cmd/compile,cmd/preprofile: move logic to shared common package

The processing performed in cmd/preprofile is a simple version of the
same initial processing performed by cmd/compile/internal/pgo. Refactor
this processing into the new IR-independent cmd/internal/pgo package.

Now cmd/preprofile and cmd/compile run the same code for initial
processing of a pprof profile, guaranteeing that they always stay in
sync.

Since it is now trivial, this CL makes one change to the serialization
format: the entries are ordered by weight. This allows us to avoid
sorting ByWeight on deserialization.

Impact on PGO parsing when compiling cmd/compile with PGO:

* Without preprocessing: PGO parsing ~13.7% of CPU time
* With preprocessing (unsorted): ~2.9% of CPU time (sorting ~1.7%)
* With preprocessing (sorted): ~1.3% of CPU time

The remaining 1.3% of CPU time approximately breaks down as:

* ~0.5% parsing the preprocessed profile
* ~0.7% building weighted IR call graph
  * ~0.5% walking function IR to find direct calls
  * ~0.2% performing lookups for indirect calls targets

For #58102.

Change-Id: Iaba425ea30b063ca195fb2f7b29342961c8a64c2
Reviewed-on: https://go-review.googlesource.com/c/go/+/569337
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Auto-Submit: Michael Pratt <mpratt@google.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>

16 files changed, 685 insertions, 443 deletions

diff --git a/src/cmd/compile/internal/devirtualize/pgo_test.go b/src/cmd/compile/internal/devirtualize/pgo_test.go
index 84c96df122..6ba8e9f907 100644
--- a/src/cmd/compile/internal/devirtualize/pgo_test.go
+++ b/src/cmd/compile/internal/devirtualize/pgo_test.go
@@ -7,10 +7,11 @@ package devirtualize
 import (
 	"cmd/compile/internal/base"
 	"cmd/compile/internal/ir"
-	"cmd/compile/internal/pgo"
+	pgoir "cmd/compile/internal/pgo"
 	"cmd/compile/internal/typecheck"
 	"cmd/compile/internal/types"
 	"cmd/internal/obj"
+	"cmd/internal/pgo"
 	"cmd/internal/src"
 	"testing"
 )
@@ -32,32 +33,32 @@ func makePos(b *src.PosBase, line, col uint) src.XPos {
 }
 
 type profileBuilder struct {
-	p *pgo.Profile
+	p *pgoir.Profile
 }
 
 func newProfileBuilder() *profileBuilder {
-	// findHotConcreteCallee only uses pgo.Profile.WeightedCG, so we're
+	// findHotConcreteCallee only uses pgoir.Profile.WeightedCG, so we're
 	// going to take a shortcut and only construct that.
 	return &profileBuilder{
-		p: &pgo.Profile{
-			WeightedCG: &pgo.IRGraph{
-				IRNodes: make(map[string]*pgo.IRNode),
+		p: &pgoir.Profile{
+			WeightedCG: &pgoir.IRGraph{
+				IRNodes: make(map[string]*pgoir.IRNode),
 			},
 		},
 	}
 }
 
 // Profile returns the constructed profile.
-func (p *profileBuilder) Profile() *pgo.Profile {
+func (p *profileBuilder) Profile() *pgoir.Profile {
 	return p.p
 }
 
 // NewNode creates a new IRNode and adds it to the profile.
 //
 // fn may be nil, in which case the node will set LinkerSymbolName.
-func (p *profileBuilder) NewNode(name string, fn *ir.Func) *pgo.IRNode {
-	n := &pgo.IRNode{
-		OutEdges: make(map[pgo.NamedCallEdge]*pgo.IREdge),
+func (p *profileBuilder) NewNode(name string, fn *ir.Func) *pgoir.IRNode {
+	n := &pgoir.IRNode{
+		OutEdges: make(map[pgo.NamedCallEdge]*pgoir.IREdge),
 	}
 	if fn != nil {
 		n.AST = fn
@@ -69,13 +70,13 @@ func (p *profileBuilder) NewNode(name string, fn *ir.Func) *pgo.IRNode {
 }
 
 // Add a new call edge from caller to callee.
-func addEdge(caller, callee *pgo.IRNode, offset int, weight int64) {
+func addEdge(caller, callee *pgoir.IRNode, offset int, weight int64) {
 	namedEdge := pgo.NamedCallEdge{
 		CallerName:     caller.Name(),
 		CalleeName:     callee.Name(),
 		CallSiteOffset: offset,
 	}
-	irEdge := &pgo.IREdge{
+	irEdge := &pgoir.IREdge{
 		Src:            caller,
 		Dst:            callee,
 		CallSiteOffset: offset,
diff --git a/src/cmd/compile/internal/inline/inl.go b/src/cmd/compile/internal/inline/inl.go
index dd300bbd51..33f454083f 100644
--- a/src/cmd/compile/internal/inline/inl.go
+++ b/src/cmd/compile/internal/inline/inl.go
@@ -36,10 +36,11 @@ import (
 	"cmd/compile/internal/inline/inlheur"
 	"cmd/compile/internal/ir"
 	"cmd/compile/internal/logopt"
-	"cmd/compile/internal/pgo"
+	pgoir "cmd/compile/internal/pgo"
 	"cmd/compile/internal/typecheck"
 	"cmd/compile/internal/types"
 	"cmd/internal/obj"
+	"cmd/internal/pgo"
 )
 
 // Inlining budget parameters, gathered in one place
@@ -58,11 +59,11 @@ const (
 var (
 	// List of all hot callee nodes.
 	// TODO(prattmic): Make this non-global.
-	candHotCalleeMap = make(map[*pgo.IRNode]struct{})
+	candHotCalleeMap = make(map[*pgoir.IRNode]struct{})
 
 	// List of all hot call sites. CallSiteInfo.Callee is always nil.
 	// TODO(prattmic): Make this non-global.
-	candHotEdgeMap = make(map[pgo.CallSiteInfo]struct{})
+	candHotEdgeMap = make(map[pgoir.CallSiteInfo]struct{})
 
 	// Threshold in percentage for hot callsite inlining.
 	inlineHotCallSiteThresholdPercent float64
@@ -78,7 +79,7 @@ var (
 )
 
 // PGOInlinePrologue records the hot callsites from ir-graph.
-func PGOInlinePrologue(p *pgo.Profile) {
+func PGOInlinePrologue(p *pgoir.Profile) {
 	if base.Debug.PGOInlineCDFThreshold != "" {
 		if s, err := strconv.ParseFloat(base.Debug.PGOInlineCDFThreshold, 64); err == nil && s >= 0 && s <= 100 {
 			inlineCDFHotCallSiteThresholdPercent = s
@@ -103,7 +104,7 @@ func PGOInlinePrologue(p *pgo.Profile) {
 		}
 		// mark hot call sites
 		if caller := p.WeightedCG.IRNodes[n.CallerName]; caller != nil && caller.AST != nil {
-			csi := pgo.CallSiteInfo{LineOffset: n.CallSiteOffset, Caller: caller.AST}
+			csi := pgoir.CallSiteInfo{LineOffset: n.CallSiteOffset, Caller: caller.AST}
 			candHotEdgeMap[csi] = struct{}{}
 		}
 	}
@@ -120,7 +121,7 @@ func PGOInlinePrologue(p *pgo.Profile) {
 // (currently only used in debug prints) (in case of equal weights,
 // comparing with the threshold may not accurately reflect which nodes are
 // considered hot).
-func hotNodesFromCDF(p *pgo.Profile) (float64, []pgo.NamedCallEdge) {
+func hotNodesFromCDF(p *pgoir.Profile) (float64, []pgo.NamedCallEdge) {
 	cum := int64(0)
 	for i, n := range p.NamedEdgeMap.ByWeight {
 		w := p.NamedEdgeMap.Weight[n]
@@ -136,7 +137,7 @@ func hotNodesFromCDF(p *pgo.Profile) (float64, []pgo.NamedCallEdge) {
 }
 
 // CanInlineFuncs computes whether a batch of functions are inlinable.
-func CanInlineFuncs(funcs []*ir.Func, profile *pgo.Profile) {
+func CanInlineFuncs(funcs []*ir.Func, profile *pgoir.Profile) {
 	if profile != nil {
 		PGOInlinePrologue(profile)
 	}
@@ -224,7 +225,7 @@ func GarbageCollectUnreferencedHiddenClosures() {
 // possibility that a call to the function might have its score
 // adjusted downwards. If 'verbose' is set, then print a remark where
 // we boost the budget due to PGO.
-func inlineBudget(fn *ir.Func, profile *pgo.Profile, relaxed bool, verbose bool) int32 {
+func inlineBudget(fn *ir.Func, profile *pgoir.Profile, relaxed bool, verbose bool) int32 {
 	// Update the budget for profile-guided inlining.
 	budget := int32(inlineMaxBudget)
 	if profile != nil {
@@ -246,7 +247,7 @@ func inlineBudget(fn *ir.Func, profile *pgo.Profile, relaxed bool, verbose bool)
 // CanInline determines whether fn is inlineable.
 // If so, CanInline saves copies of fn.Body and fn.Dcl in fn.Inl.
 // fn and fn.Body will already have been typechecked.
-func CanInline(fn *ir.Func, profile *pgo.Profile) {
+func CanInline(fn *ir.Func, profile *pgoir.Profile) {
 	if fn.Nname == nil {
 		base.Fatalf("CanInline no nname %+v", fn)
 	}
@@ -451,7 +452,7 @@ type hairyVisitor struct {
 	extraCallCost int32
 	usedLocals    ir.NameSet
 	do            func(ir.Node) bool
-	profile       *pgo.Profile
+	profile       *pgoir.Profile
 }
 
 func (v *hairyVisitor) tooHairy(fn *ir.Func) bool {
@@ -768,7 +769,7 @@ func IsBigFunc(fn *ir.Func) bool {
 
 // TryInlineCall returns an inlined call expression for call, or nil
 // if inlining is not possible.
-func TryInlineCall(callerfn *ir.Func, call *ir.CallExpr, bigCaller bool, profile *pgo.Profile) *ir.InlinedCallExpr {
+func TryInlineCall(callerfn *ir.Func, call *ir.CallExpr, bigCaller bool, profile *pgoir.Profile) *ir.InlinedCallExpr {
 	if base.Flag.LowerL == 0 {
 		return nil
 	}
@@ -804,7 +805,7 @@ func TryInlineCall(callerfn *ir.Func, call *ir.CallExpr, bigCaller bool, profile
 
 // inlCallee takes a function-typed expression and returns the underlying function ONAME
 // that it refers to if statically known. Otherwise, it returns nil.
-func inlCallee(caller *ir.Func, fn ir.Node, profile *pgo.Profile) (res *ir.Func) {
+func inlCallee(caller *ir.Func, fn ir.Node, profile *pgoir.Profile) (res *ir.Func) {
 	fn = ir.StaticValue(fn)
 	switch fn.Op() {
 	case ir.OMETHEXPR:
@@ -877,8 +878,8 @@ func inlineCostOK(n *ir.CallExpr, caller, callee *ir.Func, bigCaller bool) (bool
 	// We'll also allow inlining of hot functions below inlineHotMaxBudget,
 	// but only in small functions.
 
-	lineOffset := pgo.NodeLineOffset(n, caller)
-	csi := pgo.CallSiteInfo{LineOffset: lineOffset, Caller: caller}
+	lineOffset := pgoir.NodeLineOffset(n, caller)
+	csi := pgoir.CallSiteInfo{LineOffset: lineOffset, Caller: caller}
 	if _, ok := candHotEdgeMap[csi]; !ok {
 		// Cold
 		return false, maxCost, metric
@@ -1188,9 +1189,9 @@ func isAtomicCoverageCounterUpdate(cn *ir.CallExpr) bool {
 	return v
 }
 
-func PostProcessCallSites(profile *pgo.Profile) {
+func PostProcessCallSites(profile *pgoir.Profile) {
 	if base.Debug.DumpInlCallSiteScores != 0 {
-		budgetCallback := func(fn *ir.Func, prof *pgo.Profile) (int32, bool) {
+		budgetCallback := func(fn *ir.Func, prof *pgoir.Profile) (int32, bool) {
 			v := inlineBudget(fn, prof, false, false)
 			return v, v == inlineHotMaxBudget
 		}
@@ -1198,7 +1199,7 @@ func PostProcessCallSites(profile *pgo.Profile) {
 	}
 }
 
-func analyzeFuncProps(fn *ir.Func, p *pgo.Profile) {
+func analyzeFuncProps(fn *ir.Func, p *pgoir.Profile) {
 	canInline := func(fn *ir.Func) { CanInline(fn, p) }
 	budgetForFunc := func(fn *ir.Func) int32 {
 		return inlineBudget(fn, p, true, false)
diff --git a/src/cmd/compile/internal/pgo/irgraph.go b/src/cmd/compile/internal/pgo/irgraph.go
index 814c40f172..418066f8ff 100644
--- a/src/cmd/compile/internal/pgo/irgraph.go
+++ b/src/cmd/compile/internal/pgo/irgraph.go
@@ -46,14 +46,9 @@ import (
 	"cmd/compile/internal/ir"
 	"cmd/compile/internal/typecheck"
 	"cmd/compile/internal/types"
-	"errors"
+	"cmd/internal/pgo"
 	"fmt"
-	"internal/profile"
-	"io"
 	"os"
-	"sort"
-	"strconv"
-	"strings"
 )
 
 // IRGraph is a call graph with nodes pointing to IRs of functions and edges
@@ -82,7 +77,7 @@ type IRNode struct {
 
 	// Set of out-edges in the callgraph. The map uniquely identifies each
 	// edge based on the callsite and callee, for fast lookup.
-	OutEdges map[NamedCallEdge]*IREdge
+	OutEdges map[pgo.NamedCallEdge]*IREdge
 }
 
 // Name returns the symbol name of this function.
@@ -102,23 +97,6 @@ type IREdge struct {
 	CallSiteOffset int // Line offset from function start line.
 }
 
-// NamedCallEdge identifies a call edge by linker symbol names and call site
-// offset.
-type NamedCallEdge struct {
-	CallerName     string
-	CalleeName     string
-	CallSiteOffset int // Line offset from function start line.
-}
-
-// NamedEdgeMap contains all unique call edges in the profile and their
-// edge weight.
-type NamedEdgeMap struct {
-	Weight map[NamedCallEdge]int64
-
-	// ByWeight lists all keys in Weight, sorted by edge weight.
-	ByWeight []NamedCallEdge
-}
-
 // CallSiteInfo captures call-site information and its caller/callee.
 type CallSiteInfo struct {
 	LineOffset int // Line offset from function start line.
@@ -129,33 +107,14 @@ type CallSiteInfo struct {
 // Profile contains the processed PGO profile and weighted call graph used for
 // PGO optimizations.
 type Profile struct {
-	// Aggregated edge weights across the profile. This helps us determine
-	// the percentage threshold for hot/cold partitioning.
-	TotalWeight int64
-
-	// NamedEdgeMap contains all unique call edges in the profile and their
-	// edge weight.
-	NamedEdgeMap NamedEdgeMap
+	// Profile is the base data from the raw profile, without IR attribution.
+	*pgo.Profile
 
 	// WeightedCG represents the IRGraph built from profile, which we will
 	// update as part of inlining.
 	WeightedCG *IRGraph
 }
 
-var wantHdr = "GO PREPROFILE V1\n"
-
-func isPreProfileFile(r *bufio.Reader) (bool, error) {
-	hdr, err := r.Peek(len(wantHdr))
-	if err == io.EOF {
-		// Empty file.
-		return false, nil
-	} else if err != nil {
-		return false, fmt.Errorf("error reading profile header: %w", err)
-	}
-
-	return string(hdr) == wantHdr, nil
-}
-
 // New generates a profile-graph from the profile or pre-processed profile.
 func New(profileFile string) (*Profile, error) {
 	f, err := os.Open(profileFile)
@@ -163,240 +122,42 @@ func New(profileFile string) (*Profile, error) {
 		return nil, fmt.Errorf("error opening profile: %w", err)
 	}
 	defer f.Close()
-
 	r := bufio.NewReader(f)
 
-	isPreProf, err := isPreProfileFile(r)
+	isSerialized, err := pgo.IsSerialized(r)
 	if err != nil {
 		return nil, fmt.Errorf("error processing profile header: %w", err)
 	}
 
-	if isPreProf {
-		profile, err := processPreprof(r)
+	var base *pgo.Profile
+	if isSerialized {
+		base, err = pgo.FromSerialized(r)
 		if err != nil {
-			return nil, fmt.Errorf("error processing preprocessed PGO profile: %w", err)
+			return nil, fmt.Errorf("error processing serialized PGO profile: %w", err)
 		}
-		return profile, nil
-	}
-
-	profile, err := processProto(r)
-	if err != nil {
-		return nil, fmt.Errorf("error processing pprof PGO profile: %w", err)
-	}
-	return profile, nil
-
-}
-
-// processProto generates a profile-graph from the profile.
-func processProto(r io.Reader) (*Profile, error) {
-	p, err := profile.Parse(r)
-	if errors.Is(err, profile.ErrNoData) {
-		// Treat a completely empty file the same as a profile with no
-		// samples: nothing to do.
-		return nil, nil
-	} else if err != nil {
-		return nil, fmt.Errorf("error parsing profile: %w", err)
-	}
-
-	if len(p.Sample) == 0 {
-		// We accept empty profiles, but there is nothing to do.
-		return nil, nil
-	}
-
-	valueIndex := -1
-	for i, s := range p.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
+	} else {
+		base, err = pgo.FromPProf(r)
+		if err != nil {
+			return nil, fmt.Errorf("error processing pprof PGO profile: %w", err)
 		}
 	}
 
-	if valueIndex == -1 {
-		return nil, fmt.Errorf(`profile does not contain a sample index with value/type "samples/count" or cpu/nanoseconds"`)
-	}
-
-	g := profile.NewGraph(p, &profile.Options{
-		SampleValue: func(v []int64) int64 { return v[valueIndex] },
-	})
-
-	namedEdgeMap, totalWeight, err := createNamedEdgeMap(g)
-	if err != nil {
-		return nil, err
-	}
-
-	if totalWeight == 0 {
+	if base.TotalWeight == 0 {
 		return nil, nil // accept but ignore profile with no samples.
 	}
 
 	// Create package-level call graph with weights from profile and IR.
-	wg := createIRGraph(namedEdgeMap)
+	wg := createIRGraph(base.NamedEdgeMap)
 
 	return &Profile{
-		TotalWeight:  totalWeight,
-		NamedEdgeMap: namedEdgeMap,
-		WeightedCG:   wg,
+		Profile:    base,
+		WeightedCG: wg,
 	}, nil
 }
 
-// processPreprof generates a profile-graph from the pre-processed profile.
-func processPreprof(r io.Reader) (*Profile, error) {
-	namedEdgeMap, totalWeight, err := createNamedEdgeMapFromPreprocess(r)
-	if err != nil {
-		return nil, err
-	}
-
-	if totalWeight == 0 {
-		return nil, nil // accept but ignore profile with no samples.
-	}
-
-	// Create package-level call graph with weights from profile and IR.
-	wg := createIRGraph(namedEdgeMap)
-
-	return &Profile{
-		TotalWeight:  totalWeight,
-		NamedEdgeMap: namedEdgeMap,
-		WeightedCG:   wg,
-	}, nil
-}
-
-func postProcessNamedEdgeMap(weight map[NamedCallEdge]int64, weightVal int64) (edgeMap NamedEdgeMap, totalWeight int64, err error) {
-	if weightVal == 0 {
-		return NamedEdgeMap{}, 0, nil // accept but ignore profile with no samples.
-	}
-	byWeight := make([]NamedCallEdge, 0, len(weight))
-	for namedEdge := range weight {
-		byWeight = append(byWeight, namedEdge)
-	}
-	sort.Slice(byWeight, func(i, j int) bool {
-		ei, ej := byWeight[i], byWeight[j]
-		if wi, wj := weight[ei], weight[ej]; wi != wj {
-			return wi > wj // want larger weight first
-		}
-		// same weight, order by name/line number
-		if ei.CallerName != ej.CallerName {
-			return ei.CallerName < ej.CallerName
-		}
-		if ei.CalleeName != ej.CalleeName {
-			return ei.CalleeName < ej.CalleeName
-		}
-		return ei.CallSiteOffset < ej.CallSiteOffset
-	})
-
-	edgeMap = NamedEdgeMap{
-		Weight:   weight,
-		ByWeight: byWeight,
-	}
-
-	totalWeight = weightVal
-
-	return edgeMap, totalWeight, nil
-}
-
-// restore NodeMap information from a preprocessed profile.
-// The reader can refer to the format of preprocessed profile in cmd/preprofile/main.go.
-func createNamedEdgeMapFromPreprocess(r io.Reader) (edgeMap NamedEdgeMap, totalWeight int64, err error) {
-	fileScanner := bufio.NewScanner(r)
-	fileScanner.Split(bufio.ScanLines)
-	weight := make(map[NamedCallEdge]int64)
-
-	if !fileScanner.Scan() {
-		if err := fileScanner.Err(); err != nil {
-			return NamedEdgeMap{}, 0, fmt.Errorf("error reading preprocessed profile: %w", err)
-		}
-		return NamedEdgeMap{}, 0, fmt.Errorf("preprocessed profile missing header")
-	}
-	if gotHdr := fileScanner.Text() + "\n"; gotHdr != wantHdr {
-		return NamedEdgeMap{}, 0, fmt.Errorf("preprocessed profile malformed header; got %q want %q", gotHdr, wantHdr)
-	}
-
-	for fileScanner.Scan() {
-		readStr := fileScanner.Text()
-
-		callerName := readStr
-
-		if !fileScanner.Scan() {
-			if err := fileScanner.Err(); err != nil {
-				return NamedEdgeMap{}, 0, fmt.Errorf("error reading preprocessed profile: %w", err)
-			}
-			return NamedEdgeMap{}, 0, fmt.Errorf("preprocessed profile entry missing callee")
-		}
-		calleeName := fileScanner.Text()
-
-		if !fileScanner.Scan() {
-			if err := fileScanner.Err(); err != nil {
-				return NamedEdgeMap{}, 0, fmt.Errorf("error reading preprocessed profile: %w", err)
-			}
-			return NamedEdgeMap{}, 0, fmt.Errorf("preprocessed profile entry missing weight")
-		}
-		readStr = fileScanner.Text()
-
-		split := strings.Split(readStr, " ")
-
-		if len(split) != 2 {
-			return NamedEdgeMap{}, 0, fmt.Errorf("preprocessed profile entry got %v want 2 fields", split)
-		}
-
-		co, _ := strconv.Atoi(split[0])
-
-		namedEdge := NamedCallEdge{
-			CallerName:     callerName,
-			CalleeName:     calleeName,
-			CallSiteOffset: co,
-		}
-
-		EWeight, _ := strconv.ParseInt(split[1], 10, 64)
-
-		weight[namedEdge] += EWeight
-		totalWeight += EWeight
-	}
-
-	return postProcessNamedEdgeMap(weight, totalWeight)
-
-}
-
-// createNamedEdgeMap builds a map of callsite-callee edge weights from the
-// profile-graph.
-//
-// Caller should ignore the profile if totalWeight == 0.
-func createNamedEdgeMap(g *profile.Graph) (edgeMap NamedEdgeMap, totalWeight int64, err error) {
-	seenStartLine := false
-
-	// Process graph and build various node and edge maps which will
-	// be consumed by AST walk.
-	weight := make(map[NamedCallEdge]int64)
-	for _, n := range g.Nodes {
-		seenStartLine = seenStartLine || n.Info.StartLine != 0
-
-		canonicalName := n.Info.Name
-		// Create the key to the nodeMapKey.
-		namedEdge := NamedCallEdge{
-			CallerName:     canonicalName,
-			CallSiteOffset: n.Info.Lineno - n.Info.StartLine,
-		}
-
-		for _, e := range n.Out {
-			totalWeight += e.WeightValue()
-			namedEdge.CalleeName = e.Dest.Info.Name
-			// Create new entry or increment existing entry.
-			weight[namedEdge] += e.WeightValue()
-		}
-	}
-
-	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 NamedEdgeMap{}, 0, fmt.Errorf("profile missing Function.start_line data (Go version of profiled application too old? Go 1.20+ automatically adds this to profiles)")
-	}
-	return postProcessNamedEdgeMap(weight, totalWeight)
-}
-
 // initializeIRGraph builds the IRGraph by visiting all the ir.Func in decl list
 // of a package.
-func createIRGraph(namedEdgeMap NamedEdgeMap) *IRGraph {
+func createIRGraph(namedEdgeMap pgo.NamedEdgeMap) *IRGraph {
 	g := &IRGraph{
 		IRNodes: make(map[string]*IRNode),
 	}
@@ -425,7 +186,7 @@ func createIRGraph(namedEdgeMap NamedEdgeMap) *IRGraph {
 
 // visitIR traverses the body of each ir.Func adds edges to g from ir.Func to
 // any called function in the body.
-func visitIR(fn *ir.Func, namedEdgeMap NamedEdgeMap, g *IRGraph) {
+func visitIR(fn *ir.Func, namedEdgeMap pgo.NamedEdgeMap, g *IRGraph) {
 	name := ir.LinkFuncName(fn)
 	node, ok := g.IRNodes[name]
 	if !ok {
@@ -442,7 +203,7 @@ func visitIR(fn *ir.Func, namedEdgeMap NamedEdgeMap, g *IRGraph) {
 // 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 createIRGraphEdge(fn *ir.Func, callernode *IRNode, name string, namedEdgeMap NamedEdgeMap, g *IRGraph) {
+func createIRGraphEdge(fn *ir.Func, callernode *IRNode, name string, namedEdgeMap pgo.NamedEdgeMap, g *IRGraph) {
 	ir.VisitList(fn.Body, func(n ir.Node) {
 		switch n.Op() {
 		case ir.OCALLFUNC:
@@ -471,7 +232,7 @@ func NodeLineOffset(n ir.Node, fn *ir.Func) int {
 
 // addIREdge adds an edge between caller and new node that points to `callee`
 // based on the profile-graph and NodeMap.
-func addIREdge(callerNode *IRNode, callerName string, call ir.Node, callee *ir.Func, namedEdgeMap NamedEdgeMap, g *IRGraph) {
+func addIREdge(callerNode *IRNode, callerName string, call ir.Node, callee *ir.Func, namedEdgeMap pgo.NamedEdgeMap, g *IRGraph) {
 	calleeName := ir.LinkFuncName(callee)
 	calleeNode, ok := g.IRNodes[calleeName]
 	if !ok {
@@ -481,7 +242,7 @@ func addIREdge(callerNode *IRNode, callerName string, call ir.Node, callee *ir.F
 		g.IRNodes[calleeName] = calleeNode
 	}
 
-	namedEdge := NamedCallEdge{
+	namedEdge := pgo.NamedCallEdge{
 		CallerName:     callerName,
 		CalleeName:     calleeName,
 		CallSiteOffset: NodeLineOffset(call, callerNode.AST),
@@ -496,7 +257,7 @@ func addIREdge(callerNode *IRNode, callerName string, call ir.Node, callee *ir.F
 	}
 
 	if callerNode.OutEdges == nil {
-		callerNode.OutEdges = make(map[NamedCallEdge]*IREdge)
+		callerNode.OutEdges = make(map[pgo.NamedCallEdge]*IREdge)
 	}
 	callerNode.OutEdges[namedEdge] = edge
 }
@@ -519,7 +280,7 @@ var LookupFunc = func(fullName string) (*ir.Func, error) {
 // 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 addIndirectEdges(g *IRGraph, namedEdgeMap NamedEdgeMap) {
+func addIndirectEdges(g *IRGraph, namedEdgeMap pgo.NamedEdgeMap) {
 	// 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
@@ -616,17 +377,12 @@ func addIndirectEdges(g *IRGraph, namedEdgeMap NamedEdgeMap) {
 		}
 
 		if callerNode.OutEdges == nil {
-			callerNode.OutEdges = make(map[NamedCallEdge]*IREdge)
+			callerNode.OutEdges = make(map[pgo.NamedCallEdge]*IREdge)
 		}
 		callerNode.OutEdges[key] = edge
 	}
 }
 
-// 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")
@@ -688,7 +444,7 @@ func (p *Profile) PrintWeightedCallGraphDOT(edgeThreshold float64) {
 						style = "dashed"
 					}
 					color := "black"
-					edgepercent := WeightInPercentage(e.Weight, p.TotalWeight)
+					edgepercent := pgo.WeightInPercentage(e.Weight, p.TotalWeight)
 					if edgepercent > edgeThreshold {
 						color = "red"
 					}
diff --git a/src/cmd/compile/internal/test/testdata/pgo/devirtualize/devirt.pprof.node_map b/src/cmd/compile/internal/test/testdata/pgo/devirtualize/devirt.pprof.node_map
index c55f990e84..1bb6a80f06 100644
--- a/src/cmd/compile/internal/test/testdata/pgo/devirtualize/devirt.pprof.node_map
+++ b/src/cmd/compile/internal/test/testdata/pgo/devirtualize/devirt.pprof.node_map
@@ -1,52 +1,52 @@
 GO PREPROFILE V1
-example.com/pgo/devirtualize.ExerciseFuncClosure
-example.com/pgo/devirtualize/mult%2epkg.MultClosure.func1
-18 93
-example.com/pgo/devirtualize.ExerciseIface
-example.com/pgo/devirtualize/mult%2epkg.NegMult.Multiply
-49 4
 example.com/pgo/devirtualize.ExerciseFuncConcrete
 example.com/pgo/devirtualize.AddFn
 48 103
 example.com/pgo/devirtualize.ExerciseFuncField
-example.com/pgo/devirtualize/mult%2epkg.NegMultFn
-23 8
+example.com/pgo/devirtualize.AddFn
+23 101
 example.com/pgo/devirtualize.ExerciseFuncField
 example.com/pgo/devirtualize/mult%2epkg.MultFn
 23 94
-example.com/pgo/devirtualize.ExerciseIface
-example.com/pgo/devirtualize/mult%2epkg.Mult.Multiply
-49 40
+example.com/pgo/devirtualize.ExerciseFuncClosure
+example.com/pgo/devirtualize/mult%2epkg.MultClosure.func1
+18 93
+example.com/pgo/devirtualize.ExerciseFuncClosure
+example.com/pgo/devirtualize.Add.Add
+18 92
+example.com/pgo/devirtualize.ExerciseFuncConcrete
+example.com/pgo/devirtualize/mult%2epkg.MultFn
+48 91
 example.com/pgo/devirtualize.ExerciseIface
 example.com/pgo/devirtualize.Add.Add
 49 55
-example.com/pgo/devirtualize.ExerciseFuncConcrete
-example.com/pgo/devirtualize/mult%2epkg.NegMultFn
-48 8
+example.com/pgo/devirtualize.ExerciseIface
+example.com/pgo/devirtualize/mult%2epkg.Mult.Multiply
+49 40
+example.com/pgo/devirtualize.ExerciseFuncClosure
+example.com/pgo/devirtualize.Sub.Add
+18 14
+example.com/pgo/devirtualize.ExerciseFuncField
+example.com/pgo/devirtualize.SubFn
+23 12
 example.com/pgo/devirtualize.ExerciseFuncClosure
 example.com/pgo/devirtualize/mult%2epkg.NegMultClosure.func1
 18 10
+example.com/pgo/devirtualize.ExerciseFuncConcrete
+example.com/pgo/devirtualize/mult%2epkg.NegMultFn
+48 8
+example.com/pgo/devirtualize.ExerciseFuncField
+example.com/pgo/devirtualize/mult%2epkg.NegMultFn
+23 8
 example.com/pgo/devirtualize.ExerciseIface
 example.com/pgo/devirtualize.Sub.Add
 49 7
-example.com/pgo/devirtualize.ExerciseFuncField
-example.com/pgo/devirtualize.AddFn
-23 101
-example.com/pgo/devirtualize.ExerciseFuncField
+example.com/pgo/devirtualize.ExerciseFuncConcrete
 example.com/pgo/devirtualize.SubFn
-23 12
+48 5
+example.com/pgo/devirtualize.ExerciseIface
+example.com/pgo/devirtualize/mult%2epkg.NegMult.Multiply
+49 4
 example.com/pgo/devirtualize.BenchmarkDevirtFuncConcrete
 example.com/pgo/devirtualize.ExerciseFuncConcrete
 1 2
-example.com/pgo/devirtualize.ExerciseFuncConcrete
-example.com/pgo/devirtualize/mult%2epkg.MultFn
-48 91
-example.com/pgo/devirtualize.ExerciseFuncConcrete
-example.com/pgo/devirtualize.SubFn
-48 5
-example.com/pgo/devirtualize.ExerciseFuncClosure
-example.com/pgo/devirtualize.Add.Add
-18 92
-example.com/pgo/devirtualize.ExerciseFuncClosure
-example.com/pgo/devirtualize.Sub.Add
-18 14
diff --git a/src/cmd/compile/internal/test/testdata/pgo/inline/inline_hot.pprof.node_map b/src/cmd/compile/internal/test/testdata/pgo/inline/inline_hot.pprof.node_map
index 6e5f937a50..98a05dd12a 100644
--- a/src/cmd/compile/internal/test/testdata/pgo/inline/inline_hot.pprof.node_map
+++ b/src/cmd/compile/internal/test/testdata/pgo/inline/inline_hot.pprof.node_map
@@ -1,13 +1,13 @@
 GO PREPROFILE V1
-example.com/pgo/inline.benchmarkB
 example.com/pgo/inline.A
-18 1
+example.com/pgo/inline.(*BS).NS
+7 129
 example.com/pgo/inline.(*BS).NS
 example.com/pgo/inline.T
 8 3
 example.com/pgo/inline.(*BS).NS
 example.com/pgo/inline.T
 13 2
+example.com/pgo/inline.benchmarkB
 example.com/pgo/inline.A
-example.com/pgo/inline.(*BS).NS
-7 129
+18 1
diff --git a/src/cmd/dist/buildtool.go b/src/cmd/dist/buildtool.go
index 3232896f26..8949949bd2 100644
--- a/src/cmd/dist/buildtool.go
+++ b/src/cmd/dist/buildtool.go
@@ -47,6 +47,7 @@ var bootstrapDirs = []string{
 	"cmd/internal/notsha256",
 	"cmd/internal/obj/...",
 	"cmd/internal/objabi",
+	"cmd/internal/pgo",
 	"cmd/internal/pkgpath",
 	"cmd/internal/quoted",
 	"cmd/internal/src",
@@ -316,7 +317,7 @@ func bootstrapFixImports(srcFile string) string {
 			continue
 		}
 		if strings.HasPrefix(line, `import "`) || strings.HasPrefix(line, `import . "`) ||
-			inBlock && (strings.HasPrefix(line, "\t\"") || strings.HasPrefix(line, "\t. \"") || strings.HasPrefix(line, "\texec \"") || strings.HasPrefix(line, "\trtabi \"")) {
+			inBlock && (strings.HasPrefix(line, "\t\"") || strings.HasPrefix(line, "\t. \"") || strings.HasPrefix(line, "\texec \"") || strings.HasPrefix(line, "\trtabi \"") || strings.HasPrefix(line, "\tpgoir \"")) {
 			line = strings.Replace(line, `"cmd/`, `"bootstrap/cmd/`, -1)
 			for _, dir := range bootstrapDirs {
 				if strings.HasPrefix(dir, "cmd/") {
diff --git a/src/cmd/internal/pgo/deserialize.go b/src/cmd/internal/pgo/deserialize.go
new file mode 100644
index 0000000000..4b075b8daf
--- /dev/null
+++ b/src/cmd/internal/pgo/deserialize.go
@@ -0,0 +1,102 @@
+// Copyright 2024 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.
+
+package pgo
+
+import (
+	"bufio"
+	"fmt"
+	"io"
+	"strings"
+	"strconv"
+)
+
+// IsSerialized returns true if r is a serialized Profile.
+//
+// IsSerialized only peeks at r, so seeking back after calling is not
+// necessary.
+func IsSerialized(r *bufio.Reader) (bool, error) {
+	hdr, err := r.Peek(len(serializationHeader))
+	if err == io.EOF {
+		// Empty file.
+		return false, nil
+	} else if err != nil {
+		return false, fmt.Errorf("error reading profile header: %w", err)
+	}
+
+	return string(hdr) == serializationHeader, nil
+}
+
+// FromSerialized parses a profile from serialization output of Profile.WriteTo.
+func FromSerialized(r io.Reader) (*Profile, error) {
+	d := emptyProfile()
+
+	scanner := bufio.NewScanner(r)
+	scanner.Split(bufio.ScanLines)
+
+	if !scanner.Scan() {
+		if err := scanner.Err(); err != nil {
+			return nil, fmt.Errorf("error reading preprocessed profile: %w", err)
+		}
+		return nil, fmt.Errorf("preprocessed profile missing header")
+	}
+	if gotHdr := scanner.Text() + "\n"; gotHdr != serializationHeader {
+		return nil, fmt.Errorf("preprocessed profile malformed header; got %q want %q", gotHdr, serializationHeader)
+	}
+
+	for scanner.Scan() {
+		readStr := scanner.Text()
+
+		callerName := readStr
+
+		if !scanner.Scan() {
+			if err := scanner.Err(); err != nil {
+				return nil, fmt.Errorf("error reading preprocessed profile: %w", err)
+			}
+			return nil, fmt.Errorf("preprocessed profile entry missing callee")
+		}
+		calleeName := scanner.Text()
+
+		if !scanner.Scan() {
+			if err := scanner.Err(); err != nil {
+				return nil, fmt.Errorf("error reading preprocessed profile: %w", err)
+			}
+			return nil, fmt.Errorf("preprocessed profile entry missing weight")
+		}
+		readStr = scanner.Text()
+
