cmd/compile: allow more inlining of functions that construct closures

[gostls13.git] / src / cmd / compile / internal / inline / inl.go

// Copyright 2011 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.
//
// The inlining facility makes 2 passes: first CanInline determines which
// functions are suitable for inlining, and for those that are it
// saves a copy of the body. Then InlineCalls walks each function body to
// expand calls to inlinable functions.
//
// The Debug.l flag controls the aggressiveness. Note that main() swaps level 0 and 1,
// making 1 the default and -l disable. Additional levels (beyond -l) may be buggy and
// are not supported.
//      0: disabled
//      1: 80-nodes leaf functions, oneliners, panic, lazy typechecking (default)
//      2: (unassigned)
//      3: (unassigned)
//      4: allow non-leaf functions
//
// At some point this may get another default and become switch-offable with -N.
//
// The -d typcheckinl flag enables early typechecking of all imported bodies,
// which is useful to flush out bugs.
//
// The Debug.m flag enables diagnostic output.  a single -m is useful for verifying
// which calls get inlined or not, more is for debugging, and may go away at any point.

package inline

import (
        "fmt"
        "go/constant"
        "sort"
        "strconv"

        "cmd/compile/internal/base"
        "cmd/compile/internal/ir"
        "cmd/compile/internal/logopt"
        "cmd/compile/internal/pgo"
        "cmd/compile/internal/typecheck"
        "cmd/compile/internal/types"
        "cmd/internal/obj"
)

// Inlining budget parameters, gathered in one place
const (
        inlineMaxBudget       = 80
        inlineExtraAppendCost = 0
        // default is to inline if there's at most one call. -l=4 overrides this by using 1 instead.
        inlineExtraCallCost  = 57              // 57 was benchmarked to provided most benefit with no bad surprises; see https://github.com/golang/go/issues/19348#issuecomment-439370742
        inlineExtraPanicCost = 1               // do not penalize inlining panics.
        inlineExtraThrowCost = inlineMaxBudget // with current (2018-05/1.11) code, inlining runtime.throw does not help.

        inlineBigFunctionNodes   = 5000 // Functions with this many nodes are considered "big".
        inlineBigFunctionMaxCost = 20   // Max cost of inlinee when inlining into a "big" function.
)

var (
        // List of all hot callee nodes.
        // TODO(prattmic): Make this non-global.
        candHotCalleeMap = make(map[*pgo.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{})

        // List of inlined call sites. CallSiteInfo.Callee is always nil.
        // TODO(prattmic): Make this non-global.
        inlinedCallSites = make(map[pgo.CallSiteInfo]struct{})

        // Threshold in percentage for hot callsite inlining.
        inlineHotCallSiteThresholdPercent float64

        // Threshold in CDF percentage for hot callsite inlining,
        // that is, for a threshold of X the hottest callsites that
        // make up the top X% of total edge weight will be
        // considered hot for inlining candidates.
        inlineCDFHotCallSiteThresholdPercent = float64(99)

        // Budget increased due to hotness.
        inlineHotMaxBudget int32 = 2000
)

// pgoInlinePrologue records the hot callsites from ir-graph.
func pgoInlinePrologue(p *pgo.Profile, decls []ir.Node) {
        if base.Debug.PGOInlineCDFThreshold != "" {
                if s, err := strconv.ParseFloat(base.Debug.PGOInlineCDFThreshold, 64); err == nil && s >= 0 && s <= 100 {
                        inlineCDFHotCallSiteThresholdPercent = s
                } else {
                        base.Fatalf("invalid PGOInlineCDFThreshold, must be between 0 and 100")
                }
        }
        var hotCallsites []pgo.NodeMapKey
        inlineHotCallSiteThresholdPercent, hotCallsites = hotNodesFromCDF(p)
        if base.Debug.PGOInline > 0 {
                fmt.Printf("hot-callsite-thres-from-CDF=%v\n", inlineHotCallSiteThresholdPercent)
        }

        if x := base.Debug.PGOInlineBudget; x != 0 {
                inlineHotMaxBudget = int32(x)
        }

        for _, n := range hotCallsites {
                // mark inlineable callees from hot edges
                if callee := p.WeightedCG.IRNodes[n.CalleeName]; callee != nil {
                        candHotCalleeMap[callee] = struct{}{}
                }
                // mark hot call sites
                if caller := p.WeightedCG.IRNodes[n.CallerName]; caller != nil {
                        csi := pgo.CallSiteInfo{LineOffset: n.CallSiteOffset, Caller: caller.AST}
                        candHotEdgeMap[csi] = struct{}{}
                }
        }

        if base.Debug.PGOInline >= 2 {
                fmt.Printf("hot-cg before inline in dot format:")
                p.PrintWeightedCallGraphDOT(inlineHotCallSiteThresholdPercent)
        }
}

// hotNodesFromCDF computes an edge weight threshold and the list of hot
// nodes that make up the given percentage of the CDF. The threshold, as
// a percent, is the lower bound of weight for nodes to be considered hot
// (currently only used in debug prints) (in case of equal weights,
// comparing with the threshold may not accurately reflect which nodes are
// considiered hot).
func hotNodesFromCDF(p *pgo.Profile) (float64, []pgo.NodeMapKey) {
        nodes := make([]pgo.NodeMapKey, len(p.NodeMap))
        i := 0
        for n := range p.NodeMap {
                nodes[i] = n
                i++
        }
        sort.Slice(nodes, func(i, j int) bool {
                ni, nj := nodes[i], nodes[j]
                if wi, wj := p.NodeMap[ni].EWeight, p.NodeMap[nj].EWeight; wi != wj {
                        return wi > wj // want larger weight first
                }
                // same weight, order by name/line number
                if ni.CallerName != nj.CallerName {
                        return ni.CallerName < nj.CallerName
                }
                if ni.CalleeName != nj.CalleeName {
                        return ni.CalleeName < nj.CalleeName
                }
                return ni.CallSiteOffset < nj.CallSiteOffset
        })
        cum := int64(0)
        for i, n := range nodes {
                w := p.NodeMap[n].EWeight
                cum += w
                if pgo.WeightInPercentage(cum, p.TotalEdgeWeight) > inlineCDFHotCallSiteThresholdPercent {
                        // nodes[:i+1] to include the very last node that makes it to go over the threshold.
                        // (Say, if the CDF threshold is 50% and one hot node takes 60% of weight, we want to
                        // include that node instead of excluding it.)
                        return pgo.WeightInPercentage(w, p.TotalEdgeWeight), nodes[:i+1]
                }
        }
        return 0, nodes
}

// pgoInlineEpilogue updates IRGraph after inlining.
func pgoInlineEpilogue(p *pgo.Profile, decls []ir.Node) {
        if base.Debug.PGOInline >= 2 {
                ir.VisitFuncsBottomUp(decls, func(list []*ir.Func, recursive bool) {
                        for _, f := range list {
                                name := ir.LinkFuncName(f)
                                if n, ok := p.WeightedCG.IRNodes[name]; ok {
                                        p.RedirectEdges(n, inlinedCallSites)
                                }
                        }
                })
                // Print the call-graph after inlining. This is a debugging feature.
                fmt.Printf("hot-cg after inline in dot:")
                p.PrintWeightedCallGraphDOT(inlineHotCallSiteThresholdPercent)
        }
}

// InlinePackage finds functions that can be inlined and clones them before walk expands them.
func InlinePackage(p *pgo.Profile) {
        InlineDecls(p, typecheck.Target.Decls, true)

        // Perform a garbage collection of hidden closures functions that
        // are no longer reachable from top-level functions following
        // inlining. See #59404 and #59638 for more context.
        garbageCollectUnreferencedHiddenClosures()
}

// InlineDecls applies inlining to the given batch of declarations.
func InlineDecls(p *pgo.Profile, decls []ir.Node, doInline bool) {
        if p != nil {
                pgoInlinePrologue(p, decls)
        }

        doCanInline := func(n *ir.Func, recursive bool, numfns int) {
                if !recursive || numfns > 1 {
                        // We allow inlining if there is no
                        // recursion, or the recursion cycle is
                        // across more than one function.
                        CanInline(n, p)
                } else {
                        if base.Flag.LowerM > 1 && n.OClosure == nil {
                                fmt.Printf("%v: cannot inline %v: recursive\n", ir.Line(n), n.Nname)
                        }
                }
        }

        ir.VisitFuncsBottomUp(decls, func(list []*ir.Func, recursive bool) {
                numfns := numNonClosures(list)
                // We visit functions within an SCC in fairly arbitrary order,
                // so by computing inlinability for all functions in the SCC
                // before performing any inlining, the results are less
                // sensitive to the order within the SCC (see #58905 for an
                // example).
                if base.Debug.InlineSCCOnePass == 0 {
                        // Compute inlinability for all functions in the SCC ...
                        for _, n := range list {
                                doCanInline(n, recursive, numfns)
                        }
                        // ... then make a second pass to do inlining of calls.
                        if doInline {
                                for _, n := range list {
                                        InlineCalls(n, p)
                                }
                        }
                } else {
                        // Legacy ordering to make it easier to triage any bugs
                        // or compile time issues that might crop up.
                        for _, n := range list {
                                doCanInline(n, recursive, numfns)
                                if doInline {
                                        InlineCalls(n, p)
                                }
                        }
                }
        })

        if p != nil {
                pgoInlineEpilogue(p, decls)
        }
}

// garbageCollectUnreferencedHiddenClosures makes a pass over all the
// top-level (non-hidden-closure) functions looking for nested closure
// functions that are reachable, then sweeps through the Target.Decls
// list and marks any non-reachable hidden closure function as dead.
// See issues #59404 and #59638 for more context.
func garbageCollectUnreferencedHiddenClosures() {

        liveFuncs := make(map[*ir.Func]bool)

        var markLiveFuncs func(fn *ir.Func)
        markLiveFuncs = func(fn *ir.Func) {
                if liveFuncs[fn] {
                        return
                }
                liveFuncs[fn] = true
                ir.Visit(fn, func(n ir.Node) {
                        if clo, ok := n.(*ir.ClosureExpr); ok {
                                markLiveFuncs(clo.Func)
                        }
                })
        }

        for i := 0; i < len(typecheck.Target.Decls); i++ {
                if fn, ok := typecheck.Target.Decls[i].(*ir.Func); ok {
                        if fn.IsHiddenClosure() {
                                continue
                        }
                        markLiveFuncs(fn)
                }
        }

        for i := 0; i < len(typecheck.Target.Decls); i++ {
                if fn, ok := typecheck.Target.Decls[i].(*ir.Func); ok {
                        if !fn.IsHiddenClosure() {
                                continue
                        }
                        if fn.IsDeadcodeClosure() {
                                continue
                        }
                        if liveFuncs[fn] {
                                continue
                        }
                        fn.SetIsDeadcodeClosure(true)
                        if base.Flag.LowerM > 2 {
                                fmt.Printf("%v: unreferenced closure %v marked as dead\n", ir.Line(fn), fn)
                        }
                        if fn.Inl != nil && fn.LSym == nil {
                                ir.InitLSym(fn, true)
                        }
                }
        }
}

// 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) {
        if fn.Nname == nil {
                base.Fatalf("CanInline no nname %+v", fn)
        }

        var reason string // reason, if any, that the function was not inlined
        if base.Flag.LowerM > 1 || logopt.Enabled() {
                defer func() {
                        if reason != "" {
                                if base.Flag.LowerM > 1 {
                                        fmt.Printf("%v: cannot inline %v: %s\n", ir.Line(fn), fn.Nname, reason)
                                }
                                if logopt.Enabled() {
                                        logopt.LogOpt(fn.Pos(), "cannotInlineFunction", "inline", ir.FuncName(fn), reason)
                                }
                        }
                }()
        }

