This patch implements codegen for the reduction clause on
any parallel construct for elementary data types. An efficient
implementation requires hierarchical reduction within a
warp and a threadblock. It is complicated by the fact that
variables declared in the stack of a CUDA thread cannot be
shared with other threads.
The patch creates a struct to hold reduction variables and
a number of helper functions. The OpenMP runtime on the GPU
implements reduction algorithms that uses these helper
functions to perform reductions within a team. Variables are
shared between CUDA threads using shuffle intrinsics.
An implementation of reductions on the NVPTX device is
substantially different to that of CPUs. However, this patch
is written so that there are minimal changes to the rest of
OpenMP codegen.
The implemented design allows the compiler and runtime to be
decoupled, i.e., the runtime does not need to know of the
reduction operation(s), the type of the reduction variable(s),
or the number of reductions. The design also allows reuse of
host codegen, with appropriate specialization for the NVPTX
device.
While the patch does introduce a number of abstractions, the
expected use case calls for inlining of the GPU OpenMP runtime.
After inlining and optimizations in LLVM, these abstractions
are unwound and performance of OpenMP reductions is comparable
to CUDA-canonical code.
Patch by Tian Jin in collaboration with Arpith Jacob
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D29758
llvm-svn: 295333
SHF_LINK_ORDER sections adds special ordering requirements.
Such sections references other sections. Previously we would crash
if section that other were referenced to was discarded by script.
Patch fixes that by discarding all dependent sections in that case.
It supports chained dependencies, testcase is provided.
Differential revision: https://reviews.llvm.org/D30033
llvm-svn: 295332
Regression test neon-diagnostics.s needed changing because it now
produces a more specific diagnostic about the immediate ranges. One
change in the expected error message is not obvious, but there multiple
candidate and it happens to pick the immediate diagnostic.
Differential Revision: https://reviews.llvm.org/D29939
llvm-svn: 295331
On Windows, we were using `Sleep` which is not alertable. This means
that if the thread was used for a user APC or WinProc handling and
thread::sleep was used, we could potentially dead lock. Use `SleepEx`
with an alertable sleep, resuming until the time has expired if we are
awoken early.
llvm-svn: 295329
Unfortunately, the common way of writing linker scripts seems to be
to get the output of ld.bfd --verbose and edit it a bit.
Also unfortunately, the bfd default script contains things like
.rela.dyn : { *(... .rela.data ...) }
but bfd actually ignores that for -emit-relocs, so we have to do the
same.
llvm-svn: 295324
This patch implements codegen for the reduction clause on
any parallel construct for elementary data types. An efficient
implementation requires hierarchical reduction within a
warp and a threadblock. It is complicated by the fact that
variables declared in the stack of a CUDA thread cannot be
shared with other threads.
The patch creates a struct to hold reduction variables and
a number of helper functions. The OpenMP runtime on the GPU
implements reduction algorithms that uses these helper
functions to perform reductions within a team. Variables are
shared between CUDA threads using shuffle intrinsics.
An implementation of reductions on the NVPTX device is
substantially different to that of CPUs. However, this patch
is written so that there are minimal changes to the rest of
OpenMP codegen.
The implemented design allows the compiler and runtime to be
decoupled, i.e., the runtime does not need to know of the
reduction operation(s), the type of the reduction variable(s),
or the number of reductions. The design also allows reuse of
host codegen, with appropriate specialization for the NVPTX
device.
While the patch does introduce a number of abstractions, the
expected use case calls for inlining of the GPU OpenMP runtime.
After inlining and optimizations in LLVM, these abstractions
are unwound and performance of OpenMP reductions is comparable
to CUDA-canonical code.
Patch by Tian Jin in collaboration with Arpith Jacob
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D29758
llvm-svn: 295319
In D28836, we added a way to tag heap objects and thus provide object types into report. This patch exposes this information into the debugging API.
Differential Revision: https://reviews.llvm.org/D30023
llvm-svn: 295318
Support {a|s}ext, {a|z|s}ext load nodes as a part of load combine patters.
Reviewed By: filcab
Differential Revision: https://reviews.llvm.org/D29591
llvm-svn: 295314
Added description of a new feature that allows to specify
vendor extension in flexible way using compiler pragma instead
of modifying source code directly (committed in clang@r289979).
Review: D29829
llvm-svn: 295313
Removed ndrange_t as Clang builtin type and added
as a struct type in the OpenCL header.
