sync compilers with Plan 9

remove 1[acl] 2[acl]

13 files changed, 496 insertions, 3271 deletions

diff --git a/utils/5l/Nt.c b/utils/5l/Nt.c
deleted file mode 100644
index 73c6f795..00000000
--- a/utils/5l/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/5l/Plan9.c b/utils/5l/Plan9.c
deleted file mode 100644
index f4cf23f4..00000000
--- a/utils/5l/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/5l/Posix.c b/utils/5l/Posix.c
deleted file mode 100644
index 0da0ee0c..00000000
--- a/utils/5l/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 fails 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/5l/asm.c b/utils/5l/asm.c
index 66a7f0c6..53d51c45 100644
--- a/utils/5l/asm.c
+++ b/utils/5l/asm.c
@@ -43,7 +43,6 @@ asmb(void)
 	seek(cout, OFFSET, 0);
 	pc = INITTEXT;
 	for(p = firstp; p != P; p = p->link) {
-		setarch(p);
 		if(p->as == ATEXT) {
 			curtext = p;
 			autosize = p->to.offset + 4;
@@ -57,16 +56,9 @@ asmb(void)
 		}
 		curp = p;
 		o = oplook(p);	/* could probably avoid this call */
-		if(thumb)
-			thumbasmout(p, o);
-		else
-			asmout(p, o);
+		asmout(p, o);
 		pc += o->size;
 	}
-	while(pc-INITTEXT < textsize) {
-		cput(0);
-		pc++;
-	}
 
 	if(debug['a'])
 		Bprint(&bso, "\n");
@@ -88,10 +80,12 @@ asmb(void)
 	case 1:
 	case 2:
 	case 5:
+	case 7:
 		OFFSET = HEADR+textsize;
 		seek(cout, OFFSET, 0);
 		break;
 	case 3:
+	case 6:	/* no header, padded segments */
 		OFFSET = rnd(HEADR+textsize, 4096);
 		seek(cout, OFFSET, 0);
 		break;
@@ -108,7 +102,6 @@ asmb(void)
 		else
 			datblk(t, datsize-t, 0);
 	}
-	cflush();
 
 	symsize = 0;
 	lcsize = 0;
@@ -128,9 +121,12 @@ asmb(void)
 			seek(cout, OFFSET, 0);
 			break;
 		case 3:
+		case 6:	/* no header, padded segments */
 			OFFSET += rnd(datsize, 4096);
 			seek(cout, OFFSET, 0);
 			break;
+		case 7:
+			break;
 		}
 		if(!debug['s'])
 			asmsym();
@@ -139,8 +135,6 @@ asmb(void)
 		Bflush(&bso);
 		if(!debug['s'])
 			asmlc();
-		if(!debug['s'])
-			asmthumbmap();
 		if(dlm)
 			asmdyn();
 		cflush();
@@ -151,7 +145,6 @@ asmb(void)
 		cflush();
 	}
 
-	curtext = P;
 	if(debug['v'])
 		Bprint(&bso, "%5.2f header\n", cputime());
 	Bflush(&bso);
@@ -159,6 +152,7 @@ asmb(void)
 	seek(cout, OFFSET, 0);
 	switch(HEADTYPE) {
 	case 0:	/* no header */
+	case 6:	/* no header, padded segments */
 		break;
 	case 1:	/* aif for risc os */
 		lputl(0xe1a00000);		/* NOP - decompress code */
@@ -221,16 +215,12 @@ asmb(void)
 		lputl(0xe3300000);		/* nop */
 		lputl(0xe3300000);		/* nop */
 		break;
+	case 7:	/* elf */
+		debug['S'] = 1;			/* symbol table */
+		elf32(ARM, ELFDATA2LSB, 0, nil);
+		break;
 	}
 	cflush();
-	if(debug['c']){
-		print("textsize=%ld\n", textsize);
-		print("datsize=%ld\n", datsize);
-		print("bsssize=%ld\n", bsssize);
-		print("symsize=%ld\n", symsize);
-		print("lcsize=%ld\n", lcsize);
-		print("total=%ld\n", textsize+datsize+bsssize+symsize+lcsize);
-	}
 }
 
 void
@@ -254,16 +244,6 @@ cput(int c)
 		cflush();
 }
 
-/*
-void
-cput(long c)
-{
-	*cbp++ = c;
-	if(--cbc <= 0)
-		cflush();
-}
-*/
-
 void
 wput(long l)
 {
@@ -277,11 +257,11 @@ wput(long l)
 }
 
 void
-hput(long l)
+wputl(long l)
 {
 
-	cbp[0] = l>>8;
-	cbp[1] = l;
+	cbp[0] = l;
+	cbp[1] = l>>8;
 	cbp += 2;
 	cbc -= 2;
 	if(cbc <= 0)
@@ -317,11 +297,24 @@ lputl(long l)
 }
 
 void
+llput(vlong v)
+{
+	lput(v>>32);
+	lput(v);
+}
+
+void
+llputl(vlong v)
+{
+	lputl(v);
+	lputl(v>>32);
+}
+
+void
 cflush(void)
 {
 	int n;
 
-	/* no bug if cbc < 0 since obuf(cbuf) followed by ibuf in buf! */
 	n = sizeof(buf.cbuf) - cbc;
 	if(n)
 		write(cout, buf.cbuf, n);
@@ -463,16 +456,15 @@ asmlc(void)
 	oldpc = INITTEXT;
 	oldlc = 0;
 	for(p = firstp; p != P; p = p->link) {
-		setarch(p);
 		if(p->line == oldlc || p->as == ATEXT || p->as == ANOP) {
 			if(p->as == ATEXT)
 				curtext = p;
-			if(debug['L'])
+			if(debug['V'])
 				Bprint(&bso, "%6lux %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) {
@@ -480,7 +472,7 @@ 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++;
@@ -494,7 +486,7 @@ 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);
@@ -509,14 +501,14 @@ 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, "%6lux %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, "%6lux %P\n",
 					p->pc, p);
@@ -529,62 +521,11 @@ asmlc(void)
 		cput(s);
 		lcsize++;
 	}
-	if(debug['v'] || debug['L'])
+	if(debug['v'] || debug['V'])
 		Bprint(&bso, "lcsize = %ld\n", lcsize);
 	Bflush(&bso);
 }
 
-static void
-outt(long f, long l)
-{
-	if(debug['L'])
-		Bprint(&bso, "tmap: %lux-%lux\n", f, l);
-	lput(f);
-	lput(l);
-}
-
-void
-asmthumbmap(void)
-{
-	long pc, lastt;
-	Prog *p;
-
-	if(!seenthumb)
-		return;
-	pc = 0;
-	lastt = -1;
-	for(p = firstp; p != P; p = p->link){
-		pc = p->pc - INITTEXT;
-		if(p->as == ATEXT){
-			setarch(p);
-			if(thumb){
-				if(p->from.sym->foreign){	// 8 bytes of ARM first
-					if(lastt >= 0){
-						outt(lastt, pc-1);
-						lastt = -1;
-					}
-					pc += 8;
-				}
-				if(lastt < 0)
-					lastt = pc;
-			}
-			else{
-				if(p->from.sym->foreign){	// 4 bytes of THUMB first
-					if(lastt < 0)
-						lastt = pc;
-					pc += 4;
-				}
-				if(lastt >= 0){
-					outt(lastt, pc-1);
-					lastt = -1;
-				}
-			}
-		}
-	}
-	if(lastt >= 0)
-		outt(lastt, pc+1);
-}
-
 void
 datblk(long s, long n, int str)
 {
@@ -661,25 +602,14 @@ datblk(long s, long n, int str)
 				switch(v->type) {
 				case SUNDEF:
 					ckoff(v, d);
-					d += v->value;
-					break;
 				case STEXT:
 				case SLEAF:
-					d += v->value;
-#ifdef CALLEEBX
-					d += fnpinc(v);
-#else
-					if(v->thumb)
-						d++;		// T bit
-#endif
-					break;
 				case SSTRING:
-					d += v->value;
+					d += p->to.sym->value;
 					break;
 				case SDATA:
 				case SBSS:
-					d += v->value + INITDAT;
-					break;
+					d += p->to.sym->value + INITDAT;
 				}
 				if(dlm)
 					dynreloc(v, a+INITDAT, 1);
@@ -728,8 +658,6 @@ PP = p;
 	o4 = 0;
 	o5 = 0;
 	o6 = 0;
-	armsize += o->size;
-if(debug['P']) print("%ulx: %P	type %d\n", (ulong)(p->pc), p, o->type);
 	switch(o->type) {
 	default:
 		diag("unknown asm %d", o->type);
@@ -737,7 +665,6 @@ if(debug['P']) print("%ulx: %P	type %d\n", (ulong)(p->pc), p, o->type);
 		break;
 
 	case 0:		/* pseudo ops */
-if(debug['G']) print("%ulx: %s: arm %d %d %d %d\n", (ulong)(p->pc), p->from.sym->name, p->from.sym->thumb, p->from.sym->foreign, p->from.sym->fnptr, p->from.sym->used);
 		break;
 
 	case 1:		/* op R,[R],R */
@@ -807,10 +734,6 @@ if(debug['G']) print("%ulx: %s: arm %d %d %d %d\n", (ulong)(p->pc), p->from.sym-
 		}
 		else if(p->cond != P)
 			v = (p->cond->pc - pc) - 8;
-#ifdef CALLEEBX
-		if(p->as == ABL)
-			v += fninc(p->to.sym);
-#endif
 		o1 = opbra(p->as, p->scond);
 		o1 |= (v >> 2) & 0xffffff;
 		break;
@@ -1237,6 +1160,7 @@ if(debug['G']) print("%ulx: %s: arm %d %d %d %d\n", (ulong)(p->pc), p->from.sym-
 			o1 |= rf | (rt<<12);
 		break;
 
+	/* old arm 7500 fp using coproc 1 (1<<8) */
 	case 56:	/* move to FP[CS]R */
 		o1 = ((p->scond & C_SCOND) << 28) | (0xe << 24) | (1<<8) | (1<<4);
 		o1 |= ((p->to.reg+1)<<21) | (p->from.reg << 12);
@@ -1417,49 +1341,58 @@ if(debug['G']) print("%ulx: %s: arm %d %d %d %d\n", (ulong)(p->pc), p->from.sym-
 		else if(p->as == AMOVH)
 			o2 ^= (1<<6);
 		break;
-	case 74:	/* bx $I */
-#ifdef CALLEEBX
-		diag("bx $i case (arm)");
-#endif
-		if(!seenthumb)
-			diag("ABX $I and seenthumb==0");
-		v = p->cond->pc;
-		if(p->to.sym->thumb)
-			v |= 1;	// T bit
-		o1 = olr(8, REGPC, REGTMP, p->scond&C_SCOND);	// mov 8(PC), Rtmp
-		o2 = 	oprrr(AADD, p->scond) | immrot(8) | (REGPC<<16) | (REGLINK<<12);	// add 8,PC, LR
-		o3 = ((p->scond&C_SCOND)<<28) | (0x12fff<<8) | (1<<4) | REGTMP;		// bx Rtmp
-		o4 = opbra(AB, 14);	// B over o6
-		o5 = v;
-		break;
-	case 75:	/* bx O(R) */
-		aclass(&p->to);
-		if(instoffset != 0)
-			diag("non-zero offset in ABX");
-/*
-		o1 = 	oprrr(AADD, p->scond) | immrot(0) | (REGPC<<16) | (REGLINK<<12);	// mov PC, LR
-		o2 = ((p->scond&C_SCOND)<<28) | (0x12fff<<8) | (1<<4) | p->to.reg;		// BX R
-*/
-		// p->to.reg may be REGLINK
-		o1 = oprrr(AADD, p->scond);
-		o1 |= immrot(instoffset);
-		o1 |= p->to.reg << 16;
-		o1 |= REGTMP << 12;
-		o2 = 	oprrr(AADD, p->scond) | immrot(0) | (REGPC<<16) | (REGLINK<<12);	// mov PC, LR
-		o3 = ((p->scond&C_SCOND)<<28) | (0x12fff<<8) | (1<<4) | REGTMP;		// BX Rtmp
+
+	/* VFP ops: */
+	case 74:	/* vfp floating point arith */
+		o1 = opvfprrr(p->as, p->scond);
+		rf = p->from.reg;
+		if(p->from.type == D_FCONST) {
+			diag("invalid floating-point immediate\n%P", p);
+			rf = 0;
+		}
+		rt = p->to.reg;
+		r = p->reg;
+		if(r == NREG)
+			r = rt;
+		o1 |= rt<<12;
+		if(((o1>>20)&0xf) == 0xb)
+			o1 |= rf<<0;
+		else
+			o1 |= r<<16 | rf<<0;
 		break;
-	case 76:	/* bx O(R) when returning from fn*/
-		if(!seenthumb)
-			diag("ABXRET and seenthumb==0");
-		aclass(&p->to);
-// print("ARM BXRET %d(R%d)\n", instoffset, p->to.reg);
-		if(instoffset != 0)
-			diag("non-zero offset in ABXRET");
-		// o1 = olr(instoffset, p->to.reg, REGTMP, p->scond);	// mov O(R), Rtmp
-		o1 = ((p->scond&C_SCOND)<<28) | (0x12fff<<8) | (1<<4) | p->to.reg;		// BX R
+	case 75:	/* vfp floating point compare */
+		o1 = opvfprrr(p->as, p->scond);
+		rf = p->from.reg;
+		if(p->from.type == D_FCONST) {
+			if(p->from.ieee->h != 0 || p->from.ieee->l != 0)
+				diag("invalid floating-point immediate\n%P", p);
+			o1 |= 1<<16;
+			rf = 0;
+		}
+		rt = p->reg;
+		o1 |= rt<<12 | rf<<0;
+		o2 = 0x0ef1fa10;	/* MRS APSR_nzcv, FPSCR */
+		o2 |= (p->scond & C_SCOND) << 28;
+		break;
+	case 76:	/* vfp floating point fix and float */
+		o1 = opvfprrr(p->as, p->scond);
+		rf = p->from.reg;
+		rt = p->to.reg;
+		if(p->from.type == D_REG) {
+			o2 = o1 | rt<<12 | rt<<0;
+			o1 = 0x0e000a10;	/* VMOV F,R */
+			o1 |= (p->scond & C_SCOND) << 28 | rt<<16 | rf<<12;
+		} else {
+			o1 |= FREGTMP<<12 | rf<<0;
+			o2 = 0x0e100a10;	/* VMOV R,F */
+			o2 |= (p->scond & C_SCOND) << 28 | FREGTMP<<16 | rt<<12;
+		}
 		break;
 	}
 
+	if(debug['a'] > 1)
+		Bprint(&bso, "%2d ", o->type);
+
 	v = p->pc;
 	switch(o->size) {
 	default:
@@ -1556,6 +1489,7 @@ oprrr(int a, int sc)
 	case ASRA:	return o | (0xd<<21) | (2<<5);
 	case ASWI:	return o | (0xf<<24);
 
+	/* old arm 7500 fp using coproc 1 (1<<8) */
 	case AADDD:	return o | (0xe<<24) | (0x0<<20) | (1<<8) | (1<<7);
 	case AADDF:	return o | (0xe<<24) | (0x0<<20) | (1<<8);
 	case AMULD:	return o | (0xe<<24) | (0x1<<20) | (1<<8) | (1<<7);
@@ -1583,6 +1517,40 @@ oprrr(int a, int sc)
 }
 
 long
+opvfprrr(int a, int sc)
+{
+	long o;
+
+	o = (sc & C_SCOND) << 28;
+	if(sc & (C_SBIT|C_PBIT|C_WBIT))
+		diag(".S/.P/.W on vfp instruction");
+	o |= 0xe<<24;
+	switch(a) {
+	case AMOVWD:	return o | 0xb<<8 | 0xb<<20 | 1<<6 | 0x8<<16 | 1<<7;
+	case AMOVWF:	return o | 0xa<<8 | 0xb<<20 | 1<<6 | 0x8<<16 | 1<<7;
+	case AMOVDW:	return o | 0xb<<8 | 0xb<<20 | 1<<6 | 0xD<<16 | 1<<7;
+	case AMOVFW:	return o | 0xa<<8 | 0xb<<20 | 1<<6 | 0xD<<16 | 1<<7;
+	case AMOVFD:	return o | 0xa<<8 | 0xb<<20 | 1<<6 | 0x7<<16 | 1<<7;
+	case AMOVDF:	return o | 0xb<<8 | 0xb<<20 | 1<<6 | 0x7<<16 | 1<<7;
+	case AMOVF:	return o | 0xa<<8 | 0xb<<20 | 1<<6 | 0x0<<16 | 0<<7;
+	case AMOVD:	return o | 0xb<<8 | 0xb<<20 | 1<<6 | 0x0<<16 | 0<<7;
+	case ACMPF:	return o | 0xa<<8 | 0xb<<20 | 1<<6 | 0x4<<16 | 0<<7;
+	case ACMPD:	return o | 0xb<<8 | 0xb<<20 | 1<<6 | 0x4<<16 | 0<<7;
+	case AADDF:	return o | 0xa<<8 | 0x3<<20;
+	case AADDD:	return o | 0xb<<8 | 0x3<<20;
+	case ASUBF:	return o | 0xa<<8 | 0x3<<20 | 1<<6;
+	case ASUBD:	return o | 0xb<<8 | 0x3<<20 | 1<<6;
+	case AMULF:	return o | 0xa<<8 | 0x2<<20;
+	case AMULD:	return o | 0xb<<8 | 0x2<<20;
+	case ADIVF:	return o | 0xa<<8 | 0x8<<20;
+	case ADIVD:	return o | 0xb<<8 | 0x8<<20;
+	}
+	diag("bad vfp rrr %d", a);
+	prasm(curp);
+	return 0;
+}
+
+long
 opbra(int a, int sc)
 {
 
@@ -1637,7 +1605,7 @@ olr(long v, int b, int r, int sc)
 		o ^= 1 << 23;
 	}
 	if(v >= (1<<12))
-		diag("literal span too large: %d (R%d)\n%P", v, b, PP);
+		diag("literal span too large: %ld (R%d)\n%P", v, b, PP);
 	o |= v;
 	o |= b << 16;
 	o |= r << 12;
@@ -1662,7 +1630,7 @@ olhr(long v, int b, int r, int sc)
 		o ^= 1 << 23;
 	}
 	if(v >= (1<<8))
-		diag("literal span too large: %d (R%d)\n%P", v, b, PP);
+		diag("literal span too large: %ld (R%d)\n%P", v, b, PP);
 	o |= (v&0xf)|((v>>4)<<8)|(1<<22);
 	o |= b << 16;
 	o |= r << 12;
@@ -1717,10 +1685,45 @@ olhrr(int i, int b, int r, int sc)
 }
 
 long
+ovfpmem(int a, int r, long v, int b, int sc, Prog *p)
+{
+	long o;
+
+	if(sc & (C_SBIT|C_PBIT|C_WBIT))
+		diag(".S/.P/.W on VLDR/VSTR instruction");
+	o = (sc & C_SCOND) << 28;
+	o |= 0xd<<24 | (1<<23);
+	if(v < 0) {
+		v = -v;
+		o ^= 1 << 23;
+	}
+	if(v & 3)
+		diag("odd offset for floating point op: %ld\n%P", v, p);
+	else if(v >= (1<<10))
+		diag("literal span too large: %ld\n%P", v, p);
+	o |= (v>>2) & 0xFF;
+	o |= b << 16;
+	o |= r << 12;
+	switch(a) {
+	default:
+		diag("bad fst %A", a);
+	case AMOVD:
+		o |= 0xb<<8;
+		break;
+	case AMOVF:
+		o |= 0xa<<8;
+		break;
+	}
+	return o;
+}
+
+long
 ofsr(int a, int r, long v, int b, int sc, Prog *p)
 {
 	long o;
 
+	if(vfp)
+		return ovfpmem(a, r, v, b, sc, p);
 	if(sc & C_SBIT)
 		diag(".S on FLDR/FSTR instruction");
 	o = (sc & C_SCOND) << 28;
@@ -1734,9 +1737,9 @@ ofsr(int a, int r, long v, int b, int sc, Prog *p)
 		o ^= 1 << 23;
 	}
 	if(v & 3)
-		diag("odd offset for floating point op: %d\n%P", v, p);
+		diag("odd offset for floating point op: %ld\n%P", v, p);
 	else if(v >= (1<<10))
-		diag("literal span too large: %d\n%P", v, p);
+		diag("literal span too large: %ld\n%P", v, p);
 	o |= (v>>2) & 0xFF;
 	o |= b << 16;
 	o |= r << 12;
@@ -1792,6 +1795,8 @@ chipfloat(Ieee *e)
 	Ieee *p;
 	int n;
 
