codegen: create DIVariables ahead of using them with llvm.dbg.declare.
Instead of having `rustc_codegen_llvm` bundle creation of a `DIVariable` and `llvm.dbg.declare` into a single operation, they are now two separate methods, and the `DIVariable` is created earlier, specifically when `mir::VarDebugInfo`s are being partitioned into locals.
While this isn't currently needed, it's a prerequisite for #48300, which adds fragmented debuginfo, where one `mir::VarDebugInfo` has multiple parts of itself mapped to different `mir::Place`s.
For debuggers to see one composite variable instead of several ones with the same name, we need to create a single `DIVariable` and share it between multiple `llvm.dbg.declare` calls, which are likely pointing to different MIR locals.
That makes the `per_local_var_debug_info` partitioning a good spot to do this in, as we can create *exactly* one `DIVariable` per `mir::VarDebugInfo`, and refer to it as many things as needed.
I'm opening this PR separately because I want to test its perf impact in isolation (see https://github.com/rust-lang/rust/pull/48300#issuecomment-580121438).
r? @nagisa or @oli-obk cc @michaelwoerister @nikomatsakis
Avoid exponential behaviour when relating types
When equating bound types we check subtyping in both directions. Since closures are invariant in their substs, we end up comparing the two types an exponential number of times. If there are no bound variables this isn't needed.
Closes#68061
Changes the error handler for inner attributes that replace the root
with a non-module. Previously it would emit a fatal error. It now emits
an empty expasion and a non-fatal error like the existing handler for a
failed expansion.
This function has a variable `changed` that is erroneously used for two
related-but-different purpose:
- to detect if the current element has changed;
- to detect if any elements have changed.
As a result, its use for the first purpose is broken, because if any
prior element changed then the code always thinks the current element
has changed. This is only a performance bug, not a correctness bug,
because we frequently end up calling `assign_unpacked` unnecessarily to
overwrite the element with itself.
This commit adds `any_changed` to correctly distinguish between the two
purposes. This is a small perf win for some benchmarks.
parser: syntactically allow `self` in all `fn` contexts
Part of https://github.com/rust-lang/rust/pull/68728.
`self` parameters are now *syntactically* allowed as the first parameter irrespective of item context (and in function pointers). Instead, semantic validation (`ast_validation`) is used.
r? @petrochenkov
