src/cmd/compile/internal/walk/switch.go

// Copyright 2009 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 walk

import (
	"go/constant"
	"go/token"
	"sort"

	"cmd/compile/internal/base"
	"cmd/compile/internal/ir"
	"cmd/compile/internal/ssagen"
	"cmd/compile/internal/typecheck"
	"cmd/compile/internal/types"
	"cmd/internal/src"
)

// walkSwitch walks a switch statement.
func walkSwitch(sw *ir.SwitchStmt) {
	// Guard against double walk, see #25776.
	if sw.Walked() {
		return // Was fatal, but eliminating every possible source of double-walking is hard
	}
	sw.SetWalked(true)

	if sw.Tag != nil && sw.Tag.Op() == ir.OTYPESW {
		walkSwitchType(sw)
	} else {
		walkSwitchExpr(sw)
	}
}

// walkSwitchExpr generates an AST implementing sw.  sw is an
// expression switch.
func walkSwitchExpr(sw *ir.SwitchStmt) {
	lno := ir.SetPos(sw)

	cond := sw.Tag
	sw.Tag = nil

	// convert switch {...} to switch true {...}
	if cond == nil {
		cond = ir.NewBool(true)
		cond = typecheck.Expr(cond)
		cond = typecheck.DefaultLit(cond, nil)
	}

	// Given "switch string(byteslice)",
	// with all cases being side-effect free,
	// use a zero-cost alias of the byte slice.
	// Do this before calling walkExpr on cond,
	// because walkExpr will lower the string
	// conversion into a runtime call.
	// See issue 24937 for more discussion.
	if cond.Op() == ir.OBYTES2STR && allCaseExprsAreSideEffectFree(sw) {
		cond := cond.(*ir.ConvExpr)
		cond.SetOp(ir.OBYTES2STRTMP)
	}

	cond = walkExpr(cond, sw.PtrInit())
	if cond.Op() != ir.OLITERAL && cond.Op() != ir.ONIL {
		cond = copyExpr(cond, cond.Type(), &sw.Compiled)
	}

	base.Pos = lno

	s := exprSwitch{
		pos:      lno,
		exprname: cond,
	}

	var defaultGoto ir.Node
	var body ir.Nodes
	for _, ncase := range sw.Cases {
		label := typecheck.AutoLabel(".s")
		jmp := ir.NewBranchStmt(ncase.Pos(), ir.OGOTO, label)

		// Process case dispatch.
		if len(ncase.List) == 0 {
			if defaultGoto != nil {
				base.Fatalf("duplicate default case not detected during typechecking")
			}
			defaultGoto = jmp
		}

		for i, n1 := range ncase.List {
			var rtype ir.Node
			if i < len(ncase.RTypes) {
				rtype = ncase.RTypes[i]
			}
			s.Add(ncase.Pos(), n1, rtype, jmp)
		}

		// Process body.
		body.Append(ir.NewLabelStmt(ncase.Pos(), label))
		body.Append(ncase.Body...)
		if fall, pos := endsInFallthrough(ncase.Body); !fall {
			br := ir.NewBranchStmt(base.Pos, ir.OBREAK, nil)
			br.SetPos(pos)
			body.Append(br)
		}
	}
	sw.Cases = nil

	if defaultGoto == nil {
		br := ir.NewBranchStmt(base.Pos, ir.OBREAK, nil)
		br.SetPos(br.Pos().WithNotStmt())
		defaultGoto = br
	}

	s.Emit(&sw.Compiled)
	sw.Compiled.Append(defaultGoto)
	sw.Compiled.Append(body.Take()...)
	walkStmtList(sw.Compiled)
}

