8c, 8l dynamic loading support.

better mach binaries.
cgo working on darwin+linux amd64+386.
eliminated context switches - pi is 30x faster.
add libcgo to build.

on snow leopard:
  - non-cgo binaries work; all tests pass.
  - cgo binaries work on amd64 but not 386.

R=r
DELTA=2031  (1316 added, 626 deleted, 89 changed)
OCL=35264
CL=35304

17 files changed, 474 insertions, 204 deletions

diff --git a/src/cmd/6l/asm.c b/src/cmd/6l/asm.c
index 5562ee4e3..a0f8524e7 100644
--- a/src/cmd/6l/asm.c
+++ b/src/cmd/6l/asm.c
@@ -425,16 +425,11 @@ asmb(void)
 	int32 v, magic;
 	int a, dynsym;
 	uchar *op1;
-	vlong vl, va, startva, fo, w, symo;
+	vlong vl, va, startva, fo, w, symo, machlink;
 	vlong symdatva = 0x99LL<<32;
 	ElfEhdr *eh;
 	ElfPhdr *ph, *pph;
 	ElfShdr *sh;
-	MachoHdr *mh;
-	MachoSect *msect;
-	MachoSeg *ms;
-	MachoDebug *md;
-	MachoLoad *ml;
 
 	if(debug['v'])
 		Bprint(&bso, "%5.2f asmb\n", cputime());
@@ -523,6 +518,10 @@ asmb(void)
 			datblk(v, datsize-v);
 	}
 
+	machlink = 0;
+	if(HEADTYPE == 6)
+		machlink = domacholink();
+
 	symsize = 0;
 	spsize = 0;
 	lcsize = 0;
@@ -539,7 +538,7 @@ asmb(void)
 			symo = HEADR+textsize+datsize;
 			break;
 		case 6:
-			symo = rnd(HEADR+textsize, INITRND)+rnd(datsize, INITRND);
+			symo = rnd(HEADR+textsize, INITRND)+rnd(datsize, INITRND)+machlink;
 			break;
 		case 7:
 			symo = rnd(HEADR+textsize, INITRND)+datsize;
@@ -607,92 +606,8 @@ asmb(void)
 		lputb(lcsize);			/* line offsets */
 		break;
 	case 6:
-		/* apple MACH */
-		va = HEADR;
-		mh = getMachoHdr();
-		mh->cpu = MACHO_CPU_AMD64;
-		mh->subcpu = MACHO_SUBCPU_X86;
-
-		/* segment for zero page */
-		ms = newMachoSeg("__PAGEZERO", 0);
-		ms->vsize = va;
-
-		/* text */
-		v = rnd(HEADR+textsize, INITRND);
-		ms = newMachoSeg("__TEXT", 1);
-		ms->vaddr = va;
-		ms->vsize = v;
-		ms->filesize = v;
-		ms->prot1 = 7;
-		ms->prot2 = 5;
-
-		msect = newMachoSect(ms, "__text");
-		msect->addr = va+HEADR;
-		msect->size = v - HEADR;
-		msect->off = HEADR;
-		msect->flag = 0x400;	/* flag - some instructions */
-
-		/* data */
-		w = datsize+bsssize;
-		ms = newMachoSeg("__DATA", 2);
-		ms->vaddr = va+v;
-		ms->vsize = w;
-		ms->fileoffset = v;
-		ms->filesize = datsize;
-		ms->prot1 = 7;
-		ms->prot2 = 3;
-
-		msect = newMachoSect(ms, "__data");
-		msect->addr = va+v;
-		msect->size = datsize;
-		msect->off = v;
-
-		msect = newMachoSect(ms, "__bss");
-		msect->addr = va+v+datsize;
-		msect->size = bsssize;
-		msect->flag = 1;	/* flag - zero fill */
-
-		ml = newMachoLoad(5, 42+2);	/* unix thread */
-		ml->data[0] = 4;	/* thread type */
-		ml->data[1] = 42;	/* word count */
-		ml->data[2+32] = entryvalue();	/* start pc */
-		ml->data[2+32+1] = entryvalue()>>32;
-
-		if(!debug['d']) {
-			ml = newMachoLoad(2, 4);	/* LC_SYMTAB */
-			USED(ml);
-
-			ml = newMachoLoad(11, 18);	/* LC_DYSYMTAB */
-			USED(ml);
-
-			ml = newMachoLoad(14, 6);	/* LC_LOAD_DYLINKER */
-			ml->data[0] = 12;	/* offset to string */
-			strcpy((char*)&ml->data[1], "/usr/lib/dyld");
-		}
-
-		if(!debug['s']) {
-			ms = newMachoSeg("__SYMDAT", 1);
-			ms->vaddr = symdatva;
-			ms->vsize = 8+symsize+lcsize;
-			ms->fileoffset = symo;
-			ms->filesize = 8+symsize+lcsize;
-			ms->prot1 = 7;
-			ms->prot2 = 5;
-
-			md = newMachoDebug();
-			md->fileoffset = symo+8;
-			md->filesize = symsize;
-
-			md = newMachoDebug();
-			md->fileoffset = symo+8+symsize;
-			md->filesize = lcsize;
-		}
-
-		a = machowrite();
-		if(a > MACHORESERVE)
-			diag("MACHORESERVE too small: %d > %d", a, MACHORESERVE);
+		asmbmacho(symdatva, symo);
 		break;
-
 	case 7:
 		/* elf amd-64 */
 
