1 // Copyright 2011 The Go Authors. All rights reserved.
2 // Use of this source code is governed by a BSD-style
3 // license that can be found in the LICENSE file.
5 // The inlining facility makes 2 passes: first CanInline determines which
6 // functions are suitable for inlining, and for those that are it
7 // saves a copy of the body. Then InlineCalls walks each function body to
8 // expand calls to inlinable functions.
10 // The Debug.l flag controls the aggressiveness. Note that main() swaps level 0 and 1,
11 // making 1 the default and -l disable. Additional levels (beyond -l) may be buggy and
14 // 1: 80-nodes leaf functions, oneliners, panic, lazy typechecking (default)
17 // 4: allow non-leaf functions
19 // At some point this may get another default and become switch-offable with -N.
21 // The -d typcheckinl flag enables early typechecking of all imported bodies,
22 // which is useful to flush out bugs.
24 // The Debug.m flag enables diagnostic output. a single -m is useful for verifying
25 // which calls get inlined or not, more is for debugging, and may go away at any point.
32 "internal/goexperiment"
36 "cmd/compile/internal/base"
37 "cmd/compile/internal/inline/inlheur"
38 "cmd/compile/internal/ir"
39 "cmd/compile/internal/logopt"
40 "cmd/compile/internal/pgo"
41 "cmd/compile/internal/typecheck"
42 "cmd/compile/internal/types"
46 // Inlining budget parameters, gathered in one place
49 inlineExtraAppendCost = 0
50 // default is to inline if there's at most one call. -l=4 overrides this by using 1 instead.
51 inlineExtraCallCost = 57 // 57 was benchmarked to provided most benefit with no bad surprises; see https://github.com/golang/go/issues/19348#issuecomment-439370742
52 inlineExtraPanicCost = 1 // do not penalize inlining panics.
53 inlineExtraThrowCost = inlineMaxBudget // with current (2018-05/1.11) code, inlining runtime.throw does not help.
55 inlineBigFunctionNodes = 5000 // Functions with this many nodes are considered "big".
56 inlineBigFunctionMaxCost = 20 // Max cost of inlinee when inlining into a "big" function.
60 // List of all hot callee nodes.
61 // TODO(prattmic): Make this non-global.
62 candHotCalleeMap = make(map[*pgo.IRNode]struct{})
64 // List of all hot call sites. CallSiteInfo.Callee is always nil.
65 // TODO(prattmic): Make this non-global.
66 candHotEdgeMap = make(map[pgo.CallSiteInfo]struct{})
68 // Threshold in percentage for hot callsite inlining.
69 inlineHotCallSiteThresholdPercent float64
71 // Threshold in CDF percentage for hot callsite inlining,
72 // that is, for a threshold of X the hottest callsites that
73 // make up the top X% of total edge weight will be
74 // considered hot for inlining candidates.
75 inlineCDFHotCallSiteThresholdPercent = float64(99)
77 // Budget increased due to hotness.
78 inlineHotMaxBudget int32 = 2000
81 // pgoInlinePrologue records the hot callsites from ir-graph.
82 func pgoInlinePrologue(p *pgo.Profile, funcs []*ir.Func) {
83 if base.Debug.PGOInlineCDFThreshold != "" {
84 if s, err := strconv.ParseFloat(base.Debug.PGOInlineCDFThreshold, 64); err == nil && s >= 0 && s <= 100 {
85 inlineCDFHotCallSiteThresholdPercent = s
87 base.Fatalf("invalid PGOInlineCDFThreshold, must be between 0 and 100")
90 var hotCallsites []pgo.NodeMapKey
91 inlineHotCallSiteThresholdPercent, hotCallsites = hotNodesFromCDF(p)
92 if base.Debug.PGODebug > 0 {
93 fmt.Printf("hot-callsite-thres-from-CDF=%v\n", inlineHotCallSiteThresholdPercent)
96 if x := base.Debug.PGOInlineBudget; x != 0 {
97 inlineHotMaxBudget = int32(x)
100 for _, n := range hotCallsites {
101 // mark inlineable callees from hot edges
102 if callee := p.WeightedCG.IRNodes[n.CalleeName]; callee != nil {
103 candHotCalleeMap[callee] = struct{}{}
105 // mark hot call sites
106 if caller := p.WeightedCG.IRNodes[n.CallerName]; caller != nil && caller.AST != nil {
107 csi := pgo.CallSiteInfo{LineOffset: n.CallSiteOffset, Caller: caller.AST}
108 candHotEdgeMap[csi] = struct{}{}
112 if base.Debug.PGODebug >= 3 {
113 fmt.Printf("hot-cg before inline in dot format:")
114 p.PrintWeightedCallGraphDOT(inlineHotCallSiteThresholdPercent)
118 // hotNodesFromCDF computes an edge weight threshold and the list of hot
119 // nodes that make up the given percentage of the CDF. The threshold, as
120 // a percent, is the lower bound of weight for nodes to be considered hot
121 // (currently only used in debug prints) (in case of equal weights,
122 // comparing with the threshold may not accurately reflect which nodes are
124 func hotNodesFromCDF(p *pgo.Profile) (float64, []pgo.NodeMapKey) {
125 nodes := make([]pgo.NodeMapKey, len(p.NodeMap))
127 for n := range p.NodeMap {
131 sort.Slice(nodes, func(i, j int) bool {
132 ni, nj := nodes[i], nodes[j]
133 if wi, wj := p.NodeMap[ni].EWeight, p.NodeMap[nj].EWeight; wi != wj {
134 return wi > wj // want larger weight first
136 // same weight, order by name/line number
137 if ni.CallerName != nj.CallerName {
138 return ni.CallerName < nj.CallerName
140 if ni.CalleeName != nj.CalleeName {
141 return ni.CalleeName < nj.CalleeName
143 return ni.CallSiteOffset < nj.CallSiteOffset
146 for i, n := range nodes {
147 w := p.NodeMap[n].EWeight
149 if pgo.WeightInPercentage(cum, p.TotalEdgeWeight) > inlineCDFHotCallSiteThresholdPercent {
150 // nodes[:i+1] to include the very last node that makes it to go over the threshold.
