diff options
Diffstat (limited to 'sysdeps/x86_64/strcmp.S')
| -rw-r--r-- | sysdeps/x86_64/strcmp.S | 415 |
1 files changed, 211 insertions, 204 deletions
diff --git a/sysdeps/x86_64/strcmp.S b/sysdeps/x86_64/strcmp.S index 650ec173b6..ac3fe14679 100644 --- a/sysdeps/x86_64/strcmp.S +++ b/sysdeps/x86_64/strcmp.S @@ -1,5 +1,5 @@ /* Highly optimized version for x86-64. - Copyright (C) 1999, 2000, 2002, 2003, 2005, 2009 + Copyright (C) 1999, 2000, 2002, 2003, 2005, 2009, 2010 Free Software Foundation, Inc. This file is part of the GNU C Library. Based on i686 version contributed by Ulrich Drepper @@ -33,6 +33,13 @@ #endif #ifdef USE_AS_STRNCMP +/* The simplified code below is not set up to handle strncmp() so far. + Should this become necessary it has to be implemented. For now + just report the problem. */ +# ifdef NOT_IN_lib +# error "strncmp not implemented so far" +# endif + /* Since the counter, %r11, is unsigned, we branch to strcmp_exitz if the new counter > the old one or is 0. */ # define UPDATE_STRNCMP_COUNTER \ @@ -54,7 +61,7 @@ #ifndef USE_SSSE3 .text #else - .section .text.ssse3,"ax",@progbits + .section .text.ssse3,"ax",@progbits #endif ENTRY (BP_SYM (STRCMP)) @@ -80,13 +87,13 @@ END (BP_SYM (STRCMP)) /* * This implementation uses SSE to compare up to 16 bytes at a time. */ -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP test %rdx, %rdx je LABEL(strcmp_exitz) cmp $1, %rdx je LABEL(Byte0) mov %rdx, %r11 -#endif +# endif mov %esi, %ecx mov %edi, %eax /* Use 64bit AND here to avoid long NOP padding. */ @@ -107,10 +114,10 @@ END (BP_SYM (STRCMP)) pmovmskb %xmm1, %edx sub $0xffff, %edx /* if first 16 bytes are same, edx == 0xffff */ jnz LABEL(less16bytes) /* If not, find different value or null char */ -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub $16, %r11 jbe LABEL(strcmp_exitz) /* finish comparision */ -#endif +# endif add $16, %rsi /* prepare to search next 16 bytes */ add $16, %rdi /* prepare to search next 16 bytes */ @@ -184,10 +191,10 @@ LABEL(loop_ashr_0): sub $0xffff, %edx jnz LABEL(exit) /* mismatch or null char seen */ -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub $16, %r11 jbe LABEL(strcmp_exitz) -#endif +# endif add $16, %rcx movdqa (%rsi, %rcx), %xmm1 movdqa (%rdi, %rcx), %xmm2 @@ -198,10 +205,10 @@ LABEL(loop_ashr_0): pmovmskb %xmm1, %edx sub $0xffff, %edx jnz LABEL(exit) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub $16, %r11 jbe LABEL(strcmp_exitz) -#endif +# endif add $16, %rcx jmp LABEL(loop_ashr_0) @@ -249,13 +256,13 @@ LABEL(gobble_ashr_1): movdqa (%rdi, %rcx), %xmm2 movdqa %xmm2, %xmm4 /* store for next cycle */ -#ifndef USE_SSSE3 +# ifndef USE_SSSE3 psrldq $1, %xmm3 pslldq $15, %xmm2 por %xmm3, %xmm2 /* merge into one 16byte value */ -#else +# else palignr $1, %xmm3, %xmm2 /* merge into one 16byte value */ -#endif +# endif pcmpeqb %xmm1, %xmm0 pcmpeqb %xmm2, %xmm1 @@ -264,10 +271,10 @@ LABEL(gobble_ashr_1): sub $0xffff, %edx jnz LABEL(exit) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub $16, %r11 jbe LABEL(strcmp_exitz) -#endif +# endif add $16, %rcx movdqa %xmm4, %xmm3 @@ -278,13 +285,13 @@ LABEL(gobble_ashr_1): movdqa (%rdi, %rcx), %xmm2 movdqa %xmm2, %xmm4 /* store for next cycle */ -#ifndef USE_SSSE3 +# ifndef USE_SSSE3 psrldq $1, %xmm3 pslldq $15, %xmm2 por %xmm3, %xmm2 /* merge into one 16byte value */ -#else +# else palignr $1, %xmm3, %xmm2 /* merge into one 16byte value */ -#endif +# endif pcmpeqb %xmm1, %xmm0 pcmpeqb %xmm2, %xmm1 @@ -293,10 +300,10 @@ LABEL(gobble_ashr_1): sub $0xffff, %edx jnz LABEL(exit) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub $16, %r11 jbe LABEL(strcmp_exitz) -#endif +# endif add $16, %rcx movdqa %xmm4, %xmm3 jmp LABEL(loop_ashr_1) @@ -312,10 +319,10 @@ LABEL(nibble_ashr_1): test $0xfffe, %edx jnz LABEL(ashr_1_exittail) /* find null char*/ -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP cmp $14, %r11 jbe LABEL(ashr_1_exittail) -#endif +# endif pxor %xmm0, %xmm0 sub $0x1000, %r10 /* substract 4K from %r10 */ @@ -334,7 +341,7 @@ LABEL(ashr_1_exittail): /* * The following cases will be handled by ashr_2 - * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case + * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case * n(14~15) n -14 1(15 +(n-14) - n) ashr_2 */ .p2align 4 @@ -376,13 +383,13 @@ LABEL(gobble_ashr_2): movdqa (%rdi, %rcx), %xmm2 movdqa %xmm2, %xmm4 -#ifndef USE_SSSE3 +# ifndef USE_SSSE3 psrldq $2, %xmm3 pslldq $14, %xmm2 por %xmm3, %xmm2 /* merge into one 16byte value */ -#else +# else palignr $2, %xmm3, %xmm2 /* merge into one 16byte value */ -#endif +# endif pcmpeqb %xmm1, %xmm0 pcmpeqb %xmm2, %xmm1 @@ -391,10 +398,10 @@ LABEL(gobble_ashr_2): sub $0xffff, %edx jnz LABEL(exit) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub $16, %r11 jbe LABEL(strcmp_exitz) -#endif +# endif add $16, %rcx movdqa %xmm4, %xmm3 @@ -406,13 +413,13 @@ LABEL(gobble_ashr_2): movdqa (%rdi, %rcx), %xmm2 movdqa %xmm2, %xmm4 -#ifndef USE_SSSE3 +# ifndef USE_SSSE3 psrldq $2, %xmm3 pslldq $14, %xmm2 por %xmm3, %xmm2 /* merge into one 16byte value */ -#else +# else palignr $2, %xmm3, %xmm2 /* merge into one 16byte value */ -#endif +# endif pcmpeqb %xmm1, %xmm0 pcmpeqb %xmm2, %xmm1 @@ -421,10 +428,10 @@ LABEL(gobble_ashr_2): sub $0xffff, %edx jnz LABEL(exit) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub $16, %r11 jbe LABEL(strcmp_exitz) -#endif +# endif add $16, %rcx movdqa %xmm4, %xmm3 @@ -437,10 +444,10 @@ LABEL(nibble_ashr_2): test $0xfffc, %edx jnz LABEL(ashr_2_exittail) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP cmp $13, %r11 jbe LABEL(ashr_2_exittail) -#endif +# endif pxor %xmm0, %xmm0 sub $0x1000, %r10 @@ -498,13 +505,13 @@ LABEL(gobble_ashr_3): movdqa (%rdi, %rcx), %xmm2 movdqa %xmm2, %xmm4 -#ifndef USE_SSSE3 +# ifndef USE_SSSE3 psrldq $3, %xmm3 pslldq $13, %xmm2 por %xmm3, %xmm2 /* merge into one 16byte value */ -#else +# else palignr $3, %xmm3, %xmm2 /* merge into one 16byte value */ -#endif +# endif pcmpeqb %xmm1, %xmm0 pcmpeqb %xmm2, %xmm1 @@ -513,10 +520,10 @@ LABEL(gobble_ashr_3): sub $0xffff, %edx jnz LABEL(exit) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub $16, %r11 jbe LABEL(strcmp_exitz) -#endif +# endif add $16, %rcx movdqa %xmm4, %xmm3 @@ -528,13 +535,13 @@ LABEL(gobble_ashr_3): movdqa (%rdi, %rcx), %xmm2 movdqa %xmm2, %xmm4 -#ifndef USE_SSSE3 +# ifndef USE_SSSE3 psrldq $3, %xmm3 pslldq $13, %xmm2 por %xmm3, %xmm2 /* merge into one 16byte value */ -#else +# else palignr $3, %xmm3, %xmm2 /* merge into one 16byte value */ -#endif +# endif pcmpeqb %xmm1, %xmm0 pcmpeqb %xmm2, %xmm1 @@ -543,10 +550,10 @@ LABEL(gobble_ashr_3): sub $0xffff, %edx jnz LABEL(exit) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub $16, %r11 jbe LABEL(strcmp_exitz) -#endif +# endif add $16, %rcx movdqa %xmm4, %xmm3 @@ -559,10 +566,10 @@ LABEL(nibble_ashr_3): test $0xfff8, %edx jnz LABEL(ashr_3_exittail) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP cmp $12, %r11 jbe LABEL(ashr_3_exittail) -#endif +# endif pxor %xmm0, %xmm0 sub $0x1000, %r10 @@ -620,13 +627,13 @@ LABEL(gobble_ashr_4): movdqa (%rdi, %rcx), %xmm2 movdqa %xmm2, %xmm4 -#ifndef USE_SSSE3 +# ifndef USE_SSSE3 psrldq $4, %xmm3 pslldq $12, %xmm2 por %xmm3, %xmm2 /* merge into one 16byte value */ -#else +# else palignr $4, %xmm3, %xmm2 /* merge into one 16byte value */ -#endif +# endif pcmpeqb %xmm1, %xmm0 pcmpeqb %xmm2, %xmm1 @@ -635,10 +642,10 @@ LABEL(gobble_ashr_4): sub $0xffff, %edx jnz LABEL(exit) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub $16, %r11 jbe LABEL(strcmp_exitz) -#endif +# endif add $16, %rcx movdqa %xmm4, %xmm3 @@ -650,13 +657,13 @@ LABEL(gobble_ashr_4): movdqa (%rdi, %rcx), %xmm2 movdqa %xmm2, %xmm4 -#ifndef USE_SSSE3 +# ifndef USE_SSSE3 psrldq $4, %xmm3 pslldq $12, %xmm2 por %xmm3, %xmm2 /* merge into one 16byte value */ -#else +# else palignr $4, %xmm3, %xmm2 /* merge into one 16byte value */ -#endif +# endif pcmpeqb %xmm1, %xmm0 pcmpeqb %xmm2, %xmm1 @@ -665,10 +672,10 @@ LABEL(gobble_ashr_4): sub $0xffff, %edx jnz LABEL(exit) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub $16, %r11 jbe LABEL(strcmp_exitz) -#endif +# endif add $16, %rcx movdqa %xmm4, %xmm3 @@ -681,10 +688,10 @@ LABEL(nibble_ashr_4): test $0xfff0, %edx jnz LABEL(ashr_4_exittail) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP cmp $11, %r11 jbe LABEL(ashr_4_exittail) -#endif +# endif pxor %xmm0, %xmm0 sub $0x1000, %r10 @@ -700,7 +707,7 @@ LABEL(ashr_4_exittail): /* * The following cases will be handled by ashr_5 * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case - * n(11~15) n - 11 4(15 +(n-11) - n) ashr_5 + * n(11~15) n - 11 4(15 +(n-11) - n) ashr_5 */ .p2align 4 LABEL(ashr_5): @@ -742,13 +749,13 @@ LABEL(gobble_ashr_5): movdqa (%rdi, %rcx), %xmm2 movdqa %xmm2, %xmm4 -#ifndef USE_SSSE3 +# ifndef USE_SSSE3 psrldq $5, %xmm3 pslldq $11, %xmm2 por %xmm3, %xmm2 /* merge into one 16byte value */ -#else +# else palignr $5, %xmm3, %xmm2 /* merge into one 16byte value */ -#endif +# endif pcmpeqb %xmm1, %xmm0 pcmpeqb %xmm2, %xmm1 @@ -757,10 +764,10 @@ LABEL(gobble_ashr_5): sub $0xffff, %edx jnz LABEL(exit) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub $16, %r11 jbe LABEL(strcmp_exitz) -#endif +# endif add $16, %rcx movdqa %xmm4, %xmm3 @@ -772,13 +779,13 @@ LABEL(gobble_ashr_5): movdqa (%rdi, %rcx), %xmm2 movdqa %xmm2, %xmm4 -#ifndef USE_SSSE3 +# ifndef USE_SSSE3 psrldq $5, %xmm3 pslldq $11, %xmm2 por %xmm3, %xmm2 /* merge into one 16byte value */ -#else +# else palignr $5, %xmm3, %xmm2 /* merge into one 16byte value */ -#endif +# endif pcmpeqb %xmm1, %xmm0 pcmpeqb %xmm2, %xmm1 @@ -787,10 +794,10 @@ LABEL(gobble_ashr_5): sub $0xffff, %edx jnz LABEL(exit) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub $16, %r11 jbe LABEL(strcmp_exitz) -#endif +# endif add $16, %rcx movdqa %xmm4, %xmm3 @@ -803,10 +810,10 @@ LABEL(nibble_ashr_5): test $0xffe0, %edx jnz LABEL(ashr_5_exittail) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP cmp $10, %r11 jbe LABEL(ashr_5_exittail) -#endif +# endif pxor %xmm0, %xmm0 sub $0x1000, %r10 @@ -822,7 +829,7 @@ LABEL(ashr_5_exittail): /* * The following