// An exprSwitch walks an expression switch.
type exprSwitch struct {
	pos      src.XPos
	exprname ir.Node // value being switched on

	done    ir.Nodes
	clauses []exprClause
}

type exprClause struct {
	pos    src.XPos
	lo, hi ir.Node
	rtype  ir.Node // *runtime._type for OEQ node
	jmp    ir.Node
}

func (s *exprSwitch) Add(pos src.XPos, expr, rtype, jmp ir.Node) {
	c := exprClause{pos: pos, lo: expr, hi: expr, rtype: rtype, jmp: jmp}
	if types.IsOrdered[s.exprname.Type().Kind()] && expr.Op() == ir.OLITERAL {
		s.clauses = append(s.clauses, c)
		return
	}

	s.flush()
	s.clauses = append(s.clauses, c)
	s.flush()
}

func (s *exprSwitch) Emit(out *ir.Nodes) {
	s.flush()
	out.Append(s.done.Take()...)
}

func (s *exprSwitch) flush() {
	cc := s.clauses
	s.clauses = nil
	if len(cc) == 0 {
		return
	}

	// Caution: If len(cc) == 1, then cc[0] might not an OLITERAL.
	// The code below is structured to implicitly handle this case
	// (e.g., sort.Slice doesn't need to invoke the less function
	// when there's only a single slice element).

	if s.exprname.Type().IsString() && len(cc) >= 2 {
		// Sort strings by length and then by value. It is
		// much cheaper to compare lengths than values, and
		// all we need here is consistency. We respect this
		// sorting below.
		sort.Slice(cc, func(i, j int) bool {
			si := ir.StringVal(cc[i].lo)
			sj := ir.StringVal(cc[j].lo)
			if len(si) != len(sj) {
				return len(si) < len(sj)
			}
			return si < sj
		})

		// runLen returns the string length associated with a
		// particular run of exprClauses.
		runLen := func(run []exprClause) int64 { return int64(len(ir.StringVal(run[0].lo))) }

		// Collapse runs of consecutive strings with the same length.
		var runs [][]exprClause
		start := 0
		for i := 1; i < len(cc); i++ {
			if runLen(cc[start:]) != runLen(cc[i:]) {
				runs = append(runs, cc[start:i])
				start = i
			}
		}
		runs = append(runs, cc[start:])

		// We have strings of more than one length. Generate an
		// outer switch which switches on the length of the string
		// and an inner switch in each case which resolves all the
		// strings of the same length. The code looks something like this:

		// goto outerLabel
		// len5:
		//   ... search among length 5 strings ...
		//   goto endLabel
		// len8:
		//   ... search among length 8 strings ...
		//   goto endLabel
		// ... other lengths ...
		// outerLabel:
		// switch len(s) {
		//   case 5: goto len5
		//   case 8: goto len8
		//   ... other lengths ...
		// }
		// endLabel:

		outerLabel := typecheck.AutoLabel(".s")
		endLabel := typecheck.AutoLabel(".s")

		// Jump around all the individual switches for each length.
		s.done.Append(ir.NewBranchStmt(s.pos, ir.OGOTO, outerLabel))

		var outer exprSwitch
		outer.exprname = ir.NewUnaryExpr(s.pos, ir.OLEN, s.exprname)
		outer.exprname.SetType(types.Types[types.TINT])

		for _, run := range runs {
			// Target label to jump to when we match this length.
			label := typecheck.AutoLabel(".s")

			// Search within this run of same-length strings.
			pos := run[0].pos
			s.done.Append(ir.NewLabelStmt(pos, label))
			stringSearch(s.exprname, run, &s.done)
			s.done.Append(ir.NewBranchStmt(pos, ir.OGOTO, endLabel))

			// Add length case to outer switch.
			cas := ir.NewBasicLit(pos, constant.MakeInt64(runLen(run)))
			jmp := ir.NewBranchStmt(pos, ir.OGOTO, label)
			outer.Add(pos, cas, nil, jmp)
		}
		s.done.Append(ir.NewLabelStmt(s.pos, outerLabel))
		outer.Emit(&s.done)
		s.done.Append(ir.NewLabelStmt(s.pos, endLabel))
		return
	}

	sort.Slice(cc, func(i, j int) bool {
		return constant.Compare(cc[i].lo.Val(), token.LSS, cc[j].lo.Val())
	})

	// Merge consecutive integer cases.
	if s.exprname.Type().IsInteger() {
		consecutive := func(last, next constant.Value) bool {
			delta := constant.BinaryOp(next, token.SUB, last)
			return constant.Compare(delta, token.EQL, constant.MakeInt64(1))
		}

		merged := cc[:1]
		for _, c := range cc[1:] {
			last := &merged[len(merged)-1]
			if last.jmp == c.jmp && consecutive(last.hi.Val(), c.lo.Val()) {
				last.hi = c.lo
			} else {
				merged = append(merged, c)
			}
		}
		cc = merged
	}

	s.search(cc, &s.done)
}

func (s *exprSwitch) search(cc []exprClause, out *ir.Nodes) {
	if s.tryJumpTable(cc, out) {
		return
	}
	binarySearch(len(cc), out,
		func(i int) ir.Node {
			return ir.NewBinaryExpr(base.Pos, ir.OLE, s.exprname, cc[i-1].hi)
		},
		func(i int, nif *ir.IfStmt) {
			c := &cc[i]
			nif.Cond = c.test(s.exprname)
			nif.Body = []ir.Node{c.jmp}
		},
	)
}

// Try to implement the clauses with a jump table. Returns true if successful.
func (s *exprSwitch) tryJumpTable(cc []exprClause, out *ir.Nodes) bool {
	const minCases = 8   // have at least minCases cases in the switch
	const minDensity = 4 // use at least 1 out of every minDensity entries

	if base.Flag.N != 0 || !ssagen.Arch.LinkArch.CanJumpTable || base.Ctxt.Retpoline {
		return false
	}
	if len(cc) < minCases {
		return false // not enough cases for it to be worth it
	}
	if cc[0].lo.Val().Kind() != constant.Int {
		return false // e.g. float
	}
	if s.exprname.Type().Size() > int64(types.PtrSize) {
		return false // 64-bit switches on 32-bit archs
	}
	min := cc[0].lo.Val()
	max := cc[len(cc)-1].hi.Val()
	width := constant.BinaryOp(constant.BinaryOp(max, token.SUB, min), token.ADD, constant.MakeInt64(1))
	limit := constant.MakeInt64(int64(len(cc)) * minDensity)
	if constant.Compare(width, token.GTR, limit) {
		// We disable jump tables if we use less than a minimum fraction of the entries.
		// i.e. for switch x {case 0: case 1000: case 2000:} we don't want to use a jump table.
		return false
	}
	jt := ir.NewJumpTableStmt(base.Pos, s.exprname)
	for _, c := range cc {
		jmp := c.jmp.(*ir.BranchStmt)
		if jmp.Op() != ir.OGOTO || jmp.Label == nil {
			panic("bad switch case body")
		}
		for i := c.lo.Val(); constant.Compare(i, token.LEQ, c.hi.Val()); i = constant.BinaryOp(i, token.ADD, constant.MakeInt64(1)) {
			jt.Cases = append(jt.Cases, i)
			jt.Targets = append(jt.Targets, jmp.Label)
		}
	}
	out.Append(jt)
	return true
}

func (c *exprClause) test(exprname ir.Node) ir.Node {
	// Integer range.
	if c.hi != c.lo {
		low := ir.NewBinaryExpr(c.pos, ir.OGE, exprname, c.lo)
		high := ir.NewBinaryExpr(c.pos, ir.OLE, exprname, c.hi)
		return ir.NewLogicalExpr(c.pos, ir.OANDAND, low, high)
	}

	// Optimize "switch true { ...}" and "switch false { ... }".
	if ir.IsConst(exprname, constant.Bool) && !c.lo.Type().IsInterface() {
		if ir.BoolVal(exprname) {
			return c.lo
		} else {
			return ir.NewUnaryExpr(c.pos, ir.ONOT, c.lo)
		}
	}

	n := ir.NewBinaryExpr(c.pos, ir.OEQ, exprname, c.lo)
	n.RType = c.rtype
	return n
}

func allCaseExprsAreSideEffectFree(sw *ir.SwitchStmt) bool {
	// In theory, we could be more aggressive, allowing any
	// side-effect-free expressions in cases, but it's a bit
	// tricky because some of that information is unavailable due
	// to the introduction of temporaries during order.
	// Restricting to constants is simple and probably powerful
	// enough.

