platform/upstream/llvm.git
23 months ago[MCA][X86][AVX512] Add test coverage for unsigned<->fp conversion instructions
Simon Pilgrim [Sat, 12 Nov 2022 13:45:16 +0000 (13:45 +0000)]
[MCA][X86][AVX512] Add test coverage for unsigned<->fp conversion instructions

23 months ago[X86] SkylakeServerModel - conversion instructions don't use Port015
Simon Pilgrim [Sat, 12 Nov 2022 12:39:59 +0000 (12:39 +0000)]
[X86] SkylakeServerModel - conversion instructions don't use Port015

Fixes a lot of throughput mismatches - the more complicated conversion instructions use SKXPort5+SKXPort01, not SKXPort5+SKXPort015 (SKXPort015 is mainly used for basic Logic + blend ops)

Fixing this should allow us to remove a lot of unnecessary scheduler overrides from SkylakeServerModel

Confirmed by both Agner + uops.info

23 months ago[X86] Replace unnecessary SKL CVTPD2DQ overrides with better base class defs
Simon Pilgrim [Sat, 12 Nov 2022 12:15:56 +0000 (12:15 +0000)]
[X86] Replace unnecessary SKL CVTPD2DQ overrides with better base class defs

Also fixes some AVX missing folded instructions

23 months ago[X86] Tweak Alderlake instregex to match CodeGen-only and public scalar instruction ops
Simon Pilgrim [Sat, 12 Nov 2022 10:37:33 +0000 (10:37 +0000)]
[X86] Tweak Alderlake instregex to match CodeGen-only and public scalar instruction ops

As detailed on #58792 the _Int postfix needs to be optional in the instregex to match both instructions - fixes mismatch warnings on a scheduler model verifier I'm working on

23 months ago[X86] Replace unnecessary SKL conversion overrides with better base class defs
Simon Pilgrim [Sat, 12 Nov 2022 10:23:52 +0000 (10:23 +0000)]
[X86] Replace unnecessary SKL conversion overrides with better base class defs

Split various conversion instructions that use different scheduler pipes for the reg-reg and reg-mem variants (and not an additional Port23 uop for memory folding) - declare the classes separately instead of using the SKLWriteResPair helper

23 months ago[Support/ELF] - Add OpenBSD PT_OPENBSD_MUTABLE constant.
Brad Smith [Sat, 12 Nov 2022 09:54:41 +0000 (04:54 -0500)]
[Support/ELF] - Add OpenBSD PT_OPENBSD_MUTABLE constant.

OpenBSD commit for reference:
https://github.com/openbsd/src/commit/bd249b5664da50f0178adea78250a7a0d8ea6566

23 months ago[lldb] [cmake] Fix another typo in third-party/unittest path
Michał Górny [Sat, 12 Nov 2022 09:34:57 +0000 (10:34 +0100)]
[lldb] [cmake] Fix another typo in third-party/unittest path

23 months ago[LoongArch] Implement MCTargetExpr::fixELFSymbolsInTLSFixups hook
wanglei [Sat, 12 Nov 2022 08:35:59 +0000 (16:35 +0800)]
[LoongArch] Implement MCTargetExpr::fixELFSymbolsInTLSFixups hook

Reviewed By: SixWeining, MaskRay

Differential Revision: https://reviews.llvm.org/D137628

23 months ago[clang-format] Treats &/&& as reference when followed by ',' or ')'
Micah Weston [Sat, 12 Nov 2022 08:51:35 +0000 (00:51 -0800)]
[clang-format] Treats &/&& as reference when followed by ',' or ')'

Ran into an issue where function declarations inside function
scopes or uses of sizeof inside a function would treat the && in
'sizeof(Type &&)' as a binary operator.

Attempt to fix this by assuming reference when followed by ',' or
')'. Also adds tests for these.

Also hit an edge case in another test that treated "and" the same
as "&&" since it parses as C++. Changed the "and" to "also" so it
is no longer a keyword.

Fixes #58923.

Differential Revision: https://reviews.llvm.org/D137755

23 months ago[clang-format] Correctly annotate function names before attributes
Owen Pan [Sat, 5 Nov 2022 11:09:13 +0000 (04:09 -0700)]
[clang-format] Correctly annotate function names before attributes

Fixes #58827.

Differential Revision: https://reviews.llvm.org/D137486

23 months ago[RISCV] Rename template parameter. NFC
Craig Topper [Sat, 12 Nov 2022 08:31:31 +0000 (00:31 -0800)]
[RISCV] Rename template parameter. NFC

23 months ago[RISCV] Use template to reduce some code. NFC
Craig Topper [Sat, 12 Nov 2022 07:05:34 +0000 (23:05 -0800)]
[RISCV] Use template to reduce some code. NFC

23 months agoAdd back single quotes when dontcall attribute was split into dontcall-error/dontcall...
Fangrui Song [Sat, 12 Nov 2022 06:59:51 +0000 (22:59 -0800)]
Add back single quotes when dontcall attribute was split into dontcall-error/dontcall-warn

Single quotes were accidentally dropped in D110364.

23 months ago[X86] Reduce unnecessary instregex for AlderlakeP schedule model
Haohai Wen [Sat, 12 Nov 2022 04:30:04 +0000 (12:30 +0800)]
[X86] Reduce unnecessary instregex for AlderlakeP schedule model

Using instregex for simple instruction opcode is much slower than
instrs. This patch replaces them with instrs.

Github issue: 35303

Reviewed By: RKSimon

Differential Revision: https://reviews.llvm.org/D137841

23 months ago[AArch64][GlobalISel] Select TBZ for icmp sge x, 0.
Amara Emerson [Sat, 12 Nov 2022 03:54:39 +0000 (19:54 -0800)]
[AArch64][GlobalISel] Select TBZ for icmp sge x, 0.

This results in some nice size improvements on -Os CTMark:
Program                                       size.__text
                                              sdag           gisel           diff
consumer-typeset/consumer-typeset             414124.00      414052.00      -0.0%
tramp3d-v4/tramp3d-v4                         356840.00      356732.00      -0.0%
lencod/lencod                                 427560.00      427396.00      -0.0%
7zip/7zip-benchmark                           568400.00      568172.00      -0.0%
Bullet/bullet                                 455660.00      455428.00      -0.1%
mafft/pairlocalalign                          248236.00      248040.00      -0.1%
sqlite3/sqlite3                               284404.00      284176.00      -0.1%
ClamAV/clamscan                               381052.00      380604.00      -0.1%
SPASS/SPASS                                   411932.00      411296.00      -0.2%
kimwitu++/kc                                  439696.00      438992.00      -0.2%
                           Geomean difference                               -0.1%

23 months ago[mlir][sparse] fix incorrect coordinates ordering computed by the foreach operation.
Peiming Liu [Sat, 12 Nov 2022 01:00:44 +0000 (01:00 +0000)]
[mlir][sparse] fix incorrect coordinates ordering computed by the foreach operation.

Reviewed By: aartbik

Differential Revision: https://reviews.llvm.org/D137877

23 months agoRemoved tabs.
Raman Tenneti [Sat, 12 Nov 2022 02:11:06 +0000 (18:11 -0800)]
Removed tabs.

23 months ago[libc] Implement gettimeofday
Raman Tenneti [Sat, 12 Nov 2022 02:00:01 +0000 (18:00 -0800)]
[libc] Implement gettimeofday

Implement gettimeofday per
.../onlinepubs/9699919799/functions/gettimeofday.html.
This call clock_gettime to implement gettimeofday function.

Tested:
Limited unit test: This makes a call and checks that no error was
returned. Used nanosleep for 100 microseconds and verfified it
returns a value that elapses more than 100 microseconds and less
than 300 microseconds.

Co-authored-by: Jeff Bailey <jeffbailey@google.com>
Differential Revision: https://reviews.llvm.org/D137881

23 months ago[lldb] Rewrite to assertEqual/assertNotEqual (NFC)
Dave Lee [Sat, 12 Nov 2022 00:19:03 +0000 (16:19 -0800)]
[lldb] Rewrite to assertEqual/assertNotEqual (NFC)

Using the more specific assert* methods results in more useful error message.

23 months agoAMDGPU: Fold llvm.amdgcn.sqrt(undef)
Matt Arsenault [Fri, 26 Jun 2020 03:57:30 +0000 (23:57 -0400)]
AMDGPU: Fold llvm.amdgcn.sqrt(undef)

23 months ago[libc++] Add a libc++ CI pipeline specific to Clang changes
Louis Dionne [Thu, 10 Nov 2022 03:00:32 +0000 (17:00 -1000)]
[libc++] Add a libc++ CI pipeline specific to Clang changes

This will ensure that Clang changes get tested against libc++.

Differential Revision: https://reviews.llvm.org/D137759

23 months agoAnalysis: Reorder code in isDereferenceableAndAlignedPointer
Matt Arsenault [Mon, 19 Sep 2022 18:04:12 +0000 (14:04 -0400)]
Analysis: Reorder code in isDereferenceableAndAlignedPointer

GEPs should be the most common and basic case, so try that first.

23 months agoWebAssembly: Remove MachineFunction reference from MFI
Matt Arsenault [Wed, 2 Nov 2022 21:46:00 +0000 (14:46 -0700)]
WebAssembly: Remove MachineFunction reference from MFI

The MachineFunctionInfo here is a bit awkward because
WasmEHInfo is in the MachineFunction but handled from
the target code. Either everything should move into WebAssembly
or into the MachineFunction for MIR serialization.

23 months agoclang: Fix unnecessary truncation of resource limit values
Matt Arsenault [Sat, 12 Nov 2022 00:13:40 +0000 (16:13 -0800)]
clang: Fix unnecessary truncation of resource limit values

23 months agoDisable OMPD tests. Causing CMake issue.
Ye Luo [Sat, 12 Nov 2022 00:14:33 +0000 (18:14 -0600)]
Disable OMPD tests. Causing CMake issue.

23 months ago[lldb] Fix SBFileSpec.fullpath for Windows
Dave Lee [Fri, 11 Nov 2022 22:32:04 +0000 (14:32 -0800)]
[lldb] Fix SBFileSpec.fullpath for Windows

Fix `fullpath` to not assume a `/` path separator. This was discovered when
D133130 failed on Windows. Use `os.path.join()` to fix the issue.

Reviewed By: mib

Differential Revision: https://reviews.llvm.org/D133366

23 months ago[libc] move fork into threads folder
Michael Jones [Fri, 11 Nov 2022 21:50:01 +0000 (13:50 -0800)]
[libc] move fork into threads folder

Fork, as a thread function, should go in the threads folder.
Additionally, it depends on the thread mutex, and it was causing build
issues for targets where we don't support threads.

Reviewed By: lntue

Differential Revision: https://reviews.llvm.org/D137867

23 months ago[ORC-RT][MachO] Unlock JDStatesMutex during push-initializers to avoid deadlock.
Lang Hames [Fri, 11 Nov 2022 21:34:29 +0000 (13:34 -0800)]
[ORC-RT][MachO] Unlock JDStatesMutex during push-initializers to avoid deadlock.

During __orc_rt_macho_jit_dlopen the ORC runtime will make a request to the JIT
to push any new initializers. Since this call may add new JD-state to the
runtime (and is expected to in general) we need to unlock the JDStatesMutex
during this operation (and similarly when running initializers and atexits, as
these may call trigger push-initializers recursively).

No testcase yet: I haven't been able to reproduce the deadlock when running
llvm-jitlink in in-process mode, and we don't support out-of-process mode in
regression tests yet.

23 months ago[ORC] Capture JD by value in MachOPlatform::pushInitializersLoop.
Lang Hames [Fri, 11 Nov 2022 05:32:31 +0000 (21:32 -0800)]
[ORC] Capture JD by value in MachOPlatform::pushInitializersLoop.

The lambda may run after pushInitializersLoop returns.

23 months agodocs: add instructions for stand-alone builds of lld
Tom Stellard [Fri, 11 Nov 2022 23:17:25 +0000 (15:17 -0800)]
docs: add instructions for stand-alone builds of lld

Reviewed By: kwk, MaskRay

Differential Revision: https://reviews.llvm.org/D124405

23 months ago[mlir][Linalg] Avoid using `tensor.cast` by default while folding `fill` with `pad`.
Mahesh Ravishankar [Fri, 11 Nov 2022 23:09:08 +0000 (23:09 +0000)]
[mlir][Linalg] Avoid using `tensor.cast` by default while folding `fill` with `pad`.

This is unnecessary if the generated operation type already matches
the type of the replaced value.  Also use `OpFoldResult` to reduce the
number of cases the casts are needed.

Reviewed By: springerm, hanchung, antiagainst

Differential Revision: https://reviews.llvm.org/D137479

23 months agoRevert "[Hexagon] Use default attributes for intrinsics"
Alina Sbirlea [Fri, 11 Nov 2022 22:57:10 +0000 (14:57 -0800)]
Revert "[Hexagon] Use default attributes for intrinsics"

This reverts commit 8a8983b279dd5e4dceabe1fadbb8980b6adb88f9.

Uncovers existing regalloc issue in Hexagon backend - blocking for Halide
Hexagon users. Reverting to unblock, to be recommitted when underlying issue is resolved.
Reproducer available shortly.

23 months agollvm-diff: Add failing testcase for issue 58629
Matt Arsenault [Sun, 23 Oct 2022 19:29:18 +0000 (12:29 -0700)]
llvm-diff: Add failing testcase for issue 58629

23 months ago[NFC] Remove unused var Op
Jordan Rupprecht [Fri, 11 Nov 2022 22:47:57 +0000 (14:47 -0800)]
[NFC] Remove unused var Op

23 months agollvm-reduce: Minor code cleanups
Matt Arsenault [Mon, 24 Oct 2022 17:56:18 +0000 (10:56 -0700)]
llvm-reduce: Minor code cleanups

23 months agollvm-reduce: Use DenseSet
Matt Arsenault [Mon, 24 Oct 2022 23:19:40 +0000 (16:19 -0700)]
llvm-reduce: Use DenseSet

23 months ago[lldb] Allow flexible importing of in_call_stack
Dave Lee [Fri, 11 Nov 2022 20:18:37 +0000 (12:18 -0800)]
[lldb] Allow flexible importing of in_call_stack

Allow `in_call_stack` to be imported in either of the following ways:

```
command script import path/to/in_call_stack.py
command script import lldb.utils.in_call_stack
```

rdar://102249295

Differential Revision: https://reviews.llvm.org/D137860

23 months ago[AArch64] Select BFI/BFXIL to ORR with shifted operand when one operand is the left...
Mingming Liu [Wed, 9 Nov 2022 06:46:42 +0000 (22:46 -0800)]
[AArch64] Select BFI/BFXIL to ORR with shifted operand when one operand is the left or right shift of another operand

Use right shift [1] as an example
- Before, bfxil is generated (https://godbolt.org/z/EfzWMszPn)
- After, orr with right-shifted operand is generated (added test cases in `CodeGen/AArch64/bitfield-insert.ll`)

[1]
```
define i64 @test_orr_not_bfxil_i64(i64 %0) {
  %2 = and i64 %0, 1044480 ; 0xff000
  %3 = lshr i64 %2, 12
  %4 = or i64 %2, %3
  ret i64 %4
}
```

Differential Revision: https://reviews.llvm.org/D137689

23 months ago[NFC][AArch64]Precommit test cases to show ORR is better when one operand is a shift...
Mingming Liu [Wed, 9 Nov 2022 06:28:07 +0000 (22:28 -0800)]
[NFC][AArch64]Precommit test cases to show ORR is better when one operand is a shift of the other operand

In `bfi-not-orr` tests, bfi/bfxil are better since they simplifies away two instructions (extracting bits into destination directly)

In `orr-not-bfi` tests, orr is better since both orr and bfm would simplify away one instruction (the shl node), orr has higher throughput and shorter latency than bfm.

23 months ago[mlir][sparse] fix crash when calling getTuple on non-sparse tensors.
Peiming Liu [Fri, 11 Nov 2022 18:51:25 +0000 (18:51 +0000)]
[mlir][sparse] fix crash when calling getTuple on non-sparse tensors.

This enables full sparse convolution codegen in D137298

Reviewed By: bixia

Differential Revision: https://reviews.llvm.org/D137853

23 months agollvm-reduce: Report number of new chunks
Matt Arsenault [Thu, 20 Oct 2022 21:19:59 +0000 (14:19 -0700)]
llvm-reduce: Report number of new chunks

23 months ago[LoopFuse] Ensure inner loops are in loop simplified form under new PM
Mengxuan Cai [Fri, 11 Nov 2022 17:42:45 +0000 (12:42 -0500)]
[LoopFuse] Ensure inner loops are in loop simplified form under new PM

LoopInfo doesn't give all loops in a loop nest, it gives top level loops
only. While isLoopSimplifyForm() only checkes for the outter most loop of a
loop nest. As a result, inner loops that are not in simplied form can
not be simplified with the original code.

Reviewed By: aeubanks

Differential Revision: https://reviews.llvm.org/D137672

23 months ago[InstCombine] allow more folds more multi-use selects
Sanjay Patel [Fri, 11 Nov 2022 20:26:54 +0000 (15:26 -0500)]
[InstCombine] allow more folds more multi-use selects

The 'and' case showed up in a recent bug report and prevented
more follow-on transforms from happening.

We could handle more patterns (for example, the select arms
simplified, but not to constant values), but this seems
like a safe, conservative enhancement. The backend can
convert select-of-constants to math/logic in many cases
if it is profitable.

There is a lot of overlapping logic for these kinds of patterns
(see SimplifySelectsFeedingBinaryOp() and FoldOpIntoSelect()),
so there may be some opportunity to improve efficiency.

There are also optimization gaps/inconsistency because we do
not call this code for all bin-opcodes (see TODO for ashr test).

23 months ago[InstCombine] add tests for binop with select operand; NFC
Sanjay Patel [Fri, 11 Nov 2022 16:17:07 +0000 (11:17 -0500)]
[InstCombine] add tests for binop with select operand; NFC

23 months ago[mlir][arith] Add `arith.cmpi` support to WIE
Jakub Kuderski [Fri, 11 Nov 2022 20:13:48 +0000 (15:13 -0500)]
[mlir][arith] Add `arith.cmpi` support to WIE

This inludes both LIT tests over IR and runtime checks.

Reviewed By: antiagainst

Differential Revision: https://reviews.llvm.org/D137846

23 months ago[objcopy] Fix order of Mach-O LINKEDIT pieces during layout
Daniel Rodríguez Troitiño [Fri, 11 Nov 2022 19:56:46 +0000 (11:56 -0800)]
[objcopy] Fix order of Mach-O LINKEDIT pieces during layout

The exports trie and the chained fixups where in the opposite order, and
function starts happenned before them, instead of after them.

Restore the correct order and rewrite the code to make it easier to move
around in the future if needed by reusing the `Offset` variable and
keeping both the `StartOf...` and the size of each piece together.

This was found out while trying to use the system strip in a binary
already stripped by LLVM and receiving errors around chained fixups when
we enabled those in the linker.

Reviewed By: alexander-shaposhnikov

Differential Revision: https://reviews.llvm.org/D133974

23 months ago[Hexagon] Place aligned loads closer to users
Krzysztof Parzyszek [Tue, 8 Nov 2022 22:57:21 +0000 (14:57 -0800)]
[Hexagon] Place aligned loads closer to users

Vector alignment code was grouping all aligned loads together. In some
cases the groups could become quite large causing a lot of spill to be
generated. This will place the loads closer to where they are used,
reducing the register pressure.

23 months ago[lldb] Don't assume name of libc++ inline namespace in LibCxxUnorderedMap
Dave Lee [Fri, 11 Nov 2022 19:44:06 +0000 (11:44 -0800)]
[lldb] Don't assume name of libc++ inline namespace in LibCxxUnorderedMap

Follow up to D117383, fixing the assumption that libc++ always uses `__1` as
its inline namespace name.

Reviewed By: rupprecht

Differential Revision: https://reviews.llvm.org/D133259

23 months ago[LinkerWrapper] Fix -Wpessimizing-move
Fangrui Song [Fri, 11 Nov 2022 19:53:05 +0000 (11:53 -0800)]
[LinkerWrapper] Fix -Wpessimizing-move

23 months ago[LinkerWrapper] Perform device linking steps in parallel
Joseph Huber [Tue, 25 Oct 2022 17:28:28 +0000 (12:28 -0500)]
[LinkerWrapper] Perform device linking steps in parallel

This patch changes the device linking steps to be performed in parallel
when multiple offloading architectures are being used. We use the LLVM
parallelism support to accomplish this by simply doing each inidividual
device linking job in a single thread. This change required re-parsing
the input arguments as these arguments have internal state that would
not be properly shared between the threads otherwise.

By default, the parallelism uses all threads availible. But this can be
controlled with the `--wrapper-jobs=` option. This was required in a few
tests to ensure the ordering was still deterministic.

Reviewed By: tra

Differential Revision: https://reviews.llvm.org/D136701

23 months ago[lldb] Update regex to be less fragile in TestDataFormatterGenericUnordered
Dave Lee [Fri, 11 Nov 2022 19:07:37 +0000 (11:07 -0800)]
[lldb] Update regex to be less fragile in TestDataFormatterGenericUnordered

Follow up to D129386 where libc++ naming conventions were made consistent.

This changes the pattern to not rely on the internal name (`__cc` or `__cc_`),
and instead uses a pattern to check that the child has the form:

```
[0] = {
  first = ...
```

Thanks to @rupprecht for pointing out this issue: https://reviews.llvm.org/D133259#3773120

Reviewed By: rupprecht

Differential Revision: https://reviews.llvm.org/D133395

23 months ago[lldb] [cmake] Fix typo in unittest directory path
Michał Górny [Fri, 11 Nov 2022 19:38:56 +0000 (20:38 +0100)]
[lldb] [cmake] Fix typo in unittest directory path

Fix a typo in a11cd0d94ed3cabf0998a0289aead05da94c86eb that resulted
in additional "}" in unittest directory path, e.g.:

    CMake Error at cmake/modules/LLDBStandalone.cmake:104 (add_subdirectory):
      add_subdirectory given source
      "/var/tmp/portage/dev-util/lldb-16.0.0_pre20221111/work/lldb/../third-party}/utils/unittest"
      which is not an existing directory.
    Call Stack (most recent call first):
      CMakeLists.txt:29 (include)

23 months agoRevert "[Clang][AArch64][Darwin] Enable GlobalISel by default for Darwin ARM64 platfo...
Alex Brachet [Fri, 11 Nov 2022 19:40:08 +0000 (19:40 +0000)]
Revert "[Clang][AArch64][Darwin] Enable GlobalISel by default for Darwin ARM64 platforms."

This reverts commit f64802e8d3e9db299cad913ffcb734c8d35dc5f0.

23 months agoAdd a const version of SDUse::getUser [nfc]
Philip Reames [Fri, 11 Nov 2022 19:10:29 +0000 (11:10 -0800)]
Add a const version of SDUse::getUser [nfc]

23 months agoModel UB in integer division operations in the arith dialect
Sanjoy Das [Fri, 11 Nov 2022 05:31:33 +0000 (21:31 -0800)]
Model UB in integer division operations in the arith dialect

Before this commit `arith.{ceil}div{u|s}i` were marked `Pure` which is
incorrect because these operations invoke UB on certain inputs.

