[llvm] [GlobalISel] An optimizing legal MIR builder (PR #132282)

[Amara Emerson via llvm-commits]

aemerson wrote:

> Should we remove them for compile time?

I actually asked you a question twice now for which you haven't provided a straight answer. Judging from the redirection with another question I'm assuming that the answer is "no". I also asked you earlier about your claimed compile time improvement and you also didn't reply to that. Did you actually see a compile time improvement or was it theoretical?
> 
> ```
> grep isLegalOrBeforeLegalizer llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp | wc -l
> 43
> ```
> 
> We have Legalizers for each target that precisely state what they consider legal. Sapphire Rapids and Thumb are different. I am confused about the it just works mentality.
> 
> What are the issues? Fail to select?
> 
> Edit: I opened a PR with a **strict** constant folder.

As this is your first ever PR to LLVM I understand that you are unclear about the idea that we may trade off theoretical purity for pragmatism in some cases. You may be surprised to know that this actually happens throughout the compiler. There are some optimization corner cases (e.g. pointer provenance comes to mind) where a full theoretical fix in the past would have been to disable an entire class of optimizations, but we don't do that because it comes at an unacceptable cost of performance. And that's a case where you actually *can* construct a test case to show the problem. Here we don't even know if this can ever be an issue on any sane target.


https://github.com/llvm/llvm-project/pull/132282