Eli Friedman [Thu, 15 Sep 2011 23:15:27 +0000 (23:15 +0000)]
Tweak *mmintrin.h so that they don't make any bad assumptions about alignment (which probably has little effect in practice, but better to get it right). Make the load in _mm_loadh_pi and _mm_loadl_pi a single LLVM IR instruction to make optimizing easier for CodeGen.
Douglas Gregor [Thu, 15 Sep 2011 22:00:41 +0000 (22:00 +0000)]
Add an experimental flag -fauto-module-import that automatically turns
#include or #import direcctives of framework headers into module
imports of the corresponding framework module.
Anna Zaks [Thu, 15 Sep 2011 20:06:34 +0000 (20:06 +0000)]
[analyzer] Refactor: make PathDiagnosticLocation responsible for validation of SourceLocations (commit 4 of ?):
- The closing brace is always a single location, not a range.
- The test case previously had a location key 57:1 followed by a range [57:1 - 57:1].
Douglas Gregor [Thu, 15 Sep 2011 19:48:59 +0000 (19:48 +0000)]
Eliminate the list of modules from the preprocessor options. This was
used back when we had an -import-module command-line option, but it's
no longer used (or useful).
Anna Zaks [Thu, 15 Sep 2011 18:56:07 +0000 (18:56 +0000)]
[analyzer] Refactor: make PathDiagnosticLocation responsible for validation of SourceLocations (commit 2 of ?):
- Fix a fixme and move the logic of creating a PathDiagnosticLocation corresponding to a ProgramPoint into a PathDiagnosticLocation constructor.
- Rename PathDiagnosticLocation::create to differentiate from the added constructor.
Douglas Gregor [Thu, 15 Sep 2011 18:47:32 +0000 (18:47 +0000)]
When we load the first module, make sure that we wire up the ASTConsumer to the newly-created ASTReader. This makes sure that CodeGen sees the declarations it is interested in
[PCH] Overhaul how preprocessed entities are [de]serialized.
-Use an array of offsets for all preprocessed entities
-Get rid of the separate array of offsets for just macro definitions;
for references to macro definitions use an index inside the preprocessed
entities array.
-Deserialize each preprocessed entity lazily, at first request; not in bulk.
Paves the way for binary searching of preprocessed entities that will offer
efficiency and will simplify things on the libclang side a lot.
John McCall [Thu, 15 Sep 2011 06:49:18 +0000 (06:49 +0000)]
Rewrite this loop to use partial destruction; I'm not sure it's
possible for that to matter right now, but eventually I think we'll
need to unify this better, and then it might. Also, use a more
efficient looping structure.
Anna Zaks [Thu, 15 Sep 2011 01:08:34 +0000 (01:08 +0000)]
[analyzer] Refactor: make PathDiagnosticLocation responsible for validation of SourceLocations (commit 2 of ?):
- Modify all PathDiagnosticLocation constructors that take Stmt to also requre LocationContext.
- Add a constructor which should be used in case there is no valid statement/location (it will grab the location of the enclosing function).
Douglas Gregor [Wed, 14 Sep 2011 20:28:46 +0000 (20:28 +0000)]
Teach the driver to always pass down a module cache path. If none is
supplied, use something derived from the system's temporary
directory. Depends on LLVM r139725.
Thread safety: adding additional documentation to the main thread safety interface, and making the destructor for the thread safety handler pure virtual
Anna Zaks [Wed, 14 Sep 2011 17:48:01 +0000 (17:48 +0000)]
[analyzer] After CFG has been linearized, we can have a situation where an ExpoledNode has an invalid SourceLocation (which has no correspondence in the source code). This commit is the first step to solve this problem.
- It adds LocationContext to the PathDiagnosticLocation object and uses it to lookup the enclosing statement with a valid location.
- So far, the LocationContext is only available when the object is constructed from the ExplodedNode.
- Already found some subtle bugs(in plist-output-alternate.m) where the intermediate diagnostic steps were not previously shown.
Anna Zaks [Wed, 14 Sep 2011 00:25:17 +0000 (00:25 +0000)]
[analyzer] Refactor: Make PathDiagnosticLocation responsible for creating a valid object given an ExploadedNode (the same logic can be reused by other checkers).
Douglas Gregor [Tue, 13 Sep 2011 23:15:45 +0000 (23:15 +0000)]
For modules, use a hash of the compiler version, language options, and
target triple to separate modules built under different
conditions. The hash is used to create a subdirectory in the module
cache path where other invocations of the compiler (with the same
version, language options, etc.) can find the precompiled modules.
