llvm dev - Jun 2008 - [LLVMdev] Plans considering first class structs and multiple return values

Hi Dan,
> Yes, the intention is that getresult will be removed once first-class
> aggregates are a ready replacement. This won't leave LLVM missing the
> concept of returning multiple values; a struct can be thought of as
> a container for multiple values.
I'm not saying we don't have some way of modeling multiple return
values, I'm
sayin the explicit concept disappears. We can use a struct return type to
create a function that effectively returns multiple values, but that is still
a function returning a single value: A struct. In particular, it will be
impossible to distinguish between a function returning a single struct and a
function returning multiple values.

I'm not sure this is a big problem, but it makes adding a return value to a
function harder. I'm not sure this is really a problem though. Would adding
a
function attribute returns_multiple or something like that be useful?
returns_multiple would mean to interpret the returned struct as multiple
return values and in particular forbids to use the resulting value in any way
but as an operand to extractvalue. The main goal of this is to make
adding/removing an argument easier, because you only need to modify the
extractvalues. On the other hand, this limitation sounds a lot like the
current getresult approach and might not be all to useful.
> > Additionally, the current form of the ret instruction is still  
> > useful, for
> > making multiple return values readable. In particular, writing
> > 	ret i32 1, i32 2
> > is a lot more readable than the (functionally identical) three line  
> > return
> > statement above. However, to make the first class aggregrates even  
> > more
> > usable, it might be better to remove the multi operand return  
> > instruction and
> > add support for literal aggregrates. Currently, I think these are only
> > supported as global constants. It would be useful if the following  
> > was valid:
> > 	ret { i32, i32 } { i32 1, i32 2 }
> 
> I think this form should be valid once first-class struct support is  more
> complete. If it's not accepted today it may be a conflict with the 
current
> multiple-return-value syntax, or it may be a bug.
It doesn't seem to be accepted by llvm-as. I think you might be able to
build
this in memory, since a ConstantStruct is just a Value*.
> > Even more, one would also like to be able to build non constant 
structs
> > in a similar manner. i.e., writing
> > 	ret { i32, i32 } { i32 %a, i32 %b }
> > would be a lot more useful than the current
> > 	ret i32 %a, i32 %b
> > form, since in the first form the ret instruction still has a single
> > operand that is easy to work with.
> 
> The current design will have it looking like this:
> 
>         %t0 = insertvalue { i32, i32 } undef, i32 %a, 0
>         %t1 = insertvalue { i32, i32 } %t0,   i32 %b, 1
>         ret { i32, i32 } %t1
> 
> once first-class structs take over from the current multiple-return- value
> support. It's significantly more syntax, but it's a significantly 
simpler
> IR schema.
On the other hand, anyone looking to support multiple return values but other
(potentially complicated) uses for first class structs would have a harder
time trying to find out what these nested insertvalues actually do. The main
difference here is that using insertvalue you can do something like:

	%a = phi { i32, i32 } [ %a.0, %foo ], [ %a.1, %bar ]
	%b = insertvalue { i32, i32 } %a, i32 0, 0

which you can't do directly using a literal { } or buildagg kind of
instruction. OTOH, you can still do things like this using nested structs
then, so having a builddag will probably not improve things much.

Anyhow, so much for my blabbering of incoherent thoughts. I think that simply
using insertvalue for now and not having an explicit multiple return function
attribute should work fine.

Whenever I want to add a function argument, I will just let it return a struct
of two elements (current value and the new value). 

I'll also add a feature to the sretpromotion pass that flattens out nested
struct return types as much as possible, without having to reconstruct structs
at the caller end (ie, preserver struct types that are used in any way other
than extractvalue in any caller and flatten out all other elements). Would
this be the right place for that? Or should sretpromotion really only take
care of promotion sret pointer arguments to multiple return values, and have
second pass for flattening them out? A problem of that seems to be that it is
hard for sretpromotion to simply add return values, since it doesn't know
whether to add to the existing struct return type (if any) or create a new
struct return type. I guess integrating these two passes makes the most sense,
then.

I guess the argumentpromotion pass also needs to be adapted to promote first
class struct arguments (probably handle them identical to how byval
pointer-to-struct are handled now), but I don't currently have need of that.

Lastly, I'll modify IPConstProp and DeadArgElim to properly handle multiple
return values and do constprop/removal on each of them individually, instead
of only working when all of them are constant/dead as is done currently.

