[llvm-dev] CFG normalization: avoiding `br i1 false`

[Ariel Ben-Yehuda via llvm-dev]

Hi,

I was looking at Rust programs which are poorly optimized by LLVM, and
one occasional factor is that LLVM allows for long-lived `br i1 false`
instructions.

The problem is that if a propagation pass discovers the condition of a
branch, the branch itself will not be eliminated until a SimpllfyCfg
pass is reached, of which there are few in the pipeline.

One example is https://github.com/rust-lang/rust/issues/44041 - the
branch conditions are found by loop unrolling, but the `br i1 false`
instruction leaks to codegen.

Another example is https://github.com/rust-lang/rust/issues/45466.

In that case, induction variable simplification discovers that an
integer overflow check is unnecessary, but the remaining `br i1 false`
still blocks the following LoopIdiomRecognize pass and prevents the
loop from being optimized to a memset.

While it is possible to have point-fixes for both problems, this might
be widespread enough problem that a more general solution might be
better - for example, to have some sort of canonicalization that
automatically removes these branches in some situations (this requires
things such as the dominator tree to be rebuilt, so it shouldn't be
done literally everywhere, but it can be done fairly often).

What are your opinions?

Thanks,
 - Ariel