platform/upstream/llvm.git
20 months ago[mlir][tensor] Do not use affine ops on non-perfect-tiling unpack cases.
Hanhan Wang [Thu, 16 Feb 2023 01:36:21 +0000 (17:36 -0800)]
[mlir][tensor] Do not use affine ops on non-perfect-tiling unpack cases.

Do not create an Affine ops for expanded size because the affine op is
too complicated which would hit an assertion in affine ops
simplification.

Reviewed By: mravishankar

Differential Revision: https://reviews.llvm.org/D144151

20 months ago[RISCV] For rv32, accept constants like 0xfffff800 as a valid simm12.
Craig Topper [Fri, 17 Feb 2023 18:53:44 +0000 (10:53 -0800)]
[RISCV] For rv32, accept constants like 0xfffff800 as a valid simm12.

Internally we store constants in int64_t after parsing, but this is
kind of an implementation detail. If we only supported rv32, we might
have chosen int32_t.

For rv32, I think it makes sense to accept the constants that we
would accept if int32_t was the internal type. In fact we already
do this for the `li` alias. This patch extends this to sign
extended constants for other instructions.

This matches the GNU assembler. The difference between LLVM and gcc
was previously noted here. https://github.com/riscv-non-isa/riscv-asm-manual/pull/71

Reviewed By: reames

Differential Revision: https://reviews.llvm.org/D144166

20 months agoRevert "[RISCV] Add performMULcombine to perform strength-reduction"
Philipp Tomsich [Fri, 17 Feb 2023 18:46:13 +0000 (19:46 +0100)]
Revert "[RISCV] Add performMULcombine to perform strength-reduction"

This reverts commit 20cc23c708f04ca3fbc4289a68302a4b684ce448.

20 months agoRevert "[RISCV] Add vendor-defined XTheadMemPair (two-GPR Memory Operations) extension"
Philipp Tomsich [Fri, 17 Feb 2023 18:45:55 +0000 (19:45 +0100)]
Revert "[RISCV] Add vendor-defined XTheadMemPair (two-GPR Memory Operations) extension"

This reverts commit d2918544a7fc4b5443879fe12f32a712e6dfe325.

20 months ago[RISCV] xtheadmac: fix commutativity issue for the in/out register
Manolis Tsamis [Fri, 17 Feb 2023 18:32:07 +0000 (19:32 +0100)]
[RISCV] xtheadmac: fix commutativity issue for the in/out register

The instructions in the XTHeadMac extension (multiply accumulate
instructions) were marked as commutative but because the destination
register was also an input (accumulate) register and was connected to
the destination register with a register allocator constraint, all
three operands (instead of two) were incorrectly considered
commutative. To fix that an appropriate fixCommutedOpIndices call was
added for these instructions in findCommutedOpIndices

New test functions have been added to test the correct behaviour in
xtheadmac.ll.

Reviewed By: craig.topper

Differential Revision: https://reviews.llvm.org/D144278

20 months ago[RISCV] Add performMULcombine to perform strength-reduction
Philipp Tomsich [Fri, 3 Feb 2023 21:03:07 +0000 (22:03 +0100)]
[RISCV] Add performMULcombine to perform strength-reduction

The RISC-V backend thus far does not provide strength-reduction, which
causes a long (but not complete) list of 3-instruction patterns listed
to utilize the shift-and-add instruction from Zba and XTHeadBa in
strength-reduction.

This adds the logic to perform strength-reduction through the DAG
combine for ISD::MUL.  Initially, we wire this up for XTheadBa only,
until this has had some time to settle and get real-world test
exposure.

The following strength-reductions strategies are currently supported:
  - XTheadBa
    - C = (n + 1)           // th.addsl
    - C = (n + 1)k          // th.addsl, slli
    - C = (n + 1)(m + 1)    // th.addsl, th.addsl
    - C = (n + 1)(m + 1)k   // th.addsl, th.addsl, slli
    - C = ((n + 1)m + 1)    // th.addsl, th.addsl
    - C = ((n + 1)m + 1)k   // th.addslm th.addsl, slli
  - base ISA
    - C being 2 set-bits    // slli, slli, add
       (possibly slli, th.addsl)

Even though the slli+slli+add sequence would we supported without
XTheadBa, this currently is gated to avoid having to update a large
number of test cases (i.e., anything that has a multiplication with a
constant where only 2 bits are set) in this commit.

With the strength reduction now being performed in performMUL combine,
we drop the (now redundant) patterns from RISCVInstrInfoXTHead.td.

Depends on D143029

Differential Revision: https://reviews.llvm.org/D143394

20 months ago[RISCV] Add vendor-defined XTheadMemPair (two-GPR Memory Operations) extension
Manolis Tsamis [Mon, 13 Feb 2023 13:07:45 +0000 (14:07 +0100)]
[RISCV] Add vendor-defined XTheadMemPair (two-GPR Memory Operations) extension

The vendor-defined XTHeadMemPair (no comparable standard extension exists
at the time of writing) extension adds two-GPR load/store pair instructions.

It is supported by the C9xx cores (e.g., found in the wild in the
Allwinner D1) by Alibaba T-Head.

The current (as of this commit) public documentation for this
extension is available at:
  https://github.com/T-head-Semi/thead-extension-spec/releases/download/2.2.2/xthead-2023-01-30-2.2.2.pdf

Support for these instructions has already landed in GNU Binutils:
  https://sourceware.org/git/?p=binutils-gdb.git;a=commit;h=6e17ae625570ff8f3c12c8765b8d45d4db8694bd

Depends on D143847

Differential Revision: https://reviews.llvm.org/D144002

20 months ago[libc++] Add missing include in <__atomic/cxx_atomic_impl.h>
Nikolas Klauser [Fri, 17 Feb 2023 18:33:16 +0000 (19:33 +0100)]
[libc++] Add missing include in <__atomic/cxx_atomic_impl.h>

20 months ago[RISCV][CodeGen] Add codegen patterns for experimental zfa extension (try 2)
Jun Sha (Joshua) [Fri, 17 Feb 2023 17:36:18 +0000 (09:36 -0800)]
[RISCV][CodeGen] Add codegen patterns for experimental zfa extension (try 2)

Recommit by preames with commit message, various style cleanups, and unaddressed review comments corrected.

This patch implements experimental codegen support for the RISCV Zfa extension as specified here: https://github.com/riscv/riscv-isa-manual/releases/download/draft-20221119-5234c63/riscv-spec.pdf, Ch. 25. This extension has not been ratified.

This change does not include support for FLI (upcoming in a follow up change) or FCVTMOD (not relevant for C/C++).

Differential Revision: https://reviews.llvm.org/D143982

20 months ago[IR] Fix a warning
Kazu Hirata [Fri, 17 Feb 2023 18:21:10 +0000 (10:21 -0800)]
[IR] Fix a warning

This patch fixes:

  llvm/lib/IR/Instruction.cpp:141:20: warning: unused variable ‘CB’ [-Wunused-variable]

20 months ago[mlir] Reintroduce API for creating operations with a DictionaryAttr
Jeff Niu [Thu, 16 Feb 2023 18:27:10 +0000 (10:27 -0800)]
[mlir] Reintroduce API for creating operations with a DictionaryAttr

This patch reintroduces an API to create operations with a pre-existing
DictionaryAttr. This API does not populate the attributes with any
default attributes the operation may have, like the API that takes a
NamedAttrList does. NamedAttrList is effective at not re-hashing the
attributes if no default attributes were added, but this new API speeds
up clone-heavy workloads slightly (~5%).

