[IndVars] Incorrect comparison folding

[nikic]

In the following test (https://llvm.godbolt.org/z/rT64Yb1K6), -indvars currently folds %cmp2 to false:

declare void @use(i1)

define void @test(ptr %arr) {
entry:
  br label %loop.header

loop.header:
  %iv = phi i64 [ %iv.inc, %loop.latch ], [ 0, %entry ]
  %prev = phi i32 [ %v, %loop.latch ], [ 0, %entry ]
  %ptr = getelementptr inbounds i32, ptr %arr, i64 %iv
  %v = load i32, ptr %ptr
  %cmp1 = icmp sgt i32 %v, 0
  br i1 %cmp1, label %if, label %loop.latch

if:
  %cmp2 = icmp slt i32 %prev, 0
  call void @use(i1 %cmp2)
  br label %loop.latch

loop.latch:
  %iv.inc = add nuw nsw i64 %iv, 1
  %cmp = icmp ult i64 %iv.inc, 16
  br i1 %cmp, label %loop.header, label %exit

exit:
  ret void
}

This is incorrect: While we have a guard that %v > 0, the value of %v in the %prev phi comes from the previous iteration, so the used context is incorrect.

[nikic]

I believe https://reviews.llvm.org/D101829 previously tried to fix this issue, but the fix is incomplete. Here the incoming phi block is loop.latch, and %v does properly dominate the latch.

[nikic]

Candidate patch: https://reviews.llvm.org/D128640