<div dir="ltr"><br><div class="gmail_extra"><br><div class="gmail_quote">On Thu, Jan 12, 2017 at 2:55 PM, Xin Tong <span dir="ltr"><<a href="mailto:trent.xin.tong@gmail.com" target="_blank">trent.xin.tong@gmail.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><span class="">On Thu, Jan 12, 2017 at 2:40 PM, Dehao Chen <<a href="mailto:danielcdh@gmail.com">danielcdh@gmail.com</a>> wrote:<br>
> Are you suggesting changing it to:<br>
><br>
> rotate_update_probability(<wbr>float orig_prob) {<br>
> if (orig_prob >= 0 && orig_prob < 0.5) {<br>
> branch_prob_inside = 0;<br>
> branch_prob_outside = orig_prob;<br>
> } else {<br>
> branch_prob_inside = 2 - 1/orig_prob;<br>
> branch_prob_outside = 1;<br>
> }<br>
> }<br>
<br>
</span>This would not work for the case which the orig_prob is bigger than<br>
0.5. branch_prob_outside tells me when the guard block is hit, it will<br>
always go to the exit.<br>
<span class="">><br>
> int estimate_trip_count(float latch_probability) {<br>
> if (latch_probability == 0)<br>
> return 0;<br>
> else<br>
> return 1/(1-latch_probability)<br>
> }<br>
><br>
> My gut feeling tells me this is a little dangerous. Micheal, what do you<br>
> think?<br>
><br>
> Dehao<br>
<br>
</span>My general feeling is.<br>
<br>
1. We do not have all the information to do a very precise<br>
distribution of the weight, and we will unlikely to be able to get<br>
such information in short term (we need value profiling at least I<br>
think or some other more clever approach). we made a decision to use<br>
average trip count for loop peeling, assuming the loop tripcounts are<br>
not polarized. Loop peeling will probably not be as effective if it<br>
is. So we make the same assumption when we rotate the loops and<br>
update the !prof metadata.<br></blockquote><div><br></div><div>I agree.</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
2. We also consider the possibility of this assumption falling out, in<br>
the case of loop peel, we basically peeled for some gains when the<br>
iteration has a high trip count and some loss (increase code size)<br>
when the iterations has low trip count. I feel this is acceptable. In<br>
case of metadata update in loop rotation, if loop does have a very<br>
<span class="">polarized distribution. i.e. cases where a few iterations takes all<br>
</span>the trip counts and rest have close to zero. We basically<br>
<span class="">underestimate the weight of the branch from guard block to exit block,<br>
as there should be more iterations of the loop that exits before even<br>
one iteration is executed. This will propagate into the CFG and result<br>
in us underestimating the average tripcount of the loops when it<br>
</span>actually executes (i.e. we have fewer iterations responsible for all<br>
<span class="">these body weight). This is OK, we become less aggressive in peeling,<br>
</span>we left some performance opportunities on the table, but only if the<br>
loop distribution is polarized. And we may do a slightly worse job in<br>
block placement.<br></blockquote><div><br></div><div>That's not exactly right. Underestimating can make us more aggressive, not less.</div><div>Because we use a hard threshold for when to peel, underestimating the trip count</div><div>can cause us to be more aggressive. I'm not sure this is a big problem, because</div><div>if the threshold is chosen correctly, choosing to peel or not to peel when we're very</div><div>close to the threshold should be a wash.</div><div>Admittedly, that's a pretty big if. :-) But I don't think it's a huge problem.</div><div><br></div><div>Dehao, why do you think this is dangerous? It's certainly somewhat ugly,</div><div>but that's a separate issue.</div><div> <br></div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
3. In case the loop distribution is close to each other, the patch<br>
handles it and distributes the weight in a reasonable way I feel.<br>
<div class="HOEnZb"><div class="h5"><br>
-Xin<br>
><br>
> On Thu, Jan 12, 2017 at 2:33 PM, Xin Tong <<a href="mailto:trent.xin.tong@gmail.com">trent.xin.tong@gmail.com</a>> wrote:<br>
>><br>
>> On Thu, Jan 12, 2017 at 2:19 PM, Dehao Chen <<a href="mailto:danielcdh@gmail.com">danielcdh@gmail.com</a>> wrote:<br>
>> ><br>
>> ><br>
>> > On Thu, Jan 12, 2017 at 2:14 PM, Xin Tong <<a href="mailto:trent.xin.tong@gmail.com">trent.xin.tong@gmail.com</a>><br>
>> > wrote:<br>
>> >><br>
>> >> On Thu, Jan 12, 2017 at 2:10 PM, Dehao Chen <<a href="mailto:danielcdh@gmail.com">danielcdh@gmail.com</a>><br>
>> >> wrote:<br>
>> >> ><br>
>> >> ><br>
>> >> > On Thu, Jan 12, 2017 at 2:02 PM, Xin Tong <<a href="mailto:trent.xin.tong@gmail.com">trent.xin.tong@gmail.com</a>><br>
>> >> > wrote:<br>
>> >> >><br>
>> >> >> On Thu, Jan 12, 2017 at 1:34 PM, Dehao Chen <<a href="mailto:danielcdh@gmail.com">danielcdh@gmail.com</a>><br>
>> >> >> wrote:<br>
>> >> >> > If we assume that the trip count will always be the same, then if<br>
>> >> >> > back<br>
>> >> >> > edge<br>
>> >> >> > taken probability falls in the range of (0, 0.5), then it already<br>
>> >> >> > broke<br>
>> >> >> > this<br>
>> >> >> > assumption. So we do not have enough info to infer both branch<br>
>> >> >> > probabilities. So probably we should simply set the inner<br>
>> >> >> > probability<br>
>> >> >> > as<br>
>> >> >> > "0". And when we estimate loop trip count, we will always return 0<br>
>> >> >> > if<br>
>> >> >> > the<br>
>> >> >> > latch probability is 0. Something like:<br>
>> >> >> ><br>
>> >> >> > rotate_update_probability(<wbr>float orig_prob) {<br>
>> >> >> > branch_prob_outside = orig_prob;<br>
>> >> >> > if (orig_prob >= 0 && orig_prob < 0.5) {<br>
>> >> >> > branch_prob_inside = 0;<br>
>> >> >> > } else {<br>
>> >> >> > branch_prob_inside = 2 - 1/orig_prob;<br>
>> >> >> > }<br>
>> >> >> > }<br>
>> >> >> ><br>
>> >> >> > int estimate_trip_count(float latch_probability) {<br>
>> >> >> > if (latch_probability == 0)<br>
>> >> >> > return 0;<br>
>> >> >> > else<br>
>> >> >> > return 1/(1-latch_probability)<br>
>> >> >> > }<br>
>> >> >><br>
>> >> >> We assume the trip count will be close to each other, they do not<br>
>> >> >> have<br>
>> >> >> to be identical =). And that assumption has implications, I<br>
>> >> >> addressed<br>
>> >> >> the implications when we have polarized loop tripcounts in an<br>
>> >> >> earlier<br>
>> >> >> response. I feel we need to make assumptions and also reason about<br>
>> >> >> the<br>
>> >> >> fallouts if that assumption turns out to be not true as Michael<br>
>> >> >> said.<br>
>> >> ><br>
>> >> ><br>
>> >> > Yes, the above code is trying to address one case where the<br>
>> >> > assumption<br>
>> >> > is<br>
>> >> > untrue.<br>
>> >> ><br>
>> >> >><br>
>> >> >><br>
>> >> >> Back to your example, I am not sure i understand it fully, is<br>
>> >> >> orig_prob the probability of branching outside before rotation ?<br>
>> >> ><br>
>> >> ><br>
>> >> > Yes, orig_prob is the original probability of the branch before loop<br>
>> >> > rotation.<br>
>> >><br>
>> >> If orig_prob == 0.5, the branch_prob_inside is 0. (branch_prob_inside<br>
>> >> is the probability that the loop body is branched to before rotation<br>
>> >> ?).<br>
>> ><br>
>> ><br>
>> > Sorry about the confusion. Let me use my original example:<br>
>> ><br>
>> > Before loop rotation<br>
>> ><br>
>> > L1:<br>
>> > if (cond) { // "orig_prob"<br>
>> > stmt;<br>
>> > goto L1;<br>
>> > }<br>
>> ><br>
>> > After loop rotation:<br>
>> ><br>
>> > if (cond) { // "branch_prob_outside"<br>
>> > L1:<br>
>> > stmt;<br>
>> > if (cond) { // "branch_prob_inside" // and also "latch_probability"<br>
>> > goto L1;<br>
>> > }<br>
>> > }<br>
>> ><br>
>> > Note that all probability are for when "cond" is true.<br>
>> ><br>
>> > Thanks,<br>
>> > Dehao<br>
>><br>
>> Thanks for the code. One thing I notice is that branch_prob_outside<br>
>> probably should not be orig_prob. orig_prob before rotation is the (#<br>
>> of times the exit is taken)/(# of times the header branch is hit<br>
>> through forward and back edges). So if the loop has a cmp i32 %index,<br>
>> 127 and recurrence of +1 (index = index + 1), i.e. the loop has a trip<br>
>> count of 128, then branch_prob_outside = orig_prob = 1/128. However,<br>
>> branch_prob_outside should really be 0 as the loop will always<br>
>> execute, i..e branch_prob_outside in the guard block will never be<br>
>> taken.<br>
>> ><br>
>> ><br>
>> >><br>
>> >> Thanks,<br>
>> >> -Xin<br>
>> >> ><br>
>> >> >><br>
>> >> >><br>
>> >> >> Thanks,<br>
>> >> >> -Xin<br>
>> >> >><br>
>> >> >> ><br>
>> >> >> ><br>
>> >> >> > On Thu, Jan 12, 2017 at 12:59 PM, Michael Kuperstein via<br>
>> >> >> > Phabricator<br>
>> >> >> > <<a href="mailto:reviews@reviews.llvm.org">reviews@reviews.llvm.org</a>> wrote:<br>
>> >> >> >><br>
>> >> >> >> mkuper added inline comments.<br>
>> >> >> >><br>
>> >> >> >><br>
>> >> >> >> ================<br>
>> >> >> >> Comment at: lib/Transforms/Scalar/<wbr>LoopRotation.cpp:476<br>
>> >> >> >> + // data.<br>
>> >> >> >> + if (!<wbr>updateLoopEstimatedBranchWeigh<wbr>t(L, GBI, BI /* OrigHeader<br>
>> >> >> >> BR<br>
>> >> >> >> */)) {<br>
>> >> >> >> + BI->setMetadata(LLVMContext::<wbr>MD_prof, nullptr);<br>
>> >> >> >> ----------------<br>
>> >> >> >> Also, regardless of the rest of the discussion - I don't think we<br>
>> >> >> >> should<br>
>> >> >> >> drop the metadata on the floor if we fail.<br>
>> >> >> >><br>
>> >> >> >> I don't think "No data is better than imprecise data" is right in<br>
>> >> >> >> the<br>
>> >> >> >> general case, but that's arguable. Specifically here, though,<br>
>> >> >> >> we're<br>
>> >> >> >> imprecise even if updateLoopEstimatedBranchWeigh<wbr>t() succeeds,<br>
>> >> >> >> because<br>
>> >> >> >> of the<br>
>> >> >> >> assumptions we make on the distribution.<br>
>> >> >> >><br>
>> >> >> >><br>
>> >> >> >> <a href="https://reviews.llvm.org/D28593" rel="noreferrer" target="_blank">https://reviews.llvm.org/<wbr>D28593</a><br>
>> >> >> >><br>
>> >> >> >><br>
>> >> >> >><br>
>> >> >> ><br>
>> >> ><br>
>> >> ><br>
>> ><br>
>> ><br>
><br>
><br>
</div></div></blockquote></div><br></div></div>