sync compilers with Plan 9

remove 1[acl] 2[acl]

10 files changed, 225 insertions, 350 deletions

diff --git a/utils/6l/Nt.c b/utils/6l/Nt.c
deleted file mode 100644
index 73c6f795..00000000
--- a/utils/6l/Nt.c
+++ /dev/null
@@ -1,77 +0,0 @@
-#include <windows.h>
-#include "l.h"
-
-/*
- * fake malloc
- */
-void*
-malloc(uint n)
-{
-	void *p;
-
-	while(n & 7)
-		n++;
-	while(nhunk < n)
-		gethunk();
-	p = hunk;
-	nhunk -= n;
-	hunk += n;
-	return p;
-}
-
-void
-free(void *p)
-{
-	USED(p);
-}
-
-void*
-calloc(uint m, uint n)
-{
-	void *p;
-
-	n *= m;
-	p = malloc(n);
-	memset(p, 0, n);
-	return p;
-}
-
-void*
-realloc(void *p, uint n)
-{
-	void *new;
-
-	new = malloc(n);
-	if(new && p)
-		memmove(new, p, n);
-	return new;
-}
-
-#define	Chunk	(1*1024*1024)
-
-void*
-mysbrk(ulong size)
-{
-	void *v;
-	static int chunk;
-	static uchar *brk;
-
-	if(chunk < size) {
-		chunk = Chunk;
-		if(chunk < size)
-			chunk = Chunk + size;
-		brk = VirtualAlloc(NULL, chunk, MEM_COMMIT, PAGE_EXECUTE_READWRITE); 	
-		if(brk == 0)
-			return (void*)-1;
-	}
-	v = brk;
-	chunk -= size;
-	brk += size;
-	return v;
-}
-
-double
-cputime(void)
-{
-	return ((double)0);
-}
diff --git a/utils/6l/Plan9.c b/utils/6l/Plan9.c
deleted file mode 100644
index f4cf23f4..00000000
--- a/utils/6l/Plan9.c
+++ /dev/null
@@ -1,57 +0,0 @@
-#include	"l.h"
-
-/*
- * fake malloc
- */
-void*
-malloc(ulong n)
-{
-	void *p;
-
-	while(n & 7)
-		n++;
-	while(nhunk < n)
-		gethunk();
-	p = hunk;
-	nhunk -= n;
-	hunk += n;
-	return p;
-}
-
-void
-free(void *p)
-{
-	USED(p);
-}
-
-void*
-calloc(ulong m, ulong n)
-{
-	void *p;
-
-	n *= m;
-	p = malloc(n);
-	memset(p, 0, n);
-	return p;
-}
-
-void*
-realloc(void *p, ulong n)
-{
-	USED(p);
-	USED(n);
-	fprint(2, "realloc called\n");
-	abort();
-	return 0;
-}
-
-void*
-mysbrk(ulong size)
-{
-	return sbrk(size);
-}
-
-void
-setmalloctag(void*, ulong)
-{
-}
diff --git a/utils/6l/Posix.c b/utils/6l/Posix.c
deleted file mode 100644
index 7c3a661f..00000000
--- a/utils/6l/Posix.c
+++ /dev/null
@@ -1,80 +0,0 @@
-#include	"l.h"
-#include <sys/types.h>
-#include <sys/times.h>
-#undef getwd
-#include <unistd.h>	/* For sysconf() and _SC_CLK_TCK */
-
-/*
- * fake malloc
- */
-void*
-malloc(size_t n)
-{
-	void *p;
-
-	while(n & 7)
-		n++;
-	while(nhunk < n)
-		gethunk();
-	p = hunk;
-	nhunk -= n;
-	hunk += n;
-	return p;
-}
-
-void
-free(void *p)
-{
-	USED(p);
-}
-
-void*
-calloc(size_t m, size_t n)
-{
-	void *p;
-
-	n *= m;
-	p = malloc(n);
-	memset(p, 0, n);
-	return p;
-}
-
-void*
-realloc(void *p, size_t n)
-{
-	fprint(2, "realloc called\n", p, n);
-	abort();
-	return 0;
-}
-
-void*
-mysbrk(ulong size)
-{
-	return (void*)sbrk(size);
-}
-
-double
-cputime(void)
-{
-
-	struct tms tmbuf;
-	double	ret_val;
-
-	/*
-	 * times() only fials if &tmbuf is invalid.
-	 */
-	(void)times(&tmbuf);
-	/*
-	 * Return the total time (in system clock ticks)
-	 * spent in user code and system
-	 * calls by both the calling process and its children.
-	 */
-	ret_val = (double)(tmbuf.tms_utime + tmbuf.tms_stime +
-			tmbuf.tms_cutime + tmbuf.tms_cstime);
-	/*
-	 * Convert to seconds.
-	 */
-	ret_val *= sysconf(_SC_CLK_TCK);
-	return ret_val;
-	
-}
diff --git a/utils/6l/README b/utils/6l/README
deleted file mode 100644
index 602f5378..00000000
--- a/utils/6l/README
+++ /dev/null
@@ -1 +0,0 @@
-this is an intermediate version not yet ready for use
diff --git a/utils/6l/asm.c b/utils/6l/asm.c
index c4d46633..a4afaa2a 100644
--- a/utils/6l/asm.c
+++ b/utils/6l/asm.c
@@ -2,7 +2,7 @@
 
