[llvm-dev] Thoughts on speculatively creating IR and throwing it away if not needed

[Florian Hahn via llvm-dev]

Hi Jonas,

> On Jan 18, 2021, at 18:01, Jonas Paulsson via llvm-dev <llvm-dev at lists.llvm.org> wrote:
> 
> > Given that this might push the envelop of the pattern a bit, I’d like to check if there are any concerns with using this pattern more widely.
> 
> Is the idea here to generate instructions which are collectively cheaper than their individual costs (due to reuse/combinations)? Or is it a matter of not really knowing what instructions will be generated and no?


This is a great way to frame it.

The motivation is a bit of both. SCEV expansion will try to re-use existing expressions (including ones inserted by earlier expansion steps), so the overall cost will be cheaper collectively than when looking at the cost of each expression individually. There potentially are a number of SCEV expansion happening (e.g. SCEV assumptions, one for each runtime check bound).

And without traversing the full SCEV expression, we cannot tell which instructions will be generated. So to estimate the cost, we would have to traverse the whole expression, just to traverse it again when we generate the actual code.

And we are also able to directly & accurately compare the cost of the generated instructions to the computed loop cost, using exactly the same metric.

Cheers,
Florian