	for _, ncase := range sw.Cases {
		for _, v := range ncase.List {
			if v.Op() != ir.OLITERAL {
				return false
			}
		}
	}
	return true
}

// endsInFallthrough reports whether stmts ends with a "fallthrough" statement.
func endsInFallthrough(stmts []ir.Node) (bool, src.XPos) {
	if len(stmts) == 0 {
		return false, src.NoXPos
	}
	i := len(stmts) - 1
	return stmts[i].Op() == ir.OFALL, stmts[i].Pos()
}

// walkSwitchType generates an AST that implements sw, where sw is a
// type switch.
func walkSwitchType(sw *ir.SwitchStmt) {
	var s typeSwitch
	s.facename = sw.Tag.(*ir.TypeSwitchGuard).X
	sw.Tag = nil

	s.facename = walkExpr(s.facename, sw.PtrInit())
	s.facename = copyExpr(s.facename, s.facename.Type(), &sw.Compiled)
	s.okname = typecheck.Temp(types.Types[types.TBOOL])

	// Get interface descriptor word.
	// For empty interfaces this will be the type.
	// For non-empty interfaces this will be the itab.
	itab := ir.NewUnaryExpr(base.Pos, ir.OITAB, s.facename)

	// For empty interfaces, do:
	//     if e._type == nil {
	//         do nil case if it exists, otherwise default
	//     }
	//     h := e._type.hash
	// Use a similar strategy for non-empty interfaces.
	ifNil := ir.NewIfStmt(base.Pos, nil, nil, nil)
	ifNil.Cond = ir.NewBinaryExpr(base.Pos, ir.OEQ, itab, typecheck.NodNil())
	base.Pos = base.Pos.WithNotStmt() // disable statement marks after the first check.
	ifNil.Cond = typecheck.Expr(ifNil.Cond)
	ifNil.Cond = typecheck.DefaultLit(ifNil.Cond, nil)
	// ifNil.Nbody assigned at end.
	sw.Compiled.Append(ifNil)

	// Load hash from type or itab.
	dotHash := typeHashFieldOf(base.Pos, itab)
	s.hashname = copyExpr(dotHash, dotHash.Type(), &sw.Compiled)

	br := ir.NewBranchStmt(base.Pos, ir.OBREAK, nil)
	var defaultGoto, nilGoto ir.Node
	var body ir.Nodes
	for _, ncase := range sw.Cases {
		caseVar := ncase.Var

		// For single-type cases with an interface type,
		// we initialize the case variable as part of the type assertion.
		// In other cases, we initialize it in the body.
		var singleType *types.Type
		if len(ncase.List) == 1 && ncase.List[0].Op() == ir.OTYPE {
			singleType = ncase.List[0].Type()
		}
		caseVarInitialized := false

		label := typecheck.AutoLabel(".s")
		jmp := ir.NewBranchStmt(ncase.Pos(), ir.OGOTO, label)

		if len(ncase.List) == 0 { // default:
			if defaultGoto != nil {
				base.Fatalf("duplicate default case not detected during typechecking")
			}
			defaultGoto = jmp
		}

		for _, n1 := range ncase.List {
			if ir.IsNil(n1) { // case nil:
				if nilGoto != nil {
					base.Fatalf("duplicate nil case not detected during typechecking")
				}
				nilGoto = jmp
				continue
			}

			if singleType != nil && singleType.IsInterface() {
				s.Add(ncase.Pos(), n1, caseVar, jmp)
				caseVarInitialized = true
			} else {
				s.Add(ncase.Pos(), n1, nil, jmp)
			}
		}

