1
0
mirror of https://github.com/golang/go synced 2024-10-01 13:28:37 -06:00
go/ssa/builder.go

2519 lines
67 KiB
Go
Raw Normal View History

// 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.
package ssa
// This file implements the BUILD phase of SSA construction.
//
// SSA construction has two phases, CREATE and BUILD. In the CREATE phase
// (create.go), all packages are constructed and type-checked and
// definitions of all package members are created, method-sets are
// computed, and wrapper methods are synthesized. The create phase
go.tools/ssa: fix computation of set of types requiring method sets. Motivation: Previously, we assumed that the set of types for which a complete method set (containing all synthesized wrapper functions) is required at runtime was the set of types used as operands to some *ssa.MakeInterface instruction. In fact, this is an underapproximation because types can be derived from other ones via reflection, and some of these may need methods. The reflect.Type API allows *T to be derived from T, and these may have different method sets. Reflection also allows almost any subcomponent of a type to be accessed (with one exception: given T, defined 'type T struct{S}', you can reach S but not struct{S}). As a result, the pointer analysis was unable to generate all necessary constraints before running the solver, causing a crash when reflection derives types whose methods are unavailable. (A similar problem would afflict an ahead-of-time compiler based on ssa. The ssa/interp interpreter was immune only because it does not require all wrapper methods to be created before execution begins.) Description: This change causes the SSA builder to record, for each package, the set of all types with non-empty method sets that are referenced within that package. This set is accessed via Packages.TypesWithMethodSets(). Program.TypesWithMethodSets() returns its union across all packages. The set of references that matter are: - types of operands to some MakeInterface instruction (as before) - types of all exported package members - all subcomponents of the above, recursively. This is a conservative approximation to the set of types whose methods may be called dynamically. We define the owning package of a type as follows: - the owner of a named type is the package in which it is defined; - the owner of a pointer-to-named type is the owner of that named type; - the owner of all other types is nil. A package must include the method sets for all types that it owns, and all subcomponents of that type that are not owned by another package, recursively. Types with an owner appear in exactly one package; types with no owner (such as struct{T}) may appear within multiple packages. (A typical Go compiler would emit multiple copies of these methods as weak symbols; a typical linker would eliminate duplicates.) Also: - go/types/typemap: implement hash function for *Tuple. - pointer: generate nodes/constraints for all of ssa.Program.TypesWithMethodSets(). Add rtti.go regression test. - Add API test of Package.TypesWithMethodSets(). - Set Function.Pkg to nil (again) for wrapper functions, since these may be shared by many packages. - Remove a redundant logging statement. - Document that ssa CREATE phase is in fact sequential. Fixes golang/go#6605 R=gri CC=golang-dev https://golang.org/cl/14920056
2013-10-23 15:07:52 -06:00
// occurs sequentially (order is unimportant) as the client calls
// Program.CreatePackage.
//
// In the BUILD phase (builder.go), the builder traverses the AST of
// each Go source function and generates SSA instructions for the
// function body.
// Within each package, building proceeds in a topological order over
// the intra-package symbol reference graph, whose roots are the set
// of package-level declarations in lexical order. The BUILD phases
// for distinct packages are independent and are executed in parallel.
//
// The builder's and Program's indices (maps) are populated and
// mutated during the CREATE phase, but during the BUILD phase they
go.tools/ssa: (another) major refactoring of method-set logic. We now use LookupFieldOrMethod for all SelectorExprs, and simplify the logic to discriminate the various cases. We inline static calls to promoted/indirected functions, dramatically reducing the number of functions created. More tests are needed, but I'd like to submit this as-is. In this CL, we: - rely less on Id strings. Internally we now use *types.Method (and its components) almost everywhere. - stop thinking of types.Methods as objects. They don't have stable identities. (Hopefully they will become plain-old structs soon.) - eliminate receiver indirection wrappers: indirection and promotion are handled together by makeWrapper. - Handle the interactions of promotion, indirection and abstract methods much more cleanly. - support receiver-bound interface method closures. - break up builder.selectField so we can re-use parts (emitFieldSelection). - add importer.PackageInfo.classifySelector utility. - delete interfaceMethodIndex() - delete namedTypeMethodIndex() - delete isSuperInterface() (replaced by types.IsAssignable) - call memberFromObject on each declared concrete method's *types.Func, not on every Method frem each method set, in the CREATE phase for packages loaded by gcimporter. go/types: - document Func, Signature.Recv() better. - use fmt in {Package,Label}.String - reimplement Func.String to be prettier and to include method receivers. API changes: - Function.method now holds the types.Method (soon to be not-an-object) for synthetic wrappers. - CallCommon.Method now contains an abstract (interface) method object; was an abstract method index. - CallCommon.MethodId() gone. - Program.LookupMethod now takes a *Method not an Id string. R=gri CC=golang-dev https://golang.org/cl/11674043
2013-07-26 09:22:34 -06:00
// remain constant. The sole exception is Prog.methodSets and its
// related maps, which are protected by a dedicated mutex.
import (
"fmt"
"go/ast"
"go/token"
"os"
"sync"
"sync/atomic"
"code.google.com/p/go.tools/go/exact"
"code.google.com/p/go.tools/go/types"
)
type opaqueType struct {
types.Type
name string
}
func (t *opaqueType) String() string { return t.name }
var (
varOk = types.NewVar(token.NoPos, nil, "ok", tBool)
varIndex = types.NewVar(token.NoPos, nil, "index", tInt)
// Type constants.
tBool = types.Typ[types.Bool]
tByte = types.Typ[types.Byte]
tInt = types.Typ[types.Int]
tInvalid = types.Typ[types.Invalid]
tUntypedNil = types.Typ[types.UntypedNil]
tRangeIter = &opaqueType{nil, "iter"} // the type of all "range" iterators
tEface = new(types.Interface)
// SSA Value constants.
vZero = intConst(0)
vOne = intConst(1)
vTrue = NewConst(exact.MakeBool(true), tBool)
vFalse = NewConst(exact.MakeBool(false), tBool)
)
// builder holds state associated with the package currently being built.
// Its methods contain all the logic for AST-to-SSA conversion.
type builder struct {
nTo1Vars map[*ast.ValueSpec]bool // set of n:1 ValueSpecs already built
}
// lookup returns the package-level *Function or *Global for the named
// object obj, building it if necessary.
//
// Intra-package references are edges in the initialization dependency
// graph. If the result v is a Function or Global belonging to
// 'from', the package on whose behalf this lookup occurs, then lookup
// emits initialization code into from.init if not already done.
//
func (b *builder) lookup(from *Package, obj types.Object) Value {
go.tools/ssa: add debug information for all ast.Idents. This CL adds three new functions to determine the SSA Value for a given syntactic var, func or const object: Program.{Const,Func,Var}Value. Since constants and functions are immutable, the first two only need a types.Object; but each distinct reference to a var may return a distinct Value, so the third requires an ast.Ident parameter too. Debug information for local vars is encoded in the instruction stream in the form of DebugRef instructions, which are a no-op but relate their operand to a particular ident in the AST. The beauty of this approach is that it naturally stays consistent during optimisation passes (e.g. lifting) without additional bookkeeping. DebugRef instructions are only generated if the DebugMode builder flag is set; I plan to make the policy more fine- grained (per function). DebugRef instructions are inserted for: - expr(Ident) for rvalue idents - address.store() for idents that update an lvalue - address.address() for idents that take address of lvalue (this new method replaces all uses of lval.(address).addr) - expr() for all constant expressions - local ValueSpecs with implicit zero initialization (no RHS) (this case doesn't call store() or address()) To ensure we don't forget to emit debug info for uses of Idents, we must use the lvalue mechanism consistently. (Previously, many simple cases had effectively inlined these functions.) Similarly setCallFunc no longer inlines expr(Ident). Also: - Program.Value() has been inlined & specialized. - Program.Package() has moved nearer the new lookup functions. - refactoring: funcSyntax has lost paramFields, resultFields; gained funcType, which provides access to both. - add package-level constants to Package.values map. - opt: don't call localValueSpec for constants. (The resulting code is always optimised away.) There are a number of comments asking whether Literals should have positions. Will address in a follow-up. Added tests of all interesting cases. R=gri CC=golang-dev https://golang.org/cl/11259044
2013-07-15 11:56:46 -06:00
v := from.Prog.packages[obj.Pkg()].values[obj]
switch v := v.(type) {
case *Function:
if from == v.Pkg {
b.buildFunction(v)
}
case *Global:
if from == v.Pkg {
b.buildGlobal(v, obj)
}
}
return v
}
// cond emits to fn code to evaluate boolean condition e and jump
// to t or f depending on its value, performing various simplifications.
//
// Postcondition: fn.currentBlock is nil.
//
func (b *builder) cond(fn *Function, e ast.Expr, t, f *BasicBlock) {
switch e := e.(type) {
case *ast.ParenExpr:
b.cond(fn, e.X, t, f)
return
case *ast.BinaryExpr:
switch e.Op {
case token.LAND:
ltrue := fn.newBasicBlock("cond.true")
b.cond(fn, e.X, ltrue, f)
fn.currentBlock = ltrue
b.cond(fn, e.Y, t, f)
return
case token.LOR:
lfalse := fn.newBasicBlock("cond.false")
b.cond(fn, e.X, t, lfalse)
fn.currentBlock = lfalse
b.cond(fn, e.Y, t, f)
return
}
case *ast.UnaryExpr:
if e.Op == token.NOT {
b.cond(fn, e.X, f, t)
return
}
}
switch cond := b.expr(fn, e).(type) {
case *Const:
// Dispatch constant conditions statically.
if exact.BoolVal(cond.Value) {
emitJump(fn, t)
} else {
emitJump(fn, f)
}
default:
emitIf(fn, cond, t, f)
}
}
// logicalBinop emits code to fn to evaluate e, a &&- or
// ||-expression whose reified boolean value is wanted.
// The value is returned.
//
func (b *builder) logicalBinop(fn *Function, e *ast.BinaryExpr) Value {
rhs := fn.newBasicBlock("binop.rhs")
done := fn.newBasicBlock("binop.done")
// T(e) = T(e.X) = T(e.Y) after untyped constants have been
// eliminated.
go.tools/ssa: implement correct control flow for recovered panic. A function such as this: func one() (x int) { defer func() { recover() }() x = 1 panic("return") } that combines named return parameters (NRPs) with deferred calls that call recover, may return non-zero values despite the fact it doesn't even contain a return statement. (!) This requires a change to the SSA API: all functions' control-flow graphs now have a second entry point, called Recover, which is the block at which control flow resumes after a recovered panic. The Recover block simply loads the NRPs and returns them. As an optimization, most functions don't need a Recover block, so it is omitted. In fact it is only needed for functions that have NRPs and defer a call to another function that _may_ call recover. Dataflow analysis of SSA now requires extra work, since every may-panic instruction has an implicit control-flow edge to the Recover block. The only dataflow analysis so far implemented is SSA renaming, for which we make the following simplifying assumption: the Recover block only loads the NRPs and returns. This means we don't really need to analyze it, we can just skip the "lifting" of such NRPs. We also special-case the Recover block in the dominance computation. Rejected alternative approaches: - Specifying a Recover block for every defer instruction (like a traditional exception handler). This seemed like excessive generality, since Go programs only need the same degenerate form of Recover block. - Adding an instruction to set the Recover block immediately after the named return values are set up, so that dominance can be computed without special-casing. This didn't seem worth the effort. Interpreter: - This CL completely reimplements the panic/recover/ defer logic in the interpreter. It's clearer and simpler and closer to the model in the spec. - Some runtime panic messages have been changed to be closer to gc's, since tests depend on it. - The interpreter now requires that the runtime.runtimeError type be part of the SSA program. This requires that clients import this package prior to invoking the interpreter. This in turn requires (Importer).ImportPackage(path string), which this CL adds. - All $GOROOT/test/recover{,1,2,3}.go tests are now passing. NB, the bug described in coverage.go (defer/recover in a concatenated init function) remains. Will be fixed in a follow-up. Fixes golang/go#6381 R=gri CC=crawshaw, golang-dev https://golang.org/cl/13844043
2013-10-14 13:38:56 -06:00
// TODO(adonovan): not true; MyBool==MyBool yields UntypedBool.
t := fn.Pkg.typeOf(e)
var short Value // value of the short-circuit path
switch e.Op {
case token.LAND:
b.cond(fn, e.X, rhs, done)
short = NewConst(exact.MakeBool(false), t)
case token.LOR:
b.cond(fn, e.X, done, rhs)
short = NewConst(exact.MakeBool(true), t)
}
// Is rhs unreachable?
if rhs.Preds == nil {
// Simplify false&&y to false, true||y to true.
fn.currentBlock = done
return short
}
// Is done unreachable?
if done.Preds == nil {
// Simplify true&&y (or false||y) to y.
fn.currentBlock = rhs
return b.expr(fn, e.Y)
}
// All edges from e.X to done carry the short-circuit value.
var edges []Value
for _ = range done.Preds {
edges = append(edges, short)
}
// The edge from e.Y to done carries the value of e.Y.
fn.currentBlock = rhs
edges = append(edges, b.expr(fn, e.Y))
emitJump(fn, done)
fn.currentBlock = done
phi := &Phi{Edges: edges, Comment: e.Op.String()}
phi.pos = e.OpPos
phi.typ = t
return done.emit(phi)
}
// exprN lowers a multi-result expression e to SSA form, emitting code
// to fn and returning a single Value whose type is a *types.Tuple.
// The caller must access the components via Extract.
//
// Multi-result expressions include CallExprs in a multi-value
// assignment or return statement, and "value,ok" uses of
// TypeAssertExpr, IndexExpr (when X is a map), and UnaryExpr (when Op
// is token.ARROW).
//
func (b *builder) exprN(fn *Function, e ast.Expr) Value {
var typ types.Type
var tuple Value
switch e := e.(type) {
case *ast.ParenExpr:
return b.exprN(fn, e.X)
case *ast.CallExpr:
// Currently, no built-in function nor type conversion
// has multiple results, so we can avoid some of the
// cases for single-valued CallExpr.
var c Call
b.setCall(fn, e, &c.Call)
c.typ = fn.Pkg.typeOf(e)
return fn.emit(&c)
case *ast.IndexExpr:
mapt := fn.Pkg.typeOf(e.X).Underlying().(*types.Map)
typ = mapt.Elem()
lookup := &Lookup{
X: b.expr(fn, e.X),
Index: emitConv(fn, b.expr(fn, e.Index), mapt.Key()),
CommaOk: true,
}
lookup.setPos(e.Lbrack)
tuple = fn.emit(lookup)
case *ast.TypeAssertExpr:
t := fn.Pkg.typeOf(e).(*types.Tuple).At(0).Type()
return emitTypeTest(fn, b.expr(fn, e.X), t, e.Lparen)
case *ast.UnaryExpr: // must be receive <-
typ = fn.Pkg.typeOf(e.X).Underlying().(*types.Chan).Elem()
unop := &UnOp{
Op: token.ARROW,
X: b.expr(fn, e.X),
CommaOk: true,
}
unop.setPos(e.OpPos)
tuple = fn.emit(unop)
default:
panic(fmt.Sprintf("unexpected exprN: %T", e))
}
// The typechecker sets the type of the expression to just the
// asserted type in the "value, ok" form, not to *types.Tuple
// (though it includes the valueOk operand in its error messages).
tuple.(interface {
setType(types.Type)
}).setType(types.NewTuple(
types.NewVar(token.NoPos, nil, "value", typ),
varOk,
))
return tuple
}
// builtin emits to fn SSA instructions to implement a call to the
// built-in function obj with the specified arguments
// and return type. It returns the value defined by the result.
//
// The result is nil if no special handling was required; in this case
// the caller should treat this like an ordinary library function
// call.
//
func (b *builder) builtin(fn *Function, obj *types.Builtin, args []ast.Expr, typ types.Type, pos token.Pos) Value {
switch obj.Name() {
case "make":
switch typ.Underlying().(type) {
case *types.Slice:
n := b.expr(fn, args[1])
m := n
if len(args) == 3 {
m = b.expr(fn, args[2])
}
v := &MakeSlice{
Len: n,
Cap: m,
}
v.setPos(pos)
v.setType(typ)
return fn.emit(v)
case *types.Map:
var res Value
if len(args) == 2 {
res = b.expr(fn, args[1])
}
v := &MakeMap{Reserve: res}
v.setPos(pos)
v.setType(typ)
return fn.emit(v)
case *types.Chan:
var sz Value = vZero
if len(args) == 2 {
sz = b.expr(fn, args[1])
}
v := &MakeChan{Size: sz}
v.setPos(pos)
v.setType(typ)
return fn.emit(v)
}
case "new":
alloc := emitNew(fn, deref(typ), pos)
alloc.Comment = "new"
return alloc
case "len", "cap":
// Special case: len or cap of an array or *array is
// based on the type, not the value which may be nil.
// We must still evaluate the value, though. (If it
// was side-effect free, the whole call would have
// been constant-folded.)
t := deref(fn.Pkg.typeOf(args[0])).Underlying()
if at, ok := t.(*types.Array); ok {
b.expr(fn, args[0]) // for effects only
return intConst(at.Len())
}
// Otherwise treat as normal.
case "panic":
fn.emit(&Panic{
X: emitConv(fn, b.expr(fn, args[0]), tEface),
pos: pos,
})
fn.currentBlock = fn.newBasicBlock("unreachable")
return vFalse // any non-nil Value will do
}
return nil // treat all others as a regular function call
}
// addr lowers a single-result addressable expression e to SSA form,
// emitting code to fn and returning the location (an lvalue) defined
// by the expression.
//
// If escaping is true, addr marks the base variable of the
// addressable expression e as being a potentially escaping pointer
// value. For example, in this code:
//
// a := A{
// b: [1]B{B{c: 1}}
// }
// return &a.b[0].c
//
// the application of & causes a.b[0].c to have its address taken,
// which means that ultimately the local variable a must be
// heap-allocated. This is a simple but very conservative escape
// analysis.
//
// Operations forming potentially escaping pointers include:
// - &x, including when implicit in method call or composite literals.
// - a[:] iff a is an array (not *array)
// - references to variables in lexically enclosing functions.
