Imported Upstream version 1upstream/1

9 files changed, 182 insertions, 142 deletions

diff --git a/src/cmd/8l/8.out.h b/src/cmd/8l/8.out.h
index 9a8483aaf..9d2751cf0 100644
--- a/src/cmd/8l/8.out.h
+++ b/src/cmd/8l/8.out.h
@@ -34,6 +34,7 @@
 #define DUPOK	(1<<1)
 #define NOSPLIT	(1<<2)
 #define RODATA	(1<<3)
+#define NOPTR	(1<<4)
 
 enum	as
 {
@@ -115,7 +116,8 @@ enum	as
 	AIRETW,
 	AJCC,
 	AJCS,
-	AJCXZ,
+	AJCXZL,
+	AJCXZW,
 	AJEQ,
 	AJGE,
 	AJGT,
@@ -394,7 +396,9 @@ enum	as
 	ACMPXCHGL,
 	ACMPXCHGW,
 	ACMPXCHG8B,
-	
+
+	ARDTSC,
+
 	AXADDB,
 	AXADDL,
 	AXADDW,
@@ -442,6 +446,12 @@ enum	as
 	AFCMOVNU,
 	AFCMOVUN,
 
+	ALFENCE,
+	AMFENCE,
+	ASFENCE,
+
+	AEMMS,
+
 	ALAST
 };
 
diff --git a/src/cmd/8l/Makefile b/src/cmd/8l/Makefile
index 7d34e1704..3f528d751 100644
--- a/src/cmd/8l/Makefile
+++ b/src/cmd/8l/Makefile
@@ -1,49 +1,5 @@
-# Copyright 2009 The Go Authors. All rights reserved.
+# Copyright 2012 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.
 
-include ../../Make.inc
-O:=$(HOST_O)
-
-TARG=8l
-
-OFILES=\
-	asm.$O\
-	data.$O\
-	dwarf.$O\
-	elf.$O\
-	enam.$O\
-	go.$O\
-	ldelf.$O\
-	ldmacho.$O\
-	ldpe.$O\
-	lib.$O\
-	list.$O\
-	macho.$O\
-	obj.$O\
-	optab.$O\
-	pass.$O\
-	pe.$O\
-	prof.$O\
-	span.$O\
-	symtab.$O\
-
-
-HFILES=\
-	l.h\
-	8.out.h\
-	../ld/dwarf.h\
-	../ld/elf.h\
-	../ld/macho.h\
-	../ld/pe.h\
-
-include ../../Make.ccmd
-
-enam.c: 8.out.h
-	sh mkenam
-
-CLEANFILES+=enam.c
-
-
-%.$O: ../ld/%.c
-	$(HOST_CC) $(HOST_CFLAGS) -c -I. ../ld/$*.c
+include ../../Make.dist
diff --git a/src/cmd/8l/asm.c b/src/cmd/8l/asm.c
index 22abd8049..25ffc786f 100644
--- a/src/cmd/8l/asm.c
+++ b/src/cmd/8l/asm.c
@@ -37,10 +37,10 @@
 #include	"../ld/macho.h"
 #include	"../ld/pe.h"
 
-#define	Dbufslop	100
-
 char linuxdynld[] = "/lib/ld-linux.so.2";
 char freebsddynld[] = "/usr/libexec/ld-elf.so.1";
+char openbsddynld[] = "/usr/libexec/ld.so";
+char netbsddynld[] = "/usr/libexec/ld.elf_so";
 
 int32
 entryvalue(void)
@@ -90,6 +90,9 @@ enum {
 	ElfStrPlt,
 	ElfStrGnuVersion,
 	ElfStrGnuVersionR,
+	ElfStrNoteNetbsdIdent,
+	ElfStrNoPtrData,
+	ElfStrNoPtrBss,
 	NElfStr
 };
 
