Replace per-target ABI denylist with an allowlist
It makes very little sense to maintain denylists of ABIs when, as far as
non-generic ABIs are concerned, targets usually only support a small
subset of the available ABIs.
This has historically been a cause of bugs such as us allowing use of
the platform-specific ABIs on x86 targets – these in turn would cause
LLVM errors or assertions to fire.
In this PR we got rid of the per-target ABI denylists, and instead compute
which ABIs are supported with a simple match based on, mostly, the
`Target::arch` field. Among other things, this makes it impossible to
forget to consider this problem (in either direction) and forces one to
consider what the ABI support looks like when adding an ABI (rarely)
rather than target (often), which should hopefully also reduce the
cognitive load on both contributors as well as reviewers.
Fixes#57182
Sponsored by: standard.ai
---
## Summary for teams
One significant user-facing change after this PR is that there's now a future compat warning when building…
* `stdcall`, `fastcall`, `thiscall` using code with targets other than 32-bit x86 (i386...i686) or *-windows-*;
* `vectorcall` using code when building for targets other than x86 (either 32 or 64 bit) or *-windows-*.
Previously these ABIs have been accepted much more broadly, even for architectures and targets where this made no sense (e.g. on wasm32) and would fall back to the C ABI. In practice this doesn't seem to be used too widely and the [breakages in crater](https://github.com/rust-lang/rust/pull/86231#issuecomment-866300943) that we see are mostly about Windows-specific code that was missing relevant `cfg`s and just happened to successfully `check` on Linux for one reason or another.
The intention is that this warning becomes a hard error after some time.
It makes very little sense to maintain denylists of ABIs when, as far as
non-generic ABIs are concerned, targets usually only support a small
subset of the available ABIs.
This has historically been a cause of bugs such as us allowing use of
the platform-specific ABIs on x86 targets – these in turn would cause
LLVM errors or assertions to fire.
Fixes#57182
Sponsored by: standard.ai
Query-ify global limit attribute handling
Currently, we read various 'global limits' from inner attributes the crate root (`recursion_limit`, `move_size_limit`, `type_length_limit`, `const_eval_limit`). These limits are then stored in `Sessions`, allowing them to be access from a `TyCtxt` without registering a dependency on the crate root attributes.
This PR moves the calculation of these global limits behind queries, so that we properly track dependencies on crate root attributes. During the setup of macro expansion (before we've created a `TyCtxt`), we need to access the recursion limit, which is now done by directly calling into the code shared by the normal query implementations.
E0716: clarify that equivalent code example is erroneous
In E0716, there is a code block that is equivalent to the erroneous
code example. Especially when viewed with `rustc --explain`, it's
not obvious that it is also erroneous, and some users have been
confused when they try to change their code to match the erroneous
equivalent.
`@rustbot` label +A-diagnostics +D-newcomer-roadblock +T-compiler
Hack: Ignore inference variables in certain queries
Fixes#84841Fixes#86753
Some queries are not built to accept types with inference variables, which can lead to ICEs. These queries probably ought to be converted to canonical form, but as a quick workaround, we can return conservative results in the case that inference variables are found.
We should file a follow-up issue (and update the FIXMEs...) to do the proper refactoring.
cc `@arora-aman`
r? `@oli-obk`
In E0716, there is a code block that is equivalent to the erroneous
code example. Especially when viewed with `rustc --explain`, it's
not obvious that it is also erroneous, and some users have been
confused when they try to change their code to match the erroneous
equivalent.
