[Mlir-commits] [mlir] [mlir] Allow folding dynamic full size subviews (PR #140619)

[Matthias Springer]

================
@@ -3206,7 +3215,7 @@ class TrivialSubViewOpFolder final : public OpRewritePattern<SubViewOp> {
 
   LogicalResult matchAndRewrite(SubViewOp subViewOp,
                                 PatternRewriter &rewriter) const override {
-    if (!isTrivialSubViewOp(subViewOp))
+    if (!isTrivialSubViewOp(rewriter, subViewOp))
       return failure();
----------------
matthias-springer wrote:

I wouldn't use `ValueBoundsOpInterface` during canonicalization. It is quite expensive and should probably removed from the canonicalization patterns that use it today.

I don't have a good solution. Maybe we can build a new pass that folds various index computations and view-like ops based on a single `ValueBoundsOpInterface` analysis. Maybe that pass could then also replace the "reify ranked shaped ..." pass. Not sure if it's a good idea, just trying to think of a solution...




https://github.com/llvm/llvm-project/pull/140619