[Mlir-commits] [mlir] [mlir][ArithToLLVM] Do not lower sitofp to unsupported float types (PR #230975)

[llvmlistbot at llvm.org]

llvmorg-github-actions[bot] wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-mlir

Author: SomnusGG (SomnusGG)

<details>
<summary>Changes</summary>

The LLVM type converter maps float types with no LLVM equivalent (for example
f8E4M3FN, f8E5M2 and f8E8M0FNU) to an integer of the same width. Unlike
`UIToFPOpLowering`, `FPToSIOpLowering` and `FPToUIOpLowering`,
`SIToFPOpLowering` did not set `FailOnUnsupportedFP`. As a result, it
produced an invalid op that the verifier rejects:

  'llvm.sitofp' op result #<!-- -->0 must be floating point LLVM type or LLVM
  dialect-compatible vector of floating point LLVM type, but got 'i8'

Reproducer (`mlir-opt --convert-arith-to-llvm`, or `--convert-to-llvm`):

  %0 = arith.sitofp %arg0 : i32 to f8E4M3FN

Set `FailOnUnsupportedFP` so that the op stays unconverted, as the other
int/float conversions already do. Users can emulate the op first with
`arith-emulate-unsupported-floats`.



🤖 Generated with [Claude Code](https://claude.com/claude-code)

---
Full diff: https://github.com/llvm/llvm-project/pull/230975.diff


2 Files Affected:

- (modified) mlir/lib/Conversion/ArithToLLVM/ArithToLLVM.cpp (+3-1) 
- (modified) mlir/test/Conversion/ArithToLLVM/arith-to-llvm.mlir (+19) 


``````````diff
diff --git a/mlir/lib/Conversion/ArithToLLVM/ArithToLLVM.cpp b/mlir/lib/Conversion/ArithToLLVM/ArithToLLVM.cpp
index bac19144d6ea7..e61374cfcfda9 100644
--- a/mlir/lib/Conversion/ArithToLLVM/ArithToLLVM.cpp
+++ b/mlir/lib/Conversion/ArithToLLVM/ArithToLLVM.cpp
@@ -198,7 +198,9 @@ using ShRSIOpLowering =
 using ShRUIOpLowering =
     VectorConvertToLLVMPattern<arith::ShRUIOp, LLVM::LShrOp>;
 using SIToFPOpLowering =
-    VectorConvertToLLVMPattern<arith::SIToFPOp, LLVM::SIToFPOp>;
+    VectorConvertToLLVMPattern<arith::SIToFPOp, LLVM::SIToFPOp,
+                               AttrConvertPassThrough,
+                               /*FailOnUnsupportedFP=*/true>;
 using SubFOpLowering =
     ConstrainedVectorConvertToLLVMPattern<arith::SubFOp, LLVM::FSubOp,
                                           /*HasRoundingMode=*/false,
diff --git a/mlir/test/Conversion/ArithToLLVM/arith-to-llvm.mlir b/mlir/test/Conversion/ArithToLLVM/arith-to-llvm.mlir
index 372561d9653c8..aea4be2195274 100644
--- a/mlir/test/Conversion/ArithToLLVM/arith-to-llvm.mlir
+++ b/mlir/test/Conversion/ArithToLLVM/arith-to-llvm.mlir
@@ -1008,6 +1008,25 @@ func.func @unsupported_fp_type(%arg0: f4E2M1FN, %arg1: vector<4xf4E2M1FN>, %arg2
 
 // -----
 
+// Integer to float conversions with a result type that has no LLVM equivalent
+// are not converted. sitofp used to be lowered to an invalid `llvm.sitofp` with
+// an integer result.
+
+// CHECK-LABEL: func @unsupported_fp_type_int_to_fp
+//       CHECK:   arith.sitofp {{.*}} : i32 to f8E4M3FN
+//       CHECK:   arith.sitofp {{.*}} : vector<4xi32> to vector<4xf8E5M2>
+//       CHECK:   arith.sitofp {{.*}} : vector<2x4xi32> to vector<2x4xf8E8M0FNU>
+//       CHECK:   arith.uitofp {{.*}} : i32 to f8E4M3FN
+func.func @unsupported_fp_type_int_to_fp(%arg0: i32, %arg1: vector<4xi32>, %arg2: vector<2x4xi32>) -> (f8E4M3FN, vector<4xf8E5M2>, vector<2x4xf8E8M0FNU>, f8E4M3FN) {
+  %0 = arith.sitofp %arg0 : i32 to f8E4M3FN
+  %1 = arith.sitofp %arg1 : vector<4xi32> to vector<4xf8E5M2>
+  %2 = arith.sitofp %arg2 : vector<2x4xi32> to vector<2x4xf8E8M0FNU>
+  %3 = arith.uitofp %arg0 : i32 to f8E4M3FN
+  return %0, %1, %2, %3 : f8E4M3FN, vector<4xf8E5M2>, vector<2x4xf8E8M0FNU>, f8E4M3FN
+}
+
+// -----
+
 // CHECK-LABEL: func @supported_fp_type
 //       CHECK:   llvm.fadd {{.*}} : f32
 //       CHECK:   llvm.fadd {{.*}} : vector<4xf32>

``````````

</details>


https://github.com/llvm/llvm-project/pull/230975