update 8a, 8c, 8l to use new object format.

add "extern register" support to 8c.
extern register means allocate in the FS-relative segment.

make 8l generate segmented stack checks.

R=ken
OCL=26600
CL=26606

18 files changed, 543 insertions, 144 deletions

diff --git a/src/cmd/8a/a.h b/src/cmd/8a/a.h
index 8b5f7d9ef..dbe439929 100644
--- a/src/cmd/8a/a.h
+++ b/src/cmd/8a/a.h
@@ -32,7 +32,6 @@
 #include <libc.h>
 #include <bio.h>
 #include "../8l/8.out.h"
-#include "compat.h"
 
 
 #ifndef	EXTERN
@@ -104,6 +103,7 @@ struct	Gen
 	double	dval;
 	char	sval[8];
 	int32	offset;
+	int32	offset2;
 	Sym*	sym;
 	short	type;
 	short	index;
diff --git a/src/cmd/8a/a.y b/src/cmd/8a/a.y
index 505defd28..fced3efa4 100644
--- a/src/cmd/8a/a.y
+++ b/src/cmd/8a/a.y
@@ -34,6 +34,10 @@
 %union	{
 	Sym	*sym;
 	int32	lval;
+	struct {
+		int32 v1;
+		int32 v2;
+	} con2;
 	double	dval;
 	char	sval[8];
 	Gen	gen;
@@ -46,16 +50,17 @@
 %left	'+' '-'
 %left	'*' '/' '%'
 %token	<lval>	LTYPE0 LTYPE1 LTYPE2 LTYPE3 LTYPE4
-%token	<lval>	LTYPEC LTYPED LTYPEN LTYPER LTYPET LTYPES LTYPEM LTYPEI
+%token	<lval>	LTYPEC LTYPED LTYPEN LTYPER LTYPET LTYPES LTYPEM LTYPEI LTYPEG
 %token	<lval>	LCONST LFP LPC LSB
 %token	<lval>	LBREG LLREG LSREG LFREG
 %token	<dval>	LFCONST
 %token	<sval>	LSCONST LSP
 %token	<sym>	LNAME LLAB LVAR
 %type	<lval>	con expr pointer offset
-%type	<gen>	mem imm reg nam rel rem rim rom omem nmem
+%type	<con2>	con2
+%type	<gen>	mem imm imm2 reg nam rel rem rim rom omem nmem
 %type	<gen2>	nonnon nonrel nonrem rimnon rimrem remrim
-%type	<gen2>	spec1 spec2 spec3 spec4 spec5 spec6 spec7
+%type	<gen2>	spec1 spec2 spec3 spec4 spec5 spec6 spec7 spec8
 %%
 prog:
 |	prog line
@@ -103,6 +108,7 @@ inst:
 |	LTYPES spec5	{ outcode($1, &$2); }
 |	LTYPEM spec6	{ outcode($1, &$2); }
 |	LTYPEI spec7	{ outcode($1, &$2); }
+|	LTYPEG spec8	{ outcode($1, &$2); }
 
 nonnon:
 	{
@@ -174,12 +180,12 @@ spec1:	/* DATA */
 	}
 
 spec2:	/* TEXT */
-	mem ',' imm
+	mem ',' imm2
 	{
 		$$.from = $1;
 		$$.to = $3;
 	}
-|	mem ',' con ',' imm
+|	mem ',' con ',' imm2
 	{
 		$$.from = $1;
 		$$.from.scale = $3;
@@ -249,6 +255,19 @@ spec7:
 		$$.to = $3;
 	}
 
+spec8:	/* GLOBL */
+	mem ',' imm
+	{
+		$$.from = $1;
+		$$.to = $3;
+	}
+|	mem ',' con ',' imm
+	{
+		$$.from = $1;
+		$$.from.scale = $3;
+		$$.to = $5;
+	}
+
 rem:
 	reg
 |	mem
@@ -365,6 +384,37 @@ imm:
 		$$.dval = -$3;
 	}
 
+imm2:
+	'$' con2
+	{
+		$$ = nullgen;
+		$$.type = D_CONST2;
+		$$.offset = $2.v1;
+		$$.offset2 = $2.v2;
+	}
+
+con2:
+	LCONST
+	{
+		$$.v1 = $1;
+		$$.v2 = 0;
+	}
+|	'-' LCONST
+	{
+		$$.v1 = -$2;
+		$$.v2 = 0;
+	}
+|	LCONST '-' LCONST
+	{
+		$$.v1 = $1;
+		$$.v2 = $3;
+	}
+|	'-' LCONST '-' LCONST
+	{
+		$$.v1 = -$2;
+		$$.v2 = $4;
+	}
+
 mem:
 	omem
 |	nmem
@@ -416,6 +466,12 @@ omem:
 		$$ = nullgen;
 		$$.type = D_INDIR+D_SP;
 	}
+|	con '(' LSREG ')'
+	{
+		$$ = nullgen;
+		$$.type = D_INDIR+$3;
+		$$.offset = $1;
+	}
 |	'(' LLREG '*' con ')'
 	{
 		$$ = nullgen;
diff --git a/src/cmd/8a/lex.c b/src/cmd/8a/lex.c
index 8576f01b7..fcc4cf964 100644
--- a/src/cmd/8a/lex.c
+++ b/src/cmd/8a/lex.c
@@ -165,6 +165,9 @@ assemble(char *file)
 
 	pass = 1;
 	pinit(file);
+	
+	Bprint(&obuf, "%s\n", thestring);
+
 	for(i=0; i<nDlist; i++)
 		dodefine(Dlist[i]);
 	yyparse();
@@ -172,6 +175,8 @@ assemble(char *file)
 		cclean();
 		return nerrors;
 	}
+	
+	Bprint(&obuf, "\n!\n");
 
