Generic metadata prespecialization: enums #29345

nate-chandler · 2020-01-22T00:06:01Z

Continuing the work started in #28610 , prespecializes generic metadata for enums under limited circumstances. Specifically, prespecialization only occurs for usages within the same module where the type itself is defined and for specializations all of whose type arguments are themselves concrete types.

nate-chandler · 2020-01-22T00:06:31Z

@swift-ci please smoke test

nate-chandler · 2020-01-22T00:06:38Z

@swift-ci please test windows platform

slavapestov · 2020-01-22T03:57:47Z

lib/IRGen/GenMeta.cpp

@@ -3789,10 +3806,34 @@ namespace {

      return flags;
    }
+  };
+
+  // FIXME: Without this template typealias, the following errors are produced 


Interesting, is there a radar for this?

Unfortunately, I wasn't able to minimally recreate the problem, but rdar://problem/58884416 covers this.

slavapestov · 2020-01-22T03:59:12Z

lib/IRGen/MetadataRequest.cpp

@@ -684,11 +684,6 @@ bool irgen::isNominalGenericContextTypeMetadataAccessTrivial(
    return false;


I think this is too conservative. nominal.isResilient(..., ResilienceExpansion::Minimal) ignores the module argument and returns true if the module is resilient.

But if its resilient and defined in the same module, you can still pre-specialize it. The resilience means the metadata is opaque across module boundaries, not within the module itself.

slavapestov · 2020-01-22T04:00:55Z

I can't comment on unchanged lines in the diff, but about this part:

    // For now, to avoid statically specializing generic protocol witness
    // tables, don't statically specialize metadata for types any of whose
    // arguments are generic.
    //
    // TODO: This is more pessimistic than necessary.  Specialize even in
    //       the face of generic arguments so long as those arguments
    //       aren't required to conform to any protocols.
    //
    // TODO: Once witness tables are statically specialized, check whether the
    //       ConformanceInfo returns nullptr from tryGetConstantTable.
    //       early return.

Can you change this to use the same predicates as the code for emitting witness table references, where it decides if it should reference a witness table directly or call an accessor function? I think isDependentConformance() is the right check, or at least a starting point.

slavapestov · 2020-01-22T04:02:14Z

test/IRGen/prespecialized-metadata/enum-inmodule-1argument-1distinct_use.swift

+
+// CHECK: @"$s4main5ValueOySiGWV" = linkonce_odr hidden constant %swift.enum_vwtable { i8* bitcast ({{(%swift.opaque\* \(\[[0-9]+ x i8\]\*, \[[0-9]+ x i8\]\*, %swift.type\*\)\* @"\$[a-zA-Z0-9_]+" to i8\*|i8\* \(i8\*, i8\*, %swift.type\*\)\* @__swift_memcpy[0-9]+_[0-9]+ to i8\*)}}), i8* bitcast (void (i8*, %swift.type*)* @__swift_noop_void_return to i8*), i8* bitcast (i8* (i8*, i8*, %swift.type*)* @__swift_memcpy{{[0-9]+}}_{{[0-9]+}} to i8*), i8* bitcast (i8* (i8*, i8*, %swift.type*)* @__swift_memcpy{{[0-9]+}}_{{[0-9]+}} to i8*), i8* bitcast (i8* (i8*, i8*, %swift.type*)* @__swift_memcpy{{[0-9]+}}_{{[0-9]+}} to i8*), i8* bitcast (i8* (i8*, i8*, %swift.type*)* @__swift_memcpy{{[0-9]+}}_{{[0-9]+}} to i8*), i8* bitcast (i32 (%swift.opaque*, i32, %swift.type*)* @"$s4main5ValueOySiGwet" to i8*), i8* bitcast (void (%swift.opaque*, i32, i32, %swift.type*)* @"$s4main5ValueOySiGwst" to i8*), [[INT]] [[ALIGNMENT]], [[INT]] [[ALIGNMENT]], i32 {{[0-9]+}}, i32 0, i8* bitcast (i32 (%swift.opaque*, %swift.type*)* @"$s4main5ValueOySiGwug" to i8*), i8* bitcast (void (%swift.opaque*, %swift.type*)* @"$s4main5ValueOySiGwup" to i8*), i8* bitcast (void (%swift.opaque*, i32, %swift.type*)* @"$s4main5ValueOySiGwui" to i8*) }, align [[ALIGNMENT]]
+// CHECK: @"$s4main5ValueOySiGMf" = internal constant <{ i8**, [[INT]], %swift.type_descriptor*, %swift.type*, i64 }> <{ i8** getelementptr inbounds (%swift.enum_vwtable, %swift.enum_vwtable* @"$s4main5ValueOySiGWV", i32 0, i32 0), [[INT]] 513, %swift.type_descriptor* bitcast (<{ i32, i32, i32, i32, i32, i32, i32, i32, i32, i16, i16, i16, i16, i8, i8, i8, i8 }>* @"$s4main5ValueOMn" to %swift.type_descriptor*), %swift.type* @"$sSiN", i64 3 }>, align [[ALIGNMENT]]
+enum Value<First> {