@@ -965,6 +880,8 @@ datblk(int32 s, int32 n)
 		curp = p;
 		if(!p->from.sym->reachable)
 			diag("unreachable symbol in datblk - %s", p->from.sym->name);
+		if(p->from.sym->type == SMACHO)
+			continue;
 		l = p->from.sym->value + p->from.offset - s;
 		c = p->from.scale;
 		i = 0;
diff --git a/src/cmd/6l/l.h b/src/cmd/6l/l.h
index a1c2ec527..28c37a82c 100644
--- a/src/cmd/6l/l.h
+++ b/src/cmd/6l/l.h
@@ -160,6 +160,8 @@ enum
 	SIMPORT,
 	SEXPORT,
 
+	SMACHO,
+
 	NHASH		= 10007,
 	NHUNK		= 100000,
 	MINSIZ		= 8,
@@ -362,6 +364,7 @@ EXTERN	Prog	undefp;
 EXTERN	vlong	textstksiz;
 EXTERN	vlong	textarg;
 extern	char	thechar;
+EXTERN	int	dynptrsize;
 
 #define	UP	(&undefp)
 
@@ -403,6 +406,7 @@ void	dobss(void);
 void	dodata(void);
 void	doelf(void);
 void	doinit(void);
+void	domacho(void);
 void	doprof1(void);
 void	doprof2(void);
 void	dostkoff(void);
diff --git a/src/cmd/6l/obj.c b/src/cmd/6l/obj.c
index ba2dec3b2..4b40cce61 100644
--- a/src/cmd/6l/obj.c
+++ b/src/cmd/6l/obj.c
@@ -191,12 +191,12 @@ main(int argc, char *argv[])
 	case 6:	/* apple MACH */
 		machoinit();
 		HEADR = MACHORESERVE;
+		if(INITRND == -1)
+			INITRND = 4096;
 		if(INITTEXT == -1)
 			INITTEXT = 4096+HEADR;
 		if(INITDAT == -1)
 			INITDAT = 0;
-		if(INITRND == -1)
-			INITRND = 4096;
 		break;
 	case 7:	/* elf64 executable */
 		elfinit();
@@ -393,6 +393,8 @@ main(int argc, char *argv[])
 	patch();
 	follow();
 	doelf();
+	if(HEADTYPE == 6)
+		domacho();
 	dodata();
 	dobss();
 	dostkoff();
diff --git a/src/cmd/6l/pass.c b/src/cmd/6l/pass.c
index c2f560500..2da88bac1 100644
--- a/src/cmd/6l/pass.c
+++ b/src/cmd/6l/pass.c
@@ -142,6 +142,11 @@ dobss(void)
 	Sym *s;
 	int32 t;
 
+	if(dynptrsize > 0) {
+		/* dynamic pointer section between data and bss */
+		datsize = rnd(datsize, 8);
+	}
+
 	/* now the bss */
 	bsssize = 0;
 	for(i=0; i<NHASH; i++)
@@ -154,12 +159,13 @@ dobss(void)
 		s->size = t;
 		if(t >= 8)
 			bsssize = rnd(bsssize, 8);
-		s->value = bsssize + datsize;
+		s->value = bsssize + dynptrsize + datsize;
 		bsssize += t;
 	}
+
 	xdefine("data", SBSS, 0);
 	xdefine("edata", SBSS, datsize);
-	xdefine("end", SBSS, bsssize + datsize);
+	xdefine("end", SBSS, dynptrsize + bsssize + datsize);
 }
 
 Prog*
diff --git a/src/cmd/6l/span.c b/src/cmd/6l/span.c
index f1eafff00..18bf8cc0c 100644
--- a/src/cmd/6l/span.c
+++ b/src/cmd/6l/span.c
@@ -237,6 +237,12 @@ asmsym(void)
 				putsymb(s->name, 'D', s->value+INITDAT, s->version, s->gotype);
 				continue;
 
+			case SMACHO:
+				if(!s->reachable)
+					continue;
+				putsymb(s->name, 'D', s->value+INITDAT+datsize+bsssize, s->version, s->gotype);
+				continue;
+
 			case SBSS:
 				if(!s->reachable)
 					continue;
@@ -715,6 +721,11 @@ vaddr(Adr *a)
 					v += INITTEXT;	/* TO DO */
 				v += s->value;
 				break;
+			case SMACHO:
+				if(!s->reachable)
+					sysfatal("unreachable symbol in vaddr - %s", s->name);
+				v += INITDAT + datsize + s->value;
+				break;
 			default:
 				if(!s->reachable)
 					diag("unreachable symbol in vaddr - %s", s->name);
diff --git a/src/cmd/8l/asm.c b/src/cmd/8l/asm.c
index 449467a5c..c70af7072 100644
--- a/src/cmd/8l/asm.c
+++ b/src/cmd/8l/asm.c
@@ -257,7 +257,7 @@ needlib(char *name)
 	Sym *s;
 
