Enable GVN for `AggregateKind::RawPtr`
Looks like I was worried for nothing; this seems like it's much easier than I was originally thinking it would be.
r? `@cjgillot`
This should be useful for `x[..4]`-like things, should those start inlining enough to expose the lengths.
The only non-obvious changes:
- `building/storage_live_dead_in_statics.rs` has a `#[rustfmt::skip]`
attribute to avoid reformating a table of data.
- Two `.mir` files have slight changes involving line numbers.
- In `unusual_item_types.rs` an `EMIT_MIR` annotation is moved to
outside a function, which is the usual spot, because `tidy` complains
if such a comment is indented.
The commit also tweaks the comments in `rustfmt.toml`.
The `mir!` macro has multiple parts:
- An optional return type annotation.
- A sequence of zero or more local declarations.
- A mandatory starting anonymous basic block, which is brace-delimited.
- A sequence of zero of more additional named basic blocks.
Some `mir!` invocations use braces with a "block" style, like so:
```
mir! {
let _unit: ();
{
let non_copy = S(42);
let ptr = std::ptr::addr_of_mut!(non_copy);
// Inside `callee`, the first argument and `*ptr` are basically
// aliasing places!
Call(_unit = callee(Move(*ptr), ptr), ReturnTo(after_call), UnwindContinue())
}
after_call = {
Return()
}
}
```
Some invocations use parens with a "block" style, like so:
```
mir!(
let x: [i32; 2];
let one: i32;
{
x = [42, 43];
one = 1;
x = [one, 2];
RET = Move(x);
Return()
}
)
```
And some invocations uses parens with a "tighter" style, like so:
```
mir!({
SetDiscriminant(*b, 0);
Return()
})
```
This last style is generally used for cases where just the mandatory
starting basic block is present. Its braces are placed next to the
parens.
This commit changes all `mir!` invocations to use braces with a "block"
style. Why?
- Consistency is good.
- The contents of the invocation is a block of code, so it's odd to use
parens. They are more normally used for function-like macros.
- Most importantly, the next commit will enable rustfmt for
`tests/mir-opt/`. rustfmt is more aggressive about formatting macros
that use parens than macros that use braces. Without this commit's
changes, rustfmt would break a couple of `mir!` macro invocations that
use braces within `tests/mir-opt` by inserting an extraneous comma.
E.g.:
```
mir!(type RET = (i32, bool);, { // extraneous comma after ';'
RET.0 = 1;
RET.1 = true;
Return()
})
```
Switching those `mir!` invocations to use braces avoids that problem,
resulting in this, which is nicer to read as well as being valid
syntax:
```
mir! {
type RET = (i32, bool);
{
RET.0 = 1;
RET.1 = true;
Return()
}
}
```
Enable DestinationPropagation by default.
~~Based on https://github.com/rust-lang/rust/pull/115291.~~
This PR proposes to enable the destination propagation pass by default.
This pass is meant to reduce the amount of copies present in MIR.
At the same time, this PR removes the `RenameReturnPlace` pass, as it is currently unsound.
`DestinationPropagation` is not limited to `_0`, but does not handle borrowed locals.
[ACP 362] genericize `ptr::from_raw_parts`
This implements https://github.com/rust-lang/libs-team/issues/362
As such, it can partially undo https://github.com/rust-lang/rust/pull/124795 , letting `slice_from_raw_parts` just call `from_raw_parts` again without re-introducing the unnecessary cast to MIR.
By doing this it also removes a spurious cast from `str::from_raw_parts`. And I think it does a good job of showing the value of the ACP, since the only thing that needed new turbofishing because of this is inside `ptr::null(_mut)`, but only because `ptr::without_provenance(_mut)` doesn't support pointers to extern types, which it absolutely could (without even changing the implementation).
don't inhibit random field reordering on repr(packed(1))
`inhibit_struct_field_reordering_opt` being false means we exclude this type from random field shuffling. However, `packed(1)` types can still be shuffled! The logic was added in https://github.com/rust-lang/rust/pull/48528 since it's pointless to reorder fields in packed(1) types (there's no padding that could be saved) -- but that shouldn't inhibit `-Zrandomize-layout` (which did not exist at the time).
