[llvm-dev] [RFC] Delaying phi-to-select transformation until later in the pass pipeline

[John Brawn via llvm-dev]

Summary
=======

I'm planning on adjusting SimplifyCFG so that it doesn't turn two-entry phi
nodes into selects until later in the pass pipeline, to give passes which can
understand phis but not selects more opportunity to optimize. The thing I'm
trying to do which made me think of doing this is described below, but from the
benchmarking I've done it looks like this is overall a good idea regardless of
if I manage to get that done or not.

Motivation
==========

My goal is to get clang to optimize some code containing a call to
std::min_element which is dereferenced, so something like:

float min_element_example(float *data, int size)
{
return *std::min_element(data, data+size);
}

which, after inlining a specialization, looks like:

float min_element_example_inlined(float *first, float *last)
{
for (float *p = first; p != last; ++p)
{
if (*p < *first)
first = p;
}
return *first;
}

There are two loads in the loop, *p and *first, but actually the load *p can be
eliminated by using either the previous load *p or the previous *first,
depending on if the if-condition was taken or not. However the
"if (*p < *first) first = p" gets turned by simplifycfg into a select and this
makes optimizing this a lot harder because you no longer have distinct paths
through the CFG.

I have some ideas on how to do the optimization (see my previous RFC "Making GVN
able to visit the same block more than once" posted in April, though I've
decided that the specific idea presented there isn't the right way to do it),
but I think the first step is to make sure we don't have a select when we try
to optimise this.

Approach
========

I've posted a patch to https://reviews.llvm.org/D50723 showing what I'm
intending to do. An extra parameter is added to SimplifyCFG to control whether
two-entry phi nodes are converted into select, and this is set to false in all
instances before the end of module simplification. At the end of module
simplification we do SimplifyCFG, then Instcombine to optimise the selects that
are introduced, then EarlyCSE to eliminate common subexpressions introduced by
instcombine.

Benchmark Results
=================

These are performance differences reported by LNT when running llvm-test-suite,
spec2000, and spec2006 at -O3 with and without the patch linked above (using
trunk llvm from a week or so ago).

AArch64 results on ARM Cortex-A72:

Performance Regressions - execution_time Change
SingleSource/Benchmarks/Shootout/Shootout-ary3 9.48%
MultiSource/Benchmarks/TSVC/Packing-flt/Packing-flt 3.79%
SingleSource/Benchmarks/CoyoteBench/huffbench 1.40%

Performance Improvements - execution_time Change
MultiSource/Benchmarks/TSVC/Searching-dbl/Searching-dbl -23.74%
External/SPEC/CINT2000/256.bzip2/256.bzip2 -9.82%
MultiSource/Benchmarks/TSVC/Searching-flt/Searching-flt -9.57%
MultiSource/Benchmarks/TSVC/Equivalencing-flt/Equivalencing-flt -4.38%
MultiSource/Benchmarks/TSVC/LinearDependence-flt/LinearDependence-flt -3.94%
MultiSource/Benchmarks/TSVC/Packing-dbl/Packing-dbl -3.44%
External/SPEC/CFP2006/453.povray/453.povray -2.50%
SingleSource/Benchmarks/Adobe-C++/stepanov_vector -1.49%

X86_64 results on Intel Xeon E5-2690:

Performance Regressions - execution_time Change
MultiSource/Benchmarks/Ptrdist/yacr2/yacr2 5.62%

Performance Improvements - execution_time Change
SingleSource/Benchmarks/Misc-C++/Large/sphereflake -4.43%
External/SPEC/CINT2006/456.hmmer/456.hmmer -2.50%
External/SPEC/CINT2006/464.h264ref/464.h264ref -1.60%
MultiSource/Benchmarks/nbench/nbench -1.19%
SingleSource/Benchmarks/Adobe-C++/functionobjects -1.07%

I had a brief look at the regressions and they all look to be caused by
getting bad luck with branch mispredictions: I looked into the Shootout-ary3 and
yacr2 cases and in both the hot code path was the same with and without the
patch, but with more mispredictions probably caused by changes elsewhere.

John

_______________________________________________
LLVM Developers mailing list
llvm...@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

[Krzysztof Parzyszek via llvm-dev]

I think it would be good to have the CFG simplification options provided
by TTI, including the limit on phi->select transformation. We have some
code on Hexagon that would benefit not only from converting 1 phi to
select (as is the limit now), but 4 of them.

-Krzysztof

--
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
hosted by The Linux Foundation

[Sanjay Patel via llvm-dev]

I didn't look closely at the new patch, but this appears to be a small extension to:

https://reviews.llvm.org/D38566

...and the GVN-based reasons for delaying transformation to 'select' are discussed in detail in the motivating bug for that patch:

https://bugs.llvm.org/show_bug.cgi?id=34603

So this sounds like the right direction to me. Note that there was objection to the implementation (a pile of pass options vs. uniquely named passes).

Here's another motivating bug where early transform to select prevents optimization:

https://bugs.llvm.org/show_bug.cgi?id=36760

Is that case affected by this patch?

