更新

deatil · Sep 23, 2024 · 10ecd04 · 10ecd04
1 parent 96cf22b
commit 10ecd04
Show file tree

Hide file tree

Showing 4 changed files with 11 additions and 5 deletions.
diff --git a/.github/workflows/test_qemu.yml b/.github/workflows/test_qemu.yml
@@ -14,7 +14,7 @@ jobs:
   test:
     strategy:
       matrix:
-        go-version: [1.20]
+        go-version: [1.21.x]
         arch: [arm64]
     runs-on: ubuntu-latest
     steps:

diff --git a/.github/workflows/test_sm_ni.yml b/.github/workflows/test_sm_ni.yml
@@ -14,7 +14,7 @@ jobs:
   test:
     strategy:
       matrix:
-        go-version: [1.20]
+        go-version: [1.21.x]
         arch: [arm64]
     runs-on: ubuntu-latest
     steps:

diff --git a/sm3/blocks_avx2_amd64.s b/sm3/blocks_avx2_amd64.s
@@ -104,7 +104,7 @@ GLOBL r08_mask<>(SB), 8, $32
 	VMOVDQU (i*32)(srcPtr5), e; \
 	VMOVDQU (i*32)(srcPtr6), f; \
 	VMOVDQU (i*32)(srcPtr7), g; \
-	VMOVDQU (i*32)(srcPtr8), h; \    
+	VMOVDQU (i*32)(srcPtr8), h; \
 	; \
 	TRANSPOSE_MATRIX(a, b, c, d, e, f, g, h, TMP1, TMP2, TMP3, TMP4); \
 	REV32(a, b, c, d, e, f, g, h); \
@@ -172,7 +172,7 @@ GLOBL r08_mask<>(SB), 8, $32
 // DST = (Y XOR Z) AND X XOR Z
 #define GG1(X, Y, Z, DST) \
 	VPXOR Y, Z, DST; \
-	VPAND X, DST, DST; \ 
+	VPAND X, DST, DST; \
 	VPXOR Z, DST, DST
 
 #define SS1SS2(index, a, e, SS1, SS2) \
@@ -196,7 +196,7 @@ GLOBL r08_mask<>(SB), 8, $32
 	SS1SS2(index, a, e, Y12, Y13); \
 	; \
 	FF0(a, b, c, Y14); \
-	VPADDD d, Y14, Y14; \ // (a XOR b XOR c) + d 
+	VPADDD d, Y14, Y14; \ // (a XOR b XOR c) + d
 	loadWord(Y10, index); \
 	loadWord(Y11, index+4); \
 	VPXOR Y10, Y11, Y11; \ //Wt XOR Wt+4

diff --git a/sm3/blocks_avx2_test.go b/sm3/blocks_avx2_test.go
@@ -7,6 +7,12 @@ import (
 	"testing"
 )
 
+// p || state || words
+// p = 64 * 8 * 2 = 1024
+// state = 8 * 32 = 256
+// words = 68 * 32 = 2176
+const preallocSizeBy8 = 3456
+
 func initState8() [8]*[8]uint32 {
 	d := new(digest)
 	d.Reset()