AVX types (`__m256{,i,d}`) should be castable to each other

[btolsch]

Version of emscripten/emsdk:

emcc (Emscripten gcc/clang-like replacement + linker emulating GNU ld) 5.0.1-git (8c5f43157a3f069ade75876e23061330521eabde)
clang version 23.0.0git (/startdir/llvm-project b447f5d9763010f8c6806c578533291aef2bd484)
Target: wasm32-unknown-emscripten
Thread model: posix
InstalledDir: /opt/emscripten-llvm/bin

Failing command line in full:
For the file test.c:

#include <immintrin.h>

int
main(int argc, char** argv) {
  __m256i data, mask;
  data = _mm256_setzero_si256();
  mask = _mm256_setzero_si256();
  data = (__m256i)_mm256_and_ps((__m256)data, (__m256)mask);
  return 0;
}

$ emcc -msimd128 -mavx test.c
test.c:8:33: error: used type '__m256' where arithmetic or pointer type is required
    8 |   data = (__m256i)_mm256_and_ps((__m256)data, (__m256)mask);
      |                                 ^       ~~~~
test.c:8:47: error: used type '__m256' where arithmetic or pointer type is required
    8 |   data = (__m256i)_mm256_and_ps((__m256)data, (__m256)mask);
      |                                               ^       ~~~~
2 errors generated.
emcc: error: '/opt/emscripten-llvm/bin/clang -target wasm32-unknown-emscripten -fignore-exceptions -mllvm -combiner-global-alias-analysis=false -mllvm -enable-emscripten-sjlj -mllvm -disable-lsr
 --sysroot=/home/btolsch/.cache/emscripten/sysroot -DEMSCRIPTEN -D__SSE__=1 -D__SSE2__=1 -D__SSE3__=1 -D__SSSE3__=1 -D__SSE4_1__=1 -D__SSE4_2__=1 -D__AVX__=1 -Xclang -iwithsysroot/include/fakesd
l -Xclang -iwithsysroot/include/compat -msimd128 -c test.c -o /tmp/emscripten_temp_xnre9uu1/test.o' failed (returned 1)

Desired behavior
Both clang and gcc allow casting between __m256, __m256i, and __m256d. This is because they are declared as vector types (i.e. __attribute__((__vector_size(n)))) and not structs or unions. The emscripten definitions are structs like:

typedef struct {
  __m128d v0;
  __m128d v1;
} __m256d;

These can't be cast to each other, so instead you have to do something like *(__m256*)&data or memcpy to a new variable (which is hopefully optimized out). Moreover, the __m128* definitions are vectors, so this problem is specific to the AVX types.

The struct definitions are convenient for the implementation, which entirely falls back to 128-bit operations, but can break existing AVX code which has no SSE or scalar fallback of its own.