With Honzas change to make the x86 backend consider the actual operation
for costing vector stmts it becomes apparent that
vect_compute_single_scalar_iteration_cost uses the old-style target
cost hook which doesn't get enough information to distinguish different
operations. This means instead of actual scalar multiplication cost
we cost a general scalar-stmt cost for the testcase for the scalar
iteration but cost a vector multiplication for the vectorized body
resulting in an apples-to-oranges comparison in the end.
Fixed as follows.
Bootstrap and regtest running on x86_64-unknown-linux-gnu.
Richard.
2018-02-27 Richard Biener <rguent...@suse.de>
PR tree-optimization/84512
* tree-vect-loop.c (vect_compute_single_scalar_iteration_cost):
Do not use the estimate returned from record_stmt_cost for
the scalar iteration cost but sum properly using add_stmt_cost.
* gcc.dg/tree-ssa/pr84512.c: New testcase.
Index: gcc/tree-vect-loop.c
===================================================================
--- gcc/tree-vect-loop.c (revision 258030)
+++ gcc/tree-vect-loop.c (working copy)
@@ -1384,16 +1384,10 @@ vect_compute_single_scalar_iteration_cos
{
struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo);
- int nbbs = loop->num_nodes, factor, scalar_single_iter_cost = 0;
+ int nbbs = loop->num_nodes, factor;
int innerloop_iters, i;
- /* Count statements in scalar loop. Using this as scalar cost for a single
- iteration for now.
-
- TODO: Add outer loop support.
-
- TODO: Consider assigning different costs to different scalar
- statements. */
+ /* Gather costs for statements in the scalar loop. */
/* FORNOW. */
innerloop_iters = 1;
@@ -1437,13 +1431,28 @@ vect_compute_single_scalar_iteration_cos
else
kind = scalar_stmt;
- scalar_single_iter_cost
- += record_stmt_cost (&LOOP_VINFO_SCALAR_ITERATION_COST (loop_vinfo),
- factor, kind, stmt_info, 0, vect_prologue);
+ record_stmt_cost (&LOOP_VINFO_SCALAR_ITERATION_COST (loop_vinfo),
+ factor, kind, stmt_info, 0, vect_prologue);
}
}
- LOOP_VINFO_SINGLE_SCALAR_ITERATION_COST (loop_vinfo)
- = scalar_single_iter_cost;
+
+ /* Now accumulate cost. */
+ void *target_cost_data = init_cost (loop);
+ stmt_info_for_cost *si;
+ int j;
+ FOR_EACH_VEC_ELT (LOOP_VINFO_SCALAR_ITERATION_COST (loop_vinfo),
+ j, si)
+ {
+ struct _stmt_vec_info *stmt_info
+ = si->stmt ? vinfo_for_stmt (si->stmt) : NULL;
+ (void) add_stmt_cost (target_cost_data, si->count,
+ si->kind, stmt_info, si->misalign,
+ vect_body);
+ }
+ unsigned dummy, body_cost = 0;
+ finish_cost (target_cost_data, &dummy, &body_cost, &dummy);
+ destroy_cost_data (target_cost_data);
+ LOOP_VINFO_SINGLE_SCALAR_ITERATION_COST (loop_vinfo) = body_cost;
}
Index: gcc/testsuite/gcc.dg/tree-ssa/pr84512.c
===================================================================
--- gcc/testsuite/gcc.dg/tree-ssa/pr84512.c (nonexistent)
+++ gcc/testsuite/gcc.dg/tree-ssa/pr84512.c (working copy)
@@ -0,0 +1,15 @@
+/* { dg-do compile } */
+/* { dg-options "-O3 -fdump-tree-optimized" } */
+
+int foo()
+{
+ int a[10];
+ for(int i = 0; i < 10; ++i)
+ a[i] = i*i;
+ int res = 0;
+ for(int i = 0; i < 10; ++i)
+ res += a[i];
+ return res;
+}
+
+/* { dg-final { scan-tree-dump "return 285;" "optimized" } } */
