[Mlir-commits] [mlir] [MLIR][SCF] Add canonicalization pattern to fold away iter args of scf.forall (PR #90189)

[llvmlistbot at llvm.org]

================
@@ -1509,6 +1510,203 @@ class ForallOpControlOperandsFolder : public OpRewritePattern<ForallOp> {
   }
 };
 
+/// The following canonicalization pattern folds the iter arguments of
+/// scf.forall op if :-
+/// 1. The corresponding result has zero uses.
+/// 2. The iter argument is NOT being modified within the loop body.
+/// uses.
+///
+/// Example of first case :-
+///  INPUT:
+///   %res:3 = scf.forall ... shared_outs(%arg0 = %a, %arg1 = %b, %arg2 = %c)
+///            {
+///                ...
+///                <SOME USE OF %arg0>
----------------
MaheshRavishankar wrote:

I think I mean both. If you update/use the bbArg within the body of the loop and in the terminator, that seems like an inherent race condition. If one thread is reading/writing data to bbarg or init value then it is inherently reading a non-deterministic state of the result being computed in the loop?

EDIT: Ok, I was mistaken. I guess you can use the bbArg in the body cause that represents the value "before the iteration". But really its only used for destination of operations. All other uses are invalid.

https://github.com/llvm/llvm-project/pull/90189