aboutsummaryrefslogtreecommitdiff
path: root/sysdeps/x86_64/multiarch/strcmp.S
diff options
context:
space:
mode:
authorUlrich Drepper <drepper@gmail.com>2011-10-19 12:42:38 -0400
committerUlrich Drepper <drepper@gmail.com>2011-10-19 12:42:38 -0400
commitd9a4d2ab278ab50637e383b6174e9ec42db84327 (patch)
treec5107d9a43e560658bf3b65fcde55659d3292200 /sysdeps/x86_64/multiarch/strcmp.S
parent8f3b1ffefa61275315891770944bdbf5079e04ed (diff)
downloadglibc-d9a4d2ab278ab50637e383b6174e9ec42db84327.tar
glibc-d9a4d2ab278ab50637e383b6174e9ec42db84327.tar.gz
glibc-d9a4d2ab278ab50637e383b6174e9ec42db84327.tar.bz2
glibc-d9a4d2ab278ab50637e383b6174e9ec42db84327.zip
Add optimized str{,n}casecmp for AVX on x86-64
Diffstat (limited to 'sysdeps/x86_64/multiarch/strcmp.S')
-rw-r--r--sysdeps/x86_64/multiarch/strcmp.S1828
1 files changed, 31 insertions, 1797 deletions
diff --git a/sysdeps/x86_64/multiarch/strcmp.S b/sysdeps/x86_64/multiarch/strcmp.S
index 8879855d92..f93c83d7d4 100644
--- a/sysdeps/x86_64/multiarch/strcmp.S
+++ b/sysdeps/x86_64/multiarch/strcmp.S
@@ -1,5 +1,5 @@
/* strcmp with SSE4.2
- Copyright (C) 2009, 2010 Free Software Foundation, Inc.
+ Copyright (C) 2009, 2010, 2011 Free Software Foundation, Inc.
Contributed by Intel Corporation.
This file is part of the GNU C Library.
@@ -28,9 +28,9 @@
/* calculate left number to compare */ \
lea -16(%rcx, %r11), %r9; \
cmp %r9, %r11; \
- jb LABEL(strcmp_exitz_sse4_2); \
+ jb LABEL(strcmp_exitz); \
test %r9, %r9; \
- je LABEL(strcmp_exitz_sse4_2); \
+ je LABEL(strcmp_exitz); \
mov %r9, %r11
# define STRCMP_SSE42 __strncmp_sse42
@@ -42,6 +42,7 @@
# define UPDATE_STRNCMP_COUNTER
+# define STRCMP_AVX __strcasecmp_l_avx
# define STRCMP_SSE42 __strcasecmp_l_sse42
# define STRCMP_SSSE3 __strcasecmp_l_ssse3
# define STRCMP_SSE2 __strcasecmp_l_sse2
@@ -55,11 +56,12 @@
/* calculate left number to compare */ \
lea -16(%rcx, %r11), %r9; \
cmp %r9, %r11; \
- jb LABEL(strcmp_exitz_sse4_2); \
+ jb LABEL(strcmp_exitz); \
test %r9, %r9; \
- je LABEL(strcmp_exitz_sse4_2); \
+ je LABEL(strcmp_exitz); \
mov %r9, %r11
+# define STRCMP_AVX __strncasecmp_l_avx
# define STRCMP_SSE42 __strncasecmp_l_sse42
# define STRCMP_SSSE3 __strncasecmp_l_ssse3
# define STRCMP_SSE2 __strncasecmp_l_sse2
@@ -75,10 +77,6 @@
# endif
#endif
-#ifndef LABEL
-# define LABEL(l) L(l)
-#endif
-
/* Define multiple versions only for the definition in libc. Don't
define multiple versions for strncmp in static library since we
need strncmp before the initialization happened. */
@@ -107,6 +105,11 @@ ENTRY(__strcasecmp)
jne 1f
call __init_cpu_features
1:
+# ifdef HAVE_AVX_SUPPORT
+ leaq __strcasecmp_avx(%rip), %rax
+ testl $bit_AVX, __cpu_features+CPUID_OFFSET+index_AVX(%rip)
+ jnz 2f
+# endif
leaq __strcasecmp_sse42(%rip), %rax
testl $bit_SSE4_2, __cpu_features+CPUID_OFFSET+index_SSE4_2(%rip)
jnz 2f
@@ -125,6 +128,11 @@ ENTRY(__strncasecmp)
jne 1f
call __init_cpu_features
1:
+# ifdef HAVE_AVX_SUPPORT
+ leaq __strncasecmp_avx(%rip), %rax
+ testl $bit_AVX, __cpu_features+CPUID_OFFSET+index_AVX(%rip)
+ jnz 2f
+# endif
leaq __strncasecmp_sse42(%rip), %rax
testl $bit_SSE4_2, __cpu_features+CPUID_OFFSET+index_SSE4_2(%rip)
jnz 2f
@@ -137,1798 +145,24 @@ END(__strncasecmp)
weak_alias (__strncasecmp, strncasecmp)
# endif
-/* We use 0x1a:
- _SIDD_SBYTE_OPS
- | _SIDD_CMP_EQUAL_EACH
- | _SIDD_NEGATIVE_POLARITY
- | _SIDD_LEAST_SIGNIFICANT
- on pcmpistri to find out if two 16byte data elements are the same
- and the offset of the first different byte. There are 4 cases:
-
- 1. Both 16byte data elements are valid and identical.
- 2. Both 16byte data elements have EOS and identical.
- 3. Both 16byte data elements are valid and they differ at offset X.
- 4. At least one 16byte data element has EOS at offset X. Two 16byte
- data elements must differ at or before offset X.
-
- Here is the table of ECX, CFlag, ZFlag and SFlag for 4 cases:
-
- case ECX CFlag ZFlag SFlag
- 1 16 0 0 0
- 2 16 0 1 1
- 3 X 1 0 0
- 4 0 <= X 1 0/1 0/1
-
- We exit from the loop for cases 2, 3 and 4 with jbe which branches
- when either CFlag or ZFlag is 1. If CFlag == 0, we return 0 for
- case 2. */
-
- /* Put all SSE 4.2 functions together. */
- .section .text.sse4.2,"ax",@progbits
- .align 16
- .type STRCMP_SSE42, @function
-# ifdef USE_AS_STRCASECMP_L
-ENTRY (__strcasecmp_sse42)
- movq __libc_tsd_LOCALE@gottpoff(%rip),%rax
- movq %fs:(%rax),%rdx
-
- // XXX 5 byte should be before the function
- /* 5-byte NOP. */
- .byte 0x0f,0x1f,0x44,0x00,0x00
-END (__strcasecmp_sse42)
- /* FALLTHROUGH to strcasecmp_l. */
-# endif
-# ifdef USE_AS_STRNCASECMP_L
-ENTRY (__strncasecmp_sse42)
- movq __libc_tsd_LOCALE@gottpoff(%rip),%rax
- movq %fs:(%rax),%rcx
-
- // XXX 5 byte should be before the function
- /* 5-byte NOP. */
- .byte 0x0f,0x1f,0x44,0x00,0x00
-END (__strncasecmp_sse42)
- /* FALLTHROUGH to strncasecmp_l. */
-# endif
+# undef LABEL
+# define LABEL(l) .L##l##_sse42
+# define GLABEL(l) l##_sse42
+# define SECTION sse4.2
+# include "strcmp-sse42.S"
-STRCMP_SSE42:
- cfi_startproc
- CALL_MCOUNT
-/*
- * This implementation uses SSE to compare up to 16 bytes at a time.
