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"
56 // IRGraph is the key data structure that is built from profile. It is
57 // essentially a call graph with nodes pointing to IRs of functions and edges
58 // carrying weights and callsite information. The graph is bidirectional that
59 // helps in removing nodes efficiently.
62 IRNodes map[string]*IRNode
67 // IRNode represents a node in the IRGraph.
69 // Pointer to the IR of the Function represented by this node.
71 // Flat weight of the IRNode, obtained from profile.
73 // Cumulative weight of the IRNode.
77 // IREdgeMap maps an IRNode to its successors.
78 type IREdgeMap map[*IRNode][]*IREdge
80 // IREdge represents a call edge in the IRGraph with source, destination,
81 // weight, callsite, and line number information.
83 // Source and destination of the edge in IRNode.
86 CallSiteOffset int // Line offset from function start line.
89 // NodeMapKey represents a hash key to identify unique call-edges in profile
90 // and in IR. Used for deduplication of call edges found in profile.
91 type NodeMapKey struct {
94 CallSiteOffset int // Line offset from function start line.
97 // Weights capture both node weight and edge weight.
104 // CallSiteInfo captures call-site information and its caller/callee.
105 type CallSiteInfo struct {
106 LineOffset int // Line offset from function start line.
111 // Profile contains the processed PGO profile and weighted call graph used for
112 // PGO optimizations.
113 type Profile struct {
114 // Aggregated NodeWeights and EdgeWeights across the profile. This
115 // helps us determine the percentage threshold for hot/cold
117 TotalNodeWeight int64
118 TotalEdgeWeight int64
120 // NodeMap contains all unique call-edges in the profile and their
121 // aggregated weight.
122 NodeMap map[NodeMapKey]*Weights
124 // WeightedCG represents the IRGraph built from profile, which we will
125 // update as part of inlining.
129 // New generates a profile-graph from the profile.
130 func New(profileFile string) *Profile {
131 f, err := os.Open(profileFile)
133 log.Fatal("failed to open file " + profileFile)
137 profile, err := profile.Parse(f)
139 log.Fatal("failed to Parse profile file.")
143 samplesCountIndex := -1
144 for i, s := range profile.SampleType {
145 // Samples count is the raw data collected, and CPU nanoseconds is just
146 // a scaled version of it, so either one we can find is fine.
147 if (s.Type == "samples" && s.Unit == "count") ||
148 (s.Type == "cpu" && s.Unit == "nanoseconds") {
149 samplesCountIndex = i
154 if samplesCountIndex == -1 {
155 log.Fatal("failed to find CPU samples count or CPU nanoseconds value-types in profile.")
159 g := newGraph(profile, &Options{
161 SampleValue: func(v []int64) int64 { return v[samplesCountIndex] },
165 NodeMap: make(map[NodeMapKey]*Weights),
166 WeightedCG: &IRGraph{
167 IRNodes: make(map[string]*IRNode),
171 // Build the node map and totals from the profile graph.
172 if !p.processprofileGraph(g) {
176 // Create package-level call graph with weights from profile and IR.
177 p.initializeIRGraph()
182 // processprofileGraph builds various maps from the profile-graph.
184 // It initializes NodeMap and Total{Node,Edge}Weight based on the name and
185 // callsite to compute node and edge weights which will be used later on to
186 // create edges for WeightedCG.
187 // Returns whether it successfully processed the profile.
188 func (p *Profile) processprofileGraph(g *Graph) bool {
189 nFlat := make(map[string]int64)
190 nCum := make(map[string]int64)
191 seenStartLine := false
193 // Accummulate weights for the same node.
194 for _, n := range g.Nodes {
195 canonicalName := n.Info.Name
196 nFlat[canonicalName] += n.FlatValue()
197 nCum[canonicalName] += n.CumValue()
200 // Process graph and build various node and edge maps which will
201 // be consumed by AST walk.
202 for _, n := range g.Nodes {
203 seenStartLine = seenStartLine || n.Info.StartLine != 0
205 p.TotalNodeWeight += n.FlatValue()
206 canonicalName := n.Info.Name
207 // Create the key to the nodeMapKey.
