Shafik Yaghmour [Wed, 31 Oct 2018 21:53:15 +0000 (21:53 +0000)]
Revert "[ASTImporter][Structural Eq] Check for isBeingDefined"
This reverts commit r345760
because it caused an assertion in the lldb test suite. This is the log from the build bot: http://green.lab.llvm.org/green/view/LLDB/job/lldb-cmake/12003/
Richard Smith [Wed, 31 Oct 2018 20:39:26 +0000 (20:39 +0000)]
Part of PR39508: Emit an @llvm.invariant.start after storing to
__tls_guard.
__tls_guard can only ever transition from 0 to 1, and only once. This
permits LLVM to remove repeated checks for TLS initialization and
repeated initialization code in cases like:
int g();
thread_local int n = g();
int a = n + n;
where we could not prove that __tls_guard was still 'true' when checking
it for the second reference to 'n' in the initializer of 'a'.
[clang-format] tweaked another case of lambda formatting
Summary:
This is done in order to improve cases where the lambda's body is moved too far to the right. Consider the following snippet with column limit set to 79:
The tool favors extra indentation for the lambda body and so the code incurs extra wrapping and adjacent calls are indented to a different level. I find this behavior annoying and I'd like the tool to favor new lines and, thus, use the extra width.
The fix, reduced, brings the following formatting.
George Karpenkov [Wed, 31 Oct 2018 17:38:29 +0000 (17:38 +0000)]
[analyzer] RetainCountChecker: for now, do not trust the summaries of inlined code
Trusting summaries of inlined code would require a more thorough work,
as the current approach was causing too many false positives, as the new
example in test. The culprit lies in the fact that we currently escape
all variables written into a field (but not passed off to unknown
functions!), which can result in inconsistent behavior.
Kristof Umann [Wed, 31 Oct 2018 14:54:27 +0000 (14:54 +0000)]
[analyzer][PlistMacroExpansion] Part 1.: New expand-macros flag
This is the first part of the implementation of the inclusion of macro
expansions into the plist output. It adds a new flag that adds a new
"macro_expansions" entry to each report that has PathDiagnosticPieces that were
expanded from a macro. While there's an entry for each macro expansion, both
the name of the macro and what it expands to is missing, and will be implemented
in followup patches.
Hans Wennborg [Wed, 31 Oct 2018 10:34:46 +0000 (10:34 +0000)]
Follow-up to r345699: Call CheckStaticLocalForDllExport later for templates
Calling it too early might cause dllimport to get inherited onto the
VarDecl before the initializer got attached. See the test case for an
example where this broke things.
Hans Wennborg [Wed, 31 Oct 2018 08:38:48 +0000 (08:38 +0000)]
[clang-cl] Inherit dllexport to static locals also in template instantiations (PR39496)
In the course of D51340, @takuto.ikuta discovered that Clang fails to put
dllexport/import attributes on static locals during template instantiation.
For regular functions, this happens in Sema::FinalizeDeclaration(), however for
template instantiations we need to do something in or around
TemplateDeclInstantiator::VisitVarDecl(). This patch does that, and extracts
the code to a utility function.
Bill Wendling [Wed, 31 Oct 2018 03:48:47 +0000 (03:48 +0000)]
Create ConstantExpr class
A ConstantExpr class represents a full expression that's in a context where a
constant expression is required. This class reflects the path the evaluator
took to reach the expression rather than the syntactic context in which the
expression occurs.
In the future, the class will be expanded to cache the result of the evaluated
expression so that it's not needlessly re-evaluated
Reid Kleckner [Tue, 30 Oct 2018 23:58:41 +0000 (23:58 +0000)]
[Win64] Handle passing i128 by value
For arguments, pass it indirectly, since the ABI doc says pretty clearly
that arguments larger than 8 bytes are passed indirectly. This makes
va_list handling easier, anyway.
When returning, GCC returns in XMM0, and we match them.
This is the second half of Implicit Integer Conversion Sanitizer.
It completes the first half, and finally makes the sanitizer
fully functional! Only the bitfield handling is missing.
Summary:
C and C++ are interesting languages. They are statically typed, but weakly.
