Clang has the default FP contraction setting of “-ffp-contract=on”, which
doesn't really mean “on” in the conventional sense of the word, but rather
really means “according to the per-statement effective value of the relevant
pragma”.
Before this patch, Clang has that pragma defaulting to “off”. Since the
“-ffp-contract=on” mode is really an AND of two booleans and the second of them
defaults to “off”, the whole thing effectively defaults to “off”. This patch
changes the default value of the pragma to “on”, thus making the default pair of
booleans (on, on) rather than (on, off). This makes FP optimization slightly
more aggressive than before when not using either “-Ofast”, “-ffast-math”, or
“-ffp-contract=fast”. Even with this patch the compiler still respects
“-ffp-contract=off”.
As per a suggestion by Steve Canon, the added code does _not_ require “-O3” or
higher. This is so as to try our best to preserve identical floating-point
results for unchanged source code compiling for an unchanged target when only
changing from any optimization level in the set (“-O0”, “-O1”, “-O2”, “-O3”) to
any other optimization level in that set. “-Os” and “-Oz” seem to be behaving
identically, i.e. should probably be considered a part of the aforementioned
set, but I have not reviewed this rigorously. “-Ofast” is explicitly _not_ a
member of that set.
Patch authored by Abe Skolnik [a.skolnik@samsung.com] and Stephen Canon [scanon@apple.com].
Differential Revision: https://reviews.llvm.org/D24481
llvm-svn: 282259
Currently, the Clang version is computed as follows:
1. LLVM defines major, minor, and patch versions, all statically set. Today,
these are 4, 0, and 0, respectively.
2. The static version numbers are combined into PACKAGE_VERSION along with a
suffix, so the result today looks like "4.0.0svn".
3. Clang extracts CLANG_VERSION from PACKAGE_VERSION using a regexp. The regexp
allows the patch level to omitted, and drops any non-digit trailing values.
Today, this result looks like "4.0.0".
4. CLANG_VERSION is then split further into CLANG_VERSION_MAJOR and
CLANG_VERSION_MINOR. Today, these resolve to 4 and 0, respectively.
5. If CLANG_VERSION matches a regexp with three version components, then
CLANG_VERSION_PATCHLEVEL is extracted and the CLANG_HAS_VERSION_PATCHLEVEL
variable is set to 1. Today, these values are 0 and 1, respectively.
6. The CLANG_VERSION_* variables (and CLANG_HAS_VERSION_PATCHLEVEL) are
configured into [llvm/tools/clang/]include/clang/Basic/Version.inc
verbatim by CMake.
7. In [llvm/tools/clang/]include/clang/Basic/Version.h, macros are defined
conditionally, based on CLANG_HAS_VERSION_PATCHLEVEL, to compute
CLANG_VERSION_STRING as either a two- or three-level version number. Today,
this value is "4.0.0", because despite the patchlevel being 0, it was
matched by regexp and is thus "HAS"ed by the preprocessor. This string is
then used wherever Clang's "version" is needed [*].
[*] Including, notably, by compiler-rt, for computing its installation path.
This change collapses steps 2-5 by defaulting Clang to use LLVM's (non-string)
version components for the Clang version (see [*] for why not PACKAGE_VERSION),
and collapses steps 6 and 7 by simply writing CLANG_VERSION_STRING into
Version.inc. The Clang version today always uses the patchlevel form, so the
collapsed Version.inc does not have logic for a version without a patch level.
Historically speaking, this technique began with the VER file in r82085 (which
survives in the form of the regexp in #3). The major, minor, and patchlevel
versions were introduced by r106863 (which remains in #4-6). The VER file itself
was deleted in favor of the LLVM version number in r106914. On the LLVM side,
the individual LLVM_VERSION_MAJOR, LLVM_VERSION_MINOR, and PACKAGE_VERSION
weren't introduced for nearly two more years, until r150405.
llvm-svn: 281666
The class PTHWriter is in lib/Frontend/CacheTokens.cpp
inside the anonymous namespace.
This diff changes the order of fields an removes excessive padding.
Test plan: make -j8 check-clang
Differential revision: https://reviews.llvm.org/D23902
llvm-svn: 281385
Original commit message:
Add -fdiagnostics-show-hotness
Summary:
I've recently added the ability for optimization remarks to include the
hotness of the corresponding code region. This uses PGO and allows
filtering of the optimization remarks by relevance. The idea was first
discussed here:
http://thread.gmane.org/gmane.comp.compilers.llvm.devel/98334
The general goal is to produce a YAML file with the remarks. Then, an
external tool could dynamically filter these by hotness and perhaps by
other things.
That said it makes sense to also expose this at the more basic level
where we just include the hotness info with each optimization remark.
For example, in D22694, the clang flag was pretty useful to measure the
overhead of the additional analyses required to include hotness.
(Without the flag we don't even run the analyses.)
For the record, Hal has already expressed support for the idea of this
patch on IRC.
Differential Revision: https://reviews.llvm.org/D23284
llvm-svn: 281293
Summary:
I've recently added the ability for optimization remarks to include the
hotness of the corresponding code region. This uses PGO and allows
filtering of the optimization remarks by relevance. The idea was first
discussed here:
http://thread.gmane.org/gmane.comp.compilers.llvm.devel/98334
The general goal is to produce a YAML file with the remarks. Then, an
external tool could dynamically filter these by hotness and perhaps by
other things.
That said it makes sense to also expose this at the more basic level
where we just include the hotness info with each optimization remark.
For example, in D22694, the clang flag was pretty useful to measure the
overhead of the additional analyses required to include hotness.
(Without the flag we don't even run the analyses.)