 	/* reuse hash code in symbol table */
-	p = smprint(".elfload.%s", name);
+	p = smprint(".dynlib.%s", name);
 	s = lookup(p, 0);
 	if(s->type == 0) {
 		s->type = 100;	// avoid SDATA, etc.
@@ -414,8 +414,8 @@ asmb(void)
 {
 	Prog *p;
 	int32 v, magic;
-	int a, dynsym;
-	uint32 va, fo, w, symo, startva;
+	int a, i, dynsym;
+	uint32 va, fo, w, symo, startva, machlink;
 	uchar *op1;
 	ulong expectpc;
 	ElfEhdr *eh;
@@ -547,6 +547,10 @@ asmb(void)
 			datblk(v, datsize-v);
 	}
 
+	machlink = 0;
+	if(HEADTYPE == 6)
+		machlink = domacholink();
+
 	symsize = 0;
 	spsize = 0;
 	lcsize = 0;
@@ -572,7 +576,7 @@ asmb(void)
 			symo = HEADR+textsize+datsize;
 			break;
 		case 6:
-			symo = rnd(HEADR+textsize, INITRND)+rnd(datsize, INITRND);
+			symo = rnd(HEADR+textsize, INITRND)+rnd(datsize, INITRND)+machlink;
 			break;
 		case 7:
 		case 8:
@@ -740,89 +744,7 @@ asmb(void)
 		break;
 
 	case 6:
-		/* apple MACH */
-		va = HEADR;
-		mh = getMachoHdr();
-		mh->cpu = MACHO_CPU_386;
-		mh->subcpu = MACHO_SUBCPU_X86;
-
-		/* segment for zero page */
-		ms = newMachoSeg("__PAGEZERO", 0);
-		ms->vsize = va;
-
-		/* text */
-		v = rnd(HEADR+textsize, INITRND);
-		ms = newMachoSeg("__TEXT", 1);
-		ms->vaddr = va;
-		ms->vsize = v;
-		ms->filesize = v;
-		ms->prot1 = 7;
-		ms->prot2 = 5;
-
-		msect = newMachoSect(ms, "__text");
-		msect->addr = va+HEADR;
-		msect->size = v - HEADR;
-		msect->off = HEADR;
-		msect->flag = 0x400;	/* flag - some instructions */
-
-		/* data */
-		w = datsize+bsssize;
-		ms = newMachoSeg("__DATA", 2);
-		ms->vaddr = va+v;
-		ms->vsize = w;
-		ms->fileoffset = v;
-		ms->filesize = datsize;
-		ms->prot1 = 7;
-		ms->prot2 = 3;
-
-		msect = newMachoSect(ms, "__data");
-		msect->addr = va+v;
-		msect->size = datsize;
-		msect->off = v;
-
-		msect = newMachoSect(ms, "__bss");
-		msect->addr = va+v+datsize;
-		msect->size = bsssize;
-		msect->flag = 1;	/* flag - zero fill */
-
-		ml = newMachoLoad(5, 16+2);	/* unix thread */
-		ml->data[0] = 1;	/* thread type */
-		ml->data[1] = 16;	/* word count */
-		ml->data[2+10] = entryvalue();	/* start pc */
-
-		if(!debug['d']) {
-			ml = newMachoLoad(2, 4);	/* LC_SYMTAB */
-			USED(ml);
-
-			ml = newMachoLoad(11, 18);	/* LC_DYSYMTAB */
-			USED(ml);
-
-			ml = newMachoLoad(14, 6);	/* LC_LOAD_DYLINKER */
-			ml->data[0] = 12;	/* offset to string */
-			strcpy((char*)&ml->data[1], "/usr/lib/dyld");
-		}
-
-		if(!debug['s']) {
-			ms = newMachoSeg("__SYMDAT", 1);
-			ms->vaddr = symdatva;
-			ms->vsize = 8+symsize+lcsize;
-			ms->fileoffset = symo;
-			ms->filesize = 8+symsize+lcsize;
-			ms->prot1 = 7;
-			ms->prot2 = 5;
-
-			md = newMachoDebug();
-			md->fileoffset = symo+8;
-			md->filesize = symsize;
-
-			md = newMachoDebug();
-			md->fileoffset = symo+8+symsize;
-			md->filesize = lcsize;
-		}
-
-		a = machowrite();
-		if(a > MACHORESERVE)
-			diag("MACHORESERVE too small: %d > %d", a, MACHORESERVE);
+		asmbmacho(symdatva, symo);
 		break;
 
 	case 7:
@@ -963,6 +885,17 @@ asmb(void)
 			ph->type = PT_DYNAMIC;
 			ph->flags = PF_R + PF_W;
 			phsh(ph, sh);
+
+			/*
+			 * Thread-local storage segment (really just size).
+			 */
+			if(tlsoffset != 0) {
+				ph = newElfPhdr();
+				ph->type = PT_TLS;
+				ph->flags = PF_R;
+				ph->memsz = -tlsoffset;
+				ph->align = 4;
+			}
 		}
 
 		ph = newElfPhdr();
@@ -1105,6 +1038,8 @@ datblk(int32 s, int32 n)
 		curp = p;
 		if(!p->from.sym->reachable)
 			diag("unreachable symbol in datblk - %s", p->from.sym->name);
+		if(p->from.sym->type == SMACHO)
+			continue;
 		l = p->from.sym->value + p->from.offset - s;
 		c = p->from.scale;
 		i = 0;
diff --git a/src/cmd/8l/l.h b/src/cmd/8l/l.h
index cc5901fcb..1959b2c74 100644
--- a/src/cmd/8l/l.h
+++ b/src/cmd/8l/l.h
@@ -151,6 +151,8 @@ enum
 	SIMPORT,
 	SEXPORT,
 
+	SMACHO,	/* pointer to mach-o imported symbol */
+
 	NHASH		= 10007,
 	NHUNK		= 100000,
 	MINSIZ		= 4,
@@ -272,6 +274,7 @@ EXTERN	Prog*	curtext;
 EXTERN	Prog*	datap;
 EXTERN	Prog*	edatap;
 EXTERN	int32	datsize;
+EXTERN	int32	dynptrsize;
 EXTERN	char	debug[128];
 EXTERN	char	literal[32];
 EXTERN	Prog*	etextp;
@@ -311,6 +314,7 @@ EXTERN	int	version;
 EXTERN	Prog	zprg;
 EXTERN	int	dtype;
 EXTERN	char	thechar;
+EXTERN	int	tlsoffset;
 
 EXTERN	Adr*	reloca;
 EXTERN	int	doexp, dlm;
diff --git a/src/cmd/8l/obj.c b/src/cmd/8l/obj.c
index 4b6532568..aa197be53 100644
--- a/src/cmd/8l/obj.c
+++ b/src/cmd/8l/obj.c
@@ -213,6 +213,11 @@ main(int argc, char *argv[])
 			Bprint(&bso, "HEADR = 0x%ld\n", HEADR);
 		break;
 	case 6:	/* apple MACH */
+		/*
+		 * OS X system constant - offset from %gs to our TLS.
+		 * Explained in ../../libcgo/darwin_386.c.
+		 */
+		tlsoffset = 0x468;
 		machoinit();
 		HEADR = MACHORESERVE;
 		if(INITTEXT == -1)
@@ -223,6 +228,13 @@ main(int argc, char *argv[])
 			INITRND = 4096;
 		break;
 	case 7:	/* elf32 executable */
+		/*
+		 * Linux ELF uses TLS offsets negative from %gs.
+		 * Translate 0(GS) and 4(GS) into -8(GS) and -4(GS).
+		 * Also known to ../../pkg/runtime/linux/386/sys.s
+		 * and ../../libcgo/linux_386.c.
+		 */
+		tlsoffset = -8;
 		elfinit();
 		HEADR = ELFRESERVE;
 		if(INITTEXT == -1)
@@ -373,6 +385,8 @@ main(int argc, char *argv[])
 	patch();
 	follow();
 	doelf();
+	if(HEADTYPE == 6)
+		domacho();
 	dodata();
 	dostkoff();
 	if(debug['p'])
@@ -592,11 +606,14 @@ zaddr(Biobuf *f, Adr *a, Sym *h[])
 		a->type = Bgetc(f);
 	if(t & T_GOTYPE)
 		a->gotype = h[Bgetc(f)];
+
+	t = a->type;
+	if(t == D_INDIR+D_GS)
+		a->offset += tlsoffset;
+
 	s = a->sym;
 	if(s == S)
 		return;
-
-	t = a->type;
 	if(t != D_AUTO && t != D_PARAM) {
 		if(a->gotype)
 			s->gotype = a->gotype;
diff --git a/src/cmd/8l/pass.c b/src/cmd/8l/pass.c
index 7ce419e8f..c624f750a 100644
--- a/src/cmd/8l/pass.c
+++ b/src/cmd/8l/pass.c
@@ -125,6 +125,11 @@ dodata(void)
 		datsize += u;
 	}
 
+	if(dynptrsize > 0) {
+		/* dynamic pointer section between data and bss */
+		datsize = rnd(datsize, 4);
+	}
+
 	/* now the bss */
 	bsssize = 0;
 	for(i=0; i<NHASH; i++)
@@ -135,12 +140,13 @@ dodata(void)
 			continue;
 		t = s->value;
 		s->size = t;
-		s->value = bsssize + datsize;
+		s->value = bsssize + dynptrsize + datsize;
 		bsssize += t;
 	}
+
 	xdefine("data", SBSS, 0);
 	xdefine("edata", SBSS, datsize);
-	xdefine("end", SBSS, bsssize + datsize);
+	xdefine("end", SBSS, dynptrsize + bsssize + datsize);
 }
 
 Prog*
@@ -570,7 +576,7 @@ dostkoff(void)
 				p = appendp(p);	// load g into CX
 				p->as = AMOVL;
 				p->from.type = D_INDIR+D_GS;
-				p->from.offset = 0;
+				p->from.offset = tlsoffset + 0;
 				p->to.type = D_CX;
 
 				if(debug['K']) {
diff --git a/src/cmd/8l/span.c b/src/cmd/8l/span.c
index 71607fcf2..6f62f9b4d 100644
--- a/src/cmd/8l/span.c
+++ b/src/cmd/8l/span.c
@@ -219,6 +219,12 @@ asmsym(void)
 				putsymb(s->name, 'D', s->value+INITDAT, s->version, s->gotype);
 				continue;
 
+			case SMACHO:
+				if(!s->reachable)
+					continue;
+				putsymb(s->name, 'D', s->value+INITDAT+datsize+bsssize, s->version, s->gotype);
+				continue;
+
 			case SBSS:
 				if(!s->reachable)
 					continue;
@@ -611,6 +617,11 @@ vaddr(Adr *a)
 					sysfatal("unreachable symbol in vaddr - %s", s->name);
 				v += s->value;
 				break;
+			case SMACHO:
+				if(!s->reachable)
+					sysfatal("unreachable symbol in vaddr - %s", s->name);
+				v += INITDAT + datsize + s->value;
+				break;
 			default:
 				if(!s->reachable)
 					sysfatal("unreachable symbol in vaddr - %s", s->name);
diff --git a/src/cmd/cgo/ast.go b/src/cmd/cgo/ast.go
index 9b122676c..57500680b 100644
--- a/src/cmd/cgo/ast.go
+++ b/src/cmd/cgo/ast.go
@@ -36,6 +36,7 @@ type Prog struct {
 	Vardef map[string]*Type;
 	Funcdef map[string]*FuncType;
 	PtrSize int64;
+	GccOptions []string;
 }
 
 // A Type collects information about a type in both the C and Go worlds.
diff --git a/src/cmd/cgo/gcc.go b/src/cmd/cgo/gcc.go
index e3f526845..f573b98cb 100644
--- a/src/cmd/cgo/gcc.go
+++ b/src/cmd/cgo/gcc.go
@@ -172,6 +172,17 @@ func (p *Prog) loadDebugInfo() {
 	p.Typedef = conv.typedef;
 }
 
+func concat(a, b []string) []string {
+	c := make([]string, len(a)+len(b));
+	for i, s := range a {
+		c[i] = s;
+	}
+	for i, s := range b {
+		c[i+len(a)] = s;
+	}
+	return c;
+}
+
 // gccDebug runs gcc -gdwarf-2 over the C program stdin and
 // returns the corresponding DWARF data and any messages
 // printed to standard error.
@@ -182,7 +193,7 @@ func (p *Prog) gccDebug(stdin []byte) (*dwarf.Data, string) {
 	}
 
 	tmp := "_cgo_.o";
-	_, stderr, ok := run(stdin, []string{
+	base := []string{
 		"gcc",
 		machine,
 		"-Wall",	// many warnings
@@ -192,7 +203,8 @@ func (p *Prog) gccDebug(stdin []byte) (*dwarf.Data, string) {
 		"-c",	// do not link
 		"-xc", 	// input language is C
 		"-",	// read input from standard input
-	});
+	};
+	_, stderr, ok := run(stdin, concat(base, p.GccOptions));
 	if !ok {
 		return nil, string(stderr);
 	}
diff --git a/src/cmd/cgo/main.go b/src/cmd/cgo/main.go
index b629f0a22..eb04fa77d 100644
--- a/src/cmd/cgo/main.go
+++ b/src/cmd/cgo/main.go
@@ -11,15 +11,13 @@
 package main
 
 import (
-	"flag";
 	"fmt";
 	"go/ast";
 	"os";
 )
 
 func usage() {
-	fmt.Fprint(os.Stderr, "usage: cgo file.cgo\n");
-	flag.PrintDefaults();
+	fmt.Fprint(os.Stderr, "usage: cgo [compiler options] file.go\n");
 }
 
 var ptrSizeMap = map[string]int64 {
@@ -28,9 +26,24 @@ var ptrSizeMap = map[string]int64 {
 	"arm": 4
 }
 
+var expandName = map[string]string {
+	"schar": "signed char",
+	"uchar": "unsigned char",
+	"ushort": "unsigned short",
+	"uint": "unsigned int",
+	"ulong": "unsigned long",
+	"longlong": "long long",
+	"ulonglong": "unsigned long long",
+}
+
 func main() {
-	flag.Usage = usage;
-	flag.Parse();
+	args := os.Args;
+	if len(args) < 2 {
+		usage();
+		os.Exit(2);
+	}
+	gccOptions := args[1:len(args)-1];
+	input := args[len(args)-1];
 
 	arch := os.Getenv("GOARCH");
 	if arch == "" {
@@ -41,14 +54,17 @@ func main() {
 		fatal("unknown architecture %s", arch);
 	}
 
