platform/upstream/llvm.git
22 months ago[CLANG] Fix typo in test/CodeGen/c-strings.c
Muhammad Omair Javaid [Tue, 20 Dec 2022 08:56:03 +0000 (13:56 +0500)]
[CLANG] Fix typo in test/CodeGen/c-strings.c

This patch fixes a typo in test/CodeGen/c-strings.c. Test was failing
on AArch64/Windows.

22 months agoFix an unused-variable warning in release build, NFC
Haojian Wu [Tue, 20 Dec 2022 08:38:03 +0000 (09:38 +0100)]
Fix an unused-variable warning in release build, NFC

22 months agomlir/{SPIRV,Bufferization}: use std::optional in .td files (NFC)
Ramkumar Ramachandra [Mon, 19 Dec 2022 18:50:54 +0000 (19:50 +0100)]
mlir/{SPIRV,Bufferization}: use std::optional in .td files (NFC)

This is part of an effort to migrate from llvm::Optional to
std::optional. 22426110c5ef changed the way mlir-tblgen generates .inc
files, emitting std::optional when an Optional attribute is specified in
a .td file. It also changed several .td files hard-coding llvm::Optional
to use std::optional. However, the patch excluded a few .td files in
SPIRV and Bufferization hard-coding llvm::Optional. This patch fixes
that defect, and after this patch, references to llvm::Optional in .cpp
and .h files can be replaced mechanically.

See also: https://discourse.llvm.org/t/deprecating-llvm-optional-x-hasvalue-getvalue-getvalueor/63716

Signed-off-by: Ramkumar Ramachandra <r@artagnon.com>
Differential Revision: https://reviews.llvm.org/D140329

22 months ago[flang] Finish substring lowering
Jean Perier [Tue, 20 Dec 2022 07:46:45 +0000 (08:46 +0100)]
[flang] Finish substring lowering

Hlfir.designate was made to support substrings but so far substrings
were not yet lowered to it. Implement support for them.

Differential Revision: https://reviews.llvm.org/D140310

22 months ago[AMDGPU] Replace getPhysRegClass with getPhysRegBaseClass
Carl Ritson [Tue, 20 Dec 2022 06:47:13 +0000 (15:47 +0900)]
[AMDGPU] Replace getPhysRegClass with getPhysRegBaseClass

Accelerate finding the base class for a physical register by
building a statically mapping table from physical registers
to base classes using TableGen.

Replace uses of SIRegisterInfo::getPhysRegClass with
TargetRegisterInfo::getPhysRegBaseClass in order to use
the computed table.

Reviewed By: arsenm, foad

Differential Revision: https://reviews.llvm.org/D139422

22 months ago[TableGen] Emit table mapping physical registers to base classes
Carl Ritson [Tue, 20 Dec 2022 06:21:51 +0000 (15:21 +0900)]
[TableGen] Emit table mapping physical registers to base classes

Allow targets to define a mapping from registers to register
classes such that each register has exactly one base class.
As registers may be in multiple register classes the base class
is determined by the container class with the lowest BaseClassOrder.

Only register classes with BaseClassOrder set are considered
when determining the base classes.  By default BaseClassOrder is
unset in RegisterClass so no code is generated unless a target
explicit defines one or more base register classes.

Reviewed By: arsenm, foad

Differential Revision: https://reviews.llvm.org/D139616

22 months ago[UniformityAnalysis][NFC] Make checks independent of order using CHECK-DAG.
Sameer Sahasrabuddhe [Tue, 20 Dec 2022 06:05:45 +0000 (11:35 +0530)]
[UniformityAnalysis][NFC] Make checks independent of order using CHECK-DAG.

This fixes the buildbot failure on reverse iteration:
https://lab.llvm.org/buildbot/#/builders/54/builds/3034

22 months ago[RISCV] Replace i64:$r in tablegen patterns with GPR:$r. NFC
Craig Topper [Tue, 20 Dec 2022 04:52:33 +0000 (20:52 -0800)]
[RISCV] Replace i64:$r in tablegen patterns with GPR:$r. NFC

It's much more common to use a register class rather than a type.

Add an additional i64 cast to the patterns where needed to avoid
increasing isel table size.

22 months ago[RISCV] Add Predicates to the XVentanaCondOps isel patterns.
Craig Topper [Tue, 20 Dec 2022 03:28:56 +0000 (19:28 -0800)]
[RISCV] Add Predicates to the XVentanaCondOps isel patterns.

Missed this in the review of D139394. Not a functional issue because
select only makes it to isel with XVentanaCondOps.

22 months ago[RISCV] Add +xventanacondops command line to select-binop-identity.ll. NFC
Craig Topper [Tue, 20 Dec 2022 03:12:35 +0000 (19:12 -0800)]
[RISCV] Add +xventanacondops command line to select-binop-identity.ll. NFC

22 months ago[RISCV] Add zeroext attribute to i1 arguments in xventanacondops.ll. NFC
Craig Topper [Tue, 20 Dec 2022 03:09:24 +0000 (19:09 -0800)]
[RISCV] Add zeroext attribute to i1 arguments in xventanacondops.ll. NFC

Removes some extra andi instructions.

22 months ago[OpenMP] Account for dynamic shared memory in the AMDGPU nextgen plugin
Johannes Doerfert [Tue, 20 Dec 2022 03:09:07 +0000 (19:09 -0800)]
[OpenMP] Account for dynamic shared memory in the AMDGPU nextgen plugin

22 months ago[OpenMP] Improve AMDGPU Plugin
Johannes Doerfert [Sat, 17 Dec 2022 07:46:59 +0000 (23:46 -0800)]
[OpenMP] Improve AMDGPU Plugin

With this patch we:
- pick more sensible defaults for the number of teams, inspired by the
  old plugin, and configured via LIBOMPTARGET_AMDGPU_TEAMS_PER_CU.
- check the input signal of a kernel launch late, after the queue lock
  was taken, to avoid a barrier packet more often.
- copy the kernel arguments in one swoop into the appropriate memory.
- manually specialize the callbacks to avoid potential indirect calls.

