Hi all,
TL;DR: Should the loop guard branch be part of the canonical loop structure?
Background:
Kit,
I'm a bit unsure about having the guarded loop be canonical form. In
particular, I'm hesitant about introducing a conditional branch on
constant for loops which aren't found to have an obvious guard block.
(Note the distinction: no obvious guard does not imply no guard, just
one we didn't find.)
I'd suggest that we maybe start with a small step in this direction,
rather than everything at once?
Introducing utilities on Loop to return the loop guard block if one is
available seems like a reasonable and necessary first step. Once that's
done, we could consider extending LoopSimplify to make guarded loops
"more obvious" by moving code out of guarded blocks and the like. Note
that my "more obvious" here is about exposing existing guarded loops,
and not introducing new guards.
I think we need both of these pieces if we were to go all the way
towards your proposal right? If so, let's start incremental and see if
we have enough of a problem left to justify the canonicalization.
Philip
Hi all,
TL;DR: Should the loop guard branch be part of the canonical loop structure?
Background:
-----------
While working on the next set of improvements to loop fusion, we discovered that
loop rotation will put a guard around a loop if it cannot prove the loop will
execute at least once, as seen in the following (simplified) example:entry:
br i1 %cmp1, label %for.body.lr.ph, label %for.cond.cleanupfor.body.lr.ph: ; preds = %entry
br label %for.bodyfor.body: ; preds = %for.body.lr.ph, %for.inc
br label %for.incfor.inc: ; preds = %for.body
br i1 %cmp, label %for.body, label %for.cond.for.cond.cleanup_crit_edgefor.cond.for.cond.cleanup_crit_edge: ; preds = %for.inc
br label %for.cond.cleanupfor.cond.cleanup: ; preds = %for.cond.for.cond.cleanup_crit_edge, %entry
br label %for.endfor.end: ; preds = %for.cond.cleanup
ret voidHowever, for loops that can be proven to execute at least once, there is no
guard.This creates complications for several loop optimizations because they need to
handle both guarded loops and non-guarded loops. For example, the current loop
fusion pass needs to check whether two loops are control flow equivalent before
fusing them (i.e., if the first loop executes, the second loop is guaranteed to
execute). This is currently done by checking that the preheader of the first
loop dominates the preheader of the second loop, and the preheader of the second
loop post-dominates the preheader of the first loop. When one (or both) of the
loops have a guard, then this check no longer works. If the loop guard was part
of the canonical form, then this check could be done using the loop guard,
instead of the preheader.Similarly, when checking for perfect loop nests, a loop guard around the inner
loop will make the nest appear imperfect. However, if the guard is considered to
be part of the inner loop, it is easier to prove a perfect nest.Both of these examples have alternative ways to deal with the loop guards, if
they are present. However, I would like to have the discussion regarding making
guards part of the canonical loop structure. If people are supportive of this,
we can start doing the work to implement the guards and start modify loop
optimizations to deal with them appropriately. However, if this is not the
direction we want to go, we can deal with the guards in the specific
optimizations where they are relevant.Proposed Solution:
------------------
Ensure that all loops in canonical form have a guard branch that dominates the
preheader. If such a guard branch does not exist, a trivial guard will be
created.I have not looked at the implementation of this
extensively yet because I wanted feedback on this direction first. However, my
initial thought is to modify LoopSimplify to add the concept of a loop guard,
and provide similar guarantees that it currently provides for preheaders and exit
blocks. Specifically, if a guard block is not found, one is created and inserted
immediately before the loop preheader to guarantee a structure similar to the
one in the example above. Note that as with the preheader and exit blocks, it is
possible that subsequent passes (e.g., SimplifyCFG) may remove these blocks thus
taking it out of canonical form. However, this is an existing situation that
loop optimizations already deal with.
Hi, Kit,
Maybe I'm missing something obvious, but how does this help? I
understand how guards can make checking for perfect nesting, etc.
harder, but does always having guards make this easier?
