[all-commits] [llvm/llvm-project] 682f0b: [InstCombine] use select-of-constants with set/cle...

[RotateRight via All-commits]

  Branch: refs/heads/master
  Home:   https://github.com/llvm/llvm-project
  Commit: 682f0b366be266b9473507f12be21021e3e4912f
      https://github.com/llvm/llvm-project/commit/682f0b366be266b9473507f12be21021e3e4912f
  Author: Sanjay Patel <spatel at rotateright.com>
  Date:   2020-05-03 (Sun, 03 May 2020)

  Changed paths:
    M llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp
    M llvm/test/Transforms/InstCombine/cast.ll
    M llvm/test/Transforms/InstCombine/select-with-bitwise-ops.ll

  Log Message:
  -----------
  [InstCombine] use select-of-constants with set/clear bit mask patterns

Cond ? (X & ~C) : (X | C) --> (X & ~C) | (Cond ? 0 : C)
Cond ? (X | C) : (X & ~C) --> (X & ~C) | (Cond ? C : 0)

The select-of-constants form results in better codegen.
There's an existing test diff that shows a transform that
results in an extra IR instruction, but that's an existing
problem.

This is motivated by code seen in LLVM itself - see PR37581:
https://bugs.llvm.org/show_bug.cgi?id=37581

define i8 @src(i8 %x, i8 %C, i1 %b)  {
  %notC = xor i8 %C, -1
  %and = and i8 %x, %notC
  %or = or i8 %x, %C
  %cond = select i1 %b, i8 %or, i8 %and
  ret i8 %cond
}

define i8 @tgt(i8 %x, i8 %C, i1 %b)  {
  %notC = xor i8 %C, -1
  %and = and i8 %x, %notC
  %mul = select i1 %b, i8 %C, i8 0
  %or = or i8 %mul, %and
  ret i8 %or
}

http://volta.cs.utah.edu:8080/z/Vt2WVm

Differential Revision: https://reviews.llvm.org/D78880