closures - runtime and debugger support, test case

R=r
DELTA=257  (250 added, 1 deleted, 6 changed)
OCL=24509
CL=24565

src/libmach_amd64/8db.c

@@ -167,7 +167,23 @@ i386trace(Map *map, uvlong pc, uvlong sp, uvlong link, Tracer trace)
 	USED(link);
 	osp = 0;
 	i = 0;
-	while(findsym(pc, CTEXT, &s)) {
+
+	for(;;) {
+		if(!findsym(pc, CTEXT, &s)) {
+			// check for closure return sequence
+			uchar buf[8];
+			if(get1(map, pc, buf, 8) < 0)
+				break;
+			// ADDQ $xxx, SP; RET
+			if(buf[0] != 0x48 || buf[1] != 0x81 || buf[2] != 0xc4 || buf[7] != 0xc3)
+				break;
+			sp += buf[3] | (buf[4]<<8) | (buf[5]<<16) | (buf[6]<<24);
+			if(geta(map, sp, &pc) < 0)
+				break;
+			sp += mach->szaddr;
+			continue;
+		}
+
 		if (osp == sp)
 			break;
 		osp = sp;
@@ -186,7 +202,6 @@ i386trace(Map *map, uvlong pc, uvlong sp, uvlong link, Tracer trace)
 			(*trace)(map, pc, sp +  8, &s1);
 			sp += 16;  // two irrelevant calls on stack - morestack, plus the call morestack made
 			continue;
-			break;
 		}
 		s1 = s;
 		pc1 = 0;

src/runtime/malloc.c

@@ -203,7 +203,7 @@ void*
 SysAlloc(uintptr n)
 {
 	mstats.sys += n;
-	return sys_mmap(nil, n, PROT_READ|PROT_WRITE, MAP_ANON|MAP_PRIVATE, 0, 0);
+	return sys_mmap(nil, n, PROT_READ|PROT_WRITE|PROT_EXEC, MAP_ANON|MAP_PRIVATE, 0, 0);
 }
 
 void

src/runtime/mem.c

@@ -29,7 +29,7 @@ brk(uint32 n)
 {
 	byte *v;
 
-	v = sys_mmap(nil, n, PROT_READ|PROT_WRITE, MAP_ANON|MAP_PRIVATE, 0, 0);
+	v = sys_mmap(nil, n, PROT_READ|PROT_WRITE|PROT_EXEC, MAP_ANON|MAP_PRIVATE, 0, 0);
 	m->mem.nmmap += n;
 	return v;
 }
@@ -64,7 +64,7 @@ oldmal(uint32 n)
 			// so we have to call sys_mmap directly - it is written
 			// in assembly and tagged not to grow the stack.
 			m->mem.hunk =
-				sys_mmap(nil, NHUNK, PROT_READ|PROT_WRITE,
+				sys_mmap(nil, NHUNK, PROT_READ|PROT_WRITE|PROT_EXEC,
 					MAP_ANON|MAP_PRIVATE, 0, 0);
 			m->mem.nhunk = NHUNK;
 			m->mem.nmmap += NHUNK;

src/runtime/rt2_amd64.c

@@ -15,6 +15,7 @@ traceback(byte *pc0, byte *sp, G *g)
 	uint64 pc;
 	int32 i, n;
 	Func *f;
+	byte *p;
 
 	pc = (uint64)pc0;
 
@@ -36,6 +37,16 @@ traceback(byte *pc0, byte *sp, G *g)
 		}
 		f = findfunc(pc);
 		if(f == nil) {
+			// dangerous, but poke around to see if it is a closure
+			p = (byte*)pc;
+			// ADDQ $xxx, SP; RET
+			if(p[0] == 0x48 && p[1] == 0x81 && p[2] == 0xc4 && p[7] == 0xc3) {
+				sp += *(uint32*)(p+3) + 8;
+				pc = *(uint64*)(sp - 8);
+				if(pc <= 0x1000)
+					return;
+				continue;
+			}
 			printf("%p unknown pc\n", pc);
 			return;
 		}
@@ -76,6 +87,7 @@ sys·Caller(int32 n, uint64 retpc, string retfile, int32 retline, bool retbool)
 {
 	uint64 pc;
 	byte *sp;
+	byte *p;
 	Stktop *stk;
 	Func *f;
 
@@ -110,9 +122,19 @@ sys·Caller(int32 n, uint64 retpc, string retfile, int32 retline, bool retbool)
 		else
 			sp += f->frame;
 
+	loop:
 		pc = *(uint64*)(sp-8);
-		if(pc <= 0x1000 || (f = findfunc(pc)) == nil)
+		if(pc <= 0x1000 || (f = findfunc(pc)) == nil) {
+			// dangerous, but let's try this.
+			// see if it is a closure.
+			p = (byte*)pc;
+			// ADDQ $xxx, SP; RET
+			if(p[0] == 0x48 && p[1] == 0x81 && p[2] == 0xc4 && p[7] == 0xc3) {
+				sp += *(uint32*)(p+3) + 8;
+				goto loop;
+			}
 			goto error;
+		}
 	}
 
 	retpc = pc;
@@ -124,3 +146,119 @@ sys·Caller(int32 n, uint64 retpc, string retfile, int32 retline, bool retbool)
 	FLUSH(&retline);
 	FLUSH(&retbool);
 }
+
+#pragma textflag 7
+// func closure(siz int32,
+//	fn func(arg0, arg1, arg2 *ptr, callerpc uintptr, xxx) yyy,
+//	arg0, arg1, arg2 *ptr) (func(xxx) yyy)
+void
+sys·closure(int32 siz, byte *fn, byte *arg0)
+{
+	byte *p, *q, **ret;
+	int32 i, n;
+	int64 pcrel;
+
+	if(siz < 0 || siz%8 != 0)
+		throw("bad closure size");
+
+	ret = (byte**)((byte*)&arg0 + siz);
+
+	if(siz > 100) {
+		// TODO(rsc): implement stack growth preamble?
+		throw("closure too big");
+	}
+
+	// compute size of new fn.
+	// must match code laid out below.
+	n = 7+10+3;	// SUBQ MOVQ MOVQ
+	if(siz <= 4*8)
+		n += 2*siz/8;	// MOVSQ MOVSQ...
+	else
+		n += 7+3;	// MOVQ REP MOVSQ
+	n += 12;	// CALL worst case; sometimes only 5
+	n += 7+1;	// ADDQ RET
+
+	// store args aligned after code, so gc can find them.
+	n += siz;
+	if(n%8)
+		n += 8 - n%8;
+
+	p = mal(n);
+	*ret = p;
+	q = p + n - siz;
+	mcpy(q, (byte*)&arg0, siz);
+
+	// SUBQ $siz, SP
+	*p++ = 0x48;
+	*p++ = 0x81;
+	*p++ = 0xec;
+	*(uint32*)p = siz;
+	p += 4;
+
+	// MOVQ $q, SI
+	*p++ = 0x48;
+	*p++ = 0xbe;
+	*(byte**)p = q;
+	p += 8;
+
+	// MOVQ SP, DI
+	*p++ = 0x48;
+	*p++ = 0x89;
+	*p++ = 0xe7;
+
+	if(siz <= 4*8) {
+		for(i=0; i<siz; i+=8) {
+			// MOVSQ
+			*p++ = 0x48;
+			*p++ = 0xa5;
+		}
+	} else {
+		// MOVQ $(siz/8), CX  [32-bit immediate siz/8]
+		*p++ = 0x48;
+		*p++ = 0xc7;
+		*p++ = 0xc1;
+		*(uint32*)p = siz/8;
+		p += 4;
+
+		// REP; MOVSQ
+		*p++ = 0xf3;
+		*p++ = 0x48;
+		*p++ = 0xa5;
+	}
+
+
+	// call fn
+	pcrel = fn - (p+5);
+	if((int32)pcrel == pcrel) {
+		// can use direct call with pc-relative offset
+		// CALL fn
+		*p++ = 0xe8;
+		*(int32*)p = pcrel;
+		p += 4;
+	} else {
+		// MOVQ $fn, CX  [64-bit immediate fn]
+		*p++ = 0x48;
+		*p++ = 0xb9;
+		*(byte**)p = fn;
+		p += 8;
+
+		// CALL *CX
+		*p++ = 0xff;
+		*p++ = 0xd1;
+	}
+
+	// ADDQ $siz, SP
+	*p++ = 0x48;
+	*p++ = 0x81;
+	*p++ = 0xc4;
+	*(uint32*)p = siz;
+	p += 4;
+
+	// RET
+	*p++ = 0xc3;
+
+	if(p > q)
+		throw("bad math in sys.closure");
+}
+
+

test/closure.go

@@ -0,0 +1,88 @@
+// $G $D/$F.go && $L $F.$A && ./$A.out
+
+// Copyright 2009 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package main
+
+var c = make(chan int);
+
+func check(a []int) {
+	for i := 0; i < len(a); i++ {
+		n := <-c;
+		if n != a[i] {
+			panicln("want", a[i], "got", n, "at", i);
+		}
+	}
+}
+
+func f() {
+	var i, j int;
+
+	i = 1;
+	j = 2;
+	f := func() {
+		c <- i;
+		i = 4;
+		g := func() {
+			c <- i;
+			c <- j;
+		};
+		g();
+		c <- i;
+	};
+	j = 5;
+	f();
+}
+
+// Accumulator generator
+func accum(n int) (func(int) int) {
+	return func(i int) int {
+		n += i;
+		return n;
+	}
+}
+
+func g(a, b func(int) int) {
+	c <- a(2);
+	c <- b(3);
+	c <- a(4);
+	c <- b(5);
+}
+
+func h() {
+	var x8 byte = 100;
+	var x64 int64 = 200;
+
+	c <- int(x8);
+	c <- int(x64);
+	f := func(z int) {
+		g := func() {
+			c <- int(x8);
+			c <- int(x64);
+			c <- z;
+		};
+		g();
+		c <- int(x8);
+		c <- int(x64);
+		c <- int(z);
+	};
+	x8 = 101;
+	x64 = 201;
+	f(500);
+}
+
+
+func main() {
+	go f();
+	check([]int{1,4,5,4});
+
+	a := accum(0);
+	b := accum(1);
+	go g(a, b);
+	check([]int{2,4,6,9});
+
+	go h();
+	check([]int{100,200,101,201,500,101,201,500});
+}

test/stack.go

@@ -5,7 +5,7 @@
 // license that can be found in the LICENSE file.
 
 // Try to tickle stack splitting bugs by doing
-// go and defer at different stack depths.
+// go, defer, and closure calls at different stack depths.
 
 package main
 
@@ -38,6 +38,18 @@ func recur(n int) {
 	if s != len(t) {
 		panicln("bad go", s);
 	}
+	f := func(t T) int {
+		s := 0;
+		for i := 0; i < len(t); i++ {
+			s += t[i];
+		}
+		s += n;
+		return s;
+	};
+	s = f(t);
+	if s != len(t) + n {
+		panicln("bad func", s, "at level", n);
+	}
 	if n > 0 {
 		recur(n-1);
 	}