platform/upstream/llvm.git
16 months ago[clang] Replace uses of CreateElementBitCast (NFC)
Youngsuk Kim [Fri, 30 Jun 2023 21:33:28 +0000 (17:33 -0400)]
[clang] Replace uses of CreateElementBitCast (NFC)

Partial progress towards replacing uses of CreateElementBitCast, as it
no longer does what its name suggests.

Reviewed By: barannikov88

Differential Revision: https://reviews.llvm.org/D154229

16 months ago[mlir][sparse] make UnpackOp return the actual filled length of unpacked memory
Peiming Liu [Fri, 30 Jun 2023 18:07:21 +0000 (18:07 +0000)]
[mlir][sparse] make UnpackOp return the actual filled length of unpacked memory

This might simplify frontend implementation by avoiding recomputation for the same value.

Reviewed By: aartbik

Differential Revision: https://reviews.llvm.org/D154244

16 months ago[gn build] Port c4e98722ca79
LLVM GN Syncbot [Fri, 30 Jun 2023 21:17:57 +0000 (21:17 +0000)]
[gn build] Port c4e98722ca79

16 months ago[libc] Add missing cast in x86 big_endian_cmp_mask
Roland McGrath [Fri, 30 Jun 2023 20:27:52 +0000 (13:27 -0700)]
[libc] Add missing cast in x86 big_endian_cmp_mask

Implicit narrowing conversions from int to uint16_t
get a compiler warning with the warning settings used
in the Fuchsia build.

Reviewed By: michaelrj

Differential Revision: https://reviews.llvm.org/D154256

16 months ago[LV] Pass bool to requiresScalarEpilogue (NFC).
Florian Hahn [Fri, 30 Jun 2023 21:08:27 +0000 (22:08 +0100)]
[LV] Pass bool to requiresScalarEpilogue (NFC).

requiresScalarEpilogue only checks if the selected VF is vectorizing
(and not scalar). Update it to just take a boolean, to make it clearer
what information is used and to allow callers without a VF (used in a
follow-up patch).

16 months agoAdd Int4 support for tosa::ConstOp
Jerry Ge [Fri, 30 Jun 2023 20:49:58 +0000 (13:49 -0700)]
Add Int4 support for tosa::ConstOp

- Also added Tosa_Weight and Tosa_WeightTensorXD specifically for weights

Signed-off-by: Jerry Ge <jerry.ge@arm.com>
Reviewed By: eric-k256

Differential Revision: https://reviews.llvm.org/D153646

16 months ago[scudo] Improve Secondary Cache Dump
fernandosalas [Fri, 30 Jun 2023 20:55:40 +0000 (20:55 +0000)]
[scudo] Improve Secondary Cache Dump

Exclude cached blocks with invalid start address. Mainly concerned with
cached blocks that are still available/unused.

Reviewed By: Chia-hungDuan, cferris

Differential Revision: https://reviews.llvm.org/D154148

16 months ago[scudo] Restore Freelist when we fail on PageMap creation
Chia-hung Duan [Thu, 29 Jun 2023 04:23:21 +0000 (04:23 +0000)]
[scudo] Restore Freelist when we fail on PageMap creation

Differential Revision: https://reviews.llvm.org/D154041

16 months ago[msan] Add lit test for dladdr
Thurston Dang [Fri, 30 Jun 2023 20:18:53 +0000 (20:18 +0000)]
[msan] Add lit test for dladdr

This adds a simple lit test for dladdr. (There is currently a dladdr testcase in
compiler-rt/lib/msan/tests/msan_test.cpp, but not a stand-alone lit test.)

This is an incremental step towards adding a test case (and, eventually, an interceptor)
for the more complicated dladdr1 function.

Differential Revision: https://reviews.llvm.org/D154255

16 months agocorrect misspelled _LIBCPP___TYPE_TRAITS_IS_EQUAILITY_COMPARABLE_H to _LIBCPP___TYPE_...
Amirreza Ashouri [Fri, 30 Jun 2023 20:50:27 +0000 (13:50 -0700)]
correct misspelled _LIBCPP___TYPE_TRAITS_IS_EQUAILITY_COMPARABLE_H to _LIBCPP___TYPE_TRAITS_IS_EQUALITY_COMPARABLE_H

I just corrected a mistakenly misspelled macro

Reviewed By: philnik, #libc

Spies: libcxx-commits, #llvm, philnik

Differential Revision: https://reviews.llvm.org/D154252

16 months ago[libc++] Fix std::copy and std::move for ranges with potentially overlapping tail...
Nikolas Klauser [Fri, 30 Jun 2023 20:47:39 +0000 (13:47 -0700)]
[libc++] Fix std::copy and std::move for ranges with potentially overlapping tail padding

This fixes thr bug reported in https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108846.

Reviewed By: ldionne, #libc

Spies: mstorsjo, libcxx-commits

Differential Revision: https://reviews.llvm.org/D151953

16 months ago[RISCV] Fix vfwcvt/vfncvt pseudos w/ rounding mode lowering
Luke Lau [Fri, 30 Jun 2023 16:44:47 +0000 (17:44 +0100)]
[RISCV] Fix vfwcvt/vfncvt pseudos w/ rounding mode lowering

Some signed opcodes were being lowered to their unsigned counterparts and
vice-versa.

Reviewed By: craig.topper

Differential Revision: https://reviews.llvm.org/D154234

16 months ago[test] Add a couple of missing REQUIRES
Justin Bogner [Fri, 30 Jun 2023 20:41:23 +0000 (14:41 -0600)]
[test] Add a couple of missing REQUIRES

These tests only work if the targets in the input are available.

16 months ago[flang][openmp] Parallel reduction FIR lowering
Ethan Luis McDonough [Fri, 30 Jun 2023 20:32:21 +0000 (15:32 -0500)]
[flang][openmp] Parallel reduction FIR lowering

This patch extends the logic for lowering loop construct reductions to parallel block reductions.

Reviewed By: kiranchandramohan

Differential Revision: https://reviews.llvm.org/D154182

16 months ago[flang][NFC] Fix typos
Valentin Clement [Fri, 30 Jun 2023 20:15:25 +0000 (13:15 -0700)]
[flang][NFC] Fix typos

16 months ago[UBSan] Fix shared library linking in test for Darwin.
usama hameed [Fri, 30 Jun 2023 19:58:55 +0000 (12:58 -0700)]
[UBSan] Fix shared library linking in test for Darwin.

Without rpath information, the test fails on some Darwin platforms since the binary fails to find the shared library.

rdar://110010041

Differential Revision: https://reviews.llvm.org/D154003

16 months ago[OpenMP][NVPTX] Handle additional invalid PTX characters
Jeffrey Sandoval [Fri, 30 Jun 2023 19:57:05 +0000 (14:57 -0500)]
[OpenMP][NVPTX] Handle additional invalid PTX characters

For OpenMP offload, Clang emits global symbols containing the string
'<captured>', which contains characters that are invalid in PTX.
Extend the existing pass that replaces '.'  and '@' characters with
'_$_' to also replace '<' and '>' characters.

Reviewed By: cchen

Differential Revision: https://reviews.llvm.org/D154241

16 months ago[openacc][NFC] Bump parser support number to OpenACC 3.3
Valentin Clement [Fri, 30 Jun 2023 19:56:00 +0000 (12:56 -0700)]
[openacc][NFC] Bump parser support number to OpenACC 3.3

Parser support reached OpenACC 3.3 specification.
Bump the numbers to reflect the latest specs.

Reviewed By: razvanlupusoru, awarzynski

Differential Revision: https://reviews.llvm.org/D154249

16 months ago[llvm-exegesis] Make sure auxiliary memory is unlinked after usage
Aiden Grossman [Fri, 30 Jun 2023 19:42:37 +0000 (12:42 -0700)]
[llvm-exegesis] Make sure auxiliary memory is unlinked after usage

Currently the shm_unlink is never called on the auxiliary memory shared
memory, so it stays around after running llvm-exegesis (including
running the llvm-exegesis tests). This patch adds the auxiliary memory
shared memory name to the SharedMemoryNames list so that it gets
unlinked when the destructor of SubprocessMemory is called.

16 months agoRevert "[SimplifyCFG] Allow dropping block that only contains ephemeral values"
Nikita Popov [Fri, 30 Jun 2023 19:05:59 +0000 (21:05 +0200)]
Revert "[SimplifyCFG] Allow dropping block that only contains ephemeral values"

This reverts commit 20f0c68fd83a0147a8ec1722bd2e848180610288.

https://reviews.llvm.org/D153966#4464594 reports an optimization
regression in Rust.

Additionally this change has caused an unexpected 0.3% compile-time
regression.

16 months ago[mlir][Vector] Update the lowering of `vector.transfer_write` to SCF
Andrzej Warzynski [Fri, 30 Jun 2023 15:04:28 +0000 (16:04 +0100)]
[mlir][Vector] Update the lowering of `vector.transfer_write` to SCF

This change updates the lowering of `vector.transfer_write` to SCF when
scalable vectors are used. Specifically, when lowering
`vector.transfer_write` to a loop of `vector.extractelement` ops, make
sure that the upper bound of the generated loop is scaled by
`vector.vscale`:
```
    %10 = vector.vscale
    %11 = arith.muli %10, %c16 : index
    scf.for %arg2 = %c0 to %11 step %c1
```

For reference, this is the current version (i.e. before this change):
```
    scf.for %arg2 = %c0 to %c16 step %c1
```
Note that this only valid for fixed-width vectors.

Differential Revision: https://reviews.llvm.org/D154226

16 months ago[mlir][sparse][gpu] Add explaining string to three static_assert stmts
Kun Wu [Fri, 30 Jun 2023 18:32:13 +0000 (13:32 -0500)]
[mlir][sparse][gpu] Add explaining string to three static_assert stmts

Differential Revision: https://reviews.llvm.org/D154243

16 months ago[OpenMP] Added memory scope to atomic::inc API and used the device scope in reduction.
Dhruva Chakrabarti [Fri, 30 Jun 2023 18:02:12 +0000 (14:02 -0400)]
[OpenMP] Added memory scope to atomic::inc API and used the device scope in reduction.

With https://reviews.llvm.org/D137524, memory scope and ordering
attributes are being used to generate the required instructions for
atomic inc/dec on AMDGPU. This patch adds the memory scope attribute to
the atomic::inc API and uses the device scope in reduction. Without
the device scope in atomic_inc, the default system scope leads to
unnecessary L2 write-backs/invalidates.

Reviewed By: arsenm

Differential Revision: https://reviews.llvm.org/D154172

16 months agoOpenMP: Don't use target regions in library function test
Matt Arsenault [Tue, 13 Jun 2023 21:22:29 +0000 (17:22 -0400)]
OpenMP: Don't use target regions in library function test

Use pragma omp declare target to reduce the IR noise around the call
we're actually trying to test.

16 months agoAMDGPU: Fix opaque pointer conversion error in test
Matt Arsenault [Mon, 26 Jun 2023 22:36:06 +0000 (18:36 -0400)]
AMDGPU: Fix opaque pointer conversion error in test

The * was in the wrong place so this was missed by the script.

16 months agoAMDGPU: Fold away mbcnt.hi in wave32 mode
Matt Arsenault [Thu, 29 Jun 2023 14:07:28 +0000 (10:07 -0400)]
AMDGPU: Fold away mbcnt.hi in wave32 mode

This will allow libraries to drop some of the special casing based on
wave size.

16 months agoAMDGPU: Add baseline tests for mbcnt.hi combine
Matt Arsenault [Thu, 29 Jun 2023 14:05:20 +0000 (10:05 -0400)]
AMDGPU: Add baseline tests for mbcnt.hi combine

16 months ago[SLP] Fix crash on attempt to access on invalid iterator state.
Valery N Dmitriev [Fri, 30 Jun 2023 01:15:57 +0000 (18:15 -0700)]
[SLP] Fix crash on attempt to access on invalid iterator state.

The patch fixes corner case when no of scalar instructions
required scheduling for vectorized node.

Differential Revision: https://reviews.llvm.org/D154175

16 months ago[WebAssembly] Fix incorrect assertion in SIMD reduction codegen
Thomas Lively [Fri, 30 Jun 2023 18:30:17 +0000 (11:30 -0700)]
[WebAssembly] Fix incorrect assertion in SIMD reduction codegen

The codegen routine introduced in 18077e9fd688 did not account for vectors with
more than 16 lanes. Remove the incorrect assertion and bail out of the
optimization when encountering this case. Add test cases that previously
triggered the assertion. Unfortunately, these test cases now have terrible
codegen, but that is at least better than crashing.

Fixes #63500.

Differential Revision: https://reviews.llvm.org/D154124

16 months ago[flang][hlfir] probe ultimate intrinsic module procedure name
Jean Perier [Fri, 30 Jun 2023 18:26:59 +0000 (20:26 +0200)]
[flang][hlfir] probe ultimate intrinsic module procedure name

HLFIR lowering was currently hitting errors like "not yet implemented:
intrinsic: z0$iso_c_binding$c_associated_c_ptr". These error were caused
by HLFIR lowering using the name with the full context instead of
the ultimate one.
I do not want to change proc.GetName() because I think it may be best
if it retains the full context for its usages in semantics, so I fixed
the helper to get the name in lowering to HLFIR.

Differential Revision: https://reviews.llvm.org/D154230

16 months ago[flang] add nested DEC STRUCTURE in DerivedTypeDetails component names
Jean Perier [Fri, 30 Jun 2023 18:24:46 +0000 (20:24 +0200)]
[flang] add nested DEC STRUCTURE in DerivedTypeDetails component names

Currently, when a (legacy) DEC structure contained other DEC structure
declarations, the related component names were not added to the
containing DerivedTypeDetails component_names. This lead to bugs in
later phase when visiting the components (like in when lowering the
type to FIR/MLIR).

When an EntityDecl is visited and the scope is a DEC structure, add
the entity to the component names of this DEC structure.

Differential Revision: https://reviews.llvm.org/D154216

16 months ago[flang][openacc] Lower default clause to acc.data operation attributes
Valentin Clement [Fri, 30 Jun 2023 18:21:32 +0000 (11:21 -0700)]
[flang][openacc] Lower default clause to acc.data operation attributes

Lowering of the default clause was missing on the acc.data operation. This
patch lowers the default clause to its corresponding attribute.

Reviewed By: razvanlupusoru

Differential Revision: https://reviews.llvm.org/D154162

16 months ago[flang][openacc] Lower async/wait clauses for acc.data
Valentin Clement [Fri, 30 Jun 2023 18:12:14 +0000 (11:12 -0700)]
[flang][openacc] Lower async/wait clauses for acc.data

Lower async and wait clauses to the correponding
attributes and operands on the acc.data operation.

Depends on D154131

Reviewed By: razvanlupusoru

Differential Revision: https://reviews.llvm.org/D154155

16 months ago[X86] Add common SSE2/SSSE3 check prefix to vector truncation tests
Simon Pilgrim [Fri, 30 Jun 2023 17:48:50 +0000 (18:48 +0100)]
[X86] Add common SSE2/SSSE3 check prefix to vector truncation tests

Reduced duplicate checks where we can

16 months ago[dfsan] Add missing functions to done_abilist.txt
Jin Xin Ng [Sat, 24 Jun 2023 01:42:21 +0000 (01:42 +0000)]
[dfsan] Add missing functions to done_abilist.txt

__sanitizer_get_allocated_begin and __sanitizer_get_allocated_size_fast
were introduced recently in D147005 and D151360, but not added here,
leading to linker errors.

Differential Revision: https://reviews.llvm.org/D153680

16 months ago[flang][openacc] Lower the default clause on compute constructs
Valentin Clement [Fri, 30 Jun 2023 17:24:21 +0000 (10:24 -0700)]
[flang][openacc] Lower the default clause on compute constructs

Lower default clause to the correct attribute for compute
operation.

Reviewed By: razvanlupusoru

Differential Revision: https://reviews.llvm.org/D154165

16 months ago[mlir][openacc] Add wait attribute and variadic operand
Valentin Clement [Fri, 30 Jun 2023 17:04:17 +0000 (10:04 -0700)]
[mlir][openacc] Add wait attribute and variadic operand

OpenACC 3.2 allowed the wait clause to the data construct. This patch
adds a unit attribute and a variadic operand to the data operation to model
the wait clause in a similar way it was added to other data operation.
The attribute models the presence of the clause without any argument. When
arguments are provided they are placed in the wait operand.

Depends on D154111

Reviewed By: razvanlupusoru

Differential Revision: https://reviews.llvm.org/D154131

16 months ago[mlir][openacc] Add async attribute and optional operand
Valentin Clement [Fri, 30 Jun 2023 17:03:24 +0000 (10:03 -0700)]
[mlir][openacc] Add async attribute and optional operand

OpenACC 3.2 allowed the async clause to the data construct. This patch
adds a unit attribute and an optional operand to the data operation to model
the data clause in a similar way it was added to other data operation.
The attribute models the presence of the clause without any argument. When
an argument is provided it is placed in the async operand.

Reviewed By: razvanlupusoru

Differential Revision: https://reviews.llvm.org/D154111

16 months ago[gn build] Port 62e4c22c95a9
LLVM GN Syncbot [Fri, 30 Jun 2023 16:58:05 +0000 (16:58 +0000)]
[gn build] Port 62e4c22c95a9

16 months ago[clang] Fix ASTUnit working directory handling
Hamish Knight [Fri, 30 Jun 2023 16:46:52 +0000 (09:46 -0700)]
[clang] Fix ASTUnit working directory handling

Fix a couple of issues with the handling of the current working directory in ASTUnit:

- Use `createPhysicalFileSystem` instead of `getRealFileSystem` to avoid affecting the process' current working directory, and set it at the top of `ASTUnit::LoadFromCommandLine` such that the driver used for argument parsing and the ASTUnit share the same VFS. This ensures that '-working-directory' correctly sets the VFS working directory in addition to the FileManager working directory.
- Ensure we preserve the FileSystemOptions set on the FileManager when re-creating it (as `ASTUnit::Reparse` will clear the currently set FileManager).

rdar://110697657

Reviewed By: bnbarham, benlangmuir

Differential Revision: https://reviews.llvm.org/D154134

16 months ago[HIP] Fix -mllvm option for device lld linker
Yaxun (Sam) Liu [Thu, 29 Jun 2023 20:02:13 +0000 (16:02 -0400)]
[HIP] Fix -mllvm option for device lld linker

currently clang passes -mllvm options to the device lld linker plugin
when compiling HIP. This is against default clang behavior
which is only passing -mllvm options to linker plugin specified through -Wl
options. This patch lets clang only pass -Xoffload-linker -mllvm= options
to device lld linker plugin.

Fixes: https://github.com/llvm/llvm-project/issues/63604

Reviewed by: Joseph Huber, Matt Arsenault

Differential Revision: https://reviews.llvm.org/D154145

16 months agoclang: Add start of header test for __clang_hip_libdevice_declares
Matt Arsenault [Tue, 13 Jun 2023 14:05:44 +0000 (10:05 -0400)]
clang: Add start of header test for __clang_hip_libdevice_declares

It was shockingly difficult to get identical checks out of the 3
languages. The visibility, dso_local and fp-contract settings seem to
differ unless I force every test function to
static. update_cc_test_checks also isn't smart enough when the
attribute groups slightly differ.

16 months agoVerifier: Verify absolute_symbol metadata
Matt Arsenault [Wed, 28 Jun 2023 12:57:19 +0000 (08:57 -0400)]
Verifier: Verify absolute_symbol metadata

This is the same as !range except for one edge case.

16 months agoVerifier: Fix assertion on range metadata with equal bounds
Matt Arsenault [Wed, 28 Jun 2023 13:14:06 +0000 (09:14 -0400)]
Verifier: Fix assertion on range metadata with equal bounds

This only worked if the same values were the min or max. We also seem
to be missing proper assembler tests for this.

16 months agoAnalysis: Fix assertion when load alignment exceeds address space size
Matt Arsenault [Mon, 19 Jun 2023 22:03:33 +0000 (18:03 -0400)]
Analysis: Fix assertion when load alignment exceeds address space size

Apparently the maximum alignment no longer fits in 32-bits now, which
overflows a 32-bit offset and would fail on the isPowerOf2 assert.

16 months ago[libc] Make the RPC server target always available
Joseph Huber [Thu, 29 Jun 2023 16:23:47 +0000 (11:23 -0500)]
[libc] Make the RPC server target always available

This patch makes sure that we always build the RPC server. The proposed
used for this is to begin integrating this server implementation into
`libomptarget`. That requires that we build this server ahead of time
when using a `LLVM_ENABLE_PROJECTS` build. Make a few tweaks to ensure
that the GCC compiler which may be used for this build doesn't complain.

Reviewed By: JonChesterfield

Differential Revision: https://reviews.llvm.org/D154105

16 months ago[libc] Add other RPC callback methods to the RPC server
Joseph Huber [Fri, 30 Jun 2023 14:48:13 +0000 (09:48 -0500)]
[libc] Add other RPC callback methods to the RPC server

This patch adds the other two methods to the server so the external
users can use the interface through the obfuscated interface.

Reviewed By: JonChesterfield

Differential Revision: https://reviews.llvm.org/D154224

16 months agoReland "[BOLT][Instrumentation] Don't share counters when using append-pid"
Amir Ayupov [Fri, 30 Jun 2023 16:14:58 +0000 (09:14 -0700)]
Reland "[BOLT][Instrumentation] Don't share counters when using append-pid"

This reverts commit c15e9b6814e53bccb0194268a826c1213a84b395.

The issue appears unrelated as the crash happened in the BOLTed binary, not
instrumented binary.

16 months agoMachineFunction: -fsanitize={function,kcfi}: ensure 4-byte alignment
Fangrui Song [Fri, 30 Jun 2023 16:13:19 +0000 (09:13 -0700)]
MachineFunction: -fsanitize={function,kcfi}: ensure 4-byte alignment

Fix https://github.com/llvm/llvm-project/issues/63579
```
% cat a.c
void foo() {}
% clang --target=arm-none-eabi -mthumb -mno-unaligned-access -fsanitize=kcfi a.c -S -o - | grep p2align
        .p2align        1
% clang --target=armv6m-none-eabi -fsanitize=function a.c -S -o - | grep p2align
        .p2align        1
```

Ensure that -fsanitize={function,kcfi} instrumented functions are aligned by at
least 4, so that loading the type hash before the function label will not cause
a misaligned access. This is especially important for -mno-unaligned-access
configurations that don't set `setMinFunctionAlignment` to 4 or greater.

With this patch, the generated assembly for the examples above will contain `.p2align 2`
before the type hash.

If `__attribute__((aligned(N)))` or `-falign-functions=N` is specified, the
larger alignment will be used.

Reviewed By: simon_tatham, samitolvanen

Differential Revision: https://reviews.llvm.org/D154125

16 months ago[clang] Update cxx_status for 632dd6a4ca0036009f
Yuanfang Chen [Fri, 30 Jun 2023 16:09:31 +0000 (09:09 -0700)]
[clang] Update cxx_status for 632dd6a4ca0036009f

missed this after rebase

16 months ago[NVPTX] added a hidden option to control NVPTXLowerUnreachable pass.
Artem Belevich [Thu, 29 Jun 2023 16:31:31 +0000 (09:31 -0700)]
[NVPTX] added a hidden option to control NVPTXLowerUnreachable pass.

We've ran into an issue where the pass breaks a handful of our internal tests
and need a way to temporarily disable the pass while we're investigating.

Differential Revision: https://reviews.llvm.org/D154106

16 months ago[AMDGPU][Codegen] Clean up legalizeOpWithMove().
Ivan Kosarev [Fri, 30 Jun 2023 15:43:02 +0000 (16:43 +0100)]
[AMDGPU][Codegen] Clean up legalizeOpWithMove().

The removed logic was added in
<https://reviews.llvm.org/rG0c93c9ecee0624f8469f5a971a09fbc9e9cc1061>,
but now doesn't seem to be needed.

Reviewed By: arsenm

Differential Revision: https://reviews.llvm.org/D154220

16 months ago[RISCV] Custom lower FP_TO_FP16 and FP16_TO_FP to correct ABI of of libcall
Alex Bradbury [Fri, 30 Jun 2023 15:39:15 +0000 (16:39 +0100)]
[RISCV] Custom lower FP_TO_FP16 and FP16_TO_FP to correct ABI of of libcall

As introduced in D99148, RISC-V uses the softPromoteHalf legalisation
for fp16 values without zfh, with logic ensuring that f16 values are
passed in lower bits of FPRs (see D98670) when F or D support is
present. This legalisation produces ISD::FP_TO_FP16 and ISD::FP16_TO_FP
nodes which (as described in ISDOpcodes.h) provide a "semi-softened
interface for dealing with f16 (as an i16)". i.e. the return type of the
FP_TO_FP16 is an integer rather than a float (and the arg of FP16_TO_FP
is an integer). The remainder of the description focuses primarily on
FP_TO_FP16 for ease of explanation.

FP_TO_FP16 is lowered to a libcall to `__truncsfhf2 (float)` or
`__truncdfhf2 (double)`. As of D92241, `_Float16` is used as the return
type of these libcalls if the host compiler accepts `_Float16` in a test
input (i.e. dst_t is set to `_Float16`). `_Float16` is enabled for the
RISC-V target as of D105001 and so the return value should be passed in
an FPR on hard float ABIs.

This patch fixes the ABI issue in what appears to be a minimally
invasive way - leaving the softPromoteHalf logic undisturbed, and
lowering FP_TO_FP16 to an f32-returning libcall, converting its result
to an XLen integer value.

As can be seen in the test changes, the custom lowering for FP16_TO_FP
means the libcall is no longer tail-callable.

Although this patch fixes the issue, there are two open items:
* Redundant fmv.x.w and fmv.w.x pairs are now somtimes produced during
  lowering (not a correctness issue).
* Now coverage for STRICT variants of FP16 conversion opcodes.

Differential Revision: https://reviews.llvm.org/D151284

16 months ago[RISCV][test] Add additional RUN lines to half-convert.ll in preparation for D151824
Alex Bradbury [Fri, 30 Jun 2023 15:30:15 +0000 (16:30 +0100)]
[RISCV][test] Add additional RUN lines to half-convert.ll in preparation for D151824

There wasn't previous coverage for rv32id-ilp32, rv64id-lp64,
rv32id-ilp32d, or rv64id-lp64d. This is needed as D151284 fixes a bug
related to the ABI used for libcalls for fp<->fp16 conversion when hard
FP support is present.

16 months ago[AMDGPU][AsmParser][NFC] Simplify parsing of sopp_brtarget operands.
Ivan Kosarev [Fri, 30 Jun 2023 15:28:40 +0000 (16:28 +0100)]
[AMDGPU][AsmParser][NFC] Simplify parsing of sopp_brtarget operands.

Also refine the definitions while there.

Part of <https://github.com/llvm/llvm-project/issues/62629>.

Reviewed By: mbrkusanin

Differential Revision: https://reviews.llvm.org/D154061

16 months ago[Headers][doc] Add load/store/cmp/cvt intrinsic descriptions to avx2intrin.h
Paul Robinson [Wed, 28 Jun 2023 18:25:03 +0000 (11:25 -0700)]
[Headers][doc] Add load/store/cmp/cvt intrinsic descriptions to avx2intrin.h

Differential Revision: https://reviews.llvm.org/D153993

16 months ago[ConstraintElim] Add phi tests with multiple GEP indices.
Florian Hahn [Fri, 30 Jun 2023 15:28:37 +0000 (16:28 +0100)]
[ConstraintElim] Add phi tests with multiple GEP indices.

Extra tests for D152730.

16 months ago[mlir][NFC] Fully qualify TableGen generated code handling properties
Markus Böck [Fri, 30 Jun 2023 15:28:15 +0000 (17:28 +0200)]
[mlir][NFC] Fully qualify TableGen generated code handling properties

This makes it possible to compile the TableGen generated files without having to use `using namespace mlir;` or being in the `mlir` namespace.

16 months ago[analyzer][NFC] Fix dangling StringRef in barely used code
Donát Nagy [Tue, 27 Jun 2023 15:49:44 +0000 (17:49 +0200)]
[analyzer][NFC] Fix dangling StringRef in barely used code

CheckerContext::getNoteTag has a shorthand version that takes a
plain 'StringRef Note' instead of a lambda that calculates the note.

The old implementation of this method was incorrect because it created a
lambda that captured the StringRef, which was dereferenced much later,
when the NoteTags were visited.

In the current codebase this does not cause errors because this method
is called only once, and there the `Note` argument is a string literal
that remains valid. However, I tried to use this method in a checker
that I was prototyping, and there it printed random memory junk (instead
of the message that I composed in a local variable).

Differential Revision: https://reviews.llvm.org/D153889

16 months ago[scudo] SizeClassAllocator64: dispatch all the classes in init()
Fabio D'Urso [Fri, 30 Jun 2023 14:35:09 +0000 (16:35 +0200)]
[scudo] SizeClassAllocator64: dispatch all the classes in init()

This will make it easier to verify in a future patch that the ReservedMemory object has been fully dispatched.

Reviewed By: Chia-hungDuan

Differential Revision: https://reviews.llvm.org/D153019

16 months ago[Metarenamer] Introduce option to only change inst names
Anna Thomas [Fri, 30 Jun 2023 14:21:16 +0000 (10:21 -0400)]
[Metarenamer] Introduce option to only change inst names

This is useful when needing to modify IR and test some optimizations on
them, while keeping BB names and function names intact. If
the IR uses ordered number naming (%1, %2, %3 etc), then we cannot just
remove or reorder specific instructions since the verifier expects the
numbers to be in order.

16 months agoReverting commit 0d8d718171192301f2beb10bd08ce62e70281a5e as it broke libomptarget...
Akash Banerjee [Fri, 30 Jun 2023 15:03:57 +0000 (16:03 +0100)]
Reverting commit 0d8d718171192301f2beb10bd08ce62e70281a5e as it broke libomptarget tests

16 months ago[CSKY] Optimize IR pattern icmp-select with INCT32/INCF32
Ben Shi [Wed, 21 Jun 2023 10:05:24 +0000 (18:05 +0800)]
[CSKY] Optimize IR pattern icmp-select with INCT32/INCF32

Reviewed By: zixuan-wu

Differential Revision: https://reviews.llvm.org/D153436

16 months ago[CSKY][test][NFC] Add tests of IR pattern icmp-select
Ben Shi [Tue, 20 Jun 2023 10:58:00 +0000 (18:58 +0800)]
[CSKY][test][NFC] Add tests of IR pattern icmp-select

These tests will be optimized with INCT32/INCF32/DECT32/DECF32
in the future.

Reviewed By: zixuan-wu

Differential Revision: https://reviews.llvm.org/D153434

16 months ago[GVN] Add tests for PR63059 (NFC)
ManuelJBrito [Thu, 29 Jun 2023 17:23:55 +0000 (18:23 +0100)]
[GVN] Add tests for PR63059 (NFC)

Besides the example from the issue, this adds some other related wrong transformations.

Differential Revision: https://reviews.llvm.org/D154103

16 months ago[OpenMP][OpenMPIRBuilder] Migrate kernel launch code and host fallback code generatio...
Jan Sjodin [Thu, 29 Jun 2023 14:41:11 +0000 (10:41 -0400)]
[OpenMP][OpenMPIRBuilder] Migrate kernel launch code and host fallback code generation from Clang to the OpenMPIRBuilder

This patch refactors the code generation that emits the offloading kernel
launch and moves the core portion to the OpenMPIRBuilder so that it can be used
from flang in the future.

Reviewed By: jdoerfert

Differential Revision: https://reviews.llvm.org/D151035

16 months ago[mlir][LLVM] Add section attribute to LLVMFuncOp
Christian Ulmann [Fri, 30 Jun 2023 14:19:31 +0000 (14:19 +0000)]
[mlir][LLVM] Add section attribute to LLVMFuncOp

This commit adds an optional section attribute to the `LLVMFuncOp` and
adds import and export functionality for it.

Reviewed By: gysit

Differential Revision: https://reviews.llvm.org/D154219

16 months ago[dataflow] Fix complie on gcc7 part 2
Donát Nagy [Fri, 30 Jun 2023 13:43:30 +0000 (15:43 +0200)]
[dataflow] Fix complie on gcc7 part 2

Change c2bb68078eb9 triggered a bug in gcc7 by disabling a copy
constructor. Commit 2f7d30dee826 fixed one occurrence of this issue, but
after that the compilation still failed with another very similar error.

This commit eliminates this second compilation error, and ensures that
`ninja check-clang-analysis` can be successfully executed with gcc 7.5.

16 months agollvm-reduce: Fix not preserving uselistorder with bitcode
Matt Arsenault [Sun, 25 Jun 2023 12:55:52 +0000 (08:55 -0400)]
llvm-reduce: Fix not preserving uselistorder with bitcode

Fix accidentally passing pointer to bool argument This was supposed to
be writing bitcode with preserved uselistorder, but instead was only
enabling it with LTO module summaries.

16 months ago[ConstraintElim] Add extra tests and split up test file.
Florian Hahn [Fri, 30 Jun 2023 14:16:23 +0000 (15:16 +0100)]
[ConstraintElim] Add extra tests and split up test file.

Extra tests for D152730, including more tests where the step GEP doesn't
use the phi and inbounds are missing.

Also split up test file to make it easier to verify for Alive2.

16 months ago[OpenMP] Change clang emitTargetDataCalls to use OMPIRBuilder
Akash Banerjee [Tue, 27 Jun 2023 12:50:17 +0000 (13:50 +0100)]
[OpenMP] Change clang emitTargetDataCalls to use OMPIRBuilder

This patch changes the emitTargetDataCalls function in clang to make use of the OpenMPIRBuilder::createTargetData function for Target Data directive code gen.

Depends on D146557

Differential Revision: https://reviews.llvm.org/D150860

16 months ago[SystemZ] Improve error messages for unsupported relocations
Ulrich Weigand [Fri, 30 Jun 2023 14:02:56 +0000 (16:02 +0200)]
[SystemZ] Improve error messages for unsupported relocations

In the SystemZMCObjectWriter, we currently just abort in case
some unsupported relocation in requested.  However, as this
situation can be triggered by invalid (inline) assembler input,
we should really get a regular error message instead.

16 months ago[Attributor] Convert test to opaque pointers (NFC)
Nikita Popov [Fri, 30 Jun 2023 13:52:55 +0000 (15:52 +0200)]
[Attributor] Convert test to opaque pointers (NFC)

The loads and stores of the global are now optimized, as the relevant
parts of the function signatures match.

16 months ago[DAG] Fold (*ext (*_extend_vector_inreg x)) -> (*_extend_vector_inreg x)
Simon Pilgrim [Fri, 30 Jun 2023 13:42:38 +0000 (14:42 +0100)]
[DAG] Fold (*ext (*_extend_vector_inreg x)) -> (*_extend_vector_inreg x)

16 months ago[gn build] Port f13e1a65cabb
LLVM GN Syncbot [Fri, 30 Jun 2023 13:26:44 +0000 (13:26 +0000)]
[gn build] Port f13e1a65cabb

16 months ago[SimplifyCFG] Allow dropping block that only contains ephemeral values
Nikita Popov [Wed, 28 Jun 2023 09:58:16 +0000 (11:58 +0200)]
[SimplifyCFG] Allow dropping block that only contains ephemeral values

Perform the TryToSimplifyUncondBranchFromEmptyBlock() transform if
the block is empty except for ephemeral values. The ephemeral values
will be dropped in that case.

This makes sure that assumes don't block this transforms, as reported
in https://discourse.llvm.org/t/llvm-assume-blocks-optimization/71609.

Differential Revision: https://reviews.llvm.org/D153966

16 months ago[libc++] Revert __uninitialized_buffer changes
Louis Dionne [Thu, 29 Jun 2023 22:09:58 +0000 (18:09 -0400)]
[libc++] Revert __uninitialized_buffer changes

This patch reverts the following commits:

    015cd317eaed28a923d14a33c9d6739012a688be (add missing HIDE_FROM_ABI)
    420a204d52205f1277a8d5df3dbafac6082e02e2 (add _LIBCPP_NO_CFI)
    31eeba3f7c0e2ef4a21c07da9326a4ae1a8de7e2 (add __uninitialized_buffer)

It also reverts a small part of b935ab8e747cf52ff12471879460206a9f433eea
which is required to make the stable_partition.pass.cpp test pass on GCC.

Some issues were pointed out in https://reviews.llvm.org/D152208 and
in https://reviews.llvm.org/D154017, so I am reverting this patch
until we have time to weigh the various solutions and get consensus
on the design of the API.

Differential Revision: https://reviews.llvm.org/D154161

16 months agoDAG: Really fix patch split
Matt Arsenault [Fri, 30 Jun 2023 13:13:44 +0000 (09:13 -0400)]
DAG: Really fix patch split

16 months ago[X86][FP16] Pre-commit test to show a mis-combination
Phoebe Wang [Fri, 30 Jun 2023 13:07:20 +0000 (21:07 +0800)]
[X86][FP16] Pre-commit test to show a mis-combination

16 months agoReapply "clang: Use new frexp intrinsic for builtins and add f16 version"
Matt Arsenault [Fri, 30 Jun 2023 13:04:00 +0000 (09:04 -0400)]
Reapply "clang: Use new frexp intrinsic for builtins and add f16 version"

This reverts commit 0c545a441285a73e00b859dd52f1a85cb9eeeefc.

ARM libcall expansion was fixed in 160d7227e043cd8f491cb706916da953cabfd905

16 months agoDAG: Fix broken patch split
Matt Arsenault [Fri, 30 Jun 2023 13:07:09 +0000 (09:07 -0400)]
DAG: Fix broken patch split

16 months agoDAG: Fix libcall expansion for frexp on ARM
Matt Arsenault [Fri, 30 Jun 2023 12:17:11 +0000 (08:17 -0400)]
DAG: Fix libcall expansion for frexp on ARM

The ExpandLibcallResult result was a bitcast and not the direct call
result, so we couldn't find the chain. Use the new separate chain
return value instead.

16 months agoDAG: Return the chain from ExpandLibCall
Matt Arsenault [Fri, 30 Jun 2023 12:11:12 +0000 (08:11 -0400)]
DAG: Return the chain from ExpandLibCall

If the libcall expansion requires use of the inserted call's result
chain, it's unreliable to query it from the main result. The call
lowering may have added additional casts or other obscuring operations
we don't want to parse through.

16 months ago[libc] Improve memcmp latency and codegen
Guillaume Chatelet [Fri, 30 Jun 2023 12:55:54 +0000 (12:55 +0000)]
[libc] Improve memcmp latency and codegen

This is based on ideas from @nafi to:
 - use a branchless version of 'cmp' for 'uint32_t',
 - completely resolve the lexicographic comparison through vector
   operations when wide types are available. We also get rid of byte
   reloads and serializing '__builtin_ctzll'.

I did not include the suggestion to replace comparisons of 'uint16_t'
with two 'uint8_t' as it did not seem to help the codegen. This can
be revisited in sub-sequent patches.

The code been rewritten to reduce nested function calls, making the
job of the inliner easier and preventing harmful code duplication.

Reviewed By: nafi3000

Differential Revision: https://reviews.llvm.org/D148717

16 months ago[llvm-readobj] Add support for dumping CHPE metadata.
Jacek Caban [Fri, 7 Apr 2023 21:31:16 +0000 (23:31 +0200)]
[llvm-readobj] Add support for dumping CHPE metadata.

CHPE metadata is used by ARM64EC/ARM64X PE files to provide metadata for
emulator/loader. Most of this metadata will need to be generated by LLD.

Differential Revision: https://reviews.llvm.org/D149089

16 months ago[mlir][spirv] Allow custom mangling of SPIRV built-in global variables
Victor Perez [Wed, 28 Jun 2023 08:18:20 +0000 (09:18 +0100)]
[mlir][spirv] Allow custom mangling of SPIRV built-in global variables

The SPIR-V spec does not specify the mangling for these variables, so
the conversion to SPIR-V should be flexible enough to allow adding a
custom prefix and suffix to the core name.

Differential Revision: https://reviews.llvm.org/D153951

Signed-off-by: Victor Perez <victor.perez@codeplay.com>
16 months ago[SystemZ][z/OS] Add required options/macro/etc for z/os compilation step
Sean Perry [Fri, 30 Jun 2023 12:04:43 +0000 (08:04 -0400)]
[SystemZ][z/OS] Add required options/macro/etc  for z/os compilation step

Add the required options and macros to the compilation step for z/os.

Reviewed By: abhina.sreeskantharajan, fanbo-meng

Differential Revision: https://reviews.llvm.org/D153582

16 months ago[mlir][tensor] More static padding sizes for PadOp
Matthias Springer [Fri, 30 Jun 2023 10:44:48 +0000 (12:44 +0200)]
[mlir][tensor] More static padding sizes for PadOp

Improve `tensor::createPadHighOp`: more static padding sizes when possible.

Differential Revision: https://reviews.llvm.org/D154200

16 months ago[mlir][Affine] Delete duplicate code: `applyMapToValues`
Matthias Springer [Fri, 30 Jun 2023 10:44:33 +0000 (12:44 +0200)]
[mlir][Affine] Delete duplicate code: `applyMapToValues`

The same functionality is provided by `makeComposedFoldedAffineApply`.