@@ -107,6 +110,7 @@ needlib(char *name)
 	/* reuse hash code in symbol table */
 	p = smprint(".dynlib.%s", name);
 	s = lookup(p, 0);
+	free(p);
 	if(s->type == 0) {
 		s->type = 100;	// avoid SDATA, etc.
 		return 1;
@@ -416,7 +420,7 @@ adddynsym(Sym *s)
 		return;
 	
 	if(s->dynimpname == nil)
-		diag("adddynsym: no dynamic name for %s", s->name, *(int32*)0);
+		diag("adddynsym: no dynamic name for %s", s->name);
 
 	if(iself) {
 		s->dynid = nelfsym++;
@@ -524,8 +528,12 @@ doelf(void)
 
 	elfstr[ElfStrEmpty] = addstring(shstrtab, "");
 	elfstr[ElfStrText] = addstring(shstrtab, ".text");
+	elfstr[ElfStrNoPtrData] = addstring(shstrtab, ".noptrdata");
 	elfstr[ElfStrData] = addstring(shstrtab, ".data");
 	elfstr[ElfStrBss] = addstring(shstrtab, ".bss");
+	elfstr[ElfStrNoPtrBss] = addstring(shstrtab, ".noptrbss");
+	if(HEADTYPE == Hnetbsd)
+		elfstr[ElfStrNoteNetbsdIdent] = addstring(shstrtab, ".note.netbsd.ident");
 	addstring(shstrtab, ".elfdata");
 	addstring(shstrtab, ".rodata");
 	addstring(shstrtab, ".gosymtab");
@@ -606,7 +614,7 @@ doelf(void)
 		/* define dynamic elf table */
 		s = lookup(".dynamic", 0);
 		s->reachable = 1;
-		s->type = SELFROSECT;
+		s->type = SELFSECT; // writable
 
 		/*
 		 * .dynamic table
@@ -626,6 +634,8 @@ doelf(void)
 		elfwritedynentsymsize(s, DT_PLTRELSZ, lookup(".rel.plt", 0));
 		elfwritedynentsym(s, DT_JMPREL, lookup(".rel.plt", 0));
 
+		elfwritedynent(s, DT_DEBUG, 0);
+
 		// Do not write DT_NULL.  elfdynhash will finish it.
 	}
 }
@@ -657,7 +667,7 @@ asmb(void)
 {
 	int32 v, magic;
 	int a, dynsym;
-	uint32 symo, startva, machlink;
+	uint32 symo, startva, dwarfoff, machlink, resoff;
 	ElfEhdr *eh;
 	ElfPhdr *ph, *pph;
 	ElfShdr *sh;
@@ -688,8 +698,19 @@ asmb(void)
 	datblk(segdata.vaddr, segdata.filelen);
 
 	machlink = 0;
-	if(HEADTYPE == Hdarwin)
+	if(HEADTYPE == Hdarwin) {
+		if(debug['v'])
+			Bprint(&bso, "%5.2f dwarf\n", cputime());
+
+		dwarfoff = rnd(HEADR+segtext.len, INITRND) + rnd(segdata.filelen, INITRND);
+		cseek(dwarfoff);
+
+		segdwarf.fileoff = cpos();
+		dwarfemitdebugsections();
+		segdwarf.filelen = cpos() - segdwarf.fileoff;
+
 		machlink = domacholink();
+	}
 
 	if(iself) {
 		/* index of elf text section; needed by asmelfsym, double-checked below */
@@ -700,6 +721,8 @@ asmb(void)
 			if(elfverneed)
 				elftextsh += 2;
 		}
+		if(HEADTYPE == Hnetbsd)
+			elftextsh += 1;
 	}
 
 	symsize = 0;
@@ -746,7 +769,7 @@ asmb(void)
 		default:
 			if(iself) {
 				if(debug['v'])
-				       Bprint(&bso, "%5.2f elfsym\n", cputime());
+					Bprint(&bso, "%5.2f elfsym\n", cputime());
 				asmelfsym();
 				cflush();
 				cwrite(elfstrdat, elfstrsize);
@@ -769,7 +792,6 @@ asmb(void)
 				cflush();
 			}
 			break;
-		case Hdarwin:
 		case Hwindows:
 			if(debug['v'])
 				Bprint(&bso, "%5.2f dwarf\n", cputime());
@@ -918,6 +940,7 @@ asmb(void)
 	Elfput:
 		eh = getElfEhdr();
 		startva = INITTEXT - HEADR;
+		resoff = ELFRESERVE;
 
