platform/upstream/llvm.git
2 years ago[OpenMP][libomp] Fix accidental removal of else for core attributes
Jonathan Peyton [Thu, 19 May 2022 18:57:02 +0000 (13:57 -0500)]
[OpenMP][libomp] Fix accidental removal of else for core attributes

2 years ago[ARM] Cost modelling for scalar fptoi_sat
David Green [Thu, 19 May 2022 18:53:21 +0000 (19:53 +0100)]
[ARM] Cost modelling for scalar fptoi_sat

Similar to D124357, this adds some cost modelling for fptoi_sat for Arm
targets. Where VFP2 is available (and FP64/FP16 for the relevant types),
the operations are legal as the Arm instructions naturally saturate.
Otherwise they will need an extra smin/smax clamp, similar to AArch64.

Differential Revision: https://reviews.llvm.org/D125665

2 years ago[InstCombine] NEW Baseline tests for InstCombine optimization to merge GEP instructio...
William Huang [Thu, 12 May 2022 00:32:03 +0000 (00:32 +0000)]
[InstCombine] NEW Baseline tests for InstCombine optimization to merge GEP instructions with constant indices

Splitted the merge constant-indexed GEP optimization into two smaller transformations: 1. Merging GEP of GEP if both are constant-indexed. 2. Swapping constant indexed GEP in a chain of (non-constant) GEP to the end, so that 1 can be applied repeatedly.
There is existing code to partially  handle transformation 1, but it only deals with limited cases

Unit tests are breaking down into two parts for the 2 transformations.

Reviewed By: nikic

Differential Revision: https://reviews.llvm.org/D125438

2 years ago[Office Hours] add initial guidance for hosts
Kristof Beyls [Thu, 19 May 2022 10:25:01 +0000 (12:25 +0200)]
[Office Hours] add initial guidance for hosts

This includes adding guidance to announce an office hours session on the
Discord channel and/or IRC, as discussed at the office hours round table at
EuroLLVM 2022, see
https://discourse.llvm.org/t/office-hours-eurollvm-round-table-summary/62480.

Fixes #55423

2 years ago[ValueTracking] Added support to deduce PHI Nodes values being a power of 2
William Huang [Tue, 10 May 2022 19:47:10 +0000 (19:47 +0000)]
[ValueTracking] Added support to deduce PHI Nodes values being a power of 2

 Add Value Tracking support to deduce induction variable being a power of 2, allowing urem optimizations

Reviewed By: spatel

Differential Revision: https://reviews.llvm.org/D125332

2 years ago[llvm-dis][test] Fix error case on Windows
Keith Smiley [Thu, 19 May 2022 18:29:37 +0000 (11:29 -0700)]
[llvm-dis][test] Fix error case on Windows

The `N` case in the error differs across platforms.

2 years ago[ValueTracking] Baseline tests for Power-of-2 value tracking on PHI nodes
William Huang [Tue, 3 May 2022 22:19:01 +0000 (22:19 +0000)]
[ValueTracking] Baseline tests for Power-of-2 value tracking on PHI nodes

Reviewed By: davidxl

Differential Revision: https://reviews.llvm.org/D124885

2 years ago[Libomptarget] Add `leaf` attribute to `vprintf` declaration
Joseph Huber [Thu, 19 May 2022 18:18:51 +0000 (14:18 -0400)]
[Libomptarget] Add `leaf` attribute to `vprintf` declaration

Summary:
This patch adds the `leaf` attribute to the `vprintf` declaration in the
OpenMP runtime. This attribute allows us to determine that the `vprintf`
function will not call any functions within the translation unit,
allowing us to deduce `norecurse` attributes on the caller.

2 years ago[WebAssembly] Fix register use-def in FixIrreducibleControlFlow
Heejin Ahn [Fri, 13 May 2022 01:46:53 +0000 (18:46 -0700)]
[WebAssembly] Fix register use-def in FixIrreducibleControlFlow

FixIrreducibleControlFlow pass adds dispatch blocks with a `br_table`
that has multiple predecessors and successors, because it serves as
something like a traffic hub for BBs. As a result of this, there can be
register uses that are not dominated by a def in every path from the
entry block. For example, suppose register %a is defined in BB1 and used
in BB2, and there is a single path from BB1 and BB2:
```
BB1 -> ... -> BB2
```
After FixIrreducibleControlFlow runs, there can be a dispatch block
between these two BBs:
```
BB1 -> ... -> Dispatch -> ... -> BB2
```
And this dispatch block has multiple predecessors, now
there is a path to BB2 that does not first visit BB1, and in that path
%a is not dominated by a def anymore.

To fix this problem, we have been adding `IMPLICIT_DEF`s to all
registers in PrepareForLiveInternals pass, and then remove unnecessary
ones in OptimizeLiveIntervals pass after computing `LiveIntervals`. But
FixIrreducibleControlFlow pass itself ends up violating register use-def
relationship, resulting in invalid code. This was OK so far because
MIR verifier apparently didn't check this in validation. But @arsenm
fixed this and it caught this bug in validation
(https://github.com/llvm/llvm-project/issues/55249).

This CL moves the `IMPLICIT_DEF` adding routine from
PrepareForLiveInternals to FixIrreducibleControlFlow. We only run it
when FixIrreducibleControlFlow changes the code. And then
PrepareForLiveInternals doesn't do anything other than setting
`TracksLiveness` property, which is a prerequisite for running
`LiveIntervals` analysis, which is required by the next pass
OptimizeLiveIntervals.

But in our backend we don't seem to do anything that invalidates this up
until OptimizeLiveIntervals, and I'm not sure why we are calling
`invalidateLiveness` in ReplacePhysRegs pass, because what that pass
does is to replace physical registers with virtual ones 1-to-1. I
deleted the `invalidateLiveness` call there and we don't need to set
that flag explicitly, which obviates all the need for
PrepareForLiveInternals.

(By the way, This 'Liveness' here is different from `LiveIntervals`
analysis. Setting this only means BBs' live-in info is correct, all uses
are dominated by defs, `kill` flag is conservatively correct, which
means if there is a `kill` flag set it should be the last use. See
https://github.com/llvm/llvm-project/blob/2a0837aab1489c88efb03784e34c4dc9f2e28302/llvm/include/llvm/CodeGen/MachineFunction.h#L125-L134
for details.)

