go/ssa: wait for shared functions to finish building

Package.Build and Program.MethodValue can create and build
Functions that may be depended on by other calls to
Package.Build and Program.MethodValue. Both of these
APIs now wait for any Functions that are shared and depended
on to finish building before returning.

This fixes some data races observed when accessing the Blocks
field of shared functions.

Functions that are members of Packages (fn.Pkg != nil)
are still always built by fn.Pkg.Build().

These dependencies can be create cycles between public API
calls that both depend on Functions the other is building.
For example, both the "net/http" and "internal/trace/testtrace"
packages both depend on slices.Contains[[]string,string] and
slices.Index[[]string, string]. Under some schedules, building
both packages may need to wait for the other package to finish
building before finishing.

This introduces a new internal type task. tasks
allows for waiting on the transitive closure of builders iterating
to be marked as done before proceedind.

Also:
- Add a test that reliably reproduces the data race.
- Refactors creator.add calls to happen after the function is created.
- Merges creator into builder.
- Updates the builder to mark when it is done building and wait
  for dependencies.
- Define type unit = struct{} and clean up.

Fixes golang/go#67079

Change-Id: I34d2b8730b19609f5fbf6b55b3db33f59818a17b
Reviewed-on: https://go-review.googlesource.com/c/tools/+/591775
LUCI-TryBot-Result: Go LUCI <[email protected]>
Reviewed-by: Alan Donovan <[email protected]>

Author: timothy-king

go/ssa/builder.go

@@ -127,10 +127,48 @@ var (
 
 // builder holds state associated with the package currently being built.
 // Its methods contain all the logic for AST-to-SSA conversion.
+//
+// All Functions belong to the same Program.
+//
+// builders are not thread-safe.
 type builder struct {
-	// Invariant: 0 <= rtypes <= finished <= created.Len()
-	created  *creator // functions created during building
-	finished int      // Invariant: create[i].built holds for i in [0,finished)
+	fns []*Function // Functions that have finished their CREATE phases.
+
+	finished int // finished is the length of the prefix of fns containing built functions.
+
+	// The task of building shared functions within the builder.
+	// Shared functions are ones the the builder may either create or lookup.
+	// These may be built by other builders in parallel.
+	// The task is done when the builder has finished iterating, and it
+	// waits for all shared functions to finish building.
+	// nil implies there are no hared functions to wait on.
+	buildshared *task
+}
+
+// shared is done when the builder has built all of the
+// enqueued functions to a fixed-point.
+func (b *builder) shared() *task {
+	if b.buildshared == nil { // lazily-initialize
+		b.buildshared = &task{done: make(chan unit)}
+	}
+	return b.buildshared
+}
+
+// enqueue fn to be built by the builder.
+func (b *builder) enqueue(fn *Function) {
+	b.fns = append(b.fns, fn)
+}
+
+// waitForSharedFunction indicates that the builder should wait until
+// the potentially shared function fn has finished building.
+//
+// This should include any functions that may be built by other
+// builders.
+func (b *builder) waitForSharedFunction(fn *Function) {
+	if fn.buildshared != nil { // maybe need to wait?
+		s := b.shared()
+		s.addEdge(fn.buildshared)
+	}
 }
 
 // cond emits to fn code to evaluate boolean condition e and jump
@@ -779,7 +817,7 @@ func (b *builder) expr0(fn *Function, e ast.Expr, tv types.TypeAndValue) Value {
 			callee := v.(*Function) // (func)
 			if callee.typeparams.Len() > 0 {
 				targs := fn.subst.types(instanceArgs(fn.info, e))
-				callee = callee.instance(targs, b.created)
+				callee = callee.instance(targs, b)
 			}
 			return callee
 		}
@@ -800,7 +838,8 @@ func (b *builder) expr0(fn *Function, e ast.Expr, tv types.TypeAndValue) Value {
 		case types.MethodExpr:
 			// (*T).f or T.f, the method f from the method-set of type T.
 			// The result is a "thunk".
-			thunk := createThunk(fn.Prog, sel, b.created)
+			thunk := createThunk(fn.Prog, sel)
+			b.enqueue(thunk)
 			return emitConv(fn, thunk, fn.typ(tv.Type))
 
