#include	"l.h"

void
span(void)
{
	Prog *p, *q;
	long v, c, idat;
	int m, n, again;

	xdefine("etext", STEXT, 0L);
	idat = INITDAT;
	for(p = firstp; p != P; p = p->link) {
		if(p->as == ATEXT)
			curtext = p;
		n = 0;
		if(p->to.type == D_BRANCH)
			if(p->cond == P)
				p->cond = p;
		if((q = p->cond) != P)
			if(q->back != 2)
				n = 1;
		p->back = n;
		if(p->as == AADJSP) {
			p->to.type = D_SP;
			v = -p->from.offset;
			p->from.offset = v;
			p->as = AADDL;
			if(v < 0) {
				p->as = ASUBL;
				v = -v;
				p->from.offset = v;
			}
			if(v == 0)
				p->as = ANOP;
		}
	}
	n = 0;

start:
	if(debug['v'])
		Bprint(&bso, "%5.2f span\n", cputime());
	Bflush(&bso);
	c = INITTEXT;
	for(p = firstp; p != P; p = p->link) {
		if(p->as == ATEXT)
			curtext = p;
		if(p->to.type == D_BRANCH)
			if(p->back)
				p->pc = c;
		asmins(p);
		p->pc = c;
		m = andptr-and;
		p->mark = m;
		c += m;
	}

loop:
	n++;
	if(debug['v'])
		Bprint(&bso, "%5.2f span %d\n", cputime(), n);
	Bflush(&bso);
	if(n > 50) {
		print("span must be looping\n");
		errorexit();
	}
	again = 0;
	c = INITTEXT;
	for(p = firstp; p != P; p = p->link) {
		if(p->as == ATEXT)
			curtext = p;
		if(p->to.type == D_BRANCH) {
			if(p->back)
				p->pc = c;
			asmins(p);
			m = andptr-and;
			if(m != p->mark) {
				p->mark = m;
				again++;
			}
		}
		p->pc = c;
		c += p->mark;
	}
	if(again) {
		textsize = c;
		goto loop;
	}
	if(INITRND) {
		INITDAT = rnd(c, INITRND);
		if(INITDAT != idat) {
			idat = INITDAT;
			goto start;
		}
	}
	xdefine("etext", STEXT, c);
	if(debug['v'])
		Bprint(&bso, "etext = %lux\n", c);
	Bflush(&bso);
	for(p = textp; p != P; p = p->cond)
		p->from.sym->value = p->pc;
	textsize = c - INITTEXT;
}

void
xdefine(char *p, int t, long v)
{
	Sym *s;

	s = lookup(p, 0);
	if(s->type == 0 || s->type == SXREF) {
		s->type = t;
		s->value = v;
	}
	if(s->type == STEXT && s->value == 0)
		s->value = v;
}

void
putsymb(Sym *s, int t, long v)
{
	int i, f;
	char *n, str[STRINGSZ];

	n = s->name;
	if(t == 'f')
		n++;
	lput(v);
	if(s->version)
		t += 'a' - 'A';
	CPUT(t);
	for(i=0; i<NNAME; i++)
		CPUT(n[i]);
	CPUT(0);
	CPUT(0);
	CPUT(0);
	symsize += 4 + 1 + NNAME + 3;
	if(debug['n']) {
		if(t == 'z' || t == 'Z') {
			str[0] = 0;
			for(i=1; i<NNAME; i+=2) {
				f = ((n[i]&0xff) << 8) | (n[i+1]&0xff);
				if(f == 0)
					break;
				sprint(strchr(str, 0), "/%x", f);
			}
			Bprint(&bso, "%c %.6lux %s\n", t, v, str);
			return;
		}
		if(s->version)
			Bprint(&bso, "%c %.6lux %s<%d>\n", t, v, n, s->version);
		else
			Bprint(&bso, "%c %.6lux %s\n", t, v, n);
	}
}

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

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

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

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

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

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

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

		putsymb(s, 'T', s->value);