So this CL removes PrepareForLiveInternals pass altogether. Something
similar to this was attempted by D56091 long ago but that came short of
actually removing the pass, and I couldn't land it because
FixIrreducibleControlFlow violated use-def relationship, which this CL
fixes.

This doesn't change output in any meaningful way. All test changes
except `irreducible-cfg.mir` are register numbering.

Also this will likely to reduce compilation time, because we have been
adding `IMPLICIT_DEF` for all registers every time `-O2` is given, but
now we do that only when there is irreducible control flow, which is
rare.

Fixes https://github.com/llvm/llvm-project/issues/55249.

Reviewed By: dschuff, kripken

Differential Revision: https://reviews.llvm.org/D125515

2 years ago[WebAssembly] Use CHECK-NEXT for irreducible-cfg.mir
Heejin Ahn [Fri, 13 May 2022 02:30:55 +0000 (19:30 -0700)]
[WebAssembly] Use CHECK-NEXT for irreducible-cfg.mir

Reviewed By: dschuff

Differential Revision: https://reviews.llvm.org/D125514

2 years ago[llvm-dis] Improve missing file error message
Keith Smiley [Tue, 15 Mar 2022 04:50:46 +0000 (21:50 -0700)]
[llvm-dis] Improve missing file error message

Previously the error message didn't include the failing path, which made
it hard to tell what went wrong.

Differential Revision: https://reviews.llvm.org/D121665

2 years ago[docs][tools] Remove old llvm-bcanalyzer options
Keith Smiley [Wed, 23 Mar 2022 23:29:07 +0000 (16:29 -0700)]
[docs][tools] Remove old llvm-bcanalyzer options

These no longer exist. A few have been added since but I'm not enough of
an expert to provide a useful blurb on them outside of what you see with
`--help`.

Differential Revision: https://reviews.llvm.org/D122361

2 years ago[Object] Fix updating darwin archives
Keith Smiley [Wed, 4 May 2022 01:43:46 +0000 (18:43 -0700)]
[Object] Fix updating darwin archives

When creating an archive, llvm-ar looks at the host to determine the
archive format to use, on Apple platforms this means it uses the
K_DARWIN format. K_DARWIN is _virtually_ equivalent to K_BSD, expect for
some very slight differences around padding, timestamps in deterministic
mode, and 64 bit formats. When updating an archive using llvm-ar, or
llvm-objcopy, Archive would try to determine the kind, but it was not
possible to get K_DARWIN in the initialization of the archive, because
they're virtually inciting usable from K_BSD, especially since the
slight differences only apply in very specific cases. This leads to
linker failures when the alignment workaround is not applied to an
archive copied with llvm-objcopy. This change teaches Archive to infer
the K_DARWIN type in the cases where it's possible and the first object
in the archive is a macho object. This avoids using the host triple to
determine this to not affect cross compiling.

Ideally we would eliminate the separate K_DARWIN type entirely since
it's not a truly separate archive type, but then we'd have to force the
macho workarounds on the BSD format generally. This might be acceptable
but then it would be unclear how to handle this case without forcing the
K_DARWIN64 format on all BSD users:

```
if (LastOffset >= Sym64Threshold) {
  if (Kind == object::Archive::K_DARWIN)
    Kind = object::Archive::K_DARWIN64;
  else
    Kind = object::Archive::K_GNU64;
}
```

The logic used to determine if the object is macho is derived from the
logic llvm-ar uses.

Previous context:

111cd669e90e5b2132187d36f8b141b11a671a8b
23a76be5adcaa768ba538f8a4514a7afccf61988

Differential Revision: https://reviews.llvm.org/D124895

2 years ago[ORC] Avoid more SymbolStringPtr copies.
Lang Hames [Thu, 19 May 2022 02:19:57 +0000 (19:19 -0700)]
[ORC] Avoid more SymbolStringPtr copies.

2 years ago[ORC] Add a FIXME.
Lang Hames [Thu, 19 May 2022 01:58:14 +0000 (18:58 -0700)]
[ORC] Add a FIXME.

2 years ago[ORC] Add missing std::moves, pass SymbolLookupSet by value.
Lang Hames [Thu, 19 May 2022 01:39:33 +0000 (18:39 -0700)]
[ORC] Add missing std::moves, pass SymbolLookupSet by value.

Avoids some unnecessary SymbolStringPtr copies.

2 years ago[llvm-jitlink] Print session report even if entry-point lookup errors out.
Lang Hames [Sun, 17 Apr 2022 23:38:01 +0000 (16:38 -0700)]
[llvm-jitlink] Print session report even if entry-point lookup errors out.

2 years ago[mlir][vector] Fix crash in DropInnerMostUnitDims pattern
Thomas Raoux [Thu, 19 May 2022 17:38:04 +0000 (17:38 +0000)]
[mlir][vector] Fix crash in DropInnerMostUnitDims pattern

Fix number of dimensions when incrementally replacing dimensions in
affine map.

Differential Revision: https://reviews.llvm.org/D125984

2 years ago[mlir][tensor] Add canonicalization for tensor.cast from extract_slice
Thomas Raoux [Thu, 19 May 2022 13:55:18 +0000 (13:55 +0000)]
[mlir][tensor] Add canonicalization for tensor.cast from extract_slice

Propagate static size information into extract_slice producer if
possible.

Differential Revision: https://reviews.llvm.org/D125972

2 years ago[NFC] Fix a couple of whitespace issues.
Paul Walker [Thu, 19 May 2022 11:09:36 +0000 (11:09 +0000)]
[NFC] Fix a couple of whitespace issues.

2 years agoDrop qualifiers from return types in C (DR423)
Aaron Ballman [Thu, 19 May 2022 17:05:34 +0000 (13:05 -0400)]
Drop qualifiers from return types in C (DR423)

WG14 DR423 (https://www.open-std.org/jtc1/sc22/wg14/www/docs/n2148.htm#dr_423),
resolved during the C11 time frame, changed the way qualifiers are
handled on function return types and in cast expressions after it was
noted that these types are now directly observable via generic
selection expressions. In C, the function declarator is adjusted to
ignore all qualifiers (including _Atomic qualifiers).

Clang already handles the cast expression case correctly (by performing
the lvalue conversion, which drops the qualifiers as well), but with
these changes it will now also handle function declarations
appropriately.

