[AsmWriter/APFloat] FP constant printing: Avoid usage of locale dependent snprinf

[skatkov]

Summary

This should fix the bug https://bugs.llvm.org/show_bug.cgi?id=12906

To print the FP constant AsmWriter does the following:

1. convert FP value to String (actually using snprintf function which is locale dependent).
2. Convert String back to FP Value
3. Compare original and got FP values. If they are not equal just dump as hex.

The problem happens on the 2nd step when APFloat does not expect group delimiter or
fraction delimiter other than period symbol and so on, which can be produced on the
first step if LLVM library is used in an environment with corresponding locale set.

To fix this issue the locale independent APFloat:toString function is used.
However it prints FP values slightly differently than snprintf does. Specifically
it suppress trailing zeros in significant, use capital E and so on.
It results in 117 test failures during make check.
To avoid this I've also updated APFloat.toString a bit to pass make check at least.

[skatkov]

lib/IR/AsmWriter.cpp

1125–1126

Do we really need this check? To me it can be safely removed. Does anyone see the reason not to do that?

[timshen]

Have you considered making llvm::write_double() not to follow the user locale?

include/llvm/ADT/APFloat.h

415

Can you document TruncateZero?

lib/IR/AsmWriter.cpp

1125–1126

Alternatively we can turn it into an assertion.

[skatkov]

Actually I considered the re-writing llvm::write_double() to be locale independent. I decided to go with solution due to:

1. to be consistent with Lexer which also uses APFloat
2. llvm::write_double() is used in raw_ostream which might be widely used and may be someone really want it to be locale specific.
3. Actually re-writing llvm::write_double() will require me to just re-use printf functionality for double just only remove locale specific code. I'm not sure about licensing but it is a big piece of code to write it from scratch. However most of the code should look like APFloat::toString.

So I decided to go with this solution. Does it make sense?

[rnk]

In D32276#731864, @skatkov wrote:

Actually I considered the re-writing llvm::write_double() to be locale independent. I decided to go with solution due to:

1. to be consistent with Lexer which also uses APFloat

To me, this is a really good reason to do it this way.

Looks good to me. Tim, did you have anything else?

[timshen]

LGTM!

include/llvm/ADT/APFloat.h

public:
   ///   zero padding will not be used unless the value is
   ///   integral and small enough to be expressed in
   ///   FormatPrecision digits.  0 means to use the natural
   ///   precision of the number.
   /// \param FormatMaxPadding The maximum number of zeros to
   ///   consider inserting before falling back to scientific
   ///   notation.  0 means to always use scientific notation.
   ///
+  /// \param TruncateZero Indicate whether to remove the trailing zero in
+  ///   fraction part or not. Also setting this parameter to false forcing
+  ///   producing of output more similar to default printf behavior.
+  ///   Specifically the lower e is used as exponent delimiter and exponent
+  ///   always contains no less than two digits.
+  ///
   /// Number       Precision    MaxPadding      Result
   /// ------       ---------    ----------      ------
   /// 1.01E+4              5             2       10100
   /// 1.01E+4              4             2       1.01E+4
   /// 1.01E+4              5             1       1.01E+4
   /// 1.01E-2              5             2       0.0101
   /// 1.01E-2              4             2       0.0101
   /// 1.01E-2              4             1       1.01E-2
   void toString(SmallVectorImpl<char> &Str, unsigned FormatPrecision = 0,
-                unsigned FormatMaxPadding = 3) const;
+                unsigned FormatMaxPadding = 3, bool TruncateZero = true) const;
 
   /// If this value has an exact multiplicative inverse, store it in inv and
   /// return true.
   bool getExactInverse(APFloat *inv) const;
 
   /// Returns the exponent of the internal representation of the APFloat.
   ///
   /// Because the radix of APFloat is 2, this is equivalent to floor(log2(x)).
unsigned int convertToHexString(char *DST, unsigned int HexDigits,
                                   bool UpperCase, roundingMode RM) const;
 
   bool isDenormal() const;
   bool isSmallest() const;
   bool isLargest() const;
   bool isInteger() const;
 
   void toString(SmallVectorImpl<char> &Str, unsigned FormatPrecision,
-                unsigned FormatMaxPadding) const;
+                unsigned FormatMaxPadding, bool TruncateZero = true) const;
 
   bool getExactInverse(APFloat *inv) const;
 
   friend int ilogb(const DoubleAPFloat &Arg);
   friend DoubleAPFloat scalbn(DoubleAPFloat X, int Exp, roundingMode);
   friend DoubleAPFloat frexp(const DoubleAPFloat &X, int &Exp, roundingMode);
   friend hash_code hash_value(const DoubleAPFloat &Arg);
 };
public:
   bool isSmallest() const { APFLOAT_DISPATCH_ON_SEMANTICS(isSmallest()); }
   bool isLargest() const { APFLOAT_DISPATCH_ON_SEMANTICS(isLargest()); }
   bool isInteger() const { APFLOAT_DISPATCH_ON_SEMANTICS(isInteger()); }
 
