Re: [patch tree-ssa-math-opts]: Improve bswap detection

[Richard Guenther <richard.guenther@gmail.com>]

On Mon, Jun 27, 2011 at 11:50 AM, Kai Tietz <ktietz70@googlemail.com> wrote:
> 2011/6/27 Richard Guenther <richard.guenther@gmail.com>:
>> On Mon, Jun 27, 2011 at 10:06 AM, Kai Tietz <ktietz@redhat.com> wrote:
>>> Hello,
>>>
>>> this patch improves the bswap (32/64) detection and fixes an issue about shifted bit values out of type-size precision.
>>>
>>> ChangeLog
>>>
>>> 2011-06-27  Kai Tietz  <ktietz@redhat.com>
>>>
>>>        * tree-ssa-math-opts.c (do_shift_rotate): Zero bits
>>>        out of type precision after operation.
>>>        (find_bswap): Take for limit value the integer auto-
>>>        promotion into account.
>>>
>>> ChangeLog
>>>
>>> 2011-06-27  Kai Tietz  <ktietz@redhat.com>
>>>
>>>        * gcc.dg/optimize-bswapdi-2.c: New test.
>>>
>>> Bootstrapped and regression tested for x86_64-pc-linux-gnu. Ok for apply?
>>
>> The do_shift_rotate hunk is ok.  Where did you get the testcase from?
>> I can easily construct testcases that go via an intermediate SImode step
>> and the still won't be recongized.  Thus, is the testcase practically
>> relevant?  If so we should add 1 + log2 (byte-size) instead of 3 to also
>> cover the SImode -> HImode intermediate case.
>>
>> Thanks,
>> Richard.
>
> Ok, I can adjust it to use here 1 + 3 (for byte-size covering the
> SImode -> HImode intermediate case).  I was seen a regression for
> bswap detection by doing in forward-propagate a type-sinking on (type)
> X op (type) Y (for X and Y with compatible type), and (type) X op CST
> (for ((type) (type-x) CST) == CST case.  I will sent soon a patch for
> this optimization related thing. By this I detected that bswap
> algorithm didn't zero'ed upper bits out of size-range.
>
> Ok with that change?

Ok with using 1 + ceil_log2 (TREE_INT_CST_LOW (TYPE_SIZE_UNIT (...))),
CSE that TREE_INT_CST_LOW.

Thanks,
Richard.

> Regards,
> Kai
>