Firstly, each offloading entry must have a unique name or the
linker will complain if there are multiple files with target
regions. Secondly, the compiler must not introduce padding so
mark the struct with a PackedAttr.
Differential Revision: https://reviews.llvm.org/D42168
llvm-svn: 322858
simd`.
Added host codegen + codegen for devices with default codegen for
`#pragma omp target teams distribute parallel for simd` directive.
llvm-svn: 322515
getAssociatedStmt() returns the outermost captured statement for the
OpenMP directive. It may return incorrect region in case of combined
constructs. Reworked the code to reduce the number of calls of
getAssociatedStmt() and used getInnermostCapturedStmt() and
getCapturedStmt() functions instead.
In case of firstprivate variables it may lead to an extra allocas
generation for private copies even if the variable is passed by value
into outlined function and could be used directly as private copy.
llvm-svn: 322393
While updating clang tests for having clang set dso_local I noticed
that:
- There are *a lot* of tests to update.
- Many of the updates are redundant.
They are redundant because a GV is "obviously dso_local". This patch
starts formalizing that a bit by requiring that internal and private
GVs be dso_local too. Since they all are, we don't have to print
dso_local to the textual representation, making it a bit more compact
and easier to read.
llvm-svn: 322318
only.
Added support for -fopenmp-simd option that allows compilation of
simd-based constructs without emission of OpenMP runtime calls.
llvm-svn: 321560
The adjustment is calculated with CreatePtrDiff() which returns
the difference in (base) elements. This is passed to CreateGEP()
so make sure that the GEP base has the correct pointer type:
It needs to be a pointer to the base type, not a pointer to a
constant sized array.
Differential Revision: https://reviews.llvm.org/D40911
llvm-svn: 319931
Previously we emitted `__tgt_target_teams` only for standalone teams
directives. This patch allows emit this function for all teams-based
directives.
llvm-svn: 319585
This clang patch changes the __tgt_* API function signatures in preparation for the new map interface.
Changes are: Device IDs 32bits --> 64bits, Flags 32bits --> 64bits
Differential revision: https://reviews.llvm.org/D40281
llvm-svn: 318789
This patch renames some of the flag names of the clang/libomptarget map interface. The old names are slightly misleading, whereas the new ones describe in a better way what each flag is about.
Only the macros within the enumeration are renamed, there is no change in functionality therefore there are no updated regression tests.
Differential Revision: https://reviews.llvm.org/D39745
llvm-svn: 317598
If the thread id is requested in windows mode within funclets, we may
generate incorrect function call that could lead to broken codegen.
llvm-svn: 317208
This patch fixes various places in clang to propagate may-alias
TBAA access descriptors during construction of lvalues, thus
eliminating the need for the LValueBaseInfo::MayAlias flag.
This is part of D38126 reworked to be a separate patch to
simplify review.
Differential Revision: https://reviews.llvm.org/D39008
llvm-svn: 316988
In some cases the compiler can deduce the length of an array section
as constants. With this information, VLAs can be avoided in place of
a constant sized array or even a scalar value if the length is 1.
Example:
int a[4], b[2];
pragma omp parallel reduction(+: a[1:2], b[1:1])
{ }
For chained array sections, this optimization is restricted to cases
where all array sections except the last have a constant length 1.
This trivially guarantees that there are no holes in the memory region
that needs to be privatized.
Example:
int c[3][4];
pragma omp parallel reduction(+: c[1:1][1:2])
{ }
This relands commit r316229 that I reverted in r316235 because it
failed on some bots. During investigation I found that this was because
Clang and GCC evaluate the two arguments to emplace_back() in
ReductionCodeGen::emitSharedLValue() in a different order, hence
leading to a different order of generated instructions in the final
LLVM IR. Fix this by passing in the arguments from temporary variables
that are evaluated in a defined order.
Differential Revision: https://reviews.llvm.org/D39136
llvm-svn: 316362
In some cases the compiler can deduce the length of an array section
as constants. With this information, VLAs can be avoided in place of
a constant sized array or even a scalar value if the length is 1.
Example:
int a[4], b[2];
pragma omp parallel reduction(+: a[1:2], b[1:1])
{ }
For chained array sections, this optimization is restricted to cases
where all array sections except the last have a constant length 1.
This trivially guarantees that there are no holes in the memory region
that needs to be privatized.
Example:
int c[3][4];
pragma omp parallel reduction(+: c[1:1][1:2])
{ }
Differential Revision: https://reviews.llvm.org/D39136
llvm-svn: 316229
reduction.
If the reduction is an array or an array section and reduction operation
is declare reduction without initializer, it may lead to crash.
llvm-svn: 315611
in C.