//
func (b *builder) addr(fn *Function, e ast.Expr, escaping bool) lvalue {
switch e := e.(type) {
case *ast.Ident:
go.tools/ssa: add debug information for all ast.Idents. This CL adds three new functions to determine the SSA Value for a given syntactic var, func or const object: Program.{Const,Func,Var}Value. Since constants and functions are immutable, the first two only need a types.Object; but each distinct reference to a var may return a distinct Value, so the third requires an ast.Ident parameter too. Debug information for local vars is encoded in the instruction stream in the form of DebugRef instructions, which are a no-op but relate their operand to a particular ident in the AST. The beauty of this approach is that it naturally stays consistent during optimisation passes (e.g. lifting) without additional bookkeeping. DebugRef instructions are only generated if the DebugMode builder flag is set; I plan to make the policy more fine- grained (per function). DebugRef instructions are inserted for: - expr(Ident) for rvalue idents - address.store() for idents that update an lvalue - address.address() for idents that take address of lvalue (this new method replaces all uses of lval.(address).addr) - expr() for all constant expressions - local ValueSpecs with implicit zero initialization (no RHS) (this case doesn't call store() or address()) To ensure we don't forget to emit debug info for uses of Idents, we must use the lvalue mechanism consistently. (Previously, many simple cases had effectively inlined these functions.) Similarly setCallFunc no longer inlines expr(Ident). Also: - Program.Value() has been inlined & specialized. - Program.Package() has moved nearer the new lookup functions. - refactoring: funcSyntax has lost paramFields, resultFields; gained funcType, which provides access to both. - add package-level constants to Package.values map. - opt: don't call localValueSpec for constants. (The resulting code is always optimised away.) There are a number of comments asking whether Literals should have positions. Will address in a follow-up. Added tests of all interesting cases. R=gri CC=golang-dev https://golang.org/cl/11259044
2013-07-15 11:56:46 -06:00
if isBlankIdent(e) {
return blank{}
}
obj := fn.Pkg.objectOf(e)
v := b.lookup(fn.Pkg, obj) // var (address)
if v == nil {
v = fn.lookup(obj, escaping)
}
return &address{addr: v, expr: e}
case *ast.CompositeLit:
t := deref(fn.Pkg.typeOf(e))
var v *Alloc
if escaping {
v = emitNew(fn, t, e.Lbrace)
} else {
v = fn.addLocal(t, e.Lbrace)
}
v.Comment = "complit"
b.compLit(fn, v, e, t) // initialize in place
return &address{addr: v, expr: e}
case *ast.ParenExpr:
return b.addr(fn, e.X, escaping)
case *ast.SelectorExpr:
switch sel := fn.Pkg.info.Selections[e]; sel.Kind() {
case types.PackageObj:
obj := sel.Obj()
if v := b.lookup(fn.Pkg, obj); v != nil {
return &address{addr: v, expr: e}
}
panic("undefined package-qualified name: " + obj.Name())
go.tools/ssa: (another) major refactoring of method-set logic. We now use LookupFieldOrMethod for all SelectorExprs, and simplify the logic to discriminate the various cases. We inline static calls to promoted/indirected functions, dramatically reducing the number of functions created. More tests are needed, but I'd like to submit this as-is. In this CL, we: - rely less on Id strings. Internally we now use *types.Method (and its components) almost everywhere. - stop thinking of types.Methods as objects. They don't have stable identities. (Hopefully they will become plain-old structs soon.) - eliminate receiver indirection wrappers: indirection and promotion are handled together by makeWrapper. - Handle the interactions of promotion, indirection and abstract methods much more cleanly. - support receiver-bound interface method closures. - break up builder.selectField so we can re-use parts (emitFieldSelection). - add importer.PackageInfo.classifySelector utility. - delete interfaceMethodIndex() - delete namedTypeMethodIndex() - delete isSuperInterface() (replaced by types.IsAssignable) - call memberFromObject on each declared concrete method's *types.Func, not on every Method frem each method set, in the CREATE phase for packages loaded by gcimporter. go/types: - document Func, Signature.Recv() better. - use fmt in {Package,Label}.String - reimplement Func.String to be prettier and to include method receivers. API changes: - Function.method now holds the types.Method (soon to be not-an-object) for synthetic wrappers. - CallCommon.Method now contains an abstract (interface) method object; was an abstract method index. - CallCommon.MethodId() gone. - Program.LookupMethod now takes a *Method not an Id string. R=gri CC=golang-dev https://golang.org/cl/11674043
2013-07-26 09:22:34 -06:00
case types.FieldVal:
wantAddr := true
v := b.receiver(fn, e.X, wantAddr, escaping, sel)
last := len(sel.Index()) - 1
return &address{
addr: emitFieldSelection(fn, v, sel.Index()[last], true, e.Sel.Pos()),
expr: e.Sel,
}
}
case *ast.IndexExpr:
var x Value
var et types.Type
switch t := fn.Pkg.typeOf(e.X).Underlying().(type) {
case *types.Array:
go.tools/ssa: add debug information for all ast.Idents. This CL adds three new functions to determine the SSA Value for a given syntactic var, func or const object: Program.{Const,Func,Var}Value. Since constants and functions are immutable, the first two only need a types.Object; but each distinct reference to a var may return a distinct Value, so the third requires an ast.Ident parameter too. Debug information for local vars is encoded in the instruction stream in the form of DebugRef instructions, which are a no-op but relate their operand to a particular ident in the AST. The beauty of this approach is that it naturally stays consistent during optimisation passes (e.g. lifting) without additional bookkeeping. DebugRef instructions are only generated if the DebugMode builder flag is set; I plan to make the policy more fine- grained (per function). DebugRef instructions are inserted for: - expr(Ident) for rvalue idents - address.store() for idents that update an lvalue - address.address() for idents that take address of lvalue (this new method replaces all uses of lval.(address).addr) - expr() for all constant expressions - local ValueSpecs with implicit zero initialization (no RHS) (this case doesn't call store() or address()) To ensure we don't forget to emit debug info for uses of Idents, we must use the lvalue mechanism consistently. (Previously, many simple cases had effectively inlined these functions.) Similarly setCallFunc no longer inlines expr(Ident). Also: - Program.Value() has been inlined & specialized. - Program.Package() has moved nearer the new lookup functions. - refactoring: funcSyntax has lost paramFields, resultFields; gained funcType, which provides access to both. - add package-level constants to Package.values map. - opt: don't call localValueSpec for constants. (The resulting code is always optimised away.) There are a number of comments asking whether Literals should have positions. Will address in a follow-up. Added tests of all interesting cases. R=gri CC=golang-dev https://golang.org/cl/11259044
2013-07-15 11:56:46 -06:00
x = b.addr(fn, e.X, escaping).address(fn)
et = types.NewPointer(t.Elem())
case *types.Pointer: // *array
x = b.expr(fn, e.X)
et = types.NewPointer(t.Elem().Underlying().(*types.Array).Elem())
case *types.Slice:
x = b.expr(fn, e.X)
et = types.NewPointer(t.Elem())
case *types.Map:
return &element{
m: b.expr(fn, e.X),
k: emitConv(fn, b.expr(fn, e.Index), t.Key()),
t: t.Elem(),
pos: e.Lbrack,
}
default:
panic("unexpected container type in IndexExpr: " + t.String())
}
v := &IndexAddr{
X: x,
Index: emitConv(fn, b.expr(fn, e.Index), tInt),
}
v.setPos(e.Lbrack)
v.setType(et)
return &address{addr: fn.emit(v), expr: e}
case *ast.StarExpr:
return &address{addr: b.expr(fn, e.X), starPos: e.Star, expr: e}
}
panic(fmt.Sprintf("unexpected address expression: %T", e))
}
// exprInPlace emits to fn code to initialize the lvalue loc with the
// value of expression e.
//
// This is equivalent to loc.store(fn, b.expr(fn, e)) but may
// generate better code in some cases, e.g. for composite literals
// in an addressable location.
//
func (b *builder) exprInPlace(fn *Function, loc lvalue, e ast.Expr) {
if e, ok := e.(*ast.CompositeLit); ok {
// A CompositeLit never evaluates to a pointer,
// so if the type of the location is a pointer,
// an &-operation is implied.
if _, ok := loc.(blank); !ok { // avoid calling blank.typ()
if isPointer(loc.typ()) {
go.tools/ssa: add debug information for all ast.Idents. This CL adds three new functions to determine the SSA Value for a given syntactic var, func or const object: Program.{Const,Func,Var}Value. Since constants and functions are immutable, the first two only need a types.Object; but each distinct reference to a var may return a distinct Value, so the third requires an ast.Ident parameter too. Debug information for local vars is encoded in the instruction stream in the form of DebugRef instructions, which are a no-op but relate their operand to a particular ident in the AST. The beauty of this approach is that it naturally stays consistent during optimisation passes (e.g. lifting) without additional bookkeeping. DebugRef instructions are only generated if the DebugMode builder flag is set; I plan to make the policy more fine- grained (per function). DebugRef instructions are inserted for: - expr(Ident) for rvalue idents - address.store() for idents that update an lvalue - address.address() for idents that take address of lvalue (this new method replaces all uses of lval.(address).addr) - expr() for all constant expressions - local ValueSpecs with implicit zero initialization (no RHS) (this case doesn't call store() or address()) To ensure we don't forget to emit debug info for uses of Idents, we must use the lvalue mechanism consistently. (Previously, many simple cases had effectively inlined these functions.) Similarly setCallFunc no longer inlines expr(Ident). Also: - Program.Value() has been inlined & specialized. - Program.Package() has moved nearer the new lookup functions. - refactoring: funcSyntax has lost paramFields, resultFields; gained funcType, which provides access to both. - add package-level constants to Package.values map. - opt: don't call localValueSpec for constants. (The resulting code is always optimised away.) There are a number of comments asking whether Literals should have positions. Will address in a follow-up. Added tests of all interesting cases. R=gri CC=golang-dev https://golang.org/cl/11259044
2013-07-15 11:56:46 -06:00
ptr := b.addr(fn, e, true).address(fn)
loc.store(fn, ptr) // copy address
return
}
}
if _, ok := loc.(*address); ok {
typ := loc.typ()
if _, ok := typ.Underlying().(*types.Interface); ok {
// e.g. var x interface{} = T{...}
// Can't in-place initialize an interface value.
// Fall back to copying.
} else {
go.tools/ssa: add debug information for all ast.Idents. This CL adds three new functions to determine the SSA Value for a given syntactic var, func or const object: Program.{Const,Func,Var}Value. Since constants and functions are immutable, the first two only need a types.Object; but each distinct reference to a var may return a distinct Value, so the third requires an ast.Ident parameter too. Debug information for local vars is encoded in the instruction stream in the form of DebugRef instructions, which are a no-op but relate their operand to a particular ident in the AST. The beauty of this approach is that it naturally stays consistent during optimisation passes (e.g. lifting) without additional bookkeeping. DebugRef instructions are only generated if the DebugMode builder flag is set; I plan to make the policy more fine- grained (per function). DebugRef instructions are inserted for: - expr(Ident) for rvalue idents - address.store() for idents that update an lvalue - address.address() for idents that take address of lvalue (this new method replaces all uses of lval.(address).addr) - expr() for all constant expressions - local ValueSpecs with implicit zero initialization (no RHS) (this case doesn't call store() or address()) To ensure we don't forget to emit debug info for uses of Idents, we must use the lvalue mechanism consistently. (Previously, many simple cases had effectively inlined these functions.) Similarly setCallFunc no longer inlines expr(Ident). Also: - Program.Value() has been inlined & specialized. - Program.Package() has moved nearer the new lookup functions. - refactoring: funcSyntax has lost paramFields, resultFields; gained funcType, which provides access to both. - add package-level constants to Package.values map. - opt: don't call localValueSpec for constants. (The resulting code is always optimised away.) There are a number of comments asking whether Literals should have positions. Will address in a follow-up. Added tests of all interesting cases. R=gri CC=golang-dev https://golang.org/cl/11259044
2013-07-15 11:56:46 -06:00
b.compLit(fn, loc.address(fn), e, typ) // in place
return
}
}
}
loc.store(fn, b.expr(fn, e)) // copy value
}
// expr lowers a single-result expression e to SSA form, emitting code
// to fn and returning the Value defined by the expression.
//
func (b *builder) expr(fn *Function, e ast.Expr) Value {
// Is expression a constant?
if v := fn.Pkg.info.ValueOf(e); v != nil {
return NewConst(v, fn.Pkg.typeOf(e))
}
e = unparen(e)
v := b.expr0(fn, e)
if fn.debugInfo() {
emitDebugRef(fn, e, v)
}
return v
}
func (b *builder) expr0(fn *Function, e ast.Expr) Value {
switch e := e.(type) {
case *ast.BasicLit:
panic("non-constant BasicLit") // unreachable
case *ast.FuncLit:
posn := fn.Prog.Fset.Position(e.Type.Func)
fn2 := &Function{
name: fmt.Sprintf("func@%d.%d", posn.Line, posn.Column),
Signature: fn.Pkg.typeOf(e.Type).Underlying().(*types.Signature),
pos: e.Type.Func,
Enclosing: fn,
Pkg: fn.Pkg,
Prog: fn.Prog,
syntax: &funcSyntax{
go.tools/ssa: add debug information for all ast.Idents. This CL adds three new functions to determine the SSA Value for a given syntactic var, func or const object: Program.{Const,Func,Var}Value. Since constants and functions are immutable, the first two only need a types.Object; but each distinct reference to a var may return a distinct Value, so the third requires an ast.Ident parameter too. Debug information for local vars is encoded in the instruction stream in the form of DebugRef instructions, which are a no-op but relate their operand to a particular ident in the AST. The beauty of this approach is that it naturally stays consistent during optimisation passes (e.g. lifting) without additional bookkeeping. DebugRef instructions are only generated if the DebugMode builder flag is set; I plan to make the policy more fine- grained (per function). DebugRef instructions are inserted for: - expr(Ident) for rvalue idents - address.store() for idents that update an lvalue - address.address() for idents that take address of lvalue (this new method replaces all uses of lval.(address).addr) - expr() for all constant expressions - local ValueSpecs with implicit zero initialization (no RHS) (this case doesn't call store() or address()) To ensure we don't forget to emit debug info for uses of Idents, we must use the lvalue mechanism consistently. (Previously, many simple cases had effectively inlined these functions.) Similarly setCallFunc no longer inlines expr(Ident). Also: - Program.Value() has been inlined & specialized. - Program.Package() has moved nearer the new lookup functions. - refactoring: funcSyntax has lost paramFields, resultFields; gained funcType, which provides access to both. - add package-level constants to Package.values map. - opt: don't call localValueSpec for constants. (The resulting code is always optimised away.) There are a number of comments asking whether Literals should have positions. Will address in a follow-up. Added tests of all interesting cases. R=gri CC=golang-dev https://golang.org/cl/11259044
2013-07-15 11:56:46 -06:00
functype: e.Type,
body: e.Body,
},
}
fn.AnonFuncs = append(fn.AnonFuncs, fn2)
b.buildFunction(fn2)
if fn2.FreeVars == nil {
return fn2
}
v := &MakeClosure{Fn: fn2}
v.setType(fn.Pkg.typeOf(e))
for _, fv := range fn2.FreeVars {
v.Bindings = append(v.Bindings, fv.outer)
fv.outer = nil
}
return fn.emit(v)
case *ast.TypeAssertExpr: // single-result form only
return emitTypeAssert(fn, b.expr(fn, e.X), fn.Pkg.typeOf(e), e.Lparen)
case *ast.CallExpr:
typ := fn.Pkg.typeOf(e)
if fn.Pkg.info.IsType(e.Fun) {
// Explicit type conversion, e.g. string(x) or big.Int(x)
x := b.expr(fn, e.Args[0])
y := emitConv(fn, x, typ)
if y != x {
switch y := y.(type) {
case *Convert:
y.pos = e.Lparen
case *ChangeType:
y.pos = e.Lparen
case *MakeInterface:
y.pos = e.Lparen
}
}
return y
}
// Call to "intrinsic" built-ins, e.g. new, make, panic.
if id, ok := e.Fun.(*ast.Ident); ok {
if obj, ok := fn.Pkg.objectOf(id).(*types.Builtin); ok {
if v := b.builtin(fn, obj, e.Args, typ, e.Lparen); v != nil {
return v
}
}
}
// Regular function call.
var v Call
b.setCall(fn, e, &v.Call)
v.setType(typ)
return fn.emit(&v)
case *ast.UnaryExpr:
switch e.Op {
case token.AND: // &X --- potentially escaping.
addr := b.addr(fn, e.X, true)
if _, ok := unparen(e.X).(*ast.StarExpr); ok {
// &*p must panic if p is nil (http://golang.org/s/go12nil).
// For simplicity, we'll just (suboptimally) rely
// on the side effects of a load.
addr.load(fn)
}
return addr.address(fn)
case token.ADD:
return b.expr(fn, e.X)
case token.NOT, token.ARROW, token.SUB, token.XOR: // ! <- - ^
v := &UnOp{
Op: e.Op,
X: b.expr(fn, e.X),
}
v.setPos(e.OpPos)
v.setType(fn.Pkg.typeOf(e))
return fn.emit(v)
default:
panic(e.Op)
}
case *ast.BinaryExpr:
switch e.Op {
case token.LAND, token.LOR:
return b.logicalBinop(fn, e)
case token.SHL, token.SHR:
fallthrough
case token.ADD, token.SUB, token.MUL, token.QUO, token.REM, token.AND, token.OR, token.XOR, token.AND_NOT:
return emitArith(fn, e.Op, b.expr(fn, e.X), b.expr(fn, e.Y), fn.Pkg.typeOf(e), e.OpPos)
case token.EQL, token.NEQ, token.GTR, token.LSS, token.LEQ, token.GEQ:
// TODO(gri): we shouldn't need DefaultType here.
cmp := emitCompare(fn, e.Op, b.expr(fn, e.X), b.expr(fn, e.Y), e.OpPos)
return emitConv(fn, cmp, DefaultType(fn.Pkg.typeOf(e)))
default:
panic("illegal op in BinaryExpr: " + e.Op.String())
}
case *ast.SliceExpr:
var low, high Value
var x Value
switch fn.Pkg.typeOf(e.X).Underlying().(type) {
case *types.Array:
// Potentially escaping.
go.tools/ssa: add debug information for all ast.Idents. This CL adds three new functions to determine the SSA Value for a given syntactic var, func or const object: Program.{Const,Func,Var}Value. Since constants and functions are immutable, the first two only need a types.Object; but each distinct reference to a var may return a distinct Value, so the third requires an ast.Ident parameter too. Debug information for local vars is encoded in the instruction stream in the form of DebugRef instructions, which are a no-op but relate their operand to a particular ident in the AST. The beauty of this approach is that it naturally stays consistent during optimisation passes (e.g. lifting) without additional bookkeeping. DebugRef instructions are only generated if the DebugMode builder flag is set; I plan to make the policy more fine- grained (per function). DebugRef instructions are inserted for: - expr(Ident) for rvalue idents - address.store() for idents that update an lvalue - address.address() for idents that take address of lvalue (this new method replaces all uses of lval.(address).addr) - expr() for all constant expressions - local ValueSpecs with implicit zero initialization (no RHS) (this case doesn't call store() or address()) To ensure we don't forget to emit debug info for uses of Idents, we must use the lvalue mechanism consistently. (Previously, many simple cases had effectively inlined these functions.) Similarly setCallFunc no longer inlines expr(Ident). Also: - Program.Value() has been inlined & specialized. - Program.Package() has moved nearer the new lookup functions. - refactoring: funcSyntax has lost paramFields, resultFields; gained funcType, which provides access to both. - add package-level constants to Package.values map. - opt: don't call localValueSpec for constants. (The resulting code is always optimised away.) There are a number of comments asking whether Literals should have positions. Will address in a follow-up. Added tests of all interesting cases. R=gri CC=golang-dev https://golang.org/cl/11259044
2013-07-15 11:56:46 -06:00
x = b.addr(fn, e.X, true).address(fn)
case *types.Basic, *types.Slice, *types.Pointer: // *array
x = b.expr(fn, e.X)
default:
unreachable()
}
if e.High != nil {
high = b.expr(fn, e.High)
}
if e.Low != nil {
low = b.expr(fn, e.Low)
}
v := &Slice{
X: x,
Low: low,
High: high,
}
v.setPos(e.Lbrack)
v.setType(fn.Pkg.typeOf(e))
return fn.emit(v)
case *ast.Ident:
obj := fn.Pkg.objectOf(e)
// Universal built-in or nil?
switch obj := obj.(type) {
case *types.Builtin:
return fn.Prog.builtins[obj]
case *types.Nil:
return nilConst(fn.Pkg.typeOf(e))
}
// Package-level func or var?
if v := b.lookup(fn.Pkg, obj); v != nil {
if _, ok := obj.(*types.Var); ok {
return emitLoad(fn, v) // var (address)
}
return v // (func)
}
go.tools/ssa: add debug information for all ast.Idents. This CL adds three new functions to determine the SSA Value for a given syntactic var, func or const object: Program.{Const,Func,Var}Value. Since constants and functions are immutable, the first two only need a types.Object; but each distinct reference to a var may return a distinct Value, so the third requires an ast.Ident parameter too. Debug information for local vars is encoded in the instruction stream in the form of DebugRef instructions, which are a no-op but relate their operand to a particular ident in the AST. The beauty of this approach is that it naturally stays consistent during optimisation passes (e.g. lifting) without additional bookkeeping. DebugRef instructions are only generated if the DebugMode builder flag is set; I plan to make the policy more fine- grained (per function). DebugRef instructions are inserted for: - expr(Ident) for rvalue idents - address.store() for idents that update an lvalue - address.address() for idents that take address of lvalue (this new method replaces all uses of lval.(address).addr) - expr() for all constant expressions - local ValueSpecs with implicit zero initialization (no RHS) (this case doesn't call store() or address()) To ensure we don't forget to emit debug info for uses of Idents, we must use the lvalue mechanism consistently. (Previously, many simple cases had effectively inlined these functions.) Similarly setCallFunc no longer inlines expr(Ident). Also: - Program.Value() has been inlined & specialized. - Program.Package() has moved nearer the new lookup functions. - refactoring: funcSyntax has lost paramFields, resultFields; gained funcType, which provides access to both. - add package-level constants to Package.values map. - opt: don't call localValueSpec for constants. (The resulting code is always optimised away.) There are a number of comments asking whether Literals should have positions. Will address in a follow-up. Added tests of all interesting cases. R=gri CC=golang-dev https://golang.org/cl/11259044
2013-07-15 11:56:46 -06:00
// Local var.
return emitLoad(fn, fn.lookup(obj, false)) // var (address)
case *ast.SelectorExpr:
switch sel := fn.Pkg.info.Selections[e]; sel.Kind() {
case types.PackageObj:
return b.expr(fn, e.Sel)
case types.MethodExpr:
// (*T).f or T.f, the method f from the method-set of type T.
go.tools/ssa: (another) major refactoring of method-set logic. We now use LookupFieldOrMethod for all SelectorExprs, and simplify the logic to discriminate the various cases. We inline static calls to promoted/indirected functions, dramatically reducing the number of functions created. More tests are needed, but I'd like to submit this as-is. In this CL, we: - rely less on Id strings. Internally we now use *types.Method (and its components) almost everywhere. - stop thinking of types.Methods as objects. They don't have stable identities. (Hopefully they will become plain-old structs soon.) - eliminate receiver indirection wrappers: indirection and promotion are handled together by makeWrapper. - Handle the interactions of promotion, indirection and abstract methods much more cleanly. - support receiver-bound interface method closures. - break up builder.selectField so we can re-use parts (emitFieldSelection). - add importer.PackageInfo.classifySelector utility. - delete interfaceMethodIndex() - delete namedTypeMethodIndex() - delete isSuperInterface() (replaced by types.IsAssignable) - call memberFromObject on each declared concrete method's *types.Func, not on every Method frem each method set, in the CREATE phase for packages loaded by gcimporter. go/types: - document Func, Signature.Recv() better. - use fmt in {Package,Label}.String - reimplement Func.String to be prettier and to include method receivers. API changes: - Function.method now holds the types.Method (soon to be not-an-object) for synthetic wrappers. - CallCommon.Method now contains an abstract (interface) method object; was an abstract method index. - CallCommon.MethodId() gone. - Program.LookupMethod now takes a *Method not an Id string. R=gri CC=golang-dev https://golang.org/cl/11674043
2013-07-26 09:22:34 -06:00
// For declared methods, a simple conversion will suffice.
return emitConv(fn, fn.Prog.Method(sel), fn.Pkg.typeOf(e))
case types.MethodVal:
go.tools/ssa: (another) major refactoring of method-set logic. We now use LookupFieldOrMethod for all SelectorExprs, and simplify the logic to discriminate the various cases. We inline static calls to promoted/indirected functions, dramatically reducing the number of functions created. More tests are needed, but I'd like to submit this as-is. In this CL, we: - rely less on Id strings. Internally we now use *types.Method (and its components) almost everywhere. - stop thinking of types.Methods as objects. They don't have stable identities. (Hopefully they will become plain-old structs soon.) - eliminate receiver indirection wrappers: indirection and promotion are handled together by makeWrapper. - Handle the interactions of promotion, indirection and abstract methods much more cleanly. - support receiver-bound interface method closures. - break up builder.selectField so we can re-use parts (emitFieldSelection). - add importer.PackageInfo.classifySelector utility. - delete interfaceMethodIndex() - delete namedTypeMethodIndex() - delete isSuperInterface() (replaced by types.IsAssignable) - call memberFromObject on each declared concrete method's *types.Func, not on every Method frem each method set, in the CREATE phase for packages loaded by gcimporter. go/types: - document Func, Signature.Recv() better. - use fmt in {Package,Label}.String - reimplement Func.String to be prettier and to include method receivers. API changes: - Function.method now holds the types.Method (soon to be not-an-object) for synthetic wrappers. - CallCommon.Method now contains an abstract (interface) method object; was an abstract method index. - CallCommon.MethodId() gone. - Program.LookupMethod now takes a *Method not an Id string. R=gri CC=golang-dev https://golang.org/cl/11674043
2013-07-26 09:22:34 -06:00
// e.f where e is an expression and f is a method.
// The result is a bound method closure.
obj := sel.Obj().(*types.Func)
go.tools/ssa: (another) major refactoring of method-set logic. We now use LookupFieldOrMethod for all SelectorExprs, and simplify the logic to discriminate the various cases. We inline static calls to promoted/indirected functions, dramatically reducing the number of functions created. More tests are needed, but I'd like to submit this as-is. In this CL, we: - rely less on Id strings. Internally we now use *types.Method (and its components) almost everywhere. - stop thinking of types.Methods as objects. They don't have stable identities. (Hopefully they will become plain-old structs soon.) - eliminate receiver indirection wrappers: indirection and promotion are handled together by makeWrapper. - Handle the interactions of promotion, indirection and abstract methods much more cleanly. - support receiver-bound interface method closures. - break up builder.selectField so we can re-use parts (emitFieldSelection). - add importer.PackageInfo.classifySelector utility. - delete interfaceMethodIndex() - delete namedTypeMethodIndex() - delete isSuperInterface() (replaced by types.IsAssignable) - call memberFromObject on each declared concrete method's *types.Func, not on every Method frem each method set, in the CREATE phase for packages loaded by gcimporter. go/types: - document Func, Signature.Recv() better. - use fmt in {Package,Label}.String - reimplement Func.String to be prettier and to include method receivers. API changes: - Function.method now holds the types.Method (soon to be not-an-object) for synthetic wrappers. - CallCommon.Method now contains an abstract (interface) method object; was an abstract method index. - CallCommon.MethodId() gone. - Program.LookupMethod now takes a *Method not an Id string. R=gri CC=golang-dev https://golang.org/cl/11674043
2013-07-26 09:22:34 -06:00
wantAddr := isPointer(recvType(obj))
escaping := true
v := b.receiver(fn, e.X, wantAddr, escaping, sel)
c := &MakeClosure{
go.tools/ssa: (another) major refactoring of method-set logic. We now use LookupFieldOrMethod for all SelectorExprs, and simplify the logic to discriminate the various cases. We inline static calls to promoted/indirected functions, dramatically reducing the number of functions created. More tests are needed, but I'd like to submit this as-is. In this CL, we: - rely less on Id strings. Internally we now use *types.Method (and its components) almost everywhere. - stop thinking of types.Methods as objects. They don't have stable identities. (Hopefully they will become plain-old structs soon.) - eliminate receiver indirection wrappers: indirection and promotion are handled together by makeWrapper. - Handle the interactions of promotion, indirection and abstract methods much more cleanly. - support receiver-bound interface method closures. - break up builder.selectField so we can re-use parts (emitFieldSelection). - add importer.PackageInfo.classifySelector utility. - delete interfaceMethodIndex() - delete namedTypeMethodIndex() - delete isSuperInterface() (replaced by types.IsAssignable) - call memberFromObject on each declared concrete method's *types.Func, not on every Method frem each method set, in the CREATE phase for packages loaded by gcimporter. go/types: - document Func, Signature.Recv() better. - use fmt in {Package,Label}.String - reimplement Func.String to be prettier and to include method receivers. API changes: - Function.method now holds the types.Method (soon to be not-an-object) for synthetic wrappers. - CallCommon.Method now contains an abstract (interface) method object; was an abstract method index. - CallCommon.MethodId() gone. - Program.LookupMethod now takes a *Method not an Id string. R=gri CC=golang-dev https://golang.org/cl/11674043
2013-07-26 09:22:34 -06:00
Fn: boundMethodWrapper(fn.Prog, obj),
Bindings: []Value{v},
}
c.setPos(e.Sel.Pos())
c.setType(fn.Pkg.typeOf(e))
return fn.emit(c)
go.tools/ssa: (another) major refactoring of method-set logic. We now use LookupFieldOrMethod for all SelectorExprs, and simplify the logic to discriminate the various cases. We inline static calls to promoted/indirected functions, dramatically reducing the number of functions created. More tests are needed, but I'd like to submit this as-is. In this CL, we: - rely less on Id strings. Internally we now use *types.Method (and its components) almost everywhere. - stop thinking of types.Methods as objects. They don't have stable identities. (Hopefully they will become plain-old structs soon.) - eliminate receiver indirection wrappers: indirection and promotion are handled together by makeWrapper. - Handle the interactions of promotion, indirection and abstract methods much more cleanly. - support receiver-bound interface method closures. - break up builder.selectField so we can re-use parts (emitFieldSelection). - add importer.PackageInfo.classifySelector utility. - delete interfaceMethodIndex() - delete namedTypeMethodIndex() - delete isSuperInterface() (replaced by types.IsAssignable) - call memberFromObject on each declared concrete method's *types.Func, not on every Method frem each method set, in the CREATE phase for packages loaded by gcimporter. go/types: - document Func, Signature.Recv() better. - use fmt in {Package,Label}.String - reimplement Func.String to be prettier and to include method receivers. API changes: - Function.method now holds the types.Method (soon to be not-an-object) for synthetic wrappers. - CallCommon.Method now contains an abstract (interface) method object; was an abstract method index. - CallCommon.MethodId() gone. - Program.LookupMethod now takes a *Method not an Id string. R=gri CC=golang-dev https://golang.org/cl/11674043
2013-07-26 09:22:34 -06:00
case types.FieldVal:
indices := sel.Index()
go.tools/ssa: (another) major refactoring of method-set logic. We now use LookupFieldOrMethod for all SelectorExprs, and simplify the logic to discriminate the various cases. We inline static calls to promoted/indirected functions, dramatically reducing the number of functions created. More tests are needed, but I'd like to submit this as-is. In this CL, we: - rely less on Id strings. Internally we now use *types.Method (and its components) almost everywhere. - stop thinking of types.Methods as objects. They don't have stable identities. (Hopefully they will become plain-old structs soon.) - eliminate receiver indirection wrappers: indirection and promotion are handled together by makeWrapper. - Handle the interactions of promotion, indirection and abstract methods much more cleanly. - support receiver-bound interface method closures. - break up builder.selectField so we can re-use parts (emitFieldSelection). - add importer.PackageInfo.classifySelector utility. - delete interfaceMethodIndex() - delete namedTypeMethodIndex() - delete isSuperInterface() (replaced by types.IsAssignable) - call memberFromObject on each declared concrete method's *types.Func, not on every Method frem each method set, in the CREATE phase for packages loaded by gcimporter. go/types: - document Func, Signature.Recv() better. - use fmt in {Package,Label}.String - reimplement Func.String to be prettier and to include method receivers. API changes: - Function.method now holds the types.Method (soon to be not-an-object) for synthetic wrappers. - CallCommon.Method now contains an abstract (interface) method object; was an abstract method index. - CallCommon.MethodId() gone. - Program.LookupMethod now takes a *Method not an Id string. R=gri CC=golang-dev https://golang.org/cl/11674043
2013-07-26 09:22:34 -06:00
last := len(indices) - 1
v := b.expr(fn, e.X)
v = emitImplicitSelections(fn, v, indices[:last])
v = emitFieldSelection(fn, v, indices[last], false, e.Sel.Pos())
emitDebugRef(fn, e.Sel, v)
return v
}
go.tools/ssa: (another) major refactoring of method-set logic. We now use LookupFieldOrMethod for all SelectorExprs, and simplify the logic to discriminate the various cases. We inline static calls to promoted/indirected functions, dramatically reducing the number of functions created. More tests are needed, but I'd like to submit this as-is. In this CL, we: - rely less on Id strings. Internally we now use *types.Method (and its components) almost everywhere. - stop thinking of types.Methods as objects. They don't have stable identities. (Hopefully they will become plain-old structs soon.) - eliminate receiver indirection wrappers: indirection and promotion are handled together by makeWrapper. - Handle the interactions of promotion, indirection and abstract methods much more cleanly. - support receiver-bound interface method closures. - break up builder.selectField so we can re-use parts (emitFieldSelection). - add importer.PackageInfo.classifySelector utility. - delete interfaceMethodIndex() - delete namedTypeMethodIndex() - delete isSuperInterface() (replaced by types.IsAssignable) - call memberFromObject on each declared concrete method's *types.Func, not on every Method frem each method set, in the CREATE phase for packages loaded by gcimporter. go/types: - document Func, Signature.Recv() better. - use fmt in {Package,Label}.String - reimplement Func.String to be prettier and to include method receivers. API changes: - Function.method now holds the types.Method (soon to be not-an-object) for synthetic wrappers. - CallCommon.Method now contains an abstract (interface) method object; was an abstract method index. - CallCommon.MethodId() gone. - Program.LookupMethod now takes a *Method not an Id string. R=gri CC=golang-dev https://golang.org/cl/11674043
2013-07-26 09:22:34 -06:00
panic("unexpected expression-relative selector")
case *ast.IndexExpr:
switch t := fn.Pkg.typeOf(e.X).Underlying().(type) {
case *types.Array:
// Non-addressable array (in a register).
v := &Index{
X: b.expr(fn, e.X),
Index: emitConv(fn, b.expr(fn, e.Index), tInt),
}
v.setPos(e.Lbrack)
v.setType(t.Elem())
return fn.emit(v)
case *types.Map:
// Maps are not addressable.
mapt := fn.Pkg.typeOf(e.X).Underlying().(*types.Map)
v := &Lookup{
X: b.expr(fn, e.X),
Index: emitConv(fn, b.expr(fn, e.Index), mapt.Key()),
}
v.setPos(e.Lbrack)
v.setType(mapt.Elem())
return fn.emit(v)
case *types.Basic: // => string
// Strings are not addressable.
v := &Lookup{
X: b.expr(fn, e.X),
Index: b.expr(fn, e.Index),
}
v.setPos(e.Lbrack)
v.setType(tByte)
return fn.emit(v)
case *types.Slice, *types.Pointer: // *array
// Addressable slice/array; use IndexAddr and Load.
return b.addr(fn, e, false).load(fn)
default:
panic("unexpected container type in IndexExpr: " + t.String())
}
case *ast.CompositeLit, *ast.StarExpr:
// Addressable types (lvalues)
return b.addr(fn, e, false).load(fn)
}
panic(fmt.Sprintf("unexpected expr: %T", e))
}
// stmtList emits to fn code for all statements in list.
func (b *builder) stmtList(fn *Function, list []ast.Stmt) {
for _, s := range list {
b.stmt(fn, s)
}
}
go.tools/ssa: (another) major refactoring of method-set logic. We now use LookupFieldOrMethod for all SelectorExprs, and simplify the logic to discriminate the various cases. We inline static calls to promoted/indirected functions, dramatically reducing the number of functions created. More tests are needed, but I'd like to submit this as-is. In this CL, we: - rely less on Id strings. Internally we now use *types.Method (and its components) almost everywhere. - stop thinking of types.Methods as objects. They don't have stable identities. (Hopefully they will become plain-old structs soon.) - eliminate receiver indirection wrappers: indirection and promotion are handled together by makeWrapper. - Handle the interactions of promotion, indirection and abstract methods much more cleanly. - support receiver-bound interface method closures. - break up builder.selectField so we can re-use parts (emitFieldSelection). - add importer.PackageInfo.classifySelector utility. - delete interfaceMethodIndex() - delete namedTypeMethodIndex() - delete isSuperInterface() (replaced by types.IsAssignable) - call memberFromObject on each declared concrete method's *types.Func, not on every Method frem each method set, in the CREATE phase for packages loaded by gcimporter. go/types: - document Func, Signature.Recv() better. - use fmt in {Package,Label}.String - reimplement Func.String to be prettier and to include method receivers. API changes: - Function.method now holds the types.Method (soon to be not-an-object) for synthetic wrappers. - CallCommon.Method now contains an abstract (interface) method object; was an abstract method index. - CallCommon.MethodId() gone. - Program.LookupMethod now takes a *Method not an Id string. R=gri CC=golang-dev https://golang.org/cl/11674043
2013-07-26 09:22:34 -06:00
// receiver emits to fn code for expression e in the "receiver"
// position of selection e.f (where f may be a field or a method) and
// returns the effective receiver after applying the implicit field
// selections of sel.
go.tools/ssa: (another) major refactoring of method-set logic. We now use LookupFieldOrMethod for all SelectorExprs, and simplify the logic to discriminate the various cases. We inline static calls to promoted/indirected functions, dramatically reducing the number of functions created. More tests are needed, but I'd like to submit this as-is. In this CL, we: - rely less on Id strings. Internally we now use *types.Method (and its components) almost everywhere. - stop thinking of types.Methods as objects. They don't have stable identities. (Hopefully they will become plain-old structs soon.) - eliminate receiver indirection wrappers: indirection and promotion are handled together by makeWrapper. - Handle the interactions of promotion, indirection and abstract methods much more cleanly. - support receiver-bound interface method closures. - break up builder.selectField so we can re-use parts (emitFieldSelection). - add importer.PackageInfo.classifySelector utility. - delete interfaceMethodIndex() - delete namedTypeMethodIndex() - delete isSuperInterface() (replaced by types.IsAssignable) - call memberFromObject on each declared concrete method's *types.Func, not on every Method frem each method set, in the CREATE phase for packages loaded by gcimporter. go/types: - document Func, Signature.Recv() better. - use fmt in {Package,Label}.String - reimplement Func.String to be prettier and to include method receivers. API changes: - Function.method now holds the types.Method (soon to be not-an-object) for synthetic wrappers. - CallCommon.Method now contains an abstract (interface) method object; was an abstract method index. - CallCommon.MethodId() gone. - Program.LookupMethod now takes a *Method not an Id string. R=gri CC=golang-dev https://golang.org/cl/11674043
2013-07-26 09:22:34 -06:00
//
// wantAddr requests that the result is an an address. If
// !sel.Indirect(), this may require that e be build in addr() mode; it
go.tools/ssa: (another) major refactoring of method-set logic. We now use LookupFieldOrMethod for all SelectorExprs, and simplify the logic to discriminate the various cases. We inline static calls to promoted/indirected functions, dramatically reducing the number of functions created. More tests are needed, but I'd like to submit this as-is. In this CL, we: - rely less on Id strings. Internally we now use *types.Method (and its components) almost everywhere. - stop thinking of types.Methods as objects. They don't have stable identities. (Hopefully they will become plain-old structs soon.) - eliminate receiver indirection wrappers: indirection and promotion are handled together by makeWrapper. - Handle the interactions of promotion, indirection and abstract methods much more cleanly. - support receiver-bound interface method closures. - break up builder.selectField so we can re-use parts (emitFieldSelection). - add importer.PackageInfo.classifySelector utility. - delete interfaceMethodIndex() - delete namedTypeMethodIndex() - delete isSuperInterface() (replaced by types.IsAssignable) - call memberFromObject on each declared concrete method's *types.Func, not on every Method frem each method set, in the CREATE phase for packages loaded by gcimporter. go/types: - document Func, Signature.Recv() better. - use fmt in {Package,Label}.String - reimplement Func.String to be prettier and to include method receivers. API changes: - Function.method now holds the types.Method (soon to be not-an-object) for synthetic wrappers. - CallCommon.Method now contains an abstract (interface) method object; was an abstract method index. - CallCommon.MethodId() gone. - Program.LookupMethod now takes a *Method not an Id string. R=gri CC=golang-dev https://golang.org/cl/11674043
2013-07-26 09:22:34 -06:00
// must thus be addressable.
//
go.tools/ssa: (another) major refactoring of method-set logic. We now use LookupFieldOrMethod for all SelectorExprs, and simplify the logic to discriminate the various cases. We inline static calls to promoted/indirected functions, dramatically reducing the number of functions created. More tests are needed, but I'd like to submit this as-is. In this CL, we: - rely less on Id strings. Internally we now use *types.Method (and its components) almost everywhere. - stop thinking of types.Methods as objects. They don't have stable identities. (Hopefully they will become plain-old structs soon.) - eliminate receiver indirection wrappers: indirection and promotion are handled together by makeWrapper. - Handle the interactions of promotion, indirection and abstract methods much more cleanly. - support receiver-bound interface method closures. - break up builder.selectField so we can re-use parts (emitFieldSelection). - add importer.PackageInfo.classifySelector utility. - delete interfaceMethodIndex() - delete namedTypeMethodIndex() - delete isSuperInterface() (replaced by types.IsAssignable) - call memberFromObject on each declared concrete method's *types.Func, not on every Method frem each method set, in the CREATE phase for packages loaded by gcimporter. go/types: - document Func, Signature.Recv() better. - use fmt in {Package,Label}.String - reimplement Func.String to be prettier and to include method receivers. API changes: - Function.method now holds the types.Method (soon to be not-an-object) for synthetic wrappers. - CallCommon.Method now contains an abstract (interface) method object; was an abstract method index. - CallCommon.MethodId() gone. - Program.LookupMethod now takes a *Method not an Id string. R=gri CC=golang-dev https://golang.org/cl/11674043
2013-07-26 09:22:34 -06:00
// escaping is defined as per builder.addr().
//
func (b *builder) receiver(fn *Function, e ast.Expr, wantAddr, escaping bool, sel *types.Selection) Value {
go.tools/ssa: (another) major refactoring of method-set logic. We now use LookupFieldOrMethod for all SelectorExprs, and simplify the logic to discriminate the various cases. We inline static calls to promoted/indirected functions, dramatically reducing the number of functions created. More tests are needed, but I'd like to submit this as-is. In this CL, we: - rely less on Id strings. Internally we now use *types.Method (and its components) almost everywhere. - stop thinking of types.Methods as objects. They don't have stable identities. (Hopefully they will become plain-old structs soon.) - eliminate receiver indirection wrappers: indirection and promotion are handled together by makeWrapper. - Handle the interactions of promotion, indirection and abstract methods much more cleanly. - support receiver-bound interface method closures. - break up builder.selectField so we can re-use parts (emitFieldSelection). - add importer.PackageInfo.classifySelector utility. - delete interfaceMethodIndex() - delete namedTypeMethodIndex() - delete isSuperInterface() (replaced by types.IsAssignable) - call memberFromObject on each declared concrete method's *types.Func, not on every Method frem each method set, in the CREATE phase for packages loaded by gcimporter. go/types: - document Func, Signature.Recv() better. - use fmt in {Package,Label}.String - reimplement Func.String to be prettier and to include method receivers. API changes: - Function.method now holds the types.Method (soon to be not-an-object) for synthetic wrappers. - CallCommon.Method now contains an abstract (interface) method object; was an abstract method index. - CallCommon.MethodId() gone. - Program.LookupMethod now takes a *Method not an Id string. R=gri CC=golang-dev https://golang.org/cl/11674043
2013-07-26 09:22:34 -06:00
var v Value
if wantAddr && !sel.Indirect() && !isPointer(fn.Pkg.typeOf(e)) {
go.tools/ssa: (another) major refactoring of method-set logic. We now use LookupFieldOrMethod for all SelectorExprs, and simplify the logic to discriminate the various cases. We inline static calls to promoted/indirected functions, dramatically reducing the number of functions created. More tests are needed, but I'd like to submit this as-is. In this CL, we: - rely less on Id strings. Internally we now use *types.Method (and its components) almost everywhere. - stop thinking of types.Methods as objects. They don't have stable identities. (Hopefully they will become plain-old structs soon.) - eliminate receiver indirection wrappers: indirection and promotion are handled together by makeWrapper. - Handle the interactions of promotion, indirection and abstract methods much more cleanly. - support receiver-bound interface method closures. - break up builder.selectField so we can re-use parts (emitFieldSelection). - add importer.PackageInfo.classifySelector utility. - delete interfaceMethodIndex() - delete namedTypeMethodIndex() - delete isSuperInterface() (replaced by types.IsAssignable) - call memberFromObject on each declared concrete method's *types.Func, not on every Method frem each method set, in the CREATE phase for packages loaded by gcimporter. go/types: - document Func, Signature.Recv() better. - use fmt in {Package,Label}.String - reimplement Func.String to be prettier and to include method receivers. API changes: - Function.method now holds the types.Method (soon to be not-an-object) for synthetic wrappers. - CallCommon.Method now contains an abstract (interface) method object; was an abstract method index. - CallCommon.MethodId() gone. - Program.LookupMethod now takes a *Method not an Id string. R=gri CC=golang-dev https://golang.org/cl/11674043
2013-07-26 09:22:34 -06:00
v = b.addr(fn, e, escaping).address(fn)
} else {
v = b.expr(fn, e)
}
last := len(sel.Index()) - 1
v = emitImplicitSelections(fn, v, sel.Index()[:last])
if !wantAddr && isPointer(v.Type()) {
v = emitLoad(fn, v)
}
return v
}
// setCallFunc populates the function parts of a CallCommon structure
// (Func, Method, Recv, Args[0]) based on the kind of invocation
// occurring in e.
//
func (b *builder) setCallFunc(fn *Function, e *ast.CallExpr, c *CallCommon) {
c.pos = e.Lparen
c.HasEllipsis = e.Ellipsis != 0
// Is this a method call?
if selector, ok := unparen(e.Fun).(*ast.SelectorExpr); ok {
switch sel := fn.Pkg.info.Selections[selector]; sel.Kind() {
case types.PackageObj:
// e.g. fmt.Println
case types.MethodExpr:
// T.f() or (*T).f(): a statically dispatched
// call to the method f in the method-set of T
// or *T. T may be an interface.
// e.Fun would evaluate to a concrete method,
// interface wrapper function, or promotion
// wrapper.
//
// For now, we evaluate it in the usual way.
// TODO(adonovan): opt: inline expr() here, to
// make the call static and to avoid
// generation of wrappers. It's somewhat
// tricky as it may consume the first actual
// parameter if the call is "invoke" mode.
//
// Examples:
// type T struct{}; func (T) f() {} // "call" mode
// type T interface { f() } // "invoke" mode
//
// type S struct{ T }
//
// var s S
// S.f(s)
// (*S).f(&s)
//
// Suggested approach:
// - consume the first actual parameter expression
// and build it with b.expr().
// - apply implicit field selections.
// - use MethodVal logic to populate fields of c.
case types.FieldVal:
// A field access, not a method call.
case types.MethodVal:
obj := sel.Obj().(*types.Func)
wantAddr := isPointer(recvType(obj))
escaping := true
v := b.receiver(fn, selector.X, wantAddr, escaping, sel)
if _, ok := deref(v.Type()).Underlying().(*types.Interface); ok {
// Invoke-mode call.
c.Value = v
c.Method = obj
} else {
// "Call"-mode call.
// TODO(adonovan): fix: in -build=G
// mode, declaredFunc panics for
// cross-package calls.
c.Value = fn.Prog.declaredFunc(obj)
c.Args = append(c.Args, v)
}
return
default:
panic(fmt.Sprintf("illegal (%s).%s() call; X:%T",
fn.Pkg.typeOf(selector.X), selector.Sel.Name, selector.X))
go.tools/ssa: (another) major refactoring of method-set logic. We now use LookupFieldOrMethod for all SelectorExprs, and simplify the logic to discriminate the various cases. We inline static calls to promoted/indirected functions, dramatically reducing the number of functions created. More tests are needed, but I'd like to submit this as-is. In this CL, we: - rely less on Id strings. Internally we now use *types.Method (and its components) almost everywhere. - stop thinking of types.Methods as objects. They don't have stable identities. (Hopefully they will become plain-old structs soon.) - eliminate receiver indirection wrappers: indirection and promotion are handled together by makeWrapper. - Handle the interactions of promotion, indirection and abstract methods much more cleanly. - support receiver-bound interface method closures. - break up builder.selectField so we can re-use parts (emitFieldSelection). - add importer.PackageInfo.classifySelector utility. - delete interfaceMethodIndex() - delete namedTypeMethodIndex() - delete isSuperInterface() (replaced by types.IsAssignable) - call memberFromObject on each declared concrete method's *types.Func, not on every Method frem each method set, in the CREATE phase for packages loaded by gcimporter. go/types: - document Func, Signature.Recv() better. - use fmt in {Package,Label}.String - reimplement Func.String to be prettier and to include method receivers. API changes: - Function.method now holds the types.Method (soon to be not-an-object) for synthetic wrappers. - CallCommon.Method now contains an abstract (interface) method object; was an abstract method index. - CallCommon.MethodId() gone. - Program.LookupMethod now takes a *Method not an Id string. R=gri CC=golang-dev https://golang.org/cl/11674043
2013-07-26 09:22:34 -06:00
}
}
go.tools/ssa: (another) major refactoring of method-set logic. We now use LookupFieldOrMethod for all SelectorExprs, and simplify the logic to discriminate the various cases. We inline static calls to promoted/indirected functions, dramatically reducing the number of functions created. More tests are needed, but I'd like to submit this as-is. In this CL, we: - rely less on Id strings. Internally we now use *types.Method (and its components) almost everywhere. - stop thinking of types.Methods as objects. They don't have stable identities. (Hopefully they will become plain-old structs soon.) - eliminate receiver indirection wrappers: indirection and promotion are handled together by makeWrapper. - Handle the interactions of promotion, indirection and abstract methods much more cleanly. - support receiver-bound interface method closures. - break up builder.selectField so we can re-use parts (emitFieldSelection). - add importer.PackageInfo.classifySelector utility. - delete interfaceMethodIndex() - delete namedTypeMethodIndex() - delete isSuperInterface() (replaced by types.IsAssignable) - call memberFromObject on each declared concrete method's *types.Func, not on every Method frem each method set, in the CREATE phase for packages loaded by gcimporter. go/types: - document Func, Signature.Recv() better. - use fmt in {Package,Label}.String - reimplement Func.String to be prettier and to include method receivers. API changes: - Function.method now holds the types.Method (soon to be not-an-object) for synthetic wrappers. - CallCommon.Method now contains an abstract (interface) method object; was an abstract method index. - CallCommon.MethodId() gone. - Program.LookupMethod now takes a *Method not an Id string. R=gri CC=golang-dev https://golang.org/cl/11674043
2013-07-26 09:22:34 -06:00
// Evaluate the function operand in the usual way.
c.Value = b.expr(fn, e.Fun)
}
// emitCallArgs emits to f code for the actual parameters of call e to
// a (possibly built-in) function of effective type sig.
// The argument values are appended to args, which is then returned.
//
func (b *builder) emitCallArgs(fn *Function, sig *types.Signature, e *ast.CallExpr, args []Value) []Value {
// f(x, y, z...): pass slice z straight through.
if e.Ellipsis != 0 {
for i, arg := range e.Args {
v := emitConv(fn, b.expr(fn, arg), sig.Params().At(i).Type())
args = append(args, v)
}
return args
}
offset := len(args) // 1 if call has receiver, 0 otherwise
// Evaluate actual parameter expressions.
//
// If this is a chained call of the form f(g()) where g has
// multiple return values (MRV), they are flattened out into
// args; a suffix of them may end up in a varargs slice.
for _, arg := range e.Args {
v := b.expr(fn, arg)
if ttuple, ok := v.Type().(*types.Tuple); ok { // MRV chain
for i, n := 0, ttuple.Len(); i < n; i++ {
args = append(args, emitExtract(fn, v, i, ttuple.At(i).Type()))
}
} else {
args = append(args, v)
}
}
// Actual->formal assignability conversions for normal parameters.
np := sig.Params().Len() // number of normal parameters
if sig.IsVariadic() {
np--
}
for i := 0; i < np; i++ {
args[offset+i] = emitConv(fn, args[offset+i], sig.Params().At(i).Type())
}
// Actual->formal assignability conversions for variadic parameter,
// and construction of slice.
if sig.IsVariadic() {
varargs := args[offset+np:]
st := sig.Params().At(np).Type().(*types.Slice)
vt := st.Elem()
if len(varargs) == 0 {
args = append(args, nilConst(st))
} else {
// Replace a suffix of args with a slice containing it.
at := types.NewArray(vt, int64(len(varargs)))
go.tools/ssa: (another) major refactoring of method-set logic. We now use LookupFieldOrMethod for all SelectorExprs, and simplify the logic to discriminate the various cases. We inline static calls to promoted/indirected functions, dramatically reducing the number of functions created. More tests are needed, but I'd like to submit this as-is. In this CL, we: - rely less on Id strings. Internally we now use *types.Method (and its components) almost everywhere. - stop thinking of types.Methods as objects. They don't have stable identities. (Hopefully they will become plain-old structs soon.) - eliminate receiver indirection wrappers: indirection and promotion are handled together by makeWrapper. - Handle the interactions of promotion, indirection and abstract methods much more cleanly. - support receiver-bound interface method closures. - break up builder.selectField so we can re-use parts (emitFieldSelection). - add importer.PackageInfo.classifySelector utility. - delete interfaceMethodIndex() - delete namedTypeMethodIndex() - delete isSuperInterface() (replaced by types.IsAssignable) - call memberFromObject on each declared concrete method's *types.Func, not on every Method frem each method set, in the CREATE phase for packages loaded by gcimporter. go/types: - document Func, Signature.Recv() better. - use fmt in {Package,Label}.String - reimplement Func.String to be prettier and to include method receivers. API changes: - Function.method now holds the types.Method (soon to be not-an-object) for synthetic wrappers. - CallCommon.Method now contains an abstract (interface) method object; was an abstract method index. - CallCommon.MethodId() gone. - Program.LookupMethod now takes a *Method not an Id string. R=gri CC=golang-dev https://golang.org/cl/11674043
2013-07-26 09:22:34 -06:00
// Don't set pos for implicit Allocs.
a := emitNew(fn, at, token.NoPos)
a.Comment = "varargs"
for i, arg := range varargs {
iaddr := &IndexAddr{
X: a,
Index: intConst(int64(i)),
}
iaddr.setType(types.NewPointer(vt))
fn.emit(iaddr)
emitStore(fn, iaddr, arg)
}
s := &Slice{X: a}
s.setType(st)
args[offset+np] = fn.emit(s)
args = args[:offset+np+1]
}
}
return args
}
// setCall emits to fn code to evaluate all the parameters of a function
// call e, and populates *c with those values.
