[LLVMdev] InstCombine adds bit masks, confuses self, others

[Jakob Stoklund Olesen]

On Apr 17, 2012, at 6:35 AM, Rafael Espíndola <rafael.espindola at gmail.com> wrote:

> I also agree that the pipeline should first make the IL canonical, and
> then decide what is the best way to compute the canonical form. The
> one thing I find unfortunate is that in LLVM the second step is almost
> all in DAG and MI levels. There is no point were we have IL passes
> that known more about the target.

Note that adding the bit masks does make sense as a canonical form, it is not just a premature optimization.

Expressions like x >> 2 << 3 >> 5 << 6 get canonicalized to a single shift + a mask that way.

I think it would make sense to leave the last shl alone, so the above expression gets canonicalized to:

  ((x >> 4) & mask) << 6

DAGCombine can reduce it completely if it wants.

This also matches the code generated for the bit field access in my small2.c example:

entry:
  %and = lshr i64 %d, 2
  %div = and i64 %and, 63
  %add.ptr.sum = shl nuw nsw i64 %div, 1

/jakob