runtime.cmd/ld: Add ARM external linking and implement -shared in terms of external linking

This CL is an aggregate of 10271047, 10499043, 9733044. Descriptions of each follow:

10499043
runtime,cmd/ld: Merge TLS symbols and teach 5l about ARM TLS

This CL prepares for external linking support to ARM.

The pseudo-symbols runtime.g and runtime.m are merged into a single
runtime.tlsgm symbol. When external linking, the offset of a thread local
variable is stored at a memory location instead of being embedded into a offset
of a ldr instruction. With a single runtime.tlsgm symbol for both g and m, only
one such offset is needed.

The larger part of this CL moves TLS code from gcc compiled to internally
compiled. The TLS code now uses the modern MRC instruction, and 5l is taught
about TLS fallbacks in case the instruction is not available or appropriate.

10271047
This CL adds support for -linkmode external to 5l.

For 5l itself, use addrel to allow for D_CALL relocations to be handled by the
host linker. Of the cases listed in rsc's comment in issue 4069, only case 5 and
63 needed an update. One of the TODO: addrel cases was since replaced, and the
rest of the cases are either covered by indirection through addpool (cases with
LTO or LFROM flags) or stubs (case 74). The addpool cases are covered because
addpool emits AWORD instructions, which in turn are handled by case 11.

In the runtime, change the argv argument in the rt0* functions slightly to be a
pointer to the argv list, instead of relying on a particular location of argv.

9733044
The -shared flag to 6l outputs a shared library, implemented in Go
and callable from non-Go programs such as C.

The main part of this CL change the thread local storage model.
Go uses the fastest and least general mode, local exec. TLS data in shared
libraries normally requires at least the local dynamic mode, however, this CL
instead opts for using the initial exec mode. Initial exec mode is faster than
local dynamic mode and can be used in linux since the linker has reserved a
limited amount of TLS space for performance sensitive TLS code.

Initial exec mode requires an extra load from the GOT table to determine the
TLS offset. This penalty will not be paid if ld is not in -shared mode, since
TLS accesses will be reduced to local exec.

The elf sections .init_array and .rela.init_array are added to register the Go
runtime entry with cgo at library load time.

The "hidden" attribute is added to Cgo functions called from Go, since Go
does not generate call through the GOT table, and adding non-GOT relocations for
a global function is not supported by gcc. Cgo symbols don't need to be global
and avoiding the GOT table is also faster.

The changes to 8l are only removes code relevant to the old -shared mode where
internal linking was used.

This CL only address the low level linker work. It can be submitted by itself,
but to be useful, the runtime changes in CL 9738047 is also needed.

Design discussion at
https://groups.google.com/forum/?fromgroups#!topic/golang-nuts/zmjXkGrEx6Q

Fixes #5590.

R=rsc
CC=golang-dev
https://golang.org/cl/12871044

src/cmd/5l/5.out.h

@@ -273,7 +273,7 @@ enum	as
 #define	D_PLT1		(D_NONE+44) // R_ARM_PLT32, 2nd inst: add ip, ip, #0xNN000
 #define	D_PLT2		(D_NONE+45) // R_ARM_PLT32, 3rd inst: ldr pc, [ip, #0xNNN]!
 #define	D_CALL		(D_NONE+46) // R_ARM_PLT32/R_ARM_CALL/R_ARM_JUMP24, bl xxxxx or b yyyyy
-#define	D_TLS		(D_NONE+47)
+#define	D_TLS		(D_NONE+47) // R_ARM_TLS_LE32
 
 /*
  * this is the ranlib header

src/cmd/5l/asm.c

@@ -93,12 +93,6 @@ braddoff(int32 a, int32 b)
 	return (((uint32)a) & 0xff000000U) | (0x00ffffffU & (uint32)(a + b));
 }
 
-Sym *
-lookuprel(void)
-{
-	return lookup(".rel", 0);
-}
-
 void
 adddynrela(Sym *rel, Sym *s, Reloc *r)
 {
@@ -264,6 +258,26 @@ elfreloc1(Reloc *r, vlong sectoff)
 		else
 			return -1;
 		break;
+
+	case D_CALL:
+		if(r->siz == 4) {
+			if((r->add & 0xff000000) == 0xeb000000) // BL
+				LPUT(R_ARM_CALL | elfsym<<8);
+			else
+				LPUT(R_ARM_JUMP24 | elfsym<<8);
+		} else
+			return -1;
+		break;
+
+	case D_TLS:
+		if(r->siz == 4) {
+			if(flag_shared)
+				LPUT(R_ARM_TLS_IE32 | elfsym<<8);
+			else
+				LPUT(R_ARM_TLS_LE32 | elfsym<<8);
+		} else
+			return -1;
+		break;
 	}
 
 	return 0;
@@ -308,6 +322,34 @@ machoreloc1(Reloc *r, vlong sectoff)
 int
 archreloc(Reloc *r, Sym *s, vlong *val)
 {
+	Sym *rs;
+
+	if(linkmode == LinkExternal) {
+		switch(r->type) {
+		case D_CALL:
+			r->done = 0;
+
+			// set up addend for eventual relocation via outer symbol.
+			rs = r->sym;
+			r->xadd = r->add;
+			if(r->xadd & 0x800000)
+				r->xadd |= ~0xffffff;
+			r->xadd *= 4;
+			while(rs->outer != nil) {
+				r->xadd += symaddr(rs) - symaddr(rs->outer);
+				rs = rs->outer;
+			}
+
+			if(rs->type != SHOSTOBJ && rs->sect == nil)
+				diag("missing section for %s", rs->name);
+			r->xsym = rs;
+
+			*val = braddoff((0xff000000U & (uint32)r->add), 
+							(0xffffff & (uint32)(r->xadd / 4)));
+			return 0;
+		}
+		return -1;
+	}
 	switch(r->type) {
 	case D_CONST:
 		*val = r->add;
@@ -766,7 +808,7 @@ nopstat(char *f, Count *c)
 }
 
 void
-asmout(Prog *p, Optab *o, int32 *out)
+asmout(Prog *p, Optab *o, int32 *out, Sym *gmsym)
 {
 	int32 o1, o2, o3, o4, o5, o6, v;
 	int r, rf, rt, rt2;
@@ -849,11 +891,19 @@ if(debug['G']) print("%ux: %s: arm %d\n", (uint32)(p->pc), p->from.sym->name, p-
 		break;
 
 	case 5:		/* bra s */
+		o1 = opbra(p->as, p->scond);
 		v = -8;
-		// TODO: Use addrel.
+		if(p->to.sym != S && p->to.sym->type != 0) {
+			rel = addrel(cursym);
+			rel->off = pc - cursym->value;
+			rel->siz = 4;
+			rel->sym = p->to.sym;
+			rel->add = o1 | ((v >> 2) & 0xffffff);
+			rel->type = D_CALL;
+			break;
+		}
 		if(p->cond != P)
 			v = (p->cond->pc - pc) - 8;
-		o1 = opbra(p->as, p->scond);
 		o1 |= (v >> 2) & 0xffffff;
 		break;
 
@@ -911,7 +961,13 @@ if(debug['G']) print("%ux: %s: arm %d\n", (uint32)(p->pc), p->from.sym->name, p-
 			rel->siz = 4;
 			rel->sym = p->to.sym;
 			rel->add = p->to.offset;
-			if(flag_shared) {
+			if(rel->sym == gmsym) {
+				rel->type = D_TLS;
+				if(flag_shared)
+					rel->add += pc - p->pcrel->pc - 8 - rel->siz;
+				rel->xadd = rel->add;
+				rel->xsym = rel->sym;
+			} else if(flag_shared) {
 				rel->type = D_PCREL;
 				rel->add += pc - p->pcrel->pc - 8;
 			} else
@@ -1242,9 +1298,22 @@ if(debug['G']) print("%ux: %s: arm %d\n", (uint32)(p->pc), p->from.sym->name, p-
 
 	case 63:	/* bcase */
 		if(p->cond != P) {
-			o1 = p->cond->pc;
-			if(flag_shared)
-				o1 = o1 - p->pcrel->pc - 16;
+			rel = addrel(cursym);
+			rel->off = pc - cursym->value;
+			rel->siz = 4;
+			if(p->to.sym != S && p->to.sym->type != 0) {
+				rel->sym = p->to.sym;
+				rel->add = p->to.offset;
+			} else {
+				rel->sym = cursym;
+				rel->add = p->cond->pc - cursym->value;
+			}
+			if(o->flag & LPCREL) {
+				rel->type = D_PCREL;
+				rel->add += pc - p->pcrel->pc - 16 + rel->siz;
+			} else
+				rel->type = D_ADDR;
+			o1 = 0;
 		}
 		break;
 

src/cmd/5l/l.h

@@ -183,7 +183,6 @@ struct	Sym
 	Reloc*	r;
 	int32	nr;
 	int32	maxr;
-	int 	rel_ro;
 };
 
