[llvm-dev] RFC: alloca -- specify address space for allocation

[Joseph Tremoulet via llvm-dev]

MM> It is not clear to me if these GC pointers are actually living in a different address space when allocated on the stack (aka. you have two different stacks)
HF> Is the mental model here that a function using this feature actually has multiple stacks, in different address spaces, and this allows creating local stack allocations in those non-addrspace-0 stacks?
PR> the reason we originally went with addrspaces to mark GC pointers was that the managed and non-managed heaps for us are semantically non-overlapping
PR> we might be better off introducing an orthogonal notion for tracking gc references entirely.  The addrspace mechanism has worked, but it is a little bit of a hack
CC> these use cases seem really straight forward for address spaces... But this is very different from the idea of using an alloca with an address spaces for GC pointers to stack objects


I think it's an important question whether address spaces are a good fit for what we're trying to model here.  So I'll explain my mental model of LLILC's address spaces, and I'd be very interested in feedback on whether this seems like a good fit for LLVM's address space concept, or a bastardization thereof:

[preface: it's been my understanding that dereferences of pointers in different address spaces can alias is semantically meaningful ways; i.e. that it's appropriate to use different address spaces to model different means of indexing the same memory.  Some of the comments/questions on this thread seemed to imply instead an expectation that distinct address spaces always reference semantically disjoint storage; if that's a hard assumption, then nothing I'm about to say will make sense and we'll almost certainly need a different mechanism to model GC pointers]

1. The value of an unmanaged/addrspace(0) pointer is an address (in the virtual memory available to the process)
2. The value of a managed/addrspace(1) pointer is conceptually an (ID, offset) pair.  The first component is the "identity" of the object that the offset component is relative to.  Identity is distinct from address; you could imagine the GC heap allocator having a counter that it increments with each allocation, and that an object's identity is the value the counter had when it was allocated.
3. There are two special reserved IDs for pseudo-objects:
  3a. one reserved ID for the null pseudo-object (i.e. what nullptr points to)
  3b. one reserved ID for the "everything-outside-the-GC-heap" pseudo-object.  Conceptually this is an infinitely-large object and the data at offset N in this object is the data at address N (in the virtual memory available to the process, and with a requirement that N does not correspond to an address in the GC heap)
4. The benefit of this conceptual model is that garbage collections are value-preserving, which is why we don't need to model safepoints as rewriting GC pointers (until the RewriteStatepointsForGC pass where we rewrite the IR in terms of addresses rather than object identity)
5. We know that the representation of an addrspace(1) pointer into the outside-the-GC-heap pseudo-object is bit-identical to the representation of an addrspace(0) pointer to the corresponding address  (question: Doesn't that mean we can/should be using bitcast instead of addrspacecast when we know the value in question is an outside-the-GC-heap pointer?)
6. Our source language includes constructs that expose a managed pointer's address (which have well-defined semantics only if the object in question is "pinned").  These naturally correspond to addrspace casts (or maybe need to be additionally constrained?) and are a function of not only the input pointer but also of the layout of the GC heap and the runtime's pinning state.  This is where we can get a mix of addrspace(0) and addrspace(1) pointers whose dereferences alias (in semantically well-defined ways).

Again, I'd love feedback on how sane that sounds to those of you familiar with LLVM's address space notion.

I think that 3b is the part that seems to generate the most surprise/consternation.  The reason our source language includes the "outside-the-GC-heap" pseudo-object is that it allows more powerful code (you could think of it as allowing polymorphism over GC-heap-ness of inputs) with no adverse effects to the system's typesafety guarantees.

Relating this back to the question of what address space an alloca belongs in (which I'm doing to contextualize, not in an attempt to continue the debate), as the stack (of which there is only one) is outside-the-GC-heap, the question is one of whether you want to conceptualize an alloca as producing an unmanaged address or as producing a pointer into the outside-the-GC-heap pseudo-object.  I'd argue that our source language conceptualizes it as the latter[1], which (as Swaroop points out below) is why it would feel natural to model it that way in LLVM, but of course (as Swaroop also points out below) we could also decompose the source construct into two steps in LLVM IR.  As far as compiler-introduced allocas, they of course wouldn't be referenced in the source, and so we wouldn't have a hard requirement either way.

