1 // Copyright 2022 The Go Authors. All rights reserved.
2 // Use of this source code is governed by a BSD-style
3 // license that can be found in the LICENSE file.
7 // A note on line numbers: when working with line numbers, we always use the
8 // binary-visible relative line number. i.e., the line number as adjusted by
9 // //line directives (ctxt.InnermostPos(ir.Node.Pos()).RelLine()). Use
10 // NodeLineOffset to compute line offsets.
12 // If you are thinking, "wait, doesn't that just make things more complex than
13 // using the real line number?", then you are 100% correct. Unfortunately,
14 // pprof profiles generated by the runtime always contain line numbers as
15 // adjusted by //line directives (because that is what we put in pclntab). Thus
16 // for the best behavior when attempting to match the source with the profile
17 // it makes sense to use the same line number space.
19 // Some of the effects of this to keep in mind:
21 // - For files without //line directives there is no impact, as RelLine() ==
23 // - For functions entirely covered by the same //line directive (i.e., a
24 // directive before the function definition and no directives within the
25 // function), there should also be no impact, as line offsets within the
26 // function should be the same as the real line offsets.
27 // - Functions containing //line directives may be impacted. As fake line
28 // numbers need not be monotonic, we may compute negative line offsets. We
29 // should accept these and attempt to use them for best-effort matching, as
30 // these offsets should still match if the source is unchanged, and may
31 // continue to match with changed source depending on the impact of the
32 // changes on fake line numbers.
33 // - Functions containing //line directives may also contain duplicate lines,
34 // making it ambiguous which call the profile is referencing. This is a
35 // similar problem to multiple calls on a single real line, as we don't
36 // currently track column numbers.
38 // Long term it would be best to extend pprof profiles to include real line
39 // numbers. Until then, we have to live with these complexities. Luckily,
40 // //line directives that change line numbers in strange ways should be rare,
41 // and failing PGO matching on these files is not too big of a loss.
46 "cmd/compile/internal/base"
47 "cmd/compile/internal/ir"
48 "cmd/compile/internal/typecheck"
49 "cmd/compile/internal/types"
55 // IRGraph is the key data structure that is built from profile. It is
56 // essentially a call graph with nodes pointing to IRs of functions and edges
57 // carrying weights and callsite information. The graph is bidirectional that
58 // helps in removing nodes efficiently.
61 IRNodes map[string]*IRNode
66 // IRNode represents a node in the IRGraph.
68 // Pointer to the IR of the Function represented by this node.
70 // Flat weight of the IRNode, obtained from profile.
72 // Cumulative weight of the IRNode.
76 // IREdgeMap maps an IRNode to its successors.
77 type IREdgeMap map[*IRNode][]*IREdge
79 // IREdge represents a call edge in the IRGraph with source, destination,
80 // weight, callsite, and line number information.
82 // Source and destination of the edge in IRNode.
85 CallSiteOffset int // Line offset from function start line.
88 // NodeMapKey represents a hash key to identify unique call-edges in profile
89 // and in IR. Used for deduplication of call edges found in profile.
90 type NodeMapKey struct {
93 CallSiteOffset int // Line offset from function start line.
96 // Weights capture both node weight and edge weight.
103 // CallSiteInfo captures call-site information and its caller/callee.
104 type CallSiteInfo struct {
105 LineOffset int // Line offset from function start line.
110 // Profile contains the processed PGO profile and weighted call graph used for
111 // PGO optimizations.
112 type Profile struct {
113 // Aggregated NodeWeights and EdgeWeights across the profile. This
114 // helps us determine the percentage threshold for hot/cold
116 TotalNodeWeight int64
117 TotalEdgeWeight int64
119 // NodeMap contains all unique call-edges in the profile and their
120 // aggregated weight.
121 NodeMap map[NodeMapKey]*Weights
123 // WeightedCG represents the IRGraph built from profile, which we will
124 // update as part of inlining.
128 // New generates a profile-graph from the profile.
129 func New(profileFile string) (*Profile, error) {
130 f, err := os.Open(profileFile)
132 return nil, fmt.Errorf("error opening profile: %w", err)
135 profile, err := profile.Parse(f)
137 return nil, fmt.Errorf("error parsing profile: %w", err)
140 if len(profile.Sample) == 0 {
141 // We accept empty profiles, but there is nothing to do.
146 for i, s := range profile.SampleType {
147 // Samples count is the raw data collected, and CPU nanoseconds is just
148 // a scaled version of it, so either one we can find is fine.
149 if (s.Type == "samples" && s.Unit == "count") ||
150 (s.Type == "cpu" && s.Unit == "nanoseconds") {
156 if valueIndex == -1 {
157 return nil, fmt.Errorf(`profile does not contain a sample index with value/type "samples/count" or cpu/nanoseconds"`)
160 g := newGraph(profile, &Options{
162 SampleValue: func(v []int64) int64 { return v[valueIndex] },
166 NodeMap: make(map[NodeMapKey]*Weights),
167 WeightedCG: &IRGraph{
168 IRNodes: make(map[string]*IRNode),
172 // Build the node map and totals from the profile graph.
173 if err := p.processprofileGraph(g); err != nil {
177 if p.TotalNodeWeight == 0 || p.TotalEdgeWeight == 0 {
178 return nil, nil // accept but ignore profile with no samples.
181 // Create package-level call graph with weights from profile and IR.
182 p.initializeIRGraph()
187 // processprofileGraph builds various maps from the profile-graph.
189 // It initializes NodeMap and Total{Node,Edge}Weight based on the name and
190 // callsite to compute node and edge weights which will be used later on to
191 // create edges for WeightedCG.
193 // Caller should ignore the profile if p.TotalNodeWeight == 0 || p.TotalEdgeWeight == 0.
194 func (p *Profile) processprofileGraph(g *Graph) error {
195 nFlat := make(map[string]int64)
196 nCum := make(map[string]int64)
197 seenStartLine := false
199 // Accummulate weights for the same node.
200 for _, n := range g.Nodes {
201 canonicalName := n.Info.Name
202 nFlat[canonicalName] += n.FlatValue()
203 nCum[canonicalName] += n.CumValue()
206 // Process graph and build various node and edge maps which will
207 // be consumed by AST walk.
208 for _, n := range g.Nodes {
209 seenStartLine = seenStartLine || n.Info.StartLine != 0
211 p.TotalNodeWeight += n.FlatValue()
212 canonicalName := n.Info.Name
213 // Create the key to the nodeMapKey.
214 nodeinfo := NodeMapKey{
215 CallerName: canonicalName,
216 CallSiteOffset: n.Info.Lineno - n.Info.StartLine,
219 for _, e := range n.Out {
220 p.TotalEdgeWeight += e.WeightValue()
221 nodeinfo.CalleeName = e.Dest.Info.Name
222 if w, ok := p.NodeMap[nodeinfo]; ok {
223 w.EWeight += e.WeightValue()
225 weights := new(Weights)
226 weights.NFlat = nFlat[canonicalName]
227 weights.NCum = nCum[canonicalName]
228 weights.EWeight = e.WeightValue()
229 p.NodeMap[nodeinfo] = weights
234 if p.TotalNodeWeight == 0 || p.TotalEdgeWeight == 0 {
235 return nil // accept but ignore profile with no samples.
239 // TODO(prattmic): If Function.start_line is missing we could
240 // fall back to using absolute line numbers, which is better
242 return fmt.Errorf("profile missing Function.start_line data (Go version of profiled application too old? Go 1.20+ automatically adds this to profiles)")
248 // initializeIRGraph builds the IRGraph by visiting all the ir.Func in decl list
250 func (p *Profile) initializeIRGraph() {
251 // Bottomup walk over the function to create IRGraph.
252 ir.VisitFuncsBottomUp(typecheck.Target.Decls, func(list []*ir.Func, recursive bool) {
253 for _, n := range list {
259 // VisitIR traverses the body of each ir.Func and use NodeMap to determine if
260 // we need to add an edge from ir.Func and any node in the ir.Func body.
261 func (p *Profile) VisitIR(fn *ir.Func) {
264 if g.IRNodes == nil {
265 g.IRNodes = make(map[string]*IRNode)
267 if g.OutEdges == nil {
268 g.OutEdges = make(map[*IRNode][]*IREdge)
270 if g.InEdges == nil {
271 g.InEdges = make(map[*IRNode][]*IREdge)
273 name := ir.PkgFuncName(fn)
276 if g.IRNodes[name] == nil {
277 g.IRNodes[name] = node
279 // Create the key for the NodeMapKey.
280 nodeinfo := NodeMapKey{
285 // If the node exists, then update its node weight.
286 if weights, ok := p.NodeMap[nodeinfo]; ok {
287 g.IRNodes[name].Flat = weights.NFlat
288 g.IRNodes[name].Cum = weights.NCum
291 // Recursively walk over the body of the function to create IRGraph edges.
292 p.createIRGraphEdge(fn, g.IRNodes[name], name)
295 // NodeLineOffset returns the line offset of n in fn.
296 func NodeLineOffset(n ir.Node, fn *ir.Func) int {
297 // See "A note on line numbers" at the top of the file.
298 line := int(base.Ctxt.InnermostPos(n.Pos()).RelLine())
299 startLine := int(base.Ctxt.InnermostPos(fn.Pos()).RelLine())
300 return line - startLine
303 // addIREdge adds an edge between caller and new node that points to `callee`
304 // based on the profile-graph and NodeMap.
305 func (p *Profile) addIREdge(caller *IRNode, callername string, call ir.Node, callee *ir.Func) {
308 // Create an IRNode for the callee.
309 calleenode := new(IRNode)
310 calleenode.AST = callee
311 calleename := ir.PkgFuncName(callee)
313 // Create key for NodeMapKey.
314 nodeinfo := NodeMapKey{
315 CallerName: callername,
316 CalleeName: calleename,
317 CallSiteOffset: NodeLineOffset(call, caller.AST),
320 // Create the callee node with node weight.
321 if g.IRNodes[calleename] == nil {
322 g.IRNodes[calleename] = calleenode
323 nodeinfo2 := NodeMapKey{
324 CallerName: calleename,
328 if weights, ok := p.NodeMap[nodeinfo2]; ok {
329 g.IRNodes[calleename].Flat = weights.NFlat
330 g.IRNodes[calleename].Cum = weights.NCum
334 if weights, ok := p.NodeMap[nodeinfo]; ok {
335 caller.Flat = weights.NFlat
336 caller.Cum = weights.NCum
338 // Add edge in the IRGraph from caller to callee.
339 info := &IREdge{Src: caller, Dst: g.IRNodes[calleename], Weight: weights.EWeight, CallSiteOffset: nodeinfo.CallSiteOffset}
340 g.OutEdges[caller] = append(g.OutEdges[caller], info)
341 g.InEdges[g.IRNodes[calleename]] = append(g.InEdges[g.IRNodes[calleename]], info)
343 nodeinfo.CalleeName = ""
344 nodeinfo.CallSiteOffset = 0
345 if weights, ok := p.NodeMap[nodeinfo]; ok {
346 caller.Flat = weights.NFlat
347 caller.Cum = weights.NCum
348 info := &IREdge{Src: caller, Dst: g.IRNodes[calleename], Weight: 0, CallSiteOffset: nodeinfo.CallSiteOffset}
349 g.OutEdges[caller] = append(g.OutEdges[caller], info)
350 g.InEdges[g.IRNodes[calleename]] = append(g.InEdges[g.IRNodes[calleename]], info)
352 info := &IREdge{Src: caller, Dst: g.IRNodes[calleename], Weight: 0, CallSiteOffset: nodeinfo.CallSiteOffset}
353 g.OutEdges[caller] = append(g.OutEdges[caller], info)
354 g.InEdges[g.IRNodes[calleename]] = append(g.InEdges[g.IRNodes[calleename]], info)
359 // createIRGraphEdge traverses the nodes in the body of ir.Func and add edges between callernode which points to the ir.Func and the nodes in the body.
360 func (p *Profile) createIRGraphEdge(fn *ir.Func, callernode *IRNode, name string) {
361 var doNode func(ir.Node) bool
362 doNode = func(n ir.Node) bool {
365 ir.DoChildren(n, doNode)
367 call := n.(*ir.CallExpr)
368 // Find the callee function from the call site and add the edge.
369 callee := inlCallee(call.X)
371 p.addIREdge(callernode, name, n, callee)
374 call := n.(*ir.CallExpr)
375 // Find the callee method from the call site and add the edge.
376 callee := ir.MethodExprName(call.X).Func
377 p.addIREdge(callernode, name, n, callee)
384 // WeightInPercentage converts profile weights to a percentage.
385 func WeightInPercentage(value int64, total int64) float64 {
386 return (float64(value) / float64(total)) * 100
389 // PrintWeightedCallGraphDOT prints IRGraph in DOT format.
390 func (p *Profile) PrintWeightedCallGraphDOT(edgeThreshold float64) {
391 fmt.Printf("\ndigraph G {\n")
392 fmt.Printf("forcelabels=true;\n")
394 // List of functions in this package.
395 funcs := make(map[string]struct{})
396 ir.VisitFuncsBottomUp(typecheck.Target.Decls, func(list []*ir.Func, recursive bool) {
397 for _, f := range list {
398 name := ir.PkgFuncName(f)
399 funcs[name] = struct{}{}
403 // Determine nodes of DOT.