Fixes #39595

Differential Revision: https://reviews.llvm.org/D125919

2 years ago[DeadArgElim] Use poison instead of undef as placeholder for dead arguments
Nuno Lopes [Thu, 19 May 2022 17:00:24 +0000 (18:00 +0100)]
[DeadArgElim] Use poison instead of undef as placeholder for dead arguments

It doesn't matter which value we use for dead args, so let's switch
to poison, so we can eventually kill undef.

Reviewed By: aeubanks, fhahn

Differential Revision: https://reviews.llvm.org/D125983

2 years ago[gn build] Port ca7c307d1816
LLVM GN Syncbot [Thu, 19 May 2022 16:42:14 +0000 (16:42 +0000)]
[gn build] Port ca7c307d1816

2 years ago[SelectOpti][1/5] Setup new select-optimize pass
Sotiris Apostolakis [Fri, 13 May 2022 22:29:21 +0000 (22:29 +0000)]
[SelectOpti][1/5] Setup new select-optimize pass

This is the first commit for the cmov-vs-branch optimization pass.
The goal is to develop a new profile-guided and target-independent cost/benefit analysis
for selecting conditional moves over branches when optimizing for performance.

Initially, this new pass is expected to be enabled only for instrumentation-based PGO.

RFC: https://discourse.llvm.org/t/rfc-cmov-vs-branch-optimization/6040

Reviewed By: tejohnson

Differential Revision: https://reviews.llvm.org/D120230

2 years ago[NVVM] Update intrinsic defintions to include the `nocallback` attribute
Joseph Huber [Wed, 18 May 2022 23:33:17 +0000 (19:33 -0400)]
[NVVM] Update intrinsic defintions to include the `nocallback` attribute

This patch adds the `nocallback` attribute to the NVVM intrinsics that
did not use the `DefaultAttrsIntrinsic` method that includes it already.
The `nocallback` attribute states that the intrinsic function cannot
enter back into the caller's translation-unit. This allows as to
determine that a function calling a `nocallback` function can have the
`norecurse` attribute.  This should be safe for all the NVVM intrinsics
because they do not call other functions within the translation unit.

Reviewed By: jdoerfert

Differential Revision: https://reviews.llvm.org/D125937

2 years ago[PowerPC] Implement XL compat __fnabs and __fnabss builtins.
Amy Kwan [Thu, 19 May 2022 14:38:34 +0000 (09:38 -0500)]
[PowerPC] Implement XL compat __fnabs and __fnabss builtins.

This patch implements the following floating point negative absolute value
builtins that required for compatibility with the XL compiler:
```
double __fnabs(double);
float __fnabss(float);
```

These builtins will emit :
- fnabs on PWR6 and below, or if VSX is disabled.
- xsnabsdp on PWR7 and above, if VSX is enabled.

Differential Revision: https://reviews.llvm.org/D125506

2 years ago[AMDGPU] emit macro __GFX9__ etc
Yaxun (Sam) Liu [Wed, 18 May 2022 16:44:59 +0000 (12:44 -0400)]
[AMDGPU] emit macro __GFX9__ etc

Emit predefined macros for GPU family. e.g.
for GPU gfx9xx emit __GFX9__, etc.

Reviewed by: Artem Belevich

Differential Revision: https://reviews.llvm.org/D125909

2 years ago[SimpleLoopUnswitch] Skip trivial selects during trivial unswitching.
Florian Hahn [Thu, 19 May 2022 16:01:11 +0000 (17:01 +0100)]
[SimpleLoopUnswitch] Skip trivial selects during trivial unswitching.

Update the remaining places in unswitchTrivialBranch to properly skip
trivial selects.

Fixes #55526.

2 years ago[AMDGPU] Allow multiple uses of the same literal in SOP2/SOPC
Jay Foad [Thu, 19 May 2022 14:23:10 +0000 (15:23 +0100)]
[AMDGPU] Allow multiple uses of the same literal in SOP2/SOPC

AMDGPUAsmParser::validateSOPLiteral already knew about this but
SIInstrInfo::verifyInstruction did not.

Differential Revision: https://reviews.llvm.org/D125976

2 years ago[lldb] Add non-address bit improvements to release notes
David Spickett [Thu, 19 May 2022 15:35:46 +0000 (15:35 +0000)]
[lldb] Add non-address bit improvements to release notes

This summarises the changes made by d9398a91e2a6b8837a47a5fda2164c9160e86199.
Which forms the bulk of the fixes needed for non-address bit handling.

Note that in the previous releases we noted memory tagging support,
which is a subset of non-address bits. The recent changes enable
debugging of programs using memory tagging, pointer authentication
and top byte ignore (all at once) on AArch64.

2 years ago[clang] Fix __has_builtin
Yaxun (Sam) Liu [Tue, 17 May 2022 19:12:03 +0000 (15:12 -0400)]
[clang] Fix __has_builtin

Fix __has_builtin to return 1 only if the requested target features
of a builtin are enabled by refactoring the code for checking
required target features of a builtin and use it in evaluation
of __has_builtin.

Reviewed by: Artem Belevich

Differential Revision: https://reviews.llvm.org/D125829

2 years ago[LoopVectorize] Don't interleave when the number of runtime checks exceeds the threshold
Tiehu Zhang [Thu, 19 May 2022 15:24:14 +0000 (23:24 +0800)]
[LoopVectorize] Don't interleave when the number of runtime checks exceeds the threshold

The runtime check threshold should also restrict interleave count.
Otherwise, too many runtime checks will be generated for some cases.

Reviewed By: fhahn, dmgreen

Differential Revision: https://reviews.llvm.org/D122126

2 years ago[LoopVectorize] Precommit a test for D122126
Tiehu Zhang [Thu, 19 May 2022 15:14:59 +0000 (23:14 +0800)]
[LoopVectorize] Precommit a test for D122126

2 years ago[VPlan] Update VPWidenMemoryInstruction to not inherit from VPValue.
Florian Hahn [Thu, 19 May 2022 15:24:38 +0000 (16:24 +0100)]
[VPlan] Update VPWidenMemoryInstruction to not inherit from VPValue.

VPWidenMemoryInstruction also models stores which may not produce a value.
This can trip over analyses. Improve the modeling by only adding
VPValues for VPWidenMemoryInstructionRecipes modeling loads.

2 years ago[libc++] Override the value of LIBCXX_CXX_ABI in the cache
Louis Dionne [Thu, 19 May 2022 15:20:26 +0000 (11:20 -0400)]
[libc++] Override the value of LIBCXX_CXX_ABI in the cache

This will allow us to remove this entirely once the commit has propagated
through all CI and hence changed the value in the cache.

