- consistently set underlying types of incoming named types in typInternal
- use underlying() helper function to resolve forward chains
- related consistency cleanups
LGTM=adonovan
R=adonovan
CC=golang-codereviews
https://golang.org/cl/53870044
This is (a) more efficient and (b) avoids the need for
constant error handling, since buffer writes can't fail.
Also:
- added WriteFunction and WritePackage functions,
similar to types.WriteFoo.
- *Function and *Package now implement io.WriterTo.
LGTM=gri
R=gri
CC=golang-codereviews
https://golang.org/cl/57930044
This results in significant improvement to type-checking time:
it reduces by 4% the entire running time of ssa/stdlib_test
(GOMAXPROCS=8, n=7).
LGTM=gri
R=gri
CC=golang-codereviews
https://golang.org/cl/57770043
Since Put() makes a dynamic function call, this CL includes a long-overdue change to supply a *frame (and thus the call
stack and the interpreter) to intrinsics too.
This fixes the tests that were broken by (sound) revision e4a4cb47c141.
LGTM=gri
R=gri
CC=golang-codereviews
https://golang.org/cl/57350043
CL 49530047 made the (over-)simplifying assumption that the
Path of an ad-hoc (Created) package should default to its
Name, i.e. package declaration.
With this change, the Name is still always computed from the
package declaration (by go/types) but the Path may be
specified by the loader.Config. If "", the value of the Name
is used, which is not globally unique.
R=gri, axwalk
CC=golang-codereviews
https://golang.org/cl/55180043
Was: Now:
call.Graph callgraph.Graph
call.GraphNode callgraph.Node
call.Edge callgraph.Edge
Though call.Graph was cute, the original naming was a mistake:
'call' is too useful a var name to waste on a package.
R=gri, crawshaw
CC=golang-codereviews
https://golang.org/cl/53190043
The goal is to allow clients to distinguish function-local objects
from package-local objects and identify the specific function by
examining its scope. For example, a compiler may need to mangle
object names according to the outer function name.
R=gri
CC=golang-codereviews, golang-dev
https://golang.org/cl/52280043
Both gc and gccgo appear to use this size, so it seems appropriate
to use it here.
R=gri
CC=golang-codereviews, golang-dev, iant, rsc
https://golang.org/cl/52290043
The issue is addressed with the change in conversions.go.
Also:
- added corresponding API test
- made names for untyped ints, bools consistent with others
- use *Basic with names byte/rune instead of uint8/int32 for better output
- minor cleanups
Fixesgolang/go#6949.
R=adonovan
CC=golang-codereviews
https://golang.org/cl/50520043
Now that inner functions are processed "in line", usage errors
can be detected immediately after each function is processed.
R=adonovan
CC=golang-codereviews
https://golang.org/cl/49900044
Inner function bodies must be type-checked "in-place" for
them to see the correct state of the surrounding scopes.
Fixesgolang/go#7035.
R=adonovan
CC=golang-codereviews
https://golang.org/cl/49350043
This can import all of the standard library, and has been tested
by using gotype to type check libgo with gccgo's export data (this
would be nice to automate, but I can't see a good way to do it,
not least because system-specific source files cause errors which
I needed to identify manually).
It includes a builtin export locator. Unfortunately I can't see a
more reliable way to locate the builtin export files than to parse
the output of 'gccgo -###'.
R=gri, iant, gri
CC=golang-codereviews, golang-dev
https://golang.org/cl/31860043
- if a named type was imported before, read it again
and throw it away in favor of the existing type
(the old code did the same, but more circuitously)
- better tag name for int64 values
R=adonovan
CC=golang-codereviews
https://golang.org/cl/47650044
Anonymous field names are emitted as "" in the export data
since the actual name can be reconstructed easily from the
field's type name. But "" names are not exported names and
thus the respective qualified name emits complete package
information even if the actual field name is exported. Fix
the package upon import.
R=adonovan
CC=golang-codereviews
https://golang.org/cl/42090044
The package of a struct field is independent of the type of the
struct field - the old code was technically not correct. That said,
it does not seem possible (or very difficult) to create a test case
because for exported anonymous fields that field package doesn't matter
(it's not needed for name identity), and non-exported anonymous fields
cannot be accessed from an imported package.
R=adonovan
CC=golang-codereviews
https://golang.org/cl/47560043
Also: Provide GcCompatibilityMode for printing types
(intended for testing with gc-generated export data
only).
(TBR adonovan)
R=adonovan
TBR=adonovan
CC=golang-codereviews
https://golang.org/cl/44780043
By using a simple (graph-based) serialization algorithm
and binary encoding, a significantly more compact export
data format is achieved than what the current compilers
use. Furthermore, the exporter and importer are completely
symmetric algorithms that are compact, and much easier to
change/expand.
R=adonovan
CC=golang-dev
https://golang.org/cl/42960043
Obviously in that mode, we can't correctly diagnose such
errors, so we shouldn't attempt it (and emit false positives).
R=gri
CC=golang-dev
https://golang.org/cl/41080043
This test generates a program that declares the constant
elements of an n*n Hilbert matrix, its inverse, and the
constant elements of the explicit product of the two.
The product should be the identity matrix; that check is
also expressed as a constant expression. Type-checking
verifies that the product is indeed the identity matrix
by asserting the result of the identity check (using the
assert built-in which is available for type-check tests).
The test is run for n = 5. Other values can be tested via
the -H flag, say: go test -run=Hilbert -H=100
The generated program can be written to a file for testing
the constant arithmetic of a compiler: go test -out=test.go
Because of the mathematically precise constant arithmetic
of go/types, this test should always succeed and is only
limited by the size of the matrix. It does run successfully
from n = 0 to values larger than 100.
The Hilbert matrix is famous for being ill-conditioned and
thus exposes arithmetic imprecision very quickly. The gc
compiler only produces a correct result for n = 0 (trivially),
and n = 1 at the moment.
R=adonovan, rsc
CC=golang-dev
https://golang.org/cl/35840043
The benchmarks don't permit clear-cut apples-to-apples
comparisons since they depend on the very source code
they are testing. But they do give an indication of
the approximate performance - as a sanity test.
Using 3 different code bases makes it apparent that
there's some difference in performance per code base;
i.e., the lines/s speed varies pretty strongly. This
may be due to setup costs, or other issues. We should
investigate eventually.
R=adonovan
CC=golang-dev
https://golang.org/cl/30650043
Also: Use defer to make sure scopes are always closed
even in case of early exits via bailout (and don't cause
an imbalanced scope error in debug mode).
R=adonovan, gri
CC=golang-dev
https://golang.org/cl/29300043
- consolidate them in (expr|type)string[_test].go
- support for printing all ast.Expr
- fix printing of type: chan (<-chan int) (parentheses)
- more consistent names, comments, and exports
R=adonovan
CC=golang-dev
https://golang.org/cl/28680043
These are variants of Type.String(), Object.String() that
accept a 'from *Package' argument. If provided, package
qualification is omitted when printing named types belonging
to that package.
This is useful for UIs where a package is implied by context
e.g. ssadump disassembly, oracle output.
+ Test.
R=gri, gri, gordon.klaus
CC=golang-dev
https://golang.org/cl/22190048
- fixed a couple of TODOs
- various cleanups along the way
- adjusted clients
Once submitted, clients of go/types that don't explicitly
specify Config.Import will need to add the extra import:
import _ "code.google.com/p/go.tools/go/gcimporter"
to install the default (gc) importer in go/types.
R=adonovan, gri
CC=golang-dev
https://golang.org/cl/26390043
Depending on the context, printing only the package name can
be ambiguous or even incorrect since it is valid only within
the environment of a given file's import specs.
(The standard library packages are mostly unique in their last
segment, but this is not the case for proprietary repos.)
R=gri, gri
CC=golang-dev
https://golang.org/cl/26300043
This check is currently done in go/parser as well but
eventually can be removed from there (after Go 1.2).
R=adonovan
CC=golang-dev
https://golang.org/cl/22240045
Initialization cycles need to reported for cycles
that contain variables, even if they don't end in
a variable.
This fixes the last known issue with the existing
std library tests.
R=adonovan, gri
CC=golang-dev
https://golang.org/cl/22200049
The gc compiler is inconsistent how it handles method
"mentions" with respect to initialization cycle detection
(see issue 6703 for details). Pending a spec clarification,
this CL assumes that for a method to be "mentioned", it
must be mentioned as a method expression rather than a
method value (closer in intent to "syntactic" mention).
R=adonovan
CC=golang-dev
https://golang.org/cl/22050044
- Info.InitOrder now provides list of Initializers
(vars + init expr), handling n:1 decls and blank
identifiers
- added respective API test
- cycles detected through function "mentions"
Missing: cycles through method "mentions" and via
closures
R=adonovan
CC=golang-dev
https://golang.org/cl/21810043
Missing:
- dependencies via functions (incl. closures and methods)
- more tests (incl. API test)
R=adonovan
CC=golang-dev
https://golang.org/cl/20510043
- better error messages
- in contrast to a long-standing TODO, comparisons
between interface and non-interface types always
worked correctly
R=adonovan
CC=golang-dev
https://golang.org/cl/17310043
The spec does not exclude blank _ method names in interfaces
from the uniqueness criteria; i.e., at most one blank method
may appear in an interface type.
Arguably the spec is vague (and possibly incorrect) here.
gccgo handles it the same way. gc crashes with an internal
compiler error.
R=adonovan
CC=golang-dev
https://golang.org/cl/16380043
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
This will make it a bit easier to create commonly used "custom" sizes for types.
With this CL, interfaces are now by default 2*WordSize (= 16) instead of 1*WordSize
as before.
Also: minor unrelated cleanups.
R=adonovan
CC=golang-dev
https://golang.org/cl/14719043
Given a built-in call f(args), Info.Types now maps f to the call-site
specific type of f (by looking at the argument types) if the built-in
call is not producing a constant (at typecheck time) result. If the
result is constant, the recorded type is invalid (a back-end won't
need it).
R=adonovan
CC=golang-dev
https://golang.org/cl/14598045
Catch ... errors earlier and in case of error, type-check all
arguments anyway for better type reporting and fewer "declared
but not used" errors.
R=adonovan
CC=golang-dev
https://golang.org/cl/14600043
Assignments to "comma, ok" expressions on the lhs of an
assignment are not permitted unless we have map index
"comma, ok" expression. Created new operand mode 'mapindex'
to distinguish this case. Renamed mode 'valueok' to the more
commonly used 'commaok' term, which also makes it easier to
distinguish from simply 'value'.
Added corresponding tests.
Fixes a TODO.
R=adonovan
CC=golang-dev
https://golang.org/cl/14526049
- removed support for nil constants from go/exact
- instead define a singleton Nil Object (the nil _value_)
- in assignments, follow more closely spec wording
(pending spec CL 14415043)
- removed use of goto in checker.unary
- cleanup around handling of isRepresentable for
constants, with better error messages
- fix missing checks in checker.convertUntyped
- added isTyped (== !isUntyped) and isInterface predicates
- fixed hasNil predicate: unsafe.Pointer also has nil
- adjusted ssa per adonovan
- implememted types.Implements (wrapper arounfd types.MissingMethod)
- use types.Implements in vet (and fix a bug)
R=adonovan, r
CC=golang-dev
https://golang.org/cl/14438052
- factor out argument extraction logic
- cleaned up error handling in builtin.go (no need for goto's anymore)
- lots of additional test cases
- various cleanups, better documentation
Fixesgolang/go#5795.
R=adonovan
CC=golang-dev
https://golang.org/cl/14312044
It would be nice to be able to use this package
as a dependency (or other go utilities in the
ecosystem that depend on this package) in
environments which have not (or cannot) for
whatever reason upgraded to newer versions of
golang.
R=golang-dev, adg
CC=golang-dev
https://golang.org/cl/14283043
Revision f280b8a485fd of the std library changed the
gc export format: anonymous fields may be qualified
with a package.
R=rsc
TBR=rsc
CC=golang-dev
https://golang.org/cl/14312043
- permit ERROR markers to be in full or line comments
- don't require ""s in /* ERROR "foo" */
- enable more std tests
- some minor cleanups
R=adonovan
CC=golang-dev
https://golang.org/cl/14169044
In general, if a break or continue label is not found, we don't
know if a correspondingly named label was not declared, was declared
but is not visible, or will be declared (and won't be visible).
Complain about "invalid" rather than "not declared" label.
Added more tests.
R=adonovan
CC=golang-dev
https://golang.org/cl/14149043
This CL temporarily removes some preliminary label checks.
They will be implemented completely in a subsequent CL.
R=adonovan
CC=golang-dev
https://golang.org/cl/14055043
This change affects the API: Func objects now always have a *Signature
as type (never a *Builtin). Instead, built-ins now appear as *Builtin
objects. Only the built-in name is exposed, other fields are now private
to go/types.
Several bugs are fixed:
- correctly checking for built-ins permitted in statement context
- expression statements that are calls are not type-checked twice anymore
- go/defer statements report call types and provide good error messages now
This CL will briefly break the build until CL 13848043 is submitted.
R=adonovan
CC=golang-dev
https://golang.org/cl/13813043
also:
- initial code for unused label errors
- some cleanups, better names
- additional tests
TODO: Dot-imported packages are not handled yet; i.e., they
are always considered used for now.
R=adonovan
CC=golang-dev
https://golang.org/cl/13768043
Also:
- added more tests
- removed Var.Used accessor: it's not meaningful for clients since
it does not reflect actual use/def information
- fixed position for short variable declaration errors
R=adonovan
CC=golang-dev
https://golang.org/cl/13240051
- updated all tests to conform to stricter rules
- TODO: check for implicitly declared variables in type switches
R=adonovan
CC=golang-dev
https://golang.org/cl/13695046
1. handle return statements with zero (but expected) return values
2. indices provided for array or slice composite literals must be integer constants
Added additional test cases.
R=adonovan
CC=golang-dev
https://golang.org/cl/13734043
Robert Griesemer
Alan Donovan
Peter Collingbourne
Richard Musiol
Matt Reiferson
Russ Cox