Kristina Brooks [Mon, 12 Nov 2018 01:19:16 +0000 (01:19 +0000)]
[CodeGen][CXX]: Fix no_destroy CG bug under specific circumstances
Summary:
Class with no user-defined destructor that has an inherited member that has a
non-trivial destructor and a non-default constructor will attempt to emit a
destructor despite being marked as __attribute((no_destroy)) in which case it
would trigger an assertion due to an incorrect assumption.
In addition this adds missing test coverage for IR generation for no_destroy.
(Note that here use of no_destroy is synonymous with its global flag
counterpart `-fno-c++-static-destructors` being enabled)
Pass the function type instead of the return type to FunctionDecl::Create
Fix places where the return type of a FunctionDecl was being used in
place of the function type
FunctionDecl::Create() takes as its T parameter the type of function
that should be created, not the return type. Passing in the return type
looks to have been copypasta'd around a bit, but the number of correct
usages outweighs the incorrect ones so I've opted for keeping what T is
the same and fixing up the call sites instead.
This fixes a crash in Clang when attempting to compile the following
snippet of code with -fblocks -fsanitize=function -x objective-c++ (my
original repro case):
void g(void(^)());
void f()
{
__block int a = 0;
g(^(){ a++; });
}
as well as the following which only requires -fsanitize=function -x c++:
Michal Gorny [Sat, 10 Nov 2018 11:41:36 +0000 (11:41 +0000)]
[python] Support PathLike filenames and directories
Python 3.6 introduced a file system path protocol (PEP 519[1]).
The standard library APIs accepting file system paths now accept path
objects too. It could be useful to add this here as well
for convenience.
Alexey Bataev [Fri, 9 Nov 2018 20:03:19 +0000 (20:03 +0000)]
[OPENMP][NVPTX]Extend number of constructs executed in SPMD mode.
If the statements between target|teams|distribute directives does not
require execution in master thread, like constant expressions, null
statements, simple declarations, etc., such construct can be xecuted in
SPMD mode.
Simon Pilgrim [Fri, 9 Nov 2018 19:42:53 +0000 (19:42 +0000)]
Revert rL346454: Fix a use-after-free introduced by r344915.
r344915 added a call to ApplyDebugLocation to the sanitizer check
function emitter. Some of the sanitizers are emitted in the function
epilogue though and the LexicalScopeStack is emptied out before. By
detecting this situation and early-exiting from ApplyDebugLocation the
fallback location is used, which is equivalent to the return location.
Dylan McKay [Fri, 9 Nov 2018 19:42:05 +0000 (19:42 +0000)]
Use the correct address space when bitcasting func pointer to int pointer
When we cast a function pointer to an int pointer, at some pointer later
it gets bitcasted back to a function and called.
In backends that have a nonzero program memory address space specified
in the data layout, the old code would lose the address space data. When
LLVM later attempted to generate the bitcast from i8* to i8(..)*
addrspace(1), it would fail because the pointers are not in the same
address space.
With this patch, the address space of the function will carry on to the
address space of the i8* pointer. This is because all function pointers
in Harvard architectures need to be assigned to the correct address
space.
This has no effect to any in-tree backends except AVR.
Aaron Ballman [Fri, 9 Nov 2018 19:37:18 +0000 (19:37 +0000)]
Allow a double-underscore spelling of Clang attributes using double square bracket syntax.
This matches a similar behavior with GCC accepting [[gnu::__attr__]] as a alias for [[gnu::attr]] in that clang attributes can now be spelled with two leading and trailing underscores.
I had always intended for this to work, but missed the critical bit. We already had an existing test in test/Preprocessor/has_attribute.cpp for [[clang::__fallthrough__]] but using that spelling would still give an "unknown attribute" diagnostic.
Adrian Prantl [Fri, 9 Nov 2018 19:17:56 +0000 (19:17 +0000)]
Fix a nondeterminism in the debug info for VLA size expressions.
The artificial variable describing the array size is supposed to be
called "__vla_expr", but this was implemented by retrieving the name
of the associated alloca, which isn't a reliable source for the name,
since nonassert compilers may drop names from LLVM IR.
Yaxun Liu [Fri, 9 Nov 2018 18:52:05 +0000 (18:52 +0000)]
[HIP] Remove useless sections in linked files
clang-offload-bundler creates __CLANG_OFFLOAD_BUNDLE__* sections in the bundles,
which get into the linked files. These sections are useless after linking. They waste disk
space and cause confusion for clang when directly linked with other object files, therefore
should be removed.