 		case types.MethodVal:
@@ -842,8 +881,11 @@ func (b *builder) expr0(fn *Function, e ast.Expr, tv types.TypeAndValue) Value {
 				// obj is generic.
 				obj = fn.Prog.canon.instantiateMethod(obj, fn.subst.types(targs), fn.Prog.ctxt)
 			}
+			bound := createBound(fn.Prog, obj)
+			b.enqueue(bound)
+
 			c := &MakeClosure{
-				Fn:       createBound(fn.Prog, obj, b.created),
+				Fn:       bound,
 				Bindings: []Value{v},
 			}
 			c.setPos(e.Sel.Pos())
@@ -976,7 +1018,7 @@ func (b *builder) setCallFunc(fn *Function, e *ast.CallExpr, c *CallCommon) {
 				c.Method = obj
 			} else {
 				// "Call"-mode call.
-				c.Value = fn.Prog.objectMethod(obj, b.created)
+				c.Value = fn.Prog.objectMethod(obj, b)
 				c.Args = append(c.Args, v)
 			}
 			return
@@ -2838,11 +2880,16 @@ type buildFunc = func(*builder, *Function)
 // iterate causes all created but unbuilt functions to be built. As
 // this may create new methods, the process is iterated until it
 // converges.
+//
+// Waits for any dependencies to finish building.
 func (b *builder) iterate() {
-	for ; b.finished < b.created.Len(); b.finished++ {
-		fn := b.created.At(b.finished)
+	for ; b.finished < len(b.fns); b.finished++ {
+		fn := b.fns[b.finished]
 		b.buildFunction(fn)
 	}
+
+	b.buildshared.markDone()
+	b.buildshared.wait()
 }
 
 // buildFunction builds SSA code for the body of function fn.  Idempotent.
@@ -3084,7 +3131,7 @@ func (prog *Program) Build() {
 			p.Build()
 		} else {
 			wg.Add(1)
-			cpuLimit <- struct{}{} // acquire a token
+			cpuLimit <- unit{} // acquire a token
 			go func(p *Package) {
 				p.Build()
 				wg.Done()
@@ -3096,7 +3143,7 @@ func (prog *Program) Build() {
 }
 
 // cpuLimit is a counting semaphore to limit CPU parallelism.
-var cpuLimit = make(chan struct{}, runtime.GOMAXPROCS(0))
+var cpuLimit = make(chan unit, runtime.GOMAXPROCS(0))
 
 // Build builds SSA code for all functions and vars in package p.
 //
@@ -3117,7 +3164,7 @@ func (p *Package) build() {
 		defer logStack("build %s", p)()
 	}
 
-	b := builder{created: &p.created}
+	b := builder{fns: p.created}
 	b.iterate()
 
 	// We no longer need transient information: ASTs or go/types deductions.

go/ssa/builder_test.go

@@ -20,6 +20,7 @@ import (
 	"strings"
 	"testing"
 
