We need to restrict what we handle as last operation in a nested
cycle because vectorizable_reduction performs the code-generation
in the end.
Boostrap and regtest in progress on x86_64-unknown-linux-gnu.
Richard.
2018-11-13 Richard Biener <rguent...@suse.de>
PR tree-optimization/87931
* tree-vect-loop.c (vect_is_simple_reduction): Restrict
nested cycles we support to latch computations vectorizable_reduction
handles.
* gcc.dg/graphite/pr87931.c: New testcase.
Index: gcc/tree-vect-loop.c
===================================================================
--- gcc/tree-vect-loop.c (revision 266061)
+++ gcc/tree-vect-loop.c (working copy)
@@ -2983,6 +2976,22 @@ vect_is_simple_reduction (loop_vec_info
if (nested_in_vect_loop && !check_reduction)
{
+ /* FIXME: Even for non-reductions code generation is funneled
+ through vectorizable_reduction for the stmt defining the
+ PHI latch value. So we have to artificially restrict ourselves
+ for the supported operations. */
+ switch (get_gimple_rhs_class (code))
+ {
+ case GIMPLE_BINARY_RHS:
+ case GIMPLE_TERNARY_RHS:
+ break;
+ default:
+ /* Not supported by vectorizable_reduction. */
+ if (dump_enabled_p ())
+ report_vect_op (MSG_MISSED_OPTIMIZATION, def_stmt,
+ "nested cycle: not handled operation: ");
+ return NULL;
+ }
if (dump_enabled_p ())
report_vect_op (MSG_NOTE, def_stmt, "detected nested cycle: ");
return def_stmt_info;
Index: gcc/testsuite/gcc.dg/graphite/pr87931.c
===================================================================
--- gcc/testsuite/gcc.dg/graphite/pr87931.c (nonexistent)
+++ gcc/testsuite/gcc.dg/graphite/pr87931.c (working copy)
@@ -0,0 +1,22 @@
+/* { dg-do compile } */
+/* { dg-options "-O3 -fno-tree-copy-prop -fgraphite-identity" } */
+
+#define N 40
+#define M 128
+float in[N+M];
+float coeff[M];
+float fir_out[N];
+
+void fir ()
+{
+ int i,j,k;
+ float diff;
+
+ for (i = 0; i < N; i++) {
+ diff = 0;
+ for (j = 0; j < M; j++) {
+ diff += in[j+i]*coeff[j];
+ }
+ fir_out[i] = diff;
+ }
+}
