[llvm-branch-commits] [clang] [flang] [flang][OpenMP] Map simple `do concurrent` loops to OpenMP host constructs (PR #127633)

[Sergio Afonso via llvm-branch-commits]

================
@@ -24,7 +25,82 @@ namespace flangomp {
 
 namespace {
 namespace looputils {
-using LoopNest = llvm::SetVector<fir::DoLoopOp>;
+/// Stores info needed about the induction/iteration variable for each `do
+/// concurrent` in a loop nest. This includes only for now:
+/// * the operation allocating memory for iteration variable,
+struct InductionVariableInfo {
+  mlir::Operation *iterVarMemDef;
+};
+
+using LoopNestToIndVarMap =
+    llvm::MapVector<fir::DoLoopOp, InductionVariableInfo>;
+
+/// Given an operation `op`, this returns true if one of `op`'s operands is
+/// "ultimately" the loop's induction variable. This helps in cases where the
+/// induction variable's use is "hidden" behind a convert/cast.
+///
+/// For example, give the following loop:
+/// ```
+///   fir.do_loop %ind_var = %lb to %ub step %s unordered {
+///     %ind_var_conv = fir.convert %ind_var : (index) -> i32
+///     fir.store %ind_var_conv to %i#1 : !fir.ref<i32>
+///     ...
+///   }
+/// ```
+///
+/// If \p op is the `fir.store` operation, then this function will return true
+/// since the IV is the "ultimate" opeerand to the `fir.store` op through the
----------------
skatrak wrote:

```suggestion
/// since the IV is the "ultimate" operand to the `fir.store` op through the
```

https://github.com/llvm/llvm-project/pull/127633