[PATCH] D104870: [SimplifyCFG] Tail-merging all blocks with `unreachable` terminator

[Reid Kleckner via Phabricator via llvm-commits]

rnk added a subscriber: aeubanks.
rnk added a comment.

I think all my high level concerns are addressed then. I think it was worth letting all those folks know about this change. Users will also probably notice this change: it will affect the source location of lots of noreturn calls.



================
Comment at: llvm/test/Transforms/SimplifyCFG/tail-merge-noreturn.ll:129-130
 ; CHECK:       a2:
 ; CHECK-NEXT:    call void @assert_fail_1_alt(i32 0)
 ; CHECK-NEXT:    unreachable
 ;
----------------
lebedev.ri wrote:
> rnk wrote:
> > lebedev.ri wrote:
> > > rnk wrote:
> > > > I expected your code to fire on this test case. Can you explain why this example isn't getting tail merged?
> > > > 
> > > > Consider this example: https://gcc.godbolt.org/z/ox16a9P1z
> > > > ```
> > > > [[noreturn]] void abort1();
> > > > [[noreturn]] void abort2();
> > > > [[noreturn]] void abort3();
> > > > bool cond();
> > > > void doAsserts() {
> > > >     if (cond()) abort1();
> > > >     if (cond()) abort2();
> > > >     if (cond()) abort3();
> > > > }
> > > > ```
> > > > 
> > > > I think it is more canonical to leave these unreachable terminators in place after the calls to noreturn functions, rather than merging the unreachables together.
> > > > 
> > > > I just want to make sure your transform isn't firing, creating BBs, and then a later part of simplifycfg rolls the unreachables back up into place after the calls.
> > > > I expected your code to fire on this test case. Can you explain why this example isn't getting tail merged?
> > > 
> > > It fired, we didn't sink anything, and `SimplifyCFGOpt::simplifyUnreachable()` decided to undo it.
> > Got it, and we want to avoid that because otherwise it will make the overall pass return true to indicate that it changed something, which will make the parent pass manager re-run more passes.
> True.
> 
> As far as i'm aware that only results in potentially invalidating analysises,
> i'm not aware of that triggering another optimization pass runs.
> 
> IIRC that part of `SimplifyCFGOpt::simplifyUnreachable()` is a pretty important canonicalization,
> because e.g. instcombine can't modify cfg.
Even if it only invalidates analyses, I think this is worth addressing before landing this. Ideally this code would directly call the heuristic that "sink from common predecessors" uses, but if that isn't available, I think you could approximate it by not merging unreachable terminators when the previous non-debug instruction is a noreturn call with distinct callees. We know that is unprofitable, and that accounts for most blocks ending in unreachable. It saves compile time from IR churn too.

@aeubanks, what are the consequences of passes indicating that they changed the IR when they actually didn't?


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D104870/new/

https://reviews.llvm.org/D104870