[PATCH] D21449: Target independent codesize heuristics for Loop Idiom Recognition

[Sunita_Marathe via llvm-commits]

Sunita_Marathe added a comment.

As reported in an earlier note, the multi-block loop heuristic, just on its own, affects only the 176.gcc benchmark, which shows a small code size reduction of 39 bytes resulting from avoiding recognition of a single memset idiom in a top-level multi-block loop in function find_reloads.   The before/after function size is 12354/12315 bytes.  The before/after results for the CPU2000 suite are:

BEFORE:
177.mesa total size = 392061
179.art total size = 12803
183.equake total size = 13739
188.ammp total size = 109140
164.gzip total size = 28385
175.vpr total size = 92008
176.gcc total size = 928659
181.mcf total size = 7964
186.crafty total size = 115463
197.parser total size = 64733
253.perlbmk total size = 378322
254.gap total size = 340958
255.vortex total size = 185553
256.bzip2 total size = 20673
300.twolf total size = 156159
Total Size =  2846620

AFTER:
177.mesa total size = 392061
179.art total size = 12803
183.equake total size = 13739
188.ammp total size = 109140
164.gzip total size = 28385
175.vpr total size = 92008
176.gcc total size = 928620
181.mcf total size = 7964
186.crafty total size = 115463
197.parser total size = 64733
253.perlbmk total size = 378322
254.gap total size = 340958
255.vortex total size = 185553
256.bzip2 total size = 20673
300.twolf total size = 156159
Total Size =  2846581


http://reviews.llvm.org/D21449