 		/* This null SHdr must appear before all others */
 		newElfShdr(elfstr[ElfStrEmpty]);
@@ -956,9 +979,15 @@ asmb(void)
 				case Hfreebsd:
 					interpreter = freebsddynld;
 					break;
+				case Hnetbsd:
+					interpreter = netbsddynld;
+					break;
+				case Hopenbsd:
+					interpreter = openbsddynld;
+					break;
 				}
 			}
-			elfinterp(sh, startva, interpreter);
+			resoff -= elfinterp(sh, startva, resoff, interpreter);
 
 			ph = newElfPhdr();
 			ph->type = PT_INTERP;
@@ -966,11 +995,27 @@ asmb(void)
 			phsh(ph, sh);
 		}
 
+		if(HEADTYPE == Hnetbsd) {
+			sh = newElfShdr(elfstr[ElfStrNoteNetbsdIdent]);
+			sh->type = SHT_NOTE;
+			sh->flags = SHF_ALLOC;
+			sh->addralign = 4;
+			resoff -= elfnetbsdsig(sh, startva, resoff);
+
+			ph = newElfPhdr();
+			ph->type = PT_NOTE;
+			ph->flags = PF_R;
+			phsh(ph, sh);
+		}
+
+		// Additions to the reserved area must be above this line.
+		USED(resoff);
+
 		elfphload(&segtext);
 		elfphload(&segdata);
 
 		/* Dynamic linking sections */
-		if (!debug['d']) {	/* -d suppresses dynamic loader format */
+		if(!debug['d']) {	/* -d suppresses dynamic loader format */
 			/* S headers for dynamic linking */
 			sh = newElfShdr(elfstr[ElfStrGot]);
 			sh->type = SHT_PROGBITS;
@@ -1094,7 +1139,7 @@ asmb(void)
 		for(sect=segdata.sect; sect!=nil; sect=sect->next)
 			elfshbits(sect);
 
-		if (!debug['s']) {
+		if(!debug['s']) {
 			sh = newElfShdr(elfstr[ElfStrSymtab]);
 			sh->type = SHT_SYMTAB;
 			sh->off = symo;
@@ -1122,7 +1167,13 @@ asmb(void)
 		eh->ident[EI_VERSION] = EV_CURRENT;
 		switch(HEADTYPE) {
 		case Hfreebsd:
-			eh->ident[EI_OSABI] = 9;
+			eh->ident[EI_OSABI] = ELFOSABI_FREEBSD;
+			break;
+		case Hnetbsd:
+			eh->ident[EI_OSABI] = ELFOSABI_NETBSD;
+			break;
+		case Hopenbsd:
+			eh->ident[EI_OSABI] = ELFOSABI_OPENBSD;
 			break;
 		}
 
@@ -1141,8 +1192,10 @@ asmb(void)
 		a += elfwritehdr();
 		a += elfwritephdrs();
 		a += elfwriteshdrs();
-		cflush();
-		if(a+elfwriteinterp() > ELFRESERVE)	
+		a += elfwriteinterp(elfstr[ElfStrInterp]);
+		if(HEADTYPE == Hnetbsd)
+			a += elfwritenetbsdsig(elfstr[ElfStrNoteNetbsdIdent]);
+		if(a > ELFRESERVE)	
 			diag("ELFRESERVE too small: %d > %d", a, ELFRESERVE);
 		break;
 
@@ -1205,12 +1258,16 @@ genasmsym(void (*put)(Sym*, char*, int, vlong, vlong, int, Sym*))
 			case SSTRING:
 			case SGOSTRING:
 			case SWINDOWS:
+			case SNOPTRDATA:
+			case SSYMTAB:
+			case SPCLNTAB:
 				if(!s->reachable)
 					continue;
 				put(s, s->name, 'D', symaddr(s), s->size, s->version, s->gotype);
 				continue;
 
 			case SBSS:
+			case SNOPTRBSS:
 				if(!s->reachable)
 					continue;
 				put(s, s->name, 'B', symaddr(s), s->size, s->version, s->gotype);
diff --git a/src/cmd/8l/doc.go b/src/cmd/8l/doc.go
index b70888907..12301d4f2 100644
--- a/src/cmd/8l/doc.go
+++ b/src/cmd/8l/doc.go
@@ -4,47 +4,10 @@
 
