1#include <math.h>
 2
 3long double remquol(long double x, long double y, int *quo)
 4{
 5	signed char *cx = (void *)&x, *cy = (void *)&y;
 6	/* By ensuring that addresses of x and y cannot be discarded,
 7	 * this empty asm guides GCC into representing extraction of
 8	 * their sign bits as memory loads rather than making x and y
 9	 * not-address-taken internally and using bitfield operations,
10	 * which in the end wouldn't work out, as extraction from FPU
11	 * registers needs to go through memory anyway. This way GCC
12	 * should manage to use incoming stack slots without spills. */
13	__asm__ ("" :: "X"(cx), "X"(cy));
14
15	long double t = x;
16	unsigned fpsr;
17	do __asm__ ("fprem1; fnstsw %%ax" : "+t"(t), "=a"(fpsr) : "u"(y));
18	while (fpsr & 0x400);
19	/* C0, C1, C3 flags in x87 status word carry low bits of quotient:
20	 * 15 14 13 12 11 10  9  8
21	 *  . C3  .  .  . C2 C1 C0
22	 *  . b1  .  .  .  0 b0 b2 */
23	unsigned char i = fpsr >> 8;
24	i = i>>4 | i<<4;
25	/* i[5:2] is now {b0 b2 ? b1}. Retrieve {0 b2 b1 b0} via
26	 * in-register table lookup. */
27	unsigned qbits = 0x7575313164642020 >> (i & 60);
28	qbits &= 7;
29
30	*quo = (cx[9]^cy[9]) < 0 ? -qbits : qbits;
31	return t;
32}