diff --git a/src/cmd/compile/internal/gc/phi.go b/src/cmd/compile/internal/gc/phi.go
new file mode 100644
index 0000000000..ea9e5b10aa
--- /dev/null
+++ b/src/cmd/compile/internal/gc/phi.go
@@ -0,0 +1,521 @@
+// Copyright 2016 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package gc
+
+import (
+	"cmd/compile/internal/ssa"
+	"container/heap"
+	"fmt"
+)
+
+// This file contains the algorithm to place phi nodes in a function.
+// For small functions, we use Braun, Buchwald, Hack, Leißa, Mallon, and Zwinkau.
+// http://pp.info.uni-karlsruhe.de/uploads/publikationen/braun13cc.pdf
+// For large functions, we use Sreedhar & Gao: A Linear Time Algorithm for Placing Φ-Nodes.
+// http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.8.1979&rep=rep1&type=pdf
+
+const smallBlocks = 500
+
+const debugPhi = false
+
+// insertPhis finds all the places in the function where a phi is
+// necessary and inserts them.
+// Uses FwdRef ops to find all uses of variables, and s.defvars to find
+// all definitions.
+// Phi values are inserted, and all FwdRefs are changed to a Copy
+// of the appropriate phi or definition.
+// TODO: make this part of cmd/compile/internal/ssa somehow?
+func (s *state) insertPhis() {
+	if len(s.f.Blocks) <= smallBlocks && false {
+		sps := simplePhiState{s: s, f: s.f, defvars: s.defvars}
+		sps.insertPhis()
+		return
+	}
+	ps := phiState{s: s, f: s.f, defvars: s.defvars}
+	ps.insertPhis()
+}
+
+type phiState struct {
+	s       *state                 // SSA state
+	f       *ssa.Func              // function to work on
+	defvars []map[*Node]*ssa.Value // defined variables at end of each block
+
+	varnum map[*Node]int32 // variable numbering
+
+	// properties of the dominator tree
+	idom  []*ssa.Block // dominator parents
+	tree  []domBlock   // dominator child+sibling
+	level []int32      // level in dominator tree (0 = root or unreachable, 1 = children of root, ...)
+
+	// scratch locations
+	priq   blockHeap    // priority queue of blocks, higher level (toward leaves) = higher priority
+	q      []*ssa.Block // inner loop queue
+	queued *sparseSet   // has been put in q
+	hasPhi *sparseSet   // has a phi
+	hasDef *sparseSet   // has a write of the variable we're processing
+
+	// miscellaneous
+	placeholder *ssa.Value // dummy value to use as a "not set yet" placeholder.
+}
+
+func (s *phiState) insertPhis() {
+	if debugPhi {
+		fmt.Println(s.f.String())
+	}
+
+	// Find all the variables for which we need to match up reads & writes.
+	// This step prunes any basic-block-only variables from consideration.
+	// Generate a numbering for these variables.
+	s.varnum = map[*Node]int32{}
+	var vars []*Node
+	var vartypes []ssa.Type
+	for _, b := range s.f.Blocks {
+		for _, v := range b.Values {
+			if v.Op != ssa.OpFwdRef {
+				continue
+			}
+			var_ := v.Aux.(*Node)
+
+			// Optimization: look back 1 block for the definition.
+			if len(b.Preds) == 1 {
+				c := b.Preds[0].Block()
+				if w := s.defvars[c.ID][var_]; w != nil {
+					v.Op = ssa.OpCopy
+					v.Aux = nil
+					v.AddArg(w)
+					continue
+				}
+			}
+
+			if _, ok := s.varnum[var_]; ok {
+				continue
+			}
+			s.varnum[var_] = int32(len(vartypes))
+			if debugPhi {
+				fmt.Printf("var%d = %v\n", len(vartypes), var_)
+			}
+			vars = append(vars, var_)
+			vartypes = append(vartypes, v.Type)
+		}
+	}
+
+	if len(vartypes) == 0 {
+		return
+	}
+
+	// Find all definitions of the variables we need to process.
+	// defs[n] contains all the blocks in which variable number n is assigned.
+	defs := make([][]*ssa.Block, len(vartypes))
+	for _, b := range s.f.Blocks {
+		for var_ := range s.defvars[b.ID] { // TODO: encode defvars some other way (explicit ops)? make defvars[n] a slice instead of a map.
+			if n, ok := s.varnum[var_]; ok {
+				defs[n] = append(defs[n], b)
+			}
+		}
+	}
+
+	// Make dominator tree.
+	s.idom = s.f.Idom()
+	s.tree = make([]domBlock, s.f.NumBlocks())
+	for _, b := range s.f.Blocks {
+		p := s.idom[b.ID]
+		if p != nil {
+			s.tree[b.ID].sibling = s.tree[p.ID].firstChild
+			s.tree[p.ID].firstChild = b
+		}
+	}
+	// Compute levels in dominator tree.
+	// With parent pointers we can do a depth-first walk without
+	// any auxiliary storage.