+	"golang.org/x/sync/errgroup"
 	"golang.org/x/tools/go/analysis/analysistest"
 	"golang.org/x/tools/go/buildutil"
 	"golang.org/x/tools/go/loader"
@@ -1211,3 +1212,51 @@ func TestFixedBugs(t *testing.T) {
 		})
 	}
 }
+
+func TestIssue67079(t *testing.T) {
+	// This test reproduced a race in the SSA builder nearly 100% of the time.
+
+	// Load the package.
+	const src = `package p; type T int; func (T) f() {}; var _ = (*T).f`
+	conf := loader.Config{Fset: token.NewFileSet()}
+	f, err := parser.ParseFile(conf.Fset, "p.go", src, 0)
+	if err != nil {
+		t.Fatal(err)
+	}
+	conf.CreateFromFiles("p", f)
+	iprog, err := conf.Load()
+	if err != nil {
+		t.Fatal(err)
+	}
+	pkg := iprog.Created[0].Pkg
+
+	// Create and build SSA program.
+	prog := ssautil.CreateProgram(iprog, ssa.BuilderMode(0))
+	prog.Build()
+
+	var g errgroup.Group
+
+	// Access bodies of all functions.
+	g.Go(func() error {
+		for fn := range ssautil.AllFunctions(prog) {
+			for _, b := range fn.Blocks {
+				for _, instr := range b.Instrs {
+					if call, ok := instr.(*ssa.Call); ok {
+						call.Common().StaticCallee() // access call.Value
+					}
+				}
+			}
+		}
+		return nil
+	})
+
+	// Force building of wrappers.
+	g.Go(func() error {
+		ptrT := types.NewPointer(pkg.Scope().Lookup("T").Type())
+		ptrTf := types.NewMethodSet(ptrT).At(0) // (*T).f symbol
+		prog.MethodValue(ptrTf)
+		return nil
+	})
+
+	g.Wait() // ignore error
+}

go/ssa/create.go

@@ -96,8 +96,9 @@ func memberFromObject(pkg *Package, obj types.Object, syntax ast.Node, goversion
 			pkg.ninit++
 			name = fmt.Sprintf("init#%d", pkg.ninit)
 		}
-		fn := createFunction(pkg.Prog, obj, name, syntax, pkg.info, goversion, &pkg.created)
+		fn := createFunction(pkg.Prog, obj, name, syntax, pkg.info, goversion)
 		fn.Pkg = pkg
+		pkg.created = append(pkg.created, fn)
 		pkg.objects[obj] = fn
 		if name != "_" && sig.Recv() == nil {
 			pkg.Members[name] = fn // package-level function
@@ -111,7 +112,7 @@ func memberFromObject(pkg *Package, obj types.Object, syntax ast.Node, goversion
 // createFunction creates a function or method. It supports both
 // CreatePackage (with or without syntax) and the on-demand creation
 // of methods in non-created packages based on their types.Func.
-func createFunction(prog *Program, obj *types.Func, name string, syntax ast.Node, info *types.Info, goversion string, cr *creator) *Function {
+func createFunction(prog *Program, obj *types.Func, name string, syntax ast.Node, info *types.Info, goversion string) *Function {
 	sig := obj.Type().(*types.Signature)
 
 	// Collect type parameters.
@@ -143,7 +144,6 @@ func createFunction(prog *Program, obj *types.Func, name string, syntax ast.Node
 	if tparams.Len() > 0 {
 		fn.generic = new(generic)
 	}
-	cr.Add(fn)
 	return fn
 }
 
@@ -184,19 +184,6 @@ func membersFromDecl(pkg *Package, decl ast.Decl, goversion string) {
 	}
 }
 
-// creator tracks functions that have finished their CREATE phases.
-//
-// All Functions belong to the same Program. May have differing packages.
-//
-// creators are not thread-safe.
-type creator []*Function
-
-func (c *creator) Add(fn *Function) {
-	*c = append(*c, fn)
-}
-func (c *creator) At(i int) *Function { return (*c)[i] }
-func (c *creator) Len() int           { return len(*c) }
-
 // CreatePackage creates and returns an SSA Package from the
 // specified type-checked, error-free file ASTs, and populates its
 // Members mapping.
@@ -238,7 +225,7 @@ func (prog *Program) CreatePackage(pkg *types.Package, files []*ast.File, info *
 		goversion: "", // See Package.build for details.
 	}
 	p.Members[p.init.name] = p.init
-	p.created.Add(p.init)
+	p.created = append(p.created, p.init)
 