151 // (Say, if the CDF threshold is 50% and one hot node takes 60% of weight, we want to
152 // include that node instead of excluding it.)
153 return pgo.WeightInPercentage(w, p.TotalEdgeWeight), nodes[:i+1]
159 // InlinePackage finds functions that can be inlined and clones them before walk expands them.
160 func InlinePackage(p *pgo.Profile) {
161 if base.Debug.PGOInline == 0 {
165 InlineDecls(p, typecheck.Target.Funcs, true)
167 // Perform a garbage collection of hidden closures functions that
168 // are no longer reachable from top-level functions following
169 // inlining. See #59404 and #59638 for more context.
170 garbageCollectUnreferencedHiddenClosures()
172 if base.Debug.DumpInlFuncProps != "" {
173 inlheur.DumpFuncProps(nil, base.Debug.DumpInlFuncProps, nil)
177 // InlineDecls applies inlining to the given batch of declarations.
178 func InlineDecls(p *pgo.Profile, funcs []*ir.Func, doInline bool) {
180 pgoInlinePrologue(p, funcs)
183 doCanInline := func(n *ir.Func, recursive bool, numfns int) {
184 if !recursive || numfns > 1 {
185 // We allow inlining if there is no
186 // recursion, or the recursion cycle is
187 // across more than one function.
190 if base.Flag.LowerM > 1 && n.OClosure == nil {
191 fmt.Printf("%v: cannot inline %v: recursive\n", ir.Line(n), n.Nname)
196 ir.VisitFuncsBottomUp(funcs, func(list []*ir.Func, recursive bool) {
197 numfns := numNonClosures(list)
198 // We visit functions within an SCC in fairly arbitrary order,
199 // so by computing inlinability for all functions in the SCC
200 // before performing any inlining, the results are less
201 // sensitive to the order within the SCC (see #58905 for an
204 // First compute inlinability for all functions in the SCC ...
205 for _, n := range list {
206 doCanInline(n, recursive, numfns)
208 // ... then make a second pass to do inlining of calls.
210 for _, n := range list {
217 // garbageCollectUnreferencedHiddenClosures makes a pass over all the
218 // top-level (non-hidden-closure) functions looking for nested closure
219 // functions that are reachable, then sweeps through the Target.Decls
220 // list and marks any non-reachable hidden closure function as dead.
221 // See issues #59404 and #59638 for more context.
222 func garbageCollectUnreferencedHiddenClosures() {
224 liveFuncs := make(map[*ir.Func]bool)
226 var markLiveFuncs func(fn *ir.Func)
227 markLiveFuncs = func(fn *ir.Func) {
232 ir.Visit(fn, func(n ir.Node) {
233 if clo, ok := n.(*ir.ClosureExpr); ok {
234 markLiveFuncs(clo.Func)
239 for i := 0; i < len(typecheck.Target.Funcs); i++ {
240 fn := typecheck.Target.Funcs[i]
241 if fn.IsHiddenClosure() {
247 for i := 0; i < len(typecheck.Target.Funcs); i++ {
248 fn := typecheck.Target.Funcs[i]
249 if !fn.IsHiddenClosure() {
252 if fn.IsDeadcodeClosure() {
258 fn.SetIsDeadcodeClosure(true)
259 if base.Flag.LowerM > 2 {
260 fmt.Printf("%v: unreferenced closure %v marked as dead\n", ir.Line(fn), fn)
262 if fn.Inl != nil && fn.LSym == nil {
263 ir.InitLSym(fn, true)
268 // inlineBudget determines the max budget for function 'fn' prior to
269 // analyzing the hairyness of the body of 'fn'. We pass in the pgo
270 // profile if available, which can change the budget. If 'verbose' is
271 // set, then print a remark where we boost the budget due to PGO.
272 func inlineBudget(fn *ir.Func, profile *pgo.Profile, verbose bool) int32 {
273 // Update the budget for profile-guided inlining.
274 budget := int32(inlineMaxBudget)
276 if n, ok := profile.WeightedCG.IRNodes[ir.LinkFuncName(fn)]; ok {
277 if _, ok := candHotCalleeMap[n]; ok {
278 budget = int32(inlineHotMaxBudget)
280 fmt.Printf("hot-node enabled increased budget=%v for func=%v\n", budget, ir.PkgFuncName(fn))
288 // CanInline determines whether fn is inlineable.
