cmd/compile: reduce allocations in prove by reusing posets

In prove, reuse posets between different functions by storing them
in the per-worker cache.

Allocation count regression caused by prove improvements is down
from 5% to 3% after this CL.

Updates #25179

Change-Id: I6d14003109833d9b3ef5165fdea00aa9c9e952e8
Reviewed-on: https://go-review.googlesource.com/110455
Run-TryBot: Giovanni Bajo <[email protected]>
TryBot-Result: Gobot Gobot <[email protected]>
Reviewed-by: David Chase <[email protected]>

Author: rasky

src/cmd/compile/internal/ssa/cache.go

@@ -24,6 +24,7 @@ type Cache struct {
 	domblockstore []ID         // scratch space for computing dominators
 	scrSparseSet  []*sparseSet // scratch sparse sets to be re-used.
 	scrSparseMap  []*sparseMap // scratch sparse maps to be re-used.
+	scrPoset      []*poset     // scratch poset to be reused
 
 	ValueToProgAfter []*obj.Prog
 	debugState       debugState

src/cmd/compile/internal/ssa/func.go

@@ -130,6 +130,21 @@ func (f *Func) retSparseMap(ss *sparseMap) {
 	f.Cache.scrSparseMap = append(f.Cache.scrSparseMap, ss)
 }
 
+// newPoset returns a new poset from the internal cache
+func (f *Func) newPoset() *poset {
+	if len(f.Cache.scrPoset) > 0 {
+		po := f.Cache.scrPoset[len(f.Cache.scrPoset)-1]
+		f.Cache.scrPoset = f.Cache.scrPoset[:len(f.Cache.scrPoset)-1]
+		return po
+	}
+	return newPoset()
+}
+
+// retPoset returns a poset to the internal cache
+func (f *Func) retPoset(po *poset) {
+	f.Cache.scrPoset = append(f.Cache.scrPoset, po)
+}
+
 // newValue allocates a new Value with the given fields and places it at the end of b.Values.
 func (f *Func) newValue(op Op, t *types.Type, b *Block, pos src.XPos) *Value {
 	var v *Value

src/cmd/compile/internal/ssa/poset.go

@@ -152,13 +152,8 @@ type poset struct {
 	undo      []posetUndo   // undo chain
 }
 
-func newPoset(unsigned bool) *poset {
-	var flags uint8
-	if unsigned {
-		flags |= posetFlagUnsigned
-	}
+func newPoset() *poset {
 	return &poset{
-		flags:     flags,
 		values:    make(map[ID]uint32),
 		constants: make([]*Value, 0, 8),
 		nodes:     make([]posetNode, 1, 16),
@@ -168,6 +163,14 @@ func newPoset(unsigned bool) *poset {
 	}
 }
 
+func (po *poset) SetUnsigned(uns bool) {
+	if uns {
+		po.flags |= posetFlagUnsigned
+	} else {
+		po.flags &^= posetFlagUnsigned
+	}
+}
+
 // Handle children
 func (po *poset) setchl(i uint32, l posetEdge) { po.nodes[i].l = l }
 func (po *poset) setchr(i uint32, r posetEdge) { po.nodes[i].r = r }

src/cmd/compile/internal/ssa/poset_test.go

@@ -64,7 +64,8 @@ func testPosetOps(t *testing.T, unsigned bool, ops []posetTestOp) {
 		}
 	}
 
-	po := newPoset(unsigned)
+	po := newPoset()
+	po.SetUnsigned(unsigned)
 	for idx, op := range ops {
 		t.Logf("op%d%v", idx, op)
 		switch op.typ {

src/cmd/compile/internal/ssa/prove.go

@@ -181,10 +181,12 @@ type factsTable struct {
 var checkpointFact = fact{}
 var checkpointBound = limitFact{}
 
-func newFactsTable() *factsTable {
+func newFactsTable(f *Func) *factsTable {
 	ft := &factsTable{}
-	ft.order[0] = newPoset(false) // signed
-	ft.order[1] = newPoset(true)  // unsigned
+	ft.order[0] = f.newPoset() // signed
+	ft.order[1] = f.newPoset() // unsigned
+	ft.order[0].SetUnsigned(false)
+	ft.order[1].SetUnsigned(true)
 	ft.facts = make(map[pair]relation)
 	ft.stack = make([]fact, 4)
 	ft.limits = make(map[ID]limit)
@@ -666,7 +668,8 @@ var (
 // its negation. If either leads to a contradiction, it can trim that
 // successor.
 func prove(f *Func) {
-	ft := newFactsTable()
+	ft := newFactsTable(f)
+	ft.checkpoint()
 
 	// Find length and capacity ops.
 	var zero *Value
@@ -794,6 +797,20 @@ func prove(f *Func) {
 			ft.restore()
 		}
 	}
+
+	ft.restore()
+
+	// Return the posets to the free list
+	for _, po := range ft.order {
+		// Make sure it's empty as it should be. A non-empty poset
+		// might cause errors and miscompilations if reused.
+		if checkEnabled {
+			if err := po.CheckEmpty(); err != nil {
+				f.Fatalf("prove poset not empty after function %s: %v", f.Name, err)
+			}
+		}
+		f.retPoset(po)
+	}
 }
 
 // getBranch returns the range restrictions added by p