Remove some Core.Boxes

[giordano]

Using script from JuliaLang/julia#60478
The seven four remaining boxes are

var	func	sig	location
changed	#emit_llvm#92	Tuple{GPUCompiler.var"##emit_llvm#92", Base.Pairs{Symbol, V, Nothing, NT} where {V, NT<:NamedTuple}, typeof(GPUCompiler.emit_llvm), CompilerJob}	src/driver.jl:182
config	#emit_llvm#92	Tuple{GPUCompiler.var"##emit_llvm#92", Base.Pairs{Symbol, V, Nothing, NT} where {V, NT<:NamedTuple}, typeof(GPUCompiler.emit_llvm), CompilerJob}	src/driver.jl:182
dyn_entry_fn	#emit_llvm#92	Tuple{GPUCompiler.var"##emit_llvm#92", Base.Pairs{Symbol, V, Nothing, NT} where {V, NT<:NamedTuple}, typeof(GPUCompiler.emit_llvm), CompilerJob}	src/driver.jl:182
localUnique	@device_code	Tuple{GPUCompiler.var"#@device_code", LineNumberNode, Module, Vararg{Any}}	src/reflection.jl:364

I attempted to fix the first three in src/driver.jl, but it was a mess because changed, config and dyn_entry_fn are captured by a get!() do in

GPUCompiler.jl/src/driver.jl

Lines 252 to 263 in da0be16

	dyn_entry_fn = get!(jobs, dyn_job) do
	target = nest_target(dyn_job.config.target, job.config.target)
	params = nest_params(dyn_job.config.params, job.config.params)
	config = CompilerConfig(dyn_job.config; toplevel=false, target, params)
	dyn_ir, dyn_meta = codegen(:llvm, CompilerJob(dyn_job; config))
	dyn_entry_fn = LLVM.name(dyn_meta.entry)
	merge!(compiled, dyn_meta.compiled)
	@assert context(dyn_ir) == context(ir)
	link!(ir, dyn_ir)
	changed = true
	dyn_entry_fn
	end

and I couldn't come up with a solution. I don't understand what's wrong with job. runtime_fns and runtime_intrinsics should be fixable [Edit: job, runtime_fns, and runtime_intrinsics are all fixed now], but localUnique in

GPUCompiler.jl/src/reflection.jl

Lines 364 to 396 in da0be16

	macro device_code(ex...)
	localUnique = 1
	function hook(job::CompilerJob; dir::AbstractString)
	name = job.source.def.name
	fn = "$(name)_$(localUnique)"
	mkpath(dir)
	open(joinpath(dir, "$fn.lowered.jl"), "w") do io
	code = only(code_lowered(job))
	println(io, code)
	end
	open(joinpath(dir, "$fn.typed.jl"), "w") do io
	code = only(code_typed(job; debuginfo=:source))
	println(io, code)
	end
	open(joinpath(dir, "$fn.unopt.ll"), "w") do io
	code_llvm(io, job; dump_module=true, raw=true, optimize=false)
	end
	open(joinpath(dir, "$fn.opt.ll"), "w") do io
	code_llvm(io, job; dump_module=true, raw=true)
	end
	open(joinpath(dir, "$fn.asm"), "w") do io
	code_native(io, job; dump_module=true, raw=true)
	end
	localUnique += 1
	end
	emit_hooked_compilation(hook, ex...)
	end

is a bit evil because it's purposefully captured, no clue of what to do with it without invasive changes to the hook business.

Note: most tests pass for me when I run them with Pkg.test, except for native/kwcall inference + overlay method (which fails for me also on main so I don't think it's my fault) and ptx, which crashes the worker without an error message, but when I run the script test/ptx.jl standalone (with the test init code added at the beginning) the test is successful, so I have little clue of what's the issue there:

% julia --startup-file=no --project --check-bounds=yes ptx.jl
Test Summary: | Pass  Total  Time
IR            |   18     18  5.8s

Edit: resolved.

[github-actions]

Your PR no longer requires formatting changes. Thank you for your contribution!

[codecov]

Codecov Report

❌ Patch coverage is 53.23741% with 65 lines in your changes missing coverage. Please review.
✅ Project coverage is 73.28%. Comparing base (da0be16) to head (32c335b).

Files with missing lines	Patch %	Lines
src/irgen.jl	52.43%	39 Missing ⚠️
src/mcgen.jl	0.00%	15 Missing ⚠️
src/metal.jl	71.87%	9 Missing ⚠️
src/driver.jl	75.00%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #751      +/-   ##
==========================================
- Coverage   74.79%   73.28%   -1.52%     
==========================================
  Files          24       24              
  Lines        3666     3668       +2     
==========================================
- Hits         2742     2688      -54     
- Misses        924      980      +56     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[maleadt]

Do these hurt performance? I intentionally wrote the code as such, e.g., keeping the check_user lambda close to its only relevant use case instead of defining a global function (there's no possibility for reuse). Same with scan_uses! and rewrite_uses!. As much as I appreciate performance work, code maintainability is important as well...