constraints were added to _mm256_broadcast_{pd,ps} intel intrinsics.
The spec for these intrinics is ... pretty much silent on alignment.
This is especially frustrating considering the amount of discussion of
alignment in the load and store instrinsics. So I was forced to rely on
the specification for the VBROADCASTF128 instruction.
That instruction's spec is *also* completely silent on alignment.
Fortunately, when it comes to the instruction's spec, silence is enough.
There is no #GP fault option for an underaligned address so this
instruction, and by inference the intrinsic, can read any alignment.
As it happens, the old code worked exactly this way and in fact we have
plenty of code that hands pointers with less than 16-byte alignment to
these intrinsics. This code broke pretty spectacularly with this commit.
Fortunately, the fix is super simple! Change a 16 to a 1, and ta da!
Anyways, a lot of debugging for a really boring fix. =]
llvm-svn: 278202
Summary:
In order to re-define OpenCL built-in functions
'to_{private,local,global}' in OpenCL run-time library LLVM names must
be different from the clang built-in function names.
Reviewers: yaxunl, Anastasia
Subscribers: cfe-commits
Differential Revision: https://reviews.llvm.org/D23120
llvm-svn: 277743
As discussed on D22460, I've updated the vbroadcastf128 pd256/ps256 builtins to map directly to generic IR - load+splat a 128-bit vector to both lanes of a 256-bit vector.
Fix for PR28657.
llvm-svn: 276417
- Added new Builtins: enqueue_kernel, get_kernel_work_group_size
and get_kernel_preferred_work_group_size_multiple.
These Builtins use custom check to diagnose parameters of the passed Blocks
i. e. variable number of 'local void*' type params, and check different
overloads specified in Table 6.31 of OpenCL v2.0.
- IR is generated as an internal library call for each OpenCL Builtin,
reusing ObjC Block implementation.
Review: http://reviews.llvm.org/D20249
llvm-svn: 274540
Currently we only have OpenCL 2.0 Builtins i.e. pipes or address space conversions.
They have to be added only in the version 2.0 compilation mode to make the identifiers
available for use in the other versions.
Review: http://reviews.llvm.org/D20249
llvm-svn: 274509
This is important for building libclc. Since r273039 tests are failing
due to now emitting calls to these functions instead of emitting the
DAG node. The libm function names are implemented for OpenCL, and should
call the locally defined versions, so -fno-builtin is used. The IR
Some functions use the __builtins and expect the intrinsics to be
emitted. Without this we end up with nobuiltin calls to intrinsics
or to unsupported library calls.
llvm-svn: 274370
Reapplying patch in r272777 which was reverted
because the llvm patch which added support
for generating the mcrr/mcrr2 instructions
from the intrinsic was causing an assertion
failure. This has now been fixed in llvm.
llvm-svn: 272983
As noted in the code comment, a potential follow-on would be to remove
the builtins themselves. Other than ord/unord, this already works as
expected. Eg:
typedef float v4sf __attribute__((__vector_size__(16)));
v4sf fcmpgt(v4sf a, v4sf b) { return a > b; }
Differential Revision: http://reviews.llvm.org/D21268
llvm-svn: 272840
Patch adds intrinsics for mrrc/mrrc2. The
intrinsics for mrrc/mrrc2 return a single
uint64_t to represent two 32 bit values.
The mcrr/mcrr2 intrinsic was changed to
accept a single uint64_t instead of two
32 bit values as the input for consistency.
Differential Revision: http://reviews.llvm.org/D21179
llvm-svn: 272777
We can now use __builtin_nontemporal_store instead of target specific builtins for naturally aligned nontemporal stores which avoids the need for handling in CGBuiltin.cpp
The scalar integer nontemporal (unaligned) store builtins will have to wait as __builtin_nontemporal_store currently assumes natural alignment and doesn't accept the 'packed struct' trick that we use for normal unaligned load/stores.
The nontemporal loads require further backend support before we can safely convert them to __builtin_nontemporal_load
Differential Revision: http://reviews.llvm.org/D21272
llvm-svn: 272540
_InterlockedIncrement and _InterlockedDecrement have 'long' in their
prototypes. We assumed 'long' was the same size as an i32 which is
incorrect for other targets.
This fixes PR27892.
llvm-svn: 270953
OpenCL builtin functions to_{global|local|private} accepts argument of pointer type to arbitrary pointee type, and return a pointer to the same pointee type in different addr space, i.e.
global gentype *to_global(gentype *p);
It is not desirable to declare it as
global void *to_global(void *);
in opencl header file since it misses diagnostics.
This patch implements these builtin functions as Clang builtin functions. In the builtin def file they are defined to have signature void*(void*). When handling call expressions, their declarations are re-written to have correct parameter type and return type corresponding to the call argument.