cases will be handled by ashr_6 * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case - * n(10~15) n - 10 5(15 +(n-10) - n) ashr_6 + * n(10~15) n - 10 5(15 +(n-10) - n) ashr_6 */ .p2align 4 LABEL(ashr_6): @@ -864,13 +871,13 @@ LABEL(gobble_ashr_6): movdqa (%rdi, %rcx), %xmm2 movdqa %xmm2, %xmm4 -#ifndef USE_SSSE3 +# ifndef USE_SSSE3 psrldq $6, %xmm3 pslldq $10, %xmm2 por %xmm3, %xmm2 /* merge into one 16byte value */ -#else +# else palignr $6, %xmm3, %xmm2 /* merge into one 16byte value */ -#endif +# endif pcmpeqb %xmm1, %xmm0 pcmpeqb %xmm2, %xmm1 @@ -879,10 +886,10 @@ LABEL(gobble_ashr_6): sub $0xffff, %edx jnz LABEL(exit) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub $16, %r11 jbe LABEL(strcmp_exitz) -#endif +# endif add $16, %rcx movdqa %xmm4, %xmm3 @@ -894,13 +901,13 @@ LABEL(gobble_ashr_6): movdqa (%rdi, %rcx), %xmm2 movdqa %xmm2, %xmm4 -#ifndef USE_SSSE3 +# ifndef USE_SSSE3 psrldq $6, %xmm3 pslldq $10, %xmm2 por %xmm3, %xmm2 /* merge into one 16byte value */ -#else +# else palignr $6, %xmm3, %xmm2 /* merge into one 16byte value */ -#endif +# endif pcmpeqb %xmm1, %xmm0 pcmpeqb %xmm2, %xmm1 @@ -909,10 +916,10 @@ LABEL(gobble_ashr_6): sub $0xffff, %edx jnz LABEL(exit) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub $16, %r11 jbe LABEL(strcmp_exitz) -#endif +# endif add $16, %rcx movdqa %xmm4, %xmm3 @@ -925,10 +932,10 @@ LABEL(nibble_ashr_6): test $0xffc0, %edx jnz LABEL(ashr_6_exittail) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP cmp $9, %r11 jbe LABEL(ashr_6_exittail) -#endif +# endif pxor %xmm0, %xmm0 sub $0x1000, %r10 @@ -944,7 +951,7 @@ LABEL(ashr_6_exittail): /* * The following cases will be handled by ashr_7 * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case - * n(9~15) n - 9 6(15 +(n - 9) - n) ashr_7 + * n(9~15) n - 9 6(15 +(n - 9) - n) ashr_7 */ .p2align 4 LABEL(ashr_7): @@ -986,13 +993,13 @@ LABEL(gobble_ashr_7): movdqa (%rdi, %rcx), %xmm2 movdqa %xmm2, %xmm4 -#ifndef USE_SSSE3 +# ifndef USE_SSSE3 psrldq $7, %xmm3 pslldq $9, %xmm2 por %xmm3, %xmm2 /* merge into one 16byte value */ -#else +# else palignr $7, %xmm3, %xmm2 /* merge into one 16byte value */ -#endif +# endif pcmpeqb %xmm1, %xmm0 pcmpeqb %xmm2, %xmm1 @@ -1001,10 +1008,10 @@ LABEL(gobble_ashr_7): sub $0xffff, %edx jnz LABEL(exit) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub $16, %r11 jbe LABEL(strcmp_exitz) -#endif +# endif add $16, %rcx movdqa %xmm4, %xmm3 @@ -1016,13 +1023,13 @@ LABEL(gobble_ashr_7): movdqa (%rdi, %rcx), %xmm2 movdqa %xmm2, %xmm4 -#ifndef USE_SSSE3 +# ifndef USE_SSSE3 psrldq $7, %xmm3 pslldq $9, %xmm2 por %xmm3, %xmm2 /* merge into one 16byte value */ -#else +# else palignr $7, %xmm3, %xmm2 /* merge into one 16byte value */ -#endif +# endif pcmpeqb %xmm1, %xmm0 pcmpeqb %xmm2, %xmm1 @@ -1031,10 +1038,10 @@ LABEL(gobble_ashr_7): sub $0xffff, %edx jnz LABEL(exit) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub $16, %r11 jbe LABEL(strcmp_exitz) -#endif +# endif add $16, %rcx movdqa %xmm4, %xmm3 @@ -1047,10 +1054,10 @@ LABEL(nibble_ashr_7): test $0xff80, %edx jnz LABEL(ashr_7_exittail) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP cmp $8, %r11 jbe LABEL(ashr_7_exittail) -#endif +# endif pxor %xmm0, %xmm0 sub $0x1000, %r10 @@ -1066,7 +1073,7 @@ LABEL(ashr_7_exittail): /* * The following cases will be handled by ashr_8 * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case - * n(8~15) n - 8 7(15 +(n - 8) - n) ashr_8 + * n(8~15) n - 8 7(15 +(n - 8) - n) ashr_8 */ .p2align 4 LABEL(ashr_8): @@ -1108,13 +1115,13 @@ LABEL(gobble_ashr_8): movdqa (%rdi, %rcx), %xmm2 movdqa %xmm2, %xmm4 -#ifndef USE_SSSE3 +# ifndef USE_SSSE3 psrldq $8, %xmm3 pslldq $8, %xmm2 por %xmm3, %xmm2 /* merge into one 16byte value */ -#else +# else palignr $8, %xmm3, %xmm2 /* merge into one 16byte value */ -#endif +# endif pcmpeqb %xmm1, %xmm0 pcmpeqb %xmm2, %xmm1 @@ -1123,10 +1130,10 @@ LABEL(gobble_ashr_8): sub $0xffff, %edx jnz LABEL(exit) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub $16, %r11 jbe LABEL(strcmp_exitz) -#endif +# endif add $16, %rcx movdqa %xmm4, %xmm3 @@ -1138,13 +1145,13 @@ LABEL(gobble_ashr_8): movdqa (%rdi, %rcx), %xmm2 movdqa %xmm2, %xmm4 -#ifndef USE_SSSE3 +# ifndef USE_SSSE3 psrldq $8, %xmm3 pslldq $8, %xmm2 por %xmm3, %xmm2 /* merge into one 16byte value */ -#else +# else palignr $8, %xmm3, %xmm2 /* merge into one 16byte value */ -#endif +# endif pcmpeqb %xmm1, %xmm0 pcmpeqb %xmm2, %xmm1 @@ -1153,10 +1160,10 @@ LABEL(gobble_ashr_8): sub $0xffff, %edx jnz LABEL(exit) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub $16, %r11 jbe LABEL(strcmp_exitz) -#endif +# endif add $16, %rcx movdqa %xmm4, %xmm3 @@ -1169,10 +1176,10 @@ LABEL(nibble_ashr_8): test $0xff00, %edx jnz LABEL(ashr_8_exittail) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP cmp $7, %r11 jbe LABEL(ashr_8_exittail) -#endif +# endif pxor %xmm0, %xmm0 sub $0x1000, %r10 @@ -1188,7 +1195,7 @@ LABEL(ashr_8_exittail): /* * The following cases will be handled by ashr_9 * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case - * n(7~15) n - 7 8(15 +(n - 7) - n) ashr_9 + * n(7~15) n - 7 8(15 +(n - 7) - n) ashr_9 */ .p2align 4 LABEL(ashr_9): @@ -1230,13 +1237,13 @@ LABEL(gobble_ashr_9): movdqa (%rdi, %rcx), %xmm2 movdqa %xmm2, %xmm4 -#ifndef USE_SSSE3 +# ifndef USE_SSSE3 psrldq $9, %xmm3 pslldq $7, %xmm2 por %xmm3, %xmm2 /* merge into one 16byte value */ -#else +# else palignr $9, %xmm3, %xmm2 /* merge into one 16byte value */ -#endif +# endif pcmpeqb %xmm1, %xmm0 pcmpeqb %xmm2, %xmm1 @@ -1245,10 +1252,10 @@ LABEL(gobble_ashr_9): sub $0xffff, %edx jnz LABEL(exit) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub $16, %r11 jbe LABEL(strcmp_exitz) -#endif +# endif add $16, %rcx movdqa %xmm4, %xmm3 @@ -1260,13 +1267,13 @@ LABEL(gobble_ashr_9): movdqa (%rdi, %rcx), %xmm2 movdqa %xmm2, %xmm4 -#ifndef USE_SSSE3 +# ifndef USE_SSSE3 psrldq $9, %xmm3 pslldq $7, %xmm2 por %xmm3, %xmm2 /* merge into one 16byte value */ -#else +# else palignr $9, %xmm3, %xmm2 /* merge into one 16byte value */ -#endif +# endif pcmpeqb %xmm1, %xmm0 pcmpeqb %xmm2, %xmm1 @@ -1275,10 +1282,10 @@ LABEL(gobble_ashr_9): sub $0xffff, %edx jnz LABEL(exit) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub $16, %r11 jbe LABEL(strcmp_exitz) -#endif +# endif add $16, %rcx movdqa %xmm4, %xmm3 /* store for next cycle */ @@ -1291,10 +1298,10 @@ LABEL(nibble_ashr_9): test $0xfe00, %edx jnz LABEL(ashr_9_exittail) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP cmp $6, %r11 jbe LABEL(ashr_9_exittail) -#endif +# endif pxor %xmm0, %xmm0 sub $0x1000, %r10 @@ -1310,7 +1317,7 @@ LABEL(ashr_9_exittail): /* * The following cases will be handled by ashr_10 * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case - * n(6~15) n - 6 9(15 +(n - 6) - n) ashr_10 + * n(6~15) n - 6 9(15 +(n - 6) - n) ashr_10 */ .p2align 4 LABEL(ashr_10): @@ -1352,13 +1359,13 @@ LABEL(gobble_ashr_10): movdqa (%rdi, %rcx), %xmm2 movdqa %xmm2, %xmm4 -#ifndef USE_SSSE3 +# ifndef USE_SSSE3 psrldq $10, %xmm3 pslldq $6, %xmm2 por %xmm3, %xmm2 /* merge into one 16byte value */ -#else +# else palignr $10, %xmm3, %xmm2 /* merge into one 16byte value */ -#endif +# endif pcmpeqb %xmm1, %xmm0 pcmpeqb %xmm2, %xmm1 @@ -1367,10 +1374,10 @@ LABEL(gobble_ashr_10): sub $0xffff, %edx jnz LABEL(exit) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub $16, %r11 jbe LABEL(strcmp_exitz) -#endif +# endif add $16, %rcx movdqa %xmm4, %xmm3 @@ -1382,13 +1389,13 @@ LABEL(gobble_ashr_10): movdqa (%rdi, %rcx), %xmm2 movdqa %xmm2, %xmm4 -#ifndef USE_SSSE3 +# ifndef USE_SSSE3 psrldq $10, %xmm3 pslldq $6, %xmm2 por %xmm3, %xmm2 /* merge into one 16byte value */ -#else +# else palignr $10, %xmm3, %xmm2 /* merge into one 16byte value */ -#endif +# endif pcmpeqb %xmm1, %xmm0 pcmpeqb %xmm2, %xmm1 @@ -1397,10 +1404,10 @@ LABEL(gobble_ashr_10): sub $0xffff, %edx jnz LABEL(exit) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub $16, %r11 jbe LABEL(strcmp_exitz) -#endif +# endif add $16, %rcx movdqa %xmm4, %xmm3 @@ -1413,10 +1420,10 @@ LABEL(nibble_ashr_10): test $0xfc00, %edx jnz LABEL(ashr_10_exittail) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP cmp $5, %r11 jbe LABEL(ashr_10_exittail) -#endif +# endif pxor %xmm0, %xmm0 sub $0x1000, %r10 @@ -1432,7 +1439,7 @@ LABEL(ashr_10_exittail): /* * The following cases will be handled by ashr_11 * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case - * n(5~15) n - 5 10(15 +(n - 5) - n) ashr_11 + * n(5~15) n - 5 10(15 +(n - 5) - n) ashr_11 */ .p2align 4 LABEL(ashr_11): @@ -1474,13 +1481,13 @@ LABEL(gobble_ashr_11): movdqa (%rdi, %rcx), %xmm2 movdqa %xmm2, %xmm4 -#ifndef USE_SSSE3 +# ifndef USE_SSSE3 psrldq $11, %xmm3 pslldq $5, %xmm2 por %xmm3, %xmm2 /* merge into one 16byte value */ -#else +# else palignr $11, %xmm3, %xmm2 /* merge into one 16byte value */ -#endif +# endif pcmpeqb %xmm1, %xmm0 pcmpeqb %xmm2, %xmm1 @@ -1489,10 +1496,10 @@ LABEL(gobble_ashr_11): sub $0xffff, %edx jnz LABEL(exit) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub $16, %r11 jbe LABEL(strcmp_exitz) -#endif +# endif add $16, %rcx movdqa %xmm4, %xmm3 @@ -1504,13 +1511,13 @@ LABEL(gobble_ashr_11): movdqa (%rdi, %rcx), %xmm2 movdqa %xmm2, %xmm4 -#ifndef USE_SSSE3 +# ifndef USE_SSSE3 psrldq $11, %xmm3 pslldq $5, %xmm2 por %xmm3, %xmm2 /* merge into one 16byte value */ -#else +# else palignr $11, %xmm3, %xmm2 /* merge into one 16byte value */ -#endif +# endif pcmpeqb %xmm1, %xmm0 pcmpeqb %xmm2, %xmm1 @@ -1519,10 +1526,10 @@ LABEL(gobble_ashr_11): sub $0xffff, %edx jnz LABEL(exit) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub $16, %r11 jbe LABEL(strcmp_exitz) -#endif +# endif add $16, %rcx movdqa %xmm4, %xmm3 @@ -1535,10 +1542,10 @@ LABEL(nibble_ashr_11): test $0xf800, %edx jnz LABEL(ashr_11_exittail) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP cmp $4, %r11 jbe LABEL(ashr_11_exittail) -#endif +# endif pxor %xmm0, %xmm0 sub $0x1000, %r10 @@ -1554,7 +1561,7 @@ LABEL(ashr_11_exittail): /* * The following cases will be handled by ashr_12 * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case - * n(4~15) n - 4 11(15 +(n - 4) - n) ashr_12 + * n(4~15) n - 4 11(15 +(n - 4) - n) ashr_12 */ .p2align 4 LABEL(ashr_12): @@ -1596,13 +1603,13 @@ LABEL(gobble_ashr_12): movdqa (%rdi, %rcx), %xmm2 movdqa %xmm2, %xmm4 -#ifndef USE_SSSE3 +# ifndef USE_SSSE3 psrldq $12, %xmm3 pslldq $4, %xmm2 por %xmm3, %xmm2 /* merge into one 16byte value */ -#else +# else palignr $12, %xmm3, %xmm2 /* merge into one 16byte value */ -#endif +# endif pcmpeqb %xmm1, %xmm0 pcmpeqb %xmm2, %xmm1 @@ -1611,10 +1618,10 @@ LABEL(gobble_ashr_12): sub $0xffff, %edx jnz LABEL(exit) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub $16, %r11 jbe LABEL(strcmp_exitz) -#endif +# endif add $16, %rcx movdqa %xmm4, %xmm3 @@ -1626,13 +1633,13 @@ LABEL(gobble_ashr_12): movdqa (%rdi, %rcx), %xmm2 movdqa %xmm2, %xmm4 -#ifndef USE_SSSE3 +# ifndef USE_SSSE3 psrldq $12, %xmm3 pslldq $4, %xmm2 por %xmm3, %xmm2 /* merge into one 16byte value */ -#else +# else palignr $12, %xmm3, %xmm2 /* merge into one 16byte value */ -#endif +# endif pcmpeqb %xmm1, %xmm0 pcmpeqb %xmm2, %xmm1 @@ -1641,10 +1648,10 @@ LABEL(gobble_ashr_12): sub $0xffff, %edx jnz LABEL(exit) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub $16, %r11 jbe LABEL(strcmp_exitz) -#endif +# endif add $16, %rcx movdqa %xmm4, %xmm3 @@ -1657,10 +1664,10 @@ LABEL(nibble_ashr_12): test $0xf000, %edx jnz LABEL(ashr_12_exittail) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP cmp $3, %r11 jbe LABEL(ashr_12_exittail) -#endif +# endif pxor %xmm0, %xmm0 sub $0x1000, %r10 @@ -1676,7 +1683,7 @@ LABEL(ashr_12_exittail): /* * The following cases will be handled by ashr_13 * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case - * n(3~15) n - 3 12(15 +(n - 3) - n) ashr_13 + * n(3~15) n - 3 12(15 +(n - 3) - n) ashr_13 */ .p2align 4 LABEL(ashr_13): @@ -1718,13 +1725,13 @@ LABEL(gobble_ashr_13): movdqa (%rdi, %rcx), %xmm2 movdqa %xmm2, %xmm4 -#ifndef USE_SSSE3 +# ifndef USE_SSSE3 psrldq $13, %xmm3 pslldq $3, %xmm2 por %xmm3, %xmm2 /* merge into one 16byte value */ -#else +# else palignr $13, %xmm3, %xmm2 /* merge into one 16byte value */ -#endif +# endif pcmpeqb %xmm1, %xmm0 pcmpeqb %xmm2, %xmm1 @@ -1733,10 +1740,10 @@ LABEL(gobble_ashr_13): sub $0xffff, %edx jnz LABEL(exit) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub $16, %r11 jbe LABEL(strcmp_exitz) -#endif +# endif add $16, %rcx movdqa %xmm4, %xmm3 @@ -1748,13 +1755,13 @@ LABEL(gobble_ashr_13): movdqa (%rdi, %rcx), %xmm2 movdqa %xmm2, %xmm4 -#ifndef USE_SSSE3 +# ifndef USE_SSSE3 psrldq $13, %xmm3 pslldq $3, %xmm2 por %xmm3, %xmm2 /* merge into one 16byte value */ -#else +# else palignr $13, %xmm3, %xmm2 /* merge into one 16byte value */ -#endif +# endif pcmpeqb %xmm1, %xmm0 pcmpeqb %xmm2, %xmm1 @@ -1763,10 +1770,10 @@ LABEL(gobble_ashr_13): sub $0xffff, %edx jnz LABEL(exit) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub $16, %r11 jbe LABEL(strcmp_exitz) -#endif +# endif add $16, %rcx movdqa %xmm4, %xmm3 @@ -1779,10 +1786,10 @@ LABEL(nibble_ashr_13): test $0xe000, %edx jnz LABEL(ashr_13_exittail) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP cmp $2, %r11 jbe LABEL(ashr_13_exittail) -#endif +# endif pxor %xmm0, %xmm0 sub $0x1000, %r10 @@ -1798,7 +1805,7 @@ LABEL(ashr_13_exittail): /* * The following cases will be handled by ashr_14 * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case - * n(2~15) n - 2 13(15 +(n - 2) - n) ashr_14 + * n(2~15) n - 2 13(15 +(n - 2) - n) ashr_14 */ .p2align 4 LABEL(ashr_14): @@ -1840,13 +1847,13 @@ LABEL(gobble_ashr_14): movdqa (%rdi, %rcx), %xmm2 movdqa %xmm2, %xmm4 -#ifndef USE_SSSE3 +# ifndef USE_SSSE3 psrldq $14, %xmm3 pslldq $2, %xmm2 por %xmm3, %xmm2 /* merge into one 16byte value */ -#else +# else palignr $14, %xmm3, %xmm2 /* merge into one 16byte value */ -#endif +# endif pcmpeqb %xmm1, %xmm0 pcmpeqb %xmm2, %xmm1 @@ -1855,10 +1862,10 @@ LABEL(gobble_ashr_14): sub $0xffff, %edx jnz LABEL(exit) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub $16, %r11 jbe LABEL(strcmp_exitz) -#endif +# endif add $16, %rcx movdqa %xmm4, %xmm3 @@ -1870,13 +1877,13 @@ LABEL(gobble_ashr_14): movdqa (%rdi, %rcx), %xmm2 movdqa %xmm2, %xmm4 -#ifndef USE_SSSE3 +# ifndef USE_SSSE3 psrldq $14, %xmm3 pslldq $2, %xmm2 por %xmm3, %xmm2 /* merge into one 16byte value */ -#else +# else palignr $14, %xmm3, %xmm2 /* merge into one 16byte value */ -#endif +# endif pcmpeqb %xmm1, %xmm0 pcmpeqb %xmm2, %xmm1 @@ -1885,10 +1892,10 @@ LABEL(gobble_ashr_14): sub $0xffff, %edx jnz LABEL(exit) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub $16, %r11 jbe LABEL(strcmp_exitz) -#endif +# endif add $16, %rcx movdqa %xmm4, %xmm3 @@ -1901,10 +1908,10 @@ LABEL(nibble_ashr_14): test $0xc000, %edx jnz LABEL(ashr_14_exittail) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP cmp $1, %r11 jbe LABEL(ashr_14_exittail) -#endif +# endif pxor %xmm0, %xmm0 sub $0x1000, %r10 @@ -1920,7 +1927,7 @@ LABEL(ashr_14_exittail): /* * The following cases will be handled by ashr_15 * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case - * n(1~15) n - 1 14(15 +(n - 1) - n) ashr_15 + * n(1~15) n - 1 14(15 +(n - 1) - n) ashr_15 */ .p2align 4 LABEL(ashr_15): @@ -1964,13 +1971,13 @@ LABEL(gobble_ashr_15): movdqa (%rdi, %rcx), %xmm2 movdqa %xmm2, %xmm4 -#ifndef USE_SSSE3 +# ifndef USE_SSSE3 psrldq $15, %xmm3 pslldq $1, %xmm2 por %xmm3, %xmm2 /* merge into one 16byte value */ -#else +# else palignr $15, %xmm3, %xmm2 /* merge into one 16byte value */ -#endif +# endif pcmpeqb %xmm1, %xmm0 pcmpeqb %xmm2, %xmm1 @@ -1979,10 +1986,10 @@ LABEL(gobble_ashr_15): sub $0xffff, %edx jnz LABEL(exit) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub $16, %r11 jbe LABEL(strcmp_exitz) -#endif +# endif add $16, %rcx movdqa %xmm4, %xmm3 @@ -1994,13 +2001,13 @@ LABEL(gobble_ashr_15): movdqa (%rdi, %rcx), %xmm2 movdqa %xmm2, %xmm4 -#ifndef USE_SSSE3 +# ifndef USE_SSSE3 psrldq $15, %xmm3 pslldq $1, %xmm2 por %xmm3, %xmm2 /* merge into one 16byte value */ -#else +# else palignr $15, %xmm3, %xmm2 /* merge into one 16byte value */ -#endif +# endif pcmpeqb %xmm1, %xmm0 pcmpeqb %xmm2, %xmm1 @@ -2009,10 +2016,10 @@ LABEL(gobble_ashr_15): sub $0xffff, %edx jnz LABEL(exit) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub $16, %r11 jbe LABEL(strcmp_exitz) -#endif +# endif add $16, %rcx movdqa %xmm4, %xmm3 @@ -2025,10 +2032,10 @@ LABEL(nibble_ashr_15): test $0x8000, %edx jnz LABEL(ashr_15_exittail) -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP test %r11, %r11 je LABEL(ashr_15_exittail) -#endif +# endif pxor %xmm0, %xmm0 sub $0x1000, %r10 @@ -2062,10 +2069,10 @@ LABEL(ret): LABEL(less16bytes): bsf %rdx, %rdx /* find and store bit index in %rdx */ -#ifdef USE_AS_STRNCMP +# ifdef USE_AS_STRNCMP sub %rdx, %r11 jbe LABEL(strcmp_exitz) -#endif +# endif movzbl (%rsi, %rdx), %ecx movzbl (%rdi, %rdx), %eax |