[PATCH] D19950: Use frequency info to guide Loop Invariant Code Motion.
Hal Finkel via llvm-commits
llvm-commits at lists.llvm.org
Thu May 12 18:56:16 PDT 2016
----- Original Message -----
> From: "Xinliang David Li" <davidxl at google.com>
> To: "Dehao Chen" <danielcdh at gmail.com>
> Cc: "Hal Finkel" <hfinkel at anl.gov>,
> reviews+D19950+public+38ba22078c2035b8 at reviews.llvm.org, "David
> Majnemer" <david.majnemer at gmail.com>, "Junbum Lim"
> <junbuml at codeaurora.org>, mcrosier at codeaurora.org, "llvm-commits"
> <llvm-commits at lists.llvm.org>, "amara emerson"
> <amara.emerson at arm.com>
> Sent: Wednesday, May 11, 2016 6:01:32 PM
> Subject: Re: [PATCH] D19950: Use frequency info to guide Loop
> Invariant Code Motion.
> This is probably just a concern in theory -- current store motion
> only does downward code motion (sink and merge).
I'm not sure I understand the example. To host the cold_load, it must not alias with any stores in the loop. To hoist the store past the loop, it also must not alias with anything in the loop.
> While upward code motion for stores is also possible (e.g. to shrink
> live ranges of the stored value and address val), it is not likely
> done as an IR optimization pass.
This reminds me of a conversation I was having with Philip some weeks ago about how InstCombine has a somewhat-unfortunate heuristic of sinking instructions with a single use in only one predecessor into that predecessor. It does this even for instructions with potential side effects, and so we lose the ability to hoist them back again. Hosting them back might be important later for scheduling, etc. For what it wants to do, however, the heurisitic is also not strong enough because it a) does not handle multiple uses in the predecessor and b) only looks in direct predecessors, not any dominated block. Sort of the worst of both worlds ;)
-Hal
> David
> On Wed, May 11, 2016 at 3:44 PM, Dehao Chen < danielcdh at gmail.com >
> wrote:
> > hoist-early sink-later may also introduces hoisted instructions
> > that
> > is not sinkable later.
>
> > e.g.
>
> > orig code:
>
> > for() {
>
> > if (cond) {
>
> > cold_load;
>
> > cold_code;
>
> > }
>
> > }
>
> > store;
>
> > after hoisting:
>
> > cold_load;
>
> > for() {
>
> > if (cond) {
>
> > cold_code;
>
> > }
>
> > }
>
> > store;
>
> > after other code motion:
>
> > cold_load;
>
> > store;
>
> > for() {
>
> > if (cond) {
>
> > cold_code;
>
> > }
>
> > }
>
> > then later in cgp, when you want to sink cold_load to its uses, the
> > store may prevent the sinking due to aliasing.
>
> > On Wed, May 11, 2016 at 10:58 AM, Xinliang David Li <
> > davidxl at google.com > wrote:
>
> > > On Tue, May 10, 2016 at 3:14 PM, Hal Finkel < hfinkel at anl.gov >
> > > wrote:
> >
>
> > > > ----- Original Message -----
> > >
> >
>
> > > > > From: "Dehao Chen" < danielcdh at gmail.com >
> > >
> >
>
> > > > > To: "Hal Finkel" < hfinkel at anl.gov >
> > >
> >
>
> > > > > Cc: "Xinliang David Li" < davidxl at google.com >,
> > > > > reviews+D19950+public+38ba22078c2035b8 at reviews.llvm.org ,
> > > > > "David
> > >
> >
>
> > > > > Majnemer" < david.majnemer at gmail.com >, "Junbum Lim" <
> > > > > junbuml at codeaurora.org >, mcrosier at codeaurora.org ,
> > > > > "llvm-commits"
> > >
> >
>
> > > > > < llvm-commits at lists.llvm.org >, "amara emerson" <
> > > > > amara.emerson at arm.com >
> > >
> >
>
> > > > > Sent: Tuesday, May 10, 2016 4:10:49 PM
> > >
> >
>
> > > > > Subject: Re: [PATCH] D19950: Use frequency info to guide Loop
> > > > > Invariant Code Motion.
