forked from OSchip/llvm-project
a0c6ae02a5
scalarizePHI only looked for phis that have exactly two uses - the "latch" use, and an extract. Unfortunately, we can not assume all equivalent extracts are CSE'd, since InstCombine itself may create an extract which is a duplicate of an existing one. This extends it to handle several distinct extracts from the same index. This should fix at least some of the performance regressions from PR27988. Differential Revision: http://reviews.llvm.org/D20983 llvm-svn: 271961 |
||
|---|---|---|
| .. | ||
| agg-interleave-a2.ll | ||
| large-loop-rdx.ll | ||
| lit.local.cfg | ||
| small-loop-rdx.ll | ||
| stride-vectorization.ll | ||
| vectorize-only-for-real.ll | ||
| vsx-tsvc-s173.ll | ||