Aaron Ballman [Fri, 9 Nov 2018 17:19:45 +0000 (17:19 +0000)]
Introduce the _Clang scoped attribute token.
Currently, we only accept clang as the scoped attribute identifier for double square bracket attributes provided by Clang, but this has the potential to conflict with user-defined macros. To help alleviate these concerns, this introduces the _Clang scoped attribute identifier as an alias for clang. It also introduces a warning with a fixit on the off chance someone attempts to use __clang__ as the scoped attribute (which is a predefined compiler identification macro).
Dylan McKay [Fri, 9 Nov 2018 17:15:06 +0000 (17:15 +0000)]
Use the correct address space when emitting the ctor function list
This patch modifies clang so that, if compiling for a target that
explicitly specifies a nonzero program memory address space, the
constructor list global will have the same address space as the
functions it contains.
AVR is the only in-tree backend which has a nonzero program memory
address space.
Without this, the IR verifier would always fail if a constructor
was used on a Harvard architecture backend.
This has no functional change to any in-tree backends except AVR.
Alexey Bataev [Fri, 9 Nov 2018 16:18:04 +0000 (16:18 +0000)]
[OPENMP][NVPTX]Allow to use shared memory for the
target|teams|distribute variables.
If the total size of the variables, declared in target|teams|distribute
regions, is less than the maximal size of shared memory available, the
buffer is allocated in the shared memory.
Zi Xuan Wu [Fri, 9 Nov 2018 03:35:32 +0000 (03:35 +0000)]
[PowerPC] [Clang] [AltiVec] The second parameter of vec_sr function should be modulo the number of bits in the element
The second parameter of vec_sr function is representing shift bits and it should be modulo the number of bits in the element like what vec_sl does now.
This is actually required by the ABI:
Each element of the result vector is the result of logically right shifting the corresponding
element of ARG1 by the number of bits specified by the value of the corresponding
element of ARG2, modulo the number of bits in the element. The bits that are shifted out
are replaced by zeros.
Bill Wendling [Fri, 9 Nov 2018 00:41:36 +0000 (00:41 +0000)]
Compound literals, enums, et al require const expr
Summary:
Compound literals, enums, file-scoped arrays, etc. require their
initializers and size specifiers to be constant. Wrap the initializer
expressions in a ConstantExpr so that we can easily check for this later
on.
Adrian Prantl [Fri, 9 Nov 2018 00:26:15 +0000 (00:26 +0000)]
Fix a use-after-free introduced by r344915.
r344915 added a call to ApplyDebugLocation to the sanitizer check
function emitter. Some of the sanitizers are emitted in the function
epilogue though and the LexicalScopeStack is emptied out before. By
detecting this situation and early-exiting from ApplyDebugLocation the
fallback location is used, which is equivalent to the return location.
David Blaikie [Thu, 8 Nov 2018 20:47:30 +0000 (20:47 +0000)]
[Frontend/Modules] Show diagnostics on prebuilt module configuration mismatch too
The current version only emits the below error for a module (attempted to be loaded) from the `prebuilt-module-path`:
```
error: module file blabla.pcm cannot be loaded due to a configuration mismatch with the current compilation [-Wmodule-file-config-mismatch]
```
With this change, if the prebuilt module is used, we allow the proper diagnostic behind the configuration mismatch to be shown.
```
error: POSIX thread support was disabled in PCH file but is currently enabled
error: module file blabla.pcm cannot be loaded due to a configuration mismatch with the current compilation [-Wmodule-file-config-mismatch]
```
(A few lines later an error is emitted anyways, so there is no reason not to complain for configuration mismatches if a config mismatch is found and kills the build.)
Alexey Bataev [Thu, 8 Nov 2018 15:47:39 +0000 (15:47 +0000)]
[OPENMP]Make lambda mapping follow reqs for PTR_AND_OBJ mapping.
The base pointer for the lambda mapping must point to the lambda capture
placement and pointer must point to the captured variable itself. Patch
fixes this problem.
Hans Wennborg [Thu, 8 Nov 2018 11:27:04 +0000 (11:27 +0000)]
clang-cl: Add "/clang:" pass-through arg support.
The clang-cl driver disables access to command line options outside of the
"Core" and "CLOption" sets of command line arguments. This filtering makes it
impossible to pass arguments that are interpreted by the clang driver and not
by either 'cc1' (the frontend) or one of the other tools invoked by the driver.
