llvm dev - Jul 2021 - Question about Unrolling Loop with Multiple Exits

A) Are you measuring on tip of tree?  There were changes for multiple 
exit unrolling which landed very recently.

B) One of your exits does not dominate your latch.  Those are generally 
hard.

C) This example does not seem to require gotos.  I strongly suggest 
reducing your test cases if you want more informed commentary.

Philip

On 7/16/21 7:42 AM, Jingu Kang via llvm-dev wrote:
>
> Hi All,
>
> While I am investigating benchmarks, I have found loops which llvm 
> fails to unroll because the loops have multiple exits.
>
> For example,
>
> char *foo(void);
>
> int boo(char *s);
>
> int test(char *s, char *end, char *check1, char *check2) {
>
>   while (s <= end) {
>
>     if (*s++ != '\n')
>
>       continue;
>
>     if (check1 || check2) {
>
>       s = foo();
>
>       if (!s)
>
>         goto ret1;
>
>     }
>
>     if (boo(s))
>
>       goto ret0;
>
>   }
>
>   goto ret1;
>
> ret0:
>
>   return 0;
>
> ret1:
>
>   return 1;
>
> }
>
> Above code causes below messages from LoopUnroll pass.
>
> Bailout for multi-exit handling when latch exit has >1 predecessor.
>
> Multiple exit/exiting blocks in loop and multi-exit unrolling not enabled!
>
> I can see the option `unroll-runtime-multi-exit` and comments about 
> it. I wonder there are already reviews for the work on phabriactor or 
> some people are working on it.
>
> If someone knows information about it, please share it.
>
> Thanks
>
> JinGu Kang
>
>
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20210716/abcd8085/attachment.html>

Hi Philip,

Thanks for your kind reply.
> A) Are you measuring on tip of tree?� There were changes for multiple
exit unrolling which landed very recently.
Yep, I am investigating benchmarks with llvm tip and I can see the llvm fails to
unroll some loops with multiple exits.
> B) One of your exits does not dominate your latch.� Those are generally
hard
> C) This example does not seem to require gotos.� I strongly suggest
reducing your test cases if you want more informed commentary.�
I am looking at perlbench recently and it has `goto` statements inside loop. The
example is a reduced case. When I look at the gcc's output of the example,
it looks like gcc unrolls only the below `if` statement block...

    if (*s++ != '\n')
      continue;

Thanks
JinGu Kang

From: llvm-dev <llvm-dev-bounces at lists.llvm.org> On Behalf Of Philip
Reames via llvm-dev
Sent: 16 July 2021 15:52
To: Jingu Kang <Jingu.Kang at arm.com>; llvm-dev at lists.llvm.org
Subject: Re: [llvm-dev] Question about Unrolling Loop with Multiple Exits


A) Are you measuring on tip of tree?� There were changes for multiple exit
unrolling which landed very recently.

B) One of your exits does not dominate your latch.� Those are generally
hard.�

C) This example does not seem to require gotos.� I strongly suggest reducing
your test cases if you want more informed commentary.�

Philip
On 7/16/21 7:42 AM, Jingu Kang via llvm-dev wrote:
Hi All,
�
While I am investigating benchmarks, I have found loops which llvm fails to
unroll because the loops have multiple exits.
For example,
�
char *foo(void);
int boo(char *s);
�
int test(char *s, char *end, char *check1, char *check2) {
� while (s <= end) {
��� if (*s++ != '\n')
����� continue;
��� if (check1 || check2) {
����� s = foo();
����� if (!s)
������� goto ret1;
��� }��
����if (boo(s))
����� goto ret0;
� }
� goto ret1;
�
ret0:
� return 0;
ret1:
� return 1;
}
�
Above code causes below messages from LoopUnroll pass.
�
Bailout for multi-exit handling when latch exit has >1 predecessor.
Multiple exit/exiting blocks in loop and multi-exit unrolling not enabled!
�
I can see the option `unroll-runtime-multi-exit` and comments about it. I wonder
there are already reviews for the work on phabriactor or some people are working
on it.
If someone knows information about it, please share it.
�
Thanks
JinGu Kang
�
�
�
�
�
�



_______________________________________________

LLVM Developers mailing list

llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>

https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20210716/e2f0c72a/attachment.html>