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 var funcProps *inlheur.FuncProps
297 if goexperiment.NewInliner || inlheur.UnitTesting() {
298 funcProps = inlheur.AnalyzeFunc(fn,
299 func(fn *ir.Func) { CanInline(fn, profile) })
302 var reason string // reason, if any, that the function was not inlined
303 if base.Flag.LowerM > 1 || logopt.Enabled() {
306 if base.Flag.LowerM > 1 {
307 fmt.Printf("%v: cannot inline %v: %s\n", ir.Line(fn), fn.Nname, reason)
309 if logopt.Enabled() {
310 logopt.LogOpt(fn.Pos(), "cannotInlineFunction", "inline", ir.FuncName(fn), reason)
316 reason = InlineImpossible(fn)
320 if fn.Typecheck() == 0 {
321 base.Fatalf("CanInline on non-typechecked function %v", fn)
325 if n.Func.InlinabilityChecked() {
328 defer n.Func.SetInlinabilityChecked(true)
330 cc := int32(inlineExtraCallCost)
331 if base.Flag.LowerL == 4 {
332 cc = 1 // this appears to yield better performance than 0.
335 // Compute the inline budget for this function.
336 budget := inlineBudget(fn, profile, base.Debug.PGODebug > 0)
338 // At this point in the game the function we're looking at may
339 // have "stale" autos, vars that still appear in the Dcl list, but
340 // which no longer have any uses in the function body (due to
341 // elimination by deadcode). We'd like to exclude these dead vars
342 // when creating the "Inline.Dcl" field below; to accomplish this,
343 // the hairyVisitor below builds up a map of used/referenced
344 // locals, and we use this map to produce a pruned Inline.Dcl
345 // list. See issue 25459 for more context.
347 visitor := hairyVisitor{
354 if visitor.tooHairy(fn) {
355 reason = visitor.reason
359 n.Func.Inl = &ir.Inline{
360 Cost: budget - visitor.budget,
361 Dcl: pruneUnusedAutos(n.Func.Dcl, &visitor),
364 CanDelayResults: canDelayResults(fn),
366 if goexperiment.NewInliner {
367 n.Func.Inl.Properties = funcProps.SerializeToString()
370 if base.Flag.LowerM > 1 {
371 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))
372 } else if base.Flag.LowerM != 0 {
373 fmt.Printf("%v: can inline %v\n", ir.Line(fn), n)
375 if logopt.Enabled() {
376 logopt.LogOpt(fn.Pos(), "canInlineFunction", "inline", ir.FuncName(fn), fmt.Sprintf("cost: %d", budget-visitor.budget))
380 // InlineImpossible returns a non-empty reason string if fn is impossible to
381 // inline regardless of cost or contents.
382 func InlineImpossible(fn *ir.Func) string {
383 var reason string // reason, if any, that the function can not be inlined.
389 // If marked "go:noinline", don't inline.
390 if fn.Pragma&ir.Noinline != 0 {
391 reason = "marked go:noinline"
395 // If marked "go:norace" and -race compilation, don't inline.
396 if base.Flag.Race && fn.Pragma&ir.Norace != 0 {
397 reason = "marked go:norace with -race compilation"
401 // If marked "go:nocheckptr" and -d checkptr compilation, don't inline.
402 if base.Debug.Checkptr != 0 && fn.Pragma&ir.NoCheckPtr != 0 {
403 reason = "marked go:nocheckptr"
407 // If marked "go:cgo_unsafe_args", don't inline, since the function
408 // makes assumptions about its argument frame layout.
409 if fn.Pragma&ir.CgoUnsafeArgs != 0 {
410 reason = "marked go:cgo_unsafe_args"
414 // If marked as "go:uintptrkeepalive", don't inline, since the keep
415 // alive information is lost during inlining.
417 // TODO(prattmic): This is handled on calls during escape analysis,
418 // which is after inlining. Move prior to inlining so the keep-alive is
419 // maintained after inlining.