//
func (b *builder) setCall(fn *Function, e *ast.CallExpr, c *CallCommon) {
// First deal with the f(...) part and optional receiver.
b.setCallFunc(fn, e, c)
// Then append the other actual parameters.
sig, _ := fn.Pkg.typeOf(e.Fun).Underlying().(*types.Signature)
if sig == nil {
panic(fmt.Sprintf("no signature for call of %s", e.Fun))
}
c.Args = b.emitCallArgs(fn, sig, e, c.Args)
}
// assignOp emits to fn code to perform loc += incr or loc -= incr.
func (b *builder) assignOp(fn *Function, loc lvalue, incr Value, op token.Token) {
oldv := loc.load(fn)
loc.store(fn, emitArith(fn, op, oldv, emitConv(fn, incr, oldv.Type()), loc.typ(), token.NoPos))
}
// buildGlobal emits code to the g.Pkg.init function for the variable
// definition(s) of g. Effects occur out of lexical order; see
// explanation at globalValueSpec.
go.tools/ssa: add debug information for all ast.Idents. This CL adds three new functions to determine the SSA Value for a given syntactic var, func or const object: Program.{Const,Func,Var}Value. Since constants and functions are immutable, the first two only need a types.Object; but each distinct reference to a var may return a distinct Value, so the third requires an ast.Ident parameter too. Debug information for local vars is encoded in the instruction stream in the form of DebugRef instructions, which are a no-op but relate their operand to a particular ident in the AST. The beauty of this approach is that it naturally stays consistent during optimisation passes (e.g. lifting) without additional bookkeeping. DebugRef instructions are only generated if the DebugMode builder flag is set; I plan to make the policy more fine- grained (per function). DebugRef instructions are inserted for: - expr(Ident) for rvalue idents - address.store() for idents that update an lvalue - address.address() for idents that take address of lvalue (this new method replaces all uses of lval.(address).addr) - expr() for all constant expressions - local ValueSpecs with implicit zero initialization (no RHS) (this case doesn't call store() or address()) To ensure we don't forget to emit debug info for uses of Idents, we must use the lvalue mechanism consistently. (Previously, many simple cases had effectively inlined these functions.) Similarly setCallFunc no longer inlines expr(Ident). Also: - Program.Value() has been inlined & specialized. - Program.Package() has moved nearer the new lookup functions. - refactoring: funcSyntax has lost paramFields, resultFields; gained funcType, which provides access to both. - add package-level constants to Package.values map. - opt: don't call localValueSpec for constants. (The resulting code is always optimised away.) There are a number of comments asking whether Literals should have positions. Will address in a follow-up. Added tests of all interesting cases. R=gri CC=golang-dev https://golang.org/cl/11259044
2013-07-15 11:56:46 -06:00
// Precondition: g == g.Prog.value(obj)
//
func (b *builder) buildGlobal(g *Global, obj types.Object) {
spec := g.spec
if spec == nil {
return // already built (or in progress)
}
b.globalValueSpec(g.Pkg.init, spec, g, obj)
}
// globalValueSpec emits to init code to define one or all of the vars
// in the package-level ValueSpec spec.
//
// It implements the build phase for a ValueSpec, ensuring that all
// vars are initialized if not already visited by buildGlobal during
// the reference graph traversal.
//
// This function may be called in two modes:
// A) with g and obj non-nil, to initialize just a single global.
// This occurs during the reference graph traversal.
// B) with g and obj nil, to initialize all globals in the same ValueSpec.
// This occurs during the left-to-right traversal over the ast.File.
//
go.tools/ssa: add debug information for all ast.Idents. This CL adds three new functions to determine the SSA Value for a given syntactic var, func or const object: Program.{Const,Func,Var}Value. Since constants and functions are immutable, the first two only need a types.Object; but each distinct reference to a var may return a distinct Value, so the third requires an ast.Ident parameter too. Debug information for local vars is encoded in the instruction stream in the form of DebugRef instructions, which are a no-op but relate their operand to a particular ident in the AST. The beauty of this approach is that it naturally stays consistent during optimisation passes (e.g. lifting) without additional bookkeeping. DebugRef instructions are only generated if the DebugMode builder flag is set; I plan to make the policy more fine- grained (per function). DebugRef instructions are inserted for: - expr(Ident) for rvalue idents - address.store() for idents that update an lvalue - address.address() for idents that take address of lvalue (this new method replaces all uses of lval.(address).addr) - expr() for all constant expressions - local ValueSpecs with implicit zero initialization (no RHS) (this case doesn't call store() or address()) To ensure we don't forget to emit debug info for uses of Idents, we must use the lvalue mechanism consistently. (Previously, many simple cases had effectively inlined these functions.) Similarly setCallFunc no longer inlines expr(Ident). Also: - Program.Value() has been inlined & specialized. - Program.Package() has moved nearer the new lookup functions. - refactoring: funcSyntax has lost paramFields, resultFields; gained funcType, which provides access to both. - add package-level constants to Package.values map. - opt: don't call localValueSpec for constants. (The resulting code is always optimised away.) There are a number of comments asking whether Literals should have positions. Will address in a follow-up. Added tests of all interesting cases. R=gri CC=golang-dev https://golang.org/cl/11259044
2013-07-15 11:56:46 -06:00
// Precondition: g == g.Prog.value(obj)
//
// Package-level var initialization order is quite subtle.
// The side effects of:
// var a, b = f(), g()
// are not observed left-to-right if b is referenced before a in the
// reference graph traversal. So, we track which Globals have been
// initialized by setting Global.spec=nil.
//
// Blank identifiers make things more complex since they don't have
// associated types.Objects or ssa.Globals yet we must still ensure
// that their corresponding side effects are observed at the right
// moment. Consider:
// var a, _, b = f(), g(), h()
// Here, the relative ordering of the call to g() is unspecified but
// it must occur exactly once, during mode B. So globalValueSpec for
// blanks must special-case n:n assigments and just evaluate the RHS
// g() for effect.
//
// In a n:1 assignment:
// var a, _, b = f()
// a reference to either a or b causes both globals to be initialized
// at the same time. Furthermore, no further work is required to
// ensure that the effects of the blank assignment occur. We must
// keep track of which n:1 specs have been evaluated, independent of
// which Globals are on the LHS (possibly none, if all are blank).
//
// See also localValueSpec.
//
func (b *builder) globalValueSpec(init *Function, spec *ast.ValueSpec, g *Global, obj types.Object) {
switch {
case len(spec.Values) == len(spec.Names):
// e.g. var x, y = 0, 1
// 1:1 assignment.
// Only the first time for a given GLOBAL has any effect.
for i, id := range spec.Names {
var lval lvalue = blank{}
if g != nil {
// Mode A: initialize only a single global, g
if isBlankIdent(id) || init.Pkg.objectOf(id) != obj {
continue
}
g.spec = nil
lval = &address{addr: g}
} else {
// Mode B: initialize all globals.
if !isBlankIdent(id) {
go.tools/ssa: add debug information for all ast.Idents. This CL adds three new functions to determine the SSA Value for a given syntactic var, func or const object: Program.{Const,Func,Var}Value. Since constants and functions are immutable, the first two only need a types.Object; but each distinct reference to a var may return a distinct Value, so the third requires an ast.Ident parameter too. Debug information for local vars is encoded in the instruction stream in the form of DebugRef instructions, which are a no-op but relate their operand to a particular ident in the AST. The beauty of this approach is that it naturally stays consistent during optimisation passes (e.g. lifting) without additional bookkeeping. DebugRef instructions are only generated if the DebugMode builder flag is set; I plan to make the policy more fine- grained (per function). DebugRef instructions are inserted for: - expr(Ident) for rvalue idents - address.store() for idents that update an lvalue - address.address() for idents that take address of lvalue (this new method replaces all uses of lval.(address).addr) - expr() for all constant expressions - local ValueSpecs with implicit zero initialization (no RHS) (this case doesn't call store() or address()) To ensure we don't forget to emit debug info for uses of Idents, we must use the lvalue mechanism consistently. (Previously, many simple cases had effectively inlined these functions.) Similarly setCallFunc no longer inlines expr(Ident). Also: - Program.Value() has been inlined & specialized. - Program.Package() has moved nearer the new lookup functions. - refactoring: funcSyntax has lost paramFields, resultFields; gained funcType, which provides access to both. - add package-level constants to Package.values map. - opt: don't call localValueSpec for constants. (The resulting code is always optimised away.) There are a number of comments asking whether Literals should have positions. Will address in a follow-up. Added tests of all interesting cases. R=gri CC=golang-dev https://golang.org/cl/11259044
2013-07-15 11:56:46 -06:00
g2 := init.Pkg.values[init.Pkg.objectOf(id)].(*Global)
if g2.spec == nil {
continue // already done
}
g2.spec = nil
lval = &address{addr: g2}
}
}
if init.Prog.mode&LogSource != 0 {
fmt.Fprintln(os.Stderr, "build global", id.Name)
}
b.exprInPlace(init, lval, spec.Values[i])
if g != nil {
break
}
}
case len(spec.Values) == 0:
// e.g. var x, y int
// Globals are implicitly zero-initialized.
default:
// e.g. var x, _, y = f()
// n:1 assignment.
// Only the first time for a given SPEC has any effect.
if !b.nTo1Vars[spec] {
b.nTo1Vars[spec] = true
if init.Prog.mode&LogSource != 0 {
defer logStack("build globals %s", spec.Names)()
}
tuple := b.exprN(init, spec.Values[0])
result := tuple.Type().(*types.Tuple)
for i, id := range spec.Names {
if !isBlankIdent(id) {
go.tools/ssa: add debug information for all ast.Idents. This CL adds three new functions to determine the SSA Value for a given syntactic var, func or const object: Program.{Const,Func,Var}Value. Since constants and functions are immutable, the first two only need a types.Object; but each distinct reference to a var may return a distinct Value, so the third requires an ast.Ident parameter too. Debug information for local vars is encoded in the instruction stream in the form of DebugRef instructions, which are a no-op but relate their operand to a particular ident in the AST. The beauty of this approach is that it naturally stays consistent during optimisation passes (e.g. lifting) without additional bookkeeping. DebugRef instructions are only generated if the DebugMode builder flag is set; I plan to make the policy more fine- grained (per function). DebugRef instructions are inserted for: - expr(Ident) for rvalue idents - address.store() for idents that update an lvalue - address.address() for idents that take address of lvalue (this new method replaces all uses of lval.(address).addr) - expr() for all constant expressions - local ValueSpecs with implicit zero initialization (no RHS) (this case doesn't call store() or address()) To ensure we don't forget to emit debug info for uses of Idents, we must use the lvalue mechanism consistently. (Previously, many simple cases had effectively inlined these functions.) Similarly setCallFunc no longer inlines expr(Ident). Also: - Program.Value() has been inlined & specialized. - Program.Package() has moved nearer the new lookup functions. - refactoring: funcSyntax has lost paramFields, resultFields; gained funcType, which provides access to both. - add package-level constants to Package.values map. - opt: don't call localValueSpec for constants. (The resulting code is always optimised away.) There are a number of comments asking whether Literals should have positions. Will address in a follow-up. Added tests of all interesting cases. R=gri CC=golang-dev https://golang.org/cl/11259044
2013-07-15 11:56:46 -06:00
g := init.Pkg.values[init.Pkg.objectOf(id)].(*Global)
g.spec = nil // just an optimization
emitStore(init, g, emitExtract(init, tuple, i, result.At(i).Type()))
}
}
}
}
}
// localValueSpec emits to fn code to define all of the vars in the
// function-local ValueSpec, spec.
//
// See also globalValueSpec: the two routines are similar but local
// ValueSpecs are much simpler since they are encountered once only,
// in their entirety, in lexical order.
//
func (b *builder) localValueSpec(fn *Function, spec *ast.ValueSpec) {
switch {
case len(spec.Values) == len(spec.Names):
// e.g. var x, y = 0, 1
// 1:1 assignment
for i, id := range spec.Names {
if !isBlankIdent(id) {
go.tools/ssa: add debug information for all ast.Idents. This CL adds three new functions to determine the SSA Value for a given syntactic var, func or const object: Program.{Const,Func,Var}Value. Since constants and functions are immutable, the first two only need a types.Object; but each distinct reference to a var may return a distinct Value, so the third requires an ast.Ident parameter too. Debug information for local vars is encoded in the instruction stream in the form of DebugRef instructions, which are a no-op but relate their operand to a particular ident in the AST. The beauty of this approach is that it naturally stays consistent during optimisation passes (e.g. lifting) without additional bookkeeping. DebugRef instructions are only generated if the DebugMode builder flag is set; I plan to make the policy more fine- grained (per function). DebugRef instructions are inserted for: - expr(Ident) for rvalue idents - address.store() for idents that update an lvalue - address.address() for idents that take address of lvalue (this new method replaces all uses of lval.(address).addr) - expr() for all constant expressions - local ValueSpecs with implicit zero initialization (no RHS) (this case doesn't call store() or address()) To ensure we don't forget to emit debug info for uses of Idents, we must use the lvalue mechanism consistently. (Previously, many simple cases had effectively inlined these functions.) Similarly setCallFunc no longer inlines expr(Ident). Also: - Program.Value() has been inlined & specialized. - Program.Package() has moved nearer the new lookup functions. - refactoring: funcSyntax has lost paramFields, resultFields; gained funcType, which provides access to both. - add package-level constants to Package.values map. - opt: don't call localValueSpec for constants. (The resulting code is always optimised away.) There are a number of comments asking whether Literals should have positions. Will address in a follow-up. Added tests of all interesting cases. R=gri CC=golang-dev https://golang.org/cl/11259044
2013-07-15 11:56:46 -06:00
fn.addLocalForIdent(id)
}
go.tools/ssa: add debug information for all ast.Idents. This CL adds three new functions to determine the SSA Value for a given syntactic var, func or const object: Program.{Const,Func,Var}Value. Since constants and functions are immutable, the first two only need a types.Object; but each distinct reference to a var may return a distinct Value, so the third requires an ast.Ident parameter too. Debug information for local vars is encoded in the instruction stream in the form of DebugRef instructions, which are a no-op but relate their operand to a particular ident in the AST. The beauty of this approach is that it naturally stays consistent during optimisation passes (e.g. lifting) without additional bookkeeping. DebugRef instructions are only generated if the DebugMode builder flag is set; I plan to make the policy more fine- grained (per function). DebugRef instructions are inserted for: - expr(Ident) for rvalue idents - address.store() for idents that update an lvalue - address.address() for idents that take address of lvalue (this new method replaces all uses of lval.(address).addr) - expr() for all constant expressions - local ValueSpecs with implicit zero initialization (no RHS) (this case doesn't call store() or address()) To ensure we don't forget to emit debug info for uses of Idents, we must use the lvalue mechanism consistently. (Previously, many simple cases had effectively inlined these functions.) Similarly setCallFunc no longer inlines expr(Ident). Also: - Program.Value() has been inlined & specialized. - Program.Package() has moved nearer the new lookup functions. - refactoring: funcSyntax has lost paramFields, resultFields; gained funcType, which provides access to both. - add package-level constants to Package.values map. - opt: don't call localValueSpec for constants. (The resulting code is always optimised away.) There are a number of comments asking whether Literals should have positions. Will address in a follow-up. Added tests of all interesting cases. R=gri CC=golang-dev https://golang.org/cl/11259044
2013-07-15 11:56:46 -06:00
lval := b.addr(fn, id, false) // non-escaping
b.exprInPlace(fn, lval, spec.Values[i])
}
case len(spec.Values) == 0:
// e.g. var x, y int
// Locals are implicitly zero-initialized.
