code *movxmmconst(unsigned xreg, unsigned sz, targ_size_t value, regm_t flags)
{
/* Generate:
* MOV reg,value
* MOV xreg,reg
* Not so efficient. We should at least do a PXOR for 0.
*/
assert(mask[xreg] & XMMREGS);
assert(sz == 4 || sz == 8);
CodeBuilder cdb;
if (I32 && sz == 8)
{
unsigned r;
regm_t rm = ALLREGS;
cdb.append(allocreg(&rm,&r,TYint)); // allocate scratch register
union { targ_size_t s; targ_long l[2]; } u;
u.l[1] = 0;
u.s = value;
targ_long *p = &u.l[0];
cdb.append(movregconst(CNIL,r,p[0],0));
cdb.genfltreg(STO,r,0); // MOV floatreg,r
cdb.append(movregconst(CNIL,r,p[1],0));
cdb.genfltreg(STO,r,4); // MOV floatreg+4,r
unsigned op = xmmload(TYdouble, true);
cdb.genxmmreg(op,xreg,0,TYdouble); // MOVSD XMMreg,floatreg
}
else
{
unsigned reg;
cdb.append(regwithvalue(CNIL,ALLREGS,value,®,(sz == 8) ? 64 : 0));
cdb.gen2(LODD,modregxrmx(3,xreg-XMM0,reg)); // MOVD xreg,reg
if (sz == 8)
code_orrex(cdb.last(), REX_W);
checkSetVex(cdb.last(), TYulong);
}
return cdb.finish();
}
code *movxmmconst(unsigned xreg, unsigned sz, targ_size_t value, regm_t flags)
{
/* Generate:
* MOV reg,value
* MOV xreg,reg
* Not so efficient. We should at least do a PXOR for 0.
*/
assert(mask[xreg] & XMMREGS);
assert(sz == 4 || sz == 8);
code *c;
if (I32 && sz == 8)
{
unsigned r;
regm_t rm = ALLREGS;
c = allocreg(&rm,&r,TYint); // allocate scratch register
union { targ_size_t s; targ_long l[2]; } u;
u.l[1] = 0;
u.s = value;
targ_long *p = &u.l[0];
c = movregconst(c,r,p[0],0);
c = genfltreg(c,0x89,r,0); // MOV floatreg,r
c = movregconst(c,r,p[1],0);
c = genfltreg(c,0x89,r,4); // MOV floatreg+4,r
unsigned op = xmmload(TYdouble);
c = genfltreg(c,op,xreg - XMM0,0); // MOVSD XMMreg,floatreg
}
else
{
unsigned reg;
c = regwithvalue(CNIL,ALLREGS,value,®,(sz == 8) ? 64 : 0);
c = gen2(c,LODD,modregxrmx(3,xreg-XMM0,reg)); // MOVD xreg,reg
if (sz == 8)
code_orrex(c, REX_W);
}
return c;
}