platform/upstream/llvm.git
2 years ago[MLIR][Presburger] Disambiguate call to floor
Rainer Orth [Wed, 9 Feb 2022 21:01:55 +0000 (22:01 +0100)]
[MLIR][Presburger] Disambiguate call to floor

While testing LLVM 14.0.0 rc1 on Solaris, compilation of `FAIL`ed with

  /var/llvm/llvm-14.0.0-rc1/rc1/llvm-project/mlir/lib/Analysis/Presburger/Utils.cpp: In lambda function:
  /var/llvm/llvm-14.0.0-rc1/rc1/llvm-project/mlir/lib/Analysis/Presburger/Utils.cpp:48:58: error: call of overloaded ‘floor(int64_t)’ is ambiguous
     48 |                  [gcd](int64_t &n) { return floor(n / gcd); });
        |                                                          ^
  ...
  /usr/gcc/10/lib/gcc/sparcv9-sun-solaris2.11/10.3.0/include-fixed/iso/math_iso.h:201:21:
note: candidate: ‘long double std::floor(long double)’
    201 |  inline long double floor(long double __X) { return __floorl(__X); }
        |                     ^~~~~
  /usr/gcc/10/lib/gcc/sparcv9-sun-solaris2.11/10.3.0/include-fixed/iso/math_iso.h:165:15:
note: candidate: ‘float std::floor(float)’
    165 |  inline float floor(float __X) { return __floorf(__X); }
        |               ^~~~~
  /usr/gcc/10/lib/gcc/sparcv9-sun-solaris2.11/10.3.0/include-fixed/iso/math_iso.h:78:15:
note: candidate: ‘double std::floor(double)’
     78 | extern double floor __P((double));
        |               ^~~~~

The same issue had already occured in the past, cf. D108750
<https://reviews.llvm.org/D108750>, and the solution is the same: cast the
`floor` arg to `double`.

Tested on `amd64-pc-solaris2.11` and `sparcv9-sun-solaris2.11`.

Differential Revision: https://reviews.llvm.org/D119324

2 years ago[libunwind] Avoid a warning in 32 bit builds. NFC.
Martin Storsjö [Wed, 9 Feb 2022 17:36:58 +0000 (19:36 +0200)]
[libunwind] Avoid a warning in 32 bit builds. NFC.

The warning was introduced with the recently merged SPARCv9
support in 2b9554b8850192bdd86c02eb671de1d866df8d87.

The cast matches the existing surrounding cases.

Differential Revision: https://reviews.llvm.org/D119353

2 years ago[mlir][sparse] Rename index_t to index_type again
Rainer Orth [Wed, 9 Feb 2022 20:59:52 +0000 (21:59 +0100)]
[mlir][sparse] Rename index_t to index_type again

While testing LLVM 14.0.0 rc1 on Solaris, I ran into a compile failure:

                   from /var/llvm/llvm-14.0.0-rc1/rc1/llvm-project/mlir/lib/ExecutionEngine/SparseTensorUtils.cpp:22:
  /usr/include/sys/types.h:103:16: error: conflicting declaration ‘typedef short int index_t’
    103 | typedef short  index_t;
        |                ^~~~~~~
  In file included from
/var/llvm/llvm-14.0.0-rc1/rc1/llvm-project/mlir/lib/ExecutionEngine/SparseTensorUtils.cpp:17:
  /var/llvm/llvm-14.0.0-rc1/rc1/llvm-project/mlir/include/mlir/ExecutionEngine/SparseTensorUtils.h:26:7:
note: previous declaration as ‘using index_t = uint64_t’
     26 | using index_t = uint64_t;
        |       ^~~~~~~

The same issue had already occured in the past and fixed in D72619
<https://reviews.llvm.org/D72619>.  More detailed explanation can also be
found there.

Tested on `amd64-pc-solaris2.11` and `sparcv9-solaris2.11`.

Differential Revision: https://reviews.llvm.org/D119323

2 years ago[SCEV] Unwind SCEVUnionPredicate from getPredicatedBackedgeTakenCount [NFC]
Philip Reames [Wed, 9 Feb 2022 20:16:10 +0000 (12:16 -0800)]
[SCEV] Unwind SCEVUnionPredicate from getPredicatedBackedgeTakenCount [NFC]

For those curious, the whole reason for tracking the predicate set seperately as opposed to just immediately registering the dependencies appears to be allowing the printing code to print a result without changing the PSE state.  It's slightly questionable if this justifies the complexity, but since we can preserve it with local ugliness, I did so.

2 years ago[libc] undefined reference in LibcTest.cpp
Guillaume Chatelet [Tue, 8 Feb 2022 14:52:50 +0000 (14:52 +0000)]
[libc] undefined reference in LibcTest.cpp

GCC complains about undefined reference in LibcTest.cpp and indeed the wrong version of the template has been explicitly instanciated.
This is necessary to get llvm-libc compile with GCC.

Mentionning D119002 here for navigability.

Differential Revision: https://reviews.llvm.org/D119242

2 years ago[flang] Upstream runtime changes for inquiry intrinsics
Peter Steinfeld [Wed, 9 Feb 2022 19:17:18 +0000 (11:17 -0800)]
[flang] Upstream runtime changes for inquiry intrinsics

This change adds runtime routines and tests for LBOUND when passed a DIM argument, SIZE, and UBOUND when not passed a DIM argument.

Associated changes for lowering have already been merged into fir-dev.

Differential Revision: https://reviews.llvm.org/D119360

2 years ago[Libomptarget] Increase stack size for bug49779 test
Joseph Huber [Wed, 9 Feb 2022 18:43:56 +0000 (13:43 -0500)]
[Libomptarget] Increase stack size for bug49779 test

The 'bug49779.cpp' test has been failing recently. This is because the
runtime is sufficiently complex when using nested parallelism without
optimizations that the CUDA tools cannot statically determine the stack
size. Because of this the kernel can exceed the thread stack size and
crash. Work around this using the 'LIBOMPTARGET_STACK_SIZE' environment
variable and add an FAQ entry for this situation.

Fixes #53670

Reviewed By: Meinersbur

Differential Revision: https://reviews.llvm.org/D119357

2 years ago[RISCV] Teach RISCVDAGToDAGISel::selectShiftMask to replace sub from constant with...
Craig Topper [Wed, 9 Feb 2022 20:03:22 +0000 (12:03 -0800)]
[RISCV] Teach RISCVDAGToDAGISel::selectShiftMask to replace sub from constant with neg.

