Problem with gnu conditional extension with missing
LHS and when conditional expression is an array. Since
it will be decayed, saved expression must be saved with
decayed expression. This is necessary to preserve semantics
of this extension (and prevent an IRGen crash which expects
an array to always be decayed). I am sure there will be other
cases in c++ (aggregate conditionals for example) when saving of the
expression must happen after some transformation on conditional
expression has happened.
Doug, please review. Fixes // rdar://8446940
John McCall [Sat, 18 Sep 2010 05:25:11 +0000 (05:25 +0000)]
static local variables with destructors don't require a global destructor
unless we're on a platform without __cxa_atexit (or use thereof has been
disabled). This patch actually just disables the check completely for
static locals, but I've filed http://llvm.org/bugs/show_bug.cgi?id=8176 to
track the platform-specific fix.
John McCall [Sat, 18 Sep 2010 02:24:39 +0000 (02:24 +0000)]
Adjust a fixup's starting branch if it's being resolved because
it reached the outermost scope and it hasn't yet been forwarded
to a cleanup. Fixed PR8175.
Tom Care [Sat, 18 Sep 2010 00:06:34 +0000 (00:06 +0000)]
IdempotentOperationChecker no longer reports errors that are post-dominated by a sink. This fixes some false positives that were flagged because a path was incomplete (usually in a loop).
Douglas Gregor [Fri, 17 Sep 2010 22:25:06 +0000 (22:25 +0000)]
When we run into an error parsing or type-checking the left-hand side
of a binary expression, continue on and parse the right-hand side of
the binary expression anyway, but don't call the semantic actions to
type-check. Previously, we would see the error and then, effectively,
skip tokens until the end of the statement.
The result should be more useful recovery, both in the normal case
(we'll actually see errors beyond the first one in a statement), but
it also helps code completion do a much better job, because we do
"real" code completion on the right-hand side of an invalid binary
expression rather than completing with the recovery completion. For
example, given
x = p->y
if there is no variable named "x", we can still complete after the p->
as a member expression. Along the recovery path, we would have
completed after the "->" as if we were in an expression context, which
is mostly useless.
David Chisnall [Fri, 17 Sep 2010 18:29:54 +0000 (18:29 +0000)]
Add a -ftrapv-handler= option which allows a handler to invoke instead of simply aborting when a signed operation overflows. This mirrors the (GCC-incompatible) behaviour from clang 1.0 and 1.1 when -ftrapv was specified, but allows the handler to be defined for each compilation unit.
Use a temporary file for output which gets renamed after all the writing is finished.
This mainly prevents failures and/or crashes when multiple processes try to read/write the same PCH file. (rdar://8392711&8294781); suggestion & review by Daniel!
Patch to add IRgen support for Gnu's conditional operator
extension when missing LHS. This patch covers scalar
conditionals only. Others are wip.
(pr7726, radar 8353567).
Daniel Dunbar [Fri, 17 Sep 2010 08:22:12 +0000 (08:22 +0000)]
Driver/DarwinClang: Add GCC libexec paths so we can find the cc1 binaries if we
ever use fallback.
- Not necessary for clang as a compiler, but useful for testing purposes.
Daniel Dunbar [Fri, 17 Sep 2010 03:46:37 +0000 (03:46 +0000)]
Revert "Driver/Darwin: Time for another spin at switching to the DarwinClang
toolchain.", while I investigate the totally non-surprising ensuing breakage.
John McCall [Fri, 17 Sep 2010 02:31:44 +0000 (02:31 +0000)]
Currently we're initializing the vtable pointers of a class only after
the bases are completely initialized. This won't work --- base
initializer expressions can rely on the vtables having been set up.
Check for uses of 'this' in the initializers and force a vtable
initialization if found.
This might not be good enough; we might need to extend this to handle
the possibility of arbitrary code finding an external reference to this
(not yet completely-constructed!) object and accessing through it,
in which case we'll probably find ourselves doing a lot more unnecessary
stores.
