Currently clang assumes the temporary variables emitted during
codegen of atomic builtins have address space 0, which
is not true for target triple amdgcn---amdgiz and causes invalid
bitcasts.
This patch fixes that.
Differential Revision: https://reviews.llvm.org/D38966
llvm-svn: 316000
In OpenCL the kernel function and non-kernel function has different calling conventions.
For certain targets they have different argument ABIs. Also kernels have special function
attributes and metadata for runtime to launch them.
The blocks passed to enqueue_kernel is supposed to be executed as kernels. As such,
the block invoke function should be emitted as kernel with proper calling convention and
argument ABI.
This patch emits enqueued block as kernel. If a block is both called directly and passed
to enqueue_kernel, separate functions will be generated.
Differential Revision: https://reviews.llvm.org/D38134
llvm-svn: 315804
Currently Clang uses default address space (0) to represent private address space for OpenCL
in AST. There are two issues with this:
Multiple address spaces including private address space cannot be diagnosed.
There is no mangling for default address space. For example, if private int* is emitted as
i32 addrspace(5)* in IR. It is supposed to be mangled as PUAS5i but it is mangled as
Pi instead.
This patch attempts to represent OpenCL private address space explicitly in AST. It adds
a new enum LangAS::opencl_private and adds it to the variable types which are implicitly
private:
automatic variables without address space qualifier
function parameter
pointee type without address space qualifier (OpenCL 1.2 and below)
Differential Revision: https://reviews.llvm.org/D35082
llvm-svn: 315668
This was done for CUDA functions in r261779, and for the same
reason this also needs to be done for OpenCL. An arbitrary
function could have a barrier() call in it, which in turn
requires the calling function to be convergent.
llvm-svn: 315094
Currently block is translated to a structure equivalent to
struct Block {
void *isa;
int flags;
int reserved;
void *invoke;
void *descriptor;
};
Except invoke, which is the pointer to the block invoke function,
all other fields are useless for OpenCL, which clutter the IR and
also waste memory since the block struct is passed to the block
invoke function as argument.
On the other hand, the size and alignment of the block struct is
not stored in the struct, which causes difficulty to implement
__enqueue_kernel as library function, since the library function
needs to know the size and alignment of the argument which needs
to be passed to the kernel.
This patch removes the useless fields from the block struct and adds
size and align fields. The equivalent block struct will become
struct Block {
int size;
int align;
generic void *invoke;
/* custom fields */
};
It also changes the pointer to the invoke function to be
a generic pointer since the address space of a function
may not be private on certain targets.
Differential Revision: https://reviews.llvm.org/D37822
llvm-svn: 314932
Added missing addrspacecast case in alignment computation
logic of pointer type emission in IR generation.
Differential Revision: https://reviews.llvm.org/D37804
llvm-svn: 314304
Add tests for different address spaces and insert some blank lines to make them more readable.
Differential Revision: https://reviews.llvm.org/D37742
llvm-svn: 313172
Not all targets support vararg (e.g. amdgpu). Instead of using vararg in the emitted functions for enqueue_kernel,
this patch creates a temporary array of size_t, stores the size arguments in the temporary array
and passes it to the emitted functions for enqueue_kernel.
Differential Revision: https://reviews.llvm.org/D36678
llvm-svn: 312441
Summary:
Most DIExpressions are empty or very simple. When they are complex, they
tend to be unique, so checking them inline is reasonable.
This also avoids the need for CodeGen passes to append to the
llvm.dbg.mir named md node.
See also PR22780, for making DIExpression not be an MDNode.
Reviewers: aprantl, dexonsmith, dblaikie
Subscribers: qcolombet, javed.absar, eraman, hiraditya, llvm-commits
Differential Revision: https://reviews.llvm.org/D37075
llvm-svn: 311594
Generalize getOpenCLImageAddrSpace into getOpenCLTypeAddrSpace, such
that targets can select the address space per type.
No functional changes intended.
Initial patch by Simon Perretta.
Differential Revision: https://reviews.llvm.org/D33989
llvm-svn: 310911
This is an improvement over always using byval for
structs.
This will use registers until ~16 are used, and then
switch back to byval. This needs more work, since I'm
not sure it ever really makes sense to use byval. If
the register limit is exceeded, the arguments still
end up passed on the stack, but with a different ABI.
It also may make sense to base this on number of
registers used for non-struct arguments, rather than
just arguments that appear first in the argument list.
llvm-svn: 310527
OpenCL 2.0 atomic builtin functions have a scope argument which is ideally
represented as synchronization scope argument in LLVM atomic instructions.
Clang supports translating Clang atomic builtin functions to LLVM atomic
instructions. However it currently does not support synchronization scope
of LLVM atomic instructions. Without this, users have to use LLVM assembly
code to implement OpenCL atomic builtin functions.
This patch adds OpenCL 2.0 atomic builtin functions as Clang builtin
functions, which supports generating LLVM atomic instructions with
synchronization scope operand.
Currently only constant memory scope argument is supported. Support of
non-constant memory scope argument will be added later.
Differential Revision: https://reviews.llvm.org/D28691
llvm-svn: 310082
Currently AMDGPUTargetInfo does not initialize AddrSpaceMap in constructor, which causes regressions in mesa/clover with libclc.
This patch fixes that.
Differential Revision: https://reviews.llvm.org/D34987
llvm-svn: 307105
Clang assumes coerced function argument is in address space 0, which is not always true and results in invalid bitcasts.
This patch fixes failure in OpenCL conformance test api/get_kernel_arg_info with amdgcn---amdgizcl triple, where non-zero alloca address space is used.
Differential Revision: https://reviews.llvm.org/D34777
llvm-svn: 306721
Summary: OpenCL and SPIR version metadata must be generated once per module instead of once per mangled global value.
Reviewers: Anastasia, yaxunl
Reviewed By: Anastasia
Subscribers: ahatanak, cfe-commits
Differential Revision: https://reviews.llvm.org/D34235
llvm-svn: 305796
Rationale: OpenCL kernels are called via an explicit runtime API
with arguments set with clSetKernelArg(), not as normal sub-functions.
Return SPIR_KERNEL by default as the kernel calling convention to ensure
the fingerprint is fixed such way that each OpenCL argument gets one
matching argument in the produced kernel function argument list to enable
feasible implementation of clSetKernelArg() with aggregates etc. In case
we would use the default C calling conv here, clSetKernelArg() might
break depending on the target-specific conventions; different targets
might split structs passed as values to multiple function arguments etc.
https://reviews.llvm.org/D33639
llvm-svn: 304389
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
A recent change requires opencl triple environment for compiling OpenCL
program, which causes regressions in libclc.
This patch fixes that. Instead of deducing language based on triple
environment, it checks LangOptions.
Differential Revision: https://reviews.llvm.org/D33445
llvm-svn: 303644
Alloca always returns a pointer in alloca address space, which may
be different from the type defined by the language. For example,
in C++ the auto variables are in the default address space. Therefore
cast alloca to the expected address space when necessary.
Differential Revision: https://reviews.llvm.org/D32248
llvm-svn: 303370
LLVM has changed the semantics of dbg.declare for describing function
arguments. After this patch a dbg.declare always takes the *address*
of a variable as the first argument, even if the argument is not an
alloca.
https://bugs.llvm.org/show_bug.cgi?id=32382
rdar://problem/31205000
llvm-svn: 300523
For OpenCL, the private address space qualifier is 0 in AST. Before this change, 0 address space qualifier
is always mapped to target address space 0. As now target private address space is specified by
alloca address space in data layout, address space qualifier 0 needs to be mapped to alloca addr space specified by the data layout.
This change has no impact on targets whose alloca addr space is 0.
With contributions from Matt Arsenault, Tony Tye and Wen-Heng (Jack) Chung
Differential Revision: https://reviews.llvm.org/D31404
llvm-svn: 299965
Change constant address space from 4 to 2 for the new address space mapping in Clang.
Differential Revision: https://reviews.llvm.org/D31771
llvm-svn: 299691
These two attributes specify the same info in a different way.
AMGPU BE only checks the latter as a target specific attribute
as opposed to language specific reqd_work_group_size.
This change produces amdgpu_flat_work_group_size out of
reqd_work_group_size if specified.
Differential Revision: https://reviews.llvm.org/D31728
llvm-svn: 299678
Summary:
"kernel_arg_type_qual" metadata should contain const/volatile/restrict
tags only for pointer types to match the corresponding requirement of
the OpenCL specification.
OpenCL 2.0 spec 5.9.3 Kernel Object Queries:
CL_KERNEL_ARG_TYPE_VOLATILE is returned if the argument is a pointer
and the referenced type is declared with the volatile qualifier.
[...]
Similarly, CL_KERNEL_ARG_TYPE_CONST is returned if the argument is a
pointer and the referenced type is declared with the restrict or const
qualifier.
[...]
CL_KERNEL_ARG_TYPE_RESTRICT will be returned if the pointer type is
marked restrict.
Reviewers: Anastasia, cfe-commits
Reviewed By: Anastasia
Subscribers: bader, yaxunl
Differential Revision: https://reviews.llvm.org/D31321
llvm-svn: 299192
For target environment amdgiz and amdgizcl (giz means Generic Is Zero), AMDGPU will use new address space mapping where generic address space is 0 and private address space is 5. The data layout is also changed correspondingly.