420 if fn.Pragma&ir.UintptrKeepAlive != 0 {
421 reason = "marked as having a keep-alive uintptr argument"
425 // If marked as "go:uintptrescapes", don't inline, since the escape
426 // information is lost during inlining.
427 if fn.Pragma&ir.UintptrEscapes != 0 {
428 reason = "marked as having an escaping uintptr argument"
432 // The nowritebarrierrec checker currently works at function
433 // granularity, so inlining yeswritebarrierrec functions can confuse it
434 // (#22342). As a workaround, disallow inlining them for now.
435 if fn.Pragma&ir.Yeswritebarrierrec != 0 {
436 reason = "marked go:yeswritebarrierrec"
440 // If a local function has no fn.Body (is defined outside of Go), cannot inline it.
441 // Imported functions don't have fn.Body but might have inline body in fn.Inl.
442 if len(fn.Body) == 0 && !typecheck.HaveInlineBody(fn) {
443 reason = "no function body"
450 // canDelayResults reports whether inlined calls to fn can delay
451 // declaring the result parameter until the "return" statement.
452 func canDelayResults(fn *ir.Func) bool {
453 // We can delay declaring+initializing result parameters if:
454 // (1) there's exactly one "return" statement in the inlined function;
455 // (2) it's not an empty return statement (#44355); and
456 // (3) the result parameters aren't named.
459 ir.VisitList(fn.Body, func(n ir.Node) {
460 if n, ok := n.(*ir.ReturnStmt); ok {
462 if len(n.Results) == 0 {
463 nreturns++ // empty return statement (case 2)
469 return false // not exactly one return statement (case 1)
472 // temporaries for return values.
473 for _, param := range fn.Type().Results() {
474 if sym := param.Sym; sym != nil && !sym.IsBlank() {
475 return false // found a named result parameter (case 3)
482 // hairyVisitor visits a function body to determine its inlining
483 // hairiness and whether or not it can be inlined.
484 type hairyVisitor struct {
485 // This is needed to access the current caller in the doNode function.
491 usedLocals ir.NameSet
492 do func(ir.Node) bool
496 func (v *hairyVisitor) tooHairy(fn *ir.Func) bool {
497 v.do = v.doNode // cache closure
498 if ir.DoChildren(fn, v.do) {
502 v.reason = fmt.Sprintf("function too complex: cost %d exceeds budget %d", v.maxBudget-v.budget, v.maxBudget)
508 // doNode visits n and its children, updates the state in v, and returns true if
509 // n makes the current function too hairy for inlining.
510 func (v *hairyVisitor) doNode(n ir.Node) bool {
516 // Call is okay if inlinable and we have the budget for the body.
518 n := n.(*ir.CallExpr)
519 // Functions that call runtime.getcaller{pc,sp} can not be inlined
520 // because getcaller{pc,sp} expect a pointer to the caller's first argument.
522 // runtime.throw is a "cheap call" like panic in normal code.
524 if n.X.Op() == ir.ONAME {
525 name := n.X.(*ir.Name)
526 if name.Class == ir.PFUNC {
527 switch fn := types.RuntimeSymName(name.Sym()); fn {
528 case "getcallerpc", "getcallersp":
529 v.reason = "call to " + fn
532 v.budget -= inlineExtraThrowCost
535 // Special case for reflect.noescape. It does just type
536 // conversions to appease the escape analysis, and doesn't
538 if types.ReflectSymName(name.Sym()) == "noescape" {
542 // Special case for coverage counter updates; although
543 // these correspond to real operations, we treat them as
544 // zero cost for the moment. This is due to the existence
545 // of tests that are sensitive to inlining-- if the
546 // insertion of coverage instrumentation happens to tip a
547 // given function over the threshold and move it from
548 // "inlinable" to "not-inlinable", this can cause changes
549 // in allocation behavior, which can then result in test
550 // failures (a good example is the TestAllocations in
552 if isAtomicCoverageCounterUpdate(n) {
556 if n.X.Op() == ir.OMETHEXPR {
557 if meth := ir.MethodExprName(n.X); meth != nil {
558 if fn := meth.Func; fn != nil {
560 if types.RuntimeSymName(s) == "heapBits.nextArena" {
561 // Special case: explicitly allow mid-stack inlining of
562 // runtime.heapBits.next even though it calls slow-path
563 // runtime.heapBits.nextArena.
