This function has some shared code for the thin LTO and fat LTO cases,
but those cases have so little in common that it's actually clearer to
treat them fully separately.
PR #112946 tweaked the naming of LLVM threads, but messed things up
slightly, resulting in threads on Windows having names like `optimize
module {} regex.f10ba03eb5ec7975-cgu.0`.
This commit removes the extraneous `{} `.
The main loop has a *very* complex condition, which includes two
mentions of `codegen_state`. The body of the loop then immediately
switches on the `codegen_state`.
I find it easier to understand if it's a `loop` and we check for exit
conditions after switching on `codegen_state`. We end up with a tiny bit
of code duplication, but it's clear that (a) we never exit in the
`Ongoing` case, (b) we exit in the `Completed` state only if several
things are true (and there's interaction with LTO there), and (c) we
exit in the `Aborted` state if a couple of things are true. Also, the
exit conditions are all simple conjunctions.
This loop condition involves `codegen_state`, `work_items`, and
`running_with_own_token`. But the body of the loop cannot modify
`codegen_state`, so repeatedly checking it is unnecessary.
`CodegenContext` is immutable except for the `worker` field - we clone
`CodegenContext` in multiple places, changing the `worker` field each
time. It's simpler to move the `worker` field out of `CodegenContext`.
It took me some time to understand how the main thread can lend a
jobserver token to an LLVM thread. This commit renames a couple of
things to make it clearer.
- Rename the `LLVMing` variant as `Lending`, because that is a clearer
description of what is happening.
- Rename `running` as `running_with_own_token`, which makes it clearer
that there might be one additional LLVM thread running (with a loaned
token). Also add a comment to its definition.
And rename the `Compiled` variant as `Finished`, because that name makes
it clearer there is nothing left to do, contrasting nicely with the
`Needs*` variants.
Resurrect: rustc_llvm: Add a -Z `print-codegen-stats` option to expose LLVM statistics.
This resurrects PR https://github.com/rust-lang/rust/pull/104000, which has sat idle for a while. And I want to see the effect of stack-move optimizations on LLVM (like https://reviews.llvm.org/D153453) :).
I have applied the changes requested by `@oli-obk` and `@nagisa` https://github.com/rust-lang/rust/pull/104000#discussion_r1014625377 and https://github.com/rust-lang/rust/pull/104000#discussion_r1014642482 in the latest commits.
r? `@oli-obk`
-----
LLVM has a neat [statistics](https://llvm.org/docs/ProgrammersManual.html#the-statistic-class-stats-option) feature that tracks how often optimizations kick in. It's very handy for optimization work. Since we expose the LLVM pass timings, I thought it made sense to expose the LLVM statistics too.
-----
(Edit: fix broken link
(Edit2: fix segmentation fault and use malloc
If `rustc` is built with
```toml
[llvm]
assertions = true
```
Then you can see like
```
rustc +stage1 -Z print-codegen-stats -C opt-level=3 tmp.rs
===-------------------------------------------------------------------------===
... Statistics Collected ...
===-------------------------------------------------------------------------===
3 aa - Number of MayAlias results
193 aa - Number of MustAlias results
531 aa - Number of NoAlias results
...
```
And the current default build emits only
```
$ rustc +stage1 -Z print-codegen-stats -C opt-level=3 tmp.rs
===-------------------------------------------------------------------------===
... Statistics Collected ...
===-------------------------------------------------------------------------===
$
```
This might be better to emit the message to tell assertion flag necessity, but now I can't find how to do that...
Implement rust-lang/compiler-team#578.
When an ICE is encountered on nightly releases, the new rustc panic
handler will also write the contents of the backtrace to disk. If any
`delay_span_bug`s are encountered, their backtrace is also added to the
file. The platform and rustc version will also be collected.
LLVM has a neat [statistics] feature that tracks how often optimizations kick
in. It's very handy for optimization work. Since we expose the LLVM pass
timings, I thought it made sense to expose the LLVM statistics too.
[statistics]: https://llvm.org/docs/ProgrammersManual.html#the-statistic-class-stats-option
Add `-Zremark-dir` unstable flag to write LLVM optimization remarks to YAML
This PR adds an option for `rustc` to emit LLVM optimization remarks to a set of YAML files, which can then be digested by existing tools, like https://github.com/OfekShilon/optview2. When `-Cremark-dir` is passed, and remarks are enabled (`-Cremark=all`), the remarks will be now written to the specified directory, **instead** of being printed to standard error output. The files are named based on the CGU from which they are being generated.
Currently, the remarks are written using the LLVM streaming machinery, directly in the diagnostics handler. It seemed easier than going back to Rust and then form there back to C++ to use the streamer from the diagnostics handler. But there are many ways to implement this, of course, so I'm open to suggestions :)
I included some comments with questions into the code. Also, I'm not sure how to test this.
r? `@tmiasko`
For non-incremental builds on Unix, currently all the thread names look
like `opt regex.f10ba03eb5ec7975-cgu.0`. But they are truncated by
`pthread_setname` to `opt regex.f10ba`, hiding the numeric suffix that
distinguishes them. This is really annoying when using a profiler like
Samply.
This commit changes these thread names to a form like `opt cgu.0`, which
is much better.
The codegen main loop has two bools, `codegen_done` and
`codegen_aborted`. There are only three valid combinations: `(false,
false)`, `(true, false)`, `(true, true)`.
