Revert "Revert "Merge pull request #15860 from JuliaLang/kf/tracing""

[Keno]

This was silently reverted by #15779.

[vtjnash]

i'm concerned about the usage of this now that i can see JuliaDebug/Gallium.jl@a14a67b. the one thing you must not do with the returned lambda object is look at its llvm object (there are a few code paths where that can assert). so i don't see how it can be safely used for debugging.

also you need to update jltypes.c now

also, 810e53f was unrelated, but I'm not sure it's a good idea to assume without testing that jl_tparam0 in jl_get_llvmf is a datatype.

[Keno]

i'm concerned about the usage of this now that i can see JuliaDebug/Gallium.jl@a14a67b. the one thing you must not do with the returned lambda object is look at its llvm object (there are a few code paths where that can assert). so i don't see how it can be safely used for debugging.

Returned lambda object where?

also you need to update jltypes.c now

in what way?

also, 810e53f was unrelated, but I'm not sure it's a good idea to assume without testing that jl_tparam0 in jl_get_llvmf is a datatype.

Adding an assert to this extent seems reasonable.

[vtjnash]

from cache_method / jl_linfo_tracer

in what way?

there's an extra field in jl_method_t, it needs to be added to the description of the type

[Keno]

Which code path is it that would assert. Would it be reasonable to add an extra tracing flag to the lambda info that gets triggered when that lambda info gets compiled? That flags could be set from jl_linfo_tracer, and only set the breakpoint after the thing is actually compiled?

I'm not sure I want this flag to be accessible from julia.

[vtjnash]

you can give it a name of "" to make it harder to access

that's sounds like a possibility. we might be able to attach NotifyObjectEmitted to Julia? that's a tricky time too though, since calling jl_get_llvmf in there will hit llvm assertions (recursive compilation and double-finalizing the RTDyld)

codegen may decide to assert if inference hasn't decided to finish yet (for example, if cache_method is called from within inference), or if the function is being compiled recursively and forms part of a cycle (because it would force resolution of the function before the rest of the cycle was ready)

[Keno]

We could record all traced linfos in NotifyObjectEmitted and delay calling the callback until we're out of the JIT again, no?

[vtjnash]

that sounds plausible. i don't know off-hand how many entry points we have into the JIT, but I don't believe it could be that many (c-function, c-alias, and fptr?)

[Keno]

Don't they all go through getAddressForFunction?