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"
35 "cmd/compile/internal/base"
36 "cmd/compile/internal/inline/inlheur"
37 "cmd/compile/internal/ir"
38 "cmd/compile/internal/logopt"
39 "cmd/compile/internal/pgo"
40 "cmd/compile/internal/typecheck"
41 "cmd/compile/internal/types"
45 // Inlining budget parameters, gathered in one place
48 inlineExtraAppendCost = 0
49 // default is to inline if there's at most one call. -l=4 overrides this by using 1 instead.
50 inlineExtraCallCost = 57 // 57 was benchmarked to provided most benefit with no bad surprises; see https://github.com/golang/go/issues/19348#issuecomment-439370742
51 inlineExtraPanicCost = 1 // do not penalize inlining panics.
52 inlineExtraThrowCost = inlineMaxBudget // with current (2018-05/1.11) code, inlining runtime.throw does not help.
54 inlineBigFunctionNodes = 5000 // Functions with this many nodes are considered "big".
55 inlineBigFunctionMaxCost = 20 // Max cost of inlinee when inlining into a "big" function.
59 // List of all hot callee nodes.
60 // TODO(prattmic): Make this non-global.
61 candHotCalleeMap = make(map[*pgo.IRNode]struct{})
63 // List of all hot call sites. CallSiteInfo.Callee is always nil.
64 // TODO(prattmic): Make this non-global.
65 candHotEdgeMap = make(map[pgo.CallSiteInfo]struct{})
67 // Threshold in percentage for hot callsite inlining.
68 inlineHotCallSiteThresholdPercent float64
70 // Threshold in CDF percentage for hot callsite inlining,
71 // that is, for a threshold of X the hottest callsites that
72 // make up the top X% of total edge weight will be
73 // considered hot for inlining candidates.
74 inlineCDFHotCallSiteThresholdPercent = float64(99)
76 // Budget increased due to hotness.
77 inlineHotMaxBudget int32 = 2000
80 // pgoInlinePrologue records the hot callsites from ir-graph.
81 func pgoInlinePrologue(p *pgo.Profile, funcs []*ir.Func) {
82 if base.Debug.PGOInlineCDFThreshold != "" {
83 if s, err := strconv.ParseFloat(base.Debug.PGOInlineCDFThreshold, 64); err == nil && s >= 0 && s <= 100 {
84 inlineCDFHotCallSiteThresholdPercent = s
86 base.Fatalf("invalid PGOInlineCDFThreshold, must be between 0 and 100")
89 var hotCallsites []pgo.NamedCallEdge
90 inlineHotCallSiteThresholdPercent, hotCallsites = hotNodesFromCDF(p)
91 if base.Debug.PGODebug > 0 {
92 fmt.Printf("hot-callsite-thres-from-CDF=%v\n", inlineHotCallSiteThresholdPercent)
95 if x := base.Debug.PGOInlineBudget; x != 0 {
96 inlineHotMaxBudget = int32(x)
99 for _, n := range hotCallsites {
100 // mark inlineable callees from hot edges
101 if callee := p.WeightedCG.IRNodes[n.CalleeName]; callee != nil {
102 candHotCalleeMap[callee] = struct{}{}
104 // mark hot call sites
105 if caller := p.WeightedCG.IRNodes[n.CallerName]; caller != nil && caller.AST != nil {
106 csi := pgo.CallSiteInfo{LineOffset: n.CallSiteOffset, Caller: caller.AST}
107 candHotEdgeMap[csi] = struct{}{}
111 if base.Debug.PGODebug >= 3 {
112 fmt.Printf("hot-cg before inline in dot format:")
113 p.PrintWeightedCallGraphDOT(inlineHotCallSiteThresholdPercent)
117 // hotNodesFromCDF computes an edge weight threshold and the list of hot
118 // nodes that make up the given percentage of the CDF. The threshold, as
119 // a percent, is the lower bound of weight for nodes to be considered hot
120 // (currently only used in debug prints) (in case of equal weights,
121 // comparing with the threshold may not accurately reflect which nodes are
123 func hotNodesFromCDF(p *pgo.Profile) (float64, []pgo.NamedCallEdge) {
125 for i, n := range p.NamedEdgeMap.ByWeight {
126 w := p.NamedEdgeMap.Weight[n]
128 if pgo.WeightInPercentage(cum, p.TotalWeight) > inlineCDFHotCallSiteThresholdPercent {
129 // nodes[:i+1] to include the very last node that makes it to go over the threshold.
130 // (Say, if the CDF threshold is 50% and one hot node takes 60% of weight, we want to
131 // include that node instead of excluding it.)
132 return pgo.WeightInPercentage(w, p.TotalWeight), p.NamedEdgeMap.ByWeight[:i+1]
135 return 0, p.NamedEdgeMap.ByWeight
138 // InlinePackage finds functions that can be inlined and clones them before walk expands them.
