shithub: loong

ref: a2ad59b340bf3339dbba105310002474b98d0184
dir: /sys/src/cmd/zl/asm.c/

View raw version
#include	"l.h"

long	OFFSET;

#define PADDR(a)	((a) & ~0xfffffffff0000000ull)

#define VPUT(v)	VLEPUT(v)

#define VLEPUT(c)\
	{\
		cbp[0] = (c);\
		cbp[1] = (c)>>8;\
		cbp[2] = (c)>>16;\
		cbp[3] = (c)>>24;\
		cbp[4] = (c)>>32;\
		cbp[5] = (c)>>40;\
		cbp[6] = (c)>>48;\
		cbp[7] = (c)>>56;\
		cbp += 8;\
		cbc -= 8;\
		if(cbc <= 0)\
			cflush();\
	}

#define VBEPUT(c)\
	{\
		cbp[0] = (c)>>56;\
		cbp[1] = (c)>>48;\
		cbp[2] = (c)>>40;\
		cbp[3] = (c)>>32;\
		cbp[4] = (c)>>24;\
		cbp[5] = (c)>>16;\
		cbp[6] = (c)>>8;\
		cbp[7] = (c);\
		cbp += 8;\
		cbc -= 8;\
		if(cbc <= 0)\
			cflush();\
	}

#define LPUT(l)	LLEPUT(l)

#define	LLEPUT(c)\
	{\
		cbp[0] = (c);\
		cbp[1] = (c)>>8;\
		cbp[2] = (c)>>16;\
		cbp[3] = (c)>>24;\
		cbp += 4;\
		cbc -= 4;\
		if(cbc <= 0)\
			cflush();\
	}

#define	LBEPUT(c)\
	{\
		cbp[0] = (c)>>24;\
		cbp[1] = (c)>>16;\
		cbp[2] = (c)>>8;\
		cbp[3] = (c);\
		cbp += 4;\
		cbc -= 4;\
		if(cbc <= 0)\
			cflush();\
	}

#define HPUT(h)	HLEPUT(h)

#define	HLEPUT(c)\
	{\
		cbp[0] = (c);\
		cbp[1] = (c)>>8;\
		cbp += 2;\
		cbc -= 2;\
		if(cbc <= 0)\
			cflush();\
	}

#define	HBEPUT(c)\
	{\
		cbp[0] = (c)>>8;\
		cbp[1] = (c);\
		cbp += 2;\
		cbc -= 2;\
		if(cbc <= 0)\
			cflush();\
	}


#define	CPUT(c)\
	{\
		cbp[0] = (c);\
		cbp++;\
		cbc--;\
		if(cbc <= 0)\
			cflush();\
	}

void
cput(long l)
{
	CPUT(l);
}

void
objput(long l)	/* emit long in byte order appropriate to object machine */
{
	LPUT(l);
}

void
objhput(short s)
{
	HPUT(s);
}

void
wput(long l)
{

	cbp[0] = l>>8;
	cbp[1] = l;
	cbp += 2;
	cbc -= 2;
	if(cbc <= 0)
		cflush();
}

void
wputl(long l)
{

	cbp[0] = l;
	cbp[1] = l>>8;
	cbp += 2;
	cbc -= 2;
	if(cbc <= 0)
		cflush();
}

void
lput(long l)		/* emit long in big-endian byte order */
{
	LBEPUT(l);
}

void
lputl(long l)		/* emit long in big-endian byte order */
{
	LLEPUT(l);
}

void
llput(vlong v)
{
	lput(v>>32);
	lput(v);
}

void
llputl(vlong v)
{
	lputl(v);
	lputl(v>>32);
}

vlong
entryvalue(void)
{
	char *a;
	Sym *s;

	a = INITENTRY;
	if(*a >= '0' && *a <= '9')
		return atolwhex(a);
	s = lookup(a, 0);
	if(s->type == 0)
		return INITTEXT;
	if(s->type != STEXT && s->type != SLEAF)
		diag("entry not text: %s", s->name);
	return s->value;
}

void
asmb(void)
{
	Prog *p;
	long magic;
	vlong vl, t, etext;
	Optab *o;

	if(debug['v'])
		Bprint(&bso, "%5.2f asm\n", cputime());
	Bflush(&bso);
	OFFSET = HEADR;
	seek(cout, OFFSET, 0);
	pc = INITTEXT;
	for(p = firstp; p != P; p = p->link) {
		if(p->as == ATEXT) {
			curtext = p;
			autosize = p->to.offset + 8;
		}
		if(p->pc != pc) {
			diag("phase error %llux sb %llux", p->pc, pc);
			if(!debug['a'])
				prasm(curp);
			pc = p->pc;
		}
		curp = p;
		o = oplook(p);	/* could probably avoid this call */
		if(asmout(p, o, 0)) {
			p = p->link;
			pc += 4;
		}
		pc += o->size;
	}
	if(debug['a'])
		Bprint(&bso, "\n");
	Bflush(&bso);
	cflush();

	etext = INITTEXT + textsize;
	for(t = pc; t < etext; t += sizeof(buf)-100) {
		if(etext-t > sizeof(buf)-100)
			datblk(t, sizeof(buf)-100, 1);
		else
			datblk(t, etext-t, 1);
	}

	Bflush(&bso);
	cflush();

	curtext = P;
	switch(HEADTYPE) {
	case 0:
	case 2:
		OFFSET = HEADR+textsize;
		seek(cout, OFFSET, 0);
		break;
	case 6:	/* no header, padded */
		OFFSET = rnd(HEADR+textsize, INITRND);
		seek(cout, OFFSET, 0);
		break;
	}
	for(t = 0; t < datsize; t += sizeof(buf)-100) {
		if(datsize-t > sizeof(buf)-100)
			datblk(t, sizeof(buf)-100, 0);
		else
			datblk(t, datsize-t, 0);
	}

	symsize = 0;
	lcsize = 0;
	if(!debug['s']) {
		if(debug['v'])
			Bprint(&bso, "%5.2f sym\n", cputime());
		Bflush(&bso);
		switch(HEADTYPE) {
		case 0:
			debug['s'] = 1;
			break;
		case 2:
			OFFSET = HEADR+textsize+datsize;
			seek(cout, OFFSET, 0);
			break;
		case 6:	/* no header, padded */
			OFFSET += rnd(datsize, INITRND);
			seek(cout, OFFSET, 0);
			break;
		}
		if(!debug['s'])
			asmsym();
		if(debug['v'])
			Bprint(&bso, "%5.2f pc\n", cputime());
		Bflush(&bso);
		if(!debug['s'])
			asmlc();
		cflush();
	}

	if(debug['v'])
		Bprint(&bso, "%5.2f header\n", cputime());
	Bflush(&bso);
	OFFSET = 0;
	seek(cout, OFFSET, 0);
	switch(HEADTYPE) {
	case 0:	/* no header */
	case 6:	/* no header, padded */
		break;
	case 2:
		magic = 4*29*29+7;
		magic |= 0x00008000;
		lput(magic);		/* magic */
		lput(textsize);		/* sizes */
		lput(datsize);
		lput(bsssize);
		lput(symsize);		/* nsyms */
		vl = entryvalue();
		lput(PADDR(vl));	/* va of entry */
		lput(0L);
		lput(lcsize);
		llput(vl);		/* va of entry */
		break;
	}
	cflush();
}

void
strnput(char *s, int n)
{
	for(; *s; s++){
		CPUT(*s);
		n--;
	}
	for(; n > 0; n--)
		CPUT(0);
}

void
cflush(void)
{
	int n;

	n = sizeof(buf.cbuf) - cbc;
	if(n)
		write(cout, buf.cbuf, n);
	cbp = buf.cbuf;
	cbc = sizeof(buf.cbuf);
}

void
nopstat(char *f, Count *c)
{
	if(c->outof)
	Bprint(&bso, "%s delay %ld/%ld (%.2f)\n", f,
		c->outof - c->count, c->outof,
		(double)(c->outof - c->count)/c->outof);
}

void
asmsym(void)
{
	Prog *p;
	Auto *a;
	Sym *s;
	int h;

	s = lookup("etext", 0);
	if(s->type == STEXT)
		putsymb(s->name, 'T', s->value, s->version);

	for(h=0; h<NHASH; h++)
		for(s=hash[h]; s!=S; s=s->link)
			switch(s->type) {
			case SCONST:
				putsymb(s->name, 'D', s->value, s->version);
				continue;

			case SSTRING:
				putsymb(s->name, 'T', s->value, s->version);
				continue;

			case SDATA:
				putsymb(s->name, 'D', s->value+INITDAT, s->version);
				continue;

			case SBSS:
				putsymb(s->name, 'B', s->value+INITDAT, s->version);
				continue;

			case SFILE:
				putsymb(s->name, 'f', s->value, s->version);
				continue;
			}

	for(p=textp; p!=P; p=p->cond) {
		s = p->from.sym;
		if(s->type != STEXT && s->type != SLEAF)
			continue;

		/* filenames first */
		for(a=p->to.autom; a; a=a->link)
			if(a->type == D_FILE)
				putsymb(a->asym->name, 'z', a->aoffset, 0);
			else
			if(a->type == D_FILE1)
				putsymb(a->asym->name, 'Z', a->aoffset, 0);

		if(s->type == STEXT)
			putsymb(s->name, 'T', s->value, s->version);
		else
			putsymb(s->name, 'L', s->value, s->version);

