diff options
Diffstat (limited to 'mac-x86_64/crypto/fipsmodule/p256-x86_64-asm.S')
| -rw-r--r-- | mac-x86_64/crypto/fipsmodule/p256-x86_64-asm.S | 140 |
1 files changed, 139 insertions, 1 deletions
diff --git a/mac-x86_64/crypto/fipsmodule/p256-x86_64-asm.S b/mac-x86_64/crypto/fipsmodule/p256-x86_64-asm.S index 2c58d48c..f7875772 100644 --- a/mac-x86_64/crypto/fipsmodule/p256-x86_64-asm.S +++ b/mac-x86_64/crypto/fipsmodule/p256-x86_64-asm.S @@ -516,6 +516,10 @@ __ecp_nistz256_sqr_montq: .p2align 5 _ecp_nistz256_select_w5: + leaq _OPENSSL_ia32cap_P(%rip),%rax + movq 8(%rax),%rax + testl $32,%eax + jnz L$avx2_select_w5 movdqa L$One(%rip),%xmm0 movd %edx,%xmm1 @@ -576,6 +580,10 @@ L$select_loop_sse_w5: .p2align 5 _ecp_nistz256_select_w7: + leaq _OPENSSL_ia32cap_P(%rip),%rax + movq 8(%rax),%rax + testl $32,%eax + jnz L$avx2_select_w7 movdqa L$One(%rip),%xmm8 movd %edx,%xmm1 @@ -617,12 +625,142 @@ L$select_loop_sse_w7: movdqu %xmm5,48(%rdi) .byte 0xf3,0xc3 + + + +.p2align 5 +ecp_nistz256_avx2_select_w5: +L$avx2_select_w5: + vzeroupper + vmovdqa L$Two(%rip),%ymm0 + + vpxor %ymm2,%ymm2,%ymm2 + vpxor %ymm3,%ymm3,%ymm3 + vpxor %ymm4,%ymm4,%ymm4 + + vmovdqa L$One(%rip),%ymm5 + vmovdqa L$Two(%rip),%ymm10 + + vmovd %edx,%xmm1 + vpermd %ymm1,%ymm2,%ymm1 + + movq $8,%rax +L$select_loop_avx2_w5: + + vmovdqa 0(%rsi),%ymm6 + vmovdqa 32(%rsi),%ymm7 + vmovdqa 64(%rsi),%ymm8 + + vmovdqa 96(%rsi),%ymm11 + vmovdqa 128(%rsi),%ymm12 + vmovdqa 160(%rsi),%ymm13 + + vpcmpeqd %ymm1,%ymm5,%ymm9 + vpcmpeqd %ymm1,%ymm10,%ymm14 + + vpaddd %ymm0,%ymm5,%ymm5 + vpaddd %ymm0,%ymm10,%ymm10 + leaq 192(%rsi),%rsi + + vpand %ymm9,%ymm6,%ymm6 + vpand %ymm9,%ymm7,%ymm7 + vpand %ymm9,%ymm8,%ymm8 + vpand %ymm14,%ymm11,%ymm11 + vpand %ymm14,%ymm12,%ymm12 + vpand %ymm14,%ymm13,%ymm13 + + vpxor %ymm6,%ymm2,%ymm2 + vpxor %ymm7,%ymm3,%ymm3 + vpxor %ymm8,%ymm4,%ymm4 + vpxor %ymm11,%ymm2,%ymm2 + vpxor %ymm12,%ymm3,%ymm3 + vpxor %ymm13,%ymm4,%ymm4 + + decq %rax + jnz L$select_loop_avx2_w5 + + vmovdqu %ymm2,0(%rdi) + vmovdqu %ymm3,32(%rdi) + vmovdqu %ymm4,64(%rdi) + vzeroupper + .byte 0xf3,0xc3 + + + + .globl _ecp_nistz256_avx2_select_w7 .private_extern _ecp_nistz256_avx2_select_w7 .p2align 5 _ecp_nistz256_avx2_select_w7: -.byte 0x0f,0x0b +L$avx2_select_w7: + vzeroupper + vmovdqa L$Three(%rip),%ymm0 + + vpxor %ymm2,%ymm2,%ymm2 + vpxor %ymm3,%ymm3,%ymm3 + + vmovdqa L$One(%rip),%ymm4 + vmovdqa L$Two(%rip),%ymm8 + vmovdqa L$Three(%rip),%ymm12 + + vmovd %edx,%xmm1 + vpermd %ymm1,%ymm2,%ymm1 + + + movq $21,%rax +L$select_loop_avx2_w7: + + vmovdqa 0(%rsi),%ymm5 + vmovdqa 32(%rsi),%ymm6 + + vmovdqa 64(%rsi),%ymm9 + vmovdqa 96(%rsi),%ymm10 + + vmovdqa 128(%rsi),%ymm13 + vmovdqa 160(%rsi),%ymm14 + + vpcmpeqd %ymm1,%ymm4,%ymm7 + vpcmpeqd %ymm1,%ymm8,%ymm11 + vpcmpeqd %ymm1,%ymm12,%ymm15 + + vpaddd %ymm0,%ymm4,%ymm4 + vpaddd %ymm0,%ymm8,%ymm8 + vpaddd %ymm0,%ymm12,%ymm12 + leaq 192(%rsi),%rsi + + vpand %ymm7,%ymm5,%ymm5 + vpand %ymm7,%ymm6,%ymm6 + vpand %ymm11,%ymm9,%ymm9 + vpand %ymm11,%ymm10,%ymm10 + vpand %ymm15,%ymm13,%ymm13 + vpand %ymm15,%ymm14,%ymm14 + + vpxor %ymm5,%ymm2,%ymm2 + vpxor %ymm6,%ymm3,%ymm3 + vpxor %ymm9,%ymm2,%ymm2 + vpxor %ymm10,%ymm3,%ymm3 + vpxor %ymm13,%ymm2,%ymm2 + vpxor %ymm14,%ymm3,%ymm3 + + decq %rax + jnz L$select_loop_avx2_w7 + + + vmovdqa 0(%rsi),%ymm5 + vmovdqa 32(%rsi),%ymm6 + + vpcmpeqd %ymm1,%ymm4,%ymm7 + + vpand %ymm7,%ymm5,%ymm5 + vpand %ymm7,%ymm6,%ymm6 + + vpxor %ymm5,%ymm2,%ymm2 + vpxor %ymm6,%ymm3,%ymm3 + + vmovdqu %ymm2,0(%rdi) + vmovdqu %ymm3,32(%rdi) + vzeroupper .byte 0xf3,0xc3 |