https://github.com/victor-eds updated https://github.com/llvm/llvm-project/pull/228771
>From e9c665f8dbcf1db8b568f11e7a4dab74e0375f2f Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?V=C3=ADctor=20P=C3=A9rez=20Carrasco?= <[email protected]> Date: Fri, 2 Oct 2026 18:48:55 -0700 Subject: [PATCH] [mlir][shape] Report the in-place change of the assuming_all fold MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit The `shape.assuming_all` fold visits its inputs in reverse order. It erases each constant input from the operands, and it returns null at the first input that is not constant. If the fold already erased an input at that point, it changed the op but reported a failure, so no driver knew about the change. The fold now returns the result of the op in this case, which is the in-place signal of the single-result fold form. No test fails on main. The later patch "[mlir] Apply partial folds in the greedy pattern rewrite driver" adds a fingerprint check of each fold to the greedy driver. The check runs only in a build with `-DMLIR_ENABLE_EXPENSIVE_PATTERN_API_CHECKS=ON`. In such a build of that patch, with this fix reverted, the existing test `Dialect/Shape/canonicalize.mlir` fails: $ mlir-opt -split-input-file -allow-unregistered-dialect \ -canonicalize="test-convergence" \ mlir/test/Dialect/Shape/canonicalize.mlir LLVM ERROR: fold returned failure but changed the operation Signed-off-by: Víctor Pérez Carrasco <[email protected]> --- mlir/lib/Dialect/Shape/IR/Shape.cpp | 13 +++++++++---- 1 file changed, 9 insertions(+), 4 deletions(-) diff --git a/mlir/lib/Dialect/Shape/IR/Shape.cpp b/mlir/lib/Dialect/Shape/IR/Shape.cpp index e01b705e5610a..4487c2e652276 100644 --- a/mlir/lib/Dialect/Shape/IR/Shape.cpp +++ b/mlir/lib/Dialect/Shape/IR/Shape.cpp @@ -615,11 +615,16 @@ void AssumingAllOp::getCanonicalizationPatterns(RewritePatternSet &patterns, OpFoldResult AssumingAllOp::fold(FoldAdaptor adaptor) { // Iterate in reverse to first handle all constant operands. They are // guaranteed to be the tail of the inputs because this is commutative. - for (int idx = adaptor.getInputs().size() - 1; idx >= 0; idx--) { + int numInputs = adaptor.getInputs().size(); + for (int idx = numInputs - 1; idx >= 0; idx--) { Attribute a = adaptor.getInputs()[idx]; - // Cannot fold if any inputs are not constant; - if (!a) - return nullptr; + // Cannot fold if any inputs are not constant; if the loop already erased + // constant inputs, the op's own result reports an in-place fold. + if (!a) { + if (idx == numInputs - 1) + return nullptr; + return getResult(); + } // We do not need to keep statically known values after handling them in // this method. _______________________________________________ llvm-branch-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits
