platform/upstream/llvm.git
2 years ago[mlir] Rename mlir::SmallVector -> llvm::SmallVector
Mogball [Tue, 24 May 2022 15:03:12 +0000 (15:03 +0000)]
[mlir] Rename mlir::SmallVector -> llvm::SmallVector

2 years ago[mlir] Breakdown diagnostic string literals
Logan Chien [Wed, 18 May 2022 20:59:08 +0000 (13:59 -0700)]
[mlir] Breakdown diagnostic string literals

This commit breaks down diagnostic string literals so that the attribute
name and enumurator names can be shared with the stringify utility
function and the "expected ", " to be one of ", and ", " can be shared
between different enum-related diagnostic.

Differential Revision: https://reviews.llvm.org/D125938

2 years ago[flang][runtime] Clean up asynchronous I/O APIs
Peter Klausler [Thu, 12 May 2022 00:08:21 +0000 (17:08 -0700)]
[flang][runtime] Clean up asynchronous I/O APIs

Now that the requirements and implementation of asynchronous I/O are
better understood, adjust their I/O runtime APIs.  In particular:
1) Remove the BeginAsynchronousOutput/Input APIs; they're not needed,
   since any data transfer statement might have ASYNCHRONOUS= and
   (if ASYNCHRONOUS='YES') ID= control list specifiers that need to
   at least be checked.
2) Add implementations for BeginWait(All) to check for the error
   case of a bad unit number and nonzero ID=.
3) Rearrange and comment SetAsynchronous so that it's clear that
   it can be called for READ/WRITE as well as for OPEN.

The implementation remains completely synchronous, but should be conforming.
Where opportunities make sense for true asynchronous implementations of
some big block transfers without SIZE= in the future, we'll need to add
a GetAsynchronousId API to capture ID= on a READ or WRITE; add sourceFile
and sourceLine arguments to BeginWait(All) for good error reporting;
track pending operations in unit.h; and add code to force synchronization
to non-asynchronous I/O operations.

Lowering should call SetAsynchronous when ASYNCHRONOUS= appears as
a control list specifier.  It should also set ID=x variables to 0
until such time as we support asynchronous operations, if ever.
This patch only removes the removed APIs from lowering.

Differential Revision: https://reviews.llvm.org/D126143

2 years agoFix behavior of is_fp_class on empty class set
Serge Pavlov [Tue, 24 May 2022 11:30:13 +0000 (18:30 +0700)]
Fix behavior of is_fp_class on empty class set

The second argument to is_fp_class specifies the set of floating-point
class to test against. It can be zero, in this case the intrinsic is
expected to return zero value.

Differential Revision: https://reviews.llvm.org/D112025

2 years ago[DAG] Unroll vectorized FPOW instructions before widening that will scalarize to...
Simon Pilgrim [Tue, 24 May 2022 14:44:44 +0000 (15:44 +0100)]
[DAG] Unroll vectorized FPOW instructions before widening that will scalarize to libcalls anyway

Followup to D125988 - FPOW is similar to FREM and will most likely scalarize to libcalls, so unroll before widening to prevent use making additional libcalls with UNDEF args.

2 years ago[libcxx] Add sort.bench.cpp to libcxx/benchmarks/CMakeLists.txt
Hans Wennborg [Tue, 24 May 2022 13:47:41 +0000 (15:47 +0200)]
[libcxx] Add sort.bench.cpp to libcxx/benchmarks/CMakeLists.txt

It was forgotten in D124740.

Differential revision: https://reviews.llvm.org/D126297

2 years ago[TypePromotion] Avoid unnecessary trunc zext pairs
Sam Parker [Tue, 24 May 2022 14:33:21 +0000 (15:33 +0100)]
[TypePromotion] Avoid unnecessary trunc zext pairs

Any zext 'sink' should already have an operand that is in the legal
value, so avoid using a trunc and just use the trunc operand instead.

Differential Revision: https://reviews.llvm.org/D118905

2 years ago[mlir][vector] Add new lowering mode to vector.contractionOp
Thomas Raoux [Tue, 24 May 2022 14:16:00 +0000 (14:16 +0000)]
[mlir][vector] Add new lowering mode to vector.contractionOp

Add lowering for cases where the reduction dimension is fully unrolled.
It is common to unroll the reduction dimension, therefore we would want
to lower the contractions to an elementwise vector op in this case.

Differential Revision: https://reviews.llvm.org/D126120

2 years ago[CostModel][X86] getScalarizationOverhead - improve extraction costs for > 128-bit...
Simon Pilgrim [Tue, 24 May 2022 14:17:59 +0000 (15:17 +0100)]
[CostModel][X86] getScalarizationOverhead - improve extraction costs for > 128-bit vectors

We were using the default getScalarizationOverhead expansion for extraction costs, which adds up all the individual element extraction costs.

This is fine for 128-bit vectors, but for 256/512-bit vectors each element extraction also has to account for extracting the upper 128-bit subvector extraction before it can handle the element. For scalarization costs we only need to extract each demanded subvector once.

Differential Revision: https://reviews.llvm.org/D125527

2 years ago[AMDGPU][MC][GFX11] Support base+soffset+offset SMEM loads.
Ivan Kosarev [Tue, 24 May 2022 14:12:45 +0000 (15:12 +0100)]
[AMDGPU][MC][GFX11] Support base+soffset+offset SMEM loads.

Reviewed By: dp

Differential Revision: https://reviews.llvm.org/D126207

2 years ago[InstCombine] Strip bitcasts in GEP diff fold
Nikita Popov [Tue, 24 May 2022 14:08:21 +0000 (16:08 +0200)]
[InstCombine] Strip bitcasts in GEP diff fold

Bitcasts were stripped in one case, but not the other. Of course,
this no longer really matters with opaque pointers, but as I went
through the trouble of tracking this down, we may as well remove
one typed vs opaque pointer optimization discrepancy.

2 years ago[InstCombine] Add test for GEP difference with bitcasts (NFC)
Nikita Popov [Tue, 24 May 2022 14:09:26 +0000 (16:09 +0200)]
[InstCombine] Add test for GEP difference with bitcasts (NFC)

2 years ago[libc++] Rename the generic-singlethreaded CI job to generic-no-threads for consistency
Louis Dionne [Tue, 24 May 2022 13:58:04 +0000 (09:58 -0400)]
[libc++] Rename the generic-singlethreaded CI job to generic-no-threads for consistency

2 years ago[InstCombine] Use IRBuilder in freeze pushing transform (PR55619)
Nikita Popov [Tue, 24 May 2022 13:48:28 +0000 (15:48 +0200)]
[InstCombine] Use IRBuilder in freeze pushing transform (PR55619)

Use IRBuilder so that the newly created freeze instructions
automatically gets inserted back into the IC worklist.