		/* frame, auto and param after */
		putsymb(".frame", 'm', p->to.offset+8, 0);
		for(a=p->to.autom; a; a=a->link)
			if(a->type == D_AUTO)
				putsymb(a->asym->name, 'a', -a->aoffset, 0);
			else
			if(a->type == D_PARAM)
				putsymb(a->asym->name, 'p', a->aoffset, 0);
	}
	if(debug['v'] || debug['n'])
		Bprint(&bso, "symsize = %lud\n", symsize);
	Bflush(&bso);
}

void
putsymb(char *s, int t, vlong v, int ver)
{
	int i, f;
	long l;

	if(t == 'f')
		s++;
	if(HEADTYPE == 2) {
		l = v >> 32;
		LBEPUT(l);
	}
	l = v;
	LBEPUT(l);
	if(ver)
		t += 'a' - 'A';
	CPUT(t+0x80);			/* 0x80 is variable length */

	if(t == 'Z' || t == 'z') {
		CPUT(s[0]);
		for(i=1; s[i] != 0 || s[i+1] != 0; i += 2) {
			CPUT(s[i]);
			CPUT(s[i+1]);
		}
		CPUT(0);
		CPUT(0);
		i++;
	}
	else {
		for(i=0; s[i]; i++)
			CPUT(s[i]);
		CPUT(0);
	}
	symsize += 4 + 1 + i + 1;
	if(HEADTYPE == 2)
		symsize += 4;

	if(debug['n']) {
		if(t == 'z' || t == 'Z') {
			Bprint(&bso, "%c %.8llux ", t, v);
			for(i=1; s[i] != 0 || s[i+1] != 0; i+=2) {
				f = ((s[i]&0xff) << 8) | (s[i+1]&0xff);
				Bprint(&bso, "/%x", f);
			}
			Bprint(&bso, "\n");
			return;
		}
		if(ver)
			Bprint(&bso, "%c %.8llux %s<%d>\n", t, v, s, ver);
		else
			Bprint(&bso, "%c %.8llux %s\n", t, v, s);
	}
}

#define	MINLC	4
void
asmlc(void)
{
	long oldlc, v, s;
	vlong oldpc;
	Prog *p;

	oldpc = INITTEXT;
	oldlc = 0;
	for(p = firstp; p != P; p = p->link) {
		if(p->line == oldlc || p->as == ATEXT || p->as == ANOP) {
			if(p->as == ATEXT)
				curtext = p;
			if(debug['V'])
				Bprint(&bso, "%6llux %P\n", p->pc, p);
			continue;
		}
		if(debug['V'])
			Bprint(&bso, "\t\t%6ld", lcsize);
		v = (p->pc - oldpc) / MINLC;
		while(v) {
			s = 127;
			if(v < 127)
				s = v;
			CPUT(s+128);	/* 129-255 +pc */
			if(debug['V'])
				Bprint(&bso, " pc+%ld*%d(%ld)", s, MINLC, s+128);
			v -= s;
			lcsize++;
		}
		s = p->line - oldlc;
		oldlc = p->line;
		oldpc = p->pc + MINLC;
		if(s > 64 || s < -64) {
			CPUT(0);	/* 0 vv +lc */
			CPUT(s>>24);
			CPUT(s>>16);
			CPUT(s>>8);
			CPUT(s);
			if(debug['V']) {
				if(s > 0)
					Bprint(&bso, " lc+%ld(%d,%ld)\n",
						s, 0, s);
				else
					Bprint(&bso, " lc%ld(%d,%ld)\n",
						s, 0, s);
				Bprint(&bso, "%6llux %P\n", p->pc, p);
			}
			lcsize += 5;
			continue;
		}
		if(s > 0) {
			CPUT(0+s);	/* 1-64 +lc */
			if(debug['V']) {
				Bprint(&bso, " lc+%ld(%ld)\n", s, 0+s);
				Bprint(&bso, "%6llux %P\n", p->pc, p);
			}
		} else {
			CPUT(64-s);	/* 65-128 -lc */
			if(debug['V']) {
				Bprint(&bso, " lc%ld(%ld)\n", s, 64-s);
				Bprint(&bso, "%6llux %P\n", p->pc, p);
			}
		}
		lcsize++;
	}
	while(lcsize & 1) {
		s = 129;
		CPUT(s);
		lcsize++;
	}
	if(debug['v'] || debug['V'])
		Bprint(&bso, "lcsize = %ld\n", lcsize);
	Bflush(&bso);
}

void
datblk(long s, long n, int str)
{
	Prog *p;
	char *cast;
	long l, fl, j;
	vlong d;
	int i, c;

	memset(buf.dbuf, 0, n+100);
	for(p = datap; p != P; p = p->link) {
		curp = p;
		if(str != (p->from.sym->type == SSTRING))
			continue;
		l = p->from.sym->value + p->from.offset - s;
		c = p->reg;
		i = 0;
		if(l < 0) {
			if(l+c <= 0)
				continue;
			while(l < 0) {
				l++;
				i++;
			}
		}
		if(l >= n)
			continue;
		if(p->as != AINIT && p->as != ADYNT) {
			for(j=l+(c-i)-1; j>=l; j--)
				if(buf.dbuf[j]) {
					print("%P\n", p);
					diag("multiple initialization");
					break;
				}
		}
		switch(p->to.type) {
		default:
			diag("unknown mode in initialization\n%P", p);
			break;

