fix the last known rounding bug in floating point printing
authorRich Felker <dalias@aerifal.cx>
Wed, 11 May 2011 23:58:03 +0000 (19:58 -0400)
committerRich Felker <dalias@aerifal.cx>
Wed, 11 May 2011 23:58:03 +0000 (19:58 -0400)
the observed symptom was that the code was incorrectly rounding up
1.0625 to 1.063 despite the rounding mode being round-to-nearest with
ties broken by rounding to even last place. however, the code was just
not right in many respects, and i'm surprised it worked as well as it
did. this time i tested the values that end up in the variables round,
small, and the expression round+small, and all look good.

src/stdio/vfprintf.c

index f19058d3c0cb04a97ca81c0befec93c838cab172..f13fbe10576d157684ecfc9d4e8bcf448010cd6d 100644 (file)
@@ -326,9 +326,10 @@ static int fmt_fp(FILE *f, long double y, int w, int p, int fl, int t)
                if (x || d+1!=z) {
                        long double round = CONCAT(0x1p,LDBL_MANT_DIG);
                        long double small;
-                       if (x<i/2) small=0x01p-1;
-                       else if (i==i/2 && d+1==z) small=0x10p-1;
-                       else small=0x11p-1;
+                       if (*d/i & 1) round += 2;
+                       if (x<i/2) small=0x0.8p0;
+                       else if (x==i/2 && d+1==z) small=0x1.0p0;
+                       else small=0x1.8p0;
                        if (pl && *prefix=='-') round*=-1, small*=-1;
                        *d -= x;
                        /* Decide whether to round by probing round+small */