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

void
asmsp(void)
{
	long oldpc, oldsp;
	Prog *p;
	int s;
	long v;

	oldpc = INITTEXT;
	oldsp = 0;
	for(p = firstp; p != P; p = p->link) {
		if(p->stkoff == oldsp || p->as == ATEXT || p->as == ANOP) {
			if(p->as == ATEXT)
				curtext = p;
			if(debug['G'])
				Bprint(&bso, "%6lux %4ld%P\n",
					p->pc, p->stkoff, p);
			continue;
		}
		if(debug['G'])
			Bprint(&bso, "\t\t%6ld", spsize);
		v = (p->pc - oldpc) / MINLC;
		while(v) {
			s = 127;
			if(v < 127)
				s = v;
			CPUT(s+128);	/* 129-255 +pc */
			if(debug['G'])
				Bprint(&bso, " pc+%d*2(%d)", s, s+128);
			v -= s;
			spsize++;
		}
		v = p->stkoff - oldsp;
		oldsp = p->stkoff;
		oldpc = p->pc + MINLC;
		if(v & 3 || v > 64L*4L || v < -64L*4L) {
			CPUT(0);	/* 0 vvvv +sp */
			lput(v);
			if(debug['G']) {
				if(v > 0)
					Bprint(&bso, " sp+%ld*1(%d,%ld)\n",
						v, 0, v);
				else
					Bprint(&bso, " sp%ld*1(%d,%ld)\n",
						v, 0, v);
				Bprint(&bso, "%6lux %4ld%P\n",
					p->pc, p->stkoff, p);
			}
			spsize += 5;
			continue;
		}
		s = v/4;
		if(s > 0) {
			CPUT(0+s);	/* 1-64 +sp */
			if(debug['G']) {
				Bprint(&bso, " sp+%d*4(%d)\n", s, 0+s);
				Bprint(&bso, "%6lux %4ld%P\n",
					p->pc, p->stkoff, p);
			}
		} else {
			CPUT(64-s);	/* 65-128 -sp */
			if(debug['G']) {
				Bprint(&bso, " sp%d*4(%d)\n", s, 64-s);
				Bprint(&bso, "%6lux %4ld%P\n",
					p->pc, p->stkoff, p);
			}
		}
		spsize++;
	}
	while(spsize & 1) {
		s = 129;
		CPUT(s);
		spsize++;
	}
	if(debug['v'] || debug['G'])
		Bprint(&bso, "stsize = %ld\n", spsize);
	Bflush(&bso);
}

void
asmlc(void)
{
	long oldpc, oldlc;
	Prog *p;
	long v, s;

	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['L'])
				Bprint(&bso, "%6lux %P\n",
					p->pc, p);
			continue;
		}
		if(debug['L'])
			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['L'])
				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['L']) {
				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, "%6lux %P\n",
					p->pc, p);
			}
			lcsize += 5;
			continue;
		}
		if(s > 0) {
			CPUT(0+s);	/* 1-64 +lc */
			if(debug['L']) {
				Bprint(&bso, " lc+%ld(%ld)\n", s, 0+s);
				Bprint(&bso, "%6lux %P\n",
					p->pc, p);
			}
		} else {
			CPUT(64-s);	/* 65-128 -lc */
			if(debug['L']) {
				Bprint(&bso, " lc%ld(%ld)\n", s, 64-s);
				Bprint(&bso, "%6lux %P\n",
					p->pc, p);
			}
		}
		lcsize++;
	}
	while(lcsize & 1) {
		s = 129;
		CPUT(s);
		lcsize++;
	}
	if(debug['v'] || debug['L'])
		Bprint(&bso, "lcsize = %ld\n", lcsize);
	Bflush(&bso);
}

int
oclass(Adr *a)
{
	long v;

	if(a->type >= D_INDIR || a->index != D_NONE) {
		if(a->index != D_NONE && a->scale == 0) {
			if(a->type >= D_ADDR)
				return Yi32;
			return Ycol;
		}
		return Ym;
	}
	switch(a->type)
	{
	case D_AL:
		return Yal;

	case D_AX:
		return Yax;

	case D_CL:
	case D_DL:
	case D_BL:
	case D_AH:
	case D_CH:
	case D_DH:
	case D_BH:
		return Yrb;

	case D_CX:
		return Ycx;

	case D_DX:
	case D_BX:
		return Yrx;

	case D_SP:
	case D_BP:
	case D_SI:
	case D_DI:
		return Yrl;

	case D_F0+0:
		return	Yf0;

	case D_F0+1:
	case D_F0+2:
	case D_F0+3:
	case D_F0+4:
	case D_F0+5:
	case D_F0+6:
	case D_F0+7:
		return	Yrf;

	case D_NONE:
		return Ynone;

	case D_CS:	return	Ycs;
	case D_SS:	return	Yss;
	case D_DS:	return	Yds;
	case D_ES:	return	Yes;
	case D_FS:	return	Yfs;
	case D_GS:	return	Ygs;

	case D_GDTR:	return	Ygdtr;
	case D_IDTR:	return	Yidtr;
	case D_LDTR:	return	Yldtr;
	case D_MSW:	return	Ymsw;
	case D_TASK:	return	Ytask;

	case D_CR+0:	return	Ycr0;
	case D_CR+1:	return	Ycr1;
	case D_CR+2:	return	Ycr2;
	case D_CR+3:	return	Ycr3;
	case D_CR+4:	return	Ycr4;
	case D_CR+5:	return	Ycr5;
	case D_CR+6:	return	Ycr6;
	case D_CR+7:	return	Ycr7;

	case D_DR+0:	return	Ydr0;
	case D_DR+1:	return	Ydr1;
	case D_DR+2:	return	Ydr2;
	case D_DR+3:	return	Ydr3;
	case D_DR+4:	return	Ydr4;
	case D_DR+5:	return	Ydr5;
	case D_DR+6:	return	Ydr6;
	case D_DR+7:	return	Ydr7;

	case D_TR+0:	return	Ytr0;
	case D_TR+1:	return	Ytr1;
	case D_TR+2:	return	Ytr2;
	case D_TR+3:	return	Ytr3;
	case D_TR+4:	return	Ytr4;
	case D_TR+5:	return	Ytr5;
	case D_TR+6:	return	Ytr6;
	case D_TR+7:	return	Ytr7;

	case D_EXTERN:
	case D_STATIC:
	case D_AUTO:
	case D_PARAM:
		return Ym;

	case D_CONST:
	case D_ADDR:
		if(a->sym == S) {
			v = a->offset;
			if(v == 0)
				return Yi0;
			if(v == 1)
				return Yi1;
			if(v >= -128 && v <= 127)
				return Yi8;
		}
		return Yi32;

	case D_BRANCH:
		return Ybr;
	}
	return Yxxx;
}

void
asmidx(Adr *a, int base)
{
	int i;

	switch(a->index) {
	default:
		goto bad;

	case D_NONE:
		i = 4 << 3;
		goto bas;

	case D_AX:
	case D_CX:
	case D_DX:
	case D_BX:
	case D_BP:
	case D_SI:
	case D_DI:
		i = reg[a->index] << 3;
		break;
	}
	switch(a->scale) {
	default:
		goto bad;
	case 1:
		break;
	case 2:
		i |= (1<<6);
		break;
	case 4:
		i |= (2<<6);
		break;
	case 8:
		i |= (3<<6);
		break;
	}
bas:
	switch(base) {
	default:
		goto bad;
	case D_NONE:	/* must be mod=00 */
		i |= 5;
		break;
	case D_AX:
	case D_CX:
	case D_DX:
	case D_BX:
	case D_SP:
	case D_SI:
	case D_DI:
		i |= reg[base];
		break;
	}
	*andptr++ = i;
	return;
bad:
	diag("asmidx: bad address %D\n", a);
	*andptr++ = 0;
	return;
}

long
vaddr(Adr *a)
{
	int t;
	long v;

	t = a->type;
	v = a->offset;
	if(t == D_ADDR)
		t = a->index;
	switch(t) {
	case D_STATIC:
	case D_EXTERN:
		if(a->sym) {
			v += a->sym->value;
			if(a->sym->type != STEXT)
				v += INITDAT;
		}
	}
	return v;
}