- */
-# ifdef USE_AS_STRCASECMP_L
- /* We have to fall back on the C implementation for locales
- with encodings not matching ASCII for single bytes. */
-# if LOCALE_T___LOCALES != 0 || LC_CTYPE != 0
- movq LOCALE_T___LOCALES+LC_CTYPE*8(%rdx), %rax
-# else
- movq (%rdx), %rax
+# ifdef HAVE_AVX_SUPPORT
+# if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L
+# define LABEL(l) .L##l##_avx
+# define GLABEL(l) l##_avx
+# define USE_AVX 1
+# undef STRCMP_SSE42
+# define STRCMP_SSE42 STRCMP_AVX
+# define SECTION avx
+# include "strcmp-sse42.S"
# endif
- testl $0, LOCALE_DATA_VALUES+_NL_CTYPE_NONASCII_CASE*SIZEOF_VALUES(%rax)
- jne __strcasecmp_l_nonascii
-# endif
-# ifdef USE_AS_STRNCASECMP_L
- /* We have to fall back on the C implementation for locales
- with encodings not matching ASCII for single bytes. */
-# if LOCALE_T___LOCALES != 0 || LC_CTYPE != 0
- movq LOCALE_T___LOCALES+LC_CTYPE*8(%rcx), %rax
-# else
- movq (%rcx), %rax
-# endif
- testl $0, LOCALE_DATA_VALUES+_NL_CTYPE_NONASCII_CASE*SIZEOF_VALUES(%rax)
- jne __strncasecmp_l_nonascii
-# endif
-
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- test %rdx, %rdx
- je LABEL(strcmp_exitz_sse4_2)
- cmp $1, %rdx
- je LABEL(Byte0_sse4_2)
- mov %rdx, %r11
-# endif
- mov %esi, %ecx
- mov %edi, %eax
-/* Use 64bit AND here to avoid long NOP padding. */
- and $0x3f, %rcx /* rsi alignment in cache line */
- and $0x3f, %rax /* rdi alignment in cache line */
-# if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L
- .section .rodata.cst16,"aM",@progbits,16
- .align 16
-.Lbelowupper_sse4:
- .quad 0x4040404040404040
- .quad 0x4040404040404040
-.Ltopupper_sse4:
- .quad 0x5b5b5b5b5b5b5b5b
- .quad 0x5b5b5b5b5b5b5b5b
-.Ltouppermask_sse4:
- .quad 0x2020202020202020
- .quad 0x2020202020202020
- .previous
- movdqa .Lbelowupper_sse4(%rip), %xmm4
-# define UCLOW_reg %xmm4
- movdqa .Ltopupper_sse4(%rip), %xmm5
-# define UCHIGH_reg %xmm5
- movdqa .Ltouppermask_sse4(%rip), %xmm6
-# define LCQWORD_reg %xmm6
-# endif
- cmp $0x30, %ecx
- ja LABEL(crosscache_sse4_2)/* rsi: 16-byte load will cross cache line */
- cmp $0x30, %eax
- ja LABEL(crosscache_sse4_2)/* rdi: 16-byte load will cross cache line */
- movdqu (%rdi), %xmm1
- movdqu (%rsi), %xmm2
-# if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L
-# define TOLOWER(reg1, reg2) \
- movdqa reg1, %xmm7; \
- movdqa UCHIGH_reg, %xmm8; \
- movdqa reg2, %xmm9; \
- movdqa UCHIGH_reg, %xmm10; \
- pcmpgtb UCLOW_reg, %xmm7; \
- pcmpgtb reg1, %xmm8; \
- pcmpgtb UCLOW_reg, %xmm9; \
- pcmpgtb reg2, %xmm10; \
- pand %xmm8, %xmm7; \
- pand %xmm10, %xmm9; \
- pand LCQWORD_reg, %xmm7; \
- pand LCQWORD_reg, %xmm9; \
- por %xmm7, reg1; \
- por %xmm9, reg2
- TOLOWER (%xmm1, %xmm2)
-# else
-# define TOLOWER(reg1, reg2)
-# endif
- pxor %xmm0, %xmm0 /* clear %xmm0 for null char checks */
- pcmpeqb %xmm1, %xmm0 /* Any null chars? */
- pcmpeqb %xmm2, %xmm1 /* compare first 16 bytes for equality */
- psubb %xmm0, %xmm1 /* packed sub of comparison results*/
- pmovmskb %xmm1, %edx
- sub $0xffff, %edx /* if first 16 bytes are same, edx == 0xffff */
- jnz LABEL(less16bytes_sse4_2)/* If not, find different value or null char */
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub $16, %r11
- jbe LABEL(strcmp_exitz_sse4_2)/* finish comparision */
-# endif
- add $16, %rsi /* prepare to search next 16 bytes */
- add $16, %rdi /* prepare to search next 16 bytes */
-
- /*
- * Determine source and destination string offsets from 16-byte alignment.
- * Use relative offset difference between the two to determine which case
- * below to use.
- */
- .p2align 4
-LABEL(crosscache_sse4_2):
- and $0xfffffffffffffff0, %rsi /* force %rsi is 16 byte aligned */
- and $0xfffffffffffffff0, %rdi /* force %rdi is 16 byte aligned */
- mov $0xffff, %edx /* for equivalent offset */
- xor %r8d, %r8d
- and $0xf, %ecx /* offset of rsi */
- and $0xf, %eax /* offset of rdi */
- cmp %eax, %ecx
- je LABEL(ashr_0_sse4_2) /* rsi and rdi relative offset same */
- ja LABEL(bigger_sse4_2)
- mov %edx, %r8d /* r8d is offset flag for exit tail */
- xchg %ecx, %eax
- xchg %rsi, %rdi
-LABEL(bigger_sse4_2):
- lea 15(%rax), %r9
- sub %rcx, %r9
- lea LABEL(unaligned_table_sse4_2)(%rip), %r10
- movslq (%r10, %r9,4), %r9
- lea (%r10, %r9), %r10
- jmp *%r10 /* jump to corresponding case */
-
-/*
- * The following cases will be handled by ashr_0
- * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case
- * n(0~15) n(0~15) 15(15+ n-n) ashr_0
- */
- .p2align 4
-LABEL(ashr_0_sse4_2):
-
- movdqa (%rsi), %xmm1
- pxor %xmm0, %xmm0 /* clear %xmm0 for null char check */
- pcmpeqb %xmm1, %xmm0 /* Any null chars? */
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpeqb (%rdi), %xmm1 /* compare 16 bytes for equality */
-# else
- movdqa (%rdi), %xmm2
- TOLOWER (%xmm1, %xmm2)
- pcmpeqb %xmm2, %xmm1 /* compare 16 bytes for equality */
-# endif
- psubb %xmm0, %xmm1 /* packed sub of comparison results*/
- pmovmskb %xmm1, %r9d
- shr %cl, %edx /* adjust 0xffff for offset */
- shr %cl, %r9d /* adjust for 16-byte offset */
- sub %r9d, %edx
- /*
- * edx must be the same with r9d if in left byte (16-rcx) is equal to
- * the start from (16-rax) and no null char was seen.
- */
- jne LABEL(less32bytes_sse4_2) /* mismatch or null char */
- UPDATE_STRNCMP_COUNTER
- mov $16, %rcx
- mov $16, %r9
- pxor %xmm0, %xmm0 /* clear xmm0, may have changed above */
-
- /*
- * Now both strings are aligned at 16-byte boundary. Loop over strings
- * checking 32-bytes per iteration.
- */
- mov %rcx, %rdx /* only for offset of sse4 instruction loop*/
- .p2align 4
-LABEL(ashr_0_use_sse4_2):
- movdqa (%rdi,%rdx), %xmm0
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpistri $0x1a,(%rsi,%rdx), %xmm0
-# else
- movdqa (%rsi,%rdx), %xmm1
- TOLOWER (%xmm0, %xmm1)
- pcmpistri $0x1a, %xmm1, %xmm0
-# endif
- lea 16(%rdx), %rdx
- jbe LABEL(ashr_0_use_sse4_2_exit)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub $16, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
-
- movdqa (%rdi,%rdx), %xmm0
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpistri $0x1a,(%rsi,%rdx), %xmm0
-# else
- movdqa (%rsi,%rdx), %xmm1
- TOLOWER (%xmm0, %xmm1)
- pcmpistri $0x1a, %xmm1, %xmm0
-# endif
- lea 16(%rdx), %rdx
- jbe LABEL(ashr_0_use_sse4_2_exit)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub $16, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
- jmp LABEL(ashr_0_use_sse4_2)
-
-
- .p2align 4
-LABEL(ashr_0_use_sse4_2_exit):
- jnc LABEL(strcmp_exitz_sse4_2)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub %rcx, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
- lea -16(%rdx, %rcx), %rcx
- movzbl (%rdi, %rcx), %eax
- movzbl (%rsi, %rcx), %edx
-# if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L
- leaq _nl_C_LC_CTYPE_tolower+128*4(%rip), %rcx
- movl (%rcx,%rax,4), %eax
- movl (%rcx,%rdx,4), %edx
-# endif
- sub %edx, %eax
- ret
-
-
-
-/*
- * The following cases will be handled by ashr_1
- * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case
- * n(15) n -15 0(15 +(n-15) - n) ashr_1
- */
- .p2align 4
-LABEL(ashr_1_sse4_2):
- pxor %xmm0, %xmm0
- movdqa (%rdi), %xmm2
- movdqa (%rsi), %xmm1
- pcmpeqb %xmm1, %xmm0 /* Any null chars? */
- pslldq $15, %xmm2 /* shift first string to align with second */
- TOLOWER (%xmm1, %xmm2)
- pcmpeqb %xmm1, %xmm2 /* compare 16 bytes for equality */
- psubb %xmm0, %xmm2 /* packed sub of comparison results*/
- pmovmskb %xmm2, %r9d
- shr %cl, %edx /* adjust 0xffff for offset */
- shr %cl, %r9d /* adjust for 16-byte offset */
- sub %r9d, %edx
- jnz LABEL(less32bytes_sse4_2)/* mismatch or null char seen */
- movdqa (%rdi), %xmm3
- UPDATE_STRNCMP_COUNTER
-
- pxor %xmm0, %xmm0
- mov $16, %rcx /* index for loads*/
- mov $1, %r9d /* byte position left over from less32bytes case */
- /*
- * Setup %r10 value allows us to detect crossing a page boundary.
- * When %r10 goes positive we have crossed a page boundary and
- * need to do a nibble.
- */
- lea 1(%rdi), %r10
- and $0xfff, %r10 /* offset into 4K page */
- sub $0x1000, %r10 /* subtract 4K pagesize */
- mov %rcx, %rdx /* only for offset of sse4 instruction loop*/
-
- .p2align 4
-LABEL(loop_ashr_1_use_sse4_2):
- add $16, %r10
- jg LABEL(nibble_ashr_1_use_sse4_2)
-
-LABEL(nibble_ashr_1_use_sse4_2_restart):
- movdqa (%rdi, %rdx), %xmm0
- palignr $1, -16(%rdi, %rdx), %xmm0
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpistri $0x1a,(%rsi,%rdx), %xmm0
-# else
- movdqa (%rsi,%rdx), %xmm1
- TOLOWER (%xmm0, %xmm1)
- pcmpistri $0x1a, %xmm1, %xmm0
-# endif
- jbe LABEL(use_sse4_2_exit)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub $16, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
-
- add $16, %rdx
- add $16, %r10
- jg LABEL(nibble_ashr_1_use_sse4_2)
-
- movdqa (%rdi, %rdx), %xmm0
- palignr $1, -16(%rdi, %rdx), %xmm0
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpistri $0x1a,(%rsi,%rdx), %xmm0
-# else
- movdqa (%rsi,%rdx), %xmm1
- TOLOWER (%xmm0, %xmm1)
- pcmpistri $0x1a, %xmm1, %xmm0
-# endif
- jbe LABEL(use_sse4_2_exit)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub $16, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
- add $16, %rdx
- jmp LABEL(loop_ashr_1_use_sse4_2)
-
- .p2align 4
-LABEL(nibble_ashr_1_use_sse4_2):
- sub $0x1000, %r10
- movdqa -16(%rdi, %rdx), %xmm0
- psrldq $1, %xmm0
- pcmpistri $0x3a,%xmm0, %xmm0
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- cmp %r11, %rcx
- jae LABEL(nibble_ashr_use_sse4_2_exit)
-# endif
- cmp $14, %ecx
- ja LABEL(nibble_ashr_1_use_sse4_2_restart)
-
- jmp LABEL(nibble_ashr_use_sse4_2_exit)
-
-/*
- * The following cases will be handled by ashr_2
- * 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
-LABEL(ashr_2_sse4_2):
- pxor %xmm0, %xmm0
- movdqa (%rdi), %xmm2
- movdqa (%rsi), %xmm1
- pcmpeqb %xmm1, %xmm0
- pslldq $14, %xmm2
- TOLOWER (%xmm1, %xmm2)
- pcmpeqb %xmm1, %xmm2
- psubb %xmm0, %xmm2
- pmovmskb %xmm2, %r9d
- shr %cl, %edx
- shr %cl, %r9d
- sub %r9d, %edx
- jnz LABEL(less32bytes_sse4_2)
- movdqa (%rdi), %xmm3
- UPDATE_STRNCMP_COUNTER
-
- pxor %xmm0, %xmm0
- mov $16, %rcx /* index for loads */
- mov $2, %r9d /* byte position left over from less32bytes case */
- /*
- * Setup %r10 value allows us to detect crossing a page boundary.
- * When %r10 goes positive we have crossed a page boundary and
- * need to do a nibble.
- */
- lea 2(%rdi), %r10
- and $0xfff, %r10 /* offset into 4K page */
- sub $0x1000, %r10 /* subtract 4K pagesize */
- mov %rcx, %rdx /* only for offset of sse4 instruction loop*/
-
- .p2align 4
-LABEL(loop_ashr_2_use_sse4_2):
- add $16, %r10
- jg LABEL(nibble_ashr_2_use_sse4_2)
-
-LABEL(nibble_ashr_2_use_sse4_2_restart):
- movdqa (%rdi, %rdx), %xmm0
- palignr $2, -16(%rdi, %rdx), %xmm0
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpistri $0x1a,(%rsi,%rdx), %xmm0
-# else
- movdqa (%rsi,%rdx), %xmm1
- TOLOWER (%xmm0, %xmm1)
- pcmpistri $0x1a, %xmm1, %xmm0
-# endif
- jbe LABEL(use_sse4_2_exit)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub $16, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
-
- add $16, %rdx
- add $16, %r10
- jg LABEL(nibble_ashr_2_use_sse4_2)
-
- movdqa (%rdi, %rdx), %xmm0
- palignr $2, -16(%rdi, %rdx), %xmm0
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpistri $0x1a,(%rsi,%rdx), %xmm0
-# else
- movdqa (%rsi,%rdx), %xmm1
- TOLOWER (%xmm0, %xmm1)
- pcmpistri $0x1a, %xmm1, %xmm0
-# endif
- jbe LABEL(use_sse4_2_exit)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub $16, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
- add $16, %rdx
- jmp LABEL(loop_ashr_2_use_sse4_2)
-
- .p2align 4
-LABEL(nibble_ashr_2_use_sse4_2):
- sub $0x1000, %r10
- movdqa -16(%rdi, %rdx), %xmm0
- psrldq $2, %xmm0
- pcmpistri $0x3a,%xmm0, %xmm0
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- cmp %r11, %rcx
- jae LABEL(nibble_ashr_use_sse4_2_exit)
-# endif
- cmp $13, %ecx
- ja LABEL(nibble_ashr_2_use_sse4_2_restart)
-
- jmp LABEL(nibble_ashr_use_sse4_2_exit)
-
-/*
- * The following cases will be handled by ashr_3
- * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case
- * n(13~15) n -13 2(15 +(n-13) - n) ashr_3
- */
- .p2align 4
-LABEL(ashr_3_sse4_2):
- pxor %xmm0, %xmm0
- movdqa (%rdi), %xmm2
- movdqa (%rsi), %xmm1
- pcmpeqb %xmm1, %xmm0
- pslldq $13, %xmm2
- TOLOWER (%xmm1, %xmm2)
- pcmpeqb %xmm1, %xmm2
- psubb %xmm0, %xmm2
- pmovmskb %xmm2, %r9d
- shr %cl, %edx
- shr %cl, %r9d
- sub %r9d, %edx
- jnz LABEL(less32bytes_sse4_2)
- movdqa (%rdi), %xmm3
-
- UPDATE_STRNCMP_COUNTER
-
- pxor %xmm0, %xmm0
- mov $16, %rcx /* index for loads */
- mov $3, %r9d /* byte position left over from less32bytes case */
- /*
- * Setup %r10 value allows us to detect crossing a page boundary.
- * When %r10 goes positive we have crossed a page boundary and
- * need to do a nibble.
- */
- lea 3(%rdi), %r10
- and $0xfff, %r10 /* offset into 4K page */
- sub $0x1000, %r10 /* subtract 4K pagesize */
- mov %rcx, %rdx /* only for offset of sse4 instruction loop*/
-
-LABEL(loop_ashr_3_use_sse4_2):
- add $16, %r10
- jg LABEL(nibble_ashr_3_use_sse4_2)
-
-LABEL(nibble_ashr_3_use_sse4_2_restart):
- movdqa (%rdi, %rdx), %xmm0
- palignr $3, -16(%rdi, %rdx), %xmm0
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpistri $0x1a,(%rsi,%rdx), %xmm0
-# else
- movdqa (%rsi,%rdx), %xmm1
- TOLOWER (%xmm0, %xmm1)
- pcmpistri $0x1a, %xmm1, %xmm0
-# endif
- jbe LABEL(use_sse4_2_exit)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub $16, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
-
- add $16, %rdx
- add $16, %r10
- jg LABEL(nibble_ashr_3_use_sse4_2)
-
- movdqa (%rdi, %rdx), %xmm0
- palignr $3, -16(%rdi, %rdx), %xmm0
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpistri $0x1a,(%rsi,%rdx), %xmm0
-# else
- movdqa (%rsi,%rdx), %xmm1
- TOLOWER (%xmm0, %xmm1)
- pcmpistri $0x1a, %xmm1, %xmm0
-# endif
- jbe LABEL(use_sse4_2_exit)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub $16, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
- add $16, %rdx
- jmp LABEL(loop_ashr_3_use_sse4_2)
-
- .p2align 4
-LABEL(nibble_ashr_3_use_sse4_2):
- sub $0x1000, %r10
- movdqa -16(%rdi, %rdx), %xmm0
- psrldq $3, %xmm0
- pcmpistri $0x3a,%xmm0, %xmm0
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- cmp %r11, %rcx
- jae LABEL(nibble_ashr_use_sse4_2_exit)
-# endif
- cmp $12, %ecx
- ja LABEL(nibble_ashr_3_use_sse4_2_restart)
-
- jmp LABEL(nibble_ashr_use_sse4_2_exit)
-
-/*
- * The following cases will be handled by ashr_4
- * rcx(offset of rsi) rax(offset of rdi) relative offset corresponding case
- * n(12~15) n -12 3(15 +(n-12) - n) ashr_4
- */
- .p2align 4
-LABEL(ashr_4_sse4_2):
- pxor %xmm0, %xmm0
- movdqa (%rdi), %xmm2
- movdqa (%rsi), %xmm1
- pcmpeqb %xmm1, %xmm0
- pslldq $12, %xmm2
- TOLOWER (%xmm1, %xmm2)
- pcmpeqb %xmm1, %xmm2
- psubb %xmm0, %xmm2
- pmovmskb %xmm2, %r9d
- shr %cl, %edx
- shr %cl, %r9d
- sub %r9d, %edx
- jnz LABEL(less32bytes_sse4_2)
- movdqa (%rdi), %xmm3
-
- UPDATE_STRNCMP_COUNTER
-
- pxor %xmm0, %xmm0
- mov $16, %rcx /* index for loads */
- mov $4, %r9d /* byte position left over from less32bytes case */
- /*
- * Setup %r10 value allows us to detect crossing a page boundary.