For the record, Hal has already expressed support for the idea of this
patch on IRC.
Differential Revision: https://reviews.llvm.org/D23284
llvm-svn: 281276
OpenCL requires __ENDIAN_LITTLE__ be set for little endian targets.
The default for targets was also apparently big endian, so AMDGPU
was incorrectly reported as big endian. Set this from the triple
so targets don't have another place to set the endianness.
llvm-svn: 280787
-fprofile-dir=path allows the user to specify where .gcda files should be
emitted when the program is run. In particular, this is the first flag that
causes the .gcno and .o files to have different paths, LLVM is extended to
support this. -fprofile-dir= does not change the file name in the .gcno (and
thus where lcov looks for the source) but it does change the name in the .gcda
(and thus where the runtime library writes the .gcda file). It's different from
a GCOV_PREFIX because a user can observe that the GCOV_PREFIX_STRIP will strip
paths off of -fprofile-dir= but not off of a supplied GCOV_PREFIX.
To implement this we split -coverage-file into -coverage-data-file and
-coverage-notes-file to specify the two different names. The !llvm.gcov
metadata node grows from a 2-element form {string coverage-file, node dbg.cu}
to 3-elements, {string coverage-notes-file, string coverage-data-file, node
dbg.cu}. In the 3-element form, the file name is already "mangled" with
.gcno/.gcda suffixes, while the 2-element form left that to the middle end
pass.
llvm-svn: 280306
r280133. Original commit message:
C++ Modules TS: driver support for building modules.
This works as follows: we add --precompile to the existing gamut of options for
specifying how far to go when compiling an input (-E, -c, -S, etc.). This flag
specifies that an input is taken to the precompilation step and no further, and
this can be specified when building a .pcm from a module interface or when
building a .pch from a header file.
The .cppm extension (and some related extensions) are implicitly recognized as
C++ module interface files. If --precompile is /not/ specified, the file is
compiled (via a .pcm) to a .o file containing the code for the module (and then
potentially also assembled and linked, if -S, -c, etc. are not specified). We
do not yet suppress the emission of object code for other users of the module
interface, so for now this will only work if everything in the .cppm file has
vague linkage.
As with the existing support for module-map modules, prebuilt modules can be
provided as compiler inputs either via the -fmodule-file= command-line argument
or via files named ModuleName.pcm in one of the directories specified via
-fprebuilt-module-path=.
This also exposes the -fmodules-ts cc1 flag in the driver. This is still
experimental, and in particular, the concrete syntax is subject to change as
the Modules TS evolves in the C++ committee. Unlike -fmodules, this flag does
not enable support for implicitly loading module maps nor building modules via
the module cache, but those features can be turned on separately and used in
conjunction with the Modules TS support.
llvm-svn: 280134
to CC1, which are translated to function attributes and can e.g. be mapped on
build attributes FP_exceptions and FP_denormal. Setting these build attributes
allows better selection of floating point libraries.
Differential Revision: https://reviews.llvm.org/D23840
llvm-svn: 280064
This works as follows: we add --precompile to the existing gamut of options for
specifying how far to go when compiling an input (-E, -c, -S, etc.). This flag
specifies that an input is taken to the precompilation step and no further, and
this can be specified when building a .pcm from a module interface or when
building a .pch from a header file.
The .cppm extension (and some related extensions) are implicitly recognized as
C++ module interface files. If --precompile is /not/ specified, the file is
compiled (via a .pcm) to a .o file containing the code for the module (and then
potentially also assembled and linked, if -S, -c, etc. are not specified). We
do not yet suppress the emission of object code for other users of the module
interface, so for now this will only work if everything in the .cppm file has
vague linkage.
As with the existing support for module-map modules, prebuilt modules can be
provided as compiler inputs either via the -fmodule-file= command-line argument
or via files named ModuleName.pcm in one of the directories specified via
-fprebuilt-module-path=.
This also exposes the -fmodules-ts cc1 flag in the driver. This is still
experimental, and in particular, the concrete syntax is subject to change as
the Modules TS evolves in the C++ committee. Unlike -fmodules, this flag does
not enable support for implicitly loading module maps nor building modules via
the module cache, but those features can be turned on separately and used in
conjunction with the Modules TS support.
llvm-svn: 280035
In cases where .dwo/.dwp files are guaranteed to be available, skipping
the extra online (in the .o file) inline info can save a substantial
amount of space - see the original r221306 for more details there.
llvm-svn: 279651
In this mode, there is no need to load any module map and the programmer can
simply use "@import" syntax to load the module directly from a prebuilt
module path. When loading from prebuilt module path, we don't support
rebuilding of the module files and we ignore compatible configuration
mismatches.
rdar://27290316
Differential Revision: http://reviews.llvm.org/D23125
llvm-svn: 279096
standard's Annex B. We now attempt to increase the process's stack rlimit to
8MiB on startup, which appears to be enough to allow this to work reliably.
(And if it turns out not to be, we can investigate increasing it further.)
llvm-svn: 278983
Add 'ignore-non-existent-contents' to tell the VFS whether an invalid path
obtained via 'external-contents' should cause iteration on the VFS to stop.
If 'true', the VFS should ignore the entry and continue with the next. Allows
YAML files to be shared across multiple compiler invocations regardless of
prior existent paths in 'external-contents'. This global value is overridable
on a per-file basis.
This adds the parsing and write test part, but use by VFS comes next.
Differential Revision: https://reviews.llvm.org/D23422
rdar://problem/27531549
llvm-svn: 278456
Let the driver pass the option to frontend. Do not set precision metadata for division instructions when this option is set. Set function attribute "correctly-rounded-divide-sqrt-fp-math" based on this option.
Differential Revision: https://reviews.llvm.org/D22940
llvm-svn: 278155
Adjust target features for amdgcn target when -cl-denorms-are-zero is set.
Denormal support is controlled by feature strings fp32-denormals fp64-denormals in amdgcn target. If -cl-denorms-are-zero is not set and the command line does not set fp32/64-denormals feature string, +fp32-denormals +fp64-denormals will be on for GPU's supporting them.
A new virtual function virtual void TargetInfo::adjustTargetOptions(const CodeGenOptions &CGOpts, TargetOptions &TargetOpts) const is introduced to allow adjusting target option by codegen option.
Differential Revision: https://reviews.llvm.org/D22815
llvm-svn: 278151
This patch adds a command line option to list the checkers that were enabled
by analyzer-checker and not disabled by -analyzer-disable-checker.
It can be very useful to debug long command lines when it is not immediately
apparent which checkers are turned on and which checkers are turned off.
Differential Revision: https://reviews.llvm.org/D23060
llvm-svn: 278006
This differs from the previous version by being more careful about template
instantiation/specialization in order to prevent errors when building with
clang -Werror. Specifically:
* begin is not defined in the template and is instead instantiated when Head
is. I think the warning when we don't do that is wrong (PR28815) but for now
at least do it this way to avoid the warning.
* Instead of performing template specializations in LLVM_INSTANTIATE_REGISTRY
instead provide a template definition then do explicit instantiation. No
compiler I've tried has problems with doing it the other way, but strictly
speaking it's not permitted by the C++ standard so better safe than sorry.
Original commit message:
Currently the Registry class contains the vestiges of a previous attempt to
allow plugins to be used on Windows without using BUILD_SHARED_LIBS, where a
plugin would have its own copy of a registry and export it to be imported by
the tool that's loading the plugin. This only works if the plugin is entirely
self-contained with the only interface between the plugin and tool being the
registry, and in particular this conflicts with how IR pass plugins work.
This patch changes things so that instead the add_node function of the registry
is exported by the tool and then imported by the plugin, which solves this
problem and also means that instead of every plugin having to export every
registry they use instead LLVM only has to export the add_node functions. This
allows plugins that use a registry to work on Windows if
LLVM_EXPORT_SYMBOLS_FOR_PLUGINS is used.
llvm-svn: 277806
Currently Clang use int32 to represent sampler_t, which have been a source of issue for some backends, because in some backends sampler_t cannot be represented by int32. They have to depend on kernel argument metadata and use IPA to find the sampler arguments and global variables and transform them to target specific sampler type.
This patch uses opaque pointer type opencl.sampler_t* for sampler_t. For each use of file-scope sampler variable, it generates a function call of __translate_sampler_initializer. For each initialization of function-scope sampler variable, it generates a function call of __translate_sampler_initializer.
Each builtin library can implement its own __translate_sampler_initializer(). Since the real sampler type tends to be architecture dependent, allowing it to be initialized by a library function simplifies backend design. A typical implementation of __translate_sampler_initializer could be a table lookup of real sampler literal values. Since its argument is always a literal, the returned pointer is known at compile time and easily optimized to finally become some literal values directly put into image read instructions.
This patch is partially based on Alexey Sotkin's work in Khronos Clang (3d4eec6162).
Differential Revision: https://reviews.llvm.org/D21567
llvm-svn: 277024
This version has two fixes compared to the original:
* In Registry.h the template static members are instantiated before they are
used, as clang gives an error if you do it the other way around.
* The use of the Registry template in clang-tidy is updated in the same way as
has been done everywhere else.
Original commit message:
Currently the Registry class contains the vestiges of a previous attempt to
allow plugins to be used on Windows without using BUILD_SHARED_LIBS, where a
plugin would have its own copy of a registry and export it to be imported by
the tool that's loading the plugin. This only works if the plugin is entirely
self-contained with the only interface between the plugin and tool being the
registry, and in particular this conflicts with how IR pass plugins work.
This patch changes things so that instead the add_node function of the registry
is exported by the tool and then imported by the plugin, which solves this
problem and also means that instead of every plugin having to export every
registry they use instead LLVM only has to export the add_node functions. This
allows plugins that use a registry to work on Windows if
LLVM_EXPORT_SYMBOLS_FOR_PLUGINS is used.
llvm-svn: 276973
Currently the Registry class contains the vestiges of a previous attempt to
allow plugins to be used on Windows without using BUILD_SHARED_LIBS, where a
plugin would have its own copy of a registry and export it to be imported by
the tool that's loading the plugin. This only works if the plugin is entirely
self-contained with the only interface between the plugin and tool being the
registry, and in particular this conflicts with how IR pass plugins work.
This patch changes things so that instead the add_node function of the registry
is exported by the tool and then imported by the plugin, which solves this
problem and also means that instead of every plugin having to export every
registry they use instead LLVM only has to export the add_node functions. This
allows plugins that use a registry to work on Windows if
LLVM_EXPORT_SYMBOLS_FOR_PLUGINS is used.
Differential Revision: http://reviews.llvm.org/D21385
llvm-svn: 276856
With PCH+Module, sometimes compiler gives a hard error:
Module file ‘<some-file path>.pcm' is out of date and needs to be rebuilt
This happens when we have a pch importing a module and the module gets
overwritten by another compiler instance after we build the pch (one example is
that both compiler instances hash to the same pcm file but use different
diagnostic options). When we try to load the pch later on, the compiler notices
that the imported module is out of date (modification date, size do not match)
but it can't handle this out of date pcm (i.e it does not know how to rebuild
the pch).
