llvm dev - Oct 2013 - [LLVMdev] Interfacing llvm with a precise, relocating GC

On Thu, Oct 24, 2013 at 6:42 PM, Sanjoy Das <sanjoy at azulsystems.com>
wrote:
> Hi Rafael, Andrew,
>
> Thank you for the prompt reply.
>
> One approach we've been considering involves representing the
> constraint "pointers to heap objects are invalidated at every
> safepoint" somehow in the IR itself.  So, if %a and %b are values the
> GC is interested in, the safepoint at the back-edge of a loop might
> look like:
>
>   ; <label>: body
>     %a = phi i32 [ %a.relocated, %body ] [ %a.initial_value, %pred ]
>     %b = phi i32 [ %b.relocated, %body ] [ %b.initial_value, %pred ]
>     ;; Use %a and %b
>
>     ;; The safepoint starts here
>     %a.relocated = @llvm.gc_relocate(%a)
>     %b.relocated = @llvm.gc_relocate(%b)
>     br %body
>
> This allows us to not bother with relocating derived pointers pointing
> inside %a and %b, since it is semantically incorrect for llvm to reuse
> them in the next iteration of the loop.

This is the right general idea, but you can already express this constraint
in LLVM as it exists today, by using llvm.gcroot().  As you noted, this
also solves the interior-pointer problem by making it the front end's job
to convey to LLVM when it would/would not be safe to cache interior
pointers across loop iterations. The invariant that a front-end must
maintain is that any pointer which is live across a given
potential-GC-point must be reloaded from its root after a (relocating) GC
might have occurred. This falls naturally out of the viewpoint that %a is
not an object pointer, it's the name of an object pointer's value at a
given point in time. So, of course, whenever that object pointer's value
might change, there must be a new name.

The fact that the mutable memory associated with a gcroot() is allocated on
the stack (rather than, say, a machine register) is an implementation
detail; fixing it doesn't require altering the (conceptual) interface for
LLVM's existing GC support, AFAICT.

> We lower gc_relocate to a
> pseudo opcode which lowered into nothing after register allocation.
>
> The problem is, of course, the performance penalty.  Does it make
> sense to get the register allocator "see" the gc_relocate
instruction
> as a copy so that they get the same register / slot?  Will that
> violate the intended semantics of gc_relocate (for example, after
> assigning the same register / slot to %a and %a.relocated, are there
> passes that will try to cache derived pointers across loop
> iterations)?
>
> Thanks,
> -- Sanjoy
>
>
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20131025/d77b69ff/attachment.html>

On 10/25/13 1:10 PM, Ben Karel wrote:
>
>
>
> On Thu, Oct 24, 2013 at 6:42 PM, Sanjoy Das <sanjoy at azulsystems.com 
> <mailto:sanjoy at azulsystems.com>> wrote:
>
>     Hi Rafael, Andrew,
>
>     Thank you for the prompt reply.
>
>     One approach we've been considering involves representing the
>     constraint "pointers to heap objects are invalidated at every
>     safepoint" somehow in the IR itself.  So, if %a and %b are values
the
>     GC is interested in, the safepoint at the back-edge of a loop might
>     look like:
>
>       ; <label>: body
>         %a = phi i32 [ %a.relocated, %body ] [ %a.initial_value, %pred ]
>         %b = phi i32 [ %b.relocated, %body ] [ %b.initial_value, %pred ]
>         ;; Use %a and %b
>
>         ;; The safepoint starts here
>         %a.relocated = @llvm.gc_relocate(%a)
>         %b.relocated = @llvm.gc_relocate(%b)
>         br %body
>
>     This allows us to not bother with relocating derived pointers pointing
>     inside %a and %b, since it is semantically incorrect for llvm to reuse
>     them in the next iteration of the loop. 
>
>
> This is the right general idea, but you can already express this 
> constraint in LLVM as it exists today, by using llvm.gcroot().  As you 
> noted, this also solves the interior-pointer problem by making it the 
> front end's job to convey to LLVM when it would/would not be safe to 
> cache interior pointers across loop iterations. The invariant that a 
> front-end must maintain is that any pointer which is live across a 
> given potential-GC-point must be reloaded from its root after a 
> (relocating) GC might have occurred. This falls naturally out of the 
> viewpoint that %a is not an object pointer, it's the name of an object 
> pointer's value at a given point in time. So, of course, whenever that 
> object pointer's value might change, there must be a new name.
To rephrase, you're saying that the current gcroot mechanism is 
analogous to a boxed object pointer.  We could model a safepoint by 
storing the unboxed value into the box, applying an opaque operation to 
the box, and reloading the raw object pointer from the box.  (The opaque 
operator is key to prevent the store/load pair from being removed.  It 
also implements the actual safepoint operation.)  Is this a correct 
restatement?

Here's an example:
gcroot a = ...; b = ...;
object* ax = *a, bx = *b;
repeat 500000 iterations {
   spin for 50 instructions
   *a = ax;
   *b = bx;
   safepoint(a, b)
   ax = *a;
   bx = *b;
}

If so, I would agree with you that this is a correct encoding of object 
relocation.  I think what got me confused was using the in-tree examples 
to reverse engineer the semantics.  Both the Erlang and Ocaml examples 
insert safepoints after all the LLVM IR passes are run and derived 
pointers were inserted.  This was the part that got me thinking that the 
gcroot semantics were unsuitable for a precise relocating collector.  If 
you completely ignored these implementations and inserted the full 
box/safepoint call/unbox implementation at each safepoint before 
invoking any optimizations, I think this would be correct.

One problem with this encoding is that there does not currently exist an 
obvious mechanism to describe a safepoint in the IR itself.  (i.e. there 
is no llvm.gcsafepoint)  You could model this with a "well known" 
function which a custom GCStrategy recorded a stackmap on every call.  
It would be good to extend the existing intrinsics with such a mechanism.

At least if I'm reading things right, the current *implementation* is 
not a correct implementation of the intended semantics.  In particular, 
the safepoint operation which provides the opaque manipulation of the 
boxed gcroots is not preserved into the SelectionDAG.  As a result, 
optimizations on the SelectionDAG could eliminate the now adjacent 
box/unbox pairs.  This would break a relocating collector.

Leaving this aside, there are also a number of performance problems with 
the current implementation.  I'll summarize them here for now, but will 
expand into approaches for resolving each if this looks like the most 
likely implementation strategy.
1) Explicitly adding box/safepoint/unbox prevents creation of derived 
pointers.  This prevents optimizations such as loop-blocking.  Ideally, 
we'd rather let the derived pointers be created and capture them for 
updating at the safepoint.
2) The redundant loads and stores required for box/unbox.  (These might 
be partially eliminated by other optimization passes.)
3) The inability to allocate GC roots to registers.
4) Frame sizes are implicitly doubled since both an unboxed and boxed 
representation of every object pointer is required.

Frankly, I think that (3) is likely solvable, but that (1) and (4) could 
be fatal for a high performance implementation.  I don't have hard 
numbers on that; I'm simply stating my intuition.
>
> The fact that the mutable memory associated with a gcroot() is 
> allocated on the stack (rather than, say, a machine register) is an 
> implementation detail; fixing it doesn't require altering the 
> (conceptual) interface for LLVM's existing GC support, AFAICT.
While in principal, I agree with you here, in practice the difference is 
actually quite significant.  I am not convinced that the implementation 
would be straight forward.  Let's set this aside for the moment and 
focus on the more fundamental questions.
>
>     We lower gc_relocate to a
>     pseudo opcode which lowered into nothing after register allocation.
>
>     The problem is, of course, the performance penalty.  Does it make
>     sense to get the register allocator "see" the gc_relocate
instruction
>     as a copy so that they get the same register / slot?  Will that
>     violate the intended semantics of gc_relocate (for example, after
>     assigning the same register / slot to %a and %a.relocated, are there
>     passes that will try to cache derived pointers across loop
>     iterations)?
>
>     Thanks,
>     -- Sanjoy
>
>
>     _______________________________________________
>     LLVM Developers mailing list
>     LLVMdev at cs.uiuc.edu <mailto:LLVMdev at cs.uiuc.edu>
>     http://llvm.cs.uiuc.edu
>     http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>
>
>
>
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu          http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20131025/c75e46b2/attachment.html>

I'm also highly interested in relocating-GC support from LLVM. Up until now
my GC implementation has been non-relocating which is obviously kind of a
bummer given that it inhibits certain classes of memory
allocation/deallocation tricks.


I wrote up a bunch of my findings on the implementation of my GC here:
https://code.google.com/p/epoch-language/wiki/GarbageCollectionScheme


Frankly I haven't had time to get much further than idle pondering of how
I'd go about implementing relocation, but it seems to me like the existing
read/write barrier intrinsics might be sufficient if abused properly by the
front-end and lowered carefully. My operating hypothesis has been to stash
barriers at key points - identified by the front-end itself - and possibly
elide them during lowering if it's safe to do so. If my understanding is
correct, it should be possible to lower the barriers into exactly the kind
of boxing/unboxing procedure described in this thread.

Based on my experimentation so far, which is admittedly minimal, I think
the overhead of updating relocated pointers is actually not as bad as it
sounds. It isn't strictly necessary to store both a boxed *and* unboxed
pointer in every frame. In fact, in the current incarnation of gcroot,
marking an alloca as a gcroot effectively forces a stack spill for that
alloca; this means that the relocating collector just updates the single
existing pointer on the stack when it does its magic, and you're done. With
proper barriers in place, and/or careful location of safepoints by the
front-end, it's not that hard to make sure that a relocated pointer gets
updated.

The real trick, as near as I can tell, would be updating registers that
have live roots during a collection. But again based on my past
investigations this should just be a matter of ensuring that
post-collection you have a barrier that is lowered into a register refresh
based on the on-stack relocated pointer value.


One thing I've been meaning to try and do is use this barrier-abuse scheme
to work around the existing lack of support for tracing roots in machine
registers, by effectively setting up an artificial stack spill just prior
to a collection, and otherwise operating on registers as much as possible
instead of gcroot'ed allocas. Again, I've only considered this as a
thought
exercise until now, so I apologize if there's some obvious flaw I'm not
aware of in my reasoning. I haven't actually done any of this stuff yet so
it's more speculative than anything - just trying to creatively engineer
around the existing limitations :-)