 #define SIGNINTERN	(1729*325*1729)
@@ -293,7 +292,6 @@ EXTERN	int32	INITDAT;		/* data location */
 EXTERN	int32	INITRND;		/* data round above text location */
 EXTERN	int32	INITTEXT;		/* text location */
 EXTERN	char*	INITENTRY;		/* entry point */
-EXTERN	char*	LIBINITENTRY;		/* shared library entry point */
 EXTERN	int32	autosize;
 EXTERN	Auto*	curauto;
 EXTERN	Auto*	curhist;
@@ -364,7 +362,7 @@ int	aclass(Adr*);
 void	addhist(int32, int);
 Prog*	appendp(Prog*);
 void	asmb(void);
-void	asmout(Prog*, Optab*, int32*);
+void	asmout(Prog*, Optab*, int32*, Sym*);
 int32	atolwhex(char*);
 Prog*	brloop(Prog*);
 void	buildop(void);

src/cmd/5l/noop.c

@@ -60,13 +60,14 @@ noops(void)
 	int o;
 	int32 arg;
 	Prog *pmorestack;
-	Sym *symmorestack;
+	Sym *symmorestack, *tlsfallback, *gmsym;
 
 	/*
 	 * find leaf subroutines
 	 * strip NOPs
 	 * expand RET
 	 * expand BECOME pseudo
+	 * fixup TLS
 	 */
 
 	if(debug['v'])
@@ -81,6 +82,10 @@ noops(void)
 	pmorestack = symmorestack->text;
 	pmorestack->reg |= NOSPLIT;
 
+	tlsfallback = lookup("runtime.read_tls_fallback", 0);
+	gmsym = S;
+	if(linkmode == LinkExternal)
+		gmsym = lookup("runtime.tlsgm", 0);
 	q = P;
 	for(cursym = textp; cursym != nil; cursym = cursym->next) {
 		for(p = cursym->text; p != P; p = p->link) {
@@ -145,6 +150,82 @@ noops(void)
 					}
 				}
 				break;
+			case AWORD:
+				// Rewrite TLS register fetch: MRC 15, 0, <reg>, C13, C0, 3
+				if((p->to.offset & 0xffff0fff) == 0xee1d0f70) {
+					if(HEADTYPE == Hopenbsd) {
+						p->as = ARET;
+					} else if(goarm < 7) {
+						if(tlsfallback->type != STEXT) {
+							diag("runtime·read_tls_fallback not defined");
+							errorexit();
+						}
+						// BL runtime.read_tls_fallback(SB)
+						p->as = ABL;
+						p->to.type = D_BRANCH;
+						p->to.sym = tlsfallback;
+						p->cond = tlsfallback->text;
+						p->to.offset = 0;
+						cursym->text->mark &= ~LEAF;
+					}
+					if(linkmode == LinkExternal) {
+						// runtime.tlsgm is relocated with R_ARM_TLS_LE32
+						// and $runtime.tlsgm will contain the TLS offset.
+						//
+						// MOV $runtime.tlsgm+tlsoffset(SB), REGTMP
+						// ADD REGTMP, <reg>
+						//
+						// In shared mode, runtime.tlsgm is relocated with
+						// R_ARM_TLS_IE32 and runtime.tlsgm(SB) will point
+						// to the GOT entry containing the TLS offset.
+						//
+						// MOV runtime.tlsgm(SB), REGTMP
+						// ADD REGTMP, <reg>
+						// SUB -tlsoffset, <reg>
+						//
+						// The SUB compensates for tlsoffset
+						// used in runtime.save_gm and runtime.load_gm.
+						q = p;
+						p = appendp(p);
+						p->as = AMOVW;
+						p->scond = 14;
+						p->reg = NREG;
+						if(flag_shared) {
+							p->from.type = D_OREG;
+							p->from.offset = 0;
+						} else {
+							p->from.type = D_CONST;
+							p->from.offset = tlsoffset;
+						}
+						p->from.sym = gmsym;
+						p->from.name = D_EXTERN;
+						p->to.type = D_REG;
+						p->to.reg = REGTMP;
+						p->to.offset = 0;
+
+						p = appendp(p);
+						p->as = AADD;
+						p->scond = 14;
+						p->reg = NREG;
+						p->from.type = D_REG;
+						p->from.reg = REGTMP;
+						p->to.type = D_REG;
+						p->to.reg = (q->to.offset & 0xf000) >> 12;
+						p->to.offset = 0;
+
+						if(flag_shared) {
+							p = appendp(p);
+							p->as = ASUB;
+							p->scond = 14;
+							p->reg = NREG;
+							p->from.type = D_CONST;
+							p->from.offset = -tlsoffset;
+							p->to.type = D_REG;
+							p->to.reg = (q->to.offset & 0xf000) >> 12;
+							p->to.offset = 0;
+						}
+					}
+				}
 			}
 			q = p;
 		}

src/cmd/5l/obj.c

@@ -81,8 +81,7 @@ main(int argc, char *argv[])
 	INITDAT = -1;
 	INITRND = -1;
 	INITENTRY = 0;
-	LIBINITENTRY = 0;
-	linkmode = LinkInternal; // TODO: LinkAuto once everything works.
+	linkmode = LinkAuto;
 	nuxiinit();
 	
 	p = getgoarm();
@@ -126,34 +125,43 @@ main(int argc, char *argv[])
 	flagstr("r", "dir1:dir2:...: set ELF dynamic linker search path", &rpath);
 	flagcount("race", "enable race detector", &flag_race);
 	flagcount("s", "disable symbol table", &debug['s']);
+	flagcount("shared", "generate shared object (implies -linkmode external)", &flag_shared);
 	flagstr("tmpdir", "leave temporary files in this directory", &tmpdir);
 	flagcount("u", "reject unsafe packages", &debug['u']);
 	flagcount("v", "print link trace", &debug['v']);
 	flagcount("w", "disable DWARF generation", &debug['w']);
-	flagcount("shared", "generate shared object", &flag_shared);
-	// TODO: link mode flag
 	
 	flagparse(&argc, &argv, usage);
 
 	if(argc != 1)
 		usage();
 
+	if(flag_shared)
+		linkmode = LinkExternal;
+
+	mywhatsys();
+
+	if(HEADTYPE == -1)
+		HEADTYPE = headtype(goos);
+
 	// getgoextlinkenabled is based on GO_EXTLINK_ENABLED when
 	// Go was built; see ../../make.bash.
 	if(linkmode == LinkAuto && strcmp(getgoextlinkenabled(), "0") == 0)
 		linkmode = LinkInternal;
 
-	if(linkmode == LinkExternal) {
-		diag("only -linkmode=internal is supported");
-		errorexit();
-	} else if(linkmode == LinkAuto) {
-		linkmode = LinkInternal;
+	switch(HEADTYPE) {
+	default:
+		if(linkmode == LinkAuto)
+			linkmode = LinkInternal;
+		if(linkmode == LinkExternal && strcmp(getgoextlinkenabled(), "1") != 0)
+			sysfatal("cannot use -linkmode=external with -H %s", headstr(HEADTYPE));
+		break;
+	case Hlinux:
+		break;
 	}
 
 	libinit();
 
-	if(HEADTYPE == -1)
-		HEADTYPE = headtype(goos);
 	switch(HEADTYPE) {
 	default:
 		diag("unknown -H option");
@@ -208,7 +216,7 @@ main(int argc, char *argv[])
 	case Hnetbsd:
 		debug['d'] = 0;	// with dynamic linking
 		tlsoffset = -8; // hardcoded number, first 4-byte word for g, and then 4-byte word for m
-		                // this number is known to ../../pkg/runtime/cgo/gcc_linux_arm.c
+		                // this number is known to ../../pkg/runtime/rt0_*_arm.s
 		elfinit();
 		HEADR = ELFRESERVE;
 		if(INITTEXT == -1)
@@ -253,6 +261,7 @@ main(int argc, char *argv[])
 	// mark some functions that are only referenced after linker code editing
 	if(debug['F'])
 		mark(rlookup("_sfloat", 0));
+	mark(lookup("runtime.read_tls_fallback", 0));
 	deadcode();
 	if(textp == nil) {
 		diag("no code");

src/cmd/5l/optab.c

@@ -191,7 +191,7 @@ Optab	optab[] =
 	{ AMOVBU,	C_REG,	C_NONE,	C_SHIFT,	61, 4, 0 },
 
 	{ ACASE,	C_REG,	C_NONE,	C_NONE,		62, 4, 0, LPCREL, 8 },
-	{ ABCASE,	C_NONE, C_NONE, C_SBRA,		63, 4, 0 },
+	{ ABCASE,	C_NONE, C_NONE, C_SBRA,		63, 4, 0, LPCREL, 0 },
 
