| Age | Commit message (Collapse) | Author | Lines |
|---|
|
r=oli-obk
Added llvm lifetime annotations to function call argument temporaries.
The goal of this change is to ensure that llvm will do stack slot
optimization on these temporaries. This ensures that in code like:
```rust
const A: [u8; 1024] = [0; 1024];
fn copy_const() {
f(A);
f(A);
}
```
we only use 1024 bytes of stack space, instead of 2048 bytes.
I am new to developing for the rust compiler, and as such not entirely sure, but I believe this should be sufficient to close #98156.
Also, this does not contain a test case to ensure this keeps working, primarily because I am not sure how to go about testing this. I would love some suggestions as to how that could be approached.
|
|
|
|
Make RwLockReadGuard covariant
Hi, first time contributor here, if anything is not as expected, please let me know.
`RwLockReadGoard`'s type constructor is invariant. Since it behaves like a smart pointer to an immutable reference, there is no reason that it should not be covariant. Take e.g.
```
fn test_read_guard_covariance() {
fn do_stuff<'a>(_: RwLockReadGuard<'_, &'a i32>, _: &'a i32) {}
let j: i32 = 5;
let lock = RwLock::new(&j);
{
let i = 6;
do_stuff(lock.read().unwrap(), &i);
}
drop(lock);
}
```
where the compiler complains that &i doesn't live long enough. If `RwLockReadGuard` is covariant, then the above code is accepted because the lifetime can be shorter than `'a`.
In order for `RwLockReadGuard` to be covariant, it can't contain a full reference to the `RwLock`, which can never be covariant (because it exposes a mutable reference to the underlying data structure). By reducing the data structure to the required pieces of `RwLock`, the rest falls in place.
If there is a better way to do a test that tests successful compilation, please let me know.
Fixes #80392
|
|
Remove dereferencing of Box from codegen
Through #94043, #94414, #94873, and #95328, I've been fixing issues caused by Box being treated like a pointer when it is not a pointer. However, these PRs just introduced special cases for Box. This PR removes those special cases and instead transforms a deref of Box into a deref of the pointer it contains.
Hopefully, this is the end of the Box<T, A> ICEs.
|
|
|
|
|
|
|
|
|
|
Use unchecked mul to compute slice sizes
This allows LLVM to realize that `slice.len() > 0` iff `slice.len() * size_of::<T>() > 0`, allowing a branch on the latter to be folded into the former when dropping vecs and boxed slices, in some cases.
Fixes (partially) #96497
|
|
...since slice sizes can't signed wrap
see https://doc.rust-lang.org/std/slice/fn.from_raw_parts.html
> The total size len * mem::size_of::<T>() of the slice must be no larger than isize::MAX.
|
|
The offset in the llvm.type.checked.load intrinsic differs on 32 bit platforms
|
|
|
|
Specify DWARF alignment in bits, not bytes.
In DWARF, alignment of types is specified in bits, as is made clear by the
parameter name `AlignInBits`. However, `rustc` was incorrectly passing a byte
alignment. This commit fixes that.
This was noticed in upstream LLVM when I tried to check in a test consisting of
LLVM IR generated from `rustc` and it triggered assertions [1].
[1]: https://reviews.llvm.org/D126835
|
|
In DWARF, alignment of types is specified in bits, as is made clear by the
parameter name `AlignInBits`. However, `rustc` was incorrectly passing a byte
alignment. This commit fixes that.
This was noticed in upstream LLVM when I tried to check in a test consisting of
LLVM IR generated from `rustc` and it triggered assertions [1].
[1]: https://reviews.llvm.org/D126835
|
|
Add support for emitting functions with `coldcc` to LLVM
The eventual goal is to try using this for things like the internal panicking stuff, to see whether it helps.
|
|
|
|
The eventual goal is to try using this for things like the internal panicking stuff, to see whether it helps.
|
|
|
|
|
|
Use pointers in `cell::{Ref,RefMut}` to avoid `noalias`
When `Ref` and `RefMut` were based on references, they would get LLVM `noalias` attributes that were incorrect, because that alias guarantee is only true until the guard drops. A `&RefCell` on the same value can get a new borrow that aliases the previous guard, possibly leading to miscompilation. Using `NonNull` pointers in `Ref` and `RefCell` avoids `noalias`.
Fixes the library side of #63787, but we still might want to explore language solutions there.
|
|
|
|
Add tmm_reg clobbers
This adds support for naming the 8 tile registers from intel AMX as clobbers from `asm!` invocations on x86_64 (only). It does not add the registers as input or output operands.
|
|
Prevent unwinding when `-C panic=abort` is used regardless declared ABI
Ensures that Rust code will abort with `-C panic=abort` regardless ABI used.
