platform/upstream/llvm.git
13 months ago[AArch64][SME] Rename strided load/store enums
Matt Devereau [Tue, 20 Jun 2023 15:05:29 +0000 (15:05 +0000)]
[AArch64][SME] Rename strided load/store enums

This patch renames load/store enums to be equivalent
to the contiguous addressing modes with _STRIDED and _STRIDED_IMM
suffixed.

13 months ago[Attributor] Convert test to opaque pointers (NFC)
Nikita Popov [Tue, 20 Jun 2023 15:38:08 +0000 (17:38 +0200)]
[Attributor] Convert test to opaque pointers (NFC)

13 months ago[Attributor] Name instructions in test (NFC)
Nikita Popov [Tue, 20 Jun 2023 15:34:55 +0000 (17:34 +0200)]
[Attributor] Name instructions in test (NFC)

13 months ago[Attributor] Convert some tests to opaque pointers (NFC)
Nikita Popov [Tue, 20 Jun 2023 15:24:55 +0000 (17:24 +0200)]
[Attributor] Convert some tests to opaque pointers (NFC)

13 months ago[clang-format] Add InsertNewlineAtEOF to .clang-format files
Owen Pan [Sat, 17 Jun 2023 22:28:49 +0000 (15:28 -0700)]
[clang-format] Add InsertNewlineAtEOF to .clang-format files

Also, reformat all clang-format related files.

Differential Revision: https://reviews.llvm.org/D153208

13 months ago[libc++] Add missing [[maybe_unused]] attribute in format tests
Louis Dionne [Mon, 19 Jun 2023 18:40:19 +0000 (14:40 -0400)]
[libc++] Add missing [[maybe_unused]] attribute in format tests

Otherwise, Clang complains about format_ctx being unused in the tests
when exceptions are disabled in Freestanding mode. I don't know why it
doesn't complain not in freestanding mode.

Differential Revision: https://reviews.llvm.org/D153301

13 months ago[LoopVersioning] Convert tests to opaque pointers (NFC)
Nikita Popov [Tue, 20 Jun 2023 15:22:19 +0000 (17:22 +0200)]
[LoopVersioning] Convert tests to opaque pointers (NFC)

13 months ago[LoopVersioning] Regenerate test checks (NFC)
Nikita Popov [Tue, 20 Jun 2023 15:22:07 +0000 (17:22 +0200)]
[LoopVersioning] Regenerate test checks (NFC)

13 months ago[NFC][libc++] Addresses LWG3927.
Mark de Wever [Mon, 19 Jun 2023 15:06:43 +0000 (17:06 +0200)]
[NFC][libc++] Addresses LWG3927.

Changes to preconditions have no impact on the library.

Implements
- LWG3927 Unclear preconditions for operator[] for sequence containers

Reviewed By: #libc, philnik

Differential Revision: https://reviews.llvm.org/D153286

13 months ago[NFC][libc++] Addresses LWG3935.
Mark de Wever [Mon, 19 Jun 2023 15:06:43 +0000 (17:06 +0200)]
[NFC][libc++] Addresses LWG3935.

Note libc++ implemented this in its initial version.

Implements
- LWG3935 template<class X> constexpr complex& operator=(const complex<X>&) has no specification

Reviewed By: #libc, philnik, ldionne

Differential Revision: https://reviews.llvm.org/D153287

13 months ago[SafepointIRVerifier] Convert test to opaque pointers (NFC)
Nikita Popov [Tue, 20 Jun 2023 15:14:52 +0000 (17:14 +0200)]
[SafepointIRVerifier] Convert test to opaque pointers (NFC)

13 months ago[DAG] Add getExtOrTrunc helper. NFC.
Simon Pilgrim [Tue, 20 Jun 2023 15:03:07 +0000 (16:03 +0100)]
[DAG] Add getExtOrTrunc helper. NFC.

Wrap the getSExtOrTrunc/getZExtOrTrunc calls behind an IsSigned argument.

13 months ago[LTO] Avoid -opaque-pointers=0 in test (NFC)
Nikita Popov [Tue, 20 Jun 2023 14:56:09 +0000 (16:56 +0200)]
[LTO] Avoid -opaque-pointers=0 in test (NFC)

Commit the bitcode file instead of generating it dynamically, as
this will no longer be possible in the future.

13 months ago[libc++] Add incomplete availability markup for std::pmr
Louis Dionne [Wed, 29 Mar 2023 20:48:20 +0000 (16:48 -0400)]
[libc++] Add incomplete availability markup for std::pmr

This fixes rdar://110330781, which asked for the feature-test macro
for std::pmr to take into account the deployment target. It doesn't
fix https://llvm.org/PR62212, though, because the availability markup
itself must be disabled until some Clang bugs have been fixed.

This is pretty vexing, however at least everything should work once
those Clang bugs have been fixed. In the meantime, this patch at least
adds the required markup (as disabled) and ensures that the feature-test
macro for std::pmr is aware of the deployment target requirement.

Differential Revision: https://reviews.llvm.org/D135813

13 months ago[DebugInfo] Convert tests to opaque pointers (NFC)
Nikita Popov [Tue, 20 Jun 2023 14:44:16 +0000 (16:44 +0200)]
[DebugInfo] Convert tests to opaque pointers (NFC)

13 months ago[Bitcode] Remove -opaque-pointer=0 check lines (NFC)
Nikita Popov [Tue, 20 Jun 2023 14:43:14 +0000 (16:43 +0200)]
[Bitcode] Remove -opaque-pointer=0 check lines (NFC)

These tests were testing both typed an opaque pointers. Only keep
opaque pointers tests.

13 months ago[llvm-nm] Avoid -opaque-pointers option in test (NFC)
Nikita Popov [Tue, 20 Jun 2023 14:34:58 +0000 (16:34 +0200)]
[llvm-nm] Avoid -opaque-pointers option in test (NFC)

Commit the typed pointer bitcode file instead of producing it,
as this will not be possible in the future anymore.

13 months ago[DAG] Fold (abds x, y) -> (abdu x, y) iff both args are known positive
Simon Pilgrim [Tue, 20 Jun 2023 14:10:32 +0000 (15:10 +0100)]
[DAG] Fold (abds x, y) -> (abdu x, y) iff both args are known positive

This is a generic DAG combine version of D151055 which recognizes when a signed ABDS can be safely replaced with a unsigned ABDU instruction if it is legal.

