Hi!
This is the middle-end for the same PR. While supporting min/max for
avx512f V8DI generates of course better code, when the backend claims it is
not available, we should not ICE on that by trying to expand the vector
min/max by expanding comparison and jump.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
2017-03-06 Jakub Jelinek <ja...@redhat.com>
PR rtl-optimization/79901
* expr.c (expand_expr_real_2): For vector MIN/MAX, if there is no
min/max expander, expand it using expand_vec_cond_expr.
--- gcc/expr.c.jj 2017-03-02 10:15:32.000000000 +0100
+++ gcc/expr.c 2017-03-06 16:07:56.159868534 +0100
@@ -8943,6 +8943,18 @@ expand_expr_real_2 (sepops ops, rtx targ
if (temp != 0)
return temp;
+ /* For vector MIN <x, y>, expand it a VEC_COND_EXPR <x <= y, x, y>
+ and similarly for MAX <x, y>. */
+ if (VECTOR_TYPE_P (type))
+ {
+ tree t0 = make_tree (type, op0);
+ tree t1 = make_tree (type, op1);
+ tree comparison = build2 (code == MIN_EXPR ? LE_EXPR : GE_EXPR,
+ type, t0, t1);
+ return expand_vec_cond_expr (type, comparison, t0, t1,
+ original_target);
+ }
+
/* At this point, a MEM target is no longer useful; we will get better
code without it. */
Jakub
