Using --no-codegen blocks --emit options

[bew]

When we do: crystal build hello.cr --emit llvm-ir --no-codegen
It'll not generate the hello.ll file even though it was asked for.

This is because the emit options are handled all-in-one-place in the CompilationUnit#emit method

crystal/src/compiler/crystal/compiler.cr

Lines 676 to 687 in 80cbe66

	def emit(value : String, output_filename)
	case value
	when "asm"
	compiler.target_machine.emit_asm_to_file llvm_mod, "#{output_filename}.s"
	when "llvm-bc"
	FileUtils.cp(bc_name, "#{output_filename}.bc")
	when "llvm-ir"
	llvm_mod.print_to_file "#{output_filename}.ll"
	when "obj"
	FileUtils.cp(object_name, "#{output_filename}.o")
	end
	end

Which is called in the codegen block:

crystal/src/compiler/crystal/compiler.cr

Lines 339 to 353 in 80cbe66

	@progress_tracker.stage("Codegen (bc+obj)") do
	@progress_tracker.stage_progress_total = units.size
	if units.size == 1
	first_unit = units.first
	first_unit.compile
	reused << first_unit.name if first_unit.reused_previous_compilation?
	if emit = @emit
	first_unit.emit(emit, emit_base_filename || output_filename)
	end
	else
	reused = codegen_many_units(program, units, target_triple)
	end
	end

At line 348.

I think it should emit what was asked for as soon as it get it, so the llvm-ir after it was built, the obj in the codegen phase, etc..

[bew]

This also means that there is no way to dump llvm IR when there is a module validation failed error, that we would like to debug.

[straight-shoota]

In order to debug #5972 I had to modify the compiler manually to disable module validation. So +1 to that!

[bew]

While looking into that, I noticed that all the codegen phases (mainly Crystal to LLVM IR and LLVM IR to BC+OBJ) are done in the codegen method.
I suggest to separate the Crystal to LLVM IR codegen phase from the binary codegen (LLVM-IR to BC+OBJ), maybe name it IR codegen ?

Also the current emit options are in 2 categories:

• after llvm ir generation: for llvm-ir (should be processed even on --no-codegen)
• after binary codegen: for asm, llvm-bc, obj (not processed on --no-codegen)

Those are currently all handled and processed after binary codegen. By having the IR codegen separated, we could handle the different emit options at different time.

WDYT?

[bew]

Also, when doing --no-codegen this would disable all codegen ? or only binary codegen? or how to configure that?

Maybe if there is --emit llvm-ir, it would only disable binary codegen (IR codegen will still be done, so that the IR can be dumped), and when not given, all codegen would be disabled.

[asterite]

What's so bad about not using --no-codegen when using --emit? No codegen means the codegen phase doesn't run, and this trivially means no artifacts (.ll, .s, .o, final executable) will be produced.

[straight-shoota]

@asterite True, these are all different stages of codegen. But what do you do if you want the .ll (or .s) but not .o?

--no-codegen --emit llvm-ir makes sense for this: It essentially says "skip codegen, but emit LLVM-IR" which can be interpreted as "do codegen but abort after LLVM-IR is dumped". There is no other logical interpretation of the combination of these two flags other than ignoring or failing when --emit is presented together with --no-codegen. But if it can be expressed that way, why shouldn't it be usable?

[asterite]

True, these are all different stages of codegen. But what do you do if you want the .ll (or .s) but not .o?

You just wait a little more? :-)

It essentially says "skip codegen, but emit LLVM-IR" which can be interpreted as "do codegen but abort after LLVM-IR is dumped".

There's the confusion. Codegen means "run code to create the LLVM in memory". Emit comes after that. Of course we could change that, but I don't see the point. Just wait a few more seconds and you'll have it.

[asterite]

Or, well, put another one, if someone wants to implement it, please send a PR (I won't)

[bew]

Also, there could be a --no-binary-codegen in addition to --no-codegen, where the latter disables all codegen (ir & binary), and the former disable the binary codegen only, but allow the ir codegen to be done.

I think it allows more control, and remove weird edge cases like "always do IR codegen when --emit is llvm-ir even on --no-codegen".

[HertzDevil]

For the Compiler Explorer it would be very useful if the assembly and the IR can both be generated without actually compiling / linking that into a binary or .o file. One reason is that link failures should be tolerated there unless the code is ultimately executed as requested by the user.

[mattrberry]

Agree with HertzDevil above. It would be useful to be able to emit and inspect the IR without needing to link