Reviewed By: rriddle

Differential Revision: https://reviews.llvm.org/D144204

20 months ago[ADT] Fix tests for `StringMap::at` and `DenseMap::at`
Ryan Guo [Fri, 17 Feb 2023 17:07:35 +0000 (09:07 -0800)]
[ADT] Fix tests for `StringMap::at` and `DenseMap::at`

These methods won't assert for release build.

20 months ago[RISCV] Improve isInterleaveShuffle to handle interleaving the high half and low...
Craig Topper [Fri, 17 Feb 2023 18:00:40 +0000 (10:00 -0800)]
[RISCV] Improve isInterleaveShuffle to handle interleaving the high half and low half of the same source.

This is needed to support the new interleave intrinsics from D141924 for
fixed vectors.

I've reworked the core loop to operate in terms of half of a source. Making 4
possible half sources. The first element of the half is used to indicate which
source using the same numbering as the shuffle where the second source elements
are numbered after the first source.

I've added restrictions to only match the first half of two vectors or the
first and second half of a single vector. This was done to prevent regressions
on the cases we have coverage for. I saw cases where generic DAG combine split
a single interleave into 2 smaller interleaves a concat. We can revisit in the
future.

Reviewed By: reames

Differential Revision: https://reviews.llvm.org/D144143

20 months ago[ELF][llvm-objcopy] Reject duplicate SHT_SYMTAB sections
Moshe Berman [Fri, 17 Feb 2023 17:18:14 +0000 (17:18 +0000)]
[ELF][llvm-objcopy] Reject duplicate SHT_SYMTAB sections

The gABI prohibits multiple SH_SYMTAB sections. As a result,
llvm-objcopy was crashing in SymbolTableSection::removeSymbols(). This
patch fixes the issue by emitting an error if multiple SH_SYMTAB
sections are encountered when building an ELF object.

Fixes: https://github.com/llvm/llvm-project/issues/60448

Differential Revision: https://reviews.llvm.org/D143508

20 months agoReland [gn build] Support linux/win compiler-rt cross compilation
Arthur Eubanks [Wed, 8 Feb 2023 02:11:35 +0000 (18:11 -0800)]
Reland [gn build] Support linux/win compiler-rt cross compilation

Differential Revision: https://reviews.llvm.org/D143545

20 months agoRevert "[gn build] Support linux/win compiler-rt cross compilation"
Arthur Eubanks [Fri, 17 Feb 2023 17:56:51 +0000 (09:56 -0800)]
Revert "[gn build] Support linux/win compiler-rt cross compilation"

This reverts commit d73eb92cc673acf532b396537d78fe041dc9a9c2.

Breaks win/mac, e.g. http://45.33.8.238/win/75236/step_4.txt

20 months ago[lib++][format] Uses the new exception test macros.
Mark de Wever [Sat, 21 Jan 2023 12:35:30 +0000 (13:35 +0100)]
[lib++][format] Uses the new exception test macros.

Reviewed By: #libc, ldionne

Differential Revision: https://reviews.llvm.org/D143393

20 months ago[RISCV][MC] Add support for experimental zfa extension (FLI instruction not included...
Jun Sha (Joshua) [Fri, 17 Feb 2023 17:02:15 +0000 (09:02 -0800)]
[RISCV][MC] Add support for experimental zfa extension (FLI instruction not included) (try 3)

(Opening comment by preames - submitter, not patch author)

Try 1 was reverted because it had been landed in pieces, none of which had a meaningful commit message.

Try 2 was reverted because of a buildbot failure.  It turns out this was a spurious revert on my (preames) part.  The buildbots were failing, and it was from one of my changes in this sequence, but not the recommit.  I had reverted both of the try 1 commits in a single git push, but in separate *commits*.  Some of the builders decided to build the broken state between the original try 1 commit, and I'd not noticed that.  I'd assumed the breakage was because of the try2 re-commit.  So the revert of the try 2 recommit was spurious.

Original commit message:

This implements experimental support for the RISCV Zfa extension as specified here: https://github.com/riscv/riscv-isa-manual/releases/download/draft-20221119-5234c63/riscv-spec.pdf, Ch. 25. This extension has not been ratified. Once ratified, it'll move out of experimental status.

This change adds assembly support for all instructions except load-immediate instructions (fli.s/fli.d/fli.h).  Assembly support for that instruction and codegen support will follow in separate patches.

Differential Revision: https://reviews.llvm.org/D141984

20 months ago[gn build] Support linux/win compiler-rt cross compilation
Arthur Eubanks [Wed, 8 Feb 2023 02:11:35 +0000 (18:11 -0800)]
[gn build] Support linux/win compiler-rt cross compilation

Differential Revision: https://reviews.llvm.org/D143545

20 months ago[M68k] fix test regression introduced by D140180
Nick Desaulniers [Fri, 17 Feb 2023 17:19:36 +0000 (09:19 -0800)]
[M68k] fix test regression introduced by D140180

I added a new pass, callbrprepare, to the pass pipelines in
commit a3a84c9e2511 ("[llvm] add CallBrPrepare pass to pipelines")
but did not test experimental backends.

20 months ago[libc++] Fixes a flaky test.
Mark de Wever [Sat, 11 Feb 2023 15:45:35 +0000 (16:45 +0100)]
[libc++] Fixes a flaky test.

While investigating the flaky tests on the mac backdeployment targets
it seems the tests are different than suggested in the LWG issue.
That tests "does work", grabs the mutex, marks the task as done, and
finally calls the notifier.
Our test emulated "does work" after the notification, effectively
sleeping with a lock held. This has been fixed.

A second improvement is that the test fails when, due to OS
scheduling, the condition variable is not used in the main thread.

During discussing the approach of the patch with @ldionne, he
noticed the real reason why the test is flaky; the Apple
backdeployment targets use a dylib, which does not contain the
fix. So the test can't be tested on that platform; it only
proves the LWG issue was correct and the Standard contained
a bug.

Reviewed By: ldionne, #libc

Differential Revision: https://reviews.llvm.org/D143816

20 months ago[gn build] Support building x86/64 Android libraries
Arthur Eubanks [Wed, 8 Feb 2023 19:05:38 +0000 (11:05 -0800)]
[gn build] Support building x86/64 Android libraries

Reviewed By: hans

Differential Revision: https://reviews.llvm.org/D143598

20 months agoPort PlaceSafepoints pass to the new pass manager
Jan Dupej [Thu, 16 Feb 2023 20:37:16 +0000 (12:37 -0800)]
Port PlaceSafepoints pass to the new pass manager

This patch ports the PlaceSafepoints pass to the new pass manager as it is used by .NET/Mono. Compatibility with the legacy pass manager is maintained by adding PlaceSafepointsLegacyPass. This pass also depends on PlaceBackedgeSafepointsLegacyPass, which has been kept in the legacy-only variant, since it is apparently used only from PlaceSafepointsPass. It has been renamed, though, to indicate its legacy interface.

Reviewed By: aeubanks

Differential Revision: https://reviews.llvm.org/D136163

20 months ago[flang][runtime] Support DT edit descriptor in result of READ(SIZE=)
Peter Klausler [Fri, 17 Feb 2023 00:30:38 +0000 (16:30 -0800)]
[flang][runtime] Support DT edit descriptor in result of READ(SIZE=)

When a formatted I/O READ statement processes a DT edit descriptor to call a
user-defined I/O subroutine to read a derived type data item, all of the
characters that that subroutine reads via child I/O count as charecters
read by an edit descriptor and should accumulate in the result returned
by a SIZE= item in the original READ statement's control list.

Differential Revision: https://reviews.llvm.org/D144234

20 months ago[libc++][CI] Switches to clang-format-16.
Mark de Wever [Wed, 15 Feb 2023 20:00:06 +0000 (21:00 +0100)]
[libc++][CI] Switches to clang-format-16.

