Re: patch: CONST_VECTOR (was clear_storage)

[Aldy Hernandez <aldyh at redhat dot com>]

>> *************** rtx_equal_p (x, y)
>> *** 626,631 ****
>> --- 639,645 ----
>>       case SCRATCH:
>>       case CONST_DOUBLE:
>>       case CONST_INT:
>> +     case CONST_VECTOR:
>>         return 0;
>
> Err, except that it is necessary if you do this.  The reason we can
> return 0 here is that we've already handled pointer equality above.

uhh, i don't understand.  should i remove the case for
CONST_VECTOR?

>> *************** purge_addressof_1 (loc, insn, force, sto
>> *** 3087,3092 ****
>> --- 3088,3094 ----
>>   	  /* Don't even consider working with paradoxical subregs,
>>   	     or the moral equivalent seen here.  */
>>   	  if (size_x <= size_sub
>> + 	      && !VECTOR_MODE_P (GET_MODE (sub))
>>   	      && int_mode_for_mode (GET_MODE (sub)) != BLKmode)
>
> I don't understand this.  You say "disallow paradoxical subregs of
> vectors", yet SIZE_X is the outer and  SIZE_SUB is the inner, thus
> inside the if we have non-paradoxical subregs.
>
> Leave this out and we can discuss this independently.

ok

>> + static rtx
>> + gen_const_vector_0 (type, mode)
>> +      enum mode_class type;
>> +      enum machine_mode mode;
>> + {
>> +   rtx tem;
>> +   rtvec v;
>> +   int units, i;
>> +   enum machine_mode inner;
>> +
>> +   units = GET_MODE_NUNITS (mode);
>> +   inner = GET_MODE_INNER (mode);
>> +
>> +   v = rtvec_alloc (units);
>> +
>> +   if (type == MODE_VECTOR_INT)
>> +     {
>> +       for (i = 0; i < units; ++i)
>> + 	RTVEC_ELT (v, i)
>> + 	  = gen_rtx_CONST_INT (VOIDmode, 0);
>> +     }
>> +   else if (type == MODE_VECTOR_FLOAT)
>> +     {
>> +       for (i = 0; i < units; ++i)
>> + 	RTVEC_ELT (v, i)
>> + 	  = gen_rtx_CONST_DOUBLE (inner, 0, 0);
>> +     }
>
> Note that if this is run after int and fp const0 is generated,
> this simplfies to
>
> 	RTVEC_ELT (v, i) = CONST0_RTX (inner);

yup.  noticed that too after i submitted.  it should be ok.
i'll test.

>> ! /* Make strings word-aligned so strcpy from constants will be faster.
>> !    Make vector constants quadword aligned.  */
>> ! #define CONSTANT_ALIGNMENT(EXP, ALIGN)                           \
>> !   (TREE_CODE (EXP) == STRING_CST	                         \
>> !    && (ALIGN) < BITS_PER_WORD                                    \
>> !    ? BITS_PER_WORD                                               \
>> !    : (TREE_CODE (EXP) == VECTOR_CST && (ALIGN) < 128) ? 128      \
>> !    : (ALIGN))
>
> I'm surprised you need this.  I would have thought this would have
> been handled by the alignment of the type.  Notice, for instance,

that is what started me on my goose chase.

force_const_mem() does this nastiness to get the alignment right:

   /* Align the location counter as required by EXP's data type.  */
   align = GET_MODE_ALIGNMENT (mode == VOIDmode ? word_mode : mode);
#ifdef CONSTANT_ALIGNMENT
   align = CONSTANT_ALIGNMENT (make_tree (type_for_mode (mode, 0), x), 
align);
#endif

get_mode_alignment() returns 32 for vectors:

unsigned int
get_mode_alignment (mode)
      enum machine_mode mode;
{
   unsigned int alignment = GET_MODE_UNIT_SIZE (mode);

   /* Extract the LSB of the size.  */
   alignment = alignment & -alignment;
   alignment *= BITS_PER_UNIT;

   alignment = MIN (BIGGEST_ALIGNMENT, MAX (1, alignment));
   return alignment;
}

so we defer to CONSTANT_ALIGNMENT.  and this, believe it or not,
is the ony place where we need a VECTOR_CST.

1500 lines of a patch later, we get correct alignment :)

--
Aldy Hernandez                                E-mail: aldyh@redhat.com
Professional Gypsy Lost in Australia
Red Hat, Inc.