[prev in list] [next in list] [prev in thread] [next in thread] 

List:       openjdk-hotspot-compiler-dev
Subject:    Re: Suboptimal code generation in C2?
From:       Vitaly Davidovich <vitalyd () gmail ! com>
Date:       2015-09-24 1:31:34
Message-ID: CAHjP37Gv_=1GM7+2bJNtB0M9cW4_09eh8x0g+-g+m1v8OyyANw () mail ! gmail ! com
[Download RAW message or body]

Great, thank you guys.

sent from my phone
On Sep 23, 2015 9:09 PM, "Vladimir Kozlov" <vladimir.kozlov@oracle.com>
wrote:

> Could be loop predicates or C2 splits loop (pre-main-post) before we
> collapsed it. But I agree that we could do better.
>
> Aleksey, thank you for filing RFE.
>
> Thanks,
> Vladimir
>
> On 9/24/15 3:04 AM, Vitaly Davidovich wrote:
>
>> Thanks Aleksey.
>>
>> The cmov is probably a non-issue here since both dependencies are in
>> registers.  The arithmetic dance, jumps, and register allocation seems
>> off though.  For instance, esi is cleared with xor then it jumps to code
>> that adds edx to esi, then it moves esi into eax - this could be
>> replaced with moving edx into eax.
>>
>> sent from my phone
>>
>> Hi,
>>
>> On 09/23/2015 08:49 PM, Vitaly Davidovich wrote:
>>  > Consider this simple method:
>>  >
>>  >     private static int test(final int ops, int start) {
>>  >         for (int i = 0; i < ops; ++i) {
>>  >             start++;
>>  >         }
>>  >         return start;
>>  >     }
>>
>>  > test(int, int):
>>  > leal(%rsi,%rdi), %edx
>>  > movl%esi, %eax
>>  > testl%edi, %edi
>>  > cmovg%edx, %eax
>>  > ret
>>  >
>>  > Any ideas on why C2 doesn't generate roughly the same thing? Clang 3.7
>>  > and ICC 13.0.1 are almost the same as the GCC output above.
>>
>> Yup, could be better:
>> https://bugs.openjdk.java.net/browse/JDK-8137049
>> <https://bugs.openjdk.java.net/browse/JDK-8137049>
>>
>> I am not entirely sure conditional move is better than a branch in this
>> particular case, since $ops is probably always non-negative in practice.
>> HotSpot generates a branch there, which is good. The arithmetic dance is
>> indeed not very clean -- I'd speculate that's a leftover from loop
>> peeling.
>>
>> Thanks,
>> -Aleksey
>>
>>
>> sent from my phone
>>
>>

[Attachment #3 (text/html)]

<p dir="ltr">Great, thank you guys. </p>
<p dir="ltr">sent from my phone</p>
<div class="gmail_quote">On Sep 23, 2015 9:09 PM, &quot;Vladimir Kozlov&quot; &lt;<a \
href="mailto:vladimir.kozlov@oracle.com">vladimir.kozlov@oracle.com</a>&gt; wrote:<br \
type="attribution"><blockquote class="gmail_quote" style="margin:0 0 0 \
.8ex;border-left:1px #ccc solid;padding-left:1ex">Could be loop predicates or C2 \
splits loop (pre-main-post) before we collapsed it. But I agree that we could do \
better.<br> <br>
Aleksey, thank you for filing RFE.<br>
<br>
Thanks,<br>
Vladimir<br>
<br>
On 9/24/15 3:04 AM, Vitaly Davidovich wrote:<br>
<blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc \
solid;padding-left:1ex"> Thanks Aleksey.<br>
<br>
The cmov is probably a non-issue here since both dependencies are in<br>
registers.   The arithmetic dance, jumps, and register allocation seems<br>
off though.   For instance, esi is cleared with xor then it jumps to code<br>
that adds edx to esi, then it moves esi into eax - this could be<br>
replaced with moving edx into eax.<br>
<br>
sent from my phone<br>
<br>
Hi,<br>
<br>
On 09/23/2015 08:49 PM, Vitaly Davidovich wrote:<br>
  &gt; Consider this simple method:<br>
  &gt;<br>
  &gt;        private static int test(final int ops, int start) {<br>
  &gt;              for (int i = 0; i &lt; ops; ++i) {<br>
  &gt;                    start++;<br>
  &gt;              }<br>
  &gt;              return start;<br>
  &gt;        }<br>
<br>
  &gt; test(int, int):<br>
  &gt; leal(%rsi,%rdi), %edx<br>
  &gt; movl%esi, %eax<br>
  &gt; testl%edi, %edi<br>
  &gt; cmovg%edx, %eax<br>
  &gt; ret<br>
  &gt;<br>
  &gt; Any ideas on why C2 doesn&#39;t generate roughly the same thing? Clang 3.7<br>
  &gt; and ICC 13.0.1 are almost the same as the GCC output above.<br>
<br>
Yup, could be better:<br>
<a href="https://bugs.openjdk.java.net/browse/JDK-8137049" rel="noreferrer" \
target="_blank">https://bugs.openjdk.java.net/browse/JDK-8137049</a><br> &lt;<a \
href="https://bugs.openjdk.java.net/browse/JDK-8137049" rel="noreferrer" \
target="_blank">https://bugs.openjdk.java.net/browse/JDK-8137049</a>&gt;<br> <br>
I am not entirely sure conditional move is better than a branch in this<br>
particular case, since $ops is probably always non-negative in practice.<br>
HotSpot generates a branch there, which is good. The arithmetic dance is<br>
indeed not very clean -- I&#39;d speculate that&#39;s a leftover from loop \
peeling.<br> <br>
Thanks,<br>
-Aleksey<br>
<br>
<br>
sent from my phone<br>
<br>
</blockquote>
</blockquote></div>



[prev in list] [next in list] [prev in thread] [next in thread]