Uses an absolute path to the selected binary.
Updates the formatting of two files to match clang-format-16 style.

Depends on D144126

Reviewed By: #libc, philnik

Differential Revision: https://reviews.llvm.org/D144132

20 months ago[libc++] Improve the format ignorelist generation.
Mark de Wever [Wed, 15 Feb 2023 18:50:55 +0000 (19:50 +0100)]
[libc++] Improve the format ignorelist generation.

Several improvements
- Only add files that we actually want to format.
- Sort according to a fixed locale.

Some drive-by fixes
- Rename a text file, this avoids a filter exception.
- Adds a some missing source files extensions.
- Removes unused extensions hh, hxx, cc, and cxx from clang-format.

Reviewed By: philnik, #libc

Differential Revision: https://reviews.llvm.org/D144126

20 months ago[RISCV] Add preferred function and loop alignment RISCVSubtarget. NFC
Craig Topper [Fri, 17 Feb 2023 16:43:40 +0000 (08:43 -0800)]
[RISCV] Add preferred function and loop alignment RISCVSubtarget. NFC

These seem like properties we will want to adjust based on -mtune.
Move them to subtarget like is done on ARM and AArch64. Don't add
any overrides yet.

Note there's a slight change here. We are now passing Align(1) for
preferred function alignment where we previously passed the minimum
alignment. As far as I could tell, it will be maxed with min when
it used so this should be ok.

Reviewed By: reames

Differential Revision: https://reviews.llvm.org/D144048

20 months ago[libc++][NFC] Rename _LIBCPP_NO_EXCEPTIONS to _LIBCPP_HAS_NO_EXCEPTIONS
Nikolas Klauser [Thu, 2 Feb 2023 10:47:01 +0000 (11:47 +0100)]
[libc++][NFC] Rename _LIBCPP_NO_EXCEPTIONS to _LIBCPP_HAS_NO_EXCEPTIONS

Other macros that disable parts of the library are named `_LIBCPP_HAS_NO_WHATEVER`.

Reviewed By: ldionne, Mordante, #libc

Spies: libcxx-commits, smeenai

Differential Revision: https://reviews.llvm.org/D143163

20 months ago[ADT] Add `at` method (assertive lookup) to DenseMap and StringMap
Ryan Guo [Tue, 14 Feb 2023 04:06:44 +0000 (20:06 -0800)]
[ADT] Add `at` method (assertive lookup) to DenseMap and StringMap

This patch makes it easier for users when they want to use validated
lookup on DenseMap/StringMap as a composable C++ expression. For
instance:

```
// instead of
if (auto val = map.lookup(key))
   return val;
assert("...");

// we can write
return map.at(key);
```

Differential Revision: https://reviews.llvm.org/D143976

20 months ago[clang-tidy] Clarify bugprone-branch-clone diagnostic message
Donát Nagy [Mon, 13 Feb 2023 14:35:24 +0000 (15:35 +0100)]
[clang-tidy] Clarify bugprone-branch-clone diagnostic message

This simple commit inserts "body" into the message "repeated branch _body_ in
conditional chain". This is motivated by feedback from a user who (at first
glance) thought that clang-tidy complained about a repeated branch _condition_.

Differential Revision: https://reviews.llvm.org/D143917

20 months ago[libc][bazel] Export .bzl files for libc/test and libc/test/src/math.
Tue Ly [Fri, 17 Feb 2023 16:02:22 +0000 (11:02 -0500)]
[libc][bazel] Export .bzl files for libc/test and libc/test/src/math.

20 months ago[libc++][test] Adds more generic test macros.
Mark de Wever [Sat, 21 Jan 2023 12:35:30 +0000 (13:35 +0100)]
[libc++][test] Adds more generic test macros.

These macros are intended to replace the macros in rapid-cxx-test.h.

Reviewed By: #libc, ldionne

Differential Revision: https://reviews.llvm.org/D142808

20 months ago[gn build] Port bf0f94a5cf82
LLVM GN Syncbot [Fri, 17 Feb 2023 15:40:33 +0000 (15:40 +0000)]
[gn build] Port bf0f94a5cf82

20 months ago[mlir] add option to multi-buffering
Thomas Raoux [Fri, 17 Feb 2023 15:05:23 +0000 (15:05 +0000)]
[mlir] add option to multi-buffering

Allow user to apply multi-buffering transformation for cases where proving
that there is no loop carried dependency is not trivial. In this case user needs
to ensure that the data are written and read in the same iteration otherwise the
result is incorrect.

Differential Revision: https://reviews.llvm.org/D144227

20 months ago[InstCombine] create a pass options container and add "use-loop-info" argument
Sanjay Patel [Fri, 17 Feb 2023 15:11:40 +0000 (10:11 -0500)]
[InstCombine] create a pass options container and add "use-loop-info" argument

This is a cleanup/modernization patch requested in D144045 to make loop
analysis a proper optional parameter to the pass rather than a
semi-arbitrary value inherited via the pass pipeline.

It's a bit more complicated than the recent patch I started copying from
(D143980) because InstCombine already has an option for MaxIterations
(added with D71145).

I debated just deleting that option, but it was used by a pair of existing
tests, so I put it into a struct (code largely copied from SimplifyCFG's
implementation) to make the code more flexible for future options
enhancements.

I didn't alter the pass manager invocations of InstCombine in this patch
because the patch was already getting big, but that will be a small
follow-up as noted with the TODO comment.

Differential Revision: https://reviews.llvm.org/D144199

20 months ago[SLP]Do not reduce repeated values, use scalar red ops instead.
Alexey Bataev [Fri, 17 Jun 2022 17:15:23 +0000 (10:15 -0700)]
[SLP]Do not reduce repeated values, use scalar red ops instead.

Metric: size..text

                                                     size..text                 results     results0    diff
SingleSource/Regression/C/gcc-c-torture/execute/GCC-C-execute-980605-1.test      445.00      461.00  3.6%
SingleSource/Benchmarks/Adobe-C++/loop_unroll.test                               428477.00   428445.00 -0.0%
External/SPEC/CFP2006/447.dealII/447.dealII.test                                 618849.00   618785.00 -0.0%

For all tests some extra code was optimized, GCC-C-execute has some more
inlining after

Differential Revision: https://reviews.llvm.org/D132261

20 months agoNew SetOperations and unittesting for all SetOperations
Teresa Johnson [Thu, 16 Feb 2023 22:37:07 +0000 (14:37 -0800)]
New SetOperations and unittesting for all SetOperations

New set operations split out of D140908 as suggested, and I have added
unit testing for all set operations.

This adds a set_intersection, which returns the intersection instead of
updating the first set like set_intersect (using a different name
analogous to set_difference vs set_subtract).

Also adds a variant of set_subtract that updates two additional set
arguments to note which members of the subtrahend were removed from
the minuend and which were not.

Differential Revision: https://reviews.llvm.org/D144220

20 months agoTighten up a modules test
Paul Robinson [Thu, 16 Feb 2023 15:20:42 +0000 (07:20 -0800)]
Tighten up a modules test

Differential Revision: https://reviews.llvm.org/D144188

20 months ago[mlir][llvm] Fuse access_group & loop export (NFC)
Christian Ulmann [Fri, 17 Feb 2023 14:31:06 +0000 (15:31 +0100)]
[mlir][llvm] Fuse access_group & loop export (NFC)

This commit moves the access group translation into the
LoopAnnotationTranslation class as these two metadata kinds only appear
together.

Drops the access group cleanup from `ModuleTranslation::forgetMapping`
as this is only used on function regions. Access groups only appear in the
region of a global metadata operation and will thus not be cleaned here.

