Regression: Missed fold for x * 0

[jeremy-rifkin]

Bugzilla Link	51286
Resolution	FIXED
Resolved on	Sep 15, 2021 06:10
Version	trunk
OS	Windows NT
Blocks	#49274
CC	@nerh,@RKSimon,@rotateright
Fixed by commit(s)	f5d8952

Extended Description

Clang trunk is not generating optimal code for a (very poor) multiplication routine https://godbolt.org/z/GrjMhsTcv.

int mul(int x, int y) {
    int p = 0;
    while(x--) p += y;
    return p;
}

Should compile to:

mul(int, int):
        mov     eax, edi
        imul    eax, esi
        ret

And it does in clang 12 but not in clang trunk.

If I'm not mistaken it appears a select i1 (icmp eq i32 %x, 0), i32 0, i32 (mul i32 %x, %y) -> mul i32 %x, %y fold is not firing.

[jeremy-rifkin]

assigned to @nerh

[ghost]

Aforementioned code snipped is no longer folded into single multiplication instruction after changes in "simplifyWithOpReplaced".

If there are no objections I'd like to take this issue fix the optimization.

[RKSimon]

SGTM

Current IR:

define i32 @mul(i32 %0, i32 %1) {
  %3 = icmp eq i32 %0, 0
  %4 = mul i32 %1, %0
  %5 = select i1 %3, i32 0, i32 %4
  ret i32 %5
}

To get rid of that select we will need freeze: https://alive2.llvm.org/ce/z/maX0k-

[ghost]

Candidate patch: https://reviews.llvm.org/D108408

[rotateright]

We have the expected codegen again (just a multiply, no cmov) after:
https://reviews.llvm.org/rGf5d89523567b