[llvm-dev] Question about a AA result and its use in Dependence Analysis
De Azevedo Piovezan, Felipe via llvm-dev
llvm-dev at lists.llvm.org
Thu Jun 6 05:38:51 PDT 2019
Hi David,
Would you mind sharing the link to the bug report? I'm not familiar with the code in AA/DA, but I would like to follow the development on this.
Thanks
-----Original Message-----
From: David Green [mailto:David.Green at arm.com]
Sent: Wednesday, June 5, 2019 1:02 PM
To: Eli Friedman <efriedma at quicinc.com>; Doerfert, Johannes <jdoerfert at anl.gov>; De Azevedo Piovezan, Felipe <felipe.de.azevedo.piovezan at intel.com>; llvm-dev <llvm-dev at lists.llvm.org>; Finkel, Hal J. <hfinkel at anl.gov>
Cc: nd <nd at arm.com>
Subject: Re: [llvm-dev] Question about a AA result and its use in Dependence Analysis
Oh. So it wasn't just the memory location size that was incorrect on those aliasing checks.
I took a look at the tests and it appears that the alias call also catches some cases of restrict/noalias attributes, along with the tbaa info. My first idea of constructing an alias call with no Ptr's but valid tbaa info won't catch that at least.
I will create a bug report and try and see if I can come up with anything more sensible.
Dave
From: llvm-dev <llvm-dev-bounces at lists.llvm.org> on behalf of Finkel, Hal J. via llvm-dev <llvm-dev at lists.llvm.org>
Sent: 03 June 2019 22:01
To: Eli Friedman; Doerfert, Johannes; De Azevedo Piovezan, Felipe; llvm-dev
Subject: Re: [llvm-dev] Question about a AA result and its use in Dependence Analysis
On 6/3/19 1:45 PM, Eli Friedman via llvm-dev wrote:
Alias analysis is figuring out the relationship between two pointer expressions, at some location in the program. At a given point in the program, do two expressions always refer to the same location? At a given point in the program, do two expressions never refer to the same location?
AliasAnalysis::alias() doesn't explicitly take a "point" in the program because we don't implement any alias analysis that cares about the precise point in the program; the answer applies to any point in the program dominated by the definitions of both values. This produces useful results because LLVM uses SSA form. Working from that definition, it's easy to see that the logic used in aliasSelect is valid, and that aliasPHI is returning the correct result for the given example. Maybe it's worth calling this out in the alias analysis documentation.
+1
The flaw here seems to be in the way that DA is using AA. DependenceAnalysis's underlyingObjectsAlias is doing this:
// Check the original locations (minus size) for noalias, which can happen for
// tbaa, incompatible underlying object locations, etc.
MemoryLocation LocAS(LocA.Ptr, LocationSize::unknown(), LocA.AATags);
MemoryLocation LocBS(LocB.Ptr, LocationSize::unknown(), LocB.AATags);
if (AA->alias(LocAS, LocBS) == NoAlias)
return NoAlias;
// Check the underlying objects are the same
const Value *AObj = GetUnderlyingObject(LocA.Ptr, DL);
const Value *BObj = GetUnderlyingObject(LocB.Ptr, DL);
// If the underlying objects are the same, they must alias
if (AObj == BObj)
return MustAlias;
// We may have hit the recursion limit for underlying objects, or have
// underlying objects where we don't know they will alias.
if (!isIdentifiedObject(AObj) || !isIdentifiedObject(BObj))
return MayAlias;
// Otherwise we know the objects are different and both identified objects so
// must not alias.
return NoAlias;
and I recall suggesting this use of AA to pick up the metadata-based AA results. We shouldn't need both the AA check here and the underlying-object check also, as the AA check includes an underlying-object check (and not much else, aside from checks on metadata, when both sizes are unknown). However, as this example shows, the fact that the two SSA values in question refer to different, identified underlying objects, doesn't prevent them from having a cross-iteration dependence as the identify of those underlying objects can change across different iterations (and, thus, the underlying object of one variable in one iteration might be the underlying object used by another variable in another iteration).