 	pass = 2;
 	outhist();
@@ -304,6 +309,9 @@ struct
 	"CMPSB",	LTYPE0,	ACMPSB,
 	"CMPSL",	LTYPE0,	ACMPSL,
 	"CMPSW",	LTYPE0,	ACMPSW,
+	"CMPXCHGB",	LTYPE3,	ACMPXCHGB,
+	"CMPXCHGL",	LTYPE3,	ACMPXCHGL,
+	"CMPXCHGW",	LTYPE3,	ACMPXCHGW,	
 	"DAA",		LTYPE0,	ADAA,
 	"DAS",		LTYPE0,	ADAS,
 	"DATA",		LTYPED,	ADATA,
@@ -315,7 +323,7 @@ struct
 	"DIVW",		LTYPE2,	ADIVW,
 	"END",		LTYPE0,	AEND,
 	"ENTER",	LTYPE2,	AENTER,
-	"GLOBL",	LTYPET,	AGLOBL,
+	"GLOBL",	LTYPEG,	AGLOBL,
 	"HLT",		LTYPE0,	AHLT,
 	"IDIVB",	LTYPE2,	AIDIVB,
 	"IDIVL",	LTYPE2,	AIDIVL,
@@ -714,7 +722,7 @@ zname(char *n, int t, int s)
 void
 zaddr(Gen *a, int s)
 {
-	long l;
+	int32 l;
 	int i, t;
 	char *n;
 	Ieee e;
@@ -734,6 +742,9 @@ zaddr(Gen *a, int s)
 	case D_FCONST:
 		t |= T_FCONST;
 		break;
+	case D_CONST2:
+		t |= T_OFFSET|T_OFFSET2;
+		break;
 	case D_SCONST:
 		t |= T_SCONST;
 		break;
@@ -753,6 +764,13 @@ zaddr(Gen *a, int s)
 		Bputc(&obuf, l>>16);
 		Bputc(&obuf, l>>24);
 	}
+	if(t & T_OFFSET2) {
+		l = a->offset2;
+		Bputc(&obuf, l);
+		Bputc(&obuf, l>>8);
+		Bputc(&obuf, l>>16);
+		Bputc(&obuf, l>>24);
+	}
 	if(t & T_SYM)		/* implies sym */
 		Bputc(&obuf, s);
 	if(t & T_FCONST) {
diff --git a/src/cmd/8c/gc.h b/src/cmd/8c/gc.h
index c30eb55bc..55a37c42c 100644
--- a/src/cmd/8c/gc.h
+++ b/src/cmd/8c/gc.h
@@ -66,6 +66,7 @@ EXTERN	struct
 struct	Adr
 {
 	int32	offset;
+	int32	offset2;
 	double	dval;
 	char	sval[NSNAME];
 
diff --git a/src/cmd/8c/list.c b/src/cmd/8c/list.c
index d700b63af..ec5ac9d60 100644
--- a/src/cmd/8c/list.c
+++ b/src/cmd/8c/list.c
@@ -154,6 +154,10 @@ Dconv(Fmt *fp)
 		sprint(str, "$%ld", a->offset);
 		break;
 
+	case D_CONST2:
+		sprint(str, "$%ld-%ld", a->offset, a->offset2);
+		break;
+
 	case D_FCONST:
 		sprint(str, "$(%.17e)", a->dval);
 		break;
diff --git a/src/cmd/8c/sgen.c b/src/cmd/8c/sgen.c
index 14bd32adb..c4d5665d8 100644
--- a/src/cmd/8c/sgen.c
+++ b/src/cmd/8c/sgen.c
@@ -30,6 +30,31 @@
 
 #include "gc.h"
 
+int32
+argsize(void)
+{
+	Type *t;
+	int32 s;
+
+//print("t=%T\n", thisfn);
+	s = 0;
+	for(t=thisfn->down; t!=T; t=t->down) {
+		switch(t->etype) {
+		case TVOID:
+			break;
+		case TDOT:
+			s += 64;
+			break;
+		default:
+			s = align(s, t, Aarg1);
+			s = align(s, t, Aarg2);
+			break;
+		}
+//print("	%d %T\n", s, t);
+	}
+	return (s+7) & ~7;
+}
+
 void
 codgen(Node *n, Node *nn)
 {
@@ -52,7 +77,10 @@ codgen(Node *n, Node *nn)
 			break;
 	}
 	nearln = nn->lineno;
+
 	gpseudo(ATEXT, n1->sym, nodconst(stkoff));
+	p->to.type = D_CONST2;
+	p->to.offset2 = argsize();
 
 	/*
 	 * isolate first argument
@@ -494,6 +522,10 @@ xcom(Node *n)
 			n->addable = 11;
 		break;
 
+	case OEXREG:
+		n->addable = 10;
+		break;
+
 	case OREGISTER:
 		n->addable = 12;
 		break;
diff --git a/src/cmd/8c/swt.c b/src/cmd/8c/swt.c
index 13e9a51e3..6d693337a 100644
--- a/src/cmd/8c/swt.c
+++ b/src/cmd/8c/swt.c
@@ -328,7 +328,10 @@ outcode(void)
 		return;
 	}
 	Binit(&b, f, OWRITE);
-	Bseek(&b, 0L, 2);
+
+	Bprint(&b, "%s\n", thestring);
+	Bprint(&b, "!\n");
+
 	outhist(&b);
 	for(sym=0; sym<NSYM; sym++) {
 		h[sym].sym = S;
@@ -530,6 +533,9 @@ zaddr(Biobuf *b, Adr *a, int s)
 	case D_SCONST:
 		t |= T_SCONST;
 		break;
+	case D_CONST2:
+		t |= T_OFFSET|T_OFFSET2;
+		break;
 	}
 	Bputc(b, t);
 
