[PATCH] D11442: Provide an interface normalizeSuccWeights in MachineBasicBlock to normalize its successors' weights and use it in other places.
Cong Hou via llvm-commits
llvm-commits at lists.llvm.org
Wed Aug 12 15:30:04 PDT 2015
On Wed, Aug 5, 2015 at 6:35 PM, Xinliang David Li <davidxl at google.com>
wrote:
> I'd like to summarize my proposal (interfaces) so we are on the same
> page. I have also checked the client usages of the MachineCFG update
> APIs, and found many potential problems. There are no evidences to
> show that using fixed point representation for branch probability will
> be a problem when successor edge is removed.
>
> Proposal summary:
> ===============
>
> 1. For BranchProbabilityInfo class
>
> 1) Keep getEdgeProbablity(..) interface as it is
> 2) Make getEdgeWeight(..) interfaces simply a wrapper using 1);
> deprecate its usage in the longer term
> 3) Remove setEdgeWeight interface
> 4) Add a new interface: SetEdgeProbability(..., const BranchProbability&
> prob);
>
> (Note this proposal does not cover how BranchProbabilty representation
> should be changed or not).
>
> 2. For MachineBranchProbabilityInfo class
>
> 1) Keep getEdgeProbability interface
> 2) make getEdgeWeight a wrapper to 1) and deprecate them in the future
> 3) Remove getSumForBlock() interface. No need to keep this interface
> in order to 'recalculate' branch probability on the fly.
>
> 3. MachineBasicBlock class
>
> 1) change addSuccessor(.., uint32_t weight) to addSuccessor(...,
> BranchProbability prob)
> 2) Remove setSuccWeight(..)
> 3) Add a new interface: setEdgeProbabiltity(..., const BranchProbability&)
>
>
> How Machine CFG update interfaces are used and potential issues with
> profile update
> =================================================================
>
> The interfaces include:
> MachineBasicBlock::addSuccessor(...)
> MachineBasicBlock::removeSuccessor(..)
> MachineBasicBlock::replaceSuccssor(..)
> MachineBasicBlock::transferSuccessor(..)
> MachineBasicBlock::transferSuccessorAndUpdatePHIs(..)
>
>
> 1. MachineBasicBlock::addSuccessor
>
> The interface allows client to omit a weight by defining the default
> weight to be 0. I think this is a mistake. When control flow is
> synthensized during target lowering, the client has better knowledge
> of the branch bias. The current design allows client simply do:
>
> allocMBB->addSuccessor(&PrologueMBB);
>
> checkMBB->addSuccessor(allocMBB);
> checkMBB->addSuccessor(&PrologueMBB);
>
> There are many such cases.
>
> Another big problem is that the CFG update code freely transfer
> weight from one parent BB to another BB which can be totally wrong if
> their original weight scales are out of sync:
>
> if (FuncInfo.BPI)
> BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
>
> TrueMBB->getBasicBlock());
> FuncInfo.MBB->addSuccessor(TrueMBB, BranchWeight);
>
>
> 2. MachineBasicBlock::removeSuccessor
>
When we remove a successor, should we adjust the probabilities of other
successors? If yes, what if we later add another successor with a given
probability? If not, the sum of all successor's probabilities will not be 1
and probabilities are more like weights (and probabilities don't make sense
anymore).
The similar question is on addSuccessor: what if we add a successor when
the sum of other successors' probabilities is already 1? Do we always need
to check this?
Here are some candidate designs:
1. addSuccessor(MachineBasicBlock*, BranchProbability);
a. Just assign the given probability to the internal one, without doing
any check or normalization.
Pros:
- Fast as there is no additional overhead.
- More natural when the user uses this interface to add several
successors with precalculated probabilities.
Cons:
- The user need to make sure the sum of all successors' probabilities
is one.
- Fails to adjust probabilities if the user adds a successor while the
sum of other successors' probabilities is one.
b. Always normalize all successors' probabilities so that the sum of them
is one. The given probability of the new successor should be guaranteed,
and other successors' probabilities are scaled down.
Pros:
- It is guaranteed that the sum of all successors' probabilities is one.
Cons:
- Has overhead when adding successors, though the overhead is moderate.
c. Provide a new interface addSuccessors() to add several successors at
one time with precalculated probabilities. Let addSuccessor() behave as in
b.
Pros:
- Faster than b when adding more than one successor.
Cons:
- The use has to make sure that the sum of all successors'
probabilities is one when using this new interface.
d. Add a new parameter to addSuccessor() to indicate whether to
normalize probabilities.
Pros:
- If used properly, it is faster than b in some cases.
Cons:
- Add the complexity of this interface, and make it harder to be used
by users.
2. removeSuccessor()
This will have similar designs as a/b/d of addSuccessor() shown above.
Any suggestion?
thanks,
Cong
>
> 1) A common pattern of using removeSuccessor is to insert a trampoline
> BB with conditional jump to the old target.
>
> JTBB->removeSuccessor(BB);
> JTBB->addSuccessor(NewBB);
>
> After this, the outgoing edge probabilities of JTBB will be wrong. If
> we enforce the interface to take branch probability, then problems
> like this won't occur.
>
> 2) removing edge to landing pad -- in this case, it does not matter
> whether the edge probability is stored as weight over sum or as fixed
> point representation of the real probability.
>
> 3) In IfConvert, remove successor edge to TBB. The branch is gone
> after the transformation, so there is no profile update issue
>
> 4) Superfluous edge removal in MachineBasicBlock::CorrectExtraCFGEdges
> should assert that superfluous edges have zero probability.
>
> 5) Dead Block Removal in various passes such as BranchFolding,
> UnreachableBlockElim, removeSuccessor is invoked on the dead BB, so
> there is no update issue there.
>
> 6) Block Merging in BranchFolding -- after merging, the branch is
> gone, so there is no profile update issue
>
> 7) TailDuplication: current profile update is correct. Using
> probability based interface will work fine.
>
> uint32_t Weight = MBPI->getEdgeWeight(PredBB, TailBB);
> PredBB->removeSuccessor(TailBB);
> unsigned NumSuccessors = PredBB->succ_size();
> assert(NumSuccessors <= 1);
> if (NumSuccessors == 0 || *PredBB->succ_begin() != NewTarget)
> PredBB->addSuccessor(NewTarget, Weight);
>
>
> 3. MachineBasicBlock::transferSuccessors (MachineBasicBlock *FromBB)
>
> The implementation uses removeSuccessor and addSuccessor and tries to
> update weight. The implementation in general sense is wrong, as the
> succ edge weights fetched from FromBB may have a different scale than
> weights owned by 'this' BB. The profile update is guaranteed to be
> right only when 'this' BB's successor list is empty. I have checked
> the sites that invoke this interface -- this condition (empty
> successor list) does seem to be always satisfied.
>
> In conclusion, it is safe to get rid of getSumForBlock. The potential
> bugs of missing profile/profile update in CodeGen also need to be
> further examined and fixed.
>
> thanks,
>
> David
>
>
>
> On Wed, Aug 5, 2015 at 1:59 PM, Duncan P. N. Exon Smith
> <dexonsmith at apple.com> wrote:
> >
> >> On 2015-Aug-05, at 13:48, Cong Hou <congh at google.com> wrote:
> >>
> >> On Wed, Aug 5, 2015 at 12:24 PM, Duncan P. N. Exon Smith
> >> <dexonsmith at apple.com> wrote:
> >>>
> >>> In terms of changing branch weights to use fixed point arithmetic (and
> >>> always be normalized): I like this idea. I'd prefer a number like
> >>> UINT32_MAX instead of 100000, but I agree with David that 32 bits ought
> >>> to be enough. Something similar to the `BlockMass` design seems about
> >>> right.
> >>>
> >>> The major downside is that passes that modify the CFG will always have
> >>> to update edge weights to renormalize them. Currently, if a pass
> >>> removes a successor edge, it needn't worry about normalization at all.
> >>> Although that's convenient, it's probably not important. (Or maybe
> >>> they could be normalized lazily...?)
> >>
> >> One important idea from David's design is that we should pass branch
> >> probability instead of edge weight when adding new edges or updating
> >> edge weights. If we decide to use this design, then when edge weights
> >> (or probabilities) are updated, we should always update the weights on
> >> other out-edges from the same block. Removing edge doesn't need
> >> normalization but if we want a fast getEdgeProbability()
> >> implementation without storing the sum of edge weights, we should
> >> always keep edge weights normalized.
> >
> > Wouldn't caching in BPI be fast enough for `getEdgeProbability()`?
> > And it slows down update speed (since, as you point out, every update
> > to one edge requires updates to all the others).
> >
> >>> If you go ahead with this, then you should first change
> >>> `BranchProbability` to use this design, and then you can reuse that
> >>> class for branch weights.
> >>
> >> OK. I think we can allow BranchProbability to be built either from
> >> uint32_t or BlockMass-like class.
> >
> > I was suggesting: change `BranchProbability` to be a BlockMass-like
> > class. (Or are we on the same page already?)
> >
> >>> BTW, another approach would be to change `getSumForBlock()` to return a
> >>> 64-bit number, and just use 64-bit math everywhere (weights would still
> >>> be 32-bit). Sounds like a lot less work. Thoughts on that?
> >>
> >> As BranchProbability only accepts uint32_t parameters, we still need
> >> to scale down the sum of 64-bit type when constructing probabilities.
> >
> > Sure, but BPI could easily cache that result. It caches the weight
> > right now, and could just as easily cache the probability instead (or
> > as well).
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150812/b1462295/attachment.html>
More information about the llvm-commits
mailing list