This is the mail archive of the
gcc-bugs@gcc.gnu.org
mailing list for the GCC project.
[Bug target/70429] Wrong code with -O1.
- From: "jakub at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>
- To: gcc-bugs at gcc dot gnu dot org
- Date: Tue, 29 Mar 2016 08:40:50 +0000
- Subject: [Bug target/70429] Wrong code with -O1.
- Auto-submitted: auto-generated
- References: <bug-70429-4 at http dot gcc dot gnu dot org/bugzilla/>
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70429
Jakub Jelinek <jakub at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Target Milestone|--- |4.9.4
--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
(In reply to Kirill Yukhin from comment #4)
> Seems like combiner performs invalid reassociation. This trivial addition to
> Jakub's PR70222 fix makes test work:
>
> --- a/gcc/combine.c
> +++ b/gcc/combine.c
> @@ -10526,7 +10526,7 @@ simplify_shift_const_1 (enum rtx_code code,
> machine_mode result_mode,
> {
> /* For ((unsigned) (cstULL >> count)) >> cst2 we have to make
> sure the result will be masked. See PR70222. */
> - if (code == LSHIFTRT
> + if ((code == LSHIFTRT || code == ASHIFTRT)
> && mode != result_mode
> && !merge_outer_ops (&outer_op, &outer_const, AND,
> GET_MODE_MASK (result_mode)
That looks wrong.
Consider following testcase:
__attribute__((noinline, noclone)) int
foo (int a)
{
return (int) (0x14ff6e2207db5d1fLL >> a) >> 4;
}
int
main ()
{
if (sizeof (int) != 4 || sizeof (long long) != 8 || __CHAR_BIT__ != 8)
return 0;
if (foo (1) != 0x3edae8 || foo (2) != -132158092)
__builtin_abort ();
return 0;
}
With your above patch, it will try to combine:
Trying 6, 7 -> 8:
Failed to match this instruction:
(parallel [
(set (reg:SI 91)
(and:SI (subreg:SI (ashiftrt:DI (const_int 94565568266221009
[0x14ff6e2207db5d1])
(subreg:QI (reg/v:SI 90 [ a ]) 0)) 0)
(const_int 268435455 [0xfffffff])))
(clobber (reg:CC 17 flags))
])
Failed to match this instruction:
(set (reg:SI 91)
(and:SI (subreg:SI (ashiftrt:DI (const_int 94565568266221009
[0x14ff6e2207db5d1])
(subreg:QI (reg/v:SI 90 [ a ]) 0)) 0)
(const_int 268435455 [0xfffffff])))
and only because there is no such insn on x86, it will not miscompile it.
ASHIFTRT needs corresponding sign extension, not AND.