Coverage instrumentation sometimes thinks the negation operator isn't executable

[Zalathar]

#111752 introduced this behaviour in coverage reports:

   LL|       |    if
   LL|       |        !
   LL|      1|        is_true
   LL|      0|    {

Notice that the line with ! no longer has a coverage count, even though it should have the same count as is_true.

---

This is caused by the fact that MIR building specifically handles if ! by flipping the then/else arms, so the original condition expression (containing !) does not have its span represented by any statement in MIR.

[Zalathar]

(This issue is based on my original notes at #111752 (comment).)

[Zalathar]

This seems to be the result of if !cond { a } else { b } being lowered as though it were if cond { b } else { a }.

When that happens, the ! operation no longer exists as a MIR node, so InstrumentCoverage can’t see it normally, and doesn’t have special-case handling to look back at HIR and see that this has happened.

For now, it’s probably not worth trying to add another special case, since the difference is hardly noticeable most of the time. It might be worth revisiting this if we ever implement deeper coverage integration into HIR-to-MIR lowering.

[Zalathar]

@rustbot label -needs-triage

[Zalathar]

@rustbot label +T-compiler

[Zalathar]

As I continue to investigate branch coverage, this has become more of a concern.

When the user writes if !cond, I want the branch region to cover all of !cond (not just cond), and I want its true/false arms to reflect the value of !cond (rather than the value of cond).