Use dragonbox to print floats (#481)

* add new float printer

* more includes

* compare as floats

common/CMakeLists.txt

@@ -22,6 +22,7 @@ add_library(common
         util/FileUtil.cpp
         util/json_util.cpp
         util/Timer.cpp
+        util/print_float.cpp
         )
 
 target_link_libraries(common fmt lzokay replxx)

common/goos/Object.cpp

@@ -42,6 +42,7 @@
 #include "Object.h"
 #include "common/util/FileUtil.h"
 #include "third-party/fmt/core.h"
+#include "common/util/print_float.h"
 
 namespace goos {
 
@@ -83,51 +84,13 @@ std::string object_type_to_string(ObjectType type) {
 }
 
 /*!
- * Special case to print a float with the %g format specifier
+ * Special case to print a float
  */
 template <>
 std::string fixed_to_string(FloatType x) {
-  char buff[256];
-  s64 rounded = x;
-  bool exact_int = ((float)rounded) == x;
-
-  // it's an integer number, so let's just get this over with asap
-  if (exact_int) {
-    sprintf(buff, "%" PRId64 ".0", rounded);
-    return {buff};
-  } else {
-    // not an integer - see how many decimal cases we need
-    // i'm not sure what happens if x is a NaN/inf...
-
-    // buffer for format string
-    char fmt_buf[256];
-
-    // we are going to try our hardest to make sure the output is re-parseable
-    // for what it's worth, the lowest representable 32-bit floating point number has almost 50
-    // decimal cases, although:
-    // - by that point we should just be using scientific notation instead?
-    // - the PS2 DOES NOT DENORMALIZE FLOATS, or handle NaNs/infs! so the representation wouldn't
-    // be accurate anyway
-    // - we might not ever encounter numbers like that. the pretty printer has a "banned" floats
-    // list just in case
-
-    // 99 might seem high, but we are gonna get the result in under 10 or so most of the time, so
-    // it's fine.
-    // maybe there's some math principles or other tricks to optimize this?
-    for (int i = 1; i <= 99; ++i) {
-      // generate the format string
-      sprintf(fmt_buf, "%%.%df", i);
-      sprintf(buff, fmt_buf, x);
-
-      auto float_from_string = float(std::stod(buff));
-      float value_as_float = float(x);
-      bool are_they_equal = float_from_string == value_as_float;
-
-      if (are_they_equal)
-        return {buff};
-    }
-    throw std::runtime_error("a float could not be represented accurately");
-  }
+  auto result = float_to_string(x);
+  assert((float)x == (float)std::stod(result));
+  return result;
 }
 
 /*!

common/util/print_float.cpp

@@ -0,0 +1,130 @@
+#include <cmath>
+
+#include "third-party/fmt/core.h"
+#include "common/common_types.h"
+#include "third-party/dragonbox.h"
+#include "print_float.h"
+
+/*!
+ * Convert a float to a string. The string is _always_ in this format:
+ * [negative_sign] [at least 1 digit] [decimal point] [at least 1 digit]
+ * and, if you trust the dragonbox library, should be the shortest possible representation
+ * that round-trips through a properly implemented string -> float conversion.
+ */
+std::string float_to_string(float value) {
+  constexpr int buff_size = 128;
+  char buff[buff_size];
+  float_to_cstr(value, buff);
+  return {buff};
+}
+
+int float_to_cstr(float value, char* buffer) {
+  assert(std::isfinite(value));
+  // dragonbox gives us:
+  //  - an integer, representing the decimal value
+  //  - sign
+  //  - exponent
+
+  int i = 0;
+
+  // the exponent/significand representation of dragonbox is ambiguous with how it represents 0,
+  // so just handle that as a special case
+  if (value == 0) {
+    buffer[i++] = '0';
+    buffer[i++] = '.';
+    buffer[i++] = '0';
+    buffer[i++] = '\0';
+  }
+
+  auto decimal = jkj::dragonbox::to_decimal(value);
+
+  // in all cases, we need to convert the decimal to characters.
+  char digit_buff[64];
+  int num_digits = 0;
+  u64 significand = decimal.significand;
+  while (significand) {
+    digit_buff[num_digits++] = '0' + (significand % 10);
+    significand /= 10;
+  }
+
+  if (decimal.exponent >= 0) {
+    // needs 0 or more trailing zeros before decimal (no nonzeros after decimal).
+
+    // print in four parts:
+    // negative sign | digits | 000's | .0
+
+    // part 1
+    if (decimal.is_negative) {
+      buffer[i++] = '-';
+    }
+
+    // part 2
+    for (int digit = 0; digit < num_digits; digit++) {
+      buffer[i++] = digit_buff[num_digits - (digit + 1)];
+    }
+
+    // part 3
+    for (int j = 0; j < decimal.exponent; j++) {
+      buffer[i++] = '0';
+    }
+
+    // part 4
+    buffer[i++] = '.';
+    buffer[i++] = '0';
+    buffer[i++] = '\0';
+  } else {
+    // some nonzero digits after decimal.
+
+    if (num_digits <= -decimal.exponent) {
+      // all after the decimal
+      // negative sign | 0. | 000's | digits
+
+      // part 1
+      if (decimal.is_negative) {
+        buffer[i++] = '-';
+      }
+
+      // part 2
+      buffer[i++] = '0';
+      buffer[i++] = '.';
+
+      // part 3
+      int zeros = -decimal.exponent - num_digits;
+      for (int j = 0; j < zeros; j++) {
+        buffer[i++] = '0';
+      }
+
+      // part 4
+      for (int digit = 0; digit < num_digits; digit++) {
+        buffer[i++] = digit_buff[num_digits - (digit + 1)];
+      }
+      buffer[i++] = '\0';
+
+    } else {
+      // some before, some after.
+      // negative sign | digits | . | digits
+
+      // part 1
+      if (decimal.is_negative) {
+        buffer[i++] = '-';
+      }
+
+      // part 2
+      int digits_before_decimal = num_digits + decimal.exponent;
+      int digit = 0;
+      for (; digit < digits_before_decimal; digit++) {
+        buffer[i++] = digit_buff[num_digits - (digit + 1)];
+      }
+
+      // part 3
+      buffer[i++] = '.';
+
+      // part 4
+      for (; digit < num_digits; digit++) {
+        buffer[i++] = digit_buff[num_digits - (digit + 1)];
+      }
+      buffer[i++] = '\0';
+    }
+  }
+  return i;
+}

common/util/print_float.h

@@ -0,0 +1,6 @@
+#pragma once
+
+#include <string>
+
+std::string float_to_string(float value);
+int float_to_cstr(float value, char* buffer);

test/test_common_util.cpp

@@ -1,5 +1,6 @@
 #include <string>
 #include <vector>
+#include <limits>
 
 #include "common/util/FileUtil.h"
 #include "common/util/Trie.h"
@@ -9,6 +10,7 @@
 #include "common/util/json_util.h"
 #include "common/util/Range.h"
 #include "third-party/fmt/core.h"
+#include "common/util/print_float.h"
 
 TEST(CommonUtil, get_file_path) {
   std::vector<std::string> test = {"cabbage", "banana", "apple"};
@@ -95,3 +97,38 @@ TEST(CommonUtil, BitRange) {
   EXPECT_EQ(get_bit_range(UINT64_MAX - 1), Range<int>(1, 64));
   EXPECT_EQ(get_bit_range(UINT64_MAX / 2), Range<int>(0, 63));
 }
+
+TEST(CommonUtil, FloatToString) {
+  float test_floats[] = {0.f,
+                         1.f,
+                         -1.f,
+                         0.1f,
+                         -0.1f,
+                         1234,
+                         12340,
+                         123400,
+                         -1234000,
+                         0.00342f,
+                         -0.003423f,
+                         std::numeric_limits<float>::min(),
+                         std::numeric_limits<float>::max(),
+                         std::numeric_limits<float>::lowest(),
+                         std::numeric_limits<float>::epsilon(),
+                         std::numeric_limits<float>::denorm_min(),
+                         -std::numeric_limits<float>::min(),
+                         -std::numeric_limits<float>::max(),
+                         -std::numeric_limits<float>::lowest(),
+                         -std::numeric_limits<float>::epsilon(),
+                         -std::numeric_limits<float>::denorm_min()};
+
+  for (auto x : test_floats) {
+    EXPECT_TRUE(x == (float)std::stod(float_to_string(x)));
+  }
+
+  // all three of these constants should become _exactly_ 1460961.25 when converted to a float.
+  // to break a tie, dragonbox defaults to round to even, which is nice because that's the
+  // default rounding mode.
+  EXPECT_EQ("1460961.2", float_to_string(1460961.25));
+  EXPECT_EQ("1460961.2", float_to_string(1460961.20));
+  EXPECT_EQ("1460961.2", float_to_string(1460961.30));
+}