Re: How to add 256 byte sized opaque mode in gcc

On Thu, 20 Aug 2026, Avinash Jayakar wrote:

> On Wed, 2026-08-19 at 08:50 +0200, Richard Biener via Gcc wrote:
> > On Wed, 19 Aug 2026, Avinash Jayakar wrote:
> > 
> > > Hi Jakub/Richard,
> > > 
> > > Thank you for these suggestions and review.
> > > 
> > > I have attached the proposed patch based on your input to
> > > dynamically
> > > size the mode_unit_size array based on maximum bytesize declared in
> > > machine modes.
> > > 
> > > Please do let me know if any tests are required for this before
> > > sending
> > > to gcc-patches?
> > 
> > +  unsigned int max_type_size = 0;
> > +  for_all_modes (c, m)
> > +    {
> > +      if (max_type_size < m->bytesize)
> > +       max_type_size = m->bytesize;
> > +    }
> > +  if (max_type_size < (1 << 8))
> > +    puts ("#define MODE_UNIT_SIZE_TYPE unsigned char\n");
> > +  else if (max_type_size < (1 << 16))
> > +    puts ("#define MODE_UNIT_SIZE_TYPE unsigned short\n");
> > +  else
> > +    puts ("#define MODE_UNIT_SIZE_TYPE unsigned int\n");
> > +
> > 
> > I think you need to account for bits_per_unit here, meaning
> > you want to track max_unit_size which would be
> > ((size_t)m->bytesize * 8) / bits_per_unit
> > 
> Yes, I think I missed this part. But had a doubt here in the following
> function
> static void
> emit_mode_unit_size (void)
> { 
>   int c;
>   struct mode_data *m;
>   
>   print_maybe_const_decl ("%sMODE_UNIT_SIZE_TYPE", "mode_unit_size",
>                           "NUM_MACHINE_MODES", adj_bytesize);
> 
>   for_all_modes (c, m)
>     tagged_printf ("%u",
>                    c != MODE_PARTIAL_INT && m->component
>                    ? m->component->bytesize : m->bytesize, m->name);
>   
>   print_closer ();
> }  
> 
> we print the component->bytesize if component is not null, else just
> the bytesize. I do not see bits_per_unit accounted for here. Could this
> be a potential issue?
> Can I directly use whatever number is being printed here to calculate
> the max, since we need to size the type according to what numbers we
> print here?

So it looks like mode_size[] contains byte size and mode_unit_size[]
contains unit size.  I assumed that m->bytesize is really bytesize.

The difficulty is, of course, that we have no target with BITS_PER_UNIT
!= 8 and that support might have bit-rotted quite a bit.

That said, until we officially declare BITS_PER_UNIT == 8 forever
we should see to honor != 8 values correctly.

Richard.

> Thanks,
> Avinash
> 
> > Richard.
> > 
> > > Also as you mentioned these lookups happen quite often in
> > > compilation,
> > > so introducing this new mode could potentially increase compile
> > > times
> > > for powerpc64le right?
> > > 
> > > Thanks and regards,
> > > Avinash Jayakar
> > > 
> > > On Tue, 2026-08-18 at 13:06 +0200, Jakub Jelinek via Gcc wrote:
> > > > On Tue, Aug 18, 2026 at 12:48:55PM +0200, Richard Biener via Gcc
> > > > wrote:
> > > > > I'd use unsigned short, it's important to not use an overly
> > > > > large
> > > > > data
> > > > > structure for those common lookups to reduce cache footprint. 
> > > > > We
> > > > > could possibly even dynamically size the component from
> > > > > genmodes?
> > > > 
> > > > Yeah, I'd prefer that, emit a typedef/macro in insn-modes.h, say
> > > > MODE_UNIT_SIZE_TYPE, unsigned char or unsigned short depending in
> > > > whether all modes fit or don't fit into 255 bytes and use it for
> > > > the
> > > > generated array too.  Similar to how e.g. the CONST_MODE_* macros
> > > > are defined.
> > > > 
> > > > > > @@ -1884,6 +1885,11 @@ rs6000_hard_regno_mode_ok_uncached
> > > > > > (int
> > > > > > regno, machine_mode mode)
> > > > > >        else
> > > > > >  	return 0;
> > > > > >      }
> > > > > > +  /* QDOmode needs even/odd DMR register pairs.  */
> > > > > > +  if (mode == QDOmode)
> > > > > > +    {
> > > > > > +      return (TARGET_DMF && DMR_REGNO_P (regno) && (regno &
> > > > > > 1)
> > > > > > == 0);
> > > > > > +    }
> > > > 
> > > > The GCC coding style here is to avoid the {}s around a single
> > > > statement
> > > > body.
> > > Sure I will update this.
> > > > 
> > > > 	Jakub
> > > 
> 

-- 
Richard Biener <[email protected]>
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Jochen Jaser, Andrew McDonald, Abhinav Puri; (HRB 36809, AG Nuernberg)