This commit introduces a new command line option so for PCH + module, we can
turn on this option and if two compiler instances only differ in diagnostic
options, the latter instance will not invalidate the original pcm.
rdar://26675801
Differential Revision: http://reviews.llvm.org/D22773
llvm-svn: 276769
Summary:
This patch replaces the CUDA specific action by a generic offload action. The offload action may have multiple dependences classier in “host” and “device”. The way this generic offloading action is used is very similar to what is done today by the CUDA implementation: it is used to set a specific toolchain and architecture to its dependences during the generation of jobs.
This patch also proposes propagating the offloading information through the action graph so that that information can be easily retrieved at any time during the generation of commands. This allows e.g. the "clang tool” to evaluate whether CUDA should be supported for the device or host and ptas to easily retrieve the target architecture.
This is an example of how the action graphs would look like (compilation of a single CUDA file with two GPU architectures)
```
0: input, "cudatests.cu", cuda, (host-cuda)
1: preprocessor, {0}, cuda-cpp-output, (host-cuda)
2: compiler, {1}, ir, (host-cuda)
3: input, "cudatests.cu", cuda, (device-cuda, sm_35)
4: preprocessor, {3}, cuda-cpp-output, (device-cuda, sm_35)
5: compiler, {4}, ir, (device-cuda, sm_35)
6: backend, {5}, assembler, (device-cuda, sm_35)
7: assembler, {6}, object, (device-cuda, sm_35)
8: offload, "device-cuda (nvptx64-nvidia-cuda:sm_35)" {7}, object
9: offload, "device-cuda (nvptx64-nvidia-cuda:sm_35)" {6}, assembler
10: input, "cudatests.cu", cuda, (device-cuda, sm_37)
11: preprocessor, {10}, cuda-cpp-output, (device-cuda, sm_37)
12: compiler, {11}, ir, (device-cuda, sm_37)
13: backend, {12}, assembler, (device-cuda, sm_37)
14: assembler, {13}, object, (device-cuda, sm_37)
15: offload, "device-cuda (nvptx64-nvidia-cuda:sm_37)" {14}, object
16: offload, "device-cuda (nvptx64-nvidia-cuda:sm_37)" {13}, assembler
17: linker, {8, 9, 15, 16}, cuda-fatbin, (device-cuda)
18: offload, "host-cuda (powerpc64le-unknown-linux-gnu)" {2}, "device-cuda (nvptx64-nvidia-cuda)" {17}, ir
19: backend, {18}, assembler
20: assembler, {19}, object
21: input, "cuda", object
22: input, "cudart", object
23: linker, {20, 21, 22}, image
```
The changes in this patch pass the existent regression tests (keeps the existent functionality) and resulting binaries execute correctly in a Power8+K40 machine.
Reviewers: echristo, hfinkel, jlebar, ABataev, tra
Subscribers: guansong, andreybokhanko, tcramer, mkuron, cfe-commits, arpith-jacob, carlo.bertolli, caomhin
Differential Revision: https://reviews.llvm.org/D18171
llvm-svn: 275645
This changes the CompilerInstance::createOutputFile function to return
a std::unique_ptr<llvm::raw_ostream>, rather than an llvm::raw_ostream
implicitly owned by the CompilerInstance. This in most cases required that
I move ownership of the output stream to the relevant ASTConsumer.
The motivation for this change is to allow BackendConsumer to be a client
of interfaces such as D20268 which take ownership of the output stream.
Differential Revision: http://reviews.llvm.org/D21537
llvm-svn: 275507
-fxray-instrument: enables XRay annotation of IR
-fxray-instruction-threshold: configures the threshold for function size (looking at IR instructions), and allow LLVM to decide whether to add the nop sleds later on in the process.
Also implements the related xray_always_instrument and xray_never_instrument function attributes.
Patch by Dean Michael Berris.
llvm-svn: 275330
Also fixes strict-aliasing option to only be allowed when OpenCL Version 1.0. Added testcase in test/Frontend/opencl-blocks.cl.
Patch by Aaron En Ye Shi.
Differential Revision: http://reviews.llvm.org/D22170
llvm-svn: 275318
This is to allow distributed build systems, that do not preserve time stamps, to use PCH files.
Second and last part of the patch proposed at:
Differential Revision: http://reviews.llvm.org/D20867
llvm-svn: 275267
Add OCL option -cl-no-signed-zeros to driver options.
Also added to opencl.cl testcases.
Patch by Aaron En Ye Shi.
Differential Revision: http://reviews.llvm.org/D22067
llvm-svn: 274923
Summary: This patch changes the options used by offloading to start with -fopenmp instead of -fomp. This makes the option naming more consistent and materializes a suggestion by Richard Smith in http://reviews.llvm.org/D9888.
Reviewers: hfinkel, carlo.bertolli, arpith-jacob, ABataev
Subscribers: kkwli0, cfe-commits, caomhin
Differential Revision: http://reviews.llvm.org/D21841
llvm-svn: 274283
Allow -cl-std and other standard -cl- options from cc1 to driver.
Added a test for the options moved.
Patch by Aaron En Ye Shi.
Differential Revision: http://reviews.llvm.org/D21031
llvm-svn: 274150
DiagnosticNoteRenderer asserts trying to emit its "while building
module Foo imported from bar.h:5" note when the presumed location
of the import is invalid. This assertion was added in r267914,
where most uses of 'getFilename' were updated to test 'isValid'
instead. This one must have been missed.
I can't come up with a test because this location is always valid
in C-based code, but external clients that manually import modules
(*cough*Swift*cough*) sometimes provide invalid SourceLocations.
rdar://problem/26099576
http://reviews.llvm.org/D21111
llvm-svn: 273976
The PIC and PIE levels are not independent. In fact, if PIE is defined
it is always the same as PIC.
This is clear in the driver where ParsePICArgs returns a PIC level and
a IsPIE boolean. Unfortunately that is currently lost and we pass two
redundant levels down the pipeline.
This patch keeps a bool and a PIC level all the way down to codegen.
llvm-svn: 273566
Add support for /Ob1 (and equivalent -finline-hint-functions), which enable
inlining only for functions marked inline, either explicitly (via inline
keyword, for example), or implicitly (function definition in class body,
for example).
This works by enabling inlining pass, and adding noinline attribute to
every function not marked inline.
Patch by Rudy Pons <rudy.pons@ilod.org>!
Differential Revision: http://reviews.llvm.org/D20647
llvm-svn: 273440
Include opencl-c.h by default as a module to utilize the automatic AST caching mechanism of clang modules.
Add an option -finclude-default-header to enable default header for OpenCL, which is off by default.
Differential Revision: http://reviews.llvm.org/D20444
llvm-svn: 273191
Fix a regression which forbids using -std=cl|CL1.1|CL1.2|CL2.0 in driver.
Allow -std and -cl-std={cl|CL}{|1.1|1.2|2.0}.
Differential Revision: http://reviews.llvm.org/D20630
llvm-svn: 273015
Summary:
Create a new Frontend LangOpt to specify the renderscript language. It
is enabled by the "-x renderscript" option from the driver.
Add a "kernel" function attribute only for RenderScript (an "ignored
attribute" warning is generated otherwise).
Make the NativeHalfType and NativeHalfArgsAndReturns LangOpts be implied
by the RenderScript LangOpt.
Reviewers: rsmith
Subscribers: cfe-commits, srhines
Differential Revision: http://reviews.llvm.org/D21198
llvm-svn: 272342
Diagnostics that happen during driver time do not have color output support
unless -fcolor-diagonostic is explicitly passed into the driver. This is not a
problem for cc1 since dianostic arguments are properly handled and color is
enabled by default if the terminal supports it.
Make the driver behave like CC1. There are tests that already check for these
flags, but for the color itself there's no sensible way to test it.
Differential Revision: http://reviews.llvm.org/D20404
rdar://problem/26290980
llvm-svn: 271042
OpenMP version.
If '-fopenmp' option is provided '-fopenmp-version=' allows to control,
which version of OpenMP must be supported. Currently it affects only the
value of _OPENMP define.
llvm-svn: 270838
OpenMP 4.5.
According to OpenMP 4.5 the _OPENMP macro name is defined to have the decimal value yyyymm where yyyy and mm are the year and month designations of the version of the OpenMP API that the implementation supports. Clang supports OpenMP 4.5 so updated value of _OPENMP macro to 201511.
llvm-svn: 270822
-finline-functions and /Ob2 are currently ignored by Clang. The only way to
enable inlining is to use the global O flags, which also enable other options,
or to emit LLVM bitcode using Clang, then running opt by hand with the inline
pass.
This patch allows to simply use the -finline-functions flag (same as GCC) or
/Ob2 in clang-cl mode to enable inlining without other optimizations.
This is the first patch of a serie to improve support for the /Ob flags.
Patch by Rudy Pons <rudy.pons@ilod.org>!
Differential Revision: http://reviews.llvm.org/D20576
llvm-svn: 270609
This matches default nvcc behavior and gives substantial
performance boost on GPU where fmad is much cheaper compared to add+mul.
Differential Revision: http://reviews.llvm.org/D20341
llvm-svn: 270094
When remapped files were changed, they would not always cause the preamble's PCH to be invalidated, because the remapped path didn't necessarily match the include path (e.g. slash direction -- this happens a lot on Windows). I fixed this by moving to a llvm::sys::fs::UniqueID-based map instead of comparing paths stringwise.
Differential Revision: http://reviews.llvm.org/D20137
llvm-svn: 269769
licensees actually see in the toolchain we deliver to them. This will
reduce the set of local patches we have to maintain. The triple is
not changing. (The term ORBIS is an internal code name for PS4.)
llvm-svn: 269671
Add supported OpenCL extensions to target info. It serves as default values to save the users of the burden setting each supported extensions and optional core features in command line.
Re-commit after fixing build error due to missing override attribute.
Differential Revision: http://reviews.llvm.org/D19484
llvm-svn: 269670
Clang performs directory walk while searching headers inside modules by
using the ::sys::fs instead of ::vfs. This prevents any code that uses
the VFS (e.g, reproducer scripts) to actually find such headers, since
the VFS will never be searched for those.
Change these places to use vfs::recursive_directory_iterator and
vfs::directory_iterator instead.
Differential Revision: http://reviews.llvm.org/D20266
rdar://problem/25880368
llvm-svn: 269661
This reverts commit r269492 as the corresponding LLVM commit was
reverted due to lots of warnings. See the review thread for the original
LLVM commit (r269491) for details.
llvm-svn: 269549
(1) Collect headers under inner frameworks (frameworks inside other
other frameworks).
(2) Make sure we also collect the right header files inside them.
More info on (2):
Consider a dummy framework module B, with header Frameworks/B/B.h. Now
consider that another framework A, with header Frameworks/A/A.h, has a
layout with a inner framework Frameworks/A/Frameworks/B/B.h, where the
"B/B.h" part is a symlink for Frameworks/B/B.h. Also assume that
Frameworks/A/A.h includes <B/B.h>.
When parsing header Frameworks/A/A.h, framework module lookup is
performed in search for B, and it happens that
"Frameworks/A/Frameworks/B/B.h" path is registered in the module instead
of real "Frameworks/B/B.h". This occurs because
"Frameworks/A/Frameworks/B/B.h" is scanned first by the FileManager,
when looking for inner framework modules under Frameworks/A/Frameworks.
