platform/upstream/llvm.git
17 months ago[-Wunsafe-buffer-usage] Add unsafe buffer checking opt-out pragmas
Ziqing Luo [Wed, 8 Feb 2023 00:45:44 +0000 (16:45 -0800)]
[-Wunsafe-buffer-usage] Add unsafe buffer checking opt-out pragmas

Add a pair of clang pragmas:
- `#pragma clang unsafe_buffer_usage begin` and
- `#pragma clang unsafe_buffer_usage end`,
which specify the start and end of an (unsafe buffer checking) opt-out
region, respectively.

Behaviors of opt-out regions conform to the following rules:

- No nested nor overlapped opt-out regions are allowed. One cannot
  start an opt-out region with `... unsafe_buffer_usage begin` but never
  close it with `... unsafe_buffer_usage end`. Mis-use of the pragmas
  will be warned.
- Warnings raised from unsafe buffer operations inside such an opt-out
  region will always be suppressed. This behavior CANNOT be changed by
  `clang diagnostic` pragmas or command-line flags.
- Warnings raised from unsafe operations outside of such opt-out
  regions may be reported on declarations inside opt-out
  regions. These warnings are NOT suppressed.
- An un-suppressed unsafe operation warning may be attached with
  notes. These notes are NOT suppressed as well regardless of whether
  they are in opt-out regions.

The implementation maintains a separate sequence of location pairs
representing opt-out regions in `Preprocessor`.  The `UnsafeBufferUsage`
analyzer reads the region sequence to check if an unsafe operation is
in an opt-out region. If it is, discard the warning raised from the
operation immediately.

Reviewed by: NoQ

Differential revision: https://reviews.llvm.org/D140179

17 months ago[libc++][NFC] Use _LIBCPP_HIDE_FROM_ABI everywhere in pair
Louis Dionne [Wed, 8 Feb 2023 00:52:28 +0000 (16:52 -0800)]
[libc++][NFC] Use _LIBCPP_HIDE_FROM_ABI everywhere in pair

17 months ago[gn build] Port 692da6245d71
LLVM GN Syncbot [Wed, 8 Feb 2023 00:16:16 +0000 (00:16 +0000)]
[gn build] Port 692da6245d71

17 months ago[-Wunsafe-buffer-usage] Filter out conflicting fix-its
Ziqing Luo [Wed, 8 Feb 2023 00:04:09 +0000 (16:04 -0800)]
[-Wunsafe-buffer-usage] Filter out conflicting fix-its

Two fix-its conflict if they have overlapping source ranges. We shall
not emit conflicting fix-its.  This patch checks conflicts in fix-its
generated for one variable (including variable declaration fix-its and
variable usage fix-its). If there is any, we do NOT emit any fix-it
for that variable.

Reviewed by: NoQ

Differential revision: https://reviews.llvm.org/D141338

17 months agoOnly run the weird new try-to-read-too-much test on Darwin
Jason Molenda [Wed, 8 Feb 2023 00:04:14 +0000 (16:04 -0800)]
Only run the weird new try-to-read-too-much test on Darwin

I'm still getting linux CI bot failures for this test.  It's not
critical, and it depends on a failure mode that is true on Darwin
but I was always gambling that it might fail in the same way on
other systems.

17 months agoFix TestProcessAPI.py to only allocate sys.maxsize buffer
Jason Molenda [Tue, 7 Feb 2023 23:25:51 +0000 (15:25 -0800)]
Fix TestProcessAPI.py to only allocate sys.maxsize buffer

I hardcoded nearly a UINT64_MAX number in this test case,
and python is not able to convert it to a long on some
platforms.  Use sys.maxsize instead; this also would have
failed if the testsuite was run on a 32-bit system.

17 months ago[emacs] Add llvm-mir-mode
Justin Bogner [Tue, 7 Feb 2023 23:54:49 +0000 (15:54 -0800)]
[emacs] Add llvm-mir-mode

Adds an emacs mode for .mir files. For the most part this just
consists of keyword rules for various MIR constructs and then
appending the llvm-mode keywords to that. This doesn't currently
attempt to do anything to be aware of the YAML structure or
differentiate between machine IR and embedded LLVM IR.

17 months agoAdd `set(CMAKE_CXX_STANDARD 17)` to MLIR CMakeLists.txt
Mehdi Amini [Tue, 7 Feb 2023 23:54:27 +0000 (15:54 -0800)]
Add `set(CMAKE_CXX_STANDARD 17)` to MLIR CMakeLists.txt

This is only useful when building the project in a "standalone" way: that is by
invoking cmake pointing at mlir/ to build against an already built LLVM.

Fixes #60574

17 months ago[mlir][memref] fix typo in documentation (NFC)
Peiming Liu [Tue, 7 Feb 2023 23:49:29 +0000 (23:49 +0000)]
[mlir][memref] fix typo in documentation (NFC)

Reviewed By: aartbik

Differential Revision: https://reviews.llvm.org/D143532

17 months ago[bazel] Fix libc
Fangrui Song [Tue, 7 Feb 2023 23:51:06 +0000 (15:51 -0800)]
[bazel] Fix libc

17 months ago[-Wunsafe-buffer-usage] Generate fix-it for local variable declarations
Ziqing Luo [Tue, 7 Feb 2023 23:38:21 +0000 (15:38 -0800)]
[-Wunsafe-buffer-usage] Generate fix-it for local variable declarations

Use clang fix-its to transform declarations of local variables, which
are used for buffer access , to be of std::span type.

We placed a few limitations to keep the solution simple:
- it only transforms local variable declarations (no parameter declaration);
- it only considers single level pointers, i.e., pointers of type T * regardless of whether T is again a pointer;
- it only transforms to std::span types (no std::array, or std::span::iterator, or ...);
    - it can only transform a VarDecl that belongs to a DeclStmt whose has a single child.

One of the purposes of keeping this patch simple enough is to first
evaluate if fix-it is an appropriate approach to do the
transformation.

This commit was reverted by 622be09c815266632e204eaf1c7a35f050220459
for a compilation warning and now it is fixed.

