[InstCombine][WIP] Don't canonicalize select of bools if it costs an instruction

[goldstein.w.n]

Summary

Proofs for new transforms: https://alive2.llvm.org/ce/z/j9o0D3

We current canonicalize true to the TrueVal operand and false to
the FalseVal operand in selects. This makes recognizing LogicalAnd /
LogicalOr patterns easier but often costs an instruction.

[goldstein.w.n]

Rebase

[nikic]

I don't think we can do this. Logical and/or is a canonical form that lots of passes rely on, not just InstCombine. We would have to teach all passes that use LogicalAnd/LogicalOr matchers to also support the inverted forms. I don't think that would be worthwhile.

[goldstein.w.n]

In D159328#4645746, @nikic wrote:

I don't think we can do this. Logical and/or is a canonical form that lots of passes rely on, not just InstCombine. We would have to teach all passes that use LogicalAnd/LogicalOr matchers to also support the inverted forms. I don't think that would be worthwhile.

In D159328#4645746, @nikic wrote:

I don't think we can do this. Logical and/or is a canonical form that lots of passes rely on, not just InstCombine. We would have to teach all passes that use LogicalAnd/LogicalOr matchers to also support the inverted forms. I don't think that would be worthwhile.

I'm happy to drop this. It was mostly motivated by wanting to fix the regression caused by expanding isFreeToInvert.
I'll see if there is another better way to go about it.

[goldstein.w.n]

Abandoning and resubmitting on GH