[dev.ssa] cmd/compile: make cse faster

It is one of the slowest compiler phases right now, and we
run two of them.

Instead of using a map to make the initial partition, use a sort.
It is much less memory intensive.

Do a few optimizations to avoid work for size-1 equivalence classes.

Implement -N.

Change-Id: I1d2d85d3771abc918db4dd7cc30b0b2d854b15e1
Reviewed-on: https://go-review.googlesource.com/19024
Reviewed-by: David Chase <drchase@google.com>

10 files changed, 206 insertions, 276 deletions

diff --git a/src/cmd/compile/internal/gc/ssa.go b/src/cmd/compile/internal/gc/ssa.go
index de00fe9651..203de6421c 100644
--- a/src/cmd/compile/internal/gc/ssa.go
+++ b/src/cmd/compile/internal/gc/ssa.go
@@ -121,7 +121,7 @@ func buildssa(fn *Node) *ssa.Func {
 
 	var e ssaExport
 	e.log = printssa
-	s.config = ssa.NewConfig(Thearch.Thestring, &e, Ctxt)
+	s.config = ssa.NewConfig(Thearch.Thestring, &e, Ctxt, Debug['N'] == 0)
 	s.f = s.config.NewFunc()
 	s.f.Name = name
 	s.exitCode = fn.Func.Exit
diff --git a/src/cmd/compile/internal/ssa/compile.go b/src/cmd/compile/internal/ssa/compile.go
index 7a515f898c..048f189ffe 100644
--- a/src/cmd/compile/internal/ssa/compile.go
+++ b/src/cmd/compile/internal/ssa/compile.go
@@ -40,6 +40,9 @@ func Compile(f *Func) {
 	checkFunc(f)
 	const logMemStats = false
 	for _, p := range passes {
+		if !f.Config.optimize && !p.required {
+			continue
+		}
 		phaseName = p.name
 		f.Logf("  pass %s begin\n", p.name)
 		// TODO: capture logging during this pass, add it to the HTML
@@ -75,38 +78,39 @@ func Compile(f *Func) {
 }
 
 type pass struct {
-	name string
-	fn   func(*Func)
+	name     string
+	fn       func(*Func)
+	required bool
 }
 
 // list of passes for the compiler
 var passes = [...]pass{
 	// TODO: combine phielim and copyelim into a single pass?
-	{"early phielim", phielim},
-	{"early copyelim", copyelim},
-	{"early deadcode", deadcode}, // remove generated dead code to avoid doing pointless work during opt
-	{"decompose", decompose},
-	{"opt", opt},
-	{"opt deadcode", deadcode}, // remove any blocks orphaned during opt
-	{"generic cse", cse},
-	{"nilcheckelim", nilcheckelim},
-	{"generic deadcode", deadcode},
-	{"fuse", fuse},
-	{"dse", dse},
-	{"tighten", tighten}, // move values closer to their uses
-	{"lower", lower},
-	{"lowered cse", cse},
-	{"lowered deadcode", deadcode},
-	{"checkLower", checkLower},
-	{"late phielim", phielim},
-	{"late copyelim", copyelim},
-	{"late deadcode", deadcode},
-	{"critical", critical},   // remove critical edges
-	{"layout", layout},       // schedule blocks
-	{"schedule", schedule},   // schedule values
-	{"flagalloc", flagalloc}, // allocate flags register
-	{"regalloc", regalloc},   // allocate int & float registers
-	{"trim", trim},           // remove empty blocks
+	{"early phielim", phielim, false},
+	{"early copyelim", copyelim, false},
+	{"early deadcode", deadcode, false}, // remove generated dead code to avoid doing pointless work during opt
+	{"decompose", decompose, true},
+	{"opt", opt, true},                // TODO: split required rules and optimizing rules
+	{"opt deadcode", deadcode, false}, // remove any blocks orphaned during opt
+	{"generic cse", cse, false},
+	{"nilcheckelim", nilcheckelim, false},
+	{"generic deadcode", deadcode, false},
+	{"fuse", fuse, false},
+	{"dse", dse, false},
+	{"tighten", tighten, false}, // move values closer to their uses
+	{"lower", lower, true},
+	{"lowered cse", cse, false},
+	{"lowered deadcode", deadcode, true},
+	{"checkLower", checkLower, true},
+	{"late phielim", phielim, false},
+	{"late copyelim", copyelim, false},
+	{"late deadcode", deadcode, false},
+	{"critical", critical, true},   // remove critical edges
+	{"layout", layout, true},       // schedule blocks
+	{"schedule", schedule, true},   // schedule values
+	{"flagalloc", flagalloc, true}, // allocate flags register
+	{"regalloc", regalloc, true},   // allocate int & float registers + stack slots
+	{"trim", trim, false},          // remove empty blocks
 }
 
