From 1ac2cfc72042ffbe17155b9407907d6686dc4664 Mon Sep 17 00:00:00 2001
From: Kai Backman <kaib@golang.org>
Date: Tue, 16 Jun 2009 11:25:58 -0700
Subject: [PATCH] grab bag of changes aimed at getting stack splitting to work:
 - morestack support for 5l and arm runtime - argsize support in 5c, 5l, ar
 and nm. assembly code from 5a   will break in interesting ways unless NOSPLIT
 is specified - explicit cond execution constants - fix 5l output to use %d
 instead of %ld so that negative   values show. - added a lot of code to
 arm/asm.s. runtime entry code almost   working currently aborts at gogo not
 implemented

R=rsc
APPROVED=rsc
DELTA=305  (125 added, 29 deleted, 151 changed)
OCL=30246
CL=30347
---
 src/cmd/5c/gc.h           |  11 +-
 src/cmd/5c/list.c         |  14 ++-
 src/cmd/5c/sgen.c         |  30 ++++-
 src/cmd/5c/swt.c          |   9 +-
 src/cmd/5l/5.out.h        |  25 ++++-
 src/cmd/5l/l.h            |   1 +
 src/cmd/5l/list.c         |  40 ++++---
 src/cmd/5l/noop.c         | 106 ++++++++++--------
 src/cmd/5l/obj.c          |   2 +
 src/cmd/5l/span.c         |  17 +--
 src/libmach_amd64/5obj.c  |   5 +
 src/pkg/runtime/arm/asm.s | 229 ++++++++++++++++++++------------------
 12 files changed, 293 insertions(+), 196 deletions(-)

diff --git a/src/cmd/5c/gc.h b/src/cmd/5c/gc.h
index 4ddff6d9a2..297a6073e1 100644
--- a/src/cmd/5c/gc.h
+++ b/src/cmd/5c/gc.h
@@ -61,6 +61,7 @@ typedef	struct	Rgn	Rgn;
 struct	Adr
 {
 	int32	offset;
+	int32	offset2;
 	double	dval;
 	char	sval[NSNAME];
 	Ieee	ieee;
@@ -140,7 +141,7 @@ struct	Reg
 	int32	regu;
 	int32	loop;		/* could be shorter */
 
-	
+
 	Reg*	log5;
 	int32	active;
 
@@ -362,10 +363,10 @@ int32	FtoB(int);
 int	BtoR(int32);
 int	BtoF(int32);
 
-void	predicate(void); 
-int	isbranch(Prog *); 
-int	predicable(Prog *p); 
-int	modifiescpsr(Prog *p); 
+void	predicate(void);
+int	isbranch(Prog *);
+int	predicable(Prog *p);
+int	modifiescpsr(Prog *p);
 
 #pragma	varargck	type	"A"	int
 #pragma	varargck	type	"B"	Bits
diff --git a/src/cmd/5c/list.c b/src/cmd/5c/list.c
index 14454abb05..c792c130da 100644
--- a/src/cmd/5c/list.c
+++ b/src/cmd/5c/list.c
@@ -28,6 +28,7 @@
 // OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
 // THE SOFTWARE.
 
+
 #define	EXTERN
 #include "gc.h"
 
@@ -71,9 +72,9 @@ Bconv(Fmt *fp)
 }
 
 char *extra [] = {
-	".EQ", ".NE", ".CS", ".CC", 
-	".MI", ".PL", ".VS", ".VC", 
-	".HI", ".LS", ".GE", ".LT", 
+	".EQ", ".NE", ".CS", ".CC",
+	".MI", ".PL", ".VS", ".VC",
+	".HI", ".LS", ".GE", ".LT",
 	".GT", ".LE", "", ".NV",
 };
 
@@ -86,7 +87,7 @@ Pconv(Fmt *fp)
 
 	p = va_arg(fp->args, Prog*);
 	a = p->as;
-	s = p->scond; 
+	s = p->scond;
 	strcpy(sc, extra[s & C_SCOND]);
 	if(s & C_SBIT)
 		strcat(sc, ".S");
@@ -162,6 +163,10 @@ Dconv(Fmt *fp)
 			sprint(str, "$%N", a);
 		break;
 
+	case D_CONST2:
+		sprint(str, "$%d-%d", a->offset, a->offset2);
+		break;
+
 	case D_SHIFT:
 		v = a->offset;
 		op = "<<>>->@>" + (((v>>5) & 3) << 1);
@@ -224,6 +229,7 @@ Rconv(Fmt *fp)
 	sprint(str, "GOK-reglist");
 	switch(a->type) {
 	case D_CONST:
+	case D_CONST2:
 		if(a->reg != NREG)
 			break;
 		if(a->sym != S)
diff --git a/src/cmd/5c/sgen.c b/src/cmd/5c/sgen.c
index b7f61bb7ea..e18cb61680 100644
--- a/src/cmd/5c/sgen.c
+++ b/src/cmd/5c/sgen.c
@@ -30,6 +30,31 @@
 
 #include "gc.h"
 
+int32
+argsize(void)
+{
+	Type *t;
+	int32 s;
+
+//print("t=%T\n", thisfn);
+	s = 0;
+	for(t=thisfn->down; t!=T; t=t->down) {
+		switch(t->etype) {
+		case TVOID:
+			break;
+		case TDOT:
+			s += 64;
+			break;
+		default:
+			s = align(s, t, Aarg1);
+			s = align(s, t, Aarg2);
+			break;
+		}
+//print("	%d %T\n", s, t);
+	}
+	return (s+7) & ~7;
+}
+
 void
 codgen(Node *n, Node *nn)
 {
@@ -53,6 +78,9 @@ codgen(Node *n, Node *nn)
 	}
 	nearln = nn->lineno;
 	gpseudo(ATEXT, n1->sym, nodconst(stkoff));
+	p->to.type = D_CONST2;
+	p->to.offset2 = argsize();
+
 	sp = p;
 
 	/*
@@ -313,7 +341,7 @@ loop:
 
 		patch(spc, pc);
 		gen(l->right->right);	/* inc */
-		patch(sp, pc);	
+		patch(sp, pc);
 		if(l->left != Z) {	/* test */
 			bcomplex(l->left, Z);
 			patch(p, breakpc);
diff --git a/src/cmd/5c/swt.c b/src/cmd/5c/swt.c
index 83f7f5621c..28314dacaa 100644
--- a/src/cmd/5c/swt.c
+++ b/src/cmd/5c/swt.c
@@ -338,7 +338,7 @@ loop:
 		if(vs < 0) {
 			gopcode(OAS, &nod1, Z, &nod1);
 			gopcode(OSUB, &nod1, nodconst(0), nn);
-		} else 
+		} else
 			gopcode(OAS, &nod1, Z, nn);
 		regfree(&nod1);
 		return 1;
@@ -649,6 +649,13 @@ zaddr(char *bp, Adr *a, int s)
 	case D_PSR:
 		break;
 
+	case D_CONST2:
+		l = a->offset2;
+		bp[0] = l;
+		bp[1] = l>>8;
+		bp[2] = l>>16;
+		bp[3] = l>>24;
+		bp += 4;	// fall through
 	case D_OREG:
 	case D_CONST:
 	case D_BRANCH:
diff --git a/src/cmd/5l/5.out.h b/src/cmd/5l/5.out.h
index 1a0da76952..45f30fa5d5 100644
--- a/src/cmd/5l/5.out.h
+++ b/src/cmd/5l/5.out.h
@@ -84,10 +84,10 @@ enum	as
 	AB,
 	ABL,
 
-/* 
- * Do not reorder or fragment the conditional branch 
- * opcodes, or the predication code will break 
- */ 
+/*
+ * Do not reorder or fragment the conditional branch
+ * opcodes, or the predication code will break
+ */
 	ABEQ,
 	ABNE,
 	ABCS,
@@ -186,6 +186,22 @@ enum	as
 #define	C_FBIT	(1<<7)	/* psr flags-only */
 #define	C_UBIT	(1<<7)	/* up bit */
 
+#define C_SCOND_EQ	0
+#define C_SCOND_NE	1
+#define C_SCOND_HS	2
+#define C_SCOND_LO	3
+#define C_SCOND_MI	4
+#define C_SCOND_PL	5
+#define C_SCOND_VS	6
+#define C_SCOND_VC	7
+#define C_SCOND_HI	8
+#define C_SCOND_LS	9
+#define C_SCOND_GE	10
+#define C_SCOND_LT	11
+#define C_SCOND_GT	12
+#define C_SCOND_LE	13
+#define C_SCOND_N	15
+
 /* type/name */
 #define	D_GOK	0
 #define	D_NONE	1
@@ -209,6 +225,7 @@ enum	as
 #define D_ADDR		(D_NONE+22)
 
 #define D_SBIG		(D_NONE+23)
+#define	D_CONST2	(D_NONE+24)
 
 /* name */
 #define	D_EXTERN	(D_NONE+3)
diff --git a/src/cmd/5l/l.h b/src/cmd/5l/l.h
index fdb0488275..36ada96d29 100644
--- a/src/cmd/5l/l.h
+++ b/src/cmd/5l/l.h
@@ -72,6 +72,7 @@ struct	Adr
 	uchar	index; // not used on arm, required by ld/go.c
 	char	reg;
 	char	name;
+	int32	offset2; // argsize
 	char	class;
 };
 
diff --git a/src/cmd/5l/list.c b/src/cmd/5l/list.c
index 5375829185..c3153938c1 100644
--- a/src/cmd/5l/list.c
+++ b/src/cmd/5l/list.c
@@ -61,7 +61,7 @@ Pconv(Fmt *fp)
 	switch(a) {
 	default:
 		s = str;
-		s += sprint(s, "(%ld)", p->line);
+		s += sprint(s, "(%d)", p->line);
 		if(p->reg == NREG)
 			sprint(s, "	%A%C	%D,%D",
 				a, p->scond, &p->from, &p->to);
@@ -76,23 +76,23 @@ Pconv(Fmt *fp)
 
 	case ASWPW:
 	case ASWPBU:
-		sprint(str, "(%ld)	%A%C	R%d,%D,%D",
+		sprint(str, "(%d)	%A%C	R%d,%D,%D",
 			p->line, a, p->scond, p->reg, &p->from, &p->to);
 		break;
 
 	case ADATA:
 	case AINIT:
 	case ADYNT:
-		sprint(str, "(%ld)	%A%C	%D/%d,%D",
+		sprint(str, "(%d)	%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);
+		sprint(str, "WORD %x", p->to.offset);
 		break;
 
 	case ADWORD:
-		sprint(str, "DWORD %ld %ld", p->from.offset, p->to.offset);
+		sprint(str, "DWORD %x %x", p->from.offset, p->to.offset);
 		break;
 	}
 	return fmtstrcpy(fp, str);
@@ -178,13 +178,17 @@ Dconv(Fmt *fp)
 			sprint(str, "$%N(R%d)", a, a->reg);
 		break;
 
+	case D_CONST2:
+		sprint(str, "$%d-%d", a->offset, a->offset2);
+		break;
+
 	case D_SHIFT:
 		v = a->offset;
 		op = "<<>>->@>" + (((v>>5) & 3) << 1);
 		if(v & (1<<4))
-			sprint(str, "R%ld%c%cR%ld", v&15, op[0], op[1], (v>>8)&15);
+			sprint(str, "R%d%c%cR%d", v&15, op[0], op[1], (v>>8)&15);
 		else
-			sprint(str, "R%ld%c%c%ld", v&15, op[0], op[1], (v>>7)&31);
+			sprint(str, "R%d%c%c%d", v&15, op[0], op[1], (v>>7)&31);
 		if(a->reg != NREG)
 			sprint(str+strlen(str), "(R%d)", a->reg);
 		break;
@@ -262,9 +266,9 @@ Dconv(Fmt *fp)
 				sprint(str, "%.5lux(BRANCH)", v);
 		} else
 			if(a->sym != S)
-				sprint(str, "%s+%ld(APC)", a->sym->name, a->offset);
+				sprint(str, "%s+%d(APC)", a->sym->name, a->offset);
 			else
-				sprint(str, "%ld(APC)", a->offset);
+				sprint(str, "%d(APC)", a->offset);
 		break;
 
 	case D_FCONST:
@@ -293,35 +297,35 @@ Nconv(Fmt *fp)
 		break;
 
 	case D_NONE:
-		sprint(str, "%ld", a->offset);
+		sprint(str, "%d", a->offset);
 		break;
 
 	case D_EXTERN:
 		if(s == S)
-			sprint(str, "%ld(SB)", a->offset);
+			sprint(str, "%d(SB)", a->offset);
 		else
-			sprint(str, "%s+%ld(SB)", s->name, a->offset);
+			sprint(str, "%s+%d(SB)", s->name, a->offset);
 		break;
 
 	case D_STATIC:
 		if(s == S)
-			sprint(str, "<>+%ld(SB)", a->offset);
+			sprint(str, "<>+%d(SB)", a->offset);
 		else
-			sprint(str, "%s<>+%ld(SB)", s->name, a->offset);
+			sprint(str, "%s<>+%d(SB)", s->name, a->offset);
 		break;
 
 	case D_AUTO:
 		if(s == S)
-			sprint(str, "%ld(SP)", a->offset);
+			sprint(str, "%d(SP)", a->offset);
 		else
-			sprint(str, "%s-%ld(SP)", s->name, -a->offset);
+			sprint(str, "%s-%d(SP)", s->name, -a->offset);
 		break;
 
 	case D_PARAM:
 		if(s == S)
-			sprint(str, "%ld(FP)", a->offset);
+			sprint(str, "%d(FP)", a->offset);
 		else
-			sprint(str, "%s+%ld(FP)", s->name, a->offset);
+			sprint(str, "%s+%d(FP)", s->name, a->offset);
 		break;
 	}
 	return fmtstrcpy(fp, str);
diff --git a/src/cmd/5l/noop.c b/src/cmd/5l/noop.c
index 333a3999a3..f4de0a0eba 100644
--- a/src/cmd/5l/noop.c
+++ b/src/cmd/5l/noop.c
@@ -128,7 +128,7 @@ noops(void)
 	Bflush(&bso);
 
 	pmorestack = P;
-	symmorestack = lookup("sys·morestack", 0);
+	symmorestack = lookup("sys·morestackx", 0);
 
 	if(symmorestack->type == STEXT)
 	for(p = firstp; p != P; p = p->link) {
@@ -342,8 +342,7 @@ noops(void)
 				break;
 			}
 
-// 			if(p->reg & NOSPLIT) {
-			if(1) {
+			if(p->reg & NOSPLIT) {
 				q1 = prg();
 				q1->as = AMOVW;
 				q1->scond |= C_WBIT;
@@ -355,61 +354,78 @@ noops(void)
 				q1->to.reg = REGSP;
 				q1->link = p->link;
 				p->link = q1;
-			} else { // !NOSPLIT
-				// split stack check
-				if(autosize < StackBig) {
-					p = appendp(p); // load G.stackguard into R1
-					p->as = AMOVW;
-					p->from.type = D_OREG;
-					p->from.reg = REGG;
-					p->to.type = D_REG;
-					p->to.reg = 1;
-
-					p = appendp(p);
-					p->as = ACMP;
-					p->from.type = D_REG;
-					p->from.reg = 1;
-					p->from.offset = -autosize;
-					p->reg = REGSP;
-				}
-
-				// TODO(kaib): Optimize the heck out of this
-				p = appendp(p); // store autosize in M.morearg
+			} else if (autosize < StackBig) {
+				// split stack check for small functions
+				// MOVW			(REGG), R1
+				// CMP			R1, $-autosize(SP)
+				// MOVW.W.LT	R14,$-autosize(SP)
+				// MOVW.W.GE	R14,$-4(SP)
+				// MOVW.GE		$(args << 24 | autosize), R1
+				// BL.GE		callmorestack(SB)
+
+				// TODO(kaib): double check we allocate autosize after
+				// 				stack has been split
+				// TODO(kaib): add error in case autosize doesn't pack
+				// TODO(kaib): add more trampolines
+				// TODO(kaib): put stackguard in register
+				// TODO(kaib): add support for -K and underflow detection
+
+				p = appendp(p); // load G.stackguard into R1
 				p->as = AMOVW;
-				p->from.type = D_CONST;
-				if(autosize+160 > 4096)
-					p->from.offset = (autosize+160) & ~7LL;
+				p->from.type = D_OREG;
+				p->from.reg = REGG;
 				p->to.type = D_REG;
-				p->to.reg = REGTMP;
+				p->to.reg = 1;
+
+				p = appendp(p);
+				p->as = ACMP;
+				p->from.type = D_REG;
+				p->from.reg = 1;
+				p->from.offset = -autosize;
+				p->reg = REGSP;
 
 				p = appendp(p);
 				p->as = AMOVW;
+ 				p->scond = C_SCOND_GE | C_WBIT;
 				p->from.type = D_REG;
-				p->from.reg = REGTMP;
+				p->from.reg = REGLINK;
 				p->to.type = D_OREG;
-				p->to.reg = REGM;
-				p->to.offset = 4;
+				p->to.offset = -autosize;
+				p->to.reg = REGSP;
 
 				p = appendp(p);
 				p->as = AMOVW;
+				p->scond = C_SCOND_LT | C_WBIT;
+				p->from.type = D_REG;
+				p->from.reg = REGLINK;
+				p->to.type = D_OREG;
+				p->to.offset = -4;
+				p->to.reg = REGSP;
+
+				p = appendp(p); // packs args and autosize
+				p->as = AMOVW;
+				p->scond = C_SCOND_LT;
 				p->from.type = D_CONST;
-// 				p->from.offset = curtext->to.offset2;
+				// top 8 bits are arg count, lower 24 bits number of 4 byte
+				// words
+				p->from.offset =
+					(curtext->to.offset2 & ~7) << 21 |
+					(autosize & ~7) >> 3;
 				p->to.type = D_REG;
-				p->to.reg = REGTMP;
+				p->to.reg = 1;
 
 				p = appendp(p);
-				p->as = AMOVW;
-				p->from.type = D_REG;
-				p->from.reg = REGTMP;
-				p->to.type = D_OREG;
-				p->to.reg = REGM;
-				p->to.offset = 8;
- 
-// 				p = appendp(p);
-// 				p->as = ABL;
-// 				p->to.type = D_BRANCH;
-// 				p->to.sym = symmorestack;
-// 				p->cond = pmorestack;
+				p->as = ABL;
+				p->scond = C_SCOND_LT;
+ 				p->to.type = D_BRANCH;
+				p->to.sym = symmorestack;
+				p->cond = pmorestack;
+			} else { // > StackBig
+				// MOVW.W		R14,$-4(SP)
+				// MOVW			$(args << 24 | autosize), R1
+				// BL			callmorestack(SB)
+				// TODO(kaib): Fix large stacks, don't use packing
+				diag("StackBig broken");
 			}
 			break;
 
@@ -803,7 +819,7 @@ noops(void)
 					p->link = q;
 				}
 			}
-			if(seenthumb && !thumb && p->to.type == D_OREG && p->to.reg == REGLINK){	
+			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;
 			}
diff --git a/src/cmd/5l/obj.c b/src/cmd/5l/obj.c
index a34a20ebcd..bcb2110f01 100644
--- a/src/cmd/5l/obj.c
+++ b/src/cmd/5l/obj.c
@@ -546,6 +546,8 @@ zaddr(Biobuf *f, Adr *a, Sym *h[])
 		c++;
 		break;
 
+	case D_CONST2:
+		a->offset2 = Bget4(f);	// fall through
 	case D_BRANCH:
 	case D_OREG:
 	case D_CONST:
diff --git a/src/cmd/5l/span.c b/src/cmd/5l/span.c
index c60b5478cc..4ca8e01f74 100644
--- a/src/cmd/5l/span.c
+++ b/src/cmd/5l/span.c
@@ -87,12 +87,12 @@ fninc(Sym *s)
 				return 4;
 			else
 				return 0;
-		}			
+		}
 	}
 	return 0;
 }
 
-int 
+int
 fnpinc(Sym *s)
 {
 	if(!s->fnptr){	// a simplified case BX O(R) -> BL O(R)
@@ -318,7 +318,7 @@ span(void)
 				else
 					m = 2;
 				p->align = 0;
-			}	
+			}
 			if(p->align){
 				if((p->align == 4 && (c&3)) || (p->align == 2 && !(c&3))){
 					if(ispad(op)){
@@ -374,9 +374,9 @@ span(void)
 		// print("%d bytes removed (padding)\n", d);
 		c -= d;
 	}
-	
+
 	if(debug['t']) {
-		/* 
+		/*
 		 * add strings to text segment
 		 */
 		c = rnd(c, 8);
@@ -391,7 +391,7 @@ span(void)
 			c += v;
 		}
 	}
-			
+
 	c = rnd(c, 8);
 
 	setext = lookup("etext", 0);
@@ -726,6 +726,7 @@ aclass(Adr *a)
 		return C_FCON;
 
 	case D_CONST:
+	case D_CONST2:
 		switch(a->name) {
 
 		case D_NONE:
@@ -1072,7 +1073,7 @@ buildop(void)
 			oprange[AMOVFD] = oprange[r];
 			oprange[AMOVDF] = oprange[r];
 			break;
-			
+
 		case ACMPF:
 			oprange[ACMPD] = oprange[r];
 			break;
@@ -1246,7 +1247,7 @@ asmdyn()
 				t += 4;
 				t += sput(s->name);
 			}
-	
+
 	la = 0;
 	r = &rels;
 	n = r->n;
diff --git a/src/libmach_amd64/5obj.c b/src/libmach_amd64/5obj.c
index fa7be5abd4..08a7738d2a 100644
--- a/src/libmach_amd64/5obj.c
+++ b/src/libmach_amd64/5obj.c
@@ -123,6 +123,11 @@ addr(Biobuf *bp)
 	case D_PSR:
 	case D_FPCR:
 		break;
+	case D_CONST2:
+		Bgetc(bp);
+		Bgetc(bp);
+		Bgetc(bp);
+		Bgetc(bp);	// fall through
 	case D_OREG:
 	case D_CONST:
 	case D_BRANCH:
diff --git a/src/pkg/runtime/arm/asm.s b/src/pkg/runtime/arm/asm.s
index e47ab86e31..5e68b72ffe 100644
--- a/src/pkg/runtime/arm/asm.s
+++ b/src/pkg/runtime/arm/asm.s
@@ -3,88 +3,80 @@
 // license that can be found in the LICENSE file.
 
 TEXT _rt0_arm(SB),7,$0
-// copy arguments forward on an even stack
-//     	MOVW	$0(SP), R0
-//	MOVL	0(SP), R1		// argc
-//	LEAL	4(SP), R1		// argv
-//	SUBL	$128, SP		// plenty of scratch
-//	ANDL	$~7, SP
-//	MOVL	AX, 120(SP)		// save argc, argv away
-//	MOVL	BX, 124(SP)
-
-
-//	// write "go386\n"
-//	PUSHL	$6
-//	PUSHL	$hello(SB)
-//	PUSHL	$1
-//	CALL	sys·write(SB)
-//	POPL	AX
-//	POPL	AX
-//	POPL	AX
-
-
-//	CALL	ldt0setup(SB)
-
-	// set up %fs to refer to that ldt entry
-//	MOVL	$(7*8+7), AX
-//	MOVW	AX, FS
-
-//	// store through it, to make sure it works
-//	MOVL	$0x123, 0(FS)
-//	MOVL	tls0(SB), AX
-//	CMPL	AX, $0x123
-//	JEQ	ok
-//	MOVL	AX, 0
-// ok:
-
-//	// set up m and g "registers"
-//	// g is 0(FS), m is 4(FS)
-//	LEAL	g0(SB), CX
-//	MOVL	CX, 0(FS)
-//	LEAL	m0(SB), AX
-//	MOVL	AX, 4(FS)
-
-//	// save m->g0 = g0
-//	MOVL	CX, 0(AX)
-
-//	// create istack out of the OS stack
-//	LEAL	(-8192+104)(SP), AX	// TODO: 104?
-//	MOVL	AX, 0(CX)	// 8(g) is stack limit (w 104b guard)
-//	MOVL	SP, 4(CX)	// 12(g) is base
-//	CALL	emptyfunc(SB)	// fault if stack check is wrong
-
-//	// convention is D is always cleared
-//	CLD
-
-//	CALL	check(SB)
-
-//	// saved argc, argv
-//	MOVL	120(SP), AX
-//	MOVL	AX, 0(SP)
-//	MOVL	124(SP), AX
-//	MOVL	AX, 4(SP)
-//	CALL	args(SB)
-//	CALL	osinit(SB)
-//	CALL	schedinit(SB)
-
-//	// create a new goroutine to start program
-//	PUSHL	$mainstart(SB)	// entry
-//	PUSHL	$8	// arg size
-//	CALL	sys·newproc(SB)
-//	POPL	AX
-//	POPL	AX
+	MOVW $setR12(SB), R12
+
+	// copy arguments forward on an even stack
+	MOVW	0(SP), R0		// argc
+	MOVW	4(SP), R1		// argv
+	SUB	$128, SP		// plenty of scratch
+	AND	$~7, SP
+	MOVW	R0, 120(SP)		// save argc, argv away
+	MOVW	R1, 124(SP)
+
+	// set up m and g registers
+	// g is R10, m is R9
+	MOVW	$g0(SB), R10
+	MOVW	$m0(SB), R9
+
+	// save m->g0 = g0
+	MOVW	R10, 0(R9)
+
+	// create istack out of the OS stack
+	MOVW	$(-8192+104)(SP), R0
+	MOVW	R0, 0(R10)	// 0(g) is stack limit (w 104b guard)
+	MOVW	SP, 4(R10)	// 4(g) is base
+	BL	emptyfunc(SB)	// fault if stack check is wrong
+
+	BL	check(SB)
+
+	// saved argc, argv
+	MOVW	120(SP), R0
+	MOVW	R0, 0(SP)
+	MOVW	124(SP), R0
+	MOVW	R0, 4(SP)
+	BL	args(SB)
+	BL	osinit(SB)
+	BL	schedinit(SB)
+
+	// create a new goroutine to start program
+	MOVW	$mainstart(SB), R0
+	MOVW.W	R0, -4(SP)
+	MOVW	$8, R0
+	MOVW.W	R0, -4(SP)
+	MOVW	$0, R0
+	MOVW.W	R0, -4(SP)	// push $0 as guard
+	BL	sys·newproc(SB)
+	MOVW	$12(SP), SP	// pop args and LR
+
+	// start this M
+	BL	mstart(SB)
+
+	MOVW	$0, R0
+	SWI	$0x00900001
+	B	_dep_dummy(SB)	// Never reached
 
-//	// start this M
-//	CALL	mstart(SB)
 
+TEXT mainstart(SB),7,$0
+	BL	main·init(SB)
+	BL	initdone(SB)
 	BL	main·main(SB)
-	MOVW	$99, R0
-	SWI	$0x00900001
+	MOVW	$0, R0
+	MOVW.W	R0, -4(SP)
+	MOVW.W	R14, -4(SP)	// Push link as well
+	BL	exit(SB)
+	MOVW	$8(SP), SP	// pop args and LR
+	RET
 
 // TODO(kaib): remove these once linker works properly
 // pull in dummy dependencies
-// TEXT _dep_dummy(SB),7,$0
-//	BL	sys·morestack(SB)
+TEXT _dep_dummy(SB),7,$0
+	BL	sys·morestack(SB)
+	BL	sys·morestackx(SB)
+	BL	_div(SB)
+	BL	_divu(SB)
+	BL	_mod(SB)
+	BL	_modu(SB)
+	BL	_modu(SB)
 
 
 TEXT	breakpoint(SB),7,$0
@@ -114,33 +106,38 @@ TEXT gosave(SB), 7, $0
 // support for morestack
 
 // return point when leaving new stack.
-// save AX, jmp to lesstack to switch back
+// save R0, jmp to lesstack to switch back
 TEXT	retfromnewstack(SB),7,$0
-	BL	abort(SB)
-//	MOVL	4(FS), BX	// m
-//	MOVL	AX, 12(BX)	// save AX in m->cret
-//	JMP	lessstack(SB)
+	MOVW	R0,12(R9)	// m->cret
+	B	lessstack(SB)
 
 // gogo, returning 2nd arg instead of 1
 TEXT gogoret(SB), 7, $0
-	BL	abort(SB)
-//	MOVL	8(SP), AX	// return 2nd arg
-//	MOVL	4(SP), BX	// gobuf
-//	MOVL	0(BX), SP	// restore SP
-//	MOVL	4(BX), BX
-//	MOVL	BX, 0(SP)	// put PC on the stack
-//	RET
+	MOVW	8(SP), R0	// return 2nd arg
+	MOVW	4(SP), R1	// gobuf
+	MOVW	0(R1), SP	// restore SP
+	MOVW	4(R1), PC	// restore PC
 
 TEXT setspgoto(SB), 7, $0
-	BL	abort(SB)
-//	MOVL	4(SP), AX	// SP
-//	MOVL	8(SP), BX	// fn to call
-//	MOVL	12(SP), CX	// fn to return
-//	MOVL	AX, SP
-//	PUSHL	CX
-//	JMP	BX
-//	POPL	AX	// not reached
-//	RET
+	MOVW	4(SP), R0	// SP
+	MOVW	8(SP), R1	// fn to call
+	MOVW	12(SP), R2	// fn to return into
+	MOVW	R2, R14		// restore LR
+	MOVW	R0, SP
+	MOVW	R1, PC		// goto
+
+// Optimization to make inline stack splitting code smaller
+// R0 is original first argument
+// R1 is arg_num << 24 | autosize >> 3
+TEXT sys·morestackx(SB), 7, $0
+	MOVW	R0, 4(SP)	// Save arg0
+	MOVW	R1<<8, R2
+	MOVW	R2>>5, R2
+	MOVW	R2, 4(R10)	// autooffset into g
+	MOVW	R1>>24, R2
+	MOVW	R2<<3, R2
+	MOVW	R2, 8(R10)	// argsize into g
+	B	sys·morestack(SB)
 
 // bool cas(int32 *val, int32 old, int32 new)
 // Atomically:
@@ -149,18 +146,26 @@ TEXT setspgoto(SB), 7, $0
 //		return 1;
 //	}else
 //		return 0;
-TEXT cas(SB), 7, $0
-	BL	abort(SB)
-//	MOVL	4(SP), BX
-//	MOVL	8(SP), AX
-//	MOVL	12(SP), CX
-//	LOCK
-//	CMPXCHGL	CX, 0(BX)
-//	JZ 3(PC)
-//	MOVL	$0, AX
-//	RET
-//	MOVL	$1, AX
-//	RET
+#define	LDREX(a,r)	WORD	$(0xe<<28|0x01900f9f | (a)<<16 | (r)<<12)
+#define	STREX(a,v,r)	WORD	$(0xe<<28|0x01800f90 | (a)<<16 | (r)<<12 | (v)<<0)
+
+TEXT	cas+0(SB),0,$12		/* r0 holds p */
+	MOVW	ov+4(FP), R1
+	MOVW	nv+8(FP), R2
+spin:
+/*	LDREX	0(R0),R3	*/
+	LDREX(0,3)
+	CMP.S	R3, R1
+	BNE	fail
+/*	STREX	0(R0),R2,R4	*/
+	STREX(0,2,4)
+	CMP.S	$0, R4
+	BNE	spin
+	MOVW	$1, R0
+	RET
+fail:
+	MOVW	$0, R0
+	RET
 
 // void jmpdefer(fn, sp);
 // called from deferreturn.
@@ -200,6 +205,10 @@ TEXT	sys·setcallerpc+0(SB),7,$0
 //	MOVL	BX, -4(AX)		// set calling pc
 //	RET
 
+TEXT emptyfunc(SB),0,$0
+	RET
+
 TEXT abort(SB),7,$0
-	WORD	$0
+	MOVW	$0, R0
+	MOVW	(R0), R1
 
-- 
2.48.1