[PATCH] D115406: [SystemZ] Improve eliminateFrameIndex() to use STDY/STEY/LDY for VR32/VR64 when possible.
Jonas Paulsson via Phabricator via llvm-commits
llvm-commits at lists.llvm.org
Sat May 7 10:54:00 PDT 2022
jonpa added a comment.
Herald added a project: All.
This patch seems to have a much lesser impact now: In December last year this gave ~6000 less LAYs, and now only ~250 less. This seems to be due to the added ordering of the frame objects - if I disable that ordering I get some 6000 more LAYs in the output...
I ran the patch with and without the use of LEY and this time I did not see that it made any difference (earlier there was a regression with LEY). Now there are only a handful of extra LEYs emitted anyway...
I also made the experiment again with substituting all LEY:s with LAY+(LDE or LEY). In one build an LEY would result in LAY+LDE, and in another LAY+LEY. This would for all else the same, including alignments. In the "LEY build", the number LEYs went from 390 to 38, so most were gone. I could not see in any benchmark on either default or off that LEYs would be worse.
So overall, this seems to not give any noticable performance impact on SPEC to me. Given the above I would probably suggest putting back the LEY emission to the patch, or?
"default fp-contract":
lay : 55499 55231 -268
std : 65122 64967 -155
stdy : 865 1020 +155
ld : 98755 98666 -89
ldy : 3031 3120 +89
ste : 51716 51698 -18
stey : 286 304 +18
ley : 423 429 +6
lde : 68442 68436 -6
lg : 976008 976007 -1
stg : 369621 369620 -1
OPCDIFFS: -270
Even less for some reason with "fp-contract off":
lay : 55247 55193 -54
...
OPCDIFFS: -54
CHANGES SINCE LAST ACTION
https://reviews.llvm.org/D115406/new/
https://reviews.llvm.org/D115406
More information about the llvm-commits
mailing list