 /*
 
-8l is a modified version of the Plan 9 linker.  The original is documented at
-
-	http://plan9.bell-labs.com/magic/man2html/1/2l
-
-Its target architecture is the x86, referred to by these tools for historical reasons as 386.
-It reads files in .8 format generated by 8g, 8c, and 8a and emits
-a binary called 8.out by default.
-
-Major changes include:
-	- support for ELF and Mach-O binary files
-	- support for segmented stacks (this feature is implemented here, not in the compilers).
-
-
-Original options are listed in the link above.
-
-Options new in this version:
-
--d
-	Elide the dynamic linking header.  With this option, the binary
-	is statically linked and does not refer to dynld.  Without this option
-	(the default), the binary's contents are identical but it is loaded with dynld.
--Hplan9
-	Write Plan 9 32-bit format binaries (default when $GOOS is plan9)
--Hdarwin
-	Write Apple Mach-O binaries (default when $GOOS is darwin)
--Hlinux
-	Write Linux ELF binaries (default when $GOOS is linux)
--Hfreebsd
-	Write FreeBSD ELF binaries (default when $GOOS is freebsd)
--Hwindows
-	Write Windows PE32 binaries (default when $GOOS is windows)
--I interpreter
-	Set the ELF dynamic linker to use.
--L dir1 -L dir2
-	Search for libraries (package files) in dir1, dir2, etc.
-	The default is the single location $GOROOT/pkg/$GOOS_386.
--r dir1:dir2:...
-	Set the dynamic linker search path when using ELF.
--V
-	Print the linker version.
+8l is the linker for the 32-bit x86.
+The $GOARCH for these tools is 386.
 
+The flags are documented in ../ld/doc.go.
 
 */
 package documentation
diff --git a/src/cmd/8l/l.h b/src/cmd/8l/l.h
index 4ee0db967..b974f464b 100644
--- a/src/cmd/8l/l.h
+++ b/src/cmd/8l/l.h
@@ -134,6 +134,7 @@ struct	Sym
 	int32	dynid;
 	int32	plt;
 	int32	got;
+	int32	align;	// if non-zero, required alignment in bytes
 	Sym*	hash;	// in hash table
 	Sym*	allsym;	// in all symbol list
 	Sym*	next;	// in text or data list
@@ -209,6 +210,7 @@ enum
 	Zbr,
 	Zcall,
 	Zcallcon,
+	Zcallind,
 	Zib_,
 	Zib_rp,
 	Zibo_m,
diff --git a/src/cmd/8l/obj.c b/src/cmd/8l/obj.c
index a8e1c34a5..af4bc844f 100644
--- a/src/cmd/8l/obj.c
+++ b/src/cmd/8l/obj.c
@@ -47,17 +47,19 @@ char	*noname		= "<none>";
 char	*thestring 	= "386";
 
 Header headers[] = {
-   "garbunix", Hgarbunix,
-   "unixcoff", Hunixcoff,
-   "plan9", Hplan9x32,
-   "msdoscom", Hmsdoscom,
-   "msdosexe", Hmsdosexe,
-   "darwin", Hdarwin,
-   "linux", Hlinux,
-   "freebsd", Hfreebsd,
-   "windows", Hwindows,
-   "windowsgui", Hwindows,
-   0, 0
+	"garbunix", Hgarbunix,
+	"unixcoff", Hunixcoff,
+	"plan9", Hplan9x32,
+	"msdoscom", Hmsdoscom,
+	"msdosexe", Hmsdosexe,
+	"darwin", Hdarwin,
+	"linux", Hlinux,
+	"freebsd", Hfreebsd,
+	"netbsd", Hnetbsd,
+	"openbsd", Hopenbsd,
+	"windows", Hwindows,
+	"windowsgui", Hwindows,
+	0, 0
 };
 
 /*
@@ -69,6 +71,8 @@ Header headers[] = {
  *	-Hdarwin -Tx -Rx			is Apple Mach-O
  *	-Hlinux -Tx -Rx				is Linux ELF32
  *	-Hfreebsd -Tx -Rx			is FreeBSD ELF32
+ *	-Hnetbsd -Tx -Rx			is NetBSD ELF32
+ *	-Hopenbsd -Tx -Rx			is OpenBSD ELF32
  *	-Hwindows -Tx -Rx			is MS Windows PE32
  */
 