Thanks
-Joseph

[1] - To be more pedantic/precise:
 - Static allocas are implied by local variable declarations, and those declarations are not annotated with a pointer type
 - Our source language includes a compact form that can be used to describe a load or store of a local variable; these compact forms are not annotated with a pointer type
 - Everywhere else that our source language refers to the address of a local variable, it uses the managed pointer type




-----Original Message-----
From: Swaroop Sridhar 
Sent: Saturday, August 29, 2015 12:30 AM
To: Philip Reames <listmail at philipreames.com>; llvm-dev <llvm-dev at lists.llvm.org>; Sanjoy Das <sanjoy at playingwithpointers.com>
Cc: Joseph Tremoulet <jotrem at microsoft.com>; Andy Ayers <andya at microsoft.com>; Russell Hadley <rhadley at microsoft.com>
Subject: RE: RFC: alloca -- specify address space for allocation

> -----Original Message-----
> From: Philip Reames [mailto:listmail at philipreames.com]
> Sent: Friday, August 28, 2015 9:38 AM
> To: Swaroop Sridhar <Swaroop.Sridhar at microsoft.com>; llvm-dev <llvm- 
> dev at lists.llvm.org>; Sanjoy Das <sanjoy at playingwithpointers.com>
> Cc: Joseph Tremoulet <jotrem at microsoft.com>; Andy Ayers 
> <andya at microsoft.com>; Russell Hadley <rhadley at microsoft.com>
> Subject: Re: RFC: alloca -- specify address space for allocation
> 

>> I think for the use case you are outlining, an addrspacecast is the 
>> correct IR model -- you're specifically saying that it is OK in this 
>> case to turn a pointer from addrspace 0 into one for addrspace N 
>> because N is your "managed pointer" set that can be *either* a GC-pointer or a non-GC-pointer.

>> What the FE is saying is that this is an *acceptable* transition of 
>> addrspace, because your language and runtime semantics have provided for it.
>> I think the proper way to say that is with a cast.

> The key bit here is that I think Chandler is right.  You are 
> effectively casting a stack allocation *into* a managed pointer. 
> Having something to mark that transition seems reasonable.

I think there are two views here:

(1) MSIL level view:
In CLR, the stack, is a part of "managed memory" (which is not the same as gc-heap, which is managed and garbage-collected memory).
Therefore, all *references* to stack locations are "managed addresses,"  in the sense that the compiler/runtime exercises certain control over (values that are) managed-address:
For example: it enforces certain restrictions to guarantee safety -- ex: lifetime restrictions, non-null requirement in certain contexts, etc.

This is different from a notion of "unmanaged memory" which is for interoperability with native code.
*Pointers* to unmanaged memory are not controlled by the runtime (ex: do not provide any safety guarantees).

So, from the language semantics point of view, stack addresses are created as managed pointers.
Which is why the proposal is to have alloca directly in the managed address-space seemed natural.

Joseph has written more details in the document that Philip shared out in this thread.

 (2) A more Lower level IR view:
LLVM creates all stack locations in addrespace(0) for all code, whether it comes from managed-code or native code. 
Of these, Stack locations corresponding to the managed-stack are promoted to managed-addresses via addrspacecast.
As an optimization, the FrontEnd inserts the addrspace casts only for those stack locations that are actually address-taken.

If I understand correctly, the recommendation (by Philip, Chandler and David) is approach (2) because:
(a) No change to Instruction-Set is necessary when the semantics is achievable via existing instructions.
(b) It saves changing the optimizer to allocate in the correct address-space. 
Looks like the problem here is that: the optimizer is expected to create type-preserving transformation by allocating in the correct address-space, but blindly allocates in the default address space today.
I don't know the LLVM optimizer well enough to have a good estimate of the magnitude of changes necessary here. But, I agree that (avoiding) substantial changes to the optimizer is a strong consideration.

>> You might need N to be a distinct address space from the one used for 
>> GC-pointers and to have similar casts emitted by the frontend.

Yes, eventually we'll need to differentiate between:
(i) Pointers to unmanaged memory -- which will never be reported to the runtime
(ii) Pointers to GC-heap objects -- which will always be reported to the runtime
(iii) Generic managed pointer -- which may need to be reported if we cannot establish that it points outside the GC heap.

Currently we report all pointers to the runtime as managed pointers.
This is inefficient because the GC then needs to do extra work to figure out what kind of pointer it is:
Pointer to a heap object, pointer within a heap object, or outside the heap.

> Of course, having said that all, I'm back to thinking that having a 
> marker on the alloca would be somewhat reasonable too.  However, I 
> think we need a much stronger justification to change the IR than has 
> been provided.  If you can show that the cast based model doesn't work 
> for some reason, we can re-evaluate.

I don't think we can say that the cast-based model will not work. 
The question is whether alloca addrspace(1)* is a better fit for MSIL semantics, analysis phases, and managed-code specific optimizations.

I'm OK if we conclude that we'll keep using the cast model until we hit a concrete case where it does not work, or seems architecturally misfit.

> Worth noting is that we might be better off introducing an orthogonal 
> notion for tracking gc references entirely.  The addrspace mechanism 
> has worked, but it is a little bit of a hack. We've talked about the 
> need for an opaque pointer type.  Maybe when we actually get around to 
> defining that, the alloca case is one we should consider.

Yes, I'm mainly concerned about getting the right types on the different kinds of pointers. If adders space annotation implies more constraints (ex: on layout) than what's already necessitated by the type distinction, we should use a separate mechanism.

Again, I'm OK if we want to keep using addrspacecast until we hit a concrete case where it breaks down.

Swaroop.