platform/upstream/llvm.git
21 months ago[NFC][AArch64]Precommit test for D135102
Mingming Liu [Thu, 3 Nov 2022 17:45:18 +0000 (10:45 -0700)]
[NFC][AArch64]Precommit test for D135102

- Run auto updater for 'trunc-to-tbl.ll' and 'build-pair-isel.ll'
- Add the motivating test of D135102 to 'bitfield-insert.ll'

Reviewed By: fhahn

Differential Revision: https://reviews.llvm.org/D137296

21 months agoHave GetSupportedArchitectures report all supported arches
Jason Molenda [Thu, 3 Nov 2022 17:46:48 +0000 (10:46 -0700)]
Have GetSupportedArchitectures report all supported arches

PlatformDarwinKernel::GetSupportedArchitectures returns a list
of architectures that are possible for this platform; it was using
a compile-time check for the debug host to decide the list of arches
that were valid.  This was copied from a codepath doing native process
debugging, and was clearly wrong for kernel debugging, but it had
not happened to cause problems so it went unnoticed for a long time.

Small NFC change to the logging messages of Target::SetArchitecture
to make them a little more explicit about how the architecture is
being modified/replaced.

Differential Revision: https://reviews.llvm.org/D137301
rdar://101690111

21 months ago[mlir][bufferize] Add filterFn option to BufferResultsToOutParams
Emilio Cota [Thu, 3 Nov 2022 14:47:25 +0000 (10:47 -0400)]
[mlir][bufferize] Add filterFn option to BufferResultsToOutParams

This allows users to restrict the transformation to a
subset of the functions in a module.

For example, a user might want to apply the transformation to
a module's entry point, but not to the calls in the module
because those calls might refer to external C functions outside
of their control.

Reviewed By: springerm, nicolasvasilache
Differential Revision: https://reviews.llvm.org/D137264

21 months agoFixing flang's definition of MLIR_MAIN_SRC_DIR
Renaud-K [Wed, 2 Nov 2022 23:02:04 +0000 (16:02 -0700)]
Fixing flang's definition of MLIR_MAIN_SRC_DIR

21 months agoGlobalISel: Fix artifact combine value finder look through copy
Petar Avramovic [Thu, 3 Nov 2022 16:41:41 +0000 (17:41 +0100)]
GlobalISel: Fix artifact combine value finder look through copy

Search for COPY source in instruction we get from look through (not copy dst).

Differential Revision: https://reviews.llvm.org/D137273

21 months ago[Flang] Run clang-format on all flang files
Peter Steinfeld [Wed, 2 Nov 2022 21:31:14 +0000 (14:31 -0700)]
[Flang] Run clang-format on all flang files

This will make it easier for me to do reviews.

Differential Revision: https://reviews.llvm.org/D137291

21 months ago[mlir][llvmir] Add support for llvm.signext/zeroext function attributes.
Slava Zakharin [Thu, 3 Nov 2022 16:07:44 +0000 (09:07 -0700)]
[mlir][llvmir] Add support for llvm.signext/zeroext function attributes.

This change-set adds basic support for llvm.signext and llvm.zeroext
attributes, and makes sure that the attributes are translated to LLVM IR
when attached to arguments.

This is needed for https://github.com/llvm/llvm-project/issues/58579

Differential Revision: https://reviews.llvm.org/D137048

21 months ago[RISCV][CodeGen] Account for LMUL for Vector Integer Arithmetic Instructions
Michael Maitland [Wed, 26 Oct 2022 01:43:00 +0000 (18:43 -0700)]
[RISCV][CodeGen] Account for LMUL for Vector Integer Arithmetic Instructions

It is likley that subtargets act differently for a vector integer arithmetic instruction based on the LMUL.
This patch creates seperate SchedRead, SchedWrite, WriteRes, ReadAdvance for each relevant LMUL. It also
introduces the concept of an "UpperBound LMUL" which allows us to describe how an instruction should behave
when the LMUL is unknown. All base instructions use the UpperBound resources because they are not tied to
a specific LMUL. This gives subtargetes the flexibility to describe their own upper bounds on each vector
instruction.

I have a series of patches for the rest of the vector instruction set ready to go, but I would like to first
get feedback on the first one of the series (this one).

Differential Revision: https://reviews.llvm.org/D136730

21 months ago[AArch64] Alter arm_fp16.h to be target-based, not preprocessor based.
David Green [Thu, 3 Nov 2022 16:06:35 +0000 (16:06 +0000)]
[AArch64] Alter arm_fp16.h to be target-based, not preprocessor based.

As the other recent patches, this alters the arm_fp16 intrinsics to be
target based, not preprocessor based. Apparently arm_fp16.h is AArch64
only under clang, making this mostly trivial with the TargetGuard
infrastructure.

Differential Revision: https://reviews.llvm.org/D137256

21 months ago[mlir][llvmir] Convert attributes for functions without bodies.
Slava Zakharin [Wed, 2 Nov 2022 19:47:50 +0000 (12:47 -0700)]
[mlir][llvmir] Convert attributes for functions without bodies.

So far the function argument attributes were only translated
for functions with bodies. This change makes sure that this
happens for functions without bodies (declarations) as well.

This is needed for https://github.com/llvm/llvm-project/issues/58579

Reviewed By: ftynse

Differential Revision: https://reviews.llvm.org/D137047

21 months ago[mlir][python] Include anchor op in PassManager constructor
rkayaith [Thu, 20 Oct 2022 05:04:34 +0000 (01:04 -0400)]
[mlir][python] Include anchor op in PassManager constructor

This adds an extra argument for specifying the pass manager's anchor op,
with a default of `any`. Previously the anchor was always defaulted to
`builtin.module`.

Reviewed By: mehdi_amini

Differential Revision: https://reviews.llvm.org/D136406

21 months ago[mlir][python] Include anchor op in PassManager.parse
rkayaith [Thu, 20 Oct 2022 03:36:15 +0000 (23:36 -0400)]
[mlir][python] Include anchor op in PassManager.parse

The pipeline string must now include the pass manager's anchor op. This
makes the parse API properly roundtrip the printed form of a pass
manager.

Reviewed By: mehdi_amini

Differential Revision: https://reviews.llvm.org/D136405

21 months ago[mlir][CAPI] Include anchor op in mlirParsePassPipeline
rkayaith [Thu, 20 Oct 2022 02:37:12 +0000 (22:37 -0400)]
[mlir][CAPI] Include anchor op in mlirParsePassPipeline

The pipeline string must now include the pass manager's anchor op. This
makes the parse API properly roundtrip the printed form of a pass
manager. Since this is already an API break, I also added an extra
callback argument which is used for reporting errors.