Can you add some tests that use protocol conformances, both the supported and unsupported cases?

Yes, I'm in the process of writing more.

nate-chandler · 2020-01-24T02:50:41Z

@swift-ci please test

swift-ci · 2020-01-24T02:51:04Z

Build failed
Swift Test Linux Platform
Git Sha - a745581219fe577a2421647048fe588ef842af5a

swift-ci · 2020-01-24T03:00:10Z

Build failed
Swift Test OS X Platform
Git Sha - a745581219fe577a2421647048fe588ef842af5a

nate-chandler · 2020-01-24T21:35:56Z

@swift-ci please test linux platform

nate-chandler · 2020-01-24T21:36:01Z

@swift-ci please test windows platform

swift-ci · 2020-01-24T21:40:23Z

Build failed
Swift Test Linux Platform
Git Sha - 7e8d20c67154b899f16598950b765cb69b3e0a23

nate-chandler · 2020-01-24T21:46:33Z

@swift-ci please clean test linux platform

nate-chandler · 2020-01-24T21:52:31Z

@swift-ci please clean test windows platform

swift-ci · 2020-01-24T22:21:40Z

Build failed
Swift Test Linux Platform
Git Sha - 568d6dae24d88cd96983f41cd51d6cf32694e988

nate-chandler · 2020-01-24T23:12:49Z

@swift-ci please clean test linux platform

nate-chandler · 2020-01-24T23:12:57Z

@swift-ci please clean test linux platform

swift-ci · 2020-01-24T23:15:28Z

Build failed
Swift Test Linux Platform
Git Sha - 568d6dae24d88cd96983f41cd51d6cf32694e988

nate-chandler · 2020-01-26T04:28:57Z

Can you change this to use the same predicates as the code for emitting witness table references, where it decides if it should reference a witness table directly or call an accessor function? I think isDependentConformance() is the right check, or at least a starting point.

@slavapestov, thanks for pointing me at that! Is this along the lines of what you have in mind? b716235

nate-chandler · 2020-02-06T19:22:55Z

@swift-ci please test

swift-ci · 2020-02-06T19:55:14Z

Build failed
Swift Test Linux Platform
Git Sha - 40bc9fe4a2d7b90388da22b856e03f9082be43b1

swift-ci · 2020-02-06T20:04:32Z

Build failed
Swift Test OS X Platform
Git Sha - 40bc9fe4a2d7b90388da22b856e03f9082be43b1

swift-ci · 2020-02-06T22:56:57Z

Build failed
Swift Test OS X Platform
Git Sha - 40bc9fe4a2d7b90388da22b856e03f9082be43b1