22 months ago[fuchsia] Tune PrimaryGroupSizeLog to reduce fragmentation.
Luke Nicholson [Tue, 20 Dec 2022 02:33:51 +0000 (02:33 +0000)]
[fuchsia] Tune PrimaryGroupSizeLog to reduce fragmentation.

21U is the default group size, which demonstrates ~15mb reduction
in heap size for some highly fragmented heaps on Fuchsia, and
a general 5mb savings when devices are under no load.

Microbenchmarks show no performance regressions, but most of our
benchmarks perform no significant mallocs. So we are choosing the
default setting, and monitoring for potential performance
issues.

Reviewed By: Chia-hungDuan

Differential Revision: https://reviews.llvm.org/D140311

22 months agoReapply "[OpenMP][FIX] Restrict more unsound assmptions about threading"
Johannes Doerfert [Sat, 17 Dec 2022 23:10:51 +0000 (15:10 -0800)]
Reapply "[OpenMP][FIX] Restrict more unsound assmptions about threading"

This reverts commit 3b052558125cbedf18c2ddb65780b50d6f437d54.

This patch got reverted due to an unrelated memory leak that has been
fixed.

22 months agoGlobalISel: Enable CSE of G_SELECT
Matt Arsenault [Mon, 21 Nov 2022 20:38:10 +0000 (15:38 -0500)]
GlobalISel: Enable CSE of G_SELECT

Stop trying to delete a select in one combine since it would
be deleting the CSE'd instruction if that happened.

22 months ago[gn build] Port 475ce4c200ca
LLVM GN Syncbot [Tue, 20 Dec 2022 01:52:50 +0000 (01:52 +0000)]
[gn build] Port 475ce4c200ca

22 months agoRFC: Uniformity Analysis for Irreducible Control Flow
Sameer Sahasrabuddhe [Tue, 20 Dec 2022 01:19:30 +0000 (06:49 +0530)]
RFC: Uniformity Analysis for Irreducible Control Flow

Uniformity analysis is a generalization of divergence analysis to
include irreducible control flow:

  1. The proposed spec presents a notion of "maximal convergence" that
     captures the existing convention of converging threads at the
     headers of natual loops.

  2. Maximal convergence is then extended to irreducible cycles. The
     identity of irreducible cycles is determined by the choices made
     in a depth-first traversal of the control flow graph. Uniformity
     analysis uses criteria that depend only on closed paths and not
     cycles, to determine maximal convergence. This makes it a
     conservative analysis that is independent of the effect of DFS on
     CycleInfo.

  3. The analysis is implemented as a template that can be
     instantiated for both LLVM IR and Machine IR.

Validation:
  - passes existing tests for divergence analysis
  - passes new tests with irreducible control flow
  - passes equivalent tests in MIR and GMIR

Based on concepts originally outlined by
Nicolai Haehnle <nicolai.haehnle@amd.com>

With contributions from Ruiling Song <ruiling.song@amd.com> and
Jay Foad <jay.foad@amd.com>.

Support for GMIR and lit tests for GMIR/MIR added by
Yashwant Singh <yashwant.singh@amd.com>.

Differential Revision: https://reviews.llvm.org/D130746

22 months ago[mlir][tensor][linalg] Add a pattern that generalizes tensor.unpack op.
Hanhan Wang [Fri, 16 Dec 2022 21:03:10 +0000 (13:03 -0800)]
[mlir][tensor][linalg] Add a pattern that generalizes tensor.unpack op.

The pattern generalizes a tensor::UnPackOp into a sequence of tensor +
Linalg ops, when the outer dims are all 1s. It uses the trick of
rank-reduced tensor.extract_slice to get the tile; transpose the tile;
extract sub tile for incomplete cases if needed; use tensor.insert_slice
to insert it to the destination tensor.

Reviewed By: tyb0807, chelini

Differential Revision: https://reviews.llvm.org/D140254

22 months ago[PowerPC][NFC] add explicit check for ctr loop pseudo expansion.
Chen Zheng [Mon, 19 Dec 2022 11:15:04 +0000 (06:15 -0500)]
[PowerPC][NFC] add explicit check for ctr loop pseudo expansion.

22 months ago[PowerPC] don't generate hardware loop.
Chen Zheng [Mon, 19 Dec 2022 10:58:09 +0000 (05:58 -0500)]
[PowerPC] don't generate hardware loop.

If the candidate loop already has hardware loop related intrinsics,
don't generate hardware loop on PPC. PPC does not support nested
hardware loops.

22 months ago[PowerPC][NFC] reuse a case for checking hardware loop intrinsic input
Chen Zheng [Mon, 19 Dec 2022 10:20:23 +0000 (05:20 -0500)]
[PowerPC][NFC] reuse a case for checking hardware loop intrinsic input

22 months ago[OpenMP][FIX] Ensure to inline `ompx::` functions after the rename in D140334
Johannes Doerfert [Tue, 20 Dec 2022 00:41:04 +0000 (16:41 -0800)]
[OpenMP][FIX] Ensure to inline `ompx::` functions after the rename in D140334

22 months ago[libc] Add support for standalone cross compilation of libc.
Siva Chandra Reddy [Sat, 17 Dec 2022 00:04:23 +0000 (00:04 +0000)]
[libc] Add support for standalone cross compilation of libc.

One should be able to do a cross build of the libc now. For example, using
clang on a x86_64 linux host, one can build for an aarch64 linux target by
specifying -DLIBC_TARGET_TRIPLE=aarch64-linux-gnu.

Follow up changes will add a baremetal config and also appropriate
documentation about cross compiling the libc for CPU targets.

Reviewed By: jhuber6

Differential Revision: https://reviews.llvm.org/D140351

22 months ago[OpenMP] Change the nextgen plugin kernel thread count scheme as old plugins'
Ye Luo [Tue, 20 Dec 2022 00:26:38 +0000 (18:26 -0600)]
[OpenMP] Change the nextgen plugin kernel thread count scheme as old plugins'

Reviewed By: jdoerfert

Differential Revision: https://reviews.llvm.org/D140352

22 months ago[gn build] Port 2916b9918275
LLVM GN Syncbot [Tue, 20 Dec 2022 00:03:21 +0000 (00:03 +0000)]
[gn build] Port 2916b9918275

22 months agoRevert "LTO: always parse modules in opaque pointer mode."
Steven Wu [Tue, 20 Dec 2022 00:02:18 +0000 (16:02 -0800)]
Revert "LTO: always parse modules in opaque pointer mode."

This reverts commit 8ba9a5218782fa4f94b5c516d513a4259992c254.

22 months ago[ADT] Alias llvm::Optional to std::optional
Benjamin Kramer [Mon, 19 Dec 2022 23:17:25 +0000 (00:17 +0100)]
[ADT] Alias llvm::Optional to std::optional

This avoids the continuous API churn when upgrading things to use
std::optional and makes trivial string replace upgrades possible.

I tested this with GCC 7.5, the oldest supported GCC I had around.

Differential Revision: https://reviews.llvm.org/D140332

22 months ago[NFC] use Optional instead of separate bool
Florian Mayer [Mon, 19 Dec 2022 21:46:19 +0000 (13:46 -0800)]
[NFC] use Optional instead of separate bool

22 months ago[Clang] Prepare for llvm::Optional becoming std::optional.
Benjamin Kramer [Mon, 19 Dec 2022 23:15:11 +0000 (00:15 +0100)]
[Clang] Prepare for llvm::Optional becoming std::optional.

The needed tweaks are mostly trivial, the one nasty bit is Clang's usage
of OptionalStorage. To keep this working old Optional stays around as
clang::CustomizableOptional, with the default Storage removed.
Optional<File/DirectoryEntryRef> is replaced with a typedef.

I tested this with GCC 7.5, the oldest supported GCC I had around.

Differential Revision: https://reviews.llvm.org/D140332

22 months ago[NFC] use bitwise or instead of addition
Florian Mayer [Mon, 19 Dec 2022 21:24:34 +0000 (13:24 -0800)]
[NFC] use bitwise or instead of addition

as the bits are all distinct, these two operations have the same result,
but the bitwise operation is more explicit about what's happening.

Reviewed By: hctim

Differential Revision: https://reviews.llvm.org/D140346

22 months ago[SDAG] neg x with only low bit demanded is x
Philip Reames [Mon, 19 Dec 2022 23:24:45 +0000 (15:24 -0800)]
[SDAG] neg x with only low bit demanded is x

We have a version of this transform in InstCombine, but surprisingly not in SDAG.  Even more surprisingly, this benefits RISCV, but no other target. This was surprising enough I double checked my build configuration to make sure all targets were enabled; they appear to be.

Differential Revision: https://reviews.llvm.org/D140324

22 months ago[ORC] Add JITDylib argument to ResourceManager notify-removing/transferring ops.
Lang Hames [Mon, 14 Nov 2022 06:12:09 +0000 (22:12 -0800)]
[ORC] Add JITDylib argument to ResourceManager notify-removing/transferring ops.

In some cases it's helpful to group trackers by JITDylib. E.g. Platform classes
may want to track initializer symbols with a `JITDylib -> Tracker -> [ Symbol ]`
map. This makes it easy to collect all symbols for the JITDylib, while still
allowing efficient removal of a single tracker. Passing the JITDylib as an
argument to ResourceManager::notifyRemovingResources and
ResourceManager::notifyTransferringResources supports such use-cases.

22 months ago[ORC] Add withResourceKeyDo method to ResourceTracker.
Lang Hames [Mon, 14 Nov 2022 00:41:04 +0000 (16:41 -0800)]
[ORC] Add withResourceKeyDo method to ResourceTracker.

This method behaves the same as MaterializationResponsibility::withResourceKeyDo
(which now forwards to the new method): It locks the session while providing
access to the ResourceKey associated with the tracker.

Adding this method to ResourceTracker allows resources to be allocated and
tracked for a given MaterializationUnit prior to that MaterializationUnit being
materialized. E.g. Platforms can now track and remove initializers and other
symbols.

22 months ago[OpenMP] Export `ompx::` symbols from the device runtime
Johannes Doerfert [Mon, 19 Dec 2022 19:25:05 +0000 (11:25 -0800)]
[OpenMP] Export `ompx::` symbols from the device runtime

Differential Revision: https://reviews.llvm.org/D140335

22 months ago[OpenMP] Rename the `_OMP` namespace in the device runtime to `ompx`
Johannes Doerfert [Mon, 19 Dec 2022 18:54:15 +0000 (10:54 -0800)]
[OpenMP] Rename the `_OMP` namespace in the device runtime to `ompx`

Differential Revision: https://reviews.llvm.org/D140334

22 months agoDrop unused forward declarations of llvm::Optional. NFC.
Benjamin Kramer [Mon, 19 Dec 2022 22:42:08 +0000 (23:42 +0100)]
Drop unused forward declarations of llvm::Optional. NFC.

22 months ago[mlir] Add operator!= to WalkResult, for completeness.
Benjamin Kramer [Mon, 19 Dec 2022 22:38:55 +0000 (23:38 +0100)]
[mlir] Add operator!= to WalkResult, for completeness.

22 months ago[mlir][linalg] Downscale 2D pooling with unit dimensions for height to 1D pooling
Murali Vijayaraghavan [Fri, 16 Dec 2022 04:51:01 +0000 (04:51 +0000)]
[mlir][linalg] Downscale 2D pooling with unit dimensions for height to 1D pooling

Differential Revision: https://reviews.llvm.org/D140187

22 months ago[OpenMP][NFC] Cleanup clause boilerplate
Johannes Doerfert [Mon, 19 Dec 2022 07:19:48 +0000 (23:19 -0800)]
[OpenMP][NFC] Cleanup clause boilerplate

- Add helper classes to avoid duplicated boilerplate.
- Add `final` for classes inheriting from those helpers.
- Fix a `struct` vs `class` mismatch.

Differential Revision: https://reviews.llvm.org/D140336

22 months ago[RISCV] Match neg (and x, 1) to two shifts to improve codesize
Philip Reames [Mon, 19 Dec 2022 21:56:01 +0000 (13:56 -0800)]
[RISCV] Match neg (and x, 1) to two shifts to improve codesize

The negate operation is never compressible (as the destination and rs1 register must differ). The two shift versions will be equal size if the input GPR is reused, or smaller if this is the only use of the input.

For clarity, the operation being performed is (select (low-bit-of x), -1, 0).

Differential Revision: https://reviews.llvm.org/D140319

22 months ago[mlir] Drop Optional::value() references from tblgen files
Benjamin Kramer [Mon, 19 Dec 2022 21:38:37 +0000 (22:38 +0100)]
[mlir] Drop Optional::value() references from tblgen files

std::optional::value() has undesired exception checking semantics and is
unavailable in older Xcode (see _LIBCPP_AVAILABILITY_BAD_OPTIONAL_ACCESS).
The call sites block std::optional migration.

22 months ago[mlir][sparse] minor merger code cleanup
Aart Bik [Mon, 19 Dec 2022 20:55:22 +0000 (12:55 -0800)]
[mlir][sparse] minor merger code cleanup

Moved larger constructor from header to CPP file.
Used toMLIRString() for DimLvlType debug.
Minor layout changes.

Reviewed By: Peiming

Differential Revision: https://reviews.llvm.org/D140342

22 months agoFix gcc 7.5 build break after TypeSize ctors became constexpr
Paul Robinson [Mon, 19 Dec 2022 21:16:04 +0000 (13:16 -0800)]
Fix gcc 7.5 build break after TypeSize ctors became constexpr

As seen here: https://lab.llvm.org/staging/#/builders/235/builds/1135

22 months ago[clang] Don't spuriously pass -stdlib=libc++ to CC1 on Darwin
Louis Dionne [Tue, 13 Dec 2022 15:10:35 +0000 (10:10 -0500)]
[clang] Don't spuriously pass -stdlib=libc++ to CC1 on Darwin

Previously, we would be passing down -stdlib=libc++ from the Driver
to CC1 whenever the default standard library on the platform was libc++,
even if -stdlib= had not been passed to the Driver. This meant that we
would pass -stdlib=libc++ in nonsensical circumstances, such as when
compiling C code.

This logic had been added in b534ce46bd40 to make sure that header
search paths were set up properly. However, since libc++ is now the
default Standard Library on Darwin, passing this explicitly is not
required anymore. Indeed, if no -stdlib= is specified, CC1 will end
up using libc++ if it queries which standard library to use, without
having to be told.

Not passing -stdlib= at all to CC1 on Darwin should become possible
once CC1 stops relying on it to set up framework search paths.

Furthermore, this commit also removes a diagnostic checking whether the
deployment target is too old to support libc++. Nowadays, all supported
deployment targets use libc++ and compiling with libstdc++ is not
supported anymore. The Driver was the wrong place to issue this
diagnostic since it doesn't know whether libc++ will actually be linked
against (e.g. C vs C++), which would lead to spurious diagnostics.
Given that these targets are not supported anymore, we simply drop
the diagnostic instead of trying to refactor it into CC1.

rdar://103198514

Differential Revision: https://reviews.llvm.org/D139938

22 months ago[clangd] Support type hints for `decltype(expr)`
v1nh1shungry [Mon, 19 Dec 2022 21:01:30 +0000 (16:01 -0500)]
[clangd] Support type hints for `decltype(expr)`

Reviewed By: nridge

Differential Revision: https://reviews.llvm.org/D138300

22 months ago[clang][dataflow][NFC] Fix comments related to widening.
Yitzhak Mandelbaum [Mon, 19 Dec 2022 14:47:35 +0000 (14:47 +0000)]
[clang][dataflow][NFC] Fix comments related to widening.

The comments describing the API for analysis `widen` and the environment `widen`
were overly strict in the preconditions they assumed for the operation. In
particular, both assumed that the previous value preceded the current value in
the relevant ordering. However, that's not generally how widen operators work
and widening itself can violate this property. That is, when the previous value
is the result of a widening, it can easily be "greater" than the current value.

This patch updates the comments to accurately reflect the expectations.

Differential Revision: https://reviews.llvm.org/D140308

22 months ago[AArch64] Add additional umull tests cases. NFC
David Green [Mon, 19 Dec 2022 21:01:01 +0000 (21:01 +0000)]
[AArch64] Add additional umull tests cases. NFC

22 months ago[NFC] fix some clang-tidy warnings for hwasan pass
Florian Mayer [Mon, 19 Dec 2022 20:56:08 +0000 (12:56 -0800)]
[NFC] fix some clang-tidy warnings for hwasan pass

22 months ago[LV] Convert some tests to use opaque pointers (NFC).
Florian Hahn [Mon, 19 Dec 2022 20:55:44 +0000 (20:55 +0000)]
[LV] Convert some tests to use opaque pointers (NFC).

22 months ago[LV] Convert some tests to use opaque pointers (NFC).
Florian Hahn [Mon, 19 Dec 2022 20:44:44 +0000 (20:44 +0000)]
[LV] Convert some tests to use opaque pointers (NFC).

22 months ago[libc] Fix full build check for the GPU build
Joseph Huber [Mon, 19 Dec 2022 20:42:01 +0000 (14:42 -0600)]
[libc] Fix full build check for the GPU build

Summary:
This condition should fire when the full build is set OFF instead of
when it is ON.

22 months ago[LV] Remove unused AAResults argument (NFC).
Florian Hahn [Mon, 19 Dec 2022 20:37:47 +0000 (20:37 +0000)]
[LV] Remove unused AAResults argument (NFC).

AAResults is passed to LoopVectorizationLegality but no longer used.
Remove the dead code.

22 months ago[mlir][AMDGPU] Make header build standalone. NFC.
Benjamin Kramer [Mon, 19 Dec 2022 20:10:19 +0000 (21:10 +0100)]
[mlir][AMDGPU] Make header build standalone. NFC.

22 months ago[ubsan] Convert tests to check 'target=...'
Paul Robinson [Mon, 19 Dec 2022 20:03:55 +0000 (12:03 -0800)]
[ubsan] Convert tests to check 'target=...'

Part of the project to eliminate special handling for triples in lit
expressions.

22 months agoReapply "[Attributor][FIX] Allow negative offsets for ranges"
Johannes Doerfert [Sat, 17 Dec 2022 21:15:43 +0000 (13:15 -0800)]
Reapply "[Attributor][FIX] Allow negative offsets for ranges"

This reverts commit d57a3443f3e2b423fd7f5402f017dc7c0dff8cdf.

This patch was never part of the memory leak problem that lead to the
revert, just an innocent bystander caught in the middle...

Also, added a second reproducer reported after the revert.

22 months ago[tsan] Convert tests to check 'target=...'
Paul Robinson [Mon, 19 Dec 2022 19:32:24 +0000 (11:32 -0800)]
[tsan] Convert tests to check 'target=...'

Part of the project to eliminate special handling for triples in lit
expressions.

22 months ago[FileCheck] Remove the last llvm::Optional
Benjamin Kramer [Mon, 19 Dec 2022 19:21:15 +0000 (20:21 +0100)]
[FileCheck] Remove the last llvm::Optional

It doesn't seem necessary and it's blocking std::optional.

22 months agoMake evaluation of nested requirement consistent with requires expr.
Utkarsh Saxena [Tue, 29 Nov 2022 13:05:53 +0000 (14:05 +0100)]
Make evaluation of nested requirement consistent with requires expr.

Fixes: https://github.com/llvm/llvm-project/issues/45563
```
template<class T>  concept True = true;

template <class T>
concept C1 = requires (T) {
   requires True<typename T::value> || True<T>;
};

template <class T>
constexpr bool foo()
requires True<typename T::value> || True<T> {
    return true;
}
static_assert(C1<double>); // Previously failed due to SFINAE error
static_assert(foo<int>()); // but this works fine.
```
The issue here is the discrepancy between how a [nested requirement is evaluated](https://github.com/llvm/llvm-project/blob/main/clang/lib/Sema/SemaTemplateInstantiate.cpp#L2331) Vs how a [non-nested requirement is evaluated](https://github.com/llvm/llvm-project/blob/main/clang/lib/Sema/SemaConcept.cpp#L167-L200).

This patch makes constraint checking consistent for nested requirement
and trailing requires expressions by reusing the same evaluator.

Differential Revision: https://reviews.llvm.org/D138914

22 months agoHIP: Directly call fabs builtins
Matt Arsenault [Sun, 20 Nov 2022 22:36:02 +0000 (14:36 -0800)]
HIP: Directly call fabs builtins

All the attribute numbering changes are due to the ocml function being
assumed convergent, and the direct call is not.

This begins a series of patches to avoid using the thinnest
ocml wrappers around llvm intrinsics. These wrapper functions
aren't buying anything and add complexity. As we do not propagate
fast math flags into the library on linking, and cannot generally
safely do so, we're losing information by calling into them. Directly
call the builtins so we get the fast math flags set appropriate for the
translation unit naturally.

22 months agoRevert "[AIX][BigArchive] Treat the archive is empty if the first child member offset...
Mitch Phillips [Mon, 19 Dec 2022 19:02:14 +0000 (11:02 -0800)]
Revert "[AIX][BigArchive] Treat the archive is empty if the first child member offset is zero"

This reverts commit b26fe2a3e520a520e785dc6b4e5758b7fde6af9d.

Reason: Broke the ASan build bots. See more info in the original
phabricator review: https://reviews.llvm.org/D138986

22 months ago[sanitizer-common] Convert tests to check 'target=...'
Paul Robinson [Mon, 19 Dec 2022 19:00:50 +0000 (11:00 -0800)]
[sanitizer-common] Convert tests to check 'target=...'

Part of the project to eliminate special handling for triples in lit
expressions.

22 months ago[lit] Fix a few issues in relative_lines.py
Haojian Wu [Mon, 19 Dec 2022 14:10:49 +0000 (15:10 +0100)]
[lit] Fix a few issues in relative_lines.py

- fixes error: Line: 91  result += contents[pos:m.start(index)] TypeError: can only concatenate str (not "bytes") to str
- fixes the "-code-completion-at=%s:%(lineb'-7')" rewritten results

Differential Revision: https://reviews.llvm.org/D140300

22 months agoRevert "Reland "[TargetLowering] Teach DemandedBits about VSCALE""
Saleem Abdulrasool [Mon, 19 Dec 2022 18:12:08 +0000 (18:12 +0000)]
Revert "Reland "[TargetLowering] Teach DemandedBits about VSCALE""

This reverts commit 3010f60381bcd828d1b409cfaa576328bcd05bbc.

This change introduced undefined behaviour (reported at
https://reviews.llvm.org/D138508#inline-1352840).  Additionally, it
appears to be responsible for a mis-compilation on RISCV64 with the
vector extension (https://github.com/llvm/llvm-project/issues/59594).
The commit message indicates that this is meant to be ARM64 specific
though is a generic selection change.

22 months ago[include-cleaner] Base-type usage from member exprs is implicit.
Haojian Wu [Mon, 19 Dec 2022 08:22:42 +0000 (09:22 +0100)]
[include-cleaner] Base-type usage from member exprs is implicit.

Per the discussion on https://reviews.llvm.org/D140095#inline-1352956

Differential Revision: https://reviews.llvm.org/D140284

22 months agoReland "[pgo] Avoid introducing relocations by using private alias"
Paul Kirth [Fri, 9 Dec 2022 04:35:23 +0000 (04:35 +0000)]
Reland "[pgo] Avoid introducing relocations by using private alias"

In many cases, we can use an alias to avoid a symbolic relocations,
instead of using the public, interposable symbol. When the instrumented
function is in a COMDAT, we can use a hidden alias, and still avoid
references to discarded sections.

New compiler-rt tests are Linux only for now.

Previous versions of this patch allowed the compiler to name the
generated alias, but that would only be valid when the functions were
local. Since the alias may be used across TUs we use a more
deterministic naming convention, and add a `.local` suffix to the alias
name just as we do for relative vtables aliases.

Reviewed By: phosek

Differential Revision: https://reviews.llvm.org/D137982

22 months ago[SPIR-V] Support memory(...) function attributes
Michal Paszkowski [Mon, 19 Dec 2022 18:36:32 +0000 (19:36 +0100)]
[SPIR-V] Support memory(...) function attributes

Adds support for memory(...) function attributes in SPIR-V function
control info lowering.

Differential Revision: https://reviews.llvm.org/D139133

22 months ago[SPIR-V] Disable MachineLateInstrsCleanup pass
Michal Paszkowski [Mon, 19 Dec 2022 18:05:30 +0000 (19:05 +0100)]
[SPIR-V] Disable MachineLateInstrsCleanup pass

This patch disables MachineLateInstrsCleanup pass (since it does not
work with virtual registers) and modifies LIT tests to run with -O0.

Differential Revision: https://reviews.llvm.org/D140103

22 months ago[X86] Fix SLM uops counts for WriteBitTestSetRegRMW instructions
Simon Pilgrim [Mon, 19 Dec 2022 11:23:02 +0000 (11:23 +0000)]
[X86] Fix SLM uops counts for WriteBitTestSetRegRMW instructions

The set/reset/complement RMW variants use +1uop compared to the BT read-only instructions

Based off llvm-exegesis captures, confirmed with Agner + uops.info

22 months ago[libc][gpu][NFC] Move CMake logic to prepare GPU build to a separate file.
Siva Chandra Reddy [Sat, 17 Dec 2022 17:02:00 +0000 (17:02 +0000)]
[libc][gpu][NFC] Move CMake logic to prepare GPU build to a separate file.

Reviewed By: jhuber6

Differential Revision: https://reviews.llvm.org/D140309

22 months ago[RISCV] Add codegen support for RISCV XVentanaCondOps Extension
Kautuk Consul [Mon, 19 Dec 2022 17:42:24 +0000 (09:42 -0800)]
[RISCV] Add codegen support for RISCV XVentanaCondOps Extension

This patch adds codegen support for part of XVentanaCondOps extension.
This extension is designed to reduce the number of branches in
the generated RISCV assembly by replacing branches with conditional
move instructions as defined by XVentanaCondOps specification.

The specification for XVentanaCondOps extension can be found at:
https://github.com/ventanamicro/ventana-custom-extensions/releases/download/v1.0.1/ventana-custom-extensions-v1.0.1.pdf

Co-authored-by: Mikhail Gudim <mgudim@ventanamicro.com>
Differential Revision: https://reviews.llvm.org/D139394

22 months agoFix: [DebugInfo] Add function to test debug values for equivalence
Stephen Tozer [Mon, 19 Dec 2022 17:53:46 +0000 (17:53 +0000)]
Fix: [DebugInfo] Add function to test debug values for equivalence

Fixes a test in 6d169089 which had unspecified platform dependencies
and platform-dependent output.

22 months agomlir/DialectConversion: use std::optional (NFC)
Ramkumar Ramachandra [Mon, 19 Dec 2022 13:25:14 +0000 (14:25 +0100)]
mlir/DialectConversion: use std::optional (NFC)

This is part of an effort to migrate from llvm::Optional to
std::optional. This patch touches DialectConversion, and modifies
existing conversions and tests appropriately.

See also: https://discourse.llvm.org/t/deprecating-llvm-optional-x-hasvalue-getvalue-getvalueor/63716

Signed-off-by: Ramkumar Ramachandra <r@artagnon.com>
Differential Revision: https://reviews.llvm.org/D140303

22 months ago[profile] Convert tests to check 'target=...'
Paul Robinson [Mon, 19 Dec 2022 17:43:59 +0000 (09:43 -0800)]
[profile] Convert tests to check 'target=...'

Part of the project to eliminate special handling for triples in lit
expressions.

22 months ago[DebugInfo] Add function to test debug values for equivalence
Stephen Tozer [Mon, 19 Dec 2022 15:44:16 +0000 (15:44 +0000)]
[DebugInfo] Add function to test debug values for equivalence

This patch adds a new function that can be used to check all the
properties, other than the machine values, of a pair of debug values for
equivalence. This is done by folding the "directness" into the
expression, converting the expression to variadic form if it is not
already in that form, and then comparing directly. In a few places which
check whether two debug values are identical to see if their ranges can
be merged, this function will correctly identify cases where two debug
values are expressed differently but have the same meaning, allowing
those ranges to be correctly merged.

Differential Revision: https://reviews.llvm.org/D136173

22 months agoReapply "GlobalISel: Make LLT constructors constexpr"
Matt Arsenault [Mon, 19 Dec 2022 00:29:59 +0000 (19:29 -0500)]
Reapply "GlobalISel: Make LLT constructors constexpr"

I initially attempted this in 5a95be22d248be654b992dfb25e3850dbb182a14. It
was reverted in 81cbe0ca83c2f912ff612ddb65629a108197b0d1 since it
crashed GCC 5.3. That has dropped out of the list of supported host
compilers, so try again.

