Ted Kremenek [Thu, 13 Nov 2008 17:11:24 +0000 (17:11 +0000)]
Using llvm::OwningPtr<> for CurLexer and CurTokenLexer. This makes both the ownership semantics of these objects explicit within the Preprocessor and also tightens up the code (explicit deletes not needed).
Ted Kremenek [Thu, 13 Nov 2008 16:51:03 +0000 (16:51 +0000)]
Use PushIncludeMacroStack/PopMacroStack instead of manually pushing/popping from IncludeMacroStack. This is both cleaner and makes the include stack transparently extensible.
Zhongxing Xu [Thu, 13 Nov 2008 07:58:20 +0000 (07:58 +0000)]
Lift the pointer to alloca'ed region to the pointer to its first element.
This is required by some operations, e.g., *p = 1; p[0] = 1;.
Also set the AllocaRegion's type during the cast.
Ted Kremenek [Thu, 13 Nov 2008 06:10:40 +0000 (06:10 +0000)]
GRExprEngine/CFRefCount/GRSimpleVals: We don't do any special handling (yet) of vector types. Add explicit checks that when we process integers that they really are scalars.
Eli Friedman [Thu, 13 Nov 2008 06:09:17 +0000 (06:09 +0000)]
Fix for crash issues with comma operators with a void first operand, and
some more bullet-proofing/enhancements for tryEvaluate. This shouldn't
cause any behavior changes except for handling cases where we were
crashing before and being able to evaluate a few more cases in tryEvaluate.
This should settle the minor mess surrounding r59196.
Daniel Dunbar [Thu, 13 Nov 2008 05:29:02 +0000 (05:29 +0000)]
Obey the FIXMES!
- Resume running the always inliner pass always now that LLVM has
been improved and functions with debug info can be inlined.
- Remove unused header.
Daniel Dunbar [Thu, 13 Nov 2008 05:09:21 +0000 (05:09 +0000)]
[LLVM up] Update for raw_fd_ostream change. This fixes a FIXME that
the Backend output should be done in binary mode.
- I'd appreciate it if someone who has a Windows build could verify
this.
Eli Friedman [Thu, 13 Nov 2008 02:13:11 +0000 (02:13 +0000)]
Backout of r59196, plus a new ICE test. Sorry if this is a
little rude; I figure it's cleaner to just back this out now so
it doesn't get forgotten or mixed up with other checkins.
The modification to isICE is simply wrong; I've added a test that the
change to isICE breaks.
I'm pretty sure the modification to tryEvaluate is also wrong.
At the very least, there's some serious miscommunication going on here,
as this is going in exactly the opposite direction of r59105. My
understanding is that tryEvaluate is not supposed to care about side
effects. That said, a lot of the clients to tryEvaluate are
expecting it to enforce a no-side-effects policy, so we probably need
another method that provides that guarantee.
Daniel Dunbar [Thu, 13 Nov 2008 01:54:24 +0000 (01:54 +0000)]
Supply finished flag to EmitBlock for common statements which use
landing pads.
- Primarily a cleanliness issue instead of a performance issue (this
eliminates all blocks w/o predecessors on 176.gcc/expr.c), but this
also allows subsequent code to recognize it is unreachable and
potentially avoid IRgen.
Daniel Dunbar [Thu, 13 Nov 2008 01:24:05 +0000 (01:24 +0000)]
Add IsFinished arg to EmitBlock.
- Indicates that caller is done with the block and it can be dropped
if it has no predecessors. Useful for callers who need to make
landing pads but which may not be reached.
Daniel Dunbar [Thu, 13 Nov 2008 00:03:19 +0000 (00:03 +0000)]
Fix bug in constant evaluation exposed by 176.gcc.
- Evaluation of , operator used bogus assumption that LHS could be
evaluated as an integral expression even though its type is
unspecified.
This change is making isICE very permissive of the LHS in non-evaluated
contexts because it is not clear what predicate we would use to reject
code here. The standard didn't offer me any guidance; opinions?
Douglas Gregor [Wed, 12 Nov 2008 23:21:09 +0000 (23:21 +0000)]
Don't build identifiers for C++ constructors, destructors, or
conversion functions. Instead, we just use a placeholder identifier
for these (e.g., "<constructor>") and override NamedDecl::getName() to
provide a human-readable name.
This is one potential solution to the problem; another solution would
be to replace the use of IdentifierInfo* in NamedDecl with a different
class that deals with identifiers better. I'm also prototyping that to
see how it compares, but this commit is better than what we had
previously.
Ted Kremenek [Wed, 12 Nov 2008 22:43:05 +0000 (22:43 +0000)]
Move LexIncludeFilename from Lexer to PreprocessorLexer.
PreprocessorLexer now has a virtual method "IndirectLex" which allows it to call the lex method of its subclasses. This is not for performance intensive operations.
Ted Kremenek [Wed, 12 Nov 2008 22:10:22 +0000 (22:10 +0000)]
Add Preprocessor::PushIncludeMacroStack() and Preprocessor::PopIncludeMacroStack(), two utility methods for manipulating the Preprocessor stack. These will be used to remove manually manipulation of IncludeMacroStack from the rest of the Preprocessor implementation.
Ted Kremenek [Wed, 12 Nov 2008 21:33:59 +0000 (21:33 +0000)]
Move pieces of Lexer that the Preprocessor mutates to a new base class 'PreprocessorLexer'. This will also be the base class of the new Preprocessed-Token-Header (PTH) lexer. No functionality change.
