David Tweed [Fri, 13 Sep 2013 12:04:22 +0000 (12:04 +0000)]
Certain multi-platform languages, such as OpenCL, have the concept of
address spaces which is both (1) a "semantic" concept and
(2) possibly a hardware level restriction. It is desirable to
be able to discard/merge the LLVM-level address spaces on arguments for which
there is no difference to the current backend while keeping
track of the semantic address spaces in a funciton prototype. To do this
enable addition of the address space into the name-mangling process. Add
some tests to document this behaviour against inadvertent changes.
David Majnemer [Fri, 13 Sep 2013 09:40:55 +0000 (09:40 +0000)]
[-cxx-abi microsoft] Mangle user defined entry points properly
Summary:
Functions named "main", "wmain", "WinMain", "wWinMain", and "DllMain"
are never mangled regardless of linkage, even when compiling for kernel
mode.
Depends on D1655
Summary:
This is a first step to getting extern "C" working properly inside
clang. There are a number of quirks but mangling declarations inside
such a function are a good first step.
Richard Trieu [Fri, 13 Sep 2013 03:20:53 +0000 (03:20 +0000)]
Refactor the uninitialized field visitor. Also moved the calls to the visitor
later in the code so that the expressions will have addition processing first.
This catches a few additional cases of uninitialized uses of class fields.
James Dennett [Fri, 13 Sep 2013 02:46:09 +0000 (02:46 +0000)]
Documentation cleanup: Fixing \brief comments, migrating away from old style
that duplicated the name of the entity being documented at the start of its
comment, and other minor tidyups.
Hal Finkel [Thu, 12 Sep 2013 23:57:55 +0000 (23:57 +0000)]
Restore the sqrt -> llvm.sqrt mapping in fast-math mode
This restores the sqrt -> llvm.sqrt mapping, but only in fast-math mode
(specifically, when the UnsafeFPMath or NoNaNsFPMath CodeGen options are
enabled). The @llvm.sqrt* intrinsics have slightly different semantics from the
libm call, specifically, they are undefined when given a non-zero negative
number (the libm calls will always return NaN for any negative number).
This mapping was removed in r100613, and replaced with a TODO, but at that time
the fast-math flags were not yet implemented. Now that we have these, restoring
this mapping is important because it will enable autovectorization of sqrt
calls in loops (at least in fast-math mode).
Richard Smith [Thu, 12 Sep 2013 23:28:08 +0000 (23:28 +0000)]
PR13657 (and duplicates):
When a comma occurs in a default argument or default initializer within a
class, disambiguate whether it is part of the initializer or whether it ends
the initializer.
The way this works (which I will be proposing for standardization) is to treat
the comma as ending the default argument or default initializer if the
following token sequence matches the syntactic constraints of a
parameter-declaration-clause or init-declarator-list (respectively).
This is both consistent with the disambiguation rules elsewhere (where entities
are treated as declarations if they can be), and should have no regressions
over our old behavior. I think it might also disambiguate all cases correctly,
but I don't have a proof of that.
There is an annoyance here: because we're performing a tentative parse in a
situation where we may not have seen declarations of all relevant entities (if
the comma is part of the initializer, lookup may find entites declared later in
the class), we need to turn off typo-correction and diagnostics during the
tentative parse, and in the rare case that we decide the comma is part of the
initializer, we need to revert all token annotations we performed while
disambiguating.
Any diagnostics that occur outside of the immediate context of the tentative
parse (for instance, if we trigger the implicit instantiation of a class
template) are *not* suppressed, mirroring the usual rules for a SFINAE context.
Richard Smith [Thu, 12 Sep 2013 18:49:10 +0000 (18:49 +0000)]
PR16054: Slight strengthening for -Wsometimes-uninitialized: if we use a
variable uninitialized every time we reach its (reachable) declaration, or
every time we call the surrounding function, promote the warning from
-Wmaybe-uninitialized to -Wsometimes-uninitialized.
This is still slightly weaker than desired: we should, in general, warn
if a use is uninitialized the first time it is evaluated.
Hans Wennborg [Thu, 12 Sep 2013 18:23:34 +0000 (18:23 +0000)]
Move Compilation::PrintJob and PrintDiagnosticJob into Job::Print.
This moves the code to Job.cpp, which seems like a more natural fit,
and replaces the "is this a JobList? is this a Command?" logic with
a virtual function call.
It also removes the code duplication between PrintJob and
PrintDiagnosticJob and simplifies the code a little.
There's no functionality change here, except that the Executable is
now always printed within quotes, whereas it would previously not be
quoted in crash reports, which I think was a bug.
Jordan Rose [Thu, 12 Sep 2013 16:17:41 +0000 (16:17 +0000)]
[CMake] Put controversial always-recheck-revision-number behind an option.
CMake does not have the ability to perform actions before calculating
dependencies, so it can't know whether it needs to rebuild clangBasic
to update for a new revision number. CLANG_ALWAYS_CHECK_VC_REV (off by
default) will cause clangBasic to always be dirty by deleting the
generated SVNVersion.inc after use; otherwise, SVNVersion.inc will
always be updated, but only included in the final binary when clangBasic
is rebuilt.
