cmd/compile: enable PGO-driven call devirtualization

[gostls13.git] / src / cmd / compile / internal / pgo / irgraph.go

// Copyright 2022 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.

// A note on line numbers: when working with line numbers, we always use the
// binary-visible relative line number. i.e., the line number as adjusted by
// //line directives (ctxt.InnermostPos(ir.Node.Pos()).RelLine()). Use
// NodeLineOffset to compute line offsets.
//
// If you are thinking, "wait, doesn't that just make things more complex than
// using the real line number?", then you are 100% correct. Unfortunately,
// pprof profiles generated by the runtime always contain line numbers as
// adjusted by //line directives (because that is what we put in pclntab). Thus
// for the best behavior when attempting to match the source with the profile
// it makes sense to use the same line number space.
//
// Some of the effects of this to keep in mind:
//
//  - For files without //line directives there is no impact, as RelLine() ==
//    Line().
//  - For functions entirely covered by the same //line directive (i.e., a
//    directive before the function definition and no directives within the
//    function), there should also be no impact, as line offsets within the
//    function should be the same as the real line offsets.
//  - Functions containing //line directives may be impacted. As fake line
//    numbers need not be monotonic, we may compute negative line offsets. We
//    should accept these and attempt to use them for best-effort matching, as
//    these offsets should still match if the source is unchanged, and may
//    continue to match with changed source depending on the impact of the
//    changes on fake line numbers.
//  - Functions containing //line directives may also contain duplicate lines,
//    making it ambiguous which call the profile is referencing. This is a
//    similar problem to multiple calls on a single real line, as we don't
//    currently track column numbers.
//
// Long term it would be best to extend pprof profiles to include real line
// numbers. Until then, we have to live with these complexities. Luckily,
// //line directives that change line numbers in strange ways should be rare,
// and failing PGO matching on these files is not too big of a loss.

package pgo

import (
        "cmd/compile/internal/base"
        "cmd/compile/internal/ir"
        "cmd/compile/internal/pgo/internal/graph"
        "cmd/compile/internal/typecheck"
        "cmd/compile/internal/types"
        "fmt"
        "internal/profile"
        "os"
)

// IRGraph is a call graph with nodes pointing to IRs of functions and edges
// carrying weights and callsite information.
//
// Nodes for indirect calls may have missing IR (IRNode.AST == nil) if the node
// is not visible from this package (e.g., not in the transitive deps). Keeping
// these nodes allows determining the hottest edge from a call even if that
// callee is not available.
//
// TODO(prattmic): Consider merging this data structure with Graph. This is
// effectively a copy of Graph aggregated to line number and pointing to IR.
type IRGraph struct {
        // Nodes of the graph
        IRNodes map[string]*IRNode
}

// IRNode represents a node (function) in the IRGraph.
type IRNode struct {
        // Pointer to the IR of the Function represented by this node.
        AST *ir.Func
        // Linker symbol name of the Function represented by this node.
        // Populated only if AST == nil.
        LinkerSymbolName string

        // Set of out-edges in the callgraph. The map uniquely identifies each
        // edge based on the callsite and callee, for fast lookup.
        OutEdges map[NodeMapKey]*IREdge
}

// Name returns the symbol name of this function.
func (i *IRNode) Name() string {
        if i.AST != nil {
                return ir.LinkFuncName(i.AST)
        }
        return i.LinkerSymbolName
}

// IREdge represents a call edge in the IRGraph with source, destination,
// weight, callsite, and line number information.
type IREdge struct {
        // Source and destination of the edge in IRNode.
        Src, Dst       *IRNode
        Weight         int64
        CallSiteOffset int // Line offset from function start line.
}

// NodeMapKey represents a hash key to identify unique call-edges in profile
// and in IR. Used for deduplication of call edges found in profile.
//
// TODO(prattmic): rename to something more descriptive.
type NodeMapKey struct {
        CallerName     string
        CalleeName     string
        CallSiteOffset int // Line offset from function start line.
}

// Weights capture both node weight and edge weight.
type Weights struct {
        NFlat   int64
        NCum    int64
        EWeight int64
}

// CallSiteInfo captures call-site information and its caller/callee.
type CallSiteInfo struct {
        LineOffset int // Line offset from function start line.
        Caller     *ir.Func
        Callee     *ir.Func
}

// Profile contains the processed PGO profile and weighted call graph used for
// PGO optimizations.
type Profile struct {
        // Aggregated NodeWeights and EdgeWeights across the profile. This
        // helps us determine the percentage threshold for hot/cold
        // partitioning.
        TotalNodeWeight int64
        TotalEdgeWeight int64

        // NodeMap contains all unique call-edges in the profile and their
        // aggregated weight.
        NodeMap map[NodeMapKey]*Weights

        // WeightedCG represents the IRGraph built from profile, which we will
        // update as part of inlining.
        WeightedCG *IRGraph
}

// New generates a profile-graph from the profile.
func New(profileFile string) (*Profile, error) {
        f, err := os.Open(profileFile)
        if err != nil {
                return nil, fmt.Errorf("error opening profile: %w", err)
        }
        defer f.Close()
        profile, err := profile.Parse(f)
        if err != nil {
                return nil, fmt.Errorf("error parsing profile: %w", err)
        }

        if len(profile.Sample) == 0 {
                // We accept empty profiles, but there is nothing to do.
                return nil, nil
        }

        valueIndex := -1
        for i, s := range profile.SampleType {
                // Samples count is the raw data collected, and CPU nanoseconds is just
                // a scaled version of it, so either one we can find is fine.
                if (s.Type == "samples" && s.Unit == "count") ||
                        (s.Type == "cpu" && s.Unit == "nanoseconds") {
                        valueIndex = i
                        break
                }
        }

        if valueIndex == -1 {
                return nil, fmt.Errorf(`profile does not contain a sample index with value/type "samples/count" or cpu/nanoseconds"`)
        }

        g := graph.NewGraph(profile, &graph.Options{
                SampleValue: func(v []int64) int64 { return v[valueIndex] },
        })

        p := &Profile{
                NodeMap: make(map[NodeMapKey]*Weights),
                WeightedCG: &IRGraph{
                        IRNodes: make(map[string]*IRNode),
                },
        }

        // Build the node map and totals from the profile graph.
        if err := p.processprofileGraph(g); err != nil {
                return nil, err
        }

        if p.TotalNodeWeight == 0 || p.TotalEdgeWeight == 0 {
                return nil, nil // accept but ignore profile with no samples.
        }

        // Create package-level call graph with weights from profile and IR.
        p.initializeIRGraph()

        return p, nil
}

// processprofileGraph builds various maps from the profile-graph.
//
// It initializes NodeMap and Total{Node,Edge}Weight based on the name and
// callsite to compute node and edge weights which will be used later on to
// create edges for WeightedCG.
//
// Caller should ignore the profile if p.TotalNodeWeight == 0 || p.TotalEdgeWeight == 0.
func (p *Profile) processprofileGraph(g *graph.Graph) error {
        nFlat := make(map[string]int64)
        nCum := make(map[string]int64)
        seenStartLine := false

        // Accummulate weights for the same node.
        for _, n := range g.Nodes {
                canonicalName := n.Info.Name
                nFlat[canonicalName] += n.FlatValue()
                nCum[canonicalName] += n.CumValue()
        }

        // Process graph and build various node and edge maps which will
        // be consumed by AST walk.
        for _, n := range g.Nodes {
                seenStartLine = seenStartLine || n.Info.StartLine != 0

                p.TotalNodeWeight += n.FlatValue()
                canonicalName := n.Info.Name
                // Create the key to the nodeMapKey.
                nodeinfo := NodeMapKey{
                        CallerName:     canonicalName,
                        CallSiteOffset: n.Info.Lineno - n.Info.StartLine,
                }

                for _, e := range n.Out {
                        p.TotalEdgeWeight += e.WeightValue()
                        nodeinfo.CalleeName = e.Dest.Info.Name
                        if w, ok := p.NodeMap[nodeinfo]; ok {
                                w.EWeight += e.WeightValue()
                        } else {
                                weights := new(Weights)
                                weights.NFlat = nFlat[canonicalName]
                                weights.NCum = nCum[canonicalName]
                                weights.EWeight = e.WeightValue()
                                p.NodeMap[nodeinfo] = weights
                        }
                }
        }

        if p.TotalNodeWeight == 0 || p.TotalEdgeWeight == 0 {
                return nil // accept but ignore profile with no samples.
        }

        if !seenStartLine {
                // TODO(prattmic): If Function.start_line is missing we could
                // fall back to using absolute line numbers, which is better
                // than nothing.
                return fmt.Errorf("profile missing Function.start_line data (Go version of profiled application too old? Go 1.20+ automatically adds this to profiles)")
        }

        return nil
}

// initializeIRGraph builds the IRGraph by visiting all the ir.Func in decl list
// of a package.
func (p *Profile) initializeIRGraph() {
        // Bottomup walk over the function to create IRGraph.
        ir.VisitFuncsBottomUp(typecheck.Target.Decls, func(list []*ir.Func, recursive bool) {
                for _, fn := range list {
                        p.VisitIR(fn)
                }
        })

        // Add additional edges for indirect calls. This must be done second so
        // that IRNodes is fully populated (see the dummy node TODO in
        // addIndirectEdges).
        //
        // TODO(prattmic): VisitIR above populates the graph via direct calls
        // discovered via the IR. addIndirectEdges populates the graph via
        // calls discovered via the profile. This combination of opposite
        // approaches is a bit awkward, particularly because direct calls are
        // discoverable via the profile as well. Unify these into a single
        // approach.
        p.addIndirectEdges()
}

// VisitIR traverses the body of each ir.Func and use NodeMap to determine if
// we need to add an edge from ir.Func and any node in the ir.Func body.
func (p *Profile) VisitIR(fn *ir.Func) {
        g := p.WeightedCG

        if g.IRNodes == nil {
                g.IRNodes = make(map[string]*IRNode)
        }

        name := ir.LinkFuncName(fn)
        node, ok := g.IRNodes[name]
        if !ok {
                node = &IRNode{
                        AST: fn,
                }
                g.IRNodes[name] = node
        }

        // Recursively walk over the body of the function to create IRGraph edges.
        p.createIRGraphEdge(fn, node, name)
}

// NodeLineOffset returns the line offset of n in fn.
func NodeLineOffset(n ir.Node, fn *ir.Func) int {
        // See "A note on line numbers" at the top of the file.
        line := int(base.Ctxt.InnermostPos(n.Pos()).RelLine())
        startLine := int(base.Ctxt.InnermostPos(fn.Pos()).RelLine())
        return line - startLine
}

// addIREdge adds an edge between caller and new node that points to `callee`
// based on the profile-graph and NodeMap.
func (p *Profile) addIREdge(callerNode *IRNode, callerName string, call ir.Node, callee *ir.Func) {
        g := p.WeightedCG

        calleeName := ir.LinkFuncName(callee)
        calleeNode, ok := g.IRNodes[calleeName]
        if !ok {
                calleeNode = &IRNode{
                        AST: callee,
                }
                g.IRNodes[calleeName] = calleeNode
        }

        nodeinfo := NodeMapKey{
                CallerName:     callerName,
                CalleeName:     calleeName,
                CallSiteOffset: NodeLineOffset(call, callerNode.AST),
        }

        var weight int64
        if weights, ok := p.NodeMap[nodeinfo]; ok {
                weight = weights.EWeight
        }

        // Add edge in the IRGraph from caller to callee.
        edge := &IREdge{
                Src:            callerNode,
                Dst:            calleeNode,
                Weight:         weight,
                CallSiteOffset: nodeinfo.CallSiteOffset,
        }

        if callerNode.OutEdges == nil {
                callerNode.OutEdges = make(map[NodeMapKey]*IREdge)
        }
        callerNode.OutEdges[nodeinfo] = edge
}

