summaryrefslogtreecommitdiffstats
path: root/third_party/boringssl/mac-x86_64/crypto/poly1305/poly1305-x86_64.S
diff options
context:
space:
mode:
Diffstat (limited to 'third_party/boringssl/mac-x86_64/crypto/poly1305/poly1305-x86_64.S')
-rw-r--r--third_party/boringssl/mac-x86_64/crypto/poly1305/poly1305-x86_64.S88
1 files changed, 44 insertions, 44 deletions
diff --git a/third_party/boringssl/mac-x86_64/crypto/poly1305/poly1305-x86_64.S b/third_party/boringssl/mac-x86_64/crypto/poly1305/poly1305-x86_64.S
index 78c887dd..13f028d 100644
--- a/third_party/boringssl/mac-x86_64/crypto/poly1305/poly1305-x86_64.S
+++ b/third_party/boringssl/mac-x86_64/crypto/poly1305/poly1305-x86_64.S
@@ -1101,6 +1101,20 @@ L$short_tail_avx:
+ vpsrldq $8,%xmm14,%xmm9
+ vpsrldq $8,%xmm13,%xmm8
+ vpsrldq $8,%xmm11,%xmm6
+ vpsrldq $8,%xmm10,%xmm5
+ vpsrldq $8,%xmm12,%xmm7
+ vpaddq %xmm8,%xmm13,%xmm13
+ vpaddq %xmm9,%xmm14,%xmm14
+ vpaddq %xmm5,%xmm10,%xmm10
+ vpaddq %xmm6,%xmm11,%xmm11
+ vpaddq %xmm7,%xmm12,%xmm12
+
+
+
+
vpsrlq $26,%xmm13,%xmm3
vpand %xmm15,%xmm13,%xmm13
vpaddq %xmm3,%xmm14,%xmm14
@@ -1132,25 +1146,11 @@ L$short_tail_avx:
vpand %xmm15,%xmm13,%xmm13
vpaddq %xmm3,%xmm14,%xmm14
-
-
-
- vpsrldq $8,%xmm12,%xmm7
- vpsrldq $8,%xmm10,%xmm5
- vpsrldq $8,%xmm11,%xmm6
- vpsrldq $8,%xmm13,%xmm8
- vpsrldq $8,%xmm14,%xmm9
- vpaddq %xmm7,%xmm12,%xmm2
- vpaddq %xmm5,%xmm10,%xmm0
- vpaddq %xmm6,%xmm11,%xmm1
- vpaddq %xmm8,%xmm13,%xmm3
- vpaddq %xmm9,%xmm14,%xmm4
-
- vmovd %xmm0,-112(%rdi)
- vmovd %xmm1,-108(%rdi)
- vmovd %xmm2,-104(%rdi)
- vmovd %xmm3,-100(%rdi)
- vmovd %xmm4,-96(%rdi)
+ vmovd %xmm10,-112(%rdi)
+ vmovd %xmm11,-108(%rdi)
+ vmovd %xmm12,-104(%rdi)
+ vmovd %xmm13,-100(%rdi)
+ vmovd %xmm14,-96(%rdi)
leaq 88(%r11),%rsp
vzeroupper
.byte 0xf3,0xc3
@@ -1746,6 +1746,31 @@ L$tail_avx2:
+ vpsrldq $8,%ymm12,%ymm8
+ vpsrldq $8,%ymm2,%ymm9
+ vpsrldq $8,%ymm3,%ymm10
+ vpsrldq $8,%ymm4,%ymm6
+ vpsrldq $8,%ymm0,%ymm7
+ vpaddq %ymm8,%ymm12,%ymm12
+ vpaddq %ymm9,%ymm2,%ymm2
+ vpaddq %ymm10,%ymm3,%ymm3
+ vpaddq %ymm6,%ymm4,%ymm4
+ vpaddq %ymm7,%ymm0,%ymm0
+
+ vpermq $0x2,%ymm3,%ymm10
+ vpermq $0x2,%ymm4,%ymm6
+ vpermq $0x2,%ymm0,%ymm7
+ vpermq $0x2,%ymm12,%ymm8
+ vpermq $0x2,%ymm2,%ymm9
+ vpaddq %ymm10,%ymm3,%ymm3
+ vpaddq %ymm6,%ymm4,%ymm4
+ vpaddq %ymm7,%ymm0,%ymm0
+ vpaddq %ymm8,%ymm12,%ymm12
+ vpaddq %ymm9,%ymm2,%ymm2
+
+
+
+
vpsrlq $26,%ymm3,%ymm14
vpand %ymm5,%ymm3,%ymm3
vpaddq %ymm14,%ymm4,%ymm4
@@ -1777,31 +1802,6 @@ L$tail_avx2:
vpand %ymm5,%ymm3,%ymm3
vpaddq %ymm14,%ymm4,%ymm4
-
-
-
- vpsrldq $8,%ymm2,%ymm9
- vpsrldq $8,%ymm0,%ymm7
- vpsrldq $8,%ymm1,%ymm8
- vpsrldq $8,%ymm3,%ymm10
- vpsrldq $8,%ymm4,%ymm6
- vpaddq %ymm9,%ymm2,%ymm2
- vpaddq %ymm7,%ymm0,%ymm0
- vpaddq %ymm8,%ymm1,%ymm1
- vpaddq %ymm10,%ymm3,%ymm3
- vpaddq %ymm6,%ymm4,%ymm4
-
- vpermq $0x2,%ymm2,%ymm9
- vpermq $0x2,%ymm0,%ymm7
- vpermq $0x2,%ymm1,%ymm8
- vpermq $0x2,%ymm3,%ymm10
- vpermq $0x2,%ymm4,%ymm6
- vpaddq %ymm9,%ymm2,%ymm2
- vpaddq %ymm7,%ymm0,%ymm0
- vpaddq %ymm8,%ymm1,%ymm1
- vpaddq %ymm10,%ymm3,%ymm3
- vpaddq %ymm6,%ymm4,%ymm4
-
vmovd %xmm0,-112(%rdi)
vmovd %xmm1,-108(%rdi)
vmovd %xmm2,-104(%rdi)