If the shift amount is (sub C, X) where C is 0 modulo the size of
the shift, we can replace it with neg or negw.

Similar is is done for AArch64 and X86.

Reviewed By: khchen

Differential Revision: https://reviews.llvm.org/D119089

2 years ago[mlir] Use sparse-compiler pass in sparse benchmark
Saurabh Jha [Wed, 9 Feb 2022 20:10:30 +0000 (20:10 +0000)]
[mlir] Use sparse-compiler pass in sparse benchmark

This patch uses sparse-compiler option in mlir sparse tensor benchmark

Reviewed By: aartbik

Differential Revision: https://reviews.llvm.org/D118579

2 years ago[SCEV] Remove conversion to SCEVUnionPredicate in ExitNotTakenInfo [NFC]
Philip Reames [Wed, 9 Feb 2022 20:04:25 +0000 (12:04 -0800)]
[SCEV] Remove conversion to SCEVUnionPredicate in ExitNotTakenInfo [NFC]

This removes one of the places where we mutate an existing union predicate.

2 years ago[InstCombine] improve test name and cleanup; NFC
Sanjay Patel [Wed, 9 Feb 2022 20:05:12 +0000 (15:05 -0500)]
[InstCombine] improve test name and cleanup; NFC

The last test was intended to be a negative test with D114272,
but a more general fold can reduce it, so that is renamed.

These are all single-block tests, so remove the unnecessary labels.

2 years agoRevert "[LV] Remove `LoopVectorizationCostModel::useEmulatedMaskMemRefHack()`"
David Green [Wed, 9 Feb 2022 20:02:54 +0000 (20:02 +0000)]
Revert "[LV] Remove `LoopVectorizationCostModel::useEmulatedMaskMemRefHack()`"

This reverts commit 77a0da926c9ea86afa9baf28158d79c7678fc6b9 as we've
received multiple reports of this significantly impacting performance,
in ways that don't seem to just be target specific cost models going
wrong. I would offer some reproducers, but the test changes here seem to
be full of them!

Reverting for now and hopefully we can remove the "hack" more carefully
as we go.

2 years ago[DWARF][codegen] Fix for Aranges when split inlining is present
Alexander Yermolovich [Wed, 9 Feb 2022 19:51:32 +0000 (11:51 -0800)]
[DWARF][codegen] Fix for Aranges when split inlining is present

When we enable -fsplit-dwarf-inlining we end up with two entries
in .debug_aranges for each CU. Because it processes Skeleton CU
inline information and DWO CU.

Furthermore address calculations were incorrect because we were processing sections in Skeleton CU.

Reviewed By: dblaikie

Differential Revision: https://reviews.llvm.org/D118857

2 years ago[RISCV] Add a really basic cost model for SK_Splice.
Craig Topper [Wed, 9 Feb 2022 19:37:26 +0000 (11:37 -0800)]
[RISCV] Add a really basic cost model for SK_Splice.

While testing scalable vectors I found that if we generate a
vector splice intrinsic and run the code through the loop unroller,
we'll crash due to an invalid cost.

This adds a basic cost based on the 2 slide instructions used by the
lowering in D119303.

We probably need to factor LMUL into this, but that's true for
arithmetic instructions too. So I've ignored for the moment.

Reviewed By: ABataev

Differential Revision: https://reviews.llvm.org/D119316

2 years ago[lldb] Disable failing test on Windows
Andy Yankovsky [Wed, 9 Feb 2022 19:40:50 +0000 (20:40 +0100)]
[lldb] Disable failing test on Windows

Test was introduced in https://reviews.llvm.org/D113498.

2 years ago[MC] clang-format llvm/MC/MCContext.h. NFC
Fangrui Song [Wed, 9 Feb 2022 19:38:28 +0000 (11:38 -0800)]
[MC] clang-format llvm/MC/MCContext.h. NFC

People get clang-format lints when changing declarations in the large block,
e.g. ef736a1c39f27ef4777bbe6cf3af68c1fd588c8a Cleanup LLVMMC headers.
Just format it to reduce hassle in the future.

2 years ago[RISCV][VP] Lower VP_FMA to RVV instructions.
Craig Topper [Wed, 9 Feb 2022 19:08:24 +0000 (11:08 -0800)]
[RISCV][VP] Lower VP_FMA to RVV instructions.

We already had FMA_VL node, but we didn't have masked patterns.
I have not added the fneg variations. I'll do those after I add
llvm.vp.fneg.

Reviewed By: frasercrmck

Differential Revision: https://reviews.llvm.org/D119196

2 years ago[NFC] Fix sign-compare warning in GrammarBNF thanks to int promotion
Erich Keane [Wed, 9 Feb 2022 19:25:58 +0000 (11:25 -0800)]
[NFC] Fix sign-compare warning in GrammarBNF thanks to int promotion

2 years ago[RISCV] Lower VP_FNEG to RVV instructions
Craig Topper [Wed, 9 Feb 2022 18:56:39 +0000 (10:56 -0800)]
[RISCV] Lower VP_FNEG to RVV instructions

Reviewed By: frasercrmck

Differential Revision: https://reviews.llvm.org/D119269

2 years ago[ELF][docs] Remove ignore -dc from ld.lld.1
Fangrui Song [Wed, 9 Feb 2022 18:38:36 +0000 (10:38 -0800)]
[ELF][docs] Remove ignore -dc from ld.lld.1

2 years ago[ELF] Remove obscure -dp and GNU ld incompatible --[no-]define-common, ignore -d/-dc
Fangrui Song [Wed, 9 Feb 2022 18:35:53 +0000 (10:35 -0800)]
[ELF] Remove obscure -dp and GNU ld incompatible --[no-]define-common, ignore -d/-dc

https://maskray.me/blog/2022-02-06-all-about-common-symbols#no-define-common

In GNU ld, -dc only affects -r links and causes COMMON symbols to be allocated.
--no-define-common is defined to make COMMON symbols undefined for -shared.
AIUI --no-define-common is a workaround around glibc 2.1 time and not really useful.

gold confuses --define-common with -d/FORCE_COMMON_ALLOCATION and implements
--define-common with -d semantics. Its --no-define-common is incompatible with
GNU ld.

In ld.lld, b2a23cf3c08cee45614f27eb2c6d044e506aa6a6 fixed the default -r
behavior for COMMON symbols but ported the incompatible gold
--[no-]define-common. To the best of my knowledge, no project uses -dp
--[no-]define-common. So just remove these options.

