Currently we assume that if function may write to memory and it has MRI_ModRef behaviour in relation to the requested memory location then it means that this function writes to the memory location. In fact it's possible that it only reads from it. For example we can achieve this by using readonly attribute on the function operand.
Details
Diff Detail
- Repository
- rL LLVM
Event Timeline
So
A. This added test still fails with this patch for me :)
B. I see this is only used in the case where the code is trying to eliminate redundant stores, so seems okay otherwise.
@dberlin
A: Sorry, I should have mentioned that this change depends the alias analysis change: https://reviews.llvm.org/D29989
B: Yes, this function specifically checks for memory modifications, so it shouldn't be used to determine if store is dead.
test/Transforms/DeadStoreElimination/operand-bundles.ll | ||
---|---|---|
48 ↗ | (On Diff #88565) | Looks like today -basicaa -dse does not elide the second store in declare void @foo() declare void @bar(i8*) define void @test4(i8* %ptr) { store i8 0, i8* %ptr, align 4 call void @foo() readonly store i8 0, i8* %ptr, align 4 call void @bar(i8* nocapture %ptr) ret void } Will it be able to do that with your changes? If so, adding that as a test case will be nice. |
The actual code change looks obviously correct.
LGTM once the AA change lands.
test/Transforms/DeadStoreElimination/operand-bundles.ll | ||
---|---|---|
48 ↗ | (On Diff #88565) | It doesn't look like we have the framework for this in DSE which is unfortunate. Oddly enough earlycse probably will get this example. |