I'm open to suggestions on how to best resolve this situation. I'm leaning toward suggesting that we have a metadata-only AA query, and use that here, instead of calling something that will invoke BasicAA's logic. My metadata-only AA query, I specifically mean a fundamental dependency-analysis query, likely implemented by the AA subsystem for code-reuse reasons, which answers the question: is there some reason to believe that any instances of these two values never take the same value along any dynamic path through the CFG (or something similar).
Thanks again,
Hal
For information across loop iterations, the appropriate analysis is probably LoopAccessAnalysis.
-Eli
-----Original Message-----
From: llvm-dev <llvm-dev-bounces at lists.llvm.org> On Behalf Of Doerfert, Johannes via llvm-dev
Sent: Monday, June 3, 2019 9:04 AM
To: De Azevedo Piovezan, Felipe <felipe.de.azevedo.piovezan at intel.com>
Cc: llvm-dev at lists.llvm.org
Subject: [EXT] Re: [llvm-dev] Question about a AA result and its use in Dependence Analysis
Thanks for investigating this further.
In BasicAAResult::aliasSelect the "same" logic is used:
// If the values are Selects with the same condition, we can do a more precise
// check: just check for aliases between the values on corresponding arms.
That is what happens for PHI nodes as well. So, if either is broken both are.
Now it remains to definitively decide if this is broken (which I still think).
---------------------------------------
Johannes Doerfert
Researcher
Argonne National Laboratory
Lemont, IL 60439, USA
jdoerfert at anl.gov
________________________________________
From: De Azevedo Piovezan, Felipe <felipe.de.azevedo.piovezan at intel.com>
Sent: Monday, June 3, 2019 11:00
To: Doerfert, Johannes
Cc: llvm-dev at lists.llvm.org; Finkel, Hal J.
Subject: RE: [llvm-dev] Question about a AA result and its use in Dependence Analysis
It seems the same bug is there if we do pointer swapping with selects. Do you agree? (see example below)
define void @f() {
entry:
%a1 = alloca float, align 4
%a2 = alloca float, align 4
br label %loop
end:
ret void
loop:
%phi = phi i32 [ 0, %entry ], [ 1, %loop ]
%select_cond = icmp eq i32 %phi, 0
%ptr1 = select i1 %select_cond, float* %a1, float* %a2
%ptr2 = select i1 %select_cond, float* %a2, float* %a1
store float 0.000000e+00, float* %ptr1, align 4
store float 1.000000e+00, float* %ptr2, align 4
br i1 %select_cond, label %end, label %loop }
Alias Set Tracker: 2 alias sets for 2 pointer values.
AliasSet[0x55bd786b8300, 1] must alias, Mod Pointers: (float* %ptr1,
LocationSize::precise(4))
AliasSet[0x55bd786b96b0, 1] must alias, Mod Pointers: (float* %ptr2,
LocationSize::precise(4))
-----Original Message-----
From: Doerfert, Johannes [mailto:jdoerfert at anl.gov]
Sent: Saturday, June 1, 2019 2:29 PM
To: De Azevedo Piovezan, Felipe <felipe.de.azevedo.piovezan at intel.com>
Cc: llvm-dev at lists.llvm.org; Finkel, Hal J. <hfinkel at anl.gov>
Subject: Re: [llvm-dev] Question about a AA result and its use in Dependence Analysis
Hi Felipe,
(+ Hal)
I think the reasoning in `BasicAAResult::aliasPHI(...)` is flawed but I want someone else to take a look as well.
As far as I can tell, the reasoning there is as follows:
- Two phi nodes in the same block do not alias by default.
- They do alias if the pair of inputs from the same predecessor alias.
- If all inputs are pair-wise alias free, they do not alias.
Now in the case below, %g and %h do clearly not alias, which is what the query for %entry will determine. Since %p and %q are initially assumed to be not aliasing, the query for the incoming values from %for.body will return NoAlias.