-d/-dc are used by the following projects:

* grub grub-core/genmod.sh.in uses -Wl,-r,-d (https://lists.gnu.org/archive/html/grub-devel/2022-02/msg00088.html)
* FreeBSD crunchgen uses -Wl,-dc (https://reviews.freebsd.org/D34215)

A no-op implementation works for them. Only when a program inspects relocatable
output by itself and does not recognize COMMON symbols, there may be a problem.
This is an extremely unlikely case.

Reviewed By: peter.smith

Differential Revision: https://reviews.llvm.org/D119108

2 years ago[SCEV] Add interface for constructing generic SCEVComparePredicate [NFC}
Philip Reames [Tue, 8 Feb 2022 19:39:22 +0000 (11:39 -0800)]
[SCEV] Add interface for constructing generic SCEVComparePredicate [NFC}

2 years ago[ConstraintElimination] Remove redundant lookup (NFC).
Florian Hahn [Wed, 9 Feb 2022 18:00:03 +0000 (18:00 +0000)]
[ConstraintElimination] Remove redundant lookup (NFC).

2 years ago[mlir][GPU] Add ShuffleOp builder for constant offset/width
Matthias Springer [Wed, 9 Feb 2022 16:21:55 +0000 (01:21 +0900)]
[mlir][GPU] Add ShuffleOp builder for constant offset/width

Differential Revision: https://reviews.llvm.org/D119345

2 years ago[gn build] Port f1984b143367
LLVM GN Syncbot [Wed, 9 Feb 2022 17:34:07 +0000 (17:34 +0000)]
[gn build] Port f1984b143367

2 years ago[gn build] Port 46a6f5ae148a
LLVM GN Syncbot [Wed, 9 Feb 2022 17:34:06 +0000 (17:34 +0000)]
[gn build] Port 46a6f5ae148a

2 years ago[ConstraintElimination] Move some definitions closer to uses (NFC).
Florian Hahn [Wed, 9 Feb 2022 17:29:48 +0000 (17:29 +0000)]
[ConstraintElimination] Move some definitions closer to uses (NFC).

2 years ago[LoopLoadElim] Support opaque pointers
Arthur Eubanks [Wed, 9 Feb 2022 17:22:21 +0000 (09:22 -0800)]
[LoopLoadElim] Support opaque pointers

With typed pointers the pointer operand type checks the address space
and the load/store type. With opaque pointers we have to check the
load/store type separately.

2 years ago[OpaquePtr][LoopAccessAnalysis] Support opaque pointers
Arthur Eubanks [Tue, 8 Feb 2022 02:16:08 +0000 (18:16 -0800)]
[OpaquePtr][LoopAccessAnalysis] Support opaque pointers

Previously we relied on the pointee type to determine what type we need
to do runtime pointer access checks.

With opaque pointers, we can access a pointer with more than one type,
so now we keep track of all the types we're accessing a pointer's
memory with.

Also some other minor getPointerElementType() removals.

Reviewed By: #opaque-pointers, nikic

Differential Revision: https://reviews.llvm.org/D119047

2 years ago[gn build] Port a1862d78eb45
Arthur Eubanks [Wed, 9 Feb 2022 17:10:54 +0000 (09:10 -0800)]
[gn build] Port a1862d78eb45

2 years ago[AMDGPU] Remove dead code from shrinkScalarLogicOp
Jay Foad [Wed, 9 Feb 2022 17:05:19 +0000 (17:05 +0000)]
[AMDGPU] Remove dead code from shrinkScalarLogicOp

It looks like this code has been dead since shrinkScalarLogicOp
was introduced in svn r348601.

2 years ago[lldb] Account for extra threads in TestGdbRemoteThreadsInStopReply on windows
Pavel Labath [Wed, 9 Feb 2022 16:38:27 +0000 (17:38 +0100)]
[lldb] Account for extra threads in TestGdbRemoteThreadsInStopReply on windows

After 9611282c, TestGdbRemoteThreadsInStopReply is not non-deterministic
-- instead it deterministically fails due to extra threads created by
std::thread thread pool.

Adjust the tests to account for that.

2 years ago[lldb] Constant-resolve operands to `getelementptr`
Andy Yankovsky [Tue, 9 Nov 2021 18:41:56 +0000 (19:41 +0100)]
[lldb] Constant-resolve operands to `getelementptr`

Operands to `getelementptr` can be constants or constant expressions. Check
that all operands can be constant-resolved and resolve them during the
evaluation. If some operands can't be resolved as constants -- the expression
evaluation will fallback to JIT.

Fixes: https://bugs.llvm.org/show_bug.cgi?id=52449

Reviewed By: #lldb, shafik

Differential Revision: https://reviews.llvm.org/D113498

2 years ago[libc++] Enables put_long_double test for glibc.
Mark de Wever [Tue, 8 Feb 2022 17:32:08 +0000 (18:32 +0100)]
[libc++] Enables put_long_double test for glibc.

Glibc adds a +-sign for NaN-values when showpos fmtflags are set.

[tab:ios.fmtflags]
showpos generates a + sign in non-negative generated numeric output

Since NaNs aren't negative this behaviour seems correct. Enable the test
for glibc and add ifdefs to make sure the existing tests still pass.

This was noticed while working on D118971.

Reviewed By: #libc, ldionne

Differential Revision: https://reviews.llvm.org/D119260

2 years ago[libc++][format][nfc] Header cleanup.
Mark de Wever [Thu, 3 Feb 2022 18:02:36 +0000 (19:02 +0100)]
[libc++][format][nfc] Header cleanup.

Remove the unneeded macro protection, forward declarations, and
includes.

Reviewed By: #libc, Quuxplusone, ldionne, philnik

Differential Revision: https://reviews.llvm.org/D118925

2 years ago[RISCV] Pre-process integer ISD::SPLAT_VECTOR to RISCISD::VMV_V_X_VL before isel.
Craig Topper [Wed, 9 Feb 2022 16:05:42 +0000 (08:05 -0800)]
[RISCV] Pre-process integer ISD::SPLAT_VECTOR to RISCISD::VMV_V_X_VL before isel.

This allows us to remove some isel patterns that exist for both
operations. Saving nearly 3000 bytes from the isel table.

