[llvm-commits] [STATS] Compile time measurements r107882 vs. r108240

[Gabor Greif]

Hi all,

I was curious what effect in compile performance my commits
throughout the last few days did cause. So I ran a loop three
times (after a warmup round) on several revisions of the "opt"
tool on a raw 403.gcc bitcode file ("-std-compile-opts").

Here are the results:

gabor at google7:~/llvm-build$ grep "Total Execution" */1.info
r107882/1.info:  Total Execution Time: 32.1460 seconds (32.1268 wall clock)
r107990/1.info:  Total Execution Time: 32.0500 seconds (32.0840 wall clock)
r108150/1.info:  Total Execution Time: 32.0700 seconds (32.1464 wall clock)
r108239/1.info:  Total Execution Time: 32.0780 seconds (31.9802 wall clock)
r108240/1.info:  Total Execution Time: 31.8140 seconds (31.7415 wall clock)
gabor at google7:~/llvm-build$ grep "Total Execution" */2.info
r107882/2.info:  Total Execution Time: 32.4140 seconds (32.1649 wall clock)
r107990/2.info:  Total Execution Time: 32.0940 seconds (31.9776 wall clock)
r108150/2.info:  Total Execution Time: 32.0220 seconds (32.0771 wall clock)
r108239/2.info:  Total Execution Time: 32.1420 seconds (32.0734 wall clock)
r108240/2.info:  Total Execution Time: 31.7900 seconds (31.7341 wall clock)
gabor at google7:~/llvm-build$ grep "Total Execution" */3.info
r107882/3.info:  Total Execution Time: 32.2140 seconds (32.1837 wall clock)
r107990/3.info:  Total Execution Time: 32.0900 seconds (32.1469 wall clock)
r108150/3.info:  Total Execution Time: 32.0980 seconds (32.0384 wall clock)
r108239/3.info:  Total Execution Time: 32.0540 seconds (32.0528 wall clock)
r108240/3.info:  Total Execution Time: 31.8500 seconds (31.7337 wall clock)


Dividing the fastest "after" by the fastest "before" gives:

gabor at google7:~/llvm-build$ expr 317900000 / 32146
9889

This is more than 1% speedup.

Just looking at the "CallInst operand rotation" (r108240) gives:

gabor at google7:~/llvm-build$ expr 317900000 / 32054
9917

About 0.83%. Not too bad.

Hope you enjoyed these numbers,

	Gabor


PS: any ideas what options for heavy-duty passes in "opt" could I test?