Reviewed by: NoQ, jkorous

Differential revision: https://reviews.llvm.org/D139737

17 months ago[Sanitizers] Fix read buffer overrun in scanning loader commands
Mariusz Borsa [Tue, 7 Feb 2023 22:19:54 +0000 (14:19 -0800)]
[Sanitizers] Fix read buffer overrun in scanning loader commands

The fix only affects Darwin, but to write the test I had to modify
the MemoryMappingLayout class which is used by all OSes,
to allow for mocking of image header (this change should be NFC). Hence no [Darwin] in the subject
so I can get more eyes on it.

While looking for a memory gap to put the shadow area into, the sanitizer code
scans through the loaded images, and for each image it scans through its
loader command to determine the occupied memory ranges.

While doing so, if the 'segment load' (kLCSegment) loader comand is encountered, the command scanning function
returns success (true), but does not decrement the command list iterator counter.
The result is that the function is called again and again, with the iterator counter
now being too high. The command scanner keeps updating the loader command pointer,
by using the command size field.

If the loop counter is too high, the command pointer
lands into unintended area ( beyond
+sizeof(mac_header64)+header->sizeofcmds ),
and result depends on the random content found there.

The random content interpreted as loader command might contain a large integer value in the
cmdsize field - this value is added to the current loader command pointer,
which might now point to an inaccessible memory address. It can occasionally result
in a crash if it happens to run beyond the mapped memory segment.

Note that when the area after the loader command list
contains zeros or small integers only, the loop will end normally and the problem
will go unnoticed. So it happened until now since having a some big value
after the header area, falling into command size field is a pretty rare situation.

The fix makes sure that the iterator counter gets updated when the segment load (kLCSegment)
loader command is found too, and in the same code location so the updates will always go together.

Undo the changes in the sanitizer_procmaps_mac.cpp to see the test failing.

rdar://101161047
rdar://102819707

Differential Revision: https://reviews.llvm.org/D142164

17 months ago[Clang][CMake] Set up distribution target for Clang-BOLT
Amir Ayupov [Tue, 7 Feb 2023 23:05:34 +0000 (15:05 -0800)]
[Clang][CMake] Set up distribution target for Clang-BOLT

Provide a way to install usable BOLT-optimized Clang
(clang + resource headers) using
`ninja clang-bolt install-distribution` with BOLT.cmake cache file
or `ninja stage2-clang-bolt stage2-install-distribution`
with BOLT-PGO.cmake cache file.

Reviewed By: phosek

Differential Revision: https://reviews.llvm.org/D140565

17 months ago[mlir][gpu] Add support for integer types in gpu.subgroup_mma ops
Quinn Dawkins [Mon, 6 Feb 2023 04:53:38 +0000 (23:53 -0500)]
[mlir][gpu] Add support for integer types in gpu.subgroup_mma ops

The signedness is carried by `!gpu.mma_matrix` types to most closely
match the Cooperative Matrix specification which determines signedness
with the type (and sometimes the operation).

See: https://htmlpreview.github.io/?https://github.com/KhronosGroup/SPIRV-Registry/blob/master/extensions/NV/SPV_NV_cooperative_matrix.html

To handle the lowering from vector to gpu, ops such as arith.extsi are
pattern matched next to `vector.transfer_read` and `vector.contract` to
determine the signedness of the matrix type.

Enables s8 and u8 WMMA types in NVVM for the GPUToNVVM conversion.

Reviewed By: ThomasRaoux

Differential Revision: https://reviews.llvm.org/D143223

17 months agoRevert "[-Wunsafe-buffer-usage] Generate fix-it for local variable declarations"
Ziqing Luo [Tue, 7 Feb 2023 22:47:43 +0000 (14:47 -0800)]
Revert "[-Wunsafe-buffer-usage] Generate fix-it for local variable declarations"

This reverts commit a29e67614c3b7018287e5f68c57bba7618aa880e.

17 months ago[mlir] Relax version requirement for PyYAML in mlir
Stella Stamenova [Tue, 7 Feb 2023 22:24:55 +0000 (14:24 -0800)]
[mlir] Relax version requirement for PyYAML in mlir

Some Ubuntu 20.04 images come with PyYAML 5.3.1 pre-installed through distutils. This makes pip very angry. See https://github.com/yaml/pyyaml/issues/349.

Since older versions of PyYAML should work for mlir, relax the version requirement to ease developer setup.

Reviewed By: stellaraccident

Differential Revision: https://reviews.llvm.org/D143523

17 months ago[NVPTX] Lower extraction of upper half of i32/i64 as partial move.
Artem Belevich [Tue, 7 Feb 2023 00:32:22 +0000 (16:32 -0800)]
[NVPTX] Lower extraction of upper half of i32/i64 as partial move.

This produces better SASS than right-shift + truncate and is fairly common for
CUDA code that operates on __half2 values represented as opaque integer.

Differential Revision: https://reviews.llvm.org/D143448

17 months agoCheck if null buffer handed to SBProcess::ReadMemory
Jason Molenda [Tue, 7 Feb 2023 21:54:41 +0000 (13:54 -0800)]
Check if null buffer handed to SBProcess::ReadMemory

Add a check for a null destination buffer in SBProcess::ReadMemory,
and return an error if that happens.  If a Python SB API script
tries to allocate a huge amount of memory, the malloc done by the
intermediate layers will fail and will hand a null pointer to
ReadMemory.  lldb will eventually crash trying to write in to that
buffer.

Also add a test that tries to allocate an impossibly large amount
of memory, and hopefully should result in a failed malloc and hitting
this error codepath.

Differential Revision: https://reviews.llvm.org/D143012
rdar://104846609

17 months ago[VPlan] Mark load VPWidenMemoryInstruction as not having side-effects.
Florian Hahn [Tue, 7 Feb 2023 22:02:50 +0000 (22:02 +0000)]
[VPlan] Mark load VPWidenMemoryInstruction as not having side-effects.

Also add an assert using the underlying instruction to catch any
potential violations.

17 months ago[bazel] Make libc:__support_common depend on :libc_root
Fangrui Song [Tue, 7 Feb 2023 21:59:22 +0000 (13:59 -0800)]
[bazel] Make libc:__support_common depend on :libc_root

17 months ago[FuncSpec] Prevent assertion failure when no store value is found
Jonathon Penix [Tue, 7 Feb 2023 03:05:09 +0000 (19:05 -0800)]
[FuncSpec] Prevent assertion failure when no store value is found

If the only user of the Alloca argument provided to getPromotableAlloca()
is the same as the Call argument, StoreValue is never set and results
in an assertion failure that isa<> was used on a nullptr when passed into
getCandidateConstant().

