Summary:
To use new/delete in NVPTX code we need to define them. Implementation
copied from CUDA wrappers.
Reviewers: hfinkel, jdoerfert
Subscribers: mgorny, guansong, kkwli0, caomhin, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D73128
Summary:
For now, this ABI simply expands all possible aggregate arguments and
returns all possible aggregates directly. This ABI will change rapidly
as we prototype and benchmark a new ABI that takes advantage of
multivalue return and possibly other changes from the MVP ABI.
Reviewers: aheejin, dschuff
Subscribers: sbc100, jgravelle-google, sunfish, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D72972
Summary:
Moves a batch of instructions from unimplemented-simd128 to simd128
because they have recently become available in V8.
Reviewers: aheejin
Subscribers: dschuff, sbc100, jgravelle-google, hiraditya, sunfish, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D73926
This extends the RemarkStreamer to allow for other emitters (e.g.
frontends, SIL, etc.) to emit remarks through a common interface.
See changes in llvm/docs/Remarks.rst for motivation and design choices.
Differential Revision: https://reviews.llvm.org/D73676
To factor the error checking, use importChecked instead of importSeq.
This avoids repeating the names of all of the imported child nodes once,
and allows errors to be checked with a single conditional as it is with
importSeq.
After:
peak memory: 601.63MB
real: 0m19.172s
obj size: 8,352kb
Before:
peak memory: 954.11MB
real: 0m26.188s
obj size: 10,000kb
The speed is not as impressive as I hoped, but the memory use reduction
is impressive, and seems worth it.
Reviewed By: martong, shafik
Differential Revision: https://reviews.llvm.org/D73675
We did not have a CXXThisScope around constraint checking of functions and
function template specializations, causing a crash when checking a constraint
that had a 'this' (bug 44689).
Recommit after fixing test.
We did not have a CXXThisScope around constraint checking of functions and
function template specializations, causing a crash when checking a constraint
that had a 'this' (bug 44689)
Add support for Flush in the OMPIRBuilder. This patch also adds changes
to clang to use the OMPIRBuilder when '-fopenmp-enable-irbuilder'
commandline option is used.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D70712
Summary:
We were linking all the clang objects and shared libraries into
libclang-cpp.so, which was causing the command line options to be
registered twice.
Reviewers: beanz, mgorny
Reviewed By: beanz, mgorny
Subscribers: merge_guards_bot, mgorny, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D68520
Summary:
- The device compilation needs to have a consistent source code compared
to the corresponding host compilation. If macros based on the
host-specific target processor is not properly populated, the device
compilation may fail due to the inconsistent source after the
preprocessor. So far, only the host triple is used to build the
macros. If a detailed host CPU target or certain features are
specified, macros derived from them won't be populated properly, e.g.
`__SSE3__` won't be added unless `+sse3` feature is present. On
Windows compilation compatible with MSVC, that missing macros result
in that intrinsics are not included and cause device compilation
failure on the host-side source.
- This patch addresses this issue by introducing two `cc1` options,
i.e., `-aux-target-cpu` and `-aux-target-feature`. If a specific host
CPU target or certain features are specified, the compiler driver will
append them during the construction of the offline compilation
actions. Then, the toolchain in `cc1` phase will populate macros
accordingly.
- An internal option `--gpu-use-aux-triple-only` is added to fall back
the original behavior to help diagnosing potential issues from the new
behavior.
Reviewers: tra, yaxunl
Subscribers: cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D73942
Summary:
Changes:
- Calls to consteval function are now evaluated in constant context but IR is still generated for them.
- Add diagnostic for taking address of a consteval function in non-constexpr context.
- Add diagnostic for address of consteval function accessible at runtime.
- Add tests
Reviewers: rsmith, aaron.ballman
Reviewed By: rsmith
Subscribers: mgrang, riccibruno, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D63960
Summary: Add comments to the list of tokens that can follow the ']' at the end of a C# attribute specifier to prevent comments after attribute specifiers from being formatted as continuations.
Reviewers: MyDeveloperDay, krasimir
Reviewed By: MyDeveloperDay
Tags: #clang-format
Differential Revision: https://reviews.llvm.org/D73977
Summary: Spells out some `auto`s explicitly and adds another test for the matcher `isExpandedFromMacro`.
Reviewers: aaron.ballman
Subscribers: gribozavr, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D73975
AMDGPU and x86 at least both have separate controls for whether
denormal results are flushed on output, and for whether denormals are
implicitly treated as 0 as an input. The current DAGCombiner use only
really cares about the input treatment of denormals.
Linux commit
1cf5b23988 (diff-289313b9fec99c6f0acfea19d9cfd949)
uses "#pragma clang attribute push (__attribute__((preserve_access_index)),
apply_to = record)"
to apply CO-RE relocations to all records including the following pattern:
#pragma clang attribute push (__attribute__((preserve_access_index)), apply_to = record)
typedef struct {
int a;
} __t;
#pragma clang attribute pop
int test(__t *arg) { return arg->a; }
The current approach to use struct type in the relocation record will
result in an anonymous struct, which make later type matching difficult
in bpf loader. In fact, current BPF backend will fail the above program
with assertion:
clang: ../lib/Target/BPF/BPFAbstractMemberAccess.cpp:796: ...
Assertion `TypeName.size()' failed.
The patch use the base lvalue type for the "base" value to annotate
preservee_{struct,union}_access_index intrinsics. In the above example,
the type will be "__t" which preserved the type name.
Differential Revision: https://reviews.llvm.org/D73900
Summary:
This revision adds a matcher `isExpandedFromMacro` that determines whether a
statement is (transitively) expanded from a given macro.
Reviewers: gribozavr
Subscribers: cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D73965
Summary:
This patch changes the underlying type of the ARM::ArchExtKind
enumeration to uint64_t and adjusts the related code.
The goal of the patch is to prepare the code base for a new
architecture extension.
Reviewers: simon_tatham, eli.friedman, ostannard, dmgreen
Reviewed By: dmgreen
Subscribers: merge_guards_bot, kristof.beyls, hiraditya, cfe-commits, llvm-commits, pbarrio
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D73906
Align `Type` names to their actual OpenCL C type names, such that we
can print OpenCL C types from TableGen. This is particularly helpful
for completeness comparisons with `opencl-c.h`.
Drop the "Name" field from `TypeList`, as it is unused.
Minor formatting changes.
The tool is now looked for in the source directory rather than in the
install directory, which should exclude the problems with not being able
to find it.
The tests still aren't being run on Windows, but they hopefully will run
on other platforms that have shell, which hopefully also means Perl.
Differential Revision: https://reviews.llvm.org/D69781
Summary:
Clang -fpic defaults to -fno-semantic-interposition (GCC -fpic defaults
to -fsemantic-interposition).
Users need to specify -fsemantic-interposition to get semantic
interposition behavior.
Semantic interposition is currently a best-effort feature. There may
still be some cases where it is not handled well.
Reviewers: peter.smith, rnk, serge-sans-paille, sfertile, jfb, jdoerfert
Subscribers: dschuff, jyknight, dylanmckay, nemanjai, jvesely, kbarton, fedor.sergeev, asb, rbar, johnrusso, simoncook, sabuasal, niosHD, jrtc27, zzheng, edward-jones, atanasyan, rogfer01, MartinMosbeck, brucehoult, the_o, arphaman, PkmX, jocewei, jsji, Jim, lenary, s.egerton, pzheng, sameer.abuasal, apazos, luismarques, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D73865
We previously instantiated type-constraints of template type parameters along with the type parameter itself,
this caused problems when the type-constraints created by abbreviated templates refreneced other parameters
in the abbreviated templates.
When encountering a template type parameter with a type constraint, if it is implicit, delay instantiation of
the type-constraint until the function parameter which created the invented template type parameter is
instantiated.
Reland after fixing bug caused by another flow reaching SubstParmVarDecl and instantiating the TypeConstraint
a second time.
Add support for Master and Critical directive in the OMPIRBuilder. Both make use of a new common interface for emitting inlined OMP regions called `emitInlinedRegion` which was added in this patch as well.
Also this patch modifies clang to use the new directives when `-fopenmp-enable-irbuilder` commandline option is passed.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D72304
We previously instantiated type-constraints of template type parameters along with the type parameter itself,
this caused problems when the type-constraints created by abbreviated templates refreneced other parameters
in the abbreviated templates.
When encountering a template type parameter with a type constraint, if it is implicit, delay instantiation of
the type-constraint until the function parameter which created the invented template type parameter is
instantiated.
Summary:
In big-endian MVE, the simple vector load/store instructions (i.e.
both contiguous and non-widening) don't all store the bytes of a
register to memory in the same order: it matters whether you did a
VSTRB.8, VSTRH.16 or VSTRW.32. Put another way, the in-register
formats of different vector types relate to each other in a different
way from the in-memory formats.
So, if you want to 'bitcast' or 'reinterpret' one vector type as
another, you have to carefully specify which you mean: did you want to
reinterpret the //register// format of one type as that of the other,
or the //memory// format?
The ACLE `vreinterpretq` intrinsics are specified to reinterpret the
register format. But I had implemented them as LLVM IR bitcast, which
is specified for all types as a reinterpretation of the memory format.
So a `vreinterpretq` intrinsic, applied to values already in registers,
would code-generate incorrectly if compiled big-endian: instead of
emitting no code, it would emit a `vrev`.
To fix this, I've introduced a new IR intrinsic to perform a
register-format reinterpretation: `@llvm.arm.mve.vreinterpretq`. It's
implemented by a trivial isel pattern that expects the input in an
MQPR register, and just returns it unchanged.
In the clang codegen, I only emit this new intrinsic where it's
actually needed: I prefer a bitcast wherever it will have the right
effect, because LLVM understands bitcasts better. So we still generate
bitcasts in little-endian mode, and even in big-endian when you're
casting between two vector types with the same lane size.
For testing, I've moved all the codegen tests of vreinterpretq out
into their own file, so that they can have a different set of RUN
lines to check both big- and little-endian.
Reviewers: dmgreen, MarkMurrayARM, miyuki, ostannard
Reviewed By: dmgreen
Subscribers: kristof.beyls, hiraditya, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D73786
Summary:
These instructions generate a vector of consecutive elements starting
from a given base value and incrementing by 1, 2, 4 or 8. The `wdup`
versions also wrap the values back to zero when they reach a given
limit value. The instruction updates the scalar base register so that
another use of the same instruction will continue the sequence from
where the previous one left off.
At the IR level, I've represented these instructions as a family of
target-specific intrinsics with two return values (the constructed
vector and the updated base). The user-facing ACLE API provides a set
of intrinsics that throw away the written-back base and another set
that receive it as a pointer so they can update it, plus the usual
predicated versions.
Because the intrinsics return two values (as do the underlying
instructions), the isel has to be done in C++.
This is the first family of MVE intrinsics that use the `imm_1248`
immediate type in the clang Tablegen framework, so naturally, I found
I'd given it the wrong C integer type. Also added some tests of the
check that the immediate has a legal value, because this is the first
time those particular checks have been exercised.
Finally, I also had to fix a bug in MveEmitter which failed an
assertion when I nested two `seq` nodes (the inner one used to extract
the two values from the pair returned by the IR intrinsic, and the
outer one put on by the predication multiclass).
Reviewers: dmgreen, MarkMurrayARM, miyuki, ostannard
Reviewed By: dmgreen
Subscribers: kristof.beyls, hiraditya, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D73357
Summary:
The unpredicated case of this is trivial: the clang codegen just makes
a vector splat of the input, and LLVM isel is already prepared to
handle that. For the predicated version, I've generated a `select`
between the same vector splat and the `inactive` input parameter, and
added new Tablegen isel rules to match that pattern into a predicated
`MVE_VDUP` instruction.
Reviewers: dmgreen, MarkMurrayARM, miyuki, ostannard
Reviewed By: dmgreen
Subscribers: kristof.beyls, hiraditya, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D73356
During the review of D73007 Aaron Puchert mentioned
`warn_for_range_variable_always_copy` shouldn't be part of -Wall since
some coding styles require `for(const auto &bar : bars)`. This warning
would cause false positives for these users. Based on Aaron's proposal
refactored the warnings:
* -Wrange-loop-construct warns about possibly unintended constructor
calls. This is part of -Wall. It contains
* warn_for_range_copy: loop variable A of type B creates a copy from
type C
* warn_for_range_const_reference_copy: loop variable A is initialized
with a value of a different type resulting in a copy
* -Wrange-loop-bind-reference warns about misleading use of reference
types. This is not part of -Wall. It contains
* warn_for_range_variable_always_copy: loop variable A is always a copy
because the range of type B does not return a reference
Differential Revision: https://reviews.llvm.org/D73434
Summary:
From `clang-format` version 3.7.0 and up, , there is no way to keep following format of ObjectiveC block:
```
- (void)_aMethod
{
[self.test1 t:self w:self callback:^(typeof(self) self, NSNumber *u, NSNumber *v) {
u = c;
}]
}
```
Regardless of the change in `.clang-format` configuration file, all parameters will be lined up so that colons will be on the same column, like following:
```
- (void)_aMethod
{
[self.test1 t:self
w:self
callback:^(typeof(self) self, NSNumber *u, NSNumber *v) {
u = c;
}]
}
```
Considering with ObjectiveC, the first code style is cleaner & more readable for some people, I've added a config option: `ObjCDontBreakBeforeNestedBlockParam` (boolean) so that if it is enable, the first code style will be favored.
Reviewed By: MyDeveloperDay
Patch By: ghvg1313
Tags: #clang, #clang-format
Differential Revision: https://reviews.llvm.org/D70926
Driver errors if -fomit-frame-pointer is used together with -pg.
useFramePointerForTargetByDefault() returns true if -pg is specified.
=>
(!OmitFP && useFramePointerForTargetByDefault(Args, Triple)) is true
=>
We cannot get FramePointerKind::None
When T is a class type, only nvsize(T) bytes need be accessible through
the reference. We had matching bugs in the application of the
dereferenceable attribute and in -fsanitize=undefined.
Summary: Just like templates, they are excepted from the ODR rule.
Reviewed By: aaron.ballman, rsmith
Differential Revision: https://reviews.llvm.org/D68923
instead of the "enabled by default" color.
It may be technically correct to list unimplemented diagnostics as
"enabled by default" but it's quite misleading.
types are needed to compute the return type of a defaulted operator<=>.
This raises the question of what to do if return type deduction fails.
The standard doesn't say, and implementations vary, so for now reject
that case eagerly to keep our options open.
isDeclarationSpecifiers did not handle some cases of placeholder-type-specifiers with
type-constraints, causing parsing bugs in abbreviated constructor templates.
Add comprehensive handling of type-constraints to isDeclarationSpecifier.
Summary: Every call to a main like function in llvm and clang lib violates the naming convention for parameters. This prevents clang-tidy warning on such breaches.
Reviewers: alexfh, hokein
Reviewed By: hokein
Subscribers: merge_guards_bot, aheejin, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D73715
We previously would not correctly for the initial parameter mapping for variadic template parameters in Concepts.
Testing this lead to the discovery that with the normalization process we would need to substitute into already-substituted-into
template arguments, which means we need to add NonTypeTemplateParmExpr support to TemplateInstantiator.
We do that by substituting into the replacement and the type separately, and then re-checking the expression against the NTTP
with the new type, in order to form any new required implicit casts (for cases where the type of the NTTP was dependent).
First attempt at implementing -fsemantic-interposition.
Rely on GlobalValue::isInterposable that already captures most of the expected
behavior.
Rely on a ModuleFlag to state whether we should respect SemanticInterposition or
not. The default remains no.
So this should be a no-op if -fsemantic-interposition isn't used, and if it is,
isInterposable being already used in most optimisation, they should honor it
properly.
Note that it only impacts architecture compiled with -fPIC and no pie.
Differential Revision: https://reviews.llvm.org/D72829
Add fixits for messaging self in MRR or using super, as the intent is
clear, and it turns out people do that a lot more than expected.
Allow for objc_direct_members on main interfaces, it's extremely useful
for internal only classes, and proves to be quite annoying for adoption.
Add some better warnings around properties direct/non-direct clashes (it
was done for methods but properties were a miss).
Radar-Id: rdar://problem/58355212
Signed-off-by: Pierre Habouzit <phabouzit@apple.com>
For non direct methods, the codegen uses the type of the Implementation.
Because Objective-C rules allow some differences between the Declaration
and Implementation return types, when the Implementation is in this
translation unit, the type of the Implementation should be preferred to
emit the Function over the Declaration.
Radar-Id: rdar://problem/58797748
Signed-off-by: Pierre Habouzit <phabouzit@apple.com>
Differential Revision: https://reviews.llvm.org/D73208
A constrained function with an auto return type would have it's definition
instantiated in order to deduce the auto return type before the constraints
are checked.
Move the constraints check after the return type deduction.
when building a defaulted comparison.
As a convenient way of asking whether `x @ y` is valid and building it,
we previouly always performed overload resolution and built an
overloaded expression, which would both end up picking a builtin
operator candidate when given a non-overloadable type. But that's not
quite right, because it can result in our finding a user-declared
operator overload, which we should never do when applying operators
non-overloadable types.
Handle this more correctly: skip overload resolution when building
`x @ y` if the operands are not overloadable. But still perform overload
resolution (considering only builtin candidates) when checking validity,
as we don't have any other good way to ask whether a binary operator
expression would be valid.
In passing, split it up into three values (no explicit functions /
explicit conversion functions only / any explicit functions) in
preparation for using that in a future change.
This is never appropriate on Fuchsia and any future needs for
system library dependencies of compiler-supplied runtimes will
be addressed via .deplibs instead of driver hacks.
Patch By: mcgrathr
Differential Revision: https://reviews.llvm.org/D73734
Avoid recursively instantiating importSeq. Use initializer list
expansion to stamp out a single instantiation of std::tuple of the
deduced sequence of types, and thread the error around that tuple type.
Avoids needlessly instantiating std::tuple N-1 times.
new time to compile: 0m25.985s
old time to compile: 0m35.563s
new obj size: 10,000kb
old obj size: 12,332kb
I found the slow TU by looking at ClangBuildAnalyzer results, and looked
at -ftime-trace for the file in chrome://tracing to find this.
Tested with: clang-cl, MSVC, and GCC.
Reviewed By: martong
Differential Revision: https://reviews.llvm.org/D73667
We previously checked for containsUnexpandedParameterPack in CSEs by observing the property
in the converted arguments of the CSE. This may not work if the argument is an expanded
type-alias that contains a pack-expansion (see added test).
Check the as-written arguments when determining containsUnexpandedParameterPack and isInstantiationDependent.
Summary: With OpenMP offloading host compilation is done in two phases to capture host IR that is passed to all device compilations as input. But it turns out that we currently run entire LLVM optimization pipeline on host IR on both compilations which may have unpredictable effects on the resulting code. This patch fixes this problem by disabling LLVM passes on the first compilation, so the host IR that is passed to device compilations will be captured right after front end.
Reviewers: ABataev, jdoerfert, hfinkel
Reviewed By: ABataev
Subscribers: guansong, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D73721
regions.
If the lastprivate conditional is passed as shared in inner region, we
shall check if it was ever changed and use this updated value after exit
from the inner region as an update value.
Summary:
Fat object size has significantly increased after D65819 which changed bundler tool to add host object as a normal bundle to the fat output which almost doubled its size. That patch was fixing the following issues
1. Problems associated with the partial linking - global constructors were not called for partially linking objects which clearly resulted in incorrect behavior.
2. Eliminating "junk" target object sections from the linked binary on the host side.
The first problem is no longer relevant because we do not use partial linking for creating fat objects anymore. Target objects sections are now inserted into the resulting fat object with a help of llvm-objcopy tool.
The second issue, "junk" sections in the linked host binary, has been fixed in D73408 by adding "exclude" flag to the fat object's sections which contain target objects. This flag tells linker to drop section from the inputs when linking executable or shared library, therefore these sections will not be propagated in the linked binary.
Since both problems have been solved, we can revert D65819 changes to reduce fat object size and this patch essentially is doing that.
Reviewers: ABataev, alexshap, jdoerfert
Reviewed By: ABataev
Subscribers: cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D73642
Summary:
This patch hooks the `Preprocessor` trough `BugReporter` to the
`CheckerContext` so the checkers could look for macro definitions.
Reviewed By: NoQ
Differential Revision: https://reviews.llvm.org/D69731
Summary:
This patch uses the new `DynamicSize.cpp` to serve dynamic information.
Previously it was static and probably imprecise data.
Reviewed By: NoQ
Differential Revision: https://reviews.llvm.org/D69599
Summary:
This patch introduces a placeholder for representing the dynamic size of
regions. It also moves the `getExtent()` method of `SubRegions` to the
`MemRegionManager` as `getStaticSize()`.
Reviewed By: NoQ
Differential Revision: https://reviews.llvm.org/D69540
The constrained fcmp intrinsics don't allow the TRUE/FALSE predicates.
Using them will assert. To workaround this I'm emitting the old X86 specific intrinsics that were never removed from the backend when we switched to using fcmp in IR. We have no way to mark them as being strict, but that's true of all target specific intrinsics so doesn't seem like we need to solve that here.
I've also added support for selecting between signaling and quiet.
Still need to support SAE which will require using a target specific
intrinsic. Also need to fix masking to not use an AND instruction
after the compare.
Differential Revision: https://reviews.llvm.org/D72906
Summary: This flag tells link editor to exclude section from linker inputs when linking executable or shared library.
Reviewers: ABataev, alexshap, jdoerfert
Reviewed By: ABataev
Subscribers: cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D73408
Iterator modeling depends on container modeling,
but not vice versa. This enables the possibility
to arrange these two modeling checkers into
separate layers.
There are several advantages for doing this: the
first one is that this way we can keep the
respective modeling checkers moderately simple
and small. Furthermore, this enables creation of
checkers on container operations which only
depend on the container modeling. Thus iterator
modeling can be disabled together with the
iterator checkers if they are not needed.
Since many container operations also affect
iterators, container modeling also uses the
iterator library: it creates iterator positions
upon calling the `begin()` or `end()` method of
a containter (but propagation of the abstract
position is left to the iterator modeling),
shifts or invalidates iterators according to the
rules upon calling a container modifier and
rebinds the iterator to a new container upon
`std::move()`.
Iterator modeling propagates the abstract
iterator position, handles the relations between
iterator positions and models iterator
operations such as increments and decrements.
Differential Revision: https://reviews.llvm.org/D73547
Summary:
Currently, sqdmulh_lane and friends from the ACLE (implemented in arm_neon.h),
are represented in LLVM IR as a (by vector) sqdmulh and a vector of (repeated)
indices, like so:
%shuffle = shufflevector <4 x i16> %v, <4 x i16> undef, <4 x i32> <i32 3, i32 3, i32 3, i32 3>
%vqdmulh2.i = tail call <4 x i16> @llvm.aarch64.neon.sqdmulh.v4i16(<4 x i16> %a, <4 x i16> %shuffle)
When %v's values are known, the shufflevector is optimized away and we are no
longer able to select the lane variant of sqdmulh in the backend.
This defeats a (hand-coded) optimization that packs several constants into a
single vector and uses the lane intrinsics to reduce register pressure and
trade-off materialising several constants for a single vector load from the
constant pool, like so:
int16x8_t v = {2,3,4,5,6,7,8,9};
a = vqdmulh_laneq_s16(a, v, 0);
b = vqdmulh_laneq_s16(b, v, 1);
c = vqdmulh_laneq_s16(c, v, 2);
d = vqdmulh_laneq_s16(d, v, 3);
[...]
In one microbenchmark from libjpeg-turbo this accounts for a 2.5% to 4%
performance difference.
We could teach the compiler to recover the lane variants, but this would likely
require its own pass. (Alternatively, "volatile" could be used on the constants
vector, but this is a bit ugly.)
This patch instead implements the following LLVM IR intrinsics for AArch64 to
maintain the original structure through IR optmization and into instruction
selection:
- sqdmulh_lane
- sqdmulh_laneq
- sqrdmulh_lane
- sqrdmulh_laneq.
These 'lane' variants need an additional register class. The second argument
must be in the lower half of the 64-bit NEON register file, but only when
operating on i16 elements.
Note that the existing patterns for shufflevector and sqdmulh into sqdmulh_lane
(etc.) remain, so code that does not rely on NEON intrinsics to generate these
instructions is not affected.
This patch also changes clang to emit these IR intrinsics for the corresponding
NEON intrinsics (AArch64 only).
Reviewers: SjoerdMeijer, dmgreen, t.p.northover, rovka, rengolin, efriedma
Reviewed By: efriedma
Subscribers: kristof.beyls, hiraditya, jdoerfert, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D71469
Alexey Bataev