Issue 243906
Summary [mlir][Linalg] convert-arith-to-apfloat invalidates named linalg.matmul regions
Labels mlir
Assignees
Reporter lin334141-cell
    ## Summary

`convert-arith-to-apfloat` rewrites `arith.mulf` and `arith.addf` inside a named `linalg.matmul` region into APFloat runtime calls, leaving the operation unverifiable:

```
error: expected add/mul op in the body
```

This is another pass-specific manifestation of the underlying named-region issue discussed in #217967, but it involves a different conversion implementation.

## Minimal Reproducer

```mlir
module {
 func.func @f(%a: tensor<2x3xf32>, %b: tensor<3x2xf32>, %c: tensor<2x2xf32>) -> tensor<2x2xf32> {
    %0 = linalg.matmul ins(%a, %b : tensor<2x3xf32>, tensor<3x2xf32>) outs(%c : tensor<2x2xf32>) -> tensor<2x2xf32>
    return %0 : tensor<2x2xf32>
  }
}
```

**Commands:**

```bash
mlir-opt minimal.mlir --verify-each -o /dev/null
# Success

mlir-opt minimal.mlir --convert-arith-to-apfloat
# error: expected add/mul op in the body
```

## Expected Behavior

The conversion should preserve the validity of the named `linalg.matmul` operation, or avoid rewriting its region until the named operation has been lowered.

## Actual Behavior

`BinaryArithOpToAPFloatConversion` replaces the region's multiplication and addition with calls to `_mlir_apfloat_multiply` and `_mlir_apfloat_add`, together with operand and result conversions.

This breaks the named Linalg operation's required multiplication/addition structure, causing verification to fail.

Lowering the named Linalg operation to loops before applying the APFloat conversion avoids the failure.

## Tested Versions

Reproduced on:

- LLVM main `c1856996d677c1b3578752e6781d0df46617fd68` (November 6, 2026)
- LLVM 24.0.0 development nightly associated with `cecb42f82e628dbdeeeee586086c2006844716aa` (November 16, 2026)

During a 24-hour MLIR fuzzing campaign, 47 generated programs triggered this APFloat verifier failure. Of these, 45 also triggered the analogous `convert-arith-to-llvm` failure.

## Related Issue

#217967 reports the same underlying named Linalg region constraint violation through `convert-arith-to-llvm`.

This report documents a second affected pass, `convert-arith-to-apfloat`, implemented in `mlir/lib/Conversion/ArithAndMathToAPFloat/ArithToAPFloat.cpp`.

Both cases share the same verifier failure and underlying root cause. The APFloat path is reported separately to provide a concrete reproducer and track this additional affected conversion.
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs

Reply via email to