Alive2: https://alive2.llvm.org/ce/z/pb5BjG

Differential Revision: https://reviews.llvm.org/D153328

13 months ago[AArch64] Use ISD::isExtOpcode. NFC
David Green [Tue, 20 Jun 2023 14:26:01 +0000 (15:26 +0100)]
[AArch64] Use ISD::isExtOpcode. NFC

13 months ago[AArch64] Try to fold uaddlv and uaddlp
Jingu Kang [Mon, 19 Jun 2023 16:48:00 +0000 (17:48 +0100)]
[AArch64] Try to fold uaddlv and uaddlp

Add tablegen pattern for uaddlv(uaddlp(x)) ==> uaddlv(x).

Differential Revision: https://reviews.llvm.org/D153323

13 months ago[AMDGPU] Remove unused method Waitcnt::dominates(). NFC
Stephen Thomas [Tue, 20 Jun 2023 08:59:08 +0000 (09:59 +0100)]
[AMDGPU] Remove unused method Waitcnt::dominates(). NFC

Differential Revision: https://reviews.llvm.org/D153322

13 months ago[LoongArch] Optimize conditional selection of integer
Weining Lu [Tue, 20 Jun 2023 13:50:54 +0000 (21:50 +0800)]
[LoongArch] Optimize conditional selection of integer

This patch optimizes code generation by leveraging the zeroing behavior of the `maskeqz`/`masknez` instructions.

```
int sel(int a, int b)
{
    return (a < b) ? a : 0;
}
```

```
slt $a1,$a0,$a1
masknez $a2,$r0,$a1
maskeqz $a0,$a0,$a1
or $a0,$a0,$a2
```

=>

```
slt $a1,$a0,$a1
maskeqz $a0,$a0,$a1
```

Reviewed By: SixWeining

Differential Revision: https://reviews.llvm.org/D153193

13 months ago[AMDGPU] Use verify<domtree> instead of intra-pass asserts.
Pravin Jagtap [Tue, 20 Jun 2023 13:52:58 +0000 (09:52 -0400)]
[AMDGPU] Use verify<domtree> instead of intra-pass asserts.

Verifying dominator tree is expensive using intra-pass
asserts. Asserts added during D147408 are
increasing the build time of libc significantly. This change
does the verification after the atomic optimizer pass
and should fix the regression reported in D153232.

Reviewed By: arsenm, #amdgpu

Differential Revision: https://reviews.llvm.org/D153261

13 months ago[LoongArch] Indent LoongArchInstrInfo.td a little bit. NFC
Weining Lu [Tue, 20 Jun 2023 13:49:28 +0000 (21:49 +0800)]
[LoongArch] Indent LoongArchInstrInfo.td a little bit. NFC

13 months ago[libc][math] Improve exp2f performance.
Tue Ly [Fri, 16 Jun 2023 13:22:00 +0000 (09:22 -0400)]
[libc][math] Improve exp2f performance.

Re-organize special cases and add a special case when `|x| < 2^-5`.

Reviewed By: michaelrj

Differential Revision: https://reviews.llvm.org/D153134

13 months ago[libc][math] Slightly improve sinhf and coshf performance.
Tue Ly [Thu, 15 Jun 2023 18:56:08 +0000 (14:56 -0400)]
[libc][math] Slightly improve sinhf and coshf performance.

Re-order exceptional branches and slightly adjust the evaluation.
Depends on https://reviews.llvm.org/D153026 .

Reviewed By: michaelrj

Differential Revision: https://reviews.llvm.org/D153062

13 months ago[libc][math] Improve tanhf performance.
Tue Ly [Thu, 15 Jun 2023 05:21:57 +0000 (01:21 -0400)]
[libc][math] Improve tanhf performance.

Re-order exceptional branches and slightly adjust the evaluation.

Performance tested with the CORE-MATH project on AMD EPYC 7B12 (clocks/op)

Reciprocal throughputs:
```
--- BEFORE ---

$ CORE_MATH_PERF_MODE=rdtsc ./perf.sh tanhf
[####################] 100 %  (with -mavx2 -mfma)
Ntrial = 20 ; Min = 7.794 + 0.102 clc/call; Median-Min = 0.066 clc/call; Max = 8.267 clc/call;
[####################] 100 %. (with -msse4.2)
Ntrial = 20 ; Min = 10.783 + 0.172 clc/call; Median-Min = 0.144 clc/call; Max = 11.446 clc/call;
[####################] 100 %. (SSE2)
Ntrial = 20 ; Min = 18.926 + 0.381 clc/call; Median-Min = 0.342 clc/call; Max = 19.623 clc/call;

--- AFTER ---

$ CORE_MATH_PERF_MODE=rdtsc ./perf.sh tanhf
[####################] 100 %  (with -mavx2 -mfma)
Ntrial = 20 ; Min = 6.598 + 0.085 clc/call; Median-Min = 0.052 clc/call; Max = 6.868 clc/call;
[####################] 100 %  (with -msse4.2)
Ntrial = 20 ; Min = 9.245 + 0.304 clc/call; Median-Min = 0.248 clc/call; Max = 10.675 clc/call;
[####################] 100 %. (SSE2)
Ntrial = 20 ; Min = 11.724 + 0.440 clc/call; Median-Min = 0.444 clc/call; Max = 12.262 clc/call;
```

Latency:
```
--- BEFORE ---

$ PERF_ARGS="--latency" CORE_MATH_PERF_MODE=rdtsc ./perf.sh tanhf
[####################] 100 %  (with -mavx2 -mfma)
Ntrial = 20 ; Min = 38.821 + 0.157 clc/call; Median-Min = 0.122 clc/call; Max = 39.539 clc/call;
[####################] 100 %. (with -msse4.2)
Ntrial = 20 ; Min = 44.767 + 0.766 clc/call; Median-Min = 0.681 clc/call; Max = 45.951 clc/call;
[####################] 100 %. (SSE2)
Ntrial = 20 ; Min = 55.055 + 1.512 clc/call; Median-Min = 1.571 clc/call; Max = 57.039 clc/call;

--- AFTER ---

$ PERF_ARGS="--latency" CORE_MATH_PERF_MODE=rdtsc ./perf.sh tanhf
[####################] 100 %  (with -mavx2 -mfma)
Ntrial = 20 ; Min = 36.147 + 0.194 clc/call; Median-Min = 0.181 clc/call; Max = 36.536 clc/call;
[####################] 100 %  (with -msse4.2)
Ntrial = 20 ; Min = 40.904 + 0.728 clc/call; Median-Min = 0.557 clc/call; Max = 42.231 clc/call;
[####################] 100 %. (SSE2)
Ntrial = 20 ; Min = 55.776 + 0.557 clc/call; Median-Min = 0.542 clc/call; Max = 56.551 clc/call;
```

Reviewed By: michaelrj

Differential Revision: https://reviews.llvm.org/D153026

13 months ago[LoongArch] Add missing chains and remove unnecessary `SDNPSideEffect` property for...
Weining Lu [Tue, 20 Jun 2023 09:21:31 +0000 (17:21 +0800)]
[LoongArch] Add missing chains and remove unnecessary `SDNPSideEffect` property for some intrinsic nodes

13 months ago[AArch64] Add tablegen patterns for fp16 fcvtn2.
David Green [Tue, 20 Jun 2023 13:10:25 +0000 (14:10 +0100)]
[AArch64] Add tablegen patterns for fp16 fcvtn2.

Similar to the existing f32 pattern, this adds a tablegen pattern for the fp16
fcvtn2.

13 months ago[lld][Arm] Big Endian - Byte invariant support.
Simi Pallipurath [Thu, 18 May 2023 11:29:07 +0000 (12:29 +0100)]
[lld][Arm] Big Endian - Byte invariant support.

Arm has BE8 big endian configuration called a byte-invariant(every byte has the same address on little and big-endian systems).

When in BE8 mode:
  1. Instructions are big-endian in relocatable objects but
     little-endian in executables and shared objects.
  2. Data is big-endian.
  3. The data encoding of the ELF file is ELFDATA2MSB.

To support BE8 without an ABI break for relocatable objects,the linker takes on the responsibility of changing the endianness of instructions. At a high level the only difference between BE32 and BE8 in the linker is that for BE8:
  1. The linker sets the flag EF_ARM_BE8 in the ELF header.
  2. The linker endian reverses the instructions, but not data.

This patch adds BE8 big endian support for Arm. To endian reverse the instructions we'll need access to the mapping symbols. Code sections can contain a mix of Arm, Thumb and literal data. We need to endian reverse Arm instructions as words, Thumb instructions
as half-words and ignore literal data.The only way to find these transitions precisely is by using mapping symbols. The instruction reversal will need to take place after relocation. For Arm BE8 code sections (Section has SHF_EXECINSTR flag ) we inserted a step after relocation to endian reverse the instructions. The implementation strategy i have used here is to write all sections BE32  including SyntheticSections then endian reverse all code in InputSections via mapping symbols.

Reviewed By: peter.smith

Differential Revision: https://reviews.llvm.org/D150870

13 months ago[LinkerWrapper] Support linking vendor bitcode late
Joseph Huber [Thu, 8 Jun 2023 13:28:07 +0000 (08:28 -0500)]
[LinkerWrapper] Support linking vendor bitcode late

The GPU vendors currently provide bitcode files for their device
runtime. These files need to be handled specially as they are not built
to be linked in with a standard `llvm-link` call or through LTO linking.
This patch adds an alternative to use the existing clang handling of
these libraries that does the necessary magic to make this work.

We do this by causing the LTO backend to emit bitcode before running the
backend. We then pass this through to clang which uses the existing
support which has been fixed to support this by D152391. The backend
will then be run with the merged module.

This patch adds the `--builtin-bitcode=<triple>=file.bc` to specify a single
file, or just `--clang-backend` to let the toolchain handle its defaults
(currently nothing for NVPTX and the ROCm device libs for AMDGPU). This may have
a performance impact due to running the optimizations again, we could
potentially disable optimizations in LTO and only do the linking if this is an
issue.

This should allow us to resolve issues when relying on the `linker-wrapper` to
do a late linking that may depend on vendor libraries.

Depends on D152391

Reviewed By: JonChesterfield

Differential Revision: https://reviews.llvm.org/D152442

13 months ago[Clang] Allow bitcode linking when the input is LLVM-IR
Joseph Huber [Wed, 7 Jun 2023 18:11:19 +0000 (13:11 -0500)]
[Clang] Allow bitcode linking when the input is LLVM-IR

Clang provides the `-mlink-bitcode-file` and `-mlink-builtin-bitcode`
options to insert LLVM-IR into the current TU. These are usefuly
primarily for including LLVM-IR files that require special handling to
be correct and cannot be linked normally, such as GPU vendor libraries
like `libdevice.10.bc`. Currently these options can only be used if the
source input goes through the AST consumer path. This patch makes the
changes necessary to also support this when the input is LLVM-IR. This
will allow the following operation:

```
clang in.bc -Xclang -mlink-builtin-bitcode -Xclang libdevice.10.bc
```

Reviewed By: yaxunl

Differential Revision: https://reviews.llvm.org/D152391

13 months ago[mlir][Pass] Check supported op types before running a pass
Matthias Springer [Tue, 20 Jun 2023 12:39:40 +0000 (14:39 +0200)]
[mlir][Pass] Check supported op types before running a pass

Add extra error checking to prevent passes from being run on unsupported ops through the pass manager infrastructure.

Differential Revision: https://reviews.llvm.org/D153144

13 months ago[include-cleaner] Ignore the ParmVarDecl itself in WalkAST.cpp
Haojian Wu [Tue, 20 Jun 2023 11:10:45 +0000 (13:10 +0200)]
[include-cleaner] Ignore the ParmVarDecl itself in WalkAST.cpp

This fixes a false positive where a ParamVarDecl happend to be the
same name of some C standard symbol and has a global namespace.

```
using A = int(int time); // we suggest <ctime> for the `int time`.
```

Differential Revision: https://reviews.llvm.org/D153330

13 months agoRevert "Reland "[DebugMetadata][DwarfDebug] Support function-local types in lexical...
Vladislav Dzhidzhoev [Tue, 20 Jun 2023 11:08:47 +0000 (13:08 +0200)]
Revert "Reland "[DebugMetadata][DwarfDebug] Support function-local types in lexical block scopes (4/7)" (2)"

This reverts commit cb9ac7051589ea0d05507f9370d0716bef86b4ae.
It causes an assert in clang:
virtual void llvm::DwarfDebug::endFunctionImpl(const llvm::MachineFunction*): Assertion `LScopes.getAbstractScopesList().size() == NumAbstractSubprograms && "getOrCreateAbstractScope() inserted an abstract subprogram scope"' failed.
https://bugs.chromium.org/p/chromium/issues/detail?id=1456288#c2

13 months ago[DWARFLinker] add DWARFUnit::getIndexedAddressOffset().
Alexey Lapshin [Mon, 19 Jun 2023 15:04:19 +0000 (17:04 +0200)]
[DWARFLinker] add DWARFUnit::getIndexedAddressOffset().

This patch is a followup for D153162. It cures one more place
where indexed address was incorrectly read. It also moves handling
of indexed address into DWARFUnit.

Differential Revision: https://reviews.llvm.org/D153297

13 months ago[mlir] mark libraries in mlir/examples as such
Alex Zinenko [Mon, 19 Jun 2023 11:21:33 +0000 (13:21 +0200)]
[mlir] mark libraries in mlir/examples as such

LLVM build system separates between `add_llvm_example_library` and
`add_llvm_library`, which is presumably used to package examples
separately from the regular library. Introduce a similar approach to
building example libraries in MLIR and use it for the transform dialect
tutorial.

Reviewed By: mehdi_amini

Differential Revision: https://reviews.llvm.org/D153265

13 months ago[AArch64] Add Cortex-A510 specific scheduling
Harvin Iriawan [Tue, 20 Jun 2023 10:48:18 +0000 (11:48 +0100)]
[AArch64] Add Cortex-A510 specific scheduling

- Update the Cortex-A510 mcpu target to use A510 scheduling info instead of
  A55. Values taken are based on the A510 software optimisation guide
  https://developer.arm.com/documentation/PJDOC-466751330-536816/latest
- Make latency of most integer ops to 1. CPU uarch is able to resolve most
  integer ops in 1 cycle

Differential Revision: https://reviews.llvm.org/D152688

13 months ago[clang][NFC] Drop alignment in builtin-nondeterministic-value test
ManuelJBrito [Sun, 11 Jun 2023 19:39:45 +0000 (20:39 +0100)]
[clang][NFC] Drop alignment in builtin-nondeterministic-value test

Drop alignment to allow test to run in different platforms.

Differential Revision: https://reviews.llvm.org/D152547

13 months ago[Coroutines] Store the index for final suspend point in the exception path
Chuanqi Xu [Tue, 20 Jun 2023 10:35:58 +0000 (18:35 +0800)]
[Coroutines] Store the index for final suspend point in the exception path

Try to address part of
https://github.com/llvm/llvm-project/issues/61900.

It is not completely addressed since the original reproducer is not
fixed due to the final suspend point is optimized out in its special
case. But that is a relatively independent issue.

13 months ago[AMDGPU] Drop GFX11 runs for dagcombine-fma-fmad.ll and fma.f16.ll.
Ivan Kosarev [Tue, 20 Jun 2023 10:29:40 +0000 (11:29 +0100)]
[AMDGPU] Drop GFX11 runs for dagcombine-fma-fmad.ll and fma.f16.ll.

They cause failures on the llvm-clang-x86_64-expensive-checks-debian
buildbot.

This partially reverts
D153269 [AMDGPU][GFX11] Add test coverage for FMA instructions.

13 months ago[clang][index] Fix cast warning
Jan Svoboda [Tue, 20 Jun 2023 10:21:59 +0000 (12:21 +0200)]
[clang][index] Fix cast warning

This is a follow-up to D151938 that should fix GCC's -Wcast-qual warning.

13 months ago[CodeGen][test] Add missing `REQUIRES`.
Francesco Petrogalli [Tue, 20 Jun 2023 09:56:03 +0000 (11:56 +0200)]
[CodeGen][test] Add missing `REQUIRES`.

Differential Revision: https://reviews.llvm.org/D153325

13 months ago[AMDGPU][GFX11] Add test coverage for FMA instructions.
Ivan Kosarev [Tue, 20 Jun 2023 09:32:12 +0000 (10:32 +0100)]
[AMDGPU][GFX11] Add test coverage for FMA instructions.

Reviewed By: arsenm

Differential Revision: https://reviews.llvm.org/D153269

13 months ago[llc][MISched] Add `-misched-detail-resource-booking` to llc.
Francesco Petrogalli [Tue, 20 Jun 2023 09:42:01 +0000 (11:42 +0200)]
[llc][MISched] Add `-misched-detail-resource-booking` to llc.

The option `-misched-detail-resource-booking` prints the following
information every time the method
`SchedBoundary::getNextResourceCycle` is invoked:

1. counters of the resources that have already been booked;

2. the values returned by `getNextResourceCycle`, which is the next
available cycle in which a resource can be booked.

The method is useful to debug low-level checks inside the machine
scheduler that make decisions based on the values returned by
`getNextResourceCycle`.

Reviewed By: andreadb

Differential Revision: https://reviews.llvm.org/D153116

13 months agoRevert "[llc][MISched] Add `-misched-detail-resource-booking` to llc."
Francesco Petrogalli [Tue, 20 Jun 2023 09:28:45 +0000 (11:28 +0200)]
Revert "[llc][MISched] Add `-misched-detail-resource-booking` to llc."

Reverting because of https://lab.llvm.org/buildbot#builders/75/builds/32485:

llvm-project/llvm/lib/CodeGen/MachineScheduler.cpp:2374:7: error: use of undeclared identifier 'MischedDetailResourceBooking'
 if (MischedDetailResourceBooking)

This reverts commit fc06262c1c365777e71207b6a5de281cba927c96.

13 months ago[llc][MISched] Add `-misched-detail-resource-booking` to llc.
Francesco Petrogalli [Tue, 20 Jun 2023 09:13:39 +0000 (11:13 +0200)]
[llc][MISched] Add `-misched-detail-resource-booking` to llc.

The option `-misched-detail-resource-booking` prints the following
information every time the method
`SchedBoundary::getNextResourceCycle` is invoked:

1. counters of the resources that have already been booked;

2. the values returned by `getNextResourceCycle`, which is the next
available cycle in which a resource can be booked.

The method is useful to debug low-level checks inside the machine
scheduler that make decisions based on the values returned by
`getNextResourceCycle`.

Reviewed By: andreadb

Differential Revision: https://reviews.llvm.org/D153116

13 months ago[mlir][transform] Add TransformRewriter
Matthias Springer [Tue, 20 Jun 2023 08:48:40 +0000 (10:48 +0200)]
[mlir][transform] Add TransformRewriter

All `apply` functions now have a `TransformRewriter &` parameter. This rewriter should be used to modify the IR. It has a `TrackingListener` attached and updates the internal handle-payload mappings based on rewrites.

Implementations no longer need to create their own `TrackingListener` and `IRRewriter`. Error checking is integrated into `applyTransform`. Tracking listener errors are reported only for ops with the `ReportTrackingListenerFailuresOpTrait` trait attached, allowing for a gradual migration. Furthermore, errors can be silenced with an op attribute.

Additional API will be added to `TransformRewriter` in subsequent revisions. This revision just adds an "empty" `TransformRewriter` class and updates all `apply` implementations.

Differential Revision: https://reviews.llvm.org/D152427

13 months ago[SCEVNormalization] Short circuit case with no loops (NFC)
Nikita Popov [Tue, 20 Jun 2023 07:33:49 +0000 (09:33 +0200)]
[SCEVNormalization] Short circuit case with no loops (NFC)

If there are no post-inc loops, normalization is a no-op. Don't
bother rewriting the SCEV in that case.

13 months ago[AMDGPU] Document amdgpu_cs_chain[_preserve] CCs. NFC
Diana Picus [Wed, 7 Jun 2023 12:42:02 +0000 (14:42 +0200)]
[AMDGPU] Document amdgpu_cs_chain[_preserve] CCs. NFC

Co-authored-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Differential Revision: https://reviews.llvm.org/D151997

13 months ago[AMDGPU] Start documenting calling conventions. NFC
Diana Picus [Wed, 7 Jun 2023 11:52:36 +0000 (13:52 +0200)]
[AMDGPU] Start documenting calling conventions. NFC

Add a section to AMDGPUUsage.rst about calling conventions and list the
ones from the CallingConv enum. Full descriptions can come later (help
appreciated).

Differential Revision: https://reviews.llvm.org/D151996

13 months ago[llvm-profdata] Fix llvm-profdata help and make sure it remains in sync
serge-sans-paille [Mon, 19 Jun 2023 16:35:03 +0000 (18:35 +0200)]
[llvm-profdata] Fix llvm-profdata help and make sure it remains in sync

This makes the new `order` subcommand part of the help.
As a side effect, also make llvm::map_range compatible with plain
arrays.

Differential Revision: https://reviews.llvm.org/D153303

13 months ago[clang][DebugInfo] Emit DW_AT_deleted on any deleted member function
Michael Buch [Mon, 19 Jun 2023 12:59:32 +0000 (13:59 +0100)]
[clang][DebugInfo] Emit DW_AT_deleted on any deleted member function

Currently we emit `DW_AT_deleted` for `deleted` special-member
functions (i.e., ctors/dtors). However, in C++ one can mark any
member function as deleted. This patch expands the set of member
functions for which we emit `DW_AT_deleted`.

The DWARFv5 spec section 5.7.8 says:
```
<non-normative>
In C++, a member function may be declared as deleted. This prevents the compiler from
generating a default implementation of a special member function such as a constructor
or destructor, and can affect overload resolution when used on other member functions.
</non-normative>

If the member function entry has been declared as deleted, then that entry has a
DW_AT_deleted attribute.
```

Thus this change is conforming.

Differential Revision: https://reviews.llvm.org/D153282

13 months ago[docs] Fix GEP faq references to undefined behavior [NFC]
Nuno Lopes [Tue, 20 Jun 2023 08:07:08 +0000 (09:07 +0100)]
[docs] Fix GEP faq references to undefined behavior [NFC]

13 months ago[CSKY] Optimize multiplication with immediates
Ben Shi [Mon, 19 Jun 2023 09:14:41 +0000 (17:14 +0800)]
[CSKY] Optimize multiplication with immediates

Try to break a multiplication with a specific immediate to
an/a addition/subtraction of left shifts.

Reviewed By: zixuan-wu

Differential Revision: https://reviews.llvm.org/D153106

13 months ago[CSKY][test][NFC] Add more tests of multiplication with immediates
Ben Shi [Mon, 19 Jun 2023 09:12:25 +0000 (17:12 +0800)]
[CSKY][test][NFC] Add more tests of multiplication with immediates

Reviewed By: zixuan-wu

Differential Revision: https://reviews.llvm.org/D153105

13 months agoRevert "[AMDGPU] Start documenting calling conventions. NFC"
Diana Picus [Tue, 20 Jun 2023 07:54:40 +0000 (09:54 +0200)]
Revert "[AMDGPU] Start documenting calling conventions. NFC"

This reverts commit aa7b127cb7314e326457d7f790d36db1cb74f63c.

...because I really ought to install sphinx.

13 months agoRevert "Prevent deadlocks in death tests."
Martin Braenne [Tue, 20 Jun 2023 07:15:31 +0000 (07:15 +0000)]
Revert "Prevent deadlocks in death tests."

This reverts commit dfbcee286b9b96751014ebc5ba5290e42796be37.

This was causing unit tests to fail on Gentoo, see comments on
https://reviews.llvm.org/D152696.

13 months agoFixup D151996
Diana Picus [Tue, 20 Jun 2023 07:37:16 +0000 (09:37 +0200)]
Fixup D151996

13 months ago[AMDGPU] Start documenting calling conventions. NFC
Diana Picus [Wed, 7 Jun 2023 11:52:36 +0000 (13:52 +0200)]
[AMDGPU] Start documenting calling conventions. NFC

Add a section to AMDGPUUsage.rst about calling conventions and list the
ones from the CallingConv enum. Full descriptions can come later (help
appreciated).

Differential Revision: https://reviews.llvm.org/D151996

13 months ago[clangd] Index the type of a non-type template parameter
Nathan Ridge [Tue, 20 Jun 2023 06:57:36 +0000 (02:57 -0400)]
[clangd] Index the type of a non-type template parameter

Fixes https://github.com/clangd/clangd/issues/1666

Differential Revision: https://reviews.llvm.org/D153251

13 months ago[include-cleaner] Bailout on invalid code for the command-line tool
Haojian Wu [Mon, 19 Jun 2023 12:57:09 +0000 (14:57 +0200)]
[include-cleaner] Bailout on invalid code for the command-line tool

The binary tool only works on working source code, if the source code is
not compilable, don't perform any analysis and edits.

Differential Revision: https://reviews.llvm.org/D153271

13 months ago[mlir][transform] Add ApplyRegisteredPassOp transform op
Matthias Springer [Tue, 20 Jun 2023 06:54:49 +0000 (08:54 +0200)]
[mlir][transform] Add ApplyRegisteredPassOp transform op

This transform op runs a pass on the target op.

Differential Revision: https://reviews.llvm.org/D153143

13 months ago[tools] Use llvm::is_contained (NFC)
Kazu Hirata [Tue, 20 Jun 2023 06:36:14 +0000 (23:36 -0700)]
[tools] Use llvm::is_contained (NFC)

13 months ago[xray][AArch64] Rewrite trampoline
Fangrui Song [Tue, 20 Jun 2023 06:02:45 +0000 (23:02 -0700)]
[xray][AArch64] Rewrite trampoline

Optimize (cmp+beq => cbz), duduplicate code (SAVE_REGISTERS/RESTORE_REGISTERS),
improve portability (use ASM_SYMBOL to be compatible with Mach-O), and fix style
issues.
Also, port D37965 (x86 tail call) to __xray_FunctionTailExit.

13 months ago[lldb] Make the test for D153043 linux-only
Jaroslav Sevcik [Tue, 20 Jun 2023 05:56:16 +0000 (07:56 +0200)]
[lldb] Make the test for D153043 linux-only

13 months ago[lldb] Make test for D153043 independent of external symbols
Jaroslav Sevcik [Tue, 20 Jun 2023 05:19:37 +0000 (07:19 +0200)]
[lldb] Make test for D153043 independent of external symbols

This removes dependence on the libc abort function.

13 months ago[LV] Add cost model for simd vector select instructions of type float
Zhongyunde [Tue, 20 Jun 2023 05:12:02 +0000 (13:12 +0800)]
[LV] Add cost model for simd vector select instructions of type float

For simd vector selects, use cmeq + bsl for v2f32/v4f32/v2f64, so their cost are cheep.
Fix https://github.com/llvm/llvm-project/issues/63082

Reviewed By: dmgreen
Differential Revision: https://reviews.llvm.org/D152523

13 months ago[X86][AMX] set Stride to Tile's Col when doing combine amxcast and store into tilestore
Bing1 Yu [Mon, 19 Jun 2023 07:57:01 +0000 (15:57 +0800)]
[X86][AMX] set Stride to Tile's Col when doing combine amxcast and store into tilestore

%tile = call x86_amx @llvm.x86.tileloadd64.internal(i16 8, i16 32, i8* %src_ptr, i64 64)
%vec = call <256 x i8> @llvm.x86.cast.tile.to.vector.v256i8(x86_amx...%tile)
store <256 x i8> %vec, <256 x i8>* %dst_ptr, align 256
=>
%tile = call x86_amx @llvm.x86.tileloadd64.internal(i16 8, i16 32, i8* %src_ptr, i64 64)
%stride = sext i16 32 to i64
call void @llvm.x86.tilestored64.internal(i16 8, i16 32, i8* %dst_ptr, i64 32, x86_amx %tile)

Reviewed By: LuoYuanke

Differential Revision: https://reviews.llvm.org/D153002

13 months ago[XRay] Make llvm.xray.customevent parameter type match __xray_customevent
Fangrui Song [Tue, 20 Jun 2023 03:38:16 +0000 (20:38 -0700)]
[XRay] Make llvm.xray.customevent parameter type match __xray_customevent

The intrinsic has a smaller integer type than the parameter type of
builtin-function/API. Fix this similar to commit 3fa3cb408d8d0f1365b322262e501b6945f7ead9.

13 months ago[XRay] Make llvm.xray.typedevent parameter type match __xray_typedevent
Fangrui Song [Tue, 20 Jun 2023 03:28:39 +0000 (20:28 -0700)]
[XRay] Make llvm.xray.typedevent parameter type match __xray_typedevent

The Clang built-in function is void __xray_typedevent(size_t, const void *, size_t),
but the LLVM intrinsics has smaller integer types. Since we only allow
64-bit ELF/Mach-O targets, we can change llvm.xray.typedevent to
i64/ptr/i64.

This allows encoding more information and avoids i16 legalization for
many non-X86 targets.

fdrLoggingHandleTypedEvent only supports uint16_t event type.

13 months ago[BOLT] Fix a warning in release builds
Kazu Hirata [Tue, 20 Jun 2023 03:02:47 +0000 (20:02 -0700)]
[BOLT] Fix a warning in release builds

This patch fixes:

  bolt/lib/Core/BinarySection.cpp:120:24: error: unused variable
  'Relocation' [-Werror,-Wunused-variable]

13 months agoReland "[DebugMetadata][DwarfDebug] Support function-local types in lexical block...
Vladislav Dzhidzhoev [Mon, 19 Jun 2023 14:42:05 +0000 (16:42 +0200)]
Reland "[DebugMetadata][DwarfDebug] Support function-local types in lexical block scopes (4/7)" (2)

Test "local-type-as-template-parameter.ll" is now enabled only for
x86_64.

Authored-by: Kristina Bessonova <kbessonova@accesssoftek.com>
Differential Revision: https://reviews.llvm.org/D144006

Depends on D144005

13 months agoRevert "Reland "[DebugMetadata][DwarfDebug] Support function-local types in lexical...
Vladislav Dzhidzhoev [Mon, 19 Jun 2023 23:54:48 +0000 (01:54 +0200)]
Revert "Reland "[DebugMetadata][DwarfDebug] Support function-local types in lexical block scopes (4/7)""

This reverts commit 2da45172c4bcd42f704c57c656926f56f32fc5ce.
Test local-type-as-template-parameter.ll fails on ppc64-aix.

13 months ago[XRay][X86] Remove sled version 0 support from patchCustomEvent
Fangrui Song [Mon, 19 Jun 2023 22:11:26 +0000 (15:11 -0700)]
[XRay][X86] Remove sled version 0 support from patchCustomEvent

This is remnant after D140739.

13 months ago[xray][test] Test __xray_typedevent after D43668
Fangrui Song [Mon, 19 Jun 2023 21:53:22 +0000 (14:53 -0700)]
[xray][test] Test __xray_typedevent after D43668

13 months ago[DebugInfo] Add DW_OP_LLVM_user extension point
Scott Linder [Wed, 17 May 2023 21:33:09 +0000 (21:33 +0000)]
[DebugInfo] Add DW_OP_LLVM_user extension point

The extension codespace for DWARF expressions (DW_OP_LLVM_{lo,hi}_user)
has shrunk over time, as no extension is ever "retired" in practice. To
facilitate future extensions, this patch reserves one open opcode as an extension
point (0xfe), which is followed by a ULEB128-encoded SubOperation, and
then by the subop's operands.

There is some prior-art, namely DW_OP_AARCH64_operation
(see https://github.com/ARM-software/abi-aa/blob/edd7460d87493fff124b8b5713acf71ffc06ee91/aadwarf64/aadwarf64.rst#45dwarf-expression-operations).

This version makes some different tradeoffs, opting to use a ULEB128 for
the subop encoding for future-proofing.

Reviewed By: #debug-info, dblaikie

Differential Revision: https://reviews.llvm.org/D147271

13 months ago[AMDGPU] Do not release VGPRs if there may be pending scratch stores
Jay Foad [Mon, 19 Jun 2023 15:54:17 +0000 (16:54 +0100)]
[AMDGPU] Do not release VGPRs if there may be pending scratch stores

Differential Revision: https://reviews.llvm.org/D153295

13 months ago[AMDGPU] Remove unused macro CNT_MASK
Jay Foad [Mon, 19 Jun 2023 20:08:35 +0000 (21:08 +0100)]
[AMDGPU] Remove unused macro CNT_MASK

13 months ago[Driver] Correct -fnoxray-link-deps to -fno-xray-link-deps
Fangrui Song [Mon, 19 Jun 2023 19:48:33 +0000 (12:48 -0700)]
[Driver] Correct -fnoxray-link-deps to -fno-xray-link-deps

and removed unused CC1Option.
Also change -whole-archive to the canonical spelling and improve tests.

13 months ago[DebugInfo] Support more than 2 operands in DWARF operations
Scott Linder [Mon, 19 Jun 2023 17:13:08 +0000 (17:13 +0000)]
[DebugInfo] Support more than 2 operands in DWARF operations

Update DWARFExpression::Operation and LVOperation to support more than
2 operands.

Take the opportunity to use a SmallVector, which will handle at least 2
operands without allocation anyway, and removes the static limit
completely.

As there is no longer the concept of an "unused operand", remove
Operation::Encoding::SizeNA. Any use of it is now replaced with explicit
checks for how many operands an operation has.

There are still places where the limit remains 2, namely in the
DWARFLinker and in DIExpressions, but these can be updated in later
patches as-needed.

There are no explicit tests as this is nearly NFC: no new operation is
added which makes use of the additional operand capacity yet. A future
patch adding a new DWARF extension point will include operations which
require the support.

Reviewed By: Orlando, CarlosAlbertoEnciso

Differential Revision: https://reviews.llvm.org/D147270

13 months ago[libc] Remove the requirement of a platform-flush operation in File abstraction.
Siva Chandra Reddy [Fri, 16 Jun 2023 23:23:33 +0000 (23:23 +0000)]
[libc] Remove the requirement of a platform-flush operation in File abstraction.

The libc flush operation is not supposed to trigger a platform level
flush operation. See "Notes" on this Linux man page:
    https://man7.org/linux/man-pages/man3/fflush.3.html

Reviewed By: michaelrj

Differential Revision: https://reviews.llvm.org/D153182

13 months agoReland "[DebugMetadata][DwarfDebug] Support function-local types in lexical block...
Vladislav Dzhidzhoev [Mon, 19 Jun 2023 14:42:05 +0000 (16:42 +0200)]
Reland "[DebugMetadata][DwarfDebug] Support function-local types in lexical block scopes (4/7)"

Test "local-type-as-template-parameter.ll" now requires linux-system.

Authored-by: Kristina Bessonova <kbessonova@accesssoftek.com>
Differential Revision: https://reviews.llvm.org/D144006

Depends on D144005

13 months ago[libc++] Add missing 'return 0' from main functions in tests
Louis Dionne [Mon, 19 Jun 2023 17:43:17 +0000 (13:43 -0400)]
[libc++] Add missing 'return 0' from main functions in tests

13 months ago[SpecialCaseList] Remove TrigramIndex
Ellis Hoag [Mon, 19 Jun 2023 17:39:13 +0000 (10:39 -0700)]
[SpecialCaseList] Remove TrigramIndex

`TrigramIndex` was added back in https://reviews.llvm.org/D27188 as an optimization to make `SpecialCaseList::match()` faster. I've found that `TrigramIndex` actually makes the function slower and it has no functional use, so we can remove it.

I grabbed the list of queries passed to `SpecialCaseList::match()` on a random very large file (`AArch64ISelLowering.cpp`) and measured the runtime to call `match()` on all of them with [this line](https://github.com/llvm/llvm-project/blob/8e1f820bb4eadf5c0704818f6063e0db1006e32d/llvm/lib/Support/SpecialCaseList.cpp#L64) disabled and then enabled.

```
$ hyperfine --warmup 3 'GTEST_FILTER="SpecialCaseListTest.Large" USE_TRIGRAMS=1 build/unittests/Support/SupportTests' 'GTEST_FILTER="SpecialCaseListTest.Large" USE_TRIGRAMS=0 build/unittests/Support/SupportTests'
Benchmark 1: GTEST_FILTER="SpecialCaseListTest.Large" USE_TRIGRAMS=1 build/unittests/Support/SupportTests
  Time (mean ± σ):     575.9 ms ±  20.3 ms    [User: 573.1 ms, System: 2.7 ms]
  Range (min … max):   555.5 ms … 620.0 ms    10 runs

Benchmark 2: GTEST_FILTER="SpecialCaseListTest.Large" USE_TRIGRAMS=0 build/unittests/Support/SupportTests
  Time (mean ± σ):     283.4 ms ±   6.7 ms    [User: 280.3 ms, System: 3.0 ms]
  Range (min … max):   277.0 ms … 294.9 ms    10 runs

Summary
  'GTEST_FILTER="SpecialCaseListTest.Large" USE_TRIGRAMS=0 build/unittests/Support/SupportTests' ran
    2.03 ± 0.09 times faster than 'GTEST_FILTER="SpecialCaseListTest.Large" USE_TRIGRAMS=1 build/unittests/Support/SupportTests'
```

Using `perf` I found that most of the runtime in `TrigramIndex::isDefinitelyOut()` comes from a division operation that seems to come from `std::unordered_map`: https://github.com/llvm/llvm-project/blob/8e1f820bb4eadf5c0704818f6063e0db1006e32d/llvm/include/llvm/Support/TrigramIndex.h#L62

Removing `TrigramIndex` will make it easier to potentially switch to using `GlobPattern` instead of a full regex for `SpecialCaseList`. See discussion in https://reviews.llvm.org/D152762 for details.

Reviewed By: MaskRay, #sanitizers, vitalybuka

Differential Revision: https://reviews.llvm.org/D153171

13 months ago[AIX][TLS] Generate 64-bit local-exec access code sequence
Amy Kwan [Sat, 17 Jun 2023 05:33:38 +0000 (00:33 -0500)]
[AIX][TLS] Generate 64-bit local-exec access code sequence

This patch adds support for the TLS local-exec access model on AIX to allow
for the ability to generate the 64-bit (specifically, non-optimized) code sequence.

For this patch in particular, the sequence that is generated involves a load of the
variable offset, followed by an add of the loaded variable offset to r13 (which is
thread pointer, respectively). This code sequence looks like the following:
```
ld reg1,var[TC](2)
add reg2, reg1, r13     // r13 contains the thread pointer
```
The TOC (.tc pseudo-op) entries generated in the assembly files are also
changed where we add the @le relocation for the variable offset.

Differential Revision: https://reviews.llvm.org/D149722

13 months ago[AIX][TLS] Relax front end diagnostics to accept the local-exec TLS model
Amy Kwan [Sat, 17 Jun 2023 05:31:35 +0000 (00:31 -0500)]
[AIX][TLS] Relax front end diagnostics to accept the local-exec TLS model

This patch relaxes the front end AIX diagnostics added in D102070 to accept the
local-exec TLS model, as we plan to support this model in a series of future patches.

The diagnostics are relaxed when local-exec is used as a compiler option to
`-ftls-model=*` and in the `__attribute__((tls_model("local-exec")))` attribute.

Differential Revision: https://reviews.llvm.org/D149596

13 months agoRevert "[DebugMetadata][DwarfDebug] Support function-local types in lexical block...
Vladislav Dzhidzhoev [Mon, 19 Jun 2023 17:16:13 +0000 (19:16 +0200)]
Revert "[DebugMetadata][DwarfDebug] Support function-local types in lexical block scopes (4/7)"

This reverts commit 66511b401042f28c74d2ded3aac76d19a53bd7c4.
llvm/test/DebugInfo/Generic/local-type-as-template-parameter.ll is
broken.

13 months agoRecommit "[LSR] Consider post-inc form when creating extends/truncates."
Florian Hahn [Mon, 19 Jun 2023 16:57:05 +0000 (17:57 +0100)]
Recommit "[LSR] Consider post-inc form when creating extends/truncates."

This reverts the revert commit 1797ab36efc9c90c921cd725831f8c3f6a7125a2.

The recommitted version now checks the PostIncLoopSets for all fixups
and returns nullptr if the result doesn't match for all fixups.

13 months ago[AMDGPU] Add basic support for extended i8 perm matching
Jeffrey Byrnes [Mon, 8 May 2023 18:58:14 +0000 (11:58 -0700)]
[AMDGPU] Add basic support for extended i8 perm matching

Differential Revision: https://reviews.llvm.org/D142782

Change-Id: Ibb95224f7885839e8b77a705f487f10b47a258a6

13 months ago[mlir] Fix a rare use-after free in dialect loading
Benjamin Kramer [Mon, 19 Jun 2023 16:18:07 +0000 (18:18 +0200)]
[mlir] Fix a rare use-after free in dialect loading

applyExtensions can load further dialects, invalidating the reference to
the dialect pointer in the dialects DenseMap. Capture the pointer to
prevent that from happening.

13 months ago[gn build] Port eb7491769a51
LLVM GN Syncbot [Mon, 19 Jun 2023 16:13:28 +0000 (16:13 +0000)]
[gn build] Port eb7491769a51

13 months ago[AMDGPU] Reimplement the GFX11 early release VGPRs optimization
Jay Foad [Mon, 19 Jun 2023 14:39:45 +0000 (15:39 +0100)]
[AMDGPU] Reimplement the GFX11 early release VGPRs optimization

Implement this optimization in SIInsertWaitcnts, where we already have
information about whether there might be outstanding VMEM store
instructions. This has the following advantages:
- Correctly handles atomics-with-return.
- Correctly handles call instructions.
- Should be faster because it does not require running a separate pass.

Differential Revision: https://reviews.llvm.org/D153279

13 months ago[mlir][sparse][gpu] remove tuple as one of the spmm_buffer_size output type
Kun Wu [Mon, 19 Jun 2023 15:57:39 +0000 (15:57 +0000)]
[mlir][sparse][gpu] remove tuple as one of the spmm_buffer_size output type

Reviewed By: aartbik

Differential Revision: https://reviews.llvm.org/D153188

13 months ago[DAGCombiner][NFC] Factor out ByteProvider
Jeffrey Byrnes [Mon, 8 May 2023 18:43:15 +0000 (11:43 -0700)]
[DAGCombiner][NFC] Factor out ByteProvider

Differential Revision: https://reviews.llvm.org/D143018

Change-Id: I3dc03787a3382c0c3fe6b869f869c2946f450874

13 months ago[libc++][NFC] Sort header list in header_information.py
Louis Dionne [Mon, 19 Jun 2023 15:53:53 +0000 (11:53 -0400)]
[libc++][NFC] Sort header list in header_information.py

13 months agoAMDGPU: Remove amdgpu-waves-per-eu support in old attribute pass
Matt Arsenault [Fri, 10 Dec 2021 23:48:16 +0000 (18:48 -0500)]
AMDGPU: Remove amdgpu-waves-per-eu support in old attribute pass

AMDGPUAttributor now handles this attribute with value merging, so
delete the old approach which could only apply this to functions which
did not set it, or cloned the function.

13 months agoclang: Add __builtin_elementwise_round
Matt Arsenault [Thu, 5 Jan 2023 19:50:11 +0000 (14:50 -0500)]
clang: Add __builtin_elementwise_round

13 months agoValueTracking: Handle compare to nan and -inf constants in fcmpToClassTest
Matt Arsenault [Wed, 24 May 2023 11:47:57 +0000 (12:47 +0100)]
ValueTracking: Handle compare to nan and -inf constants in fcmpToClassTest

This will help enable a cleanup of simplifyFCmpInst

13 months ago[Hexagon] Fix range checks for immediate operands
Krzysztof Parzyszek [Mon, 19 Jun 2023 15:18:05 +0000 (08:18 -0700)]
[Hexagon] Fix range checks for immediate operands

The output assembly (textual) contains the instruction
  r29 = add(r29,#4294967136)
The value 4294967136 is -160 when interpreted as a signed 32-bit
integer, so it fits in the range of the immediate operand without
a constant extender. The range check in HexagonInstrInfo was putting
the operand value into an int variable, reporting no need for an
extender. This resulted in a packet with 4 instructions, including
the "add". The corresponding check in HexagonMCInstrInfo was using
an int64_t variable, causing the range check to fail, and an extender
to be emitted when lowering to MCInst, resulting in a packet with
too many instructions.