This was originally seen when trying to build SPEC 2006 416.gamess using
flang with lto enabled.

Differential Revision: https://reviews.llvm.org/D143457

17 months ago[LV] Remove unused load from RISCV test (NFC).
Florian Hahn [Tue, 7 Feb 2023 21:55:43 +0000 (21:55 +0000)]
[LV] Remove unused load from RISCV test (NFC).

The test contained a unused load that appears unrelated to the test
(store of vector of i1). Remove it to avoid test changes in follow-up
change which will lead to dead loads being removed.

17 months agoRemove textual headers from module map
Adrian Prantl [Tue, 7 Feb 2023 21:24:47 +0000 (13:24 -0800)]
Remove textual headers from module map

since headers were moved into the lib/ directory.

17 months ago[-Wunsafe-buffer-usage] Generate fix-it for local variable declarations
Ziqing Luo [Tue, 7 Feb 2023 21:14:09 +0000 (13:14 -0800)]
[-Wunsafe-buffer-usage] Generate fix-it for local variable declarations

Use clang fix-its to transform declarations of local variables, which are used for buffer access , to be of std::span type.

We placed a few limitations to keep the solution simple:
- it only transforms local variable declarations (no parameter declaration);
- it only considers single level pointers, i.e., pointers of type T * regardless of whether T is again a pointer;
- it only transforms to std::span types (no std::array, or std::span::iterator, or ...);
- it can only transform a VarDecl that belongs to a DeclStmt whose has a single child.

One of the purposes of keeping this patch simple enough is to first
evaluate if fix-it is an appropriate approach to do the
transformation.

Reviewed by: NoQ, jkorous

Differential revision: https://reviews.llvm.org/D139737

17 months ago[RPOFunctionAttrs] Use LazyCallGraph instead of CallGraph
Samuel [Tue, 7 Feb 2023 18:05:36 +0000 (10:05 -0800)]
[RPOFunctionAttrs] Use LazyCallGraph instead of CallGraph

There are a few more uses of CallGraph that should be replaced with LazyCallGraph

Also delete legacy version of RPOFunctionAttrs since it is deprecated and LazyCallGraph is not available under the legacy pass manager.

Reviewed By: aeubanks

Differential Revision: https://reviews.llvm.org/D143358

17 months ago[BOLT] Search section based on relocation symbol
yavtuk [Tue, 7 Feb 2023 20:23:44 +0000 (23:23 +0300)]
[BOLT] Search section based on relocation symbol

We need to search referenced section based on relocations symbol section
to properly match end section symbols. For example on some binaries we
can observe that init_array_end/fini_array_end might be "placed" in to
the gap and since no section could be found for address the relocation
would be skipped resulting in wrong ADRP imm after emitting new text
resulting in binary sigsegv.

Credits for the test to Vladislav Khmelevskii aka yota9.

17 months ago[bazel] Fix llvm/unittests after TargetParser change
Fangrui Song [Tue, 7 Feb 2023 21:06:23 +0000 (13:06 -0800)]
[bazel] Fix llvm/unittests after TargetParser change

17 months ago[bazel] Fix dependencies after TargetParser change
Fangrui Song [Tue, 7 Feb 2023 21:01:38 +0000 (13:01 -0800)]
[bazel] Fix dependencies after TargetParser change

17 months ago[Clang] Refactor "Designators" into a unified implementation [NFC]
Bill Wendling [Tue, 7 Feb 2023 18:01:44 +0000 (10:01 -0800)]
[Clang] Refactor "Designators" into a unified implementation [NFC]

The interfaces for designators (i.e. C99 designated initializers) was
done in two slightly different ways. This was rather wasteful as the
differences could be combined into one.

Reviewed By: rsmith

Differential Revision: https://reviews.llvm.org/D140584

17 months ago[libc] Fix sanitizer
Guillaume Chatelet [Tue, 7 Feb 2023 20:56:12 +0000 (20:56 +0000)]
[libc] Fix sanitizer

17 months agoReland [LegacyPM] Remove some legacy passes
Arthur Eubanks [Tue, 20 Dec 2022 21:36:20 +0000 (13:36 -0800)]
Reland [LegacyPM] Remove some legacy passes

These are part of the optimization pipeline, of which the legacy pass manager version is deprecated.

Namely
* Internalize
* StripSymbols
* StripNonDebugSymbols
* StripDeadDebugInfo
* StripDeadPrototypes
* VectorCombine
* WarnMissedTransformations

Fixed previously failing ocaml tests (one of them seems to already be failing?)

17 months ago[libc][NFC] Rename macros
Guillaume Chatelet [Tue, 7 Feb 2023 20:41:30 +0000 (20:41 +0000)]
[libc][NFC] Rename macros

17 months ago[libc][NFC] Move code to sanitizer.h + more consistent naming
Guillaume Chatelet [Tue, 7 Feb 2023 15:23:40 +0000 (15:23 +0000)]
[libc][NFC] Move code to sanitizer.h + more consistent naming

17 months ago[AArch64][GlobalISel] Legalize wide s8/s16 vectors G_ADD/G_MUL/G_OR/...
Vladislav Dzhidzhoev [Tue, 7 Feb 2023 16:48:58 +0000 (17:48 +0100)]
[AArch64][GlobalISel] Legalize wide s8/s16 vectors G_ADD/G_MUL/G_OR/...

Clamp the max number of elements of s8/s16 vectors when legalizing G_ADD,
G_SUB, G_MUL, G_AND, G_OR, G_XOR, in order to support some wide vectors.

Fixes https://github.com/llvm/llvm-project/issues/58156.

Differential Revision: https://reviews.llvm.org/D143517

17 months agoAdd codegen for llvm log2/log10 elementwise builtins
Joshua Batista [Tue, 7 Feb 2023 19:26:08 +0000 (11:26 -0800)]
Add codegen for llvm log2/log10 elementwise builtins

Add codegen for llvm log2 / log10 elementwise builtin
The log2/log10 elementwise builtin is necessary for HLSL codegen.
Tests were added to make sure that the expected errors are encountered when these functions are given inputs of incompatible types.
The new builtins are restricted to floating point types only.

Reviewed By: fhahn

Differential Revision: https://reviews.llvm.org/D143207

17 months ago[BOLT][TEST] Fix is-strip test for NFC testing
Amir Ayupov [Tue, 7 Feb 2023 20:16:45 +0000 (12:16 -0800)]
[BOLT][TEST] Fix is-strip test for NFC testing

17 months ago[mlir][LLVMIR] Add lowering support for vp.muladd intrinsic
Diego Caballero [Tue, 7 Feb 2023 20:08:46 +0000 (20:08 +0000)]
[mlir][LLVMIR] Add lowering support for vp.muladd intrinsic

Reviewed By: gysit

Differential Revision: https://reviews.llvm.org/D143449

17 months ago[OpenMP 5.2] Deprecate 'destroy' clause without argument for 'depobj' construct
Fazlay Rabbi [Tue, 7 Feb 2023 19:52:48 +0000 (11:52 -0800)]
[OpenMP 5.2] Deprecate 'destroy' clause without argument for 'depobj' construct

Deprecate and diagnose uses of the 'destroy' clauses without an argument
in 'omp depobj' directives.

Differential Revision: https://reviews.llvm.org/D143021

17 months ago[mlir][Vector] Add lowering support for 1-D masked multi-reductions
Diego Caballero [Tue, 7 Feb 2023 19:25:19 +0000 (19:25 +0000)]
[mlir][Vector] Add lowering support for 1-D masked multi-reductions

1-D multi-reductions follow a different lowering path (they are
converted to 2-D multi-reductions) so masked variants need to be
supported explicitly.

Reviewed By: nicolasvasilache

Differential Revision: https://reviews.llvm.org/D143453

17 months ago[flang] Add AMDGPU target in flang
Jan Sjodin [Wed, 1 Feb 2023 20:09:59 +0000 (15:09 -0500)]
[flang] Add AMDGPU target in flang

This is the first patch of several that will enable generating code for AMD
GPUs. It adds the AMDGPU target so it can be used with the --target and -mcpu
options.

Reviewed By: kiranchandramohan

Differential Revision: https://reviews.llvm.org/D143102

17 months ago[libc][NFC] Move UnitTest and IntegrationTest to the 'test' directory.
Siva Chandra Reddy [Mon, 6 Feb 2023 07:07:23 +0000 (07:07 +0000)]
[libc][NFC] Move UnitTest and IntegrationTest to the 'test' directory.

This part of the effort to make all test related pieces into the `test`
directory. This helps is excluding test related pieces in a straight
forward manner if LLVM_INCLUDE_TESTS is OFF. Future patches will also move
the MPFR wrapper and testutils into the 'test' directory.

17 months ago[Serialization] Add support for (de)serializing #pragma pack
Dustin Howett [Tue, 7 Feb 2023 19:02:59 +0000 (11:02 -0800)]
[Serialization] Add support for (de)serializing #pragma pack

Serialization of tokens is required when PCH is used with late parsed
templates, including annotation tokens used for pragmas.

This patch implements the serialization for annot_pragma_pack.

Fixes https://github.com/llvm/llvm-project/issues/60543

Differential Revision: https://reviews.llvm.org/D143410

17 months ago[CodeGen] Kill off a few more direct uses of EVT::getEVTString [nfc]
Philip Reames [Tue, 7 Feb 2023 19:22:13 +0000 (11:22 -0800)]
[CodeGen] Kill off a few more direct uses of EVT::getEVTString [nfc]

17 months ago[lit] don't skip test times when executions are instantaneous
Ashay Rane [Tue, 7 Feb 2023 19:27:23 +0000 (11:27 -0800)]
[lit] don't skip test times when executions are instantaneous

If a test executes quickly or if the timing resolution is too coarse,
the running time associated with a test could be zero. This happens
often for the //llvm/utils/lit/tests/reorder.py test on Windows, which
causes the test to fail non-deterministically.

This patch modifies the existing check, so that instead of skipping
zeros, lit now skips if the timing is None (presumably due to an error
in measurement). Recurring executions of the reorder.py test on Windows
seem to indicate that the test no longer exhibits occasional failure.

Reviewed By: stella.stamenova

Differential Revision: https://reviews.llvm.org/D143504

17 months ago[AArch64] Recurse further into performUADDVCombine
David Green [Tue, 7 Feb 2023 19:14:06 +0000 (19:14 +0000)]
[AArch64] Recurse further into performUADDVCombine

This adjusts the performUADDVCombine to recurse further into adds from
reductions, which can help produce more UADDLP when the order of lanes is
unimportant. It also prevents some regressions from subsequent patches.

17 months ago[CodeGen] Add standard print/debug utilities to MVT
Philip Reames [Tue, 7 Feb 2023 18:41:18 +0000 (10:41 -0800)]
[CodeGen] Add standard print/debug utilities to MVT

Doing so makes it easier to do printf style debugging in idiomatic manner. I followed the code structure of Value with only the definition of dump being #ifdef out in non-debug builds. Not sure if this is the "right" option; we don't seem to have any single consistent scheme on how dump is handled.

Note: This is a follow up to D143454 which did the same for EVT.

Differential Revision: https://reviews.llvm.org/D143511

17 months ago[llvm][IfConversion] update successor list when merging INLINEASM_BR
Nick Desaulniers [Tue, 7 Feb 2023 18:22:24 +0000 (10:22 -0800)]
[llvm][IfConversion] update successor list when merging INLINEASM_BR

If this successor list is not correct, then branch-folding may
incorrectly think that the indirect target is dead and remove it. This
results in a dangling reference to the removed block as an operand to
the INLINEASM_BR, which later will get AsmPrinted into code that doesn't
assemble.

This was made more obvious by, but is not a regression of
https://reviews.llvm.org/D130316.

Fixes: https://github.com/llvm/llvm-project/issues/60346

Reviewed By: efriedma, void

Differential Revision: https://reviews.llvm.org/D142924

17 months agoprecommit test for pr60346
Nick Desaulniers [Tue, 7 Feb 2023 18:19:12 +0000 (10:19 -0800)]
precommit test for pr60346

Link: https://github.com/llvm/llvm-project/issues/60346
Reviewed By: efriedma, void

Differential Revision: https://reviews.llvm.org/D142923

17 months ago[NFC][Clang] Regenerate test output for SVE ACLE tests.
Paul Walker [Tue, 7 Feb 2023 18:17:51 +0000 (18:17 +0000)]
[NFC][Clang] Regenerate test output for SVE ACLE tests.

Just a rerun of update_cc_test_checks.py to capture some changes to
variable names after their reliance on instcombine was removed.

17 months ago[X86] combineX86ShufflesRecursively - peek through insert_subvector(undef, x, 0)...
Simon Pilgrim [Tue, 7 Feb 2023 18:24:50 +0000 (18:24 +0000)]
[X86] combineX86ShufflesRecursively - peek through insert_subvector(undef, x, 0) vector widening nodes

If the node is inert then just peek through it and canonicalize the shuffle mask to ensure any referenced upper elements are set to undef.

17 months agoRevert "[LegacyPM] Remove some legacy passes"
Arthur Eubanks [Tue, 7 Feb 2023 18:17:24 +0000 (10:17 -0800)]
Revert "[LegacyPM] Remove some legacy passes"

This reverts commit a4b4f62beb0bf40123181e5f5bdf32ef54f87166.

Ocaml bindings tests failing.

17 months ago[mlir][Linalg] NFC - Extract inferGemmDims - Amend forgotten changes
Nicolas Vasilache [Tue, 7 Feb 2023 18:16:46 +0000 (10:16 -0800)]
[mlir][Linalg] NFC - Extract inferGemmDims - Amend forgotten changes

17 months ago[mlir][Linalg] NFC - Extract inferGemmDims
Nicolas Vasilache [Tue, 7 Feb 2023 17:58:02 +0000 (09:58 -0800)]
[mlir][Linalg] NFC - Extract inferGemmDims

17 months ago[LegacyPM] Remove some legacy passes
Arthur Eubanks [Tue, 20 Dec 2022 21:36:20 +0000 (13:36 -0800)]
[LegacyPM] Remove some legacy passes

These are part of the optimization pipeline, of which the legacy pass manager version is deprecated.

Namely
* Internalize
* StripSymbols
* StripNonDebugSymbols
* StripDeadDebugInfo
* StripDeadPrototypes
* VectorCombine
* WarnMissedTransformations

17 months ago[CostModel][X86] Update LV UTC_ARGS regexp filter to match ptr instead of i8* etc.
Simon Pilgrim [Tue, 7 Feb 2023 17:46:34 +0000 (17:46 +0000)]
[CostModel][X86] Update LV UTC_ARGS regexp filter to match ptr instead of i8* etc.

17 months ago[libc] Add startup code implementation for GPU targets
Joseph Huber [Thu, 2 Feb 2023 21:05:54 +0000 (15:05 -0600)]
[libc] Add startup code implementation for GPU targets

This patch introduces startup code for executing `main` on a device
compiled for the GPU. We will primarily use this to run standalone
integration tests on the GPU. The actual execution of this routine will
need to be provided by a `loader` utility to bootstrap execution on the
GPU.

Reviewed By: sivachandra

Differential Revision: https://reviews.llvm.org/D143212

17 months ago[clang] NFC: Remove GCC 5.1 workaround in FileEntry
Jan Svoboda [Mon, 6 Feb 2023 19:39:40 +0000 (11:39 -0800)]
[clang] NFC: Remove GCC 5.1 workaround in FileEntry

We no longer support GCC 5.1 (D122976), so let's remove the workaround in FileEntry.

Reviewed By: benlangmuir

Differential Revision: https://reviews.llvm.org/D143427

17 months ago[clang][deps] Fix race condition
Jan Svoboda [Mon, 6 Feb 2023 19:36:51 +0000 (11:36 -0800)]
[clang][deps] Fix race condition

D140176 introduced new `FullDeps` API that's not thread-safe, breaking the class invariant. This was causing race condition when `clang-scan-deps` was run with multiple threads.

Reviewed By: steven_wu, akyrtzi

Differential Revision: https://reviews.llvm.org/D143428

17 months ago[Hexagon] Add release note for LLVM 16
Krzysztof Parzyszek [Tue, 7 Feb 2023 17:24:06 +0000 (09:24 -0800)]
[Hexagon] Add release note for LLVM 16

17 months ago[X86] combineX86ShuffleChainWithExtract - refactor to remove need to widen all vectors
Simon Pilgrim [Tue, 7 Feb 2023 13:03:16 +0000 (13:03 +0000)]
[X86] combineX86ShuffleChainWithExtract - refactor to remove need to widen all vectors

combineX86ShuffleChain no longer needs the candidate shuffle inputs to all have been widened to root size, but combineX86ShuffleChainWithExtract still assumes that they have.

This refactor peeks through all inputs' extract_subvector nodes to find the widest legal vector type and widens the shuffle mask accordingly - it no longer bails if any of the inputs can't be widened to the new width.

combineX86ShuffleChainWithExtract now follows a similar process to combineX86ShufflesRecursively, so I think we should eventually be able to get rid of it and just let combineX86ShufflesRecursively handle widening the shuffle width as well.

Another step toward removing subvector widening for Issue #45319

17 months ago[ARM][AArch64][RISCV] Add tests for various double reductions. NFC
David Green [Tue, 7 Feb 2023 16:39:20 +0000 (16:39 +0000)]
[ARM][AArch64][RISCV] Add tests for various double reductions. NFC

17 months ago[CodeGen] Add standard print/debug utilities to EVT
Philip Reames [Tue, 7 Feb 2023 16:02:43 +0000 (08:02 -0800)]
[CodeGen] Add standard print/debug utilities to EVT

Doing so makes it easier to do printf style debugging in idiomatic manner. I followed the code structure of Value with only the definition of dump being #ifdef out in non-debug builds. Not sure if this is the "right" option; we don't seem to have any single consistent scheme on how dump is handled.

Differential Revision: https://reviews.llvm.org/D143454

17 months ago[libc] Cleanup the CMake infrastructure to add startup objects.
Siva Chandra Reddy [Tue, 7 Feb 2023 06:46:09 +0000 (06:46 +0000)]
[libc] Cleanup the CMake infrastructure to add startup objects.

Instead of using a custom target to copy the startup object file to a
file with the desired name, a normal object library with a special
property is used.

Follow up patches will do more cleanup wrt how the startup objects are
used in integration tests.

Reviewed By: jhuber6

Differential Revision: https://reviews.llvm.org/D143464

17 months ago[llvm] Add tests for upcoming fixes to pointer/index type confusion.
Krzysztof Drewniak [Mon, 6 Feb 2023 21:45:44 +0000 (21:45 +0000)]
[llvm] Add tests for upcoming fixes to pointer/index type confusion.

Various parts of the codebase are using getIntPtrType() and its
relatives when getting the type of the offset argument to GEP. Most
such code has been updated to use the pointer index type field from
the data layout, but there is code that still assumes these two types
are the same in certain optimizaiton passes.

This commit adds regression tests to capture the old behavior.

Reviewed By: #amdgpu, arsenm

Differential Revision: https://reviews.llvm.org/D143435

17 months ago[mlir][doc] Remove outdated section about a deprecated `OpBuilder` use
Markus Böck [Tue, 7 Feb 2023 16:04:28 +0000 (17:04 +0100)]
[mlir][doc] Remove outdated section about a deprecated `OpBuilder` use

This syntax has been removed two years ago in https://github.com/llvm/llvm-project/commit/2f5569f6f67a30f7774f7c2d2f3d726752a862ae yet the doc still had a section marking them as deprecated

17 months ago[Flang] [Semantics] More descript warning message for same named external statements...
Paul Scoropan [Mon, 6 Feb 2023 19:09:10 +0000 (19:09 +0000)]
[Flang] [Semantics] More descript warning message for same named external statements and interfaces

For issue: https://github.com/llvm/llvm-project/issues/56605. Previous error messages are not descriptive of problem. Issuing warnings instead that do not halt compilation but offer better description of problem.

Reviewed By: klausler

Differential Revision: https://reviews.llvm.org/D143426

17 months ago[OpenMP][libomp] Remove false positive for memory sanitizer
Jonathan Peyton [Mon, 6 Feb 2023 19:32:35 +0000 (13:32 -0600)]
[OpenMP][libomp] Remove false positive for memory sanitizer

The memory sanitizer intercepts the memcpy() call but not the direct
assignment of last byte to 0. This leads the sanitizer to believe the
last byte of a string based on the kmp_str_buf_t type is uninitialized.
Hence, the eventual strlen() inside __kmp_env_dump() leads to an
use-of-uninitialized-value warning.

Using strncat() instead gives the sanitizer the information it needs.

Differential Revision: https://reviews.llvm.org/D143401

Fixes #60501

17 months ago[WebAssembly][test][NFC] Add coverage of non-void funcref calls
Alex Bradbury [Tue, 7 Feb 2023 15:56:10 +0000 (15:56 +0000)]
[WebAssembly][test][NFC] Add coverage of non-void funcref calls

This is trickier to handle in some other representations of funcrefs
that are being explored, so it makes sense to ensure we have some
coverage of this requirement.

17 months ago[HWASAN] Remove GetRequestedSize check from Metadata::IsAllocated
Kirill Stoimenov [Tue, 7 Feb 2023 00:02:17 +0000 (00:02 +0000)]
[HWASAN] Remove GetRequestedSize check from Metadata::IsAllocated

After D143438 there is no point in this check as the size will never 0. Also this keeps it consistent with ASAN where there is not size check in LsanMetadata::allocated.

Reviewed By: MaskRay, vitalybuka

Differential Revision: https://reviews.llvm.org/D143442

17 months ago[mlir][ods] add mechanism for deprecating an `OpBuilder` in ODS
Markus Böck [Thu, 2 Feb 2023 15:35:02 +0000 (16:35 +0100)]
[mlir][ods] add mechanism for deprecating an `OpBuilder` in ODS

This allows discouraging the use of specific build methods of an op by having TableGen generate C++ code, instructing the C++ compiler to warn on use of the `build` method.
The implementation uses the C++14 `[[deprecated(...)]]`` for this purpose. I considered using `LLVM_DEPRECATED`, but thought requiring a fix-it was not necassery, nor would the syntax in ODS have been very nice.

My motivation for this change is that in the future I'd like to deprecate the use `build` methods in the LLVM Dialect, not using explicit pointer types for ops such as `llvm.load` or `llvm.alloca`, which makes the code not future proof for opaque pointers. In-tree has to be clean first before I could commit such a change of course, but I thought the initial infrastructure change could already be submitted.

Differential Revision: https://reviews.llvm.org/D143190

17 months ago[mlir] Switch default Fold API to using FoldAdaptors
Markus Böck [Tue, 7 Feb 2023 15:38:28 +0000 (16:38 +0100)]
[mlir] Switch default Fold API to using FoldAdaptors

As mentioned in https://discourse.llvm.org/t/psa-new-improved-fold-method-signature-has-landed-please-update-your-downstream-projects/67618, it has now been about one month since the new fold API has been introduced.

Any downstream users that have not yet switched, nor explicitly set useFoldAPI to kEmitRawAttributesFolder will experience a build break after this patch, requiring them to either switch to the new API or explicitly use the old one and live with the warning set by TableGen.

Differential Revision: https://reviews.llvm.org/D143497

17 months ago[WebAssembly][test][NFC] Regenerate funcref-call.ll using update_llc_test_checks.py
Alex Bradbury [Tue, 7 Feb 2023 15:44:24 +0000 (15:44 +0000)]
[WebAssembly][test][NFC] Regenerate funcref-call.ll using update_llc_test_checks.py

In preparation for some slight expansion of the tests.

17 months ago[libc++] Remove #error about _LIBCPP_DEBUG
Louis Dionne [Tue, 7 Feb 2023 15:42:30 +0000 (07:42 -0800)]
[libc++] Remove #error about _LIBCPP_DEBUG

This doesn't require a release note since this used to be an error,
so nobody should have been using it anymore.