139 func InlinePackage(p *pgo.Profile) {
140 if base.Debug.PGOInline == 0 {
144 InlineDecls(p, typecheck.Target.Funcs, true)
146 // Perform a garbage collection of hidden closures functions that
147 // are no longer reachable from top-level functions following
148 // inlining. See #59404 and #59638 for more context.
149 garbageCollectUnreferencedHiddenClosures()
151 if base.Debug.DumpInlFuncProps != "" {
152 inlheur.DumpFuncProps(nil, base.Debug.DumpInlFuncProps, nil, inlineMaxBudget)
154 if goexperiment.NewInliner {
155 postProcessCallSites(p)
159 // InlineDecls applies inlining to the given batch of declarations.
160 func InlineDecls(p *pgo.Profile, funcs []*ir.Func, doInline bool) {
162 pgoInlinePrologue(p, funcs)
165 doCanInline := func(n *ir.Func, recursive bool, numfns int) {
166 if !recursive || numfns > 1 {
167 // We allow inlining if there is no
168 // recursion, or the recursion cycle is
169 // across more than one function.
172 if base.Flag.LowerM > 1 && n.OClosure == nil {
173 fmt.Printf("%v: cannot inline %v: recursive\n", ir.Line(n), n.Nname)
178 ir.VisitFuncsBottomUp(funcs, func(list []*ir.Func, recursive bool) {
179 numfns := numNonClosures(list)
180 // We visit functions within an SCC in fairly arbitrary order,
181 // so by computing inlinability for all functions in the SCC
182 // before performing any inlining, the results are less
183 // sensitive to the order within the SCC (see #58905 for an
186 // First compute inlinability for all functions in the SCC ...
187 for _, n := range list {
188 doCanInline(n, recursive, numfns)
190 // ... then make a second pass to do inlining of calls.
192 for _, n := range list {
199 // garbageCollectUnreferencedHiddenClosures makes a pass over all the
200 // top-level (non-hidden-closure) functions looking for nested closure
201 // functions that are reachable, then sweeps through the Target.Decls
202 // list and marks any non-reachable hidden closure function as dead.
203 // See issues #59404 and #59638 for more context.
204 func garbageCollectUnreferencedHiddenClosures() {
206 liveFuncs := make(map[*ir.Func]bool)
208 var markLiveFuncs func(fn *ir.Func)
209 markLiveFuncs = func(fn *ir.Func) {
214 ir.Visit(fn, func(n ir.Node) {
215 if clo, ok := n.(*ir.ClosureExpr); ok {
216 markLiveFuncs(clo.Func)
221 for i := 0; i < len(typecheck.Target.Funcs); i++ {
222 fn := typecheck.Target.Funcs[i]
223 if fn.IsHiddenClosure() {
229 for i := 0; i < len(typecheck.Target.Funcs); i++ {
230 fn := typecheck.Target.Funcs[i]
231 if !fn.IsHiddenClosure() {
234 if fn.IsDeadcodeClosure() {
240 fn.SetIsDeadcodeClosure(true)
241 if base.Flag.LowerM > 2 {
242 fmt.Printf("%v: unreferenced closure %v marked as dead\n", ir.Line(fn), fn)
244 if fn.Inl != nil && fn.LSym == nil {
245 ir.InitLSym(fn, true)
250 // inlineBudget determines the max budget for function 'fn' prior to
251 // analyzing the hairyness of the body of 'fn'. We pass in the pgo
252 // profile if available (which can change the budget), also a
253 // 'relaxed' flag, which expands the budget slightly to allow for the
254 // possibility that a call to the function might have its score
255 // adjusted downwards. If 'verbose' is set, then print a remark where
256 // we boost the budget due to PGO.
257 func inlineBudget(fn *ir.Func, profile *pgo.Profile, relaxed bool, verbose bool) int32 {
258 // Update the budget for profile-guided inlining.
259 budget := int32(inlineMaxBudget)
261 if n, ok := profile.WeightedCG.IRNodes[ir.LinkFuncName(fn)]; ok {
262 if _, ok := candHotCalleeMap[n]; ok {
263 budget = int32(inlineHotMaxBudget)
265 fmt.Printf("hot-node enabled increased budget=%v for func=%v\n", budget, ir.PkgFuncName(fn))
271 budget += inlineMaxBudget
276 // CanInline determines whether fn is inlineable.
277 // If so, CanInline saves copies of fn.Body and fn.Dcl in fn.Inl.
278 // fn and fn.Body will already have been typechecked.