		case D_FCONST:
			switch(c) {
			default:
			case 4:
				fl = ieeedtof(p->to.ieee);
				cast = (char*)&fl;
				for(; i<c; i++) {
					buf.dbuf[l] = cast[fnuxi4[i]];
					l++;
				}
				break;
			case 8:
				cast = (char*)p->to.ieee;
				for(; i<c; i++) {
					buf.dbuf[l] = cast[fnuxi8[i]];
					l++;
				}
				break;
			}
			break;

		case D_SCONST:
			for(; i<c; i++) {
				buf.dbuf[l] = p->to.sval[i];
				l++;
			}
			break;

		case D_CONST:
			d = p->to.offset;
			if(p->to.sym) {
				switch(p->to.sym->type) {
				case STEXT:
				case SLEAF:
				case SSTRING:
					d += p->to.sym->value;
					break;
				case SDATA:
				case SBSS:
					d += p->to.sym->value + INITDAT;
					break;
				}
			}
			cast = (char*)&d;
			switch(c) {
			default:
				diag("bad nuxi %d %d\n%P", c, i, curp);
				break;
			case 1:
				for(; i<c; i++) {
					buf.dbuf[l] = cast[inuxi1[i]];
					l++;
				}
				break;
			case 2:
				for(; i<c; i++) {
					buf.dbuf[l] = cast[inuxi2[i]];
					l++;
				}
				break;
			case 4:
				for(; i<c; i++) {
					buf.dbuf[l] = cast[inuxi4[i]];
					l++;
				}
				break;
			case 8:
				for(; i<c; i++) {
					buf.dbuf[l] = cast[inuxi8[i]];
					l++;
				}
				break;
			}
			break;
		}
	}
	write(cout, buf.dbuf, n);
}

#define	OP_RRR(op,r1,r2,r3)\
	(op|(((r1)&0x1FL)<<10)|(((r2)&0x1FL)<<5)|(((r3)&0x1FL)<<0))
#define	OP_RR(op,r2,r3)\
	(op|(((r2)&0x1FL)<<5)|(((r3)&0x1FL)<<0))
#define	OP_16IR_5I(op,i,r2)\
	(op|(((i)&0xFFFFL)<<10)|(((r2)&0x1FL)<<5)|(((i)>>16)&0x1FL))
#define	OP_16IRR(op,i,r2,r3)\
	(op|(((i)&0xFFFFL)<<10)|(((r2)&0x1FL)<<5)|(((r3)&0x1FL)<<0))
#define	OP_12IRR(op,i,r2,r3)\
	(op|(((i)&0xFFFL)<<10)|(((r2)&0x1FL)<<5)|(((r3)&0x1FL)<<0))
#define	OP_IR(op,i,r2)\
	(op|(((i)&0xFFFFFL)<<5)|(((r2)&0x1FL)<<0))
#define	OP_B_BL(op,i)\
	(op|(((i)&0xFFFFL)<<10)|(((i)>>16)&0x3FFL))

#define	OP(x,y)\
	(((x)<<3)|((y)<<0))
#define	SP(x,y)\
	(((x)<<29)|((y)<<26))
#define	OP_TEN(x,y)\
	(((x)<<21)|((y)<<10))

int vshift(int);

int
asmout(Prog *p, Optab *o, int aflag)
{
	long o1, o2, o3, o4, o5;
	vlong v;
	Prog *ct;
	int r, a;

	o1 = 0;
	o2 = 0;
	o3 = 0;
	o4 = 0;
	o5 = 0;
	switch(o->type) {
	default:
		diag("unknown type %d", o->type);
		if(!debug['a'])
			prasm(p);
		break;

	case 0:		/* pseudo ops */
		if(aflag) {
			if(p->link) {
				if(p->as == ATEXT) {
					ct = curtext;
					o2 = autosize;
					curtext = p;
					autosize = p->to.offset + 8;
					o1 = asmout(p->link, oplook(p->link), aflag);
					curtext = ct;
					autosize = o2;
				} else
					o1 = asmout(p->link, oplook(p->link), aflag);
			}
			return o1;
		}
		break;

	case 1:		/* mov[v] r1,r2 ==> OR r1,r0,r2 */
		o1 = OP_RRR(oprrr(AOR), REGZERO, p->from.reg, p->to.reg);
		break;

	case 2:		/* add/sub r1,[r2],r3 */
		r = p->reg;
		if(r == NREG)
			r = p->to.reg;
		o1 = OP_RRR(oprrr(p->as), p->from.reg, r, p->to.reg);
		break;