        // If marked "go:noinline", don't inline
        if fn.Pragma&ir.Noinline != 0 {
                reason = "marked go:noinline"
                return
        }

        // If marked "go:norace" and -race compilation, don't inline.
        if base.Flag.Race && fn.Pragma&ir.Norace != 0 {
                reason = "marked go:norace with -race compilation"
                return
        }

        // If marked "go:nocheckptr" and -d checkptr compilation, don't inline.
        if base.Debug.Checkptr != 0 && fn.Pragma&ir.NoCheckPtr != 0 {
                reason = "marked go:nocheckptr"
                return
        }

        // If marked "go:cgo_unsafe_args", don't inline, since the
        // function makes assumptions about its argument frame layout.
        if fn.Pragma&ir.CgoUnsafeArgs != 0 {
                reason = "marked go:cgo_unsafe_args"
                return
        }

        // If marked as "go:uintptrkeepalive", don't inline, since the
        // keep alive information is lost during inlining.
        //
        // TODO(prattmic): This is handled on calls during escape analysis,
        // which is after inlining. Move prior to inlining so the keep-alive is
        // maintained after inlining.
        if fn.Pragma&ir.UintptrKeepAlive != 0 {
                reason = "marked as having a keep-alive uintptr argument"
                return
        }

        // If marked as "go:uintptrescapes", don't inline, since the
        // escape information is lost during inlining.
        if fn.Pragma&ir.UintptrEscapes != 0 {
                reason = "marked as having an escaping uintptr argument"
                return
        }

        // The nowritebarrierrec checker currently works at function
        // granularity, so inlining yeswritebarrierrec functions can
        // confuse it (#22342). As a workaround, disallow inlining
        // them for now.
        if fn.Pragma&ir.Yeswritebarrierrec != 0 {
                reason = "marked go:yeswritebarrierrec"
                return
        }

        // If fn has no body (is defined outside of Go), cannot inline it.
        if len(fn.Body) == 0 {
                reason = "no function body"
                return
        }

        // If fn is synthetic hash or eq function, cannot inline it.
        // The function is not generated in Unified IR frontend at this moment.
        if ir.IsEqOrHashFunc(fn) {
                reason = "type eq/hash function"
                return
        }

        if fn.Typecheck() == 0 {
                base.Fatalf("CanInline on non-typechecked function %v", fn)
        }

        n := fn.Nname
        if n.Func.InlinabilityChecked() {
                return
        }
        defer n.Func.SetInlinabilityChecked(true)

        cc := int32(inlineExtraCallCost)
        if base.Flag.LowerL == 4 {
                cc = 1 // this appears to yield better performance than 0.
        }

        // Update the budget for profile-guided inlining.
        budget := int32(inlineMaxBudget)
        if profile != nil {
                if n, ok := profile.WeightedCG.IRNodes[ir.LinkFuncName(fn)]; ok {
                        if _, ok := candHotCalleeMap[n]; ok {
                                budget = int32(inlineHotMaxBudget)
                                if base.Debug.PGOInline > 0 {
                                        fmt.Printf("hot-node enabled increased budget=%v for func=%v\n", budget, ir.PkgFuncName(fn))
                                }
                        }
                }
        }

        // At this point in the game the function we're looking at may
        // have "stale" autos, vars that still appear in the Dcl list, but
        // which no longer have any uses in the function body (due to
        // elimination by deadcode). We'd like to exclude these dead vars
        // when creating the "Inline.Dcl" field below; to accomplish this,
        // the hairyVisitor below builds up a map of used/referenced
        // locals, and we use this map to produce a pruned Inline.Dcl
        // list. See issue 25249 for more context.

        visitor := hairyVisitor{
                curFunc:       fn,
                budget:        budget,
                maxBudget:     budget,
                extraCallCost: cc,
                profile:       profile,
        }
        if visitor.tooHairy(fn) {
                reason = visitor.reason
                return
        }

        n.Func.Inl = &ir.Inline{
                Cost: budget - visitor.budget,
                Dcl:  pruneUnusedAutos(n.Defn.(*ir.Func).Dcl, &visitor),
                Body: inlcopylist(fn.Body),

                CanDelayResults: canDelayResults(fn),
        }

        if base.Flag.LowerM > 1 {
                fmt.Printf("%v: can inline %v with cost %d as: %v { %v }\n", ir.Line(fn), n, budget-visitor.budget, fn.Type(), ir.Nodes(n.Func.Inl.Body))
        } else if base.Flag.LowerM != 0 {
                fmt.Printf("%v: can inline %v\n", ir.Line(fn), n)
        }
        if logopt.Enabled() {
                logopt.LogOpt(fn.Pos(), "canInlineFunction", "inline", ir.FuncName(fn), fmt.Sprintf("cost: %d", budget-visitor.budget))
        }
}

// canDelayResults reports whether inlined calls to fn can delay
// declaring the result parameter until the "return" statement.
func canDelayResults(fn *ir.Func) bool {
        // We can delay declaring+initializing result parameters if:
        // (1) there's exactly one "return" statement in the inlined function;
        // (2) it's not an empty return statement (#44355); and
        // (3) the result parameters aren't named.

        nreturns := 0
        ir.VisitList(fn.Body, func(n ir.Node) {
                if n, ok := n.(*ir.ReturnStmt); ok {
                        nreturns++
                        if len(n.Results) == 0 {
                                nreturns++ // empty return statement (case 2)
                        }
                }
        })

        if nreturns != 1 {
                return false // not exactly one return statement (case 1)
        }