 	// Allocate all package members: vars, funcs, consts and types.
 	if len(files) > 0 {

go/ssa/instantiate.go

@@ -23,7 +23,7 @@ type generic struct {
 // Any created instance is added to cr.
 //
 // Acquires fn.generic.instancesMu.
-func (fn *Function) instance(targs []types.Type, cr *creator) *Function {
+func (fn *Function) instance(targs []types.Type, b *builder) *Function {
 	key := fn.Prog.canon.List(targs)
 
 	gen := fn.generic
@@ -32,20 +32,24 @@ func (fn *Function) instance(targs []types.Type, cr *creator) *Function {
 	defer gen.instancesMu.Unlock()
 	inst, ok := gen.instances[key]
 	if !ok {
-		inst = createInstance(fn, targs, cr)
+		inst = createInstance(fn, targs)
+		inst.buildshared = b.shared()
+		b.enqueue(inst)
+
 		if gen.instances == nil {
 			gen.instances = make(map[*typeList]*Function)
 		}
 		gen.instances[key] = inst
+	} else {
+		b.waitForSharedFunction(inst)
 	}
 	return inst
 }
 
 // createInstance returns the instantiation of generic function fn using targs.
-// If the instantiation is created, this is added to cr.
 //
 // Requires fn.generic.instancesMu.
-func createInstance(fn *Function, targs []types.Type, cr *creator) *Function {
+func createInstance(fn *Function, targs []types.Type) *Function {
 	prog := fn.Prog
 
 	// Compute signature.
@@ -89,7 +93,7 @@ func createInstance(fn *Function, targs []types.Type, cr *creator) *Function {
 	}
 
 	/* generic instance or instantiation wrapper */
-	instance := &Function{
+	return &Function{
 		name:           fmt.Sprintf("%s%s", fn.Name(), targs), // may not be unique
 		object:         obj,
 		Signature:      sig,
@@ -106,8 +110,6 @@ func createInstance(fn *Function, targs []types.Type, cr *creator) *Function {
 		typeargs:       targs,
 		subst:          subst,
 	}
-	cr.Add(instance)
-	return instance
 }
 
 // isParameterized reports whether any of the specified types contains

go/ssa/instantiate_test.go

@@ -96,11 +96,11 @@ func LoadPointer(addr *unsafe.Pointer) (val unsafe.Pointer)
 
 		meth := prog.FuncValue(obj)
 
-		var cr creator
+		b := &builder{}
 		intSliceTyp := types.NewSlice(types.Typ[types.Int])
-		instance := meth.instance([]types.Type{intSliceTyp}, &cr)
-		if len(cr) != 1 {
-			t.Errorf("Expected first instance to create a function. got %d created functions", len(cr))
+		instance := meth.instance([]types.Type{intSliceTyp}, b)
+		if len(b.fns) != 1 {
+			t.Errorf("Expected first instance to create a function. got %d created functions", len(b.fns))
 		}
 		if instance.Origin() != meth {
 			t.Errorf("Expected Origin of %s to be %s. got %s", instance, meth, instance.Origin())
@@ -114,13 +114,13 @@ func LoadPointer(addr *unsafe.Pointer) (val unsafe.Pointer)
 		}
 
 		// A second request with an identical type returns the same Function.
-		second := meth.instance([]types.Type{types.NewSlice(types.Typ[types.Int])}, &cr)
-		if second != instance || len(cr) != 1 {
+		second := meth.instance([]types.Type{types.NewSlice(types.Typ[types.Int])}, b)
+		if second != instance || len(b.fns) != 1 {
 			t.Error("Expected second identical instantiation to not create a function")
 		}
 
 		// Add a second instance.
-		inst2 := meth.instance([]types.Type{types.NewSlice(types.Typ[types.Uint])}, &cr)
+		inst2 := meth.instance([]types.Type{types.NewSlice(types.Typ[types.Uint])}, b)
 		instances = allInstances(meth)
 
 		// Note: instance.Name() < inst2.Name()
@@ -134,7 +134,6 @@ func LoadPointer(addr *unsafe.Pointer) (val unsafe.Pointer)
 		// TODO(adonovan): tests should not rely on unexported functions.
 
 		// build and sanity check manually created instance.
-		var b builder
 		b.buildFunction(instance)
 		var buf bytes.Buffer
 		if !sanityCheck(instance, &buf) {