Use an interpreter in MIR jump threading
This allows to understand assignments of aggregate constants. This case appears more frequently with GVN promoting aggregates to constants.
Sandwich MIR optimizations between DSE.
This PR reorders MIR optimization passes in an attempt to increase their efficiency.
- Stop running CopyProp before GVN, it's useless as GVN will do the same thing anyway. Instead, we perform CopyProp at the end of the pipeline, to ensure we do not emit copy/move chains.
- Run DSE before GVN, as it increases the probability to have single-assignment locals.
- Run DSE after the final CopyProp to turn copies into moves.
r? `@ghost`
Avoid some redundant work in GVN
The first 2 commits are about reducing the perf effect.
Third commit avoids doing redundant work: is a local is SSA, it already has been simplified, and the resulting value is in `self.locals`. No need to call any code on it.
The last commit avoids removing some storage statements.
r? wg-mir-opt
coverage: Add enums to accommodate other kinds of coverage mappings
Extracted from #118305.
LLVM supports several different kinds of coverage mapping regions, but currently we only ever emit ordinary “code” regions. This PR performs the plumbing required to add other kinds of regions as enum variants, but does not add any specific variants other than `Code`.
The main motivation for this change is branch coverage, but it will also allow separate experimentation with gap regions and skipped regions, which might help in producing more accurate and useful coverage reports.
---
``@rustbot`` label +A-code-coverage
Reorder early post-inlining passes.
`RemoveZsts`, `RemoveUnneededDrops` and `UninhabitedEnumBranching` only depend on types, so they should be executed together early after MIR inlining introduces those types.
This does not change the end-result, but this makes the pipeline a bit more consistent.
Merge dead bb pruning and unreachable bb deduplication.
Both routines share the same basic structure: iterate on all bbs to identify work, and then renumber bbs.
We can do both at once.
Run Miri and mir-opt tests without a target linker
Normally, we need a linker for the target to build the standard library. That's only because `std` declares crate-type lib and dylib; building the dylib is what creates a need for the linker.
But for mir-opt tests (and for Miri) we do not need to build a `libstd.so`. So with this PR, when we build the standard library for mir-opt tests, instead of `cargo build` we run `cargo rustc --crate-type=lib` which overrides the configured crate types in `std`'s manifest.
I've also swapped in what seems to me a better hack than `BOOTSTRAP_SKIP_TARGET_SANITY` to prevent cross-interpreting with Miri from checking for a target linker and expanded it to mir-opt tests too. Whether it's actually better is up to a reviewer.
By using FxIndexMap instead of FxHashMap, so that the order of visiting
of locals is deterministic.
We also need to bless
copy_propagation_arg.foo.DestinationPropagation.panic*.diff. Do not
review the diff of the diff. Instead look at the diff file before and
after this commit. Both before and after this commit, 3 statements are
replaced with nop. It's just that due to change in ordering, different
statements are replaced. But the net result is the same.
Migrate memory overlap check from validator to lint
The check attempts to identify potential undefined behaviour, rather
than whether MIR is well-formed. It belongs in the lint not validator.
Follow up to changes from #119077.
Remove `-Zdump-mir-spanview`
The `-Zdump-mir-spanview` flag was added back in #76074, as a development/debugging aid for the initial work on what would eventually become `-Cinstrument-coverage`. It causes the compiler to emit an HTML file containing a function's source code, with various spans highlighted based on the contents of MIR.
When the suggestion was made to [triage and remove unnecessary `-Z` flags (Zulip)](https://rust-lang.zulipchat.com/#narrow/stream/131828-t-compiler/topic/.60-Z.60.20option.20triage), I noted that this flag could potentially be worth removing, but I wanted to keep it around to see whether I found it useful for my own coverage work.
But when I actually tried to use it, I ran into various issues (e.g. it crashes on `tests/coverage/closure.rs`). If I can't trust it to work properly without a full overhaul, then instead of diving down a rabbit hole of trying to fix arcane span-handling bugs, it seems better to just remove this obscure old code entirely.
---
````@rustbot```` label +A-code-coverage
Reevaluate `body.should_skip()` after updating the MIR phase to ensure
that injected MIR is processed correctly.
Update a few custom MIR tests that were ill-formed for the injected
phase.
custom mir: make it clear what the return block is
Custom MIR recently got support for specifying the "unwind action", so now there's two things coming after the actual call part of `Call` terminators. That's not very self-explaining so I propose we change the syntax to imitate keyword arguments:
```
Call(popped = Vec::pop(v), ReturnTo(drop), UnwindContinue())
```
Also fix some outdated docs and add some docs to `Call` and `Drop`.
This involves lots of breaking changes. There are two big changes that
force changes. The first is that the bitflag types now don't
automatically implement normal derive traits, so we need to derive them
manually.
Additionally, bitflags now have a hidden inner type by default, which
breaks our custom derives. The bitflags docs recommend using the impl
form in these cases, which I did.
Implement constant propagation on top of MIR SSA analysis
This implements the idea I proposed in https://github.com/rust-lang/rust/pull/110719#issuecomment-1718324700
Based on https://github.com/rust-lang/rust/pull/109597
The value numbering "GVN" pass formulates each rvalue that appears in MIR with an abstract form (the `Value` enum), and assigns an integer `VnIndex` to each. This abstract form can be used to deduplicate values, reusing an earlier local that holds the same value instead of recomputing. This part is proposed in #109597.
From this abstract representation, we can perform more involved simplifications, for example in https://github.com/rust-lang/rust/pull/111344.
With the abstract representation `Value`, we can also attempt to evaluate each to a constant using the interpreter. This builds a `VnIndex -> OpTy` map. From this map, we can opportunistically replace an operand or a rvalue with a constant if their value has an associated `OpTy`.
The most relevant commit is [Evaluated computed values to constants.](2767c4912e)"
r? `@oli-obk`
rework `-Zverbose`
implements the changes described in https://github.com/rust-lang/compiler-team/issues/706
the first commit is only a name change from `-Zverbose` to `-Zverbose-internals` and does not change behavior. the second commit changes diagnostics.
possible follow up work:
- `ty::pretty` could print more info with `--verbose` than it does currently. `-Z verbose-internals` shows too much info in a way that's not helpful to users. michael had ideas about this i didn't fully understand: https://rust-lang.zulipchat.com/#narrow/stream/233931-t-compiler.2Fmajor-changes/topic/uplift.20some.20-Zverbose.20calls.20and.20rename.20to.E2.80.A6.20compiler-team.23706/near/408984200
- `--verbose` should imply `-Z write-long-types-to-disk=no`. the code in `ty_string_with_limit` should take `--verbose` into account (apparently this affects `Ty::sort_string`, i'm not familiar with this code). writing a file to disk should suggest passing `--verbose`.
r? `@compiler-errors` cc `@estebank`
match lowering: Remove the `make_target_blocks` hack
This hack was introduced 4 years ago in [`a1d0266` (#60730)](a1d0266878) to improve LLVM optimization time, specifically noticed in the `encoding` benchmark. Measurements today indicate it is no longer needed.
r? `@matthewjasper`
State transforms retains storage statements for locals that are not
stored inside a coroutine. It ensures those locals are live when
resuming by inserting StorageLive as appropriate. It forgot to end the
storage of those locals when suspending, which is fixed here.
While the end of live range is implicit when executing return, it is
nevertheless useful for inliner which would otherwise extend the live
range beyond return.