Merge pull request #33296 from hborla/merge-joined-literal-typevars

[ConstraintSystem] Add a type variable merging heuristic to addJoinConstraint

Author: hborla

lib/Sema/CSDiagnostics.cpp

@@ -4954,6 +4954,9 @@ bool CollectionElementContextualFailure::diagnoseAsError() {
 
   Optional<InFlightDiagnostic> diagnostic;
   if (isExpr<ArrayExpr>(anchor)) {
+    if (diagnoseMergedLiteralElements())
+      return true;
+
     diagnostic.emplace(emitDiagnostic(diag::cannot_convert_array_element,
                                       eltType, contextualType));
   }
@@ -5004,6 +5007,30 @@ bool CollectionElementContextualFailure::diagnoseAsError() {
   return true;
 }
 
+bool CollectionElementContextualFailure::diagnoseMergedLiteralElements() {
+  auto elementAnchor = simplifyLocatorToAnchor(getLocator());
+  if (!elementAnchor)
+    return false;
+
+  auto *typeVar = getRawType(elementAnchor)->getAs<TypeVariableType>();
+  if (!typeVar || !typeVar->getImpl().getAtomicLiteralKind())
+    return false;
+
+  // This element is a literal whose type variable could have been merged with others,
+  // but the conversion constraint to the array element type was only placed on one
+  // of them. So, we want to emit the error for each element whose type variable is in
+  // this equivalence class.
+  auto &cs = getConstraintSystem();
+  auto node = cs.getRepresentative(typeVar)->getImpl().getGraphNode();
+  for (const auto *typeVar : node->getEquivalenceClass()) {
+    auto anchor = typeVar->getImpl().getLocator()->getAnchor();
+    emitDiagnosticAt(constraints::getLoc(anchor), diag::cannot_convert_array_element,
+                     getFromType(), getToType());
+  }
+
+  return true;
+}
+
 bool MissingContextualConformanceFailure::diagnoseAsError() {
   auto anchor = getAnchor();
   auto path = getLocator()->getPath();

lib/Sema/CSDiagnostics.h

@@ -1637,6 +1637,8 @@ class CollectionElementContextualFailure final : public ContextualFailure {
       : ContextualFailure(solution, eltType, contextualType, locator) {}
 
   bool diagnoseAsError() override;
+
+  bool diagnoseMergedLiteralElements();
 };
 
 class MissingContextualConformanceFailure final : public ContextualFailure {

lib/Sema/CSGen.cpp

@@ -1803,31 +1803,28 @@ namespace {
 
       auto locator = CS.getConstraintLocator(expr);
       auto contextualType = CS.getContextualType(expr);
-      Type contextualArrayType = nullptr;
-      Type contextualArrayElementType = nullptr;
-      
+
+      auto joinElementTypes = [&](Optional<Type> elementType) {
+        const auto elements = expr->getElements();
+        unsigned index = 0;
+
+        using Iterator = decltype(elements)::iterator;
+        CS.addJoinConstraint<Iterator>(locator, elements.begin(), elements.end(),
+                                       elementType, [&](const auto it) {
+          auto *locator = CS.getConstraintLocator(expr, LocatorPathElt::TupleElement(index++));
+          return std::make_pair(CS.getType(*it), locator);
+        });
+      };
+
       // If a contextual type exists for this expression, apply it directly.
       Optional<Type> arrayElementType;
       if (contextualType &&
           (arrayElementType = ConstraintSystem::isArrayType(contextualType))) {
-        // Is the array type a contextual type
-        contextualArrayType = contextualType;
-        contextualArrayElementType = *arrayElementType;
-
         CS.addConstraint(ConstraintKind::LiteralConformsTo, contextualType,
                          arrayProto->getDeclaredInterfaceType(),
                          locator);
-        
-        unsigned index = 0;
-        for (auto element : expr->getElements()) {
-          CS.addConstraint(ConstraintKind::Conversion,
-                           CS.getType(element),
-                           contextualArrayElementType,
-                           CS.getConstraintLocator(
-                               expr, LocatorPathElt::TupleElement(index++)));
-        }
-        
-        return contextualArrayType;
+        joinElementTypes(arrayElementType);
+        return contextualType;
       }
 
       // Produce a specialized diagnostic if this is an attempt to initialize
@@ -1893,14 +1890,7 @@ namespace {
 
       // Introduce conversions from each element to the element type of the
       // array.
-      unsigned index = 0;
-      for (auto element : expr->getElements()) {
-        CS.addConstraint(ConstraintKind::Conversion,
-                         CS.getType(element),
-                         arrayElementTy,
-                         CS.getConstraintLocator(
-                           expr, LocatorPathElt::TupleElement(index++)));
-      }
+      joinElementTypes(arrayElementTy);
 
       // The array element type defaults to 'Any'.
       CS.addConstraint(ConstraintKind::Defaultable, arrayElementTy,

lib/Sema/CSSimplify.cpp

@@ -4220,6 +4220,7 @@ bool ConstraintSystem::repairFailures(
 
       conversionsOrFixes.push_back(CollectionElementContextualMismatch::create(
           *this, lhs, rhs, getConstraintLocator(locator)));
+      break;
     }
 
     // Drop the `tuple element` locator element so that all tuple element
@@ -10435,35 +10436,6 @@ void ConstraintSystem::addContextualConversionConstraint(
                 convertTypeLocator, /*isFavored*/ true);
 }
 
-Type ConstraintSystem::addJoinConstraint(
-    ConstraintLocator *locator,
-    ArrayRef<std::pair<Type, ConstraintLocator *>> inputs) {
-  switch (inputs.size()) {
-  case 0:
-    return Type();
-
-  case 1:
-    return inputs.front().first;
-
-  default:
-    // Produce the join below.
-    break;
-  }
-
-  // Create a type variable to capture the result of the join.
-  Type resultTy = createTypeVariable(locator,
-                                     (TVO_PrefersSubtypeBinding |
-                                      TVO_CanBindToNoEscape));
-
-  // Introduce conversions from each input type to the type variable.
-  for (const auto &input : inputs) {
-    addConstraint(
-      ConstraintKind::Conversion, input.first, resultTy, input.second);
-  }
-
-  return resultTy;
-}
-
 void ConstraintSystem::addFixConstraint(ConstraintFix *fix, ConstraintKind kind,
                                         Type first, Type second,
                                         ConstraintLocatorBuilder locator,

lib/Sema/ConstraintSystem.h

@@ -306,6 +306,11 @@ class TypeVariableType::Implementation {
   /// Retrieve the generic parameter opened by this type variable.
   GenericTypeParamType *getGenericParameter() const;
 
+  /// Returns the \c ExprKind of this type variable if it's the type of an
+  /// atomic literal expression, meaning the literal can't be composed of subexpressions.
+  /// Otherwise, returns \c None.
+  Optional<ExprKind> getAtomicLiteralKind() const;
+
   /// Determine whether this type variable represents a closure type.
   bool isClosureType() const;
 
@@ -3076,16 +3081,78 @@ class ConstraintSystem {
       Expr *expr, Type conversionType, ContextualTypePurpose purpose,
       bool isOpaqueReturnType);
 