// addIndirectEdges adds indirect call edges found in the profile to the graph,
// to be used for devirtualization.
//
// targetDeclFuncs is the set of functions in typecheck.Target.Decls. Only
// edges from these functions will be added.
//
// Devirtualization is only applied to typecheck.Target.Decls functions, so there
// is no need to add edges from other functions.
//
// N.B. despite the name, addIndirectEdges will add any edges discovered via
// the profile. We don't know for sure that they are indirect, but assume they
// are since direct calls would already be added. (e.g., direct calls that have
// been deleted from source since the profile was taken would be added here).
//
// TODO(prattmic): Devirtualization runs before inlining, so we can't devirtualize
// calls inside inlined call bodies. If we did add that, we'd need edges from
// inlined bodies as well.
func (p *Profile) addIndirectEdges() {
        g := p.WeightedCG

        // g.IRNodes is populated with the set of functions in the local
        // package build by VisitIR. We want to filter for local functions
        // below, but we also add unknown callees to IRNodes as we go. So make
        // an initial copy of IRNodes to recall just the local functions.
        localNodes := make(map[string]*IRNode, len(g.IRNodes))
        for k, v := range g.IRNodes {
                localNodes[k] = v
        }

        for key, weights := range p.NodeMap {
                // All callers in the local package build were added to IRNodes
                // in VisitIR. If a caller isn't in the local package build we
                // can skip adding edges, since we won't be devirtualizing in
                // them anyway. This keeps the graph smaller.
                callerNode, ok := localNodes[key.CallerName]
                if !ok {
                        continue
                }

                // Already handled this edge?
                if _, ok := callerNode.OutEdges[key]; ok {
                        continue
                }

                calleeNode, ok := g.IRNodes[key.CalleeName]
                if !ok {
                        // IR is missing for this callee. Most likely this is
                        // because the callee isn't in the transitive deps of
                        // this package.
                        //
                        // Record this call anyway. If this is the hottest,
                        // then we want to skip devirtualization rather than
                        // devirtualizing to the second most common callee.
                        //
                        // TODO(prattmic): VisitIR populates IRNodes with all
                        // of the functions discovered via local package
                        // function declarations and calls. Thus we could miss
                        // functions that are available in export data of
                        // transitive deps, but aren't directly reachable. We
                        // need to do a lookup directly from package export
                        // data to get complete coverage.
                        calleeNode = &IRNode{
                                LinkerSymbolName: key.CalleeName,
                                // TODO: weights? We don't need them.
                        }
                        // Add dummy node back to IRNodes. We don't need this
                        // directly, but PrintWeightedCallGraphDOT uses these
                        // to print nodes.
                        g.IRNodes[key.CalleeName] = calleeNode
                }
                edge := &IREdge{
                        Src:            callerNode,
                        Dst:            calleeNode,
                        Weight:         weights.EWeight,
                        CallSiteOffset: key.CallSiteOffset,
                }

                if callerNode.OutEdges == nil {
                        callerNode.OutEdges = make(map[NodeMapKey]*IREdge)
                }
                callerNode.OutEdges[key] = edge
        }
}

// createIRGraphEdge traverses the nodes in the body of ir.Func and adds edges
// between the callernode which points to the ir.Func and the nodes in the
// body.
func (p *Profile) createIRGraphEdge(fn *ir.Func, callernode *IRNode, name string) {
        ir.VisitList(fn.Body, func(n ir.Node) {
                switch n.Op() {
                case ir.OCALLFUNC:
                        call := n.(*ir.CallExpr)
                        // Find the callee function from the call site and add the edge.
                        callee := DirectCallee(call.X)
                        if callee != nil {
                                p.addIREdge(callernode, name, n, callee)
                        }
                case ir.OCALLMETH:
                        call := n.(*ir.CallExpr)
                        // Find the callee method from the call site and add the edge.
                        callee := ir.MethodExprName(call.X).Func
                        p.addIREdge(callernode, name, n, callee)
                }
        })
}

// WeightInPercentage converts profile weights to a percentage.
func WeightInPercentage(value int64, total int64) float64 {
        return (float64(value) / float64(total)) * 100
}

// PrintWeightedCallGraphDOT prints IRGraph in DOT format.
func (p *Profile) PrintWeightedCallGraphDOT(edgeThreshold float64) {
        fmt.Printf("\ndigraph G {\n")
        fmt.Printf("forcelabels=true;\n")

        // List of functions in this package.
        funcs := make(map[string]struct{})
        ir.VisitFuncsBottomUp(typecheck.Target.Decls, func(list []*ir.Func, recursive bool) {
                for _, f := range list {
                        name := ir.LinkFuncName(f)
                        funcs[name] = struct{}{}
                }
        })

        // Determine nodes of DOT.
        //
        // Note that ir.Func may be nil for functions not visible from this
        // package.
        nodes := make(map[string]*ir.Func)
        for name := range funcs {
                if n, ok := p.WeightedCG.IRNodes[name]; ok {
                        for _, e := range n.OutEdges {
                                if _, ok := nodes[e.Src.Name()]; !ok {
                                        nodes[e.Src.Name()] = e.Src.AST
                                }
                                if _, ok := nodes[e.Dst.Name()]; !ok {
                                        nodes[e.Dst.Name()] = e.Dst.AST
                                }
                        }
                        if _, ok := nodes[n.Name()]; !ok {
                                nodes[n.Name()] = n.AST
                        }
                }
        }

        // Print nodes.
        for name, ast := range nodes {
                if _, ok := p.WeightedCG.IRNodes[name]; ok {
                        style := "solid"
                        if ast == nil {
                                style = "dashed"
                        }

                        if ast != nil && ast.Inl != nil {
                                fmt.Printf("\"%v\" [color=black, style=%s, label=\"%v,inl_cost=%d\"];\n", name, style, name, ast.Inl.Cost)
                        } else {
                                fmt.Printf("\"%v\" [color=black, style=%s, label=\"%v\"];\n", name, style, name)
                        }
                }
        }
        // Print edges.
        ir.VisitFuncsBottomUp(typecheck.Target.Decls, func(list []*ir.Func, recursive bool) {
                for _, f := range list {
                        name := ir.LinkFuncName(f)
                        if n, ok := p.WeightedCG.IRNodes[name]; ok {
                                for _, e := range n.OutEdges {
                                        style := "solid"
                                        if e.Dst.AST == nil {
                                                style = "dashed"
                                        }
                                        color := "black"
                                        edgepercent := WeightInPercentage(e.Weight, p.TotalEdgeWeight)
                                        if edgepercent > edgeThreshold {
                                                color = "red"
                                        }

                                        fmt.Printf("edge [color=%s, style=%s];\n", color, style)
                                        fmt.Printf("\"%v\" -> \"%v\" [label=\"%.2f\"];\n", n.Name(), e.Dst.Name(), edgepercent)
                                }
                        }
                }
        })
        fmt.Printf("}\n")
}

// DirectCallee takes a function-typed expression and returns the underlying
// function that it refers to if statically known. Otherwise, it returns nil.
//
// Equivalent to inline.inlCallee without calling CanInline on closures.
func DirectCallee(fn ir.Node) *ir.Func {
        fn = ir.StaticValue(fn)
        switch fn.Op() {
        case ir.OMETHEXPR:
                fn := fn.(*ir.SelectorExpr)
                n := ir.MethodExprName(fn)
                // Check that receiver type matches fn.X.
                // TODO(mdempsky): Handle implicit dereference
                // of pointer receiver argument?
                if n == nil || !types.Identical(n.Type().Recv().Type, fn.X.Type()) {
                        return nil
                }
                return n.Func
        case ir.ONAME:
                fn := fn.(*ir.Name)
                if fn.Class == ir.PFUNC {
                        return fn.Func
                }
        case ir.OCLOSURE:
                fn := fn.(*ir.ClosureExpr)
                c := fn.Func
                return c
        }
        return nil
}