cmd/gc: shorten temporary lifetimes when possible

The new channel and map runtime routines take pointers
to values, typically temporaries. Without help, the compiler
cannot tell when those temporaries stop being needed,
because it isn't sure what happened to the pointer.
Arrange to insert explicit VARKILL instructions for these
temporaries so that the liveness analysis can avoid seeing
them as "ambiguously live".

The change is made in order.c, which was already in charge of
introducing temporaries to preserve the order-of-evaluation
guarantees. Now its job has expanded to include introducing
temporaries as needed by runtime routines, and then also
inserting the VARKILL annotations for all these temporaries,
so that their lifetimes can be shortened.

In order to do its job for the map runtime routines, order.c arranges
that all map lookups or map assignments have the form:

        x = m[k]
        x, y = m[k]
        m[k] = x

where x, y, and k are simple variables (often temporaries).
Likewise, receiving from a channel is now always:

        x = <-c

In order to provide the map guarantee, order.c is responsible for
rewriting x op= y into x = x op y, so that m[k] += z becomes

        t = m[k]
        t2 = t + z
        m[k] = t2

While here, fix a few bugs in order.c's traversal: it was failing to
walk into select and switch case bodies, so order of evaluation
guarantees were not preserved in those situations.
Added tests to test/reorder2.go.

Fixes #7671.

In gc/popt's temporary-merging optimization, allow merging
of temporaries with their address taken as long as the liveness
ranges do not intersect. (There is a good chance of that now
that we have VARKILL annotations to limit the liveness range.)

Explicitly killing temporaries cuts the number of ambiguously
live temporaries that must be zeroed in the godoc binary from
860 to 711, or -17%. There is more work to be done, but this
is a good checkpoint.

Update #7345

LGTM=khr
R=khr
CC=golang-codereviews
https://golang.org/cl/81940043

src/cmd/5g/ggen.c

@@ -89,7 +89,7 @@ void
 markautoused(Prog* p)
 {
 	for (; p; p = p->link) {
-		if (p->as == ATYPE || p->as == AVARDEF)
+		if (p->as == ATYPE || p->as == AVARDEF || p->as == AVARKILL)
 			continue;
 
 		if (p->from.node)
@@ -111,7 +111,7 @@ fixautoused(Prog* p)
 			*lp = p->link;
 			continue;
 		}
-		if (p->as == AVARDEF && p->to.node && !p->to.node->used) {
+		if ((p->as == AVARDEF || p->as == AVARKILL) && p->to.node && !p->to.node->used) {
 			// Cannot remove VARDEF instruction, because - unlike TYPE handled above -
 			// VARDEFs are interspersed with other code, and a jump might be using the
 			// VARDEF as a target. Replace with a no-op instead. A later pass will remove

src/cmd/5g/peep.c

@@ -287,7 +287,7 @@ subprop(Flow *r0)
 		if(uniqs(r) == nil)
 			break;
 		p = r->prog;
-		if(p->as == AVARDEF)
+		if(p->as == AVARDEF || p->as == AVARKILL)
 			continue;
 		proginfo(&info, p);
 		if(info.flags & Call)
@@ -1167,6 +1167,7 @@ copyu(Prog *p, Adr *v, Adr *s)
 	case APCDATA:
 	case AFUNCDATA:
 	case AVARDEF:
+	case AVARKILL:
 		return 0;
 	}
 }

src/cmd/5g/prog.c

@@ -30,6 +30,7 @@ static ProgInfo progtable[ALAST] = {
 	[AUSEFIELD]=	{OK},
 	[ACHECKNIL]=	{LeftRead},
 	[AVARDEF]=	{Pseudo | RightWrite},
+	[AVARKILL]=	{Pseudo | RightWrite},
 
 	// NOP is an internal no-op that also stands
 	// for USED and SET annotations, not the Intel opcode.

src/cmd/5g/reg.c

@@ -213,7 +213,7 @@ regopt(Prog *firstp)
 
 	for(r = firstr; r != R; r = (Reg*)r->f.link) {
 		p = r->f.prog;
-		if(p->as == AVARDEF)
+		if(p->as == AVARDEF || p->as == AVARKILL)
 			continue;
 		proginfo(&info, p);
 

src/cmd/5l/5.out.h

@@ -199,6 +199,7 @@ enum	as
 	APCDATA,
 	ACHECKNIL,
 	AVARDEF,
+	AVARKILL,
 
 	AMRC, // MRC/MCR
 

src/cmd/6g/ggen.c

@@ -83,7 +83,7 @@ void
 markautoused(Prog* p)
 {
 	for (; p; p = p->link) {
-		if (p->as == ATYPE || p->as == AVARDEF)
+		if (p->as == ATYPE || p->as == AVARDEF || p->as == AVARKILL)
 			continue;
 
 		if (p->from.node)
@@ -105,7 +105,7 @@ fixautoused(Prog *p)
 			*lp = p->link;
 			continue;
 		}
-		if (p->as == AVARDEF && p->to.node && !p->to.node->used) {
+		if ((p->as == AVARDEF || p->as == AVARKILL) && p->to.node && !p->to.node->used) {
 			// Cannot remove VARDEF instruction, because - unlike TYPE handled above -
 			// VARDEFs are interspersed with other code, and a jump might be using the
 			// VARDEF as a target. Replace with a no-op instead. A later pass will remove

src/cmd/6g/peep.c

@@ -573,7 +573,7 @@ subprop(Flow *r0)
 			break;
 		}
 		p = r->prog;
-		if(p->as == AVARDEF)
+		if(p->as == AVARDEF || p->as == AVARKILL)
 			continue;
 		proginfo(&info, p);
 		if(info.flags & Call) {
@@ -790,7 +790,7 @@ copyu(Prog *p, Adr *v, Adr *s)
 		return 0;
 	}
 
-	if(p->as == AVARDEF)
+	if(p->as == AVARDEF || p->as == AVARKILL)
 		return 0;
 	proginfo(&info, p);
 

src/cmd/6g/prog.c

@@ -42,6 +42,7 @@ static ProgInfo progtable[ALAST] = {
 	[AUSEFIELD]=	{OK},
 	[ACHECKNIL]=	{LeftRead},
 	[AVARDEF]=	{Pseudo | RightWrite},
+	[AVARKILL]=	{Pseudo | RightWrite},
 
 	// NOP is an internal no-op that also stands
 	// for USED and SET annotations, not the Intel opcode.

src/cmd/6g/reg.c

@@ -199,7 +199,7 @@ regopt(Prog *firstp)
 
 	for(r = firstr; r != R; r = (Reg*)r->f.link) {
 		p = r->f.prog;
-		if(p->as == AVARDEF)
+		if(p->as == AVARDEF || p->as == AVARKILL)
 			continue;
 		proginfo(&info, p);
 

src/cmd/6l/6.out.h

@@ -763,6 +763,7 @@ enum	as
 	APCDATA,
 	ACHECKNIL,
 	AVARDEF,
+	AVARKILL,
 	
 	ALAST
 };

src/cmd/8g/ggen.c

@@ -82,7 +82,7 @@ void
 markautoused(Prog* p)
 {
 	for (; p; p = p->link) {
-		if (p->as == ATYPE || p->as == AVARDEF)
+		if (p->as == ATYPE || p->as == AVARDEF || p->as == AVARKILL)
 			continue;
 
 		if (p->from.node)
@@ -104,7 +104,7 @@ fixautoused(Prog* p)
 			*lp = p->link;
 			continue;
 		}
-		if (p->as == AVARDEF && p->to.node && !p->to.node->used) {
+		if ((p->as == AVARDEF || p->as == AVARKILL) && p->to.node && !p->to.node->used) {
 			// Cannot remove VARDEF instruction, because - unlike TYPE handled above -
 			// VARDEFs are interspersed with other code, and a jump might be using the
 			// VARDEF as a target. Replace with a no-op instead. A later pass will remove

src/cmd/8g/peep.c

@@ -387,7 +387,7 @@ subprop(Flow *r0)
 		if(uniqs(r) == nil)
 			break;
 		p = r->prog;
-		if(p->as == AVARDEF)
+		if(p->as == AVARDEF || p->as == AVARKILL)
 			continue;
 		proginfo(&info, p);
 		if(info.flags & Call)
@@ -586,7 +586,7 @@ copyu(Prog *p, Adr *v, Adr *s)
 		return 0;
 	}
 
-	if(p->as == AVARDEF)
+	if(p->as == AVARDEF || p->as == AVARKILL)
 		return 0;
 	proginfo(&info, p);
 

src/cmd/8g/prog.c

@@ -42,6 +42,7 @@ static ProgInfo progtable[ALAST] = {
 	[AUSEFIELD]=	{OK},
 	[ACHECKNIL]=	{LeftRead},
 	[AVARDEF]=	{Pseudo | RightWrite},
+	[AVARKILL]=	{Pseudo | RightWrite},
 
 	// NOP is an internal no-op that also stands
 	// for USED and SET annotations, not the Intel opcode.

