[Mlir-commits] [mlir] [MLIR][Arith] Add denormal attribute to binary/unary operations (PR #112700)

[lorenzo chelini]

================
@@ -53,22 +53,49 @@ struct ConstrainedVectorConvertToLLVMPattern
   }
 };
 
+template <typename SourceOp, typename TargetOp,
+          template <typename, typename> typename AttrConvert =
+              AttrConvertPassThrough>
+struct DenormalOpConversionToLLVMPattern
+    : public VectorConvertToLLVMPattern<SourceOp, TargetOp, AttrConvert> {
+  using VectorConvertToLLVMPattern<SourceOp, TargetOp,
+                                   AttrConvert>::VectorConvertToLLVMPattern;
+
+  LogicalResult
+  matchAndRewrite(SourceOp op, typename SourceOp::Adaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const override {
+    // TODO: Here, we need a legalization step. LLVM provides a function-level
+    // attribute for denormal; here, we need to move this information from the
+    // operation to the function, making sure all the operations in the same
+    // function are consistent.
+    if (op.getDenormalModeAttr().getValue() != arith::DenormalMode::ieee)
+      return rewriter.notifyMatchFailure(
+          op, "only ieee denormal mode is supported at the moment");
+
+    StringRef arithDenormalAttrName = SourceOp::getDenormalModeAttrName();
+    op->removeAttr(arithDenormalAttrName);
+    return VectorConvertToLLVMPattern<SourceOp, TargetOp,
+                                      AttrConvert>::matchAndRewrite(op, adaptor,
+                                                                    rewriter);
----------------
chelini wrote:

In this PR, we bail out if a non-default denormal mode is specified (default is ieee). We then remove it from the operation since we do not have a mapping to LLVM yet. The next step would be to add an LLVM attribute to model denormals and implement the legalization step mentioned in the TODOs.

https://github.com/llvm/llvm-project/pull/112700