 	{ AMOVH,	C_REG,	C_NONE, C_HAUTO,	70, 4, REGSP,	0 },
 	{ AMOVH,	C_REG,	C_NONE,	C_HOREG,	70, 4, 0,	0 },

src/cmd/5l/pass.c

@@ -246,6 +246,13 @@ patch(void)
 			p->cond = q;
 		}
 	}
+	if(flag_shared) {
+		s = lookup("init_array", 0);
+		s->type = SINITARR;
+		s->reachable = 1;
+		s->hide = 1;
+		addaddr(s, lookup(INITENTRY, 0));
+	}
 
 	for(cursym = textp; cursym != nil; cursym = cursym->next) {
 		for(p = cursym->text; p != P; p = p->link) {

src/cmd/5l/span.c

@@ -90,7 +90,7 @@ span(void)
 	int32 c, otxt, out[6];
 	Section *sect;
 	uchar *bp;
-	Sym *sub;
+	Sym *sub, *gmsym;
 
 	if(debug['v'])
 		Bprint(&bso, "%5.2f span\n", cputime());
@@ -237,6 +237,9 @@ span(void)
 	 * code references to be relocated too, and then
 	 * perhaps we'd be able to parallelize the span loop above.
 	 */
+	gmsym = S;
+	if(linkmode == LinkExternal)
+		gmsym = lookup("runtime.tlsgm", 0);
 	for(cursym = textp; cursym != nil; cursym = cursym->next) {
 		p = cursym->text;
 		if(p == P || p->link == P)
@@ -249,7 +252,7 @@ span(void)
 			pc = p->pc;
 			curp = p;
 			o = oplook(p);
-			asmout(p, o, out);
+			asmout(p, o, out, gmsym);
 			for(i=0; i<o->size/4; i++) {
 				v = out[i];
 				*bp++ = v;
@@ -574,10 +577,7 @@ aclass(Adr *a)
 			if(s == S)
 				break;
 			instoffset = 0;	// s.b. unused but just in case
-			if(flag_shared)
-				return C_LCONADDR;
-			else
-				return C_LCON;
+			return C_LCONADDR;
 
 		case D_AUTO:
 			instoffset = autosize + a->offset;

src/cmd/6a/a.y

@@ -518,6 +518,15 @@ omem:
 		$$.scale = $8;
 		checkscale($$.scale);
 	}
+|	con '(' LLREG ')' '(' LSREG '*' con ')'
+	{
+		$$ = nullgen;
+		$$.type = D_INDIR+$3;
+		$$.offset = $1;
+		$$.index = $6;
+		$$.scale = $8;
+		checkscale($$.scale);
+	}
 |	'(' LLREG ')'
 	{
 		$$ = nullgen;

src/cmd/6l/asm.c

@@ -99,12 +99,6 @@ int nelfsym = 1;
 static void addpltsym(Sym*);
 static void addgotsym(Sym*);
 
-Sym *
-lookuprel(void)
-{
-	return lookup(".rela", 0);
-}
-
 void
 adddynrela(Sym *rela, Sym *s, Reloc *r)
 {
@@ -312,9 +306,12 @@ elfreloc1(Reloc *r, vlong sectoff)
 		break;
 	
 	case D_TLS:
-		if(r->siz == 4)
-			VPUT(R_X86_64_TPOFF32 | (uint64)elfsym<<32);
-		else
+		if(r->siz == 4) {
+			if(flag_shared)
+				VPUT(R_X86_64_GOTTPOFF | (uint64)elfsym<<32);
+			else
+				VPUT(R_X86_64_TPOFF32 | (uint64)elfsym<<32);
+		} else
 			return -1;
 		break;		
 	}

src/cmd/6l/l.h

@@ -187,7 +187,6 @@ struct	Sym
 	Reloc*	r;
 	int32	nr;
 	int32	maxr;
-	int 	rel_ro;
 };
 struct	Optab
 {
@@ -329,7 +328,6 @@ EXTERN	int32	INITRND;
 EXTERN	int64	INITTEXT;
 EXTERN	int64	INITDAT;
 EXTERN	char*	INITENTRY;		/* entry point */
-EXTERN	char*	LIBINITENTRY;		/* shared library entry point */
 EXTERN	char*	pcstr;
 EXTERN	Auto*	curauto;
 EXTERN	Auto*	curhist;

src/cmd/6l/obj.c

@@ -82,7 +82,6 @@ main(int argc, char *argv[])
 	INITDAT = -1;
 	INITRND = -1;
 	INITENTRY = 0;
-	LIBINITENTRY = 0;
 	linkmode = LinkAuto;
 	nuxiinit();
 
@@ -119,7 +118,7 @@ main(int argc, char *argv[])
 	flagstr("r", "dir1:dir2:...: set ELF dynamic linker search path", &rpath);
 	flagcount("race", "enable race detector", &flag_race);
 	flagcount("s", "disable symbol table", &debug['s']);
-	flagcount("shared", "generate shared object", &flag_shared);
+	flagcount("shared", "generate shared object (implies -linkmode external)", &flag_shared);
 	flagstr("tmpdir", "leave temporary files in this directory", &tmpdir);
 	flagcount("u", "reject unsafe packages", &debug['u']);
 	flagcount("v", "print link trace", &debug['v']);
@@ -140,6 +139,9 @@ main(int argc, char *argv[])
 	if(linkmode == LinkAuto && strcmp(getgoextlinkenabled(), "0") == 0)
 		linkmode = LinkInternal;
 
+	if(flag_shared)
+		linkmode = LinkExternal;
+
 	switch(HEADTYPE) {
 	default:
 		if(linkmode == LinkAuto)
@@ -365,7 +367,7 @@ zaddr(char *pn, Biobuf *f, Adr *a, Sym *h[])
 		adrgotype = zsym(pn, f, h);
 	s = a->sym;
 	t = a->type;
-	if(t == D_INDIR+D_GS)
+	if(t == D_INDIR+D_GS || a->index == D_GS)
 		a->offset += tlsoffset;
 	if(t != D_AUTO && t != D_PARAM) {
 		if(s && adrgotype)

src/cmd/6l/pass.c

@@ -271,7 +271,7 @@ patch(void)
 {
 	int32 c;
 	Prog *p, *q;
-	Sym *s;
+	Sym *s, *gmsym;
 	int32 vexit;
 
 	if(debug['v'])
@@ -282,6 +282,17 @@ patch(void)
 		Bprint(&bso, "%5.2f patch\n", cputime());
 	Bflush(&bso);
 
+	if(flag_shared) {
+		s = lookup("init_array", 0);
+		s->type = SINITARR;
+		s->reachable = 1;
+		s->hide = 1;
+		addaddr(s, lookup(INITENTRY, 0));
+	}
+
+	gmsym = lookup("runtime.tlsgm", 0);
+	if(linkmode != LinkExternal)
+		gmsym->reachable = 0;
 	s = lookup("exit", 0);
 	vexit = s->value;
 	for(cursym = textp; cursym != nil; cursym = cursym->next)
@@ -317,6 +328,59 @@ patch(void)
 				p->from.type = D_INDIR+D_FS;
 			if(p->to.type == D_INDIR+D_GS)
 				p->to.type = D_INDIR+D_FS;
+			if(p->from.index == D_GS)
+				p->from.index = D_FS;
+			if(p->to.index == D_GS)
+				p->to.index = D_FS;
+		}
+		if(!flag_shared) {
+			// Convert g() or m() accesses of the form
+			//   op n(reg)(GS*1), reg
+			// to
+			//   op n(GS*1), reg
+			if(p->from.index == D_FS || p->from.index == D_GS) {
+				p->from.type = D_INDIR + p->from.index;
+				p->from.index = D_NONE;
+			}
+			// Convert g() or m() accesses of the form
+			//   op reg, n(reg)(GS*1)
+			// to
+			//   op reg, n(GS*1)
+			if(p->to.index == D_FS || p->to.index == D_GS) {
+				p->to.type = D_INDIR + p->to.index;
+				p->to.index = D_NONE;
+			}
+			// Convert get_tls access of the form
+			//   op runtime.tlsgm(SB), reg
+			// to
+			//   NOP
+			if(gmsym != S && p->from.sym == gmsym) {
+				p->as = ANOP;
+				p->from.type = D_NONE;
+				p->to.type = D_NONE;
+				p->from.sym = nil;
+				p->to.sym = nil;
+				continue;
+			}
+		} else {
+			// Convert TLS reads of the form
+			//   op n(GS), reg
+			// to
+			//   MOVQ $runtime.tlsgm(SB), reg
+			//   op n(reg)(GS*1), reg
+			if((p->from.type == D_INDIR+D_FS || p->from.type == D_INDIR + D_GS) && p->to.type >= D_AX && p->to.type <= D_DI) {
+				q = appendp(p);
+				q->to = p->to;
+				q->as = p->as;
+				q->from.type = D_INDIR+p->to.type;
+				q->from.index = p->from.type - D_INDIR;
+				q->from.scale = 1;
+				q->from.offset = p->from.offset;
+				p->as = AMOVQ;
+				p->from.type = D_EXTERN;
+				p->from.sym = gmsym;
+				p->from.offset = 0;
+			}
 		}
 		if(p->as == ACALL || (p->as == AJMP && p->to.type != D_BRANCH) || (p->as == ARET && p->to.sym != nil)) {
 			s = p->to.sym;
@@ -411,7 +475,10 @@ dostkoff(void)
 	int32 autoffset, deltasp;
 	int a, pcsize;
 	uint32 moreconst1, moreconst2, i;
+	Sym *gmsym;
 
