Fix PR94043 by making vect_live_op generate lc-phi

Message ID af33d56b-4ccf-f449-94dc-dcb3a0f62623@linux.ibm.com
State New
Headers show
Series
  • Fix PR94043 by making vect_live_op generate lc-phi
Related show

Commit Message

Marek Polacek via Gcc-patches March 30, 2020, 10:24 a.m.
Hi,

As PR94043 shows, my commit r10-4524 exposed one issue in
vectorizable_live_operation, which inserts one extra BB
before the single exit, leading unexpected operand expansion
and unexpected loop depth assertion.  As Richi suggested,
this patch is to teach vectorizable_live_operation to
generate loop closed phi for vec_lhs, it looks like:
     loop;
     # lhs' = PHI <lhs>
=>
     loop;
     # vec_lhs' = PHI <vec_lhs>
     new_tree = BIT_FIELD_REF <vec_lhs', ...>;
     lhs' = new_tree;

I noticed that there are some SLP cases that have same lhs
and vec_lhs but different offsets, which can make us have
more PHIs for the same vec_lhs there.  But I think it would
be fine since only one of them is actually live, the others
should be eliminated by the following dce.  So the patch
doesn't check whether there is one phi for vec_lhs, just
create one directly instead.

Bootstrapped/regtested on powerpc64le-linux-gnu (LE) P8.

Is it ok for trunk?

BR,
Kewen
-----------

gcc/ChangeLog

2020-MM-DD  Kewen Lin  <linkw@gcc.gnu.org>

	PR tree-optimization/94043
	* tree-vect-loop.c (vectorizable_live_operation): Generate loop-closed
	phi for vec_lhs and use it for lane extraction.

gcc/testsuite/ChangeLog

2020-MM-DD  Kewen Lin  <linkw@gcc.gnu.org>

	PR tree-optimization/94043
	* gfortran.dg/graphite/vect-pr94043.f90: New test.

Comments

Marek Polacek via Gcc-patches March 30, 2020, 10:38 a.m. | #1
On Mon, Mar 30, 2020 at 12:24 PM Kewen.Lin <linkw@linux.ibm.com> wrote:
>

> Hi,

>

> As PR94043 shows, my commit r10-4524 exposed one issue in

> vectorizable_live_operation, which inserts one extra BB

> before the single exit, leading unexpected operand expansion

> and unexpected loop depth assertion.  As Richi suggested,

> this patch is to teach vectorizable_live_operation to

> generate loop closed phi for vec_lhs, it looks like:

>      loop;

>      # lhs' = PHI <lhs>

> =>

>      loop;

>      # vec_lhs' = PHI <vec_lhs>

>      new_tree = BIT_FIELD_REF <vec_lhs', ...>;

>      lhs' = new_tree;

>

> I noticed that there are some SLP cases that have same lhs

> and vec_lhs but different offsets, which can make us have

> more PHIs for the same vec_lhs there.  But I think it would

> be fine since only one of them is actually live, the others

> should be eliminated by the following dce.  So the patch

> doesn't check whether there is one phi for vec_lhs, just

> create one directly instead.

>

> Bootstrapped/regtested on powerpc64le-linux-gnu (LE) P8.

>

> Is it ok for trunk?


OK.

Thanks,
Richard.

> BR,

> Kewen

> -----------

>

> gcc/ChangeLog

>

> 2020-MM-DD  Kewen Lin  <linkw@gcc.gnu.org>

>

>         PR tree-optimization/94043

>         * tree-vect-loop.c (vectorizable_live_operation): Generate loop-closed

>         phi for vec_lhs and use it for lane extraction.

>

> gcc/testsuite/ChangeLog

>

> 2020-MM-DD  Kewen Lin  <linkw@gcc.gnu.org>

>

>         PR tree-optimization/94043

>         * gfortran.dg/graphite/vect-pr94043.f90: New test.

>
Marek Polacek via Gcc-patches April 1, 2020, 10:51 p.m. | #2
On Mon, Mar 30, 2020 at 4:09 AM Richard Biener via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>

> On Mon, Mar 30, 2020 at 12:24 PM Kewen.Lin <linkw@linux.ibm.com> wrote:

> >

> > Hi,

> >

> > As PR94043 shows, my commit r10-4524 exposed one issue in

> > vectorizable_live_operation, which inserts one extra BB

> > before the single exit, leading unexpected operand expansion

> > and unexpected loop depth assertion.  As Richi suggested,

> > this patch is to teach vectorizable_live_operation to

> > generate loop closed phi for vec_lhs, it looks like:

> >      loop;

> >      # lhs' = PHI <lhs>

> > =>

> >      loop;

> >      # vec_lhs' = PHI <vec_lhs>

> >      new_tree = BIT_FIELD_REF <vec_lhs', ...>;

> >      lhs' = new_tree;

> >

> > I noticed that there are some SLP cases that have same lhs

> > and vec_lhs but different offsets, which can make us have

> > more PHIs for the same vec_lhs there.  But I think it would

> > be fine since only one of them is actually live, the others

> > should be eliminated by the following dce.  So the patch

> > doesn't check whether there is one phi for vec_lhs, just

> > create one directly instead.

> >

> > Bootstrapped/regtested on powerpc64le-linux-gnu (LE) P8.

> >

> > Is it ok for trunk?

>

> OK.

>

> Thanks,

> Richard.

>

> > BR,

> > Kewen

> > -----------

> >

> > gcc/ChangeLog

> >

> > 2020-MM-DD  Kewen Lin  <linkw@gcc.gnu.org>

> >

> >         PR tree-optimization/94043

> >         * tree-vect-loop.c (vectorizable_live_operation): Generate loop-closed

> >         phi for vec_lhs and use it for lane extraction.