John McCall [Tue, 13 Sep 2011 23:08:34 +0000 (23:08 +0000)]
In general, don't look through explicit casts when trying
to find the called declaration. Explicit casts can radically
change the semantics of a call, and it's no longer really a
builtin call any more than it would be a builtin call if you stored
the function pointer into a variable and called that.
[libclang] Introduce clang_getPresumedLocation which works like clang_getExpansionLocation
but takes into account #line directives coming from preprocessed files.
[PCH] Fix a regression that r139441 introduced (decls were getting passed
to the consumer without being fully deserialized).
The regression was on compiling boost.python and it was too difficult to get a reduced
test case unfortunately.
Also modify the logic of how objc methods are getting passed to the consumer;
codegen depended on receiving objc methods before the implementation decl.
Since the interesting objc methods are ones with a body and such methods only
exist inside an ObjCImplDecl, deserialize and pass to consumer all the methods
of ObCImplDecl when we see one.
John McCall [Tue, 13 Sep 2011 18:49:24 +0000 (18:49 +0000)]
A strong property of block type has "copy" setter semantics, not "retain".
This is consistent with the behavior of assigning into a __strong l-value,
and it's also necessary for ensuring that the ivar doesn't end up a dangling
reference. We decided not to change the behavior of "retain" properties, but
just to make them warnings/errors when of block type.
Douglas Gregor [Tue, 13 Sep 2011 18:26:39 +0000 (18:26 +0000)]
Switch the serialization of LangOptions over to use the .def file. We
should no longer have the serialization of LangOptions out of sync
with the structure itself (yay).
[libclang]
-Allow cursor visitation of an attribute using its source range
-Add C++ 'final' and 'override' attributes as cursor kinds
-Simplify the logic that marks 'final' and 'override' attributes as tokens.
Douglas Gregor [Tue, 13 Sep 2011 17:21:33 +0000 (17:21 +0000)]
Switch LangOptions over to a .def file that describes header of the
language options. Use that .def file to declare the LangOptions class
and initialize all of its members, eliminating a source of annoying
initialization bugs.
Switch -Wreturn-type to completely rely on the CFG model of no-return.
This deletes a bunch of crufty code, and allows more logic sharing
between the analyzer and the warnings.
John McCall [Tue, 13 Sep 2011 07:33:34 +0000 (07:33 +0000)]
Don't use native atomics on ivars whose size is not a power of two,
even on architectures that support unaligned access (which is the
only way this is otherwise legal, given that ivars apparently do
not honor alignment attributes).
Enhance the CFG construction to detect no-return destructors for
temporary objects and local variables. When detected, these split the
block, marking the new one as having only the exit block as a successor.
This prevents a large number of false positives in warnings sensitive to
no-return constructs such as -Wreturn-type, and fixes the remainder of
PR10063 along with several variations of this bug that had not been
reported. The test cases are extended across the board to cover these
patterns.
This also checks in a stress test for these types of CFGs. The stress
test declares some 32k variables, a mixture of no-return and normal
destructors. Previously, this resulted in roughly 2500 CFG blocks, but
didn't model any of the no-return destructors. With this patch, it
results in over 33k blocks, many of them now unreachable.
The nice thing about how the analyzer is set up? This causes *no*
regression in performance of building the CFG. It actually in some cases
makes it faster, as best I can benchmark. The analysis for -Wreturn-type
(and any other that cares about no-return code paths) is technically
slower now as it has to look at many more candidate blocks, but it
computes the correct answer. I have more test cases to follow, I think
they all work now. Also I have further work that should dramatically
simplify analyses in the presence of no-return.
John McCall [Tue, 13 Sep 2011 03:34:09 +0000 (03:34 +0000)]
Unify the decision of how to emit property getters and setters into a
single code path. Use atomic loads and stores where necessary. Load and
store anything of the appropriate size and alignment with primitive
operations instead of going through the call.
Douglas Gregor [Tue, 13 Sep 2011 01:26:44 +0000 (01:26 +0000)]
When compiling a module on-demand, re-use the diagnostics client
already provided. This required a little bit of clean-up in the way
that VerifyDiagnosticsClient managed ownership of its underlying
"primary" client, because now it will no longer always take ownership.
Douglas Gregor [Mon, 12 Sep 2011 23:31:24 +0000 (23:31 +0000)]
When an import statement fails to find a module in the module cache,
but there is a corresponding umbrella header in a framework, build the
module on-the-fly so it can be immediately loaded at the import
statement. This is very much proof-of-concept code, with details to be
fleshed out over time.
projects / clang / log