+
+	gmsym = lookup("runtime.tlsgm", 0);
 	for(i=0; i<nelem(morename); i++) {
 		symmorestack[i] = lookup(morename[i], 0);
 		if(symmorestack[i]->type != STEXT)
@@ -443,6 +510,14 @@ dostkoff(void)
 			diag("nosplit func likely to overflow stack");
 
 		if(!(p->from.scale & NOSPLIT)) {
+			if(flag_shared) {
+				// Load TLS offset with MOVQ $runtime.tlsgm(SB), CX
+				p = appendp(p);
+				p->as = AMOVQ;
+				p->from.type = D_EXTERN;
+				p->from.sym = gmsym;
+				p->to.type = D_CX;
+			}
 			p = appendp(p);	// load g into CX
 			p->as = AMOVQ;
 			if(HEADTYPE == Hlinux || HEADTYPE == Hfreebsd
@@ -451,6 +526,11 @@ dostkoff(void)
 				p->from.type = D_INDIR+D_FS;
 			else
 				p->from.type = D_INDIR+D_GS;
+			if(flag_shared) {
+				// Add TLS offset stored in CX
+				p->from.index = p->from.type - D_INDIR;
+				p->from.type = D_INDIR + D_CX;
+			}
 			p->from.offset = tlsoffset+0;
 			p->to.type = D_CX;
 			if(HEADTYPE == Hwindows) {

src/cmd/6l/span.c

@@ -358,6 +358,18 @@ prefixof(Adr *a)
 	case D_INDIR+D_GS:
 		return 0x65;
 	}
+	switch(a->index) {
+	case D_CS:
+		return 0x2e;
+	case D_DS:
+		return 0x3e;
+	case D_ES:
+		return 0x26;
+	case D_FS:
+		return 0x64;
+	case D_GS:
+		return 0x65;
+	}
 	return 0;
 }
 
@@ -735,15 +747,20 @@ vaddr(Adr *a, Reloc *r)
 			diag("need reloc for %D", a);
 			errorexit();
 		}
-		if(flag_shared)
-			r->type = D_PCREL;
-		else
-			r->type = D_ADDR;
 		r->siz = 4;	// TODO: 8 for external symbols
 		r->off = -1;	// caller must fill in
 		r->sym = s;
 		r->add = v;
 		v = 0;
+		if(flag_shared) {
+			if(s->type == STLSBSS) {
+				r->xadd = r->add - r->siz;
+				r->type = D_TLS;
+				r->xsym = s;
+			} else
+				r->type = D_PCREL;
+		} else
+			r->type = D_ADDR;
 	}
 	return v;
 }
@@ -760,7 +777,7 @@ asmandsz(Adr *a, int r, int rex, int m64)
 	v = a->offset;
 	t = a->type;
 	rel.siz = 0;
-	if(a->index != D_NONE) {
+	if(a->index != D_NONE && a->index != D_FS && a->index != D_GS) {
 		if(t < D_INDIR) { 
 			switch(t) {
 			default:
@@ -888,18 +905,11 @@ putrelv:
 		
 		r = addrel(cursym);
 		r->off = curp->pc + andptr - and;
-		r->add = 0;
-		r->xadd = 0;
+		r->add = a->offset-tlsoffset;
+		r->xadd = r->add;
 		r->siz = 4;
 		r->type = D_TLS;
-		if(a->offset == tlsoffset+0)
-			s = lookup("runtime.g", 0);
-		else
-			s = lookup("runtime.m", 0);
-		s->type = STLSBSS;
-		s->reachable = 1;
-		s->size = PtrSize;
-		s->hide = 1;
+		s = lookup("runtime.tlsgm", 0);
 		r->sym = s;
 		r->xsym = s;
 		v = 0;

src/cmd/8l/asm.c

@@ -95,12 +95,6 @@ int	nelfsym = 1;
 static void	addpltsym(Sym*);
 static void	addgotsym(Sym*);
 
-Sym *
-lookuprel(void)
-{
-	return lookup(".rel", 0);
-}
-
 void
 adddynrela(Sym *rela, Sym *s, Reloc *r)
 {
@@ -366,6 +360,8 @@ int
 archreloc(Reloc *r, Sym *s, vlong *val)
 {
 	USED(s);
+	if(linkmode == LinkExternal)
+		return -1;
 	switch(r->type) {
 	case D_CONST:
 		*val = r->add;

src/cmd/8l/l.h

@@ -169,7 +169,6 @@ struct	Sym
 	Reloc*	r;
 	int32	nr;
 	int32	maxr;
-	int 	rel_ro;
 };
 struct	Optab
 {
@@ -285,7 +284,6 @@ EXTERN	int32	INITRND;
 EXTERN	int32	INITTEXT;
 EXTERN	int32	INITDAT;
 EXTERN	char*	INITENTRY;		/* entry point */
-EXTERN	char*	LIBINITENTRY;		/* shared library entry point */
 EXTERN	char*	pcstr;
 EXTERN	Auto*	curauto;
 EXTERN	Auto*	curhist;

src/cmd/8l/obj.c

@@ -89,7 +89,6 @@ main(int argc, char *argv[])
 	INITDAT = -1;
 	INITRND = -1;
 	INITENTRY = 0;
-	LIBINITENTRY = 0;
 	linkmode = LinkAuto;
 	nuxiinit();
 

src/cmd/8l/span.c

@@ -695,18 +695,11 @@ putrelv:
 
 		r = addrel(cursym);
 		r->off = curp->pc + andptr - and;
-		r->add = 0;
-		r->xadd = 0;
+		r->add = a->offset-tlsoffset;
+		r->xadd = r->add;
 		r->siz = 4;
 		r->type = D_TLS;
-		if(a->offset == tlsoffset+0)
-			s = lookup("runtime.g", 0);
-		else
-			s = lookup("runtime.m", 0);
-		s->type = STLSBSS;
-		s->reachable = 1;
-		s->hide = 1;
-		s->size = PtrSize;
+		s = lookup("runtime.tlsgm", 0);
 		r->sym = s;
 		r->xsym = s;
 		v = 0;

src/cmd/cgo/out.go

@@ -494,7 +494,7 @@ func (p *Package) writeOutputFunc(fgcc *os.File, n *Name) {
 
 	// Gcc wrapper unpacks the C argument struct
 	// and calls the actual C function.
-	fmt.Fprintf(fgcc, "void\n")
+	fmt.Fprintf(fgcc, "__attribute__ ((visibility (\"hidden\"))) void\n")
 	fmt.Fprintf(fgcc, "_cgo%s%s(void *v)\n", cPrefix, n.Mangle)
 	fmt.Fprintf(fgcc, "{\n")
 	if n.AddError {

src/cmd/dist/buildruntime.c

@@ -162,15 +162,19 @@ static struct {
 		"#define	m(r) 8(GS)\n"
 		"#define	procid(r) 16(GS)\n"
 	},
+	// The TLS accessors here are defined here to use initial exec model.
+	// If the linker is not outputting a shared library, it will reduce
+	// the TLS accessors to the local exec model, effectively removing
+	// get_tls().
 	{"amd64", "",
 		"// The offsets 0 and 8 are known to:\n"
 		"//	../../cmd/6l/pass.c:/D_GS\n"
 		"//	cgo/gcc_linux_amd64.c:/^threadentry\n"
 		"//	cgo/gcc_darwin_amd64.c:/^threadentry\n"
 		"//\n"
-		"#define	get_tls(r)\n"
-		"#define	g(r) 0(GS)\n"
-		"#define	m(r) 8(GS)\n"
+		"#define	get_tls(r) MOVQ runtime·tlsgm(SB), r\n"
+		"#define	g(r) 0(r)(GS*1)\n"
+		"#define	m(r) 8(r)(GS*1)\n"
 	},
 	
 	{"arm", "",

src/cmd/ld/data.c

@@ -178,12 +178,14 @@ relocsym(Sym *s)
 		switch(r->type) {
 		default:
 			o = 0;
-			if(linkmode == LinkExternal || archreloc(r, s, &o) < 0)
+			if(archreloc(r, s, &o) < 0)
 				diag("unknown reloc %d", r->type);
 			break;
 		case D_TLS:
 			r->done = 0;
 			o = 0;
+			if(thechar != '6')
+				o = r->add;
 			break;
 		case D_ADDR:
 			if(linkmode == LinkExternal && r->sym->type != SCONST) {
@@ -305,8 +307,6 @@ void
 dynrelocsym(Sym *s)
 {
 	Reloc *r;
-	Sym *rel;
-	Sym *got;
 	
 	if(HEADTYPE == Hwindows) {
 		Sym *rel, *targ;
@@ -343,22 +343,9 @@ dynrelocsym(Sym *s)
 		return;
 	}
 
-	got = rel = nil;
-	if(flag_shared) {
-		rel = lookuprel();
-		got = lookup(".got", 0);
-	}
-	s->rel_ro = 0;
 	for(r=s->r; r<s->r+s->nr; r++) {
 		if(r->sym != S && r->sym->type == SDYNIMPORT || r->type >= 256)
 			adddynrel(s, r);
-		if(flag_shared && r->sym != S && s->type != SDYNIMPORT && r->type == D_ADDR
-				&& (s == got || s->type == SDATA || s->type == SGOSTRING || s->type == STYPE || s->type == SRODATA)) {
-			// Create address based RELATIVE relocation
-			adddynrela(rel, s, r);
-			if(s->type < SNOPTRDATA)
-				s->rel_ro = 1;
-		}
 	}
 }
 
@@ -1099,12 +1086,6 @@ dodata(void)
 	}
 	*l = nil;
 
-	if(flag_shared) {
-		for(s=datap; s != nil; s = s->next) {
-			if(s->rel_ro)
-				s->type = SDATARELRO;
-		}
-	}
 	datap = listsort(datap, datcmp, offsetof(Sym, next));
 