279 func CanInline(fn *ir.Func, profile *pgo.Profile) {
281 base.Fatalf("CanInline no nname %+v", fn)
284 var funcProps *inlheur.FuncProps
285 if goexperiment.NewInliner || inlheur.UnitTesting() {
286 callCanInline := func(fn *ir.Func) { CanInline(fn, profile) }
287 funcProps = inlheur.AnalyzeFunc(fn, callCanInline, inlineMaxBudget)
288 budgetForFunc := func(fn *ir.Func) int32 {
289 return inlineBudget(fn, profile, true, false)
291 defer func() { inlheur.RevisitInlinability(fn, budgetForFunc) }()
294 var reason string // reason, if any, that the function was not inlined
295 if base.Flag.LowerM > 1 || logopt.Enabled() {
298 if base.Flag.LowerM > 1 {
299 fmt.Printf("%v: cannot inline %v: %s\n", ir.Line(fn), fn.Nname, reason)
301 if logopt.Enabled() {
302 logopt.LogOpt(fn.Pos(), "cannotInlineFunction", "inline", ir.FuncName(fn), reason)
308 reason = InlineImpossible(fn)
312 if fn.Typecheck() == 0 {
313 base.Fatalf("CanInline on non-typechecked function %v", fn)
317 if n.Func.InlinabilityChecked() {
320 defer n.Func.SetInlinabilityChecked(true)
322 cc := int32(inlineExtraCallCost)
323 if base.Flag.LowerL == 4 {
324 cc = 1 // this appears to yield better performance than 0.
327 // Used a "relaxed" inline budget if goexperiment.NewInliner is in
328 // effect, or if we're producing a debugging dump.
329 relaxed := goexperiment.NewInliner ||
330 (base.Debug.DumpInlFuncProps != "" ||
331 base.Debug.DumpInlCallSiteScores != 0)
333 // Compute the inline budget for this func.
334 budget := inlineBudget(fn, profile, relaxed, base.Debug.PGODebug > 0)
336 // At this point in the game the function we're looking at may
337 // have "stale" autos, vars that still appear in the Dcl list, but
338 // which no longer have any uses in the function body (due to
339 // elimination by deadcode). We'd like to exclude these dead vars
340 // when creating the "Inline.Dcl" field below; to accomplish this,
341 // the hairyVisitor below builds up a map of used/referenced
342 // locals, and we use this map to produce a pruned Inline.Dcl
343 // list. See issue 25459 for more context.
345 visitor := hairyVisitor{
347 isBigFunc: isBigFunc(fn),
353 if visitor.tooHairy(fn) {
354 reason = visitor.reason
358 n.Func.Inl = &ir.Inline{
359 Cost: budget - visitor.budget,
360 Dcl: pruneUnusedAutos(n.Func.Dcl, &visitor),
363 CanDelayResults: canDelayResults(fn),
365 if goexperiment.NewInliner {
366 n.Func.Inl.Properties = funcProps.SerializeToString()
369 if base.Flag.LowerM > 1 {
370 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))
371 } else if base.Flag.LowerM != 0 {
372 fmt.Printf("%v: can inline %v\n", ir.Line(fn), n)
374 if logopt.Enabled() {
375 logopt.LogOpt(fn.Pos(), "canInlineFunction", "inline", ir.FuncName(fn), fmt.Sprintf("cost: %d", budget-visitor.budget))
379 // InlineImpossible returns a non-empty reason string if fn is impossible to
380 // inline regardless of cost or contents.
381 func InlineImpossible(fn *ir.Func) string {
382 var reason string // reason, if any, that the function can not be inlined.
388 // If marked "go:noinline", don't inline.
389 if fn.Pragma&ir.Noinline != 0 {
390 reason = "marked go:noinline"
394 // If marked "go:norace" and -race compilation, don't inline.
395 if base.Flag.Race && fn.Pragma&ir.Norace != 0 {
396 reason = "marked go:norace with -race compilation"
400 // If marked "go:nocheckptr" and -d checkptr compilation, don't inline.
401 if base.Debug.Checkptr != 0 && fn.Pragma&ir.NoCheckPtr != 0 {
402 reason = "marked go:nocheckptr"
406 // If marked "go:cgo_unsafe_args", don't inline, since the function
407 // makes assumptions about its argument frame layout.
408 if fn.Pragma&ir.CgoUnsafeArgs != 0 {
409 reason = "marked go:cgo_unsafe_args"
413 // If marked as "go:uintptrkeepalive", don't inline, since the keep
414 // alive information is lost during inlining.
416 // TODO(prattmic): This is handled on calls during escape analysis,
417 // which is after inlining. Move prior to inlining so the keep-alive is
418 // maintained after inlining.
419 if fn.Pragma&ir.UintptrKeepAlive != 0 {
420 reason = "marked as having a keep-alive uintptr argument"
424 // If marked as "go:uintptrescapes", don't inline, since the escape
425 // information is lost during inlining.
426 if fn.Pragma&ir.UintptrEscapes != 0 {
427 reason = "marked as having an escaping uintptr argument"
431 // The nowritebarrierrec checker currently works at function
432 // granularity, so inlining yeswritebarrierrec functions can confuse it
433 // (#22342). As a workaround, disallow inlining them for now.
434 if fn.Pragma&ir.Yeswritebarrierrec != 0 {
435 reason = "marked go:yeswritebarrierrec"
439 // If a local function has no fn.Body (is defined outside of Go), cannot inline it.
440 // Imported functions don't have fn.Body but might have inline body in fn.Inl.
441 if len(fn.Body) == 0 && !typecheck.HaveInlineBody(fn) {
442 reason = "no function body"
449 // canDelayResults reports whether inlined calls to fn can delay
450 // declaring the result parameter until the "return" statement.