-	args := flag.Args();
-	if len(args) != 1 {
-		usage();
-		os.Exit(2);
+	p := openProg(input);
+	for _, cref := range p.Crefs {
+		// Convert C.ulong to C.unsigned long, etc.
+		if expand, ok := expandName[cref.Name]; ok {
+			cref.Name = expand;
+		}
 	}
-	p := openProg(args[0]);
+
 	p.PtrSize = ptrSize;
 	p.Preamble = p.Preamble + "\n" + builtinProlog;
+	p.GccOptions = gccOptions;
 	p.loadDebugInfo();
 	p.Vardef = make(map[string]*Type);
 	p.Funcdef = make(map[string]*FuncType);
@@ -83,5 +99,5 @@ func main() {
 	}
 
 	p.PackagePath = p.Package;
-	p.writeOutput(args[0]);
+	p.writeOutput(input);
 }
diff --git a/src/cmd/cgo/out.go b/src/cmd/cgo/out.go
index 91473abeb..d2eedc331 100644
--- a/src/cmd/cgo/out.go
+++ b/src/cmd/cgo/out.go
@@ -198,7 +198,7 @@ const cProlog = `
 #include "cgocall.h"
 
 #pragma dynld initcgo initcgo "%s/libcgo.so"
-#pragma dynld cgo cgo "%s/libcgo.so"
+#pragma dynld libcgo_thread_start libcgo_thread_start "%s/libcgo.so"
 #pragma dynld _cgo_malloc _cgo_malloc "%s/libcgo.so"
 #pragma dynld _cgo_free free "%s/libcgo.so"
 
diff --git a/src/cmd/ld/elf.h b/src/cmd/ld/elf.h
index cb4857248..e0c2bd1b6 100644
--- a/src/cmd/ld/elf.h
+++ b/src/cmd/ld/elf.h
@@ -243,7 +243,7 @@ typedef struct {
 #define PT_NOTE		4	/* Auxiliary information. */
 #define PT_SHLIB	5	/* Reserved (not used). */
 #define PT_PHDR		6	/* Location of program header itself. */
-#define	PT_TLS		7	/* Thread local storage segment */
+#define PT_TLS		7	/* Thread local storage segment */
 #define PT_LOOS		0x60000000	/* First OS-specific. */
 #define PT_HIOS		0x6fffffff	/* Last OS-specific. */
 #define PT_LOPROC	0x70000000	/* First processor-specific type. */
diff --git a/src/cmd/ld/macho.c b/src/cmd/ld/macho.c
index 159aceb9e..e4fe963ac 100644
--- a/src/cmd/ld/macho.c
+++ b/src/cmd/ld/macho.c
@@ -92,6 +92,22 @@ newMachoDebug(void)
 	return &xdebug[ndebug++];
 }
 
+
+// Generic linking code.
+
+static uchar *linkdata;
+static uint32 nlinkdata;
+static uint32 mlinkdata;
+
+static uchar *strtab;
+static uint32 nstrtab;
+static uint32 mstrtab;
+
+static char **dylib;
+static int ndylib;
+
+static vlong linkoff;
+
 int
 machowrite(void)
 {
@@ -205,3 +221,308 @@ machowrite(void)
 
 	return Boffset(&bso) - o1;
 }
+
+static void*
+grow(uchar **dat, uint32 *ndat, uint32 *mdat, uint32 n)
+{
+	uchar *p;
+	uint32 old;
+
+	if(*ndat+n > *mdat) {
+		old = *mdat;
+		*mdat = (*ndat+n)*2 + 128;
+		*dat = realloc(*dat, *mdat);
+		if(*dat == 0) {
+			diag("out of memory");
+			errorexit();
+		}
+		memset(*dat+old, 0, *mdat-old);
+	}
+	p = *dat + *ndat;
+	*ndat += n;
+	return p;
+}
+
+static int
+needlib(char *name)
+{
+	char *p;
+	Sym *s;
+
+	/* reuse hash code in symbol table */
+	p = smprint(".machoload.%s", name);
+	s = lookup(p, 0);
+	if(s->type == 0) {
+		s->type = 100;	// avoid SDATA, etc.
+		return 1;
+	}
+	return 0;
+}
+
+void
+domacho(void)
+{
+	int h, nsym, ptrsize;
+	char *p;
+	uchar *dat;
+	uint32 x;
+	Sym *s;
+
+	ptrsize = 4;
+	if(macho64)
+		ptrsize = 8;
+
+	// empirically, string table must begin with " \x00".
+	if(!debug['d'])
+		*(char*)grow(&strtab, &nstrtab, &mstrtab, 2) = ' ';
+
+	nsym = 0;
+	for(h=0; h<NHASH; h++) {
+		for(s=hash[h]; s!=S; s=s->link) {
+			if(!s->reachable || (s->type != SDATA && s->type != SBSS) || s->dynldname == nil)
+				continue;
+			if(debug['d']) {
+				diag("cannot use dynamic loading and -d");
+				errorexit();
+			}
+			s->type = SMACHO;
+			s->value = nsym*ptrsize;
+
+			/* symbol table entry - darwin still puts _ prefixes on all C symbols */
+			x = nstrtab;
+			p = grow(&strtab, &nstrtab, &mstrtab, 1+strlen(s->dynldname)+1);
+			*p++ = '_';
+			strcpy(p, s->dynldname);
+
+			dat = grow(&linkdata, &nlinkdata, &mlinkdata, 8+ptrsize);
+			dat[0] = x;
+			dat[1] = x>>8;
+			dat[2] = x>>16;
+			dat[3] = x>>24;
+			dat[4] = 0x01;	// type: N_EXT - external symbol
+
+			if(needlib(s->dynldlib)) {
+				if(ndylib%32 == 0) {
+					dylib = realloc(dylib, (ndylib+32)*sizeof dylib[0]);
+					if(dylib == nil) {
+						diag("out of memory");
+						errorexit();
+					}
+				}
+				dylib[ndylib++] = s->dynldlib;
+			}
+			nsym++;
+		}
+	}
+
+	/*
+	 * list of symbol table indexes.
+	 * we don't take advantage of the opportunity
+	 * to order the symbol table differently from
+	 * this list, so it is boring: 0 1 2 3 4 ...
+	 */
+	for(x=0; x<nsym; x++) {
+		dat = grow(&linkdata, &nlinkdata, &mlinkdata, 4);
+		dat[0] = x;
+		dat[1] = x>>8;
+		dat[2] = x>>16;
+		dat[3] = x>>24;
+	}
+
+	dynptrsize = nsym*ptrsize;
+}
+
+vlong
+domacholink(void)
+{
+	linkoff = 0;
+	if(nlinkdata > 0) {
+		linkoff = rnd(HEADR+textsize, INITRND) + rnd(datsize, INITRND);
+		seek(cout, linkoff, 0);
+		write(cout, linkdata, nlinkdata);
+		write(cout, strtab, nstrtab);
+	}
+	return rnd(nlinkdata+nstrtab, INITRND);
+}
+
+void
+asmbmacho(vlong symdatva, vlong symo)
+{
+	vlong v, w;
+	vlong va;
+	int a, i, ptrsize;
+	MachoHdr *mh;
+	MachoSect *msect;
+	MachoSeg *ms;
+	MachoDebug *md;
+	MachoLoad *ml;
+
+	/* apple MACH */
+	va = INITTEXT - HEADR;
+	mh = getMachoHdr();
+	switch(thechar){
+	default:
+		diag("unknown mach architecture");
+		errorexit();
+	case '6':
+		mh->cpu = MACHO_CPU_AMD64;
+		mh->subcpu = MACHO_SUBCPU_X86;
+		ptrsize = 8;
+		break;
+	case '8':
+		mh->cpu = MACHO_CPU_386;
+		mh->subcpu = MACHO_SUBCPU_X86;
+		ptrsize = 4;
+		break;
+	}
+
+	/* segment for zero page */
+	ms = newMachoSeg("__PAGEZERO", 0);
+	ms->vsize = va;
+
+	/* text */
+	v = rnd(HEADR+textsize, INITRND);
+	ms = newMachoSeg("__TEXT", 1);
+	ms->vaddr = va;
+	ms->vsize = v;
+	ms->filesize = v;
+	ms->prot1 = 7;
+	ms->prot2 = 5;
+
+	msect = newMachoSect(ms, "__text");
+	msect->addr = INITTEXT;
+	msect->size = textsize;
+	msect->off = INITTEXT - va;
+	msect->flag = 0x400;	/* flag - some instructions */
+
+	/* data */
+	w = datsize+dynptrsize+bsssize;
+	ms = newMachoSeg("__DATA", 2+(dynptrsize>0));
+	ms->vaddr = va+v;
+	ms->vsize = w;
+	ms->fileoffset = v;
+	ms->filesize = datsize;
+	ms->prot1 = 7;
+	ms->prot2 = 3;
+
+	msect = newMachoSect(ms, "__data");
+	msect->addr = va+v;
+	msect->size = datsize;
+	msect->off = v;
+
+	if(dynptrsize > 0) {
+		msect = newMachoSect(ms, "__nl_symbol_ptr");
+		msect->addr = va+v+datsize;
