[Bug rtl-optimization/25972] New: pack and unpack of long doubles via union generates poor code

["amodra at bigpond dot net dot au" <gcc-bugzilla at gcc dot gnu dot org>]

Compile the following with -m64 -mlong-double-128 -O2 -S

long double
pack (double a, double aa)
{
  union { long double ld; double d[2]; } u;
  u.d[0] = a;
  u.d[1] = aa;
  return u.ld;
}

void
unpack (long double x, double *a, double *aa)
{
  union { long double ld; double d[2]; } u;
  u.ld = x;
  *a = u.d[0];
  *aa = u.d[1];
}

Results in
[snip]
.L.pack:
        stfd 1,-176(1)
        stfd 14,-144(1)
        nop
        nop
        ld 9,-176(1)
        stfd 2,-176(1)
        std 9,-160(1)
        nop
        ld 10,-176(1)
        std 10,-152(1)
        lfd 13,-160(1)
        fmr 1,13
        lfd 14,-152(1)
        fmr 2,14
        lfd 14,-144(1)
        blr
[snip]
.L.unpack:
        stfd 1,-16(1)
        stfd 2,-8(1)
        nop
        nop
        ld 11,-16(1)
        ld 12,-8(1)
        std 11,0(5)
        std 12,0(6)
        blr

This ought to be
.L.pack:
        blr

.L.unpack:
        stfd 1,0(5)
        stfd 2,0(6)
        blr

ie. packing and unpacking long doubles ought to be nops.  Instead we see
needless moving between fprs and gprs via memory.  This problem affects
functions like __gcc_qadd from rs6000/darwin-ldouble.c


-- 
           Summary: pack and unpack of long doubles via union generates poor
                    code
           Product: gcc
           Version: 4.2.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: rtl-optimization
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: amodra at bigpond dot net dot au
GCC target triplet: powerpc64-linux


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=25972