Cheers,
Johannes
On 06/01, De Azevedo Piovezan, Felipe wrote:
Hi Johannes,
I followed your advice and got the same result: NoAlias and No dependence.
Would you say AA is faulty for saying NoAlias or DA is faulty for saying no dependence? Or both?
(revised example below)
Thanks!
define float @f() {
entry:
%g = alloca float, align 4
%h = alloca float, align 4
br label %for.body
for.cond.cleanup: ; preds = %for.body
ret float undef
for.body: ; preds = %for.body, %entry
%p = phi float* [ %g, %entry ], [ %q, %for.body ]
%q = phi float* [ %h, %entry ], [ %p, %for.body ]
%0 = load float, float* %p, align 4
store float undef, float* %q, align 4
%branch_cond = fcmp ugt float %0, 0.0
br i1 %branch_cond, label %for.cond.cleanup, label %for.body }
Alias Set Tracker: 2 alias sets for 2 pointer values.
AliasSet[0x83e1fe0, 1] must alias, Ref Pointers: (float* %p,
LocationSize::precise(4))
AliasSet[0x83e3390, 1] must alias, Mod Pointers: (float* %q,
LocationSize::precise(4))
da analyze -
%0 = load float, float* %p, align 4
store float undef, float* %q, align 4 none!
-----Original Message-----
From: Doerfert, Johannes [mailto:jdoerfert at anl.gov]
Sent: Friday, May 31, 2019 9:07 PM
To: De Azevedo Piovezan, Felipe <felipe.de.azevedo.piovezan at intel.com>
Cc: llvm-dev at lists.llvm.org
Subject: Re: [llvm-dev] Question about a AA result and its use in Dependence Analysis
Can you try it without the undef branch condition. If you get still the same
result I'd argue its a bug. In the program as it is, I'd say it is not because the back edge is never taken.
On 05/31, De Azevedo Piovezan, Felipe via llvm-dev wrote:
Hello llvm-dev,
I would appreciate your feedback on the following problem. We're trying to
determine whether this is a bug in LLVM or not.
In the IR snippet below, we have two pointers (p and q) which initially point
to two completely non-overlapping locations. Then, on every iteration of a loop, we swap the pointers and load from the first, followed by a store to the second.
1) AA says the two pointers are NoAlias, even though they do point to the
same location if we consider them in distinct loop iterations. Is this right?
2) Dependence Analysis says there is no dependence between the load and
the store. Is this right?
define float @f() {
entry:
%g = alloca float, align 4
%h = alloca float, align 4
br label %for.body
for.cond.cleanup:
ret float undef
for.body:
%p = phi float* [ %g, %entry ], [ %q, %for.body ]
%q = phi float* [ %h, %entry ], [ %p, %for.body ]
%0 = load float, float* %p, align 4
store float undef, float* %q, align 4
br i1 undef, label %for.cond.cleanup, label %for.body }
AliasSet[0x872d800, 1] must alias, Ref Pointers: (float* %p,
LocationSize::precise(4))
AliasSet[0x872d8b0, 1] must alias, Mod Pointers: (float* %q,
LocationSize::precise(4))
da analyze -
%0 = load float, float* %p, align 4 ; I added these two debug statements,
DA doesn't print the values it is looking at...
store float undef, float* %q, align 4 none!
--
Felipe
_______________________________________________
LLVM Developers mailing list
llvm-dev at lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
--
Johannes Doerfert
Researcher
Argonne National Laboratory
Lemont, IL 60439, USA
jdoerfert at anl.gov
--
Johannes Doerfert
Researcher
Argonne National Laboratory
Lemont, IL 60439, USA
jdoerfert at anl.gov
_______________________________________________
LLVM Developers mailing list
llvm-dev at lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
_______________________________________________
LLVM Developers mailing list
llvm-dev at lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
--
Hal Finkel
Lead, Compiler Technology and Programming Languages Leadership Computing Facility Argonne National Laboratory
More information about the llvm-dev
mailing list