This is the second part loosely extracted from D71179 and cleaned up.
This patch provides semantic analysis support for `omp begin/end declare
variant`, mostly as defined in OpenMP technical report 8 (TR8) [0].
The sema handling makes code generation obsolete as we generate "the
right" calls that can just be handled as usual. This handling also
applies to the existing, albeit problematic, `omp declare variant
support`. As a consequence a lot of unneeded code generation and
complexity is removed.
A major purpose of this patch is to provide proper `math.h`/`cmath`
support for OpenMP target offloading. See PR42061, PR42798, PR42799. The
current code was developed with this feature in mind, see [1].
The logic is as follows:
If we have seen a `#pragma omp begin declare variant match(<SELECTOR>)`
but not the corresponding `end declare variant`, and we find a function
definition we will:
1) Create a function declaration for the definition we were about to generate.
2) Create a function definition but with a mangled name (according to
`<SELECTOR>`).
3) Annotate the declaration with the `OMPDeclareVariantAttr`, the same
one used already for `omp declare variant`, using and the mangled
function definition as specialization for the context defined by
`<SELECTOR>`.
When a call is created we inspect it. If the target has an
`OMPDeclareVariantAttr` attribute we try to specialize the call. To this
end, all variants are checked, the best applicable one is picked and a
new call to the specialization is created. The new call is used instead
of the original one to the base function. To keep the AST printing and
tooling possible we utilize the PseudoObjectExpr. The original call is
the syntactic expression, the specialized call is the semantic
expression.
[0] https://www.openmp.org/wp-content/uploads/openmp-TR8.pdf
[1] https://reviews.llvm.org/D61399#change-496lQkg0mhRN
Reviewers: kiranchandramohan, ABataev, RaviNarayanaswamy, gtbercea, grokos, sdmitriev, JonChesterfield, hfinkel, fghanim, aaron.ballman
Subscribers: bollu, guansong, openmp-commits, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D75779
Summary:
- Even though the bindless surface/texture interfaces are promoted,
there are still code using surface/texture references. For example,
[PR#26400](https://bugs.llvm.org/show_bug.cgi?id=26400) reports the
compilation issue for code using `tex2D` with texture references. For
better compatibility, this patch proposes the support of
surface/texture references.
- Due to the absent documentation and magic headers, it's believed that
`nvcc` does use builtins for texture support. From the limited NVVM
documentation[^nvvm] and NVPTX backend texture/surface related
tests[^test], it's believed that surface/texture references are
supported by replacing their reference types, which are annotated with
`device_builtin_surface_type`/`device_builtin_texture_type`, with the
corresponding handle-like object types, `cudaSurfaceObject_t` or
`cudaTextureObject_t`, in the device-side compilation. On the host
side, that global handle variables are registered and will be
established and updated later when corresponding binding/unbinding
APIs are called[^bind]. Surface/texture references are most like
device global variables but represented in different types on the host
and device sides.
- In this patch, the following changes are proposed to support that
behavior:
+ Refine `device_builtin_surface_type` and
`device_builtin_texture_type` attributes to be applied on `Type`
decl only to check whether a variable is of the surface/texture
reference type.
+ Add hooks in code generation to replace that reference types with
the correponding object types as well as all accesses to them. In
particular, `nvvm.texsurf.handle.internal` should be used to load
object handles from global reference variables[^texsurf] as well as
metadata annotations.
+ Generate host-side registration with proper template argument
parsing.
---
[^nvvm]: https://docs.nvidia.com/cuda/pdf/NVVM_IR_Specification.pdf
[^test]: https://raw.githubusercontent.com/llvm/llvm-project/master/llvm/test/CodeGen/NVPTX/tex-read-cuda.ll
[^bind]: See section 3.2.11.1.2 ``Texture reference API` in [CUDA C Programming Guide](https://docs.nvidia.com/cuda/pdf/CUDA_C_Programming_Guide.pdf).
[^texsurf]: According to NVVM IR, `nvvm.texsurf.handle` should be used. But, the current backend doesn't have that supported. We may revise that later.
Reviewers: tra, rjmccall, yaxunl, a.sidorin
Subscribers: cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D76365
Passing small data limit to RISCVELFTargetObjectFile by module flag,
So the backend can set small data section threshold by the value.
The data will be put into the small data section if the data smaller than
the threshold.
Differential Revision: https://reviews.llvm.org/D57497
Summary:
- https://reviews.llvm.org/D68578 revises the `GlobalDecl` constructors
to ensure all GPU kernels have `ReferenceKenelKind` initialized
properly with an explicit constructor and static one. But, there are
lots of places using the implicit constructor triggering the assertion
on non-GPU kernels. That's found in compilation of many tests and
workloads.
- Fixing all of them may change more code and, more importantly, all of
them assumes the default kernel reference kind. This patch changes
that constructor to tell `CUDAGlobalAttr` and construct `GlobalDecl`
properly.
Reviewers: yaxunl
Subscribers: cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D76344
Summary:
[Clang] Attribute to allow defining undef global variables
Initializing global variables is very cheap on hosted implementations. The
C semantics of zero initializing globals work very well there. It is not
necessarily cheap on freestanding implementations. Where there is no loader
available, code must be emitted near the start point to write the appropriate
values into memory.
At present, external variables can be declared in C++ and definitions provided
in assembly (or IR) to achive this effect. This patch provides an attribute in
order to remove this reason for writing assembly for performance sensitive
freestanding implementations.
A close analogue in tree is LDS memory for amdgcn, where the kernel is
responsible for initializing the memory after it starts executing on the gpu.
Uninitalized variables in LDS are observably cheaper than zero initialized.
Patch is loosely based on the cuda __shared__ and opencl __local variable
implementation which also produces undef global variables.
Reviewers: kcc, rjmccall, rsmith, glider, vitalybuka, pcc, eugenis, vlad.tsyrklevich, jdoerfert, gregrodgers, jfb, aaron.ballman
Reviewed By: rjmccall, aaron.ballman
Subscribers: Anastasia, aaron.ballman, davidb, Quuxplusone, dexonsmith, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D74361
Most clients of SourceManager.h need to do things like turning source
locations into file & line number pairs, but this doesn't require
bringing in FileManager.h and LLVM's FS headers.
The main code change here is to sink SM::createFileID into the cpp file.
I reason that this is not performance critical because it doesn't happen
on the diagnostic path, it happens along the paths of macro expansion
(could be hot) and new includes (less hot).
Saves some includes:
309 - /usr/local/google/home/rnk/llvm-project/clang/include/clang/Basic/FileManager.h
272 - /usr/local/google/home/rnk/llvm-project/clang/include/clang/Basic/FileSystemOptions.h
271 - /usr/local/google/home/rnk/llvm-project/llvm/include/llvm/Support/VirtualFileSystem.h
267 - /usr/local/google/home/rnk/llvm-project/llvm/include/llvm/Support/FileSystem.h
266 - /usr/local/google/home/rnk/llvm-project/llvm/include/llvm/Support/Chrono.h
Differential Revision: https://reviews.llvm.org/D75406
HIP emits a device stub function for each kernel in host code.
The HIP debugger requires device stub function to have a different unmangled name as the kernel.
Currently the name of the device stub function is the mangled name with a postfix .stub. However,
this does not work with the HIP debugger since the unmangled name is the same as the kernel.
This patch adds prefix __device__stub__ to the unmangled name of the device stub before mangling,
therefore the device stub function has a valid mangled name which is different than the device kernel
name. The device side kernel name is kept unchanged. kernels with extern "C" also gets the prefix added
to the corresponding device stub function.
Differential Revision: https://reviews.llvm.org/D68578
The initial implementation only did 'first declaration renaming' when
a default version came after. This is insufficient in cases where a
default does not exist, so this patch makes sure that we do the renaming
in all cases.
This renaming is necessary because we emit the first declaration before
knowing that it IS a target multiversion function, which would change
its name. The second declaration (the one that caused the
multiversioning) then needs to make sure that the first one has its name
changed to be consistent with the resolver usage.
Summary:
hip-pinned-shadow global var should remain in the final code object irrespective
of whether it is used or not within the code. Add it to used list, so that it
will not get eliminated when it is unused.
Reviewers: yaxunl, tra, hliao
Reviewed By: yaxunl
Subscribers: hliao, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D75402
WebAssembly enforces a rule that caller and callee signatures must
match. This means that the traditional technique of passing `main`
`argc` and `argv` even when it doesn't need them doesn't work.
Currently the backend renames `main` to `__original_main`, however this
doesn't interact well with LTO'ing libc, and the name isn't intuitive.
This patch allows us to transition to `__main_argc_argv` instead.
This implements the proposal in
https://github.com/WebAssembly/tool-conventions/pull/134
with a flag to disable it when targeting Emscripten, though this is
expected to be temporary, as discussed in the proposal comments.
Differential Revision: https://reviews.llvm.org/D70700
Summary:
Right now we annotate C++'s `operator new` with `noalias` attribute,
which very much is healthy for optimizations.
