Use pragma(inline, true)?

[p0nce]

Code written with intel-intrinsics is often way slower without optimizations, because of no-inlining. Try to improve that for DMD 32-bit.

[p0nce]

With clang intrinsics use __nodebug__, find something similar. (EDIT: doesnt change build times very much)

[p0nce]

Ideas:

• could be to avoid defining vector types if D_SIMD is here? (EDIT no because of int2)
• Perhaps using v.array[0] instead of v[0] who calls opIndex (EDIT: needed for GDC 32-bit support anyway)
• pragma(inline, true)
• do not call loadUnaligned when emulating SIMD vectors (EDIT: sometimes loadUnaligned produce better code with ARM)

[p0nce]

• D_SIMD is an obvious option too

[p0nce]

• loadUnaligned actually leads to different (often better) optimized code! useful for ARM + LDC

[p0nce]

Close in favour of #42 and #59