Use larger span for adjustment THIR expressions
Currently, we use a relatively 'small' span for THIR
expressions generated by an 'adjustment' (e.g. an autoderef,
autoborrow, unsizing). As a result, if a borrow generated
by an adustment ends up causing a borrowcheck error, for example:
```rust
let mut my_var = String::new();
let my_ref = &my_var
my_var.push('a');
my_ref;
```
then the span for the mutable borrow may end up referring
to only the base expression (e.g. `my_var`), rather than
the method call which triggered the mutable borrow
(e.g. `my_var.push('a')`)
Due to a quirk of the MIR borrowck implementation,
this doesn't always get exposed in migration mode,
but it does in many cases.
This commit makes THIR building consistently use 'larger'
spans for adjustment expressions. These spans are recoded
when we first create the adjustment during typecheck. For
example, an autoref adjustment triggered by a method call
will record the span of the entire method call.
The intent of this change it make it clearer to users
when it's the specific way in which a variable is
used (for example, in a method call) that produdes
a borrowcheck error. For example, an error message
claiming that a 'mutable borrow occurs here' might
be confusing if it just points at a usage of a variable
(e.g. `my_var`), when no `&mut` is in sight. Pointing
at the entire expression should help to emphasize
that the method call itself is responsible for
the mutable borrow.
In several cases, this makes the `#![feature(nll)]` diagnostic
output match up exactly with the default (migration mode) output.
As a result, several `.nll.stderr` files end up getting removed
entirely.
Currently, we use a relatively 'small' span for THIR
expressions generated by an 'adjustment' (e.g. an autoderef,
autoborrow, unsizing). As a result, if a borrow generated
by an adustment ends up causing a borrowcheck error, for example:
```rust
let mut my_var = String::new();
let my_ref = &my_var
my_var.push('a');
my_ref;
```
then the span for the mutable borrow may end up referring
to only the base expression (e.g. `my_var`), rather than
the method call which triggered the mutable borrow
(e.g. `my_var.push('a')`)
Due to a quirk of the MIR borrowck implementation,
this doesn't always get exposed in migration mode,
but it does in many cases.
This commit makes THIR building consistently use 'larger'
spans for adjustment expressions
The intent of this change it make it clearer to users
when it's the specific way in which a variable is
used (for example, in a method call) that produdes
a borrowcheck error. For example, an error message
claiming that a 'mutable borrow occurs here' might
be confusing if it just points at a usage of a variable
(e.g. `my_var`), when no `&mut` is in sight. Pointing
at the entire expression should help to emphasize
that the method call itself is responsible for
the mutable borrow.
In several cases, this makes the `#![feature(nll)]` diagnostic
output match up exactly with the default (migration mode) output.
As a result, several `.nll.stderr` files end up getting removed
entirely.
Introduce NullOp::AlignOf
This PR introduces `Rvalue::NullaryOp(NullOp::AlignOf, ty)`, which will be lowered from `align_of`, similar to `size_of` lowering to `Rvalue::NullaryOp(NullOp::SizeOf, ty)`.
The changes are originally part of #88700 but since it's not dependent on other changes and could have performance impact on its own, it's separated into its own PR.
rustc: use more correct span data in for loop desugaring
Fixes#82462
Before:
help: consider adding semicolon after the expression so its temporaries are dropped sooner, before the local variables declared by the block are dropped
|
LL | for x in DroppingSlice(&*v).iter(); {
| +
After:
help: consider adding semicolon after the expression so its temporaries are dropped sooner, before the local variables declared by the block are dropped
|
LL | };
| +
This seems like a reasonable fix: since the desugared "expr_drop_temps_mut" contains the entire desugared loop construct, its span should contain the entire loop construct as well.
Each pattern in a match arm has its own copy of the match guard in MIR,
with its own temporary, so it has to be dropped before the the guards
are joined to the single copy of the arm.
MIR lowering for `if let` expressions is now more complicated now that
`if let` exists in HIR. This PR adds a scope for the variables bound in
an `if let` expression and then uses an approach similar to how we
handle loops to ensure that we reliably drop the correct variables.
