go/src/cmd/compile/internal/ssa/func.go
David Chase 6b99fb5bea cmd/compile: use sparse algorithm for phis in large program
This adds a sparse method for locating nearest ancestors
in a dominator tree, and checks blocks with more than one
predecessor for differences and inserts phi functions where
there are.

Uses reversed post order to cut number of passes, running
it from first def to last use ("last use" for paramout and
mem is end-of-program; last use for a phi input from a
backedge is the source of the back edge)

Includes a cutover from old algorithm to new to avoid paying
large constant factor for small programs.  This keeps normal
builds running at about the same time, while not running
over-long on large machine-generated inputs.

Add "phase" flags for ssa/build -- ssa/build/stats prints
number of blocks, values (before and after linking references
and inserting phis, so expansion can be measured), and their
product; the product governs the cutover, where a good value
seems to be somewhere between 1 and 5 million.

Among the files compiled by make.bash, this is the shape of
the tail of the distribution for #blocks, #vars, and their
product:

	 #blocks	#vars	    product
 max	6171	28180	173,898,780
99.9%	1641	 6548	 10,401,878
  99%	 463	 1909	    873,721
  95%	 152	  639	     95,235
  90%	  84	  359	     30,021

The old algorithm is indeed usually fastest, for 99%ile
values of usually.

The fix to LookupVarOutgoing
( https://go-review.googlesource.com/#/c/22790/ )
deals with some of the same problems addressed by this CL,
but on at least one bug ( #15537 ) this change is still
a significant help.

With this CL:
/tmp/gopath$ rm -rf pkg bin
/tmp/gopath$ time go get -v -gcflags -memprofile=y.mprof \
   github.com/gogo/protobuf/test/theproto3/combos/...
...
real	4m35.200s
user	13m16.644s
sys	0m36.712s
and pprof reports 3.4GB allocated in one of the larger profiles

With tip:
/tmp/gopath$ rm -rf pkg bin
/tmp/gopath$ time go get -v -gcflags -memprofile=y.mprof \
   github.com/gogo/protobuf/test/theproto3/combos/...
...
real	10m36.569s
user	25m52.286s
sys	4m3.696s
and pprof reports 8.3GB allocated in the same larger profile

With this CL, most of the compilation time on the benchmarked
input is spent in register/stack allocation (cumulative 53%)
and in the sparse lookup algorithm itself (cumulative 20%).

Fixes #15537.

Change-Id: Ia0299dda6a291534d8b08e5f9883216ded677a00
Reviewed-on: https://go-review.googlesource.com/22342
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: David Chase <drchase@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
2016-05-16 21:08:05 +00:00

429 lines
12 KiB
Go

// Copyright 2015 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 ssa
import (
"fmt"
"math"
)
// A Func represents a Go func declaration (or function literal) and
// its body. This package compiles each Func independently.
type Func struct {
Config *Config // architecture information
pass *pass // current pass information (name, options, etc.)
Name string // e.g. bytes·Compare
Type Type // type signature of the function.
StaticData interface{} // associated static data, untouched by the ssa package
Blocks []*Block // unordered set of all basic blocks (note: not indexable by ID)
Entry *Block // the entry basic block
bid idAlloc // block ID allocator
vid idAlloc // value ID allocator
scheduled bool // Values in Blocks are in final order
// when register allocation is done, maps value ids to locations
RegAlloc []Location
// map from LocalSlot to set of Values that we want to store in that slot.
NamedValues map[LocalSlot][]*Value
// Names is a copy of NamedValues.Keys. We keep a separate list
// of keys to make iteration order deterministic.
Names []LocalSlot
freeValues *Value // free Values linked by argstorage[0]. All other fields except ID are 0/nil.
freeBlocks *Block // free Blocks linked by succstorage[0].b. All other fields except ID are 0/nil.
idom []*Block // precomputed immediate dominators
sdom SparseTree // precomputed dominator tree
constants map[int64][]*Value // constants cache, keyed by constant value; users must check value's Op and Type
}
// NumBlocks returns an integer larger than the id of any Block in the Func.
func (f *Func) NumBlocks() int {
return f.bid.num()
}
// NumValues returns an integer larger than the id of any Value in the Func.
func (f *Func) NumValues() int {
return f.vid.num()
}
// newSparseSet returns a sparse set that can store at least up to n integers.
func (f *Func) newSparseSet(n int) *sparseSet {
for i, scr := range f.Config.scrSparse {
if scr != nil && scr.cap() >= n {
f.Config.scrSparse[i] = nil
scr.clear()
return scr
}
}
return newSparseSet(n)
}
// retSparseSet returns a sparse set to the config's cache of sparse sets to be reused by f.newSparseSet.
func (f *Func) retSparseSet(ss *sparseSet) {
for i, scr := range f.Config.scrSparse {
if scr == nil {
f.Config.scrSparse[i] = ss
return
}
}
f.Config.scrSparse = append(f.Config.scrSparse, ss)
}
// newValue allocates a new Value with the given fields and places it at the end of b.Values.
func (f *Func) newValue(op Op, t Type, b *Block, line int32) *Value {
var v *Value
if f.freeValues != nil {
v = f.freeValues
f.freeValues = v.argstorage[0]
v.argstorage[0] = nil
} else {
ID := f.vid.get()
if int(ID) < len(f.Config.values) {
v = &f.Config.values[ID]
} else {
v = &Value{ID: ID}
}
}
v.Op = op
v.Type = t
v.Block = b
v.Line = line
b.Values = append(b.Values, v)
return v
}
// logPassStat writes a string key and int value as a warning in a
// tab-separated format easily handled by spreadsheets or awk.
// file names, lines, and function names are included to provide enough (?)
// context to allow item-by-item comparisons across runs.
// For example:
// awk 'BEGIN {FS="\t"} $3~/TIME/{sum+=$4} END{print "t(ns)=",sum}' t.log
func (f *Func) LogStat(key string, args ...interface{}) {
value := ""
for _, a := range args {
value += fmt.Sprintf("\t%v", a)
}
n := "missing_pass"
if f.pass != nil {
n = f.pass.name
}
f.Config.Warnl(f.Entry.Line, "\t%s\t%s%s\t%s", n, key, value, f.Name)
}
// freeValue frees a value. It must no longer be referenced.
func (f *Func) freeValue(v *Value) {
if v.Block == nil {
f.Fatalf("trying to free an already freed value")
}
if v.Uses != 0 {
f.Fatalf("value %s still has %d uses", v, v.Uses)
}
// Clear everything but ID (which we reuse).
id := v.ID
// Zero argument values might be cached, so remove them there.
nArgs := opcodeTable[v.Op].argLen
if nArgs == 0 {
vv := f.constants[v.AuxInt]
for i, cv := range vv {
if v == cv {
vv[i] = vv[len(vv)-1]
f.constants[v.AuxInt] = vv[0 : len(vv)-1]
break
}
}
}
*v = Value{}
v.ID = id
v.argstorage[0] = f.freeValues
f.freeValues = v
}
// newBlock allocates a new Block of the given kind and places it at the end of f.Blocks.
func (f *Func) NewBlock(kind BlockKind) *Block {
var b *Block
if f.freeBlocks != nil {
b = f.freeBlocks
f.freeBlocks = b.succstorage[0].b
b.succstorage[0].b = nil
} else {
ID := f.bid.get()
if int(ID) < len(f.Config.blocks) {
b = &f.Config.blocks[ID]
} else {
b = &Block{ID: ID}
}
}
b.Kind = kind
b.Func = f
b.Preds = b.predstorage[:0]
b.Succs = b.succstorage[:0]
b.Values = b.valstorage[:0]
f.Blocks = append(f.Blocks, b)
return b
}
func (f *Func) freeBlock(b *Block) {
if b.Func == nil {
f.Fatalf("trying to free an already freed block")
}
// Clear everything but ID (which we reuse).
id := b.ID
*b = Block{}
b.ID = id
b.succstorage[0].b = f.freeBlocks
f.freeBlocks = b
}
// NewValue0 returns a new value in the block with no arguments and zero aux values.
func (b *Block) NewValue0(line int32, op Op, t Type) *Value {
v := b.Func.newValue(op, t, b, line)
v.AuxInt = 0
v.Args = v.argstorage[:0]
return v
}
// NewValue returns a new value in the block with no arguments and an auxint value.
func (b *Block) NewValue0I(line int32, op Op, t Type, auxint int64) *Value {
v := b.Func.newValue(op, t, b, line)
v.AuxInt = auxint
v.Args = v.argstorage[:0]
return v
}
// NewValue returns a new value in the block with no arguments and an aux value.
func (b *Block) NewValue0A(line int32, op Op, t Type, aux interface{}) *Value {
if _, ok := aux.(int64); ok {
// Disallow int64 aux values. They should be in the auxint field instead.
// Maybe we want to allow this at some point, but for now we disallow it
// to prevent errors like using NewValue1A instead of NewValue1I.
b.Fatalf("aux field has int64 type op=%s type=%s aux=%v", op, t, aux)
}
v := b.Func.newValue(op, t, b, line)
v.AuxInt = 0
v.Aux = aux
v.Args = v.argstorage[:0]
return v
}
// NewValue returns a new value in the block with no arguments and both an auxint and aux values.
func (b *Block) NewValue0IA(line int32, op Op, t Type, auxint int64, aux interface{}) *Value {
v := b.Func.newValue(op, t, b, line)
v.AuxInt = auxint
v.Aux = aux
v.Args = v.argstorage[:0]
return v
}
// NewValue1 returns a new value in the block with one argument and zero aux values.
func (b *Block) NewValue1(line int32, op Op, t Type, arg *Value) *Value {
v := b.Func.newValue(op, t, b, line)
v.AuxInt = 0
v.Args = v.argstorage[:1]
v.argstorage[0] = arg
arg.Uses++
return v
}
// NewValue1I returns a new value in the block with one argument and an auxint value.
func (b *Block) NewValue1I(line int32, op Op, t Type, auxint int64, arg *Value) *Value {
v := b.Func.newValue(op, t, b, line)
v.AuxInt = auxint
v.Args = v.argstorage[:1]
v.argstorage[0] = arg
arg.Uses++
return v
}
// NewValue1A returns a new value in the block with one argument and an aux value.
func (b *Block) NewValue1A(line int32, op Op, t Type, aux interface{}, arg *Value) *Value {
v := b.Func.newValue(op, t, b, line)
v.AuxInt = 0
v.Aux = aux
v.Args = v.argstorage[:1]
v.argstorage[0] = arg
arg.Uses++
return v
}
// NewValue1IA returns a new value in the block with one argument and both an auxint and aux values.
func (b *Block) NewValue1IA(line int32, op Op, t Type, auxint int64, aux interface{}, arg *Value) *Value {
v := b.Func.newValue(op, t, b, line)
v.AuxInt = auxint
v.Aux = aux
v.Args = v.argstorage[:1]
v.argstorage[0] = arg
arg.Uses++
return v
}
// NewValue2 returns a new value in the block with two arguments and zero aux values.
func (b *Block) NewValue2(line int32, op Op, t Type, arg0, arg1 *Value) *Value {
v := b.Func.newValue(op, t, b, line)
v.AuxInt = 0
v.Args = v.argstorage[:2]
v.argstorage[0] = arg0
v.argstorage[1] = arg1
arg0.Uses++
arg1.Uses++
return v
}
// NewValue2I returns a new value in the block with two arguments and an auxint value.
func (b *Block) NewValue2I(line int32, op Op, t Type, auxint int64, arg0, arg1 *Value) *Value {
v := b.Func.newValue(op, t, b, line)
v.AuxInt = auxint
v.Args = v.argstorage[:2]
v.argstorage[0] = arg0
v.argstorage[1] = arg1
arg0.Uses++
arg1.Uses++
return v
}
