[llvm-dev] GlobalISel round table follow-up: multi-stage legalization

[Daniel Sanders via llvm-dev]

> On 7 Oct 2020, at 05:07, Dominik Montada <dominik.montada at hightec-rt.com> wrote:
> 
> Hi all,
> 
> first of all thanks to Matt for hosting the round table yesterday and to everyone else for their time and the valuable discussions. As promised in the chat yesterday, here is the email regarding the legalizer / artifact combiner issue we talked about. I will send a separate email regarding the regbankselect discussion.
> 
> In case someone on the list didn't attend the round table yesterday, here is quick summary:
> 
> One problem that some of us are seeing in the legalizer is that sometimes instructions get expanded before they get folded away in the combiner. IIRC Matt sees this happening with division, while I am seeing something similar with unmerge.

IIRC there were two main manifestations of it. For one it ends up at the same MIR but takes a slower-than-necessary route to get there. For the other it ends up with worse code which potentially can't be folded post-legalization.

> To my particular problem: due to the nature of the architecture I'm working with, we have pretty strict legalization rules. One example is that we only allow unmerges from 64 to 32 bit. Unmerges of 32-bit or less get lowered to bit-arithmetic. However if we would do the same for anything bigger than 64-bit, the legalization of the resulting bit arithmetic would introduce illegal unmerges again, which then cause an endless loop in the legalizer.

Are there particular >s64 cases that are the problem or is it all of them? I'd expect s128, s256, etc. to G_UNMERGE fairly simply but non-powers-of-2 are more likely to be tricky

> So our rules are carefully designed to create exactly the unmerges we can handle and let everything else get handled by the artifact combiner.
> 
> The artifact combiner usually cannot get rid of the illegal unmerges on the first go and must do some intermediary combines first. Some of those try to create other illegal unmerges but ask the legalizer whether those are supported. If we only provide rules which work on the unmerges that we can handle, then this query would return false and the combine would not work. But we also cannot create a rule for those unmerges without causing endless loops through legalization.
> 
> The way we get around this problem is to add a custom rule, in which we simply return false (i.e. UnableToLegalize). This is enough for the combiner to produce the intermediary unmerge (because it is now support through the custom rule), while also getting around the problem of expanding something we don't want in the legalizer (because we say we didn't legalize anything). This is obviously a hack.
> 
> One of the ideas that was floated around yesterday sounded quite interesting to me: multi-stage legalization where you could specify which of your rules apply at which stage. I'm pretty sure this would solve our problem. In our case we would declare all artifacts as legal in the first stage to not hinder the combiner and in the second stage we could then focus on actually legalizing any left-over artifacts we have.
> 
> I do however see the problem that this could clutter up the existing legalization info. Due to the amount of instructions and rules, it already is quite complex and if rules could apply to different stages in the same file, it could make it quite difficult to understand what exactly is happening now.

It would definitely add some clutter but I suspect it would be manageable. Essentially it would be a common ruleset for most operations and each pass would add its own version of the merge/unmerge rules.

> I think Aditya pointed out that multi-stage legalization might be already possible by just having two legalizer passes with different legalization info and I feel like this might be the better approach of the two. I guess this would still require some tweeks as currently in llc we can only say `-stop-before/after=legalizer` but not which one of those.

That's right. Each legalizer pass owns it's own ruleset so two passes is a possibility. The -stop-before/after bit has been solved for some other passes but it does need a bit of boilerplate. Each subclass needs it's own pass id and INITIALIZE_* macros and getPassName() needs to be overridable.

> Another thing I was thinking of when I implemented this hack for our use-case was that we need some kind of rule which tells the combiner that something is supported but is actually not doing any legalization in the legalizer (something like a `.combine()`, `.combineFor({s96})`).

I'm not quite sure what you mean here. Are you thinking of legalization rules for combining or something like legalization rules but for the combiner? Are you thinking of artifact combines in particular or more generally for combines?

> Cheers,
> 
> Dominik
>