Teach ValueTracking that compare-and-branch on poison is UB

[sanjoy]

Summary

This lets us optimize the interesting case of

do {
  I = PHI(I_Begin, I_Next);
  I_Next = I +nsw 1
  Use(Sext(I_Next))
} while (I_Next < N);

Note: we "already get" the above case, but the way we do so is
incorrect. Making our poison value analysis smarter about this will let
us get rid of the buggy behavior (in a later patch) without regressing
too much performance across the board.

[majnemer]

lib/Analysis/ValueTracking.cpp

3413–3416

Our langref states:

An instruction control-depends on a terminator instruction if the terminator instruction has multiple successors and the instruction is always executed when control transfers to one of the successors, and may not be executed when control is transferred to another.

The code, as written, sounds like it will do the wrong thing if the branch instruction branches to the same basic block regardless of the condition operand. Maybe this is just a bug in the langref? Equally possible is that I misunderstand the behavior that this function is trying to suss out.

[sanjoy]

lib/Analysis/ValueTracking.cpp

3413–3416

You're right; as written this code is not correct and it can't easily be fixed either. Going by the langref, this code fragment (the motivation for this change)

loop:
  %iv = phi(0, %iv.inc)
  %iv.inc = add nsw %iv, 1
  %cmp = %iv.inc s< %L
  br i1 %cmp, label %leave, label %loop
leave:
  ...

is well defined, even when %iv.inc is poison.

/sadtrombone

[sanjoy]

As @majnemer pointed out, this change is wrong.