Because `PassMode::Cast` is by far the largest variant, but is
relatively rare.
This requires making `PassMode` not impl `Copy`, and `Clone` is no
longer necessary. This causes lots of sigil adjusting, but nothing very
notable.
Improve size assertions.
- For any file with four or more size assertions, move them into a
separate module (as is already done for `hir.rs`).
- Add some more for AST nodes and THIR nodes.
- Put the `hir.rs` ones in alphabetical order.
r? `@lqd`
- For any file with four or more size assertions, move them into a
separate module (as is already done for `hir.rs`).
- Add some more for AST nodes and THIR nodes.
- Put the `hir.rs` ones in alphabetical order.
interpret, ptr_offset_from: refactor and test too-far-apart check
We didn't have any tests for the "too far apart" message, and indeed that check mostly relied on the in-bounds check and was otherwise probably not entirely correct... so I rewrote that check, and it is before the in-bounds check so we can test it separately.
interpret: rename Tag/PointerTag to Prov/Provenance
We were pretty inconsistent with calling this the "tag" vs the "provenance" of the pointer; I think we should consistently call it "provenance".
r? `@oli-obk`
Let's avoid using two different terms for the same thing -- let's just call it "provenance" everywhere.
In Miri, provenance consists of an AllocId and an SbTag (Stacked Borrows tag), which made this even more confusing.
interpret: make some large types not Copy
Also remove some unused trait impls (mostly HashStable).
This didn't find any unnecessary copies that I managed to avoid, but it might still be better to require explicit clone for these types? Not sure.
r? `@oli-obk`
Use constant eval to do strict mem::uninit/zeroed validity checks
I'm not sure about the code organisation here, I just dumped the check in rustc_const_eval at the root. Not hard to move it elsewhere, in any case.
Also, this means cranelift codegen intrinsics lose the strict checks, since they don't seem to depend on rustc_const_eval, and I didn't see a point in keeping around two copies.
I also left comments in the is_zero_valid methods about "uhhh help how do i do this", those apply to both methods equally.
Also rustc_codegen_ssa now depends on rustc_const_eval... is this okay?
Pinging `@RalfJung` since you were the one who mentioned this to me, so I'm assuming you're interested.
Haven't had a chance to run full tests on this since it's really warm, and it's 1AM, I'll check out any failures/comments in the morning :)
interpret/visitor: support visiting with a PlaceTy
Finally we can visit a `PlaceTy` in a way that will only do `force_allocation` when needed ti visit a field. :)
r? `@oli-obk`
interpret: get rid of MemPlaceMeta::Poison
This is achieved by refactoring the projection code (`{mplace,place,operand}_{downcast,field,index,...}`) so that we no longer need to call `assert_mem_place` in the operand handling.
Pull Derefer before ElaborateDrops
_Follow up work to #97025#96549#96116#95887 #95649_
This moves `Derefer` before `ElaborateDrops` and creates a new `Rvalue` called `VirtualRef` that allows us to bypass many constraints for `DerefTemp`.
r? `@oli-obk`
Moves our projection handling code into a common file, and avoids the use of a
general mplace-based fallback function by have more specialized implementations.
mplace_index (and the other slice-related functions) could be more efficient by
copy-pasting the body of operand_index. Or we could do some trait magic to share
the code between them. But for now this is probably fine.
don't allow ZST in ScalarInt
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 I propose we stop using ScalarInt to represent ZST (which are clearly not integers). Instead, we can add new ZST variants to those types that did not have other variants which could be used for this purpose.
Based on https://github.com/rust-lang/rust/pull/98831. Only the commits starting from "don't allow ZST in ScalarInt" are new.
r? `@oli-obk`
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.
interpret: use AllocRange in UninitByteAccess
also use nice new format string syntax in `interpret/error.rs`, and use the `#` flag to add `0x` prefixes where applicable.
r? ``@oli-obk``
Operand::Uninit is an *allocated* operand that is fully uninitialized.
This lets us lazily allocate the actual backing store of *all* locals (no matter their ABI).
I also reordered things in pop_stack_frame at the same time.
I should probably have made that a separate commit...
Change enum->int casts to not go through MIR casts.
follow-up to https://github.com/rust-lang/rust/pull/96814
this simplifies all backends and even gives LLVM more information about the return value of `Rvalue::Discriminant`, enabling optimizations in more cases.
fix interpreter validity check on Box
Follow-up to https://github.com/rust-lang/rust/pull/98554: avoid walking over parts of the value twice.
And then move all that logic into the general visitor so not each visitor implementation has to deal with it...
Interpret: AllocRange Debug impl, and use it more consistently
The two commits are pretty independent but it did not seem worth having two PRs for them.
r? ``@oli-obk``
interpret: don't rely on ScalarPair for overflowed arithmetic
This is for https://github.com/rust-lang/rust/pull/97861.
Cc `@eddyb`
I would like to avoid making this depend on `dest.layout.abi` to avoid a branch that we are not usually covering both sides of. Though OTOH this seems like fairly straight-forward code. But let's benchmark this option first to see how bad that extra `force_allocation` really is.
CTFE interning: don't walk allocations that don't need it
The interning of const allocations visits the mplace looking for references to intern. Walking big aggregates like big static arrays can be costly, so we only do it if the allocation we're interning contains references or interior mutability.
Walking ZSTs was avoided before, and this optimization is now applied to cases where there are no references/relocations either.
---
While initially looking at this in the context of #93215, I've been testing with smaller allocations than the 16GB one in that issue, and with different init/uninit patterns (esp. via padding).
In that example, by default, `eval_to_allocation_raw` is the heaviest query followed by `incr_comp_serialize_result_cache`. So I'll show numbers when incremental compilation is disabled, to focus on the const allocations themselves at 95% of the compilation time, at bigger array sizes on these minimal examples like `static ARRAY: [u64; LEN] = [0; LEN];`.
That is a close construction to parts of the `ctfe-stress-test-5` benchmark, which has const allocations in the megabytes, while most crates usually have way smaller ones. This PR will have the most impact in these situations, as the walk during the interning starts to dominate the runtime.
Unicode crates (some of which are present in our benchmarks) like `ucd`, `encoding_rs`, etc come to mind as having bigger than usual allocations as well, because of big tables of code points (in the hundreds of KB, so still an order of magnitude or 2 less than the stress test).
In a check build, for a single static array shown above, from 100 to 10^9 u64s (for lengths in powers of ten), the constant factors are lowered:
(log scales for easier comparisons)
(linear scale for absolute diff at higher Ns)
For one of the alternatives of that issue
```rust
const ROWS: usize = 100_000;
const COLS: usize = 10_000;
static TWODARRAY: [[u128; COLS]; ROWS] = [[0; COLS]; ROWS];
```
we can see a similar reduction of around 3x (from 38s to 12s or so).
For the same size, the slowest case IIRC is when there are uninitialized bytes e.g. via padding
```rust
const ROWS: usize = 100_000;
const COLS: usize = 10_000;
static TWODARRAY: [[(u64, u8); COLS]; ROWS] = [[(0, 0); COLS]; ROWS];
```
then interning/walking does not dominate anymore (but means there is likely still some interesting work left to do here).
Compile times in this case rise up quite a bit, and avoiding interning walks has less impact: around 23%, from 730s on master to 568s with this PR.
