Add "algebraic" fast-math intrinsics, based on fast-math ops that cannot return poison
Setting all of LLVM's fast-math flags makes our fast-math intrinsics very dangerous, because some inputs are UB. This set of flags permits common algebraic transformations, but according to the [LangRef](https://llvm.org/docs/LangRef.html#fastmath), only the flags `nnan` (no nans) and `ninf` (no infs) can produce poison.
And this uses the algebraic float ops to fix https://github.com/rust-lang/rust/issues/120720
cc `@orlp`
Implement intrinsics with fallback bodies
fixes#93145 (though we can port many more intrinsics)
cc #63585
The way this works is that the backend logic for generating custom code for intrinsics has been made fallible. The only failure path is "this intrinsic is unknown". The `Instance` (that was `InstanceDef::Intrinsic`) then gets converted to `InstanceDef::Item`, which represents the fallback body. A regular function call to that body is then codegenned. This is currently implemented for
* codegen_ssa (so llvm and gcc)
* codegen_cranelift
other backends will need to adjust, but they can just keep doing what they were doing if they prefer (though adding new intrinsics to the compiler will then require them to implement them, instead of getting the fallback body).
cc `@scottmcm` `@WaffleLapkin`
### todo
* [ ] miri support
* [x] default intrinsic name to name of function instead of requiring it to be specified in attribute
* [x] make sure that the bodies are always available (must be collected for metadata)
This test started failing on LLVM 18 after change
61118ffd04. As far as I can tell, it's
just good fortune that LLVM is able to sniff out the new noalias here,
and it's correct.
Remove an unneeded helper from the tuple library code
Thanks to https://github.com/rust-lang/rust/pull/107022, this is just what `==` does, so we don't need the helper here anymore.
Rollup of 8 pull requests
Successful merges:
- #120484 (Avoid ICE when is_val_statically_known is not of a supported type)
- #120516 (pattern_analysis: cleanup manual impls)
- #120517 (never patterns: It is correct to lower `!` to `_`.)
- #120523 (Improve `io::Read::read_buf_exact` error case)
- #120528 (Store SHOULD_CAPTURE as AtomicU8)
- #120529 (Update data layouts in custom target tests for LLVM 18)
- #120531 (Remove a bunch of `has_errors` checks that have no meaningful or the wrong effect)
- #120533 (Correct paths for hexagon-unknown-none-elf platform doc)
r? `@ghost`
`@rustbot` modify labels: rollup
Avoid ICE when is_val_statically_known is not of a supported type
2 ICE with 1 stone!
1. Implement `llvm.is.constant.ptr` to avoid first ICE in linked issue.
2. return `false` when the argument is not one of `i*`/`f*`/`ptr` to avoid second ICE.
fixes#120480
Replacement of #114390: Add new intrinsic `is_var_statically_known` and optimize pow for powers of two
This adds a new intrinsic `is_val_statically_known` that lowers to [``@llvm.is.constant.*`](https://llvm.org/docs/LangRef.html#llvm-is-constant-intrinsic).` It also applies the intrinsic in the int_pow methods to recognize and optimize the idiom `2isize.pow(x)`. See #114390 for more discussion.
While I have extended the scope of the power of two optimization from #114390, I haven't added any new uses for the intrinsic. That can be done in later pull requests.
Note: When testing or using the library, be sure to use `--stage 1` or higher. Otherwise, the intrinsic will be a noop and the doctests will be skipped. If you are trying out edits, you may be interested in [`--keep-stage 0`](https://rustc-dev-guide.rust-lang.org/building/suggested.html#faster-builds-with---keep-stage).
Fixes#47234Resolves#114390
`@Centri3`
Use `assert_unchecked` instead of `assume` intrinsic in the standard library
Now that a public wrapper for the `assume` intrinsic exists, we can use it in the standard library.
CC #119131
Fix overflow check
Make MIRI choose the path randomly and rename the intrinsic
Add back test
Add miri test and make it operate on `ptr`
Define `llvm.is.constant` for primitives
Update MIRI comment and fix test in stage2
Add const eval test
Clarify that both branches must have the same side effects
guaranteed non guarantee
use immediate type instead
Co-Authored-By: Ralf Jung <post@ralfj.de>
Tune the inlinability of `unwrap`
Fixes#115463
cc `@thomcc`
This tweaks `unwrap` on ~~`Option` &~~ `Result` to be two parts:
- `#[inline(always)]` for checking the discriminant
- `#[cold]` for actually panicking
The idea here is that checking the discriminant on a `Result` ~~or `Option`~~ should always be trivial enough to be worth inlining, even in `opt-level=z`, especially compared to passing it to a function.
As seen in the issue and codegen test, this will hopefully help particularly for things like `.try_into().unwrap()`s that are actually infallible, but in a way that's only visible with the inlining.
EDIT: I've restricted this to `Result` to avoid combining effects
Separate immediate and in-memory ScalarPair representation
Currently, we assume that ScalarPair is always represented using a two-element struct, both as an immediate value and when stored in memory.
This currently works fairly well, but runs into problems with https://github.com/rust-lang/rust/pull/116672, where a ScalarPair involving an i128 type can no longer be represented as a two-element struct in memory. For example, the tuple `(i32, i128)` needs to be represented in-memory as `{ i32, [3 x i32], i128 }` to satisfy alignment requirements. Using `{ i32, i128 }` instead will result in the second element being stored at the wrong offset (prior to LLVM 18).
Resolve this issue by no longer requiring that the immediate and in-memory type for ScalarPair are the same. The in-memory type will now look the same as for normal struct types (and will include padding filler and similar), while the immediate type stays a simple two-element struct type. This also means that booleans in immediate ScalarPair are now represented as i1 rather than i8, just like we do everywhere else.
