Make the lifetime parameters of tcx consistent.

[gabi-250]

I have implemented codegen_allocator for my backend, but I've had to make a small change to its signature in ExtraBackendMethods. I wonder if this change is justified, or if it is too specific to my use case to be useful to anyone else.

write_metadata and codegen_allocator are both called from codegen_crate (in librustc_codegen_ssa/base.rs), and they both receive the same tcx as an argument:

rust/src/librustc_codegen_ssa/base.rs

Lines 555 to 557 in c196097

	let metadata = time(tcx.sess, "write metadata", || {
	backend.write_metadata(tcx, &mut metadata_llvm_module)
	});

and:

rust/src/librustc_codegen_ssa/base.rs

Lines 640 to 642 in c196097

	time(tcx.sess, "write allocator module", || {
	backend.codegen_allocator(tcx, &mut modules, kind)
	});

However, codegen_allocator accepts a TyCtxt with any lifetime parameters (tcx: TyCtxt<'_, '_, '_>), while write_metadata requires that the tcx argument is of type TyCtxt<'b, 'gcx, 'gcx>. In my implementation, I've found that it's necessary for tcx in codegen_allocator to also have the <'b, 'gcx, 'gcx> lifetime parameters.

Have I misunderstood the purpose of the parameters of TyCtxt? I've read here that the last two parameters only need to be distinct if the function needs to be used during type inference, but I don't think that is the case here.

[rust-highfive]

r? @petrochenkov

(rust_highfive has picked a reviewer for you, use r? to override)

[petrochenkov]

write_metadata doesn't seem to technically require the parameters to be the same, but I guess this is still the intent given the stage at which write_metadata works.
Then it seems reasonable to adjust codegen_allocator.

@bors r+ rollup

[bors]

📌 Commit cd9a0cf has been approved by petrochenkov