Disable some VEC_COND_EXPR transformations after vector lowering

ARM understands VEC_COND_EXPR<v == w, -1, 0> but not a plain v == w which is
fed to something other than VEC_COND_EXPR (say BIT_IOR_EXPR). This patch avoids
introducing the second kind of statement after the vector lowering pass, which
is the last chance to turn v == w back into something the target handles.

This is just a workaround to avoid ICEs, a v == w produced before vector
lowering will yield pretty bad code. Either the arm target needs to learn to
handle vector comparisons (aarch64 already does), or the middle-end needs to
fall back to vcond when plain comparisons are not supported (or ...).

2020-08-07  Marc Glisse  <marc.glisse@inria.fr>

	* generic-match-head.c (optimize_vectors_before_lowering_p): New
	function.
	* gimple-match-head.c (optimize_vectors_before_lowering_p):
	Likewise.
	* match.pd ((v ? w : 0) ? a : b, c1 ? c2 ? a : b : b): Use it.

1 files changed, 10 insertions, 0 deletions

diff --git a/gcc/generic-match-head.c b/gcc/generic-match-head.c
index 2454baa..fdb528d 100644
--- a/gcc/generic-match-head.c
+++ b/gcc/generic-match-head.c
@@ -80,6 +80,16 @@ canonicalize_math_after_vectorization_p ()
   return false;
 }
 
+/* Return true if we can still perform transformations that may introduce
+   vector operations that are not supported by the target. Vector lowering
+   normally handles those, but after that pass, it becomes unsafe.  */
+
+static inline bool
+optimize_vectors_before_lowering_p ()
+{
+  return true;
+}
+
 /* Return true if successive divisions can be optimized.
    Defer to GIMPLE opts.  */