17 months ago[libc++] Use __builtin_FILE and __builtin_LINE in _LIBCPP_ASSERT
Louis Dionne [Tue, 7 Feb 2023 00:11:14 +0000 (16:11 -0800)]
[libc++] Use __builtin_FILE and __builtin_LINE in _LIBCPP_ASSERT

This avoids theoretical potential ODR violations since __LINE__ and
__FILE__ expand to different things depending on where they are.

Differential Revision: https://reviews.llvm.org/D143443

17 months ago[libc++][NFC] Reformat __assert with clang-format
Louis Dionne [Tue, 7 Feb 2023 15:34:29 +0000 (07:34 -0800)]
[libc++][NFC] Reformat __assert with clang-format

17 months ago[Flang] Restrict emit-asm-from-mlir.mlir to be ran on specific targets using 'target...
Amy Kwan [Tue, 7 Feb 2023 15:03:24 +0000 (09:03 -0600)]
[Flang] Restrict emit-asm-from-mlir.mlir to be ran on specific targets using 'target=...'

This patch aims to restrict the emit-asm-from-mlir.mlir test to the targets that
the test cases expect using the 'target=...' LIT expression.

rG9067de2a43145411eff8b7782a404c5834249717 has previously attempted to resolve
the LIT failure on the ppc64le-flang-rhel-clang buildbot coming from the
emit-asm-from-mlir.mlir test by restricting the test case for AArch64, x86 and
ARM (as long as these targets are built). The original solution of adding
"*-registered-target" works on ppc64le-flang-rhel-clang because this bot only
builds PowerPC in LLVM_TARGETS_TO_BUILD.

When we build all targets and also build the flang project, this test case will
fail because despite being on PowerPC, we are still building the AArch64, x86
and ARM targets which will cause this test to run. These targets expect the ret
instruction to be produced in this test case, while PowerPC produces a blr instead.

Differential Revision: https://reviews.llvm.org/D143366

17 months ago[libc++][format] Fixes constexpr validation.
Mark de Wever [Mon, 6 Feb 2023 15:29:30 +0000 (16:29 +0100)]
[libc++][format] Fixes constexpr validation.

The constexpr validation parsed parts of the format string that didn't
belong to the specific replacement field.

Fixes https://llvm.org/PR60536

Reviewed By: #libc, ldionne

Differential Revision: https://reviews.llvm.org/D143402

17 months ago[flang][hlfir] Emit hlfir.declare inside internal procedures
Jean Perier [Tue, 7 Feb 2023 14:29:35 +0000 (15:29 +0100)]
[flang][hlfir] Emit hlfir.declare inside internal procedures

Captured variables inside internal procedure do not go though
Fortran::lower::instantiateVar because the specification expressions
should no be lowered again, and instead, all the information must be
taken from the host link argument.

There is nothing very special to do for HLFIR, but the hlfir.declare
should be emitted for the instantiated captured variable and mapped
to the symbol.

Differential Revision: https://reviews.llvm.org/D143481

17 months ago[flang][hlfir] Lower evaluate::DescriptorInquiry
Jean Perier [Tue, 7 Feb 2023 14:27:19 +0000 (15:27 +0100)]
[flang][hlfir] Lower evaluate::DescriptorInquiry

Lower extents and lower bounds inquiries with a compile time
constant DIM that cannot be folded by the front-end.

Differential Revision: https://reviews.llvm.org/D143476

17 months ago[ARM] Perform lane interleaving from reductions.
David Green [Tue, 7 Feb 2023 14:12:51 +0000 (14:12 +0000)]
[ARM] Perform lane interleaving from reductions.

We have a pass for MVE to perform lane interleaving to make use of top/bottom
instructions, that adds shuffles before extends and after truncates. This
extends it to also start from add reductions, where the order of lanes does not
matter so the shuffle is not needed. We need to be careful about not breaking
the form of existing reductions, but otherwise can save some instructions and
awkward extends.

Differential Revision: https://reviews.llvm.org/D143396

17 months agoFix broken Bazel targets
Guillaume Chatelet [Tue, 7 Feb 2023 14:00:56 +0000 (14:00 +0000)]
Fix broken Bazel targets

17 months ago[clangd] Fix a crash in semantic highlighting.
Haojian Wu [Tue, 7 Feb 2023 11:40:00 +0000 (12:40 +0100)]
[clangd] Fix a crash in semantic highlighting.

We encounter a few internal reports that we're dereference a nullptr.
Unfortunately, no small reproduce testcase for this crash yet, but it makes the
clangd more robost on broken code.

Differential Revision: https://reviews.llvm.org/D143486

17 months ago[include-mapping] Regenerate the StdSymbolMap.inc from the 20220730 html book.
Viktoriia Bakalova [Fri, 3 Feb 2023 09:02:01 +0000 (10:02 +0100)]
[include-mapping] Regenerate the StdSymbolMap.inc from the 20220730 html book.

Merge the zoombie names into the StdSymbols as well.

Differential Revision: https://reviews.llvm.org/D143054

17 months ago[mlir][scf] Fix builder of WhileOp with region builder arguments.
Ingo Müller [Tue, 31 Jan 2023 06:39:31 +0000 (06:39 +0000)]
[mlir][scf] Fix builder of WhileOp with region builder arguments.

The overload of WhileOp::build with arguments for builder functions for
the regions of the op was broken: It did not compute correctly the types
(and locations) of the region arguments, which lead to failed assertions
when the result types were different from the operand types.
Specifically, it used the result types (and operand locations) for *both*
regions, instead of the operand types (and locations) for the 'before'
region and the result types (and loecations) for the 'after' region.

Reviewed By: Mogball, mehdi_amini

Differential Revision: https://reviews.llvm.org/D142952

17 months ago[include-mapping] Better #includes support for std input/output symbols
Haojian Wu [Tue, 7 Feb 2023 09:37:44 +0000 (10:37 +0100)]
[include-mapping] Better #includes support for std input/output symbols

Reviewed By: kadircet

Differential Revision: https://reviews.llvm.org/D143280

17 months ago[Tooling/Inclusion] Support multiple headers in the stdlib APIs.
Haojian Wu [Fri, 3 Feb 2023 22:50:25 +0000 (23:50 +0100)]
[Tooling/Inclusion] Support multiple headers in the stdlib APIs.