	case 3:		/* mov $soreg, r ==> or/add $i,o,r */
		v = regoff(&p->from);
		r = p->from.reg;
		if(r == NREG)
			r = o->param;
		a = AADDVU;
		if(o->a1 == C_ANDCON)
			a = AOR;
		o1 = OP_12IRR(opirr(a), v, r, p->to.reg);
		break;

	case 4:		/* add $scon,[r1],r2 */
		v = regoff(&p->from);
		r = p->reg;
		if(r == NREG)
			r = p->to.reg;
		o1 = OP_12IRR(opirr(p->as), v, r, p->to.reg);
		break;

	case 5:		/* syscall */
		if(aflag)
			return 0;
		o1 = oprrr(p->as);
		break;

	case 6:		/* beq r1,[r2],sbra */
		if(aflag)
			return 0;
		v = 0;
		if(p->cond != P)
			v = (p->cond->pc - pc) >> 2;
		if(((v << 16) >> 16) != v)
			diag("short branch too far: %lld\n%P", v, p);
		o1 = OP_16IRR(opirr(p->as), v, p->from.reg, p->reg);
		break;

	case 7:		/* mov r, soreg ==> sw o(r) */
		r = p->to.reg;
		if(r == NREG)
			r = o->param;
		v = regoff(&p->to);
		o1 = OP_12IRR(opirr(p->as), v, r, p->from.reg);
		break;

	case 8:		/* mov soreg, r ==> lw o(r) */
		r = p->from.reg;
		if(r == NREG)
			r = o->param;
		v = regoff(&p->from);
		o1 = OP_12IRR(opirr(p->as+ALAST), v, r, p->to.reg);
		break;

	case 9:		/* asl r1,[r2],r3 */
		r = p->reg;
		if(r == NREG)
			r = p->to.reg;
		o1 = OP_RRR(oprrr(p->as), r, p->from.reg, p->to.reg);
		break;

	case 10:	/* add $con,[r1],r2 ==> mov $con,t; add t,[r1],r2 */
		v = regoff(&p->from);
		r = AOR;
		if(v < 0)
			r = AADDU;
		o1 = OP_12IRR(opirr(r), v, 0, REGTMP);
		r = p->reg;
		if(r == NREG)
			r = p->to.reg;
		o2 = OP_RRR(oprrr(p->as), REGTMP, r, p->to.reg);
		break;

	case 11:	/* jmp lbra */
		if(aflag)
			return 0;
		v = 0;
		if(p->cond != P)
			v = (p->cond->pc - pc) >> 2;
		o1 = OP_B_BL(opirr(p->as), v);
		break;

	case 12:	/* movbs r,r */
		v = 16;
		if(p->as == AMOVB)
			v = 24;
		o1 = OP_16IRR(opirr(ASLL), v, p->from.reg, p->to.reg);
		o2 = OP_16IRR(opirr(ASRA), v, p->to.reg, p->to.reg);
		break;

	case 13:	/* movbu r,r */
		if(p->as == AMOVBU)
			o1 = OP_12IRR(opirr(AAND), 0xffL, p->from.reg, p->to.reg);
		else
			o1 = (0x33c0<<10) | ((p->from.reg&0x1FL)<<5) | (p->to.reg&0x1FL);
		break;

	case 14:	/* movwu r,r */
		o1 = OP_16IRR(opirr(ASLLV+ALAST), 32, p->from.reg, p->to.reg);
		o2 = OP_16IRR(opirr(ASRLV+ALAST), 32, p->to.reg, p->to.reg);
		break;

	case 16:	/* sll $c,[r1],r2 */
		v = regoff(&p->from);
		r = p->reg;
		if(r == NREG)
			r = p->to.reg;

		/* OP_SRR will use only the low 5 bits of the shift value */
		if(v >= 32 && vshift(p->as))
			o1 = OP_16IRR(opirr(p->as+ALAST), v&0x3FL, r, p->to.reg);
		else
			o1 = OP_16IRR(opirr(p->as), v&0x1FL, r, p->to.reg);
		break;

	case 18:	/* jmp [r1],0(r2) */
		if(aflag)
			return 0;
		r = p->reg;
		if(r == NREG)
			r = o->param;
		o1 = OP_RRR(oprrr(p->as), 0, p->to.reg, r);
		break;

	case 19:	/* mov $lcon,r ==> lu+or */
		v = regoff(&p->from);
		o1 = OP_IR(opir(ALU12IW), v>>12, p->to.reg);
		o2 = OP_12IRR(opirr(AOR), v, p->to.reg, p->to.reg);
		break;

	case 23:	/* add $lcon,r1,r2 ==> lu+or+add */
		v = regoff(&p->from);
		if(p->to.reg == REGTMP || p->reg == REGTMP)
			diag("cant synthesize large constant\n%P", p);
		o1 = OP_IR(opir(ALU12IW), v>>12, REGTMP);
		o2 = OP_12IRR(opirr(AOR), v, REGTMP, REGTMP);
		r = p->reg;
		if(r == NREG)
			r = p->to.reg;
		o3 = OP_RRR(oprrr(p->as), REGTMP, r, p->to.reg);
		break;

	case 24:	/* mov $ucon,,r ==> lu r */
		v = regoff(&p->from);
		o1 = OP_IR(opir(ALU12IW), v>>12, p->to.reg);
		break;

	case 25:	/* add/and $ucon,[r1],r2 ==> lu $con,t; add t,[r1],r2 */
		v = regoff(&p->from);
		o1 = OP_IR(opir(ALU12IW), v>>12, REGTMP);
		r = p->reg;
		if(r == NREG)
			r = p->to.reg;
		o2 = OP_RRR(oprrr(p->as), REGTMP, r, p->to.reg);
		break;

	case 26:	/* mov $lsext/auto/oreg,,r2 ==> lu+or+add */
		v = regoff(&p->from);
		if(p->to.reg == REGTMP)
			diag("cant synthesize large constant\n%P", p);
		o1 = OP_IR(opir(ALU12IW), v>>12, REGTMP);
		o2 = OP_12IRR(opirr(AOR), v, REGTMP, REGTMP);
		r = p->from.reg;
		if(r == NREG)
			r = o->param;
		o3 = OP_RRR(oprrr(AADDVU), REGTMP, r, p->to.reg);
		break;

	case 27:		/* mov [sl]ext/auto/oreg,fr ==> lwc1 o(r) */
		r = p->from.reg;
		if(r == NREG)
			r = o->param;
		v = regoff(&p->from);
		a = AMOVF+ALAST;
		if(p->as == AMOVD)
			a = AMOVD+ALAST;
		switch(o->size) {
		case 12:
			o1 = OP_IR(opir(ALU12IW), (v+1<<11)>>12, REGTMP);
			o2 = OP_RRR(oprrr(AADDVU), r, REGTMP, REGTMP);
			o3 = OP_12IRR(opirr(a), v, REGTMP, p->to.reg);
			break;
		case 4:
			o1 = OP_12IRR(opirr(a), v, r, p->to.reg);
			break;
		default:
			diag("bad 27 op size");
		}
		break;

	case 28:		/* mov fr,[sl]ext/auto/oreg ==> swc1 o(r) */
		r = p->to.reg;
		if(r == NREG)
			r = o->param;
		v = regoff(&p->to);
		a = AMOVF;
		if(p->as == AMOVD)
			a = AMOVD;
		switch(o->size) {
		case 12:
			if(r == REGTMP)
				diag("cant synthesize large constant\n%P", p);
			o1 = OP_IR(opir(ALU12IW), (v+1<<11)>>12, REGTMP);
			o2 = OP_RRR(oprrr(AADDVU), r, REGTMP, REGTMP);
			o3 = OP_12IRR(opirr(a), v, REGTMP, p->from.reg);
			break;
		case 4:
			o1 = OP_12IRR(opirr(a), v, r, p->from.reg);
			break;
		default:
			diag("bad 28 op size");
		}
		break;

	case 30:	/* movw r,fr */
		a = OP_TEN(8, 1321);	/* movgr2fr.w */
		o1 = OP_RR(a, p->from.reg, p->to.reg);
		break;

	case 31:	/* movw fr,r */
		a = OP_TEN(8, 1325);		/* movfr2gr.s */
		o1 = OP_RR(a, p->to.reg, p->from.reg);
		break;

	case 32:	/* fadd fr1,[fr2],fr3 */
		r = p->reg;
		if(r == NREG)
			r = p->to.reg;
		o1 = OP_RRR(oprrr(p->as), p->from.reg, r, p->to.reg);
		break;

	case 33:	/* fabs fr1,fr3 */
		o1 = OP_RRR(oprrr(p->as), 0, p->from.reg, p->to.reg);
		break;

	case 34:	/* mov $con,fr ==> or/add $i,r,r2 */
		v = regoff(&p->from);
		r = AADDU;
		if(o->a1 == C_ANDCON)
			r = AOR;
		o1 = OP_12IRR(opirr(r), v, 0, REGTMP);
		o2 = OP_RRR(OP_TEN(8, 1321), REGTMP, 0, p->to.reg);	/* movgr2fr.w */
		break;

	case 35:	/* mov r,lext/luto/oreg ==> sw o(r) */
		v = regoff(&p->to);
		r = p->to.reg;
		if(r == NREG)
			r = o->param;
		if(r == REGTMP)
			diag("cant synthesize large constant\n%P", p);
		o1 = OP_IR(opir(ALU12IW), (v+1<<11)>>12, REGTMP);
		o2 = OP_RRR(oprrr(AADDVU), r, REGTMP, REGTMP);
		o3 = OP_12IRR(opirr(p->as), r, REGTMP, p->from.reg);
		break;

