From afcd11d7816d4d3369c79bbcf8c462b614210d51 Mon Sep 17 00:00:00 2001 From: greatroar <61184462+greatroar@users.noreply.github.com> Date: Tue, 6 Dec 2022 21:04:46 +0100 Subject: [PATCH] zstd: Import xxhash v2.2.0 Unsafe parts omitted, noasm build tag added back in. writeBlocks argument renamed s to please go vet. --- zstd/internal/xxhash/README.md | 49 ++-- zstd/internal/xxhash/xxhash.go | 47 ++-- zstd/internal/xxhash/xxhash_amd64.s | 336 +++++++++++++-------------- zstd/internal/xxhash/xxhash_arm64.s | 138 ++++++----- zstd/internal/xxhash/xxhash_asm.go | 2 +- zstd/internal/xxhash/xxhash_other.go | 19 +- zstd/internal/xxhash/xxhash_test.go | 6 +- 7 files changed, 299 insertions(+), 298 deletions(-) diff --git a/zstd/internal/xxhash/README.md b/zstd/internal/xxhash/README.md index 69aa3bb587..777290d44c 100644 --- a/zstd/internal/xxhash/README.md +++ b/zstd/internal/xxhash/README.md @@ -2,12 +2,7 @@ VENDORED: Go to [github.com/cespare/xxhash](https://github.com/cespare/xxhash) for original package. - -[![GoDoc](https://godoc.org/github.com/cespare/xxhash?status.svg)](https://godoc.org/github.com/cespare/xxhash) -[![Build Status](https://travis-ci.org/cespare/xxhash.svg?branch=master)](https://travis-ci.org/cespare/xxhash) - -xxhash is a Go implementation of the 64-bit -[xxHash](http://cyan4973.github.io/xxHash/) algorithm, XXH64. This is a +xxhash is a Go implementation of the 64-bit [xxHash] algorithm, XXH64. This is a high-quality hashing algorithm that is much faster than anything in the Go standard library. @@ -28,31 +23,49 @@ func (*Digest) WriteString(string) (int, error) func (*Digest) Sum64() uint64 ``` -This implementation provides a fast pure-Go implementation and an even faster -assembly implementation for amd64. +The package is written with optimized pure Go and also contains even faster +assembly implementations for amd64 and arm64. If desired, the `purego` build tag +opts into using the Go code even on those architectures. + +[xxHash]: http://cyan4973.github.io/xxHash/ + +## Compatibility + +This package is in a module and the latest code is in version 2 of the module. +You need a version of Go with at least "minimal module compatibility" to use +github.com/cespare/xxhash/v2: + +* 1.9.7+ for Go 1.9 +* 1.10.3+ for Go 1.10 +* Go 1.11 or later + +I recommend using the latest release of Go. ## Benchmarks Here are some quick benchmarks comparing the pure-Go and assembly implementations of Sum64. -| input size | purego | asm | -| --- | --- | --- | -| 5 B | 979.66 MB/s | 1291.17 MB/s | -| 100 B | 7475.26 MB/s | 7973.40 MB/s | -| 4 KB | 17573.46 MB/s | 17602.65 MB/s | -| 10 MB | 17131.46 MB/s | 17142.16 MB/s | +| input size | purego | asm | +| ---------- | --------- | --------- | +| 4 B | 1.3 GB/s | 1.2 GB/s | +| 16 B | 2.9 GB/s | 3.5 GB/s | +| 100 B | 6.9 GB/s | 8.1 GB/s | +| 4 KB | 11.7 GB/s | 16.7 GB/s | +| 10 MB | 12.0 