Fixes #114, #115: Special-casing powers close to the extremal ones.

* Special-casing the power-of-10 function for 10^308 and 10^-308.
* Special-casing `get_normalized_components()` to not strive for better precision for near-extremal powers and settle for the "rougher" way to obtain the integral and decimal parts.
* Added a test-case of a near-extremal-power which fails due to the re-normalization even if the powering is correct.

src/printf/printf.c

@@ -650,11 +650,20 @@ static struct scaling_factor update_normalization(struct scaling_factor sf, doub
   return result;
 }
 
-static struct double_components get_normalized_components(bool negative, printf_size_t precision, double non_normalized, struct scaling_factor normalization)
+static struct double_components get_normalized_components(bool negative, printf_size_t precision, double non_normalized, struct scaling_factor normalization, int floored_exp10)
 {
   struct double_components components;
   components.is_negative = negative;
-  components.integral = (int_fast64_t) apply_scaling(non_normalized, normalization);
+  double scaled = apply_scaling(non_normalized, normalization);
+
+  bool close_to_representation_extremum = ( (-floored_exp10 + (int) precision) >= DBL_MAX_10_EXP - 1 );
+  if (close_to_representation_extremum) {
+    // We can't have a normalization factor which also accounts for the precision, i.e. moves
+    // some decimal digits into the mantissa, since it's unrepresentable, or nearly unrepresentable.
+    // So, we'll give up early on getting extra precision...
+    return get_components(negative ? -scaled : scaled, precision);
+  }
+  components.integral = (int_fast64_t) scaled;
   double remainder = non_normalized - unapply_scaling((double) components.integral, normalization);
   double prec_power_of_10 = powers_of_10[precision];
   struct scaling_factor account_for_precision = update_normalization(normalization, prec_power_of_10);
@@ -819,6 +828,8 @@ static double log10_of_positive(double positive_number)
 
 static double pow10_of_int(int floored_exp10)
 {
+  if (floored_exp10 == 308) { return 1e308; }
+  if (floored_exp10 == -308) { return 1e-308; }
   // Compute 10^(floored_exp10) but (try to) make sure that doesn't overflow
   double_with_bit_access dwba;
   int exp2 = bastardized_floor(floored_exp10 * 3.321928094887362 + 0.5);
@@ -887,7 +898,7 @@ static void print_exponential_number(output_gadget_t* output, double number, pri
   struct double_components decimal_part_components =
     should_skip_normalization ?
     get_components(negative ? -abs_number : abs_number, precision) :
-    get_normalized_components(negative, precision, abs_number, normalization);
+    get_normalized_components(negative, precision, abs_number, normalization, floored_exp10);
 
   // Account for roll-over, e.g. rounding from 9.99 to 100.0 - which effects
   // the exponent and may require additional tweaking of the parts

test/test_suite.cpp

@@ -1012,6 +1012,9 @@ TEST_CASE("floating-point specifiers, precision and flags", "[]" ) {
   PRINTING_CHECK("-4e+04",           ==, sprintf_, buffer, "%.1g", -40661.5);
   PRINTING_CHECK("-4.e+04",          ==, sprintf_, buffer, "%#.1g", -40661.5);
   PRINTING_CHECK("100.",             ==, sprintf_, buffer, "%#.3g", 99.998580932617187500);
+  // Ensure high precision does make us exceed the representation threshold
+  // internally (e.g. with normalization factors and such
+  PRINTING_CHECK("1.2345678901e-308", ==, sprintf_, buffer, "%.10e", 1.2345678901e-308);
   // Rounding-focused checks
   PRINTING_CHECK("4.895512e+04",     ==, sprintf_, buffer, "%e", 48955.125);
   PRINTING_CHECK("9.2524e+04",       ==, sprintf_, buffer, "%.4e", 92523.5);