@@ -544,6 +550,13 @@ zaddr(Biobuf *b, Adr *a, int s)
 		Bputc(b, l>>16);
 		Bputc(b, l>>24);
 	}
+	if(t & T_OFFSET2) {	/* implies offset2 */
+		l = a->offset2;
+		Bputc(b, l);
+		Bputc(b, l>>8);
+		Bputc(b, l>>16);
+		Bputc(b, l>>24);
+	}
 	if(t & T_SYM)		/* implies sym */
 		Bputc(b, s);
 	if(t & T_FCONST) {
diff --git a/src/cmd/8c/txt.c b/src/cmd/8c/txt.c
index 4e95f3c7a..5cc43e0e9 100644
--- a/src/cmd/8c/txt.c
+++ b/src/cmd/8c/txt.c
@@ -449,6 +449,10 @@ naddr(Node *n, Adr *a)
 		a->sym = S;
 		break;
 
+	case OEXREG:
+		a->type = D_INDIR + D_FS;
+		a->offset = n->reg - 1;
+		break;
 
 	case OIND:
 		naddr(n->left, a);
@@ -1389,6 +1393,15 @@ sconst(Node *n)
 int32
 exreg(Type *t)
 {
+	int32 o;
+
+	if(typechlp[t->etype]){
+		if(exregoffset >= 32)
+			return 0;
+		o = exregoffset;
+		exregoffset += 4;
+		return o+1;	// +1 to avoid 0 == failure; naddr case OEXREG will -1.
+	}
 
 	USED(t);
 	return 0;
diff --git a/src/cmd/8l/8.out.h b/src/cmd/8l/8.out.h
index 647c1dd59..f7750129d 100644
--- a/src/cmd/8l/8.out.h
+++ b/src/cmd/8l/8.out.h
@@ -32,6 +32,7 @@
 #define	NSNAME	8
 #define NOPROF	(1<<0)
 #define DUPOK	(1<<1)
+#define NOSPLIT	(1<<2)
 
 enum	as
 {
@@ -78,6 +79,9 @@ enum	as
 	ACMPSB,
 	ACMPSL,
 	ACMPSW,
+	ACMPXCHGB,
+	ACMPXCHGL,
+	ACMPXCHGW,
 	ADAA,
 	ADAS,
 	ADATA,
@@ -437,6 +441,7 @@ enum
 	D_FCONST	= 66,
 	D_SCONST	= 67,
 	D_ADDR		= 68,
+	D_CONST2	= 69,
 
 	D_FILE,
 	D_FILE1,
@@ -449,6 +454,7 @@ enum
 	T_FCONST	= 1<<3,
 	T_SYM		= 1<<4,
 	T_SCONST	= 1<<5,
+	T_OFFSET2	= 1<<6,
 
 	REGARG		= 0,
 	REGRET		= D_AX,
diff --git a/src/cmd/8l/l.h b/src/cmd/8l/l.h
index fabf89034..114c66369 100644
--- a/src/cmd/8l/l.h
+++ b/src/cmd/8l/l.h
@@ -67,6 +67,7 @@ struct	Adr
 	short	type;
 	uchar	index;
 	char	scale;
+	int32	offset2;
 };
 
 #define	offset	u0.u0offset
@@ -342,7 +343,7 @@ void	histtoauto(void);
 double	ieeedtod(Ieee*);
 int32	ieeedtof(Ieee*);
 void	import(void);
-void	ldobj(int, int32, char*);
+void	ldobj(Biobuf*, int32, char*);
 void	loadlib(void);
 void	listinit(void);
 Sym*	lookup(char*, int);
@@ -368,7 +369,7 @@ int32	vaddr(Adr*);
 void	wput(ushort);
 void	xdefine(char*, int, int32);
 void	xfol(Prog*);
-int	zaddr(uchar*, Adr*, Sym*[]);
+void	zaddr(Biobuf*, Adr*, Sym*[]);
 void	zerosig(char*);
 uint32	machheadr(void);
 uint32	elfheadr(void);
diff --git a/src/cmd/8l/list.c b/src/cmd/8l/list.c
index ebd8b864b..873f7c1f8 100644
--- a/src/cmd/8l/list.c
+++ b/src/cmd/8l/list.c
@@ -82,6 +82,14 @@ Aconv(Fmt *fp)
 	return fmtstrcpy(fp, anames[i]);
 }
 
+char*
+xsymname(Sym *s)
+{
+	if(s == nil)
+		return "!!noname!!";
+	return s->name;
+}
+
 int
 Dconv(Fmt *fp)
 {
@@ -120,16 +128,16 @@ Dconv(Fmt *fp)
 		break;
 
 	case D_EXTERN:
-		sprint(str, "%s+%ld(SB)", a->sym->name, a->offset);
+		sprint(str, "%s+%ld(SB)", xsymname(a->sym), a->offset);
 		break;
 
 	case D_STATIC:
-		sprint(str, "%s<%d>+%ld(SB)", a->sym->name,
+		sprint(str, "%s<%d>+%ld(SB)", xsymname(a->sym),
 			a->sym->version, a->offset);
 		break;
 
 	case D_AUTO:
-		sprint(str, "%s+%ld(SP)", a->sym->name, a->offset);
+		sprint(str, "%s+%ld(SP)", xsymname(a->sym), a->offset);
 		break;
 
 	case D_PARAM:
@@ -143,6 +151,10 @@ Dconv(Fmt *fp)
 		sprint(str, "$%ld", a->offset);
 		break;
 
+	case D_CONST2:
+		sprint(str, "$%ld-%ld", a->offset, a->offset2);
+		break;
+
 	case D_FCONST:
 		sprint(str, "$(%.8lux,%.8lux)", a->ieee.h, a->ieee.l);
 		break;
diff --git a/src/cmd/8l/obj.c b/src/cmd/8l/obj.c
index a1d14318d..022f315d2 100644
--- a/src/cmd/8l/obj.c
+++ b/src/cmd/8l/obj.c
@@ -419,14 +419,15 @@ void
 objfile(char *file)
 {
 	int32 off, esym, cnt, l;
-	int f, work;
+	int work;
+	Biobuf *f;
 	Sym *s;
 	char magbuf[SARMAG];
 	char name[100], pname[150];
 	struct ar_hdr arhdr;
 	char *e, *start, *stop;
 
-	if(file[0] == '-' && file[1] == 'l') {
+	if(file[0] == '-' && file[1] == 'l') {	// TODO: fix this
 		if(debug['9'])
 			sprint(name, "/%s/lib/lib", thestring);
 		else
@@ -438,22 +439,22 @@ objfile(char *file)
 	if(debug['v'])
 		Bprint(&bso, "%5.2f ldobj: %s\n", cputime(), file);
 	Bflush(&bso);
-	f = open(file, 0);
-	if(f < 0) {
+	f = Bopen(file, 0);
+	if(f == nil) {
 		diag("cannot open file: %s", file);
 		errorexit();
 	}
-	l = read(f, magbuf, SARMAG);
+	l = Bread(f, magbuf, SARMAG);
 	if(l != SARMAG || strncmp(magbuf, ARMAG, SARMAG)){
 		/* load it as a regular file */
-		l = seek(f, 0L, 2);
-		seek(f, 0L, 0);
+		l = Bseek(f, 0L, 2);
+		Bseek(f, 0L, 0);
 		ldobj(f, l, file);
-		close(f);
+		Bterm(f);
 		return;
 	}
 
-	l = read(f, &arhdr, SAR_HDR);
+	l = Bread(f, &arhdr, SAR_HDR);
 	if(l != SAR_HDR) {
 		diag("%s: short read on archive file symbol header", file);
 		goto out;
@@ -469,12 +470,12 @@ objfile(char *file)
 	/*
 	 * just bang the whole symbol file into memory
 	 */
-	seek(f, off, 0);
+	Bseek(f, off, 0);
 	cnt = esym - off;
 	start = malloc(cnt + 10);
-	cnt = read(f, start, cnt);
+	cnt = Bread(f, start, cnt);
 	if(cnt <= 0){
-		close(f);
+		Bterm(f);
 		return;
 	}
 	stop = &start[cnt];
@@ -498,12 +499,16 @@ objfile(char *file)
 			l |= (e[2] & 0xff) << 8;
 			l |= (e[3] & 0xff) << 16;
 			l |= (e[4] & 0xff) << 24;
-			seek(f, l, 0);
-			l = read(f, &arhdr, SAR_HDR);
+			Bseek(f, l, 0);
+			l = Bread(f, &arhdr, SAR_HDR);
 			if(l != SAR_HDR)
 				goto bad;
 			if(strncmp(arhdr.fmag, ARFMAG, sizeof(arhdr.fmag)))
 				goto bad;
+			l = SARNAME;
+			while(l > 0 && arhdr.name[l-1] == ' ')
+				l--;
+			sprint(pname, "%s(%.*s)", file, l, arhdr.name);
 			l = atolwhex(arhdr.size);
 			ldobj(f, l, pname);
 			if(s->type == SXREF) {
@@ -519,85 +524,87 @@ objfile(char *file)
 bad:
 	diag("%s: bad or out of date archive", file);
 out:
-	close(f);
+	Bterm(f);
 }
 
-int
-zaddr(uchar *p, Adr *a, Sym *h[])
+int32
+Bget4(Biobuf *f)
+{
+	uchar p[4];
+
+	if(Bread(f, p, 4) != 4)
+		return 0;
+	return p[0] | (p[1] << 8) | (p[2] << 16) | (p[3] << 24);
+}
+
+void
+zaddr(Biobuf *f, Adr *a, Sym *h[])
 {
-	int c, t, i;
+	int t;
 	int32 l;
 	Sym *s;
 	Auto *u;
 
-	t = p[0];
-
-	c = 1;
+	t = Bgetc(f);
 	if(t & T_INDEX) {
-		a->index = p[c];
-		a->scale = p[c+1];
-		c += 2;
+		a->index = Bgetc(f);
+		a->scale = Bgetc(f);
 	} else {
 		a->index = D_NONE;
 		a->scale = 0;
 	}
+	a->type = D_NONE;
 	a->offset = 0;
-	if(t & T_OFFSET) {
-		a->offset = p[c] | (p[c+1]<<8) | (p[c+2]<<16) | (p[c+3]<<24);
-		c += 4;
+	if(t & T_OFFSET)
+		a->offset = Bget4(f);
+	a->offset2 = 0;
+	if(t & T_OFFSET2) {
+		a->offset2 = Bget4(f);
+		a->type = D_CONST2;
 	}
 	a->sym = S;
-	if(t & T_SYM) {
-		a->sym = h[p[c]];
-		c++;
-	}
-	a->type = D_NONE;
+	if(t & T_SYM)
+		a->sym = h[Bgetc(f)];
 	if(t & T_FCONST) {
-		a->ieee.l = p[c] | (p[c+1]<<8) | (p[c+2]<<16) | (p[c+3]<<24);
-		a->ieee.h = p[c+4] | (p[c+5]<<8) | (p[c+6]<<16) | (p[c+7]<<24);
-		c += 8;
+		a->ieee.l = Bget4(f);
+		a->ieee.h = Bget4(f);
 		a->type = D_FCONST;
 	} else
 	if(t & T_SCONST) {
-		for(i=0; i<NSNAME; i++)
-			a->scon[i] = p[c+i];
-		c += NSNAME;
+		Bread(f, a->scon, NSNAME);
 		a->type = D_SCONST;
 	}
-	if(t & T_TYPE) {
-		a->type = p[c];
-		c++;
-	}
+	if(t & T_TYPE)
+		a->type = Bgetc(f);
 	s = a->sym;
 	if(s == S)
-		return c;
+		return;
 
 	t = a->type;
 	if(t != D_AUTO && t != D_PARAM)
-		return c;
+		return;
 	l = a->offset;
 	for(u=curauto; u; u=u->link) {
 		if(u->asym == s)
 		if(u->type == t) {
 			if(u->aoffset > l)
 				u->aoffset = l;
-			return c;
+			return;
 		}
 	}
 
