platform/upstream/llvm.git
2 years ago[pseudo] A basic implementation of compiling cxx grammar at build time.
Haojian Wu [Tue, 24 May 2022 18:21:45 +0000 (20:21 +0200)]
[pseudo] A basic implementation of compiling cxx grammar at build time.

The main idea is to compile the cxx grammar at build time, and construct
the core pieces (Grammar, LRTable) of the pseudoparse based on the compiled
data sources.

This is a tiny implementation, which is good for start:

- defines how the public API should look like;
- integrates the cxx grammar compilation workflow with the cmake system.
- onlynonterminal symbols of the C++ grammar are compiled, anything
  else are still doing the real compilation work at runtime, we can opt-in more
  bits in the future;
- splits the monolithic clangPsuedo library for better layering;

Reviewed By: sammccall

Differential Revision: https://reviews.llvm.org/D125667

2 years ago[mlir] Fix warning `missing base in copy ctor`
Javed Absar [Mon, 23 May 2022 13:48:57 +0000 (14:48 +0100)]
[mlir] Fix warning `missing base in copy ctor`

This suppresse annoying warning when building mlir.
```
 warning: base class ‘class mlir::PassWrapper<{anonymous}::TestStatisticPass,
  mlir::OperationPass<void> >’ should be explicitly initialized in the copy constructor [-Wextra]
```

Reviewed By: rriddle

Differential Revision: https://reviews.llvm.org/D126209

2 years ago[XCORE][CodeGen][NFC] Revert: Use ArrayRef in TargetLowering functions
Nigel Perks [Wed, 25 May 2022 09:09:59 +0000 (10:09 +0100)]
[XCORE][CodeGen][NFC] Revert: Use ArrayRef in TargetLowering functions

Revert 6365bde6585651b7813010dc63df7984a5b8ad41

Restore more readable version in line with other targets,
which did not apply the change.

Differential Revision: https://reviews.llvm.org/D123661

2 years ago[OpenCL] Remove argument names from async copy builtins
Sven van Haastregt [Wed, 25 May 2022 09:05:25 +0000 (10:05 +0100)]
[OpenCL] Remove argument names from async copy builtins

This simplifies completeness comparisons against OpenCLBuiltins.td and
also makes the header no longer "claim" the argument name identifiers.

Continues the direction set out in D119560.

2 years ago[analyzer] Fix symbol simplification assertion failure
Gabor Marton [Tue, 24 May 2022 08:01:18 +0000 (10:01 +0200)]
[analyzer] Fix symbol simplification assertion failure

Fixes https://github.com/llvm/llvm-project/issues/55546

The assertion mentioned in the issue is triggered because an
inconsistency is formed in the Sym->Class and Class->Sym relations. A
simpler but similar inconsistency is demonstrated here:
https://reviews.llvm.org/D114887 .

Previously in `removeMember`, we didn't remove the old symbol's
Sym->Class relation. Back then, we explained it with the following two
bullet points:
> 1) This way constraints for the old symbol can still be found via it's
> equivalence class that it used to be the member of.
> 2) Performance and resource reasons. We can spare one removal and thus one
> additional tree in the forest of `ClassMap`.

This patch do remove the old symbol's Sym->Class relation in order to
keep the Sym->Class relation consistent with the Class->Sym relations.
Point 2) above has negligible performance impact, empirical measurements
do not show any noticeable difference in the run-time. Point 1) above
seems to be a not well justified statement. This is because we cannot
create a new symbol that would be equal to the old symbol after the
simplification had happened. The reason for this is that the SValBuilder
uses the available constant constraints for each sub-symbol.

Differential Revision: https://reviews.llvm.org/D126281

2 years ago[gn build] Port 7af89a379cce
LLVM GN Syncbot [Wed, 25 May 2022 08:31:45 +0000 (08:31 +0000)]
[gn build] Port 7af89a379cce

2 years ago[gn build] Port 29a5a7c6d47a
LLVM GN Syncbot [Wed, 25 May 2022 08:31:44 +0000 (08:31 +0000)]
[gn build] Port 29a5a7c6d47a

2 years ago[VPlan] Use MapVector for LiveOuts for deterministic iteration.
Florian Hahn [Wed, 25 May 2022 08:29:52 +0000 (09:29 +0100)]
[VPlan] Use MapVector for LiveOuts for deterministic iteration.

During code-gen, we iterate over the LiveOuts and the differences in
iteration order can cause slightly different outputs.

2 years ago[libc++] Implement ranges::fill{, _n}
Nikolas Klauser [Sat, 21 May 2022 16:26:29 +0000 (18:26 +0200)]
[libc++] Implement ranges::fill{, _n}

Reviewed By: var-const, #libc

Spies: libcxx-commits, mgorny

Differential Revision: https://reviews.llvm.org/D123462

2 years ago[RISCV] Add pre-emit pass to make more instructions compressible
Lewis Revill [Mon, 25 Apr 2022 11:24:09 +0000 (12:24 +0100)]
[RISCV] Add pre-emit pass to make more instructions compressible

When optimizing for size, this pass searches for instructions that are
prevented from being compressed by one of the following:

1. The use of a single uncompressed register.
2. A base register + offset where the offset is too large to be
   compressed and the base register may or may not already be compressed.

In the first case, if there is a compressed register available, then the
uncompressed register is copied to the compressed register and its uses
replaced. This is only done if there are enough uses that code size
would be improved.

In the second case, if a compressed register is available, then the
original base register is copied and adjusted such that:

new_base_register = base_register + adjustment
base_register + large_offset = new_base_register + small_offset

and the uses of the base register are replaced with the new base
register. Again this is only done if there are enough uses for code size
to be improved.

This pass was authored by Lewis Revill, with large offset optimization
added by Craig Blackmore.

Differential Revision: https://reviews.llvm.org/D92105

2 years ago[mlir] Add `complex.atan2` operation.
Alexander Belyaev [Wed, 25 May 2022 07:58:00 +0000 (09:58 +0200)]
[mlir] Add `complex.atan2` operation.

Differential Revision: https://reviews.llvm.org/D126357

2 years ago[ValueTracking] Loads with !dereferenceable metadata cannot be undef/poison
Nikita Popov [Tue, 24 May 2022 13:20:01 +0000 (15:20 +0200)]
[ValueTracking] Loads with !dereferenceable metadata cannot be undef/poison

A load with !dereferenceable or !dereferenceable_or_null metadata
must return a well-defined (non-undef/poison) value. Effectively
they imply !noundef. This is the same as we do for the
dereferenceable(N) attribute.

This should fix https://github.com/llvm/llvm-project/issues/55672,
or at least the specific case discussed there.