451 func canDelayResults(fn *ir.Func) bool {
452 // We can delay declaring+initializing result parameters if:
453 // (1) there's exactly one "return" statement in the inlined function;
454 // (2) it's not an empty return statement (#44355); and
455 // (3) the result parameters aren't named.
458 ir.VisitList(fn.Body, func(n ir.Node) {
459 if n, ok := n.(*ir.ReturnStmt); ok {
461 if len(n.Results) == 0 {
462 nreturns++ // empty return statement (case 2)
468 return false // not exactly one return statement (case 1)
471 // temporaries for return values.
472 for _, param := range fn.Type().Results() {
473 if sym := param.Sym; sym != nil && !sym.IsBlank() {
474 return false // found a named result parameter (case 3)
481 // hairyVisitor visits a function body to determine its inlining
482 // hairiness and whether or not it can be inlined.
483 type hairyVisitor struct {
484 // 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.Fun.Op() == ir.ONAME {
525 name := n.Fun.(*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.Fun.Op() == ir.OMETHEXPR {
557 if meth := ir.MethodExprName(n.Fun); 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 if ir.IsIntrinsicCall(n) {
589 // Treat like any other node.
593 if callee := inlCallee(v.curFunc, n.Fun, v.profile); callee != nil && typecheck.HaveInlineBody(callee) {
594 // Check whether we'd actually inline this call. Set
595 // log == false since we aren't actually doing inlining
597 if canInlineCallExpr(v.curFunc, n, callee, v.isBigFunc, false) {
598 // mkinlcall would inline this call [1], so use
599 // the cost of the inline body as the cost of
600 // the call, as that is what will actually
601 // appear in the code.
603 // [1] This is almost a perfect match to the
604 // mkinlcall logic, except that
605 // canInlineCallExpr considers inlining cycles
606 // by looking at what has already been inlined.
607 // Since we haven't done any inlining yet we
609 v.budget -= callee.Inl.Cost
614 // Call cost for non-leaf inlining.
615 v.budget -= v.extraCallCost
618 base.FatalfAt(n.Pos(), "OCALLMETH missed by typecheck")
620 // Things that are too hairy, irrespective of the budget
621 case ir.OCALL, ir.OCALLINTER:
622 // Call cost for non-leaf inlining.
623 v.budget -= v.extraCallCost
626 n := n.(*ir.UnaryExpr)
627 if n.X.Op() == ir.OCONVIFACE && n.X.(*ir.ConvExpr).Implicit() {
628 // Hack to keep reflect.flag.mustBe inlinable for TestIntendedInlining.
629 // Before CL 284412, these conversions were introduced later in the
630 // compiler, so they didn't count against inlining budget.
633 v.budget -= inlineExtraPanicCost
636 base.FatalfAt(n.Pos(), "ORECOVER missed typecheck")
638 // recover matches the argument frame pointer to find
639 // the right panic value, so it needs an argument frame.
640 v.reason = "call to recover"
644 if base.Debug.InlFuncsWithClosures == 0 {
645 v.reason = "not inlining functions with closures"
649 // TODO(danscales): Maybe make budget proportional to number of closure
651 //v.budget -= int32(len(n.(*ir.ClosureExpr).Func.ClosureVars) * 3)
652 // TODO(austin): However, if we're able to inline this closure into
653 // v.curFunc, then we actually pay nothing for the closure captures. We
654 // should try to account for that if we're going to account for captures.
657 case ir.OGO, ir.ODEFER, ir.OTAILCALL:
658 v.reason = "unhandled op " + n.Op().String()
662 v.budget -= inlineExtraAppendCost
665 n := n.(*ir.AddrExpr)
666 // Make "&s.f" cost 0 when f's offset is zero.
667 if dot, ok := n.X.(*ir.SelectorExpr); ok && (dot.Op() == ir.ODOT || dot.Op() == ir.ODOTPTR) {
668 if _, ok := dot.X.(*ir.Name); ok && dot.Selection.Offset == 0 {
669 v.budget += 2 // undo ir.OADDR+ir.ODOT/ir.ODOTPTR
674 // *(*X)(unsafe.Pointer(&x)) is low-cost
675 n := n.(*ir.StarExpr)
678 for ptr.Op() == ir.OCONVNOP {
679 ptr = ptr.(*ir.ConvExpr).X
681 if ptr.Op() == ir.OADDR {
682 v.budget += 1 // undo half of default cost of ir.ODEREF+ir.OADDR
686 // This doesn't produce code, but the children might.
687 v.budget++ // undo default cost
689 case ir.OFALL, ir.OTYPE:
690 // These nodes don't produce code; omit from inlining budget.
695 if ir.IsConst(n.Cond, constant.Bool) {
696 // This if and the condition cost nothing.
697 if doList(n.Init(), v.do) {
700 if ir.BoolVal(n.Cond) {
701 return doList(n.Body, v.do)
703 return doList(n.Else, v.do)
709 if n.Class == ir.PAUTO {
714 // The only OBLOCK we should see at this point is an empty one.
