suoyuan666 wrote:
> I ran GLM 5.2 on your PR here's a few issues it found that I agree with.
>
> 1. There's a crash for this kind of code:
>
>
> ```c++
> void f() {
> int* p;
> while (*p) {
> int x;
> p = &x;
> }
> }
> ```
>
> I can't say I totally understand why, but I think it is related to the usage
> of that pointer in the loop condition and it getting reassigned to a local in
> the body that dies each iteration (perhaps we are missing a block?).
Thanks for the test case!
It looks like I should use `CFGBlock->preds()` to traverse blocks in reverse
order instead of replying on `PostOrderCFGView`. In the provided example, the
`UseFact` block (B4) comes before the `IssueFact`/`OriginFlowFact` block (B3)
in post-order, which explains the incorrect behavior in the current
implementation.
```txt
==========================================
Lifetime Analysis Facts:
==========================================
Function: f
Block B6:
End of Block
Block B5:
End of Block
Block B4:
Use (1 (Decl: p, Type : int *), Read)
Issue (0 (Path: p), ToOrigin: 0 (Expr: DeclRefExpr, Decl: p))
OriginFlow:
Dest: 2 (Expr: ImplicitCastExpr, Type : int *)
Src: 1 (Decl: p, Type : int *)
OriginFlow:
Dest: 3 (Expr: UnaryOperator, Type : int &)
Src: 2 (Expr: ImplicitCastExpr, Type : int *)
End of Block
Block B3:
Issue (1 (Path: x), ToOrigin: 4 (Expr: DeclRefExpr, Decl: x))
OriginFlow:
Dest: 5 (Expr: UnaryOperator, Type : int *)
Src: 4 (Expr: DeclRefExpr, Decl: x)
Use (1 (Decl: p, Type : int *), Write)
Issue (2 (Path: p), ToOrigin: 6 (Expr: DeclRefExpr, Decl: p))
Use (5 (Expr: UnaryOperator, Type : int *), Read)
OriginFlow:
Dest: 1 (Decl: p, Type : int *)
Src: 5 (Expr: UnaryOperator, Type : int *)
OriginFlow:
Dest: 7 (Expr: BinaryOperator, Type : int *&)
Src: 6 (Expr: DeclRefExpr, Decl: p)
OriginFlow:
Dest: 8 (Expr: BinaryOperator, Type : int *)
Src: 1 (Decl: p, Type : int *)
Expire (x)
End of Block
Block B2:
End of Block
Block B1:
Expire (p, Origin: 1 (Decl: p, Type : int *))
End of Block
Block B0:
End of Block
```
https://github.com/llvm/llvm-project/pull/204592
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