This makes Frameworks/A/Frameworks/B/B.h the default cached named inside
the FileManager for the B.h file UID.
This leads to modules being built without consistent paths to underlying
header files. This is usually not a problem in regular compilation flow,
but it's an issue when running the crash reproducer. The issue is that
clangs collect "Frameworks/A/Frameworks/B/B.h" but not
"Frameworks/B/B.h" into the VFS, leading to err_mmap_umbrella_clash. So
make sure we also collect the original header.
Differential Revision: http://reviews.llvm.org/D20194
rdar://problem/25880368
llvm-svn: 269502
Revert r269431 due to build failure caused by warning msg:
llvm/tools/clang/lib/Basic/Targets.cpp:2090:9: error: 'setSupportedOpenCLOpts' overrides a member function but is not marked 'override' [-Werror,-Winconsistent-missing-override]
void setSupportedOpenCLOpts() {
llvm-svn: 269435
Add supported OpenCL extensions to target info. It serves as default values to save the users of the burden setting each supported extensions and optional core features in command line.
Differential Revision: http://reviews.llvm.org/D19484
llvm-svn: 269431
Summary:
There virtual destructor can't be found and cause a compilation error
on a shared build.
To repro: [Release + Shared]
```
-DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=ON
```
Which produce this error:
```
tools/clang/unittests/Tooling/CMakeFiles/ToolingTests.dir/ToolingTest.cpp.o: In function `clang::tooling::newFrontendActionFactory_CreatesFrontendActionFactoryFromType_Test::TestBody()':
ToolingTest.cpp:(.text._ZN5clang7tooling66newFrontendActionFactory_CreatesFrontendActionFactoryFromType_Test8TestBodyEv+0x49): undefined reference to `clang::SyntaxOnlyAction::~SyntaxOnlyAction()'
```
Reviewers: alexfh
Subscribers: cfe-commits
Differential Revision: http://reviews.llvm.org/D20218
llvm-svn: 269334
Summary:
This option allows the user to control how much of the file name is
emitted by UBSan. Tuning this option allows one to save space in the
resulting binary, which is helpful for restricted execution
environments.
With a positive N, UBSan skips the first N path components.
With a negative N, UBSan only keeps the last N path components.
Reviewers: rsmith
Subscribers: cfe-commits
Differential Revision: http://reviews.llvm.org/D19666
llvm-svn: 269309
Summary:
Teach clang to embed bitcode inside bitcode. When -fembed-bitcode cc1
option is used, clang will embed both the input bitcode and cc1
commandline into the bitcode in special sections before compiling to
the object file. Using -fembed-bitcode-marker will only introduce a
marker in both sections.
Depends on D17390
Reviewers: rsmith
Subscribers: yaron.keren, vsk, cfe-commits
Differential Revision: http://reviews.llvm.org/D17392
llvm-svn: 269202
[ Copied from https://llvm.org/bugs/show_bug.cgi?id=26404 ]
clang support on Haiku is lagging a bit, and missing on x86_64.
This patch updates support for x86 and add support for x86_64. It should
apply directly to trunk and it's harmless in the sense that it only
affects Haiku.
Reviewers: rnk, rsmith
Patch by Jérôme Duval
Differential Revision: http://reviews.llvm.org/D16797
llvm-svn: 269201
Support OpenMP version 4.5 syntax for #pragma omp declare target.
Syntax:
#pragma omp declare target (extended-list) new-line
or
#pragma omp declare target clause[ [,] clause ... ] new-line
Where clause is one of the following:
to(extended-list)
link(list)
Differential Revision: http://reviews.llvm.org/D20011
llvm-svn: 268925
When running reproducer scripts we need that original symlinks from the
source filesystem are reproduced in the VFS so that different virtual
paths can map to the same file, allowing the FileManager to share the
same UID between these virtual entries. This avoids all sorts of module
redefinition errors when using frameworks.
llvm-svn: 268825
Use a StringRef instead of a FileEntry in the moduleMapAddHeader
callback to allow more flexibility on what to collect on further
patches. This changes the interface I introduced in r264971.
llvm-svn: 268819
From the GCC manpage:
-I dir
... If dir begins with =, then the = will be replaced by the sysroot prefix;
see --sysroot and -isysroot.
Add support to expand the `=` as a prefix of the include path with the sysroot
if specified. `-isysroot` takes precedence over `--sysroot` as the normal
argument behaviour occurs. The ordering of the `-isysroot` is relevant to the
path substituted. If no `--sysroot=` or `-isysroot` option is present, the = is
not expanded.
Resolves PR26965!
llvm-svn: 268777
This option, like the corresponding gcc option, is SystemZ-specific and
enables storing frame backchain links, as specified in the ABI.
Differential Revision: http://reviews.llvm.org/D19891
llvm-svn: 268575
Let the loop unroll pass handle /Os. It already checks that option and adjust its thresholds accordingly. Also, will allow the #pragma unroll to have an effect in /Os.
Differential Revision: http://reviews.llvm.org/D19827
llvm-svn: 268509
The -malign-double flag causes i64 and f64 types to have alignment 8
instead of 4. On x86-64, the behavior of -malign-double is enabled by default.
Rebases and cleans phosek's work here: http://reviews.llvm.org/D12860
Patch by Sean Klein
Reviewers: rnk
Subscribers: rnk, jfb, dschuff, phosek
Differential Revision: http://reviews.llvm.org/D19734
llvm-svn: 268473
Summary:
Host and device types must match, otherwise when we pass values back and
forth between the host and device, we will get the wrong result.
