f7f70bf1b
Denys Vlasenko
2015-02-02
Subject
gzip: speed up and shrink put_16bit()
Binary
No binary (build failed)
Architecture & Compiler Flags
—
Build Log
(log yok)
Commit Diff GitHub →
archival/gzip.c+34 -7
| @@ -417,19 +417,46 @@ static void flush_outbuf(void) | |
| #define put_8bit(c) \ | |
| do { \ | |
| G1.outbuf[G1.outcnt++] = (c); \ | |
| - | if (G1.outcnt == OUTBUFSIZ) flush_outbuf(); \ |
| + | if (G1.outcnt == OUTBUFSIZ) \ |
| + | flush_outbuf(); \ |
| } while (0) | |
| /* Output a 16 bit value, lsb first */ | |
| static void put_16bit(ush w) | |
| { | |
| - | if (G1.outcnt < OUTBUFSIZ - 2) { |
| - | G1.outbuf[G1.outcnt++] = w; |
| - | G1.outbuf[G1.outcnt++] = w >> 8; |
| - | } else { |
| - | put_8bit(w); |
| - | put_8bit(w >> 8); |
| + | /* GCC 4.2.1 won't optimize out redundant loads of G1.outcnt |
| + | * (probably because of fear of aliasing with G1.outbuf[] |
| + | * stores), do it explicitly: |
| + | */ |
| + | unsigned outcnt = G1.outcnt; |
| + | uch *dst = &G1.outbuf[outcnt]; |
| + | |
| + | #if BB_UNALIGNED_MEMACCESS_OK && BB_LITTLE_ENDIAN |
| + | if (outcnt < OUTBUFSIZ-2) { |
| + | /* Common case */ |
| + | ush *dst16 = (void*) dst; |
| + | *dst16 = w; /* unalinged LSB 16-bit store */ |
| + | G1.outcnt = outcnt + 2; |
| + | return; |
| + | } |
| + | *dst = (uch)w; |
| + | w >>= 8; |
| + | #else |
| + | *dst++ = (uch)w; |
| + | w >>= 8; |
| + | if (outcnt < OUTBUFSIZ-2) { |
| + | /* Common case */ |
| + | *dst = w; |
| + | G1.outcnt = outcnt + 2; |
| + | return; |
| } | |
| + | #endif |
| + | |
| + | /* Slowpath: we will need to do flush_outbuf() */ |
| + | G1.outcnt++; |
| + | if (G1.outcnt == OUTBUFSIZ) |
| + | flush_outbuf(); |
| + | put_8bit(w); |
| } | |
| static void put_32bit(ulg n) | |
include/platform.h+2 -0
| @@ -217,6 +217,7 @@ typedef uint64_t bb__aliased_uint64_t FIX_ALIASING; | |
| * a lvalue. This makes it more likely to not swap them by mistake | |
| */ | |
| #if defined(i386) || defined(__x86_64__) || defined(__powerpc__) | |
| + | # define BB_UNALIGNED_MEMACCESS_OK 1 |
| # define move_from_unaligned_int(v, intp) ((v) = *(bb__aliased_int*)(intp)) | |
| # define move_from_unaligned_long(v, longp) ((v) = *(bb__aliased_long*)(longp)) | |
| # define move_from_unaligned16(v, u16p) ((v) = *(bb__aliased_uint16_t*)(u16p)) | |
| @@ -225,6 +226,7 @@ typedef uint64_t bb__aliased_uint64_t FIX_ALIASING; | |
| # define move_to_unaligned32(u32p, v) (*(bb__aliased_uint32_t*)(u32p) = (v)) | |
| /* #elif ... - add your favorite arch today! */ | |
| #else | |
| + | # define BB_UNALIGNED_MEMACCESS_OK 0 |
| /* performs reasonably well (gcc usually inlines memcpy here) */ | |
| # define move_from_unaligned_int(v, intp) (memcpy(&(v), (intp), sizeof(int))) | |
| # define move_from_unaligned_long(v, longp) (memcpy(&(v), (longp), sizeof(long))) | |