for _, id := range spec.Names {
if !isBlankIdent(id) {
go.tools/ssa: add debug information for all ast.Idents. This CL adds three new functions to determine the SSA Value for a given syntactic var, func or const object: Program.{Const,Func,Var}Value. Since constants and functions are immutable, the first two only need a types.Object; but each distinct reference to a var may return a distinct Value, so the third requires an ast.Ident parameter too. Debug information for local vars is encoded in the instruction stream in the form of DebugRef instructions, which are a no-op but relate their operand to a particular ident in the AST. The beauty of this approach is that it naturally stays consistent during optimisation passes (e.g. lifting) without additional bookkeeping. DebugRef instructions are only generated if the DebugMode builder flag is set; I plan to make the policy more fine- grained (per function). DebugRef instructions are inserted for: - expr(Ident) for rvalue idents - address.store() for idents that update an lvalue - address.address() for idents that take address of lvalue (this new method replaces all uses of lval.(address).addr) - expr() for all constant expressions - local ValueSpecs with implicit zero initialization (no RHS) (this case doesn't call store() or address()) To ensure we don't forget to emit debug info for uses of Idents, we must use the lvalue mechanism consistently. (Previously, many simple cases had effectively inlined these functions.) Similarly setCallFunc no longer inlines expr(Ident). Also: - Program.Value() has been inlined & specialized. - Program.Package() has moved nearer the new lookup functions. - refactoring: funcSyntax has lost paramFields, resultFields; gained funcType, which provides access to both. - add package-level constants to Package.values map. - opt: don't call localValueSpec for constants. (The resulting code is always optimised away.) There are a number of comments asking whether Literals should have positions. Will address in a follow-up. Added tests of all interesting cases. R=gri CC=golang-dev https://golang.org/cl/11259044
2013-07-15 11:56:46 -06:00
lhs := fn.addLocalForIdent(id)
if fn.debugInfo() {
emitDebugRef(fn, id, emitLoad(fn, lhs))
}
}
}
default:
// e.g. var x, y = pos()
tuple := b.exprN(fn, spec.Values[0])
result := tuple.Type().(*types.Tuple)
for i, id := range spec.Names {
if !isBlankIdent(id) {
go.tools/ssa: add debug information for all ast.Idents. This CL adds three new functions to determine the SSA Value for a given syntactic var, func or const object: Program.{Const,Func,Var}Value. Since constants and functions are immutable, the first two only need a types.Object; but each distinct reference to a var may return a distinct Value, so the third requires an ast.Ident parameter too. Debug information for local vars is encoded in the instruction stream in the form of DebugRef instructions, which are a no-op but relate their operand to a particular ident in the AST. The beauty of this approach is that it naturally stays consistent during optimisation passes (e.g. lifting) without additional bookkeeping. DebugRef instructions are only generated if the DebugMode builder flag is set; I plan to make the policy more fine- grained (per function). DebugRef instructions are inserted for: - expr(Ident) for rvalue idents - address.store() for idents that update an lvalue - address.address() for idents that take address of lvalue (this new method replaces all uses of lval.(address).addr) - expr() for all constant expressions - local ValueSpecs with implicit zero initialization (no RHS) (this case doesn't call store() or address()) To ensure we don't forget to emit debug info for uses of Idents, we must use the lvalue mechanism consistently. (Previously, many simple cases had effectively inlined these functions.) Similarly setCallFunc no longer inlines expr(Ident). Also: - Program.Value() has been inlined & specialized. - Program.Package() has moved nearer the new lookup functions. - refactoring: funcSyntax has lost paramFields, resultFields; gained funcType, which provides access to both. - add package-level constants to Package.values map. - opt: don't call localValueSpec for constants. (The resulting code is always optimised away.) There are a number of comments asking whether Literals should have positions. Will address in a follow-up. Added tests of all interesting cases. R=gri CC=golang-dev https://golang.org/cl/11259044
2013-07-15 11:56:46 -06:00
fn.addLocalForIdent(id)
lhs := b.addr(fn, id, false) // non-escaping
lhs.store(fn, emitExtract(fn, tuple, i, result.At(i).Type()))
}
}
}
}
// assignStmt emits code to fn for a parallel assignment of rhss to lhss.
// isDef is true if this is a short variable declaration (:=).
//
// Note the similarity with localValueSpec.
//
func (b *builder) assignStmt(fn *Function, lhss, rhss []ast.Expr, isDef bool) {
// Side effects of all LHSs and RHSs must occur in left-to-right order.
var lvals []lvalue
for _, lhs := range lhss {
var lval lvalue = blank{}
if !isBlankIdent(lhs) {
if isDef {
// Local may be "redeclared" in the same
// scope, so don't blindly create anew.
obj := fn.Pkg.objectOf(lhs.(*ast.Ident))
if _, ok := fn.objects[obj]; !ok {
fn.addNamedLocal(obj)
}
}
lval = b.addr(fn, lhs, false) // non-escaping
}
lvals = append(lvals, lval)
}
if len(lhss) == len(rhss) {
// e.g. x, y = f(), g()
if len(lhss) == 1 {
// x = type{...}
// Optimization: in-place construction
// of composite literals.
b.exprInPlace(fn, lvals[0], rhss[0])
} else {
// Parallel assignment. All reads must occur
// before all updates, precluding exprInPlace.
// TODO(adonovan): opt: is it sound to
// perform exprInPlace if !isDef?
var rvals []Value
for _, rval := range rhss {
rvals = append(rvals, b.expr(fn, rval))
}
for i, lval := range lvals {
lval.store(fn, rvals[i])
}
}
} else {
// e.g. x, y = pos()
tuple := b.exprN(fn, rhss[0])
result := tuple.Type().(*types.Tuple)
for i, lval := range lvals {
lval.store(fn, emitExtract(fn, tuple, i, result.At(i).Type()))
}
}
}
// arrayLen returns the length of the array whose composite literal elements are elts.
func (b *builder) arrayLen(fn *Function, elts []ast.Expr) int64 {
var max int64 = -1
var i int64 = -1
for _, e := range elts {
if kv, ok := e.(*ast.KeyValueExpr); ok {
i = b.expr(fn, kv.Key).(*Const).Int64()
} else {
i++
}
if i > max {
max = i
}
}
return max + 1
}
// compLit emits to fn code to initialize a composite literal e at
// address addr with type typ, typically allocated by Alloc.
// Nested composite literals are recursively initialized in place
// where possible.
//
func (b *builder) compLit(fn *Function, addr Value, e *ast.CompositeLit, typ types.Type) {
// TODO(adonovan): document how and why typ ever differs from
// fn.Pkg.typeOf(e).
switch t := typ.Underlying().(type) {
case *types.Struct:
for i, e := range e.Elts {
fieldIndex := i
if kv, ok := e.(*ast.KeyValueExpr); ok {
fname := kv.Key.(*ast.Ident).Name
for i, n := 0, t.NumFields(); i < n; i++ {
sf := t.Field(i)
if sf.Name() == fname {
fieldIndex = i
e = kv.Value
break
}
}
}
sf := t.Field(fieldIndex)
faddr := &FieldAddr{
X: addr,
Field: fieldIndex,
}
faddr.setType(types.NewPointer(sf.Type()))
fn.emit(faddr)
b.exprInPlace(fn, &address{addr: faddr, expr: e}, e)
}
case *types.Array, *types.Slice:
var at *types.Array
var array Value
switch t := t.(type) {
case *types.Slice:
at = types.NewArray(t.Elem(), b.arrayLen(fn, e.Elts))
alloc := emitNew(fn, at, e.Lbrace)
alloc.Comment = "slicelit"
array = alloc
case *types.Array:
at = t
array = addr
}
var idx *Const
for _, e := range e.Elts {
if kv, ok := e.(*ast.KeyValueExpr); ok {
idx = b.expr(fn, kv.Key).(*Const)
e = kv.Value
} else {
var idxval int64
if idx != nil {
idxval = idx.Int64() + 1
}
idx = intConst(idxval)
}
iaddr := &IndexAddr{
X: array,
Index: idx,
}
iaddr.setType(types.NewPointer(at.Elem()))
fn.emit(iaddr)
b.exprInPlace(fn, &address{addr: iaddr, expr: e}, e)
}
if t != at { // slice
s := &Slice{X: array}
s.setPos(e.Lbrace)
s.setType(t)
emitStore(fn, addr, fn.emit(s))
}
case *types.Map:
m := &MakeMap{Reserve: intConst(int64(len(e.Elts)))}
m.setPos(e.Lbrace)
m.setType(typ)
emitStore(fn, addr, fn.emit(m))
for _, e := range e.Elts {
e := e.(*ast.KeyValueExpr)
loc := &element{
m: m,
k: emitConv(fn, b.expr(fn, e.Key), t.Key()),
t: t.Elem(),
pos: e.Colon,
}
b.exprInPlace(fn, loc, e.Value)
}
case *types.Pointer:
// Pointers can only occur in the recursive case; we
// strip them off in addr() before calling compLit
// again, so that we allocate space for a T not a *T.
panic("compLit(fn, addr, e, *types.Pointer")
default:
panic("unexpected CompositeLit type: " + t.String())
}
}
// switchStmt emits to fn code for the switch statement s, optionally
// labelled by label.
//
func (b *builder) switchStmt(fn *Function, s *ast.SwitchStmt, label *lblock) {
// We treat SwitchStmt like a sequential if-else chain.
// More efficient strategies (e.g. multiway dispatch)
// are possible if all cases are free of side effects.
if s.Init != nil {
b.stmt(fn, s.Init)
}
var tag Value = vTrue
if s.Tag != nil {
tag = b.expr(fn, s.Tag)
}
done := fn.newBasicBlock("switch.done")
if label != nil {
label._break = done
}
// We pull the default case (if present) down to the end.
// But each fallthrough label must point to the next
// body block in source order, so we preallocate a
// body block (fallthru) for the next case.
// Unfortunately this makes for a confusing block order.
var dfltBody *[]ast.Stmt
var dfltFallthrough *BasicBlock
var fallthru, dfltBlock *BasicBlock
ncases := len(s.Body.List)
for i, clause := range s.Body.List {
body := fallthru
if body == nil {
body = fn.newBasicBlock("switch.body") // first case only
}
// Preallocate body block for the next case.
fallthru = done
if i+1 < ncases {
fallthru = fn.newBasicBlock("switch.body")
}
cc := clause.(*ast.CaseClause)
if cc.List == nil {
// Default case.
dfltBody = &cc.Body
dfltFallthrough = fallthru
dfltBlock = body
continue
}
var nextCond *BasicBlock
for _, cond := range cc.List {
nextCond = fn.newBasicBlock("switch.next")
// TODO(adonovan): opt: when tag==vTrue, we'd
// get better much code if we use b.cond(cond)
// instead of BinOp(EQL, tag, b.expr(cond))
// followed by If. Don't forget conversions
// though.
cond := emitCompare(fn, token.EQL, tag, b.expr(fn, cond), token.NoPos)
emitIf(fn, cond, body, nextCond)
fn.currentBlock = nextCond
}
fn.currentBlock = body
fn.targets = &targets{
tail: fn.targets,
_break: done,
_fallthrough: fallthru,
}
b.stmtList(fn, cc.Body)
fn.targets = fn.targets.tail
emitJump(fn, done)
fn.currentBlock = nextCond
}
if dfltBlock != nil {
emitJump(fn, dfltBlock)
fn.currentBlock = dfltBlock
fn.targets = &targets{
tail: fn.targets,
_break: done,
_fallthrough: dfltFallthrough,
}
b.stmtList(fn, *dfltBody)
fn.targets = fn.targets.tail
}
emitJump(fn, done)
fn.currentBlock = done
}
// typeSwitchStmt emits to fn code for the type switch statement s, optionally
// labelled by label.
//
func (b *builder) typeSwitchStmt(fn *Function, s *ast.TypeSwitchStmt, label *lblock) {
// We treat TypeSwitchStmt like a sequential if-else
// chain. More efficient strategies (e.g. multiway
// dispatch) are possible.
// Typeswitch lowering:
//
// var x X
// switch y := x.(type) {
// case T1, T2: S1 // >1 (y := x)
// case nil: SN // nil (y := x)
// default: SD // 0 types (y := x)
// case T3: S3 // 1 type (y := x.(T3))
// }
//
// ...s.Init...
// x := eval x
// .caseT1:
// t1, ok1 := typeswitch,ok x <T1>
// if ok1 then goto S1 else goto .caseT2
// .caseT2:
// t2, ok2 := typeswitch,ok x <T2>
// if ok2 then goto S1 else goto .caseNil
// .S1:
// y := x
// ...S1...
// goto done
// .caseNil:
// if t2, ok2 := typeswitch,ok x <T2>
// if x == nil then goto SN else goto .caseT3
// .SN:
// y := x
// ...SN...
// goto done
// .caseT3:
// t3, ok3 := typeswitch,ok x <T3>
// if ok3 then goto S3 else goto default
// .S3:
// y := t3
// ...S3...
// goto done
// .default:
// y := x
// ...SD...
// goto done
// .done:
if s.Init != nil {
b.stmt(fn, s.Init)
}
var x Value
switch ass := s.Assign.(type) {
case *ast.ExprStmt: // x.(type)
x = b.expr(fn, unparen(ass.X).(*ast.TypeAssertExpr).X)
case *ast.AssignStmt: // y := x.(type)
x = b.expr(fn, unparen(ass.Rhs[0]).(*ast.TypeAssertExpr).X)
}
done := fn.newBasicBlock("typeswitch.done")
if label != nil {
label._break = done
}
var default_ *ast.CaseClause
for _, clause := range s.Body.List {
cc := clause.(*ast.CaseClause)
if cc.List == nil {
default_ = cc
continue
}
body := fn.newBasicBlock("typeswitch.body")
var next *BasicBlock
var casetype types.Type
var ti Value // ti, ok := typeassert,ok x <Ti>
for _, cond := range cc.List {
next = fn.newBasicBlock("typeswitch.next")
casetype = fn.Pkg.typeOf(cond)
var condv Value
if casetype == tUntypedNil {
condv = emitCompare(fn, token.EQL, x, nilConst(x.Type()), token.NoPos)
ti = x
} else {
yok := emitTypeTest(fn, x, casetype, cc.Case)
ti = emitExtract(fn, yok, 0, casetype)
condv = emitExtract(fn, yok, 1, tBool)
}
emitIf(fn, condv, body, next)
fn.currentBlock = next
}
if len(cc.List) != 1 {
ti = x
}
fn.currentBlock = body
b.typeCaseBody(fn, cc, ti, done)
fn.currentBlock = next
}
if default_ != nil {
b.typeCaseBody(fn, default_, x, done)
} else {
emitJump(fn, done)
}
fn.currentBlock = done
}
func (b *builder) typeCaseBody(fn *Function, cc *ast.CaseClause, x Value, done *BasicBlock) {
if obj := fn.Pkg.info.TypeCaseVar(cc); obj != nil {
// In a switch y := x.(type), each case clause
// implicitly declares a distinct object y.
// In a single-type case, y has that type.
// In multi-type cases, 'case nil' and default,
// y has the same type as the interface operand.
emitStore(fn, fn.addNamedLocal(obj), x)
}
fn.targets = &targets{
tail: fn.targets,
_break: done,
}
b.stmtList(fn, cc.Body)
fn.targets = fn.targets.tail
emitJump(fn, done)
}
// selectStmt emits to fn code for the select statement s, optionally
// labelled by label.
//
func (b *builder) selectStmt(fn *Function, s *ast.SelectStmt, label *lblock) {
// A blocking select of a single case degenerates to a
// simple send or receive.
// TODO(adonovan): opt: is this optimization worth its weight?
if len(s.Body.List) == 1 {
clause := s.Body.List[0].(*ast.CommClause)
if clause.Comm != nil {
b.stmt(fn, clause.Comm)
done := fn.newBasicBlock("select.done")
if label != nil {
label._break = done
}
fn.targets = &targets{
tail: fn.targets,
_break: done,
}
b.stmtList(fn, clause.Body)
fn.targets = fn.targets.tail
emitJump(fn, done)
fn.currentBlock = done
return
}
}
// First evaluate all channels in all cases, and find
// the directions of each state.
var states []*SelectState
blocking := true
debugInfo := fn.debugInfo()
for _, clause := range s.Body.List {
var st *SelectState
switch comm := clause.(*ast.CommClause).Comm.(type) {
case nil: // default case
blocking = false
continue
case *ast.SendStmt: // ch<- i
ch := b.expr(fn, comm.Chan)
st = &SelectState{
Dir: ast.SEND,
Chan: ch,
Send: emitConv(fn, b.expr(fn, comm.Value),
ch.Type().Underlying().(*types.Chan).Elem()),
Pos: comm.Arrow,
}
if debugInfo {
st.DebugNode = comm
}
case *ast.AssignStmt: // x := <-ch
recv := unparen(comm.Rhs[0]).(*ast.UnaryExpr)
st = &SelectState{
Dir: ast.RECV,
Chan: b.expr(fn, recv.X),
Pos: recv.OpPos,
}
if debugInfo {
st.DebugNode = recv
}
case *ast.ExprStmt: // <-ch
recv := unparen(comm.X).(*ast.UnaryExpr)
st = &SelectState{
Dir: ast.RECV,
Chan: b.expr(fn, recv.X),
Pos: recv.OpPos,
}
if debugInfo {
st.DebugNode = recv
}
}
states = append(states, st)
}
// We dispatch on the (fair) result of Select using a
// sequential if-else chain, in effect:
//
// idx, recvOk, r0...r_n-1 := select(...)
// if idx == 0 { // receive on channel 0 (first receive => r0)
// x, ok := r0, recvOk
// ...state0...
// } else if v == 1 { // send on channel 1
// ...state1...
// } else {
// ...default...
// }
sel := &Select{
States: states,
Blocking: blocking,
}
sel.setPos(s.Select)
var vars []*types.Var
vars = append(vars, varIndex, varOk)
for _, st := range states {
if st.Dir == ast.RECV {
tElem := st.Chan.Type().Underlying().(*types.Chan).Elem()
vars = append(vars, types.NewVar(token.NoPos, nil, "", tElem))
}
}
sel.setType(types.NewTuple(vars...))
fn.emit(sel)
idx := emitExtract(fn, sel, 0, tInt)
done := fn.newBasicBlock("select.done")
if label != nil {
label._break = done
}
var defaultBody *[]ast.Stmt
state := 0
r := 2 // index in 'sel' tuple of value; increments if st.Dir==RECV
for _, cc := range s.Body.List {
clause := cc.(*ast.CommClause)
if clause.Comm == nil {
defaultBody = &clause.Body
continue
}
body := fn.newBasicBlock("select.body")
next := fn.newBasicBlock("select.next")
emitIf(fn, emitCompare(fn, token.EQL, idx, intConst(int64(state)), token.NoPos), body, next)
fn.currentBlock = body
fn.targets = &targets{
tail: fn.targets,
_break: done,
}
switch comm := clause.Comm.(type) {
case *ast.ExprStmt: // <-ch
if debugInfo {
v := emitExtract(fn, sel, r, vars[r].Type())
emitDebugRef(fn, states[state].DebugNode.(ast.Expr), v)
}
r++
case *ast.AssignStmt: // x := <-states[state].Chan
if comm.Tok == token.DEFINE {
fn.addLocalForIdent(comm.Lhs[0].(*ast.Ident))
}
x := b.addr(fn, comm.Lhs[0], false) // non-escaping
v := emitExtract(fn, sel, r, vars[r].Type())
if debugInfo {
emitDebugRef(fn, states[state].DebugNode.(ast.Expr), v)
}
x.store(fn, v)
if len(comm.Lhs) == 2 { // x, ok := ...
if comm.Tok == token.DEFINE {
fn.addLocalForIdent(comm.Lhs[1].(*ast.Ident))
}
ok := b.addr(fn, comm.Lhs[1], false) // non-escaping
ok.store(fn, emitExtract(fn, sel, 1, deref(ok.typ())))
}
r++
}
b.stmtList(fn, clause.Body)
fn.targets = fn.targets.tail
emitJump(fn, done)
fn.currentBlock = next
state++
}
if defaultBody != nil {
fn.targets = &targets{
tail: fn.targets,
_break: done,
}
b.stmtList(fn, *defaultBody)
fn.targets = fn.targets.tail
}
emitJump(fn, done)
fn.currentBlock = done
}
// forStmt emits to fn code for the for statement s, optionally
// labelled by label.
//
func (b *builder) forStmt(fn *Function, s *ast.ForStmt, label *lblock) {
// ...init...
// jump loop
// loop:
// if cond goto body else done
// body:
// ...body...
// jump post
// post: (target of continue)
// ...post...
// jump loop
// done: (target of break)
if s.Init != nil {
b.stmt(fn, s.Init)
}
body := fn.newBasicBlock("for.body")
done := fn.newBasicBlock("for.done") // target of 'break'
loop := body // target of back-edge
if s.Cond != nil {
loop = fn.newBasicBlock("for.loop")
}
cont := loop // target of 'continue'
if s.Post != nil {
cont = fn.newBasicBlock("for.post")
}
if label != nil {
label._break = done
label._continue = cont
}
emitJump(fn, loop)
fn.currentBlock = loop
if loop != body {
b.cond(fn, s.Cond, body, done)
fn.currentBlock = body
}
fn.targets = &targets{
tail: fn.targets,
_break: done,
_continue: cont,
}
b.stmt(fn, s.Body)
fn.targets = fn.targets.tail
emitJump(fn, cont)
if s.Post != nil {
fn.currentBlock = cont
b.stmt(fn, s.Post)
emitJump(fn, loop) // back-edge
}
fn.currentBlock = done
}
// rangeIndexed emits to fn the header for an integer indexed loop
// over array, *array or slice value x.
// The v result is defined only if tv is non-nil.
//
func (b *builder) rangeIndexed(fn *Function, x Value, tv types.Type) (k, v Value, loop, done *BasicBlock) {
//
// length = len(x)
// index = -1
// loop: (target of continue)
// index++
// if index < length goto body else done
// body:
// k = index
// v = x[index]
// ...body...
// jump loop
// done: (target of break)
// Determine number of iterations.
var length Value
if arr, ok := deref(x.Type()).Underlying().(*types.Array); ok {
// For array or *array, the number of iterations is
// known statically thanks to the type. We avoid a
// data dependence upon x, permitting later dead-code
// elimination if x is pure, static unrolling, etc.
// Ranging over a nil *array may have >0 iterations.
length = intConst(arr.Len())
} else {
// length = len(x).
var c Call
c.Call.Value = fn.Prog.builtins[types.Universe.Lookup("len").(*types.Builtin)]
c.Call.Args = []Value{x}
c.setType(tInt)
length = fn.emit(&c)
}
index := fn.addLocal(tInt, token.NoPos)
emitStore(fn, index, intConst(-1))
loop = fn.newBasicBlock("rangeindex.loop")
emitJump(fn, loop)
fn.currentBlock = loop
incr := &BinOp{
Op: token.ADD,
X: emitLoad(fn, index),
Y: vOne,
}
incr.setType(tInt)
emitStore(fn, index, fn.emit(incr))
body := fn.newBasicBlock("rangeindex.body")
done = fn.newBasicBlock("rangeindex.done")
emitIf(fn, emitCompare(fn, token.LSS, incr, length, token.NoPos), body, done)
fn.currentBlock = body
k = emitLoad(fn, index)
if tv != nil {
switch t := x.Type().Underlying().(type) {
case *types.Array:
instr := &Index{
X: x,
Index: k,
}
instr.setType(t.Elem())
v = fn.emit(instr)
case *types.Pointer: // *array
instr := &IndexAddr{
X: x,
Index: k,
}
instr.setType(types.NewPointer(t.Elem().(*types.Array).Elem()))
v = emitLoad(fn, fn.emit(instr))
case *types.Slice:
instr := &IndexAddr{
X: x,
Index: k,
}
instr.setType(types.NewPointer(t.Elem()))
v = emitLoad(fn, fn.emit(instr))
default:
panic("rangeIndexed x:" + t.String())
}
}
return
}
// rangeIter emits to fn the header for a loop using
// Range/Next/Extract to iterate over map or string value x.
// tk and tv are the types of the key/value results k and v, or nil
// if the respective component is not wanted.
//
func (b *builder) rangeIter(fn *Function, x Value, tk, tv types.Type, pos token.Pos) (k, v Value, loop, done *BasicBlock) {
//
// it = range x
// loop: (target of continue)
// okv = next it (ok, key, value)
// ok = extract okv #0
// if ok goto body else done
// body:
// k = extract okv #1
// v = extract okv #2
// ...body...
// jump loop
// done: (target of break)
//
if tk == nil {
tk = tInvalid
}
if tv == nil {
tv = tInvalid
}
rng := &Range{X: x}
rng.setPos(pos)
rng.setType(tRangeIter)
it := fn.emit(rng)
loop = fn.newBasicBlock("rangeiter.loop")
emitJump(fn, loop)
fn.currentBlock = loop
_, isString := x.Type().Underlying().(*types.Basic)
okv := &Next{
Iter: it,
IsString: isString,
}
okv.setType(types.NewTuple(
varOk,
types.NewVar(token.NoPos, nil, "k", tk),
types.NewVar(token.NoPos, nil, "v", tv),
))
fn.emit(okv)
body := fn.newBasicBlock("rangeiter.body")
done = fn.newBasicBlock("rangeiter.done")
emitIf(fn, emitExtract(fn, okv, 0, tBool), body, done)
fn.currentBlock = body
if tk != tInvalid {
k = emitExtract(fn, okv, 1, tk)
}
if tv != tInvalid {
v = emitExtract(fn, okv, 2, tv)
}
return
}
// rangeChan emits to fn the header for a loop that receives from
// channel x until it fails.
// tk is the channel's element type, or nil if the k result is
// not wanted
// pos is the position of the '=' or ':=' token.
//
func (b *builder) rangeChan(fn *Function, x Value, tk types.Type, pos token.Pos) (k Value, loop, done *BasicBlock) {
//
// loop: (target of continue)
// ko = <-x (key, ok)
// ok = extract ko #1
// if ok goto body else done
// body:
// k = extract ko #0
// ...
// goto loop
// done: (target of break)
loop = fn.newBasicBlock("rangechan.loop")
emitJump(fn, loop)
fn.currentBlock = loop
recv := &UnOp{
Op: token.ARROW,
X: x,
CommaOk: true,
}
recv.setPos(pos)
recv.setType(types.NewTuple(
types.NewVar(token.NoPos, nil, "k", x.Type().Underlying().(*types.Chan).Elem()),
varOk,
))
ko := fn.emit(recv)
body := fn.newBasicBlock("rangechan.body")
done = fn.newBasicBlock("rangechan.done")
emitIf(fn, emitExtract(fn, ko, 1, tBool), body, done)
fn.currentBlock = body
if tk != nil {
k = emitExtract(fn, ko, 0, tk)
}
return
}
// rangeStmt emits to fn code for the range statement s, optionally
// labelled by label.
//
func (b *builder) rangeStmt(fn *Function, s *ast.RangeStmt, label *lblock) {
var tk, tv types.Type
if !isBlankIdent(s.Key) {
tk = fn.Pkg.typeOf(s.Key)
}
if s.Value != nil && !isBlankIdent(s.Value) {
tv = fn.Pkg.typeOf(s.Value)
}
// If iteration variables are defined (:=), this
// occurs once outside the loop.
//
// Unlike a short variable declaration, a RangeStmt
// using := never redeclares an existing variable; it
// always creates a new one.
if s.Tok == token.DEFINE {
if tk != nil {
fn.addLocalForIdent(s.Key.(*ast.Ident))
}
if tv != nil {
fn.addLocalForIdent(s.Value.(*ast.Ident))
}
}
x := b.expr(fn, s.X)
var k, v Value
var loop, done *BasicBlock
switch rt := x.Type().Underlying().(type) {
case *types.Slice, *types.Array, *types.Pointer: // *array
k, v, loop, done = b.rangeIndexed(fn, x, tv)
case *types.Chan:
k, loop, done = b.rangeChan(fn, x, tk, s.TokPos)
case *types.Map, *types.Basic: // string
k, v, loop, done = b.rangeIter(fn, x, tk, tv, s.For)
default:
panic("Cannot range over: " + rt.String())
}
// Evaluate both LHS expressions before we update either.
var kl, vl lvalue
if tk != nil {
kl = b.addr(fn, s.Key, false) // non-escaping
}
if tv != nil {
vl = b.addr(fn, s.Value, false) // non-escaping
}
if tk != nil {
kl.store(fn, k)
}
if tv != nil {
vl.store(fn, v)
}
if label != nil {
label._break = done
label._continue = loop
}
fn.targets = &targets{
tail: fn.targets,
_break: done,
_continue: loop,
}
b.stmt(fn, s.Body)
fn.targets = fn.targets.tail
emitJump(fn, loop) // back-edge
fn.currentBlock = done
}
// stmt lowers statement s to SSA form, emitting code to fn.
func (b *builder) stmt(fn *Function, _s ast.Stmt) {
// The label of the current statement. If non-nil, its _goto
// target is always set; its _break and _continue are set only
// within the body of switch/typeswitch/select/for/range.
// It is effectively an additional default-nil parameter of stmt().
var label *lblock
start:
switch s := _s.(type) {
case *ast.EmptyStmt:
// ignore. (Usually removed by gofmt.)
case *ast.DeclStmt: // Con, Var or Typ
d := s.Decl.(*ast.GenDecl)
go.tools/ssa: add debug information for all ast.Idents. This CL adds three new functions to determine the SSA Value for a given syntactic var, func or const object: Program.{Const,Func,Var}Value. Since constants and functions are immutable, the first two only need a types.Object; but each distinct reference to a var may return a distinct Value, so the third requires an ast.Ident parameter too. Debug information for local vars is encoded in the instruction stream in the form of DebugRef instructions, which are a no-op but relate their operand to a particular ident in the AST. The beauty of this approach is that it naturally stays consistent during optimisation passes (e.g. lifting) without additional bookkeeping. DebugRef instructions are only generated if the DebugMode builder flag is set; I plan to make the policy more fine- grained (per function). DebugRef instructions are inserted for: - expr(Ident) for rvalue idents - address.store() for idents that update an lvalue - address.address() for idents that take address of lvalue (this new method replaces all uses of lval.(address).addr) - expr() for all constant expressions - local ValueSpecs with implicit zero initialization (no RHS) (this case doesn't call store() or address()) To ensure we don't forget to emit debug info for uses of Idents, we must use the lvalue mechanism consistently. (Previously, many simple cases had effectively inlined these functions.) Similarly setCallFunc no longer inlines expr(Ident). Also: - Program.Value() has been inlined & specialized. - Program.Package() has moved nearer the new lookup functions. - refactoring: funcSyntax has lost paramFields, resultFields; gained funcType, which provides access to both. - add package-level constants to Package.values map. - opt: don't call localValueSpec for constants. (The resulting code is always optimised away.) There are a number of comments asking whether Literals should have positions. Will address in a follow-up. Added tests of all interesting cases. R=gri CC=golang-dev https://golang.org/cl/11259044
2013-07-15 11:56:46 -06:00
if d.Tok == token.VAR {
for _, spec := range d.Specs {
if vs, ok := spec.(*ast.ValueSpec); ok {
b.localValueSpec(fn, vs)
}
}
}
case *ast.LabeledStmt:
label = fn.labelledBlock(s.Label)
emitJump(fn, label._goto)
fn.currentBlock = label._goto
_s = s.Stmt
goto start // effectively: tailcall stmt(fn, s.Stmt, label)
case *ast.ExprStmt:
b.expr(fn, s.X)
case *ast.SendStmt:
fn.emit(&Send{
Chan: b.expr(fn, s.Chan),
X: emitConv(fn, b.expr(fn, s.Value),
fn.Pkg.typeOf(s.Chan).Underlying().(*types.Chan).Elem()),
pos: s.Arrow,
})
case *ast.IncDecStmt:
op := token.ADD
if s.Tok == token.DEC {
op = token.SUB
}
b.assignOp(fn, b.addr(fn, s.X, false), vOne, op)
case *ast.AssignStmt:
switch s.Tok {
case token.ASSIGN, token.DEFINE:
b.assignStmt(fn, s.Lhs, s.Rhs, s.Tok == token.DEFINE)
default: // +=, etc.
op := s.Tok + token.ADD - token.ADD_ASSIGN
b.assignOp(fn, b.addr(fn, s.Lhs[0], false), b.expr(fn, s.Rhs[0]), op)
}
case *ast.GoStmt:
// The "intrinsics" new/make/len/cap are forbidden here.
// panic is treated like an ordinary function call.
v := Go{pos: s.Go}
b.setCall(fn, s.Call, &v.Call)
fn.emit(&v)
case *ast.DeferStmt:
// The "intrinsics" new/make/len/cap are forbidden here.
// panic is treated like an ordinary function call.
v := Defer{pos: s.Defer}
b.setCall(fn, s.Call, &v.Call)
fn.emit(&v)
go.tools/ssa: implement correct control flow for recovered panic. A function such as this: func one() (x int) { defer func() { recover() }() x = 1 panic("return") } that combines named return parameters (NRPs) with deferred calls that call recover, may return non-zero values despite the fact it doesn't even contain a return statement. (!) This requires a change to the SSA API: all functions' control-flow graphs now have a second entry point, called Recover, which is the block at which control flow resumes after a recovered panic. The Recover block simply loads the NRPs and returns them. As an optimization, most functions don't need a Recover block, so it is omitted. In fact it is only needed for functions that have NRPs and defer a call to another function that _may_ call recover. Dataflow analysis of SSA now requires extra work, since every may-panic instruction has an implicit control-flow edge to the Recover block. The only dataflow analysis so far implemented is SSA renaming, for which we make the following simplifying assumption: the Recover block only loads the NRPs and returns. This means we don't really need to analyze it, we can just skip the "lifting" of such NRPs. We also special-case the Recover block in the dominance computation. Rejected alternative approaches: - Specifying a Recover block for every defer instruction (like a traditional exception handler). This seemed like excessive generality, since Go programs only need the same degenerate form of Recover block. - Adding an instruction to set the Recover block immediately after the named return values are set up, so that dominance can be computed without special-casing. This didn't seem worth the effort. Interpreter: - This CL completely reimplements the panic/recover/ defer logic in the interpreter. It's clearer and simpler and closer to the model in the spec. - Some runtime panic messages have been changed to be closer to gc's, since tests depend on it. - The interpreter now requires that the runtime.runtimeError type be part of the SSA program. This requires that clients import this package prior to invoking the interpreter. This in turn requires (Importer).ImportPackage(path string), which this CL adds. - All $GOROOT/test/recover{,1,2,3}.go tests are now passing. NB, the bug described in coverage.go (defer/recover in a concatenated init function) remains. Will be fixed in a follow-up. Fixes golang/go#6381 R=gri CC=crawshaw, golang-dev https://golang.org/cl/13844043
2013-10-14 13:38:56 -06:00
// A deferred call can cause recovery from panic.
// If the panicking function has named results,
// control resumes at the Recover block to load those
// locals (which may be mutated by the deferred call)
// and return them.
if fn.namedResults != nil {
// Optimization: if we can prove the deferred call
// won't cause recovery from panic, we can avoid a
// Recover block.
// We scan the callee for calls to recover() iff:
// - it's a static call
// - to a function in the same package
// (other packages' SSA building happens concurrently)
// - whose SSA building has started (Blocks != nil)
// - and finished (i.e. not this function)
// NB, this is always true for: defer func() { ... } ()
//
// TODO(adonovan): optimize interpackage cases, e.g.
// (sync.Mutex).Unlock(), (io.Closer).Close
if callee, ok := v.Call.Value.(*Function); ok && callee.Pkg == fn.Pkg && callee != fn && callee.Blocks != nil && !callsRecover(callee) {
// Deferred call cannot cause recovery from panic.
} else {
createRecoverBlock(fn)
}
}
case *ast.ReturnStmt:
var results []Value
if len(s.Results) == 1 && fn.Signature.Results().Len() > 1 {
// Return of one expression in a multi-valued function.
tuple := b.exprN(fn, s.Results[0])
ttuple := tuple.Type().(*types.Tuple)
for i, n := 0, ttuple.Len(); i < n; i++ {
results = append(results,
emitConv(fn, emitExtract(fn, tuple, i, ttuple.At(i).Type()),
fn.Signature.Results().At(i).Type()))
}
} else {
// 1:1 return, or no-arg return in non-void function.
for i, r := range s.Results {
v := emitConv(fn, b.expr(fn, r), fn.Signature.Results().At(i).Type())
results = append(results, v)
}
}
if fn.namedResults != nil {
// Function has named result parameters (NRPs).
