[SEH] Delete the 32-bit IR lowering for __finally blocks and use x64
32-bit finally funclets are intended to be called both directly from the
parent function and indirectly from the EH runtime. Because we aren't
contorting LLVM's X86 prologue to match MSVC's, calling the finally
block directly passes in a different value of EBP than the one that the
runtime provides. We need an adapter thunk to adjust EBP to the expected
value. However, WinEHPrepare already has to solve this problem when
cleanups are not pre-outlined, so we can go ahead and rely on it rather
than duplicating work.
Now we only do the llvm.x86.seh.recoverfp dance for 32-bit SEH filter
functions.
Richard Smith [Wed, 1 Jul 2015 19:32:54 +0000 (19:32 +0000)]
[modules] Don't make out-of-line member specializations of an instantiated
class template specialization visible just because the class template
specialization's definition is visible.
This re-lands r236052 and adds support for __exception_code().
In 32-bit SEH, the exception code is not available in eax. It is only
available in the filter function, and now we arrange to load it and
store it into an escaped variable in the parent frame.
As a consequence, we have to disable the "catch i8* null" optimization
on 32-bit and always generate a filter function. We can re-enable the
optimization if we detect an __except block that doesn't use the
exception code, but this probably isn't worth optimizing.
[DebugInfo] Fix debug info generation for function static variables, typedefs, and records
Function static variables, typedefs and records (class, struct or union) declared inside
a lexical scope were associated with the function as their parent scope, rather than the
lexical scope they are defined or declared in.
-frewrite-includes: Rework how includes and modules are differentiated
The map of FileChange structs here was storing two disjoint types of
information:
1. A pointer to the Module that an #include directive implicitly
imported
2. A FileID and FileType for an included file. These would be left
uninitialized in the Module case.
This change splits these two kinds of information into their own maps,
which both simplifies how we access either and avoids the undefined
behaviour we were hitting due to the uninitialized fields in the
included file case.
Mostly NFC, but fixes some errors found by self-host with ubsan.
Richard Smith [Wed, 1 Jul 2015 02:29:35 +0000 (02:29 +0000)]
[modules] Before checking whether the controlling macro of a header is defined,
update the identifier in case we've imported a definition of the macro (and
thus the contents of the header) from a module.
Also fold ExternalIdentifierLookup into ExternalPreprocessorSource; it no longer
makes sense to keep these separate now that the only user of the former also
needs the latter.
CodeGen: Do not give local-linkage functions externally available linkage, even temporarily.
When an internal-linkage thunk is code gen'd, CodeGenVTables::emitThunk
will first be called with ForVTable=true (which incorrectly set the
thunk's linkage to available_externally under the Itanium ABI) and later
with ForVTable=false (which reset it to internal). Because we will always
see a call with ForVTable=false, this incorrect linkage never ended up in
the final IR. However, the temporary presence of this linkage caused us
to give such functions a comdat as a result of code introduced in r241102.
To avoid this, check that the thunk is externally visible before giving it
available_externally linkage.
Richard Smith [Wed, 1 Jul 2015 01:51:38 +0000 (01:51 +0000)]
[modules] Make the include guard optimization fire a bit more when considering
re-entering a modular header.
When we do the include guard check, we're in the visibility state for the file
with the #include; the include guard may not be visible there, but we don't
actually need it to be: if we've already parsed the submodule we're considering
entering, it's always safe to skip it.
Use a stable sort to guarantee target feature ordering in the IR
in order to make testing somewhat more feasible. Has the advantage
of making it easier to find target features as well.
CodeGen: Resize LifetimeExtendedCleanupHeader to avoid alignment issues
The LifetimeExtendedCleanupHeader is carefully fit into 32 bytes,
meaning that cleanups on the LifetimeExtendedCleanupStack are *always*
allocated at a misaligned address and cause undefined behaviour.
There are two ways to solve this - add padding after the header when
we allocated our cleanups, or just simplify the header and let it use
64 bits in the first place. I've opted for the latter, and added a
static assert to avoid the issue in the future.
Fix sse4 for target attribute feature additions.
This reinstates part of the hack removed in r233223, by special
casing sse4 as part of the feature additions. The notable change
here is that we consider it only as part of setting the SSE level
and not as part of the actual target features set which handles
setting the rest of the masks.
Fix a TODO dealing with canonicalizing attributes on functions by
using a string map to canonicalize. Fix up a couple of testcases
that needed changing since we are no longer simply appending features
to the list, but all of their mask dependencies as well.