Before:
help: consider adding semicolon after the expression so its temporaries are dropped sooner, before the local variables declared by the block are dropped
|
LL | for x in DroppingSlice(&*v).iter(); {
| +
After:
help: consider adding semicolon after the expression so its temporaries are dropped sooner, before the local variables declared by the block are dropped
|
LL | };
| +
This seems like a reasonable fix: since the desugared "expr_drop_temps_mut"
contains the entire desugared loop construct, its span should contain the
entire loop construct as well.
Previously, we would set up the source lines for `match` expressions so
that the code generated to perform the test of the scrutinee was matched
to the line of the arm that required the test and then jump from the arm
block to the "next" block was matched to all of the lines in the `match`
expression.
While that makes sense, it has the side effect of causing strange
stepping behavior in debuggers.
I've changed the source information so that all of the generated tests
are sourced to `match {scrutinee}` and the jumps are sourced to the last
line of the block they are inside. This resolves the weird stepping
behavior in all debuggers and resolves some instances of "ambiguous
symbol" errors in WinDbg preventing the user from setting breakpoints at
`match` expressions.
MIR opt: separate constant predecessors of a switch
For each block S ending with a switch, this pass copies S for each of S's predecessors that seem to assign the value being switched over as a const. This is done using a somewhat simple heuristic to determine what seems to be a const transitively.
More precisely, this is what the pass does:
- find a block that ends in a switch
- track if there is an unique place set before the current basic block that determines the result of the switch (this is the part that resolves switching over discriminants)
- if there is, iterate over the parents that have a reasonable terminator and find if the found determining place is likely to be (transitively) set from a const within that parent block
- if so, add the corresponding edge to a vector of edges to duplicate
- once this is done, iterate over the found edges: copy the target block and replace the reference to the target block in the origin block with the new block
This pass is not optimal and could probably duplicate in more cases, but the intention was mostly to address cases like in #85133 or #85365, to avoid creating new enums that get destroyed immediately afterwards (notably making the new try v2 `?` desugar zero-cost).
A benefit of this pass working the way it does is that it is easy to ensure its correctness: the worst that can happen is for it to needlessly copy a basic block, which is likely to be destroyed by cleanup passes afterwards. The complex parts where aliasing matters are only heuristics and the hard work is left to further passes like ConstProp.
# LLVM blocker
Unfortunately, I believe it would be unwise to enable this optimization by default for now. Indeed, currently switch lowering passes like SimplifyCFG in LLVM lose the information on the set of possible variant values, which means it tends to actually generate worse code with this optimization enabled. A fix would have to be done in LLVM itself. This is something I also want to look into. I have opened [a bug report at the LLVM bug tracker](https://bugs.llvm.org/show_bug.cgi?id=50455).
When this is done, I hope we can enable this pass by default. It should be fairly fast and I think it is beneficial in many cases. Notably, it should be a sound alternative to simplify-arm-identity. By the way, ConstProp only seems to pick up the optimization in functions that are not generic. This is however most likely an issue in ConstProp that I will look into afterwards.
This is my first contribution to rustc, and I would like to thank everyone on the Zulip mir-opt chat for the help and support, and especially `@scottmcm` for the guidance.
Peephole optimize `x == false` and `x != true`
This adds peephole optimizations to make `x == false`, `false == x`, `x != true`, and `true != x` get optimized to `!x` in the `instcombine` MIR pass. That pass currently handles `x == true` -> `x` already.
Implement the new desugaring from `try_trait_v2`
~~Currently blocked on https://github.com/rust-lang/rust/issues/84782, which has a PR in https://github.com/rust-lang/rust/pull/84811~~ Rebased atop that fix.
`try_trait_v2` tracking issue: https://github.com/rust-lang/rust/issues/84277
Unfortunately this is already touching a ton of things, so if you have suggestions for good ways to split it up, I'd be happy to hear them. (The combination between the use in the library, the compiler changes, the corresponding diagnostic differences, even MIR tests mean that I don't really have a great plan for it other than trying to have decently-readable commits.
r? `@ghost`
~~(This probably shouldn't go in during the last week before the fork anyway.)~~ Fork happened.