208 nodeinfo := NodeMapKey{
209 CallerName: canonicalName,
210 CallSiteOffset: n.Info.Lineno - n.Info.StartLine,
213 for _, e := range n.Out {
214 p.TotalEdgeWeight += e.WeightValue()
215 nodeinfo.CalleeName = e.Dest.Info.Name
216 if w, ok := p.NodeMap[nodeinfo]; ok {
217 w.EWeight += e.WeightValue()
219 weights := new(Weights)
220 weights.NFlat = nFlat[canonicalName]
221 weights.NCum = nCum[canonicalName]
222 weights.EWeight = e.WeightValue()
223 p.NodeMap[nodeinfo] = weights
228 if p.TotalNodeWeight == 0 || p.TotalEdgeWeight == 0 {
229 return false // accept but ignore profile with no sample
233 // TODO(prattic): If Function.start_line is missing we could
234 // fall back to using absolute line numbers, which is better
236 log.Fatal("PGO profile missing Function.start_line data (Go version of profiled application too old? Go 1.20+ automatically adds this to profiles)")
242 // initializeIRGraph builds the IRGraph by visiting all the ir.Func in decl list
244 func (p *Profile) initializeIRGraph() {
245 // Bottomup walk over the function to create IRGraph.
246 ir.VisitFuncsBottomUp(typecheck.Target.Decls, func(list []*ir.Func, recursive bool) {
247 for _, n := range list {
248 p.VisitIR(n, recursive)
253 // VisitIR traverses the body of each ir.Func and use NodeMap to determine if
254 // we need to add an edge from ir.Func and any node in the ir.Func body.
255 func (p *Profile) VisitIR(fn *ir.Func, recursive bool) {
258 if g.IRNodes == nil {
259 g.IRNodes = make(map[string]*IRNode)
261 if g.OutEdges == nil {
262 g.OutEdges = make(map[*IRNode][]*IREdge)
264 if g.InEdges == nil {
265 g.InEdges = make(map[*IRNode][]*IREdge)
267 name := ir.PkgFuncName(fn)
270 if g.IRNodes[name] == nil {
271 g.IRNodes[name] = node
273 // Create the key for the NodeMapKey.
274 nodeinfo := NodeMapKey{
279 // If the node exists, then update its node weight.
280 if weights, ok := p.NodeMap[nodeinfo]; ok {
281 g.IRNodes[name].Flat = weights.NFlat
282 g.IRNodes[name].Cum = weights.NCum
285 // Recursively walk over the body of the function to create IRGraph edges.
286 p.createIRGraphEdge(fn, g.IRNodes[name], name)
289 // NodeLineOffset returns the line offset of n in fn.
290 func NodeLineOffset(n ir.Node, fn *ir.Func) int {
291 // See "A note on line numbers" at the top of the file.
292 line := int(base.Ctxt.InnermostPos(n.Pos()).RelLine())
293 startLine := int(base.Ctxt.InnermostPos(fn.Pos()).RelLine())
294 return line - startLine
297 // addIREdge adds an edge between caller and new node that points to `callee`
298 // based on the profile-graph and NodeMap.
299 func (p *Profile) addIREdge(caller *IRNode, callername string, call ir.Node, callee *ir.Func) {
302 // Create an IRNode for the callee.
303 calleenode := new(IRNode)
304 calleenode.AST = callee
305 calleename := ir.PkgFuncName(callee)
307 // Create key for NodeMapKey.
308 nodeinfo := NodeMapKey{
309 CallerName: callername,
310 CalleeName: calleename,
311 CallSiteOffset: NodeLineOffset(call, caller.AST),
314 // Create the callee node with node weight.
315 if g.IRNodes[calleename] == nil {
316 g.IRNodes[calleename] = calleenode
317 nodeinfo2 := NodeMapKey{
318 CallerName: calleename,
322 if weights, ok := p.NodeMap[nodeinfo2]; ok {
323 g.IRNodes[calleename].Flat = weights.NFlat
324 g.IRNodes[calleename].Cum = weights.NCum
328 if weights, ok := p.NodeMap[nodeinfo]; ok {
329 caller.Flat = weights.NFlat
330 caller.Cum = weights.NCum
332 // Add edge in the IRGraph from caller to callee.
333 info := &IREdge{Src: caller, Dst: g.IRNodes[calleename], Weight: weights.EWeight, CallSiteOffset: nodeinfo.CallSiteOffset}
334 g.OutEdges[caller] = append(g.OutEdges[caller], info)
335 g.InEdges[g.IRNodes[calleename]] = append(g.InEdges[g.IRNodes[calleename]], info)
337 nodeinfo.CalleeName = ""
338 nodeinfo.CallSiteOffset = 0
339 if weights, ok := p.NodeMap[nodeinfo]; ok {
340 caller.Flat = weights.NFlat
341 caller.Cum = weights.NCum
342 info := &IREdge{Src: caller, Dst: g.IRNodes[calleename], Weight: 0, CallSiteOffset: nodeinfo.CallSiteOffset}
343 g.OutEdges[caller] = append(g.OutEdges[caller], info)
344 g.InEdges[g.IRNodes[calleename]] = append(g.InEdges[g.IRNodes[calleename]], info)
346 info := &IREdge{Src: caller, Dst: g.IRNodes[calleename], Weight: 0, CallSiteOffset: nodeinfo.CallSiteOffset}
347 g.OutEdges[caller] = append(g.OutEdges[caller], info)
348 g.InEdges[g.IRNodes[calleename]] = append(g.InEdges[g.IRNodes[calleename]], info)
353 // 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.
354 func (p *Profile) createIRGraphEdge(fn *ir.Func, callernode *IRNode, name string) {
355 var doNode func(ir.Node) bool
356 doNode = func(n ir.Node) bool {
359 ir.DoChildren(n, doNode)
361 call := n.(*ir.CallExpr)
362 // Find the callee function from the call site and add the edge.
363 callee := inlCallee(call.X)
365 p.addIREdge(callernode, name, n, callee)
368 call := n.(*ir.CallExpr)
369 // Find the callee method from the call site and add the edge.
370 callee := ir.MethodExprName(call.X).Func
371 p.addIREdge(callernode, name, n, callee)
378 // WeightInPercentage converts profile weights to a percentage.
379 func WeightInPercentage(value int64, total int64) float64 {
380 return (float64(value) / float64(total)) * 100
383 // PrintWeightedCallGraphDOT prints IRGraph in DOT format.
384 func (p *Profile) PrintWeightedCallGraphDOT(edgeThreshold float64) {
385 fmt.Printf("\ndigraph G {\n")
386 fmt.Printf("forcelabels=true;\n")
388 // List of functions in this package.
389 funcs := make(map[string]struct{})
390 ir.VisitFuncsBottomUp(typecheck.Target.Decls, func(list []*ir.Func, recursive bool) {
391 for _, f := range list {
392 name := ir.PkgFuncName(f)
393 funcs[name] = struct{}{}
397 // Determine nodes of DOT.