+	if(vfp)
+		return -1;
 	for(n = sizeof(chipfloats)/sizeof(chipfloats[0]); --n >= 0;){
 		p = &chipfloats[n];
 		if(p->l == e->l && p->h == e->h)
diff --git a/utils/5l/l.h b/utils/5l/l.h
index a7c322be..bd734f1e 100644
--- a/utils/5l/l.h
+++ b/utils/5l/l.h
@@ -1,13 +1,17 @@
 #include	<lib9.h>
 #include	<bio.h>
 #include	"../5c/5.out.h"
+#include	"../8l/elf.h"
 
 #ifndef	EXTERN
 #define	EXTERN	extern
 #endif
 
-/* do not undefine this - code will be removed eventually */
-#define	CALLEEBX
+#define	LIBNAMELEN	300
+
+void	addlibpath(char*);
+int	fileexists(char*);
+char*	findlib(char*);
 
 typedef	struct	Adr	Adr;
 typedef	struct	Sym	Sym;
@@ -17,11 +21,9 @@ typedef	struct	Optab	Optab;
 typedef	struct	Oprang	Oprang;
 typedef	uchar	Opcross[32][2][32];
 typedef	struct	Count	Count;
-typedef	struct	Use	Use;
 
 #define	P		((Prog*)0)
 #define	S		((Sym*)0)
-#define	U		((Use*)0)
 #define	TNAME		(curtext&&curtext->from.sym?curtext->from.sym->name:noname)
 
 struct	Adr
@@ -68,7 +70,6 @@ struct	Prog
 	uchar	as;
 	uchar	scond;
 	uchar	reg;
-	uchar	align;
 };
 #define	regused	u0.u0regused
 #define	forwd	u0.u0forwd
@@ -81,14 +82,10 @@ struct	Sym
 	short	become;
 	short	frame;
 	uchar	subtype;
+	uchar	used;
 	ushort	file;
 	long	value;
 	long	sig;
-	uchar	used;
-	uchar	thumb;	// thumb code
-	uchar	foreign;	// called by arm if thumb, by thumb if arm
-	uchar	fnptr;	// used as fn ptr
-	Use*		use;
 	Sym*	link;
 };
 
@@ -122,12 +119,6 @@ struct	Count
 	long	count;
 	long	outof;
 };
-struct	Use
-{
-	Prog*	p;	/* use */
-	Prog*	ct;	/* curtext */
-	Use*		link;
-};
 
 enum
 {
@@ -141,7 +132,6 @@ enum
 	SCONST,
 	SSTRING,
 	SUNDEF,
-	SREMOVED,
 
 	SIMPORT,
 	SEXPORT,
@@ -150,6 +140,7 @@ enum
 	LTO		= 1<<1,
 	LPOOL		= 1<<2,
 	V4		= 1<<3,	/* arm v4 arch */
+	VFP		= 1<<4,	/* arm vfpv3 floating point */
 
 	C_NONE		= 0,
 	C_REG,
@@ -162,22 +153,17 @@ enum
 	C_RCON,		/* 0xff rotated */
 	C_NCON,		/* ~RCON */
 	C_SCON,		/* 0xffff */
-	C_BCON,		/* thumb */
 	C_LCON,
 	C_FCON,
-	C_GCON,		/* thumb */
 
 	C_RACON,
-	C_SACON,	/* thumb */
 	C_LACON,
-	C_GACON,	/* thumb */
 
 	C_RECON,
 	C_LECON,
 
 	C_SBRA,
 	C_LBRA,
-	C_GBRA,		/* thumb */
 
 	C_HAUTO,	/* halfword insn offset (-0xff to 0xff) */
 	C_FAUTO,	/* float insn offset (0 to 0x3fc, word aligned) */
@@ -198,12 +184,6 @@ enum
 	C_ROREG,
 	C_SROREG,	/* both S and R */
 	C_LOREG,
-	C_GOREG,		/* thumb */
-
-	C_PC,
-	C_SP,
-	C_HREG,
-	C_OFFPC,		/* thumb */
 
 	C_ADDR,		/* relocatable address */
 
@@ -240,9 +220,6 @@ EXTERN union
 #define	cbuf	u.obuf
 #define	xbuf	u.ibuf
 
-#define	setarch(p)		if((p)->as==ATEXT) thumb=(p)->reg&ALLTHUMBS
-#define	setthumb(p)	if((p)->as==ATEXT) seenthumb|=(p)->reg&ALLTHUMBS
-
 #ifndef COFFCVT
 
 EXTERN	long	HEADR;			/* length of header */
@@ -250,6 +227,7 @@ EXTERN	int	HEADTYPE;		/* type of header */
 EXTERN	long	INITDAT;		/* data location */
 EXTERN	long	INITRND;		/* data round above text location */
 EXTERN	long	INITTEXT;		/* text location */
+EXTERN	long	INITTEXTP;		/* text location (physical) */
 EXTERN	char*	INITENTRY;		/* entry point */
 EXTERN	long	autosize;
 EXTERN	Biobuf	bso;
@@ -289,7 +267,6 @@ EXTERN	long	nhunk;
 EXTERN	long	instoffset;
 EXTERN	Opcross	opcross[8];
 EXTERN	Oprang	oprange[ALAST];
-EXTERN	Oprang	thumboprange[ALAST];
 EXTERN	char*	outfile;
 EXTERN	long	pc;
 EXTERN	uchar	repop[ALAST];
@@ -302,9 +279,7 @@ EXTERN	char	xcmp[C_GOK+1][C_GOK+1];
 EXTERN	Prog	zprg;
 EXTERN	int	dtype;
 EXTERN	int	armv4;
-EXTERN	int	thumb;
-EXTERN	int	seenthumb;
-EXTERN	int	armsize;
+EXTERN	int vfp;
 
 EXTERN	int	doexp, dlm;
 EXTERN	int	imports, nimports;
@@ -316,7 +291,6 @@ EXTERN	Prog	undefp;
 
 extern	char*	anames[];
 extern	Optab	optab[];
-extern	Optab	thumboptab[];
 
 void	addpool(Prog*, Adr*);
 EXTERN	Prog*	blitrl;
@@ -329,12 +303,15 @@ EXTERN	Prog*	prog_mod;
 EXTERN	Prog*	prog_modu;
 
 #pragma	varargck	type	"A"	int
+#pragma	varargck	type	"A"	uint
 #pragma	varargck	type	"C"	int
 #pragma	varargck	type	"D"	Adr*
 #pragma	varargck	type	"N"	Adr*
 #pragma	varargck	type	"P"	Prog*
 #pragma	varargck	type	"S"	char*
 
+#pragma	varargck	argpos	diag 1
+
 int	Aconv(Fmt*);
 int	Cconv(Fmt*);
 int	Dconv(Fmt*);
@@ -342,20 +319,17 @@ int	Nconv(Fmt*);
 int	Pconv(Fmt*);
 int	Sconv(Fmt*);
 int	aclass(Adr*);
-int	thumbaclass(Adr*, Prog*);
 void	addhist(long, int);
+void	addlibpath(char*);
 void	append(Prog*, Prog*);
 void	asmb(void);
 void	asmdyn(void);
 void	asmlc(void);
-void	asmthumbmap(void);
 void	asmout(Prog*, Optab*);
-void	thumbasmout(Prog*, Optab*);
 void	asmsym(void);
 long	atolwhex(char*);
 Prog*	brloop(Prog*);
 void	buildop(void);
-void	thumbbuildop(void);
 void	buildrep(int, int);
 void	cflush(void);
 void	ckoff(Sym*, long);
@@ -374,11 +348,12 @@ long	entryvalue(void);
 void	errorexit(void);
 void	exchange(Prog*);
 void	export(void);
+int	fileexists(char*);
 int	find1(long, int);
+char*	findlib(char*);
 void	follow(void);
 void	gethunk(void);
 void	histtoauto(void);
-void	hputl(int);
 double	ieeedtod(Ieee*);
 long	ieeedtof(Ieee*);
 void	import(void);
@@ -388,7 +363,8 @@ void	loadlib(void);
 void	listinit(void);
 Sym*	lookup(char*, int);
 void	cput(int);
-void	hput(long);
+void	llput(vlong);
+void	llputl(vlong);
 void	lput(long);
 void	lputl(long);
 void	mkfwd(void);
@@ -397,10 +373,11 @@ void	names(void);
 void	nocache(Prog*);
 void	nuxiinit(void);
 void	objfile(char*);
-int	ocmp(const void*, const void*);
+int	ocmp(void*, void*);
 long	opirr(int);
 Optab*	oplook(Prog*);
 long	oprrr(int, int);
+long	opvfprrr(int, int);
 long	olr(long, int, int, int);
 long	olhr(long, int, int, int);
 long	olrr(int, int, int, int);
@@ -426,6 +403,7 @@ void	strnput(char*, int);
 void	undef(void);
 void	undefsym(Sym*);
 void	wput(long);
+void	wputl(long);
 void	xdefine(char*, int, long);
 void	xfol(Prog*);
 void	zerosig(char*);
@@ -433,12 +411,5 @@ void	noops(void);
 long	immrot(ulong);
 long	immaddr(long);
 long	opbra(int, int);
-int	brextra(Prog*);
-int	isbranch(Prog*);
-int	fnpinc(Sym *);
-int	fninc(Sym *);
-void	thumbcount(void);
-void reachable(void);
-void fnptrs(void);
 
 #endif
diff --git a/utils/5l/list.c b/utils/5l/list.c
index 5d2e7b49..173f3031 100644
--- a/utils/5l/list.c
+++ b/utils/5l/list.c
@@ -56,14 +56,6 @@ Pconv(Fmt *fp)
 		sprint(str, "(%ld)	%A%C	%D/%d,%D",
 			p->line, a, p->scond, &p->from, p->reg, &p->to);
 		break;
-
-	case AWORD:
-		sprint(str, "WORD %ld", p->to.offset);
-		break;
-
-	case ADWORD:
-		sprint(str, "DWORD %ld %ld", p->from.offset, p->to.offset);
-		break;
 	}
 	return fmtstrcpy(fp, str);
 }
@@ -75,7 +67,7 @@ Aconv(Fmt *fp)
 	int a;
 
 	a = va_arg(fp->args, int);
-	s = "???";
+	s = "?";
 	if(a >= AXXX && a < ALAST)
 		s = anames[a];
 	return fmtstrcpy(fp, s);
diff --git a/utils/5l/mkfile b/utils/5l/mkfile
index 4be81d08..75fa0678 100644
--- a/utils/5l/mkfile
+++ b/utils/5l/mkfile
@@ -12,7 +12,7 @@ OFILES=\
 	span.$O\
 	enam.$O\
 	$TARGMODEL.$O\
-	thumb.$O\
+	elf.$O\
 
 HFILES=\
 	l.h\
@@ -21,7 +21,7 @@ HFILES=\
 
 LIBS=bio 9			# order is important
 
-CFLAGS=$CFLAGS -I../include
+CFLAGS=$CFLAGS -I../include -I.
 
 BIN=$ROOT/$OBJDIR/bin
 
@@ -30,3 +30,8 @@ BIN=$ROOT/$OBJDIR/bin
 enam.$O:	../5c/enam.c
 	$CC $CFLAGS ../5c/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/5l/noop.c b/utils/5l/noop.c
index 08ef6dbe..e68c08f6 100644
--- a/utils/5l/noop.c
+++ b/utils/5l/noop.c
@@ -5,76 +5,11 @@ static	Sym*	sym_divu;
 static	Sym*	sym_mod;
 static	Sym*	sym_modu;
 
-static void setdiv(int);
-
-static Prog *
-movrr(Prog *q, int rs, int rd, Prog *p)
-{
-	if(q == nil)
-		q = prg();
-	q->as = AMOVW;
-	q->line = p->line;
-	q->from.type = D_REG;
-	q->from.reg = rs;
-	q->to.type = D_REG;
-	q->to.reg = rd;
-	q->link = p->link;
-	return q;
-}
-
-static Prog *
-fnret(Prog *q, int rs, int foreign, Prog *p)
-{
-	q = movrr(q, rs, REGPC, p);
-	if(foreign){	// BX rs
-		q->as = ABXRET;
-		q->from.type = D_NONE;
-		q->from.reg = NREG;
-		q->to.reg = rs;
-	}
-	return q;
-}
-
-static Prog *
-aword(long w, Prog *p)
-{
-	Prog *q;
-
-	q = prg();
-	q->as = AWORD;
-	q->line = p->line;
-	q->from.type = D_NONE;
-	q->reg = NREG;
-	q->to.type = D_CONST;
-	q->to.offset = w;
-	q->link = p->link;
-	p->link = q;
-	return q;
-}
-
-static Prog *
-adword(long w1, long w2, Prog *p)
-{
-	Prog *q;
-
-	q = prg();
-	q->as = ADWORD;
-	q->line = p->line;
-	q->from.type = D_CONST;
-	q->from.offset = w1;
-	q->reg = NREG;
-	q->to.type = D_CONST;
-	q->to.offset = w2;
-	q->link = p->link;
-	p->link = q;
-	return q;
-}
-
 void
 noops(void)
 {
-	Prog *p, *q, *q1, *q2;
-	int o, curframe, curbecome, maxbecome, foreign;
+	Prog *p, *q, *q1;
+	int o, curframe, curbecome, maxbecome;
 
 	/*
 	 * find leaf subroutines
@@ -96,7 +31,6 @@ noops(void)
 
 	q = P;
 	for(p = firstp; p != P; p = p->link) {
-		setarch(p);
 
 		/* find out how much arg space is used in this TEXT */
 		if(p->to.type == D_OREG && p->to.reg == REGSP)
@@ -134,7 +68,6 @@ noops(void)
 				initdiv();
 			if(curtext != P)
 				curtext->mark &= ~LEAF;
-			setdiv(p->as);
 			continue;
 
 		case ANOP:
@@ -144,7 +77,6 @@ noops(void)
 			continue;
 
 		case ABL:
-		case ABX:
 			if(curtext != P)
 				curtext->mark &= ~LEAF;
 
@@ -193,13 +125,11 @@ noops(void)
 
 	curtext = 0;
 	for(p = firstp; p != P; p = p->link) {
-		setarch(p);
 		switch(p->as) {
 		case ATEXT:
 			curtext = p;
 			break;
 		case ABL:
-		// case ABX:
 			if(curtext != P && curtext->from.sym != S && curtext->to.offset >= 0) {
 				o = maxbecome - curtext->from.sym->frame;
 				if(o <= 0)
@@ -219,7 +149,6 @@ noops(void)
 	}
 
 	for(p = firstp; p != P; p = p->link) {
-		setarch(p);
 		o = p->as;
 		switch(o) {
 		case ATEXT:
@@ -238,16 +167,7 @@ noops(void)
 				Bflush(&bso);
 				curtext->mark |= LEAF;
 			}
-#ifdef CALLEEBX
-			if(p->from.sym->foreign){
-				if(thumb)
-					// don't allow literal pool to seperate these
-					p = adword(0xe28f7001, 0xe12fff17, p); // arm add 1, pc, r7 and bx r7
-					// p = aword(0xe12fff17, aword(0xe28f7001, p)); // arm add 1, pc, r7 and bx r7
-				else
-					p = aword(0x4778, p);	// thumb bx pc and 2 bytes padding
-			}
-#endif
+
 			if(curtext->mark & LEAF) {
 				if(curtext->from.sym)
 					curtext->from.sym->type = SLEAF;
@@ -271,36 +191,6 @@ noops(void)
 #endif
 			}
 
-			if(thumb){
-				if(!(curtext->mark & LEAF)){
-					q = movrr(nil, REGLINK, REGTMPT-1, p);
-					p->link = q;
-					q1 = prg();
-					q1->as = AMOVW;
-					q1->line = p->line;
-					q1->from.type = D_REG;
-					q1->from.reg = REGTMPT-1;
-					q1->to.type = D_OREG;
-					q1->to.name = D_NONE;
-					q1->to.reg = REGSP;
-					q1->to.offset = 0;
-					q1->link = q->link;
-					q->link = q1;
-				}
-				if(autosize){
-					q2 = prg();
-					q2->as = ASUB;
-					q2->line = p->line;
-					q2->from.type = D_CONST;
-					q2->from.offset = autosize;
-					q2->to.type = D_REG;
-					q2->to.reg = REGSP;
-					q2->link = p->link;
-					p->link = q2;
-				}
-				break;
-			}
-
 			q1 = prg();
 			q1->as = AMOVW;
 			q1->scond |= C_WBIT;
@@ -310,24 +200,18 @@ noops(void)
 			q1->to.type = D_OREG;
 			q1->to.offset = -autosize;
 			q1->to.reg = REGSP;
+
 			q1->link = p->link;
 			p->link = q1;
 			break;
 
 		case ARET:
 			nocache(p);
-			foreign = seenthumb && curtext->from.sym != S && (curtext->from.sym->foreign || curtext->from.sym->fnptr);
-// print("%s %d %d\n", curtext->from.sym->name, curtext->from.sym->foreign, curtext->from.sym->fnptr);
 			if(p->from.type == D_CONST)
 				goto become;
 			if(curtext->mark & LEAF) {
 				if(!autosize) {
-					if(thumb){
-						p = fnret(p, REGLINK, foreign, p);
-						break;
-					}
-// if(foreign) print("ABXRET 1 %s\n", curtext->from.sym->name);
-					p->as = foreign ? ABXRET : AB;
+					p->as = AB;
 					p->from = zprg.from;
 					p->to.type = D_OREG;
 					p->to.offset = 0;
@@ -341,13 +225,9 @@ noops(void)
 				p->from.offset = autosize;
 				p->to.type = D_REG;
 				p->to.reg = REGSP;
-				if(thumb){
-					p->link = fnret(nil, REGLINK, foreign, p);
-					break;
-				}
+
 				q = prg();
-// if(foreign) print("ABXRET 2 %s\n", curtext->from.sym->name);
-				q->as = foreign ? ABXRET : AB;
+				q->as = AB;
 				q->scond = p->scond;
 				q->line = p->line;
 				q->to.type = D_OREG;
@@ -360,100 +240,16 @@ noops(void)
 				break;
 #endif
 			}
-			if(thumb){
-				if(curtext->mark & LEAF){
-					if(autosize){
-						p->as = AADD;
-						p->from.type = D_CONST;
-						p->from.offset = autosize;
-						p->to.type = D_REG;
-						p->to.reg = REGSP;
-						q = nil;
-					}
-					else
-						q = p;
-					q = fnret(q, REGLINK, foreign, p);
-					if(q != p)
-						p->link = q;
-				}
-				else{
-					p->as = AMOVW;
-					p->from.type = D_OREG;
-					p->from.name = D_NONE;
-					p->from.reg = REGSP;
-					p->from.offset = 0;
-					p->to.type = D_REG;
-					p->to.reg = REGTMPT-1;
-					if(autosize){
-						q = prg();
-						q->as = AADD;
-						q->from.type = D_CONST;
-						q->from.offset = autosize;
-						q->to.type = D_REG;
-						q->to.reg = REGSP;
-						q->link = p->link;
-						p->link = 	q;
-					}
-					else
-						q = p;
-					q1 = fnret(nil, REGTMPT-1, foreign, p);
-					q1->link = q->link;
-					q->link = q1;
-				}
-				break;
-			}
-			if(foreign) {
-// if(foreign) print("ABXRET 3 %s\n", curtext->from.sym->name);
-#define	R	1
-				p->as = AMOVW;
-				p->from.type = D_OREG;
-				p->from.name = D_NONE;
-				p->from.reg = REGSP;
-				p->from.offset = 0;
-				p->to.type = D_REG;
-				p->to.reg = R;
-				q = prg();
-				q->as = AADD;
-				q->scond = p->scond;
-				q->line = p->line;
-				q->from.type = D_CONST;
-				q->from.offset = autosize;
-				q->to.type = D_REG;
-				q->to.reg = REGSP;
-				q->link = p->link;
-				p->link = q;
-				q1 = prg();
-				q1->as = ABXRET;
-				q1->scond = p->scond;
-				q1->line = p->line;
-				q1->to.type = D_OREG;
-				q1->to.offset = 0;
-				q1->to.reg = R;
-				q1->link = q->link;
-				q->link = q1;
-#undef	R
-			}
-			else {
-				p->as = AMOVW;
-				p->scond |= C_PBIT;
-				p->from.type = D_OREG;
-				p->from.offset = autosize;
-				p->from.reg = REGSP;
-				p->to.type = D_REG;
-				p->to.reg = REGPC;
-			}
+			p->as = AMOVW;
+			p->scond |= C_PBIT;
+			p->from.type = D_OREG;
+			p->from.offset = autosize;
+			p->from.reg = REGSP;
+			p->to.type = D_REG;
+			p->to.reg = REGPC;
 			break;
 
 		become:
-			if(foreign){
-				diag("foreign become - help");
-				break;
-			}
-			if(thumb){
-				diag("thumb become - help");
-				break;
-			}
-			print("arm become\n");
 			if(curtext->mark & LEAF) {
 
 				if(!autosize) {
@@ -493,29 +289,7 @@ noops(void)
 			q->cond = p->cond;
 			q->link = p->link;
 			p->link = q;
-			if(thumb){
-				q1 = prg();
-				q1->line = p->line;
-				q1->as = AADD;
-				q1->from.type = D_CONST;
-				q1->from.offset = autosize;
-				q1->to.type = D_REG;
-				q1->to.reg = REGSP;
-				p->as = AMOVW;
-				p->line = p->line;
-				p->from.type = D_OREG;
-				p->from.name = D_NONE;
-				p->from.reg = REGSP;
-				p->from.offset = 0;
-				p->to.type = D_REG;
-				p->to.reg = REGTMPT-1;
-				q1->link = q;
-				p->link = q1;
-				q2 = movrr(nil, REGTMPT-1, REGLINK, p);
-				q2->link = q;
-				q1->link = q2;
-				break;
-			}
+
 			p->as = AMOVW;
 			p->scond |= C_PBIT;
 			p->from = zprg.from;
@@ -528,6 +302,30 @@ noops(void)
 
 			break;
 
