platform/upstream/llvm.git
18 months ago[lldb][DataFormatter] Fix libcxx std::deque formatter for references and pointers
Michael Buch [Mon, 17 Apr 2023 14:04:57 +0000 (15:04 +0100)]
[lldb][DataFormatter] Fix libcxx std::deque formatter for references and pointers

(Addresses GH#62153)

The `SBType` APIs to retrieve details about template arguments,
such as `GetTemplateArgumentType` or `GetTemplateArgumentKind`
don't "desugar" LValueReferences/RValueReferences or pointers.
So when we try to format a `std::deque&`, the python call to
`GetTemplateArgumentType` fails to get a type, leading to
an `element_size` of `0` and a division-by-zero python exception
(which gets caught by the summary provider silently). This leads
to the contents of such `std::deque&` to be printed incorrectly.

This patch dereferences the reference/pointer before calling
into the above SBAPIs.

**Testing**

* Add API test

Differential Revision: https://reviews.llvm.org/D148531

18 months ago[gn build] Port b0b2b2e047ca
LLVM GN Syncbot [Mon, 17 Apr 2023 20:21:50 +0000 (20:21 +0000)]
[gn build] Port b0b2b2e047ca

18 months ago[arcanist] rm -f for write protected temp files
Nick Desaulniers [Mon, 17 Apr 2023 20:18:16 +0000 (13:18 -0700)]
[arcanist] rm -f for write protected temp files

When running `arc diff` on a newly added file,
utils/arcanist/clang-format.sh is run. For new files there is no change,
but the script will bail. Without the use of the -f flag, we get
interactive prompts like:
rm: remove write-protected regular file '/tmp/tmp.ReMybrBw35'?

Reviewed By: MaskRay

Differential Revision: https://reviews.llvm.org/D148555

18 months ago[COFF] Add MC support for emitting IMAGE_WEAK_EXTERN_ANTI_DEPENDENCY symbols
Eli Friedman [Mon, 17 Apr 2023 20:15:46 +0000 (13:15 -0700)]
[COFF] Add MC support for emitting IMAGE_WEAK_EXTERN_ANTI_DEPENDENCY symbols

This is mostly useful for ARM64EC, which uses such symbols extensively.

One interesting quirk of ARM64EC is that we need to be able to emit weak
symbols that point at each other (so if either symbol is defined
elsewhere, both symbols point at the definition).  This required a few
changes to the way we handle weak symbols on Windows.

Differential Revision: https://reviews.llvm.org/D145208

18 months agoRevert "[ADT] add StringViewExtras llvm::starts_with for std::string_view"
Nick Desaulniers [Mon, 17 Apr 2023 20:10:16 +0000 (13:10 -0700)]
Revert "[ADT] add StringViewExtras llvm::starts_with for std::string_view"

This reverts commit d6d30dd9590b41fbcdc88b43767b27a31bfb1c73.

These helpers were added to be used by LLVMDemangle, which would be a
layering violation for that to depend on LLVMSupport. Undo this commit,
and I'll these these directly in LLVMDemangle.

Reviewed By: MaskRay

Differential Revision: https://reviews.llvm.org/D148547

18 months ago[scudo] Tune FuchsiaConfig for RiscV
Caslyn Tonelli [Sun, 16 Apr 2023 19:07:09 +0000 (19:07 +0000)]
[scudo] Tune FuchsiaConfig for RiscV

Reduce `PrimaryRegionSizeLog` to 28U to be compatible with a 38bit user
address space on the fuchsia-riscv platform. `PrimaryGroupSizeLog` is
reduced to 19 to preserve 512 BatchGroups per region.

This change can be tested on Fuchsia with:

```
fx set --auto-dir bringup.riscv64 --with //bundles:boot_tests \
fx build bundles:boot_tests \
fx run-boot-test --args={-s,1} boot-libc-unittests \
--cmdline='--gtest_filter=-*DeathTest*:PthreadGetSet*:ScudoSecondaryTest*'
```
The gtest filter ignores pthread and death tests due to non-scudo
related issues on fuchsia-riscv (ScudoSecondaryTest includes a death
check).

Related Ticket: https://fxbug.dev/125263

Differential Revision: https://reviews.llvm.org/D148475

18 months ago[InstCombine] support fold select(X|Y,X|Y,X) to X|Y
Congcong Cai [Mon, 17 Apr 2023 20:06:59 +0000 (22:06 +0200)]
[InstCombine] support fold select(X|Y,X|Y,X) to X|Y

Fixed: https://github.com/llvm/llvm-project/issues/62113
Add addtional check in `visitSelectInst` to:
1. match `select(X|Y==0, X, X|Y)` and replaced with `X|Y`
2. match `select(X&Y==-1, X, X&Y)` and replaced with `X&Y`

alive proof:
https://alive2.llvm.org/ce/z/4qHmv-
https://alive2.llvm.org/ce/z/c2MBGy

Reviewed By: nikic

Differential Revision: https://reviews.llvm.org/D148275

18 months ago[InstCombine][NFC] Add tests for simplifying select(X|Y,X|Y,X) to X|Y
Congcong Cai [Mon, 17 Apr 2023 20:06:43 +0000 (22:06 +0200)]
[InstCombine][NFC] Add tests for simplifying select(X|Y,X|Y,X) to X|Y

Reviewed By: nikic

Differential Revision: https://reviews.llvm.org/D148562

18 months ago[libc] Partial revert of "[libc] Update supported riscv libs"
Mikhail R. Gadelha [Mon, 17 Apr 2023 19:45:46 +0000 (16:45 -0300)]
[libc] Partial revert of "[libc] Update supported riscv libs"

Partial revert of commit ee917996708c936dd7f275de06b7895f5e3c11ec.

I can't reproduce the error locally, so let's revert it to keep the bots
happy.

18 months ago[mlir][openacc][NFC] Use assembly format for acc.loop
Valentin Clement [Mon, 17 Apr 2023 19:44:36 +0000 (12:44 -0700)]
[mlir][openacc][NFC] Use assembly format for acc.loop

Use the assembly format with custom parser/printer
for specific clauses instead of a full custom parser/printer.

Reviewed By: PeteSteinfeld

Differential Revision: https://reviews.llvm.org/D148391

18 months ago[lldb] fix build issue on MSVC because of missing byte-swap builtins
Ashay Rane [Mon, 17 Apr 2023 16:17:43 +0000 (11:17 -0500)]
[lldb] fix build issue on MSVC because of missing byte-swap builtins

The `__builtin_bswap{32,64}()` builtins (introduced in commit e07a421d)
are missing from MSVC, which causes build errors when compiling LLDB on
Windows (tested with MSVC 19.34.31943.0).  This patch replaces the
builtins with either MSVC's `_byteswap_u{long,64}()` or the original
builtins, or the `bswap_{32,64}()` functions from byteswap.h, depending
on which ones are available.

Reviewed By: bulbazord

Differential Revision: https://reviews.llvm.org/D148541

18 months agoRevert "[VPlan] Switch to checking sinking legality for recurrences in VPlan."
Manoj Gupta [Mon, 17 Apr 2023 18:39:07 +0000 (11:39 -0700)]
Revert "[VPlan] Switch to checking sinking legality for recurrences in VPlan."

This reverts commit 7fc0b3049df532fce726d1ff6869a9f6e3183780.

Causes a clang hang when building xz utils, github issue #62187.

18 months ago[libc] Update supported riscv libs
Mikhail R. Gadelha [Mon, 17 Apr 2023 19:13:14 +0000 (16:13 -0300)]
[libc] Update supported riscv libs

This patch updates the list of supported libs according to the latest changes to x86_64

Reviewed By: sivachandra

Differential Revision: https://reviews.llvm.org/D147961

18 months agogithub: Add manual workflow to build and upload release binaries
Jack Grigg [Mon, 17 Apr 2023 18:59:56 +0000 (11:59 -0700)]
github: Add manual workflow to build and upload release binaries

Reviewed By: kwk

Differential Revision: https://reviews.llvm.org/D143535

18 months ago[lldb] Fix TestObjectFileJSON and TestSymbolFileJSON
Jonas Devlieghere [Mon, 17 Apr 2023 18:58:27 +0000 (11:58 -0700)]
[lldb] Fix TestObjectFileJSON and TestSymbolFileJSON

 - Separate the two test and only have TestSymbolFileJSON rely on strip.
 - Use different file names to make sure LLDB reloads the module.

This should address all the post commit review from D148062.

18 months ago[JITLink][ELF] Check that ELF buffer has at least ELF::EI_NIDENT bytes.
Lang Hames [Mon, 17 Apr 2023 17:58:29 +0000 (10:58 -0700)]
[JITLink][ELF] Check that ELF buffer has at least ELF::EI_NIDENT bytes.

This allows the subsequent code to access the E_IDENT fields, which will be
used in the upcoming JITLink ppc64 backend (https://reviews.llvm.org/D148192).

18 months ago[nfc][llvm] Replace pointer cast functions in PointerUnion by llvm casting functions.
Shraiysh Vaishay [Sun, 16 Apr 2023 02:16:16 +0000 (21:16 -0500)]
[nfc][llvm] Replace pointer cast functions in PointerUnion by llvm casting functions.

This patch replaces the uses of PointerUnion.is function by llvm::isa,
PointerUnion.get function by llvm::cast, and PointerUnion.dyn_cast by
llvm::dyn_cast_if_present. This is according to the FIXME in
the definition of the class PointerUnion.

This patch does not remove them as they are being used in other
subprojects.

Reviewed By: mehdi_amini

Differential Revision: https://reviews.llvm.org/D148449

18 months ago[clang-format] CSharp don't allow there not to be a space between `is` and `[`
mydeveloperday [Mon, 17 Apr 2023 18:38:53 +0000 (19:38 +0100)]
[clang-format] CSharp don't allow there not to be a space between `is` and `[`

as `is` is a keyword in C# ensure there is always a space before the `[` regardless of `SpaceBeforeSquareBrackets` setting

Fixes: #61965

https://github.com/llvm/llvm-project/issues/61965

Reviewed By: owenpan, HazardyKnusperkeks

Differential Revision: https://reviews.llvm.org/D148472

18 months agoRevert "[lldb] Add support for MSP430 in LLDB."
Anton Korobeynikov [Mon, 17 Apr 2023 18:29:59 +0000 (11:29 -0700)]
Revert "[lldb] Add support for MSP430 in LLDB."

This reverts commit 82c02b733c7736507a41a26bebd37d3f8e88bd4e.

Apparently, the original patch was not rebased onto `main

18 months ago[mlir][vector] Fix integer promotion type mismatch
Lei Zhang [Mon, 17 Apr 2023 18:23:27 +0000 (11:23 -0700)]
[mlir][vector] Fix integer promotion type mismatch

We need to create a new type with transposed shape after
transposing the operand in `CanonicalizeContractMatmulToMMT`.

Reviewed By: kuhar, dcaballe

Differential Revision: https://reviews.llvm.org/D148470

18 months agoRevert "[ADT] Apply fixes from modernize-type-traits (NFC)"
Jonas Devlieghere [Mon, 17 Apr 2023 18:26:56 +0000 (11:26 -0700)]
Revert "[ADT] Apply fixes from modernize-type-traits (NFC)"

This reverts commit 9395cf063a013003704118deccf7633533170a5b because it
breaks the modules build:

  STLFunctionalExtras.h:54:3: error:
  'llvm::function_ref<std::__1::optional<std::__1::basic_string<char>>
  (llvm::StringRef, llvm::StringRef)>::function_ref' from module
  'LLVM_Utils.ADT.STLFunctionalExtras' is not present in definition of
  'llvm::function_ref<std::__1::optional<std::__1::basic_string<char>>
  (llvm::StringRef, llvm::StringRef)>' in module
  'LLVM_Utils.ADT.STLFunctionalExtras'

https://green.lab.llvm.org/green/view/LLDB/job/lldb-cmake/53806/

18 months ago[lldb] Add support for MSP430 in LLDB.
Anton Korobeynikov [Mon, 17 Apr 2023 17:56:01 +0000 (10:56 -0700)]
[lldb] Add support for MSP430 in LLDB.

Add MSP430 to the list of available targets, implement MSP430 ABI, add support for debugging targets with 16-bit address size.

The update is intended for use with MSPDebug, a GDB server implementation for MSP430.

Reviewed By: bulbazord, DavidSpickett

Differential Revision: https://reviews.llvm.org/D146965

18 months ago[clang][dataflow] Drop optional model's dependency on libc++ internals.
Yitzhak Mandelbaum [Fri, 14 Apr 2023 21:00:24 +0000 (21:00 +0000)]
[clang][dataflow] Drop optional model's dependency on libc++ internals.

Adjusts the matchers in the optional model to avoid dependency on internal
implementation details of libc++'s `std::optional`. In the process, factors out
the code to check the name of these types so that it's shared throughout.

Differential Revision: https://reviews.llvm.org/D148377

18 months ago[clang][dataflow] Refine matching of optional types to anchor at top level.
Yitzhak Mandelbaum [Fri, 14 Apr 2023 14:21:41 +0000 (14:21 +0000)]
[clang][dataflow] Refine matching of optional types to anchor at top level.

This patch refines the matching of the relevant optional types to anchor on the
global namespace. Previously, we could match anything with the right name
(e.g. `base::Optional`) even if nested within other namespaces. This over
matching resulted in an assertion violation when _different_ `base::Optional`
was encountered nested inside another namespace.

Fixes issue #57036.

Differential Revision: https://reviews.llvm.org/D148344

18 months ago[flang] Complex numbers in function arguments on Windows
Markus Mützel [Mon, 17 Apr 2023 16:15:54 +0000 (09:15 -0700)]
[flang] Complex numbers in function arguments on Windows

Function arguments or return values that are complex floating point values
aren't correctly lowered for Windows x86 32-bit and 64-bit targets.
See: https://github.com/llvm/llvm-project/issues/61976

Add targets that are specific for these platforms and OS.

With thanks to @mstorsjo for pointing out the fix.

Reviewed By: vzakhari

Differential Revision: https://reviews.llvm.org/D147768

18 months agoSecond try: [clang][cmake] Add options to pass in vcs repo and revision info
Han Zhu [Mon, 10 Apr 2023 22:40:15 +0000 (15:40 -0700)]
Second try: [clang][cmake] Add options to pass in vcs repo and revision info

V2: Fix cmake error when -DLLVM_APPEND_VC_REV=OFF

Clang may be built in an environment where Git is not available. In our case,
Clang is part of a larger monorepo which is not Git-based, and
GenerateVersionFromVCS was not able to get source info.

Provide options to pass in repo and revision info from cmake.
```
cmake \
  -DCLANG_VC_REPOSITORY=abc://repo.url.com \
  -DCLANG_VC_REVISION=abcd1234 \
  ...
```
This would allow us to prepare the source info beforehand and pass it to the
clang binary.

Differential Revision: https://reviews.llvm.org/D148262

18 months ago[memprof] Print out profile build ids in the error message.
Snehasish Kumar [Fri, 14 Apr 2023 05:12:09 +0000 (05:12 +0000)]
[memprof] Print out profile build ids in the error message.

When no --profiled-binary flag is provided we can print out the build
ids of the modules in the profile. This can help the user fetch the
correct binary from e.g. remote object store.

Reviewed By: tejohnson

Differential Revision: https://reviews.llvm.org/D148301

18 months ago[libc][NFC] Remove the unused FDReader testutil.
Siva Chandra Reddy [Sun, 16 Apr 2023 07:28:41 +0000 (07:28 +0000)]
[libc][NFC] Remove the unused FDReader testutil.

Differential Revision: https://reviews.llvm.org/D148454

18 months ago[mlir][sparse] enable more sparse convolution kernels.
Peiming Liu [Tue, 4 Apr 2023 18:41:00 +0000 (18:41 +0000)]
[mlir][sparse] enable more sparse convolution kernels.

Reviewed By: aartbik

Differential Revision: https://reviews.llvm.org/D147670

18 months ago[FuncSpec] Consider constant struct arguments when specializing.
Alexandros Lamprineas [Mon, 20 Mar 2023 12:17:23 +0000 (12:17 +0000)]
[FuncSpec] Consider constant struct arguments when specializing.

Optionally enabled just like integer and floating point arguments.

Differential Revision: https://reviews.llvm.org/D145374

18 months ago[scudo] Descriptive Fucshia errors
Caslyn Tonelli [Sun, 16 Apr 2023 02:03:58 +0000 (02:03 +0000)]
[scudo] Descriptive Fucshia errors

`dieOnError` in fuchsia.cpp takes the syscall name, status code, and
relevant size parameter to describe what failed, the reason, and size
context. Several different Fuchsia syscalls can fail underlying Scudo's
map operation and gets hidden if using `dieOnMapUnmapError`.

The reason for this change is to provide an error status code and more
helpful information to debug failures.

18 months ago[libc++][NFC] Remove some dead code in common_type
Nikolas Klauser [Sun, 16 Apr 2023 14:16:34 +0000 (16:16 +0200)]
[libc++][NFC] Remove some dead code in common_type

Reviewed By: #libc, Mordante

Spies: Mordante, libcxx-commits

Differential Revision: https://reviews.llvm.org/D148466

18 months ago[Docs] Added RTTI, Run-time Type Information
John-Earnshaw [Mon, 17 Apr 2023 17:11:20 +0000 (10:11 -0700)]
[Docs] Added RTTI, Run-time Type Information

Reviewed By: MaskRay

Differential Revision: https://reviews.llvm.org/D148538

18 months agoRevert "[clang][cmake] Add options to pass in vcs repo and revision info"
Han Zhu [Mon, 17 Apr 2023 17:08:34 +0000 (10:08 -0700)]
Revert "[clang][cmake] Add options to pass in vcs repo and revision info"

This reverts commit 1f5e737fc135bf991889a1364b8f8c5edc3953d2.

18 months ago[BPF] Make sure ALU32 feature is set in MCSubtargetInfo for mcpu=v3
Eduard Zingerman [Tue, 11 Apr 2023 18:08:45 +0000 (21:08 +0300)]
[BPF] Make sure ALU32 feature is set in MCSubtargetInfo for mcpu=v3

`BPF.td` is used to generate (among other things) `MCSubtargetInfo`
setup function for BPF target.
Specifically, the `BPFGenSubtargetInfo.inc` file:

    enum {
      ALU32 = 0,
      ...
    };
    ...
    extern const llvm::SubtargetSubTypeKV BPFSubTypeKV[] = {
      { "generic", { { { 0x0ULL, ... } } }, ... },
      { "probe",   { { { 0x0ULL, ... } } }, ... },
      { "v1",      { { { 0x0ULL, ... } } }, ... },
      { "v2",      { { { 0x0ULL, ... } } }, ... },
      { "v3",      { { { 0x1ULL, ... } } }, ... },
    };
    ...
    static inline MCSubtargetInfo *createBPFMCSubtargetInfoImpl(...) {
      return new BPFGenMCSubtargetInfo(..., BPFSubTypeKV, ...);
    }

The `SubtargetSubTypeKV` is defined in `MCSubtargetInfo.h` as:

    /// Used to provide key value pairs for feature and CPU bit flags.
    struct SubtargetSubTypeKV {
      const char *Key;                      ///< K-V key string
      FeatureBitArray Implies;              ///< K-V bit mask
      FeatureBitArray TuneImplies;          ///< K-V bit mask
      const MCSchedModel *SchedModel;
      ...
    }

The first bit array specifies features enabled by default for a
specific CPU. This commit makes sure that this information is
communicated to `tablegen` and correct `BPFSubTypeKV` table is
generated. This allows tools like `objdump` to detect available
features when `--mcpu` flag is specified.

