llvm dev - Aug 2014 - [LLVMdev] Any Optimization Suggestion to Get Rid of AddrSpaceCast around PHI

In the following example, for some reasons, the input pointer entering the
loop was casted to generic pointer. How can the backend get rid of the
addrspacecast and use local store in the loop?


for.body.lr.ph: ; preds = %entry

%0 = addrspacecast i32 addrspace(3)* %in to i32 addrspace(4)*

br label %for.body

for.body: ; preds = %for.body, %for.body.lr.ph

%i.03 = phi i32 [ 0, %for.body.lr.ph ], [ %inc, %for.body ]

%ptr.02 = phi i32 addrspace(4)* [ %0, %for.body.lr.ph ], [ %add.ptr,
%for.body ]

store i32 %i.03, i32 addrspace(4)* %ptr.02, align 4

%add.ptr = getelementptr inbounds i32 addrspace(4)* %ptr.02, i64 4

%inc = add i32 %i.03, 1

%exitcond = icmp eq i32 %inc, %numElems

br i1 %exitcond, label %for.end, label %for.body

for.end: ; preds = %


Thanks;

Changpeng
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20140821/a0d9e48b/attachment.html>

It's fairly simple to find all of the casts and then walk down the use
chains, identifying whether the pointers escape, and if they don't rewrite
all of the instructions to use the other address space.  I've done this in
an (extremely hacky) pass to allow allocas to be in an address space other than
0 (which I'm slowly replacing with much less hacky code).

As I understand it, for your target stores in AS 4 are more expensive than
stores in AS 3, but both can refer to the same memory?

David

On 21 Aug 2014, at 23:11, Changpeng Fang <cfang.llvm at gmail.com> wrote:
> In the following example, for some reasons, the input pointer entering the
loop was casted to generic pointer. How can the backend get rid of the
> addrspacecast and use local store in the loop?
>  
>  
> for.body.lr.ph: ; preds = %entry
> %0 = addrspacecast i32 addrspace(3)* %in to i32 addrspace(4)*
> 
> br label %for.body
> 
> for.body: ; preds = %for.body, %for.body.lr.ph
> 
> %i.03 = phi i32 [ 0, %for.body.lr.ph ], [ %inc, %for.body ]
> 
> %ptr.02 = phi i32 addrspace(4)* [ %0, %for.body.lr.ph ], [ %add.ptr,
%for.body ]
> 
> store i32 %i.03, i32 addrspace(4)* %ptr.02, align 4
> 
> %add.ptr = getelementptr inbounds i32 addrspace(4)* %ptr.02, i64 4
> 
> %inc = add i32 %i.03, 1
> 
> %exitcond = icmp eq i32 %inc, %numElems
> 
> br i1 %exitcond, label %for.end, label %for.body
> 
> for.end: ; preds = %
> 
> 
> 
> Thanks;
>  
> Changpeng 
> 
>  
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev

On Fri, Aug 22, 2014 at 1:11 AM, David Chisnall <David.Chisnall at
cl.cam.ac.uk
> wrote:
> It's fairly simple to find all of the casts and then walk down the use
> chains, identifying whether the pointers escape, and if they don't
rewrite
> all of the instructions to use the other address space.  I've done this
in
> an (extremely hacky) pass to allow allocas to be in an address space other
> than 0 (which I'm slowly replacing with much less hacky code).
>
But I got lost when a PHI node (as in the test case) is encountered.

> As I understand it, for your target stores in AS 4 are more expensive than
> stores in AS 3, but both can refer to the same memory?
>
> Yes. Load/store in generic address space in expansive. The casts also add
a small overhead in address computation.

Thanks;

Changpeng


> David
>
> On 21 Aug 2014, at 23:11, Changpeng Fang <cfang.llvm at gmail.com>
wrote:
>
> > In the following example, for some reasons, the input pointer entering
> the loop was casted to generic pointer. How can the backend get rid of the
> > addrspacecast and use local store in the loop?
> >
> >
> > for.body.lr.ph: ; preds = %entry
> > %0 = addrspacecast i32 addrspace(3)* %in to i32 addrspace(4)*
> >
> > br label %for.body
> >
> > for.body: ; preds = %for.body, %for.body.lr.ph
> >
> > %i.03 = phi i32 [ 0, %for.body.lr.ph ], [ %inc, %for.body ]
> >
> > %ptr.02 = phi i32 addrspace(4)* [ %0, %for.body.lr.ph ], [ %add.ptr,
> %for.body ]
> >
> > store i32 %i.03, i32 addrspace(4)* %ptr.02, align 4
> >
> > %add.ptr = getelementptr inbounds i32 addrspace(4)* %ptr.02, i64 4
> >
> > %inc = add i32 %i.03, 1
> >
> > %exitcond = icmp eq i32 %inc, %numElems
> >
> > br i1 %exitcond, label %for.end, label %for.body
> >
> > for.end: ; preds = %
> >
> >
> >
> > Thanks;
> >
> > Changpeng
> >
> >
> > _______________________________________________
> > LLVM Developers mailing list
> > LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> > http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20140822/8ccb24f5/attachment.html>