We could add an optimization elsewhere to not bother sorting the fields for `repr(packed)` types, but I don't think that's worth the effort.
This *does* change the behavior in that we may now reorder fields of `packed(1)` structs (e.g. if there are niches, we'll try to move them to the start/end, according to `NicheBias`). We were always allowed to do that but so far we didn't. Quoting the [reference](https://doc.rust-lang.org/reference/type-layout.html):
> On their own, align and packed do not provide guarantees about the order of fields in the layout of a struct or the layout of an enum variant, although they may be combined with representations (such as C) which do provide such guarantees.
coverage: Memoize and simplify counter expressions
When creating coverage counter expressions as part of coverage instrumentation, we often end up creating obviously-redundant expressions like `c1 + (c0 - c1)`, which is equivalent to just `c0`.
To avoid doing so, this PR checks when we would create an expression matching one of 5 patterns, and uses the simplified form instead:
- `(a - b) + b` → `a`.
- `(a + b) - b` → `a`.
- `(a + b) - a` → `b`.
- `a + (b - a)` → `b`.
- `a - (a - b)` → `b`.
Of all the different ways to combine 3 operands and 2 operators, these are the patterns that allow simplification.
(Some of those patterns currently don't occur in practice, but are included anyway for completeness, to avoid having to add them later as branch coverage and MC/DC coverage support expands.)
---
This PR also adds memoization for newly-created (or newly-simplified) counter expressions, to avoid creating duplicates.
This currently makes no difference to the final mappings, but is expected to be useful for MC/DC coverage of match expressions, as proposed by https://github.com/rust-lang/rust/pull/124278#issuecomment-2106754753.
Some of these cases currently don't occur in practice, but are included for
completeness, and to avoid having to add them later as branch coverage and
MC/DC coverage start building more complex expressions.
Avoid a cast in `ptr::slice_from_raw_parts(_mut)`
Casting to `*const ()` or `*mut ()` is no longer needed after https://github.com/rust-lang/rust/pull/123840 so let's make the MIR smaller (and more inline-able, as seen in the tests).
If [ACP#362](https://github.com/rust-lang/libs-team/issues/362) goes through we can keep calling `ptr::from_raw_parts(_mut)` in these also without the cast, but that hasn't had any libs-api attention yet, so I'm not waiting on it.
never patterns: lower never patterns to `Unreachable` in MIR
This lowers a `!` pattern to "goto Unreachable". Ideally I'd like to read from the place to make it clear that the UB is coming from an invalid value, but that's tricky so I'm leaving it for later.
r? `@compiler-errors` how do you feel about a lil bit of MIR lowering
Casting to `*const ()` or `*mut ()` just bloats the MIR, so let's not.
If ACP#362 goes through we can keep calling `ptr::from_raw_parts(_mut)` in these also without the cast, but that hasn't had any libs-api attention yet, so I'm not waiting on it.
Account for immutably borrowed locals in MIR copy-prop and GVN
For the most part, we consider that immutably borrowed `Freeze` locals still fulfill SSA conditions. As the borrow is immutable, any use of the local will have the value given by the single assignment, and there can be no surprise.
This allows copy-prop to merge a non-borrowed local with a borrowed local. We chose to keep copy-classes heads unborrowed, as those may be easier to optimize in later passes.
This also allows to GVN the value behind an immutable borrow. If a SSA local is borrowed, dereferencing that borrow is equivalent to copying the local's value: re-executing the assignment between the borrow and the dereference would be UB.
r? `@ghost` for perf
deref patterns: lower deref patterns to MIR
This lowers deref patterns to MIR. This is a bit tricky because this is the first kind of pattern that requires storing a value in a temporary. Thanks to https://github.com/rust-lang/rust/pull/123324 false edges are no longer a problem.
The thing I'm not confident about is the handling of fake borrows. This PR ignores any fake borrows inside a deref pattern. We are guaranteed to at least fake borrow the place of the first pointer value, which could be enough, but I'm not certain.