715 // In any event, let the visitList(n.List()) below take care of the statements,
716 // and don't charge for the OBLOCK itself. The ++ undoes the -- below.
719 case ir.OMETHVALUE, ir.OSLICELIT:
720 v.budget-- // Hack for toolstash -cmp.
723 v.budget++ // Hack for toolstash -cmp.
726 n := n.(*ir.AssignListStmt)
728 // Unified IR unconditionally rewrites:
739 // so that it can insert implicit conversions as necessary. To
740 // minimize impact to the existing inlining heuristics (in
741 // particular, to avoid breaking the existing inlinability regress
742 // tests), we need to compensate for this here.
744 // See also identical logic in isBigFunc.
745 if init := n.Rhs[0].Init(); len(init) == 1 {
746 if _, ok := init[0].(*ir.AssignListStmt); ok {
747 // 4 for each value, because each temporary variable now
748 // appears 3 times (DCL, LHS, RHS), plus an extra DCL node.
750 // 1 for the extra "tmp1, tmp2 = f()" assignment statement.
751 v.budget += 4*int32(len(n.Lhs)) + 1
756 // Special case for coverage counter updates and coverage
757 // function registrations. Although these correspond to real
758 // operations, we treat them as zero cost for the moment. This
759 // is primarily due to the existence of tests that are
760 // sensitive to inlining-- if the insertion of coverage
761 // instrumentation happens to tip a given function over the
762 // threshold and move it from "inlinable" to "not-inlinable",
763 // this can cause changes in allocation behavior, which can
764 // then result in test failures (a good example is the
765 // TestAllocations in crypto/ed25519).
766 n := n.(*ir.AssignStmt)
767 if n.X.Op() == ir.OINDEX && isIndexingCoverageCounter(n.X) {
774 // When debugging, don't stop early, to get full cost of inlining this function
775 if v.budget < 0 && base.Flag.LowerM < 2 && !logopt.Enabled() {
776 v.reason = "too expensive"
780 return ir.DoChildren(n, v.do)
783 func isBigFunc(fn *ir.Func) bool {
784 budget := inlineBigFunctionNodes
785 return ir.Any(fn, func(n ir.Node) bool {
786 // See logic in hairyVisitor.doNode, explaining unified IR's
787 // handling of "a, b = f()" assignments.
788 if n, ok := n.(*ir.AssignListStmt); ok && n.Op() == ir.OAS2 {
789 if init := n.Rhs[0].Init(); len(init) == 1 {
790 if _, ok := init[0].(*ir.AssignListStmt); ok {
791 budget += 4*len(n.Lhs) + 1
801 // InlineCalls/inlnode walks fn's statements and expressions and substitutes any
802 // calls made to inlineable functions. This is the external entry point.
803 func InlineCalls(fn *ir.Func, profile *pgo.Profile) {
804 if goexperiment.NewInliner && !fn.Wrapper() {
805 inlheur.ScoreCalls(fn)
807 if base.Debug.DumpInlFuncProps != "" && !fn.Wrapper() {
808 inlheur.DumpFuncProps(fn, base.Debug.DumpInlFuncProps,
809 func(fn *ir.Func) { CanInline(fn, profile) }, inlineMaxBudget)
813 bigCaller := isBigFunc(fn)
814 if bigCaller && base.Flag.LowerM > 1 {
815 fmt.Printf("%v: function %v considered 'big'; reducing max cost of inlinees\n", ir.Line(fn), fn)
817 var inlCalls []*ir.InlinedCallExpr
818 var edit func(ir.Node) ir.Node
819 edit = func(n ir.Node) ir.Node {
820 return inlnode(fn, n, bigCaller, &inlCalls, edit, profile)
822 ir.EditChildren(fn, edit)
824 // If we inlined any calls, we want to recursively visit their
825 // bodies for further inlining. However, we need to wait until
826 // *after* the original function body has been expanded, or else
827 // inlCallee can have false positives (e.g., #54632).
828 for len(inlCalls) > 0 {
830 inlCalls = inlCalls[1:]
831 ir.EditChildren(call, edit)
837 // inlnode recurses over the tree to find inlineable calls, which will
838 // be turned into OINLCALLs by mkinlcall. When the recursion comes
839 // back up will examine left, right, list, rlist, ninit, ntest, nincr,
840 // nbody and nelse and use one of the 4 inlconv/glue functions above
841 // to turn the OINLCALL into an expression, a statement, or patch it
842 // in to this nodes list or rlist as appropriate.
843 // NOTE it makes no sense to pass the glue functions down the
844 // recursion to the level where the OINLCALL gets created because they
845 // have to edit /this/ n, so you'd have to push that one down as well,
846 // but then you may as well do it here. so this is cleaner and
847 // shorter and less complicated.
848 // The result of inlnode MUST be assigned back to n, e.g.