It'd be great to find a better way to do this, but hopefully this is
still an improvement over the complete lack of version information before.
Jordan Rose [Wed, 11 Sep 2013 22:46:46 +0000 (22:46 +0000)]
[CMake] Always include the Clang repo version, just like the autoconf build.
Now that LLVM's helper script GetSVN.cmake actually works consistently,
there's no reason not to use it. This does mean that the clangBasic target
is potentially always dirty, because CMake-generated projects do not
necessarily recalculate dependencies after running each target.
This should end the issues of the AST format changing and breaking old
module files; CMake-Clang should now detect that the version changed just
like Autoconf-Clang has.
Tablegen now generates a StringSwitch for attributes containing enumeration arguments to map strings to the proper enumeration value. This makes error checking more consistent and reduces the amount of hand-written code required.
ObjectiveC migrator. Modify inferred property name
such that it does not lower case the staring property
name if getter name (after "get" prefix) starts with
two upper case letters.
Test for correct usage of columnWidth in clang fixit hints.
Summary:
This test only works on systems capable of outputting UTF-8 encoded
text on the standard output (tested on linux and OS X, should XFAIL on windows,
if I haven't messed up the XFAIL line).
Jordan Rose [Wed, 11 Sep 2013 16:46:50 +0000 (16:46 +0000)]
[analyzer] Handle zeroing constructors for fields of structs with empty bases.
RegionStore tries to protect against accidentally initializing the same
region twice, but it doesn't take subregions into account very well. If
the outer region being initialized is a struct with an empty base class,
the offset of the first field in the struct will be 0. When we initialize
the base class, we may invalidate the contents of the struct by providing
a default value of Unknown (or some new symbol). We then go to initialize
the member with a zeroing constructor, only to find that the region at
that offset in the struct already has a value. The best we can do here is
to invalidate that value and continue; neither the old default value nor
the new 0 is correct for the entire struct after the member constructor call.
The correct solution for this is to track region extents in the store.
Michael Han [Wed, 11 Sep 2013 15:53:29 +0000 (15:53 +0000)]
Teach RAV to visit parameter variable declarations of implicit functions. Fixes PR16182.
Normally RAV visits parameter variable declarations of a function by traversing the TypeLoc of
the parameter declarations. However, for implicit functions, their parameters don't have any
TypeLoc, because they are implicit.
So for implicit functions, we visit their parameter variable declarations by traversing them through
the function declaration, and visit them accordingly.
Summary:
reformat() tries to determine the newline style used in the input
(either LF or CR LF), and uses it for the output. Maybe not every single case is
supported, but at least the bug described in http://llvm.org/PR17182 should be
resolved.
Daniel Jasper [Wed, 11 Sep 2013 10:37:35 +0000 (10:37 +0000)]
Split -Wunused-variable warning.
With r190382, -Wunused-variable warns about unused const variables when
appropriate. For codebases that use -Werror, this poses a problem as
existing unused const variables need to be cleaned up first. To make the
transistion easier, this patch splits -Wunused-variable by pulling out
an additional -Wunused-const-variable (by default activated along with
-Wunused-variable).
SemaTemplateDeduction.cpp: Prune two stray \param(s), TPOC and NumCallArguments in Sema::getMostSpecialized(). They were removed since r190444. [-Wdocumentation]
Daniel Jasper [Wed, 11 Sep 2013 07:20:44 +0000 (07:20 +0000)]
Support for modular module-map-files
This patch is the first step to make module-map-files modular (instead
of requiring a single "module.map"-file per include directory). This
step adds a new "extern module" declaration that enables
module-map-files to reference one another along with a very basic
implementation.
The next steps are:
* Combine this with the use-declaration (from
http://llvm-reviews.chandlerc.com/D1546) in order to only load module
map files required for a specific compilation.
* Add an additional flag to start with a specific module-map-file (instead
of requiring there to be at least one "module.map").
Summary:
More accurately characterize the nature of array parameters. Doing this
removes false back-reference opportunities. Remove some hacks now that
we characterize these better.
Richard Smith [Wed, 11 Sep 2013 00:52:39 +0000 (00:52 +0000)]
PR17075: When performing partial ordering of a member function against a
non-member function, the number of arguments in the two candidate calls
will be different (the non-member call will have one extra argument).
We used to get confused by this, and fail to compare the last argument
when testing whether the member is better, resulting in us always
thinking it is, even if the non-member is more specialized in the last
argument.
Some build systems use pipes for stdin/stderr. On nix-ish platforms colored
output can be forced by -fcolor-diagnostics. On Windows this option has
no effect in these cases because LLVM uses the console API (which only
operates on the console buffer) even if a console wrapper capable of
interpreting ANSI escape codes is used.
The -fansi-escape-codes option allows switching from the console API to
ANSI escape codes. It has no effect on other platforms.