398 nodes := make(map[string]*ir.Func)
399 for name := range funcs {
400 if n, ok := p.WeightedCG.IRNodes[name]; ok {
401 for _, e := range p.WeightedCG.OutEdges[n] {
402 if _, ok := nodes[ir.PkgFuncName(e.Src.AST)]; !ok {
403 nodes[ir.PkgFuncName(e.Src.AST)] = e.Src.AST
405 if _, ok := nodes[ir.PkgFuncName(e.Dst.AST)]; !ok {
406 nodes[ir.PkgFuncName(e.Dst.AST)] = e.Dst.AST
409 if _, ok := nodes[ir.PkgFuncName(n.AST)]; !ok {
410 nodes[ir.PkgFuncName(n.AST)] = n.AST
416 for name, ast := range nodes {
417 if n, ok := p.WeightedCG.IRNodes[name]; ok {
418 nodeweight := WeightInPercentage(n.Flat, p.TotalNodeWeight)
421 fmt.Printf("\"%v\" [color=%v,label=\"%v,freq=%.2f,inl_cost=%d\"];\n", ir.PkgFuncName(ast), color, ir.PkgFuncName(ast), nodeweight, ast.Inl.Cost)
423 fmt.Printf("\"%v\" [color=%v, label=\"%v,freq=%.2f\"];\n", ir.PkgFuncName(ast), color, ir.PkgFuncName(ast), nodeweight)
428 ir.VisitFuncsBottomUp(typecheck.Target.Decls, func(list []*ir.Func, recursive bool) {
429 for _, f := range list {
430 name := ir.PkgFuncName(f)
431 if n, ok := p.WeightedCG.IRNodes[name]; ok {
432 for _, e := range p.WeightedCG.OutEdges[n] {
433 edgepercent := WeightInPercentage(e.Weight, p.TotalEdgeWeight)
434 if edgepercent > edgeThreshold {
435 fmt.Printf("edge [color=red, style=solid];\n")
437 fmt.Printf("edge [color=black, style=solid];\n")
440 fmt.Printf("\"%v\" -> \"%v\" [label=\"%.2f\"];\n", ir.PkgFuncName(n.AST), ir.PkgFuncName(e.Dst.AST), edgepercent)
448 // RedirectEdges deletes and redirects out-edges from node cur based on
449 // inlining information via inlinedCallSites.
451 // CallSiteInfo.Callee must be nil.
452 func (p *Profile) RedirectEdges(cur *IRNode, inlinedCallSites map[CallSiteInfo]struct{}) {
455 for i, outEdge := range g.OutEdges[cur] {
456 if _, found := inlinedCallSites[CallSiteInfo{LineOffset: outEdge.CallSiteOffset, Caller: cur.AST}]; !found {
457 for _, InEdge := range g.InEdges[cur] {
458 if _, ok := inlinedCallSites[CallSiteInfo{LineOffset: InEdge.CallSiteOffset, Caller: InEdge.Src.AST}]; ok {
459 weight := g.calculateWeight(InEdge.Src, cur)
460 g.redirectEdge(InEdge.Src, cur, outEdge, weight, i)
469 // redirectEdges deletes the cur node out-edges and redirect them so now these
470 // edges are the parent node out-edges.
471 func (g *IRGraph) redirectEdges(parent *IRNode, cur *IRNode) {
472 for _, outEdge := range g.OutEdges[cur] {
474 g.OutEdges[parent] = append(g.OutEdges[parent], outEdge)
476 delete(g.OutEdges, cur)
479 // redirectEdge deletes the cur-node's out-edges and redirect them so now these
480 // edges are the parent node out-edges.
481 func (g *IRGraph) redirectEdge(parent *IRNode, cur *IRNode, outEdge *IREdge, weight int64, idx int) {
483 outEdge.Weight = weight * outEdge.Weight
484 g.OutEdges[parent] = append(g.OutEdges[parent], outEdge)
488 // remove deletes the cur-node's out-edges at index idx.
489 func (g *IRGraph) remove(cur *IRNode, i int) {
490 if len(g.OutEdges[cur]) >= 2 {
491 g.OutEdges[cur][i] = g.OutEdges[cur][len(g.OutEdges[cur])-1]
492 g.OutEdges[cur] = g.OutEdges[cur][:len(g.OutEdges[cur])-1]
494 delete(g.OutEdges, cur)
498 // calculateWeight calculates the weight of the new redirected edge.
499 func (g *IRGraph) calculateWeight(parent *IRNode, cur *IRNode) int64 {
502 for _, InEdge := range g.InEdges[cur] {
504 if InEdge.Src == parent {
517 // inlCallee is same as the implementation for inl.go with one change. The change is that we do not invoke CanInline on a closure.
518 func inlCallee(fn ir.Node) *ir.Func {
519 fn = ir.StaticValue(fn)
522 fn := fn.(*ir.SelectorExpr)
523 n := ir.MethodExprName(fn)
524 // Check that receiver type matches fn.X.
525 // TODO(mdempsky): Handle implicit dereference
526 // of pointer receiver argument?
527 if n == nil || !types.Identical(n.Type().Recv().Type, fn.X.Type()) {
533 if fn.Class == ir.PFUNC {
537 fn := fn.(*ir.ClosureExpr)