std.fmt.parseFloat does not parse 9007199254740993.0 correctly

[andrewrk]

Implementation is here: https://github.com/ziglang/zig/blob/master/lib/std/fmt/parse_float.zig

--- a/lib/std/fmt/parse_float.zig
+++ b/lib/std/fmt/parse_float.zig
@@ -425,4 +425,6 @@ test "fmt.parseFloat" {
             try expect(approxEqAbs(T, try parseFloat(T, "2.71828182845904523536"), @as(T, 2.718281828459045), epsilon));
         }
     }
+
+    try expectEqual(try parseFloat(f128, "9007199254740993.0"), 9007199254740993.0);
 }

$ ./zig test ../lib/std/fmt/parse_float.zig 
Test [1/1] test "parseFloat"... expected 9.007199254740994e+15, found 9.007199254740992e+15
Test [1/1] test "parseFloat"... FAIL (TestExpectedEqual)

This is causing a behavior test failure in the self-hosted compiler, which relies on std.fmt.parseFloat:

zig/test/behavior/floatop.zig

Line 679 in 9eceba2

try expect(9007199254740992.0 + 1.0 == 9007199254740993.0);

[curtiswilkinson]

I've done a little bit of digging:

This code was ported from https://github.com/grzegorz-kraszewski/stringtofloat

When I run that in C on our example float string I'm getting the 4 at the end that we're seeing with the zig port (vs the 2 from stage1, which I believe uses a port of musl for float parsing?).

int main() {
  printf("%fn\n", str2dbl("9007199254740993.0"));
  return 0;
};

This could potentially be what the important comment in the C reference implementation is talking about: https://github.com/grzegorz-kraszewski/stringtofloat/blob/master/str2dbl.c#L7

/* The code works in "round towards zero" mode. This is different from  */
/* GCC standard library strtod(), which uses "round half to even" rule. */
/* Therefore it cannot be used as a direct drop-in replacement, as in   */
/* some cases results will be different on the least significant bit of */
/* mantissa. Read more in the README.md file.

[andrewrk]

Here's the best path forward in my opinion:

• ditch grzegorz-kraszewski/stringtofloat, forget about it
• port musl's f128 parsing code. I already did half the work here: parse_f128.c, it just needs to be ported to zig

[danielchasehooper]

I'm working on this now. Unable to assign it to myself.

[matu3ba]

@danielchasehooper I have a wip though I should have used more vim macros for porting to get it over finish line:
https://gist.github.com/matu3ba/040ef3ad23f53213ad277999eb847586

Feel free to continue or rip out what you need.