platform/upstream/llvm.git
15 months ago[flang] Emit error when a positional actual argument follows an argument with a keyword
Peter Klausler [Wed, 15 Mar 2023 00:11:48 +0000 (17:11 -0700)]
[flang] Emit error when a positional actual argument follows an argument with a keyword

A positional (non-keyword) actual argument or alternate return label is not
allowed to follow an actual argument with a keyword.

Differential Revision: https://reviews.llvm.org/D146575

15 months ago[flang] Warn about inconsistent external procedure interfaces
Peter Klausler [Tue, 14 Mar 2023 23:02:30 +0000 (16:02 -0700)]
[flang] Warn about inconsistent external procedure interfaces

When multiple scopes in a compilation define interfaces (explicit
or implicit) for an external procedure, warn when those interfaces
are inconsistent.

Differential Revision: https://reviews.llvm.org/D146574

15 months ago[flang] Catch attempt to apply ASYNCHRONOUS attribute to a non-variable
Peter Klausler [Tue, 14 Mar 2023 20:10:25 +0000 (13:10 -0700)]
[flang] Catch attempt to apply ASYNCHRONOUS attribute to a non-variable

Complain about the ASYNCHRONOUS attribute on symbols that aren't
variables.

Differential Revision: https://reviews.llvm.org/D146573

15 months ago[mlir] tosa.concat - fix isCompatibleReturnTypes()
Maya Amrami [Mon, 27 Mar 2023 21:59:25 +0000 (21:59 +0000)]
[mlir] tosa.concat - fix isCompatibleReturnTypes()

Reviewed By: eric-k256

Differential Revision: https://reviews.llvm.org/D146901

15 months ago[flang] Diagnose bad attributes of implicit interface externals
Peter Klausler [Tue, 14 Mar 2023 19:38:26 +0000 (12:38 -0700)]
[flang] Diagnose bad attributes of implicit interface externals

Procedures with the EXTERNAL attribute (explicit or implied) are
not permitted to be ALLOCATABLE or be arrays if their interfaces
are implicit.

Differential Revision: https://reviews.llvm.org/D146572

15 months ago[mlir] Add support for non-f32 polynomial approximation
Robert Suderman [Mon, 27 Mar 2023 20:55:32 +0000 (20:55 +0000)]
[mlir] Add support for non-f32 polynomial approximation

Polynomial approximations assume F32 values. We can convert all non-f32
cases to operate on f32s with intermediate casts.

Reviewed By: jpienaar

Differential Revision: https://reviews.llvm.org/D146677

15 months ago[docs] Fix code example: the JITLink APIs use ExecutorAddr now.
Lang Hames [Mon, 27 Mar 2023 21:56:04 +0000 (14:56 -0700)]
[docs] Fix code example: the JITLink APIs use ExecutorAddr now.

15 months ago[flang] Handle dynamic and remotely scoped non-type-bound UDDTIO subroutines
Peter Klausler [Fri, 10 Mar 2023 23:06:01 +0000 (15:06 -0800)]
[flang] Handle dynamic and remotely scoped non-type-bound UDDTIO subroutines

The present I/O infrastructure for user-defined derived type I/O
subroutines works fine for type-bound I/O generic bindings.  It also works
for explicit INTERFACE blocks and GENERIC statements that define
UDDIO subroutines in the same scope as the definition of the derived type,
so long as the specific procedures in those bindings are module procedures
or external procedures.

For non-type-bound UDDTIO specific procedures that are dummy procedures,
thunks of inner procedures, or procedure pointers, or that are defined with
interfaces or GENERIC outside the scope of the definition of the derived
type, a new runtime I/O API is needed so that lowering can generate
a call that supplies the appropriate procedure as well as the defined
type instance.

This patch specifies and implements this new runtime API and provides
utility routines for lowering to use to determine whether it should be
called for any particular OutputItem or InputItem in the parse tree.

Differential Revision: https://reviews.llvm.org/D146571

15 months ago[NFC] Remove unused member variable `PolicyBitMask` in SemaRecord
Chittireddy, Sindhu [Fri, 24 Mar 2023 21:12:17 +0000 (17:12 -0400)]
[NFC] Remove unused member variable `PolicyBitMask` in SemaRecord

Coverity static analysis tool found PolicyBitMask being used
uninitialized in push_back call in RVVEmitter::createRVVIntrinsics()
but this variable has no uses.

Differential revision: https://reviews.llvm.org/D146847

15 months ago[gn build] Port 158a431227a8
LLVM GN Syncbot [Mon, 27 Mar 2023 21:13:05 +0000 (21:13 +0000)]
[gn build] Port 158a431227a8

15 months agoRevert ExtractAPI from https://reviews.llvm.org/D146656
Daniel Grumberg [Mon, 27 Mar 2023 21:12:11 +0000 (22:12 +0100)]
Revert ExtractAPI from https://reviews.llvm.org/D146656

15 months ago[clang-format] Don't annotate left brace of struct as FunctionLBrace
Owen Pan [Sun, 26 Mar 2023 02:31:57 +0000 (19:31 -0700)]
[clang-format] Don't annotate left brace of struct as FunctionLBrace

Related to a02c3af9f19d. Fixes #61700.

Differential Revision: https://reviews.llvm.org/D146895

15 months agoFix incorrect .mailmap entry
Min-Yih Hsu [Mon, 27 Mar 2023 21:02:10 +0000 (14:02 -0700)]
Fix incorrect .mailmap entry

Of my contact email. Sorry for the churn.

15 months ago[CODE_OWNERS] Update my contact information
Min-Yih Hsu [Mon, 27 Mar 2023 20:51:53 +0000 (13:51 -0700)]
[CODE_OWNERS] Update my contact information

As well as updating .mailmap

15 months ago[RISCV] Replace an 'else if' with 'else'+assert. NFC
Craig Topper [Mon, 27 Mar 2023 20:48:06 +0000 (13:48 -0700)]
[RISCV] Replace an 'else if' with 'else'+assert. NFC

There are only two cases here. Using an assert ensures there is
no handled third case.

Also move comment to avoid odd formatting.

Reviewed By: reames

Differential Revision: https://reviews.llvm.org/D146998

15 months ago[gn build] Port 41b6d5863fef
LLVM GN Syncbot [Mon, 27 Mar 2023 20:45:22 +0000 (20:45 +0000)]
[gn build] Port 41b6d5863fef

15 months agoRevert "[lldb] Move UnixSignals creation into Platform plugins"
Alex Langford [Mon, 27 Mar 2023 20:33:06 +0000 (13:33 -0700)]
Revert "[lldb] Move UnixSignals creation into Platform plugins"

This reverts commit ee232506b870ce5282cc4da5ca493d41d361feb3.
As discussed in https://reviews.llvm.org/D146668 we'll find another way
forward.