In codegen call to addr void *to_addr(void*) is generated with addrcasts or bitcasts to facilitate implementation in builtin library.
Differential Revision: http://reviews.llvm.org/D19932
llvm-svn: 270261
This is matching what trunk gcc is accepting. Also adds a missing ssse3
case. PR27779. The amount of duplication here is annoying, maybe it
should be factored into a separate .def file?
llvm-svn: 270224
Summary:
Previously it was implemented as inline asm in the CUDA headers.
This change allows us to use the [addr+imm] addressing mode when
executing ld.global.nc instructions. This translates into a 1.3x
speedup on some benchmarks that call this instruction from within an
unrolled loop.
Reviewers: tra, rsmith
Subscribers: jhen, cfe-commits, jholewinski
Differential Revision: http://reviews.llvm.org/D19990
llvm-svn: 270150
The intrinsic is now called llvm.thread.pointer, not
llvm.aarch64.thread.pointer. Also, the code handling it in CGBuiltin.cpp
is dead - it's already covered by GCCBuiltin. Remove it.
Differential Revision: http://reviews.llvm.org/D19099
llvm-svn: 266817
r259537 added vfma/vfms to armv7, but the builtin was only lowered
on the AArch64 side. Instead of supporting it on ARM, get rid of it.
The vfms builtin lowered to:
%nb = fsub float -0.0, %b
%r = @llvm.fma.f32(%a, %nb, %c)
Instead, define the operation in terms of vfma, and swap the
multiplicands. It now lowers to:
%na = fsub float -0.0, %a
%r = @llvm.fma.f32(%na, %b, %c)
This matches the instruction more closely, and lets current LLVM
generate the "natural" operand ordering:
fmls.2s v0, v1, v2
instead of the crooked (but equivalent):
fmls.2s v0, v2, v1
Except for theses changes, assembly is identical.
LLVM accepts both commutations, and the LLVM tests in:
test/CodeGen/AArch64/arm64-fmadd.ll
test/CodeGen/AArch64/fp-dp3.ll
test/CodeGen/AArch64/neon-fma.ll
test/CodeGen/ARM/fusedMAC.ll
already check either the new one only, or both.
Also verified against the test-suite unittests.
llvm-svn: 266807
isinf (is infinite) and isfinite should be implemented with the same function
except we change the comparison operator.
See PR27145 for more details:
https://llvm.org/bugs/show_bug.cgi?id=27145
Ref: forked off of the discussion in D18513.
Differential Revision: http://reviews.llvm.org/D18648
llvm-svn: 265675
Summary: See LLVM change D18775 for details, this change depends on it.
Reviewers: jyknight, reames
Subscribers: cfe-commits
Differential Revision: http://reviews.llvm.org/D18776
llvm-svn: 265569
These functions cannot be implemented as atomicrmw or cmpxchg
instructions, so they are implemented as a call to the NVVM intrinsics
@llvm.nvvm.atomic.load.inc.32.p0i32 and
@llvm.nvvm.atomic.load.dec.32.p0i32.
Patch by Jason Henline.
Reviewers: jlebar
Differential Revision: http://reviews.llvm.org/D18322
llvm-svn: 264009
Includes new built-in, conversion of built-in to target-independent intrinsic
and update in the header file. Tests are also updated. There is a second part in
the backend for which I will post a separate code-review. BACKEND PART SHOULD BE
COMMITTED FIRST.
Phabricator: http://reviews.llvm.org/D17816
llvm-svn: 263051
Fix arc patch fuzz error.
Summary:
Support for the pipe built-in functions for OpenCL 2.0.
The pipe builtin functions may have infinite kinds of element types, one approach
would be to just generate calls that would always use generic types such as void*.
This patch is based on bader's opencl support patch on SPIR-V branch.
Reviewers: Anastasia, pekka.jaaskelainen
Subscribers: keryell, bader, cfe-commits
Differential Revision: http://reviews.llvm.org/D15914
llvm-svn: 258782
Summary:
Support for the pipe built-in functions for OpenCL 2.0.
The pipe builtin functions may have infinite kinds of element types, one approach
would be to just generate calls that would always use generic types such as void*.
This patch is based on bader's opencl support patch on SPIR-V branch.
Reviewers: Anastasia, pekka.jaaskelainen
Subscribers: keryell, bader, cfe-commits
Differential Revision: http://reviews.llvm.org/D15914
llvm-svn: 258773
Keep the ones still used by libclc around for now.
Emit the new amdgcn intrinsic name if not targeting r600,
in which case the old AMDGPU name is still used.
llvm-svn: 258560
`pass_object_size` is our way of enabling `__builtin_object_size` to
produce high quality results without requiring inlining to happen
everywhere.
