llvm dev - Jan 2014 - [LLVMdev] How to update LiveInterval information of newly inserted machine basic block

At 2014-01-01 04:36:21,"Andrew Trick" <atrick at apple.com>
wrote:


On Dec 31, 2013, at 3:52 AM, Haishan <hndxvon at 163.com> wrote:


Hi,
  I insert a new machine basic block(MBB) before Greedy Register Allocation,
after Simple Register Coalescing. But I encounter a fatal
error "regalloc = ... not currently supported with -O0". I use command
line with opt level O2, not O0.
The probable reason of this error is that no LiveInterval information for newly
MBB which is used by Register Allocation.
And, LiveIntervals depend on LiveVariables which require the machine function to
be in SSA form.
However, in this case, the modified machine function(including newly inserted
MBB) is not SSA form any more.
So, how to update LiveInterval information of newly MBB?
Could someone help me with that?
Thank you very much in advance.
-Haishan


Yes, I think error that means LiveIntervals have been invalidated. It is
supposed to be user-friendly but should probably be changed to be more specific.


If you schedule a new pass between coalescing and regalloc, it needs to preserve
LiveIntervals analysis. There are utilities to recompute live intervals for
individual virtual registers. See LiveIntervals::computeVirtRegs. LiveVariables
is not actually needed to update live intervals. It is only added as a
requirement to make sure it runs before the 2-address pass.


-Andy


  Thank you for your reply.
  I do it as you tell me, and you are right.
  My update steps are shown following:
  LiveIntervals *LIS = &getAnalysis<LiveIntervals>();
    1. LIS->insertMBBInMaps(NewMBB)
    2. LIS->InsertMachineInstrRangeInMaps(NewMBB->begin(),
NewMBB->end())
    3. Due to some machine instructions of NewMBB copy from OldMBB, I have to
removeInterval firstly. Then I use LIS->createAndComputeVirtRegInterval() to
update LiveInterval info of NewMBB.
    4. Then I want to delete OldMBB in the current MachineFuncion. At the same
time, delete LiveIntervals info of OldMBB. So, firstly, Use
RemoveMachineInstrFromMaps() to erase SlotIndexes info of OldMBB. And, delete
OldMBB from current CFG. Lastly, remove all registers' interval, create and
recompute virtual register interval info.
  However, I get an unreachable error "Use not jointly dominated by
defs" in LiveRangeCalc.cpp:188.
So, comparing with Step3, Step4 generates an unreachable error, Why?
  Valid information of OldMBB is deleted when updating CFG? Is there something
wrong with my updating steps?
  Thank you very much again.
  -Haishan
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20140103/8f91a78c/attachment.html>

On Jan 3, 2014, at 4:58 AM, Haishan <hndxvon at 163.com> wrote:
> 
> At 2014-01-01 04:36:21,"Andrew Trick" <atrick at apple.com>
wrote:
> 
> On Dec 31, 2013, at 3:52 AM, Haishan <hndxvon at 163.com> wrote:
> 
>> Hi,
>>   I insert a new machine basic block(MBB) before Greedy Register
Allocation, after Simple Register Coalescing. But I encounter a fatal
>> error "regalloc = ... not currently supported with -O0". I
use command line with opt level O2, not O0.
>> The probable reason of this error is that no LiveInterval information
for newly MBB which is used by Register Allocation.
>> And, LiveIntervals depend on LiveVariables which require the machine
function to be in SSA form.
>> However, in this case, the modified machine function(including newly
inserted MBB) is not SSA form any more.
>> So, how to update LiveInterval information of newly MBB?
>> Could someone help me with that?
>> Thank you very much in advance.
>> -Haishan
> 
> Yes, I think error that means LiveIntervals have been invalidated. It is
supposed to be user-friendly but should probably be changed to be more specific.
> 
> If you schedule a new pass between coalescing and regalloc, it needs to
preserve LiveIntervals analysis. There are utilities to recompute live intervals
for individual virtual registers. See LiveIntervals::computeVirtRegs.
LiveVariables is not actually needed to update live intervals. It is only added
as a requirement to make sure it runs before the 2-address pass.
> 
> -Andy
> 
>   Thank you for your reply.
>   I do it as you tell me, and you are right.
>   My update steps are shown following:
>   LiveIntervals *LIS = &getAnalysis<LiveIntervals>();
>     1. LIS->insertMBBInMaps(NewMBB)
>     2. LIS->InsertMachineInstrRangeInMaps(NewMBB->begin(),
NewMBB->end())
>     3. Due to some machine instructions of NewMBB copy from OldMBB, I have
to removeInterval firstly. Then I use LIS->createAndComputeVirtRegInterval()
to update LiveInterval info of NewMBB.
>     4. Then I want to delete OldMBB in the current MachineFuncion. At the
same time, delete LiveIntervals info of OldMBB. So, firstly, Use
RemoveMachineInstrFromMaps() to erase SlotIndexes info of OldMBB. And, delete
OldMBB from current CFG. Lastly, remove all registers' interval, create and
recompute virtual register interval info.
>   However, I get an unreachable error "Use not jointly dominated by
defs" in LiveRangeCalc.cpp:188.
> So, comparing with Step3, Step4 generates an unreachable error, Why?
>   Valid information of OldMBB is deleted when updating CFG? Is there
something wrong with my updating steps?
>   Thank you very much again.
I’m haven’t done anything like this so I can only guess.

The error means that some path from function entry does not provide a reaching
def for a vreg. You’ll have to analyze the CFG and debug to understand why.

My suggestion would be to completely update the CFG first, with no connections
to the old block. Remove the old block, it’s instructions and vregs from maps.
Then add the new block to maps and recompute.

You might need to keep track of any vregs referenced in the old block but not in
the new block to ensure they are updated.

Hope it works.

-Andy
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20140103/2652b285/attachment.html>

On Jan 3, 2014, at 11:01 AM, Andrew Trick <atrick at apple.com> wrote:
> 
> On Jan 3, 2014, at 4:58 AM, Haishan <hndxvon at 163.com> wrote:
> 
>> 
>> At 2014-01-01 04:36:21,"Andrew Trick" <atrick at
apple.com> wrote:
>> 
>> On Dec 31, 2013, at 3:52 AM, Haishan <hndxvon at 163.com> wrote:
>>   My update steps are shown following:
>>   LiveIntervals *LIS = &getAnalysis<LiveIntervals>();
>>     1. LIS->insertMBBInMaps(NewMBB)
>>     2. LIS->InsertMachineInstrRangeInMaps(NewMBB->begin(),
NewMBB->end())
>>     3. Due to some machine instructions of NewMBB copy from OldMBB, I
have to removeInterval firstly. Then I use
LIS->createAndComputeVirtRegInterval() to update LiveInterval info of NewMBB.
>>     4. Then I want to delete OldMBB in the current MachineFuncion. At
the same time, delete LiveIntervals info of OldMBB. So, firstly, Use
RemoveMachineInstrFromMaps() to erase SlotIndexes info of OldMBB. And, delete
OldMBB from current CFG. Lastly, remove all registers' interval, create and
recompute virtual register interval info.
>>   However, I get an unreachable error "Use not jointly dominated
by defs" in LiveRangeCalc.cpp:188.
>> So, comparing with Step3, Step4 generates an unreachable error, Why?
>>   Valid information of OldMBB is deleted when updating CFG? Is there
something wrong with my updating steps?
>>   Thank you very much again.
> 
> I’m haven’t done anything like this so I can only guess.
> 
> The error means that some path from function entry does not provide a
reaching def for a vreg. You’ll have to analyze the CFG and debug to understand
why.
> 
> My suggestion would be to completely update the CFG first, with no
connections to the old block. Remove the old block, it’s instructions and vregs
from maps. Then add the new block to maps and recompute.
> 
> You might need to keep track of any vregs referenced in the old block but
not in the new block to ensure they are updated.
Those LIS->insert…() functions are only proxies for the corresponding
functions on SlotIndexes. They only add the necessary slot indexes, they don’t
update any liveness information.

We don’t have any support for updating liveness after inserting a new basic
block. There is no reason you couldn’t do it, but the code doesn’t exist.

Thanks,
/jakob
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20140103/216b094c/attachment.html>