If we try to get the lvalue for thread_id variables in inlined regions,
we did not use the correct version of function. Fixed this bug by adding
overrided version of the function getThreadIDVariableLValue for inlined
regions.
llvm-svn: 315578
This patch enables explicit generation of TBAA information in all
cases where LValue base info is propagated or constructed in
non-trivial ways. Eventually, we will consider each of these
cases to make sure the TBAA information is correct and not too
conservative. For now, we just fall back to generating TBAA info
from the access type.
This patch should not bring in any functional changes.
This is part of D38126 reworked to be a separate patch to
simplify review.
Differential Revision: https://reviews.llvm.org/D38733
llvm-svn: 315575
If both taskloop and task directives are used at the same time in one
program, we may ran into the situation when the particular type for task
directive is reused for taskloop directives. Patch fixes this problem.
llvm-svn: 315464
Besides obvious code simplification, avoiding explicit creation
of LValueBaseInfo objects makes it easier to make TBAA
information to be part of such objects.
This is part of D38126 reworked to be a separate patch to
simplify review.
Differential Revision: https://reviews.llvm.org/D38695
llvm-svn: 315289
Simplified and generalized codegen for non-offloading part that works if
offloading is failed or condition of the `if` clause is `false`.
llvm-svn: 314670
directives.
If the variable is used in the target-based region but is not found in
any private|mapping clause, then generate implicit firstprivate|map
clauses for these implicitly mapped variables.
llvm-svn: 314205
__kmpc_for_static_fini().
Added special flags for calls of __kmpc_for_static_fini(), like previous
ly for __kmpc_for_static_init(). Added flag OMP_IDENT_WORK_DISTRIBUTE
for distribute cnstruct, OMP_IDENT_WORK_SECTIONS for sections-based
constructs and OMP_IDENT_WORK_LOOP for loop-based constructs in
location flags.
llvm-svn: 312642
move constructor.
Previously user-defined reduction initializer was considered as an
assignment expression, not as initializer. Fixed this by treating the
initializer expression as an initializer.
llvm-svn: 312638
If exceptions are enabled, there may be a problem with the codegen of
the finalization functions from OpenMP runtime. It happens because of
the problem with the getting of thread identifier value. Patch tries to
fix it by using the result of the call of function
__kmpc_global_thread_num() rather than loading of value of outlined
function parameter.
llvm-svn: 311007
__kmpc_for_static_init().
OpenMP 5.0 will include OpenMP Tools interface that requires distinguishing different worksharing constructs.
Since the same entry point (__kmp_for_static_init(ident_t *loc,
kmp_int32 global_tid,........)) is called in case static
loop/sections/distribute it is suggested using 'flags' field of the
ident_t structure to pass the type of the construct.
llvm-svn: 310865
Arguments, passed to the outlined function, must have correct address
space info for proper Debug info support. Patch sets global address
space for arguments that are mapped and passed by reference.
Also, cuda-gdb does not handle reference types correctly, so reference
arguments are represented as pointers.
llvm-svn: 310387
If taskloop directive has no associated nogroup clause, it must emitted
inside implicit taskgroup block. Runtime supports it, but we need to
generate implicit taskgroup block explicitly to support future
reductions codegen.
llvm-svn: 307822
Currently, if the some of the parameters are captured by value, this
argument is converted to uintptr_t type and thus we loosing the debug
info about real type of the argument (captured variable):
```
void @.outlined_function.(uintptr %par);
...
%a = alloca i32
%a.casted = alloca uintptr
%cast = bitcast uintptr* %a.casted to i32*
%a.val = load i32, i32 *%a
store i32 %a.val, i32 *%cast
%a.casted.val = load uintptr, uintptr* %a.casted
call void @.outlined_function.(uintptr %a.casted.val)
...
```
To resolve this problem, in debug mode a speciall external wrapper
function is generated, that calls the outlined function with the correct
parameters types:
```
void @.wrapper.(uintptr %par) {
%a = alloca i32
%cast = bitcast i32* %a to uintptr*
store uintptr %par, uintptr *%cast
%a.val = load i32, i32* %a
call void @.outlined_function.(i32 %a)
ret void
}
void @.outlined_function.(i32 %par);
...
%a = alloca i32
%a.casted = alloca uintptr
%cast = bitcast uintptr* %a.casted to i32*
%a.val = load i32, i32 *%a
store i32 %a.val, i32 *%cast
%a.casted.val = load uintptr, uintptr* %a.casted
call void @.wrapper.(uintptr %a.casted.val)
...