566 // Special case: on architectures that can do unaligned loads,
567 // explicitly mark encoding/binary methods as cheap,
568 // because in practice they are, even though our inlining
569 // budgeting system does not see that. See issue 42958.
570 if base.Ctxt.Arch.CanMergeLoads && s.Pkg.Path == "encoding/binary" {
572 case "littleEndian.Uint64", "littleEndian.Uint32", "littleEndian.Uint16",
573 "bigEndian.Uint64", "bigEndian.Uint32", "bigEndian.Uint16",
574 "littleEndian.PutUint64", "littleEndian.PutUint32", "littleEndian.PutUint16",
575 "bigEndian.PutUint64", "bigEndian.PutUint32", "bigEndian.PutUint16",
576 "littleEndian.AppendUint64", "littleEndian.AppendUint32", "littleEndian.AppendUint16",
577 "bigEndian.AppendUint64", "bigEndian.AppendUint32", "bigEndian.AppendUint16":
585 break // treat like any other node, that is, cost of 1
588 // Determine if the callee edge is for an inlinable hot callee or not.
589 if v.profile != nil && v.curFunc != nil {
590 if fn := inlCallee(v.curFunc, n.X, v.profile); fn != nil && typecheck.HaveInlineBody(fn) {
591 lineOffset := pgo.NodeLineOffset(n, fn)
592 csi := pgo.CallSiteInfo{LineOffset: lineOffset, Caller: v.curFunc}
593 if _, o := candHotEdgeMap[csi]; o {
594 if base.Debug.PGODebug > 0 {
595 fmt.Printf("hot-callsite identified at line=%v for func=%v\n", ir.Line(n), ir.PkgFuncName(v.curFunc))
601 if ir.IsIntrinsicCall(n) {
602 // Treat like any other node.
606 if fn := inlCallee(v.curFunc, n.X, v.profile); fn != nil && typecheck.HaveInlineBody(fn) {
607 v.budget -= fn.Inl.Cost
611 // Call cost for non-leaf inlining.
612 v.budget -= v.extraCallCost
615 base.FatalfAt(n.Pos(), "OCALLMETH missed by typecheck")
617 // Things that are too hairy, irrespective of the budget
618 case ir.OCALL, ir.OCALLINTER:
619 // Call cost for non-leaf inlining.
620 v.budget -= v.extraCallCost
623 n := n.(*ir.UnaryExpr)
624 if n.X.Op() == ir.OCONVIFACE && n.X.(*ir.ConvExpr).Implicit() {
625 // Hack to keep reflect.flag.mustBe inlinable for TestIntendedInlining.
626 // Before CL 284412, these conversions were introduced later in the
627 // compiler, so they didn't count against inlining budget.
630 v.budget -= inlineExtraPanicCost
633 base.FatalfAt(n.Pos(), "ORECOVER missed typecheck")
635 // recover matches the argument frame pointer to find
636 // the right panic value, so it needs an argument frame.
637 v.reason = "call to recover"
641 if base.Debug.InlFuncsWithClosures == 0 {
642 v.reason = "not inlining functions with closures"
646 // TODO(danscales): Maybe make budget proportional to number of closure
648 //v.budget -= int32(len(n.(*ir.ClosureExpr).Func.ClosureVars) * 3)
649 // TODO(austin): However, if we're able to inline this closure into
650 // v.curFunc, then we actually pay nothing for the closure captures. We
651 // should try to account for that if we're going to account for captures.
654 case ir.OGO, ir.ODEFER, ir.OTAILCALL:
655 v.reason = "unhandled op " + n.Op().String()
659 v.budget -= inlineExtraAppendCost
662 n := n.(*ir.AddrExpr)
663 // Make "&s.f" cost 0 when f's offset is zero.
664 if dot, ok := n.X.(*ir.SelectorExpr); ok && (dot.Op() == ir.ODOT || dot.Op() == ir.ODOTPTR) {
665 if _, ok := dot.X.(*ir.Name); ok && dot.Selection.Offset == 0 {
666 v.budget += 2 // undo ir.OADDR+ir.ODOT/ir.ODOTPTR
671 // *(*X)(unsafe.Pointer(&x)) is low-cost
672 n := n.(*ir.StarExpr)
675 for ptr.Op() == ir.OCONVNOP {
676 ptr = ptr.(*ir.ConvExpr).X
678 if ptr.Op() == ir.OADDR {
679 v.budget += 1 // undo half of default cost of ir.ODEREF+ir.OADDR
683 // This doesn't produce code, but the children might.
684 v.budget++ // undo default cost
686 case ir.OFALL, ir.OTYPE:
687 // These nodes don't produce code; omit from inlining budget.
692 if ir.IsConst(n.Cond, constant.Bool) {
693 // This if and the condition cost nothing.
694 if doList(n.Init(), v.do) {
697 if ir.BoolVal(n.Cond) {
698 return doList(n.Body, v.do)
700 return doList(n.Else, v.do)
706 if n.Class == ir.PAUTO {
711 // The only OBLOCK we should see at this point is an empty one.
712 // In any event, let the visitList(n.List()) below take care of the statements,
713 // and don't charge for the OBLOCK itself. The ++ undoes the -- below.
716 case ir.OMETHVALUE, ir.OSLICELIT:
717 v.budget-- // Hack for toolstash -cmp.
720 v.budget++ // Hack for toolstash -cmp.
723 n := n.(*ir.AssignListStmt)
725 // Unified IR unconditionally rewrites:
736 // so that it can insert implicit conversions as necessary. To
737 // minimize impact to the existing inlining heuristics (in
738 // particular, to avoid breaking the existing inlinability regress
739 // tests), we need to compensate for this here.
741 // See also identical logic in isBigFunc.
742 if init := n.Rhs[0].Init(); len(init) == 1 {
743 if _, ok := init[0].(*ir.AssignListStmt); ok {
744 // 4 for each value, because each temporary variable now
745 // appears 3 times (DCL, LHS, RHS), plus an extra DCL node.
747 // 1 for the extra "tmp1, tmp2 = f()" assignment statement.
748 v.budget += 4*int32(len(n.Lhs)) + 1
753 // Special case for coverage counter updates and coverage
754 // function registrations. Although these correspond to real
755 // operations, we treat them as zero cost for the moment. This
756 // is primarily due to the existence of tests that are
757 // sensitive to inlining-- if the insertion of coverage
758 // instrumentation happens to tip a given function over the
759 // threshold and move it from "inlinable" to "not-inlinable",
760 // this can cause changes in allocation behavior, which can
761 // then result in test failures (a good example is the
762 // TestAllocations in crypto/ed25519).
763 n := n.(*ir.AssignStmt)
764 if n.X.Op() == ir.OINDEX && isIndexingCoverageCounter(n.X) {
771 // When debugging, don't stop early, to get full cost of inlining this function
772 if v.budget < 0 && base.Flag.LowerM < 2 && !logopt.Enabled() {
773 v.reason = "too expensive"
777 return ir.DoChildren(n, v.do)
780 func isBigFunc(fn *ir.Func) bool {
781 budget := inlineBigFunctionNodes
782 return ir.Any(fn, func(n ir.Node) bool {
783 // See logic in hairyVisitor.doNode, explaining unified IR's
784 // handling of "a, b = f()" assignments.