15 months agoRevert "[lldb-server] Use Platform plugin corresponding to the host"
Alex Langford [Mon, 27 Mar 2023 20:32:48 +0000 (13:32 -0700)]
Revert "[lldb-server] Use Platform plugin corresponding to the host"

This reverts commit 0c5cee779929f840f4f286c5894a01f583ee7b4a.
As discussed in https://reviews.llvm.org/D146668 we'll find another way
forward.

15 months ago[lld-macho] Support re-exports of individual symbols
Jez Ng [Mon, 27 Mar 2023 20:39:37 +0000 (16:39 -0400)]
[lld-macho] Support re-exports of individual symbols

Specifically, we support this:

  ld64.lld -dylib foo.o libbar.dylib -exported_symbol _bar -o libfoo.dylib

Where `_bar` is defined in libbar.dylib.

Reviewed By: #lld-macho, smeenai

Differential Revision: https://reviews.llvm.org/D144153

15 months ago[lldb] Remove NO_PLUGIN_DEPENDENCIES from lldbTarget
Alex Langford [Mon, 27 Mar 2023 20:35:46 +0000 (13:35 -0700)]
[lldb] Remove NO_PLUGIN_DEPENDENCIES from lldbTarget

As discussed in https://reviews.llvm.org/D146668 I will be reverting
0c5cee779929f840f4f286c5894a01f583ee7b4a and ee232506b870ce5282cc4da5ca493d41d361feb3.
I'm removing NO_PLUGIN_DEPENDENCIES from lldbTarget first so that these
reverts will not leave LLDB in a bad state afterwards.

15 months ago[RISCV] Rename FeatureExtZc* to FeatureStdExtZc*. NFC
Craig Topper [Mon, 27 Mar 2023 20:19:01 +0000 (13:19 -0700)]
[RISCV] Rename FeatureExtZc* to FeatureStdExtZc*. NFC

Even for experimental extensions, I think we always include "Std"
in the feature name.

Reviewed By: asb

Differential Revision: https://reviews.llvm.org/D146997

15 months agowrite semantics tests for CHANGE TEAM statement.
Hussain Kadhem [Mon, 20 Mar 2023 17:33:01 +0000 (10:33 -0700)]
write semantics tests for CHANGE TEAM statement.

Added semantics tests for aspects of CHANGE TEAM statements which are not duplicates from FORM TEAM statements.

Reviewed By: ktras

Differential Revision: https://reviews.llvm.org/D137908

15 months agoRecommit "[RISCV] Add .insn support for compressed formats."
Craig Topper [Mon, 27 Mar 2023 19:30:37 +0000 (12:30 -0700)]
Recommit "[RISCV] Add .insn support for compressed formats."

I think the failure was caused by a mistake in an earlier patch.

Original commit message:

We've supported .insn for non-compressed for a while. This finishes the compressed supported.

Differential Revision: https://reviews.llvm.org/D146663

15 months ago[M68k] Add support for lowering i1 SIGN_EXTEND_INREG
Min-Yih Hsu [Mon, 27 Mar 2023 19:12:39 +0000 (12:12 -0700)]
[M68k] Add support for lowering i1 SIGN_EXTEND_INREG

Lowering i1 (inreg) sext with `(NEG (AND %val, 1))`.

15 months ago[RISCV] Handle non-recursive muls of strides in gather/scatter lowering
Luke Lau [Mon, 27 Mar 2023 16:13:15 +0000 (17:13 +0100)]
[RISCV] Handle non-recursive muls of strides in gather/scatter lowering

The gather scatter lowering pass can fold multiplies of a step vector
into the stride for the recursive case, so this extends it for the
non-recursive case.
The logic can probably be shared between the two at some point to extend
it to shls and ors.

Reviewed By: reames

Differential Revision: https://reviews.llvm.org/D146983

15 months ago[RISCV] Add tests for gather scatter pass for non-recursive mul strides
Luke Lau [Mon, 27 Mar 2023 16:02:16 +0000 (17:02 +0100)]
[RISCV] Add tests for gather scatter pass for non-recursive mul strides

Reviewed By: craig.topper

Differential Revision: https://reviews.llvm.org/D146982

15 months agoEnsure I'm listed as MSP430 code owner
Anton Korobeynikov [Mon, 27 Mar 2023 19:10:12 +0000 (12:10 -0700)]
Ensure I'm listed as MSP430 code owner

15 months agoFix bazel build for mlir-opt
Chenguang Wang [Mon, 27 Mar 2023 18:44:15 +0000 (11:44 -0700)]
Fix bazel build for mlir-opt

Tested locally with `bazel build --repo_env=CC=clang --repo_env=BAZEL_CXXOPTS="-std=c++17" @llvm-project//mlir:mlir-opt`.

Reviewed By: jyknight

Differential Revision: https://reviews.llvm.org/D146992

15 months agoAdded checking for completeness of lvalue in conditional operator
Aditya Singh [Mon, 27 Mar 2023 18:45:43 +0000 (14:45 -0400)]
Added checking for completeness of lvalue in conditional operator

Given:
```
struct x y;
int main(void)
{
    (void)(1 ? y : y);
}
struct x {int i;};
```
The conditional operator(?:) requires the second and third operands to
be of compatible types. To be compatible, they also need to be
complete (however, both can be void). Therefore, the expected response
from clang after running the above code as a C program should be error
dialogue pointing out that both the types are incomplete hence
incompatible, but the code compiled without any errors.

The patch ensures the completeness in the CheckCondtionalOperand
function present in llvm-project/clang/lib/Sema/SemaChecking.cpp.

Fixes https://github.com/llvm/llvm-project/issues/59718
Differential Revision: https://reviews.llvm.org/D144358

15 months ago[libc] Simplify enabling the GPU build for libc
Joseph Huber [Mon, 27 Mar 2023 16:05:11 +0000 (11:05 -0500)]
[libc] Simplify enabling the GPU build for libc

Currently the GPU build requires the `LLVM_LIBC_FULL_BUILD` option to be
set. This patch changes the logic so that it is always enabled when
targeting the GPU. Also, this patch allows `LIBC_GPU_BUILD` and
`LIBC_GPU_ARCHITECTURES` to both enable a GPU build. Now, enabling the
GPU support should only require the following CMake:

```
  -DLLVM_ENABLE_RUNTIMES=libc -DLIBC_GPU_ARCHITECTURES=gfx1030
```

Reviewed By: jdoerfert, sivachandra

Differential Revision: https://reviews.llvm.org/D146979

15 months ago[test] Avoid legacy -target and move 2>&1 to conventional location
Fangrui Song [Mon, 27 Mar 2023 18:22:03 +0000 (11:22 -0700)]
[test] Avoid legacy -target and move 2>&1 to conventional location

The 2>&1 moving part is similar to commit 731264b0c2af7aa46bd39625202a99e06cfccff9.

15 months agoRecommit "[RISCV] Relax opcode restriction for .insn."
Craig Topper [Mon, 27 Mar 2023 17:59:08 +0000 (10:59 -0700)]
Recommit "[RISCV] Relax opcode restriction for .insn."

Original commit message:
I recently added a check to make sure the lower 2 bits of the opcode
were 0x3. A post commit comment pointed out this would be allowed
for a custom extension when C is not supported.

Note, binutils does have this restriction.

15 months ago[mlir] Add missing cast functions to mlir namespace
Rahul Kayaith [Fri, 24 Mar 2023 18:25:18 +0000 (14:25 -0400)]
[mlir] Add missing cast functions to mlir namespace

Reviewed By: rriddle

Differential Revision: https://reviews.llvm.org/D146843

15 months ago[M68k] Add support for lowering ATOMIC_SWAP
Min-Yih Hsu [Mon, 27 Mar 2023 17:48:10 +0000 (10:48 -0700)]
[M68k] Add support for lowering ATOMIC_SWAP

Lower to calling __sync_lock_test_and_set_* for target < M68020.

15 months agoRevert "foo"
Craig Topper [Mon, 27 Mar 2023 17:53:14 +0000 (10:53 -0700)]
Revert "foo"

This reverts commit 4fde20b8a62ed3d850e4e770ae640039ca1a9ed8.

I failed to squash this

15 months agoRevert "[RISCV] Relax opcode restriction for .insn."
Craig Topper [Mon, 27 Mar 2023 17:52:56 +0000 (10:52 -0700)]
Revert "[RISCV] Relax opcode restriction for .insn."

This reverts commit 5c5fe3afff12b41d34477977c07c71337fa4ba9d.

This is failing on the build bots

15 months agoRevert "[RISCV] Add .insn support for compressed formats."
Craig Topper [Mon, 27 Mar 2023 17:52:41 +0000 (10:52 -0700)]
Revert "[RISCV] Add .insn support for compressed formats."

This reverts commit 5e2445ae65f0a8f1ac38049bdd2f0fd5572be488.

This is failing on the build bots.

15 months ago[AArch64][PeepholeOpt]Optimize ALU + compare to flag-setting ALU
Mingming Liu [Thu, 23 Mar 2023 20:31:26 +0000 (13:31 -0700)]
[AArch64][PeepholeOpt]Optimize ALU + compare to flag-setting ALU

The motivating example is in https://godbolt.org/z/45nbdYMK9
- For this example, `subs` is generated for the good case; `sub` followed by `cmp` is generated for the bad case. Since signed overflow is undefined behavior in C/C++ (indicated as `nsw` flag in LLVM IR), `subs` should be generated for the good case as well.

This patch relaxes one restriction from "quit optimization when V is used" to "continue if MI produces poison value when signed overflow occurs". This is not meant to be C/C++ specific since it looks at 'NoSWrap' since it looks at MachineInstr flags.

Reviewed By: dmgreen

Differential Revision: https://reviews.llvm.org/D146820

15 months ago[NFC]Pre-commit test case
Mingming Liu [Thu, 23 Mar 2023 07:41:19 +0000 (00:41 -0700)]
[NFC]Pre-commit test case

Pre-commit test cases to show missed icmp-opt with flag-setting ALUs

Reviewed By: dmgreen

Differential Revision: https://reviews.llvm.org/D146754

15 months ago[SCCIterator] Fix an issue in scc_member_iterator sorting
Hongtao Yu [Tue, 21 Mar 2023 17:57:11 +0000 (10:57 -0700)]
[SCCIterator] Fix an issue in scc_member_iterator sorting

Members in an scc are supposed to be sorted in a top-down or topological order based on edge weights. Previously this is achived by building a MST out of the SCC and enforcing an BFS walk on the MST. A BFS on a tree does give a top-down topological order, however, the MST built here isn't really a tree. This is becuase of a trick done to avoid expansive detection of a cycle on a directed graph when an edge is added. When the MST is built, its edges are considered undirected. But in reality they are directed, thus a BST walk doesn't necessarily give a topological order. I'm tweaking the BFS walk slightly to yield a topological order.

Basically I'm using Kahn's algorithm on MST to compute a topological traversal order. The algorithm starts from nodes that have no incoming edge. These nodes are "roots" of the MST forest. This ensures that nodes are visited before their descendants are, thus ensures a topological traversal order of the MST.

Reviewed By: wenlei

Differential Revision: https://reviews.llvm.org/D130717

15 months ago[RISCV] Add .insn support for compressed formats.
Craig Topper [Mon, 27 Mar 2023 17:27:59 +0000 (10:27 -0700)]
[RISCV] Add .insn support for compressed formats.

We've supported .insn for non-compressed for a while. This finishes the compressed supported.

Reviewed By: asb

Differential Revision: https://reviews.llvm.org/D146663

15 months ago[RISCV] Relax opcode restriction for .insn.
Craig Topper [Mon, 27 Mar 2023 17:24:52 +0000 (10:24 -0700)]
[RISCV] Relax opcode restriction for .insn.

I recently added a check to make sure the lower 2 bits of the opcode
were 0x3. A post commit comment pointed out this would be allowed
for a custom extension when C is not supported.

Note, binutils does have this restriction.

15 months agofoo
Craig Topper [Mon, 27 Mar 2023 17:06:34 +0000 (10:06 -0700)]
foo

15 months ago[LegalizeTypes][TargetLowering][RISCV] Fix regressions from D146786.
Craig Topper [Mon, 27 Mar 2023 16:55:09 +0000 (09:55 -0700)]
[LegalizeTypes][TargetLowering][RISCV] Fix regressions from D146786.

Add some special cases for UADDO to recover codegen after D146786.

Reviewed By: reames, liaolucy

Differential Revision: https://reviews.llvm.org/D146789

15 months ago[CodeGenPrepare][RISCV] Correct the MathUsed flag for shouldFormOverflowOp
Craig Topper [Mon, 27 Mar 2023 16:51:46 +0000 (09:51 -0700)]
[CodeGenPrepare][RISCV] Correct the MathUsed flag for shouldFormOverflowOp

For add, if we match the constant edge case the add isn't used by
the compare so we shouldn't check for 2 users.

For sub, the compare is not a user of the sub so the math is
used if the sub has any users.

This regresses RISC-V which I will work on other patches for.

Reviewed By: RKSimon

Differential Revision: https://reviews.llvm.org/D146786

15 months ago[X86] select-constant-lea.ll - cleanup test name. NFC.
Simon Pilgrim [Mon, 27 Mar 2023 16:56:05 +0000 (17:56 +0100)]
[X86] select-constant-lea.ll - cleanup test name. NFC.

Drop the trailing j function signature artifact.

15 months ago[SLP] Add test to check for GEP vectorization
Zain Jaffal [Mon, 27 Mar 2023 16:52:07 +0000 (17:52 +0100)]
[SLP] Add test to check for GEP vectorization

add a test to check for gep vectorization after the change from D144128 where the gep vectorization is dependant on the target hook `prefersVectorizedAddressing()`

Reviewed By: fhahn

Differential Revision: https://reviews.llvm.org/D146540

15 months ago[RISCV] Replace RISCV->RISC-V in strings.
Craig Topper [Mon, 27 Mar 2023 16:20:07 +0000 (09:20 -0700)]
[RISCV] Replace RISCV->RISC-V in strings.

To be consistent with RISC-V branding guidelines
https://riscv.org/about/risc-v-branding-guidelines/
Think we should be using RISC-V where possible.

D146449 already updated comments. Strings may have more user impact.

Reviewed By: asb

Differential Revision: https://reviews.llvm.org/D146451

15 months ago[RISCV] Replace RISCV -> RISC-V in comments. NFC
Craig Topper [Mon, 27 Mar 2023 16:15:53 +0000 (09:15 -0700)]
[RISCV] Replace RISCV -> RISC-V in comments. NFC

To be consistent with RISC-V branding guidelines
https://riscv.org/about/risc-v-branding-guidelines/
Think we should be using RISC-V where possible.

More patches will follow.

Reviewed By: asb

Differential Revision: https://reviews.llvm.org/D146449

15 months ago[X86] Regenerate avx2-vector-shifts.ll check prefixes
Simon Pilgrim [Mon, 27 Mar 2023 16:46:21 +0000 (17:46 +0100)]
[X86] Regenerate avx2-vector-shifts.ll check prefixes

Add common CHECK prefix

15 months ago[libc++] Use named Lit features to flag back-deployment XFAILs
Louis Dionne [Sat, 18 Mar 2023 14:05:19 +0000 (10:05 -0400)]
[libc++] Use named Lit features to flag back-deployment XFAILs

Instead of writing something like `XFAIL: use_system_cxx_lib && target=...`
to XFAIL back-deployment tests, introduce named Lit features like
`availability-shared_mutex-missing` to represent those. This makes the
XFAIL annotations leaner, and solves the problem of XFAIL comments
potentially getting out of sync. This would also make it easier for
another vendor to add their own annotations to the test suite by simply
changing how the feature is defined for their OS releases, instead
of having to modify hundreds of tests to add repetitive annotations.

This doesn't touch *all* annotations -- only annotations that were widely
duplicated are given named features (e.g. when filesystem or shared_mutex
were introduced). I still think it probably doesn't make sense to have a
named feature for every single fix we make to the dylib.

This is in essence a revert of 2659663, but since then the test suite
has changed significantly. Back when I did 2659663, the configuration
files we have for the test suite right now were being bootstrapped and
it wasn't clear how to provide these features for back-deployment in
that context. Since then, we have a streamlined way of defining these
features in `features.py` and that doesn't impact the ability for a
configuration file to stay minimal.

The original motivation for this change was that I am about to propose
a change that would touch essentially all XFAIL annotations for back-deployment
in the test suite, and this greatly reduces the number of lines changed
by that upcoming change, in addition to making the test suite generally
better.

Differential Revision: https://reviews.llvm.org/D146359

15 months ago[diagtool] explain that yellow is used to denote disabled-by-default warnings
Nick Desaulniers [Mon, 27 Mar 2023 16:38:05 +0000 (09:38 -0700)]
[diagtool] explain that yellow is used to denote disabled-by-default warnings

This is already implied, but let's just be explicit about it.

Reviewed By: MaskRay

Differential Revision: https://reviews.llvm.org/D146827

15 months ago[LinkerWrapper] Do not extract globals with no offloading language
Joseph Huber [Fri, 17 Mar 2023 20:10:25 +0000 (15:10 -0500)]
[LinkerWrapper] Do not extract globals with no offloading language

The linker wrapper needs to reinvent its own special static library
handling for static libraries containing fatbinaries. This is primarily
because offloading languages expect certain global symbols to be visible
to the host so we must consider them used symbols. However we should be
able to remove this requirement if we are linking in "freestanding" code
that was not created by an offloading language.

The motivation for this is to support the work-in-progress `libc` for
GPUs. It is provided as a static library with no offloading language
set. This logic will let us only import used `libc` symbols always.

Reviewed By: yaxunl

Differential Revision: https://reviews.llvm.org/D146326

15 months ago[X86] Regenerate avx2-shift.ll check prefixes
Simon Pilgrim [Mon, 27 Mar 2023 16:34:41 +0000 (17:34 +0100)]
[X86] Regenerate avx2-shift.ll check prefixes

Add common CHECK prefix, and rename X32 -> X86 (we try to use X32 for gnux32 triples)

15 months ago[gn build] Port ea35740e7e18
LLVM GN Syncbot [Mon, 27 Mar 2023 16:24:59 +0000 (16:24 +0000)]
[gn build] Port ea35740e7e18

15 months agoAddress code review feedback
Daniel Grumberg [Wed, 22 Mar 2023 23:22:48 +0000 (23:22 +0000)]
Address code review feedback

15 months ago[clang][ExtractAPI] Refactor ExtractAPIVisitor to make it more extensible
Daniel Grumberg [Wed, 22 Mar 2023 18:50:59 +0000 (18:50 +0000)]
[clang][ExtractAPI] Refactor ExtractAPIVisitor to make it more extensible

Use CRTP to enable creating statically dispatched subclasses of
ExtractAPIVisitor.
This enables adding extension points and customising the behavior more
easily.

This is used in CXExtractAPI.cpp to create a specialized visitor for
Libclang as well as streamlining the batch implementation in ExtractAPIConsumer.cpp

[clang][ExtractAPI] Improve tests for clang_getSymbolGraphForCursor

Adds a new mode to c-index-test that can fetch a single symbol symbol
graph for a given source location. This way we can be more precise when
writing tests for clang_getSymbolGraphForCursor.
Additionaly this makes it easier to debug the function.

Differential Revision: https://reviews.llvm.org/D146656

15 months ago[clang][ExtractAPI] Refactor ExtractAPIVisitor to make it more extensible
Daniel Grumberg [Wed, 22 Mar 2023 18:43:24 +0000 (18:43 +0000)]
[clang][ExtractAPI] Refactor ExtractAPIVisitor to make it more extensible

Use CRTP to enable creating statically dispatched subclasses of
ExtractAPIVisitor.
This enables adding extension points and customising the behavior more
easily.

This is used in CXExtractAPI.cpp to create a specialized visitor for
Libclang as well as streamlining the batch implementation in ExtractAPIConsumer.cpp

15 months ago[mlir] Implement pass utils for 1:N type conversions.
Ingo Müller [Tue, 21 Feb 2023 11:21:25 +0000 (11:21 +0000)]
[mlir] Implement pass utils for 1:N type conversions.

The current dialect conversion does not support 1:N type conversions.
This commit implements a (poor-man's) dialect conversion pass that does
just that. To keep the pass independent of the "real" dialect conversion
infrastructure, it provides a specialization of the TypeConverter class
that allows for N:1 target materializations, a specialization of the
RewritePattern and PatternRewriter classes that automatically add
appropriate unrealized casts supporting 1:N type conversions and provide
converted operands for implementing subclasses, and a conversion driver
that applies the provided patterns and replaces the unrealized casts
that haven't folded away with user-provided materializations.

The current pass is powerful enough to express many existing manual
solutions for 1:N type conversions or extend transforms that previously
didn't support them, out of which this patch implements call graph type
decomposition (which is currently implemented with a ValueDecomposer
that is only used there).

The goal of this pass is to illustrate the effect that 1:N type
conversions could have, gain experience in how patterns should be
written that achieve that effect, and get feedback on how the APIs of
the dialect conversion should be extended or changed to support such
patterns. The hope is that the "real" dialect conversion eventually
supports such patterns, at which point, this pass could be removed
again.

Reviewed By: springerm

Differential Revision: https://reviews.llvm.org/D144469

15 months ago[X86] Teach computeKnownBitsForTargetNode about MUL_IMM
Kazu Hirata [Mon, 27 Mar 2023 16:02:19 +0000 (09:02 -0700)]
[X86] Teach computeKnownBitsForTargetNode about MUL_IMM

This patch teaches computeKnownBitsForTargetNode about MUL_IMM.

MUL_IMM comes up in certain select of constants.  Specifically, it is
used to multiply the result of SETCC.  Computing the known zero bits
of MUL_IMM allows matchAddressRecursively us to convert some OR into
ADD, which eventually becomes a part of LEA.

This patch fixes:

https://github.com/llvm/llvm-project/issues/61365

Differential Revision: https://reviews.llvm.org/D146787

15 months ago[clang][ASTImporter] Import typedefs to distinct records as distinct nodes.
Balázs Kéri [Mon, 27 Mar 2023 14:53:02 +0000 (16:53 +0200)]
[clang][ASTImporter] Import typedefs to distinct records as distinct nodes.

When a typedef node is imported, ASTImporter should not find an existing similar
typedef node for it that comes from different context (translation unit or scope).
This should avoid a situation where an existing typedef declaration is returned
at import of a typedef, but the underlying type was already imported as a new
type object.

Reviewed By: vabridgers

Differential Revision: https://reviews.llvm.org/D145479

15 months ago[X86] Move combineVectorSizedSetCCEquality above MatchVectorAllZeroTest. NFC.
Simon Pilgrim [Mon, 27 Mar 2023 15:03:04 +0000 (16:03 +0100)]
[X86] Move combineVectorSizedSetCCEquality above MatchVectorAllZeroTest. NFC.

The plan is to merge most of the functionality of both of these into a single 'match vector sized data' function.

15 months ago[AIX][Driver] Implement -mxcoff-build-id option
Wael Yehia [Tue, 28 Feb 2023 04:26:52 +0000 (23:26 -0500)]
[AIX][Driver] Implement -mxcoff-build-id option

The -mxcoff-build-id=0xHEXSTRING option is an alternative to the
--build-id=0xHEXSTRING linker option that is not currently available in
the AIX linker.

If HEXSTRING is an odd number of hex digits then a '0' character is prepended.
The characters ':' and '-' are not allowed (unlike the GNU linker option).
The given build-id will be saved in the string table of the loader section.

A subsequent commit will teach the profile runtime to read and use the embedded id.

Reviewed By: daltenty, qiongsiwu1, stephenpeckham

Differential Revision: https://reviews.llvm.org/D146431

15 months ago[llvm-exegesis] if lit can't find llvm-exegesis, fail gracefully
Ties Stuij [Mon, 27 Mar 2023 15:31:50 +0000 (16:31 +0100)]
[llvm-exegesis] if lit can't find llvm-exegesis, fail gracefully

currently in can_use_perfcounters() in
llvm/test/tools/llvm-exegesis/lit.local.cfg, we check if we can execute
llvm-exegesis. But we don't check if we did actually find llvm-exegesis. If we
didn't, lit.util.which() will return None, and we try to execute that.

Reviewed By: courbet

Differential Revision: https://reviews.llvm.org/D146964

15 months ago[mlir] FuncToLLVM: fail gracefully in bare pointer multi-result covnersion
Alex Zinenko [Mon, 27 Mar 2023 14:53:11 +0000 (14:53 +0000)]
[mlir] FuncToLLVM: fail gracefully in bare pointer multi-result covnersion

When type conversion fails, return pattern failure instead of crashing.

Closes #61717.

15 months ago[ADT] Work around MSVC bug affecting `get(enumerator_result)`
Jakub Kuderski [Mon, 27 Mar 2023 14:28:36 +0000 (10:28 -0400)]
[ADT] Work around MSVC bug affecting `get(enumerator_result)`

This happened on a small number of MSVC releases (19.31.31xxx, Visual Studio 2022 17.1.x), and worked fine on everything else.

The issue seemed to be related to return type deduction on a function with and `if constexpr`; the compiler got confused and deduced different function return type from the type of the return statement.

The workaround is to split `get` into two functions using `enable_if`.

Reviewed By: dstuttard

Differential Revision: https://reviews.llvm.org/D146893

15 months ago[clangd] Use expansion location for missing include diagnostics.
Viktoriia Bakalova [Thu, 23 Mar 2023 15:15:32 +0000 (15:15 +0000)]
[clangd] Use expansion location for missing include diagnostics.

Differential Revision: https://reviews.llvm.org/D146727

15 months ago[clang-format] Improve QualifierAlignment
Alexander Hederstaf [Mon, 27 Mar 2023 14:16:11 +0000 (15:16 +0100)]
[clang-format] Improve QualifierAlignment

Qualifiers were not moved for non-pointer non-simple types.
Add additional support for many special cases such as templates,
requires clauses, long qualified names.

Fixes https://github.com/llvm/llvm-project/issues/57154 and
https://github.com/llvm/llvm-project/issues/60898

Reviewed By: MyDeveloperDay, HazardyKnusperkeks

Differential Revision: https://reviews.llvm.org/D144709

15 months ago[RISCV] Remove isCall instruction flag from JAL/JALR
Job Noorman [Mon, 27 Mar 2023 14:10:48 +0000 (16:10 +0200)]
[RISCV] Remove isCall instruction flag from JAL/JALR

As far as I can tell, the instruction flags set in the various
RISCVInstrInfo*.td files are always under-approximations. That is,
something like isTerminator will only be set for opcodes that are
*always* terminators.