This patch makes NVPTXTargetInfo inherit most of its type information
from the host's target info.
Reviewers: rsmith
Subscribers: cfe-commits, jhen, tra
Differential Revision: http://reviews.llvm.org/D19346
llvm-svn: 268131
It makes compiler-rt tests fail if the gold plugin is enabled.
Revert "Rework interface for bitset-using features to use a notion of LTO visibility."
Revert "Driver: only produce CFI -fvisibility= error when compiling."
Revert "clang/test/CodeGenCXX/cfi-blacklist.cpp: Exclude ms targets. They would be non-cfi."
llvm-svn: 267871
Bitsets, and the compiler features they rely on (vtable opt, CFI),
only have visibility within the LTO'd part of the linkage unit. Therefore,
only enable these features for classes with hidden LTO visibility. This
notion is based on object file visibility or (on Windows)
dllimport/dllexport attributes.
We provide the [[clang::lto_visibility_public]] attribute to override the
compiler's LTO visibility inference in cases where the class is defined
in the non-LTO'd part of the linkage unit, or where the ABI supports
calling classes derived from abstract base classes with hidden visibility
in other linkage units (e.g. COM on Windows).
If the cross-DSO CFI mode is enabled, bitset checks are emitted even for
classes with public LTO visibility, as that mode uses a separate mechanism
to cause bitsets to be exported.
This mechanism replaces the whole-program-vtables blacklist, so remove the
-fwhole-program-vtables-blacklist flag.
Because __declspec(uuid()) now implies [[clang::lto_visibility_public]], the
support for the special attr:uuid blacklist entry is removed.
Differential Revision: http://reviews.llvm.org/D18635
llvm-svn: 267784
OpenCL spec requires __OPENCL_C_VERSION__ to be defined based on -cl-std option. This patch implements that.
The patch also defines __FAST_RELAXED_MATH__ based on -cl-fast-relaxed-math option.
Also fixed a test using -std=c99 for OpenCL program. Limit allowed language standard of OpenCL to be OpenCL standards.
Differential Revision: http://reviews.llvm.org/D19071
llvm-svn: 267590
Hide the real paths when rebuilding from VFS by setting up the crash
reproducer to use 'use-external-names' = false. This way we avoid
module redifinition errors and consistently use the same paths against
all modules.
With this change on Darwin we are able to simulate a crash for a simple
application using "Foundation/Foundation.h" (which relies on a bunch of
different frameworks and headers) and successfully rebuild all the
modules by relying solely at the VFS overlay.
llvm-svn: 266234
It is possible to argue that the EABIVersion field is similar in spirit to the
ABI field in TargetOptions. It represents the embedded ABI that the target
follows. This will allow us to thread this information into the target
information construction.
llvm-svn: 265807
Revert the two changes to thread CodeGenOptions into the TargetInfo allocation
and to fix the layering violation by moving CodeGenOptions into Basic.
Code Generation is arguably not particularly "basic". This addresses Richard's
post-commit review comments. This change purely does the mechanical revert and
will be followed up with an alternate approach to thread the desired information
into TargetInfo.
llvm-svn: 265806
CodeGen-level implementation. Instead of adding an attribute to clang's
FunctionDecl, add the IR attribute directly. This means a module built with
this flag is now compatible with code built without it and vice versa.
This change also results in the 'noalias' attribute no longer being added to
calls to operator new in the IR; it's now only added to the declaration. It
also fixes a bug where we failed to add the attribute to the 'nothrow' versions
(because we didn't implicitly declare them, there was no good time to inject a
fake attribute).
llvm-svn: 265728
This is a mechanical move of CodeGenOptions from libFrontend to libBasic. This
fixes the layering violation introduced earlier by threading CodeGenOptions into
TargetInfo. It should also fix the modules based self-hosting builds. NFC.
llvm-svn: 265702
This threads CodeGenOptions into the TargetInfo hierarchy. This is motivated by
ARM which can change some target information based on the EABI selected
(-meabi). Similar options exist for other platforms (e.g. MIPS) and thus is
generally useful. NFC.
llvm-svn: 265640
This is a SmallVector anyway, and so the exact size doesn't matter.
clang\lib\Frontend\ModuleDependencyCollector.cpp(83) : error C2065: 'PATH_MAX' : undeclared identifier
clang\lib\Frontend\ModuleDependencyCollector.cpp(83) : error C2975: 'InternalLen' : invalid template argument for 'llvm::SmallString', expected compile-time constant expression
llvm\include\llvm/ADT/SmallString.h(24) : see declaration of 'InternalLen'
llvm-svn: 265634
The crash reproducer was not setting up case sensitivity in the
VFS yaml files, which defaults to true. Make the crash reproducer
explicitly set that flag based on the case sensitivity of the .cache
path where vfs and modules are dumped.
llvm-svn: 265622
Add parsing, sema analysis for 'declare target' construct for OpenMP 4.0
(4.5 support will be added in separate patch).
The declare target directive specifies that variables, functions (C, C++
and Fortran), and subroutines (Fortran) are mapped to a device. The declare
target directive is a declarative directive. In Clang declare target is
implemented as implicit attribute for the declaration.
The syntax of the declare target directive is as follows:
#pragma omp declare target
declarations-definition-seq
#pragma omp end declare target
Based on patch from Michael Wong http://reviews.llvm.org/D15321
llvm-svn: 265530
Summary:
Setting this flag causes all functions are annotated with the
"nvvm-f32ftz" = "true" attribute.
In addition, we annotate the module with "nvvm-reflect-ftz" set
to 0 or 1, depending on whether -cuda-flush-denormals-to-zero is set.
This is read by the NVVMReflect pass.
