[PATCH] Avoid changing DR for scatter/gather refs

[Richard Biener]

On Sun, 24 Jun 2018, Eric Botcazou wrote:

> > 2018-06-21  Richard Biener  <rguenther@suse.de>
> > 
> > 	* tree-data-ref.c (dr_step_indicator): Handle NULL DR_STEP.
> > 	* tree-vect-data-refs.c (vect_analyze_possibly_independent_ddr):
> > 	Avoid calling vect_mark_for_runtime_alias_test with gathers or scatters.
> > 	(vect_analyze_data_ref_dependence): Re-order checks to deal with
> > 	NULL DR_STEP.
> > 	(vect_record_base_alignments): Do not record base alignment
> > 	for gathers or scatters.
> > 	(vect_compute_data_ref_alignment): Drop return value that is always
> > 	true.  Bail out early for gathers or scatters.
> > 	(vect_enhance_data_refs_alignment): Bail out early for gathers
> > 	or scatters.
> > 	(vect_find_same_alignment_drs): Likewise.
> > 	(vect_analyze_data_refs_alignment): Remove dead code.
> > 	(vect_slp_analyze_and_verify_node_alignment): Likewise.
> > 	(vect_analyze_data_refs): For possible gathers or scatters do
> > 	not create an alternate DR, just check their possible validity
> > 	and mark them.  Adjust DECL_NONALIASED handling to not rely
> > 	on DR_BASE_ADDRESS.
> > 	* tree-vect-loop-manip.c (vect_update_inits_of_drs): Do not
> > 	update inits of gathers or scatters.
> > 	* tree-vect-patterns.c (vect_recog_mask_conversion_pattern):
> > 	Also copy gather/scatter flag to pattern vinfo.
> 
> This breaks the attached testcase sso9.adb compiled at -O3:
> 
> +===========================GNAT BUG DETECTED==============================+
> | 9.0.0 20180621 (experimental) [trunk revision 261832] (x86_64-suse-linux) 
> GCC error:|
> | in vect_check_gather_scatter, at tree-vect-data-refs.c:3733              |
> | Error detected around /home/eric/svn/gcc/gcc/testsuite/gnat.dg/sso9.adb:6:1|
> | Please submit a bug report; see https://gcc.gnu.org/bugs/ .              |

Ah, ok.  So currently data-ref analysis in dr_analyze_innermost punts
on reverse storage order accesses.  Before this patch the scatter/gather
case re-analyzed the ref w/o loop context and the of course failed
again because of that.  Now we treat it as possibly scatter/gather
and ask vect_check_gather_scatter whether it is really ok.  That does

  base = get_inner_reference (base, &pbitsize, &pbitpos, &off, &pmode,
                              &punsignedp, &reversep, &pvolatilep);
  gcc_assert (base && !reversep);

but of course both reversep can now happen.  (I think get_inner_reference
can never return NULL)

So the following should fix it, testing in progress.

Richard.

2018-06-25  Richard Biener  <rguenther@suse.de>

	* tree-vect-data-refs.c (vect_check_gather_scatter): Fail
	for reverse storage order accesses rather than asserting
	they cannot happen here.

Index: gcc/tree-vect-data-refs.c
===================================================================
--- gcc/tree-vect-data-refs.c	(revision 262005)
+++ gcc/tree-vect-data-refs.c	(working copy)
@@ -3730,7 +3730,9 @@ vect_check_gather_scatter (gimple *stmt,
      that can be gimplified before the loop.  */
   base = get_inner_reference (base, &pbitsize, &pbitpos, &off, &pmode,
 			      &punsignedp, &reversep, &pvolatilep);
-  gcc_assert (base && !reversep);
+  if (reversep)
+    return false;
+
   poly_int64 pbytepos = exact_div (pbitpos, BITS_PER_UNIT);
 
   if (TREE_CODE (base) == MEM_REF)