Richard Smith [Tue, 30 Jun 2015 21:29:55 +0000 (21:29 +0000)]
[modules] Make sure macros get made visible in the top-level file if we've got
local submodule visibility enabled; that top-level file might not actually be
the module includes buffer if use of prebuilt modules is disabled.
Previously we were not assigning a comdat to thunks in the Microsoft ABI,
which would have required us to emit these functions outside of a comdat.
(Due to an inconsistency in how we were emitting objects, we were getting this
right most of the time, but only when compiling with function sections.) This
code generator change causes us to create a comdat for each thunk.
Adrian Prantl [Tue, 30 Jun 2015 17:39:43 +0000 (17:39 +0000)]
Add a function to ExternalASTSource that returns a descriptor that
abstracts the commonalities between modules and PCH files that are
needed to emit debug info for a module or precompiled header.
Reid Kleckner [Tue, 30 Jun 2015 16:32:04 +0000 (16:32 +0000)]
[clang-cl] Use /arch: to set the base target CPU
The main effect of this change is that /arch:IA32 will use i386 as the
CPU, while clang-cl will continue to default to pentium4 (aka SSE2 plus
the usual other features).
/arch:AVX and /arch:AVX2 will also now enable the other features
available in sandybridge and haswell respectively, which is consistent
with MSDN.
Hubert Tong [Tue, 30 Jun 2015 12:14:52 +0000 (12:14 +0000)]
[CONCEPTS] Parsing of concept keyword
Summary: This change adds parsing for the concept keyword in a
declaration and tracks the location. Diagnostic testing added for
invalid use of concept keyword.
David Majnemer [Tue, 30 Jun 2015 04:41:18 +0000 (04:41 +0000)]
[CodeGen] Tweak isTriviallyRecursive further
isTriviallyRecursive is a hack used to bridge a gap between the
expectations that source code assumes and the semantics that LLVM IR can
provide. Specifically, asm labels on functions are treated as an
explicit name for a GlobalObject in Clang but treated like an
output-processing step in GCC. Tweak this hack a little further to emit
calls to library functions instead of emitting an incorrect definition.
The definition in question would have available_externally linkage (this
is OK) but result in a call to itself which will either result in an
infinite loop or stack overflow.
David Majnemer [Tue, 30 Jun 2015 03:30:26 +0000 (03:30 +0000)]
[MS ABI] Workaround corner-case bug in the ABI for operator delete
MSVC only genreates array cookies if the class has a destructor. This
is problematic when having to call T::operator delete[](void *, size_t)
because the second argument's argument is impossible to synthesize
correctly if the class has no destructor (because there will be no array
cookie).
Instead, MSVC passes the size of the class. Do the same, for
compatibility, instead of crashing.
Andrew Wilkins [Tue, 30 Jun 2015 02:52:38 +0000 (02:52 +0000)]
Sphinx-based clang man pages
Summary:
This diff introduces .rst files, Sphinx config, and a CMake target
for building clang man pages. This will deprecate the existing .pod-
based man page, and will integrate nicely with CMake. This diff does
not remove the existing man page; that will be done in a follow-up
once packagers have had a chance to react to the change.
For now, only clang(1) has been done; others can be added over time
by dropping additional files into the docs/CommandGuide directory.
The index page for CommandGuide has been copied from LLVM's
docs/CommandGuide.
The man page itself is mostly the same, with a few minor cosmetic
changes. The only major change is the SYNOPSIS section. I was unable
to get .rst/Sphinx produce the same style as in the existing man page.
Instead, I changed it to match the LLVM tools' relatively simple style.
To build the man pages, use the "docs-clang-man" target if building
with CMake. Otherwise, use "make -f Makefile.sphinx man".
Adrian Prantl [Tue, 30 Jun 2015 02:26:03 +0000 (02:26 +0000)]
Pass HeaderSearchOptions and PreprocessorOptions into CodeGenModule.
In order to produce debug info for clang modules CGDebugInfo it needs
access to macros passed on the command line and the isysroot.
Eric Christopher [Mon, 29 Jun 2015 21:00:05 +0000 (21:00 +0000)]
Add support for the x86 builtin __builtin_cpu_supports.
This matches the implementation of the gcc support for the same
feature, including checking the values set up by libgcc at runtime.
The structure looks like this:
unsigned int __cpu_vendor;
unsigned int __cpu_type;
unsigned int __cpu_subtype;
unsigned int __cpu_features[1];
with a set of enums to match various fields that are field out after
parsing the output of the cpuid instruction.
