Add -fload-plugin and -mllvm flags to gllgo

pcc commented 10 years ago

(In case you were curious, I was using this to run the Souper pass in llgo.)

axw commented 10 years ago

LGTM. Any interesting results to share?

pcc commented 10 years ago

Not yet, I'm afraid. I wanted to see what it did with libgo, but there are bugs that sometimes cause the compiler to hang inside the solver and prevent the build from finishing. Once those bugs are fixed I'll try to remember to share some results.

pcc commented 9 years ago

So I finally got around to running Souper over libgo:

The libgo build took about 11 minutes on my machine (as compared to about 1 minute normally).
benchcomp reports a geo-mean 1.8% improvement in performance for the libgo benchmarks (1.3% - 2.3% @ 95% CI).
The size of libgo.so decreased from 24538168 to 24524928 bytes.
The full list of optimizations found by Souper: http://www.pcc.me.uk/~peter/libgo-results.txt
Some of the biggest wins seem to come from image processing, compression and sorting benchmarks.

The top 20 performance improvements:

# benchmark-name without-souper-ns-op with-souper-ns-op ratio
BenchmarkStableInt64K 7497470.000000 6062380.000000 1.236721
BenchmarkAppendSpecialCase 44.300000 35.800000 1.237430
BenchmarkAfter 15362750.000000 12353980.000000 1.243547
BenchmarkAppendFloat 766.000000 614.000000 1.247557
BenchmarkNRGBAOver 1007819.000000 794335.000000 1.268758
BenchmarkAllocation 52325.000000 40988.000000 1.276593
BenchmarkDecodeDigits 9079210.000000 6884975.000000 1.318699
BenchmarkStable1e6 6632502000.000000 4992931000.000000 1.328378
BenchmarkDecodeDigitsCompress1e6 21680700.000000 15807110.000000 1.371579
BenchmarkDecodeDigitsCompress1e4 281471.000000 202906.000000 1.387199
BenchmarkDecodeDigitsDefault1e4 271866.000000 193607.000000 1.404216
BenchmarkAppendFloat32Integer 666.000000 473.000000 1.408034
BenchmarkDecodeDigitsSpeed1e5 2674337.000000 1878701.000000 1.423503
BenchmarkWriterFlush 37.800000 26.500000 1.426415
BenchmarkDecodeDigitsCompress1e5 2347250.000000 1642850.000000 1.428767
BenchmarkIndexRuneFastPath 17.600000 12.200000 1.442623
BenchmarkDecodeTwain 32195220.000000 22295950.000000 1.443994
BenchmarkAppendFloat32NegExp 844.000000 575.000000 1.467826
BenchmarkStableInt1K 116959.000000 78026.000000 1.498975
BenchmarkNRGBASrc 727460.000000 394435.000000 1.844309

axw commented 9 years ago

Some impressive numbers in there. Thanks for sharing.

go-llvm / llgo

Add -fload-plugin and -mllvm flags to gllgo #185