[PATCH] D130839: [DAG] FoldConstantArithmetic - add initial support for undef elements in bitcasted binop constant folding

[Xiang Zhang via Phabricator via llvm-commits]

xiangzhangllvm added inline comments.


================
Comment at: llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp:5589
+          if (!(IsUndef1 || IsUndef2))
+            Fold = FoldValue(Opcode, C1, C2);
+          else if (Opcode == ISD::AND || Opcode == ISD::MUL)
----------------
RKSimon wrote:
> xiangzhangllvm wrote:
> > RKSimon wrote:
> > > xiangzhangllvm wrote:
> > > > Do we need to make sure the C1 C2 has no other uses ?
> > > I don't think we have any one use limits on constant folding?
> > Yes, sorry my mistake at first.
> > take following case for example:
> > ```
> > A = and(x, undef)
> > C = A op B
> > ```
> > If we directly let "A = 0" (I think A=undef before), seems it may break the undef propagation in other use. for example C when "op" can do undef propagation (but not constant zero propagation).
> ```
> A = and(x, undef)
> C = A op B
> ```
> A can't (or shouldn't) propagate undef: https://alive2.llvm.org/ce/z/0EB60y - we almost certainly have cases of that in code, which need addressing
> 
Yes, thanks for your cases! it really not propagate.

But sorry for I am still some puzzle here:
Do you know why we can't (or shouldn't) propagate undef ? 

To me, I can't see the real difference between
```
 A = undef 
```
 and 
```
 A = and(x, undef)
```


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D130839/new/

https://reviews.llvm.org/D130839