[llvm-dev] Function LICM for readonly, nocapture functions

[Xin Tong via llvm-dev]

Hi William

LICM needs some improvements to be able to move strlen out of the loop
here. More specifically, LICM is currently only able to hoist these cases
for calls.

1. readnone - in such case we can move the call, granted other conditions
are met, e.g. its speculatable, etc
2. readonly+onlyaccessargumentpointee - in such case, we make sure the
argument pointees are not modified in the loop.
3. readonly+nowrites in the loop - writes are checked using aliassets here.

Basically, we need to use mod-ref information to tell whether we can hoist
a call or not.

-Xin
Software Engineer - Compiler Toolchain.
Employee of Facebook Inc.

On Wed, Apr 26, 2017 at 11:42 PM, William Moses via llvm-dev <
llvm-dev at lists.llvm.org> wrote:

> Hey all,
>
> I was doing some investigation of LICM and I ran into something that seems
> a bit odd to me.
>
> Suppose I was looking at the following code snippet:
>
> #define N 1000
>
> int main() {
>
>   int B[N];
>
>   char A[N];
>
>   for(int i=0; i<N; i++) {
>
>     B[i] = strlen(A);
>
>   }
>
>   return B[0]+B[N-1];
>
> }
>
> Among other optimizations that I may want to happen, I'd hope that the
> call to strlen could be moved outside the loop via LICM. Unfortunately,
> this is not the case. Looking at the IR I found the following.
>
> The function strlen is marked readonly and the argument is marked
> nocapture.
>
> ; Function Attrs: nounwind readonly
>
> declare i64 @strlen(i8* nocapture) local_unnamed_addr #2
>
> attributes #2 = { nounwind readonly ... }
>
>
> Presumably this should be enough to permit LICM to move it out of the
> loop. Upon a further investigation, the issue appeared to be that
> canSinkOrHoistInst didn't figure out that the call could be moved.
>
>
> Curious as to see what I might be able to do to be able to do to force it
> to happen, I declared my own version of the strlen as follows (which,
> though technically incorrect as it read memory, did confirm that LICM isn't
> being stopped for some other reason). In the IR, strlen was marked as
> readnone which presumably hit the "does not access memory" fast-exit case
> in canSinkOrHoistInst.
>
> __attribute__((const)) int strlen(char* A);
>
> Is there a reason why LICM isn't occurring here -- or is just that there's
> some more work to be done on alias analysis?
>
> Cheers,
> Billy Moses
>
> ᐧ
>
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://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/20170519/b6157604/attachment.html>