nate-chandler · 2020-02-11T01:03:57Z

@swift-ci please test

nate-chandler · 2020-02-11T01:04:19Z

@swift-ci please test windows platform

swift-ci · 2020-02-11T02:15:04Z

Build failed
Swift Test OS X Platform
Git Sha - 63cd291415a97e882a5943444756334c9b23eb1b

swift-ci · 2020-02-11T02:26:33Z

Build failed
Swift Test Linux Platform
Git Sha - 63cd291415a97e882a5943444756334c9b23eb1b

nate-chandler · 2020-02-11T03:38:39Z

@swift-ci please test

nate-chandler · 2020-02-11T03:39:04Z

@swift-ci please test windows platform

swift-ci · 2020-02-11T03:41:09Z

Build failed
Swift Test Linux Platform
Git Sha - 63cd291415a97e882a5943444756334c9b23eb1b

swift-ci · 2020-02-11T03:43:03Z

Build failed
Swift Test OS X Platform
Git Sha - 63cd291415a97e882a5943444756334c9b23eb1b

Extracted implementation of SpecializedGenericStructMetadataBuilder into SpecializedGenericNominalMetadataBuilderBase, a CRTP with a template template argument for the CRTP superclass and a template argument for the implementation. That new type is now subclassed by SpecializedGenericStructMetadataBuilder. Additionally, this new type is also subclassed by the newly added SpecializedGenericEnumMetadataBuilder which is responsible for build the prespecialization of generic enum metadata. rdar://problem/56960887

nate-chandler · 2020-02-11T22:55:49Z

@swift-ci please test

nate-chandler · 2020-02-11T22:55:57Z

@swift-ci please test windows platform

swift-ci · 2020-02-11T22:58:18Z

Build failed
Swift Test Linux Platform
Git Sha - 521d8cb318a555973c89cfab23ced37879d0a6e8

swift-ci · 2020-02-11T23:00:07Z

Build failed
Swift Test OS X Platform
Git Sha - 521d8cb318a555973c89cfab23ced37879d0a6e8

nate-chandler · 2020-02-12T01:30:14Z

@swift-ci please test

swift-ci · 2020-02-12T01:32:46Z

Build failed
Swift Test Linux Platform
Git Sha - 76eaa6d74aa67ba494e4442d6a2f124e0bcac439

swift-ci · 2020-02-12T01:33:36Z

Build failed
Swift Test OS X Platform
Git Sha - 76eaa6d74aa67ba494e4442d6a2f124e0bcac439

When a specialized usage of a generic enum occurs in the same module where the enum was defined, directly reference the prespecialized metadata. rdar://problem/56994321

nate-chandler · 2020-02-12T18:58:31Z

@swift-ci please test

swift-ci · 2020-02-12T19:01:03Z

Build failed
Swift Test Linux Platform
Git Sha - a2e5db24f34500cd57f36f7467b1ae5b41108a61

swift-ci · 2020-02-12T19:01:58Z

Build failed
Swift Test OS X Platform
Git Sha - a2e5db24f34500cd57f36f7467b1ae5b41108a61

nate-chandler changed the title ~~Generic metadata prespecialization components/enums~~ Generic metadata prespecialization: enums Jan 22, 2020

slavapestov reviewed Jan 22, 2020

View reviewed changes

nate-chandler force-pushed the generic-metadata-prespecialization-components/enums branch from a745581 to 7e8d20c Compare January 24, 2020 02:50

nate-chandler force-pushed the generic-metadata-prespecialization-components/enums branch from 7e8d20c to 568d6da Compare January 24, 2020 21:34

nate-chandler force-pushed the generic-metadata-prespecialization-components/enums branch from 568d6da to a8b7f6c Compare January 24, 2020 23:11

nate-chandler force-pushed the generic-metadata-prespecialization-components/enums branch from a8b7f6c to 40bc9fe Compare February 6, 2020 19:21

nate-chandler force-pushed the generic-metadata-prespecialization-components/enums branch 3 times, most recently from f554099 to 324f52a Compare February 6, 2020 20:35

nate-chandler force-pushed the generic-metadata-prespecialization-components/enums branch from cca4be0 to 63cd291 Compare February 11, 2020 01:01

nate-chandler force-pushed the generic-metadata-prespecialization-components/enums branch 2 times, most recently from d957ebd to 521d8cb Compare February 11, 2020 03:38

nate-chandler force-pushed the generic-metadata-prespecialization-components/enums branch 2 times, most recently from 07954ac to 76eaa6d Compare February 11, 2020 22:55

nate-chandler force-pushed the generic-metadata-prespecialization-components/enums branch from 76eaa6d to a2e5db2 Compare February 12, 2020 01:30

[metadata prespecialization] Direct refs to enums.

40e17d9

When a specialized usage of a generic enum occurs in the same module where the enum was defined, directly reference the prespecialized metadata. rdar://problem/56994321

nate-chandler force-pushed the generic-metadata-prespecialization-components/enums branch from a2e5db2 to 40e17d9 Compare February 12, 2020 18:57

nate-chandler merged commit b628710 into swiftlang:master Feb 12, 2020

nate-chandler deleted the generic-metadata-prespecialization-components/enums branch February 12, 2020 21:10

		@@ -684,11 +684,6 @@ bool irgen::isNominalGenericContextTypeMetadataAccessTrivial(
		return false;

Generic metadata prespecialization: enums #29345

Generic metadata prespecialization: enums #29345

Uh oh!

Conversation

nate-chandler commented Jan 22, 2020

Uh oh!

nate-chandler commented Jan 22, 2020

Uh oh!

nate-chandler commented Jan 22, 2020

Uh oh!

slavapestov Jan 22, 2020

Choose a reason for hiding this comment

Uh oh!

nate-chandler Jan 24, 2020

Choose a reason for hiding this comment

Uh oh!

slavapestov Jan 22, 2020

Choose a reason for hiding this comment

Uh oh!

slavapestov commented Jan 22, 2020

Uh oh!

slavapestov Jan 22, 2020

Choose a reason for hiding this comment

Uh oh!

nate-chandler Jan 22, 2020

Choose a reason for hiding this comment

Uh oh!

nate-chandler commented Jan 24, 2020

Uh oh!

swift-ci commented Jan 24, 2020

Uh oh!

swift-ci commented Jan 24, 2020

Uh oh!

nate-chandler commented Jan 24, 2020

Uh oh!

nate-chandler commented Jan 24, 2020

Uh oh!

swift-ci commented Jan 24, 2020

Uh oh!

nate-chandler commented Jan 24, 2020

Uh oh!

nate-chandler commented Jan 24, 2020

Uh oh!

swift-ci commented Jan 24, 2020

Uh oh!

nate-chandler commented Jan 24, 2020

Uh oh!

nate-chandler commented Jan 24, 2020

Uh oh!

swift-ci commented Jan 24, 2020

Uh oh!

nate-chandler commented Jan 26, 2020

Uh oh!

nate-chandler commented Feb 6, 2020

Uh oh!

swift-ci commented Feb 6, 2020

Uh oh!

swift-ci commented Feb 6, 2020

Uh oh!

swift-ci commented Feb 6, 2020

Uh oh!

nate-chandler commented Feb 11, 2020

Uh oh!

nate-chandler commented Feb 11, 2020

Uh oh!

swift-ci commented Feb 11, 2020

Uh oh!

swift-ci commented Feb 11, 2020

Uh oh!

nate-chandler commented Feb 11, 2020

Uh oh!

nate-chandler commented Feb 11, 2020

Uh oh!

swift-ci commented Feb 11, 2020

Uh oh!

swift-ci commented Feb 11, 2020

Uh oh!

nate-chandler commented Feb 11, 2020

Uh oh!

nate-chandler commented Feb 11, 2020