I'm also thinking of adding a transformation that makes return values dead
if
they only return an unmodified argument, this can probably go into
DeadArgElim.

Gr.

Matthijs
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 189 bytes
Desc: Digital signature
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20080602/963e1959/attachment.sig>

On Jun 2, 2008, at 8:45 AM, Matthijs Kooijman wrote:
> Hi Dan,
>
>> Yes, the intention is that getresult will be removed once first-class
>> aggregates are a ready replacement. This won't leave LLVM missing
the
>> concept of returning multiple values; a struct can be thought of as
>> a container for multiple values.
> I'm not saying we don't have some way of modeling multiple return  
> values, I'm
> sayin the explicit concept disappears. We can use a struct return  
> type to
> create a function that effectively returns multiple values, but that  
> is still
> a function returning a single value: A struct. In particular, it  
> will be
> impossible to distinguish between a function returning a single  
> struct and a
> function returning multiple values.
>
> I'm not sure this is a big problem, but it makes adding a return  
> value to a
> function harder. I'm not sure this is really a problem though. Would  
> adding a
> function attribute returns_multiple or something like that be useful?
> returns_multiple would mean to interpret the returned struct as  
> multiple
> return values and in particular forbids to use the resulting value  
> in any way
> but as an operand to extractvalue. The main goal of this is to make
> adding/removing an argument easier, because you only need to modify  
> the
> extractvalues. On the other hand, this limitation sounds a lot like  
> the
> current getresult approach and might not be all to useful.
The requirement to update all callers' call instructions when a callee
gets a new return value is also present in the current MRV-mechanism
with getresult. It's not been a problem we've worried about so far.
Can you give some background about what kinds of things you're thinking
about for this?
>
>
>>> Additionally, the current form of the ret instruction is still
>>> useful, for
>>> making multiple return values readable. In particular, writing
>>> 	ret i32 1, i32 2
>>> is a lot more readable than the (functionally identical) three line
>>> return
>>> statement above. However, to make the first class aggregrates even
>>> more
>>> usable, it might be better to remove the multi operand return
>>> instruction and
>>> add support for literal aggregrates. Currently, I think these are  
>>> only
>>> supported as global constants. It would be useful if the following
>>> was valid:
>>> 	ret { i32, i32 } { i32 1, i32 2 }
>>
>> I think this form should be valid once first-class struct support  
>> is  more
>> complete. If it's not accepted today it may be a conflict with the
>> current
>> multiple-return-value syntax, or it may be a bug.
> It doesn't seem to be accepted by llvm-as. I think you might be able  
> to build
> this in memory, since a ConstantStruct is just a Value*.
Ok. I'll look into this after some more of the basics of
first-class aggregates are in place.
>
>
>>> Even more, one would also like to be able to build non constant   
>>> structs
>>> in a similar manner. i.e., writing
>>> 	ret { i32, i32 } { i32 %a, i32 %b }
>>> would be a lot more useful than the current
>>> 	ret i32 %a, i32 %b
>>> form, since in the first form the ret instruction still has a
single
>>> operand that is easy to work with.
>>
>> The current design will have it looking like this:
>>
>>        %t0 = insertvalue { i32, i32 } undef, i32 %a, 0
>>        %t1 = insertvalue { i32, i32 } %t0,   i32 %b, 1
>>        ret { i32, i32 } %t1
>>
>> once first-class structs take over from the current multiple- 
>> return- value
>> support. It's significantly more syntax, but it's a
significantly
>> simpler
>> IR schema.
> On the other hand, anyone looking to support multiple return values  
> but other
> (potentially complicated) uses for first class structs would have a  
> harder
> time trying to find out what these nested insertvalues actually do.  
> The main
> difference here is that using insertvalue you can do something like:
>
> 	%a = phi { i32, i32 } [ %a.0, %foo ], [ %a.1, %bar ]
> 	%b = insertvalue { i32, i32 } %a, i32 0, 0
>
> which you can't do directly using a literal { } or buildagg kind of
> instruction. OTOH, you can still do things like this using nested  
> structs
> then, so having a builddag will probably not improve things much.
>
> Anyhow, so much for my blabbering of incoherent thoughts. I think  
> that simply
> using insertvalue for now and not having an explicit multiple return  
> function
> attribute should work fine.
Ok. And as I mentioned before, we can add buildagg (maybe with a
different name ;-)) later if we find it would be of significant
use or convenience. In any case, I'm glad to have someone with a
different perspective thinking about this feature :-).
>
>
> Whenever I want to add a function argument, I will just let it  
> return a struct
> of two elements (current value and the new value).
>
> I'll also add a feature to the sretpromotion pass that flattens out  
> nested
> struct return types as much as possible, without having to  
> reconstruct structs
> at the caller end (ie, preserver struct types that are used in any  
> way other
> than extractvalue in any caller and flatten out all other elements).  
> Would
> this be the right place for that? Or should sretpromotion really  
> only take
> care of promotion sret pointer arguments to multiple return values,  
> and have
> second pass for flattening them out? A problem of that seems to be  
> that it is
> hard for sretpromotion to simply add return values, since it doesn't  
> know
> whether to add to the existing struct return type (if any) or create  
> a new
> struct return type. I guess integrating these two passes makes the  
> most sense,
> then.
I'm not sure what you're saying here. What do you mean by flattening out
nested structs? If you mean moving all the members in nested structs to
be members of a single non-nested struct, that doesn't really buy
anything, because extractvalue and insertvalue can index directly into
nested structs.
>
> I guess the argumentpromotion pass also needs to be adapted to  
> promote first
> class struct arguments (probably handle them identical to how byval
> pointer-to-struct are handled now), but I don't currently have need  
> of that.
>
> Lastly, I'll modify IPConstProp and DeadArgElim to properly handle  
> multiple
> return values and do constprop/removal on each of them individually,  
> instead
> of only working when all of them are constant/dead as is done  
> currently.
>
> I'm also thinking of adding a transformation that makes return  
> values dead if
> they only return an unmodified argument, this can probably go into
> DeadArgElim.
Sounds interesting. One thing to keep in mind here is the tradeoff
between teaching existing optimizations special things about
aggregate values, versus having a separate pass that just promotes
first-class aggregate arguments to a bunch of individual
non-aggregate arguments. The latter would let many existing passes
achieve many of the same results as the former without having to be
burdened with special aggregate knowledge, which would be nice.
But it might be less aggressive in some cases.

