llvm dev - Nov 2018 - [RFC] Tablegen-erated GlobalISel Combine Rules

Hi Daniel,

Thanks for the reply!

Le lun. 26 nov. 2018 à 10:00, Daniel Sanders
<daniel_l_sanders at apple.com> a écrit :
>
> Hi Quentin,
>
> Sorry for the slow reply.
>
> > On Nov 16, 2018, at 09:25, Quentin Colombet <quentin.colombet at
gmail.com> wrote:
> >
> > Hi Daniel,
> >
> > I finally read the proposal.
> >
> > I have a couple of comments that re-enforce what I was thinking
> > initially: it is too early IMHO to develop a full tablegen backend for
> > this. I believe we still miss ways to represent things that matters
> > (see my comments below) and I am afraid we still don't know all
that
> > needs to be represented.
> >
> >> It can reach across basic blocks when it's safe to do so. We
may need to introduce a means to limit the inter-basic-block reach as I can
foresee situations where it may pull expensive operations into a loop if
it's not careful about this
> >
> > I think we would need a syntax for that and in particular, I don't
> > think a generic heuristic based on frequency can solely solve the
> > problem (i.e., what you're hinting in your RFC).
>
> That's something I'm intending to look into properly as part of the
algorithm work. The syntax can support whatever mechanism we need for this
through GIMatchPredicate subclasses and there's room for syntactic sugar by
defining new operator classes.
That's the kind of things that worries me. My cynical view makes me
think that more syntactic sugar sounds like we develop a hack on top
of the suggested syntax, thus I'm pushing for more forward thinking. I
admit I may be wrong and that additional syntactic sugar would fit
well, but I haven't seen evidence one way or another. Therefore, again
I am hesitant to stamp a new tablegen backend based on the few
elements we have.
>
> > Another thing that we miss has to do with the number of users. E.g.,
> > should a pattern only be applied if there is only one user? Is it
> > valid if there are more that one users, if we cannot move something
> > into the destination block can we still apply the pattern is we move
> > the result in the source block, etc.
>
> The existing hasOneUse() checks can be done via by defining a
GIMatchPredicate like so:
>     def hasOneUse : GIMatchPredicate<(ins reg:$A), bool, [{
>       return MRI.hasOneUse(${A});
>     }]>;
> and then using it in the match section. Improving on hasOneUse() to support
the cases where some/all uses are combinable will require algorithm changes and
I'm intending to look into that as part of that work.
>
> > The number of users may not be that relevant, but this gives use is a
> > notion of profitability (it is just how SDISel does this). This
> > actually ties back to the inter-basic-block case: is this pattern
> > profitable if it is between basic blocks with different frequencies.
> > In other words, I believe we should start to think in terms of how
> > profitable is a pattern. E.g., the source pattern as some cost and the
> > destination patterns as some other cost. Applying the pattern should
> > give us a profitability score, which would be higher when it produces
> > dead code (i.e., what the oneUse check tries to identify in SDISel).
> >
> > Following that logic, we may get away with all the disabling specific
> > patterns kind of thing. I.e., if something is not profitable it does
> > not get applied.
>
> I agree. We'll need to decide on precise means to measure profitability
but the underlying mechanism to support this is the metadata. I'd expect a
portion of the cost measurement to be gathered implicitly (more nodes =>
higher cost, some contribution from the scheduler model, etc.) and a portion to
be gathered explicitly via GIMatchPredicate. That value(s) would then be
delivered to the apply step via metadata.
>
> > All in all, the syntax that your suggesting is reasonable (though IMHO
> > MIR is going to limit our ability to share more logic with IR, but
> > maybe that's fine to give up on that goal?), but I feel it
actually
> > distracts us from thinking about the problems we are solving and thus,
> > I am hesitant to stamp a work that would write a new tablegen backend.
>
> I think on that last point we're drawing opposite conclusions from the
same thing :-). I think you see the syntax as defining the Combine algorithm (is
that right?)
No, I really worry that we tie ourselves to a syntax that is not the
right abstraction to define the work. Again, where I would like to go
is a description that can be used for high level combines as well and
MIR seems wrong for this.
> but I see the syntax as abstracting it away from the rules. One of my key
goals with this syntax is to allow the reorganization of the match and apply
steps so that I can try out a different Combine algorithm without blocking the
development of Combines with the current algorithm.
>
> > That said, again, I am not the one doing the work!
> >
> > Cheers,
> > -Quentin
> > [snip]
>