Reviewers: tra, rnk
Subscribers: cfe-commits
Differential Revision: http://reviews.llvm.org/D18671
llvm-svn: 265435
Add no-jump-tables flag to disable use of jump tables when lowering
switch statements
Reviewers: echristo, hans
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D18407
llvm-svn: 265425
The current ModuleDependencyCollector has a AST listener to collect
header files present in loaded modules, but this isn't enough to collect
all headers needed in the crash reproducer. One of the reasons is that
the AST writer doesn't write symbolic link header paths in the pcm modules,
this makes the listeners on the reader only able to collect the real files.
Since the module maps could contain submodules that use headers which
are symbolic links, not collecting those forbid the reproducer scripts
to regen the modules.
For instance:
usr/include/module.map:
...
module pthread {
header "pthread.h"
export *
module impl {
header "pthread_impl.h"
export *
}
}
...
usr/include/pthread/pthread_impl.h
usr/include/pthread_impl.h -> pthread/pthread_impl.h
The AST dump for the module above:
<SUBMODULE_HEADER abbrevid=6/> blob data = 'pthread_impl.h'
<SUBMODULE_TOPHEADER abbrevid=7/> blob data = '/<path_to_sdk>/usr/include/pthread/pthread_impl.h'
Note that we don't have "usr/include/pthread_impl.h" which is requested
by the module.map in case we want to reconstruct the module in the
reproducer. The reason the original symbolic link path isn't used is
because the headers are kept by name and requested through the
FileManager, which unique files and returns the real path only.
To fix that, add a callback to be invoked everytime a header is added
while parsing module maps and hook that up to the module dependecy
collector. This callback is only registered when generating the
reproducer.
Differential Revision: http://reviews.llvm.org/D18585
rdar://problem/24499339
llvm-svn: 264971
With this patch, by a constexpr function is implicitly host+device
unless:
a) it's a variadic function (variadic functions are not allowed on the
device side), or
b) it's preceeded by a __device__ overload in a system header.
The restriction on overloading __host__ __device__ functions on the
basis of their CUDA attributes remains in place, but we use (b) to allow
us to define __device__ overloads for constexpr functions in cmath,
which would otherwise be __host__ __device__ and thus not overloadable.
You can disable this behavior with -fno-cuda-host-device-constexpr.
Reviewers: tra, rnk, rsmith
Subscribers: cfe-commits
Differential Revision: http://reviews.llvm.org/D18380
llvm-svn: 264964
...as that is apparently what MSVC does. This is an updated version of r263738,
which had to be reverted in r263740 due to test failures. The original version
had erroneously emitted functions that are defined in class templates, too (see
the updated "Handle friend functions" code in EmitDeferredDecls,
lib/CodeGen/ModuleBuilder.cpp). (The updated tests needed to be split out into
their own dllexport-ms-friend.cpp because of the CHECK-NOTs which would have
interfered with subsequent CHECK-DAGs in dllexport.cpp.)
Differential Revision: http://reviews.llvm.org/D18430
llvm-svn: 264841
- Make ModuleDependencyCollector use the DependencyCollector interface
- Move some methods from ModuleDependencyListener to ModuleDependencyCollector
in order to share common functionality with other future possible
callbacks.
llvm-svn: 264808
Summary:
* -fcuda-target-overloads
Previously unconditionally set to true by the driver. Necessary for
correct functioning of the compiler -- our CUDA headers wrapper won't
compile without this.
* -fcuda-disable-target-call-checks
Previously unconditionally set to true by the driver. Necessary to
compile almost any external CUDA code -- almost all libraries assume
that host+device code can call host or device functions.
* -fcuda-allow-host-calls-from-host-device
No effect when target overloading is enabled.
Reviewers: tra
Subscribers: rsmith, cfe-commits
Differential Revision: http://reviews.llvm.org/D18416
llvm-svn: 264739
include altivec.h has come and gone.
Rationale: This causes modules, rewrite-includes, etc to be sad and
people should just include altivec.h in their source.
llvm-svn: 264235
Remove tests that have neither a triple nor an explicit -fmsc-version flag,
since in the absence of an -fmsc-version flag, the implicit value of the flag
is 17 (MSVC2013) with MSVC triples but 0 (not set) for other triples, and
the default triple is platform dependent.
This relands r263974 with a test fix.
llvm-svn: 264210
Instead of putting the /Yc header into ExtraDeps, give DependencyOutputOptions
a dedicated field for /Yc mode, and let HeaderIncludesCallback hang on to the
full DependencyOutputOptions object, not just ExtraDeps.
Reverts parts of r263352 that are now no longer needed.
llvm-svn: 264182
-H in gcc mode doesn't print -include headers, but they are included in
depfiles written by MMD and friends. Since /showIncludes is what's used instead
of depfiles, printing /FI there seems important (and matches cl.exe).
Instead of giving HeaderIncludeGen more options, just switch on ShowAllHeaders
in clang-cl mode and let clang::InitializePreprocessor() not put -include flags
in the <command line> block. This changes the behavior of -E slightly, and it
removes the <command line> flag from the output triggered by setting the
obscure CC_PRINT_HEADERS=1 env var to true while running clang. Both of these
seem ok to change.
http://reviews.llvm.org/D18401
llvm-svn: 264174
This reverts commit r263607.
This change caused more objc_retain/objc_release calls in the IR but those
are then incorrectly optimized by the ARC optimizer. Work is going to have
to be done to ensure the ARC optimizer doesn't optimize user written RR, but
that should land before this change.
This change will also need to be updated to take account for any changes required
to ensure that user written calls to RR are distinct from those inserted by ARC.
llvm-svn: 263984
Renato Golin