+		/*
+		 * 5c code generation for unsigned -> double made the
+		 * unfortunate assumption that single and double floating
+		 * point registers are aliased - true for emulated 7500
+		 * but not for vfp.  Now corrected, but this test is
+		 * insurance against old 5c compiled code in libraries.
+		 */
+		case AMOVWD:
+			if((q = p->link) != P && q->as == ACMP)
+			if((q = q->link) != P && q->as == AMOVF)
+			if((q1 = q->link) != P && q1->as == AADDF)
+			if(q1->to.type == D_FREG && q1->to.reg == p->to.reg) {
+				q1->as = AADDD;
+				q1 = prg();
+				q1->scond = q->scond;
+				q1->line = q->line;
+				q1->as = AMOVFD;
+				q1->from = q->to;
+				q1->to = q1->from;
+				q1->link = q->link;
+				q->link = q1;
+			}
+			break;
+
 		case ADIV:
 		case ADIVU:
 		case AMOD:
@@ -567,7 +365,7 @@ noops(void)
 			if(q1->reg == NREG)
 				p->from.reg = q1->to.reg;
 			p->to.type = D_REG;
-			p->to.reg = prog_div != UP && prog_div->from.sym->thumb ? REGTMPT : REGTMP;
+			p->to.reg = REGTMP;
 			p->to.offset = 0;
 
 			/* CALL appropriate */
@@ -576,14 +374,7 @@ noops(void)
 			p->link = q;
 			p = q;
 
-#ifdef CALLEEBX
 			p->as = ABL;
-#else
-			if(prog_div != UP && prog_div->from.sym->thumb)
-				p->as = thumb ? ABL : ABX;
-			else
-				p->as = thumb ? ABX : ABL;
-#endif
 			p->line = q1->line;
 			p->to.type = D_BRANCH;
 			p->cond = p;
@@ -615,7 +406,7 @@ noops(void)
 			p->as = AMOVW;
 			p->line = q1->line;
 			p->from.type = D_REG;
-			p->from.reg = prog_div != UP && prog_div->from.sym->thumb ? REGTMPT : REGTMP;
+			p->from.reg = REGTMP;
 			p->from.offset = 0;
 			p->to.type = D_REG;
 			p->to.reg = q1->to.reg;
@@ -642,145 +433,6 @@ noops(void)
 			q1->reg = NREG;
 			q1->to.type = D_REG;
 			q1->to.reg = REGSP;
-
-			break;
-		case AMOVW:
-			if(thumb){
-				Adr *a = &p->from;
-
-				if(a->type == D_CONST && ((a->name == D_NONE && a->reg == REGSP) || a->name == D_AUTO || a->name == D_PARAM) && (a->offset & 3))
-					diag("SP offset not multiple of 4");
-			}
-			break;
-		case AMOVB:
-		case AMOVBU:
-		case AMOVH:
-		case AMOVHU:
-			if(thumb){
-				if(p->from.type == D_OREG && (p->from.name == D_AUTO || p->from.name == D_PARAM || (p->from.name == D_CONST && p->from.reg == REGSP))){
-					q = prg();
-					*q = *p;
-					if(p->from.name == D_AUTO)
-						q->from.offset += autosize;
-					else if(p->from.name == D_PARAM)
-						q->from.offset += autosize+4;
-					q->from.name = D_NONE;
-					q->from.reg = REGTMPT;
-					p = movrr(p, REGSP, REGTMPT, p);
-					q->link = p->link;
-					p->link = q;
-				}
-				if(p->to.type == D_OREG && (p->to.name == D_AUTO || p->to.name == D_PARAM || (p->to.name == D_CONST && p->to.reg == REGSP))){
-					q = prg();
-					*q = *p;
-					if(p->to.name == D_AUTO)
-						q->to.offset += autosize;
-					else if(p->to.name == D_PARAM)
-						q->to.offset += autosize+4;
-					q->to.name = D_NONE;
-					q->to.reg = REGTMPT;
-					p = movrr(p, REGSP, REGTMPT, p);
-					q->link = p->link;
-					p->link = q;
-					if(q->to.offset < 0 || q->to.offset > 255){	// complicated
-						p->to.reg = REGTMPT+1;			// mov sp, r8
-						q1 = prg();
-						q1->line = p->line;
-						q1->as = AMOVW;
-						q1->from.type = D_CONST;
-						q1->from.offset = q->to.offset;
-						q1->to.type = D_REG;
-						q1->to.reg = REGTMPT;			// mov $o, r7
-						p->link = q1;
-						q1->link = q;
-						q1 = prg();
-						q1->line = p->line;
-						q1->as = AADD;
-						q1->from.type = D_REG;
-						q1->from.reg = REGTMPT+1;
-						q1->to.type = D_REG;
-						q1->to.reg = REGTMPT;			// add r8, r7
-						p->link->link = q1;
-						q1->link = q;
-						q->to.offset = 0;				// mov* r, 0(r7)
-						/* phew */
-					}
-				}
-			}
-			break;
-		case AMOVM:
-			if(thumb){
-				if(p->from.type == D_OREG){
-					if(p->from.offset == 0)
-						p->from.type = D_REG;
-					else
-						diag("non-zero AMOVM offset");
-				}
-				else if(p->to.type == D_OREG){
-					if(p->to.offset == 0)
-						p->to.type = D_REG;
-					else
-						diag("non-zero AMOVM offset");
-				}
-			}
-			break;
-		case AB:
-			if(thumb && p->to.type == D_OREG){
-				if(p->to.offset == 0){
-					p->as = AMOVW;
-					p->from.type = D_REG;
-					p->from.reg = p->to.reg;
-					p->to.type = D_REG;
-					p->to.reg = REGPC;
-				}
-				else{
-					p->as = AADD;
-					p->from.type = D_CONST;
-					p->from.offset = p->to.offset;
-					p->reg = p->to.reg;
-					p->to.type = D_REG;
-					p->to.reg = REGTMPT-1;
-					q = prg();
-					q->as = AMOVW;
-					q->line = p->line;
-					q->from.type = D_REG;
-					q->from.reg = REGTMPT-1;
-					q->to.type = D_REG;
-					q->to.reg = REGPC;
-					q->link = p->link;
-					p->link = q;
-				}
-			}
-			if(seenthumb && !thumb && p->to.type == D_OREG && p->to.reg == REGLINK){	
-				// print("warn %s:	b	(R%d)	assuming a return\n", curtext->from.sym->name, p->to.reg);
-				p->as = ABXRET;
-			}
-			break;
-		case ABL:
-		case ABX:
-			if(thumb && p->to.type == D_OREG){
-				if(p->to.offset == 0){
-					p->as = o;
-					p->from.type = D_NONE;
-					p->to.type = D_REG;
-				}
-				else{
-					p->as = AADD;
-					p->from.type = D_CONST;
-					p->from.offset = p->to.offset;
-					p->reg = p->to.reg;
-					p->to.type = D_REG;
-					p->to.reg = REGTMPT-1;
-					q = prg();
-					q->as = o;
-					q->line = p->line;
-					q->from.type = D_NONE;
-					q->to.type = D_REG;
-					q->to.reg = REGTMPT-1;
-					q->link = p->link;
-					p->link = q;
-				}
-			}
 			break;
 		}
 	}
@@ -870,21 +522,6 @@ initdiv(void)
 	}
 }
 
-static void
-setdiv(int as)
-{
-	Prog *p = nil;
-
-	switch(as){
-	case ADIV: p = prog_div; break;
-	case ADIVU: p = prog_divu; break;
-	case AMOD: p = prog_mod; break;
-	case AMODU: p = prog_modu; break;
-	}
-	if(p != UP && thumb != p->from.sym->thumb)
-		p->from.sym->foreign = 1;
-}
-
 void
 nocache(Prog *p)
 {
diff --git a/utils/5l/obj.c b/utils/5l/obj.c
index 35c5f558..202f8eec 100644
--- a/utils/5l/obj.c
+++ b/utils/5l/obj.c
@@ -11,14 +11,28 @@ char	symname[]	= SYMDEF;
 char	thechar		= '5';
 char	*thestring 	= "arm";
 
+char**	libdir;
+int	nlibdir	= 0;
+static	int	maxlibdir = 0;
+
 /*
+ *	-H0				no header
  *	-H1 -T0x10005000 -R4		is aif for risc os
  *	-H2 -T4128 -R4096		is plan9 format
  *	-H3 -T0xF0000020 -R4		is NetBSD format
  *	-H4				is IXP1200 (raw)
  *	-H5 -T0xC0008010 -R1024		is ipaq
+ *	-H6 -R4096			no header with segments padded to pages
+ *	-H7				is elf
  */
 
+void
+usage(void)
+{
+	diag("usage: %s [-options] objects", argv0);
+	errorexit();
+}
+
 static int
 isobjfile(char *f)
 {
@@ -46,9 +60,9 @@ main(int argc, char *argv[])
 {
 	int c;
 	char *a;
+	char name[LIBNAMELEN];
 
 	Binit(&bso, 1, OWRITE);
-	srand(time(0));
 	cout = -1;
 	listinit();
 	outfile = 0;
@@ -56,6 +70,7 @@ main(int argc, char *argv[])
 	curtext = P;
 	HEADTYPE = -1;
 	INITTEXT = -1;
+	INITTEXTP = -1;
 	INITDAT = -1;
 	INITRND = -1;
 	INITENTRY = 0;
@@ -74,11 +89,19 @@ main(int argc, char *argv[])
 		if(a)
 			INITENTRY = 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)
@@ -102,7 +125,6 @@ main(int argc, char *argv[])
 		break;
 	case 'u':	/* produce dynamically loadable module */
 		dlm = 1;
-		debug['l']++;
 		if(argv[1] != nil && argv[1][0] != '-' && !isobjfile(argv[1]))
 			readundefs(ARGF(), SIMPORT);
 		break;
@@ -110,12 +132,20 @@ main(int argc, char *argv[])
 
 	USED(argc);
 
-	if(*argv == 0) {
-		diag("usage: 5l [-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['U'])
 			HEADTYPE = 0;
@@ -129,6 +159,7 @@ main(int argc, char *argv[])
 		diag("unknown -H option");
 		errorexit();
 	case 0:	/* no header */
+	case 6:	/* no header, padded segments */
 		HEADR = 0L;
 		if(INITTEXT == -1)
 			INITTEXT = 0;
@@ -182,7 +213,18 @@ main(int argc, char *argv[])
 		if(INITRND == -1)
 			INITRND = 1024;
 		break;
+	case 7:	/* elf executable */
+		HEADR = rnd(Ehdr32sz+3*Phdr32sz, 16);
+		if(INITTEXT == -1)
+			INITTEXT = 4096+HEADR;
+		if(INITDAT == -1)
+			INITDAT = 0;
+		if(INITRND == -1)
+			INITRND = 4;
+		break;
 	}
+	if (INITTEXTP == -1)
+		INITTEXTP = INITTEXT;
 	if(INITDAT != 0 && INITRND != 0)
 		print("warning: -D0x%lux is ignored because of -R0x%lux\n",
 			INITDAT, INITRND);
@@ -198,7 +240,6 @@ main(int argc, char *argv[])
 	zprg.from.reg = NREG;
 	zprg.to = zprg.from;
 	buildop();
-	thumbbuildop();	// could build on demand
 	histgen = 0;
 	textp = P;
 	datap = P;
@@ -208,7 +249,7 @@ main(int argc, char *argv[])
 		outfile = "5.out";
 	cout = create(outfile, 1, 0775);
 	if(cout < 0) {
-		diag("%s: cannot create", outfile);
+		diag("cannot create %s: %r", outfile);
 		errorexit();
 	}
 	nuxiinit();
@@ -225,7 +266,7 @@ main(int argc, char *argv[])
 			INITENTRY = "_mainp";
 		if(!debug['l'])
 			lookup(INITENTRY, 0)->type = SXREF;
-	} else
+	} else if(!(*INITENTRY >= '0' && *INITENTRY <= '9'))
 		lookup(INITENTRY, 0)->type = SXREF;
 
 	while(*argv)
@@ -259,11 +300,7 @@ main(int argc, char *argv[])
 			doprof1();
 		else
 			doprof2();
-	if(debug['u'])
-		reachable();
 	dodata();
-	if(seenthumb && debug['f'])
-		fnptrs();
 	follow();
 	if(firstp == P)
 		goto out;
@@ -273,10 +310,6 @@ main(int argc, char *argv[])
 	undef();
 
 out:
-	if(debug['c']){
-		thumbcount();
-		print("ARM size = %d\n", armsize);
-	}
 	if(debug['v']) {
 		Bprint(&bso, "%5.2f cpu time\n", cputime());
 		Bprint(&bso, "%ld memory used\n", thunk);
@@ -288,6 +321,42 @@ out:
 }
 
 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;
@@ -298,7 +367,7 @@ loop:
 	xrefresolv = 0;
 	for(i=0; i<libraryp; i++) {
 		if(debug['v'])
-			Bprint(&bso, "%5.2f autolib: %s (from %s)\n", cputime(), library[i], libraryobj[i]);
+			Bprint(&bso, "%5.2f autolib: %s\n", cputime(), library[i]);
 		objfile(library[i]);
 	}
 	if(xrefresolv)
@@ -312,7 +381,6 @@ void
 errorexit(void)
 {
 
-	Bflush(&bso);
 	if(nerrors) {
 		if(cout >= 0)
 			remove(outfile);
@@ -328,25 +396,26 @@ 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);
@@ -407,7 +476,8 @@ 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)))
@@ -434,7 +504,7 @@ int
 zaddr(uchar *p, Adr *a, Sym *h[])
 {
 	int i, c;
-	long l;
+	int l;
 	Sym *s;
 	Auto *u;
 
@@ -456,17 +526,6 @@ zaddr(uchar *p, Adr *a, Sym *h[])
 		return 0;	/*  force real diagnostic */
 	}
 
-	if(a->type == D_CONST || a->type == D_OCONST) {
-		if(a->name == D_EXTERN || a->name == D_STATIC) {
-			s = a->sym;
-			if(s != S && (s->type == STEXT || s->type == SLEAF || s->type == SCONST || s->type == SXREF)) {
-				if(0 && !s->fnptr && s->name[0] != '.')
-					print("%s used as function pointer\n", s->name);
-				s->fnptr = 1;	// over the top cos of SXREF
-			}
-		}
-	}
-
 	switch(a->type) {
 	default:
 		print("unknown type %d\n", a->type);
@@ -553,25 +612,24 @@ 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++) {
@@ -594,13 +652,25 @@ 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;
@@ -719,8 +789,6 @@ readsome(int f, uchar *buf, uchar *good, uchar *stop, int max)
 	return stop + n;
 }
 
-static void puntfp(Prog *);
-
 void
 ldobj(int f, long c, char *pn)
 {
@@ -955,9 +1023,6 @@ loop:
 		break;
 
 	case ATEXT:
-		setarch(p);
-		setthumb(p);
-		p->align = 4;
 		if(curtext != P) {
 			histtoauto();
 			curtext->to.autom = curauto;
@@ -982,7 +1047,6 @@ loop:
 		}
 		s->type = STEXT;
 		s->value = pc;
-		s->thumb = thumb;
 		lastp->link = p;
 		lastp = p;
 		p->pc = pc;
@@ -1014,31 +1078,7 @@ loop:
 		}
 		goto casedef;
 
-	case AMOVWD:
-	case AMOVWF:
-	case AMOVDW:
-	case AMOVFW:
-	case AMOVFD:
-	case AMOVDF:
-	// case AMOVF:
-	// case AMOVD:
-	case ACMPF:
-	case ACMPD:
-	case AADDF:
-	case AADDD:
-	case ASUBF:
-	case ASUBD:
-	case AMULF:
-	case AMULD:
-	case ADIVF:
-	case ADIVD:
-		if(thumb)
-			puntfp(p);
-		goto casedef;
-
 	case AMOVF:
-		if(thumb)
-			puntfp(p);
 		if(skip)
 			goto casedef;
 
@@ -1068,8 +1108,6 @@ loop:
 		goto casedef;
 
 	case AMOVD:
-		if(thumb)
-			puntfp(p);
 		if(skip)
 			goto casedef;
 
@@ -1130,8 +1168,7 @@ 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)
@@ -1152,8 +1189,6 @@ lookup(char *symb, int v)
 	s->version = v;
 	s->value = 0;
 	s->sig = 0;
-	s->used = s->thumb = s->foreign = s->fnptr = 0;
-	s->use = nil;
 	hash[h] = s;
 	return s;
 }
@@ -1208,7 +1243,6 @@ doprof1(void)
 	s = lookup("__mcount", 0);
 	n = 1;
 	for(p = firstp->link; p != P; p = p->link) {
-		setarch(p);
 		if(p->as == ATEXT) {
 			q = prg();
 			q->line = p->line;
@@ -1235,7 +1269,7 @@ doprof1(void)
 			p->from.sym = s;
 			p->from.offset = n*4 + 4;
 			p->to.type = D_REG;
-			p->to.reg = thumb ? REGTMPT : REGTMP;
+			p->to.reg = REGTMP;
 
 			q = prg();
 			q->line = p->line;
@@ -1247,7 +1281,7 @@ doprof1(void)
 			p->from.type = D_CONST;
 			p->from.offset = 1;
 			p->to.type = D_REG;
-			p->to.reg = thumb ? REGTMPT : REGTMP;
+			p->to.reg = REGTMP;
 
 			q = prg();
 			q->line = p->line;
@@ -1257,7 +1291,7 @@ doprof1(void)
 			p = q;
 			p->as = AMOVW;
 			p->from.type = D_REG;
-			p->from.reg = thumb ? REGTMPT : REGTMP;
+			p->from.reg = REGTMP;
 			p->to.type = D_OREG;
 			p->to.name = D_EXTERN;
 			p->to.sym = s;
@@ -1284,25 +1318,36 @@ doprof1(void)
 	s->value = n*4;
 }
 
+static int brcond[] = {ABEQ, ABNE, ABCS, ABCC, ABMI, ABPL, ABVS, ABVC, ABHI, ABLS, ABGE, ABLT, ABGT, ABLE};
+
 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;
 	}
+
 	ps2 = P;
 	ps4 = P;
 	for(p = firstp; p != P; p = p->link) {
-		setarch(p);
 		if(p->as == ATEXT) {
 			if(p->from.sym == s2) {
 				ps2 = p;
@@ -1315,7 +1360,6 @@ doprof2(void)
 		}
 	}
 	for(p = firstp; p != P; p = p->link) {
-		setarch(p);
 		if(p->as == ATEXT) {
 			if(p->reg & NOPROF) {
 				for(;;) {
@@ -1330,13 +1374,26 @@ doprof2(void)
 			}
 
 			/*
-			 * BL	profin, R2
+			 * BL	profin
 			 */
 			q = prg();
 			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 = AB;
+				q2->to.type = D_BRANCH;
+				q2->to.sym = p->to.sym;
+				q2->cond = q->link;
+			}else
+				p->link = q;
 			p = q;
 			p->as = ABL;
 			p->to.type = D_BRANCH;
@@ -1347,27 +1404,64 @@ 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();
 			q->as = ARET;
 			q->from = p->from;
 			q->to = p->to;
+			q->cond = p->cond;
 			q->link = p->link;
+			q->reg = p->reg;
 			p->link = q;
 