        // temporaries for return values.
        for _, param := range fn.Type().Results().FieldSlice() {
                if sym := types.OrigSym(param.Sym); sym != nil && !sym.IsBlank() {
                        return false // found a named result parameter (case 3)
                }
        }

        return true
}

// hairyVisitor visits a function body to determine its inlining
// hairiness and whether or not it can be inlined.
type hairyVisitor struct {
        // This is needed to access the current caller in the doNode function.
        curFunc       *ir.Func
        budget        int32
        maxBudget     int32
        reason        string
        extraCallCost int32
        usedLocals    ir.NameSet
        do            func(ir.Node) bool
        profile       *pgo.Profile
}

func (v *hairyVisitor) tooHairy(fn *ir.Func) bool {
        v.do = v.doNode // cache closure
        if ir.DoChildren(fn, v.do) {
                return true
        }
        if v.budget < 0 {
                v.reason = fmt.Sprintf("function too complex: cost %d exceeds budget %d", v.maxBudget-v.budget, v.maxBudget)
                return true
        }
        return false
}

// doNode visits n and its children, updates the state in v, and returns true if
// n makes the current function too hairy for inlining.
func (v *hairyVisitor) doNode(n ir.Node) bool {
        if n == nil {
                return false
        }
        switch n.Op() {
        // Call is okay if inlinable and we have the budget for the body.
        case ir.OCALLFUNC:
                n := n.(*ir.CallExpr)
                // Functions that call runtime.getcaller{pc,sp} can not be inlined
                // because getcaller{pc,sp} expect a pointer to the caller's first argument.
                //
                // runtime.throw is a "cheap call" like panic in normal code.
                if n.X.Op() == ir.ONAME {
                        name := n.X.(*ir.Name)
                        if name.Class == ir.PFUNC && types.IsRuntimePkg(name.Sym().Pkg) {
                                fn := name.Sym().Name
                                if fn == "getcallerpc" || fn == "getcallersp" {
                                        v.reason = "call to " + fn
                                        return true
                                }
                                if fn == "throw" {
                                        v.budget -= inlineExtraThrowCost
                                        break
                                }
                        }
                        // Special case for coverage counter updates; although
                        // these correspond to real operations, we treat them as
                        // zero cost for the moment. This is due to the existence
                        // of tests that are sensitive to inlining-- if the
                        // insertion of coverage instrumentation happens to tip a
                        // given function over the threshold and move it from
                        // "inlinable" to "not-inlinable", this can cause changes
                        // in allocation behavior, which can then result in test
                        // failures (a good example is the TestAllocations in
                        // crypto/ed25519).
                        if isAtomicCoverageCounterUpdate(n) {
                                return false
                        }
                }
                if n.X.Op() == ir.OMETHEXPR {
                        if meth := ir.MethodExprName(n.X); meth != nil {
                                if fn := meth.Func; fn != nil {
                                        s := fn.Sym()
                                        var cheap bool
                                        if types.IsRuntimePkg(s.Pkg) && s.Name == "heapBits.nextArena" {
                                                // Special case: explicitly allow mid-stack inlining of
                                                // runtime.heapBits.next even though it calls slow-path
                                                // runtime.heapBits.nextArena.
                                                cheap = true
                                        }
                                        // Special case: on architectures that can do unaligned loads,
                                        // explicitly mark encoding/binary methods as cheap,
                                        // because in practice they are, even though our inlining
                                        // budgeting system does not see that. See issue 42958.
                                        if base.Ctxt.Arch.CanMergeLoads && s.Pkg.Path == "encoding/binary" {
                                                switch s.Name {
                                                case "littleEndian.Uint64", "littleEndian.Uint32", "littleEndian.Uint16",
                                                        "bigEndian.Uint64", "bigEndian.Uint32", "bigEndian.Uint16",
                                                        "littleEndian.PutUint64", "littleEndian.PutUint32", "littleEndian.PutUint16",
                                                        "bigEndian.PutUint64", "bigEndian.PutUint32", "bigEndian.PutUint16",
                                                        "littleEndian.AppendUint64", "littleEndian.AppendUint32", "littleEndian.AppendUint16",
                                                        "bigEndian.AppendUint64", "bigEndian.AppendUint32", "bigEndian.AppendUint16":
                                                        cheap = true
                                                }
                                        }
                                        if cheap {
                                                break // treat like any other node, that is, cost of 1
                                        }
                                }
                        }
                }

                // Determine if the callee edge is for an inlinable hot callee or not.
                if v.profile != nil && v.curFunc != nil {
                        if fn := inlCallee(n.X, v.profile); fn != nil && typecheck.HaveInlineBody(fn) {
                                lineOffset := pgo.NodeLineOffset(n, fn)
                                csi := pgo.CallSiteInfo{LineOffset: lineOffset, Caller: v.curFunc}
                                if _, o := candHotEdgeMap[csi]; o {
                                        if base.Debug.PGOInline > 0 {
                                                fmt.Printf("hot-callsite identified at line=%v for func=%v\n", ir.Line(n), ir.PkgFuncName(v.curFunc))
                                        }
                                }
                        }
                }

                if ir.IsIntrinsicCall(n) {
                        // Treat like any other node.
                        break
                }

                if fn := inlCallee(n.X, v.profile); fn != nil && typecheck.HaveInlineBody(fn) {
                        v.budget -= fn.Inl.Cost
                        break
                }

                // Call cost for non-leaf inlining.
                v.budget -= v.extraCallCost

        case ir.OCALLMETH:
                base.FatalfAt(n.Pos(), "OCALLMETH missed by typecheck")