Reviewed By: frasercrmck

Differential Revision: https://reviews.llvm.org/D119197

2 years ago[libc++][nfc] Use TEST_HAS_NO_WIDE_CHARACTERS.
Mark de Wever [Wed, 2 Feb 2022 18:28:04 +0000 (19:28 +0100)]
[libc++][nfc] Use TEST_HAS_NO_WIDE_CHARACTERS.

This avoids using an libc++ internal macro in our tests.

Reviewed By: #libc, Quuxplusone

Differential Revision: https://reviews.llvm.org/D119263

2 years ago[VP] llvm.vp.fneg intrinsic and LangRef
Craig Topper [Wed, 9 Feb 2022 15:54:07 +0000 (07:54 -0800)]
[VP] llvm.vp.fneg intrinsic and LangRef

Reviewed By: frasercrmck

Differential Revision: https://reviews.llvm.org/D119262

2 years ago[nfc][mlgo] Make error message macro
Mircea Trofin [Wed, 9 Feb 2022 15:25:37 +0000 (07:25 -0800)]
[nfc][mlgo] Make error message macro

`llvm_unreachable` is itself a macro, so using a constant for its
message renders that constant unused in non-debug builds.

2 years ago[lldb] Stabilize threaded windows lldb-server tests
Pavel Labath [Wed, 9 Feb 2022 15:14:15 +0000 (16:14 +0100)]
[lldb] Stabilize threaded windows lldb-server tests

The tests enabled in 9e699595 are not passing reliably -- sometimes they
report seeing fewer threads than expected.

Based on my (limited) research, this is not a lldb bug, but simply a
consequence of the operating system reporting their presence
asynchronously -- they're reported when they are scheduled to run (or
something similar), and not at the time of the CreateThread call.

To fix this, I add some additional synchronization to the test inferior,
which makes sure that the created thread is alive before continuing to
do other things.

2 years ago[lldb] Simplify SBCommandInterpreter::SourceInitFileInHomeDirectory
Pavel Labath [Wed, 9 Feb 2022 13:56:45 +0000 (14:56 +0100)]
[lldb] Simplify SBCommandInterpreter::SourceInitFileInHomeDirectory

just have it delegate to the new overload.

2 years ago[SLP][NFC]Fix comment, NFC.
Alexey Bataev [Wed, 9 Feb 2022 15:00:43 +0000 (07:00 -0800)]
[SLP][NFC]Fix comment, NFC.

2 years ago[OpenCL][Docs] Update OpenCL 3.0 status info
Sven van Haastregt [Wed, 9 Feb 2022 15:12:49 +0000 (15:12 +0000)]
[OpenCL][Docs] Update OpenCL 3.0 status info

Update the table to reflect recently committed work.

2 years ago[NFC] Fixing test requirements I broke
Chris Bieneman [Wed, 9 Feb 2022 15:11:17 +0000 (09:11 -0600)]
[NFC] Fixing test requirements I broke

I broke these in 7a0cbe11fb26, thanks @ikudrin for catching it!

2 years ago[libc++] Try to fix the quoting of -isystem on Windows bots
Louis Dionne [Wed, 9 Feb 2022 15:07:27 +0000 (10:07 -0500)]
[libc++] Try to fix the quoting of -isystem on Windows bots

2 years ago[X86] Refresh funnel/rotate AVX512 VBMI tests
Simon Pilgrim [Wed, 9 Feb 2022 15:04:27 +0000 (15:04 +0000)]
[X86] Refresh funnel/rotate AVX512 VBMI tests

Assume that if VBMI2 is enabled then VBMI is as well - correctly shows the fallback and widening code that real world targets will actually see

2 years ago[AArch64][SVE] NFC: Add test file for predicate vector reductions.
Sander de Smalen [Mon, 24 Jan 2022 17:45:39 +0000 (17:45 +0000)]
[AArch64][SVE] NFC: Add test file for predicate vector reductions.

This adds some tests for vector reductions which can and should
be implemented with ptest as opposed to promoted ANDV/ORV reduction.

2 years ago[AArch64] NFC: Autogen check lines for sve-setcc.ll
Sander de Smalen [Fri, 4 Feb 2022 08:46:26 +0000 (08:46 +0000)]
[AArch64] NFC: Autogen check lines for sve-setcc.ll

2 years ago[DAGCombiner] Fold `ty1 extract_vector(ty2 splat(V)) -> ty1 splat(V)`
Sander de Smalen [Wed, 9 Feb 2022 09:54:53 +0000 (09:54 +0000)]
[DAGCombiner] Fold `ty1 extract_vector(ty2 splat(V)) -> ty1 splat(V)`

This seems like an obvious fold, which leads to a few improvements.

Reviewed By: david-arm

Differential Revision: https://reviews.llvm.org/D118920

2 years ago[mlir][nfc] Expose linalg tiling helpers.
Alexander Belyaev [Wed, 9 Feb 2022 14:20:04 +0000 (15:20 +0100)]
[mlir][nfc] Expose linalg tiling helpers.

Differential Revision: https://reviews.llvm.org/D119330

2 years ago[flang] Add type conversion for !fir.box<none>
Valentin Clement [Wed, 9 Feb 2022 14:11:55 +0000 (15:11 +0100)]
[flang] Add type conversion for !fir.box<none>

`none` is used in `fir.box` type to specify a polymorphic type.
This patch add the conversion from `!fir.box<none>` to LLVM.

This patch is part of the upstreaming effort from fir-dev branch.

Reviewed By: awarzynski

Differential Revision: https://reviews.llvm.org/D119325

Co-authored-by: Eric Schweitz <eschweitz@nvidia.com>
Co-authored-by: Jean Perier <jperier@nvidia.com>
2 years agoUse functions with prototypes when appropriate; NFC
Aaron Ballman [Wed, 9 Feb 2022 14:10:10 +0000 (09:10 -0500)]
Use functions with prototypes when appropriate; NFC

A significant number of our tests in C accidentally use functions
without prototypes. This patch converts the function signatures to have
a prototype for the situations where the test is not specific to K&R C
declarations. e.g.,

  void func();

becomes

  void func(void);

This is the fifth batch of tests being updated (there are a significant
number of other tests left to be updated).

Note, the behavior of -ast-print is broken. It prints functions with a
prototype (void) as if they have no prototype () in C. Some tests need
to disable strict prototype checking when recompiling the results of an
-ast-print invocation.