Analogous to https://reviews.llvm.org/D143577

Reviewed By: gysit

Differential Revision: https://reviews.llvm.org/D144253

20 months ago[SVE] Add intrinsics for floating-point operations that explicitly undefine the resul...
Paul Walker [Fri, 10 Feb 2023 12:25:26 +0000 (12:25 +0000)]
[SVE] Add intrinsics for floating-point operations that explicitly undefine the result for inactive lanes.

This patch is the floating-point equivalent of D141937.

Depends on D143764.

Differential Revision: https://reviews.llvm.org/D143765

20 months ago[mlir][standalone] Extend sample with custom type
Marius Brehler [Fri, 17 Feb 2023 13:51:04 +0000 (14:51 +0100)]
[mlir][standalone] Extend sample with custom type

This extends the standalone example to illustrate how to structure the
files needed to create own types.

Reviewed By: jpienaar

Differential Revision: https://reviews.llvm.org/D143715

20 months ago[clang][CodeGen] Use base subobject type layout for potentially-overlapping fields
Vladislav Dzhidzhoev [Fri, 13 Jan 2023 23:38:10 +0000 (02:38 +0300)]
[clang][CodeGen] Use base subobject type layout for potentially-overlapping fields

RecordLayoutBuilder assumes the size of a potentially-overlapping
class/struct field with non-zero size as the size of the base subobject
type corresponding to the field type.
Make CGRecordLayoutBuilder to acknowledge that in order to avoid incorrect
padding insertion.

Differential Revision: https://reviews.llvm.org/D139741

20 months agoGH60642: Fix ICE when checking a lambda defined in a concept definition
Erich Keane [Thu, 16 Feb 2023 15:50:56 +0000 (07:50 -0800)]
GH60642: Fix ICE when checking a lambda defined in a concept definition

As reported in GH60642, we asserted when there was a lambda defined in a
template arguments inside of a concept, which caused us to not properly
set up the list of instantiation args.  This patch ensures that the
'lambda context decl' correctly falls-through the template argument
instantiation, so that it is available when instantiating the lambda,
and thus, when setting up the lambda instantiation args list.

20 months ago[Clang] Convert some tests to opaque pointers (NFC)
Nikita Popov [Fri, 17 Feb 2023 13:34:05 +0000 (14:34 +0100)]
[Clang] Convert some tests to opaque pointers (NFC)

20 months ago[Clang] Convert some tests to opaque pointers (NFC)
Nikita Popov [Fri, 17 Feb 2023 11:12:07 +0000 (12:12 +0100)]
[Clang] Convert some tests to opaque pointers (NFC)

20 months agoFix Clang sphinx build
Aaron Ballman [Fri, 17 Feb 2023 12:56:52 +0000 (07:56 -0500)]
Fix Clang sphinx build

This addresses issues found in:
https://lab.llvm.org/buildbot/#/builders/92/builds/40269

20 months ago[RISCV][NFC] Add missing immediate operand types.
Dmitry Bushev [Fri, 17 Feb 2023 12:03:40 +0000 (15:03 +0300)]
[RISCV][NFC] Add missing immediate operand types.

Some immediate types in RISCV target description lack operand type field.
This leads them being listed as OPERAND_UNKNOWN in MCOperandInfo. This patch adds this fields.
This is NFC because it does not affect flow of any current tools implementation.

Reviewed By: craig.topper

Differential Revision: https://reviews.llvm.org/D144105

20 months ago[mlir][llvm] Add DINamespace attribute
Christian Ulmann [Fri, 17 Feb 2023 12:09:17 +0000 (13:09 +0100)]
[mlir][llvm] Add DINamespace attribute

This commit introduces the DINamespaceAttr to model LLVM's DINamespace metadata.

Reviewed By: gysit

Differential Revision: https://reviews.llvm.org/D144243

20 months ago[mlir][llvm] Make DI param optional to match LLVM
Christian Ulmann [Fri, 17 Feb 2023 12:04:51 +0000 (13:04 +0100)]
[mlir][llvm] Make DI param optional to match LLVM

This commit makes DIDerivedTypeAttr's baseType parameter optional to ensure imported IR
doesn't break the printer.

Reviewed By: gysit

Differential Revision: https://reviews.llvm.org/D144247

20 months ago[gn build] Port 46db8d822ecd
LLVM GN Syncbot [Fri, 17 Feb 2023 11:44:43 +0000 (11:44 +0000)]
[gn build] Port 46db8d822ecd

20 months ago[libc++] Granularize <atomic>
Nikolas Klauser [Sat, 14 Jan 2023 01:53:52 +0000 (02:53 +0100)]
[libc++] Granularize <atomic>

Reviewed By: Mordante, #libc

Spies: arichardson, libcxx-commits, krytarowski

Differential Revision: https://reviews.llvm.org/D142972

20 months ago[AMDGPU] Simplify widenScalar condition for BigTy for G_(UN)MERGE_VALUES
Jay Foad [Fri, 17 Feb 2023 10:18:38 +0000 (10:18 +0000)]
[AMDGPU] Simplify widenScalar condition for BigTy for G_(UN)MERGE_VALUES

Differential Revision: https://reviews.llvm.org/D144250

20 months ago[Clang] Convert some tests to opaque pointers (NFC)
Nikita Popov [Fri, 17 Feb 2023 10:27:36 +0000 (11:27 +0100)]
[Clang] Convert some tests to opaque pointers (NFC)

20 months ago[analyzer] Remove unjustified assert from EquivalenceClass::simplify
Balazs Benics [Fri, 17 Feb 2023 10:37:02 +0000 (11:37 +0100)]
[analyzer] Remove unjustified assert from EquivalenceClass::simplify

One might think that by merging the equivalence classes (eqclasses) of
`Sym1` and `Sym2` symbols we would end up with a `State` in which the
eqclass of `Sym1` and `Sym2` should now be the same. Surprisingly, it's
not //always// true.

Such an example triggered the assertion enforcing this _unjustified_
invariant in https://github.com/llvm/llvm-project/issues/58677.
```lang=C++
unsigned a, b;
#define assert(cond) if (!(cond)) return

void f(unsigned c) {
    /*(1)*/ assert(c == b);
    /*(2)*/ assert((c | a) != a);
    /*(3)*/ assert(a);
    // a = 0  =>  c | 0 != 0  =>  c != 0  =>  b != 0
}
```

I believe, that this assertion hold for reasonable cases - where both
`MemberSym` and `SimplifiedMemberSym` refer to live symbols.
It can only fail if `SimplifiedMemberSym` refers to an already dead
symbol. See the reasoning at the very end.
In this context, I don't know any way of determining if a symbol is
alive/dead, so I cannot refine the assertion in that way.
So, I'm proposing to drop this unjustified assertion.

---

Let me elaborate on why I think the assertion is wrong in its current
shape.

Here is a quick reminder about equivalence classes in CSA.
We have 4 mappings:
1) `ClassMap`: map, associating `Symbols` with an `EquivalenceClass`.
2) `ClassMembers`: map, associating `EquivalenceClasses` with its
   members - basically an enumeration of the `Symbols` which are known
   to be equal.
3) `ConstraintRange`: map, associating `EquivalenceClasses` with the
   range constraint which should hold for all the members of the
   eqclass.
4) `DisequalityMap`: I'm omitting this, as it's irrelevant for our
   purposes now.