22 months agoDAG: Pull fneg out of select feeding fadd into fsub
Simon Pilgrim [Thu, 15 Dec 2022 23:35:45 +0000 (18:35 -0500)]
DAG: Pull fneg out of select feeding fadd into fsub

Enables folding fadd x, (select c, (fneg a), (fneg b))
-> fsub (select a, b), c

Avoids some regressions in a future AMDGPU change.

22 months agoAMDGPU: Add regression tests for fmin/fmax legacy matching
Matt Arsenault [Thu, 15 Dec 2022 12:44:32 +0000 (07:44 -0500)]
AMDGPU: Add regression tests for fmin/fmax legacy matching

22 months ago[llvm][ADT] Fix Any with msvc and lto
Sebastian Neubauer [Mon, 19 Dec 2022 13:25:05 +0000 (14:25 +0100)]
[llvm][ADT] Fix Any with msvc and lto

llvm::Any had and has several bugs, so we eventually want to replace it
with std::any.

Unfortunately, we cannot do that right now because of bugs in the msvc
standard library that are only fixed in VS 2022 17.4.

When lto is enabled in msvc, constant symbols end up at the same
address, breaking the TypeId implementation of llvm::Any.
Make the TypeId<T>::Id non-const to fix this.

I was able to find an easy reproducer (tried in godbolt with
x64 msvc v19.32 and `/GL` as compiler flags to enable lto):

```c++

template <typename T> struct TypeId {
  // Remove the const here and below to make it work.
  static const char Id;
};

template <typename T> const char TypeId<T>::Id = 0;

template <typename A, typename B>
bool isSame() {
  return &TypeId<A>::Id == &TypeId<B>::Id;
}

class A {};
class B {};

int main() {
  // This should output "is same 0" because the addresses of A's and B's
  // TypeId::Id should be different.
  printf("is same %d\n", isSame<A, B>());
  return 0;
}
```

Differential Revision: https://reviews.llvm.org/D139974

22 months ago[LAA] Convert test to opaque pointers (NFC)
Nikita Popov [Fri, 16 Dec 2022 12:00:59 +0000 (13:00 +0100)]
[LAA] Convert test to opaque pointers (NFC)

When converting this test to opaque pointers (and dropping bitcast),
we get improved memory checks. Per fhahn:

> It looks like the difference is due to the logic that determines
> pointer strides in LAA not handling bitcasts. Without the
> bitcasts, the logic now triggers successfully.

Differential Revision: https://reviews.llvm.org/D140204

22 months ago[AArch64] Add alias predicate-as-counter register for PFALSE
Caroline.Concatto@arm.com [Mon, 19 Dec 2022 14:01:39 +0000 (14:01 +0000)]
[AArch64] Add alias predicate-as-counter register for PFALSE

According to:
https://developer.arm.com/documentation/ddi0602/2022-09/
PFALSE should:
"...an assembler must also accept predicate-as-counter register
name for the destination predicate register."

Differential Revision: https://reviews.llvm.org/D140301

22 months agoAMDGPU: Add more fneg combine tests
Matt Arsenault [Fri, 16 Dec 2022 01:09:08 +0000 (20:09 -0500)]
AMDGPU: Add more fneg combine tests

22 months ago[clang][dataflow] Remove stray lines from `Environment::join`
Yitzhak Mandelbaum [Mon, 19 Dec 2022 14:49:03 +0000 (14:49 +0000)]
[clang][dataflow] Remove stray lines from `Environment::join`

Removes an assertion and a useless line. The assertion seems left over from
earlier debugging and the line that follows is a stray line.

Differential Revision: https://reviews.llvm.org/D140306

22 months agoARM: Add baseline tests for fadd with select combine
Matt Arsenault [Fri, 16 Dec 2022 01:08:23 +0000 (20:08 -0500)]
ARM: Add baseline tests for fadd with select combine

22 months ago[clang-tidy][NFC] Replace custom isStatic matcher with the existing isStaticStorageClass
Carlos Galvez [Mon, 19 Dec 2022 15:22:47 +0000 (15:22 +0000)]
[clang-tidy][NFC] Replace custom isStatic matcher with the existing isStaticStorageClass

22 months ago[gn build] Port ca0f4e193b05
LLVM GN Syncbot [Mon, 19 Dec 2022 15:16:21 +0000 (15:16 +0000)]
[gn build] Port ca0f4e193b05

22 months ago[NFC][llvm-exegesis] Try to fix GCC7 build
Roman Lebedev [Mon, 19 Dec 2022 15:09:30 +0000 (18:09 +0300)]
[NFC][llvm-exegesis] Try to fix GCC7 build

https://lab.llvm.org/staging/#/builders/235/builds/1121

22 months ago[AArch64] Add missing v8.8a Non-maskable Interrupts feature
Lucas Prates [Thu, 15 Dec 2022 11:26:17 +0000 (11:26 +0000)]
[AArch64] Add missing v8.8a Non-maskable Interrupts feature

This adds support for the missing Non-maskable Interrupts (FEAT_NMI)
feature from armv8.8-A, which consists of the `ALLINT` pstate register.

This is a second iteration of the patch from D131389, building on top of
the D139925 changes that enable better support for `msr (immediate)`
instructions that take 1-bit immediates.

Contributors:
* David Candler
* Tomas Matheson
* Sam Elliott

Reviewed By: lenary, tmatheson

Differential Revision: https://reviews.llvm.org/D140216

22 months ago[AArch64] Add new v9.4-A PM pstate system register
Lucas Prates [Thu, 8 Dec 2022 15:01:06 +0000 (15:01 +0000)]
[AArch64] Add new v9.4-A PM pstate system register

This adds support for the new PM pstate system register introduced by
the v9.4-A Exception-based Event Profiling extension (FEAT_EBEP).