2 years ago[OpenMP][libomp] Replace accidental VLA with KMP_ALLOCA
Jonathan Peyton [Wed, 9 Feb 2022 14:04:47 +0000 (08:04 -0600)]
[OpenMP][libomp] Replace accidental VLA with KMP_ALLOCA

MSVC does not support variable length arrays. Replace with KMP_ALLOCA
which is already used in the same file for stack-allocated variables.

2 years ago[runtimes] Remove support for standalone builds
Louis Dionne [Tue, 8 Feb 2022 16:38:29 +0000 (11:38 -0500)]
[runtimes] Remove support for standalone builds

Standalone build have been deprecated for some time now, so this
commit removes support for those builds entirely from libc++, libc++abi
and libunwind.

This, along with the removal of other legacy ways to build, will allow
for major build system simplifications.

Differential Revision: https://reviews.llvm.org/D119255

2 years ago[libc++] Avoid -Wmacro-redefined in libc++ headers in the legacy test config
Louis Dionne [Wed, 9 Feb 2022 13:38:06 +0000 (08:38 -0500)]
[libc++] Avoid -Wmacro-redefined in libc++ headers in the legacy test config

2 years agoAdd missing MC includes in bolt/
serge-sans-paille [Wed, 9 Feb 2022 13:26:30 +0000 (08:26 -0500)]
Add missing MC includes in bolt/

Changes needed after ef736a1c39f27ef4 that removes some implicit
dependencies from MrCV headers.

2 years ago[Bitcode] Fix size check for DIImportedEntity record
Nikita Popov [Wed, 9 Feb 2022 13:19:24 +0000 (14:19 +0100)]
[Bitcode] Fix size check for DIImportedEntity record

This was using && instead of ||.

2 years ago[Bitcode] Check minimum size of constant GEP record
Nikita Popov [Wed, 9 Feb 2022 13:14:04 +0000 (14:14 +0100)]
[Bitcode] Check minimum size of constant GEP record

Checking this early, because we may end up reading up to two
records before the operands.

2 years ago[LLDB] Port toolchain-msvc.test for Arm/AArch4 Windows
Muhammad Omair Javaid [Wed, 9 Feb 2022 12:35:33 +0000 (17:35 +0500)]
[LLDB] Port toolchain-msvc.test for Arm/AArch4 Windows

This patch updates toolchain-msvc.test to cater for Arm64 windows platform.

Reviewed By: labath

Differential Revision: https://reviews.llvm.org/D117676

2 years ago[X86] TCRETURNmi fix for 32bit platform
Tong Zhang [Wed, 9 Feb 2022 11:51:57 +0000 (19:51 +0800)]
[X86] TCRETURNmi fix for 32bit platform

This fix is similar to 3cf3ffce240e("Fix the TCRETURNmi64 bug differently.")

after allocating register for index+base, we will only have one register left

This bug affects linux kernel compilation for x86 target. Error happens when compiling kmod_si476x_core.

clang complains:
    error: ran out of registers during register allocation

The full command is:

clang -Wp,-MMD,drivers/mfd/.si476x-cmd.o.d  -nostdinc -isystem /opt/toolchain/main/lib/clang/14.0.0/include -I./arch/x86/include -I./arch/x86/include/generated  -I./include -I./arch/x86/include/uapi -I./arch/x86/include/generated/uapi -I./include/uapi -I./include/generated/uapi -include ./include/linux/compiler-version.h -include ./include/linux/kconfig.h -include ./include/linux/compiler_types.h -D__KERNEL__ -Qunused-arguments -fmacro-prefix-map=./= -Wall -Wundef -Werror=strict-prototypes -Wno-trigraphs -fno-strict-aliasing -fno-common -fshort-wchar -fno-PIE -Werror=implicit-function-declaration -Werror=implicit-int -Werror=return-type -Wno-format-security -std=gnu89 -no-integrated-as --prefix=/usr/bin/ -Werror=unknown-warning-option -mno-sse -mno-mmx -mno-sse2 -mno-3dnow -mno-avx -fcf-protection=none -m32 -msoft-float -mregparm=3 -freg-struct-return -fno-pic -mstack-alignment=4 -march=atom -mtune=atom -mtune=generic -Wa,-mtune=generic32 -ffreestanding -Wno-sign-compare -fno-asynchronous-unwind-tables -fno-delete-null-pointer-checks -Wno-frame-address -Wno-address-of-packed-member -O2 -Wframe-larger-than=1024 -fno-stack-protector -Wno-format-invalid-specifier -Wno-gnu -mno-global-merge -Wno-unused-but-set-variable -Wno-unused-const-variable -fomit-frame-pointer -ftrivial-auto-var-init=pattern -fno-stack-clash-protection -falign-functions=32 -Wdeclaration-after-statement -Wvla -Wno-pointer-sign -Wno-array-bounds -fno-strict-overflow -fno-stack-check -Werror=date-time -Werror=incompatible-pointer-types -Wno-initializer-overrides -Wno-format -Wno-sign-compare -Wno-format-zero-length -Wno-pointer-to-enum-cast -Wno-tautological-constant-out-of-range-compare     -DKBUILD_MODFILE='"drivers/mfd/si476x-core"' -DKBUILD_BASENAME='"si476x_cmd"' -DKBUILD_MODNAME='"si476x_core"' -D__KBUILD_MODNAME=kmod_si476x_core -c -o drivers/mfd/si476x-cmd.o drivers/mfd/si476x-cmd.c

-------------

LLVM cannot compile the following code for x86 32bit target, the reason is tail call(TCRETURNmi) is using 2 registers for index+base and we want to use more than one registers for passing function args and that is impossible.
This fix is similar to 3cf3ffce240e("Fix the TCRETURNmi64 bug differently.").
We will only use tail call when it is using <=1 registers for passing args.

```
struct BIG_PARM {
int ver;
};

static struct {
int (*foo)  (struct BIG_PARM* a, void *b);
int (*bar)  (struct BIG_PARM* a);
int (*zoo0) (void);
int (*zoo1) (void);
int (*zoo2) (void);
int (*zoo3) (void);
int (*zoo4) (void);
} vtable[] = {
[0] = {
.foo = (int (*)(struct BIG_PARM* a, void *b))0xdeadbeef,
},
};

int something(struct BIG_PARM *a, void* b) {
return vtable[a->ver].foo(a,b);
}

```

```
$ clang -std=gnu89 -m32 -mregparm=3 -mtune=generic -fno-strict-overflow -O2 -c t0.c -o t0.c.o
error: ran out of registers during register allocation
1 error generated.
```

Reviewed By: pengfei

Differential Revision: https://reviews.llvm.org/D118312

2 years agoRevert "X86: gate all vmovsh instructions on FP16 support."
Tim Northover [Wed, 9 Feb 2022 12:33:17 +0000 (12:33 +0000)]
Revert "X86: gate all vmovsh instructions on FP16 support."