Also, we already have a situation with LCSSA form where trivial
constructs are introduced, and then eliminated by local simplifications,
sometimes multiple times as the pipeline progresses. Having more of them
seems undesirable - it leads to more IR changes, which causes more
analysis invalidation, and so on.
Thanks again,
Hal
I just realized that inadvertently hit reply, instead of reply-all, when
replying to Philip. I'm going to summarize the conversation we have had
since, so everyone can see the direction we are thinking to go.
We'll start by adding a guard detection API as a loop utility that can
identify existing guards. The initial patch will be as simple as
possible, to establish the API. From there we can extend the detection
to get more difficult cases, and potentially modify some parts of loop
simplification (or other optimizations) to make the detection easier to
do, it it makes sense. At that point we will have experience and can
revisit this discussion if it still makes sense to do so.
Kit
Philip Reames <listmail@philipreames.com> writes:
Hi Hal,
You're right, for perfect loop nests I think simply knowing that the
inner loop has a guard will make things easier. This ties in with
Philip's earlier reply, that being able to reliably identify guarded
loops is a good first step.
Your comments about continually modifying the IR and thus invalidating
analysis makes a lot of sense. I just replied to the list with a summary
of a discussion that Philip and I had off the list (because I forgot to
use reply-all) and we're thinking to start by focusing on accurately
identifying existing guards before trying to add new (unnecessary)
guards. Please let me know if you're OK with this direction to start.
Kit
"Finkel, Hal J." <hfinkel@anl.gov> writes:
I just realized that inadvertently hit reply, instead of reply-all, when
replying to Philip. I'm going to summarize the conversation we have had
since, so everyone can see the direction we are thinking to go.We'll start by adding a guard detection API as a loop utility that can
identify existing guards. The initial patch will be as simple as
possible, to establish the API. From there we can extend the detection
to get more difficult cases, and potentially modify some parts of loop
simplification (or other optimizations) to make the detection easier to
do, it it makes sense. At that point we will have experience and can
revisit this discussion if it still makes sense to do so.Kit
Sounds good to me.
-Hal
This seems like potentially a good thing. There have been cases where
I've wanted to know the guard associated with the loop.
Loop versioning can create many guards, including guards nested within
other guards. I don't know how/if we want to capture that level of
complexity, but it was the first question that popped into my mind.
-David
Kit Barton via llvm-dev <llvm-dev@lists.llvm.org> writes:
This sounds great. Ignore the message I just sent. 
-David
Kit Barton via llvm-dev <llvm-dev@lists.llvm.org> writes:
On Hexagon, unguarded loops cannot be converted to hardware loops.
If the loop's latch branch alone handles the iteration count, including the possibility of 0, then the loop cannot be converted to a hardware loop, because hardware loops must iterate at least once. If the entire loop is guarded against zero iteration count, we can put the loop setup in the preheader, since at that point the loop is guaranteed to execute at least once.
I am strongly in favor of having some way to create loop guards, even if they are trivial.
Er, I'm missing something. Every loop header is guaranteed to execute
Philip
I don't remember the details of the particular case where we encountered this, but I think the loop started with the condition check and ended with an unconditional branch back to the beginning.
Seems like a case which should be handled by loop rotation. If you find
a reproducer for it, that's where I'd start looking rather than anything
specific to guarded loops.
Philip
That sounds like it wasn't being rotated then?
I don't know that what I'm proposing would help you in that case -
although I would be very happy if it did
If you can dig up some details and send it to me, I'd like to see we can
find a way to make it work. One of the things I don't have a good
feeling for right now is how many loops (in general) we are able to
successfully rotate and/or simplify. This will have an impact on the
overall effectiveness of loop optimizations in general, as we move
forward.
Krzysztof Parzyszek <kparzysz@quicinc.com> writes:
I don't really care that much about that particular case. We generate hardware loops somewhat late in codegen, so this is long past IR-level loop transformations. It's just a case where having guarded loops would make things easier to analyze, regardless of how the loop code may have been reorganized in the meantime.
Outlook removed llvm-dev from the recipient list, so re-replying.