Whenever we encounter/assume that two `SymbolRefs` are equal, we update
the `ClassMap` so that now both `SymbolRefs` are referring to the same
eqclass. This operation is known as `merge` or `union`.
Each eqclass is uniquely identified by the `representative` symbol, but
it could have been just a unique number or anything else - the point
is that an eqclass is identified by something unique.
Initially, all Symbols form - by itself - a trivial eqclass, as there
are no other Symbols to which it is assumed to be equal. A trivial
eqclass has //notionally// exactly one member, the representative symbol.
I'm emphasizing that //notionally// because for such cases we don't store
an entry in the `ClassMap` nor in the `ClassMembers` map, because it
could be deduced.

By `merging` two eqclasses, we essentially do what you would think it
does. An important thing to highlight is that the representative symbol
of the resulting eqclass will be the same as one of the two eqclasses.
This operation should be commutative, so that `merge(eq1,eq2)` and
`merge(eq2,eq1)` should result in the same eqclass - except for the
representative symbol, which is just a unique identifier of the class,
a name if you will. Using the representative symbol of `eq1` or `eq2`
should have no visible effect on the analysis overall.

When merging `eq1` into `eq2`, we take each of the `ClassMembers` of
`eq1` and add them to the `ClassMembers` of `eq2` while we also redirect
all the `Symbol` members of `eq1` to map to the `eq2` eqclass in the
`ClassMap`. This way all members of `eq1` will refer to the correct
eqclass. After these, `eq1` key is unreachable in the `ClassMembers`,
hence we can drop it.

---

Let's get back to the example.
Note that when I refer to symbols `a`, `b`, `c`, I'm referring to the
`SymbolRegionValue{VarRegion{.}}` - the value of that variable.

After `(1)`, we will have a constraint `c == b : [1,1]` and an eqclass
`c,b` with the `c` representative symbol.
After `(2)`, we will have an additional constraint `c|b != a : [1,1]`
with the same eqclass. We will also have disequality info about that
`c|a` is disequal to `a` - and the other way around.
However, after the full-expression, `c` will become dead. This kicks in
the garbage collection, which transforms the state into this:
 - We no longer have any constraints, because only `a` is alive, for
   which we don't have any constraints.
 - We have a single (non-trivial) eqclass with a single element `b` and
   representative symbol `c`. (Dead symbols can be representative
   symbols.)
 - We have the same disequality info as before.

At `(3)` within the false branch, `a` get perfectly constrained to zero.
This kicks in the simplification, so we try to substitute (simplify) the
variable in each SymExpr-tree. In our case, it means that the
`c|a != a : [1,1]` constraint gets re-evaluated as `c|0 != 0 : [1,1]`,
which is `c != 0 : [1,1]`.
Under the hood, it means that we will call `merge(c|a, c)`. where `c` is
the result of `simplifyToSVal(State, MemberSym).getAsSymbol()` inside
`EquivalenceClass::simplify()`.
Note that the result of `simplifyToSVal()` was a dead symbol. We
shouldn't have acquired an already dead symbol. AFAIK, this is the only
way we can get one at this point.
Since `c` is dead, we no longer have a mapping in `ClassMap` for it;
hence when we try to `find()` the eqclass of `c`, it will report that
it's a trivial eqclass with the representative symbol `c`.

After `merge(c|a, c)`, we will have a single (non-trivial) eqclass of
`b, c|a` with the representative symbol `c|a` - because we merged the
eqclass of `c` into the eqclass of `c|a`.

This means that `find(c|a)` will result in the eqclass with the
representative symbol `c|a`. So, we ended up having different eqclasses
for `find(c|a)` and `find(c)` after `merge(c|a, c)`, firing the
assertion.

I believe, that this assertion hold for reasonable cases - where both
`MemberSym` and `SimplifiedMemberSym` refer to live symbols.
`MemberSym` should be live in all cases here, because it is from
`ClassMembers` which is pruned in `removeDeadBindings()` so these must
be alive. In this context, I don't know any way of determining if a
symbol is alive/dead, so I cannot refine the assertion in that way.
So, I'm proposing to drop this unjustified assertion.

I'd like to thank @martong for helping me to conclude the investigation.
It was really difficult to track down.

PS: I mentioned that `merge(eq1, eq2)` should be commutative. We
actually exploit this for merging the smaller eqclass into the bigger
one within `EquivalenceClass::merge()`.
Yea, for some reason, if you swap the operands, 3 tests break (only
failures, no crashes) aside from the test which checks the state dumps.
But I believe, that is a different bug and orthogonal to this one. I
just wanted to mention that.
- `Analysis/solver-sym-simplification-adjustment.c`
- `Analysis/symbol-simplification-fixpoint-iteration-unreachable-code.cpp`
- `Analysis/symbol-simplification-reassume.cpp`

Fixes #58677

Reviewed By: vabridgers

Differential Revision: https://reviews.llvm.org/D138037

20 months ago[libc++] Granularize <bit> includes
Nikolas Klauser [Sun, 8 Jan 2023 14:26:32 +0000 (15:26 +0100)]
[libc++] Granularize <bit> includes

Reviewed By: ldionne, #libc

Spies: libcxx-commits

Differential Revision: https://reviews.llvm.org/D141228

20 months ago[MachineCombiner] Support local strategy for traces
Anton Sidorenko [Tue, 14 Feb 2023 10:32:10 +0000 (13:32 +0300)]
[MachineCombiner] Support local strategy for traces

For in-order cores MachineCombiner makes better decisions when the critical path
is calculated only for the current basic block and does not take into account
other blocks from the trace.

This patch adds a virtual method to TargetInstrInfo to allow each target decide
which strategy to use.

Depends on D140541

Reviewed By: spatel

Differential Revision: https://reviews.llvm.org/D140542

20 months ago[Clang] Convert some tests to opaque pointers (NFC)
Nikita Popov [Fri, 17 Feb 2023 09:01:18 +0000 (10:01 +0100)]
[Clang] Convert some tests to opaque pointers (NFC)

20 months ago[mlir] Add a folder for lbs, ubs, steps of scf.forall.
Alexander Belyaev [Fri, 17 Feb 2023 09:57:44 +0000 (10:57 +0100)]
[mlir] Add a folder for lbs, ubs, steps of scf.forall.

Differential Revision: https://reviews.llvm.org/D144245

20 months agoRevert "[LLD] [COFF] Don't try to detect MSVC installations in mingw mode"
Martin Storsjö [Fri, 17 Feb 2023 09:25:12 +0000 (11:25 +0200)]
Revert "[LLD] [COFF] Don't try to detect MSVC installations in mingw mode"

This reverts commit 389bfbd66d6f78b5fc60e51e620e7f767fc867f0.

This commit broke the compiler-rt/test/profile/Windows/coverage-weak-lld.cpp
testcase. This testcase builds and links in an MSVC environment,
but explicitly passes -lldmingw to the linker, to opt in to
certain mingw-style behaviours regarding weak symbols, since
effb87dfa810a28e763f914fe3e6e984782cc846. This patch broke that
test, which now errored out with these errors:

    lld-link: error: could not open 'libuuid.a': no such file or directory
    lld-link: error: could not open 'libLIBCMT.a': no such file or directory
    lld-link: error: could not open 'libOLDNAMES.a': no such file or directory

20 months ago[mlir][llvm] Verify LLVM module before import
Christian Ulmann [Fri, 17 Feb 2023 08:47:37 +0000 (09:47 +0100)]
[mlir][llvm] Verify LLVM module before import

This commit ensures that the importing of LLVM modules first verifies
that the module is even valid. As many tests did not work with valid
LLVM IR, they were fixed as part of this commit.

Some error messages were only reachable with invalid input IR, thus they
were replaced with a failures.

