Revert "hash/crc32: improve the AMD64 implementation using SSE4.2"

This reverts commit 54d7de7dd62bab764125c48fd159bb938da078e1.

It was breaking non-amd64 builds.

Change-Id: I22650e922498eeeba3d4fa08bb4ea40a210c8f97
Reviewed-on: https://go-review.googlesource.com/27925
Reviewed-by: Keith Randall <khr@golang.org>

7 files changed, 14 insertions, 292 deletions

diff --git a/src/hash/crc32/crc32.go b/src/hash/crc32/crc32.go
index 57089a700d..c3ac7b80c3 100644
--- a/src/hash/crc32/crc32.go
+++ b/src/hash/crc32/crc32.go
@@ -52,14 +52,8 @@ var castagnoliTable8 *slicing8Table
 var castagnoliOnce sync.Once
 
 func castagnoliInit() {
-	// Call the arch-specific init function and let it decide if we will need
-	// the tables for the generic implementation.
-	needGenericTables := castagnoliInitArch()
-
-	if needGenericTables {
-		castagnoliTable = makeTable(Castagnoli)
-		castagnoliTable8 = makeTable8(Castagnoli)
-	}
+	castagnoliTable = makeTable(Castagnoli)
+	castagnoliTable8 = makeTable8(Castagnoli)
 }
 
 // IEEETable is the table for the IEEE polynomial.
diff --git a/src/hash/crc32/crc32_amd64.go b/src/hash/crc32/crc32_amd64.go
index a071cbcb88..a0180a12de 100644
--- a/src/hash/crc32/crc32_amd64.go
+++ b/src/hash/crc32/crc32_amd64.go
@@ -4,8 +4,6 @@
 
 package crc32
 
-import "unsafe"
-
 // This file contains the code to call the SSE 4.2 version of the Castagnoli
 // and IEEE CRC.
 
@@ -15,20 +13,11 @@ func haveSSE41() bool
 func haveSSE42() bool
 func haveCLMUL() bool
 
-// castagnoliSSE42 is defined in crc32_amd64.s and uses the SSE4.2 CRC32
+// castagnoliSSE42 is defined in crc_amd64.s and uses the SSE4.2 CRC32
 // instruction.
 //go:noescape
 func castagnoliSSE42(crc uint32, p []byte) uint32
 
-// castagnoliSSE42Triple is defined in crc32_amd64.s and uses the SSE4.2 CRC32
-// instruction.
-//go:noescape
-func castagnoliSSE42Triple(
-	crcA, crcB, crcC uint32,
-	a, b, c []byte,
-	rounds uint32,
-) (retA uint32, retB uint32, retC uint32)
-
 // ieeeCLMUL is defined in crc_amd64.s and uses the PCLMULQDQ
 // instruction as well as SSE 4.1.
 //go:noescape
@@ -37,160 +26,15 @@ func ieeeCLMUL(crc uint32, p []byte) uint32
 var sse42 = haveSSE42()
 var useFastIEEE = haveCLMUL() && haveSSE41()
 
-const castagnoliK1 = 168
-const castagnoliK2 = 1344
-
-type sse42Table [4]Table
-
-var castagnoliSSE42TableK1 *sse42Table
-var castagnoliSSE42TableK2 *sse42Table
-
-func castagnoliInitArch() (needGenericTables bool) {
-	if !sse42 {
-		return true
-	}
-	castagnoliSSE42TableK1 = new(sse42Table)
-	castagnoliSSE42TableK2 = new(sse42Table)
-	// See description in updateCastagnoli.
-	//    t[0][i] = CRC(i000, O)
-	//    t[1][i] = CRC(0i00, O)
-	//    t[2][i] = CRC(00i0, O)
-	//    t[3][i] = CRC(000i, O)
-	// where O is a sequence of K zeros.
-	var tmp [castagnoliK2]byte
-	for b := 0; b < 4; b++ {
-		for i := 0; i < 256; i++ {
-			val := uint32(i) << uint32(b*8)
-			castagnoliSSE42TableK1[b][i] = castagnoliSSE42(val, tmp[:castagnoliK1])
-			castagnoliSSE42TableK2[b][i] = castagnoliSSE42(val, tmp[:])
-		}
-	}
-	return false
-}
-
-// castagnoliShift computes the CRC32-C of K1 or K2 zeroes (depending on the
-// table given) with the given initial crc value. This corresponds to
-// CRC(crc, O) in the description in updateCastagnoli.
-func castagnoliShift(table *sse42Table, crc uint32) uint32 {
-	return table[3][crc>>24] ^
-		table[2][(crc>>16)&0xFF] ^
-		table[1][(crc>>8)&0xFF] ^
-		table[0][crc&0xFF]
-}
-
 func updateCastagnoli(crc uint32, p []byte) uint32 {
-	if !sse42 {
-		// Use slicing-by-8 on larger inputs.
-		if len(p) >= sliceBy8Cutoff {
-			return updateSlicingBy8(crc, castagnoliTable8, p)
-		}
-		return update(crc, castagnoliTable, p)
-	}
-
-	// This method is inspired from the algorithm in Intel's white paper:
-	//    "Fast CRC Computation for iSCSI Polynomial Using CRC32 Instruction"
-	// The same strategy of splitting the buffer in three is used but the
-	// combining calculation is different; the complete derivation is explained
-	// below.
-	//
-	// -- The basic idea --
-	//
-	// The CRC32 instruction (available in SSE4.2) can process 8 bytes at a
-	// time. In recent Intel architectures the instruction takes 3 cycles;
-	// however the processor can pipeline up to three instructions if they
-	// don't depend on each other.
-	//
-	// Roughly this means that we can process three buffers in about the same
-	// time we can process one buffer.
-	//
-	// The idea is then to split the buffer in three, CRC the three pieces
-	// separately and then combine the results.
-	//
-	// Combining the results requires precomputed tables, so we must choose a
-	// fixed buffer length to optimize. The longer the length, the faster; but
-	// only buffers longer than this length will use the optimization. We choose
-	// two cutoffs and compute tables for both:
-	//  - one around 512: 168*3=504
-	//  - one around 4KB: 1344*3=4032
-	//
-	// -- The nitty gritty --
-	//
-	// Let CRC(I, X) be the non-inverted CRC32-C of the sequence X (with
-	// initial non-inverted CRC I). This function has the following properties:
