Re: [PATCH, x86] Generate vector -1 with pcmpeq

[Evan Cheng <evan dot cheng at apple dot com>]

Here is a new patch:

Attachment: patch.fsf.4283414.txt
Description: Text document

Thanks,

Evan

On Nov 1, 2005, at 1:09 PM, Richard Henderson wrote:

> On Fri, Oct 28, 2005 at 01:10:18PM -0700, Evan Cheng wrote:
>
> > + /* Return true if the rtx is a vector constant that is all  
> > ones.  */
> > + bool
> > + ix86_const_vector_all_ones_operand (rtx v, enum machine_mode mode)
>
> There's no reason to have this code outside predicates.md.
>
>
> > +           if (GET_CODE (x) != CONST_INT || INTVAL (x) != -1)
>
> x != constm1_rtx.
>
> You could probably also match the corresponding NaN, but
> perhaps it's not worth it.
>
>
> > !     case 2:
> > !       if (get_attr_mode (insn) == MODE_V2DF)
> > !         return "pcmpeqd\t%0, %0";
> > !       else
> > !         return "pcmpeqw\t%0, %0";
>
> Which brings me to this distinction.  What are you doing?  If you
> really need this in V2DFmode you'd be using cmpneqpd.  But that
> would imply you'd been matching a vector of nan("0xff...") and you
> aren't.
>
> I don't see that any processor distinguishes between integer modes,
> so I don't see any reason not to always use pcmpeqd, or always
> pcmpeqw, or for that matter pcmpeqb.
>
> Which raises the question of SSE1, which I'll get back to.
>
>
> > !  abort ();
>
> gcc_unreachable.
>
>
> > +       && !ix86_const_vector_all_ones_operand (op1, mode)
> >         && CONSTANT_P (op1) && op1 != CONST0_RTX (mode))
>
> Use standard_sse_constant_p here.
>
>
> > !   "TARGET_SSE && ix86_binary_operator_ok (UNKNOWN, V4SFmode,  
> > operands)"
> ...
>
> > !   "TARGET_SSE && ix86_binary_operator_ok (UNKNOWN, V4SFmode,  
> > operands)
> > !    && !ix86_const_vector_all_ones_operand (operands[2], V4SFmode)"
>
> I'm guessing this is locally adjusting for vector_move_operand.
> Notice how you had to replicate this 8 times.  Better to fix this
> once elsewhere.
>
> My thinking is that standard_sse_constant_p should return false
> for -1 on SSE1, and that vector_move_operand should use
> standard_sse_constant_p instead of listing out the alternatives.
>
>
> r~