diff mbox

[fortran,57..59/66] inline sum and product: Prevent regressions: Fix {min, max}{loc, val}

Message ID 20111027233322.18581.12112@gimli.local
State New
Headers show

Commit Message

Mikael Morin Oct. 27, 2011, 11:33 p.m. UTC
Patches 58 and 59 fix the {min,max}loc and {min,max}val intrinsics which use
multiple loops. See the comments in the patches for details.
Patch 57 avoids duplicated offset calculation in the code generated.
OK?
2011-10-19  Mikael Morin  <mikael@gcc.gnu.org>

	* trans-intrinsic.c (gfc_conv_intrinsic_minmaxloc): Don't calculate
	offset twice in generated code.
2011-10-19  Mikael Morin  <mikael@gcc.gnu.org>

	* trans-intrinsic.c (gfc_conv_intrinsic_minmaxloc): Set loop's
	temporary rank to the loop rank. Mark ss chains for multiple loop
	if necessary.  Use gfc_trans_scalarized_loop_boundary to end one loop
	and start another.
2011-10-19  Mikael Morin  <mikael@gcc.gnu.org>

	* trans-intrinsic.c (gfc_conv_intrinsic_minmaxval): Set loop's
	temporary rank to the loop rank. Mark ss chains for multiple loop
	if necessary.  Use gfc_trans_scalarized_loop_boundary to end one loop
	and start another.
diff mbox

Patch

diff --git a/trans-intrinsic.c b/trans-intrinsic.c
index 506cdf2..3cdc1e0 100644
--- a/trans-intrinsic.c
+++ b/trans-intrinsic.c
@@ -3522,6 +3522,22 @@  gfc_conv_intrinsic_minmaxval (gfc_se * se, gfc_expr * expr, enum tree_code op)
 
   /* Initialize the loop.  */
   gfc_conv_ss_startstride (&loop);
+
+  /* The code generated can have more than one loop in sequence (see the
+     comment at the function header).  This doesn't work well with the
+     scalarizer, which changes arrays' offset when the scalarization loops
+     are generated (see gfc_trans_preloop_setup).  Fortunately, {min,max}val
+     are  currently inlined in the scalar case only.  As there is no dependency
+     to care about in that case, there is no temporary, so that we can use the
+     scalarizer temporary code to handle multiple loops.  Thus, we set temp_dim
+     here, we call gfc_mark_ss_chain_used with flag=3 later, and we use
+     gfc_trans_scalarized_loop_boundary even later to restore offset.
+     TODO: this prevents inlining of rank > 0 minmaxval calls, so this
+     should eventually go away.  We could either create two loops properly,
+     or find another way to save/restore the array offsets between the two
+     loops (without conflicting with temporary management), or use a single
+     loop minmaxval implementation.  See PR 31067.  */
+  loop.temp_dim = loop.dimen;
   gfc_conv_loop_setup (&loop, &expr->where);
 
   if (nonempty == NULL && maskss == NULL
@@ -3553,9 +3569,9 @@  gfc_conv_intrinsic_minmaxval (gfc_se * se, gfc_expr * expr, enum tree_code op)
 	}
     }
 
-  gfc_mark_ss_chain_used (arrayss, 1);
+  gfc_mark_ss_chain_used (arrayss, lab ? 3 : 1);
   if (maskss)
-    gfc_mark_ss_chain_used (maskss, 1);
+    gfc_mark_ss_chain_used (maskss, lab ? 3 : 1);
   /* Generate the loop body.  */
   gfc_start_scalarized_body (&loop, &body);
 
@@ -3665,15 +3681,13 @@  gfc_conv_intrinsic_minmaxval (gfc_se * se, gfc_expr * expr, enum tree_code op)
 
   if (lab)
     {
-      gfc_trans_scalarized_loop_end (&loop, 0, &body);
+      gfc_trans_scalarized_loop_boundary (&loop, &body);
 
       tmp = fold_build3_loc (input_location, COND_EXPR, type, nonempty,
 			     nan_cst, huge_cst);
       gfc_add_modify (&loop.code[0], limit, tmp);
       gfc_add_expr_to_block (&loop.code[0], build1_v (LABEL_EXPR, lab));
 
-      gfc_start_block (&body);
-
       /* If we have a mask, only add this element if the mask is set.  */
       if (maskss)
 	{