The core change here is to llvm_type (which now treats ScalarPair as a normal struct) and immediate_llvm_type (which returns the two-element struct that llvm_type used to produce). The rest is fixing things up to no longer assume these are the same. In particular, this switches places that try to get pointers to the ScalarPair elements to use byte-geps instead of struct-geps.
llvm: Allow `noundef` in codegen tests
LLVM 18 will automatically infer `noundef` in some situations. Adjust codegen tests to accept this.
See llvm/llvm-project#76553 for why `noundef` is being generated now.
``@rustbot`` label:+llvm-main
LLVM 18 will automatically infer `noundef` in some situations.
Adjust codegen tests to accept this.
See llvm/llvm-project#76553 for why `noundef` is being generated now.
On LLVM 18 we get slightly different arguments here, so it's easier to
just regex those away. The important details are all still asserted as I
understand things.
Fixes#119193.
@rustbot label: +llvm-main
add more niches to rawvec
Previously RawVec only had a single niche in its `NonNull` pointer. With this change it now has `isize::MAX` niches since half the value-space of the capacity field is never needed, we can't have a capacity larger than isize::MAX.
Currently, we assume that ScalarPair is always represented using
a two-element struct, both as an immediate value and when stored
in memory.
This currently works fairly well, but runs into problems with
https://github.com/rust-lang/rust/pull/116672, where a ScalarPair
involving an i128 type can no longer be represented as a two-element
struct in memory. For example, the tuple `(i32, i128)` needs to be
represented in-memory as `{ i32, [3 x i32], i128 }` to satisfy
alignment requirement. Using `{ i32, i128 }` instead will result in
the second element being stored at the wrong offset (prior to
LLVM 18).
Resolve this issue by no longer requiring that the immediate and
in-memory type for ScalarPair are the same. The in-memory type
will now look the same as for normal struct types (and will include
padding filler and similar), while the immediate type stays a
simple two-element struct type. This also means that booleans in
immediate ScalarPair are now represented as i1 rather than i8,
just like we do everywhere else.
The core change here is to llvm_type (which now treats ScalarPair
as a normal struct) and immediate_llvm_type (which returns the
two-element struct that llvm_type used to produce). The rest is
fixing things up to no longer assume these are the same. In
particular, this switches places that try to get pointers to the
ScalarPair elements to use byte-geps instead of struct-geps.
Sets the accessibility of types and fields in DWARF using
`DW_AT_accessibility` attribute.
`DW_AT_accessibility` (public/protected/private) isn't exactly right for
Rust, but neither is `DW_AT_visibility` (local/exported/qualified), and
there's no way to set `DW_AT_visbility` in LLVM's API.
Signed-off-by: David Wood <david@davidtw.co>
Enable stack probes on aarch64 for LLVM 18
I tested this on `aarch64-unknown-linux-gnu` with LLVM main (~18).
cc #77071, to be closed once we upgrade our LLVM submodule.
Add `-Zfunction-return={keep,thunk-extern}` option
This is intended to be used for Linux kernel RETHUNK builds.
With this commit (optionally backported to Rust 1.73.0), plus a patched Linux kernel to pass the flag, I get a RETHUNK build with Rust enabled that is `objtool`-warning-free and is able to boot in QEMU and load a sample Rust kernel module.
Issue: https://github.com/rust-lang/rust/issues/116853.
This is intended to be used for Linux kernel RETHUNK builds.
With this commit (optionally backported to Rust 1.73.0), plus a
patched Linux kernel to pass the flag, I get a RETHUNK build with
Rust enabled that is `objtool`-warning-free and is able to boot in
QEMU and load a sample Rust kernel module.
Signed-off-by: Miguel Ojeda <ojeda@kernel.org>
Add thinlto support to codegen, assembly and coverage tests
Using `--emit=llvm-ir` with thinlto usually result in multiple IR files.
Resolve test case failure issue reported in #113923.
In the future Windows will enable Control-flow Enforcement Technology
(CET aka Shadow Stacks). To protect the path where the context is
updated during exception handling, the binary is required to enumerate
valid unwind entrypoints in a dedicated section which is validated when
the context is being set during exception handling.
The required support for EHCONT has already been merged into LLVM,
long ago. This change adds the Rust codegen option to enable it.
Reference:
* https://reviews.llvm.org/D40223
This also adds a new `ehcont-guard` option to the bootstrap config which
enables EHCont Guard when building std.
Add -Z llvm_module_flag
Allow adding values to the `!llvm.module.flags` metadata for a generated module. The syntax is
`-Z llvm_module_flag=<name>:<type>:<value>:<behavior>`
Currently only u32 values are supported but the type is required to be specified for forward compatibility. The `behavior` element must match one of the named LLVM metadata behaviors.viors.
This flag is expected to be perma-unstable.
This was broken by upstream
llvm/llvm-project@dc6d077396. It's easy
enough to use a regex match to support both, so we do that.
r? @nikic
@rustbot label: +llvm-main
Allow adding values to the `!llvm.module.flags` metadata for a generated
module. The syntax is
`-Z llvm_module_flag=<name>:<type>:<value>:<behavior>`
Currently only u32 values are supported but the type is required to be
specified for forward compatibility. The `behavior` element must match
one of the named LLVM metadata behaviors.viors.
This flag is expected to be perma-unstable.
