Work around strtod("nan") behavior on Windows.

Windows -- including Cygwin -- return a NaN representation with the sign bit set, whereas other platforms return one without the sign bit set. This results in different binary output in the JSON-to-binary part of capnp-test.sh.

This was uncovered since we're now runing capnp-test.sh for Cygwin builds, but the problem actually applies to *all* Windows builds. We just weren't running this test for the other builds before.

Interestingly, I *had* run this test with a MinGW cross-compile build running on WINE, and it passed there. Seemingly, WINE's strtod() directly wraps the host system's without trying to emulate Windows differences.

c++/src/kj/string-test.c++

@@ -216,6 +216,21 @@ KJ_TEST("kj::delimited() and kj::strPreallocated()") {
   }
 }
 
+KJ_TEST("parsing 'nan' returns canonical NaN value") {
+  // There are many representations of NaN. We would prefer that parsing "NaN" produces exactly the
+  // same bits that kj::nan() returns.
+  {
+    double parsedNan = StringPtr("NaN").parseAs<double>();
+    double canonicalNan = kj::nan();
+    KJ_EXPECT(memcmp(&parsedNan, &canonicalNan, sizeof(parsedNan)) == 0);
+  }
+  {
+    float parsedNan = StringPtr("NaN").parseAs<float>();
+    float canonicalNan = kj::nan();
+    KJ_EXPECT(memcmp(&parsedNan, &canonicalNan, sizeof(parsedNan)) == 0);
+  }
+}
+
 }  // namespace
 }  // namespace _ (private)
 }  // namespace kj

c++/src/kj/string.c++

@@ -82,6 +82,19 @@ double parseDouble(const StringPtr& s) {
   errno = 0;
   auto value = strtod(s.begin(), &endPtr);
   KJ_REQUIRE(endPtr == s.end(), "String does not contain valid floating number", s) { return 0; }
+#if _WIN32 || __CYGWIN__
+  // When Windows' strtod() parses "nan", it returns a value with the sign bit set. But, our
+  // preferred canonical value for NaN does not have the sign bit set, and all other platforms
+  // return one without the sign bit set. So, on Windows, detect NaN and return our preferred
+  // version.
+  //
+  // Cygwin seemingly does not try to emulate Linux behavior here, but rather allows Windows'
+  // behavior to leak through.
+  if (isNaN(value)) {
+    // NaN
+    return kj::nan();
+  }
+#endif
   return value;
 }