From a03c5cd49047ae9815f5242794537e9b5cc0a075 Mon Sep 17 00:00:00 2001
From: Cody Tapscott <84105208+topolarity@users.noreply.github.com>
Date: Tue, 19 Mar 2024 04:20:16 -0400
Subject: [PATCH] Fix handling of virtual exit node in `PostDomTree` (#53739)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

This is an alternative to https://github.com/JuliaLang/julia/pull/53642

The `dom_edges()` for an exit block in the CFG are empty when computing
the PostDomTree so the loop below this may not actually run. In that
case, the right semidominator is the ancestor from the DFSTree, which is
the "virtual" -1 block.

This resolves half of the issue in
https://github.com/JuliaLang/julia/issues/53613:
```julia
julia> let code = Any[
               # block 1
               GotoIfNot(Argument(2), 3),
               # block 2
               ReturnNode(Argument(3)),
               # block 3 (we should visit this block)
               Expr(:call, throw, "potential throw"),
               ReturnNode(), # unreachable
           ]
           ir = make_ircode(code; slottypes=Any[Any,Bool,Bool])
           visited = BitSet()
           @test !Core.Compiler.visit_conditional_successors(CC.LazyPostDomtree(ir), ir, #=bb=#1) do succ::Int
               push!(visited, succ)
               return false
           end
           @test 2 ∈ visited
           @test 3 ∈ visited
       end
Test Passed
```

This needs some tests (esp. since I don't think we have any DomTree
tests at all right now), but otherwise should be good to go.
---
 base/compiler/ssair/domtree.jl | 5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)

diff --git a/base/compiler/ssair/domtree.jl b/base/compiler/ssair/domtree.jl
index 1edb8d2d5c6d4..2f22da04f9804 100644
--- a/base/compiler/ssair/domtree.jl
+++ b/base/compiler/ssair/domtree.jl
@@ -332,10 +332,7 @@ function SNCA!(domtree::GenericDomTree{IsPostDom}, blocks::Vector{BasicBlock}, m
     ancestors = copy(D.to_parent_pre)
     relevant_blocks = IsPostDom ? (1:max_pre) : (2:max_pre)
     for w::PreNumber in reverse(relevant_blocks)
-        # LLVM initializes this to the parent, the paper initializes this to
-        # `w`, but it doesn't really matter (the parent is a predecessor, so at
-        # worst we'll discover it below). Save a memory reference here.
-        semi_w = typemax(PreNumber)
+        semi_w = ancestors[w]
         last_linked = PreNumber(w + 1)
         for v ∈ dom_edges(domtree, blocks, D.from_pre[w])
             # For the purpose of the domtree, ignore virtual predecessors into