However as per [[ http://eel.is/c++draft/basic.stc.dynamic.allocation | `[basic.stc.dynamic.allocation]` ]],
there are more promises on global `operator new`, namely:
* non-`std::nothrow_t` `operator new` *never* returns `nullptr`
* If `std::align_val_t align` parameter is taken, the pointer will also be `align`-aligned
* ~~global `operator new`-returned pointer is `__STDCPP_DEFAULT_NEW_ALIGNMENT__`-aligned ~~ It's more caveated than that.
Supplying this information may not cause immediate landslide effects
on any specific benchmarks, but it for sure will be healthy for optimizer
in the sense that the IR will better reflect the guarantees provided in the source code.
The caveat is `-fno-assume-sane-operator-new`, which currently prevents emitting `noalias`
attribute, and is automatically passed by Sanitizers ([[ https://bugs.llvm.org/show_bug.cgi?id=16386 | PR16386 ]]) - should it also cover these attributes?
The problem is that the flag is back-end-specific, as seen in `test/Modules/explicit-build-flags.cpp`.
But while it is okay to add `noalias` metadata in backend, we really should be adding at least
the alignment metadata to the AST, since that allows us to perform sema checks on it.
Reviewers: erichkeane, rjmccall, jdoerfert, eugenis, rsmith
Reviewed By: rsmith
Subscribers: xbolva00, jrtc27, atanasyan, nlopes, cfe-commits
Tags: #llvm, #clang
Differential Revision: https://reviews.llvm.org/D73380
This PR enables "XL" C++ ABI in frontend AST to IR codegen. And it is driven by
static init work. The current kind in Clang by default is Generic Itanium, which
has different behavior on static init with IBM xlclang compiler on AIX.
Differential Revision: https://reviews.llvm.org/D74015
Implement protection against the stack clash attack [0] through inline stack
probing.
Probe stack allocation every PAGE_SIZE during frame lowering or dynamic
allocation to make sure the page guard, if any, is touched when touching the
stack, in a similar manner to GCC[1].
This extends the existing `probe-stack' mechanism with a special value `inline-asm'.
Technically the former uses function call before stack allocation while this
patch provides inlined stack probes and chunk allocation.
Only implemented for x86.
[0] https://www.qualys.com/2017/06/19/stack-clash/stack-clash.txt
[1] https://gcc.gnu.org/ml/gcc-patches/2017-07/msg00556.html
This a recommit of 39f50da2a3 with proper LiveIn
declaration, better option handling and more portable testing.
Differential Revision: https://reviews.llvm.org/D68720
Implement protection against the stack clash attack [0] through inline stack
probing.
Probe stack allocation every PAGE_SIZE during frame lowering or dynamic
allocation to make sure the page guard, if any, is touched when touching the
stack, in a similar manner to GCC[1].
This extends the existing `probe-stack' mechanism with a special value `inline-asm'.
Technically the former uses function call before stack allocation while this
patch provides inlined stack probes and chunk allocation.
Only implemented for x86.
[0] https://www.qualys.com/2017/06/19/stack-clash/stack-clash.txt
[1] https://gcc.gnu.org/ml/gcc-patches/2017-07/msg00556.html
This a recommit of 39f50da2a3 with proper LiveIn
declaration, better option handling and more portable testing.
Differential Revision: https://reviews.llvm.org/D68720
Implement protection against the stack clash attack [0] through inline stack
probing.
Probe stack allocation every PAGE_SIZE during frame lowering or dynamic
allocation to make sure the page guard, if any, is touched when touching the
stack, in a similar manner to GCC[1].
This extends the existing `probe-stack' mechanism with a special value `inline-asm'.
Technically the former uses function call before stack allocation while this
patch provides inlined stack probes and chunk allocation.
Only implemented for x86.
[0] https://www.qualys.com/2017/06/19/stack-clash/stack-clash.txt
[1] https://gcc.gnu.org/ml/gcc-patches/2017-07/msg00556.html
This a recommit of 39f50da2a3 with better option
handling and more portable testing
Differential Revision: https://reviews.llvm.org/D68720
Implement protection against the stack clash attack [0] through inline stack
probing.
Probe stack allocation every PAGE_SIZE during frame lowering or dynamic
allocation to make sure the page guard, if any, is touched when touching the
stack, in a similar manner to GCC[1].
This extends the existing `probe-stack' mechanism with a special value `inline-asm'.
Technically the former uses function call before stack allocation while this
patch provides inlined stack probes and chunk allocation.
Only implemented for x86.
[0] https://www.qualys.com/2017/06/19/stack-clash/stack-clash.txt
[1] https://gcc.gnu.org/ml/gcc-patches/2017-07/msg00556.html
This a recommit of 39f50da2a3 with correct option
flags set.
