Adrian Prantl [Tue, 18 Apr 2017 01:22:01 +0000 (01:22 +0000)]
Debug Info: Remove special-casing of indirect function argument handling.
LLVM has changed the semantics of dbg.declare for describing function
arguments. After this patch a dbg.declare always takes the *address*
of a variable as the first argument, even if the argument is not an
alloca.
Richard Smith [Mon, 17 Apr 2017 23:44:51 +0000 (23:44 +0000)]
Fix mishandling of escaped newlines followed by newlines or nuls.
Previously, if an escaped newline was followed by a newline or a nul, we'd lex
the escaped newline as a bogus space character. This led to a bunch of
different broken corner cases:
For the pattern "\\\n\0#", we would then have a (horizontal) space whose
spelling ends in a newline, and would decide that the '#' is at the start of a
line, and incorrectly start preprocessing a directive in the middle of a
logical source line. If we were already in the middle of a directive, this
would result in our attempting to process multiple directives at the same time!
This resulted in crashes, asserts, and hangs on invalid input, as discovered by
fuzz-testing.
For the pattern "\\\n" at EOF (with an implicit following nul byte), we would
produce a bogus trailing space character with spelling "\\\n". This was mostly
harmless, but would lead to clang-format getting confused and misformatting in
rare cases. We now produce a trailing EOF token with spelling "\\\n",
consistent with our handling for other similar cases -- an escaped newline is
always part of the token containing the next character, if any.
For the pattern "\\\n\n", this was somewhat more benign, but would produce an
extraneous whitespace token to clients who care about preserving whitespace.
However, it turns out that our lexing for line comments was relying on this bug
due to an off-by-one error in its computation of the end of the comment, on the
slow path where the comment might contain escaped newlines.
[ubsan] Skip null checks if they are constant-folded away
The IR builder can constant-fold null checks if the pointer operand
points to a constant. If the "is-non-null" check is folded away to
"true", don't emit the null check + branch.
Testing: check-clang, check-ubsan.
This slightly reduces the amount of null checks we emit when compiling
X86ISelLowering.cpp. Here are the numbers from patched/unpatched clangs
based on r300371.
[ubsan] Skip null checks on pointers to the start of an alloca
Pointers to the start of an alloca are non-null, so we don't need to
emit runtime null checks for them.
Testing: check-clang, check-ubsan.
This significantly reduces the amount of null checks we emit when
compiling X86ISelLowering.cpp. Here are the numbers from patched /
unpatched clangs based on r300371.
Eric Fiselier [Mon, 17 Apr 2017 22:06:13 +0000 (22:06 +0000)]
[coroutines] Fix rebuilding of implicit and dependent coroutine statements.
Summary:
Certain implicitly generated coroutine statements, such as the calls to 'return_value()' or `return_void()` or `get_return_object_on_allocation_failure()`, cannot be built until the promise type is no longer dependent. This means they are not built until after the coroutine body statement has been transformed.
This patch fixes an issue where these statements would never be built for coroutine templates.
It also fixes a small issue where diagnostics about `get_return_object_on_allocation_failure()` were incorrectly suppressed.
Benjamin Kramer [Mon, 17 Apr 2017 20:57:40 +0000 (20:57 +0000)]
Revert "Address http://bugs.llvm.org/pr30994 so that a non-friend can properly replace a friend, and a visible friend can properly replace an invisible friend but not vice verse, and definitions are not replaced. This fixes the two FIXME in SemaTemplate/friend-template.cpp."
This reverts commit r300443. Breaks compiling libc++ with modules in
some configurations.
Address http://bugs.llvm.org/pr30994 so that a non-friend can properly replace a friend, and a visible friend can properly replace an invisible friend but not vice verse, and definitions are not replaced. This fixes the two FIXME in SemaTemplate/friend-template.cpp.
The code implements Richard Smith suggestion in comment 3 of the PR.
Eric Fiselier [Sun, 16 Apr 2017 09:19:59 +0000 (09:19 +0000)]
[coroutines] Fix building of new/delete expressions when get_return_object_on_allocation_failure() is present.
