runner.go

package lint

/*
Package loading

Conceptually, package loading in the runner can be imagined as a
graph-shaped work list. We iteratively pop off leaf nodes (packages
that have no unloaded dependencies) and load data from export data,
our cache, or source.

Specifically, non-initial packages are loaded from export data and the
fact cache if possible, otherwise from source. Initial packages are
loaded from export data, the fact cache and the (problems, ignores,
config) cache if possible, otherwise from source.

The appeal of this approach is that it is both simple to implement and
easily parallelizable. Each leaf node can be processed independently,
and new leaf nodes appear as their dependencies are being processed.

The downside of this approach, however, is that we're doing more work
than necessary. Imagine an initial package A, which has the following
dependency chain: A->B->C->D – in the current implementation, we will
load all 4 packages. However, if package A can be loaded fully from
cached information, then none of its dependencies are necessary, and
we could avoid loading them.


Parallelism

Runner implements parallel processing of packages by spawning one
goroutine per package in the dependency graph, without any semaphores.
Each goroutine initially waits on the completion of all of its
dependencies, thus establishing correct order of processing. Once all
dependencies finish processing, the goroutine will load the package
from export data or source – this loading is guarded by a semaphore,
sized according to the number of CPU cores. This way, we only have as
many packages occupying memory and CPU resources as there are actual
cores to process them.

This combination of unbounded goroutines but bounded package loading
means that if we have many parallel, independent subgraphs, they will
all execute in parallel, while not wasting resources for long linear
chains or trying to process more subgraphs in parallel than the system
can handle.


Caching

We make use of several caches. These caches are Go's export data, our
facts cache, and our (problems, ignores, config) cache.

Initial packages will either be loaded from a combination of all three
caches, or from source. Non-initial packages will either be loaded
from a combination of export data and facts cache, or from source.

The facts cache is separate from the (problems, ignores, config) cache
because when we process non-initial packages, we generate facts, but
we discard problems and ignores.

The facts cache is keyed by (package, analyzer), whereas the
(problems, ignores, config) cache is keyed by (package, list of
analyzes). The difference between the two exists because there are
only a handful of analyses that produce facts, but hundreds of
analyses that don't. Creating one cache entry per fact-generating
analysis is feasible, creating one cache entry per normal analysis has
significant performance and storage overheads.

The downside of keying by the list of analyzes is, naturally, that a
change in list of analyzes changes the cache key. `staticcheck -checks
A` and `staticcheck -checks A,B` will therefore need their own cache
entries and not reuse each other's work. This problem does not affect
the facts cache.

*/

import (
	"bytes"
	"encoding/gob"
	"encoding/hex"
	"fmt"
	"go/ast"
	"go/token"
	"go/types"
	"reflect"
	"regexp"
	"runtime"
	"sort"
	"strconv"
	"strings"
	"sync"
	"sync/atomic"
	"time"

	"golang.org/x/tools/go/analysis"
	"golang.org/x/tools/go/packages"
	"golang.org/x/tools/go/types/objectpath"
	"honnef.co/go/tools/config"
	"honnef.co/go/tools/facts"
	"honnef.co/go/tools/internal/cache"
	"honnef.co/go/tools/loader"
)

func init() {
	gob.Register(&FileIgnore{})
	gob.Register(&LineIgnore{})
}

// If enabled, abuse of the go/analysis API will lead to panics
const sanityCheck = true

// OPT(dh): for a dependency tree A->B->C->D, if we have cached data
// for B, there should be no need to load C and D individually. Go's
// export data for B contains all the data we need on types, and our
// fact cache could store the union of B, C and D in B.
//
// This may change unused's behavior, however, as it may observe fewer
// interfaces from transitive dependencies.

// OPT(dh): every single package will have the same value for
// canClearTypes. We could move the Package.decUse method to runner to
// eliminate this field. This is probably not worth it, though. There
// are only thousands of packages, so the field only takes up
// kilobytes of memory.

// OPT(dh): do we really need the Package.gen field? it's based
// trivially on pkg.results and merely caches the result of a type
// assertion. How often do we actually use the field?

type Package struct {
	// dependents is initially set to 1 plus the number of packages
	// that directly import this package. It is atomically decreased
	// by 1 every time a dependent has been processed or when the
	// package itself has been processed. Once the value reaches zero,
	// the package is no longer needed.
	dependents uint64

	*packages.Package
	Imports []*Package
	initial bool
	// fromSource is set to true for packages that have been loaded
	// from source. This is the case for initial packages, packages
	// with missing export data, and packages with no cached facts.
	fromSource bool
	// hash stores the package hash, as computed by packageHash
	hash     string
	actionID cache.ActionID
	done     chan struct{}

	resultsMu sync.Mutex
	// results maps analyzer IDs to analyzer results. it is
	// implemented as a deduplicating concurrent cache.
	results []*result

	cfg *config.Config
	// gen maps file names to the code generator that created them
	gen      map[string]facts.Generator
	problems []Problem
	ignores  []Ignore
	errs     []error

	// these slices are indexed by analysis
	facts    []map[types.Object][]analysis.Fact
	pkgFacts [][]analysis.Fact

	// canClearTypes is set to true if we can discard type
	// information after the package and its dependents have been
	// processed. This is the case when no cumulative checkers are
	// being run.
	canClearTypes bool
}

type cachedPackage struct {
	Problems []Problem
	Ignores  []Ignore
	Config   *config.Config
}

func (pkg *Package) decUse() {
	ret := atomic.AddUint64(&pkg.dependents, ^uint64(0))
	if ret == 0 {
		// nobody depends on this package anymore
		if pkg.canClearTypes {
			pkg.Types = nil
		}
		pkg.facts = nil
		pkg.pkgFacts = nil

		for _, imp := range pkg.Imports {
			imp.decUse()
		}
	}
}

type result struct {
	v     interface{}
	err   error
	ready chan struct{}
}

type Runner struct {
	cache           *cache.Cache
	goVersion       int
	stats           *Stats
	repeatAnalyzers uint

	analyzerIDs      analyzerIDs
	problemsCacheKey string

	// limits parallelism of loading packages
	loadSem chan struct{}
}

type analyzerIDs struct {
	m map[*analysis.Analyzer]int
}

func (ids analyzerIDs) get(a *analysis.Analyzer) int {
	id, ok := ids.m[a]
	if !ok {
		panic(fmt.Sprintf("no analyzer ID for %s", a.Name))
	}
	return id
}

type Fact struct {
	Path string
	Fact analysis.Fact
}

type analysisAction struct {
	analyzer        *analysis.Analyzer
	analyzerID      int
	pkg             *Package
	newPackageFacts []analysis.Fact
	problems        []Problem

	pkgFacts map[*types.Package][]analysis.Fact
}

func (ac *analysisAction) String() string {
	return fmt.Sprintf("%s @ %s", ac.analyzer, ac.pkg)
}

func (ac *analysisAction) allObjectFacts() []analysis.ObjectFact {
	out := make([]analysis.ObjectFact, 0, len(ac.pkg.facts[ac.analyzerID]))
	for obj, facts := range ac.pkg.facts[ac.analyzerID] {
		for _, fact := range facts {
			out = append(out, analysis.ObjectFact{
				Object: obj,
				Fact:   fact,
			})
		}
	}
	return out
}

func (ac *analysisAction) allPackageFacts() []analysis.PackageFact {
	out := make([]analysis.PackageFact, 0, len(ac.pkgFacts))
	for pkg, facts := range ac.pkgFacts {
		for _, fact := range facts {
			out = append(out, analysis.PackageFact{
				Package: pkg,
				Fact:    fact,
			})
		}
	}
	return out
}

func (ac *analysisAction) importObjectFact(obj types.Object, fact analysis.Fact) bool {
	if sanityCheck && len(ac.analyzer.FactTypes) == 0 {
		panic("analysis doesn't export any facts")
	}
	for _, f := range ac.pkg.facts[ac.analyzerID][obj] {
		if reflect.TypeOf(f) == reflect.TypeOf(fact) {
			reflect.ValueOf(fact).Elem().Set(reflect.ValueOf(f).Elem())
			return true
		}
	}
	return false
}

func (ac *analysisAction) importPackageFact(pkg *types.Package, fact analysis.Fact) bool {
	if sanityCheck && len(ac.analyzer.FactTypes) == 0 {
		panic("analysis doesn't export any facts")
	}
	for _, f := range ac.pkgFacts[pkg] {
		if reflect.TypeOf(f) == reflect.TypeOf(fact) {
			reflect.ValueOf(fact).Elem().Set(reflect.ValueOf(f).Elem())
			return true
		}
	}
	return false
}

func (ac *analysisAction) exportObjectFact(obj types.Object, fact analysis.Fact) {
	if sanityCheck && len(ac.analyzer.FactTypes) == 0 {
		panic("analysis doesn't export any facts")
	}
	ac.pkg.facts[ac.analyzerID][obj] = append(ac.pkg.facts[ac.analyzerID][obj], fact)
}

func (ac *analysisAction) exportPackageFact(fact analysis.Fact) {
	if sanityCheck && len(ac.analyzer.FactTypes) == 0 {
		panic("analysis doesn't export any facts")
	}
	ac.pkgFacts[ac.pkg.Types] = append(ac.pkgFacts[ac.pkg.Types], fact)
	ac.newPackageFacts = append(ac.newPackageFacts, fact)
}

func (ac *analysisAction) report(pass *analysis.Pass, d analysis.Diagnostic) {
	p := Problem{
		Pos:     DisplayPosition(pass.Fset, d.Pos),
		End:     DisplayPosition(pass.Fset, d.End),
		Message: d.Message,
		Check:   pass.Analyzer.Name,
	}
	for _, r := range d.Related {
		p.Related = append(p.Related, Related{
			Pos:     DisplayPosition(pass.Fset, r.Pos),
			End:     DisplayPosition(pass.Fset, r.End),
			Message: r.Message,
		})
	}
	ac.problems = append(ac.problems, p)
}

func (r *Runner) runAnalysis(ac *analysisAction) (ret interface{}, err error) {
	ac.pkg.resultsMu.Lock()
	res := ac.pkg.results[r.analyzerIDs.get(ac.analyzer)]
	if res != nil {
		ac.pkg.resultsMu.Unlock()
		<-res.ready
		return res.v, res.err
	} else {
		res = &result{
			ready: make(chan struct{}),
		}
		ac.pkg.results[r.analyzerIDs.get(ac.analyzer)] = res
		ac.pkg.resultsMu.Unlock()

		defer func() {
			res.v = ret
			res.err = err
			close(res.ready)
		}()

		pass := new(analysis.Pass)
		*pass = analysis.Pass{
			Analyzer: ac.analyzer,
			Fset:     ac.pkg.Fset,
			Files:    ac.pkg.Syntax,
			// type information may be nil or may be populated. if it is
			// nil, it will get populated later.
			Pkg:               ac.pkg.Types,
			TypesInfo:         ac.pkg.TypesInfo,
			TypesSizes:        ac.pkg.TypesSizes,
			ResultOf:          map[*analysis.Analyzer]interface{}{},
			ImportObjectFact:  ac.importObjectFact,
			ImportPackageFact: ac.importPackageFact,
			ExportObjectFact:  ac.exportObjectFact,
			ExportPackageFact: ac.exportPackageFact,
			Report: func(d analysis.Diagnostic) {
				ac.report(pass, d)
			},
			AllObjectFacts:  ac.allObjectFacts,
			AllPackageFacts: ac.allPackageFacts,
		}

		if !ac.pkg.initial {
			// Don't report problems in dependencies
			pass.Report = func(analysis.Diagnostic) {}
		}
		return r.runAnalysisUser(pass, ac)
	}
}

func (r *Runner) loadCachedPackage(pkg *Package, analyzers []*analysis.Analyzer) (cachedPackage, bool) {
	// OPT(dh): we can cache this computation, it'll be the same for all packages
	id := cache.Subkey(pkg.actionID, "data "+r.problemsCacheKey)

	b, _, err := r.cache.GetBytes(id)
	if err != nil {
		return cachedPackage{}, false
	}
	var cpkg cachedPackage
	if err := gob.NewDecoder(bytes.NewReader(b)).Decode(&cpkg); err != nil {
		return cachedPackage{}, false
	}
	return cpkg, true
}

func (r *Runner) loadCachedFacts(a *analysis.Analyzer, pkg *Package) ([]Fact, bool) {
	if len(a.FactTypes) == 0 {
		return nil, true
	}

	var facts []Fact
	// Look in the cache for facts
	aID := passActionID(pkg, a)
	aID = cache.Subkey(aID, "facts")
	b, _, err := r.cache.GetBytes(aID)
	if err != nil {
		// No cached facts, analyse this package like a user-provided one, but ignore diagnostics
		return nil, false
	}

	if err := gob.NewDecoder(bytes.NewReader(b)).Decode(&facts); err != nil {
		// Cached facts are broken, analyse this package like a user-provided one, but ignore diagnostics
		return nil, false
	}
	return facts, true
}

type dependencyError struct {
	dep string
	err error
}

func (err dependencyError) nested() dependencyError {
	if o, ok := err.err.(dependencyError); ok {
		return o.nested()
	}
	return err
}

func (err dependencyError) Error() string {
	if o, ok := err.err.(dependencyError); ok {
		return o.Error()
	}
	return fmt.Sprintf("error running dependency %s: %s", err.dep, err.err)
}

func (r *Runner) makeAnalysisAction(a *analysis.Analyzer, pkg *Package) *analysisAction {
	aid := r.analyzerIDs.get(a)
	ac := &analysisAction{
		analyzer:   a,
		analyzerID: aid,
		pkg:        pkg,
	}

	if len(a.FactTypes) == 0 {
		return ac
	}

	// Merge all package facts of dependencies
	ac.pkgFacts = map[*types.Package][]analysis.Fact{}
	seen := map[*Package]struct{}{}
	var dfs func(*Package)
	dfs = func(pkg *Package) {
		if _, ok := seen[pkg]; ok {
			return
		}
		seen[pkg] = struct{}{}
		s := pkg.pkgFacts[aid]
		ac.pkgFacts[pkg.Types] = s[0:len(s):len(s)]
		for _, imp := range pkg.Imports {
			dfs(imp)
		}
	}
	dfs(pkg)

	return ac
}

// analyzes that we always want to run, even if they're not being run
// explicitly or as dependencies. these are necessary for the inner
// workings of the runner.
var injectedAnalyses = []*analysis.Analyzer{facts.Generated, config.Analyzer}

func (r *Runner) runAnalysisUser(pass *analysis.Pass, ac *analysisAction) (interface{}, error) {
	if !ac.pkg.fromSource {
		panic(fmt.Sprintf("internal error: %s was not loaded from source", ac.pkg))
	}

	// User-provided package, analyse it
	// First analyze it with dependencies
	for _, req := range ac.analyzer.Requires {
		acReq := r.makeAnalysisAction(req, ac.pkg)
		ret, err := r.runAnalysis(acReq)
		if err != nil {
			// We couldn't run a dependency, no point in going on
			return nil, dependencyError{req.Name, err}
		}

		pass.ResultOf[req] = ret
	}

	// Then with this analyzer
	var ret interface{}
	for i := uint(0); i < r.repeatAnalyzers+1; i++ {
		var err error
		t := time.Now()
		ret, err = ac.analyzer.Run(pass)
		r.stats.MeasureAnalyzer(ac.analyzer, ac.pkg, time.Since(t))
		if err != nil {
			return nil, err
		}
	}

	if len(ac.analyzer.FactTypes) > 0 {
		// Merge new facts into the package and persist them.
		var facts []Fact
		for _, fact := range ac.newPackageFacts {
			id := r.analyzerIDs.get(ac.analyzer)
			ac.pkg.pkgFacts[id] = append(ac.pkg.pkgFacts[id], fact)
			facts = append(facts, Fact{"", fact})
		}
		for obj, afacts := range ac.pkg.facts[ac.analyzerID] {
			if obj.Pkg() != ac.pkg.Package.Types {
				continue
			}
			path, err := objectpath.For(obj)
			if err != nil {
				continue
			}
			for _, fact := range afacts {
				facts = append(facts, Fact{string(path), fact})
			}
		}

		if err := r.cacheData(facts, ac.pkg, ac.analyzer, "facts"); err != nil {
			return nil, err
		}
	}

	return ret, nil
}

func (r *Runner) cacheData(v interface{}, pkg *Package, a *analysis.Analyzer, subkey string) error {
	buf := &bytes.Buffer{}
	if err := gob.NewEncoder(buf).Encode(v); err != nil {
		return err
	}
	aID := passActionID(pkg, a)
	aID = cache.Subkey(aID, subkey)
	if err := r.cache.PutBytes(aID, buf.Bytes()); err != nil {
		return err
	}
	return nil
}

func NewRunner(stats *Stats) (*Runner, error) {
	cache, err := cache.Default()
	if err != nil {
		return nil, err
	}

	return &Runner{
		cache: cache,
		stats: stats,
	}, nil
}

// Run loads packages corresponding to patterns and analyses them with
// analyzers. It returns the loaded packages, which contain reported
// diagnostics as well as extracted ignore directives.
//
// Note that diagnostics have not been filtered at this point yet, to
// accommodate cumulative analyzes that require additional steps to
// produce diagnostics.
func (r *Runner) Run(cfg *packages.Config, patterns []string, analyzers []*analysis.Analyzer, hasCumulative bool) ([]*Package, error) {
	checkerNames := make([]string, len(analyzers))
	for i, a := range analyzers {
		checkerNames[i] = a.Name
	}
	sort.Strings(checkerNames)
	r.problemsCacheKey = strings.Join(checkerNames, " ")

	var allAnalyzers []*analysis.Analyzer
	r.analyzerIDs = analyzerIDs{m: map[*analysis.Analyzer]int{}}
	id := 0
	seen := map[*analysis.Analyzer]struct{}{}
	var dfs func(a *analysis.Analyzer)
	dfs = func(a *analysis.Analyzer) {
		if _, ok := seen[a]; ok {
			return
		}
		seen[a] = struct{}{}
		allAnalyzers = append(allAnalyzers, a)
		r.analyzerIDs.m[a] = id
		id++
		for _, f := range a.FactTypes {
			gob.Register(f)
		}
		for _, req := range a.Requires {
			dfs(req)
		}
	}
	for _, a := range analyzers {
		if v := a.Flags.Lookup("go"); v != nil {
			v.Value.Set(fmt.Sprintf("1.%d", r.goVersion))
		}
		dfs(a)
	}
	for _, a := range injectedAnalyses {
		dfs(a)
	}
	// Run all analyzers on all packages (subject to further
	// restrictions enforced later). This guarantees that if analyzer
	// A1 depends on A2, and A2 has facts, that A2 will run on the
	// dependencies of user-provided packages, even though A1 won't.
	analyzers = allAnalyzers

	var dcfg packages.Config
	if cfg != nil {
		dcfg = *cfg
	}

	atomic.StoreUint32(&r.stats.State, StateGraph)
	initialPkgs, err := loader.Graph(dcfg, patterns...)
	if err != nil {
		return nil, err
	}
	defer r.cache.Trim()

	var allPkgs []*Package
	m := map[*packages.Package]*Package{}
	packages.Visit(initialPkgs, nil, func(l *packages.Package) {
		m[l] = &Package{
			Package:  l,
			results:  make([]*result, len(r.analyzerIDs.m)),
			facts:    make([]map[types.Object][]analysis.Fact, len(r.analyzerIDs.m)),
			pkgFacts: make([][]analysis.Fact, len(r.analyzerIDs.m)),
			done:     make(chan struct{}),
			// every package needs itself
			dependents:    1,
			canClearTypes: !hasCumulative,
		}
		allPkgs = append(allPkgs, m[l])
		for i := range m[l].facts {
			m[l].facts[i] = map[types.Object][]analysis.Fact{}
		}
		for _, err := range l.Errors {
			m[l].errs = append(m[l].errs, err)
		}
		for _, v := range l.Imports {
			m[v].dependents++
			m[l].Imports = append(m[l].Imports, m[v])
		}

		m[l].hash, err = r.packageHash(m[l])
		m[l].actionID = packageActionID(m[l])
		if err != nil {
			m[l].errs = append(m[l].errs, err)
		}
	})

	pkgs := make([]*Package, len(initialPkgs))
	for i, l := range initialPkgs {
		pkgs[i] = m[l]
		pkgs[i].initial = true
	}

	atomic.StoreUint32(&r.stats.InitialPackages, uint32(len(initialPkgs)))
	atomic.StoreUint32(&r.stats.TotalPackages, uint32(len(allPkgs)))
	atomic.StoreUint32(&r.stats.State, StateProcessing)

	var wg sync.WaitGroup
	wg.Add(len(allPkgs))
	r.loadSem = make(chan struct{}, runtime.GOMAXPROCS(-1))
	atomic.StoreUint32(&r.stats.TotalWorkers, uint32(cap(r.loadSem)))
	for _, pkg := range allPkgs {
		pkg := pkg
		go func() {
			r.processPkg(pkg, analyzers)

			if pkg.initial {
				atomic.AddUint32(&r.stats.ProcessedInitialPackages, 1)
			}
			atomic.AddUint32(&r.stats.Problems, uint32(len(pkg.problems)))
			wg.Done()
		}()
	}
	wg.Wait()

	return pkgs, nil
}

var posRe = regexp.MustCompile(`^(.+?):(\d+)(?::(\d+)?)?`)

func parsePos(pos string) (token.Position, int, error) {
	if pos == "-" || pos == "" {
		return token.Position{}, 0, nil
	}
	parts := posRe.FindStringSubmatch(pos)
	if parts == nil {
		return token.Position{}, 0, fmt.Errorf("malformed position %q", pos)
	}
	file := parts[1]
	line, _ := strconv.Atoi(parts[2])
	col, _ := strconv.Atoi(parts[3])
	return token.Position{
		Filename: file,
		Line:     line,
		Column:   col,
	}, len(parts[0]), nil
}

