[Mlir-commits] [mlir] [mlir][gpu] Fix crash in transform.gpu.map_nested_forall_to_threads with zero iterations (PR #170282)

[Fabian Mora]

================
@@ -141,12 +141,14 @@ std::optional<int64_t> getConstantIntValue(OpFoldResult ofr) {
 std::optional<SmallVector<int64_t>>
 getConstantIntValues(ArrayRef<OpFoldResult> ofrs) {
   bool failed = false;
-  SmallVector<int64_t> res = llvm::map_to_vector(ofrs, [&](OpFoldResult ofr) {
+  SmallVector<int64_t> res;
+  res.reserve(ofrs.size());
+  for (OpFoldResult ofr : ofrs) {
     auto cv = getConstantIntValue(ofr);
     if (!cv.has_value())
       failed = true;
-    return cv.value_or(0);
-  });
+    res.push_back(cv.value_or(0));
+  }
----------------
fabianmcg wrote:

That sounds like a C++ compiler bug, semantically they should produce equivalent results. However, let's assume the code first code is wrong. Let's keep the loop but early return once we now it's going to fail instead of waiting till the end.

https://github.com/llvm/llvm-project/pull/170282