Different shared libraries contain different fat binary, which is stored in a global variable
__hip_gpubin_handle. Since different compilation units share the same fat binary, this
variable has linkonce linkage. However, it should not be merged across different shared
libraries.
This patch set the visibility of the global variable to be hidden, which will make it invisible
in the shared library, therefore preventing it from being merged.
Differential Revision: https://reviews.llvm.org/D50596
llvm-svn: 340056
The way address space declarations for builtins currently work
is nearly useless. The code assumes the address spaces used for
builtins is a confusingly named "target address space" from user
code using __attribute__((address_space(N))) that matches
the builtin declaration. There's no way to use this to declare
a builtin that returns a language specific address space.
The terminology used is highly cofusing since it has nothing
to do with the the address space selected by the target to use
for a language address space.
This feature is essentially unused as-is. AMDGPU and NVPTX
are the only in-tree targets attempting to use this. The AMDGPU
builtins certainly do not behave as intended (i.e. all of the
builtins returning pointers can never compile because the numbered
address space never matches the expected named address space).
The NVPTX builtins are missing tests for some, and the others
seem to rely on an implicit addrspacecast.
Change the used address space for builtins based on a target
hook to allow using a language address space for a builtin.
This allows the same builtin declaration to be used for multiple
languages with similarly purposed address spaces (e.g. the same
AMDGPU builtin can be used in OpenCL and CUDA even though the
constant address spaces are arbitarily different).
This breaks the possibility of using arbitrary numbered
address spaces alongside the named address spaces for builtins.
If this is an issue we probably need to introduce another builtin
declaration character to distinguish language address spaces from
so-called "target address spaces".
llvm-svn: 338707
CUDA 8.0 E.3.9.4 says: Within the body of a __device__ or __global__
function, only __shared__ variables or variables without any device
memory qualifiers may be declared with static storage class.
It is unclear how a function-scope non-const static variable
without device memory qualifier is implemented, therefore only static
const variable without device memory qualifier is allowed, which
can be emitted as a global variable in constant address space.
Currently clang only allows function-scope static variable with
__shared__ qualifier.
This patch also allows function-scope static const variable without
device memory qualifier and emits it as a global variable in constant
address space.
Differential Revision: https://reviews.llvm.org/D49931
llvm-svn: 338188
HIP generates one fat binary for all devices after linking. However, for each compilation
unit a ctor function is emitted which register the same fat binary. Measures need to be
taken to make sure the fat binary is only registered once.
Currently each ctor function calls __hipRegisterFatBinary and stores the returned value
to __hip_gpubin_handle. This patch changes the linkage of __hip_gpubin_handle to be linkonce
so that they are shared between LLVM modules. Then this patch adds check of value of
__hip_gpubin_handle to make sure __hipRegisterFatBinary is only called once. The code
is equivalent to
void *_gpubin_handle;
void ctor() {
if (__hip_gpubin_handle == 0) {
__hip_gpubin_handle = __hipRegisterFatBinary(...);
}
// register kernels and variables.
}
The patch also does similar change to dtors so that __hipUnregisterFatBinary
is called once.
Differential Revision: https://reviews.llvm.org/D49083
llvm-svn: 337631
This matches the way NVCC does it. Doing module cleanup at global
destructor phase used to work, but is, apparently, too late for
the CUDA runtime in CUDA-9.2, which ends up crashing with double-free.
Differential Revision: https://reviews.llvm.org/D48613
llvm-svn: 335763
There are HIP applications e.g. Tensorflow 1.3 using amdgpu kernel attributes, however
currently they are only allowed on OpenCL kernel functions.
This patch will allow amdgpu kernel attributes to be applied to CUDA/HIP __global__
functions.
Differential Revision: https://reviews.llvm.org/D47958
llvm-svn: 334561
Currently clang set kernel calling convention for CUDA/HIP after
arranging function, which causes incorrect kernel function type since
it depends on calling convention.
This patch moves setting kernel convention before arranging
function.
Differential Revision: https://reviews.llvm.org/D47733
llvm-svn: 334457
CGM.GetAddrOfConstantCString() sets the adress of the created GlobalValue
to unnamed. When emitting the object file LLVM will mark the surrounding
section as SHF_MERGE iff the string is nul-terminated and contains no
other nuls (see IsNullTerminatedString). This results in problems when
saving temporaries because LLVM doesn't set an EntrySize, so reading in
the serialized assembly file fails.
This never happened for the GPU binaries because they usually contain
a nul-character somewhere. Instead this only affected the module ID
when compiling relocatable device code.
However, this points to a potentially larger problem: If we put a
constant string into a named section, we really want the data to end
up in that section in the object file. To avoid LLVM merging sections
this patch unmarks the GlobalVariable's address as unnamed which also
fixes the problem of invalid serialized assembly files when saving
temporaries.
Differential Revision: https://reviews.llvm.org/D47902
llvm-svn: 334281
CUDA/HIP does not support RTTI on device side, therefore there
is no point of emitting type info when compiling for device.
Emitting type info for device not only clutters the IR with useless
global variables, but also causes undefined symbol at linking
since vtable for cxxabiv1::class_type_info has external linkage.
Differential Revision: https://reviews.llvm.org/D47694
llvm-svn: 334021
To support linking device code in different source files, it is necessary to
embed fat binary at host linking stage.
This patch emits an external symbol for fat binary in host codegen, then
embed the fat binary by lld through a linker script.
Differential Revision: https://reviews.llvm.org/D46472
llvm-svn: 332724
HIP is a language similar to CUDA (https://github.com/ROCm-Developer-Tools/HIP/blob/master/docs/markdown/hip_kernel_language.md ).
The language syntax is very similar, which allows a hip program to be compiled as a CUDA program by Clang. The main difference
is the host API. HIP has a set of vendor neutral host API which can be implemented on different platforms. Currently there is open source
implementation of HIP runtime on amdgpu target (https://github.com/ROCm-Developer-Tools/HIP).
This patch adds support of input kind and language standard hip.
When hip file is compiled, both LangOpts.CUDA and LangOpts.HIP is turned on. This allows compilation of hip program as CUDA
in most cases and only special handling of hip program is needed LangOpts.HIP is checked.
This patch also adds support of kernel launching of HIP program using HIP host API.
When -x hip is not specified, there is no behaviour change for CUDA.
Patch by Greg Rodgers.
Revised and lit test added by Yaxun Liu.
Differential Revision: https://reviews.llvm.org/D44984
llvm-svn: 330790
Some targets need special LLVM calling convention for CUDA kernel.
This patch does that through a TargetCodeGenInfo hook.
It only affects amdgcn target.
Patch by Greg Rodgers.
Revised and lit tests added by Yaxun Liu.
Differential Revision: https://reviews.llvm.org/D45223
llvm-svn: 330447
nvcc generates a unique registration function for each object file
that contains relocatable device code. Unique names are achieved
with a module id that is also reflected in the function's name.
Differential Revision: https://reviews.llvm.org/D42922
llvm-svn: 330425
It means the same thing as -mllvm; there isn't any reason to have two
options which do the same thing.
Differential Revision: https://reviews.llvm.org/D45109
llvm-svn: 329965
Found via codespell -q 3 -I ../clang-whitelist.txt
Where whitelist consists of:
archtype
cas
classs
checkk
compres
definit
frome
iff
inteval
ith
lod
methode
nd
optin
ot
pres
statics
te
thru
Patch by luzpaz! (This is a subset of D44188 that applies cleanly with a few
files that have dubious fixes reverted.)
Differential revision: https://reviews.llvm.org/D44188
llvm-svn: 329399
This reverts r328795 which introduced an issue with referencing __global__
function templates. More details in the original review D44747.
llvm-svn: 329099
CUDA shared variable should be initialized with undef.
Patch by Greg Rodgers.
Revised and lit test added by Yaxun Liu.
Differential Revision: https://reviews.llvm.org/D44985
llvm-svn: 328994
This patch sets target specific calling convention for CUDA kernels in IR.
Patch by Greg Rodgers.
Revised and lit test added by Yaxun Liu.
Differential Revision: https://reviews.llvm.org/D44747
llvm-svn: 328795
We should (almost) never consider a device-side declaration to match a
library builtin functio. Otherwise clang may ignore the implementation
provided by the CUDA headers and emit clang's idea of the builtin.
Differential Revision: https://reviews.llvm.org/D42319
llvm-svn: 323239
Re-commit r303463 now that LLVM is fixed and adjust some lit tests.
llvm::TargetLibraryInfo needs to know the size of wchar_t to work on
functions like `wcslen`. This patch changes clang to always emit the
wchar_size module flag (it would only do so for ARM previously).
This also adds an `assert()` to ensure the LLVM defaults based on the
target triple are in sync with clang.
Differential Revision: https://reviews.llvm.org/D32982
llvm-svn: 303478
FPContractModeKind is the codegen option flag which is already ternary (off,
on, fast). This makes it universally the type for the contractable info
across the front-end:
* In FPOptions (i.e. in the Sema + in the expression nodes).
* In LangOpts::DefaultFPContractMode which is the option that initializes
FPOptions in the Sema.
Another way to look at this change is that before fp-contractable on/off were
the only states handled to the front-end:
* For "on", FMA folding was performed by the front-end
* For "fast", we simply forwarded the flag to TargetOptions to handle it in
LLVM
Now off/on/fast are all exposed because for fast we will generate
fast-math-flags during CodeGen.
This is toward moving fp-contraction=fast from an LLVM TargetOption to a
FastMathFlag in order to fix PR25721.
---
This is a recommit of r299027 with an adjustment to the test
CodeGenCUDA/fp-contract.cu. The test assumed that even
though -ffp-contract=on is passed FE-based folding of FMA won't happen.
This is obviously wrong since the user is asking for this explicitly with the
option. CUDA is different that -ffp-contract=fast is on by default.
The test used to "work" because contract=fast and contract=on were maintained
separately and we didn't fold in the FE because contract=fast was on due to
the target-default. This patch consolidates the contract=on/fast/off state
into a ternary state hence the change in behavior.
---
Differential Revision: https://reviews.llvm.org/D31167
llvm-svn: 299033
Summary:
Now when you ask clang to link in a bitcode module, you can tell it to
set attributes on that module's functions to match what we would have
set if we'd emitted those functions ourselves.
This is particularly important for fast-math attributes in CUDA
compilations.
Each CUDA compilation links in libdevice, a bitcode library provided by
nvidia as part of the CUDA distribution. Without this patch, if we have
a user-function F that is compiled with -ffast-math that calls a
function G from libdevice, F will have the unsafe-fp-math=true (etc.)
attributes, but G will have no attributes.
Since F calls G, the inliner will merge G's attributes into F's. It
considers the lack of an unsafe-fp-math=true attribute on G to be
tantamount to unsafe-fp-math=false, so it "merges" these by setting
unsafe-fp-math=false on F.
This then continues up the call graph, until every function that
(transitively) calls something in libdevice gets unsafe-fp-math=false
set, thus disabling fastmath in almost all CUDA code.
Reviewers: echristo
Subscribers: hfinkel, llvm-commits, mehdi_amini
Differential Revision: https://reviews.llvm.org/D28538
llvm-svn: 293097
* __host__ __device__ functions are no longer considered to be
redeclarations of __host__ or __device__ functions. This prevents
unintentional merging of target attributes across them.
* Function target attributes are not considered (and must match) during
explicit instantiation and specialization of function templates.
Differential Revision: https://reviews.llvm.org/D25809
llvm-svn: 288962
Summary: This matches the idiom we use for our other CUDA wrapper headers.
Reviewers: tra
Subscribers: beanz, mgorny, cfe-commits
Differential Revision: https://reviews.llvm.org/D24978
llvm-svn: 283679
Summary:
This prevents clang from emitting 'invoke's and catch statements.
Things previously mostly worked thanks to TryToMarkNoThrow() in
CodeGenFunction. But this is not a proper IPO, and it doesn't properly
handle cases like mutual recursion.
Fixes bug 30593.
