[all-commits] [llvm/llvm-project] 68f9d8: [MLIR][Parser] Fix precision loss when parsing f80...

Chibuoyim (Wilson) Ogbonna via All-commits all-commits at lists.llvm.org
Thu Sep 10 23:51:02 PDT 2026


  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 68f9d8d45cad616e223ffe460404edd86dd972ca
      https://github.com/llvm/llvm-project/commit/68f9d8d45cad616e223ffe460404edd86dd972ca
  Author: Chibuoyim (Wilson) Ogbonna <chibuoyim.faith.ogbonna at huawei.com>
  Date:   2026-09-11 (Fri, 11 Sep 2026)

  Changed paths:
    M mlir/lib/AsmParser/AsmParserImpl.h
    M mlir/lib/AsmParser/AttributeParser.cpp
    M mlir/lib/AsmParser/Parser.cpp
    M mlir/lib/AsmParser/Token.cpp
    M mlir/lib/AsmParser/Token.h
    M mlir/test/Dialect/Builtin/Bytecode/attrs.mlir
    M mlir/test/Dialect/Builtin/Bytecode/builtin_fixed.mlir
    M mlir/test/Dialect/Builtin/Bytecode/builtin_fixed_0.mlirbc
    M mlir/test/IR/custom-float-attr-roundtrip.mlir
    M mlir/test/IR/parser.mlir

  Log Message:
  -----------
  [MLIR][Parser] Fix precision loss when parsing f80 and f128 literals (#220267)

While analyzing this CIR issue
[216422](https://github.com/llvm/llvm-project/issues/216422), I traced
the issue to the MLIR asm parser.

For the following case:
```
%0 = arith.constant 1.0E+309 : f80
```
using `mlir-opt repro.mlir`
we get (infinity):
```
module {
  %cst = arith.constant 0x7FFF8000000000000000 : f80
}
```
instead of
```
module {
  %cst = arith.constant 9.99999999999999999986E+308 : f80
}
```

Currently the parser casts floats to a double before converting to the
requested semantics, as noted in the
[comments](https://github.com/llvm/llvm-project/blob/bb3b509138f4a34cb3d5c659955ccc16745ae82e/mlir/lib/AsmParser/AsmParserImpl.h#L286).

Not entirely sure why this was the default, anyway this patch instead
parses the float with the given semantics directly, so we don't lose
precision.

I used Claude to generate tests and reproducers for this patch. I've
also updated a few existing lit tests to show the improved precision
this patch introduces.

**Note:** This fixes the CIR issue I mentioned above.



To unsubscribe from these emails, change your notification settings at https://github.com/llvm/llvm-project/settings/notifications


More information about the All-commits mailing list