        // Things that are too hairy, irrespective of the budget
        case ir.OCALL, ir.OCALLINTER:
                // Call cost for non-leaf inlining.
                v.budget -= v.extraCallCost

        case ir.OPANIC:
                n := n.(*ir.UnaryExpr)
                if n.X.Op() == ir.OCONVIFACE && n.X.(*ir.ConvExpr).Implicit() {
                        // Hack to keep reflect.flag.mustBe inlinable for TestIntendedInlining.
                        // Before CL 284412, these conversions were introduced later in the
                        // compiler, so they didn't count against inlining budget.
                        v.budget++
                }
                v.budget -= inlineExtraPanicCost

        case ir.ORECOVER:
                // recover matches the argument frame pointer to find
                // the right panic value, so it needs an argument frame.
                v.reason = "call to recover"
                return true

        case ir.OCLOSURE:
                if base.Debug.InlFuncsWithClosures == 0 {
                        v.reason = "not inlining functions with closures"
                        return true
                }

                // TODO(danscales): Maybe make budget proportional to number of closure
                // variables, e.g.:
                //v.budget -= int32(len(n.(*ir.ClosureExpr).Func.ClosureVars) * 3)
                // TODO(austin): However, if we're able to inline this closure into
                // v.curFunc, then we actually pay nothing for the closure captures. We
                // should try to account for that if we're going to account for captures.
                v.budget -= 15

        case ir.OGO,
                ir.ODEFER,
                ir.ODCLTYPE, // can't print yet
                ir.OTAILCALL:
                v.reason = "unhandled op " + n.Op().String()
                return true

        case ir.OAPPEND:
                v.budget -= inlineExtraAppendCost

        case ir.OADDR:
                n := n.(*ir.AddrExpr)
                // Make "&s.f" cost 0 when f's offset is zero.
                if dot, ok := n.X.(*ir.SelectorExpr); ok && (dot.Op() == ir.ODOT || dot.Op() == ir.ODOTPTR) {
                        if _, ok := dot.X.(*ir.Name); ok && dot.Selection.Offset == 0 {
                                v.budget += 2 // undo ir.OADDR+ir.ODOT/ir.ODOTPTR
                        }
                }

        case ir.ODEREF:
                // *(*X)(unsafe.Pointer(&x)) is low-cost
                n := n.(*ir.StarExpr)

                ptr := n.X
                for ptr.Op() == ir.OCONVNOP {
                        ptr = ptr.(*ir.ConvExpr).X
                }
                if ptr.Op() == ir.OADDR {
                        v.budget += 1 // undo half of default cost of ir.ODEREF+ir.OADDR
                }

        case ir.OCONVNOP:
                // This doesn't produce code, but the children might.
                v.budget++ // undo default cost

        case ir.ODCLCONST, ir.OFALL:
                // These nodes don't produce code; omit from inlining budget.
                return false

        case ir.OIF:
                n := n.(*ir.IfStmt)
                if ir.IsConst(n.Cond, constant.Bool) {
                        // This if and the condition cost nothing.
                        if doList(n.Init(), v.do) {
                                return true
                        }
                        if ir.BoolVal(n.Cond) {
                                return doList(n.Body, v.do)
                        } else {
                                return doList(n.Else, v.do)
                        }
                }

        case ir.ONAME:
                n := n.(*ir.Name)
                if n.Class == ir.PAUTO {
                        v.usedLocals.Add(n)
                }

        case ir.OBLOCK:
                // The only OBLOCK we should see at this point is an empty one.
                // In any event, let the visitList(n.List()) below take care of the statements,
                // and don't charge for the OBLOCK itself. The ++ undoes the -- below.
                v.budget++

        case ir.OMETHVALUE, ir.OSLICELIT:
                v.budget-- // Hack for toolstash -cmp.

        case ir.OMETHEXPR:
                v.budget++ // Hack for toolstash -cmp.

        case ir.OAS2:
                n := n.(*ir.AssignListStmt)

                // Unified IR unconditionally rewrites:
                //
                //      a, b = f()
                //
                // into:
                //
                //      DCL tmp1
                //      DCL tmp2
                //      tmp1, tmp2 = f()
                //      a, b = tmp1, tmp2
                //
                // so that it can insert implicit conversions as necessary. To
                // minimize impact to the existing inlining heuristics (in
                // particular, to avoid breaking the existing inlinability regress
                // tests), we need to compensate for this here.
                //
                // See also identical logic in isBigFunc.
                if init := n.Rhs[0].Init(); len(init) == 1 {
                        if _, ok := init[0].(*ir.AssignListStmt); ok {
                                // 4 for each value, because each temporary variable now
                                // appears 3 times (DCL, LHS, RHS), plus an extra DCL node.
                                //
                                // 1 for the extra "tmp1, tmp2 = f()" assignment statement.
                                v.budget += 4*int32(len(n.Lhs)) + 1
                        }
                }

        case ir.OAS:
                // Special case for coverage counter updates and coverage
                // function registrations. Although these correspond to real
                // operations, we treat them as zero cost for the moment. This
                // is primarily due to the existence of tests that are
                // sensitive to inlining-- if the insertion of coverage
                // instrumentation happens to tip a given function over the
                // threshold and move it from "inlinable" to "not-inlinable",
                // this can cause changes in allocation behavior, which can
                // then result in test failures (a good example is the
                // TestAllocations in crypto/ed25519).
                n := n.(*ir.AssignStmt)
                if n.X.Op() == ir.OINDEX && isIndexingCoverageCounter(n.X) {
                        return false
                }
        }

        v.budget--

        // When debugging, don't stop early, to get full cost of inlining this function
        if v.budget < 0 && base.Flag.LowerM < 2 && !logopt.Enabled() {
                v.reason = "too expensive"
                return true
        }

        return ir.DoChildren(n, v.do)
}

func isBigFunc(fn *ir.Func) bool {
        budget := inlineBigFunctionNodes
        return ir.Any(fn, func(n ir.Node) bool {
                // See logic in hairyVisitor.doNode, explaining unified IR's
                // handling of "a, b = f()" assignments.
                if n, ok := n.(*ir.AssignListStmt); ok && n.Op() == ir.OAS2 {
                        if init := n.Rhs[0].Init(); len(init) == 1 {
                                if _, ok := init[0].(*ir.AssignListStmt); ok {
                                        budget += 4*len(n.Lhs) + 1
                                }
                        }
                }

                budget--
                return budget <= 0
        })
}

// inlcopylist (together with inlcopy) recursively copies a list of nodes, except
// that it keeps the same ONAME, OTYPE, and OLITERAL nodes. It is used for copying
// the body and dcls of an inlineable function.
func inlcopylist(ll []ir.Node) []ir.Node {
        s := make([]ir.Node, len(ll))
        for i, n := range ll {
                s[i] = inlcopy(n)
        }
        return s
}