   APFloat &operator=(const APFloat &RHS) = default;
   APFloat &operator=(APFloat &&RHS) = default;
 
   void toString(SmallVectorImpl<char> &Str, unsigned FormatPrecision = 0,
-                unsigned FormatMaxPadding = 3) const {
+                unsigned FormatMaxPadding = 3, bool TruncateZero = true) const {
     APFLOAT_DISPATCH_ON_SEMANTICS(
-        toString(Str, FormatPrecision, FormatMaxPadding));
+        toString(Str, FormatPrecision, FormatMaxPadding, TruncateZero));
   }
 
   void print(raw_ostream &) const;
   void dump() const;
 
   bool getExactInverse(APFloat *inv) const {
     APFLOAT_DISPATCH_ON_SEMANTICS(getExactInverse(inv));
   }

lib/IR/AsmWriter.cpp

if (CI->getType()->isIntegerTy(1)) {
       Out << (CI->getZExtValue() ? "true" : "false");
       return;
     }
     Out << CI->getValue();
     return;
   }
 
   if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CV)) {
-    if (&CFP->getValueAPF().getSemantics() == &APFloat::IEEEsingle() ||
-        &CFP->getValueAPF().getSemantics() == &APFloat::IEEEdouble()) {
+    const APFloat &APF = CFP->getValueAPF();
+    if (&APF.getSemantics() == &APFloat::IEEEsingle() ||
+        &APF.getSemantics() == &APFloat::IEEEdouble()) {
       // We would like to output the FP constant value in exponential notation,
       // but we cannot do this if doing so will lose precision.  Check here to
       // make sure that we only output it in exponential format if we can parse
       // the value back and get the same value.
       //
       bool ignored;
-      bool isDouble = &CFP->getValueAPF().getSemantics()==&APFloat::IEEEdouble();
-      bool isInf = CFP->getValueAPF().isInfinity();
-      bool isNaN = CFP->getValueAPF().isNaN();
+      bool isDouble = &APF.getSemantics() == &APFloat::IEEEdouble();
+      bool isInf = APF.isInfinity();
+      bool isNaN = APF.isNaN();
       if (!isInf && !isNaN) {
-        double Val = isDouble ? CFP->getValueAPF().convertToDouble() :
-                                CFP->getValueAPF().convertToFloat();
+        double Val = isDouble ? APF.convertToDouble() : APF.convertToFloat();
         SmallString<128> StrVal;
-        raw_svector_ostream(StrVal) << Val;
-
+        APF.toString(StrVal, 6, 0, false);
         // Check to make sure that the stringized number is not some string like
         // "Inf" or NaN, that atof will accept, but the lexer will not.  Check
         // that the string matches the "[-+]?[0-9]" regex.
         //
-        if ((StrVal[0] >= '0' && StrVal[0] <= '9') ||
+        assert((StrVal[0] >= '0' && StrVal[0] <= '9') ||
                ((StrVal[0] == '-' || StrVal[0] == '+') &&
-             (StrVal[1] >= '0' && StrVal[1] <= '9'))) {
-          // Reparse stringized version!
-          if (APFloat(APFloat::IEEEdouble(), StrVal).convertToDouble() == Val) {
-            Out << StrVal;
-            return;
-          }
-        }
-      }
+                (StrVal[1] >= '0' && StrVal[1] <= '9')) &&
+                   "[-+]?[0-9] regex does not match!");
       // Otherwise we could not reparse it to exactly the same value, so we must
       // output the string in hexadecimal format!  Note that loading and storing
       // floating point types changes the bits of NaNs on some hosts, notably
       // x86, so we must not use these types.
       static_assert(sizeof(double) == sizeof(uint64_t),
                     "assuming that double is 64 bits!");
-      APFloat apf = CFP->getValueAPF();
+      APFloat apf = APF;
       // Floats are represented in ASCII IR as double, convert.
       if (!isDouble)
         apf.convert(APFloat::IEEEdouble(), APFloat::rmNearestTiesToEven,
                           &ignored);
       Out << format_hex(apf.bitcastToAPInt().getZExtValue(), 0, /*Upper=*/true);
       return;
     }
 
     // Either half, or some form of long double.
     // These appear as a magic letter identifying the type, then a
     // fixed number of hex digits.
     Out << "0x";
-    APInt API = CFP->getValueAPF().bitcastToAPInt();
-    if (&CFP->getValueAPF().getSemantics() == &APFloat::x87DoubleExtended()) {
+    APInt API = APF.bitcastToAPInt();
+    if (&APF.getSemantics() == &APFloat::x87DoubleExtended()) {
       Out << 'K';
       Out << format_hex_no_prefix(API.getHiBits(16).getZExtValue(), 4,
                                   /*Upper=*/true);
       Out << format_hex_no_prefix(API.getLoBits(64).getZExtValue(), 16,
                                   /*Upper=*/true);
       return;
-    } else if (&CFP->getValueAPF().getSemantics() == &APFloat::IEEEquad()) {
+    } else if (&APF.getSemantics() == &APFloat::IEEEquad()) {
       Out << 'L';
       Out << format_hex_no_prefix(API.getLoBits(64).getZExtValue(), 16,
                                   /*Upper=*/true);
       Out << format_hex_no_prefix(API.getHiBits(64).getZExtValue(), 16,
                                   /*Upper=*/true);
-    } else if (&CFP->getValueAPF().getSemantics() == &APFloat::PPCDoubleDouble()) {
+    } else if (&APF.getSemantics() == &APFloat::PPCDoubleDouble()) {
       Out << 'M';
       Out << format_hex_no_prefix(API.getLoBits(64).getZExtValue(), 16,
                                   /*Upper=*/true);
       Out << format_hex_no_prefix(API.getHiBits(64).getZExtValue(), 16,
                                   /*Upper=*/true);
-    } else if (&CFP->getValueAPF().getSemantics() == &APFloat::IEEEhalf()) {
+    } else if (&APF.getSemantics() == &APFloat::IEEEhalf()) {
       Out << 'H';
       Out << format_hex_no_prefix(API.getZExtValue(), 4,
                                   /*Upper=*/true);
     } else
       llvm_unreachable("Unsupported floating point type");
     return;
   }
 

lib/Support/APFloat.cpp

void AdjustToPrecision(SmallVectorImpl<char> &buffer,
     }
 
     exp += FirstSignificant;
     buffer.erase(&buffer[0], &buffer[FirstSignificant]);
   }
 }
 
 void IEEEFloat::toString(SmallVectorImpl<char> &Str, unsigned FormatPrecision,
-                         unsigned FormatMaxPadding) const {
+                         unsigned FormatMaxPadding, bool TruncateZero) const {
   switch (category) {
   case fcInfinity:
     if (isNegative())
       return append(Str, "-Inf");
     else
       return append(Str, "+Inf");
 
   case fcNaN: return append(Str, "NaN");
 
   case fcZero:
     if (isNegative())
       Str.push_back('-');
 
-    if (!FormatMaxPadding)
+    if (!FormatMaxPadding) {
+      if (TruncateZero)
         append(Str, "0.0E+0");
-    else
+      else {
+        append(Str, "0.0");
+        if (FormatPrecision > 1)
+          Str.append(FormatPrecision - 1, '0');
+        append(Str, "e+00");
+      }
+    } else
       Str.push_back('0');
     return;
 
   case fcNormal:
     break;
   }
 
   if (isNegative())
void IEEEFloat::toString(SmallVectorImpl<char> &Str, unsigned FormatPrecision,
   }
 
   // Scientific formatting is pretty straightforward.
   if (FormatScientific) {
     exp += (NDigits - 1);
 
     Str.push_back(buffer[NDigits-1]);
     Str.push_back('.');
-    if (NDigits == 1)
+    if (NDigits == 1 && TruncateZero)
       Str.push_back('0');
     else
       for (unsigned I = 1; I != NDigits; ++I)
         Str.push_back(buffer[NDigits-1-I]);
-    Str.push_back('E');
+    // Fill with zeros up to FormatPrecision.
+    if (!TruncateZero && FormatPrecision > NDigits - 1)
+      Str.append(FormatPrecision - NDigits + 1, '0');
+    // For !TruncateZero we use lower 'e'.
+    Str.push_back(TruncateZero ? 'E' : 'e');
 
     Str.push_back(exp >= 0 ? '+' : '-');
     if (exp < 0) exp = -exp;
     SmallVector<char, 6> expbuf;
     do {
       expbuf.push_back((char) ('0' + (exp % 10)));
       exp /= 10;
     } while (exp);
+    // Exponent always at least two digits if we do not truncate zeros.
+    if (!TruncateZero && expbuf.size() < 2)
+      expbuf.push_back('0');
     for (unsigned I = 0, E = expbuf.size(); I != E; ++I)
       Str.push_back(expbuf[E-1-I]);
     return;
   }
 
   // Non-scientific, positive exponents.
   if (exp >= 0) {
     for (unsigned I = 0; I != NDigits; ++I)
bool DoubleAPFloat::isInteger() const {
   APFloat Tmp(semPPCDoubleDoubleLegacy);
   (void)Tmp.add(Floats[0], rmNearestTiesToEven);
   (void)Tmp.add(Floats[1], rmNearestTiesToEven);
   return Tmp.isInteger();
 }
 
 void DoubleAPFloat::toString(SmallVectorImpl<char> &Str,
                              unsigned FormatPrecision,
-                             unsigned FormatMaxPadding) const {
+                             unsigned FormatMaxPadding,
+                             bool TruncateZero) const {
   assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics");
   APFloat(semPPCDoubleDoubleLegacy, bitcastToAPInt())
-      .toString(Str, FormatPrecision, FormatMaxPadding);
+      .toString(Str, FormatPrecision, FormatMaxPadding, TruncateZero);
 }
 
 bool DoubleAPFloat::getExactInverse(APFloat *inv) const {
   assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics");
   APFloat Tmp(semPPCDoubleDoubleLegacy, bitcastToAPInt());
   if (!inv)
     return Tmp.getExactInverse(nullptr);
   APFloat Inv(semPPCDoubleDoubleLegacy);

unittests/ADT/APFloatTest.cpp

 using namespace llvm;
 
 static double convertToDoubleFromString(const char *Str) {
   llvm::APFloat F(0.0);
   F.convertFromString(Str, llvm::APFloat::rmNearestTiesToEven);
   return F.convertToDouble();
 }
 
-static std::string convertToString(double d, unsigned Prec, unsigned Pad) {
+static std::string convertToString(double d, unsigned Prec, unsigned Pad,
+                                   bool Tr = true) {
   llvm::SmallVector<char, 100> Buffer;
   llvm::APFloat F(d);
-  F.toString(Buffer, Prec, Pad);
+  F.toString(Buffer, Prec, Pad, Tr);
   return std::string(Buffer.data(), Buffer.size());
 }
 
 namespace {
 
 TEST(APFloatTest, isSignaling) {
   // We test qNaN, -qNaN, +sNaN, -sNaN with and without payloads. *NOTE* The
   // positive/negative distinction is included only since the getQNaN/getSNaN
TEST(APFloatTest, toString) {
   ASSERT_EQ("0.0101", convertToString(1.01E-2, 5, 2));
   ASSERT_EQ("0.0101", convertToString(1.01E-2, 4, 2));
   ASSERT_EQ("1.01E-2", convertToString(1.01E-2, 5, 1));
   ASSERT_EQ("0.78539816339744828", convertToString(0.78539816339744830961, 0, 3));
   ASSERT_EQ("4.9406564584124654E-324", convertToString(4.9406564584124654e-324, 0, 3));
   ASSERT_EQ("873.18340000000001", convertToString(873.1834, 0, 1));
   ASSERT_EQ("8.7318340000000001E+2", convertToString(873.1834, 0, 0));
   ASSERT_EQ("1.7976931348623157E+308", convertToString(1.7976931348623157E+308, 0, 0));
+  ASSERT_EQ("10", convertToString(10.0, 6, 3, false));
+  ASSERT_EQ("1.000000e+01", convertToString(10.0, 6, 0, false));
+  ASSERT_EQ("10100", convertToString(1.01E+4, 5, 2, false));
+  ASSERT_EQ("1.0100e+04", convertToString(1.01E+4, 4, 2, false));
+  ASSERT_EQ("1.01000e+04", convertToString(1.01E+4, 5, 1, false));
+  ASSERT_EQ("0.0101", convertToString(1.01E-2, 5, 2, false));
+  ASSERT_EQ("0.0101", convertToString(1.01E-2, 4, 2, false));
+  ASSERT_EQ("1.01000e-02", convertToString(1.01E-2, 5, 1, false));
+  ASSERT_EQ("0.78539816339744828",
+            convertToString(0.78539816339744830961, 0, 3, false));
+  ASSERT_EQ("4.94065645841246540e-324",
+            convertToString(4.9406564584124654e-324, 0, 3, false));
+  ASSERT_EQ("873.18340000000001", convertToString(873.1834, 0, 1, false));
+  ASSERT_EQ("8.73183400000000010e+02", convertToString(873.1834, 0, 0, false));
+  ASSERT_EQ("1.79769313486231570e+308",
+            convertToString(1.7976931348623157E+308, 0, 0, false));
 }
 
 TEST(APFloatTest, toInteger) {
   bool isExact = false;
   APSInt result(5, /*isUnsigned=*/true);
 
   EXPECT_EQ(APFloat::opOK,
             APFloat(APFloat::IEEEdouble(), "10")