The implicit conversions are allowed. This is nice, allows to write code
while balancing between getting drowned in everything being convertible,
and nothing being convertible. As usual, this comes with a price:
```
void consume(unsigned int val);
void test(int val) {
consume(val);
// The 'val' is `signed int`, but `consume()` takes `unsigned int`.
// If val is negative, then consume() will be operating on a large
// unsigned value, and you may or may not have a bug.
// But yes, sometimes this is intentional.
// Making the conversion explicit silences the sanitizer.
consume((unsigned int)val);
}
```
Yes, there is a `-Wsign-conversion`` diagnostic group, but first, it is kinda
noisy, since it warns on everything (unlike sanitizers, warning on an
actual issues), and second, likely there are cases where it does **not** warn.
The actual detection is pretty easy. We just need to check each of the values
whether it is negative, and equality-compare the results of those comparisons.
The unsigned value is obviously non-negative. Zero is non-negative too.
https://godbolt.org/g/w93oj2
We do not have to emit the check *always*, there are obvious situations
where we can avoid emitting it, since it would **always** get optimized-out.
But i do think the tautological IR (`icmp ult %x, 0`, which is always false)
should be emitted, and the middle-end should cleanup it.
This sanitizer is in the `-fsanitize=implicit-conversion` group,
and is a logical continuation of D48958 `-fsanitize=implicit-integer-truncation`.
As for the ordering, i'we opted to emit the check **after**
`-fsanitize=implicit-integer-truncation`. At least on these simple 16 test cases,
this results in 1 of the 12 emitted checks being optimized away,
as compared to 0 checks being optimized away if the order is reversed.
This is a clang part.
The compiler-rt part is D50251.
Erik Pilkington [Tue, 30 Oct 2018 20:31:30 +0000 (20:31 +0000)]
NFC: Remove the ObjC1/ObjC2 distinction from clang (and related projects)
We haven't supported compiling ObjC1 for a long time (and never will again), so
there isn't any reason to keep these separate. This patch replaces
LangOpts::ObjC1 and LangOpts::ObjC2 with LangOpts::ObjC.
Aaron Ballman [Tue, 30 Oct 2018 19:49:17 +0000 (19:49 +0000)]
Speculatively attempt to fix a failing testbot.
A testbot ( http://green.lab.llvm.org/green/job/clang-stage1-cmake-RA-incremental/54690/) was failing with a complaint about an obsolete option that wasn't present in the command line in the first place. This replaces my guess at the "obsolete option" with a different spelling that will hopefully be more acceptable to this bot without breaking other bots.
Aaron Ballman [Tue, 30 Oct 2018 18:55:38 +0000 (18:55 +0000)]
Add the ability to output static analysis results to SARIF.
This allows users to specify SARIF (https://github.com/oasis-tcs/sarif-spec) as the output from the clang static analyzer so that the results can be read in by other tools, such as extensions to Visual Studio and VSCode, as well as static analyzers like CodeSonar.
Alexey Bataev [Tue, 30 Oct 2018 15:50:12 +0000 (15:50 +0000)]
[OPENMP] Support for mapping of the lambdas in target regions.
Added support for mapping of lambdas in the target regions. It scans all
the captures by reference in the lambda, implicitly maps those variables
in the target region and then later reinstate the addresses of
references in lambda to the correct addresses of the captured|privatized
variables.
Bruno Ricci [Tue, 30 Oct 2018 14:40:49 +0000 (14:40 +0000)]
[AST] Only store data for the NRVO candidate in ReturnStmt if needed
Only store the NRVO candidate if needed in ReturnStmt.
A good chuck of all of the ReturnStmt have no NRVO candidate
(more than half when parsing all of Boost). For all of them
this saves one pointer. This has no impact on children().
Bruno Ricci [Tue, 30 Oct 2018 13:42:41 +0000 (13:42 +0000)]
[AST] Only store the needed data in WhileStmt
Don't store the data for the condition variable if not needed.
This cuts the size of WhileStmt by up to a pointer.
The order of the children is kept the same.
[analyzer] Allow padding checker to traverse simple class hierarchies
The existing padding checker skips classes that have any base classes.
This patch allows the checker to traverse very simple cases:
classes that have no fields and have exactly one base class.