850 // n.Left = inlnode(n.Left)
851 func inlnode(callerfn *ir.Func, n ir.Node, bigCaller bool, inlCalls *[]*ir.InlinedCallExpr, edit func(ir.Node) ir.Node, profile *pgo.Profile) ir.Node {
857 case ir.ODEFER, ir.OGO:
858 n := n.(*ir.GoDeferStmt)
859 switch call := n.Call; call.Op() {
861 base.FatalfAt(call.Pos(), "OCALLMETH missed by typecheck")
863 call := call.(*ir.CallExpr)
867 n := n.(*ir.TailCallStmt)
868 n.Call.NoInline = true // Not inline a tail call for now. Maybe we could inline it just like RETURN fn(arg)?
870 // TODO do them here (or earlier),
871 // so escape analysis can avoid more heapmoves.
875 base.FatalfAt(n.Pos(), "OCALLMETH missed by typecheck")
877 n := n.(*ir.CallExpr)
878 if n.Fun.Op() == ir.OMETHEXPR {
879 // Prevent inlining some reflect.Value methods when using checkptr,
880 // even when package reflect was compiled without it (#35073).
881 if meth := ir.MethodExprName(n.Fun); meth != nil {
883 if base.Debug.Checkptr != 0 {
884 switch types.ReflectSymName(s) {
885 case "Value.UnsafeAddr", "Value.Pointer":
895 ir.EditChildren(n, edit)
897 // with all the branches out of the way, it is now time to
898 // transmogrify this node itself unless inhibited by the
899 // switch at the top of this function.
902 base.FatalfAt(n.Pos(), "OCALLMETH missed by typecheck")
905 call := n.(*ir.CallExpr)
909 if base.Flag.LowerM > 3 {
910 fmt.Printf("%v:call to func %+v\n", ir.Line(n), call.Fun)
912 if ir.IsIntrinsicCall(call) {
915 if fn := inlCallee(callerfn, call.Fun, profile); fn != nil && typecheck.HaveInlineBody(fn) {
916 n = mkinlcall(callerfn, call, fn, bigCaller, inlCalls)
925 // inlCallee takes a function-typed expression and returns the underlying function ONAME
926 // that it refers to if statically known. Otherwise, it returns nil.
927 func inlCallee(caller *ir.Func, fn ir.Node, profile *pgo.Profile) (res *ir.Func) {
928 fn = ir.StaticValue(fn)
931 fn := fn.(*ir.SelectorExpr)
932 n := ir.MethodExprName(fn)
933 // Check that receiver type matches fn.X.
934 // TODO(mdempsky): Handle implicit dereference
935 // of pointer receiver argument?
936 if n == nil || !types.Identical(n.Type().Recv().Type, fn.X.Type()) {
942 if fn.Class == ir.PFUNC {
946 fn := fn.(*ir.ClosureExpr)
948 if len(c.ClosureVars) != 0 && c.ClosureVars[0].Outer.Curfn != caller {
949 return nil // inliner doesn't support inlining across closure frames
951 CanInline(c, profile)
959 // SSADumpInline gives the SSA back end a chance to dump the function
960 // when producing output for debugging the compiler itself.
961 var SSADumpInline = func(*ir.Func) {}
963 // InlineCall allows the inliner implementation to be overridden.
964 // If it returns nil, the function will not be inlined.
965 var InlineCall = func(callerfn *ir.Func, call *ir.CallExpr, fn *ir.Func, inlIndex int) *ir.InlinedCallExpr {
966 base.Fatalf("inline.InlineCall not overridden")
970 // inlineCostOK returns true if call n from caller to callee is cheap enough to
971 // inline. bigCaller indicates that caller is a big function.
973 // If inlineCostOK returns false, it also returns the max cost that the callee
975 func inlineCostOK(n *ir.CallExpr, caller, callee *ir.Func, bigCaller bool) (bool, int32) {
976 maxCost := int32(inlineMaxBudget)
978 // We use this to restrict inlining into very big functions.
979 // See issue 26546 and 17566.
980 maxCost = inlineBigFunctionMaxCost
983 metric := callee.Inl.Cost
984 if goexperiment.NewInliner {
985 score, ok := inlheur.GetCallSiteScore(caller, n)
987 metric = int32(score)
992 if metric <= maxCost {
993 // Simple case. Function is already cheap enough.
997 // We'll also allow inlining of hot functions below inlineHotMaxBudget,
998 // but only in small functions.
1000 lineOffset := pgo.NodeLineOffset(n, caller)
1001 csi := pgo.CallSiteInfo{LineOffset: lineOffset, Caller: caller}
1002 if _, ok := candHotEdgeMap[csi]; !ok {
1004 return false, maxCost
1010 if base.Debug.PGODebug > 0 {
1011 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))
1013 return false, maxCost
1016 if metric > inlineHotMaxBudget {
1017 return false, inlineHotMaxBudget
1020 if !base.PGOHash.MatchPosWithInfo(n.Pos(), "inline", nil) {
1021 // De-selected by PGO Hash.
