GCC 14 build failure on ARM/NEON: `FS::Cast<int32_t>(Mask<32>)` cannot convert `uint32x4_t` to `Register<int, 4, ...>`

[a12e]

Description

Building FastNoise2 with the FastSIMD ARM backends enabled fails with GCC 14 when compiling the generated FastSIMD_FastNoise_AARCH64.cpp and FastSIMD_FastNoise_NEON.cpp translation units.

The error appears to come from FastSIMD/ToolSet/ARM/NEON.h, where FS::Cast<int32_t>(...) on a Mask<32> returns a native uint32x4_t, but GCC 14 does not implicitly convert that to FS::Register<int, 4, ...>.

Environment

• OS: Linux
• Compiler: GCC / G++ 14
• C++ standard: GNU++17
• Build type: Release
• Relevant flags:
  • -O2
  • -DNDEBUG
  • -flto=auto
  • -ffast-math
  • -DFASTSIMD_IS_RELAXED=1
  • -DFASTSIMD_STATIC_LIB
• Target/features involved:
  • FastSIMD::FeatureSet::AARCH64
  • FastSIMD::FeatureSet::NEON

Actual result

The build fails while compiling both ARM SIMD backends:

• FastSIMD_FastNoise_AARCH64.cpp
• FastSIMD_FastNoise_NEON.cpp

FAILED: src/CMakeFiles/FastSIMD_FastNoise.dir/fastsimd/FastSIMD_FastNoise/FastSIMD_FastNoise_AARCH64.cpp.o

<compiler> \
  -DFASTSIMD_EXPORT \
  -DFASTSIMD_IS_RELAXED=1 \
  -DFASTSIMD_LIBRARY_NAME=FastSIMD_FastNoise \
  -DFASTSIMD_STATIC_LIB \
  -I<build-dir>/src/fastsimd/FastSIMD_FastNoise/include \
  -I<source-dir>/include \
  -I<build-dir>/_deps/fastsimd-src/include \
  -O2 -g -DNDEBUG \
  -flto=auto \
  -Wno-ignored-attributes \
  -Wno-nan-infinity-disabled \
  -ffast-math \
  -std=gnu++17 \
  -c <build-dir>/src/fastsimd/FastSIMD_FastNoise/FastSIMD_FastNoise_AARCH64.cpp

In file included from <build-dir>/_deps/fastsimd-src/include/FastSIMD/ToolSet/ARM/ARM.h:4,
                 from <build-dir>/_deps/fastsimd-src/include/FastSIMD/ToolSet.h:31,
                 from <build-dir>/_deps/fastsimd-src/dispatch/impl/DispatchClassImpl.h:2,
                 from <build-dir>/src/fastsimd/FastSIMD_FastNoise/FastSIMD_FastNoise_AARCH64.cpp:7:

<build-dir>/_deps/fastsimd-src/include/FastSIMD/ToolSet/ARM/NEON.h:
In instantiation of
'FS::Register<U, 4, SIMD> FS::Cast(const Register<T, 4, SIMD>&, TypeDummy<U>)
 [with U = int;
       T = Mask<32, false>;
       FastSIMD::FeatureSet SIMD = FastSIMD::FeatureSet::AARCH64;
       <template-parameter-1-4> = __Uint32x4_t]':

<build-dir>/_deps/fastsimd-src/include/FastSIMD/ToolSet/Generic/Functions.h:121:42:
required from
'FS::Register<T, N, SIMD> FS::impl::Cast(const FS::Register<U, N, SIMD>&, FS::TypeDummy<U>)
 [with U = int;
       T = FS::Mask<32, false>;
       long unsigned int N = 8;
       FastSIMD::FeatureSet SIMD = FastSIMD::FeatureSet::AARCH64]'

<source-dir>/include/FastNoise/Generators/DomainWarpSimplex.inl:569:47:
required from
'void FastNoise::ApplyOrthogonalGradientMatrixVectorProductSimplex(...)
 [with FastSIMD::FeatureSet SIMD = FastSIMD::FeatureSet::AARCH64;
       int32v = FS::Register<int, 8>;
       float32v = FS::Register<float, 8>]'

<source-dir>/include/FastNoise/Generators/DomainWarpSimplex.inl:569:47:
    indexPositive |= FS::Cast<int32_t>( singleNonExtraCase );
                     ~~~~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~

<build-dir>/_deps/fastsimd-src/include/FastSIMD/ToolSet/ARM/NEON.h:39:31:
error: could not convert
'(& a)->FS::Register<FS::Mask<32, false>, 4, FastSIMD::FeatureSet::AARCH64, void>::GetNative()'
from
'FS::Register<FS::Mask<32, false>, 4, FastSIMD::FeatureSet::AARCH64, void>::NativeType'
{aka 'uint32x4_t'}
to
'FS::Register<int, 4, FastSIMD::FeatureSet::AARCH64, void>'

   39 |             return a.GetNative();
      |                    ~~~~~~~~~~~^~
      |                               |
      |                               uint32x4_t

The same error also occurs for the NEON backend:

FAILED: src/CMakeFiles/FastSIMD_FastNoise.dir/fastsimd/FastSIMD_FastNoise/FastSIMD_FastNoise_NEON.cpp.o

<compiler> \
  -DFASTSIMD_EXPORT \
  -DFASTSIMD_IS_RELAXED=1 \
  -DFASTSIMD_LIBRARY_NAME=FastSIMD_FastNoise \
  -DFASTSIMD_STATIC_LIB \
  -I<build-dir>/src/fastsimd/FastSIMD_FastNoise/include \
  -I<source-dir>/include \
  -I<build-dir>/_deps/fastsimd-src/include \
  -O2 -g -DNDEBUG \
  -flto=auto \
  -Wno-ignored-attributes \
  -Wno-nan-infinity-disabled \
  -ffast-math \
  -std=gnu++17 \
  -c <build-dir>/src/fastsimd/FastSIMD_FastNoise/FastSIMD_FastNoise_NEON.cpp

In file included from <build-dir>/_deps/fastsimd-src/include/FastSIMD/ToolSet/ARM/ARM.h:4,
                 from <build-dir>/_deps/fastsimd-src/include/FastSIMD/ToolSet.h:31,
                 from <build-dir>/_deps/fastsimd-src/dispatch/impl/DispatchClassImpl.h:2,
                 from <build-dir>/src/fastsimd/FastSIMD_FastNoise/FastSIMD_FastNoise_NEON.cpp:7:

<build-dir>/_deps/fastsimd-src/include/FastSIMD/ToolSet/ARM/NEON.h:
In instantiation of
'FS::Register<U, 4, SIMD> FS::Cast(const Register<T, 4, SIMD>&, TypeDummy<U>)
 [with U = int;
       T = Mask<32, false>;
       FastSIMD::FeatureSet SIMD = FastSIMD::FeatureSet::NEON;
       <template-parameter-1-4> = __Uint32x4_t]':

<source-dir>/include/FastNoise/Generators/DomainWarpSimplex.inl:569:47:
required from
'void FastNoise::ApplyOrthogonalGradientMatrixVectorProductSimplex(...)
 [with FastSIMD::FeatureSet SIMD = FastSIMD::FeatureSet::NEON;
       int32v = FS::Register<int, 8>;
       float32v = FS::Register<float, 8>]'

<source-dir>/include/FastNoise/Generators/DomainWarpSimplex.inl:569:47:
    indexPositive |= FS::Cast<int32_t>( singleNonExtraCase );
                     ~~~~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~

<build-dir>/_deps/fastsimd-src/include/FastSIMD/ToolSet/ARM/NEON.h:39:31:
error: could not convert
'(& a)->FS::Register<FS::Mask<32, false>, 4, FastSIMD::FeatureSet::NEON, void>::GetNative()'
from
'FS::Register<FS::Mask<32, false>, 4, FastSIMD::FeatureSet::NEON, void>::NativeType'
{aka 'uint32x4_t'}
to
'FS::Register<int, 4, FastSIMD::FeatureSet::NEON, void>'

   39 |             return a.GetNative();
      |                    ~~~~~~~~~~~^~
      |                               |
      |                               uint32x4_t

A similar failure also occurs for Mask<32> without the false template parameter:

<build-dir>/_deps/fastsimd-src/include/FastSIMD/ToolSet/ARM/NEON.h:39:31:
error: could not convert
'(& a)->FS::Register<FS::Mask<32>, 4, FastSIMD::FeatureSet::NEON, void>::GetNative()'
from
'FS::Register<FS::Mask<32>, 4, FastSIMD::FeatureSet::NEON, void>::NativeType'
{aka 'uint32x4_t'}
to
'FS::Register<int, 4, FastSIMD::FeatureSet::NEON, void>'

There is also this compiler note, which may be unrelated:

cc1plus: note: unrecognized command-line option '-Wno-nan-infinity-disabled' may have been intended to silence earlier diagnostics

Expected result

The ARM/NEON FastSIMD backends should compile successfully with GCC 14.

Suspected cause

The problematic code seems to be in:

FastSIMD/ToolSet/ARM/NEON.h:39

where the implementation returns:

return a.GetNative();

For Mask<32> / Mask<32, false>, GetNative() returns uint32x4_t, but the function return type is:

FS::Register<int, 4, SIMD>

GCC 14 rejects the implicit conversion from the unsigned native NEON vector type to the signed integer Register.

The failure is triggered through calls such as:

indexPositive |= FS::Cast<int32_t>( singleNonExtraCase );
indexA        |= FS::Cast<int32_t>( equal );

in:

FastNoise/Generators/DomainWarpSimplex.inl