This is important mostly in the case of array declarations.
Aleksei Sidorin [Mon, 29 Oct 2018 21:46:18 +0000 (21:46 +0000)]
[ASTImporter] Reorder fields after structure import is finished
There are multiple reasons why field structures can be imported
in wrong order. The simplest is the ability of field initializers
and method bodies to refer fields not in order they are listed in.
Unfortunately, there is no clean solution for that currently
so I'm leaving a FIXME.
Petr Hosek [Mon, 29 Oct 2018 20:37:52 +0000 (20:37 +0000)]
[Driver] Support sanitized libraries on Fuchsia
When using sanitizers, add <resource_dir>/<target>/lib/<sanitizer>
to the list of library paths to support using sanitized version of
runtime libraries if available.
Alexey Bataev [Mon, 29 Oct 2018 20:17:42 +0000 (20:17 +0000)]
[OPENMP]Fix PR39372: Does not complain about loop bound variable not
being shared.
According to the standard, the variables with unspecified data-sharing
attributes in presence of `default(none)` clause must be reported to
users. Compiler did not generate error reports for the variables used in
other OpenMP regions. Patch fixes this.
Kristof Umann [Mon, 29 Oct 2018 20:06:30 +0000 (20:06 +0000)]
[analyzer][NFC] Refactor PlistDiagnostics to use a class instead of passing 9 parameters around
This has been a long time coming. Note the usage of AnalyzerOptions: I'll need
it for D52742, and added it in rC343620. The main motivation for this was that
I'll need to add yet another parameter to every single function, and some
functions would reach their 10th parameter with that change.
Bruno Ricci [Mon, 29 Oct 2018 16:12:37 +0000 (16:12 +0000)]
[AST] Only store the needed data in SwitchStmt
Don't store the data for the init statement and condition variable
if not needed. This cuts the size of SwitchStmt by up to 2 pointers.
The order of the children is intentionally kept the same.
Also use the newly available space in the bit-fields of Stmt
to store the bit representing whether all enums have been covered
instead of using a PointerIntPair.
[OpenMP][NVPTX] Use single loops when generating code for distribute parallel for
Summary: This patch adds a new code generation path for bound sharing directives containing distribute parallel for. The new code generation scheme applies to chunked schedules on distribute and parallel for directives. The scheme simplifies the code that is being generated by eliminating the need for an outer for loop over chunks for both distribute and parallel for directives. In the case of distribute it applies to any sized chunk while in the parallel for case it only applies when chunk size is 1.
[OpenCL] Fix serialization of OpenCLExtensionDecls
Summary:
I recently discovered that adding the following code into `opencl-c.h` causes
failure of `test/Headers/opencl-c-header.cl`:
```
#pragma OPENCL EXTENSION cl_my_ext : begin
void cl_my_ext_foobarbaz();
#pragma OPENCL EXTENSIOn cl_my_ext : end
```
Clang crashes at the assertion is `ASTReader::getGlobalSubmoduleID()`:
```
assert(I != M.SubmoduleRemap.end() && "Invalid index into submodule index remap");
```
The root cause of the problem that to deserialize `OPENCL_EXTENSION_DECLS`
section `ASTReader` needs to deserialize a Decl contained in it. In turn,
deserializing a Decl requires information about whether this declaration is
part of a (sub)module, but this information is not read yet because it is
located further in a module file.
Gabor Marton [Mon, 29 Oct 2018 10:18:28 +0000 (10:18 +0000)]
[ASTImporter] Import overrides before importing the rest of the chain
Summary:
During method import we check for structural eq of two methods.
In the structural eq check we check for their isVirtual() flag. That
flag, however, may depend on the number of overrides. Before this
change we imported the overrides *after* we had imported the rest of the
redecl chain. So, during the import of another decl from the chain
IsVirtual() gave false result.
Writing tests for this is not really possible, because there is no way
to remove an overridden method via the AST API.
(We should access the private ASTContext::OverriddenMethods container.)
Also, we should do the remove in the middle of the import process.
