[llvm] [AMDGPU] Call the `FINI_ARRAY` destructors in the correct order (PR #71815)

[Joseph Huber via llvm-commits]

================
@@ -96,15 +105,39 @@ static void createInitOrFiniCalls(Function &F, bool IsCtor) {
   // for now we just call them with no arguments.
   auto *CallBackTy = FunctionType::get(IRB.getVoidTy(), {});
 
-  IRB.CreateCondBr(IRB.CreateICmpNE(Begin, End), LoopBB, ExitBB);
+  Constant *Start = Begin;
+  Constant *Stop = End;
+  // The destructor array must be called in reverse order. Get a constant
+  // expression to the end of the array and iterate backwards instead.
+  if (!IsCtor) {
+    Type *Int64Ty = IntegerType::getInt64Ty(C);
+    auto *Offset = ConstantExpr::getSub(
+        ConstantExpr::getAShr(
+            ConstantExpr::getSub(ConstantExpr::getPtrToInt(End, Int64Ty),
+                                 ConstantExpr::getPtrToInt(Begin, Int64Ty)),
+            ConstantInt::get(Int64Ty, 3)),
+        ConstantInt::get(Int64Ty, 1));
+    Start = ConstantExpr::getGetElementPtr(
+        ArrayType::get(IRB.getPtrTy(), 0), Begin,
+        ArrayRef<Constant *>({ConstantInt::get(Int64Ty, 0), Offset}),
+        /*InBounds=*/true);
+    Stop = Begin;
----------------
jhuber6 wrote:

The patch was functional and wasn't breaking any tests, if it's a matter of improving it would've been much easier to adjust it on top of this.

What is the issue with using ConstantExpr / ASHR here? There's no deprecation warnings or anything on the implementation. These can be constant expressions and we get the exact same code from `clang` if you compile it, see https://godbolt.org/z/Mzvhas451. So I'm confused as to what's wrong here. I guess we need to fix this in `clang` as well?

https://github.com/llvm/llvm-project/pull/71815