// loadPkg loads a Go package. It may be loaded from a combination of
// caches, or from source.
func (r *Runner) loadPkg(pkg *Package, analyzers []*analysis.Analyzer) error {
	if pkg.Types != nil {
		panic(fmt.Sprintf("internal error: %s has already been loaded", pkg.Package))
	}

	if pkg.initial {
		// Try to load cached package
		cpkg, ok := r.loadCachedPackage(pkg, analyzers)
		if ok {
			pkg.problems = cpkg.Problems
			pkg.ignores = cpkg.Ignores
			pkg.cfg = cpkg.Config
		} else {
			pkg.fromSource = true
			return loader.LoadFromSource(pkg.Package)
		}
	}

	// At this point we're either working with a non-initial package,
	// or we managed to load cached problems for the package. We still
	// need export data and facts.

	// OPT(dh): we don't need type information for this package if no
	// other package depends on it. this may be the case for initial
	// packages.

	// Load package from export data
	if err := loader.LoadFromExport(pkg.Package); err != nil {
		// We asked Go to give us up to date export data, yet
		// we can't load it. There must be something wrong.
		//
		// Attempt loading from source. This should fail (because
		// otherwise there would be export data); we just want to
		// get the compile errors. If loading from source succeeds
		// we discard the result, anyway. Otherwise we'll fail
		// when trying to reload from export data later.
		//
		// FIXME(dh): we no longer reload from export data, so
		// theoretically we should be able to continue
		pkg.fromSource = true
		if err := loader.LoadFromSource(pkg.Package); err != nil {
			return err
		}
		// Make sure this package can't be imported successfully
		pkg.Package.Errors = append(pkg.Package.Errors, packages.Error{
			Pos:  "-",
			Msg:  fmt.Sprintf("could not load export data: %s", err),
			Kind: packages.ParseError,
		})
		return fmt.Errorf("could not load export data: %s", err)
	}

	failed := false
	seen := make([]bool, len(r.analyzerIDs.m))
	var dfs func(*analysis.Analyzer)
	dfs = func(a *analysis.Analyzer) {
		if seen[r.analyzerIDs.get(a)] {
			return
		}
		seen[r.analyzerIDs.get(a)] = true

		if len(a.FactTypes) > 0 {
			facts, ok := r.loadCachedFacts(a, pkg)
			if !ok {
				failed = true
				return
			}

			for _, f := range facts {
				if f.Path == "" {
					// This is a package fact
					pkg.pkgFacts[r.analyzerIDs.get(a)] = append(pkg.pkgFacts[r.analyzerIDs.get(a)], f.Fact)
					continue
				}
				obj, err := objectpath.Object(pkg.Types, objectpath.Path(f.Path))
				if err != nil {
					// Be lenient about these errors. For example, when
					// analysing io/ioutil from source, we may get a fact
					// for methods on the devNull type, and objectpath
					// will happily create a path for them. However, when
					// we later load io/ioutil from export data, the path
					// no longer resolves.
					//
					// If an exported type embeds the unexported type,
					// then (part of) the unexported type will become part
					// of the type information and our path will resolve
					// again.
					continue
				}
				pkg.facts[r.analyzerIDs.get(a)][obj] = append(pkg.facts[r.analyzerIDs.get(a)][obj], f.Fact)
			}
		}

		for _, req := range a.Requires {
			dfs(req)
		}
	}
	for _, a := range analyzers {
		dfs(a)
	}

	if !failed {
		return nil
	}

	// We failed to load some cached facts
	pkg.fromSource = true
	// XXX we added facts to the maps, we need to get rid of those
	return loader.LoadFromSource(pkg.Package)
}

type analysisError struct {
	analyzer *analysis.Analyzer
	pkg      *Package
	err      error
}

func (err analysisError) Error() string {
	return fmt.Sprintf("error running analyzer %s on %s: %s", err.analyzer, err.pkg, err.err)
}

// processPkg processes a package. This involves loading the package,
// either from export data or from source. For packages loaded from
// source, the provides analyzers will be run on the package.
func (r *Runner) processPkg(pkg *Package, analyzers []*analysis.Analyzer) {
	defer func() {
		// Clear information we no longer need. Make sure to do this
		// when returning from processPkg so that we clear
		// dependencies, not just initial packages.
		pkg.TypesInfo = nil
		pkg.Syntax = nil
		pkg.results = nil

		atomic.AddUint32(&r.stats.ProcessedPackages, 1)
		pkg.decUse()
		close(pkg.done)
	}()

