[C++] Build failure with ARROW_SIMD_LEVEL=NONE

[lahwaacz]

Describe the bug, including details regarding any error messages, version, and platform.

Building arrow 22.0.0 with the -DARROW_SIMD_LEVEL=NONE CMake option leads to these errors:

FAILED: [code=1] src/arrow/CMakeFiles/arrow_util.dir/util/byte_stream_split_internal_avx2.cc.o
/usr/bin/c++ -DARROW_HAVE_RUNTIME_AVX2 -DARROW_HAVE_RUNTIME_AVX512 -DARROW_HAVE_RUNTIME_BMI2 -DARROW_HAVE_RUNTIME_SSE4_2 -DARROW_WITH_TIMING_TESTS -DGFLAGS_IS_A_DLL=0 -DGLOG_USE_GFLAGS -DGLOG_USE_GLOG_EXPORT -DURI_STATIC_BUILD -I/build/arrow/src/build/src -I/build/arrow/src/apache-arrow-22.0.0/cpp/src -I/build/arrow/src/apache-arrow-22.0.0/cpp/src/generated -Wredundant-move -Wno-noexcept-type -Wno-self-move -march=x86-64 -mtune=generic -O2 -pipe -fno-plt -fexceptions         -Wp,-D_FORTIFY_SOURCE=3 -Wformat -Werror=format-security         -fstack-clash-protection -fcf-protection         -fno-omit-frame-pointer -mno-omit-leaf-frame-pointer -Wp,-D_GLIBCXX_ASSERTIONS -g -ffile-prefix-map=/build/arrow/src=/usr/src/debug/arrow -flto=auto -fdiagnostics-color=always  -Wall -Wno-conversion -Wno-sign-conversion -Wdate-time -Wimplicit-fallthrough -Wunused-result -fno-semantic-interposition  -O2 -g -DNDEBUG -ftree-vectorize -ggdb  -std=c++17 -fPIC -march=haswell -mavx2 -MD -MT src/arrow/CMakeFiles/arrow_util.dir/util/byte_stream_split_internal_avx2.cc.o -MF src/arrow/CMakeFiles/arrow_util.dir/util/byte_stream_split_internal_avx2.cc.o.d -o src/arrow/CMakeFiles/arrow_util.dir/util/byte_stream_split_internal_avx2.cc.o -c /build/arrow/src/apache-arrow-22.0.0/cpp/src/arrow/util/byte_stream_split_internal_avx2.cc
In file included from /usr/include/xsimd/types/xsimd_avx2_register.hpp:15,
                 from /build/arrow/src/apache-arrow-22.0.0/cpp/src/arrow/util/byte_stream_split_internal_avx2.cc:22:
/usr/include/xsimd/types/xsimd_avx_register.hpp:48:63: error: expected constructor, destructor, or type conversion before ‘;’ token
   48 |         XSIMD_DECLARE_SIMD_REGISTER(signed char, avx, __m256i);
      |                                                               ^
/usr/include/xsimd/types/xsimd_avx_register.hpp:49:65: error: expected constructor, destructor, or type conversion before ‘;’ token
   49 |         XSIMD_DECLARE_SIMD_REGISTER(unsigned char, avx, __m256i);
      |                                                                 ^
/usr/include/xsimd/types/xsimd_avx_register.hpp:50:56: error: expected constructor, destructor, or type conversion before ‘;’ token
   50 |         XSIMD_DECLARE_SIMD_REGISTER(char, avx, __m256i);
      |                                                        ^
/usr/include/xsimd/types/xsimd_avx_register.hpp:51:66: error: expected constructor, destructor, or type conversion before ‘;’ token
   51 |         XSIMD_DECLARE_SIMD_REGISTER(unsigned short, avx, __m256i);
      |                                                                  ^
/usr/include/xsimd/types/xsimd_avx_register.hpp:52:57: error: expected constructor, destructor, or type conversion before ‘;’ token
   52 |         XSIMD_DECLARE_SIMD_REGISTER(short, avx, __m256i);
      |                                                         ^
/usr/include/xsimd/types/xsimd_avx_register.hpp:53:64: error: expected constructor, destructor, or type conversion before ‘;’ token
   53 |         XSIMD_DECLARE_SIMD_REGISTER(unsigned int, avx, __m256i);
      |                                                                ^
/usr/include/xsimd/types/xsimd_avx_register.hpp:54:55: error: expected constructor, destructor, or type conversion before ‘;’ token
   54 |         XSIMD_DECLARE_SIMD_REGISTER(int, avx, __m256i);