Dan

Hi Dan,
> The requirement to update all callers' call instructions when a callee
> gets a new return value is also present in the current MRV-mechanism
> with getresult. It's not been a problem we've worried about so far.
I didn't mean you can get away without updating your calllers, I'm just
saying
it could be a bit easier.
> Can you give some background about what kinds of things you're thinking
> about for this?
For example, when I have a function returning {i32, i32} and I want to add
another i32 to that. If this was a function that simply returns two i32
values, any caller will only use extractvalue on the result to get the
seperate values (since the struct as a whole has no meaning). In this case, I
can make the function return {i32, i32, i32}, add an extra extractvalue in
each caller and be done.

If, on the other hand, the struct currently returns a complex number, for
example (or any other struct of two elements), things get more interesting.
In particular, there will probably be callers that don't only use the
individual values, but also the struct as a whole (for example to pass the
resulting complex number to another function.

Now assume we add another i32 using the same approach as above. Then we make
the function return {i32, i32, i32}. Now assume we just modified foo and there
is some caller that does something like:

	%complex = @foo () ; {i32, i32}
	@bar ( {i32, i32 } %complex)

Then after adding an argument we get something like:

	%tmp = @foo () ; {i32, i32, i32}
	%im  = extractvalue {i32, i32, i32} %tmp, 0
	%re  = extractvalue {i32, i32, i32} %tmp, 1
	%c0  = insertvalue {i32, i32} undef, %im, 0
	%c1  = insertvalue {i32, i32} %c0, %re, 1
	@bar ( {i32, i32 } %c1)

Which isn't pretty code and needs quite some lines of code to generate. In
this case, we're better off creating a new struct, so a function that
returns
{ {i32, i32}, i32}, which means we get something like:

	%tmp = @foo () ; { {i32, i32}, i32}
	%complex  = extractvalue { {i32, i32}, i32} %tmp, 0
	@bar ( {i32, i32 } %complex)

which is a lot nicer and (a bit) easier to generate. It would be nice if, to
add an argument to a function, you wouldn't need to support both of the
above
ways (since the former way is a lot better for a function that returns two
seperate i32's as I described in the first example).

However, the only way to really do this is to make all functions return a
struct, possibly of only a single element. This also requires a guarantee that
nothing special happens to the return value as a whole, but only the
individual elements are accessed through extractvalue.

