Summary:
Unlike other outlined regions in OpenMP, offloading entry points have to have be visible (external linkage) for the device side. Using dots in the names of the entries can be therefore problematic for some toolchains, e.g. NVPTX.
Also the patch drops the column information in the unique name of the entry points. The parsing of directives ignore unknown tokens, preventing several target regions to be implemented in the same line. Therefore, the line information is sufficient for the name to be unique. Also, the preprocessor printer does not preserve the column information, causing offloading-entry detection issues if the host uses an integrated preprocessor and the target doesn't (or vice versa).
Reviewers: hfinkel, arpith-jacob, carlo.bertolli, kkwli0, ABataev
Subscribers: cfe-commits, fraggamuffin, caomhin
Differential Revision: http://reviews.llvm.org/D17179
llvm-svn: 260837
Codegen for array sections/array subscripts worked only for expressions with arrays as base. Patch fixes codegen for bases with pointer/reference types.
llvm-svn: 259776
If 'sections' directive has only one sub-section, the code for 'single'-based directive was emitted. Removed this codegen, because it causes crashes in different cases.
llvm-svn: 258495
This patch attempts to fix the regressions identified when the patch was committed initially.
Thanks to Michael Liao for identifying the fix in the offloading metadata generation
related with side effects in evaluation of function arguments.
llvm-svn: 256933
Summary:
In order to offloading work properly two things need to be in place:
- a descriptor with all the offloading information (device entry functions, and global variable) has to be created by the host and registered in the OpenMP offloading runtime library.
- all the device functions need to be emitted for the device and a convention has to be in place so that the runtime library can easily map the host ID of an entry point with the actual function in the device.
This patch adds support for these two things. However, only entry functions are being registered given that 'declare target' directive is not yet implemented.
About offloading descriptor:
The details of the descriptor are explained with more detail in http://goo.gl/L1rnKJ. Basically the descriptor will have fields that specify the number of devices, the pointers to where the device images begin and end (that will be defined by the linker), and also pointers to a the begin and end of table whose entries contain information about a specific entry point. Each entry has the type:
```
struct __tgt_offload_entry{
void *addr;
char *name;
int64_t size;
};
```
and will be implemented in a pre determined (ELF) section `.omp_offloading.entries` with 1-byte alignment, so that when all the objects are linked, the table is in that section with no padding in between entries (will be like a C array). The code generation ensures that all `__tgt_offload_entry` entries are emitted in the same order for both host and device so that the runtime can have the corresponding entries in both host and device in same index of the table, and efficiently implement the mapping.
The resulting descriptor is registered/unregistered with the runtime library using the calls `__tgt_register_lib` and `__tgt_unregister_lib`. The registration is implemented in a high priority global initializer so that the registration happens always before any initializer (that can potentially include target regions) is run.
The driver flag -omptargets= was created to specify a comma separated list of devices the user wants to support so that the new functionality can be exercised. Each device is specified with its triple.
About target codegen:
The target codegen is pretty much straightforward as it reuses completely the logic of the host version for the same target region. The tricky part is to identify the meaningful target regions in the device side. Unlike other programming models, like CUDA, there are no already outlined functions with attributes that mark what should be emitted or not. So, the information on what to emit is passed in the form of metadata in host bc file. This requires a new option to pass the host bc to the device frontend. Then everything is similar to what happens in CUDA: the global declarations emission is intercepted to check to see if it is an "interesting" declaration. The difference is that instead of checking an attribute, the metadata information in checked. Right now, there is only a form of metadata to pass information about the device entry points (target regions). A class `OffloadEntriesInfoManagerTy` was created to manage all the information and queries related with the metadata. The metadata looks like this:
```
!omp_offload.info = !{!0, !1, !2, !3, !4, !5, !6}
!0 = !{i32 0, i32 52, i32 77426347, !"_ZN2S12r1Ei", i32 479, i32 13, i32 4}
!1 = !{i32 0, i32 52, i32 77426347, !"_ZL7fstatici", i32 461, i32 11, i32 5}
!2 = !{i32 0, i32 52, i32 77426347, !"_Z9ftemplateIiET_i", i32 444, i32 11, i32 6}
!3 = !{i32 0, i32 52, i32 77426347, !"_Z3fooi", i32 99, i32 11, i32 0}
!4 = !{i32 0, i32 52, i32 77426347, !"_Z3fooi", i32 272, i32 11, i32 3}
!5 = !{i32 0, i32 52, i32 77426347, !"_Z3fooi", i32 127, i32 11, i32 1}
!6 = !{i32 0, i32 52, i32 77426347, !"_Z3fooi", i32 159, i32 11, i32 2}
```
The fields in each metadata entry are (in sequence):
Entry 1) an ID of the type of metadata - right now only zero is used meaning "OpenMP target region".