Differential Revision: https://reviews.llvm.org/D31210
llvm-svn: 298767
For variables in generic address spaces, for example:
```
unsigned char V[6442450944];
...
```
the address space is not yet known when we get into
*getConstantArrayType*, it is 0. AMDGCN target's
address space 0 has 32 bits pointers, so when we
call *getPointerWidth* with 0, the array size is
trimmed to 32 bits, which is not right.
Differential Revision: https://reviews.llvm.org/D30845
llvm-svn: 298420
Summary: I added a new rank to ImplicitConversionRank enum to resolve the function overload ambiguity with vector types. Rank of scalar types conversion is lower than vector splat. So, we can choose which function should we call. See test for more details.
Reviewers: Anastasia, cfe-commits
Reviewed By: Anastasia
Subscribers: bader, yaxunl
Differential Revision: https://reviews.llvm.org/D30816
llvm-svn: 298366
We can't actually pretend that 0 is valid for address space 0.
r295877 added a workaround to stop allocating user objects
there, so we can use 0 as the invalid pointer.
Some of the tests seemed to be using private as the non-0 null
test address space, so add copies using local to make sure
this is still stressed.
llvm-svn: 297659
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
Modify ObjC blocks impl wrt address spaces as follows:
- keep default private address space for blocks generated
as local variables (with captures);
- add global address space for global block literals (no captures);
- make the block invoke function and enqueue_kernel prototype with
the generic AS block pointer parameter to accommodate both
private and global AS cases from above;
- add block handling into default AS because it's implemented as
a special pointer type (BlockPointer) in the frontend and therefore
it is used as a pointer everywhere. This is also needed to accommodate
both private and global AS blocks for the two cases above.
- removes ObjC RT specific symbols (NSConcreteStackBlock and
NSConcreteGlobalBlock) in the OpenCL mode.
Review: https://reviews.llvm.org/D28814
llvm-svn: 293286
Summary:
We compile user opencl kernel code with spir triple. But built-ins are written in OpenCL and we compile it with triple x86_64 to be able to use x86 intrinsics. And we need address spaces to match in both cases. So, we change fake address space map in OpenCL for matching with spir.
On CPU address spaces are not really important but we'd like to preserve address space information in order to perform optimizations relying on this info like enhanced alias analysis.
Reviewers: pekka.jaaskelainen, Anastasia
Subscribers: pekka.jaaskelainen, yaxunl, bader, cfe-commits
Differential Revision: https://reviews.llvm.org/D28048
llvm-svn: 290436
-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
This is a recommit of r290149, which was reverted in r290169 due to msan
failures. msan was failing because we were calling
`isMostDerivedAnUnsizedArray` on an invalid designator, which caused us
to read uninitialized memory. To fix this, the logic of the caller of
said function was simplified, and we now have a `!Invalid` assert in
`isMostDerivedAnUnsizedArray`, so we can catch this particular bug more
easily in the future.
Fingers crossed that this patch sticks this time. :)
Original commit message:
This patch does three things:
- Gives us the alloc_size attribute in clang, which lets us infer the
number of bytes handed back to us by malloc/realloc/calloc/any user
functions that act in a similar manner.
- Teaches our constexpr evaluator that evaluating some `const` variables
is OK sometimes. This is why we have a change in
test/SemaCXX/constant-expression-cxx11.cpp and other seemingly
unrelated tests. Richard Smith okay'ed this idea some time ago in
person.
- Uniques some Blocks in CodeGen, which was reviewed separately at
D26410. Lack of uniquing only really shows up as a problem when
combined with our new eagerness in the face of const.
llvm-svn: 290297
This reverts commit r290171. It triggers a bunch of warnings, because
the new enumerator isn't handled in all switches. We want a warning-free
build.
Replied on the commit with more details.
llvm-svn: 290173
Summary: Enabling the compression of CLK_NULL_QUEUE to variable of type queue_t.
Reviewers: Anastasia
Subscribers: cfe-commits, yaxunl, bader
Differential Revision: https://reviews.llvm.org/D27569
llvm-svn: 290171
This commit fails MSan when running test/CodeGen/object-size.c in
a confusing way. After some discussion with George, it isn't really
clear what is going on here. We can make the MSan failure go away by
testing for the invalid bit, but *why* things are invalid isn't clear.
And yet, other code in the surrounding area is doing precisely this and
testing for invalid.
George is going to take a closer look at this to better understand the
nature of the failure and recommit it, for now backing it out to clean
up MSan builds.
llvm-svn: 290169
This patch does three things:
- Gives us the alloc_size attribute in clang, which lets us infer the
number of bytes handed back to us by malloc/realloc/calloc/any user
functions that act in a similar manner.
- Teaches our constexpr evaluator that evaluating some `const` variables
is OK sometimes. This is why we have a change in
test/SemaCXX/constant-expression-cxx11.cpp and other seemingly
unrelated tests. Richard Smith okay'ed this idea some time ago in
person.
- Uniques some Blocks in CodeGen, which was reviewed separately at
D26410. Lack of uniquing only really shows up as a problem when
combined with our new eagerness in the face of const.
Differential Revision: https://reviews.llvm.org/D14274
llvm-svn: 290149
Added a map to associate types and declarations with extensions.
Refactored existing diagnostic for disabled types associated with extensions and extended it to declarations for generic situation.
Fixed some bugs for types associated with extensions.
Allow users to use pragma to declare types and functions for supported extensions, e.g.
#pragma OPENCL EXTENSION the_new_extension_name : begin
// declare types and functions associated with the extension here
#pragma OPENCL EXTENSION the_new_extension_name : end
Differential Revision: https://reviews.llvm.org/D21698
llvm-svn: 289979
This change makes sure single-precision floating point types are used if the
cl_fp64 extension is not supported by the target.
Also removed the check to see whether the OpenCL version is >= 1.2, as this has
been incorporated into the extension setting code.
Differential Revision: https://reviews.llvm.org/D24235
llvm-svn: 289544
Summary: Although the feature was introduced only in OpenCL C v2.0 spec., it's useful for OpenCL 1.x too and doesn't require HW support.
Reviewers: Anastasia
Subscribers: yaxunl, cfe-commits, bader
Differential Revision: https://reviews.llvm.org/D27453
llvm-svn: 289535
In amdgcn target, null pointers in global, constant, and generic address space take value 0 but null pointers in private and local address space take value -1. Currently LLVM assumes all null pointers take value 0, which results in incorrectly translated IR. To workaround this issue, instead of emit null pointers in local and private address space, a null pointer in generic address space is emitted and casted to local and private address space.
Tentative definition of global variables with non-zero initializer will have weak linkage instead of common linkage since common linkage requires zero initializer and does not have explicit section to hold the non-zero value.
Virtual member functions getNullPointer and performAddrSpaceCast are added to TargetCodeGenInfo which by default returns ConstantPointerNull and emitting addrspacecast instruction. A virtual member function getNullPointerValue is added to TargetInfo which by default returns 0. Each target can override these virtual functions to get target specific null pointer and the null pointer value for specific address space, and perform specific translations for addrspacecast.
Wrapper functions getNullPointer is added to CodegenModule and getTargetNullPointerValue is added to ASTContext to facilitate getting the target specific null pointers and their values.
This change has no effect on other targets except amdgcn target. Other targets can provide support of non-zero null pointer in a similar way.
This change only provides support for non-zero null pointer for C and OpenCL. Supporting for other languages will be added later incrementally.
Differential Revision: https://reviews.llvm.org/D26196
llvm-svn: 289252
Avoid using shortcut for const qualified non-constant address space
aggregate variables while generating them on the stack such that
the alloca object is used instead of a global variable containing
initializer.
Review: https://reviews.llvm.org/D27109
llvm-svn: 288163
The wave barrier represents the discardable barrier. Its main purpose is to
carry convergent attribute, thus preventing illegal CFG optimizations. All lanes
in a wave come to convergence point simultaneously with SIMT, thus no special
instruction is needed in the ISA. The barrier is discarded during code generation.
Differential Revision: https://reviews.llvm.org/D26584
llvm-svn: 287006
Make handling integer parameters more flexible:
- For the number of events argument allow to pass larger
integers than 32 bits as soon as compiler can prove that
the range fits in 32 bits. If not, the diagnostic will be given.
- Change type of the arguments specifying the sizes of
the corresponding block arguments to be size_t.
Review: https://reviews.llvm.org/D26509
llvm-svn: 286849
- Accept NULL pointer as a valid parameter value for clk_event.
- Generate clk_event_t arguments of internal
__enqueue_kernel_XXX function as pointers in generic address space.
Review: https://reviews.llvm.org/D26507
llvm-svn: 286836
It doesn't make sense to use the target's address space ids in this context as
this is metadata that should be referring to the "logical" OpenCL address spaces.
For flat AS machines like all "CPUs" in general, the logical AS info gets lost as
there's only one address space (0).
This commit changes the logic such that we always use the SPIR address space
ids for the argument metadata. It thus allows implementing the clGetKernelArgInfo()
and the other detection needs.
https://reviews.llvm.org/D26157
llvm-svn: 286819
Yaxun Liu