https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65832
--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> --- Another example where the vectorizer thinks vectorization is profitable: #define N 16 unsigned int out[N]; unsigned int in[N] = {0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15}; __attribute__ ((noinline)) int main1 (unsigned int x, unsigned int y) { int i; unsigned int a0, a1, a2, a3; a0 = in[0]; a1 = in[1]; a2 = in[2]; a3 = in[3]; out[0] = a0 * x; out[1] = a1 * y; out[2] = a2 * x; out[3] = a3 * y; } generates main1: .LFB0: .cfi_startproc movl %edi, -12(%rsp) movd -12(%rsp), %xmm0 movl %esi, -12(%rsp) movd -12(%rsp), %xmm3 movdqa in(%rip), %xmm2 punpckldq %xmm3, %xmm0 psrlq $32, %xmm2 punpcklqdq %xmm0, %xmm0 movdqa %xmm0, %xmm1 psrlq $32, %xmm0 pmuludq %xmm2, %xmm0 pshufd $8, %xmm0, %xmm0 pmuludq in(%rip), %xmm1 pshufd $8, %xmm1, %xmm1 punpckldq %xmm0, %xmm1 movaps %xmm1, out(%rip) ret slightly less obfuscated when we allow gpr xmm moves with -mtune=intel: main1: .LFB0: .cfi_startproc movd %edi, %xmm0 movd %esi, %xmm3 movdqa in(%rip), %xmm2 punpckldq %xmm3, %xmm0 punpcklqdq %xmm0, %xmm0 psrlq $32, %xmm2 movdqa %xmm0, %xmm1 psrlq $32, %xmm0 pmuludq in(%rip), %xmm1 pmuludq %xmm2, %xmm0 pshufd $8, %xmm1, %xmm1 pshufd $8, %xmm0, %xmm0 punpckldq %xmm0, %xmm1 movdqa %xmm1, out(%rip) ret so for { x, y, x, y } construction we generate movd %edi, %xmm0 movd %esi, %xmm3 punpckldq %xmm3, %xmm0 punpcklqdq %xmm0, %xmm0 no f*** idea where all the shifting and shuffling comes from... This is just vect_cst_.7_18 = {x_6(D), y_9(D), x_6(D), y_9(D)}; vect_a0_2.5_17 = MEM[(unsigned int *)&in]; vect__7.6_19 = vect_a0_2.5_17 * vect_cst_.7_18; MEM[(unsigned int *)&out] = vect__7.6_19;