1022 return false, maxCost
1025 if base.Debug.PGODebug > 0 {
1026 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))
1032 // canInlineCallsite returns true if the call n from caller to callee can be
1033 // inlined. bigCaller indicates that caller is a big function. log indicates
1034 // that the 'cannot inline' reason should be logged.
1036 // Preconditions: CanInline(callee) has already been called.
1037 func canInlineCallExpr(callerfn *ir.Func, n *ir.CallExpr, callee *ir.Func, bigCaller bool, log bool) bool {
1038 if callee.Inl == nil {
1039 // callee is never inlinable.
1040 if log && logopt.Enabled() {
1041 logopt.LogOpt(n.Pos(), "cannotInlineCall", "inline", ir.FuncName(callerfn),
1042 fmt.Sprintf("%s cannot be inlined", ir.PkgFuncName(callee)))
1047 if ok, maxCost := inlineCostOK(n, callerfn, callee, bigCaller); !ok {
1048 // callee cost too high for this call site.
1049 if log && logopt.Enabled() {
1050 logopt.LogOpt(n.Pos(), "cannotInlineCall", "inline", ir.FuncName(callerfn),
1051 fmt.Sprintf("cost %d of %s exceeds max caller cost %d", callee.Inl.Cost, ir.PkgFuncName(callee), maxCost))
1056 if callee == callerfn {
1057 // Can't recursively inline a function into itself.
1058 if log && logopt.Enabled() {
1059 logopt.LogOpt(n.Pos(), "cannotInlineCall", "inline", fmt.Sprintf("recursive call to %s", ir.FuncName(callerfn)))
1064 if base.Flag.Cfg.Instrumenting && types.IsNoInstrumentPkg(callee.Sym().Pkg) {
1065 // Runtime package must not be instrumented.
1066 // Instrument skips runtime package. However, some runtime code can be
1067 // inlined into other packages and instrumented there. To avoid this,
1068 // we disable inlining of runtime functions when instrumenting.
1069 // The example that we observed is inlining of LockOSThread,
1070 // which lead to false race reports on m contents.
1071 if log && logopt.Enabled() {
1072 logopt.LogOpt(n.Pos(), "cannotInlineCall", "inline", ir.FuncName(callerfn),
1073 fmt.Sprintf("call to runtime function %s in instrumented build", ir.PkgFuncName(callee)))
1078 if base.Flag.Race && types.IsNoRacePkg(callee.Sym().Pkg) {
1079 if log && logopt.Enabled() {
1080 logopt.LogOpt(n.Pos(), "cannotInlineCall", "inline", ir.FuncName(callerfn),
1081 fmt.Sprintf(`call to into "no-race" package function %s in race build`, ir.PkgFuncName(callee)))
1086 // Check if we've already inlined this function at this particular
1087 // call site, in order to stop inlining when we reach the beginning
1088 // of a recursion cycle again. We don't inline immediately recursive
1089 // functions, but allow inlining if there is a recursion cycle of
1090 // many functions. Most likely, the inlining will stop before we
1091 // even hit the beginning of the cycle again, but this catches the
1093 parent := base.Ctxt.PosTable.Pos(n.Pos()).Base().InliningIndex()
1094 sym := callee.Linksym()
1095 for inlIndex := parent; inlIndex >= 0; inlIndex = base.Ctxt.InlTree.Parent(inlIndex) {
1096 if base.Ctxt.InlTree.InlinedFunction(inlIndex) == sym {
1098 if base.Flag.LowerM > 1 {
1099 fmt.Printf("%v: cannot inline %v into %v: repeated recursive cycle\n", ir.Line(n), callee, ir.FuncName(callerfn))
1101 if logopt.Enabled() {
1102 logopt.LogOpt(n.Pos(), "cannotInlineCall", "inline", ir.FuncName(callerfn),
1103 fmt.Sprintf("repeated recursive cycle to %s", ir.PkgFuncName(callee)))
1113 // If n is a OCALLFUNC node, and fn is an ONAME node for a
1114 // function with an inlinable body, return an OINLCALL node that can replace n.
1115 // The returned node's Ninit has the parameter assignments, the Nbody is the
1116 // inlined function body, and (List, Rlist) contain the (input, output)
1118 // The result of mkinlcall MUST be assigned back to n, e.g.
1120 // n.Left = mkinlcall(n.Left, fn, isddd)
1121 func mkinlcall(callerfn *ir.Func, n *ir.CallExpr, fn *ir.Func, bigCaller bool, inlCalls *[]*ir.InlinedCallExpr) ir.Node {
1122 if !canInlineCallExpr(callerfn, n, fn, bigCaller, true) {
1125 typecheck.AssertFixedCall(n)
1127 parent := base.Ctxt.PosTable.Pos(n.Pos()).Base().InliningIndex()
1129 inlIndex := base.Ctxt.InlTree.Add(parent, n.Pos(), sym, ir.FuncName(fn))
1131 closureInitLSym := func(n *ir.CallExpr, fn *ir.Func) {
1132 // The linker needs FuncInfo metadata for all inlined
1133 // functions. This is typically handled by gc.enqueueFunc
1134 // calling ir.InitLSym for all function declarations in
1135 // typecheck.Target.Decls (ir.UseClosure adds all closures to
1138 // However, non-trivial closures in Decls are ignored, and are
1139 // insteaded enqueued when walk of the calling function
1142 // This presents a problem for direct calls to closures.
1143 // Inlining will replace the entire closure definition with its
1144 // body, which hides the closure from walk and thus suppresses
1147 // Explicitly create a symbol early in this edge case to ensure
1148 // we keep this metadata.
1150 // TODO: Refactor to keep a reference so this can all be done
1153 if n.Op() != ir.OCALLFUNC {
1154 // Not a standard call.
1157 if n.Fun.Op() != ir.OCLOSURE {
1158 // Not a direct closure call.
1162 clo := n.Fun.(*ir.ClosureExpr)
1163 if ir.IsTrivialClosure(clo) {
1164 // enqueueFunc will handle trivial closures anyways.
1168 ir.InitLSym(fn, true)
1171 closureInitLSym(n, fn)
1173 if base.Flag.GenDwarfInl > 0 {
1174 if !sym.WasInlined() {
1175 base.Ctxt.DwFixups.SetPrecursorFunc(sym, fn)
1176 sym.Set(obj.AttrWasInlined, true)
1180 if base.Flag.LowerM != 0 {
1181 fmt.Printf("%v: inlining call to %v\n", ir.Line(n), fn)
1183 if base.Flag.LowerM > 2 {
1184 fmt.Printf("%v: Before inlining: %+v\n", ir.Line(n), n)
1187 res := InlineCall(callerfn, n, fn, inlIndex)
1190 base.FatalfAt(n.Pos(), "inlining call to %v failed", fn)
1193 if base.Flag.LowerM > 2 {
1194 fmt.Printf("%v: After inlining %+v\n\n", ir.Line(res), res)
1197 *inlCalls = append(*inlCalls, res)
1202 // CalleeEffects appends any side effects from evaluating callee to init.
1203 func CalleeEffects(init *ir.Nodes, callee ir.Node) {
1205 init.Append(ir.TakeInit(callee)...)
1207 switch callee.Op() {
1208 case ir.ONAME, ir.OCLOSURE, ir.OMETHEXPR:
1212 conv := callee.(*ir.ConvExpr)
1216 ic := callee.(*ir.InlinedCallExpr)
1217 init.Append(ic.Body.Take()...)
1218 callee = ic.SingleResult()
1221 base.FatalfAt(callee.Pos(), "unexpected callee expression: %v", callee)
1226 func pruneUnusedAutos(ll []*ir.Name, vis *hairyVisitor) []*ir.Name {
1227 s := make([]*ir.Name, 0, len(ll))
1228 for _, n := range ll {
1229 if n.Class == ir.PAUTO {
1230 if !vis.usedLocals.Has(n) {
1231 // TODO(mdempsky): Simplify code after confident that this
1232 // never happens anymore.
1233 base.FatalfAt(n.Pos(), "unused auto: %v", n)
1242 // numNonClosures returns the number of functions in list which are not closures.
1243 func numNonClosures(list []*ir.Func) int {
1245 for _, fn := range list {
1246 if fn.OClosure == nil {
1253 func doList(list []ir.Node, do func(ir.Node) bool) bool {
1254 for _, x := range list {
1264 // isIndexingCoverageCounter returns true if the specified node 'n' is indexing
1265 // into a coverage counter array.
1266 func isIndexingCoverageCounter(n ir.Node) bool {
1267 if n.Op() != ir.OINDEX {
1270 ixn := n.(*ir.IndexExpr)
1271 if ixn.X.Op() != ir.ONAME || !ixn.X.Type().IsArray() {
1274 nn := ixn.X.(*ir.Name)
1275 return nn.CoverageCounter()
1278 // isAtomicCoverageCounterUpdate examines the specified node to
1279 // determine whether it represents a call to sync/atomic.AddUint32 to
1280 // increment a coverage counter.
1281 func isAtomicCoverageCounterUpdate(cn *ir.CallExpr) bool {
1282 if cn.Fun.Op() != ir.ONAME {
1285 name := cn.Fun.(*ir.Name)
1286 if name.Class != ir.PFUNC {
1289 fn := name.Sym().Name
1290 if name.Sym().Pkg.Path != "sync/atomic" ||
1291 (fn != "AddUint32" && fn != "StoreUint32") {
1294 if len(cn.Args) != 2 || cn.Args[0].Op() != ir.OADDR {
1297 adn := cn.Args[0].(*ir.AddrExpr)
1298 v := isIndexingCoverageCounter(adn.X)
1302 func postProcessCallSites(profile *pgo.Profile) {
1303 if base.Debug.DumpInlCallSiteScores != 0 {
1304 budgetCallback := func(fn *ir.Func, prof *pgo.Profile) (int32, bool) {
1305 v := inlineBudget(fn, prof, false, false)
1306 return v, v == inlineHotMaxBudget
1308 inlheur.DumpInlCallSiteScores(profile, budgetCallback)