289 // If so, CanInline saves copies of fn.Body and fn.Dcl in fn.Inl.
290 // fn and fn.Body will already have been typechecked.
291 func CanInline(fn *ir.Func, profile *pgo.Profile) {
293 base.Fatalf("CanInline no nname %+v", fn)
296 canInline := func(fn *ir.Func) { CanInline(fn, profile) }
298 var funcProps *inlheur.FuncProps
299 if goexperiment.NewInliner {
300 funcProps = inlheur.AnalyzeFunc(fn, canInline)
303 if base.Debug.DumpInlFuncProps != "" {
304 inlheur.DumpFuncProps(fn, base.Debug.DumpInlFuncProps, canInline)
307 var reason string // reason, if any, that the function was not inlined
308 if base.Flag.LowerM > 1 || logopt.Enabled() {
311 if base.Flag.LowerM > 1 {
312 fmt.Printf("%v: cannot inline %v: %s\n", ir.Line(fn), fn.Nname, reason)
314 if logopt.Enabled() {
315 logopt.LogOpt(fn.Pos(), "cannotInlineFunction", "inline", ir.FuncName(fn), reason)
321 reason = InlineImpossible(fn)
325 if fn.Typecheck() == 0 {
326 base.Fatalf("CanInline on non-typechecked function %v", fn)
330 if n.Func.InlinabilityChecked() {
333 defer n.Func.SetInlinabilityChecked(true)
335 cc := int32(inlineExtraCallCost)
336 if base.Flag.LowerL == 4 {
337 cc = 1 // this appears to yield better performance than 0.
340 // Compute the inline budget for this function.
341 budget := inlineBudget(fn, profile, base.Debug.PGODebug > 0)
343 // At this point in the game the function we're looking at may
344 // have "stale" autos, vars that still appear in the Dcl list, but
345 // which no longer have any uses in the function body (due to
346 // elimination by deadcode). We'd like to exclude these dead vars
347 // when creating the "Inline.Dcl" field below; to accomplish this,
348 // the hairyVisitor below builds up a map of used/referenced
349 // locals, and we use this map to produce a pruned Inline.Dcl
350 // list. See issue 25459 for more context.
352 visitor := hairyVisitor{
359 if visitor.tooHairy(fn) {
360 reason = visitor.reason
364 n.Func.Inl = &ir.Inline{
365 Cost: budget - visitor.budget,
366 Dcl: pruneUnusedAutos(n.Func.Dcl, &visitor),
369 CanDelayResults: canDelayResults(fn),
371 if goexperiment.NewInliner {
372 n.Func.Inl.Properties = funcProps.SerializeToString()
375 if base.Flag.LowerM > 1 {
376 fmt.Printf("%v: can inline %v with cost %d as: %v { %v }\n", ir.Line(fn), n, budget-visitor.budget, fn.Type(), ir.Nodes(fn.Body))
377 } else if base.Flag.LowerM != 0 {
378 fmt.Printf("%v: can inline %v\n", ir.Line(fn), n)
380 if logopt.Enabled() {
381 logopt.LogOpt(fn.Pos(), "canInlineFunction", "inline", ir.FuncName(fn), fmt.Sprintf("cost: %d", budget-visitor.budget))
385 // InlineImpossible returns a non-empty reason string if fn is impossible to
386 // inline regardless of cost or contents.
387 func InlineImpossible(fn *ir.Func) string {
388 var reason string // reason, if any, that the function can not be inlined.
394 // If marked "go:noinline", don't inline.
395 if fn.Pragma&ir.Noinline != 0 {
396 reason = "marked go:noinline"
400 // If marked "go:norace" and -race compilation, don't inline.
401 if base.Flag.Race && fn.Pragma&ir.Norace != 0 {
402 reason = "marked go:norace with -race compilation"
406 // If marked "go:nocheckptr" and -d checkptr compilation, don't inline.
407 if base.Debug.Checkptr != 0 && fn.Pragma&ir.NoCheckPtr != 0 {
408 reason = "marked go:nocheckptr"
412 // If marked "go:cgo_unsafe_args", don't inline, since the function
413 // makes assumptions about its argument frame layout.
414 if fn.Pragma&ir.CgoUnsafeArgs != 0 {
415 reason = "marked go:cgo_unsafe_args"
419 // If marked as "go:uintptrkeepalive", don't inline, since the keep
420 // alive information is lost during inlining.
422 // TODO(prattmic): This is handled on calls during escape analysis,
423 // which is after inlining. Move prior to inlining so the keep-alive is
424 // maintained after inlining.
425 if fn.Pragma&ir.UintptrKeepAlive != 0 {
426 reason = "marked as having a keep-alive uintptr argument"
430 // If marked as "go:uintptrescapes", don't inline, since the escape
431 // information is lost during inlining.
432 if fn.Pragma&ir.UintptrEscapes != 0 {
433 reason = "marked as having an escaping uintptr argument"
437 // The nowritebarrierrec checker currently works at function
438 // granularity, so inlining yeswritebarrierrec functions can confuse it
439 // (#22342). As a workaround, disallow inlining them for now.
440 if fn.Pragma&ir.Yeswritebarrierrec != 0 {
441 reason = "marked go:yeswritebarrierrec"
445 // If a local function has no fn.Body (is defined outside of Go), cannot inline it.
446 // Imported functions don't have fn.Body but might have inline body in fn.Inl.
447 if len(fn.Body) == 0 && !typecheck.HaveInlineBody(fn) {
448 reason = "no function body"
455 // canDelayResults reports whether inlined calls to fn can delay
456 // declaring the result parameter until the "return" statement.
457 func canDelayResults(fn *ir.Func) bool {
458 // We can delay declaring+initializing result parameters if:
459 // (1) there's exactly one "return" statement in the inlined function;
460 // (2) it's not an empty return statement (#44355); and
461 // (3) the result parameters aren't named.
464 ir.VisitList(fn.Body, func(n ir.Node) {
465 if n, ok := n.(*ir.ReturnStmt); ok {
467 if len(n.Results) == 0 {
468 nreturns++ // empty return statement (case 2)
474 return false // not exactly one return statement (case 1)
477 // temporaries for return values.
478 for _, param := range fn.Type().Results() {
479 if sym := types.OrigSym(param.Sym); sym != nil && !sym.IsBlank() {
480 return false // found a named result parameter (case 3)
487 // hairyVisitor visits a function body to determine its inlining
488 // hairiness and whether or not it can be inlined.
489 type hairyVisitor struct {
490 // This is needed to access the current caller in the doNode function.
496 usedLocals ir.NameSet
497 do func(ir.Node) bool
501 func (v *hairyVisitor) tooHairy(fn *ir.Func) bool {
502 v.do = v.doNode // cache closure
503 if ir.DoChildren(fn, v.do) {
507 v.reason = fmt.Sprintf("function too complex: cost %d exceeds budget %d", v.maxBudget-v.budget, v.maxBudget)
513 // doNode visits n and its children, updates the state in v, and returns true if
514 // n makes the current function too hairy for inlining.
515 func (v *hairyVisitor) doNode(n ir.Node) bool {
521 // Call is okay if inlinable and we have the budget for the body.
523 n := n.(*ir.CallExpr)
524 // Functions that call runtime.getcaller{pc,sp} can not be inlined
525 // because getcaller{pc,sp} expect a pointer to the caller's first argument.
527 // runtime.throw is a "cheap call" like panic in normal code.
529 if n.X.Op() == ir.ONAME {
530 name := n.X.(*ir.Name)
531 if name.Class == ir.PFUNC {
532 switch fn := types.RuntimeSymName(name.Sym()); fn {
533 case "getcallerpc", "getcallersp":
534 v.reason = "call to " + fn
537 v.budget -= inlineExtraThrowCost
540 // Special case for reflect.noescape. It does just type
541 // conversions to appease the escape analysis, and doesn't
543 if types.ReflectSymName(name.Sym()) == "noescape" {
547 // Special case for coverage counter updates; although
548 // these correspond to real operations, we treat them as
549 // zero cost for the moment. This is due to the existence
550 // of tests that are sensitive to inlining-- if the
551 // insertion of coverage instrumentation happens to tip a
552 // given function over the threshold and move it from
553 // "inlinable" to "not-inlinable", this can cause changes
554 // in allocation behavior, which can then result in test
555 // failures (a good example is the TestAllocations in
557 if isAtomicCoverageCounterUpdate(n) {
561 if n.X.Op() == ir.OMETHEXPR {
562 if meth := ir.MethodExprName(n.X); meth != nil {
563 if fn := meth.Func; fn != nil {
565 if types.RuntimeSymName(s) == "heapBits.nextArena" {
566 // Special case: explicitly allow mid-stack inlining of
567 // runtime.heapBits.next even though it calls slow-path
568 // runtime.heapBits.nextArena.
571 // Special case: on architectures that can do unaligned loads,
572 // explicitly mark encoding/binary methods as cheap,
573 // because in practice they are, even though our inlining
574 // budgeting system does not see that. See issue 42958.
575 if base.Ctxt.Arch.CanMergeLoads && s.Pkg.Path == "encoding/binary" {
577 case "littleEndian.Uint64", "littleEndian.Uint32", "littleEndian.Uint16",
578 "bigEndian.Uint64", "bigEndian.Uint32", "bigEndian.Uint16",
579 "littleEndian.PutUint64", "littleEndian.PutUint32", "littleEndian.PutUint16",
580 "bigEndian.PutUint64", "bigEndian.PutUint32", "bigEndian.PutUint16",
581 "littleEndian.AppendUint64", "littleEndian.AppendUint32", "littleEndian.AppendUint16",
582 "bigEndian.AppendUint64", "bigEndian.AppendUint32", "bigEndian.AppendUint16":
590 break // treat like any other node, that is, cost of 1
593 // Determine if the callee edge is for an inlinable hot callee or not.
594 if v.profile != nil && v.curFunc != nil {
595 if fn := inlCallee(v.curFunc, n.X, v.profile); fn != nil && typecheck.HaveInlineBody(fn) {
596 lineOffset := pgo.NodeLineOffset(n, fn)
597 csi := pgo.CallSiteInfo{LineOffset: lineOffset, Caller: v.curFunc}
598 if _, o := candHotEdgeMap[csi]; o {
599 if base.Debug.PGODebug > 0 {
600 fmt.Printf("hot-callsite identified at line=%v for func=%v\n", ir.Line(n), ir.PkgFuncName(v.curFunc))
606 if ir.IsIntrinsicCall(n) {
607 // Treat like any other node.
611 if fn := inlCallee(v.curFunc, n.X, v.profile); fn != nil && typecheck.HaveInlineBody(fn) {
612 v.budget -= fn.Inl.Cost
616 // Call cost for non-leaf inlining.
617 v.budget -= v.extraCallCost
620 base.FatalfAt(n.Pos(), "OCALLMETH missed by typecheck")
622 // Things that are too hairy, irrespective of the budget
623 case ir.OCALL, ir.OCALLINTER:
624 // Call cost for non-leaf inlining.
625 v.budget -= v.extraCallCost
628 n := n.(*ir.UnaryExpr)
629 if n.X.Op() == ir.OCONVIFACE && n.X.(*ir.ConvExpr).Implicit() {
630 // Hack to keep reflect.flag.mustBe inlinable for TestIntendedInlining.
631 // Before CL 284412, these conversions were introduced later in the
632 // compiler, so they didn't count against inlining budget.
635 v.budget -= inlineExtraPanicCost
638 base.FatalfAt(n.Pos(), "ORECOVER missed typecheck")
640 // recover matches the argument frame pointer to find
641 // the right panic value, so it needs an argument frame.
642 v.reason = "call to recover"
646 if base.Debug.InlFuncsWithClosures == 0 {
647 v.reason = "not inlining functions with closures"
651 // TODO(danscales): Maybe make budget proportional to number of closure
653 //v.budget -= int32(len(n.(*ir.ClosureExpr).Func.ClosureVars) * 3)
654 // TODO(austin): However, if we're able to inline this closure into
655 // v.curFunc, then we actually pay nothing for the closure captures. We
656 // should try to account for that if we're going to account for captures.
659 case ir.OGO, ir.ODEFER, ir.OTAILCALL:
660 v.reason = "unhandled op " + n.Op().String()
664 v.budget -= inlineExtraAppendCost
667 n := n.(*ir.AddrExpr)
668 // Make "&s.f" cost 0 when f's offset is zero.
669 if dot, ok := n.X.(*ir.SelectorExpr); ok && (dot.Op() == ir.ODOT || dot.Op() == ir.ODOTPTR) {
670 if _, ok := dot.X.(*ir.Name); ok && dot.Selection.Offset == 0 {
671 v.budget += 2 // undo ir.OADDR+ir.ODOT/ir.ODOTPTR
676 // *(*X)(unsafe.Pointer(&x)) is low-cost
677 n := n.(*ir.StarExpr)
680 for ptr.Op() == ir.OCONVNOP {
681 ptr = ptr.(*ir.ConvExpr).X
683 if ptr.Op() == ir.OADDR {
684 v.budget += 1 // undo half of default cost of ir.ODEREF+ir.OADDR
688 // This doesn't produce code, but the children might.
689 v.budget++ // undo default cost
691 case ir.OFALL, ir.OTYPE:
692 // These nodes don't produce code; omit from inlining budget.
697 if ir.IsConst(n.Cond, constant.Bool) {
698 // This if and the condition cost nothing.
699 if doList(n.Init(), v.do) {
702 if ir.BoolVal(n.Cond) {
703 return doList(n.Body, v.do)
705 return doList(n.Else, v.do)
711 if n.Class == ir.PAUTO {
716 // The only OBLOCK we should see at this point is an empty one.
717 // In any event, let the visitList(n.List()) below take care of the statements,
718 // and don't charge for the OBLOCK itself. The ++ undoes the -- below.
721 case ir.OMETHVALUE, ir.OSLICELIT:
722 v.budget-- // Hack for toolstash -cmp.
725 v.budget++ // Hack for toolstash -cmp.
728 n := n.(*ir.AssignListStmt)
730 // Unified IR unconditionally rewrites:
741 // so that it can insert implicit conversions as necessary. To
742 // minimize impact to the existing inlining heuristics (in
743 // particular, to avoid breaking the existing inlinability regress
744 // tests), we need to compensate for this here.
746 // See also identical logic in isBigFunc.
747 if init := n.Rhs[0].Init(); len(init) == 1 {
748 if _, ok := init[0].(*ir.AssignListStmt); ok {
749 // 4 for each value, because each temporary variable now
750 // appears 3 times (DCL, LHS, RHS), plus an extra DCL node.
752 // 1 for the extra "tmp1, tmp2 = f()" assignment statement.
753 v.budget += 4*int32(len(n.Lhs)) + 1
758 // Special case for coverage counter updates and coverage
759 // function registrations. Although these correspond to real
760 // operations, we treat them as zero cost for the moment. This
761 // is primarily due to the existence of tests that are
762 // sensitive to inlining-- if the insertion of coverage
763 // instrumentation happens to tip a given function over the
764 // threshold and move it from "inlinable" to "not-inlinable",
765 // this can cause changes in allocation behavior, which can
766 // then result in test failures (a good example is the
767 // TestAllocations in crypto/ed25519).
