beneficial optimization of undef examples needed

Project Infrastructure LLVM Dev List Archives
1

All,
     These discussions seem to be based on the premise that there is a
need for the compiler to exploit undefined behavior for performance
optimization reasons.

So far the only beneficial optimization I am aware of that relies on some
form of “undefined” is Dan Gohman’s original project for LP64 targets of
promoting i32 induction variables to i64 and hoisting sign-extension out
of the loop.

But “undef” / “poison” never appears in either the original or the transformed
IR for these types of loops, instead properties of “+nsw” are used to
justify the transformation. The transformation does not just fall out because
we’ve done a good job at defining “undef” / “poison” IR nodes.

So I’d like to see some concrete examples of where the compiler can
do useful optimization based on “undef” / “poison” appearing explicitly
In the IR, finding some would surely advance this discussion.

Peter Lawrence.

2

I'll repeat that open-ended requests would that end up generating lots of work for other people probably aren't going to get great results here.

John

3

Luckily someone already did the work writing a bunch of examples down:
http://blog.llvm.org/2011/05/what-every-c-programmer-should-know.html

And +1 for keeping this on-topic on how to implement poison.

  • Matthias
4

Hi Peter,

Why we need an undef value is covered here:
http://sunfishcode.github.io/blog/2014/07/14/undef-introduction.html
(in short, to do SSA construction well).

For a while we did not have a literal representation for poison.
However, in practice having both undef and poison was problematic (see
the paper), so we decided to ditch undef and keep poison.

However for the new poison to provide the same functionality as the
old undef (which is now going away), we need a literal representation
for the new poison.

-- Sanjoy

5

Sanjoy,
            You have changed the subject. We still need real world examples
showing how taking advantage of “undef” results in beneficial optimization.

My belief is that they don’t exist, my reasoning is this: real world programmers
are likely to run UBSan before compiling (or if they don’t they should), therefore
it is highly unlikely that any “undef” will actually exist during compilation of their
source code.

Yes, “Undef” can be created during SSA construction, but we already discussed
That in a previous email (its a register allocation problem).

The only other way I can see of “undef” occurring in IR is if we go out of our
way to optimize for example

  if ( ((i64)X + (i64)Y)) < INT_MIN || ((i64)X) + (i64)Y) > INT_MAX ) {
    . . .
    Z = X "+nsw” Y;
    . . .
  }

Here “Z” could in theory be replaced with “undef”, but there is no point in doing so.

Similarly with provably out-of-bounds GEP, similarly with provably invalid pointers,
etc, but again there is no point in doing so.

So is there any other way of having “undef” appear in the IR ?

Peter Lawrence.

6

Sanjoy,
You have changed the subject. We still need real world examples
showing how taking advantage of “undef” results in beneficial optimization.

My belief is that they don’t exist, my reasoning is this: real world programmers
are likely to run UBSan before compiling (or if they don’t they should), therefore
it is highly unlikely that any “undef” will actually exist during compilation of their
source code.

Wait - that ^ doesn’t seem to follow. The point of undefined behavior optimizations isn’t to only break programs that would trigger UBSan. The point is to optimize on the assumption that they won’t break UBSan/exhibit undefined behavior.

Yes, “Undef” can be created during SSA construction, but we already discussed
That in a previous email (its a register allocation problem).

The only other way I can see of “undef” occurring in IR is if we go out of our
way to optimize for example

if ( ((i64)X + (i64)Y)) < INT_MIN || ((i64)X) + (i64)Y) > INT_MAX ) {
. . .
Z = X "+nsw” Y;
. . .
}

Here “Z” could in theory be replaced with “undef”, but there is no point in doing so.

Similarly with provably out-of-bounds GEP, similarly with provably invalid pointers,
etc, but again there is no point in doing so.

Why isn’t there any point in doing so? One common situation where exploiting UB (aka: “assuming UB can’t happen”) may be beneficial is when the code in isolation is fine, but when inlined it’s possible to prove certain paths create contradictions. That way even if the conditions for those paths can’t be analyzed to prove they are unreachable (or values are not needed, etc), the compiler can skip all that and just collapse the code (remove the unreachable code, delete the value computation, etc) anyway.

7

David,
I’m not asking for the logic behind optimizing UB,
I’m asking for real world examples, a SPEC benchmark would do,
where optimizing UB actually proves to be beneficial.

I believe that in the real world it doesn’t happen, and I disagree
with the generalization of “constant-propagation after inlining is
beneficial” to “UB-propagation after inlining is beneficial”.

rather I believe that undef-propagation is the root cause of our problems.

But even if you disagree with my logic, we still don’t have any real world
examples showing the benefit of “UB-propagation after inlining".

Peter Lawrence.

8

Hi Peter,

            You have changed the subject. We still need real world examples

Not sure how -- I though I was pretty clear on how my answer was
related to you question.

showing how taking advantage of “undef” results in beneficial optimization.

My belief is that they don’t exist, my reasoning is this: real world programmers
are likely to run UBSan before compiling (or if they don’t they should), therefore
it is highly unlikely that any “undef” will actually exist during compilation of their
source code.

The canonical example motivating undef is this:

  int a;
  boil initialized;
  if (condition) {
    a = 5;
    initialized = true;
  }
  ...
  int result = 0;
  if (condition) {
    result += a;
  }

You can't optimize the PHI node that you'll get in side the block
guarded by 'condition' to 5 without something like undef.

Yes, “Undef” can be created during SSA construction, but we already discussed
That in a previous email (its a register allocation problem).

The only other way I can see of “undef” occurring in IR is if we go out of our
way to optimize for example

        if ( ((i64)X + (i64)Y)) < INT_MIN || ((i64)X) + (i64)Y) > INT_MAX ) {
                . . .
                Z = X "+nsw” Y;
                . . .
        }

Here “Z” could in theory be replaced with “undef”, but there is no point in doing so.

Similarly with provably out-of-bounds GEP, similarly with provably invalid pointers,
etc, but again there is no point in doing so.

These would be poison, not undef?

Just to re-iterate what I've already said before:

- If you're willing to have both undef and poison, then you don't
need a literal representation of poison to do the kind of
transformations poison allows (e.g. A s< (A +nsw 1)).
- However, having both undef and poison is a bad idea for several
reasons outlined in the paper.
- Since we want the new poison to "fill in" for the undef too, we
need a literal representation for the new poison to addresses cases
like the PHI node case above.

We don't care about optimizing code that actually has undefined
behavior -- we only care about optimizing code *under the assumption*
that it does not have UB. Both the old undef and the new poison have
been carefully designed to address this use case.

So is there any other way of having “undef” appear in the IR ?

Loading from uninitalized memory is another way you can have undef
appear in the IR today.

-- Sanjoy