 // Double-check phase ordering constraints.
diff --git a/src/cmd/compile/internal/ssa/config.go b/src/cmd/compile/internal/ssa/config.go
index fb0d886b88..7325873a15 100644
--- a/src/cmd/compile/internal/ssa/config.go
+++ b/src/cmd/compile/internal/ssa/config.go
@@ -15,6 +15,7 @@ type Config struct {
 	fe         Frontend                   // callbacks into compiler frontend
 	HTML       *HTMLWriter                // html writer, for debugging
 	ctxt       *obj.Link                  // Generic arch information
+	optimize   bool                       // Do optimization
 
 	// TODO: more stuff.  Compiler flags of interest, ...
 }
@@ -80,7 +81,7 @@ type GCNode interface {
 }
 
 // NewConfig returns a new configuration object for the given architecture.
-func NewConfig(arch string, fe Frontend, ctxt *obj.Link) *Config {
+func NewConfig(arch string, fe Frontend, ctxt *obj.Link, optimize bool) *Config {
 	c := &Config{arch: arch, fe: fe}
 	switch arch {
 	case "amd64":
@@ -97,6 +98,7 @@ func NewConfig(arch string, fe Frontend, ctxt *obj.Link) *Config {
 		fe.Unimplementedf(0, "arch %s not implemented", arch)
 	}
 	c.ctxt = ctxt
+	c.optimize = optimize
 
 	return c
 }
diff --git a/src/cmd/compile/internal/ssa/cse.go b/src/cmd/compile/internal/ssa/cse.go
index 58c52f23e6..7603e17ecf 100644
--- a/src/cmd/compile/internal/ssa/cse.go
+++ b/src/cmd/compile/internal/ssa/cse.go
@@ -25,56 +25,29 @@ func cse(f *Func) {
 	// It starts with a coarse partition and iteratively refines it
 	// until it reaches a fixed point.
 
-	// Make initial partition based on opcode, type-name, aux, auxint, nargs, phi-block, and the ops of v's first args
-	type key struct {
-		op     Op
-		typ    string
-		aux    interface{}
-		auxint int64
-		nargs  int
-		block  ID // block id for phi vars, -1 otherwise
-		arg0op Op // v.Args[0].Op if len(v.Args) > 0, OpInvalid otherwise
-		arg1op Op // v.Args[1].Op if len(v.Args) > 1, OpInvalid otherwise
-	}
-	m := map[key]eqclass{}
+	// Make initial coarse partitions by using a subset of the conditions above.
+	a := make([]*Value, 0, f.NumValues())
 	for _, b := range f.Blocks {
 		for _, v := range b.Values {
-			bid := ID(-1)
-			if v.Op == OpPhi {
-				bid = b.ID
+			if v.Type.IsMemory() {
+				continue // memory values can never cse
 			}
-			arg0op := OpInvalid
-			if len(v.Args) > 0 {
-				arg0op = v.Args[0].Op
-			}
-			arg1op := OpInvalid
-			if len(v.Args) > 1 {
-				arg1op = v.Args[1].Op
-			}
-
-			// This assumes that floats are stored in AuxInt
-			// instead of Aux. If not, then we need to use the
-			// float bits as part of the key, otherwise since 0.0 == -0.0
-			// this would incorrectly treat 0.0 and -0.0 as identical values
-			k := key{v.Op, v.Type.String(), v.Aux, v.AuxInt, len(v.Args), bid, arg0op, arg1op}
-			m[k] = append(m[k], v)
+			a = append(a, v)
 		}
 	}
-
-	// A partition is a set of disjoint eqclasses.
-	var partition []eqclass
-	for _, v := range m {
-		partition = append(partition, v)
-	}
-	// TODO: Sort partition here for perfect reproducibility?
-	// Sort by what? Partition size?
-	// (Could that improve efficiency by discovering splits earlier?)
+	partition := partitionValues(a)
 
 	// map from value id back to eqclass id
-	valueEqClass := make([]int, f.NumValues())
+	valueEqClass := make([]ID, f.NumValues())
+	for _, b := range f.Blocks {
+		for _, v := range b.Values {
+			// Use negative equivalence class #s for unique values.
+			valueEqClass[v.ID] = -v.ID
+		}
+	}
 	for i, e := range partition {
 		for _, v := range e {
-			valueEqClass[v.ID] = i
+			valueEqClass[v.ID] = ID(i)
 		}
 	}
 
@@ -104,7 +77,7 @@ func cse(f *Func) {
 						// move it to the end and shrink e.
 						e[j], e[len(e)-1] = e[len(e)-1], e[j]
 						e = e[:len(e)-1]
-						valueEqClass[w.ID] = len(partition)
+						valueEqClass[w.ID] = ID(len(partition))
 						changed = true
 						continue eqloop
 					}
@@ -131,7 +104,6 @@ func cse(f *Func) {
 	// if v and w are in the same equivalence class and v dominates w.
 	rewrite := make([]*Value, f.NumValues())
 	for _, e := range partition {
-		sort.Sort(e) // ensure deterministic ordering
 		for len(e) > 1 {
 			// Find a maximal dominant element in e
 			v := e[0]
@@ -197,7 +169,141 @@ func dom(b, c *Block, idom []*Block) bool {
 // final equivalence classes.
 type eqclass []*Value
 
-// Sort an equivalence class by value ID.
-func (e eqclass) Len() int           { return len(e) }
-func (e eqclass) Swap(i, j int)      { e[i], e[j] = e[j], e[i] }
-func (e eqclass) Less(i, j int) bool { return e[i].ID < e[j].ID }
+// partitionValues partitions the values into equivalence classes
+// based on having all the following features match:
+//  - opcode
+//  - type
+//  - auxint
+//  - aux
+//  - nargs
+//  - block # if a phi op
+//  - first two arg's opcodes
+// partitionValues returns a list of equivalence classes, each
+// being a sorted by ID list of *Values.  The eqclass slices are
+// backed by the same storage as the input slice.
+// Equivalence classes of size 1 are ignored.
+func partitionValues(a []*Value) []eqclass {
+	typNames := map[Type]string{}
+	auxIDs := map[interface{}]int32{}
+	sort.Sort(sortvalues{a, typNames, auxIDs})
+
+	var partition []eqclass
+	for len(a) > 0 {
+		v := a[0]
+		j := 1
+		for ; j < len(a); j++ {
+			w := a[j]
+			if v.Op != w.Op ||
+				v.AuxInt != w.AuxInt ||
+				len(v.Args) != len(w.Args) ||
+				v.Op == OpPhi && v.Block != w.Block ||
+				v.Aux != w.Aux ||
+				len(v.Args) >= 1 && v.Args[0].Op != w.Args[0].Op ||
+				len(v.Args) >= 2 && v.Args[1].Op != w.Args[1].Op ||
+				typNames[v.Type] != typNames[w.Type] {
+				break
+			}
+		}
+		if j > 1 {
+			partition = append(partition, a[:j])
+		}
+		a = a[j:]
+	}
+
+	return partition
+}
+
+// Sort values to make the initial partition.
+type sortvalues struct {
+	a        []*Value              // array of values
+	typNames map[Type]string       // type -> type ID map
+	auxIDs   map[interface{}]int32 // aux -> aux ID map
+}
+
+func (sv sortvalues) Len() int      { return len(sv.a) }
+func (sv sortvalues) Swap(i, j int) { sv.a[i], sv.a[j] = sv.a[j], sv.a[i] }
+func (sv sortvalues) Less(i, j int) bool {
+	v := sv.a[i]
+	w := sv.a[j]
+	if v.Op != w.Op {
+		return v.Op < w.Op
+	}
+	if v.AuxInt != w.AuxInt {
+		return v.AuxInt < w.AuxInt
+	}
+	if v.Aux == nil && w.Aux != nil { // cheap aux check - expensive one below.
+		return true
+	}
+	if v.Aux != nil && w.Aux == nil {
+		return false
+	}
+	if len(v.Args) != len(w.Args) {
+		return len(v.Args) < len(w.Args)
+	}
+	if v.Op == OpPhi && v.Block.ID != w.Block.ID {
+		return v.Block.ID < w.Block.ID
+	}
+	if len(v.Args) >= 1 {
+		x := v.Args[0].Op
+		y := w.Args[0].Op
+		if x != y {
+			return x < y
+		}
+		if len(v.Args) >= 2 {
+			x = v.Args[1].Op
+			y = w.Args[1].Op
+			if x != y {
+				return x < y
+			}
+		}
+	}
+
+	// Sort by type.  Types are just interfaces, so we can't compare
+	// them with < directly.  Instead, map types to their names and
+	// sort on that.
+	if v.Type != w.Type {
+		x := sv.typNames[v.Type]
+		if x == "" {
+			x = v.Type.String()
+			sv.typNames[v.Type] = x
+		}
+		y := sv.typNames[w.Type]
+		if y == "" {
+			y = w.Type.String()
+			sv.typNames[w.Type] = y
+		}
+		if x != y {
+			return x < y
+		}
+	}
+
+	// Same deal for aux fields.
+	if v.Aux != w.Aux {
+		x := sv.auxIDs[v.Aux]
+		if x == 0 {
+			x = int32(len(sv.auxIDs)) + 1
+			sv.auxIDs[v.Aux] = x
+		}
+		y := sv.auxIDs[w.Aux]
+		if y == 0 {
+			y = int32(len(sv.auxIDs)) + 1
+			sv.auxIDs[w.Aux] = y
+		}
+		if x != y {
+			return x < y
+		}
+	}
+
+	// TODO(khr): is the above really ok to do?  We're building
+	// the aux->auxID map online as sort is asking about it.  If
+	// sort has some internal randomness, then the numbering might
+	// change from run to run.  That will make the ordering of
+	// partitions random.  It won't break the compiler but may
+	// make it nondeterministic.  We could fix this by computing
+	// the aux->auxID map ahead of time, but the hope is here that
+	// we won't need to compute the mapping for many aux fields
+	// because the values they are in are otherwise unique.
+
+	// Sort by value ID last to keep the sort result deterministic.
+	return v.ID < w.ID
+}
diff --git a/src/cmd/compile/internal/ssa/dom_test.go b/src/cmd/compile/internal/ssa/dom_test.go
index 84e0093799..7174f10e4d 100644
--- a/src/cmd/compile/internal/ssa/dom_test.go
+++ b/src/cmd/compile/internal/ssa/dom_test.go
@@ -160,7 +160,7 @@ func genMaxPredValue(size int) []bloc {
 var domBenchRes []*Block
 
 func benchmarkDominators(b *testing.B, size int, bg blockGen) {
-	c := NewConfig("amd64", DummyFrontend{b}, nil)
+	c := NewConfig("amd64", DummyFrontend{b}, nil, true)
 	fun := Fun(c, "entry", bg(size)...)
 
 	CheckFunc(fun.f)
diff --git a/src/cmd/compile/internal/ssa/export_test.go b/src/cmd/compile/internal/ssa/export_test.go
index badafadd70..962dc52a5f 100644
--- a/src/cmd/compile/internal/ssa/export_test.go
+++ b/src/cmd/compile/internal/ssa/export_test.go
@@ -16,7 +16,7 @@ var Deadcode = deadcode
 
 func testConfig(t *testing.T) *Config {
 	testCtxt := &obj.Link{}
-	return NewConfig("amd64", DummyFrontend{t}, testCtxt)
+	return NewConfig("amd64", DummyFrontend{t}, testCtxt, true)
 }
 
 // DummyFrontend is a test-only frontend.
diff --git a/src/cmd/compile/internal/ssa/nilcheck_test.go b/src/cmd/compile/internal/ssa/nilcheck_test.go
index d4a55c0855..c4aff58d76 100644
--- a/src/cmd/compile/internal/ssa/nilcheck_test.go
+++ b/src/cmd/compile/internal/ssa/nilcheck_test.go
@@ -40,7 +40,7 @@ func benchmarkNilCheckDeep(b *testing.B, depth int) {
 		Bloc("exit", Exit("mem")),
 	)
 
-	c := NewConfig("amd64", DummyFrontend{b}, nil)
+	c := NewConfig("amd64", DummyFrontend{b}, nil, true)
 	fun := Fun(c, "entry", blocs...)
 