> On Nov 26, 2018, at 10:11, Quentin Colombet <quentin.colombet at
gmail.com> wrote:
> 
> Hi Daniel,
> 
> Thanks for the reply!
> 
> Le lun. 26 nov. 2018 à 10:00, Daniel Sanders
> <daniel_l_sanders at apple.com <mailto:daniel_l_sanders at
apple.com>> a écrit :
>> 
>> Hi Quentin,
>> 
>> Sorry for the slow reply.
>> 
>>> On Nov 16, 2018, at 09:25, Quentin Colombet <quentin.colombet at
gmail.com> wrote:
>>> 
>>> Hi Daniel,
>>> 
>>> I finally read the proposal.
>>> 
>>> I have a couple of comments that re-enforce what I was thinking
>>> initially: it is too early IMHO to develop a full tablegen backend
for
>>> this. I believe we still miss ways to represent things that matters
>>> (see my comments below) and I am afraid we still don't know all
that
>>> needs to be represented.
>>> 
>>>> It can reach across basic blocks when it's safe to do so.
We may need to introduce a means to limit the inter-basic-block reach as I can
foresee situations where it may pull expensive operations into a loop if
it's not careful about this
>>> 
>>> I think we would need a syntax for that and in particular, I
don't
>>> think a generic heuristic based on frequency can solely solve the
>>> problem (i.e., what you're hinting in your RFC).
>> 
>> That's something I'm intending to look into properly as part of
the algorithm work. The syntax can support whatever mechanism we need for this
through GIMatchPredicate subclasses and there's room for syntactic sugar by
defining new operator classes.
> 
> That's the kind of things that worries me. My cynical view makes me
> think that more syntactic sugar sounds like we develop a hack on top
> of the suggested syntax, thus I'm pushing for more forward thinking. I
> admit I may be wrong and that additional syntactic sugar would fit
> well, but I haven't seen evidence one way or another. Therefore, again
> I am hesitant to stamp a new tablegen backend based on the few
> elements we have.
I don't think I understand what you're getting at. At the moment we have
C++ code fragments which can be given names via GIMatchPredicate and are invoked
like a macro with '(GIMatchPredicateDef arg1, arg2)' and everything in
the current match section can be implemented in terms of that. We have
instruction matching which is invoked with '(InstructionDef op1, op2,
op3)' so we can analyze the structure of the DAG and decide how best to
match it. We have room for different kinds of code blocks using '[{ ... some
code ... }]' and choose what they mean. We can also have other kinds of
matching invoked by '(SomethingWeHaventThoughtOfYet ...)' which tblgen
can distinguish from instruction matching by checking the base class of the
'SomethingWeHaventThoughtOfYet' def.

