From 0c6786c6fad563f181e66c90df2a74597ce6d18b Mon Sep 17 00:00:00 2001 From: Joshua Haberman Date: Sun, 20 Feb 2011 08:29:36 -0800 Subject: Split varint decoders into separate .h file. This makes it easier to benchmark and test the multiple possible implementations of varint decoding. --- src/upb_varint_decoder.h | 120 +++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 120 insertions(+) create mode 100644 src/upb_varint_decoder.h (limited to 'src') diff --git a/src/upb_varint_decoder.h b/src/upb_varint_decoder.h new file mode 100644 index 0000000..8619596 --- /dev/null +++ b/src/upb_varint_decoder.h @@ -0,0 +1,120 @@ +/* + * upb - a minimalist implementation of protocol buffers. + * + * A number of routines for varint decoding (we keep them all around to have + * multiple approaches available for benchmarking). All of these functions + * require the buffer to have at least 10 bytes available; if we don't know + * for sure that there are 10 bytes, then there is only one viable option + * (branching on every byte). + * + * Copyright (c) 2011 Joshua Haberman. See LICENSE for details. + */ + +#ifndef UPB_VARINT_DECODER_H_ +#define UPB_VARINT_DECODER_H_ + +#include "upb.h" +#include +#include + +#ifdef __cplusplus +extern "C" { +#endif + +// All decoding functions return this struct by value. +typedef struct { + const char *p; // NULL if the varint was unterminated. + uint64_t val; +} upb_decoderet; + +// A basic branch-based decoder, uses 32-bit values to get good performance +// on 32-bit architectures (but performs well on 64-bits also). +INLINE upb_decoderet upb_decode_varint_branch32(const char *p) { + upb_decoderet r = {NULL, 0}; + uint32_t low, high = 0; + uint32_t b; + b = *(p++); low = (b & 0x7f) ; if(!(b & 0x80)) goto done; + b = *(p++); low |= (b & 0x7f) << 7; if(!(b & 0x80)) goto done; + b = *(p++); low |= (b & 0x7f) << 14; if(!(b & 0x80)) goto done; + b = *(p++); low |= (b & 0x7f) << 21; if(!(b & 0x80)) goto done; + b = *(p++); low |= (b & 0x7f) << 28; + high = (b & 0x7f) >> 4; if(!(b & 0x80)) goto done; + b = *(p++); high |= (b & 0x7f) << 3; if(!(b & 0x80)) goto done; + b = *(p++); high |= (b & 0x7f) << 10; if(!(b & 0x80)) goto done; + b = *(p++); high |= (b & 0x7f) << 17; if(!(b & 0x80)) goto done; + b = *(p++); high |= (b & 0x7f) << 24; if(!(b & 0x80)) goto done; + b = *(p++); high |= (b & 0x7f) << 31; if(!(b & 0x80)) goto done; + return r; + +done: + r.val = ((uint64_t)high << 32) | low; + r.p = p; + return r; +} + +// Like the previous, but uses 64-bit values. +INLINE upb_decoderet upb_decode_varint_branch64(const char *p) { + uint64_t val; + uint64_t b; + upb_decoderet r = {(void*)0, 0}; + b = *(p++); val = (b & 0x7f) ; if(!(b & 0x80)) goto done; + b = *(p++); val |= (b & 0x7f) << 7; if(!(b & 0x80)) goto done; + b = *(p++); val |= (b & 0x7f) << 14; if(!(b & 0x80)) goto done; + b = *(p++); val |= (b & 0x7f) << 21; if(!(b & 0x80)) goto done; + b = *(p++); val |= (b & 0x7f) << 28; if(!(b & 0x80)) goto done; + b = *(p++); val |= (b & 0x7f) << 35; if(!(b & 0x80)) goto done; + b = *(p++); val |= (b & 0x7f) << 42; if(!(b & 0x80)) goto done; + b = *(p++); val |= (b & 0x7f) << 49; if(!(b & 0x80)) goto done; + b = *(p++); val |= (b & 0x7f) << 56; if(!(b & 0x80)) goto done; + b = *(p++); val |= (b & 0x7f) << 63; if(!(b & 0x80)) goto done; + return r; + +done: + r.val = val; + r.p = p; + return r; +} + +#ifdef __SSE__ + +#include + +// Avoids branches (this can very likely be improved). Requires SSE. +INLINE upb_decoderet upb_decode_varint_nobranch(const char *p) { + upb_decoderet r = {(void*)0, 0}; + __m128i val128 = _mm_loadu_si128((void*)p); + unsigned int continuation_bits = _mm_movemask_epi8(val128); + unsigned int bsr_val = ~continuation_bits; + int varint_length = __builtin_ffs(bsr_val); + if (varint_length > 10) return r; + + uint16_t twob; + memcpy(&twob, p, 2); + twob &= 0x7f7f; + twob = ((twob & 0xff00) >> 1) | (twob & 0xff); + + uint64_t eightb; + memcpy(&eightb, p + 2, 8); + eightb &= 0x7f7f7f7f7f7f7f7f; + eightb = ((eightb & 0xff00ff00ff00ff00) >> 1) | (eightb & 0x00ff00ff00ff00ff); + eightb = ((eightb & 0xffff0000ffff0000) >> 2) | (eightb & 0x0000ffff0000ffff); + eightb = ((eightb & 0xffffffff00000000) >> 4) | (eightb & 0x00000000ffffffff); + + uint64_t all_bits = twob | (eightb << 14); + int varint_bits = varint_length * 7; + uint64_t mask = varint_bits == 70 ? (uint64_t)-1 : (1ULL << (varint_bits)) - 1; + r.val = all_bits & mask; + r.p = p + varint_length; + return r; +} + +#endif + +// For now, always use the branch32 decoder. +#define upb_decode_varint_fast upb_decode_varint_branch32 + +#ifdef __cplusplus +} /* extern "C" */ +#endif + +#endif /* UPB_VARINT_DECODER_H_ */ -- cgit v1.2.3