Erik Pilkington [Mon, 29 Oct 2018 02:29:21 +0000 (02:29 +0000)]
Support for groups of attributes in #pragma clang attribute
This commit enables pushing an empty #pragma clang attribute push, then adding
multiple attributes to it, then popping them all with #pragma clang attribute
pop, just like #pragma clang diagnostic. We still support the current way of
adding these, #pragma clang attribute push(__attribute__((...))), by treating it
like a combined push/attribute. This is needed to create macros like:
AST: extend MS decoration handling for extended vectors
We correctly handled extended vectors of non-floating point types.
However, we have the Intel style builtins which MSVC also supports which
do overlap in sizes with the floating point extended vectors. This
would result in overloading of floating point extended vector types
which matched sizes (e.g. <3 x float> would be backed by a <4 x float>
and thus match sizes) to be mangled similarly. Extended vectors are a
clang extension which live outside of the builtins, so mangle them all
similarly. This change just extends the current scheme to treat
floating point types similar to the way that we treat other types
currently.
This now allows the swift runtime to be built for Windows again.
Bruno Ricci [Sun, 28 Oct 2018 12:30:53 +0000 (12:30 +0000)]
[AST] Don't store data for GNU range case statement if not needed
Don't store the data for case statements of the form LHS ... RHS if not
needed. This cuts the size of CaseStmt by 1 pointer + 1 SourceLocation in
the common case.
Also use the newly available space in the bit-fields of Stmt to store the
keyword location of SwitchCase and move the small accessor
SwitchCase::getSubStmt to the header.
Bruno Ricci [Sat, 27 Oct 2018 21:12:20 +0000 (21:12 +0000)]
[AST] Only store the needed data in IfStmt
Only store the needed data in IfStmt. This cuts the size of IfStmt
by up to 3 pointers + 1 SourceLocation. The order of the children
is intentionally kept the same even though it would be more
convenient to put the optional trailing objects last. Additionally
use the newly available space in the bit-fields of Stmt to store
the location of the "if".
The result of this is that for the common case of an
if statement of the form:
if (some_cond)
some_statement
the size of IfStmt is brought down to 8 bytes + 2 pointers,
instead of 8 bytes + 5 pointers + 2 SourceLocation.
Bruno Ricci [Sat, 27 Oct 2018 19:21:19 +0000 (19:21 +0000)]
[AST] Refactor PredefinedExpr
Make the following changes to PredefinedExpr:
1. Move PredefinedExpr below StringLiteral so that it can use its definition.
2. Rename IdentType to IdentKind to be more in line with clang's conventions,
and propagate the change to its users.
3. Move the location and the IdentKind into the newly available space of
the bit-fields of Stmt.
4. Only store the function name when needed. When parsing all of Boost,
of the 1357 PredefinedExpr 919 have no function name.
Bruno Ricci [Sat, 27 Oct 2018 18:43:27 +0000 (18:43 +0000)]
[AST] Widen the bit-fields of Stmt to 8 bytes.
Although some classes are using the tail padding of Stmt, most of
them are not. In particular the expression classes are not using it
since there is Expr in between, and Expr contains a single pointer.
This patch widen the bit-fields to Stmt to 8 bytes and move some
data from NullStmt, CompoundStmt, LabelStmt, AttributedStmt, SwitchStmt,
WhileStmt, DoStmt, ForStmt, GotoStmt, ContinueStmt, BreakStmt
and ReturnStmt to the newly available space.
In itself this patch do not achieve much but I plan to go through each of
the classes in the statement/expression hierarchy and use this newly
available space. A quick estimation gives me that this should shrink the
size of the statement/expression hierarchy by >10% when parsing all of Boost.
Volodymyr Sapsai [Fri, 26 Oct 2018 22:16:24 +0000 (22:16 +0000)]
[VFS] Add property 'fallthrough' that controls fallback to real file system.
Default property value 'true' preserves current behavior. Value 'false' can be
used to create VFS "root", file system that gives better control over which
files compiler can use during compilation as there are no unpredictable
accesses to real file system.
Non-fallthrough use case changes how we treat multiple VFS overlay
files. Instead of all of them being at the same level just above a real
file system, now they are nested and subsequent overlays can refer to
files in previous overlays.
Change is done both in LLVM and Clang, corresponding LLVM commit is r345431.