src/cmd/8g/reg.c

@@ -169,7 +169,7 @@ regopt(Prog *firstp)
 
 	for(r = firstr; r != R; r = (Reg*)r->f.link) {
 		p = r->f.prog;
-		if(p->as == AVARDEF)
+		if(p->as == AVARDEF || p->as == AVARKILL)
 			continue;
 		proginfo(&info, p);
 

src/cmd/8l/8.out.h

@@ -580,6 +580,7 @@ enum	as
 	APCDATA,
 	ACHECKNIL,
 	AVARDEF,
+	AVARKILL,
 	
 	ALAST
 };

src/cmd/gc/gen.c

@@ -495,6 +495,11 @@ gen(Node *n)
 	
 	case OCHECKNIL:
 		cgen_checknil(n->left);
+		break;
+	
+	case OVARKILL:
+		gvarkill(n->left);
+		break;
 	}
 
 ret:

src/cmd/gc/go.h

@@ -581,6 +581,7 @@ enum
 	OCLOSUREVAR, // variable reference at beginning of closure function
 	OCFUNC,	// reference to c function pointer (not go func value)
 	OCHECKNIL, // emit code to ensure pointer/interface not nil
+	OVARKILL, // variable is dead
 
 	// arch-specific registers
 	OREGISTER,	// a register, such as AX.
@@ -1502,6 +1503,7 @@ void	gdatastring(Node*, Strlit*);
 void	ggloblnod(Node *nam);
 void	ggloblsym(Sym *s, int32 width, int dupok, int rodata);
 void	gvardef(Node*);
+void	gvarkill(Node*);
 Prog*	gjmp(Prog*);
 void	gused(Node*);
 void	movelarge(NodeList*);

src/cmd/gc/order.c

@@ -5,79 +5,349 @@
 // Rewrite tree to use separate statements to enforce
 // order of evaluation.  Makes walk easier, because it
 // can (after this runs) reorder at will within an expression.
+//
+// Rewrite x op= y into x = x op y.
+//
+// Introduce temporaries as needed by runtime routines.
+// For example, the map runtime routines take the map key
+// by reference, so make sure all map keys are addressable
+// by copying them to temporaries as needed.
+// The same is true for channel operations.
+//
+// Arrange that map index expressions only appear in direct
+// assignments x = m[k] or m[k] = x, never in larger expressions.
+//
+// Arrange that receive expressions only appear in direct assignments
+// x = <-c or as standalone statements <-c, never in larger expressions.
+
+// TODO(rsc): Temporaries are not cleaned in for, if, select, and swtch
+// statements. The cleaning needs to be introduced aggressively, so
+// that for example a temporary introduced during evaluation of an
+// if condition is killed in both the 'if' and 'else' bodies, not delayed
+// until after the entire if statement has completed.
+
+// TODO(rsc): Goto and multilevel break/continue can jump over
+// inserted VARKILL annotations. Work out a way to handle these.
+// The current implementation is safe, in that it will execute correctly.
+// But it won't reuse temporaries as aggressively as it might, and
+// it can result in unnecessary zeroing of those variables in the function
+// prologue.
 
 #include	<u.h>
 #include	<libc.h>
 #include	"go.h"
 
-static void	orderstmt(Node*, NodeList**);
-static void	orderstmtlist(NodeList*, NodeList**);
-static void	orderblock(NodeList **l);
-static void	orderexpr(Node**, NodeList**);
-static void	orderexprlist(NodeList*, NodeList**);
+// Order holds state during the ordering process.
+typedef struct Order Order;
+struct Order
+{
+	NodeList *out; // list of generated statements
+	NodeList *temp; // head of stack of temporary variables
+	NodeList *free; // free list of NodeList* structs (for use in temp)
+};
 
+static void	orderstmt(Node*, Order*);
+static void	orderstmtlist(NodeList*, Order*);
+static void	orderblock(NodeList **l);
+static void	orderexpr(Node**, Order*);
+static void orderexprinplace(Node**, Order*);
+static void	orderexprlist(NodeList*, Order*);
+static void	orderexprlistinplace(NodeList*, Order*);
+
+// Order rewrites fn->nbody to apply the ordering constraints
+// described in the comment at the top of the file.
 void
 order(Node *fn)
 {
 	orderblock(&fn->nbody);
 }
 
-static void
-orderstmtlist(NodeList *l, NodeList **out)
+// Ordertemp allocates a new temporary with the given type,
+// pushes it onto the temp stack, and returns it.
+// If clear is true, ordertemp emits code to zero the temporary.
+static Node*
+ordertemp(Type *t, Order *order, int clear)
 {
-	for(; l; l=l->next)
-		orderstmt(l->n, out);
+	Node *var, *a;
+	NodeList *l;
+
+	var = temp(t);
+	if(clear) {
+		a = nod(OAS, var, N);
+		typecheck(&a, Etop);
+		order->out = list(order->out, a);
+	}
+	if((l = order->free) == nil)
+		l = mal(sizeof *l);
+	order->free = l->next;
+	l->next = order->temp;
+	l->n = var;
+	order->temp = l;
+	return var;
 }
 
-// Order the block of statements *l onto a new list,
-// and then replace *l with that list.
+// Ordercopyexpr behaves like ordertemp but also emits
+// code to initialize the temporary to the value n.
+//
+// The clear argument is provided for use when the evaluation
+// of tmp = n turns into a function call that is passed a pointer
+// to the temporary as the output space. If the call blocks before
+// tmp has been written, the garbage collector will still treat the
+// temporary as live, so we must zero it before entering that call.
+// Today, this only happens for channel receive operations.
+// (The other candidate would be map access, but map access
+// returns a pointer to the result data instead of taking a pointer
+// to be filled in.)
+static Node*
+ordercopyexpr(Node *n, Type *t, Order *order, int clear)
+{
+	Node *a, *var;
+
+	var = ordertemp(t, order, clear);
+	a = nod(OAS, var, n);
+	typecheck(&a, Etop);
+	order->out = list(order->out, a);
+	return var;
+}
+
+// Ordercheapexpr returns a cheap version of n.
+// The definition of cheap is that n is a variable or constant.
+// If not, ordercheapexpr allocates a new tmp, emits tmp = n,
+// and then returns tmp.
+static Node*
+ordercheapexpr(Node *n, Order *order)
+{
+	switch(n->op) {
+	case ONAME:
+	case OLITERAL:
+		return n;
+	}
+	return ordercopyexpr(n, n->type, order, 0);
+}
+
+// Ordersafeexpr returns a safe version of n.
+// The definition of safe is that n can appear multiple times
+// without violating the semantics of the original program,
+// and that assigning to the safe version has the same effect
+// as assigning to the original n.
+//
+// The intended use is to apply to x when rewriting x += y into x = x + y.
+static Node*
+ordersafeexpr(Node *n, Order *order)
+{
+	Node *l, *r, *a;
+	
+	switch(n->op) {
+	default:
+		fatal("ordersafeexpr %O", n->op);
+
+	case ONAME:
+	case OLITERAL:
+		return n;
+
+	case ODOT:
+		l = ordersafeexpr(n->left, order);
+		if(l == n->left)
+			return n;
+		a = nod(OXXX, N, N);
+		*a = *n;
+		a->orig = a;
+		a->left = l;
+		typecheck(&a, Erv);
+		return a;
+
+	case ODOTPTR:
+	case OIND:
+		l = ordercheapexpr(n->left, order);
+		if(l == n->left)
+			return n;
+		a = nod(OXXX, N, N);
+		*a = *n;
+		a->orig = a;
+		a->left = l;
+		typecheck(&a, Erv);
+		return a;
+		
+	case OINDEX:
+	case OINDEXMAP:
+		if(isfixedarray(n->left->type))
+			l = ordersafeexpr(n->left, order);
+		else
+			l = ordercheapexpr(n->left, order);
+		r = ordercheapexpr(n->right, order);
+		if(l == n->left && r == n->right)
+			return n;
+		a = nod(OXXX, N, N);
+		*a = *n;
+		a->orig = a;
+		a->left = l;
+		a->right = r;
+		typecheck(&a, Erv);
+		return a;
+	}
+}		
+
+// Istemp reports whether n is a temporary variable.
+static int
+istemp(Node *n)
+{
+	if(n->op != ONAME)
+		return 0;
+	return strncmp(n->sym->name, "autotmp_", 8) == 0;
+}
+
+// Isaddrokay reports whether it is okay to pass n's address to runtime routines.
+// Taking the address of a variable makes the liveness and optimization analyses
+// lose track of where the variable's lifetime ends. To avoid hurting the analyses
+// of ordinary stack variables, those are not 'isaddrokay'. Temporaries are okay,
+// because we emit explicit VARKILL instructions marking the end of those
+// temporaries' lifetimes.
+static int
+isaddrokay(Node *n)
+{
+	return islvalue(n) && (n->op != ONAME || n->class == PEXTERN || istemp(n));
+}
+
+// Orderaddrtemp ensures that *np is okay to pass by address to runtime routines.
+// If the original argument *np is not okay, orderaddrtemp creates a tmp, emits
+// tmp = *np, and then sets *np to the tmp variable.
+static void
+orderaddrtemp(Node **np, Order *order)
+{
+	Node *n;
+	
+	n = *np;
+	if(isaddrokay(n))
+		return;
+	*np = ordercopyexpr(n, n->type, order, 0);
+}
+
+// Marktemp returns the top of the temporary variable stack.
+static NodeList*
+marktemp(Order *order)
+{
+	return order->temp;
+}
+
+// Poptemp pops temporaries off the stack until reaching the mark,
+// which must have been returned by marktemp.
+static void
+poptemp(NodeList *mark, Order *order)
+{
+	NodeList *l;
+
+	while((l = order->temp) != mark) {
+		order->temp = l->next;
+		l->next = order->free;
+		order->free = l;
+	}
+}
+
+// Cleantempnopop emits VARKILL instructions for each temporary
+// above the mark on the temporary stack, but it does not pop them
+// from the stack.
+static void
+cleantempnopop(NodeList *mark, Order *order)
+{
+	NodeList *l;
+	Node *kill;
+
+	for(l=order->temp; l != mark; l=l->next) {
+		kill = nod(OVARKILL, l->n, N);
+		typecheck(&kill, Etop);
+		order->out = list(order->out, kill);
+	}
+}
+
+// Cleantemp emits VARKILL instructions for each temporary above the
+// mark on the temporary stack and removes them from the stack.
+static void
+cleantemp(NodeList *top, Order *order)
+{
+	cleantempnopop(top, order);
+	poptemp(top, order);
+}
+
+// Orderstmtlist orders each of the statements in the list.
+static void
+orderstmtlist(NodeList *l, Order *order)
+{
+	for(; l; l=l->next)
+		orderstmt(l->n, order);
+}
+
+// Orderblock orders the block of statements *l onto a new list,
+// and then replaces *l with that list.
 static void
 orderblock(NodeList **l)
 {
-	NodeList *out;
+	Order order;
 	