768 n := n.(*ir.AssignStmt)
769 if n.X.Op() == ir.OINDEX && isIndexingCoverageCounter(n.X) {
776 // When debugging, don't stop early, to get full cost of inlining this function
777 if v.budget < 0 && base.Flag.LowerM < 2 && !logopt.Enabled() {
778 v.reason = "too expensive"
782 return ir.DoChildren(n, v.do)
785 func isBigFunc(fn *ir.Func) bool {
786 budget := inlineBigFunctionNodes
787 return ir.Any(fn, func(n ir.Node) bool {
788 // See logic in hairyVisitor.doNode, explaining unified IR's
789 // handling of "a, b = f()" assignments.
790 if n, ok := n.(*ir.AssignListStmt); ok && n.Op() == ir.OAS2 {
791 if init := n.Rhs[0].Init(); len(init) == 1 {
792 if _, ok := init[0].(*ir.AssignListStmt); ok {
793 budget += 4*len(n.Lhs) + 1
803 // InlineCalls/inlnode walks fn's statements and expressions and substitutes any
804 // calls made to inlineable functions. This is the external entry point.
805 func InlineCalls(fn *ir.Func, profile *pgo.Profile) {
808 bigCaller := isBigFunc(fn)
809 if bigCaller && base.Flag.LowerM > 1 {
810 fmt.Printf("%v: function %v considered 'big'; reducing max cost of inlinees\n", ir.Line(fn), fn)
812 var inlCalls []*ir.InlinedCallExpr
813 var edit func(ir.Node) ir.Node
814 edit = func(n ir.Node) ir.Node {
815 return inlnode(fn, n, bigCaller, &inlCalls, edit, profile)
817 ir.EditChildren(fn, edit)
819 // If we inlined any calls, we want to recursively visit their
820 // bodies for further inlining. However, we need to wait until
821 // *after* the original function body has been expanded, or else
822 // inlCallee can have false positives (e.g., #54632).
823 for len(inlCalls) > 0 {
825 inlCalls = inlCalls[1:]
826 ir.EditChildren(call, edit)
832 // inlnode recurses over the tree to find inlineable calls, which will
833 // be turned into OINLCALLs by mkinlcall. When the recursion comes
834 // back up will examine left, right, list, rlist, ninit, ntest, nincr,
835 // nbody and nelse and use one of the 4 inlconv/glue functions above
836 // to turn the OINLCALL into an expression, a statement, or patch it
837 // in to this nodes list or rlist as appropriate.
838 // NOTE it makes no sense to pass the glue functions down the
839 // recursion to the level where the OINLCALL gets created because they
840 // have to edit /this/ n, so you'd have to push that one down as well,
841 // but then you may as well do it here. so this is cleaner and
842 // shorter and less complicated.
843 // The result of inlnode MUST be assigned back to n, e.g.
845 // n.Left = inlnode(n.Left)
846 func inlnode(callerfn *ir.Func, n ir.Node, bigCaller bool, inlCalls *[]*ir.InlinedCallExpr, edit func(ir.Node) ir.Node, profile *pgo.Profile) ir.Node {
852 case ir.ODEFER, ir.OGO:
853 n := n.(*ir.GoDeferStmt)
854 switch call := n.Call; call.Op() {
856 base.FatalfAt(call.Pos(), "OCALLMETH missed by typecheck")
858 call := call.(*ir.CallExpr)
862 n := n.(*ir.TailCallStmt)
863 n.Call.NoInline = true // Not inline a tail call for now. Maybe we could inline it just like RETURN fn(arg)?
865 // TODO do them here (or earlier),
866 // so escape analysis can avoid more heapmoves.
870 base.FatalfAt(n.Pos(), "OCALLMETH missed by typecheck")
872 n := n.(*ir.CallExpr)
873 if n.X.Op() == ir.OMETHEXPR {
874 // Prevent inlining some reflect.Value methods when using checkptr,
875 // even when package reflect was compiled without it (#35073).
876 if meth := ir.MethodExprName(n.X); meth != nil {
878 if base.Debug.Checkptr != 0 {
879 switch types.ReflectSymName(s) {
880 case "Value.UnsafeAddr", "Value.Pointer":
890 ir.EditChildren(n, edit)
892 // with all the branches out of the way, it is now time to
893 // transmogrify this node itself unless inhibited by the
894 // switch at the top of this function.
897 base.FatalfAt(n.Pos(), "OCALLMETH missed by typecheck")
900 call := n.(*ir.CallExpr)
904 if base.Flag.LowerM > 3 {
905 fmt.Printf("%v:call to func %+v\n", ir.Line(n), call.X)
907 if ir.IsIntrinsicCall(call) {
910 if fn := inlCallee(callerfn, call.X, profile); fn != nil && typecheck.HaveInlineBody(fn) {
911 n = mkinlcall(callerfn, call, fn, bigCaller, inlCalls)
920 // inlCallee takes a function-typed expression and returns the underlying function ONAME
921 // that it refers to if statically known. Otherwise, it returns nil.
922 func inlCallee(caller *ir.Func, fn ir.Node, profile *pgo.Profile) (res *ir.Func) {
923 fn = ir.StaticValue(fn)
926 fn := fn.(*ir.SelectorExpr)
927 n := ir.MethodExprName(fn)
928 // Check that receiver type matches fn.X.
929 // TODO(mdempsky): Handle implicit dereference
930 // of pointer receiver argument?
931 if n == nil || !types.Identical(n.Type().Recv().Type, fn.X.Type()) {
937 if fn.Class == ir.PFUNC {
941 fn := fn.(*ir.ClosureExpr)
943 if len(c.ClosureVars) != 0 && c.ClosureVars[0].Outer.Curfn != caller {
944 return nil // inliner doesn't support inlining across closure frames
946 CanInline(c, profile)
954 // SSADumpInline gives the SSA back end a chance to dump the function
955 // when producing output for debugging the compiler itself.
956 var SSADumpInline = func(*ir.Func) {}
958 // InlineCall allows the inliner implementation to be overridden.
959 // If it returns nil, the function will not be inlined.
960 var InlineCall = func(callerfn *ir.Func, call *ir.CallExpr, fn *ir.Func, inlIndex int) *ir.InlinedCallExpr {
961 base.Fatalf("inline.InlineCall not overridden")
965 // inlineCostOK returns true if call n from caller to callee is cheap enough to
966 // inline. bigCaller indicates that caller is a big function.
968 // If inlineCostOK returns false, it also returns the max cost that the callee
970 func inlineCostOK(n *ir.CallExpr, caller, callee *ir.Func, bigCaller bool) (bool, int32) {
971 maxCost := int32(inlineMaxBudget)
973 // We use this to restrict inlining into very big functions.
974 // See issue 26546 and 17566.
975 maxCost = inlineBigFunctionMaxCost
978 if callee.Inl.Cost <= maxCost {
979 // Simple case. Function is already cheap enough.
983 // We'll also allow inlining of hot functions below inlineHotMaxBudget,
984 // but only in small functions.
986 lineOffset := pgo.NodeLineOffset(n, caller)
987 csi := pgo.CallSiteInfo{LineOffset: lineOffset, Caller: caller}
988 if _, ok := candHotEdgeMap[csi]; !ok {
990 return false, maxCost
996 if base.Debug.PGODebug > 0 {
997 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))
999 return false, maxCost
1002 if callee.Inl.Cost > inlineHotMaxBudget {
1003 return false, inlineHotMaxBudget
1006 if base.Debug.PGODebug > 0 {
1007 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))
1013 // If n is a OCALLFUNC node, and fn is an ONAME node for a
1014 // function with an inlinable body, return an OINLCALL node that can replace n.
1015 // The returned node's Ninit has the parameter assignments, the Nbody is the
1016 // inlined function body, and (List, Rlist) contain the (input, output)
1018 // The result of mkinlcall MUST be assigned back to n, e.g.
1020 // n.Left = mkinlcall(n.Left, fn, isddd)
1021 func mkinlcall(callerfn *ir.Func, n *ir.CallExpr, fn *ir.Func, bigCaller bool, inlCalls *[]*ir.InlinedCallExpr) ir.Node {
1023 if logopt.Enabled() {
1024 logopt.LogOpt(n.Pos(), "cannotInlineCall", "inline", ir.FuncName(callerfn),
1025 fmt.Sprintf("%s cannot be inlined", ir.PkgFuncName(fn)))
1030 if ok, maxCost := inlineCostOK(n, callerfn, fn, bigCaller); !ok {
1031 if logopt.Enabled() {
1032 logopt.LogOpt(n.Pos(), "cannotInlineCall", "inline", ir.FuncName(callerfn),
1033 fmt.Sprintf("cost %d of %s exceeds max caller cost %d", fn.Inl.Cost, ir.PkgFuncName(fn), maxCost))
1039 // Can't recursively inline a function into itself.
1040 if logopt.Enabled() {
1041 logopt.LogOpt(n.Pos(), "cannotInlineCall", "inline", fmt.Sprintf("recursive call to %s", ir.FuncName(callerfn)))
1046 if base.Flag.Cfg.Instrumenting && types.IsNoInstrumentPkg(fn.Sym().Pkg) {
1047 // Runtime package must not be instrumented.
1048 // Instrument skips runtime package. However, some runtime code can be
1049 // inlined into other packages and instrumented there. To avoid this,
1050 // we disable inlining of runtime functions when instrumenting.
1051 // The example that we observed is inlining of LockOSThread,
1052 // which lead to false race reports on m contents.