Summary:
This patch implements [dcl.fct.def.coroutine]p8:
> The unqualified-id get_return_object_on_allocation_failure is looked up in the scope of
> class P by class member access lookup (3.4.5). If a declaration is found, ..., and if a
> global allocation function is selected, the ::operator new(size_t, nothrow_t) form shall be used.
> [...]
> The allocation function used in this case must have a non-throwing noexcept-specification.
Eric Christopher [Sat, 15 Apr 2017 06:15:00 +0000 (06:15 +0000)]
When we turn on vsx it should also turn on altivec explicitly, same
with disabling it as well as disabling all vsx specific features when
turning off altivec.
Do not run tests for crash recovery if libstdc++ safe mode is enabled
If expensive checks are enabled, safe mode of libstdc++ is enabled too.
In this mode the library uses more complex data that allow additional
checks, for instance, a container may keep list of iterators that points
to it. If a code crashes it can leave these complex library objects in
inconsistent state. It occurs in a few tests that check error recovery
if compiler crashes. These test hang in expensive check mode, as the
library tries to synchronize access to the iterators pointing to some
container, but corresponding mutex remains locked after the crash.
This fix marks these tests as unsupported if clang is built with
libstdc++ safe mode enabled.
The modules side of r299226, which serializes #pragma pack state,
doesn't work well.
The main purpose was to make -include and -include-pch match semantics
(the PCH side). We also started serializing #pragma pack in PCMs, in
the hopes of making modules and non-modules builds more consistent. But
consider:
As of r299226, the #pragma pack(show) gives "2", even though we've only
included "a.h".
- With -fmodules-local-submodule-visibility, this is clearly wrong. We
should get the default state (8 on x86_64).
- Without -fmodules-local-submodule-visibility, this kind of matches how
other things work (as if include-the-whole-module), but it's still
really terrible, and it doesn't actually make modules and non-modules
builds more consistent.
This commit disables the serialization for modules, essentially a
partial revert of r299226.
Going forward:
1. Having this #pragma pack stuff escape is terrible design (or, more
often, a horrible bug). We should prioritize adding warnings (maybe
-Werror by default?).
2. If we eventually reintroduce this for modules, it should only apply
to -fmodules-local-submodule-visibility, and it should be tracked on
a per-submodule basis.
[ubsan] Don't check alignment if the alignment is 1
If a pointer is 1-byte aligned, there's no use in checking its
alignment. Somewhat surprisingly, ubsan can spend a significant amount
of time doing just that!
This loosely depends on D30283.
Testing: check-clang, check-ubsan, and a stage2 ubsan build.
[ubsan] Reduce alignment checking of C++ object pointers
This patch teaches ubsan to insert an alignment check for the 'this'
pointer at the start of each method/lambda. This allows clang to emit
significantly fewer alignment checks overall, because if 'this' is
aligned, so are its fields.
This is essentially the same thing r295515 does, but for the alignment
check instead of the null check. One difference is that we keep the
alignment checks on member expressions where the base is a DeclRefExpr.
There's an opportunity to diagnose unaligned accesses in this situation
(as pointed out by Eli, see PR32630).
Testing: check-clang, check-ubsan, and a stage2 ubsan build.
Along with the patch from D30285, this roughly halves the amount of
alignment checks we emit when compiling X86FastISel.cpp. Here are the
numbers from patched/unpatched clangs based on r298160.
Kuba Mracek [Fri, 14 Apr 2017 16:53:25 +0000 (16:53 +0000)]
[ObjC] Fix lifetime markers of loop variable in EmitObjCForCollectionStmt [take 2]
CodeGenFunction::EmitObjCForCollectionStmt currently emits lifetime markers for the loop variable in an inconsistent way: lifetime.start is emitted before the loop is entered, but lifetime.end is emitted inside the loop. AddressSanitizer uses these markers to track out-of-scope accesses to local variables, and we get false positives in Obj-C foreach loops (in the 2nd iteration of the loop). This patch keeps the loop variable alive for the whole loop by extending ForScope and registering the cleanup function inside EmitAutoVarAlloca.
Erich Keane [Fri, 14 Apr 2017 15:21:18 +0000 (15:21 +0000)]
Make Gentoo GNU GCC Config override whitespace tolerant
The config-*triple* file handling isn't tolerant of
leading/trailing whitespace, making it not terribly
obvious when a single extraneous tab/space/etc will
cause the override to be ignored. This patch simply
trims the lines to ensure that it is tolerant of
whitespace.