// inlcopy is like DeepCopy(), but does extra work to copy closures.
func inlcopy(n ir.Node) ir.Node {
        var edit func(ir.Node) ir.Node
        edit = func(x ir.Node) ir.Node {
                switch x.Op() {
                case ir.ONAME, ir.OTYPE, ir.OLITERAL, ir.ONIL:
                        return x
                }
                m := ir.Copy(x)
                ir.EditChildren(m, edit)
                if x.Op() == ir.OCLOSURE {
                        x := x.(*ir.ClosureExpr)
                        // Need to save/duplicate x.Func.Nname,
                        // x.Func.Nname.Ntype, x.Func.Dcl, x.Func.ClosureVars, and
                        // x.Func.Body for iexport and local inlining.
                        oldfn := x.Func
                        newfn := ir.NewFunc(oldfn.Pos())
                        m.(*ir.ClosureExpr).Func = newfn
                        newfn.Nname = ir.NewNameAt(oldfn.Nname.Pos(), oldfn.Nname.Sym())
                        // XXX OK to share fn.Type() ??
                        newfn.Nname.SetType(oldfn.Nname.Type())
                        newfn.Body = inlcopylist(oldfn.Body)
                        // Make shallow copy of the Dcl and ClosureVar slices
                        newfn.Dcl = append([]*ir.Name(nil), oldfn.Dcl...)
                        newfn.ClosureVars = append([]*ir.Name(nil), oldfn.ClosureVars...)
                }
                return m
        }
        return edit(n)
}

// InlineCalls/inlnode walks fn's statements and expressions and substitutes any
// calls made to inlineable functions. This is the external entry point.
func InlineCalls(fn *ir.Func, profile *pgo.Profile) {
        savefn := ir.CurFunc
        ir.CurFunc = fn
        bigCaller := isBigFunc(fn)
        if bigCaller && base.Flag.LowerM > 1 {
                fmt.Printf("%v: function %v considered 'big'; reducing max cost of inlinees\n", ir.Line(fn), fn)
        }
        var inlCalls []*ir.InlinedCallExpr
        var edit func(ir.Node) ir.Node
        edit = func(n ir.Node) ir.Node {
                return inlnode(n, bigCaller, &inlCalls, edit, profile)
        }
        ir.EditChildren(fn, edit)

        // If we inlined any calls, we want to recursively visit their
        // bodies for further inlining. However, we need to wait until
        // *after* the original function body has been expanded, or else
        // inlCallee can have false positives (e.g., #54632).
        for len(inlCalls) > 0 {
                call := inlCalls[0]
                inlCalls = inlCalls[1:]
                ir.EditChildren(call, edit)
        }

        ir.CurFunc = savefn
}

// inlnode recurses over the tree to find inlineable calls, which will
// be turned into OINLCALLs by mkinlcall. When the recursion comes
// back up will examine left, right, list, rlist, ninit, ntest, nincr,
// nbody and nelse and use one of the 4 inlconv/glue functions above
// to turn the OINLCALL into an expression, a statement, or patch it
// in to this nodes list or rlist as appropriate.
// NOTE it makes no sense to pass the glue functions down the
// recursion to the level where the OINLCALL gets created because they
// have to edit /this/ n, so you'd have to push that one down as well,
// but then you may as well do it here.  so this is cleaner and
// shorter and less complicated.
// The result of inlnode MUST be assigned back to n, e.g.
//
//      n.Left = inlnode(n.Left)
func inlnode(n ir.Node, bigCaller bool, inlCalls *[]*ir.InlinedCallExpr, edit func(ir.Node) ir.Node, profile *pgo.Profile) ir.Node {
        if n == nil {
                return n
        }

        switch n.Op() {
        case ir.ODEFER, ir.OGO:
                n := n.(*ir.GoDeferStmt)
                switch call := n.Call; call.Op() {
                case ir.OCALLMETH:
                        base.FatalfAt(call.Pos(), "OCALLMETH missed by typecheck")
                case ir.OCALLFUNC:
                        call := call.(*ir.CallExpr)
                        call.NoInline = true
                }
        case ir.OTAILCALL:
                n := n.(*ir.TailCallStmt)
                n.Call.NoInline = true // Not inline a tail call for now. Maybe we could inline it just like RETURN fn(arg)?

