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.
Fixesgolang/go#6605
R=gri
CC=golang-dev
https://golang.org/cl/14920056
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.
Fixesgolang/go#6381
R=gri
CC=crawshaw, golang-dev
https://golang.org/cl/13844043
It is needed after all, as I discovered in the pointer analysis.
(ssa needs more API test coverage.)
Also:
- sanity: remove debugging cruft
- promote: don't redundantly include the function's
own name in its Synthetic string.
- print: IntuitiveMethodSet utility fixes a bug in the
printing logic (for interface types, mset(*T) is empty).
The function is also used by the Oracle.
R=crawshaw
CC=golang-dev
https://golang.org/cl/13116043
Fix bug: the Signature for an interface method wrapper
erroneously had a non-nil receiver.
Function:
- Set Pkg field non-nil even for wrappers.
It is equal to that of the wrapped function.
Only wrappers of error.Error
(and its embeddings in other interfaces) may have nil.
Sanity checker now asserts this.
- FullName() now uses .Synthetic field to discriminate
synthetic methods, not Pkg==nil.
- Fullname() uses new relType() utility to print receiver type
name unqualified if it belongs to the same package.
(Alloc.String also uses relType utility.)
CallCommon:
- Description(): fix switch logic broken when we
eliminated the Recv field.
- better docs.
R=david.crawshaw, crawshaw, gri
CC=golang-dev
https://golang.org/cl/13057043
e.g.
type T int
func (T) f() {}
var t T
_ = t.f // method value: should have signature "func()", no receiver
Also:
- ssa: add sanity check that helped diagnose this.
R=gri
CC=golang-dev
https://golang.org/cl/12283043
(Motivation: "Literal" is a syntactic property, not a semantic one.)
Also: delete a "TODO: opt" that the lifting pass already does for us.
R=gri
CC=golang-dev
https://golang.org/cl/11351043
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
Before, all values received on some channel by Select would
flow to an empty interface, creating a spurious confluence for
flow analyses. Now, the tuple returned by Select has one
component for each 'receive' case.
Also, fixes:
- Removed workarounds for now-fixed typechecker bug in FuncLit+TypeAssert.
- sanity check that all Value Instructions have non-nil Type().
- Convert: document and sanity-check that at least one of the types is basic.
Also, other things to help clients:
- Define CallInstruction interface: common parts of Call, Go, Defer.
- Add CallCommon.Signature() method.
- Literal.Pos() is now populated.
R=gri
CC=golang-dev
https://golang.org/cl/10505043
Method sets:
- Simplify CallCommon.
Avoid the implicit copy when calling a T method on a *T
receiver. This simplifies clients. Instead we generate
"indirection wrapper" functions that do this (like gc does).
New invariant:
m's receiver type is exactly T for all m in MethodSet(T)
- MakeInterface no longer holds the concrete type's MethodSet.
We can defer its computation this way.
- ssa.Type now just wraps a types.TypeName object.
MethodSets are computed as needed, not eagerly.
Position info:
- new CanonicalPos utility maps ast.Expr to canonical
token.Pos, as returned by {Instruction,Value}.Pos() methods.
- Don't set posn for implicit operations (e.g. varargs array alloc)
- Set position info for ChangeInterface and Slice instructions.
Cosmetic:
- add Member.Token() method
- simplify isPointer
- Omit words "interface", "slice" when printing MakeInterface,
MakeSlice; the type is enough.
- Comments on PathEnclosingInterval.
- Remove Function.FullName() where implicit String() suffices.
Also:
- Exposed NewLiteral to clients.
- Added ssa.Instruction.Parent() *Function
Added ssa.BasicBlock.Parent() *Function.
Added Sanity checks for above.
R=golang-dev, gri
CC=golang-dev
https://golang.org/cl/10166045