The old functionality of appending to an existing pass manager is
available through `mlirOpPassManagerAddPipeline`.

Reviewed By: mehdi_amini, ftynse

Differential Revision: https://reviews.llvm.org/D136403

21 months ago[mlir] Include anchor op in reproducer pipeline string
rkayaith [Wed, 21 Sep 2022 00:20:44 +0000 (20:20 -0400)]
[mlir] Include anchor op in reproducer pipeline string

Including the anchor op ensures that all pass manager settings are fully
specified, and makes the string consistent with the printed form.

Depends on D134622

Reviewed By: rriddle

Differential Revision: https://reviews.llvm.org/D134623

21 months ago[OpenMP][OMPIRBuilder] Migrate createOffloadEntriesAndInfoMetadata from clang to...
Jan Sjodin [Thu, 3 Nov 2022 15:41:21 +0000 (11:41 -0400)]
[OpenMP][OMPIRBuilder] Migrate createOffloadEntriesAndInfoMetadata from clang to OpenMPIRBuilder

Remove unused [this] capture in lambda.

21 months ago[mlir][Pass] Include anchor op in -pass-pipeline
rkayaith [Tue, 18 Oct 2022 18:44:11 +0000 (14:44 -0400)]
[mlir][Pass] Include anchor op in -pass-pipeline

In D134622 the printed form of a pass manager is changed to include the
name of the op that the pass manager is anchored on. This updates the
`-pass-pipeline` argument format to include the anchor op as well, so
that the printed form of a pipeline can be directly passed to
`-pass-pipeline`. In most cases this requires updating
`-pass-pipeline='pipeline'` to
`-pass-pipeline='builtin.module(pipeline)'`.

This also fixes an outdated assert that prevented running a
`PassManager` anchored on `'any'`.

Reviewed By: rriddle

Differential Revision: https://reviews.llvm.org/D134900

21 months ago[AMDGPU] Fix duplicated words in comments
Jay Foad [Thu, 3 Nov 2022 15:33:07 +0000 (15:33 +0000)]
[AMDGPU] Fix duplicated words in comments

21 months ago[AArch64] Add a baseline test for fp16 target intrinsics. NFC
David Green [Thu, 3 Nov 2022 15:18:36 +0000 (15:18 +0000)]
[AArch64] Add a baseline test for fp16 target intrinsics. NFC

21 months ago[LoopVectorize] Fix crash on "Cannot dereference end iterator!"(PR56627)
LiDongjin [Tue, 18 Oct 2022 15:28:34 +0000 (23:28 +0800)]
[LoopVectorize] Fix crash on "Cannot dereference end iterator!"(PR56627)

Check hasOneUser before user_back().

Differential Revision: https://reviews.llvm.org/D136227

21 months ago[InstCombine] Perform memset -> load forwarding
Nikita Popov [Thu, 3 Nov 2022 09:54:16 +0000 (10:54 +0100)]
[InstCombine] Perform memset -> load forwarding

InstCombine does some basic store to load forwarding. One case it
currently misses is the case where the store is actually a memset.
This patch adds support for this case. This is a minimal
implementation that only handles a load at the memset base address,
without an offset.

GVN is already capable of performing this optimization. Having it
in InstCombine can help with phase ordering issues, similar to the
existing store to load forwarding.

Differential Revision: https://reviews.llvm.org/D137323

21 months ago[SimplifyCFG] Add tests for block speculation with assumes (NFC)
Nikita Popov [Thu, 3 Nov 2022 14:46:23 +0000 (15:46 +0100)]
[SimplifyCFG] Add tests for block speculation with assumes (NFC)

21 months ago[XCOFF] change the decoding of External symbol's function auxiliary entry in XCOFF32...
zhijian [Thu, 3 Nov 2022 14:36:44 +0000 (10:36 -0400)]
[XCOFF] change the decoding of External symbol's function auxiliary entry in XCOFF32 for llvm-readobj

Summary:

llvm-readobj decide whether to decode the external symbol's function auxiliary entry based on whether symbol is function or not currently. But the XCOFFSymbolRef::isFunction() do not work properly when -ffunction-sections is enabled. we will not decode the function auxiliary entry based on the XCOFFSymbolRef::isFunction()
we will decode the function auxiliary entry based on following:
According to the https://www.ibm.com/docs/en/aix/7.2?topic=formats-xcoff-object-file-format#XCOFF__c0f91ad419jbau

In XCOFF32, there are only "one csect Auxiliary Entry" and "a function auxiliary symbol table entry" for the C_EXT, C_WEAKEXT, and C_HIDEXT Symbols. and By convention, the csect auxiliary entry in an XCOFF32 file must be the last auxiliary entry for any external symbol that has more than one auxiliary entry( that means for the C_EXT, C_WEAKEXT, and C_HIDEXT Symbols. if there more than one auxiliary Entries. we look the last one as csect auxiliary entry. and others auxiliary entries as function entries).

Reviewers: Hubert Tong, James Henderson
Differential Revision: https://reviews.llvm.org/D136950

21 months ago[CVP] Simplify comparisons without constant operand
Nikita Popov [Wed, 2 Nov 2022 10:25:42 +0000 (11:25 +0100)]
[CVP] Simplify comparisons without constant operand

CVP currently only tries to simplify comparisons if there is a
constant operand. However, even if both are non-constant, we may
be able to determine the result of the comparison based on range
information.

IPSCCP is already capable of doing this, but because it runs very
early, it may miss some cases.

Differential Revision: https://reviews.llvm.org/D137253

21 months ago[AggressiveInstCombine] Avoid load merge/widen if stores are present b/w loads
bipmis [Thu, 3 Nov 2022 14:32:07 +0000 (14:32 +0000)]
[AggressiveInstCombine] Avoid load merge/widen if stores are present b/w loads

This patch is to address the test cases in which the load has to be inserted at a right point. This happens when there is a store b/w the loads.

This patch reverts the loads merge in all cases when stores are present b/w loads and will eventually be replaced with proper fix and test cases.

