cmd/compile: trim more blocks

 - trim blocks with multiple predecessors
 - trim blocks, which contain only phi-functions
 - trim blocks, which can be merged into the successor block

As an example, compiling the following source:

---8<------
package p

type Node struct {
        Key         int
        Left, Right *Node
}

func Search(r *Node, k int) *Node {
        for r != nil {
                switch {
                case k == r.Key:
                        return r
                case k < r.Key:
                        r = r.Left
                default:
                        r = r.Right
                }
        }
        return nil
}
---8<------

with `GOSSAFUNC=Search" go tool compile t.go`, results in the following
code:

---8<------
genssa

      00000 (t.go:8)	TEXT	"".Search(SB), $0
      00001 (t.go:8)	FUNCDATA	$0, "".gcargs·0(SB)
      00002 (t.go:8)	FUNCDATA	$1, "".gclocals·1(SB)
      00003 (t.go:8)	TYPE	"".r(FP)type.*"".Node, $8
      00004 (t.go:8)	TYPE	"".k+8(FP)type.int, $8
      00005 (t.go:8)	TYPE	"".~r2+16(FP)type.*"".Node, $8
v40   00006 (t.go:9)	MOVQ	"".k+8(FP), AX
v34   00007 (t.go:9)	MOVQ	"".r(FP), CX
v33   00008 (t.go:9)	TESTQ	CX, CX
b2    00009 (t.go:9)	JEQ	$0, 22
v16   00010 (t.go:11)	MOVQ	(CX), DX
v21   00011 (t.go:11)	CMPQ	DX, AX
b9    00012 (t.go:11)	JEQ	$0, 19
v64   00013 (t.go:13)	CMPQ	AX, DX
b13   00014 (t.go:13)	JGE	17
v36   00015 (t.go:14)	MOVQ	8(CX), CX
b4    00016 (t.go:9)	JMP	8                  <---+
v42   00017 (t.go:16)	MOVQ	16(CX), CX             |
b21   00018 (t.go:10)	JMP	16                 ----+
v28   00019 (t.go:12)	VARDEF	"".~r2+16(FP)
v29   00020 (t.go:12)	MOVQ	CX, "".~r2+16(FP)
b10   00021 (t.go:12)	RET
v44   00022 (t.go:19)	VARDEF	"".~r2+16(FP)
v45   00023 (t.go:19)	MOVQ	$0, "".~r2+16(FP)
b5    00024 (t.go:19)	RET
00025 (<unknown line number>)	END
---8<------

Note the jump at 18 jumps to another jump at 16.

Looking at the function after trimming:

--8<------
after trim [199 ns]

b1:
v1 = InitMem <mem>
v2 = SP <uintptr> : SP
v67 = Arg <*Node> {r} : r[*Node]
v59 = Arg <int> {k} : k[int]
v40 = LoadReg <int> v59 : AX
v34 = LoadReg <*Node> v67 : CX
Plain → b2

b2: ← b1 b4
v8 = Phi <*Node> v34 v68 : CX
v33 = TESTQ <flags> v8 v8
NE v33 → b9 b5 (likely)

b9: ← b2
v16 = MOVQload <int> v8 v1 : DX
v21 = CMPQ <flags> v16 v40
EQ v21 → b10 b13 (unlikely)

b13: ← b9
v64 = CMPQ <flags> v40 v16
LT v64 → b19 b21

b19: ← b13
v36 = MOVQload <*Node> [8] v8 v1 : CX
Plain → b4

b4: ← b21 b19                       <
v68 = Phi <*Node> v42 v36 : CX      <- no actual code
Plain → b2                          <

b21: ← b13
v42 = MOVQload <*Node> [16] v8 v1 : CX
Plain → b4

b10: ← b9
v28 = VarDef <mem> {~r2} v1
v29 = MOVQstore <mem> {~r2} v2 v8 v28
v30 = Copy <mem> v29
Ret v30

b5: ← b2
v44 = VarDef <mem> {~r2} v1
v45 = MOVQstoreconst <mem> {~r2} [val=0,off=0] v2 v44
v47 = Copy <mem> v45
Ret v47

--8<------

The jump at 16 corresponds to the edge b21 -> b4. The block b4 contains
only phi-ops, i.e. no actual code besides the jump to b2. However b4 is
not trimmed, because it a) has more than one predecessor, and b) it is
not empty.

This change enhances trim.go to remove more blocks, subject to the
following criteria:

 - block has predecessors (i.e. not the start block)

 - block is BlockPlain

 - block does not loop back to itself

 - block is the single predecessor of its successor; the instructions of
   the block are merged into the successor

 - block does no emit actual code, besides a possible unconditional
   jump.
     Currently only OpPhi are considered to not be actual code,
   perhaps OpKeepAlive/others should be considered too?

As an example, after the change, the block b4 is trimmed and the jump at
18 jumps directly to 8.

Revision 1: Adjust phi-ops arguments after merge

Ensure the number of phi-ops arguments matches the new number of
predecessors in the merged block.
When moving values, make them refer to the merged block.