+		split := strings.Split(readStr, " ")
+
+		if len(split) != 2 {
+			return nil, fmt.Errorf("preprocessed profile entry got %v want 2 fields", split)
+		}
+
+		co, err := strconv.Atoi(split[0])
+		if err != nil {
+			return nil, fmt.Errorf("preprocessed profile error processing call line: %w", err)
+		}
+
+		edge := NamedCallEdge{
+			CallerName:     callerName,
+			CalleeName:     calleeName,
+			CallSiteOffset: co,
+		}
+
+		weight, err := strconv.ParseInt(split[1], 10, 64)
+		if err != nil {
+			return nil, fmt.Errorf("preprocessed profile error processing call weight: %w", err)
+		}
+
+		if _, ok := d.NamedEdgeMap.Weight[edge]; ok {
+			return nil, fmt.Errorf("preprocessed profile contains duplicate edge %+v", edge)
+		}
+
+		d.NamedEdgeMap.ByWeight = append(d.NamedEdgeMap.ByWeight, edge) // N.B. serialization is ordered.
+		d.NamedEdgeMap.Weight[edge] += weight
+		d.TotalWeight += weight
+	}
+
+	return d, nil
+
+}
diff --git a/src/cmd/internal/pgo/pgo.go b/src/cmd/internal/pgo/pgo.go
new file mode 100644
index 0000000000..1d2cb880f7
--- /dev/null
+++ b/src/cmd/internal/pgo/pgo.go
@@ -0,0 +1,55 @@
+// Copyright 2024 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.
+
+// Package pgo contains the compiler-agnostic portions of PGO profile handling.
+// Notably, parsing pprof profiles and serializing/deserializing from a custom
+// intermediate representation.
+package pgo
+
+// Profile contains the processed data from the PGO profile.
+type Profile struct {
+	// TotalWeight is the aggregated edge weights across the profile. This
+	// helps us determine the percentage threshold for hot/cold
+	// partitioning.
+	TotalWeight int64
+
+	// NamedEdgeMap contains all unique call edges in the profile and their
+	// edge weight.
+	NamedEdgeMap NamedEdgeMap
+}
+
+// NamedCallEdge identifies a call edge by linker symbol names and call site
+// offset.
+type NamedCallEdge struct {
+	CallerName     string
+	CalleeName     string
+	CallSiteOffset int // Line offset from function start line.
+}
+
+// NamedEdgeMap contains all unique call edges in the profile and their
+// edge weight.
+type NamedEdgeMap struct {
+	Weight map[NamedCallEdge]int64
+
+	// ByWeight lists all keys in Weight, sorted by edge weight from
+	// highest to lowest.
+	ByWeight []NamedCallEdge
+}
+
+func emptyProfile() *Profile {
+	// Initialize empty maps/slices for easier use without a requiring a
+	// nil check.
+	return &Profile{
+		NamedEdgeMap: NamedEdgeMap{
+			ByWeight: make([]NamedCallEdge, 0),
+			Weight:   make(map[NamedCallEdge]int64),
+		},
+	}
+}
+
+// WeightInPercentage converts profile weights to a percentage.
+func WeightInPercentage(value int64, total int64) float64 {
+	return (float64(value) / float64(total)) * 100
+}
+
diff --git a/src/cmd/internal/pgo/pprof.go b/src/cmd/internal/pgo/pprof.go
new file mode 100644
index 0000000000..5e61a11141
--- /dev/null
+++ b/src/cmd/internal/pgo/pprof.go
@@ -0,0 +1,140 @@
+// Copyright 2024 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.
+
+// Package pgo contains the compiler-agnostic portions of PGO profile handling.
+// Notably, parsing pprof profiles and serializing/deserializing from a custom
+// intermediate representation.
+package pgo
+
+import (
+	"errors"
+	"fmt"
+	"internal/profile"
+	"io"
+	"sort"
+)
+
+// FromPProf parses Profile from a pprof profile.
+func FromPProf(r io.Reader) (*Profile, error) {
+	p, err := profile.Parse(r)
+	if errors.Is(err, profile.ErrNoData) {
+		// Treat a completely empty file the same as a profile with no
+		// samples: nothing to do.
+		return emptyProfile(), nil
+	} else if err != nil {
+		return nil, fmt.Errorf("error parsing profile: %w", err)
+	}
+
+	if len(p.Sample) == 0 {
+		// We accept empty profiles, but there is nothing to do.
+		return emptyProfile(), nil
+	}
+
+	valueIndex := -1
+	for i, s := range p.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 := profile.NewGraph(p, &profile.Options{
+		SampleValue: func(v []int64) int64 { return v[valueIndex] },
+	})
+
+	namedEdgeMap, totalWeight, err := createNamedEdgeMap(g)
+	if err != nil {
+		return nil, err
+	}
+
+	if totalWeight == 0 {
+		return emptyProfile(), nil // accept but ignore profile with no samples.
+	}
+
+	return &Profile{
+		TotalWeight:  totalWeight,
+		NamedEdgeMap: namedEdgeMap,
+	}, nil
+}
+
+// createNamedEdgeMap builds a map of callsite-callee edge weights from the
+// profile-graph.
+//
+// Caller should ignore the profile if totalWeight == 0.
+func createNamedEdgeMap(g *profile.Graph) (edgeMap NamedEdgeMap, totalWeight int64, err error) {
+	seenStartLine := false
+
+	// Process graph and build various node and edge maps which will
+	// be consumed by AST walk.
+	weight := make(map[NamedCallEdge]int64)
+	for _, n := range g.Nodes {
+		seenStartLine = seenStartLine || n.Info.StartLine != 0
+
+		canonicalName := n.Info.Name
+		// Create the key to the nodeMapKey.
+		namedEdge := NamedCallEdge{
+			CallerName:     canonicalName,
+			CallSiteOffset: n.Info.Lineno - n.Info.StartLine,
+		}
+
+		for _, e := range n.Out {
+			totalWeight += e.WeightValue()
+			namedEdge.CalleeName = e.Dest.Info.Name
+			// Create new entry or increment existing entry.
+			weight[namedEdge] += e.WeightValue()
+		}
+	}
+
+	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 NamedEdgeMap{}, 0, fmt.Errorf("profile missing Function.start_line data (Go version of profiled application too old? Go 1.20+ automatically adds this to profiles)")
+	}
+	return postProcessNamedEdgeMap(weight, totalWeight)
+}
+
+func sortByWeight(edges []NamedCallEdge, weight map[NamedCallEdge]int64) {
+	sort.Slice(edges, func(i, j int) bool {
+		ei, ej := edges[i], edges[j]
+		if wi, wj := weight[ei], weight[ej]; wi != wj {
+			return wi > wj // want larger weight first
+		}
+		// same weight, order by name/line number
+		if ei.CallerName != ej.CallerName {
+			return ei.CallerName < ej.CallerName
+		}
+		if ei.CalleeName != ej.CalleeName {
+			return ei.CalleeName < ej.CalleeName
+		}
+		return ei.CallSiteOffset < ej.CallSiteOffset
+	})
+}
+
+func postProcessNamedEdgeMap(weight map[NamedCallEdge]int64, weightVal int64) (edgeMap NamedEdgeMap, totalWeight int64, err error) {
+	if weightVal == 0 {
+		return NamedEdgeMap{}, 0, nil // accept but ignore profile with no samples.
+	}
+	byWeight := make([]NamedCallEdge, 0, len(weight))
+	for namedEdge := range weight {
+		byWeight = append(byWeight, namedEdge)
+	}
+	sortByWeight(byWeight, weight)
+
+	edgeMap = NamedEdgeMap{
+		Weight:   weight,
+		ByWeight: byWeight,
+	}
+
+	totalWeight = weightVal
+
+	return edgeMap, totalWeight, nil
+}
diff --git a/src/cmd/internal/pgo/serialize.go b/src/cmd/internal/pgo/serialize.go
new file mode 100644
index 0000000000..caf67ce485
--- /dev/null
+++ b/src/cmd/internal/pgo/serialize.go
@@ -0,0 +1,79 @@
+// Copyright 2024 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.
+
+package pgo
+
+import (
+	"bufio"
+	"fmt"
+	"io"
+)
+
+// Serialization of a Profile allows go tool preprofile to construct the edge
+// map only once (rather than once per compile process). The compiler processes
+// then parse the pre-processed data directly from the serialized format.
+//
+// The format of the serialized output is as follows.
+//
+//      GO PREPROFILE V1
+//      caller_name
+//      callee_name
+//      "call site offset" "call edge weight"
+//      ...
+//      caller_name
+//      callee_name
+//      "call site offset" "call edge weight"
+//
