gopkg/fastxor
2
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

fix SSE segfault

Browse Source
This commit is contained in:
lukechampine
2018-07-06 17:10:20 -04:00
parent 58103ea98e
commit 62a404d90e
2 changed files with 33 additions and 54 deletions
Download Patch File Download Diff File Expand all files Collapse all files
xor_amd64.s
+19 -52
View File
@@ -8,68 +8,34 @@
#define N R12 #define N R12
// func xorBytesSSE(dst, a, b []byte, n int) // func xorBytesSSE(dst, a, b []byte, n int)
TEXT ·xorBytesSSE(SB), NOSPLIT ,$0 TEXT ·xorBytesSSE(SB), NOSPLIT, $0
MOVQ dst_data+0(FP), Dst MOVQ dst_data+0(FP), Dst
MOVQ a_data+24(FP), A MOVQ a_data+24(FP), A
MOVQ b_data+48(FP), B MOVQ b_data+48(FP), B
MOVQ n+72(FP), N MOVQ n+72(FP), N
XOR_LOOP_128_SSE:
CMPQ N, $128
JB XOR_LOOP_64_SSE
MOVOU (A), X0
MOVOU 16(A), X1
MOVOU 32(A), X2
MOVOU 48(A), X3
MOVOU 64(A), X4
MOVOU 80(A), X5
MOVOU 96(A), X6
MOVOU 112(A), X7
PXOR (B), X0
PXOR 16(B), X1
PXOR 32(B), X2
PXOR 48(B), X3
PXOR 64(B), X4
PXOR 80(B), X5
PXOR 96(B), X6
PXOR 112(B), X7
MOVOU X0, (Dst)
MOVOU X1, 16(Dst)
MOVOU X2, 32(Dst)
MOVOU X3, 48(Dst)
MOVOU X4, 64(Dst)
MOVOU X5, 80(Dst)
MOVOU X6, 96(Dst)
MOVOU X7, 112(Dst)
ADDQ $128, A
ADDQ $128, B
ADDQ $128, Dst
SUBQ $128, N
JNZ XOR_LOOP_128_SSE
RET
XOR_LOOP_64_SSE: XOR_LOOP_64_SSE:
CMPQ N, $64 CMPQ N, $64
JB XOR_LOOP_16_SSE JB XOR_LOOP_16_SSE
MOVOU (A), X0 MOVOU 0*16(A), X0
MOVOU 16(A), X1 MOVOU 1*16(A), X1
MOVOU 32(A), X2 MOVOU 2*16(A), X2
MOVOU 48(A), X3 MOVOU 3*16(A), X3
MOVOU 0*16(B), X4
MOVOU 1*16(B), X5
MOVOU 2*16(B), X6
MOVOU 3*16(B), X7
PXOR (B), X0 PXOR X4, X0
PXOR 16(B), X1 PXOR X5, X1
PXOR 32(B), X2 PXOR X6, X2
PXOR 48(B), X3 PXOR X7, X3
MOVOU X0, (Dst) MOVOU X0, 0*16(Dst)
MOVOU X1, 16(Dst) MOVOU X1, 1*16(Dst)
MOVOU X2, 32(Dst) MOVOU X2, 2*16(Dst)
MOVOU X3, 48(Dst) MOVOU X3, 3*16(Dst)
ADDQ $64, A ADDQ $64, A
ADDQ $64, B ADDQ $64, B
@@ -82,7 +48,8 @@ XOR_LOOP_16_SSE:
CMPQ N, $16 CMPQ N, $16
JB XOR_LOOP_FINAL_SSE JB XOR_LOOP_FINAL_SSE
MOVOU (A), X0 MOVOU (A), X0
PXOR (B), X0 MOVOU (B), X1
PXOR X1, X0
MOVOU X0, (Dst) MOVOU X0, (Dst)
ADDQ $16, A ADDQ $16, A
ADDQ $16, B ADDQ $16, B
xor_test.go
+14 -2
View File
@@ -56,8 +56,10 @@ func refBlock(dst, a, b []byte) {
func TestBytes(t *testing.T) { func TestBytes(t *testing.T) {
err := quick.Check(func(a, b []byte) bool { err := quick.Check(func(a, b []byte) bool {
// double size to increase chances of reaching 64 bytes // quadruple size to increase coverage
a = append(a, a...) a = append(a, a...)
a = append(a, a...)
b = append(b, b...)
b = append(b, b...) b = append(b, b...)
if len(a) < 8 { if len(a) < 8 {
return true return true
@@ -72,7 +74,17 @@ func TestBytes(t *testing.T) {
return bytes.Equal(dst1, dst2) return bytes.Equal(dst1, dst2)
}, &quick.Config{MaxCount: 10000}) }, &quick.Config{MaxCount: 10000})
if err != nil { if err != nil {
t.Fatal(err) t.Error(err)
}
// test a large slice
a := make([]byte, 512-1)
for i := range a {
a[i] = byte(i)
}
Bytes(a, a, a)
if !bytes.Equal(a, make([]byte, len(a))) {
t.Error("bad xor of large slice")
} }
} }