Re: [musl] Wrong formatting of doubles?

[Markus Wichmann <nullplan@gmx.net>]

Am Wed, Jul 02, 2025 at 10:58:23AM -0400 schrieb Rich Felker:
> On Tue, Jul 01, 2025 at 09:46:10PM -0400, Rich Felker wrote:
> >  static int fmt_fp(FILE *f, long double y, int w, int p, int fl, int t, int ps)
> >  {
> > -	int bufsize = (ps==BIGLPRE)
> > -		? (LDBL_MANT_DIG+28)/29 + 1 +          // mantissa expansion
> > -		  (LDBL_MAX_EXP+LDBL_MANT_DIG+28+8)/9  // exponent expansion
> > -		: (DBL_MANT_DIG+28)/29 + 1 +
> > -		  (DBL_MAX_EXP+DBL_MANT_DIG+28+8)/9;
> > +	int max_mant_dig = (ps==BIGLPRE) ? LDBL_MANT_DIG : DBL_MANT_DIG;
> > +	int max_exp = (ps==BIGLPRE) ? LDBL_MAX_EXP : DBL_MAX_EXP;
> > +	int max_mant_slots = (max_mant_dig+28)/29 + 1;
> > +	int max_exp_slots = (max_exp+max_mant_dig+28+8)/9;
> > +	int bufsize = max_mant_slots + max_exp_slots;
> >  	uint32_t big[bufsize];
> >  	uint32_t *a, *d, *r, *z;
> >  	int e2=0, e, i, j, l;
> > @@ -266,7 +266,7 @@ static int fmt_fp(FILE *f, long double y, int w, int p, int fl, int t, int ps)
> >  	if (y) y *= 0x1p28, e2-=28;
> >  
> >  	if (e2<0) a=r=z=big;
> > -	else a=r=z=big+sizeof(big)/sizeof(*big) - LDBL_MANT_DIG - 1;
> > +	else a=r=z=big+sizeof(big)/sizeof(*big) - max_mant_slots - 1;
> >  
> >  	do {
> >  		*z = y;
> > -- 
> > 2.21.0
> > 
> 
> While nothing should change in the case of negative exponents, I also
> did an empirical check with DBL_TRUE_MIN to check the space for
> denormals, and we have 5 slots to spare.

That tracks with my calculations. I would just look at it from the
result: The buffer only needs to be big enough to contain the largest
end result, right? The largest decimal expansion occurs at TRUE_MIN,
which is 2^(MIN_EXP - MANT_DIG), which has MANT_DIG - MIN_EXP decimal
places. So you need enough slots for that many digits, +1 for the
pre-radix slot you always have, so in the end

bufsize = (mant_dig - min_exp + 8)/9 + 1

And that results in a little bit less than what you calculate.

type | yours | mine
-----|-------|-----
ld64 | 126   | 121
ld80 | 1835  | 1828
ld128| 1842  | 1833

> 
> The multiplication by 0x1p28 above is handled by the +28 in the
> expression for max_exp_slots.
> 

See, that is an intermediate result. The multiplication is not big
enough to make the mantissa expansion require more pre-radix slots, so
it doesn't change how much space is needed. It only changes how many
slots the mantissa expansion might take, which is not important to
printing TRUE_MIN. It is important to printing numbers with positive
exponents, but only for keeping the right distance from the end of the
buffer.

Ciao,
Markus