The exception seems to be isCall; this is set for JAL/JALR while these
instruction are not always calls. This patch proposes to remove isCall
from JAL/JALR for consistency.

A follow-up patch will add more detailed analysis of some instruction
properties based on MCInstrAnalysis.

Reviewed By: asb, craig.topper

Differential Revision: https://reviews.llvm.org/D146437

15 months ago[SVE][LoopVectorize] Add option to disable tail-folding for reverse loops
David Sherwood [Tue, 14 Mar 2023 18:15:03 +0000 (18:15 +0000)]
[SVE][LoopVectorize] Add option to disable tail-folding for reverse loops

If we use tail-folding for reverse loops that contain loads
and stores then we will need to reverse the loop predicate.
This patch adds a new 'reverse' sve-tail-folding option and
ensures they are not considered 'simple'.

I did this by adding a function called
containsDecreasingPointers to AArch64TargetTransformInfo.cpp
that searches all instructions in the loop for loads or
stores with negative strides.

Differential Revision: https://reviews.llvm.org/D146128

15 months ago[gn build] Port 42058eea7912
LLVM GN Syncbot [Mon, 27 Mar 2023 13:58:17 +0000 (13:58 +0000)]
[gn build] Port 42058eea7912

15 months agoDocument the Clang policies on claiming support for a feature
Aaron Ballman [Mon, 27 Mar 2023 13:54:29 +0000 (09:54 -0400)]
Document the Clang policies on claiming support for a feature

We do not currently have this written down anywhere, and as a result,
we're sometimes inconsistent with how we handle feature test macros and
the feature status pages. This is an attempt to document what I
understand our existing policies to be instead of defining a new policy.

Note, this is being added to the Clang internals manual because we
don't have a separate document for Clang developer policies. At some
point, I think we may want a standalone document for that, and this
content can be moved there at that time.

Differential Revision: https://reviews.llvm.org/D146420

15 months ago[reland][ADT] add ConcurrentHashtable class.
Alexey Lapshin [Wed, 22 Mar 2023 16:37:15 +0000 (17:37 +0100)]
[reland][ADT] add ConcurrentHashtable class.

ConcurrentHashTable - is a resizeable concurrent hashtable.
The range of resizings is limited up to x2^31.
The hashtable allows only concurrent insertions.

Concurrent hashtable is necessary for the D96035 patch.

Reviewed By: JDevlieghere

Differential Revision: https://reviews.llvm.org/D132455

15 months ago[mlir][AsmPrinter] Fallback to using qualified printer if unqualified output is empty
Markus Böck [Mon, 27 Mar 2023 12:13:44 +0000 (14:13 +0200)]
[mlir][AsmPrinter] Fallback to using qualified printer if unqualified output is empty

The current behaviour of always writing the unqualified form of an attribute or type is problematic for any type or attribute that might output an empty string, making it impossible to parse and therefore roundtrip. This is commonly the case for any types or attributes with optional parameters. One would have to currently woarkaround the issue by either changing ones syntax to not be completetly empty or by explicitly using `qualified` in ALL ops using that type or attribute.

This patch fixes that issue by simply checking whether anything was written to the output. In the case there wasn't, it simply falls back to using the normal printer with the dialect prefix. This also makes the default of unqualified printing always correct and safe. The implementation could theoretically still be tricked if the user were to print just a space or similar. I'd argue this'd be user error and not worth handling.

Fixes https://github.com/llvm/llvm-project/issues/61701

Differential Revision: https://reviews.llvm.org/D146944

15 months ago[libc] Enable integration tests targeting NVIDIA GPUs
Joseph Huber [Sat, 25 Mar 2023 02:00:25 +0000 (21:00 -0500)]
[libc] Enable integration tests targeting NVIDIA GPUs

This patch adds the necessary build infrastructure to build and run the
integration tests on NVIDIA GPUs. The NVIDIA `nvlink` linker utility is
what is ultimately used to combine these files into a single executable
image. Unfortunately, their tool does not support static libraries. So
we need to link with every object directly instead. This could be solved
by impelementing a "wrapper" utility around `nvlink` like we used to use
for OpenMP. But for now this should be sufficient.

Reviewed By: sivachandra

Differential Revision: https://reviews.llvm.org/D146861

15 months ago[RISCV] Remove UpperBound from SchedMxList and create WorstCaseBehavior records
Michael Maitland [Sat, 25 Mar 2023 00:16:41 +0000 (17:16 -0700)]
[RISCV] Remove UpperBound from SchedMxList and create WorstCaseBehavior records

Prior to this patch, UpperBound refered to the largest LMUL
supported. Instructions used UpperBound to assign the worst
case behaviour to records, since Instructions are not LMUL
specific. This forced the largest LMUL to have the worst case
behavior, even if that wasn't true for a subtarget.

Now that SchedWrites, SchedReads, WriteRes, and ReadAdvances
are created for (name, LMUL) pairs and (name, LMUL, SEW)
tuples, it becomes even less clear which pair should correspond
to the worst case behavior. Additionally, it no longer makes sense
for the UpperBound to belong to LMUL list and not to the SEW list.

Instead of creating a special UpperBound LMUL and an UpperBound
SEW, this patch renames UpperBound to WorstCaseBehavior, removes
it from the SchedMxList, and defines a WorstCaseBehavior SchedWrite,
SchedRead, WriteRes, and ReadAdvance for each name.

This gives subtargets the ability to describe the worst case behavior
of a record without forcing it to be the largest LMUL or the smallest
SEW.

Differential Revision: https://reviews.llvm.org/D146855

15 months ago[MC] Quick fix for Windows build failures after D142218
Jay Foad [Mon, 27 Mar 2023 12:29:00 +0000 (13:29 +0100)]
[MC] Quick fix for Windows build failures after D142218

15 months agoRevert "[RISCV] Remove UpperBound from SchedMxList and create WorstCaseBehavior records"
Michael Maitland [Mon, 27 Mar 2023 13:26:25 +0000 (06:26 -0700)]
Revert "[RISCV] Remove UpperBound from SchedMxList and create WorstCaseBehavior records"

This reverts commit 74c0bd240499683078d4f15d31de690950da8231.

15 months ago[PowerPC][GISel] Add initial GlobalISel support for vector functions.
Amy Kwan [Sun, 26 Mar 2023 04:04:53 +0000 (23:04 -0500)]
[PowerPC][GISel] Add initial GlobalISel support for vector functions.

This patch adds the initial support for vector functions and register banks
within GlobalISel. With this patch, we are able to support simple functions that
return vectors, and also functions that perform simple operations.

This patch also:
- Legalizes vector types for G_AND, G_OR, G_XOR, G_ADD, G_SUB, G_BITCAST, G_FADD, G_FSUB
- Introduce initial support for bitcasting (that will need to be extended upon)
- Add various different test cases to for test vector support within GlobalISel

Differential Revision: https://reviews.llvm.org/D137785

15 months ago[PowerPC] Implement 64-bit ELFv2 Calling Convention in TableGen (for integers/floats...
Amy Kwan [Sun, 26 Mar 2023 04:03:10 +0000 (23:03 -0500)]
[PowerPC] Implement 64-bit ELFv2 Calling Convention in TableGen (for integers/floats/vectors in registers)

This patch partially implements the parameter passing rules outlined in the
ELFv2 ABI within TableGen. Specifically, it implements the parameter assignment
of integers, floats, and vectors within registers - where the GPR numbering will
be "skipped" depending on the ordering of floats and vectors that appear within
a parameter list.

As we begin to adopt GlobalISel to the PowerPC backend, there is a need for a
TableGen definition that encapsulates the ELFv2 parameter passing rules. Thus,
this patch also changes the default calling convention that is returned within
the ccAssignFnForCall() function used in our GlobalISel implementation, and also
adds some additional testing of the calling convention that is implemented.

Future patches that build on top of this initial TableGen definition will aim to
add more of the ABI complexities, including support for additional types and
also in-memory arguments.

Differential Revision: https://reviews.llvm.org/D137504

15 months ago[JITLink] Drop using namespace
Benjamin Kramer [Mon, 27 Mar 2023 12:19:18 +0000 (14:19 +0200)]
[JITLink] Drop using namespace

I don't think these are holding their weight and cause nasty to debug
compiler issues with some versions of MSVC as seen in
https://reviews.llvm.org/D144083#4220866

Just qualify the types with the full name.

15 months ago[RISCV] Remove UpperBound from SchedMxList and create WorstCaseBehavior records
Michael Maitland [Sat, 25 Mar 2023 00:16:41 +0000 (17:16 -0700)]
[RISCV] Remove UpperBound from SchedMxList and create WorstCaseBehavior records

Prior to this patch, UpperBound refered to the largest LMUL
supported. Instructions used UpperBound to assign the worst
case behaviour to records, since Instructions are not LMUL
specific. This forced the largest LMUL to have the worst case
behavior, even if that wasn't true for a subtarget.

Now that SchedWrites, SchedReads, WriteRes, and ReadAdvances
are created for (name, LMUL) pairs and (name, LMUL, SEW)
tuples, it becomes even less clear which pair should correspond
to the worst case behavior. Additionally, it no longer makes sense
for the UpperBound to belong to LMUL list and not to the SEW list.

Instead of creating a special UpperBound LMUL and an UpperBound
SEW, this patch renames UpperBound to WorstCaseBehavior, removes
it from the SchedMxList, and defines a WorstCaseBehavior SchedWrite,
SchedRead, WriteRes, and ReadAdvance for each name.

This gives subtargets the ability to describe the worst case behavior
of a record without forcing it to be the largest LMUL or the smallest
SEW.

Differential Revision: https://reviews.llvm.org/D146855

15 months agoRevert "[C++20][Modules] Introduce an implementation module."
Mitch Phillips [Mon, 27 Mar 2023 11:33:53 +0000 (04:33 -0700)]
Revert "[C++20][Modules] Introduce an implementation module."

This reverts commit c6e9823724ef6bdfee262289ee34d162db436af0.

Reason: Broke the ASan buildbots, see https://reviews.llvm.org/D126959
(the original phabricator review) for more info.

15 months agoRevert "Silence unused variable warning in NDEBUG builds"
Mitch Phillips [Mon, 27 Mar 2023 11:33:09 +0000 (04:33 -0700)]
Revert "Silence unused variable warning in NDEBUG builds"

This reverts commit 8c7c1f11ffaacf762e612c65440fd2cbb58ee426.

Reason: Dependent change https://reviews.llvm.org/D126959 broke the ASan
buildbots. See that phabricator review for more comments.

15 months ago[llvm-exegesis] Improve test flexibility by using available_features
Pavel Kosov [Mon, 27 Mar 2023 11:43:08 +0000 (14:43 +0300)]
[llvm-exegesis] Improve test flexibility by using available_features

Currently, all llvm-exegesis regression tests are target-specific and
they are organized so that the entire subdirectories of
llvm/test/tools/llvm-exegesis are enabled or disabled as a whole.

After the commit e0ad2af691 ("Skip codegen" dry-run mode), at least two
kinds of tests are possible:
1) ensure that a snippet can be *generated* for the particular opcode
2) tests involving actual snippet execution inside the llvm-exegesis
   process

Thus, for the particular target subdirectory, some tests should run only
on the particular host architecture and other tests only need the target
being among LLVM_TARGETS_TO_BUILD.

This commit defines a bunch of exegesis-specific features that can be
referenced in REQUIRES, UNSUPPORTED, etc. on a test-by-test basis.

~~

Huawei RRI

Reviewed By: courbet

Differential Revision: https://reviews.llvm.org/D146302

15 months agoAdd Release Note for -mcode-object-v3 removal
pvanhout [Mon, 27 Mar 2023 11:27:31 +0000 (13:27 +0200)]
Add Release Note for -mcode-object-v3 removal

Reviewed By: aaron.ballman

Differential Revision: https://reviews.llvm.org/D146945

15 months ago[Flang][OpenMP] Added TODO checks for unsupported map types
Akash Banerjee [Tue, 21 Mar 2023 12:12:11 +0000 (12:12 +0000)]
[Flang][OpenMP] Added TODO checks for unsupported map types

This patch adds TODO checks for unspported types in the map clause for OpenMP Target directives.

Example of unsupported code:

implicit none
character(len=10) :: str1, str2(5,5)

type t
  character(len=10) :: str1, str2(5,5)
end type t
type(t) :: v

!$omp target enter data map(to: str2(2,5))
!$omp target enter data map(to: v%str1)
!$omp target enter data map(to: v%str2)
!$omp target enter data map(to: v%str2(1,2))

end

Differential Revision: https://reviews.llvm.org/D146292

15 months ago[TableGen][GlobalISel] Support EXTRACT_SUBREGs for ComplexPattern suboperands.
Ivan Kosarev [Mon, 27 Mar 2023 10:37:16 +0000 (11:37 +0100)]
[TableGen][GlobalISel] Support EXTRACT_SUBREGs for ComplexPattern suboperands.

This makes it possible to write GlobalISel patterns generating
EXTRACT_SUBREG instructions applied to suboperands of ComplexPattern
operands. Currently, TableGen complains that such operands are not
declared in matcher.

Reviewed By: arsenm

Differential Revision: https://reviews.llvm.org/D146800

15 months ago[mlir][ClangTidy] Remove redundant return (NFC).
Adrian Kuegel [Mon, 27 Mar 2023 11:03:53 +0000 (13:03 +0200)]
[mlir][ClangTidy] Remove redundant return (NFC).