Differential Revision: https://reviews.llvm.org/D137333

21 months ago[OpenMP][OMPIRBuilder] Migrate createOffloadEntriesAndInfoMetadata from clang to...
Jan Sjodin [Wed, 2 Nov 2022 14:18:48 +0000 (10:18 -0400)]
[OpenMP][OMPIRBuilder] Migrate createOffloadEntriesAndInfoMetadata from clang to OpenMPIRBuilder

This patch moves the createOffloadEntriesAndInfoMetadata to OpenMPIRBuilder,
the createOffloadEntry helper function. The clang specific error handling is
invoked using a callback. This code will also be used by flang in the future.

21 months agoSilence a "not all control paths return" MSVC warning; NFC
Aaron Ballman [Thu, 3 Nov 2022 14:14:30 +0000 (10:14 -0400)]
Silence a "not all control paths return" MSVC warning; NFC

21 months ago[SimplifyCFG] Use range based for loop (NFC)
Nikita Popov [Thu, 3 Nov 2022 12:16:24 +0000 (13:16 +0100)]
[SimplifyCFG] Use range based for loop (NFC)

21 months ago[JITLink] Fix check-line in Windows X86 COFF COMDAT test
Stefan Gränitz [Thu, 3 Nov 2022 14:02:11 +0000 (15:02 +0100)]
[JITLink] Fix check-line in Windows X86 COFF COMDAT test

This test caught my attention because it's the only one in JITLink that XFAILs. Running it in isolation showed that the output doesn't meet the CHECK-LINES, i.e. the block address didn't match:

```
error: CHECK-NEXT: expected string not found in input
  CHECK-NEXT: block 0xfff02000 size = 0x00000001, align = 16, alignment-offset = 0

<stdin>:22:2: note: possible intended match here
              block 0xfff01000 size = 0x00000001, align = 16, alignment-offset = 0
```

Though, that doesn't appear to be the reason the test XFAILs. What we really want to check here is that llvm-jitlink doesn't fail with a duplicate section error yet.
In order to avoid issues like this in the future we can match a placeholder to check for some valid address within the slab (64Kb == last 4 digits).

The patch also drops the duplicate -noexec argument, removes an empty RUN-line, fixes indentation and adds a newline at EOF.

Reviewed By: sunho

Differential Revision: https://reviews.llvm.org/D137148

21 months agoRevert "[Assignment Tracking][3/*] Add DIAssignID metadata boilerplate"
OCHyams [Thu, 3 Nov 2022 13:46:00 +0000 (13:46 +0000)]
Revert "[Assignment Tracking][3/*] Add DIAssignID metadata boilerplate"

This reverts commit c285df77e9b78f971f9cd9d025248c20b030cc2a.

A sanitizer bot found an issue:
https://lab.llvm.org/buildbot/#/builders/5/builds/28809/steps/13/logs/stdio

21 months ago[lldb][Test] Make TestFrameFormatNameWithArgs.test more compatible across platforms
Michael Buch [Wed, 2 Nov 2022 17:18:57 +0000 (10:18 -0700)]
[lldb][Test] Make TestFrameFormatNameWithArgs.test more compatible across platforms

On Linux the `std::function` behaved differently to that on Darwin.

This patch removes usage of `std::function` in the test but attempts
to retain the test-coverage. We mainly want function types appearing
in the template argument and function argument lists.

Also add a `char const*` overload to one of the test functions to
cover the "format function argument using ValueObject formatter" code-path.

Differential Revision: https://reviews.llvm.org/D137272

21 months ago[PowerPC] Add new DMR register classes to Future CPU.
Stefan Pintilie [Wed, 2 Nov 2022 14:55:54 +0000 (09:55 -0500)]
[PowerPC] Add new DMR register classes to Future CPU.

A new register class as well as a number of related subregisters are being added
to Future CPU. These registers are Dense Math Registers (DMR) and are 1024 bits
long. These regsiters can also be used in consecutive pairs which leads to a
register that is 2048 bits.

This patch also adds 7 new instructions that use these registers. More
instructions will be added in future patches.

Reviewed By: amyk, saghir

Differential Revision: https://reviews.llvm.org/D136366

21 months ago[SimplifyCFG] Extract code for tracking ephemeral values (NFC)
Nikita Popov [Thu, 3 Nov 2022 12:08:14 +0000 (13:08 +0100)]
[SimplifyCFG] Extract code for tracking ephemeral values (NFC)

To allow reusing this in more places in SimplifyCFG.

21 months agoRemove an unused local variable, NFC.
Haojian Wu [Thu, 3 Nov 2022 12:57:46 +0000 (13:57 +0100)]
Remove an unused local variable, NFC.

21 months ago[SLP]Fix write after bounds.
Alexey Bataev [Thu, 3 Nov 2022 12:30:41 +0000 (05:30 -0700)]
[SLP]Fix write after bounds.

Need to use comma instead of + symbol to prevent writing after bounds.

21 months ago[SLP][NFC]Formatting and reduce number of iterations, NFC.
Alexey Bataev [Wed, 2 Nov 2022 21:29:24 +0000 (14:29 -0700)]
[SLP][NFC]Formatting and reduce number of iterations, NFC.

21 months ago[C++20] Diagnose invalid and reserved module names
Aaron Ballman [Thu, 3 Nov 2022 12:27:56 +0000 (08:27 -0400)]
[C++20] Diagnose invalid and reserved module names

[module.unit]p1 specifies that module and import are invalid components
of a module name, that module names cannot contain reserved
identifiers, and that std followed by zero or more digits is reserved.

The first issue (module and import pseudo-keywords) requires a
diagnostic, the second issue (use of reserved identifiers) does not
require a diagnostic. We diagnose both the same -- the code is ill-
formed unless the module declaration is in a system "header". This
allows STL implementations to use the reserved module names while
preventing users from stealing them out from under us.

Differential Revision: https://reviews.llvm.org/D136953

21 months agoAdd another test which breaks the load insert point
bipmis [Thu, 3 Nov 2022 12:28:24 +0000 (12:28 +0000)]
Add another test which breaks the load insert point

21 months ago[Clang] Implement CWG2358 Explicit capture of value
Corentin Jabot [Tue, 1 Nov 2022 16:37:06 +0000 (17:37 +0100)]
[Clang] Implement CWG2358 Explicit capture of value

Reviewed By: aaron.ballman

Differential Revision: https://reviews.llvm.org/D137172

21 months ago[ConstraintElim] Use ConstantInt::getTrue to create constants (NFC).
Florian Hahn [Thu, 3 Nov 2022 12:20:23 +0000 (12:20 +0000)]
[ConstraintElim] Use ConstantInt::getTrue to create constants (NFC).