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > > Thanks for the comment. I spent quite a while to think, but
> > > > > still
> > >
> >
>
> > > > > cannot think of an optimization that could be unblocked by
> > >
> >
>
> > > > > speculatively hoisting an loop invariant from an unlikely
> > > > > executed
> > >
> >
>
> > > > > path. Can you give some hint (or an example) on what type of
> > >
> >
>
> > > > > optimization can benefit from this case?
> > >
> >
>
> > > > I'm specifically thinking about this case (although I suspect
> > > > there
> > > > are others):
> > >
> >
>
> > > > for (...) {
> > >
> >
>
> > > > if (...) {
> > >
> >
>
> > > > hoistable
> > >
> >
>
> > > > cold_stuff
> > >
> >
>
> > > > }
> > >
> >
>
> > > > }
> > >
> >
>
> > > > for (...) {
> > >
> >
>
> > > > if (...) {
> > >
> >
>
> > > > hoistable
> > >
> >
>
> > > > hot_stuff
> > >
> >
>
> > > > }
> > >
> >
>
> > > > }
> > >
> >
>
> > > > I expect that 'hoistable' will be hoisted by LICM out of both
> > > > loops,
> > > > and then CSE'd by GVN.
> > >
> >
>
> > > I think this case can be/should be handled by more general
> > > profile
> > > driver speculative PRE. The above case may not be profitable even
> > > after GVN CSEed two expressions. On the other hand,
> >
>
> > > ... = a * b;
> >
>
> > > for (...) {
> >
>
> > > if (cold) {
> >
>
> > > .... = a * b;
> >
>
> > > }
> >
>
> > > }
> >
>
> > > It will be good to hoist and CSE. Though in this case, we do not
> > > need
> > > LICM to enable this CSE. Another case:
> >
>
> > > if (....) {
> >
>
> > > ... = a*b;
> >
>
> > > }
> >
>
> > > for (....) {
> >
>
> > > if (cold) {
> >
>
> > > ... = a * b;
> >
>
> > > }
> >
>
> > > }
> >
>
> > > Depending on the profile, it might be profitable to do:
> >
>
> > > t = a * b;
> >
>
> > > if (...) {
> >
>
> > > .. = t;
> >
>
> > > }
> >
>
> > > for (...) {
> >
>
> > > if (cold) {
> >
>
> > > .. = t ;
> >
>
> > > }
> >
>
> > > }
> >
>
> > > Again, LICM won't be necessary to enable this.
> >
>
> > > > One might also imagine cases where the two hoistable sections
> > > > are
> > > > SLP
> > > > vectorized.
> > >
> >
>
> > > Will that make it harder to undo the damage later ?
> >
>
> > > > Failing to host the code might also prevent loop unswitching
> > > > (by
> > > > failing to reduce the size of the loop body below the threshold
> > > > size).
> > >
> >
>
> > > There are always existing cleanups that can only happen after
> > > loop-unswitching happens. IMO, loop unswiitching, like inliner
> > > should also look at the code state if the transformation happens.
> >
>
> > > > Another potential issue is that the hoistable code might be
> > > > cold,
> > > > and
> > > > relatively cheap to hoist, but expensive to vectorize. As a
> > > > result,
> > > > failing to hoist the code might block otherwise-profitable
> > > > vectorization. Which reminds me, we need to fix the
> > > > vectorizer's
> > > > if-conversion heuristic to use profiling information too ;)
> > >
> >
>
> > > SLP vectorize? Any example like this? Can vectorizor be enhanced
> > > so
> > > that it can be done in absence of the hoisting?
> >
>
> > > thanks,
> >
>
> > > David
> >
>
> > > > Thanks again,
> > >
> >
>
> > > > Hal
> > >
> >
>
> > > > >
> > >
> >
>
> > > > > Thanks,
> > >
> >
>
> > > > > Dehao
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > > On Tue, May 10, 2016 at 1:58 PM, Hal Finkel < hfinkel at anl.gov
> > > > > >
> > >
> >
>
> > > > > wrote:
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > > From: "Xinliang David Li" < davidxl at google.com >
> > >
> >
>
> > > > > To: "Dehao Chen" < danielcdh at gmail.com >
> > >
> >
>
> > > > > Cc: reviews+D19950+public+38ba22078c2035b8 at reviews.llvm.org ,
> > > > > "David
> > >
> >
>
> > > > > Majnemer" < david.majnemer at gmail.com >, "Hal Finkel" <
> > >
> >
>
> > > > > hfinkel at anl.gov >, "Junbum Lim" < junbuml at codeaurora.org >,
> > >
> >
>
> > > > > mcrosier at codeaurora.org , "llvm-commits" <
> > >
> >
>
> > > > > llvm-commits at lists.llvm.org >, "amara emerson" <
> > >
> >
>
> > > > > amara.emerson at arm.com >
> > >
> >
>
> > > > > Sent: Tuesday, May 10, 2016 3:15:24 PM
> > >
> >
>
> > > > > Subject: Re: [PATCH] D19950: Use frequency info to guide Loop
> > >
> >
>
> > > > > Invariant Code Motion.
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > > On Tue, May 10, 2016 at 1:03 PM, Dehao Chen <
> > > > > danielcdh at gmail.com
> > > > > >
> > >
> >
>
> > > > > wrote:
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > > On Tue, May 10, 2016 at 11:48 AM, Xinliang David Li <
> > >
> >
>
> > > > > davidxl at google.com > wrote:
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > > On Tue, May 10, 2016 at 11:01 AM, Dehao Chen <
> > > > > danielcdh at gmail.com
> > > > > >
> > >
> >
>
> > > > > wrote:
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > > danielcdh added a comment.
> > >
> >
>
> > > > >
> > >
> >
>
> > > > > In http://reviews.llvm.org/D19950#425287 , @hfinkel wrote:
> > >
> >
>
> > > > >
> > >
> >
>
> > > > > > In http://reviews.llvm.org/D19950#425286 , @hfinkel wrote:
> > >
> >
>
> > > > > >
> > >
> >
>
> > > > > > > In http://reviews.llvm.org/D19950#425285 , @davidxl
> > > > > > > wrote:
> > >
> >
>
> > > > > > >
> > >
> >
>
> > > > > > > > Static prediction has been conservative in estimating
> > > > > > > > loop
> > > > > > > > trip
> > >
> >
>
> > > > > > > > count -- it produces something like 30ish iterations.
> > > > > > > > If
> > > > > > > > the
> > > > > > > > a
> > >
> >
>
> > > > > > > > very hot loop has a big if-then-else (or switch), it is
> > > > > > > > very
> > >
> >
>
> > > > > > > > likely to mark many bbs' to be colder than the loop
> > > > > > > > header.
> > >
> >
>
> > > > > > > > Turning on this for static prediction really depends on
> > > > > > > > the
> > >
> >
>
> > > > > > > > false rate. It seems to be this can get wrong pretty
> > > > > > > > easily
> > >
> >
>
> > > > > > > > for very hot loops (which is also the most important
> > > > > > > > thing
> > > > > > > > to
> > >
> >
>
> > > > > > > > optimize for).
> > >
> >
>
> > > > > > >
> > >
> >
>
> > > > > > >
> > >
> >
>
> > > > > > > This is a good point. There's no universal conservative
> > > > > > > choice
> > >
> >
>
> > > > > > > (assuming a small trip count is conservative in some
> > > > > > > cases,
> > > > > > > and
> > >
> >
>
> > > > > > > assuming a large trip count is conservative in other
> > > > > > > cases).
> > >
> >
>
> > > > > >
> > >
> >
>
> > > > > >
> > >
> >
>
> > > > > > Would it be better (and practical) if there were some way
> > > > > > for
> > > > > > the
> > >
> >
>
> > > > > > BFI client to specify which kind of 'conservative' is
> > > > > > desired?
