The following testcase ICEs (and only without -g), because we don't replace
one VEC_COND_EXPR with .VCOND* call.
We don't do that because gimple_expand_vec_set_expr adds some stmts before
*gsi and then uses gsi_remove to remove it. gsi_remove moves the iterator
to the next stmt and in the caller we then do gsi_next before looking at
another stmt, which means we can skip processing of one stmt, which in this
case happened to be a VEC_COND_EXPR but with -g is some debug stmt in
between. As we always emit some stmts before it, it is easy to update the
iterator to the last stmt emitted there, so that caller continues really
with the next stmt.
2022-05-10 Jakub Jelinek <jakub@redhat.com>
PR tree-optimization/105528
* gimple-isel.cc (gimple_expand_vec_set_expr): After gsi_remove
set *gsi to gsi_for_stmt (ass_stmt). Fix up function comment.
* gcc.dg/pr105528.c: New test.
(cherry picked from commit
ddd46293e2b508b260eefe1adbad6720df7a5dd2)
/* Expand all ARRAY_REF(VIEW_CONVERT_EXPR) gimple assignments into calls to
internal function based on vector type of selected expansion.
i.e.:
- VIEW_CONVERT_EXPR<int[4]>(u)[_1] = = i_4(D);
+ VIEW_CONVERT_EXPR<int[4]>(u)[_1] = i_4(D);
=>
_7 = u;
_8 = .VEC_SET (_7, i_4(D), _1);
if (gsi_remove (gsi, true)
&& gimple_purge_dead_eh_edges (bb))
cfg_changed = true;
+ *gsi = gsi_for_stmt (ass_stmt);
}
}
--- /dev/null
+/* PR tree-optimization/105528 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -Wno-psabi -fcompare-debug" } */
+/* { dg-additional-options "-mavx512f" { target i?86-*-* x86_64-*-* } } */
+
+typedef unsigned V __attribute__((__vector_size__ (64)));
+V g;
+
+V
+bar (V v)
+{
+ V w;
+ v <<= (V){(V){}[53]} >= v & 5;
+ w[w[5]] -= ~0;
+ v %= ~0;
+ return v + w;
+}
+
+void
+foo (void)
+{
+ g -= (V){bar((V){~0})[3]};
+}