Error: no arguments to '_mm_malloc' or '_mm_free' that depend on a template parameter

[Millani]

Hi there,

Using g++ (version 7) in Ubuntu 18.04, I tried to compile a single simple Levenshtein distance calculation with:

#include <algorithm>
#include <string>
#include "levenshtein-sse.hpp"
#include <iostream>

int main() {
    std::cout << levenshteinSSE::levenshtein("baaad", "bad") <<std::endl;
    return 0;
}

But I get the following error:

levenshtein-sse.hpp: In member function ‘T* levenshteinSSE::AlignmentAllocator<T, N>::allocate(levenshteinSSE::AlignmentAllocator<T, N>::size_type)’:
levenshtein-sse.hpp:102:62: error: there are no arguments to ‘_mm_malloc’ that depend on a template parameter, so a declaration of ‘_mm_malloc’ must be available [-fpermissive]
     return reinterpret_cast<pointer>(reinterpret_cast<char*>(_mm_malloc(n * sizeof(value_type) + 8 * N, N)) + 4*N);
                                                              ^~~~~~~~~~
levenshtein-sse.hpp:102:62: note: (if you use ‘-fpermissive’, G++ will accept your code, but allowing the use of an undeclared name is deprecated)
levenshtein-sse.hpp: In member function ‘void levenshteinSSE::AlignmentAllocator<T, N>::deallocate(levenshteinSSE::AlignmentAllocator<T, N>::pointer, levenshteinSSE::AlignmentAllocator<T, N>::size_type)’:
levenshtein-sse.hpp:106:5: error: there are no arguments to ‘_mm_free’ that depend on a template parameter, so a declaration of ‘_mm_free’ must be available [-fpermissive]
     _mm_free(reinterpret_cast<char*>(p) - 4*N);

Indeed, compiling with -fpermissive bypasses the issue and gets the code compiled and the distance calculation performed. However, in the real application setup I am working with, I cannot leave the -fpermissive flag enabled.

Hence, my question: could you please give me any hint at how to solve that error? If it helps to know, my CPU does have SSE2 and AVX instructions enabled.

Thanks!

[addaleax]

Which compiler flags are you using here?

[Millani]

I actually tried with many combinations, and the results were always the same: it only worked with -fpermissive enabled. Are there any flags in specific that you want me to avoid and\or try for the purpose of testing?

[addaleax]

Yeah, I’m talking about the CPU flags you might have enabled, i.e. -msomething – in particular, do you compile for at least SSSE3?

Does it help to add #include <xmmintrin.h> (or #include <mm_malloc.h>) to the list of headers included by levenshtein-sse.hpp?