We have lots of extension points we can use to address things beyond
match/apply. What kind of things do you want to consider?
>>> Another thing that we miss has to do with the number of users.
E.g.,
>>> should a pattern only be applied if there is only one user? Is it
>>> valid if there are more that one users, if we cannot move something
>>> into the destination block can we still apply the pattern is we
move
>>> the result in the source block, etc.
>> 
>> The existing hasOneUse() checks can be done via by defining a
GIMatchPredicate like so:
>>    def hasOneUse : GIMatchPredicate<(ins reg:$A), bool, [{
>>      return MRI.hasOneUse(${A});
>>    }]>;
>> and then using it in the match section. Improving on hasOneUse() to
support the cases where some/all uses are combinable will require algorithm
changes and I'm intending to look into that as part of that work.
>> 
>>> The number of users may not be that relevant, but this gives use is
a
>>> notion of profitability (it is just how SDISel does this). This
>>> actually ties back to the inter-basic-block case: is this pattern
>>> profitable if it is between basic blocks with different
frequencies.
>>> In other words, I believe we should start to think in terms of how
>>> profitable is a pattern. E.g., the source pattern as some cost and
the
>>> destination patterns as some other cost. Applying the pattern
should
>>> give us a profitability score, which would be higher when it
produces
>>> dead code (i.e., what the oneUse check tries to identify in
SDISel).
>>> 
>>> Following that logic, we may get away with all the disabling
specific
>>> patterns kind of thing. I.e., if something is not profitable it
does
>>> not get applied.
>> 
>> I agree. We'll need to decide on precise means to measure
profitability but the underlying mechanism to support this is the metadata.
I'd expect a portion of the cost measurement to be gathered implicitly (more
nodes => higher cost, some contribution from the scheduler model, etc.) and a
portion to be gathered explicitly via GIMatchPredicate. That value(s) would then
be delivered to the apply step via metadata.
>> 
>>> All in all, the syntax that your suggesting is reasonable (though
IMHO
>>> MIR is going to limit our ability to share more logic with IR, but
>>> maybe that's fine to give up on that goal?), but I feel it
actually
>>> distracts us from thinking about the problems we are solving and
thus,
>>> I am hesitant to stamp a work that would write a new tablegen
backend.
>> 
>> I think on that last point we're drawing opposite conclusions from
the same thing :-). I think you see the syntax as defining the Combine algorithm
(is that right?)
> 
> No, I really worry that we tie ourselves to a syntax that is not the
> right abstraction to define the work. Again, where I would like to go
> is a description that can be used for high level combines as well and
> MIR seems wrong for this.
There's two things that get locked in. The first is the overall structure:
      (defs ...)
      (match ...)
      (apply ...)
'match' and 'apply' directly correspond to the two major steps
in a combine algorithm: Recognizing existing IR and changing it to something
else. I think these two are safe to lock in as they're already key actions
that any combiner implementation must do. 'defs' is less clear since
it's not as directly connected to a combiners actions but I still think
it's safe to lock in as a concept since the match step needs to tell the
apply step what it matched.

The other bit that gets locked in is that the contents of these sections must
fit within tblgen's 'dag' syntax. That doesn't seem like much of
a burden as code blocks and DagArg/DagArgList
(https://llvm.org/docs/TableGen/LangRef.html#values) are very lenient about
their contents. The operator DagArg can be any def within tablegen and different
backends can act on those defs as they see fit.

The idea of sharing code between InstCombine and MIR for combines that make
sense to both isn't something I intend to work on but it does fit within the
syntax proposed. It would be something along the lines of:
    class IRIndependentBinaryOp;
    def IRIndependentAdd;
    def IRIndependentMul;
    def : GICombineRule<(defs ...), (match (IRIndependentMul $C, $A, 2)),
(apply (IRIndependentAdd $C, $A, $A))>;
The backend(s) would need to know how to match and create LLVM-IR and MIR
versions of IRIndependentMul and IRIndependentAdd much like the GlobalISel
Combiner backend needs to know how to match and create defs that inherit from
Instruction.

* Just as an aside, it's technically an extension on the original syntax and
could co-exist with the MIR code blocks but there's little point in
supporting both methods.
>> but I see the syntax as abstracting it away from the rules. One of my
key goals with this syntax is to allow the reorganization of the match and apply
steps so that I can try out a different Combine algorithm without blocking the
development of Combines with the current algorithm.
>> 
>>> That said, again, I am not the one doing the work!
>>> 
>>> Cheers,
>>> -Quentin
>>> [snip]
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20181127/49be798b/attachment.html>

Hi Daniel,

Let me try to clarify my concern.