Differential Revision: https://reviews.llvm.org/D148037

18 months ago[DebugInfo] Fix file path separator when targeting windows.
Zequan Wu [Thu, 30 Mar 2023 19:01:21 +0000 (15:01 -0400)]
[DebugInfo] Fix file path separator when targeting windows.

This fixes two problems:

1. When crossing compiling for windows on linux, source file path in debug info is concatenated with directory by host native separator ('/'). For windows local build, they are concatenated by '\'. This causes non-determinism bug.

  The solution here is to let `LangOptions.UseTargetPathSeparator` to control if we should use host native separator or not.

2. Objectfile path in CodeView also uses host native separator when generated.

 It's fixed by changing the path separator in `/Fo` to '\' if the path is not an absolute path when adding the `-object-file-name=` flag.

Reviewed By: hans

Differential Revision: https://reviews.llvm.org/D147256

18 months ago[Parse] Remove TimeTraceScope for "ParseTemplate"
Fangrui Song [Mon, 17 Apr 2023 16:57:53 +0000 (09:57 -0700)]
[Parse] Remove TimeTraceScope for "ParseTemplate"

Fix https://github.com/llvm/llvm-project/issues/56554

```
#include "1.h"
#include "2.h"
int foo();
```

Suppose that 1.h ends with a template function. When parsing the function,
the `ParseFunctionDefinition` call after the TimeTraceScope object
may consume a `r_brace` token and lex the end of file (1.h), resulting
in an ExitFile event in SemaPPCallbacks::FileChanged. This event will call `llvm::timeTraceProfilerEnd();`,
which incorrectly ends "ParseTemplate" instead of "Source" (1.h).
Once 2.h has been fully parsed, the destructor of 1.h's TimeTraceScope object
will end "Source" (1.h).

This behavior erroneously extends the end of "Source" (1.h), which makes
"Source" (2.h) appear to be nested inside "Source" (1.h). This bug is difficult
to fix correctly in an elegant way, and we have two options: either end
"ParseTemplate" when ending "Source" (1.h), or delay the ExitFile event.

However, both approaches require complex code. For now, we can remove the
"ParseTemplate" TimeTraceScope. This can be re-added if properly repaired.

Reviewed By: anton-afanasyev

Differential Revision: https://reviews.llvm.org/D148410

18 months ago[mlir][sparse] use macro to build common operations in LoopEmitter
Peiming Liu [Tue, 4 Apr 2023 18:41:00 +0000 (18:41 +0000)]
[mlir][sparse] use macro to build common operations in LoopEmitter

Reviewed By: aartbik

Differential Revision: https://reviews.llvm.org/D147553

18 months agoRevert "[test][sanitizer] Add another stress test for pthread_create"
Thurston Dang [Mon, 17 Apr 2023 16:55:24 +0000 (16:55 +0000)]
Revert "[test][sanitizer] Add another stress test for pthread_create"

Reverting 364884e088d45b162ecb47d093f955a2333eeee1 because it appears the sanitizers are not yet ready to be stressed out (build failures starting from https://lab.llvm.org/buildbot/#/builders/37/builds/21483).

18 months ago[libc] Add special handling for CUDA PTX features
Joseph Huber [Mon, 17 Apr 2023 14:11:09 +0000 (09:11 -0500)]
[libc] Add special handling for CUDA PTX features

The NVIDIA compilation path requires some special options. This is
mostly because compilation is dependent on having a valid CUDA
toolchain. We don't actually need the CUDA toolchain to create the
exported `libcgpu.a` library because it's pure LLVM-IR. However, for
some language features we need the PTX version to be set. This is
normally set by checking the CUDA version, but without one installed it
will fail to build. We instead choose a minimum set of features on the
desired target, inferred from
https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#release-notes
and the PTX refernece for functions like `nanosleep`.

Reviewed By: tianshilei1992

Differential Revision: https://reviews.llvm.org/D148532

18 months ago[clang][cmake] Add options to pass in vcs repo and revision info
Han Zhu [Mon, 10 Apr 2023 22:40:15 +0000 (15:40 -0700)]
[clang][cmake] Add options to pass in vcs repo and revision info

Clang may be built in an environment where Git is not available. In our case,
Clang is part of a larger monorepo which is not Git-based, and
GenerateVersionFromVCS was not able to get source info.

Provide options to pass in repo and revision info from cmake.
```
cmake \
  -DCLANG_VC_REPOSITORY=abc://repo.url.com \
  -DCLANG_VC_REVISION=abcd1234 \
  ...
```
This would allow us to prepare the source info beforehand and pass it to the
clang binary.

Differential Revision: https://reviews.llvm.org/D148262

18 months ago[NFC][Assignment Tracking] Rename flag variable in Options.td
OCHyams [Mon, 17 Apr 2023 16:18:36 +0000 (17:18 +0100)]
[NFC][Assignment Tracking] Rename flag variable in Options.td

Prefix the variable with 'f' to match other 'f' options.

18 months ago[ORC][LLJIT] Use JITLink by default on ELF/x86-64.
Lang Hames [Mon, 17 Apr 2023 15:26:28 +0000 (15:26 +0000)]
[ORC][LLJIT] Use JITLink by default on ELF/x86-64.

This patch switches LLJIT's default JIT linker for ELF/x86-64 from RuntimeDyld
to JITLink.

Most clients should not be affected, but if you were explicitly accessing the
old RTDyldObjectLinkingLayer (e.g. to install JITEventListeners) you will need
to either force use of RuntimeDyld (following the example in
llvm/examples/OrcV2Examples/LLJITWithCustomObjectLinkingLayer), or switch to
using JITLink plugins instead.

18 months ago[mlir][sparse] implement index redution on dense level (for CSR)
Peiming Liu [Fri, 24 Mar 2023 16:16:21 +0000 (16:16 +0000)]
[mlir][sparse] implement index redution on dense level (for CSR)

Reviewed By: aartbik

Differential Revision: https://reviews.llvm.org/D147550

18 months ago[Bazel] Move InstCombineTableGen from InstCombine to AMDGPUCodeGen
NAKAMURA Takumi [Mon, 17 Apr 2023 16:17:38 +0000 (01:17 +0900)]
[Bazel] Move InstCombineTableGen from InstCombine to AMDGPUCodeGen

18 months ago[AArch64][GISel] Add FP16 fcmp lowering
David Green [Mon, 17 Apr 2023 16:22:46 +0000 (17:22 +0100)]
[AArch64][GISel] Add FP16 fcmp lowering

This adds v4f16 and v8f16 lowering for fp16 vector compares. It splits the
getActionDefinitionsBuilder of G_FCMP from G_ICMP, as they are quite different
operations, and adds fp16 vector lowering.

