On January 3, 2018 9:46:14 PM GMT+01:00, Jakub Jelinek <ja...@redhat.com> wrote:
>Hi!
>
>The ICE here is when calling multiple_p with GET_MODE_BITSIZE
>(BLKmode),
>so SIGFPE because it is 0. BLKmode can leak into the debug stmts
>through
>generic vectors without HW support, but doing say (plus:BLK ...) etc.
>just doesn't look like a valid RTL, so instead of just not trying to
>subreg
>it and not call multiple_p, this patch punts when unary/binary/ternary
>ops have BLKmode.
>
>Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
OK.
Richard.
>2018-01-03 Jakub Jelinek <ja...@redhat.com>
>
> PR debug/83621
> * cfgexpand.c (expand_debug_expr): Return NULL if mode is
> BLKmode for ternary, binary or unary expressions.
>
> * gcc.dg/pr83621.c: New test.
>
>--- gcc/cfgexpand.c.jj 2018-01-03 10:19:54.000000000 +0100
>+++ gcc/cfgexpand.c 2018-01-03 16:56:28.375179714 +0100
>@@ -4208,6 +4208,8 @@ expand_debug_expr (tree exp)
>
> binary:
> case tcc_binary:
>+ if (mode == BLKmode)
>+ return NULL_RTX;
> op1 = expand_debug_expr (TREE_OPERAND (exp, 1));
> if (!op1)
> return NULL_RTX;
>@@ -4232,6 +4234,8 @@ expand_debug_expr (tree exp)
>
> unary:
> case tcc_unary:
>+ if (mode == BLKmode)
>+ return NULL_RTX;
> inner_mode = TYPE_MODE (TREE_TYPE (TREE_OPERAND (exp, 0)));
> op0 = expand_debug_expr (TREE_OPERAND (exp, 0));
> if (!op0)
>--- gcc/testsuite/gcc.dg/pr83621.c.jj 2018-01-03 17:01:21.244249712
>+0100
>+++ gcc/testsuite/gcc.dg/pr83621.c 2018-01-03 17:00:56.414243769 +0100
>@@ -0,0 +1,12 @@
>+/* PR debug/83621 */
>+/* { dg-do compile } */
>+/* { dg-options "-O -g" } */
>+
>+typedef int __attribute__ ((__vector_size__ (64))) V;
>+V v;
>+
>+void
>+foo ()
>+{
>+ V u = v >> 1;
>+}
>
> Jakub