Richard Smith [Fri, 26 Oct 2018 19:26:45 +0000 (19:26 +0000)]
PR26547: alignof should return ABI alignment, not preferred alignment
Summary:
- Add `UETT_PreferredAlignOf` to account for the difference between `__alignof` and `alignof`
- `AlignOfType` now returns ABI alignment instead of preferred alignment iff clang-abi-compat > 7, and one uses _Alignof or alignof
Bjorn Pettersson [Fri, 26 Oct 2018 16:12:12 +0000 (16:12 +0000)]
[Fixed Point Arithmetic] Refactor fixed point casts
Summary:
- Added names for some emitted values (such as "tobool" for
the result of a cast to boolean).
- Replaced explicit IRBuilder request for doing sext/zext/trunc
by using CreateIntCast instead.
- Simplify code for emitting satuation into one if-statement
for clamping to max, and one if-statement for clamping to min.
Martin Storsjo [Fri, 26 Oct 2018 08:33:29 +0000 (08:33 +0000)]
Revert "Reapply: [Driver] Use forward slashes in most linker arguments"
This reverts commit r345370, as it uncovered even more issues in
tests with partial/inconsistent path normalization:
http://lab.llvm.org:8011/builders/llvm-clang-x86_64-expensive-checks-win/builds/13562
http://lab.llvm.org:8011/builders/clang-x64-windows-msvc/builds/886
http://lab.llvm.org:8011/builders/llvm-clang-lld-x86_64-scei-ps4-windows10pro-fast/builds/20994
In particular, these tests seem to have failed:
Clang :: CodeGen/thinlto-diagnostic-handler-remarks-with-hotness.ll
Clang :: CodeGen/thinlto-multi-module.ll
Clang :: Driver/cuda-external-tools.cu
Clang :: Driver/cuda-options.cu
Clang :: Driver/hip-toolchain-no-rdc.hip
Clang :: Driver/hip-toolchain-rdc.hip
Clang :: Driver/openmp-offload-gpu.c
At least the Driver tests could potentially be fixed by extending
the path normalization to even more places, but the issues with the
CodeGen tests are still unknown.
In addition, a number of other tests seem to have been broken in
other clang dependent tools such as clang-tidy and clangd.
Martin Storsjo [Fri, 26 Oct 2018 07:01:59 +0000 (07:01 +0000)]
Reapply: [Driver] Use forward slashes in most linker arguments
libtool inspects the output of $CC -v to detect what object files and
libraries are linked in by default. When clang is built as a native
windows executable, all paths are formatted with backslashes, and
the backslashes cause each argument to be enclosed in quotes. The
backslashes and quotes break further processing within libtool (which
is implemented in shell script, running in e.g. msys) pretty badly.
Between unix style pathes (that only work in tools that are linked
to the msys runtime, essentially the same as cygwin) and proper windows
style paths (with backslashes, that can easily break shell scripts
and msys environments), the best compromise is to use windows style
paths (starting with e.g. c:) but with forward slashes, which both
msys based tools, shell scripts and native windows executables can
cope with. This incidentally turns out to be the form of paths that
GCC prints out when run with -v on windows as well.
This change potentially makes the output from clang -v a bit more
inconsistent, but it is isn't necessarily very consistent to begin with.
Compared to the previous attempt in SVN r345004, this now does
the same transformation on more paths, hopefully on the right set
of paths so that all tests pass (previously some tests failed, where
path fragments that were required to be identical turned out to
use different path separators in different places). This now also
is done only for non-windows, or cygwin/mingw targets, to preserve
all backslashes for MSVC cases (where the paths can end up e.g. embedded
into PDB files. (The transformation function itself,
llvm::sys::path::convert_to_slash only has an effect when run on windows.)
Bryan Chan [Thu, 25 Oct 2018 23:47:00 +0000 (23:47 +0000)]
[AArch64] Implement FP16FML intrinsics
Generate the FP16FML intrinsics into arm_neon.h (AArch64 only for now).
Add two new type modifiers to NeonEmitter to handle the new prototypes.
Define __ARM_FEATURE_FP16FML when +fp16fml is enabled and guard the
intrinsics with the macro in arm_neon.h.