This reverts commit 3fc40b6e6628cd3206d5aaf2bfb864add7d74fe1.

It was pushed unintentionally.

2 years agoAArch64: do not use xzr for ldxp -> stxp dataflow.
Tim Northover [Wed, 9 Feb 2022 11:42:36 +0000 (11:42 +0000)]
AArch64: do not use xzr for ldxp -> stxp dataflow.

If the result of a cmpxchg is unused, regalloc chooses `xzr` for the defs of
CMP_SWAP_128*. However, on the failure path this gets expanded to a LDXP ->
STXP to store the original value (to ensure no tearing occurred). This
unintentionally nulls out half of the value.

So instead use GPR64common for these defs, so regalloc has to choose a real
one.

2 years agoX86: gate all vmovsh instructions on FP16 support.
Tim Northover [Mon, 7 Feb 2022 11:19:59 +0000 (11:19 +0000)]
X86: gate all vmovsh instructions on FP16 support.

Previously the `let Predicates = ...` line only applied to the rr version, and
so VMOVSH was being emitted whenever HasAVX512 (the default) applied. This is
not right.

2 years ago[AMDGPU] Remove irrelevant comments on V_BFE_I32 instructions
Jay Foad [Wed, 9 Feb 2022 12:02:06 +0000 (12:02 +0000)]
[AMDGPU] Remove irrelevant comments on V_BFE_I32 instructions

These comments explain the encoding of the immediate operand of S_BFE_*
which is not relevant for V_BFE_I32.

2 years ago[mlir][bufferize] OpOperands can have multiple aliasing OpResults
Matthias Springer [Wed, 9 Feb 2022 11:55:25 +0000 (20:55 +0900)]
[mlir][bufferize] OpOperands can have multiple aliasing OpResults

This makes getAliasingOpResult symmetric to getAliasingOpOperand. The previous implementation was confusing for users and implemented in such a way only because there are currently no bufferizable ops that have multiple aliasing OpResults.

Differential Revision: https://reviews.llvm.org/D119259

2 years ago[mlir][linalg][bufferize] Print results of FuncOp read/write analysis
Matthias Springer [Wed, 9 Feb 2022 11:44:02 +0000 (20:44 +0900)]
[mlir][linalg][bufferize] Print results of FuncOp read/write analysis

Print more information with test-analysis-only.

Differential Revision: https://reviews.llvm.org/D119118

2 years ago[cross-project-tests] Make GDB version string parsing more robust
OCHyams [Wed, 9 Feb 2022 11:32:29 +0000 (11:32 +0000)]
[cross-project-tests] Make GDB version string parsing more robust

Follow up to D118468 (5257efdc5b30212b62a9d68857dc8e66d0e1a863).

When built from source, gdb's version string looks like this:

    GNU gdb (GDB) 9.2
    ...

But for installed versions it looks different. E.g.

    GNU gdb (Ubuntu 9.2-0ubuntu1~20.04.1) 9.2
    ...

Use a regex rather than str.parition in the version string parsing in order to
handle this case too.

2 years ago[lldb] Restore original meaning to test_qThreadStopInfo_only_reports_one_thread_stop_...
Pavel Labath [Wed, 9 Feb 2022 11:30:57 +0000 (12:30 +0100)]
[lldb] Restore original meaning to test_qThreadStopInfo_only_reports_one_thread_stop_reason_during_interrupt

D119167 changed the meaning of that test by removing the use of the
interrupt packet. I did not notice this because the interrupting
happened in a shared utility function.

This patch restores the original meaning of the test, but (almost)
avoids sleeps by using process stdout to synchronize. Sadly, this means
the test stays disabled on windows, as it does not implement output
forwarding.

2 years ago[clang] Fix sphinx "start-string without end-string" warning
Simon Pilgrim [Wed, 9 Feb 2022 11:31:24 +0000 (11:31 +0000)]
[clang] Fix sphinx "start-string without end-string" warning

This looks to introduced in D110869

2 years ago[clang] Remove duplicate doc index for ClangLinkerWrapper
Simon Pilgrim [Wed, 9 Feb 2022 11:21:16 +0000 (11:21 +0000)]
[clang] Remove duplicate doc index for ClangLinkerWrapper

This looks to be a merge error from D118815

2 years ago[RISCV][NFC] Remove useless code
Lian Wang [Wed, 9 Feb 2022 11:16:35 +0000 (19:16 +0800)]
[RISCV][NFC] Remove useless code

Reviewed By: craig.topper, asb

Differential Revision: https://reviews.llvm.org/D119317

2 years ago[LV] Pass step to emitTransformedIndex (NFC).
Florian Hahn [Wed, 9 Feb 2022 11:12:44 +0000 (11:12 +0000)]
[LV] Pass step to emitTransformedIndex (NFC).

Move out the induction step creation from emitTransformedIndex to the
callers. In some places (e.g. widenIntOrFpInduction) the step is already
created. Passing the step in ensures the steps are kept in sync.

2 years ago[lldb] Adjust windows xfails for D119167
Pavel Labath [Wed, 9 Feb 2022 11:01:30 +0000 (12:01 +0100)]
[lldb] Adjust windows xfails for D119167

A couple of additional tests pass with that patch. One new test fails
(because it's not testing a slightly different thing). I'll update it
later to restore the original meaning (I don't want to revert as the net
effect is still very positive), but for now this gets the bot green.

2 years ago[cross-project-tests] XFAIL llgdb-tests when gdb can't read clang's DWARF
OCHyams [Wed, 9 Feb 2022 10:47:07 +0000 (10:47 +0000)]
[cross-project-tests] XFAIL llgdb-tests when gdb can't read clang's DWARF

Tests in the `cross-project-tests/debuginfo-tests/llgdb-tests` directory run
gdb on non-darwin platforms. gdb versions less than 10.1 cannot parse the DWARF
v5 emitted by clang, and DWARF v5 is now the default, so these tests fail on
Linux with gdb versions less than 10.1. This patch lets us XFAIL the tests
under these conditions.