-			/*
-			 * BL	profout
-			 */
-			p->as = ABL;
-			p->from = zprg.from;
-			p->to = zprg.to;
-			p->to.type = D_BRANCH;
-			p->cond = ps4;
-			p->to.sym = s4;
-
-			p = q;
-
+			if(p->scond != 14) {
+				q = prg();
+				q->as = ABL;
+				q->from = zprg.from;
+				q->to = zprg.to;
+				q->to.type = D_BRANCH;
+				q->cond = ps4;
+				q->to.sym = s4;
+				q->link = p->link;
+				p->link = q;
+
+				p->as = brcond[p->scond^1];	/* complement */
+				p->scond = 14;
+				p->from = zprg.from;
+				p->to = zprg.to;
+				p->to.type = D_BRANCH;
+				p->cond = q->link->link;	/* successor of RET */
+				p->to.offset = q->link->link->pc;
+
+				p = q->link->link;
+			} else {
+
+				/*
+				 * BL	profout
+				 */
+				p->as = ABL;
+				p->from = zprg.from;
+				p->to = zprg.to;
+				p->to.type = D_BRANCH;
+				p->cond = ps4;
+				p->to.sym = s4;
+				p->scond = 14;
+
+				p = q;
+			}
 			continue;
 		}
 	}
@@ -1387,7 +1481,7 @@ nuxiinit(void)
 			inuxi1[i] = c;
 		inuxi4[i] = c;
 		fnuxi4[i] = c;
-		if(!debug['d']){
+		if(debug['d'] == 0){
 			fnuxi8[i] = c;
 			fnuxi8[i+4] = c+4;
 		}
@@ -1481,17 +1575,6 @@ ieeedtod(Ieee *ieeep)
 	return ldexp(fr, exp);
 }
 
-static void
-puntfp(Prog *p)
-{
-	USED(p);
-	/* floating point - punt for now */
-	curtext->reg = NREG;	/* ARM */
-	curtext->from.sym->thumb = 0;
-	thumb = 0;
-	// print("%s: generating ARM code (contains floating point ops %d)\n", curtext->from.sym->name, p->line);
-}
-
 void
 undefsym(Sym *s)
 {
@@ -1544,7 +1627,7 @@ readundefs(char *f, int t)
 			diag("%s: bad format", f);
 			errorexit();
 		}
-		for(i = 0; i < n; i++){
+		for(i = 0; i < n; i++) {
 			s = lookup(fields[i], 0);
 			s->type = SXREF;
 			s->subtype = t;
diff --git a/utils/5l/optab.c b/utils/5l/optab.c
index 65ad3447..e9a566fd 100644
--- a/utils/5l/optab.c
+++ b/utils/5l/optab.c
@@ -29,13 +29,10 @@ Optab	optab[] =
 
 	{ AB,		C_NONE,	C_NONE,	C_SBRA,		 5, 4, 0,	LPOOL },
 	{ ABL,		C_NONE,	C_NONE,	C_SBRA,		 5, 4, 0 },
-	{ ABX,		C_NONE,	C_NONE,	C_SBRA,		 74, 20, 0 },
 	{ ABEQ,		C_NONE,	C_NONE,	C_SBRA,		 5, 4, 0 },
 
 	{ AB,		C_NONE,	C_NONE,	C_ROREG,	 6, 4, 0,	LPOOL },
 	{ ABL,		C_NONE,	C_NONE,	C_ROREG,	 7, 8, 0 },
-	{ ABX,		C_NONE,	C_NONE,	C_ROREG,	 75, 12, 0 },
-	{ ABXRET,		C_NONE,	C_NONE,	C_ROREG,	 76, 4, 0 },
 
 	{ ASLL,		C_RCON,	C_REG,	C_REG,		 8, 4, 0 },
 	{ ASLL,		C_RCON,	C_NONE,	C_REG,		 8, 4, 0 },
@@ -47,7 +44,6 @@ Optab	optab[] =
 	{ ASWI,		C_NONE,	C_NONE,	C_LOREG,	10, 4, 0 },
 
 	{ AWORD,	C_NONE,	C_NONE,	C_LCON,		11, 4, 0 },
-	{ AWORD,	C_NONE,	C_NONE,	C_GCON,		11, 4, 0 },
 	{ AWORD,	C_NONE,	C_NONE,	C_LEXT,		11, 4, 0 },
 	{ AWORD,	C_NONE,	C_NONE,	C_ADDR,		11, 4, 0 },
 
@@ -215,6 +211,14 @@ Optab	optab[] =
 	{ ACASE,	C_REG,	C_NONE,	C_NONE,		62, 4, 0 },
 	{ ABCASE,	C_NONE, C_NONE, C_SBRA,		63, 4, 0 },
 
+	{ AADDF,	C_FREG,	C_NONE,	C_FREG,		74, 4, 0, VFP },
+	{ AADDF,	C_FREG,	C_REG,	C_FREG,		74, 4, 0, VFP },
+	{ AMOVF,	C_FREG, C_NONE, C_FREG,		74, 4, 0, VFP },
+	{ ACMPF,	C_FREG,	C_REG,	C_NONE,		75, 8, 0, VFP },
+	{ ACMPF,	C_FCON,	C_REG,	C_NONE,		75, 8, 0, VFP },
+	{ AMOVFW,	C_FREG,	C_NONE,	C_REG,		76, 8, 0, VFP },
+	{ AMOVFW,	C_REG,	C_NONE,	C_FREG,		76, 8, 0, VFP },
+
 	{ AMOVH,	C_REG,	C_NONE,	C_HEXT,		70, 4, REGSB,	V4 },
 	{ AMOVH,	C_REG,	C_NONE, C_HAUTO,	70, 4, REGSP,	V4 },
 	{ AMOVH,	C_REG,	C_NONE,	C_HOREG,	70, 4, 0,	V4 },
diff --git a/utils/5l/pass.c b/utils/5l/pass.c
index e4322df5..8ce7b6d0 100644
--- a/utils/5l/pass.c
+++ b/utils/5l/pass.c
@@ -24,11 +24,6 @@ dodata(void)
 		if(v > s->value)
 			diag("initialize bounds (%ld): %s\n%P",
 				s->value, s->name, p);
-		if((s->type == SBSS || s->type == SDATA) && (p->to.type == D_CONST || p->to.type == D_OCONST) && (p->to.name == D_EXTERN || p->to.name == D_STATIC)){
-			s = p->to.sym;
-			if(s != S && (s->type == STEXT || s->type == SLEAF || s->type == SCONST || s->type == SXREF))
-				s->fnptr = 1;
-		}
 	}
 
 	if(debug['t']) {
@@ -187,7 +182,6 @@ xfol(Prog *p)
 loop:
 	if(p == P)
 		return;
-	setarch(p);
 	a = p->as;
 	if(a == ATEXT)
 		curtext = p;
@@ -260,7 +254,7 @@ loop:
 		return;
 	}
 	if(p->cond != P)
-	if(a != ABL && a != ABX && p->link != P) {
+	if(a != ABL && p->link != P) {
 		q = brchain(p->link);
 		if(a != ATEXT && a != ABCASE)
 		if(q != P && (q->mark&FOLL)) {
@@ -288,7 +282,7 @@ patch(void)
 {
 	long c, vexit;
 	Prog *p, *q;
-	Sym *s, *s1;
+	Sym *s;
 	int a;
 
 	if(debug['v'])
@@ -298,17 +292,10 @@ patch(void)
 	s = lookup("exit", 0);
 	vexit = s->value;
 	for(p = firstp; p != P; p = p->link) {
-		setarch(p);
 		a = p->as;
 		if(a == ATEXT)
 			curtext = p;
-		if(seenthumb && a == ABL){
-			// if((s = p->to.sym) != S && (s1 = curtext->from.sym) != S)
-			//	print("%s calls %s\n", s1->name, s->name);
-			 if((s = p->to.sym) != S && (s1 = curtext->from.sym) != S && s->thumb != s1->thumb)
-				s->foreign = 1;
-		}
-		if((a == ABL || a == ABX || a == AB || a == ARET) &&
+		if((a == ABL || a == AB || a == ARET) &&
 		   p->to.type != D_BRANCH && p->to.sym != S) {
 			s = p->to.sym;
 			switch(s->type) {
@@ -351,21 +338,8 @@ patch(void)
 	}
 
 	for(p = firstp; p != P; p = p->link) {
-		setarch(p);
-		a = p->as;
 		if(p->as == ATEXT)
 			curtext = p;
-		if(seenthumb && a == ABL) {
-#ifdef CALLEEBX
-			if(0)
-				{}
-#else
-			if((s = p->to.sym) != S && (s->foreign || s->fnptr))
-				p->as = ABX;
-#endif
-			else if(p->to.type == D_OREG)
-				p->as = ABX;
-		}
 		if(p->cond != P && p->cond != UP) {
 			p->cond = brloop(p->cond);
 			if(p->cond != P)
@@ -483,327 +457,6 @@ rnd(long v, long r)
 	return v;
 }
 
-#define Reachable(n)	if((s = lookup(n, 0)) != nil) s->used++
-
-static void
-rused(Adr *a)
-{
-	Sym *s = a->sym;
-
-	if(s == S)
-		return;
-	if(a->type == D_OREG || a->type == D_OCONST || a->type == D_CONST){
-		if(a->name == D_EXTERN || a->name == D_STATIC){
-			if(s->used == 0)
-				s->used = 1;
-		}
-	}
-	else if(a->type == D_BRANCH){
-		if(s->used == 0)
-			s->used = 1;
-	}
-}
-
-void
-reachable()
-{
-	Prog *p, *prev, *prevt, *nextt, *q;
-	Sym *s, *s0;
-	int i, todo;
-	char *a;
-
-	Reachable("_div");
-	Reachable("_divu");
-	Reachable("_mod");
-	Reachable("_modu");
-	a = INITENTRY;
-	if(*a >= '0' && *a <= '9')
-		return;
-	s = lookup(a, 0);
-	if(s == nil)
-		return;
-	if(s->type == 0){
-		s->used = 1;	// to stop asm complaining
-		for(p = firstp; p != P && p->as != ATEXT; p = p->link)
-			;
-		if(p == nil)
-			return;
-		s = p->from.sym;
-	}	
-	s->used = 1;
-	do{
-		todo = 0;
-		for(p = firstp; p != P; p = p->link){
-			if(p->as == ATEXT && (s0 = p->from.sym)->used == 1){
-				todo = 1;
-				for(q = p->link; q != P && q->as != ATEXT; q = q->link){
-					rused(&q->from);
-					rused(&q->to);
-				}
-				s0->used = 2;
-			}
-		}
-		for(p = datap; p != P; p = p->link){
-			if((s0 = p->from.sym)->used == 1){
-				todo = 1;
-				for(q = p; q != P; q = q->link){	// data can be scattered
-					if(q->from.sym == s0)
-						rused(&q->to);
-				}
-				s0->used = 2;
-			}
-		}
-	}while(todo);
-	prev = nil;
-	prevt = nextt = nil;
-	for(p = firstp; p != P; ){
-		if(p->as == ATEXT){
-			prevt = nextt;
-			nextt = p;
-		}
-		if(p->as == ATEXT && (s0 = p->from.sym)->used == 0){
-			s0->type = SREMOVED;
-			for(q = p->link; q != P && q->as != ATEXT; q = q->link)
-				;
-			if(q != p->cond)
-				diag("bad ptr in reachable()");
-			if(prev == nil)
-				firstp = q;
-			else
-				prev->link = q;
-			if(q == nil)
-				lastp = prev;
-			if(prevt == nil)
-				textp = q;
-			else
-				prevt->cond = q;
-			if(q == nil)
-				etextp = prevt;
-			nextt = prevt;
-			if(debug['V'])
-				print("%s unused\n", s0->name);
-			p = q;
-		}
-		else{
-			prev = p;
-			p = p->link;
-		}
-	}
-	prevt = nil;
-	for(p = datap; p != nil; ){
-		if((s0 = p->from.sym)->used == 0){
-			s0->type = SREMOVED;
-			prev = prevt;
-			for(q = p; q != nil; q = q->link){
-				if(q->from.sym == s0){
-					if(prev == nil)
-						datap = q->link;
-					else
-						prev->link = q->link;
-				}
-				else
-					prev = q;
-			}
-			if(debug['V'])
-				print("%s unused (data)\n", s0->name);
-			p = prevt->link;
-		}
-		else{
-			prevt = p;
-			p = p->link;
-		}
-	}
-	for(i=0; i<NHASH; i++){
-		for(s = hash[i]; s != S; s = s->link){
-			if(s->used == 0)
-				s->type = SREMOVED;
-		}
-	}
-}
-
-static void
-fused(Adr *a, Prog *p, Prog *ct)
-{
-	Sym *s = a->sym;
-	Use *u;
-
-	if(s == S)
-		return;
-	if(a->type == D_OREG || a->type == D_OCONST || a->type == D_CONST){
-		if(a->name == D_EXTERN || a->name == D_STATIC){
-			u = malloc(sizeof(Use));
-			u->p = p;
-			u->ct = ct;
-			u->link = s->use;
-			s->use = u;
-		}
-	}
-	else if(a->type == D_BRANCH){
-		u = malloc(sizeof(Use));
-		u->p = p;
-		u->ct = ct;
-		u->link = s->use;
-		s->use = u;
-	}
-}
-
-static int
-ckfpuse(Prog *p, Prog *ct, Sym *fp, Sym *r)
-{
-	int reg;
-
-	USED(fp);
-	USED(ct);
-	if(p->from.sym == r && p->as == AMOVW && (p->from.type == D_CONST || p->from.type == D_OREG) && p->reg == NREG && p->to.type == D_REG){
-		reg = p->to.reg;
-		for(p = p->link; p != P && p->as != ATEXT; p = p->link){
-			if((p->as == ABL || p->as == ABX) && p->to.type == D_OREG && p->to.reg == reg)
-				return 1;
-			if(!debug['F'] && (isbranch(p) || p->as == ARET)){
-				// print("%s: branch %P in %s\n", fp->name, p, ct->from.sym->name);
-				return 0;
-			}
-			if((p->from.type == D_REG || p->from.type == D_OREG) && p->from.reg == reg){
-				if(!debug['F'] && p->to.type != D_REG){
-					// print("%s: store %P in %s\n", fp->name, p, ct->from.sym->name);
-					return 0;
-				}
-				reg = p->to.reg;
-			}
-		}
-	}
-	// print("%s: no MOVW O(R), R\n", fp->name);
-	return debug['F'];
-}
-
-static void
-setfpuse(Prog *p, Sym *fp, Sym *r)
-{
-	int reg;
-
-	if(p->from.sym == r && p->as == AMOVW && (p->from.type == D_CONST || p->from.type == D_OREG) && p->reg == NREG && p->to.type == D_REG){
-		reg = p->to.reg;
-		for(p = p->link; p != P && p->as != ATEXT; p = p->link){
-			if((p->as == ABL || p->as == ABX) && p->to.type == D_OREG && p->to.reg == reg){
-				fp->fnptr = 0;
-				p->as = ABL;	// safe to do so
-// print("simplified %s call\n", fp->name);
-				break;
-			}
-			if(!debug['F'] && (isbranch(p) || p->as == ARET))
-				diag("bad setfpuse call");
-			if((p->from.type == D_REG || p->from.type == D_OREG) && p->from.reg == reg){
-				if(!debug['F'] && p->to.type != D_REG)
-					diag("bad setfpuse call");
-				reg = p->to.reg;
-			}
-		}
-	}
-}
-
-static int
-cksymuse(Sym *s, int t)
-{
-	Prog *p;
-
-	for(p = datap; p != P; p = p->link){
-		if(p->from.sym == s && p->to.sym != nil && strcmp(p->to.sym->name, ".string") != 0 && p->to.sym->thumb != t){
-			// print("%s %s %d %d ", p->from.sym->name, p->to.sym->name, p->to.sym->thumb, t);
-			return 0;
-		}
-	}
-	return 1;
-}
-
-/* check the use of s at the given point */
-static int
-ckuse(Sym *s, Sym *s0, Use *u)
-{
-	Sym *s1;
-
-	s1 = u->p->from.sym;
-// print("ckuse %s %s %s\n", s->name, s0->name, s1 ? s1->name : "nil");
-	if(u->ct == nil){	/* in data area */
-		if(s0 == s && !cksymuse(s1, s0->thumb)){
-			// print("%s: cksymuse fails\n", s0->name);
-			return 0;
-		}
-		for(u = s1->use; u != U; u = u->link)
-			if(!ckuse(s1, s0, u))
-				return 0;
-	}
-	else{		/* in text area */
-		if(u->ct->from.sym->thumb != s0->thumb){
-			// print("%s(%d): foreign call %s(%d)\n", s0->name, s0->thumb, u->ct->from.sym->name, u->ct->from.sym->thumb);
-			return 0;
-		}
-		return ckfpuse(u->p, u->ct, s0, s);
-	}
-	return 1;
-}
-		
-static void
-setuse(Sym *s, Sym *s0, Use *u)
-{
-	Sym *s1;
-
-	s1 = u->p->from.sym;
-	if(u->ct == nil){	/* in data area */
-		for(u = s1->use; u != U; u = u->link)
-			setuse(s1, s0, u);
-	}
-	else{		/* in text area */
-		setfpuse(u->p, s0, s);
-	}
-}
-		
-/* detect BX O(R) which can be done as BL O(R) */
-void
-fnptrs()
-{
-	int i;
-	Sym *s;
-	Prog *p;
-	Use *u;
-	
-	for(i=0; i<NHASH; i++){
-		for(s = hash[i]; s != S; s = s->link){
-			if(s->fnptr && (s->type == STEXT || s->type == SLEAF || s->type == SCONST)){
-				// print("%s : fnptr %d %d\n", s->name, s->thumb, s->foreign);
-			}
-		}
-	}	
-	/* record use of syms */
-	for(p = firstp; p != P; p = p->link){
-		if(p->as == ATEXT)
-			curtext = p;
-		else{
-			fused(&p->from, p, curtext);
-			fused(&p->to, p, curtext);
-		}
-	}
-	for(p = datap; p != P; p = p->link)
-		fused(&p->to, p, nil);
-
-	/* now look for fn ptrs */
-	for(i=0; i<NHASH; i++){
-		for(s = hash[i]; s != S; s = s->link){
-			if(s->fnptr && (s->type == STEXT || s->type == SLEAF || s->type == SCONST)){
-				for(u = s->use; u != U; u = u->link){
-					if(!ckuse(s, s, u))
-						break;
-				}
-				if(u == U){		// can simplify
-					for(u = s->use; u != U; u = u->link)
-						setuse(s, s, u);
-				}
-			}	
-		}
-	}
-
-	/*  now free Use structures */
-}
-
 void
 import(void)
 {
diff --git a/utils/5l/span.c b/utils/5l/span.c
index 0b69a96f..d086f592 100644
--- a/utils/5l/span.c
+++ b/utils/5l/span.c
@@ -3,140 +3,19 @@
 static struct {
 	ulong	start;
 	ulong	size;
-	ulong	extra;
 } pool;
 
-int	checkpool(Prog*, int);
-int 	flushpool(Prog*, int, int);
-
-int
-isbranch(Prog *p)
-{
-	int as = p->as;
-	return (as >= ABEQ && as <= ABLE) || as == AB || as == ABL || as == ABX;
-}
-
-static int
-ispad(Prog *p)
-{
-	if(p->as != AMOVW)
-		return 0;
-	if(p->from.type != D_REG || p->from.reg != REGSB)
-		return 0;
-	if(p->to.type != D_REG || p->to.reg != REGSB)
-		return 0;
-	return 1;
-}
-
-int
-fninc(Sym *s)
-{
-	if(thumb){
-		if(s->thumb){
-			if(s->foreign)
-				return 8;
-			else
-				return 0;
-		}
-		else{
-			if(s->foreign)
-				return 0;
-			else
-				diag("T A !foreign in fninc");
-		}
-	}
-	else{
-		if(s->thumb){
-			if(s->foreign)
-				return 0;
-			else
-				diag("A T !foreign in fninc");
-		}
-		else{
-			if(s->foreign)
-				return 4;
-			else
-				return 0;
-		}			
-	}
-	return 0;
-}
-
-int 
-fnpinc(Sym *s)
-{
-	if(!s->fnptr){	// a simplified case BX O(R) -> BL O(R)
-		if(!debug['f'])
-			diag("fnptr == 0 in fnpinc");
-		if(s->foreign)
-			diag("bad usage in fnpinc %s %d %d %d", s->name, s->used, s->foreign, s->thumb);
-		return 0;
-	}
-	/* 0, 1, 2, 3 squared */
-	if(s->thumb)
-		return s->foreign ? 9 : 1;
-	else
-		return s->foreign ? 4 : 0;
-}
-
-static Prog *
-pad(Prog *p, int pc)
-{
-	Prog *q;
-
-	q = prg();
-	q->as = AMOVW;
-	q->line = p->line;
-	q->from.type = D_REG;
-	q->from.reg = REGSB;
-	q->to.type = D_REG;
-	q->to.reg = REGSB;
-	q->pc = pc;
-	q->link = p->link;
-	return q;
-}
-
-static int
-scan(Prog *op, Prog *p, int c)
-{
-	Prog *q;
-
-	for(q = op->link; q != p; q = q->link){
-		q->pc = c;
-		c += oplook(q)->size;
-		nocache(q);
-	}
-	return c;
-}
-
-/* size of a case statement including jump table */
-static long
-casesz(Prog *p)
-{
-	int jt = 0;
-	long n = 0;
-	Optab *o;
-
-	for( ; p != P; p = p->link){
-		if(p->as == ABCASE)
-			jt = 1;
-		else if(jt)
-			break;
-		o = oplook(p);
-		n += o->size;
-	}
-	return n;
-}
+void	checkpool(Prog*);
+void 	flushpool(Prog*, int);
 
 void
 span(void)
 {
-	Prog *p, *op;
+	Prog *p;
 	Sym *setext, *s;
 	Optab *o;
 	int m, bflag, i;
 	long c, otxt, v;
-	int lastthumb = -1;
 
 	if(debug['v'])
 		Bprint(&bso, "%5.2f span\n", cputime());
@@ -144,27 +23,13 @@ span(void)
 
 	bflag = 0;
 	c = INITTEXT;
-	op = nil;
 	otxt = c;
-	for(p = firstp; p != P; op = p, p = p->link) {
-		setarch(p);
+	for(p = firstp; p != P; p = p->link) {
 		p->pc = c;
 		o = oplook(p);
 		m = o->size;
-		// must check literal pool here in case p generates many instructions
-		if(blitrl){
-			if(thumb && isbranch(p))
-				pool.extra += brextra(p);
-			if(checkpool(op, p->as == ACASE ? casesz(p) : m))
-				c = p->pc = scan(op, p, c);
-		}
 		if(m == 0) {
 			if(p->as == ATEXT) {
-				if(blitrl && lastthumb != -1 && lastthumb != thumb){	// flush literal pool
-					if(flushpool(op, 0, 1))
-						c = p->pc = scan(op, p, c);
-				}
-				lastthumb = thumb;
 				curtext = p;
 				autosize = p->to.offset + 4;
 				if(p->from.sym != S)
@@ -173,8 +38,6 @@ span(void)
 				if(c-otxt >= 1L<<17)
 					bflag = 1;
 				otxt = c;
-				if(thumb && blitrl)
-					pool.extra += brextra(p);
 				continue;
 			}
 			diag("zero-width instruction\n%P", p);
@@ -189,17 +52,14 @@ span(void)
 			break;
 		case LPOOL:
 			if ((p->scond&C_SCOND) == 14)
-				flushpool(p, 0, 0);
+				flushpool(p, 0);
 			break;
 		}
 		if(p->as==AMOVW && p->to.type==D_REG && p->to.reg==REGPC && (p->scond&C_SCOND) == 14)
-			flushpool(p, 0, 0);
+			flushpool(p, 0);
 		c += m;
-		if(blitrl && p->link == P){
-			if(thumb && isbranch(p))
-				pool.extra += brextra(p);
-			checkpool(p, 0);
-		}
+		if(blitrl)
+			checkpool(p);
 	}
 
 	/*
@@ -214,10 +74,7 @@ span(void)
 		bflag = 0;
 		c = INITTEXT;
 		for(p = firstp; p != P; p = p->link) {
-			setarch(p);
 			p->pc = c;
-			if(thumb && isbranch(p))
-				nocache(p);
 			o = oplook(p);
 /* very larg branches
 			if(o->type == 6 && p->cond) {
@@ -258,93 +115,6 @@ span(void)
 		}
 	}
 
-	if(seenthumb){		// branch resolution
-		int passes = 0;
-		int lastc = 0;
-		int again;
-		Prog *oop;
-
-	loop:
-		passes++;
-		if(passes > 150){
-			diag("span looping !");
-			errorexit();
-		}
-		c = INITTEXT;
-		oop = op = nil;
-		again = 0;
-		for(p = firstp; p != P; oop = op, op = p, p = p->link){
-			setarch(p);
-			if(p->pc != c)
-				again = 1;
-			p->pc = c;
-			if(thumb && isbranch(p))
-				nocache(p);
-			o = oplook(p);
-			m = o->size;
-			if(passes == 1 && thumb && isbranch(p)){	// start conservative so unneeded alignment is not added
-				if(p->as == ABL)
-					m = 4;
-				else
-					m = 2;
-				p->align = 0;
-			}	
-			if(p->align){
-				if((p->align == 4 && (c&3)) || (p->align == 2 && !(c&3))){
-					if(ispad(op)){
-						oop->link = p;
-						op = oop;
-						c -= 2;
-						p->pc = c;
-					}
-					else{
-						op->link = pad(op, c);
-						op = op->link;
-						c += 2;
-						p->pc = c;
-					}
-					again = 1;
-				}
-			}
-			if(m == 0) {
-				if(p->as == ATEXT) {
-					curtext = p;
-					autosize = p->to.offset + 4;
-					if(p->from.sym != S)
-						p->from.sym->value = c;
-					continue;
-				}
-			}
-			c += m;
-		}
-		if(c != lastc || again){
-			lastc = c;
-			goto loop;
-		}
-	}
-
-	if(0 && seenthumb){		// rm redundant padding - obsolete
-		int d;
-
-		op = nil;
-		d = 0;
-		for(p = firstp; p != P; op = p, p = p->link){
-			p->pc -= d;
-			if(p->as == ATEXT){
-				if(p->from.sym != S)
-					p->from.sym->value -= d;
-// if(p->from.sym != S) print("%s %ux %d %d %d\n", p->from.sym->name ? p->from.sym->name : "?", p->from.sym->value, p->from.sym->thumb, p->from.sym->foreign, p->from.sym->fnptr);
-			}
-			if(ispad(p) && p->link != P && ispad(p->link)){
-				op->link = p->link->link;
-				d += 4;
-				p = op;
-			}
-		}
-		// print("%d bytes removed (padding)\n", d);
-		c -= d;
-	}
-	
 	if(debug['t']) {
 		/* 
 		 * add strings to text segment
@@ -361,7 +131,7 @@ span(void)
 			c += v;
 		}
 	}
-			
+
 	c = rnd(c, 8);
 
 	setext = lookup("etext", 0);
@@ -382,31 +152,24 @@ span(void)
  * drop the pool now, and branch round it.
  * this happens only in extended basic blocks that exceed 4k.
  */
-int
-checkpool(Prog *p, int sz)
+void
+checkpool(Prog *p)
 {
-	if(thumb){
-		if(pool.size >= 0x3fc || (p->pc+sz+pool.extra+2+2)+(pool.size-4)-pool.start-4 >= 0x3fc)
-			return flushpool(p, 1, 0);
-		else if(p->link == P)
-			return flushpool(p, 2, 0);
-		return 0;
-	}
-	if(pool.size >= 0xffc || immaddr((p->pc+sz+4)+4+pool.size - pool.start+8) == 0)
-		return flushpool(p, 1, 0);
+	if(pool.size >= 0xffc || immaddr((p->pc+4)+4+pool.size - pool.start+8) == 0)
+		flushpool(p, 1);
 	else if(p->link == P)
-		return flushpool(p, 2, 0);
-	return 0;
+		flushpool(p, 2);
 }
 
-int
-flushpool(Prog *p, int skip, int force)
+void
+flushpool(Prog *p, int skip)
 {
 	Prog *q;
 
 	if(blitrl) {
 		if(skip){
-			if(0 && skip==1)print("note: flush literal pool at %lux: len=%lud ref=%lux\n", p->pc+4, pool.size, pool.start);
+			if(debug['v'] && skip == 1)
+				print("note: flush literal pool at %lux: len=%lud ref=%lux\n", p->pc+4, pool.size, pool.start);
 			q = prg();
 			q->as = AB;
 			q->to.type = D_BRANCH;
@@ -414,18 +177,15 @@ flushpool(Prog *p, int skip, int force)
 			q->link = blitrl;
 			blitrl = q;
 		}
-		else if(!force && (p->pc+pool.size-pool.start < (thumb ? 0x3fc+4-pool.extra : 2048)))
-			return 0;
+		else if(p->pc+pool.size-pool.start < 2048)
+			return;
 		elitrl->link = p->link;
 		p->link = blitrl;
 		blitrl = 0;	/* BUG: should refer back to values until out-of-range */
 		elitrl = 0;
 		pool.size = 0;
 		pool.start = 0;
-		pool.extra = 0;
-		return 1;
 	}
-	return 0;
 }
 
 void
@@ -434,10 +194,7 @@ addpool(Prog *p, Adr *a)
 	Prog *q, t;
 	int c;
 
-	if(thumb)
-		c = thumbaclass(a, p);
-	else
-		c = aclass(a);
+	c = aclass(a);
 
 	t = zprg;
 	t.as = AWORD;
@@ -447,18 +204,15 @@ addpool(Prog *p, Adr *a)
 		t.to = *a;
 		break;
 
-	case	C_SROREG:
+	case C_SROREG:
 	case C_LOREG:
 	case C_ROREG:
 	case C_FOREG:
 	case C_SOREG:
-	case C_HOREG:
-	case C_GOREG:
 	case C_FAUTO:
 	case C_SAUTO:
 	case C_LAUTO:
 	case C_LACON:
-	case C_GACON:
 		t.to.type = D_CONST;
 		t.to.offset = instoffset;
 		break;
@@ -477,7 +231,6 @@ addpool(Prog *p, Adr *a)
 	if(blitrl == P) {
 		blitrl = q;
 		pool.start = p->pc;
-		q->align = 4;
 	} else
 		elitrl->link = q;
 	elitrl = q;
@@ -678,16 +431,8 @@ aclass(Adr *a)
 				s->type = SDATA;
 			}
 			instoffset = s->value + a->offset + INITDAT;
-			if(s->type == STEXT || s->type == SLEAF || s->type == SUNDEF) {
+			if(s->type == STEXT || s->type == SLEAF || s->type == SUNDEF)
 				instoffset = s->value + a->offset;
-#ifdef CALLEEBX
-				instoffset += fnpinc(s);
-#else
-				if(s->thumb)
-					instoffset++;	// T bit
-#endif
-				return C_LCON;
-			}
 			return C_LCON;
 		}
 		return C_GOK;
@@ -730,12 +475,6 @@ aclass(Adr *a)
 			case SCONST:
 			case SLEAF:
 				instoffset = s->value + a->offset;
-#ifdef CALLEEBX
-				instoffset += fnpinc(s);
-#else
-				if(s->thumb)
-					instoffset++;	// T bit
-#endif
 				return C_LCON;
 			}
 			if(!dlm) {
@@ -773,35 +512,19 @@ oplook(Prog *p)
 	int a1, a2, a3, r;
 	char *c1, *c3;
 	Optab *o, *e;
-	Optab *otab;
-	Oprang *orange;
 
-	if(thumb){
-		otab = thumboptab;
-		orange = thumboprange;
-	}
-	else{
-		otab = optab;
-		orange = oprange;
-	}
 	a1 = p->optab;
 	if(a1)
-		return otab+(a1-1);
+		return optab+(a1-1);
 	a1 = p->from.class;
 	if(a1 == 0) {
-		if(thumb)
-			a1 = thumbaclass(&p->from, p) + 1;
-		else
-			a1 = aclass(&p->from) + 1;
+		a1 = aclass(&p->from) + 1;
 		p->from.class = a1;
 	}
 	a1--;
 	a3 = p->to.class;
 	if(a3 == 0) {
-		if(thumb)
-			a3 = thumbaclass(&p->to, p) + 1;
-		else
-			a3 = aclass(&p->to) + 1;
+		a3 = aclass(&p->to) + 1;
 		p->to.class = a3;
 	}
 	a3--;
@@ -809,35 +532,35 @@ oplook(Prog *p)
 	if(p->reg != NREG)
 		a2 = C_REG;
 	r = p->as;
-	o = orange[r].start;
+	o = oprange[r].start;
 	if(o == 0) {
 		a1 = opcross[repop[r]][a1][a2][a3];
 		if(a1) {
 			p->optab = a1+1;
-			return otab+a1;
+			return optab+a1;
 		}
-		o = orange[r].stop; /* just generate an error */
+		o = oprange[r].stop; /* just generate an error */
 	}
 	if(0) {
 		print("oplook %A %d %d %d\n",
 			(int)p->as, a1, a2, a3);
 		print("		%d %d\n", p->from.type, p->to.type);
 	}
-	e = orange[r].stop;
+	e = oprange[r].stop;
 	c1 = xcmp[a1];
 	c3 = xcmp[a3];
 	for(; o<e; o++)
 		if(o->a2 == a2)
 		if(c1[o->a1])
 		if(c3[o->a3]) {
-			p->optab = (o-otab)+1;
+			p->optab = (o-optab)+1;
 			return o;
 		}
 	diag("illegal combination %A %d %d %d",
 		p->as, a1, a2, a3);
 	prasm(p);
 	if(o == 0)
-		o = otab;
+		o = optab;
 	return o;
 }
 
@@ -898,19 +621,12 @@ cmp(int a, int b)
 		if(b == C_SBRA)
 			return 1;
 		break;
-	case C_GBRA:
-		if(b == C_SBRA || b == C_LBRA)
-			return 1;
-
-	case C_HREG:
-		return cmp(C_SP, b) || cmp(C_PC, b);
-
 	}
 	return 0;
 }
 
 int
-ocmp(const void *a1, const void *a2)
+ocmp(void *a1, void *a2)
 {
 	Optab *p1, *p2;
 	int n;
@@ -923,6 +639,9 @@ ocmp(const void *a1, const void *a2)
 	n = (p2->flag&V4) - (p1->flag&V4);	/* architecture version */
 	if(n)
 		return n;
+	n = (p2->flag&VFP) - (p1->flag&VFP);	/* floating point arch */
+	if(n)
+		return n;
 	n = p1->a1 - p2->a1;
 	if(n)
 		return n;
@@ -941,14 +660,18 @@ buildop(void)
 	int i, n, r;
 