Use existing ConstantInt::getTrue/getFalse functionality instead of
custom getScalarConstOrSplat as suggested by @nikic.

21 months ago[Flang][Driver]Add datalayout before doing LLVM-IR transformation
Mats Petersson [Fri, 2 Sep 2022 09:18:12 +0000 (10:18 +0100)]
[Flang][Driver]Add datalayout before doing LLVM-IR transformation

The earlier available datalyaout allows MLIR to LLVM-IR transformation
to use the datalayout for decisions, such as comparing sizes for
different types of integers.

This should solve https://github.com/llvm/llvm-project/issues/57230

Reviewed By: awarzynski, vzakhari

Differential Revision: https://reviews.llvm.org/D133568

21 months ago[clang-format][NFC] Fix document of AlignTrailingComments
Yusuke Kadowaki [Thu, 3 Nov 2022 12:10:48 +0000 (13:10 +0100)]
[clang-format][NFC] Fix document of AlignTrailingComments

The documentation of the patch https://reviews.llvm.org/D132131 looks
disorganized on the website
https://clang.llvm.org/docs/ClangFormatStyleOptions.html.
This patch tries to fix that.

Differential Revision: https://reviews.llvm.org/D137075

21 months ago[clang-format] Don't misannotate in CTor init list
Björn Schäpers [Mon, 24 Oct 2022 19:21:31 +0000 (21:21 +0200)]
[clang-format] Don't misannotate in CTor init list

They were annotated with TrailingAnnotation, which they are not. And
that resulted in some quirky formatting in some cases.

Differential Revision: https://reviews.llvm.org/D136635

21 months ago[clang-format] Fix lambda formatting in conditional
Björn Schäpers [Thu, 13 Oct 2022 21:19:16 +0000 (23:19 +0200)]
[clang-format] Fix lambda formatting in conditional

Without the patch UnwrappedLineFormatter::analyzeSolutionSpace just ran
out of possible formattings and would put everything just on one line.
The problem was that the the line break was forbidden, but putting the
conditional colon on the same line is also forbidden.

Differential Revision: https://reviews.llvm.org/D135918

21 months ago[Assignment Tracking][3/*] Add DIAssignID metadata boilerplate
OCHyams [Thu, 3 Nov 2022 09:50:31 +0000 (09:50 +0000)]
[Assignment Tracking][3/*] Add DIAssignID metadata boilerplate

The Assignment Tracking debug-info feature is outlined in this RFC:

https://discourse.llvm.org/t/
rfc-assignment-tracking-a-better-way-of-specifying-variable-locations-in-ir

Add the DIAssignID metadata attachment boilerplate. Includes a textual-bitcode
roundtrip test and tests that the verifier and parser catch badly formed IR.

This piece of metadata links together stores (used as an attachment) and the
yet-to-be-added llvm.dbg.assign debug intrinsic (used as an operand).

Reviewed By: jmorse

Differential Revision: https://reviews.llvm.org/D132222

21 months ago[AMDGPU] Fix GCNDownwardRPTracker::advanceBeforeNext at the end of MBB
Valery Pykhtin [Thu, 27 Oct 2022 12:38:00 +0000 (14:38 +0200)]
[AMDGPU] Fix GCNDownwardRPTracker::advanceBeforeNext at the end of MBB

The problem with GCNDownwardRPTracker::advanceBeforeNext is that it doesn't allow to get register pressure
after the last instruction in a MBB.

However when we track RP through the boundary of a MBB we need the state that is after the last instruction
of the MBB and before the first instruction of the successor MBB. Currently we stop traking RP in the state
 'at' the last instruction of the MBB which is incorrect.

This patch fixes 27 lit tests with EXPENSIVE_CHECKS enabled.

Reviewed By: rampitec, arsenm

Differential Revision: https://reviews.llvm.org/D136927

21 months ago[flang] Add -f[no-]honor-infinities and -menable-no-infs
Tom Eccles [Mon, 17 Oct 2022 17:16:24 +0000 (17:16 +0000)]
[flang] Add -f[no-]honor-infinities and -menable-no-infs

Only add the option processing and store the result. No attributes are
added to FIR yet.

This patch follows Clang in forwarding -fno-honor-infinities as
-menable-no-infs.

Reviewed By: kiranchandramohan awarzynski vzakhari

Differential Revision: https://reviews.llvm.org/D137072

21 months ago[InstCombine] Add more memset->load forwarding tests (NFC)
Nikita Popov [Thu, 3 Nov 2022 09:55:30 +0000 (10:55 +0100)]
[InstCombine] Add more memset->load forwarding tests (NFC)

21 months ago[AArch64] Add support for the Cortex-A715 CPU
Simi Pallipurath [Fri, 28 Oct 2022 15:06:00 +0000 (16:06 +0100)]
[AArch64] Add support for the Cortex-A715 CPU

Cortex-A715 is an Armv9-A AArch64 CPU.

This patch introduces support for Cortex-A715.

Technical Reference Manual: https://developer.arm.com/documentation/101590/latest.

Reviewed By: vhscampos

Differential Revision: https://reviews.llvm.org/D136957

21 months ago[InstCombine] Add tests for memset -> load forward with offset (NFC)
Nikita Popov [Thu, 3 Nov 2022 09:20:38 +0000 (10:20 +0100)]
[InstCombine] Add tests for memset -> load forward with offset (NFC)

21 months ago[InstCombine] Add tests for memset -> load forwarding (NFC)
Nikita Popov [Thu, 3 Nov 2022 09:15:10 +0000 (10:15 +0100)]
[InstCombine] Add tests for memset -> load forwarding (NFC)

We currently only forward store -> load, but could do the same
for memset as well.

21 months ago[CVP] Add vector icmp test (NFC)
Nikita Popov [Thu, 3 Nov 2022 08:56:29 +0000 (09:56 +0100)]
[CVP] Add vector icmp test (NFC)

21 months ago[AArch64] SME2 -Fix failing buildbots because of warning
Caroline Concatto [Thu, 3 Nov 2022 08:30:09 +0000 (08:30 +0000)]
[AArch64] SME2 -Fix failing buildbots because of warning

This patch is to solve this:
    https://lab.llvm.org/buildbot#builders/36/builds/26801
Created by this patch: a20112a74cb34f

[AArch64]SME2 instructions that use ZTO operand

21 months ago[DebugInfo][Metadata] Make AllEnumTypes holding TrackingMDNodeRef
Kristina Bessonova [Thu, 3 Nov 2022 08:27:10 +0000 (10:27 +0200)]
[DebugInfo][Metadata] Make AllEnumTypes holding TrackingMDNodeRef

Having AllEnumtypes to be a vector of TrackingMDNodeRef makes it possible
to reflect changes in metadata in the vector if they took place before DIBuilder
being finalized.