Le mar. 27 nov. 2018 à 14:23, Daniel Sanders
<daniel_l_sanders at apple.com> a écrit :
>
>
>
> On Nov 26, 2018, at 10:11, Quentin Colombet <quentin.colombet at
gmail.com> wrote:
>
> Hi Daniel,
>
> Thanks for the reply!
>
> Le lun. 26 nov. 2018 à 10:00, Daniel Sanders
> <daniel_l_sanders at apple.com> a écrit :
>
>
> Hi Quentin,
>
> Sorry for the slow reply.
>
> On Nov 16, 2018, at 09:25, Quentin Colombet <quentin.colombet at
gmail.com> wrote:
>
> Hi Daniel,
>
> I finally read the proposal.
>
> I have a couple of comments that re-enforce what I was thinking
> initially: it is too early IMHO to develop a full tablegen backend for
> this. I believe we still miss ways to represent things that matters
> (see my comments below) and I am afraid we still don't know all that
> needs to be represented.
>
> It can reach across basic blocks when it's safe to do so. We may need
to introduce a means to limit the inter-basic-block reach as I can foresee
situations where it may pull expensive operations into a loop if it's not
careful about this
>
>
> I think we would need a syntax for that and in particular, I don't
> think a generic heuristic based on frequency can solely solve the
> problem (i.e., what you're hinting in your RFC).
>
>
> That's something I'm intending to look into properly as part of the
algorithm work. The syntax can support whatever mechanism we need for this
through GIMatchPredicate subclasses and there's room for syntactic sugar by
defining new operator classes.
>
>
> That's the kind of things that worries me. My cynical view makes me
> think that more syntactic sugar sounds like we develop a hack on top
> of the suggested syntax, thus I'm pushing for more forward thinking. I
> admit I may be wrong and that additional syntactic sugar would fit
> well, but I haven't seen evidence one way or another. Therefore, again
> I am hesitant to stamp a new tablegen backend based on the few
> elements we have.
>
>
> I don't think I understand what you're getting at.
What I am getting at is I feel we are defining a model for a problem
we don't fully understand yet.
I understand the model is flexible enough to inject C++ code wherever
we want but I question why we should have to do that if we had
captured the problem properly in the first place. I am not saying we
don't capture it properly with this design, I am again saying that I
don't have evidence to know.

The reason I am so hesitant here is because it sounds like this syntax
is the final solution whereas I think we could explore other
directions (for instance extending tablegen itself for instance).
However, by committing to this syntax/tablegen backend, when people
start to adopt it, we will have to support them and more importantly
migrate them if we decide to change something, that's why I believe
this is premature.

Now, again you're the one doing the work, so if you believe this step
brings us closer to the final design, I have no issue with that :).

