race: gc changes

This is the first part of a bigger change that adds data race detection feature:
https://golang.org/cl/6456044
This change makes gc compiler instrument memory accesses when supplied with -b flag.

R=rsc, nigeltao, lvd
CC=golang-dev
https://golang.org/cl/6497074

src/cmd/gc/builtin.c

@@ -104,6 +104,10 @@ char *runtimeimport =
 	"func @\"\".int64tofloat64(? int64) (? float64)\n"
 	"func @\"\".uint64tofloat64(? uint64) (? float64)\n"
 	"func @\"\".complex128div(@\"\".num complex128, @\"\".den complex128) (@\"\".quo complex128)\n"
+	"func @\"\".racefuncenter()\n"
+	"func @\"\".racefuncexit()\n"
+	"func @\"\".raceread(? uintptr)\n"
+	"func @\"\".racewrite(? uintptr)\n"
 	"\n"
 	"$$\n";
 char *unsafeimport =

src/cmd/gc/doc.go

@@ -54,6 +54,8 @@ Flags:
 		disallow importing packages not marked as safe
 	-V
 		print the compiler version
+	-b
+		compile with race detection enabled
 
 There are also a number of debugging flags; run the command with no arguments
 to get a usage message.

src/cmd/gc/go.h

@@ -838,6 +838,7 @@ EXTERN	Pkg*	builtinpkg;	// fake package for builtins
 EXTERN	Pkg*	gostringpkg;	// fake pkg for Go strings
 EXTERN	Pkg*	itabpkg;	// fake pkg for itab cache
 EXTERN	Pkg*	runtimepkg;	// package runtime
+EXTERN	Pkg*	racepkg;	// package runtime/race
 EXTERN	Pkg*	stringpkg;	// fake package for C strings
 EXTERN	Pkg*	typepkg;	// fake package for runtime type info
 EXTERN	Pkg*	weaktypepkg;	// weak references to runtime type info
@@ -1442,3 +1443,8 @@ void	zname(Biobuf *b, Sym *s, int t);
 #pragma	varargck	type	"V"	Val*
 #pragma	varargck	type	"Y"	char*
 #pragma	varargck	type	"Z"	Strlit*
+
+/*
+ *	racewalk.c
+ */
+void	racewalk(Node *fn);

src/cmd/gc/lex.c

@@ -264,7 +264,12 @@ main(int argc, char *argv[])
 		print("%cg version %s%s%s\n", thechar, getgoversion(), *p ? " " : "", p);
 		exits(0);
 	} ARGEND
-	
+
+	if(debug['b']) {
+		racepkg = mkpkg(strlit("runtime/race"));
+		racepkg->name = "race";
+	}
+
 	// enable inlining.  for now:
 	//	default: inlining on.  (debug['l'] == 1)
 	//	-l: inlining off  (debug['l'] == 0)
@@ -530,7 +535,7 @@ static int
 findpkg(Strlit *name)
 {
 	Idir *p;
-	char *q;
+	char *q, *race;
 
 	if(islocalname(name)) {
 		if(safemode)
@@ -568,10 +573,13 @@ findpkg(Strlit *name)
 			return 1;
 	}
 	if(goroot != nil) {
-		snprint(namebuf, sizeof(namebuf), "%s/pkg/%s_%s/%Z.a", goroot, goos, goarch, name);
+		race = "";
+		if(debug['b'])
+			race = "_race";
+		snprint(namebuf, sizeof(namebuf), "%s/pkg/%s_%s%s/%Z.a", goroot, goos, goarch, race, name);
 		if(access(namebuf, 0) >= 0)
 			return 1;
-		snprint(namebuf, sizeof(namebuf), "%s/pkg/%s_%s/%Z.%c", goroot, goos, goarch, name, thechar);
+		snprint(namebuf, sizeof(namebuf), "%s/pkg/%s_%s%s/%Z.%c", goroot, goos, goarch, race, name, thechar);
 		if(access(namebuf, 0) >= 0)
 			return 1;
 	}

src/cmd/gc/pgen.c

@@ -63,6 +63,10 @@ compile(Node *fn)
 	walk(curfn);
 	if(nerrors != 0)
 		goto ret;
+	if(debug['b'])
+		racewalk(curfn);
+	if(nerrors != 0)
+		goto ret;
 
 	continpc = P;
 	breakpc = P;

