[all-commits] [llvm/llvm-project] d067b8: [SPIRV] Add missing OpenCL atomic_fetch_min/max bu...

[Paulius Velesko via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: d067b836b6c8ddab8abc9441b771aa793b543916
      https://github.com/llvm/llvm-project/commit/d067b836b6c8ddab8abc9441b771aa793b543916
  Author: Paulius Velesko <pvelesko at pglc.io>
  Date:   2026-04-29 (Wed, 29 Apr 2026)

  Changed paths:
    M llvm/lib/Target/SPIRV/SPIRVBuiltins.cpp
    M llvm/lib/Target/SPIRV/SPIRVBuiltins.td
    A llvm/test/CodeGen/SPIRV/transcoding/OpenCL/atomic_fetch_min_max.ll

  Log Message:
  -----------
  [SPIRV] Add missing OpenCL atomic_fetch_min/max builtin mappings (#190443)

## Summary

The SPIR-V backend maps OpenCL `atomic_fetch_add`/`sub`/`or`/`xor`/`and`
(and their `_explicit` variants) to SPIR-V atomic opcodes, but was
missing support for `atomic_fetch_min`/`atomic_fetch_max`, their
`_explicit` variants, and the legacy `atom_min`/`atom_max` builtins.
This caused OpenCL programs using these atomics to emit unresolved
function calls instead of the correct
`OpAtomicSMin`/`OpAtomicSMax`/`OpAtomicUMin`/`OpAtomicUMax`
instructions.

### Approach

Unlike add/sub/or/xor/and (which are sign-agnostic), min/max require
distinct signed vs unsigned SPIR-V opcodes. Rather than inspecting the
`OpTypeInt` signedness bit at runtime (which is always 0 in this
backend), this patch uses the existing prefix-based builtin lookup
mechanism in `lookupBuiltin`: the itanium demangler exposes the argument
type (`int` vs `unsigned int`), and the lookup adds an `s_` or `u_`
prefix accordingly, matching the correct `.td` entry.

### Changes

- **`SPIRVBuiltins.td`**: Add 12 prefixed `DemangledNativeBuiltin`
entries (`s_`/`u_` variants for `atomic_fetch_min`, `atomic_fetch_max`,
`atomic_fetch_min_explicit`, `atomic_fetch_max_explicit`, `atom_min`,
`atom_max`)
- **`SPIRVBuiltins.cpp`**: Add
`OpAtomicSMin`/`OpAtomicSMax`/`OpAtomicUMin`/`OpAtomicUMax` cases to the
`generateAtomicInst` switch to route these opcodes to
`buildAtomicRMWInst` (also fixes the pre-existing
`__spirv_Atomic{S,U}{Min,Max}` builtins that were defined in `.td` but
not handled in the switch)
- **New test**: `atomic_fetch_min_max.ll` covering all 6 builtins for
both signed and unsigned `i32` types



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