1055 if base.Flag.Race && types.IsNoRacePkg(fn.Sym().Pkg) {
1059 parent := base.Ctxt.PosTable.Pos(n.Pos()).Base().InliningIndex()
1062 // Check if we've already inlined this function at this particular
1063 // call site, in order to stop inlining when we reach the beginning
1064 // of a recursion cycle again. We don't inline immediately recursive
1065 // functions, but allow inlining if there is a recursion cycle of
1066 // many functions. Most likely, the inlining will stop before we
1067 // even hit the beginning of the cycle again, but this catches the
1069 for inlIndex := parent; inlIndex >= 0; inlIndex = base.Ctxt.InlTree.Parent(inlIndex) {
1070 if base.Ctxt.InlTree.InlinedFunction(inlIndex) == sym {
1071 if base.Flag.LowerM > 1 {
1072 fmt.Printf("%v: cannot inline %v into %v: repeated recursive cycle\n", ir.Line(n), fn, ir.FuncName(callerfn))
1078 typecheck.AssertFixedCall(n)
1080 inlIndex := base.Ctxt.InlTree.Add(parent, n.Pos(), sym, ir.FuncName(fn))
1082 closureInitLSym := func(n *ir.CallExpr, fn *ir.Func) {
1083 // The linker needs FuncInfo metadata for all inlined
1084 // functions. This is typically handled by gc.enqueueFunc
1085 // calling ir.InitLSym for all function declarations in
1086 // typecheck.Target.Decls (ir.UseClosure adds all closures to
1089 // However, non-trivial closures in Decls are ignored, and are
1090 // insteaded enqueued when walk of the calling function
1093 // This presents a problem for direct calls to closures.
1094 // Inlining will replace the entire closure definition with its
1095 // body, which hides the closure from walk and thus suppresses
1098 // Explicitly create a symbol early in this edge case to ensure
1099 // we keep this metadata.
1101 // TODO: Refactor to keep a reference so this can all be done
1104 if n.Op() != ir.OCALLFUNC {
1105 // Not a standard call.
1108 if n.X.Op() != ir.OCLOSURE {
1109 // Not a direct closure call.
1113 clo := n.X.(*ir.ClosureExpr)
1114 if ir.IsTrivialClosure(clo) {
1115 // enqueueFunc will handle trivial closures anyways.
1119 ir.InitLSym(fn, true)
1122 closureInitLSym(n, fn)
1124 if base.Flag.GenDwarfInl > 0 {
1125 if !sym.WasInlined() {
1126 base.Ctxt.DwFixups.SetPrecursorFunc(sym, fn)
1127 sym.Set(obj.AttrWasInlined, true)
1131 if base.Flag.LowerM != 0 {
1132 fmt.Printf("%v: inlining call to %v\n", ir.Line(n), fn)
1134 if base.Flag.LowerM > 2 {
1135 fmt.Printf("%v: Before inlining: %+v\n", ir.Line(n), n)
1138 res := InlineCall(callerfn, n, fn, inlIndex)
1141 base.FatalfAt(n.Pos(), "inlining call to %v failed", fn)
1144 if base.Flag.LowerM > 2 {
1145 fmt.Printf("%v: After inlining %+v\n\n", ir.Line(res), res)
1148 *inlCalls = append(*inlCalls, res)
1153 // CalleeEffects appends any side effects from evaluating callee to init.
1154 func CalleeEffects(init *ir.Nodes, callee ir.Node) {
1156 init.Append(ir.TakeInit(callee)...)
1158 switch callee.Op() {
1159 case ir.ONAME, ir.OCLOSURE, ir.OMETHEXPR:
1163 conv := callee.(*ir.ConvExpr)
1167 ic := callee.(*ir.InlinedCallExpr)
1168 init.Append(ic.Body.Take()...)
1169 callee = ic.SingleResult()
1172 base.FatalfAt(callee.Pos(), "unexpected callee expression: %v", callee)
1177 func pruneUnusedAutos(ll []*ir.Name, vis *hairyVisitor) []*ir.Name {
1178 s := make([]*ir.Name, 0, len(ll))
1179 for _, n := range ll {
1180 if n.Class == ir.PAUTO {
1181 if !vis.usedLocals.Has(n) {
1182 // TODO(mdempsky): Simplify code after confident that this
1183 // never happens anymore.
1184 base.FatalfAt(n.Pos(), "unused auto: %v", n)
1193 // numNonClosures returns the number of functions in list which are not closures.
1194 func numNonClosures(list []*ir.Func) int {
1196 for _, fn := range list {
1197 if fn.OClosure == nil {
1204 func doList(list []ir.Node, do func(ir.Node) bool) bool {
1205 for _, x := range list {
1215 // isIndexingCoverageCounter returns true if the specified node 'n' is indexing
1216 // into a coverage counter array.
1217 func isIndexingCoverageCounter(n ir.Node) bool {
1218 if n.Op() != ir.OINDEX {
1221 ixn := n.(*ir.IndexExpr)
1222 if ixn.X.Op() != ir.ONAME || !ixn.X.Type().IsArray() {
1225 nn := ixn.X.(*ir.Name)
1226 return nn.CoverageCounter()
1229 // isAtomicCoverageCounterUpdate examines the specified node to
1230 // determine whether it represents a call to sync/atomic.AddUint32 to
1231 // increment a coverage counter.
1232 func isAtomicCoverageCounterUpdate(cn *ir.CallExpr) bool {
1233 if cn.X.Op() != ir.ONAME {
1236 name := cn.X.(*ir.Name)
1237 if name.Class != ir.PFUNC {
1240 fn := name.Sym().Name
1241 if name.Sym().Pkg.Path != "sync/atomic" ||
1242 (fn != "AddUint32" && fn != "StoreUint32") {
1245 if len(cn.Args) != 2 || cn.Args[0].Op() != ir.OADDR {
1248 adn := cn.Args[0].(*ir.AddrExpr)
1249 v := isIndexingCoverageCounter(adn.X)