Otherwise, we end up with heap-use-after-free because AllEnumTypes contains
metadata that no longer valid.

Consider a case where we have a class containing a definition of a enum,
so this enum has the class as a scope. For some reason (doesn't matter for
the current issue), we create a temporary debug metadata for this class, and
then resolve it while finalizing CGDebugInfo.

In the case of collision during uniqifying the temporary, we then need
to replace its uses with a new pointer. If a temporary's user is unique
(this is the enum mentioned above), we may need re-uniquefying it,
which may return a new pointer in the case of another collision.
If so, the pointer we stored in AllEnumTypes vector become dangling.
Making AllEnumTypes hodling TrackingMDNodeRef should solve this problem
(see debug-info-enum-metadata-collision.cpp test for details).

Reviewed By: dblaikie

Differential Revision: https://reviews.llvm.org/D137067

21 months ago[llvm-diff] Precommit: Add loop test case with forward reference
Jannik Silvanus [Wed, 2 Nov 2022 15:51:01 +0000 (16:51 +0100)]
[llvm-diff] Precommit: Add loop test case with forward reference

Diffing phi nodes was recently added to llvm-diff.
However, there currently is a limitation where equivalent values
cannot be detected as such, leading to false positive diff reports.

If a phi node refers a value defined in a basic block dominated by
the current basic block, for example a phi node in a loop header referring
a value defined in the loop body, we cannot prove equivalence of the referred
values, because the basic block containing the variable definition has
not yet been processed.

This commit adds a test case showing this behavior, serving as a precommit
for an upcoming fix of the above.

Differential Revision: https://reviews.llvm.org/D137262

21 months ago[GISel] Add (fsub +-0.0, X) -> fneg combine
Pierre van Houtryve [Fri, 21 Oct 2022 11:23:55 +0000 (11:23 +0000)]
[GISel] Add (fsub +-0.0, X) -> fneg combine

Allows for better matching of VOP3 mods.

Reviewed By: arsenm

Differential Revision: https://reviews.llvm.org/D136442

21 months agoRevert "[InstCombine] Remove redundant splats in InstCombineVectorOps"
Peter Waller [Thu, 3 Nov 2022 07:56:03 +0000 (07:56 +0000)]
Revert "[InstCombine] Remove redundant splats in InstCombineVectorOps"

This reverts commit 957eed0b1af2cb88edafe1ff2643a38165c67a40.

21 months ago[AArch64] Install arm_neon_sve_bridge.h
Peter Waller [Wed, 2 Nov 2022 10:19:34 +0000 (10:19 +0000)]
[AArch64] Install arm_neon_sve_bridge.h

arm_neon_sve_bridge.h is not generated, so the rules which ensure the
generated files get copied into the installation prefix don't apply to
this one.

Add it to the aarch64_only_files set instead, which ensures it ends up
both in the build directory and the installation directory.

Tested with build targets `clang-resource-headers` and
`install-clang-resource-headers`.

Differential Revision: https://reviews.llvm.org/D137239

21 months ago[Clang] Check for response file existence prior to check for recursion
Serge Pavlov [Thu, 3 Nov 2022 05:26:49 +0000 (12:26 +0700)]
[Clang] Check for response file existence prior to check for recursion

As now errors in file operation are handled, check for file existence
must be done prior to check for recursion, otherwise reported errors are
misleading.

Differential Revision: https://reviews.llvm.org/D136090

21 months ago[AArch64]SME2 instructions that use ZTO operand
Caroline Concatto [Fri, 28 Oct 2022 14:51:20 +0000 (15:51 +0100)]
[AArch64]SME2 instructions that use ZTO operand

This patch adds the assembly/disassembly for the following instructions:
  ZERO (ZT0): Zero ZT0.
  LDR (ZT0): Load ZT0 register.
  STR (ZT0): Store ZT0 register.
  MOVT (scalar to ZT0): Move 8 bytes from general-purpose register to ZT0.
       (ZT0 to scalar): Move 8 bytes from ZT0 to general-purpose register.
 Consecutive:
   LUTI2 (single): Lookup table read with 2-bit indexes.
         (two registers): Lookup table read with 2-bit indexes.
         (four registers): Lookup table read with 2-bit indexes.
   LUTI4 (single): Lookup table read with 4-bit indexes.
         (two registers): Lookup table read with 4-bit indexes.
         (four registers): Lookup table read with 4-bit indexes.

The reference can be found here:

https://developer.arm.com/documentation/ddi0602/2022-09

This patch also adds a new register class and operand for zt0
and a another index operand uimm3s8

Differential Revision: https://reviews.llvm.org/D136088

21 months ago[mlir][Bazel] Add mlir-translate as data for dialect tests.
Adrian Kuegel [Thu, 3 Nov 2022 07:12:19 +0000 (08:12 +0100)]
[mlir][Bazel] Add mlir-translate as data for dialect tests.

21 months agoRevert "[AAPointerInfo] refactor how offsets and Access objects are tracked"
Ye Luo [Thu, 3 Nov 2022 05:01:51 +0000 (00:01 -0500)]
Revert "[AAPointerInfo] refactor how offsets and Access objects are tracked"

This reverts commit b756096b0cbef0918394851644649b3c28a886e2.
See regression https://github.com/llvm/llvm-project/issues/58774

21 months ago[NFC][flang] Lowering options clean-up.
Slava Zakharin [Tue, 1 Nov 2022 22:12:43 +0000 (15:12 -0700)]
[NFC][flang] Lowering options clean-up.

This change-set defines the LoweringOptions the same way
other options are defined in Flang.

Differential Revision: https://reviews.llvm.org/D137207

21 months ago[LoongArch] Fix codegen for [su]itofp instructions
wanglei [Thu, 3 Nov 2022 03:10:41 +0000 (11:10 +0800)]
[LoongArch] Fix codegen for [su]itofp instructions

This patch fixes codegen for `[su]itofp` instructions.

