D17779: host-side shadow variables of external declarations of device-side
global variables have internal linkage and are referenced by
`__cuda_register_globals`.
nvcc from CUDA 11 does not allow `__device__ inline` or `__device__ constexpr`
(C++17 inline variables) but clang has incorrectly supported them for a while:
```
error: A __device__ variable cannot be marked constexpr
error: An inline __device__/__constant__/__managed__ variable must have internal linkage when the program is compiled in whole program mode (-rdc=false)
```
If such a variable (which has a comdat group) is discarded (a copy from another
translation unit is prevailing and selected), accessing the variable from
outside the section group (`__cuda_register_globals`) is a violation of the ELF
specification and will be rejected by linkers:
> A symbol table entry with STB_LOCAL binding that is defined relative to one of a group's sections, and that is contained in a symbol table section that is not part of the group, must be discarded if the group members are discarded. References to this symbol table entry from outside the group are not allowed.
As a workaround, don't register such inline variables for now.
(If we register the variables in all TUs, we will keep multiple instances of the shadow and break the C++ semantics for inline variables).
We should reject such variables in Sema but our internal users need some time to migrate.
Reviewed By: tra
Differential Revision: https://reviews.llvm.org/D88786
To facilitate faster loading of device binaries and share them among processes,
HIP runtime favors their alignment being 4096 bytes. HIP runtime can load
unaligned device binaries, however, aligning them at 4096 bytes results in
faster loading and less shared memory usage.
This patch adds an option -bundle-align to clang-offload-bundler which allows
bundles to be aligned at specified alignment. By default it is 1, which is NFC
compared to existing format.
This patch then aligns embedded fat binary and device binary inside fat binary
at 4096 bytes.
It has been verified this change does not cause significant overall file size increase
for typical HIP applications (less than 1%).
Differential Revision: https://reviews.llvm.org/D88734
A static device variable may be accessed in host code through
cudaMemCpyFromSymbol etc. Currently clang does not
emit the static device variable if it is only referenced by
host code, which causes host code to fail at run time.
This patch fixes that.
Differential Revision: https://reviews.llvm.org/D88115
- After loading builtin bitcode for linking, skip adding default
function attributes on LLVM intrinsics as their attributes are
well-defined and retrieved directly from internal definitions. Adding
extra attributes on intrinsics results in inconsistent result when
`-save-temps` is present. Also, that makes few optimizations
conservative.
Differential Revision: https://reviews.llvm.org/D87761
Temporarily revert commit 04abbb3a78
due to regressions in some HIP apps due backend issues revealed by
this change.
Will re-commit it when backend issues are fixed.
- Skip generating profile data on `__global__` function in the host
compilation. It's a host-side stub function only and don't have
profile instrumentation generated on the real function body. The extra
profile data results in the malformed instrumentation profile data.
- Skip generating region mapping on functions in the wrong-side, i.e.,
+ For the device compilation, skip host-only functions; and,
+ For the host compilation, skip device-only functions (including
`__global__` functions.)
- As the device-side profiling is not ready yet, only host-side profile
code generation is checked.
Differential Revision: https://reviews.llvm.org/D85276
Add address space to indirect abi info and use it for kernels.
Previously, indirect arguments assumed assumed a stack passed object
in the alloca address space using byval. A stack pointer is unsuitable
for kernel arguments, which are passed in a separate, constant buffer
with a different address space.
Start using the new byref for aggregate kernel arguments. Previously
these were emitted as raw struct arguments, and turned into loads in
the backend. These will lower identically, although with byref you now
have the option of applying an explicit alignment. In the future, a
reasonable implementation would use byref for all kernel arguments
(this would be a practical problem at the moment due to losing things
like noalias on pointer arguments).
This is mostly to avoid fighting the optimizer's treatment of
aggregate load/store. SROA and instcombine both turn aggregate loads
and stores into a long sequence of element loads and stores, rather
than the optimizable memcpy I would expect in this situation. Now an
explicit memcpy will be introduced up-front which is better understood
and helps eliminate the alloca in more situations.
This skips using byref in the case where HIP kernel pointer arguments
in structs are promoted to global pointers. At minimum an additional
patch is needed to allow coercion with indirect arguments. This also
skips using it for OpenCL due to the current workaround used to
support kernels calling kernels. Distinct function bodies would need
to be generated up front instead of emitting an illegal call.
nvcc supports accessing file-scope static device variables in host code by host APIs
like cudaMemcpyToSymbol etc.
CUDA/HIP let users access device variables in host code by shadow variables. In host compilation,
clang emits a shadow variable for each device variable, and calls __*RegisterVariable to
register it in init function. The address of the shadow variable and the device side mangled
name of the device variable is passed to __*RegisterVariable. Runtime looks up the symbol
by name in the device binary to find the address of the device variable.
The problem with static device variables is that they have internal linkage, therefore their
name may be changed by the linker if there are multiple symbols with the same name. Also
they end up as local symbols in the elf file, whereas the runtime only looks up the global symbols.
Another reason for making the static device variables external linkage is that they may be
initialized externally by host code and their final value may be accessed by host code
after kernel execution, therefore they actually have external linkage. Giving them internal
linkage will cause incorrect optimizations on them.
To support accessing static device var in host code for -fno-gpu-rdc mode, change the intnernal
linkage to external linkage. The name does not need change since there is only one TU for
-fno-gpu-rdc mode. Also the externalization is done only if the device static var is referenced
by host code.
Differential Revision: https://reviews.llvm.org/D80858
This patch let lambda be host device by default and adds diagnostics for
capturing host variable by reference in device lambda.
Differential Revision: https://reviews.llvm.org/D78655
Summary:
- `ptrtoint` and `inttoptr` are defined as no-op casts if the integer
value as the same size as the pointer value. The pair of
`ptrtoint`/`inttoptr` is in fact a no-op cast sequence between
different address spaces. Teach `infer-address-spaces` to handle them
like a `bitcast`.
Reviewers: arsenm, chandlerc
Subscribers: jvesely, wdng, nhaehnle, hiraditya, kerbowa, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D81938
constexpr variables are compile time constants and implicitly const, therefore
they are safe to emit on both device and host side. Besides, in many cases
they are intended for both device and host, therefore it makes sense
to emit them on both device and host sides if necessary.
In most cases constexpr variables are used as rvalue and the variables
themselves do not need to be emitted. However if their address is taken,
then they need to be emitted.
For C++14, clang is able to handle that since clang emits them with
available_externally linkage together with the initializer.
However for C++17, the constexpr static data member of a class or template class
become inline variables implicitly. Therefore they become definitions with
linkonce_odr or weak_odr linkages. As such, they can not have available_externally
linkage.
This patch fixes that by adding implicit constant attribute to
file scope constexpr variables and constexpr static data members
in device compilation.
Differential Revision: https://reviews.llvm.org/D79237
Canonicalize on storing FP options in LangOptions instead of
redundantly in CodeGenOptions. Incorporate -ffast-math directly
into the values of those LangOptions rather than considering it
separately when building FPOptions. Build IR attributes from
those options rather than a mix of sources.
We should really simplify the driver/cc1 interaction here and have
the driver pass down options that cc1 directly honors. That can
happen in a follow-up, though.
Patch by Michele Scandale!
https://reviews.llvm.org/D80315
NoDebug attr does not totally eliminate debug info about a function when
inlining is enabled. This is inconsistent with when inlining is disabled.
This patch fixes that.
Differential Revision: https://reviews.llvm.org/D79967
If we're going to assume references are dereferenceable, we should also
assume they're aligned: otherwise, we can't actually dereference them.
See also D80072.
Differential Revision: https://reviews.llvm.org/D80166
The stub function is generated by compiler and its instructions have nothing
to do with the kernel source code.
Currently clang generates debug info for the stub function, which causes
confusion for the HIP debugger. For example, when users set break point
on a line of a kernel, the debugger should break on that line when the kernel is
executed and reaches that line, but instead the debugger breaks in the stub function.
This patch disables debug info for stub function for HIP.
Differential Revision: https://reviews.llvm.org/D79866
Summary:
- If the coerced type is still a pointer, it should be set with proper
parameter attributes, such as `noalias`, `nonnull`, and etc. Hoist
that (pointer) parameter attribute setting so that the coerced pointer
parameter could be marked properly.
Depends on D79394
Reviewers: rjmccall, kerbowa, yaxunl
Subscribers: jvesely, nhaehnle, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D79395
Move function emitDeferredDiags from Sema to DeferredDiagsEmitter since it
is only used by DeferredDiagsEmitter.
Also skip visited functions to avoid exponential compile time.
Differential Revision: https://reviews.llvm.org/D77028
Summary:
- `RegisterVar` has `void` return type and `size_t` in its variable size
parameter in HIP or CUDA 9.0+.
Reviewers: tra, yaxunl
Subscribers: cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D77398
The main purpose of introducing these builtins is to add a range
metadata [1, 1025) on the work group size loaded from dispatch
ptr, which cannot be done by source code.
Differential Revision: https://reviews.llvm.org/D76772
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
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
This reverts commit 737394c490.
The fp-model test was failing on platforms that enable denormal flushing
based on -ffast-math. This needs to reset to IEEE, not the default in
these cases.
Change-Id: Ibbad32f66d0d0b89b9c1173a3a96fb1a570ddd89
The IR hasn't switched the default yet, so explicitly add the ieee
attributes.
I'm still not really sure how the target default denormal mode should
interact with -fno-unsafe-math-optimizations. The target may have
selected the default mode to be non-IEEE based on the flags or based
on its true behavior, but we don't know which is the case. Since the
only users of a non-IEEE mode without a flag still support IEEE mode,
just reset to IEEE.
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
norecurse function attr indicates the function is not called recursively
directly or indirectly.
Add norecurse to OpenCL functions, SYCL functions in device compilation
and CUDA/HIP kernels.
Although there is LLVM pass adding norecurse to functions, it only works
for whole-program compilation. Also FE adding norecurse can make that
pass run faster since functions with norecurse do not need to be checked
again.
Differential Revision: https://reviews.llvm.org/D73651
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.
The CUDA builtin library is apparently compiled in C++ mode, so the
assumption of convergent needs to be made in a typically non-SPMD
language. The functions in the library should still be assumed
convergent. Currently they are not, which is potentially incorrect and
this happens to work after the library is linked.
Summary:
- Fix a bug which misses the change for a variable to be set with
target-specific attributes.
Reviewers: yaxunl
Subscribers: jvesely, nhaehnle, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D63020
The linker options (e.g. pragma detect_mismatch) are intended for host
compilation only, therefore disable it for device compilation.
Differential Revision: https://reviews.llvm.org/D57829
Add this option to allow device side class type global variables
with non-trivial ctor/dtor. device side init/fini functions will
be emitted, which will be executed by HIP runtime when
the fat binary is loaded/unloaded.
This feature is to facilitate implementation of device side
sanitizer which requires global vars with non-trival ctors.
By default this option is disabled.
Differential Revision: https://reviews.llvm.org/D69268
Summary:
- HIP/CUDA host side needs to use device kernel symbol name to match the
device side binaries. Without a consistent naming between host- and
device-side compilations, it's risky that wrong device binaries are
executed. Consistent naming is usually not an issue until unnamed
types are used, especially the lambda. In this patch, the consistent
name mangling is addressed for the extended lambdas, i.e. the lambdas
annotated with `__device__`.
- In [Itanium C++ ABI][1], the mangling of the lambda is generally
unspecified unless, in certain cases, ODR rule is required to ensure
consisent naming cross TUs. The extended lambda is such a case as its
name may be part of a device kernel function, e.g., the extended
lambda is used as a template argument and etc. Thus, we need to force
ODR for extended lambdas as they are referenced in both device- and
host-side TUs. Furthermore, if a extended lambda is nested in other
(extended or not) lambdas, those lambdas are required to follow ODR
naming as well. This patch revises the current lambda mangle numbering
to force ODR from an extended lambda to all its parent lambdas.
- On the other side, the aforementioned ODR naming should not change
those lambdas' original linkages, i.e., we cannot replace the original
`internal` with `linkonce_odr`; otherwise, we may violate ODR in
general. This patch introduces a new field `HasKnownInternalLinkage`
in lambda data to decouple the current linkage calculation based on
mangling number assigned.
[1]: https://itanium-cxx-abi.github.io/cxx-abi/abi.html
Reviewers: tra, rsmith, yaxunl, martong, shafik
Subscribers: cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D68818
llvm-svn: 375309
Fangrui Song