hash/crc32: fix incorrect checksums with avx512+race

CRC value loaded from incorrect register, which happened
to line up with actual register on default compile.
Therefore failures would only show up with -race.

Add regression test with longer payloads.

Fix regression from CL 689435.

Fixes #74767.

Change-Id: Ib63ddade998a5630297b285f3d566361f36f28e5
GitHub-Last-Rev: 5074012d2cbc8fdb78c1fb2ada72e3bd98c985c2
GitHub-Pull-Request: golang/go#74775
Reviewed-on: https://go-review.googlesource.com/c/go/+/690855
Reviewed-by: Keith Randall <khr@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Auto-Submit: Keith Randall <khr@golang.org>
Reviewed-by: David Chase <drchase@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>

diff --git a/src/hash/crc32/crc32_amd64.s b/src/hash/crc32/crc32_amd64.s
index 4c482dc4a7f1d357952d04421ee3810293018a67..85d8b67a716de55c599a14dfd9d9a214b197afcf 100644 (file)
--- a/src/hash/crc32/crc32_amd64.s
+++ b/src/hash/crc32/crc32_amd64.s
@@ -173,11 +173,11 @@ TEXT ·ieeeCLMUL(SB),NOSPLIT,$0
        CMPQ CX, $1024
        JL   useSSE42
 
-       // Use AVX512
-       VPXORQ    Z0, Z0, Z0
-       VMOVQ     AX, X0
+       // Use AVX512. Zero upper and Z10 and load initial CRC into lower part of Z10.
+       VPXORQ    Z10, Z10, Z10
+       VMOVAPS   X0, X10
        VMOVDQU64 (SI), Z1
-       VPXORQ    Z0, Z1, Z1 // Merge initial CRC value into Z1
+       VPXORQ    Z10, Z1, Z1 // Merge initial CRC value into Z1
        ADDQ      $64, SI    // buf+=64
        SUBQ      $64, CX    // len-=64
 

diff --git a/src/hash/crc32/crc32_test.go b/src/hash/crc32/crc32_test.go
index 40acd7da4f84ec1b738fc45888f508fc38a228d7..eb5e73c3943b1f441c90a6166486025a999cf167 100644 (file)
--- a/src/hash/crc32/crc32_test.go
+++ b/src/hash/crc32/crc32_test.go
@@ -11,6 +11,7 @@ import (
        "internal/testhash"
        "io"
        "math/rand"
+       "strings"
        "testing"
 )
 
@@ -67,6 +68,8 @@ var golden = []test{
        {0x8c79fd79, 0x297a88ed, "Even if I could be Shakespeare, I think I should still choose to be Faraday. - A. Huxley", "crc\x01ʇ\x91Ml+\xb8\xa7", "crc\x01wB\x84\x81\xbf\xd6S\xdd"},
        {0xa20b7167, 0x66ed1d8b, "The fugacity of a constituent in a mixture of gases at a given temperature is proportional to its mole fraction.  Lewis-Randall Rule", "crc\x01ʇ\x91M<lR[", "crc\x01wB\x84\x81{\xaco\xb1"},
        {0x8e0bb443, 0xdcded527, "How can you write a big system without C++?  -Paul Glick", "crc\x01ʇ\x91M\x0e\x88\x89\xed", "crc\x01wB\x84\x813\xd7C\u007f"},
+       {0x1010dab0, 0x8a11661f, strings.Repeat("01234567", 1024), "crc\x01ʇ\x91M\x92\xe5\xba\xf3", "crc\x01wB\x84\x81\x1a\x02\x88Y"},
+       {0x772d04d7, 0x5a6f5c45, strings.Repeat("a", 1024+65), "crc\x01ʇ\x91M\xe7Љ\xd1", "crc\x01wB\x84\x81\x95B\xa9("},
 }
 
 // testGoldenIEEE verifies that the given function returns