[llvm] [SROA] Rewrite invariant group intrinsics after splitting alloca (PR #107557)

[Antonio Frighetto via llvm-commits]

================
@@ -3531,8 +3531,19 @@ class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
       return true;
     }
 
-    if (II.isLaunderOrStripInvariantGroup())
+    if (II.isLaunderOrStripInvariantGroup()) {
+      Value *AdjustedPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
+      Value *New = nullptr;
+      if (II.getIntrinsicID() == Intrinsic::launder_invariant_group)
+        New = IRB.CreateLaunderInvariantGroup(AdjustedPtr);
+      else if (II.getIntrinsicID() == Intrinsic::strip_invariant_group)
+        New = IRB.CreateStripInvariantGroup(AdjustedPtr);
+
+      New->takeName(&II);
+      II.replaceAllUsesWith(New);
----------------
antoniofrighetto wrote:

Would you mind showing an example that could lead to get incorrect pointers? I had in mind something as follows, where we memcpy each field separately, but I suspect this is not what you were referring to:
```llvm
define void @memcpy_after_laundering_allocas_field(ptr %ptr) {
  %alloca = alloca { i64, i64 }, align 8
  %field1_ptr = getelementptr inbounds { i64, i64 }, ptr %alloca, i32 0, i32 0
  %field2_ptr = getelementptr inbounds { i64, i64 }, ptr %alloca, i32 0, i32 1
  %launder_field1 = call ptr @llvm.launder.invariant.group.p0(ptr %field1_ptr)
  %launder_field2 = call ptr @llvm.launder.invariant.group.p0(ptr %field2_ptr)
  call void @llvm.memcpy.p0.p0.i64(ptr %launder_field1, ptr %ptr, i64 8, i1 false)
  call void @llvm.memcpy.p0.p0.i64(ptr %launder_field2, ptr %ptr, i64 8, i1 false)
  ret void
}
```

https://github.com/llvm/llvm-project/pull/107557