Alternatively, by adding a function attribute like multiple_return, you could
making the choice between the two ways above easier.

Anyway, I don't think any of the above suggestions are really worth
implementing, I'm just trying to explain my thought process now.

In practice, I guess the second way can be used pretty much everywhere (i.e.,
to add a return values, you simply make the function return a struct of two
elements, regardless of what it returned before [with the exception of void
functions, of course...]).

This approach could possibly result in ugly nested structs like
{{{{i32},i32},i32},i32} after adding three values, for example. Considering
that adding return values is really not used anywhere except for
sretpromotion (and there not really either), this is probably not a big deal.
However, for our particular applications we like the function results to be as
flat as possible, to simplify our codegeneration.
> Ok. And as I mentioned before, we can add buildagg (maybe with a
> different name ;-)) 
Yeah, builddag is an ugly name :-p
> later if we find it would be of significant use or convenience.
By then, we will probably have though our backend to read insertvalue chains,
so it won't be really necessary anymore :-) But let's keep it in mind.
> In any case, I'm glad to have someone with a different perspective
thinking
> about this feature :-).
:-)
> > Whenever I want to add a function argument, I will just let it  
> > return a struct
> > of two elements (current value and the new value).
> >
> > I'll also add a feature to the sretpromotion pass that flattens
out
> > nested
> > struct return types as much as possible, without having to  
> > reconstruct structs
> > at the caller end (ie, preserver struct types that are used in any  
> > way other
> > than extractvalue in any caller and flatten out all other elements).  
> > Would
> > this be the right place for that? Or should sretpromotion really  
> > only take
> > care of promotion sret pointer arguments to multiple return values,  
> > and have
> > second pass for flattening them out? A problem of that seems to be  
> > that it is
> > hard for sretpromotion to simply add return values, since it
doesn't
> > know
> > whether to add to the existing struct return type (if any) or create  
> > a new
> > struct return type. I guess integrating these two passes makes the  
> > most sense,
> > then.
> 
> I'm not sure what you're saying here. What do you mean by
flattening out
> nested structs? If you mean moving all the members in nested structs to
> be members of a single non-nested struct, that doesn't really buy
> anything, because extractvalue and insertvalue can index directly into
> nested structs.
Yeah, that is what I meant. I hadn't thought about the multiple indexing,
good
that you mention it. Still, for our backend, having structs flat whenever
possible will probably make things easier, not sure if that also holds for
other backends? I guess this depends a bit on what codegenerators (will) do
with (nested) struct returns, but I'm not really into that.

Reconsidering, the sretpromotion pass might not be the best place for this. It
currently promotes the special sret pointer arguments to multiple return
values (I should probably modify it to use a struct return and insertvalue
instead?). Since an sret function will, by definition, have a void return
type, the new return values will never have to be merged.

In that light, I will be probably building an internal (i.e., to our company)
pass that flattens struct return values.
> > I guess the argumentpromotion pass also needs to be adapted to  
> > promote first
> > class struct arguments (probably handle them identical to how byval
> > pointer-to-struct are handled now), but I don't currently have
need
> > of that.
> >
> > Lastly, I'll modify IPConstProp and DeadArgElim to properly handle
> > multiple
> > return values and do constprop/removal on each of them individually,  
> > instead
> > of only working when all of them are constant/dead as is done  
> > currently.
> >
> > I'm also thinking of adding a transformation that makes return  
> > values dead if
> > they only return an unmodified argument, this can probably go into
> > DeadArgElim.
> 
> Sounds interesting. One thing to keep in mind here is the tradeoff
> between teaching existing optimizations special things about
> aggregate values, versus having a separate pass that just promotes
> first-class aggregate arguments to a bunch of individual
> non-aggregate arguments.
Promoting struct arguments to multiple seperate values is exactly what I said
that needed happening (but since I'm currently using byval struct*
arguments,
which are already handled by argpromotion) I'm probably not going to change
this.

However, that will not help for return values, as the passes I mentioned
(IPConstProp and DeadArgElim) do not look at individual return values
currently, only the returned struct as a whole. I do not intend to modify the
argument propagation/elimination of these passes, for I agree that that should
be handled by running argumentpromotion first.

Thanks for your insights!

Matthijs
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 189 bytes
Desc: Digital signature
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20080602/13705efc/attachment.sig>