On Thu, Mar 3, 2016 at 9:16 PM, Jakub Jelinek <ja...@redhat.com> wrote:
> Hi!
>
> Before my recent decide_alg change, *dynamic_check == -1 was indeed
> guaranteed, because any_alg_usable_p doesn't depend on the arguments of
> decide_alg that might change during recursive call, so we'd only recurse if
> it wouldn't set *dynamic_check. But, if we give up because we'd otherwise
> recurse infinitely, we can set *dynamic_check to 128.
>
> Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux, ok for
> trunk?
>
> 2016-03-03 Jakub Jelinek <ja...@redhat.com>
>
> PR target/70062
> * config/i386/i386.c (decide_alg): If
> TARGET_INLINE_STRINGOPS_DYNAMICALLY, allow *dynamic_check to be also
> 128 from the recursive call.
>
> * gcc.target/i386/pr70062.c: New test.
I don't like the fact that *dynamic_check is set to max (which is 0
with your testcase) when recursion avoidance code already set it to
"something reasonable", together with loop_1_byte alg. What do you
think about attached (lightly tested) patch?
Uros.
diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c
index 8a026ae..ded9951 100644
--- a/gcc/config/i386/i386.c
+++ b/gcc/config/i386/i386.c
@@ -26170,11 +26170,22 @@ decide_alg (HOST_WIDE_INT count, HOST_WIDE_INT
expected_size,
}
alg = decide_alg (count, new_expected_size, min_size, max_size, memset,
zero_memset, have_as, dynamic_check, noalign);
- gcc_assert (*dynamic_check == -1);
+
if (TARGET_INLINE_STRINGOPS_DYNAMICALLY)
- *dynamic_check = max;
+ {
+ /* *dynamic_check could be set to 128 above because we avoided
+ infinite recursion. */
+ if (*dynamic_check == 128)
+ gcc_assert (alg == loop_1_byte);
+ else
+ {
+ gcc_assert (*dynamic_check == -1);
+ *dynamic_check = max;
+ }
+ }
else
- gcc_assert (alg != libcall);
+ gcc_assert (alg != libcall && *dynamic_check == -1);
+
return alg;
}
return (alg_usable_p (algs->unknown_size, memset, have_as)
