cmd/compile: adjust PGO inlining default parameters

Adjust PGO inlining default parameters to 99% CDF threshold and
2000 budget. Benchmark results (mostly from Sweet) show that this
set of parameters performs reasonably well, with a few percent
speedup at the cost of a few percent binary size increase.

Also rename the debug flags to start with "pgo", to make it clear
that they are related to PGO.

Change-Id: I0749249b1298d1dc55a28993c37b3185f9d7639d
Reviewed-on: https://go-review.googlesource.com/c/go/+/449477
Run-TryBot: Cherry Mui <[email protected]>
Reviewed-by: Michael Pratt <[email protected]>
TryBot-Result: Gopher Robot <[email protected]>

Author: cherrymui

src/cmd/compile/internal/base/debug.go

@@ -16,41 +16,41 @@ var Debug DebugFlags
 // The -d option takes a comma-separated list of settings.
 // Each setting is name=value; for ints, name is short for name=1.
 type DebugFlags struct {
-	Append                        int    `help:"print information about append compilation"`
-	Checkptr                      int    `help:"instrument unsafe pointer conversions\n0: instrumentation disabled\n1: conversions involving unsafe.Pointer are instrumented\n2: conversions to unsafe.Pointer force heap allocation" concurrent:"ok"`
-	Closure                       int    `help:"print information about closure compilation"`
-	DclStack                      int    `help:"run internal dclstack check"`
-	Defer                         int    `help:"print information about defer compilation"`
-	DisableNil                    int    `help:"disable nil checks" concurrent:"ok"`
-	DumpPtrs                      int    `help:"show Node pointers values in dump output"`
-	DwarfInl                      int    `help:"print information about DWARF inlined function creation"`
-	Export                        int    `help:"print export data"`
-	Fmahash                       string `help:"hash value for use in debugging platform-dependent multiply-add use" concurrent:"ok"`
-	GCAdjust                      int    `help:"log adjustments to GOGC" concurrent:"ok"`
-	GCProg                        int    `help:"print dump of GC programs"`
-	Gossahash                     string `help:"hash value for use in debugging the compiler"`
-	InlFuncsWithClosures          int    `help:"allow functions with closures to be inlined" concurrent:"ok"`
-	Libfuzzer                     int    `help:"enable coverage instrumentation for libfuzzer"`
-	LocationLists                 int    `help:"print information about DWARF location list creation"`
-	Nil                           int    `help:"print information about nil checks"`
-	NoOpenDefer                   int    `help:"disable open-coded defers" concurrent:"ok"`
-	NoRefName                     int    `help:"do not include referenced symbol names in object file" concurrent:"ok"`
-	PCTab                         string `help:"print named pc-value table\nOne of: pctospadj, pctofile, pctoline, pctoinline, pctopcdata"`
-	Panic                         int    `help:"show all compiler panics"`
-	Reshape                       int    `help:"print information about expression reshaping"`
-	Shapify                       int    `help:"print information about shaping recursive types"`
-	Slice                         int    `help:"print information about slice compilation"`
-	SoftFloat                     int    `help:"force compiler to emit soft-float code" concurrent:"ok"`
-	SyncFrames                    int    `help:"how many writer stack frames to include at sync points in unified export data"`
-	TypeAssert                    int    `help:"print information about type assertion inlining"`
-	TypecheckInl                  int    `help:"eager typechecking of inline function bodies" concurrent:"ok"`
-	Unified                       int    `help:"enable unified IR construction"`
-	WB                            int    `help:"print information about write barriers"`
-	ABIWrap                       int    `help:"print information about ABI wrapper generation"`
-	MayMoreStack                  string `help:"call named function before all stack growth checks" concurrent:"ok"`
-	InlineHotCallSiteCDFThreshold string `help:"cummulative threshold percentage for determining call sites as hot candidates for inlining" concurrent:"ok"`
-	InlineHotBudget               int    `help:"inline budget for hot functions" concurrent:"ok"`
-	PGOInline                     int    `help:"debug profile-guided inlining"`
+	Append                int    `help:"print information about append compilation"`
+	Checkptr              int    `help:"instrument unsafe pointer conversions\n0: instrumentation disabled\n1: conversions involving unsafe.Pointer are instrumented\n2: conversions to unsafe.Pointer force heap allocation" concurrent:"ok"`
+	Closure               int    `help:"print information about closure compilation"`
+	DclStack              int    `help:"run internal dclstack check"`
+	Defer                 int    `help:"print information about defer compilation"`
+	DisableNil            int    `help:"disable nil checks" concurrent:"ok"`
+	DumpPtrs              int    `help:"show Node pointers values in dump output"`
+	DwarfInl              int    `help:"print information about DWARF inlined function creation"`
+	Export                int    `help:"print export data"`
+	Fmahash               string `help:"hash value for use in debugging platform-dependent multiply-add use" concurrent:"ok"`
+	GCAdjust              int    `help:"log adjustments to GOGC" concurrent:"ok"`
+	GCProg                int    `help:"print dump of GC programs"`
+	Gossahash             string `help:"hash value for use in debugging the compiler"`
+	InlFuncsWithClosures  int    `help:"allow functions with closures to be inlined" concurrent:"ok"`
+	Libfuzzer             int    `help:"enable coverage instrumentation for libfuzzer"`
+	LocationLists         int    `help:"print information about DWARF location list creation"`
+	Nil                   int    `help:"print information about nil checks"`
+	NoOpenDefer           int    `help:"disable open-coded defers" concurrent:"ok"`
+	NoRefName             int    `help:"do not include referenced symbol names in object file" concurrent:"ok"`
+	PCTab                 string `help:"print named pc-value table\nOne of: pctospadj, pctofile, pctoline, pctoinline, pctopcdata"`
+	Panic                 int    `help:"show all compiler panics"`
+	Reshape               int    `help:"print information about expression reshaping"`
+	Shapify               int    `help:"print information about shaping recursive types"`
+	Slice                 int    `help:"print information about slice compilation"`
+	SoftFloat             int    `help:"force compiler to emit soft-float code" concurrent:"ok"`
+	SyncFrames            int    `help:"how many writer stack frames to include at sync points in unified export data"`
+	TypeAssert            int    `help:"print information about type assertion inlining"`
+	TypecheckInl          int    `help:"eager typechecking of inline function bodies" concurrent:"ok"`
+	Unified               int    `help:"enable unified IR construction"`
+	WB                    int    `help:"print information about write barriers"`
+	ABIWrap               int    `help:"print information about ABI wrapper generation"`
+	MayMoreStack          string `help:"call named function before all stack growth checks" concurrent:"ok"`
+	PGOInlineCDFThreshold string `help:"cummulative threshold percentage for determining call sites as hot candidates for inlining" concurrent:"ok"`
+	PGOInlineBudget       int    `help:"inline budget for hot functions" concurrent:"ok"`
+	PGOInline             int    `help:"debug profile-guided inlining"`
 