	case 36:	/* mov lext/lauto/lreg,r ==> lw o(r30) */
		v = regoff(&p->from);
		r = p->from.reg;
		if(r == NREG)
			r = o->param;
		if(r == REGTMP)
			diag("cant synthesize large constant\n%P", p);
		o1 = OP_IR(opir(ALU12IW), (v+1<<11)>>12, REGTMP);
		o2 = OP_RRR(oprrr(AADDVU), r, REGTMP, REGTMP);
		o3 = OP_12IRR(opirr(p->as+ALAST), v, REGTMP, p->to.reg);
		break;

	case 40:	/* word */
		if(aflag)
			return 0;
		o1 = regoff(&p->to);
		break;

	case 47:	/* movv r,fr */
		r = OP_TEN(8, 1322);		/* movgr2fr.d */
		o1 = OP_RR(r, p->from.reg, p->to.reg);
		break;

	case 48:	/* movv fr,r */
		r = OP_TEN(8, 1326);		/* movfr2gr.d */
		o1 = OP_RR(r, p->from.reg, p->to.reg);
		break;
	}
	if(aflag)
		return o1;
	v = p->pc;
	switch(o->size) {
	default:
		if(debug['a'])
			Bprint(&bso, " %.16llux:\t\t%P\n", v, p);
		break;
	case 4:
		if(debug['a'])
			Bprint(&bso, " %.16llux: %.8lux\t%P\n", v, o1, p);
		LPUT(o1);
		break;
	case 8:
		if(debug['a'])
			Bprint(&bso, " %.16llux: %.8lux %.8lux\t%P\n", v, o1, o2, p);
		LPUT(o1);
		LPUT(o2);
		break;
	case 12:
		if(debug['a'])
			Bprint(&bso, " %.16llux: %.8lux %.8lux %.8lux\t%P\n", v, o1, o2, o3, p);
		LPUT(o1);
		LPUT(o2);
		LPUT(o3);
		break;
	case 16:
		if(debug['a'])
			Bprint(&bso, " %.16llux: %.8lux %.8lux %.8lux %.8lux%P\n",
				v, o1, o2, o3, o4, p);
		LPUT(o1);
		LPUT(o2);
		LPUT(o3);
		LPUT(o4);
		break;
	case 20:
		if(debug['a'])
			Bprint(&bso, " %.16llux: %.8lux %.8lux %.8lux %.8lux %.8lux%P\n",
				v, o1, o2, o3, o4, o5, p);
		LPUT(o1);
		LPUT(o2);
		LPUT(o3);
		LPUT(o4);
		LPUT(o5);
		break;
	}
	return 0;
}

int
isnop(Prog *p)
{
	if(p->as != ANOR)
		return 0;
	if(p->reg != REGZERO && p->reg != NREG)
		return 0;
	if(p->from.type != D_REG || p->from.reg != REGZERO)
		return 0;
	if(p->to.type != D_REG || p->to.reg != REGZERO)
		return 0;
	return 1;
}

long
oprrr(int a)
{
	switch(a) {
	case AADD:
	case AADDU:	return 0x20<<15;
	case ASGT:	return 0x24<<15;
	case ASGTU:	return 0x25<<15;
	case AAND:	return 0x29<<15;
	case AOR:	return 0x2a<<15;
	case AXOR:	return 0x2b<<15;
	case ASUB:
	case ASUBU:	return 0x22<<15;
	case ANOR:	return 0x28<<15;
	case ASLL:	return 0x2e<<15;
	case ASRL:	return 0x2f<<15;
	case ASRA:	return 0x30<<15;

	case ASLLV:	return 0x31<<15;
	case ASRLV:	return 0x32<<15;
	case ASRAV:	return 0x33<<15;

	case AADDV:
	case AADDVU:	return 0x21<<15;
	case ASUBV:
	case ASUBVU:	return 0x23<<15;
	case AREM:	return 0x41<<15;
	case ADIV:	return 0x40<<15;
	case AREMU:	return 0x43<<15;
	case ADIVU:	return 0x42<<15;
	case AMUL:
	case AMULU:	return 0x38<<15;

	case AREMV:	return 0x45<<15;
	case ADIVV:	return 0x44<<15;
	case AREMVU:	return 0x47<<15;
	case ADIVVU:	return 0x46<<15;
	case AMULV:
	case AMULVU:	return 0x3b<<15;

	case AJMP:	return 0x13<<26;
	case AJAL:	return (0x13<<26)|1;

	case ABREAK:	return 0x54<<15;
	case ASYSCALL:	return 0x56<<15;

	case ADIVF:	return 0x20d<<15;
	case ADIVD:	return 0x20e<<15;
	case AMULF:	return 0x209<<15;
	case AMULD:	return 0x20a<<15;
	case ASUBF:	return 0x205<<15;
	case ASUBD:	return 0x206<<15;
	case AADDF:	return 0x201<<15;
	case AADDD:	return 0x202<<15;

