cc'ing @​reames

This was caused/exposed by:
https://reviews.llvm.org/rG412eb07edd4e83afeb1e727e91213e2cbd771f29
https://reviews.llvm.org/D111836

define i32 @​#52276 (i32 %a, i32* %p) {
entry:
%cmp = icmp sgt i32 %a, 0
%conv = zext i1 %cmp to i32
%neg = xor i32 %a, -1
%cmp1 = icmp slt i32 %conv, %neg
br i1 %cmp1, label %ph, label %exit

ph:
%b = load i32, i32* %p, align 4
br label %loop

loop:
%inc1 = phi i32 [ %b, %ph ], [ %inc, %loop ]
%inc = add nsw i32 %inc1, 1
br i1 %cmp1, label %loop, label %crit_edge, !llvm.loop !​0

crit_edge:
%inc2 = phi i32 [ %inc, %loop ]
store i32 %inc2, i32* %p, align 4
br label %exit

exit:
ret i32 0
}

!​0 = distinct !{#0, !​1}
!​1 = !{!"llvm.loop.mustprogress"}

Agreed, definitely a regression from my patch. I'd missed a hasOneUse check somewhere in the process of splitting and rebasing this. Will have a fix landed shortly, sorry for the breakage.

p.s. The bug is actually introduced before the patch fingered. That's simply the first one which exposes this particular use case.

Should be fixed by:

commit f82cf61 (HEAD -> main, origin/main)
Author: Philip Reames listmail@philipreames.com
Date: Mon Oct 25 09:25:00 2021 -0700

[indvars] Fix pr52276 (missing one use check)

The recently added logic to canonicalize exit conditions to unsigned relies on facts which hold about the use (i.e. exit test).  Applying this blindly to the icmp is not legal, as there may be another use which never reaches the exit.  Restrict ourselves to case where we have a single use.

Please reopen if needed. I confirmed the reduced test case was fixed, but not the original C example.