llvm dev - Jun 2014 - [LLVMdev] Failed to Unroll a Seemingly Simple Loop

Hi,

I found LLVM cannot unroll the loop in the example below, while gcc can.
Before I dig more about this issue, is this behavior as designed?

bool bar(int i);

void foo(int *a, int x, int y) {
  for (int i = 0; i < 4; ++i) {
    if (bar(i)) {
      break;
    }
    a[i] = i;
  }
}

Btw, if s/break/continue, LLVM is able to unroll it.

Thanks,
Jingyue
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20140622/5cad23cb/attachment.html>

On Sun, Jun 22, 2014 at 11:05 AM, Jingyue Wu <jingyue at google.com>
wrote:
> I found LLVM cannot unroll the loop in the example below, while gcc can.
> Before I dig more about this issue, is this behavior as designed?
It doesn't unroll the loop because the trip count is not known at
compile time (it could exit early because of the break), and by
default loops with runtime loop counts are not unrolled.  However,
even specifying -unroll-runtime still doesn't get the loop to unroll.
Looking deeper it looks like runtime loop unrolling is only performed
if the loop has a single exit.  See:
Transforms/Util/LoopUnroll.cpp:211 and
Transforms/Util/LoopUnrollRuntime.cpp:231.

It'd be nice if an optimization missed remark was emitted.  Certainly
a remark should be emitted in the case where you give a unroll pragma
to specifically unroll the loop.  Can you file a PR about this remark
issue and assign it to me?

thanks,
Mark
>
> bool bar(int i);
>
> void foo(int *a, int x, int y) {
>   for (int i = 0; i < 4; ++i) {
>     if (bar(i)) {
>       break;
>     }
>     a[i] = i;
>   }
> }
>
> Btw, if s/break/continue, LLVM is able to unroll it.
>
> Thanks,
> Jingyue
>

Nice find, Mark! Thanks a lot.

I filed http://llvm.org/bugs/show_bug.cgi?id=20100 to track the remark
issue.

Jingyue


On Sun, Jun 22, 2014 at 11:53 AM, Mark Heffernan <meheff at google.com>
wrote:
> On Sun, Jun 22, 2014 at 11:05 AM, Jingyue Wu <jingyue at google.com>
wrote:
> > I found LLVM cannot unroll the loop in the example below, while gcc
can.
> > Before I dig more about this issue, is this behavior as designed?
>
> It doesn't unroll the loop because the trip count is not known at
> compile time (it could exit early because of the break), and by
> default loops with runtime loop counts are not unrolled.  However,
> even specifying -unroll-runtime still doesn't get the loop to unroll.
> Looking deeper it looks like runtime loop unrolling is only performed
> if the loop has a single exit.  See:
> Transforms/Util/LoopUnroll.cpp:211 and
> Transforms/Util/LoopUnrollRuntime.cpp:231.
>
> It'd be nice if an optimization missed remark was emitted.  Certainly
> a remark should be emitted in the case where you give a unroll pragma
> to specifically unroll the loop.  Can you file a PR about this remark
> issue and assign it to me?
>
> thanks,
> Mark
>
> >
> > bool bar(int i);
> >
> > void foo(int *a, int x, int y) {
> >   for (int i = 0; i < 4; ++i) {
> >     if (bar(i)) {
> >       break;
> >     }
> >     a[i] = i;
> >   }
> > }
> >
> > Btw, if s/break/continue, LLVM is able to unroll it.
> >
> > Thanks,
> > Jingyue
> >
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20140622/bcda70ce/attachment.html>