Use type name to do the Sema checking in enqueue_kernel
and modify IR generation accordingly.
Review: D28058
Patch by Dmitry Borisenkov!
llvm-svn: 295311
Since they're only used for passing around double precision floating point
values into the general purpose registers, we'll lower them to VMOVDRR and
VMOVRRD.
llvm-svn: 295310
Summary:
This patch adds onTypeFormatting to clangd.
The trigger character is '}' and it works by scanning for the matching '{' and formatting the range in-between.
There are problems with ';' as a trigger character, the cursor position is before the `|`:
```
int main() {
int i;|
}
```
becomes:
```
int main() { int i;| }
```
which is not likely what the user intended.
Also formatting at semicolon in a non-properly closed scope puts the following tokens in the same unwrapped line, which doesn't reformat nicely.
Reviewers: bkramer
Reviewed By: bkramer
Subscribers: cfe-commits
Differential Revision: https://reviews.llvm.org/D29990
llvm-svn: 295304
Modules/preambles/PCH files can contain diagnostics, which, when used,
are added to the current ASTUnit. For that to work, they are translated
to use the current FileManager's FileIDs. When the entry is not the
main file, all local source locations will be checked by a linear
search. Now this is a problem, when there are lots of diagnostics (say,
25000) and lots of local source locations (say, 440000), and end up
taking seconds when using such a preamble.
The fix is to cache the last FileID, because many subsequent diagnostics
refer to the same file. This reduces the time spent in
ASTUnit::TranslateStoredDiagnostics from seconds to a few milliseconds
for files with many slocs/diagnostics.
This fixes PR31353.
Differential Revision: https://reviews.llvm.org/D29755
llvm-svn: 295301
For now we just mark them as legal all the time and let the other passes bail
out if they can't handle it. In the future, we'll want to move more of the
brains into the legalizer.
llvm-svn: 295300
That fixes a case when section has more than one metadata
section. Previously GC would collect one of such sections
because we had implementation that stored only last one as
dependent.
Differential revision: https://reviews.llvm.org/D29981
llvm-svn: 295298
For the hard float calling convention, we just use the D registers.
For the soft-fp calling convention, we use the R registers and move values
to/from the D registers by means of G_SEQUENCE/G_EXTRACT. While doing so, we
make sure to honor the endianness of the target, since the CCAssignFn doesn't do
that for us.
For pure soft float targets, we still bail out because we don't support the
libcalls yet.
llvm-svn: 295295
The new 512-bit unmasked intrinsics will make it easy to handle these with the SSE/AVX intrinsics in InstCombine where we currently have a TODO.
llvm-svn: 295290
This patch removes NeedsCopyOrPltAddr and instead add two variables,
NeedsCopy and NeedsPltAddr. This uses one more bit in Symbol class,
but the actual size doesn't increase because we had unused bits.
This should improve code readability.
llvm-svn: 295287
libunwind depends on C++ library headers. When building libunwind
as part of LLVM and libc++ is available, use its headers.
Differential Revision: https://reviews.llvm.org/D29997
llvm-svn: 295285
Recommit r293585 that was reverted in r293611 with new fixes. The previous
issue was determined to be an overly aggressive AST visitor from forward
declared objects. The visitor will now only deeply visit certain Decl's and
only do a shallow information extraction from all other Decl's.
When objects are imported for modules, there is a chance that a name collision
will cause an ODR violation. Previously, only a small number of such
violations were detected. This patch provides a stronger check based on
AST nodes.
The information needed to uniquely identify an object is taken from the AST and
put into a one-dimensional byte stream. This stream is then hashed to give
a value to represent the object, which is stored with the other object data
in the module.
When modules are loaded, and Decl's are merged, the hash values of the two
Decl's are compared. Only Decl's with matched hash values will be merged.
Mismatch hashes will generate a module error, and if possible, point to the
first difference between the two objects.
The transform from AST to byte stream is a modified depth first algorithm.
Due to references between some AST nodes, a pure depth first algorithm could
generate loops. For Stmt nodes, a straight depth first processing occurs.
For Type and Decl nodes, they are replaced with an index number and only on
first visit will these nodes be processed. As an optimization, boolean
values are saved and stored together in reverse order at the end of the
byte stream to lower the ammount of data that needs to be hashed.
Compile time impact was measured at 1.5-2.0% during module building, and
negligible during builds without module building.
Differential Revision: https://reviews.llvm.org/D21675
llvm-svn: 295284
Arpith Chacko Jacob