In LoongArch, a legal int-to-float conversion is done in two steps:

1. Move the data from `GPR` to `FPR`. (FRLen >= GRLen)
2. Conversion in `FPR`. (the data in `FPR` is treated as a signed value)

Based on the above features, when the type's BitWidth meets the
requirements, all `SINT_TO_FP` are legal, all `UINT_TO_FP` are expand
and lowered to libcall when appropriate.
The only special case is, LoongArch64 with `+f,-d` features. At this
point, custom processing is required for `[SU]INT_TO_FP`. Of course, we
can also ignore it and use libcall directly.

Differential Revision: https://reviews.llvm.org/D136916

21 months ago[X86] Remove `IntrArgMemOnly` from cmpccxadd intrinsics
Phoebe Wang [Thu, 3 Nov 2022 02:39:31 +0000 (10:39 +0800)]
[X86] Remove `IntrArgMemOnly` from cmpccxadd intrinsics

CMPSXADD will modify memory, so we can't use `IntrArgMemOnly` here.
Found it during review D137250.

21 months ago[sanitizer] Fix vfork interception on loongarch64
Xi Ruoyao [Thu, 3 Nov 2022 03:07:53 +0000 (11:07 +0800)]
[sanitizer] Fix vfork interception on loongarch64

Fix a brown paper bag error made by me in D129418.  I didn't set
ASAN_INTERCEPT_VFORK correctly for loongarch64, but created an all-zero
object for __interception::real_vfork.  This caused anything calling
vfork() to die instantly.

Fix this issue by setting ASAN_INTERCEPT_VFORK and remove the bad
all-zero definition.  Other ports have an all-zero common definition but
we don't need it at least for now.

And, enable ASAN vfork test for loongarch64 to prevent regression in the
future.

Differential Revision: https://reviews.llvm.org/D137160

21 months ago[asan] Use proper shadow offset for loongarch64 in instrumentation passes
Youling Tang [Thu, 3 Nov 2022 03:03:37 +0000 (11:03 +0800)]
[asan] Use proper shadow offset for loongarch64 in instrumentation passes

Instrumentation passes now use the proper shadow offset. There will be many
asan test failures without this patch. For example:

```
$ ./lib/asan/tests/LOONGARCH64LinuxConfig/Asan-loongarch64-calls-Test
AddressSanitizer:DEADLYSIGNAL
=================================================================
==651209==ERROR: AddressSanitizer: SEGV on unknown address 0x1ffffe2dfa9b (pc 0x5555585e151c bp 0x7ffffb9ec070 sp 0x7ffffb9ebfd0 T0)
==651209==The signal is caused by a UNKNOWN memory access.
```

Before the patch:

```
$ make check-asan
Testing Time: 36.13s
  Unsupported      : 205
  Passed           :  83
  Expectedly Failed:   1
  Failed           : 239
```

After the patch:
```
$ make check-asan
Testing Time: 58.98s
  Unsupported      : 205
  Passed           : 421
  Expectedly Failed:   1
  Failed           :  89
```

Differential Revision: https://reviews.llvm.org/D137013

21 months ago[asan] Default to -fsanitize-address-use-odr-indicator for non-Windows
Fangrui Song [Thu, 3 Nov 2022 02:21:33 +0000 (19:21 -0700)]
[asan] Default to -fsanitize-address-use-odr-indicator for non-Windows

This enables odr indicators on all platforms and private aliases on non-Windows.
Note that GCC also uses private aliases: this fixes bogus
`The following global variable is not properly aligned.` errors for interposed global variables

Fix https://github.com/google/sanitizers/issues/398
Fix https://github.com/google/sanitizers/issues/1017
Fix https://github.com/llvm/llvm-project/issues/36893 (we can restore D46665)

Global variables of non-hasExactDefinition() linkages (i.e.
linkonce/linkonce_odr/weak/weak_odr/common/external_weak) are not instrumented.
If an instrumented variable gets interposed to an uninstrumented variable due to
symbol interposition (e.g. in issue 36893, _ZTS1A in foo.so is resolved to _ZTS1A in
the executable), there may be a bogus error.

With private aliases, the register code will not resolve to a definition in
another module, and thus prevent the issue.

Cons: minor size increase. This is mainly due to extra `__odr_asan_gen_*` symbols.
(ELF) In addition, in relocatable files private aliases replace some relocations
referencing global symbols with .L symbols and may introduce some STT_SECTION symbols.

For lld, with -g0, the size increase is 0.07~0.09% for many configurations I
have tested: -O0, -O1, -O2, -O3, -O2 -ffunction-sections -fdata-sections
-Wl,--gc-sections. With -g1 or above, the size increase ratio will be even smaller.

This patch obsoletes D92078.

Don't migrate Windows for now: the static data member of a specialization
`std::num_put<char>::id` is a weak symbol, as well as its ODR indicator.
Unfortunately, link.exe (and lld without -lldmingw) generally doesn't support
duplicate weak definitions (weak symbols in different TUs likely pick different
defined external symbols and conflict).

Differential Revision: https://reviews.llvm.org/D137227

21 months ago[flang][RFC] Add lowering design for procdure pointers
Peixin-Qiao [Thu, 3 Nov 2022 01:00:52 +0000 (09:00 +0800)]
[flang][RFC] Add lowering design for procdure pointers

This document aims to give insights at the representation of procdure
pointers in FIR.

Reviewed By: PeteSteinfeld, jeanPerier, kiranchandramohan

Differential Revision: https://reviews.llvm.org/D136840

21 months agoLangRef: Fix typo in backtick placement
Matt Arsenault [Wed, 2 Nov 2022 23:13:12 +0000 (16:13 -0700)]
LangRef: Fix typo in backtick placement

21 months ago[Driver] Don't preprocess source files when reproducing linker crashes
Alex Brachet [Thu, 3 Nov 2022 00:12:10 +0000 (00:12 +0000)]
[Driver] Don't preprocess source files when reproducing linker crashes

It's not necessary to redo the source file preprocessing for reproducing linker
crashes because we must have successfully created the object file by this point.
Skip this step, and also don't report the preprocessed source file or create
the clang invocation shell script. The latter is no longer sensible without the
preprocessed source, or helpful given the linker reproducer will have it's own
shell script.

