Unconditionally call really_init on GNU/Linux

[workingjubilee]

This makes miri not diverge in behavior, it fixes running Rust linux-gnu binaries on musl with gcompat, it fixes dlopen edge-cases that cranelift somehow hits, etc.

Fixes #124126

thou hast gazed into this abyss with me:
r? @ChrisDenton

[ChrisDenton]

Making the gnu case the same as the musl case makes sense to me. It is just setting two atomics so I don't think it saves much to skip init from main

[ChrisDenton]

@bors r+ rollup

[bors]

📌 Commit fa73ebb has been approved by ChrisDenton

It is now in the queue for this repository.

[joshtriplett]

Could we check, first, whether we've already had these initialized, and only call this in the unusual case where they're not initialized? That would avoid the extra work in the common case, and do the work in the unusual case.

[workingjubilee]

@joshtriplett how, exactly, is updating an atomic a single extra time in a program's entire lifetime "extra work"?

[workingjubilee]

Doing a branch and only then updating is quite plausibly more work because it impacts the branch prediction tables.

[ChrisDenton]

I'm happy to look at any PR changing this but, as jubilee says, I'm really not clear how we could possibly do less work then a couple of relaxed stores? I mean other than doing nothing, which I'm assuming is still undesirable.

[workingjubilee]

I can add a comment explaining that the reason the scare-quotes are there is because the "duplicate work" is so enormously trivial that on almost all systems it is actually more costly to write the if, as anything that doesn't have a branch predictor also makes its atomic locking implementation dead-simple.

[joshtriplett]

@workingjubilee I read the couple of GitHub issues too quickly, didn't dig into the body of really_init, and forgot that really_init is only saving off argc and argv as values, not doing the parsing that turns them into the form provided by args()/args_os(). Doing a "duplicate" initialization of two relaxed stores is definitely fine to duplicate, and as you said, would cost more to try to avoid.