RE: [PATCH]middle-end: check if target can do extract first for early breaks [PR113199]

[Tamar Christina <Tamar.Christina@arm.com>]

> -----Original Message-----
> From: Richard Biener <rguenther@suse.de>
> Sent: Monday, January 8, 2024 12:48 PM
> To: Tamar Christina <Tamar.Christina@arm.com>
> Cc: gcc-patches@gcc.gnu.org; nd <nd@arm.com>; jlaw@ventanamicro.com
> Subject: Re: [PATCH]middle-end: check if target can do extract first for early breaks
> [PR113199]
> 
> On Tue, 2 Jan 2024, Tamar Christina wrote:
> 
> > Hi All,
> >
> > I was generating the vector reverse mask without checking if the target
> > actually supported such an operation.
> >
> > It also seems like more targets implement VEC_EXTRACT than permute on mask
> > registers.
> >
> > So this adds a check for IFN_VEC_EXTRACT support when required and changes
> > the select first code to use it.
> >
> > This is good for now since masks always come from whilelo.  But in the future
> > when masks can come from other sources we will need the old code back.
> >
> > Bootstrapped Regtested on aarch64-none-linux-gnu, x86_64-pc-linux-gnu
> > and no issues with --enable-checking=release --enable-lto
> > --with-build-config=bootstrap-O3 --enable-checking=yes,rtl,extra.
> > tested on cross cc1 for amdgcn-amdhsa and issue fixed.
> >
> > Ok for master?
> >
> > Thanks,
> > Tamar
> >
> > gcc/ChangeLog:
> >
> > 	PR tree-optimization/113199
> > 	* tree-vect-loop.cc (vectorizable_live_operation_1): Use
> > 	IFN_VEC_EXTRACT.
> > 	(vectorizable_live_operation): Check for IFN_VEC_EXTRACT support.
> >
> > gcc/testsuite/ChangeLog:
> >
> > 	PR tree-optimization/113199
> > 	* gcc.target/gcn/pr113199.c: New test.
> >
> > --- inline copy of patch --
> > diff --git a/gcc/testsuite/gcc.target/gcn/pr113199.c
> b/gcc/testsuite/gcc.target/gcn/pr113199.c
> > new file mode 100644
> > index
> 0000000000000000000000000000000000000000..8a641e5536e80e207ca01
> 63cac66c0f4f6ca93f7
> > --- /dev/null
> > +++ b/gcc/testsuite/gcc.target/gcn/pr113199.c
> > @@ -0,0 +1,44 @@
> > +/* { dg-do compile } */
> > +/* { dg-additional-options "-O2" } */
> > +
> > +typedef long unsigned int size_t;
> > +typedef int wchar_t;
> > +struct tm
> > +{
> > +  int tm_mon;
> > +  int tm_year;
> > +};
> > +int abs (int);
> > +struct lc_time_T { const char *month[12]; };
> > +struct __locale_t * __get_current_locale (void) { }
> > +const struct lc_time_T * __get_time_locale (struct __locale_t *locale) { }
> > +const wchar_t * __ctloc (wchar_t *buf, const char *elem, size_t *len_ret) {
> return buf; }
> > +size_t
> > +__strftime (wchar_t *s, size_t maxsize, const wchar_t *format,
> > +     const struct tm *tim_p, struct __locale_t *locale)
> > +{
> > +  size_t count = 0;
> > +  const wchar_t *ctloc;
> > +  wchar_t ctlocbuf[256];
> > +  size_t i, ctloclen;
> > +  const struct lc_time_T *_CurrentTimeLocale = __get_time_locale (locale);
> > +    {
> > +      switch (*format)
> > + {
> > + case L'B':
> > +   (ctloc = __ctloc (ctlocbuf, _CurrentTimeLocale->month[tim_p->tm_mon],
> &ctloclen));
> > +   for (i = 0; i < ctloclen; i++)
> > +     {
> > +       if (count < maxsize - 1)
> > +  s[count++] = ctloc[i];
> > +       else
> > +  return 0;
> > +       {
> > +  int century = tim_p->tm_year >= 0
> > +    ? tim_p->tm_year / 100 + 1900 / 100
> > +    : abs (tim_p->tm_year + 1900) / 100;
> > +       }
> > +   }
> > + }
> > +    }
> > +}
> > diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc
> > index
> 37f1be1101ffae779214056a0886411e0683e887..5aa92e67444e7aacf458fffa14
> 28f1983c482374 100644
> > --- a/gcc/tree-vect-loop.cc
> > +++ b/gcc/tree-vect-loop.cc
> > @@ -10648,36 +10648,18 @@ vectorizable_live_operation_1 (loop_vec_info
> loop_vinfo,
> >  				      &LOOP_VINFO_MASKS (loop_vinfo),
> >  				      1, vectype, 0);
> >        tree scalar_res;
> > +      gimple_seq_add_seq (&stmts, tem);
> >
> >        /* For an inverted control flow with early breaks we want EXTRACT_FIRST
> > -	 instead of EXTRACT_LAST.  Emulate by reversing the vector and mask. */
> > +	 instead of EXTRACT_LAST.  For now since the mask always comes from a
> > +	 WHILELO we can get the first element ignoring the mask since CLZ of the
> > +	 mask will always be zero.  */
> >        if (restart_loop && LOOP_VINFO_EARLY_BREAKS (loop_vinfo))
> > -	{
> > -	  /* First create the permuted mask.  */
> > -	  tree perm_mask = perm_mask_for_reverse (TREE_TYPE (mask));
> > -	  tree perm_dest = copy_ssa_name (mask);
> > -	  gimple *perm_stmt
> > -		= gimple_build_assign (perm_dest, VEC_PERM_EXPR, mask,
> > -				       mask, perm_mask);
> > -	  vect_finish_stmt_generation (loop_vinfo, stmt_info, perm_stmt,
> > -				       &gsi);
> > -	  mask = perm_dest;
> > -
> > -	  /* Then permute the vector contents.  */
> > -	  tree perm_elem = perm_mask_for_reverse (vectype);
> > -	  perm_dest = copy_ssa_name (vec_lhs_phi);
> > -	  perm_stmt
> > -		= gimple_build_assign (perm_dest, VEC_PERM_EXPR, vec_lhs_phi,
> > -				       vec_lhs_phi, perm_elem);
> > -	  vect_finish_stmt_generation (loop_vinfo, stmt_info, perm_stmt,
> > -				       &gsi);
> > -	  vec_lhs_phi = perm_dest;
> > -	}
> > -
> > -      gimple_seq_add_seq (&stmts, tem);
> > -
> > -      scalar_res = gimple_build (&stmts, CFN_EXTRACT_LAST, scalar_type,
> > -				 mask, vec_lhs_phi);
> > +	scalar_res = gimple_build (&stmts, CFN_VEC_EXTRACT, TREE_TYPE
> (vectype),
> > +				   vec_lhs_phi, bitstart);
> 
> So bitstart is always zero?  I wonder why using CFN_VEC_EXTRACT over
> BIT_FIELD_REF here which wouldn't need any additional target support.
> 