Disable SimplifyToExp in MatchBranchSimplification
Due to the miscompilation mentioned in #124150, We need to disable MatchBranchSimplification temporarily.
To fully resolve this issue, my plan is:
1. Disable SimplifyToExp in MatchBranchSimplification (this PR).
2. Remove all potentially unclear transforms in #124122.
3. Gradually add back the removed transforms (possibly multiple PRs).
r? `@Nilstrieb` or `@oli-obk`
Make `checked` ops emit *unchecked* LLVM operations where feasible
For things with easily pre-checked overflow conditions -- shifts and unsigned subtraction -- write the checked methods in such a way that we stop emitting wrapping versions of them.
For example, today <https://rust.godbolt.org/z/qM9YK8Txb> neither
```rust
a.checked_sub(b).unwrap()
```
nor
```rust
a.checked_sub(b).unwrap_unchecked()
```
actually optimizes to `sub nuw`. After this PR they do.
cc #103299
For things with easily pre-checked overflow conditions -- shifts and unsigned subtraction -- write then checked methods in such a way that we stop emitting wrapping versions of them.
For example, today <https://rust.godbolt.org/z/qM9YK8Txb> neither
```rust
a.checked_sub(b).unwrap()
```
nor
```rust
a.checked_sub(b).unwrap_unchecked()
```
actually optimizes to `sub nuw`. After this PR they do.
Re-enable the early otherwise branch optimization
Closes#95162. Fixes#119014.
This is the first part of #121397.
An invalid enum discriminant can come from anywhere. We have to check to see if all successors contain the discriminant statement. This should have a pass to hoist instructions.
r? cjgillot
Pass list of defineable opaque types into canonical queries
This eliminates `DefiningAnchor::Bubble` for good and brings the old solver closer to the new one wrt cycles and nested obligations. At that point the difference between `DefiningAnchor::Bind([])` and `DefiningAnchor::Error` was academic. We only used the difference for some sanity checks, which actually had to be worked around in places, so I just removed `DefiningAnchor` entirely and just stored the list of opaques that may be defined.
fixes#108498
fixes https://github.com/rust-lang/rust/issues/116877
* [x] run crater
- https://github.com/rust-lang/rust/pull/122077#issuecomment-2013293931
match lowering: make false edges more precise
When lowering match expressions, we add false edges to hide details of the lowering from borrowck. Morally we pretend we're testing the patterns (and guards) one after the other in order. See the tests for examples. Problem is, the way we implement this today is too coarse for deref patterns.
In deref patterns, a pattern like `deref [1, x]` matches on a `Vec` by creating a temporary to store the output of the call to `deref()` and then uses that to continue matching. Here the pattern has a binding, which we set up after the pre-binding block. Problem is, currently the false edges tell borrowck that the pre-binding block can be reached from a previous arm as well, so the `deref()` temporary may not be initialized. This triggers an error when we try to use the binding `x`.
We could call `deref()` a second time, but this opens the door to soundness issues if the deref impl is weird. Instead in this PR I rework false edges a little bit.
What we need from false edges is a (fake) path from each candidate to the next, specifically from candidate C's pre-binding block to next candidate D's pre-binding block. Today, we link the pre-binding blocks directly. In this PR, I link them indirectly by choosing an earlier node on D's success path. Specifically, I choose the earliest block on D's success path that doesn't make a loop (if I chose e.g. the start block of the whole match (which is on the success path of all candidates), that would make a loop). This turns out to be rather straightforward to implement.
r? `@matthewjasper` if you have the bandwidth, otherwise let me know
Rename `UninhabitedEnumBranching` to `UnreachableEnumBranching`
Per [#120268](https://github.com/rust-lang/rust/pull/120268#discussion_r1517492060), I rename `UninhabitedEnumBranching` to `UnreachableEnumBranching` .
I solved some nits to add some comments.
I adjusted the workaround restrictions. This should be useful for `a <= b` and `if let Some/Ok(v)`. For enum with few variants, `early-tailduplication` should not cause compile time overhead.
r? RalfJung
rename ptr::from_exposed_addr -> ptr::with_exposed_provenance
As discussed on [Zulip](https://rust-lang.zulipchat.com/#narrow/stream/136281-t-opsem/topic/To.20expose.20or.20not.20to.20expose/near/427757066).
The old name, `from_exposed_addr`, makes little sense as it's not the address that is exposed, it's the provenance. (`ptr.expose_addr()` stays unchanged as we haven't found a better option yet. The intended interpretation is "expose the provenance and return the address".)
The new name nicely matches `ptr::without_provenance`.
De-LLVM the unchecked shifts [MCP#693]
This is just one part of the MCP (https://github.com/rust-lang/compiler-team/issues/693), but it's the one that IMHO removes the most noise from the standard library code.
Seems net simpler this way, since MIR already supported heterogeneous shifts anyway, and thus it's not more work for backends than before.
r? WaffleLapkin
Add `Ord::cmp` for primitives as a `BinOp` in MIR
Update: most of this OP was written months ago. See https://github.com/rust-lang/rust/pull/118310#issuecomment-2016940014 below for where we got to recently that made it ready for review.
---
There are dozens of reasonable ways to implement `Ord::cmp` for integers using comparison, bit-ops, and branches. Those differences are irrelevant at the rust level, however, so we can make things better by adding `BinOp::Cmp` at the MIR level:
1. Exactly how to implement it is left up to the backends, so LLVM can use whatever pattern its optimizer best recognizes and cranelift can use whichever pattern codegens the fastest.
2. By not inlining those details for every use of `cmp`, we drastically reduce the amount of MIR generated for `derive`d `PartialOrd`, while also making it more amenable to MIR-level optimizations.
Having extremely careful `if` ordering to μoptimize resource usage on broadwell (#63767) is great, but it really feels to me like libcore is the wrong place to put that logic. Similarly, using subtraction [tricks](https://graphics.stanford.edu/~seander/bithacks.html#CopyIntegerSign) (#105840) is arguably even nicer, but depends on the optimizer understanding it (https://github.com/llvm/llvm-project/issues/73417) to be practical. Or maybe [bitor is better than add](https://discourse.llvm.org/t/representing-in-ir/67369/2?u=scottmcm)? But maybe only on a future version that [has `or disjoint` support](https://discourse.llvm.org/t/rfc-add-or-disjoint-flag/75036?u=scottmcm)? And just because one of those forms happens to be good for LLVM, there's no guarantee that it'd be the same form that GCC or Cranelift would rather see -- especially given their very different optimizers. Not to mention that if LLVM gets a spaceship intrinsic -- [which it should](https://rust-lang.zulipchat.com/#narrow/stream/131828-t-compiler/topic/Suboptimal.20inlining.20in.20std.20function.20.60binary_search.60/near/404250586) -- we'll need at least a rustc intrinsic to be able to call it.
As for simplifying it in Rust, we now regularly inline `{integer}::partial_cmp`, but it's quite a large amount of IR. The best way to see that is with 8811efa88b (diff-d134c32d028fbe2bf835fef2df9aca9d13332dd82284ff21ee7ebf717bfa4765R113) -- I added a new pre-codegen MIR test for a simple 3-tuple struct, and this PR change it from 36 locals and 26 basic blocks down to 24 locals and 8 basic blocks. Even better, as soon as the construct-`Some`-then-match-it-in-same-BB noise is cleaned up, this'll expose the `Cmp == 0` branches clearly in MIR, so that an InstCombine (#105808) can simplify that to just a `BinOp::Eq` and thus fix some of our generated code perf issues. (Tracking that through today's `if a < b { Less } else if a == b { Equal } else { Greater }` would be *much* harder.)
---
r? `@ghost`
But first I should check that perf is ok with this
~~...and my true nemesis, tidy.~~
This is just one part of the MCP, but it's the one that IMHO removes the most noise from the standard library code.
Seems net simpler this way, since MIR already supported heterogeneous shifts anyway, and thus it's not more work for backends than before.
