[LLVM] New Dead Code Elimination

I have a rewrite of the aggressive dead code elimination pass which handles control flow and allows may-be-infinite loops to be removed under optional flag. Chandler suggested rather than a large change a series of small changes and, while that may not get us to small changes easily, there has been no further commentary on the diff: https://reviews.llvm.org/D18762. Given that, I will plan on proposing a series of diffs and would like to recruit some people who are interested in the change to help review those changes.

I expect to stage as follows:

  1. Rewrite base algorithm to introduce data structures needed to hold extra information (no functional change)

  2. Rewrite base algorithm to allow removing of control flow (under option, off by default). This variant will not be correct in the general case.

  3. Code to rewrite the control flow graph and remove dead basic blocks

  4. Identify Phi nodes with non-instruction reaching definitions and keep the associated source block live. Identify Phi nodes with multiple-live values from dead blocks

  5. Remove unreachable code discovered by post-order traversal to avoid.

(code is functionally correct at this point)

  1. Use post-order traversal to identify loop bottoms. By default this will be kept live but include switch to allow loops to be removed.

  2. Improve handling of live values out of dead regions

Please respond if you are willing to help or have suggestions on staging or approach.

Thanks

David

Hi David,

Thanks for working on this. I’d like to help reviewing your work!

How much code from the original patch are you leaving around? If you are reusing a lot of code from the original pass then you can refactor, clean and add new functionality. But if you are not keeping much then maybe it would be useful to develop a new pass (with small incremental reviewable changes), and when the work is done flip the switch and start using the new pass. What do you prefer?

I think that introducing the new flag that allows the removal of infinite loops is a great idea. LLVM is used for compilation of non-C languages, like Swift, Java, Python, etc, and these languages may benefit from the new functionality.

-Nadav

Hi David,

Feel free to add me as a reviewer (mehdi_amini on Phabricator). I’ll do my best to help with it.

Hi David,

Thanks for working on this. I’d like to help reviewing your work!

How much code from the original patch are you leaving around? If you are
reusing a lot of code from the original pass then you can refactor, clean
and add new functionality. But if you are not keeping much then maybe it
would be useful to develop a new pass (with small incremental reviewable
changes), and when the work is done flip the switch and start using the new
pass. What do you prefer?

I think that introducing the new flag that allows the removal of infinite
loops is a great idea. LLVM is used for compilation of non-C languages,
like Swift, Java, Python, etc, and these languages may benefit from the new
functionality.

LLVM doesn't handle "infinite loops are a side effect" in a principled way:
- The LLVM IR corresponding to an infinite loop doesn't contain side
effects which means that the loop itself doesn't have side effects. Calls
to functions which merely infinite loop can (and will) be removed by LLVM
(because they are readnone, nounwind, etc.).

A proper design for "infinite loops are a side effect" would likely be a
function attribute which indicates what behavior the optimizer may assume
when it sees a loop with unknown bound in a function.

While this doesn't mean that the pass shouldn't contain logic for leaving
infinite loops alone, it likely means that parameterizing the pass on a
flag is wrong.

In any case, I'd be willing to pitch in on these reviews.

Hi David,

Thanks for working on this. I’d like to help reviewing your work!

How much code from the original patch are you leaving around? If you are
reusing a lot of code from the original pass then you can refactor, clean
and add new functionality. But if you are not keeping much then maybe it
would be useful to develop a new pass (with small incremental reviewable
changes), and when the work is done flip the switch and start using the new
pass. What do you prefer?

I think that introducing the new flag that allows the removal of infinite
loops is a great idea. LLVM is used for compilation of non-C languages,
like Swift, Java, Python, etc, and these languages may benefit from the new
functionality.

LLVM doesn't handle "infinite loops are a side effect" in a principled way:
- The LLVM IR corresponding to an infinite loop doesn't contain side
effects which means that the loop itself doesn't have side effects. Calls
to functions which merely infinite loop can (and will) be removed by LLVM
(because they are readnone, nounwind, etc.).

A proper design for "infinite loops are a side effect" would likely be a
function attribute which indicates what behavior the optimizer may assume
when it sees a loop with unknown bound in a function.

True. FWIW, at least the previous versions of the pass are entirely based
on control flow.

That is, they are variants of the traditional control-dependence based DCE
removal. They will remove output-free infinite loops, regardless of bound.
They will not detect loops whose bound is infinite.

Essentially, it will remove any loop it is not necessary to go through to
get to an instruction that is not dead.

What kind of perf. uplifts are expected and in what benchmarks/apps ?

You should not expect pretty much any perf uplifts from any DCE, ever.
If it does, it’s mostly luck (better cache behavior, etc).

You should expect size improvements :wink:

I don’t think its true for all kinds of dead code elimination because there may be instructions which do execute but can be proven to be redundant/dead.

Lets say you do a bit-width analysis ( something along the lines of http://www.cs.cmu.edu/~seth/papers/budiu-tr00.pdf ). You may be able to remove some instructions and get runtime benefit ( which may have nothing to do with IC miss behavior because these were instructions that were getting executed ).

Anyway, the advanced analyses for DCE may not apply to this RFC. So we may not see any runtime benefits.

I don’t think its true for all kinds of dead code elimination

because there may be instructions which do execute but can be proven to be

redundant/dead.

This definition of dead code covers all redundant code (IE subsumes
redundancy elimination of literally all forms).
I'd agree that under this definition, you can improve runtime.

Lets say you do a bit-width analysis ( something along the lines of
http://www.cs.cmu.edu/~seth/papers/budiu-tr00.pdf ). You may be able to
remove some instructions and get runtime benefit ( which may have nothing
to do with IC miss behavior because these were instructions that were
getting executed ).

I personally would not class this as dead code elimination :slight_smile:

Anyway, the advanced analyses for DCE may not apply to this RFC. So we may
not see any runtime benefits.

It would not apply in all versions of this patch so far, so unless David
has other tricks up his sleeve :slight_smile: