qbit/go
1
0
Fork 0
mirror of https://github.com/golang/go synced 2024-09-24 19:40:12 -06:00
Code Releases Activity

cmd/6g: change sbop swap logic

Browse Source 
I added the nl->op == OLITERAL case during the recent
performance round, and while it helps for small integer constants,
it hurts for floating point constants.  In the Mandelbrot benchmark
it causes 2*Zr*Zi to compile like Zr*2*Zi:

        0x000000000042663d <+249>:	movsd  %xmm6,%xmm0
        0x0000000000426641 <+253>:	movsd  $2,%xmm1
        0x000000000042664a <+262>:	mulsd  %xmm1,%xmm0
        0x000000000042664e <+266>:	mulsd  %xmm5,%xmm0

instead of:

        0x0000000000426835 <+276>:	movsd  $2,%xmm0
        0x000000000042683e <+285>:	mulsd  %xmm6,%xmm0
        0x0000000000426842 <+289>:	mulsd  %xmm5,%xmm0

It is unclear why that has such a dramatic performance effect
in a tight loop, but it's obviously slightly better code, so go with it.

benchmark                 old ns/op    new ns/op    delta
BenchmarkBinaryTree17    5957470000   5973924000   +0.28%
BenchmarkFannkuch11      3811295000   3869128000   +1.52%
BenchmarkGobDecode         26001900     25670500   -1.27%
BenchmarkGobEncode         12051430     11948590   -0.85%
BenchmarkGzip                177432       174821   -1.47%
BenchmarkGunzip               10967        10756   -1.92%
BenchmarkJSONEncode        78924750     79746900   +1.04%
BenchmarkJSONDecode       313606400    307081600   -2.08%
BenchmarkMandelbrot200     13670860      8200725  -40.01%  !!!
BenchmarkRevcomp25M      1179194000   1206539000   +2.32%
BenchmarkTemplate         447931200    443948200   -0.89%
BenchmarkMD5Hash1K             2856         2873   +0.60%
BenchmarkMD5Hash8K            22083        22029   -0.24%

benchmark                  old MB/s     new MB/s  speedup
BenchmarkGobDecode            29.52        29.90    1.01x
BenchmarkGobEncode            63.69        64.24    1.01x
BenchmarkJSONEncode           24.59        24.33    0.99x
BenchmarkJSONDecode            6.19         6.32    1.02x
BenchmarkRevcomp25M          215.54       210.66    0.98x
BenchmarkTemplate              4.33         4.37    1.01x
BenchmarkMD5Hash1K           358.54       356.31    0.99x
BenchmarkMD5Hash8K           370.95       371.86    1.00x

R=ken2
CC=golang-dev
https://golang.org/cl/6261051
This commit is contained in:
Russ Cox 2012-05-30 10:22:33 -04:00
parent dbcdce5866
commit de96df1b02
1 changed files with 19 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

20
src/cmd/6g/cgen.c
View File

@ -396,7 +396,25 @@ cgen(Node *n, Node *res)
goto ret;
sbop: // symmetric binary
if(nl->ullman < nr->ullman || nl->op == OLITERAL) {
/*
* put simplest on right - we'll generate into left
* and then adjust it using the computation of right.
* constants and variables have the same ullman
* count, so look for constants specially.
*
* an integer constant we can use as an immediate
* is simpler than a variable - we can use the immediate
* in the adjustment instruction directly - so it goes
* on the right.
*
* other constants, like big integers or floating point
* constants, require a mov into a register, so those
* might as well go on the left, so we can reuse that
* register for the computation.
*/
if(nl->ullman < nr->ullman ||
(nl->ullman == nr->ullman &&
(smallintconst(nl) || (nr->op == OLITERAL && !smallintconst(nr))))) {
r = nl;
nl = nr;
nr = r;