[llvm-dev] Early CSE clobbering llvm.assume

[Hal Finkel via llvm-dev]

----- Original Message -----

> From: "Peter Lawrence" <c_plawre at qca.qualcomm.com>
> To: "Daniel Berlin" <dberlin at dberlin.org>, "Hal Finkel"
> <hfinkel at anl.gov>
> Cc: "llvm-dev" <llvm-dev at lists.llvm.org>
> Sent: Tuesday, June 14, 2016 9:04:16 PM
> Subject: RE: [llvm-dev] Early CSE clobbering llvm.assume

> Daniel,
> This first part is to whoever you are quoting, I can’t tell from the
> email,

> The more information made available to the optimizers the better the
> optimizations,
> Asserts provide more information,
> You * should * expect better code with asserts enabled.

> And this is * not * a bad thing !!!

> And IMHO there is no winning argument that says we should not use
> this information.

> In other words saying “asserts aren’t for optimization” isn’t a
> winning argument,
> Its information and it isn’t useful to throw that information away.
Unfortunately, it is not that simple. When asserts are enabled, the conditions they encapsulate need to be checked at runtime. This checking has a cost (especially if the asserted condition is not cheap to compute), and also introduces extra control flow which can inhibit optimization. It might also provide additional information which can help optimization, but generally speaking, in part because of usage conventions (e.g. using expensive-to-compute expressions in the assert), enabling asserts does not generally help performance. 

> It’s not very far removed from saying “if, while, and for conditions”
> aren’t for optimization.

It is quite far removed. 

Unfortunately, asserts are a feature designed to help debugging, and not optimization. This is why defining NDEBUG turns them off. We would like a feature, similar in spirit, that is intended to help optimization. This is what assume is for. The contracts feature being discussed for C++ also has specific allowances for usage for optimization. 

> This second part is for your response,

> Hmmm, I don’t get how you go from “assert having a call to abort()”
> to a bunch of talk
> About “branch around unreachable” ? the condition isn’t unreachable,
> the abort isn’t
> Unreachable, what’s unreachable ???
I think we're confusing several branches of the conversation here. There actually is an unreachable after the abort because the abort does not return. We were talking, however, about using an alternate representation for assumes which used control flow and unreachable. In this case, there would be no abort, just the unreachable. 

-Hal 

> --Peter.

> From: Daniel Berlin [mailto:dberlin at dberlin.org]
> Sent: Tuesday, June 14, 2016 11:23 AM
> To: Hal Finkel <hfinkel at anl.gov>
> Cc: Lawrence, Peter <c_plawre at qca.qualcomm.com>; llvm-dev
> <llvm-dev at lists.llvm.org>
> Subject: Re: [llvm-dev] Early CSE clobbering llvm.assume

> > > > Sanjoy’s argument is faulty, if it were true we would also find
> > > > our
> > > > handling of “assert” to be unacceptable
> > > 
> > 
> 
> > > > but this is not the case, no one is arguing that we need to
> > > > re-design
> > > > “assert”
> > > 
> > 
> 
> > Sure, but no one should make this argument anyway: assert is not
> > for
> > optimization. In fact, we don't really want it to be used for
> > optimization, because if we do, then we might end up in a situation
> > where the -DNDEBUG build generates worse code than the build with
> > asserts enabled.
> 
> :)

> Also, I'll note that the reason that assume is an intrinsic, instead
> of a branch around unreachable, is that we aggressively remove
> branches around unreachable as part of our canonicalization process.
> We do this in order to simplify code, and this is important in order
> to remove abstraction penalties. Note that unreachables are also
> generated from undefined behavior, and one of the ways we use
> undefined behavior is to assume it is unreachable, enabling us to
> eliminate what should be dead code. This is an important technique
> for limiting abstraction penalties from, for example, heavily
> templated C++ code.

> Thus, somewhat unfortunately, Sanjoy's argument is not faulty.

> Asserts occur much more often than assumes, it may or may not be
> sensible to handle them the same way.
> I would argue it is sensible, but it's also reasonable to argue it is
> not.
> We need to be careful what we mean by "in the same way".

> Yes, i simply meant "extract information from the control flow
> structure and comparisons they generate when they are enabled".

> You are correct with all of your observations :)

> We can certainly improve the representations of assumes, perhaps as
> Danny has suggested by converting their control dependencies into
> extended SSA token vales, and better capture knowledge from
> conditional branches, but the tradeoffs here are not trivial.

> 100% agreed, this is something that requires some exploration and
> messing around, and then a design document going through the
> tradeoffs of various approaches with real data.

-- 

Hal Finkel 
Assistant Computational Scientist 
Leadership Computing Facility 
Argonne National Laboratory 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20160614/29839e41/attachment.html>