-	out = nil;
-	orderstmtlist(*l, &out);
-	*l = out;
+	memset(&order, 0, sizeof order);
+	orderstmtlist(*l, &order);
+	*l = order.out;
 }
 
-// Order the side effects in *np and leave them as
-// the init list of the final *np.
+// Orderexprinplace orders the side effects in *np and
+// leaves them as the init list of the final *np.
 static void
-orderexprinplace(Node **np)
+orderexprinplace(Node **np, Order *TODO)
 {
 	Node *n;
-	NodeList *out;
+	Order order;
+	
+	// TODO(rsc): Decide how much of the passed-in order to use.
+	// For example, should the temporaries created during the
+	// ordering of expr be added onto the caller's order temp list
+	// for freeing? Probably.
+	USED(TODO);
 
 	n = *np;
-	out = nil;
-	orderexpr(&n, &out);
-	addinit(&n, out);
+	memset(&order, 0, sizeof order);
+	orderexpr(&n, &order);
+	addinit(&n, order.out);
 	*np = n;
 }
 
-// Like orderblock, but applied to a single statement.
+// Orderexprtolist orders the side effects in *np and
+// appends them to *out.
+static void
+orderexprtolist(Node **np, NodeList **out)
+{
+	Node *n;
+	Order order;
+	
+	n = *np;
+	memset(&order, 0, sizeof order);
+	orderexpr(&n, &order);
+	*out = concat(*out, order.out);
+	*np = n;
+}
+
+// Orderstmtinplace orders the side effects of the single statement *np
+// and replaces it with the resulting statement list.
 static void
 orderstmtinplace(Node **np)
 {
 	Node *n;
-	NodeList *out;
+	Order order;
 
 	n = *np;
-	out = nil;
-	orderstmt(n, &out);
-	*np = liststmt(out);
+	memset(&order, 0, sizeof order);
+	orderstmt(n, &order);
+	*np = liststmt(order.out);
 }
 
-// Move n's init list to *out.
+// Orderinit moves n's init list to order->out.
 static void
-orderinit(Node *n, NodeList **out)
+orderinit(Node *n, Order *order)
 {
-	orderstmtlist(n->ninit, out);
+	orderstmtlist(n->ninit, order);
 	n->ninit = nil;
 }
 
-// Is the list l actually just f() for a multi-value function?
+// Ismulticall reports whether the list l is f() for a multi-value function.
+// Such an f() could appear as the lone argument to a multi-arg function.
 static int
 ismulticall(NodeList *l)
 {
@@ -102,10 +372,10 @@ ismulticall(NodeList *l)
 	return n->left->type->outtuple > 1;
 }
 
-// n is a multi-value function call.  Add t1, t2, .. = n to out
-// and return the list t1, t2, ...
+// Copyret emits t1, t2, ... = n, where n is a function call,
+// and then returns the list t1, t2, ....
 static NodeList*
-copyret(Node *n, NodeList **out)
+copyret(Node *n, Order *order)
 {
 	Type *t;
 	Node *tmp, *as;
@@ -127,86 +397,216 @@ copyret(Node *n, NodeList **out)
 	as->list = l1;
 	as->rlist = list1(n);
 	typecheck(&as, Etop);
-	orderstmt(as, out);
+	orderstmt(as, order);
 
 	return l2;
 }
 
+// Ordercallargs orders the list of call arguments *l.
 static void
-ordercallargs(NodeList **l, NodeList **out)
+ordercallargs(NodeList **l, Order *order)
 {
 	if(ismulticall(*l)) {
 		// return f() where f() is multiple values.
-		*l = copyret((*l)->n, out);
+		*l = copyret((*l)->n, order);
 	} else {
-		orderexprlist(*l, out);
+		orderexprlist(*l, order);
 	}
 }
 
+// Ordercall orders the call expression n.
+// n->op is  OCALLMETH/OCALLFUNC/OCALLINTER.
 static void
-ordercall(Node *n, NodeList **out)
+ordercall(Node *n, Order *order)
 {
-	orderexpr(&n->left, out);
-	ordercallargs(&n->list, out);
+	orderexpr(&n->left, order);
+	ordercallargs(&n->list, order);
 }
 
+// Ordermapassign appends n to order->out, introducing temporaries
+// to make sure that all map assignments have the form m[k] = x,
+// where x is adressable.
+// (Orderexpr has already been called on n, so we know k is addressable.)
+//
+// If n is m[k] = x where x is not addressable, the rewrite is:
+//	tmp = x
+//	m[k] = tmp
+//
+// If n is the multiple assignment form ..., m[k], ... = ..., the rewrite is
+//	t1 = m
+//	t2 = k
+//	...., t3, ... = x
+//	t1[t2] = t3
+//
+// The temporaries t1, t2 are needed in case the ... being assigned
+// contain m or k. They are usually unnecessary, but in the unnecessary
+// cases they are also typically registerizable, so not much harm done.
+// And this only applies to the multiple-assignment form.
+// We could do a more precise analysis if needed, like in walk.c.
 static void
-orderstmt(Node *n, NodeList **out)
+ordermapassign(Node *n, Order *order)
+{
+	Node *m, *a;
+	NodeList *l;
+	NodeList *post;
+
+	switch(n->op) {
+	default:
+		fatal("ordermapassign %O", n->op);
+
+	case OAS:
+		order->out = list(order->out, n);
+		if(n->left->op == OINDEXMAP && !isaddrokay(n->right)) {
+			m = n->left;
+			n->left = ordertemp(m->type, order, 0);
+			a = nod(OAS, m, n->left);
+			typecheck(&a, Etop);
+			order->out = list(order->out, a);
+		}
+		break;
+
+	case OAS2:
+	case OAS2DOTTYPE:
+	case OAS2MAPR:
+	case OAS2FUNC:
+		post = nil;
+		for(l=n->list; l != nil; l=l->next) {
+			if(l->n->op == OINDEXMAP) {
+				m = l->n;
+				if(!istemp(m->left))
+					m->left = ordercopyexpr(m->left, m->left->type, order, 0);
+				if(!istemp(m->right))
+					m->right = ordercopyexpr(m->left, m->left->type, order, 0);
+				l->n = ordertemp(m->type, order, 0);
+				a = nod(OAS, m, l->n);
+				typecheck(&a, Etop);
+				post = list(post, a);
+			}
+		}
+		order->out = list(order->out, n);
+		order->out = concat(order->out, post);
+		break;
+	}
+}
+
+// Orderstmt orders the statement n, appending to order->out.
+// Temporaries created during the statement are cleaned
+// up using VARKILL instructions as possible.
+static void
+orderstmt(Node *n, Order *order)
 {
 	int lno;
-	NodeList *l;
-	Node *r;
+	NodeList *l, *t, *t1;
+	Node *r, *tmp1, *tmp2, **np;
+	Type *ch;
 
 	if(n == N)
 		return;
 
 	lno = setlineno(n);
 
-	orderinit(n, out);
+	orderinit(n, order);
 
 	switch(n->op) {
 	default:
 		fatal("orderstmt %O", n->op);
 
+	case OAS:
 	case OAS2:
 	case OAS2DOTTYPE:
-	case OAS2MAPR:
-	case OAS:
-	case OASOP:
 	case OCLOSE:
 	case OCOPY:
-	case ODELETE:
 	case OPANIC:
 	case OPRINT:
 	case OPRINTN:
 	case ORECOVER:
 	case ORECV:
-	case OSEND:
-		orderexpr(&n->left, out);
-		orderexpr(&n->right, out);
-		orderexprlist(n->list, out);
-		orderexprlist(n->rlist, out);
-		*out = list(*out, n);
+		t = marktemp(order);
+		orderexpr(&n->left, order);
+		orderexpr(&n->right, order);
+		orderexprlist(n->list, order);
+		orderexprlist(n->rlist, order);
+		switch(n->op) {
+		case OAS:
+		case OAS2:
+		case OAS2DOTTYPE:
+			ordermapassign(n, order);
+			break;
+		default:
+			order->out = list(order->out, n);
+			break;
+		}
+		cleantemp(t, order);
+		break;
+
+	case OASOP:
+		// Special: rewrite l op= r into l = l op r.
+		// This simplies quite a few operations;
+		// most important is that it lets us separate
+		// out map read from map write when l is
+		// a map index expression.
+		t = marktemp(order);
+		orderexpr(&n->left, order);
+		orderexpr(&n->right, order);
+		n->left = ordersafeexpr(n->left, order);
+		tmp1 = treecopy(n->left);
+		if(tmp1->op == OINDEXMAP)
+			tmp1->etype = 0; // now an rvalue not an lvalue
+		tmp1 = ordercopyexpr(tmp1, n->left->type, order, 0);
+		n->right = nod(n->etype, tmp1, n->right);
+		typecheck(&n->right, Erv);
+		n->etype = 0;
+		n->op = OAS;
+		ordermapassign(n, order);
+		cleantemp(t, order);
+		break;
+
+	case OAS2MAPR:
+		// Special: make sure key is addressable,
+		// and make sure OINDEXMAP is not copied out.
+		t = marktemp(order);
+		orderexprlist(n->list, order);
+		orderexpr(&n->rlist->n->left, order);
+		orderexpr(&n->rlist->n->right, order);
+		orderaddrtemp(&n->rlist->n->right, order);
+		ordermapassign(n, order);
+		cleantemp(t, order);
 		break;
 
 	case OAS2FUNC:
 		// Special: avoid copy of func call n->rlist->n.
-		orderexprlist(n->list, out);
-		ordercall(n->rlist->n, out);
-		*out = list(*out, n);
+		t = marktemp(order);
+		orderexprlist(n->list, order);
+		ordercall(n->rlist->n, order);
+		ordermapassign(n, order);
+		cleantemp(t, order);
 		break;
 
 	case OAS2RECV:
 		// Special: avoid copy of receive.
-		orderexprlist(n->list, out);
-		orderexpr(&n->rlist->n->left, out);  // arg to recv
-		*out = list(*out, n);
+		// Use temporary variables to hold result,
+		// so that chanrecv can take address of temporary.
+		t = marktemp(order);
+		orderexprlist(n->list, order);
+		orderexpr(&n->rlist->n->left, order);  // arg to recv
+		ch = n->rlist->n->left->type;
+		tmp1 = ordertemp(ch->type, order, haspointers(ch->type));
+		tmp2 = ordertemp(types[TBOOL], order, 0);
+		order->out = list(order->out, n);
+		r = nod(OAS, n->list->n, tmp1);
+		typecheck(&r, Etop);
+		ordermapassign(r, order);
+		r = nod(OAS, n->list->next->n, tmp2);
+		typecheck(&r, Etop);
+		ordermapassign(r, order);
+		n->list = list(list1(tmp1), tmp2);
+		cleantemp(t, order);
 		break;
 
 	case OBLOCK:
 	case OEMPTY:
 		// Special: does not save n onto out.
-		orderstmtlist(n->list, out);
+		orderstmtlist(n->list, order);
 		break;
 
 	case OBREAK:
@@ -215,107 +615,158 @@ orderstmt(Node *n, NodeList **out)
 	case ODCLCONST:
 	case ODCLTYPE:
 	case OFALL:
-	case_OFALL:
+	case OXFALL:
 	case OGOTO:
 	case OLABEL:
 	case ORETJMP:
 		// Special: n->left is not an expression; save as is.
-		*out = list(*out, n);
+		order->out = list(order->out, n);
 		break;
 
 	case OCALLFUNC:
 	case OCALLINTER:
 	case OCALLMETH:
 		// Special: handle call arguments.
-		ordercall(n, out);
-		*out = list(*out, n);
+		t = marktemp(order);
+		ordercall(n, order);
+		order->out = list(order->out, n);
+		cleantemp(t, order);
 		break;
 
 	case ODEFER:
 	case OPROC:
 		// Special: order arguments to inner call but not call itself.
-		ordercall(n->left, out);
-		*out = list(*out, n);
+		t = marktemp(order);
+		switch(n->left->op) {
+		case ODELETE:
+			// Delete will take the address of the key.
+			// Copy key into new temp and do not clean it
+			// (it persists beyond the statement).
+			orderexprlist(n->left->list, order);
+			t1 = marktemp(order);
+			np = &n->left->list->next->n; // map key
+			*np = ordercopyexpr(*np, (*np)->type, order, 0);
+			poptemp(t1, order);
+			break;
+		default:
+			ordercall(n->left, order);
+			break;
+		}
+		order->out = list(order->out, n);
+		cleantemp(t, order);
+		break;
+
+	case ODELETE:
+		t = marktemp(order);
+		orderexpr(&n->list->n, order);
+		orderexpr(&n->list->next->n, order);
+		orderaddrtemp(&n->list->next->n, order); // map key
+		order->out = list(order->out, n);
+		cleantemp(t, order);
 		break;
 
 	case OFOR:
-		orderexprinplace(&n->ntest);
+		// TODO(rsc): Clean temporaries.
+		orderexprinplace(&n->ntest, order);
 		orderstmtinplace(&n->nincr);
 		orderblock(&n->nbody);
-		*out = list(*out, n);
+		order->out = list(order->out, n);
 		break;
 		
 	case OIF:
-		orderexprinplace(&n->ntest);
+		// TODO(rsc): Clean temporaries.
+		orderexprinplace(&n->ntest, order);
 		orderblock(&n->nbody);
 		orderblock(&n->nelse);
-		*out = list(*out, n);
+		order->out = list(order->out, n);
 		break;
 
 	case ORANGE:
-		orderexpr(&n->right, out);
+		// TODO(rsc): Clean temporaries.
+		orderexpr(&n->right, order);
 		for(l=n->list; l; l=l->next)
-			orderexprinplace(&l->n);
+			orderexprinplace(&l->n, order);
 		orderblock(&n->nbody);
-		*out = list(*out, n);
+		order->out = list(order->out, n);
 		break;
 
 	case ORETURN:
-		ordercallargs(&n->list, out);
-		*out = list(*out, n);
+		ordercallargs(&n->list, order);
+		order->out = list(order->out, n);
 		break;
 	
 	case OSELECT:
+		// TODO(rsc): Clean temporaries.
 		for(l=n->list; l; l=l->next) {
 			if(l->n->op != OXCASE)
 				fatal("order select case %O", l->n->op);
 			r = l->n->left;
-			if(r == nil)
-				continue;
-			switch(r->op) {
-			case OSELRECV:
-			case OSELRECV2:
-				orderexprinplace(&r->left);
-				orderexprinplace(&r->ntest);
-				orderexpr(&r->right->left, &l->n->ninit);
-				break;
-			case OSEND:
-				orderexpr(&r->left, &l->n->ninit);
-				orderexpr(&r->right, &l->n->ninit);
-				break;
+			if(r != nil) {
+				switch(r->op) {
+				case OSELRECV:
+				case OSELRECV2:
+					orderexprinplace(&r->left, order);
+					orderexprinplace(&r->ntest, order);
+					orderexprtolist(&r->right->left, &l->n->ninit);
+					break;
+				case OSEND:
+					orderexprtolist(&r->left, &l->n->ninit);
+					orderexprtolist(&r->right, &l->n->ninit);
+					break;
+				}
 			}
+			orderblock(&l->n->nbody);
 		}
-		*out = list(*out, n);
+		order->out = list(order->out, n);
+		break;
+
+	case OSEND:
+		// Special: value being sent is passed as a pointer; make it addressable.
+		t = marktemp(order);
+		orderexpr(&n->left, order);
+		orderexpr(&n->right, order);
+		orderaddrtemp(&n->right, order);
+		order->out = list(order->out, n);
+		cleantemp(t, order);
 		break;
 
 	case OSWITCH:
-		orderexpr(&n->ntest, out);
+		// TODO(rsc): Clean temporaries.
+		orderexpr(&n->ntest, order);
 		for(l=n->list; l; l=l->next) {
 			if(l->n->op != OXCASE)
 				fatal("order switch case %O", l->n->op);
-			orderexpr(&l->n->left, &l->n->ninit);
+			orderexprlistinplace(l->n->list, order);
+			orderblock(&l->n->nbody);
 		}
-		*out = list(*out, n);
+		order->out = list(order->out, n);
 		break;
-
-	case OXFALL:
-		yyerror("fallthrough statement out of place");
-		n->op = OFALL;
-		goto case_OFALL;
 	}
 	
 	lineno = lno;
 }
 
+// Orderexprlist orders the expression list l into order.
 static void
-orderexprlist(NodeList *l, NodeList **out)
+orderexprlist(NodeList *l, Order *order)
 {
 	for(; l; l=l->next)
-		orderexpr(&l->n, out);
+		orderexpr(&l->n, order);
 }
 
+// Orderexprlist orders the expression list l but saves
+// the side effects on the individual expression ninit lists.
 static void
-orderexpr(Node **np, NodeList **out)
+orderexprlistinplace(NodeList *l, Order *order)
+{
+	for(; l; l=l->next)
+		orderexprinplace(&l->n, order);
+}
+
+// Orderexpr orders a single expression, appending side
+// effects to order->out as needed.
+static void
+orderexpr(Node **np, Order *order)
 {
 	Node *n;
 	int lno;
@@ -325,20 +776,32 @@ orderexpr(Node **np, NodeList **out)
 		return;
 
 	lno = setlineno(n);
-	orderinit(n, out);
+	orderinit(n, order);
 
 	switch(n->op) {
 	default:
-		orderexpr(&n->left, out);
-		orderexpr(&n->right, out);
-		orderexprlist(n->list, out);
-		orderexprlist(n->rlist, out);
+		orderexpr(&n->left, order);
+		orderexpr(&n->right, order);
+		orderexprlist(n->list, order);
+		orderexprlist(n->rlist, order);
+		break;
+	
+	case OINDEXMAP:
+		// key must be addressable
+		orderexpr(&n->left, order);
+		orderexpr(&n->right, order);
+		orderaddrtemp(&n->right, order);
+		if(n->etype == 0) {
+			// use of value (not being assigned);
+			// make copy in temporary.
+			n = ordercopyexpr(n, n->type, order, 0);
+		}
 		break;
 	
 	case OANDAND:
 	case OOROR:
-		orderexpr(&n->left, out);
-		orderexprinplace(&n->right);
+		orderexpr(&n->left, order);
+		orderexprinplace(&n->right, order);
 		break;
 	
 	case OCALLFUNC:
@@ -346,12 +809,12 @@ orderexpr(Node **np, NodeList **out)
 	case OCALLINTER:
 	case OAPPEND:
 	case OCOMPLEX:
-		ordercall(n, out);
-		n = copyexpr(n, n->type, out);
+		ordercall(n, order);
+		n = ordercopyexpr(n, n->type, order, 0);
 		break;
 
 	case ORECV:
-		n = copyexpr(n, n->type, out);
+		n = ordercopyexpr(n, n->type, order, 1);
 		break;
 	}
 	

src/cmd/gc/pgen.c

@@ -79,10 +79,16 @@ makefuncdatasym(char *namefmt, int64 funcdatakind)
 // wants to work on individual variables, which might be multi-word
 // aggregates. It might make sense at some point to look into letting
 // the liveness analysis work on single-word values as well, although
-// there are complications around interface values, which cannot be
-// treated as individual words.
-void
-gvardef(Node *n)
+// there are complications around interface values, slices, and strings,
+// all of which cannot be treated as individual words.
+//
+// VARKILL is the opposite of VARDEF: it marks a value as no longer needed,
+// even if its address has been taken. That is, a VARKILL annotation asserts
+// that its argument is certainly dead, for use when the liveness analysis
+// would not otherwise be able to deduce that fact.
+
+static void
+gvardefx(Node *n, int as)
 {
 	if(n == N)
 		fatal("gvardef nil");
@@ -94,20 +100,32 @@ gvardef(Node *n)
 	case PAUTO:
 	case PPARAM:
 	case PPARAMOUT:
-		gins(AVARDEF, N, n);
+		gins(as, N, n);
 	}
 }
 
+void
+gvardef(Node *n)
+{
+	gvardefx(n, AVARDEF);
+}
+
+void
+gvarkill(Node *n)
+{
+	gvardefx(n, AVARKILL);
+}
+
 static void
 removevardef(Prog *firstp)
 {
 	Prog *p;
 
 	for(p = firstp; p != P; p = p->link) {
-		while(p->link != P && p->link->as == AVARDEF)
+		while(p->link != P && (p->link->as == AVARDEF || p->link->as == AVARKILL))
 			p->link = p->link->link;
 		if(p->to.type == D_BRANCH)
-			while(p->to.u.branch != P && p->to.u.branch->as == AVARDEF)
+			while(p->to.u.branch != P && (p->to.u.branch->as == AVARDEF || p->to.u.branch->as == AVARKILL))
 				p->to.u.branch = p->to.u.branch->link;
 	}
 }

src/cmd/gc/plive.c

@@ -743,8 +743,9 @@ Next:
 				if(pos == -1)
 					goto Next1;
 				if(to->node->addrtaken) {
-					bvset(avarinit, pos);
-					if(prog->as == AVARDEF)
+					if(prog->as != AVARKILL)
+						bvset(avarinit, pos);
+					if(prog->as == AVARDEF || prog->as == AVARKILL)
 						bvset(varkill, pos);
 				} else {
 					if(info.flags & (RightRead | RightAddr))
@@ -1310,7 +1311,7 @@ livenessprologue(Liveness *lv)
 {
 	BasicBlock *bb;
 	Bvec *uevar, *varkill, *avarinit;
-	Prog *prog;
+	Prog *p;
 	int32 i;
 	int32 nvars;
 
@@ -1322,14 +1323,25 @@ livenessprologue(Liveness *lv)
 		bb = *(BasicBlock**)arrayget(lv->cfg, i);
 		// Walk the block instructions backward and update the block
 		// effects with the each prog effects.
-		for(prog = bb->last; prog != nil; prog = prog->opt) {
-			progeffects(prog, lv->vars, uevar, varkill, avarinit);
+		for(p = bb->last; p != nil; p = p->opt) {
+			progeffects(p, lv->vars, uevar, varkill, avarinit);
 			if(debuglive >= 3)
-				printeffects(prog, uevar, varkill, avarinit);
+				printeffects(p, uevar, varkill, avarinit);
 			bvor(lv->varkill[i], lv->varkill[i], varkill);
 			bvandnot(lv->uevar[i], lv->uevar[i], varkill);
 			bvor(lv->uevar[i], lv->uevar[i], uevar);			
+		}
+		// Walk the block instructions forward to update avarinit bits.
+		// avarinit describes the effect at the end of the block, not the beginning.
+		bvresetall(varkill);
+		for(p = bb->first;; p = p->link) {
+			progeffects(p, lv->vars, uevar, varkill, avarinit);
+			if(debuglive >= 3)
+				printeffects(p, uevar, varkill, avarinit);
+			bvandnot(lv->avarinit[i], lv->avarinit[i], varkill);
 			bvor(lv->avarinit[i], lv->avarinit[i], avarinit);
+			if(p == bb->last)
+				break;
 		}
 	}
 	free(uevar);
@@ -1385,6 +1397,8 @@ livenesssolve(Liveness *lv)
 					bvand(all, all, lv->avarinitall[pred->rpo]);
 				}
 			}
+			bvandnot(any, any, lv->varkill[rpo]);
+			bvandnot(all, all, lv->varkill[rpo]);
 			bvor(any, any, lv->avarinit[rpo]);
 			bvor(all, all, lv->avarinit[rpo]);
 			if(bvcmp(any, lv->avarinitany[rpo])) {
@@ -1511,6 +1525,8 @@ livenessepilogue(Liveness *lv)
 		// Seed the maps with information about the addrtaken variables.
 		for(p = bb->first;; p = p->link) {
 			progeffects(p, lv->vars, uevar, varkill, avarinit);
+			bvandnot(any, any, varkill);
+			bvandnot(all, all, varkill);
 			bvor(any, any, avarinit);
 			bvor(all, all, avarinit);
 
@@ -1572,7 +1588,7 @@ livenessepilogue(Liveness *lv)
 				break;
 		}
 		
-		if(debuglive >= 1 && strcmp(curfn->nname->sym->name, "init") != 0) {
+		if(debuglive >= 1 && strcmp(curfn->nname->sym->name, "init") != 0 && curfn->nname->sym->name[0] != '.') {
 			nmsg = arraylength(lv->livepointers);
 			startmsg = nmsg;
 			msg = xmalloc(nmsg*sizeof msg[0]);

src/cmd/gc/popt.c

@@ -531,10 +531,11 @@ startcmp(const void *va, const void *vb)
 static int
 canmerge(Node *n)
 {
-	return n->class == PAUTO && !n->addrtaken && strncmp(n->sym->name, "autotmp", 7) == 0;
+	return n->class == PAUTO && strncmp(n->sym->name, "autotmp", 7) == 0;
 }
 
 static void mergewalk(TempVar*, TempFlow*, uint32);
+static void varkillwalk(TempVar*, TempFlow*, uint32);
 
 void
 mergetemp(Prog *firstp)
@@ -651,6 +652,11 @@ mergetemp(Prog *firstp)
 		gen++;
 		for(r = v->use; r != nil; r = r->uselink)
 			mergewalk(v, r, gen);
+		if(v->addr) {
+			gen++;
+			for(r = v->use; r != nil; r = r->uselink)
+				varkillwalk(v, r, gen);
+		}
 	}
 
 	// Sort variables by start.
@@ -670,7 +676,7 @@ mergetemp(Prog *firstp)
 	nfree = nvar;
 	for(i=0; i<nvar; i++) {
 		v = bystart[i];
-		if(v->addr || v->removed)
+		if(v->removed)
 			continue;
 
 		// Expire no longer in use.
@@ -683,7 +689,12 @@ mergetemp(Prog *firstp)
 		t = v->node->type;
 		for(j=nfree; j<nvar; j++) {
 			v1 = inuse[j];
-			if(eqtype(t, v1->node->type)) {
+			// Require the types to match but also require the addrtaken bits to match.
+			// If a variable's address is taken, that disables registerization for the individual
+			// words of the variable (for example, the base,len,cap of a slice).
+			// We don't want to merge a non-addressed var with an addressed one and
+			// inhibit registerization of the former.
+			if(eqtype(t, v1->node->type) && v->node->addrtaken == v1->node->addrtaken) {
 				inuse[j] = inuse[nfree++];
 				if(v1->merge)
 					v->merge = v1->merge;
@@ -776,6 +787,29 @@ mergewalk(TempVar *v, TempFlow *r0, uint32 gen)
 			mergewalk(v, r2, gen);
 }
 
+static void
+varkillwalk(TempVar *v, TempFlow *r0, uint32 gen)
+{
+	Prog *p;
+	TempFlow *r1, *r;
+	
+	for(r1 = r0; r1 != nil; r1 = (TempFlow*)r1->f.s1) {
+		if(r1->f.active == gen)
+			break;
+		r1->f.active = gen;
+		p = r1->f.prog;
+		if(v->end < p->pc)
+			v->end = p->pc;
+		if(v->start > p->pc)
+			v->start = p->pc;
+		if(p->as == ARET || (p->as == AVARKILL && p->to.node == v->node))
+			break;
+	}
+	
+	for(r = r0; r != r1; r = (TempFlow*)r->f.s1)
+		varkillwalk(v, (TempFlow*)r->f.s2, gen);
+}
+
 // Eliminate redundant nil pointer checks.
 //
 // The code generation pass emits a CHECKNIL for every possibly nil pointer.

src/cmd/gc/racewalk.c

@@ -166,6 +166,7 @@ racewalknode(Node **np, NodeList **init, int wr, int skip)
 		goto ret;
 
 	case OCFUNC:
+	case OVARKILL:
 		// can't matter
 		goto ret;
 

src/cmd/gc/range.c

@@ -164,8 +164,7 @@ walkrange(Node *n)
 		}
 
 		n->ntest = nod(OLT, hv1, hn);
-		n->nincr = nod(OASOP, hv1, nodintconst(1));
-		n->nincr->etype = OADD;
+		n->nincr = nod(OAS, hv1, nod(OADD, hv1, nodintconst(1)));
 		if(v2 == N)
 			body = list1(nod(OAS, v1, hv1));
 		else {

src/cmd/gc/select.c

@@ -110,7 +110,9 @@ walkselect(Node *sel)
 	}
 
 	// optimization: one-case select: single op.
-	if(i == 1) {
+	// TODO(rsc): Reenable optimization once order.c can handle it.
+	// golang.org/issue/7672.
+	if(0 && i == 1) {
 		cas = sel->list->n;
 		setlineno(cas);
 		l = cas->ninit;

src/cmd/gc/sinit.c

@@ -838,7 +838,7 @@ maplit(int ctxt, Node *n, Node *var, NodeList **init)
 	int nerr;
 	int64 b;
 	Type *t, *tk, *tv, *t1;
-	Node *vstat, *index, *value;
+	Node *vstat, *index, *value, *key, *val;
 	Sym *syma, *symb;
 
 USED(ctxt);
@@ -952,8 +952,7 @@ ctxt = 0;
 
 		a->ninit = list1(nod(OAS, index, nodintconst(0)));
 		a->ntest = nod(OLT, index, nodintconst(t->bound));
-		a->nincr = nod(OASOP, index, nodintconst(1));
-		a->nincr->etype = OADD;
+		a->nincr = nod(OAS, index, nod(OADD, index, nodintconst(1)));
 
 		typecheck(&a, Etop);
 		walkstmt(&a);
@@ -961,6 +960,8 @@ ctxt = 0;
 	}
 
 	// put in dynamic entries one-at-a-time
+	key = nil;
+	val = nil;
 	for(l=n->list; l; l=l->next) {
 		r = l->n;
 
@@ -972,14 +973,36 @@ ctxt = 0;
 		if(isliteral(index) && isliteral(value))
 			continue;
 			
-		// build list of var[c] = expr
-		a = nod(OINDEX, var, r->left);
-		a = nod(OAS, a, r->right);
+		// build list of var[c] = expr.
+		// use temporary so that mapassign1 can have addressable key, val.
+		if(key == nil) {
+			key = temp(var->type->down);
+			val = temp(var->type->type);
+		}
+		a = nod(OAS, key, r->left);
 		typecheck(&a, Etop);
-		walkexpr(&a, init);
+		walkstmt(&a);
+		*init = list(*init, a);
+		a = nod(OAS, val, r->right);
+		typecheck(&a, Etop);
+		walkstmt(&a);
+		*init = list(*init, a);
+
+		a = nod(OAS, nod(OINDEX, var, key), val);
+		typecheck(&a, Etop);
+		walkstmt(&a);
+		*init = list(*init, a);
+
 		if(nerr != nerrors)
 			break;
+	}
 	
+	if(key != nil) {
+		a = nod(OVARKILL, key, N);
+		typecheck(&a, Etop);
+		*init = list(*init, a);
+		a = nod(OVARKILL, var, N);
+		typecheck(&a, Etop);
 		*init = list(*init, a);
 	}
 }

src/cmd/gc/typecheck.c

@@ -950,7 +950,7 @@ reswitch:
 		r = n->right;
 		if(r->type == T)
 			goto error;
-		r = assignconv(r, l->type->type, "send");
+		n->right = assignconv(r, l->type->type, "send");
 		// TODO: more aggressive
 		n->etype = 0;
 		n->type = T;
@@ -1655,6 +1655,7 @@ reswitch:
 	case OGOTO:
 	case OLABEL:
 	case OXFALL:
+	case OVARKILL:
 		ok |= Etop;
 		goto ret;
 

src/cmd/gc/walk.c

@@ -202,6 +202,7 @@ walkstmt(Node **np)
 	case ODCLCONST:
 	case ODCLTYPE:
 	case OCHECKNIL:
+	case OVARKILL:
 		break;
 
 	case OBLOCK:
@@ -361,8 +362,8 @@ void
 walkexpr(Node **np, NodeList **init)
 {
 	Node *r, *l, *var, *a;
-	Node *map, *key, *keyvar;
-	NodeList *ll, *lr, *lpost;
+	Node *map, *key;
+	NodeList *ll, *lr;
 	Type *t;
 	int et, old_safemode;
 	int64 v;
@@ -476,7 +477,6 @@ walkexpr(Node **np, NodeList **init)
 	case ORSH:
 		walkexpr(&n->left, init);
 		walkexpr(&n->right, init);
-	shiftwalked:
 		t = n->left->type;
 		n->bounded = bounded(n->right, 8*t->width);
 		if(debug['m'] && n->etype && !isconst(n->right, CTINT))
@@ -602,13 +602,32 @@ walkexpr(Node **np, NodeList **init)
 	case OAS:
 		*init = concat(*init, n->ninit);
 		n->ninit = nil;
+
 		walkexpr(&n->left, init);
 		n->left = safeexpr(n->left, init);
 
 		if(oaslit(n, init))
 			goto ret;
 
-		walkexpr(&n->right, init);
+		if(n->right == N)
+			goto ret;
+
+		switch(n->right->op) {
+		default:
+			walkexpr(&n->right, init);
+			break;
+		
+		case ORECV:
+			// x = <-c; n->left is x, n->right->left is c.
+			// orderstmt made sure x is addressable.
+			walkexpr(&n->right->left, init);
+			n1 = nod(OADDR, n->left, N);
+			r = n->right->left; // the channel
+			n = mkcall1(chanfn("chanrecv1", 2, r->type), T, init, typename(r->type), r, n1);
+			walkexpr(&n, init);
+			goto ret;
+		}
+
 		if(n->left != N && n->right != N) {
 			r = convas(nod(OAS, n->left, n->right), init);
 			r->dodata = n->dodata;
@@ -618,7 +637,6 @@ walkexpr(Node **np, NodeList **init)
 		goto ret;
 
 	case OAS2:
-	as2:
 		*init = concat(*init, n->ninit);
 		n->ninit = nil;
 		walkexprlistsafe(n->list, init);
@@ -637,52 +655,26 @@ walkexpr(Node **np, NodeList **init)
 		walkexpr(&r, init);
 		l = n->list->n;
 
-		// all the really hard stuff - explicit function calls and so on -
-		// is gone, but map assignments remain.
-		// if there are map assignments here, assign via
-		// temporaries, because ascompatet assumes
-		// the targets can be addressed without function calls
-		// and map index has an implicit one.
-		lpost = nil;
-		if(l->op == OINDEXMAP) {
-			var = temp(l->type);
-			n->list->n = var;
-			a = nod(OAS, l, var);
-			typecheck(&a, Etop);
-			lpost = list(lpost, a);
-		}
 		l = n->list->next->n;
-		if(l->op == OINDEXMAP) {
-			var = temp(l->type);
-			n->list->next->n = var;
-			a = nod(OAS, l, var);
-			typecheck(&a, Etop);
-			lpost = list(lpost, a);
-		}
 		ll = ascompatet(n->op, n->list, &r->type, 0, init);
-		walkexprlist(lpost, init);
-		n = liststmt(concat(concat(list1(r), ll), lpost));
+		n = liststmt(concat(list1(r), ll));
 		goto ret;
 
 	case OAS2RECV:
+		// x, y = <-c
+		// orderstmt made sure x is addressable.
 		*init = concat(*init, n->ninit);
 		n->ninit = nil;
 		r = n->rlist->n;
 		walkexprlistsafe(n->list, init);
 		walkexpr(&r->left, init);
-		var = temp(r->left->type->type);
-		if(haspointers(var->type)) {
-			// clear for garbage collector - var is live during chanrecv2 call.
-			a = nod(OAS, var, N);
-			typecheck(&a, Etop);
-			*init = concat(*init, list1(a));
-		}
-		n1 = nod(OADDR, var, N);
+		n1 = nod(OADDR, n->list->n, N);
+		n1->etype = 1; // addr does not escape
 		fn = chanfn("chanrecv2", 2, r->left->type);
 		r = mkcall1(fn, types[TBOOL], init, typename(r->left->type), r->left, n1);
-		n->op = OAS2;
-		n->rlist = concat(list1(var), list1(r));
-		goto as2;
+		n = nod(OAS, n->list->next->n, r);
+		typecheck(&n, Etop);
+		goto ret;
 
 	case OAS2MAPR:
 		// a,b = m[i];
@@ -714,15 +706,8 @@ walkexpr(Node **np, NodeList **init)
 			key = r->right;
 		} else {
 			// standard version takes key by reference
-			if(islvalue(r->right)) {
-				key = nod(OADDR, r->right, N);
-			} else {
-				keyvar = temp(t->down);
-				n1 = nod(OAS, keyvar, r->right);
-				typecheck(&n1, Etop);
-				*init = list(*init, n1);
-				key = nod(OADDR, keyvar, N);
-			}
+			// orderexpr made sure key is addressable.
+			key = nod(OADDR, r->right, N);
 			p = "mapaccess2";
 		}
 