Simon Pilgrim [Fri, 14 Apr 2017 15:05:57 +0000 (15:05 +0000)]
[X86][SSE] Update MOVNTDQA non-temporal loads to generic implementation (clang)
MOVNTDQA non-temporal aligned vector loads can be correctly represented using generic builtin loads, allowing us to remove the existing x86 intrinsics.
[docs] UBSan: Mention that print_stacktrace=1 is unsupported on Darwin
Printing out stack traces along with UBSan diagnostics is unsupported on
Darwin. That's because it isn't possible to use the fast unwinder or the
slow unwinder.
Apparently, it's inappropriate to use the fast unwinder for UBSan
issues. I'm not exactly sure why (see the comment in ubsan_diag.cc).
Forcing use of the fast unwinder produces decent results, AFAICT.
Darwin also does not appear to have a slow unwinder suitable for use
with the sanitizers. Apparently that's because of PR20800 [1][2]. But
that bug has been fixed. I'm not sure if there is anything preventing
use of the slow unwinder now.
Currently, passing UBSAN_OPTIONS=print_stacktrace=1 does nothing on
Darwin. This isn't good, but it might be a while before we can fix the
situation, so we should at least document it.
[1] https://github.com/google/sanitizers/issues/137
"We can't use the slow unwinder on OSX now, because Clang produces
incorrect unwind info for the ASan runtime functions on OSX
(http://llvm.org/PR20800)."
[2] https://bugs.llvm.org/show_bug.cgi?id=20800
Bug 20800 - Invalid compact unwind info generated for a function without
frame pointers on OSX
Kuba Mracek [Fri, 14 Apr 2017 00:32:43 +0000 (00:32 +0000)]
[ObjC] Fix lifetime markers of loop variable in EmitObjCForCollectionStmt
CodeGenFunction::EmitObjCForCollectionStmt currently emits lifetime markers for the loop variable in an inconsistent way: lifetime.start is emitted before the loop is entered, but lifetime.end is emitted inside the loop. AddressSanitizer uses these markers to track out-of-scope accesses to local variables, and we get false positives in Obj-C foreach loops (in the 2nd iteration of the loop). The markers of the loop variable need to be either both inside the loop (so that we poison and unpoison the variable in each iteration), or both outside. This patch implements the "both inside" approach.
Richard Smith [Thu, 13 Apr 2017 21:37:24 +0000 (21:37 +0000)]
PR32185: Revert r291512 and add a testcase for PR32185.
This reverts an attempt to check that types match when matching a
dependently-typed non-type template parameter. (This comes up when matching the
parameters of a template template parameter against the parameters of a
template template argument.)
The matching rules here are murky at best. Our behavior after this revert is
definitely wrong for certain C++17 features (for 'auto' template parameter
types within the parameter list of a template template argument in particular),
but our behavior before this revert is wrong for some pre-existing testcases,
so reverting to our prior behavior seems like our best option.
Hans Wennborg [Thu, 13 Apr 2017 17:42:45 +0000 (17:42 +0000)]
Warning-free clang-format plugin install for VS 15.0
With the new release of VS, it's required that all plugins migrate to
the new VSIX manifest format. The new format is backwards compatible
with all versions newer that Visual Studio 2012, so this migration
effectively drops support for older versions of the IDE.
It's also required that these new extensions are built with Visual
Studio 2017, so unfortunately it was necessary to migrate the project
and solution. Also removed COM references to EnvDTE and
Microsoft.VisualStudio.TextManager.Interop from the csproj, as they seem
to both be unnecessary and would trigger build warnings because of
changes to GAC.
Re-land "[clang-cl] Make all sanitizer flags available in clang-cl"
Adding RUN lines with %clang_cl was causing these tests to fail on Mac
because absolute paths there tend to start with "/User/", which is
recognized as the "/U" flag.
[analyzer] Simplify values in binary operations a bit more aggressively.