 #define	Dbufslop	100
 
-#define PADDR(a)	((ulong)(a) & ~0x80000000)
+#define PADDR(a)	((a) & ~0xfffffffff0000000ull)
 
 vlong
 entryvalue(void)
@@ -28,15 +28,15 @@ entryvalue(void)
 	return s->value;
 }
 
-void
-wputl(ushort w)
+/* these need to take long arguments to be compatible with elf.c */void
+wputl(long w)
 {
 	cput(w);
 	cput(w>>8);
 }
 
 void
-wput(ushort w)
+wput(long w)
 {
 	cput(w>>8);
 	cput(w);
@@ -68,6 +68,13 @@ lputl(long l)
 }
 
 void
+llputl(vlong v)
+{
+	lputl(v);
+	lputl(v>>32);
+}
+
+void
 strnput(char *s, int n)
 {
 	for(; *s && n > 0; s++){
@@ -133,6 +140,7 @@ asmb(void)
 		diag("unknown header type %ld", HEADTYPE);
 	case 2:
 	case 5:
+	case 6:
 		seek(cout, HEADR+textsize, 0);
 		break;
 	}
@@ -166,6 +174,7 @@ asmb(void)
 		default:
 		case 2:
 		case 5:
+		case 6:
 			seek(cout, HEADR+textsize+datsize, 0);
 			break;
 		}
@@ -210,65 +219,11 @@ asmb(void)
 		lput(lcsize);			/* line offsets */
 		llput(vl);			/* va of entry */
 		break;
-	case 3:	/* plan9 */
-		magic = 4*26*26+7;
-		if(dlm)
-			magic |= 0x80000000;
-		lput(magic);			/* magic */
-		lput(textsize);			/* sizes */
-		lput(datsize);
-		lput(bsssize);
-		lput(symsize);			/* nsyms */
-		lput(entryvalue());		/* va of entry */
-		lput(spsize);			/* sp offsets */
-		lput(lcsize);			/* line offsets */
-		break;
 	case 5:
-		strnput("\177ELF", 4);		/* e_ident */
-		cput(1);			/* class = 32 bit */
-		cput(1);			/* data = LSB */
-		cput(1);			/* version = CURRENT */
-		strnput("", 9);
-		wputl(2);			/* type = EXEC */
-		wputl(62);			/* machine = AMD64 */
-		lputl(1L);			/* version = CURRENT */
-		lputl(PADDR(entryvalue()));	/* entry vaddr */
-		lputl(52L);			/* offset to first phdr */
-		lputl(0L);			/* offset to first shdr */
-		lputl(0L);			/* processor specific flags */
-		wputl(52);			/* Ehdr size */
-		wputl(32);			/* Phdr size */
-		wputl(3);			/* # of Phdrs */
-		wputl(0);			/* Shdr size */
-		wputl(0);			/* # of Shdrs */
-		wputl(0);			/* Shdr string size */
-
-		lputl(1L);			/* text - type = PT_LOAD */
-		lputl(HEADR);			/* file offset */
-		lputl(INITTEXT);		/* vaddr */
-		lputl(PADDR(INITTEXT));		/* paddr */
-		lputl(textsize);		/* file size */
-		lputl(textsize);		/* memory size */
-		lputl(0x05L);			/* protections = RX */
-		lputl(INITRND);			/* alignment */
-
-		lputl(1L);			/* data - type = PT_LOAD */
-		lputl(HEADR+textsize);		/* file offset */
-		lputl(INITDAT);			/* vaddr */
-		lputl(PADDR(INITDAT));		/* paddr */
-		lputl(datsize);			/* file size */
-		lputl(datsize+bsssize);		/* memory size */
-		lputl(0x06L);			/* protections = RW */
-		lputl(INITRND);			/* alignment */
-
-		lputl(0L);			/* data - type = PT_NULL */
-		lputl(HEADR+textsize+datsize);	/* file offset */
-		lputl(0L);
-		lputl(0L);
-		lputl(symsize);			/* symbol table size */
-		lputl(lcsize);			/* line number size */
-		lputl(0x04L);			/* protections = R */
-		lputl(0x04L);			/* alignment */
+		elf32(debug['8']? I386: AMD64, ELFDATA2LSB, 0, nil);
+		break;
+	case 6:
+		elf64(AMD64, ELFDATA2LSB, 0, nil);
 		break;
 	}
 	cflush();
diff --git a/utils/6l/l.h b/utils/6l/l.h
index 115aa69d..5a1c6b95 100644
--- a/utils/6l/l.h
+++ b/utils/6l/l.h
@@ -1,6 +1,7 @@
 #include	<lib9.h>
 #include	<bio.h>
 #include	"../6c/6.out.h"
+#include	"../8l/elf.h"
 