Hint optimizer about try-reserved capacity
This is #116568, but limited only to the less-common `try_reserve` functions to reduce bloat in debug binaries from debug info, while still addressing the main use-case #116570
Clean up unchecked_math, separate out unchecked_shifts
Tracking issue: #85122
Changes:
1. Remove `const_inherent_unchecked_arith` flag and make const-stability flags the same as the method feature flags. Given the number of other unsafe const fns already stabilised, it makes sense to just stabilise these in const context when they're stabilised.
2. Move `unchecked_shl` and `unchecked_shr` into a separate `unchecked_shifts` flag, since the semantics for them are unclear and they'll likely be stabilised separately as a result.
3. Add an `unchecked_neg` method exclusively to signed integers, under the `unchecked_neg` flag. This is because it's a new API and probably needs some time to marinate before it's stabilised, and while it *would* make sense to have a similar version for unsigned integers since `checked_neg` also exists for those there is absolutely no case where that would be a good idea, IMQHO.
The longer-term goal here is to prepare the `unchecked_math` methods for an FCP and stabilisation since they've existed for a while, their semantics are clear, and people seem in favour of stabilising them.
Decompose singular `matches!` with or-patterns to individual `matches!`
statements to enable branchless code output. The following functions
were changed:
- `is_ascii_alphanumeric`
- `is_ascii_hexdigit`
- `is_ascii_punctuation`
Add codegen tests
Co-authored-by: George Bateman <george.bateman16@gmail.com>
Co-authored-by: scottmcm <scottmcm@users.noreply.github.com>
Copy 1-element arrays as scalars, not vectors
For `[T; 1]` it's silly to copy as `<1 x T>` when we can just copy as `T`.
Inspired by https://github.com/rust-lang/rust/issues/101210#issuecomment-1732470941, which pointed out that `Option<[u8; 1]>` was codegenning worse than `Option<u8>`.
(I'm not sure *why* LLVM doesn't optimize out `<1 x u8>`, but might as well just not emit it in the first place in this codepath.)
---
I think I bit off too much in #116479; let me try just the scalar case first.
r? `@ghost`
Raise minimum supported Apple OS versions
This implements the proposal to raise the minimum supported Apple OS versions as laid out in the now-completed MCP (https://github.com/rust-lang/compiler-team/issues/556).
As of this PR, rustc and the stdlib now support these versions as the baseline:
- macOS: 10.12 Sierra
- iOS: 10
- tvOS: 10
- watchOS: 5 (Unchanged)
In addition to everything this breaks indirectly, these changes also erase the `armv7-apple-ios` target (currently tier 3) because the oldest supported iOS device now uses ARMv7s. Not sure what the policy around tier3 target removal is but shimming it is not an option due to the linker refusing.
[Per comment](https://github.com/rust-lang/compiler-team/issues/556#issuecomment-1297175073), this requires a FCP to merge. cc `@wesleywiser.`
Enable -Zdrop-tracking-mir by default
This PR enables the `drop-tracking-mir` flag by default. This flag was initially implemented in https://github.com/rust-lang/rust/pull/101692.
This flag computes auto-traits on generators based on their analysis MIR, instead of trying to compute on the HIR body. This removes the need for HIR-based drop-tracking, as we can now reuse the same code to compute generator witness types and to compute generator interior fields.
Add codegen test to guard against VecDeque optimization regression
Very small PR that adds a codegen test to guard against regression for the `VecDeque` optimization addressed in #80836. Ensures that Rustc optimizes away the panic when unwrapping the result of `.get(0)` because of the `!is_empty()` condition.
Use the same DISubprogram for each instance of the same inlined function within a caller
# Issue Details:
The call to `panic` within a function like `Option::unwrap` is translated to LLVM as a `tail call` (as it will never return), when multiple calls to the same function like this are inlined LLVM will notice the common `tail call` block (i.e., loading the same panic string + location info and then calling `panic`) and merge them together.
When merging these instructions together, LLVM will also attempt to merge the debug locations as well, but this fails (i.e., debug info is dropped) as Rust emits a new `DISubprogram` at each inline site thus LLVM doesn't recognize that these are actually the same function and so thinks that there isn't a common debug location.
As an example of this, consider the following program:
```rust
#[no_mangle]
fn add_numbers(x: &Option<i32>, y: &Option<i32>) -> i32 {
let x1 = x.unwrap();
let y1 = y.unwrap();
x1 + y1
}
```
When building for x86_64 Windows using 1.72 it generates (note the lack of `.cv_loc` before the call to `panic`, thus it will be attributed to the same line at the `addq` instruction):
```llvm
.cv_loc 0 1 3 0 # src\lib.rs:3:0
addq $40, %rsp
retq
leaq .Lalloc_f570dea0a53168780ce9a91e67646421(%rip), %rcx
leaq .Lalloc_629ace53b7e5b76aaa810d549cc84ea3(%rip), %r8
movl $43, %edx
callq _ZN4core9panicking5panic17h12e60b9063f6dee8E
int3
```
# Fix Details:
Cache the `DISubprogram` emitted for each inlined function instance within a caller so that this can be reused if that instance is encountered again.
Ideally, we would also deduplicate child scopes and variables, however my attempt to do that with #114643 resulted in asserts when building for Linux (#115156) which would require some deep changes to Rust to fix (#115455).
Instead, when using an inlined function as a debug scope, we will also create a new child scope such that subsequent child scopes and variables do not collide (from LLVM's perspective).
