[all-commits] [llvm/llvm-project] ba61e4: [AMDGPU] Implement int div/rem by fp recip more ef...

[LU-JOHN via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: ba61e4d1c110199ef07d7ef4c34c23bcc64b4b54
      https://github.com/llvm/llvm-project/commit/ba61e4d1c110199ef07d7ef4c34c23bcc64b4b54
  Author: LU-JOHN <John.Lu at amd.com>
  Date:   2026-07-13 (Mon, 13 Jul 2026)

  Changed paths:
    M llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp
    M llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-idiv.ll
    M llvm/test/CodeGen/AMDGPU/divrem24-assume.ll
    M llvm/test/CodeGen/AMDGPU/idiv-licm.ll
    M llvm/test/CodeGen/AMDGPU/permute_i8.ll
    M llvm/test/CodeGen/AMDGPU/sdiv.ll
    M llvm/test/CodeGen/AMDGPU/sdiv64.ll
    M llvm/test/CodeGen/AMDGPU/srem64.ll
    M llvm/test/CodeGen/AMDGPU/udiv.ll
    M llvm/test/CodeGen/AMDGPU/udivrem24.ll

  Log Message:
  -----------
  [AMDGPU] Implement int div/rem by fp recip more efficiently (#204950)

Integer division q = a/b can be implemented by fp reciprocal with:
`fq = fa * recip(fb)`
fq is truncated to produce q. Due to fp rounding and reciprocal accuracy
issues fq can be too small and truncation can produce a value too small
by one.

If abs(a)<=0x400000, this underestimate can be guarded more efficiently
by calculating:
`fq=fa+1ulp/b`
If abs(a)<=0x400000, adding 1 ulp will increase a by at most 0.5, so the
calculated q will be the same. Adding 1ulp can be done with one integer
add.

Incrementing `fa` by 1 ulp was exhaustively tested for all dividend and
divisor combinations with values between [-0x400000,0x3FFFFF] with no
errors.

---------

Signed-off-by: John Lu <John.Lu at amd.com>



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