Differential Revision: https://reviews.llvm.org/D126296

2 years ago[libc++] Make sure that all headers can be included with modules enabled
Louis Dionne [Tue, 10 May 2022 19:27:52 +0000 (15:27 -0400)]
[libc++] Make sure that all headers can be included with modules enabled

This commit ensures that we can include all libc++ headers with modules
enabled. It adds a test to ensure that this doesn't regress, which is
necessary because our modules CI job does not build in all Standard modes.

Differential Revision: https://reviews.llvm.org/D125331

2 years agoFix warning by handling OMPC_fail in switch statement.
Adrian Kuegel [Wed, 25 May 2022 07:33:41 +0000 (09:33 +0200)]
Fix warning by handling OMPC_fail in switch statement.

2 years ago[Clang][OpenMP] Support for omp nothing
Sunil Kuravinakop [Wed, 25 May 2022 04:59:19 +0000 (23:59 -0500)]
[Clang][OpenMP] Support for omp nothing

Patch to support "#pragma omp nothing"

Reviewed By: tianshilei1992

Differential Revision: https://reviews.llvm.org/D123286

2 years ago[OpenMP] atomic compare fail : Parser & AST support
Sunil Kuravinakop [Wed, 25 May 2022 04:55:08 +0000 (23:55 -0500)]
[OpenMP] atomic compare fail : Parser & AST support

This is a support for " #pragma omp atomic compare fail ". It has Parser & AST support for now.

Reviewed By: tianshilei1992

Differential Revision: https://reviews.llvm.org/D123235

2 years ago[RISCV] Fix vnsrl/vnsra isel patterns that are dropping VL.
Craig Topper [Wed, 25 May 2022 02:10:40 +0000 (19:10 -0700)]
[RISCV] Fix vnsrl/vnsra isel patterns that are dropping VL.

We were incorrectly using VLMax instead of the passed VL.

Reviewed By: khchen, reames

Differential Revision: https://reviews.llvm.org/D126319

2 years ago[flang][MSVC] Fix building with `/permissive-` flag
Mehdi Chinoune [Mon, 23 May 2022 21:25:41 +0000 (16:25 -0500)]
[flang][MSVC] Fix building with `/permissive-` flag

CLOCK_REALTIME is POSIX defined and never available with MSVC, even without /permissive-.
The difference is that the template is never instantiated and the compiler ignores the undefined identifier.

Reviewed By: Meinersbur

Differential Revision: https://reviews.llvm.org/D125262

2 years ago[RISCV] Replace untested code with assert
Fraser Cormack [Tue, 24 May 2022 06:20:55 +0000 (07:20 +0100)]
[RISCV] Replace untested code with assert

We found untested code where negative frame indices were ostensibly
handled despite it being in a block guarded by !MFI.isFixedObjectIndex.

While the implementation of MachineFrameInfo::isFixedObjectIndex
suggests this is possible (i.e., if a frame index was more negative - less than the
number of fixed objects), I couldn't find any test in tree -- for any
target -- where a negative frame index wasn't also a fixed object
offset. I couldn't find a way of creating such a object with the
public MachineFrameInfo creation APIs. Even
MachineFrameInfo::getObjectIndexBegin starts counting at the negative
number of fixed objects, so such frame indices wouldn't be covered by
loops using the provided begin/end methods.

Given all this, an assert that any object encountered in the block is
non-negative seems reasonable.

Reviewed By: StephenFan, kito-cheng

Differential Revision: https://reviews.llvm.org/D126278

2 years agoMove GCC-compatible pod-packing change to v15/old behavior available at v14 and below
David Blaikie [Tue, 24 May 2022 22:14:56 +0000 (22:14 +0000)]
Move GCC-compatible pod-packing change to v15/old behavior available at v14 and below

Since this didn't make it into the v14 release - anyone requesting the
v14 ABI shouldn't get this GCC-compatible change that isn't backwards
compatible with v14 Clang.

Differential Revision: https://reviews.llvm.org/D126334

2 years ago[libc++][NFC] Add more tests to `move_{iterator,sentinel}`.
Konstantin Varlamov [Wed, 25 May 2022 02:56:02 +0000 (19:56 -0700)]
[libc++][NFC] Add more tests to `move_{iterator,sentinel}`.

More test coverage for the parts added by the One Ranges Proposal.

Differential Revision: https://reviews.llvm.org/D124906

2 years agoRevert "[MachineSink] replace MachineLoop with MachineCycle"
Chen Zheng [Wed, 25 May 2022 02:43:37 +0000 (22:43 -0400)]
Revert "[MachineSink] replace MachineLoop with MachineCycle"

This reverts commit 62a9b36fcf728b104ea87e6eb84c0be69b779df7.
Cause build failure on lldb incremental buildbot:
https://green.lab.llvm.org/green/view/LLDB/job/lldb-cmake/43994/changes

2 years ago[C++20] [Coroutines] Conform the updates for CWG issue 2585
Chuanqi Xu [Wed, 25 May 2022 02:30:32 +0000 (10:30 +0800)]
[C++20] [Coroutines] Conform the updates for CWG issue 2585

According to the updates in CWG issue 2585
https://cplusplus.github.io/CWG/issues/2585.html, we shouldn't find an
allocation function with (size, p0, …, pn) in global scope.

Reviewed By: erichkeane

Differential Revision: https://reviews.llvm.org/D126187

2 years ago[InstCombine] [NFC] Move transforms for truncated shifts into narrowBinOp
Chenbing Zheng [Wed, 25 May 2022 02:21:39 +0000 (10:21 +0800)]
[InstCombine] [NFC] Move transforms for truncated shifts into narrowBinOp

Reviewed By: spatel

Differential Revision: https://reviews.llvm.org/D126056

2 years ago[InstCombine] add test for trunc-shl-trunc ; NFC
Chenbing Zheng [Wed, 25 May 2022 02:09:58 +0000 (10:09 +0800)]
[InstCombine] add test for trunc-shl-trunc ; NFC

2 years ago[clang-format][NFC] Insert/remove braces in clang/lib/Format/
owenca [Sun, 22 May 2022 04:51:19 +0000 (21:51 -0700)]
[clang-format][NFC] Insert/remove braces in clang/lib/Format/

Differential Revision: https://reviews.llvm.org/D126157

2 years ago[BOLT][NFC] Use ListSeparator in BinaryFunction print methods
Amir Ayupov [Wed, 25 May 2022 01:28:42 +0000 (18:28 -0700)]
[BOLT][NFC] Use ListSeparator in BinaryFunction print methods

Reviewed By: rafauler

Differential Revision: https://reviews.llvm.org/D126243

2 years ago[BOLT][NFC] Customize endline character for printInstruction(s)
Amir Ayupov [Wed, 25 May 2022 01:25:40 +0000 (18:25 -0700)]
[BOLT][NFC] Customize endline character for printInstruction(s)

This would be used in `BF::dumpGraph` to dump left-justified text.

Reviewed By: rafauler

Differential Revision: https://reviews.llvm.org/D126232

2 years ago[BOLT][NFC] Use for_each to simplify printLoopInfo
Amir Ayupov [Wed, 25 May 2022 01:04:42 +0000 (18:04 -0700)]
[BOLT][NFC] Use for_each to simplify printLoopInfo

Reviewed By: rafauler

Differential Revision: https://reviews.llvm.org/D126242

2 years ago[BOLT][NFC] Remove unused BF::computeLocalUDChain method definition
Amir Ayupov [Wed, 25 May 2022 01:01:20 +0000 (18:01 -0700)]
[BOLT][NFC] Remove unused BF::computeLocalUDChain method definition

The function is only used inside AArch64MCPlusBuilder class, there are no uses
of it as a BinaryFunction method.

Reviewed By: rafauler

Differential Revision: https://reviews.llvm.org/D126220

2 years ago[lldb] Improve TestAppleSimulatorOSType.py error message
Jonas Devlieghere [Wed, 25 May 2022 00:17:26 +0000 (17:17 -0700)]
[lldb] Improve TestAppleSimulatorOSType.py error message

This was inspired by D109336 which got reverted because we didn't want
the test to fail silently. This patch prints a more informative error
message when we fail to parse the simctl output while still failing the
test.

Differential revision: https://reviews.llvm.org/D126217

2 years ago[lldb] Disable modules in Apple-lldb-base
Jonas Devlieghere [Wed, 25 May 2022 00:07:36 +0000 (17:07 -0700)]
[lldb] Disable modules in Apple-lldb-base

The LLDB website recommends using the CMake caches to build on macOS.
Although modules result in a faster build, this configuration tends to
break occasionally because it's specific to our platform. I don't expect
newcomers to be able to deal with those kind of breakages so don't
enable them by default.

2 years ago[clang-tidy] Fix #55134 (regression introduced by 5da7c04)
Salman Javed [Tue, 24 May 2022 22:15:12 +0000 (10:15 +1200)]
[clang-tidy] Fix #55134 (regression introduced by 5da7c04)

5da7c04 introduced a regression in the NOLINT macro checking loop, replacing the
call to `getImmediateExpansionRange().getBegin()` with
`getImmediateMacroCallerLoc()`, which has similar but subtly different
behaviour.

The consequence is that NOLINTs cannot suppress diagnostics when they are
attached to a token that came from a macro **argument**, rather than elsewhere
in the macro expansion.

Revert to pre-patch behaviour and add test cases to cover this issue.

Differential Revision: https://reviews.llvm.org/D126138

2 years ago[DFSan] Add option to specify individual library files, and an option to exit with...
Andrew Browne [Tue, 24 May 2022 21:42:41 +0000 (14:42 -0700)]
[DFSan] Add option to specify individual library files, and an option to exit with an error code if any library file was not found.

Reviewed By: vitalybuka

Differential Revision: https://reviews.llvm.org/D126336

2 years ago[LLVM][IR] Fix assert in ConstantExpr::getPtrToInt so all vector types are supported.
Paul Walker [Tue, 24 May 2022 22:49:59 +0000 (23:49 +0100)]
[LLVM][IR] Fix assert in ConstantExpr::getPtrToInt so all vector types are supported.

Fixes: #55410

2 years ago[mlir][sparse] complex lowering
Aart Bik [Tue, 24 May 2022 22:17:39 +0000 (15:17 -0700)]
[mlir][sparse] complex lowering

Reviewed By: bixia

Differential Revision: https://reviews.llvm.org/D126335

2 years ago[InstCombine] Fold memrchr calls with sequences of identical bytes.
Martin Sebor [Tue, 24 May 2022 23:00:11 +0000 (17:00 -0600)]
[InstCombine] Fold memrchr calls with sequences of identical bytes.

Reviewed By: nikic

Differential Revision: https://reviews.llvm.org/D123631

2 years ago[libc] Add the pthread_mutex_t type.
Siva Chandra Reddy [Fri, 20 May 2022 06:44:35 +0000 (06:44 +0000)]
[libc] Add the pthread_mutex_t type.

Simple implementations of the functions pthread_mutex_init,
pthread_mutex_destroy, pthread_mutex_lock and pthread_mutex_unlock have
have also been added. Future patches will extend these functions to add
features required by the POSIX specification.

Reviewed By: lntue

Differential Revision: https://reviews.llvm.org/D126235

2 years ago[OpenMP] Add codegen for 'omp_all_memory' reserved locator.
Mike Rice [Tue, 24 May 2022 18:48:42 +0000 (11:48 -0700)]
[OpenMP] Add codegen for 'omp_all_memory' reserved locator.

This creates an entry with address=nullptr and flag=0x80.
When an 'omp_all_memory' entry is specified any other 'out' or
'inout' entries are not needed and are not passed to the runtime.

Differential Revision: https://reviews.llvm.org/D126321

2 years ago[RISCV] Ensure the forwarded AVL register is alive
Philip Reames [Tue, 24 May 2022 21:58:09 +0000 (14:58 -0700)]
[RISCV] Ensure the forwarded AVL register is alive

When the AVL value does not fit in 5 bits, the register in which this value is stored may be dead when we want to forward it. This patch ensure the kill flags on the register are cleared before forwarding.

Patch by: loralb
Differential Revision: https://reviews.llvm.org/D125971

2 years ago[mlir][sparse] add new complex ops to reduction recognition
Aart Bik [Tue, 24 May 2022 18:49:47 +0000 (11:49 -0700)]
[mlir][sparse] add new complex ops to reduction recognition

Reviewed By: bixia

Differential Revision: https://reviews.llvm.org/D126318

2 years ago[RISCV] Add an operand kind to the opcode/imm returned from RISCVMatInt.
Craig Topper [Tue, 24 May 2022 21:54:57 +0000 (14:54 -0700)]
[RISCV] Add an operand kind to the opcode/imm returned from RISCVMatInt.

Instead of matching opcodes to know the format to emit, use an
enum value that we can get from the RISCVMatInt::Inst class.

Change the consumers to use fully covered switches so that we get
a compiler warning if a new kind is added. With the opcode checks
it was easier to forget to update one of the 3 consumers.

