Ted Kremenek [Wed, 14 Nov 2007 17:46:35 +0000 (17:46 +0000)]
SourceManager, IdentifierTable, Selectors are now serialized in their own
block separate from ASTContext. This block is serialized out AFTER writing out
ASTContext, but deserialized BEFORE reading in ASTContext. This permits the
optimization of the serialization of the IdentifierTable where we only write
out identifiers that are used.
This was needed because TagDecls are owned by Types, and TagDecls contain
identifiers. Thus types need to be written out first to register with the
serializer any identifiers they refer to (and hence need to be serialized out
with IdentifierTable).
Ted Kremenek [Wed, 14 Nov 2007 08:06:37 +0000 (08:06 +0000)]
Implemented serialization of RecordDecls. Changed serialization of TagType to
have an owning pointer to the referred TagDecl. This should hopefully fix a
bug where TagDecls (including decls from structs, etc.) were not serialized.
Ted Kremenek [Wed, 14 Nov 2007 00:03:20 +0000 (00:03 +0000)]
Modified ASTContext::getTagDeclType() to accept a NULL pointer for the passed
in TagDecl*. This allows the deserializer to use ASTContext to create the
TagTypes. Deserialize TagTypes then rely on pointer-backpatching to resolve
the decls.
This may not be the interface that we want, but as the implementation of
TagTypes will potentially change significantly in the future, I'm leaving this
for now. An appropriate FIXME is in place.
Ted Kremenek [Tue, 13 Nov 2007 22:02:55 +0000 (22:02 +0000)]
Rewrote type serialization to used the same methodology as we do for Decls.
Removed tons of dead code in ASTContext concerning how types use to be
serialized.
Removed serialization methods from QualType that are no longer used.
Ted Kremenek [Tue, 13 Nov 2007 00:15:39 +0000 (00:15 +0000)]
Restructured serialization code for decls to make it cleaner, easier to
understand, and batched the emission owned subobjects (using
BatchEmitOwnedPtr) to get a smaller output bitcode size.
Steve Naroff [Mon, 12 Nov 2007 19:18:37 +0000 (19:18 +0000)]
Add an error diagnostic to Parse::ParseObjCMessageExpression().
This now exposes the following bug...
******************** TEST 'Sema/message.m' FAILED! ********************
Command:
clang -fsyntax-only -verify Sema/message.m
Output:
Errors seen but not expected:
Line 9: invalid receiver to message expression
******************** TEST 'Sema/message.m' FAILED! ********************
As far as I can tell, all messages to method agruments fail.
The method arguments are built by Sema::ObjcActOnStartOfMethodDef().
Steve Naroff [Mon, 12 Nov 2007 13:56:41 +0000 (13:56 +0000)]
- Minor cleanup to yesterday's changes to Sema::ObjcActOnStartOfMethodDef();
- Add Sema::CurMethodDecl, in preparation for adding ObjcIvarRefExpr.
- Add ObjcInterfaceDecl::lookupInstanceVariable(), in prep for adding ivars.
- A couple renames in ObjcInterfaceDecl, while I was in the vicinity:-)
Steve Naroff [Mon, 12 Nov 2007 03:44:46 +0000 (03:44 +0000)]
Remove Sema::ObjcBuildMethodParameter().
Modify Sema::ParseParamDeclarator() to work for both ActOnStartOfFunctionDef() and ObjcActOnStartOfMethodDef().
Steve Naroff [Sun, 11 Nov 2007 23:20:51 +0000 (23:20 +0000)]
Replace 2 method definition actions (ActOnFunctionDefBody, ActOnMethodDefBody) with 1 method definition action (ActOnFinishFunctionBody). I can't think of any reason that we would need two action hooks.
Steve Naroff [Sun, 11 Nov 2007 17:19:15 +0000 (17:19 +0000)]
This is the last 5% of the solution to teaching Sema::ActOnInstanceMessage() about private methods (r43989).
While the diff is large, the idea is very simple. When we parse method definitions (in an @implementation), we need to add them incrementally (rather than wait until the @end).
Other details...
- Renamed Sema::ActOnAddMethodsToObjcDecl() to Sema::ActOnAtEnd(). The methods are now optional arguments.
- Removed Parser::AllImplMethods (a nice cleanup).
- Added location info to ObjcImplementationDecl (since we will need it very soon:-)
- Modified message.m test to no longer allow the bogus diagnostic.
Steve Naroff [Sun, 11 Nov 2007 00:10:47 +0000 (00:10 +0000)]
Teach Sema::ActOnInstanceMessage() about private methods. That is, methods declared in an implementation (but not listed in the interface).
This commit is only 95% of the bug fix. The last piece to this puzzle is to add the method decls to the implementation incrementally (as we encounter them). At the moment, the methods aren't added until we see an @end (which is too late).
Steve Naroff [Sat, 10 Nov 2007 18:06:36 +0000 (18:06 +0000)]
Since we are all working on Leopard, change the built-in MacOS X version number (__ENVIRONMENT_MAC_OS_X_VERSION_MIN_REQUIRED__) to target Leopard. This is only a short term solution, until the clang driver implements the "-mmacosx-version-min" flag (which will compute this value based on the target).
Ted Kremenek [Sat, 10 Nov 2007 02:11:55 +0000 (02:11 +0000)]
Changed the serialization of IdentifierTable to only serialize out entries
that are referenced in the ASTs. This assumes that we serialize out the
decls/stmts first, and use the pointer-tracking logic in the Serializer to
determine if an IdentifierInfo (or its string key) is ever referenced.
This is a significant space optimization for serialized ASTs.
Consider the following program:
void foo(int x,int y) {
return x > y ? x : y+1;
}
Here are the sizes of the files for the serialized ASTs:
Full IdentifierTable: 23676 bytes
Only-referenced Identifiers: 304 bytes.
For this simple program, this is a 77% reduction in the file size of the
serialized ASTs.
Ted Kremenek [Sat, 10 Nov 2007 02:07:12 +0000 (02:07 +0000)]
Modifed the test serialization driver to...
(1) serialize out top-level decls BEFORE serializing out translation unit
structures like ASTContext.
(2) deserialize out translation unit structures like ASTContext before
top-level decls by first skipping the decls in the bitstream, deserializing
ASTContext and friends, and then jumping back to the bitstream block with the
decls and then deserializing them.
Change (1) allows us to utilize the pointer-tracking system in the Serializer
to only serialize out metadata that is actually referenced by the ASTS.
Change (2) allows us to deserialize the metadata first as before, which
signficantly reduces the amount of pointer backpatching the deserializer
would have to do if the decls were deserialized first.
Chris Lattner [Fri, 9 Nov 2007 23:52:16 +0000 (23:52 +0000)]
change source location to have two bits for macros, tracking
whether the location is the start and/or end of an expansion.
These are currently not set or used by anything.
Ted Kremenek [Thu, 8 Nov 2007 19:52:41 +0000 (19:52 +0000)]
Rewrote serialization of IdentifierInfo and IdentifierTable to use methods Emit
and Materialize/Read instead of using specializations of SerializeTrait<>. The
resulting code is much cleaner. We are also setting the stage so that only the
parts of the IdentifierTable that are ever referenced within the ASTs are
serialized, and not the whole table.
Ted Kremenek [Thu, 8 Nov 2007 16:32:00 +0000 (16:32 +0000)]
Implemented serialization of AddrLabelExpr.
Implemented serialization of ConditionalOperator.
Implemented serialization of StmtExpr.
Fixed bug in serialization of IndirectGoto (did not properly serialize subexpression).
Chris Lattner [Thu, 8 Nov 2007 07:35:14 +0000 (07:35 +0000)]
Add a new RewriteRope data structure which is a smarter way to represent the text
backing a rewrite buffer than using an std::vector<char>. This class was hacked
together very quickly and needs to be cleaned up, but it seems to work. It speeds
up rewriting a a 7M file from 6.43s to 0.24s on my machine. The impl could also
be made to be a lot more algorithmically sound.
This produces identical output to using vector on this testcase, if it causes a
problems or bugs are encountered, it can be disabled by changing the
RewriteBuffer::Buffer typedef back.
projects / clang / log