GB/s | 17.3 GB/s | -These numbers were generated on Ubuntu 18.04 with an Intel i7-8700K CPU using -the following commands under Go 1.11.2: +These numbers were generated on Ubuntu 20.04 with an Intel Xeon Platinum 8252C +CPU using the following commands under Go 1.19.2: ``` -$ go test -tags purego -benchtime 10s -bench '/xxhash,direct,bytes' -$ go test -benchtime 10s -bench '/xxhash,direct,bytes' +benchstat <(go test -tags purego -benchtime 500ms -count 15 -bench 'Sum64$') +benchstat <(go test -benchtime 500ms -count 15 -bench 'Sum64$') ``` ## Projects using this package - [InfluxDB](https://github.com/influxdata/influxdb) - [Prometheus](https://github.com/prometheus/prometheus) +- [VictoriaMetrics](https://github.com/VictoriaMetrics/VictoriaMetrics) - [FreeCache](https://github.com/coocood/freecache) +- [FastCache](https://github.com/VictoriaMetrics/fastcache) diff --git a/zstd/internal/xxhash/xxhash.go b/zstd/internal/xxhash/xxhash.go index 2c112a0ab1..fc40c82001 100644 --- a/zstd/internal/xxhash/xxhash.go +++ b/zstd/internal/xxhash/xxhash.go @@ -18,19 +18,11 @@ const ( prime5 uint64 = 2870177450012600261 ) -// NOTE(caleb): I'm using both consts and vars of the primes. Using consts where -// possible in the Go code is worth a small (but measurable) performance boost -// by avoiding some MOVQs. Vars are needed for the asm and also are useful for -// convenience in the Go code in a few places where we need to intentionally -// avoid constant arithmetic (e.g., v1 := prime1 + prime2 fails because the -// result overflows a uint64). -var ( - prime1v = prime1 - prime2v = prime2 - prime3v = prime3 - prime4v = prime4 - prime5v = prime5 -) +// Store the primes in an array as well. +// +// The consts are used when possible in Go code to avoid MOVs but we need a +// contiguous array of the assembly code. +var primes = [...]uint64{prime1, prime2, prime3, prime4, prime5} // Digest implements hash.Hash64. type Digest struct { @@ -52,10 +44,10 @@ func New() *Digest { // Reset clears the Digest's state so that it can be reused. func (d *Digest) Reset() { - d.v1 = prime1v + prime2 + d.v1 = primes[0] + prime2 d.v2 = prime2 d.v3 = 0 - d.v4 = -prime1v + d.v4 = -primes[0] d.total = 0 d.n = 0 } @@ -71,21 +63,23 @@ func (d *Digest) Write(b []byte) (n int, err error) { n = len(b) d.total += uint64(n) + memleft := d.mem[d.n&(len(d.mem)-1):] + if d.n+n < 32 { // This new data doesn't even fill the current block. - copy(d.mem[d.n:], b) + copy(memleft, b) d.n += n return } if d.n > 0 { // Finish off the partial block. - copy(d.mem[d.n:], b) + c := copy(memleft, b) d.v1 = round(d.v1, u64(d.mem[0:8])) d.v2 = round(d.v2, u64(d.mem[8:16])) d.v3 = round(d.v3, u64(d.mem[16:24])) d.v4 = round(d.v4, u64(d.mem[24:32])) - b = b[32-d.n:] + b = b[c:] d.n = 0 } @@ -135,21 +129,20 @@ func (d *Digest) Sum64() uint64 { h += d.total - i, end := 0, d.n - for ; i+8 <= end; i += 