diff --git a/llvm/include/llvm/Transforms/IPO/Attributor.h b/llvm/include/llvm/Transforms/IPO/Attributor.h
--- a/llvm/include/llvm/Transforms/IPO/Attributor.h
+++ b/llvm/include/llvm/Transforms/IPO/Attributor.h
@@ -5068,7 +5068,8 @@
     }
 
     /// Constant used to represent unknown offset or sizes.
-    static constexpr int64_t Unknown = 1 << 31;
+    static constexpr int64_t Unassigned = -1;
+    static constexpr int64_t Unknown = -2;
   };
 
   /// Call \p CB on all accesses that might interfere with \p OAS and return
diff --git a/llvm/lib/Transforms/IPO/AttributorAttributes.cpp b/llvm/lib/Transforms/IPO/AttributorAttributes.cpp
--- a/llvm/lib/Transforms/IPO/AttributorAttributes.cpp
+++ b/llvm/lib/Transforms/IPO/AttributorAttributes.cpp
@@ -1226,7 +1226,7 @@
 
   /// Helper struct, will support ranges eventually.
   struct OffsetInfo {
-    int64_t Offset = OffsetAndSize::Unknown;
+    int64_t Offset = OffsetAndSize::Unassigned;
 
     bool operator==(const OffsetInfo &OI) const { return Offset == OI.Offset; }
   };
@@ -1243,6 +1243,7 @@
 
     auto HandlePassthroughUser = [&](Value *Usr, OffsetInfo PtrOI,
                                      bool &Follow) {
+      assert(PtrOI.Offset != OffsetAndSize::Unassigned);
       OffsetInfo &UsrOI = OffsetInfoMap[Usr];
       UsrOI = PtrOI;
       Follow = true;
@@ -1283,11 +1284,15 @@
         APInt GEPOffset(DL.getIndexTypeSizeInBits(GEP->getType()), 0);
         if (PtrOI.Offset == OffsetAndSize::Unknown ||
             !GEP->accumulateConstantOffset(DL, GEPOffset)) {
+          LLVM_DEBUG(dbgs() << "[AAPointerInfo] GEP offset not constant "
+                            << *GEP << "\n");
           UsrOI.Offset = OffsetAndSize::Unknown;
           Follow = true;
           return true;
         }
 
+        LLVM_DEBUG(dbgs() << "[AAPointerInfo] GEP offset is constant " << *GEP
+                          << "\n");
         UsrOI.Offset = PtrOI.Offset + GEPOffset.getZExtValue();
         Follow = true;
         return true;
@@ -1306,15 +1311,22 @@
         bool IsFirstPHIUser = !OffsetInfoMap.count(Usr);
         OffsetInfo &UsrOI = OffsetInfoMap[Usr];
         OffsetInfo &PtrOI = OffsetInfoMap[CurPtr];
-        // Check if the PHI is invariant (so far).
-        if (UsrOI == PtrOI)
-          return true;
 
         // Check if the PHI operand has already an unknown offset as we can't
         // improve on that anymore.
         if (PtrOI.Offset == OffsetAndSize::Unknown) {
+          LLVM_DEBUG(dbgs() << "[AAPointerInfo] PHI operand offset unknown "
+                            << *CurPtr << " in " << *Usr << "\n");
+          Follow = UsrOI.Offset == OffsetAndSize::Unassigned ||
+                   UsrOI.Offset != OffsetAndSize::Unknown;
           UsrOI = PtrOI;
-          Follow = true;
+          return true;
+        }
+
+        // Check if the PHI is invariant (so far).
+        if (UsrOI == PtrOI) {
+          assert(PtrOI.Offset != OffsetAndSize::Unassigned);
+          LLVM_DEBUG(dbgs() << "[AAPointerInfo] PHI is invariant (so far)");
           return true;
         }
 
diff --git a/llvm/test/CodeGen/AMDGPU/attributor.ll b/llvm/test/CodeGen/AMDGPU/attributor.ll
new file mode 100644
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/attributor.ll
@@ -0,0 +1,42 @@
+; RUN: llc -o - < %s | FileCheck %s
+
+target triple = "amdgcn-amd-amdhsa"
+
+; The call to intrinsic implicitarg_ptr reaches a load through a phi. The
+; offsets of the phi cannot be determined, and hence the attirbutor assumes that
+; hostcall is in use.
+
+; CHECK: .value_kind:     hidden_hostcall_buffer
+; CHECK: .value_kind:     hidden_multigrid_sync_arg
+
+define amdgpu_kernel void @the_kernel(i32 addrspace(1)* %a, i64 %index1, i64 %index2, i1 %cond)  {
+entry:
+  %tmp7 = tail call i8 addrspace(4)* @llvm.amdgcn.implicitarg.ptr()
+  br i1 %cond, label %old, label %new
+
+old:                                              ; preds = %entry
+  %tmp4 = getelementptr i8, i8 addrspace(4)* %tmp7, i64 %index1
+  br label %join
+
+new:                                              ; preds = %entry
+  %tmp12 = getelementptr inbounds i8, i8 addrspace(4)* %tmp7, i64 %index2
+  br label %join
+
+join:                                             ; preds = %new, %old
+  %.in.in.in = phi i8 addrspace(4)* [ %tmp12, %new ], [ %tmp4, %old ]
+  %.in.in = bitcast i8 addrspace(4)* %.in.in.in to i16 addrspace(4)*
+
+  ;;; THIS USE is where the offset into implicitarg_ptr is unknown
+  %.in = load i16, i16 addrspace(4)* %.in.in, align 2
+
+  %idx.ext = sext i16 %.in to i64
+  %add.ptr3 = getelementptr inbounds i32, i32 addrspace(1)* %a, i64 %idx.ext
+  %tmp16 = atomicrmw add i32 addrspace(1)* %add.ptr3, i32 15 syncscope("agent-one-as") monotonic, align 4
+  ret void
+}
+
+declare i32 @llvm.amdgcn.workitem.id.x()
+
+declare align 4 i8 addrspace(4)* @llvm.amdgcn.implicitarg.ptr()
+
+declare i32 @llvm.amdgcn.workgroup.id.x()