+	s.level = make([]int32, s.f.NumBlocks())
+	b := s.f.Entry
+levels:
+	for {
+		if p := s.idom[b.ID]; p != nil {
+			s.level[b.ID] = s.level[p.ID] + 1
+			if debugPhi {
+				fmt.Printf("level %s = %d\n", b, s.level[b.ID])
+			}
+		}
+		if c := s.tree[b.ID].firstChild; c != nil {
+			b = c
+			continue
+		}
+		for {
+			if c := s.tree[b.ID].sibling; c != nil {
+				b = c
+				continue levels
+			}
+			b = s.idom[b.ID]
+			if b == nil {
+				break levels
+			}
+		}
+	}
+
+	// Allocate scratch locations.
+	s.priq.level = s.level
+	s.q = make([]*ssa.Block, 0, s.f.NumBlocks())
+	s.queued = newSparseSet(s.f.NumBlocks())
+	s.hasPhi = newSparseSet(s.f.NumBlocks())
+	s.hasDef = newSparseSet(s.f.NumBlocks())
+	s.placeholder = s.s.entryNewValue0(ssa.OpUnknown, ssa.TypeInvalid)
+
+	// Generate phi ops for each variable.
+	for n := range vartypes {
+		s.insertVarPhis(n, vars[n], defs[n], vartypes[n])
+	}
+
+	// Resolve FwdRefs to the correct write or phi.
+	s.resolveFwdRefs()
+
+	// Erase variable numbers stored in AuxInt fields of phi ops. They are no longer needed.
+	for _, b := range s.f.Blocks {
+		for _, v := range b.Values {
+			if v.Op == ssa.OpPhi {
+				v.AuxInt = 0
+			}
+		}
+	}
+}
+
+func (s *phiState) insertVarPhis(n int, var_ *Node, defs []*ssa.Block, typ ssa.Type) {
+	priq := &s.priq
+	q := s.q
+	queued := s.queued
+	queued.clear()
+	hasPhi := s.hasPhi
+	hasPhi.clear()
+	hasDef := s.hasDef
+	hasDef.clear()
+
+	// Add defining blocks to priority queue.
+	for _, b := range defs {
+		priq.a = append(priq.a, b)
+		hasDef.add(b.ID)
+		if debugPhi {
+			fmt.Printf("def of var%d in %s\n", n, b)
+		}
+	}
+	heap.Init(priq)
+
+	// Visit blocks defining variable n, from deepest to shallowest.
+	for len(priq.a) > 0 {
+		currentRoot := heap.Pop(priq).(*ssa.Block)
+		if debugPhi {
+			fmt.Printf("currentRoot %s\n", currentRoot)
+		}
+		// Walk subtree below definition.
+		// Skip subtrees we've done in previous iterations.
+		// Find edges exiting tree dominated by definition (the dominance frontier).
+		// Insert phis at target blocks.
+		if queued.contains(currentRoot.ID) {
+			s.s.Fatalf("root already in queue")
+		}
+		q = append(q, currentRoot)
+		queued.add(currentRoot.ID)
+		for len(q) > 0 {
+			b := q[len(q)-1]
+			q = q[:len(q)-1]
+			if debugPhi {
+				fmt.Printf("  processing %s\n", b)
+			}
+
+			for _, e := range b.Succs {
+				c := e.Block()
+				// TODO: if the variable is dead at c, skip it.
+				if s.level[c.ID] > s.level[currentRoot.ID] {
+					// a D-edge, or an edge whose target is in currentRoot's subtree.
+					continue
+				}
+				if !hasPhi.contains(c.ID) {
+					// Add a phi to block c for variable n.
+					hasPhi.add(c.ID)
+					v := c.NewValue0I(currentRoot.Line, ssa.OpPhi, typ, int64(n)) // TODO: line number right?
+					// Note: we store the variable number in the phi's AuxInt field. Used temporarily by phi building.
+					s.s.addNamedValue(var_, v)
+					for i := 0; i < len(c.Preds); i++ {
+						v.AddArg(s.placeholder) // Actual args will be filled in by resolveFwdRefs.
+					}
+					if debugPhi {
+						fmt.Printf("new phi for var%d in %s: %s\n", n, c, v)
+					}
+					if !hasDef.contains(c.ID) {
+						// There's now a new definition of this variable in block c.
+						// Add it to the priority queue to explore.
+						heap.Push(priq, c)
+						hasDef.add(c.ID)
+					}
+				}
+			}
+
+			// Visit children if they have not been visited yet.
+			for c := s.tree[b.ID].firstChild; c != nil; c = s.tree[c.ID].sibling {
+				if !queued.contains(c.ID) {
+					q = append(q, c)
+					queued.add(c.ID)
+				}
+			}
+		}
+	}
+}
+
+// resolveFwdRefs links all FwdRef uses up to their nearest dominating definition.
+func (s *phiState) resolveFwdRefs() {
+	// Do a depth-first walk of the dominator tree, keeping track
+	// of the most-recently-seen value for each variable.
+
+	// Map from variable ID to SSA value at the current point of the walk.
