Compile time measurement is currently measuring some non-compilation things?

[NHDaly]

@vtjnash noted in #41733 that there are a couple places currently counted towards compilation time that maybe shouldn't be. Let's discuss! :)

Though it is actually unclear to me why this function counts against the compile-time. It is only used by the GPU stack, and is therefore not quite JIT time.

Originally posted by @vtjnash in #41733 (comment)

Same here: this is the implementation of code_llvm, which is not part of compile time, as that is run at runtime.

Originally posted by @vtjnash in #41733 (comment)

[NHDaly]

My gut is that for the first one, which is this:

julia/src/aotcompile.cpp

Lines 272 to 291 in 9deb6ed

	// takes the running content that has collected in the shadow module and dump it to disk
	// this builds the object file portion of the sysimage files for fast startup, and can
	// also be used be extern consumers like GPUCompiler.jl to obtain a module containing
	// all reachable & inferrrable functions. The `policy` flag switches between the default
	// mode `0`, the extern mode `1`, and imaging mode `2`.
	extern "C" JL_DLLEXPORT
	void *jl_create_native(jl_array_t *methods, const jl_cgparams_t cgparams, int _policy)
	{
	jl_native_code_desc_t *data = new jl_native_code_desc_t;
	jl_codegen_params_t params;
	params.params = &cgparams;
	std::map<jl_code_instance_t*, jl_compile_result_t> emitted;
	jl_method_instance_t *mi = NULL;
	jl_code_info_t *src = NULL;
	JL_GC_PUSH1(&src);
	JL_LOCK(&codegen_lock);
	uint64_t compiler_start_time = 0;
	uint8_t measure_compile_time_enabled = jl_atomic_load(&jl_measure_compile_time_enabled);
	if (measure_compile_time_enabled)
	compiler_start_time = jl_hrtime();

that we probably should count it. Since if your code is doing GPU compilation, you probably want to know about that still if you're running @time? I would think so.

---

For the second one, I'm less sure. I think you're probably right that we wouldn't want to measure that... But i'm not sure if there's some case where we'd want to report the compilation time separately there.. i don't think so though, yeah.

[NHDaly]

CC: @IanButterworth, @Sacha0, @janrous-rai

[vchuravy]

GPUCompiler has it's own separate accounting. So I would be okay with removing it here and making it mean: CPU compile time