indirect immutable freeze by-value function parameters.
Right now, `rustc` only examines function signatures and the platform ABI when
determining the LLVM attributes to apply to parameters. This results in missed
optimizations, because there are some attributes that can be determined via
analysis of the MIR making up the function body. In particular, `readonly`
could be applied to most indirectly-passed by-value function arguments
(specifically, those that are freeze and are observed not to be mutated), but
it currently is not.
This patch introduces the machinery that allows `rustc` to determine those
attributes. It consists of a query, `deduced_param_attrs`, that, when
evaluated, analyzes the MIR of the function to determine supplementary
attributes. The results of this query for each function are written into the
crate metadata so that the deduced parameter attributes can be applied to
cross-crate functions. In this patch, we simply check the parameter for
mutations to determine whether the `readonly` attribute should be applied to
parameters that are indirect immutable freeze by-value. More attributes could
conceivably be deduced in the future: `nocapture` and `noalias` come to mind.
Adding `readonly` to indirect function parameters where applicable enables some
potential optimizations in LLVM that are discussed in [issue 103103] and [PR
103070] around avoiding stack-to-stack memory copies that appear in functions
like `core::fmt::Write::write_fmt` and `core::panicking::assert_failed`. These
functions pass a large structure unchanged by value to a subfunction that also
doesn't mutate it. Since the structure in this case is passed as an indirect
parameter, it's a pointer from LLVM's perspective. As a result, the
intermediate copy of the structure that our codegen emits could be optimized
away by LLVM's MemCpyOptimizer if it knew that the pointer is `readonly
nocapture noalias` in both the caller and callee. We already pass `nocapture
noalias`, but we're missing `readonly`, as we can't determine whether a
by-value parameter is mutated by examining the signature in Rust. I didn't have
much success with having LLVM infer the `readonly` attribute, even with fat
LTO; it seems that deducing it at the MIR level is necessary.
No large benefits should be expected from this optimization *now*; LLVM needs
some changes (discussed in [PR 103070]) to more aggressively use the `noalias
nocapture readonly` combination in its alias analysis. I have some LLVM patches
for these optimizations and have had them looked over. With all the patches
applied locally, I enabled LLVM to remove all the `memcpy`s from the following
code:
```rust
fn main() {
println!("Hello {}", 3);
}
```
which is a significant codegen improvement over the status quo. I expect that
if this optimization kicks in in multiple places even for such a simple
program, then it will apply to Rust code all over the place.
[issue 103103]: https://github.com/rust-lang/rust/issues/103103
[PR 103070]: https://github.com/rust-lang/rust/pull/103070
The compiler currently has `-Ztime` and `-Ztime-passes`. I've used
`-Ztime-passes` for years but only recently learned about `-Ztime`.
What's the difference? Let's look at the `-Zhelp` output:
```
-Z time=val -- measure time of rustc processes (default: no)
-Z time-passes=val -- measure time of each rustc pass (default: no)
```
The `-Ztime-passes` description is clear, but the `-Ztime` one is less so.
Sounds like it measures the time for the entire process?
No. The real difference is that `-Ztime-passes` prints out info about passes,
and `-Ztime` does the same, but only for a subset of those passes. More
specifically, there is a distinction in the profiling code between a "verbose
generic activity" and an "extra verbose generic activity". `-Ztime-passes`
prints both kinds, while `-Ztime` only prints the first one. (It took me
a close reading of the source code to determine this difference.)
In practice this distinction has low value. Perhaps in the past the "extra
verbose" output was more voluminous, but now that we only print stats for a
pass if it exceeds 5ms or alters the RSS, `-Ztime-passes` is less spammy. Also,
a lot of the "extra verbose" cases are for individual lint passes, and you need
to also use `-Zno-interleave-lints` to see those anyway.
Therefore, this commit removes `-Ztime` and the associated machinery. One thing
to note is that the existing "extra verbose" activities all have an extra
string argument, so the commit adds the ability to accept an extra argument to
the "verbose" activities.
This simplifies things, but requires making `CacheEncoder` non-generic.
(This was previously merged as commit 4 in #94732 and then was reverted
in #97905 because it caused a perf regression.)
This avoids the name clash with `rustc_serialize::Encoder` (a trait),
and allows lots qualifiers to be removed and imports to be simplified
(e.g. fewer `as` imports).
(This was previously merged as commit 5 in #94732 and then was reverted
in #97905 because of a perf regression caused by commit 4 in #94732.)
This avoids the name clash with `rustc_serialize::Encoder` (a trait),
and allows lots qualifiers to be removed and imports to be simplified
(e.g. fewer `as` imports).
There are two impls of the `Encoder` trait: `opaque::Encoder` and
`opaque::FileEncoder`. The former encodes into memory and is infallible, the
latter writes to file and is fallible.