Richard Smith [Tue, 10 Sep 2013 22:59:25 +0000 (22:59 +0000)]
getMostSpecialized for function template sets is never used in the context of a
call; remove its 'number of explicit arguments' and 'what kind of call'
parameters.
Eli Friedman [Tue, 10 Sep 2013 22:57:15 +0000 (22:57 +0000)]
Add unused markings to suppress warnings.
trunk clang is a bit more aggressive about emitting unused-declaration
warnings, so adjust some AST code to match. Specifically, use
LLVM_ATTRIBUTE_UNUSED for declarations which are never supposed to be
referenced, and turn references to declarations which are supposed to be
referenced into odr-uses.
Richard Smith [Tue, 10 Sep 2013 21:34:14 +0000 (21:34 +0000)]
PR5683: Issue a warning when subtracting pointers to types of zero size, and
treat such subtractions as being non-constant. Patch by Serge Pavlov! With a
few tweaks by me.
Hans Wennborg [Tue, 10 Sep 2013 20:18:04 +0000 (20:18 +0000)]
clang-cl: Support building DLLs (PR17083)
This adds driver support for building DLLs (the /LD and /LDd flags).
It basically does two things: runtime selection and passing -dll and
-implib to the linker.
[ms-cxxabi] Implement guard variables for static initialization
Static locals requiring initialization are not thread safe on Windows.
Unfortunately, it's possible to create static locals that are actually
externally visible with inline functions and templates. As a result, we
have to implement an initialization guard scheme that is compatible with
TUs built by MSVC, which makes thread safety prohibitively difficult.
MSVC's scheme is that every function that requires a guard gets an i32
bitfield. Each static local is assigned a bit that indicates if it has
been initialized, up to 32 bits, at which point a new bitfield is
created. MSVC rejects inline functions with more than 32 static locals,
and the externally visible mangling (?_B) only allows for one guard
variable per function.
On Eli's recommendation, I used MangleNumberingContext to track which
bit each static corresponds to.
Hans Wennborg [Tue, 10 Sep 2013 15:41:12 +0000 (15:41 +0000)]
Allow _clang-format as alternative to .clang-format config filename
Dotfiles are impractical on Windows. This makes clang-format search
for the style configuration file as '_clang-format' in addition to
the usual '.clang-format'. This is similar to how VIM searches for
'_vimrc' on Windows.
PR16752: 'mode' attribute for unusual targets doesn't work properly
Description:
Troubles could be happened due to some assumptions in handleModeAttr function (see SemaDeclAttr.cpp).
For example, it assumes that 32 bit integer is 'int', while it could be 16 bit only.
Instead of asking target: 'which type do you want to use for int32_t ?' it just hardcodes general opinion. That doesn't looks pretty correct.
Please consider the next solution:
1. In Basic/TargetInfo add getIntTypeByWidth and getRealTypeByWidth methods. Methods asks target for proper type for given bit width.
2. Fix handleModeAttr according to new methods in TargetInfo.
Fixes:
1st Commit (Done): Add new methods for TargetInfo:
getRealTypeByWidth and getIntTypeByWidth
for ASTContext names are almost same(invokes new methods from TargetInfo):
getIntTypeForBitwidth and getRealTypeForBitwidth
Separate popcnt and sse4.2 feature control somewhat to match gcc behavior.
Enabling sse4.2 will implicitly enable popcnt unless popcnt is explicitly disabled.
Disabling sse4.2 will not disable popcnt if popcnt is explicitly enabled.
Eli Friedman [Tue, 10 Sep 2013 03:05:56 +0000 (03:05 +0000)]
Make -Wunused warning rules more consistent.
This patch does a few different things.
This patch improves unused var diags for const vars: we no longer
unconditionally suppress diagnostics for const vars, instead only suppressing
the diagnostic when the declaration appears to be useful.
This patch also makes us more consistently use whether a variable/function
is declared in the main file to suppress diagnostics where appropriate.
Ignore calling conventions when checking function template specializations
Summary:
Calling conventions are inherited during decl merging. Before this
change, deduction would fail due to a type mismatch between the template
and the specialization. This change adjusts the CCs to match before
deduction, and lets the decl merging logic diagnose mismatch or inherit
the CC from the template.
This allows specializations of static member function templates in the
Microsoft C++ ABI.
Attribute tablegen now understands that attribute arguments can be optional. This allows for automated checking of the number of arguments expected vs number of arguments given for attributes. Greatly reduces the amount of manual checking required.
Modern ObjectiveC translator. Fix translation of
message sent to aggregate-valued methods. Fix
visibility of trampoline type used in translation
of such expressions. // rdar://14932320
Sean Silva [Mon, 9 Sep 2013 19:50:47 +0000 (19:50 +0000)]
[docs] End users should see just "Clang"
Half the internet has wildly incorrect ideas about what LLVM is (in
particular, many get the impression that LLVM is some magical
cross-platform runtime), so saying "LLVM" near "cross-compilation" in a
user-facing doc might provoke confusion.
projects / clang / log