	case ATRUNCFV:	return 0x46a9<<10;
	case ATRUNCDV:	return 0x46aa<<10;
	case ATRUNCFW:	return 0x46a1<<10;
	case ATRUNCDW:	return 0x46a2<<10;
	case AMOVFV:	return 0x46c9<<10;
	case AMOVDV:	return 0x46ca<<10;
	case AMOVVF:	return 0x4746<<10;
	case AMOVVD:	return 0x474a<<10;

	case AMOVFW:	return 0x46c1<<10;
	case AMOVDW:	return 0x46c2<<10;
	case AMOVWF:	return 0x4744<<10;
	case AMOVDF:	return 0x4646<<10;
	case AMOVWD:	return 0x4748<<10;
	case AMOVFD:	return 0x4649<<10;
	case AABSF:	return 0x4501<<10;
	case AABSD:	return 0x4502<<10;
	case AMOVF:	return 0x4525<<10;
	case AMOVD:	return 0x4526<<10;
	case ANEGF:	return 0x4505<<10;
	case ANEGD:	return 0x4506<<10;

	case ACMPEQF:	return (0xc1<<20) | (0x4<<15);
	case ACMPEQD:	return (0xc2<<20) | (0x4<<15);
	case ACMPGTF:	return (0xc1<<20) | (0x3<<15);
	case ACMPGTD:	return (0xc2<<20) | (0x3<<15);
	case ACMPGEF:	return (0xc1<<20) | (0x7<<15);
	case ACMPGED:	return (0xc2<<20) | (0x7<<15);
	}
	if(a >= ALAST)
		diag("bad rrr %A+ALAST", a-ALAST);
	else
		diag("bad rrr %A", a);
	return 0;
}

long
opir(int a)
{
	switch(a){
	case ALU12IW:	return 0xa<<25;
	case ALU32ID:	return 0xb<<25;
	}
	diag("bad ir %d", a);
	return 0;
}

long
opirr(int a)
{
	switch(a) {
	case AADD:
	case AADDU:		return 0xa<<22;
	case ASGT:		return 0x8<<22;
	case ASGTU:		return 0x9<<22;
	case AAND:		return 0xd<<22;
	case AOR:		return 0xe<<22;
	case AXOR:		return 0xf<<22;
	case ASLL:		return 0x81<<15;
	case ASRL:		return 0x89<<15;
	case ASRA:		return 0x91<<15;

	case AADDV:
	case AADDVU:		return 0xb<<22;

	case AJMP:		return 0x14<<26;
	case AJAL:		return 0x15<<26;
	case ABEQ:		return 0x16<<26;
	case ABEQ+ALAST:	return 0x10<<26;			/* likely */
	case ABNE:		return 0x17<<26;
	case ABNE+ALAST:	return 0x11<<26;			/* likely */

	case ABLEZ:
	case ABGEZ:		return 0x19<<26;
	case ABLTZ:
	case ABGTZ:		return 0x18<<26;

	case ABFPT:		return (0x12<<26) | (0x1<<8);
	case ABFPF:		return 0x12<<26;

	case AMOVB:
	case AMOVBU:		return 0xa4<<22;
	case AMOVH:
	case AMOVHU:		return 0xa5<<22;
	case AMOVW:		return 0xa6<<22;
	case AMOVV:		return 0xa7<<22;
	case AMOVF:		return 0xad<<22;
	case AMOVD:		return 0xaf<<22;
	case AMOVWL:		return 0xbc<<22;
	case AMOVWR:		return 0xbd<<22;
	case AMOVVL:		return 0xbe<<22;
	case AMOVVR:		return 0xbf<<22;

	case ABREAK:		return 0x18<<22;

	case AMOVWL+ALAST:	return 0xb8<<22;
	case AMOVWR+ALAST:	return 0xb9<<22;
	case AMOVVL+ALAST:	return 0xba<<22;
	case AMOVVR+ALAST:	return 0xbb<<22;
	case AMOVB+ALAST:	return 0xa0<<22;
	case AMOVBU+ALAST:	return 0xa8<<22;
	case AMOVH+ALAST:	return 0xa1<<22;
	case AMOVHU+ALAST:	return 0xa9<<22;
	case AMOVW+ALAST:	return 0xa2<<22;
	case AMOVV+ALAST:	return 0xa3<<22;
	case AMOVF+ALAST:	return 0xac<<22;
	case AMOVD+ALAST:	return 0xae<<22;

	case ASLLV:
	case ASLLV+ALAST:	return 0x41<<16;
	case ASRLV:
	case ASRLV+ALAST:	return 0x45<<16;
	case ASRAV:
	case ASRAV+ALAST:	return 0x49<<16;

	case  ALU52ID:		return 0xc<<22;
	}
	if(a >= ALAST)
		diag("bad irr %A+ALAST", a-ALAST);
	else
		diag("bad irr %A", a);
	return 0;
}

int
vshift(int a)
{
	switch(a){
	case ASLLV:		return 1;
	case ASRLV:		return 1;
	case ASRAV:		return 1;
	}
	return 0;
}