c - Intel SSE: Why does `_mm_extract_ps` return `int` instead of `float`?

Question

Why does _mm_extract_ps return an int instead of a float?

What's the proper way to read a single float from an XMM register in C?

Or rather, a different way to ask it is: What's the opposite of the _mm_set_ps instruction?

Answer 1

None of the answers appear to actually answer the question, why does it return int.

The reason is, the extractps instruction actually copies a component of the vector to a general register. It does seem pretty silly for it to return an int but that's what's actually happening - the raw floating point value ends up in a general register (which hold integers).

If your compiler is configured to generate SSE for all floating point operations, then the closest thing to "extracting" a value to a register would be to shuffle the value into the low component of the vector, then cast it to a scalar float. This should cause that component of the vector to remain in an SSE register:

/* returns the second component of the vector */
float foo(__m128 b)
{
    return _mm_cvtss_f32(_mm_shuffle_ps(b, b, _MM_SHUFFLE(0, 0, 0, 2)));
}

The _mm_cvtss_f32 intrinsic is free, it does not generate instructions, it only makes the compiler reinterpret the xmm register as a float so it can be returned as such.

The _mm_shuffle_ps gets the desired value into the lowest component. The _MM_SHUFFLE macro generates an immediate operand for the resulting shufps instruction.

The 2 in the example gets the float from bit 95:64 of the 127:0 register (the 3rd 32 bit component from the beginning, in memory order) and places it in the 31:0 component of the register (the beginning, in memory order).

The resulting generated code will most likely return the value naturally in a register, like any other floating point value return, with no inefficient writing out to memory and reading it back.

If you're generating code that uses the x87 FPU for floating point (for normal C code that isn't SSE optimized), this would probably result in inefficient code being generated - the compiler would probably store out the component of the SSE vector then use fld to read it back into the x87 register stack. In general 64-bit platforms don't use x87 (they use SSE for all floating point, mostly scalar instructions unless the compiler is vectorizing).

I should add that I always use C++, so I'm not sure whether it is more efficient to pass __m128 by value or by pointer in C. In C++ I would use a const __m128 & and this kind of code would be in a header, so the compiler can inline.