void
asmand(Adr *a, int r)
{
	long v;
	int t;
	Adr aa;

	v = a->offset;
	t = a->type;
	if(a->index != D_NONE) {
		if(t >= D_INDIR) {
			t -= D_INDIR;
			if(t == D_NONE) {
				*andptr++ = (0 << 6) | (4 << 0) | (r << 3);
				asmidx(a, t);
				andptr[0] = v;
				andptr[1] = v>>8;
				andptr[2] = v>>16;
				andptr[3] = v>>24;
				andptr += 4;
				return;
			}
			if(v == 0) {
				*andptr++ = (0 << 6) | (4 << 0) | (r << 3);
				asmidx(a, t);
				return;
			}
			if(v >= -128 && v < 128) {
				*andptr++ = (1 << 6) | (4 << 0) | (r << 3);
				asmidx(a, t);
				*andptr++ = v;
				return;
			}
			*andptr++ = (2 << 6) | (4 << 0) | (r << 3);
			asmidx(a, t);
			andptr[0] = v;
			andptr[1] = v>>8;
			andptr[2] = v>>16;
			andptr[3] = v>>24;
			andptr += 4;
			return;
		}
		switch(t) {
		default:
			goto bad;
		case D_STATIC:
		case D_EXTERN:
			aa.type = D_NONE+D_INDIR;
			break;
		case D_AUTO:
		case D_PARAM:
			aa.type = D_SP+D_INDIR;
			break;
		}
		aa.offset = vaddr(a);
		aa.index = a->index;
		aa.scale = a->scale;
		asmand(&aa, r);
		return;
	}
	if(t >= D_AL && t <= D_F0+7) {
		if(v)
			goto bad;
		*andptr++ = (3 << 6) | (reg[t] << 0) | (r << 3);
		return;
	}
	if(t >= D_INDIR) {
		t -= D_INDIR;
		if(t == D_NONE) {
			andptr[0] = (0 << 6) | (5 << 0) | (r << 3);
			andptr[1] = v;
			andptr[2] = v>>8;
			andptr[3] = v>>16;
			andptr[4] = v>>24;
			andptr += 5;
			return;
		}
		if(t == D_SP) {
			if(v == 0) {
				*andptr++ = (0 << 6) | (4 << 0) | (r << 3);
				asmidx(a, D_SP);
				return;
			}
			if(v >= -128 && v < 128) {
				*andptr++ = (1 << 6) | (4 << 0) | (r << 3);
				asmidx(a, D_SP);
				*andptr++ = v;
				return;
			}
			*andptr++ = (2 << 6) | (4 << 0) | (r << 3);
			asmidx(a, D_SP);
			andptr[0] = v;
			andptr[1] = v>>8;
			andptr[2] = v>>16;
			andptr[3] = v>>24;
			andptr += 4;
			return;
		}
		if(t >= D_AX && t <= D_DI) {
			if(v == 0 && t != D_BP) {
				*andptr++ = (0 << 6) | (reg[t] << 0) | (r << 3);
				return;
			}
			if(v >= -128 && v < 128) {
				andptr[0] = (1 << 6) | (reg[t] << 0) | (r << 3);
				andptr[1] = v;
				andptr += 2;
				return;
			}
			andptr[0] = (2 << 6) | (reg[t] << 0) | (r << 3);
			andptr[1] = v;
			andptr[2] = v>>8;
			andptr[3] = v>>16;
			andptr[4] = v>>24;
			andptr += 5;
			return;
		}
		goto bad;
	}
	switch(a->type) {
	default:
		goto bad;
	case D_STATIC:
	case D_EXTERN:
		aa.type = D_NONE+D_INDIR;
		break;
	case D_AUTO:
	case D_PARAM:
		aa.type = D_SP+D_INDIR;
		break;
	}
	aa.index = D_NONE;
	aa.scale = 1;
	aa.offset = vaddr(a);
	asmand(&aa, r);
	return;
bad:
	diag("asmand: bad address %D\n", a);
	return;
}

#define	E	0xff
uchar	ymovtab[] =
{
/* push */
	APUSHL,	Ycs,	Ynone,	0,	0x0e,E,0,0,
	APUSHL,	Yss,	Ynone,	0,	0x16,E,0,0,
	APUSHL,	Yds,	Ynone,	0,	0x1e,E,0,0,
	APUSHL,	Yes,	Ynone,	0,	0x06,E,0,0,
	APUSHL,	Yfs,	Ynone,	0,	0x0f,0xa0,E,0,
	APUSHL,	Ygs,	Ynone,	0,	0x0f,0xa8,E,0,

	APUSHW,	Ycs,	Ynone,	0,	Pe,0x0e,E,0,
	APUSHW,	Yss,	Ynone,	0,	Pe,0x16,E,0,
	APUSHW,	Yds,	Ynone,	0,	Pe,0x1e,E,0,
	APUSHW,	Yes,	Ynone,	0,	Pe,0x06,E,0,
	APUSHW,	Yfs,	Ynone,	0,	Pe,0x0f,0xa0,E,
	APUSHW,	Ygs,	Ynone,	0,	Pe,0x0f,0xa8,E,

/* pop */
	APOPL,	Ynone,	Yds,	0,	0x1f,E,0,0,
	APOPL,	Ynone,	Yes,	0,	0x07,E,0,0,
	APOPL,	Ynone,	Yss,	0,	0x17,E,0,0,
	APOPL,	Ynone,	Yfs,	0,	0x0f,0xa1,E,0,
	APOPL,	Ynone,	Ygs,	0,	0x0f,0xa9,E,0,

	APOPW,	Ynone,	Yds,	0,	Pe,0x1f,E,0,
	APOPW,	Ynone,	Yes,	0,	Pe,0x07,E,0,
	APOPW,	Ynone,	Yss,	0,	Pe,0x17,E,0,
	APOPW,	Ynone,	Yfs,	0,	Pe,0x0f,0xa1,E,
	APOPW,	Ynone,	Ygs,	0,	Pe,0x0f,0xa9,E,

/* mov seg */
	AMOVW,	Yes,	Yml,	1,	0x8c,0,0,0,
	AMOVW,	Ycs,	Yml,	1,	0x8c,1,0,0,
	AMOVW,	Yss,	Yml,	1,	0x8c,2,0,0,
	AMOVW,	Yds,	Yml,	1,	0x8c,3,0,0,
	AMOVW,	Yfs,	Yml,	1,	0x8c,4,0,0,
	AMOVW,	Ygs,	Yml,	1,	0x8c,5,0,0,

	AMOVW,	Yml,	Yes,	2,	0x8e,0,0,0,
	AMOVW,	Yml,	Ycs,	2,	0x8e,1,0,0,
	AMOVW,	Yml,	Yss,	2,	0x8e,2,0,0,
	AMOVW,	Yml,	Yds,	2,	0x8e,3,0,0,
	AMOVW,	Yml,	Yfs,	2,	0x8e,4,0,0,
	AMOVW,	Yml,	Ygs,	2,	0x8e,5,0,0,

/* mov cr */
	AMOVL,	Ycr0,	Yml,	3,	0x0f,0x20,0,0,
	AMOVL,	Ycr2,	Yml,	3,	0x0f,0x20,2,0,
	AMOVL,	Ycr3,	Yml,	3,	0x0f,0x20,3,0,

	AMOVL,	Yml,	Ycr0,	4,	0x0f,0x22,0,0,
	AMOVL,	Yml,	Ycr2,	4,	0x0f,0x22,2,0,
	AMOVL,	Yml,	Ycr3,	4,	0x0f,0x22,3,0,

/* mov dr */
	AMOVL,	Ydr0,	Yml,	3,	0x0f,0x21,0,0,
	AMOVL,	Ydr6,	Yml,	3,	0x0f,0x21,6,0,
	AMOVL,	Ydr7,	Yml,	3,	0x0f,0x21,7,0,

	AMOVL,	Yml,	Ydr0,	4,	0x0f,0x23,0,0,
	AMOVL,	Yml,	Ydr6,	4,	0x0f,0x23,6,0,
	AMOVL,	Yml,	Ydr7,	4,	0x0f,0x23,7,0,

