Add ccalls to enable dumping LLVM_OPT time per method instance.

[NHDaly]

Adds these functions to toggle the output:

• ccall(:jl_dump_emitted_mi_name, Nothing, (Ptr{Nothing},), e_io.handle)
• ccall(:jl_dump_llvm_opt, Nothing, (Ptr{Nothing},), l_io.handle)

This will then be controlled from SnoopCompileCore, via timholy/SnoopCompile.jl#135.

This logs a CSV file with mappings from C function name to Julia
MethodInstance name, and a YAML file with timings statistics about the
LLVM module being optimized, including:

• Total time to optimize the llvm module
• The optlevel used for the llvm module
• The name of each function in the module, before and after
• Number of instructions for each function, before and after
• Number of basic blocks for each function, before and after

Co-Authored-By: @Sacha0 @vchuravy
Fixes #31616

[NHDaly]

CC: @Sacha0 @vchuravy

[NHDaly]

(I've opened this branch which just applies these same changes back to julia 1.5, so we can test our codebase which currently doesn't support 1.6 because of some of our package dependencies):
v1.5.0...NHDaly:nhd-emit-per-method-llvm_timings--1.5

[vtjnash]

SGTM

[NHDaly]

Thanks for the review, @vtjnash! Sorry for my delay in responding.

As a short-term bandaid, -O1 on julia 1.5 made a big improvement, so we stopped looking at this. But now that we're working through the inference timings, we'd like to go back to -O2, so we'll pick up this investigation again soon.

I'm sorry I let this languish for a bit, but do you think we'll be able to get this merged into 1.6? It would be really great to have this there so that our engineers don't need a custom build of julia to debug compilation times.

[NHDaly]

And, similarly, I'm going to try to wrap up the per-method-instance inference timings PR as well: #37749.

Ideally we'll be able to merge both of these this week so that they can make it into the v1.6 release! 😊 Does that seem possible to you? Sorry to be coming in so last-minute! ❤️ thanks for your help

[Sacha0]

This patch looks great! Thanks @NHDaly! :)

It looks like one comment from Jameson remains outstanding (rewriting a conditional in terms of isDeclaration and startsWith), which I noted applies to similar code later in the patch as well. (Perhaps the duplicated code snippet should be factored into a little helper function? (It's sufficiently little duplication that I don't mind either way.) Otherwise this patch looks ready to go! :)

[NHDaly]

Thanks @Sacha0! Addressed comments, rebased on top of master, and pushed.

[Sacha0]

Looks great! Thanks @NHDaly! :)

[vchuravy]

PR needs squashing.

[NHDaly]

Done. Thanks! :)