Remove `Session::crate_types_opt`?

[ecstatic-morse]

See #72256 (comment) for context.

There's only one place in the compiler that can't call Session::crate_types because that field is sometimes uninitialized.

rust/src/librustc_codegen_llvm/context.rs

Lines 100 to 103 in 31add7e

	/// PIE is potentially more effective than PIC, but can only be used in executables.
	/// If all our outputs are executables, then we can relax PIC to PIE when producing object code.
	/// If the list of crate types is not yet known we conservatively return `false`.
	pub fn all_outputs_are_pic_executables(sess: &Session) -> bool {

Perhaps this indicates that we want to force initialization of crate_types somewhere?

[petrochenkov]

all_outputs_are_pic_executables is used during creation of LLVM TargetMachine.
Sometimes we create an "informational" target machine to query supported LLVM target features, for example.
We don't need to force initialize crate_types in this case because the queried information doesn't depend on them, but we also can't do it in some cases because initialization of crate_types requires running parsing and expansion, which would create cycles because the LLVM target feature list is an input for expansion.

[Mark-Simulacrum]

Hm, it sounds like maybe we should not attempt a read of the crate types of we're creating it for that informational use case?

I presume the crate type can influence LLVM in some way, right?

[petrochenkov]

Ha, the PositionIndependentExecutable parameter in LLVMRustCreateTargetMachine is not even used.
We need to enable more warnings in C++ code.

So it should be possible to remove crate_types_opt from all_outputs_are_pic_executables, I'll prepare a PR now.

[petrochenkov]

Fixed in #72435.