@@ -83,6 +87,7 @@ void
 main(int argc, char *argv[])
 {
 	int c;
+	char *name, *val;
 
 	Binit(&bso, 1, OWRITE);
 	listinit();
@@ -94,6 +99,7 @@ main(int argc, char *argv[])
 	INITDAT = -1;
 	INITRND = -1;
 	INITENTRY = 0;
+	nuxiinit();
 
 	ARGBEGIN {
 	default:
@@ -133,6 +139,11 @@ main(int argc, char *argv[])
 	case 'V':
 		print("%cl version %s\n", thechar, getgoversion());
 		errorexit();
+	case 'X':
+		name = EARGF(usage());
+		val = EARGF(usage());
+		addstrdata(name, val);
+		break;
 	} ARGEND
 
 	if(argc != 1)
@@ -209,7 +220,7 @@ main(int argc, char *argv[])
 	case Hdarwin:	/* apple MACH */
 		/*
 		 * OS X system constant - offset from %gs to our TLS.
-		 * Explained in ../../libcgo/darwin_386.c.
+		 * Explained in ../../pkg/runtime/cgo/gcc_darwin_386.c.
 		 */
 		tlsoffset = 0x468;
 		machoinit();
@@ -223,11 +234,13 @@ main(int argc, char *argv[])
 		break;
 	case Hlinux:	/* elf32 executable */
 	case Hfreebsd:
+	case Hnetbsd:
+	case Hopenbsd:
 		/*
 		 * 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.
+		 * Also known to ../../pkg/runtime/sys_linux_386.s
+		 * and ../../pkg/runtime/cgo/gcc_linux_386.c.
 		 */
 		tlsoffset = -8;
 		elfinit();
@@ -269,7 +282,6 @@ main(int argc, char *argv[])
 	zprg.to = zprg.from;
 
 	pcstr = "%.6ux ";
-	nuxiinit();
 	histgen = 0;
 	pc = 0;
 	dtype = 4;
@@ -477,7 +489,7 @@ loop:
 			sig = 1729;
 		if(sig != 0){
 			if(s->sig != 0 && s->sig != sig)
-				diag("incompatible type signatures"
+				diag("incompatible type signatures "
 					"%ux(%s) and %ux(%s) for %s",
 					s->sig, s->file, sig, pn, s->name);
 			s->sig = sig;
@@ -551,7 +563,7 @@ loop:
 			s->type = SBSS;
 			s->size = 0;
 		}
-		if(s->type != SBSS && !s->dupok) {
+		if(s->type != SBSS && s->type != SNOPTRBSS && !s->dupok) {
 			diag("%s: redefinition: %s in %s",
 				pn, s->name, TNAME);
 			s->type = SBSS;
@@ -563,6 +575,8 @@ loop:
 			s->dupok = 1;
 		if(p->from.scale & RODATA)
 			s->type = SRODATA;
+		else if(p->from.scale & NOPTR)
+			s->type = SNOPTRBSS;
 		goto loop;
 