2 years ago[NFC] Fix typos in X86CmovConversion
Sotiris Apostolakis [Thu, 19 May 2022 05:37:22 +0000 (05:37 +0000)]
[NFC] Fix typos in X86CmovConversion

2 years ago[libunwind] Remove unused _LIBUNWIND_HAS_NO_THREADS macro in tests
Louis Dionne [Thu, 19 May 2022 14:57:13 +0000 (10:57 -0400)]
[libunwind] Remove unused _LIBUNWIND_HAS_NO_THREADS macro in tests

The _LIBUNWIND_HAS_NO_THREADS macro is only picked up by libunwind
inside its sources, so it is only required when it builds. It doesn't
need to be defined when running the tests.

2 years ago[AMDGPU] gfx11 scalar memory instructions
Joe Nash [Mon, 25 Apr 2022 17:33:24 +0000 (13:33 -0400)]
[AMDGPU] gfx11 scalar memory instructions

Contributors:
Mirko Brkusanin <Mirko.Brkusanin@amd.com>

Patch 9/N for upstreaming of AMDGPU gfx11 architecture.

Depends on D125820

Reviewed By: kosarev, #amdgpu, arsenm

Differential Revision: https://reviews.llvm.org/D125822

2 years ago[runtimes] Fix the build of merged ABI/unwinder libraries
Louis Dionne [Wed, 18 May 2022 16:05:45 +0000 (12:05 -0400)]
[runtimes] Fix the build of merged ABI/unwinder libraries

Also, add a CI job that tests this configuration. The exact configuration
is that we build a shared libc++ and merge objects for the ABI library
and the unwinder library into it.

Differential Revision: https://reviews.llvm.org/D125903

2 years ago[flang][driver] Add support for generating executables on MacOSX/Darwin
Andrzej Warzynski [Sun, 15 May 2022 11:35:37 +0000 (12:35 +0100)]
[flang][driver] Add support for generating executables on MacOSX/Darwin

This patch basically extends https://reviews.llvm.org/D122008 with
support for MacOSX/Darwin.

To facilitate this, I've added `MacOSX` to the list of supported OSes in
Target.cpp. Flang already supports `Darwin` and it doesn't really do
anything OS-specific there (it could probably safely skip checking the
OS for now).

Note that generating executables remains hidden behind the
`-flang-experimental-exec` flag. Also, we don't need to add `-lm` on
MacOSX as `libm` is effectively included in `libSystem` (which is linked
in unconditionally).

Differential Revision: https://reviews.llvm.org/D125628

2 years ago[flang][OpenMP] Support for Collapse
Mats Petersson [Wed, 7 Jul 2021 15:58:32 +0000 (16:58 +0100)]
[flang][OpenMP] Support for Collapse

Convert Fortran parse-tree into MLIR for collapse-clause.

Includes simple Fortran to LLVM-IR test, with auto-generated
check-lines (some of which have been edited by hand).

Reviewed By: kiranchandramohan, shraiysh, peixin

Differential Revision: https://reviews.llvm.org/D125302

2 years ago[AMDGPU] gfx11 LDSDIR instructions MC support
Joe Nash [Fri, 22 Apr 2022 19:18:40 +0000 (15:18 -0400)]
[AMDGPU] gfx11 LDSDIR instructions MC support

Contributors:
Carl Ritson <carl.ritson@amd.com>

Patch 8/N for upstreaming of AMDGPU gfx11 architecture.

Depends on D125498

Reviewed By: critson, rampitec, #amdgpu

Differential Revision: https://reviews.llvm.org/D125820

2 years ago[libc++] Granularize algorithm benchmarks
Nikolas Klauser [Thu, 19 May 2022 10:50:02 +0000 (12:50 +0200)]
[libc++] Granularize algorithm benchmarks

Reviewed By: ldionne, #libc

Spies: libcxx-commits, mgorny, mgrang

Differential Revision: https://reviews.llvm.org/D124740

2 years ago[flang][NFC] Allow whitespaces before `ERROR`
Daniil Dudkin [Thu, 19 May 2022 14:11:51 +0000 (17:11 +0300)]
[flang][NFC] Allow whitespaces before `ERROR`

This change allows to write whitespaces before the `ERROR` keyword
in semantic tests for consistency with other testing infrastructure.

Also, one test is changed in order to test if the change works
correctly.

Reviewed By: Meinersbur

Differential Revision: https://reviews.llvm.org/D125884

2 years ago[libc++] Enable move semantics for vector in C++03
Nikolas Klauser [Thu, 19 May 2022 10:46:09 +0000 (12:46 +0200)]
[libc++] Enable move semantics for vector in C++03

We require move semantics in C++03 anyways, so let's enable them for the containers.

Reviewed By: ldionne, #libc

Spies: libcxx-commits

Differential Revision: https://reviews.llvm.org/D123802

2 years ago[AArch64][SVE] Convert SRSHL to LSL when the fed from an ABS intrinsic
Bradley Smith [Fri, 6 May 2022 14:45:56 +0000 (14:45 +0000)]
[AArch64][SVE] Convert SRSHL to LSL when the fed from an ABS intrinsic

Differential Revision: https://reviews.llvm.org/D125233

2 years agoAdd an option to fill container for ref
Utkarsh Saxena [Wed, 18 May 2022 20:08:53 +0000 (22:08 +0200)]
Add an option to fill container for ref

This allows index implementations to fill container details when required specially when computing containerID is expensive.

Differential Revision: https://reviews.llvm.org/D125925

2 years ago[SLP][NFC] Pre-commit test showing vectorization preventing FMA
William Schmidt [Wed, 18 May 2022 20:51:47 +0000 (13:51 -0700)]
[SLP][NFC] Pre-commit test showing vectorization preventing FMA

When we generate a horizontal reduction of floating adds fed by a vectorized
tree rooted at floating multiplies, we should account for the cost of no
longer being able to generate scalar FMAs.  Similarly, if we vectorize a
list of floating multiplies that each feeds a single floating add, we should
again account for this cost.

The first test was reduced from a case where the vectorizable tree looked
barely profitable (cost -1) with a horizontal reduction, but produced
substantially worse code than allowing the FMAs to be generated.  The second
test was derived from the first: we again generate a horizontal reduction
here, but even if the horizontal reduction is forced to be unprofitable, we
try to vectorize the multiplies.  I have follow-up patches to address these
issues.

