User talk:Bluefoxicy/gcc optimizations
cachegrind
I tried cachegrind on povbench to see if I can attribute the cost to I1 misses.. Ignore L2 miss rate; L2 refs are L1 miss rate. I'm interested primarily in I1 miss rate, not D1.
CFLAGS="-march=pentiumpro -O2" valgrind --tool=cachegrind --I1=16384,4,32 --D1=16384,4,32 --L2=64,2,32 povbench/bin/povray benchmark.ini benchmark.pov
==1319== I refs: 41,083,755,757 ==1319== I1 misses: 182,725,637 ==1319== L2i misses: 152,387,115 ==1319== I1 miss rate: 0.44% ==1319== L2i miss rate: 0.37% ==1319== ==1319== D refs: 21,345,148,925 (15,955,324,459 rd + 5,389,824,466 wr) ==1319== D1 misses: 221,554,710 ( 174,862,567 rd + 46,692,143 wr) ==1319== L2d misses: 221,064,369 ( 174,419,792 rd + 46,644,577 wr) ==1319== D1 miss rate: 1.0% ( 1.0% + 0.8% ) ==1319== L2d miss rate: 1.0% ( 1.0% + 0.8% ) ==1319== ==1319== L2 refs: 404,280,347 ( 357,588,204 rd + 46,692,143 wr) ==1319== L2 misses: 373,451,484 ( 326,806,907 rd + 46,644,577 wr) ==1319== L2 miss rate: 0.5% ( 0.5% + 0.8% )
The below is with -fno-tree-pre. I don't see the problem; there's only a 0.05% increased miss rate, which should translate to 1.25% difference in I1... not doing the math right now.
CFLAGS="-march=pentiumpro -O2 -fno-tree-pre" valgrind --tool=cachegrind --I1=16384,4,32 --D1=16384,4,32 --L2=64,2,32 povbench/bin/povray benchmark.ini benchmark.pov
==11679== I refs: 25,339,345,936 ==11679== I1 misses: 126,151,035 ==11679== L2i misses: 104,445,090 ==11679== I1 miss rate: 0.49% ==11679== L2i miss rate: 0.41% ==11679== ==11679== D refs: 13,199,180,344 (9,804,287,156 rd + 3,394,893,188 wr) ==11679== D1 misses: 147,020,945 ( 115,976,223 rd + 31,044,722 wr) ==11679== L2d misses: 146,782,461 ( 115,759,617 rd + 31,022,844 wr) ==11679== D1 miss rate: 1.1% ( 1.1% + 0.9% ) ==11679== L2d miss rate: 1.1% ( 1.1% + 0.9% ) ==11679== ==11679== L2 refs: 273,171,980 ( 242,127,258 rd + 31,044,722 wr) ==11679== L2 misses: 251,227,551 ( 220,204,707 rd + 31,022,844 wr) ==11679== L2 miss rate: 0.6% ( 0.6% + 0.9% )
