Now that #149310 has restricted lifetime intrinsics to only work on
allocas, we can also drop the explicit size argument. Instead, the size
is implied by the alloca.
This removes the ability to only mark a prefix of an alloca alive/dead.
We never used that capability, so we should remove the need to handle
that possibility everywhere (though many key places, including stack
coloring, did not actually respect this).
__enqueue_kernel_varargs' last parameter is in addrspace(5), but CodeGen
currently misses this qualifier. This commit fixes the code to preserve
the qualifier by referencing Alloca, which has its casts removed, rather
than TmpPtr.
This fixes emitting undefined behavior where a 64-bit generic
pointer is written to a 32-bit slot allocated for a private pointer.
This can be seen in test/CodeGenOpenCL/amdgcn-automatic-variable.cl's
wrong_pointer_alloca.
OpenCL Kernels body is emitted as stubs and the kernel is emitted as
call to respective stub.
(https://github.com/llvm/llvm-project/pull/115821).
The stub function should be alwaysinlined, since call to stub can cause
performance drop.
Co-authored-by: anikelal <anikelal@amd.com>
This feature is currently not supported in the compiler.
To facilitate this we emit a stub version of each kernel
function body with different name mangling scheme, and
replaces the respective kernel call-sites appropriately.
Fixes https://github.com/llvm/llvm-project/issues/60313
D120566 was an earlier attempt made to upstream a solution
for this issue.
---------
Co-authored-by: anikelal <anikelal@amd.com>
Summary:
When we were first porting to COV5, this lead to some ABI issues due to
a change in how we looked up the work group size. Bitcode libraries
relied on the builtins to emit code, but this was changed between
versions. This prevented the bitcode libraries, like OpenMP or libc,
from being used for both COV4 and COV5. The solution was to have this
'none' functionality which effectively emitted code that branched off of
a global to resolve to either version.
This isn't a great solution because it forced every TU to have this
variable in it. The patch in
https://github.com/llvm/llvm-project/pull/131033 removed support for
COV4 from OpenMP, which was the only consumer of this functionality.
Other users like HIP and OpenCL did not use this because they linked the
ROCm Device Library directly which has its own handling (The name was
borrowed from it after all).
So, now that we don't need to worry about backward compatibility with
COV4, we can remove this special handling. Users can still emit COV4
code, this simply removes the special handling used to make the OpenMP
device runtime bitcode version agnostic.
The previous implementation wasn't maintaining a faithful IR
representation of how this really works. The value returned by
createEnqueuedBlockKernel wasn't actually used as a function, and
hacked up later to be a pointer to the runtime handle global
variable. In reality, the enqueued block is a struct where the first
field is a pointer to the kernel descriptor, not the kernel itself. We
were also relying on passing around a reference to a global using a
string attribute containing its name. It's better to base this on a
proper IR symbol reference during final emission.
This now avoids using a function attribute on kernels and avoids using
the additional "runtime-handle" attribute to populate the final
metadata. Instead, associate the runtime handle reference to the
kernel with the !associated global metadata. We can then get a final,
correctly mangled name at the end.
I couldn't figure out how to get rename-with-external-symbol behavior
using a combination of comdats and aliases, so leaves an IR pass to
externalize the runtime handles for codegen. If anything breaks, it's
most likely this, so leave avoiding this for a later step. Use a
special section name to enable this behavior. This also means it's
possible to declare enqueuable kernels in source without going through
the dedicated block syntax or other dedicated compiler support.
We could move towards initializing the runtime handle in the
compiler/linker. I have a working patch where the linker sets up the
first field of the handle, avoiding the need to export the block
kernel symbol for the runtime. We would need new relocations to get
the private and group sizes, but that would avoid the runtime's
special case handling that requires the device_enqueue_symbol metadata
field.
https://reviews.llvm.org/D141700
EmitAggExprToLValue started wrapping the temporary alloca in an
addrspacecast
at some point. We take the direct type from this as the pointer argument
for the
runtime function type, but this isn't correct. Technically, we should be
querying
the target's ABI for what IR to produce for this sequence. The
assumption seems to
always have been that this will be indirectly passed with byval (or
byref).
I started working on a patch to go through the ABI handling, but it
seems to
require more time and/or clang expertise than I have at the moment.
AMDGPU disabled the use of `byval` for struct argument passing in commit
d77c620. However, when emitting `__enqueue_kernel_basic`, Clang still
adds the
`byval` attribute by default. Emitting the `byval` attribute by default
in this
context doesn’t seem like a good idea, as argument-passing conventions
are
highly target-dependent, and assumptions here could lead to issues. This
PR
removes the addition of the `byval` attribute, aligning the behavior
with other
`__enqueue_kernel_*` functions.
