[llvm-dev] When can the dominator tree not contain a node for a basic block?

Mon Sep 21 18:11:22 PDT 2015

<pedantry>
FWIW, I've always understood "unreachable nodes are dominated by everything and dominate nothing but other unreachable nodes" to be the conventional interpretation of the formal definition, since universal quantification over an empty set is vacuously true; i.e. " of the 0 paths that exist, *all* of them go through any other node ".

> we claim … that *every* path from the entry node goes to the unreachable node
I don't think that follows; saying that d doesn't dominate n says that a path exists form start to n that doesn't go through d, but saying that d does dominate n just says something about all paths in a possibly empty set.

> most dominance algorithms will never consider an unreachable node to dominate anything
Agreed, but I think here we're talking about the inverse; whether anything dominates an unreachable node.
</pedantry>

As to the more important point, I've certainly seen dominance queries on unreachable code cause trouble before and agree there's just not a good answer.  I don't know if this would be acceptable for the API in LLVM, but you'd almost want dominates(d, n) to assert that n is reachable and force callers to decide what they want to do with unreachable code.
I agree with your point that confusion/problems arise when people expect pairwise dominance queries to equate to reachability in a tree and "extra" edges to unreachable nodes defeat their expectations.  I think what happens if you go all the way to the other extreme (and effectively use the predicate "d dominates n and n is reachable" in place of "d dominates n") is that sometimes a pass will be processing a fragment of the CFG (like the diamond of an if/else) and expect certain dominance properties on it (like that the head dominates both arms and the join point), which would then be reported as false if the whole fragment is unreachable.  Likewise, one would expect entry to dominate all nodes, etc.
I suppose that many unreachable code fragments have natural heads and you could figure out the "expected" answers for those, but cycle-breaking would still be arbitrary and I think usually the right answer is that the caller should be ignoring or pre-removing unreachable code.

Thanks
-Joseph

From: llvm-dev [mailto:llvm-dev-bounces at lists.llvm.org] On Behalf Of Daniel Berlin via llvm-dev
Sent: Monday, September 21, 2015 12:48 PM
To: Philip Reames <listmail at philipreames.com>
Cc: llvm-dev <llvm-dev at lists.llvm.org>
Subject: Re: [llvm-dev] When can the dominator tree not contain a node for a basic block?

So, there is no good answer.
It's not documented anywhere, and genericdomtree.h makes assumptions both ways.

For example, :
 void getDescendants(NodeT *R, SmallVectorImpl<NodeT *> &Result) const {
...

 if (!RN)
       return; // If R is unreachable, it will not be present in the DOM tree. :)
...
}
vs

dominates()
{
...
  // Any unreachable use is dominated, even if Def == User.
   if (!isReachableFromEntry(UseBB))
     return true;

   // Unreachable definitions don't dominate anything.
   if (!isReachableFromEntry(DefBB))
     return false;
}
...

From what I can tell, the split here is because we expect that nodes that *become unreachable* from entry will still be in the dom tree, but nodes that are completely unreachable generally don't end up in the domtree.  GetNode will work on the former, but not the latter.  This seems ... wrong.

This is *dominators*, mind you.
post-dominators goes through explicit machinations (and fails in some cases) when building to try to make sure even unreachable nodes end up in the dominator tree.

My suggestion would be that we clearly think about this, document and define what should happen, declare anything violating these invariants to be a bug, and go fix them.

As to what "the right answers are", this seems harder.

I've previously pointed out why giving dominance answers on unreachable nodes rarely, if ever, makes sense given the formal definition of a dominator:
 "a node<https://na01.safelinks.protection.outlook.com/?url=https%3a%2f%2fen.wikipedia.org%2fwiki%2fBasic_block&data=01%7c01%7cjotrem%40microsoft.com%7c0eaa02c7e0ca4450610008d2c2a47ef5%7c72f988bf86f141af91ab2d7cd011db47%7c1&sdata=ZIpgDcoJDbRUlp1xT%2bvQTEdSshYDO3cT2CLRsVDikB4%3d> d dominates a node n if every path from the entry node to n must go through d"

If n is unreachable, it is not dominated  by anything, since no path from the entry node to n exist, and of the 0 paths that exist, none of them go through any other node.  Essentially, right now we claim the opposite - that *every* path from the entry node goes to the unreachable node, and in fact, that it goes through *every* block.

