Implement two more transforms of atomicrmw in InstCombine.
- We can convert an atomicrmw which produces a known value in memory into an xchg instead.
- We can convert an atomicrmw xchg w/o users into a store for some orderings.
Details
Details
Diff Detail
Diff Detail
- Repository
- rL LLVM
Event Timeline
Comment Actions
A few nits, LGTM otherwise.
| lib/Transforms/InstCombine/InstCombineAtomicRMW.cpp | ||
|---|---|---|
| 52 ↗ | (On Diff #187031) | "its value's" |
| 59 ↗ | (On Diff #187031) | You can move RMWI.getOperation() in here. |
| 73 ↗ | (On Diff #187031) | FAdd / FSub with any NaN are also saturating. Technically xchg is also saturating, pre your description above... |
| 99 ↗ | (On Diff #187031) | Do you want to factor this out and run it after isSaturating succeeds above? That way instcombine doesn't have to do two passes to do this transform. |
| 133 ↗ | (On Diff #187031) | I guess here you could have done return eraseInstFromFunction(RMWI); as well? |
| lib/Transforms/InstCombine/InstCombineAtomicRMW.cpp | ||
|---|---|---|
| 59 ↗ | (On Diff #187031) | Will do this and the other site of the same in a follow up nfc |
| 99 ↗ | (On Diff #187031) | Saving a cycle through the iteration didn't seem worth code complexity. |
| 133 ↗ | (On Diff #187031) | InstCombine does that automatically for non-void instructions. |