Currently, for AMDGPU, when compiling for OpenCL, we unconditionally use
`private` as the default address space. This is wrong for cases where
the `generic` address space is available, and is corrected via this
patch. In general, this AS map abuse is a bad hack and we should re-work
it altogether, but at least after this patch we will stop being
incorrect for e.g. OpenCL 2.0.
Generate nuw GEPs for struct member accesses, as inbounds + non-negative
implies nuw.
Regression tests are updated using update scripts where possible, and by
find + replace where not.
The previous name 'amdgpu_code_object_version', was misleading since
this is really a property of the HSA OS. The new spelling also matches
the asm directive I added in bc82cfb.
This was missing important environment context, like denormal-fp-math
and target-features. Curiously this seems to be losing nounwind. Note
this only fixes the actual invoke kernel. The invoke function is
already setting the default attribute set for internal
functions. However that is still buggy since it's not applying any use
function attributes (it's also missing uniform-work-group-size).
There seem to be too many different functions for setting attributes
with inconsistent behavior. The Function overload of
addDefaultFunctionAttributes seems to miss the target-cpu and
target-features. The AttrBuilder one seems to miss optnone (but that
seems to be disallowed on blocks anyway). Neither one calls
setTargetAttributes, when it probably should. uniform-work-group-size
is also set through AMDGPU code when it should be emitting generically
as a language property.
I also noticed update_cc_test_checks for attributes seem to not
connect the captured attribute variables to the attributes at the end
(although I think the numbers happen to work out correctly).
Neither OpenCL nor C++ for OpenCL support exceptions, so add the
`nounwind` attribute unconditionally for those languages.
Differential Revision: https://reviews.llvm.org/D142033
This adds -no-opaque-pointers to clang tests whose output will
change when opaque pointers are enabled by default. This is
intended to be part of the migration approach described in
https://discourse.llvm.org/t/enabling-opaque-pointers-by-default/61322/9.
The patch has been produced by replacing %clang_cc1 with
%clang_cc1 -no-opaque-pointers for tests that fail with opaque
pointers enabled. Worth noting that this doesn't cover all tests,
there's a remaining ~40 tests not using %clang_cc1 that will need
a followup change.
Differential Revision: https://reviews.llvm.org/D123115
For a default visibility external linkage definition, dso_local is set for ELF
-fno-pic/-fpie and COFF and Mach-O. Since default clang -cc1 for ELF is similar
to -fpic ("PIC Level" is not set), this nuance causes unneeded binary format differences.
To make emitted IR similar, ELF -cc1 -fpic will default to -fno-semantic-interposition,
which sets dso_local for default visibility external linkage definitions.
To make this flip smooth and enable future (dso_local as definition default),
this patch replaces (function) `define ` with `define{{.*}} `,
(variable/constant/alias) `= ` with `={{.*}} `, or inserts appropriate `{{.*}} `.
For consistency with normal instructions and clarity when reading IR,
it's best to print the %0, %1, ... names of function arguments in
definitions.
Also modifies the parser to accept IR in that form for obvious reasons.
llvm-svn: 367755
This reverts r326937 as it broke block argument handling in OpenCL.
See the discussion on https://reviews.llvm.org/D43783 .
The next commit will add a test case that revealed the issue.
llvm-svn: 343582
OpenCL runtime tracks the invoke function emitted for
any block expression. Due to restrictions on blocks in
OpenCL (v2.0 s6.12.5), it is always possible to know the
block invoke function when emitting call of block expression
or __enqueue_kernel builtin functions. Since __enqueu_kernel
already has an argument for the invoke function, it is redundant
to have invoke function member in the llvm block literal structure.
This patch removes invoke function from the llvm block literal
structure. It also removes the bitcast of block invoke function
to the generic block literal type which is useless for OpenCL.
This will save some space for the kernel argument, and also
eliminate some store instructions.
Differential Revision: https://reviews.llvm.org/D43783
llvm-svn: 326937
The following test case causes issue with codegen of __enqueue_block
void (^block)(void) = ^{ callee(id, out); };
enqueue_kernel(queue, 0, ndrange, block);
Clang first does codegen for block expression in the first line and deletes its block info.
Clang then tries to do codegen for the same block expression again for the second line,
and fails because the block info is gone.
The fix is to do normal codegen for both lines. Introduce an API to OpenCL runtime to
record llvm block invoke function and llvm block literal emitted for each AST block
expression, and use the recorded information for generating the wrapper kernel.
The EmitBlockLiteral APIs are cleaned up to minimize changes to the normal codegen
of blocks.
Another minor issue is that some clean up AST expression is generated for block
with captures, which can be stripped by IgnoreImplicit.
Differential Revision: https://reviews.llvm.org/D43240
llvm-svn: 325264
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