Differential Revision: https://reviews.llvm.org/D124867

2 years ago[lldb] Add --show-tags option to "memory find"
David Spickett [Wed, 13 Apr 2022 13:32:18 +0000 (14:32 +0100)]
[lldb] Add --show-tags option to "memory find"

This is off by default. If you get a result and that
memory has memory tags, when --show-tags is given you'll
see the tags inline with the memory content.

```
(lldb) memory read mte_buf mte_buf+64 --show-tags
<...>
0xfffff7ff8020: 00 00 00 00 00 00 00 00 0d f0 fe ca 00 00 00 00 ................ (tag: 0x2)
<...>
(lldb) memory find -e 0xcafef00d mte_buf mte_buf+64 --show-tags
data found at location: 0xfffff7ff8028
0xfffff7ff8028: 0d f0 fe ca 00 00 00 00 00 00 00 00 00 00 00 00 ................ (tags: 0x2 0x3)
0xfffff7ff8038: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ (tags: 0x3 0x4)
```

The logic for handling alignments is the same as for memory read
so in the above example because the line starts misaligned to the
granule it covers 2 granules.

Depends on D125089

Reviewed By: omjavaid

Differential Revision: https://reviews.llvm.org/D125090

2 years ago[M68k] Fix a bug in disassembler
Sheng [Thu, 19 May 2022 13:17:29 +0000 (21:17 +0800)]
[M68k] Fix a bug in disassembler

Sorry for my reckless patch. In some cases `RoundUp` is less than
the bit width of APInt. We need to check this before we do zext.

2 years ago[AArch64] Fix zero element TBL indices
David Green [Thu, 19 May 2022 12:54:35 +0000 (13:54 +0100)]
[AArch64] Fix zero element TBL indices

A TBL instruction will fill out-of-range values with 0's, something used
in D121139 to turn tbl2 with a zero input into tbl1s. This works OK for
v16i8, but for v8i8 the input is still treated as a v16i8, so
out-of-range values (like a lane index of 8) would end up loading values
from the top half of the input register. Clean this up by detecting the
out of range values and making sure they really use out of range values.
There is a fix for swapped indices of 64bit input vectors too, which
could be incorrectly adjusted if the zerovector was the first operand.

Fixes #55545

Differential Revision: https://reviews.llvm.org/D125865

2 years ago[NFC][M68k] Replace `APInt::zextOrSelf` with `APInt::zext`
Sheng [Thu, 19 May 2022 12:43:56 +0000 (20:43 +0800)]
[NFC][M68k] Replace `APInt::zextOrSelf` with `APInt::zext`

This is a follow up to D125558

2 years agoReland "[lldb] Add --all option to "memory region""
David Spickett [Thu, 19 May 2022 11:57:44 +0000 (12:57 +0100)]
Reland "[lldb] Add --all option to "memory region""

This reverts commit 3e928c4b9dfb01efd2cb968795e605760828e873.

This fixes an issue seen on Windows where we did not properly
get the section names of regions if they overlapped. Windows
has regions like:
[0x00007fff928db000-0x00007fff949a0000) ---
[0x00007fff949a0000-0x00007fff949a1000) r-- PECOFF header
[0x00007fff949a0000-0x00007fff94a3d000) r-x .hexpthk
[0x00007fff949a0000-0x00007fff94a85000) r-- .rdata
[0x00007fff949a0000-0x00007fff94a88000) rw- .data
[0x00007fff949a0000-0x00007fff94a94000) r-- .pdata
[0x00007fff94a94000-0x00007fff95250000) ---

I assumed that you could just resolve the address and get the section
name using the start of the region but here you'd always get
"PECOFF header" because they all have the same start point.

The usual command repeating loop used the end address of the previous
region when requesting the next, or getting the section name.
So I've matched this in the --all scenario.

In the example above, somehow asking for the region at
0x00007fff949a1000 would get you a region that starts at
0x00007fff949a0000 but has a different end point. Using the load
address you get (what I assume is) the correct section name.

2 years ago[AArch64] Extend zero vector TBL codegen tests. NFC
David Green [Thu, 19 May 2022 12:01:55 +0000 (13:01 +0100)]
[AArch64] Extend zero vector TBL codegen tests. NFC

2 years ago[flang][driver] Make driver accept `-module-dir<value>`
Andrzej Warzynski [Thu, 19 May 2022 09:31:03 +0000 (09:31 +0000)]
[flang][driver] Make driver accept `-module-dir<value>`

`-module-dir` is Flang's equivalent for `-J` from GFortran (in fact,
`-J` is an alias for `-module-dir` in Flang). Currently, only
`-module-dir <value>` is accepted. However, `-J` (and other options for
specifying various paths) accepts `-J<value>` as well as `-J <value>`.
This patch makes sure that `-module-dir` behaves consistently with other
such flags.

Differential Revision: https://reviews.llvm.org/D125957

2 years ago[AMDGPU][MC][GFX940] Add SMFMAC aliases
Dmitry Preobrazhensky [Thu, 19 May 2022 10:40:48 +0000 (13:40 +0300)]
[AMDGPU][MC][GFX940] Add SMFMAC aliases

Differential Revision: https://reviews.llvm.org/D125888

2 years ago[APInt] Deprecate truncOrSelf, zextOrSelf and sextOrSelf
Jay Foad [Wed, 6 Oct 2021 11:04:03 +0000 (12:04 +0100)]
[APInt] Deprecate truncOrSelf, zextOrSelf and sextOrSelf

Differential Revision: https://reviews.llvm.org/D125558

2 years ago[APInt] Remove all uses of zextOrSelf, sextOrSelf and truncOrSelf
Jay Foad [Wed, 6 Oct 2021 09:54:07 +0000 (10:54 +0100)]
[APInt] Remove all uses of zextOrSelf, sextOrSelf and truncOrSelf

Most clients only used these methods because they wanted to be able to
extend or truncate to the same bit width (which is a no-op). Now that
the standard zext, sext and trunc allow this, there is no reason to use
the OrSelf versions.

The OrSelf versions additionally have the strange behaviour of allowing
extending to a *smaller* width, or truncating to a *larger* width, which
are also treated as no-ops. A small amount of client code relied on this
(ConstantRange::castOp and MicrosoftCXXNameMangler::mangleNumber) and
needed rewriting.