      |                                                       ^
/usr/include/xsimd/types/xsimd_avx_register.hpp:55:69: error: expected constructor, destructor, or type conversion before ‘;’ token
   55 |         XSIMD_DECLARE_SIMD_REGISTER(unsigned long int, avx, __m256i);
      |                                                                     ^
/usr/include/xsimd/types/xsimd_avx_register.hpp:56:60: error: expected constructor, destructor, or type conversion before ‘;’ token
   56 |         XSIMD_DECLARE_SIMD_REGISTER(long int, avx, __m256i);
      |                                                            ^
/usr/include/xsimd/types/xsimd_avx_register.hpp:57:74: error: expected constructor, destructor, or type conversion before ‘;’ token
   57 |         XSIMD_DECLARE_SIMD_REGISTER(unsigned long long int, avx, __m256i);
      |                                                                          ^
/usr/include/xsimd/types/xsimd_avx_register.hpp:58:65: error: expected constructor, destructor, or type conversion before ‘;’ token
   58 |         XSIMD_DECLARE_SIMD_REGISTER(long long int, avx, __m256i);
      |                                                                 ^
/usr/include/xsimd/types/xsimd_avx_register.hpp:59:56: error: expected constructor, destructor, or type conversion before ‘;’ token
   59 |         XSIMD_DECLARE_SIMD_REGISTER(float, avx, __m256);
      |                                                        ^
/usr/include/xsimd/types/xsimd_avx_register.hpp:60:58: error: expected constructor, destructor, or type conversion before ‘;’ token
   60 |         XSIMD_DECLARE_SIMD_REGISTER(double, avx, __m256d);
      |                                                          ^
/usr/include/xsimd/types/xsimd_avx2_register.hpp:39:53: error: expected constructor, destructor, or type conversion before ‘;’ token
   39 |         XSIMD_DECLARE_SIMD_REGISTER_ALIAS(avx2, avx);
      |                                                     ^
/build/arrow/src/apache-arrow-22.0.0/cpp/src/arrow/util/byte_stream_split_internal_avx2.cc:33:15: error: variable or field ‘ByteStreamSplitDecodeSimd’ declared void
   33 | template void ByteStreamSplitDecodeSimd<xsimd::avx2, 2>(const uint8_t*, int, int64_t,
      |               ^~~~~~~~~~~~~~~~~~~~~~~~~
/build/arrow/src/apache-arrow-22.0.0/cpp/src/arrow/util/byte_stream_split_internal_avx2.cc:33:40: error: expected ‘;’ before ‘<’ token
   33 | template void ByteStreamSplitDecodeSimd<xsimd::avx2, 2>(const uint8_t*, int, int64_t,
      |                                        ^
      |                                        ;
/build/arrow/src/apache-arrow-22.0.0/cpp/src/arrow/util/byte_stream_split_internal_avx2.cc:35:15: error: variable or field ‘ByteStreamSplitDecodeSimd’ declared void
   35 | template void ByteStreamSplitDecodeSimd<xsimd::avx2, 4>(const uint8_t*, int, int64_t,
      |               ^~~~~~~~~~~~~~~~~~~~~~~~~
/build/arrow/src/apache-arrow-22.0.0/cpp/src/arrow/util/byte_stream_split_internal_avx2.cc:35:40: error: expected ‘;’ before ‘<’ token
   35 | template void ByteStreamSplitDecodeSimd<xsimd::avx2, 4>(const uint8_t*, int, int64_t,
      |                                        ^
      |                                        ;
/build/arrow/src/apache-arrow-22.0.0/cpp/src/arrow/util/byte_stream_split_internal_avx2.cc:37:15: error: variable or field ‘ByteStreamSplitDecodeSimd’ declared void
   37 | template void ByteStreamSplitDecodeSimd<xsimd::avx2, 8>(const uint8_t*, int, int64_t,
      |               ^~~~~~~~~~~~~~~~~~~~~~~~~
/build/arrow/src/apache-arrow-22.0.0/cpp/src/arrow/util/byte_stream_split_internal_avx2.cc:37:40: error: expected ‘;’ before ‘<’ token
   37 | template void ByteStreamSplitDecodeSimd<xsimd::avx2, 8>(const uint8_t*, int, int64_t,
      |                                        ^
      |                                        ;
/build/arrow/src/apache-arrow-22.0.0/cpp/src/arrow/util/byte_stream_split_internal_avx2.cc: In function ‘void arrow::util::internal::ByteStreamSplitEncodeAvx2Impl4(const uint8_t*, int, int64_t, uint8_t*)’:
/build/arrow/src/apache-arrow-22.0.0/cpp/src/arrow/util/byte_stream_split_internal_avx2.cc:52:12: error: ‘ByteStreamSplitEncodeSimd’ was not declared in this scope; did you mean ‘ByteStreamSplitEncode’?
   52 |     return ByteStreamSplitEncodeSimd<xsimd::sse4_2, kNumStreams>(
      |            ^~~~~~~~~~~~~~~~~~~~~~~~~
      |            ByteStreamSplitEncode
/build/arrow/src/apache-arrow-22.0.0/cpp/src/arrow/util/byte_stream_split_internal_avx2.cc:52:51: error: expected primary-expression before ‘,’ token
   52 |     return ByteStreamSplitEncodeSimd<xsimd::sse4_2, kNumStreams>(
      |                                                   ^
/build/arrow/src/apache-arrow-22.0.0/cpp/src/arrow/util/byte_stream_split_internal_avx2.cc:53:9: warning: left operand of comma operator has no effect [-Wunused-value]
   53 |         raw_values, width, num_values, output_buffer_raw);
      |         ^~~~~~~~~~
/build/arrow/src/apache-arrow-22.0.0/cpp/src/arrow/util/byte_stream_split_internal_avx2.cc:53:28: warning: right operand of comma operator has no effect [-Wunused-value]
   53 |         raw_values, width, num_values, output_buffer_raw);
      |                            ^~~~~~~~~~
/build/arrow/src/apache-arrow-22.0.0/cpp/src/arrow/util/byte_stream_split_internal_avx2.cc:53:40: warning: right operand of comma operator has no effect [-Wunused-value]
   53 |         raw_values, width, num_values, output_buffer_raw);
      |                                        ^~~~~~~~~~~~~~~~~
/build/arrow/src/apache-arrow-22.0.0/cpp/src/arrow/util/byte_stream_split_internal_avx2.cc:52:64: error: ISO C++ forbids comparison between pointer and integer [-fpermissive]
   52 |     return ByteStreamSplitEncodeSimd<xsimd::sse4_2, kNumStreams>(
      |                                                     ~~~~~~~~~~~^~
   53 |         raw_values, width, num_values, output_buffer_raw);
      |         ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/build/arrow/src/apache-arrow-22.0.0/cpp/src/arrow/util/byte_stream_split_internal_avx2.cc: In function ‘void arrow::util::internal::ByteStreamSplitEncodeAvx2(const uint8_t*, int, int64_t, uint8_t*)’:
/build/arrow/src/apache-arrow-22.0.0/cpp/src/arrow/util/byte_stream_split_internal_avx2.cc:128:12: error: ‘ByteStreamSplitEncodeSimd’ was not declared in this scope; did you mean ‘ByteStreamSplitEncodeAvx2’? [-Wtemplate-body]
  128 |     return ByteStreamSplitEncodeSimd<xsimd::avx2, kNumStreams>(
      |            ^~~~~~~~~~~~~~~~~~~~~~~~~
      |            ByteStreamSplitEncodeAvx2
/build/arrow/src/apache-arrow-22.0.0/cpp/src/arrow/util/byte_stream_split_internal_avx2.cc:128:49: error: expected primary-expression before ‘,’ token [-Wtemplate-body]
  128 |     return ByteStreamSplitEncodeSimd<xsimd::avx2, kNumStreams>(
      |                                                 ^
/build/arrow/src/apache-arrow-22.0.0/cpp/src/arrow/util/byte_stream_split_internal_avx2.cc:128:62: error: ISO C++ forbids comparison between pointer and integer [-Wtemplate-body]
  128 |     return ByteStreamSplitEncodeSimd<xsimd::avx2, kNumStreams>(
      |                                                   ~~~~~~~~~~~^~
  129 |         raw_values, width, num_values, output_buffer_raw);
      |         ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

