Apply UnsignedWhenEquivalent at the ModuleOp level. (#21743)
TSan revealed a race condition when UnsignedWhenEquivalent was being
applied. UnsignedWhenEquivalent will call DeadCodeAnalysis which reads
data of IREE::Util::FuncOp, however, a different thread will be
modifying these operations.
Signed-off-by: Erick Ochoa <erick.ochoalopez@amd.com>
diff --git a/compiler/src/iree/compiler/Dialect/VM/Transforms/Passes.cpp b/compiler/src/iree/compiler/Dialect/VM/Transforms/Passes.cpp
index a5c185f..ff0e57c 100644
--- a/compiler/src/iree/compiler/Dialect/VM/Transforms/Passes.cpp
+++ b/compiler/src/iree/compiler/Dialect/VM/Transforms/Passes.cpp
@@ -107,8 +107,9 @@
// Remove if it is added there.
// https://github.com/llvm/llvm-project/issues/78458
.addPass(affine::createAffineExpandIndexOpsPass)
- .addPass(mlir::createLowerAffinePass)
- .addPass(mlir::arith::createArithUnsignedWhenEquivalentPass);
+ .addPass(mlir::createLowerAffinePass);
+
+ passManager.addPass(mlir::arith::createArithUnsignedWhenEquivalentPass());
// Propagate buffer subranges throughout the program - this should remove any
// remaining subspans and give us a smaller surface area during conversion.