8 { - k1 := round(0, u64(d.mem[i:i+8])) + b := d.mem[:d.n&(len(d.mem)-1)] + for ; len(b) >= 8; b = b[8:] { + k1 := round(0, u64(b[:8])) h ^= k1 h = rol27(h)*prime1 + prime4 } - if i+4 <= end { - h ^= uint64(u32(d.mem[i:i+4])) * prime1 + if len(b) >= 4 { + h ^= uint64(u32(b[:4])) * prime1 h = rol23(h)*prime2 + prime3 - i += 4 + b = b[4:] } - for i < end { - h ^= uint64(d.mem[i]) * prime5 + for ; len(b) > 0; b = b[1:] { + h ^= uint64(b[0]) * prime5 h = rol11(h) * prime1 - i++ } h ^= h >> 33 diff --git a/zstd/internal/xxhash/xxhash_amd64.s b/zstd/internal/xxhash/xxhash_amd64.s index cea1785619..ddb63aa91b 100644 --- a/zstd/internal/xxhash/xxhash_amd64.s +++ b/zstd/internal/xxhash/xxhash_amd64.s @@ -1,3 +1,4 @@ +//go:build !appengine && gc && !purego && !noasm // +build !appengine // +build gc // +build !purego @@ -5,212 +6,205 @@ #include "textflag.h" -// Register allocation: -// AX h -// SI pointer to advance through b -// DX n -// BX loop end -// R8 v1, k1 -// R9 v2 -// R10 v3 -// R11 v4 -// R12 tmp -// R13 prime1v -// R14 prime2v -// DI prime4v - -// round reads from and advances the buffer pointer in SI. -// It assumes that R13 has prime1v and R14 has prime2v. -#define round(r) \ - MOVQ (SI), R12 \ - ADDQ $8, SI \ - IMULQ R14, R12 \ - ADDQ R12, r \ - ROLQ $31, r \ - IMULQ R13, r - -// mergeRound applies a merge round on the two registers acc and val. -// It assumes that R13 has prime1v, R14 has prime2v, and DI has prime4v. -#define mergeRound(acc, val) \ - IMULQ R14, val \ - ROLQ $31, val \ - IMULQ R13, val \ - XORQ val, acc \ - IMULQ R13, acc \ - ADDQ DI, acc +// Registers: +#define h AX +#define d AX +#define p SI // pointer to advance through b +#define n DX +#define end BX // loop end +#define v1 R8 +#define v2 R9 +#define v3 R10 +#define v4 R11 +#define x R12 +#define prime1 R13 +#define prime2 R14 +#define prime4 DI + +#define round(acc, x) \ + IMULQ prime2, x \ + ADDQ x, acc \ + ROLQ $31, acc \ + IMULQ prime1, acc + +// round0 performs the operation x = round(0, x). +#define round0(x) \ + IMULQ prime2, x \ + ROLQ $31, x \ + IMULQ prime1, x + +// mergeRound applies a merge round on the two registers acc and x. +// It assumes that prime1, prime2, and prime4 have been loaded. +#define mergeRound(acc, x) \ + round0(x) \ + XORQ x, acc \ + IMULQ prime1, acc \ + ADDQ prime4, acc + +// blockLoop processes as many 32-byte blocks as possible, +// updating v1, v2, v3, and v4. It assumes that there is at least one block +// to process. +#define blockLoop() \ +loop: \ + MOVQ +0(p), x \ + round(v1, x) \ + MOVQ +8(p), x \ + round(v2, x) \ + MOVQ +16(p), x \ + round(v3, x) \ + MOVQ +24(p), x \ + round(v4, x) \ + ADDQ $32, p \ + CMPQ p, end \ + JLE loop // func Sum64(b []byte) uint64 -TEXT ·Sum64(SB), NOSPLIT, $0-32 +TEXT ·Sum64(SB), NOSPLIT|NOFRAME, $0-32 // Load fixed primes. - MOVQ ·prime1v(SB), R13 - MOVQ ·prime2v(SB), R14 - MOVQ ·prime4v(SB), DI + MOVQ ·primes+0(SB), prime1 + MOVQ ·primes+8(SB), prime2 + MOVQ ·primes+24(SB), prime4 // Load slice. - MOVQ b_base+0(FP), SI - MOVQ b_len+8(FP), DX - LEAQ (SI)(DX*1), BX + MOVQ b_base+0(FP), p + MOVQ b_len+8(FP), n + LEAQ (p)(n*1), end // The first loop limit will be len(b)-32. - SUBQ $32, BX + SUBQ $32, end // Check whether we have at least one block. - CMPQ DX, $32 + CMPQ n, $32 JLT noBlocks // Set up initial state (v1, v2, v3, v4). - MOVQ R13, R8 - ADDQ R14, R8 - MOVQ R14, R9 - XORQ R10, R10 - XORQ R11, R11 - SUBQ R13, R11 - - // Loop until SI > BX. -blockLoop: - round(R8) - round(R9) - round(R10) - round(R11) - - CMPQ SI, BX - JLE blockLoop - - MOVQ R8, AX - ROLQ $1, AX - MOVQ R9, R12 - ROLQ $7, R12 - ADDQ R12, AX - MOVQ R10, R12 - ROLQ $12, R12 - ADDQ R12, AX - MOVQ R11, R12 - ROLQ $18, R12 - ADDQ R12, AX - - mergeRound(AX, R8) - mergeRound(AX, R9) - mergeRound(AX, R10) - mergeRound(AX, R11) + MOVQ prime1, v1 + ADDQ prime2, v1 + MOVQ prime2, v2 + XORQ v3, v3 + XORQ v4, v4 + SUBQ prime1, v4 + + blockLoop() + + MOVQ v1, h + ROLQ $1, h + MOVQ v2, x + ROLQ $7, x + ADDQ x, h + MOVQ v3, x + ROLQ $12, x + ADDQ x, h + MOVQ v4, x + ROLQ $18, x + ADDQ x, h + + mergeRound(h, v1) + mergeRound(h, v2) + mergeRound(h, v3) + mergeRound(h, v4) JMP afterBlocks noBlocks: - MOVQ ·prime5v(SB), AX + MOVQ ·primes+32(SB), h afterBlocks: - ADDQ DX, AX - - // Right now BX has len(b)-32, and we want to loop until SI > len(b)-8. - ADDQ $24, BX - - CMPQ SI, BX - JG fourByte - -wordLoop: - // Calculate k1. - MOVQ (SI), R8 - ADDQ $8, SI - IMULQ R14, R8 - ROLQ $31, R8 - IMULQ R13, R8 - - XORQ R8, AX - ROLQ $27, AX - IMULQ R13, AX - ADDQ DI, AX - - CMPQ SI, BX - JLE wordLoop - -fourByte: - ADDQ $4, BX - CMPQ SI, BX - JG singles - - MOVL (SI), R8 - ADDQ $4, SI - IMULQ R13, R8 - XORQ R8, AX - - ROLQ $23, AX - IMULQ R14, AX - ADDQ ·prime3v(SB), AX - -singles: - ADDQ $4, BX - CMPQ SI, BX + ADDQ n, h + + ADDQ $24, end + CMPQ p, end + JG try4 + +loop8: + MOVQ (p), x + ADDQ $8, p + round0(x) + XORQ x, h + ROLQ $27, h + IMULQ prime1, h + ADDQ prime4, h + + CMPQ p, end + JLE loop8 + +try4: + ADDQ $4, end + CMPQ p, end + JG try1 + + MOVL (p), x + ADDQ $4, p + IMULQ prime1, x + XORQ x, h + + ROLQ $23, h + IMULQ prime2, h + ADDQ ·primes+16(SB), h + +try1: + ADDQ $4, end + CMPQ p, end JGE finalize -singlesLoop: - MOVBQZX (SI), R12 - ADDQ $1, SI - IMULQ ·prime5v(SB), R12 - XORQ R12, AX +loop1: + MOVBQZX (p), x + ADDQ $1, p + IMULQ ·primes+32(SB), x + XORQ x, h + ROLQ $11, h + IMULQ prime1, h - ROLQ $11, AX - IMULQ R13, AX - - CMPQ SI, BX - JL singlesLoop + CMPQ p, end + JL loop1 finalize: - MOVQ AX, R12 - SHRQ $33, R12 - XORQ R12, AX - IMULQ R14, AX - MOVQ AX, R12 - SHRQ $29, R12 - XORQ R12, AX - IMULQ ·prime3v(SB), AX - MOVQ AX, R12 - SHRQ $32, R12 - XORQ R12, AX - - MOVQ AX, ret+24(FP) + MOVQ h, x + SHRQ $33, x + XORQ x, h + IMULQ prime2, h + MOVQ h, x + SHRQ $29, x + XORQ x, h + IMULQ ·primes+16(SB), h + MOVQ h, x + SHRQ $32, x + XORQ x, h + + MOVQ h, ret+24(FP) RET -// writeBlocks uses the same registers as above except that it uses AX to store -// the d pointer. - // func writeBlocks(d *Digest, b []byte) int -TEXT ·writeBlocks(SB), NOSPLIT, $0-40 +TEXT ·writeBlocks(SB), NOSPLIT|NOFRAME, $0-40 // Load fixed primes needed for round. - MOVQ ·prime1v(SB), R13 - MOVQ ·prime2v(SB), R14 + MOVQ ·primes+0(SB), prime1 + MOVQ ·primes+8(SB), prime2 // Load slice. - MOVQ b_base+8(FP), SI - MOVQ b_len+16(FP), DX - LEAQ (SI)(DX*1), BX - SUBQ $32, BX + MOVQ b_base+8(FP), p + MOVQ b_len+16(FP), n + LEAQ (p)(n*1), end + SUBQ $32, end // Load vN from d. - MOVQ d+0(FP), AX - MOVQ 0(AX), R8 // v1 - MOVQ 8(AX), R9 // v2 - MOVQ 16(AX), R10 // v3 - MOVQ 24(AX), R11 // v4 + MOVQ s+0(FP), d + MOVQ 0(d), v1 + MOVQ 8(d), v2 + MOVQ 16(d), v3 + MOVQ 24(d), v4 // We don't need to check the loop condition here; this function is // always called with at least one block of data to process. -blockLoop: - round(R8) - round(R9) - round(R10) - round(R11) - - CMPQ SI, BX - JLE blockLoop + blockLoop() // Copy vN back to d. - MOVQ R8, 0(AX) - MOVQ R9, 8(AX) - MOVQ R10, 16(AX) - MOVQ R11, 24(AX) - - // The number of bytes written is SI minus the old base pointer. - SUBQ b_base+8(FP), SI - MOVQ SI, ret+32(FP) + MOVQ v1, 0(d) + MOVQ v2, 8(d) + MOVQ v3, 16(d) + MOVQ v4, 24(d) + + // The number of bytes written is p minus the old base pointer. + SUBQ b_base+8(FP), p + MOVQ p, ret+32(FP) RET diff --git a/zstd/internal/xxhash/xxhash_arm64.s b/zstd/internal/xxhash/xxhash_arm64.s index 36f2effee0..17901e0804 100644 --- a/zstd/internal/xxhash/xxhash_arm64.s +++ b/zstd/internal/xxhash/xxhash_arm64.s @@ -1,13 +1,17 @@ -// +build gc,!purego,!noasm +//go:build !appengine && gc && !purego && !noasm +// +build !appengine +// +build gc +// +build !purego +// +build !noasm #include "textflag.h" -// Register allocation. +// Registers: #define digest R1 -#define h R2 // Return value. -#define p R3 // Input pointer. -#define len R4 -#define nblocks R5 // len / 32. +#define h R2 // return value +#define p R3 // input pointer +#define n R4 // input length +#define nblocks R5 // n / 32 #define prime1 R7 #define prime2 R8 #define prime3 R9 @@ -25,52 +29,44 @@ #define round(acc, x) \ MADD prime2, acc, x, acc \ ROR $64-31, acc \ - MUL prime1, acc \ + MUL prime1, acc -// x = round(0, x). +// round0 performs the operation x = round(0, x). #define round0(x) \ MUL prime2, x \ ROR $64-31, x \ - MUL prime1, x \ - -#define mergeRound(x) \ - round0(x) \ - EOR x, h \ - MADD h, prime4, prime1, h \ - -// Update v[1-4] with 32-byte blocks. Assumes len >= 32. -#define blocksLoop() \ - LSR $5, len, nblocks \ - PCALIGN $16 \ - loop: \ - LDP.P 32(p), (x1, x2) \ - round(v1, x1) \ - LDP -16(p), (x3, x4) \ - round(v2, x2) \ - round(v3, x3) \ - round(v4, x4) \ - SUB $1, nblocks \ - CBNZ nblocks, loop \ - -// The primes are repeated here to ensure that they're stored -// in a contiguous array, so we can load them with LDP. -DATA primes<> +0(SB)/8, $11400714785074694791 -DATA primes<> +8(SB)/8, $14029467366897019727 -DATA primes<>+16(SB)/8, $1609587929392839161 -DATA primes<>+24(SB)/8, $9650029242287828579 -DATA primes<>+32(SB)/8, $2870177450012600261 -GLOBL primes<>(SB), NOPTR+RODATA, $40 + MUL prime1, x + +#define mergeRound(acc, x) \ + round0(x) \ + EOR x, acc \ + MADD acc, prime4, prime1, acc + +// blockLoop processes as many 32-byte blocks as possible, +// updating v1, v2, v3, and v4. It assumes that n >= 32. +#define blockLoop() \ + LSR $5, n, nblocks \ + PCALIGN $16 \ + loop: \ + LDP.P 16(p), (x1, x2) \ + LDP.P 16(p), (x3, x4) \ + round(v1, x1) \ + round(v2, x2) \ + round(v3, x3) \ + round(v4, x4) \ + SUB $1, nblocks \ + CBNZ nblocks, loop // func Sum64(b []byte) uint64 -TEXT ·Sum64(SB), NOFRAME+NOSPLIT, $0-32 - LDP b_base+0(FP), (p, len) +TEXT ·Sum64(SB), NOSPLIT|NOFRAME, $0-32 + LDP b_base+0(FP), (p, n) - LDP primes<> +0(SB), (prime1, prime2) - LDP primes<>+16(SB), (prime3, prime4) - MOVD primes<>+32(SB), prime5 + LDP ·primes+0(SB), (prime1, prime2) + LDP ·primes+16(SB), (prime3, prime4) + MOVD ·primes+32(SB), prime5 - CMP $32, len - CSEL LT, prime5, ZR, h // if len < 32 { h = prime5 } else { h = 0 } + CMP $32, n + CSEL LT, prime5, ZR, h // if n < 32 { h = prime5 } else { h = 0 } BLT afterLoop ADD prime1, prime2, v1 @@ -78,7 +74,7 @@ TEXT ·Sum64(SB), NOFRAME+NOSPLIT, $0-32 MOVD $0, v3 NEG prime1, v4 - blocksLoop() + blockLoop() ROR $64-1, v1, x1 ROR $64-7, v2, x2 @@ -88,71 +84,75 @@ TEXT ·Sum64(SB), NOFRAME+NOSPLIT, $0-32 ADD x3, x4 ADD x2, x4, h - mergeRound(v1) - mergeRound(v2) - mergeRound(v3) - mergeRound(v4) + mergeRound(h, v1) + mergeRound(h, v2) + mergeRound(h, v3) + mergeRound(h, v4) afterLoop: - ADD len, h + ADD n, h - TBZ $4, len, try8 + TBZ $4, n, try8 LDP.P 16(p), (x1, x2) round0(x1) + + // NOTE: here and below, sequencing the EOR after the ROR (using a + // rotated register) is worth a small but measurable speedup for small + // inputs. ROR $64-27, h EOR x1 @> 64-27, h, h MADD h, prime4, prime1, h round0(x2) ROR $64-27, h - EOR x2 @> 64-27, h + EOR x2 @> 64-27, h, h MADD h, prime4, prime1, h try8: - TBZ $3, len, try4 + TBZ $3, n, try4 MOVD.P 8(p), x1 round0(x1) ROR $64-27, h - EOR x1 @> 64-27, h + EOR x1 @> 64-27, h, h MADD h, prime4, prime1, h try4: - TBZ $2, len, try2 + TBZ $2, n, try2 MOVWU.P 4(p), x2 MUL prime1, x2 ROR $64-23, h - EOR x2 @> 64-23, h + EOR x2 @> 64-23, h, h MADD h, prime3, prime2, h try2: - TBZ $1, len, try1 + TBZ $1, n, try1 MOVHU.P 2(p), x3 AND $255, x3, x1 LSR $8, x3, x2 MUL prime5, x1 ROR $64-11, h - EOR x1 @> 64-11, h + EOR x1 @> 64-11, h, h MUL prime1, h MUL prime5, x2 ROR $64-11, h - EOR x2 @> 64-11, h + EOR x2 @> 64-11, h, h MUL prime1, h try1: - TBZ $0, len, end + TBZ $0, n, finalize MOVBU (p), x4 MUL prime5, x4 ROR $64-11, h - EOR x4 @> 64-11, h + EOR x4 @> 64-11, h, h MUL prime1, h -end: +finalize: EOR h >> 33, h MUL prime2, h EOR h >> 29, h @@ -163,24 +163,22 @@ end: RET // func writeBlocks(d *Digest, b []byte) int -// -// Assumes len(b) >= 32. -TEXT ·writeBlocks(SB), NOFRAME+NOSPLIT, $0-40 - LDP primes<>(SB), (prime1, prime2) +TEXT ·writeBlocks(SB), NOSPLIT|NOFRAME, $0-40 + LDP ·primes+0(SB), (prime1, prime2) // Load state. Assume v[1-4] are stored contiguously. MOVD d+0(FP), digest LDP 0(digest), (v1, v2) LDP 16(digest), (v3, v4) - LDP b_base+8(FP), (p, len) + LDP b_base+8(FP), (p, n) - blocksLoop() + blockLoop() // Store updated state. STP (v1, v2), 0(digest) STP (v3, v4), 16(digest) - BIC $31, len - MOVD len, ret+32(FP) + BIC $31, n + MOVD n, ret+32(FP) RET diff --git a/zstd/internal/xxhash/xxhash_asm.go b/zstd/internal/xxhash/xxhash_asm.go index 1a1fac9c26..d4221edf4f 100644 --- a/zstd/internal/xxhash/xxhash_asm.go +++ b/zstd/internal/xxhash/xxhash_asm.go @@ -13,4 +13,4 @@ package xxhash func Sum64(b []byte) uint64 //go:noescape -func writeBlocks(d *Digest, b []byte) int +func writeBlocks(s *Digest, b []byte) int diff --git a/zstd/internal/xxhash/xxhash_other.go b/zstd/internal/xxhash/xxhash_other.go index 209cb4a999..0be16cefc7 100644 --- a/zstd/internal/xxhash/xxhash_other.go +++ b/zstd/internal/xxhash/xxhash_other.go @@ -15,10 +15,10 @@ func Sum64(b []byte) uint64 { var h uint64 if n >= 32 { - v1 := prime1v + prime2 + v1 := primes[0] + prime2 v2 := prime2 v3 := uint64(0) - v4 := -prime1v + v4 := -primes[0] for len(b) >= 32 { v1 = round(v1, u64(b[0:8:len(b)])) v2 = round(v2, u64(b[8:16:len(b)])) @@ -37,19 +37,18 @@ func Sum64(b []byte) uint64 { h += uint64(n) - i, end := 0, len(b) - for ; i+8 <= end; i += 8 { - k1 := round(0, u64(b[i:i+8:len(b)])) + for ; len(b) >= 8; b = b[8:] { + k1 := round(0, u64(b[:8])) h ^= k1 h = rol27(h)*prime1 + prime4 } - if i+4 <= end { - h ^= uint64(u32(b[i:i+4:len(b)])) * prime1 + if len(b) >= 4 { + h ^= uint64(u32(b[:4])) * prime1 h = rol23(h)*prime2 + prime3 - i += 4 + b = b[4:] } - for ; i < end; i++ { - h ^= uint64(b[i]) * prime5 + for ; len(b) > 0; b = b[1:] { + h ^= uint64(b[0]) * prime5 h = rol11(h) * prime1 } diff --git a/zstd/internal/xxhash/xxhash_test.go b/zstd/internal/xxhash/xxhash_test.go index 414e34ab55..6330f19123 100644 --- a/zstd/internal/xxhash/xxhash_test.go +++ b/zstd/internal/xxhash/xxhash_test.go @@ -26,7 +26,11 @@ func TestAll(t *testing.T) { 0x02a2e85470d6fd96, }, } { - for chunkSize := 1; chunkSize <= len(tt.input); chunkSize++ { + lastChunkSize := len(tt.input) + if lastChunkSize == 0 { + lastChunkSize = 1 + } + for chunkSize := 1; chunkSize <= lastChunkSize; chunkSize++ { name := fmt.Sprintf("%s,chunkSize=%d", tt.name, chunkSize) t.Run(name, func(t *testing.T) { testDigest(t, tt.input, chunkSize, tt.want)