Differential Revision: https://reviews.llvm.org/D147947

18 months ago[flang] REAL(KIND=3) and COMPLEX(KIND=3) descriptors
V Donaldson [Fri, 14 Apr 2023 18:54:07 +0000 (11:54 -0700)]
[flang] REAL(KIND=3) and COMPLEX(KIND=3) descriptors

Update descriptor generation to correctly set the `type` field for
REAL(3) and COMPLEX(3) objects.

18 months ago[libc][NFC] Remove the StreamWrapper class and use the new test logger.
Siva Chandra Reddy [Fri, 14 Apr 2023 22:11:33 +0000 (22:11 +0000)]
[libc][NFC] Remove the StreamWrapper class and use the new test logger.

Reviewed By: lntue

Differential Revision: https://reviews.llvm.org/D148452

18 months agoFix the indentation error.
aabhinavg [Mon, 17 Apr 2023 15:43:37 +0000 (21:13 +0530)]
Fix the indentation error.

Differential Revision: https://reviews.llvm.org/D148535

18 months ago[libc][NFC] Move RandUtils.h to test/src/math.
Siva Chandra Reddy [Mon, 17 Apr 2023 06:41:22 +0000 (06:41 +0000)]
[libc][NFC] Move RandUtils.h to test/src/math.

It is currently used only by math tests so moving it to test/src/math
keeps it closer to where it is used.

Reviewed By: lntue

Differential Revision: https://reviews.llvm.org/D148497

18 months ago[clang] Replace find_executable with shutil.which in creduce script
David Spickett [Mon, 17 Apr 2023 13:40:16 +0000 (13:40 +0000)]
[clang] Replace find_executable with shutil.which in creduce script

distutils is deprecated and shutil.which is the suggested
replacement for this function.

https://peps.python.org/pep-0632/#migration-advice
https://docs.python.org/3/library/shutil.html#shutil.which

which was added in 3.3 (https://docs.python.org/3/library/shutil.html#shutil.which)
and LLVM requires at least 3.6 (https://llvm.org/docs/GettingStarted.html#software).

There is one small differnce here that shutil.which ignores the PATH
when given a path argument. However in this case I think that's actually
the behaviour we want.

Reviewed By: zequanwu

Differential Revision: https://reviews.llvm.org/D148529

18 months ago[AMDGPU] Don't select _SGPR forms of SMEM instructions on GFX9+
Jay Foad [Fri, 10 Mar 2023 14:14:06 +0000 (14:14 +0000)]
[AMDGPU] Don't select _SGPR forms of SMEM instructions on GFX9+

On GFX9+, SMEM instructions have an _SGPR_IMM form which is strictly
more powerful than the _SGPR form. It simplifies codegen if we always
select the _SGPR_IMM form with an immediate offset of 0 instead of the
_SGPR form.

Note that this patch just makes minimal changes to the selection
patterns to prove the concept. Further simplifications are possible to
reduced the number of selection patterns.

On GFX9 the _SGPR form of the Real instruction is still required for
assembly/disassembly but on GFX10+ it can be removed completely.

Differential Revision: https://reviews.llvm.org/D147334

18 months ago[LV] Precommit test for D147963.
Florian Hahn [Mon, 17 Apr 2023 15:19:12 +0000 (16:19 +0100)]
[LV] Precommit test for D147963.

Reduced test case for #58811.

18 months ago[docs][LangRef] Added minor update inside the `frem`. Fix : #61653
aabhinavg [Mon, 17 Apr 2023 14:48:38 +0000 (20:18 +0530)]
[docs][LangRef] Added minor update inside the `frem`. Fix : #61653

Added minor update inside the `frem`. Fix : #61653

Differential Revision: https://reviews.llvm.org/D146900

18 months agoRevert "[mlir][openacc][NFC] Use assembly format for acc.loop"
Valentin Clement [Mon, 17 Apr 2023 15:05:51 +0000 (08:05 -0700)]
Revert "[mlir][openacc][NFC] Use assembly format for acc.loop"

This reverts commit 7960993e4638c6243c4b37f05bb61b080fe13d43.

18 months ago[libcxx] Replace find_executable with shutil.which
David Spickett [Mon, 17 Apr 2023 13:28:22 +0000 (13:28 +0000)]
[libcxx] Replace find_executable with shutil.which

distutils is deprecated and shutil.which is the suggested
replacement for this function.

https://peps.python.org/pep-0632/#migration-advice
https://docs.python.org/3/library/shutil.html#shutil.which

which was added in 3.3 (https://docs.python.org/3/library/shutil.html#shutil.which)
and LLVM requires at least 3.6 (https://llvm.org/docs/GettingStarted.html#software).

Reviewed By: #libc, philnik

Differential Revision: https://reviews.llvm.org/D148527

18 months ago[AArch64] Add a basic pattern for FACGE and FACGT
David Green [Mon, 17 Apr 2023 15:00:57 +0000 (16:00 +0100)]
[AArch64] Add a basic pattern for FACGE and FACGT

These instructions take the absolute of two values and perform a FCMGE/FCMGT
between them. This patch adds some basic patterns for them.

Differential Revision: https://reviews.llvm.org/D147945

18 months ago[clangd] Use all inputs to SystemIncludeExtractor in cache key
Kadir Cetinkaya [Mon, 27 Mar 2023 10:29:22 +0000 (12:29 +0200)]
[clangd] Use all inputs to SystemIncludeExtractor in cache key

Instead of passing in a tooling::CompileCommand into system include
extraction, pass a limited set, whose elements are used as keys.

Also fix the issue around accepting `-isysroot=/foo` which isn't a valid
argument (or the directory should be `=/foo` not `/foo`).

Fixes https://github.com/clangd/clangd/issues/1404
Fixes https://github.com/clangd/clangd/issues/1403

This should also unblock https://reviews.llvm.org/D138546

Differential Revision: https://reviews.llvm.org/D146941

18 months agoPECOFF: consume errors properly
Saleem Abdulrasool [Mon, 17 Apr 2023 14:52:38 +0000 (07:52 -0700)]
PECOFF: consume errors properly

We would not ensure that the error is consumed in the case that logging
is disabled. Ensure that we properly drop the error on the floor or we
would re-trigger the checked failure.

Differential Revision: https://reviews.llvm.org/D147669
Reviewed By: sgraenitz

18 months ago[mlir][openacc][NFC] Use assembly format for acc.loop
Valentin Clement [Mon, 17 Apr 2023 14:52:50 +0000 (07:52 -0700)]
[mlir][openacc][NFC] Use assembly format for acc.loop

Use the assembly format with custom parser/printer
for specific clauses instead of a full custom parser/printer.