@@ -741,7 +726,6 @@ walkexpr(Node **np, NodeList **init)
 		n->list->n = var;
 		walkexpr(&n, init);
 		*init = list(*init, n);
-		
 		n = nod(OAS, a, nod(OIND, var, N));
 		typecheck(&n, Etop);
 		walkexpr(&n, init);
@@ -758,15 +742,8 @@ walkexpr(Node **np, NodeList **init)
 		key = n->list->next->n;
 		walkexpr(&map, init);
 		walkexpr(&key, init);
-		if(islvalue(key)) {
-			key = nod(OADDR, key, N);
-		} else {
-			keyvar = temp(key->type);
-			n1 = nod(OAS, keyvar, key);
-			typecheck(&n1, Etop);
-			*init = list(*init, n1);
-			key = nod(OADDR, keyvar, N);
-		}
+		// orderstmt made sure key is addressable.
+		key = nod(OADDR, key, N);
 		t = map->type;
 		n = mkcall1(mapfndel("mapdelete", t), T, init, typename(t), map, key);
 		goto ret;
@@ -981,51 +958,6 @@ walkexpr(Node **np, NodeList **init)
 		walkexpr(&n->left, init);
 		goto ret;
 
-	case OASOP:
-		if(n->etype == OANDNOT) {
-			n->etype = OAND;
-			n->right = nod(OCOM, n->right, N);
-			typecheck(&n->right, Erv);
-		}
-		n->left = safeexpr(n->left, init);
-		walkexpr(&n->left, init);
-		l = n->left;
-		walkexpr(&n->right, init);
-
-		/*
-		 * on 32-bit arch, rewrite 64-bit ops into l = l op r.
-		 * on 386, rewrite float ops into l = l op r.
-		 * everywhere, rewrite map ops into l = l op r.
-		 * everywhere, rewrite string += into l = l op r.
-		 * everywhere, rewrite integer/complex /= into l = l op r.
-		 * TODO(rsc): Maybe this rewrite should be done always?
-		 */
-		et = n->left->type->etype;
-		if((widthptr == 4 && (et == TUINT64 || et == TINT64)) ||
-		   (thechar == '8' && isfloat[et]) ||
-		   l->op == OINDEXMAP ||
-		   et == TSTRING ||
-		   (!isfloat[et] && n->etype == ODIV) ||
-		   n->etype == OMOD) {
-			l = safeexpr(n->left, init);
-			a = l;
-			if(a->op == OINDEXMAP) {
-				// map index has "lhs" bit set in a->etype.
-				// make a copy so we can clear it on the rhs.
-				a = nod(OXXX, N, N);
-				*a = *l;
-				a->etype = 0;
-			}
-			r = nod(OAS, l, nod(n->etype, a, n->right));
-			typecheck(&r, Etop);
-			walkexpr(&r, init);
-			n = r;
-			goto ret;
-		}
-		if(n->etype == OLSH || n->etype == ORSH)
-			goto shiftwalked;
-		goto ret;
-
 	case OANDNOT:
 		walkexpr(&n->left, init);
 		n->op = OAND;
@@ -1159,16 +1091,9 @@ walkexpr(Node **np, NodeList **init)
 			// fast versions take key by value
 			key = n->right;
 		} else {
-			// standard version takes key by reference
-			if(islvalue(n->right)) {
-				key = nod(OADDR, n->right, N);
-			} else {
-				keyvar = temp(t->down);
-				n1 = nod(OAS, keyvar, n->right);
-				typecheck(&n1, Etop);
-				*init = list(*init, n1);
-				key = nod(OADDR, keyvar, N);
-			}
+			// standard version takes key by reference.
+			// orderexpr made sure key is addressable.
+			key = nod(OADDR, n->right, N);
 			p = "mapaccess1";
 		}
 		n = mkcall1(mapfn(p, t), ptrto(t->type), init, typename(t), n->left, key);
@@ -1181,20 +1106,7 @@ walkexpr(Node **np, NodeList **init)
 		goto ret;
 
 	case ORECV:
-		walkexpr(&n->left, init);
-		var = temp(n->left->type->type);
-		if(haspointers(var->type)) {
-			// clear for garbage collector - var is live during chanrecv1 call.
-			a = nod(OAS, var, N);
-			typecheck(&a, Etop);
-			*init = concat(*init, list1(a));
-		}
-		n1 = nod(OADDR, var, N);
-		n = mkcall1(chanfn("chanrecv1", 2, n->left->type), T, init, typename(n->left->type), n->left, n1);
-		walkexpr(&n, init);
-		*init = list(*init, n);
-		n = var;
-		goto ret;
+		fatal("walkexpr ORECV"); // should see inside OAS only
 
 	case OSLICE:
 		if(n->right != N && n->right->left == N && n->right->right == N) { // noop
@@ -1462,15 +1374,7 @@ walkexpr(Node **np, NodeList **init)
 		n1 = n->right;
 		n1 = assignconv(n1, n->left->type->type, "chan send");
 		walkexpr(&n1, init);
-		if(islvalue(n1)) {
-			n1 = nod(OADDR, n1, N);
-		} else {
-			var = temp(n1->type);
-			n1 = nod(OAS, var, n1);
-			typecheck(&n1, Etop);
-			*init = list(*init, n1);
-			n1 = nod(OADDR, var, N);
-		}
+		n1 = nod(OADDR, n1, N);
 		n = mkcall1(chanfn("chansend1", 2, n->left->type), T, init, typename(n->left->type), n->left, n1);
 		goto ret;
 
@@ -2008,8 +1912,7 @@ static Node*
 convas(Node *n, NodeList **init)
 {
 	Type *lt, *rt;
-	Node *map, *key, *keyvar, *val, *valvar;
-	Node *n1;
+	Node *map, *key, *val;
 
 	if(n->op != OAS)
 		fatal("convas: not OAS %O", n->op);
@@ -2036,24 +1939,9 @@ convas(Node *n, NodeList **init)
 		walkexpr(&map, init);
 		walkexpr(&key, init);
 		walkexpr(&val, init);
-		if(islvalue(key)) {
-			key = nod(OADDR, key, N);
-		} else {
-			keyvar = temp(key->type);
-			n1 = nod(OAS, keyvar, key);
-			typecheck(&n1, Etop);
-			*init = list(*init, n1);
-			key = nod(OADDR, keyvar, N);
-		}
-		if(islvalue(val)) {
-			val = nod(OADDR, val, N);
-		} else {
-			valvar = temp(val->type);
-			n1 = nod(OAS, valvar, val);
-			typecheck(&n1, Etop);
-			*init = list(*init, n1);
-			val = nod(OADDR, valvar, N);
-		}
+		// orderexpr made sure key and val are addressable.
+		key = nod(OADDR, key, N);
+		val = nod(OADDR, val, N);
 		n = mkcall1(mapfn("mapassign1", map->type), T, init,
 			typename(map->type), map, key, val);
 		goto out;

