exp/ssa: build fully pruned SSA form.

Overview: Function.finish() now invokes the "lifting" pass which replaces local allocs and loads and stores to such cells by SSA registers.  We use very standard machinery:

(1) we build the dominator tree for the function's control flow graph (CFG) using the "Simple" Lengauer-Tarjan algorithm.  (Very "simple" in fact: even simple path compression is not yet implemented.)

In sanity-checking mode, we cross check the dominator tree against an alternative implementation using a simple iterative dataflow algorithm.
This all lives in dom.go, along with some diagnostic printing routines.

(2) we build the dominance frontier for the entire CFG using the Cytron et al algorithm.  The DF is represented as a slice of slices, keyed by block index.  See buildDomFrontier() in lift.go.

(3) we determine for each Alloc whether it can be lifted: is it only subject to loads and stores?  If so, we traverse the iterated dominance frontier (IDF) creating φ-nodes; they are not prepended to the blocks yet.
See liftAlloc() in lift.go.

(4) we perform the SSA renaming algorithm from Cytron et al, replacing all loads to lifted Alloc cells by the value stored by the dominating store operation, and deleting the stores and allocs.  See rename() in lift.go.

(5) we eliminate unneeded φ-nodes, then concatenate the remaining ones with the non-deleted instructions of the block into a new slice.  We eliminate any lifted allocs from Function.Locals.

To ease reviewing, I have avoided almost all optimisations at this point, though there are many opportunities to explore.  These will be easier to understand as follow-up changes.

All the existing tests (pending CL 7313062) pass.  (Faster!)

Details:

"NaiveForm" BuilderMode flag suppresses all the new logic.
Exposed as 'ssadump -build=N'.

BasicBlock:
- add .Index field (b.Func[b.Index]==b), simplifying
  algorithms such as Kildall-style dataflow with bitvectors.
- rename the Name field to Comment to better reflect its
  reduced purpose.  It now has a String() method.
- 'dom' field holds dominator tree node; private for now.
- new predIndex method.
- hasPhi is now a method

dom.go:
- domTree: a new struct for a node in a dominator tree.
- buildDomTree builds the dominator tree using the simple
  variant Lengauer/Tarjan algorithm with Georgiadis'
  bucket optimizations.
- sanityCheckDomTree builds dominance relation using
  Kildall-style dataflow and ensures the same result is
  obtained.
- printDomTreeDot prints the CFG/DomTree in GraphViz format.

blockopt.go:
- perform a mark/sweep pass to eliminate unreachable
  cycles; the previous prune() opt would only eliminate
  trivially dead blocks.  (Needed for LT algo.)
- using .Index, fuseblocks can now delete fused blocks directly.
- delete prune().

sanity.go: more consistency checks:
- Phi with missing edge value
- local Alloc instructions must appear in Function.Locals.
- BasicBlock.Index, Func consistency
- CFG edges are all intraprocedural.
- detect nils in BasicBlock.Instrs.
- detect Function.Locals with Heap flag set.
- check fn.Blocks is nil if empty.

Also:
- Phi now has Comment field for debugging.
- Fixed bug in Select.Operands()
  (took address of temporary copy of field)
- new Literal constructor zeroLiteral().
- algorithms steal private fields Alloc.index,
  BasicBlock.gaps to avoid allocating maps.
- We print Function.Locals in DumpTo.
- added profiling support to ssadump.

R=iant, gri
CC=golang-dev
https://golang.org/cl/7229074

src/pkg/exp/ssa/blockopt.go

 package ssa
 
-// Simple block optimisations to simplify the control flow graph.
+// Simple block optimizations to simplify the control flow graph.
 
 // TODO(adonovan): instead of creating several "unreachable" blocks
 // per function in the Builder, reuse a single one (e.g. at Blocks[1])
@@ -15,47 +15,52 @@ import (
 // successive iteration of optimizeBlocks.  Very verbose.
 const debugBlockOpt = false
 
-func hasPhi(b *BasicBlock) bool {
-	_, ok := b.Instrs[0].(*Phi)
-	return ok
+// markReachable sets Index=-1 for all blocks reachable from b.
+func markReachable(b *BasicBlock) {
+	b.Index = -1
+	for _, succ := range b.Succs {
+		if succ.Index == 0 {
+			markReachable(succ)
+		}
+	}
 }
 
-// prune attempts to prune block b if it is unreachable (i.e. has no
-// predecessors other than itself), disconnecting it from the CFG.
-// The result is true if the optimisation was applied.  i is the block
-// index within the function.
+// deleteUnreachableBlocks marks all reachable blocks of f and
+// eliminates (nils) all others, including possibly cyclic subgraphs.
 //
-func prune(f *Function, i int, b *BasicBlock) bool {
-	if i == 0 {
-		return false // don't prune entry block
-	}
-	if len(b.Preds) == 0 || len(b.Preds) == 1 && b.Preds[0] == b {
-		// Disconnect it from its successors.
-		for _, c := range b.Succs {
-			c.removePred(b)
-		}
-		if debugBlockOpt {
-			fmt.Fprintln(os.Stderr, "prune", b.Name)
+func deleteUnreachableBlocks(f *Function) {
+	const white, black = 0, -1
+	// We borrow b.Index temporarily as the mark bit.
+	for _, b := range f.Blocks {
+		b.Index = white
+	}
+	markReachable(f.Blocks[0])
+	for i, b := range f.Blocks {
+		if b.Index == white {
+			for _, c := range b.Succs {
+				if c.Index == black {
+					c.removePred(b) // delete white->black edge
+				}
+			}
+			if debugBlockOpt {
+				fmt.Fprintln(os.Stderr, "unreachable", b)
+			}
+			f.Blocks[i] = nil // delete b
 		}
-
-		// Delete b.
-		f.Blocks[i] = nil
-		return true
 	}
-	return false
+	f.removeNilBlocks()
 }
 