Add `gdb-clang-incompatibility` to the `available_features` in
`cross-project-tests/lit.cfg.py` when clang's default DWARF version is 5 or
greater and the gdb (if found) version is less than 10.1.

Discourse discussion:
https://llvm.discourse.group/t/gdb-10-1-cant-read-clangs-dwarf-v5/6035

Reviewed By: jmorse

Differential Revision: https://reviews.llvm.org/D118468

2 years ago[clang] Fix the tooling build after D119130
Kirill Bobyrev [Wed, 9 Feb 2022 10:52:03 +0000 (11:52 +0100)]
[clang] Fix the tooling build after D119130

New StandardLibrary.cpp depends on Clang AST, add the dependency to
CMakeLists.txt

Broken builbot: https://lab.llvm.org/buildbot/#/builders/57/builds/14892

2 years ago[OpenCL] Fix atomic_fetch_add/sub with half type
Sven van Haastregt [Wed, 9 Feb 2022 10:47:45 +0000 (10:47 +0000)]
[OpenCL] Fix atomic_fetch_add/sub with half type

An error in the tablegen description affects the declarations
provided by `-fdeclare-opencl-builtins` for `atomic_fetch_add` and
`atomic_fetch_sub`.

The atomic argument should be an atomic_half, not an atomic_float.

2 years ago[bazel] Sync llvm-config.h.cmake after a1862d78eb45
Benjamin Kramer [Wed, 9 Feb 2022 10:27:48 +0000 (11:27 +0100)]
[bazel] Sync llvm-config.h.cmake after a1862d78eb45

2 years ago[bazel] Port 46a6f5a
Benjamin Kramer [Wed, 9 Feb 2022 10:25:33 +0000 (11:25 +0100)]
[bazel] Port 46a6f5a

2 years ago[flang][NFC] Reorder conversions
Valentin Clement [Wed, 9 Feb 2022 10:21:27 +0000 (02:21 -0800)]
[flang][NFC] Reorder conversions

During upstreaming the alphabetical order was not respected.

2 years ago[NFC] test commit
Sheng [Wed, 9 Feb 2022 10:22:01 +0000 (18:22 +0800)]
[NFC] test commit

Empty test commit, check commit access

2 years ago[pseudo] Implement LRGraph
Haojian Wu [Mon, 7 Feb 2022 14:21:45 +0000 (15:21 +0100)]
[pseudo] Implement LRGraph

LRGraph is the key component of the clang pseudo parser, it is a
deterministic handle-finding finite-state machine, which is used to
generated the LR parsing table.

Separate from https://reviews.llvm.org/D118196.

Differential Revision: https://reviews.llvm.org/D119172

2 years ago[mlir][linalg][bufferize][NFC] Allow passing custom BufferizationOptions to pass
Matthias Springer [Wed, 9 Feb 2022 10:14:42 +0000 (19:14 +0900)]
[mlir][linalg][bufferize][NFC] Allow passing custom BufferizationOptions to pass

Differential Revision: https://reviews.llvm.org/D118891

2 years agoCleanup LLVMMC headers
serge-sans-paille [Tue, 8 Feb 2022 14:32:21 +0000 (15:32 +0100)]
Cleanup LLVMMC headers

There's a few relevant forward declarations in there that may require downstream
adding explicit includes:

llvm/MC/MCContext.h no longer includes llvm/BinaryFormat/ELF.h, llvm/MC/MCSubtargetInfo.h, llvm/MC/MCTargetOptions.h
llvm/MC/MCObjectStreamer.h no longer include llvm/MC/MCAssembler.h
llvm/MC/MCAssembler.h no longer includes llvm/MC/MCFixup.h, llvm/MC/MCFragment.h

Counting preprocessed lines required to rebuild llvm-project on my setup:
before: 1052436830
after:  1049293745

Which is significant and backs up the change in addition to the usual benefits of
decreasing coupling between headers and compilation units.

Discourse thread: https://discourse.llvm.org/t/include-what-you-use-include-cleanup
Differential Revision: https://reviews.llvm.org/D119244

2 years agoPrepare for LLVMMC headers cleanup
serge-sans-paille [Wed, 9 Feb 2022 08:53:00 +0000 (09:53 +0100)]
Prepare for LLVMMC headers cleanup

Be more explicit about which headers should be included in MC files generated by
tblgen.

See also: https://reviews.llvm.org/D119244

2 years ago[X86] Specify Undef for the registers we xor
Bill Wendling [Wed, 9 Feb 2022 10:05:13 +0000 (02:05 -0800)]
[X86] Specify Undef for the registers we xor

Fixes expensive check failures from D110869.

2 years ago[llvm] Add IWYU pragmas to GTest
Kirill Bobyrev [Wed, 9 Feb 2022 10:05:58 +0000 (11:05 +0100)]
[llvm] Add IWYU pragmas to GTest

This will allow using tools like Include-What-You-Use and clangd
IncludeCleaner. The tools will correctly identify the public headers
responsible for importing symbols in the testing code.

This is a backport of https://github.com/google/googletest/commit/100f6fbf5f81a82d163c1e29735e8a2936eacd4f

The only untouched file from that patch is
googletest/include/gtest/gtest_prod.h because the change is unrelated.

Reviewed By: sammccall

Differential Revision: https://reviews.llvm.org/D119320

2 years ago[clangd] NFC: Move stdlib headers handling to Clang
Kirill Bobyrev [Wed, 9 Feb 2022 10:04:58 +0000 (11:04 +0100)]
[clangd] NFC: Move stdlib headers handling to Clang

This will allow moving the IncludeCleaner library essentials to Clang
and decoupling them from the majority of clangd.

The patch itself just moves the code, it doesn't change existing
functionality.

Reviewed By: sammccall

Differential Revision: https://reviews.llvm.org/D119130

2 years ago[lldb/test] Remove sleeps from some lldb-server tests
Pavel Labath [Mon, 7 Feb 2022 19:04:42 +0000 (20:04 +0100)]
[lldb/test] Remove sleeps from some lldb-server tests

Instead of using sleeps, have the inferior notify us (via a trap opcode) that
the requested number of threads have been created.

This allows us to get rid of some fairly dodgy test utility code --
wait_for_thread_count seemed like it was waiting for the threads to
appear, but it never actually let the inferior run, so it only succeeded
if the threads were already started when the function was called. Since
the function was called after a fairly small delay (1s, usually), this
is probably the reason why the tests were failing on some bots.

