Missed optimization: bitwise OR

[zamazan4ik]

Bugzilla Link	44136
Resolution	FIXED
Resolved on	Apr 12, 2021 05:11
Version	trunk
OS	Linux
Blocks	#49274
CC	@topperc,@LebedevRI,@RKSimon,@zygoloid,@rotateright
Fixed by commit(s)	2450369, d8bc4de

Extended Description

For the following code:

bool foo(int A, int B) {
    return (A | B) == A;
}

clang(trunk) with '-O3 -march=haswell' produces:

foo(int, int): # @foo(int, int)
  or esi, edi
  cmp esi, edi
  sete al
  ret

but GCC(trunk) with '-O3 -march=haswell' produces:

foo(int, int):
  andn eax, edi, esi
  sete al
  ret

Godbolt playground: https://godbolt.org/z/MZNdWW

[LebedevRI]

https://godbolt.org/z/P5-bZ5
https://rise4fun.com/Alive/3d9

Looks like we are missing an undo middle-end canonicalization too?

[RKSimon]

Candidate X86 backend patch: https://reviews.llvm.org/D100177

[RKSimon]

X86 Backend support:

BMI: rG245036950a7a
non-BMI: rGd8bc4de3cfe0

[RKSimon]

The middle-end reverse fold would require a freeze to be valid:

https://alive2.llvm.org/ce/z/PgAUx5

define i1 @src(i8 %a, i8 %b) {
  %n = xor i8 %a, -1
  %o = and i8 %n, %b
  %c = icmp eq i8 %o, 0
  ret i1 %c
}

define i1 @tgt(i8 %a, i8 %b) {
  %aa = freeze i8 %a
  %o = or i8 %aa, %b
  %c = icmp eq i8 %o, %aa
  ret i1 %c
}

[RKSimon]

Candidate middle-end patch: https://reviews.llvm.org/D100211

[RKSimon]

Candidate middle-end patch: https://reviews.llvm.org/D100211

There's some concern about introducing freeze() for this middle-end fold, so this has been abandoned, we can maybe revisit this in the future.

Resolving as there's nothing else we're intending to do anytime soon.

[RKSimon]

mentioned in issue #49274