> > >
> >
>
> > > > > >
> > >
> >
>
> > > > > > Also, why are we doing this instead of sinking later (in
> > > > > > CGP
> > > > > > or
> > >
> >
>
> > > > > > similar)? LICM can expose optimization opportunities, plus
> > >
> >
>
> > > > > > represents a code pattern the user might input manually.
> > > > > > Sinking
> > >
> >
>
> > > > > > later seems more robust.
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > > I looked at CGP pass, looks like it's handling the sinking
> > >
> >
>
> > > > > case-by-case (e.g. there is separate routine to handle
> > > > > sinking
> > > > > of
> > >
> >
>
> > > > > load, gep, etc. I'm afraid this would miss opportunities.
> > >
> >
>
> > > > > Additionally, the file-level comment of CGP pass says "This
> > > > > works
> > >
> >
>
> > > > > around limitations in it's basic-block-at-a-time approach. It
> > > > > should
> > >
> >
>
> > > > > eventually be removed."
> > >
> >
>
> > > > > Yes, but it will be "removed" when the entire subsystem is
> > > > > replaced
> > >
> >
>
> > > > > by GlobalISel, and we'll certainly need to make GlobalISel
> > >
> >
>
> > > > > profiling-data aware, so I expect this is the right path
> > > > > forward
> > >
> >
>
> > > > > regardless. I agree, however, that we want a general sinking
> > > > > here
> > >
> >
>
> > > > > based on profiling data, not just the specific existing
> > > > > heuristics
> > >
> >
>
> > > > > for loads, GEPs, etc.
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > > Perhaps you can do profile driven sinking CGP separately to
> > > > > handle
> > >
> >
>
> > > > > manually hoisted code situation mentioned by Hal.
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > > Do you mean we still use frequency to decide whether to hoist
> > > > > code
> > > > > in
> > >
> >
>
> > > > > LICM, additionally use frequency info to check if we want to
> > > > > sink
> > >
> >
>
> > > > > instructions in CGP?
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > > yes -- that is the suggestion. I'd prefer that we try to sink
> > > > > late
> > >
> >
>
> > > > > first, and only if there are use cases that we can't handle
> > > > > this
> > >
> >
>
> > > > > way, we consider throttling hoisting early. If we come across
> > > > > such
> > >
> >
>
> > > > > use cases, I'd like to understand them better. Hoisting can
> > > > > expose
> > >
> >
>
> > > > > other optimization opportunities, and you lose those
> > > > > opportunities
> > >
> >
>
> > > > > if you don't hoist in the first place.
> > >
> >
>
> > > > >
> > >
> >
>
> > > > > -Hal
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > > David
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > > Dehao
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > > David
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > > I'm not quite clear why it helps to move code out of loop
> > > > > early
> > > > > and
> > >
> >
>
> > > > > later sink it inside. Could you give an example or some more
> > >
> >
>
> > > > > context?
> > >
> >
>
> > > > >
> > >
> >
>
> > > > > Thanks,
> > >
> >
>
> > > > > Dehao
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > > http://reviews.llvm.org/D19950
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > > --
> > >
> >
>
> > > > >
> > >
> >
>
> > > > > Hal Finkel
> > >
> >
>
> > > > > Assistant Computational Scientist
> > >
> >
>
> > > > > Leadership Computing Facility
> > >
> >
>
> > > > > Argonne National Laboratory
> > >
> >
>
> > > > >
> > >
> >
>
> > > > >
> > >
> >
>
> > > > --
> > >
> >
>
> > > > Hal Finkel
> > >
> >
>
> > > > Assistant Computational Scientist
> > >
> >
>
> > > > Leadership Computing Facility
> > >
> >
>
> > > > Argonne National Laboratory
> > >
> >
>
--
Hal Finkel
Assistant Computational Scientist
Leadership Computing Facility
Argonne National Laboratory
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20160512/34b68870/attachment.html>
More information about the llvm-commits
mailing list