        // TODO do them here (or earlier),
        // so escape analysis can avoid more heapmoves.
        case ir.OCLOSURE:
                return n
        case ir.OCALLMETH:
                base.FatalfAt(n.Pos(), "OCALLMETH missed by typecheck")
        case ir.OCALLFUNC:
                n := n.(*ir.CallExpr)
                if n.X.Op() == ir.OMETHEXPR {
                        // Prevent inlining some reflect.Value methods when using checkptr,
                        // even when package reflect was compiled without it (#35073).
                        if meth := ir.MethodExprName(n.X); meth != nil {
                                s := meth.Sym()
                                if base.Debug.Checkptr != 0 && types.IsReflectPkg(s.Pkg) && (s.Name == "Value.UnsafeAddr" || s.Name == "Value.Pointer") {
                                        return n
                                }
                        }
                }
        }

        lno := ir.SetPos(n)

        ir.EditChildren(n, edit)

        // with all the branches out of the way, it is now time to
        // transmogrify this node itself unless inhibited by the
        // switch at the top of this function.
        switch n.Op() {
        case ir.OCALLMETH:
                base.FatalfAt(n.Pos(), "OCALLMETH missed by typecheck")

        case ir.OCALLFUNC:
                call := n.(*ir.CallExpr)
                if call.NoInline {
                        break
                }
                if base.Flag.LowerM > 3 {
                        fmt.Printf("%v:call to func %+v\n", ir.Line(n), call.X)
                }
                if ir.IsIntrinsicCall(call) {
                        break
                }
                if fn := inlCallee(call.X, profile); fn != nil && typecheck.HaveInlineBody(fn) {
                        n = mkinlcall(call, fn, bigCaller, inlCalls)
                }
        }

        base.Pos = lno

        return n
}

// 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(fn ir.Node, profile *pgo.Profile) *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
                CanInline(c, profile)
                return c
        }
        return nil
}

var inlgen int

// SSADumpInline gives the SSA back end a chance to dump the function
// when producing output for debugging the compiler itself.
var SSADumpInline = func(*ir.Func) {}

// InlineCall allows the inliner implementation to be overridden.
// If it returns nil, the function will not be inlined.
var InlineCall = func(call *ir.CallExpr, fn *ir.Func, inlIndex int) *ir.InlinedCallExpr {
        base.Fatalf("inline.InlineCall not overridden")
        panic("unreachable")
}

// inlineCostOK returns true if call n from caller to callee is cheap enough to
// inline. bigCaller indicates that caller is a big function.
//
// If inlineCostOK returns false, it also returns the max cost that the callee
// exceeded.
func inlineCostOK(n *ir.CallExpr, caller, callee *ir.Func, bigCaller bool) (bool, int32) {
        maxCost := int32(inlineMaxBudget)
        if bigCaller {
                // We use this to restrict inlining into very big functions.
                // See issue 26546 and 17566.
                maxCost = inlineBigFunctionMaxCost
        }

        if callee.Inl.Cost <= maxCost {
                // Simple case. Function is already cheap enough.
                return true, 0
        }

        // 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}
        if _, ok := candHotEdgeMap[csi]; !ok {
                // Cold
                return false, maxCost
        }

        // Hot

        if bigCaller {
                if base.Debug.PGOInline > 0 {
                        fmt.Printf("hot-big check disallows inlining for call %s (cost %d) at %v in big function %s\n", ir.PkgFuncName(callee), callee.Inl.Cost, ir.Line(n), ir.PkgFuncName(caller))
                }
                return false, maxCost
        }

        if callee.Inl.Cost > inlineHotMaxBudget {
                return false, inlineHotMaxBudget
        }

        if base.Debug.PGOInline > 0 {
                fmt.Printf("hot-budget check allows inlining for call %s (cost %d) at %v in function %s\n", ir.PkgFuncName(callee), callee.Inl.Cost, ir.Line(n), ir.PkgFuncName(caller))
        }

        return true, 0
}

// If n is a OCALLFUNC node, and fn is an ONAME node for a
// function with an inlinable body, return an OINLCALL node that can replace n.
// The returned node's Ninit has the parameter assignments, the Nbody is the
// inlined function body, and (List, Rlist) contain the (input, output)
// parameters.
// The result of mkinlcall MUST be assigned back to n, e.g.
//
//      n.Left = mkinlcall(n.Left, fn, isddd)
func mkinlcall(n *ir.CallExpr, fn *ir.Func, bigCaller bool, inlCalls *[]*ir.InlinedCallExpr) ir.Node {
        if fn.Inl == nil {
                if logopt.Enabled() {
                        logopt.LogOpt(n.Pos(), "cannotInlineCall", "inline", ir.FuncName(ir.CurFunc),
                                fmt.Sprintf("%s cannot be inlined", ir.PkgFuncName(fn)))
                }
                return n
        }

        if ok, maxCost := inlineCostOK(n, ir.CurFunc, fn, bigCaller); !ok {
                if logopt.Enabled() {
                        logopt.LogOpt(n.Pos(), "cannotInlineCall", "inline", ir.FuncName(ir.CurFunc),
                                fmt.Sprintf("cost %d of %s exceeds max caller cost %d", fn.Inl.Cost, ir.PkgFuncName(fn), maxCost))
                }
                return n
        }

        if fn == ir.CurFunc {
                // Can't recursively inline a function into itself.
                if logopt.Enabled() {
                        logopt.LogOpt(n.Pos(), "cannotInlineCall", "inline", fmt.Sprintf("recursive call to %s", ir.FuncName(ir.CurFunc)))
                }
                return n
        }

        if base.Flag.Cfg.Instrumenting && types.IsRuntimePkg(fn.Sym().Pkg) {
                // Runtime package must not be instrumented.
                // Instrument skips runtime package. However, some runtime code can be
                // inlined into other packages and instrumented there. To avoid this,
                // we disable inlining of runtime functions when instrumenting.
                // The example that we observed is inlining of LockOSThread,
                // which lead to false race reports on m contents.
                return n
        }

        parent := base.Ctxt.PosTable.Pos(n.Pos()).Base().InliningIndex()
        sym := fn.Linksym()

