[PATCH] D32998: [SROA] enable splitting for non-whole-alloca loads and stores

[Hiroshi Inoue via Phabricator via llvm-commits]

inouehrs added a comment.

I conducted more performance measurements.
Overall, I did not find significant degradetions or improvements except for the iterotor example above.

tcmalloc (on ppc64le)
This patch makes additional splitting but no visible changes in malloc_bench score

gipfeli (on ppc64le) 
This patch makes additional splitting in the benchmark harness (for istreambuf_iterator) but not in compressor/decompressor.

snappy (on ppc64le)
This patch makes additional splitting but no visible changes in snappy-unittest score

SPECCPU2006 (on x86_64)
No additional splitting observed in 12 out of 19 C/C++ benchmarks.
Even for benchmarks with additional splitting the changes in the score is within range of fluctuation.

  403.gcc	0.00%
  462.libquantum	0.00%
  471.omnetpp	-0.42%
  473.astar	0.00%
  483.xalancbmk	0.20%
  447.dealII	0.13%
  453.povray	-0.06%
  SPECINT (5)	-0.04%
  SPECFP  (2)	0.04%
  TOTAL   (7)	-0.02%




https://reviews.llvm.org/D32998