Hi, I have worked a little on the PBQP register allocator, and it is quite clear (at least to me) that it is not even a serious alternative to RegAllocGreedy at the moment, due to the poor handling of spilling. As Arnaud wrote below, it is not optimizing spilling at all, but rather just spills anything that does not get an assignment. The result is a lot more spill/reload instructions than needed. In RegAllocBase.h it says "...Register allocation complexity, and generated code performance is determined by the effectiveness of live range splitting rather than optimal coloring...". I would then think that any register allocation algorithm should benefit from this, but find that only RegAllocGreedy is doing live range splitting, and that the code for doing this is local to that allocator. I would like to suggest a refactoring to make RAGreedy::trySplit() and its sub functions callable from any register allocator. Perhaps part of SplitEditor? What do you think about this? /Jonas From: Arnaud A. de Grandmaison [mailto:arnaud.degrandmaison at arm.com] Sent: den 4 mars 2015 15:43 To: Jonas Paulsson; Lang Hames Cc: llvmdev at cs.uiuc.edu Subject: RE: PBQP spilling Yes, for now the spilling is done in the most basic way, i.e. it's functionally correct --- but not efficient. The focus was on the allocator itself, not on the spilling. As you noticed, the work still to be done in this area is live range splitting, and smarter spill code insertion. Another area is improving the reduction order, to make the allocator less sensitive to the reduction order. There is no official plan; we started to discuss that with Lang some time ago, but none of us had time to dive into it yet. Any help appreciated there :). Cheers, Arnaud From: Jonas Paulsson [mailto:jonas.paulsson at ericsson.com] Sent: 04 March 2015 13:51 To: Lang Hames; Arnaud De Grandmaison Cc: llvmdev at cs.uiuc.edu<mailto:llvmdev at cs.uiuc.edu> Subject: PBQP spilling Hi, I would like to ask about PBQPs use of InlineSpiller. The code output when using PBQP gets a lot bigger compared to when using RegAllocGreedy. PBQP does not split the live intervals, and a lot more (often redundant) reload instructions are emitted as a result, it seems. I wonder why this is, and if there are any plans to improve on this point? /Jonas Paulsson -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150306/102169de/attachment.html>
Hi Jonas,> On Mar 6, 2015, at 12:31 AM, Jonas Paulsson <jonas.paulsson at ericsson.com> wrote: > > Hi, > > I have worked a little on the PBQP register allocator, and it is quite clear (at least to me) that it is not even a serious alternative to RegAllocGreedy at the moment, due to the poor handling of spilling. As Arnaud wrote below, it is not optimizing spilling at all, but rather just spills anything that does not get an assignment. The result is a lot more spill/reload instructions than needed. > > In RegAllocBase.h it says “…Register allocation complexity, and generated code performance is determined by the effectiveness of live range splitting rather than optimal coloring…”. I would then think that any register allocation algorithm should benefit from this, but find that only RegAllocGreedy is doing live range splitting, and that the code for doing this is local to that allocator. > > I would like to suggest a refactoring to make RAGreedy::trySplit() and its sub functions callable from any register allocator. Perhaps part of SplitEditor?What do you expect from sure refactoring? In the current form, live-range splitting for the PBQP implies to rebuild part of the graph and I suspect it would be easier to rebuild it from scratch than trying to update it if we would want to use it. I believe that to have an efficient implementation, the PBQP splitting should work directly on the graph and not on the program as it is the case for the Greedy Allocator. Now, regarding the fast register allocator, I am not sure it could cope at all with live-range splitting as it would change some allocation decisions that were supposed to be final. The bottom line is I do not believe there is much to share here. Cheers, -Quentin> > What do you think about this? > > /Jonas > > > From: Arnaud A. de Grandmaison [mailto:arnaud.degrandmaison at arm.com] > Sent: den 4 mars 2015 15:43 > To: Jonas Paulsson; Lang Hames > Cc: llvmdev at cs.uiuc.edu > Subject: RE: PBQP spilling > > Yes, for now the spilling is done in the most basic way, i.e. it’s functionally correct --- but not efficient. The focus was on the allocator itself, not on the spilling. As you noticed, the work still to be done in this area is live range splitting, and smarter spill code insertion. Another area is improving the reduction order, to make the allocator less sensitive to the reduction order. > > There is no official plan; we started to discuss that with Lang some time ago, but none of us had time to dive into it yet. Any help appreciated there J. > > Cheers, > Arnaud > > From: Jonas Paulsson [mailto:jonas.paulsson at ericsson.com <mailto:jonas.paulsson at ericsson.com>] > Sent: 04 March 2015 13:51 > To: Lang Hames; Arnaud De Grandmaison > Cc: llvmdev at cs.uiuc.edu <mailto:llvmdev at cs.uiuc.edu> > Subject: PBQP spilling > > Hi, > > I would like to ask about PBQPs use of InlineSpiller. The code output when using PBQP gets a lot bigger compared to when using RegAllocGreedy. PBQP does not split the live intervals, and a lot more (often redundant) reload instructions are emitted as a result, it seems. I wonder why this is, and if there are any plans to improve on this point? > > /Jonas Paulsson-------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150306/4cd89035/attachment.html>
Hi Quentin, Jonas, Splitting fits in with PBQP reasonably well, at least conceptually. The PBQP graph is designed to be mutable, so there is no problem with updating it when splitting. As I see it, there are two logical places to integrate splitting into PBQP: 1) Split during spilling -- If a PBQP solution selects the spill option for a node, rather than spill immediately, split the interval first. This seems (superficially) like an opportunity to use trySplit, but I haven't looked at the code for it. If trySplit is looking at the existing allocation state to make those decisions we may have trouble: In RAGreedy some "current" colorings may be revisited, but during spilling in PBQP we're guaranteed that *all* colorings will be revisited in the next round, which may throw trySplit's heuristics out. 2) Pre-split -- We could have a pre-pass split live ranges at sensible points (e.g. around loops) before we enter the coloring/spilling loop. PBQP tries to coalesce as it colors, so it will try to undo these splits where possible. I assume trySplit would not be a good fit here, but we may be able to reuse other parts of the splitting code. We do want to add splitting support to PBQP, so we should be on the lookout for opportunities to share code where it makes sense. Cheers, Lang. On Sat, Mar 7, 2015 at 5:01 AM, Quentin Colombet <qcolombet at apple.com> wrote:> Hi Jonas, > > On Mar 6, 2015, at 12:31 AM, Jonas Paulsson <jonas.paulsson at ericsson.com> > wrote: > > Hi, > > I have worked a little on the PBQP register allocator, and it is quite > clear (at least to me) that it is not even a serious alternative to > RegAllocGreedy at the moment, due to the poor handling of spilling. As > Arnaud wrote below, it is not optimizing spilling at all, but rather just > spills anything that does not get an assignment. The result is a lot more > spill/reload instructions than needed. > > In RegAllocBase.h it says “…Register allocation complexity, and generated > code performance is determined by the effectiveness of live range splitting > rather than optimal coloring…”. I would then think that any register > allocation algorithm should benefit from this, but find that only > RegAllocGreedy is doing live range splitting, and that the code for doing > this is local to that allocator. > > I would like to suggest a refactoring to make RAGreedy::trySplit() and its > sub functions callable from any register allocator. Perhaps part of > SplitEditor? > > > What do you expect from sure refactoring? > > In the current form, live-range splitting for the PBQP implies to rebuild > part of the graph and I suspect it would be easier to rebuild it from > scratch than trying to update it if we would want to use it. > I believe that to have an efficient implementation, the PBQP splitting > should work directly on the graph and not on the program as it is the case > for the Greedy Allocator. > > Now, regarding the fast register allocator, I am not sure it could cope at > all with live-range splitting as it would change some allocation decisions > that were supposed to be final. > > The bottom line is I do not believe there is much to share here. > > Cheers, > -Quentin > > > What do you think about this? > > /Jonas > > > *From:* Arnaud A. de Grandmaison [mailto:arnaud.degrandmaison at arm.com > <arnaud.degrandmaison at arm.com>] > *Sent:* den 4 mars 2015 15:43 > *To:* Jonas Paulsson; Lang Hames > *Cc:* llvmdev at cs.uiuc.edu > *Subject:* RE: PBQP spilling > > Yes, for now the spilling is done in the most basic way, i.e. it’s > functionally correct --- but not efficient. The focus was on the allocator > itself, not on the spilling. As you noticed, the work still to be done in > this area is live range splitting, and smarter spill code insertion. > Another area is improving the reduction order, to make the allocator less > sensitive to the reduction order. > > There is no official plan; we started to discuss that with Lang some time > ago, but none of us had time to dive into it yet. Any help appreciated there > J. > > Cheers, > Arnaud > > *From:* Jonas Paulsson [mailto:jonas.paulsson at ericsson.com > <jonas.paulsson at ericsson.com>] > *Sent:* 04 March 2015 13:51 > *To:* Lang Hames; Arnaud De Grandmaison > *Cc:* llvmdev at cs.uiuc.edu > *Subject:* PBQP spilling > > Hi, > > I would like to ask about PBQPs use of InlineSpiller. The code output when > using PBQP gets a lot bigger compared to when using RegAllocGreedy. PBQP > does not split the live intervals, and a lot more (often redundant) reload > instructions are emitted as a result, it seems. I wonder why this is, and > if there are any plans to improve on this point? > > /Jonas Paulsson > > >-------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150310/7f5f3dd7/attachment.html>