Reviewers: tra
Subscribers: cfe-commits
Differential Revision: https://reviews.llvm.org/D25166
llvm-svn: 283272
Summary:
We'd attempted to allow this, but turns out we were doing a very bad
job. :)
Making this work properly would be a giant change in clang. For
example, we'd need to make CXXRecordDecl::getDestructor()
context-sensitive, because the destructor you end up with depends on
where you're calling it from.
For now (and hopefully for ever), just disallow overloading of
destructors in CUDA.
Reviewers: rsmith
Subscribers: cfe-commits, tra
Differential Revision: https://reviews.llvm.org/D24571
llvm-svn: 283120
Summary:
Some function calls in CUDA are allowed to appear in
semantically-correct programs but are an error if they're ever
codegen'ed. Specifically, a host+device function may call a host
function, but it's an error if such a function is ever codegen'ed in
device mode (and vice versa).
Previously, clang made no attempt to catch these errors. For the most
part, they would be caught by ptxas, and reported as "call to unknown
function 'foo'".
Now we catch these errors and report them the same as we report other
illegal calls (e.g. a call from a host function to a device function).
This has a small change in error-message behavior for calls that were
previously disallowed (e.g. calls from a host to a device function).
Previously, we'd catch disallowed calls fairly early, before doing
additional semantic checking e.g. of the call's arguments. Now we catch
these illegal calls at the very end of our semantic checks, so we'll
only emit a "illegal CUDA call" error if the call is otherwise
well-formed.
Reviewers: tra, rnk
Subscribers: cfe-commits
Differential Revision: https://reviews.llvm.org/D23242
llvm-svn: 278759
This matches the way nvcc encapsulates GPU binaries into host object file.
Now cuobjdump can deal with clang-compiled object files.
Differential Revision: https://reviews.llvm.org/D23429
llvm-svn: 278549
Summary:
Before this patch, we computed the offsets in memory of args passed to
GPU kernel functions by throwing all of the args into an LLVM struct.
clang emits packed llvm structs basically whenever it feels like it, and
packed structs have alignment 1. So we cannot rely on the llvm type's
alignment matching the C++ type's alignment.
This patch fixes our codegen so we always respect the clang types'
alignments.
Reviewers: rnk
Subscribers: cfe-commits, tra
Differential Revision: https://reviews.llvm.org/D22879
llvm-svn: 276927
Summary:
This lets LLVM perform IPO over these functions. In particular, it
allows LLVM to emit ld.global.nc for loads to __restrict pointers in
kernels that are never written to.
Reviewers: rsmith
Subscribers: cfe-commits, tra
Differential Revision: http://reviews.llvm.org/D21337
llvm-svn: 274261
Summary:
This is particularly important because a some convergent CUDA intrinsics
(e.g. __shfl_down) are implemented in terms of inline asm.
Reviewers: tra
Subscribers: cfe-commits
Differential Revision: http://reviews.llvm.org/D20836
llvm-svn: 271336
Some people have weird CI systems that run each test subdirectory
independently without access to other parallel trees.
Unfortunately, this means we have to suffer some duplication until Art
can sort out how to share these types.
llvm-svn: 270164
According to Cuda Programming guide (v7.5, E2.3.1):
> __device__, __constant__ and __shared__ variables defined in namespace
> scope, that are of class type, cannot have a non-empty constructor or a
> non-empty destructor.
Clang already deals with device-side constructors (see D15305).
This patch enforces similar rules for destructors.
Differential Revision: http://reviews.llvm.org/D20140
llvm-svn: 270108
Codegen tests for device-side variable initialization are subset of test
cases used to verify Sema's part of the job.
Including CodeGenCUDA/device-var-init.cu from SemaCUDA makes it easier to
keep both sides in sync.
Differential Revision: http://reviews.llvm.org/D20139
llvm-svn: 270107
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
Allow only empty constructors for local __shared__ variables in a way
identical to restrictions imposed on dynamic initializers for global
variables on device.
Differential Revision: http://reviews.llvm.org/D20039
llvm-svn: 268982
According to CUDA programming guide (v7.5):
> E.2.9.4: Within the body of a device or global function, only
> shared variables may be declared with static storage class.
Differential Revision: http://reviews.llvm.org/D20034
llvm-svn: 268962
Matt Arsenault