Evgeniy Stepanov [Wed, 24 Oct 2012 14:05:29 +0000 (14:05 +0000)]
Pass LLVM_ANDROID_TOOLCHAIN_DIR if set.
This lets one build ASan runtime for ARM/Android by running
make -C tools/clang/runtime/ \
LLVM_ANDROID_TOOLCHAIN_DIR=/path/to/ndk/toolchain
in an existing build tree.
Ulrich Weigand [Wed, 24 Oct 2012 12:22:56 +0000 (12:22 +0000)]
A number of test cases assume that an "int" parameter or return value
will be represented in the IR as a plain "i32" type. This causes the
tests to spuriously fail on platforms where int is not a 32-bit type,
or where the ABI requires attributes like "signext" or "zeroext" to
be used.
This patch adds -triple or -target parameters to force those tests
to use the i386-unknown-unknown target.
Nadav Rotem [Wed, 24 Oct 2012 03:52:31 +0000 (03:52 +0000)]
Clang now attempts to create a TargetMachine whenever a triple is given.
Many of our tests specify triples that are not built into clang.
In this commit we allow clang to fail loading the triple if we are only
using clang to emit llvm ir.
Nadav Rotem [Wed, 24 Oct 2012 00:53:38 +0000 (00:53 +0000)]
Change EmitAssemblyHelper to create the target machine early
and use it to initialize the TargetTransformInfo analysis pass.
We need the TTI information for the loop vectorizer.
Jordan Rose [Tue, 23 Oct 2012 23:59:05 +0000 (23:59 +0000)]
[analyzer] Replace -analyzer-no-eagerly-trim-egraph with graph-trim-interval.
After every 1000 CFGElements processed, the ExplodedGraph trims out nodes
that satisfy a number of criteria for being "boring" (single predecessor,
single successor, and more). Rather than controlling this with a cc1 option,
which can only disable this behavior, we now have an analyzer-config option,
'graph-trim-interval', which can change this interval from 1000 to something
else. Setting the value to 0 disables reclamation.
The next commit relies on this behavior to actually test anything.
Douglas Gregor [Tue, 23 Oct 2012 22:26:28 +0000 (22:26 +0000)]
Make DiagnosticOptions intrusively reference-counted, and make sure
the various stakeholders bump up the reference count. In particular,
the diagnostics engine now keeps the DiagnosticOptions object alive.
Jordan Rose [Tue, 23 Oct 2012 21:54:03 +0000 (21:54 +0000)]
CMake: Fix public header search for generating Xcode/MSVC projects.
Previously, we only had support for one level of library under lib/,
with the existence of the two-level lib/StaticAnalyzer/* hardcoded in
the top-level CMakeLists.txt. This became a problem with split of
libRewrite into several libraries -- with the same sub-names as the
libraries in lib/StaticAnalyzer/.
Now, we match up anything under lib/ to the corresponding directory
in include/clang/.
Eli Friedman [Tue, 23 Oct 2012 20:19:32 +0000 (20:19 +0000)]
Add a new warning -Wmissing-variable-declarations, to warn about variables
defined without a previous declaration. This is similar to
-Wmissing-prototypes, but for variables instead of functions.
Richard Smith [Tue, 23 Oct 2012 19:56:01 +0000 (19:56 +0000)]
When rebuilding a DependentScopeDeclRefExpr, perform a lookup into the scope
even if it's dependent, in case it now names a member of the current instantiation.
Douglas Gregor [Tue, 23 Oct 2012 06:18:24 +0000 (06:18 +0000)]
If the precompiled header named by "-include" is actually a directory,
check each of the files within that directory to determine if any of
them is an AST file that matches the language and target options. If
so, the first matching AST file is loaded. This fixes a longstanding
discrepency with GCC's precompiled header implementation.
Richard Smith [Tue, 23 Oct 2012 00:32:41 +0000 (00:32 +0000)]
Ugly ugly hack for libstdc++-4.6 and libstdc++-4.7 compatibility. These
libraries have an incorrect definition of std::common_type (inherited from a
bug in the standard -- see LWG issue 2141), whereby they produce reference
types when they should not.
If we instantiate a typedef named std::common_type<...>::type, which is defined
in a system header as decltype(... ? ... : ...), and the decltype produces a
reference type, convert it to the non-reference type. (This doesn't affect any
LWG2141-conforming implementation of common_type, such as libc++'s, because the
default implementation of common_type<...>::type isn't supposed to produce a
reference type.)
This is horrible. I'm really sorry. :( Better ideas appreciated!
Douglas Gregor [Mon, 22 Oct 2012 23:59:45 +0000 (23:59 +0000)]
Handle implicitly-included PCH files the same way as
implicitly-included PTH files during initialization, delaying the
mapping down to the "original source file" until after later in the
initialization process.
