[Bug target/5360] [SH] Useless address inc-dec sequence for DFmode access

[olegendo at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=5360

--- Comment #6 from Oleg Endo <olegendo at gcc dot gnu.org> ---
When compiling the following test with -m2a -mb -O2

double test (void)
{
  return 123.123;
}

we get

        mova    .L2,r0
        fmov.s  @r0+,fr0
        fmov.s  @r0,fr1
        rts
        add     #-4,r0
.L3:
        .align 2
.L2:
        .long   1079953375
        .long   996432413

Although the address reg is dead after the last load, the add #-4 insn is not
eliminated.  This is probably because the DFmode move patterns are split at a
late stage during compilation.  The add #-4 has to be put there initially
because for the register allocator the DFmode constant looks like it doesn't
modify the address register.  In some cases this allows optimizations of
consecutive constant loads, as in:

void test (double* out)
{
  out[0] = 123.123;
  out[1] = 222;
}

        mova    .L2,r0
        fmov.s  @r0+,fr2
        fmov.s  @r0+,fr3
        add     #4,r4
        fmov.s  fr3,@r4
        fmov.s  fr2,@-r4
        fmov.s  @r0+,fr2
        fmov.s  @r0,fr3
        add     #-4,r0
        fmov.s  fr2,@(8,r4)
        fmov.s  fr3,@(12,r4)
        rts/n
.L3:
        .align 2
.L2:
        .long   1079953375
        .long   996432413
        .long   1080803328
        .long   0