GRState:
- Environment now binds to Stmt* instead of Expr*. This is needed for processing ObjCForCollectionStmt (essentially the declaration of the the 'element' variable can have an SVal attached to it).
- BindDecl no longer accepts Expr* for the initialization value; use SVal* instead.
Ted Kremenek [Wed, 12 Nov 2008 19:18:35 +0000 (19:18 +0000)]
StoreManager::BindDecl now takes an SVal* for the initialization value instead of an Expr* (which can be null). Lazy symbolication of conjured symbols is now the sole responsibility of GRExprEngine.
Douglas Gregor [Wed, 12 Nov 2008 17:17:38 +0000 (17:17 +0000)]
Implement support for operator overloading using candidate operator
functions for built-in operators, e.g., the builtin
bool operator==(int const*, int const*)
can be used for the expression "x1 == x2" given:
struct X {
operator int const*();
} x1, x2;
The scheme for handling these built-in operators is relatively simple:
for each candidate required by the standard, create a special kind of
candidate function for the built-in. If overload resolution picks the
built-in operator, we perform the appropriate conversions on the
arguments and then let the normal built-in operator take care of it.
There may be some optimization opportunity left: if we can reduce the
number of built-in operator overloads we generate, overload resolution
for these cases will go faster. However, one must be careful when
doing this: GCC generates too few operator overloads in our little
test program, and fails to compile it because none of the overloads it
generates match.
Note that we only support operator overload for non-member binary
operators at the moment. The other operators will follow.
As part of this change, ImplicitCastExpr can now be an lvalue.
Eli Friedman [Wed, 12 Nov 2008 09:44:48 +0000 (09:44 +0000)]
Some additions to tryEvaluate I've had sitting around for a while.
This pushes it a lot closer to being able to deal with most of the stuff
CodeGen's constant expression evaluator knows how to deal with. This
also fixes PR3003.
The test could possibly use some improvement, but this'll work for now.
Test 6 is inspired by PR3003; the other tests are mostly just designed
to exercise the new code. The reason for the funny structure of the
tests is that type fixing for arrays inside of structs is the only place
in Sema that calls tryEvaluate, at least for the moment.
Daniel Dunbar [Wed, 12 Nov 2008 08:21:33 +0000 (08:21 +0000)]
IRgen improvements on unreachable code:
- Split out "simple" statements which can easily handle IR generation
when there is no insert point. These are generally statements which
start by emitting a new block or are only containers for other
statements.
- This fixes a regression in emitting dummy blocks, notably for case
statements.
- This also fixes spurious emission of a number of debug stoppoint
intrinsic instructions.
Remove unneeded sw.body block, just clear the insertion point.
Lift out CodeGenFunction::EmitStopPoint which calls into the
CGDebugInfo class when generating debug info.
Normalize definitions of Emit{Break,Continue}Stmt and usage of
ErrorUnsupported.
Chris Lattner [Wed, 12 Nov 2008 08:04:58 +0000 (08:04 +0000)]
Move EmitBranchOnBoolExpr and ConstantFoldsToSimpleInteger to
CodeGenFunction.cpp. Change VisitConditionalOperator to use
constant fold instead of codegen'ing a constant conditional.
Change ForStmt to use EmitBranchOnBoolExpr, this shrinks
expr.c very slightly to 40239 lines.
Chris Lattner [Wed, 12 Nov 2008 07:43:42 +0000 (07:43 +0000)]
fix a crash analyzing constants in 176.gcc/expr.c with my next patch. It was
crashing because we errors are ignored in subexpressions that are not evaluated,
but we still evaluate the result of parents. This would cause an assertion
because the erroneous subexpr didn't have its result value set to the right type.
Daniel Dunbar [Tue, 11 Nov 2008 23:11:34 +0000 (23:11 +0000)]
Rework IRgen invariant w.r.t. current insert point.
- EmitStmt is no longer required to finish with a current insertion
point defined (i.e. it does not need to make dummy
blocks). Instead, it can clear the insertion point in the builder
which indicates that the current insertion point is unreachable.
- CodeGenFunction provides HaveInsertPoint and EnsureInsertPoint
which respectively test if there is an insert point and ensure an
insertion point exists (by making a dummy block).
- Clearly mark functions in CodeGenFunction which can be called with
no insertion point defined. Currently this is a limited set, and
EmitStmt simply EnsureInsertPoint()s before emitting subsequent IR.
Remove EmitDummyBlock, which is no longer needed. Clients who haven't
already cleared the insertion point (typically via EmitBranch) can do
so by hand.
Remove isDummyBlock, which has effectively been renamed to
HaveInsertPoint.
The main thrust of this change is that we no longer have create dummy
blocks just to destroy them a short time later in EmitBlock in the
common case that there is no unreachable code following something like
a goto.
Additionally, this means that we are not using the hokey condition in
isDummyBlock that a block without a name is a dummy block. Guess how
well that works when we never emit block names!
Ted Kremenek [Tue, 11 Nov 2008 19:41:42 +0000 (19:41 +0000)]
Flow-sensitive uninitialized values analysis:
- Added support for ObjCForCollectionStmt
- Fixed bug where expression values would be always set to uninitialized when loops were involved
Chris Lattner [Tue, 11 Nov 2008 07:41:27 +0000 (07:41 +0000)]
short circuit && and || when possible. This substantially reduces
the size of the -O0 output on some cases. For example, on expr.c from
176.gcc, it shrinks the .ll file from 43164 to 42835 lines, and removed
references to two external symbols.
projects / clang / log