 	/*
@@ -1138,12 +1119,12 @@ dodata(void)
 
 	/* pointer-free data */
 	sect = addsection(&segdata, ".noptrdata", 06);
-	sect->align = maxalign(s, SDATARELRO-1);
+	sect->align = maxalign(s, SINITARR-1);
 	datsize = rnd(datsize, sect->align);
 	sect->vaddr = datsize;
 	lookup("noptrdata", 0)->sect = sect;
 	lookup("enoptrdata", 0)->sect = sect;
-	for(; s != nil && s->type < SDATARELRO; s = s->next) {
+	for(; s != nil && s->type < SINITARR; s = s->next) {
 		datsize = aligndatsize(datsize, s);
 		s->sect = sect;
 		s->type = SDATA;
@@ -1152,18 +1133,15 @@ dodata(void)
 	}
 	sect->len = datsize - sect->vaddr;
 
-	/* dynamic relocated rodata */
+	/* shared library initializer */
 	if(flag_shared) {
-		sect = addsection(&segdata, ".data.rel.ro", 06);
-		sect->align = maxalign(s, SDATARELRO);
+		sect = addsection(&segdata, ".init_array", 06);
+		sect->align = maxalign(s, SINITARR);
 		datsize = rnd(datsize, sect->align);
 		sect->vaddr = datsize;
-		lookup("datarelro", 0)->sect = sect;
-		lookup("edatarelro", 0)->sect = sect;
-		for(; s != nil && s->type == SDATARELRO; s = s->next) {
+		for(; s != nil && s->type == SINITARR; s = s->next) {
 			datsize = aligndatsize(datsize, s);
 			s->sect = sect;
-			s->type = SDATA;
 			s->value = datsize - sect->vaddr;
 			growdatsize(&datsize, s);
 		}
@@ -1178,7 +1156,7 @@ dodata(void)
 	lookup("data", 0)->sect = sect;
 	lookup("edata", 0)->sect = sect;
 	for(; s != nil && s->type < SBSS; s = s->next) {
-		if(s->type == SDATARELRO) {
+		if(s->type == SINITARR) {
 			cursym = s;
 			diag("unexpected symbol type %d", s->type);
 		}
@@ -1423,7 +1401,7 @@ textaddress(void)
 void
 address(void)
 {
-	Section *s, *text, *data, *rodata, *symtab, *pclntab, *noptr, *bss, *noptrbss, *datarelro;
+	Section *s, *text, *data, *rodata, *symtab, *pclntab, *noptr, *bss, *noptrbss;
 	Section *typelink;
 	Sym *sym, *sub;
 	uvlong va;
@@ -1473,7 +1451,6 @@ address(void)
 	noptr = nil;
 	bss = nil;
 	noptrbss = nil;
-	datarelro = nil;
 	for(s=segdata.sect; s != nil; s=s->next) {
 		vlen = s->len;
 		if(s->next)
@@ -1489,8 +1466,6 @@ address(void)
 			bss = s;
 		if(strcmp(s->name, ".noptrbss") == 0)
 			noptrbss = s;
-		if(strcmp(s->name, ".data.rel.ro") == 0)
-			datarelro = s;
 	}
 	segdata.filelen = bss->vaddr - segdata.vaddr;
 
@@ -1516,10 +1491,6 @@ address(void)
 	xdefine("erodata", SRODATA, rodata->vaddr + rodata->len);
 	xdefine("typelink", SRODATA, typelink->vaddr);
 	xdefine("etypelink", SRODATA, typelink->vaddr + typelink->len);
-	if(datarelro != nil) {
-		xdefine("datarelro", SRODATA, datarelro->vaddr);
-		xdefine("edatarelro", SRODATA, datarelro->vaddr + datarelro->len);
-	}
 
 	sym = lookup("gcdata", 0);
 	xdefine("egcdata", SRODATA, symaddr(sym) + sym->size);

src/cmd/ld/elf.c

@@ -905,8 +905,6 @@ doelf(void)
 	addstring(shstrtab, ".elfdata");
 	addstring(shstrtab, ".rodata");
 	addstring(shstrtab, ".typelink");
-	if(flag_shared)
-		addstring(shstrtab, ".data.rel.ro");
 	addstring(shstrtab, ".gosymtab");
 	addstring(shstrtab, ".gopclntab");
 	
@@ -936,6 +934,14 @@ doelf(void)
 		addstring(shstrtab, ".note.GNU-stack");
 	}
 
+	if(flag_shared) {
+		addstring(shstrtab, ".init_array");
+		if(thechar == '6')
+			addstring(shstrtab, ".rela.init_array");
+		else
+			addstring(shstrtab, ".rel.init_array");
+	}
+
 	if(!debug['s']) {
 		addstring(shstrtab, ".symtab");
 		addstring(shstrtab, ".strtab");
@@ -1064,13 +1070,6 @@ doelf(void)
 		
 		elfwritedynent(s, DT_DEBUG, 0);
 
-		if(flag_shared) {
-			Sym *init_sym = lookup(LIBINITENTRY, 0);
-			if(init_sym->type != STEXT)
-				diag("entry not text: %s", init_sym->name);
-			elfwritedynentsym(s, DT_INIT, init_sym);
-		}
-
 		// Do not write DT_NULL.  elfdynhash will finish it.
 	}
 }
@@ -1469,9 +1468,7 @@ elfobj:
 	eh->ident[EI_DATA] = ELFDATA2LSB;
 	eh->ident[EI_VERSION] = EV_CURRENT;
 
-	if(flag_shared)
-		eh->type = ET_DYN;
-	else if(linkmode == LinkExternal)
+	if(linkmode == LinkExternal)
 		eh->type = ET_REL;
 	else
 		eh->type = ET_EXEC;

src/cmd/ld/elf.h

@@ -569,6 +569,8 @@ typedef struct {
 #define	R_ARM_GOT_PREL		96
 #define	R_ARM_GNU_VTENTRY	100
 #define	R_ARM_GNU_VTINHERIT	101
+#define	R_ARM_TLS_IE32		107
+#define	R_ARM_TLS_LE32		108
 #define	R_ARM_RSBREL32		250
 #define	R_ARM_THM_RPC22		251
 #define	R_ARM_RREL32		252
@@ -576,7 +578,7 @@ typedef struct {
 #define	R_ARM_RPC24		254
 #define	R_ARM_RBASE		255
 
-#define	R_ARM_COUNT		37	/* Count of defined relocation types. */
+#define	R_ARM_COUNT		38	/* Count of defined relocation types. */
 
 
 #define	R_386_NONE	0	/* No relocation. */

src/cmd/ld/go.c

@@ -499,6 +499,9 @@ loadcgo(char *file, char *pkg, char *p, int n)
 			local = expandpkg(local, pkg);
 			s = lookup(local, 0);
 
+			if(flag_shared && s == lookup("main", 0))
+				continue;
+
 			// export overrides import, for openbsd/cgo.
 			// see issue 4878.
 			if(s->dynimplib != nil) {
@@ -680,8 +683,6 @@ deadcode(void)
 		Bprint(&bso, "%5.2f deadcode\n", cputime());
 
 	mark(lookup(INITENTRY, 0));
-	if(flag_shared)
-		mark(lookup(LIBINITENTRY, 0));
 	for(i=0; i<nelem(markextra); i++)
 		mark(lookup(markextra[i], 0));
 

src/cmd/ld/lib.c

@@ -119,17 +119,14 @@ libinit(void)
 	}
 
 	if(INITENTRY == nil) {
-		INITENTRY = mal(strlen(goarch)+strlen(goos)+10);
-		sprint(INITENTRY, "_rt0_%s_%s", goarch, goos);
-	}
-	lookup(INITENTRY, 0)->type = SXREF;
-	if(flag_shared) {
-		if(LIBINITENTRY == nil) {
-			LIBINITENTRY = mal(strlen(goarch)+strlen(goos)+20);
-			sprint(LIBINITENTRY, "_rt0_%s_%s_lib", goarch, goos);
+		INITENTRY = mal(strlen(goarch)+strlen(goos)+20);
+		if(!flag_shared) {
+			sprint(INITENTRY, "_rt0_%s_%s", goarch, goos);
+		} else {
+			sprint(INITENTRY, "_rt0_%s_%s_lib", goarch, goos);
 		}
-		lookup(LIBINITENTRY, 0)->type = SXREF;
 	}
+	lookup(INITENTRY, 0)->type = SXREF;
 }
 
 void
@@ -308,7 +305,13 @@ void
 loadlib(void)
 {
 	int i, w, x;
-	Sym *s;
+	Sym *s, *gmsym;
+
+	if(flag_shared) {
+		s = lookup("runtime.islibrary", 0);
+		s->dupok = 1;
+		adduint8(s, 1);
+	}
 
 	loadinternal("runtime");
 	if(thechar == '5')
@@ -357,7 +360,15 @@ loadlib(void)
 				} else
 					s->type = 0;
 			}
-	}
+	} 
+	gmsym = lookup("runtime.tlsgm", 0);
+	gmsym->type = STLSBSS;
+	gmsym->size = 2*PtrSize;
+	gmsym->hide = 1;
+	if(linkmode == LinkExternal)
+		gmsym->reachable = 1;
+	else
+		gmsym->reachable = 0;
 	
