diff --git a/src/cmd/5l/asm.c b/src/cmd/5l/asm.c
index 4d9264c9140..0b38e0b29b5 100644
--- a/src/cmd/5l/asm.c
+++ b/src/cmd/5l/asm.c
@@ -291,13 +291,11 @@ phsh(Elf64_Phdr *ph, Elf64_Shdr *sh)
 void
 asmb(void)
 {
-	Prog *p;
 	int32 t;
 	int a, dynsym;
 	uint32 va, fo, w, symo, startva;
 	uint32 symdatva = SYMDATVA;
 	int strtabsize;
-	Optab *o;
 	ElfEhdr *eh;
 	ElfPhdr *ph, *pph;
 	ElfShdr *sh;
@@ -312,28 +310,8 @@ asmb(void)
 	OFFSET = HEADR;
 	seek(cout, OFFSET, 0);
 	pc = INITTEXT;
-	for(cursym = textp; cursym != nil; cursym = cursym->next) {
-		for(p = cursym->text; p != P; p = p->link) {
-			setarch(p);
-			if(p->as == ATEXT)
-				autosize = p->to.offset + 4;
-			if(p->pc != pc) {
-				diag("phase error %ux sb %ux",
-					p->pc, pc);
-				if(!debug['a'])
-					prasm(curp);
-				pc = p->pc;
-			}
-			curp = p;
-			o = oplook(p);	/* could probably avoid this call */
-			if(thumb)
-				thumbasmout(p, o);
-			else
-				asmout(p, o);
-			pc += o->size;
-		}
-	}
-	cflush();
+	codeblk(pc, segtext.sect->len);
+	pc += segtext.sect->len;
 	if(seek(cout, 0, 1) != pc - segtext.vaddr + segtext.fileoff)
 		diag("text phase error");
 
@@ -846,10 +824,11 @@ asmthumbmap(void)
 }
 
 void
-asmout(Prog *p, Optab *o)
+asmout(Prog *p, Optab *o, int32 *out)
 {
 	int32 o1, o2, o3, o4, o5, o6, v;
 	int r, rf, rt, rt2;
+	Reloc *rel;
 
 PP = p;
 	o1 = 0;
@@ -991,6 +970,15 @@ if(debug['G']) print("%ux: %s: arm %d %d %d %d\n", (uint32)(p->pc), p->from.sym-
 	case 11:	/* word */
 		aclass(&p->to);
 		o1 = instoffset;
+		if(p->to.sym != S) {
+			rel = addrel(cursym);
+			rel->off = pc - cursym->value;
+			rel->siz = 4;
+			rel->type = D_ADDR;
+			rel->sym = p->to.sym;
+			rel->add = p->to.offset;
+			o1 = 0;
+		}
 		break;
 
 	case 12:	/* movw $lcon, reg */
@@ -1588,6 +1576,14 @@ if(debug['G']) print("%ux: %s: arm %d %d %d %d\n", (uint32)(p->pc), p->from.sym-
 		o1 |= (p->scond & C_SCOND) << 28;
 		break;
 	}
+	
+	out[0] = o1;
+	out[1] = o2;
+	out[2] = o3;
+	out[3] = o4;
+	out[4] = o5;
+	out[5] = o6;
+	return;
 
 	v = p->pc;
 	switch(o->size) {
diff --git a/src/cmd/5l/l.h b/src/cmd/5l/l.h
index 1e720f3bea5..4f7ec522356 100644
--- a/src/cmd/5l/l.h
+++ b/src/cmd/5l/l.h
@@ -379,7 +379,7 @@ void	addhist(int32, int);
 Prog*	appendp(Prog*);
 void	asmb(void);
 void	asmthumbmap(void);
-void	asmout(Prog*, Optab*);
+void	asmout(Prog*, Optab*, int32*);
 void	thumbasmout(Prog*, Optab*);
 void	asmsym(void);
 int32	atolwhex(char*);
diff --git a/src/cmd/5l/obj.c b/src/cmd/5l/obj.c
index 6392d93ca81..5a508b4f4aa 100644
--- a/src/cmd/5l/obj.c
+++ b/src/cmd/5l/obj.c
@@ -70,7 +70,7 @@ main(int argc, char *argv[])
 {
 	int c, i;
 