Entry 2) a unique ID of the device where the input source file that contain the target region lives.
Entry 3) a unique ID of the file where the input source file that contain the target region lives.
Entry 4) a mangled name of the function that encloses the target region.
Entries 5) and 6) line and column number where the target region was found.
Entry 7) is the order the entry was emitted.
Entry 2) and 3) are required to distinguish files that have the same function name.
Entry 4) is required to distinguish different instances of the same declaration (usually templated ones)
Entries 5) and 6) are required to distinguish the particular target region in body of the function (it is possible that a given target region is not an entry point - if clause can evaluate always to zero - and therefore we need to identify the "interesting" target regions. )
This patch replaces http://reviews.llvm.org/D12306.
Reviewers: ABataev, hfinkel, tra, rjmccall, sfantao
Subscribers: FBrygidyn, piotr.rak, Hahnfeld, cfe-commits
Differential Revision: http://reviews.llvm.org/D12614
llvm-svn: 256842
OpenMP 4.5 adds 'depend(sink:vec)' in 'ordered' directive for doacross loop synchronization. Patch adds parsing and semantic analysis for this clause.
llvm-svn: 256330
OpenMP 4.5 adds 'depend(sink:vec)' in 'ordered' directive for doacross loop synchronization. Patch adds parsing and semantic analysis for this clause.
llvm-svn: 256238
OpenMP codegen tried to emit the code for its constructs even if it was detected as a dead-code. Added checks to ensure that the code is emitted if the code is not dead.
llvm-svn: 255990
OpenMP 4.5 adds 'depend(source)' clause for 'ordered' directive to support cross-iteration dependence. Patch adds parsing and semantic analysis for this construct.
llvm-svn: 255986
Summary:
This patch implements the 4.5 specification for the implicit data maps. OpenMP 4.5 specification changes the default way data is captured into a target region. All the non-aggregate kinds are passed by value by default. This required activating the capturing by value during SEMA for the target region. All the non-aggregate values that can be encoded in the size of a pointer are properly casted and forwarded to the runtime library. On top of fixing the previous weird behavior for mapping pointers in nested data regions (an explicit map was always required), this also improves performance as the number of allocations/transactions to the device per non-aggregate map are reduced from two to only one - instead of passing a reference and the value, only the value passed.
Explicit maps will be added later on once firstprivate, private, and map clauses' SEMA and parsing are available.
Reviewers: hfinkel, rjmccall, ABataev
Subscribers: cfe-commits, carlo.bertolli
Differential Revision: http://reviews.llvm.org/D14940
llvm-svn: 254521
Runtime library requires, that codegen for 'depend' clause for 'out' dependency kind must be the same as codegen for 'depend' clause with 'inout' dependency.
llvm-svn: 253866
attributes to internal functions.
This patch fixes CodeGenModule::CreateGlobalInitOrDestructFunction to
use SetInternalFunctionAttributes instead of SetLLVMFunctionAttributes
to attach function attributes to internal functions.
Also, make sure the correct CGFunctionInfo is passed instead of always
passing what arrangeNullaryFunction returns.
rdar://problem/20828324
Differential Revision: http://reviews.llvm.org/D13610
llvm-svn: 251734
functions.
This commit fixes a bug in CGOpenMPRuntime.cpp and CGObjC.cpp where
some of the function attributes are not attached to newly created
functions.
rdar://problem/20828324
Differential Revision: http://reviews.llvm.org/D13928
llvm-svn: 251476
This patch implements the outlining for offloading functions for code
annotated with the OpenMP target directive. It uses a temporary naming
of the outlined functions that will have to be updated later on once
target side codegen and registration of offloading libraries is
implemented - the naming needs to be made unique in the produced
library.
llvm-svn: 249148
Description.
If the simd clause is specified, the ordered regions encountered by any thread will use only a single SIMD lane to execute the ordered regions in the order of the loop iterations.
Restrictions.
An ordered construct with the simd clause is the only OpenMP construct that can appear in the simd region.
An ordered directive with ‘simd’ clause is generated as an outlined function and corresponding function call to prevent this part of code from vectorization later in backend.
llvm-svn: 248772
Patch improves codegen for OpenMP constructs. If the OpenMP region does not have internal 'cancel' construct, a call to 'void __kmpc_barrier()' runtime function is generated for all implicit/explicit barriers. If the region has inner 'cancel' directive, then
```
if (__kmpc_cancel_barrier())
exit from outer construct;
```
code is generated.
Also, the code for 'canellation point' directive is not generated if parent directive does not have 'cancel' directive.
llvm-svn: 247681
Current implementation may end up emitting an undefined reference for
an "inline __attribute__((always_inline))" function by generating an
"available_externally alwaysinline" IR function for it and then failing to
inline all the calls. This happens when a call to such function is in dead
code. As the inliner is an SCC pass, it does not process dead code.
Libc++ relies on the compiler never emitting such undefined reference.
With this patch, we emit a pair of
1. internal alwaysinline definition (called F.alwaysinline)
2a. A stub F() { musttail call F.alwaysinline }
-- or, depending on the linkage --
2b. A declaration of F.
The frontend ensures that F.inlinefunction is only used for direct
calls, and the stub is used for everything else (taking the address of
the function, really). Declaration (2b) is emitted in the case when
"inline" is meant for inlining only (like __gnu_inline__ and some
other cases).
This approach, among other nice properties, ensures that alwaysinline
functions are always internal, making it impossible for a direct call
to such function to produce an undefined symbol reference.
This patch is based on ideas by Chandler Carruth and Richard Smith.
llvm-svn: 247494
Current implementation may end up emitting an undefined reference for
an "inline __attribute__((always_inline))" function by generating an
"available_externally alwaysinline" IR function for it and then failing to
inline all the calls. This happens when a call to such function is in dead
code. As the inliner is an SCC pass, it does not process dead code.
Libc++ relies on the compiler never emitting such undefined reference.
With this patch, we emit a pair of
1. internal alwaysinline definition (called F.alwaysinline)
2a. A stub F() { musttail call F.alwaysinline }
-- or, depending on the linkage --
2b. A declaration of F.
The frontend ensures that F.inlinefunction is only used for direct
calls, and the stub is used for everything else (taking the address of
the function, really). Declaration (2b) is emitted in the case when
"inline" is meant for inlining only (like __gnu_inline__ and some
other cases).
This approach, among other nice properties, ensures that alwaysinline
functions are always internal, making it impossible for a direct call
to such function to produce an undefined symbol reference.
This patch is based on ideas by Chandler Carruth and Richard Smith.
llvm-svn: 247465
Currently all variables used in OpenMP regions are captured into a record and passed to outlined functions in this record. It may result in some poor performance because of too complex analysis later in optimization passes. Patch makes to emit outlined functions for parallel-based regions with a list of captured variables. It reduces code for 2*n GEPs, stores and loads at least.
Codegen for task-based regions remains unchanged because runtime requires that all captured variables are passed in captured record.
llvm-svn: 247251
Seems it broke the Polly build.
From http://lab.llvm.org:8011/builders/perf-x86_64-penryn-O3-polly-fast/builds/11687/steps/compile/logs/stdio:
In file included from /home/grosser/buildslave/perf-x86_64-penryn-O3-polly-fast/llvm.src/lib/TableGen/Record.cpp:14:0:
/home/grosser/buildslave/perf-x86_64-penryn-O3-polly-fast/llvm.src/include/llvm/TableGen/Record.h:369:3: error: looser throw specifier for 'virtual llvm::TypedInit::~TypedInit()'
/home/grosser/buildslave/perf-x86_64-penryn-O3-polly-fast/llvm.src/include/llvm/TableGen/Record.h:270:11: error: overriding 'virtual llvm::Init::~Init() noexcept (true)'
llvm-svn: 247222
Introduce an Address type to bundle a pointer value with an
alignment. Introduce APIs on CGBuilderTy to work with Address
values. Change core APIs on CGF/CGM to traffic in Address where
appropriate. Require alignments to be non-zero. Update a ton
of code to compute and propagate alignment information.
As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment
helper function to CGF and made use of it in a number of places in
the expression emitter.
The end result is that we should now be significantly more correct
when performing operations on objects that are locally known to
be under-aligned. Since alignment is not reliably tracked in the
type system, there are inherent limits to this, but at least we
are no longer confused by standard operations like derived-to-base
conversions and array-to-pointer decay. I've also fixed a large
number of bugs where we were applying the complete-object alignment
to a pointer instead of the non-virtual alignment, although most of
these were hidden by the very conservative approach we took with
member alignment.
Also, because IRGen now reliably asserts on zero alignments, we
should no longer be subject to an absurd but frustrating recurring
bug where an incomplete type would report a zero alignment and then
we'd naively do a alignmentAtOffset on it and emit code using an
alignment equal to the largest power-of-two factor of the offset.
We should also now be emitting much more aggressive alignment
attributes in the presence of over-alignment. In particular,
field access now uses alignmentAtOffset instead of min.
Several times in this patch, I had to change the existing
code-generation pattern in order to more effectively use
the Address APIs. For the most part, this seems to be a strict
improvement, like doing pointer arithmetic with GEPs instead of
ptrtoint. That said, I've tried very hard to not change semantics,
but it is likely that I've failed in a few places, for which I
apologize.
ABIArgInfo now always carries the assumed alignment of indirect and
indirect byval arguments. In order to cut down on what was already
a dauntingly large patch, I changed the code to never set align
attributes in the IR on non-byval indirect arguments. That is,
we still generate code which assumes that indirect arguments have
the given alignment, but we don't express this information to the
backend except where it's semantically required (i.e. on byvals).
This is likely a minor regression for those targets that did provide
this information, but it'll be trivial to add it back in a later
patch.
I partially punted on applying this work to CGBuiltin. Please
do not add more uses of the CreateDefaultAligned{Load,Store}
APIs; they will be going away eventually.
llvm-svn: 246985
Added codegen for array section in 'depend' clause of 'task' directive. It emits to pointers, one for the begin of array section and another for the end of array section. Size of the section is calculated as (end + 1 - start) * sizeof(basic_element_type).
llvm-svn: 246422
Added codegen for array section in 'depend' clause of 'task' directive. It emits to pointers, one for the begin of array section and another for the end of array section. Size of the section is calculated as (end + 1 - start) * sizeof(basic_element_type).
llvm-svn: 246278
Summary:
float_cast_overflow is the only UBSan check without a source location attached.
This patch propagates SourceLocations where necessary to get them to the
EmitCheck() call.
Reviewers: rsmith, ABataev, rjmccall
Subscribers: cfe-commits
Differential Revision: http://reviews.llvm.org/D11757
llvm-svn: 244568
Alexey Bataev