[llvm] [IR] Avoid self-referencing values caused by PHI node removal (PR #129501)

[Robert Imschweiler via llvm-commits]

ro-i wrote:

Ok, I now created a third approach :) Let me first show the situation more in detail:
```
define i32 @constant_phi_leads_to_self_reference(ptr %ptr) {
  %A9 = alloca i1, align 1
  br label %F6
 
BB4:                                              ; No predecessors!
  unreachable
 
F1:                                               ; No predecessors!
  unreachable
 
T3:                                               ; preds = %T4, %BB6
  %L6 = phi i1 [ %L6.pr, %T4 ], [ %C4, %BB6 ]
  br label %BB5
 
BB5:                                              ; preds = %F7, %T3
  %L10 = load i1, ptr %A9, align 1
  br i1 %L10, label %BB6, label %F6
 
BB6:                                              ; preds = %BB5
  %LGV3 = load i1, ptr %ptr, align 1
  %C4 = icmp sle i1 %L6, true
  store i1 %C4, ptr %ptr, align 1
  br i1 %L6, label %F6, label %T3
 
T4:                                               ; No predecessors!
  %L6.pr = load i1, ptr %ptr, align 1
  br label %T3
 
F6:                                               ; preds = %0, %BB6, %BB5
  ret i32 0
 
F7:                                               ; No predecessors!
  br label %BB5
}
```
This is the state of the function before the problem occurs while handling T4. Why is T4 a problem? It's the pre-header of T3 (which is a loop header). L6.pr is used by L6. The pass now detects that T4 is unreachable and removes it. Consequently, L6 become `phi i1 [%C4, %BB6]`. At a later step, this constant phi node will then be simplified. As a result, every use of L6 will be replaced by C4, which leads to a self-reference in C4.
IMHO, there are 3 good approaches to handle this problem:
1. Avoid creating invalid IR at a general level. This was the first approach of my PR. This prevents replacing the phi node with its only value if that would lead to invalid IR.
2. Avoid handling unreachable blocks. This was the second approach I pushed. However, this would require some analysis (especially since DT should not be used during execution of the pass, as @nikic mentioned [here](https://github.com/llvm/llvm-project/issues/79175#issuecomment-1906813122). Since @arsenm felt that would be too much. I tried to figure out another way without trying to avoid working on blocks that became unreachable during the execution of the pass.
3. Avoid creating invalid IR at an early stage by avoiding deletion of loop pre-headers whenever they define instructions used by phi nodes in the loop header if removing these values from the phi nodes could lead to the phi node being removed and thus creating self-references (see the code for the details of the checks).

https://github.com/llvm/llvm-project/pull/129501