[Philip Reames via llvm-dev]

I'm concerned that we're focusing on one side of this.  Let me point out a few concerns w/changing the canonical form here:

1. LICM does not know how to hoist or sink regions.  It does know how to hoist and sink selects.
2. InstCombine has limited support for triangles/diamonds, but fairly extensive support for selects.
3. EarlyCSE and GVN do not know how to eliminate fully redundant triangles/diamonds.  PRE *may* get some of these cases, but that is by no means guaranteed or likely to be robust.

My personal opinion is that selects are the appropriate canonical form.  

For the one of the specific cases mentioned, teaching GVN to do FRE and PRE for loads from selects of pointers just doesn't seem that painful.  I would be really tempted to do that instead.  Similarly, walking facts back from select uses in CVP seems generally useful since we have use dependent facts in a bunch of contexts, not just selects.  (Call arguments for instance,  non-null from unconditional deref, etc..)

To be clear, I am raising concerns, not actively objecting to this.  If you want to move forward and commit to work through all of the issues identified I wont actively stand in the way.

Philip

[Hal Finkel via llvm-dev]

On 08/15/2018 02:38 PM, Philip Reames via llvm-dev wrote:

I'm concerned that we're focusing on one side of this.  Let me point out a few concerns w/changing the canonical form here:

1. LICM does not know how to hoist or sink regions.  It does know how to hoist and sink selects.
2. InstCombine has limited support for triangles/diamonds, but fairly extensive support for selects.
3. EarlyCSE and GVN do not know how to eliminate fully redundant triangles/diamonds.  PRE *may* get some of these cases, but that is by no means guaranteed or likely to be robust.

My personal opinion is that selects are the appropriate canonical form.  

For the one of the specific cases mentioned, teaching GVN to do FRE and PRE for loads from selects of pointers just doesn't seem that painful.  I would be really tempted to do that instead.  Similarly, walking facts back from select uses in CVP seems generally useful since we have use dependent facts in a bunch of contexts, not just selects.  (Call arguments for instance,  non-null from unconditional deref, etc..)

To be clear, I am raising concerns, not actively objecting to this.  If you want to move forward and commit to work through all of the issues identified I wont actively stand in the way.

As I've expressed in the past, I think that not using select as part of our canonical form is potentially a superior design. However, it would be a major change. In addition to the issues that Philip mentions, there's also the fact that we'll just have more basic blocks and that, in itself, could lead to an increase in compile time. However, working through these issues will likely leave us with a more-robust optimizer.

See also: https://bugs.llvm.org/show_bug.cgi?id=34603#c19

 -Hal

-- 
Hal Finkel
Lead, Compiler Technology and Programming Languages
Leadership Computing Facility
Argonne National Laboratory

[Amara Emerson via llvm-dev]

On Aug 15, 2018, at 10:57 PM, Hal Finkel via llvm-dev <llvm...@lists.llvm.org> wrote:

On 08/15/2018 02:38 PM, Philip Reames via llvm-dev wrote:

I'm concerned that we're focusing on one side of this.  Let me point out a few concerns w/changing the canonical form here:

1. LICM does not know how to hoist or sink regions.  It does know how to hoist and sink selects.

I think we should teach LICM to do this eventually.

2. InstCombine has limited support for triangles/diamonds, but fairly extensive support for selects.
3. EarlyCSE and GVN do not know how to eliminate fully redundant triangles/diamonds.  PRE *may* get some of these cases, but that is by no means guaranteed or likely to be robust.

Agreed, we’ll need a plan to deal with these issues.

My personal opinion is that selects are the appropriate canonical form.  

For the one of the specific cases mentioned, teaching GVN to do FRE and PRE for loads from selects of pointers just doesn't seem that painful.  I would be really tempted to do that instead.  Similarly, walking facts back from select uses in CVP seems generally useful since we have use dependent facts in a bunch of contexts, not just selects.  (Call arguments for instance,  non-null from unconditional deref, etc..)

To be clear, I am raising concerns, not actively objecting to this.  If you want to move forward and commit to work through all of the issues identified I wont actively stand in the way.

As I've expressed in the past, I think that not using select as part of our canonical form is potentially a superior design. However, it would be a major change. In addition to the issues that Philip mentions, there's also the fact that we'll just have more basic blocks and that, in itself, could lead to an increase in compile time. However, working through these issues will likely leave us with a more-robust optimizer.

See also: https://bugs.llvm.org/show_bug.cgi?id=34603#c19

Canonicalizing to phis is my preference too.

Is scheduling another simplifycfg, instcombine + earlycse pass really necessary? I’m concerned about the compile time impact.

Cheers,

Amara

[Philip Reames via llvm-dev]

On 08/17/2018 09:02 AM, Amara Emerson wrote:

On Aug 15, 2018, at 10:57 PM, Hal Finkel via llvm-dev <llvm...@lists.llvm.org> wrote:

On 08/15/2018 02:38 PM, Philip Reames via llvm-dev wrote:

I'm concerned that we're focusing on one side of this.  Let me point out a few concerns w/changing the canonical form here:

1. LICM does not know how to hoist or sink regions.  It does know how to hoist and sink selects.

I think we should teach LICM to do this eventually.

Agreed.  I just don't currently see patches to do so.  Once LICM supports region hoisting, much of my concern disappears.

2. InstCombine has limited support for triangles/diamonds, but fairly extensive support for selects.
3. EarlyCSE and GVN do not know how to eliminate fully redundant triangles/diamonds.  PRE *may* get some of these cases, but that is by no means guaranteed or likely to be robust.

Agreed, we’ll need a plan to deal with these issues.

Again, I'd love to see these issues fixed.  Once that's done, the convert about phi vs select as canonical form is much less relevant. 

[John Brawn via llvm-dev]

> Here's another motivating bug where early transform to select prevents optimization:

> https://bugs.llvm.org/show_bug.cgi?id=36760

 

> Is that case affected by this patch?

From a quick reading of that bug it looks like CVP needs to be enhanced to do a certain transformation, but the presence
of a select means that even with that enhancement it wouldn’t do anything. With this patch we have a phi and so I think
that if the CVP enhancement were done it should work.

>>> 1. LICM does not know how to hoist or sink regions.  It does know how to hoist and sink selects.

>>I think we should teach LICM to do this eventually.
>Agreed.  I just don't currently see patches to do so.  Once LICM supports region hoisting, much of my concern disappears.

I’ve looked into what happens with LICM with this change, and it looks like everything except the phi gets hoisted, then
later the phi is turned into a select, then a later LICM pass hoists the select. For the very simple tests I tried we end up with
the same result in the end, but I wouldn’t be surprised if indeed LICM not doing this makes things worse in more complex
cases.

I’ve hacked together a modification to LICM to make it able to hoist phis (by essentially hoisting entire blocks instead of
moving everything into a single preheader block) and it seemed to work OK, so I’ll get to work on getting that working
properly and committed before I touch phi-to-select transformation.

> InstCombine has limited support for triangles/diamonds, but fairly extensive support for selects.

Currently I have InstCombine running directly after phis have been turned into selects, so this should only be a problem if
something before that relies on the select instcombine having already happened. I’ll look into that to see if it is indeed a
problem.

>>>EarlyCSE and GVN do not know how to eliminate fully redundant triangles/diamonds.  PRE *may* get some of these cases, but that is by no means guaranteed or likely to be robust.
>>Agreed, we’ll need a plan to deal with these issues.
>Again, I'd love to see these issues fixed.  Once that's done, the convert about phi vs select as canonical form is much less relevant. 

I’ll look into this.

> For the one of the specific cases mentioned, teaching GVN to do FRE and PRE for loads from selects of pointers just doesn't seem that painful.  I would be really tempted to do that
> instead.  Similarly, walking facts back from select uses in CVP seems generally useful since we have use dependent facts in a bunch of contexts, not just selects.  (Call arguments for
> instance,  non-null from unconditional deref, etc..)

Actually making GVN able to handle selects being painful is exactly the reason I’m proposing doing this. GVN has a hard
constraint that a single block can only have a single available value, and attempting to change that (which you have to do
when selecting between values defined in the same block) causes a ton of problems that have to be resolved.

John

[Philip Reames via llvm-dev]

Sorry for the delayed response. 

On 08/23/2018 10:01 AM, John Brawn wrote:

> Here's another motivating bug where early transform to select prevents optimization:

> https://bugs.llvm.org/show_bug.cgi?id=36760

 

> Is that case affected by this patch?

From a quick reading of that bug it looks like CVP needs to be enhanced to do a certain transformation, but the presence
of a select means that even with that enhancement it wouldn’t do anything. With this patch we have a phi and so I think
that if the CVP enhancement were done it should work.

>>> 1. LICM does not know how to hoist or sink regions.  It does know how to hoist and sink selects.
>>I think we should teach LICM to do this eventually.
>Agreed.  I just don't currently see patches to do so.  Once LICM supports region hoisting, much of my concern disappears.

I’ve looked into what happens with LICM with this change, and it looks like everything except the phi gets hoisted, then
later the phi is turned into a select, then a later LICM pass hoists the select. For the very simple tests I tried we end up with
the same result in the end, but I wouldn’t be surprised if indeed LICM not doing this makes things worse in more complex
cases.

I’ve hacked together a modification to LICM to make it able to hoist phis (by essentially hoisting entire blocks instead of
moving everything into a single preheader block) and it seemed to work OK, so I’ll get to work on getting that working
properly and committed before I touch phi-to-select transformation.

This would be a really useful extension to LICM.  I look forward to the patch.  I was half considering an approach recently where LICM itself did phi-to-select if all of the operands to the resulting phi would be loop invariant.  It was just ugly enough I didn't move forward with it, but we have seen cases where it would resolve pass ordering problems leading to strictly better results overall.  I suspect your region hoister would have the same overall effect.