Reviewed By: gysit

Differential Revision: https://reviews.llvm.org/D144186

20 months ago[JT] Always create BPI/BFI when running in legacy PM
Benjamin Kramer [Fri, 17 Feb 2023 09:09:34 +0000 (10:09 +0100)]
[JT] Always create BPI/BFI when running in legacy PM

This is wasteful, but only affects the legacy pass manager. Otherwise
a1b78fb929fccf96acaa0212cf68fee82298e747 would crash JT when running
with that PM. There are still a few users of the legacy PM out there
that are reluctant to migrate, numba in this case.

No test as we don't test legacy PM anymore.

20 months ago[clang-repl] Support compound statement as a top-level statement.
Vassil Vassilev [Sun, 11 Dec 2022 21:08:30 +0000 (21:08 +0000)]
[clang-repl] Support compound statement as a top-level statement.

This patch teaches our incremental compilation infrastructure to push and pop a
fake function scope making the Parser happy when parsing compound statements as
part of a top-leve statement declaration.

Differential revision: https://reviews.llvm.org/D139798

20 months ago[mlir][scf] Rename ForeachThreadOp->ForallOp, PerformConcurrentlyOp->InParallelOp.
Alexander Belyaev [Fri, 17 Feb 2023 08:24:33 +0000 (09:24 +0100)]
[mlir][scf] Rename ForeachThreadOp->ForallOp, PerformConcurrentlyOp->InParallelOp.

Differential Revision: https://reviews.llvm.org/D144242

20 months ago[ConstantRange] Handle `Intrinsic::ctlz`
Antonio Frighetto [Fri, 17 Feb 2023 08:50:49 +0000 (09:50 +0100)]
[ConstantRange] Handle `Intrinsic::ctlz`

Introduce ConstantRange support for ctlz intrinsic, including
exhaustive testing. Among other things, LVI may now be able to
propagate information about cltz constant ranges lattice values.

Differential Revision: https://reviews.llvm.org/D142234

20 months ago[Support] [Windows] Don't check file access time in equivalent(file_status, file_status)
R. Voggenauer [Tue, 14 Feb 2023 14:21:06 +0000 (15:21 +0100)]
[Support] [Windows] Don't check file access time in equivalent(file_status, file_status)

The sys::fs::equivalent(file_status, file_status) function is meant to
judge whether two file_status structs denote the same individual file.
On Unix, this is implemented by only comparing the st_dev and st_ino
numbers (from stat), ignoring all other fields.

On Windows, lacking reliable fields corresponding to st_dev and st_ino,
equivalent(file_status, file_status) compares essentially all fields.
However, since 1e39ef331b2b78baec84fdb577d497511cc46bd5
(https://reviews.llvm.org/D18456), file_status also contains the file
access time.

Including the file access time in equivalent(file_status, file_status)
makes it possible to spuriously break. In particular, when invoking
equivalent(Twine, Twine), with two paths, there's a race condition - the
function calls status() on both input paths. Even if the two paths
are the same, the comparison can fail, if the file was accessed
between the two status() calls.

Thus, it seems like the inclusion of the access time in
equivalent(file_status, file_status) was a mistake.

This race condition can cause spurious failures when linking with
LLD/ELF, where LLD uses equivalent() to check whether a file
exists within a specific sysroot, and that sysroot is accessed by other
processes concurrently.

This fixes downstream issue
https://github.com/msys2/MINGW-packages/issues/15695.

Differential Revision: https://reviews.llvm.org/D144172

20 months ago[LLD] [COFF] Don't try to detect MSVC installations in mingw mode
Martin Storsjö [Wed, 15 Feb 2023 10:06:45 +0000 (12:06 +0200)]
[LLD] [COFF] Don't try to detect MSVC installations in mingw mode

In mingw mode, all linker paths are passed explicitly to the linker
by the compiler driver. Don't try to implicitly add linker paths
from the LIB environment variable or by detecting an MSVC
installation directory.

If the /winsysroot command line parameter is explicitly passed to
lld-link while /lldmingw is specified, it could be considered reasonable
to actually include those paths. However, modifying the code to
handle only the /winsysroot case but not the other ones, when the
mingw mode has been enabled, seems like much more code complexity
for a mostly hypothetical case.

Add a test for this when case when using LIB. (The code paths for
trying to detect an MSVC installation aren't really regression tested.)

Also fix an issue in the existing test for "Check that when /winsysroot
is specified, %LIB% is ignored.", where the LIB variable pointed
to a nonexistent directory, so the test would pass even if /winsysroot
wouldn't be specified.

Differential Revision: https://reviews.llvm.org/D144084

20 months ago[Clang] Convert some tests to opaque pointers (NFC)
Nikita Popov [Fri, 17 Feb 2023 08:29:41 +0000 (09:29 +0100)]
[Clang] Convert some tests to opaque pointers (NFC)

20 months ago[mlir][llvm] Add atomic support to the LoadOp.
Tobias Gysi [Fri, 17 Feb 2023 08:20:08 +0000 (09:20 +0100)]
[mlir][llvm] Add atomic support to the LoadOp.

This revision adds atomic support to the LoadOp. It chooses
to print the atomic keywords together with the syncscope and
ordering arguments, which simplifies parsing and printing compared
to the LLVM IR printer that puts the atomic keyword at the beginning.
It uses the ordering attribute to check if the load is atomic.

The revision also implements verifiers to ensure the constraints
that apply to atomic load operations are checked.

Reviewed By: Dinistro

Differential Revision: https://reviews.llvm.org/D144112

20 months ago[Clang] Convert some tests to opaque pointers (NFC)
Nikita Popov [Thu, 16 Feb 2023 16:06:50 +0000 (17:06 +0100)]
[Clang] Convert some tests to opaque pointers (NFC)

20 months ago[AMDGPU] Add cross-project-tests for WMMA builtins
Diana Picus [Thu, 16 Feb 2023 10:10:54 +0000 (11:10 +0100)]
[AMDGPU] Add cross-project-tests for WMMA builtins

Add a few tests to make sure we get the expected instruction for the
WMMA builtins (and generally that our builtins and intrinsics are on the
same page and won't blow up).

Differential Revision: https://reviews.llvm.org/D144176

20 months ago[MLIR][LLVM] Only disallow inlining for selected function attributes.
Johannes de Fine Licht [Fri, 17 Feb 2023 08:04:43 +0000 (09:04 +0100)]
[MLIR][LLVM] Only disallow inlining for selected function attributes.

This loosens the requirement of no passthrough function attribute being
present to checking for specific attributes that prevent inlining. Since
these attributes are no longer strictly passthrough, they should
eventually be upgraded to some form of addressable attributes.

Drops the expensive StringSwitches over call and function attributes in
favor of selectively disallowing attributes that prevent inlining
(similiar to the LLVM inliner), thereby moving to a less conservative
approach.

Reviewed By: gysit

Differential Revision: https://reviews.llvm.org/D144104

20 months ago[Clang] Convert update_cc_test_checks tests to opaque pointers (NFC)
Nikita Popov [Fri, 17 Feb 2023 08:06:02 +0000 (09:06 +0100)]
[Clang] Convert update_cc_test_checks tests to opaque pointers (NFC)

20 months ago[DAGCombiner] Teach MatchContextClass classes to use TargetLowering::isOperationLegal...
Yeting Kuo [Wed, 15 Feb 2023 07:17:26 +0000 (15:17 +0800)]
[DAGCombiner] Teach MatchContextClass classes to use TargetLowering::isOperationLegalOrCustom().

Some of TargetLowering functions needed opcodes are often used in DAGCombiner.
The patch make those MatchContextClass classes have TargetLowering members and
pass specific opcodes for those TargetLowering functions.

