[LLVMdev] [RFC] Scoped no-alias metadata

[Hal Finkel]

----- Original Message -----
> From: "Chandler Carruth" <chandlerc at google.com>
> To: "Hal Finkel" <hfinkel at anl.gov>
> Cc: "LLVM Developers Mailing List" <llvmdev at cs.uiuc.edu>, "Clang Developers" <cfe-dev at cs.uiuc.edu>, "Dan Gohman"
> <dan433584 at gmail.com>
> Sent: Sunday, December 2, 2012 7:02:52 PM
> Subject: Re: [RFC] Scoped no-alias metadata
> 
> 
> 
> On Sun, Dec 2, 2012 at 12:48 PM, Hal Finkel < hfinkel at anl.gov >
> wrote:
> 
> 
> 
> 
> Hello,
> 
> I'd like to propose a new form of memory-aliasing metadata: scoped
> no-alias metadata. This can be used to model local 'restrict'
> pointers in C99, but should also be useful for other frontends (I
> think, for example, that Fortran will want this as well). Currently,
> we only use the restrict qualifier on function arguments in Clang
> where we translate the restrict qualifier as a 'noalias' function
> parameter attribute.
> 
> In C99, the 'restrict' guarantee holds only for the lifetime of the
> pointer, and different lifetime regions can exist within a single
> function. Furthermore, these lifetime regions can be nested. As a
> result, we'll need to assign a unique identifier to each lifetime
> region (roughly a block in C99, although this may be only a partial
> block if a restrict pointer is declared in the middle of the block).
> Also, during optimization, instructions from these different
> lifetime regions can be merged, so a particular pointer value may
> end up associated with multiple lifetime regions.
> 
> 
> 
> I think a good way to think about this is through the inliner. When
> we inline a function that has a noalias parameter, we should
> preserve that noalias for the region of code inlined, no? This might
> be simpler than describing it in terms of C99 initially, but I think
> we'll be able to support everything we need... Anyways, not really
> relevant to the proposal...

This is also a good use case. When a function is inlined, you'd like to preserve the 'noalias', but only when comparing pointers from the callee, not when comparing pointers from the callee to pointers in the caller.

> 
> 
> Similar to TBAA, the scoped lifetime regions are arranged into
> disjoint tree structures:
> !0 = metadata !{ metadata !"scope of foo()" }
> !1 = metadata !{ metadata !"scope 1", metadata !0 }
> !2 = metadata !{ metadata !"scope 2", metadata !0 }
> !3 = metadata !{ metadata !"scope 2.1", metadata !2 }
> !4 = metadata !{ metadata !"scope 2.2", metadata !2 }
> 
> and these are used to mark the pointer values:
> %arrayidx = getelementptr inbounds %struct.ST* %s, i64 1, i32 2, i32
> 1, i64 5, i64 13, !sna !3
> 
> in C99 this corresponds to something like:
> void foo(...) {
> // begin root scope for foo()
> int *restrict a = ...;
> ...
> {
> // things in this block have scope 1
> int *restrict b = ...;
> ...
> }
> ...
> {
> // things here have scope 2
> int *restrict b = ...;
> ...
> {
> // a new scope 2.1
> }
> ...
> *b += 1; // some use of a restrict pointer
> ...
> // more restrict pointers are declared, start a new nested scope 2.2
> int *restrict c = ...;
> ...
> }
> }
> 
> When BasicAA compares to pointers for aliasing, as it recurses to
> find each pointer's base pointer, it collects a list of scope ids
> associated with each pointer. If one of the pointers is associated
> with a scope id that is identical to one associated with the other
> pointer, or is a descendant or parent to one associated with the
> other pointer, then the two pointers are assumed not to alias. When
> merging two pointer values, if both have scope ids, then the
> resulting value can carry both scope ids. If one does not have a
> scope id, then the resulting pointer must carry none (we can
> preserve optimization ability by moving the metadata from the value
> to be merged to its uses that are also pointer-manipulation
> instructions).
> 
> 
> 
> Maybe I'm misunderstanding, but why can you assume that a particular
> scope will have a distinct GEP to produce the pointer? Or are you
> really relying on the propagation from the pre-optimization copy of
> the pointer the frontend emits to the load/store instructions during
> optimization?

I'm not completely sure I understand your point. A particular GEP could carry multiple scope ids, the real danger comes with sharing a GEP between a restrict pointer in one scope and a non-restrict pointer from a different scope. From my understanding of how Clang CodeGen works, this won't happen (because each statement and initializer gets its own set of instructions; there is no effective CSE).

> 
> 
> Why not just directly attach !noalias metadata to the loads and
> stores? It feels like that would simplify the rules here, and more
> closely match the way TBAA metadata works.

Fair enough; this would make the LLVM-side of the implementation easier too (I suspect). But it involves moving the 'dataflow analysis' part of this into Clang. Any load or store would need to be marked with metadata corresponding to the scope of the underlying restrict pointer (if any). This will either requiring walking up the expression chain on each load/store (which seems inefficient), or some kind of forward propagation phase in Clang prior to CodeGen. Is there currently a good place to do this?

> 
> 
> I think that the most difficult part of implementing this proposal
> would be updating the various transformation passes to preserve this
> metadata.
> 
> 
> 
> I think by directly attaching the aliasing information to loads and
> stores, this is made somewhat easier.

Good point.

Thanks again,
Hal

> 
> 
> Dan, thoughts?

-- 
Hal Finkel
Postdoctoral Appointee
Leadership Computing Facility
Argonne National Laboratory