[Mlir-commits] [mlir] [mlir][memref] Fix hoist-static-allocs option of buffer-results-to-out-params when function parameters are returned (PR #102093)

[llvmlistbot at llvm.org]

================
@@ -120,10 +128,15 @@ static LogicalResult updateReturnOps(func::FuncOp func,
     }
     OpBuilder builder(op);
     for (auto [orig, arg] : llvm::zip(copyIntoOutParams, appendedEntryArgs)) {
-      if (hoistStaticAllocs && isa<memref::AllocOp>(orig.getDefiningOp()) &&
-          mlir::cast<MemRefType>(orig.getType()).hasStaticShape()) {
+      bool mayHoistStaticAlloc =
+          hoistStaticAllocs &&
+          mlir::cast<MemRefType>(orig.getType()).hasStaticShape();
+      if (mayHoistStaticAlloc &&
+          isa_and_nonnull<memref::AllocOp>(orig.getDefiningOp())) {
         orig.replaceAllUsesWith(arg);
         orig.getDefiningOp()->erase();
+      } else if (mayHoistStaticAlloc && isFunctionArgument(orig)) {
----------------
Menooker wrote:

It indeed changes the behavior of the program. It is not sure to me what is the expected behavior of this pass without `hoistStaticAllocs`, if we return the memref from an arg. If we should copy the data from arg to arg?

https://github.com/llvm/llvm-project/pull/102093