[all-commits] [llvm/llvm-project] 185ff0: [X86] Fix i64-to-bf16 conversion crash on 32-bit t...

Tim Besard via All-commits all-commits at lists.llvm.org
Fri Sep 11 08:05:17 PDT 2026


  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 185ff0c972cbdf0f221fbf1f45b8554810da79eb
      https://github.com/llvm/llvm-project/commit/185ff0c972cbdf0f221fbf1f45b8554810da79eb
  Author: Tim Besard <tim.besard at gmail.com>
  Date:   2026-09-11 (Fri, 11 Sep 2026)

  Changed paths:
    M llvm/lib/Target/X86/X86ISelLowering.cpp
    A llvm/test/CodeGen/X86/bfloat-int64-cvt-i686.ll

  Log Message:
  -----------
  [X86] Fix i64-to-bf16 conversion crash on 32-bit targets (#222934)

Converting a loaded `i64` to `bfloat` can crash instruction selection on
i686:

```llvm
define bfloat @convert(ptr %p) {
  %value = load i64, ptr %p
  %result = sitofp i64 %value to bfloat
  ret bfloat %result
}
```

With `llc -mtriple=i686-linux-gnu`, this hits:

```text
Do not know how to custom type legalize this operation!
```

The X86 DAG combine folds the load and conversion into an x87 `FILD`
node
before type legalization. It already excludes `f16` and `f128`, but was
missing
`bf16`, leaving the legalizer with an unsupported result type.

Add `bf16` to that exclusion. This lets the existing legalization
promote the
conversion through `f32` and then round to bf16. The change also covers
the
constrained signed conversion, which uses the same combine.

Assisted by: Fable 5.1, Astra 6



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