src/cmd/gc/racewalk.c

@@ -0,0 +1,431 @@
+// Copyright 2012 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.
+
+// The racewalk pass modifies the code tree for the function as follows:
+//
+// 1. It inserts a call to racefuncenter at the beginning of each function.
+// 2. It inserts a call to racefuncexit at the end of each function.
+// 3. It inserts a call to raceread before each memory read.
+// 4. It inserts a call to racewrite before each memory write.
+//
+// The rewriting is not yet complete. Certain nodes are not rewritten
+// but should be.
+
+#include <u.h>
+#include <libc.h>
+#include "go.h"
+#include "opnames.h"
+
+//TODO: do not instrument initialization as writes:
+// a := make([]int, 10)
+
+static void racewalklist(NodeList *l, NodeList **init);
+static void racewalknode(Node **np, NodeList **init, int wr, int skip);
+static void callinstr(Node *n, NodeList **init, int wr, int skip);
+static Node* uintptraddr(Node *n);
+static Node* basenod(Node *n);
+
+static const char *omitPkgs[] = {"runtime", "runtime/race", "sync", "sync/atomic"};
+
+void
+racewalk(Node *fn)
+{
+	int i;
+	Node *nd;
+	char s[1024];
+
+	if(myimportpath) {
+		for(i=0; i<nelem(omitPkgs); i++) {
+			if(strcmp(myimportpath, omitPkgs[i]) == 0)
+				return;
+		}
+	}
+
+	nd = mkcall("racefuncenter", T, nil);
+	fn->enter = list(fn->enter, nd);
+	nd = mkcall("racefuncexit", T, nil);
+	fn->exit = list(fn->exit, nd); // works fine if (!fn->exit)
+	racewalklist(curfn->nbody, nil);
+
+	if(debug['W']) {
+		snprint(s, sizeof(s), "after racewalk %S", curfn->nname->sym);
+		dumplist(s, curfn->nbody);
+	}
+}
+
+static void
+racewalklist(NodeList *l, NodeList **init)
+{
+	NodeList *instr;
+
+	for(; l; l = l->next) {
+		instr = nil;
+		racewalknode(&l->n, &instr, 0, 0);
+		if(init == nil)
+			l->n->ninit = concat(l->n->ninit, instr);
+		else
+			*init = concat(*init, instr);
+	}
+}
+
+// walkexpr and walkstmt combined
+// walks the tree and adds calls to the
+// instrumentation code to top-level (statement) nodes' init
+static void
+racewalknode(Node **np, NodeList **init, int wr, int skip)
+{
+	Node *n, *n1;
+
+	n = *np;
+
+	if(n == N)
+		return;
+	if(0)
+		print("op=%s, left=[ %N ], right=[ %N ], right's type=%T, n's type=%T, n's class=%d\n",
+			opnames[n->op], n->left, n->right, n->right ? n->right->type : nil, n->type, n->class);
+	setlineno(n);
+
+	switch(n->op) {
+	default:
+		fatal("racewalk: unknown node type %O", n->op);
+
+	case OASOP:
+	case OAS:
+	case OAS2:
+	case OAS2DOTTYPE:
+	case OAS2RECV:
+	case OAS2FUNC:
+	case OAS2MAPR:
+		racewalklist(n->ninit, init);
+		racewalknode(&n->left, init, 1, 0);
+		racewalknode(&n->right, init, 0, 0);
+		goto ret;
+
+	case OBLOCK:
+		// leads to crashes.
+		//racewalklist(n->list, nil);
+		goto ret;
+
+	case ODEFER:
+		racewalknode(&n->left, init, 0, 0);
+		goto ret;
+
+	case OFOR:
+		racewalklist(n->ninit, nil);
+		if(n->ntest != N)
+			racewalklist(n->ntest->ninit, nil);
+		racewalknode(&n->nincr, init, wr, 0);
+		racewalklist(n->nbody, nil);
+		goto ret;
+
+	case OIF:
+		racewalklist(n->ninit, nil);
+		racewalknode(&n->ntest, &n->ninit, wr, 0);
+		racewalklist(n->nbody, nil);
+		racewalklist(n->nelse, nil);
+		goto ret;
+
+	case OPROC:
+		racewalknode(&n->left, init, 0, 0);
+		goto ret;
+
+	case OCALLINTER:
+		racewalknode(&n->left, init, 0, 0);
+		racewalklist(n->list, init);
+		goto ret;
+
+	case OCALLFUNC:
+		racewalknode(&n->left, init, 0, 0);
+		racewalklist(n->ninit, init);
+		racewalklist(n->list, init);
+		goto ret;
+
+	case OCALLMETH:
+		racewalklist(n->list, init);
+		goto ret;
+
+	case ORETURN:
+		racewalklist(n->list, nil);
+		goto ret;
+
+	case OSELECT:
+		// n->nlist is nil by now because this code
+		// is running after walkselect
+		racewalklist(n->nbody, nil);
+		goto ret;
+
+	case OSWITCH:
+		racewalklist(n->ninit, nil);
+		if(n->ntest->op == OTYPESW)
+			// don't bother, we have static typization
+			return;
+		racewalknode(&n->ntest, &n->ninit, 0, 0);
+		racewalklist(n->nbody, nil);
+		goto ret;
+
+	case OEMPTY:
+		racewalklist(n->ninit, nil);
+		goto ret;
+
+	case ONOT:
+	case OMINUS:
+	case OPLUS:
+	case OREAL:
+	case OIMAG:
+		racewalknode(&n->left, init, wr, 0);
+		goto ret;
+
+	case ODOTINTER:
+		racewalknode(&n->left, init, 0, 0);
+		goto ret;
+
+	case ODOT:
+		callinstr(n, init, wr, skip);
+		racewalknode(&n->left, init, 0, 1);
+		goto ret;
+
+	case ODOTPTR: // dst = (*x).f with implicit *; otherwise it's ODOT+OIND
+		callinstr(n, init, wr, skip);
+		racewalknode(&n->left, init, 0, 0);
+		goto ret;
+
+	case OIND: // *p
+		callinstr(n, init, wr, skip);
+		racewalknode(&n->left, init, 0, 0);
+		goto ret;
+
+	case OLEN:
+	case OCAP:
+		racewalknode(&n->left, init, 0, 0);
+		if(istype(n->left->type, TMAP)) {
+			// crashes on len(m[0]) or len(f())
+			USED(&n1);
+			/*
+			n1 = nod(OADDR, n->left, N);
+			n1 = conv(n1, types[TUNSAFEPTR]);
+			n1 = conv(n1, ptrto(ptrto(types[TINT8])));
+			n1 = nod(OIND, n1, N);
+			n1 = nod(OIND, n1, N);
+			typecheck(&n1, Erv);
+			callinstr(n1, init, 0, skip);
+			*/
+		}
+		goto ret;
+
+	case OLSH:
+	case ORSH:
+	case OAND:
+	case OANDNOT:
+	case OOR:
+	case OXOR:
+	case OSUB:
+	case OMUL:
+	case OEQ:
+	case ONE:
+	case OLT:
+	case OLE:
+	case OGE:
+	case OGT:
+	case OADD:
+	case OCOMPLEX:
+		racewalknode(&n->left, init, wr, 0);
+		racewalknode(&n->right, init, wr, 0);
+		goto ret;
+
+	case OANDAND:
+	case OOROR:
+		racewalknode(&n->left, init, wr, 0);
+		// It requires more complex tree transformation,
+		// because we don't know whether it will be executed or not.
+		//racewalknode(&n->right, init, wr, 0);
+		goto ret;
+
+	case ONAME:
+		callinstr(n, init, wr, skip);
+		goto ret;
+
+	case OCONV:
+		racewalknode(&n->left, init, wr, 0);
+		goto ret;
+
+	case OCONVNOP:
+		racewalknode(&n->left, init, wr, 0);
+		goto ret;
+
+	case ODIV:
+	case OMOD:
+		// TODO(dvyukov): add a test for this
+		racewalknode(&n->left, init, wr, 0);
+		racewalknode(&n->right, init, wr, 0);
+		goto ret;
+
+	case OINDEX:
+		if(n->left->type->etype != TSTRING)
+			callinstr(n, init, wr, skip);
+		if(!isfixedarray(n->left->type))
+			racewalknode(&n->left, init, 0, 0);
+		racewalknode(&n->right, init, 0, 0);
+		goto ret;
+
+	case OSLICE:
+	case OSLICEARR:
+		// Seems to only lead to double instrumentation.
+		//racewalklist(n->ninit, init);
+		//racewalknode(&n->left, init, 0, 0);
+		//racewalklist(n->list, init);
+		goto ret;
+
+	case OADDR:
+		racewalknode(&n->left, init, 0, 1);
+		goto ret;
+
+	// should not appear in AST by now
+	case OSEND:
+	case ORECV:
+	case OCLOSE:
+	case ONEW:
+	case OXCASE:
+	case OXFALL:
+	case OCASE:
+	case OPANIC:
+	case ORECOVER:
+		yyerror("racewalk: %O must be lowered by now", n->op);
+		goto ret;
+
+	// does not require instrumentation
+	case OINDEXMAP:  // implemented in runtime
+	case OPRINT:  // don't bother instrumenting it
+	case OPRINTN:  // don't bother instrumenting it
+		goto ret;
+
+	// unimplemented
+	case OCMPSTR:
+	case OADDSTR:
+	case OSLICESTR:
+	case OAPPEND:
+	case OCOPY:
+	case OMAKECHAN:
+	case OMAKEMAP:
+	case OMAKESLICE:
+	case ORUNESTR:
+	case OARRAYBYTESTR:
+	case OARRAYRUNESTR:
+	case OSTRARRAYBYTE:
+	case OSTRARRAYRUNE:
+	case OCMPIFACE:
+	case OARRAYLIT:
+	case OMAPLIT:
+	case OSTRUCTLIT:
+	case OCLOSURE:
+	case ODOTTYPE:
+	case ODOTTYPE2:
+	case OCONVIFACE:
+	case OCALL:
+	case OBREAK:
+	case ODCL:
+	case OCONTINUE:
+	case OFALL:
+	case OGOTO:
+	case OLABEL:
+	case ODCLCONST:
+	case ODCLTYPE:
+	case OLITERAL:
+	case ORANGE:
+	case OTYPE:
+	case ONONAME:
+	case OINDREG:
+	case OCOM:
+	case ODOTMETH:
+	case OEFACE:
+	case OITAB:
+	case OEXTEND:
+	case OHMUL:
+	case OLROT:
+	case ORROTC:
+		goto ret;
+	}
+
+ret:
+	*np = n;
+}
+
+static void
+callinstr(Node *n, NodeList **init, int wr, int skip)
+{
+	Node *f, *b;
+	Type *t, *t1;
+	int class;
+
+	//print("callinstr for %N [ %s ] etype=%d class=%d\n",
+	//	  n, opnames[n->op], n->type ? n->type->etype : -1, n->class);
+
+	if(skip || n->type == T || n->type->etype >= TIDEAL)
+		return;
+	t = n->type;
+	if(n->op == ONAME) {
+		if(n->sym != S) {
+			if(n->sym->name != nil) {
+				if(strncmp(n->sym->name, "_", sizeof("_")-1) == 0)
+					return;
+				if(strncmp(n->sym->name, "autotmp_", sizeof("autotmp_")-1) == 0)
+					return;
+				if(strncmp(n->sym->name, "statictmp_", sizeof("statictmp_")-1) == 0)
+					return;
+			}
+		}
+	}
+	if (t->etype == TSTRUCT) {
+		for(t1=t->type; t1; t1=t1->down) {
+			if(t1->sym && strncmp(t1->sym->name, "_", sizeof("_")-1)) {
+				n = treecopy(n);
+				f = nod(OXDOT, n, newname(t1->sym));
+				typecheck(&f, Erv);
+				callinstr(f, init, wr, 0);
+			}
+		}
+		return;
+	}
+
+	b = basenod(n);
+	class = b->class;
+	// BUG: we _may_ want to instrument PAUTO sometimes
+	// e.g. if we've got a local variable/method receiver
+	// that has got a pointer inside. Whether it points to
+	// the heap or not is impossible to know at compile time
+	if((class&PHEAP) || class == PPARAMREF || class == PEXTERN
+		|| b->type->etype == TARRAY || b->op == ODOTPTR || b->op == OIND || b->op == OXDOT) {
+		n = treecopy(n);
+		f = mkcall(wr ? "racewrite" : "raceread", T, nil, uintptraddr(n));
+		//typecheck(&f, Etop);
+		*init = list(*init, f);
+	}
+}
+
+static Node*
+uintptraddr(Node *n)
+{
+	Node *r;
+
+	r = nod(OADDR, n, N);
+	r = conv(r, types[TUNSAFEPTR]);
+	r = conv(r, types[TUINTPTR]);
+	return r;
+}
+
+static Node*
+basenod(Node *n)
+{
+	for(;;) {
+		if(n->op == ODOT || n->op == OPAREN) {
+			n = n->left;
+			continue;
+		}
+		if(n->op == OINDEX) {
+			n = n->left;
+			continue;
+		}
+		break;
+	}
+	return n;
+}

src/cmd/gc/reflect.c

@@ -934,6 +934,8 @@ dumptypestructs(void)
 
 		// add paths for runtime and main, which 6l imports implicitly.
 		dimportpath(runtimepkg);
+		if(debug['b'])
+			dimportpath(racepkg);
 		dimportpath(mkpkg(strlit("main")));
 	}
 }

src/cmd/gc/runtime.go

@@ -138,3 +138,9 @@ func int64tofloat64(int64) float64
 func uint64tofloat64(uint64) float64
 
 func complex128div(num complex128, den complex128) (quo complex128)
+
+// race detection
+func racefuncenter()
+func racefuncexit()
+func raceread(uintptr)
+func racewrite(uintptr)