Currently, standard `Result`/`?`/`unwrap` error handling is used, but this is a
bit verbose and has non-trivial cost, which is annoying given how rare failures
are (especially in the infallible `opaque::Encoder` case).
This commit changes how `Encoder` fallibility is handled. All the `emit_*`
methods are now infallible. `opaque::Encoder` requires no great changes for
this. `opaque::FileEncoder` now implements a delayed error handling strategy.
If a failure occurs, it records this via the `res` field, and all subsequent
encoding operations are skipped if `res` indicates an error has occurred. Once
encoding is complete, the new `finish` method is called, which returns a
`Result`. In other words, there is now a single `Result`-producing method
instead of many of them.
This has very little effect on how any file errors are reported if
`opaque::FileEncoder` has any failures.
Much of this commit is boring mechanical changes, removing `Result` return
values and `?` or `unwrap` from expressions. The more interesting parts are as
follows.
- serialize.rs: The `Encoder` trait gains an `Ok` associated type. The
`into_inner` method is changed into `finish`, which returns
`Result<Vec<u8>, !>`.
- opaque.rs: The `FileEncoder` adopts the delayed error handling
strategy. Its `Ok` type is a `usize`, returning the number of bytes
written, replacing previous uses of `FileEncoder::position`.
- Various methods that take an encoder now consume it, rather than being
passed a mutable reference, e.g. `serialize_query_result_cache`.
`SourceFile::lines` is a big part of metadata. It's stored in a compressed form
(a difference list) to save disk space. Decoding it is a big fraction of
compile time for very small crates/programs.
This commit introduces a new type `SourceFileLines` which has a `Lines`
form and a `Diffs` form. The latter is used when the metadata is first
read, and it is only decoded into the `Lines` form when line data is
actually needed. This avoids the decoding cost for many files,
especially in `std`. It's a performance win of up to 15% for tiny
crates/programs where metadata decoding is a high part of compilation
costs.
A `Lock` is needed because the methods that access lines data (which can
trigger decoding) take `&self` rather than `&mut self`. To allow for this,
`SourceFile::lines` now takes a `FnMut` that operates on the lines slice rather
than returning the lines slice.
`Decoder` has two impls:
- opaque: this impl is already partly infallible, i.e. in some places it
currently panics on failure (e.g. if the input is too short, or on a
bad `Result` discriminant), and in some places it returns an error
(e.g. on a bad `Option` discriminant). The number of places where
either happens is surprisingly small, just because the binary
representation has very little redundancy and a lot of input reading
can occur even on malformed data.
- json: this impl is fully fallible, but it's only used (a) for the
`.rlink` file production, and there's a `FIXME` comment suggesting it
should change to a binary format, and (b) in a few tests in
non-fundamental ways. Indeed #85993 is open to remove it entirely.
And the top-level places in the compiler that call into decoding just
abort on error anyway. So the fallibility is providing little value, and
getting rid of it leads to some non-trivial performance improvements.
Much of this commit is pretty boring and mechanical. Some notes about
a few interesting parts:
- The commit removes `Decoder::{Error,error}`.
- `InternIteratorElement::intern_with`: the impl for `T` now has the same
optimization for small counts that the impl for `Result<T, E>` has,
because it's now much hotter.
- Decodable impls for SmallVec, LinkedList, VecDeque now all use
`collect`, which is nice; the one for `Vec` uses unsafe code, because
that gave better perf on some benchmarks.
Fixes#92163Fixes#92014
When writing to the incremental cache, we encode all `Span`s
we encounter, regardless of whether or not their `SourceFile`
comes from the local crate, or from a foreign crate.
When we decode a `Span`, we use the `StableSourceFileId` we encoded
to locate the matching `SourceFile` in the current session. If this
id corresponds to a `SourceFile` from another crate, then we need to
have already imported that `SourceFile` into our current session.
This usually happens automatically during resolution / macro expansion,
when we try to resolve definitions from other crates. In certain cases,
however, we may try to load a `Span` from a transitive dependency
without having ever imported the `SourceFile`s from that crate, leading
to an ICE.
This PR fixes the issue by calling `imported_source_files()`
when we encounter a `SourceFile` with a foreign `CrateNum`.
This ensures that all `SourceFile`s from that crate are imported
into the current session.
This encoding allows for random access without an expensive upfront decoding
state which in turn allows simplifying the DefPathIndex lookup logic without
regressing performance.
generic_const_exprs: use thir for abstract consts instead of mir
Changes `AbstractConst` building to use `thir` instead of `mir` so that there's less chance of consts unifying when they shouldn't because lowering to mir dropped information (see `abstract-consts-as-cast-5.rs` test)
r? `@lcnr`
