Re: [PATCH v2 00/14] GLIBC LoongArch PATCHES

[Xi Ruoyao <xry111@mengyan1223.wang>]

On Sun, 2022-01-30 at 08:43 +0800, Xi Ruoyao wrote:
> On Tue, 2022-01-04 at 13:27 +0000, Adhemerval Zanella via Libc-alpha
> wrote:
> > On 31/12/2021 03:44, caiyinyu wrote:
> 
> /* snip */
> 
> > > FAIL: math/test-double-acos
> > > FAIL: math/test-double-asin
> > > FAIL: math/test-float32x-acos
> > > FAIL: math/test-float32x-asin
> > > FAIL: math/test-float64-acos
> > > FAIL: math/test-float64-asin
> > > 
> > > These fails are caused by gcc optimizations. if we use -O0
> > > options, these fails
> > > will pass.
> 
> /* snip */
> 
> > > 338       else {
> > > 339         u.i[HIGH_HALF]=0x7ff00000;
> > > 340         v.i[HIGH_HALF]=0x7ff00000;
> > > 341         u.i[LOW_HALF]=0;
> > > 342         v.i[LOW_HALF]=0;
> > > 343         return u.x/v.x;                     ///////// 
> > > optimized out
> > > 
> > >    0x00007ffff7f4dad8 <+1432>:  pcaddu12i       $t0, 63(0x3f)
> > >    0x00007ffff7f4dadc <+1436>:  addi.d          $t0, $t0, -
> > > 1248(0xb20)
> > >    0x00007ffff7f4dae0 <+1440>:  fld.d           $fa0, $t0, 0
> > >    0x00007ffff7f4dae4 <+1444>:  b               -812(0xffffcd4) #
> > > 0x7ffff7f4d7b8 <__ieee754_acos+632>
> 
> /* snip */
> 
> > Is this being tracked by a GCC bug report? We need to understand if
> > we
> > require to use math_force_eval to avoid such issue on other ports as
> > well.
> 
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95115
> 
> This one has been haunting RISC-V for a while.

Fixed for GCC 12 at r12-6959.  Now the generated LoongArch assembly is:

.L98:
 # ../sysdeps/ieee754/dbl-64/e_asin.c:343:     return u.x/v.x;
    la.local    $r12,.LC19   # tmp848,
    fld.d   $f0,$r12,0   # tmp847,
    fdiv.d  $f0,$f0,$f0  # <retval>, tmp847, tmp847
    b   .L63         #

# ... ...

.LC19:
    .word   0
    .word   2146435072
    .align  3

Which is, literally, inf/inf.
-- 
Xi Ruoyao <xry111@mengyan1223.wang>
School of Aerospace Science and Technology, Xidian University