diff options
| author | Pirama Arumuga Nainar <pirama@google.com> | 2018-04-10 13:23:10 -0700 |
|---|---|---|
| committer | Pirama Arumuga Nainar <pirama@google.com> | 2018-04-10 13:23:10 -0700 |
| commit | e8594d5184f397e9cad4e7803f47894107bdd3f4 (patch) | |
| tree | b90a3c9167826a02ab0352d535089d1a9e397df5 | |
| parent | 111055a9fe97192942044a08b63f9b3fae896d0c (diff) | |
| download | llvm-e8594d5184f397e9cad4e7803f47894107bdd3f4.tar.gz | |
[MemorySSA] Be less aggressive with @llvm.lifetime.start
This is a workaround to keep LLVM from incorrectly removing stores in
rare cases. It has not yet been committed, and -- due to the direction
we seem to want to take MSSA in -- probably never will be.
It should be sufficient to work around the problem until we agree on a
good way forward.
Additional context: https://reviews.llvm.org/D43269
This is rebased to r327695 by gbiv@chromium.org.
Change-Id: Ib78a9e12845a28a48bbc4154e89775b1ef4b5319
| -rw-r--r-- | lib/Analysis/MemorySSA.cpp | 4 |
1 files changed, 2 insertions, 2 deletions
diff --git a/lib/Analysis/MemorySSA.cpp b/lib/Analysis/MemorySSA.cpp index c8fef6a1f16..0c5bff09ba8 100644 --- a/lib/Analysis/MemorySSA.cpp +++ b/lib/Analysis/MemorySSA.cpp @@ -257,13 +257,13 @@ static ClobberAlias instructionClobbersQuery(MemoryDef *MD, if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(DefInst)) { // These intrinsics will show up as affecting memory, but they are just - // markers. + // markers, mostly. switch (II->getIntrinsicID()) { case Intrinsic::lifetime_start: if (UseCS) return {false, NoAlias}; AR = AA.alias(MemoryLocation(II->getArgOperand(1)), UseLoc); - return {AR == MustAlias, AR}; + return {AR != NoAlias, AR}; case Intrinsic::lifetime_end: case Intrinsic::invariant_start: case Intrinsic::invariant_end: |