SValBuilder tries to constant-fold symbols in the left-hand side of the symbolic
expression whenever it fails to evaluate the expression directly. However, it
only constant-folds them when they are atomic expressions, not when they are
complicated expressions themselves. This patch adds recursive constant-folding
to the left-hand side subexpression (there's a lack of symmetry because we're
trying to have symbols on the left and constants on the right). As an example,
we'd now be able to handle operations similar to "$x + 1 < $y", when $x is
constrained to a constant.
[IR] Take func, ret, and arg attrs separately in AttributeList::get
This seems like a much more natural API, based on Derek Schuff's
comments on r300015. It further hides the implementation detail of
AttributeList that function attributes come last and appear at index
~0U, which is easy for the user to screw up. git diff says it saves code
as well: 97 insertions(+), 137 deletions(-)
This also makes it easier to change the implementation, which I want to
do next.
Richard Smith [Wed, 12 Apr 2017 23:21:25 +0000 (23:21 +0000)]
Update to match LLVM r300135.
Remove "REQUIRES: long_tests" from test/Driver/response-file.c since it is now about 10x faster. (We can add that back if it's still too slow for some buildbot.)
[clang-cl] Make all sanitizer flags available in clang-cl
Summary:
Use a tablegen let {} block so that new sanitizer flags are available by
default in all driver modes. This should cut down on time wasted with
bugs like http://crbug.com/710928.
Eric Fiselier [Wed, 12 Apr 2017 22:12:15 +0000 (22:12 +0000)]
[Sema] Add __is_aggregate type-trait
Summary:
[LWG 2911](http://cplusplus.github.io/LWG/lwg-defects.html#2911) adds `std::is_aggregate` to the library, which requires a new builtin trait. This patch implements `__is_aggregate`.
[analyzer] Add a check for IvarRegion in getExtraInvalidatedValues
This diff adds a defensive check in getExtraInvalidatedValues
for the case when there are no regions for the ivar associated with
a property. Corresponding test case added.
Test plan:
make check-clang
make check-clang-analysis
[modules] Delay calling DeclMustBeEmitted until it's safe.
This patch implements the suggestion in D29753 that calling DeclMustBeEmitted in
the middle of deserialization should be avoided and that the actual check should
be deferred until it's safe to do so.
This patch fixes a crash when accessing the invalid redecl chains while trying
to evaluate the value of a const VarDecl that contains a function call.
[Driver] Add compiler option to generate a reproducer
One way to currently test the reproducers is to setup
"FORCE_CLANG_DIAGNOSTICS_CRASH=1" before invoking clang. This simulates
a crash and produces the same contents needed by the reproducers. The
reproducers are specially useful when triaging Modules issues, not only
on crashes, but also for reproducing misleading warnings, errors, etc.
Add a '-gen-reproducer' driver option to clang (or any similar name) and
give users a flag option.
Note that clang already has a -fno-crash-diagnostics, which disables the
crash reproducers. I've decided not to propose "-fcrash-diagnostics"
since it doesn't convey the ideia of reproduction despite a crash.
David Blaikie [Wed, 12 Apr 2017 20:58:33 +0000 (20:58 +0000)]
Modular Codegen: Separate flags for function and debug info support
This allows using and testing these two features separately. (noteably,
debug info is, so far as I know, always a win (basically). But function
modular codegen is currently a loss for highly optimized code - where
most of the linkonce_odr definitions are optimized away, so providing
weak_odr definitions is only overhead)
This copies the text used in the #define statements to the code comments.
The conflicting text comes from AMD manuals, but those are wrong. Sadly,
that FP cmp text has not been updated even after some docs were updated
for Zen:
http://support.amd.com/en-us/search/tech-docs
( AMD64 Architecture Programmer's Manual Volume 4 )
See PR28110 for more discussion:
https://bugs.llvm.org/show_bug.cgi?id=28110
[Modules] Remove darwin specific code to check for SystemVersion.plist
This isn't need anymore and modules options -fbuild-session-file and
-fmodules-validate-once-per-build-session already provide a sane
mechanism to validate the system headers.
Serialization: Simulate -Werror settings in implicit modules
r293123 started serializing diagnostic pragma state for modules. This
makes the serialization work properly for implicit modules.
An implicit module build (using Clang's internal build system) uses the
same PCM file location for different `-Werror` levels.
