[clang] [llvm] [LoopUnroll] Add flag to enforce loop unroll pragma regardless of expensive trip count (PR #180961)

[Adel Ejjeh via cfe-commits]

================
@@ -1102,7 +1110,8 @@ bool llvm::computeUnrollCount(
   }
   if (UP.Count == 0)
     UP.Count = UP.DefaultUnrollRuntimeCount;
-
+  if (PragmaEnableUnroll && ForcePragmaUnroll)
+    UP.AllowExpensiveTripCount = true;
----------------
adelejjeh wrote:

Agreed. And this is the longer-term intent with bullet (1) in my PR description. However, we don't want to change the default behavior of the compiler without further testing of how that will affect users. this will also have significant ramifications on other backends as well. This is why this PR is intended as a short-term stopgap to unblock some AMDGPU users. And overall, an opt-in flag will not cause issues for other targets by changing the default compiler behavior.

https://github.com/llvm/llvm-project/pull/180961