On Wed, Nov 3, 2010 at 6:12 PM, Andrew Haley <a...@redhat.com> wrote:
> On 11/03/2010 04:49 PM, Bingfeng Mei wrote:
>> Hello,
>> I came across an issue with function "optimize" attribute. The code is like:
>> __attribute__((optimize("-fno-strict-aliasing")))
>> void foo()
>> {
>> ...
>> }
>>
>> When compiling with -O2, we expect this function is compiled without
>> following
>> strict aliasing rule, whereas other code does. However, I found this
>> function still
>> has strict aliasing flag turned on during compilation. After a little
>> investigation,
>> I found the following lines of code in parse_optimize_options of c-common.c
>> (4.5, 4.6 is similar)
>>
>> ...
>> saved_flag_strict_aliasing = flag_strict_aliasing;
>>
>> /* Now parse the options. */
>> decode_options (opt_argc, opt_argv);
>>
>> targetm.override_options_after_change();
>>
>> /* Don't allow changing -fstrict-aliasing. */
>> flag_strict_aliasing = saved_flag_strict_aliasing;
>> ...
>>
>> -fstrict-aliasing is excluded from function specific optimize option. I
>> checked
>> both internal manual and gcc manual. It seems not to be documented. I wonder
>> whether there is a good reason for this? If yes, at least we should document
>> it.
>> BTW, the code certainly works in 4.4.
>
> As far as I can tell the problem was that it didn't work.
>
> See http://gcc.gnu.org/ml/gcc-patches/2008-11/msg00073.html
A proper way to implement it nowadays would be to make sure that
during gimplification we make the alias type in all MEM-REFs
ref-all if no-strict-aliasing is in effect. This would of course also require
to wrap all memory accesses in a MEM_REF tree.
Richard.
> Andrew.
>