	// Ensure all packages have the generated map and config. This is
	// required by internals of the runner. Analyses that themselves
	// make use of either have an explicit dependency so that other
	// runners work correctly, too.
	analyzers = append(analyzers[0:len(analyzers):len(analyzers)], injectedAnalyses...)

	if len(pkg.errs) != 0 {
		return
	}

	for _, imp := range pkg.Imports {
		<-imp.done
		if len(imp.errs) > 0 {
			if imp.initial {
				// Don't print the error of the dependency since it's
				// an initial package and we're already printing the
				// error.
				pkg.errs = append(pkg.errs, fmt.Errorf("could not analyze dependency %s of %s", imp, pkg))
			} else {
				var s string
				for _, err := range imp.errs {
					s += "\n\t" + err.Error()
				}
				pkg.errs = append(pkg.errs, fmt.Errorf("could not analyze dependency %s of %s: %s", imp, pkg, s))
			}
			return
		}
	}
	if pkg.PkgPath == "unsafe" {
		pkg.Types = types.Unsafe
		return
	}

	r.loadSem <- struct{}{}
	atomic.AddUint32(&r.stats.ActiveWorkers, 1)
	defer func() {
		<-r.loadSem
		atomic.AddUint32(&r.stats.ActiveWorkers, ^uint32(0))
	}()
	if err := r.loadPkg(pkg, analyzers); err != nil {
		pkg.errs = append(pkg.errs, err)
		return
	}

	// A package's object facts is the union of all of its dependencies.
	for _, imp := range pkg.Imports {
		for ai, m := range imp.facts {
			for obj, facts := range m {
				pkg.facts[ai][obj] = facts[0:len(facts):len(facts)]
			}
		}
	}

	if !pkg.fromSource {
		// Nothing left to do for the package.
		return
	}

	// Run analyses on initial packages and those missing facts
	var wg sync.WaitGroup
	wg.Add(len(analyzers))
	errs := make([]error, len(analyzers))
	var acs []*analysisAction
	for i, a := range analyzers {
		i := i
		a := a
		ac := r.makeAnalysisAction(a, pkg)
		acs = append(acs, ac)
		go func() {
			defer wg.Done()
			// Only initial packages and packages with missing
			// facts will have been loaded from source.
			if pkg.initial || len(a.FactTypes) > 0 {
				if _, err := r.runAnalysis(ac); err != nil {
					errs[i] = analysisError{a, pkg, err}
					return
				}
			}
		}()
	}
	wg.Wait()

	depErrors := map[dependencyError]int{}
	for _, err := range errs {
		if err == nil {
			continue
		}
		switch err := err.(type) {
		case analysisError:
			switch err := err.err.(type) {
			case dependencyError:
				depErrors[err.nested()]++
			default:
				pkg.errs = append(pkg.errs, err)
			}
		default:
			pkg.errs = append(pkg.errs, err)
		}
	}
	for err, count := range depErrors {
		pkg.errs = append(pkg.errs,
			fmt.Errorf("could not run %s@%s, preventing %d analyzers from running: %s", err.dep, pkg, count, err.err))
	}

	// We can't process ignores at this point because `unused` needs
	// to see more than one package to make its decision.
	//
	// OPT(dh): can't we guard this block of code by pkg.initial?
	ignores, problems := parseDirectives(pkg.Package)
	pkg.ignores = append(pkg.ignores, ignores...)
	pkg.problems = append(pkg.problems, problems...)
	for _, ac := range acs {
		pkg.problems = append(pkg.problems, ac.problems...)
	}

	if pkg.initial {
		// Only initial packages have these analyzers run, and only
		// initial packages need these.
		if pkg.results[r.analyzerIDs.get(config.Analyzer)].v != nil {
			pkg.cfg = pkg.results[r.analyzerIDs.get(config.Analyzer)].v.(*config.Config)
		}
		pkg.gen = pkg.results[r.analyzerIDs.get(facts.Generated)].v.(map[string]facts.Generator)
	}

	// In a previous version of the code, we would throw away all type
	// information and reload it from export data. That was
	// nonsensical. The *types.Package doesn't keep any information
	// live that export data wouldn't also. We only need to discard
	// the AST and the TypesInfo maps; that happens after we return
	// from processPkg.
}

func parseDirective(s string) (cmd string, args []string) {
	if !strings.HasPrefix(s, "//lint:") {
		return "", nil
	}
	s = strings.TrimPrefix(s, "//lint:")
	fields := strings.Split(s, " ")
	return fields[0], fields[1:]
}

