Fix for 56175

Richard Biener richard.guenther@gmail.com
Thu Feb 21 11:06:00 GMT 2013


On Wed, Feb 20, 2013 at 4:41 PM, Yuri Rumyantsev <ysrumyan@gmail.com> wrote:
> Richard,
>
> First of all, your proposal to move type sinking to the end of
> function does not work since we handle each statement in function and
> we want that 1st type folding of X & C will not happen.
> Note that we have the following sequence of gimple before forwprop1:
>
>    x.0_10 = (signed char) x_8;
>   _11 = x.0_10 & 1;
>   _12 = (signed char) y_9;
>   _13 = _12 & 1;
>   _14 = _11 ^ _13;

Ah, indeed.  Reminds me of some of my dead patches that separated
forwprop into a forward and backward stage.  Of course then you have
the ordering issue of whether to first forward or backward.

Which means that I bet you can construct a testcase that with
your change is no longer optimized (just make pushing the conversion
make the types _match_).  Which is always the case
with this kind of local pattern-matching transforms.

Currently forwprop processes leafs of expression trees first (well, inside
a basic-block), similar to how fold () is supposed to be operated, based
on the idea that simplified / canonicalized leafs helps keeping pattern
recognition simple and cost considerations more accurate.

When one order works better than another you always have to consider
that the user could already have written the code in a way that results
in the input that isn't well handled.

Not that this helps very much for the situation ;)

But I don't like the use of first_pass_instance ... and the fix isn't
an improvement but just a hack for the benchmark.

Richard.

> I also added comment to my fix and create new test for it. I also
> checked that this test is passed with patched compiler  only. So
> Change Log was also modified:
>
> ChangeLog
>
> 2013-02-20  Yuri Rumyantsev  <ysrumyan@gmail.com>
>
>         PR tree-optimization/56175
>         * tree-ssa-forwprop.c (simplify_bitwise_binary): Avoid type sinking
>         at 1st forwprop pass to recognize (A & C) ^ (B & C) -> (A ^ B) & C
>         for short integer types.
>         * gcc.dg/pr56175.c: New test.
>
>
>
>
> 2013/2/20 Richard Biener <richard.guenther@gmail.com>:
>> On Wed, Feb 20, 2013 at 1:00 PM, Yuri Rumyantsev <ysrumyan@gmail.com> wrote:
>>> Hi All,
>>>
>>> This patch is aimed to recognize (A & C) ^ (B & C) -> (A ^ B) & C
>>> pattern in simpify_bitwise_binary for short integer types.
>>> The fix is very simple - we simply turn off short type sinking at the
>>> first pass of forward propagation allows to get
>>> +10% speedup for important benchmark Coremark 1.0 at x86 Atom and
>>> +5-7% for other x86 platforms too.
>>> Bootstrapping and regression testing were successful on x86-64.
>>>
>>> Is it Ok for trunk?
>>
>> It definitely needs a comment before the checks.
>>
>> Also I think it simply shows that the code is placed at the wrong spot.
>> Simply moving it down in simplify_bitwise_binary to be done the very last
>> should get both of the effects done.
>>
>> Can you rework the patch according to that?
>>
>> You also miss a testcase, we should make sure to not regress again here.
>>
>> Thanks,
>> Richard.
>>
>>> ChangeLog.
>>>
>>> 2013-02-20  Yuri Rumyantsev  <ysrumyan@gmail.com>
>>>
>>>         PR tree-optimization/56175
>>>         * tree-ssa-forwprop.c (simplify_bitwise_binary) : Avoid type sinking
>>>         at 1st forwprop pass to recognize (A & C) ^ (B & C) -> (A ^ B) & C
>>>         for short integer types.



More information about the Gcc-patches mailing list