The changed worklist processing order leads to some cosmetic
differences in tests.

Fixes https://github.com/llvm/llvm-project/issues/55619.

2 years ago[InstCombine] Add tests for freeze of load with metadata (NFC)
Nikita Popov [Tue, 24 May 2022 13:17:19 +0000 (15:17 +0200)]
[InstCombine] Add tests for freeze of load with metadata (NFC)

2 years ago[pseudo] (trivial) bracket-matching
Sam McCall [Wed, 18 May 2022 17:24:07 +0000 (19:24 +0200)]
[pseudo] (trivial) bracket-matching

Error-tolerant bracket matching enables our error-tolerant parsing strategies.
The implementation here is *not* yet error tolerant: this patch sets up the APIs
and plumbing, and describes the planned approach.

Differential Revision: https://reviews.llvm.org/D125911

2 years ago[OpenMP] Add `-Xoffload-linker` to forward input to the device linker
Joseph Huber [Mon, 23 May 2022 21:51:30 +0000 (17:51 -0400)]
[OpenMP] Add `-Xoffload-linker` to forward input to the device linker

We use the clang-linker-wrapper to perform device linking of embedded
offloading object files. This is done by generating those jobs inside of
the linker-wrapper itself. This patch adds an argument in Clang and the
linker-wrapper that allows users to forward input to the device linking
phase. This can either be done for every device linker, or for a
specific target triple. We use the `-Xoffload-linker <arg>` and the
`-Xoffload-linker-<triple> <arg>` syntax to accomplish this.

Reviewed By: markdewing, tra

Differential Revision: https://reviews.llvm.org/D126226

2 years ago[SLP][NFC]Make isFirstInsertElement a weak strict ordering comparator.
Alexey Bataev [Tue, 24 May 2022 13:01:17 +0000 (06:01 -0700)]
[SLP][NFC]Make isFirstInsertElement a weak strict ordering comparator.

To be used correctly in a sort-like function, isFirstInsertElement
function must follow weak strict ordering rule, i.e.
isFirstInsertElement(IE1, IE1) should return false.

2 years ago[x86][DAG] Unroll vectorized FREMs that will become libcalls
Nabeel Omer [Tue, 24 May 2022 12:30:47 +0000 (13:30 +0100)]
[x86][DAG] Unroll vectorized FREMs that will become libcalls

Currently, two element vectors produced as the result of a binary op are
widened to four element vectors on x86 by
DAGTypeLegalizer::WidenVecRes_BinaryCanTrap. If the op still isn't legal
after widening it is unrolled into 4 scalar ops in SelectionDAG before
being converted into a libcall. This way we end up with 4 libcalls (two of
them on known undef elements) instead of the original two libcalls.

This patch modifies DAGTypeLegalizer::WidenVectorResult to ensure that if
it is known that a binary op will be tunred into a libcall, it is unrolled
instead of being widened. This prevents the creation of the extra scalar
instructions on known undef elements and (eventually) libacalls with known
undef parameters which would otherwise be created when the op gets expanded
post widening.

Differential Revision: https://reviews.llvm.org/D125988

2 years agoRevert "[TableGen] Remove code beads"
Sylvestre Ledru [Tue, 24 May 2022 12:17:49 +0000 (14:17 +0200)]
Revert "[TableGen] Remove code beads"

It is breaking the build with:

/build/llvm-toolchain-snapshot-15~++20220524114008+96323c9f4c10/llvm/lib/Target/M68k/MCTargetDesc/M68kMCCodeEmitter.cpp:478:10: fatal error: 'M68kGenMCCodeBeads.inc' file not found
         ^~~~~~~~~~~~~~~~~~~~~~~~
1 error generated.

Remove the #include causes:
error: undefined reference to 'llvm::M68k::getMCInstrBeads(unsigned int)'

This reverts commit f50be3d21808ae113c40a68a9ac6581f203d92d2.

2 years ago[Clang][CodeGen] Fix the cmse-clear-return.c test.
Ivan Kosarev [Tue, 24 May 2022 11:32:48 +0000 (12:32 +0100)]
[Clang][CodeGen] Fix the cmse-clear-return.c test.

Caught with D125604.

Reviewed By: nikic

Differential Revision: https://reviews.llvm.org/D126191

2 years agoMinutes for security group sync-ups have moved to Discourse.
Kristof Beyls [Tue, 24 May 2022 11:46:08 +0000 (13:46 +0200)]
Minutes for security group sync-ups have moved to Discourse.

2 years ago[X86] scalar_widen_div.ll - remove non-generated CHECKs
Simon Pilgrim [Tue, 24 May 2022 10:55:36 +0000 (11:55 +0100)]
[X86] scalar_widen_div.ll - remove non-generated CHECKs

These were left over from when we converted to the update_llc_test_checks script and were just matching some asm/cfi directives - we have CHECK-LABEL to do this properly now.

2 years ago[OpenCL] Make -cl-ext a driver option.
Anastasia Stulova [Tue, 24 May 2022 10:26:43 +0000 (11:26 +0100)]
[OpenCL] Make -cl-ext a driver option.

For generic targets such as SPIR-V clang sets all OpenCL
extensions/features as supported by default. However
concrete targets are unlikely to support all extensions
features, which creates a problem when such generic SPIR-V
binary is compiled for a specific target later on.

To allow compile time diagnostics for unsupported features
this flag is now being exposed in the clang driver.

Differential Revision: https://reviews.llvm.org/D125243

2 years agoFix comment in test. NFC
Max Kazantsev [Tue, 24 May 2022 10:22:16 +0000 (17:22 +0700)]
Fix comment in test. NFC

2 years ago[Test] Add LICM test for PR55672 showing problem with freeze instruction
Max Kazantsev [Tue, 24 May 2022 10:17:05 +0000 (17:17 +0700)]
[Test] Add LICM test for PR55672 showing problem with freeze instruction

2 years ago[X86][AMX] Reduce the compiling time for non-amx code.
Luo, Yuanke [Tue, 24 May 2022 08:46:27 +0000 (16:46 +0800)]
[X86][AMX] Reduce the compiling time for non-amx code.

Differential Revision: https://reviews.llvm.org/D126280

2 years ago[X86] Add test showing failure to expand <2 x float> fpow without widening to <4...
Simon Pilgrim [Tue, 24 May 2022 09:45:12 +0000 (10:45 +0100)]
[X86] Add test showing failure to expand <2 x float> fpow without widening to <4 x float>

Similar to D125988 (and I have a pending follow up patch to handle fpow).

2 years ago[TableGen] Remove code beads
Sheng [Tue, 24 May 2022 02:53:48 +0000 (10:53 +0800)]
[TableGen] Remove code beads

Code beads is useless since the only user, M68k, has moved on to
a new encoding/decoding infrastructure.

2 years ago[RISCV][NFC] Change interface of RVVIntrinsic::getSuffixStr
Kito Cheng [Fri, 20 May 2022 08:24:37 +0000 (16:24 +0800)]
[RISCV][NFC] Change interface of RVVIntrinsic::getSuffixStr

This NFC patch is splited from D111617.

Using llvm::ArrayRef rather than llvm::SmallVector, ArrayRef is more generic
interface that could accept both llvm::ArrayRef and llvm::SmallVector.

Reviewed By: reames

Differential Revision: https://reviews.llvm.org/D125893

2 years ago[InstCombine] Support logical and in masked icmp fold
Nikita Popov [Tue, 24 May 2022 08:45:29 +0000 (10:45 +0200)]
[InstCombine] Support logical and in masked icmp fold

Most of the folds implemented in this function work fine with
logical operations. We only need to be careful for the cases that
work on non-constant masks, where the RHS operand shouldn't be
poison.

This is a conservative implementation that bails out of illegal
transforms, but we could also change these to insert freeze instead.

2 years ago[InstCombine] Add tests for masked icmps with bitwise+logical and (NFC)
Nikita Popov [Tue, 24 May 2022 09:10:36 +0000 (11:10 +0200)]
[InstCombine] Add tests for masked icmps with bitwise+logical and (NFC)

2 years ago[InstCombine] Use m_APInt() in asymmetric masked icmp fold
Nikita Popov [Tue, 24 May 2022 08:56:16 +0000 (10:56 +0200)]
[InstCombine] Use m_APInt() in asymmetric masked icmp fold

This is mostly intended as code cleanup, but it does also add
support for splat vectors to this fold.

2 years ago[InstCombine] Add splat vector test for asymmetric masked icmp fold
Nikita Popov [Tue, 24 May 2022 08:49:08 +0000 (10:49 +0200)]
[InstCombine] Add splat vector test for asymmetric masked icmp fold
(NFC)

2 years ago[gn build] Port 1d1a191edcfa
LLVM GN Syncbot [Tue, 24 May 2022 08:37:12 +0000 (08:37 +0000)]
[gn build] Port 1d1a191edcfa

2 years ago[libc++] Implement ranges::reverse
Nikolas Klauser [Tue, 24 May 2022 08:32:50 +0000 (10:32 +0200)]
[libc++] Implement ranges::reverse

Reviewed By: var-const, #libc

Spies: libcxx-commits, mgorny

Differential Revision: https://reviews.llvm.org/D125752

2 years ago[libc++] Replace modulus operations in std::seed_seq::generate with conditional checks.
Laramie Leavitt [Tue, 24 May 2022 08:27:37 +0000 (10:27 +0200)]
[libc++] Replace modulus operations in std::seed_seq::generate with conditional checks.

Abseil benchmarks suggest that the conditional checks result in faster code (4-5x)
as they are compiled into conditional move instructions (cmov on x86).

Reviewed By: #libc, philnik, Mordante

Spies: pengfei, Mordante, philnik, libcxx-commits

Differential Revision: https://reviews.llvm.org/D125329

2 years ago[LegalizeTypes][NFC] Fix node name in assertion message
Fraser Cormack [Tue, 24 May 2022 08:16:18 +0000 (09:16 +0100)]
[LegalizeTypes][NFC] Fix node name in assertion message

This was probably copy/pasted from the MSCATTER widening.

2 years ago[libc++] type_traits: use __is_core_convertible in __invokable_r.
Aaron Jacobs [Tue, 24 May 2022 08:21:05 +0000 (10:21 +0200)]
[libc++] type_traits: use __is_core_convertible in __invokable_r.

This fixes incorrect handling of non-moveable types, adding tests for this case.
See [issue 55346](https://github.com/llvm/llvm-project/issues/55346).

The current implementation is based on is_convertible, which is
[defined](https://timsong-cpp.github.io/cppwp/n4659/meta.rel#5) in terms of
validity of the following function:

```
To test() {
  return declval<From>();
}
```

But this doesn't work if To and From are both some non-moveable type, which the
[definition](https://timsong-cpp.github.io/cppwp/n4659/conv#3) of implicit
conversions says should work due to guaranteed copy elision:

```
To to = E;  // E has type From
```

It is this latter definition that is used in the
[definition](https://timsong-cpp.github.io/cppwp/n4659/function.objects#func.require-2)
of INVOKE<R>. Make __invokable_r use __is_core_convertible, which
captures the ability to use guaranteed copy elision, making the
definition correct for non-moveable types.

Fixes llvm/llvm-project#55346.

Reviewed By: #libc, philnik, EricWF

Spies: EricWF, jloser, ldionne, philnik, libcxx-commits

Differential Revision: https://reviews.llvm.org/D125300

2 years ago[InstCombine] Handle logical and/or in recursive and/or of icmps fold
Nikita Popov [Tue, 24 May 2022 08:07:00 +0000 (10:07 +0200)]
[InstCombine] Handle logical and/or in recursive and/or of icmps fold

The and/or of icmps fold is also applied in reassociated form.
However, this currently only happens for bitwise and of bitwise
and, but not for bitwise and of logical and (or other combinations,
but this is the one being addressed here).

We can do this for bitwise+logical combinations as well, but need
to be a bit careful about which of the resulting ands are logical:
https://alive2.llvm.org/ce/z/WYSjGh
https://alive2.llvm.org/ce/z/guxYnz
https://alive2.llvm.org/ce/z/S5SYxY
https://alive2.llvm.org/ce/z/2rAWeW

2 years ago[InstCombine] Use different icmp pattern in test (NFC)
Nikita Popov [Tue, 24 May 2022 08:04:24 +0000 (10:04 +0200)]
[InstCombine] Use different icmp pattern in test (NFC)

Use an and/or of icmp pattern that produces different code
depending on whether it is part of a logical or bitwise and/or.

2 years agollvm-reduce: improve basic-blocks removal pass
Markus Lavin [Tue, 24 May 2022 07:42:07 +0000 (09:42 +0200)]
llvm-reduce: improve basic-blocks removal pass

When the single branch target of a block has been removed try updating
it to target a  block that is kept (by scanning forward in the sequence)
instead of replacing the branch with a return instruction. Doing so
reduces the risk of breaking loop structures meaning that when the loop
is 'interesting' these reductions should have more blocks eliminated.

Differential Revision: https://reviews.llvm.org/D125766

2 years ago[LoopUnroll] Freeze tripcount rather than condition
Nikita Popov [Wed, 18 May 2022 15:09:36 +0000 (17:09 +0200)]
[LoopUnroll] Freeze tripcount rather than condition

This is a followup to D125754. We introduce two branches, one
before the unrolled loop and one before the epilogue (and similar
for the prologue case). The previous patch only froze the
condition on the first branch.

Rather than independently freezing the second condition, this patch
instead freezes TripCount and bases BECount on it. These are the
two quantities involved in the conditions, and this ensures that
both work on a consistent, non-poisonous trip count.

Differential Revision: https://reviews.llvm.org/D125896

2 years ago[LegalizeTypes][VP] Fix OpNo in WidenVecOp_VP_SCATTER
Lian Wang [Tue, 24 May 2022 07:12:31 +0000 (07:12 +0000)]
[LegalizeTypes][VP] Fix OpNo in WidenVecOp_VP_SCATTER

Reviewed By: craig.topper

Differential Revision: https://reviews.llvm.org/D126276

2 years ago[RISCV] Ensure the entire stack is aligned to the RVV stack alignment
Fraser Cormack [Thu, 19 May 2022 13:47:40 +0000 (14:47 +0100)]
[RISCV] Ensure the entire stack is aligned to the RVV stack alignment

This patch fixes another bug in the RVV frame lowering. While some frame
objects with non-default stack IDs (such scalable-vector alloca
instructions) are considered in the target-independent max alignment
calculations, others (for example, during calling-convention lowering)
are not. This means we'd occasionally align the base of the stack to
only 16 bytes, with no way to ensure that the RVV section contained
within that is aligned to anything higher.

Reviewed By: StephenFan

Differential Revision: https://reviews.llvm.org/D125973

2 years ago[RISCV] Fix RVV stack frame alignment bugs
Fraser Cormack [Mon, 16 May 2022 09:57:32 +0000 (10:57 +0100)]
[RISCV] Fix RVV stack frame alignment bugs

This patch addresses several alignment issues in the stack frame when
RVV objects are taken into account.

One bug is that the RVV stack was never guaranteed to keep the alignment
of the stack *as a whole*. We must maintain a 16-byte aligned stack at
all times, especially when calling other functions. With the standard V
extension, this is conveniently happening since VLEN is at least 128 and
always 16-byte aligned. However, we support Zvl64b which does not
guarantee this. To fix this, the RVV stack size is rounded up to be
aligned to 16 bytes. This in practice generally makes us allocate a
stack sized at least 2*VLEN in size, and a multiple of 2.

    |------------------------------| -- <-- FP
    | 8-byte callee-save           | |      |
    |------------------------------| |      |
    | one VLENB-sized RVV object   | |      |
    |------------------------------| |      |
    | 8-byte local variable        | |      |
    |------------------------------| -- <-- SP (must be aligned to 16)

In the example above, with Zvl64b we are decrementing SP by 12 bytes
which does not leave SP correctly aligned. We therefore introduce an
extra VLENB-sized amount used for alignment. This would therefore ensure
the total stack size was 16 bytes (48 for Zvl128b, 80 for Zvl256b, etc):

    |------------------------------| -- <-- FP
    | 8-byte callee-save           | |      |
    |------------------------------| |      |
    | one VLENB-sized padding obj  | |      |
    | one VLENB-sized RVV object   | |      |
    |------------------------------| |      |
    | 8-byte local variable        | |      |
    |------------------------------| -- <-- SP

A new RVV invariant has been introduced in this patch, which is that the
base of the RVV stack itself is now always aligned to 16 bytes, not 8 as
before. This keeps us more in line with the scalar stack and should be
easier to reason about. The calculation of the RVV padding has thus
changed to be the amount required to align the scalar local variable
section to the RVV section's alignment. This amount is further rounded
up when setting up the initial stack to keep everything aligned:

    |------------------------------| -- <-- FP
    | 8-byte callee-save           |
    |------------------------------|
    |                              |
    | RVV objects                  |
    | (aligned to at least 16)     |
    |                              |
    |------------------------------|
    | RVV padding of 8 bytes       |
    |------------------------------|
    | 8-byte local variable        |
    |------------------------------| -- <-- SP

In the example above, it's clear that we need 8 bytes of padding to keep
the RVV section aligned to 16 when using SP. But to keep SP *itself*
aligned to 16 we can't decrement the initial stack pointer by 24 - we
have to round up to 32.

With the RVV section correctly aligned, the second bug fixed by
this patch is that RVV objects themselves are now correctly aligned. We
were previously only guaranteeing an alignment of 8 bytes, even if they
required a higher alignment. This is relatively simple and in practice
we see more rounding up of VLEN amounts to account for alignment in
between objects:

    |------------------------------|
    | RVV object (aligned to 16)   |
    |------------------------------|
    | no padding necessary         |
    |------------------------------|
    | 2*VLENB RVV object (align 16)|
    |------------------------------|
    | VLENB alignment padding      |
    |------------------------------|
    | RVV object (align 32)        |
    |------------------------------|
    | 3*VLENB alignment padding    |
    |------------------------------|
    | VLENB RVV object (align 32)  |
    |------------------------------| -- <-- base of RVV section

Note that a lot of the regressions in codegen owing to the new alignment
rules are correct but actually only strictly necessary for Zvl64b (and
Zvl32b but that's not really supported). I plan a follow-up patch to
take the known VLEN into account when padding for alignment.

Reviewed By: StephenFan

Differential Revision: https://reviews.llvm.org/D125787

2 years ago[gn build] Port 496156ac57da
LLVM GN Syncbot [Tue, 24 May 2022 05:44:48 +0000 (05:44 +0000)]
[gn build] Port 496156ac57da

2 years ago[X86][AMX] Multiple configure for AMX register.
Luo, Yuanke [Tue, 3 May 2022 10:57:25 +0000 (18:57 +0800)]
[X86][AMX] Multiple configure for AMX register.

The previous solution depends on variable name to record the shape
information. However it is not reliable, because in release build
compiler would not set the variable name. It can be accomplished with an
additional option `fno-discard-value-names`, but it is not acceptable
for users.
This patch is to preconfigure the tile register with machine
instruction. It follow the same way what sigle configure does. In the
future we can fall back to multiple configure when single configure
fails due to the shape dependency issue.
The algorithm to configure the tile register is simple in the patch. We
may improve it in the future. It configure tile register based on basic
block. Compiler would spill the tile register if it live out the basic
block. After the configure there should be no spill across tile
confgiure in the register alloction. Just like fast register allocation
the algorithm walk the instruction in reverse order. When the shape
dependency doesn't meet, it insert ldtilecfg after the last instruction
that define the shape.
In post configuration compiler also walk the basic block to collect the
physical tile register number and generate instruction to fill the stack
slot for the correponding shape information.
TODO: There is some following work in D125602. The risk is modifying the
fast RA may cause regression as fast RA is usded for different targets.
We may create an independent RA for tile register.

Differential Revision: https://reviews.llvm.org/D125075

2 years ago[MachineSink] replace MachineLoop with MachineCycle
Chen Zheng [Tue, 19 Apr 2022 07:40:17 +0000 (03:40 -0400)]
[MachineSink] replace MachineLoop with MachineCycle

MachineCycle can handle irreducible loop. Natural loop
analysis (MachineLoop) can not return correct loop depth if
the loop is irreducible loop. And MachineSink is sensitive
to the loop depth, see MachineSinking::isProfitableToSinkTo().

This patch tries to use MachineCycle so that we can handle
irreducible loop better.

Reviewed By: sameerds, MatzeB

Differential Revision: https://reviews.llvm.org/D123995

2 years ago[OpenMP][IRBuilder] `omp task` support
Shraiysh Vaishay [Tue, 24 May 2022 04:23:33 +0000 (09:53 +0530)]
[OpenMP][IRBuilder] `omp task` support

This patch adds basic support for `omp task` to the OpenMPIRBuilder.

The outlined function after code extraction is called from a wrapper function with appropriate arguments. This wrapper function is passed to the runtime calls for task allocation.

This approach is different from the Clang approach - clang directly emits the runtime call to the outlined function. The outlining utility (OutlineInfo) simply outlines the code and generates a function call to the outlined function. After the function has been generated by the outlining utility, there is no easy way to alter the function arguments without meddling with the outlining itself. Hence the wrapper function approach is taken.

Reviewed By: Meinersbur

Differential Revision: https://reviews.llvm.org/D71989

2 years ago[flang] Allow more forward references to ENTRY names
Peter Klausler [Tue, 17 May 2022 01:10:27 +0000 (18:10 -0700)]
[flang] Allow more forward references to ENTRY names

Forward references to ENTRY names to pass them as actual procedure arguments
don't work in all cases, exposing some basic ordering problems in
name resolution for these symbols.  Refactor; create all the
necessary procedure symbols, and either function result or host association
symbols (for subroutines), at the time that the subprogrma scope is
created, so that the names exist in the scope as text "before"
the ENTRY is processed in name resolution.  Some processing
remains in PostEntryStmt() so that we can check that an ENTRY with
an explicit distinct RESULT doesn't also have declarations for the
ENTRY name.

Differential Revision: https://reviews.llvm.org/D126142

2 years agoRevert "[SelectOpti][5/5] Optimize select-to-branch transformation"
Sotiris Apostolakis [Tue, 24 May 2022 04:02:00 +0000 (00:02 -0400)]
Revert "[SelectOpti][5/5] Optimize select-to-branch transformation"

This reverts commit a111fb960108df910a864500f3b98d75d37f083c.

2 years ago[SelectOpti][5/5] Optimize select-to-branch transformation
Sotiris Apostolakis [Tue, 24 May 2022 03:04:20 +0000 (23:04 -0400)]
[SelectOpti][5/5] Optimize select-to-branch transformation

This patch optimizes the transformation of selects to a branch when the heuristics deemed it profitable.
It aggressively sinks eligible instructions to the newly created true/false blocks to prevent their
execution on the common path and interleaves dependence slices to maximize ILP.

Depends on D120232

Reviewed By: davidxl

Differential Revision: https://reviews.llvm.org/D120233

2 years ago[Hexagon] Fix test on OpenBSD
Brad Smith [Tue, 24 May 2022 03:26:14 +0000 (23:26 -0400)]
[Hexagon] Fix test on OpenBSD

The test specifies a CPU arch but not a particular OS. So if run on
OpenBSD it acts as if it's an OpenBSD/hexagon system. OpenBSD uses
__guard_local instead of __stack_chk_guard so the test will fail. So
specify an OS other than OpenBSD fixes the test.

Reviewed By: MaskRay

Differential Revision: https://reviews.llvm.org/D126265

2 years agoExposes interface to free up caching data structure in DWARFDebugLine and DWARFUnit...
Hyoun Kyu Cho [Tue, 24 May 2022 03:04:40 +0000 (03:04 +0000)]
Exposes interface to free up caching data structure in DWARFDebugLine and DWARFUnit for memory management

This is minimum changes extracted from https://reviews.llvm.org/D78950. The old patch tried to add LRU eviction of caching data structure. Due to multiple layers of interfaces that users could be using, it was not clear where to put the functionality. While we work out on where to put that functionality, it'll be great to add this minimum interface change so that the user could implement their own memory management. More specifically:

* Add a clearLineTable method for DWARFDebugLine which erases the given offset from the LineTableMap.
* DWARFDebugContext adds the clearLineTableForUnit method that leverages clearLineTable to remove the object corresponding to a given compile unit, for memory management purposes. When it is referred to again, the line table object will be repopulated.

Reviewed By: dblaikie

Differential Revision: https://reviews.llvm.org/D90006

2 years agoupdated canResolveToExpr to accept both statements and expressions. Removed unnecessa...
usama hameed [Mon, 23 May 2022 23:05:15 +0000 (16:05 -0700)]
updated canResolveToExpr to accept both statements and expressions. Removed unnecessary code

2 years agobugfix in InfiniteLoopCheck to not print warnings for unevaluated loops
usama hameed [Mon, 23 May 2022 21:52:14 +0000 (14:52 -0700)]
bugfix in InfiniteLoopCheck to not print warnings for unevaluated loops

Added a separate check for unevaluated statements. Updated InfiniteLoopCheck to use new check

Differential Revision: https://reviews.llvm.org/D126246

2 years agobugfix in InfiniteLoopCheck to not print warnings for unevaluated loops
usama hameed [Thu, 19 May 2022 23:51:34 +0000 (16:51 -0700)]
bugfix in InfiniteLoopCheck to not print warnings for unevaluated loops

Differential Revision: https://reviews.llvm.org/D126034

2 years ago[NFC][Metadata] Define move constructor and move assignment operator for MDOperand.
Wolfgang Pieb [Tue, 24 May 2022 00:08:01 +0000 (17:08 -0700)]
[NFC][Metadata] Define move constructor and move assignment operator for MDOperand.

This is a preparatory patch for the MDNode resize functionality.

Reviewed By: dexonsmith

Differential Revision: https://reviews.llvm.org/D125994

2 years ago[SelectOpti][4/5] Loop Heuristics
Sotiris Apostolakis [Tue, 24 May 2022 02:05:41 +0000 (22:05 -0400)]
[SelectOpti][4/5] Loop Heuristics

This patch adds the loop-level heuristics for determining whether branches are more profitable than conditional moves.
These heuristics apply to only inner-most loops.

Depends on D120231

Reviewed By: davidxl

Differential Revision: https://reviews.llvm.org/D120232

2 years ago[SelectOpti][3/5] Base Heuristics
Sotiris Apostolakis [Mon, 23 May 2022 20:26:09 +0000 (16:26 -0400)]
[SelectOpti][3/5] Base Heuristics

This patch adds the base heuristics for determining whether branches are more profitable than conditional moves.
Base heuristics apply to all code apart from inner-most loops.

Depends on D122259

Reviewed By: davidxl

Differential Revision: https://reviews.llvm.org/D120231

2 years ago[lld-macho][nfc] Run clang-format on lld/MachO/*.{h,cpp}
Vy Nguyen [Tue, 24 May 2022 00:59:18 +0000 (07:59 +0700)]
[lld-macho][nfc] Run clang-format on lld/MachO/*.{h,cpp}

- fixed inconsistent indents and spaces
- prevent extraneous formatting changes in other patches

Differential Revision: https://reviews.llvm.org/D126262

2 years ago[flang] Ignore BIND(C) binding name conflicts of inner procedures
Peter Klausler [Wed, 11 May 2022 21:32:59 +0000 (14:32 -0700)]
[flang] Ignore BIND(C) binding name conflicts of inner procedures

The binding names of inner procedures with BIND(C) are not exposed
to the loader and should be ignored for potential conflict errors.

Differential Revision: https://reviews.llvm.org/D126141

2 years ago[flang] Allow global scope names that clash with intrinsic modules
Peter Klausler [Wed, 11 May 2022 21:13:50 +0000 (14:13 -0700)]
[flang] Allow global scope names that clash with intrinsic modules

Intrinsic module names are not in the user's namespace, so they
are free to declare global names that conflict with intrinsic
modules.

Differential Revision: https://reviews.llvm.org/D126140

2 years ago[RISCV] Make old JIT ExecutionEngine tests unsupported
Xeonacid [Tue, 24 May 2022 00:58:23 +0000 (02:58 +0200)]
[RISCV] Make old JIT ExecutionEngine tests unsupported

Make old JIT ExecutionEngine tests unsupported for RISCV, like many other architectures included.

Reviewed By: reames

Differential Revision: https://reviews.llvm.org/D126188

2 years ago[flang] Fix character length calculation for Unicode component
Peter Klausler [Wed, 11 May 2022 20:15:59 +0000 (13:15 -0700)]
[flang] Fix character length calculation for Unicode component

The character length value in the derived type component information table
entry is already in units of characters, not bytes, so don't divide by the
per-character byte size.

Differential Revision: https://reviews.llvm.org/D126139

2 years ago[lld][WebAssembly] Allow use of statically allocated TLS region.
Sam Clegg [Fri, 20 May 2022 21:39:33 +0000 (14:39 -0700)]
[lld][WebAssembly] Allow use of statically allocated TLS region.

It turns out we were already allocating static address space for TLS
data along with the non-TLS static data, but this space was going
unused/ignored.

With this change, we include the TLS segment in `__wasm_init_memory`
(which does the work of loading the passive segments into memory when a
module is first loaded).  We also set the `__tls_base` global to point
to the start of this segment.

This means that the runtime can use this static copy of the TLS data for
the first/primary thread if it chooses, rather than doing a runtime
allocation prior to calling `__wasm_init_tls`.

Practically speaking, this will allow emscripten to avoid dynamic
allocation of TLS region on the main thread.

Differential Revision: https://reviews.llvm.org/D126107

2 years ago[BasicBlockUtils] Do not move loop metadata if outer loop header.
Hendrik Greving [Fri, 13 May 2022 17:53:13 +0000 (10:53 -0700)]
[BasicBlockUtils] Do not move loop metadata if outer loop header.

Fixes a bug preventing moving the loop's metadata to an outer loop's header,
which happens if the loop's exit is also the header of an outer loop.

Adjusts test for above.

Fixes #55416.

Differential Revision: https://reviews.llvm.org/D125574

2 years ago[BasicBlockUtils] Add corner case test for loop metadata.
Hendrik Greving [Mon, 16 May 2022 14:34:04 +0000 (07:34 -0700)]
[BasicBlockUtils] Add corner case test for loop metadata.

Adds a test to expose #55416.

Differential Revision: https://reviews.llvm.org/D125696

2 years agoApply clang-tidy fixes for modernize-use-bool-literals in Parser.cpp (NFC)
Mehdi Amini [Mon, 16 May 2022 10:33:00 +0000 (10:33 +0000)]
Apply clang-tidy fixes for modernize-use-bool-literals in Parser.cpp (NFC)

2 years agoApply clang-tidy fixes for modernize-use-override in SparseTensorUtils.cpp (NFC)
Mehdi Amini [Mon, 16 May 2022 10:24:43 +0000 (10:24 +0000)]
Apply clang-tidy fixes for modernize-use-override in SparseTensorUtils.cpp (NFC)

2 years agoApply clang-tidy fixes for performance-unnecessary-value-param in Utils.cpp (NFC)
Mehdi Amini [Mon, 16 May 2022 10:09:28 +0000 (10:09 +0000)]
Apply clang-tidy fixes for performance-unnecessary-value-param in Utils.cpp (NFC)

2 years ago[test][clang] Move -O3 in command line
Vitaly Buka [Mon, 23 May 2022 22:56:35 +0000 (15:56 -0700)]
[test][clang] Move -O3 in command line

2 years agoAdd new hidden option -print-on-crash that prints out IR that caused opt pipeline...
Jamie Schmeiser [Thu, 7 Oct 2021 19:02:19 +0000 (15:02 -0400)]
Add new hidden option -print-on-crash that prints out IR that caused opt pipeline to crash

A new hidden option -print-on-crash that prints the IR as it was upon entering
the last pass when there is a crash.

The IR is saved in its print form before each pass is started and a
signal handler is registered.  If the compilation crashes, the signal
handler will print the saved IR to dbgs().  This option
can be modified using -print-module-scope to get the IR for the complete
module.  Note that this option only works with the new pass manager.

Reviewed By: yrouban

Differential Revision: https://reviews.llvm.org/D86657

2 years agogithub: Switch release PR repository to llvm/llvm-project-release-prs
Tom Stellard [Mon, 23 May 2022 22:09:26 +0000 (15:09 -0700)]
github: Switch release PR repository to llvm/llvm-project-release-prs

As discussed in https://discourse.llvm.org/t/creating-a-new-repository-for-release-branch-pull-requests/61339

Reviewed By: asl

Differential Revision: https://reviews.llvm.org/D125851

2 years ago[libc][docs] Use same formatting for headers in source_layout
Alex Brachet [Mon, 23 May 2022 21:47:22 +0000 (21:47 +0000)]
[libc][docs] Use same formatting for headers in source_layout

utils looks different from the other directory names
in the docs, see
https://libc.llvm.org/source_layout.html#the-utils-directory

Differential revision: https://reviews.llvm.org/D126211

2 years ago[TableGen] emitStringLiteralDef: Pad trailing '\0' at the end of char array.
NAKAMURA Takumi [Sat, 21 May 2022 23:52:03 +0000 (08:52 +0900)]
[TableGen] emitStringLiteralDef: Pad trailing '\0' at the end of char array.

Fixup for https://reviews.llvm.org/D73044
String literal has an implicit terminator '\0'. This commit adjusts char array
to long literal.

This causes difference of artifacts between -long-string-literals=true
and false.

Differential Revision: https://reviews.llvm.org/D126136

2 years agoFix lldb-vscode frame test failure
Jeffrey Tan [Mon, 23 May 2022 17:17:44 +0000 (10:17 -0700)]
Fix lldb-vscode frame test failure

Previous patch (https://reviews.llvm.org/D126013) added a new "optimized"
attribute to DAP stack frame this caused some tests, like
lldb-vscode/coreFile/TestVSCode_coreFile.py
to fail because the tests explicitly check for all attributes.

To fix the test failure I decided to remove this attribute.

Differential Revision: https://reviews.llvm.org/D126225

2 years agoemitStringLiteralDef: Return earlier here. NFC.
NAKAMURA Takumi [Sat, 21 May 2022 23:56:30 +0000 (08:56 +0900)]
emitStringLiteralDef: Return earlier here. NFC.

Differential Revision: https://reviews.llvm.org/D126135

2 years ago[symbolizer] Parse DW_TAG_variable DIs to show line info for globals
Mitch Phillips [Mon, 23 May 2022 20:11:01 +0000 (13:11 -0700)]
[symbolizer] Parse DW_TAG_variable DIs to show line info for globals

Currently, llvm-symbolizer doesn't like to parse .debug_info in order to
show the line info for global variables. addr2line does this. In the
future, I'm looking to migrate AddressSanitizer off of internal metadata
over to using debuginfo, and this is predicated on being able to get the
line info for global variables.

This patch adds the requisite support for getting the line info from the
.debug_info section for symbolizing global variables. This only happens
when you ask for a global variable to be symbolized as data.

Reviewed By: dblaikie

Differential Revision: https://reviews.llvm.org/D123538

2 years ago[SelectOpti][2/5] Select-to-branch base transformation
Sotiris Apostolakis [Mon, 23 May 2022 14:47:32 +0000 (10:47 -0400)]
[SelectOpti][2/5] Select-to-branch base transformation

This patch implements the actual transformation of selects to branches.
It includes only the base transformation without any sinking.

Depends on D120230

Reviewed By: davidxl

Differential Revision: https://reviews.llvm.org/D122259

2 years agoRemove duplicate fields in RAGreedy
Qunyan Mangus [Mon, 23 May 2022 19:48:06 +0000 (12:48 -0700)]
Remove duplicate fields in RAGreedy

RAGreedy has two fields of RegisterClassInfo, one called RCI and another RegClassInfo from its base class.
RCI is initialized without freezeReservedRegs first, while RegClassInfo does. Therefore, if reserved registers
information is changed between last time freezeReservedRegs is called and RAGreedy, it's not picked up by RCI.
Instead of having both fields in RAGreedy, remove RCI and use RegClassInfo instead. Also removed is the TRI field
which is present in its base class.

Reviewed By: MatzeB

Differential Revision: https://reviews.llvm.org/D125926

2 years ago[PS5] Make driver's PIC behavior match PS4
Paul Robinson [Mon, 23 May 2022 19:49:20 +0000 (12:49 -0700)]
[PS5] Make driver's PIC behavior match PS4

The new test is a copy of the corresponding PS4 test, with the triple
etc updated, because there's currently no good way to make one lit test
"iterate" with multiple targets.

2 years ago[libc++] Remove duplicate tests for callable concepts
Louis Dionne [Mon, 23 May 2022 19:36:35 +0000 (15:36 -0400)]
[libc++] Remove duplicate tests for callable concepts

This is essentially a revert of c7ad02009. Indeed, it seems that both
96dbdd75 and c7ad02009 were committed, but c7ad02009 seems to be only
an older version of 96dbdd75's tests.

2 years ago[mlir] Use 'native' instead of 'llvm_has_native_target' in the mlir tests
Stella Stamenova [Mon, 23 May 2022 19:38:02 +0000 (12:38 -0700)]
[mlir] Use 'native' instead of 'llvm_has_native_target' in the mlir tests

The tests actually require the target triple to match the host, rather than just having the host in the list of available targets. This change removes `llvm_has_native_target` and instead uses the `native` feature from the lit configuration.