-	u = mal(sizeof(Auto));
+	u = mal(sizeof(*u));
 	u->link = curauto;
 	curauto = u;
 	u->asym = s;
 	u->aoffset = l;
 	u->type = t;
-	return c;
 }
 
 void
-addlib(char *obj)
+addlib(char *src, char *obj)
 {
-	char name[1024], comp[256], *p;
+	char name[1024], comp[256], *p, *q;
 	int i;
 
 	if(histfrogp <= 0)
@@ -645,6 +652,23 @@ addlib(char *obj)
 		strcat(name, "/");
 		strcat(name, comp);
 	}
+	if(debug['v'])
+		Bprint(&bso, "%5.2f addlib: %s %s pulls in %s\n", cputime(), obj, src, name);
+
+	p = strrchr(src, '/');
+	q = strrchr(name, '/');
+	if(p != nil && q != nil && p - src == q - name && memcmp(src, name, p - src) == 0) {
+		// leading paths are the same.
+		// if the source file refers to an object in its own directory
+		// and we are inside an archive, ignore the reference, in the hope
+		// that the archive contains that object too.
+		if(strchr(obj, '(')) {
+			if(debug['v'])
+				Bprint(&bso, "%5.2f ignored srcdir object %s\n", cputime(), name);
+			return;
+		}
+	}
+
 	for(i=0; i<libraryp; i++)
 		if(strcmp(name, library[i]) == 0)
 			return;
@@ -663,6 +687,22 @@ addlib(char *obj)
 }
 
 void
+copyhistfrog(char *buf, int nbuf)
+{
+	char *p, *ep;
+	int i;
+
+	p = buf;
+	ep = buf + nbuf;
+	i = 0;
+	for(i=0; i<histfrogp; i++) {
+		p = seprint(p, ep, "%s", histfrog[i]->name+1);
+		if(i+1<histfrogp && (p == buf || p[-1] != '/'))
+			p = seprint(p, ep, "/");
+	}
+}
+
+void
 addhist(int32 line, int type)
 {
 	Auto *u;
@@ -679,6 +719,7 @@ addhist(int32 line, int type)
 	u->link = curhist;
 	curhist = u;
 
+	s->name[0] = 0;
 	j = 1;
 	for(i=0; i<histfrogp; i++) {
 		k = histfrog[i]->value;
@@ -686,6 +727,8 @@ addhist(int32 line, int type)
 		s->name[j+1] = k;
 		j += 2;
 	}
+	s->name[j] = 0;
+	s->name[j+1] = 0;
 }
 
 void
@@ -746,35 +789,27 @@ nopout(Prog *p)
 	p->to.type = D_NONE;
 }
 
-uchar*
-readsome(int f, uchar *buf, uchar *good, uchar *stop, int max)
-{
-	int n;
-
-	n = stop - good;
-	memmove(buf, good, stop - good);
-	stop = buf + n;
-	n = MAXIO - n;
-	if(n > max)
-		n = max;
-	n = read(f, stop, n);
-	if(n <= 0)
-		return 0;
-	return stop + n;
-}
-
 void
-ldobj(int f, int32 c, char *pn)
+ldobj(Biobuf *f, int32 len, char *pn)
 {
 	int32 ipc;
 	Prog *p, *t;
-	uchar *bloc, *bsize, *stop;
 	int v, o, r, skip;
 	Sym *h[NSYM], *s, *di;
 	uint32 sig;
 	static int files;
 	static char **filen;
 	char **nfilen;
+	int ntext, n, c1, c2, c3;
+	int32 eof;
+	int32 import0, import1;
+	char *line, *name;
+	char src[1024];
+
+	src[0] = '\0';
+	eof = Boffset(f) + len;
+
+	ntext = 0;
 
 	if((files&15) == 0){
 		nfilen = malloc((files+16)*sizeof(char*));
@@ -782,12 +817,48 @@ ldobj(int f, int32 c, char *pn)
 		free(filen);
 		filen = nfilen;
 	}
-	filen[files++] = strdup(pn);
+	pn = strdup(pn);
+	filen[files++] = pn;
 
-	bsize = buf.xbuf;
-	bloc = buf.xbuf;
 	di = S;
 
+	/* check the header */
+	line = Brdline(f, '\n');
+	if(line == nil) {
+		if(Blinelen(f) > 0) {
+			diag("%s: malformed object file", pn);
+			return;
+		}
+		goto eof;
+	}
+	n = Blinelen(f) - 1;
+	if(n != strlen(thestring) || strncmp(line, thestring, n) != 0) {
+		if(line)
+			line[n] = '\0';
+		diag("file not %s [%s]\n", thestring, line);
+		return;
+	}
+
+	/* skip over exports and other info -- ends with \n!\n */
+	import0 = Boffset(f);
+	c1 = '\n';	// the last line ended in \n
+	c2 = Bgetc(f);
+	c3 = Bgetc(f);
+	while(c1 != '\n' || c2 != '!' || c3 != '\n') {
+		c1 = c2;
+		c2 = c3;
+		c3 = Bgetc(f);
+		if(c3 == Beof)
+			goto eof;
+	}
+	import1 = Boffset(f);
+
+	Bseek(f, import0, 0);
+//	ldpkg(f, import1 - import0 - 2, pn);	// -2 for !\n
+	Bseek(f, import1, 0);
+
+print("import %ld-%ld\n", import0, import1);
+
 newloop:
 	memset(h, 0, sizeof(h));
 	version++;
@@ -796,61 +867,46 @@ newloop:
 	skip = 0;
 
 loop:
-	if(c <= 0)
+	if(f->state == Bracteof || Boffset(f) >= eof)
 		goto eof;
-	r = bsize - bloc;
-	if(r < 100 && r < c) {		/* enough for largest prog */
-		bsize = readsome(f, buf.xbuf, bloc, bsize, c);
-		if(bsize == 0)
-			goto eof;
-		bloc = buf.xbuf;
-		goto loop;
-	}
-	o = bloc[0] | (bloc[1] << 8);
+	o = Bgetc(f);
+	if(o == Beof)
+		goto eof;
+	o |= Bgetc(f) << 8;
 	if(o <= AXXX || o >= ALAST) {
 		if(o < 0)
 			goto eof;
-		diag("%s: opcode out of range %d", pn, o);
-		print("	probably not a .8 file\n");
+		diag("%s:#%lld: opcode out of range: %#ux", pn, Boffset(f), o);
+		print("	probably not a .%c file\n", thechar);
 		errorexit();
 	}
 
 	if(o == ANAME || o == ASIGNAME) {
 		sig = 0;
-		if(o == ASIGNAME) {
-			sig = bloc[2] | (bloc[3]<<8) | (bloc[4]<<16) | (bloc[5]<<24);
-			bloc += 4;
-			c -= 4;
-		}
-		stop = memchr(&bloc[4], 0, bsize-&bloc[4]);
-		if(stop == 0){
-			bsize = readsome(f, buf.xbuf, bloc, bsize, c);
-			if(bsize == 0)
-				goto eof;
-			bloc = buf.xbuf;
-			stop = memchr(&bloc[4], 0, bsize-&bloc[4]);
-			if(stop == 0){
+		if(o == ASIGNAME)
+			sig = Bget4(f);
+		v = Bgetc(f);	/* type */
+		o = Bgetc(f);	/* sym */
+		r = 0;
+		if(v == D_STATIC)
+			r = version;
+		name = Brdline(f, '\0');
+		if(name == nil) {
+			if(Blinelen(f) > 0) {
 				fprint(2, "%s: name too long\n", pn);
 				errorexit();
 			}
+			goto eof;
 		}
-		v = bloc[2];	/* type */
-		o = bloc[3];	/* sym */
-		bloc += 4;
-		c -= 4;
-
-		r = 0;
-		if(v == D_STATIC)
-			r = version;
-		s = lookup((char*)bloc, r);
-		c -= &stop[1] - bloc;
-		bloc = stop + 1;
+		s = lookup(name, r);
 
 		if(debug['S'] && r == 0)
 			sig = 1729;
 		if(sig != 0){
 			if(s->sig != 0 && s->sig != sig)
-				diag("incompatible type signatures %lux(%s) and %lux(%s) for %s", s->sig, filen[s->file], sig, pn, s->name);
+				diag("incompatible type signatures"
+					"%lux(%s) and %lux(%s) for %s",
+					s->sig, filen[s->file], sig, pn, s->name);
 			s->sig = sig;
 			s->file = files-1;
 		}
@@ -877,12 +933,10 @@ loop:
 
 	p = mal(sizeof(*p));
 	p->as = o;
-	p->line = bloc[2] | (bloc[3] << 8) | (bloc[4] << 16) | (bloc[5] << 24);
+	p->line = Bget4(f);
 	p->back = 2;
-	r = zaddr(bloc+6, &p->from, h) + 6;
-	r += zaddr(bloc+r, &p->to, h);
-	bloc += r;
-	c -= r;
+	zaddr(f, &p->from, h);
+	zaddr(f, &p->to, h);
 
 	if(debug['W'])
 		print("%P\n", p);
@@ -890,10 +944,12 @@ loop:
 	switch(p->as) {
 	case AHISTORY:
 		if(p->to.offset == -1) {
-			addlib(pn);
+			addlib(src, pn);
 			histfrogp = 0;
 			goto loop;
 		}
+		if(src[0] == '\0')
+			copyhistfrog(src, sizeof src);
 		addhist(p->line, D_FILE);		/* 'z' */
 		if(p->to.offset)
 			addhist(p->to.offset, D_FILE1);	/* 'Z' */
@@ -906,9 +962,9 @@ loop:
 			curtext->to.autom = curauto;
 		curauto = 0;
 		curtext = P;
-		if(c)
-			goto newloop;
-		return;
+		if(Boffset(f) == eof)
+			return;
+		goto newloop;
 
 	case AGLOBL:
 		s = p->from.sym;
@@ -983,6 +1039,17 @@ loop:
 		goto loop;
 