This means, for example, that if you tried to reconstruct the dominator tree by calling dominates on pairs of instructions, you'd get wrong answers (you'd simultaneously try to make the unreachable block a child of every node in the dominator tree, as well as no node in the dominator tree)

In fact, you can show that most dominance algorithms will never consider an unreachable node to dominate anything, even going back to the Lowry and Medlock algorithm from 1969, or even the original Prosser description of a dominator on flow diagrams from 1959:
"We say box i dominates box j if every path (leading from input to output through the diagram) which passes through box j must also pass through box i. Thus box i dominates box j if box j is subordinate to box i in the program"

Now, whether we care about this or not, it's up in the air.

(Note that forward/reverse unreachable for dom/post-dom is different from "never exits", since never-exiting nodes are reachable)

On Mon, Sep 21, 2015 at 9:26 AM, Philip Reames <listmail at philipreames.com<mailto:listmail at philipreames.com>> wrote:
When looking into https://llvm.org/bugs/show_bug.cgi?id=24866<https://na01.safelinks.protection.outlook.com/?url=https%3a%2f%2fllvm.org%2fbugs%2fshow_bug.cgi%3fid%3d24866&data=01%7c01%7cjotrem%40microsoft.com%7c0eaa02c7e0ca4450610008d2c2a47ef5%7c72f988bf86f141af91ab2d7cd011db47%7c1&sdata=bNpncZVa7gkv5td%2bmmcfAkf3t%2bffFuZftsg6kJ2Ztpk%3d>, I discovered that the root cause of the crash is that I was expecting every basic block to have a corresponding Node in the dominator tree.  Apparently, the "while.end" basic block in the example does not have a Node in the Dominator Tree.  Can anyone tell me if this is expected?   If so, under what circumstances?

Interestingly, the crash in question appears to be fairly recently introduced.  A checkout from week before last didn't crash, whereas a checkout from this morning does.  I have yet tried to isolate the change in question; before doing so, I wanted to determine if this was a newly introduced problem (i.e. every BB should have a Node), or merely a newly exposed problem (i.e. not every BB does).

Philip

For reference, here's the reproduction command:
opt -S <input> -value-tracking-dom-conditions -licm -load-combine

And here's the IR:
; ModuleID = '../bugpoint-reduced-simplified.bc'
target datalayout = "e-m:w-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-pc-windows-msvc18.0.0"

%struct.c_derived_tbl.2.5.8.11.14.17.23.38.59.80.92.98.104.107.155.183 = type { [256 x i32], [256 x i8] }

; Function Attrs: nounwind uwtable
define void @encode_one_blockX2(%struct.c_derived_tbl.2.5.8.11.14.17.23.38.59.80.92.98.104.107.155.183* nocapture readonly %actbl) #0 {
entry:
  br i1 undef, label %L_KLOOP_01, label %L_KLOOP.preheader

L_KLOOP_01:                                       ; preds = %while.end, %entry
  br label %L_KLOOP.preheader

L_KLOOP_08:                                       ; preds = %while.end
  br label %L_KLOOP.preheader

L_KLOOP.preheader:                                ; preds = %L_KLOOP_08, %L_KLOOP_01, %entry
  %r.2.ph<https://na01.safelinks.protection.outlook.com/?url=http%3a%2f%2fr.2.ph&data=01%7c01%7cjotrem%40microsoft.com%7c0eaa02c7e0ca4450610008d2c2a47ef5%7c72f988bf86f141af91ab2d7cd011db47%7c1&sdata=jXpaIGmv6qhI7eaSNnoEvdGICfE53T553k6j7cGti94%3d> = phi i32 [ undef, %L_KLOOP_08 ], [ 0, %entry ], [ undef, %L_KLOOP_01 ]
  br label %L_KLOOP