+		msect->size = dynptrsize;
+		msect->align = 2;
+		msect->flag = 6;	/* section with nonlazy symbol pointers */
+		/*
+		 * The reserved1 field is supposed to be the index of
+		 * the first entry in the list of symbol table indexes
+		 * in isymtab for the symbols we need.  We only use
+		 * pointers, so we need the entire list, so the index
+		 * here should be 0, which luckily is what the Mach-O
+		 * writing code emits by default for this not really reserved field.
+		msect->reserved1 = 0; - first indirect symbol table entry we need
+		 */
+	}
+
+	msect = newMachoSect(ms, "__bss");
+	msect->addr = va+v+datsize+dynptrsize;
+	msect->size = bsssize;
+	msect->flag = 1;	/* flag - zero fill */
+
+	switch(thechar) {
+	default:
+		diag("unknown macho architecture");
+		errorexit();
+	case '6':
+		ml = newMachoLoad(5, 42+2);	/* unix thread */
+		ml->data[0] = 4;	/* thread type */
+		ml->data[1] = 42;	/* word count */
+		ml->data[2+32] = entryvalue();	/* start pc */
+		ml->data[2+32+1] = entryvalue()>>32;
+		break;
+	case '8':
+		ml = newMachoLoad(5, 16+2);	/* unix thread */
+		ml->data[0] = 1;	/* thread type */
+		ml->data[1] = 16;	/* word count */
+		ml->data[2+10] = entryvalue();	/* start pc */
+		break;
+	}
+
+	if(!debug['d']) {
+		int nsym;
+
+		nsym = dynptrsize/ptrsize;
+
+		ms = newMachoSeg("__LINKEDIT", 0);
+		ms->vaddr = va+v+rnd(datsize+dynptrsize+bsssize, INITRND);
+		ms->vsize = nlinkdata+nstrtab;
+		ms->fileoffset = linkoff;
+		ms->filesize = nlinkdata+nstrtab;
+		ms->prot1 = 7;
+		ms->prot2 = 3;
+
+		ml = newMachoLoad(2, 4);	/* LC_SYMTAB */
+		ml->data[0] = linkoff;	/* symoff */
+		ml->data[1] = nsym;	/* nsyms */
+		ml->data[2] = linkoff + nlinkdata;	/* stroff */
+		ml->data[3] = nstrtab;	/* strsize */
+
+		ml = newMachoLoad(11, 18);	/* LC_DYSYMTAB */
+		ml->data[0] = 0;	/* ilocalsym */
+		ml->data[1] = 0;	/* nlocalsym */
+		ml->data[2] = 0;	/* iextdefsym */
+		ml->data[3] = 0;	/* nextdefsym */
+		ml->data[4] = 0;	/* iundefsym */
+		ml->data[5] = nsym;	/* nundefsym */
+		ml->data[6] = 0;	/* tocoffset */
+		ml->data[7] = 0;	/* ntoc */
+		ml->data[8] = 0;	/* modtaboff */
+		ml->data[9] = 0;	/* nmodtab */
+		ml->data[10] = 0;	/* extrefsymoff */
+		ml->data[11] = 0;	/* nextrefsyms */
+		ml->data[12] = linkoff + nlinkdata - nsym*4;	/* indirectsymoff */
+		ml->data[13] = nsym;	/* nindirectsyms */
+		ml->data[14] = 0;	/* extreloff */
+		ml->data[15] = 0;	/* nextrel */
+		ml->data[16] = 0;	/* locreloff */
+		ml->data[17] = 0;	/* nlocrel */
+
+		ml = newMachoLoad(14, 6);	/* LC_LOAD_DYLINKER */
+		ml->data[0] = 12;	/* offset to string */
+		strcpy((char*)&ml->data[1], "/usr/lib/dyld");
+
+		for(i=0; i<ndylib; i++) {
+			ml = newMachoLoad(12, 4+(strlen(dylib[i])+1+7)/8*2);	/* LC_LOAD_DYLIB */
+			ml->data[0] = 24;	/* offset of string from beginning of load */
+			ml->data[1] = 0;	/* time stamp */
+			ml->data[2] = 0;	/* version */
+			ml->data[3] = 0;	/* compatibility version */
+			strcpy((char*)&ml->data[4], dylib[i]);
+		}
+	}
+
+	if(!debug['s']) {
+		ms = newMachoSeg("__SYMDAT", 1);
+		ms->vaddr = symdatva;
+		ms->vsize = 8+symsize+lcsize;
+		ms->fileoffset = symo;
+		ms->filesize = 8+symsize+lcsize;
+		ms->prot1 = 7;
+		ms->prot2 = 5;
+
+		md = newMachoDebug();
+		md->fileoffset = symo+8;
+		md->filesize = symsize;
+
+		md = newMachoDebug();
+		md->fileoffset = symo+8+symsize;
+		md->filesize = lcsize;
+	}
+
+	a = machowrite();
+	if(a > MACHORESERVE)
+		diag("MACHORESERVE too small: %d > %d", a, MACHORESERVE);
+}
diff --git a/src/cmd/ld/macho.h b/src/cmd/ld/macho.h
index 747adac2d..a96b2a383 100644
--- a/src/cmd/ld/macho.h
+++ b/src/cmd/ld/macho.h
@@ -61,10 +61,17 @@ void	machoinit(void);
  * for Header, PHeaders, and SHeaders.
  * May waste some.
  */
-#define	MACHORESERVE	4096
+#define	MACHORESERVE	3*1024
 
 enum {
 	MACHO_CPU_AMD64 = (1<<24)|7,
 	MACHO_CPU_386 = 7,
 	MACHO_SUBCPU_X86 = 3,
+
+	MACHO32SYMSIZE = 12,
+	MACHO64SYMSIZE = 16,
 };
+
+void	domacho(void);
+vlong	domacholink(void);
+void	asmbmacho(vlong, vlong);