It (introduced by 556d713c70) appears to be
related to the removed dragonegg project. In addition, the feature was a bit
misnamed and may lur users to unnecessarily use it.
Introduce -fgpu-default-stream={legacy|per-thread} option to
support per-thread default stream for HIP runtime.
When -fgpu-default-stream=per-thread, HIP kernels are
launched through hipLaunchKernel_spt instead of
hipLaunchKernel. Also HIP_API_PER_THREAD_DEFAULT_STREAM=1
is defined by the preprocessor to enable other per-thread stream
API's.
Reviewed by: Artem Belevich
Differential Revision: https://reviews.llvm.org/D120298
Diagnose -fopenmp-targets for HIP programs since
dual HIP and OpenMP offloading in the same compilation
is currently not supported by HIP toolchain.
Reviewed by: Artem Belevich
Differential Revision: https://reviews.llvm.org/D109718
A recent change (D99683) to support ThinLTO for HIP caused a regression
when compiling cuda code with -flto=thin -fwhole-program-vtables.
Specifically, we now get an error:
error: invalid argument '-fwhole-program-vtables' only allowed with '-flto'
This error is coming from the device offload cc1 action being set up for
the cuda compile, for which -flto=thin doesn't apply and gets dropped.
This is a regression, but points to a potential issue that was silently
occurring before the patch, details below.
Before D99683, the check for fwhole-program-vtables in the driver looked
like:
if (WholeProgramVTables) {
if (!D.isUsingLTO())
D.Diag(diag::err_drv_argument_only_allowed_with)
<< "-fwhole-program-vtables"
<< "-flto";
CmdArgs.push_back("-fwhole-program-vtables");
}
And D.isUsingLTO() returned true since we have -flto=thin. However,
because the cuda cc1 compile is doing device offloading, which didn't
support any LTO, there was other code that suppressed -flto* options
from being passed to the cc1 invocation. So the cc1 invocation silently
had -fwhole-program-vtables without any -flto*. This seems potentially
problematic, since if we had any virtual calls we would get type test
assume sequences without the corresponding LTO pass that handles them.
However, with the patch, which adds support for device offloading LTO
option -foffload-lto=thin, the code has changed so that we set a bool
IsUsingLTO based on either -flto* or -foffload-lto*, depending on
whether this is the device offloading action. For the device offload
action in our compile, since we don't have -foffload-lto, IsUsingLTO is
false, and the check for LTO with -fwhole-program-vtables now fails.
What we should do is only pass through -fwhole-program-vtables to the
cc1 invocation that has LTO enabled (either the device offload action
with -foffload-lto, or the non-device offload action with -flto), and
otherwise drop the -fwhole-program-vtables for the non-LTO action.
Then we should error only if we have -fwhole-program-vtables without any
-f*lto* options.
Differential Revision: https://reviews.llvm.org/D103579
Add options -[no-]offload-lto and -foffload-lto=[thin,full] for controlling
LTO for offload compilation. Allow LTO for AMDGPU target.
AMDGPU target does not support codegen of object files containing
call of external functions, therefore the LLVM module passed to
AMDGPU backend needs to contain definitions of all the callees.
An LLVM option is added to allow function importer to import
functions with noinline attribute.
HIP toolchain passes proper LLVM options to lld to make sure
function importer imports definitions of all the callees.
Reviewed by: Teresa Johnson, Artem Belevich
Differential Revision: https://reviews.llvm.org/D99683
A bug was introduced when adding -munsafe-fp-atomics.
By default it should be off.
Reviewed by: Artem Belevich
Differential Revision: https://reviews.llvm.org/D97967
Like nvptx and some other targets, -mconstructor-aliases does not work well with amdgpu,
therefore we disable it in the same approach.
Reviewed by: Artem Belevich
Differential Revision: https://reviews.llvm.org/D97959
This patch implements correct hostness based overloading resolution
in isBetterOverloadCandidate.
Based on hostness, if one candidate is emittable whereas the other
candidate is not emittable, the emittable candidate is better.
If both candidates are emittable, or neither is emittable based on hostness, then
other rules should be used to determine which is better. This is because
hostness based overloading resolution is mostly for determining
viability of a function. If two functions are both viable, other factors
should take precedence in preference.
If other rules cannot determine which is better, CUDA preference will be
used again to determine which is better.
However, correct hostness based overloading resolution
requires overloading resolution diagnostics to be deferred,
which is not on by default. The rationale is that deferring
overloading resolution diagnostics may hide overloading reslolutions
issues in header files.
An option -fgpu-exclude-wrong-side-overloads is added, which is off by
default.
When -fgpu-exclude-wrong-side-overloads is off, keep the original behavior,
that is, exclude wrong side overloads only if there are same side overloads.
This may result in incorrect overloading resolution when there are no
same side candates, but is sufficient for most CUDA/HIP applications.
When -fgpu-exclude-wrong-side-overloads is on, enable deferring
overloading resolution diagnostics and enable correct hostness
based overloading resolution, i.e., always exclude wrong side overloads.
Differential Revision: https://reviews.llvm.org/D80450
Add an option -munsafe-fp-atomics for AMDGPU target.
When enabled, clang adds function attribute "amdgpu-unsafe-fp-atomics"
to any functions for amdgpu target. This allows amdgpu backend to use
unsafe fp atomic instructions in these functions.
Differential Revision: https://reviews.llvm.org/D91546
The argument after -Xarch_device will be added to the arguments for CUDA/HIP
device compilation and will be removed for host compilation.
The argument after -Xarch_host will be added to the arguments for CUDA/HIP
host compilation and will be removed for device compilation.
Differential Revision: https://reviews.llvm.org/D76520
HIPToolChain::TranslateArgs call TranslateArgs of host toolchain with
the input args to get a list of derived args called DAL, then
go through the input args by itself and append them to DAL.
This assumes that the host toolchain should not append any unchanged
args to DAL, otherwise there will be duplicates since
HIPToolChain will append it again.
This works for GNU toolchain since it returns an empty list for DAL.
However, MSVC toolchain will append unchanged args to DAL, which
causes duplicate args.
This patch let MSVC toolchain not append unchanged args for HIP
offloading kind, which fixes this issue.
Differential Revision: https://reviews.llvm.org/D76032
Fangrui Song