Currently rocm detector expects device library bitcodes named as *.bc
instead of *.amdgcn.bc. However in rocm3.5 the device library bitcodes
are named as *.amdgcn.bc, which causes rocm3.5 not detected.
This patch fixes that.
Differential Revision: https://reviews.llvm.org/D81713
It's useful for using clang from tools that may need need to provide SDK files
from non-standard locations.
Clang CLI only provides a way to specify VFS for include files, so there's no
good way to test this yet.
Differential Revision: https://reviews.llvm.org/D81771
To support std::complex and some other standard C/C++ functions in HIP device code,
they need to be forced to be __host__ __device__ functions by pragmas. This is done
by some clang standard C++ wrapper headers which are shared between cuda-clang and hip-Clang.
For these standard C++ wapper headers to work properly, specific include path order
has to be enforced:
clang C++ wrapper include path
standard C++ include path
clang include path
Also, these C++ wrapper headers require device version of some standard C/C++ functions
must be declared before including them. This needs to be done by including a default
header which declares or defines these device functions. The default header is always
included before any other headers are included by users.
This patch adds the the default header and include path for HIP.
Differential Revision: https://reviews.llvm.org/D81176
The various HIP builds are all inconsistent.
The default llvm install goes to ${INSTALL_PREFIX}/bin/clang, but the
rocm packaging scripts move this under
${INSTALL_PREFIX}/llvm/bin/clang. Some other builds further pollute
this with ${INSTALL_PREFIX}/bin/x86_64/clang. These should really be
consolidated, but try to handle them for now.
-nogpulib makes sense when there is a host (where -nostdlib would
apply) and offload target. Accept nostdlib when there is no offload
target as an alias.
Merge with the new --rocm-path handling used for OpenCL. This looks
for a usable set of device libraries upfront, rather than giving a
generic "no such file or directory error". If any of the required
bitcode libraries are missing, this will now produce a "cannot find
ROCm installation." error. This differs from the existing hip specific
flags by pointing to a rocm root install instead of a single directory
with bitcode files.
This tries to maintain compatibility with the existing the
--hip-device-lib and --hip-device-lib-path flags, as well as the
HIP_DEVICE_LIB_PATH environment variable, or at least the range of
uses with testcases. The existing range of uses and behavior doesn't
entirely make sense to me, so some of the untested edge cases change
behavior. Currently the two path forms seem to have the double purpose
of a search path for an arbitrary --hip-device-lib, and for finding
the stock set of libraries. Since the stock set of libraries This also
changes the behavior when multiple paths are specified, and only takes
the last one (and the environment variable only handles a single
path).
If --hip-device-lib is used, it now only treats --hip-device-lib-path
as the search path for it, and does not attempt to find the rocm
installation. If not, --hip-device-lib-path and the environment
variable are used as the directory to search instead of the rocm root
based path.
This should also automatically fix handling of the options to use
wave64.
Since the default logic was based on having fast denormal/fma
features, and the default target has no features, we assumed flushing
by default. This fixes incorrectly assuming flushing in builds for
"generic" IR libraries.
The handling for no specified --cuda-gpu-arch in HIP is kind of
broken. Somewhere else forces a default target of gfx803, which does
not enable denormal handling by default. We don't see this default
switching here, so you'll end up with a different denormal mode
depending on whether you explicitly requested gfx803, or used it by
default.
I didn't realize HIP was a distinct offloading kind, so the subtarget
was looking for -march, which isn't correct for HIP. We also have the
possibility of different denormal defaults in the case of multiple
offload targets, so we need to thread the JobAction through the target
hook.
Currently the library is separately linked, but this isn't correct to
implement fast math flags correctly. Each module should get the
version of the library appropriate for its combination of fast math
and related flags, with the attributes propagated into its functions
and internalized.
HIP already maintains the list of libraries, but this is not used for
OpenCL. Unfortunately, HIP uses a separate --hip-device-lib argument,
despite both languages using the same bitcode library. Eventually
these two searches need to be merged.
An additional problem is there are 3 different locations the libraries
are installed, depending on which build is used. This also needs to be
consolidated (or at least the search logic needs to deal with this
unnecessary complexity).
Apparently HIPToolChain does not subclass from AMDGPUToolChain, so
this was not applying the new denormal attributes. I'm not sure why
this doesn't subclass. Just copy the implementation for now.
AMDGPU and x86 at least both have separate controls for whether
denormal results are flushed on output, and for whether denormals are
implicitly treated as 0 as an input. The current DAGCombiner use only
really cares about the input treatment of denormals.
Currently there are 4 different mechanisms for controlling denormal
flushing behavior, and about as many equivalent frontend controls.
- AMDGPU uses the fp32-denormals and fp64-f16-denormals subtarget features
- NVPTX uses the nvptx-f32ftz attribute
- ARM directly uses the denormal-fp-math attribute
- Other targets indirectly use denormal-fp-math in one DAGCombine
- cl-denorms-are-zero has a corresponding denorms-are-zero attribute
AMDGPU wants a distinct control for f32 flushing from f16/f64, and as
far as I can tell the same is true for NVPTX (based on the attribute
name).
