[all-commits] [llvm/llvm-project] 816165: HIP: Directly call fabs builtins

[Matt Arsenault via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 81616561c1b376af85365c9eaf94d49ad184c623
      https://github.com/llvm/llvm-project/commit/81616561c1b376af85365c9eaf94d49ad184c623
  Author: Matt Arsenault <Matthew.Arsenault at amd.com>
  Date:   2022-12-19 (Mon, 19 Dec 2022)

  Changed paths:
    M clang/lib/Headers/__clang_hip_math.h
    M clang/test/Headers/__clang_hip_math.hip
    M clang/test/Headers/amdgcn_openmp_device_math.c

  Log Message:
  -----------
  HIP: Directly call fabs builtins

All the attribute numbering changes are due to the ocml function being
assumed convergent, and the direct call is not.

This begins a series of patches to avoid using the thinnest
ocml wrappers around llvm intrinsics. These wrapper functions
aren't buying anything and add complexity. As we do not propagate
fast math flags into the library on linking, and cannot generally
safely do so, we're losing information by calling into them. Directly
call the builtins so we get the fast math flags set appropriate for the
translation unit naturally.