Reviewed By: stellaraccident

Differential Revision: https://reviews.llvm.org/D126011

2 years ago[AArch64] Add tests with free shuffles for indexed fma variants.
Florian Hahn [Mon, 23 May 2022 19:27:42 +0000 (20:27 +0100)]
[AArch64] Add tests with free shuffles for indexed fma variants.

The new tests contain examples where shuffles are free, because indexed
fma instructions can be used.

2 years ago[SVEInstrFormats] Ensure scatter instructions are named consistently.
Paul Walker [Mon, 23 May 2022 18:07:10 +0000 (19:07 +0100)]
[SVEInstrFormats] Ensure scatter instructions are named consistently.

2 years ago[SLP][NFC]Improve compile time, NFC.
Alexey Bataev [Mon, 23 May 2022 15:09:55 +0000 (08:09 -0700)]
[SLP][NFC]Improve compile time, NFC.

Builds UserIgnore list only once as a SmallDenseSet without rebuilding
it between the runs, iterate over gathers instead list of reduction ops,
do some checks in the buildTree_rec only if the corresponding containers
  are not empty.

2 years ago[Sanitizer][Darwin] Add explanation for Apple platform macros
Julian Lettner [Mon, 23 May 2022 18:32:38 +0000 (11:32 -0700)]
[Sanitizer][Darwin] Add explanation for Apple platform macros

Differential Revision: https://reviews.llvm.org/D126229

2 years ago[gn build] Port eebc1fb772c5
LLVM GN Syncbot [Mon, 23 May 2022 18:52:16 +0000 (18:52 +0000)]
[gn build] Port eebc1fb772c5

2 years ago[libc++] Add ranges::max_element to the synopsis and ADL-proof the __min_element_impl...
Nikolas Klauser [Sun, 22 May 2022 11:43:37 +0000 (13:43 +0200)]
[libc++] Add ranges::max_element to the synopsis and ADL-proof the __min_element_impl calls

Reviewed By: ldionne, #libc

Spies: sstefan1, libcxx-commits

Differential Revision: https://reviews.llvm.org/D126167

2 years ago[libc++] Add auto to the list of required extensions in C++03
Nikolas Klauser [Sun, 22 May 2022 11:34:22 +0000 (13:34 +0200)]
[libc++] Add auto to the list of required extensions in C++03

We use `auto` in C++03, so we shouldn't say that we aren't.

Reviewed By: ldionne, #libc

Spies: libcxx-commits

Differential Revision: https://reviews.llvm.org/D126165

2 years ago[libc++] Assume that push_macro and pop_macro are available
Nikolas Klauser [Fri, 20 May 2022 15:11:58 +0000 (17:11 +0200)]
[libc++] Assume that push_macro and pop_macro are available

All compilers that libc++ supports support `push_macro` and `pop_macro`. So let's remove it.

Reviewed By: ldionne, #libc

Spies: libcxx-commits, mgorny

Differential Revision: https://reviews.llvm.org/D126073

2 years ago[libc++] Always enable the ranges concepts
Nikolas Klauser [Thu, 12 May 2022 13:46:18 +0000 (15:46 +0200)]
[libc++] Always enable the ranges concepts

The ranges concepts were already available in libc++13, so we shouldn't guard them with `_LIBCPP_HAS_NO_INCOMPLETE_RANGES`.
Fixes https://github.com/llvm/llvm-project/issues/54765

Reviewed By: #libc, ldionne

Spies: ldionne, libcxx-commits

Differential Revision: https://reviews.llvm.org/D124011

2 years ago[libc++] Granularize parts of <type_traits>
Nikolas Klauser [Fri, 20 May 2022 21:31:13 +0000 (23:31 +0200)]
[libc++] Granularize parts of <type_traits>

`<type_traits>` is quite a large header, so I'll granularize it in a few steps.

Reviewed By: ldionne, #libc

Spies: libcxx-commits, mgorny

Differential Revision: https://reviews.llvm.org/D124755

2 years agoRemove `friend` classes from TypeCategoryMap
Jorge Gorbe Moya [Mon, 23 May 2022 18:29:10 +0000 (11:29 -0700)]
Remove `friend` classes from TypeCategoryMap

As far as I can tell, the only thing those friend classes access is the
`ValueSP` typedef.

Given that this is a map-ish class, with "Map" in its name, it doesn't
seem like a stretch to make `KeyType`, `ValueType` and `ValueSP` public.
More so when the public methods of the class have `KeyType` and
`ValueSP` arguments and clearly `ValueSP` needs to be accessed from the
outside.

`friend` complicates local reasoning about who can access private
members, which is valuable in a class like this that has every method
locking a mutex to prevent concurrent access.

Differential Revision: https://reviews.llvm.org/D126103

2 years ago[mlir][tosa] Change tosa.depthwise_conv2d's ending reshape to a collapse.
natashaknk [Mon, 23 May 2022 17:58:33 +0000 (10:58 -0700)]
[mlir][tosa] Change tosa.depthwise_conv2d's ending reshape to a collapse.

TOSAs depthwise_conv2d operation includes a reshape to include the implicit x1 dimension.

Reviewed By: rsuderman

Differential Revision: https://reviews.llvm.org/D126212

2 years ago[Sanitizer][Darwin] Add SANITIZER_DRIVERKIT platform macro
Julian Lettner [Mon, 23 May 2022 18:18:15 +0000 (11:18 -0700)]
[Sanitizer][Darwin] Add SANITIZER_DRIVERKIT platform macro

2 years ago[IR] add and use pattern match specialization for sqrt intrinsic; NFC
Sanjay Patel [Mon, 23 May 2022 17:31:00 +0000 (13:31 -0400)]
[IR] add and use pattern match specialization for sqrt intrinsic; NFC

This was included in D126190 originally, but it's
independent and a useful change for readability.

2 years ago[DAGCombiner][AArch64] Don't fold (smulo x, 2) -> (saddo x, x) if VT is i2.
Craig Topper [Mon, 23 May 2022 05:38:04 +0000 (22:38 -0700)]
[DAGCombiner][AArch64] Don't fold (smulo x, 2) -> (saddo x, x) if VT is i2.

If the VT is i2, then 2 is really -2.

Test has not been commited yet, but diff shows the change.

Fixes PR55644.

Differential Revision: https://reviews.llvm.org/D126213