[NFC] Improve the interfaces for integer widths and parsing integers

Author: rjmccall

include/swift/AST/Expr.h

@@ -789,10 +789,12 @@ class IntegerLiteralExpr : public NumberLiteralExpr {
   static IntegerLiteralExpr *
   createFromUnsigned(ASTContext &C, unsigned value);
 
+  /// Returns the value of the literal, appropriately constructed in the
+  /// target type.
   APInt getValue() const;
-  static APInt getValue(StringRef Text, unsigned BitWidth, bool Negative);
 
-  APInt getRawMagnitude() const;
+  /// Returns the raw value of the literal without any truncation.
+  APInt getRawValue() const;
 
   static bool classof(const Expr *E) {
     return E->getKind() == ExprKind::IntegerLiteral;

include/swift/AST/Types.h

@@ -1346,13 +1346,17 @@ END_CAN_TYPE_WRAPPER(BuiltinVectorType, BuiltinType)
 class BuiltinIntegerWidth {
   /// Tag values for abstract integer sizes.
   enum : unsigned {
-    Least_SpecialValue = ~2U,
+    /// Inhabitants stolen for use as DenseMap special values.
+    DenseMapEmpty = ~0U,
+    DenseMapTombstone = ~1U,
+
+    /// An arbitrary-precision integer.
+    ArbitraryWidth = ~2U,
+
     /// The size of a pointer on the target system.
-    PointerWidth = ~0U,
+    PointerWidth = ~3U,
 
-    /// Inhabitants stolen for use as DenseMap special values.
-    DenseMapEmpty = ~1U,
-    DenseMapTombstone = ~2U,
+    Least_SpecialValue = ~3U,
   };
 
   unsigned RawValue;
@@ -1372,6 +1376,10 @@ class BuiltinIntegerWidth {
   static BuiltinIntegerWidth pointer() {
     return BuiltinIntegerWidth(PointerWidth);
   }
+
+  static BuiltinIntegerWidth arbitrary() {
+    return BuiltinIntegerWidth(ArbitraryWidth);
+  }
 
   /// Is this a fixed width?
   bool isFixedWidth() const { return RawValue < Least_SpecialValue; }
@@ -1384,6 +1392,9 @@ class BuiltinIntegerWidth {
 
   /// Is this the abstract target pointer width?
   bool isPointerWidth() const { return RawValue == PointerWidth; }
+
+  /// Is this the abstract arbitrary-width value?
+  bool isArbitraryWidth() const { return RawValue == ArbitraryWidth; }
 
   /// Get the least supported value for the width.
   ///
@@ -1393,6 +1404,8 @@ class BuiltinIntegerWidth {
       return getFixedWidth();
     if (isPointerWidth())
       return 32;
+    if (isArbitraryWidth())
+      return 1;
     llvm_unreachable("impossible width value");
   }
 
@@ -1404,8 +1417,16 @@ class BuiltinIntegerWidth {
       return getFixedWidth();
     if (isPointerWidth())
       return 64;
+    if (isArbitraryWidth())
+      return ~0U;
     llvm_unreachable("impossible width value");
   }
+
+  /// Parse a value of this bit-width.
+  ///
+  /// If the radix is 0, it is autosensed.
+  APInt parse(StringRef text, unsigned radix, bool negate,
+              bool *hadError = nullptr) const;
 
   friend bool operator==(BuiltinIntegerWidth a, BuiltinIntegerWidth b) {
     return a.RawValue == b.RawValue;
@@ -1440,7 +1461,8 @@ class BuiltinIntegerType : public BuiltinType {
     return get(BuiltinIntegerWidth::pointer(), C);
   }
 
-  /// Return the bit width of the integer.
+  /// Return the bit width of the integer.  Always returns a non-arbitrary
+  /// width.
   BuiltinIntegerWidth getWidth() const {
     return Width;
   }

lib/AST/ASTContext.cpp

@@ -3032,6 +3032,7 @@ Type ErrorType::get(Type originalType) {
 
 BuiltinIntegerType *BuiltinIntegerType::get(BuiltinIntegerWidth BitWidth,
                                             const ASTContext &C) {
+  assert(!BitWidth.isArbitraryWidth());
   BuiltinIntegerType *&Result = C.getImpl().IntegerTypes[BitWidth];
   if (Result == nullptr)
     Result = new (C, AllocationArena::Permanent) BuiltinIntegerType(BitWidth,C);

lib/AST/Expr.cpp

@@ -849,46 +849,81 @@ IntegerLiteralExpr * IntegerLiteralExpr::createFromUnsigned(ASTContext &C, unsig
   return new (C) IntegerLiteralExpr(Text, SourceLoc(), /*implicit*/ true);
 }
 
-/// A wrapper around LLVM::getAsInteger that can be used on Swift interger
-/// literals. It avoids misinterpreting decimal numbers prefixed with 0 as
-/// octal numbers.
-static bool getAsInteger(StringRef Text, llvm::APInt &Value) {
-  // swift encodes octal differently from C
-  bool IsCOctal = Text.size() > 1 && Text[0] == '0' && isdigit(Text[1]);
-  return Text.getAsInteger(IsCOctal ? 10 : 0, Value);
+APInt IntegerLiteralExpr::getRawValue() const {
+  return BuiltinIntegerWidth::arbitrary().parse(getDigitsText(), /*radix*/0,
+                                                isNegative());
 }
 
-static APInt getIntegerLiteralValue(bool IsNegative, StringRef Text,
-                                    unsigned BitWidth) {
-  llvm::APInt Value(BitWidth, 0);
-  bool Error = getAsInteger(Text, Value);
-  assert(!Error && "Invalid IntegerLiteral formed"); (void)Error;
-  if (IsNegative)
-    Value = -Value;
-  if (Value.getBitWidth() != BitWidth)
-    Value = Value.sextOrTrunc(BitWidth);
-  return Value;
+APInt IntegerLiteralExpr::getValue() const {
+  assert(!getType().isNull() && "Semantic analysis has not completed");
+  assert(!getType()->hasError() && "Should have a valid type");
+  auto width = getType()->castTo<BuiltinIntegerType>()->getWidth();
+  return width.parse(getDigitsText(), /*radix*/ 0, isNegative());
 }
 
-APInt IntegerLiteralExpr::getValue(StringRef Text, unsigned BitWidth, bool Negative) {
-  return getIntegerLiteralValue(Negative, Text, BitWidth);
-}
+APInt BuiltinIntegerWidth::parse(StringRef text, unsigned radix, bool negate,
+                                 bool *hadError) const {
+  if (hadError) *hadError = false;
 
-/// Returns the raw magnitude of the literal text without any truncation.
-APInt IntegerLiteralExpr::getRawMagnitude() const {
-  llvm::APInt Value(64, 0);
-  bool Error = getAsInteger(getDigitsText(), Value);
-  assert(!Error && "Invalid IntegerLiteral formed");
-  (void)Error;
-  return Value;
-}
+  // Parse an unsigned value from the string.
+  APInt value;
 
-APInt IntegerLiteralExpr::getValue() const {
-  assert(!getType().isNull() && "Semantic analysis has not completed");
-  assert(!getType()->hasError() && "Should have a valid type");
-  return getIntegerLiteralValue(
-      isNegative(), getDigitsText(),
-      getType()->castTo<BuiltinIntegerType>()->getGreatestWidth());
+  // Swift doesn't treat a leading zero as signifying octal, but
+  // StringRef::getAsInteger does.  Force decimal parsing in this case.
+  if (radix == 0 && text.size() >= 2 && text[0] == '0' && isdigit(text[1]))
+    radix = 10;
+
+  bool error = text.getAsInteger(radix, value);
+  if (error) {
+    if (hadError) *hadError = true;
+    return value;
+  }
+
+  // If we're producing an arbitrary-precision value, we don't need to do
+  // much additional processing.
+  if (isArbitraryWidth()) {
+    // The parser above always produces a non-negative value, so if the sign
+    // bit is set we need to give it some breathing room.
+    if (value.isNegative())
+      value = value.zext(value.getBitWidth() + 1);
+    assert(!value.isNegative());
+
+    // Now we can safely negate.
+    if (negate) {
+      value = -value;
+      assert(value.isNegative() || value.isNullValue());
+    }
+
+    // Truncate down to the minimum number of bits required to express
+    // this value exactly.
+    auto requiredBits = value.getMinSignedBits();
+    if (value.getBitWidth() > requiredBits)
+      value = value.trunc(requiredBits);
+
+  // If we have a fixed-width type (including abstract ones), we need to do
+  // fixed-width transformations, which can overflow.
+  } else {
+    unsigned width = getGreatestWidth();
+
+    // The overflow diagnostics in this case can't be fully correct because
+    // we don't know whether we're supposed to be producing a signed number
+    // or an unsigned one.
+
+    if (hadError && value.getActiveBits() > width)
+      *hadError = true;
+    value = value.zextOrTrunc(width);
+
+    if (negate) {
+      value = -value;
+
+      if (hadError && !value.isNegative())
+        *hadError = true;
+    }
+
+    assert(value.getBitWidth() == width);
+  }
+
+  return value;
 }
 
 static APFloat getFloatLiteralValue(bool IsNegative, StringRef Text,

lib/IRGen/GenEnum.cpp

@@ -1077,9 +1077,9 @@ namespace {
       auto intExpr = cast<IntegerLiteralExpr>(target->getRawValueExpr());
       auto intType = getDiscriminatorType();
 
-      APInt intValue = IntegerLiteralExpr::getValue(intExpr->getDigitsText(),
-                                                    intType->getBitWidth(),
-                                                    intExpr->isNegative());
+      APInt intValue =
+        BuiltinIntegerWidth::fixed(intType->getBitWidth())
+          .parse(intExpr->getDigitsText(), /*radix*/ 0, intExpr->isNegative());
 
       return intValue.getZExtValue();
     }

lib/Sema/CSGen.cpp

@@ -1220,17 +1220,7 @@ namespace {
         if (IntegerLiteralExpr *intLit = dyn_cast<IntegerLiteralExpr>(expr)) {
           unsigned maxWidth =
               maxIntType->castTo<BuiltinIntegerType>()->getGreatestWidth();
-          APInt magnitude = intLit->getRawMagnitude();
-          unsigned magWidth = magnitude.getActiveBits();
-          bool isNegative = intLit->isNegative();
-
-          // Compute the literal bit width in the signed two's complement form.
-          // This is generally one more than the magnitude width, but is the
-          // same when the literal is of the form -2^i (for some Nat `i`).
-          unsigned signedLitWidth =
-              (isNegative && (magnitude.countTrailingZeros() == magWidth - 1))
-                  ? magWidth
-                  : (magWidth + 1);
+          unsigned signedLitWidth = intLit->getRawValue().getMinSignedBits();
 
           if (signedLitWidth > maxWidth) { // overflow?
             CS.TC.diagnose(expr->getLoc(),

lib/Sema/TypeCheckSwitchStmt.cpp

@@ -870,12 +870,8 @@ namespace {
         return !cacheVal.second;
       }
       case ExprKind::IntegerLiteral: {
-        // FIXME: The magic number 128 is bad and we should actually figure out
-        // the bitwidth.  But it's too early in Sema to get it.
         auto *ILE = cast<IntegerLiteralExpr>(EL);
-        auto cacheVal =
-            IntLiteralCache.insert(
-                {ILE->getValue(ILE->getDigitsText(), 128, ILE->isNegative()), ILE});
+        auto cacheVal = IntLiteralCache.insert({ILE->getRawValue(), ILE});
         PrevPattern = (cacheVal.first != IntLiteralCache.end())
                     ? cacheVal.first->getSecond()
                     : nullptr;