// parseDirectives extracts all linter directives from the source
// files of the package. Malformed directives are returned as problems.
func parseDirectives(pkg *packages.Package) ([]Ignore, []Problem) {
	var ignores []Ignore
	var problems []Problem

	for _, f := range pkg.Syntax {
		found := false
	commentLoop:
		for _, cg := range f.Comments {
			for _, c := range cg.List {
				if strings.Contains(c.Text, "//lint:") {
					found = true
					break commentLoop
				}
			}
		}
		if !found {
			continue
		}
		cm := ast.NewCommentMap(pkg.Fset, f, f.Comments)
		for node, cgs := range cm {
			for _, cg := range cgs {
				for _, c := range cg.List {
					if !strings.HasPrefix(c.Text, "//lint:") {
						continue
					}
					cmd, args := parseDirective(c.Text)
					switch cmd {
					case "ignore", "file-ignore":
						if len(args) < 2 {
							p := Problem{
								Pos:      DisplayPosition(pkg.Fset, c.Pos()),
								Message:  "malformed linter directive; missing the required reason field?",
								Severity: Error,
								Check:    "compile",
							}
							problems = append(problems, p)
							continue
						}
					default:
						// unknown directive, ignore
						continue
					}
					checks := strings.Split(args[0], ",")
					pos := DisplayPosition(pkg.Fset, node.Pos())
					var ig Ignore
					switch cmd {
					case "ignore":
						ig = &LineIgnore{
							File:   pos.Filename,
							Line:   pos.Line,
							Checks: checks,
							Pos:    DisplayPosition(pkg.Fset, c.Pos()),
						}
					case "file-ignore":
						ig = &FileIgnore{
							File:   pos.Filename,
							Checks: checks,
						}
					}
					ignores = append(ignores, ig)
				}
			}
		}
	}

	return ignores, problems
}

// packageHash computes a package's hash. The hash is based on all Go
// files that make up the package, as well as the hashes of imported
// packages.
func (r *Runner) packageHash(pkg *Package) (string, error) {
	key := cache.NewHash("package hash")
	fmt.Fprintf(key, "pkgpath %s\n", pkg.PkgPath)
	fmt.Fprintf(key, "go %d\n", r.goVersion)
	for _, f := range pkg.CompiledGoFiles {
		h, err := cache.FileHash(f)
		if err != nil {
			return "", err
		}
		fmt.Fprintf(key, "file %s %x\n", f, h)
	}

	// Actually load the configuration to calculate its hash. This
	// will take into consideration inheritance of configuration
	// files, as well as the default configuration.
	//
	// OPT(dh): doing this means we'll load the config twice: once for
	// computing the hash, and once when analyzing the package from
	// source.
	cdir := config.Dir(pkg.GoFiles)
	if cdir == "" {
		fmt.Fprintf(key, "file %s %x\n", config.ConfigName, [cache.HashSize]byte{})
	} else {
		cfg, err := config.Load(cdir)
		if err != nil {
			return "", err
		}
		h := cache.NewHash(config.ConfigName)
		if _, err := h.Write([]byte(cfg.String())); err != nil {
			return "", err
		}
		fmt.Fprintf(key, "file %s %x\n", config.ConfigName, h.Sum())
	}

	imps := make([]*Package, len(pkg.Imports))
	copy(imps, pkg.Imports)
	sort.Slice(imps, func(i, j int) bool {
		return imps[i].PkgPath < imps[j].PkgPath
	})
	for _, dep := range imps {
		if dep.PkgPath == "unsafe" {
			continue
		}

		fmt.Fprintf(key, "import %s %s\n", dep.PkgPath, dep.hash)
	}
	h := key.Sum()
	return hex.EncodeToString(h[:]), nil
}

func packageActionID(pkg *Package) cache.ActionID {
	key := cache.NewHash("package ID")
	fmt.Fprintf(key, "pkgpath %s\n", pkg.PkgPath)
	fmt.Fprintf(key, "pkghash %s\n", pkg.hash)
	return key.Sum()
}

// passActionID computes an ActionID for an analysis pass.
func passActionID(pkg *Package, analyzer *analysis.Analyzer) cache.ActionID {
	return cache.Subkey(pkg.actionID, fmt.Sprintf("analyzer %s", analyzer.Name))
}