- * When %r10 goes positive we have crossed a page boundary and
- * need to do a nibble.
- */
- lea 4(%rdi), %r10
- and $0xfff, %r10 /* offset into 4K page */
- sub $0x1000, %r10 /* subtract 4K pagesize */
- mov %rcx, %rdx /* only for offset of sse4 instruction loop*/
-
- .p2align 4
-LABEL(loop_ashr_4_use_sse4_2):
- add $16, %r10
- jg LABEL(nibble_ashr_4_use_sse4_2)
-
-LABEL(nibble_ashr_4_use_sse4_2_restart):
- movdqa (%rdi, %rdx), %xmm0
- palignr $4, -16(%rdi, %rdx), %xmm0
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpistri $0x1a,(%rsi,%rdx), %xmm0
-# else
- movdqa (%rsi,%rdx), %xmm1
- TOLOWER (%xmm0, %xmm1)
- pcmpistri $0x1a, %xmm1, %xmm0
-# endif
- jbe LABEL(use_sse4_2_exit)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub $16, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
-
- add $16, %rdx
- add $16, %r10
- jg LABEL(nibble_ashr_4_use_sse4_2)
-
- movdqa (%rdi, %rdx), %xmm0
- palignr $4, -16(%rdi, %rdx), %xmm0
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpistri $0x1a,(%rsi,%rdx), %xmm0
-# else
- movdqa (%rsi,%rdx), %xmm1
- TOLOWER (%xmm0, %xmm1)
- pcmpistri $0x1a, %xmm1, %xmm0
-# endif
- jbe LABEL(use_sse4_2_exit)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub $16, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
- add $16, %rdx
- jmp LABEL(loop_ashr_4_use_sse4_2)
-
- .p2align 4
-LABEL(nibble_ashr_4_use_sse4_2):
- sub $0x1000, %r10
- movdqa -16(%rdi, %rdx), %xmm0
- psrldq $4, %xmm0
- pcmpistri $0x3a,%xmm0, %xmm0
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- cmp %r11, %rcx
- jae LABEL(nibble_ashr_use_sse4_2_exit)
-# endif
- cmp $11, %ecx
- ja LABEL(nibble_ashr_4_use_sse4_2_restart)
-
- jmp LABEL(nibble_ashr_use_sse4_2_exit)
-
-/*
- * 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
- */
- .p2align 4
-LABEL(ashr_5_sse4_2):
- pxor %xmm0, %xmm0
- movdqa (%rdi), %xmm2
- movdqa (%rsi), %xmm1
- pcmpeqb %xmm1, %xmm0
- pslldq $11, %xmm2
- TOLOWER (%xmm1, %xmm2)
- pcmpeqb %xmm1, %xmm2
- psubb %xmm0, %xmm2
- pmovmskb %xmm2, %r9d
- shr %cl, %edx
- shr %cl, %r9d
- sub %r9d, %edx
- jnz LABEL(less32bytes_sse4_2)
- movdqa (%rdi), %xmm3
-
- UPDATE_STRNCMP_COUNTER
-
- pxor %xmm0, %xmm0
- mov $16, %rcx /* index for loads */
- mov $5, %r9d /* byte position left over from less32bytes case */
- /*
- * Setup %r10 value allows us to detect crossing a page boundary.
- * When %r10 goes positive we have crossed a page boundary and
- * need to do a nibble.
- */
- lea 5(%rdi), %r10
- and $0xfff, %r10 /* offset into 4K page */
- sub $0x1000, %r10 /* subtract 4K pagesize */
- mov %rcx, %rdx /* only for offset of sse4 instruction loop*/
-
- .p2align 4
-LABEL(loop_ashr_5_use_sse4_2):
- add $16, %r10
- jg LABEL(nibble_ashr_5_use_sse4_2)
-
-LABEL(nibble_ashr_5_use_sse4_2_restart):
- movdqa (%rdi, %rdx), %xmm0
- palignr $5, -16(%rdi, %rdx), %xmm0
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpistri $0x1a,(%rsi,%rdx), %xmm0
-# else
- movdqa (%rsi,%rdx), %xmm1
- TOLOWER (%xmm0, %xmm1)
- pcmpistri $0x1a, %xmm1, %xmm0
-# endif
- jbe LABEL(use_sse4_2_exit)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub $16, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
-
- add $16, %rdx
- add $16, %r10
- jg LABEL(nibble_ashr_5_use_sse4_2)
-
- movdqa (%rdi, %rdx), %xmm0
-
- palignr $5, -16(%rdi, %rdx), %xmm0
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpistri $0x1a,(%rsi,%rdx), %xmm0
-# else
- movdqa (%rsi,%rdx), %xmm1
- TOLOWER (%xmm0, %xmm1)
- pcmpistri $0x1a, %xmm1, %xmm0
-# endif
- jbe LABEL(use_sse4_2_exit)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub $16, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
- add $16, %rdx
- jmp LABEL(loop_ashr_5_use_sse4_2)
-
- .p2align 4
-LABEL(nibble_ashr_5_use_sse4_2):
- sub $0x1000, %r10
- movdqa -16(%rdi, %rdx), %xmm0
- psrldq $5, %xmm0
- pcmpistri $0x3a,%xmm0, %xmm0
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- cmp %r11, %rcx
- jae LABEL(nibble_ashr_use_sse4_2_exit)
-# endif
- cmp $10, %ecx
- ja LABEL(nibble_ashr_5_use_sse4_2_restart)
-
- jmp LABEL(nibble_ashr_use_sse4_2_exit)
-
-/*
- * 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
- */
- .p2align 4
-LABEL(ashr_6_sse4_2):
- pxor %xmm0, %xmm0
- movdqa (%rdi), %xmm2
- movdqa (%rsi), %xmm1
- pcmpeqb %xmm1, %xmm0
- pslldq $10, %xmm2
- TOLOWER (%xmm1, %xmm2)
- pcmpeqb %xmm1, %xmm2
- psubb %xmm0, %xmm2
- pmovmskb %xmm2, %r9d
- shr %cl, %edx
- shr %cl, %r9d
- sub %r9d, %edx
- jnz LABEL(less32bytes_sse4_2)
- movdqa (%rdi), %xmm3
-
- UPDATE_STRNCMP_COUNTER
-
- pxor %xmm0, %xmm0
- mov $16, %rcx /* index for loads */
- mov $6, %r9d /* byte position left over from less32bytes case */
- /*
- * Setup %r10 value allows us to detect crossing a page boundary.
- * When %r10 goes positive we have crossed a page boundary and
- * need to do a nibble.
- */
- lea 6(%rdi), %r10
- and $0xfff, %r10 /* offset into 4K page */
- sub $0x1000, %r10 /* subtract 4K pagesize */
- mov %rcx, %rdx /* only for offset of sse4 instruction loop*/
-
- .p2align 4
-LABEL(loop_ashr_6_use_sse4_2):
- add $16, %r10
- jg LABEL(nibble_ashr_6_use_sse4_2)
-
-LABEL(nibble_ashr_6_use_sse4_2_restart):
- movdqa (%rdi, %rdx), %xmm0
- palignr $6, -16(%rdi, %rdx), %xmm0
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpistri $0x1a,(%rsi,%rdx), %xmm0
-# else
- movdqa (%rsi,%rdx), %xmm1
- TOLOWER (%xmm0, %xmm1)
- pcmpistri $0x1a, %xmm1, %xmm0
-# endif
- jbe LABEL(use_sse4_2_exit)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub $16, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
-
- add $16, %rdx
- add $16, %r10
- jg LABEL(nibble_ashr_6_use_sse4_2)
-
- movdqa (%rdi, %rdx), %xmm0
- palignr $6, -16(%rdi, %rdx), %xmm0
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpistri $0x1a,(%rsi,%rdx), %xmm0
-# else
- movdqa (%rsi,%rdx), %xmm1
- TOLOWER (%xmm0, %xmm1)
- pcmpistri $0x1a, %xmm1, %xmm0
-# endif
- jbe LABEL(use_sse4_2_exit)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub $16, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
- add $16, %rdx
- jmp LABEL(loop_ashr_6_use_sse4_2)
-
- .p2align 4
-LABEL(nibble_ashr_6_use_sse4_2):
- sub $0x1000, %r10
- movdqa -16(%rdi, %rdx), %xmm0
- psrldq $6, %xmm0
- pcmpistri $0x3a,%xmm0, %xmm0
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- cmp %r11, %rcx
- jae LABEL(nibble_ashr_use_sse4_2_exit)
-# endif
- cmp $9, %ecx
- ja LABEL(nibble_ashr_6_use_sse4_2_restart)
-
- jmp LABEL(nibble_ashr_use_sse4_2_exit)
-
-/*
- * 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
- */
- .p2align 4
-LABEL(ashr_7_sse4_2):
- pxor %xmm0, %xmm0
- movdqa (%rdi), %xmm2
- movdqa (%rsi), %xmm1
- pcmpeqb %xmm1, %xmm0
- pslldq $9, %xmm2
- TOLOWER (%xmm1, %xmm2)
- pcmpeqb %xmm1, %xmm2
- psubb %xmm0, %xmm2
- pmovmskb %xmm2, %r9d
- shr %cl, %edx
- shr %cl, %r9d
- sub %r9d, %edx
- jnz LABEL(less32bytes_sse4_2)
- movdqa (%rdi), %xmm3
-
- UPDATE_STRNCMP_COUNTER
-
- pxor %xmm0, %xmm0
- mov $16, %rcx /* index for loads */
- mov $7, %r9d /* byte position left over from less32bytes case */
- /*
- * Setup %r10 value allows us to detect crossing a page boundary.
- * When %r10 goes positive we have crossed a page boundary and
- * need to do a nibble.
- */
- lea 7(%rdi), %r10
- and $0xfff, %r10 /* offset into 4K page */
- sub $0x1000, %r10 /* subtract 4K pagesize */
- mov %rcx, %rdx /* only for offset of sse4 instruction loop*/
-
- .p2align 4
-LABEL(loop_ashr_7_use_sse4_2):
- add $16, %r10
- jg LABEL(nibble_ashr_7_use_sse4_2)
-
-LABEL(nibble_ashr_7_use_sse4_2_restart):
- movdqa (%rdi, %rdx), %xmm0
- palignr $7, -16(%rdi, %rdx), %xmm0
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpistri $0x1a,(%rsi,%rdx), %xmm0
-# else
- movdqa (%rsi,%rdx), %xmm1
- TOLOWER (%xmm0, %xmm1)
- pcmpistri $0x1a, %xmm1, %xmm0
-# endif
- jbe LABEL(use_sse4_2_exit)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub $16, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
-
- add $16, %rdx
- add $16, %r10
- jg LABEL(nibble_ashr_7_use_sse4_2)
-
- movdqa (%rdi, %rdx), %xmm0
- palignr $7, -16(%rdi, %rdx), %xmm0
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpistri $0x1a,(%rsi,%rdx), %xmm0
-# else
- movdqa (%rsi,%rdx), %xmm1
- TOLOWER (%xmm0, %xmm1)
- pcmpistri $0x1a, %xmm1, %xmm0
-# endif
- jbe LABEL(use_sse4_2_exit)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub $16, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
- add $16, %rdx
- jmp LABEL(loop_ashr_7_use_sse4_2)
-
- .p2align 4
-LABEL(nibble_ashr_7_use_sse4_2):
- sub $0x1000, %r10
- movdqa -16(%rdi, %rdx), %xmm0
- psrldq $7, %xmm0
- pcmpistri $0x3a,%xmm0, %xmm0
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- cmp %r11, %rcx
- jae LABEL(nibble_ashr_use_sse4_2_exit)
-# endif
- cmp $8, %ecx
- ja LABEL(nibble_ashr_7_use_sse4_2_restart)
-
- jmp LABEL(nibble_ashr_use_sse4_2_exit)
-
-/*
- * 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
- */
- .p2align 4
-LABEL(ashr_8_sse4_2):
- pxor %xmm0, %xmm0
- movdqa (%rdi), %xmm2
- movdqa (%rsi), %xmm1
- pcmpeqb %xmm1, %xmm0
- pslldq $8, %xmm2
- TOLOWER (%xmm1, %xmm2)
- pcmpeqb %xmm1, %xmm2
- psubb %xmm0, %xmm2
- pmovmskb %xmm2, %r9d
- shr %cl, %edx
- shr %cl, %r9d
- sub %r9d, %edx
- jnz LABEL(less32bytes_sse4_2)
- movdqa (%rdi), %xmm3
-
- UPDATE_STRNCMP_COUNTER
-
- pxor %xmm0, %xmm0
- mov $16, %rcx /* index for loads */
- mov $8, %r9d /* byte position left over from less32bytes case */
- /*
- * Setup %r10 value allows us to detect crossing a page boundary.
- * When %r10 goes positive we have crossed a page boundary and
- * need to do a nibble.
- */
- lea 8(%rdi), %r10
- and $0xfff, %r10 /* offset into 4K page */
- sub $0x1000, %r10 /* subtract 4K pagesize */
- mov %rcx, %rdx /* only for offset of sse4 instruction loop*/
-
- .p2align 4
-LABEL(loop_ashr_8_use_sse4_2):
- add $16, %r10
- jg LABEL(nibble_ashr_8_use_sse4_2)
-
-LABEL(nibble_ashr_8_use_sse4_2_restart):
- movdqa (%rdi, %rdx), %xmm0
- palignr $8, -16(%rdi, %rdx), %xmm0
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpistri $0x1a, (%rsi,%rdx), %xmm0
-# else
- movdqa (%rsi,%rdx), %xmm1
- TOLOWER (%xmm0, %xmm1)
- pcmpistri $0x1a, %xmm1, %xmm0
-# endif
- jbe LABEL(use_sse4_2_exit)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub $16, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
-
- add $16, %rdx
- add $16, %r10
- jg LABEL(nibble_ashr_8_use_sse4_2)
-
- movdqa (%rdi, %rdx), %xmm0
- palignr $8, -16(%rdi, %rdx), %xmm0
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpistri $0x1a, (%rsi,%rdx), %xmm0
-# else
- movdqa (%rsi,%rdx), %xmm1
- TOLOWER (%xmm0, %xmm1)
- pcmpistri $0x1a, %xmm1, %xmm0
-# endif
- jbe LABEL(use_sse4_2_exit)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub $16, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
- add $16, %rdx
- jmp LABEL(loop_ashr_8_use_sse4_2)
-
- .p2align 4
-LABEL(nibble_ashr_8_use_sse4_2):
- sub $0x1000, %r10
- movdqa -16(%rdi, %rdx), %xmm0
- psrldq $8, %xmm0
- pcmpistri $0x3a,%xmm0, %xmm0
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- cmp %r11, %rcx
- jae LABEL(nibble_ashr_use_sse4_2_exit)
-# endif
- cmp $7, %ecx
- ja LABEL(nibble_ashr_8_use_sse4_2_restart)
-
- jmp LABEL(nibble_ashr_use_sse4_2_exit)
-
-/*
- * 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
- */
- .p2align 4
-LABEL(ashr_9_sse4_2):
- pxor %xmm0, %xmm0
- movdqa (%rdi), %xmm2
- movdqa (%rsi), %xmm1
- pcmpeqb %xmm1, %xmm0
- pslldq $7, %xmm2
- TOLOWER (%xmm1, %xmm2)
- pcmpeqb %xmm1, %xmm2
- psubb %xmm0, %xmm2
- pmovmskb %xmm2, %r9d
- shr %cl, %edx
- shr %cl, %r9d
- sub %r9d, %edx
- jnz LABEL(less32bytes_sse4_2)
- movdqa (%rdi), %xmm3
-
- UPDATE_STRNCMP_COUNTER
-
- pxor %xmm0, %xmm0
- mov $16, %rcx /* index for loads */
- mov $9, %r9d /* byte position left over from less32bytes case */
- /*
- * Setup %r10 value allows us to detect crossing a page boundary.
- * When %r10 goes positive we have crossed a page boundary and
- * need to do a nibble.
- */
- lea 9(%rdi), %r10
- and $0xfff, %r10 /* offset into 4K page */
- sub $0x1000, %r10 /* subtract 4K pagesize */
- mov %rcx, %rdx /* only for offset of sse4 instruction loop*/
-
- .p2align 4
-LABEL(loop_ashr_9_use_sse4_2):
- add $16, %r10
- jg LABEL(nibble_ashr_9_use_sse4_2)
-
-LABEL(nibble_ashr_9_use_sse4_2_restart):
- movdqa (%rdi, %rdx), %xmm0
-
- palignr $9, -16(%rdi, %rdx), %xmm0
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpistri $0x1a, (%rsi,%rdx), %xmm0
-# else
- movdqa (%rsi,%rdx), %xmm1
- TOLOWER (%xmm0, %xmm1)
- pcmpistri $0x1a, %xmm1, %xmm0
-# endif
- jbe LABEL(use_sse4_2_exit)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub $16, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
-
- add $16, %rdx
- add $16, %r10
- jg LABEL(nibble_ashr_9_use_sse4_2)
-
- movdqa (%rdi, %rdx), %xmm0
- palignr $9, -16(%rdi, %rdx), %xmm0
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpistri $0x1a, (%rsi,%rdx), %xmm0
-# else
- movdqa (%rsi,%rdx), %xmm1
- TOLOWER (%xmm0, %xmm1)
- pcmpistri $0x1a, %xmm1, %xmm0
-# endif
- jbe LABEL(use_sse4_2_exit)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub $16, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
- add $16, %rdx
- jmp LABEL(loop_ashr_9_use_sse4_2)
-
- .p2align 4
-LABEL(nibble_ashr_9_use_sse4_2):
- sub $0x1000, %r10
- movdqa -16(%rdi, %rdx), %xmm0
- psrldq $9, %xmm0
- pcmpistri $0x3a,%xmm0, %xmm0
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- cmp %r11, %rcx
- jae LABEL(nibble_ashr_use_sse4_2_exit)
-# endif
- cmp $6, %ecx
- ja LABEL(nibble_ashr_9_use_sse4_2_restart)
-
- jmp LABEL(nibble_ashr_use_sse4_2_exit)
-
-/*
- * 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
- */
- .p2align 4
-LABEL(ashr_10_sse4_2):
- pxor %xmm0, %xmm0
- movdqa (%rdi), %xmm2
- movdqa (%rsi), %xmm1
- pcmpeqb %xmm1, %xmm0
- pslldq $6, %xmm2
- TOLOWER (%xmm1, %xmm2)
- pcmpeqb %xmm1, %xmm2
- psubb %xmm0, %xmm2
- pmovmskb %xmm2, %r9d
- shr %cl, %edx
- shr %cl, %r9d
- sub %r9d, %edx
- jnz LABEL(less32bytes_sse4_2)
- movdqa (%rdi), %xmm3
-
- UPDATE_STRNCMP_COUNTER
-
- pxor %xmm0, %xmm0
- mov $16, %rcx /* index for loads */
- mov $10, %r9d /* byte position left over from less32bytes case */
- /*
- * Setup %r10 value allows us to detect crossing a page boundary.
- * When %r10 goes positive we have crossed a page boundary and
- * need to do a nibble.
- */
- lea 10(%rdi), %r10
- and $0xfff, %r10 /* offset into 4K page */
- sub $0x1000, %r10 /* subtract 4K pagesize */
- mov %rcx, %rdx /* only for offset of sse4 instruction loop*/
-
- .p2align 4
-LABEL(loop_ashr_10_use_sse4_2):
- add $16, %r10
- jg LABEL(nibble_ashr_10_use_sse4_2)
-
-LABEL(nibble_ashr_10_use_sse4_2_restart):
- movdqa (%rdi, %rdx), %xmm0
- palignr $10, -16(%rdi, %rdx), %xmm0
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpistri $0x1a, (%rsi,%rdx), %xmm0
-# else
- movdqa (%rsi,%rdx), %xmm1
- TOLOWER (%xmm0, %xmm1)
- pcmpistri $0x1a, %xmm1, %xmm0
-# endif
- jbe LABEL(use_sse4_2_exit)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub $16, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
-
- add $16, %rdx
- add $16, %r10
- jg LABEL(nibble_ashr_10_use_sse4_2)
-
- movdqa (%rdi, %rdx), %xmm0
- palignr $10, -16(%rdi, %rdx), %xmm0
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpistri $0x1a, (%rsi,%rdx), %xmm0
-# else
- movdqa (%rsi,%rdx), %xmm1
- TOLOWER (%xmm0, %xmm1)
- pcmpistri $0x1a, %xmm1, %xmm0
-# endif
- jbe LABEL(use_sse4_2_exit)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub $16, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
- add $16, %rdx
- jmp LABEL(loop_ashr_10_use_sse4_2)
-
- .p2align 4
-LABEL(nibble_ashr_10_use_sse4_2):
- sub $0x1000, %r10
- movdqa -16(%rdi, %rdx), %xmm0
- psrldq $10, %xmm0
- pcmpistri $0x3a,%xmm0, %xmm0
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- cmp %r11, %rcx
- jae LABEL(nibble_ashr_use_sse4_2_exit)
-# endif
- cmp $5, %ecx
- ja LABEL(nibble_ashr_10_use_sse4_2_restart)
-
- jmp LABEL(nibble_ashr_use_sse4_2_exit)
-
-/*
- * 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
- */
- .p2align 4
-LABEL(ashr_11_sse4_2):
- pxor %xmm0, %xmm0
- movdqa (%rdi), %xmm2
- movdqa (%rsi), %xmm1
- pcmpeqb %xmm1, %xmm0
- pslldq $5, %xmm2
- TOLOWER (%xmm1, %xmm2)
- pcmpeqb %xmm1, %xmm2
- psubb %xmm0, %xmm2
- pmovmskb %xmm2, %r9d
- shr %cl, %edx
- shr %cl, %r9d
- sub %r9d, %edx
- jnz LABEL(less32bytes_sse4_2)
- movdqa (%rdi), %xmm3
-
- UPDATE_STRNCMP_COUNTER
-
- pxor %xmm0, %xmm0
- mov $16, %rcx /* index for loads */
- mov $11, %r9d /* byte position left over from less32bytes case */
- /*
- * Setup %r10 value allows us to detect crossing a page boundary.
- * When %r10 goes positive we have crossed a page boundary and
- * need to do a nibble.
- */
- lea 11(%rdi), %r10
- and $0xfff, %r10 /* offset into 4K page */
- sub $0x1000, %r10 /* subtract 4K pagesize */
- mov %rcx, %rdx /* only for offset of sse4 instruction loop*/
-
- .p2align 4
-LABEL(loop_ashr_11_use_sse4_2):
- add $16, %r10
- jg LABEL(nibble_ashr_11_use_sse4_2)
-
-LABEL(nibble_ashr_11_use_sse4_2_restart):
- movdqa (%rdi, %rdx), %xmm0
- palignr $11, -16(%rdi, %rdx), %xmm0
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpistri $0x1a, (%rsi,%rdx), %xmm0
-# else
- movdqa (%rsi,%rdx), %xmm1
- TOLOWER (%xmm0, %xmm1)
- pcmpistri $0x1a, %xmm1, %xmm0
-# endif
- jbe LABEL(use_sse4_2_exit)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub $16, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
-
- add $16, %rdx
- add $16, %r10
- jg LABEL(nibble_ashr_11_use_sse4_2)
-
- movdqa (%rdi, %rdx), %xmm0
- palignr $11, -16(%rdi, %rdx), %xmm0
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpistri $0x1a, (%rsi,%rdx), %xmm0
-# else
- movdqa (%rsi,%rdx), %xmm1
- TOLOWER (%xmm0, %xmm1)
- pcmpistri $0x1a, %xmm1, %xmm0
-# endif
- jbe LABEL(use_sse4_2_exit)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub $16, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
- add $16, %rdx
- jmp LABEL(loop_ashr_11_use_sse4_2)
-
- .p2align 4
-LABEL(nibble_ashr_11_use_sse4_2):
- sub $0x1000, %r10
- movdqa -16(%rdi, %rdx), %xmm0
- psrldq $11, %xmm0
- pcmpistri $0x3a,%xmm0, %xmm0
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- cmp %r11, %rcx
- jae LABEL(nibble_ashr_use_sse4_2_exit)
-# endif
- cmp $4, %ecx
- ja LABEL(nibble_ashr_11_use_sse4_2_restart)
-
- jmp LABEL(nibble_ashr_use_sse4_2_exit)
-
-/*
- * 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
- */
- .p2align 4
-LABEL(ashr_12_sse4_2):
- pxor %xmm0, %xmm0
- movdqa (%rdi), %xmm2
- movdqa (%rsi), %xmm1
- pcmpeqb %xmm1, %xmm0
- pslldq $4, %xmm2
- TOLOWER (%xmm1, %xmm2)
- pcmpeqb %xmm1, %xmm2
- psubb %xmm0, %xmm2
- pmovmskb %xmm2, %r9d
- shr %cl, %edx
- shr %cl, %r9d
- sub %r9d, %edx
- jnz LABEL(less32bytes_sse4_2)
- movdqa (%rdi), %xmm3
-
- UPDATE_STRNCMP_COUNTER
-
- pxor %xmm0, %xmm0
- mov $16, %rcx /* index for loads */
- mov $12, %r9d /* byte position left over from less32bytes case */
- /*
- * Setup %r10 value allows us to detect crossing a page boundary.
- * When %r10 goes positive we have crossed a page boundary and
- * need to do a nibble.
- */
- lea 12(%rdi), %r10
- and $0xfff, %r10 /* offset into 4K page */
- sub $0x1000, %r10 /* subtract 4K pagesize */
- mov %rcx, %rdx /* only for offset of sse4 instruction loop*/
-
- .p2align 4
-LABEL(loop_ashr_12_use_sse4_2):
- add $16, %r10
- jg LABEL(nibble_ashr_12_use_sse4_2)
-
-LABEL(nibble_ashr_12_use_sse4_2_restart):
- movdqa (%rdi, %rdx), %xmm0
- palignr $12, -16(%rdi, %rdx), %xmm0
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpistri $0x1a, (%rsi,%rdx), %xmm0
-# else
- movdqa (%rsi,%rdx), %xmm1
- TOLOWER (%xmm0, %xmm1)
- pcmpistri $0x1a, %xmm1, %xmm0
-# endif
- jbe LABEL(use_sse4_2_exit)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub $16, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
-
- add $16, %rdx
- add $16, %r10
- jg LABEL(nibble_ashr_12_use_sse4_2)
-
- movdqa (%rdi, %rdx), %xmm0
- palignr $12, -16(%rdi, %rdx), %xmm0
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpistri $0x1a, (%rsi,%rdx), %xmm0
-# else
- movdqa (%rsi,%rdx), %xmm1
- TOLOWER (%xmm0, %xmm1)
- pcmpistri $0x1a, %xmm1, %xmm0
# endif
- jbe LABEL(use_sse4_2_exit)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub $16, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
- add $16, %rdx
- jmp LABEL(loop_ashr_12_use_sse4_2)
-
- .p2align 4
-LABEL(nibble_ashr_12_use_sse4_2):
- sub $0x1000, %r10
- movdqa -16(%rdi, %rdx), %xmm0
- psrldq $12, %xmm0
- pcmpistri $0x3a,%xmm0, %xmm0
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- cmp %r11, %rcx
- jae LABEL(nibble_ashr_use_sse4_2_exit)
-# endif
- cmp $3, %ecx
- ja LABEL(nibble_ashr_12_use_sse4_2_restart)
-
- jmp LABEL(nibble_ashr_use_sse4_2_exit)
-
-/*
- * 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
- */
- .p2align 4
-LABEL(ashr_13_sse4_2):
- pxor %xmm0, %xmm0
- movdqa (%rdi), %xmm2
- movdqa (%rsi), %xmm1
- pcmpeqb %xmm1, %xmm0
- pslldq $3, %xmm2
- TOLOWER (%xmm1, %xmm2)
- pcmpeqb %xmm1, %xmm2
- psubb %xmm0, %xmm2
- pmovmskb %xmm2, %r9d
- shr %cl, %edx
- shr %cl, %r9d
- sub %r9d, %edx
- jnz LABEL(less32bytes_sse4_2)
- movdqa (%rdi), %xmm3
-
- UPDATE_STRNCMP_COUNTER
-
- pxor %xmm0, %xmm0
- mov $16, %rcx /* index for loads */
- mov $13, %r9d /* byte position left over from less32bytes case */
- /*
- * Setup %r10 value allows us to detect crossing a page boundary.
- * When %r10 goes positive we have crossed a page boundary and
- * need to do a nibble.
- */
- lea 13(%rdi), %r10
- and $0xfff, %r10 /* offset into 4K page */
- sub $0x1000, %r10 /* subtract 4K pagesize */
-
- mov %rcx, %rdx /* only for offset of sse4 instruction loop*/
-
- .p2align 4
-LABEL(loop_ashr_13_use_sse4_2):
- add $16, %r10
- jg LABEL(nibble_ashr_13_use_sse4_2)
-
-LABEL(nibble_ashr_13_use_sse4_2_restart):
- movdqa (%rdi, %rdx), %xmm0
- palignr $13, -16(%rdi, %rdx), %xmm0
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpistri $0x1a, (%rsi,%rdx), %xmm0
-# else
- movdqa (%rsi,%rdx), %xmm1
- TOLOWER (%xmm0, %xmm1)
- pcmpistri $0x1a, %xmm1, %xmm0
-# endif
- jbe LABEL(use_sse4_2_exit)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub $16, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
-
- add $16, %rdx
- add $16, %r10
- jg LABEL(nibble_ashr_13_use_sse4_2)
-
- movdqa (%rdi, %rdx), %xmm0
- palignr $13, -16(%rdi, %rdx), %xmm0
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpistri $0x1a, (%rsi,%rdx), %xmm0
-# else
- movdqa (%rsi,%rdx), %xmm1
- TOLOWER (%xmm0, %xmm1)
- pcmpistri $0x1a, %xmm1, %xmm0
-# endif
- jbe LABEL(use_sse4_2_exit)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub $16, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
- add $16, %rdx
- jmp LABEL(loop_ashr_13_use_sse4_2)
-
- .p2align 4
-LABEL(nibble_ashr_13_use_sse4_2):
- sub $0x1000, %r10
- movdqa -16(%rdi, %rdx), %xmm0
- psrldq $13, %xmm0
- pcmpistri $0x3a,%xmm0, %xmm0
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- cmp %r11, %rcx
- jae LABEL(nibble_ashr_use_sse4_2_exit)
-# endif
- cmp $2, %ecx
- ja LABEL(nibble_ashr_13_use_sse4_2_restart)
-
- jmp LABEL(nibble_ashr_use_sse4_2_exit)
-
-/*
- * 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
- */
- .p2align 4
-LABEL(ashr_14_sse4_2):
- pxor %xmm0, %xmm0
- movdqa (%rdi), %xmm2
- movdqa (%rsi), %xmm1
- pcmpeqb %xmm1, %xmm0
- pslldq $2, %xmm2
- TOLOWER (%xmm1, %xmm2)
- pcmpeqb %xmm1, %xmm2
- psubb %xmm0, %xmm2
- pmovmskb %xmm2, %r9d
- shr %cl, %edx
- shr %cl, %r9d
- sub %r9d, %edx
- jnz LABEL(less32bytes_sse4_2)
- movdqa (%rdi), %xmm3
-
- UPDATE_STRNCMP_COUNTER
-
- pxor %xmm0, %xmm0
- mov $16, %rcx /* index for loads */
- mov $14, %r9d /* byte position left over from less32bytes case */
- /*
- * Setup %r10 value allows us to detect crossing a page boundary.
- * When %r10 goes positive we have crossed a page boundary and
- * need to do a nibble.
- */
- lea 14(%rdi), %r10
- and $0xfff, %r10 /* offset into 4K page */
- sub $0x1000, %r10 /* subtract 4K pagesize */
-
- mov %rcx, %rdx /* only for offset of sse4 instruction loop*/
-
- .p2align 4
-LABEL(loop_ashr_14_use_sse4_2):
- add $16, %r10
- jg LABEL(nibble_ashr_14_use_sse4_2)
-
-LABEL(nibble_ashr_14_use_sse4_2_restart):
- movdqa (%rdi, %rdx), %xmm0
- palignr $14, -16(%rdi, %rdx), %xmm0
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpistri $0x1a, (%rsi,%rdx), %xmm0
-# else
- movdqa (%rsi,%rdx), %xmm1
- TOLOWER (%xmm0, %xmm1)
- pcmpistri $0x1a, %xmm1, %xmm0
-# endif
- jbe LABEL(use_sse4_2_exit)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub $16, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
-
- add $16, %rdx
- add $16, %r10
- jg LABEL(nibble_ashr_14_use_sse4_2)
-
- movdqa (%rdi, %rdx), %xmm0
- palignr $14, -16(%rdi, %rdx), %xmm0
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpistri $0x1a, (%rsi,%rdx), %xmm0
-# else
- movdqa (%rsi,%rdx), %xmm1
- TOLOWER (%xmm0, %xmm1)
- pcmpistri $0x1a, %xmm1, %xmm0
-# endif
- jbe LABEL(use_sse4_2_exit)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub $16, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
- add $16, %rdx
- jmp LABEL(loop_ashr_14_use_sse4_2)
-
- .p2align 4
-LABEL(nibble_ashr_14_use_sse4_2):
- sub $0x1000, %r10
- movdqa -16(%rdi, %rdx), %xmm0
- psrldq $14, %xmm0
- pcmpistri $0x3a,%xmm0, %xmm0
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- cmp %r11, %rcx
- jae LABEL(nibble_ashr_use_sse4_2_exit)
-# endif
- cmp $1, %ecx
- ja LABEL(nibble_ashr_14_use_sse4_2_restart)
-
- jmp LABEL(nibble_ashr_use_sse4_2_exit)
-
-/*
- * 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
- */
- .p2align 4
-LABEL(ashr_15_sse4_2):
- pxor %xmm0, %xmm0
- movdqa (%rdi), %xmm2
- movdqa (%rsi), %xmm1
- pcmpeqb %xmm1, %xmm0
- pslldq $1, %xmm2
- TOLOWER (%xmm1, %xmm2)
- pcmpeqb %xmm1, %xmm2
- psubb %xmm0, %xmm2
- pmovmskb %xmm2, %r9d
- shr %cl, %edx
- shr %cl, %r9d
- sub %r9d, %edx
- jnz LABEL(less32bytes_sse4_2)
-
- movdqa (%rdi), %xmm3
-
- UPDATE_STRNCMP_COUNTER
-
- pxor %xmm0, %xmm0
- mov $16, %rcx /* index for loads */
- mov $15, %r9d /* byte position left over from less32bytes case */
- /*
- * Setup %r10 value allows us to detect crossing a page boundary.
- * When %r10 goes positive we have crossed a page boundary and
- * need to do a nibble.
- */
- lea 15(%rdi), %r10
- and $0xfff, %r10 /* offset into 4K page */
-
- sub $0x1000, %r10 /* subtract 4K pagesize */
-
- mov %rcx, %rdx /* only for offset of sse4 instruction loop*/
-
- .p2align 4
-LABEL(loop_ashr_15_use_sse4_2):
- add $16, %r10
- jg LABEL(nibble_ashr_15_use_sse4_2)
-
-LABEL(nibble_ashr_15_use_sse4_2_restart):
- movdqa (%rdi, %rdx), %xmm0
- palignr $15, -16(%rdi, %rdx), %xmm0
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpistri $0x1a, (%rsi,%rdx), %xmm0
-# else
- movdqa (%rsi,%rdx), %xmm1
- TOLOWER (%xmm0, %xmm1)
- pcmpistri $0x1a, %xmm1, %xmm0
-# endif
- jbe LABEL(use_sse4_2_exit)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub $16, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
-
- add $16, %rdx
- add $16, %r10
- jg LABEL(nibble_ashr_15_use_sse4_2)
-
- movdqa (%rdi, %rdx), %xmm0
- palignr $15, -16(%rdi, %rdx), %xmm0
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpistri $0x1a, (%rsi,%rdx), %xmm0
-# else
- movdqa (%rsi,%rdx), %xmm1
- TOLOWER (%xmm0, %xmm1)
- pcmpistri $0x1a, %xmm1, %xmm0
-# endif
- jbe LABEL(use_sse4_2_exit)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub $16, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
- add $16, %rdx
- jmp LABEL(loop_ashr_15_use_sse4_2)
-
- .p2align 4
-LABEL(nibble_ashr_15_use_sse4_2):
- sub $0x1000, %r10
- movdqa -16(%rdi, %rdx), %xmm0
- psrldq $15, %xmm0
- pcmpistri $0x3a,%xmm0, %xmm0
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- cmp %r11, %rcx
- jae LABEL(nibble_ashr_use_sse4_2_exit)
-# endif
- cmp $0, %ecx
- ja LABEL(nibble_ashr_15_use_sse4_2_restart)
-
-LABEL(nibble_ashr_use_sse4_2_exit):
-# if !defined USE_AS_STRCASECMP_L && !defined USE_AS_STRNCASECMP_L
- pcmpistri $0x1a,(%rsi,%rdx), %xmm0
-# else
- movdqa (%rsi,%rdx), %xmm1
- TOLOWER (%xmm0, %xmm1)
- pcmpistri $0x1a, %xmm1, %xmm0
-# endif
- .p2align 4
-LABEL(use_sse4_2_exit):
- jnc LABEL(strcmp_exitz_sse4_2)
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub %rcx, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
- add %rcx, %rdx
- lea -16(%rdi, %r9), %rdi
- movzbl (%rdi, %rdx), %eax
- movzbl (%rsi, %rdx), %edx
- test %r8d, %r8d
- jz LABEL(use_sse4_2_ret_sse4_2)
- xchg %eax, %edx
-LABEL(use_sse4_2_ret_sse4_2):
-# if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L
- leaq _nl_C_LC_CTYPE_tolower+128*4(%rip), %rcx
- movl (%rcx,%rdx,4), %edx
- movl (%rcx,%rax,4), %eax
-# endif
-
- sub %edx, %eax
- ret
-
-LABEL(less32bytes_sse4_2):
- lea (%rdi, %rax), %rdi /* locate the exact address for first operand(rdi) */
- lea (%rsi, %rcx), %rsi /* locate the exact address for second operand(rsi) */
- test %r8d, %r8d
- jz LABEL(ret_sse4_2)
- xchg %rsi, %rdi /* recover original order according to flag(%r8d) */
-
- .p2align 4
-LABEL(ret_sse4_2):
-LABEL(less16bytes_sse4_2):
- bsf %rdx, %rdx /* find and store bit index in %rdx */
-
-# if defined USE_AS_STRNCMP || defined USE_AS_STRNCASECMP_L
- sub %rdx, %r11
- jbe LABEL(strcmp_exitz_sse4_2)
-# endif
- movzbl (%rsi, %rdx), %ecx
- movzbl (%rdi, %rdx), %eax
-
-# if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L
- leaq _nl_C_LC_CTYPE_tolower+128*4(%rip), %rdx
- movl (%rdx,%rcx,4), %ecx
- movl (%rdx,%rax,4), %eax
-# endif
-
- sub %ecx, %eax
- ret
-
-LABEL(strcmp_exitz_sse4_2):
- xor %eax, %eax
- ret
-
- .p2align 4
- // XXX Same as code above
-LABEL(Byte0_sse4_2):
- movzx (%rsi), %ecx
- movzx (%rdi), %eax
-
-# if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L
- leaq _nl_C_LC_CTYPE_tolower+128*4(%rip), %rdx
- movl (%rdx,%rcx,4), %ecx
- movl (%rdx,%rax,4), %eax
-# endif
-
- sub %ecx, %eax
- ret
- cfi_endproc
- .size STRCMP_SSE42, .-STRCMP_SSE42
-
-# undef UCLOW_reg
-# undef UCHIGH_reg
-# undef LCQWORD_reg
-# undef TOLOWER
-
- /* Put all SSE 4.2 functions together. */
- .section .rodata.sse4.2,"a",@progbits
- .p2align 3
-LABEL(unaligned_table_sse4_2):
- .int LABEL(ashr_1_sse4_2) - LABEL(unaligned_table_sse4_2)
- .int LABEL(ashr_2_sse4_2) - LABEL(unaligned_table_sse4_2)
- .int LABEL(ashr_3_sse4_2) - LABEL(unaligned_table_sse4_2)
- .int LABEL(ashr_4_sse4_2) - LABEL(unaligned_table_sse4_2)
- .int LABEL(ashr_5_sse4_2) - LABEL(unaligned_table_sse4_2)
- .int LABEL(ashr_6_sse4_2) - LABEL(unaligned_table_sse4_2)
- .int LABEL(ashr_7_sse4_2) - LABEL(unaligned_table_sse4_2)
- .int LABEL(ashr_8_sse4_2) - LABEL(unaligned_table_sse4_2)
- .int LABEL(ashr_9_sse4_2) - LABEL(unaligned_table_sse4_2)
- .int LABEL(ashr_10_sse4_2) - LABEL(unaligned_table_sse4_2)
- .int LABEL(ashr_11_sse4_2) - LABEL(unaligned_table_sse4_2)
- .int LABEL(ashr_12_sse4_2) - LABEL(unaligned_table_sse4_2)
- .int LABEL(ashr_13_sse4_2) - LABEL(unaligned_table_sse4_2)
- .int LABEL(ashr_14_sse4_2) - LABEL(unaligned_table_sse4_2)
- .int LABEL(ashr_15_sse4_2) - LABEL(unaligned_table_sse4_2)
- .int LABEL(ashr_0_sse4_2) - LABEL(unaligned_table_sse4_2)
# undef ENTRY