+	values := make([]*ssa.Value, len(s.varnum))
+	for i := range values {
+		values[i] = s.placeholder
+	}
+
+	// Stack of work to do.
+	type stackEntry struct {
+		b *ssa.Block // block to explore
+
+		// variable/value pair to reinstate on exit
+		n int32 // variable ID
+		v *ssa.Value
+
+		// Note: only one of b or n,v will be set.
+	}
+	var stk []stackEntry
+
+	stk = append(stk, stackEntry{b: s.f.Entry})
+	for len(stk) > 0 {
+		work := stk[len(stk)-1]
+		stk = stk[:len(stk)-1]
+
+		b := work.b
+		if b == nil {
+			// On exit from a block, this case will undo any assignments done below.
+			values[work.n] = work.v
+			continue
+		}
+
+		// Process phis as new defs. They come before FwdRefs in this block.
+		for _, v := range b.Values {
+			if v.Op != ssa.OpPhi {
+				continue
+			}
+			n := int32(v.AuxInt)
+			// Remember the old assignment so we can undo it when we exit b.
+			stk = append(stk, stackEntry{n: n, v: values[n]})
+			// Record the new assignment.
+			values[n] = v
+		}
+
+		// Replace a FwdRef op with the current incoming value for its variable.
+		for _, v := range b.Values {
+			if v.Op != ssa.OpFwdRef {
+				continue
+			}
+			n := s.varnum[v.Aux.(*Node)]
+			v.Op = ssa.OpCopy
+			v.Aux = nil
+			v.AddArg(values[n])
+		}
+
+		// Establish values for variables defined in b.
+		for var_, v := range s.defvars[b.ID] {
+			n, ok := s.varnum[var_]
+			if !ok {
+				// some variable not live across a basic block boundary.
+				continue
+			}
+			// Remember the old assignment so we can undo it when we exit b.
+			stk = append(stk, stackEntry{n: n, v: values[n]})
+			// Record the new assignment.
+			values[n] = v
+		}
+
+		// Replace phi args in successors with the current incoming value.
+		for _, e := range b.Succs {
+			c, i := e.Block(), e.Index()
+			for j := len(c.Values) - 1; j >= 0; j-- {
+				v := c.Values[j]
+				if v.Op != ssa.OpPhi {
+					break // All phis will be at the end of the block during phi building.
+				}
+				v.SetArg(i, values[v.AuxInt])
+			}
+		}
+
+		// Walk children in dominator tree.
+		for c := s.tree[b.ID].firstChild; c != nil; c = s.tree[c.ID].sibling {
+			stk = append(stk, stackEntry{b: c})
+		}
+	}
+}
+
+// domBlock contains extra per-block information to record the dominator tree.
+type domBlock struct {
+	firstChild *ssa.Block // first child of block in dominator tree
+	sibling    *ssa.Block // next child of parent in dominator tree
+}
+
+// A block heap is used as a priority queue to implement the PiggyBank
+// from Sreedhar and Gao.  That paper uses an array which is better
+// asymptotically but worse in the common case when the PiggyBank
+// holds a sparse set of blocks.
+type blockHeap struct {
+	a     []*ssa.Block // block IDs in heap
+	level []int32      // depth in dominator tree (static, used for determining priority)
+}
+
+func (h *blockHeap) Len() int      { return len(h.a) }
+func (h *blockHeap) Swap(i, j int) { a := h.a; a[i], a[j] = a[j], a[i] }
+
+func (h *blockHeap) Push(x interface{}) {
+	v := x.(*ssa.Block)
+	h.a = append(h.a, v)
+}
+func (h *blockHeap) Pop() interface{} {
+	old := h.a
+	n := len(old)
+	x := old[n-1]
+	h.a = old[:n-1]
+	return x
+}
+func (h *blockHeap) Less(i, j int) bool {
+	return h.level[h.a[i].ID] > h.level[h.a[j].ID]
+}
+
+// TODO: stop walking the iterated domininance frontier when
+// the variable is dead. Maybe detect that by checking if the
+// node we're on is reverse dominated by all the reads?
+// Reverse dominated by the highest common successor of all the reads?
+
+// copy of ../ssa/sparseset.go
+// TODO: move this file to ../ssa, then use sparseSet there.
+type sparseSet struct {
+	dense  []ssa.ID
+	sparse []int32
+}
+
+// newSparseSet returns a sparseSet that can represent
+// integers between 0 and n-1
+func newSparseSet(n int) *sparseSet {
+	return &sparseSet{dense: nil, sparse: make([]int32, n)}
+}
+
+func (s *sparseSet) contains(x ssa.ID) bool {
+	i := s.sparse[x]
+	return i < int32(len(s.dense)) && s.dense[i] == x
+}
+
+func (s *sparseSet) add(x ssa.ID) {
+	i := s.sparse[x]
+	if i < int32(len(s.dense)) && s.dense[i] == x {
+		return
+	}
+	s.dense = append(s.dense, x)
+	s.sparse[x] = int32(len(s.dense)) - 1
+}
+
+func (s *sparseSet) clear() {
+	s.dense = s.dense[:0]
+}
+
+// Variant to use for small functions.
+type simplePhiState struct {
+	s       *state                 // SSA state
+	f       *ssa.Func              // function to work on
+	fwdrefs []*ssa.Value           // list of FwdRefs to be processed
+	defvars []map[*Node]*ssa.Value // defined variables at end of each block
+}
+
+func (s *simplePhiState) insertPhis() {
+	// Find FwdRef ops.
+	for _, b := range s.f.Blocks {
+		for _, v := range b.Values {
+			if v.Op != ssa.OpFwdRef {
+				continue
+			}
+			s.fwdrefs = append(s.fwdrefs, v)
+			var_ := v.Aux.(*Node)
+			if _, ok := s.defvars[b.ID][var_]; !ok {
+				s.defvars[b.ID][var_] = v // treat FwdDefs as definitions.
+			}
+		}
+	}
+
+	var args []*ssa.Value
+
+loop:
+	for len(s.fwdrefs) > 0 {
+		v := s.fwdrefs[len(s.fwdrefs)-1]
+		s.fwdrefs = s.fwdrefs[:len(s.fwdrefs)-1]
+		b := v.Block
+		var_ := v.Aux.(*Node)
+		if len(b.Preds) == 0 {
+			if b == s.f.Entry {
+				// No variable should be live at entry.
+				s.s.Fatalf("Value live at entry. It shouldn't be. func %s, node %v, value %v", s.f.Name, var_, v)
+			}
+			// This block is dead; it has no predecessors and it is not the entry block.
+			// It doesn't matter what we use here as long as it is well-formed.
+			v.Op = ssa.OpUnknown
+			v.Aux = nil
+			continue
+		}
+		// Find variable value on each predecessor.
+		args = args[:0]
+		for _, e := range b.Preds {
+			args = append(args, s.lookupVarOutgoing(e.Block(), v.Type, var_, v.Line))
+		}
+
+		// Decide if we need a phi or not. We need a phi if there
+		// are two different args (which are both not v).
+		var w *ssa.Value
+		for _, a := range args {
+			if a == v {
+				continue // self-reference
+			}
+			if a == w {
+				continue // already have this witness
+			}
+			if w != nil {
+				// two witnesses, need a phi value
+				v.Op = ssa.OpPhi
+				v.AddArgs(args...)
+				v.Aux = nil
+				continue loop
+			}
+			w = a // save witness
+		}
+		if w == nil {
+			s.s.Fatalf("no witness for reachable phi %s", v)
+		}
+		// One witness. Make v a copy of w.
+		v.Op = ssa.OpCopy
+		v.Aux = nil
+		v.AddArg(w)
+	}
+}
+
+// lookupVarOutgoing finds the variable's value at the end of block b.
+func (s *simplePhiState) lookupVarOutgoing(b *ssa.Block, t ssa.Type, var_ *Node, line int32) *ssa.Value {
+	for {
+		if v := s.defvars[b.ID][var_]; v != nil {
+			return v
+		}
+		// The variable is not defined by b and we haven't looked it up yet.
+		// If b has exactly one predecessor, loop to look it up there.
+		// Otherwise, give up and insert a new FwdRef and resolve it later.
+		if len(b.Preds) != 1 {
+			break
+		}
+		b = b.Preds[0].Block()
+	}
+	// Generate a FwdRef for the variable and return that.
+	v := b.NewValue0A(line, ssa.OpFwdRef, t, var_)
+	s.defvars[b.ID][var_] = v
+	s.s.addNamedValue(var_, v)
+	s.fwdrefs = append(s.fwdrefs, v)
+	return v
+}
diff --git a/src/cmd/compile/internal/gc/racewalk.go b/src/cmd/compile/internal/gc/racewalk.go
index b47ac9d1e0..effdcf9ea4 100644
--- a/src/cmd/compile/internal/gc/racewalk.go
+++ b/src/cmd/compile/internal/gc/racewalk.go
@@ -72,6 +72,7 @@ func instrument(fn *Node) {
 		fn.Func.Enter.Prepend(nd)
 		nd = mkcall("racefuncexit", nil, nil)
 		fn.Func.Exit.Append(nd)
+		fn.Func.Dcl = append(fn.Func.Dcl, &nodpc)
 	}
 
 	if Debug['W'] != 0 {
diff --git a/src/cmd/compile/internal/gc/sparselocatephifunctions.go b/src/cmd/compile/internal/gc/sparselocatephifunctions.go
deleted file mode 100644
index 43cc50bd92..0000000000
--- a/src/cmd/compile/internal/gc/sparselocatephifunctions.go
+++ /dev/null
@@ -1,202 +0,0 @@
-// Copyright 2016 The Go Authors. All rights reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-package gc
-
-import (
-	"cmd/compile/internal/ssa"
-	"fmt"
-	"math"
-)
-
-// sparseDefState contains a Go map from ONAMEs (*Node) to sparse definition trees, and
-// a search helper for the CFG's dominator tree in which those definitions are embedded.
-// Once initialized, given a use of an ONAME within a block, the ssa definition for
-// that ONAME can be discovered in time roughly proportional to the log of the number
-// of SSA definitions of that ONAME (thus avoiding pathological quadratic behavior for
-// very large programs).  The helper contains state (a dominator tree numbering) common
-// to all the sparse definition trees, as well as some necessary data obtained from
-// the ssa package.
-//
-// This algorithm has improved asymptotic complexity, but the constant factor is
-// rather large and thus it is only preferred for very large inputs containing
-// 1000s of blocks and variables.
-type sparseDefState struct {
-	helper         *ssa.SparseTreeHelper // contains one copy of information needed to do sparse mapping
-	defmapForOname map[*Node]*onameDefs  // for each ONAME, its definition set (normal and phi)
-}
-
-// onameDefs contains a record of definitions (ordinary and implied phi function) for a single OName.
-// stm is the set of definitions for the OName.
-// firstdef and lastuse are postorder block numberings that
-// conservatively bracket the entire lifetime of the OName.
-type onameDefs struct {
-	stm *ssa.SparseTreeMap
-	// firstdef and lastuse define an interval in the postorder numbering
-	// that is guaranteed to include the entire lifetime of an ONAME.
-	// In the postorder numbering, math.MaxInt32 is before anything,
-	// and 0 is after-or-equal all exit nodes and infinite loops.
-	firstdef int32 // the first definition of this ONAME *in the postorder numbering*
-	lastuse  int32 // the last use of this ONAME *in the postorder numbering*
-}
-
-// defsFor finds or creates-and-inserts-in-map the definition information
-// (sparse tree and live range) for a given OName.
-func (m *sparseDefState) defsFor(n *Node) *onameDefs {
-	d := m.defmapForOname[n]
-	if d != nil {
-		return d
-	}
-	// Reminder: firstdef/lastuse are postorder indices, not block indices,
-	// so these default values define an empty interval, not the entire one.
-	d = &onameDefs{stm: m.helper.NewTree(), firstdef: 0, lastuse: math.MaxInt32}
-	m.defmapForOname[n] = d
-	return d
-}
-
-// Insert adds a definition at b (with specified before/within/after adjustment)
-// to sparse tree onameDefs.  The lifetime is extended as necessary.
-func (m *sparseDefState) Insert(tree *onameDefs, b *ssa.Block, adjust int32) {
-	bponum := m.helper.Ponums[b.ID]
-	if bponum > tree.firstdef {
-		tree.firstdef = bponum
-	}
-	tree.stm.Insert(b, adjust, b, m.helper)
-}
-
-// Use updates tree to record a use within b, extending the lifetime as necessary.
-func (m *sparseDefState) Use(tree *onameDefs, b *ssa.Block) {
-	bponum := m.helper.Ponums[b.ID]
-	if bponum < tree.lastuse {
-		tree.lastuse = bponum
-	}
-}
-
-// locatePotentialPhiFunctions finds all the places where phi functions
-// will be inserted into a program and records those and ordinary definitions
-// in a "map" (not a Go map) that given an OName and use site, returns the
-// SSA definition for that OName that will reach the use site (that is,
-// the use site's nearest def/phi site in the dominator tree.)
-func (s *state) locatePotentialPhiFunctions(fn *Node) *sparseDefState {
-	// s.config.SparsePhiCutoff() is compared with product of numblocks and numvalues,
-	// if product is smaller than cutoff, use old non-sparse method.
-	// cutoff == 0 implies all sparse
-	// cutoff == uint(-1) implies all non-sparse
-	if uint64(s.f.NumValues())*uint64(s.f.NumBlocks()) < s.config.SparsePhiCutoff() {
-		return nil
-	}
-
-	helper := ssa.NewSparseTreeHelper(s.f)
-	po := helper.Po // index by block.ID to obtain postorder # of block.
-	trees := make(map[*Node]*onameDefs)
-	dm := &sparseDefState{defmapForOname: trees, helper: helper}
-
-	// Process params, taking note of their special lifetimes
-	b := s.f.Entry
-	for _, n := range fn.Func.Dcl {
-		switch n.Class {
-		case PPARAM, PPARAMOUT:
-			t := dm.defsFor(n)
-			dm.Insert(t, b, ssa.AdjustBefore) // define param at entry block
-			if n.Class == PPARAMOUT {
-				dm.Use(t, po[0]) // Explicitly use PPARAMOUT at very last block
-			}
-		default:
-		}
-	}
-
-	// Process memory variable.
-	t := dm.defsFor(&memVar)
-	dm.Insert(t, b, ssa.AdjustBefore) // define memory at entry block
-	dm.Use(t, po[0])                  // Explicitly use memory at last block
-
-	// Next load the map w/ basic definitions for ONames recorded per-block
-	// Iterate over po to avoid unreachable blocks.
