llvm dev - Jun 2010 - [LLVMdev] [llvmdev] Whole function SelectionDAG

At first I'll try the simplest strategy: put the instruction in the machine
BB corresponding to the LLVM BB where it comes from.

To implement this, I plan to add an operand to non-passive node which points
to the BasicBlockSDNode which the node belongs to.

Another idea is to use a side map to map each node to its block. But I guess
it's difficult to keep the map consistent when transforming the DAG.

On Wed, Jun 9, 2010 at 1:39 AM, Dan Gohman <gohman at apple.com> wrote:
> The first question is code placement.  How is the compiler going to decide
> which block (or blocks) to emit an instruction in?  The answer to that will
> help determine how control relationships should be represented.
>
> Dan
>
>
> On Jun 7, 2010, at 11:56 PM, Zhongxing Xu <xuzhongxing at gmail.com>
wrote:
>
>  I am trying to build a DAG for a whole function. The first problem I met
>> is to assign a user for the last instruction of each basic block, be it
>> BRCOND or other node. There is no natural user for such nodes. Without
user,
>> it will be removed in later phases. My idea is to use it as chain for
nodes
>> in the next basic block. Is this reasonable?
>>
>> -Zhongxing Xu
>> _______________________________________________
>> 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/20100609/490d7bf7/attachment.html>

I have another idea: let the BasicBlockSDNode points to the last node in its
predecessor blocks. But this will introduce cycles into the selection DAG,
making it not a DAG anymore.

So what's the impact of a non-DAG on the existing DAG legalizers and
combiner?

On Wed, Jun 9, 2010 at 9:33 AM, Zhongxing Xu <xuzhongxing at gmail.com>
wrote:
> At first I'll try the simplest strategy: put the instruction in the
machine
> BB corresponding to the LLVM BB where it comes from.
>
> To implement this, I plan to add an operand to non-passive node which
> points to the BasicBlockSDNode which the node belongs to.
>
> Another idea is to use a side map to map each node to its block. But I
> guess it's difficult to keep the map consistent when transforming the
DAG.
>
>
> On Wed, Jun 9, 2010 at 1:39 AM, Dan Gohman <gohman at apple.com>
wrote:
>
>> The first question is code placement.  How is the compiler going to
decide
>> which block (or blocks) to emit an instruction in?  The answer to that
will
>> help determine how control relationships should be represented.
>>
>> Dan
>>
>>
>> On Jun 7, 2010, at 11:56 PM, Zhongxing Xu <xuzhongxing at
gmail.com> wrote:
>>
>>  I am trying to build a DAG for a whole function. The first problem I
met
>>> is to assign a user for the last instruction of each basic block,
be it
>>> BRCOND or other node. There is no natural user for such nodes.
Without user,
>>> it will be removed in later phases. My idea is to use it as chain
for nodes
>>> in the next basic block. Is this reasonable?
>>>
>>> -Zhongxing Xu
>>> _______________________________________________
>>> 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/20100611/266f0821/attachment.html>

On Jun 8, 2010, at 8:33 PM, Zhongxing Xu <xuzhongxing at gmail.com> wrote:
> At first I'll try the simplest strategy: put the instruction in the  
> machine BB corresponding to the LLVM BB where it comes from.
>
> To implement this, I plan to add an operand to non-passive node  
> which points to the BasicBlockSDNode which the node belongs to.
This would inhibit cross-block CSE, which would be one of the main  
benefits of whole-function processing.  What problem are you trying to  
solve?
>
>
> Another idea is to use a side map to map each node to its block. But  
> I guess it's difficult to keep the map consistent when transforming  
> the DAG.
Yep.

Dan

On Jun 11, 2010, at 12:48 AM, Zhongxing Xu <xuzhongxing at gmail.com>  
wrote:
> I have another idea: let the BasicBlockSDNode points to the last  
> node in its predecessor blocks. But this will introduce cycles into  
> the selection DAG, making it not a DAG anymore.
>
> So what's the impact of a non-DAG on the existing DAG legalizers and  
> combiner?
Lots of stuff would break, for example anything that uses the  
topological sort facility.

Dan