 	ConcurrentOk bool // true if only concurrentOk flags seen
 }

src/cmd/compile/internal/inline/inl.go

@@ -76,15 +76,15 @@ var (
 	// that is, for a threshold of X the hottest callsites that
 	// make up the top X% of total edge weight will be
 	// considered hot for inlining candidates.
-	inlineCDFHotCallSiteThresholdPercent = float64(95)
+	inlineCDFHotCallSiteThresholdPercent = float64(99)
 
 	// Budget increased due to hotness.
-	inlineHotMaxBudget int32 = 160
+	inlineHotMaxBudget int32 = 2000
 )
 
 // pgoInlinePrologue records the hot callsites from ir-graph.
 func pgoInlinePrologue(p *pgo.Profile, decls []ir.Node) {
-	if s, err := strconv.ParseFloat(base.Debug.InlineHotCallSiteCDFThreshold, 64); err == nil {
+	if s, err := strconv.ParseFloat(base.Debug.PGOInlineCDFThreshold, 64); err == nil {
 		inlineCDFHotCallSiteThresholdPercent = s
 	}
 	var hotCallsites []pgo.NodeMapKey
@@ -93,8 +93,8 @@ func pgoInlinePrologue(p *pgo.Profile, decls []ir.Node) {
 		fmt.Printf("hot-callsite-thres-from-CDF=%v\n", inlineHotCallSiteThresholdPercent)
 	}
 
-	if base.Debug.InlineHotBudget != 0 {
-		inlineHotMaxBudget = int32(base.Debug.InlineHotBudget)
+	if x := base.Debug.PGOInlineBudget; x != 0 {
+		inlineHotMaxBudget = int32(x)
 	}
 
 	// mark inlineable callees from hot edges

src/cmd/compile/internal/test/pgo_inl_test.go

@@ -67,9 +67,10 @@ go 1.19
 		expectedNotInlinedList[fullName] = struct{}{}
 	}
 
-	// go test -c -o /tmp/test.exe -cpuprofile inline_hot.pprof
+	// Build the test with the profile. Use a smaller threshold to test.
+	// TODO: maybe adjust the test to work with default threshold.
 	pprof := filepath.Join(dir, "inline_hot.pprof")
-	gcflag := fmt.Sprintf("-gcflags=-m -m -pgoprofile %s", pprof)
+	gcflag := fmt.Sprintf("-gcflags=-m -m -pgoprofile=%s -d=pgoinlinebudget=160,pgoinlinecdfthreshold=90", pprof)
 	out := filepath.Join(dir, "test.exe")
 	cmd := testenv.CleanCmdEnv(exec.Command(testenv.GoToolPath(t), "test", "-c", "-o", out, gcflag, "."))
 	cmd.Dir = dir