llvm dev - Jan 2010 - [LLVMdev] <IsKill> getting from MachineOperand is just <Used> attribute from logic.

Hi, 

I have ported LLC to a risc cpu. It can pass benchmark that I have at current. 

But I want do some optimization after register alloction by adjusting 
register using. I scan MachineBasicBlock to analyze operand's IsKill, IsDead
, IsDef attribute to get a physical register's liverange. But I get a
strange case at MBB.jpg.

  R4 is marked <kill> at MBB0.  If I scan R4's liverange by
[MBB0->MBB1->MBB2]. I will find R4 first is killed, then is used. It can
not unlogisch. Attually R4 just is <Used>. It will cause my optimization
pass crash(Actually, I ingore Live In message of MBB. I recollect live in
messges at my pass.).

  1. Does <kill> attribute of R4 at MBB0 is a unimportant  and redundancy
messages, Or a little bug???
  2. Is it unreliable to get a physical register's liverange by Kill, Dead
messages from MachineBasicBlock??

  


      ___________________________________________________________ 
  好玩贺卡等你发，邮箱贺卡全新上线！ 
http://card.mail.cn.yahoo.com/
-------------- next part --------------
A non-text attachment was scrubbed...
Name: MBB.jpg
Type: image/jpeg
Size: 26396 bytes
Desc: not available
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20100115/2c00f5cc/attachment.jpg>

On Jan 14, 2010, at 6:39 PM, 任坤 wrote:
> But I want do some optimization after register alloction by adjusting 
> register using. I scan MachineBasicBlock to analyze operand's IsKill,
IsDead , IsDef attribute to get a physical register's liverange. But I get a
strange case at MBB.jpg.
You can also look at RegisterScavenging.cpp and MachineVerifier.cpp. They are
doing the same thing.
>  R4 is marked <kill> at MBB0.  If I scan R4's liverange by
[MBB0->MBB1->MBB2]. I will find R4 first is killed, then is used. It can
not unlogisch. Attually R4 just is <Used>. It will cause my optimization
pass crash(Actually, I ingore Live In message of MBB. I recollect live in
messges at my pass.).
A register should not be used after it is killed, and if it is needed by a
successor block, it should be live out.

Note that a register in the live-in list of an MBB is not always live-out from
all predecessors. A register defined by IMPLICIT_DEF can be optimized away
entirely.
>  1. Does <kill> attribute of R4 at MBB0 is a unimportant  and
redundancy messages, Or a little bug???
You have probably found a bug. Can you reproduce it with one of the normal back
ends?
>  2. Is it unreliable to get a physical register's liverange by Kill,
Dead messages from MachineBasicBlock??
You also need to use the live-in list for each MBB, but otherwise it should be
reliable. Look at how RegisterScavenger is doing it.

/jakob

-------------- next part --------------
A non-text attachment was scrubbed...
Name: smime.p7s
Type: application/pkcs7-signature
Size: 1929 bytes
Desc: not available
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20100114/b0907dbf/attachment.bin>

Hi, Jakob:

  Thanks for your answer.
  I hope to trace all physical register liverange in MachineBasicBlock.
In my test, I find LiveIn message of MBB can not give all livein 
physical register. So I write a pass to recollect livein message
by scan MBB. 
  
  Current case tell me that just to scan MachineOperand's isDef, isKill,
IsDead attribute to rebuild physical register's livein will have bug.

  If I use add missing live-in into <live-in list for each MBB>, 
Could I can know which physical register is live at any time? 

  If yes, it is easy for my pass. 

  If not, I need to treat isKill and isDead as isUse, then implement a pass to
anaylze CFG to delete unvalid livein message.


--- 10年1月15日，周五, Jakob Stoklund Olesen <stoklund at 2pi.dk> 写道：
> 发件人: Jakob Stoklund Olesen <stoklund at 2pi.dk>
> 主题: Re: [LLVMdev] <IsKill> getting from MachineOperand is just
<Used> attribute from logic.
> 收件人: "任坤" <hbrenkun at yahoo.cn>
> 抄送: "llvm" <llvmdev at cs.uiuc.edu>
> 日期: 2010年1月15日,周五,上午11:44
> 
> On Jan 14, 2010, at 6:39 PM, 任坤 wrote:
> 
> > But I want do some optimization after register
> alloction by adjusting 
> > register using. I scan MachineBasicBlock to analyze
> operand's IsKill, IsDead , IsDef attribute to get a physical
> register's liverange. But I get a strange case at MBB.jpg.
> 
> You can also look at RegisterScavenging.cpp and
> MachineVerifier.cpp. They are doing the same thing.
> 
> >  R4 is marked <kill> at MBB0.  If I
> scan R4's liverange by [MBB0->MBB1->MBB2]. I will find
> R4 first is killed, then is used. It can not unlogisch.
> Attually R4 just is <Used>. It will cause my
> optimization pass crash(Actually, I ingore Live In message
> of MBB. I recollect live in messges at my pass.).
> 
> A register should not be used after it is killed, and if it
> is needed by a successor block, it should be live out.
> 
> Note that a register in the live-in list of an MBB is not
> always live-out from all predecessors. A register defined by
> IMPLICIT_DEF can be optimized away entirely.
> 
> >  1. Does <kill> attribute of R4 at MBB0 is
> a unimportant  and redundancy messages, Or a little
> bug???
> 
> You have probably found a bug. Can you reproduce it with
> one of the normal back ends?
> 
> >  2. Is it unreliable to get a physical register's
> liverange by Kill, Dead messages from MachineBasicBlock?? 
> 
> You also need to use the live-in list for each MBB, but
> otherwise it should be reliable. Look at how
> RegisterScavenger is doing it.
> 
> /jakob
> 
> 

      ___________________________________________________________ 
  好玩贺卡等你发，邮箱贺卡全新上线！ 
http://card.mail.cn.yahoo.com/