[flang-commits] [flang] [llvm] [mlir] [Flang][OpenMP] Add nsw flags to OMPIRBuilder loop IV arithmetic (PR #214165)

[Tom Eccles via flang-commits]

================
@@ -5807,9 +5807,18 @@ CanonicalLoopInfo *OpenMPIRBuilder::createLoopSkeleton(
   Builder.CreateBr(Latch);
 
   Builder.SetInsertPoint(Latch);
-  Value *Next = Builder.CreateAdd(IndVarPHI, ConstantInt::get(IndVarTy, 1),
-                                  "omp_" + Name + ".next", /*HasNUW=*/true,
-                                  /*HasNSW=*/Config.hasNoSignedWrap());
+  bool HasNSW = Config.hasNoSignedWrap();
+  if (HasNSW) {
+    if (auto *CI = dyn_cast<ConstantInt>(TripCount)) {
+      unsigned BitWidth = CI->getType()->getIntegerBitWidth();
+      APInt SignedMax = APInt::getSignedMaxValue(BitWidth);
+      if (CI->getValue().ugt(SignedMax))
+        HasNSW = false;
+    }
----------------
tblah wrote:

I agree. It is collapse which allows the counter to go above representable values in the loop iteration variable type. Outside of collapse, NSW should be safe (that case from 0 to INT_MAX is clever but I agree we can ignore it).

For collapse, disallowing NSW for dynamic bounds is the safe option. I just wanted to check that didn't break the case you are trying to optimise.

https://github.com/llvm/llvm-project/pull/214165