Hi!
On the following testcase from the PR, dce2 turns
# .MEM_21 = PHI <.MEM_20(5), .MEM_4(D)(2)>
into
# .MEM_21 = PHI <.MEM_21(5), .MEM_4(D)(2)>
because all writes in the loop have been dead code eliminated.
But only the phicprop2 pass much later is able to optimize this
out (by changing all .MEM_21 vuses into .MEM_4(D) vuses and removing
the virtual phi stmt). In between those passes is if-conversion though,
and that just gives up when it sees it. I think the test only wants
to detect conditional stores, so the above useless virtual phis
don't really pose a problem to it.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
2015-05-29 Jakub Jelinek <ja...@redhat.com>
PR tree-optimization/66142
* tree-if-conv.c (if_convertible_phi_p): Don't give up on
virtual phis that feed themselves.
* gcc.dg/vect/pr66142.c: New test.
--- gcc/tree-if-conv.c.jj 2015-05-21 21:16:18.000000000 +0200
+++ gcc/tree-if-conv.c 2015-05-29 11:24:06.726448462 +0200
@@ -594,7 +594,8 @@ if_convertible_phi_p (struct loop *loop,
FOR_EACH_IMM_USE_FAST (use_p, imm_iter, gimple_phi_result (phi))
{
- if (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI)
+ if (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI
+ && USE_STMT (use_p) != (gimple) phi)
{
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "Difficult to handle this virtual phi.\n");
--- gcc/testsuite/gcc.dg/vect/pr66142.c.jj 2015-05-29 11:31:21.563628051
+0200
+++ gcc/testsuite/gcc.dg/vect/pr66142.c 2015-05-29 11:31:00.000000000 +0200
@@ -0,0 +1,44 @@
+/* PR middle-end/66142 */
+/* { dg-do compile } */
+/* { dg-additional-options "-ffast-math -fopenmp-simd" } */
+/* { dg-additional-options "-mavx" { target avx_runtime } } */
+
+struct A { float x, y; };
+struct B { struct A t, w; };
+
+static inline float
+bar (const struct B *x)
+{
+ struct A r;
+ float b, c, d;
+ r.x = x->t.x;
+ r.y = x->t.y;
+ b = r.x * x->w.x + r.y * x->w.y;
+ c = b + r.x * r.x + r.y * r.y;
+ if (c > 0.0f)
+ return c + 3.0f;
+ return 0.0f;
+}
+
+void
+foo (float *a, float *b, float *c)
+{
+ int i;
+ float z = 0.0f;
+ float u = *a;
+#pragma omp simd
+ for (i = 0; i < 32; i++)
+ {
+ float x = b[i];
+ float y = c[i];
+ struct B r;
+ r.t.x = 1.0f;
+ r.t.y = u;
+ r.w.x = x;
+ r.w.y = y;
+ z += bar (&r);
+ }
+ *a = z;
+}
+
+/* { dg-final { scan-tree-dump-times "vectorized 1 loops in function" 1 "vect"
{ target vect_condition } } } */
Jakub
