Fix a few API issues. There's also a newly-added file in compiler-rt
that is not supposed to be built by default but is not being excluded
properly by a glob. Added a patch to exclude that and sent
https://github.com/llvm/llvm-project/pull/208861 upstream.
This is based on an idea suggested during review of the original type
erasure PR.
The type-erased check failure path lowers each check through
`CheckFailFormat<Ts...>`, which builds one format adapter per value,
down to the out-of-line `CheckFailImpl`, which takes an array of
base-class `format_adapter*` pointers. Previously a separate variadic
`CheckFailWithAdapters<Adapters...>` template sat in between: it existed
only to bind the adapter temporaries to named parameters so that
pointers to their base class could be collected into a `std::array` and
outlive the call to `CheckFailImpl`.
This removes that layer. `CheckFailFormat` now builds the pointer array
directly, in the braced-init-list of the `CheckFailImpl` call, using a
small `CheckFailFormatAdapterAddr<T>` helper to take the base-class
address of each adapter. The adapter temporaries are materialized as
named parameters of that helper within the same full-expression as the
`CheckFailImpl` call, so they remain alive across the call without a
dedicated function for that purpose.
The win is in debug info, not code. `CheckFailWithAdapters<Adapters...>`
was instantiated once per distinct adapter-type sequence in every
translation unit that uses `CARBON_CHECK`, and each instantiation
carried its own DWARF records -- type entries, string-table offsets, and
range lists. Dropping it removes those records from every such
translation unit. The generated machine code is unchanged.
Measured impact (fastbuild, the inputs to the inner-loop links):
- First-party object files shrink by 173,728 bytes (-0.233%) across the
`carbon-busybox` link inputs; 232 of 270 objects get smaller and none
grow. The largest reductions are in the check-densest translation units
(`type_completion`, `type`, `import_ref`, `function_context`,
`constant`, ...).
- Per `bloaty`, the reduction is almost entirely DWARF (`.debug_info`,
`.debug_str`, `.rela.debug_str_offsets`, `.debug_rnglists`); the loaded
code (VM size) is unchanged.
Assisted-by: Claude Code
---------
Co-authored-by: Geoff Romer <gromer@google.com>
`CARBON_CHECK` and `CARBON_FATAL` messages are formatted with
`llvm::formatv`. Previously each check site that had a message
instantiated its own copy of the formatv machinery -- a `formatv_object`
over a tuple of per-argument format adapters, plus that tuple -- in
every translation unit, keyed on the site's file, line, condition, and
format strings. A translation unit with many checks paid for that
machinery over and over.
This restructures check failure so that the formatting machinery is
compiled exactly once, and only a single small adapter is instantiated
per distinct value type per translation unit:
- `CheckFailImpl` (out-of-line) now takes the message's format string
and an array of already-type-erased `format_adapter`s, and renders the
whole failure message -- prefix plus the extra message -- directly into
one stream. The extra message is rendered in place, so no separate
string is ever materialized for it.
- `FormatvInto` (in the `.cpp`) renders a format string over that
adapter array. Rather than instantiate `llvm::formatv`, it drives the
formatv replacement loop over the public
`formatv_object_base::parseFormatString`, so this rendering code exists
exactly once. (A TODO notes that we should add a type-erased entry point
upstream in LLVM rather than reimplement the loop here.)
- The lowering from the macro down to that out-of-line call is split so
that the only per-check-site instantiation is trivial:
- `CheckFail<...>` is instantiated once per site, since its file, line,
condition, and format template-string parameters are unique to the site.
It just lowers those compile-time strings to ordinary arguments and
forwards to `CheckFailFormat`.
- `CheckFailFormat<Ts...>` is instantiated once per distinct value-type
sequence and shared across sites; it builds one type-erased adapter per
value.
- `CheckFailWithAdapters<Adapters...>` collects pointers to those
adapters into an array and calls `CheckFailImpl`. It is a distinct
function so the adapter temporaries stay alive while pointers to their
base class are in flight.
Format semantics, including runtime format-string validation, are
unchanged, and the rendered message is byte-for-byte identical.
For `DCHECK` in optimized builds the check is dead code; its arguments
are now routed through a trivial `IgnoreDeadCheckArgs` no-op rather than
`CheckFail`. This still type-checks the arguments so they cannot bitrot,
without instantiating any formatting machinery for them and without
provoking unused-variable warnings.
Measured full-rebuild impact (353 first-party translation units,
fastbuild): -112.6s CPU, -6.9% relative to trunk.