 #ifndef	EXTERN
 #define	EXTERN	extern
@@ -14,6 +15,8 @@
 	if(--cbc <= 0)\
 		cflush(); }
 
+#define	LIBNAMELEN	300
+
 typedef	struct	Adr	Adr;
 typedef	struct	Prog	Prog;
 typedef	struct	Sym	Sym;
@@ -226,7 +229,7 @@ EXTERN union
 {
 	struct
 	{
-		char	obuf[MAXIO];			/* output buffer */
+		uchar	obuf[MAXIO];			/* output buffer */
 		uchar	ibuf[MAXIO];			/* input buffer */
 	} u;
 	char	dbuf[1];
@@ -235,22 +238,26 @@ EXTERN union
 #define	cbuf	u.obuf
 #define	xbuf	u.ibuf
 
+#pragma	varargck	type	"A"	int
 #pragma	varargck	type	"A"	uint
 #pragma	varargck	type	"D"	Adr*
 #pragma	varargck	type	"P"	Prog*
 #pragma	varargck	type	"R"	int
 #pragma	varargck	type	"S"	char*
 
+#pragma	varargck	argpos	diag 1
+
 EXTERN	long	HEADR;
 EXTERN	long	HEADTYPE;
 EXTERN	vlong	INITDAT;
 EXTERN	long	INITRND;
 EXTERN	vlong	INITTEXT;
+EXTERN	vlong	INITTEXTP;
 EXTERN	char*	INITENTRY;		/* entry point */
 EXTERN	Biobuf	bso;
 EXTERN	long	bsssize;
 EXTERN	int	cbc;
-EXTERN	char*	cbp;
+EXTERN	uchar*	cbp;
 EXTERN	char*	pcstr;
 EXTERN	int	cout;
 EXTERN	Auto*	curauto;
@@ -323,6 +330,7 @@ int	Pconv(Fmt*);
 int	Rconv(Fmt*);
 int	Sconv(Fmt*);
 void	addhist(long, int);
+void	addlibpath(char*);
 Prog*	appendp(Prog*);
 void	asmb(void);
 void	asmdyn(void);
@@ -349,8 +357,10 @@ void	dynreloc(Sym*, ulong, int);
 vlong	entryvalue(void);
 void	errorexit(void);
 void	export(void);
+int	fileexists(char*);
 int	find1(long, int);
 int	find2(long, int);
+char*	findlib(char*);
 void	follow(void);
 void	gethunk(void);
 void	histtoauto(void);
@@ -361,6 +371,8 @@ void	ldobj(int, long, char*);
 void	loadlib(void);
 void	listinit(void);
 Sym*	lookup(char*, int);
+void	llput(vlong v);
+void	llputl(vlong v);
 void	lput(long);
 void	lputl(long);
 void	main(int, char*[]);
@@ -376,17 +388,13 @@ int	relinv(int);
 long	reuse(Prog*, Sym*);
 vlong	rnd(vlong, vlong);
 void	span(void);
+void	strnput(char*, int);
 void	undef(void);
 void	undefsym(Sym*);
 vlong	vaddr(Adr*);
-void	wput(ushort);
+void	wput(long);
+void	wputl(long);
 void	xdefine(char*, int, vlong);
 void	xfol(Prog*);
 int	zaddr(uchar*, Adr*, Sym*[]);
 void	zerosig(char*);
-
-#pragma	varargck	type	"D"	Adr*
-#pragma	varargck	type	"P"	Prog*
-#pragma	varargck	type	"R"	int
-#pragma	varargck	type	"A"	int
-#pragma	varargck	argpos	diag 1
diff --git a/utils/6l/mkfile b/utils/6l/mkfile
index 19a65c8b..e0bb5f7c 100644
--- a/utils/6l/mkfile
+++ b/utils/6l/mkfile
@@ -11,6 +11,7 @@ OFILES=\
 	list.$O\
 	enam.$O\
 	$TARGMODEL.$O\
+	elf.$O\
 
