[PATCH] D101216: [AArch64] Avoid generating LL/SC for atomicrmw before regalloc at -O0.

[Eli Friedman via Phabricator via llvm-commits]

efriedma created this revision.
efriedma added reviewers: t.p.northover, arsenm, rogfer01, mgorny.
Herald added subscribers: tmatheson, danielkiss, jfb, hiraditya, kristof.beyls.
efriedma requested review of this revision.
Herald added a subscriber: wdng.
Herald added a project: LLVM.

This is a retread of the issue we had in the past with cmpxchg: fast regalloc can generate spills in weird places, so the ll/sc sequence never terminates.  To avoid this, don't use the early ll/sc expansion at -O0.

I didn't spend the time to implement individual instructions for each atomicrmw instruction; instead, it's just expanding to the known-good cmpxchg. This is inefficient, but not sure how much we care; we don't use these sequences with LSE or atomic outlining.

Fixes https://bugs.llvm.org/show_bug.cgi?id=48017


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D101216

Files:
  llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
  llvm/test/CodeGen/AArch64/atomicrmw-O0.ll

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D101216.340213.patch
Type: text/x-patch
Size: 11642 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20210424/d3f0371b/attachment.bin>