[LLVMdev] PBQP crash

[Arnaud De Grandmaison]

Thanks for the explanations !
--
Arnaud

From: Lang Hames [mailto:lhames at gmail.com]
Sent: 30 January 2015 22:28
To: Arnaud De Grandmaison
Cc: Jonas Paulsson; LLVM Developers Mailing List
Subject: Re: [LLVMdev] PBQP crash

Hi Arnaud,

The conservatively allocatable test is supposed to check two conditions, either of which would be sufficient to make a node allocatable:

(1) There exists some register that is not aliased by any register option for any neighbor. This is the "safe row" test. It is straightforward, but likely to fire only rarely.

(2) The sum of the maximum number of registers aliased by any register for each neighbor is less than the number of available registers for this node. This is the "worst-column" test. More intuitively: for each neighbor you compute the maximum number of registers that could be "taken" from this node by an allocation to that neighbor. If you assume (conservatively) that none of these "taken" registers alias one another then you can sum them to find the maximum number of registers that might be unavailable to this node. If this sum is lower than the number of registers available then you're safely allocatable.

Cheers,
Lang.


On Mon, Jan 26, 2015 at 11:42 PM, Arnaud A. de Grandmaison <arnaud.degrandmaison at arm.com<mailto:arnaud.degrandmaison at arm.com>> wrote:
> A node should never be put into the conservatively allocatable list if there is a chance of it spilling.

I can understand why the logic of NodeMetadata::isConservativelyAllocatable is necessary for the node to be allocatable, but I have not been able to convince myself this is sufficient, especially when the node degree > available registers.

Cheers,
Arnaud

From: llvmdev-bounces at cs.uiuc.edu<mailto:llvmdev-bounces at cs.uiuc.edu> [mailto:llvmdev-bounces at cs.uiuc.edu<mailto:llvmdev-bounces at cs.uiuc.edu>] On Behalf Of Lang Hames
Sent: 27 January 2015 06:06
To: Jonas Paulsson
Cc: llvmdev at cs.uiuc.edu<mailto:llvmdev at cs.uiuc.edu>
Subject: Re: [LLVMdev] PBQP crash

Hi Jonas,

> * The problematic node that was spilled again, was in the ConservativelyAllocatableNodes set during reduce(). The comment in reduce() “Conservatively allocatable nodes will never spill…” indicates that perhaps this is an incorrect insertion, as the regs did in fact run out in this case.

Arnaud is correct: A node should never be put into the conservatively allocatable list if there is a chance of it spilling. Off the top of my head I can imagine 2 things going wrong here:

(1) Conservative allocability is mostly-precomputed, and updated with deltas as nodes are removed. It is possible that there is some subtle bug in this code.

(2) I think the current conservative allocability test bakes in the assumption that all register options have non-infinite cost. If you assign infinite costs to any physical register I would expect this to blow up.

Are you able to share a test case at all? If so that would be great. If not, I can add an option to the allocator to dump abstract PBQP graphs, and I could use these to test the problem on my end.

Regards,
Lang.


On Mon, Jan 26, 2015 at 7:55 AM, Jonas Paulsson <jonas.paulsson at ericsson.com<mailto:jonas.paulsson at ericsson.com>> wrote:
Hi,

I have run into a test case on an out-of-tree target where PBQP fails to complete register allocation after “Attempting to spill already spilled value” (the triggered assert in InlineSpiller::spill().

First, the original LiveInterval is spilled. It is a load of a symbol into a narrow register class, i.e. a subset of the class of address registers. InlineSpiller decides to rematerialize the load of the symbol to lie right before its only user, which makes good sense. The original def is removed.

The new LiveInterval pushed is thus much smaller in the next PBQP round. The spill cost is marked as ‘inf’ during graph building. This small interval has also a lot of overlapping intervals and thus edges in the PBQP graph. It gets pushed on the node stack to later be popped after 17 others.
Those 17 nodes use up all registers of the narrow reg-class, and the cost vector has become all infinities. Spill option is selected again, and thus the error is a fact of spilling an already spilled value.

I wonder what has gone wrong here, and have some initial thoughts:

* The problematic node that was spilled again, was in the ConservativelyAllocatableNodes set during reduce(). The comment in reduce() “Conservatively allocatable nodes will never spill…” indicates that perhaps this is an incorrect insertion, as the regs did in fact run out in this case.
   In setup(), the node is first put into not-provably-allocatables. However, one of it’s neigbhour invoked handleDisconnectEdge(), and moves it into conservatively-allocatables, because DeniedOpts had become less than NumOpts (in isConservativelyAllocatable().
* There are lots of spillable nodes being popped before the one that can’t be spilled.  This seems intuitively wrong, as they are intervals that actually could be spilled.

I would really appreciate some help and pointers on what might be going wrong here,

Jonas Paulsson




-- IMPORTANT NOTICE: The contents of this email and any attachments are confidential and may also be privileged. If you are not the intended recipient, please notify the sender immediately and do not disclose the contents to any other person, use it for any purpose, or store or copy the information in any medium. Thank you.

ARM Limited, Registered office 110 Fulbourn Road, Cambridge CB1 9NJ, Registered in England & Wales, Company No: 2557590
ARM Holdings plc, Registered office 110 Fulbourn Road, Cambridge CB1 9NJ, Registered in England & Wales, Company No: 2548782
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150130/caf5130e/attachment.html>