-debug['s'] = 1;
+//debug['s'] = 1;  // qemu cannot handle symdat load
 	Binit(&bso, 1, OWRITE);
 	cout = -1;
 	listinit();
diff --git a/src/cmd/5l/span.c b/src/cmd/5l/span.c
index 3a1c35b62d8..048a4768b63 100644
--- a/src/cmd/5l/span.c
+++ b/src/cmd/5l/span.c
@@ -166,11 +166,12 @@ span(void)
 {
 	Prog *p, *op;
 	Optab *o;
-	int m, bflag;
-	int32 c, otxt;
+	int m, bflag, i, v;
+	int32 c, otxt, out[6];
 	int lastthumb = -1;
 	Section *rosect, *sect;
 	Sym *sym;
+	uchar *bp;
 
 	if(debug['v'])
 		Bprint(&bso, "%5.2f span\n", cputime());
@@ -187,11 +188,8 @@ span(void)
 		p = cursym->text;
 		setarch(p);
 		p->pc = c;
+		cursym->value = c;
 
-		if(blitrl && lastthumb != -1 && lastthumb != thumb){	// flush literal pool
-			if(flushpool(op, 0, 1))
-				c = p->pc = scan(op, p, c);
-		}
 		lastthumb = thumb;
 		autosize = p->to.offset + 4;
 		if(p->from.sym != S)
@@ -216,8 +214,6 @@ span(void)
 					c = p->pc = scan(op, p, c);
 			}
 			if(m == 0) {
-				if(p->as == ATEXT) {
-				}
 				diag("zero-width instruction\n%P", p);
 				continue;
 			}
@@ -237,12 +233,13 @@ span(void)
 				flushpool(p, 0, 0);
 			c += m;
 		}
-		if(blitrl && cursym->next == nil){
+		if(blitrl){
 			if(thumb && isbranch(op))
 				pool.extra += brextra(op);
 			if(checkpool(op, 0))
 				c = scan(op, P, c);
 		}
+		cursym->size = c - cursym->value;
 	}
 
 	/*
@@ -257,6 +254,7 @@ span(void)
 		bflag = 0;
 		c = INITTEXT;
 		for(cursym = textp; cursym != nil; cursym = cursym->next) {
+			cursym->value = c;
 			for(p = cursym->text; p != P; p = p->link) {
 				setarch(p);
 				p->pc = c;
@@ -299,6 +297,7 @@ span(void)
 				}
 				c += m;
 			}
+			cursym->size = c - cursym->value;
 		}
 	}
 
@@ -318,6 +317,7 @@ span(void)
 		oop = op = nil;
 		again = 0;
 		for(cursym = textp; cursym != nil; cursym = cursym->next) {
+			cursym->value = c;
 			for(p = cursym->text; p != P; oop = op, op = p, p = p->link) {
 				setarch(p);
 				if(p->pc != c)
@@ -361,6 +361,7 @@ span(void)
 				}
 				c += m;
 			}
+			cursym->size = c - cursym->value;
 		}
 		if(c != lastc || again){
 			lastc = c;
@@ -368,12 +369,39 @@ span(void)
 		}
 	}
 	c = rnd(c, 8);
-
 	xdefine("etext", STEXT, c);
-	for(cursym = textp; cursym != nil; cursym = cursym->next)
-		cursym->value = cursym->text->pc;
 	textsize = c - INITTEXT;
 	
+	/*
+	 * lay out the code.  all the pc-relative code references,
+	 * even cross-function, are resolved now;
+	 * only data references need to be relocated.
+	 * with more work we could leave cross-function
+	 * code references to be relocated too, and then
+	 * perhaps we'd be able to parallelize the span loop above.
+	 */
+	for(cursym = textp; cursym != nil; cursym = cursym->next) {
+		p = cursym->text;
+		setarch(p);
+		autosize = p->to.offset + 4;
+		symgrow(cursym, cursym->size);
+	
+		bp = cursym->p;
+		for(p = p->link; p != P; p = p->link) {
+			pc = p->pc;
+			curp = p;
+			o = oplook(p);
+			asmout(p, o, out);
+			for(i=0; i<o->size/4; i++) {
+				v = out[i];
+				*bp++ = v;
+				*bp++ = v>>8;
+				*bp++ = v>>16;
+				*bp++ = v>>24;
+			}
+		}
+	}
+
 	rosect = segtext.sect->next;
 	if(rosect) {
 		if(INITRND)