 HFILES=\
 	l.h\
@@ -23,7 +24,12 @@ BIN=$ROOT/$OBJDIR/bin
 
 <$ROOT/mkfiles/mkone-$SHELLTYPE
 
-CFLAGS=	$CFLAGS -I../include
+CFLAGS=	$CFLAGS -I../include -I.
 
 enam.$O:	../6c/enam.c
 	$CC $CFLAGS ../6c/enam.c
+elf.$O:	../ld/elf.c
+	$CC $CFLAGS ../ld/elf.c
+
+$TARGMODEL.$O:	../ld/$TARGMODEL.c
+	$CC $CFLAGS ../ld/$TARGMODEL.c
diff --git a/utils/6l/obj.c b/utils/6l/obj.c
index 9f57d612..ed52ec2b 100644
--- a/utils/6l/obj.c
+++ b/utils/6l/obj.c
@@ -12,14 +12,25 @@ char	thechar		= '6';
 char	*thestring 	= "amd64";
 char	*paramspace	= "FP";
 
+char**	libdir;
+int	nlibdir	= 0;
+static	int	maxlibdir = 0;
+
 /*
- *	-H2 -T4136 -R4096		is plan9 64-bit format
- *	-H3 -T4128 -R4096		is plan9 32-bit format
+ *	-H2 -T0x200028 -R0x200000	is plan9 format (was -T4136 -R4096)
  *	-H5 -T0x80110000 -R4096		is ELF32
+ *	-H6 -T0x2000e8 -R0x200000	is ELF64
  *
- *	options used: 189BLQSWabcjlnpsvz
+ *	options used: 189BLPQSVWabcjlnpsvz
  */
 
+void
+usage(void)
+{
+	diag("usage: %s [-options] objects", argv0);
+	errorexit();
+}
+
 static int
 isobjfile(char *f)
 {
@@ -47,6 +58,7 @@ main(int argc, char *argv[])
 {
 	int i, c;
 	char *a;
+	char name[LIBNAMELEN];
 
 	Binit(&bso, 1, OWRITE);
 	cout = -1;
@@ -56,6 +68,7 @@ main(int argc, char *argv[])
 	outfile = "6.out";
 	HEADTYPE = -1;
 	INITTEXT = -1;
+	INITTEXTP = -1;
 	INITDAT = -1;
 	INITRND = -1;
 	INITENTRY = 0;
@@ -78,11 +91,19 @@ main(int argc, char *argv[])
 		if(a)
 			HEADTYPE = atolwhex(a);
 		break;
+	case 'L':
+		addlibpath(EARGF(usage()));
+		break;
 	case 'T':
 		a = ARGF();
 		if(a)
 			INITTEXT = atolwhex(a);
 		break;
+	case 'P':
+		a = ARGF();
+		if(a)
+			INITTEXTP = atolwhex(a);
+		break;
 	case 'D':
 		a = ARGF();
 		if(a)
@@ -106,12 +127,20 @@ main(int argc, char *argv[])
 		break;
 	} ARGEND
 	USED(argc);
-	if(*argv == 0) {
-		diag("usage: 6l [-options] objects");
-		errorexit();
-	}
+	if(*argv == 0)
+		usage();
 	if(!debug['9'] && !debug['U'] && !debug['B'])
 		debug[DEFAULT] = 1;
+	a = getenv("ccroot");
+	if(a != nil && *a != '\0') {
+		if(!fileexists(a)) {
+			diag("nonexistent $ccroot: %s", a);
+			errorexit();
+		}
+	}else
+		a = "";
+	snprint(name, sizeof(name), "%s/%s/lib", a, thestring);
+	addlibpath(name);
 	if(HEADTYPE == -1) {
 		if(debug['B'])
 			HEADTYPE = 2;
@@ -125,31 +154,33 @@ main(int argc, char *argv[])
 	case 2:	/* plan 9 */
 		HEADR = 32L+8L;
 		if(INITTEXT == -1)
-			INITTEXT = 4096+HEADR;
+			INITTEXT = 0x200000+HEADR;
 		if(INITDAT == -1)
 			INITDAT = 0;
 		if(INITRND == -1)
-			INITRND = 4096;
+			INITRND = 0x200000;
 		break;
-	case 3:	/* plan 9 */
-		HEADR = 32L;
+	case 5:	/* elf32 executable */
+		HEADR = rnd(Ehdr32sz+3*Phdr32sz, 16);
 		if(INITTEXT == -1)
-			INITTEXT = 4096+32;
+			INITTEXT = 0xf0110000L;
 		if(INITDAT == -1)
 			INITDAT = 0;
 		if(INITRND == -1)
 			INITRND = 4096;
 		break;
-	case 5:	/* elf32 executable */
-		HEADR = rnd(52L+3*32L, 16);
+	case 6:	/* ELF64 executable */
+		HEADR = rnd(Ehdr64sz+3*Phdr64sz, 16);
 		if(INITTEXT == -1)
-			INITTEXT = 0x80110000L;
+			INITTEXT = 0x200000+HEADR;
 		if(INITDAT == -1)
 			INITDAT = 0;
 		if(INITRND == -1)
-			INITRND = 4096;
+			INITRND = 0x200000;
 		break;
 	}
+	if (INITTEXTP == -1)
+		INITTEXTP = INITTEXT;
 	if(INITDAT != 0 && INITRND != 0)
 		print("warning: -D0x%llux is ignored because of -R0x%lux\n",
 			INITDAT, INITRND);
@@ -283,7 +314,7 @@ main(int argc, char *argv[])
 	dtype = 4;
 	cout = create(outfile, 1, 0775);
 	if(cout < 0) {
-		diag("cannot create %s", outfile);
+		diag("cannot create %s: %r", outfile);
 		errorexit();
 	}
 	version = 0;
@@ -351,6 +382,42 @@ main(int argc, char *argv[])
 }
 
 void
+addlibpath(char *arg)
+{
+	char **p;
+
+	if(nlibdir >= maxlibdir) {
+		if(maxlibdir == 0)
+			maxlibdir = 8;
+		else
+			maxlibdir *= 2;
+		p = malloc(maxlibdir*sizeof(*p));
+		if(p == nil) {
+			diag("out of memory");
+			errorexit();
+		}
+		memmove(p, libdir, nlibdir*sizeof(*p));
+		free(libdir);
+		libdir = p;
+	}
+	libdir[nlibdir++] = strdup(arg);
+}
+
+char*
+findlib(char *file)
+{
+	int i;
+	char name[LIBNAMELEN];
+
+	for(i = 0; i < nlibdir; i++) {
+		snprint(name, sizeof(name), "%s/%s", libdir[i], file);
+		if(fileexists(name))
+			return libdir[i];
+	}
+	return nil;
+}
+
+void
 loadlib(void)
 {
 	int i;
@@ -390,25 +457,26 @@ objfile(char *file)
 	int f, work;
 	Sym *s;
 	char magbuf[SARMAG];
-	char name[100], pname[150];
+	char name[LIBNAMELEN], pname[LIBNAMELEN];
 	struct ar_hdr arhdr;
 	char *e, *start, *stop;
 