Hmm Unless it's recently changed, I believe BIT_FIELD_REF doesn't support
VLA types no?  I guess maybe it does for index 0?  But I was pretty sure it asserts.

> > +      else
> > +	scalar_res = gimple_build (&stmts, CFN_EXTRACT_LAST, scalar_type,
> > +				   mask, vec_lhs_phi);
> >
> >        /* Convert the extracted vector element to the scalar type.  */
> >        new_tree = gimple_convert (&stmts, lhs_type, scalar_res);
> > @@ -10852,9 +10834,25 @@ vectorizable_live_operation (vec_info *vinfo,
> stmt_vec_info stmt_info,
> >  	      gcc_assert (ncopies == 1 && !slp_node);
> >  	      if (direct_internal_fn_supported_p (IFN_EXTRACT_LAST, vectype,
> >  						  OPTIMIZE_FOR_SPEED))
> > -		vect_record_loop_mask (loop_vinfo,
> > -				       &LOOP_VINFO_MASKS (loop_vinfo),
> > -				       1, vectype, NULL);
> > +		{
> > +		  if (LOOP_VINFO_EARLY_BREAKS_VECT_PEELED (loop_vinfo)
> > +		      && LOOP_VINFO_EARLY_BREAKS (loop_vinfo)
> > +		      && !direct_internal_fn_supported_p (IFN_EXTRACT_LAST,
> > +							  vectype,
> > +							  OPTIMIZE_FOR_SPEED))
> 
> I don't quite understand this part - this is guarded by
> 
> direct_internal_fn_supported_p (IFN_EXTRACT_LAST, vectype,
>                                                  OPTIMIZE_FOR_SPEED)
> 
> unless I'm mis-matching this means the checked
> !direct_internal_fn_supported_p condition is always false?

Arg,, sorry should be IFN_ VEC_EXTRACT.

I'll fix and see if BIT_FIELD_REF works for 0 index.

Thanks,
Tamar
> 
> > +		    {
> > +		      if (dump_enabled_p ())
> > +			dump_printf_loc (MSG_MISSED_OPTIMIZATION,
> vect_location,
> > +				"can't operate on partial vectors "
> > +				"because the target doesn't support extract "
> > +				"first reduction.\n");
> > +		      LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) =
> false;
> > +		    }
> > +		  else
> > +		    vect_record_loop_mask (loop_vinfo,
> > +					   &LOOP_VINFO_MASKS (loop_vinfo),
> > +					   1, vectype, NULL);
> > +		}
> >  	      else if (can_vec_extract_var_idx_p (
> >  			 TYPE_MODE (vectype), TYPE_MODE (TREE_TYPE
> (vectype))))
> >  		vect_record_loop_len (loop_vinfo,
> >
> >
> >
> >
> >
> 
> --
> Richard Biener <rguenther@suse.de>
> SUSE Software Solutions Germany GmbH,
> Frankenstrasse 146, 90461 Nuernberg, Germany;
> GF: Ivo Totev, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)