E.g., if a TU has `-Werror=format` and tries to load a PCM built without
`-Werror=format`, a new PCM will be built in its place (and the new PCM
should have the same signature, since r297655). In the other direction,
if a TU does not have `-Werror=format` and tries to load a PCM built
with `-Werror=format`, it should "just work".
The idea is to evolve the PCM toward the strictest -Werror flags that
anyone tries.
r293123 started serializing the diagnostic pragma state for each PCM.
Since this encodes the -Werror settings at module-build time, it breaks
the implicit build model.
This commit filters the diagnostic state in order to simulate the
current compilation's diagnostic settings. Firstly, it ignores the
module's serialized first diagnostic state, replacing it with the state
from this compilation's command-line. Secondly, if a pragma warning was
upgraded to error/fatal when generating the PCM (e.g., due to `-Werror`
on the command-line), it checks whether it should still be upgraded in
its current context.
[IR] Add AttributeSet to hide AttributeSetNode* again, NFC
Summary:
For now, it just wraps AttributeSetNode*. Eventually, it will hold
AvailableAttrs as an inline bitset, and adding and removing enum
attributes will be super cheap.
This sinks AttributeSetNode back down to lib/IR/AttributeImpl.h.
[Sema][ObjC] Check whether a variable has a definition, rather than
checking its storage class, when determining whether casting a C pointer
to an ObjC pointer is allowed.
This change allows casting variables whose declarations are directly
contained in a linkage specification to an ObjC pointer type. Those
variables are treated as if they contain the extern specifier for the
purpose of determining whether they are definitions or not.
Richard Trieu [Tue, 11 Apr 2017 21:31:00 +0000 (21:31 +0000)]
[ODRHash] Improve handling of hash values
Calculating the hash in Sema::ActOnTagFinishDefinition could happen before
all sub-Decls were parsed or processed, which would produce the wrong hash
value. Change to calculating the hash on the first use and storing the value
instead. Also, avoid using the macros that were only for Boolean fields and
use an explicit checker during the DefintionData merge. No functional change,
but was this blocking other ODRHash patches.
David Blaikie [Tue, 11 Apr 2017 21:13:37 +0000 (21:13 +0000)]
Modular Codegen: Support homing debug info for types in modular objects
Matching the function-homing support for modular codegen. Any type
implicitly (implicit template specializations) or explicitly defined in
a module is attached to that module's object file and omitted elsewhere
(only a declaration used if necessary for references).
David Blaikie [Tue, 11 Apr 2017 20:46:34 +0000 (20:46 +0000)]
Modular Codegen: Add/use a bit in serialized function definitions to track whether they are the subject of modular codegen
Some decls are created not where they are written, but in other module
files/users (implicit special members and function template implicit
specializations). To correctly identify them, use a bit next to the definition
to track the modular codegen property.
Discussed whether the module file bit could be omitted in favor of
reconstituting from the modular codegen decls list - best guess today is that
the efficiency improvement of not having to deserialize the whole list whenever
any function is queried by a module user is worth it for the small size
increase of this redundant (list + bit-on-def) representation.
Sean Callanan [Tue, 11 Apr 2017 19:33:35 +0000 (19:33 +0000)]
[clang-import-test] Lookup inside contexts
clang-import-test has until now been only able to report top-level Decls.
This is clearly insufficient; we should be able to look inside structs
and namespaces also. This patch adds new test cases for a variety of
lookups inside existing ASTContexts, and adds the functionality necessar
to make most of these testcases work. (One testcase is known to fail
because of ASTImporter limitations when importing templates; I'll look
into that separately.)
This patch also separates the core functionality out into
ExternalASTMerger, an interface that allows clients like LLDB to make
use of it. clang-import-test now only has the machinery necessary to
set up the tests.
Yaxun Liu [Tue, 11 Apr 2017 17:24:23 +0000 (17:24 +0000)]
[OpenCL] Map default address space to alloca address space
For OpenCL, the private address space qualifier is 0 in AST. Before this change, 0 address space qualifier
is always mapped to target address space 0. As now target private address space is specified by
alloca address space in data layout, address space qualifier 0 needs to be mapped to alloca addr space specified by the data layout.
This change has no impact on targets whose alloca addr space is 0.
With contributions from Matt Arsenault, Tony Tye and Wen-Heng (Jack) Chung