1 |
On Tue, 2022-10-18 at 10:14 +0200, Ulrich Mueller wrote: |
2 |
> > > > > > On Tue, 18 Oct 2022, Mike Gilbert wrote: |
3 |
> |
4 |
> > Reference: https://gcc.gnu.org/wiki/x87note |
5 |
> |
6 |
> Which says: |
7 |
> |
8 |
> > ... the amount of worst-case error that could possibly happen using |
9 |
> > the x87 (with any amount of intermediate rounding) is at worst the |
10 |
> > same as true 64 or 32 bit arithmetic, and in practice is almost |
11 |
> > always |
12 |
> > better. |
13 |
> |
14 |
> and: |
15 |
> |
16 |
> > Note, however, that this greater repeatability comes at the cost of |
17 |
> > lost precision (i.e. SSE always gets the same precision because it |
18 |
> > always takes the equivalent of the x87's worst case: a forced round |
19 |
> > down at each step). |
20 |
> |
21 |
> So, it comes with a price, and I wonder if we shouldn't leave that |
22 |
> choice to the user, and go with the upstream GCC default? |
23 |
> |
24 |
> > -CFLAGS_x86="-m32" |
25 |
> > +CFLAGS_x86="-m32 -mfpmath=sse" |
26 |
|
27 |
-mfpmath=sse is already the default on amd64. |
28 |
|
29 |
> Also, why add the flag only to CFLAGS_x86 but not to CFLAGS_amd64? |
30 |
> They should have the same single and double precision arithmetic? |
31 |
> |
32 |
> Ulrich |