		body.Append(ir.NewLabelStmt(ncase.Pos(), label))
		if caseVar != nil && !caseVarInitialized {
			val := s.facename
			if singleType != nil {
				// We have a single concrete type. Extract the data.
				if singleType.IsInterface() {
					base.Fatalf("singleType interface should have been handled in Add")
				}
				val = ifaceData(ncase.Pos(), s.facename, singleType)
			}
			if len(ncase.List) == 1 && ncase.List[0].Op() == ir.ODYNAMICTYPE {
				dt := ncase.List[0].(*ir.DynamicType)
				x := ir.NewDynamicTypeAssertExpr(ncase.Pos(), ir.ODYNAMICDOTTYPE, val, dt.RType)
				x.ITab = dt.ITab
				x.SetType(caseVar.Type())
				x.SetTypecheck(1)
				val = x
			}
			l := []ir.Node{
				ir.NewDecl(ncase.Pos(), ir.ODCL, caseVar),
				ir.NewAssignStmt(ncase.Pos(), caseVar, val),
			}
			typecheck.Stmts(l)
			body.Append(l...)
		}
		body.Append(ncase.Body...)
		body.Append(br)
	}
	sw.Cases = nil

	if defaultGoto == nil {
		defaultGoto = br
	}
	if nilGoto == nil {
		nilGoto = defaultGoto
	}
	ifNil.Body = []ir.Node{nilGoto}

	s.Emit(&sw.Compiled)
	sw.Compiled.Append(defaultGoto)
	sw.Compiled.Append(body.Take()...)

	walkStmtList(sw.Compiled)
}

// typeHashFieldOf returns an expression to select the type hash field
// from an interface's descriptor word (whether a *runtime._type or
// *runtime.itab pointer).
func typeHashFieldOf(pos src.XPos, itab *ir.UnaryExpr) *ir.SelectorExpr {
	if itab.Op() != ir.OITAB {
		base.Fatalf("expected OITAB, got %v", itab.Op())
	}
	var hashField *types.Field
	if itab.X.Type().IsEmptyInterface() {
		// runtime._type's hash field
		if rtypeHashField == nil {
			rtypeHashField = runtimeField("hash", int64(2*types.PtrSize), types.Types[types.TUINT32])
		}
		hashField = rtypeHashField
	} else {
		// runtime.itab's hash field
		if itabHashField == nil {
			itabHashField = runtimeField("hash", int64(2*types.PtrSize), types.Types[types.TUINT32])
		}
		hashField = itabHashField
	}
	return boundedDotPtr(pos, itab, hashField)
}

var rtypeHashField, itabHashField *types.Field

// A typeSwitch walks a type switch.
type typeSwitch struct {
	// Temporary variables (i.e., ONAMEs) used by type switch dispatch logic:
	facename ir.Node // value being type-switched on
	hashname ir.Node // type hash of the value being type-switched on
	okname   ir.Node // boolean used for comma-ok type assertions

	done    ir.Nodes
	clauses []typeClause
}

type typeClause struct {
	hash uint32
	body ir.Nodes
}

func (s *typeSwitch) Add(pos src.XPos, n1 ir.Node, caseVar *ir.Name, jmp ir.Node) {
	typ := n1.Type()
	var body ir.Nodes
	if caseVar != nil {
		l := []ir.Node{
			ir.NewDecl(pos, ir.ODCL, caseVar),
			ir.NewAssignStmt(pos, caseVar, nil),
		}
		typecheck.Stmts(l)
		body.Append(l...)
	} else {
		caseVar = ir.BlankNode.(*ir.Name)
	}

	// cv, ok = iface.(type)
	as := ir.NewAssignListStmt(pos, ir.OAS2, nil, nil)
	as.Lhs = []ir.Node{caseVar, s.okname} // cv, ok =
	switch n1.Op() {
	case ir.OTYPE:
		// Static type assertion (non-generic)
		dot := ir.NewTypeAssertExpr(pos, s.facename, typ) // iface.(type)
		as.Rhs = []ir.Node{dot}
	case ir.ODYNAMICTYPE:
		// Dynamic type assertion (generic)
		dt := n1.(*ir.DynamicType)
		dot := ir.NewDynamicTypeAssertExpr(pos, ir.ODYNAMICDOTTYPE, s.facename, dt.RType)
		dot.ITab = dt.ITab
		dot.SetType(typ)
		dot.SetTypecheck(1)
		as.Rhs = []ir.Node{dot}
	default:
		base.Fatalf("unhandled type case %s", n1.Op())
	}
	appendWalkStmt(&body, as)

	// if ok { goto label }
	nif := ir.NewIfStmt(pos, nil, nil, nil)
	nif.Cond = s.okname
	nif.Body = []ir.Node{jmp}
	body.Append(nif)

	if n1.Op() == ir.OTYPE && !typ.IsInterface() {
		// Defer static, noninterface cases so they can be binary searched by hash.
		s.clauses = append(s.clauses, typeClause{
			hash: types.TypeHash(n1.Type()),
			body: body,
		})
		return
	}

	s.flush()
	s.done.Append(body.Take()...)
}

func (s *typeSwitch) Emit(out *ir.Nodes) {
	s.flush()
	out.Append(s.done.Take()...)
}

func (s *typeSwitch) flush() {
	cc := s.clauses
	s.clauses = nil
	if len(cc) == 0 {
		return
	}

	sort.Slice(cc, func(i, j int) bool { return cc[i].hash < cc[j].hash })

	// Combine adjacent cases with the same hash.
	merged := cc[:1]
	for _, c := range cc[1:] {
		last := &merged[len(merged)-1]
		if last.hash == c.hash {
			last.body.Append(c.body.Take()...)
		} else {
			merged = append(merged, c)
		}
	}
	cc = merged

	// TODO: figure out if we could use a jump table using some low bits of the type hashes.
	binarySearch(len(cc), &s.done,
		func(i int) ir.Node {
			return ir.NewBinaryExpr(base.Pos, ir.OLE, s.hashname, ir.NewInt(int64(cc[i-1].hash)))
		},
		func(i int, nif *ir.IfStmt) {
			// TODO(mdempsky): Omit hash equality check if
			// there's only one type.
			c := cc[i]
			nif.Cond = ir.NewBinaryExpr(base.Pos, ir.OEQ, s.hashname, ir.NewInt(int64(c.hash)))
			nif.Body.Append(c.body.Take()...)
		},
	)
}

// binarySearch constructs a binary search tree for handling n cases,
// and appends it to out. It's used for efficiently implementing
// switch statements.
//
// less(i) should return a boolean expression. If it evaluates true,
// then cases before i will be tested; otherwise, cases i and later.
//
// leaf(i, nif) should setup nif (an OIF node) to test case i. In
// particular, it should set nif.Cond and nif.Body.
func binarySearch(n int, out *ir.Nodes, less func(i int) ir.Node, leaf func(i int, nif *ir.IfStmt)) {
	const binarySearchMin = 4 // minimum number of cases for binary search

	var do func(lo, hi int, out *ir.Nodes)
	do = func(lo, hi int, out *ir.Nodes) {
		n := hi - lo
		if n < binarySearchMin {
			for i := lo; i < hi; i++ {
				nif := ir.NewIfStmt(base.Pos, nil, nil, nil)
				leaf(i, nif)
				base.Pos = base.Pos.WithNotStmt()
				nif.Cond = typecheck.Expr(nif.Cond)
				nif.Cond = typecheck.DefaultLit(nif.Cond, nil)
				out.Append(nif)
				out = &nif.Else
			}
			return
		}

		half := lo + n/2
		nif := ir.NewIfStmt(base.Pos, nil, nil, nil)
		nif.Cond = less(half)
		base.Pos = base.Pos.WithNotStmt()
		nif.Cond = typecheck.Expr(nif.Cond)
		nif.Cond = typecheck.DefaultLit(nif.Cond, nil)
		do(lo, half, &nif.Body)
		do(half, hi, &nif.Else)
		out.Append(nif)
	}

