[all-commits] [llvm/llvm-project] 478f9d: [SelectionDAG] Widen vector math libcalls when no ...

[Mattéo Rizza Murgier via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 478f9da04672855854811d5bca7f1fd05c52ed90
      https://github.com/llvm/llvm-project/commit/478f9da04672855854811d5bca7f1fd05c52ed90
  Author: Mattéo Rizza Murgier <matteo.rizza-murgier at sipearl.com>
  Date:   2026-09-08 (Tue, 08 Sep 2026)

  Changed paths:
    M llvm/lib/CodeGen/SelectionDAG/LegalizeVectorOps.cpp
    M llvm/test/CodeGen/AArch64/fp-veclib-expansion.ll

  Log Message:
  -----------
  [SelectionDAG] Widen vector math libcalls when no routine is available (#218948)

`tryExpandVecMathCall` currently only checks for a vector math routine
matching the node's exact vector type, unrolling when none is found.
This is suboptimal, and can lead to crashes for scalable types (which
cannot be unrolled).

This PR implements widening to first check if a routine with a wider vec
type exists before falling back to unrolling.

Example:
```
; llc -mtriple=aarch64 -mattr=+sve -vector-library=sleefgnuabi crash.ll
define <vscale x 2 x float> @frem_nxv2f32(<vscale x 2 x float> %a, <vscale x 2 x float> %b) {
  %res = frem <vscale x 2 x float> %a, %b
  ret <vscale x 2 x float> %res
}
```

NOTE: When the routines take a mask (e.g. if they're scalable variants),
only relevant lanes are activated. Vectors are padded with copies of the
operand to avoid running on poison.



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