-	for i := len(po) - 1; i >= 0; i-- {
-		b := po[i]
-		m := s.defvars[b.ID]
-		for n := range m { // no specified order, but per-node trees are independent.
-			t := dm.defsFor(n)
-			dm.Insert(t, b, ssa.AdjustWithin)
-		}
-	}
-
-	// Find last use of each variable
-	for _, v := range s.fwdRefs {
-		b := v.Block
-		name := v.Aux.(*Node)
-		t := dm.defsFor(name)
-		dm.Use(t, b)
-	}
-
-	for _, t := range trees {
-		// iterating over names in the outer loop
-		for change := true; change; {
-			change = false
-			for i := t.firstdef; i >= t.lastuse; i-- {
-				// Iterating in reverse of post-order reduces number of 'change' iterations;
-				// all possible forward flow goes through each time.
-				b := po[i]
-				// Within tree t, would a use at b require a phi function to ensure a single definition?
-				// TODO: perhaps more efficient to record specific use sites instead of range?
-				if len(b.Preds) < 2 {
-					continue // no phi possible
-				}
-				phi := t.stm.Find(b, ssa.AdjustWithin, helper) // Look for defs in earlier block or AdjustBefore in this one.
-				if phi != nil && phi.(*ssa.Block) == b {
-					continue // has a phi already in this block.
-				}
-				var defseen interface{}
-				// Do preds see different definitions? if so, need a phi function.
-				for _, e := range b.Preds {
-					p := e.Block()
-					dm.Use(t, p)                                // always count phi pred as "use"; no-op except for loop edges, which matter.
-					x := t.stm.Find(p, ssa.AdjustAfter, helper) // Look for defs reaching or within predecessors.
-					if x == nil {                               // nil def from a predecessor means a backedge that will be visited soon.
-						continue
-					}
-					if defseen == nil {
-						defseen = x
-					}
-					if defseen != x {
-						// Need to insert a phi function here because predecessors's definitions differ.
-						change = true
-						// Phi insertion is at AdjustBefore, visible with find in same block at AdjustWithin or AdjustAfter.
-						dm.Insert(t, b, ssa.AdjustBefore)
-						break
-					}
-				}
-			}
-		}
-	}
-	return dm
-}
-
-// FindBetterDefiningBlock tries to find a better block for a definition of OName name
-// reaching (or within) p than p itself.  If it cannot, it returns p instead.
-// This aids in more efficient location of phi functions, since it can skip over
-// branch code that might contain a definition of name if it actually does not.
-func (m *sparseDefState) FindBetterDefiningBlock(name *Node, p *ssa.Block) *ssa.Block {
-	if m == nil {
-		return p
-	}
-	t := m.defmapForOname[name]
-	// For now this is fail-soft, since the old algorithm still works using the unimproved block.
-	if t == nil {
-		return p
-	}
-	x := t.stm.Find(p, ssa.AdjustAfter, m.helper)
-	if x == nil {
-		return p
-	}
-	b := x.(*ssa.Block)
-	if b == nil {
-		return p
-	}
-	return b
-}
-
-func (d *onameDefs) String() string {
-	return fmt.Sprintf("onameDefs:first=%d,last=%d,tree=%s", d.firstdef, d.lastuse, d.stm.String())
-}
diff --git a/src/cmd/compile/internal/gc/ssa.go b/src/cmd/compile/internal/gc/ssa.go
index 419240eab7..741a9d4e86 100644
--- a/src/cmd/compile/internal/gc/ssa.go
+++ b/src/cmd/compile/internal/gc/ssa.go
@@ -80,6 +80,7 @@ func buildssa(fn *Node) *ssa.Func {
 	// Allocate starting values
 	s.labels = map[string]*ssaLabel{}
 	s.labeledNodes = map[*Node]*ssaLabel{}
+	s.fwdVars = map[*Node]*ssa.Value{}
 	s.startmem = s.entryNewValue0(ssa.OpInitMem, ssa.TypeMem)
 	s.sp = s.entryNewValue0(ssa.OpSP, Types[TUINTPTR]) // TODO: use generic pointer type (unsafe.Pointer?) instead
 	s.sb = s.entryNewValue0(ssa.OpSB, Types[TUINTPTR])
@@ -114,6 +115,21 @@ func buildssa(fn *Node) *ssa.Func {
 		}
 	}
 
+	// Populate arguments.
+	for _, n := range fn.Func.Dcl {
+		if n.Class != PPARAM {
+			continue
+		}
+		var v *ssa.Value
+		if s.canSSA(n) {
+			v = s.newValue0A(ssa.OpArg, n.Type, n)
+		} else {
+			// Not SSAable. Load it.