Differential Revision: https://reviews.llvm.org/D68720
This reverts commit 39f50da2a3.
The -fstack-clash-protection is being passed to the linker too, which
is not intended.
Reverting and fixing that in a later commit.
Implement protection against the stack clash attack [0] through inline stack
probing.
Probe stack allocation every PAGE_SIZE during frame lowering or dynamic
allocation to make sure the page guard, if any, is touched when touching the
stack, in a similar manner to GCC[1].
This extends the existing `probe-stack' mechanism with a special value `inline-asm'.
Technically the former uses function call before stack allocation while this
patch provides inlined stack probes and chunk allocation.
Only implemented for x86.
[0] https://www.qualys.com/2017/06/19/stack-clash/stack-clash.txt
[1] https://gcc.gnu.org/ml/gcc-patches/2017-07/msg00556.html
Differential Revision: https://reviews.llvm.org/D68720
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.
Summary:
Clang -fpic defaults to -fno-semantic-interposition (GCC -fpic defaults
to -fsemantic-interposition).
Users need to specify -fsemantic-interposition to get semantic
interposition behavior.
Semantic interposition is currently a best-effort feature. There may
still be some cases where it is not handled well.
Reviewers: peter.smith, rnk, serge-sans-paille, sfertile, jfb, jdoerfert
Subscribers: dschuff, jyknight, dylanmckay, nemanjai, jvesely, kbarton, fedor.sergeev, asb, rbar, johnrusso, simoncook, sabuasal, niosHD, jrtc27, zzheng, edward-jones, atanasyan, rogfer01, MartinMosbeck, brucehoult, the_o, arphaman, PkmX, jocewei, jsji, Jim, lenary, s.egerton, pzheng, sameer.abuasal, apazos, luismarques, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D73865
First attempt at implementing -fsemantic-interposition.
Rely on GlobalValue::isInterposable that already captures most of the expected
behavior.
Rely on a ModuleFlag to state whether we should respect SemanticInterposition or
not. The default remains no.
So this should be a no-op if -fsemantic-interposition isn't used, and if it is,
isInterposable being already used in most optimisation, they should honor it
properly.
Note that it only impacts architecture compiled with -fPIC and no pie.
Differential Revision: https://reviews.llvm.org/D72829
This is how it should've been and brings it more in line with
std::string_view. There should be no functional change here.
This is mostly mechanical from a custom clang-tidy check, with a lot of
manual fixups. It uncovers a lot of minor inefficiencies.
This doesn't actually modify StringRef yet, I'll do that in a follow-up.
Summary:
First patch to support Safe Whole Program Devirtualization Enablement,
see RFC here: http://lists.llvm.org/pipermail/llvm-dev/2019-December/137543.html
Always emit !vcall_visibility metadata under -fwhole-program-vtables,
and not just for -fvirtual-function-elimination. The vcall visibility
metadata will (in a subsequent patch) be used to communicate to WPD
which vtables are safe to devirtualize, and we will optionally convert
the metadata to hidden visibility at link time. Subsequent follow on
patches will help enable this by adding vcall_visibility metadata to the
ThinLTO summaries, and always emit type test intrinsics under
-fwhole-program-vtables (and not just for vtables with hidden
visibility).
In order to do this safely with VFE, since for VFE all vtable loads must
be type checked loads which will no longer be the case, this patch adds
a new "Virtual Function Elim" module flag to communicate to GlobalDCE
whether to perform VFE using the vcall_visibility metadata.
One additional advantage of using the vcall_visibility metadata to drive
more WPD at LTO link time is that we can use the same mechanism to
enable more aggressive VFE at LTO link time as well. The link time
option proposed in the RFC will convert vcall_visibility metadata to
hidden (aka linkage unit visibility), which combined with
-fvirtual-function-elimination will allow it to be done more
aggressively at LTO link time under the same conditions.
Reviewers: pcc, ostannard, evgeny777, steven_wu
Subscribers: mehdi_amini, Prazek, hiraditya, dexonsmith, davidxl, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D71907
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.
The option will limit debug info by only emitting complete class
type information when its constructor is emitted.
This patch changes comparisons with LimitedDebugInfo to use the new
level instead.
Differential Revision: https://reviews.llvm.org/D72427
If a system header provides an (inline) implementation of some of their
function, clang still matches on the function name and generate the appropriate
llvm builtin, e.g. memcpy. This behavior is in line with glibc recommendation «
users may not provide their own version of symbols » but doesn't account for the
fact that glibc itself can provide inline version of some functions.
