Does _MM_TRANSPOSE4_PS cause compiler errors in GCC?

Question

I collect my math library in GCC instead of MSVC for the first time and look through all the small errors, and I hit one that just doesn't make sense:

Line 284: error: lvalue required as left operand of assignment

What's on line 284? this is:

_MM_TRANSPOSE4_PS(r, u, t, _mm_setr_ps(0.0f, 0.0f, 0.0f, 1.0f));

(r, u and t are all instances of __m128 )

Those familiar with using xmmintrin.h will know that _MM_TRANSPOSE4_PS is actually not a function, but a macro that extends to:

 /* Transpose the 4x4 matrix composed of row[0-3]. */ #define _MM_TRANSPOSE4_PS(row0, row1, row2, row3) \ do { \ __v4sf __r0 = (row0), __r1 = (row1), __r2 = (row2), __r3 = (row3); \ __v4sf __t0 = __builtin_ia32_unpcklps (__r0, __r1); \ __v4sf __t1 = __builtin_ia32_unpcklps (__r2, __r3); \ __v4sf __t2 = __builtin_ia32_unpckhps (__r0, __r1); \ __v4sf __t3 = __builtin_ia32_unpckhps (__r2, __r3); \ (row0) = __builtin_ia32_movlhps (__t0, __t1); \ (row1) = __builtin_ia32_movhlps (__t1, __t0); \ (row2) = __builtin_ia32_movlhps (__t2, __t3); \ (row3) = __builtin_ia32_movhlps (__t3, __t2); \ } while (0) 

So ... what causes my compiler errors? I do not redefine anything here that I know of. This exact code compiled and worked perfectly when I used MSVC.

Answer 1

You need to change:

 _MM_TRANSPOSE4_PS(r, u, t, _mm_setr_ps(0.0f, 0.0f, 0.0f, 1.0f)); 

in

 __m128 v = _mm_setr_ps(0.0f, 0.0f, 0.0f, 1.0f); _MM_TRANSPOSE4_PS(r, u, t, v); 

since this is transposition in place, and 4 input vectors are also used for output.

Answer 2

MSVC uses its own definition:

 #define _MM_TRANSPOSE4_PS(row0, row1, row2, row3) { \ __m128 tmp3, tmp2, tmp1, tmp0; \ \ tmp0 = _mm_shuffle_ps((row0), (row1), 0x44); \ tmp2 = _mm_shuffle_ps((row0), (row1), 0xEE); \ tmp1 = _mm_shuffle_ps((row2), (row3), 0x44); \ tmp3 = _mm_shuffle_ps((row2), (row3), 0xEE); \ \ (row0) = _mm_shuffle_ps(tmp0, tmp1, 0x88); \ (row1) = _mm_shuffle_ps(tmp0, tmp1, 0xDD); \ (row2) = _mm_shuffle_ps(tmp2, tmp3, 0x88); \ (row3) = _mm_shuffle_ps(tmp2, tmp3, 0xDD); \ } 

The last line is converted to _mm_setr_ps(0.0f, 0.0f, 0.0f, 1.0f) = _mm_shuffle_ps(tmp2,tmp3, 0XDD); , which compiles in MSVC just fine, but with a lvalue error in GCC. I am not sure why MSVC allows this.

I looked at the assembly of this code in MSVC2013

 #include <immintrin.h> #include <stdio.h> int main() { __m128 rows[4]; //rows[0] = _mm_setr_ps( 1, 2, 3, 4); //rows[1] = _mm_setr_ps( 5, 6, 7, 8); rows[2] = _mm_setr_ps( 9,10,11,12); rows[3] = _mm_setr_ps(13,14,15,16); //_MM_TRANSPOSE4_PS(rows[0],rows[1],rows[2],rows[3]); //_MM_TRANSPOSE4_PS(rows[0],rows[1],rows[2],_mm_setr_ps(0.0f, 0.0f, 0.0f, 1.0f)); rows[2] = _mm_shuffle_ps(rows[2], rows[3], 0x88); _mm_setr_ps(0.0f, 0.0f, 0.0f, 1.0f) = _mm_shuffle_ps(rows[2],rows[3], 0XDD); } 

Here is the corresponding build code

 ; Line 14 mov eax, 16 imul rax, 3 mov ecx, 16 imul rcx, 2 movups xmm0, XMMWORD PTR rows$[rsp+rcx] shufps xmm0, XMMWORD PTR rows$[rsp+rax], 136 ; 00000088H movaps XMMWORD PTR $T6[rsp], xmm0 mov eax, 16 imul rax, 2 movaps xmm0, XMMWORD PTR $T6[rsp] movups XMMWORD PTR rows$[rsp+rax], xmm0 ; Line 15 mov eax, 16 imul rax, 3 mov ecx, 16 imul rcx, 2 movups xmm0, XMMWORD PTR rows$[rsp+rcx] shufps xmm0, XMMWORD PTR rows$[rsp+rax], 221 ; 000000ddH movaps XMMWORD PTR $T8[rsp], xmm0 movaps xmm0, XMMWORD PTR __xmm@3f800000000000000000000000000000 movaps XMMWORD PTR $T7[rsp], xmm0 movaps xmm0, XMMWORD PTR $T8[rsp] movaps XMMWORD PTR $T7[rsp], xmm0