 // jumpThreading attempts to apply simple jump-threading to block b,
 // in which a->b->c become a->c if b is just a Jump.
-// The result is true if the optimisation was applied.
-// i is the block index within the function.
+// The result is true if the optimization was applied.
 //
-func jumpThreading(f *Function, i int, b *BasicBlock) bool {
-	if i == 0 {
+func jumpThreading(f *Function, b *BasicBlock) bool {
+	if b.Index == 0 {
 		return false // don't apply to entry block
 	}
 	if b.Instrs == nil {
-		fmt.Println("empty block ", b.Name)
+		fmt.Println("empty block ", b)
 		return false
 	}
 	if _, ok := b.Instrs[0].(*Jump); !ok {
@@ -65,7 +70,7 @@ func jumpThreading(f *Function, i int, b *BasicBlock) bool {
 	if c == b {
 		return false // don't apply to degenerate jump-to-self.
 	}
-	if hasPhi(c) {
+	if c.hasPhi() {
 		return false // not sound without more effort
 	}
 	for j, a := range b.Preds {
@@ -87,16 +92,16 @@ func jumpThreading(f *Function, i int, b *BasicBlock) bool {
 		}
 
 		if debugBlockOpt {
-			fmt.Fprintln(os.Stderr, "jumpThreading", a.Name, b.Name, c.Name)
+			fmt.Fprintln(os.Stderr, "jumpThreading", a, b, c)
 		}
 	}
-	f.Blocks[i] = nil
+	f.Blocks[b.Index] = nil // delete b
 	return true
 }
 
-// fuseBlocks attempts to apply the block fusion optimisation to block
+// fuseBlocks attempts to apply the block fusion optimization to block
 // a, in which a->b becomes ab if len(a.Succs)==len(b.Preds)==1.
-// The result is true if the optimisation was applied.
+// The result is true if the optimization was applied.
 //
 func fuseBlocks(f *Function, a *BasicBlock) bool {
 	if len(a.Succs) != 1 {
@@ -121,11 +126,10 @@ func fuseBlocks(f *Function, a *BasicBlock) bool {
 	}
 
 	if debugBlockOpt {
-		fmt.Fprintln(os.Stderr, "fuseBlocks", a.Name, b.Name)
+		fmt.Fprintln(os.Stderr, "fuseBlocks", a, b)
 	}
 
-	// Make b unreachable.  Subsequent pruning will reclaim it.
-	b.Preds = nil
+	f.Blocks[b.Index] = nil // delete b
 	return true
 }
 
@@ -134,6 +138,8 @@ func fuseBlocks(f *Function, a *BasicBlock) bool {
 // threading.
 //
 func optimizeBlocks(f *Function) {
+	deleteUnreachableBlocks(f)
+
 	// Loop until no further progress.
 	changed := true
 	for changed {
@@ -144,43 +150,24 @@ func optimizeBlocks(f *Function) {
 			MustSanityCheck(f, nil)
 		}
 
-		for i, b := range f.Blocks {
+		for _, b := range f.Blocks {
 			// f.Blocks will temporarily contain nils to indicate
 			// deleted blocks; we remove them at the end.
 			if b == nil {
 				continue
 			}
 
-			// Prune unreachable blocks (including all empty blocks).
-			if prune(f, i, b) {
-				changed = true
-				continue // (b was pruned)
-			}
-
 			// Fuse blocks.  b->c becomes bc.
 			if fuseBlocks(f, b) {
 				changed = true
 			}
 
 			// a->b->c becomes a->c if b contains only a Jump.
-			if jumpThreading(f, i, b) {
+			if jumpThreading(f, b) {
 				changed = true
 				continue // (b was disconnected)
 			}
 		}
 	}
-
-	// Eliminate nils from Blocks.
-	j := 0
-	for _, b := range f.Blocks {
-		if b != nil {
-			f.Blocks[j] = b
-			j++
-		}
-	}
-	// Nil out b.Blocks[j:] to aid GC.
-	for i := j; i < len(f.Blocks); i++ {
-		f.Blocks[i] = nil
-	}
-	f.Blocks = f.Blocks[:j]
+	f.removeNilBlocks()
 }

src/pkg/exp/ssa/builder.go

@@ -122,6 +122,7 @@ const (
 	LogSource                                    // Show source locations as SSA builder progresses
 	SanityCheckFunctions                         // Perform sanity checking of function bodies
 	UseGCImporter                                // Ignore SourceLoader; use gc-compiled object code for all imports
+	NaiveForm                                    // Build naïve SSA form: don't replace local loads/stores with registers
 )
 
 // NewBuilder creates and returns a new SSA builder.
@@ -380,7 +381,7 @@ func (b *Builder) logicalBinop(fn *Function, e *ast.BinaryExpr) Value {
 
 	// TODO(adonovan): do we need emitConv on each edge?
 	// Test with named boolean types.
-	phi := &Phi{Edges: edges}
+	phi := &Phi{Edges: edges, Comment: e.Op.String()}
 	phi.Type_ = phi.Edges[0].Type()
 	return done.emit(phi)
 }
@@ -1340,7 +1341,7 @@ func (b *Builder) globalValueSpec(init *Function, spec *ast.ValueSpec, g *Global
 			for i, id := range spec.Names {
 				if !isBlankIdent(id) {
 					g := b.globals[b.obj(id)].(*Global)
-					g.spec = nil // just an optimisation
+					g.spec = nil // just an optimization
 					emitStore(init, g,
 						emitExtract(init, tuple, i, rtypes[i].Type))
 				}
@@ -1419,7 +1420,7 @@ func (b *Builder) assignStmt(fn *Function, lhss, rhss []ast.Expr, isDef bool) {
 		// e.g. x, y = f(), g()
 		if len(lhss) == 1 {
 			// x = type{...}
-			// Optimisation: in-place construction
+			// Optimization: in-place construction
 			// of composite literals.
 			b.exprInPlace(fn, lvals[0], rhss[0])
 		} else {
@@ -1764,7 +1765,7 @@ func (b *Builder) typeSwitchStmt(fn *Function, s *ast.TypeSwitchStmt, label *lbl
 func (b *Builder) selectStmt(fn *Function, s *ast.SelectStmt, label *lblock) {
 	// A blocking select of a single case degenerates to a
 	// simple send or receive.
-	// TODO(adonovan): is this optimisation worth its weight?
+	// TODO(adonovan): is this optimization worth its weight?
 	if len(s.Body.List) == 1 {
 		clause := s.Body.List[0].(*ast.CommClause)
 		if clause.Comm != nil {