Reviewed By: RKSimon

Differential Revision: https://reviews.llvm.org/D144075

20 months ago[mlir] Add loop bounds to scf.foreach_thread.
Alexander Belyaev [Fri, 17 Feb 2023 07:54:56 +0000 (08:54 +0100)]
[mlir] Add loop bounds to scf.foreach_thread.

https://discourse.llvm.org/t/rfc-parallel-loops-on-tensors-in-mlir/68332

Differential Revision: https://reviews.llvm.org/D144072

20 months agoRevert "[Coroutines] Stop supportting std::experimental::coroutine_traits"
Chuanqi Xu [Fri, 17 Feb 2023 07:50:54 +0000 (15:50 +0800)]
Revert "[Coroutines] Stop supportting std::experimental::coroutine_traits"

This reverts commit c4e6e771f255fb1da3d505534997b6a88195b012.

Since clang-tools-extra contains the use for
std::experimental::coroutine_traits, the previsou commit breaks the
build bot. Revert this one to make the bot green.

20 months ago[Coroutines] Stop supportting std::experimental::coroutine_traits
Chuanqi Xu [Fri, 17 Feb 2023 07:22:04 +0000 (15:22 +0800)]
[Coroutines] Stop supportting std::experimental::coroutine_traits

As we discussed before, we should stop supporting
std::experimental::coroutine_traits in clang17. Now the clang16 is
branched so we can clean them now.

All the removed tests have been duplicated before.

20 months ago[NDF] Revert earlier nfc commit to test commit access
Ryan Guo [Fri, 17 Feb 2023 06:57:23 +0000 (22:57 -0800)]
[NDF] Revert earlier nfc commit to test commit access

This patch reverts 511d55060454 by removing the empty newline.

20 months ago[NFC] Testing new commit access by adding newline
Ryan Guo [Fri, 17 Feb 2023 06:52:31 +0000 (22:52 -0800)]
[NFC] Testing new commit access by adding newline

20 months ago[lldb] Add missing decorators import in TestPoPersistentResult.py
Dave Lee [Fri, 17 Feb 2023 05:45:24 +0000 (21:45 -0800)]
[lldb] Add missing decorators import in TestPoPersistentResult.py

20 months ago[lldb] Limit TestPoPersistentResult to darwin
Dave Lee [Fri, 17 Feb 2023 05:11:38 +0000 (21:11 -0800)]
[lldb] Limit TestPoPersistentResult to darwin

20 months ago[lldb] Suppress persistent result when running po
Dave Lee [Tue, 14 Feb 2023 21:48:07 +0000 (13:48 -0800)]
[lldb] Suppress persistent result when running po

Remove the persistent result variable after executing `po`.

Without this change, the following behavior happens:

```
(lldb) p thing
(NSObject *) $0 = 0x600000008000
(lldb) po thing
<NSObject: 0x600000008000>
(lldb) p thing
(NSObject *) $2 = 0x600000008000
(lldb) p $1
(NSObject *) $1 = 0x600000008000
```

Even though `po` hides the persistent result variable, it's still created - as $1 in
this example. It can be accessed even though its existence is not evident.

With this change, the persistent result is removed after the object description has
printed. Instead, this is the behavior:

```
(lldb) p thing
(NSObject *) $0 = 0x600000008000
(lldb) po thing
<NSObject: 0x600000008000>
(lldb) p thing
(NSObject *) $1 = 0x600000008000
```

The difference here is that the `po` doens't silently create a persistent result.

Differential Revision: https://reviews.llvm.org/D144044

20 months ago[lldb] Rename SetResultIsInternal to SetSuppressPersistentResult (NFC)
Dave Lee [Tue, 14 Feb 2023 21:39:56 +0000 (13:39 -0800)]
[lldb] Rename SetResultIsInternal to SetSuppressPersistentResult (NFC)

Rename `SetResultIsInternal` and `GetResultIsInternal` to `SetSuppressPersistentResult`
and `GetSuppressPersistentResult` respectively.  Also rename `m_result_is_internal`.

This matches the naming in the SB API.

A separate change calls `SetSuppressPersistentResult`, where the name
`SetResultIsInternal` doesn't quite fit.

Differential Revision: https://reviews.llvm.org/D144042

20 months ago[llvm][test] enable/disable -verify-machineinstrs where possible for callbr
Nick Desaulniers [Fri, 17 Feb 2023 04:25:34 +0000 (20:25 -0800)]
[llvm][test] enable/disable -verify-machineinstrs where possible for callbr

I introduced new tests in
commit 5cc1016a57b3 ("[llvm][SelectionDAGBuilder] codegen callbr.landingpad intrinsic")
https://reviews.llvm.org/D140160
that fails expensive checks. Disable -verify-machineinstrs in those
tests for now. Enable it in other tests for now, since MachineVerifier
isn't on by default for assertion builds.

Link: https://github.com/llvm/llvm-project/issues/60827
20 months ago[clang-format] Add a space between an overloaded operator and '>'
Owen Pan [Fri, 10 Feb 2023 16:50:49 +0000 (08:50 -0800)]
[clang-format] Add a space between an overloaded operator and '>'

The token annotator doesn't annotate the template opener and closer
as such if they enclose an overloaded operator. This causes the
space between the operator and the closer to be removed, resulting
in invalid C++ code.

Fixes #58602.

Differential Revision: https://reviews.llvm.org/D143755

20 months ago[ADT] Provide C++20-style bit functions
Kazu Hirata [Fri, 17 Feb 2023 04:13:04 +0000 (20:13 -0800)]
[ADT] Provide C++20-style bit functions

Tihs patches adds APInt::count{l,r}_{zero,one} and APInt::popcount to
be consistent with those functions in ADT/bit.h.

Once this patch lands, I'll take care of the migration.

For now, I am intentionally leaving isPowerOf2 as is.

Differential Revision: https://reviews.llvm.org/D144165

20 months ago[CodeGen] Fix warnings
Kazu Hirata [Fri, 17 Feb 2023 04:08:35 +0000 (20:08 -0800)]
[CodeGen] Fix warnings

This patch fixes:

  llvm/lib/CodeGen/CallBrPrepare.cpp:154:14: error: unused variable
  'IsDominated' [-Werror,-Wunused-variable]

  llvm/lib/CodeGen/CallBrPrepare.cpp:150:13: error: unused function
  'PrintDebugDomInfo' [-Werror,-Wunused-function]

20 months ago[mlir][gpu] NFC let user pick the threadID values when distributing foreach_thread
Thomas Raoux [Fri, 17 Feb 2023 00:15:12 +0000 (00:15 +0000)]
[mlir][gpu] NFC let user pick the threadID values when distributing foreach_thread

Reviewed By: nicolasvasilache

Differential Revision: https://reviews.llvm.org/D144219

20 months agoRevert "[mlir][vector] Prevent duplicating operations during vector distribute"
Thomas Raoux [Fri, 17 Feb 2023 03:06:59 +0000 (03:06 +0000)]
Revert "[mlir][vector] Prevent duplicating operations during vector distribute"

This reverts commit 2fc3c5c34c4c0ce94a217717a469620e06325fb0.

20 months ago[Sema] Relax a failing assertion in TransformBlockExpr
Akira Hatanaka [Tue, 14 Feb 2023 05:06:07 +0000 (21:06 -0800)]
[Sema] Relax a failing assertion in TransformBlockExpr

The assertion fails when the expression causing the this pointer to be
captured is part of a constexpr if statement's branch and the branch
gets discarded when the enclosing method is instantiated.

Note that the test case is added to CodeGen instead of Sema since the
translation unit has to be free of errors in order for the assertion to
be checked.

