[Mlir-commits] [mlir] [mlir][math] Add FP software implementation lowering pass: math-to-apfloat (PR #171221)

[Maksim Levental]

================
@@ -32,11 +32,15 @@ namespace {
 template <typename SourceOp, typename TargetOp>
 using ConvertFastMath = arith::AttrConvertFastMathToLLVM<SourceOp, TargetOp>;
 
-template <typename SourceOp, typename TargetOp>
+template <typename SourceOp, typename TargetOp,
+          bool FailOnUnsupportedFP = false>
 using ConvertFMFMathToLLVMPattern =
-    VectorConvertToLLVMPattern<SourceOp, TargetOp, ConvertFastMath>;
+    VectorConvertToLLVMPattern<SourceOp, TargetOp, ConvertFastMath,
+                               FailOnUnsupportedFP>;
 
-using AbsFOpLowering = ConvertFMFMathToLLVMPattern<math::AbsFOp, LLVM::FAbsOp>;
+using AbsFOpLowering =
+    ConvertFMFMathToLLVMPattern<math::AbsFOp, LLVM::FAbsOp,
+                                /*FailOnUnsupportedFP=*/true>;
 using CeilOpLowering = ConvertFMFMathToLLVMPattern<math::CeilOp, LLVM::FCeilOp>;
----------------
makslevental wrote:

shouldn't we just change `FailOnUnsupportedFP` to default to `true` then? other than the patterns/paths we've been adding nothing supports fp8 right?

https://github.com/llvm/llvm-project/pull/171221