 	case ADATA:
@@ -594,6 +608,10 @@ loop:
 	case ATEXT:
 		s = p->from.sym;
 		if(s->text != nil) {
+			if(p->from.scale & DUPOK) {
+				skip = 1;
+				goto casdef;
+			}
 			diag("%s: %s: redefinition", pn, s->name);
 			return;
 		}
diff --git a/src/cmd/8l/optab.c b/src/cmd/8l/optab.c
index f5c195d75..856482290 100644
--- a/src/cmd/8l/optab.c
+++ b/src/cmd/8l/optab.c
@@ -260,6 +260,7 @@ uchar	yloop[] =
 uchar	ycall[] =
 {
 	Ynone,	Yml,	Zo_m,	2,
+	Ynone,	Ycol,	Zcallind,	2,
 	Ynone,	Ybr,	Zcall,	0,
 	Ynone,	Yi32,	Zcallcon,	1,
 	0
@@ -383,7 +384,7 @@ Optab optab[] =
 	{ ABTSL,	yml_rl,	Pm, 0xab },
 	{ ABTSW,	yml_rl,	Pq, 0xab },
 	{ ABYTE,	ybyte,	Px, 1 },
-	{ ACALL,	ycall,	Px, 0xff,(02),0xe8 },
+	{ ACALL,	ycall,	Px, 0xff,(02),0xff,(0x15),0xe8 },
 	{ ACLC,		ynone,	Px, 0xf8 },
 	{ ACLD,		ynone,	Px, 0xfc },
 	{ ACLI,		ynone,	Px, 0xfa },
@@ -430,7 +431,8 @@ Optab optab[] =
 	{ AIRETW,	ynone,	Pe, 0xcf },
 	{ AJCC,		yjcond,	Px, 0x73,0x83,(00) },
 	{ AJCS,		yjcond,	Px, 0x72,0x82 },
-	{ AJCXZ,	yloop,	Px, 0xe3 },
+	{ AJCXZL,	yloop,	Px, 0xe3 },
+	{ AJCXZW,	yloop,	Px, 0xe3 },
 	{ AJEQ,		yjcond,	Px, 0x74,0x84 },
 	{ AJGE,		yjcond,	Px, 0x7d,0x8d },
 	{ AJGT,		yjcond,	Px, 0x7f,0x8f },
@@ -705,6 +707,8 @@ Optab optab[] =
 	{ ACMPXCHGW,	yrl_ml,	Pm, 0xb1 },
 	{ ACMPXCHG8B,	yscond,	Pm, 0xc7,(01) },
 
+	{ ARDTSC,	ynone,	Pm, 0x31 },
+
 	{ AXADDB,	yrb_mb,	Pb, 0x0f,0xc0 },
 	{ AXADDL,	yrl_ml,	Pm, 0xc1 },
 	{ AXADDW,	yrl_ml,	Pe, 0x0f,0xc1 },
@@ -751,5 +755,11 @@ Optab optab[] =
 	{ AFCMOVNU,	yfcmv,	Px, 0xdb,(03) },
 	{ AFCMOVUN,	yfcmv,	Px, 0xda,(03) },
 
+	{ ALFENCE, ynone, Pm, 0xae,0xe8 },
+	{ AMFENCE, ynone, Pm, 0xae,0xf0 },
+	{ ASFENCE, ynone, Pm, 0xae,0xf8 },
+
+	{ AEMMS, ynone, Pm, 0x77 },
+
 	0
 };
diff --git a/src/cmd/8l/pass.c b/src/cmd/8l/pass.c
index 2e0990c5a..9034fdf3a 100644
--- a/src/cmd/8l/pass.c
+++ b/src/cmd/8l/pass.c
@@ -259,7 +259,7 @@ patch(void)
 				// Convert
 				//   op	  n(GS), reg
 				// to
-				//   MOVL 0x2C(FS), reg
+				//   MOVL 0x14(FS), reg
 				//   op	  n(reg), reg
 				// The purpose of this patch is to fix some accesses
 				// to extern register variables (TLS) on Windows, as
@@ -273,7 +273,7 @@ patch(void)
 					q->as = p->as;
 					p->as = AMOVL;
 					p->from.type = D_INDIR+D_FS;
-					p->from.offset = 0x2C;
+					p->from.offset = 0x14;
 				}
 			}
 			if(HEADTYPE == Hlinux) {
@@ -307,9 +307,12 @@ patch(void)
 					p->from.offset = 0;
 				}
 			}
-			if(p->as == ACALL || (p->as == AJMP && p->to.type != D_BRANCH)) {
+			if((p->as == ACALL && p->to.type != D_BRANCH) || (p->as == AJMP && p->to.type != D_BRANCH)) {
 				s = p->to.sym;
-				if(s) {
+				if(p->to.type == D_INDIR+D_ADDR) {
+					 /* skip check if this is an indirect call (CALL *symbol(SB)) */
+					 continue;
+				} else if(s) {
 					if(debug['c'])
 						Bprint(&bso, "%s calls %s\n", TNAME, s->name);
 					if((s->type&~SSUB) != STEXT) {
@@ -421,7 +424,7 @@ dostkoff(void)
 			case Hwindows:
 				p->as = AMOVL;
 				p->from.type = D_INDIR+D_FS;
-				p->from.offset = 0x2c;
+				p->from.offset = 0x14;
 				p->to.type = D_CX;
 
