I added `PlaceValue` in 123775, but kept that one line-by-line simple because it touched so many places.
This goes through to add more helpers & docs, and change some `PlaceRef` to `PlaceValue` where the type didn't need to be included.
No behaviour changes.
Avoid `alloca`s in codegen for simple `mir::Aggregate` statements
The core idea here is to remove the abstraction penalty of simple newtypes in codegen.
Even something simple like constructing a
```rust
#[repr(transparent)] struct Foo(u32);
```
forces an `alloca` to be generated in nightly right now.
Certainly LLVM can optimize that away, but it would be nice if it didn't have to.
Quick example:
```rust
#[repr(transparent)]
pub struct Transparent32(u32);
#[no_mangle]
pub fn make_transparent(x: u32) -> Transparent32 {
let a = Transparent32(x);
a
}
```
on nightly we produce <https://rust.godbolt.org/z/zcvoM79ae>
```llvm
define noundef i32 `@make_transparent(i32` noundef %x) unnamed_addr #0 {
%a = alloca i32, align 4
store i32 %x, ptr %a, align 4
%0 = load i32, ptr %a, align 4, !noundef !3
ret i32 %0
}
```
but after this PR we produce
```llvm
define noundef i32 `@make_transparent(i32` noundef %x) unnamed_addr #0 {
start:
ret i32 %x
}
```
(even before the optimizer runs).
Stop using LLVM struct types for alloca
The alloca type has no semantic meaning, only the size (and alignment, but we specify it explicitly) matter. Using `[N x i8]` is a more direct way to specify that we want `N` bytes, and avoids relying on LLVM's struct layout. It is likely that a future LLVM version will change to an untyped alloca representation.
Split out from #121577.
r? `@ghost`
I added this back in 111999, but I no longer think it's a good idea
- It had to get scaled back to only power-of-two things to not break a bunch of targets
- LLVM seems to be getting better at memcpy removal anyway
- Introducing vector instructions has seemed to sometimes (115515) make autovectorization worse
So this removes it from the codegen crates entirely, and instead just tries to use <https://doc.rust-lang.org/nightly/nightly-rustc/rustc_codegen_ssa/traits/builder/trait.BuilderMethods.html#method.typed_place_copy> instead of direct `memcpy` so things will still use load/store for immediates.
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.
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.
compile-time evaluation: detect writes through immutable pointers
This has two motivations:
- it unblocks https://github.com/rust-lang/rust/pull/116745 (and therefore takes a big step towards `const_mut_refs` stabilization), because we can now detect if the memory that we find in `const` can be interned as "immutable"
- it would detect the UB that was uncovered in https://github.com/rust-lang/rust/pull/117905, which was caused by accidental stabilization of `copy` functions in `const` that can only be called with UB
When UB is detected, we emit a future-compat warn-by-default lint. This is not a breaking change, so completely in line with [the const-UB RFC](https://rust-lang.github.io/rfcs/3016-const-ub.html), meaning we don't need t-lang FCP here. I made the lint immediately show up for dependencies since it is nearly impossible to even trigger this lint without `const_mut_refs` -- the accidentally stabilized `copy` functions are the only way this can happen, so the crates that popped up in #117905 are the only causes of such UB (in the code that crater covers), and the three cases of UB that we know about have all been fixed in their respective crates already.
The way this is implemented is by making use of the fact that our interpreter is already generic over the notion of provenance. For CTFE we now use the new `CtfeProvenance` type which is conceptually an `AllocId` plus a boolean `immutable` flag (but packed for a more efficient representation). This means we can mark a pointer as immutable when it is created as a shared reference. The flag will be propagated to all pointers derived from this one. We can then check the immutable flag on each write to reject writes through immutable pointers.
I just hope perf works out.
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.
LLVM can usually optimize these away, but especially for things like transmutes of newtypes it's silly to generate the `alloc`+`store`+`load` at all when it's actually a nop at LLVM level.
In cases where it is legal, we should prefer poison values over
undef values.
This replaces undef with poison for aggregate construction and
for uninhabited types. There are more places where we can likely
use poison, but I wanted to stay conservative to start with.
In particular the aggregate case is important for newer LLVM
versions, which are not able to handle an undef base value during
early optimization due to poison-propagation concerns.
There are several indications that we should not ZST as a ScalarInt:
- We had two ways to have ZST valtrees, either an empty `Branch` or a `Leaf` with a ZST in it.
`ValTree::zst()` used the former, but the latter could possibly arise as well.
- Likewise, the interpreter had `Immediate::Uninit` and `Immediate::Scalar(Scalar::ZST)`.
- LLVM codegen already had to special-case ZST ScalarInt.
So instead add new ZST variants to those types that did not have other variants
which could be used for this purpose.