Reviewed By: reames

Differential Revision: https://reviews.llvm.org/D126317

2 years ago[riscv] Use getFirstInstrTerminator [nfc]
Philip Reames [Tue, 24 May 2022 21:55:42 +0000 (14:55 -0700)]
[riscv] Use getFirstInstrTerminator [nfc]

2 years ago[RISCV] Hoist VSETVLI out of idiomatic fixed length vector loops
Philip Reames [Tue, 24 May 2022 21:17:43 +0000 (14:17 -0700)]
[RISCV] Hoist VSETVLI out of idiomatic fixed length vector loops

This patch teaches the VSETVLI insertion pass to perform a very limited form of partial redundancy elimination. The motivating example comes from the fixed length vectorization of a simple loop such as:

for (unsigned i = 0; i < a_len; i++)
    a[i] += b;

Without this change, the core vector loop and preheader is as follows:

.LBB0_3:                                # %vector.ph
andi a1, a6, -8
addi a4, a0, 16
mv a5, a1
.LBB0_4:                                # %vector.body
                                        # =>This Inner Loop Header: Depth=1
addi a3, a4, -16
vsetivli zero, 4, e32, m1, ta, mu
vle32.v v8, (a3)
vle32.v v9, (a4)
vadd.vx v8, v8, a2
vadd.vx v9, v9, a2
vse32.v v8, (a3)
vse32.v v9, (a4)
addi a5, a5, -8
addi a4, a4, 32
bnez a5, .LBB0_4

The key thing to note here is that, the execution of the vsetivli only needs to happen once. Since there's no tail folding happening here, the value of the vector configuration registers are invariant through the loop.

After this patch, we hoist the configuration into the preheader and perform it once.

.LBB0_3:                                # %vector.ph
andi a1, a6, -8
vsetivli zero, 4, e32, m1, ta, mu
addi a4, a0, 16
mv a5, a1
.LBB0_4:                                # %vector.body
                                        # =>This Inner Loop Header: Depth=1
addi a3, a4, -16
vle32.v v8, (a3)
vle32.v v9, (a4)
vadd.vx v8, v8, a2
vadd.vx v9, v9, a2
vse32.v v8, (a3)
vse32.v v9, (a4)
addi a5, a5, -8
addi a4, a4, 32
bnez a5, .LBB0_4

Differential Revision: https://reviews.llvm.org/D124869

2 years ago[SelectionDAG] Add support to widen ISD::STEP_VECTOR operations.
Paul Walker [Sun, 22 May 2022 12:49:01 +0000 (13:49 +0100)]
[SelectionDAG] Add support to widen ISD::STEP_VECTOR operations.

Fixes: #55165

Differential Revision: https://reviews.llvm.org/D126168

2 years ago[Flang][OpenMP] Fixes for unstructured OpenMP code
Kiran Chandramohan [Tue, 24 May 2022 21:31:57 +0000 (21:31 +0000)]
[Flang][OpenMP] Fixes for unstructured OpenMP code

Since the FIR operations are mostly structured, it is only the functions
that could contain multiple blocks inside an operation. This changes
with OpenMP since OpenMP regions can contain multiple blocks. For
unstructured code, the blocks are created in advance and belong to the
top-level function. This caused code in OpenMP region to be placed under
the function level.

In this fix, if the OpenMP region is unstructured then new blocks are
created inside it.

Note1: This is part of upstreaming from the fir-dev branch of
https://github.com/flang-compiler/f18-llvm-project. The code in this patch is a
subset of the changes in https://github.com/flang-compiler/f18-llvm-project/pull/1178.

Reviewed By: vdonaldson

Differential Revision: https://reviews.llvm.org/D126293

Co-authored-by: Val Donaldson <vdonaldson@nvidia.com>
Co-authored-by: Eric Schweitz <eschweitz@nvidia.com>
Co-authored-by: Valentin Clement <clementval@gmail.com>
2 years agoFix build failure revealed by c35ca3a1c78f693b749ad11742350b7fc6c5cd89
Amy Kwan [Tue, 24 May 2022 21:11:23 +0000 (16:11 -0500)]
Fix build failure revealed by c35ca3a1c78f693b749ad11742350b7fc6c5cd89

This commit resolves a Linux kernel build failure that was revealed by
c35ca3a1c78f693b749ad11742350b7fc6c5cd89. The patch introduces two new
intrinsics, which ultimately changes the intrinsic numbering of other PPC
intrinsics. This causes an issue introduced by
ff40fb07ad6309131c2448ca00572a078c7a2d59, as the patch checks for intrinsics
with particular values, but the addition of the fnabs/fnabss intrinsics updates
the original sqrt/sdiv intrinsic values.

2 years ago[flang][runtime] Catch decimal integer input overflow
Peter Klausler [Fri, 20 May 2022 23:16:09 +0000 (16:16 -0700)]
[flang][runtime] Catch decimal integer input overflow

B/O/Z input overflow is already caught, and real input overflow
is signalled as an IEEE arithmetic exception, but regular decimal
integer overflow was silent.

Differential Revision: https://reviews.llvm.org/D126155

2 years ago[flang] Replace crash and improve a semantics TODO message
Peter Klausler [Fri, 20 May 2022 21:21:59 +0000 (14:21 -0700)]
[flang] Replace crash and improve a semantics TODO message

The derived type information table construction code had a
crash whose root cause was replacing an expression with one
of its operands -- the deletion of the LHS of that assignment
led to the RHS being invalidated before it could be read.
Fix by cloning the RHS.  Also update a TODO message to the
new "_todo_en_US" message class and add a comment about how
it should be resolved.

Differential Revision: https://reviews.llvm.org/D126154

2 years ago[flang] Process subprogram BIND(C,NAME=...) locally
Peter Klausler [Fri, 20 May 2022 20:31:14 +0000 (13:31 -0700)]
[flang] Process subprogram BIND(C,NAME=...) locally

The scalar-default-character-expression that defines the interoperable
name of a function or subroutine (or interface) must have its names
resolved within the context of the subprogram, despite its appearance
on a function-stmt or a subroutine-stmt.  Failure to do so can lead
to bogus errors or to incorrect results.

The solution is to defer name resolution for function-stmt suffixes
(but not entry-stmt suffixes) and for subroutine-stmt language binding
specifications to EndSubprogram().  (Their resolution only need to be
deferred to the end of the specification part, but it's cleanest to
deal with it in EndSubprogram().)

Differential Revision: https://reviews.llvm.org/D126153

2 years ago[clang][dataflow] Relax `Environment` comparison operation.
Yitzhak Mandelbaum [Tue, 24 May 2022 18:45:59 +0000 (18:45 +0000)]
[clang][dataflow] Relax `Environment` comparison operation.

Ignore `MemberLocToStruct` in environment comparison. As an ancillary data
structure, including it is redundant. We also can generate environments which
differ in their `MemberLocToStruct` but are otherwise equivalent.

Differential Revision: https://reviews.llvm.org/D126314

2 years ago[flang] Fix false error for multiple defined I/O subroutines
Peter Klausler [Fri, 20 May 2022 15:45:46 +0000 (08:45 -0700)]
[flang] Fix false error for multiple defined I/O subroutines

User-defined derived type I/O subroutines need to be unique for
a given type and operation in any scope, but it is acceptable
to have more than one defined I/O subroutine so long as only one
of them is visible.

Differential Revision: https://reviews.llvm.org/D126152

2 years ago[flang] Accept defined assignment with CLASS(*) RHS
Peter Klausler [Thu, 19 May 2022 23:30:04 +0000 (16:30 -0700)]
[flang] Accept defined assignment with CLASS(*) RHS

A utility predicate in semantics was incorrectly determining that
an INTERFACE ASSIGNMENT(=) (or other form of generic) could not have
a specific procedure with an unlimited polymorphic second argument.
This led to a crash later in expression analysis.  Fix, and
extend tests.

Differential Revision: https://reviews.llvm.org/D126151

2 years ago[flang] Fix purity testing for generic calls
Peter Klausler [Thu, 19 May 2022 22:32:06 +0000 (15:32 -0700)]
[flang] Fix purity testing for generic calls

The purity or impurity of a call to a generic interface
depends on the attributes of the specific procedure or specific
binding.  Change expression analysis of calls to generic interfaces
to replace the symbol in the parse tree with the specific procedure
or binding; this ensures that later checking for purity in
DO CONCURRENT and other contexts will be accurate.

Remove an "XFAIL" from a test that now passes again with this fix.

Differential Revision: https://reviews.llvm.org/D126150

2 years ago[clang][test] mark tests added in ee8524087c78 as unsupported on AIX
Alex Lorenz [Tue, 24 May 2022 20:38:04 +0000 (13:38 -0700)]
[clang][test] mark tests added in ee8524087c78 as unsupported on AIX

These tests are failing on the PPC64 AIX CI bot, but it's unclear why,
as they pass on other CI jobs.
I marked them as unsupported on AIX for now while investigating the failure.

2 years ago[flang] Don't prematurely resolve subprogram names
Peter Klausler [Thu, 19 May 2022 21:17:17 +0000 (14:17 -0700)]
[flang] Don't prematurely resolve subprogram names

Name resolution for subprograms checks whether the name is already
present in the enclosing scope as a generic interface, so that the
case of a generic with the same name as one of its specifics can be
handled.  The particular means by which the enclosing scope is searched
for the name would resolve the name (bind a symbol to it) as a side
effect.  This turns out to be the wrong thing to do when the subprogram
is going to have its symbol created in another scope to cope with its
BIND(C,NAME="name") name, and its Fortran name is already present in the
enclosing scope for a subprogram of the same name but without
BIND(C,NAME="name").

A very long explanation for a one-line fix, sorry.  In short, change
the code to look up the name but not resolve it at that point.

Differential Revision: https://reviews.llvm.org/D126149

2 years ago[clang][dataflow] Make limit on fixpoint-algorithm iterations proportional to size...
Yitzhak Mandelbaum [Tue, 24 May 2022 19:04:54 +0000 (19:04 +0000)]
[clang][dataflow] Make limit on fixpoint-algorithm iterations proportional to size of CFG.

Currently, the maximum number of iterations of the loop for finding the fixpoint
of the dataflow analysis is set at 2^16. When things go wrong in an analysis,
this can be far too large.  This patch changes the limit to be proportional to
the size of the CFG, which will generally be far smaller than 2^16 (while still
maintaining 2^16 as the absolute limit).

Differential Revision: https://reviews.llvm.org/D126316

2 years ago[Sanitizers][Darwin] Replace SANITIZER_MAC with SANITIZER_APPLE in source files
Mariusz Borsa [Mon, 23 May 2022 21:35:42 +0000 (14:35 -0700)]
[Sanitizers][Darwin] Replace SANITIZER_MAC with SANITIZER_APPLE in source files

This is a follow up to [Sanitizers][Darwin] Rename Apple macro SANITIZER_MAC -> SANITIZER_APPLE (D125816)

Performed a global search/replace as in title against LLVM sources

Differential Revision: https://reviews.llvm.org/D126263

2 years ago[flang] Catch calls to assumed-length character functions
Peter Klausler [Wed, 18 May 2022 22:44:01 +0000 (15:44 -0700)]
[flang] Catch calls to assumed-length character functions

Semantics was allowing calls to CHARACTER(*) functions, which are odd
things -- they can be declared, and passed around, but can never actually
be called as such.  They must be redeclared with an explicit length that
ends up being passed as a hidden argument.  So check for these calls
and diagnose them, add tests, and clean up some existing tests that
were in error and now get caught.

Possible TODO for lowering: there were some test cases that used
bad calls to assumed-length CHARACTER*(*) functions and validated
their implementations.  I've removed some, and adjusted another,
but the code that somehow implemented these calls may need to be
removed and replaced with an assert about bad semantics.

Differential Revision: https://reviews.llvm.org/D126148

2 years ago[SLP] Fix crash caused by reorderBottomToTop().
Vasileios Porpodas [Tue, 24 May 2022 16:16:30 +0000 (09:16 -0700)]
[SLP] Fix crash caused by reorderBottomToTop().

The crash is caused by incorrect order set by reorderBottomToTop(), which
happens when it is reordering a TreeEntry which has a user that has already been
reordered earlier. Please see the detailed description in the lit test.

Differential Revision: https://reviews.llvm.org/D126099

2 years ago[InstCombine] fold more shuffles with FP<->Int cast operands
Sanjay Patel [Tue, 24 May 2022 19:07:24 +0000 (15:07 -0400)]
[InstCombine] fold more shuffles with FP<->Int cast operands

shuffle (cast X), (cast Y), Mask --> cast (shuffle X, Y, Mask)

This extends the transform added with 0353c2c996c5.

If the shuffle reduces vector length, the transform
reduces the width of the cast, so that should be a
win for most codegen (if not, it can be inverted).

2 years agoRevert "[compiler-rt][scudo] Add missing preprocessor token" and "[compiler-rt][scudo...
Leonard Chan [Tue, 24 May 2022 18:29:44 +0000 (11:29 -0700)]
Revert "[compiler-rt][scudo] Add missing preprocessor token" and "[compiler-rt][scudo] Simplify TBI checks"

This reverts commit 676eaa2ca967ca6ad4a84d31d6f0ebabdcf3e44b
and f6038cdca03115da22b9e6ada5c25de4df5f42d2 since builders are still
broken.

2 years ago[compiler-rt][scudo] Add missing preprocessor token
Leonard Chan [Tue, 24 May 2022 18:11:31 +0000 (11:11 -0700)]
[compiler-rt][scudo] Add missing preprocessor token

This should fix build errors seen on bots like
https://lab.llvm.org/buildbot/#/builders/57/builds/18263.

2 years agoRecommit "[SelectOpti][5/5] Optimize select-to-branch transformation"
Sotiris Apostolakis [Tue, 24 May 2022 03:04:20 +0000 (23:04 -0400)]
Recommit "[SelectOpti][5/5] Optimize select-to-branch transformation"

Use container::size_type directly to avoid type mismatch causing build failures in Windows.

Original commit message:
This patch optimizes the transformation of selects to a branch when the heuristics deemed it profitable.
It aggressively sinks eligible instructions to the newly created true/false blocks to prevent their
execution on the common path and interleaves dependence slices to maximize ILP.

Depends on D120232

Reviewed By: davidxl

Differential Revision: https://reviews.llvm.org/D120233

2 years ago[flang] Fix crash in semantics after PDT instantiation
Peter Klausler [Thu, 19 May 2022 00:48:34 +0000 (17:48 -0700)]
[flang] Fix crash in semantics after PDT instantiation

The code in semantics that reinitializes symbol table pointers in
the parse tree of a parameterized derived type prior to a new
instantiation of the type was processing the symbols of the
derived type instantiation scope in arbitrary address order,
which could fail if a reference to a type parameter inherited from
an ancestor type was processed prior to the parent component sequence.
Fix by instantiating components of PDT instantiations in declaration
order.

Differential Revision: https://reviews.llvm.org/D126147

2 years ago[flang] Alternate entry points with unused arguments
V Donaldson [Tue, 24 May 2022 17:06:24 +0000 (10:06 -0700)]
[flang] Alternate entry points with unused arguments

A dummy argument in an entry point of a subprogram with multiple
entry points need not be defined in other entry points. It is only
legal to reference such an argument when calling an entry point that
does have a definition. An entry point without such a definition
needs a local "substitute" definition sufficient to generate code.
It is nonconformant to reference such a definition at runtime.
Most such definitions and associated code will be deleted as dead
code at compile time. However, that is not always possible, as in
the following code. This code is conformant if all calls to entry
point ss set m=3, and all calls to entry point ee set n=3.

subroutine ss(a, b, m, d, k) ! no x, y, n
  integer :: a(m), b(a(m)), m, d(k)
  integer :: x(n), y(x(n)), n
  integer :: k
1 print*, m, k
  print*, a
  print*, b
  print*, d
  if (m == 3) return
entry ee(x, y, n, d, k) ! no a, b, m
  print*, n, k
  print*, x
  print*, y
  print*, d
  if (n /= 3) goto 1
end

  integer :: xx(3), yy(5), zz(3)
  xx = 5
  yy = 7
  zz = 9
  call ss(xx, yy, 3, zz, 3)
  call ss(xx, yy, 3, zz, 3)
end

Lowering currently generates fir::UndefOp's for all unused arguments.
This is usually ok, but cases such as the one here incorrectly access
unused UndefOp arguments for m and n from an entry point that doesn't
have a proper definition.

The problem is addressed by creating a more complete definition of an
unused argument in most cases. This is implemented in large part by
moving the definition of an unused argument from mapDummiesAndResults
to mapSymbolAttributes. The code in mapSymbolAttributes then chooses
one of three code generation options, depending on information
available there.

This patch deals with dummy procedures in alternate entries, and adds
a TODO for procedure pointers (the PFTBuilder is modified to analyze
procedure pointer symbol so that they are not silently ignored, and
instead hits proper TODOs).

BoxAnalyzer is also changed because assumed-sized arrays were wrongfully
categorized as constant shape arrays.  This had no impact, except when
there were unused entry points.

Co-authored-by: jeanPerier <jperier@nvidia.com>
Differential Revision: https://reviews.llvm.org/D125867

2 years ago[AMDGPU] Disable newly added gfx90a global isel image tests. NFC.
Stanislav Mekhanoshin [Tue, 24 May 2022 17:47:18 +0000 (10:47 -0700)]
[AMDGPU] Disable newly added gfx90a global isel image tests. NFC.

This fixed build failure with expensive checks after D126009.
The change has added new run lines for Global ISel which has
uncovered a pre-existing problem: it does not select a correct
flavor of these image instructions.

2 years ago[compiler-rt][scudo] Simplify TBI checks
Leonard Chan [Tue, 24 May 2022 17:51:42 +0000 (10:51 -0700)]
[compiler-rt][scudo] Simplify TBI checks

Differential Revision: https://reviews.llvm.org/D111080

2 years agoSema: adjust assertion to account for deduced types
Saleem Abdulrasool [Fri, 20 May 2022 20:35:28 +0000 (20:35 +0000)]
Sema: adjust assertion to account for deduced types

Previous changes for the BTF attributes introduced a new sub-tree
visitation.  That uncovered that when accessing the typespec location we
would assert that the type specification is either a type declaration or
`typename`.  However, `typename` was explicitly permitted.  This change
predates the introduction of newer deduced type representations such as
`__underlying_type` from C++ and the addition of the GNU `__typeof__`
expression.

Thanks to aaron.ballman for the valuable discussion and pointer to
`isTypeRep`.

Differential Revision: https://reviews.llvm.org/D126093
Reviewed By: aaron.ballman, yonghong-song

2 years ago[OpenMP] Fix file arguments for embedding bitcode in the linker wrapper
Joseph Huber [Tue, 24 May 2022 17:43:52 +0000 (13:43 -0400)]
[OpenMP] Fix file arguments for embedding bitcode in the linker wrapper

Summary:
The linker wrapper supports embedding bitcode images instead of linked
device images to facilitate JIT in the device runtime. However, we were
incorrectly passing in the file twice when this option was set. This
patch makes sure we only use the intermediate result of the LTO pass and
don't add the final output to the full job.

In the future we will want to add both of these andle handle that
accoridngly to allow the runtime to either use the AoT compiled version
or JIT compile the bitcode version if availible.

2 years ago[OpenMP] Add parsing/sema support for omp_all_memory reserved locator
Mike Rice [Tue, 17 May 2022 17:11:00 +0000 (10:11 -0700)]
[OpenMP] Add parsing/sema support for omp_all_memory reserved locator

Adds support for the reserved locator 'omp_all_memory' for use
in depend clauses with 'out' or 'inout' dependence-types.

Differential Revision: https://reviews.llvm.org/D125828

2 years ago[flang][runtime] Handle BACKSPACE after reading past EOF
Peter Klausler [Wed, 18 May 2022 20:40:33 +0000 (13:40 -0700)]
[flang][runtime] Handle BACKSPACE after reading past EOF

An external READ(END=) that hits the end of the file must
also note the virtual position of the endfile record that
has just been discovered, so that a later BACKSPACE statement
won't end up at the wrong record.