An example driver-level flag is the '-fno-slp-vectorize' flag, which is
processed by the driver in Clang::ConstructJob and used to set the cc1 flag
"-vectorize-slp". There is no negative cc1 flag or -mllvm flag, so it is not
currently possible to disable the SLP vectorizer from the clang-cl driver.
This change introduces the "/clang:" argument that is available when the
driver mode is set to CL compatibility. This option works similarly to the
"-Xclang" option, except that the option values are processed by the clang
driver rather than by 'cc1'. An example usage is:
clang-cl /clang:-fno-slp-vectorize /O2 test.c
Another example shows how "/clang:" can be used to pass a flag where there is
a conflict between a clang-cl compat option and an overlapping clang driver
option:
In the previous example, the unprefixed /MD selects the DLL version of the msvc
CRT, while the prefixed -MD flag and the -MF flags are used to create a make
dependency file for included headers.
One note about flag ordering: the /clang: flags are concatenated to the end of
the argument list, so in cases where the last flag wins, the /clang: flags
will be chosen regardless of their order relative to other flags on the driver
command line.
Shoaib Meenai [Thu, 8 Nov 2018 00:29:33 +0000 (00:29 +0000)]
[clang] Set CMP0075 to new
Make the check_include_file* macros honor CMAKE_REQUIRED_LIBRARIES. This
shouldn't cause any of the configuration checks to give different
results, and I verified that a clean configure before and after this
change resulted in identical CMake caches on my machine. The same change
was done for LLVM in r346377.
Alexey Bataev [Wed, 7 Nov 2018 19:11:14 +0000 (19:11 +0000)]
[OPENMP]Fix handling of the globals during compilation for the device.
Fixed lookup for the target regions in unused virtual functions + fixed
processing of the global variables not marked as declare target but
emitted during debug info emission.
Revert r346326 [OpenCL] Add support of cl_intel_device_side_avc_motion_estimation
This patch breaks Index/opencl-types.cl LIT test:
Script:
--
: 'RUN: at line 1'; stage1/bin/c-index-test -test-print-type llvm/tools/clang/test/Index/opencl-types.cl -cl-std=CL2.0 | stage1/bin/FileCheck llvm/tools/clang/test/Index/opencl-types.cl
--
Command Output (stderr):
--
llvm/tools/clang/test/Index/opencl-types.cl:3:26: warning: unsupported OpenCL extension 'cl_khr_fp16' - ignoring [-Wignored-pragmas]
llvm/tools/clang/test/Index/opencl-types.cl:4:26: warning: unsupported OpenCL extension 'cl_khr_fp64' - ignoring [-Wignored-pragmas]
llvm/tools/clang/test/Index/opencl-types.cl:8:9: error: use of type 'double' requires cl_khr_fp64 extension to be enabled
llvm/tools/clang/test/Index/opencl-types.cl:11:8: error: declaring variable of type 'half' is not allowed
llvm/tools/clang/test/Index/opencl-types.cl:15:3: error: use of type 'double' requires cl_khr_fp64 extension to be enabled
llvm/tools/clang/test/Index/opencl-types.cl:16:3: error: use of type 'double4' (vector of 4 'double' values) requires cl_khr_fp64 extension to be enabled
llvm/tools/clang/test/Index/opencl-types.cl:26:26: warning: unsupported OpenCL extension 'cl_khr_gl_msaa_sharing' - ignoring [-Wignored-pragmas]
llvm/tools/clang/test/Index/opencl-types.cl:35:44: error: use of type '__read_only image2d_msaa_t' requires cl_khr_gl_msaa_sharing extension to be enabled
llvm/tools/clang/test/Index/opencl-types.cl:36:49: error: use of type '__read_only image2d_array_msaa_t' requires cl_khr_gl_msaa_sharing extension to be enabled
llvm/tools/clang/test/Index/opencl-types.cl:37:49: error: use of type '__read_only image2d_msaa_depth_t' requires cl_khr_gl_msaa_sharing extension to be enabled
llvm/tools/clang/test/Index/opencl-types.cl:38:54: error: use of type '__read_only image2d_array_msaa_depth_t' requires cl_khr_gl_msaa_sharing extension to be enabled
Ilya Biryukov [Wed, 7 Nov 2018 10:02:31 +0000 (10:02 +0000)]
[CodeComplete] Do not complete self-initializations
Summary:
Removes references to initialized variable from the following completions:
int x = ^;
Handles only the trivial cases where the variable name is completed
immediately at the start of initializer or assignment, more complicated
cases aren't covered, e.g. these completions still contain 'x':
// More complicated expressions.
int x = foo(^);
int x = 10 + ^;
// Other kinds of initialization.
int x{^};
int x(^);
// Constructor initializers.
struct Foo {
Foo() : x(^) {}
int x;
};
We should address those in the future, but they are outside of the scope of
this initial change.
