[all-commits] [llvm/llvm-project] 058398: [NVPTX] Constant fold blockDim when reqntid is spe...

[Chengjun via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 058398c4ceaf880b82a8a85a087f362fd8b48d75
      https://github.com/llvm/llvm-project/commit/058398c4ceaf880b82a8a85a087f362fd8b48d75
  Author: Chengjun <chengjunp at Nvidia.com>
  Date:   2026-04-17 (Fri, 17 Apr 2026)

  Changed paths:
    M llvm/lib/Target/NVPTX/NVVMIntrRange.cpp
    M llvm/test/CodeGen/NVPTX/intr-range.ll
    A llvm/test/CodeGen/NVPTX/reqntid-const-fold.ll

  Log Message:
  -----------
  [NVPTX] Constant fold blockDim when reqntid is specified (#191575)

Currently, NVPTX cannot fold the `ntid.x/y/z` intrinsic calls into const
values when `reqntid` is specified, which prevents the code from further
optimization.
Therefore, in this change, we extend the `NVVMIntrRange` pass to:
- Tighten `ntid.x/y/z` intrinsic calls to one value range, which can be
const folded in later InstCombine pass
- Tighten `tid.x/y/z` range attributes to use per-dimension reqntid
bounds
- When .reqntid exceeds hardware limits, garbage-in/garbage-out



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