[RFC][PATCH, ASan] ASan and TSan passes place change.
Maxim Ostapenko
m.ostapenko@partner.samsung.com
Fri Aug 21 09:56:00 GMT 2015
On 07/08/15 11:42, Maxim Ostapenko wrote:
> Hi!
>
> This tiny patch moves ASan and TSan passes just before sanopt pass,
> that gives nice speedup on several SPEC2006 tests (-O2 optimization
> level was used) on x86:
>
> TESTCASE ORIGINAL PATCHED SPEEDUP
> 401.bzip2 806 798 0.99%
> 403.gcc 506 475 6.13%
> 429.mcf 337 295 12.46%
> 456.hmmer 1005 1017 -1.19%
> 458.sjeng 993 946 4.73%
> 462.libquantum 312 290 7.05%
> 473.astar 550 539 2.00%
> 483.xalancbmk 439 363 17.31%
> 410.bwaves 1720 1712 0.47%
> 416.gamess 3101 3105 -0.13%
> 433.milc 469 358 23.67%
> 434.zeusmp 1497 1499 -0.13%
> 435.gromacs 926 905 2.27%
> 436.cactusADM 2572 2612 -1.56%
> 437.leslie3d 1472 1470 0.14%
> 444.namd 474 452 4.64%
> 447.dealII 505 429 15.05%
> 459.GemsFDTD 1812 1812 0.00%
> 465.tonto 1289 1292 -0.23%
> 470.lbm 302 228 24.50%
> 482.sphinx3 1030 617 40.10%
>
> Does this change looks reasonable for trunk?
>
> -Maxim
I've rechecked testing results and found out, that such huge speedup in
appropriate tests was caused by missing instrumentation of
TARGET_MEM_REFs. After fixing this, all speedup gone, so moving ASan
pass down to pipeline doesn't make any sense. Sorry for noise.
-Maxim
More information about the Gcc-patches
mailing list