[InstCombine] Missed optimization : fold binop(select cond, C1, C2, zext(cond))

[XChy]

Alive2 proof: https://alive2.llvm.org/ce/z/-f5J6T

Motivating example

define i32 @src(i1 %c) {
entry:
  %cond_zext = zext i1 %c to i32
  %s = select i1 %c, i32 67108863, i32 0
  %or = or i32 %s, %cond_zext
  ret i32 %or
}

can be folded to:

define i32 @tgt(i1 %c) {
entry:
  %s = select i1 %c, i32 67108863, i32 0
  ret i32 %s
}

Real-world motivation

This snippet of IR is derived from qemu/system/physmem.c@phys_page_set_level (after O3 pipeline).
The example above is a reduced version. If you're interested in the original suboptimal IR and optimal IR, please email me.

Let me know if you can confirm that it's an optimization opportunity, thanks.

[XChy]

Generalized proof: https://alive2.llvm.org/ce/z/isMxFR

[gwastar]

I started working on this. Should have a patch soon.

[gwastar]

This fold already happens for add, sub, and mul and was introduced in commit f12a556. It is performed in foldBinOpOfSelectAndCastOfSelectCondition. This function seems to have two issues: It doesn't preserve flags like nuw, nsw, or disjoint and it allows multi-use selects which I don't think is profitable (see https://godbolt.org/z/vo8GY6xsa and https://alive2.llvm.org/ce/z/AtLvvW). I guess I should fix that function first so that I can use it for this.

[XChy]

@gwastar, yes, I looked into this function. One-use constraint is needed for the select at least.
You could first file a patch to add some constraints on it, and then file the second patch for this issue.