src/pkg/exp/ssa/doc.go

@@ -32,10 +32,11 @@
 //
 // The builder initially builds a naive SSA form in which all local
 // variables are addresses of stack locations with explicit loads and
-// stores.  If desired, registerisation and φ-node insertion using
-// dominance and dataflow can be performed as a later pass to improve
-// the accuracy and performance of subsequent analyses; this pass is
-// not yet implemented.
+// stores.  Registerisation of eligible locals and φ-node insertion
+// using dominance and dataflow are then performed as a second pass
+// called "lifting" to improve the accuracy and performance of
+// subsequent analyses; this pass can be skipped by setting the
+// NaiveForm builder flag.
 //
 // The program representation constructed by this package is fully
 // resolved internally, i.e. it does not rely on the names of Values,

src/pkg/exp/ssa/dom.go

+package ssa
+
+// This file defines algorithms related to dominance.
+
+// Dominator tree construction ----------------------------------------
+//
+// We use the algorithm described in Lengauer & Tarjan. 1979.  A fast
+// algorithm for finding dominators in a flowgraph.
+// http://doi.acm.org/10.1145/357062.357071
+//
+// We also apply the optimizations to SLT described in Georgiadis et
+// al, Finding Dominators in Practice, JGAA 2006,
+// http://jgaa.info/accepted/2006/GeorgiadisTarjanWerneck2006.10.1.pdf
+// to avoid the need for buckets of size > 1.
+
+import (
+	"fmt"
+	"io"
+	"math/big"
+	"os"
+)
+
+// domNode represents a node in the dominator tree.
+//
+// TODO(adonovan): export this, when ready.
+type domNode struct {
+	Block     *BasicBlock // the basic block; n.Block.dom == n
+	Idom      *domNode    // immediate dominator (parent in dominator tree)
+	Children  []*domNode  // nodes dominated by this one
+	Level     int         // level number of node within tree; zero for root
+	Pre, Post int         // pre- and post-order numbering within dominator tree
+
+	// Working state for Lengauer-Tarjan algorithm
+	// (during which Pre is repurposed for CFG DFS preorder number).
+	// TODO(adonovan): opt: measure allocating these as temps.
+	semi     *domNode // semidominator
+	parent   *domNode // parent in DFS traversal of CFG
+	ancestor *domNode // ancestor with least sdom
+}
+
+// ltDfs implements the depth-first search part of the LT algorithm.
+func ltDfs(v *domNode, i int, preorder []*domNode) int {
+	preorder[i] = v
+	v.Pre = i // For now: DFS preorder of spanning tree of CFG
+	i++
+	v.semi = v
+	v.ancestor = nil
+	for _, succ := range v.Block.Succs {
+		if w := succ.dom; w.semi == nil {
+			w.parent = v
+			i = ltDfs(w, i, preorder)
+		}
+	}
+	return i
+}
+
+// ltEval implements the EVAL part of the LT algorithm.
+func ltEval(v *domNode) *domNode {
+	// TODO(adonovan): opt: do path compression per simple LT.
+	u := v
+	for ; v.ancestor != nil; v = v.ancestor {
+		if v.semi.Pre < u.semi.Pre {
+			u = v
+		}
+	}
+	return u
+}
+
+// ltLink implements the LINK part of the LT algorithm.
+func ltLink(v, w *domNode) {
+	w.ancestor = v
+}
+
+// buildDomTree computes the dominator tree of f using the LT algorithm.
+// Precondition: all blocks are reachable (e.g. optimizeBlocks has been run).
+//
+func buildDomTree(f *Function) {
+	// The step numbers refer to the original LT paper; the
+	// reodering is due to Georgiadis.
+
+	// Initialize domNode nodes.
+	for _, b := range f.Blocks {
+		dom := b.dom
+		if dom == nil {
+			dom = &domNode{Block: b}
+			b.dom = dom
+		} else {
+			dom.Block = b // reuse
+		}
+	}
+
+	// Step 1.  Number vertices by depth-first preorder.
+	n := len(f.Blocks)
+	preorder := make([]*domNode, n)
+	root := f.Blocks[0].dom
+	ltDfs(root, 0, preorder)
+
+	buckets := make([]*domNode, n)
+	copy(buckets, preorder)
+
+	// In reverse preorder...
+	for i := n - 1; i > 0; i-- {
+		w := preorder[i]
+
+		// Step 3. Implicitly define the immediate dominator of each node.
+		for v := buckets[i]; v != w; v = buckets[v.Pre] {
+			u := ltEval(v)
+			if u.semi.Pre < i {
+				v.Idom = u
+			} else {
+				v.Idom = w
+			}
+		}
+
+		// Step 2. Compute the semidominators of all nodes.
+		w.semi = w.parent
+		for _, pred := range w.Block.Preds {
+			v := pred.dom
+			u := ltEval(v)
+			if u.semi.Pre < w.semi.Pre {
+				w.semi = u.semi
+			}
+		}
+
+		ltLink(w.parent, w)
+
+		if w.parent == w.semi {
+			w.Idom = w.parent
+		} else {
+			buckets[i] = buckets[w.semi.Pre]
+			buckets[w.semi.Pre] = w
+		}
+	}
+
+	// The final 'Step 3' is now outside the loop.
+	for v := buckets[0]; v != root; v = buckets[v.Pre] {
+		v.Idom = root
+	}
+
+	// Step 4. Explicitly define the immediate dominator of each
+	// node, in preorder.
+	for _, w := range preorder[1:] {
+		if w == root {
+			w.Idom = nil
+		} else {
+			if w.Idom != w.semi {
+				w.Idom = w.Idom.Idom
+			}
+			// Calculate Children relation as inverse of Idom.
+			w.Idom.Children = append(w.Idom.Children, w)
+		}
+
+		// Clear working state.
+		w.semi = nil
+		w.parent = nil
+		w.ancestor = nil
+	}
+
+	numberDomTree(root, 0, 0, 0)
+
+	// printDomTreeDot(os.Stderr, f)        // debugging
+	// printDomTreeText(os.Stderr, root, 0) // debugging
+
+	if f.Prog.mode&SanityCheckFunctions != 0 {
+		sanityCheckDomTree(f)
+	}
+}
+
+// numberDomTree sets the pre- and post-order numbers of a depth-first
+// traversal of the dominator tree rooted at v.  These are used to
+// answer dominance queries in constant time.  Also, it sets the level
+// numbers (zero for the root) used for frontier computation.
+//
+func numberDomTree(v *domNode, pre, post, level int) (int, int) {
+	v.Level = level
+	level++
+	v.Pre = pre
+	pre++
+	for _, child := range v.Children {
+		pre, post = numberDomTree(child, pre, post, level)
+	}
+	v.Post = post
+	post++
+	return pre, post
+}
+
+// dominates returns true if b dominates c.
+// Requires that dominance information is up-to-date.
+//
+func dominates(b, c *BasicBlock) bool {
+	return b.dom.Pre <= c.dom.Pre && c.dom.Post <= b.dom.Post
+}
+
+// Testing utilities ----------------------------------------
+
+// sanityCheckDomTree checks the correctness of the dominator tree
+// computed by the LT algorithm by comparing against the dominance
+// relation computed by a naive Kildall-style forward dataflow
+// analysis (Algorithm 10.16 from the "Dragon" book).
+//
+func sanityCheckDomTree(f *Function) {
+	n := len(f.Blocks)
+
+	// D[i] is the set of blocks that dominate f.Blocks[i],
+	// represented as a bit-set of block indices.
+	D := make([]big.Int, n)
+
+	one := big.NewInt(1)
+
+	// all is the set of all blocks; constant.
+	var all big.Int
+	all.Set(one).Lsh(&all, uint(n)).Sub(&all, one)
+
+	// Initialization.
+	for i := range f.Blocks {
+		if i == 0 {
+			// The root is dominated only by itself.
+			D[i].SetBit(&D[0], 0, 1)
+		} else {
+			// All other blocks are (initially) dominated
+			// by every block.
+			D[i].Set(&all)
+		}
+	}
+
+	// Iteration until fixed point.
+	for changed := true; changed; {
+		changed = false
+		for i, b := range f.Blocks {
+			if i == 0 {
+				continue
+			}
+			// Compute intersection across predecessors.
+			var x big.Int
+			x.Set(&all)
+			for _, pred := range b.Preds {
+				x.And(&x, &D[pred.Index])
+			}
+			x.SetBit(&x, i, 1) // a block always dominates itself.
+			if D[i].Cmp(&x) != 0 {
+				D[i].Set(&x)
+				changed = true
+			}
+		}
+	}
+
+	// Check the entire relation.  O(n^2).
+	ok := true
+	for i := 0; i < n; i++ {
+		for j := 0; j < n; j++ {
+			b, c := f.Blocks[i], f.Blocks[j]
+			actual := dominates(b, c)
+			expected := D[j].Bit(i) == 1
+			if actual != expected {
+				fmt.Fprintf(os.Stderr, "dominates(%s, %s)==%t, want %t\n", b, c, actual, expected)
+				ok = false
+			}
+		}
+	}
+	if !ok {
+		panic("sanityCheckDomTree failed for " + f.FullName())
+	}
+}
+
+// Printing functions ----------------------------------------
+
+// printDomTree prints the dominator tree as text, using indentation.
+func printDomTreeText(w io.Writer, v *domNode, indent int) {
+	fmt.Fprintf(w, "%*s%s\n", 4*indent, "", v.Block)
+	for _, child := range v.Children {
+		printDomTreeText(w, child, indent+1)
+	}
+}
+
+// printDomTreeDot prints the dominator tree of f in AT&T GraphViz
+// (.dot) format.
+func printDomTreeDot(w io.Writer, f *Function) {
+	fmt.Fprintln(w, "//", f.FullName())
+	fmt.Fprintln(w, "digraph domtree {")
+	for i, b := range f.Blocks {
+		v := b.dom
+		fmt.Fprintf(w, "\tn%d [label=\"%s (%d, %d)\",shape=\"rectangle\"];\n", v.Pre, b, v.Pre, v.Post)
+		// TODO(adonovan): improve appearance of edges
+		// belonging to both dominator tree and CFG.
+
+		// Dominator tree edge.
+		if i != 0 {
+			fmt.Fprintf(w, "\tn%d -> n%d [style=\"solid\",weight=100];\n", v.Idom.Pre, v.Pre)
+		}
+		// CFG edges.
+		for _, pred := range b.Preds {
+			fmt.Fprintf(w, "\tn%d -> n%d [style=\"dotted\",weight=0];\n", pred.dom.Pre, v.Pre)
+		}
+	}
+	fmt.Fprintln(w, "}")
+}

src/pkg/exp/ssa/func.go

@@ -16,6 +16,13 @@ func addEdge(from, to *BasicBlock) {
 	to.Preds = append(to.Preds, from)
 }
 