Cheers,
-Quentin
> At the moment we have C++ code fragments which can be given names via
GIMatchPredicate and are invoked like a macro with '(GIMatchPredicateDef
arg1, arg2)' and everything in the current match section can be implemented
in terms of that. We have instruction matching which is invoked with
'(InstructionDef op1, op2, op3)' so we can analyze the structure of the
DAG and decide how best to match it. We have room for different kinds of code
blocks using '[{ ... some code ... }]' and choose what they mean. We can
also have other kinds of matching invoked by '(SomethingWeHaventThoughtOfYet
...)' which tblgen can distinguish from instruction matching by checking the
base class of the 'SomethingWeHaventThoughtOfYet' def.
>
> We have lots of extension points we can use to address things beyond
match/apply. What kind of things do you want to consider?
>
> Another thing that we miss has to do with the number of users. E.g.,
> should a pattern only be applied if there is only one user? Is it
> valid if there are more that one users, if we cannot move something
> into the destination block can we still apply the pattern is we move
> the result in the source block, etc.
>
>
> The existing hasOneUse() checks can be done via by defining a
GIMatchPredicate like so:
>    def hasOneUse : GIMatchPredicate<(ins reg:$A), bool, [{
>      return MRI.hasOneUse(${A});
>    }]>;
> and then using it in the match section. Improving on hasOneUse() to support
the cases where some/all uses are combinable will require algorithm changes and
I'm intending to look into that as part of that work.
>
> The number of users may not be that relevant, but this gives use is a
> notion of profitability (it is just how SDISel does this). This
> actually ties back to the inter-basic-block case: is this pattern
> profitable if it is between basic blocks with different frequencies.
> In other words, I believe we should start to think in terms of how
> profitable is a pattern. E.g., the source pattern as some cost and the
> destination patterns as some other cost. Applying the pattern should
> give us a profitability score, which would be higher when it produces
> dead code (i.e., what the oneUse check tries to identify in SDISel).
>
> Following that logic, we may get away with all the disabling specific
> patterns kind of thing. I.e., if something is not profitable it does
> not get applied.
>
>
> I agree. We'll need to decide on precise means to measure profitability
but the underlying mechanism to support this is the metadata. I'd expect a
portion of the cost measurement to be gathered implicitly (more nodes =>
higher cost, some contribution from the scheduler model, etc.) and a portion to
be gathered explicitly via GIMatchPredicate. That value(s) would then be
delivered to the apply step via metadata.
>
> All in all, the syntax that your suggesting is reasonable (though IMHO
> MIR is going to limit our ability to share more logic with IR, but
> maybe that's fine to give up on that goal?), but I feel it actually
> distracts us from thinking about the problems we are solving and thus,
> I am hesitant to stamp a work that would write a new tablegen backend.
>
>
> I think on that last point we're drawing opposite conclusions from the
same thing :-). I think you see the syntax as defining the Combine algorithm (is
that right?)
>
>
> No, I really worry that we tie ourselves to a syntax that is not the
> right abstraction to define the work. Again, where I would like to go
> is a description that can be used for high level combines as well and
> MIR seems wrong for this.
>
>
> There's two things that get locked in. The first is the overall
structure:
>       (defs ...)
>       (match ...)
>       (apply ...)
> 'match' and 'apply' directly correspond to the two major
steps in a combine algorithm: Recognizing existing IR and changing it to
something else. I think these two are safe to lock in as they're already key
actions that any combiner implementation must do. 'defs' is less clear
since it's not as directly connected to a combiners actions but I still
think it's safe to lock in as a concept since the match step needs to tell
the apply step what it matched.
>
> The other bit that gets locked in is that the contents of these sections
must fit within tblgen's 'dag' syntax. That doesn't seem like
much of a burden as code blocks and DagArg/DagArgList
(https://llvm.org/docs/TableGen/LangRef.html#values) are very lenient about
their contents. The operator DagArg can be any def within tablegen and different
backends can act on those defs as they see fit.
>
> The idea of sharing code between InstCombine and MIR for combines that make
sense to both isn't something I intend to work on but it does fit within the
syntax proposed.
> It would be something along the lines of:
>     class IRIndependentBinaryOp;
>     def IRIndependentAdd;
>     def IRIndependentMul;
>     def : GICombineRule<(defs ...), (match (IRIndependentMul $C, $A,
2)), (apply (IRIndependentAdd $C, $A, $A))>;
> The backend(s) would need to know how to match and create LLVM-IR and MIR
versions of IRIndependentMul and IRIndependentAdd much like the GlobalISel
Combiner backend needs to know how to match and create defs that inherit from
Instruction.
>
> * Just as an aside, it's technically an extension on the original
syntax and could co-exist with the MIR code blocks but there's little point
in supporting both methods.
>
> but I see the syntax as abstracting it away from the rules. One of my key
goals with this syntax is to allow the reorganization of the match and apply
steps so that I can try out a different Combine algorithm without blocking the
development of Combines with the current algorithm.
>
> That said, again, I am not the one doing the work!
>
> Cheers,
> -Quentin
> [snip]
>
>