 	// Now that we know the link mode, trim the dynexp list.
 	x = CgoExportDynamic;
@@ -669,7 +680,7 @@ hostlink(void)
 		p = strchr(p + 1, ' ');
 	}
 
-	argv = malloc((10+nhostobj+nldflag+c)*sizeof argv[0]);
+	argv = malloc((13+nhostobj+nldflag+c)*sizeof argv[0]);
 	argc = 0;
 	if(extld == nil)
 		extld = "gcc";
@@ -682,7 +693,7 @@ hostlink(void)
 		argv[argc++] = "-m64";
 		break;
 	case '5':
-		// nothing required for arm
+		argv[argc++] = "-marm";
 		break;
 	}
 	if(!debug['s'] && !debug_s) {
@@ -696,6 +707,10 @@ hostlink(void)
 	if(iself && AssumeGoldLinker)
 		argv[argc++] = "-Wl,--rosegment";
 
+	if(flag_shared) {
+		argv[argc++] = "-Wl,-Bsymbolic";
+		argv[argc++] = "-shared";
+	}
 	argv[argc++] = "-o";
 	argv[argc++] = outfile;
 	

src/cmd/ld/lib.h

@@ -55,7 +55,7 @@ enum
 	SMACHO,	/* Mach-O __nl_symbol_ptr */
 	SMACHOGOT,
 	SNOPTRDATA,
-	SDATARELRO,
+	SINITARR,
 	SDATA,
 	SWINDOWS,
 	SBSS,
@@ -224,7 +224,6 @@ void	Lflag(char *arg);
 void	usage(void);
 void	adddynrel(Sym*, Reloc*);
 void	adddynrela(Sym*, Sym*, Reloc*);
-Sym*	lookuprel(void);
 void	ldobj1(Biobuf *f, char*, int64 len, char *pn);
 void	ldobj(Biobuf*, char*, int64, char*, char*, int);
 void	ldelf(Biobuf*, char*, int64, char*);

src/cmd/ld/symtab.c

@@ -181,22 +181,13 @@ asmelfsym(void)
 	genasmsym(putelfsym);
 	
 	if(linkmode == LinkExternal && HEADTYPE != Hopenbsd) {
-		s = lookup("runtime.m", 0);
+		s = lookup("runtime.tlsgm", 0);
 		if(s->sect == nil) {
 			cursym = nil;
 			diag("missing section for %s", s->name);
 			errorexit();
 		}
-		putelfsyment(putelfstr(s->name), 0, PtrSize, (STB_LOCAL<<4)|STT_TLS, s->sect->elfsect->shnum, 0);
-		s->elfsym = numelfsym++;
-
-		s = lookup("runtime.g", 0);
-		if(s->sect == nil) {
-			cursym = nil;
-			diag("missing section for %s", s->name);
-			errorexit();
-		}
-		putelfsyment(putelfstr(s->name), PtrSize, PtrSize, (STB_LOCAL<<4)|STT_TLS, s->sect->elfsect->shnum, 0);
+		putelfsyment(putelfstr(s->name), 0, 2*PtrSize, (STB_LOCAL<<4)|STT_TLS, s->sect->elfsect->shnum, 0);
 		s->elfsym = numelfsym++;
 	}
 
@@ -478,10 +469,6 @@ symtab(void)
 	xdefine("etypelink", SRODATA, 0);
 	xdefine("rodata", SRODATA, 0);
 	xdefine("erodata", SRODATA, 0);
-	if(flag_shared) {
-		xdefine("datarelro", SDATARELRO, 0);
-		xdefine("edatarelro", SDATARELRO, 0);
-	}
 	xdefine("noptrdata", SNOPTRDATA, 0);
 	xdefine("enoptrdata", SNOPTRDATA, 0);
 	xdefine("data", SDATA, 0);

src/pkg/runtime/asm_arm.s

@@ -13,7 +13,7 @@ TEXT _rt0_go(SB),NOSPLIT,$-4
 	// copy arguments forward on an even stack
 	// use R13 instead of SP to avoid linker rewriting the offsets
 	MOVW	0(R13), R0		// argc
-	MOVW	$4(R13), R1		// argv
+	MOVW	4(R13), R1		// argv
 	SUB	$64, R13		// plenty of scratch
 	AND	$~7, R13
 	MOVW	R0, 60(R13)		// save argc, argv away
@@ -35,10 +35,15 @@ TEXT _rt0_go(SB),NOSPLIT,$-4
 	BL	runtime·emptyfunc(SB)	// fault if stack check is wrong
 
 	// if there is an _cgo_init, call it.
-	MOVW	_cgo_init(SB), R2
-	CMP	$0, R2
-	MOVW.NE	g, R0 // first argument of _cgo_init is g
-	BL.NE	(R2) // will clobber R0-R3
+	MOVW	_cgo_init(SB), R4
+	CMP	$0, R4
+	B.EQ	nocgo
+	BL		runtime·save_gm(SB);
+	MOVW	g, R0 // first argument of _cgo_init is g
+	MOVW	$setmg_gcc<>(SB), R1 // second argument is address of save_gm
+	BL		(R4) // will clobber R0-R3
+
+nocgo:
 	// update stackguard after _cgo_init
 	MOVW	g_stackguard0(g), R0
 	MOVW	R0, g_stackguard(g)
@@ -119,9 +124,9 @@ TEXT runtime·gogo(SB), NOSPLIT, $-4-4
 	MOVW	0(FP), R1		// gobuf
 	MOVW	gobuf_g(R1), g
 	MOVW	0(g), R2		// make sure g != nil
-	MOVW	_cgo_save_gm(SB), R2
+	MOVB	runtime·iscgo(SB), R2
 	CMP 	$0, R2 // if in Cgo, we have to save g and m
-	BL.NE	(R2) // this call will clobber R0
+	BL.NE	runtime·save_gm(SB) // this call will clobber R0
 	MOVW	gobuf_sp(R1), SP	// restore SP
 	MOVW	gobuf_lr(R1), LR
 	MOVW	gobuf_ret(R1), R0
@@ -437,9 +442,9 @@ TEXT runtime·cgocallback(SB),NOSPLIT,$12-12
 // See cgocall.c for more details.
 TEXT	runtime·cgocallback_gofunc(SB),NOSPLIT,$8-12
 	// Load m and g from thread-local storage.
-	MOVW	_cgo_load_gm(SB), R0
+	MOVB	runtime·iscgo(SB), R0
 	CMP	$0, R0
-	BL.NE	(R0)
+	BL.NE	runtime·load_gm(SB)
 
 	// If m is nil, Go did not create the current thread.
 	// Call needm to obtain one for temporary use.
@@ -519,9 +524,9 @@ TEXT runtime·setmg(SB), NOSPLIT, $0-8
 	MOVW	gg+4(FP), g
 
 	// Save m and g to thread-local storage.
-	MOVW	_cgo_save_gm(SB), R0
+	MOVB	runtime·iscgo(SB), R0
 	CMP	$0, R0
-	BL.NE	(R0)
+	BL.NE	runtime·save_gm(SB)
 
 	RET
 
@@ -615,6 +620,34 @@ _next:
 	MOVW	$0, R0
 	RET
 
+// We have to resort to TLS variable to save g(R10) and
+// m(R9). One reason is that external code might trigger
+// SIGSEGV, and our runtime.sigtramp don't even know we
+// are in external code, and will continue to use R10/R9,
+// this might as well result in another SIGSEGV.
+// Note: all three functions will clobber R0, and the last
+// two can be called from 5c ABI code.
+
+// g (R10) at 8(TP), m (R9) at 12(TP)
+TEXT runtime·save_gm(SB),NOSPLIT,$0
+	MRC		15, 0, R0, C13, C0, 3 // Fetch TLS register
+	MOVW	g, 8(R0)
+	MOVW	m, 12(R0)
+	RET
+
+TEXT runtime·load_gm(SB),NOSPLIT,$0
+	MRC		15, 0, R0, C13, C0, 3 // Fetch TLS register
+	MOVW	8(R0), g
+	MOVW	12(R0), m
+	RET
+
+// void setmg_gcc(M*, G*); set m and g called from gcc.
+TEXT setmg_gcc<>(SB),NOSPLIT,$0
+	MOVW	R0, m
+	MOVW	R1, g
+	B		runtime·save_gm(SB)
+
+
 // TODO: share code with memeq?
 TEXT bytes·Equal(SB),NOSPLIT,$0
 	MOVW	a_len+4(FP), R1

src/pkg/runtime/cgo/asm_arm.s

@@ -14,12 +14,11 @@ TEXT crosscall2(SB),NOSPLIT,$-4
 	 *  push 2 args for fn (R1 and R2).
 	 * Also note that at procedure entry in 5c/5g world, 4(R13) will be the
 	 *  first arg, so we must push another dummy reg (R0) for 0(R13).
-	 *  Additionally, cgo_tls_set_gm will clobber R0, so we need to save R0
+	 *  Additionally, runtime·load_gm will clobber R0, so we need to save R0
 	 *  nevertheless.
 	 */
 	MOVM.WP	[R0, R1, R2, R4, R5, R6, R7, R8, m, g, R11, R12, R14], (R13)
-	MOVW	_cgo_load_gm(SB), R0
-	BL	(R0)
+	BL	runtime·load_gm(SB)
 	MOVW	PC, R14
 	MOVW	0(R13), PC
 	MOVM.IAW	(R13), [R0, R1, R2, R4, R5, R6, R7, R8, m, g, R11, R12, PC]

