[Mlir-commits] [llvm] [mlir] [MLIR][AMDGPU] Adding dynamic size check to avoid subword buffer load (PR #135014)

[Krzysztof Drewniak]

================
@@ -117,20 +121,85 @@ struct TransferReadLowering final : OpRewritePattern<vector::TransferReadOp> {
     }
 
     Location loc = readOp.getLoc();
-    Value fill = rewriter.create<vector::SplatOp>(loc, unbroadcastedVectorType,
-                                                  readOp.getPadding());
-    Value load = rewriter.create<vector::LoadOp>(
-        loc, unbroadcastedVectorType, readOp.getSource(), readOp.getIndices());
-    Value res = rewriter.create<arith::SelectOp>(loc, unbroadcastedVectorType,
-                                                 readOp.getMask(), load, fill);
-
-    // Insert a broadcasting op if required.
-    if (requiresBroadcasting) {
-      res = rewriter.create<vector::BroadcastOp>(loc, readOp.getVectorType(),
-                                                 res);
+    Value src = readOp.getSource();
+    MemRefType memRefType = cast<MemRefType>(src.getType());
+    ArrayRef<int64_t> shape = memRefType.getShape();
+
+    Value linearIndex = rewriter.create<arith::ConstantIndexOp>(loc, 0);
+    Value one = rewriter.create<arith::ConstantIndexOp>(loc, 1);
+    Value stride = one;
+
+    // Compute the linear index by linearIndex += indices[i] * stride
+    for (int i = shape.size() - 1; i >= 0; --i) {
----------------
krzysz00 wrote:

I think you want `extract_strided_metadata` here and `max_d (strides[d] * sizes[d])` unless you're assuming an identity layout, which I don't think you are

https://github.com/llvm/llvm-project/pull/135014