-	//   (a) CRC(I, AB) = CRC(CRC(I, A), B)
-	//   (b) CRC(I, A xor B) = CRC(I, A) xor CRC(0, B)
-	//
-	// Say we want to compute CRC(I, ABC) where A, B, C are three sequences of
-	// K bytes each, where K is a fixed constant. Let O be the sequence of K zero
-	// bytes.
-	//
-	// CRC(I, ABC) = CRC(I, ABO xor C)
-	//             = CRC(I, ABO) xor CRC(0, C)
-	//             = CRC(CRC(I, AB), O) xor CRC(0, C)
-	//             = CRC(CRC(I, AO xor B), O) xor CRC(0, C)
-	//             = CRC(CRC(I, AO) xor CRC(0, B), O) xor CRC(0, C)
-	//             = CRC(CRC(CRC(I, A), O) xor CRC(0, B), O) xor CRC(0, C)
-	//
-	// The castagnoliSSE42Triple function can compute CRC(I, A), CRC(0, B),
-	// and CRC(0, C) efficiently.  We just need to find a way to quickly compute
-	// CRC(uvwx, O) given a 4-byte initial value uvwx. We can precompute these
-	// values; since we can't have a 32-bit table, we break it up into four
-	// 8-bit tables:
-	//
-	//    CRC(uvwx, O) = CRC(u000, O) xor
-	//                   CRC(0v00, O) xor
-	//                   CRC(00w0, O) xor
-	//                   CRC(000x, O)
-	//
-	// We can compute tables corresponding to the four terms for all 8-bit
-	// values.
-
-	crc = ^crc
-
-	// If a buffer is long enough to use the optimization, process the first few
-	// bytes to align the buffer to an 8 byte boundary (if necessary).
-	if len(p) >= castagnoliK1*3 {
-		delta := int(uintptr(unsafe.Pointer(&p[0])) & 7)
-		if delta != 0 {
-			delta = 8 - delta
-			crc = castagnoliSSE42(crc, p[:delta])
-			p = p[delta:]
-		}
-	}
-
-	// Process 3*K2 at a time.
-	for len(p) >= castagnoliK2*3 {
-		// Compute CRC(I, A), CRC(0, B), and CRC(0, C).
-		crcA, crcB, crcC := castagnoliSSE42Triple(
-			crc, 0, 0,
-			p, p[castagnoliK2:], p[castagnoliK2*2:],
-			castagnoliK2/24)
-
-		// CRC(I, AB) = CRC(CRC(I, A), O) xor CRC(0, B)
-		crcAB := castagnoliShift(castagnoliSSE42TableK2, crcA) ^ crcB
-		// CRC(I, ABC) = CRC(CRC(I, AB), O) xor CRC(0, C)
-		crc = castagnoliShift(castagnoliSSE42TableK2, crcAB) ^ crcC
-		p = p[castagnoliK2*3:]
+	if sse42 {
+		return castagnoliSSE42(crc, p)
 	}
-
-	// Process 3*K1 at a time.
-	for len(p) >= castagnoliK1*3 {
-		// Compute CRC(I, A), CRC(0, B), and CRC(0, C).
-		crcA, crcB, crcC := castagnoliSSE42Triple(
-			crc, 0, 0,
-			p, p[castagnoliK1:], p[castagnoliK1*2:],
-			castagnoliK1/24)
-
-		// CRC(I, AB) = CRC(CRC(I, A), O) xor CRC(0, B)
-		crcAB := castagnoliShift(castagnoliSSE42TableK1, crcA) ^ crcB
-		// CRC(I, ABC) = CRC(CRC(I, AB), O) xor CRC(0, C)
-		crc = castagnoliShift(castagnoliSSE42TableK1, crcAB) ^ crcC
-		p = p[castagnoliK1*3:]
+	// Use slicing-by-8 on larger inputs.
+	if len(p) >= sliceBy8Cutoff {
+		return updateSlicingBy8(crc, castagnoliTable8, p)
 	}
-
-	// Use the simple implementation for what's left.
-	crc = castagnoliSSE42(crc, p)
-	return ^crc
+	return update(crc, castagnoliTable, p)
 }
 