 - Mike




On Fri, Oct 25, 2013 at 5:35 PM, Philip Reames <listmail at
philipreames.com>wrote:
>  On 10/25/13 1:10 PM, Ben Karel wrote:
>
>
>
>
> On Thu, Oct 24, 2013 at 6:42 PM, Sanjoy Das <sanjoy at
azulsystems.com>wrote:
>
>> Hi Rafael, Andrew,
>>
>> Thank you for the prompt reply.
>>
>> One approach we've been considering involves representing the
>> constraint "pointers to heap objects are invalidated at every
>> safepoint" somehow in the IR itself.  So, if %a and %b are values
the
>> GC is interested in, the safepoint at the back-edge of a loop might
>> look like:
>>
>>   ; <label>: body
>>     %a = phi i32 [ %a.relocated, %body ] [ %a.initial_value, %pred ]
>>     %b = phi i32 [ %b.relocated, %body ] [ %b.initial_value, %pred ]
>>     ;; Use %a and %b
>>
>>     ;; The safepoint starts here
>>     %a.relocated = @llvm.gc_relocate(%a)
>>     %b.relocated = @llvm.gc_relocate(%b)
>>     br %body
>>
>> This allows us to not bother with relocating derived pointers pointing
>> inside %a and %b, since it is semantically incorrect for llvm to reuse
>> them in the next iteration of the loop.
>
>
>  This is the right general idea, but you can already express this
> constraint in LLVM as it exists today, by using llvm.gcroot().  As you
> noted, this also solves the interior-pointer problem by making it the front
> end's job to convey to LLVM when it would/would not be safe to cache
> interior pointers across loop iterations. The invariant that a front-end
> must maintain is that any pointer which is live across a given
> potential-GC-point must be reloaded from its root after a (relocating) GC
> might have occurred. This falls naturally out of the viewpoint that %a is
> not an object pointer, it's the name of an object pointer's value
at a
> given point in time. So, of course, whenever that object pointer's
value
> might change, there must be a new name.
>
> To rephrase, you're saying that the current gcroot mechanism is
analogous
> to a boxed object pointer.  We could model a safepoint by storing the
> unboxed value into the box, applying an opaque operation to the box, and
> reloading the raw object pointer from the box.  (The opaque operator is key
> to prevent the store/load pair from being removed.  It also implements the
> actual safepoint operation.)  Is this a correct restatement?
>
> Here's an example:
> gcroot a = ...; b = ...;
> object* ax = *a, bx = *b;
> repeat 500000 iterations {
>   spin for 50 instructions
>   *a = ax;
>   *b = bx;
>   safepoint(a, b)
>   ax = *a;
>   bx = *b;
> }
>
> If so, I would agree with you that this is a correct encoding of object
> relocation.  I think what got me confused was using the in-tree examples to
> reverse engineer the semantics.  Both the Erlang and Ocaml examples insert
> safepoints after all the LLVM IR passes are run and derived pointers were
> inserted.  This was the part that got me thinking that the gcroot semantics
> were unsuitable for a precise relocating collector.  If you completely
> ignored these implementations and inserted the full box/safepoint
> call/unbox implementation at each safepoint before invoking any
> optimizations, I think this would be correct.
>
> One problem with this encoding is that there does not currently exist an
> obvious mechanism to describe a safepoint in the IR itself.  (i.e. there is
> no llvm.gcsafepoint)  You could model this with a "well known"
function
> which a custom GCStrategy recorded a stackmap on every call.  It would be
> good to extend the existing intrinsics with such a mechanism.
>
> At least if I'm reading things right, the current *implementation* is
not
> a correct implementation of the intended semantics.  In particular, the
> safepoint operation which provides the opaque manipulation of the boxed
> gcroots is not preserved into the SelectionDAG.  As a result, optimizations
> on the SelectionDAG could eliminate the now adjacent box/unbox pairs.  This
> would break a relocating collector.
>
> Leaving this aside, there are also a number of performance problems with
> the current implementation.  I'll summarize them here for now, but will
> expand into approaches for resolving each if this looks like the most
> likely implementation strategy.
> 1) Explicitly adding box/safepoint/unbox prevents creation of derived
> pointers.  This prevents optimizations such as loop-blocking.  Ideally,
> we'd rather let the derived pointers be created and capture them for
> updating at the safepoint.
> 2) The redundant loads and stores required for box/unbox.  (These might be
> partially eliminated by other optimization passes.)
> 3) The inability to allocate GC roots to registers.
> 4) Frame sizes are implicitly doubled since both an unboxed and boxed
> representation of every object pointer is required.
>
> Frankly, I think that (3) is likely solvable, but that (1) and (4) could
> be fatal for a high performance implementation.  I don't have hard
numbers
> on that; I'm simply stating my intuition.
>
>
>  The fact that the mutable memory associated with a gcroot() is allocated
> on the stack (rather than, say, a machine register) is an implementation
> detail; fixing it doesn't require altering the (conceptual) interface
for
> LLVM's existing GC support, AFAICT.
>
> While in principal, I agree with you here, in practice the difference is
> actually quite significant.  I am not convinced that the implementation
> would be straight forward.  Let's set this aside for the moment and
focus
> on the more fundamental questions.
>
>
>
>> We lower gc_relocate to a
>> pseudo opcode which lowered into nothing after register allocation.
>>
>> The problem is, of course, the performance penalty.  Does it make
>> sense to get the register allocator "see" the gc_relocate
instruction
>> as a copy so that they get the same register / slot?  Will that
>> violate the intended semantics of gc_relocate (for example, after
>> assigning the same register / slot to %a and %a.relocated, are there
>> passes that will try to cache derived pointers across loop
>> iterations)?
>>
>> Thanks,
>> -- Sanjoy
>>
>>
>> _______________________________________________
>> LLVM Developers mailing list
>> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>>
>
>
>
> _______________________________________________
> LLVM Developers mailing listLLVMdev at cs.uiuc.edu        
http://llvm.cs.uiuc.eduhttp://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>
>
>
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20131025/8ff7c861/attachment.html>

On Fri, Oct 25, 2013 at 8:35 PM, Philip Reames <listmail at
philipreames.com>wrote:
>  On 10/25/13 1:10 PM, Ben Karel wrote:
>
>
>
>
> On Thu, Oct 24, 2013 at 6:42 PM, Sanjoy Das <sanjoy at
azulsystems.com>wrote:
>
>> Hi Rafael, Andrew,
>>
>> Thank you for the prompt reply.
>>
>> One approach we've been considering involves representing the
>> constraint "pointers to heap objects are invalidated at every
>> safepoint" somehow in the IR itself.  So, if %a and %b are values
the
>> GC is interested in, the safepoint at the back-edge of a loop might
>> look like:
>>
>>   ; <label>: body
>>     %a = phi i32 [ %a.relocated, %body ] [ %a.initial_value, %pred ]
>>     %b = phi i32 [ %b.relocated, %body ] [ %b.initial_value, %pred ]
>>     ;; Use %a and %b
>>
>>     ;; The safepoint starts here
>>     %a.relocated = @llvm.gc_relocate(%a)
>>     %b.relocated = @llvm.gc_relocate(%b)
>>     br %body
>>
>> This allows us to not bother with relocating derived pointers pointing
>> inside %a and %b, since it is semantically incorrect for llvm to reuse
>> them in the next iteration of the loop.
>
>
>  This is the right general idea, but you can already express this
> constraint in LLVM as it exists today, by using llvm.gcroot().  As you
> noted, this also solves the interior-pointer problem by making it the front
> end's job to convey to LLVM when it would/would not be safe to cache
> interior pointers across loop iterations. The invariant that a front-end
> must maintain is that any pointer which is live across a given
> potential-GC-point must be reloaded from its root after a (relocating) GC
> might have occurred. This falls naturally out of the viewpoint that %a is
> not an object pointer, it's the name of an object pointer's value
at a
> given point in time. So, of course, whenever that object pointer's
value
> might change, there must be a new name.
>
> To rephrase, you're saying that the current gcroot mechanism is
analogous
> to a boxed object pointer.  We could model a safepoint by storing the
> unboxed value into the box, applying an opaque operation to the box, and
> reloading the raw object pointer from the box.  (The opaque operator is key
> to prevent the store/load pair from being removed.  It also implements the
> actual safepoint operation.)  Is this a correct restatement?
>
What does it mean to model a safepoint?