Daniel Dunbar [Fri, 17 Sep 2010 01:16:06 +0000 (01:16 +0000)]
Driver/DarwinClang: The new toolchain definition is going to drop the -L inside
the GCC dir. Unfortunately, this breaks -lstdc++ on SnowLeopard, etc. because
the libstdc++ dylib was hiding there. Workaround this by providing the path to
the right -lstdc++.6 (the only version used in recent memory) if we can't see an
obvious -lstdc++, but can find = -lstdc++.6.
John McCall [Fri, 17 Sep 2010 00:50:28 +0000 (00:50 +0000)]
When emitting a new-expression inside a conditional expression,
the cleanup might not be dominated by the allocation code.
In this case, we have to store aside all the delete arguments
in case we need them later. There's room for optimization here
in cases where we end up not actually needing the cleanup in
different branches (or being able to pop it after the
initialization code).
Also make sure we only call this operator delete along the path
where we actually allocated something.
Daniel Dunbar [Fri, 17 Sep 2010 00:45:02 +0000 (00:45 +0000)]
Driver: Add magic handling for "reserved library names", starting with
-lstdc++. This is the best gross solution for a gross problem.
This issue is that historically, GCC has add -L options to its internally
library directories. This has allowed users and platforms to end up depending on
the layout of GCC's internal library directories.
We want to correct this mistake by eliminating that -L, but this means that
existing libraries which are in the GCC lib dir won't be found. We are going to
handle this by treating those -l names as "reserved", and requiring toolchains
to know how to add the right full path to the reserved library.
The immediately side effect of this is that users trying to use -L to find their
own -lstdc++ will need to start using -nostdlib (which is a good idea
anyway). Another side effect is that -stdlib=libc++ -lstdc++ will now do the
"right" thing, for curious definitions of right.
Douglas Gregor [Thu, 16 Sep 2010 23:58:57 +0000 (23:58 +0000)]
When dealing with an anonymous enumeration declared in function
prototype scope, temporarily set the context of the enumeration
declaration to the translation unit. We do the same thing for
parameters, until we have an actual function declaration on which to
hang them. Fixes <rdar://problem/8435682>.
There is more work to do in this area, since we have existing bugs
with tags being declared/defined in function parameter lists. This fix
is correct, and we'll end up extending it when we deal with those
existing bugs.
Daniel Dunbar [Thu, 16 Sep 2010 20:42:00 +0000 (20:42 +0000)]
IRgen/ABI/x86_32/Darwin: On Darwin, only structures with SSE vector types get passed
with a non-default-stack-ABI-alignment (of 16).
- This fixes the ABI convenient, but breaks codegen since we now have
underaligned arguments. Marginal improvement overall though, and will be
fixed in next commit.
Daniel Dunbar [Thu, 16 Sep 2010 20:41:56 +0000 (20:41 +0000)]
IRgen/x86_32/Linux: Linux seems to align all stack objects to 4 bytes, unlike
Darwin. Checked vs the handiest Linux llvm-gcc I had around, someone on Linux is
welcome to investigate more.
Douglas Gregor [Thu, 16 Sep 2010 16:06:31 +0000 (16:06 +0000)]
Don't add two code-completion results for the same selector; it
doesn't add any value. Instead, we'll just take the first method with
that selector that we find and create a completion for it.
Douglas Gregor [Thu, 16 Sep 2010 15:14:18 +0000 (15:14 +0000)]
Implement code completion for Objective-C class message sends that are
missing the opening bracket '[', e.g.,
NSArray <CC>
at function scope. Previously, we would only give trivial completions
(const, volatile, etc.), because we're in a "declaration name"
scope. Now, we also provide completions for class methods of NSArray,
e.g.,
alloc
Note that we already had support for this after the first argument,
e.g.,
NSArray method:x <CC>
would get code completion for class methods of NSArray whose selector
starts with "method:". This was already present because we recover
as if NSArray method:x were a class message send missing the opening
bracket (which was committed in r114057).
John McCall [Thu, 16 Sep 2010 06:57:56 +0000 (06:57 +0000)]
Right, there are *two* cases of pr-value class-type expressions that don't
derive from temporaries of the same type. Black-list member expressions
as well.