```rust
extern "C-unwind" {
fn may_unwind();
}
// Will be nounwind with `-C panic=abort`, despite `C-unwind` ABI.
pub unsafe extern "C-unwind" fn rust_item_that_can_unwind() {
may_unwind();
}
```
Current behaviour is that unwind will propagate through. While the current behaviour won't cause unsoundness it is inconsistent with the text reading of [RFC2945](https://rust-lang.github.io/rfcs/2945-c-unwind-abi.html).
I tweaked `fn_can_unwind` instead of tweaking `AbortUnwindingCalls` because this approach would allow Rust (non-direct) callers to also see that this function is nounwind, so it can prevent excessive landing pads generation.
For more discussions: https://rust-lang.zulipchat.com/#narrow/stream/210922-project-ffi-unwind/topic/soundness.20in.20mixed.20panic.20mode.
cc `@alexcrichton,` `@BatmanAoD`
r? `@Amanieu`
`@rustbot` label: T-compiler T-lang F-c_unwind
|
|
|
|
|
|
|
|
Fix `array::IntoIter::fold` to use the optimized `Range::fold`
It was using `Iterator::by_ref` in the implementation, which ended up pessimizing it enough that, for example, it didn't vectorize when we tried it in the <https://rust-lang.zulipchat.com/#narrow/stream/257879-project-portable-simd/topic/Reducing.20sum.20into.20wider.20types> conversation.
Demonstration that the codegen test doesn't pass on the current nightly: <https://rust.godbolt.org/z/Taxev5eMn>
|
|
|
|
|
|
|
|
Like we have `add`/`sub` which are the `usize` version of `offset`, this adds the `usize` equivalent of `offset_from`. Like how `.add(d)` replaced a whole bunch of `.offset(d as isize)`, you can see from the changes here that it's fairly common that code actually knows the order between the pointers and *wants* a `usize`, not an `isize`.
As a bonus, this can do `sub nuw`+`udiv exact`, rather than `sub`+`sdiv exact`, which can be optimized slightly better because it doesn't have to worry about negatives. That's why the slice iterators weren't using `offset_from`, though I haven't updated that code in this PR because slices are so perf-critical that I'll do it as its own change.
This is an intrinsic, like `offset_from`, so that it can eventually be allowed in CTFE. It also allows checking the extra safety condition -- see the test confirming that CTFE catches it if you pass the pointers in the wrong order.
|
|
|
|
|
|
|
|
|
|
Drop support for legacy PM with LLVM 15
LLVM 15 already removes some of the legacy PM APIs we're using. This patch forces use of NewPM with LLVM 15 (with `-Z new-llvm-pass-manager=no` throwing a warning) and stubs out various FFI methods with a report_fatal_error on LLVM 15.
For LLVMPassManagerBuilderPopulateLTOPassManager() I went with adding our own wrapper, as the alternative would be to muck about with weak symbols, which seems to be non-trivial as far as cross-platform support is concerned (std has `weak!` for this purpose, but only as an internal utility.)
Fixes #96072.
Fixes #96362.
|
|
Check if call return type is visibly uninhabited when building MIR
The main motivation behind the change is to expose information about diverging
calls to the generator transform and match the precision of drop range tracking
which already understands that call expressions with visibly uninhabited types
diverges.
This change should also accept strictly more programs than before. That is
programs that were previously rejected due to errors raised by control-flow
sensitive checks in a code that is no longer considered reachable.
Fixes #93161.
|
|
The linked issue has an accurate description of the situation,
drop the inaccurate comment.
|
|
asm: Add a kreg0 register class on x86 which includes k0
Previously we only exposed a kreg register class which excludes the k0
register since it can't be used in many instructions. However k0 is a
valid register and we need to have a way of marking it as clobbered for
clobber_abi.
Fixes #94977
|
|
Previously we only exposed a kreg register class which excludes the k0
register since it can't be used in many instructions. However k0 is a
valid register and we need to have a way of marking it as clobbered for
clobber_abi.
Fixes #94977
|
|
|
|
Add codegen tests for additional cases where noop iterators get optimized away
Optimizations have improved over time and now LLVM manages to optimize more in-place-collect noop-iterators to O(1) functions. This updates the codegen test to match.
Many but not all cases reported in #79308 work now.
|
|
Speed up Vec::clear().
Currently it just calls `truncate(0)`. `truncate()` is (a) not marked as
`#[inline]`, and (b) more general than needed for `clear()`.
This commit changes `clear()` to do the work itself. This modest change
was first proposed in rust-lang#74172, where the reviewer rejected it because
there was insufficient evidence that `Vec::clear()`'s performance
mattered enough to justify the change. Recent changes within rustc have
made `Vec::clear()` hot within `macro_parser.rs`, so the change is now
clearly worthwhile.