>  Here's an example:
> gcroot a = ...; b = ...;
> object* ax = *a, bx = *b;
> repeat 500000 iterations {
>   spin for 50 instructions
>   *a = ax;
>   *b = bx;
>   safepoint(a, b)
>   ax = *a;
>   bx = *b;
> }
>
> If so, I would agree with you that this is a correct encoding of object
> relocation.  I think what got me confused was using the in-tree examples to
> reverse engineer the semantics.  Both the Erlang and Ocaml examples insert
> safepoints after all the LLVM IR passes are run and derived pointers were
> inserted.  This was the part that got me thinking that the gcroot semantics
> were unsuitable for a precise relocating collector.  If you completely
> ignored these implementations and inserted the full box/safepoint
> call/unbox implementation at each safepoint before invoking any
> optimizations, I think this would be correct.
>
> One problem with this encoding is that there does not currently exist an
> obvious mechanism to describe a safepoint in the IR itself.  (i.e. there is
> no llvm.gcsafepoint)  You could model this with a "well known"
function
> which a custom GCStrategy recorded a stackmap on every call.  It would be
> good to extend the existing intrinsics with such a mechanism.
>
> At least if I'm reading things right, the current *implementation* is
not
> a correct implementation of the intended semantics.  In particular, the
> safepoint operation which provides the opaque manipulation of the boxed
> gcroots is not preserved into the SelectionDAG.  As a result, optimizations
> on the SelectionDAG could eliminate the now adjacent box/unbox pairs.  This
> would break a relocating collector.
>
Since GC safepoints aren't explicitly represented at the IR level,
either, SelectionDAG is a red herring: regular IR-level optimizations need
the same scrutiny. But safepoints won't be inserted arbitrarily. Because
there are only four places where safe points can occur (pre/post call,
loop, return), there are essentially only two optimizations that might
break a relocating collector: a call being reordered with the stores and
loads surrounding it, or constant propagation of an alloca's contents
across a loop backedge without eliminating the backedge (iff the GC plugin
requests loop safepoints). Neither of these are performed by LLVM, AFAICT,
which implies that the implementation is not broken in the way you
suggested.

But if you can prove me wrong by producing an example that LLVM
optimizations break, that would be cool :-)

>  Leaving this aside, there are also a number of performance problems with
> the current implementation.  I'll summarize them here for now, but will
> expand into approaches for resolving each if this looks like the most
> likely implementation strategy.
> 1) Explicitly adding box/safepoint/unbox prevents creation of derived
> pointers.  This prevents optimizations such as loop-blocking.  Ideally,
> we'd rather let the derived pointers be created and capture them for
> updating at the safepoint.
> 2) The redundant loads and stores required for box/unbox.  (These might be
> partially eliminated by other optimization passes.)
> 3) The inability to allocate GC roots to registers.
> 4) Frame sizes are implicitly doubled since both an unboxed and boxed
> representation of every object pointer is required.
>
> Frankly, I think that (3) is likely solvable, but that (1) and (4) could
> be fatal for a high performance implementation.  I don't have hard
numbers
> on that; I'm simply stating my intuition.
>
>
>  The fact that the mutable memory associated with a gcroot() is allocated
> on the stack (rather than, say, a machine register) is an implementation
> detail; fixing it doesn't require altering the (conceptual) interface
for
> LLVM's existing GC support, AFAICT.
>
> While in principal, I agree with you here, in practice the difference is
> actually quite significant.  I am not convinced that the implementation
> would be straight forward.  Let's set this aside for the moment and
focus
> on the more fundamental questions.
>
>
>
>> We lower gc_relocate to a
>> pseudo opcode which lowered into nothing after register allocation.
>>
>> The problem is, of course, the performance penalty.  Does it make
>> sense to get the register allocator "see" the gc_relocate
instruction
>> as a copy so that they get the same register / slot?  Will that
>> violate the intended semantics of gc_relocate (for example, after
>> assigning the same register / slot to %a and %a.relocated, are there
>> passes that will try to cache derived pointers across loop
>> iterations)?
>>
>> Thanks,
>> -- Sanjoy
>>
>>
>> _______________________________________________
>> LLVM Developers mailing list
>> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>>
>
>
>
> _______________________________________________
> LLVM Developers mailing listLLVMdev at cs.uiuc.edu        
http://llvm.cs.uiuc.eduhttp://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20131026/98a0fec2/attachment.html>