 	CheckFunc(fun.f)
@@ -64,7 +64,7 @@ func isNilCheck(b *Block) bool {
 // TestNilcheckSimple verifies that a second repeated nilcheck is removed.
 func TestNilcheckSimple(t *testing.T) {
 	ptrType := &TypeImpl{Size_: 8, Ptr: true, Name: "testptr"} // dummy for testing
-	c := NewConfig("amd64", DummyFrontend{t}, nil)
+	c := NewConfig("amd64", DummyFrontend{t}, nil, true)
 	fun := Fun(c, "entry",
 		Bloc("entry",
 			Valu("mem", OpInitMem, TypeMem, 0, ".mem"),
@@ -101,7 +101,7 @@ func TestNilcheckSimple(t *testing.T) {
 // on the order of the dominees.
 func TestNilcheckDomOrder(t *testing.T) {
 	ptrType := &TypeImpl{Size_: 8, Ptr: true, Name: "testptr"} // dummy for testing
-	c := NewConfig("amd64", DummyFrontend{t}, nil)
+	c := NewConfig("amd64", DummyFrontend{t}, nil, true)
 	fun := Fun(c, "entry",
 		Bloc("entry",
 			Valu("mem", OpInitMem, TypeMem, 0, ".mem"),
@@ -137,7 +137,7 @@ func TestNilcheckDomOrder(t *testing.T) {
 // TestNilcheckAddr verifies that nilchecks of OpAddr constructed values are removed.
 func TestNilcheckAddr(t *testing.T) {
 	ptrType := &TypeImpl{Size_: 8, Ptr: true, Name: "testptr"} // dummy for testing
-	c := NewConfig("amd64", DummyFrontend{t}, nil)
+	c := NewConfig("amd64", DummyFrontend{t}, nil, true)
 	fun := Fun(c, "entry",
 		Bloc("entry",
 			Valu("mem", OpInitMem, TypeMem, 0, ".mem"),
@@ -170,7 +170,7 @@ func TestNilcheckAddr(t *testing.T) {
 // TestNilcheckAddPtr verifies that nilchecks of OpAddPtr constructed values are removed.
 func TestNilcheckAddPtr(t *testing.T) {
 	ptrType := &TypeImpl{Size_: 8, Ptr: true, Name: "testptr"} // dummy for testing
-	c := NewConfig("amd64", DummyFrontend{t}, nil)
+	c := NewConfig("amd64", DummyFrontend{t}, nil, true)
 	fun := Fun(c, "entry",
 		Bloc("entry",
 			Valu("mem", OpInitMem, TypeMem, 0, ".mem"),
@@ -204,7 +204,7 @@ func TestNilcheckAddPtr(t *testing.T) {
 // non-nil are removed.
 func TestNilcheckPhi(t *testing.T) {
 	ptrType := &TypeImpl{Size_: 8, Ptr: true, Name: "testptr"} // dummy for testing
-	c := NewConfig("amd64", DummyFrontend{t}, nil)
+	c := NewConfig("amd64", DummyFrontend{t}, nil, true)
 	fun := Fun(c, "entry",
 		Bloc("entry",
 			Valu("mem", OpInitMem, TypeMem, 0, ".mem"),
@@ -248,7 +248,7 @@ func TestNilcheckPhi(t *testing.T) {
 // are removed, but checks of different pointers are not.
 func TestNilcheckKeepRemove(t *testing.T) {
 	ptrType := &TypeImpl{Size_: 8, Ptr: true, Name: "testptr"} // dummy for testing
-	c := NewConfig("amd64", DummyFrontend{t}, nil)
+	c := NewConfig("amd64", DummyFrontend{t}, nil, true)
 	fun := Fun(c, "entry",
 		Bloc("entry",
 			Valu("mem", OpInitMem, TypeMem, 0, ".mem"),
@@ -296,7 +296,7 @@ func TestNilcheckKeepRemove(t *testing.T) {
 // block are *not* removed.
 func TestNilcheckInFalseBranch(t *testing.T) {
 	ptrType := &TypeImpl{Size_: 8, Ptr: true, Name: "testptr"} // dummy for testing
-	c := NewConfig("amd64", DummyFrontend{t}, nil)
+	c := NewConfig("amd64", DummyFrontend{t}, nil, true)
 	fun := Fun(c, "entry",
 		Bloc("entry",
 			Valu("mem", OpInitMem, TypeMem, 0, ".mem"),
@@ -347,7 +347,7 @@ func TestNilcheckInFalseBranch(t *testing.T) {
 // wil remove the generated nil check.
 func TestNilcheckUser(t *testing.T) {
 	ptrType := &TypeImpl{Size_: 8, Ptr: true, Name: "testptr"} // dummy for testing
-	c := NewConfig("amd64", DummyFrontend{t}, nil)
+	c := NewConfig("amd64", DummyFrontend{t}, nil, true)
 	fun := Fun(c, "entry",
 		Bloc("entry",
 			Valu("mem", OpInitMem, TypeMem, 0, ".mem"),
@@ -386,7 +386,7 @@ func TestNilcheckUser(t *testing.T) {
 // TestNilcheckBug reproduces a bug in nilcheckelim found by compiling math/big
 func TestNilcheckBug(t *testing.T) {
 	ptrType := &TypeImpl{Size_: 8, Ptr: true, Name: "testptr"} // dummy for testing
-	c := NewConfig("amd64", DummyFrontend{t}, nil)
+	c := NewConfig("amd64", DummyFrontend{t}, nil, true)
 	fun := Fun(c, "entry",
 		Bloc("entry",
 			Valu("mem", OpInitMem, TypeMem, 0, ".mem"),
diff --git a/src/cmd/compile/internal/ssa/regalloc.go b/src/cmd/compile/internal/ssa/regalloc.go
index 7cbd30311f..9238999074 100644
--- a/src/cmd/compile/internal/ssa/regalloc.go
+++ b/src/cmd/compile/internal/ssa/regalloc.go
@@ -316,6 +316,12 @@ func (s *regAllocState) assignReg(r register, v *Value, c *Value) {
 		fmt.Printf("assignReg %s %s/%s\n", registers[r].Name(), v, c)
 	}
 	if s.regs[r].v != nil {
+		if v.Op == OpSB && !v.Block.Func.Config.optimize {
+			// Rewrite rules may introduce multiple OpSB, and with
+			// -N they don't get CSEd.  Ignore the extra assignments.
+			s.f.setHome(c, &registers[r])
+			return
+		}
 		s.f.Fatalf("tried to assign register %d to %s/%s but it is already used by %s", r, v, c, s.regs[r].v)
 	}
 