The ByteStreamSplitDecodeSimd and ByteStreamSplitDecodeSimd functions are defined in byte_stream_split_internal.h only if a suitable SIMD level is configured:

arrow/cpp/src/arrow/util/byte_stream_split_internal.h

Lines 32 to 40 in 308b007

	#if defined(ARROW_HAVE_NEON) || defined(ARROW_HAVE_SSE4_2)
	# include <xsimd/xsimd.hpp>
	# define ARROW_HAVE_SIMD_SPLIT
	#endif
	namespace arrow::util::internal {
	#if defined(ARROW_HAVE_SIMD_SPLIT)

But here the functions seem to be used without any macro guards:

arrow/cpp/src/arrow/util/byte_stream_split_internal_avx2.cc

Line 52 in 308b007

return ByteStreamSplitEncodeSimd<xsimd::sse4_2, kNumStreams>(

Component(s)

C++

[kou]

@AntoinePrv Could you take a look at this? #46963 may be related.

[raulcd]

This is the same issue that was found on the Arrow feedstocks issue on conda:
conda-forge/arrow-cpp-feedstock#1875 (comment)

[AntoinePrv]

Yes, I look into it.

From: conda-forge/arrow-cpp-feedstock#1875 (comment)

Yes, seems like a bug to me if the file requires AVX2 at compile time shouldn't only use the append_runtime_avx2_src macro but should also check ARROW_SIMD_LEVEL as the append_runtime_avx2_src macro only validates the ARROW_RUNTIME_SIMD_LEVEL flag.

Should I deduce that ARROW_SIMD_LEVEL=NONE also deactivates runtime SIMD? Or that the AVX2 runtime code should be wraped in #ifdef ARROW_HAVE_RUNTIME_AVX2?

[pitrou]

Or that the AVX2 runtime code should be wraped in #ifdef ARROW_HAVE_RUNTIME_AVX2?

This.

[AntoinePrv]

Or that the AVX2 runtime code should be wraped in #ifdef ARROW_HAVE_RUNTIME_AVX2?

The append_runtime_avx2_src macro already do that apparently (it will remove the file from compilation based on that flag).

[pitrou]

Just try to replace:

#if defined(ARROW_HAVE_NEON) || defined(ARROW_HAVE_SSE4_2)

with:

#if defined(ARROW_HAVE_NEON) || defined(ARROW_HAVE_RUNTIME_SSE4_2)

[AntoinePrv]

You're right that fix works.

I find the whole gymnastic of separate files + macros + CMake macros redundant and error-prone. Wondering if we could come up with a better pattern.

[pitrou]

I agree that a simpler pattern would be nice, but I don't really know how that would work.

For example, we don't want the compiler to implicitly generate AVX2 instructions in the code path for dynamically-dispatched SSE4.2. We could use the target pragma with gcc or clang, but that is compiler-specific.

[pitrou]

Fixed by #47952