+// Entries are sorted by "call edge weight", from highest to lowest.
+
+const serializationHeader = "GO PREPROFILE V1\n"
+
+// WriteTo writes a serialized representation of Profile to w.
+//
+// FromSerialized can parse the format back to Profile.
+//
+// WriteTo implements io.WriterTo.Write.
+func (d *Profile) WriteTo(w io.Writer) (int64, error) {
+	bw := bufio.NewWriter(w)
+
+	var written int64
+
+	// Header
+	n, err := bw.WriteString(serializationHeader)
+	written += int64(n)
+	if err != nil {
+		return written, err
+	}
+
+	for _, edge := range d.NamedEdgeMap.ByWeight {
+		weight := d.NamedEdgeMap.Weight[edge]
+
+		n, err = fmt.Fprintln(bw, edge.CallerName)
+		written += int64(n)
+		if err != nil {
+			return written, err
+		}
+
+		n, err = fmt.Fprintln(bw, edge.CalleeName)
+		written += int64(n)
+		if err != nil {
+			return written, err
+		}
+
+		n, err = fmt.Fprintf(bw, "%d %d\n", edge.CallSiteOffset, weight)
+		written += int64(n)
+		if err != nil {
+			return written, err
+		}
+	}
+
+	if err := bw.Flush(); err != nil {
+		return written, err
+	}
+
+	// No need to serialize TotalWeight, it can be trivially recomputed
+	// during parsing.
+
+	return written, nil
+}
diff --git a/src/cmd/internal/pgo/serialize_test.go b/src/cmd/internal/pgo/serialize_test.go
new file mode 100644
index 0000000000..b24163d1e2
--- /dev/null
+++ b/src/cmd/internal/pgo/serialize_test.go
@@ -0,0 +1,190 @@
+// Copyright 2024 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.
+
+package pgo
+
+import (
+	"bytes"
+	"encoding/binary"
+	"fmt"
+	"reflect"
+	"strings"
+	"testing"
+)
+
+// equal returns an error if got and want are not equal.
+func equal(got, want *Profile) error {
+	if got.TotalWeight != want.TotalWeight {
+		return fmt.Errorf("got.TotalWeight %d != want.TotalWeight %d", got.TotalWeight, want.TotalWeight)
+	}
+	if !reflect.DeepEqual(got.NamedEdgeMap.ByWeight, want.NamedEdgeMap.ByWeight) {
+		return fmt.Errorf("got.NamedEdgeMap.ByWeight != want.NamedEdgeMap.ByWeight\ngot = %+v\nwant = %+v", got.NamedEdgeMap.ByWeight, want.NamedEdgeMap.ByWeight)
+	}
+	if !reflect.DeepEqual(got.NamedEdgeMap.Weight, want.NamedEdgeMap.Weight) {
+		return fmt.Errorf("got.NamedEdgeMap.Weight != want.NamedEdgeMap.Weight\ngot = %+v\nwant = %+v", got.NamedEdgeMap.Weight, want.NamedEdgeMap.Weight)
+	}
+
+	return nil
+}
+
+func testRoundTrip(t *testing.T, d *Profile) []byte {
+	var buf bytes.Buffer
+	n, err := d.WriteTo(&buf)
+	if err != nil {
+		t.Fatalf("WriteTo got err %v want nil", err)
+	}
+	if n != int64(buf.Len()) {
+		t.Errorf("WriteTo got n %d want %d", n, int64(buf.Len()))
+	}
+
+	b := buf.Bytes()
+
+	got, err := FromSerialized(&buf)
+	if err != nil {
+		t.Fatalf("processSerialized got err %v want nil", err)
+	}
+	if err := equal(got, d); err != nil {
+		t.Errorf("processSerialized output does not match input: %v", err)
+	}
+
+	return b
+}
+
+func TestEmpty(t *testing.T) {
+	d := emptyProfile()
+	b := testRoundTrip(t, d)
+
+	// Contents should consist of only a header.
+	if string(b) != serializationHeader {
+		t.Errorf("WriteTo got %q want %q", string(b), serializationHeader)
+	}
+}
+
+func TestRoundTrip(t *testing.T) {
+	d := &Profile{
+		TotalWeight: 3,
+		NamedEdgeMap: NamedEdgeMap{
+			ByWeight: []NamedCallEdge{
+				{
+					CallerName: "a",
+					CalleeName: "b",
+					CallSiteOffset: 14,
+				},
+				{
+					CallerName: "c",
+					CalleeName: "d",
+					CallSiteOffset: 15,
+				},
+			},
+			Weight: map[NamedCallEdge]int64{
+				{
+					CallerName: "a",
+					CalleeName: "b",
+					CallSiteOffset: 14,
+				}: 2,
+				{
+					CallerName: "c",
+					CalleeName: "d",
+					CallSiteOffset: 15,
+				}: 1,
+			},
+		},
+	}
+
+	testRoundTrip(t, d)
+}
+
+func constructFuzzProfile(t *testing.T, b []byte) *Profile {
+	// The fuzzer can't construct an arbitrary structure, so instead we
+	// consume bytes from b to act as our edge data.
+	r := bytes.NewReader(b)
+	consumeString := func() (string, bool) {
+		// First byte: how many bytes to read for this string? We only
+		// use a byte to avoid making humongous strings.
+		length, err := r.ReadByte()
+		if err != nil {
+			return "", false
+		}
+		if length == 0 {
+			return "", false
+		}
+
+		b := make([]byte, length)
+		_, err = r.Read(b)
+		if err != nil {
+			return "", false
+		}
+
+		return string(b), true
+	}
+	consumeInt64 := func() (int64, bool) {
+		b := make([]byte, 8)
+		_, err := r.Read(b)
+		if err != nil {
+			return 0, false
+		}
+
+		return int64(binary.LittleEndian.Uint64(b)), true
+	}
+
+	d := emptyProfile()
+
+	for {
+		caller, ok := consumeString()
+		if !ok {
+			break
+		}
+		if strings.ContainsAny(caller, " \r\n") {
+			t.Skip("caller contains space or newline")
+		}
+
+		callee, ok := consumeString()
+		if !ok {
+			break
+		}
+		if strings.ContainsAny(callee, " \r\n") {
+			t.Skip("callee contains space or newline")
+		}
+
+		line, ok := consumeInt64()
+		if !ok {
+			break
+		}
+		weight, ok := consumeInt64()
+		if !ok {
+			break
+		}
+
+		edge := NamedCallEdge{
+			CallerName: caller,
+			CalleeName: callee,
+			CallSiteOffset: int(line),
+		}
+
+		if _, ok := d.NamedEdgeMap.Weight[edge]; ok {
+			t.Skip("duplicate edge")
+		}
+
+		d.NamedEdgeMap.Weight[edge] = weight
+		d.TotalWeight += weight
+	}
+
+	byWeight := make([]NamedCallEdge, 0, len(d.NamedEdgeMap.Weight))
+	for namedEdge := range d.NamedEdgeMap.Weight {
+		byWeight = append(byWeight, namedEdge)
+	}
+	sortByWeight(byWeight, d.NamedEdgeMap.Weight)
+	d.NamedEdgeMap.ByWeight = byWeight
+
+	return d
+}
+
+func FuzzRoundTrip(f *testing.F) {
+	f.Add([]byte("")) // empty profile
+
+	f.Fuzz(func(t *testing.T, b []byte) {
+		d := constructFuzzProfile(t, b)
+		testRoundTrip(t, d)
+	})
+}
diff --git a/src/cmd/internal/pgo/testdata/fuzz/FuzzRoundTrip/12fcf136fcb7463c b/src/cmd/internal/pgo/testdata/fuzz/FuzzRoundTrip/12fcf136fcb7463c
new file mode 100644
index 0000000000..31e3552bdc
--- /dev/null
+++ b/src/cmd/internal/pgo/testdata/fuzz/FuzzRoundTrip/12fcf136fcb7463c
@@ -0,0 +1,2 @@
+go test fuzz v1
+[]byte("\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd00000000\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd\xfd0")
diff --git a/src/cmd/internal/pgo/testdata/fuzz/FuzzRoundTrip/2055d314024c8d6c b/src/cmd/internal/pgo/testdata/fuzz/FuzzRoundTrip/2055d314024c8d6c
new file mode 100644
index 0000000000..b44370f012
--- /dev/null
+++ b/src/cmd/internal/pgo/testdata/fuzz/FuzzRoundTrip/2055d314024c8d6c
@@ -0,0 +1,2 @@
+go test fuzz v1
+[]byte("\x00\x040000000000000")
diff --git a/src/cmd/internal/pgo/testdata/fuzz/FuzzRoundTrip/b615162315f7b72c b/src/cmd/internal/pgo/testdata/fuzz/FuzzRoundTrip/b615162315f7b72c
new file mode 100644
index 0000000000..094fc10fd6
--- /dev/null
+++ b/src/cmd/internal/pgo/testdata/fuzz/FuzzRoundTrip/b615162315f7b72c
@@ -0,0 +1,2 @@
+go test fuzz v1
+[]byte("\b00000000\x01\n000000000")
diff --git a/src/cmd/internal/pgo/testdata/fuzz/FuzzRoundTrip/fdc60117b431bbae b/src/cmd/internal/pgo/testdata/fuzz/FuzzRoundTrip/fdc60117b431bbae
new file mode 100644
index 0000000000..4f9af7b90a
--- /dev/null
+++ b/src/cmd/internal/pgo/testdata/fuzz/FuzzRoundTrip/fdc60117b431bbae
@@ -0,0 +1,2 @@
+go test fuzz v1
+[]byte("\x010\x01\r000000000")
diff --git a/src/cmd/preprofile/main.go b/src/cmd/preprofile/main.go
index 5b7c564081..4cb87f63c8 100644
--- a/src/cmd/preprofile/main.go
+++ b/src/cmd/preprofile/main.go
@@ -2,7 +2,9 @@
 // Use of this source code is governed by a BSD-style
 // license that can be found in the LICENSE file.
 