+  /// Convenience function to pass an \c ArrayRef to \c addJoinConstraint
+  Type addJoinConstraint(ConstraintLocator *locator,
+                         ArrayRef<std::pair<Type, ConstraintLocator *>> inputs,
+                         Optional<Type> supertype = None) {
+    return addJoinConstraint<decltype(inputs)::iterator>(
+        locator, inputs.begin(), inputs.end(), supertype, [](auto it) { return *it; });
+  }
+
   /// Add a "join" constraint between a set of types, producing the common
   /// supertype.
   ///
   /// Currently, a "join" is modeled by a set of conversion constraints to
-  /// a new type variable. At some point, we may want a new constraint kind
-  /// to cover the join.
+  /// a new type variable or a specified supertype. At some point, we may want
+  /// a new constraint kind to cover the join.
   ///
-  /// \returns the joined type, which is generally a new type variable.
+  /// \note This method will merge any input type variables for atomic literal
+  /// expressions of the same kind. It assumes that if same-kind literal type
+  /// variables are joined, there will be no differing constraints on those
+  /// type variables.
+  ///
+  /// \returns the joined type, which is generally a new type variable, unless there are
+  /// fewer than 2 input types or the \c supertype parameter is specified.
+  template<typename Iterator>
   Type addJoinConstraint(ConstraintLocator *locator,
-                         ArrayRef<std::pair<Type, ConstraintLocator *>> inputs);
+                         Iterator begin, Iterator end,
+                         Optional<Type> supertype,
+                         std::function<std::pair<Type, ConstraintLocator *>(Iterator)> getType) {
+    if (begin == end)
+      return Type();
+
+    // No need to generate a new type variable if there's only one type to join
+    if ((begin + 1 == end) && !supertype.hasValue())
+      return getType(begin).first;
+
+    // The type to capture the result of the join, which is either the specified supertype,
+    // or a new type variable.
+    Type resultTy = supertype.hasValue() ? supertype.getValue() :
+                    createTypeVariable(locator, (TVO_PrefersSubtypeBinding | TVO_CanBindToNoEscape));
+
+    using RawExprKind = uint8_t;
+    llvm::SmallDenseMap<RawExprKind, TypeVariableType *> representativeForKind;
+
+    // Join the input types.
+    while (begin != end) {
+      Type type;
+      ConstraintLocator *locator;
+      std::tie(type, locator) = getType(begin++);
+
+      // We can merge the type variables of same-kind atomic literal expressions because they
+      // will all have the same set of constraints and therefore can never resolve to anything
+      // different.
+      if (auto *typeVar = type->getAs<TypeVariableType>()) {
+        if (auto literalKind = typeVar->getImpl().getAtomicLiteralKind()) {
+          auto *&originalRep = representativeForKind[RawExprKind(*literalKind)];
+          auto *currentRep = getRepresentative(typeVar);
+
+          if (originalRep) {
+            if (originalRep != currentRep)
+              mergeEquivalenceClasses(currentRep, originalRep);
+            continue;
+          }
+
+          originalRep = currentRep;
+        }
+      }
+
+      // Introduce conversions from each input type to the supertype.
+      addConstraint(ConstraintKind::Conversion, type, resultTy, locator);
+    }
+
+    return resultTy;
+  }
 
   /// Add a constraint to the constraint system with an associated fix.
   void addFixConstraint(ConstraintFix *fix, ConstraintKind kind,

lib/Sema/TypeCheckConstraints.cpp

@@ -82,6 +82,24 @@ TypeVariableType::Implementation::getGenericParameter() const {
   return locator ? locator->getGenericParameter() : nullptr;
 }
 
+Optional<ExprKind>
+TypeVariableType::Implementation::getAtomicLiteralKind() const {
+  if (!locator || !locator->directlyAt<LiteralExpr>())
+    return None;
+
+  auto kind = getAsExpr(locator->getAnchor())->getKind();
+  switch (kind) {
+  case ExprKind::IntegerLiteral:
+  case ExprKind::FloatLiteral:
+  case ExprKind::StringLiteral:
+  case ExprKind::BooleanLiteral:
+  case ExprKind::NilLiteral:
+    return kind;
+  default:
+    return None;
+  }
+}
+
 bool TypeVariableType::Implementation::isClosureType() const {
   if (!(locator && locator->getAnchor()))
     return false;

validation-test/SILOptimizer/large_string_array.swift.gyb

@@ -1,9 +1,6 @@
 // RUN: %empty-directory(%t)
 // RUN: %gyb %s > %t/main.swift
 
-// FIXME: <rdar://problem/66219627> Re-enable SILOptimizer/large_string_array.swift.gyb
-// REQUIRES: rdar66219627
-
 // The compiler should finish in less than 1 minute. To give some slack,
 // specify a timeout of 5 minutes.
 // If the compiler needs more than 5 minutes, there is probably a real problem.

validation-test/Sema/type_checker_perf/slow/rdar30596744_2.swift.gyb renamed to validation-test/Sema/type_checker_perf/fast/rdar30596744_2.swift.gyb

@@ -1,4 +1,4 @@
-// RUN: %scale-test --invert-result --begin 1 --end 5 --step 1 --select NumLeafScopes %s --expected-exit-code 1
+// RUN: %scale-test --begin 1 --end 5 --step 1 --select NumLeafScopes %s --expected-exit-code 1
 // REQUIRES: asserts,no_asan
 
 % enum_cases = 3