-	if(file[0] == '-' && file[1] == 'l') {
-		if(debug['9'])
-			sprint(name, "/%s/lib/lib", thestring);
-		else
-			sprint(name, "/usr/%clib/lib", thechar);
-		strcat(name, file+2);
-		strcat(name, ".a");
-		file = name;
-	}
 	if(debug['v'])
 		Bprint(&bso, "%5.2f ldobj: %s\n", cputime(), file);
 	Bflush(&bso);
+	if(file[0] == '-' && file[1] == 'l') {
+		snprint(pname, sizeof(pname), "lib%s.a", file+2);
+		e = findlib(pname);
+		if(e == nil) {
+			diag("cannot find library: %s", file);
+			errorexit();
+		}
+		snprint(name, sizeof(name), "%s/%s", e, pname);
+		file = name;
+	}
 	f = open(file, 0);
 	if(f < 0) {
-		diag("cannot open file: %s", file);
+		diag("cannot open %s: %r", file);
 		errorexit();
 	}
 	l = read(f, magbuf, SARMAG);
@@ -467,7 +535,8 @@ objfile(char *file)
 			l |= (e[3] & 0xff) << 16;
 			l |= (e[4] & 0xff) << 24;
 			seek(f, l, 0);
-			l = read(f, &arhdr, SAR_HDR);
+			/* need readn to read the dumps (at least) */
+			l = readn(f, &arhdr, SAR_HDR);
 			if(l != SAR_HDR)
 				goto bad;
 			if(strncmp(arhdr.fmag, ARFMAG, sizeof(arhdr.fmag)))
@@ -494,7 +563,7 @@ int
 zaddr(uchar *p, Adr *a, Sym *h[])
 {
 	int c, t, i;
-	long l;
+	int l;
 	Sym *s;
 	Auto *u;
 
@@ -579,25 +648,24 @@ zaddr(uchar *p, Adr *a, Sym *h[])
 void
 addlib(char *obj)
 {
-	char name[1024], comp[256], *p;
-	int i;
+	char fn1[LIBNAMELEN], fn2[LIBNAMELEN], comp[LIBNAMELEN], *p, *name;
+	int i, search;
 
 	if(histfrogp <= 0)
 		return;
 
+	name = fn1;
+	search = 0;
 	if(histfrog[0]->name[1] == '/') {
 		sprint(name, "");
 		i = 1;
-	} else
-	if(histfrog[0]->name[1] == '.') {
+	} else if(histfrog[0]->name[1] == '.') {
 		sprint(name, ".");
 		i = 0;
 	} else {
-		if(debug['9'])
-			sprint(name, "/%s/lib", thestring);
-		else
-			sprint(name, "/usr/%clib", thechar);
+		sprint(name, "");
 		i = 0;
+		search = 1;
 	}
 
 	for(; i<histfrogp; i++) {
@@ -620,13 +688,25 @@ addlib(char *obj)
 			memmove(p+strlen(thestring), p+2, strlen(p+2)+1);
 			memmove(p, thestring, strlen(thestring));
 		}
-		if(strlen(name) + strlen(comp) + 3 >= sizeof(name)) {
+		if(strlen(fn1) + strlen(comp) + 3 >= sizeof(fn1)) {
 			diag("library component too long");
 			return;
 		}
-		strcat(name, "/");
-		strcat(name, comp);
+		if(i > 0 || !search)
+			strcat(fn1, "/");
+		strcat(fn1, comp);
+	}
+
+	cleanname(name);
+
+	if(search){
+		p = findlib(name);
+		if(p != nil){
+			snprint(fn2, sizeof(fn2), "%s/%s", p, name);
+			name = fn2;
+		}
 	}
+
 	for(i=0; i<libraryp; i++)
 		if(strcmp(name, library[i]) == 0)
 			return;
@@ -1138,8 +1218,7 @@ lookup(char *symb, int v)
 	for(p=symb; c = *p; p++)
 		h = h+h+h + c;
 	l = (p - symb) + 1;
-	if(h < 0)
-		h = ~h;
+	h &= 0xffffff;
 	h %= NHASH;
 	for(s = hash[h]; s != S; s = s->link)
 		if(s->version == v)
@@ -1288,16 +1367,24 @@ void
 doprof2(void)
 {
 	Sym *s2, *s4;
-	Prog *p, *q, *ps2, *ps4;
+	Prog *p, *q, *q2, *ps2, *ps4;
 
 	if(debug['v'])
 		Bprint(&bso, "%5.2f profile 2\n", cputime());
 	Bflush(&bso);
 
-	s2 = lookup("_profin", 0);
-	s4 = lookup("_profout", 0);
+	if(debug['e']){
+		s2 = lookup("_tracein", 0);
+		s4 = lookup("_traceout", 0);
+	}else{
+		s2 = lookup("_profin", 0);
+		s4 = lookup("_profout", 0);
+	}
 	if(s2->type != STEXT || s4->type != STEXT) {
-		diag("_profin/_profout not defined");
+		if(debug['e'])
+			diag("_tracein/_traceout not defined %d %d", s2->type, s4->type);
+		else
+			diag("_profin/_profout not defined");
 		return;
 	}
 