After this change the above assembly now (with <https://reviews.llvm.org/D159226> as well) shows the `panic!` was inlined from `unwrap` in `option.rs` at line 935 into the current function in `lib.rs` at line 0 (line 0 is emitted since it is ambiguous which line to use as there were two inline sites that lead to this same code):
```llvm
.cv_loc 0 1 3 0 # src\lib.rs:3:0
addq $40, %rsp
retq
.cv_inline_site_id 6 within 0 inlined_at 1 0 0
.cv_loc 6 2 935 0 # library\core\src\option.rs:935:0
leaq .Lalloc_5f55955de67e57c79064b537689facea(%rip), %rcx
leaq .Lalloc_e741d4de8cb5801e1fd7a6c6795c1559(%rip), %r8
movl $43, %edx
callq _ZN4core9panicking5panic17hde1558f32d5b1c04E
int3
```
Use `preserve_mostcc` for `extern "rust-cold"`
As experimentation in #115242 has shown looks better than `coldcc`. Notably, clang exposes `preserve_most` (https://clang.llvm.org/docs/AttributeReference.html#preserve-most) but not `cold`, so this change should put us on a better-supported path.
And *don't* use a different convention for cold on Windows, because that actually ends up making things worse. (See comment in the code.)
cc tracking issue #97544
Remove some wasm/emscripten ignores
I'm planning on landing a few PRs like this that remove ignores that aren't required. This just covers mir-opt and codegen tests.
As experimentation in 115242 has shown looks better than `coldcc`.
And *don't* use a different convention for cold on Windows, because that actually ends up making things worse.
cc tracking issue 97544
Revert "Use the same DISubprogram for each instance of the same inline function within the caller"
This reverts commit 687bffa493.
Reverting to resolve ICEs reported on nightly.
cc `@dpaoliello`
Fixes#115156
Fix#115150 by encoding f32 and f64 correctly for cross-language CFI. I
missed changing the encoding for f32 and f64 when I introduced the
integer normalization option in #105452 as integer normalization does
not include floating point. `f32` and `f64` should be always encoded as
`f` and `d` since they are both FFI safe when their representation are
the same (i.e., IEEE 754) for both the Rust compiler and Clang.
Add regression test for not `memcpy`ing padding bytes
Closes#56297
See this comparison: https://rust.godbolt.org/z/jjzfonfcE
I don't have any experience with codegen tests, I hope this is correct
Use the same DISubprogram for each instance of the same inlined function within a caller
# Issue Details:
The call to `panic` within a function like `Option::unwrap` is translated to LLVM as a `tail call` (as it will never return), when multiple calls to the same function like this is inlined LLVM will notice the common `tail call` block (i.e., loading the same panic string + location info and then calling `panic`) and merge them together.
When merging these instructions together, LLVM will also attempt to merge the debug locations as well, but this fails (i.e., debug info is dropped) as Rust emits a new `DISubprogram` at each inline site thus LLVM doesn't recognize that these are actually the same function and so thinks that there isn't a common debug location.
As an example of this when building for x86_64 Windows (note the lack of `.cv_loc` before the call to `panic`, thus it will be attributed to the same line at the `addq` instruction):
```
.cv_loc 0 1 23 0 # src\lib.rs:23:0
addq $40, %rsp
retq
leaq .Lalloc_f570dea0a53168780ce9a91e67646421(%rip), %rcx
leaq .Lalloc_629ace53b7e5b76aaa810d549cc84ea3(%rip), %r8
movl $43, %edx
callq _ZN4core9panicking5panic17h12e60b9063f6dee8E
int3
```
# Fix Details:
Cache the `DISubprogram` emitted for each inlined function instance within a caller so that this can be reused if that instance is encountered again, this also requires caching the `DILexicalBlock` and `DIVariable` objects to avoid creating duplicates.
After this change the above assembly now looks like:
```
.cv_loc 0 1 23 0 # src\lib.rs:23:0
addq $40, %rsp
retq
.cv_inline_site_id 5 within 0 inlined_at 1 0 0
.cv_inline_site_id 6 within 5 inlined_at 1 12 0
.cv_loc 6 2 935 0 # library\core\src\option.rs:935:0
leaq .Lalloc_5f55955de67e57c79064b537689facea(%rip), %rcx
leaq .Lalloc_e741d4de8cb5801e1fd7a6c6795c1559(%rip), %r8
movl $43, %edx
callq _ZN4core9panicking5panic17hde1558f32d5b1c04E
int3
```
We've investigated one reason why debugging information often goes wrong at https://reviews.llvm.org/D152095.
> LLVM can't handle IR where subprogram definitions are nested within DICompositeType when doing LTO builds,
> because there's no good way to cross the CU boundary to insert a nested DISubprogram definition in one CU into a type defined in another CU.
Add a new `compare_bytes` intrinsic instead of calling `memcmp` directly
As discussed in #113435, this lets the backends be the place that can have the "don't call the function if n == 0" logic, if it's needed for the target. (I didn't actually *add* those checks, though, since as I understood it we didn't actually need them on known targets?)
Doing this also let me make it `const` (unstable), which I don't think `extern "C" fn memcmp` can be.
cc `@RalfJung` `@Amanieu`
Nest other codegen test topics
This PR is like rust-lang/rust#114229 in that it mostly pushes codegen tests around, shoving them into their own directories, but because all of the changes are very simple cleanups I pulled them into a separate PR. The other PR might involve actually evaluating the correctness of the test after changes, but here it is mostly a matter of taste. The only "functional" change is deleting a few tests that... hinge on a version of LLVM that we don't support (as of rust-lang/rust#114148 anyways).
I considered a few different ways to group other topics but I feel the question of whether `tests/codegen/{vec,array,slice}` should exist is more subtle than these choices, as it might be better to group such related tests by other topics like bounds check elision, thus I avoided making it.