George Karpenkov [Thu, 25 Oct 2018 23:38:58 +0000 (23:38 +0000)]
[analyzer] Fix a bug in "collapsed" graph viewer
Nodes which have only one predecessor and only one successor can not
always be hidden, even if all states are the same.
An additional condition is needed: the predecessor may have only one successor.
This can be seen on this example:
```
A
/ \
B C
\ /
D
```
Nodes B and C can not be hidden even if all nodes in the graph have the
same state.
George Karpenkov [Thu, 25 Oct 2018 23:38:07 +0000 (23:38 +0000)]
[analyzer] Correct modelling of OSDynamicCast: eagerly state split
Previously, OSDynamicCast was modeled as an identity.
This is not correct: the output of OSDynamicCast may be zero even if the
input was not zero (if the class is not of desired type), and thus the
modeling led to false positives.
Instead, we are doing eager state split:
in one branch, the returned value is identical to the input parameter,
and in the other branch, the returned value is zero.
This patch required a substantial refactoring of canEval infrastructure,
as now it can return different function summaries, and not just true/false.
Reid Kleckner [Thu, 25 Oct 2018 22:37:30 +0000 (22:37 +0000)]
Avoid std::map&vector in hexagon builtin code to save code size
Constructing a global std::map requires clang to generate a linear
amount of code to construct the initializer list if the elements are not
constexpr-constructible. std::vector is not constexpr-constructible, so
this code pattern was generating large amounts of code.
Also, because of PR38829, LLVM is pathologically slow on large basic
blocks, and this causes slow compilation. This works around the bug and
reduces code size.
SemaChecking.cpp -debug-info-kind=limited:
time objsize
before: 1m45.023s 9.8M
after: 0m25.205s 6.9M
Richard Smith [Thu, 25 Oct 2018 22:35:16 +0000 (22:35 +0000)]
Avoid STMT_ and DECL_ bitcodes overlapping.
This doesn't appear to matter for deserialization purposes, because we
always know what kind of entity (declaration or statement/expression)
we're trying to load, but it makes the llvm-bcanalyzer output a lot less
mysterious.
Nicolas Lesser [Thu, 25 Oct 2018 20:15:03 +0000 (20:15 +0000)]
[C++17] Reject shadowing of capture by parameter in lambda
Summary:
This change rejects the shadowing of a capture by a parameter in lambdas in C++17.
```
int main() {
int a;
auto f = [a](int a) { return a; };
}
```
results in:
```
main.cpp:3:20: error: a lambda parameter cannot shadow an explicitly captured entity
auto f = [a](int a) { return a; };
^
main.cpp:3:13: note: variable a is explicitly captured here
auto f = [a](int a) { return a; };
^
```
Erich Keane [Thu, 25 Oct 2018 18:57:19 +0000 (18:57 +0000)]
Implement Function Multiversioning for Non-ELF Systems.
Similar to how ICC handles CPU-Dispatch on Windows, this patch uses the
resolver function directly to forward the call to the proper function.
This is not nearly as efficient as IFuncs of course, but is still quite
useful for large functions specifically developed for certain
processors.
This is unfortunately still limited to x86, since it depends on
__builtin_cpu_supports and __builtin_cpu_is, which are x86 builtins.
The naming for the resolver/forwarding function for cpu-dispatch was
taken from ICC's implementation, which uses the unmodified name for this
(no mangling additions). This is possible, since cpu-dispatch uses '.A'
for the 'default' version.
In 'target' multiversioning, this function keeps the '.resolver'
extension in order to keep the default function keeping the default
mangling.
Eric Fiselier [Thu, 25 Oct 2018 18:16:16 +0000 (18:16 +0000)]
[SemaCXX] Unconfuse Clang when std::align_val_t is unscoped in C++03
Summary:
When -faligned-allocation is specified in C++03 libc++ defines std::align_val_t as an unscoped enumeration type (because Clang didn't provide scoped enumerations as an extension until 8.0).
Unfortunately Clang confuses the `align_val_t` overloads of delete with the sized deallocation overloads which aren't enabled. This caused Clang to call the aligned deallocation function as if it were the sized deallocation overload.
projects / clang / log