 				p = appendp(p);
@@ -524,10 +527,18 @@ dostkoff(void)
 			p = appendp(p);	// save frame size in DX
 			p->as = AMOVL;
 			p->to.type = D_DX;
-			/* 160 comes from 3 calls (3*8) 4 safes (4*8) and 104 guard */
 			p->from.type = D_CONST;
-			if(autoffset+160+cursym->text->to.offset2 > 4096)
-				p->from.offset = (autoffset+160) & ~7LL;
+
+			// If we ask for more stack, we'll get a minimum of StackMin bytes.
+			// We need a stack frame large enough to hold the top-of-stack data,
+			// the function arguments+results, our caller's PC, our frame,
+			// a word for the return PC of the next call, and then the StackLimit bytes
+			// that must be available on entry to any function called from a function
+			// that did a stack check.  If StackMin is enough, don't ask for a specific
+			// amount: then we can use the custom functions and save a few
+			// instructions.
+			if(StackTop + cursym->text->to.offset2 + PtrSize + autoffset + PtrSize + StackLimit >= StackMin)
+				p->from.offset = (autoffset+7) & ~7LL;
 
 			p = appendp(p);	// save arg size in AX
 			p->as = AMOVL;
diff --git a/src/cmd/8l/span.c b/src/cmd/8l/span.c
index a4cba1257..81c1d37eb 100644
--- a/src/cmd/8l/span.c
+++ b/src/cmd/8l/span.c
@@ -83,7 +83,10 @@ span1(Sym *s)
 						loop++;
 						q->back ^= 2;
 					}
-					s->p[q->pc+1] = v;
+					if(q->as == AJCXZW)
+						s->p[q->pc+2] = v;
+					else
+						s->p[q->pc+1] = v;
 				} else {
 					bp = s->p + q->pc + q->mark - 4;
 					*bp++ = v;
@@ -282,6 +285,8 @@ oclass(Adr *a)
 				}
 				return Yxxx;
 			}
+			//if(a->type == D_INDIR+D_ADDR)
+			//	print("*Ycol\n");
 			return Ycol;
 		}
 		return Ym;
@@ -1056,9 +1061,10 @@ found:
 
 	case Zbr:
 	case Zjmp:
+	case Zloop:
 		q = p->pcond;
 		if(q == nil) {
-			diag("jmp/branch without target");
+			diag("jmp/branch/loop without target");
 			errorexit();
 		}
 		if(q->as == ATEXT) {
@@ -1084,8 +1090,12 @@ found:
 		if(p->back & 1) {
 			v = q->pc - (p->pc + 2);
 			if(v >= -128) {
+				if(p->as == AJCXZW)
+					*andptr++ = 0x67;
 				*andptr++ = op;
 				*andptr++ = v;
+			} else if(t[2] == Zloop) {
+				diag("loop too far: %P", p);
 			} else {
 				v -= 5-2;
 				if(t[2] == Zbr) {
@@ -1105,8 +1115,12 @@ found:
 		p->forwd = q->comefrom;
 		q->comefrom = p;
 		if(p->back & 2)	{ // short
+			if(p->as == AJCXZW)
+				*andptr++ = 0x67;
 			*andptr++ = op;
 			*andptr++ = 0;
+		} else if(t[2] == Zloop) {
+			diag("loop too far: %P", p);
 		} else {
 			if(t[2] == Zbr)
 				*andptr++ = 0x0f;
@@ -1131,18 +1145,17 @@ found:
 		r->add = p->to.offset;
 		put4(0);
 		break;
-
-	case Zloop:
-		q = p->pcond;
-		if(q == nil) {
-			diag("loop without target");
-			errorexit();
-		}
-		v = q->pc - p->pc - 2;
-		if(v < -128 && v > 127)
-			diag("loop too far: %P", p);
+	
+	case Zcallind:
 		*andptr++ = op;
-		*andptr++ = v;
+		*andptr++ = o->op[z+1];
+		r = addrel(cursym);
+		r->off = p->pc + andptr - and;
+		r->type = D_ADDR;
+		r->siz = 4;
+		r->add = p->to.offset;
+		r->sym = p->to.sym;
+		put4(0);
 		break;
 
 	case Zbyte: