diff options
Diffstat (limited to 'sysdeps/x86_64/multiarch/strcat-avx2.S')
-rw-r--r-- | sysdeps/x86_64/multiarch/strcat-avx2.S | 268 |
1 files changed, 6 insertions, 262 deletions
diff --git a/sysdeps/x86_64/multiarch/strcat-avx2.S b/sysdeps/x86_64/multiarch/strcat-avx2.S index d9b7fb2a43..3f914fa342 100644 --- a/sysdeps/x86_64/multiarch/strcat-avx2.S +++ b/sysdeps/x86_64/multiarch/strcat-avx2.S @@ -16,266 +16,10 @@ License along with the GNU C Library; if not, see <https://www.gnu.org/licenses/>. */ -#include <isa-level.h> - -#if ISA_SHOULD_BUILD (3) - - -# include <sysdep.h> - -# ifndef STRCAT -# define STRCAT __strcat_avx2 -# endif - -# define USE_AS_STRCAT - -/* Number of bytes in a vector register */ -# define VEC_SIZE 32 - -# ifndef SECTION -# define SECTION(p) p##.avx -# endif - - .section SECTION(.text),"ax",@progbits -ENTRY (STRCAT) - mov %rdi, %r9 -# ifdef USE_AS_STRNCAT - mov %rdx, %r8 -# endif - - xor %eax, %eax - mov %edi, %ecx - and $((VEC_SIZE * 4) - 1), %ecx - vpxor %xmm6, %xmm6, %xmm6 - cmp $(VEC_SIZE * 3), %ecx - ja L(fourth_vector_boundary) - vpcmpeqb (%rdi), %ymm6, %ymm0 - vpmovmskb %ymm0, %edx - test %edx, %edx - jnz L(exit_null_on_first_vector) - mov %rdi, %rax - and $-VEC_SIZE, %rax - jmp L(align_vec_size_start) -L(fourth_vector_boundary): - mov %rdi, %rax - and $-VEC_SIZE, %rax - vpcmpeqb (%rax), %ymm6, %ymm0 - mov $-1, %r10d - sub %rax, %rcx - shl %cl, %r10d - vpmovmskb %ymm0, %edx - and %r10d, %edx - jnz L(exit) - -L(align_vec_size_start): - vpcmpeqb VEC_SIZE(%rax), %ymm6, %ymm0 - vpmovmskb %ymm0, %edx - test %edx, %edx - jnz L(exit_null_on_second_vector) - - vpcmpeqb (VEC_SIZE * 2)(%rax), %ymm6, %ymm1 - vpmovmskb %ymm1, %edx - test %edx, %edx - jnz L(exit_null_on_third_vector) - - vpcmpeqb (VEC_SIZE * 3)(%rax), %ymm6, %ymm2 - vpmovmskb %ymm2, %edx - test %edx, %edx - jnz L(exit_null_on_fourth_vector) - - vpcmpeqb (VEC_SIZE * 4)(%rax), %ymm6, %ymm3 - vpmovmskb %ymm3, %edx - test %edx, %edx - jnz L(exit_null_on_fifth_vector) - - vpcmpeqb (VEC_SIZE * 5)(%rax), %ymm6, %ymm0 - add $(VEC_SIZE * 4), %rax - vpmovmskb %ymm0, %edx - test %edx, %edx - jnz L(exit_null_on_second_vector) - - vpcmpeqb (VEC_SIZE * 2)(%rax), %ymm6, %ymm1 - vpmovmskb %ymm1, %edx - test %edx, %edx - jnz L(exit_null_on_third_vector) - - vpcmpeqb (VEC_SIZE * 3)(%rax), %ymm6, %ymm2 - vpmovmskb %ymm2, %edx - test %edx, %edx - jnz L(exit_null_on_fourth_vector) - - vpcmpeqb (VEC_SIZE * 4)(%rax), %ymm6, %ymm3 - vpmovmskb %ymm3, %edx - test %edx, %edx - jnz L(exit_null_on_fifth_vector) - - vpcmpeqb (VEC_SIZE * 5)(%rax), %ymm6, %ymm0 - add $(VEC_SIZE * 4), %rax - vpmovmskb %ymm0, %edx - test %edx, %edx - jnz L(exit_null_on_second_vector) - - vpcmpeqb (VEC_SIZE * 2)(%rax), %ymm6, %ymm1 - vpmovmskb %ymm1, %edx - test %edx, %edx - jnz L(exit_null_on_third_vector) - - vpcmpeqb (VEC_SIZE * 3)(%rax), %ymm6, %ymm2 - vpmovmskb %ymm2, %edx - test %edx, %edx - jnz L(exit_null_on_fourth_vector) - - vpcmpeqb (VEC_SIZE * 4)(%rax), %ymm6, %ymm3 - vpmovmskb %ymm3, %edx - test %edx, %edx - jnz L(exit_null_on_fifth_vector) - - vpcmpeqb (VEC_SIZE * 5)(%rax), %ymm6, %ymm0 - add $(VEC_SIZE * 4), %rax - vpmovmskb %ymm0, %edx - test %edx, %edx - jnz L(exit_null_on_second_vector) - - vpcmpeqb (VEC_SIZE * 2)(%rax), %ymm6, %ymm1 - vpmovmskb %ymm1, %edx - test %edx, %edx - jnz L(exit_null_on_third_vector) - - vpcmpeqb (VEC_SIZE * 3)(%rax), %ymm6, %ymm2 - vpmovmskb %ymm2, %edx - test %edx, %edx - jnz L(exit_null_on_fourth_vector) - - vpcmpeqb (VEC_SIZE * 4)(%rax), %ymm6, %ymm3 - vpmovmskb %ymm3, %edx - test %edx, %edx - jnz L(exit_null_on_fifth_vector) - - test $((VEC_SIZE * 4) - 1), %rax - jz L(align_four_vec_loop) - - vpcmpeqb (VEC_SIZE * 5)(%rax), %ymm6, %ymm0 - add $(VEC_SIZE * 5), %rax - vpmovmskb %ymm0, %edx - test %edx, %edx - jnz L(exit) - - test $((VEC_SIZE * 4) - 1), %rax - jz L(align_four_vec_loop) - - vpcmpeqb VEC_SIZE(%rax), %ymm6, %ymm1 - add $VEC_SIZE, %rax - vpmovmskb %ymm1, %edx - test %edx, %edx - jnz L(exit) - - test $((VEC_SIZE * 4) - 1), %rax - jz L(align_four_vec_loop) - - vpcmpeqb VEC_SIZE(%rax), %ymm6, %ymm2 - add $VEC_SIZE, %rax - vpmovmskb %ymm2, %edx - test %edx, %edx - jnz L(exit) - - test $((VEC_SIZE * 4) - 1), %rax - jz L(align_four_vec_loop) - - vpcmpeqb VEC_SIZE(%rax), %ymm6, %ymm3 - add $VEC_SIZE, %rax - vpmovmskb %ymm3, %edx - test %edx, %edx - jnz L(exit) - - add $VEC_SIZE, %rax - - .p2align 4 -L(align_four_vec_loop): - vmovaps (%rax), %ymm4 - vpminub VEC_SIZE(%rax), %ymm4, %ymm4 - vmovaps (VEC_SIZE * 2)(%rax), %ymm5 - vpminub (VEC_SIZE * 3)(%rax), %ymm5, %ymm5 - add $(VEC_SIZE * 4), %rax - vpminub %ymm4, %ymm5, %ymm5 - vpcmpeqb %ymm5, %ymm6, %ymm5 - vpmovmskb %ymm5, %edx - test %edx, %edx - jz L(align_four_vec_loop) - - vpcmpeqb -(VEC_SIZE * 4)(%rax), %ymm6, %ymm0 - sub $(VEC_SIZE * 5), %rax - vpmovmskb %ymm0, %edx - test %edx, %edx - jnz L(exit_null_on_second_vector) - - vpcmpeqb (VEC_SIZE * 2)(%rax), %ymm6, %ymm1 - vpmovmskb %ymm1, %edx - test %edx, %edx - jnz L(exit_null_on_third_vector) - - vpcmpeqb (VEC_SIZE * 3)(%rax), %ymm6, %ymm2 - vpmovmskb %ymm2, %edx - test %edx, %edx - jnz L(exit_null_on_fourth_vector) - - vpcmpeqb (VEC_SIZE * 4)(%rax), %ymm6, %ymm3 - vpmovmskb %ymm3, %edx - sub %rdi, %rax - bsf %rdx, %rdx - add %rdx, %rax - add $(VEC_SIZE * 4), %rax - jmp L(StartStrcpyPart) - - .p2align 4 -L(exit): - sub %rdi, %rax -L(exit_null_on_first_vector): - bsf %rdx, %rdx - add %rdx, %rax - jmp L(StartStrcpyPart) - - .p2align 4 -L(exit_null_on_second_vector): - sub %rdi, %rax - bsf %rdx, %rdx - add %rdx, %rax - add $VEC_SIZE, %rax - jmp L(StartStrcpyPart) - - .p2align 4 -L(exit_null_on_third_vector): - sub %rdi, %rax - bsf %rdx, %rdx - add %rdx, %rax - add $(VEC_SIZE * 2), %rax - jmp L(StartStrcpyPart) - - .p2align 4 -L(exit_null_on_fourth_vector): - sub %rdi, %rax - bsf %rdx, %rdx - add %rdx, %rax - add $(VEC_SIZE * 3), %rax - jmp L(StartStrcpyPart) - - .p2align 4 -L(exit_null_on_fifth_vector): - sub %rdi, %rax - bsf %rdx, %rdx - add %rdx, %rax - add $(VEC_SIZE * 4), %rax - - .p2align 4 -L(StartStrcpyPart): - lea (%r9, %rax), %rdi - mov %rsi, %rcx - mov %r9, %rax /* save result */ - -# ifdef USE_AS_STRNCAT - test %r8, %r8 - jz L(ExitZero) -# define USE_AS_STRNCPY -# endif - -# include "strcpy-avx2.S" +#ifndef STRCAT +# define STRCAT __strcat_avx2 #endif + +#define USE_AS_STRCAT +#define STRCPY STRCAT +#include "strcpy-avx2.S" |