src/pkg/runtime/cgo/cgo_arm.c

-// Copyright 2013 The Go Authors.  All rights reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-#pragma cgo_import_static x_cgo_load_gm
-extern void x_cgo_load_gm(void);
-void (*_cgo_load_gm)(void) = x_cgo_load_gm;
-
-#pragma cgo_import_static x_cgo_save_gm
-extern void x_cgo_save_gm(void);
-void (*_cgo_save_gm)(void) = x_cgo_save_gm;
-

src/pkg/runtime/cgo/gcc_arm.S

@@ -12,7 +12,7 @@
 #endif
 
 /*
- * void crosscall_arm2(void (*fn)(void), void *g, void *m)
+ * void crosscall_arm2(void (*fn)(void), void (*setmg_gcc)(void *m, void *g), void *m, void *g)
  *
  * Calling into the 5c tool chain, where all registers are caller save.
  * Called from standard ARM EABI, where r4-r11 are callee-save, so they
@@ -21,12 +21,12 @@
 .globl EXT(crosscall_arm2)
 EXT(crosscall_arm2):
 	push {r4, r5, r6, r7, r8, r9, r10, r11, ip, lr}
-	mov r10, r1 // g
-	mov r9, r2 // m
-	mov r3, r0 // save r0, cgo_tls_set_gm will clobber it
-	bl EXT(x_cgo_save_gm) // save current g and m into TLS variable
-	mov lr, pc
-	mov pc, r3
+	mov r4, r0
+	mov r5, r1
+	mov r0, r2
+	mov r1, r3
+	blx r5 // setmg(m, g)
+	blx r4 // fn()
 	pop {r4, r5, r6, r7, r8, r9, r10, r11, ip, pc}
 
 .globl EXT(__stack_chk_fail_local)

src/pkg/runtime/cgo/gcc_freebsd_arm.c

@@ -8,72 +8,26 @@
 #include <string.h>
 #include "libcgo.h"
 
-static void *threadentry(void*);
-
-// We have to resort to TLS variable to save g(R10) and
-// m(R9). One reason is that external code might trigger
-// SIGSEGV, and our runtime.sigtramp don't even know we
-// are in external code, and will continue to use R10/R9,
-// this might as well result in another SIGSEGV.
-// Note: all three functions will clobber R0, and the last
-// two can be called from 5c ABI code.
-void __aeabi_read_tp(void) __attribute__((naked));
-void x_cgo_save_gm(void) __attribute__((naked));
-void x_cgo_load_gm(void) __attribute__((naked));
-
-void
-__aeabi_read_tp(void)
-{
-	__asm__ __volatile__ (
 #ifdef ARM_TP_ADDRESS
-		// ARM_TP_ADDRESS is (ARM_VECTORS_HIGH + 0x1000) or 0xffff1000
-		// GCC inline asm doesn't provide a way to provide a constant
-		// to "ldr r0, =??" pseudo instruction, so we hardcode the value
-		// and check it with cpp.
+// ARM_TP_ADDRESS is (ARM_VECTORS_HIGH + 0x1000) or 0xffff1000
+// and is known to runtime.read_tls_fallback. Verify it with
+// cpp.
 #if ARM_TP_ADDRESS != 0xffff1000
 #error Wrong ARM_TP_ADDRESS!
 #endif
-		"ldr r0, =0xffff1000\n\t"
-		"ldr r0, [r0]\n\t"
-#else
-		"mrc p15, 0, r0, c13, c0, 3\n\t"
 #endif
-		"mov pc, lr\n\t"
-	);
-}
 
-// g (R10) at 8(TP), m (R9) at 12(TP)
-void
-x_cgo_load_gm(void)
-{
-	__asm__ __volatile__ (
-		"push {lr}\n\t"
-		"bl __aeabi_read_tp\n\t"
-		"ldr r10, [r0, #8]\n\t"
-		"ldr r9, [r0, #12]\n\t"
-		"pop {pc}\n\t"
-	);
-}
+static void *threadentry(void*);
 
-void
-x_cgo_save_gm(void)
-{
-	__asm__ __volatile__ (
-		"push {lr}\n\t"
-		"bl __aeabi_read_tp\n\t"
-		"str r10, [r0, #8]\n\t"
-		"str r9, [r0, #12]\n\t"
-		"pop {pc}\n\t"
-	);
-}
+static void (*setmg_gcc)(void*, void*);
 
 void
-x_cgo_init(G *g)
+x_cgo_init(G *g, void (*setmg)(void*, void*))
 {
 	pthread_attr_t attr;
 	size_t size;
-	x_cgo_save_gm(); // save g and m for the initial thread
 
+	setmg_gcc = setmg;
 	pthread_attr_init(&attr);
 	pthread_attr_getstacksize(&attr, &size);
 	g->stackguard = (uintptr)&attr - size + 4096;
@@ -104,7 +58,7 @@ _cgo_sys_thread_start(ThreadStart *ts)
 	}
 }
 
-extern void crosscall_arm2(void (*fn)(void), void *g, void *m);
+extern void crosscall_arm2(void (*fn)(void), void (*setmg_gcc)(void*, void*), void *g, void *m);
 static void*
 threadentry(void *v)
 {
@@ -121,6 +75,6 @@ threadentry(void *v)
 	 */
 	ts.g->stackguard = (uintptr)&ts - ts.g->stackguard + 4096 * 2;
 
-	crosscall_arm2(ts.fn, (void *)ts.g, (void *)ts.m);
+	crosscall_arm2(ts.fn, setmg_gcc, (void*)ts.m, (void*)ts.g);
 	return nil;
 }

src/pkg/runtime/cgo/gcc_linux_arm.c

@@ -8,60 +8,15 @@
 
 static void *threadentry(void*);
 
-// We have to resort to TLS variable to save g(R10) and
-// m(R9). One reason is that external code might trigger
-// SIGSEGV, and our runtime.sigtramp don't even know we
-// are in external code, and will continue to use R10/R9,
-// this might as well result in another SIGSEGV.
-// Note: all three functions will clobber R0, and the last
-// two can be called from 5c ABI code.
-void __aeabi_read_tp(void) __attribute__((naked));
-void x_cgo_save_gm(void) __attribute__((naked));
-void x_cgo_load_gm(void) __attribute__((naked));
+static void (*setmg_gcc)(void*, void*);
 
 void
-__aeabi_read_tp(void)
-{
-	// b __kuser_get_tls @ 0xffff0fe0
-	__asm__ __volatile__ (
-		"mvn r0, #0xf000\n\t"
-		"sub pc, r0, #31\n\t"
-		"nop\n\tnop\n\t"
-	);
-}
-
-// g (R10) at 8(TP), m (R9) at 12(TP)
-void
-x_cgo_load_gm(void)
-{
-	__asm__ __volatile__ (
-		"push {lr}\n\t"
-		"bl __aeabi_read_tp\n\t"
-		"ldr r10, [r0, #8]\n\t"
-		"ldr r9, [r0, #12]\n\t"
-		"pop {pc}\n\t"
-	);
-}
-
-void
-x_cgo_save_gm(void)
-{
-	__asm__ __volatile__ (
-		"push {lr}\n\t"
-		"bl __aeabi_read_tp\n\t"
-		"str r10, [r0, #8]\n\t"
-		"str r9, [r0, #12]\n\t"
-		"pop {pc}\n\t"
-	);
-}
-
-void
-x_cgo_init(G *g)
+x_cgo_init(G *g, void (*setmg)(void*, void*))
 {
 	pthread_attr_t attr;
 	size_t size;
-	x_cgo_save_gm(); // save g and m for the initial thread
 
+	setmg_gcc = setmg;
 	pthread_attr_init(&attr);
 	pthread_attr_getstacksize(&attr, &size);
 	g->stackguard = (uintptr)&attr - size + 4096;
@@ -92,7 +47,7 @@ _cgo_sys_thread_start(ThreadStart *ts)
 	}
 }
 
-extern void crosscall_arm2(void (*fn)(void), void *g, void *m);
+extern void crosscall_arm2(void (*fn)(void), void (*setmg_gcc)(void*, void*), void*, void*);
 static void*
 threadentry(void *v)
 {
@@ -109,6 +64,6 @@ threadentry(void *v)
 	 */
 	ts.g->stackguard = (uintptr)&ts - ts.g->stackguard + 4096 * 2;
 
-	crosscall_arm2(ts.fn, (void *)ts.g, (void *)ts.m);
+	crosscall_arm2(ts.fn, setmg_gcc, (void*)ts.m, (void*)ts.g);
 	return nil;
 }

src/pkg/runtime/cgo/gcc_netbsd_arm.c

@@ -10,64 +10,15 @@
 
 static void *threadentry(void*);
 