Differential Revision: https://reviews.llvm.org/D119167

2 years ago[mlir][bufferize][NFC] Make PostAnalysisSteps a function
Matthias Springer [Wed, 9 Feb 2022 09:55:28 +0000 (18:55 +0900)]
[mlir][bufferize][NFC] Make PostAnalysisSteps a function

They used to be classes with a virtual `run` function. This was inconvenient because post analysis steps are stored in BufferizationOptions. Because of this design choice, BufferizationOptions were not copyable.

Differential Revision: https://reviews.llvm.org/D119258

2 years ago[flang] Add runtime interface for GET_COMMAND
Diana Picus [Wed, 26 Jan 2022 09:08:49 +0000 (09:08 +0000)]
[flang] Add runtime interface for GET_COMMAND

Use a single entry point with several optional parameters.

Differential Revision: https://reviews.llvm.org/D118776

2 years ago[Flang] Add support for lowering the goto statement
Kiran Chandramohan [Tue, 8 Feb 2022 23:01:39 +0000 (23:01 +0000)]
[Flang] Add support for lowering the goto statement

This patch adds support for lowering the Fortran goto statement from
parse-tree to MLIR. The goto statement in Fortran is a form of
unstructured control flow. The statement transfers control to the
code starting at the label specified in the statement. This can be
faithfully represented in MLIR by a branch instruction.

To assist the lowering of code with unstructured control flow, blocks
are created in advance and associated with the relevant pre-fir tree
evaluations.

This is part of the upstreaming effort from the fir-dev branch in [1].

[1] https://github.com/flang-compiler/f18-llvm-project

Reviewed By: clementval, vdonaldson, schweitz, awarzynski

Differential Revision: https://reviews.llvm.org/D118983

Co-authored-by: V Donaldson <vdonaldson@nvidia.com>
Co-authored-by: Jean Perier <jperier@nvidia.com>
Co-authored-by: Eric Schweitz <eschweitz@nvidia.com>
2 years agoSet LLVM_FORCE_USE_OLD_TOOLCHAIN to disable VS2019 checks
Muhammad Omair Javaid [Wed, 9 Feb 2022 09:26:15 +0000 (14:26 +0500)]
Set LLVM_FORCE_USE_OLD_TOOLCHAIN to disable VS2019 checks

VS2019 version 1920 in now the default and get tested in
llvm/include/llvm/Support/Compiler.h. This patch propagates
LLVM_FORCE_USE_OLD_TOOLCHAIN macro to disable testing for VS2019.

Differential Revision: https://reviews.llvm.org/D114639

2 years agoProfileDataTests: fix BUILD_SHARED_LIBS build
Roman Lebedev [Wed, 9 Feb 2022 09:27:01 +0000 (12:27 +0300)]
ProfileDataTests: fix BUILD_SHARED_LIBS build

2 years ago[analyzer][docs][NFC] Fix some broken links and some cosmetic changes
Balazs Benics [Wed, 9 Feb 2022 09:22:28 +0000 (10:22 +0100)]
[analyzer][docs][NFC] Fix some broken links and some cosmetic changes

- We should report bugs to the GitHub Issues
- We should advocate using Discourse instead of the superseded cfe-dev
  mailing list.

There are a couple of other cosmetic changes such as preferring `using`
instead of `typedef` and quoting the `-analyze-function` parameter's
payload for escaping the braces.

Thanks @loic-joly-sonarsource for reporting this on Discord!

Reviewed By: NoQ, Szelethus

Differential Revision: https://reviews.llvm.org/D119245

2 years agoReapply [sanitizers] Avoid macro clash in SignalContext::WriteFlag (NFC)
Nikita Popov [Wed, 2 Feb 2022 14:06:01 +0000 (15:06 +0100)]
Reapply [sanitizers] Avoid macro clash in SignalContext::WriteFlag (NFC)

D116208 may cause a macro clash on older versions of linux, where
fs.h defines a READ macro. This is resolved by switching to a more
typical casing style for non-macro symbols.

Reapplying with changes to the symbol names in various platform
specific code, which I missed previously.

Differential Revision: https://reviews.llvm.org/D118783

2 years ago[libc++] Add papers from February 2022 plenary meeting
Nikolas Klauser [Tue, 8 Feb 2022 23:07:14 +0000 (00:07 +0100)]
[libc++] Add papers from February 2022 plenary meeting

Reviewed By: ldionne, Quuxplusone, #libc

Spies: arichardson, libcxx-commits

Differential Revision: https://reviews.llvm.org/D119189

2 years ago[flang] Add missing CFI case for REAL and COMPLEX
Valentin Clement [Wed, 9 Feb 2022 09:07:21 +0000 (10:07 +0100)]
[flang] Add missing CFI case for REAL and COMPLEX

ISO_Fortran_binding.h was updated with missing entries for CFI
types for REAL and COMPLEX kinds 2,3,10,16. This patch updates TypeCode.h
to use these new types.

This patch is part of the upstreaming effort from fir-dev branch.

Reviewed By: jeanPerier

Differential Revision: https://reviews.llvm.org/D119283

2 years agoRevert "[sanitizers] Avoid macro clash in SignalContext::WriteFlag (NFC)"
Nikita Popov [Wed, 9 Feb 2022 09:06:56 +0000 (10:06 +0100)]
Revert "[sanitizers] Avoid macro clash in SignalContext::WriteFlag (NFC)"

This reverts commit fda29264f360820859587acdfb0ad9392c944bd6.

This breaks the sanitizer build on windows, will reapply with
additional changes.

2 years ago[NFC] Fix SSE3 intrinsics test for PowerPC
Qiu Chaofan [Wed, 9 Feb 2022 08:46:51 +0000 (16:46 +0800)]
[NFC] Fix SSE3 intrinsics test for PowerPC

Previous test in ppc-pmmintrin.c did not check IR of intrinsic function
definition. Add them and simplify.

These tests shouldn't be auto-generated, because we don't want to check
wrapper functions.

2 years ago[clang-cl] Bump default -fms-compatibility-version to 19.14
Muhammad Omair Javaid [Tue, 8 Feb 2022 15:25:41 +0000 (20:25 +0500)]
[clang-cl] Bump default -fms-compatibility-version to 19.14

clang-cl MSVC required version is 19.20 now. Update the default
-fms-compatibility-version to 19.14.

Differential Revision: https://reviews.llvm.org/D114639