from the witness tables for their associations rather than passing
them separately.
This drastically reduces the number of physical arguments required
to invoke a generic function with a complex protocol hierarchy. It's
also an important step towards allowing recursive protocol
constraints. However, it may cause some performance problems in
generic code that we'll have to figure out ways to remediate.
There are still a few places in IRGen that rely on recursive eager
expansion of associated types and protocol witnesses. For example,
passing generic arguments requires us to map from a dependent type
back to an index into the all-dependent-types list in order to
find the right Substitution; that's something we'll need to fix
more generally. Specific to IRGen, there are still a few abstractions
like NecessaryBindings that use recursive expansion and are therefore
probably extremely expensive under this patch; I intend to fix those
up in follow-ups to the greatest extent possible.
There are also still a few things that could be made lazier about
type fulfillment; for example, we eagerly project the dynamic type
metadata of class parameters rather than waiting for the first place
we actually need to do so. We should be able to be lazier about
that, at least when the parameter is @guaranteed.
Technical notes follow. Most of the basic infrastructure I set up
for this over the last few months stood up, although there were
some unanticipated complexities:
The first is that the all-dependent-types list still does not
reliably contain all the dependent types in the minimized signature,
even with my last patch, because the primary type parameters aren't
necessarily representatives. It is, unfortunately, important to
give the witness marker to the primary type parameter because
otherwise substitution won't be able to replace that parameter at all.
There are better representations for all of that, but it's not
something I wanted to condition this patch on; therefore, we have to
do a significantly more expensive check in order to figure out a
dependent type's index in the all-dependent-types list.
The second is that the ability to add requirements to associated
types in protocol refinements means that we have to find the *right*
associatedtype declaration in order to find the associated witness
table. There seems to be relatively poor AST support for this
operation; maybe I just missed it.
The third complexity (so far) is that the association between an
archetype and its parent isn't particularly more important than
any other association it has. We need to be able to recover
witness tables linked with *all* of the associations that lead
to an archetype. This is, again, not particularly well-supported
by the AST, and we may run into problems here when we eliminate
recursive associated type expansion in signatures.
Finally, it's a known fault that this potentially leaves debug
info in a bit of a mess, since we won't have any informaton for
a type parameter unless we actually needed it somewhere.
Most notably, the source caches did not respect dominance. The
simplest solution was just to drop them in favor of the ordinary
caching system; this is unfortunate because it requires walking
over the path twice instead of exploiting the trie, but it's much
easier to make this work, especially in combination with the other
caching mechanisms at play.
This will be tested by later commits that enable lazy-loading of
local type data in various contexts.
Instead of categorically forbidding caching within a conditional
scope, permit it but remember how to remove the cache entries.
This means that ad-hoc emission code can now get exactly the
right caching behavior if they use this properly. In keeping
with that, adjust a bunch of code to properly nest scopes
according to the conditional paths they enter.
There are several interesting new features here.
The first is that, when emitting a SILFunction, we're now able to
cache type data according to the full dominance structure of the
original function. For example, if we ask for type metadata, and
we've already computed it in a dominating position, we're now able
to re-use that value; previously, we were limited to only doing this
if the value was from the entry block or the LLVM basic block
matched exactly. Since this tracks the SIL dominance relationship,
things in IRGen which add their own control flow must be careful
to suppress caching within blocks that may not dominate the
fallthrough; this mechanism is currently very crude, but could be
made to allow a limited amount of caching within the
conditionally-executed blocks.
This query is done using a proper dominator tree analysis, even at -O0.
I do not expect that we will frequently need to actually build the
tree, and I expect that the code-size benefits of doing a real
analysis will be significant, especially as we move towards making
more metadata lazily computed.
The second feature is that this adds support for "abstract"
cache entries, which indicate that we know how to derive the metadata
but haven't actually done so. This code isn't yet tested, but
it's going to be the basis of making a lot of things much lazier.
A single extra inhabitant is good enough for the most important case,
that being a single level of optionality. Otherwise, we want to
reserve maximal flexibility for the implementation.
This commit also fixes a bug where I was not correctly defining
the extra-inhabitant rules for all of the existential cases.
This is a bit of a hodge-podge of related changes that I decided
weren't quite worth teasing apart:
First, rename the weak{Retain,Release} entrypoints to
unowned{Retain,Release} to better reflect their actual use
from generated code.
Second, standardize the names of the rest of the entrypoints around
unowned{operation}.
Third, standardize IRGen's internal naming scheme and API for
reference-counting so that (1) there are generic functions for
emitting operations using a given reference-counting style and
(2) all operations explicitly call out the kind and style of
reference counting.
Finally, implement a number of new entrypoints for unknown unowned
reference-counting. These entrypoints use a completely different
and incompatible scheme for working with ObjC references. The
primary difference is that the new scheme abandons the flawed idea
(which I take responsibility for) that we can simulate an unowned
reference count for ObjC references, and instead moves towards an
address-only scheme when the reference might store an ObjC reference.
(The current implementation is still trivially takable, but that is
not something we should be relying on.) These will be tested in a
follow-up commit. For now, we still rely on the bad assumption of
reference-countability.
The drivers for this change are providing a simpler API to SIL pass
authors, having a more efficient of the in-memory representation,
and ruling out an entire class of common bugs that usually result
in hard-to-debug backend crashes.
Summary
-------
SILInstruction
Old New
+---------------+ +------------------+ +-----------------+
|SILInstruction | |SILInstruction | |SILDebugLocation |
+---------------+ +------------------+ +-----------------+
| ... | | ... | | ... |
|SILLocation | |SILDebugLocation *| -> |SILLocation |
|SILDebugScope *| +------------------+ |SILDebugScope * |
+---------------+ +-----------------+
We’re introducing a new class SILDebugLocation which represents the
combination of a SILLocation and a SILDebugScope.
Instead of storing an inline SILLocation and a SILDebugScope pointer,
SILInstruction now only has one SILDebugLocation pointer. The APIs of
SILBuilder and SILDebugLocation guarantees that every SILInstruction
has a nonempty SILDebugScope.
Developer-visible changes include:
SILBuilder
----------
In the old design SILBuilder populated the InsertedInstrs list to
allow setting the debug scopes of all built instructions in bulk
at the very end (as the responsibility of the user). In the new design,
SILBuilder now carries a "current debug scope" state and immediately
sets the debug scope when an instruction is inserted.
This fixes a use-after-free issue with with SIL passes that delete
instructions before destroying the SILBuilder that created them.
Because of this, SILBuilderWithScopes no longer needs to be a template,
which simplifies its call sites.
SILInstruction
--------------
It is neither possible or necessary to manually call setDebugScope()
on a SILInstruction any more. The function still exists as a private
method, but is only used when splicing instructions from one function
to another.
Efficiency
----------
In addition to dropping 20 bytes from each SILInstruction,
SILDebugLocations are now allocated in the SILModule's bump pointer
allocator and are uniqued by SILBuilder. Unfortunately repeat compiles
of the standard library already vary by about 5% so I couldn’t yet
produce reliable numbers for how much this saves overall.
rdar://problem/22017421
The swift_unknown* entry points are not available on the Linux port.
Previously we would still attempt to use them in a couple of cases:
1) Foreign classes
2) Existentials and archetypes
3) Optionals of boxed existentials
Note that this patch changes IRGen to never emit the
swift_errorRelease/Retain entry points on Linux. We would like to
use them in the future if we ever adopt a tagged-pointer representation
for small errors. In this case, they can be brought back, and the
TypeInfo for optionals will need to be generalized to propagate the
reference counting of the payload type, instead of defaulting to
unknown if the payload type is not natively reference counted.
A similar change will need to be made to support blocks, if we ever
want to use the blocks runtime on Linux.
Fixes <rdar://problem/23335318>, <rdar://problem/23335537>,
<rdar://problem/23335453>.
This means: handling of alloc_ref [stack].
It can be configured with two new options. See Option/FrontendOptions.td.
As the [stack] attribute is not generated yet, there should be NFC.
Swift SVN r32929
Full type metadata isn't necessary to calculate the runtime layout of a dependent struct or enum; we only need the non-function data from the value witness table (size, alignment, extra inhabitant count, and POD/BT/etc. flags). This can be generated more efficiently than the type metadata for many types--if we know a specific instantiation is fixed-layout, we can regenerate the layout information, or if we know the type has the same layout as another well-known type, we can get the layout from a common value witness table. This breaks a deadlock in most (but not all) cases where a value type is recursive using classes or fixed-layout indirected structs like UnsafePointer. rdar://problem/19898165
This time, factor out the ObjC-dependent parts of the tests so they only run with ObjC interop.
Swift SVN r30266
Full type metadata isn't necessary to calculate the runtime layout of a dependent struct or enum; we only need the non-function data from the value witness table (size, alignment, extra inhabitant count, and POD/BT/etc. flags). This can be generated more efficiently than the type metadata for many types--if we know a specific instantiation is fixed-layout, we can regenerate the layout information, or if we know the type has the same layout as another well-known type, we can get the layout from a common value witness table. This breaks a deadlock in most (but not all) cases where a value type is recursive using classes or fixed-layout indirected structs like UnsafePointer. rdar://problem/19898165
Swift SVN r30243
When producing TypeInfo for a box, try to reuse instantiations for common type structures:
- any POD type with the same stride and alignment can share a fixed HeapLayout;
- any single-refcounted-pointer type can share a fixed HeapLayout with types that have the same ReferenceCounting;
- dynamically-sized types can share a runtime-based box implementation.
For the runtime implementation, use new to-be-implemented variants of allocBox/deallocBox that will produce polymorphically-projectable boxes using instantiated metadata.
Swift SVN r29612
Don't project every value witness from the metadata every time we need one; this wastes code size in a way LLVM can't really optimize since it doesn't know the metadata is immutable. The code size wins on the standard library are disappointingly small (stdlib only shrinks by 4KB), but this makes generic IR a lot more compact and easier to read.
Swift SVN r28095
Preparation to fix <rdar://problem/18151694> Add Builtin.checkUnique
to avoid lost Array copies.
This adds the following new builtins:
isUnique : <T> (inout T[?]) -> Int1
isUniqueOrPinned : <T> (inout T[?]) -> Int1
These builtins take an inout object reference and return a
boolean. Passing the reference inout forces the optimizer to preserve
a retain distinct from what’s required to maintain lifetime for any of
the reference's source-level copies, because the called function is
allowed to replace the reference, thereby releasing the referent.
Before this change, the API entry points for uniqueness checking
already took an inout reference. However, after full inlining, it was
possible for two source-level variables that reference the same object
to appear to be the same variable from the optimizer's perspective
because an address to the variable was longer taken at the point of
checking uniqueness. Consequently the optimizer could remove
"redundant" copies which were actually needed to implement
copy-on-write semantics. With a builtin, the variable whose reference
is being checked for uniqueness appears mutable at the level of an
individual SIL instruction.
The kind of reference count checking that Builtin.isUnique performs
depends on the argument type:
- Native object types are directly checked by reading the
strong reference count:
(Builtin.NativeObject, known native class reference)
- Objective-C object types require an additional check that the
dynamic object type uses native swift reference counting:
(Builtin.UnknownObject, unknown class reference, class existential)
- Bridged object types allow the dymanic object type check to be
bypassed based on the pointer encoding:
(Builtin.BridgeObject)
Any of the above types may also be wrapped in an optional. If the
static argument type is optional, then a null check is also performed.
Thus, isUnique only returns true for non-null, native swift object
references with a strong reference count of one.
isUniqueOrPinned has the same semantics as isUnique except that it
also returns true if the object is marked pinned regardless of the
reference count. This allows for simultaneous non-structural
modification of multiple subobjects.
In some cases, the standard library can dynamically determine that it
has a native reference even though the static type is a bridge or
unknown object. Unsafe variants of the builtin are available to allow
the additional pointer bit mask and dynamic class lookup to be
bypassed in these cases:
isUnique_native : <T> (inout T[?]) -> Int1
isUniqueOrPinned_native : <T> (inout T[?]) -> Int1
These builtins perform an implicit cast to NativeObject before
checking uniqueness. There’s no way at SIL level to cast the address
of a reference, so we need to encapsulate this operation as part of
the builtin.
Swift SVN r27887
Some future-proofing to let us change ErrorType's reference counting in the future, or to use various tagged pointer optimizations in its representation.
Swift SVN r27213
As part of this, re-arrange the argument order so that
generic arguments come before the context, which comes
before the error result. Be more consistent about always
adding a context parameter on thick functions, even
when it's unused. Pull out the witness-method Self
argument so that it appears last after the error
argument.
Swift SVN r26667
This avoids redundant calls of swift_get*Metadata() if the metadata is reused in the same basic block
or if it is already availabe in the function entry block.
There is still room for improvement, e.g. by not limiting the reuse to a single block.
This change reduces the dylib codesize by about 11%, mostly in protocoll witnesses (-34%).
It also results in some improvements for the -Onone performance:
RangeAssignment: +31%
StdlibSort: +20%
Swift SVN r25123
OpaqueStorageTypeInfo uses iNNN types that don't always have the correct alloc size for an expected
size at the LLVM level. This needs to be fixed before my partial apply closure fixes can hold.
Swift SVN r24551
Using the intrinsics is obnoxious because I needed them
to return Builtin.NativeObject?, but there's no reasonable
way to safely generate optional types from Builtins.cpp.
Ugh.
Dave and I also decided that there's no need for
swift_tryPin to allow a null object.
Swift SVN r23824
This is a type that has ownership of a reference while allowing access to the
spare bits inside the pointer, but which can also safely hold an ObjC tagged pointer
reference (with no spare bits of course). It additionally blesses one
Foundation-coordinated bit with the meaning of "has swift refcounting" in order
to get a faster short-circuit to native refcounting. It supports the following
builtin operations:
- Builtin.castToBridgeObject<T>(ref: T, bits: Builtin.Word) ->
Builtin.BridgeObject
Creates a BridgeObject that contains the bitwise-OR of the bit patterns of
"ref" and "bits". It is the user's responsibility to ensure "bits" doesn't
interfere with the reference identity of the resulting value. In other words,
it is undefined behavior unless:
castReferenceFromBridgeObject(castToBridgeObject(ref, bits)) === ref
This means "bits" must be zero if "ref" is a tagged pointer. If "ref" is a real
object pointer, "bits" must not have any non-spare bits set (unless they're
already set in the pointer value). The native discriminator bit may only be set
if the object is Swift-refcounted.
- Builtin.castReferenceFromBridgeObject<T>(bo: Builtin.BridgeObject) -> T
Extracts the reference from a BridgeObject.
- Builtin.castBitPatternFromBridgeObject(bo: Builtin.BridgeObject) -> Builtin.Word
Presents the bit pattern of a BridgeObject as a Word.
BridgeObject's bits are set up as follows on the various platforms:
i386, armv7:
No ObjC tagged pointers
Swift native refcounting flag bit: 0x0000_0001
Other available spare bits: 0x0000_0002
x86_64:
Reserved for ObjC tagged pointers: 0x8000_0000_0000_0001
Swift native refcounting flag bit: 0x0000_0000_0000_0002
Other available spare bits: 0x7F00_0000_0000_0004
arm64:
Reserved for ObjC tagged pointers: 0x8000_0000_0000_0000
Swift native refcounting flag bit: 0x4000_0000_0000_0000
Other available spare bits: 0x3F00_0000_0000_0007
TODO: BridgeObject doesn't present any extra inhabitants. It ought to at least provide null as an extra inhabitant for Optional.
Swift SVN r22880
llvm::Optional lives in "llvm/ADT/Optional.h". Like Clang, we can get
Optional in the 'swift' namespace by including "swift/Basic/LLVM.h".
We're now fully switched over to llvm::Optional!
Swift SVN r22477
- A spot fix in SILGen for reabstracting the result of a downcast, which fixes checked casts to function types.
- Associate the layout information in type metadata records with the most abstract representation of the type. This is the correct thing to do in cases where we need the metadata as a tag for an opaque value--if we store a value in an Any, or pass it as an unconstrained generic parameter, we must maximally reabstract it. This fixes the value semantics of existentials containing trivial metatypes.
- To ensure that we get runtime layout of structs and enums correct when they contain reabstractable types, introduce a "metadata for layout" concept, which doesn't need to describe the canonical metadata for the type, but only needs to describe a type with equivalent layout and value semantics. This is a correctness fix that allows us to correctly lay out generic types containing dependent tuples and functions, and although we don't really take advantage of it here, it's also a potential runtime performance win down the road, because we could potentially produce direct metadata for a primitive type that's layout-equivalent with a runtime-instantiated type. To aid in type safety here, push SILType deeper into IRGen in places where we potentially care about specific representations of types.
- Finally, fix an inconsistency between the runtime and IRGen's concept of what spare bits unmanaged references and thick metatypes have.
Together, these fixes address rdar://problem/16406907, rdar://problem/17822208, rdar://problem/18189508, and likely many other related issues, and also fixes crash suite cases 012 and 024.
Swift SVN r21963
I introduced a function swift_keepAlive2() which has a different signature from
swift_keepAlive() until I can verify that the stdlib is using the new
infrastructure.
The difference in signature is that swift_keepAlive2 takes just a pointer while
swift_keepAlive also takes a metadata value that is not necessary for our
purposes anymore.
Swift SVN r21718
If a type has to be passed or returned resiliently, it
will necessarily be passed indirectly, which is already
represented in SILFunctionType. There is no need to
represent this as a separate channel of information.
NFC. Also fixes a problem where the signature cache
for ExtraData::Block was writing past the end of an
array (but into the storage for an adjacent array
which was fortunately never used).
ExtraData should also disappear as a concept, but we're
still relying on that for existential protocol witnesses.
Swift SVN r21548
Failable initializers will need to clean up an uninitialized 'self' box in the case where a class initializer calls out to a 'super.init' or 'self.init' fails. Because some framework classes assume that they have a refcount of exactly one during initialization, we have to take 'self' from the box during the delegation, and reinitialize the box with the result of the delegation on success. This means we actually have to implement lowering for dealloc_box.
Swift SVN r21493
dynamicCastClass assumes that the destination type is a
Swift class type.
dynamicCastObjCClass assumes that the destination type is
an ObjC class type (represented as ObjC metadata, not type
metadata).
dynamicCastUnknownClass assumes only that the destination
type is some kind of class.
Swift SVN r18776
When we partially apply an inner-pointer method or property, the thunk or partial_apply that applies "self" needs to be the one that handles lifetime-extending "self". Verify that a partial_apply-ed inner pointer method is not inner pointer and implement lifetime extension in the partial apply thunk. Fixes <rdar://problem/16803701>.
Swift SVN r17321
Replace HeapTypeInfo::hasSwiftRefcount with a "getSwiftRefcounting" method, returning an enum indicating whether a heap object has native/ObjC/block/unknown refcounting semantics. Use _Block_copy and _Block_release for block refcounting.
Swift SVN r16041
We shouldn't store to one size and load as another, and shouldn't need
to handle things like truncation and extension of numeric types, either.
Swift SVN r12530
Previously we coerced only struct return types, but we actually need to
be able to coerce scalar types as well, e.g. differing pointer types
when the ABI type differs from the type used in the body of the
function.
Swift SVN r12527
