This commit removes the `specialization_cache` field of `TyCtxt` by moving it to
a dedicated query, which it turned out was already quite easily structured to do
so!
This allows caching closure signatures and kinds in the normal selection
and evaluation caches, and fixes the exponential worst-case in
@remram44's example, which is a part of #43787.
This improvement is complenentary to #43999 - they fix different cases.
feature error span on attribute for fn_must_use, SIMD/align reprs, macro reëxport
There were several feature-gated attributes for which the feature-not-available
error spans would point to the item annotated with the gated attribute, when it
would make more sense for the span to point to the attribute itself: if the
attribute is removed, the function/struct/_&c._ likely still makes sense and the
program will compile. (Note that we decline to make the analogous change for
the `main`, `start`, and `plugin_registrar` features, for in those cases it
makes sense for the span to implicate the entire function, of which there is
little hope of using without the gated attribute.)
std: Mark allocation functions as nounwind
This commit flags all allocation-related functions in liballoc as "this can't
unwind" which should largely resolve the size-related issues found on #42808.
The documentation on the trait was updated with such a restriction (they can't
panic) as well as some other words about the relative instability about
implementing a bullet-proof allocator.
Closes#42808
clear out projection subobligations after they are processed
After a projection was processed, its derived subobligations no longer
need any processing when encountered, and can be removed. This improves
the status of #43787.
This is actually complementary to #43938 - that PR fixes selection
caching (and @remram44's example, which "accidentally" worked because of
the buggy projection caching) while this PR fixes projection caching.
r? @nikomatsakis
Generator support
This adds experimental support for generators intended to land once https://github.com/rust-lang/rfcs/pull/2033 is approved.
This is not yet ready to be merged. Things to do:
- [x] Make closure arguments on generators an error
- [x] Spot FIXMEs
- [x] Pass make tidy
- [x] Write tests
- [x] Document the current syntax and semantics for generators somewhere
- [x] Use proper error message numbers
- [x] ~~Make the implicit argument type default to `()`~~
This commit flags all allocation-related functions in liballoc as "this can't
unwind" which should largely resolve the size-related issues found on #42808.
The documentation on the trait was updated with such a restriction (they can't
panic) as well as some other words about the relative instability about
implementing a bullet-proof allocator.
Closes#42808
Clarify that VecDeque::swap can panic
The previous documentation mentioned this, but ambiguously used the term "fail".
This clarifies that the function will panic if the index is out of bounds, instead of silently failing and not doing anything.
If there's anything else I can do to improve this PR, I'd be happy to do so! Just saw this when reading through the docs in passing - it was slightly unclear what "fail" meant.
This commit introduces the ability to create and emit `Diagnostic`
structures from proc-macros, allowing for proc-macro authors to emit
warning, error, note, and help messages just like the compiler does.
There were several feature-gated attributes for which the
feature-not-available error spans would point to the item annotated with
the gated attribute, when it would make more sense for the span to point
to the attribute itself: if the attribute is removed, the
function/struct/&c. likely still makes sense and the program will
compile. (Note that we decline to make the analogous change for the
`main`, `start`, and `plugin_registrar` features, for in those cases it
makes sense for the span to implicate the entire function, of which
there is little hope of using without the gated attribute.)
rustbuild: Rewrite the configure script in Python
This commit rewrites our ancient `./configure` script from shell into Python.
The impetus for this change is to remove `config.mk` which is just a vestige of
the old makefile build system at this point. Instead all configuration is now
solely done through `config.toml`.
The python script allows us to more flexibly program (aka we can use loops
easily) and create a `config.toml` which is based off `config.toml.example`.
This way we can preserve comments and munge various values as we see fit.
It is intended that the configure script here is a drop-in replacement for the
previous configure script, no functional change is intended. Also note that the
rationale for this is also because our build system requires Python, so having a
python script a bit earlier shouldn't cause too many problems.
Closes#40730Closes#43295Closes#42255Closes#38058Closes#32176
This commit rewrites our ancient `./configure` script from shell into Python.
The impetus for this change is to remove `config.mk` which is just a vestige of
the old makefile build system at this point. Instead all configuration is now
solely done through `config.toml`.
The python script allows us to more flexibly program (aka we can use loops
easily) and create a `config.toml` which is based off `config.toml.example`.
This way we can preserve comments and munge various values as we see fit.
It is intended that the configure script here is a drop-in replacement for the
previous configure script, no functional change is intended. Also note that the
rationale for this is also because our build system requires Python, so having a
python script a bit earlier shouldn't cause too many problems.
Closes#40730
Make the ./x.py script fail when run with an invalid command, like:
./x.py nonsense
This helps in case of chaining multiple runs, eg.:
./x.py biuld && ./x.py test
ci: Remove the need for `dumb-init`
Newer versions of Docker have a `--init` argument which spawns an init process
in the container, which we should be able to use everywhere now.
After a projection was processed, its derived subobligations no longer
need any processing when encountered, and can be removed. This improves
the status of #43787.
This is actually complementary to #43938 - that PR fixes selection
caching (and @remram44's example, which "accidentally" worked because of
the buggy projection caching) while this PR fixes projection caching
Fixes issue #43205: ICE in Rvalue::Len evaluation.
- fixes evaluation of array length for zero-sized type referenced by rvalue operand.
- adds test to verify fix.
*Cause of the issue*.
Zero-sized aggregates are handled as operands, not lvalues. Therefore while visiting `Assign` statement by `LocalAnalyser`, `mark_as_lvalue()` is not called for related `Local`. This behaviour is controlled by `rvalue_creates_operand()` method.
As result it causes error later, when rvalue operand is evaluated in `trans_rvalue_operand()` while handling `Rvalue::Len` case. Array length evaluation invokes `trans_lvalue()` which expects referenced `Local` to be value, not operand.
*How it is fixed*.
In certain cases result of `Rvalue::Len` can be evaluated without calling
`trans_lvalue()`. Method `evaluate_array_len()` is introduced to handle length
evaluation for zero-sized types referenced by Locals.
*Some concerns*.
- `trans_lvalue()` has two other entry points in `rvalue.rs`: it is invoked while handling `Rvalue::Ref` and `Rvalue::Discriminant`. There is a chance those may produce the same issue, but I've failed to write a specific test that leads to this.
- `evaluate_array_len()` performs the same check (matches lvalue and `Local`), which is performed again in `trans_lvalue()`. Without changing `trans_lvalue()` signature to make it aware that caller deals with rvalue, it seems there is no cheap solution to avoid this check.
The previous documentation mentioned this, but ambiguously used the term "fail".
This clarifies that the function will panic if the index is out of bounds, instead of silently failing and not doing anything.
Make unused-extern-crate warn-by-default
Apart from enabling the lint, this pull request also removes existing unused crates in the codebase, and fix some amount of false positives on crates with special purposes.
Now that all false positive issues are closed, it should be possible to make it available to wider users.
Quote:
> Now that macro modularization is implemented, this is true today! *https://github.com/rust-lang/rust/issues/30849#issuecomment-286573218*
Concerns: can break some `#[deny(warnings)]`.
Close#42591
