diff mbox

Fix PR77916

Message ID a5f89911-6d50-43cc-40f4-a81e8ef2a3a8@linux.vnet.ibm.com
State New
Headers show

Commit Message

Bill Schmidt Oct. 19, 2016, 1:43 p.m. UTC 
Hi,

The previous fix for https://gcc.gnu.org/bugzilla/show_bug.cgi?id=77916
ensured that record_increment would use normal multiplier initializers
for increments of -1 with unknown strides, just as for all other negative
increments.  However, we still had a problem in insert_initializers wherein
the test for whether the replacement will need a POINTER_PLUS_EXPR was not
accurate.  This patch fixes this, and again removes the stopgap solution
so that we can process -1 increments with pointer types.

Markus Trippelsdorf kindly tested all of the ffmpeg code on x86_64 with
-march=amdfam10 to ensure there are no more of these issues lurking there.
That has been a fertile environment for these kinds of negative pointer
increments.

Bootstrapped and tested on powerpc64le-unknown-linux-gnu with no new
regressions, committed.

Thanks,
Bill


2016-10-19  Bill Schmidt  <wschmidt@linux.vnet.ibm.com>

	PR tree-optimization/77916
	PR tree-optimization/77937
	* gimple-ssa-strength-reduction.c (analyze_increments): Remove
	stopgap fix.
	(insert_initializers): Requirement of initializer for -1 should be
	based on pointer-typedness of the candidate basis.
diff mbox

Patch

Index: gcc/gimple-ssa-strength-reduction.c
===================================================================
--- gcc/gimple-ssa-strength-reduction.c	(revision 241305)
+++ gcc/gimple-ssa-strength-reduction.c	(working copy)
@@ -2825,10 +2825,6 @@  analyze_increments (slsr_cand_t first_dep, machine
 		   && !POINTER_TYPE_P (first_dep->cand_type)))
 	incr_vec[i].cost = COST_NEUTRAL;
 
-      /* FIXME: Still having trouble with pointers with a -1 increment.  */
-      else if (incr == -1 && POINTER_TYPE_P (first_dep->cand_type))
-	incr_vec[i].cost = COST_INFINITE;
-
       /* FORNOW: If we need to add an initializer, give up if a cast from
 	 the candidate's type to its stride's type can lose precision.
 	 This could eventually be handled better by expressly retaining the
@@ -3115,7 +3111,7 @@  insert_initializers (slsr_cand_t c)
       if (!profitable_increment_p (i)
 	  || incr == 1
 	  || (incr == -1
-	      && gimple_assign_rhs_code (c->cand_stmt) != POINTER_PLUS_EXPR)
+	      && (!POINTER_TYPE_P (lookup_cand (c->basis)->cand_type)))
 	  || incr == 0)
 	continue;