Differential Revision: https://reviews.llvm.org/D144016

20 months agoOpenMP: Regenerate test checks
Matt Arsenault [Thu, 16 Feb 2023 19:49:32 +0000 (15:49 -0400)]
OpenMP: Regenerate test checks

20 months agoRecommit "Remove incorrect comment around `truncateAVX512SetCCNoBWI`; NFC" (2nd Try)
Noah Goldstein [Thu, 16 Feb 2023 09:15:48 +0000 (03:15 -0600)]
Recommit "Remove incorrect comment around `truncateAVX512SetCCNoBWI`; NFC" (2nd Try)

No bug here, just needed to revert to revert
8bd0e9481cfcba53946433011d841280fd456caa which had a bug.

20 months agoRecommit "Transform vector SET{LE/ULT/ULE} -> SETLT and SET{GE/UGT/UGE} -> SETGT...
Noah Goldstein [Fri, 17 Feb 2023 01:41:11 +0000 (19:41 -0600)]
Recommit "Transform vector SET{LE/ULT/ULE} -> SETLT and SET{GE/UGT/UGE} -> SETGT if possible" (2nd Try)

Original version hit assert in `incDecVectorConstant` because VT could
be EVT (as opposed to MVT). Fix is to add check for VT.isSimple() in
`incDecVectorConstant`.

Reviewed By: saugustine

Differential Revision: https://reviews.llvm.org/D142254

20 months ago[lldb] Stop generating swig bindings for SBLaunchInfo copy constructor
Alex Langford [Thu, 16 Feb 2023 23:38:56 +0000 (15:38 -0800)]
[lldb] Stop generating swig bindings for SBLaunchInfo copy constructor

Given the line
```
launch_info = lldb.SBLaunchInfo(None)
```

We see different behaviors across different versionf of swig. On some
older versions of swig (e.g. 3.0.2) this line fails because it attempts
to use the copy constructor and blows up with an invalid null reference.
On newer versions of swig, this is correctly routed to the constructor
taking a pointer.

Prior to generating the swig bindings with the API headers,
SBLaunchInfo's copy constructor was not exposed so we're effectively
going back to the old behavior anyway.

Differential Revision: https://reviews.llvm.org/D144228

20 months ago[ThinLTO/WPD] Handle function alias in vtable correctly
Teresa Johnson [Thu, 16 Feb 2023 19:19:27 +0000 (11:19 -0800)]
[ThinLTO/WPD] Handle function alias in vtable correctly

We were not summarizing a function alias in the vtable, leading to
incorrect WPD in some cases, and missing WPD in others.

Specifically, we would end up ignoring function aliases as they aren't
summarized, so we could incorrectly devirtualize if there was a single
other non-alias function in a compatible vtable. And if there was only
one implementation, but it was an alias, we would not be able to
identify and perform the single implementation devirtualization.

Handling the alias summary correctly also required fixing the handling
in mustBeUnreachableFunction, so that it is not incorrectly ignored.

Regular LTO is conservatively correct because it will skip
devirtualizing when any pointer within a vtable is not a function.
However, it needs additional work to be able to take advantage of
function alias within the vtable that is in fact the only
implementation. For that reason, the Regular LTO testing in the second
test case is currently disabled, and will be enabled along with a follow
on enhancement fix for Regular LTO WPD.

Differential Revision: https://reviews.llvm.org/D144209

20 months ago[clang] add __has_extension(gnu_asm_goto_with_outputs_full)
Nick Desaulniers [Fri, 17 Feb 2023 01:56:59 +0000 (17:56 -0800)]
[clang] add __has_extension(gnu_asm_goto_with_outputs_full)

Also move the line about __has_extension(gnu_asm_goto_with_outputs) so
that it is more generally about asm goto, not the paragraph on symbolic
references.

Reviewed By: efriedma, void

Differential Revision: https://reviews.llvm.org/D143205

20 months ago[clang] fix -Wuninitialized for asm goto outputs on indirect edges.
Nick Desaulniers [Fri, 17 Feb 2023 01:50:34 +0000 (17:50 -0800)]
[clang] fix -Wuninitialized for asm goto outputs on indirect edges.

Now that we support outputs from asm goto along indirect edges, we can
remove/revert some code that was added to help warn about the previous
limitation that outputs were not supported along indirect edges.

Reverts some code added in:
commit 72aa619a7fe0 ("Warn of uninitialized variables on asm goto's indirect branch")
commit 3a604fdbcd5f ("[Clang][CFG] check children statements of asm goto")
But keeps+updates the tests.

Link: https://github.com/llvm/llvm-project/issues/53562
Reviewed By: void

Differential Revision: https://reviews.llvm.org/D140508

20 months ago[Clang] support for outputs along indirect edges of asm goto
Nick Desaulniers [Fri, 17 Feb 2023 01:49:39 +0000 (17:49 -0800)]
[Clang] support for outputs along indirect edges of asm goto

Initial support for asm goto w/ outputs (D69876) only supported outputs
along the "default" (aka "fallthrough") edge.

We can support outputs along all edges by repeating the same pattern of
stores along the indirect edges that we allready do for the default
edge.  One complication is that these indirect edges may be critical
edges which would need to be split. Another issue is that mid-codgen of
LLVM IR, the control flow graph might not reflect the control flow of
the final function.

To avoid this "chicken and the egg" problem assume that any given
indirect edge may become a critical edge, and pro-actively split it.
This is unnecessary if the edge does not become critical, but LLVM will
optimize such cases via tail duplication.

Fixes: https://github.com/llvm/llvm-project/issues/53562

Reviewed By: void

Differential Revision: https://reviews.llvm.org/D136497

20 months ago[Clang] refactor CodeGenFunction::EmitAsmStmt NFC
Nick Desaulniers [Fri, 17 Feb 2023 01:49:16 +0000 (17:49 -0800)]
[Clang] refactor CodeGenFunction::EmitAsmStmt NFC

Prerequisite to further modifications in D136497.

Basically, there is a large body of code in CodeGenFunction::EmitAsmStmt
for emitting stores of outputs. We want to be able to repeat this logic,
for each destination of a callbr (rather than just the default
destination which is what the code currently does).

Also does some smaller cleanups like whitespace cleanups, and removing
pointless casts.

Reviewed By: void, jyknight

Differential Revision: https://reviews.llvm.org/D137113

20 months ago[llvm] add CallBrPrepare pass to pipelines
Nick Desaulniers [Fri, 17 Feb 2023 01:48:14 +0000 (17:48 -0800)]
[llvm] add CallBrPrepare pass to pipelines

Capstone of
https://discourse.llvm.org/t/rfc-syncing-asm-goto-with-outputs-with-gcc/65453/8

Clang changes are still necessary to enable the use of outputs along
indirect edges of asm goto statements.

Link: https://github.com/llvm/llvm-project/issues/53562
Reviewed By: void

Differential Revision: https://reviews.llvm.org/D140180

20 months ago[llvm][SelectionDAGBuilder] codegen callbr.landingpad intrinsic
Nick Desaulniers [Fri, 17 Feb 2023 01:47:37 +0000 (17:47 -0800)]
[llvm][SelectionDAGBuilder] codegen callbr.landingpad intrinsic

Given a CallBrInst, retain its first virtual register in SelectionDagBuilder's
FunctionLoweringInfo if there's corresponding landingpad. Walk the list
of COPY MachineInstr to find the original virtual and physical registers
defined by the INLINEASM_BR MachineInst.

Test cases from https://reviews.llvm.org/D139565.
Link: https://github.com/llvm/llvm-project/issues/59538
Part 3 from
https://discourse.llvm.org/t/rfc-syncing-asm-goto-with-outputs-with-gcc/65453/8

Follow up patches still need to wire up CallBrPrepare into the pass
pipelines.

Reviewed By: efriedma, void

Differential Revision: https://reviews.llvm.org/D140160