Differential Revision: https://reviews.llvm.org/D137289

21 months ago[ConstraintElimination] Skip compares with scalable vector types.
Florian Hahn [Wed, 2 Nov 2022 23:57:06 +0000 (23:57 +0000)]
[ConstraintElimination] Skip compares with scalable vector types.

Materializing scalable vectors with boolean values is not implemented
yet. Skip those cases for now and leave a TODO.

21 months ago[Clang] follow-up D128745, remove ClangABICompat checks
Yuanfang Chen [Wed, 2 Nov 2022 18:54:19 +0000 (11:54 -0700)]
[Clang] follow-up D128745, remove ClangABICompat checks

Per discussions in D128745, remove ClangABICompat checks for implementations
of DR692/DR1395/DR1432. This is a potentially breaking changes, so the release
note is updated accordingly.

Reviewed By: aaron.ballman

Differential Revision: https://reviews.llvm.org/D136120

21 months agoSPARC: Register null target streamer
Matt Arsenault [Sat, 29 Oct 2022 19:52:24 +0000 (12:52 -0700)]
SPARC: Register null target streamer

Fixes null dereference in emitFunctionBodyStart for 64-bit

21 months agoWebAssembly: Remove unnecessary set check
Matt Arsenault [Wed, 2 Nov 2022 21:29:14 +0000 (14:29 -0700)]
WebAssembly: Remove unnecessary set check

The empty set will be default constructed if this wasn't
in the map already.

21 months agoWebAssembly: Move exception handling code together
Matt Arsenault [Wed, 2 Nov 2022 20:44:57 +0000 (13:44 -0700)]
WebAssembly: Move exception handling code together

21 months agoFunctionLoweringInfo: Use TLI member instead of finding it
Matt Arsenault [Wed, 2 Nov 2022 20:17:30 +0000 (13:17 -0700)]
FunctionLoweringInfo: Use TLI member instead of finding it

21 months ago[mlir] Add call_intrinsic op to LLVMIIR
electriclilies [Wed, 2 Nov 2022 22:56:28 +0000 (15:56 -0700)]
[mlir] Add call_intrinsic op to LLVMIIR

The call_intrinsic op allows us to call LLVM intrinsics from the LLVMDialect without implementing a new op every time.

Reviewed By: lattner, rriddle

Differential Revision: https://reviews.llvm.org/D137187

21 months ago[libc] Add implementation of ungetc.
Siva Chandra Reddy [Wed, 2 Nov 2022 20:19:04 +0000 (20:19 +0000)]
[libc] Add implementation of ungetc.

A bug in the file read logic has also been fixed along the way. Parts
of the ungetc tests will fail without that bug fixed.

Reviewed By: michaelrj

Differential Revision: https://reviews.llvm.org/D137286

21 months ago[ARM] Fix vector ule zero lowering
David Green [Wed, 2 Nov 2022 22:34:05 +0000 (22:34 +0000)]
[ARM] Fix vector ule zero lowering

The instruction icmp ule <4 x i32> %0, zeroinitializer will usually be
simplified to icmp eq <4 x i32> %0, zeroinitializer. It is not
guaranteed though, and the code for lowering vector compares could pick
the wrong form of the instruction if this happened. I've tried to make
the code more explicit about the supported conditions.

This fixes NEON being unable to select VCMPZ with HS conditions, and
fixes some incorrect MVE patterns.

Fixes #58514.

Differential Revision: https://reviews.llvm.org/D136447

21 months ago[libc++][Android] strong_order_long_double.verify.cpp: disable on i686
Ryan Prichard [Wed, 2 Nov 2022 21:55:56 +0000 (14:55 -0700)]
[libc++][Android] strong_order_long_double.verify.cpp: disable on i686

This target (as well as 32-bit ARM Android) have sizeof(long double)
equal to sizeof(double).

Reviewed By: ldionne, #libc

Differential Revision: https://reviews.llvm.org/D137135

21 months ago[libc++][Android] XFAIL aligned_alloc and timespec_get tests
Ryan Prichard [Wed, 2 Nov 2022 21:53:51 +0000 (14:53 -0700)]
[libc++][Android] XFAIL aligned_alloc and timespec_get tests

Mark tests XFAIL that use APIs that are unsupported on old versions of
Android:
 - aligned_alloc isn't available until API 28.
 - timespec_get isn't available until API 29.

Reviewed By: ldionne, #libc

Differential Revision: https://reviews.llvm.org/D137134

21 months ago[OPENMP]Initial support for error directive.
Jennifer Yu [Tue, 1 Nov 2022 21:46:12 +0000 (14:46 -0700)]
[OPENMP]Initial support for error directive.

Differential Revision: https://reviews.llvm.org/D137209

21 months ago[Sanitizers] Modified __aarch64__ to use the 64 bit version of the allocator.
Kirill Stoimenov [Wed, 2 Nov 2022 20:58:36 +0000 (20:58 +0000)]
[Sanitizers] Modified __aarch64__ to use the 64 bit version of the allocator.

This change will switch SizeClassAllocator32 to SizeClassAllocator64 on ARM. This might potentially affect ARM platforms with 39-bit address space. This addresses [[ https://github.com/google/sanitizers/issues/703  | issues/703  ]], but unlike [[ https://reviews.llvm.org/D60243 | D60243 ]] it defaults to 64 bit allocator.

Reviewed By: vitalybuka, MaskRay

Differential Revision: https://reviews.llvm.org/D137136

21 months ago[libc++][test] Fixes transitive includes.
Mark de Wever [Wed, 2 Nov 2022 18:43:53 +0000 (19:43 +0100)]
[libc++][test] Fixes transitive includes.

These were accidentally set to generating in
243da90ea5357c1ca324f714ea4813dc9029af27

Reviewed By: #libc, philnik

Differential Revision: https://reviews.llvm.org/D137278

21 months ago[gn build] Port cf239c2f1777
LLVM GN Syncbot [Wed, 2 Nov 2022 21:01:56 +0000 (21:01 +0000)]
[gn build] Port cf239c2f1777

21 months ago[mlir] Remove eliminateBlockingAwaitOps option in AsyncToAsyncRuntime pass
yijiagu [Wed, 2 Nov 2022 20:51:48 +0000 (13:51 -0700)]
[mlir] Remove eliminateBlockingAwaitOps option in AsyncToAsyncRuntime pass

Remove the eliminateBlockingAwaitOps option in AsyncToAsyncRuntime pass

Today the AsyncToAsyncRuntime pass does two things: one is converting normal funcs with async ops to coroutine cfg; the other is lowering high level async operations to async.coro and async.runtime operations. This patch removes the converting step from AsyncToAsyncRuntime pass.