Reviewed By: PeteSteinfeld

Differential Revision: https://reviews.llvm.org/D148391

18 months ago[libcxxabi] [test] Use the correct printf formats for printing pointers
Martin Storsjö [Wed, 5 Apr 2023 11:52:30 +0000 (14:52 +0300)]
[libcxxabi] [test] Use the correct printf formats for printing pointers

Don't cast the pointers to long, as that's not large enough for
pointers on 64 bit Windows.

Differential Revision: https://reviews.llvm.org/D147640

18 months ago[libc][fix] Only use the object files when targeting NVPTX
Joseph Huber [Mon, 17 Apr 2023 14:36:13 +0000 (09:36 -0500)]
[libc][fix] Only use the object files when targeting NVPTX

Summary:
The `nvlink` linker doesn't support static libraries, so we just pass in
the object files. The condition was erroneously doing this for every
single GPU architecture and not just NVIDIA. The AMDGPU support handles
static libraries just fine.

18 months ago[libc] Add dependency on the loader for GPU tests
Joseph Huber [Mon, 17 Apr 2023 14:33:08 +0000 (09:33 -0500)]
[libc] Add dependency on the loader for GPU tests

Summary:
We need a dependency here so the loader is up-to-date whenever we run
the tests again.

18 months ago[LLVM][OHOS] Use emulated TLS for OHOS platform
Pavel Kosov [Mon, 17 Apr 2023 14:01:21 +0000 (17:01 +0300)]
[LLVM][OHOS] Use emulated TLS for OHOS platform

Both Linux and LiteOS for all OpenHarmony targets use emulated TLS

~~~

Huawei RRI, OS Lab

Reviewed By: DavidSpickett, jrtc27, MaskRay

Differential Revision: https://reviews.llvm.org/D145224

18 months ago[BOLT] Fix use-after-free in RewriteInstance::mapCodeSections
Job Noorman [Mon, 17 Apr 2023 12:55:11 +0000 (14:55 +0200)]
[BOLT] Fix use-after-free in RewriteInstance::mapCodeSections

When a cold function is too large, its section gets deregistered.
However, the section is still dereferenced later to get its RuntimeDyld
ID. This patch moves the deregistration to after the last dereference.

Reviewed By: Amir

Differential Revision: https://reviews.llvm.org/D148427

18 months agoSpeculatively fix a failing bot
Aaron Ballman [Mon, 17 Apr 2023 14:08:40 +0000 (10:08 -0400)]
Speculatively fix a failing bot

This attempts to resolve the issue found by:
https://lab.llvm.org/buildbot/#/builders/139/builds/39296

18 months ago[InstSimplify] Precommit a test
Peixin Qiao [Mon, 17 Apr 2023 13:55:51 +0000 (21:55 +0800)]
[InstSimplify] Precommit a test

This patch precommits a test for:

https://reviews.llvm.org/D148420

18 months agoConstexpr evaluator should treat [[gnu::weak]] member pointer comparisons as evaluati...
Takuya Shimizu [Mon, 17 Apr 2023 13:48:58 +0000 (09:48 -0400)]
Constexpr evaluator should treat [[gnu::weak]] member pointer comparisons as evaluation failure

This patch fixes the wrong signal from the constexpr evaluator that
[[gnu::weak]] member pointer comparison is valid, while it is emitting
notes on them.

I found a crashing case fixed by this change and added it as a test
case: https://godbolt.org/z/8391fGjGn

I noticed this while I was working on D146358.

Differential Revision: https://reviews.llvm.org/D148419

18 months ago[LoopDeletion] Don't delete loop exiting to EH pad (PR62160)
Nikita Popov [Mon, 17 Apr 2023 13:47:11 +0000 (15:47 +0200)]
[LoopDeletion] Don't delete loop exiting to EH pad (PR62160)

We can't branch directly to the EH pad, which is what the current
loop deletion code would try to do. We would need a different
approach here, which retains the invoke. This edge case does not
look worth bothering with.

Fixes https://github.com/llvm/llvm-project/issues/62160.

18 months ago[Support] Add MVT::getDoubleNumVectorElementsVT helper.
Simon Pilgrim [Mon, 17 Apr 2023 13:42:14 +0000 (14:42 +0100)]
[Support] Add MVT::getDoubleNumVectorElementsVT helper.

Matches the equivalent EVT::getDoubleNumVectorElementsVT helper.

This allows us to consistently MVT instead of EVT in the combinePTESTCC method.

18 months ago[clang] Add __is_trivially_equality_comparable
Nikolas Klauser [Wed, 29 Mar 2023 15:58:25 +0000 (17:58 +0200)]
[clang] Add __is_trivially_equality_comparable

This patch adds a new trait to allow standard libraries to forward `std::equal` calls to `memcmp` in more cases.

Reviewed By: aaron.ballman

Spies: Mordante, shafik, xbolva00, libcxx-commits, cfe-commits, ldionne

Differential Revision: https://reviews.llvm.org/D147175

18 months ago[flang][hlfir] lower hlfir.get_extent to FIR
Tom Eccles [Thu, 13 Apr 2023 10:48:12 +0000 (10:48 +0000)]
[flang][hlfir] lower hlfir.get_extent to FIR

By the ConvertToFIR pass, the hlfir.get_shape operation will have been
lowered into a fir.shape operation (during the HFLIR bufferization pass)
and so, lowering get_extent is as simple as fetching the extent from the
shape operation.

Depends on: D146833

Differential Revision: https://reviews.llvm.org/D148222

18 months ago[flang][hlfir] lower hlfir.shape_of
Tom Eccles [Fri, 17 Mar 2023 15:49:22 +0000 (15:49 +0000)]
[flang][hlfir] lower hlfir.shape_of

If possible the shape is gotten from the bufferization of the expr
argument.

The simple cases should already have been resolved during lowering. This
is mostly intended for cases where shape information is added in between
lowering and the end of bufferization (for example transformational
intrinsics with assumed shape arguments).

Depends on: D146832

Differential Revision: https://reviews.llvm.org/D146833

18 months ago[flang][hlfir] get extents from hlfir.shape_of
Tom Eccles [Thu, 2 Mar 2023 15:52:20 +0000 (15:52 +0000)]
[flang][hlfir] get extents from hlfir.shape_of

If the extents were known, this should have been canonicalised into a
fir.shape operation. Therefore, the extents at this point are not known at
compile time. Use hlfir.get_extents to delay resolving the real extent
until after the expression is bufferized.

Depends On: D146831

Differential Revision: https://reviews.llvm.org/D146832

18 months ago[flang][hlfir] get shape from HLFIR expr using hlfir.shape_of
Tom Eccles [Thu, 2 Mar 2023 15:51:31 +0000 (15:51 +0000)]
[flang][hlfir] get shape from HLFIR expr using hlfir.shape_of

Depends On: D146830

Differential Revision: https://reviews.llvm.org/D146831

18 months ago[flang][hlfir] add hlfir.get_extent
Tom Eccles [Wed, 12 Apr 2023 14:57:45 +0000 (14:57 +0000)]
[flang][hlfir] add hlfir.get_extent

This operation fetches an extent value from a fir.shape. The operation
could just as easily live in the fir namespace, but is only needed for
hlfir lowering so I put it here.

This operation is required to allow one to defer getting the extents of a shape
generated by hlfir.get_shape until after that shape has been resolved
(after bufferization of the hlfir.expr).

This operation will be lowered to FIR as an arith.constant created using
the definition of the fir.shape argument.

Depends on: D146830

Differential Revision: https://reviews.llvm.org/D148220

18 months ago[flang][hlfir] add hlfir.shape_of
Tom Eccles [Wed, 1 Mar 2023 10:20:15 +0000 (10:20 +0000)]
[flang][hlfir] add hlfir.shape_of

This is an operation which returns the fir.shape for a hlfir.expr.

A hlfir.expr can be defined by:
  - A transformational intrinsic (e.g. hlfir.matmul)
  - hlfir.as_expr
  - hlfir.elemental

hlfir.elemental is easy because there is a compulsory shape operand.
hlfir.as_expr is defined as operating on a variable (defined using a
hlfir.declare). hlfir.declare has an optional shape argument. The
transformational intrinsics do not have an associated shape.

If all extents are known at compile time, the extents for the shape can
be fetched from the hlfir.expr's type. For example, the result of a
hlfir.matmul with arguments who's extents are known at compile time will
have constant extents which can be queried from the type. In this case
the hlfir.shape_of will be canonicalised to a fir.shape operation using
those extents.

If not all extents are known at compile time, shapes have to be read
from boxes after bufferization. In the case of the transformational
intrinsics, the shape read from the result box can be queried from the
hlfir.declare operation for the buffer allocated to that hlfir.expr (via
the hlfir.as_expr).

Differential Revision: https://reviews.llvm.org/D146830

18 months ago[libc][NFC] Standardize missing syscalls error messages.
Mikhail R. Gadelha [Mon, 17 Apr 2023 13:16:56 +0000 (10:16 -0300)]
[libc][NFC] Standardize missing syscalls error messages.

This patch standardizes the error messages when a syscall is not
available to be in the format: "ABC and DEF syscalls are not available."

Reviewed By: sivachandra

Differential Revision: https://reviews.llvm.org/D148373

18 months ago[X86] combinePTESTCC - fold TESTZ(OR(LO(X),HI(X)),OR(LO(Y),HI(Y))) -> TESTZ(X,Y)...
Simon Pilgrim [Mon, 17 Apr 2023 13:09:56 +0000 (14:09 +0100)]
[X86] combinePTESTCC - fold TESTZ(OR(LO(X),HI(X)),OR(LO(Y),HI(Y))) -> TESTZ(X,Y) for TESTPS/TESTPD ops

Followup to the fix for #62171, adding support for TESTPS/TESTPD opcodes

18 months ago[RISCV] Fix RUN line in fixed-vectors-abs-vp.ll
Luke Lau [Mon, 17 Apr 2023 10:29:06 +0000 (11:29 +0100)]
[RISCV] Fix RUN line in fixed-vectors-abs-vp.ll

Reviewed By: fakepaper56

Differential Revision: https://reviews.llvm.org/D148510

18 months ago[LV] Regenreate check lines fr pr33706.ll
Florian Hahn [Mon, 17 Apr 2023 12:49:48 +0000 (13:49 +0100)]
[LV] Regenreate check lines fr pr33706.ll

This avoids conflicts when regenerating check lines.

18 months ago[VPlan] Remove uneeded NeedsVectorIV from VPWidenIntOrFpInduction.
Florian Hahn [Mon, 17 Apr 2023 12:38:00 +0000 (13:38 +0100)]
[VPlan] Remove uneeded NeedsVectorIV from VPWidenIntOrFpInduction.

After recent improvements, all instances of
VPWidenIntOrFpInductionRecipe should needs a vector IV and there's no
need for a separate field.

18 months ago[clang] Do not crash after suggesting typo correction to constexpr if condition
Mariya Podchishchaeva [Mon, 17 Apr 2023 11:53:43 +0000 (07:53 -0400)]
[clang] Do not crash after suggesting typo correction to constexpr if condition

In some cases non-null non-constant yet valid expression may reach point where
`ConditionResult` is created. For example, typo correction mechanism can return
such expression, so double check before evaluating it.

Fixes https://github.com/llvm/llvm-project/issues/61885

Reviewed By: tbaeder, aaron.ballman

Differential Revision: https://reviews.llvm.org/D148206

18 months agoRevert "[Passes] Remove legacy PM versions of InstructionNamer and MetaRenamer"
Bjorn Pettersson [Mon, 17 Apr 2023 12:23:27 +0000 (14:23 +0200)]
Revert "[Passes] Remove legacy PM versions of InstructionNamer and MetaRenamer"

This reverts commit 981ec1faeb508a364cc47c8246b72fc89dd8c1d8.

It broke polly build bots. Polly still uses -instnamer with legacy PM.

18 months agoReapply [SimplifyCFG][LICM] Preserve nonnull, range and align metadata when speculating
Nikita Popov [Wed, 22 Mar 2023 10:53:01 +0000 (11:53 +0100)]
Reapply [SimplifyCFG][LICM] Preserve nonnull, range and align metadata when speculating

This exposed a miscompile in GVN, which was fixed by D148129.

-----

After D141386, violation of nonnull, range and align metadata
results in poison rather than immediate undefined behavior,
which means that these are now safe to retain when speculating.
We only need to remove UB-implying metadata like noundef.

This is done by adding a dropUBImplyingAttrsAndMetadata() helper,
which lists the metadata which is known safe to retain on speculation.

Differential Revision: https://reviews.llvm.org/D146629

18 months ago[LV] Extend recurrence test coverage for sinking memory instructions.
Florian Hahn [Mon, 17 Apr 2023 12:08:14 +0000 (13:08 +0100)]
[LV] Extend recurrence test coverage for sinking memory instructions.

Extra coverage for D143604, D143605.

18 months ago[Clang][X86] Change X86 cast intrinsics to use __builtin_nondeterministic_value
ManuelJBrito [Mon, 17 Apr 2023 10:31:22 +0000 (11:31 +0100)]
[Clang][X86] Change X86 cast intrinsics to use __builtin_nondeterministic_value

The following intrinsics are currently implemented using a shufflevector with
an undefined mask, this is however incorrect according to intel's semantics for
undefined value which expect an unknown but consistent value.

With __builtin_nondeterministic_value we can now match intel's undefined value.

Differential Revision: https://reviews.llvm.org/D143287

18 months ago[Passes] Remove legacy PM versions of InstructionNamer and MetaRenamer
Bjorn Pettersson [Sun, 16 Apr 2023 21:17:58 +0000 (23:17 +0200)]
[Passes] Remove legacy PM versions of InstructionNamer and MetaRenamer

18 months ago[Vectorize] Clean up Transforms/Vectorize.h
Bjorn Pettersson [Sat, 15 Apr 2023 08:58:22 +0000 (10:58 +0200)]
[Vectorize] Clean up Transforms/Vectorize.h

Removed definitions of vectorizeBasicBlock and VectorizeConfig
(possibly a remnant from the BBVectorize pass that was removed
way back in 2017).

Also reduced amount of include dependencies to Transforms/Vectorize.h.

18 months agoRemove several no longer needed includes. NFCI
Bjorn Pettersson [Sat, 15 Apr 2023 08:16:09 +0000 (10:16 +0200)]
Remove several no longer needed includes. NFCI

Mostly removing includes of InitializePasses.h and Pass.h in
passes that no longer has support for the legacy PM.

18 months ago[mlir] Apply ClangTidy fix.
Adrian Kuegel [Mon, 17 Apr 2023 11:50:29 +0000 (13:50 +0200)]
[mlir] Apply ClangTidy fix.

The argument name 'useBarePtrCallConv' does not match the actual
parameter name 'useBarePointerCallConv'.

18 months ago[VPlan] Mark recurrence recipes as not having side-effects.
Florian Hahn [Mon, 17 Apr 2023 11:30:52 +0000 (12:30 +0100)]
[VPlan] Mark recurrence recipes as not having side-effects.

Add support for FirstOrderRecurrenceSplice and VPFirstOrderRecurrencePHI
recipes to mayHaveSideEffects. They both don't have side-effects.

18 months ago[mlir] Remove unused using llvm::Twine declaration (NFC).
Adrian Kuegel [Mon, 17 Apr 2023 11:15:45 +0000 (13:15 +0200)]
[mlir] Remove unused using llvm::Twine declaration (NFC).

18 months ago[Pipelines] Add LoopSink and DivRemPairs to LTO post-link pipeline
Nikita Popov [Fri, 14 Apr 2023 13:30:27 +0000 (15:30 +0200)]
[Pipelines] Add LoopSink and DivRemPairs to LTO post-link pipeline

As pointed out in D148010, these passes are missing from the LTO
post-link pipeline. They are present in the pre-link pipeline,
but LoopSink is completely useless there (it will always be fully
undone by LICM post-link) and DivRemPairs is mostly useless
(I believe most of what it does will be undone by InstCombine).

I've not added RelLookupTableConverterPass, because it's also
disabled in the LTO pre-link pipeline, with a comment that there
is an unresolved issue with full LTO.

Compile-time impact of the extra passes is minimal. Of course,
LoopSink will have a larger impact in PGO builds.

Differential Revision: https://reviews.llvm.org/D148343

18 months agoRevert "ValueTracking: Handle freeze in computeKnownFPClass"
Nikita Popov [Mon, 17 Apr 2023 10:52:58 +0000 (12:52 +0200)]
Revert "ValueTracking: Handle freeze in computeKnownFPClass"

This reverts commit 2c8d0048f03d054f13909a26f959ef95b2a0a4de.

This is incorrect: computeKnownFPClass() is only known up to
poison, and freeze poison may have any FP class.

18 months ago[LV] Add users to recurrence tests to make sure they are not removable.
Florian Hahn [Mon, 17 Apr 2023 10:56:56 +0000 (11:56 +0100)]
[LV] Add users to recurrence tests to make sure they are not removable.

This ensures VPlan-based DCE won't be able to remove the unused
recurrences.

It also adds a dedicated new test (@unused_recurrence) where an unused
recurrence can be removed.

18 months ago[GVN] Adjust metadata for coerced load CSE
Nikita Popov [Fri, 14 Apr 2023 09:48:54 +0000 (11:48 +0200)]
[GVN] Adjust metadata for coerced load CSE

When reusing a load in a way that requires coercion (i.e. casts or
bit extraction) we currently fail to adjust metadata. Unfortunately,
none of our existing tooling for this is really suitable, because
combineMetadataForCSE() expects both loads to have the same type.
In this case we may work on loads of different types and possibly
offset memory location.

As such, what this patch does is to simply drop all metadata, with
the following exceptions:

* Metadata for which violation is known to always cause UB.
* If the load is !noundef, keep all metadata, as this will turn
  poison-generating metadata into UB as well.

This fixes the miscompile that was exposed by D146629.

Differential Revision: https://reviews.llvm.org/D148129

18 months ago[LoopVectorize] Take vscale into account when deciding to create epilogues
David Sherwood [Mon, 3 Apr 2023 16:14:09 +0000 (16:14 +0000)]
[LoopVectorize] Take vscale into account when deciding to create epilogues

In LoopVectorizationCostModel::isEpilogueVectorizationProfitable we
check to see if the chosen main vector loop VF >= 16. If so, we
decide to create a vector epilogue loop. However, this doesn't
take VScaleForTuning into account because we could be targeting a
CPU where vscale > 1, and hence the runtime VF would be a multiple
of the known minimum value.

This patch multiplies scalable VFs by VScaleForTuning and several
tests have been updated that now produce vector epilogues.

Differential Revision: https://reviews.llvm.org/D147522

18 months ago[X86] combinePTESTCC - only fold TESTZ(OR(LO(X),HI(X)),OR(LO(Y),HI(Y))) -> TESTZ...
Simon Pilgrim [Mon, 17 Apr 2023 10:40:39 +0000 (11:40 +0100)]
[X86] combinePTESTCC - only fold TESTZ(OR(LO(X),HI(X)),OR(LO(Y),HI(Y))) -> TESTZ(X,Y) for TESTP ops

Stop value type mismatch assertions, we can add TESTPS/TESTPD support to this in the future.

Fixes #62171

18 months ago[libc++] Add annotation for arm64e in the strong_order test for long double
Louis Dionne [Mon, 17 Apr 2023 10:40:51 +0000 (11:40 +0100)]
[libc++] Add annotation for arm64e in the strong_order test for long double

As a fly-by, improve readability by giving a bit more space to some
comments.