-// We have to resort to TLS variable to save g(R10) and
-// m(R9). One reason is that external code might trigger
-// SIGSEGV, and our runtime.sigtramp don't even know we
-// are in external code, and will continue to use R10/R9,
-// this might as well result in another SIGSEGV.
-// Note: all three functions will clobber R0, and the last
-// two can be called from 5c ABI code.
-void __aeabi_read_tp(void) __attribute__((naked));
-void x_cgo_save_gm(void) __attribute__((naked));
-void x_cgo_load_gm(void) __attribute__((naked));
+static void (*setmg_gcc)(void*, void*);
 
 void
-__aeabi_read_tp(void)
-{
-	// this function is only allowed to clobber r0
-	__asm__ __volatile__ (
-		"mrc p15, 0, r0, c13, c0, 3\n\t"
-		"cmp r0, #0\n\t"
-		"movne pc, lr\n\t"
-		"push {r1,r2,r3,r12}\n\t"
-		"svc 0x00a0013c\n\t" // _lwp_getprivate
-		"pop {r1,r2,r3,r12}\n\t"
-		"mov pc, lr\n\t"
-	);
-}
-
-// g (R10) at 8(TP), m (R9) at 12(TP)
-void
-x_cgo_load_gm(void)
-{
-	__asm__ __volatile__ (
-		"push {lr}\n\t"
-		"bl __aeabi_read_tp\n\t"
-		"ldr r10, [r0, #8]\n\t"
-		"ldr r9, [r0, #12]\n\t"
-		"pop {pc}\n\t"
-	);
-}
-
-void
-x_cgo_save_gm(void)
-{
-	__asm__ __volatile__ (
-		"push {lr}\n\t"
-		"bl __aeabi_read_tp\n\t"
-		"str r10, [r0, #8]\n\t"
-		"str r9, [r0, #12]\n\t"
-		"pop {pc}\n\t"
-	);
-}
-
-void
-x_cgo_init(G *g)
+x_cgo_init(G *g, void (*setmg)(void*, void*))
 {
 	pthread_attr_t attr;
 	size_t size;
-	x_cgo_save_gm(); // save g and m for the initial thread
 
+	setmg_gcc = setmg;
 	pthread_attr_init(&attr);
 	pthread_attr_getstacksize(&attr, &size);
 	g->stackguard = (uintptr)&attr - size + 4096;
@@ -100,7 +51,7 @@ _cgo_sys_thread_start(ThreadStart *ts)
 	}
 }
 
-extern void crosscall_arm2(void (*fn)(void), void *g, void *m);
+extern void crosscall_arm2(void (*fn)(void), void (*setmg_gcc)(void*, void*), void *g, void *m);
 static void*
 threadentry(void *v)
 {
@@ -117,6 +68,6 @@ threadentry(void *v)
 	 */
 	ts.g->stackguard = (uintptr)&ts - ts.g->stackguard + 4096 * 2;
 
-	crosscall_arm2(ts.fn, (void *)ts.g, (void *)ts.m);
+	crosscall_arm2(ts.fn, setmg_gcc, (void*)ts.m, (void*)ts.g);
 	return nil;
 }

src/pkg/runtime/cgocall.c

@@ -87,11 +87,6 @@
 void *_cgo_init;	/* filled in by dynamic linker when Cgo is available */
 static int64 cgosync;  /* represents possible synchronization in C code */
 
-// These two are only used by the architecture where TLS based storage isn't
-// the default for g and m (e.g., ARM)
-void *_cgo_load_gm; /* filled in by dynamic linker when Cgo is available */
-void *_cgo_save_gm; /* filled in by dynamic linker when Cgo is available */
-
 static void unwindm(void);
 
 // Call from Go to C.

src/pkg/runtime/os_linux_arm.c

@@ -35,16 +35,13 @@ runtime·checkgoarm(void)
 
 #pragma textflag NOSPLIT
 void
-runtime·setup_auxv(int32 argc, void *argv_list)
+runtime·setup_auxv(int32 argc, byte **argv)
 {
-	byte **argv;
 	byte **envp;
 	byte *rnd;
 	uint32 *auxv;
 	uint32 t;
 
-	argv = &argv_list;
-
 	// skip envp to get to ELF auxiliary vector.
 	for(envp = &argv[argc+1]; *envp != nil; envp++)
 		;

src/pkg/runtime/rt0_freebsd_arm.s

@@ -7,4 +7,7 @@
 // FreeBSD and Linux use the same linkage to main
 
 TEXT _rt0_arm_freebsd(SB),NOSPLIT,$-4
+	MOVW	(R13), R0	// argc
+	MOVW	$4(R13), R1		// argv
+	MOVM.DB.W [R0-R1], (R13)
 	B	_rt0_go(SB)

src/pkg/runtime/rt0_linux_arm.s

@@ -5,6 +5,12 @@
 #include "../../cmd/ld/textflag.h"
 
 TEXT _rt0_arm_linux(SB),NOSPLIT,$-4
+	MOVW	(R13), R0	// argc
+	MOVW	$4(R13), R1		// argv
+	MOVW	$_rt0_arm_linux1(SB), R4
+	B		(R4)
+
+TEXT _rt0_arm_linux1(SB),NOSPLIT,$-4
 	// We first need to detect the kernel ABI, and warn the user
 	// if the system only supports OABI
 	// The strategy here is to call some EABI syscall to see if
@@ -14,6 +20,8 @@ TEXT _rt0_arm_linux(SB),NOSPLIT,$-4
 	// we don't know the kernel ABI... Oh, not really, we can do
 	// syscall in Thumb mode.
 
+	// Save argc and argv
+	MOVM.DB.W [R0-R1], (R13)
 	// set up sa_handler
 	MOVW	$bad_abi<>(SB), R0 // sa_handler
 	MOVW	$0, R1 // sa_flags
@@ -72,3 +80,7 @@ TEXT oabi_syscall<>(SB),NOSPLIT,$-4
 	// TODO(minux): only supports little-endian CPUs
 	WORD $0x4770df01 // swi $1; bx lr
 
+TEXT main(SB),NOSPLIT,$-4
+	MOVW	$_rt0_arm_linux1(SB), R4
+	B		(R4)
+

src/pkg/runtime/rt0_netbsd_arm.s

@@ -7,4 +7,7 @@
 // FreeBSD/NetBSD and Linux use the same linkage to main
 
 TEXT _rt0_arm_netbsd(SB),NOSPLIT,$-4
+	MOVW	(R13), R0	// argc
+	MOVW	$4(R13), R1		// argv
+	MOVM.DB.W [R0-R1], (R13)
 	B _rt0_go(SB)

src/pkg/runtime/sys_freebsd_arm.s

@@ -282,3 +282,7 @@ TEXT runtime·casp(SB),NOSPLIT,$0
 //		return 0;
 TEXT runtime·cas(SB),NOSPLIT,$0
 	B runtime·armcas(SB)
+
+TEXT runtime·read_tls_fallback(SB),NOSPLIT,$-4
+	MOVW $0xffff1000, R0
+	MOVW (R0), R0

src/pkg/runtime/sys_linux_arm.s

@@ -286,11 +286,11 @@ TEXT runtime·sigaltstack(SB),NOSPLIT,$0
 TEXT runtime·sigtramp(SB),NOSPLIT,$24
 	// this might be called in external code context,
 	// where g and m are not set.
-	// first save R0, because _cgo_load_gm will clobber it
+	// first save R0, because runtime·load_gm will clobber it
 	MOVW	R0, 4(R13)
-	MOVW	_cgo_load_gm(SB), R0
+	MOVB	runtime·iscgo(SB), R0
 	CMP 	$0, R0
-	BL.NE	(R0)
+	BL.NE	runtime·load_gm(SB)
 
 	CMP 	$0, m
 	BNE 	4(PC)
@@ -441,3 +441,8 @@ TEXT runtime·closeonexec(SB),NOSPLIT,$0
 	MOVW	$SYS_fcntl, R7
 	SWI	$0
 	RET
+
+// b __kuser_get_tls @ 0xffff0fe0
+TEXT runtime·read_tls_fallback(SB),NOSPLIT,$-4
+	MOVW	$0xffff0fe0, R0
+	B	(R0)

src/pkg/runtime/sys_netbsd_arm.s

@@ -301,3 +301,9 @@ TEXT runtime·casp(SB),NOSPLIT,$0
 //		return 0;
 TEXT runtime·cas(SB),NOSPLIT,$0
 	B runtime·armcas(SB)
+
+TEXT runtime·read_tls_fallback(SB),NOSPLIT,$-4
+	MOVM.WP [R1, R2, R3, R12], (R13)
+	SWI $0x00a0013c // _lwp_getprivate
+	MOVM.IAW    (R13), [R1, R2, R3, R12]
+	RET

src/run.bash

@@ -118,7 +118,7 @@ darwin-386 | darwin-amd64)
 	*) go test -ldflags '-linkmode=external'  || exit 1;;
 	esac
 	;;
-freebsd-386 | freebsd-amd64 | linux-386 | linux-amd64 | netbsd-386 | netbsd-amd64)
+freebsd-386 | freebsd-amd64 | linux-386 | linux-amd64 | linux-arm | netbsd-386 | netbsd-amd64)
 	go test -ldflags '-linkmode=external' || exit 1
 	go test -ldflags '-linkmode=auto' ../testtls || exit 1
 	go test -ldflags '-linkmode=external' ../testtls || exit 1