 	case ATEXT:
+		s = p->from.sym;
+		if(s == S) {
+			diag("%s: no TEXT symbol: %P", pn, p);
+			errorexit();
+		}
+		if(ntext++ == 0 && s->type != 0 && s->type != SXREF) {
+			/* redefinition, so file has probably been seen before */
+			if(debug['v'])
+				diag("skipping: %s: redefinition: %s", pn, s->name);
+			return;
+		}
 		if(curtext != P) {
 			histtoauto();
 			curtext->to.autom = curauto;
@@ -990,11 +1057,6 @@ loop:
 		}
 		skip = 0;
 		curtext = p;
-		s = p->from.sym;
-		if(s == S) {
-			diag("%s: no TEXT symbol: %P", pn, p);
-			errorexit();
-		}
 		if(s->type != 0 && s->type != SXREF) {
 			if(p->from.scale & DUPOK) {
 				skip = 1;
diff --git a/src/cmd/8l/optab.c b/src/cmd/8l/optab.c
index d88e861c5..934f14e9b 100644
--- a/src/cmd/8l/optab.c
+++ b/src/cmd/8l/optab.c
@@ -173,6 +173,11 @@ uchar	yml_rl[] =
 	Yml,	Yrl,	Zm_r,	1,
 	0
 };
+uchar	yrb_mb[] =
+{
+	Yrb,	Ymb,	Zr_m,	1,
+	0
+};
 uchar	yrl_ml[] =
 {
 	Yrl,	Yml,	Zr_m,	1,
@@ -383,6 +388,9 @@ Optab optab[] =
 	{ ACMPSB,	ynone,	Pb, 0xa6 },
 	{ ACMPSL,	ynone,	Px, 0xa7 },
 	{ ACMPSW,	ynone,	Pe, 0xa7 },
+	{ ACMPXCHGB,	yrb_mb,	Pm, 0xb0 },
+	{ ACMPXCHGL,	yrl_ml,	Pm, 0xb1 },
+	{ ACMPXCHGW,	yrl_ml,	Pm, 0xb1 },
 	{ ADAA,		ynone,	Px, 0x27 },
 	{ ADAS,		ynone,	Px, 0x2f },
 	{ ADATA },
diff --git a/src/cmd/8l/pass.c b/src/cmd/8l/pass.c
index aa69c5849..0726d1ea8 100644
--- a/src/cmd/8l/pass.c
+++ b/src/cmd/8l/pass.c
@@ -165,7 +165,7 @@ loop:
 	if(p->as == ATEXT)
 		curtext = p;
 	if(p->as == AJMP)
-	if((q = p->pcond) != P) {
+	if((q = p->pcond) != P && q->as != ATEXT) {
 		p->mark = 1;
 		p = q;
 		if(p->mark == 0)
@@ -331,14 +331,15 @@ patch(void)
 	for(p = firstp; p != P; p = p->link) {
 		if(p->as == ATEXT)
 			curtext = p;
-		if(p->as == ACALL || p->as == ARET) {
+		if(p->as == ACALL || (p->as == AJMP && p->to.type != D_BRANCH)) {
 			s = p->to.sym;
 			if(s) {
 				if(debug['c'])
 					Bprint(&bso, "%s calls %s\n", TNAME, s->name);
 				switch(s->type) {
 				default:
-					diag("undefined: %s in %s", s->name, TNAME);
+					/* diag prints TNAME first */
+					diag("%s is undefined", s->name);
 					s->type = STEXT;
 					s->value = vexit;
 					break;	/* or fall through to set offset? */
@@ -440,9 +441,27 @@ brloop(Prog *p)
 void
 dostkoff(void)
 {
-	Prog *p, *q;
+	Prog *p, *q, *q1;
 	int32 autoffset, deltasp;
 	int a, f, curframe, curbecome, maxbecome;
+	Prog *pmorestack;
+	Sym *symmorestack;
+
+	pmorestack = P;
+	symmorestack = lookup("sys·morestack", 0);
+
+	if(symmorestack->type == STEXT)
+	for(p = firstp; p != P; p = p->link) {
+		if(p->as == ATEXT) {
+			if(p->from.sym == symmorestack) {
+				pmorestack = p;
+				p->from.scale |= NOSPLIT;
+				break;
+			}
+		}
+	}
+	if(pmorestack == P)
+		diag("sys·morestack not defined");
 
 	curframe = 0;
 	curbecome = 0;
@@ -521,11 +540,122 @@ dostkoff(void)
 			autoffset = p->to.offset;
 			if(autoffset < 0)
 				autoffset = 0;
+
+			q = P;
+			q1 = P;
+			if(pmorestack != P)
+			if(!(p->from.scale & NOSPLIT)) {
+				p = appendp(p);	// load g into CX
+				p->as = AMOVL;
+				p->from.type = D_INDIR+D_FS;
+				p->from.offset = 0;
+				p->to.type = D_CX;
+
+				if(debug['K']) {
+					// 8l -K means check not only for stack
+					// overflow but stack underflow.
+					// On underflow, INT 3 (breakpoint).
+					// Underflow itself is rare but this also
+					// catches out-of-sync stack guard info.
+					p = appendp(p);
+					p->as = ACMPL;
+					p->from.type = D_INDIR+D_CX;
+					p->from.offset = 4;
+					p->to.type = D_SP;
+
+					p = appendp(p);
+					p->as = AJHI;
+					p->to.type = D_BRANCH;
+					p->to.offset = 4;
+					q1 = p;
+
+					p = appendp(p);
+					p->as = AINT;
+					p->from.type = D_CONST;
+					p->from.offset = 3;
+				}
+
+				if(autoffset < 4096) {  // do we need to call morestack
+					if(autoffset <= 75) {
+						// small stack
+						p = appendp(p);
+						p->as = ACMPL;
+						p->from.type = D_SP;
+						p->to.type = D_INDIR+D_CX;
+						if(q1) {
+							q1->pcond = p;
+							q1 = P;
+						}
+					} else {
+						// large stack
+						p = appendp(p);
+						p->as = ALEAL;
+						p->from.type = D_INDIR+D_SP;
+						p->from.offset = -(autoffset-75);
+						p->to.type = D_AX;
+						if(q1) {
+							q1->pcond = p;
+							q1 = P;
+						}
+
+						p = appendp(p);
+						p->as = ACMPL;
+						p->from.type = D_AX;
+						p->to.type = D_INDIR+D_CX;
+					}
+
+					// common
+					p = appendp(p);
+					p->as = AJHI;
+					p->to.type = D_BRANCH;
+					p->to.offset = 4;
+					q = p;
+				}
+
+				p = appendp(p);	// load m into DX
+				p->as = AMOVL;
+				p->from.type = D_INDIR+D_FS;
+				p->from.offset = 4;
+				p->to.type = D_DX;
+				if(q1) {
+					q1->pcond = p;
+					q1 = P;
+				}
+
+				p = appendp(p);	// save autoffset in 4(DX)
+				p->as = AMOVL;
+				p->to.type = D_INDIR+D_DX;
+				p->to.offset = 4;
+				/* 160 comes from 3 calls (3*8) 4 safes (4*8) and 104 guard */
+				p->from.type = D_CONST;
+				if(autoffset+160 > 4096)
+					p->from.offset = (autoffset+160) & ~7LL;
+
+				p = appendp(p);	// save textarg in 8(DX)
+				p->as = AMOVL;
+				p->to.type = D_INDIR+D_DX;
+				p->to.offset = 8;
+				p->from.type = D_CONST;
+				p->from.offset = curtext->to.offset2;
+
+				p = appendp(p);
+				p->as = ACALL;
+				p->to.type = D_BRANCH;
+				p->pcond = pmorestack;
+				p->to.sym = symmorestack;
+
+			}
+
+			if(q != P)
+				q->pcond = p->link;
+
 			if(autoffset) {
 				p = appendp(p);
 				p->as = AADJSP;
 				p->from.type = D_CONST;
 				p->from.offset = autoffset;
+				if(q != P)
+					q->pcond = p;
 			}
 			deltasp = autoffset;
 		}
diff --git a/src/cmd/8l/span.c b/src/cmd/8l/span.c
index b524c735f..81703d304 100644
--- a/src/cmd/8l/span.c
+++ b/src/cmd/8l/span.c
@@ -148,7 +148,10 @@ xdefine(char *p, int t, int32 v)
 void
 putsymb(char *s, int t, int32 v, int ver)
 {
-	int i, f;
+	int i, j, f;
+	char *go; 
+	
+	go = nil;	// TODO
 
 	if(t == 'f')
 		s++;
@@ -172,7 +175,14 @@ putsymb(char *s, int t, int32 v, int ver)
 			cput(s[i]);
 		cput(0);
 	}
-	symsize += 4 + 1 + i + 1;
+	j = 0;
+	if(go) {
+		for(j=0; go[j]; j++)
+			cput(go[j]);
+	}
+	cput(0);
+
+	symsize += 4 + 1 + i + 1 + j + 1;
 
 	if(debug['n']) {
 		if(t == 'z' || t == 'Z') {
@@ -334,6 +344,24 @@ asmlc(void)
 }
 
 int
+prefixof(Adr *a)
+{
+	switch(a->type) {
+	case D_INDIR+D_CS:
+		return 0x2e;
+	case D_INDIR+D_DS:
+		return 0x3e;
+	case D_INDIR+D_ES:
+		return 0x26;
+	case D_INDIR+D_FS:
+		return 0x64;
+	case D_INDIR+D_GS:
+		return 0x65;
+	}
+	return 0;
+}
+
+int
 oclass(Adr *a)
 {
 	int32 v;
@@ -447,6 +475,7 @@ oclass(Adr *a)
 		return Ym;
 
 	case D_CONST:
+	case D_CONST2:
 	case D_ADDR:
 		if(a->sym == S) {
 			v = a->offset;
@@ -636,7 +665,7 @@ asmand(Adr *a, int r)
 	}
 	if(t >= D_INDIR) {
 		t -= D_INDIR;
-		if(t == D_NONE) {
+		if(t == D_NONE || (D_CS <= t && t <= D_GS)) {
 			*andptr++ = (0 << 6) | (5 << 0) | (r << 3);
 			put4(v);
 			return;
@@ -851,7 +880,14 @@ doasm(Prog *p)
 	Prog *q, pp;
 	uchar *t;
 	int z, op, ft, tt;
-	int32 v;
+	int32 v, pre;
+
+	pre = prefixof(&p->from);
+	if(pre)
+		*andptr++ = pre;
+	pre = prefixof(&p->to);
+	if(pre)
+		*andptr++ = pre;
 
 	o = &optab[p->as];
 	ft = oclass(&p->from) * Ymax;
diff --git a/src/cmd/cc/cc.h b/src/cmd/cc/cc.h
index 3b13126df..42341a70c 100644
--- a/src/cmd/cc/cc.h
+++ b/src/cmd/cc/cc.h
@@ -262,6 +262,7 @@ enum
 	ODWHILE,
 	OENUM,
 	OEQ,
+	OEXREG,
 	OFOR,
 	OFUNC,
 	OGE,
diff --git a/src/cmd/cc/com.c b/src/cmd/cc/com.c
index 5ee38e882..af2509c5f 100644
--- a/src/cmd/cc/com.c
+++ b/src/cmd/cc/com.c
@@ -638,6 +638,11 @@ tcomo(Node *n, int f)
 		n->addable = 1;
 		if(n->class == CEXREG) {
 			n->op = OREGISTER;
+			// on 386, "extern register" generates
+			// memory references relative to the
+			// fs segment.
+			if(thechar == '8')	// [sic]
+				n->op = OEXREG;
 			n->reg = n->sym->offset;
 			n->xoffset = 0;
 			break;
diff --git a/src/cmd/cc/sub.c b/src/cmd/cc/sub.c
index cad5eda72..335d30bfb 100644
--- a/src/cmd/cc/sub.c
+++ b/src/cmd/cc/sub.c
@@ -1460,6 +1460,7 @@ Init	onamesinit[] =
 	ODWHILE,	0,	"DWHILE",
 	OENUM,		0,	"ENUM",
 	OEQ,		0,	"EQ",
+	OEXREG,	0,	"EXREG",
 	OFOR,		0,	"FOR",
 	OFUNC,		0,	"FUNC",
 	OGE,		0,	"GE",