cmd/gc: move genembedtramp into portable code

Requires adding new linker instruction
        RET	f(SB)
meaning return but then immediately call f.
This is what you'd use to implement a tail call after
fiddling with the arguments, but the compiler only
uses it in genwrapper.

This CL eliminates the copy-and-paste genembedtramp
functions from 5g/8g/6g and makes the code run on ARM
for the first time. It removes a small special case for function
generation, which should help Carl a bit, but at the same time
it does not bother to implement general tail call optimization,
which we do not want anyway.

Fixes #5627.

R=ken2
CC=golang-dev
https://golang.org/cl/10057044

src/cmd/5g/ggen.c

@@ -365,11 +365,19 @@ cgen_aret(Node *n, Node *res)
 void
 cgen_ret(Node *n)
 {
+	Prog *p;
+
 	genlist(n->list);		// copy out args
-	if(hasdefer || curfn->exit)
+	if(hasdefer || curfn->exit) {
 		gjmp(retpc);
-	else
-		gins(ARET, N, N);
+		return;
+	}
+	p = gins(ARET, N, N);
+	if(n->op == ORETJMP) {
+		p->to.name = D_EXTERN;
+		p->to.type = D_CONST;
+		p->to.sym = n->left->sym;
+	}
 }
 
 /*

src/cmd/5g/gobj.c

@@ -518,88 +518,6 @@ dsymptr(Sym *s, int off, Sym *x, int xoff)
 	return off;
 }
 
-
-void
-genembedtramp(Type *rcvr, Type *method, Sym *newnam, int iface)
-{
-	// TODO(kaib): re-implement genembedtramp
-	genwrapper(rcvr, method, newnam, iface);
-/*
-	Sym *e;
-	int c, d, o;
-	Prog *p;
-	Type *f;
-
-	e = method->sym;
-	for(d=0; d<nelem(dotlist); d++) {
-		c = adddot1(e, rcvr, d, nil, 0);
-		if(c == 1)
-			goto out;
-	}
-	fatal("genembedtramp %T.%S", rcvr, method->sym);
-
-out:
-	newplist()->name = newname(newnam);
-
-	//TEXT	main·S_test2(SB),7,$0
-	p = pc;
-	gins(ATEXT, N, N);
-	p->from.type = D_OREG;
-	p->from.name = D_EXTERN;
-	p->from.sym = newnam;
-	p->to.type = D_CONST2;
-	p->to.offset = 0;  // stack size
-	p->to.offset2 = rnd(method->type->argwid, widthptr);  // argument size
-	p->reg = 7;  // textflag
-	p->to.reg = NREG;
-//print("1. %P\n", p);
-
-	o = 0;
-	for(c=d-1; c>=0; c--) {
-		f = dotlist[c].field;
-		o += f->width;
-		if(!isptr[f->type->etype])
-			continue;
-
-		//MOVW	o(R0), R0
-		p = pc;
-		gins(AMOVW, N, N);
-		p->from.type = D_OREG;
-		p->from.reg = REGARG;
-		p->from.offset = o;
-		p->to.type = D_REG;
-		p->to.reg = REGARG;
-//print("2. %P\n", p);
-		o = 0;
-	}
-	if(o != 0) {
-		//MOVW	$XX(R0), R0
-		p = pc;
-		gins(AMOVW, N, N);
-		p->from.type = D_CONST;
-		p->from.reg = REGARG;
-		p->from.offset = o;
-		p->to.type = D_REG;
-		p->to.reg = REGARG;
-//print("3. %P\n", p);
-	}
-
-	f = dotlist[0].field;
-	//B	main·*Sub_test2(SB)
-	if(isptr[f->type->etype])
-		f = f->type;
-	p = pc;
-	gins(AB, N, N);
-	p->to.type = D_OREG;
-	p->to.reg = NREG;
-	p->to.name = D_EXTERN;
-	p->to.sym = methodsym(method->sym, ptrto(f->type), 0);
-//print("4. %P\n", p);
-
-	pc->as = ARET;	// overwrite AEND
-*/
-}
-
 void
 nopout(Prog *p)
 {

src/cmd/5l/noop.c

@@ -62,7 +62,7 @@ linkcase(Prog *casep)
 void
 noops(void)
 {
-	Prog *p, *q, *q1;
+	Prog *p, *q, *q1, *q2;
 	int o;
 	Prog *pmorestack;
 	Sym *symmorestack;
@@ -343,9 +343,14 @@ noops(void)
 					if(!autosize) {
 						p->as = AB;
 						p->from = zprg.from;
+						if(p->to.sym) { // retjmp
+							p->to.type = D_BRANCH;
+							p->cond = p->to.sym->text;
+						} else {
 							p->to.type = D_OREG;
 							p->to.offset = 0;
 							p->to.reg = REGLINK;
+						}
 						break;
 					}
 				}
@@ -359,6 +364,17 @@ noops(void)
 				// If there are instructions following
 				// this ARET, they come from a branch
 				// with the same stackframe, so no spadj.
+				
+				if(p->to.sym) { // retjmp
+					p->to.reg = REGLINK;
+					q2 = appendp(p);
+					q2->as = AB;
+					q2->to.type = D_BRANCH;
+					q2->to.sym = p->to.sym;
+					q2->cond = p->to.sym->text;
+					p->to.sym = nil;
+					p = q2;
+				}
 				break;
 	
 			case AADD:

src/cmd/6g/ggen.c

@@ -325,11 +325,18 @@ cgen_aret(Node *n, Node *res)
 void
 cgen_ret(Node *n)
 {
+	Prog *p;
+
 	genlist(n->list);		// copy out args
-	if(hasdefer || curfn->exit)
+	if(hasdefer || curfn->exit) {
 		gjmp(retpc);
-	else
-		gins(ARET, N, N);
+		return;
+	}
+	p = gins(ARET, N, N);
+	if(n->op == ORETJMP) {
+		p->to.type = D_EXTERN;
+		p->to.sym = n->left->sym;
+	}
 }
 
 /*

src/cmd/6g/gobj.c

@@ -497,115 +497,6 @@ dsymptr(Sym *s, int off, Sym *x, int xoff)
 	return off;
 }
 
-void
-genembedtramp(Type *rcvr, Type *method, Sym *newnam, int iface)
-{
-	Sym *e;
-	int c, d, mov, add, loaded;
-	int64 o;
-	Prog *p;
-	Type *f;
-	
-	USED(iface);
-
-	if(0 && debug['r'])
-		print("genembedtramp %T %T %S\n", rcvr, method, newnam);
-
-	e = method->sym;
-	for(d=0; d<nelem(dotlist); d++) {
-		c = adddot1(e, rcvr, d, nil, 0);
-		if(c == 1)
-			goto out;
-	}
-	fatal("genembedtramp %T.%S", rcvr, method->sym);
-
-out:
-	newplist()->name = newname(newnam);
-
-	//TEXT	main·S_test2(SB),7,$0
-	p = pc;
-	gins(ATEXT, N, N);
-	p->from.type = D_EXTERN;
-	p->from.sym = newnam;
-	p->to.type = D_CONST;
-	p->to.offset = 0; // stack size
-	p->to.offset |= rnd(method->type->argwid, widthptr) << 32;  // argument size
-	p->from.scale = 7;  // textflag
-//print("1. %P\n", p);
-
-	mov = AMOVQ;
-	add = AADDQ;
-	loaded = 0;
-	o = 0;
-	for(c=d-1; c>=0; c--) {
-		f = dotlist[c].field;
-		o += f->width;
-		if(!isptr[f->type->etype])
-			continue;
-		if(!loaded) {
-			loaded = 1;
-			//MOVQ	8(SP), AX
-			p = pc;
-			gins(mov, N, N);
-			p->from.type = D_INDIR+D_SP;
-			p->from.offset = widthptr;
-			p->to.type = D_AX;
-//print("2. %P\n", p);
-		}
-
-		//MOVQ	o(AX), AX
-		p = pc;
-		gins(mov, N, N);
-		p->from.type = D_INDIR+D_AX;
-		p->from.offset = o;
-		p->to.type = D_AX;
-//print("3. %P\n", p);
-		o = 0;
-	}
-	if(o != 0) {
-		//ADDQ	$XX, AX
-		p = pc;
-		gins(add, N, N);
-		p->from.type = D_CONST;
-		p->from.offset = o;
-		if(loaded)
-			p->to.type = D_AX;
-		else {
-			p->to.type = D_INDIR+D_SP;
-			p->to.offset = widthptr;
-		}
-//print("4. %P\n", p);
-	}
-
-	//MOVQ	AX, 8(SP)
-	if(loaded) {
-		p = pc;
-		gins(mov, N, N);
-		p->from.type = D_AX;
-		p->to.type = D_INDIR+D_SP;
-		p->to.offset = widthptr;
-//print("5. %P\n", p);
-	} else {
-		// TODO(rsc): obviously this is unnecessary,
-		// but 6l has a bug, and it can't handle
-		// JMP instructions too close to the top of
-		// a new function.
-		gins(ANOP, N, N);
-	}
-
-	f = dotlist[0].field;
-	//JMP	main·*Sub_test2(SB)
-	if(isptr[f->type->etype])
-		f = f->type;
-	p = pc;
-	gins(AJMP, N, N);
-	p->to.type = D_EXTERN;
-	p->to.sym = methodsym(method->sym, ptrto(f->type), 0);
-//print("6. %P\n", p);
-
-	pc->as = ARET;	// overwrite AEND
-}
-
 void
 nopout(Prog *p)
 {

src/cmd/6l/pass.c

@@ -318,7 +318,7 @@ patch(void)
 			if(p->to.type == D_INDIR+D_GS)
 				p->to.type = D_INDIR+D_FS;
 		}
-		if(p->as == ACALL || (p->as == AJMP && p->to.type != D_BRANCH)) {
+		if(p->as == ACALL || (p->as == AJMP && p->to.type != D_BRANCH) || (p->as == ARET && p->to.sym != nil)) {
 			s = p->to.sym;
 			if(s) {
 				if(debug['c'])
@@ -747,6 +747,8 @@ dostkoff(void)
 				// the cleanup.
 				p->spadj = +autoffset;
 			}
+			if(p->to.sym) // retjmp
+				p->as = AJMP;
 		}
 	}
 }

src/cmd/8g/ggen.c

@@ -360,11 +360,18 @@ cgen_aret(Node *n, Node *res)
 void
 cgen_ret(Node *n)
 {
+	Prog *p;
+
 	genlist(n->list);		// copy out args
-	if(retpc)
+	if(retpc) {
 		gjmp(retpc);
-	else
-		gins(ARET, N, N);
+		return;
+	}
+	p = gins(ARET, N, N);
+	if(n->op == ORETJMP) {
+		p->to.type = D_EXTERN;
+		p->to.sym = n->left->sym;
+	}
 }
 
 /*

src/cmd/8g/gobj.c

@@ -507,112 +507,6 @@ dsymptr(Sym *s, int off, Sym *x, int xoff)
 	return off;
 }
 
-void
-genembedtramp(Type *rcvr, Type *method, Sym *newnam, int iface)
-{
-	Sym *e;
-	int c, d, o, mov, add, loaded;
-	Prog *p;
-	Type *f;
-
-	USED(iface);
-
-	e = method->sym;
-	for(d=0; d<nelem(dotlist); d++) {
-		c = adddot1(e, rcvr, d, nil, 0);
-		if(c == 1)
-			goto out;
-	}
-	fatal("genembedtramp %T.%S", rcvr, method->sym);
-
-out:
-	newplist()->name = newname(newnam);
-
-	//TEXT	main·S_test2(SB),7,$0
-	p = pc;
-	gins(ATEXT, N, N);
-	p->from.type = D_EXTERN;
-	p->from.sym = newnam;
-	p->to.type = D_CONST;
-	p->to.offset = 0;  // stack skize
-	p->to.offset2 = rnd(method->type->argwid, widthptr);  // argument size
-	p->from.scale = 7;  // textflag
-//print("1. %P\n", p);
-
-	mov = AMOVL;
-	add = AADDL;
-
-	loaded = 0;
-	o = 0;
-	for(c=d-1; c>=0; c--) {
-		f = dotlist[c].field;
-		o += f->width;
-		if(!isptr[f->type->etype])
-			continue;
-		if(!loaded) {
-			loaded = 1;
-			//MOVL	4(SP), AX
-			p = pc;
-			gins(mov, N, N);
-			p->from.type = D_INDIR+D_SP;
-			p->from.offset = widthptr;
-			p->to.type = D_AX;
-//print("2. %P\n", p);
-		}
-
-		//MOVL	o(AX), AX
-		p = pc;
-		gins(mov, N, N);
-		p->from.type = D_INDIR+D_AX;
-		p->from.offset = o;
-		p->to.type = D_AX;
-//print("3. %P\n", p);
-		o = 0;
-	}
-	if(o != 0) {
-		//ADDL	$XX, AX
-		p = pc;
-		gins(add, N, N);
-		p->from.type = D_CONST;
-		p->from.offset = o;
-		if(loaded)
-			p->to.type = D_AX;
-		else {
-			p->to.type = D_INDIR+D_SP;
-			p->to.offset = widthptr;
-		}
-//print("4. %P\n", p);
-	}
-
-	//MOVL	AX, 4(SP)
-	if(loaded) {
-		p = pc;
-		gins(mov, N, N);
-		p->from.type = D_AX;
-		p->to.type = D_INDIR+D_SP;
-		p->to.offset = widthptr;
-//print("5. %P\n", p);
-	} else {
-		// TODO(rsc): obviously this is unnecessary,
-		// but 6l has a bug, and it can't handle
-		// JMP instructions too close to the top of
-		// a new function.
-		gins(ANOP, N, N);
-	}
-
-	f = dotlist[0].field;
-	//JMP	main·*Sub_test2(SB)
-	if(isptr[f->type->etype])
-		f = f->type;
-	p = pc;
-	gins(AJMP, N, N);
-	p->to.type = D_EXTERN;
-	p->to.sym = methodsym(method->sym, ptrto(f->type), 0);
-//print("6. %P\n", p);
-
-	pc->as = ARET;	// overwrite AEND
-}
-
 void
 nopout(Prog *p)
 {

src/cmd/8l/pass.c

@@ -329,7 +329,7 @@ patch(void)
 					p->from.offset = 0;
 				}
 			}
-			if((p->as == ACALL && p->to.type != D_BRANCH) || (p->as == AJMP && p->to.type != D_BRANCH)) {
+			if((p->as == ACALL && p->to.type != D_BRANCH) || (p->as == AJMP && p->to.type != D_BRANCH) || (p->as == ARET && p->to.sym != nil)) {
 				s = p->to.sym;
 				if(p->to.type == D_INDIR+D_ADDR) {
 					 /* skip check if this is an indirect call (CALL *symbol(SB)) */
@@ -692,6 +692,8 @@ dostkoff(void)
 				// the cleanup.
 				p->spadj = +autoffset;
 			}
+			if(p->to.sym) // retjmp
+				p->as = AJMP;
 		}
 	}
 }

src/cmd/gc/fmt.c

@@ -870,6 +870,10 @@ stmtfmt(Fmt *f, Node *n)
 		fmtprint(f, "return %,H", n->list);
 		break;
 
+	case ORETJMP:
+		fmtprint(f, "retjmp %S", n->sym);
+		break;
+	
 	case OPROC:
 		fmtprint(f, "go %N", n->left);
 		break;

src/cmd/gc/gen.c

@@ -489,6 +489,7 @@ gen(Node *n)
 		break;
 
 	case ORETURN:
+	case ORETJMP:
 		cgen_ret(n);
 		break;
 	

src/cmd/gc/go.h

@@ -581,6 +581,7 @@ enum
 	OHMUL, // high mul: AMUL/AIMUL for unsigned/signed (OMUL uses AIMUL for both).
 	OLROT,	// left rotate: AROL.
 	ORROTC, // right rotate-carry: ARCR.
+	ORETJMP,	// return to other function
 
 	OEND,
 };
@@ -1461,7 +1462,6 @@ void	fixautoused(Prog*);
 void	gdata(Node*, Node*, int);
 void	gdatacomplex(Node*, Mpcplx*);
 void	gdatastring(Node*, Strlit*);
-void	genembedtramp(Type*, Type*, Sym*, int iface);
 void	ggloblnod(Node *nam);
 void	ggloblsym(Sym *s, int32 width, int dupok, int rodata);
 Prog*	gjmp(Prog*);

src/cmd/gc/inl.c

@@ -197,6 +197,7 @@ ishairy(Node *n, int *budget)
 	case ODEFER:
 	case ODCLTYPE:  // can't print yet
 	case ODCLCONST:  // can't print yet
+	case ORETJMP:
 		return 1;
 
 		break;

src/cmd/gc/order.c

@@ -218,6 +218,7 @@ orderstmt(Node *n, NodeList **out)
 	case_OFALL:
 	case OGOTO:
 	case OLABEL:
+	case ORETJMP:
 		// Special: n->left is not an expression; save as is.
 		*out = list(*out, n);
 		break;

src/cmd/gc/racewalk.c

@@ -364,6 +364,7 @@ racewalknode(Node **np, NodeList **init, int wr, int skip)
 	case OIF:
 	case OCALLMETH:
 	case ORETURN:
+	case ORETJMP:
 	case OSWITCH:
 	case OSELECT:
 	case OEMPTY:

src/cmd/gc/reflect.c

@@ -208,15 +208,7 @@ methods(Type *t)
 		if(!(a->isym->flags & SymSiggen)) {
 			a->isym->flags |= SymSiggen;
 			if(!eqtype(this, it) || this->width < types[tptr]->width) {
-				// Is okay to call genwrapper here always,
-				// but we can generate more efficient code
-				// using genembedtramp if all that is necessary
-				// is a pointer adjustment and a JMP.
 				compiling_wrappers = 1;
-				if(isptr[it->etype] && isptr[this->etype]
-				&& f->embedded && !isifacemethod(f->type))
-					genembedtramp(it, f, a->isym, 1);
-				else
 				genwrapper(it, f, a->isym, 1);
 				compiling_wrappers = 0;
 			}
@@ -226,10 +218,6 @@ methods(Type *t)
 			a->tsym->flags |= SymSiggen;
 			if(!eqtype(this, t)) {
 				compiling_wrappers = 1;
-				if(isptr[t->etype] && isptr[this->etype]
-				&& f->embedded && !isifacemethod(f->type))
-					genembedtramp(t, f, a->tsym, 0);
-				else
 				genwrapper(t, f, a->tsym, 0);
 				compiling_wrappers = 0;
 			}

src/cmd/gc/subr.c

@@ -2495,13 +2495,13 @@ structargs(Type **tl, int mustname)
 void
 genwrapper(Type *rcvr, Type *method, Sym *newnam, int iface)
 {
-	Node *this, *fn, *call, *n, *t, *pad;
+	Node *this, *fn, *call, *n, *t, *pad, *dot, *as;
 	NodeList *l, *args, *in, *out;
-	Type *tpad;
+	Type *tpad, *methodrcvr;
 	int isddd;
 	Val v;
 
-	if(debug['r'])
+	if(0 && debug['r'])
 		print("genwrapper rcvrtype=%T method=%T newnam=%S\n",
 			rcvr, method, newnam);
 
@@ -2547,8 +2547,10 @@ genwrapper(Type *rcvr, Type *method, Sym *newnam, int iface)
 		isddd = l->n->left->isddd;
 	}
 	
+	methodrcvr = getthisx(method->type)->type->type;
+
 	// generate nil pointer check for better error
-	if(isptr[rcvr->etype] && rcvr->type == getthisx(method->type)->type->type) {
+	if(isptr[rcvr->etype] && rcvr->type == methodrcvr) {
 		// generating wrapper from *T to T.
 		n = nod(OIF, N, N);
 		n->ntest = nod(OEQ, this->left, nodnil());
@@ -2568,8 +2570,22 @@ genwrapper(Type *rcvr, Type *method, Sym *newnam, int iface)
 		fn->nbody = list(fn->nbody, n);
 	}
 	
+	dot = adddot(nod(OXDOT, this->left, newname(method->sym)));
+	
 	// generate call
-	call = nod(OCALL, adddot(nod(OXDOT, this->left, newname(method->sym))), N);
+	if(isptr[rcvr->etype] && isptr[methodrcvr->etype] && method->embedded && !isifacemethod(method->type)) {
+		// skip final .M
+		dot = dot->left;
+		if(!isptr[dotlist[0].field->type->etype])
+			dot = nod(OADDR, dot, N);
+		as = nod(OAS, this->left, nod(OCONVNOP, dot, N));
+		as->right->type = rcvr;
+		fn->nbody = list(fn->nbody, as);
+		n = nod(ORETJMP, N, N);
+		n->left = newname(methodsym(method->sym, methodrcvr, 0));
+		fn->nbody = list(fn->nbody, n);
+	} else {
+		call = nod(OCALL, dot, N);
 		call->list = args;
 		call->isddd = isddd;
 		if(method->type->outtuple > 0) {
@@ -2578,6 +2594,7 @@ genwrapper(Type *rcvr, Type *method, Sym *newnam, int iface)
 			call = n;
 		}
 		fn->nbody = list(fn->nbody, call);
+	}
 
 	if(0 && debug['r'])
 		dumplist("genwrapper body", fn->nbody);

src/cmd/gc/typecheck.c

@@ -1652,6 +1652,10 @@ reswitch:
 		typecheckaste(ORETURN, nil, 0, getoutargx(curfn->type), n->list, "return argument");
 		goto ret;
 	
+	case ORETJMP:
+		ok |= Etop;
+		goto ret;
+
 	case OSELECT:
 		ok |= Etop;
 		typecheckselect(n);
@@ -3282,6 +3286,7 @@ isterminating(NodeList *l, int top)
 
 	case OGOTO:
 	case ORETURN:
+	case ORETJMP:
 	case OPANIC:
 	case OXFALL:
 		return 1;

src/cmd/gc/walk.c

@@ -282,6 +282,9 @@ walkstmt(Node **np)
 		n->list = ll;
 		break;
 
+	case ORETJMP:
+		break;
+
 	case OSELECT:
 		walkselect(n);
 		break;