This also adds a set of errors checking for valid input (and cpu).
compiler-rt support for this and the other builtins in this family
(__builtin_cpu_init and __builtin_cpu_is) are forthcoming.
Serge Pavlov [Mon, 29 Jun 2015 17:50:19 +0000 (17:50 +0000)]
Instantiation of local class members.
If a function containing a local class is instantiated, instantiate
all of local class member, including default arguments and exception
specifications.
This change fixes PR21332 and thus implements DR1484.
David Blaikie [Mon, 29 Jun 2015 17:29:50 +0000 (17:29 +0000)]
Account for calling convention specifiers in function definitions in IR test cases
Several tests wouldn't pass when executed on an armv7a_pc_linux triple
due to the non-default arm_aapcs calling convention produced on the
function definitions in the IR output. Account for this with the
application of a little regex.
Summary: This makes the Mozilla style defaults more compliant with the Mozilla style guide. A few options were removed in order to use the LLVM style defaults.
Javed Absar [Mon, 29 Jun 2015 09:30:19 +0000 (09:30 +0000)]
[ARM]: Extend -mfpu options for half-precision and vfpv3xd
Some of the the permissible ARM -mfpu options, which are supported in GCC,
are currently not present in llvm/clang.This patch adds the options:
'neon-fp16', 'vfpv3-fp16', 'vfpv3-d16-fp16', 'vfpv3xd' and 'vfpv3xd-fp16.
These are related to half-precision floating-point and single precision.
David Majnemer [Mon, 29 Jun 2015 00:06:50 +0000 (00:06 +0000)]
[MS ABI] Unify constant and non-constant member pointer conversion
We had two separate paths for member pointer conversion: one which
takes a constant and another which takes an arbitrary value. In the
latter case, we are permitted to construct arbitrary instructions.
It turns out that the bulk of the member pointer conversion is sharable
if we construct an artificial IRBuilder.
Nemanja Ivanovic [Fri, 26 Jun 2015 19:27:20 +0000 (19:27 +0000)]
Add missing builtins to altivec.h for ABI compliance (vol. 1)
This patch corresponds to review:
http://reviews.llvm.org/D10637
This is the first round of additions of missing builtins listed in the ABI document. More to come (this builds onto what seurer already addes). This patch adds:
vector signed long long vec_abs(vector signed long long)
vector double vec_abs(vector double)
vector signed long long vec_add(vector signed long long, vector signed long long)
vector unsigned long long vec_add(vector unsigned long long, vector unsigned long long)
vector double vec_add(vector double, vector double)
vector double vec_and(vector bool long long, vector double)
vector double vec_and(vector double, vector bool long long)
vector double vec_and(vector double, vector double)
vector signed long long vec_and(vector signed long long, vector signed long long)
vector double vec_andc(vector bool long long, vector double)
vector double vec_andc(vector double, vector bool long long)
vector double vec_andc(vector double, vector double)
vector signed long long vec_andc(vector signed long long, vector signed long long)
vector double vec_ceil(vector double)
vector bool long long vec_cmpeq(vector double, vector double)
vector bool long long vec_cmpge(vector double, vector double)
vector bool long long vec_cmpge(vector signed long long, vector signed long long)
vector bool long long vec_cmpge(vector unsigned long long, vector unsigned long long)
vector bool long long vec_cmpgt(vector double, vector double)
vector bool long long vec_cmple(vector double, vector double)
vector bool long long vec_cmple(vector signed long long, vector signed long long)
vector bool long long vec_cmple(vector unsigned long long, vector unsigned long long)
vector bool long long vec_cmplt(vector double, vector double)
vector bool long long vec_cmplt(vector signed long long, vector signed long long)
vector bool long long vec_cmplt(vector unsigned long long, vector unsigned long long)
Anna Zaks [Fri, 26 Jun 2015 17:42:58 +0000 (17:42 +0000)]
[static analyzer] Analyzer is skipping forward declared C/C++ functions
A patch by Karthik Bhat!
This patch fixes a regression introduced by r224398. Prior to r224398
we were able to analyze the following code in test-include.c and report
a null deref in this case. But post r224398 this analysis is being skipped.
E.g.
// test-include.c
#include "test-include.h"
void test(int * data) {
data = 0;
*data = 1;
}
// test-include.h
void test(int * data);
This patch uses the function body (instead of its declaration) as the location
of the function when deciding if the Decl should be analyzed with path-sensitive
analysis. (Prior to r224398, the call graph was guaranteed to have a definition
when available.)
