[PATCH] D10979: Update the branch weight metadata in JumpThreading pass.
Duncan P. N. Exon Smith
dexonsmith at apple.com
Thu Jul 9 14:54:41 PDT 2015
> On 2015-Jul-09, at 14:28, Cong Hou <congh at google.com> wrote:
>
> On Tue, Jul 7, 2015 at 4:53 PM, Duncan P. N. Exon Smith
> <dexonsmith at apple.com> wrote:
>>
>>> On 2015-Jul-07, at 13:24, Xinliang David Li <davidxl at google.com> wrote:
>>>
>>> On Tue, Jul 7, 2015 at 10:39 AM, Duncan P. N. Exon Smith
>>> <dexonsmith at apple.com> wrote:
>>>>
>>>>> On 2015-Jul-06, at 17:22, Cong Hou <congh at google.com> wrote:
>>>>>
>>>>>
>>>>> congh added reviewers: chandlerc, davidxl.
>>>>> congh added a subscriber: llvm-commits.
>>>>>
>>>>> Currently in JumpThreading pass, the branch weight metadata is not updated after CFG modification. Consider the jump threading on PredBB, BB, and SuccBB. After jump threading, the weight on BB->SuccBB should be adjusted as some of it is contributed by the edge PredBB->BB, which doesn't exist anymore. This patch tries to update the edge weight in metadata on BB->SuccBB by scaling it by 1 - Freq(PredBB->BB) / Freq(BB->SuccBB). Two more analyses (BlockFrequencyInfo and BranchProbabilityInfo) are needed then.
>>>>
>>>> Thanks for working on this!
>>>>
>>>> Generally you don't need these analyses to keep branch weights
>>>> up-to-date. The design premise is that you don't need global
>>>> information for local updates.
>>>>
>>>
>>> This is true in general, but probably not the case for jump-threading.
>>> When a new thread is formed (from NewBB to SuccBB), the profile
>>> update delta to the original edge BB->SuccBB comes from a different
>>> edge (Pred->NewBB which is inherited from Pred->BB). Local update by
>>> only looking at BB seems impossible.
>>
>> Ah, I see. Yes, you need a global view to know predecessor
>> probabilities.
>>
>>>
>>>> You should be able to calculate the new !prof attachment based on the
>>>> old ones, without running BFI. (I'm skeptical of even running BPI -- if
>>>> there's no !prof attachment on the old block, then you have no real
>>>> information; what's the benefit in generating a new !prof attachment
>>>> based on heuristics? The new CFG will come with its own new
>>>> heuristics.)
>>>>
>>>
>>> I believe Cong's fix is targeting PGO.
>>
>> Cong, I have a few concerns with the current approach.
>>
>> Firstly, this requires BFI even when there is no profile data (as David
>> also noted in one of the Phabricator emails). I don't think it's
>> reasonable to require all users of jump threading to run an analysis
>> that will never be used. One option would be to `addOptional<>`, but
>> I'm not sure that'll get you what you want.
>>
>> To prevent everyone paying for profile data, we should do the following:
>> if the basic block in question has branch weights, then (and only then)
>> retrieve the BFI analysis so they can be updated. I think you'll either
>> have to port the analysis retrieval and caching logic from the new
>> `PassManager` over to `LegacyPassManager`, or wait for it to be used in
>> tree.
>
> It seems at this point there isn't a good way to conditionally add a
> pass dependency to another one, unless the new PassManager is
> launched. Right?
Right. AFAICT, you'll either have to port the logic over to the
`LegacyPassManager`, or wait for (or help with!) the new one.
>> (On a related point, I doubt users of PGO + jump threading with partial
>> profiles (e.g., JITs) care enough about fidelity to justify running BFI
>> for this edge. IMO, this update should be configurable.)
>>
>> Secondly, the current code requires BFI and then immediately invalidates
>> it, but it should be straightforward to incrementally update (preserve)
>> BFI here.
>
> BFI doesn't provide any interface to incrementally update itself.
> There isn't a way to directly access and modify the frequencies stored
> in BlockFrequencyInfoImpl.
Fortunately, we have the source code for BFI ;). I think this
should be trivial to add. IIRC there's some sort of `DenseMap<>`
with the frequencies there. Patches welcome!
> I think maybe we need to update it for the
> whole function (though it is expensive)?
Wouldn't you just need to (1) modify `BB` and (2) create `NewBB`?
More information about the llvm-commits
mailing list