404 nodes := make(map[string]*ir.Func)
405 for name := range funcs {
406 if n, ok := p.WeightedCG.IRNodes[name]; ok {
407 for _, e := range p.WeightedCG.OutEdges[n] {
408 if _, ok := nodes[ir.PkgFuncName(e.Src.AST)]; !ok {
409 nodes[ir.PkgFuncName(e.Src.AST)] = e.Src.AST
411 if _, ok := nodes[ir.PkgFuncName(e.Dst.AST)]; !ok {
412 nodes[ir.PkgFuncName(e.Dst.AST)] = e.Dst.AST
415 if _, ok := nodes[ir.PkgFuncName(n.AST)]; !ok {
416 nodes[ir.PkgFuncName(n.AST)] = n.AST
422 for name, ast := range nodes {
423 if n, ok := p.WeightedCG.IRNodes[name]; ok {
424 nodeweight := WeightInPercentage(n.Flat, p.TotalNodeWeight)
427 fmt.Printf("\"%v\" [color=%v,label=\"%v,freq=%.2f,inl_cost=%d\"];\n", ir.PkgFuncName(ast), color, ir.PkgFuncName(ast), nodeweight, ast.Inl.Cost)
429 fmt.Printf("\"%v\" [color=%v, label=\"%v,freq=%.2f\"];\n", ir.PkgFuncName(ast), color, ir.PkgFuncName(ast), nodeweight)
434 ir.VisitFuncsBottomUp(typecheck.Target.Decls, func(list []*ir.Func, recursive bool) {
435 for _, f := range list {
436 name := ir.PkgFuncName(f)
437 if n, ok := p.WeightedCG.IRNodes[name]; ok {
438 for _, e := range p.WeightedCG.OutEdges[n] {
439 edgepercent := WeightInPercentage(e.Weight, p.TotalEdgeWeight)
440 if edgepercent > edgeThreshold {
441 fmt.Printf("edge [color=red, style=solid];\n")
443 fmt.Printf("edge [color=black, style=solid];\n")
446 fmt.Printf("\"%v\" -> \"%v\" [label=\"%.2f\"];\n", ir.PkgFuncName(n.AST), ir.PkgFuncName(e.Dst.AST), edgepercent)
454 // RedirectEdges deletes and redirects out-edges from node cur based on
455 // inlining information via inlinedCallSites.
457 // CallSiteInfo.Callee must be nil.
458 func (p *Profile) RedirectEdges(cur *IRNode, inlinedCallSites map[CallSiteInfo]struct{}) {
462 outs := g.OutEdges[cur]
466 _, found := inlinedCallSites[CallSiteInfo{LineOffset: outEdge.CallSiteOffset, Caller: cur.AST}]
468 for _, InEdge := range g.InEdges[cur] {
469 if _, ok := inlinedCallSites[CallSiteInfo{LineOffset: InEdge.CallSiteOffset, Caller: InEdge.Src.AST}]; ok {
470 weight := g.calculateWeight(InEdge.Src, cur)
471 g.redirectEdge(InEdge.Src, outEdge, weight)
476 if found || redirected {
478 outs = g.OutEdges[cur]
485 // redirectEdge redirects a node's out-edge to one of its parent nodes, cloning is
486 // required as the node might be inlined in multiple call-sites.
487 // TODO: adjust the in-edges of outEdge.Dst if necessary
488 func (g *IRGraph) redirectEdge(parent *IRNode, outEdge *IREdge, weight int64) {
489 edge := &IREdge{Src: parent, Dst: outEdge.Dst, Weight: weight * outEdge.Weight, CallSiteOffset: outEdge.CallSiteOffset}
490 g.OutEdges[parent] = append(g.OutEdges[parent], edge)
493 // remove deletes the cur-node's out-edges at index idx.
494 func (g *IRGraph) remove(cur *IRNode, i int) {
495 if len(g.OutEdges[cur]) >= 2 {
496 g.OutEdges[cur][i] = g.OutEdges[cur][len(g.OutEdges[cur])-1]
497 g.OutEdges[cur] = g.OutEdges[cur][:len(g.OutEdges[cur])-1]
499 delete(g.OutEdges, cur)
503 // calculateWeight calculates the weight of the new redirected edge.
504 func (g *IRGraph) calculateWeight(parent *IRNode, cur *IRNode) int64 {
507 for _, InEdge := range g.InEdges[cur] {
509 if InEdge.Src == parent {
522 // inlCallee is same as the implementation for inl.go with one change. The change is that we do not invoke CanInline on a closure.
523 func inlCallee(fn ir.Node) *ir.Func {
524 fn = ir.StaticValue(fn)
527 fn := fn.(*ir.SelectorExpr)
528 n := ir.MethodExprName(fn)
529 // Check that receiver type matches fn.X.
530 // TODO(mdempsky): Handle implicit dereference
531 // of pointer receiver argument?
532 if n == nil || !types.Identical(n.Type().Recv().Type, fn.X.Type()) {
538 if fn.Class == ir.PFUNC {
542 fn := fn.(*ir.ClosureExpr)