diff --git a/test/nilcheck.go b/test/nilcheck.go
index 173fcb33a6..ab28b33d41 100644
--- a/test/nilcheck.go
+++ b/test/nilcheck.go
@@ -1,4 +1,3 @@
-// +build !amd64
 // errorcheck -0 -N -d=nil
 
 // Copyright 2013 The Go Authors.  All rights reserved.
diff --git a/test/nilcheck_ssa.go b/test/nilcheck_ssa.go
deleted file mode 100644
index a20cfd8ae6..0000000000
--- a/test/nilcheck_ssa.go
+++ /dev/null
@@ -1,187 +0,0 @@
-// +build amd64
-// errorcheck -0 -N -d=nil
-
-// Copyright 2013 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.
-
-// Test that nil checks are inserted.
-// Optimization is disabled, so redundant checks are not removed.
-
-package p
-
-type Struct struct {
-	X int
-	Y float64
-}
-
-type BigStruct struct {
-	X int
-	Y float64
-	A [1 << 20]int
-	Z string
-}
-
-type Empty struct {
-}
-
-type Empty1 struct {
-	Empty
-}
-
-var (
-	intp       *int
-	arrayp     *[10]int
-	array0p    *[0]int
-	bigarrayp  *[1 << 26]int
-	structp    *Struct
-	bigstructp *BigStruct
-	emptyp     *Empty
-	empty1p    *Empty1
-)
-
-func f1() {
-	_ = *intp    // ERROR "nil check"
-	_ = *arrayp  // ERROR "nil check"
-	_ = *array0p // ERROR "nil check"
-	_ = *array0p // ERROR "nil check"
-	_ = *intp    // ERROR "nil check"
-	_ = *arrayp  // ERROR "nil check"
-	_ = *structp // ERROR "nil check"
-	_ = *emptyp  // ERROR "nil check"
-	_ = *arrayp  // ERROR "nil check"
-}
-
-func f2() {
-	var (
-		intp       *int
-		arrayp     *[10]int
-		array0p    *[0]int
-		bigarrayp  *[1 << 20]int
-		structp    *Struct
-		bigstructp *BigStruct
-		emptyp     *Empty
-		empty1p    *Empty1
-	)
-
-	_ = *intp       // ERROR "nil check"
-	_ = *arrayp     // ERROR "nil check"
-	_ = *array0p    // ERROR "nil check"
-	_ = *array0p    // ERROR "removed nil check"
-	_ = *intp       // ERROR "removed nil check"
-	_ = *arrayp     // ERROR "removed nil check"
-	_ = *structp    // ERROR "nil check"
-	_ = *emptyp     // ERROR "nil check"
-	_ = *arrayp     // ERROR "removed nil check"
-	_ = *bigarrayp  // ERROR "nil check"
-	_ = *bigstructp // ERROR "nil check"
-	_ = *empty1p    // ERROR "nil check"
-}
-
-func fx10k() *[10000]int
-
-var b bool
-
-func f3(x *[10000]int) {
-	// Using a huge type and huge offsets so the compiler
-	// does not expect the memory hardware to fault.
-	_ = x[9999] // ERROR "nil check"
-
-	for {
-		if x[9999] != 0 { // ERROR "removed nil check"
-			break
-		}
-	}
-
-	x = fx10k()
-	_ = x[9999] // ERROR "nil check"
-	if b {
-		_ = x[9999] // ERROR "removed nil check"
-	} else {
-		_ = x[9999] // ERROR "removed nil check"
-	}
-	_ = x[9999] // ERROR "removed nil check"
-
-	x = fx10k()
-	if b {
-		_ = x[9999] // ERROR "nil check"
-	} else {
-		_ = x[9999] // ERROR "nil check"
-	}
-	_ = x[9999] // ERROR "nil check"
-
-	fx10k()
-	// SSA nilcheck removal works across calls.
-	_ = x[9999] // ERROR "removed nil check"
-}
-
-func f3a() {
-	x := fx10k()
-	y := fx10k()
-	z := fx10k()
-	_ = &x[9] // ERROR "nil check"
-	y = z
-	_ = &x[9] // ERROR "removed nil check"
-	x = y
-	_ = &x[9] // ERROR "nil check"
-}
-
-func f3b() {
-	x := fx10k()
-	y := fx10k()
-	_ = &x[9] // ERROR "nil check"
-	y = x
-	_ = &x[9] // ERROR "removed nil check"
-	x = y
-	_ = &x[9] // ERROR "removed nil check"
-}
-
-func fx10() *[10]int
-
-func f4(x *[10]int) {
-	// Most of these have no checks because a real memory reference follows,
-	// and the offset is small enough that if x is nil, the address will still be
-	// in the first unmapped page of memory.
-
-	_ = x[9] // ERROR "nil check"
-
-	for {
-		if x[9] != 0 { // ERROR "removed nil check"
-			break
-		}
-	}
-
-	x = fx10()
-	_ = x[9] // ERROR "nil check"
-	if b {
-		_ = x[9] // ERROR "removed nil check"
-	} else {
-		_ = x[9] // ERROR "removed nil check"
-	}
-	_ = x[9] // ERROR "removed nil check"
-
-	x = fx10()
-	if b {
-		_ = x[9] // ERROR "nil check"
-	} else {
-		_ = &x[9] // ERROR "nil check"
-	}
-	_ = x[9] // ERROR "nil check"
-
-	fx10()
-	_ = x[9] // ERROR "removed nil check"
-
-	x = fx10()
-	y := fx10()
-	_ = &x[9] // ERROR "nil check"
-	y = x
-	_ = &x[9] // ERROR "removed nil check"
-	x = y
-	_ = &x[9] // ERROR "removed nil check"
-}
-
-func f5(m map[string]struct{}) bool {
-	// Existence-only map lookups should not generate a nil check
-	_, ok := m[""]
-	return ok
-}