[Mlir-commits] [mlir] [mlir][amdgpu] Continue lowering make_tdm_descriptor. (PR #171498)

Erick Ochoa Lopez llvmlistbot at llvm.org
Wed Dec 10 11:08:36 PST 2025


================
@@ -2442,13 +2439,16 @@ struct AMDGPUMakeDmaDescriptorLowering
   Value setPadInterval(MakeDmaDescriptorOp op, OpAdaptor adaptor,
                        ConversionPatternRewriter &rewriter, Location loc,
                        Value sgpr0, ArrayRef<Value> consts) const {
-    bool pad_enable = op.getPadAmount() != nullptr;
-    if (!pad_enable)
+    if (!op.getPadAmount())
       return sgpr0;
 
+    // pre-condition: padInterval can be a power of two between 2 and 256.
+    // TODO: Validation if the value breaks the pre-condition.
+    // If the pre-condition fails, there is a possibility of
+    // affecting the higher bits. In a following PR add a flag
+    // that instruments conditions that need to be checked at runtime.
----------------
amd-eochoalo wrote:

For all comments so far: https://github.com/llvm/llvm-project/pull/171498/commits/c9f3558235e8194734fe4d8d985233060d7c0ce1

https://github.com/llvm/llvm-project/pull/171498


More information about the Mlir-commits mailing list