/* mov tr */
	AMOVL,	Ytr6,	Yml,	3,	0x0f,0x24,6,0,
	AMOVL,	Ytr7,	Yml,	3,	0x0f,0x24,7,0,

	AMOVL,	Yml,	Ytr6,	4,	0x0f,0x26,6,E,
	AMOVL,	Yml,	Ytr7,	4,	0x0f,0x26,7,E,

/* lgdt, sgdt, lidt, sidt */
	AMOVL,	Ym,	Ygdtr,	4,	0x0f,0x01,2,0,
	AMOVL,	Ygdtr,	Ym,	3,	0x0f,0x01,0,0,
	AMOVL,	Ym,	Yidtr,	4,	0x0f,0x01,3,0,
	AMOVL,	Yidtr,	Ym,	3,	0x0f,0x01,1,0,

/* lldt, sldt */
	AMOVW,	Yml,	Yldtr,	4,	0x0f,0x00,2,0,
	AMOVW,	Yldtr,	Yml,	3,	0x0f,0x00,0,0,

/* lmsw, smsw */
	AMOVW,	Yml,	Ymsw,	4,	0x0f,0x01,6,0,
	AMOVW,	Ymsw,	Yml,	3,	0x0f,0x01,4,0,

/* ltr, str */
	AMOVW,	Yml,	Ytask,	4,	0x0f,0x00,3,0,
	AMOVW,	Ytask,	Yml,	3,	0x0f,0x00,1,0,

/* load full pointer */
	AMOVL,	Yml,	Ycol,	5,	0,0,0,0,
	AMOVW,	Yml,	Ycol,	5,	Pe,0,0,0,

/* double shift */
	ASHLL,	Ycol,	Yml,	6,	0xa4,0xa5,0,0,
	ASHRL,	Ycol,	Yml,	6,	0xac,0xad,0,0,

	0
};

int
isax(Adr *a)
{

	switch(a->type) {
	case D_AX:
	case D_AL:
	case D_AH:
	case D_INDIR+D_AX:
		return 1;
	}
	if(a->index == D_AX)
		return 1;
	return 0;
}

void
subreg(Prog *p, int from, int to)
{

	if(debug['Q'])
		print("\n%P	s/%R/%R/\n", p, from, to);

	if(p->from.type == from)
		p->from.type = to;
	if(p->to.type == from)
		p->to.type = to;

	if(p->from.index == from)
		p->from.index = to;
	if(p->to.index == from)
		p->to.index = to;

	from += D_INDIR;
	if(p->from.type == from)
		p->from.type = to+D_INDIR;
	if(p->to.type == from)
		p->to.type = to+D_INDIR;

	if(debug['Q'])
		print("%P\n", p);
}

void
doasm(Prog *p)
{
	Optab *o;
	Prog *q, pp;
	uchar *t;
	int z, op, ft, tt;
	long v;

	o = &optab[p->as];
	ft = oclass(&p->from) * Ymax;
	tt = oclass(&p->to) * Ymax;
	t = o->ytab;
	if(t == 0) {
		diag("asmins: noproto %P\n", p);
		return;
	}
	for(z=0; *t; z+=t[3],t+=4)
		if(ycover[ft+t[0]])
		if(ycover[tt+t[1]])
			goto found;
	goto domov;

found:
	switch(o->prefix) {
	case Pq:	/* 16 bit escape and opcode escape */
		*andptr++ = Pe;
		*andptr++ = Pm;
		break;

	case Pm:	/* opcode escape */
		*andptr++ = Pm;
		break;

	case Pe:	/* 16 bit escape */
		*andptr++ = Pe;
		break;

	case Pb:	/* botch */
		break;
	}
	v = vaddr(&p->from);
	op = o->op[z];
	switch(t[2]) {
	default:
		diag("asmins: unknown z %d %P\n", t[2], p);
		return;

	case Zpseudo:
		break;

	case Zlit:
		for(; op = o->op[z]; z++)
			*andptr++ = op;
		break;

