[Mlir-commits] [mlir] [mlir][AMDGPU] Add canonicalizer for folding casts into gather_to_lds (PR #150503)

[Quinn Dawkins]

================
@@ -546,6 +550,58 @@ LogicalResult GatherToLDSOp::verify() {
   return success();
 }
 
+namespace {
+/// If the source/target of a GatherToLDSOp is a CastOp that only removes static
+/// information or changes layout, the cast can be skipped.
+struct FoldGatherToLDSOfCast final : OpRewritePattern<GatherToLDSOp> {
+  using OpRewritePattern<GatherToLDSOp>::OpRewritePattern;
+
+  LogicalResult matchAndRewrite(GatherToLDSOp gatherOp,
+                                PatternRewriter &rewriter) const override {
+    bool modified = false;
+
+    // Check source.
+    if (auto castOp = gatherOp.getSrc().getDefiningOp<memref::CastOp>()) {
+      auto fromType = dyn_cast<MemRefType>(castOp.getSource().getType());
+      auto toType = dyn_cast<MemRefType>(castOp.getSource().getType());
----------------
qedawkins wrote:

Good catch. The element type check is actually redundant with `CastOp::canFoldIntoConsumerOp` so I just dropped these lines altogether.

https://github.com/llvm/llvm-project/pull/150503