Differential Revision: https://reviews.llvm.org/D126146

2 years ago[compiler-rt][lsan] Update CanBeAHeapPointer for AArch64
Leonard Chan [Tue, 24 May 2022 17:22:46 +0000 (10:22 -0700)]
[compiler-rt][lsan] Update CanBeAHeapPointer for AArch64

While attempting to get the 64-bit lsan allocator working for Fuchsia, I
noticed this function would incorrectly return false for pointers returned
by the 64-bit allocator. On AArch64, this function attempts to get the VMA
size dynamically by counting the number of leading zeros from the function
frame address. This will fail if the frame address is significantly below an
allocated pointer (that is, the frame address has more leading zeros than an
allocated pointer). This is possible on Fuchsia and linux (when not called
from the initial thread stack).

It seems the intended use of this function is to speed up pointer scanning by
filtering out addresses that user code might not be able to access. Other
platforms this check is done on seem to hardcode the VMA size/shift, so it
seems appropriate to do this for aarch64 as well. This implies pointers on
aarch64 where the VMA size is <64 will pass through, but bad pointers will
still be caught by subsequent scan checks.

This patch also renames the function to something more fitting of what it's
trying to do.

Differential Revision: https://reviews.llvm.org/D123814

2 years ago[clangd] Handle '--' in QueryDriverDatabase
Nathan Ridge [Tue, 24 May 2022 06:12:53 +0000 (02:12 -0400)]
[clangd] Handle '--' in QueryDriverDatabase

Fixes https://github.com/clangd/clangd/issues/1100,
a regression from D116721.

Differential Revision: https://reviews.llvm.org/D126274

2 years agoReland "[PS5] Verify defaults to -fno-stack-size-section"
Paul Robinson [Tue, 24 May 2022 17:03:20 +0000 (10:03 -0700)]
Reland "[PS5] Verify defaults to -fno-stack-size-section"

This reverts commit efebb27b745a0d677ad2ea9aefff242c12aef29c.
Fixes typos (accidentally omitted %s from some RUN lines).

2 years ago[AMDGPU] Enforce alignment of image vaddr on gfx90a
Stanislav Mekhanoshin [Wed, 18 May 2022 19:22:38 +0000 (12:22 -0700)]
[AMDGPU] Enforce alignment of image vaddr on gfx90a

Even though single address image instructions only use a single VGPR
HW accesses 4 or 5 which creates alignment requirement.

Fixes: SWDEV-316648

Differential Revision: https://reviews.llvm.org/D126009

2 years ago[libclang] add supporting for indexing/visiting C++ concepts
Alex Lorenz [Thu, 19 May 2022 23:09:36 +0000 (16:09 -0700)]
[libclang] add supporting for indexing/visiting C++ concepts

This commit builds upon recently added indexing support for C++ concepts
from https://reviews.llvm.org/D124441 by extending libclang to
support indexing and visiting concepts, constraints and requires
expressions as well.

Differential Revision: https://reviews.llvm.org/D126031

2 years agoRevert "[PS5] Verify defaults to -fno-stack-size-section"
Paul Robinson [Tue, 24 May 2022 16:59:57 +0000 (09:59 -0700)]
Revert "[PS5] Verify defaults to -fno-stack-size-section"

This reverts commit 28432b0f655641df7f9d079cf69ba235038d6340.

Caused some unexpected buildbot failures.

2 years agoNFC. Clang-formatting.
Chris Bieneman [Tue, 24 May 2022 16:51:00 +0000 (11:51 -0500)]
NFC. Clang-formatting.

Since the rest of the DirectX backend is pretty well clang-format
clean, this file should be too.

2 years ago[flang][runtime] INQUIRE(UNIT=666,NUMBER=n) must set n=666
Peter Klausler [Wed, 18 May 2022 20:23:39 +0000 (13:23 -0700)]
[flang][runtime] INQUIRE(UNIT=666,NUMBER=n) must set n=666

Whether a unit number in an inquire-by-unit statement is valid or not,
it should be the value to which the NUMBER= variable is set, not -1.
-1 should be returned to NUMBER= only for an inquire-by-file statement
when the FILE= is not connected to any unit.

Differential Revision: https://reviews.llvm.org/D126145

2 years ago[PS5] Verify defaults to -fno-stack-size-section
Paul Robinson [Tue, 24 May 2022 16:47:22 +0000 (09:47 -0700)]
[PS5] Verify defaults to -fno-stack-size-section

2 years agoRecommit "[RISCV] Use selectShiftMaskXLen ComplexPattern for isel of rotates."
Craig Topper [Tue, 24 May 2022 16:41:04 +0000 (09:41 -0700)]
Recommit "[RISCV] Use selectShiftMaskXLen ComplexPattern for isel of rotates."

This reverts commit dfe513ae1bb6e788ead93b850d80d77d54cf29d3.

Tests have been changed to avoid the type legalization bug being
fixed in D126036.

Original commit message:
This will remove masks on the shift amount. We usually get this with
SimplifyDemandedBits in DAGCombine, but that's restricted to cases
where the AND has a single use. selectShiftMaskXLen does not have
that restriction.

2 years ago[RISCV] Add test cases showing failure to remove mask on rotate amounts.
Craig Topper [Tue, 24 May 2022 16:41:00 +0000 (09:41 -0700)]
[RISCV] Add test cases showing failure to remove mask on rotate amounts.

This is similar to tests I added in
e2f410feeab27a8bb2c015fc02bb8527702e401f that had to be reverted.

I've modified them to avoid the bug that is being fixed by D126036.

2 years ago[gn build] Reformat all build files
Nico Weber [Tue, 24 May 2022 15:42:34 +0000 (11:42 -0400)]
[gn build] Reformat all build files

Ran:

    git ls-files '*.gn' '*.gni' | xargs llvm/utils/gn/gn.py format

2 years ago[flang] Extension: Accept Hollerith actual arguments as if they were BOZ
Peter Klausler [Thu, 19 May 2022 16:55:29 +0000 (09:55 -0700)]
[flang] Extension: Accept Hollerith actual arguments as if they were BOZ

When a Hollerith (or short character) literal is presented as an actual
argument that corresponds to a dummy argument for which a BOZ literal
would be acceptable, treat the Hollerith as if it had been a BOZ
literal in the same way -- and with the same code -- as f18 already
does for the similar extension in DATA statements.

Differential Revision: https://reviews.llvm.org/D126144

2 years ago[Clang] Avoid misleading 'conflicting types' diagnostic with no-prototype decls.
Cyndy Ishida [Tue, 24 May 2022 13:24:02 +0000 (06:24 -0700)]
[Clang] Avoid misleading 'conflicting types' diagnostic with no-prototype decls.

Clang has recently started diagnosing prototype redeclaration errors like [rG385e7df33046](https://reviews.llvm.org/rG385e7df33046d7292612ee1e3ac00a59d8bc0441)

This flagged legitimate issues in a codebase but was confusing to resolve because it actually conflicted with a function declaration from a system header and not from the one emitted with "note: ".

This patch updates the error handling to use the canonical declaration's source location instead to avoid misleading errors like the one described.