Nest tests/codegen/sanitizer*.rs tests in sanitizer dir
The sanitizer tests are the largest and most meticulously tested set of tests in tests/codegen. That's good! They all clearly belong to a subject and thus could go in a directory, but are not, instead being placed simply in tests/codegen. That's bad! Fix this by placing them in their own directory and renaming them to be less repetitive after that move.
A few tests are brittle, and embed their filename in the test's checks. This is acceptable for the ones where it is used only two times, but one test embeds the test's mangled filename in the test *over 50 times*! This may have been one of the things discouraging anyone from moving it, and thus from moving the set. Fortunately, I have some knowledge of Itanium mangling (involuntarily), regex, and the FileCheck syntax. With a capturing variable, FileCheck allows us to now move this test around again without diffing it on ~50 lines, while still guaranteeing that the mangled substring is the same each time.
This also clarifies why the substring is repeated a zillion times, instead of being cryptic. They don't call it mangling because the result is pretty and easy to understand, but now it is slightly easier! Yay descriptive variables!
cleanup: remove pointee types
This can't be merged until the oldest LLVM version we support uses opaque pointers, which will be the case after #114148. (Also note `-Cllvm-args="-opaque-pointers=0"` can technically be used in LLVM 15, though I don't think we should support that configuration.)
I initially hoped this would provide some minor perf win, but in https://github.com/rust-lang/rust/pull/105412#issuecomment-1341224450 it had very little impact, so this is only valuable as a cleanup.
As a followup, this will enable #96242 to be resolved.
r? `@ghost`
`@rustbot` label S-blocked
In the basic case, simply do the string substitution.
For one case with many instances, capture the Itanium-
mangled filename and assert its reuse instead.
Combining revisions with only-arch allows specifying
that a test only applies to a handful of targets.
This allows removing a large amount of repetition
in the test suite for tests that do not benefit.
The revisions are suboptimal for this for some tests,
so they aren't preferred in those cases.
Slightly more complicated: also give them appropriate names
that somewhat describe the cases they are trying to cover,
using information from PR chatter in rust-lang/rust#47158
Fix #[inline(always)] on closures with target feature 1.1
Fixes#108655. I think this is the most obvious solution that isn't overly complicated. The comment includes more justification, but I think this is likely better than demoting the `#[inline(always)]` to `#[inline]`, since existing code is unaffected.
Both GCC and Clang write by default a `.comment` section with compiler
information:
```txt
$ gcc -c -xc /dev/null && readelf -p '.comment' null.o
String dump of section '.comment':
[ 1] GCC: (GNU) 11.2.0
$ clang -c -xc /dev/null && readelf -p '.comment' null.o
String dump of section '.comment':
[ 1] clang version 14.0.1 (https://github.com/llvm/llvm-project.git c62053979489ccb002efe411c3af059addcb5d7d)
```
They also implement the `-Qn` flag to avoid doing so:
```txt
$ gcc -Qn -c -xc /dev/null && readelf -p '.comment' null.o
readelf: Warning: Section '.comment' was not dumped because it does not exist!
$ clang -Qn -c -xc /dev/null && readelf -p '.comment' null.o
readelf: Warning: Section '.comment' was not dumped because it does not exist!
```
So far, `rustc` only does it for WebAssembly targets and only
when debug info is enabled:
```txt
$ echo 'fn main(){}' | rustc --target=wasm32-unknown-unknown --emit=llvm-ir -Cdebuginfo=2 - && grep llvm.ident rust_out.ll
!llvm.ident = !{!27}
```
In the RFC part of this PR it was decided to always add
the information, which gets us closer to other popular compilers.
An opt-out flag like GCC and Clang may be added later on if deemed
necessary.
Implementation-wise, this covers both `ModuleLlvm::new()` and
`ModuleLlvm::new_metadata()` cases by moving the addition to
`context::create_module` and adds a few test cases.
ThinLTO also sees the `llvm.ident` named metadata duplicated (in
temporary outputs), so this deduplicates it like it is done for
`wasm.custom_sections`. The tests also check this duplication does
not take place.
Signed-off-by: Miguel Ojeda <ojeda@kernel.org>
Turn copy into moves during DSE.
Dead store elimination computes whether removing a direct store to an unborrowed place is allowed.
Where removing a store is allowed, writing `uninit` is too.
This means that we can use this pass to transform `copy` operands into `move` operands. This is only interesting in call terminators, so we only handle those.
Special care is taken for the `use_both(_1, _1)` case:
- moving the second argument is ok, as `_1` is not live after the call;
- moving the first argument is not, as the second argument reads `_1`.
Fixes#75993
Fixes https://github.com/rust-lang/rust/issues/108068
r? `@RalfJung`
cc `@JakobDegen`
Resurrect: rustc_target: Add alignment to indirectly-passed by-value types, correcting the alignment of byval on x86 in the process.
Same as #111551, which I [accidentally closed](https://github.com/rust-lang/rust/pull/111551#issuecomment-1571222612) :/
---
This resurrects PR #103830, which has sat idle for a while.
Beyond #103830, this also:
- fixes byval alignment for types containing vectors on Darwin (see `tests/codegen/align-byval-vector.rs`)
- fixes byval alignment for overaligned types on x86 Windows (see `tests/codegen/align-byval.rs`)
- fixes ABI for types with 128bit requested alignment on ARM64 Linux (see `tests/codegen/aarch64-struct-align-128.rs`)
r? `@nikic`
---
`@pcwalton's` original PR description is reproduced below:
Commit 88e4d2c from five years ago removed
support for alignment on indirectly-passed arguments because of problems with
the `i686-pc-windows-msvc` target. Unfortunately, the `memcpy` optimizations I
recently added to LLVM 16 depend on this to forward `memcpy`s. This commit
attempts to fix the problems with `byval` parameters on that target and now
correctly adds the `align` attribute.
