From 275e1fc8a84f00534f4adcc06bc1bb0cc3500b62 Mon Sep 17 00:00:00 2001 From: Klaus Post Date: Tue, 8 Mar 2022 06:14:07 -0800 Subject: [PATCH] huff0: Add amd64 assembly for low tablelogs (#518) Solid improvement on large payload, but regression on small, so we fall back to Go for those. ``` benchmark old MB/s new MB/s speedup BenchmarkDecompress4XNoTable/digits/100-32 296.37 296.75 1.00x BenchmarkDecompress4XNoTable/digits/10000-32 693.64 919.28 1.33x BenchmarkDecompress4XNoTable/digits/262143-32 631.10 876.14 1.39x BenchmarkDecompress4XNoTable/gettysburg/100-32 343.16 342.84 1.00x BenchmarkDecompress4XNoTable/twain/100-32 297.23 296.70 1.00x BenchmarkDecompress4XNoTable/low-ent.10k/100-32 267.55 267.49 1.00x BenchmarkDecompress4XNoTable/low-ent.10k/10000-32 752.34 916.33 1.22x BenchmarkDecompress4XNoTable/low-ent.10k/262143-32 778.37 1021.03 1.31x BenchmarkDecompress4XNoTable/superlow-ent-10k/262143-32 777.72 1024.48 1.32x BenchmarkDecompress4XNoTable/case1/100-32 307.09 302.20 0.98x BenchmarkDecompress4XNoTable/case1/10000-32 693.03 912.43 1.32x BenchmarkDecompress4XNoTable/case1/262143-32 696.06 948.19 1.36x BenchmarkDecompress4XNoTable/case2/100-32 292.50 287.26 0.98x BenchmarkDecompress4XNoTable/case2/10000-32 713.88 920.04 1.29x BenchmarkDecompress4XNoTable/case2/262143-32 724.83 979.52 1.35x BenchmarkDecompress4XNoTable/case3/100-32 298.72 296.64 0.99x BenchmarkDecompress4XNoTable/case3/10000-32 704.81 928.19 1.32x BenchmarkDecompress4XNoTable/case3/262143-32 708.32 973.26 1.37x BenchmarkDecompress4XNoTable/pngdata.001/100-32 285.21 272.15 0.95x BenchmarkDecompress4XNoTable/normcount2/100-32 335.74 335.80 1.00x BenchmarkDecompress4XNoTable/normcount2/10000-32 677.11 911.46 1.35x BenchmarkDecompress4XNoTable/normcount2/262143-32 682.78 939.67 1.38x BenchmarkDecompress4XNoTableTableLog8/digits-32 678.59 931.37 1.37x ``` --- huff0/autogen.go | 1 + huff0/decompress_8b_amd64.s | 488 +++++++++++++++++++++++++++++++++ huff0/decompress_8b_amd64.s.in | 197 +++++++++++++ huff0/decompress_amd64.go | 23 +- huff0/generate.go | 4 + 5 files changed, 709 insertions(+), 4 deletions(-) create mode 100644 huff0/decompress_8b_amd64.s create mode 100644 huff0/decompress_8b_amd64.s.in diff --git a/huff0/autogen.go b/huff0/autogen.go index 110883f13f..ff2c69d60c 100644 --- a/huff0/autogen.go +++ b/huff0/autogen.go @@ -2,3 +2,4 @@ package huff0 //go:generate go run generate.go //go:generate asmfmt -w decompress_amd64.s +//go:generate asmfmt -w decompress_8b_amd64.s diff --git a/huff0/decompress_8b_amd64.s b/huff0/decompress_8b_amd64.s new file mode 100644 index 0000000000..23e9acabfe --- /dev/null +++ b/huff0/decompress_8b_amd64.s @@ -0,0 +1,488 @@ +// +build !appengine +// +build gc +// +build !noasm + +#include "textflag.h" +#include "funcdata.h" +#include "go_asm.h" + +#define bufoff 256 // see decompress.go, we're using [4][256]byte table + +// func decompress4x_main_loop_x86(pbr0, pbr1, pbr2, pbr3 *bitReaderShifted, +// peekBits uint8, buf *byte, tbl *dEntrySingle) (int, bool) +TEXT ·decompress4x_8b_loop_x86(SB), NOSPLIT, $8 +#define off R8 +#define buffer DI +#define table SI + +#define br_bits_read R9 +#define br_value R10 +#define br_offset R11 +#define peek_bits R12 +#define exhausted DX + +#define br0 R13 +#define br1 R14 +#define br2 R15 +#define br3 BP + + MOVQ BP, 0(SP) + + XORQ exhausted, exhausted // exhausted = false + XORQ off, off // off = 0 + + MOVBQZX peekBits+32(FP), peek_bits + MOVQ buf+40(FP), buffer + MOVQ tbl+48(FP), table + + MOVQ pbr0+0(FP), br0 + MOVQ pbr1+8(FP), br1 + MOVQ pbr2+16(FP), br2 + MOVQ pbr3+24(FP), br3 + +main_loop: + + // const stream = 0 + // br0.fillFast() + MOVBQZX bitReaderShifted_bitsRead(br0), br_bits_read + MOVQ bitReaderShifted_value(br0), br_value + MOVQ bitReaderShifted_off(br0), br_offset + + // if b.bitsRead >= 32 { + CMPQ br_bits_read, $32 + JB skip_fill0 + + SUBQ $32, br_bits_read // b.bitsRead -= 32 + SUBQ $4, br_offset // b.off -= 4 + + // v := b.in[b.off-4 : b.off] + // v = v[:4] + // low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24) + MOVQ bitReaderShifted_in(br0), AX + MOVL 0(br_offset)(AX*1), AX // AX = uint32(b.in[b.off:b.off+4]) + + // b.value |= uint64(low) << (b.bitsRead & 63) + MOVQ br_bits_read, CX + SHLQ CL, AX + ORQ AX, br_value + + // exhausted = exhausted || (br0.off < 4) + CMPQ br_offset, $4 + SETLT DL + ORB DL, DH + + // } +skip_fill0: + + // val0 := br0.peekTopBits(peekBits) + MOVQ br_value, AX + MOVQ peek_bits, CX + SHRQ CL, AX // AX = (value >> peek_bits) & mask + + // v0 := table[val0&mask] + MOVW 0(table)(AX*2), AX // AX - v0 + + // br0.advance(uint8(v0.entry)) + MOVB AH, BL // BL = uint8(v0.entry >> 8) + MOVBQZX AL, CX + SHLQ CL, br_value // value <<= n + ADDQ CX, br_bits_read // bits_read += n + + // val1 := br0.peekTopBits(peekBits) + MOVQ peek_bits, CX + MOVQ br_value, AX + SHRQ CL, AX // AX = (value >> peek_bits) & mask + + // v1 := table[val1&mask] + MOVW 0(table)(AX*2), AX // AX - v1 + + // br0.advance(uint8(v1.entry)) + MOVB AH, BH // BH = uint8(v1.entry >> 8) + MOVBQZX AL, CX + SHLQ CX, br_value // value <<= n + ADDQ CX, br_bits_read // bits_read += n + + // these two writes get coalesced + // buf[stream][off] = uint8(v0.entry >> 8) + // buf[stream][off+1] = uint8(v1.entry >> 8) + MOVW BX, 0(buffer)(off*1) + + // SECOND PART: + // val2 := br0.peekTopBits(peekBits) + MOVQ br_value, AX + MOVQ peek_bits, CX + SHRQ CL, AX // AX = (value >> peek_bits) & mask + + // v2 := table[val0&mask] + MOVW 0(table)(AX*2), AX // AX - v0 + + // br0.advance(uint8(v0.entry)) + MOVB AH, BL // BL = uint8(v0.entry >> 8) + MOVBQZX AL, CX + SHLQ CL, br_value // value <<= n + ADDQ CX, br_bits_read // bits_read += n + + // val3 := br0.peekTopBits(peekBits) + MOVQ peek_bits, CX + MOVQ br_value, AX + SHRQ CL, AX // AX = (value >> peek_bits) & mask + + // v3 := table[val1&mask] + MOVW 0(table)(AX*2), AX // AX - v1 + + // br0.advance(uint8(v1.entry)) + MOVB AH, BH // BH = uint8(v1.entry >> 8) + MOVBQZX AL, CX + SHLQ CX, br_value // value <<= n + ADDQ CX, br_bits_read // bits_read += n + + // these two writes get coalesced + // buf[stream][off] = uint8(v0.entry >> 8) + // buf[stream][off+1] = uint8(v1.entry >> 8) + MOVW BX, 0+2(buffer)(off*1) + + // update the bitrader reader structure + MOVB br_bits_read, bitReaderShifted_bitsRead(br0) + MOVQ br_value, bitReaderShifted_value(br0) + MOVQ br_offset, bitReaderShifted_off(br0) + + // const stream = 1 + // br1.fillFast() + MOVBQZX bitReaderShifted_bitsRead(br1), br_bits_read + MOVQ bitReaderShifted_value(br1), br_value + MOVQ bitReaderShifted_off(br1), br_offset + + // if b.bitsRead >= 32 { + CMPQ br_bits_read, $32 + JB skip_fill1 + + SUBQ $32, br_bits_read // b.bitsRead -= 32 + SUBQ $4, br_offset // b.off -= 4 + + // v := b.in[b.off-4 : b.off] + // v = v[:4] + // low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24) + MOVQ bitReaderShifted_in(br1), AX + MOVL 0(br_offset)(AX*1), AX // AX = uint32(b.in[b.off:b.off+4]) + + // b.value |= uint64(low) << (b.bitsRead & 63) + MOVQ br_bits_read, CX + SHLQ CL, AX + ORQ AX, br_value + + // exhausted = exhausted || (br1.off < 4) + CMPQ br_offset, $4 + SETLT DL + ORB DL, DH + + // } +skip_fill1: + + // val0 := br1.peekTopBits(peekBits) + MOVQ br_value, AX + MOVQ peek_bits, CX + SHRQ CL, AX // AX = (value >> peek_bits) & mask + + // v0 := table[val0&mask] + MOVW 0(table)(AX*2), AX // AX - v0 + + // br1.advance(uint8(v0.entry)) + MOVB AH, BL // BL = uint8(v0.entry >> 8) + MOVBQZX AL, CX + SHLQ CL, br_value // value <<= n + ADDQ CX, br_bits_read // bits_read += n + + // val1 := br1.peekTopBits(peekBits) + MOVQ peek_bits, CX + MOVQ br_value, AX + SHRQ CL, AX // AX = (value >> peek_bits) & mask + + // v1 := table[val1&mask] + MOVW 0(table)(AX*2), AX // AX - v1 + + // br1.advance(uint8(v1.entry)) + MOVB AH, BH // BH = uint8(v1.entry >> 8) + MOVBQZX AL, CX + SHLQ CX, br_value // value <<= n + ADDQ CX, br_bits_read // bits_read += n + + // these two writes get coalesced + // buf[stream][off] = uint8(v0.entry >> 8) + // buf[stream][off+1] = uint8(v1.entry >> 8) + MOVW BX, 256(buffer)(off*1) + + // SECOND PART: + // val2 := br1.peekTopBits(peekBits) + MOVQ br_value, AX + MOVQ peek_bits, CX + SHRQ CL, AX // AX = (value >> peek_bits) & mask + + // v2 := table[val0&mask] + MOVW 0(table)(AX*2), AX // AX - v0 + + // br1.advance(uint8(v0.entry)) + MOVB AH, BL // BL = uint8(v0.entry >> 8) + MOVBQZX AL, CX + SHLQ CL, br_value // value <<= n + ADDQ CX, br_bits_read // bits_read += n + + // val3 := br1.peekTopBits(peekBits) + MOVQ peek_bits, CX + MOVQ br_value, AX + SHRQ CL, AX // AX = (value >> peek_bits) & mask + + // v3 := table[val1&mask] + MOVW 0(table)(AX*2), AX // AX - v1 + + // br1.advance(uint8(v1.entry)) + MOVB AH, BH // BH = uint8(v1.entry >> 8) + MOVBQZX AL, CX + SHLQ CX, br_value // value <<= n + ADDQ CX, br_bits_read // bits_read += n + + // these two writes get coalesced + // buf[stream][off] = uint8(v0.entry >> 8) + // buf[stream][off+1] = uint8(v1.entry >> 8) + MOVW BX, 256+2(buffer)(off*1) + + // update the bitrader reader structure + MOVB br_bits_read, bitReaderShifted_bitsRead(br1) + MOVQ br_value, bitReaderShifted_value(br1) + MOVQ br_offset, bitReaderShifted_off(br1) + + // const stream = 2 + // br2.fillFast() + MOVBQZX bitReaderShifted_bitsRead(br2), br_bits_read + MOVQ bitReaderShifted_value(br2), br_value + MOVQ bitReaderShifted_off(br2), br_offset + + // if b.bitsRead >= 32 { + CMPQ br_bits_read, $32 + JB skip_fill2 + + SUBQ $32, br_bits_read // b.bitsRead -= 32 + SUBQ $4, br_offset // b.off -= 4 + + // v := b.in[b.off-4 : b.off] + // v = v[:4] + // low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24) + MOVQ bitReaderShifted_in(br2), AX + MOVL 0(br_offset)(AX*1), AX // AX = uint32(b.in[b.off:b.off+4]) + + // b.value |= uint64(low) << (b.bitsRead & 63) + MOVQ br_bits_read, CX + SHLQ CL, AX + ORQ AX, br_value + + // exhausted = exhausted || (br2.off < 4) + CMPQ br_offset, $4 + SETLT DL + ORB DL, DH + + // } +skip_fill2: + + // val0 := br2.peekTopBits(peekBits) + MOVQ br_value, AX + MOVQ peek_bits, CX + SHRQ CL, AX // AX = (value >> peek_bits) & mask + + // v0 := table[val0&mask] + MOVW 0(table)(AX*2), AX // AX - v0 + + // br2.advance(uint8(v0.entry)) + MOVB AH, BL // BL = uint8(v0.entry >> 8) + MOVBQZX AL, CX + SHLQ CL, br_value // value <<= n + ADDQ CX, br_bits_read // bits_read += n + + // val1 := br2.peekTopBits(peekBits) + MOVQ peek_bits, CX + MOVQ br_value, AX + SHRQ CL, AX // AX = (value >> peek_bits) & mask + + // v1 := table[val1&mask] + MOVW 0(table)(AX*2), AX // AX - v1 + + // br2.advance(uint8(v1.entry)) + MOVB AH, BH // BH = uint8(v1.entry >> 8) + MOVBQZX AL, CX + SHLQ CX, br_value // value <<= n + ADDQ CX, br_bits_read // bits_read += n + + // these two writes get coalesced + // buf[stream][off] = uint8(v0.entry >> 8) + // buf[stream][off+1] = uint8(v1.entry >> 8) + MOVW BX, 512(buffer)(off*1) + + // SECOND PART: + // val2 := br2.peekTopBits(peekBits) + MOVQ br_value, AX + MOVQ peek_bits, CX + SHRQ CL, AX // AX = (value >> peek_bits) & mask + + // v2 := table[val0&mask] + MOVW 0(table)(AX*2), AX // AX - v0 + + // br2.advance(uint8(v0.entry)) + MOVB AH, BL // BL = uint8(v0.entry >> 8) + MOVBQZX AL, CX + SHLQ CL, br_value // value <<= n + ADDQ CX, br_bits_read // bits_read += n + + // val3 := br2.peekTopBits(peekBits) + MOVQ peek_bits, CX + MOVQ br_value, AX + SHRQ CL, AX // AX = (value >> peek_bits) & mask + + // v3 := table[val1&mask] + MOVW 0(table)(AX*2), AX // AX - v1 + + // br2.advance(uint8(v1.entry)) + MOVB AH, BH // BH = uint8(v1.entry >> 8) + MOVBQZX AL, CX + SHLQ CX, br_value // value <<= n + ADDQ CX, br_bits_read // bits_read += n + + // these two writes get coalesced + // buf[stream][off] = uint8(v0.entry >> 8) + // buf[stream][off+1] = uint8(v1.entry >> 8) + MOVW BX, 512+2(buffer)(off*1) + + // update the bitrader reader structure + MOVB br_bits_read, bitReaderShifted_bitsRead(br2) + MOVQ br_value, bitReaderShifted_value(br2) + MOVQ br_offset, bitReaderShifted_off(br2) + + // const stream = 3 + // br3.fillFast() + MOVBQZX bitReaderShifted_bitsRead(br3), br_bits_read + MOVQ bitReaderShifted_value(br3), br_value + MOVQ bitReaderShifted_off(br3), br_offset + + // if b.bitsRead >= 32 { + CMPQ br_bits_read, $32 + JB skip_fill3 + + SUBQ $32, br_bits_read // b.bitsRead -= 32 + SUBQ $4, br_offset // b.off -= 4 + + // v := b.in[b.off-4 : b.off] + // v = v[:4] + // low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24) + MOVQ bitReaderShifted_in(br3), AX + MOVL 0(br_offset)(AX*1), AX // AX = uint32(b.in[b.off:b.off+4]) + + // b.value |= uint64(low) << (b.bitsRead & 63) + MOVQ br_bits_read, CX + SHLQ CL, AX + ORQ AX, br_value + + // exhausted = exhausted || (br3.off < 4) + CMPQ br_offset, $4 + SETLT DL + ORB DL, DH + + // } +skip_fill3: + + // val0 := br3.peekTopBits(peekBits) + MOVQ br_value, AX + MOVQ peek_bits, CX + SHRQ CL, AX // AX = (value >> peek_bits) & mask + + // v0 := table[val0&mask] + MOVW 0(table)(AX*2), AX // AX - v0 + + // br3.advance(uint8(v0.entry)) + MOVB AH, BL // BL = uint8(v0.entry >> 8) + MOVBQZX AL, CX + SHLQ CL, br_value // value <<= n + ADDQ CX, br_bits_read // bits_read += n + + // val1 := br3.peekTopBits(peekBits) + MOVQ peek_bits, CX + MOVQ br_value, AX + SHRQ CL, AX // AX = (value >> peek_bits) & mask + + // v1 := table[val1&mask] + MOVW 0(table)(AX*2), AX // AX - v1 + + // br3.advance(uint8(v1.entry)) + MOVB AH, BH // BH = uint8(v1.entry >> 8) + MOVBQZX AL, CX + SHLQ CX, br_value // value <<= n + ADDQ CX, br_bits_read // bits_read += n + + // these two writes get coalesced + // buf[stream][off] = uint8(v0.entry >> 8) + // buf[stream][off+1] = uint8(v1.entry >> 8) + MOVW BX, 768(buffer)(off*1) + + // SECOND PART: + // val2 := br3.peekTopBits(peekBits) + MOVQ br_value, AX + MOVQ peek_bits, CX + SHRQ CL, AX // AX = (value >> peek_bits) & mask + + // v2 := table[val0&mask] + MOVW 0(table)(AX*2), AX // AX - v0 + + // br3.advance(uint8(v0.entry)) + MOVB AH, BL // BL = uint8(v0.entry >> 8) + MOVBQZX AL, CX + SHLQ CL, br_value // value <<= n + ADDQ CX, br_bits_read // bits_read += n + + // val3 := br3.peekTopBits(peekBits) + MOVQ peek_bits, CX + MOVQ br_value, AX + SHRQ CL, AX // AX = (value >> peek_bits) & mask + + // v3 := table[val1&mask] + MOVW 0(table)(AX*2), AX // AX - v1 + + // br3.advance(uint8(v1.entry)) + MOVB AH, BH // BH = uint8(v1.entry >> 8) + MOVBQZX AL, CX + SHLQ CX, br_value // value <<= n + ADDQ CX, br_bits_read // bits_read += n + + // these two writes get coalesced + // buf[stream][off] = uint8(v0.entry >> 8) + // buf[stream][off+1] = uint8(v1.entry >> 8) + MOVW BX, 768+2(buffer)(off*1) + + // update the bitrader reader structure + MOVB br_bits_read, bitReaderShifted_bitsRead(br3) + MOVQ br_value, bitReaderShifted_value(br3) + MOVQ br_offset, bitReaderShifted_off(br3) + + ADDQ $4, off // off += 2 + + TESTB DH, DH // any br[i].ofs < 4? + JNZ end + + CMPQ off, $bufoff + JL main_loop + +end: + MOVQ 0(SP), BP + + MOVB off, ret+56(FP) + RET + +#undef off +#undef buffer +#undef table + +#undef br_bits_read +#undef br_value +#undef br_offset +#undef peek_bits +#undef exhausted + +#undef br0 +#undef br1 +#undef br2 +#undef br3 diff --git a/huff0/decompress_8b_amd64.s.in b/huff0/decompress_8b_amd64.s.in new file mode 100644 index 0000000000..6d477a2c11 --- /dev/null +++ b/huff0/decompress_8b_amd64.s.in @@ -0,0 +1,197 @@ +// +build !appengine +// +build gc +// +build !noasm + +#include "textflag.h" +#include "funcdata.h" +#include "go_asm.h" + + +#define bufoff 256 // see decompress.go, we're using [4][256]byte table + +//func decompress4x_main_loop_x86(pbr0, pbr1, pbr2, pbr3 *bitReaderShifted, +// peekBits uint8, buf *byte, tbl *dEntrySingle) (int, bool) +TEXT ·decompress4x_8b_loop_x86(SB), NOSPLIT, $8 +#define off R8 +#define buffer DI +#define table SI + +#define br_bits_read R9 +#define br_value R10 +#define br_offset R11 +#define peek_bits R12 +#define exhausted DX + +#define br0 R13 +#define br1 R14 +#define br2 R15 +#define br3 BP + + MOVQ BP, 0(SP) + + XORQ exhausted, exhausted // exhausted = false + XORQ off, off // off = 0 + + MOVBQZX peekBits+32(FP), peek_bits + MOVQ buf+40(FP), buffer + MOVQ tbl+48(FP), table + + MOVQ pbr0+0(FP), br0 + MOVQ pbr1+8(FP), br1 + MOVQ pbr2+16(FP), br2 + MOVQ pbr3+24(FP), br3 + +main_loop: +{{ define "decode_2_values_x86" }} + // const stream = {{ var "id" }} + // br{{ var "id"}}.fillFast() + MOVBQZX bitReaderShifted_bitsRead(br{{ var "id" }}), br_bits_read + MOVQ bitReaderShifted_value(br{{ var "id" }}), br_value + MOVQ bitReaderShifted_off(br{{ var "id" }}), br_offset + + // if b.bitsRead >= 32 { + CMPQ br_bits_read, $32 + JB skip_fill{{ var "id" }} + + SUBQ $32, br_bits_read // b.bitsRead -= 32 + SUBQ $4, br_offset // b.off -= 4 + + // v := b.in[b.off-4 : b.off] + // v = v[:4] + // low := (uint32(v[0])) | (uint32(v[1]) << 8) | (uint32(v[2]) << 16) | (uint32(v[3]) << 24) + MOVQ bitReaderShifted_in(br{{ var "id" }}), AX + MOVL 