test/live.go

@@ -212,3 +212,121 @@ func f15() {
 }
 
 func g15() string
+
+// Checking that various temporaries do not persist or cause
+// ambiguously live values that must be zeroed.
+// The exact temporary names are inconsequential but we are
+// trying to check that there is only one at any given site,
+// and also that none show up in "ambiguously live" messages.
+
+var m map[string]int
+
+func f16() {
+	if b {
+		delete(m, "hi") // ERROR "live at call to mapdelete: autotmp_[0-9]+$"
+	}
+	delete(m, "hi") // ERROR "live at call to mapdelete: autotmp_[0-9]+$"
+	delete(m, "hi") // ERROR "live at call to mapdelete: autotmp_[0-9]+$"
+}
+
+var m2s map[string]*byte
+var m2 map[[2]string]*byte
+var x2 [2]string
+var bp *byte
+
+func f17a() {
+	// value temporary only
+	if b {
+		m2[x2] = nil // ERROR "live at call to mapassign1: autotmp_[0-9]+$"
+	}
+	m2[x2] = nil // ERROR "live at call to mapassign1: autotmp_[0-9]+$"
+	m2[x2] = nil // ERROR "live at call to mapassign1: autotmp_[0-9]+$"
+}
+
+func f17b() {
+	// key temporary only
+	if b {
+		m2s["x"] = bp // ERROR "live at call to mapassign1: autotmp_[0-9]+$"
+	}
+	m2s["x"] = bp // ERROR "live at call to mapassign1: autotmp_[0-9]+$"
+	m2s["x"] = bp // ERROR "live at call to mapassign1: autotmp_[0-9]+$"
+}
+
+func f17c() {
+	// key and value temporaries
+	if b {
+		m2s["x"] = nil // ERROR "live at call to mapassign1: autotmp_[0-9]+ autotmp_[0-9]+$"
+	}
+	m2s["x"] = nil // ERROR "live at call to mapassign1: autotmp_[0-9]+ autotmp_[0-9]+$"
+	m2s["x"] = nil // ERROR "live at call to mapassign1: autotmp_[0-9]+ autotmp_[0-9]+$"
+}
+
+func g18() [2]string
+
+func f18() {
+	// key temporary for mapaccess.
+	// temporary introduced by orderexpr.
+	var z *byte
+	if b {
+		z = m2[g18()] // ERROR "live at call to mapaccess1: autotmp_[0-9]+$"
+	}
+	z = m2[g18()] // ERROR "live at call to mapaccess1: autotmp_[0-9]+$"
+	z = m2[g18()] // ERROR "live at call to mapaccess1: autotmp_[0-9]+$"
+	print(z)
+}
+
+var ch chan *byte
+
+func f19() {
+	// dest temporary for channel receive.
+	var z *byte
+	
+	if b {
+		z = <-ch // ERROR "live at call to chanrecv1: autotmp_[0-9]+$"
+	}
+	z = <-ch // ERROR "live at call to chanrecv1: autotmp_[0-9]+$"
+	z = <-ch // ERROR "live at call to chanrecv1: autotmp_[0-9]+$"
+	print(z)
+}
+
+func f20() {
+	// src temporary for channel send
+	if b {
+		ch <- nil // ERROR "live at call to chansend1: autotmp_[0-9]+$"
+	}
+	ch <- nil // ERROR "live at call to chansend1: autotmp_[0-9]+$"
+	ch <- nil // ERROR "live at call to chansend1: autotmp_[0-9]+$"
+}
+
+func f21() {
+	// key temporary for mapaccess using array literal key.
+	var z *byte
+	if b {
+		z = m2[[2]string{"x", "y"}] // ERROR "live at call to mapaccess1: autotmp_[0-9]+$"
+	}
+	z = m2[[2]string{"x", "y"}] // ERROR "live at call to mapaccess1: autotmp_[0-9]+$"
+	z = m2[[2]string{"x", "y"}] // ERROR "live at call to mapaccess1: autotmp_[0-9]+$"
+	print(z)
+}
+
+func f23() {
+	// key temporary for two-result map access using array literal key.
+	var z *byte
+	var ok bool
+	if b {
+		z, ok = m2[[2]string{"x", "y"}] // ERROR "live at call to mapaccess2: autotmp_[0-9]+$"
+	}
+	z, ok = m2[[2]string{"x", "y"}] // ERROR "live at call to mapaccess2: autotmp_[0-9]+$"
+	z, ok = m2[[2]string{"x", "y"}] // ERROR "live at call to mapaccess2: autotmp_[0-9]+$"
+	print(z, ok)
+}
+
+func f24() {
+	// key temporary for map access using array literal key.
+	// value temporary too.
+	if b {
+		m2[[2]string{"x", "y"}] = nil // ERROR "live at call to mapassign1: autotmp_[0-9]+ autotmp_[0-9]+$"
+	}
+	m2[[2]string{"x", "y"}] = nil // ERROR "live at call to mapassign1: autotmp_[0-9]+ autotmp_[0-9]+$"
+	m2[[2]string{"x", "y"}] = nil // ERROR "live at call to mapassign1: autotmp_[0-9]+ autotmp_[0-9]+$"
+}

test/reorder2.go

@@ -168,6 +168,175 @@ func main() {
 	}
 	log = ""
 	
+	x := 0
+	switch x {
+	case 0:
+		if a("1")("2")("3"); log != "a(1)a(2)a(3)" {
+			println("in switch, expecting a(1)a(2)a(3) , got ", log)
+			err++
+		}
+		log = ""
+	
+		if t.a("1").a(t.b("2")); log != "a(1)b(2)a(2)" {
+			println("in switch, expecting a(1)b(2)a(2), got ", log)
+			err++
+		}
+		log = ""
+		if a("3")(b("4"))(b("5")); log != "a(3)b(4)a(4)b(5)a(5)" {
+			println("in switch, expecting a(3)b(4)a(4)b(5)a(5), got ", log)
+			err++
+		}
+		log = ""
+		var i I = T1(0)
+		if i.a("6").a(i.b("7")).a(i.b("8")).a(i.b("9")); log != "a(6)b(7)a(7)b(8)a(8)b(9)a(9)" {
+			println("in switch, expecting a(6)ba(7)ba(8)ba(9), got", log)
+			err++
+		}
+		log = ""
+	}
+	
+	c := make(chan int, 1)
+	c <- 1
+	select {
+	case c <- 0:
+	case c <- 1:
+	case <-c:
+		if a("1")("2")("3"); log != "a(1)a(2)a(3)" {
+			println("in select1, expecting a(1)a(2)a(3) , got ", log)
+			err++
+		}
+		log = ""
+	
+		if t.a("1").a(t.b("2")); log != "a(1)b(2)a(2)" {
+			println("in select1, expecting a(1)b(2)a(2), got ", log)
+			err++
+		}
+		log = ""
+		if a("3")(b("4"))(b("5")); log != "a(3)b(4)a(4)b(5)a(5)" {
+			println("in select1, expecting a(3)b(4)a(4)b(5)a(5), got ", log)
+			err++
+		}
+		log = ""
+		var i I = T1(0)
+		if i.a("6").a(i.b("7")).a(i.b("8")).a(i.b("9")); log != "a(6)b(7)a(7)b(8)a(8)b(9)a(9)" {
+			println("in select1, expecting a(6)ba(7)ba(8)ba(9), got", log)
+			err++
+		}
+		log = ""
+	}
+
+	c <- 1
+	select {
+	case <-c:
+		if a("1")("2")("3"); log != "a(1)a(2)a(3)" {
+			println("in select2, expecting a(1)a(2)a(3) , got ", log)
+			err++
+		}
+		log = ""
+	
+		if t.a("1").a(t.b("2")); log != "a(1)b(2)a(2)" {
+			println("in select2, expecting a(1)b(2)a(2), got ", log)
+			err++
+		}
+		log = ""
+		if a("3")(b("4"))(b("5")); log != "a(3)b(4)a(4)b(5)a(5)" {
+			println("in select2, expecting a(3)b(4)a(4)b(5)a(5), got ", log)
+			err++
+		}
+		log = ""
+		var i I = T1(0)
+		if i.a("6").a(i.b("7")).a(i.b("8")).a(i.b("9")); log != "a(6)b(7)a(7)b(8)a(8)b(9)a(9)" {
+			println("in select2, expecting a(6)ba(7)ba(8)ba(9), got", log)
+			err++
+		}
+		log = ""
+	}
+
+	c <- 1
+	select {
+	default:
+	case c<-1:
+	case <-c:
+		if a("1")("2")("3"); log != "a(1)a(2)a(3)" {
+			println("in select3, expecting a(1)a(2)a(3) , got ", log)
+			err++
+		}
+		log = ""
+	
+		if t.a("1").a(t.b("2")); log != "a(1)b(2)a(2)" {
+			println("in select3, expecting a(1)b(2)a(2), got ", log)
+			err++
+		}
+		log = ""
+		if a("3")(b("4"))(b("5")); log != "a(3)b(4)a(4)b(5)a(5)" {
+			println("in select3, expecting a(3)b(4)a(4)b(5)a(5), got ", log)
+			err++
+		}
+		log = ""
+		var i I = T1(0)
+		if i.a("6").a(i.b("7")).a(i.b("8")).a(i.b("9")); log != "a(6)b(7)a(7)b(8)a(8)b(9)a(9)" {
+			println("in select3, expecting a(6)ba(7)ba(8)ba(9), got", log)
+			err++
+		}
+		log = ""
+	}
+
+	c <- 1
+	select {
+	default:
+	case <-c:
+		if a("1")("2")("3"); log != "a(1)a(2)a(3)" {
+			println("in select4, expecting a(1)a(2)a(3) , got ", log)
+			err++
+		}
+		log = ""
+	
+		if t.a("1").a(t.b("2")); log != "a(1)b(2)a(2)" {
+			println("in select4, expecting a(1)b(2)a(2), got ", log)
+			err++
+		}
+		log = ""
+		if a("3")(b("4"))(b("5")); log != "a(3)b(4)a(4)b(5)a(5)" {
+			println("in select4, expecting a(3)b(4)a(4)b(5)a(5), got ", log)
+			err++
+		}
+		log = ""
+		var i I = T1(0)
+		if i.a("6").a(i.b("7")).a(i.b("8")).a(i.b("9")); log != "a(6)b(7)a(7)b(8)a(8)b(9)a(9)" {
+			println("in select4, expecting a(6)ba(7)ba(8)ba(9), got", log)
+			err++
+		}
+		log = ""
+	}
+
+	select {
+	case <-c:
+	case <-c:
+	default:
+		if a("1")("2")("3"); log != "a(1)a(2)a(3)" {
+			println("in select5, expecting a(1)a(2)a(3) , got ", log)
+			err++
+		}
+		log = ""
+	
+		if t.a("1").a(t.b("2")); log != "a(1)b(2)a(2)" {
+			println("in select5, expecting a(1)b(2)a(2), got ", log)
+			err++
+		}
+		log = ""
+		if a("3")(b("4"))(b("5")); log != "a(3)b(4)a(4)b(5)a(5)" {
+			println("in select5, expecting a(3)b(4)a(4)b(5)a(5), got ", log)
+			err++
+		}
+		log = ""
+		var i I = T1(0)
+		if i.a("6").a(i.b("7")).a(i.b("8")).a(i.b("9")); log != "a(6)b(7)a(7)b(8)a(8)b(9)a(9)" {
+			println("in select5, expecting a(6)ba(7)ba(8)ba(9), got", log)
+			err++
+		}
+		log = ""
+	}
+
 	if err > 0 {
 		panic("fail")
 	}