-// Preprofile handles pprof files.
+// Preprofile creates an intermediate representation of a pprof profile for use
+// during PGO in the compiler. This transformation depends only on the profile
+// itself and is thus wasteful to perform in every invocation of the compiler.
 //
 // Usage:
 //
@@ -14,32 +16,13 @@ package main
 
 import (
 	"bufio"
+	"cmd/internal/pgo"
 	"flag"
 	"fmt"
-	"internal/profile"
 	"log"
 	"os"
-	"strconv"
 )
 
-// The current Go Compiler consumes significantly long compilation time when the PGO
-// is enabled. To optimize the existing flow and reduce build time of multiple Go
-// services, we create a standalone tool, PGO preprocessor, to extract information
-// from collected profiling files and to cache the WeightedCallGraph in one time
-// fashion. By adding the new tool to the Go compiler, it will reduce the time
-// of repeated profiling file parsing and avoid WeightedCallGraph reconstruction
-// in current Go Compiler.
-// The format of the pre-processed output is as follows.
-//
-//      Header
-//      caller_name
-//      callee_name
-//      "call site offset" "call edge weight"
-//      ...
-//      caller_name
-//      callee_name
-//      "call site offset" "call edge weight"
-
 func usage() {
 	fmt.Fprintf(os.Stderr, "usage: go tool preprofile [-v] [-o output] -i input\n\n")
 	flag.PrintDefaults()
@@ -49,109 +32,35 @@ func usage() {
 var (
 	output  = flag.String("o", "", "output file path")
 	input   = flag.String("i", "", "input pprof file path")
-	verbose = flag.Bool("v", false, "enable verbose logging")
 )
 
-type NodeMapKey struct {
-	CallerName     string
-	CalleeName     string
-	CallSiteOffset int // Line offset from function start line.
-}
-
-func preprocess(profileFile string, outputFile string, verbose bool) error {
-	// open the pprof profile file
+func preprocess(profileFile string, outputFile string) error {
 	f, err := os.Open(profileFile)
 	if err != nil {
 		return fmt.Errorf("error opening profile: %w", err)
 	}
 	defer f.Close()
-	p, err := profile.Parse(f)
+
+	r := bufio.NewReader(f)
+	d, err := pgo.FromPProf(r)
 	if err != nil {
 		return fmt.Errorf("error parsing profile: %w", err)
 	}
 
-	if len(p.Sample) == 0 {
-		// We accept empty profiles, but there is nothing to do.
-		//
-		// TODO(prattmic): write an "empty" preprocessed file.
-		return nil
-	}
-
-	valueIndex := -1
-	for i, s := range p.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 fmt.Errorf("failed to find CPU samples count or CPU nanoseconds value-types in profile.")
-	}
-
-	// The processing here is equivalent to cmd/compile/internal/pgo.createNamedEdgeMap.
-	g := profile.NewGraph(p, &profile.Options{
-		SampleValue: func(v []int64) int64 { return v[valueIndex] },
-	})
-
-	TotalEdgeWeight := int64(0)
-
-	NodeMap := make(map[NodeMapKey]int64)
-
-	for _, n := range g.Nodes {
-		canonicalName := n.Info.Name
-		// Create the key to the nodeMapKey.
-		nodeinfo := NodeMapKey{
-			CallerName:     canonicalName,
-			CallSiteOffset: n.Info.Lineno - n.Info.StartLine,
-		}
-
-		if n.Info.StartLine == 0 {
-			if verbose {
-				log.Println("[PGO] warning: " + canonicalName + " relative line number is missing from the profile")
-			}
-		}
-
-		for _, e := range n.Out {
-			TotalEdgeWeight += e.WeightValue()
-			nodeinfo.CalleeName = e.Dest.Info.Name
-			if w, ok := NodeMap[nodeinfo]; ok {
-				w += e.WeightValue()
-			} else {
-				w = e.WeightValue()
-				NodeMap[nodeinfo] = w
-			}
-		}
-	}
-
-	var fNodeMap *os.File
+	var out *os.File
 	if outputFile == "" {
-		fNodeMap = os.Stdout
+		out = os.Stdout
 	} else {
-		fNodeMap, err = os.Create(outputFile)
+		out, err = os.Create(outputFile)
 		if err != nil {
-			return fmt.Errorf("Error creating output file: %w", err)
+			return fmt.Errorf("error creating output file: %w", err)
 		}
-		defer fNodeMap.Close()
+		defer out.Close()
 	}
 
-	w := bufio.NewWriter(fNodeMap)
-	w.WriteString("GO PREPROFILE V1\n")
-	count := 1
-	separator := " "
-	for key, element := range NodeMap {
-		line := key.CallerName + "\n"
-		w.WriteString(line)
-		line = key.CalleeName + "\n"
-		w.WriteString(line)
-		line = strconv.Itoa(key.CallSiteOffset)
-		line = line + separator + strconv.FormatInt(element, 10) + "\n"
-		w.WriteString(line)
-		w.Flush()
-		count += 1
+	w := bufio.NewWriter(out)
+	if _, err := d.WriteTo(w); err != nil {
+		return fmt.Errorf("error writing output file: %w", err)
 	}
 
 	return nil
@@ -168,7 +77,7 @@ func main() {
 		usage()
 	}
 
-	if err := preprocess(*input, *output, *verbose); err != nil {
+	if err := preprocess(*input, *output); err != nil {
 		log.Fatal(err)
 	}
 }