        // Check if we've already inlined this function at this particular
        // call site, in order to stop inlining when we reach the beginning
        // of a recursion cycle again. We don't inline immediately recursive
        // functions, but allow inlining if there is a recursion cycle of
        // many functions. Most likely, the inlining will stop before we
        // even hit the beginning of the cycle again, but this catches the
        // unusual case.
        for inlIndex := parent; inlIndex >= 0; inlIndex = base.Ctxt.InlTree.Parent(inlIndex) {
                if base.Ctxt.InlTree.InlinedFunction(inlIndex) == sym {
                        if base.Flag.LowerM > 1 {
                                fmt.Printf("%v: cannot inline %v into %v: repeated recursive cycle\n", ir.Line(n), fn, ir.FuncName(ir.CurFunc))
                        }
                        return n
                }
        }

        typecheck.AssertFixedCall(n)

        inlIndex := base.Ctxt.InlTree.Add(parent, n.Pos(), sym)

        closureInitLSym := func(n *ir.CallExpr, fn *ir.Func) {
                // The linker needs FuncInfo metadata for all inlined
                // functions. This is typically handled by gc.enqueueFunc
                // calling ir.InitLSym for all function declarations in
                // typecheck.Target.Decls (ir.UseClosure adds all closures to
                // Decls).
                //
                // However, non-trivial closures in Decls are ignored, and are
                // insteaded enqueued when walk of the calling function
                // discovers them.
                //
                // This presents a problem for direct calls to closures.
                // Inlining will replace the entire closure definition with its
                // body, which hides the closure from walk and thus suppresses
                // symbol creation.
                //
                // Explicitly create a symbol early in this edge case to ensure
                // we keep this metadata.
                //
                // TODO: Refactor to keep a reference so this can all be done
                // by enqueueFunc.

                if n.Op() != ir.OCALLFUNC {
                        // Not a standard call.
                        return
                }
                if n.X.Op() != ir.OCLOSURE {
                        // Not a direct closure call.
                        return
                }

                clo := n.X.(*ir.ClosureExpr)
                if ir.IsTrivialClosure(clo) {
                        // enqueueFunc will handle trivial closures anyways.
                        return
                }

                ir.InitLSym(fn, true)
        }

        closureInitLSym(n, fn)

        if base.Flag.GenDwarfInl > 0 {
                if !sym.WasInlined() {
                        base.Ctxt.DwFixups.SetPrecursorFunc(sym, fn)
                        sym.Set(obj.AttrWasInlined, true)
                }
        }

        if base.Flag.LowerM != 0 {
                fmt.Printf("%v: inlining call to %v\n", ir.Line(n), fn)
        }
        if base.Flag.LowerM > 2 {
                fmt.Printf("%v: Before inlining: %+v\n", ir.Line(n), n)
        }

        if base.Debug.PGOInline > 0 {
                csi := pgo.CallSiteInfo{LineOffset: pgo.NodeLineOffset(n, fn), Caller: ir.CurFunc}
                if _, ok := inlinedCallSites[csi]; !ok {
                        inlinedCallSites[csi] = struct{}{}
                }
        }

        res := InlineCall(n, fn, inlIndex)

        if res == nil {
                base.FatalfAt(n.Pos(), "inlining call to %v failed", fn)
        }

        if base.Flag.LowerM > 2 {
                fmt.Printf("%v: After inlining %+v\n\n", ir.Line(res), res)
        }

        *inlCalls = append(*inlCalls, res)

        return res
}

// CalleeEffects appends any side effects from evaluating callee to init.
func CalleeEffects(init *ir.Nodes, callee ir.Node) {
        for {
                init.Append(ir.TakeInit(callee)...)

                switch callee.Op() {
                case ir.ONAME, ir.OCLOSURE, ir.OMETHEXPR:
                        return // done

                case ir.OCONVNOP:
                        conv := callee.(*ir.ConvExpr)
                        callee = conv.X

                case ir.OINLCALL:
                        ic := callee.(*ir.InlinedCallExpr)
                        init.Append(ic.Body.Take()...)
                        callee = ic.SingleResult()

                default:
                        base.FatalfAt(callee.Pos(), "unexpected callee expression: %v", callee)
                }
        }
}

func pruneUnusedAutos(ll []*ir.Name, vis *hairyVisitor) []*ir.Name {
        s := make([]*ir.Name, 0, len(ll))
        for _, n := range ll {
                if n.Class == ir.PAUTO {
                        if !vis.usedLocals.Has(n) {
                                continue
                        }
                }
                s = append(s, n)
        }
        return s
}

// numNonClosures returns the number of functions in list which are not closures.
func numNonClosures(list []*ir.Func) int {
        count := 0
        for _, fn := range list {
                if fn.OClosure == nil {
                        count++
                }
        }
        return count
}

func doList(list []ir.Node, do func(ir.Node) bool) bool {
        for _, x := range list {
                if x != nil {
                        if do(x) {
                                return true
                        }
                }
        }
        return false
}

// isIndexingCoverageCounter returns true if the specified node 'n' is indexing
// into a coverage counter array.
func isIndexingCoverageCounter(n ir.Node) bool {
        if n.Op() != ir.OINDEX {
                return false
        }
        ixn := n.(*ir.IndexExpr)
        if ixn.X.Op() != ir.ONAME || !ixn.X.Type().IsArray() {
                return false
        }
        nn := ixn.X.(*ir.Name)
        return nn.CoverageCounter()
}

// isAtomicCoverageCounterUpdate examines the specified node to
// determine whether it represents a call to sync/atomic.AddUint32 to
// increment a coverage counter.
func isAtomicCoverageCounterUpdate(cn *ir.CallExpr) bool {
        if cn.X.Op() != ir.ONAME {
                return false
        }
        name := cn.X.(*ir.Name)
        if name.Class != ir.PFUNC {
                return false
        }
        fn := name.Sym().Name
        if name.Sym().Pkg.Path != "sync/atomic" ||
                (fn != "AddUint32" && fn != "StoreUint32") {
                return false
        }
        if len(cn.Args) != 2 || cn.Args[0].Op() != ir.OADDR {
                return false
        }
        adn := cn.Args[0].(*ir.AddrExpr)
        v := isIndexingCoverageCounter(adn.X)
        return v
}