LoongArch incorrect codegen for std::byte_swap

[Xi Ruoyao]

On Sat, 2024-07-27 at 17:38 -0700, Andrew Pinski via Gcc-help wrote:
> On Sat, Jul 27, 2024 at 3:37 PM pifminns deettnta via Gcc
> <gcc@gcc.gnu.org> wrote:
> > 
> > using uint_least64_t = __UINT_LEAST64_TYPE__;
> > 
> > uint_least64_t testbswap(uint_least64_t a) noexcept
> > {
> >    return __builtin_bswap64(a);
> > }
> > 
> > clang:
> > https://godbolt.org/z/z8GTsazf4
> > 
> > 
> > _Z9testbswapm:
> >        revb.d  $a0, $a0
> >        ret
> > 
> > 
> > GCC:
> > 
> > https://godbolt.org/z/PabfxP9ve
> > 
> > _Z9testbswapm:
> >        revb.4h $r4,$r4
> >        revh.d  $r4,$r4
> >        jr      $r1
> > 
> > It should just use revb.d for bswap, not separate them into two.
> 
> The code generation is not wrong, just not as good.
> GCC swaps bytes inside 4 half words and then swaps the half words.
> 
> I looked into the history of GCC's code generation here and noticed it
> comes from the original port when committed upstream. even the bswap32
> is also done using 2 instructions. Now I am suspecting is the original
> Loongson ISA didn't have revb.d/revb.2w when the GCC port was done and
> it was added afterwards and GCC port was never updated to use the new
> instructions.

revb.2w and revb.d are available since day one.  I've no idea why they
weren't used.  (Also strangely there's no revb.w, so on the [still
insubstantial] 32-bit LoongArch we would have to use revb.2h + rotri.w.
for bswap32, but it may change before a real 32-bit CPU is launched.)

I'm preparing a patch.

-- 
Xi Ruoyao <xry111@xry111.site>
School of Aerospace Science and Technology, Xidian University