	case Zm_r:
		*andptr++ = op;
		asmand(&p->from, reg[p->to.type]);
		break;

	case Zm_o:
		*andptr++ = op;
		asmand(&p->from, o->op[z+1]);
		break;

	case Zr_m:
		*andptr++ = op;
		asmand(&p->to, reg[p->from.type]);
		break;

	case Zo_m:
		*andptr++ = op;
		asmand(&p->to, o->op[z+1]);
		break;

	case Zm_ibo:
		v = vaddr(&p->to);
		*andptr++ = op;
		asmand(&p->from, o->op[z+1]);
		*andptr++ = v;
		break;

	case Zibo_m:
		*andptr++ = op;
		asmand(&p->to, o->op[z+1]);
		*andptr++ = v;
		break;

	case Z_ib:
		v = vaddr(&p->to);
	case Zib_:
		*andptr++ = op;
		*andptr++ = v;
		break;

	case Zib_rp:
		*andptr++ = op + reg[p->to.type];
		*andptr++ = v;
		break;

	case Zil_rp:
		*andptr++ = op + reg[p->to.type];
		*andptr++ = v;
		*andptr++ = v>>8;
		if(o->prefix != Pe) {
			*andptr++ = v>>16;
			*andptr++ = v>>24;
		}
		break;

	case Z_il:
		v = vaddr(&p->to);
	case Zil_:
		*andptr++ = op;
		*andptr++ = v;
		*andptr++ = v>>8;
		if(o->prefix != Pe) {
			*andptr++ = v>>16;
			*andptr++ = v>>24;
		}
		break;

	case Zm_ilo:
		v = vaddr(&p->to);
		*andptr++ = op;
		asmand(&p->from, o->op[z+1]);
		*andptr++ = v;
		*andptr++ = v>>8;
		if(o->prefix != Pe) {
			*andptr++ = v>>16;
			*andptr++ = v>>24;
		}
		break;

	case Zilo_m:
		*andptr++ = op;
		asmand(&p->to, o->op[z+1]);
		*andptr++ = v;
		*andptr++ = v>>8;
		if(o->prefix != Pe) {
			*andptr++ = v>>16;
			*andptr++ = v>>24;
		}
		break;

	case Z_rp:
		*andptr++ = op + reg[p->to.type];
		break;

	case Zrp_:
		*andptr++ = op + reg[p->from.type];
		break;

	case Zclr:
		*andptr++ = op;
		asmand(&p->to, reg[p->to.type]);
		break;

	case Zbr:
		q = p->cond;
		if(q) {
			v = q->pc - p->pc - 2;
			if(v >= -128 && v <= 127) {
				*andptr++ = op;
				*andptr++ = v;
			} else {
				v -= 6-2;
				*andptr++ = 0x0f;
				*andptr++ = o->op[z+1];
				*andptr++ = v;
				*andptr++ = v>>8;
				*andptr++ = v>>16;
				*andptr++ = v>>24;
			}
		}
		break;

	case Zcall:
		q = p->cond;
		if(q) {
			v = q->pc - p->pc - 5;
			*andptr++ = op;
			*andptr++ = v;
			*andptr++ = v>>8;
			*andptr++ = v>>16;
			*andptr++ = v>>24;
		}
		break;

	case Zjmp:
		q = p->cond;
		if(q) {
			v = q->pc - p->pc - 2;
			if(v >= -128 && v <= 127) {
				*andptr++ = op;
				*andptr++ = v;
			} else {
				v -= 5-2;
				*andptr++ = o->op[z+1];
				*andptr++ = v;
				*andptr++ = v>>8;
				*andptr++ = v>>16;
				*andptr++ = v>>24;
			}
		}
		break;

	case Zloop:
		q = p->cond;
		if(q) {
			v = q->pc - p->pc - 2;
			if(v < -128 && v > 127)
				diag("loop too far: %P\n", p);
			*andptr++ = op;
			*andptr++ = v;
		}
		break;

	case Zbyte:
		*andptr++ = v;
		if(op > 1) {
			*andptr++ = v>>8;
			if(op > 2) {
				*andptr++ = v>>16;
				*andptr++ = v>>24;
			}
		}
		break;