In the next step we will create a new asyncfication pass for converting normal funcs to the newly added async.func operation.

Reviewed By: ezhulenev

Differential Revision: https://reviews.llvm.org/D137282

21 months ago[NFC] Make format() more amenable to format attributes
Félix Cloutier [Fri, 9 Sep 2022 21:08:19 +0000 (14:08 -0700)]
[NFC] Make format() more amenable to format attributes

This change modifies the implementation of the format() function
so that vendor forks committed to building with compilers that
support __attribute__((format)) on non-variadic functions can
check the format() function with it.

rdar://84571523

21 months ago[RISCV] Prevent autovectorization using vscale with Zvl32b.
Craig Topper [Wed, 2 Nov 2022 20:09:52 +0000 (13:09 -0700)]
[RISCV] Prevent autovectorization using vscale with Zvl32b.

RVVBitsPerBlock is 64. If VLen==32, VLen/RVVBitsPerBlock is 0.

Reviewed By: reames

Differential Revision: https://reviews.llvm.org/D137280

21 months ago[clang][pp] only __is_target_environment(unknown) should match unknown target triple...
Alex Lorenz [Wed, 2 Nov 2022 20:40:52 +0000 (13:40 -0700)]
[clang][pp] only __is_target_environment(unknown) should match unknown target triple environment

21 months ago[VectorCombine] add tests for load+shuffle and update to typeless ptr; NFC
Sanjay Patel [Wed, 2 Nov 2022 19:55:51 +0000 (15:55 -0400)]
[VectorCombine] add tests for load+shuffle and update to typeless ptr; NFC

21 months ago[OpenMP][libomp] Fix disabled affinity
Jonathan Peyton [Tue, 1 Nov 2022 17:29:17 +0000 (12:29 -0500)]
[OpenMP][libomp] Fix disabled affinity

Fix setting affinity type and topology method when affinity is disabled
and fix places that were not taking into account that affinity can be
explicitly disabled by putting proper KMP_AFFINITY_CAPABLE() check.

Differential Revision: https://reviews.llvm.org/D137176

21 months ago[clang-format] Don't skip #else/#elif of #if 0
Owen Pan [Sun, 30 Oct 2022 23:45:00 +0000 (16:45 -0700)]
[clang-format] Don't skip #else/#elif of #if 0

Fixes #58188.

Differential Revision: https://reviews.llvm.org/D137052

21 months agoRevert "[Attributor][NFCI] Move MemIntrinsic handling into the initializer"
Joseph Huber [Wed, 2 Nov 2022 20:28:50 +0000 (15:28 -0500)]
Revert "[Attributor][NFCI] Move MemIntrinsic handling into the initializer"

This was causing failures when optimizing codes with complex numbers.
Revert until a fix can be implemented.

This reverts commit 7fdf3564c04075d3e6be2d9540e5a6f1e084be9f.

21 months ago[LinkerWrapper] report on missing libraries
Joseph Huber [Tue, 1 Nov 2022 18:16:47 +0000 (18:16 +0000)]
[LinkerWrapper] report on missing libraries

The linker wrapper does its own library searching for static archives
that can contain device code. The device linking phases happen before
the host linking phases so that we can generate the necessary
registration code and link it in with the rest of the code. Previously,
If a library containing needed device code was not found the execution
would continue silently until it failed with undefined symbols. This
patch allows the linker wrapper to perform its own check beforehand to
catch these errors.

Reviewed By: jdoerfert

Differential Revision: https://reviews.llvm.org/D137180

21 months ago[clang] Fix inline builtin functions of an __asm__ renamed function with symbol prefixes
Martin Storsjö [Mon, 31 Oct 2022 10:18:21 +0000 (12:18 +0200)]
[clang] Fix inline builtin functions of an __asm__ renamed function with symbol prefixes

If a function is renamed with `__asm__`, the name provided is the
exact symbol name, without any extra implicit symbol prefixes.
If the target does use symbol prefixes, the IR level symbol gets
an `\01` prefix to indicate that it's a literal symbol name to be
taken as is.

When a builtin function is specialized by providing an inline
version of it, that inline function is named `<funcname>.inline`.

When the base function has been renamed due to `__asm__`, the inline
function ends up named `<asmname>.inline`. Up to this point,
things did work as expected before.

However, for targets with symbol prefixes, one codepath that produced
the combined name `<asmname>.inline` used the mangled `asmname` with
`\01` prefix, while others didn't. This patch fixes this.

This fixes the combination of asm renamed builtin function, with
inline override of the function, on any target with symbol
prefixes (such as i386 windows and any Darwin target).

Differential Revision: https://reviews.llvm.org/D137073

21 months ago[RISCV] Move RVVBitsPerBlock to TargetParser.h so we can use it in clang. NFC
Craig Topper [Wed, 2 Nov 2022 16:18:41 +0000 (09:18 -0700)]
[RISCV] Move RVVBitsPerBlock to TargetParser.h so we can use it in clang. NFC

Differential Revision: https://reviews.llvm.org/D137266

21 months ago[LV][RISCV] Add test showing poor choice of VF for short loop
Philip Reames [Wed, 2 Nov 2022 20:00:35 +0000 (13:00 -0700)]
[LV][RISCV] Add test showing poor choice of VF for short loop

21 months ago[wasm-ld] Update supported features in the generic CPU configuration
Dan Gohman [Tue, 18 Oct 2022 18:10:19 +0000 (11:10 -0700)]
[wasm-ld] Update supported features in the generic CPU configuration

Accompanying https://reviews.llvm.org/D125728, this updates LLVM
Codegen's "generic" CPU to enable the same new features.

Differential Revision: https://reviews.llvm.org/D125729

21 months ago[flang][NFC] Fix typo in filename
Valentin Clement [Wed, 2 Nov 2022 19:47:19 +0000 (20:47 +0100)]
[flang][NFC] Fix typo in filename

21 months ago[flang] Fix for polymoprhic pointer component
Valentin Clement [Wed, 2 Nov 2022 19:46:09 +0000 (20:46 +0100)]
[flang] Fix for polymoprhic pointer component

Fix path that generates MutableBox for
pointer component.

Reviewed By: PeteSteinfeld

Differential Revision: https://reviews.llvm.org/D137270