Let me tell you a tale...
Within some twisted maze of debug info I've ended up implementing an
insane man's Include What You Use device. When the debugger emits debug
info it really shouldn't, I find out why & then realize the code could
be improved too.
In this instance CIndexDiagnostics.cpp had a lot more debug info with
Clang than GCC. Upon inspection a major culprit was all the debug info
describing clang::Sema. This was emitted because clang::Sema is
befriended by DiagnosticEngine which was rightly required, but GCC
doesn't emit debug info for friends so it never emitted anything for
Clang. Clang does emit debug info for friends (will be fixed/changed to
reduce debug info size).
But why didn't Clang just emit a declaration of Sema if this entire TU
didn't require a definition?
1) Diagnostic.h did the right thing, only using a declaration of Sema
and not including Sema.h at all.
2) Some other dependency of CIndexDiagnostics.cpp didn't do the right
thing. ASTUnit.h, only needing a declaration, still included Sema.h
(hence this commit which removes that include and adds the necessary
includes to the cpp files that were relying on this)
3) -flimit-debug-info didn't save us because of
EnterExpressionEvaluationContext, defined inline in Sema.h which fires
the "requiresCompleteType" check/flag (since it uses nested types from
Sema and calls Sema member functions) and thus, if debug info is ever
emitted for the type, the whole type is emitted and not just a
declaration.
Improving -flimit-debug-info to account for this would be... hard.
Modifying the code so that's not 'required to be complete' might be
possible, but probably only by moving EnterExpressionEvaluationContext
either into Sema, or out of Sema.h. That might be a bit too much of a
contortion to be bothered with.
Also, this is only one of the cases where emitting debug info for
friends caused us to emit a lot more debug info (this change reduces
Clang's DWO size by 0.93%, dropping friends entirely reduces debug info
by 3.2%) - I haven't hunted down the other cases, but I assume they
might be similar (Sema or something like it). IWYU or a similar tool
might help us reduce build times a bit, but analyzing debug info to find
these differences isn't worthwhile. I'll take the 3.2% win, provide this
small improvement to the code itself, and move on.
llvm-svn: 190715
The problem was that an enum without closing semicolon could be associated as a forward enum
in an erroneous declaration, leading to the identifier being associated with the enum decl but
without a declaration actually referencing it.
This resulted in not having it serialized before serializing the identifier that is associated with.
Also prevent the ASTUnit from querying the serialized DeclID for an invalid top-level decl; it may not
have been serialized.
rdar://14539667
llvm-svn: 187914
The top-level hash is used to determine if we need to update the global code-completion results.
ImportDecls did not affect the hash so a newly introduced ImportDecl would not trigger an update of the global results.
rdar://14202797
llvm-svn: 184782
A while ago we allowed libclang to build a PCH that had compiler errors; this was to retain the performance
afforded by a PCH even if the user's code is in an intermediate state.
Extend this for the precompiled preamble as well.
rdar://14109828
llvm-svn: 183717
Previously, we would clone the current diagnostic consumer to produce
a new diagnostic consumer to use when building a module. The problem
here is that we end up losing diagnostics for important diagnostic
consumers, such as serialized diagnostics (where we'd end up with two
diagnostic consumers writing the same output file). With forwarding,
the diagnostics from all of the different modules being built get
forwarded to the one serialized-diagnostic consumer and are emitted in
a sane way.
Fixes <rdar://problem/13663996>.
llvm-svn: 181067
The global module index was querying the file manager for each of the
module files it knows about at load time, to prune out any out-of-date
information. The file manager would then cache the results of the
stat() falls used to find that module file.
Later, the same translation unit could end up trying to import one of the
module files that had previously been ignored by the module cache, but
after some other Clang instance rebuilt the module file to bring it
up-to-date. The stale stat() results in the file manager would
trigger a second rebuild of the already-up-to-date module, causing
failures down the line.
The global module index now lazily resolves its module file references
to actual AST reader module files only after the module file has been
loaded, eliminating the stat-caching race. Moreover, the AST reader
can communicate to its caller that a module file is missing (rather
than simply being out-of-date), allowing us to simplify the
module-loading logic and allowing the compiler to recover if a
dependent module file ends up getting deleted.
llvm-svn: 177367
Add an ability to specify custom documentation block comment commands via a new
class CommentOptions. The intention is that this class will hold future
customizations for comment parsing, including defining documentation comments
with specific numbers of parameters, etc.
CommentOptions instance is a member of LangOptions.
CommentOptions is controlled by a new command-line parameter
-fcomment-block-commands=Foo,Bar,Baz.
llvm-svn: 175892
Also, it was the only reason that `argc` and `argv` were being passed
into createDiagnostics, so remove those parameters and clean up callers.
llvm-svn: 172945
code-completion results, the SourceManager state may be slightly
different when code-completing.
And we don't even care for diagnostics when code-completing, anyway.
llvm-svn: 170979
uncovered.
This required manually correcting all of the incorrect main-module
headers I could find, and running the new llvm/utils/sort_includes.py
script over the files.
I also manually added quite a few missing headers that were uncovered by
shuffling the order or moving headers up to be main-module-headers.
llvm-svn: 169237
to the CodeCompletionTUInfo that is going to be used to get the results.
Previously we would use ASTUnit's CodeCompletionTUInfo which has its own allocator
that will go away when we reparse. That could result in a use-after-free bug when
getting the parent context name from a CodeCompletionString.
Addresses rdar://12568377.
llvm-svn: 168133
The stat cache became essentially useless ever since we started
validating all file entries in the PCH.
But the motivating reason for removing it now is that it also affected
correctness in this situation:
-You have a header without include guards (using "#pragma once" or #import)
-When creating the PCH:
-The same header is referenced in an #include with different filename cases.
-In the PCH, of course, we record only one file entry for the header file
-But we cache in the PCH file the stat info for both filename cases
-Then the source files are updated and the header file is updated in a way that
its size and modification time are the same but its inode changes
-When using the PCH:
-We validate the headers, we check that header file and we create a file entry with its current inode
-There's another #include with a filename with different case than the previously created file entry
-In order to get its stat info we go through the cached stat info of the PCH and we receive the old inode
-because of the different inodes, we think they are different files so we go ahead and include its contents.
Removing the stat cache will potentially break clients that are attempting to use the stat cache
as a way of avoiding having the actual input files available. If that use case is important, patches are welcome
to bring it back in a way that will actually work correctly (i.e., emit a PCH that is self-contained, coping with
literal strings, line/column computations, etc.).
This fixes rdar://5502805
llvm-svn: 167172
Argyrios Kyrtzidis