Unittest will be added in https://reviews.llvm.org/D143280

Patch by Viktoriia Bakalova.

Differential Revision: https://reviews.llvm.org/D142092

17 months ago[libc][NFC] Move attributes from common to macros folder
Guillaume Chatelet [Tue, 7 Feb 2023 13:16:33 +0000 (13:16 +0000)]
[libc][NFC] Move attributes from common to macros folder

17 months ago[libc][NFC] Sort dependencies
Guillaume Chatelet [Tue, 7 Feb 2023 12:58:26 +0000 (12:58 +0000)]
[libc][NFC] Sort dependencies

17 months ago[libc][NFC] Use LIBC_TARGET_IS_GPU instead of defining PACKAGE_FOR_GPU
Guillaume Chatelet [Tue, 7 Feb 2023 12:57:52 +0000 (12:57 +0000)]
[libc][NFC] Use LIBC_TARGET_IS_GPU instead of defining PACKAGE_FOR_GPU

17 months ago[NFC][TargetParser] Remove llvm/ADT/Triple.h
Archibald Elliott [Tue, 7 Feb 2023 12:21:51 +0000 (12:21 +0000)]
[NFC][TargetParser] Remove llvm/ADT/Triple.h

I also ran `git clang-format` to get the headers in the right order for
the new location, which has changed the order of other headers in two
files.

17 months ago[debug-info] Followup to e0374fb2f4, avoid changing codegen
Jeremy Morse [Tue, 7 Feb 2023 12:20:32 +0000 (12:20 +0000)]
[debug-info] Followup to e0374fb2f4, avoid changing codegen

As discussed on D140404, calling GetValueInMiddleOfBlock can force PHI
nodes to be created, but the patch was only supposed to fetch a Value if it
wasn't going to generate any PHIs. Fix this by using GetValueAtEndOfBlock
instead of the Middle variant, which was the original intention given the
check of HasValueForBlock.

We don't need to consider the "middle of block" scenario as we're only
dealing with blocks where there are no value definitions.

17 months ago[CodeGen] Make more use of MachineOperand::getOperandNo. NFC.
Jay Foad [Fri, 3 Feb 2023 10:04:28 +0000 (10:04 +0000)]
[CodeGen] Make more use of MachineOperand::getOperandNo. NFC.

Differential Revision: https://reviews.llvm.org/D143252

17 months ago[CodeGen] Define and use MachineOperand::getOperandNo
Jay Foad [Fri, 3 Feb 2023 09:42:25 +0000 (09:42 +0000)]
[CodeGen] Define and use MachineOperand::getOperandNo

This is a helper function to very slightly simplify many calls to
MachineInstruction::getOperandNo.

Differential Revision: https://reviews.llvm.org/D143250

17 months ago[AMDGPU] Remove the unused u8imm operand definition.
Ivan Kosarev [Tue, 7 Feb 2023 11:38:34 +0000 (11:38 +0000)]
[AMDGPU] Remove the unused u8imm operand definition.

Reviewed By: foad

Differential Revision: https://reviews.llvm.org/D142193

17 months ago[LoopVectorize] Use DataLayout::getIndexType instead of i32 for non-constant GEP...
Sander de Smalen [Mon, 6 Feb 2023 11:03:52 +0000 (11:03 +0000)]
[LoopVectorize] Use DataLayout::getIndexType instead of i32 for non-constant GEP indices.

This is specifically relevant for loops that vectorize using a scalable VF,
where the code results in:

  %vscale = call i32 llvm.vscale.i32()
  %vf.part1 = mul i32 %vscale, 4
  %gep = getelementptr  ..., i32 %vf.part1

Which InstCombine then changes into:

  %vscale = call i32 llvm.vscale.i32()
  %vf.part1 = mul i32 %vscale, 4
  %vf.part1.zext = sext i32 %vf.part1 to i64
  %gep = getelementptr  ..., i32 %vf.part1.zext

D143016 tried to remove these extends, but that only works when
the call to llvm.vscale.i32() has a single use. After doing any
kind of CSE on these calls the combine no longer kicks in.

It seems more sensible to ask DataLayout what type to use, rather
than relying on InstCombine to insert the extend and hoping it can
fold it away.

I've only changed this for indices that are not constant, because
I vaguely remember there was a reason for sticking with i32. It
would also mean patching up loads more tests.

Reviewed By: paulwalker-arm

Differential Revision: https://reviews.llvm.org/D143267

17 months ago[clang-tidy] Fix `cppcoreguidelines-init-variables` for invalid vardecl
Sockke [Tue, 7 Feb 2023 11:30:05 +0000 (19:30 +0800)]
[clang-tidy] Fix `cppcoreguidelines-init-variables` for invalid vardecl

https://godbolt.org/z/n4cK4fo3o
The checker missed a check for invalid vardecl and this could cause a false positive.

Reviewed By: carlosgalvezp

Differential Revision: https://reviews.llvm.org/D138655

17 months ago[AArch64][SME2] Add multi-vector frint intrinsics
Kerry McLaughlin [Tue, 7 Feb 2023 11:21:37 +0000 (11:21 +0000)]
[AArch64][SME2] Add multi-vector frint intrinsics

Adds x2 and x4 vector intrinsics for the following instructions:
 - frinta
 - frintm
 - frintn
 - frintp

NOTE: These intrinsics are still in development and are subject to future changes.

Reviewed By: david-arm

Differential Revision: https://reviews.llvm.org/D143058

17 months ago[ARM] Tests for lane interleaving from reductions. NFC
David Green [Tue, 7 Feb 2023 11:16:11 +0000 (11:16 +0000)]
[ARM] Tests for lane interleaving from reductions. NFC

17 months ago[libc][NFC] Rename compiler_feature macros
Guillaume Chatelet [Tue, 7 Feb 2023 11:05:53 +0000 (11:05 +0000)]
[libc][NFC] Rename compiler_feature macros

17 months ago[libc][NFC] Move sanitizer to macros folder
Guillaume Chatelet [Tue, 7 Feb 2023 10:58:13 +0000 (10:58 +0000)]
[libc][NFC] Move sanitizer to macros folder