6g: simplify trampoline by postponing load.

	TEXT tramp
		MOVQ 8(SP), AX
		ADDQ $40, AX
		MOVQ AX, 8(SP)
		JMP oldfunc

	is now

	TEXT tramp
		ADDQ $40, 8(SP)
		JMP oldfunc

	and if s/40/0/, then it simplifies to

	TEXT tramp
		JMP oldfunc

	(the tramp is still needed to satisfy
	symbol references from other object files)

R=ken
OCL=28377
CL=28381

src/cmd/6g/obj.c

@@ -553,7 +553,7 @@ void
 genembedtramp(Type *t, Sig *b)
 {
 	Sym *e;
-	int c, d, o;
+	int c, d, o, loaded;
 	Prog *p;
 	Type *f;
 
@@ -566,9 +566,6 @@ genembedtramp(Type *t, Sig *b)
 	fatal("genembedtramp %T.%s", t, b->name);
 
 out:
-	if(d == 0)
-		return;
-
 //	print("genembedtramp %d\n", d);
 //	print("	t    = %lT\n", t);
 //	print("	name = %s\n", b->name);
@@ -587,6 +584,15 @@ out:
 	p->from.scale = 7;
 //print("1. %P\n", p);
 
+	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(AMOVQ, N, N);
@@ -594,13 +600,8 @@ out:
 			p->from.offset = 8;
 			p->to.type = D_AX;
 //print("2. %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;
 		//MOVQ	o(AX), AX
 		p = pc;
 		gins(AMOVQ, N, N);
@@ -616,17 +617,31 @@ out:
 		gins(AADDQ, 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 = 8;
+		}
 //print("4. %P\n", p);
 	}
 
 	//MOVQ	AX, 8(SP)
+	if(loaded) {
 		p = pc;
 		gins(AMOVQ, N, N);
 		p->from.type = D_AX;
 		p->to.type = D_INDIR+D_SP;
 		p->to.offset = 8;
 //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.
+		p = pc;
+		gins(ANOP, N, N);
+	}
 
 	f = dotlist[0].field;
 	//JMP	main·*Sub_test2(SB)