Reviewed By: aaron.ballman

Differential Revision: https://reviews.llvm.org/D126258

2 years ago[gn build] (semi-automatically) port 0360b9f1599b
Nico Weber [Tue, 24 May 2022 15:40:40 +0000 (11:40 -0400)]
[gn build] (semi-automatically) port 0360b9f1599b

2 years ago[libc++][NFC] Whitespace refactoring of string.cpp for consistency and legibility
Louis Dionne [Tue, 24 May 2022 15:31:31 +0000 (11:31 -0400)]
[libc++][NFC] Whitespace refactoring of string.cpp for consistency and legibility

2 years ago[libc++][NFC] Move definitions around in string.cpp to reduce _LIBCPP_HAS_NO_WIDE_CHA...
Louis Dionne [Tue, 24 May 2022 15:22:43 +0000 (11:22 -0400)]
[libc++][NFC] Move definitions around in string.cpp to reduce _LIBCPP_HAS_NO_WIDE_CHARACTERS blocks

2 years ago[mlir] Rename mlir::SmallVector -> llvm::SmallVector
Mogball [Tue, 24 May 2022 15:03:12 +0000 (15:03 +0000)]
[mlir] Rename mlir::SmallVector -> llvm::SmallVector

2 years ago[mlir] Breakdown diagnostic string literals
Logan Chien [Wed, 18 May 2022 20:59:08 +0000 (13:59 -0700)]
[mlir] Breakdown diagnostic string literals

This commit breaks down diagnostic string literals so that the attribute
name and enumurator names can be shared with the stringify utility
function and the "expected ", " to be one of ", and ", " can be shared
between different enum-related diagnostic.

Differential Revision: https://reviews.llvm.org/D125938

2 years ago[flang][runtime] Clean up asynchronous I/O APIs
Peter Klausler [Thu, 12 May 2022 00:08:21 +0000 (17:08 -0700)]
[flang][runtime] Clean up asynchronous I/O APIs

Now that the requirements and implementation of asynchronous I/O are
better understood, adjust their I/O runtime APIs.  In particular:
1) Remove the BeginAsynchronousOutput/Input APIs; they're not needed,
   since any data transfer statement might have ASYNCHRONOUS= and
   (if ASYNCHRONOUS='YES') ID= control list specifiers that need to
   at least be checked.
2) Add implementations for BeginWait(All) to check for the error
   case of a bad unit number and nonzero ID=.
3) Rearrange and comment SetAsynchronous so that it's clear that
   it can be called for READ/WRITE as well as for OPEN.

The implementation remains completely synchronous, but should be conforming.
Where opportunities make sense for true asynchronous implementations of
some big block transfers without SIZE= in the future, we'll need to add
a GetAsynchronousId API to capture ID= on a READ or WRITE; add sourceFile
and sourceLine arguments to BeginWait(All) for good error reporting;
track pending operations in unit.h; and add code to force synchronization
to non-asynchronous I/O operations.

Lowering should call SetAsynchronous when ASYNCHRONOUS= appears as
a control list specifier.  It should also set ID=x variables to 0
until such time as we support asynchronous operations, if ever.
This patch only removes the removed APIs from lowering.

Differential Revision: https://reviews.llvm.org/D126143

2 years agoFix behavior of is_fp_class on empty class set
Serge Pavlov [Tue, 24 May 2022 11:30:13 +0000 (18:30 +0700)]
Fix behavior of is_fp_class on empty class set

The second argument to is_fp_class specifies the set of floating-point
class to test against. It can be zero, in this case the intrinsic is
expected to return zero value.

Differential Revision: https://reviews.llvm.org/D112025

2 years ago[DAG] Unroll vectorized FPOW instructions before widening that will scalarize to...
Simon Pilgrim [Tue, 24 May 2022 14:44:44 +0000 (15:44 +0100)]
[DAG] Unroll vectorized FPOW instructions before widening that will scalarize to libcalls anyway

Followup to D125988 - FPOW is similar to FREM and will most likely scalarize to libcalls, so unroll before widening to prevent use making additional libcalls with UNDEF args.

2 years ago[libcxx] Add sort.bench.cpp to libcxx/benchmarks/CMakeLists.txt
Hans Wennborg [Tue, 24 May 2022 13:47:41 +0000 (15:47 +0200)]
[libcxx] Add sort.bench.cpp to libcxx/benchmarks/CMakeLists.txt

It was forgotten in D124740.

Differential revision: https://reviews.llvm.org/D126297

2 years ago[TypePromotion] Avoid unnecessary trunc zext pairs
Sam Parker [Tue, 24 May 2022 14:33:21 +0000 (15:33 +0100)]
[TypePromotion] Avoid unnecessary trunc zext pairs

Any zext 'sink' should already have an operand that is in the legal
value, so avoid using a trunc and just use the trunc operand instead.

Differential Revision: https://reviews.llvm.org/D118905

2 years ago[mlir][vector] Add new lowering mode to vector.contractionOp
Thomas Raoux [Tue, 24 May 2022 14:16:00 +0000 (14:16 +0000)]
[mlir][vector] Add new lowering mode to vector.contractionOp

Add lowering for cases where the reduction dimension is fully unrolled.
It is common to unroll the reduction dimension, therefore we would want
to lower the contractions to an elementwise vector op in this case.

Differential Revision: https://reviews.llvm.org/D126120

2 years ago[CostModel][X86] getScalarizationOverhead - improve extraction costs for > 128-bit...
Simon Pilgrim [Tue, 24 May 2022 14:17:59 +0000 (15:17 +0100)]
[CostModel][X86] getScalarizationOverhead - improve extraction costs for > 128-bit vectors

We were using the default getScalarizationOverhead expansion for extraction costs, which adds up all the individual element extraction costs.

This is fine for 128-bit vectors, but for 256/512-bit vectors each element extraction also has to account for extracting the upper 128-bit subvector extraction before it can handle the element. For scalarization costs we only need to extract each demanded subvector once.

Differential Revision: https://reviews.llvm.org/D125527

2 years ago[AMDGPU][MC][GFX11] Support base+soffset+offset SMEM loads.
Ivan Kosarev [Tue, 24 May 2022 14:12:45 +0000 (15:12 +0100)]
[AMDGPU][MC][GFX11] Support base+soffset+offset SMEM loads.

Reviewed By: dp

Differential Revision: https://reviews.llvm.org/D126207

2 years ago[InstCombine] Strip bitcasts in GEP diff fold
Nikita Popov [Tue, 24 May 2022 14:08:21 +0000 (16:08 +0200)]
[InstCombine] Strip bitcasts in GEP diff fold

Bitcasts were stripped in one case, but not the other. Of course,
this no longer really matters with opaque pointers, but as I went
through the trouble of tracking this down, we may as well remove
one typed vs opaque pointer optimization discrepancy.