	case Zmov:
		goto domov;
	}
	return;

domov:
	for(t=ymovtab; *t; t+=8)
		if(p->as == t[0])
		if(ycover[ft+t[1]])
		if(ycover[tt+t[2]])
			goto mfound;
bad:
	/*
	 * here, the assembly has failed.
	 * if its a byte instruction that has
	 * unaddressable registers, try to
	 * exchange registers and reissue the
	 * instruction with the operands renamed.
	 */
	pp = *p;
	z = p->from.type;
	if(z >= D_BP && z <= D_DI) {
		if(isax(&p->to)) {
			*andptr++ = 0x87;			/* xchg lhs,bx */
			asmand(&p->from, reg[D_BX]);
			subreg(&pp, z, D_BX);
			doasm(&pp);
			*andptr++ = 0x87;			/* xchg lhs,bx */
			asmand(&p->from, reg[D_BX]);
		} else {
			*andptr++ = 0x90 + reg[z];		/* xchg lsh,ax */
			subreg(&pp, z, D_AX);
			doasm(&pp);
			*andptr++ = 0x90 + reg[z];		/* xchg lsh,ax */
		}
		return;
	}
	z = p->to.type;
	if(z >= D_BP && z <= D_DI) {
		if(isax(&p->from)) {
			*andptr++ = 0x87;			/* xchg rhs,bx */
			asmand(&p->to, reg[D_BX]);
			subreg(&pp, z, D_BX);
			doasm(&pp);
			*andptr++ = 0x87;			/* xchg rhs,bx */
			asmand(&p->to, reg[D_BX]);
		} else {
			*andptr++ = 0x90 + reg[z];		/* xchg rsh,ax */
			subreg(&pp, z, D_AX);
			doasm(&pp);
			*andptr++ = 0x90 + reg[z];		/* xchg rsh,ax */
		}
		return;
	}
	diag("doasm: notfound %P\n", p);
	return;

mfound:
	switch(t[3]) {
	default:
		diag("asmins: unknown mov %d %P\n", t[3], p);
		break;

	case 0:	/* lit */
		for(z=4; t[z]!=E; z++)
			*andptr++ = t[z];
		break;

	case 1:	/* r,m */
		*andptr++ = t[4];
		asmand(&p->to, t[5]);
		break;

	case 2:	/* m,r */
		*andptr++ = t[4];
		asmand(&p->from, t[5]);
		break;

	case 3:	/* r,m - 2op */
		*andptr++ = t[4];
		*andptr++ = t[5];
		asmand(&p->to, t[6]);
		break;

	case 4:	/* m,r - 2op */
		*andptr++ = t[4];
		*andptr++ = t[5];
		asmand(&p->from, t[6]);
		break;

	case 5:	/* load full pointer, trash heap */
		if(t[4])
			*andptr++ = t[4];
		switch(p->to.index) {
		default:
			goto bad;
		case D_DS:
			*andptr++ = 0xc5;
			break;
		case D_SS:
			*andptr++ = 0x0f;
			*andptr++ = 0xb2;
			break;
		case D_ES:
			*andptr++ = 0xc4;
			break;
		case D_FS:
			*andptr++ = 0x0f;
			*andptr++ = 0xb4;
			break;
		case D_GS:
			*andptr++ = 0x0f;
			*andptr++ = 0xb5;
			break;
		}
		asmand(&p->from, reg[p->to.type]);
		break;

	case 6:	/* load full pointer, trash heap */
		z = p->from.type;
		switch(z) {
		default:
			goto bad;
		case D_CONST:
			*andptr++ = 0x0f;
			*andptr++ = t[4];
			asmand(&p->to, reg[p->from.type]);
			*andptr++ = p->from.offset;
			break;
		case D_CL:
		case D_CX:
			*andptr++ = 0x0f;
			*andptr++ = t[5];
			asmand(&p->to, reg[p->from.type]);
			break;
		}
		break;
	}
}

void
asmins(Prog *p)
{

	andptr = and;
	doasm(p);

}