Fundamentally, this uses forward declarations of Clang types to reduce
the overall compile time cost of Clang headers across the codebase.
Tracing and profiling showed ~2s of every check TU's ~8-12s compile time
going just to parsing Clang frontend and AST headers pulled in via a few
sem_ir and check headers that only use the Clang types by pointer or
reference:
- sem_ir/cpp_file.h (reached via sem_ir/file.h by ~150 TUs) included
clang/Frontend/CompilerInstance.h, clang/CodeGen/ModuleBuilder.h,
clang/AST/Mangle.h, and llvm/IR/Module.h. CppFile's accessors move out
of line to a new cpp_file.cpp and the header now forward-declares the
Clang types.
- check/cpp/context.h (reached via check/context.h by ~100 TUs) included
clang/Frontend/FrontendAction.h and clang/Parse/Parser.h, pulling in
clang's Sema.h and ASTUnit.h.
- sem_ir/clang_decl.h included clang/AST/Decl.h; the three small
functions that need complete Clang types move out of line.
- sem_ir/cpp_overload_set.h included clang/Sema/Overload.h solely for
the three-field OverloadCandidateSet::OperatorRewriteInfo, which is now
mirrored as CppOverloadSet::OperatorRewriteInfo, and clang/AST/Decl.h
solely for a pointer.
- sem_ir/name_scope.h's clang/AST/DeclBase.h include was vestigial.
TUs (and more narrowly included headers) that genuinely use the Clang
definitions now include the Clang headers directly.
Representative compile times (fastbuild, aarch64), combined with the
preceding instantiation-cost changes, relative to trunk:
- check/eval.cpp: 11.85s -> 6.94s (-41%)
- check/handle_operator.cpp: 7.71s -> 3.30s (-57%)
- language_server.cpp: 6.68s -> 3.16s (-53%)
- lower/handle.cpp: 6.75s -> 3.66s (-46%)
- sem_ir/file.cpp: 8.60s -> 6.11s (-29%)
- driver.cpp: 6.68s -> 4.78s (-28%)
Measured full-rebuild impact (316 first-party TUs, fastbuild): -689.5s
CPU, -29.9% relative to trunk.
Assisted-by: Claude
Carbon-side thunks (for example the `Copy`/`Destroy` witness thunks
generated for imported C++ types) are mangled by Carbon, and their names
incorporate a fingerprint of the involved types. The instruction
fingerprinter identifies a class only by its name and parent scope,
which is sufficient for Carbon classes but not for imported C++ classes:
different specializations of one class template (and other cases such as
types in anonymous namespaces) share a Carbon name and parent scope. As
a result, the thunks for two distinct specializations could mangle to
the same name, producing a single LLVM function with two definitions and
failing `verifyModule` during lowering.
When fingerprinting a class imported from C++, also include the Clang
mangled name of its type.
Test: toolchain/lower/testdata/interop/cpp/thunks.carbon gains a split
with two specializations of one class template, each requiring a thunk;
their thunks now get distinct mangled names instead of colliding.
Assisted-by: Claude Code
---------
Co-authored-by: Christopher Di Bella <cjdb.ns@gmail.com>
Background:
https://docs.google.com/document/d/1wi85FRiWh4X9A-gCYMVGKR40-q5fM6-3JaSpePk-XCY/edit?usp=sharing
And specifically this work is essentially an alternative to #5543
Clang's code generation is implemented through an ASTListener
(clang::CodeGenerator) that is attached throughout Clang's
parsing/sema/code
generation phases and acts on Clang AST incrementally throughout that
process.
Prior to this patch, Carbon has only created the CodeGenerator during
Carbon's
`lower` phase, missing out on key callbacks that would be made by Clang
during
`check`. Some of these issues were addressed by #6237 and #6483 - but
there were
still remaining cases where the delayed processing lead to missing
functionality.
With #6483 much of the Clang code that made multithreaded complexity of
#5543 is
no longer present, and we have access to the point of ASTListener
registration
so we can register the CodeGenerator there and consume its resulting
llvm::Module during lower.
Examples of some of the bugs this addresses are seen in the linked doc,
and
checked in as tests in this change in
`clang_code_generator_callbacks.carbon`
An indicental bug that's also fixed, and caused all the other test case
churn,
is that the `CodeGenerator` created during `lower` wasn't getting passed
the
Clang `CodeGenOpts` and was creating its own default - so, most notably,
optimization flags were not respected. This meant that the LLVM IR from
Clang
was always -O0 style IR (optnone, no inlinehint, no TBAA, etc). With
this
change, now the Clang IRGen gets the real `CodeGenOpts` and respects
optimization/other flags specified there.
This is only meant to be a rough proof of concept - I'm totally open to
reworking this in any way (even quite substantially) if folks have ideas
about
how this should be implemented most generally/elegantly/etc.
* Move `Sema` access from `CppFile` into `CppContext`.
* Move the mangle context from `SemIR::File` into `CppContext`.
* Move source location mapping state from `Context` into `CppContext`.
Also factor out the `GenerateAst` function that builds the `CppContext`
and `CppFile` into its own file.
This intends to avoid proliferation of dependencies on the exact API of
`clang::ASTUnit`, and would enable us to more easily switch to a
different approach that gives us more control over the construction of
the Clang AST.
Also remove some unnecessary tracking of the `CppFile` and instead
always retrieve it from the `SemIR::File`.
---------
Co-authored-by: Jon Ross-Perkins <jperkins@google.com>
Chandler Carruth
David Blaikie
Richard Smith
