markused: fix for thread waiter usage

[felipensp]

Fix #24577

[huly-for-github]

Connected to Huly®: V_0.6-22945

[ctkjose]

The go statement is a coroutine. In the issue #24577 the go statement became a spawn because the option -use-coroutines was not used.

The problem is that spawn does not use v code that the builder could use to detect dependencies.

If modifying markused.v you could just do:

if table.gostmts > 0 {
	all_fn_root_names << 'free'
}

Using table.gostmts should be an acceptable way to assume a spawn without the need to modify the others.

PS In vlib/v/gen/c/cheaders.v we declare void v_free(voidptr ptr);. This is hardwired, so was it meant to be a base function always included? It was committed in 88810a7 for -gc boehm which added the C.GC_* functions. Wonder @spytheman could give us more clues.

[spytheman]

This is hardwired, so was it meant to be a base function always included?

markused evolved separately from the GC integration.
It was simpler (and buggier, measured by the % of .v files in the repository that could not be compiled with -skip-unused, or that could, at the price of still having a lot of unused code generated) initially, using a much larger set of hardcoded APIs.

Ideally, it should keep track of all of the used functionality, that the backends would then generate calls for, including free/malloc, so that compiling a .v file with the C backend, containing just pub fn f() {} should produce a simple C function, with no additional code.

[spytheman]

I agree, that using if table.gostmts > 0 { would be simpler, but it will also have a lower precision. Consider a go f() call, with fn f() { println('hi') } for example, that does not return anything, and that will not generate a _v_free() call in the V gowrapper function.

[spytheman]

Good work.

[ctkjose]

I agree, that using if table.gostmts > 0 { would be simpler, but it will also have a lower precision. Consider a go f() call, with fn f() { println('hi') } for example, that does not return anything, and that will not generate a _v_free() call in the V gowrapper function.

If I recall correctly the fix in this PR will also inject free regardless, it also adds more noise to the table.

The issue is that the spawn_and_go.v does not use actual v code for it to properly detect dependencies. Maybe an alternate version of parser.register_auto_import() that allows to register functions would avoid having to hack situations like this.