 	armv4 = !debug['h'];
+	vfp = debug['f'];
 	for(i=0; i<C_GOK; i++)
 		for(n=0; n<C_GOK; n++)
 			xcmp[i][n] = cmp(n, i);
-	for(n=0; optab[n].as != AXXX; n++)
+	for(n=0; optab[n].as != AXXX; n++) {
+		if((optab[n].flag & VFP) && !vfp)
+			optab[n].as = AXXX;
 		if((optab[n].flag & V4) && !armv4) {
 			optab[n].as = AXXX;
 			break;
 		}
+	}
 	qsort(optab, n, sizeof(optab[0]), ocmp);
 	for(i=0; i<n; i++) {
 		r = optab[i].as;
@@ -963,6 +686,8 @@ buildop(void)
 		default:
 			diag("unknown op in build: %A", r);
 			errorexit();
+		case AXXX:
+			break;
 		case AADD:
 			oprange[AAND] = oprange[r];
 			oprange[AEOR] = oprange[r];
diff --git a/utils/5l/thumb.c b/utils/5l/thumb.c
deleted file mode 100644
index f16175ea..00000000
--- a/utils/5l/thumb.c
+++ /dev/null
@@ -1,1636 +0,0 @@
-#include "l.h"
-
-static long thumboprr(int);
-static long thumboprrr(int, int);
-static long thumbopirr(int , int);
-static long thumbopri(int);
-static long thumbophh(int);
-static long thumbopbra(int);
-static long thumbopmv(int, int);
-static void lowreg(Prog *, int);
-static void mult(Prog *, int, int);
-static void numr(Prog *, int, int, int);
-static void regis(Prog *, int, int, int);
-static void dis(int, int);
-
-// build a constant using neg, add and shift - only worth it if < 6 bytes */
-static int
-immbuildcon(int c, Prog *p)
-{
-	int n = 0;
-
-	USED(p);
-	if(c >= 0 && c <= 255)
-		return 0;			// mv
-	if(c >= -255 && c < 0)	// mv, neg
-		return 1;
-	if(c >= 256 && c <= 510)	// mv, add
-		return 1;
-	if(c < 0)
-		return 0;
-	while(!(c & 1)){
-		n++;
-		c >>= 1;
-	}
-	if(c >= 0 && c <= 255)	// mv, lsl
-		return 1;
-	return 0;
-}
-
-// positive 5 bit offset from register - O(R)
-// positive 8 bit offset from register - mov O, R then [R, R]
-// otherwise O goes in literal pool - mov O1(PC), R then [R, R]
-static int
-immoreg(int off, Prog *p)
-{
-	int v = 1;
-	int as = p->as;
-
-	if(off < 0)
-		return C_GOREG;
-	if(as == AMOVW)
-		v = 4;
-	else if(as == AMOVH || as == AMOVHU)
-		v = 2;
-	else if(as == AMOVB || as == AMOVBU)
-		v = 1;
-	else
-		diag("bad op in immoreg");
-	if(off/v <= 31)
-		return C_SOREG;
-	if(off <= 255)
-		return C_LOREG;
-	return C_GOREG;
-}
-
-// positive 8 bit - mov O, R then 0(R)
-// otherwise O goes in literal pool - mov O1(PC), R then 0(R)
-static int
-immacon(int off, Prog *p, int t1, int t2)
-{
-	USED(p);
-	if(off < 0)
-		return t2;
-	if(off <= 255)
-		return t1;
-	return t2;
-}
-
-// unsigned 8 bit in words
-static int
-immauto(int off, Prog *p)
-{
-	if(p->as != AMOVW)
-		diag("bad op in immauto");
-	mult(p, off, 4);
-	if(off >= 0 && off <= 1020)
-		return C_SAUTO;
-	return C_LAUTO;
-}
-
-static int
-immsmall(int off, Prog *p, int t1, int t2, int t3)
-{
-	USED(p);
-	if(off >= 0 && off <= 7)
-		return t1;
-	if(off >= 0 && off <= 255)
-		return t2;
-	return t3;
-}
-
-static int
-immcon(int off, Prog *p)
-{
-	int as = p->as;
-
-	if(as == ASLL || as == ASRL || as == ASRA)
-		return C_SCON;
-	if(p->to.type == D_REG && p->to.reg == REGSP){
-		if(as == AADD || as == ASUB){
-			if(off >= 0 && off <= 508)
-				return C_SCON;
-			if(as == ASUB){
-				p->as = AADD;
-				p->from.offset = -p->from.offset;
-			}
-			return C_LCON;
-		}
-		diag("unknown type in immcon");
-	}
-	if(as == AADD || as == ASUB){
-		if(p->reg != NREG)
-			return immsmall(off, p, C_SCON, C_LCON, C_GCON);
-		return immacon(off, p, C_SCON, C_LCON);
-	}
-	if(as == AMOVW && p->from.type == D_CONST && p->to.type == D_REG && immbuildcon(off, p))
-		return C_BCON;
-	if(as == ACMP && p->from.type == D_CONST && immbuildcon(off, p))
-		return C_BCON;
-	if(as == ACMP || as == AMOVW)
-		return immacon(off, p, C_SCON, C_LCON);
-	return C_LCON;
-}
-
-int
-thumbaclass(Adr *a, Prog *p)
-{
-	Sym *s;
-	int t;
-
-	switch(a->type) {
-	case D_NONE:
-		return C_NONE;
-	case D_REG:
-		if(a->reg == REGSP)
-			return C_SP;
-		if(a->reg == REGPC)
-			return C_PC;
-		if(a->reg >= 8)
-			return C_HREG;
-		return C_REG;
-	case D_SHIFT:
-		diag("D_SHIFT in thumbaclass");
-		return C_SHIFT;
-	case D_FREG:
-		diag("D_FREG in thumbaclass");
-		return C_FREG;
-	case D_FPCR:
-		diag("D_FPCR in thumbaclass");
-		return C_FCR;
-	case D_OREG:
-		switch(a->name) {
-		case D_EXTERN:
-		case D_STATIC:
-			if(a->sym == 0 || a->sym->name == 0) {
-				print("null sym external\n");
-				print("%D\n", a);
-				return C_GOK;
-			}
-			t = a->sym->type;
-			if(t == 0 || t == SXREF) {
-				diag("undefined external: %s in %s\n",
-					a->sym->name, TNAME);
-				a->sym->type = SDATA;
-			}
-			instoffset = a->sym->value + a->offset + INITDAT;
-			return C_LEXT;	/* INITDAT unknown at this stage */
-			// return immacon(instoffset, p, C_SEXT, C_LEXT);
-		case D_AUTO:
-			instoffset = autosize + a->offset;
-			return immauto(instoffset, p);
-		case D_PARAM:
-			instoffset = autosize + a->offset + 4L;
-// print("D_PARAM %s %d+%d+%d = %d\n", a->sym != S ? a->sym->name : "noname", autosize, a->offset, 4, autosize+a->offset+4);
-			return immauto(instoffset, p);
-		case D_NONE:
-			instoffset = a->offset;
-			if(a->reg == REGSP)
-				return immauto(instoffset, p);
-			else
-				return immoreg(instoffset, p);
-		}
-		return C_GOK;
-	case D_PSR:
-		diag("D_PSR in thumbaclass");
-		return C_PSR;
-	case D_OCONST:
-		switch(a->name) {
-		case D_EXTERN:
-		case D_STATIC:
-			s = a->sym;
-			t = s->type;
-			if(t == 0 || t == SXREF) {
-				diag("undefined external: %s in %s\n",
-					s->name, TNAME);
-				s->type = SDATA;
-			}
-			instoffset = s->value + a->offset + INITDAT;
-			if(s->type == STEXT || s->type == SLEAF){
-				instoffset = s->value + a->offset;
-#ifdef CALLEEBX
-				instoffset += fnpinc(s);
-#else
-				if(s->thumb)
-					instoffset++;	// T bit
-#endif
-				return C_LCON;
-			}
-			return C_LCON;	/* INITDAT unknown at this stage */
-			// return immcon(instoffset, p);
-		}
-		return C_GOK;
-	case D_FCONST:
-		diag("D_FCONST in thumaclass");
-		return C_FCON;
-	case D_CONST:
-		switch(a->name) {
-		case D_NONE:
-			instoffset = a->offset;
-			if(a->reg != NREG)
-				goto aconsize;
-			return immcon(instoffset, p);
-		case D_EXTERN:
-		case D_STATIC:
-			s = a->sym;
-			if(s == S)
-				break;
-			t = s->type;
-			switch(t) {
-			case 0:
-			case SXREF:
-				diag("undefined external: %s in %s\n",
-					s->name, TNAME);
-				s->type = SDATA;
-				break;
-			case SCONST:
-			case STEXT:
-			case SLEAF:
-				instoffset = s->value + a->offset;
-#ifdef CALLEEBX
-				instoffset += fnpinc(s);
-#else
-				if(s->thumb)
-					instoffset++;	// T bit
-#endif
-				return C_LCON;
-			}
-			instoffset = s->value + a->offset + INITDAT;
-			return C_LCON;	/* INITDAT unknown at this stage */
-			// return immcon(instoffset, p);
-		case D_AUTO:
-			instoffset = autosize + a->offset;
-			goto aconsize;
-		case D_PARAM:
-			instoffset = autosize + a->offset + 4L;
-		aconsize:
-			if(p->from.reg == REGSP || p->from.reg == NREG)
-				return instoffset >= 0 && instoffset < 1024 ? C_SACON : C_GACON;
-			else if(p->from.reg == p->to.reg)
-				return immacon(instoffset, p, C_SACON, C_GACON);
-			return immsmall(instoffset, p, C_SACON, C_LACON, C_GACON);
-		}
-		return C_GOK;
-	case D_BRANCH: {
-		int v, va;
-
-		p->align = 0;
-		v = -4;
-		va = 0;
-		if(p->cond != P){
-			v = (p->cond->pc - p->pc) - 4;
-			va = p->cond->pc;
-		}
-		instoffset = v;
-		if(p->as == AB){
-			if(v >= -2048 && v <= 2046)
-				return C_SBRA;
-			p->align = 4;
-			instoffset = va;
-			return C_LBRA;
-		}
-		if(p->as == ABL){
-#ifdef CALLEEBX
-			int e;
-
-			if((e = fninc(p->to.sym))) {
-				v += e;
-				va += e;
-				instoffset += e;
-			}		
-#endif
-			if(v >= -4194304 && v <= 4194302)
-				return C_SBRA;
-			p->align = 2;
-			instoffset = va;
-			return C_LBRA;
-		}
-		if(p->as == ABX){
-			v = va;
-			if(v >= 0 && v <= 255)
-				return C_SBRA;
-			p->align = 2;
-			instoffset = va;
-			return C_LBRA;
-		}
-		if(v >= -256 && v <= 254)
-			return C_SBRA;
-		if(v >= -(2048-2) && v <= (2046+2))
-			return C_LBRA;
-		p->align = 2;
-		instoffset = va;
-		return C_GBRA;
-	}
-	}
-	return C_GOK;
-}
-
-// as a1 a2 a3 type size param lit vers
-Optab thumboptab[] =
-{
-	{ ATEXT,		C_LEXT,		C_NONE,		C_LCON,		0,	0,	0 },
-	{ ATEXT,		C_LEXT,		C_REG,		C_LCON,		0,	0,	0 },
-	{ AMVN,		C_REG,		C_NONE,		C_REG,		1,	2,	0 },
-	{ ASRL,		C_REG,		C_NONE,		C_REG,		1,	2,	0 },
-	{ ACMP,		C_REG,		C_REG,		C_NONE,		1,	2,	0 },
-	{ ACMN,		C_REG,		C_REG,		C_NONE,		1,	2,	0 },
-	{ AADD,		C_REG,		C_REG,		C_REG,		2,	2,	0 },
-	{ AADD,		C_REG,		C_NONE,		C_REG,		2,	2,	0 },
-	{ AADD,		C_SCON,		C_REG,		C_REG,		3,	2,	0 },
-	{ AADD,		C_LCON,		C_REG,		C_REG,		49,	4,	0 },
-	{ AADD,		C_GCON,		C_REG,		C_REG,		36,	4,	0,	LFROM },
-	// { AADD,		C_LCON,		C_NONE,		C_REG,		3,	2,	0,	LFROM },
-	{ ASRL,		C_SCON,		C_REG,		C_REG,		4,	2,	0 },
-	{ ASRL,		C_SCON,		C_NONE,		C_REG,		4,	2,	0 },
-	{ AADD,		C_SCON,		C_NONE,		C_REG,		5,	2,	0 },
-	{ AADD,		C_LCON,		C_NONE,		C_REG,		37,	4,	0,	LFROM },
-	{ ACMP,		C_SCON,		C_REG,		C_NONE,		5,	2,	0 },
-	{ ACMP,		C_BCON,		C_REG,		C_NONE,		48,	6,	0 },
-	{ ACMP,		C_LCON,		C_REG,		C_NONE,		39,	4,	0,	LFROM },
-	{ AMOVW,		C_SCON,		C_NONE,		C_REG,		5,	2,	0 },
-	{ AMOVW,		C_BCON,		C_NONE,		C_REG,		47,	4,	0 },
-	{ AMOVW,		C_LCON,		C_NONE,		C_REG,		38,	2,	0,	LFROM },
-	// { AADD,		C_LCON,		C_PC,		C_REG,		6,	2,	0,	LFROM },
-	// { AADD,		C_LCON,		C_SP,		C_REG,		6,	2,	0,	LFROM },
-	{ AADD,		C_SCON,		C_NONE,		C_SP,		7,	2,	0 },
-	{ AADD,		C_LCON,		C_NONE,		C_SP,		40,	4,	0,	LFROM },
-	{ AADD,		C_REG,		C_NONE,		C_HREG,		8,	2,	0 },
-	{ AADD,		C_HREG,		C_NONE,		C_REG,		8,	2,	0 },
-	{ AADD,		C_HREG,		C_NONE,		C_HREG,		8,	2,	0 },
-	{ AMOVW,		C_REG,		C_NONE,		C_HREG,		8,	2,	0 },
-	{ AMOVW,		C_HREG,		C_NONE,		C_REG,		8,	2,	0 },
-	{ AMOVW,		C_HREG,		C_NONE,		C_HREG,		8,	2,	0 },
-	{ ACMP,		C_REG,		C_HREG,		C_NONE,		8,	2,	0 },
-	{ ACMP,		C_HREG,		C_REG,		C_NONE,		8,	2,	0 },
-	{ ACMP,		C_HREG,		C_HREG,		C_NONE,		8,	2,	0 },
-	{ AB,			C_NONE,		C_NONE,		C_SBRA,		9,	2,	0,	LPOOL },
-	{ ABEQ,		C_NONE,		C_NONE,		C_SBRA,		10,	2,	0 },
-	{ ABL,		C_NONE,		C_NONE,		C_SBRA,		11,	4,	0 },
-	{ ABX,		C_NONE,		C_NONE,		C_SBRA,		12,	10,	0 },
-	{ AB,			C_NONE,		C_NONE,		C_LBRA,		41,	8,	0,	LPOOL },
-	{ ABEQ,		C_NONE,		C_NONE,		C_LBRA,		46,	4,	0 },
-	{ ABL,		C_NONE,		C_NONE,		C_LBRA,		43,	14,	0 },
-	{ ABX,		C_NONE,		C_NONE,		C_LBRA,		44,	14,	0 },
-	{ ABEQ,		C_NONE,		C_NONE,		C_GBRA,		42,  10, 	0 },
-	// { AB,		C_NONE,		C_NONE,		C_SOREG,		13,	0,	0 },
-	// { ABL,		C_NONE,		C_NONE,		C_SOREG,		14,	0,	0 },
-	{ ABL,		C_NONE,		C_NONE,		C_REG,		51,	4,	0 },
-	{ ABX,		C_NONE,		C_NONE,		C_REG,		15,	8,	0 },
-	{ ABX,		C_NONE,		C_NONE,		C_HREG,		15,	8,	0 },
-	{ ABXRET,		C_NONE,		C_NONE,		C_REG,		45,	2,	0 },
-	{ ABXRET,		C_NONE,		C_NONE,		C_HREG,		45,	2,	0 },
-	{ ASWI,		C_NONE,		C_NONE,		C_LCON,		16,	2,	0 },
-	{ AWORD,		C_NONE,		C_NONE,		C_LCON,		17,	4,	0 },
-	{ AWORD,		C_NONE,		C_NONE,		C_GCON,		17,	4,	0 },
-	{ AWORD,		C_NONE,		C_NONE,		C_LEXT,		17,	4, 	0 },
-	{ ADWORD,	C_LCON,		C_NONE,		C_LCON,		50,	8,	0 },
-	{ AMOVW,		C_SAUTO,		C_NONE,		C_REG,		18,	2,	REGSP },
-	{ AMOVW,		C_LAUTO,		C_NONE,		C_REG,		33,	6,	0,	LFROM  },
-	// { AMOVW,		C_OFFPC,		C_NONE,		C_REG,		18,	2,	REGPC,	LFROM  },
-	{ AMOVW,		C_SEXT,		C_NONE,		C_REG,		30,	4,	0 },
-	{ AMOVW,		C_SOREG,		C_NONE,		C_REG,		19,	2,	0 },
-	{ AMOVHU,	C_SEXT,		C_NONE,		C_REG,		30,	4,	0 },
-	{ AMOVHU,	C_SOREG,		C_NONE,		C_REG,		19,	2,	0 },
-	{ AMOVBU,	C_SEXT,		C_NONE,		C_REG,		30,	4,	0 },
-	{ AMOVBU,	C_SOREG,		C_NONE,		C_REG,		19,	2,	0 },
-	{ AMOVW,		C_REG,		C_NONE,		C_SAUTO,		20,	2,	0 },
-	{ AMOVW,		C_REG,		C_NONE,		C_LAUTO,		34,	6,	0,	LTO },
-	{ AMOVW,		C_REG,		C_NONE,		C_SEXT,		31,	4,	0 },
-	{ AMOVW,		C_REG,		C_NONE,		C_SOREG,		21,	2,	0 },
-	{ AMOVH,		C_REG,		C_NONE,		C_SEXT,		31,	4,	0 },
-	{ AMOVH,		C_REG,		C_NONE,		C_SOREG,		21,	2,	0 },
-	{ AMOVB,		C_REG,		C_NONE,		C_SEXT,		31,	4,	0 },
-	{ AMOVB,		C_REG,		C_NONE,		C_SOREG,		21,	2,	0 },
-	{ AMOVHU,	C_REG,		C_NONE,		C_SEXT,		31,	4,	0 },
-	{ AMOVHU,	C_REG,		C_NONE,		C_SOREG,		21,	2,	0 },
-	{ AMOVBU,	C_REG,		C_NONE,		C_SEXT,		31,	4,	0 },
-	{ AMOVBU,	C_REG,		C_NONE,		C_SOREG,		21,	2,	0 },
-	{ AMOVW,		C_REG,		C_NONE,		C_REG,		22,	2,	0 },
-	{ AMOVB,		C_REG,		C_NONE,		C_REG,		23,	4,	0 },
-	{ AMOVH,		C_REG,		C_NONE,		C_REG,		23,	4,	0 },
-	{ AMOVBU,	C_REG,		C_NONE,		C_REG,		23,	4,	0 },
-	{ AMOVHU,	C_REG,		C_NONE,		C_REG,		23,	4,	0 },
-	{ AMOVH,		C_SEXT,		C_NONE,		C_REG,		32,	6,	0 },
-	{ AMOVH,		C_SOREG,		C_NONE,		C_REG,		24,	4,	0 },
-	{ AMOVB,		C_SEXT,		C_NONE,		C_REG,		32,	6,	0 },
-	{ AMOVB,		C_SOREG,		C_NONE,		C_REG,		24,	4,	0 },
-	{ AMOVW,		C_SACON,	C_NONE,		C_REG,		25,	2,	0 },
-	{ AMOVW,		C_LACON,	C_NONE,		C_REG,		35,	4,	0 },
-	{ AMOVW,		C_GACON,	C_NONE,		C_REG,		35,	4,	0,	LFROM },
-	{ AMOVM,		C_LCON,		C_NONE,		C_REG,		26,	2,	0 },
-	{ AMOVM,		C_REG,		C_NONE,		C_LCON,		27,	2,	0 },
-	{ AMOVW,		C_LOREG,		C_NONE,		C_REG,		28,	4,	0 },
-	{ AMOVH,		C_LOREG,		C_NONE,		C_REG,		28,	4,	0 },
-	{ AMOVB,		C_LOREG,		C_NONE,		C_REG,		28,	4,	0 },
-	{ AMOVHU,	C_LOREG,		C_NONE,		C_REG,		28,	4,	0 },
-	{ AMOVBU,	C_LOREG,		C_NONE,		C_REG,		28,	4,	0 },
-	{ AMOVW,		C_REG,		C_NONE,		C_LOREG,		29,	4,	0 },
-	{ AMOVH,		C_REG,		C_NONE,		C_LOREG,		29,	4,	0 },
-	{ AMOVB,		C_REG,		C_NONE,		C_LOREG,		29,	4,	0 },
-	{ AMOVHU,	C_REG,		C_NONE,		C_LOREG,		29,	4,	0 },
-	{ AMOVBU,	C_REG,		C_NONE,		C_LOREG,		29,	4,	0 },
-	{ AMOVW,		C_GOREG,		C_NONE,		C_REG,		28,	4,	0,	LFROM },
-	{ AMOVH,		C_GOREG,		C_NONE,		C_REG,		28,	4,	0,	LFROM },
-	{ AMOVB,		C_GOREG,		C_NONE,		C_REG,		28,	4,	0,	LFROM },
-	{ AMOVHU,	C_GOREG,		C_NONE,		C_REG,		28,	4,	0,	LFROM },
-	{ AMOVBU,	C_GOREG,		C_NONE,		C_REG,		28,	4,	0,	LFROM },
-	{ AMOVW,		C_REG,		C_NONE,		C_GOREG,		29,	4,	0,	LTO },
-	{ AMOVH,		C_REG,		C_NONE,		C_GOREG,		29,	4,	0,	LTO },
-	{ AMOVB,		C_REG,		C_NONE,		C_GOREG,		29,	4,	0,	LTO },
-	{ AMOVHU,	C_REG,		C_NONE,		C_GOREG,		29,	4,	0,	LTO },
-	{ AMOVBU,	C_REG,		C_NONE,		C_GOREG,		29,	4,	0,	LTO },
-	{ AMOVW,		C_LEXT,		C_NONE,		C_REG,		30,	4,	0,	LFROM },
-	{ AMOVH,		C_LEXT,		C_NONE,		C_REG,		32,	6,	0,	LFROM },
-	{ AMOVB,		C_LEXT,		C_NONE,		C_REG,		32,	6,	0,	LFROM },
-	{ AMOVHU,	C_LEXT,		C_NONE,		C_REG,		30,	4,	0,	LFROM },
-	{ AMOVBU,	C_LEXT,		C_NONE,		C_REG,		30,	4,	0,	LFROM },
-	{ AMOVW,		C_REG,		C_NONE,		C_LEXT,		31,	4,	0,	LTO },
-	{ AMOVH,		C_REG,		C_NONE,		C_LEXT,		31,	4,	0,	LTO },
-	{ AMOVB,		C_REG,		C_NONE,		C_LEXT,		31,	4,	0,	LTO },
-	{ AMOVHU,	C_REG,		C_NONE,		C_LEXT,		31,	4,	0,	LTO },
-	{ AMOVBU,	C_REG,		C_NONE,		C_LEXT,		31,	4,	0,	LTO },
-
-	{ AXXX,		C_NONE,		C_NONE,		C_NONE,		0,	2,	0 },
-};
-
-#define OPCNTSZ	52
-int opcount[OPCNTSZ];
-
-// is this too pessimistic ?
-int
-brextra(Prog *p)
-{
-	int c;
-
-	// +2 is for padding
-	if(p->as == ATEXT)
-		return 0-0+2;
-	if(!isbranch(p))
-		diag("bad op in brextra()");
-	c = thumbaclass(&p->to, p);
-	switch(p->as){
-	case AB:
-		if(c != C_SBRA)
-			return 0;
-		return 8-2+2;
-	case ABL:
-		if(c != C_SBRA)
-			return 0;
-		return 14-4+2;
-	case ABX:
-		if(c == C_REG || c == C_HREG)
-			return 0;
-#ifdef CALLEEBX
-		diag("ABX $I in brextra");
-#endif
-		if(c != C_SBRA)
-			return 0;
-		return 14-10+2;
-	default:
-		if(c == C_GBRA)
-			return 0;
-		if(c == C_LBRA)
-			return 10-4+2;
-		return 10-2+2;
-	}
-}
-
-#define high(r)	((r)>=8)
-
-static long
-mv(Prog *p, int r, int off)
-{
-	int v, o;
-	if(p != nil && p->cond != nil){	// in literal pool
-		v = p->cond->pc - p->pc - 4;
-		if(p->cond->pc & 3)
-			diag("mv: bad literal pool alignment");
-		if(v & 3)
-			v += 2;	// ensure M(4) offset
-		mult(p, v, 4);
-		off = v/4;
-		numr(p, off, 0, 255);
-		o = 0x9<<11;
-	}
-	else{
-		numr(p, off, 0, 255);
-		o = 0x4<<11;
-	}
-	o |= (r<<8) | off;
-	return o;
-}
-
-static void
-mvcon(Prog *p, int r, int c, long *o1, long *o2)
-{
-	int op = 0, n = 0;
-
-	if(c >= 0 && c <= 255)
-		diag("bad c in mvcon");
-	if(c >= -255 && c < 0)	// mv, neg
-		c = -c;
-	else if(c >= 256 && c <= 510){	// mv, add
-		n = rand()%(511-c) + (c-255);
-		c -= n;
-		// n = c-255;
-		// c = 255;
-		op = AADD;
-	}
-	else{
-		if(c < 0)
-			diag("-ve in mvcon");
-		while(!(c & 1)){
-			n++;
-			c >>= 1;
-		}
-		if(c >= 0 && c <= 255)	// mv, lsl
-			op = ASLL;
-		else
-			diag("bad shift in mvcon");
-	}
-	*o1 = mv(p, r, c);
-	switch(op){
-		case 0:
-			*o2 = (1<<14) | (9<<6) | (r<<3) | r;
-			break;
-		case AADD:
-			*o2 = (6<<11) | (r<<8) | n;
-			break;
-		case ASLL:
-			*o2 = (n<<6) | (r<<3) | r;
-			break;
-	}
-}
-
-static long
-mvlh(int rs, int rd)
-{
-	int o = 0x46<<8;
-
-	if(high(rs)){
-		rs -= 8;
-		o |= 1<<6;
-	}
-	if(high(rd)){
-		rd -= 8;
-		o |= 1<<7;
-	}
-	o |= (rs<<3) | rd;
-	return o;
-}
-
-void
-thumbbuildop()
-{
-	int i, n, r;
-	Optab *optab = thumboptab;
-	Oprang *oprange = thumboprange;
-
-	for(n=0; optab[n].as != AXXX; n++)
-		;
-	qsort(optab, n, sizeof(optab[0]), ocmp);
-	for(i=0; i<n; i++) {
-		r = optab[i].as;
-		oprange[r].start = optab+i;
-		while(optab[i].as == r)
-			i++;
-		oprange[r].stop = optab+i;
-		i--;
-
-		switch(r)
-		{
-		default:
-			break;
-		case ABEQ:
-			oprange[ABNE] = oprange[r];
-			oprange[ABCS] = oprange[r];
-			oprange[ABHS] = oprange[r];
-			oprange[ABCC] = oprange[r];
-			oprange[ABLO] = oprange[r];
-			oprange[ABMI] = oprange[r];
-			oprange[ABPL] = oprange[r];
-			oprange[ABVS] = oprange[r];
-			oprange[ABVC] = oprange[r];
-			oprange[ABHI] = oprange[r];
-			oprange[ABLS] = oprange[r];
-			oprange[ABGE] = oprange[r];
-			oprange[ABLT] = oprange[r];
-			oprange[ABGT] = oprange[r];
-			oprange[ABLE] = oprange[r];
-			break;
-		case AMVN:
-			oprange[AADC] = oprange[r];
-			oprange[ASBC] = oprange[r];
-			oprange[AMUL] = oprange[r];
-			oprange[AAND] = oprange[r];
-			oprange[AEOR] = oprange[r];
-			oprange[AORR] = oprange[r];
-			oprange[ABIC] = oprange[r];
-			oprange[AMULU] = oprange[r];
-			break;
-		case ACMN:
-			oprange[ATST] = oprange[r];
-			break;
-		case ASRL:
-			oprange[ASRA] = oprange[r];
-			oprange[ASLL] = oprange[r];
-			break;
-		case AADD:
-			oprange[ASUB] = oprange[r];
-			break;
-		}
-	}
-}
-
-void
-thumbasmout(Prog *p, Optab *o)
-{
-	long o1, o2, o3, o4, o5, o6, o7, v;
-	int r, rf, rt;
-
-	rf = p->from.reg;
-	rt = p->to.reg;
-	r = p->reg;
-	o1 = o2 = o3 = o4 = o5 = o6 = o7 = 0;
-if(debug['P']) print("%ulx: %P	type %d %d\n", (ulong)(p->pc), p, o->type, p->align);
-	opcount[o->type] += o->size;
-	switch(o->type) {
-	default:
-		diag("unknown asm %d", o->type);
-		prasm(p);
-		break;
-	case 0:		/* pseudo ops */
-if(debug['G']) print("%ulx: %s: thumb\n", (ulong)(p->pc), p->from.sym->name);
-		break;
-	case 1:		/* op R, -, R or op R, R, - */
-		o1 = thumboprr(p->as);
-		if(rt == NREG)
-			rt = r;
-		lowreg(p, rf);
-		lowreg(p, rt);
-		o1 |= (0x10<<10) | (rf<<3) | rt;
-		break;
-	case 2:		/* add/sub R, R, R or add/sub R, -, R */
-		o1 = p->as == AADD ? 0x0<<9 : 0x1<<9;
-		if(r == NREG)
-			r = rt;
-		lowreg(p, rf);
-		lowreg(p, r);
-		lowreg(p, rt);
-		o1 |= (0x6<<10) | (rf<<6) | (r<<3) | rt;
-		break;
-	case 3:		/* add/sub $I, R, R or add/sub $I, -, R */
-		thumbaclass(&p->from, p);
-		o1 = p->as == AADD ? 0x0<<9 : 0x1<<9;
-		if(r == NREG)
-			r = rt;
-		numr(p, instoffset, 0, 7);
-		lowreg(p, r);
-		lowreg(p, rt);
-		o1 |= (0x7<<10) | (instoffset<<6) | (r<<3) | rt;
-		break;
-	case 4:		/* shift $I, R, R or shift $I, -, R */
-		thumbaclass(&p->from, p);
-		if(instoffset < 0)
-			diag("negative shift in thumbasmout");
-		instoffset %= 32;
-		o1 = thumbopri(p->as);
-		if(r == NREG)
-			r = rt;
-		numr(p, instoffset, 0, 31);
-		lowreg(p, r);
-		lowreg(p, rt);
-		o1 |= (0x0<<13) | (instoffset<<6) | (r<<3) | rt;
-		break;
-	case 5:		/* add/sub/mov $I, -, R or cmp $I, R, - */
-		thumbaclass(&p->from, p);
-		o1 = thumbopri(p->as);	
-		if(rt == NREG)
-			rt = r;
-		numr(p, instoffset, 0, 255);
-		lowreg(p, rt);
-		o1 |= (0x1<<13) | (rt<<8) | instoffset;
-		break;
-	case 6:		/* add $I, PC/SP, R */
-		if(p->as == ASUB)
-			diag("subtract in add $I, PC/SP, R");
-		thumbaclass(&p->from, p);
-		o1 = r == REGSP ? 0x1<<11 : 0x0<<11;
-		numr(p, instoffset, 0, 255);
-		regis(p, r, REGSP, REGPC);
-		lowreg(p, rt);
-		o1 |= (0xa<<12) | (rt<<8) | instoffset;
-		break;
-	case 7:		/* add, sub $I, SP */
-		thumbaclass(&p->from, p);
-		o1 = p->as == AADD ? 0x0<<7 : 0x1<<7;
-		numr(p, instoffset, 0, 508);
-		mult(p, instoffset, 4);
-		regis(p, rt, REGSP, REGSP);
-		o1 |= (0xb0<<8) | (instoffset>>2);
-		break;
-	case 8:		/* add/mov/cmp R, R where at least 1 reg is high */
-		o1 = 0;
-		if(rt == NREG)
-			rt = r;
-		if(high(rf)){
-			o1 |= 1<<6;
-			rf -= 8;
-		}
-		if(high(rt)){
-			o1 |= 2<<6;
-			rt -= 8;
-		}
-		if(o1 == 0)
-			diag("no high register(%P)", p);
-		o1 |= thumbophh(p->as);
-		o1 |= (0x11<<10) | (rf<<3) | rt;
-		break;
-	case 9:		/* B	$I */
-		thumbaclass(&p->to, p);
-		numr(p, instoffset, -2048, 2046);
-		o1 = (0x1c<<11) | ((instoffset>>1)&0x7ff);
-		break;
-	case 10:		/* Bcc $I */
-		thumbaclass(&p->to, p);
-		numr(p, instoffset, -256, 254);
-		o1 = thumbopbra(p->as);
-		o1 |= (0xd<<12) | ((instoffset>>1)&0xff);
-		break;
-	case 11:		/* BL $I */
-		thumbaclass(&p->to, p);
-		numr(p, instoffset, -4194304, 4194302);
-		o1 = (0x1e<<11) | ((instoffset>>12)&0x7ff);
-		o2 = (0x1f<<11) | ((instoffset>>1)&0x7ff);
-		break;
-	case 12:		/* BX $I */
-#ifdef CALLEEBX
-		diag("BX $I case");
-#endif
-		thumbaclass(&p->to, p);
-		if(p->to.sym->thumb)
-			instoffset  |= 1;	// T bit
-		o1 = mvlh(REGPC, REGTMPT);
-		o2 = (0x6<<11) | (REGTMPT<<8) | 7;	// add 7, RTMP	(T bit + PC offset)
-		o3 = mvlh(REGTMPT, REGLINK);
-		o4 = mv(nil, REGTMPT, instoffset);
-		o5 = (0x11c<<6) | (REGTMPT<<3);
-		// o1 = mv(nil, REGTMPT, v);
-		// o2 = (0x11b<<6) | (REGPC<<3) | REGLINK;
-		// o3 = (0x11c<<6) | (REGTMPT<<3);
-		break;
-	case 13:		/* B O(R)  */
-		diag("B O(R)");
-		break;
-	case 14:		/* BL O(R) */
-		diag("BL O(R)");
-		break;
-	case 15:		/* BX R */
-		o1 = mvlh(REGPC, REGTMPT);
-		o2 = (0x6<<11) | (REGTMPT<<8) | 5;	// add 5, RTMP (T bit + PC offset)
-		o3 = mvlh(REGTMPT, REGLINK);
-		o4 = 0;
-		if(high(rt)){
-			rt -= 8;
-			o4 |= 1<<6;
-		}
-		o4 |= (0x8e<<7) | (rt<<3);
-		// o1 = (0x11c<<6) | (rt<<3);
-		break;
-	case 16:		/* SWI $I */
-		thumbaclass(&p->to, p);
-		numr(p, instoffset, 0, 255);
-		o1 = (0xdf<<8) | instoffset;
-		break;
-	case 17:		/* AWORD */
-		thumbaclass(&p->to, p);
-		o1 = instoffset&0xffff;
-		o2 = (instoffset>>16)&0xffff;
-		break;
-	case 18:		/* AMOVW O(SP), R and AMOVW O(PC), R */
-		thumbaclass(&p->from, p);
-		rf = o->param;
-		o1 = rf == REGSP ? 0x13<<11 : 0x9<<11;
-		regis(p, rf, REGSP, REGPC);
-		lowreg(p, rt);
-		mult(p, instoffset, 4);
-		numr(p, instoffset/4, 0, 255);
-		o1 |= (rt<<8) | (instoffset/4);
-		break;
-	case 19:		/* AMOVW... O(R), R */
-		thumbaclass(&p->from, p);
-		o1 = thumbopmv(p->as, 1);
-		v = 4;
-		if(p->as == AMOVHU)
-			v = 2;
-		else if(p->as == AMOVBU)
-			v = 1;
-		mult(p, instoffset, v);
-		lowreg(p, rf);
-		lowreg(p, rt);
-		numr(p, instoffset/v, 0, 31);
-		o1 |= ((instoffset/v)<<6) | (rf<<3) | rt;
-		break;
-	case 20:		/* AMOVW R, O(SP) */
-		thumbaclass(&p->to, p);
-		o1 = 0x12<<11;
-		if(rt != NREG) regis(p, rt, REGSP, REGSP);
-		lowreg(p, rf);
-		mult(p, instoffset, 4);
-		numr(p, instoffset/4, 0, 255);
-		o1 |= (rf<<8) | (instoffset/4);
-		break;
-	case 21:		/* AMOVW... R, O(R) */
-		thumbaclass(&p->to, p);
-		o1 = thumbopmv(p->as, 0);
-		v = 4;
-		if(p->as == AMOVHU || p->as == AMOVH)
-			v = 2;
-		else if(p->as == AMOVBU || p->as == AMOVB)
-			v = 1;
-		lowreg(p, rf);
-		lowreg(p, rt);
-		mult(p, instoffset, v);
-		numr(p, instoffset/v, 0, 31);
-		o1 |= ((instoffset/v)<<6) | (rt<<3) | rf;
-		break;
-	case 22:		/* AMOVW R, R -> ASLL $0, R, R */
-		o1 = thumbopri(ASLL);
-		lowreg(p, rf);
-		lowreg(p, rt);
-		o1 |= (0x0<<13) | (rf<<3) | rt;
-		break;
-	case 23:		/* AMOVB/AMOVH/AMOVBU/AMOVHU R, R */
-		o1 = thumbopri(ASLL);
-		o2 = p->as == AMOVB || p->as == AMOVH ? thumbopri(ASRA) : thumbopri(ASRL);
-		v = p->as == AMOVB || p->as == AMOVBU ? 24 : 16;
-		lowreg(p, rf);
-		lowreg(p, rt);
-		o1 |= (0x0<<13) | (v<<6) | (rf<<3) | rt;
-		o2 |= (0x0<<13) | (v<<6) | (rt<<3) | rt;
-		break;
-	case 24:	/* AMOVH/AMOVB O(R), R -> AMOVH/AMOVB [R, R], R */
-		thumbaclass(&p->from, p);
-		lowreg(p, rf);
-		lowreg(p, rt);
-		if(rf == rt)
-			r = REGTMPT;
-		else
-			r = rt;
-		if(p->as == AMOVB)
-			numr(p, instoffset, 0, 31);
-		else{
-			mult(p, instoffset, 2);
-			numr(p, instoffset, 0, 62);
-		}
-		o1 = mv(p, r, instoffset);
-		o2 = p->as == AMOVH ? 0x2f<<9 : 0x2b<<9;
-		o2 |= (r<<6) | (rf<<3) | rt;
-		break;
-	case 25:	/* MOVW $sacon, R */
-		thumbaclass(&p->from, p);
-// print("25: %d %d %d %d\n", instoffset, rf, r, rt);
-		if(rf == NREG)
-			rf = REGSP;
-		lowreg(p, rt);
-		if(rf == REGSP){
-			mult(p, instoffset, 4);
-			numr(p, instoffset>>2, 0, 255);
-			o1 = (0x15<<11) | (rt<<8) | (instoffset>>2);	// add $O, SP, R
-		}
-		else if(rf == rt){
-			numr(p, instoffset, 0, 255);
-			o1 = (0x6<<11) | (rt<<8) | instoffset;		// add $O, R
-		}
-		else{
-			lowreg(p, rf);
-			numr(p, instoffset, 0, 7);
-			o1 = (0xe<<9) | (instoffset<<6) | (rf<<3) | rt;	// add $O, Rs, Rd
-		}
-		break;
-	case 26:	/* AMOVM $c, oreg -> stmia */
-		lowreg(p, rt);
-		numr(p, p->from.offset, -256, 255);
-		o1 = (0x18<<11) | (rt<<8) | (p->from.offset&0xff);
-		break;
-	case 27:	/* AMOVM oreg, $c ->ldmia */
-		lowreg(p, rf);
-		numr(p, p->to.offset, -256, 256);
-		o1 = (0x19<<11) | (rf<<8) | (p->to.offset&0xff);
-		break;
-	case 28:	/* AMOV* O(R), R -> AMOV* [R, R], R 	(offset large)	*/
-		thumbaclass(&p->from, p);
-		lowreg(p, rf);
-		lowreg(p, rt);
-		if(rf == rt)
-			r = REGTMPT;
-		else
-			r = rt;
-		o1 = mv(p, r, instoffset);
-		o2 = thumboprrr(p->as, 1);
-		o2 |= (r<<6) | (rf<<3) | rt;
-		break;
-	case 29:	/* AMOV* R, O(R) -> AMOV* R, [R, R]	(offset large)	*/
-		thumbaclass(&p->to, p);
-		lowreg(p, rf);
-		lowreg(p, rt);
-		if(rt == REGTMPT){	// used as tmp reg
-			if(instoffset >= 0 && instoffset <= 255){
-				o1 = (1<<13) | (2<<11) | (rt<<8) | instoffset;	// add $O, R7
-				o2 = thumbopirr(p->as, 0);
-				o2 |= (0<<6) | (rt<<3) | rf;					// mov* R, 0(R)
-			}
-			else
-				diag("big offset - case 29");
-		}
-		else{
-			o1 = mv(p, REGTMPT, instoffset);
-			o2 = thumboprrr(p->as, 0);
-			o2 |= (REGTMPT<<6) | (rt<<3) | rf;
-		}
-		break;
-	case 30:		/* AMOVW... *addr, R */
-		thumbaclass(&p->from, p);
-		o1 = mv(p, rt, instoffset);		// MOV addr, rtmp
-		o2 = thumbopmv(p->as, 1);
-		lowreg(p, rt);
-		o2 |= (rt<<3) | rt;			// MOV* 0(rtmp), R
-		break;
-	case 31:		/* AMOVW... R, *addr */
-		thumbaclass(&p->to, p);
-		o1 = mv(p, REGTMPT, instoffset);
-		o2 = thumbopmv(p->as, 0);
-		lowreg(p, rf);
-		o2 |= (REGTMPT<<3) | rf;
-		break;
-	case 32:	/* AMOVH/AMOVB *addr, R -> AMOVH/AMOVB [R, R], R */
-		thumbaclass(&p->from, p);
-		o1 = mv(p, rt, instoffset);
-		lowreg(p, rt);
-		o2 = mv(nil, REGTMPT, 0);
-		o3 = p->as == AMOVH ? 0x2f<<9 : 0x2b<<9;
-		o3 |= (REGTMPT<<6) | (rt<<3) | rt;
-		break;
-	case 33:	/* AMOVW O(SP), R	(O large) */
-		thumbaclass(&p->from, p);
-		lowreg(p, rt);
-		o1 = mv(p, rt, instoffset);
-		o2 = (0x111<<6) | (REGSP-8)<<3 | rt;	// add SP, rt
-		o3 = thumbopmv(p->as, 1);
-		o3 |= (rt<<3) | rt;
-		break;
-	case 34:	/* AMOVW R, O(SP)	(O large) */
-		thumbaclass(&p->to, p);
-		lowreg(p, rf);
-		o1 = mv(p, REGTMPT, instoffset);
-		o2 = (0x111<<6) | (REGSP-8)<<3 | REGTMPT;	// add SP, REGTMP
-		o3 = thumbopmv(p->as, 0);
-		o3 |= (REGTMPT<<3) | rf;
-		break;
-	case 35:	/* AMOVW $lacon, R */
-		thumbaclass(&p->from, p);
-		lowreg(p, rt);
-		if(rf == NREG)
-			rf = REGSP;
-		if(rf == rt)
-			rf = r = REGTMPT;
-		else
-			r = rt;
-// print("35: io=%d rf=%d rt=%d\n", instoffset, rf, rt);
-		o1 = mv(p, r, instoffset);		// mov O, Rd
-		if(high(rf))
-			o2 = (0x44<<8) | (0x1<<6) | ((rf-8)<<3) | rt;	// add Rs, Rd
-		else
-			o2 = (0x6<<10) | (rf<<6) | (rt<<3) | rt;		// add Rs, Rd
-		break;
-	case 36:	/* AADD/ASUB $i, r, r when $i too big */
-		thumbaclass(&p->from, p);
-		lowreg(p, r);
-		lowreg(p, rt);
-		o1 = mv(p, REGTMPT, instoffset);
-		o2 = p->as == AADD ? 0xc<<9 : 0xd<<9;
-		o2 |= (REGTMPT<<6) | (r<<3) | rt;
-		break;
-	case 37:	/* AADD/ASUB $i, r when $i too big */
-		thumbaclass(&p->from, p);
-		lowreg(p, rt);
-		o1 = mv(p, REGTMPT, instoffset);
-		o2 = p->as == AADD ? 0xc<<9 : 0xd<<9;
-		o2 |= (REGTMPT<<6) | (rt<<3) | rt;
-		break;
-	case 38:	/* AMOVW $i, r when $i too big */
-		thumbaclass(&p->from, p);
-		lowreg(p, rt);
-		o1 = mv(p, rt, instoffset);
-		break;
-	case 39:	/* ACMP $i, r when $i too big */
-		thumbaclass(&p->from, p);
-		lowreg(p, r);
-		o1 = mv(p, REGTMPT, instoffset);
-		o2 = (0x10a<<6) | (REGTMPT<<3) | r;
-		break;
-	case 40:		/* add, sub $I, SP when $I large*/
-		thumbaclass(&p->from, p);
-		if(p->as == ASUB)
-			instoffset = -instoffset;
-		o1 = mv(p, REGTMPT, instoffset);
-		o2 = (0x112<<6) | (REGTMPT<<3) | (REGSP-8);
-		regis(p, rt, REGSP, REGSP);
-		break;
-	case	41:		/* BL LBRA */
-		thumbaclass(&p->to, p);
-		o1 = (0x9<<11) | (REGTMPT<<8);	// mov 0(pc), r7
-		o2 = mvlh(REGTMPT, REGPC);		// mov r7, pc
-		o3 = instoffset&0xffff;			// $lab
-		o4 = (instoffset>>16)&0xffff;
-		break;
-	case 42:		/* Bcc GBRA */
-		thumbaclass(&p->to, p);
-		o1 = (0xd<<12) | thumbopbra(relinv(p->as)) | (6>>1);		// bccnot 
-		// ab lbra
-		o2 = (0x9<<11) | (REGTMPT<<8);	// mov 0(pc), r7
-		o3 = mvlh(REGTMPT, REGPC);		// mov r7, pc
-		o4 = instoffset&0xffff;			// $lab
-		o5 = (instoffset>>16)&0xffff;
-		break;
-	case 43:		/* BL LBRA */
-		thumbaclass(&p->to, p);
-		o1 = mvlh(REGPC, REGTMPT);						// mov pc, r7
-		o2 = (0x6<<11) | (REGTMPT<<8) | 10;				// add 10, r7
-		o3 = mvlh(REGTMPT, REGLINK);					// mov r7, lr
-		o4 = (0x9<<11) | (REGTMPT<<8);					// mov o(pc), r7
-		o5 = mvlh(REGTMPT, REGPC);						// mov r7, pc
-		o6 = instoffset&0xffff;							// $lab
-		o7 = (instoffset>>16)&0xffff;
-		break;
-	case 44:		/* BX LBRA */
-#ifdef CALLEEBX
-		diag("BX LBRA case");
-#endif
-		thumbaclass(&p->to, p);
-		if(p->to.sym->thumb)
-			instoffset  |= 1;	// T bit
-		o1 = mvlh(REGPC, REGTMPT);						// mov pc, r7
-		o2 = (0x6<<11) | (REGTMPT<<8) | 11;				// add 11, r7
-		o3 = mvlh(REGTMPT, REGLINK);					// mov r7, lr
-		o4 = (0x9<<11) | (REGTMPT<<8);					// mov o(pc), r7
-		o5 = (0x11c<<6) | (REGTMPT<<3);					// bx r7
-		o6 = instoffset&0xffff;							// $lab
-		o7 = (instoffset>>16)&0xffff;
-		break;
-	case 45:	/* BX R when returning from fn */
-		o1 = 0;
-		if(high(rt)){
-			rt -= 8;
-			o1 |= 1<<6;
-		}
-		o1 |= (0x8e<<7) | (rt<<3);
-		break;
-	case 46:		/* Bcc LBRA */
-		thumbaclass(&p->to, p);
-		o1 = (0xd<<12) | thumbopbra(relinv(p->as)) | (0>>1);		// bccnot 
-		// ab lbra
-		instoffset -= 2;
-		numr(p, instoffset, -2048, 2046);
-		o2 = (0x1c<<11) | ((instoffset>>1)&0x7ff);
-		break;
-	case 47:	/* mov $i, R where $i can be built */
-		thumbaclass(&p->from, p);
-		mvcon(p, rt, instoffset, &o1, &o2);
-		break;
-	case 48: /* ACMP $i, r when $i built up */
-		thumbaclass(&p->from, p);
-		lowreg(p, r);
-		mvcon(p, REGTMPT, instoffset, &o1, &o2);
-		o3 = (0x10a<<6) | (REGTMPT<<3) | r;
-		break;
-	case 49:	/* AADD $i, r, r when $i is between 0 and 255 - could merge with case 36 */
-		thumbaclass(&p->from, p);
-		lowreg(p, r);
-		lowreg(p, rt);
-		numr(p, instoffset, 0, 255);
-		o1 = mv(p, REGTMPT, instoffset);
-		o2 = p->as == AADD ? 0xc<<9 : 0xd<<9;
-		o2 |= (REGTMPT<<6) | (r<<3) | rt;
-		break;
-	case 50:		/* ADWORD */
-		thumbaclass(&p->from, p);
-		o1 = instoffset&0xffff;
-		o2 = (instoffset>>16)&0xffff;
-		thumbaclass(&p->to, p);
-		o3 = instoffset&0xffff;
-		o4 = (instoffset>>16)&0xffff;
-		break;
-	case 51:	/* BL r */
-		o1 = mvlh(REGPC, REGLINK);	// mov pc, lr
-		o2 = mvlh(rt, REGPC);		// mov r, pc
-		break;
-	}
-
-	v = p->pc;
-	switch(o->size) {
-	default:
-		if(debug['a'])
-			Bprint(&bso, " %.8lux:\t\t%P\n", v, p);
-		break;
-	case 2:
-		if(debug['a'])
-			Bprint(&bso, " %.8lux: %.8lux\t%P\n", v, o1, p);
-		hputl(o1);
-		break;
-	case 4:
-		if(debug['a'])
-			Bprint(&bso, " %.8lux: %.8lux %.8lux\t%P\n", v, o1, o2, p);
-		hputl(o1);
-		hputl(o2);
-		break;
-	case 6:
-		if(debug['a'])
-			Bprint(&bso, "%.8lux: %.8lux %.8lux %.8lux\t%P\n", v, o1, o2, o3, p);
-		hputl(o1);
-		hputl(o2);
-		hputl(o3);
-		break;
-	case 8:
-		if(debug['a'])
-			Bprint(&bso, "%.8lux: %.8lux %.8lux %.8lux %.8lux\t%P\n", v, o1, o2, o3, o4, p);
-		hputl(o1);
-		hputl(o2);
-		hputl(o3);
-		hputl(o4);
-		break;
-	case 10:
-		if(debug['a'])
-			Bprint(&bso, "%.8lux: %.8lux %.8lux %.8lux %.8lux %.8lux\t%P\n", v, o1, o2, o3, o4, o5, p);
-		hputl(o1);
-		hputl(o2);
-		hputl(o3);
-		hputl(o4);
-		hputl(o5);
-		break;
-	case 12:
-		if(debug['a'])
-			Bprint(&bso, "%.8lux: %.8lux %.8lux %.8lux %.8lux %.8lux %.8lux\t%P\n", v, o1, o2, o3, o4, o5, o6, p);
-		hputl(o1);
-		hputl(o2);
-		hputl(o3);
-		hputl(o4);
-		hputl(o5);
-		hputl(o6);
-		break;
-	case 14:
-		if(debug['a'])
-			Bprint(&bso, "%.8lux: %.8lux %.8lux %.8lux %.8lux %.8lux %.8lux %.8lux\t%P\n", v, o1, o2, o3, o4, o5, o6, o7, p);
-		hputl(o1);
-		hputl(o2);
-		hputl(o3);
-		hputl(o4);
-		hputl(o5);
-		hputl(o6);
-		hputl(o7);
-		break;
-	}
-	if(debug['G']){
-		if(o->type == 17){
-			print("%lx:	word %ld\n", p->pc, (o2<<16)+o1);
-			return;
-		}
-		if(o->type == 50){
-			print("%lx:	word %ld\n", p->pc, (o2<<16)+o1);
-			print("%lx:	word %ld\n", p->pc, (o4<<16)+o3);
-			return;
-		}
-		if(o->size > 0) dis(o1, p->pc);
-		if(o->size > 2) dis(o2, p->pc+2);
-		if(o->size > 4) dis(o3, p->pc+4);
-		if(o->size > 6) dis(o4, p->pc+6);
-		if(o->size > 8) dis(o5, p->pc+8);
-		if(o->size > 10) dis(o6, p->pc+10);
-		if(o->size > 12) dis(o7, p->pc+12);
-		// if(o->size > 14) dis(o8, p->pc+14);
-	}
-}
-
-static long
-thumboprr(int a)
-{
-	switch(a) {
-	case AMVN:	return 0xf<<6;
-	case ACMP:	return 0xa<<6;
-	case ACMN:	return 0xb<<6;
-	case ATST:	return 0x8<<6;
-	case AADC:	return 0x5<<6;
-	case ASBC:	return 0x6<<6;
-	case AMUL:
-	case AMULU:	return 0xd<<6;
-	case AAND:	return 0x0<<6;
-	case AEOR:	return 0x1<<6;
-	case AORR:	return 0xc<<6;
-	case ABIC:	return 0xe<<6;
-	case ASRL:	return 0x3<<6;
-	case ASRA:	return 0x4<<6;
-	case ASLL:	return 0x2<<6;
-	}
-	diag("bad thumbop oprr %d", a);
-	prasm(curp);
-	return 0;
-}
-
-static long
-thumbopirr(int a, int ld)
-{
-	if(ld)
-		diag("load in thumbopirr");
-	switch(a){
-		case AMOVW:	return 0xc<<11;
-		case AMOVH:
-		case AMOVHU:	return 0x10<<11;
-		case AMOVB:
-		case AMOVBU:	return 0xe<<11;
-	}
-	return 0;
-}
-	
-static long
-thumboprrr(int a, int ld)
-{
-	if(ld){
-		switch(a){
-		case AMOVW:	return 0x2c<<9;
-		case AMOVH:	return 0x2f<<9;
-		case AMOVB:	return 0x2b<<9;
-		case AMOVHU:	return 0x2d<<9;
-		case AMOVBU:	return 0x2e<<9;
-		}
-	}
-	else{
-		switch(a){
-		case AMOVW:	return 0x28<<9;
-		case AMOVHU:
-		case AMOVH:	return 0x29<<9;
-		case AMOVBU:
-		case AMOVB:	return 0x2a<<9;
-		}
-	}
-	diag("bad thumbop oprrr %d", a);
-	prasm(curp);
-	return 0;
-}
-
-static long
-thumbopri(int a)
-{
-	switch(a) {
-	case ASRL:	return 0x1<<11;
-	case ASRA:	return 0x2<<11;
-	case ASLL:	return 0x0<<11;
-	case AADD:	return 0x2<<11;
-	case ASUB:	return 0x3<<11;
-	case AMOVW:	return 0x0<<11;
-	case ACMP:	return 0x1<<11;
-	}
-	diag("bad thumbop opri %d", a);
-	prasm(curp);
-	return 0;
-}
-
-static long
-thumbophh(int a)
-{
-	switch(a) {
-	case AADD:	return 0x0<<8;
-	case AMOVW:	return 0x2<<8;
-	case ACMP:	return 0x1<<8;
-	}
-	diag("bad thumbop ophh %d", a);
-	prasm(curp);
-	return 0;
-}
-
-static long
-thumbopbra(int a)
-{
-	switch(a) {
-	case ABEQ:	return 0x0<<8;
-	case ABNE:	return 0x1<<8;
-	case ABCS:	return 0x2<<8;
-	case ABHS:	return 0x2<<8;
-	case ABCC:	return 0x3<<8;
-	case ABLO:	return 0x3<<8;
-	case ABMI:	return 0x4<<8;
-	case ABPL:	return 0x5<<8;
-	case ABVS:	return 0x6<<8;
-	case ABVC:	return 0x7<<8;
-	case ABHI:	return 0x8<<8;
-	case ABLS:	return 0x9<<8;
-	case ABGE:	return 0xa<<8;
-	case ABLT:	return 0xb<<8;
-	case ABGT:	return 0xc<<8;
-	case ABLE:	return 0xd<<8;
-	}
-	diag("bad thumbop opbra %d", a);
-	prasm(curp);
-	return 0;
-}
-
-static long
-thumbopmv(int a, int ld)
-{
-	switch(a) {
-	case AMOVW: 	return (ld ? 0xd : 0xc)<<11;
-	case AMOVH:
-	case AMOVHU:	return (ld ? 0x11: 0x10)<<11;
-	case AMOVB:
-	case AMOVBU:	return (ld ? 0xf : 0xe)<<11;
-	}
-	diag("bad thumbop opmv %d", a);
-	prasm(curp);
-	return 0;
-}
-
-static void 
-lowreg(Prog *p, int r)
-{
-	if(high(r))
-		diag("high reg [%P]", p);
-}
-
-static void
-mult(Prog *p, int n, int m)
-{
-	if(m*(n/m) != n)
-		diag("%d not M(%d) [%P]", n, m, p);
-}
-
-static void 
-numr(Prog *p, int n, int min, int max)
-{
-	if(n < min || n > max)
-		diag("%d not in %d-%d [%P]", n, min, max, p);
-}
-
-static void 
-regis(Prog *p, int r, int r1, int r2)
-{
-	if(r != r1 && r != r2)
-		diag("reg %d not %d or %d [%P]", r, r1, r2, p);
-}
-
-void
-hputl(int n)
-{
-	cbp[1] = n>>8;
-	cbp[0] = n;
-	cbp += 2;
-	cbc -= 2;
-	if(cbc <= 0)
-		cflush();
-}
-
-void
-thumbcount()
-{
-	int i, c = 0, t = 0;
-
-	for (i = 0; i < OPCNTSZ; i++)
-		t += opcount[i];
-	if(t == 0)
-		return;
-	for (i = 0; i < OPCNTSZ; i++){
-		c += opcount[i];
-		print("%d:	%d %d %d%%\n", i, opcount[i], c, (opcount[i]*100+t/2)/t);
-	}
-}
-	
-char *op1[] = { "lsl", "lsr", "asr" };
-char *op2[] = { "add", "sub" };
-char *op3[] = { "movw", "cmp", "add", "sub" };
-char *op4[] = { "and", "eor", "lsl", "lsr", "asr", "adc", "sbc", "ror",
-		        "tst", "neg", "cmp", "cmpn", "or", "mul", "bitc", "movn" };
-char *op5[] = { "add", "cmp", "movw", "bx" };
-char *op6[] = { "smovw", "smovh", "smovb", "lmovb", "lmovw", "lmovhu", "lmovbu", "lmovh" };
-char *op7[] = { "smovw", "lmovw", "smovb", "lmovbu" };
-char *op8[] = { "smovh", "lmovhu" };
-char *op9[] = { "smovw", "lmovw" };
-char *op10[] = { "push", "pop" };
-char *op11[] = { "stmia", "ldmia" };
-
-char *cond[] = { "eq", "ne", "hs", "lo", "mi", "pl", "vs", "vc",
-			 "hi", "ls", "ge", "lt", "gt", "le", "al", "nv" };
-
-#define B(h, l)		bits(i, h, l)
-#define IMM(h, l)	B(h, l)
-#define REG(h, l)	reg(B(h, l))
-#define LHREG(h, l, lh)	lhreg(B(h, l), B(lh, lh))
-#define COND(h, l)	cond[B(h, l)]
-#define OP1(h, l)	op1[B(h, l)]
-#define OP2(h, l)	op2[B(h, l)]
-#define OP3(h, l)	op3[B(h, l)]
-#define OP4(h, l)	op4[B(h, l)]
-#define OP5(h, l)	op5[B(h, l)]
-#define OP6(h, l)	op6[B(h, l)]
-#define OP7(h, l)	op7[B(h, l)]
-#define OP8(h, l)	op8[B(h, l)]
-#define OP9(h, l)	op9[B(h, l)]
-#define OP10(h, l)	op10[B(h, l)]
-#define OP11(h, l)	op11[B(h, l)]
-#define SBZ(h, l)	if(IMM(h, l) != 0) diag("%x: %x bits %d,%d not zero", pc, i, h, l)
-#define SNBZ(h, l)	if(IMM(h, l) == 0) diag("%x: %x bits %d,%d zero", pc, i, h, l)
-#define SBO(h, l)	if(IMM(h, l) != 1) diag("%x: %x bits %d,%d not one", pc, i, h, l)
-
-static int
-bits(int i, int h, int l)
-{
-	if(h < l)
-		diag("h < l in bits");
-	return (i&(((1<<(h-l+1))-1)<<l))>>l;
-}
-
-static char *
-reg(int r)
-{
-	static char s[4][4];
-	static int i = 0;
-
-	if(r < 0 || r > 7)
-		diag("register %d out of range", r);
-	i++;
-	if(i == 4)
-		i = 0;
-	sprint(s[i], "r%d", r);
-	return s[i];
-}
-
-static char *regnames[] = { "sp", "lr", "pc" };
-
-static char *
-lhreg(int r, int lh)
-{
-	static char s[4][4];
-	static int i = 0;
-
-	if(lh == 0)
-		return reg(r);
-	if(r < 0 || r > 7)
-		diag("high register %d out of range", r);
-	i++;
-	if(i == 4)
-		i = 0;
-	if(r >= 5)
-		sprint(s[i], "%s", regnames[r-5]);
-	else
-		sprint(s[i], "r%d", r+8);
-	return s[i];
-}
-	
-static void
-illegal(int i, int pc)
-{
-	diag("%x: %x illegal instruction", pc, i);
-}
-
-static void
-dis(int i, int pc)
-{
-	static int lasto;
-	int o, l;
-	char *op;
-
-	print("%x: %x:	", pc, i);
-	if(i&0xffff0000)
-		illegal(i, pc);
-	o = B(15, 13);
-	switch(o){
-	case 0:
-		o = B(12, 11);
-		switch(o){
-			case 0:
-			case 1:
-			case 2:
-				print("%s	%d, %s, %s\n", OP1(12, 11), IMM(10, 6), REG(5, 3), REG(2, 0));
-				return;
-			case 3:
-				if(B(10, 10) == 0)
-					print("%s	%s, %s, %s\n", OP2(9, 9), REG(8, 6), REG(5, 3), REG(2, 0));
-				else
-					print("%s	%d, %s, %s\n", OP2(9, 9), IMM(8, 6), REG(5, 3), REG(2, 0));
-				return;
-		}
-	case 1:
-		print("%s	%d, %s\n", OP3(12, 11), IMM(7, 0), REG(10, 8));
-		return;
-	case 2:
-		o = B(12, 10);
-		if(o == 0){
-			print("%s	%s, %s\n", OP4(9, 6), REG(5, 3), REG(2, 0));
-			return;
-		}
-		if(o == 1){
-			o = B(9, 8);
-			if(o == 3){
-				SBZ(7, 7);
-				SBZ(2, 0);
-				print("%s	%s\n", OP5(9, 8), LHREG(5, 3, 6));
-				return;
-			}
-			SNBZ(7, 6);
-			print("%s	%s, %s\n", OP5(9, 8), LHREG(5, 3, 6), LHREG(2, 0, 7));
-			return;
-		}
-		if(o == 2 || o == 3){
-			print("movw	%d(pc)[%x], %s\n", 4*IMM(7, 0), 4*IMM(7, 0)+pc+4, REG(10, 8));
-			return;
-		}
-		op = OP6(11, 9);
-		if(*op == 'l')
-			print("%s	[%s, %s], %s\n", op+1, REG(8, 6), REG(5, 3), REG(2, 0));
-		else
-			print("%s	%s, [%s, %s]\n", op+1, REG(2, 0), REG(8, 6), REG(5, 3));
-		return;
-	case 3:
-		op = OP7(12, 11);
-		if(B(12, 11) == 0 || B(12,11) == 1)
-			l = 4;
-		else
-			l = 1;
-		if(*op == 'l')
-			print("%s	%d(%s), %s\n", op+1, l*IMM(10, 6), REG(5, 3), REG(2, 0));
-		else
-			print("%s	%s, %d(%s)\n", op+1, REG(2, 0), l*IMM(10, 6), REG(5, 3));
-		return;
-	case 4:
-		if(B(12, 12) == 0){
-			op = OP8(11, 11);
-			if(*op == 'l')
-				print("%s	%d(%s), %s\n", op+1, 2*IMM(10, 6), REG(5, 3), REG(2, 0));
-			else
-				print("%s	%s, %d(%s)\n", op+1, REG(2, 0), 2*IMM(10, 6), REG(5, 3));
-			return;
-		}
-		op = OP9(11, 11);
-		if(*op == 'l')
-			print("%s	%d(sp), %s\n", op+1, 4*IMM(7, 0), REG(10, 8));
-		else
-			print("%s	%s, %d(sp)\n", op+1, REG(10, 8), 4*IMM(7, 0));
-		return;
-	case 5:
-		if(B(12, 12) == 0){
-			if(B(11, 11) == 0)
-				print("add	%d, pc, %s\n", 4*IMM(7, 0), REG(10, 8));
-			else
-				print("add	%d, sp, %s\n", 4*IMM(7, 0), REG(10, 8));
-			return;
-		}
-		if(B(11, 8) == 0){
-			print("%s	%d, sp\n", OP2(7, 7), 4*IMM(6, 0));
-			return;
-		}
-		SBO(10, 10);
-		SBZ(9, 9);
-		if(B(8, 8) == 0)
-			print("%s	sp, %d\n", OP10(11, 11), IMM(7, 0));
-		else
-			print("%s	sp, %d|15\n", OP10(11, 11), IMM(7, 0));
-		return;
-	case 6:
-		if(B(12, 12) == 0){
-			print("%s	%s, %d\n", OP11(11, 11), REG(10, 8), IMM(7, 0));
-			return;
-		}
-		if(B(11, 8) == 0xf){
-			print("swi	%d\n", IMM(7, 0));
-			return;
-		}
-		o = IMM(7, 0);
-		if(o&0x80)
-			o |= 0xffffff00;
-		o = pc+4+(o<<1);
-		print("b%s	%x\n", COND(11, 8), o);
-		return;
-	case 7:
-		o = B(12, 11);
-		switch(o){
-			case 0:
-				o = IMM(10, 0);
-				if(o&0x400)
-					o |= 0xfffff800;
-				o = pc+4+(o<<1);
-				print("b	%x\n", o);
-				return;
-			case 1:
-				illegal(i, pc);
-				return;
-			case 2:
-				lasto = IMM(10, 0);
-				print("bl\n");
-				return;
-			case 3:
-				if(lasto&0x400)
-					lasto |= 0xfffff800;
-				o = IMM(10, 0);
-				o = (pc-2)+4+(o<<1)+(lasto<<12);
-				print("bl %x\n", o);
-				return;
-		}
-	}
-}