```
llvm-svn: 306697
Summary:
If the first parameter of the function is the ImplicitParamDecl, codegen
automatically marks it as an implicit argument with `this` or `self`
pointer. Added internal kind of the ImplicitParamDecl to separate
'this', 'self', 'vtt' and other implicit parameters from other kind of
parameters.
Reviewers: rjmccall, aaron.ballman
Subscribers: cfe-commits
Differential Revision: https://reviews.llvm.org/D33735
llvm-svn: 305075
Amongst other, this will help LTO to correctly handle/honor files
compiled with O0, helping debugging failures.
It also seems in line with how we handle other options, like how
-fnoinline adds the appropriate attribute as well.
Differential Revision: https://reviews.llvm.org/D28404
llvm-svn: 304127
The functions creating LValues propagated information about alignment
source. Extend the propagated data to also include information about
possible unrestricted aliasing. A new class LValueBaseInfo will
contain both AlignmentSource and MayAlias info.
This patch should not introduce any functional changes.
Differential Revision: https://reviews.llvm.org/D33284
llvm-svn: 303358
Use variadic templates instead of relying on <cstdarg> + sentinel.
This enforces better type checking and makes code more readable.
Differential revision: https://reviews.llvm.org/D32550
llvm-svn: 302572
[OpenMP] Initial implementation of code generation for pragma 'distribute parallel for' on host
https://reviews.llvm.org/D29508
This patch makes the following additions:
It abstracts away loop bound generation code from procedures associated with pragma 'for' and loops in general, in such a way that the same procedures can be used for 'distribute parallel for' without the need for a full re-implementation.
It implements code generation for 'distribute parallel for' and adds regression tests. It includes tests for clauses.
It is important to notice that most of the clauses are implemented as part of existing procedures. For instance, firstprivate is already implemented for 'distribute' and 'for' as separate pragmas. As the implementation of 'distribute parallel for' is based on the same procedures, then we automatically obtain implementation for such clauses without the need to add new code. However, this requires regression tests that verify correctness of produced code.
llvm-svn: 301340
https://reviews.llvm.org/D29508
This patch makes the following additions:
1. It abstracts away loop bound generation code from procedures associated with pragma 'for' and loops in general, in such a way that the same procedures can be used for 'distribute parallel for' without the need for a full re-implementation.
2. It implements code generation for 'distribute parallel for' and adds regression tests. It includes tests for clauses.
It is important to notice that most of the clauses are implemented as part of existing procedures. For instance, firstprivate is already implemented for 'distribute' and 'for' as separate pragmas. As the implementation of 'distribute parallel for' is based on the same procedures, then we automatically obtain implementation for such clauses without the need to add new code. However, this requires regression tests that verify correctness of produced code.
Looking forward to comments.
llvm-svn: 301223
With tasks, the cancel may happen in another task. This has a different
region info which means that we can't find it here.
Differential Revision: https://reviews.llvm.org/D30091
llvm-svn: 295474
This resolves a deadlock with the cancel directive when there is no explicit
cancellation point. In that case, the implicit barrier acts as cancellation
point. After removing the barrier after cancel, the now unmatched barrier for
the explicit cancellation point has to go as well.
This has probably worked before rL255992: With the calls for the explicit
barrier, it was sure that all threads passed a barrier before exiting.
Reported by Simon Convent and Joachim Protze!
Differential Revision: https://reviews.llvm.org/D30088
llvm-svn: 295473
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
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
This patch adds support for codegen of 'target teams' on the host.
This combined directive has two captured statements, one for the
'teams' region, and the other for the 'parallel'.
This target teams region is offloaded using the __tgt_target_teams()
call. The patch sets the number of teams as an argument to
this call.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D29084
llvm-svn: 293005
This patch adds support for codegen of 'target teams' on the host.
This combined directive has two captured statements, one for the
'teams' region, and the other for the 'parallel'.
This target teams region is offloaded using the __tgt_target_teams()
call. The patch sets the number of teams as an argument to
this call.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D29084
llvm-svn: 293001
The num_threads-clause on the combined directive applies to the
'parallel' region of this construct. We modify the NumThreadsClause
class to capture the clause expression within the 'target' region.
The offload runtime call for 'target parallel' is changed to
__tgt_target_teams() with 1 team and the number of threads set by
this clause or a default if none.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D29082
llvm-svn: 292997
This patch adds support for codegen of 'target parallel' on the host.
It is also the first combined directive that requires two or more
captured statements. Support for this functionality is included in
the patch.