It is the case for the memcpy function when -D_FORTIFY_SOURCE=1 is on. In that
case an inline version of memcpy calls __memcpy_chk, a function that performs
extra runtime checks. Clang currently ignores the inline version and thus
provides no runtime check.
This code fixes the issue by detecting functions whose name is a builtin name
but also have an inline implementation.
Differential Revision: https://reviews.llvm.org/D71082
The validateOutputSize and validateInputSize need to check whether
AVX or AVX512 are enabled. But this can be affected by the
target attribute so we need to factor that in.
This patch moves some of the code from CodeGen to create an
appropriate feature map that we can pass to the function.
Differential Revision: https://reviews.llvm.org/D68627
Commit d77ae1552f
("[DebugInfo] Support to emit debugInfo for extern variables")
added deebugInfo for extern variables for BPF target.
The commit is reverted by 891e25b02d
as the committed tests using %clang instead of %clang_cc1 causing
test failed in certain scenarios as reported by Reid Kleckner.
This patch fixed the tests by using %clang_cc1.
Differential Revision: https://reviews.llvm.org/D71818
This is a follow up patch to use the OpenMP-IR-Builder, as discussed on
the mailing list ([1] and later) and at the US Dev Meeting'19.
[1] http://lists.flang-compiler.org/pipermail/flang-dev_lists.flang-compiler.org/2019-May/000197.html
Reviewers: kiranchandramohan, ABataev, RaviNarayanaswamy, gtbercea, grokos, sdmitriev, JonChesterfield, hfinkel, fghanim
Subscribers: ppenzin, penzn, llvm-commits, cfe-commits, jfb, guansong, bollu, hiraditya, mgorny
Tags: #clang
Differential Revision: https://reviews.llvm.org/D69922
Summary:
Attribute annotations are recorded in a special global composite variable
that points to annotation strings and the annotated objects.
As a restriction of the LLVM IR type system, those pointers are all
pointers to address space 0, so let's insert an addrspacecast when the
annotated global is in a non-0 address space.
Since this addrspacecast is only reachable from the global annotations
object, this should allow us to represent annotations on all globals
regardless of which addrspacecasts are usually legal for the target.
Reviewers: rjmccall
Subscribers: cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D71208
Extern variable usage in BPF is different from traditional
pure user space application. Recent discussion in linux bpf
mailing list has two use cases where debug info types are
required to use extern variables:
- extern types are required to have a suitable interface
in libbpf (bpf loader) to provide kernel config parameters
to bpf programs.
https://lore.kernel.org/bpf/CAEf4BzYCNo5GeVGMhp3fhysQ=_axAf=23PtwaZs-yAyafmXC9g@mail.gmail.com/T/#t
- extern types are required so kernel bpf verifier can
verify program which uses external functions more precisely.
This will make later link with actual external function no
need to reverify.
https://lore.kernel.org/bpf/87eez4odqp.fsf@toke.dk/T/#m8d5c3e87ffe7f2764e02d722cb0d8cbc136880ed
This patch added clang support to emit debuginfo for extern variables
with a TargetInfo hook to enable it. The debuginfo for the
extern variable is emitted only if that extern variable is
referenced in the current compilation unit.
Currently, only BPF target enables to generate debug info for
extern variables. The emission of such debuginfo is disabled for C++
at this moment since BPF only supports a subset of C language.
Emission with C++ can be enabled later if an appropriate use case
is identified.
-fstandalone-debug permits us to see more debuginfo with the cost
of bloated binary size. This patch did not add emission of extern
variable debug info with -fstandalone-debug. This can be
re-evaluated if there is a real need.
Differential Revision: https://reviews.llvm.org/D70696
The validateOutputSize and validateInputSize need to check whether
AVX or AVX512 are enabled. But this can be affected by the
target attribute so we need to factor that in.
This patch copies some of the code from CodeGen to create an
appropriate feature map that we can pass to the function. Probably
need some refactoring here to share more code with Codegen. Is
there a good place to do that? Also need to support the cpu_specific
attribute as well.
Differential Revision: https://reviews.llvm.org/D68627
Currently, it is a modified version of the Itanium ABI, with the only
change being that constructors and destructors return 'this'.
Differential Revision: https://reviews.llvm.org/D70575
Summary:
Currently, we ignore all locality attributes/info when building for
the device and thus all symblos are externally visible and can be
preemted at the runtime. It may lead to incorrect results. We need to
follow the same logic, compiler uses for static/pie builds. But in some
cases changing of dso locality may lead to problems with codegen, so
instead mark external symbols as hidden instead in the device code.
Reviewers: jdoerfert
Subscribers: guansong, caomhin, kkwli0, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D70549
Eli Friedman