https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80286
--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Yet another testcase, showing that we can generate very inefficient code:
typedef short V __attribute__((vector_size (8 * sizeof (short))));
__attribute__((noinline, noclone)) V
foo (V x, V y)
{
return x << y[0];
}
int
main ()
{
V x = { 1, 2, 3, 4, 5, 6, 7, 8 };
V y = { 5, 6, 7, 8, 9, 10, 11, 12 };
V z = foo (x, y);
V e = { 1 << 5, 2 << 5, 3 << 5, 4 << 5, 5 << 5, 6 << 5, 7 << 5, 8 << 5 };
if (__builtin_memcmp (&z, &e, sizeof (V)))
__builtin_abort ();
return 0;
}
vpextrw $0, %xmm1, %eax
cwtl
movl %eax, -12(%rsp)
vmovd -12(%rsp), %xmm2
vpsllw %xmm2, %xmm0, %xmm0
For SSE4.1+, we could as well emit vpmovsxwq for the sign-extension followed
immediately by vpsllw. For say SSE2 at least for the unsigned int -> unsigned
long zero extension we could use pxor (to get 0) and punpck*.