 func updateIEEE(crc uint32, p []byte) uint32 {
diff --git a/src/hash/crc32/crc32_amd64.s b/src/hash/crc32/crc32_amd64.s
index 50c0ec83aa..a775a194df 100644
--- a/src/hash/crc32/crc32_amd64.s
+++ b/src/hash/crc32/crc32_amd64.s
@@ -4,14 +4,14 @@
 
 #include "textflag.h"
 
-// castagnoliSSE42 updates the (non-inverted) crc with the given buffer.
-//
 // func castagnoliSSE42(crc uint32, p []byte) uint32
 TEXT ·castagnoliSSE42(SB),NOSPLIT,$0
 	MOVL crc+0(FP), AX  // CRC value
 	MOVQ p+8(FP), SI  // data pointer
 	MOVQ p_len+16(FP), CX  // len(p)
 
+	NOTL AX
+
 	// If there are fewer than 8 bytes to process, skip alignment.
 	CMPQ CX, $8
 	JL less_than_8
@@ -87,53 +87,10 @@ less_than_2:
 	CRC32B (SI), AX
 
 done:
+	NOTL AX
 	MOVL AX, ret+32(FP)
 	RET
 
-// castagnoliSSE42Triple updates three (non-inverted) crcs with (24*rounds)
-// bytes from each buffer.
-//
-// func castagnoliSSE42Triple(
-//     crc1, crc2, crc3 uint32,
-//     a, b, c []byte,
-//     rounds uint32,
-// ) (retA uint32, retB uint32, retC uint32)
-TEXT ·castagnoliSSE42Triple(SB),NOSPLIT,$0
-	MOVL crcA+0(FP), AX
-	MOVL crcB+4(FP), CX
-	MOVL crcC+8(FP), DX
-
-	MOVQ a+16(FP), R8   // data pointer
-	MOVQ b+40(FP), R9   // data pointer
-	MOVQ c+64(FP), R10  // data pointer
-
-	MOVL rounds+88(FP), R11
-
-loop:
-	CRC32Q (R8), AX
-	CRC32Q (R9), CX
-	CRC32Q (R10), DX
-
-	CRC32Q 8(R8), AX
-	CRC32Q 8(R9), CX
-	CRC32Q 8(R10), DX
-
-	CRC32Q 16(R8), AX
-	CRC32Q 16(R9), CX
-	CRC32Q 16(R10), DX
-
-	ADDQ $24, R8
-	ADDQ $24, R9
-	ADDQ $24, R10
-
-	DECQ R11
-	JNZ loop
-
-	MOVL AX, retA+96(FP)
-	MOVL CX, retB+100(FP)
-	MOVL DX, retC+104(FP)
-	RET
-
 // func haveSSE42() bool
 TEXT ·haveSSE42(SB),NOSPLIT,$0
 	XORQ AX, AX
diff --git a/src/hash/crc32/crc32_amd64p32.go b/src/hash/crc32/crc32_amd64p32.go
index 48d181f295..1f6cd34643 100644
--- a/src/hash/crc32/crc32_amd64p32.go
+++ b/src/hash/crc32/crc32_amd64p32.go
@@ -7,22 +7,17 @@ package crc32
 // This file contains the code to call the SSE 4.2 version of the Castagnoli
 // CRC.
 
-// haveSSE42 is defined in crc32_amd64p32.s and uses CPUID to test for SSE 4.2
+// haveSSE42 is defined in crc_amd64p32.s and uses CPUID to test for SSE 4.2
 // support.
 func haveSSE42() bool
 
-// castagnoliSSE42 is defined in crc32_amd64.s and uses the SSE4.2 CRC32
+// castagnoliSSE42 is defined in crc_amd64.s and uses the SSE4.2 CRC32
 // instruction.
 //go:noescape
 func castagnoliSSE42(crc uint32, p []byte) uint32
 
 var sse42 = haveSSE42()
 
-func castagnoliInitArch() (needGenericTables bool) {
-	// We only need the generic implementation tables if we don't have SSE4.2.
-	return !sse42
-}
-
 func updateCastagnoli(crc uint32, p []byte) uint32 {
 	if sse42 {
 		return castagnoliSSE42(crc, p)
diff --git a/src/hash/crc32/crc32_generic.go b/src/hash/crc32/crc32_generic.go
index decf973066..10a6367bc9 100644
--- a/src/hash/crc32/crc32_generic.go
+++ b/src/hash/crc32/crc32_generic.go
@@ -9,10 +9,6 @@ package crc32
 // This file contains the generic version of updateCastagnoli which does
 // slicing-by-8, or uses the fallback for very small sizes.
 
-func castagnoliInitArch() (needGenericTables bool) {
-	return true
-}
-
 func updateCastagnoli(crc uint32, p []byte) uint32 {
 	// Use slicing-by-8 on larger inputs.
 	if len(p) >= sliceBy8Cutoff {
diff --git a/src/hash/crc32/crc32_s390x.go b/src/hash/crc32/crc32_s390x.go
index 72d2648280..befb58f55f 100644
--- a/src/hash/crc32/crc32_s390x.go
+++ b/src/hash/crc32/crc32_s390x.go
@@ -25,10 +25,6 @@ func vectorizedCastagnoli(crc uint32, p []byte) uint32
 //go:noescape
 func vectorizedIEEE(crc uint32, p []byte) uint32
 
