[Runtime] Improve representation of protocol conformance records #13685

DougGregor · 2018-01-03T06:58:56Z

This pull request improves the runtime representation of protocol conformance records and type metadata records in several ways:

Move conformance kind into the low bits of the witness table reference, and be explicit about the remaining reserved kind
Reference Objective-C class objects indirectly, so the runtime can fix them up
Eliminate TypeMetadataRecordKind::UniqueDirectClass, which is no longer necessary
Use nominal type descriptors whenever we can, reducing the number of TypeMetadataRecordKind cases to 3 (nominal type descriptor, indirect Objective-C class object, and nonuniqued foreign type metadata). Pack the kind into the lower bits of type references in protocol conformance records and type metadata records.
Eliminate the "flags" word from protocol conformance records, but leave the word reserved for conditional requirements.

The protocol conformance record has two bits to describe how the witness table will be produced. There are currently three states (direct reference to witness table, witness table accessor, and conditional witness table accessor). Add a reserved case for the fourth state so the Swift 5 runtime will (silently) ignore conformances using that fourth state, when/if some future Swift uses it.

The format of protocol conformance records will be changing in Swift 5, so rename the segment (from, e.g., __swift2_proto to __swift5_proto) to allow Swift < 5 and Swift 5+ runtimes to coexist.

…nce. Move the 2-bit conformance reference kind from the conformance flags (which is meant to go away) into the lower two bits of the witness table offset.

…rds. Within conformance records, reference Objective-C class objects indirectly so the runtime can update those references appropriately. We don't need to do this for classes with Swift metadata.

Now that references to Objective-C class objects are indirected (via UniqueIndirectClass), classes with Swift type metadata can be directly referenced (via UniqueDirectType) rather than hopping through swift_getObjCClassMetadata().

DougGregor · 2018-01-03T06:59:04Z

@swift-ci please test

swift-ci · 2018-01-03T07:01:09Z

Build failed
Swift Test OS X Platform
Git Sha - b6b0697ffc2a465d637a904d2e69d26f0c861526

Effectively NFC, because the compiler is not likely to ever notice the aliasing violation here.

DougGregor · 2018-01-03T07:04:29Z

@swift-ci please test

DougGregor · 2018-01-03T07:04:34Z

@swift-ci please test

DougGregor · 2018-01-03T07:04:41Z

@swift-ci please test

swift-ci · 2018-01-03T07:04:54Z

Build failed
Swift Test Linux Platform
Git Sha - b6b0697ffc2a465d637a904d2e69d26f0c861526

swift-ci · 2018-01-03T07:04:56Z

Build failed
Swift Test OS X Platform
Git Sha - b6b0697ffc2a465d637a904d2e69d26f0c861526

Rather than emitting unique, direct type metadata for non-foreign types, emit a reference to the nominal type descriptor. This collapses the set of type metadata reference kinds to 3: nominal type descriptor, (indirect) Objective-C class object, and nonuniqued foreign type metadata.

DougGregor · 2018-01-03T09:38:10Z

@swift-ci please test

DougGregor · 2018-01-03T09:38:15Z

@swift-ci please test source compatibility

DougGregor · 2018-01-03T09:38:24Z

@swift-ci please test

DougGregor · 2018-01-03T09:38:35Z

@swift-ci please test

DougGregor · 2018-01-03T09:38:43Z

@swift-ci please test source compatibility

DougGregor · 2018-01-03T09:38:55Z

@swift-ci please test source compatibility

swift-ci · 2018-01-03T10:24:34Z

Build failed
Swift Test Linux Platform
Git Sha - 29a1ede93177a39eb2301b437ddcdd6f562c9a7a

swift-ci · 2018-01-03T10:45:02Z

Build failed
Swift Test OS X Platform
Git Sha - 29a1ede93177a39eb2301b437ddcdd6f562c9a7a

DougGregor · 2018-01-03T16:56:26Z

@jckarter, I suggest holding off on reviewing for a little bit. I have a cleaner solution coming.

…formances Use the spare bits within the type reference field to describe the kinds of type metadata records, so that we no longer need to rely on a separate "flags" field.

Eliminate the separate flags field in protocol conformance records, now that all of the information is stored in spare bits elsewhere. Reserve this 32-bit value for future use to describe conditional requirements.

…erences. The separate section of type references uses the same type reference format as in protocol conformance records. As with protocol conformance records, mangle the type reference kind into the lower two bits. Then, eliminate the separate "flags" field from the type metadata record. Finally, rename the section because the Swift 5 stable format for this section is different from prior formats, and the two runtimes need to be able to coexist.

DougGregor · 2018-01-03T19:04:07Z

@swift-ci please test

DougGregor · 2018-01-03T19:04:25Z

@swift-ci please test

DougGregor · 2018-01-03T19:04:44Z

@swift-ci please test source compatibility

DougGregor · 2018-01-03T19:04:56Z

@swift-ci please test source compatibility

DougGregor · 2018-01-03T19:05:12Z

@swift-ci please test source compatibility

jckarter · 2018-01-03T20:45:13Z

include/swift/ABI/MetadataValues.h

+
+  /// The conformance is for a nominal type referenced directly;
+  /// getNominalTypeDescriptor() points to the nominal type descriptor.
+  IndirectNominalTypeDescriptor = 0x01,


Need to update the comment here.

jckarter · 2018-01-03T20:46:38Z

include/swift/ABI/MetadataValues.h

-  /// platforms, the class object always is the type metadata.
-  UniqueDirectClass = 0xF,
+  /// and classes are emitted as UniqueDirectType.
+  UniqueIndirectClass = 0x03,


The name and comment here could be updated to indicate this kind is necessary for ObjC classes, since they don't have nominal type descriptors.

jckarter · 2018-01-03T20:48:24Z

include/swift/ABI/MetadataValues.h

  /// The conformance is for a nongeneric foreign struct or enum type.
  /// getDirectType() points to a nonunique metadata record for the type, which
  /// needs to be uniqued by the runtime.
-  NonuniqueDirectType,
+  NonuniqueDirectType = 0x02,


I guess this case goes away if CF class metadata gets nominal type descriptors (and we'd want to reserve at least one case for future expansion, for a "mangled name" representation perhaps), but the comment could be updated here too.

jckarter · 2018-01-03T20:52:17Z

include/swift/Runtime/Metadata.h

-    RelativeIndirectablePointer<TargetNominalTypeDescriptor<Runtime>>
-    TypeDescriptor;
+    RelativeIndirectablePointerIntPair<TargetNominalTypeDescriptor<Runtime>,
+                                       /*always clear=*/bool> TypeDescriptor;


Since we're using the spare bits to encode the type metadata kind, and we have separate kinds for direct and indirect type descriptor references, I think it might be clearer to use two union members here:

RelativeDirectPointerIntPair<const TargetNominalTypeDescriptor<Runtime> *, TypeMetadataRecordKind> DirectTypeDescriptor; RelativeDirectPointerIntPair<const TargetNominalTypeDescriptor<Runtime> * const *, TypeMetadataRecordKind> IndirectTypeDescriptor;

since that avoids the cleverness of the TypeMetadataRecordKind's kinds for nominal type descriptors just happening to line up with the way RelativeIndirectablePointerIntPair indicates indirection. (Furthermore, it keeps RelativeIndirectablePointerIntPair from being necessary at all.)

Yeah, I agree that's clearer. Thanks!

…ceRecord. Per JoeG's comments, thanks!

DougGregor · 2018-01-03T21:22:31Z

@swift-ci please smoke test

jckarter · 2018-01-03T21:23:38Z

include/swift/Basic/RelativePointer.h

+/// an additional tiny integer value.
+template<typename ValueTy, typename IntTy, bool Nullable = false,
+         typename Offset = int32_t>
+class RelativeIndirectablePointerIntPair {


This can go away now, right?

(Nevermind, I see it's still used to reserve the unused bit in the protocol reference.)

DougGregor · 2018-01-03T21:32:21Z

@swift-ci please smoke test and merge

DougGregor added 5 commits January 2, 2018 10:09

[Runtime] Rename the segment for protocol conformance records.

8c668ba

The format of protocol conformance records will be changing in Swift 5, so rename the segment (from, e.g., __swift2_proto to __swift5_proto) to allow Swift < 5 and Swift 5+ runtimes to coexist.

[Runtime] Move conformance kind into low bits of witness table refere…

7e6f523

…nce. Move the 2-bit conformance reference kind from the conformance flags (which is meant to go away) into the lower two bits of the witness table offset.

[Runtime] Reference ObjC class objects indirectly in conformance reco…

32ab89b

…rds. Within conformance records, reference Objective-C class objects indirectly so the runtime can update those references appropriately. We don't need to do this for classes with Swift metadata.

DougGregor requested a review from jckarter January 3, 2018 06:59

[Runtime] Cast to the correct type.

b25dc44

Effectively NFC, because the compiler is not likely to ever notice the aliasing violation here.

DougGregor force-pushed the protocol-conformance-record-cleanup branch from b6b0697 to b25dc44 Compare January 3, 2018 07:04

DougGregor added 3 commits January 3, 2018 09:23

[Runtime] Pack TypeMetadataRecordKind into spare bits of protocol con…

aee0c68

…formances Use the spare bits within the type reference field to describe the kinds of type metadata records, so that we no longer need to rely on a separate "flags" field.

[Runtime] Eliminate ProtocolConformanceFlags.

bfd2230

Eliminate the separate flags field in protocol conformance records, now that all of the information is stored in spare bits elsewhere. Reserve this 32-bit value for future use to describe conditional requirements.

DougGregor force-pushed the protocol-conformance-record-cleanup branch from 29a1ede to 310bd6b Compare January 3, 2018 18:50

jckarter requested changes Jan 3, 2018

View reviewed changes

[Runtime] Clean up TypeMetadataRecordKind and TargetProtocolConforman…

3219047

…ceRecord. Per JoeG's comments, thanks!

jckarter reviewed Jan 3, 2018

View reviewed changes

jckarter approved these changes Jan 3, 2018

View reviewed changes

DougGregor merged commit 0632d87 into swiftlang:master Jan 3, 2018

DougGregor deleted the protocol-conformance-record-cleanup branch January 3, 2018 22:11

[Runtime] Improve representation of protocol conformance records #13685

[Runtime] Improve representation of protocol conformance records #13685

Uh oh!

Conversation

DougGregor commented Jan 3, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

DougGregor commented Jan 3, 2018

Uh oh!

swift-ci commented Jan 3, 2018

Uh oh!

DougGregor commented Jan 3, 2018

Uh oh!

DougGregor commented Jan 3, 2018

Uh oh!

DougGregor commented Jan 3, 2018

Uh oh!

swift-ci commented Jan 3, 2018

Uh oh!

swift-ci commented Jan 3, 2018

Uh oh!

DougGregor commented Jan 3, 2018

Uh oh!

DougGregor commented Jan 3, 2018

Uh oh!

DougGregor commented Jan 3, 2018

Uh oh!

DougGregor commented Jan 3, 2018

Uh oh!

DougGregor commented Jan 3, 2018

Uh oh!

DougGregor commented Jan 3, 2018

Uh oh!

swift-ci commented Jan 3, 2018

Uh oh!

swift-ci commented Jan 3, 2018

Uh oh!

DougGregor commented Jan 3, 2018

Uh oh!

DougGregor commented Jan 3, 2018

Uh oh!

DougGregor commented Jan 3, 2018

Uh oh!

DougGregor commented Jan 3, 2018

Uh oh!

DougGregor commented Jan 3, 2018

Uh oh!

DougGregor commented Jan 3, 2018

Uh oh!

jckarter Jan 3, 2018

Choose a reason for hiding this comment

Uh oh!

jckarter Jan 3, 2018

Choose a reason for hiding this comment

Uh oh!

jckarter Jan 3, 2018

Choose a reason for hiding this comment

Uh oh!

jckarter Jan 3, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

DougGregor Jan 3, 2018

Choose a reason for hiding this comment

Uh oh!

DougGregor commented Jan 3, 2018

Uh oh!

jckarter Jan 3, 2018

Choose a reason for hiding this comment

Uh oh!

jckarter Jan 3, 2018

Choose a reason for hiding this comment

Uh oh!

DougGregor commented Jan 3, 2018

Uh oh!

Uh oh!

DougGregor commented Jan 3, 2018 •

edited

Loading

jckarter Jan 3, 2018 •

edited

Loading