0(br_offset)(AX*1), AX // AX = uint32(b.in[b.off:b.off+4]) + + // b.value |= uint64(low) << (b.bitsRead & 63) + MOVQ br_bits_read, CX + SHLQ CL, AX + ORQ AX, br_value + + // exhausted = exhausted || (br{{ var "id"}}.off < 4) + CMPQ br_offset, $4 + SETLT DL + ORB DL, DH + // } +skip_fill{{ var "id" }}: + + // val0 := br{{ var "id"}}.peekTopBits(peekBits) + MOVQ br_value, AX + MOVQ peek_bits, CX + SHRQ CL, AX // AX = (value >> peek_bits) & mask + + // v0 := table[val0&mask] + MOVW 0(table)(AX*2), AX // AX - v0 + + // br{{ var "id"}}.advance(uint8(v0.entry)) + MOVB AH, BL // BL = uint8(v0.entry >> 8) + MOVBQZX AL, CX + SHLQ CL, br_value // value <<= n + ADDQ CX, br_bits_read // bits_read += n + + // val1 := br{{ var "id"}}.peekTopBits(peekBits) + MOVQ peek_bits, CX + MOVQ br_value, AX + SHRQ CL, AX // AX = (value >> peek_bits) & mask + + // v1 := table[val1&mask] + MOVW 0(table)(AX*2), AX // AX - v1 + + // br{{ var "id"}}.advance(uint8(v1.entry)) + MOVB AH, BH // BH = uint8(v1.entry >> 8) + MOVBQZX AL, CX + SHLQ CX, br_value // value <<= n + ADDQ CX, br_bits_read // bits_read += n + + + // these two writes get coalesced + // buf[stream][off] = uint8(v0.entry >> 8) + // buf[stream][off+1] = uint8(v1.entry >> 8) + MOVW BX, {{ var "bufofs" }}(buffer)(off*1) + + // SECOND PART: + // val2 := br{{ var "id"}}.peekTopBits(peekBits) + MOVQ br_value, AX + MOVQ peek_bits, CX + SHRQ CL, AX // AX = (value >> peek_bits) & mask + + // v2 := table[val0&mask] + MOVW 0(table)(AX*2), AX // AX - v0 + + // br{{ var "id"}}.advance(uint8(v0.entry)) + MOVB AH, BL // BL = uint8(v0.entry >> 8) + MOVBQZX AL, CX + SHLQ CL, br_value // value <<= n + ADDQ CX, br_bits_read // bits_read += n + + // val3 := br{{ var "id"}}.peekTopBits(peekBits) + MOVQ peek_bits, CX + MOVQ br_value, AX + SHRQ CL, AX // AX = (value >> peek_bits) & mask + + // v3 := table[val1&mask] + MOVW 0(table)(AX*2), AX // AX - v1 + + // br{{ var "id"}}.advance(uint8(v1.entry)) + MOVB AH, BH // BH = uint8(v1.entry >> 8) + MOVBQZX AL, CX + SHLQ CX, br_value // value <<= n + ADDQ CX, br_bits_read // bits_read += n + + + // these two writes get coalesced + // buf[stream][off+2] = uint8(v2.entry >> 8) + // buf[stream][off+3] = uint8(v3.entry >> 8) + MOVW BX, {{ var "bufofs" }}+2(buffer)(off*1) + + // update the bitrader reader structure + MOVB br_bits_read, bitReaderShifted_bitsRead(br{{ var "id" }}) + MOVQ br_value, bitReaderShifted_value(br{{ var "id" }}) + MOVQ br_offset, bitReaderShifted_off(br{{ var "id" }}) +{{ end }} + + {{ set "id" "0" }} + {{ set "ofs" "0" }} + {{ set "bufofs" "0" }} {{/* id * bufoff */}} + {{ template "decode_2_values_x86" . }} + + {{ set "id" "1" }} + {{ set "ofs" "8" }} + {{ set "bufofs" "256" }} + {{ template "decode_2_values_x86" . }} + + {{ set "id" "2" }} + {{ set "ofs" "16" }} + {{ set "bufofs" "512" }} + {{ template "decode_2_values_x86" . }} + + {{ set "id" "3" }} + {{ set "ofs" "24" }} + {{ set "bufofs" "768" }} + {{ template "decode_2_values_x86" . }} + + ADDQ $4, off // off += 2 + + TESTB DH, DH // any br[i].ofs < 4? + JNZ end + + CMPQ off, $bufoff + JL main_loop +end: + MOVQ 0(SP), BP + + MOVB off, ret+56(FP) + RET +#undef off +#undef buffer +#undef table + +#undef br_bits_read +#undef br_value +#undef br_offset +#undef peek_bits +#undef exhausted + +#undef br0 +#undef br1 +#undef br2 +#undef br3 diff --git a/huff0/decompress_amd64.go b/huff0/decompress_amd64.go index fa83f78040..d47f6644f3 100644 --- a/huff0/decompress_amd64.go +++ b/huff0/decompress_amd64.go @@ -11,11 +11,21 @@ import ( ) // decompress4x_main_loop_x86 is an x86 assembler implementation -// of Decompress4X that uses BMI1 instructions. +// of Decompress4X when tablelog > 8. // go:noescape func decompress4x_main_loop_x86(pbr0, pbr1, pbr2, pbr3 *bitReaderShifted, peekBits uint8, buf *byte, tbl *dEntrySingle) uint8 +// decompress4x_8b_loop_x86 is an x86 assembler implementation +// of Decompress4X when tablelog <= 8 which decodes 4 entries +// per loop. +// go:noescape +func decompress4x_8b_loop_x86(pbr0, pbr1, pbr2, pbr3 *bitReaderShifted, + peekBits uint8, buf *byte, tbl *dEntrySingle) uint8 + +// fallback8BitSize is the size where using Go version is faster. +const fallback8BitSize = 800 + // Decompress4X will decompress a 4X encoded stream. // The length of the supplied input must match the end of a block exactly. // The *capacity* of the dst slice must match the destination size of @@ -27,10 +37,11 @@ func (d *Decoder) Decompress4X(dst, src []byte) ([]byte, error) { if len(src) < 6+(4*1) { return nil, errors.New("input too small") } - if use8BitTables && d.actualTableLog <= 8 { + + use8BitTables := d.actualTableLog <= 8 + if cap(dst) < fallback8BitSize && use8BitTables { return d.decompress4X8bit(dst, src) } - var br [4]bitReaderShifted // Decode "jump table" start := 6 @@ -77,7 +88,11 @@ func (d *Decoder) Decompress4X(dst, src []byte) ([]byte, error) { break } - off = decompress4x_main_loop_x86(&br[0], &br[1], &br[2], &br[3], peekBits, &buf[0][0], &single[0]) + if use8BitTables { + off = decompress4x_8b_loop_x86(&br[0], &br[1], &br[2], &br[3], peekBits, &buf[0][0], &single[0]) + } else { + off = decompress4x_main_loop_x86(&br[0], &br[1], &br[2], &br[3], peekBits, &buf[0][0], &single[0]) + } if debug { fmt.Print("DEBUG: ") fmt.Printf("off=%d,", off) diff --git a/huff0/generate.go b/huff0/generate.go index 9415cae393..95b082fd34 100644 --- a/huff0/generate.go +++ b/huff0/generate.go @@ -18,6 +18,10 @@ func main() { template: "decompress_amd64.s.in", output: "decompress_amd64.s", }, + { + template: "decompress_8b_amd64.s.in", + output: "decompress_8b_amd64.s", + }, } for i := range mapping {