-func castagnoliInitArch() (needGenericTables bool) {
-	return true
-}
-
 func genericCastagnoli(crc uint32, p []byte) uint32 {
 	// Use slicing-by-8 on larger inputs.
 	if len(p) >= sliceBy8Cutoff {
diff --git a/src/hash/crc32/crc32_test.go b/src/hash/crc32/crc32_test.go
index ccd6e59c50..067c42adf0 100644
--- a/src/hash/crc32/crc32_test.go
+++ b/src/hash/crc32/crc32_test.go
@@ -7,7 +7,6 @@ package crc32
 import (
 	"hash"
 	"io"
-	"math/rand"
 	"testing"
 )
 
@@ -86,41 +85,6 @@ func TestGolden(t *testing.T) {
 	}
 }
 
-func TestCastagnoliSSE42(t *testing.T) {
-	if !sse42 {
-		t.Skip("SSE42 not supported")
-	}
-
-	// Init the SSE42 tables.
-	MakeTable(Castagnoli)
-
-	// Manually init the software implementation to compare against.
-	castagnoliTable = makeTable(Castagnoli)
-	castagnoliTable8 = makeTable8(Castagnoli)
-
-	// The optimized SSE4.2 implementation behaves differently for different
-	// lengths (especially around multiples of K*3). Crosscheck against the
-	// software implementation for various lengths.
-	for _, base := range []int{castagnoliK1, castagnoliK2, castagnoliK1 + castagnoliK2} {
-		for _, baseMult := range []int{2, 3, 5, 6, 9, 30} {
-			for _, variation := range []int{0, 1, 2, 3, 4, 7, 10, 16, 32, 50, 128} {
-				for _, varMult := range []int{-2, -1, +1, +2} {
-					length := base*baseMult + variation*varMult
-					p := make([]byte, length)
-					_, _ = rand.Read(p)
-					crcInit := uint32(rand.Int63())
-					correct := updateSlicingBy8(crcInit, castagnoliTable8, p)
-					result := updateCastagnoli(crcInit, p)
-					if result != correct {
-						t.Errorf("SSE42 implementation = 0x%x want 0x%x (buffer length %d)",
-							result, correct, len(p))
-					}
-				}
-			}
-		}
-	}
-}
-
 func BenchmarkIEEECrc40B(b *testing.B) {
 	benchmark(b, NewIEEE(), 40, 0)
 }
@@ -149,42 +113,18 @@ func BenchmarkCastagnoliCrc40B(b *testing.B) {
 	benchmark(b, New(MakeTable(Castagnoli)), 40, 0)
 }
 
-func BenchmarkCastagnoliCrc40BMisaligned(b *testing.B) {
-	benchmark(b, New(MakeTable(Castagnoli)), 40, 1)
-}
-
-func BenchmarkCastagnoliCrc512(b *testing.B) {
-	benchmark(b, New(MakeTable(Castagnoli)), 512, 0)
-}
-
-func BenchmarkCastagnoliCrc512Misaligned(b *testing.B) {
-	benchmark(b, New(MakeTable(Castagnoli)), 512, 1)
-}
-
 func BenchmarkCastagnoliCrc1KB(b *testing.B) {
 	benchmark(b, New(MakeTable(Castagnoli)), 1<<10, 0)
 }
 
-func BenchmarkCastagnoliCrc1KBMisaligned(b *testing.B) {
-	benchmark(b, New(MakeTable(Castagnoli)), 1<<10, 1)
-}
-
 func BenchmarkCastagnoliCrc4KB(b *testing.B) {
 	benchmark(b, New(MakeTable(Castagnoli)), 4<<10, 0)
 }
 
-func BenchmarkCastagnoliCrc4KBMisaligned(b *testing.B) {
-	benchmark(b, New(MakeTable(Castagnoli)), 4<<10, 1)
-}
-
 func BenchmarkCastagnoliCrc32KB(b *testing.B) {
 	benchmark(b, New(MakeTable(Castagnoli)), 32<<10, 0)
 }
 
-func BenchmarkCastagnoliCrc32KBMisaligned(b *testing.B) {
-	benchmark(b, New(MakeTable(Castagnoli)), 32<<10, 1)
-}
-
 func benchmark(b *testing.B, h hash.Hash32, n, alignment int64) {
 	b.SetBytes(n)
 	data := make([]byte, n+alignment)