket.506: .long 0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6 - .type .L_2il0floatpacket.506,@object - .size .L_2il0floatpacket.506,16 .align 4 .L_2__STRING.3: .byte 0 .byte 0 - .type .L_2__STRING.3,@object - .size .L_2__STRING.3,2 - .section .rodata.str1.4, "aMS",@progbits,1 .align 4 .align 4 .L_2__STRING.2: @@ -11297,8 +11122,6 @@ debug: .byte 49 .byte 36 .byte 0 - .type .L_2__STRING.2,@object - .size .L_2__STRING.2,4 .align 4 .L_2__STRING.1: .byte 36 @@ -11308,8 +11131,5 @@ debug: .byte 49 .byte 36 .byte 0 - .type .L_2__STRING.1,@object - .size .L_2__STRING.1,7 .data - .section .note.GNU-stack, "" # End diff --git a/src/x86-64.h b/src/x86-64.h index 30cd502..3caa26f 100644 --- a/src/x86-64.h +++ b/src/x86-64.h @@ -163,7 +163,10 @@ #define MD4_N_STR "12x" #endif -#ifdef __GNUC__ +#if defined(__INTEL_COMPILER) || defined(USING_ICC_S_FILE) +#define SHA1_SSE_PARA 2 +#define SHA1_N_STR "8x" +#elif defined(__GNUC__) #define SHA1_SSE_PARA 2 #define SHA1_N_STR "8x" #else diff --git a/src/x86-ssei.h b/src/x86-ssei.h index d1c9be8..fc88e35 100644 --- a/src/x86-ssei.h +++ b/src/x86-ssei.h @@ -144,10 +144,10 @@ #define MD4_N_STR "12x" #endif -#if defined (_MSC_VER) +#if defined(__INTEL_COMPILER) || defined(USING_ICC_S_FILE) #define SHA1_SSE_PARA 2 #define SHA1_N_STR "8x" -#elif defined(__GNUC__) +#elif defined(__GNUC__) || defined (_MSC_VER) #define SHA1_SSE_PARA 2 #define SHA1_N_STR "8x" #else -- 1.7.5.4