On Fri, 29 May 2015, Jakub Jelinek wrote:
> 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?
Ok.
Thanks,
Richard.
> 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
>
>
--
Richard Biener <rguent...@suse.de>
SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Dilip Upmanyu, Graham
Norton, HRB 21284 (AG Nuernberg)