Differential Revision: https://reviews.llvm.org/D154199

16 months ago[mlir][NFC] Clean up builder usage around constants/non-foldable ops
Matthias Springer [Fri, 30 Jun 2023 10:44:18 +0000 (12:44 +0200)]
[mlir][NFC] Clean up builder usage around constants/non-foldable ops

* Use `create` instead of `createOrFold` for constant ops. Constants cannot be folded any further.
* Use `create` instead of `createOrFold` for ops that do not have a folder.
* Use C++ op builders that take an `int` instead of creating a `ConstantIndexOp`.
* Create `tensor::DimOp` instead of `linalg::createOrFoldDimOp` when it is certain that the operand is a tensor.

Differential Revision: https://reviews.llvm.org/D154196

16 months ago[tests] precommit test for MachineLICM subloops
Jingu Kang [Fri, 30 Jun 2023 11:34:21 +0000 (12:34 +0100)]
[tests] precommit test for MachineLICM subloops

16 months agoRevert "clang: Use new frexp intrinsic for builtins and add f16 version"
Hans Wennborg [Fri, 30 Jun 2023 11:14:17 +0000 (13:14 +0200)]
Revert "clang: Use new frexp intrinsic for builtins and add f16 version"

This caused asserts in some Android and Windows builds:

SelectionDAGNodes.h:1138: llvm::SDValue::SDValue(SDNode *, unsigned int):
Assertion `(!Node || !ResNo || ResNo < Node->getNumValues()) && "Invalid result number for the given node!"' failed.

See comment on https://github.com/llvm/llvm-project/commit/85bdea023f5116f789095b606554739403042a21

Also revert "HIP: Use frexp builtins in math headers"
which seems to depend on this change.

This reverts commit 85bdea023f5116f789095b606554739403042a21.
This reverts commit bf8e92c0e792cbe3c9cc50607a1e33c6912ffd0e.

16 months ago[AArch64] Fold tree of offset loads combine
David Green [Fri, 30 Jun 2023 11:25:07 +0000 (12:25 +0100)]
[AArch64] Fold tree of offset loads combine

This attempts to fold trees of add(ext(load p), shl(ext(load p+4)) into a
single load of twice the size, that we extract the bottom part and top part so
that the shl can start to use a shll2 instruction. The two loads in that
example can also be larger trees of instructions, which are identical except
for the leaves which are all loads offset from the LHS, including buildvectors
of multiple loads. For example:
sub(zext(buildvec(load p+4, load q+4)), zext(buildvec(load r+4, load s+4)))

Whilst it can be common for the larger loads to replace LDP instructions (which
doesn't gain anything on its own), the larger loads in buildvectors can help
create more efficient code, and prevent the need for ld1 lane inserts which can
be more expensive than continuous loads.

This creates a fairly niche, fairly large combine that attempts to be fairly
general where it is beneficial. It helps some SLP vectorized code to avoid the
use of the more expensive ld1 lane inserting loads.

Differential Revision: https://reviews.llvm.org/D153972

16 months ago[AArch64] Codegen tests for fold from D153972. NFC
David Green [Fri, 30 Jun 2023 10:21:26 +0000 (11:21 +0100)]
[AArch64] Codegen tests for fold from D153972. NFC

16 months ago[LV] Add additional tests with dead vector epilogues.
Florian Hahn [Fri, 30 Jun 2023 11:17:56 +0000 (12:17 +0100)]
[LV] Add additional tests with dead vector epilogues.

16 months ago[ARM] armv6m eXecute Only (XO) long branch Thunk
Keith Walker [Thu, 22 Jun 2023 13:53:13 +0000 (14:53 +0100)]
[ARM] armv6m eXecute Only (XO) long branch Thunk

This patch adds a thunk for Thumb long branch on V6-M for eXecute Only.

Note that there is currently no support for a position independant and
eXecute Only V6-M long branch thunk

Differential Revision: https://reviews.llvm.org/D153772

16 months ago[DAG][AArch64] Fold shuffle_vector<4,5,6,7> to extract_subvector
David Green [Fri, 30 Jun 2023 10:13:39 +0000 (11:13 +0100)]
[DAG][AArch64] Fold shuffle_vector<4,5,6,7> to extract_subvector

During legalization, we can end up with shuffles that are identity masks, so
act like extract_subvector, but do not simplify to extract_subvector. This
adjusts the profitability heuristic in foldExtractSubvectorFromShuffleVector to
allow identity vectors that do not start at element 0. Undef masks elements are
excluded as it can be more useful to keep the undef elements.

Differential Revision: https://reviews.llvm.org/D153504

16 months agoRevert "Reland [SLP] Provide an universal interface for FixedVectorType::get. NFC."
Nikita Popov [Fri, 30 Jun 2023 09:27:44 +0000 (11:27 +0200)]
Revert "Reland [SLP] Provide an universal interface for FixedVectorType::get. NFC."

This reverts commit 19b1d3bd7eeecbeb1e45045960faf325c7bc5c64.

Both the commit and the review are missing a patch description.

16 months agoAdd [mlir][nvvm] `mbarrier.test.wait`
Guray Ozen [Thu, 29 Jun 2023 12:49:19 +0000 (14:49 +0200)]
Add [mlir][nvvm] `mbarrier.test.wait`

This work adds `mbarrier.test.wait` and `mbarrier.test.wait.shared` Ops in NVVM dialect. Since they are already implemented in the LLVM kernel, it only calls createIntrinsicCall.

Reviewed By: qcolombet

Differential Revision: https://reviews.llvm.org/D154076

16 months ago[mlir] Add result to mbarrier.arrive
Guray Ozen [Thu, 29 Jun 2023 10:09:31 +0000 (12:09 +0200)]
[mlir] Add result to mbarrier.arrive

`mbarrier.arrive` returns token. This PR adds result to these ops.

Reviewed By: qcolombet

Differential Revision: https://reviews.llvm.org/D154059