Optimization for fold compare_GT(maximum(a,b), a) into compare_GT(b,a). Is it needed?

[apivovarov]

I'm working on a compiler frontend which is based on OpenXLA
Internally we have an optimization to fold compare_GT(maximum(a,b), a) into compare_GT(b,a)
The optimization is added to algebraic_simplifier.cc::HandleCompare().

Full list of similar optimizations:

GT(min(a,b), _) -> FALSE
GT(max(a,b), a) -> GT(b,a)
GT(max(a,b), b) -> GT(a,b)

GT(a, min(a,b)) -> GT(a,b)
GT(b, min(a,b)) -> GT(b,a)
GT(_, max(a,b)) -> FALSE

GE(min(a,b), a) -> GE(b,a)
GE(min(a,b), b) -> GE(a,b)
GE(max(a,b), _) -> TRUE

GE(_, min(a,b)) -> TRUE
GE(a, max(a,b)) -> GE(a,b)
GE(b, max(a,b)) -> GE(b,a)

LT(min(a,b), a) -> LT(b,a)
LT(min(a,b), b) -> LT(a,b)
LT(max(a,b), _) -> FALSE

LT(_, min(a,b)) -> FALSE
LT(a, max(a,b)) -> LT(a,b)
LT(b, max(a,b)) -> LT(b,a)

LE(a, min(a,b)) -> LE(a,b)
LE(b, min(a,b)) -> LE(b,a)
LE(_, max(a,b)) -> TRUE

LE(min(a,b), _) -> TRUE
LE(max(a,b), a) -> LE(b,a)
LE(max(a,b), b) -> LE(a,b)

How useful would it be to contribute these optimizations to OpenXLA main?

We tested GT(max(a,b), a) optimization on Resnet50 model internally
Overall, we observed the following benefits of adding this optimization:

• VmRSS usage is 14% less
• Number of instructions - 13% less
• Memory locations - 22% less

@majnemer @akuegel @lipracer @jurahul What you think?

[lipracer]

Thank you, I think this is a universal simplification.

[akuegel]

As long as it preserves NaN propagation, we are happy to accept such optimizations for AlgebraicSimplifier. I suggest to not add all at once, but in different PRs, as it is easier to review that way.

[apivovarov]

Opened PR: Simplify Gt(Max(a,b), a) -> Gt(b,a) #9201