Revision 2:
 - Make clear the intent that we do not want to trim the entry block
 - Double check that we are merging a phi operation
 - Minor code style fix
 - Fix a potentially dangerous situation when a blocks refers to the
   inline value space in another block

Change-Id: I0ab91779f931f404d11008f5c45606d985d7fbaa
Reviewed-on: https://go-review.googlesource.com/28812
Run-TryBot: Russ Cox <rsc@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>

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

@@ -9,25 +9,129 @@ package ssa
 func trim(f *Func) {
 	n := 0
 	for _, b := range f.Blocks {
-		if b.Kind != BlockPlain || len(b.Values) != 0 || len(b.Preds) != 1 {
+		if !trimmableBlock(b) {
 			f.Blocks[n] = b
 			n++
 			continue
 		}
-		// TODO: handle len(b.Preds)>1 case.
 
-		// Splice b out of the graph.
-		p := b.Preds[0].b
-		i := b.Preds[0].i
-		s := b.Succs[0].b
-		j := b.Succs[0].i
+		// Splice b out of the graph. NOTE: `mergePhi` depends on the
+		// order, in which the predecessors edges are merged here.
+		p, i := b.Preds[0].b, b.Preds[0].i
+		s, j := b.Succs[0].b, b.Succs[0].i
+		ns := len(s.Preds)
 		p.Succs[i] = Edge{s, j}
 		s.Preds[j] = Edge{p, i}
+
+		for _, e := range b.Preds[1:] {
+			p, i := e.b, e.i
+			p.Succs[i] = Edge{s, len(s.Preds)}
+			s.Preds = append(s.Preds, Edge{p, i})
+		}
+
+		// If `s` had more than one predecessor, update its phi-ops to
+		// account for the merge.
+		if ns > 1 {
+			for _, v := range s.Values {
+				if v.Op == OpPhi {
+					mergePhi(v, j, b)
+				}
+			}
+			// Remove the phi-ops from `b` if they were merged into the
+			// phi-ops of `s`.
+			k := 0
+			for _, v := range b.Values {
+				if v.Op == OpPhi {
+					if v.Uses == 0 {
+						v.resetArgs()
+						continue
+					}
+					// Pad the arguments of the remaining phi-ops, so
+					// they match the new predecessor count of `s`.
+					for len(v.Args) < len(s.Preds) {
+						v.AddArg(v.Args[0])
+					}
+				}
+				b.Values[k] = v
+				k++
+			}
+			b.Values = b.Values[:k]
+		}
+
+		// Merge the blocks' values.
+		for _, v := range b.Values {
+			v.Block = s
+		}
+		k := len(b.Values)
+		m := len(s.Values)
+		for i := 0; i < k; i++ {
+			s.Values = append(s.Values, nil)
+		}
+		copy(s.Values[k:], s.Values[:m])
+		copy(s.Values, b.Values)
+	}
+	if n < len(f.Blocks) {
 		f.invalidateCFG()
+		tail := f.Blocks[n:]
+		for i := range tail {
+			tail[i] = nil
+		}
+		f.Blocks = f.Blocks[:n]
+	}
+}
+
+// emptyBlock returns true if the block does not contain actual
+// instructions
+func emptyBlock(b *Block) bool {
+	for _, v := range b.Values {
+		if v.Op != OpPhi {
+			return false
+		}
+	}
+	return true
+}
+
+// trimmableBlock returns true if the block can be trimmed from the CFG,
+// subject to the following criteria:
+//  - it should not be the first block
+//  - it should be BlockPlain
+//  - it should not loop back to itself
+//  - it either is the single predecessor of the successor block or
+//    contains no actual instructions
+func trimmableBlock(b *Block) bool {
+	if b.Kind != BlockPlain || b == b.Func.Entry {
+		return false
 	}
-	tail := f.Blocks[n:]
-	for i := range tail {
-		tail[i] = nil
+	s := b.Succs[0].b
+	return s != b && (len(s.Preds) == 1 || emptyBlock(b))
+}
+
+// mergePhi adjusts the number of `v`s arguments to account for merge
+// of `b`, which was `i`th predecessor of the `v`s block. Returns
+// `v`.
+func mergePhi(v *Value, i int, b *Block) *Value {
+	u := v.Args[i]
+	if u.Block == b {
+		if u.Op != OpPhi {
+			b.Func.Fatalf("value %s is not a phi operation", u.LongString())
+		}
+		// If the original block contained u = φ(u0, u1, ..., un) and
+		// the current phi is
+		//    v = φ(v0, v1, ..., u, ..., vk)
+		// then the merged phi is
+		//    v = φ(v0, v1, ..., u0, ..., vk, u1, ..., un)
+		v.SetArg(i, u.Args[0])
+		v.AddArgs(u.Args[1:]...)
+	} else {
+		// If the original block contained u = φ(u0, u1, ..., un) and
+		// the current phi is
+		//    v = φ(v0, v1, ...,  vi, ..., vk)
+		// i.e. it does not use a value from the predecessor block,
+		// then the merged phi is
+		//    v = φ(v0, v1, ..., vk, vi, vi, ...)
+		for j := 1; j < len(b.Preds); j++ {
+			v.AddArg(v.Args[i])
+		}
 	}
-	f.Blocks = f.Blocks[:n]
+	return v
 }