A link to the design doc for this attribute is available at the
Differential review link below.
Differential Revision: http://reviews.llvm.org/D13263
llvm-svn: 254554
Function__builtin_signbit returns wrong value for type ppcf128 on big endian
machines. This patch fixes how value is generated in that case.
Patch by Aleksandar Beserminji.
Differential Revision: http://reviews.llvm.org/D14149
llvm-svn: 252307
only one of a group of possibilities.
This changes the syntax in the builtin files to represent:
, as the and operator
| as the or operator
The former syntax matches how the backend tablegen files represent
multiple subtarget features being required.
Updated the builtin and intrinsic headers accordingly for the new
syntax.
llvm-svn: 251388
match the feature set of the function that they're being called from.
This ensures that we can effectively diagnose some[1] code that would
instead ICE in the backend with a failure to select message.
Example:
__m128d foo(__m128d a, __m128d b) {
return __builtin_ia32_addsubps(b, a);
}
compiled for normal x86_64 via:
clang -target x86_64-linux-gnu -c
would fail to compile in the back end because the normal subtarget
features for x86_64 only include sse2 and the builtin requires sse3.
[1] We're still not erroring on:
__m128i bar(__m128i const *p) { return _mm_lddqu_si128(p); }
where we should fail and error on an always_inline function being
inlined into a function that doesn't support the subtarget features
required.
llvm-svn: 250473
Summary: __nvvm_atom_cas_* returns the old value instead of whether the swap succeeds.
Reviewers: eliben, tra
Subscribers: jholewinski, llvm-commits
Differential Revision: http://reviews.llvm.org/D13306
llvm-svn: 248951
This is the clang commit associated with llvm r248887.
This commit changes the interface of the vld[1234], vld[234]lane, and vst[1234],
vst[234]lane ARM neon intrinsics and associates an address space with the
pointer that these intrinsics take. This changes, e.g.,
<2 x i32> @llvm.arm.neon.vld1.v2i32(i8*, i32)
to
<2 x i32> @llvm.arm.neon.vld1.v2i32.p0i8(i8*, i32)
This change ensures that address spaces are fully taken into account in the ARM
target during lowering of interleaved loads and stores.
Differential Revision: http://reviews.llvm.org/D13127
llvm-svn: 248888
* adds -aux-triple option to specify target triple
* propagates aux target info to AST context and Preprocessor
* pulls in target specific preprocessor macros.
* pulls in target-specific builtins from aux target.
* sets appropriate host or device attribute on builtins.
Differential Revision: http://reviews.llvm.org/D12917
llvm-svn: 248299
Summary:
This change adds support for `__builtin_ms_va_list`, a GCC extension for
variadic `ms_abi` functions. The existing `__builtin_va_list` support is
inadequate for this because `va_list` is defined differently in the Win64
ABI vs. the System V/AMD64 ABI.
Depends on D1622.
Reviewers: rsmith, rnk, rjmccall
CC: cfe-commits
Differential Revision: http://reviews.llvm.org/D1623
llvm-svn: 247941
Summary:
Currently clang provides no general way to generate nontemporal loads/stores.
There are some architecture specific builtins for doing so (e.g. in x86), but
there is no way to generate non-temporal store on, e.g. AArch64. This patch adds
generic builtins which are expanded to a simple store with '!nontemporal'
attribute in IR.
Differential Revision: http://reviews.llvm.org/D12313
llvm-svn: 247104
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
This implements basic support for compiling (though not yet assembling
or linking) for a WebAssembly target. Note that ABI details are not yet
finalized, and may change.
Differential Revision: http://reviews.llvm.org/D12002
llvm-svn: 246814
This patch depends on r246688 (D12341).
The goal is to make LLVM generate different code for these functions for a target that
has cheap branches (see PR23827 for more details):
int foo();
int normal(int x, int y, int z) {
if (x != 0 && y != 0) return foo();
return 1;
}
int crazy(int x, int y) {
if (__builtin_unpredictable(x != 0 && y != 0)) return foo();
return 1;
}
Differential Revision: http://reviews.llvm.org/D12458
llvm-svn: 246699
GCC 4.8+ has a PowerPC-specific intrinsic, __builtin_ppc_get_timebase, to do
what Clang's __builtin_readcyclecounter does. For compatibility with code that
uses GCC's spelling (including glibc), support it as well.
Partially fixes PR23681.
llvm-svn: 246510
We had "vcvt_f16" and "VCVT_HIGH_F16": for other FP types, this naming
is used for intrinsics with integer overloads. The FP->FP conversions,
on the other hand, use the full "vcvt_f32_f64" name instead.
Use the same naming convention for the f16<->f32 conversions.
While there, reorder the definitions a little bit.
llvm-svn: 245763
Marina Yatsina