The problem is summarized in [this comment] by `@eddyb.` Briefly, 32-bit x86 has
special alignment rules for `byval` parameters: for the most part, their
alignment is forced to 4. This is not well-documented anywhere but in the Clang
source. I looked at the logic in Clang `TargetInfo.cpp` and tried to replicate
it here. The relevant methods in that file are
`X86_32ABIInfo::getIndirectResult()` and
`X86_32ABIInfo::getTypeStackAlignInBytes()`. The `align` parameter attribute
for `byval` parameters in LLVM must match the platform ABI, or miscompilations
will occur. Note that this doesn't use the approach suggested by eddyb, because
I felt it was overkill to store the alignment in `on_stack` when special
handling is really only needed for 32-bit x86.
As a side effect, this should fix#80127, because it will make the `align`
parameter attribute for `byval` parameters match the platform ABI on LLVM
x86-64.
[this comment]: #80822 (comment)
alignment of `byval` on x86 in the process.
Commit 88e4d2c291 from five years ago removed
support for alignment on indirectly-passed arguments because of problems with
the `i686-pc-windows-msvc` target. Unfortunately, the `memcpy` optimizations I
recently added to LLVM 16 depend on this to forward `memcpy`s. This commit
attempts to fix the problems with `byval` parameters on that target and now
correctly adds the `align` attribute.
The problem is summarized in [this comment] by @eddyb. Briefly, 32-bit x86 has
special alignment rules for `byval` parameters: for the most part, their
alignment is forced to 4. This is not well-documented anywhere but in the Clang
source. I looked at the logic in Clang `TargetInfo.cpp` and tried to replicate
it here. The relevant methods in that file are
`X86_32ABIInfo::getIndirectResult()` and
`X86_32ABIInfo::getTypeStackAlignInBytes()`. The `align` parameter attribute
for `byval` parameters in LLVM must match the platform ABI, or miscompilations
will occur. Note that this doesn't use the approach suggested by eddyb, because
I felt it was overkill to store the alignment in `on_stack` when special
handling is really only needed for 32-bit x86.
As a side effect, this should fix#80127, because it will make the `align`
parameter attribute for `byval` parameters match the platform ABI on LLVM
x86-64.
[this comment]: https://github.com/rust-lang/rust/pull/80822#issuecomment-829985417
Add Tests for native wasm exceptions
### Motivation
In PR #111322, I added support for native WASM exceptions. I was asked by ``@davidtwco`` to add some tests for it in a follow up PR, which seems like a very good idea.
This PR adds three tests for this feature:
* codegen: ensure the correct LLVM instructions are used
* assembly: ensure the correct WASM instructions are used
* run-make: ensure the exception handling works; the WASM code is run using a small nodejs script which demonstrates the exception handling
### Complications
There are a few changes beside adding the tests, which were necessary
* Tests for the wasm32-unknown-unknown target are (as far as I know) only run on `test-various`. Its docker image uses nodejs-15, which is very old. Experimental support for wasm-exceptions was added in nodejs16. In nodejs 18.12 (LTS), they are stable.
- --> increase nodejs to 18.12 in `test-various`
* codegen/assembly tests are not performed for the wasm32-unknown-unknown target yet
- --> add those to `test-various` as well
Due to the last point, some tests are run which have not run before (assembly+codegen tests for wasm32-unknown-unknown). I added `// ignore wasm32-bare` for those which failed
### Local testing
I run all tests locally using both `test-various` and `wasm32`. As far as I know, none of the other systems run any test for wasm32 targets.
Various impl trait in assoc tys cleanups
r? `@compiler-errors`
All commits except for the last are pure refactorings. 274dab5bd658c97886a8987340bf50ae57900c39 allows struct fields to participate in deciding whether a function has an opaque in its signature.
best reviewed commit by commit
[libs] Simplify `unchecked_{shl,shr}`
There's no need for the `const_eval_select` dance here. And while I originally wrote the `.try_into().unwrap_unchecked()` implementation here, it's kinda a mess in MIR -- this new one is substantially simpler, as shown by the old one being above the inlining threshold but the new one being below it in the `mir-opt/inline/unchecked_shifts` tests.
We don't need `u32::checked_shl` doing a dance through both `Result` *and* `Option` 🙃
There's no need for the `const_eval_select` dance here. And while I originally wrote the `.try_into().unwrap_unchecked()` implementation here, it's kinda a mess in MIR -- this new one is substantially simpler, as shown by the old one being above the inlining threshold but the new one being below it.
Fix codegen test suite for bare-metal-like targets
For Ferrocene I needed to run the test suite for custom target with no unwinding and static relocation. Running the tests uncovered ~20 failures due to the test suite not accounting for these options. This PR fixes them by:
* Fixing `CHECK`s to account for functions having extra LLVM IR attributes (in this case `nounwind`).
* Fixing `CHECK`s to account for the `dso_local` LLVM IR modifier, which is [added to every item when relocation is static](f3d597b31c/compiler/rustc_codegen_llvm/src/mono_item.rs (L139-L142)).
* Fixing `CHECK`s to account for missing `uwtables` attributes.
* Added the `needs-unwind` attributes for tests that are designed to check unwinding.
There is no part of Rust CI that checks this unfortunately, and testing whether the PR works locally is kinda hard because you need a target with std enabled but no unwinding and static relocations. Still, this works in my local testing, and if future PRs accidentally break this Ferrocene will take care of sending followup PRs.