// NewValue3 returns a new value in the block with three arguments and zero aux values.
func (b *Block) NewValue3(line int32, op Op, t Type, arg0, arg1, arg2 *Value) *Value {
v := b.Func.newValue(op, t, b, line)
v.AuxInt = 0
v.Args = v.argstorage[:3]
v.argstorage[0] = arg0
v.argstorage[1] = arg1
v.argstorage[2] = arg2
arg0.Uses++
arg1.Uses++
arg2.Uses++
return v
}
// NewValue3I returns a new value in the block with three arguments and an auxint value.
func (b *Block) NewValue3I(line int32, op Op, t Type, auxint int64, arg0, arg1, arg2 *Value) *Value {
v := b.Func.newValue(op, t, b, line)
v.AuxInt = auxint
v.Args = v.argstorage[:3]
v.argstorage[0] = arg0
v.argstorage[1] = arg1
v.argstorage[2] = arg2
arg0.Uses++
arg1.Uses++
arg2.Uses++
return v
}
// constVal returns a constant value for c.
func (f *Func) constVal(line int32, op Op, t Type, c int64, setAux bool) *Value {
if f.constants == nil {
f.constants = make(map[int64][]*Value)
}
vv := f.constants[c]
for _, v := range vv {
if v.Op == op && v.Type.Compare(t) == CMPeq {
if setAux && v.AuxInt != c {
panic(fmt.Sprintf("cached const %s should have AuxInt of %d", v.LongString(), c))
}
return v
}
}
var v *Value
if setAux {
v = f.Entry.NewValue0I(line, op, t, c)
} else {
v = f.Entry.NewValue0(line, op, t)
}
f.constants[c] = append(vv, v)
return v
}
// These magic auxint values let us easily cache non-numeric constants
// using the same constants map while making collisions unlikely.
// These values are unlikely to occur in regular code and
// are easy to grep for in case of bugs.
const (
constSliceMagic = 1122334455
constInterfaceMagic = 2233445566
constNilMagic = 3344556677
constEmptyStringMagic = 4455667788
)
// ConstInt returns an int constant representing its argument.
func (f *Func) ConstBool(line int32, t Type, c bool) *Value {
i := int64(0)
if c {
i = 1
}
return f.constVal(line, OpConstBool, t, i, true)
}
func (f *Func) ConstInt8(line int32, t Type, c int8) *Value {
return f.constVal(line, OpConst8, t, int64(c), true)
}
func (f *Func) ConstInt16(line int32, t Type, c int16) *Value {
return f.constVal(line, OpConst16, t, int64(c), true)
}
func (f *Func) ConstInt32(line int32, t Type, c int32) *Value {
return f.constVal(line, OpConst32, t, int64(c), true)
}
func (f *Func) ConstInt64(line int32, t Type, c int64) *Value {
return f.constVal(line, OpConst64, t, c, true)
}
func (f *Func) ConstFloat32(line int32, t Type, c float64) *Value {
return f.constVal(line, OpConst32F, t, int64(math.Float64bits(float64(float32(c)))), true)
}
func (f *Func) ConstFloat64(line int32, t Type, c float64) *Value {
return f.constVal(line, OpConst64F, t, int64(math.Float64bits(c)), true)
}
func (f *Func) ConstSlice(line int32, t Type) *Value {
return f.constVal(line, OpConstSlice, t, constSliceMagic, false)
}
func (f *Func) ConstInterface(line int32, t Type) *Value {
return f.constVal(line, OpConstInterface, t, constInterfaceMagic, false)
}
func (f *Func) ConstNil(line int32, t Type) *Value {
return f.constVal(line, OpConstNil, t, constNilMagic, false)
}
func (f *Func) ConstEmptyString(line int32, t Type) *Value {
v := f.constVal(line, OpConstString, t, constEmptyStringMagic, false)
v.Aux = ""
return v
}
func (f *Func) Logf(msg string, args ...interface{}) { f.Config.Logf(msg, args...) }
func (f *Func) Log() bool { return f.Config.Log() }
func (f *Func) Fatalf(msg string, args ...interface{}) { f.Config.Fatalf(f.Entry.Line, msg, args...) }
func (f *Func) Unimplementedf(msg string, args ...interface{}) {
f.Config.Unimplementedf(f.Entry.Line, msg, args...)
}
func (f *Func) Free() {
// Clear values.
n := f.vid.num()
if n > len(f.Config.values) {
n = len(f.Config.values)
}
for i := 1; i < n; i++ {
f.Config.values[i] = Value{}
f.Config.values[i].ID = ID(i)
}
// Clear blocks.
n = f.bid.num()
if n > len(f.Config.blocks) {
n = len(f.Config.blocks)
}
for i := 1; i < n; i++ {
f.Config.blocks[i] = Block{}
f.Config.blocks[i].ID = ID(i)
}
// Unregister from config.
if f.Config.curFunc != f {
f.Fatalf("free of function which isn't the last one allocated")
}
f.Config.curFunc = nil
*f = Func{} // just in case
}