785 if n, ok := n.(*ir.AssignListStmt); ok && n.Op() == ir.OAS2 {
786 if init := n.Rhs[0].Init(); len(init) == 1 {
787 if _, ok := init[0].(*ir.AssignListStmt); ok {
788 budget += 4*len(n.Lhs) + 1
798 // InlineCalls/inlnode walks fn's statements and expressions and substitutes any
799 // calls made to inlineable functions. This is the external entry point.
800 func InlineCalls(fn *ir.Func, profile *pgo.Profile) {
801 if goexperiment.NewInliner && !fn.Wrapper() {
802 inlheur.ScoreCalls(fn)
804 if base.Debug.DumpInlFuncProps != "" && !fn.Wrapper() {
805 inlheur.DumpFuncProps(fn, base.Debug.DumpInlFuncProps,
806 func(fn *ir.Func) { CanInline(fn, profile) })
810 bigCaller := isBigFunc(fn)
811 if bigCaller && base.Flag.LowerM > 1 {
812 fmt.Printf("%v: function %v considered 'big'; reducing max cost of inlinees\n", ir.Line(fn), fn)
814 var inlCalls []*ir.InlinedCallExpr
815 var edit func(ir.Node) ir.Node
816 edit = func(n ir.Node) ir.Node {
817 return inlnode(fn, n, bigCaller, &inlCalls, edit, profile)
819 ir.EditChildren(fn, edit)
821 // If we inlined any calls, we want to recursively visit their
822 // bodies for further inlining. However, we need to wait until
823 // *after* the original function body has been expanded, or else
824 // inlCallee can have false positives (e.g., #54632).
825 for len(inlCalls) > 0 {
827 inlCalls = inlCalls[1:]
828 ir.EditChildren(call, edit)
834 // inlnode recurses over the tree to find inlineable calls, which will
835 // be turned into OINLCALLs by mkinlcall. When the recursion comes
836 // back up will examine left, right, list, rlist, ninit, ntest, nincr,
837 // nbody and nelse and use one of the 4 inlconv/glue functions above
838 // to turn the OINLCALL into an expression, a statement, or patch it
839 // in to this nodes list or rlist as appropriate.
840 // NOTE it makes no sense to pass the glue functions down the
841 // recursion to the level where the OINLCALL gets created because they
842 // have to edit /this/ n, so you'd have to push that one down as well,
843 // but then you may as well do it here. so this is cleaner and
844 // shorter and less complicated.
845 // The result of inlnode MUST be assigned back to n, e.g.
847 // n.Left = inlnode(n.Left)
848 func inlnode(callerfn *ir.Func, n ir.Node, bigCaller bool, inlCalls *[]*ir.InlinedCallExpr, edit func(ir.Node) ir.Node, profile *pgo.Profile) ir.Node {
854 case ir.ODEFER, ir.OGO:
855 n := n.(*ir.GoDeferStmt)
856 switch call := n.Call; call.Op() {
858 base.FatalfAt(call.Pos(), "OCALLMETH missed by typecheck")
860 call := call.(*ir.CallExpr)
864 n := n.(*ir.TailCallStmt)
865 n.Call.NoInline = true // Not inline a tail call for now. Maybe we could inline it just like RETURN fn(arg)?
867 // TODO do them here (or earlier),
868 // so escape analysis can avoid more heapmoves.
872 base.FatalfAt(n.Pos(), "OCALLMETH missed by typecheck")
874 n := n.(*ir.CallExpr)
875 if n.X.Op() == ir.OMETHEXPR {
876 // Prevent inlining some reflect.Value methods when using checkptr,
877 // even when package reflect was compiled without it (#35073).