+// String returns a human-readable label of this block.
+// It is not guaranteed unique within the function.
+//
+func (b *BasicBlock) String() string {
+	return fmt.Sprintf("%d.%s", b.Index, b.Comment)
+}
+
 // emit appends an instruction to the current basic block.
 // If the instruction defines a Value, it is returned.
 //
@@ -26,6 +33,23 @@ func (b *BasicBlock) emit(i Instruction) Value {
 	return v
 }
 
+// predIndex returns the i such that b.Preds[i] == c or panics if
+// there is none.
+func (b *BasicBlock) predIndex(c *BasicBlock) int {
+	for i, pred := range b.Preds {
+		if pred == c {
+			return i
+		}
+	}
+	panic(fmt.Sprintf("no edge %s -> %s", c, b))
+}
+
+// hasPhi returns true if b.Instrs contains φ-nodes.
+func (b *BasicBlock) hasPhi() bool {
+	_, ok := b.Instrs[0].(*Phi)
+	return ok
+}
+
 // phis returns the prefix of b.Instrs containing all the block's φ-nodes.
 func (b *BasicBlock) phis() []Instruction {
 	for i, instr := range b.Instrs {
@@ -127,7 +151,7 @@ type funcSyntax struct {
 func (f *Function) labelledBlock(label *ast.Ident) *lblock {
 	lb := f.lblocks[label.Obj]
 	if lb == nil {
-		lb = &lblock{_goto: f.newBasicBlock("label." + label.Name)}
+		lb = &lblock{_goto: f.newBasicBlock(label.Name)}
 		f.lblocks[label.Obj] = lb
 	}
 	return lb
@@ -147,7 +171,7 @@ func (f *Function) addParam(name string, typ types.Type) *Parameter {
 
 // addSpilledParam declares a parameter that is pre-spilled to the
 // stack; the function body will load/store the spilled location.
-// Subsequent registerization will eliminate spills where possible.
+// Subsequent lifting will eliminate spills where possible.
 //
 func (f *Function) addSpilledParam(obj types.Object) {
 	name := obj.GetName()
@@ -213,6 +237,34 @@ func (f *Function) start(idents map[*ast.Ident]types.Object) {
 	}
 }
 
+// numberRegisters assigns numbers to all SSA registers
+// (value-defining Instructions) in f, to aid debugging.
+// (Non-Instruction Values are named at construction.)
+// NB: named Allocs retain their existing name.
+// TODO(adonovan): when we have source position info,
+// preserve names only for source locals.
+//
+func numberRegisters(f *Function) {
+	a, v := 0, 0
+	for _, b := range f.Blocks {
+		for _, instr := range b.Instrs {
+			switch instr := instr.(type) {
+			case *Alloc:
+				// Allocs may be named at birth.
+				if instr.Name_ == "" {
+					instr.Name_ = fmt.Sprintf("a%d", a)
+					a++
+				}
+			case Value:
+				instr.(interface {
+					setNum(int)
+				}).setNum(v)
+				v++
+			}
+		}
+	}
+}
+
 // finish() finalizes the function after SSA code generation of its body.
 func (f *Function) finish() {
 	f.objects = nil
@@ -235,27 +287,6 @@ func (f *Function) finish() {
 	}
 	f.Locals = f.Locals[:j]
 
-	// Ensure all value-defining Instructions have register names.
-	// (Non-Instruction Values are named at construction.)
-	tmp := 0
-	for _, b := range f.Blocks {
-		for _, instr := range b.Instrs {
-			switch instr := instr.(type) {
-			case *Alloc:
-				// Local Allocs may already be named.
-				if instr.Name_ == "" {
-					instr.Name_ = fmt.Sprintf("t%d", tmp)
-					tmp++
-				}
-			case Value:
-				instr.(interface {
-					setNum(int)
-				}).setNum(tmp)
-				tmp++
-			}
-		}
-	}
-
 	optimizeBlocks(f)
 
 	// Build immediate-use (referrers) graph.
@@ -273,9 +304,20 @@ func (f *Function) finish() {
 		}
 	}
 
+	if f.Prog.mode&NaiveForm == 0 {
+		// For debugging pre-state of lifting pass:
+		// numberRegisters(f)
+		// f.DumpTo(os.Stderr)
+
+		lift(f)
+	}
+
+	numberRegisters(f)
+
 	if f.Prog.mode&LogFunctions != 0 {
 		f.DumpTo(os.Stderr)
 	}
+
 	if f.Prog.mode&SanityCheckFunctions != 0 {
 		MustSanityCheck(f, nil)
 	}
@@ -284,6 +326,25 @@ func (f *Function) finish() {
 	}
 }
 
+// removeNilBlocks eliminates nils from f.Blocks and updates each
+// BasicBlock.Index.  Use this after any pass that may delete blocks.
+//
+func (f *Function) removeNilBlocks() {
+	j := 0
+	for _, b := range f.Blocks {
+		if b != nil {
+			b.Index = j
+			f.Blocks[j] = b
+			j++
+		}
+	}
+	// Nil out f.Blocks[j:] to aid GC.
+	for i := j; i < len(f.Blocks); i++ {
+		f.Blocks[i] = nil
+	}
+	f.Blocks = f.Blocks[:j]
+}
+
 // addNamedLocal creates a local variable, adds it to function f and
 // returns it.  Its name and type are taken from obj.  Subsequent
 // calls to f.lookup(obj) will return the same local.
@@ -417,6 +478,13 @@ func (f *Function) DumpTo(w io.Writer) {
 		}
 	}
 
+	if len(f.Locals) > 0 {
+		io.WriteString(w, "# Locals:\n")
+		for i, l := range f.Locals {
+			fmt.Fprintf(w, "# % 3d:\t%s %s\n", i, l.Name(), indirectType(l.Type()))
+		}
+	}
+
 	// Function Signature in declaration syntax; derived from types.Signature.String().
 	io.WriteString(w, "func ")
 	params := f.Params
@@ -450,19 +518,24 @@ func (f *Function) DumpTo(w io.Writer) {
 	}
 	io.WriteString(w, ":\n")
 
+	if f.Blocks == nil {
+		io.WriteString(w, "\t(external)\n")
+	}
+
 	for _, b := range f.Blocks {
 		if b == nil {
 			// Corrupt CFG.
 			fmt.Fprintf(w, ".nil:\n")
 			continue
 		}
-		fmt.Fprintf(w, ".%s:\t\t\t\t\t\t\t       P:%d S:%d\n", b.Name, len(b.Preds), len(b.Succs))
+		fmt.Fprintf(w, ".%s:\t\t\t\t\t\t\t       P:%d S:%d\n", b, len(b.Preds), len(b.Succs))
 		if false { // CFG debugging
-			fmt.Fprintf(w, "\t# CFG: %s --> %s --> %s\n", blockNames(b.Preds), b.Name, blockNames(b.Succs))
+			fmt.Fprintf(w, "\t# CFG: %s --> %s --> %s\n", blockNames(b.Preds), b, blockNames(b.Succs))
 		}
 		for _, instr := range b.Instrs {
 			io.WriteString(w, "\t")
-			if v, ok := instr.(Value); ok {
+			switch v := instr.(type) {
+			case Value:
 				l := 80 // for old time's sake.
 				// Left-align the instruction.
 				if name := v.Name(); name != "" {
@@ -475,7 +548,10 @@ func (f *Function) DumpTo(w io.Writer) {
 				if t := v.Type(); t != nil {
 					fmt.Fprintf(w, "%*s", l-9, t)
 				}
-			} else {
+			case nil:
+				// Be robust against bad transforms.
+				io.WriteString(w, "<deleted>")
+			default:
 				io.WriteString(w, instr.String())
 			}
 			io.WriteString(w, "\n")
@@ -484,14 +560,15 @@ func (f *Function) DumpTo(w io.Writer) {
 	fmt.Fprintf(w, "\n")
 }
 
-// newBasicBlock adds to f a new basic block with a unique name and
-// returns it.  It does not automatically become the current block for
-// subsequent calls to emit.
+// newBasicBlock adds to f a new basic block and returns it.  It does
+// not automatically become the current block for subsequent calls to emit.
+// comment is an optional string for more readable debugging output.
 //
-func (f *Function) newBasicBlock(name string) *BasicBlock {
+func (f *Function) newBasicBlock(comment string) *BasicBlock {
 	b := &BasicBlock{
-		Name: fmt.Sprintf("%d.%s", len(f.Blocks), name),
-		Func: f,
+		Index:   len(f.Blocks),
+		Comment: comment,
+		Func:    f,
 	}
 	b.Succs = b.succs2[:0]
 	f.Blocks = append(f.Blocks, b)

src/pkg/exp/ssa/literal.go

@@ -9,6 +9,8 @@ import (
 	"strconv"
 )
 
+var complexZero = types.Complex{new(big.Rat), new(big.Rat)}
+
 // newLiteral returns a new literal of the specified value and type.
 // val must be valid according to the specification of Literal.Value.
 //
@@ -28,6 +30,39 @@ func nilLiteral(typ types.Type) *Literal {
 	return newLiteral(types.NilType{}, typ)
 }
 
+// zeroLiteral returns a new "zero" literal of the specified type,
+// which must not be an array or struct type: the zero values of
+// aggregates are well-defined but cannot be represented by Literal.
+//
+func zeroLiteral(t types.Type) *Literal {
+	switch t := t.(type) {
+	case *types.Basic:
+		switch {
+		case t.Info&types.IsBoolean != 0:
+			return newLiteral(false, t)
+		case t.Info&types.IsComplex != 0:
+			return newLiteral(complexZero, t)
+		case t.Info&types.IsNumeric != 0:
+			return newLiteral(int64(0), t)
+		case t.Info&types.IsString != 0:
+			return newLiteral("", t)
+		case t.Kind == types.UnsafePointer:
+			fallthrough
+		case t.Kind == types.UntypedNil:
+			return nilLiteral(t)
+		default:
+			panic(fmt.Sprint("zeroLiteral for unexpected type:", t))
+		}
+	case *types.Pointer, *types.Slice, *types.Interface, *types.Chan, *types.Map, *types.Signature:
+		return nilLiteral(t)
+	case *types.NamedType:
+		return newLiteral(zeroLiteral(t.Underlying).Value, t)
+	case *types.Array, *types.Struct:
+		panic(fmt.Sprint("zeroLiteral applied to aggregate:", t))
+	}
+	panic(fmt.Sprint("zeroLiteral: unexpected ", t))
+}
+
 func (l *Literal) Name() string {
 	var s string
 	switch x := l.Value.(type) {

src/pkg/exp/ssa/print.go

@@ -91,13 +91,21 @@ func (v *Phi) String() string {
 		// Be robust against malformed CFG.
 		blockname := "?"
 		if v.Block_ != nil && i < len(v.Block_.Preds) {
-			blockname = v.Block_.Preds[i].Name
+			blockname = v.Block_.Preds[i].String()
 		}
 		b.WriteString(blockname)
 		b.WriteString(": ")
-		b.WriteString(relName(edge, v))
+		edgeVal := "<nil>" // be robust
+		if edge != nil {
+			edgeVal = relName(edge, v)
+		}
+		b.WriteString(edgeVal)
 	}
 	b.WriteString("]")
+	if v.Comment != "" {
+		b.WriteString(" #")
+		b.WriteString(v.Comment)
+	}
 	return b.String()
 }
 
@@ -255,7 +263,7 @@ func (s *Jump) String() string {
 	// Be robust against malformed CFG.
 	blockname := "?"
 	if s.Block_ != nil && len(s.Block_.Succs) == 1 {
-		blockname = s.Block_.Succs[0].Name
+		blockname = s.Block_.Succs[0].String()
 	}
 	return fmt.Sprintf("jump %s", blockname)
 }
@@ -264,8 +272,8 @@ func (s *If) String() string {
 	// Be robust against malformed CFG.
 	tblockname, fblockname := "?", "?"
 	if s.Block_ != nil && len(s.Block_.Succs) == 2 {
-		tblockname = s.Block_.Succs[0].Name
-		fblockname = s.Block_.Succs[1].Name
+		tblockname = s.Block_.Succs[0].String()
+		fblockname = s.Block_.Succs[1].String()
 	}
 	return fmt.Sprintf("if %s goto %s else %s", relName(s.Cond, s), tblockname, fblockname)
 }

src/pkg/exp/ssa/sanity.go

 package ssa
 
-// An optional pass for sanity checking invariants of the SSA representation.
+// An optional pass for sanity-checking invariants of the SSA representation.
 // Currently it checks CFG invariants but little at the instruction level.
 
 import (
@@ -50,7 +50,7 @@ func blockNames(blocks []*BasicBlock) string {
 		if i > 0 {
 			io.WriteString(&buf, ", ")
 		}
-		io.WriteString(&buf, b.Name)
+		io.WriteString(&buf, b.String())
 	}
 	return buf.String()
 }
@@ -58,7 +58,7 @@ func blockNames(blocks []*BasicBlock) string {
 func (s *sanity) diagnostic(prefix, format string, args ...interface{}) {
 	fmt.Fprintf(s.reporter, "%s: function %s", prefix, s.fn.FullName())
 	if s.block != nil {
-		fmt.Fprintf(s.reporter, ", block %s", s.block.Name)
+		fmt.Fprintf(s.reporter, ", block %s", s.block)
 	}
 	io.WriteString(s.reporter, ": ")
 	fmt.Fprintf(s.reporter, format, args...)
@@ -102,7 +102,7 @@ func (s *sanity) checkInstr(idx int, instr Instruction) {
 		if idx == 0 {
 			// It suffices to apply this check to just the first phi node.
 			if dup := findDuplicate(s.block.Preds); dup != nil {
-				s.errorf("phi node in block with duplicate predecessor %s", dup.Name)
+				s.errorf("phi node in block with duplicate predecessor %s", dup)
 			}
 		} else {
 			prev := s.block.Instrs[idx-1]
@@ -112,9 +112,29 @@ func (s *sanity) checkInstr(idx int, instr Instruction) {
 		}
 		if ne, np := len(instr.Edges), len(s.block.Preds); ne != np {
 			s.errorf("phi node has %d edges but %d predecessors", ne, np)
+
+		} else {
+			for i, e := range instr.Edges {
+				if e == nil {
+					s.errorf("phi node '%s' has no value for edge #%d from %s", instr.Comment, i, s.block.Preds[i])
+				}
+			}
 		}
 
 	case *Alloc:
+		if !instr.Heap {
+			found := false
+			for _, l := range s.fn.Locals {
+				if l == instr {
+					found = true
+					break
+				}
+			}
+			if !found {
+				s.errorf("local alloc %s = %s does not appear in Function.Locals", instr.Name(), instr)
+			}
+		}
+
 	case *Call:
 	case *BinOp:
 	case *UnOp:
@@ -155,7 +175,7 @@ func (s *sanity) checkFinalInstr(idx int, instr Instruction) {
 			return
 		}
 		if s.block.Succs[0] == s.block.Succs[1] {
-			s.errorf("If-instruction has same True, False target blocks: %s", s.block.Succs[0].Name)
+			s.errorf("If-instruction has same True, False target blocks: %s", s.block.Succs[0])
 			return
 		}
 
@@ -177,22 +197,30 @@ func (s *sanity) checkFinalInstr(idx int, instr Instruction) {
 	}
 }
 
-func (s *sanity) checkBlock(b *BasicBlock, isEntry bool) {
+func (s *sanity) checkBlock(b *BasicBlock, index int) {
 	s.block = b
 
+	if b.Index != index {
+		s.errorf("block has incorrect Index %d", b.Index)
+	}
+	if b.Func != s.fn {
+		s.errorf("block has incorrect Func %s", b.Func.FullName())
+	}
+
 	// Check all blocks are reachable.
 	// (The entry block is always implicitly reachable.)
-	if !isEntry && len(b.Preds) == 0 {
+	if index > 0 && len(b.Preds) == 0 {
 		s.warnf("unreachable block")
 		if b.Instrs == nil {
 			// Since this block is about to be pruned,
 			// tolerating transient problems in it
-			// simplifies other optimisations.
+			// simplifies other optimizations.
 			return
 		}
 	}
 
-	// Check predecessor and successor relations are dual.
+	// Check predecessor and successor relations are dual,
+	// and that all blocks in CFG belong to same function.
 	for _, a := range b.Preds {
 		found := false
 		for _, bb := range a.Succs {
@@ -202,7 +230,10 @@ func (s *sanity) checkBlock(b *BasicBlock, isEntry bool) {
 			}
 		}
 		if !found {
-			s.errorf("expected successor edge in predecessor %s; found only: %s", a.Name, blockNames(a.Succs))
+			s.errorf("expected successor edge in predecessor %s; found only: %s", a, blockNames(a.Succs))
+		}
+		if a.Func != s.fn {
+			s.errorf("predecessor %s belongs to different function %s", a, a.Func.FullName())
 		}
 	}
 	for _, c := range b.Succs {
@@ -214,21 +245,32 @@ func (s *sanity) checkBlock(b *BasicBlock, isEntry bool) {
 			}
 		}
 		if !found {
-			s.errorf("expected predecessor edge in successor %s; found only: %s", c.Name, blockNames(c.Preds))
+			s.errorf("expected predecessor edge in successor %s; found only: %s", c, blockNames(c.Preds))
+		}
+		if c.Func != s.fn {
+			s.errorf("successor %s belongs to different function %s", c, c.Func.FullName())
 		}
 	}
 