Akira Hatanaka [Tue, 6 Nov 2018 07:05:14 +0000 (07:05 +0000)]
os_log: Allow specifying mask type in format string.
A mask type is a 1 to 8-byte string that follows the "mask." annotation
in the format string. This enables obfuscating data in the event the
provided privacy level isn't enabled.
Akira Hatanaka [Tue, 6 Nov 2018 06:26:17 +0000 (06:26 +0000)]
os_log: Add a new privacy annotation "sensitive".
This is a stricter privacy annotation than "private", which will be used
for data that shouldn’t be logged to disk. For backward compatibility,
the "private" bit is set too.
The test hits stack overflow trying to instantiate recursive templates.
It is observed with ASAN and not with a regular build because ASAN
increases stack frame size.
Matt Davis [Mon, 5 Nov 2018 17:25:26 +0000 (17:25 +0000)]
[AST] Get aliased type info from an aliased TemplateSpecialization.
Summary:
Previously the TemplateSpecialization instance for 'template_alias', in the example below, returned the type info of the canonical type (int). This ignored the type alias if the template type happen to be aliased.
Before this patch, the assert would trigger with an alignment of 4:
```
typedef int __attribute__(( aligned( 16 ) )) aligned_int;
template < typename >
using template_alias = aligned_int;
static_assert( alignof( template_alias<void>) == 16, "" );
```
This patch checks if the TemplateSpecialization type has an alias, and if so will return the type information for the aliased type, else the canonical type's info is returned (original behavior). I believe that this is the desired behavior.
Ben Hamilton [Mon, 5 Nov 2018 16:59:33 +0000 (16:59 +0000)]
[Format] Add debugging to ObjC language guesser
Summary:
To handle diagnosing bugs where ObjCHeaderStyleGuesser guesses
wrong, this diff adds a bit more debug logging to the Objective-C
language guesser.
Benjamin Kramer [Mon, 5 Nov 2018 12:46:02 +0000 (12:46 +0000)]
Reapply "Fix regression in behavior of clang -x c++-header -fmodule-name=XXX"
This reverts commit r345963. We have a path forward now.
Original commit message:
The driver accidentally stopped passing the input filenames on to -cc1
in this mode due to confusion over what action was being requested.
This change also fixes a couple of crashes I encountered when passing
multiple files to such a -cc1 invocation.
Kristof Umann [Mon, 5 Nov 2018 03:50:37 +0000 (03:50 +0000)]
[analyzer] Restrict AnalyzerOptions' interface so that non-checker objects have to be registered
One of the reasons why AnalyzerOptions is so chaotic is that options can be
retrieved from the command line whenever and wherever. This allowed for some
options to be forgotten for a looooooong time. Have you ever heard of
"region-store-small-struct-limit"? In order to prevent this in the future, I'm
proposing to restrict AnalyzerOptions' interface so that only checker options
can be retrieved without special getters. I would like to make every option be
accessible only through a getter, but checkers from plugins are a thing, so I'll
have to figure something out for that.
This also forces developers who'd like to add a new option to register it
properly in the .def file.
This is done by
* making the third checker pointer parameter non-optional, and checked by an
assert to be non-null.
* I added new, but private non-checkers option initializers, meant only for
internal use,
* Renamed these methods accordingly (mind the consistent name for once with
getBooleanOption!):
- getOptionAsString -> getCheckerStringOption,
- getOptionAsInteger -> getCheckerIntegerOption
* The 3 functions meant for initializing data members (with the not very
descriptive getBooleanOption, getOptionAsString and getOptionAsUInt names)
were renamed to be overloads of the getAndInitOption function name.
* All options were in some way retrieved via getCheckerOption. I removed it, and
moved the logic to getStringOption and getCheckerStringOption. This did cause
some code duplication, but that's the only way I could do it, now that checker
and non-checker options are separated. Note that the non-checker version
inserts the new option to the ConfigTable with the default value, but the
checker version only attempts to find already existing entries. This is how
it always worked, but this is clunky and I might end reworking that too, so we
can eventually get a ConfigTable that contains the entire configuration of the
analyzer.
Kristof Umann [Mon, 5 Nov 2018 02:37:29 +0000 (02:37 +0000)]
Ensure the correct order of evaluation in part 2. of PlistMacroExpansion
Windows buildbots break with the previous commit '[analyzer][PlistMacroExpansion]
Part 2.: Retrieving the macro name and primitive expansion'. This patch attempts
to solve this issue.
Kristof Umann [Sun, 4 Nov 2018 13:59:44 +0000 (13:59 +0000)]
[analyzer][PlistMacroExpansion] Part 2.: Retrieving the macro name and primitive expansion
This patch adds a couple new functions to acquire the macro's name, and also
expands it, although it doesn't expand the arguments, as seen from the test files
Brian Gesiak [Sat, 3 Nov 2018 22:35:17 +0000 (22:35 +0000)]
[coroutines] Fix fallthrough warning on try/catch
Summary:
The test case added in this diff would incorrectly warn that control
flow may fall through without returning. Here's a standalone example:
https://godbolt.org/z/dCwXEi
The same program, but using `return` instead of `co_return`, does not
produce a warning: https://godbolt.org/z/mVldqQ
The issue was in how Clang analysis would structure its representation
of the control-flow graph. Specifically, when constructing the CFG,
`CFGBuilder::Visit` had special handling of a `ReturnStmt`, in which it
would place object destructors in the same CFG block as a `return` statement,
immediately after it. Doing so would allow the logic in
`lib/Sema/AnalysisBasedWarning.cpp` `CheckFallThrough` to work properly in the
program that used `return`, correctly determining that no "plain edges" preceded
the exit block of the function.
Because a `co_return` statement would not enjoy the same treatment when
it was being built into the control-flow graph, object destructors
would not be placed in the same CFG block as the `co_return`, thus
resulting in a "plain edge" preceding the exit block of the function,
and so the warning logic would be triggered.
Add special casing for `co_return` to Clang analysis, thereby
remedying the mistaken warning.
Douglas Yung [Sat, 3 Nov 2018 08:51:27 +0000 (08:51 +0000)]
Fix test on Windows.
This test checks the entire output of a help option, the problem
is that on Windows, the line break occurs in a different place
causing the CHECK to fail because it is not expecting a line break.
Takuto Ikuta [Sat, 3 Nov 2018 06:45:00 +0000 (06:45 +0000)]
Add /Zc:DllexportInlines option to clang-cl
Summary:
This CL adds /Zc:DllexportInlines flag to clang-cl.
When Zc:DllexportInlines- is specified, inline class member function is not exported if the function does not have local static variables.
By not exporting inline function, code for those functions are not generated and that reduces both compile time and obj size. Also this flag does not import inline functions from dllimported class if the function does not have local static variables.
On my 24C48T windows10 machine, build performance of chrome target in chromium repository is like below.
These stats are come with 'target_cpu="x86" enable_nacl = false is_component_build=true dcheck_always_on=true` build config and applied
* https://chromium-review.googlesource.com/c/chromium/src/+/1212379
* https://chromium-review.googlesource.com/c/v8/v8/+/1186017
| config | build time | speedup | build dir size |
| with patch, PCH on, debug | 1h10m0s | x1.13 | 35.6GB |
| without patch, PCH on, debug | 1h19m17s | | 49.0GB |
| with patch, PCH off, debug | 1h15m45s | x1.16 | 33.7GB |
| without patch, PCH off, debug | 1h28m10s | | 52.3GB |
| with patch, PCH on, release | 1h13m13s | x1.22 | 26.2GB |
| without patch, PCH on, release | 1h29m57s | | 37.5GB |
| with patch, PCH off, release | 1h23m38s | x1.32 | 23.7GB |
| without patch, PCH off, release | 1h50m50s | | 38.7GB |
This patch reduced obj size and the number of exported symbols largely, that improved link time too.
e.g. link time stats of blink_core.dll become like below
| | cold disk cache | warm disk cache |
| with patch, PCH on, debug | 71s | 30s |
| without patch, PCH on, debug | 111s | 48s |
This patch's implementation is based on Nico Weber's patch. I modified to support static local variable, added tests and took stats.
