(this is Linux, the same happens on Darwin).
This is not really a good idea, as movsd of a double-precision 1.0 is
faster.
I wonder whether fixing compress_float_constant is better. It seems
similar to the old hack in expr.c: expanding a/b to a*(1/b), and hope
that it is transformed back to a/b or CSE-ed.
Also, here is some code that tries to load 1.0 into an SSE register
without using the constant pool. I don't know if it is fast or not.
float f = 1, g;
double h = 1, i;
void main ()
{
/* Generates movl + movd */
asm volatile
("movd %1, %0" : "=x" (g) : "r" (0x3f800000));
printf ("%x %f %f\n", *(unsigned int *) &f, f, g);
/* Generates movl + xorps + pinsrwl */
asm volatile
("xorps %0, %0;\
pinsrwl $3, %1, %0" : "=x" (i) : "r" (0x3ff0));
printf ("%llx %llx %f %f\n", *(unsigned long long *) &h,
*(unsigned long long *) &i, h, i);
}
Paolo