Assisted-by: Claude
Most versions are through `pre-commit autoupdate --freeze`, clang-format
was manually updated to the latest at
https://github.com/ssciwr/clang-format-wheel
My read of the style changes here are that they seem fine, none of them
look like regressions (which has caused me to delay/adjust updates in
the past).
Noted CopyNameFromImportIR while glancing around (this one's interesting
because it's NameId, not void nor auto), did a scan just for a few other
cases. Not an exhaustive fix, and TBH assuming we'd prefer `auto ... ->
auto` since equivalent Carbon syntax would probably be `fn ... -> auto`
`toolchain/autoupdate_testdata.py --allow-check-fail` can now be used to
perform an autoupdate even if some `CARBON_CHECK`s are failing. What
this does will depend on how the toolchain behaves after the `CHECK`
failure, and of course there's no guarantees there, but this can be
useful if it's easier to debug the `CHECK` failure by looking at the
produced SemIR.
Internally, this uses `bazel build --config=non-fatal-checks`, which in
turn specifies a `--per_file_copt` for `check_internal.cpp`. The intent
here is that the rebuild required to enable or disable this mode is as
small as reasonably possible.
This mode is not compatible with `-c opt`, as it's important that check
failure calls are `[[noreturn]]` in `-c opt` mode.
---------
Co-authored-by: Jon Ross-Perkins <jperkins@google.com>
Given code like the following:
```
auto kind = ConversionTarget::Kind{0};
CARBON_CHECK(!loc_id.is_valid(), "hello {0} world", kind);
```
Currently we would print 'hello <the next line>', as the check string
would be treated as terminating at the '{0}', so it does not print the
rest of the string or a newline. This is because ConversionTarget::Kind
is an enum with underlying type `int8_t` which is a char, and
llvm::formatv does not look if the type is an enum and treat is
specially. So it prints it as a char rather than a number, which in this
case is a nul terminator.
With this change, the '{0}' value will be converted to a larger integer
before being passed through to llvm::formatv so that char-sized enums
will print as a number, and the result is that we will print 'hello 0
world\n' as the developer intended.
These appear to be causing some subtle misinteractions with MSan that we
don't understand, and may be a compiler bug. =/ Fortunately, they
weren't essential to the performance gains so just remove them for now.
When benchmarked on an x86 server, where I would expect this to be more
important due to relatively few named registers, the performance change
appears to be either an improvement or in the noise.
Huge credit to Jon for tracking down that this is related to the MSan
issues.
This switches `DCHECK` and `FATAL` as well.
The goal is to reduce the code size impact of these assertions so that
we can keep more of them enabled. Currently, the largest cost I see from
`CHECK` is not the actual check or the cold code itself, but actually
the failure to inline trivial functions due to the presence of the cold
code. This means that our goal isn't to reduce apparent code size in the
final binary but the LLVM IR cost assessed for these routines in the
inliner, which closely correlates with code size but is a bit different.
As discussed in #4283, experimentation shows that a single function call
with a minimal number of arguments is the lowest cost model for these.
This is easily achieved with a format-string API that internally uses
`llvm::formatv`. This PR is essentially the `CHECK` version of #4283.
However, the check macros are substantially harder to make work with
both format strings and streaming because they also take a condition.
Also, unexpectedly, I was very successful at devising a regular
expression based automated rewrite from the streaming to the format
string form with only low 10s of manual fixes. This includes compacting
strings broken up across lines, etc. Given how well that went, I've
prepared this PR which just directly switches to the format string API
and migrate everything to use it.
One nice side-effect is that the format string approach ends up greatly
simplifying the implementation here as well.
This is ... *shockingly* effective. Parsing speeds up by more than 3%
with just this change. And checking speeds up by **8%** with this change
alone:
```
BM_CompileAPIFileDenseDecls<Phase::Parse>/256 86.3µs ± 1% 82.9µs ± 1% -3.94% (p=0.000 n=17+19)
BM_CompileAPIFileDenseDecls<Phase::Parse>/1024 431µs ± 1% 415µs ± 1% -3.76% (p=0.000 n=18+19)
BM_CompileAPIFileDenseDecls<Phase::Parse>/4096 1.77ms ± 1% 1.71ms ± 1% -3.18% (p=0.000 n=18+19)
BM_CompileAPIFileDenseDecls<Phase::Parse>/16384 7.44ms ± 1% 7.17ms ± 2% -3.56% (p=0.000 n=18+20)
BM_CompileAPIFileDenseDecls<Phase::Parse>/65536 30.7ms ± 1% 29.7ms ± 1% -3.15% (p=0.000 n=18+20)
BM_CompileAPIFileDenseDecls<Phase::Parse>/262144 131ms ± 1% 127ms ± 1% -2.81% (p=0.000 n=18+18)
BM_CompileAPIFileDenseDecls<Phase::Check>/256 878µs ± 2% 800µs ± 1% -8.91% (p=0.000 n=19+20)
BM_CompileAPIFileDenseDecls<Phase::Check>/1024 1.88ms ± 2% 1.72ms ± 1% -8.56% (p=0.000 n=19+20)
BM_CompileAPIFileDenseDecls<Phase::Check>/4096 5.78ms ± 2% 5.28ms ± 1% -8.70% (p=0.000 n=20+18)
BM_CompileAPIFileDenseDecls<Phase::Check>/16384 21.9ms ± 1% 20.1ms ± 1% -8.02% (p=0.000 n=18+20)
BM_CompileAPIFileDenseDecls<Phase::Check>/65536 90.4ms ± 2% 83.1ms ± 1% -8.04% (p=0.000 n=19+20)
BM_CompileAPIFileDenseDecls<Phase::Check>/262144 381ms ± 2% 352ms ± 1% -7.79% (p=0.000 n=19+19)
```
---------
Co-authored-by: Richard Smith <richard@metafoo.co.uk>
Co-authored-by: josh11b <15258583+josh11b@users.noreply.github.com>
This required adding a few headers that were found transitively before,
but not too many. This is sadly a fairly manual process of opening every
file in my IDE, but I think I got everything in `//common` and
`//toolchain`.
There are a few cases where technically we don't need `foo.h` to be
included into `foo.cpp`, but I've forced those to stay with a pragma.
I've tried to catch the places where we can cut deps in Bazel as well,
but not sure I got all of those.
I had been noticing these in other PRs and it seemed better to isolate
the change.
Splits IR files into SemIR, and logic files into Check. These will be
split into separate directories as part of a later move; the namespaces
are being done first in order to vet the switch, and hopefully make
conflicts a little easier to manage due to the substantial renames.
A lot of this is just automated removal of Semantics prefixes from
names, adding namespace references where needed. A few special-cases
are:
- SemanticsIR -> SemIR::File
- A few things were discussed, like Unit, CompileUnit, or CompiledUnit.
Unit was too vague for chandlerc, and I thought CompileUnit might lead
to incorrect inferences (CompilationUnit would be more precise, but
typically written as SemIR::CompilationUnit which is pretty long). File
seemed to be a short name that we could agree on.
- SemanticsIRFormatter -> SemIR::Formatter
- FormatSemanticsIR -> SemIR::FormatFile
- SemanticsFileTest -> CheckFileTest
- It remains in the Testing namespace, where just "FileTest" might be
too broad a name.
- SemanticsDeclarationNameStack::Context ->
Check::DeclarationNameStack::NameContext
- This avoids a Check::Context name shadowing.
Changes check_internal.h to include ostream.h to improve finding of
Print/operator<< (otherwise it didn't compile).
This is part of #3070
Just opening files in vscode and seeing what clangd flags.
Some edits to ostream.h to stop it from getting flagged (the usage pattern means it's not always obviously used).
At present, CHECK/FATAL print their own stack trace. This switches to just using std::abort for the stack trace, as well as the CHECK printing more completely.
This has a few consequences:
1) I'm now buffering the FATAL strings in order to print it later.
2) We now print the bug report message and program arguments on failure. This is part of pretty printing and was elided before.
3) We can now have pretty printing on FATAL, e.g. to show the stacks we're building in the parser.
This is a variation on #1339 and also fixes#1338.
The important difference from #1339 is that this works to retain the
benefits of #1216 which seem important -- both the clarity of printing
the message last and the correctness of actually showing the correct
line number in the backtrace.
The approach in this patch is to disable LLVM's error handling just
before using `std::abort()`. This should give us roughly the best of
both worlds.
This PR also fixes an issue where we wouldn't run the file-cleanup
actions that the LLVM `std::abort()` handler does. This almost certainly
doesn't yet matter, but likely would in the future.
We should separately consider adding back information about filing bugs
that roughly corresponds to the error message that LLVM itself prints.
I've not tried to replicate that from #1339 here and just focused on
getting to `std::abort()` while preserving the desired order of messages
and stack trace locations.
Before, if the llvm hook to print a stack trace on CHECK was bound, it would trace after the message rather than before the message. I'm thinking swapping the order will generally be easier to debug.
Richard Smith
Chandler CarruthandGeoff Romer
Jon Ross-Perkins
Dana Jansens
micttyl
pk19604014
Jon Meow