+			v = s.newValue2(ssa.OpLoad, n.Type, s.decladdrs[n], s.startmem)
+		}
+		s.vars[n] = v
+	}
+
 	// Convert the AST-based IR to the SSA-based IR
 	s.stmts(fn.Func.Enter)
 	s.stmts(fn.Nbody)
@@ -151,16 +167,7 @@ func buildssa(fn *Node) *ssa.Func {
 		return nil
 	}
 
-	prelinkNumvars := s.f.NumValues()
-	sparseDefState := s.locatePotentialPhiFunctions(fn)
-
-	// Link up variable uses to variable definitions
-	s.linkForwardReferences(sparseDefState)
-
-	if ssa.BuildStats > 0 {
-		s.f.LogStat("build", s.f.NumBlocks(), "blocks", prelinkNumvars, "vars_before",
-			s.f.NumValues(), "vars_after", prelinkNumvars*s.f.NumBlocks(), "ssa_phi_loc_cutoff_score")
-	}
+	s.insertPhis()
 
 	// Don't carry reference this around longer than necessary
 	s.exitCode = Nodes{}
@@ -197,8 +204,14 @@ type state struct {
 
 	// variable assignments in the current block (map from variable symbol to ssa value)
 	// *Node is the unique identifier (an ONAME Node) for the variable.
+	// TODO: keep a single varnum map, then make all of these maps slices instead?
 	vars map[*Node]*ssa.Value
 
+	// fwdVars are variables that are used before they are defined in the current block.
+	// This map exists just to coalesce multiple references into a single FwdRef op.
+	// *Node is the unique identifier (an ONAME Node) for the variable.
+	fwdVars map[*Node]*ssa.Value
+
 	// all defined variables at the end of each block. Indexed by block ID.
 	defvars []map[*Node]*ssa.Value
 
@@ -220,12 +233,12 @@ type state struct {
 	// Used to deduplicate panic calls.
 	panics map[funcLine]*ssa.Block
 
-	// list of FwdRef values.
-	fwdRefs []*ssa.Value
-
 	// list of PPARAMOUT (return) variables.
 	returns []*Node
 
+	// A dummy value used during phi construction.
+	placeholder *ssa.Value
+
 	cgoUnsafeArgs bool
 	noWB          bool
 	WBLineno      int32 // line number of first write barrier. 0=no write barriers
@@ -292,6 +305,9 @@ func (s *state) startBlock(b *ssa.Block) {
 	}
 	s.curBlock = b
 	s.vars = map[*Node]*ssa.Value{}
+	for n := range s.fwdVars {
+		delete(s.fwdVars, n)
+	}
 }
 
 // endBlock marks the end of generating code for the current block.
@@ -2951,9 +2967,8 @@ func (s *state) addr(n *Node, bounded bool) (*ssa.Value, bool) {
 			if v != nil {
 				return v, false
 			}
-			if n.String() == ".fp" {
-				// Special arg that points to the frame pointer.
-				// (Used by the race detector, others?)
+			if n == nodfp {
+				// Special arg that points to the frame pointer (Used by ORECOVER).
 				aux := s.lookupSymbol(n, &ssa.ArgSymbol{Typ: n.Type, Node: n})
 				return s.entryNewValue1A(ssa.OpAddr, t, aux, s.sp), false
 			}
@@ -3971,12 +3986,23 @@ func (s *state) checkgoto(from *Node, to *Node) {
 // variable returns the value of a variable at the current location.
 func (s *state) variable(name *Node, t ssa.Type) *ssa.Value {
 	v := s.vars[name]
-	if v == nil {
-		v = s.newValue0A(ssa.OpFwdRef, t, name)
-		s.fwdRefs = append(s.fwdRefs, v)
-		s.vars[name] = v
-		s.addNamedValue(name, v)
+	if v != nil {
+		return v
 	}
+	v = s.fwdVars[name]
+	if v != nil {
+		return v
+	}
+
+	if s.curBlock == s.f.Entry {
+		// No variable should be live at entry.
+		s.Fatalf("Value live at entry. It shouldn't be. func %s, node %v, value %v", s.f.Name, name, v)
+	}
+	// Make a FwdRef, which records a value that's live on block input.
+	// We'll find the matching definition as part of insertPhis.
+	v = s.newValue0A(ssa.OpFwdRef, t, name)
+	s.fwdVars[name] = v
+	s.addNamedValue(name, v)
 	return v
 }
 
@@ -3984,119 +4010,6 @@ func (s *state) mem() *ssa.Value {
 	return s.variable(&memVar, ssa.TypeMem)
 }
 
-func (s *state) linkForwardReferences(dm *sparseDefState) {
-
-	// Build SSA graph. Each variable on its first use in a basic block
-	// leaves a FwdRef in that block representing the incoming value
-	// of that variable. This function links that ref up with possible definitions,
-	// inserting Phi values as needed. This is essentially the algorithm
-	// described by Braun, Buchwald, Hack, Leißa, Mallon, and Zwinkau:
-	// http://pp.info.uni-karlsruhe.de/uploads/publikationen/braun13cc.pdf
-	// Differences:
-	//   - We use FwdRef nodes to postpone phi building until the CFG is
-	//     completely built. That way we can avoid the notion of "sealed"
-	//     blocks.
-	//   - Phi optimization is a separate pass (in ../ssa/phielim.go).
-	for len(s.fwdRefs) > 0 {
-		v := s.fwdRefs[len(s.fwdRefs)-1]
-		s.fwdRefs = s.fwdRefs[:len(s.fwdRefs)-1]
-		s.resolveFwdRef(v, dm)
-	}
-}
-
-// resolveFwdRef modifies v to be the variable's value at the start of its block.
-// v must be a FwdRef op.
-func (s *state) resolveFwdRef(v *ssa.Value, dm *sparseDefState) {
-	b := v.Block
-	name := v.Aux.(*Node)
-	v.Aux = nil
-	if b == s.f.Entry {
-		// Live variable at start of function.
-		if s.canSSA(name) {
-			if strings.HasPrefix(name.Sym.Name, "autotmp_") {
-				// It's likely that this is an uninitialized variable in the entry block.
-				s.Fatalf("Treating auto as if it were arg, func %s, node %v, value %v", b.Func.Name, name, v)
-			}
-			v.Op = ssa.OpArg
-			v.Aux = name
-			return
-		}
-		// Not SSAable. Load it.
-		addr := s.decladdrs[name]
-		if addr == nil {
-			// TODO: closure args reach here.
-			s.Fatalf("unhandled closure arg %v at entry to function %s", name, b.Func.Name)
-		}
-		if _, ok := addr.Aux.(*ssa.ArgSymbol); !ok {
-			s.Fatalf("variable live at start of function %s is not an argument %v", b.Func.Name, name)
-		}
-		v.Op = ssa.OpLoad
-		v.AddArgs(addr, s.startmem)
-		return
-	}
-	if len(b.Preds) == 0 {
-		// This block is dead; we have no predecessors and we're not the entry block.
-		// It doesn't matter what we use here as long as it is well-formed.
-		v.Op = ssa.OpUnknown
-		return
-	}
-	// Find variable value on each predecessor.
-	var argstore [4]*ssa.Value
-	args := argstore[:0]
-	for _, e := range b.Preds {
-		p := e.Block()
-		p = dm.FindBetterDefiningBlock(name, p) // try sparse improvement on p
-		args = append(args, s.lookupVarOutgoing(p, v.Type, name, v.Line))
-	}
-
-	// Decide if we need a phi or not. We need a phi if there
-	// are two different args (which are both not v).
-	var w *ssa.Value
-	for _, a := range args {
-		if a == v {
-			continue // self-reference
-		}
-		if a == w {
-			continue // already have this witness
-		}
-		if w != nil {
-			// two witnesses, need a phi value
-			v.Op = ssa.OpPhi
-			v.AddArgs(args...)
-			return
-		}
-		w = a // save witness
-	}
-	if w == nil {
-		s.Fatalf("no witness for reachable phi %s", v)
-	}
-	// One witness. Make v a copy of w.
-	v.Op = ssa.OpCopy
-	v.AddArg(w)
-}
-
-// lookupVarOutgoing finds the variable's value at the end of block b.
-func (s *state) lookupVarOutgoing(b *ssa.Block, t ssa.Type, name *Node, line int32) *ssa.Value {
-	for {
-		if v, ok := s.defvars[b.ID][name]; ok {
-			return v
-		}
-		// The variable is not defined by b and we haven't looked it up yet.
-		// If b has exactly one predecessor, loop to look it up there.
-		// Otherwise, give up and insert a new FwdRef and resolve it later.
-		if len(b.Preds) != 1 {
-			break
-		}
-		b = b.Preds[0].Block()
-	}
-	// Generate a FwdRef for the variable and return that.
-	v := b.NewValue0A(line, ssa.OpFwdRef, t, name)
-	s.fwdRefs = append(s.fwdRefs, v)
-	s.defvars[b.ID][name] = v
-	s.addNamedValue(name, v)
-	return v
-}
-
 func (s *state) addNamedValue(n *Node, v *ssa.Value) {
 	if n.Class == Pxxx {
 		// Don't track our dummy nodes (&memVar etc.).
diff --git a/src/cmd/compile/internal/ssa/block.go b/src/cmd/compile/internal/ssa/block.go
index b5bedd3912..3ee27df5e7 100644
--- a/src/cmd/compile/internal/ssa/block.go
+++ b/src/cmd/compile/internal/ssa/block.go
@@ -89,6 +89,9 @@ type Edge struct {
 func (e Edge) Block() *Block {
 	return e.b
 }
+func (e Edge) Index() int {
+	return e.i
+}
 
 //     kind           control    successors
 //   ------------------------------------------
diff --git a/src/cmd/compile/internal/ssa/func.go b/src/cmd/compile/internal/ssa/func.go
index 759e19d8e6..6ba5448998 100644
--- a/src/cmd/compile/internal/ssa/func.go
+++ b/src/cmd/compile/internal/ssa/func.go
@@ -459,6 +459,9 @@ func (f *Func) idom() []*Block {
 	}
 	return f.cachedIdom
 }
+func (f *Func) Idom() []*Block {
+	return f.idom()
+}
 
 // sdom returns a sparse tree representing the dominator relationships
 // among the blocks of f.