Fixes: https://github.com/llvm/llvm-project/issues/58700

Reviewed By: kuhar

Differential Revision: https://reviews.llvm.org/D137814

23 months ago[ObjectYAML] Basic support for chained fixups.
Daniel Rodríguez Troitiño [Fri, 11 Nov 2022 18:13:37 +0000 (10:13 -0800)]
[ObjectYAML] Basic support for chained fixups.

Add basic binary support for chained fixups. This allows basic tests
with chained fixups without trying to create a format for them until the
work on the Object library is considered finished.

Reviewed By: pete

Differential Revision: https://reviews.llvm.org/D134250

23 months agoFix typo; NFC
Aaron Ballman [Fri, 11 Nov 2022 17:22:56 +0000 (12:22 -0500)]
Fix typo; NFC

Co-authored-by: Guillot Tony <tony.guillot@protonmail.com>
23 months agoRevert "[LLDB] Devirtualize coroutine promise types for `std::coroutine_handle`"
Adrian Vogelsgesang [Fri, 11 Nov 2022 17:59:08 +0000 (09:59 -0800)]
Revert "[LLDB] Devirtualize coroutine promise types for `std::coroutine_handle`"

This reverts commit 558db7787005348e2efaabb628ec36f1c461a741 due to
buildbot failures on ARM
* https://lab.llvm.org/buildbot/#/builders/96/builds/31416
* https://lab.llvm.org/buildbot/#/builders/17/builds/30086

23 months ago[NFC][AArch64]Call encoding functions for left-shift immediate (which is no-op in...
Mingming Liu [Thu, 10 Nov 2022 20:12:21 +0000 (12:12 -0800)]
[NFC][AArch64]Call encoding functions for left-shift immediate (which is no-op in terms of value but better code style)

Call encoding functions for left-shfit immidate for consistency (and
easier tracking if the encoding ever changes in the future).

Differential Revision: https://reviews.llvm.org/D137797

23 months ago[mlir][sparse] Extend more integration to run on the codegen path.
bixia1 [Fri, 11 Nov 2022 16:46:55 +0000 (08:46 -0800)]
[mlir][sparse] Extend more integration to run on the codegen path.

Reviewed By: Peiming

Differential Revision: https://reviews.llvm.org/D137850

23 months ago[X86] Split int2double and float2double scheduler classes on Haswell/Broadwell to...
Simon Pilgrim [Fri, 11 Nov 2022 17:39:14 +0000 (17:39 +0000)]
[X86] Split int2double and float2double scheduler classes on Haswell/Broadwell to remove overrides

Haswell/Broadwell have numerous conversion instructions that use different scheduler pipes for the reg-reg and reg-mem variants (and not an additional Port23 uop for memory folding) - so declare the classes separately instead of using the HWWriteResPair/BWWriteResPair helpers

23 months ago[LLDB] Devirtualize coroutine promise types for `std::coroutine_handle`
Adrian Vogelsgesang [Wed, 24 Aug 2022 03:53:00 +0000 (20:53 -0700)]
[LLDB] Devirtualize coroutine promise types for `std::coroutine_handle`

This commit teaches the `std::coroutine_handle` pretty-printer to
devirtualize type-erased promise types. This is particularly useful to
resonstruct call stacks, either of asynchronous control flow or of
recursive invocations of `std::generator`. For the example recently
introduced by https://reviews.llvm.org/D132451, printing the `__promise`
variable now shows

```
(std::__coroutine_traits_sfinae<task, void>::promise_type) __promise = {
  continuation = coro frame = 0x555555562430 {
    resume = 0x0000555555556310 (a.out`task detail::chain_fn<1>() at llvm-nested-example.cpp:66)
    destroy = 0x0000555555556700 (a.out`task detail::chain_fn<1>() at llvm-nested-example.cpp:66)
    promise = {
      continuation = coro frame = 0x5555555623e0 {
        resume = 0x0000555555557070 (a.out`task detail::chain_fn<2>() at llvm-nested-example.cpp:66)
        destroy = 0x0000555555557460 (a.out`task detail::chain_fn<2>() at llvm-nested-example.cpp:66)
        promise = {
          ...
        }
      }
      result = 0
    }
  }
  result = 0
}
```

(shortened to keep the commit message readable) instead of

```
(std::__coroutine_traits_sfinae<task, void>::promise_type) __promise = {
  continuation = coro frame = 0x555555562430 {
    resume = 0x0000555555556310 (a.out`task detail::chain_fn<1>() at llvm-nested-example.cpp:66)
    destroy = 0x0000555555556700 (a.out`task detail::chain_fn<1>() at llvm-nested-example.cpp:66)
  }
  result = 0
}
```

Note how the new debug output reveals the complete asynchronous call
stack: our own function resumes `chain_fn<1>` which in turn will resume
`chain_fn<2>` and so on. Thereby this change allows users of lldb to
inspect the logical coroutine call stack without using any custom debug
scripts (although the display is still a bit clumsy. It would be nicer
to also integrate this into lldb's backtrace feature, but I don't know
how to do so)

The devirtualization currently works by introspecting the function
pointed to by the `destroy` pointer. (The `resume` pointer is not worth
much, given that for the final suspend point `resume` is set to a
nullptr. We have to use the `destroy` pointer instead.) We then look
for a `__promise` variable inside the `destroy` function. This
`__promise` variable is synthetically generated by LLVM, and looking at
its type reveals the type-erased promise_type.

This approach only works for clang-generated code, though. While gcc
also adds a `_Coro_promise` variable to the `resume` function, it does
not do so for the `destroy` function. However, we can't use the `resume`
function, as it will be reset to a nullptr at the final suspension
point. For the time being, I am happy with de-virtualization only working
for clang. A follow-up commit will further improve devirtualization and
also expose the variables spilled to the coroutine frame. As part of
this, I will also revisit gcc support.

Differential Revision: https://reviews.llvm.org/D132624

23 months agoAMDGPU: Disable some class simplifications for strictfp
Matt Arsenault [Fri, 11 Nov 2022 16:51:24 +0000 (08:51 -0800)]
AMDGPU: Disable some class simplifications for strictfp

23 months ago[mlir][sparse] Fix a bug in rewriting dense2dense convert op.
bixia1 [Fri, 11 Nov 2022 07:13:20 +0000 (23:13 -0800)]
[mlir][sparse] Fix a bug in rewriting dense2dense convert op.

Permutation wasn't handled correctly. Add a test for the rewriting.

Extend an integration test to run with enable_runtime_library=false to
also test the rewriting.

Reviewed By: Peiming

Differential Revision: https://reviews.llvm.org/D137845

23 months agoconsistency: use spaces instead of tabs
Sylvestre Ledru [Fri, 11 Nov 2022 16:36:07 +0000 (17:36 +0100)]
consistency: use spaces instead of tabs

23 months ago[NFC] Remove unused OrigLoopID vars
Jordan Rupprecht [Fri, 11 Nov 2022 15:51:40 +0000 (07:51 -0800)]
[NFC] Remove unused OrigLoopID vars

23 months ago[LV] Remove unused OrigLoopID argument from completeLoopSekelton (NFC).
Florian Hahn [Fri, 11 Nov 2022 15:39:07 +0000 (15:39 +0000)]
[LV] Remove unused OrigLoopID argument from completeLoopSekelton (NFC).

The argument is not used any longer and can be removed.

23 months agoPrecommit for redundant and after SVE load
Benjamin Maxwell [Tue, 8 Nov 2022 11:52:55 +0000 (11:52 +0000)]
Precommit for redundant and after SVE load

23 months agoThe handling of 'funsafe-math-optimizations' doesn't update the 'MathErrno'
Zahira Ammarguellat [Mon, 7 Nov 2022 18:54:42 +0000 (13:54 -0500)]
The handling of 'funsafe-math-optimizations' doesn't update the 'MathErrno'
flag. But the driver checks for 'fno-math-errno' before passing
'funsafe-math-optimizations' to the FE. In GCC, the option
'funsafe-math-optimizations' doesn't affect the 'fmath-errno' flag.
This patch aligns clang with GCC.

'-ffast-math' sets the FPContract to 'fast'. But 'funsafe-math-optimizations'
the driver doesn't consider the FPContract when handling the option.
Unfortunately there are places in the BE that interpret unsafe math
mode as allowing FMA. This patch makes -ffast-math' and
'funsafe-math-optimizations' behave similarly in regard to the setting of the
FPContract.

Differential Revision: https://reviews.llvm.org/D137578

23 months ago[X86] Replace unnecessary CVTPS2DQ folded overrides with better base class defs
Simon Pilgrim [Fri, 11 Nov 2022 14:51:05 +0000 (14:51 +0000)]
[X86] Replace unnecessary CVTPS2DQ folded overrides with better base class defs

Broadwell just needed the load latency to be tweaked for the overrides to be unnecessary - I think this was due to Issue #38536 (underestimation of most broadwell load latencies)

23 months ago[InstSimplify] add test for fsub with inf operand; NFC
Sanjay Patel [Fri, 11 Nov 2022 13:51:13 +0000 (08:51 -0500)]
[InstSimplify] add test for fsub with inf operand; NFC

Verify that constant negation works with a partial undef vector.
Also, remove a bogus TODO comment on a related test.

23 months ago[MemCpyOpt] Avoid moving lifetime marker above def (PR58903)
Nikita Popov [Fri, 11 Nov 2022 14:05:11 +0000 (15:05 +0100)]
[MemCpyOpt] Avoid moving lifetime marker above def (PR58903)

This is unlikely to happen with opaque pointers, so just bail out
of the transform, rather than trying to move bitcasts/etc as well.

Fixes https://github.com/llvm/llvm-project/issues/58903.

23 months ago[include-cleaner] NFC, move the macro location fixme to findHeaders.
Haojian Wu [Fri, 11 Nov 2022 13:54:20 +0000 (14:54 +0100)]
[include-cleaner] NFC, move the macro location fixme to findHeaders.

23 months ago[InstSimplify] fold fsub nnan with Inf operand
Sanjay Patel [Thu, 10 Nov 2022 22:35:36 +0000 (17:35 -0500)]
[InstSimplify] fold fsub nnan with Inf operand

Similar to fbc2c8f2fbbb, but if we have a non-canonical
fsub with constant operand 1, then flip the sign of the
Infinity:
https://alive2.llvm.org/ce/z/vKWfhW

If Infinity is operand 0, then the sign remains:
https://alive2.llvm.org/ce/z/73d97C

23 months ago[include-cleaner] NFC, correct a comment in
Haojian Wu [Fri, 11 Nov 2022 13:40:08 +0000 (14:40 +0100)]
[include-cleaner] NFC, correct a comment in
PragmaIncludes::RecordPragma.

23 months ago[mlir][bufferize][NFC] Consolidate transform header files
Matthias Springer [Fri, 11 Nov 2022 12:49:02 +0000 (13:49 +0100)]
[mlir][bufferize][NFC] Consolidate transform header files

Differential Revision: https://reviews.llvm.org/D137830

23 months ago[compiler-rt] Mark $t* as clobbered for Linux/LoongArch syscalls
XingLi [Fri, 11 Nov 2022 13:22:45 +0000 (21:22 +0800)]
[compiler-rt] Mark $t* as clobbered for Linux/LoongArch syscalls

Linux/LoongArch doesn't preserve temporary registers across syscalls,
so we have to explicitly mark them as clobbered to avoid trashing local variables.

Reviewed By: xry111, xen0n, tangyouling, SixWeining

Differential Revision: https://reviews.llvm.org/D137396

23 months agofor Vignesh: land changes to disable two recent ompd random fails
Ron Lieberman [Fri, 11 Nov 2022 13:09:03 +0000 (07:09 -0600)]
for Vignesh: land changes to disable two recent ompd random fails

Differential Revision: https://reviews.llvm.org/D137831

23 months ago[clang-include-cleaner] make SymbolLocation a real class, move FindHeaders
Sam McCall [Fri, 11 Nov 2022 11:41:45 +0000 (12:41 +0100)]
[clang-include-cleaner] make SymbolLocation a real class, move FindHeaders

- replace SymbolLocation std::variant with enum-exposing version similar to
  those in types.cpp. There's no appropriate implementation file, added
  LocateSymbol.cpp in anticipation of locateDecl/locateMacro.
- FindHeaders is not part of the public Analysis interface, so should not
  be implemented/tested there (just code organization)
- rename findIncludeHeaders->findHeaders to avoid confusion with Include concept

Differential Revision: https://reviews.llvm.org/D137825

23 months ago[Clang][LoongArch] Remove duplicate declaration. NFC
wanglei [Fri, 11 Nov 2022 12:36:18 +0000 (20:36 +0800)]
[Clang][LoongArch] Remove duplicate declaration. NFC

23 months ago[include-cleaner] Provide public to_string of RefType (for HTMLReport), clean up...
Sam McCall [Fri, 11 Nov 2022 12:25:22 +0000 (13:25 +0100)]
[include-cleaner] Provide public to_string of RefType (for HTMLReport), clean up includes. NFC

23 months ago[openmp] [test] Set the right calling convention for the Windows thread start function
Martin Storsjö [Thu, 10 Nov 2022 10:37:00 +0000 (10:37 +0000)]
[openmp] [test] Set the right calling convention for the Windows thread start function

This is required on i386 Windows; this fixes 99 testcases in that
build configuration.

Differential Revision: https://reviews.llvm.org/D137776

23 months ago[openmp] [test] Use omp_testsuite.h instead of directly including pthread.h
Martin Storsjö [Wed, 2 Nov 2022 13:35:50 +0000 (13:35 +0000)]
[openmp] [test] Use omp_testsuite.h instead of directly including pthread.h

OpenMP tests that use pthread functions include this header instead.
On Unix systems, this header includes pthread.h, while it provides
minimal implementations of the used pthread functions for Windows.

Differential Revision: https://reviews.llvm.org/D137746

23 months ago[openmp] [test] Fix building the affinity/format/fields_values.c testcase on Windows
Martin Storsjö [Wed, 2 Nov 2022 11:55:39 +0000 (11:55 +0000)]
[openmp] [test] Fix building the affinity/format/fields_values.c testcase on Windows

Add a missing <process.h> include for _getpid. Don't typedef the
pid_t type on mingw, as mingw headers already provide a typedef for
it.

Differential Revision: https://reviews.llvm.org/D137745

23 months ago[openmp] Fix building in debug mode with mingw
Martin Storsjö [Sun, 6 Nov 2022 22:57:07 +0000 (00:57 +0200)]
[openmp] Fix building in debug mode with mingw

Mingw doesn't provide the _malloc_dbg/_free_dbg functions.

Differential Revision: https://reviews.llvm.org/D137743

23 months ago[include-cleaner] verbatimSpelling->verbatim, clean up some silly init-lists. NFC
Sam McCall [Fri, 11 Nov 2022 11:10:01 +0000 (12:10 +0100)]
[include-cleaner] verbatimSpelling->verbatim, clean up some silly init-lists. NFC

23 months ago[mlir] Fix asan errors in gpu transform dialect
Guray Ozen [Fri, 11 Nov 2022 10:57:00 +0000 (11:57 +0100)]
[mlir] Fix asan errors in gpu transform dialect

23 months ago[mlir][bufferize] Eliminate tensor.empty ops instead of bufferization.alloc_tensor ops
Matthias Springer [Fri, 11 Nov 2022 09:32:05 +0000 (10:32 +0100)]
[mlir][bufferize] Eliminate tensor.empty ops instead of bufferization.alloc_tensor ops

tensor.empty op elimination is an optimization that brings IR in a more bufferization-friendly form. E.g.:

```
%0 = tensor.empty()
%1 = linalg.fill(%cst, %0) {inplace = [true]}
%2 = tensor.insert_slice %1 into %t[10][20][1]
```

Is rewritten to:

```
%0 = tensor.extract_slice %t[10][20][1]
%1 = linalg.fill(%cst, %0) {inplace = [true]}
%2 = tensor.insert_slice %1 into %t[10][20][1]
```

This optimization used to operate on bufferization.alloc_tensor ops. This is not correct because the documentation of bufferization.alloc_tensor says that it always bufferizes to an allocation. Instead, this optimization should operate on tensor.empty ops, which can then be lowered to bufferization.alloc_tensor ops (if they don't get eliminated).

Differential Revision: https://reviews.llvm.org/D137162

23 months ago[LoongArch] Generate PCALAU12I + JIRL instruction pair for medium codemodel
wanglei [Fri, 11 Nov 2022 10:13:52 +0000 (18:13 +0800)]
[LoongArch] Generate PCALAU12I + JIRL instruction pair for medium codemodel

In LoongArch, when `CodeModel=Medium`, it just increases the jumping
ability of function calls relative to PC, from 2^28 to 2^32.

Depends on D137393

Reviewed By: SixWeining

Differential Revision: https://reviews.llvm.org/D137394

23 months ago[AMDGPU][MC] Disable SGPRs as src operands of VOP3 VINTRP instructions
Dmitry Preobrazhensky [Fri, 11 Nov 2022 10:14:42 +0000 (13:14 +0300)]
[AMDGPU][MC] Disable SGPRs as src operands of VOP3 VINTRP instructions

Differential Revision: https://reviews.llvm.org/D137575

23 months ago[LoongArch] Moved expansion of PseudoCALL to LoongArchPreRAExpandPseudo pass
wanglei [Fri, 11 Nov 2022 01:52:26 +0000 (09:52 +0800)]
[LoongArch] Moved expansion of PseudoCALL to LoongArchPreRAExpandPseudo pass

This patch moves the expansion of the `PseudoCALL` insturction to
`LoongArchPreRAExpandPseudo` pass. This helps to expand into different
instruction sequences according to different CodeModels.

Reviewed By: SixWeining

Differential Revision: https://reviews.llvm.org/D137393

23 months ago[Hexagon] Use default attributes for intrinsics
Nikita Popov [Tue, 8 Nov 2022 10:48:03 +0000 (11:48 +0100)]
[Hexagon] Use default attributes for intrinsics

This switches Hexagon intrinsics to use the default attributes
(nosync, nofree, nocallback and willreturn). Especially willreturn
is needed to prevent optimization regressions in the future.

The only intrinsics I've excluded here are the load/store locked
intrinsics, which presumably aren't nosync.

Differential Revision: https://reviews.llvm.org/D137623

23 months agoRevert "Revert "[mlir][linalg] Replace "string" iterator_types attr with enums in...
Oleg Shyshkov [Thu, 10 Nov 2022 10:42:48 +0000 (11:42 +0100)]
Revert "Revert "[mlir][linalg] Replace "string" iterator_types attr with enums in LinalgInterface.""

With python code fixed.

This reverts commit 41280908e43d47903960c66237ab49caa5641b4d.

23 months ago[mlir] Fix forward the fix for incorrect Optional<ArrayAttr> usage.
Alexander Belyaev [Fri, 11 Nov 2022 09:52:08 +0000 (10:52 +0100)]
[mlir] Fix forward the fix for incorrect Optional<ArrayAttr> usage.

23 months ago[Test] Add test for crash in IRCE when IV is AddRec for another loop
Dmitry Makogon [Fri, 11 Nov 2022 09:45:22 +0000 (16:45 +0700)]
[Test] Add test for crash in IRCE when IV is AddRec for another loop

This adds a test for https://github.com/llvm/llvm-project/issues/58912.
IRCE crashes when it tries to check whether it is possible to safely
calculate the bounds of a loop with IV AddRec which is in another loop.

23 months ago[mlir] Fix incorrect access to the Optional<ArrayAttr> underlying values.
Alexander Belyaev [Fri, 11 Nov 2022 09:46:04 +0000 (10:46 +0100)]
[mlir] Fix incorrect access to the Optional<ArrayAttr> underlying values.

23 months ago[include-cleaner] Initial version for the "Location=>Header" step
Haojian Wu [Fri, 11 Nov 2022 09:19:28 +0000 (10:19 +0100)]
[include-cleaner] Initial version for the "Location=>Header" step

This patch implements the initial version of "Location => Header" step:

- define the interface;
- integrate into the existing workflow, and use the PragmaIncludes;

Differential Revision: https://reviews.llvm.org/D137320

23 months ago[AArch64] Add smull sinking extract-and-splat tests and regenerate neon-vmull-high...
David Green [Fri, 11 Nov 2022 08:27:44 +0000 (08:27 +0000)]
[AArch64] Add smull sinking extract-and-splat tests and regenerate neon-vmull-high-p8.ll. NFC

23 months ago[opt] Remove support for using -O[0|1|2|3|s|z] with legacy PM in opt
Bjorn Pettersson [Tue, 8 Nov 2022 20:19:25 +0000 (21:19 +0100)]
[opt] Remove support for using -O[0|1|2|3|s|z] with legacy PM in opt

When running a default pipeline (for a specific O-level) in opt it is
now expected that the new PM should be used. Only reason to use the
legacy PM is when testing a pass that is locked to the legacy PM (or
when testing single passes, for example used by the llc backend).

If a test should run both a default pipeline plus some other passes,
the solution would be to invoke opt twice (separating the default
pipeline execution from the execution of individual passes).

Starting with this patch "opt -O0" etc. will result in an error.

Differential Revision: https://reviews.llvm.org/D137663

23 months ago[mlir] Introduce device mapper attribute for `thread_dim_map` and `mapped to dims`
Guray Ozen [Thu, 10 Nov 2022 16:55:49 +0000 (17:55 +0100)]
[mlir] Introduce device mapper attribute for `thread_dim_map` and `mapped to dims`

`scf.foreach_thread` defines mapping its loops to processors via an integer array, see an example below. A lowering can use this mapping. However, expressing mapping as an integer array is very confusing, especially when there are multiple levels of parallelism. In addition, the op does not verify the integer array. This change introduces device mapping attribute to make mapping descriptive and verifiable. Then it makes GPU transform dialect use it.

```
scf.foreach_thread (%i, %j) in (%c1, %c2) {
scf.foreach_thread (%i2, %j2) in (%c1, %c2)
{...} { thread_dim_mapping = [0, 1]}
} { thread_dim_mapping = [0, 1]}
```

It first introduces a `DeviceMappingInterface` which is an attribute interface. `scf.foreach_thread` defines its mapping via this interface. A lowering must define its attributes and implement this interface as well. This way gives us a clear validation.

The change also introduces two new attributes (`#gpu.thread<x/y/z>` and `#gpu.block<x,y,z>` ). After this change, the above code prints as below, as seen here, this way clarifies the loop mappings. The change also implements consuming of these two new attribute by the transform dialect. Transform dialect binds the outermost loops to the thread blocks and innermost loops to threads.

```
scf.foreach_thread (%i, %j) in (%c1, %c2) {
scf.foreach_thread (%i2, %j2) in (%c1, %c2)
{...} { thread_dim_mapping = [#gpu.thread<x>, #gpu.thread<y>]}
} { thread_dim_mapping = [#gpu.block<x>, #gpu.block<y>]}
```

Reviewed By: ftynse, nicolasvasilache

Differential Revision: https://reviews.llvm.org/D137413