Tom Care [Thu, 16 Sep 2010 03:50:38 +0000 (03:50 +0000)]
Refactored BugReporter to refer to EndNode as ErrorNode. We currently make the assumption that EndNode == ErrorNode, but upcoming changes will break this.
Douglas Gregor [Thu, 16 Sep 2010 01:51:54 +0000 (01:51 +0000)]
Implement automatic bracket insertion for Objective-C class message
sends. These are far trickier than instance messages, because we
typically have something like
NSArray alloc]
where it appears to be a declaration of a variable named "alloc" up
until we see the ']' (or a ':'), and at that point we can't backtrace.
So, we use a combination of syntactic and semantic disambiguation to
treat this as a message send only when the type is an Objective-C type
and it has the syntax of a class message send (which would otherwise
be ill-formed).
For self-comparison warning, check the source location of both the LHS and RHS to see if they
are expanded from macros (and if so, omit the warning). Previously we were just looking at the
location of the binary expression.
While handling change of file, check if _current_ file is already seen or not. If current file is seen then it indicates that end of previous file's lexical scope.
This fixes radar 8396182.
Douglas Gregor [Wed, 15 Sep 2010 16:23:04 +0000 (16:23 +0000)]
Improve code completion for Objective-C message sends when the opening
'[' is missing. Prior commits improving recovery also improved code
completion beyond the first selector, e.g., at or after the "to" in
calculator add:x to:y
but not after "calculator". We now provide the same completions for
calculator <CC>
that we would for
[calculator <CC>
if "calculator" is an expression whose type is something that can
receive Objective-C messages.
This code completion works for instance and super message sends, but not
class message sends.
Douglas Gregor [Wed, 15 Sep 2010 14:51:05 +0000 (14:51 +0000)]
Implement bracket insertion for Objective-C instance message sends as
part of parser recovery. For example, given:
a method1:arg];
we detect after parsing the expression "a" that we have the start of a
message send expression. We pretend we've seen a '[' prior to the a,
then parse the remainder as a message send. We'll then give a
diagnostic+fix-it such as:
fixit-objc-message.m:17:3: error: missing '[' at start of message
send expression
a method1:arg];
^
[
The algorithm here is very simple, and always assumes that the open
bracket goes at the beginning of the message send. It also only works
for non-super instance message sends at this time.
John McCall [Wed, 15 Sep 2010 10:14:12 +0000 (10:14 +0000)]
one piece of code is responsible for the lifetime of every aggregate
slot. The easiest way to do that was to bundle up the information
we care about for aggregate slots into a new structure which demands
that its creators at least consider the question.
I could probably be convinced that the ObjC 'needs GC' bit should
be rolled into this structure.
Implement generalized copy elision. The main obstacle here is that
IR-generation must be much more careful about making sure that exactly
Relax assertion in CFG builder when processing ForStmts. This fixes an assertion failure
on code containing GNU statement expressions reported in PR 8141.
Daniel Dunbar [Wed, 15 Sep 2010 03:53:45 +0000 (03:53 +0000)]
build: Start sketching code to allow grafting libc++ build into LLVM/Clang build if
libc++ is checked out into llvm/projects.
- WIP, not on by default yet.
Disallow the use of UnknownVal as the index for ElementRegions. UnknownVals can be used as
the index when the value evaluation isn't powerful enough. By creating ElementRegions with
UnknownVals as the index, this gives the false impression that they are the same element, when
they really aren't. This becomes really problematic when deriving symbols from these regions
(e.g., those representing the initial value of the index), since two different indices will
get the same symbol for their binding.
This fixes an issue with the idempotent operations checker that would cause two indices that
are clearly not the same to make it appear as if they always had the same value.
Fix Windows64 target info so pointer arithmetic is done correctly, and no sign extension code is emitted: PtrDiffType needs to be a signed long long. Add a corresponding test case.
Douglas Gregor [Tue, 14 Sep 2010 23:59:36 +0000 (23:59 +0000)]
Introduce a new code-completion context for a parenthesized
expression, e.g., after the '(' that could also be a type cast. Here,
we provide types as code-completion results in C/Objective-C (C++
already had them), although we wouldn't in a normal expression context.
projects / clang / log