AMDGPU: Re-visit nodes in performAndCombine

This fixes test regressions when i64 loads/stores are made promote.

llvm-svn: 267240
diff --git a/llvm/test/CodeGen/AMDGPU/shift-and-i64-ubfe.ll b/llvm/test/CodeGen/AMDGPU/shift-and-i64-ubfe.ll
index 29dae1b..f5ab732 100644
--- a/llvm/test/CodeGen/AMDGPU/shift-and-i64-ubfe.ll
+++ b/llvm/test/CodeGen/AMDGPU/shift-and-i64-ubfe.ll
@@ -299,9 +299,9 @@
 }
 
 ; GCN-LABEL: {{^}}and_not_mask_i64:
-; GCN: buffer_load_dwordx2 [[VAL:v\[[0-9]+:[0-9]+\]]]
-; GCN: v_lshr_b64 v{{\[}}[[SHRLO:[0-9]+]]:[[SHRHI:[0-9]+]]{{\]}}, [[VAL]], 20
-; GCN-DAG: v_and_b32_e32 v[[SHRLO]], 4, v[[SHRLO]]
+; GCN: buffer_load_dwordx2 v{{\[}}[[VALLO:[0-9]+]]:[[VALHI:[0-9]+]]{{\]}}
+; GCN: v_lshrrev_b32_e32 [[SHR:v[0-9]+]], 20, v[[VALLO]]
+; GCN-DAG: v_and_b32_e32 v[[SHRLO]], 4, [[SHR]]
 ; GCN-DAG: v_mov_b32_e32 v[[SHRHI]], 0{{$}}
 ; GCN-NOT: v[[SHRLO]]
 ; GCN-NOT: v[[SHRHI]]