> >

> > gcc/testsuite/ChangeLog

> >

> > 2020-MM-DD  Kewen Lin  <linkw@gcc.gnu.org>

> >

> >         PR tree-optimization/94043

> >         * gfortran.dg/graphite/vect-pr94043.f90: New test.

> >


This caused:

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94449

-- 
H.J.

Patch

diff --git a/gcc/testsuite/gfortran.dg/graphite/vect-pr94043.f90 b/gcc/testsuite/gfortran.dg/graphite/vect-pr94043.f90
new file mode 100644
index 00000000000..744c0f3042e
--- /dev/null
+++ b/gcc/testsuite/gfortran.dg/graphite/vect-pr94043.f90
@@ -0,0 +1,18 @@ 
+! { dg-do compile }
+! { dg-additional-options "-O3 -ftree-parallelize-loops=2 -fno-tree-dce" }
+
+! As PR94043, test it to be compiled successfully without ICE.
+
+program yw
+      integer :: hx(6, 6)
+      integer :: ps = 1, e2 = 1
+
+      do ps = 1, 6
+        do e2 = 1, 6
+            hx(e2, ps) = 0
+            if (ps >= 5 .and. e2 >= 5) then
+                hx(e2, ps) = hx(1, 1)
+            end if
+        end do
+      end do
+end program
diff --git a/gcc/tree-vect-loop.c b/gcc/tree-vect-loop.c
index 53fccb715ef..eb607f3aab0 100644
--- a/gcc/tree-vect-loop.c
+++ b/gcc/tree-vect-loop.c
@@ -7998,6 +7998,25 @@  vectorizable_live_operation (stmt_vec_info stmt_info,
       bitstart = int_const_binop (MINUS_EXPR, vec_bitsize, bitsize);
     }
 
+  /* Ensure the VEC_LHS for lane extraction stmts satisfy loop-closed PHI
+     requirement, insert one phi node for it.  It looks like:
+	 loop;
+       BB:
+	 # lhs' = PHI <lhs>
+     ==>
+	 loop;
+       BB:
+	 # vec_lhs' = PHI <vec_lhs>
+	 new_tree = lane_extract <vec_lhs', ...>;
+	 lhs' = new_tree;  */
+
+  basic_block exit_bb = single_exit (loop)->dest;
+  gcc_assert (single_pred_p (exit_bb));
+
+  tree vec_lhs_phi = copy_ssa_name (vec_lhs);
+  gimple *phi = create_phi_node (vec_lhs_phi, exit_bb);
+  SET_PHI_ARG_DEF (phi, single_exit (loop)->dest_idx, vec_lhs);
+
   gimple_seq stmts = NULL;
   tree new_tree;
   if (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo))
@@ -8010,10 +8029,10 @@  vectorizable_live_operation (stmt_vec_info stmt_info,
 	 the loop mask for the final iteration.  */
       gcc_assert (ncopies == 1 && !slp_node);
       tree scalar_type = TREE_TYPE (STMT_VINFO_VECTYPE (stmt_info));
-      tree mask = vect_get_loop_mask (gsi, &LOOP_VINFO_MASKS (loop_vinfo),
-				      1, vectype, 0);
-      tree scalar_res = gimple_build (&stmts, CFN_EXTRACT_LAST,
-				      scalar_type, mask, vec_lhs);
+      tree mask = vect_get_loop_mask (gsi, &LOOP_VINFO_MASKS (loop_vinfo), 1,
+				      vectype, 0);
+      tree scalar_res = gimple_build (&stmts, CFN_EXTRACT_LAST, scalar_type,
+				      mask, vec_lhs_phi);
 
       /* Convert the extracted vector element to the required scalar type.  */
       new_tree = gimple_convert (&stmts, lhs_type, scalar_res);
@@ -8023,13 +8042,32 @@  vectorizable_live_operation (stmt_vec_info stmt_info,
       tree bftype = TREE_TYPE (vectype);
       if (VECTOR_BOOLEAN_TYPE_P (vectype))
 	bftype = build_nonstandard_integer_type (tree_to_uhwi (bitsize), 1);
-      new_tree = build3 (BIT_FIELD_REF, bftype, vec_lhs, bitsize, bitstart);
+      new_tree = build3 (BIT_FIELD_REF, bftype, vec_lhs_phi, bitsize, bitstart);
       new_tree = force_gimple_operand (fold_convert (lhs_type, new_tree),
 				       &stmts, true, NULL_TREE);
     }
 
   if (stmts)
-    gsi_insert_seq_on_edge_immediate (single_exit (loop), stmts);
+    {
+      gimple_stmt_iterator exit_gsi = gsi_after_labels (exit_bb);
+      gsi_insert_before (&exit_gsi, stmts, GSI_CONTINUE_LINKING);
+
+      /* Remove existing phi from lhs and create one copy from new_tree.  */
+      tree lhs_phi = NULL_TREE;
+      gimple_stmt_iterator gsi;
+      for (gsi = gsi_start_phis (exit_bb); !gsi_end_p (gsi); gsi_next (&gsi))
+	{
+	  gimple *phi = gsi_stmt (gsi);
+	  if ((gimple_phi_arg_def (phi, 0) == lhs))
+	    {
+	      remove_phi_node (&gsi, false);
+	      lhs_phi = gimple_phi_result (phi);
+	      gimple *copy = gimple_build_assign (lhs_phi, new_tree);
+	      gsi_insert_after (&exit_gsi, copy, GSI_CONTINUE_LINKING);
+	      break;
+	    }
+	}
+    }
 
   /* Replace use of lhs with newly computed result.  If the use stmt is a
      single arg PHI, just replace all uses of PHI result.  It's necessary