[PATCH] D55403: [PHIElimination] Allow breaking loop backedges in certain cases.

[Eli Friedman via Phabricator via llvm-commits]

efriedma created this revision.
efriedma added reviewers: craig.topper, MatzeB, qcolombet.

This reduces the number of spills in some cases, like the new testcase phi-backedge-split.ll. It seems like it helps the allocator generate better code for PHIs where the operand is another PHI, since the implied COPY shouldn't really exist on any path other than the backedge.

It's possible the heuristic could be improved here; it seems like this makes the code slightly worse in a couple of the x86 regression tests, like CodeGen/X86/madd.ll. But I'm not really sure what else makes sense to check here; I mean, I could check if the predecessor vreg is specifically a PHI, but that seems hacky.

I'm not confident this is the best way to solve the spilling inefficiency I see in phi-backedge-split.ll, but I'm not sure what the alternative would look like.


Repository:
  rL LLVM

https://reviews.llvm.org/D55403

Files:
  lib/CodeGen/PHIElimination.cpp
  test/CodeGen/Hexagon/swp-kernel-phi1.ll
  test/CodeGen/Thumb/phi-backedge-split.ll
  test/CodeGen/X86/bug26810.ll
  test/CodeGen/X86/madd.ll
  test/CodeGen/X86/sad.ll

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D55403.177089.patch
Type: text/x-patch
Size: 93606 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20181207/1450127b/attachment.bin>