Although it doesn't show wins on CI perf runs, this seems to be because they
use PGO. But not all platforms currently use PGO. Also, local builds don't use
PGO, and `truncate` sometimes shows up in an over-represented fashion in local
profiles. So local profiling will be made easier by this change.
Note that this will also benefit `String::clear()`, because it just
calls `Vec::clear()`.
Finally, the commit removes the `vec-clear.rs` codegen test. It was
added in #52908. From before then until now, `Vec::clear()` just called
`Vec::truncate()` with a zero length. The body of Vec::truncate() has
changed a lot since then. Now that `Vec::clear()` is doing actual work
itself, and not just calling `Vec::truncate()`, it's not surprising that
its generated code includes a load and an icmp. I think it's reasonable
to remove this test.
r? `@m-ou-se`
|
|
|
|
Optimization have improved over time and now LLVM manages to optimize more
in-place-collect noop-iterators to O(1) functions. This updates the codegen test to match.
|
|
Currently it just calls `truncate(0)`. `truncate()` is (a) not marked as
`#[inline]`, and (b) more general than needed for `clear()`.
This commit changes `clear()` to do the work itself. This modest change
was first proposed in rust-lang#74172, where the reviewer rejected it because
there was insufficient evidence that `Vec::clear()`'s performance
mattered enough to justify the change. Recent changes within rustc have
made `Vec::clear()` hot within `macro_parser.rs`, so the change is now
clearly worthwhile.
Although it doesn't show wins on CI perf runs, this seems to be because they
use PGO. But not all platforms currently use PGO. Also, local builds don't use
PGO, and `truncate` sometimes shows up in an over-represented fashion in local
profiles. So local profiling will be made easier by this change.
Note that this will also benefit `String::clear()`, because it just
calls `Vec::clear()`.
Finally, the commit removes the `vec-clear.rs` codegen test. It was
added in #52908. From before then until now, `Vec::clear()` just called
`Vec::truncate()` with a zero length. The body of Vec::truncate() has
changed a lot since then. Now that `Vec::clear()` is doing actual work
itself, and not just calling `Vec::truncate()`, it's not surprising that
its generated code includes a load and an icmp. I think it's reasonable
to remove this test.
|
|
Fix miscompilation of inline assembly with outputs in cases where we emit an invoke instead of call instruction.
We ran into this bug where rustc would segfault while trying to compile certain uses of inline assembly.
Here is a simple repro that demonstrates the issue:
```rust
#![feature(asm_unwind)]
fn main() {
let _x = String::from("string here just cause we need something with a non-trivial drop");
let foo: u64;
unsafe {
std::arch::asm!(
"mov {}, 1",
out(reg) foo,
options(may_unwind)
);
}
println!("{}", foo);
}
```
([playground link](https://play.rust-lang.org/?version=nightly&mode=debug&edition=2021&gist=7d6641e83370d2536a07234aca2498ff))
But crucially `feature(asm_unwind)` is not actually needed and this can be triggered on stable as a result of the way async functions/generators are handled in the compiler. e.g.:
```rust
extern crate futures; // 0.3.21
async fn bar() {
let foo: u64;
unsafe {
std::arch::asm!(
"mov {}, 1",
out(reg) foo,
);
}
println!("{}", foo);
}
fn main() {
futures::executor::block_on(bar());
}
```
([playground link](https://play.rust-lang.org/?version=stable&mode=debug&edition=2021&gist=1c7781c34dd4a3e80ae4bd936a0c82fc))
An example of the incorrect LLVM generated:
```llvm
bb1: ; preds = %start
%1 = invoke i64 asm sideeffect alignstack inteldialect unwind "mov ${0:q}, 1", "=&r,~{dirflag},~{fpsr},~{flags},~{memory}"()
to label %bb2 unwind label %cleanup, !srcloc !9
store i64 %1, i64* %foo, align 8
bb2:
[...snip...]
```
The store should not be placed after the asm invoke but rather should be in the normal control flow basic block (`bb2` in this case).
[Here](https://gist.github.com/luqmana/be1af5b64d2cda5a533e3e23a7830b44) is a writeup of the investigation that lead to finding this.
|
|
|
|
|
|
It was using `Iterator::by_ref` in the implementation, which ended up pessimizing it enough that, for example, it didn't vectorize when we tried it in the <https://rust-lang.zulipchat.com/#narrow/stream/257879-project-portable-simd/topic/Reducing.20sum.20into.20wider.20types> conversation.
Demonstration that the codegen test doesn't pass on the current nightly: <https://rust.godbolt.org/z/Taxev5eMn>
|