Differential Revision: https://reviews.llvm.org/D125557

2 years ago[AMDGPU][NFC] Fix FileCheck directives in phi-vgpr-input-moveimm.mir.
Ivan Kosarev [Thu, 19 May 2022 10:19:26 +0000 (11:19 +0100)]
[AMDGPU][NFC] Fix FileCheck directives in phi-vgpr-input-moveimm.mir.

Discovered with D125604.

Reviewed By: #amdgpu, arsenm

Differential Revision: https://reviews.llvm.org/D125900

2 years ago[clangd] Update the test after diagnostic message change
Kirill Bobyrev [Thu, 19 May 2022 10:03:31 +0000 (12:03 +0200)]
[clangd] Update the test after diagnostic message change

2 years ago[clangd] NFC: Clarify the Include Cleaner warning
Kirill Bobyrev [Thu, 19 May 2022 09:59:00 +0000 (11:59 +0200)]
[clangd] NFC: Clarify the Include Cleaner warning

2 years ago[RISCV][SelectionDAG] Support VECREDUCE_ADD mask operation
Lian Wang [Wed, 11 May 2022 08:02:42 +0000 (08:02 +0000)]
[RISCV][SelectionDAG] Support VECREDUCE_ADD mask operation

Re-landed D125206

Reviewed By: craig.topper

Differential Revision: https://reviews.llvm.org/D125206

2 years ago[release] Add cmake as an extra tarball and not bundle it
Konrad Kleine [Thu, 19 May 2022 09:12:19 +0000 (11:12 +0200)]
[release] Add cmake as an extra tarball and not bundle it

Revert "Add cmake/ to release tarballs via concatenation"

This reverts commit 3a33664e8838e8b77acd1bbb13b1cf5e580a1077.

Revert "Add cmake to source release tarballs"

This reverts commit 32a0482a65b86cf0d84ea7e784cca9852df5c67c.

Reviewed By: tstellar, aaronpuchert

Differential Revision: https://reviews.llvm.org/D125798

2 years ago[WebAssembly][NFC] Fix errant tabs in test case in last commit
Alex Bradbury [Thu, 19 May 2022 09:09:30 +0000 (10:09 +0100)]
[WebAssembly][NFC] Fix errant tabs in test case in last commit

[4e8b2ac](https://reviews.llvm.org/rG4e8b2ac7c019) contained unintended
tabs. This commit fixes that.

2 years ago[libc] Apply no-builtin everywhere, remove unnecessary flags
Guillaume Chatelet [Wed, 18 May 2022 15:09:09 +0000 (15:09 +0000)]
[libc] Apply no-builtin everywhere, remove unnecessary flags

Some functions like `stpncpy` are implemented in terms of `memset` but are not
currently using `-fno-builtin-memset`. This is somewhat hidden by the fact that
we use `-ffreestanding` globally and that `-ffreestanding` implies
`-fno-builtin` for Clang.

This patch also removes `-mllvm -combiner-global-alias-analysis` that is Clang
specific and that does not bring substantial gains on modern processors.

Also we keep `-mllvm --tail-merge-threshold=0` for aarch64 in CMakeLists.txt
but we omit it in the Bazel config. This is because Bazel consumes the source
files directly and so it can use PGO to take optimal decisions locally.

Differential Revision: https://reviews.llvm.org/D125894

2 years ago[WebAssembly] Fix bug where -no-type-check failed to completely disable the typechecker
Alex Bradbury [Thu, 19 May 2022 09:06:02 +0000 (10:06 +0100)]
[WebAssembly] Fix bug where -no-type-check failed to completely disable the typechecker

Related to <https://github.com/llvm/llvm-project/issues/55566>.
Committing directly (per LLVM's code review policy) as this is a trivial
fix.

2 years ago[gn build] Port 4df795bff752
LLVM GN Syncbot [Thu, 19 May 2022 08:04:45 +0000 (08:04 +0000)]
[gn build] Port 4df795bff752

2 years ago[Serialization] Add missing includes for CHAR_BIT
Sam McCall [Thu, 19 May 2022 08:04:25 +0000 (10:04 +0200)]
[Serialization] Add missing includes for CHAR_BIT

2 years ago[LegalizeVectorTypes][VP] Add widen and split support for VP_SETCC
Lian Wang [Thu, 12 May 2022 03:31:13 +0000 (03:31 +0000)]
[LegalizeVectorTypes][VP] Add widen and split support for VP_SETCC

Reviewed By: craig.topper, frasercrmck

Differential Revision: https://reviews.llvm.org/D125446

2 years ago[libunwind] Remove -Wsign-conversion warning
Daniel Kiss [Thu, 19 May 2022 07:38:30 +0000 (09:38 +0200)]
[libunwind] Remove -Wsign-conversion warning

Reland after dependent change reland.

2 years ago[Serialization] Delta-encode consecutive SourceLocations in TypeLoc
Sam McCall [Wed, 11 May 2022 13:42:31 +0000 (15:42 +0200)]
[Serialization] Delta-encode consecutive SourceLocations in TypeLoc

Much of the size of PCH/PCM files comes from stored SourceLocations.
These are encoded using (almost) their raw value, VBR-encoded. Absolute
SourceLocations can be relatively large numbers, so this commonly takes
20-30 bits per location.

We can reduce this by exploiting redundancy: many "nearby" SourceLocations are
stored differing only slightly and can be delta-encoded.
Randam-access loading of AST nodes constrains how long these sequences
can be, but we can do it at least within a node that always gets
deserialized as an atomic unit.

TypeLoc is implemented in this patch as it's a relatively small change
that shows most of the API.
This saves ~3.5% of PCH size, I have local changes applying this technique
further that save another 3%, I think it's possible to get to 10% total.

Differential Revision: https://reviews.llvm.org/D125403

2 years ago[LegalizeTypes][VP] Add integer promotions support for VP_TRUNCATE
Lian Wang [Tue, 17 May 2022 01:47:32 +0000 (01:47 +0000)]
[LegalizeTypes][VP] Add integer promotions support for VP_TRUNCATE

Reviewed By: craig.topper

Differential Revision: https://reviews.llvm.org/D125739

2 years ago[LegalizeTypes][VP][NFC] Use an if and two returns instead of ?: operator
Lian Wang [Wed, 18 May 2022 07:44:26 +0000 (07:44 +0000)]
[LegalizeTypes][VP][NFC] Use an if and two returns instead of ?: operator

Reviewed By: frasercrmck

Differential Revision: https://reviews.llvm.org/D125858

2 years ago[clangd] Suppress warning: control reaches end of function
Sam McCall [Thu, 19 May 2022 06:26:13 +0000 (08:26 +0200)]
[clangd] Suppress warning: control reaches end of function

2 years ago[pseudo] Squash some warnings. NFC
Sam McCall [Thu, 19 May 2022 06:18:34 +0000 (08:18 +0200)]
[pseudo] Squash some warnings. NFC

Explicitly sizing Kind enum suggests that too-large values are allowed,
and that putting it in a bitfield is dangerous.

GCC doesn't like condition ? integer : enum.

2 years ago[gn build] Port 03ea140b3a28
LLVM GN Syncbot [Thu, 19 May 2022 06:13:53 +0000 (06:13 +0000)]
[gn build] Port 03ea140b3a28

2 years ago[clang-tidy] Fix logic of assertion
Sam McCall [Thu, 19 May 2022 06:13:02 +0000 (08:13 +0200)]
[clang-tidy] Fix logic of assertion

2 years agoReland(3) "[clangd] Indexing of standard library"
Sam McCall [Wed, 18 May 2022 17:27:43 +0000 (19:27 +0200)]
Reland(3) "[clangd] Indexing of standard library"

Tracked down the crash, which was argument-evaluation-order UB
in the wrapping indexStandardLibrary().
Sorry for the churn!

This reverts commit 77533ea443aca6e9978d7c8a6822420f8345f6af.

2 years ago[mlir][python] Add Python bindings for ml_program dialect.
Stella Laurenzo [Wed, 18 May 2022 05:42:39 +0000 (22:42 -0700)]
[mlir][python] Add Python bindings for ml_program dialect.

Differential Revision: https://reviews.llvm.org/D125852

2 years ago[mlir] Add GlobalOp, GlobalLoadConstOp to ml_program.
Stella Laurenzo [Sat, 23 Apr 2022 02:59:34 +0000 (19:59 -0700)]
[mlir] Add GlobalOp, GlobalLoadConstOp to ml_program.

The approach I took was to define a dialect 'extern' attribute that a GlobalOp can take as a value to signify external linkage. I think this approach should compose well and should also work with wherever the OpaqueElements work goes in the future (since that is just another kind of attribute). I special cased the GlobalOp parser/printer for this case because it is significantly easier on the eyes.

In the discussion, Jeff Niu had proposed an alternative syntax for GlobalOp that I ended up not taking. I did try to implement it but a) I don't think it made anything easier to read in the common case, and b) it made the parsing/printing logic a lot more complicated (I think I would need a completely custom parser/printer to do it well). Please have a look at the common cases where the global type and initial value type match: I don't think how I have it is too bad. The less common cases seem ok to me.

I chose to only implement the direct, constant load op since that is non side effecting and there was still discussion pending on that.

Differential Revision: https://reviews.llvm.org/D124318

2 years agoFix if statement in DebugInfo/GSYM/LookupResult.cpp
Alexander Pivovarov [Wed, 18 May 2022 23:04:05 +0000 (16:04 -0700)]
Fix if statement in DebugInfo/GSYM/LookupResult.cpp

2 years ago[LV] Widen freeze instead of scalarizing it
lizhijin [Thu, 19 May 2022 04:20:33 +0000 (12:20 +0800)]
[LV] Widen freeze instead of scalarizing it

This patch changes the strategy for vectorizing freeze instrucion, from
replicating multiple times to widening according to selected VF.

Fixes #54992

Reviewed By: fhahn

Differential Revision: https://reviews.llvm.org/D125016

2 years ago[NFC][RISCV] Enable TuneNoDefaultUnroll feature to control targets which use default...
Zi Xuan Wu (Zeson) [Thu, 19 May 2022 04:17:29 +0000 (12:17 +0800)]
[NFC][RISCV] Enable TuneNoDefaultUnroll feature to control targets which use default unroll preference

In RISCVTargetTransformInfo, enumerating the processor family is not a good way to predict.
Because it needs to enumerate many subtarget family and is hard to update if add new subtarget.
Instead, create a feature to distinguish whether targets want to use default unroll preference or not.

Keep TuneSiFive7 because it's flag to indicate subtarget family, which may used in other place.

Differential Revision: https://reviews.llvm.org/D125741

2 years ago[lldb/test] Fix failures caused by a previous PExpect.launch change
Med Ismail Bennani [Thu, 19 May 2022 04:12:15 +0000 (21:12 -0700)]
[lldb/test] Fix failures caused by a previous PExpect.launch change

This should fix the issues introduced by d71d1a9, which skipped all the
test setup commands.

This also fixes the test failures happening in TestAutosuggestion.py.

Signed-off-by: Med Ismail Bennani <medismail.bennani@gmail.com>
2 years agoRevert "[lldb/test] Make some tests as XFAIL while I investigate the issue"
Med Ismail Bennani [Thu, 19 May 2022 02:59:49 +0000 (19:59 -0700)]
Revert "[lldb/test] Make some tests as XFAIL while I investigate the issue"

This reverts commit 80589f272c200798b57a5151680a993bc2cc00a7.

2 years ago[InstCombine] (rot X, ?) == 0/-1 --> X == 0/-1
Chenbing Zheng [Thu, 19 May 2022 03:22:26 +0000 (11:22 +0800)]
[InstCombine] (rot X, ?) == 0/-1 --> X == 0/-1

In this patch we add a function foldICmpInstWithConstantAllowUndef
to fold integer comparisons with a constant operand: icmp Pred X, C
where X is some kind of instruction and C is AllowUndef.

We move this fold to the new function, so that it can solve undef elts in a vector.

Reviewed By: spatel, RKSimon

Differential Revision: https://reviews.llvm.org/D125220

2 years ago[InstCombine] Allow undef vectors when foldSelectToCopysign
Chenbing Zheng [Thu, 19 May 2022 02:57:49 +0000 (10:57 +0800)]
[InstCombine] Allow undef vectors when foldSelectToCopysign

Reviewed By: spatel

Differential Revision: https://reviews.llvm.org/D125671

2 years agoFix an or+and miscompile w/ GlobalISel
Jon Roelofs [Thu, 19 May 2022 01:34:49 +0000 (18:34 -0700)]
Fix an or+and miscompile w/ GlobalISel

Fixes #55284

2 years agoRevert 821ee172cdcd7196b6130321b53b6cc66bf1222b
Med Ismail Bennani [Thu, 19 May 2022 02:07:47 +0000 (19:07 -0700)]
Revert 821ee172cdcd7196b6130321b53b6cc66bf1222b

This reverts commit 821ee172cdcd7196b6130321b53b6cc66bf1222b, that
landed by mistake.