The new PM pstate register takes a 1-bit immediate and requires
different values to be specified for the higher bits of the Crm field.
To enable that, this patch creates an explicit separation between the
pstate system registers that take 4-bit and 1-bit immediate operands,
allowing each entry to specify the value for the 3 high bits of Crm.

This also updates other pstate registers to correctly accept 4-bit
immediates, matching their decoding specification from the Arm ARM.
These include: `PAN`,  `UAO`, `DIT` and `SSBS`.

More information about this extension and the new register can be found
at:
* https://developer.arm.com/documentation/ddi0601/2022-09/AArch64-Registers/PM--PMU-Exception-Mask

Contributors:
* Lucas Prates
* Sam Elliott

Reviewed By: lenary

Differential Revision: https://reviews.llvm.org/D139925

22 months ago[flang] Lowering and runtime support for F08 transformational intrinsics: BESSEL_JN...
Tarun Prabhu [Mon, 19 Dec 2022 14:51:59 +0000 (07:51 -0700)]
[flang] Lowering and runtime support for F08 transformational intrinsics: BESSEL_JN and BESSEL_YN

The runtime implementation uses the recurrence relations

`J(n-1, x) = (2.0 / x) * n * J(n, x) - J(n+1, x)`
`Y(n+1, x) = (2.0 / x) * n * Y(n, x) - Y(n-1, x)`

(see https://dlmf.nist.gov/10.74.iv and https://dlmf.nist.gov/10.6.E1).

Although the standard requires that `N1` and `N2` in `BESSEL_JN(N1, N2, x)`
and `BESSEL_YN(N1, N2, x)` be non-negative, this is not checked in the
runtime functions. This is in keeping with some other compilers which also
return some results when `N1` and/or `N2` are negative.

The special case for `x == 0` is  handled in different runtime functions
for each of `BESSEL_JN` and `BESSEL_YN`. The lowering code checks for this
case and inserts the checks and the appropriate runtime calls in FIR.

The existing tests for the two intrinsics was modified to keep the style
consistent with the additional lowering tests that were added.

22 months ago[LoopPeel] Expose ValueMap of last peeled iteration. NFC
Anna Thomas [Thu, 17 Nov 2022 18:40:33 +0000 (13:40 -0500)]
[LoopPeel] Expose ValueMap of last peeled iteration. NFC

The value map of last peeled iteration is computed within peelLoop API.
This patch exposes it for callers of peelLoop.
While this is not currently used by upstream passes, we have a usecase
downstream which benefits from this API update. Future users of peelLoop
can also use the ValueMap if needed.

Similar value maps are exposed by other loop utilities such as loop
cloning.

Differential Revision: https://reviews.llvm.org/D138228

22 months agoSupport: Make TypeSize constructors constexpr
Matt Arsenault [Mon, 19 Dec 2022 03:14:00 +0000 (22:14 -0500)]
Support: Make TypeSize constructors constexpr

This is to enable LLT to be constexpr

22 months ago[msan] Convert tests to check 'target=...'
Paul Robinson [Mon, 19 Dec 2022 14:44:27 +0000 (06:44 -0800)]
[msan] Convert tests to check 'target=...'

Part of the project to eliminate special handling for triples in lit
expressions.

22 months agoAMDGPU: Convert a few more special case tests to opaque pointers
Matt Arsenault [Mon, 19 Dec 2022 14:31:17 +0000 (09:31 -0500)]
AMDGPU: Convert a few more special case tests to opaque pointers

lower-kernargs.ll needed a switch to use update_test_checks metadata
matching.

22 months agoAMDGPU: Update some tests to use opaque pointers
Matt Arsenault [Mon, 19 Dec 2022 14:23:45 +0000 (09:23 -0500)]
AMDGPU: Update some tests to use opaque pointers

vectorize-buffer-fat-pointer.ll required a manual check line fix.
vector-alloca-addrspacecast.ll required a manual fixup of a check
line. partial-regcopy-and-spill-missed-at-regalloc.ll required
re-running update_mir_test_checks. The HSA metadata tests required
avoiding the script touching the type name in the metadata.

annotate-noclobber.ll ran into one update script bug. It deleted a
check line with a 0 offset GEP, moving the following -NEXT check
logically up one line.

22 months agoAMDGPU: Convert some amdgpu-codegenprepare tests to opaque pointers
Matt Arsenault [Mon, 5 Dec 2022 18:06:16 +0000 (13:06 -0500)]
AMDGPU: Convert some amdgpu-codegenprepare tests to opaque pointers

amdgpu-late-codegenprepare.ll required running update_test_checks
after converting.

22 months ago[lsan] Convert tests to check 'target=...'
Paul Robinson [Mon, 19 Dec 2022 14:24:50 +0000 (06:24 -0800)]
[lsan] Convert tests to check 'target=...'

Part of the project to eliminate special handling for triples in lit
expressions.

22 months ago[BPFAbstractMemberAccess] Use WeakTrackingVH for Base
Nikita Popov [Mon, 19 Dec 2022 13:47:28 +0000 (14:47 +0100)]
[BPFAbstractMemberAccess] Use WeakTrackingVH for Base

The value will be RAUWd, make sure the reference in CallInfo gets
updated.

It seems like this was not a problem without opaque pointers due
to the bitcast in between.

22 months ago[include-cleaner] Use expansion locations for macros.
Viktoriia Bakalova [Thu, 8 Dec 2022 10:20:14 +0000 (10:20 +0000)]
[include-cleaner] Use expansion locations for macros.

Use expansion locations for target symbol decls when finding headers for macros.
Fix: https://github.com/llvm/llvm-project/issues/59392
Differential Revision: https://reviews.llvm.org/D139716

22 months ago[NFC][llvm-exegesis] Try to fix GCC7 build
Roman Lebedev [Mon, 19 Dec 2022 14:14:55 +0000 (17:14 +0300)]
[NFC][llvm-exegesis] Try to fix GCC7 build

https://lab.llvm.org/staging/#/builders/235/builds/1090