Bug: https://bugs.llvm.org/show_bug.cgi?id=33628
Reviewers: hans, thakis, rnk, javed.absar
Reviewed By: hans
Subscribers: kristof.beyls, smeenai, dschuff, probinson, cfe-commits, eraman
Richard Smith [Sat, 3 Nov 2018 02:23:33 +0000 (02:23 +0000)]
Revert r345562: "PR23833, DR2140: an lvalue-to-rvalue conversion on a glvalue of type"
This exposes a (known) CodeGen bug: it can't cope with emitting lvalue
expressions that denote non-odr-used but usable-in-constant-expression
variables. See PR39528 for a testcase.
Petr Hosek [Fri, 2 Nov 2018 23:49:36 +0000 (23:49 +0000)]
[CMake][Fuchsia] Set -fuse-ld=lld explicitly for Linux runtimes
We set lld as the default linker on non-Darwin platforms, but we still
need to set -fuse-ld=lld explicitly in to support cross-compiling Linux
runtimes on Darwin.
Kristof Umann [Fri, 2 Nov 2018 19:48:56 +0000 (19:48 +0000)]
Attempt to fix 'logical operation on address of string constant'
Caused a lot of warnings for Windows:
http://lab.llvm.org:8011/builders/llvm-clang-lld-x86_64-scei-ps4-windows10pro-fast/builds/21178/steps/build/logs/warnings%20%2867%29
Kristof Umann [Fri, 2 Nov 2018 16:02:10 +0000 (16:02 +0000)]
[analyzer] Put llvm.Conventions back in alpha
Interestingly, this many year old (when I last looked I remember 2010ish)
checker was committed without any tests, so I thought I'd implement them, but I
was shocked to see how I barely managed to get it working. The code is severely
outdated, I'm not even sure it has ever been used, so I'd propose to move it
back into alpha, and possibly even remove it.
Kristof Umann [Fri, 2 Nov 2018 15:50:44 +0000 (15:50 +0000)]
[analyzer][NFC] Collect all -analyzer-config options in a .def file
I'm in the process of refactoring AnalyzerOptions. The main motivation behind
here is to emit warnings if an invalid -analyzer-config option is given from the
command line, and be able to list them all.
In this patch, I'm moving all analyzer options to a def file, and move 2 enums
to global namespace.
Kristof Umann [Fri, 2 Nov 2018 15:48:10 +0000 (15:48 +0000)]
[analyzer][NFC] Fix some incorrect uses of -analyzer-config options
I'm in the process of refactoring AnalyzerOptions. The main motivation behind
here is to emit warnings if an invalid -analyzer-config option is given from
the command line, and be able to list them all.
In this patch, I found some flags that should've been used as checker options,
or have absolutely no mention of in AnalyzerOptions, or are nonexistent.
- NonLocalizedStringChecker now uses its "AggressiveReport" flag as a checker
option
- lib/StaticAnalyzer/Frontend/ModelInjector.cpp now accesses the "model-path"
option through a getter in AnalyzerOptions
- -analyzer-config path-diagnostics-alternate=false is not a thing, I removed it,
- lib/StaticAnalyzer/Checkers/AllocationDiagnostics.cpp and
lib/StaticAnalyzer/Checkers/AllocationDiagnostics.h are weird, they actually
only contain an option getter. I deleted them, and fixed RetainCountChecker
to get it's "leak-diagnostics-reference-allocation" option as a checker option,
- "region-store-small-struct-limit" has a proper getter now.
Alexey Bataev [Fri, 2 Nov 2018 14:54:07 +0000 (14:54 +0000)]
[OPENMP][NVPTX]Improve emission of the globalized variables for
target/teams/distribute regions.
Target/teams/distribute regions exist for all the time the kernel is
executed. Thus, if the variable is declared in their context and then
escape it, we can allocate global memory statically instead of
allocating it dynamically.
Patch captures all the globalized variables in target/teams/distribute
contexts, merges them into the records, one per each target region.
Those records are then joined into the union, one per compilation unit
(to save the global memory). Those units are organized into
2 x dimensional arrays, where the first dimension is
the number of blocks per SM and the second one is the number of SMs.
Runtime functions manage this global memory space between the executing
teams.
Tim Northover [Fri, 2 Nov 2018 13:14:11 +0000 (13:14 +0000)]
Reapply Logging: make os_log buffer size an integer constant expression.
The size of an os_log buffer is known at any stage of compilation, so making it
a constant expression means that the common idiom of declaring a buffer for it
won't result in a VLA. That allows the compiler to skip saving and restoring
the stack pointer around such buffers.
This also moves the OSLog and other FormatString helpers from
libclangAnalysis to libclangAST to avoid a circular dependency.
projects / clang / log