 	// Check each instruction is sane.
+	// TODO(adonovan): check Instruction invariants:
+	// - check Operands is dual to Value.Referrers.
+	// - check all Operands that are also Instructions belong to s.fn too
+	//   (and for bonus marks, that their block dominates block b).
 	n := len(b.Instrs)
 	if n == 0 {
 		s.errorf("basic block contains no instructions")
 	}
 	for j, instr := range b.Instrs {
+		if instr == nil {
+			s.errorf("nil instruction at index %d", j)
+			continue
+		}
 		if b2 := instr.Block(); b2 == nil {
 			s.errorf("nil Block() for instruction at index %d", j)
 			continue
 		} else if b2 != b {
-			s.errorf("wrong Block() (%s) for instruction at index %d ", b2.Name, j)
+			s.errorf("wrong Block() (%s) for instruction at index %d ", b2, j)
 			continue
 		}
 		if j < n-1 {
@@ -241,21 +283,30 @@ func (s *sanity) checkBlock(b *BasicBlock, isEntry bool) {
 
 func (s *sanity) checkFunction(fn *Function) bool {
 	// TODO(adonovan): check Function invariants:
-	// - check owning Package (if any) contains this function.
+	// - check owning Package (if any) contains this (possibly anon) function
 	// - check params match signature
-	// - check locals are all !Heap
 	// - check transient fields are nil
-	// - check block labels are unique (warning)
+	// - warn if any fn.Locals do not appear among block instructions.
 	s.fn = fn
 	if fn.Prog == nil {
 		s.errorf("nil Prog")
 	}
+	for i, l := range fn.Locals {
+		if l.Heap {
+			s.errorf("Local %s at index %d has Heap flag set", l.Name(), i)
+		}
+	}
+	if fn.Blocks != nil && len(fn.Blocks) == 0 {
+		// Function _had_ blocks (so it's not external) but
+		// they were "optimized" away, even the entry block.
+		s.errorf("Blocks slice is non-nil but empty")
+	}
 	for i, b := range fn.Blocks {
 		if b == nil {
 			s.warnf("nil *BasicBlock at f.Blocks[%d]", i)
 			continue
 		}
-		s.checkBlock(b, i == 0)
+		s.checkBlock(b, i)
 	}
 	s.block = nil
 	s.fn = nil

src/pkg/exp/ssa/ssa.go

@@ -261,11 +261,14 @@ type Function struct {
 // instructions, respectively).
 //
 type BasicBlock struct {
-	Name         string         // label; no semantic significance
+	Index        int            // index of this block within Func.Blocks
+	Comment      string         // optional label; no semantic significance
 	Func         *Function      // containing function
 	Instrs       []Instruction  // instructions in order
 	Preds, Succs []*BasicBlock  // predecessors and successors
 	succs2       [2]*BasicBlock // initial space for Succs.
+	dom          *domNode       // node in dominator tree; optional.
+	gaps         int            // number of nil Instrs (transient).
 }
 
 // Pure values ----------------------------------------
@@ -372,6 +375,7 @@ type Alloc struct {
 	Type_     types.Type
 	Heap      bool
 	referrers []Instruction
+	index     int // dense numbering; for lifting
 }
 
 // Phi represents an SSA φ-node, which combines values that differ
@@ -383,7 +387,8 @@ type Alloc struct {
 //
 type Phi struct {
 	Register
-	Edges []Value // Edges[i] is value for Block().Preds[i]
+	Comment string  // a hint as to its purpose
+	Edges   []Value // Edges[i] is value for Block().Preds[i]
 }
 
 // Call represents a function or method call.
@@ -422,6 +427,8 @@ type BinOp struct {
 // UnOp yields the result of Op X.
 // ARROW is channel receive.
 // MUL is pointer indirection (load).
+// XOR is bitwise complement.
+// SUB is negation.
 //
 // If CommaOk and Op=ARROW, the result is a 2-tuple of the value above
 // and a boolean indicating the success of the receive.  The
@@ -1239,8 +1246,8 @@ func (s *Ret) Operands(rands []*Value) []*Value {
 }
 
 func (v *Select) Operands(rands []*Value) []*Value {
-	for _, st := range v.States {
-		rands = append(rands, &st.Chan, &st.Send)
+	for i := range v.States {
+		rands = append(rands, &v.States[i].Chan, &v.States[i].Send)
 	}
 	return rands
 }

src/pkg/exp/ssa/ssadump.go

@@ -11,6 +11,7 @@ import (
 	"fmt"
 	"log"
 	"os"
+	"runtime/pprof"
 	"strings"
 )
 
@@ -22,6 +23,7 @@ P	log [P]ackage inventory.
 F	log [F]unction SSA code.
 S	log [S]ource locations as SSA builder progresses.
 G	use binary object files from gc to provide imports (no code).
+N	build [N]aive SSA form: don't replace local loads/stores with registers.
 `)
 
 var runFlag = flag.Bool("run", false, "Invokes the SSA interpreter on the program.")
@@ -41,6 +43,8 @@ Examples:
 % ssadump -build=FPG hello.go         # quickly dump SSA form of a single package
 `
 
+var cpuprofile = flag.String("cpuprofile", "", "write cpu profile to file")
+
 func main() {
 	flag.Parse()
 	args := flag.Args()
@@ -58,6 +62,8 @@ func main() {
 			mode |= ssa.LogSource
 		case 'C':
 			mode |= ssa.SanityCheckFunctions
+		case 'N':
+			mode |= ssa.NaiveForm
 		case 'G':
 			mode |= ssa.UseGCImporter
 		default:
@@ -95,6 +101,16 @@ func main() {
 		log.Fatal("No *.go source files specified.")
 	}
 
+	// Profiling support.
+	if *cpuprofile != "" {
+		f, err := os.Create(*cpuprofile)
+		if err != nil {
+			log.Fatal(err)
+		}
+		pprof.StartCPUProfile(f)
+		defer pprof.StopCPUProfile()
+	}
+
 	// TODO(adonovan): permit naming a package directly instead of
 	// a list of .go files.