[llvm-dev] CFG normalization: avoiding `br i1 false`

[Anna Thomas via llvm-dev]

Hi Davide,

> On Nov 29, 2017, at 1:08 PM, Davide Italiano via llvm-dev <llvm-dev at lists.llvm.org> wrote:
> 
> On Wed, Nov 29, 2017 at 9:48 AM, Philip Reames
> <listmail at philipreames.com> wrote:
>> There's already a LoopSimplifyCFG which is a loop-pass (and thus can iterate
>> with other loop passes) and eliminates trivial branches.  Having a simple
>> pass which just strips unreachable blocks and converts conditional branches
>> to unconditional ones while updating appropriate analyzes (LoopInfo,
>> DomTree, LCSSA, etc..) seems very reasonable.
> 
> I'm not necessarily convinced having one-trick-pony pass is the best
> option here.
> It adds complexity, and there's already a pass in-tree which can
> provide this functionality.
> What are your concerns about running SimplifyCFG more often? It
> shouldn't be a compile-time sink
The main issue with running SimplifyCFG more often is if we were to place the pass 
*in between* a bunch of loop passes. Atleast in the Legacy pass manager, I think this breaks the nice
loop caching we have when placing all loop passes together.
I’ve not done any measurements on compile time to verify this theory.

LoopSimplifyCFG being a loop pass will not have this problem.

> 
>> This could also be a utility
>> function called from the end of other passes.  The hard bit is the analysis
>> preservation.  A good place to copy from might be the recent loop-unswitch
>> work Chandler did.
>> 
>> Philip
>> 
> 
> I don't think preserving the analyses is extremely hard (but I may be
> wrong on this).
> The incremental Dominator tree API makes the updates fairly easy.
> LCSSA is slightly more complicated.
> If you take a look at the new LoopUnswitch, in fact, it does call
> recalculate rather than fixing it incrementally.
> But, if LCSSA is computed right, recomputing entirely should take very
> little time.

The trouble (when I tried this transform in LoopSimplifyCFG) is updating 
loopInfo analysis: we can have cases where eliminating trivial branches in an inner loop 
will break the outer loop structure, and can cause it to no longer be a loop (consider a perfectly nested
inner loop within an outer loop, where the trivial branch is br i1 false, label %outerloopheader, label %innerloopHdr).

Another problem is if we changed the branches to unconditional ones, which then
makes some inner loop unreachable (i.e. no longer loops in LLVM terminology). 

We will need to record these kinds of loops and update the LPM. There maybe more here.

Thanks,
Anna
> 
> Thanks,
> 
> --
> Davide
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev