Missed optimization: (int)x + 1 == (int)y - 1

[Tilka]

Bugzilla Link	32814
Version	trunk
OS	All
CC	@LebedevRI,@RKSimon,@rotateright

Extended Description

int(x) + 1 == int(y) - 1

is not being optimized to

int(x) + 2 == int(y)

Proof: http://rise4fun.com/Alive/vaC

[llvmbot]

Likely a missing instcombine (need to pay some attention to wrapping/overflow)

[rotateright]

Equality should not require any special wrapping considerations...which might be the bug because we get this optimization if the constants have the same sign.

[llvmbot]

I'm a litte nervous about the possibility of introducing an overflow if x = INT_MAX - 1, can you elaborate why it's not a problem?

[llvmbot]

In order words, I'm under the impression you need to drop nsw in that case.

[Tilka]

Yes, see the proof.

[rotateright]

Right - I think the bug (of being too conservative) is in InstCombiner::foldICmpBinOp():

// if C1 has greater magnitude than C2:
// icmp (X + C1), (Y + C2) -> icmp (X + C3), Y
// s.t. C3 = C1 - C2
//
// if C2 has greater magnitude than C1:
// icmp (X + C1), (Y + C2) -> icmp X, (Y + C3)
// s.t. C3 = C2 - C1

...

    if (AP1.isNegative() == AP2.isNegative()) {

[rotateright]

Actually, that block of code is going out of its way to generate 'nsw add'...so it handled a special case and ignored the general transform. And of course it fails to work with vectors.

[RKSimon]

Current Codegen: https://gcc.godbolt.org/z/iDLR5E

Still as bad as ever

[RKSimon]

And of course it fails to work with vectors.

Vector support: https://reviews.llvm.org/D121358

[RKSimon]

Not sure what the next step should be: https://alive2.llvm.org/ce/z/cdodNQ

We can just drop the NSW/NUW requirements for eq/ne cases - which AFAICT is enough for this ticket.

Not sure what else we could be doing for other icmp predicates though? @rotateright @LebedevRI any thoughts?