Work on consolidating these into the denormal-fp-math attribute, and a
new type specific denormal-fp-math-f32 variant. Only ARM seems to
support the two different flush modes, so this is overkill for the
other use cases. Ideally we would error on the unsupported
positive-zero mode on other targets from somewhere.
Move the logic for selecting the flush mode into the compiler driver,
instead of handling it in cc1. denormal-fp-math/denormal-fp-math-f32
are now both cc1 flags, but denormal-fp-math-f32 is not yet exposed as
a user flag.
-cl-denorms-are-zero, -fcuda-flush-denormals-to-zero and
-fno-cuda-flush-denormals-to-zero will be mapped to
-fp-denormal-math-f32=ieee or preserve-sign rather than the old
attributes.
Stop emitting the denorms-are-zero attribute for the OpenCL flag. It
has no in-tree users. The meaning would also be target dependent, such
as the AMDGPU choice to treat this as only meaning allow flushing of
f32 and not f16 or f64. The naming is also potentially confusing,
since DAZ in other contexts refers to instructions implicitly treating
input denormals as zero, not necessarily flushing output denormals to
zero.
This also does not attempt to change the behavior for the current
attribute. The LangRef now states that the default is ieee behavior,
but this is inaccurate for the current implementation. The clang
handling is slightly hacky to avoid touching the existing
denormal-fp-math uses. Fixing this will be left for a future patch.
AMDGPU is still using the subtarget feature to control the denormal
mode, but the new attribute are now emitted. A future change will
switch this and remove the subtarget features.
Now that we've moved to C++14, we no longer need the llvm::make_unique
implementation from STLExtras.h. This patch is a mechanical replacement
of (hopefully) all the llvm::make_unique instances across the monorepo.
Differential revision: https://reviews.llvm.org/D66259
llvm-svn: 368942
Introduce an option to request global visibility settings be applied to
declarations without a definition or an explicit visibility, rather than
the existing behavior of giving these default visibility. When the
visibility of all or most extern definitions are known this allows for
the same optimisations -fvisibility permits without updating source code
to annotate all declarations.
Differential Revision: https://reviews.llvm.org/D56868
llvm-svn: 352391
to reflect the new license.
We understand that people may be surprised that we're moving the header
entirely to discuss the new license. We checked this carefully with the
Foundation's lawyer and we believe this is the correct approach.
Essentially, all code in the project is now made available by the LLVM
project under our new license, so you will see that the license headers
include that license only. Some of our contributors have contributed
code under our old license, and accordingly, we have retained a copy of
our old license notice in the top-level files in each project and
repository.
llvm-svn: 351636
Object linking isn't supported, so it's not useful
to emit default visibility. Default visibility requires
relocations we don't yet support for functions compiled
in another translation unit.
WebAssembly already does this, although they insert these
arguments in a different place for some reason.
llvm-svn: 341033
In current OpenCL implementation some options are set in OpenCL RT/Driver, which causes discrepancy between online and offline paths.
Implement infrastructure to move options from OpenCL RT/Driver to AMDGPUToolChain using overloaded TranslateArgs() method.
Create map for default options values, as Options.td doesn't support default values (in contrast with OPTIONS.def).
Add two driver options: -On and -mNN (like -O3, -m64).
Some minor formatting changes to follow the clang-format style.
Differential Revision: https://reviews.llvm.org/D37386
llvm-svn: 312524
Summary:
(This is a move-only refactoring patch. There are no functionality changes.)
This patch splits apart the Clang driver's tool and toolchain implementation
files. Each target platform toolchain is moved to its own file, along with the
closest-related tools. Each target platform toolchain has separate headers and
implementation files, so the hierarchy of classes is unchanged.
There are some remaining shared free functions, mostly from Tools.cpp. Several
of these move to their own architecture-specific files, similar to r296056. Some
of them are only used by a single target platform; since the tools and
toolchains are now together, some helpers now live in a platform-specific file.
The balance are helpers related to manipulating argument lists, so they are now
in a new file pair, CommonArgs.h and .cpp.
I've tried to cluster the code logically, which is fairly straightforward for
most of the target platforms and shared architectures. I think I've made
reasonable choices for these, as well as the various shared helpers; but of
course, I'm happy to hear feedback in the review.
There are some particular things I don't like about this patch, but haven't been
able to find a better overall solution. The first is the proliferation of files:
there are several files that are tiny because the toolchain is not very
different from its base (usually the Gnu tools/toolchain). I think this is
mostly a reflection of the true complexity, though, so it may not be "fixable"
in any reasonable sense. The second thing I don't like are the includes like
"../Something.h". I've avoided this largely by clustering into the current file
structure. However, a few of these includes remain, and in those cases it
doesn't make sense to me to sink an existing file any deeper.
Reviewers: rsmith, mehdi_amini, compnerd, rnk, javed.absar
Subscribers: emaste, jfb, danalbert, srhines, dschuff, jyknight, nemanjai, nhaehnle, mgorny, cfe-commits
Differential Revision: https://reviews.llvm.org/D30372
llvm-svn: 297250
Yaxun (Sam) Liu