2 years ago[lldb/test] Make some tests as XFAIL while I investigate the issue
Med Ismail Bennani [Thu, 19 May 2022 01:51:10 +0000 (18:51 -0700)]
[lldb/test] Make some tests as XFAIL while I investigate the issue

This is very likely to be caused by d71d1a947bee1247e952f22c13ad3ed3d041e36a.

Signed-off-by: Med Ismail Bennani <medismail.bennani@gmail.com>
2 years agodyld patch
Med Ismail Bennani [Mon, 16 May 2022 21:59:28 +0000 (14:59 -0700)]
dyld patch

Signed-off-by: Med Ismail Bennani <medismail.bennani@gmail.com>
2 years ago[lldb/crashlog] Fix line entries resolution in interactive mode
Med Ismail Bennani [Wed, 18 May 2022 21:18:14 +0000 (14:18 -0700)]
[lldb/crashlog] Fix line entries resolution in interactive mode

This patch subtracts 1 to the pc of any frame above frame 0 to get the
previous line entry and display the right line in the debugger.

This also rephrase some old comment from `48d157dd4`.

rdar://92686666

Differential Revision: https://reviews.llvm.org/D125928

Signed-off-by: Med Ismail Bennani <medismail.bennani@gmail.com>
2 years ago[lldb/Core] Fix "sticky" long progress messages
Med Ismail Bennani [Thu, 19 May 2022 01:16:59 +0000 (18:16 -0700)]
[lldb/Core] Fix "sticky" long progress messages

When the terminal window is too small, lldb would wrap progress messages
accross multiple lines which would break the progress event handling
code that is supposed to clear the message once the progress is completed.

This causes the progress message to remain on the screen, sometimes partially,
which can be confusing for the user.

To fix this issue, this patch trims the progress message to the terminal
width taking into account the progress counter leading the message for
finite progress events and also the trailing `...`.

rdar://91993836

Differential Revision: https://reviews.llvm.org/D124785

Signed-off-by: Med Ismail Bennani <medismail.bennani@gmail.com>
2 years ago[lldb/Test] Add `use_colors` argument to the PExpect.launch wrapper
Med Ismail Bennani [Wed, 18 May 2022 18:35:20 +0000 (11:35 -0700)]
[lldb/Test] Add `use_colors` argument to the PExpect.launch wrapper

This patch adds a new `use_colors` argument to the PExpect.launch
method.

As the name suggests, it allows the user to conditionally enable color
support in the debugger, which can be helpful to test functionalities that
rely on that, like progress reporting. It defaults to False.

Differential Revision: https://reviews.llvm.org/D125915

Signed-off-by: Med Ismail Bennani <medismail.bennani@gmail.com>
2 years ago[llvm/Support] Fallback to $TERM if terminfo has no "colors" capability
Med Ismail Bennani [Wed, 18 May 2022 18:33:31 +0000 (11:33 -0700)]
[llvm/Support] Fallback to $TERM if terminfo has no "colors" capability

It can happen on macOS that terminal doesn't report the "colors"
capability in the terminfo database, in which case `tigetnum` returns -1.

This doesn't mean however that the terminal doesn't supports color, it
just means that the capability is absent from the terminal description.

In that case, we should still fallback to the checking the $TERM
environment variable to see if it supports ANSI escapes codes.

Differential Revision: https://reviews.llvm.org/D125914

Signed-off-by: Med Ismail Bennani <medismail.bennani@gmail.com>
2 years ago[M68k][Disassembler] Fix decoding conflict
Sheng [Thu, 19 May 2022 01:09:43 +0000 (09:09 +0800)]
[M68k][Disassembler] Fix decoding conflict

This diff fixes decoding conflict between these pair of instructions:

ADD(16|32)dd / ADD(16|32)dr
SUB(16|32)dd / SUB(16|32)dr
AND(16|32)dd / AND(16|32)dr
OR(16|32)dd  / OR(16|32)dr

Reviewed By: ricky26

Differential Revision: https://reviews.llvm.org/D125861

2 years ago[Sparc] Make sure that we really don't emit quad-precision unless the "hard-quad...
Mark Kettenis [Thu, 19 May 2022 00:05:56 +0000 (20:05 -0400)]
[Sparc] Make sure that we really don't emit quad-precision unless the "hard-quad-float" feature is available

Make sure that we really don't emit quad-precision unless the "hard-quad-float"
feature is available. Add missing replacement instruction patterns that are
needed to emit alternative code for conditional moves of quad-precision floats.

Test from koakuma.

Reviewed By: koakuma

Differential Revision: https://reviews.llvm.org/D119104

2 years ago[flang] test conforming & non-conforming lcobound
Damian Rouson [Thu, 14 Apr 2022 00:03:33 +0000 (17:03 -0700)]
[flang] test conforming & non-conforming lcobound

Add a test with standard-conforming  non-conforming lcobound()
intrinsic function invocations.  Also test that several
non-conforming lcobound() invocations generate the correct error
messages.

Differential Revision: https://reviews.llvm.org/DD123747

2 years agoExtend switch condition in optimizeSwitchPhiConst when free
Matthias Braun [Tue, 3 May 2022 17:01:54 +0000 (10:01 -0700)]
Extend switch condition in optimizeSwitchPhiConst when free

In a case like:

    switch((i32)x) { case 42: phi((i64)42, ...); }

replace `(i64)42` with `zext(x)` when we can do so for free.

This fixes a part of https://github.com/llvm/llvm-project/issues/55153

Differential Revision: https://reviews.llvm.org/D124897

2 years ago[BOLT] Testcase to repro R_X86_64_REX_GOTPCRELX bug
Rafael Auler [Thu, 17 Feb 2022 01:01:19 +0000 (17:01 -0800)]
[BOLT] Testcase to repro R_X86_64_REX_GOTPCRELX bug

Add a new testcase that reproduces a bug when BOLTing current
trunk LLD bootstrapped with trunk clang. This makes it official
that we do not support this transformation but are working on
it. When the support is ready, XFAIL should be removed.

Reviewed By: maksfb, Amir, yota9

Differential Revision: https://reviews.llvm.org/D125843