[AArch64, PATCH] Improve Neon store of zero

[James Greenhalgh]

On Wed, Sep 06, 2017 at 10:02:52AM +0100, Jackson Woodruff wrote:
> Hi all,
> 
> I've attached a new patch that addresses some of the issues raised with 
> my original patch.
> 
> On 08/23/2017 03:35 PM, Wilco Dijkstra wrote:
> > Richard Sandiford wrote:
> >>
> >> Sorry for only noticing now, but the call to aarch64_legitimate_address_p
> >> is asking whether the MEM itself is a legitimate LDP/STP address.  Also,
> >> it might be better to pass false for strict_p, since this can be called
> >> before RA.  So maybe:
> >>
> >>     if (GET_CODE (operands[0]) == MEM
> >> 	&& !(aarch64_simd_imm_zero (operands[1], <MODE>mode)
> >> 	     && aarch64_mem_pair_operand (operands[0], <MODE>mode)))
> 
> There were also some issues with the choice of mode for the call the 
> aarch64_mem_pair_operand.
> 
> For a 128-bit wide mode, we want to check `aarch64_mem_pair_operand 
> (operands[0], DImode)` since that's what the stp will be.
> 
> For a 64-bit wide mode, we don't need to do that check because a normal
> `str` can be issued.
> 
> I've updated the condition as such.
> 
> > 
> > Is there any reason for doing this check at all (or at least this early during
> > expand)?
> 
> Not doing this check means that the zero is forced into a register, so 
> we then carry around a bit more RTL and rely on combine to merge things.
> 
> > 
> > There is a similar issue with this part:
> > 
> >   (define_insn "*aarch64_simd_mov<mode>"
> >     [(set (match_operand:VQ 0 "nonimmediate_operand"
> > -		"=w, m,  w, ?r, ?w, ?r, w")
> > +		"=w, Ump,  m,  w, ?r, ?w, ?r, w")
> > 
> > The Ump causes the instruction to always split off the address offset. Ump
> > cannot be used in patterns that are generated before register allocation as it
> > also calls laarch64_legitimate_address_p with strict_p set to true.
> 
> I've changed the constraint to a new constraint 'Umq', that acts the 
> same as Ump, but calls aarch64_legitimate_address_p with strict_p set to 
> false and uses DImode for the mode to pass.

This looks mostly OK to me, but this conditional:

> +  if (GET_CODE (operands[0]) == MEM
> +      && !(aarch64_simd_imm_zero (operands[1], <MODE>mode)
> +	   && ((GET_MODE_SIZE (<MODE>mode) == 16
> +		&& aarch64_mem_pair_operand (operands[0], DImode))
> +	       || GET_MODE_SIZE (<MODE>mode) == 8)))

Has grown a bit too big in such a general pattern to live without a comment
explaining what is going on.

> +(define_memory_constraint "Umq"
> +  "@internal
> +   A memory address which uses a base register with an offset small enough for
> +   a load/store pair operation in DI mode."
> +   (and (match_code "mem")
> +	(match_test "aarch64_legitimate_address_p (DImode, XEXP (op, 0),
> +						   PARALLEL, 0)")))

And here you want 'false' rather than '0'.

I'll happily merge the patch with those changes, please send an update.

Thanks,
James


> 
> ChangeLog:
> 
> gcc/
> 
> 2017-08-29  Jackson Woodruff  <jackson.woodruff@arm.com>
> 
> 	* config/aarch64/constraints.md (Umq): New constraint.
> 	* config/aarch64/aarch64-simd.md (*aarch64_simd_mov<mode>):
> 	Change to use Umq.
> 	(mov<mode>): Update condition.
> 
> gcc/testsuite
> 
> 2017-08-29  Jackson Woodruff  <jackson.woodruff@arm.com>
> 
> 	* gcc.target/aarch64/simd/vect_str_zero.c:
> 	Update testcase.