This commit replaces them with a single tri-state enum, which makes
things clearer.
`Message` is an enum with multiple variants. Four of those variants map
directly onto the four variants of `WorkItemResult`. This commit reduces
those four `Message` variants to a single variant containing a
`WorkItemResult`. This requires increasing `WorkItemResult`'s visibility
to `pub(crate)` visibility, but `WorkItem` and `Message` can also have
their visibility reduced to `pub(crate)`.
This change avoids some boilerplate enum translation code, and makes
`Message` easier to understand.
`Message` is an enum with multiple variants, for messages sent to the
coordinator thread. *Except* for `Message::CodegenItem`, which is
entirely disjoint, being for messages sent from the coordinator thread
to the main thread.
This commit move `Message::CodegenItem` into a separate type,
`CguMessage`, which makes the code much clearer.
Because tiny CGUs make compilation less efficient *and* result in worse
generated code.
We don't do this when the number of CGUs is explicitly given, because
there are times when the requested number is very important, as
described in some comments within the commit. So the commit also
introduces a `CodegenUnits` type that distinguishes between default
values and user-specified values.
This change has a roughly neutral effect on walltimes across the
rustc-perf benchmarks; there are some speedups and some slowdowns. But
it has significant wins for most other metrics on numerous benchmarks,
including instruction counts, cycles, binary size, and max-rss. It also
reduces parallelism, which is good for reducing jobserver competition
when multiple rustc processes are running at the same time. It's smaller
benchmarks that benefit the most; larger benchmarks already have CGUs
that are all larger than the minimum size.
Here are some example before/after CGU sizes for opt builds.
- html5ever
- CGUs: 16, mean size: 1196.1, sizes: [3908, 2992, 1706, 1652, 1572,
1136, 1045, 948, 946, 938, 579, 471, 443, 327, 286, 189]
- CGUs: 4, mean size: 4396.0, sizes: [6706, 3908, 3490, 3480]
- libc
- CGUs: 12, mean size: 35.3, sizes: [163, 93, 58, 53, 37, 8, 2 (x6)]
- CGUs: 1, mean size: 424.0, sizes: [424]
- tt-muncher
- CGUs: 5, mean size: 1819.4, sizes: [8508, 350, 198, 34, 7]
- CGUs: 1, mean size: 9075.0, sizes: [9075]
Note that CGUs of size 100,000+ aren't unusual in larger programs.
Write to stdout if `-` is given as output file
With this PR, if `-o -` or `--emit KIND=-` is provided, output will be written to stdout instead. Binary output (those of type `obj`, `llvm-bc`, `link` and `metadata`) being written this way will result in an error unless stdout is not a tty. Multiple output types going to stdout will trigger an error too, as they will all be mixded together.
This implements https://github.com/rust-lang/compiler-team/issues/431
The idea behind the changes is to introduce an `OutFileName` enum that represents the output - be it a real path or stdout - and to use this enum along the code paths that handle different output types.
If `-o -` or `--emit KIND=-` is provided, output will be written
to stdout instead. Binary output (`obj`, `llvm-bc`, `link` and
`metadata`) being written this way will result in an error unless
stdout is not a tty. Multiple output types going to stdout will
trigger an error too, as they will all be mixded together.
Each of `{D,Subd}iagnosticMessage::{Str,Eager}` has a comment:
```
// FIXME(davidtwco): can a `Cow<'static, str>` be used here?
```
This commit answers that question in the affirmative. It's not the most
compelling change ever, but it might be worth merging.
This requires changing the `impl<'a> From<&'a str>` impls to `impl
From<&'static str>`, which involves a bunch of knock-on changes that
require/result in call sites being a little more precise about exactly
what kind of string they use to create errors, and not just `&str`. This
will result in fewer unnecessary allocations, though this will not have
any notable perf effects given that these are error paths.
Note that I was lazy within Clippy, using `to_string` in a few places to
preserve the existing string imprecision. I could have used `impl
Into<{D,Subd}iagnosticMessage>` in various places as is done in the
compiler, but that would have required changes to *many* call sites
(mostly changing `&format("...")` to `format!("...")`) which didn't seem
worthwhile.
Currently a `{D,Subd}iagnosticMessage` can be created from any type that
impls `Into<String>`. That includes `&str`, `String`, and `Cow<'static,
str>`, which are reasonable. It also includes `&String`, which is pretty
weird, and results in many places making unnecessary allocations for
patterns like this:
```
self.fatal(&format!(...))
```
This creates a string with `format!`, takes a reference, passes the
reference to `fatal`, which does an `into()`, which clones the
reference, doing a second allocation. Two allocations for a single
string, bleh.
This commit changes the `From` impls so that you can only create a
`{D,Subd}iagnosticMessage` from `&str`, `String`, or `Cow<'static,
str>`. This requires changing all the places that currently create one
from a `&String`. Most of these are of the `&format!(...)` form
described above; each one removes an unnecessary static `&`, plus an
allocation when executed. There are also a few places where the existing
use of `&String` was more reasonable; these now just use `clone()` at
the call site.
As well as making the code nicer and more efficient, this is a step
towards possibly using `Cow<'static, str>` in
`{D,Subd}iagnosticMessage::{Str,Eager}`. That would require changing
the `From<&'a str>` impls to `From<&'static str>`, which is doable, but
I'm not yet sure if it's worthwhile.