Manuel Klimek [Fri, 26 Jun 2015 02:15:04 +0000 (02:15 +0000)]
Fix crash-on-invalid bug in template instantiation.
Get rid of code-path that (according to Richard Smith) is not needed but
leads to a crasher bug when assuming a template has been fully
instantiated and thus has a definition.
Davide Italiano [Fri, 26 Jun 2015 00:18:35 +0000 (00:18 +0000)]
[Sema] Commit a better fix for r240242
Skip calls to HasTrivialDestructorBody() in the case where the
destructor is never invoked. Alternatively, Richard proposed to change
Sema to declare a trivial destructor for anonymous union member, which
seems too wasteful.
Nico Weber [Fri, 26 Jun 2015 00:13:18 +0000 (00:13 +0000)]
Add an inttypes.h wrapper that fixes up some macros in Microsoft mode.
Before MSVS2015, MSVS's headers disagree about int32_t and PRIx32 and so on.
Provide a wrapper header to fix this, so that -Wformat can still be used.
Fixes PR23412.
Meador Inge [Fri, 26 Jun 2015 00:09:55 +0000 (00:09 +0000)]
[Sema] Maintain ellipsis location when transforming lambda captures
This patch fixes a crash caused by the following case:
template<typename T>
auto f(T x) {
auto g = [](auto ... args) {
auto h = [args...]() -> int {
return 0;
};
return h;
};
return g;
}
auto x = f(0)();
When the templated function 'f' is instantiated and the inner-most
lambda is transformed the ellipsis location on the captured variable
is lost. Then the lambda returned by 'f' is instantiated and the
tree transformer chokes on the invalid ellipsis location. The
problem is fixed by making a minor change to properly track the
ellipsis location.
Kaelyn Takata [Thu, 25 Jun 2015 23:47:39 +0000 (23:47 +0000)]
Fix a typo correction crash when resolving ambiguous corrections.
In certain cases, the tree transform would introduce new TypoExprs while
trying one of the corrections, invalidating the unique_ptr in the state
reference, and also causing a TypoExpr to exist that will never be
corrected since it doesn't exist in the final corrected expression. The
simple solution to both problems is to temporarily disable typo
correction while handling potentially ambiguous typo corrections.
Anna Zaks [Thu, 25 Jun 2015 23:36:44 +0000 (23:36 +0000)]
[docs] Several updates to the Address Sanitizer webpage.
- Added the description of the interceptor suppression.
- Re-organized a bit: grouped a few things under the Issue Suppression
section, grouped IOC and leaks under a section, placed symbolication
info into Symbolizing the Reports section..
- In supported platforms: "MacOS" -> "OS X"; added "iOS Simulator"
- Added a paragraph to the Usage section describing when DYLD_INSERT_LIBRARIES
might need to be used.
- "attribute((no_sanitize_address))" -> "__attribute__((no_sanitize("address")))"
- Updated Leak Sanitizer page with most up to date info.
....
Sean Silva [Thu, 25 Jun 2015 23:22:11 +0000 (23:22 +0000)]
Remove `requires` for x86 CPU features.
Ever since the target attributes change, we don't need to guard these
headers with `requires`. Actually it's a bit worse, because if we do
then they are included textually under the covers, causing declarations
to appear in submodules they aren't supposed to be in.
Alexey Samsonov [Thu, 25 Jun 2015 23:14:32 +0000 (23:14 +0000)]
[CFI] Diagnose when we CFI in diagnostic mode is unavailable on a toolchain.
Summary:
Namely, we must have proper C++ABI support in UBSan runtime. We don't
have a good way to check for that, so just assume that C++ABI support is
there whenever -fsanitize=vptr is supported (i.e. only on handful of
platforms).
Exact diagnostic is also tricky. It's not "cfi" that is unsupported,
just the diagnostic mode. So, I suggest to report that
"-fno-sanitize-trap=cfi-foobar" is incompatible with a given target
toolchain.
Diego Novillo [Thu, 25 Jun 2015 22:56:00 +0000 (22:56 +0000)]
Display profile file name when emitting a file not found diagnostic.
When a profile file cannot be opened, we used to display just the error
message but not the name of the profile the compiler was trying to open.
This will become useful in the next set of patches that introduce
GCC-compatible flags to specify profiles.
projects / clang / log