Douglas Gregor [Mon, 22 Oct 2012 23:51:00 +0000 (23:51 +0000)]
Allow clients of the AST reader to specify what kinds of AST load
failures they know how to tolerate, e.g., out-of-date input files or
configuration/version mismatches. Suppress the corresponding
diagnostics if the client can handle it.
Douglas Gregor [Mon, 22 Oct 2012 22:50:17 +0000 (22:50 +0000)]
Distinguish the various kinds of AST file loading failures:
file corruption, compiler version mismatch, target/language
configuration mismatch, out-of-date AST file. No functionality change
yet.
Each option has a set of prefixes. When matching an argument such as
-funroll-loops. First the leading - is removed as it is a prefix. Then
a lower_bound search for "funroll-loops" is done against the option table by
option name. From there each option prefix + option name combination is tested
against the argument.
This allows us to support Microsoft style options where both / and - are valid
prefixes. It also simplifies the cases we already have where options come in
both - and -- forms. Almost every option for gnu-ld happens to have this form.
Douglas Gregor [Mon, 22 Oct 2012 18:42:04 +0000 (18:42 +0000)]
Eliminate the redundancy between source-file information in the source
manager block and input-file information in the control block. The
source manager entries now point back into the control block. Input
files are now lazily deserialized (if validation is disabled). Reduces
Cocoa's PCH by the ~70k I added when I introduced the redundancy in
r166251.
Richard Smith [Sun, 21 Oct 2012 23:00:34 +0000 (23:00 +0000)]
PR14141 (part of DR1351): An implicitly-deduced "any" exception specification
produces an exception of 'noexcept(false)' and is thus compatible with an
explicit exception specification of 'noexcept(false)'.
Simon Atanasyan [Sun, 21 Oct 2012 11:44:57 +0000 (11:44 +0000)]
Add support of MIPS n32 ABI to the Clang driver. The fix builds correct library/object files paths and passes appropriate command line options to the linker if user provides -mabi=n32 option.
Richard Smith [Sun, 21 Oct 2012 03:28:35 +0000 (03:28 +0000)]
Unrevert r166268, reverted in r166272, with a fix for the issue which Nick
found: if an overloaded operator& is present before a template definition,
the expression &T::foo is represented as a CXXOperatorCallExpr, not as a
UnaryOperator, so we didn't notice that it's permitted to reference a non-static
data member of an unrelated class.
While investigating this, I discovered another problem in this area: we are
treating template default arguments as unevaluated contexts during substitution,
resulting in performing incorrect checks for uses of non-static data members in
C++11. That is not fixed by this patch (I'll look into this soon; it's related
to the failure to correctly instantiate constexpr function templates), but was
resulting in this bug not firing in C++11 mode (except with -Wc++98-compat).
Original message:
PR14124: When performing template instantiation of a qualified-id outside of a
class, diagnose if the qualified-id instantiates to a non-static class member.
Lang Hames [Sun, 21 Oct 2012 01:10:01 +0000 (01:10 +0000)]
When used in a compound expression FP_CONTRACT should proceed all explicit
declarations and statements. Emit an error if the FP_CONTRACT is used
later in a compound statement.
Richard Smith [Sat, 20 Oct 2012 08:26:51 +0000 (08:26 +0000)]
Rework implementation of DR1492: Apply the resolution to operator delete too,
since it also has an implicit exception specification. Downgrade the error to
an extwarn, since at least for operator delete, system headers like to declare
it as 'noexcept' whereas the implicit definition does not have an explicit
exception specification. Move the exception specification for user-declared
'operator delete' functions from the type-as-written into the type, to reflect
reality and to allow us to detect whether there was an implicit exception spec
or not.
Nico Weber [Sat, 20 Oct 2012 06:18:14 +0000 (06:18 +0000)]
Accept -Wno-arc-abi without warning for a while.
Xcode 4.5 passes -Wno-arc-abi to clang, which makes a clang newer than
r163917 warn that it doesn't understand -Wno-arc-abi. I asked if adding
this is ok at
http://lists.cs.uiuc.edu/pipermail/cfe-commits/Week-of-Mon-20121015/066433.html
and nobody objected.
Adding this makes life a bit easier for the chromium project. If you think
this is a burden on for clang, shout, and I'll revert this.
Jordan Rose [Sat, 20 Oct 2012 02:32:51 +0000 (02:32 +0000)]
[analyzer] Assume 'new' never returns NULL if it could throw an exception.
This is actually required by the C++ standard in
[basic.stc.dynamic.allocation]p3:
If an allocation function declared with a non-throwing
exception-specification fails to allocate storage, it shall return a
null pointer. Any other allocation function that fails to allocate
storage shall indicate failure only by throwing an exception of a type
that would match a handler of type std::bad_alloc.
We don't bother checking for the specific exception type, but just go off
the operator new prototype. This should help with a certain class of lazy
initalization false positives.
When associating file ranges of macro arguments with their
macro expansion ranges, make sure to check all the FileID
entries that are contained in the spelling range of the
expansion for the macro argument.
projects / clang / log