A combined directive such as 'target parallel' has two captured
statements, one for the 'target' and the other for the 'parallel'
region. Two captured statements are required because each has
different implicit parameters (see SemaOpenMP.cpp). For example,
the 'parallel' has 'global_tid' and 'bound_tid' while the 'target'
does not. The patch adds support for handling multiple captured
statements based on the combined directive.
When codegen'ing the 'target parallel' directive, the 'target'
outlined function is created using the outer captured statement
and the 'parallel' outlined function is created using the inner
captured statement.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D28753
llvm-svn: 292419
This patch adds support for codegen of 'target parallel' on the host.
It is also the first combined directive that requires two or more
captured statements. Support for this functionality is included in
the patch.
A combined directive such as 'target parallel' has two captured
statements, one for the 'target' and the other for the 'parallel'
region. Two captured statements are required because each has
different implicit parameters (see SemaOpenMP.cpp). For example,
the 'parallel' has 'global_tid' and 'bound_tid' while the 'target'
does not. The patch adds support for handling multiple captured
statements based on the combined directive.
When codegen'ing the 'target parallel' directive, the 'target'
outlined function is created using the outer captured statement
and the 'parallel' outlined function is created using the inner
captured statement.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D28753
llvm-svn: 292374
This patch refactors code that calls codegen for target regions. Currently
the codebase only supports the 'target' directive. The patch pulls out
common target processing code into a static function that can be called
by codegen for any target directive.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D28752
llvm-svn: 292134
Summary:
This patch introduces support for the execution of parallel constructs in a target
region on the NVPTX device. Parallel regions must be in the lexical scope of the
target directive.
The master thread in the master warp signals parallel work for worker threads in worker
warps on encountering a parallel region.
Note: The patch does not yet support capture of arguments in a parallel region so
the test cases are simple.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D28145
llvm-svn: 291565
Summary:
This patch adds two fields to the offload entry descriptor. One field is meant to signal Ctors/Dtors and `link` global variables, and the other is reserved for runtime library use.
Currently, these fields are only filled with zeros in the current code generation, but that will change when `declare target` is added.
The reason, we are adding these fields now is to make the code generation consistent with the runtime library proposal under review in https://reviews.llvm.org/D14031.
Reviewers: ABataev, hfinkel, carlo.bertolli, kkwli0, arpith-jacob, Hahnfeld
Subscribers: cfe-commits, caomhin, jholewinski
Differential Revision: https://reviews.llvm.org/D28298
llvm-svn: 291124
-fno-inline-functions, -O0, and optnone.
These were really, really tangled together:
- We used the noinline LLVM attribute for -fno-inline
- But not for -fno-inline-functions (breaking LTO)
- But we did use it for -finline-hint-functions (yay, LTO is happy!)
- But we didn't for -O0 (LTO is sad yet again...)
- We had weird structuring of CodeGenOpts with both an inlining
enumeration and a boolean. They interacted in weird ways and
needlessly.
- A *lot* of set smashing went on with setting these, and then got worse
when we considered optnone and other inlining-effecting attributes.
- A bunch of inline affecting attributes were managed in a completely
different place from -fno-inline.
- Even with -fno-inline we failed to put the LLVM noinline attribute
onto many generated function definitions because they didn't show up
as AST-level functions.
- If you passed -O0 but -finline-functions we would run the normal
inliner pass in LLVM despite it being in the O0 pipeline, which really
doesn't make much sense.
- Lastly, we used things like '-fno-inline' to manipulate the pass
pipeline which forced the pass pipeline to be much more
parameterizable than it really needs to be. Instead we can *just* use
the optimization level to select a pipeline and control the rest via
attributes.
Sadly, this causes a bunch of churn in tests because we don't run the
optimizer in the tests and check the contents of attribute sets. It
would be awesome if attribute sets were a bit more FileCheck friendly,
but oh well.
I think this is a significant improvement and should remove the semantic
need to change what inliner pass we run in order to comply with the
requested inlining semantics by relying completely on attributes. It
also cleans up tho optnone and related handling a bit.
One unfortunate aspect of this is that for generating alwaysinline
routines like those in OpenMP we end up removing noinline and then
adding alwaysinline. I tried a bunch of other approaches, but because we
recompute function attributes from scratch and don't have a declaration
here I couldn't find anything substantially cleaner than this.
Differential Revision: https://reviews.llvm.org/D28053
llvm-svn: 290398
The change in D26502 splits ReaderWriter.h, which contains the APIs
into both the BitReader and BitWriter libraries, into BitcodeReader.h
and BitcodeWriter.h.
Change clang uses to the appropriate split header(s).
llvm-svn: 286567
It seems the compiler was getting confused by the in-class initializers
in local struct MapInfo, so moving those to a default constructor
instead.
llvm-svn: 277256