Enable ScalarReplacementOfAggregates in optimized builds
Like MatchBranchSimplification, this pass is known to produce significant runtime improvements in Cranelift artifacts, and I believe based on the perf runs here that the primary effect of this pass is to empower MatchBranchSimplification. ScalarReplacementOfAggregates on its own has little effect on anything, but when this was rebased up to include https://github.com/rust-lang/rust/pull/112001 we started seeing significant and majority-positive results.
Based on the fact that we see most of the regressions in debug builds (https://github.com/rust-lang/rust/pull/112002#issuecomment-1566270144) and some rather significant ones in cycles and wall time, I'm only enabling this in optimized builds at the moment.
These tend to have special handling in a bunch of places anyway, so the variant helps remember that. And I think it's easier to grok than non-Scalar Aggregates sometimes being `Immediates` (like I got wrong and caused 109992). As a minor bonus, it means we don't need to generate poison LLVM values for them to pass around in `OperandValue::Immediate`s.
Only rewrite valtree-constants to patterns and keep other constants opaque
Now that we can reliably fall back to comparing constants with `PartialEq::eq` to the match scrutinee, we can
1. eagerly try to convert constants to valtrees
2. then deeply convert the valtree to a pattern
3. if the to-valtree conversion failed, create an "opaque constant" pattern.
This PR specifically avoids any behavioral changes or major cleanups. What we can now do as follow ups is
* move the two remaining call sites to `destructure_mir_constant` off that query
* make valtree to pattern conversion infallible
* this needs to be done after careful analysis of the effects. There may be user visible changes from that.
based on https://github.com/rust-lang/rust/pull/111768
Optimize scalar and scalar pair representations loaded from ByRef in llvm
in https://github.com/rust-lang/rust/pull/105653 I noticed that we were generating suboptimal LLVM IR if we had a `ConstValue::ByRef` that could be represented by a `ScalarPair`. Before https://github.com/rust-lang/rust/pull/105653 this is probably rare, but after it, every slice will go down this suboptimal code path that requires LLVM to untangle a bunch of indirections and translate static allocations that are only used once to read a scalar pair from.
`[T; N]::zip` is "eager" but most zips are mapped.
This causes poor optimization in generated code.
This is a fundamental design issue and "zip" is
"prime real estate" in terms of function names,
so let's free it up again.
Adds support for LLVM [SafeStack] which provides backward edge control
flow protection by separating the stack into two parts: data which is
only accessed in provable safe ways is allocated on the normal stack
(the "safe stack") and all other data is placed in a separate allocation
(the "unsafe stack").
SafeStack support is enabled by passing `-Zsanitizer=safestack`.
[SafeStack]: https://clang.llvm.org/docs/SafeStack.html
Support #[global_allocator] without the allocator shim
This makes it possible to use liballoc/libstd in combination with `--emit obj` if you use `#[global_allocator]`. This is what rust-for-linux uses right now and systemd may use in the future. Currently they have to depend on the exact implementation of the allocator shim to create one themself as `--emit obj` doesn't create an allocator shim.
Note that currently the allocator shim also defines the oom error handler, which is normally required too. Once `#![feature(default_alloc_error_handler)]` becomes the only option, this can be avoided. In addition when using only fallible allocator methods and either `--cfg no_global_oom_handling` for liballoc (like rust-for-linux) or `--gc-sections` no references to the oom error handler will exist.
To avoid this feature being insta-stable, you will have to define `__rust_no_alloc_shim_is_unstable` to avoid linker errors.
(Labeling this with both T-compiler and T-lang as it originally involved both an implementation detail and had an insta-stable user facing change. As noted above, the `__rust_no_alloc_shim_is_unstable` symbol requirement should prevent unintended dependence on this unstable feature.)
CFI: Fix encode_region: unexpected ReEarlyBound(0, 'a)
Fixes#111515 and complements #106547 by adding support for encoding early bound regions and also excluding projections when transforming trait objects' traits into their identities before emitting type checks.
Fixes#111515 and complements #106547 by adding support for encoding
early bound regions and also excluding projections when transforming
trait objects' traits into their identities before emitting type checks.
Fix some issues with folded AArch64 features
In #91608 the `fp` feature was removed for AArch64 and folded into the `neon` feature, however disabling the `neon` feature doesn't actually disable the `fp` feature. If my understanding on that thread is correct it should do.
While doing this, I also noticed that disabling some features would disable features that it shouldn't. For instance enabling `sve` will enable `neon`, however, when disabling `sve` it would then also disable `neon`, I wouldn't expect disabling `sve` to also disable `neon`.
cc `@workingjubilee`
When compiling with panic=abort (or using a target that doesn't have
unwinding support), the compiler adds the "nounwind" attribute to
functions. This results in a different LLVM IR, which results in a #NNN
added after the function name:
tail call void @bar() #13, !dbg !467
attributes #13 = { nounwind }
...instead of:
tail call void @bar(), !dbg !467
This commit changes the matchers to swallow the #NNN, as it's not needed
for these specific tests.
Rollup of 6 pull requests
Successful merges:
- #111461 (Fix symbol conflict diagnostic mistakenly being shown instead of missing crate diagnostic)
- #111579 (Also assume wrap-around discriminants in `as` MIR building)
- #111704 (Remove return type sized check hack from hir typeck)
- #111853 (Check opaques for mismatch during writeback)
- #111854 (rustdoc: clean up `settings.css`)
- #111860 (Don't ICE if method receiver fails to unify with `arbitrary_self_types`)
r? `@ghost`
`@rustbot` modify labels: rollup
[rustc_ty_utils] Treat `drop_in_place`'s *mut argument like &mut when adding LLVM attributes
This resurrects PR #103614, which has sat idle for a while.