// Perform parallel assignment of return operands to NRPs.
for i, r := range results {
emitStore(fn, fn.namedResults[i], r)
}
}
// Run function calls deferred in this
// function when explicitly returning from it.
fn.emit(new(RunDefers))
if fn.namedResults != nil {
// Reload NRPs to form the result tuple.
results = results[:0]
for _, r := range fn.namedResults {
results = append(results, emitLoad(fn, r))
}
}
fn.emit(&Return{Results: results, pos: s.Return})
fn.currentBlock = fn.newBasicBlock("unreachable")
case *ast.BranchStmt:
var block *BasicBlock
switch s.Tok {
case token.BREAK:
if s.Label != nil {
block = fn.labelledBlock(s.Label)._break
} else {
for t := fn.targets; t != nil && block == nil; t = t.tail {
block = t._break
}
}
case token.CONTINUE:
if s.Label != nil {
block = fn.labelledBlock(s.Label)._continue
} else {
for t := fn.targets; t != nil && block == nil; t = t.tail {
block = t._continue
}
}
case token.FALLTHROUGH:
for t := fn.targets; t != nil && block == nil; t = t.tail {
block = t._fallthrough
}
case token.GOTO:
block = fn.labelledBlock(s.Label)._goto
}
emitJump(fn, block)
fn.currentBlock = fn.newBasicBlock("unreachable")
case *ast.BlockStmt:
b.stmtList(fn, s.List)
case *ast.IfStmt:
if s.Init != nil {
b.stmt(fn, s.Init)
}
then := fn.newBasicBlock("if.then")
done := fn.newBasicBlock("if.done")
els := done
if s.Else != nil {
els = fn.newBasicBlock("if.else")
}
b.cond(fn, s.Cond, then, els)
fn.currentBlock = then
b.stmt(fn, s.Body)
emitJump(fn, done)
if s.Else != nil {
fn.currentBlock = els
b.stmt(fn, s.Else)
emitJump(fn, done)
}
fn.currentBlock = done
// TODO statement debug info.
case *ast.SwitchStmt:
b.switchStmt(fn, s, label)
case *ast.TypeSwitchStmt:
b.typeSwitchStmt(fn, s, label)
case *ast.SelectStmt:
b.selectStmt(fn, s, label)
case *ast.ForStmt:
b.forStmt(fn, s, label)
case *ast.RangeStmt:
b.rangeStmt(fn, s, label)
default:
panic(fmt.Sprintf("unexpected statement kind: %T", s))
}
}
// buildFunction builds SSA code for the body of function fn. Idempotent.
func (b *builder) buildFunction(fn *Function) {
if fn.Blocks != nil {
return // building already started
}
if fn.syntax == nil {
return // not a Go source function. (Synthetic, or from object file.)
}
if fn.syntax.body == nil {
// External function.
if fn.Params == nil {
// This condition ensures we add a non-empty
// params list once only, but we may attempt
// the degenerate empty case repeatedly.
// TODO(adonovan): opt: don't do that.
// We set Function.Params even though there is no body
// code to reference them. This simplifies clients.
if recv := fn.Signature.Recv(); recv != nil {
fn.addParamObj(recv)
}
params := fn.Signature.Params()
for i, n := 0, params.Len(); i < n; i++ {
fn.addParamObj(params.At(i))
}
}
return
}
if fn.Prog.mode&LogSource != 0 {
defer logStack("build function %s @ %s", fn, fn.Prog.Fset.Position(fn.pos))()
}
fn.startBody()
fn.createSyntacticParams()
b.stmt(fn, fn.syntax.body)
go.tools/ssa: implement correct control flow for recovered panic. A function such as this: func one() (x int) { defer func() { recover() }() x = 1 panic("return") } that combines named return parameters (NRPs) with deferred calls that call recover, may return non-zero values despite the fact it doesn't even contain a return statement. (!) This requires a change to the SSA API: all functions' control-flow graphs now have a second entry point, called Recover, which is the block at which control flow resumes after a recovered panic. The Recover block simply loads the NRPs and returns them. As an optimization, most functions don't need a Recover block, so it is omitted. In fact it is only needed for functions that have NRPs and defer a call to another function that _may_ call recover. Dataflow analysis of SSA now requires extra work, since every may-panic instruction has an implicit control-flow edge to the Recover block. The only dataflow analysis so far implemented is SSA renaming, for which we make the following simplifying assumption: the Recover block only loads the NRPs and returns. This means we don't really need to analyze it, we can just skip the "lifting" of such NRPs. We also special-case the Recover block in the dominance computation. Rejected alternative approaches: - Specifying a Recover block for every defer instruction (like a traditional exception handler). This seemed like excessive generality, since Go programs only need the same degenerate form of Recover block. - Adding an instruction to set the Recover block immediately after the named return values are set up, so that dominance can be computed without special-casing. This didn't seem worth the effort. Interpreter: - This CL completely reimplements the panic/recover/ defer logic in the interpreter. It's clearer and simpler and closer to the model in the spec. - Some runtime panic messages have been changed to be closer to gc's, since tests depend on it. - The interpreter now requires that the runtime.runtimeError type be part of the SSA program. This requires that clients import this package prior to invoking the interpreter. This in turn requires (Importer).ImportPackage(path string), which this CL adds. - All $GOROOT/test/recover{,1,2,3}.go tests are now passing. NB, the bug described in coverage.go (defer/recover in a concatenated init function) remains. Will be fixed in a follow-up. Fixes golang/go#6381 R=gri CC=crawshaw, golang-dev https://golang.org/cl/13844043
2013-10-14 13:38:56 -06:00
if cb := fn.currentBlock; cb != nil && (cb == fn.Blocks[0] || cb == fn.Recover || cb.Preds != nil) {
// Run function calls deferred in this function when
// falling off the end of the body block.
fn.emit(new(RunDefers))
fn.emit(new(Return))
}
fn.finishBody()
}
// buildFuncDecl builds SSA code for the function or method declared
// by decl in package pkg.
//
func (b *builder) buildFuncDecl(pkg *Package, decl *ast.FuncDecl) {
id := decl.Name
if isBlankIdent(id) {
return // discard
}
var fn *Function
if decl.Recv == nil && id.Name == "init" {
pkg.ninit++
fn = &Function{
name: fmt.Sprintf("init$%d", pkg.ninit),
Signature: new(types.Signature),
pos: decl.Name.NamePos,
Pkg: pkg,
Prog: pkg.Prog,
syntax: &funcSyntax{
functype: decl.Type,
recvField: decl.Recv,
body: decl.Body,
},
}
var v Call
v.Call.Value = fn
v.setType(types.NewTuple())
pkg.init.emit(&v)
} else {
fn = pkg.values[pkg.objectOf(id)].(*Function)
}
b.buildFunction(fn)
}
// BuildAll calls Package.Build() for each package in prog.
// Building occurs in parallel unless the BuildSerially mode flag was set.
//
// BuildAll is idempotent and thread-safe.
//
func (prog *Program) BuildAll() {
var wg sync.WaitGroup
go.tools/importer: generalize command-line syntax. Motivation: pointer analysis tools (like the oracle) want the user to specify a set of initial packages, like 'go test'. This change enables the user to specify a set of packages on the command line using importer.LoadInitialPackages(args). Each argument is interpreted as either: - a comma-separated list of *.go source files together comprising one non-importable ad-hoc package. e.g. "src/pkg/net/http/triv.go" gives us [main]. - an import path, denoting both the imported package and its non-importable external test package, if any. e.g. "fmt" gives us [fmt, fmt_test]. Current type-checker limitations mean that only the first import path may contribute tests: multiple packages augmented by *_test.go files could create import cycles, which 'go test' avoids by building a separate executable for each one. That approach is less attractive for static analysis. Details: (many files touched, but importer.go is the crux) importer: - PackageInfo.Importable boolean indicates whether package is importable. - un-expose Importer.Packages; expose AllPackages() instead. - CreatePackageFromArgs has become LoadInitialPackages. - imports() moved to util.go, renamed importsOf(). - InitialPackagesUsage usage message exported to clients. - the package name for ad-hoc packages now comes from the 'package' decl, not "main". ssa.Program: - added CreatePackages() method - PackagesByPath un-exposed, renamed 'imported'. - expose AllPackages and ImportedPackage accessors. oracle: - describe: explain and workaround a go/types bug. Misc: - Removed various unnecessary error.Error() calls in Printf args. R=crawshaw CC=golang-dev https://golang.org/cl/13579043
2013-09-06 16:13:57 -06:00
for _, p := range prog.packages {
if prog.mode&BuildSerially != 0 {
p.Build()
} else {
wg.Add(1)
go func(p *Package) {
p.Build()
wg.Done()
}(p)
}
}
wg.Wait()
}
// Build builds SSA code for all functions and vars in package p.
//
// Precondition: CreatePackage must have been called for all of p's
// direct imports (and hence its direct imports must have been
// error-free).
//
// Build is idempotent and thread-safe.
//
func (p *Package) Build() {
if !atomic.CompareAndSwapInt32(&p.started, 0, 1) {
return // already started
}
go.tools/ssa: fix computation of set of types requiring method sets. Motivation: Previously, we assumed that the set of types for which a complete method set (containing all synthesized wrapper functions) is required at runtime was the set of types used as operands to some *ssa.MakeInterface instruction. In fact, this is an underapproximation because types can be derived from other ones via reflection, and some of these may need methods. The reflect.Type API allows *T to be derived from T, and these may have different method sets. Reflection also allows almost any subcomponent of a type to be accessed (with one exception: given T, defined 'type T struct{S}', you can reach S but not struct{S}). As a result, the pointer analysis was unable to generate all necessary constraints before running the solver, causing a crash when reflection derives types whose methods are unavailable. (A similar problem would afflict an ahead-of-time compiler based on ssa. The ssa/interp interpreter was immune only because it does not require all wrapper methods to be created before execution begins.) Description: This change causes the SSA builder to record, for each package, the set of all types with non-empty method sets that are referenced within that package. This set is accessed via Packages.TypesWithMethodSets(). Program.TypesWithMethodSets() returns its union across all packages. The set of references that matter are: - types of operands to some MakeInterface instruction (as before) - types of all exported package members - all subcomponents of the above, recursively. This is a conservative approximation to the set of types whose methods may be called dynamically. We define the owning package of a type as follows: - the owner of a named type is the package in which it is defined; - the owner of a pointer-to-named type is the owner of that named type; - the owner of all other types is nil. A package must include the method sets for all types that it owns, and all subcomponents of that type that are not owned by another package, recursively. Types with an owner appear in exactly one package; types with no owner (such as struct{T}) may appear within multiple packages. (A typical Go compiler would emit multiple copies of these methods as weak symbols; a typical linker would eliminate duplicates.) Also: - go/types/typemap: implement hash function for *Tuple. - pointer: generate nodes/constraints for all of ssa.Program.TypesWithMethodSets(). Add rtti.go regression test. - Add API test of Package.TypesWithMethodSets(). - Set Function.Pkg to nil (again) for wrapper functions, since these may be shared by many packages. - Remove a redundant logging statement. - Document that ssa CREATE phase is in fact sequential. Fixes golang/go#6605 R=gri CC=golang-dev https://golang.org/cl/14920056
2013-10-23 15:07:52 -06:00
// Ensure we have runtime type info for all exported members.
// TODO(adonovan): ideally belongs in memberFromObject, but
// that would require package creation in topological order.
for obj := range p.values {
if obj.IsExported() {
p.needMethodsOf(obj.Type())
}
}
if p.info.Files == nil {
p.info = nil
return // nothing to do
}
if p.Prog.mode&LogSource != 0 {
defer logStack("build %s", p)()
}
init := p.init
init.startBody()
// Make init() skip if package is already initialized.
initguard := p.Var("init$guard")
doinit := init.newBasicBlock("init.start")
done := init.newBasicBlock("init.done")
emitIf(init, emitLoad(init, initguard), done, doinit)
init.currentBlock = doinit
emitStore(init, initguard, vTrue)
// Call the init() function of each package we import.
for _, pkg := range p.info.Pkg.Imports() {
prereq := p.Prog.packages[pkg]
if prereq == nil {
panic(fmt.Sprintf("Package(%q).Build(): unsatisfied import: Program.CreatePackage(%q) was not called", p.Object.Path(), pkg.Path()))
}
var v Call
v.Call.Value = prereq.init
v.Call.pos = init.pos
v.setType(types.NewTuple())
init.emit(&v)
}
b := &builder{
nTo1Vars: make(map[*ast.ValueSpec]bool),
}
// Pass 1: visit the package's var decls and in source order,
// causing init() code to be generated in topological order.
// We visit package-level vars transitively through functions
// and methods, building them as we go.
for _, file := range p.info.Files {
for _, decl := range file.Decls {
if decl, ok := decl.(*ast.GenDecl); ok && decl.Tok == token.VAR {
for _, spec := range decl.Specs {
b.globalValueSpec(init, spec.(*ast.ValueSpec), nil, nil)
}
}
}
}
// Pass 2: build all package-level functions, init functions
// and methods in source order, including unreachable/blank ones.
for _, file := range p.info.Files {
for _, decl := range file.Decls {
if decl, ok := decl.(*ast.FuncDecl); ok {
b.buildFuncDecl(p, decl)
}
}
}
// Finish up init().
emitJump(init, done)
init.currentBlock = done
init.emit(new(RunDefers))
init.emit(new(Return))
init.finishBody()
p.info = nil // We no longer need ASTs or go/types deductions.
}
// Only valid during p's create and build phases.
func (p *Package) objectOf(id *ast.Ident) types.Object {
if o := p.info.ObjectOf(id); o != nil {
return o
}
panic(fmt.Sprintf("no types.Object for ast.Ident %s @ %s",
id.Name, p.Prog.Fset.Position(id.Pos())))
}
// Only valid during p's create and build phases.
func (p *Package) typeOf(e ast.Expr) types.Type {
return p.info.TypeOf(e)
}
go.tools/ssa: fix computation of set of types requiring method sets. Motivation: Previously, we assumed that the set of types for which a complete method set (containing all synthesized wrapper functions) is required at runtime was the set of types used as operands to some *ssa.MakeInterface instruction. In fact, this is an underapproximation because types can be derived from other ones via reflection, and some of these may need methods. The reflect.Type API allows *T to be derived from T, and these may have different method sets. Reflection also allows almost any subcomponent of a type to be accessed (with one exception: given T, defined 'type T struct{S}', you can reach S but not struct{S}). As a result, the pointer analysis was unable to generate all necessary constraints before running the solver, causing a crash when reflection derives types whose methods are unavailable. (A similar problem would afflict an ahead-of-time compiler based on ssa. The ssa/interp interpreter was immune only because it does not require all wrapper methods to be created before execution begins.) Description: This change causes the SSA builder to record, for each package, the set of all types with non-empty method sets that are referenced within that package. This set is accessed via Packages.TypesWithMethodSets(). Program.TypesWithMethodSets() returns its union across all packages. The set of references that matter are: - types of operands to some MakeInterface instruction (as before) - types of all exported package members - all subcomponents of the above, recursively. This is a conservative approximation to the set of types whose methods may be called dynamically. We define the owning package of a type as follows: - the owner of a named type is the package in which it is defined; - the owner of a pointer-to-named type is the owner of that named type; - the owner of all other types is nil. A package must include the method sets for all types that it owns, and all subcomponents of that type that are not owned by another package, recursively. Types with an owner appear in exactly one package; types with no owner (such as struct{T}) may appear within multiple packages. (A typical Go compiler would emit multiple copies of these methods as weak symbols; a typical linker would eliminate duplicates.) Also: - go/types/typemap: implement hash function for *Tuple. - pointer: generate nodes/constraints for all of ssa.Program.TypesWithMethodSets(). Add rtti.go regression test. - Add API test of Package.TypesWithMethodSets(). - Set Function.Pkg to nil (again) for wrapper functions, since these may be shared by many packages. - Remove a redundant logging statement. - Document that ssa CREATE phase is in fact sequential. Fixes golang/go#6605 R=gri CC=golang-dev https://golang.org/cl/14920056
2013-10-23 15:07:52 -06:00
// needMethodsOf ensures that runtime type information (including the
// complete method set) is available for the specified type T and all
// its subcomponents.
//
// needMethodsOf must be called for at least every type that is an
// operand of some MakeInterface instruction, and for the type of
// every exported package member.
//
func (p *Package) needMethodsOf(T types.Type) {
p.needMethods(T, false)
}
// Recursive case: skip => don't call makeMethods(T).
func (p *Package) needMethods(T types.Type, skip bool) {
// Each package maintains its own set of types it has visited.
if p.needRTTI.Set(T, true) != nil {
return // already seen
}
// Prune the recursion if we find a named or *named type
// belonging to another package.
var n *types.Named
switch T := T.(type) {
case *types.Named:
n = T
case *types.Pointer:
n, _ = T.Elem().(*types.Named)
}
if n != nil {
owner := n.Obj().Pkg()
if owner == nil {
return // built-in error type
}
if owner != p.Object {
return // belongs to another package
}
}
// All the actual method sets live in the Program so that
// multiple packages can share a single copy in memory of the
// symbols that would be compiled into multiple packages (as
// weak symbols).
if !skip && p.Prog.makeMethods(T) {
p.methodSets = append(p.methodSets, T)
}
switch t := T.(type) {
case *types.Basic:
// nop
case *types.Interface:
for i, n := 0, t.NumMethods(); i < n; i++ {
p.needMethodsOf(t.Method(i).Type())
}
case *types.Pointer:
p.needMethodsOf(t.Elem())
case *types.Slice:
p.needMethodsOf(t.Elem())
case *types.Chan:
p.needMethodsOf(t.Elem())
case *types.Map:
p.needMethodsOf(t.Key())
p.needMethodsOf(t.Elem())
case *types.Signature:
if t.Recv() != nil {
p.needMethodsOf(t.Recv().Type())
}
p.needMethodsOf(t.Params())
p.needMethodsOf(t.Results())
case *types.Named:
// A pointer-to-named type can be derived from a named
// type via reflection. It may have methods too.
p.needMethodsOf(types.NewPointer(T))
// Consider 'type T struct{S}' where S has methods.
// Reflection provides no way to get from T to struct{S},
// only to S, so the method set of struct{S} is unwanted,
// so set 'skip' flag during recursion.
p.needMethods(t.Underlying(), true)
case *types.Array:
p.needMethodsOf(t.Elem())
case *types.Struct:
// TODO(adonovan): must we recur over the types of promoted methods?
for i, n := 0, t.NumFields(); i < n; i++ {
p.needMethodsOf(t.Field(i).Type())
}
case *types.Tuple:
for i, n := 0, t.Len(); i < n; i++ {
p.needMethodsOf(t.At(i).Type())
}
default:
panic(T)
}
}