Log a backtrace on entrance to type inference

[timholy]

This is needed to eliminate the Core.Compiler.Timings module while still keeping SnoopCompile working.

This is one mechanism to add a backtrace to each entrance to inference. An alternative approach is presented in #58123. Only one of these should be merged.

[vchuravy]

I have a weak preference for this over option 1.

[timholy]

I like this one better too.

[vtjnash]

I am neutral between these PR, as they seem to express entirely different operations (first-dispatch vs first-infer) which may occur at different times in different places. The one thing that must occur before merging though is adding a lock around them (c.f. jl_push_newly_inferred for possible inspiration, as it seems to be fairly common that we need a primitive "locked array push" similar to that)

[vtjnash]

Can you write a few devdocs on the intended behavior of this, just especially so we can make sure it keeps working with that behavior? Is there a reason it use CI instead of MI? It looks like we could put the calls themselves in C (in jl_type_infer) as well to keep things a bit more self-contained

[timholy]

Can you write a few devdocs on the intended behavior of this, just especially so we can make sure it keeps working with that behavior?

Done. The real link is SnoopCompile, specifically JuliaDebug/SnoopCompile.jl#419. SnoopCompileCore is the only place this ccall is used.

I don't know whether you need more detail than I provided. The test should keep things working. I added a comment that indicates the purpose of the test, to prevent someone from concluding that it doesn't seem to be used anywhere and just deleting that test.

Is there a reason it use CI instead of MI?

Not really. I just went with CI since the trend has been to use CI in logs rather than MI. Happy to switch if you'd prefer it.

It looks like we could put the calls themselves in C (in jl_type_infer) as well to keep things a bit more self-contained

I once tried to figure out how to collect the backtrace from C, and that looked a lot harder. If I can get both, then sure, moving it makes sense. But the two have to be linked, so I don't see much advantage in getting one from C and the other from Julia when this already seems pretty non-intrusive to me.

[timholy]

Bump. I'll be on vacation for two weeks starting next week. It would be really nice to get this merged, so we can get rc2 out, so I can start directing users to test timholy/Revise.jl#894