L_KLOOP:                                          ; preds = %while.end, %L_KLOOP.preheader
  %r.2 = phi i32 [ 0, %while.end ], [ %r.2.ph<https://na01.safelinks.protection.outlook.com/?url=http%3a%2f%2fr.2.ph&data=01%7c01%7cjotrem%40microsoft.com%7c0eaa02c7e0ca4450610008d2c2a47ef5%7c72f988bf86f141af91ab2d7cd011db47%7c1&sdata=jXpaIGmv6qhI7eaSNnoEvdGICfE53T553k6j7cGti94%3d>, %L_KLOOP.preheader ]
  br i1 undef, label %while.body, label %while.end

while.body:                                       ; preds = %while.body, %L_KLOOP
  br label %while.body

while.end:                                        ; preds = %L_KLOOP
  %shl105 = shl i32 %r.2, 4
  %add106 = add nsw i32 %shl105, undef ;; <-- THIS is the context instruction.  V is undef
  %idxprom107 = sext i32 %add106 to i64
  %arrayidx108 = getelementptr inbounds %struct.c_derived_tbl.2.5.8.11.14.17.23.38.59.80.92.98.104.107.155.183, %struct.c_derived_tbl.2.5.8.11.14.17.23.38.59.80.92.98.104.107.155.183* %actbl, i64 0, i32 0, i64 %idxprom107
  %0 = load i32, i32* %arrayidx108, align 4, !tbaa !2
  %arrayidx110 = getelementptr inbounds %struct.c_derived_tbl.2.5.8.11.14.17.23.38.59.80.92.98.104.107.155.183, %struct.c_derived_tbl.2.5.8.11.14.17.23.38.59.80.92.98.104.107.155.183* %actbl, i64 0, i32 1, i64 %idxprom107
  %1 = load i8, i8* %arrayidx110, align 1, !tbaa !6
  indirectbr i8* undef, [label %L_KLOOP_DONE, label %L_KLOOP_01, label %L_KLOOP_08, label %L_KLOOP]

L_KLOOP_DONE:                                     ; preds = %while.end
  ret void
}

attributes #0 = { nounwind uwtable "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+sse,+sse2" "unsafe-fp-math"="false" "use-soft-float"="false" }

!llvm.module.flags = !{!0}
!llvm.ident = !{!1}

!0 = !{i32 1, !"PIC Level", i32 2}
!1 = !{!"clang version 3.8.0 (http://llvm.org/git/clang.git<https://na01.safelinks.protection.outlook.com/?url=http%3a%2f%2fllvm.org%2fgit%2fclang.git&data=01%7c01%7cjotrem%40microsoft.com%7c0eaa02c7e0ca4450610008d2c2a47ef5%7c72f988bf86f141af91ab2d7cd011db47%7c1&sdata=5BuQ3CnZsj0Hi7fOkJrCNa9e2n6FCRfyRiMaYYvQrYw%3d> 37c860b7d0d43dde1edcd324e00a10a62b2c5776) (http://llvm.org/git/llvm.git<https://na01.safelinks.protection.outlook.com/?url=http%3a%2f%2fllvm.org%2fgit%2fllvm.git&data=01%7c01%7cjotrem%40microsoft.com%7c0eaa02c7e0ca4450610008d2c2a47ef5%7c72f988bf86f141af91ab2d7cd011db47%7c1&sdata=wjkU7x%2bTSbgGd6ikqBsqKtOMmAYVUhHV4mtB6VqKMMc%3d> fb73c0262af4183555253a5c8c13025a6e7630bc)"}
!2 = !{!3, !3, i64 0}
!3 = !{!"int", !4, i64 0}
!4 = !{!"omnipotent char", !5, i64 0}
!5 = !{!"Simple C/C++ TBAA"}
!6 = !{!4, !4, i64 0}

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150922/e7f75a01/attachment.html>