	do(0, n, out)
}

func stringSearch(expr ir.Node, cc []exprClause, out *ir.Nodes) {
	if len(cc) < 4 {
		// Short list, just do brute force equality checks.
		for _, c := range cc {
			nif := ir.NewIfStmt(base.Pos.WithNotStmt(), typecheck.DefaultLit(typecheck.Expr(c.test(expr)), nil), []ir.Node{c.jmp}, nil)
			out.Append(nif)
			out = &nif.Else
		}
		return
	}

	// The strategy here is to find a simple test to divide the set of possible strings
	// that might match expr approximately in half.
	// The test we're going to use is to do an ordered comparison of a single byte
	// of expr to a constant. We will pick the index of that byte and the value we're
	// comparing against to make the split as even as possible.
	//   if expr[3] <= 'd' { ... search strings with expr[3] at 'd' or lower  ... }
	//   else              { ... search strings with expr[3] at 'e' or higher ... }
	//
	// To add complication, we will do the ordered comparison in the signed domain.
	// The reason for this is to prevent CSE from merging the load used for the
	// ordered comparison with the load used for the later equality check.
	//   if expr[3] <= 'd' { ... if expr[0] == 'f' && expr[1] == 'o' && expr[2] == 'o' && expr[3] == 'd' { ... } }
	// If we did both expr[3] loads in the unsigned domain, they would be CSEd, and that
	// would in turn defeat the combining of expr[0]...expr[3] into a single 4-byte load.
	// See issue 48222.
	// By using signed loads for the ordered comparison and unsigned loads for the
	// equality comparison, they don't get CSEd and the equality comparisons will be
	// done using wider loads.

	n := len(ir.StringVal(cc[0].lo)) // Length of the constant strings.
	bestScore := int64(0)            // measure of how good the split is.
	bestIdx := 0                     // split using expr[bestIdx]
	bestByte := int8(0)              // compare expr[bestIdx] against bestByte
	for idx := 0; idx < n; idx++ {
		for b := int8(-128); b < 127; b++ {
			le := 0
			for _, c := range cc {
				s := ir.StringVal(c.lo)
				if int8(s[idx]) <= b {
					le++
				}
			}
			score := int64(le) * int64(len(cc)-le)
			if score > bestScore {
				bestScore = score
				bestIdx = idx
				bestByte = b
			}
		}
	}

	// The split must be at least 1:n-1 because we have at least 2 distinct strings; they
	// have to be different somewhere.
	// TODO: what if the best split is still pretty bad?
	if bestScore == 0 {
		base.Fatalf("unable to split string set")
	}

	// Convert expr to a []int8
	slice := ir.NewConvExpr(base.Pos, ir.OSTR2BYTESTMP, types.NewSlice(types.Types[types.TINT8]), expr)
	slice.SetTypecheck(1) // legacy typechecker doesn't handle this op
	// Load the byte we're splitting on.
	load := ir.NewIndexExpr(base.Pos, slice, ir.NewInt(int64(bestIdx)))
	// Compare with the value we're splitting on.
	cmp := ir.Node(ir.NewBinaryExpr(base.Pos, ir.OLE, load, ir.NewInt(int64(bestByte))))
	cmp = typecheck.DefaultLit(typecheck.Expr(cmp), nil)
	nif := ir.NewIfStmt(base.Pos, cmp, nil, nil)

	var le []exprClause
	var gt []exprClause
	for _, c := range cc {
		s := ir.StringVal(c.lo)
		if int8(s[bestIdx]) <= bestByte {
			le = append(le, c)
		} else {
			gt = append(gt, c)
		}
	}
	stringSearch(expr, le, &nif.Body)
	stringSearch(expr, gt, &nif.Else)
	out.Append(nif)

	// TODO: if expr[bestIdx] has enough different possible values, use a jump table.
}