@@ -1338,7 +1425,20 @@ doprof2(void)
 			q->line = p->line;
 			q->pc = p->pc;
 			q->link = p->link;
-			p->link = q;
+			if(debug['e']){		/* embedded tracing */
+				q2 = prg();
+				p->link = q2;
+				q2->link = q;
+
+				q2->line = p->line;
+				q2->pc = p->pc;
+
+				q2->as = AJMP;
+				q2->to.type = D_BRANCH;
+				q2->to.sym = p->to.sym;
+				q2->pcond = q->link;
+			}else
+				p->link = q;
 			p = q;
 			p->as = ACALL;
 			p->to.type = D_BRANCH;
@@ -1349,6 +1449,17 @@ doprof2(void)
 		}
 		if(p->as == ARET) {
 			/*
+			 * RET (default)
+			 */
+			if(debug['e']){		/* embedded tracing */
+				q = prg();
+				q->line = p->line;
+				q->pc = p->pc;
+				q->link = p->link;
+				p->link = q;
+				p = q;
+			}
+			/*
 			 * RET
 			 */
 			q = prg();
diff --git a/utils/6l/optab.c b/utils/6l/optab.c
index db22ca67..133bb1d2 100644
--- a/utils/6l/optab.c
+++ b/utils/6l/optab.c
@@ -15,8 +15,10 @@ uchar	ynop[] =
 	Ynone,	Ynone,	Zpseudo,1,
 	Ynone,	Yml,	Zpseudo,1,
 	Ynone,	Yrf,	Zpseudo,1,
+	Ynone,	Yxr,	Zpseudo,1,
 	Yml,	Ynone,	Zpseudo,1,
 	Yrf,	Ynone,	Zpseudo,1,
+	Yxr,	Ynone,	Zpseudo,1,
 	0
 };
 uchar	yxorb[] =
@@ -753,7 +755,7 @@ Optab optab[] =
 	{ AMOVBWSX,	ymb_rl,	Pq, 0xbe },
 	{ AMOVBWZX,	ymb_rl,	Pq, 0xb6 },
 	{ AMOVO,	yxmov,	Pe, 0x6f,0x7f },
-	{ AMOVOU,	yxmov,	Pf2, 0x6f,0x7f },
+	{ AMOVOU,	yxmov,	Pf3, 0x6f,0x7f },
 	{ AMOVHLPS,	yxr,	Pm, 0x12 },
 	{ AMOVHPD,	yxmov,	Pe, 0x16,0x17 },
 	{ AMOVHPS,	yxmov,	Pm, 0x16,0x17 },
@@ -762,7 +764,7 @@ Optab optab[] =
 	{ AMOVLPD,	yxmov,	Pe, 0x12,0x13 },
 	{ AMOVLPS,	yxmov,	Pm, 0x12,0x13 },
 	{ AMOVLQSX,	yml_rl,	Pw, 0x63 },
-	{ AMOVLQZX,	yml_rl,	Px, 0x63 },
+	{ AMOVLQZX,	yml_rl,	Px, 0x8b },
 	{ AMOVMSKPD,	yxrrl,	Pq, 0x50 },
 	{ AMOVMSKPS,	yxrrl,	Pm, 0x50 },
 	{ AMOVNTO,	yxr_ml,	Pe, 0xe7 },
@@ -877,9 +879,9 @@ Optab optab[] =
 	{ APOPQ,	ypopl,	Py, 0x58,0x8f,(00) },
 	{ APOPW,	ypopl,	Pe, 0x58,0x8f,(00) },
 	{ APOR,		ymm,	Py, 0xeb,Pe,0xeb },
-	{ APSADBW,	yxm,	Pw, Pe,0xf6 },
+	{ APSADBW,	yxm,	Pq, 0xf6 },
 	{ APSHUFHW,	yxshuf,	Pf3, 0x70 },
-	{ APSHUFL,	yxm,	Pw, Pe,0x70 },
+	{ APSHUFL,	yxshuf,	Pq, 0x70 },
 	{ APSHUFLW,	yxshuf,	Pf2, 0x70 },
 	{ APSHUFW,	ymshuf,	Pm, 0x70 },
 	{ APSLLO,	ypsdq,	Pq, 0x73,(07) },
diff --git a/utils/6l/span.c b/utils/6l/span.c
index 1ce533d0..e49d873f 100644
--- a/utils/6l/span.c
+++ b/utils/6l/span.c
@@ -127,9 +127,17 @@ putsymb(char *s, int t, vlong v, int ver)
 	if(t == 'f')
 		s++;
 	l = 4;
-	if(!debug['8']){
+	switch(HEADTYPE){
+	default:
+		break;
+	case 5:
+		if(debug['8'])
+			break;
+	case 2:
+	case 6:
 		lput(v>>32);
 		l = 8;
+		break;
 	}
 	lput(v);
 	if(ver)
@@ -245,12 +253,12 @@ asmlc(void)
 		if(p->line == oldlc || p->as == ATEXT || p->as == ANOP) {
 			if(p->as == ATEXT)
 				curtext = p;
-			if(debug['L'])
+			if(debug['V'])
 				Bprint(&bso, "%6llux %P\n",
 					p->pc, p);
 			continue;
 		}
-		if(debug['L'])
+		if(debug['V'])
 			Bprint(&bso, "\t\t%6ld", lcsize);
 		v = (p->pc - oldpc) / MINLC;
 		while(v) {
@@ -258,7 +266,7 @@ asmlc(void)
 			if(v < 127)
 				s = v;
 			cput(s+128);	/* 129-255 +pc */
-			if(debug['L'])
+			if(debug['V'])
 				Bprint(&bso, " pc+%ld*%d(%ld)", s, MINLC, s+128);
 			v -= s;
 			lcsize++;
@@ -272,7 +280,7 @@ asmlc(void)
 			cput(s>>16);
 			cput(s>>8);
 			cput(s);
-			if(debug['L']) {
+			if(debug['V']) {
 				if(s > 0)
 					Bprint(&bso, " lc+%ld(%d,%ld)\n",
 						s, 0, s);
@@ -287,14 +295,14 @@ asmlc(void)
 		}
 		if(s > 0) {
 			cput(0+s);	/* 1-64 +lc */
-			if(debug['L']) {
+			if(debug['V']) {
 				Bprint(&bso, " lc+%ld(%ld)\n", s, 0+s);
 				Bprint(&bso, "%6llux %P\n",
 					p->pc, p);
 			}
 		} else {
 			cput(64-s);	/* 65-128 -lc */
-			if(debug['L']) {
+			if(debug['V']) {
 				Bprint(&bso, " lc%ld(%ld)\n", s, 64-s);
 				Bprint(&bso, "%6llux %P\n",
 					p->pc, p);
@@ -307,7 +315,7 @@ asmlc(void)
 		cput(s);
 		lcsize++;
 	}
-	if(debug['v'] || debug['L'])
+	if(debug['v'] || debug['V'])
 		Bprint(&bso, "lcsize = %ld\n", lcsize);
 	Bflush(&bso);
 }
@@ -1385,7 +1393,7 @@ found:
 		q = p->pcond;
 		if(q) {
 			v = q->pc - p->pc - 2;
-			if(v < -128 && v > 127)
+			if(v < -128 || v > 127)
 				diag("loop too far: %P", p);
 			*andptr++ = op;
 			*andptr++ = v;
@@ -1574,7 +1582,7 @@ asmins(Prog *p)
 		/*
 		 * as befits the whole approach of the architecture,
 		 * the rex prefix must appear before the first opcode byte
-		 * (and thus after any 66/67/f2/f3 prefix bytes, but
+		 * (and thus after any 66/67/f2/f3/26/2e/3e prefix bytes, but
 		 * before the 0f opcode escape!), or it might be ignored.
 		 * note that the handbook often misleadingly shows 66/f2/f3 in `opcode'.
 		 */
@@ -1583,7 +1591,7 @@ asmins(Prog *p)
 		n = andptr - and;
 		for(np = 0; np < n; np++) {
 			c = and[np];
-			if(c != 0x66 && c != 0xf2 && c != 0xf3 && c != 0x67)
+			if(c != 0xf2 && c != 0xf3 && (c < 0x64 || c > 0x67) && c != 0x2e && c != 0x3e && c != 0x26)
 				break;
 		}
 		memmove(and+np+1, and+np, n-np);