This could probably use a new perf run, since we're on a new LLVM version now.
r? `@oli-obk`
cc `@RalfJung`
---
LLVM can make use of the `noalias` parameter attribute on the parameter to `drop_in_place` in areas like argument promotion. Because the Rust compiler fully controls the code for `drop_in_place`, it can soundly deduce parameter attributes on it.
In #103957, Miri was changed to retag `drop_in_place`'s argument as if it was `&mut`, matching this change.
Fix duplicate `arcinner_layout_for_value_layout` calls when using the uninit `Arc` constructors
What this fixes is the duplicate calls to `arcinner_layout_for_value_layout` seen here: https://godbolt.org/z/jr5Gxozhj
The issue was discovered alongside #111603 but is otherwise unrelated to the duplicate `alloca`s, which remain unsolved. Everything I tried to solve said main issue has failed.
As for the duplicate layout calculations, I also tried slapping `#[inline]` and `#[inline(always)]` on everything in sight but the only thing that worked in the end is to dedup the calls by hand.
Rather than returning an array of features from to_llvm_features, return a structure that contains
the dependencies. This also contains metadata on how the features depend on each other to allow for
the correct enabling and disabling.
Some features that are tied together only make sense to be folded
together when enabling the feature. For example on AArch64 sve and
neon are tied together, however it doesn't make sense to disable neon
when disabling sve.
CFI: Fix encode_ty: unexpected Param(B/#1)
Fixes#111510 and complements #106547 by adding support for encoding type parameters and also by transforming trait objects' traits into their identities before emitting type checks.
The CHECK, -NOT, -SAME pattern ensures that the `CHECK-NOT: noalias`
is limited to only one line, and won't match unrelated lines further
down in the file.
Explicit drop call added to preserve the `foo` argument name, since
names of unused arguments are not preserved.
We've done measurements with Miri and have determined that `noalias` shouldn't
break code. The requirements that allow us to add dereferenceable and align
have been long documented in the standard library documentation.
LLVM can make use of the `noalias` parameter attribute on the parameter to
`drop_in_place` in areas like argument promotion. Because the Rust compiler
fully controls the code for `drop_in_place`, it can soundly deduce parameter
attributes on it. In the case of a value that has a programmer-defined Drop
implementation, we know that the first thing `drop_in_place` will do is pass a
pointer to the object to `Drop::drop`. `Drop::drop` takes `&mut`, so it must be
guaranteed that there are no pointers to the object upon entering that
function. Therefore, it should be safe to mark `noalias` there.
With this patch, we mark `noalias` only when the type is a value with a
programmer-defined Drop implementation. This is possibly overly conservative,
but I thought that proceeding cautiously was best in this instance.
Fixes#111510 and complements #106547 by adding support for encoding
type parameters and also by transforming trait objects' traits into
their identities before emitting type checks.
Stop checking for the absence of something that doesn't exist
A couple of codegen tests are doing
```
// CHECK-NOT: slice_index_len_fail
```
However, that function no longer exists: [the only places](https://github.com/search?q=repo%3Arust-lang%2Frust+slice_index_len_fail&type=code) it occurs in the repo are in those tests.
So this PR updates the tests to check for the absense of the functions that are actually used today to panic for out-of-bounds indexing.
allow mutating function args through `&raw const`
Fixes https://github.com/rust-lang/rust/issues/111502 by "turning off the sketchy optimization while we figure out if this is ok", like `@JakobDegen` said.
The first commit in this PR removes some suspicious looking logic from the same method, but should have no functional changes, since it doesn't modify the `context` outside of the method. Best reviewed commit by commit.
r? opsem
A couple of codegen tests are doing
```
// CHECK-NOT: slice_index_len_fail
```
However, that function no longer exists: [the only places](https://github.com/search?q=repo%3Arust-lang%2Frust+slice_index_len_fail&type=code) it occurs in the repo are in those tests.
So this PR updates the tests to check for the absense of the functions that are actually used today to panic for out-of-bounds indexing.
CFI: Fix SIGILL reached via trait objects
Fix#106547 by transforming the concrete self into a reference to a trait object before emitting type metadata identifiers for trait methods.
You will need to add the following as replacement for the old __rust_*
definitions when not using the alloc shim.
#[no_mangle]
static __rust_no_alloc_shim_is_unstable: u8 = 0;
vec-shrink-panik: update expectations to work on LLVM 17
For some reason, the called function is `cleanup` on LLVM 17 instead of `filter`.
r? `@Amanieu`
Remove some `assume`s from slice iterators that don't do anything
Because the start pointer is iterators is already a `NonNull`, we emit the appropriate `!nonnull` metadata when loading the pointer to tell LLVM that it's non-null.
Probably the best way to see that it's the metadata that's important (and not the `assume`) is to observe that LLVM actually *removes* the `assume` from the optimized IR: <https://rust.godbolt.org/z/KhE6G963n>.
(I also checked that, yes, the if-not-ZST `assume` on `end` is still doing something: it's how there's a `!nonnull` metadata on its load, even though it's an ordinary raw pointer. The codegen test added in this PR fails if the other `assume` is removed.)
Rollup of 6 pull requests
Successful merges:
- #104070 (Prevent aborting guard from aborting the process in a forced unwind)
- #109410 (Introduce `AliasKind::Inherent` for inherent associated types)
- #111004 (Migrate `mir_transform` to translatable diagnostics)
- #111118 (Suggest struct when we get colon in fileds in enum)
- #111170 (Diagnostic args are still args if they're documented)
- #111354 (Fix miscompilation when calling default methods on `Future`)
Failed merges:
r? `@ghost`
`@rustbot` modify labels: rollup