878 if meth := ir.MethodExprName(n.X); meth != nil {
880 if base.Debug.Checkptr != 0 {
881 switch types.ReflectSymName(s) {
882 case "Value.UnsafeAddr", "Value.Pointer":
892 ir.EditChildren(n, edit)
894 // with all the branches out of the way, it is now time to
895 // transmogrify this node itself unless inhibited by the
896 // switch at the top of this function.
899 base.FatalfAt(n.Pos(), "OCALLMETH missed by typecheck")
902 call := n.(*ir.CallExpr)
906 if base.Flag.LowerM > 3 {
907 fmt.Printf("%v:call to func %+v\n", ir.Line(n), call.X)
909 if ir.IsIntrinsicCall(call) {
912 if fn := inlCallee(callerfn, call.X, profile); fn != nil && typecheck.HaveInlineBody(fn) {
913 n = mkinlcall(callerfn, call, fn, bigCaller, inlCalls)
922 // inlCallee takes a function-typed expression and returns the underlying function ONAME
923 // that it refers to if statically known. Otherwise, it returns nil.
924 func inlCallee(caller *ir.Func, fn ir.Node, profile *pgo.Profile) (res *ir.Func) {
925 fn = ir.StaticValue(fn)
928 fn := fn.(*ir.SelectorExpr)
929 n := ir.MethodExprName(fn)
930 // Check that receiver type matches fn.X.
931 // TODO(mdempsky): Handle implicit dereference
932 // of pointer receiver argument?
933 if n == nil || !types.Identical(n.Type().Recv().Type, fn.X.Type()) {
939 if fn.Class == ir.PFUNC {
943 fn := fn.(*ir.ClosureExpr)
945 if len(c.ClosureVars) != 0 && c.ClosureVars[0].Outer.Curfn != caller {
946 return nil // inliner doesn't support inlining across closure frames
948 CanInline(c, profile)
956 // SSADumpInline gives the SSA back end a chance to dump the function
957 // when producing output for debugging the compiler itself.
958 var SSADumpInline = func(*ir.Func) {}
960 // InlineCall allows the inliner implementation to be overridden.
961 // If it returns nil, the function will not be inlined.
962 var InlineCall = func(callerfn *ir.Func, call *ir.CallExpr, fn *ir.Func, inlIndex int) *ir.InlinedCallExpr {
963 base.Fatalf("inline.InlineCall not overridden")
967 // inlineCostOK returns true if call n from caller to callee is cheap enough to
968 // inline. bigCaller indicates that caller is a big function.
970 // If inlineCostOK returns false, it also returns the max cost that the callee
972 func inlineCostOK(n *ir.CallExpr, caller, callee *ir.Func, bigCaller bool) (bool, int32) {
973 maxCost := int32(inlineMaxBudget)
975 // We use this to restrict inlining into very big functions.
976 // See issue 26546 and 17566.
977 maxCost = inlineBigFunctionMaxCost
980 if callee.Inl.Cost <= maxCost {
981 // Simple case. Function is already cheap enough.
985 // We'll also allow inlining of hot functions below inlineHotMaxBudget,
986 // but only in small functions.
988 lineOffset := pgo.NodeLineOffset(n, caller)
989 csi := pgo.CallSiteInfo{LineOffset: lineOffset, Caller: caller}
990 if _, ok := candHotEdgeMap[csi]; !ok {
992 return false, maxCost
998 if base.Debug.PGODebug > 0 {
999 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))
1001 return false, maxCost
1004 if callee.Inl.Cost > inlineHotMaxBudget {
1005 return false, inlineHotMaxBudget
1008 if base.Debug.PGODebug > 0 {
1009 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))
1015 // If n is a OCALLFUNC node, and fn is an ONAME node for a
1016 // function with an inlinable body, return an OINLCALL node that can replace n.
1017 // The returned node's Ninit has the parameter assignments, the Nbody is the
1018 // inlined function body, and (List, Rlist) contain the (input, output)
1020 // The result of mkinlcall MUST be assigned back to n, e.g.
1022 // n.Left = mkinlcall(n.Left, fn, isddd)
1023 func mkinlcall(callerfn *ir.Func, n *ir.CallExpr, fn *ir.Func, bigCaller bool, inlCalls *[]*ir.InlinedCallExpr) ir.Node {
1025 if logopt.Enabled() {
1026 logopt.LogOpt(n.Pos(), "cannotInlineCall", "inline", ir.FuncName(callerfn),
1027 fmt.Sprintf("%s cannot be inlined", ir.PkgFuncName(fn)))
1032 if ok, maxCost := inlineCostOK(n, callerfn, fn, bigCaller); !ok {
1033 if logopt.Enabled() {
1034 logopt.LogOpt(n.Pos(), "cannotInlineCall", "inline", ir.FuncName(callerfn),
1035 fmt.Sprintf("cost %d of %s exceeds max caller cost %d", fn.Inl.Cost, ir.PkgFuncName(fn), maxCost))
1041 // Can't recursively inline a function into itself.
1042 if logopt.Enabled() {
1043 logopt.LogOpt(n.Pos(), "cannotInlineCall", "inline", fmt.Sprintf("recursive call to %s", ir.FuncName(callerfn)))
1048 if base.Flag.Cfg.Instrumenting && types.IsNoInstrumentPkg(fn.Sym().Pkg) {
1049 // Runtime package must not be instrumented.
1050 // Instrument skips runtime package. However, some runtime code can be
1051 // inlined into other packages and instrumented there. To avoid this,
1052 // we disable inlining of runtime functions when instrumenting.
1053 // The example that we observed is inlining of LockOSThread,
1054 // which lead to false race reports on m contents.
1057 if base.Flag.Race && types.IsNoRacePkg(fn.Sym().Pkg) {
1061 parent := base.Ctxt.PosTable.Pos(n.Pos()).Base().InliningIndex()
1064 // Check if we've already inlined this function at this particular
1065 // call site, in order to stop inlining when we reach the beginning
1066 // of a recursion cycle again. We don't inline immediately recursive
1067 // functions, but allow inlining if there is a recursion cycle of
1068 // many functions. Most likely, the inlining will stop before we
1069 // even hit the beginning of the cycle again, but this catches the
1071 for inlIndex := parent; inlIndex >= 0; inlIndex = base.Ctxt.InlTree.Parent(inlIndex) {
1072 if base.Ctxt.InlTree.InlinedFunction(inlIndex) == sym {
1073 if base.Flag.LowerM > 1 {
1074 fmt.Printf("%v: cannot inline %v into %v: repeated recursive cycle\n", ir.Line(n), fn, ir.FuncName(callerfn))
1080 typecheck.AssertFixedCall(n)
1082 inlIndex := base.Ctxt.InlTree.Add(parent, n.Pos(), sym, ir.FuncName(fn))
1084 closureInitLSym := func(n *ir.CallExpr, fn *ir.Func) {
1085 // The linker needs FuncInfo metadata for all inlined
1086 // functions. This is typically handled by gc.enqueueFunc
1087 // calling ir.InitLSym for all function declarations in
1088 // typecheck.Target.Decls (ir.UseClosure adds all closures to
1091 // However, non-trivial closures in Decls are ignored, and are
1092 // insteaded enqueued when walk of the calling function
1095 // This presents a problem for direct calls to closures.
1096 // Inlining will replace the entire closure definition with its
1097 // body, which hides the closure from walk and thus suppresses
1100 // Explicitly create a symbol early in this edge case to ensure
1101 // we keep this metadata.
1103 // TODO: Refactor to keep a reference so this can all be done
1106 if n.Op() != ir.OCALLFUNC {
1107 // Not a standard call.
1110 if n.X.Op() != ir.OCLOSURE {
1111 // Not a direct closure call.
1115 clo := n.X.(*ir.ClosureExpr)
1116 if ir.IsTrivialClosure(clo) {
1117 // enqueueFunc will handle trivial closures anyways.
1121 ir.InitLSym(fn, true)
1124 closureInitLSym(n, fn)
1126 if base.Flag.GenDwarfInl > 0 {
1127 if !sym.WasInlined() {
1128 base.Ctxt.DwFixups.SetPrecursorFunc(sym, fn)
1129 sym.Set(obj.AttrWasInlined, true)
1133 if base.Flag.LowerM != 0 {
1134 fmt.Printf("%v: inlining call to %v\n", ir.Line(n), fn)
1136 if base.Flag.LowerM > 2 {
1137 fmt.Printf("%v: Before inlining: %+v\n", ir.Line(n), n)
1140 res := InlineCall(callerfn, n, fn, inlIndex)
1143 base.FatalfAt(n.Pos(), "inlining call to %v failed", fn)
1146 if base.Flag.LowerM > 2 {
1147 fmt.Printf("%v: After inlining %+v\n\n", ir.Line(res), res)
1150 *inlCalls = append(*inlCalls, res)
1155 // CalleeEffects appends any side effects from evaluating callee to init.
1156 func CalleeEffects(init *ir.Nodes, callee ir.Node) {
1158 init.Append(ir.TakeInit(callee)...)
1160 switch callee.Op() {
1161 case ir.ONAME, ir.OCLOSURE, ir.OMETHEXPR:
1165 conv := callee.(*ir.ConvExpr)
1169 ic := callee.(*ir.InlinedCallExpr)
1170 init.Append(ic.Body.Take()...)
1171 callee = ic.SingleResult()
1174 base.FatalfAt(callee.Pos(), "unexpected callee expression: %v", callee)
1179 func pruneUnusedAutos(ll []*ir.Name, vis *hairyVisitor) []*ir.Name {
1180 s := make([]*ir.Name, 0, len(ll))
1181 for _, n := range ll {
1182 if n.Class == ir.PAUTO {
1183 if !vis.usedLocals.Has(n) {
1184 // TODO(mdempsky): Simplify code after confident that this
1185 // never happens anymore.
1186 base.FatalfAt(n.Pos(), "unused auto: %v", n)
1195 // numNonClosures returns the number of functions in list which are not closures.
1196 func numNonClosures(list []*ir.Func) int {
1198 for _, fn := range list {
1199 if fn.OClosure == nil {
1206 func doList(list []ir.Node, do func(ir.Node) bool) bool {
1207 for _, x := range list {
1217 // isIndexingCoverageCounter returns true if the specified node 'n' is indexing
1218 // into a coverage counter array.
1219 func isIndexingCoverageCounter(n ir.Node) bool {
1220 if n.Op() != ir.OINDEX {
1223 ixn := n.(*ir.IndexExpr)
1224 if ixn.X.Op() != ir.ONAME || !ixn.X.Type().IsArray() {
1227 nn := ixn.X.(*ir.Name)
1228 return nn.CoverageCounter()
1231 // isAtomicCoverageCounterUpdate examines the specified node to
1232 // determine whether it represents a call to sync/atomic.AddUint32 to
1233 // increment a coverage counter.
1234 func isAtomicCoverageCounterUpdate(cn *ir.CallExpr) bool {
1235 if cn.X.Op() != ir.ONAME {
1238 name := cn.X.(*ir.Name)
1239 if name.Class != ir.PFUNC {
1242 fn := name.Sym().Name
1243 if name.Sym().Pkg.Path != "sync/atomic" ||
1244 (fn != "AddUint32" && fn != "StoreUint32") {
1247 if len(cn.Args) != 2 || cn.Args[0].Op() != ir.OADDR {
1250 adn := cn.Args[0].(*ir.AddrExpr)
1251 v := isIndexingCoverageCounter(adn.X)