15 months ago[MLIR][LLVM] Extend argument/result/function attribute support.
Johannes de Fine Licht [Mon, 27 Mar 2023 10:38:29 +0000 (12:38 +0200)]
[MLIR][LLVM] Extend argument/result/function attribute support.

Use a disallowlist approach to deselect attributes on arguments and
return values that are not yet supported, but allow the rest to be
present when inlining.

To achieve fast lookup in `isLegalToInline`, the set of unsupported
function attributes is cached on the inliner interface object during
construction. By using a set of StringAttrs this should just amount to a
pointer lookup.

Depends on D146633

Reviewed By: gysit

Differential Revision: https://reviews.llvm.org/D146729

15 months ago[mlir][Linalg] NFC - Move transform utilities related to subcomputation inference...
Nicolas Vasilache [Mon, 27 Mar 2023 10:48:21 +0000 (03:48 -0700)]
[mlir][Linalg] NFC - Move transform utilities related to subcomputation inference to Linalg/Utils

15 months ago[MC] Store operand info immediately after the TargetInsts table. NFC.
Jay Foad [Thu, 12 Jan 2023 21:02:23 +0000 (21:02 +0000)]
[MC] Store operand info immediately after the TargetInsts table. NFC.

This shrinks MCInstrDesc (and hence the whole TargetInsts table) because
we can store a 16-bit offset value to access the operands info, instead
of a pointer. This also reduces the number of relocs that need to be
applied when LLVM is compiled as position-independent code.

Differential Revision: https://reviews.llvm.org/D142219

15 months ago[MC] Store implicit ops immediately after the TargetInsts table. NFC.
Jay Foad [Thu, 12 Jan 2023 07:38:13 +0000 (07:38 +0000)]
[MC] Store implicit ops immediately after the TargetInsts table. NFC.

This shrinks MCInstrDesc (and hence the whole TargetInsts table) because
we can store a 16-bit offset value to access the implicit operands,
instead of a pointer. This also reduces the number of relocs that need
to be applied when LLVM is compiled as position-independent code.

Differential Revision: https://reviews.llvm.org/D142218

15 months ago[clang] source range of variable template specialization should include initializer
Tomasz Kamiński [Mon, 27 Mar 2023 08:13:20 +0000 (10:13 +0200)]
[clang] source range of variable template specialization should include initializer

This patch adjust the getSourceRange() for the
VarTemplateSpecializationDecl and VarTemplatePartialSpecializationDecl,
such that the initializer is included if present:
```
template<typename T>
T temp = 1;

template<> double temp<double> = 1;
```
This patch makes it consistent with the behavior of
non-template variables with initializers and restores
behavior that was present before https://reviews.llvm.org/D139705.

n case, when the initializer is not present we still
include the template arguments in the source range,
which was required for fixing zero-initialization fix-it.

Reviewed By: erichkeane

Differential Revision: https://reviews.llvm.org/D146733

15 months ago[AMDGPU] Remove `use-legacy-divergence-analysis`
pvanhout [Mon, 27 Mar 2023 08:15:15 +0000 (10:15 +0200)]
[AMDGPU] Remove `use-legacy-divergence-analysis`

This option has been a no-op since we switched to UA.

Reviewed By: foad

Differential Revision: https://reviews.llvm.org/D146937

15 months agoRevert "[mlir] Implement pass utils for 1:N type conversions."
Ingo Müller [Mon, 27 Mar 2023 09:23:57 +0000 (09:23 +0000)]
Revert "[mlir] Implement pass utils for 1:N type conversions."

This reverts commit 9c4611f9c7a7055b18f0a30a4c9074b9917e4ab0.

15 months ago[LV][ARM][AArch64] Add multi-exit Loop Vectorizer tests. NFC
David Green [Mon, 27 Mar 2023 09:21:24 +0000 (10:21 +0100)]
[LV][ARM][AArch64] Add multi-exit Loop Vectorizer tests. NFC

These are useful to test with the various predication schemes available on
different targets.

15 months ago[documentation] Fix Typos
Priyanshi Agarwal [Mon, 27 Mar 2023 08:56:35 +0000 (09:56 +0100)]
[documentation] Fix Typos

Fixes https://github.com/llvm/llvm-project/issues/56747

Patch By: ipriyanshi1708

Differential Revision: https://reviews.llvm.org/D146644

15 months ago[mlir] Implement pass utils for 1:N type conversions.
Ingo Müller [Tue, 21 Feb 2023 11:21:25 +0000 (11:21 +0000)]
[mlir] Implement pass utils for 1:N type conversions.

The current dialect conversion does not support 1:N type conversions.
This commit implements a (poor-man's) dialect conversion pass that does
just that. To keep the pass independent of the "real" dialect conversion
infrastructure, it provides a specialization of the TypeConverter class
that allows for N:1 target materializations, a specialization of the
RewritePattern and PatternRewriter classes that automatically add
appropriate unrealized casts supporting 1:N type conversions and provide
converted operands for implementing subclasses, and a conversion driver
that applies the provided patterns and replaces the unrealized casts
that haven't folded away with user-provided materializations.

The current pass is powerful enough to express many existing manual
solutions for 1:N type conversions or extend transforms that previously
didn't support them, out of which this patch implements call graph type
decomposition (which is currently implemented with a ValueDecomposer
that is only used there).

The goal of this pass is to illustrate the effect that 1:N type
conversions could have, gain experience in how patterns should be
written that achieve that effect, and get feedback on how the APIs of
the dialect conversion should be extended or changed to support such
patterns. The hope is that the "real" dialect conversion eventually
supports such patterns, at which point, this pass could be removed
again.

Reviewed By: springerm

Differential Revision: https://reviews.llvm.org/D144469

15 months ago[LoopVectorize] Don't tail-fold for scalable VFs when there is no scalar tail
David Sherwood [Wed, 15 Mar 2023 14:25:21 +0000 (14:25 +0000)]
[LoopVectorize] Don't tail-fold for scalable VFs when there is no scalar tail

Currently in LoopVectorize we avoid tail-folding if we can
prove the trip count is always a multiple of the maximum
fixed-width VF. This works because we know the vectoriser
only ever chooses a VF that is a power of 2. However, if
we are also considering scalable VFs then we conservatively
bail out of the optimisation because we don't know the value
of vscale, which could be an odd or prime number, etc.

This patch tries to enable the same optimisation for scalable
VFs by asking if vscale is known to be a power of 2. If so,
we can then query the maximum value of vscale and use the same
logic as we do for fixed-width VFs. I've also added a new TTI
hook called isVScaleKnownToBeAPowerOfTwo that does the same
thing as the existing TargetLowering hook.

Differential Revision: https://reviews.llvm.org/D146199