Use a more efficient lowering for Unordered/Monotonic atomic load/store on Thumb1.

llvm-svn: 139865

2 files changed, 43 insertions, 2 deletions

diff --git a/llvm/lib/Target/ARM/ARMISelLowering.cpp b/llvm/lib/Target/ARM/ARMISelLowering.cpp
index f6a63f58988..02189870f8a 100644
--- a/llvm/lib/Target/ARM/ARMISelLowering.cpp
+++ b/llvm/lib/Target/ARM/ARMISelLowering.cpp
@@ -639,8 +639,10 @@ ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
     setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
     setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
     setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
-    setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Expand);
-    setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Expand);
+    // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
+    // Unordered/Monotonic case.
+    setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
+    setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
     // Since the libcalls include locking, fold in the fences
     setShouldFoldAtomicFences(true);
   }
@@ -4861,6 +4863,18 @@ static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
                      Op.getOperand(1), Op.getOperand(2));
 }
 
+static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG,
+                                    const ARMSubtarget *ST) {
+  // Monotonic load/store is legal for all targets
+  if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
+    return Op;
+
+  // Aquire/Release load/store is not legal for targets without a
+  // dmb or equivalent available.
+  return SDValue();
+}
+
+
 static void
 ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results,
                     SelectionDAG &DAG, unsigned NewOp) {
@@ -4945,6 +4959,8 @@ SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
   case ISD::ADDE:
   case ISD::SUBC:
   case ISD::SUBE:          return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
+  case ISD::ATOMIC_LOAD:
+  case ISD::ATOMIC_STORE:  return LowerAtomicLoadStore(Op, DAG, Subtarget);
   }
   return SDValue();
 }
diff --git a/llvm/lib/Target/ARM/ARMInstrThumb.td b/llvm/lib/Target/ARM/ARMInstrThumb.td
index 43fd4ed8c4f..f8198c86946 100644
--- a/llvm/lib/Target/ARM/ARMInstrThumb.td
+++ b/llvm/lib/Target/ARM/ARMInstrThumb.td
@@ -1363,6 +1363,31 @@ def : T1Pat<(sextloadi16 t_addrmode_rrs2:$addr),
 def : T1Pat<(sextloadi16 t_addrmode_is2:$addr),
             (tASRri (tLSLri (tLDRHi t_addrmode_is2:$addr), 16), 16)>;
 
+def : T1Pat<(atomic_load_8 t_addrmode_is1:$src),
+             (tLDRBi t_addrmode_is1:$src)>;
+def : T1Pat<(atomic_load_8 t_addrmode_rrs1:$src),
+             (tLDRBr t_addrmode_rrs1:$src)>;
+def : T1Pat<(atomic_load_16 t_addrmode_is2:$src),
+             (tLDRHi t_addrmode_is2:$src)>;
+def : T1Pat<(atomic_load_16 t_addrmode_rrs2:$src),
+             (tLDRHr t_addrmode_rrs2:$src)>;
+def : T1Pat<(atomic_load_32 t_addrmode_is4:$src),
+             (tLDRi t_addrmode_is4:$src)>;
+def : T1Pat<(atomic_load_32 t_addrmode_rrs4:$src),
+             (tLDRr t_addrmode_rrs4:$src)>;
+def : T1Pat<(atomic_store_8 t_addrmode_is1:$ptr, tGPR:$val),
+             (tSTRBi tGPR:$val, t_addrmode_is1:$ptr)>;
+def : T1Pat<(atomic_store_8 t_addrmode_rrs1:$ptr, tGPR:$val),
+             (tSTRBr tGPR:$val, t_addrmode_rrs1:$ptr)>;
+def : T1Pat<(atomic_store_16 t_addrmode_is2:$ptr, tGPR:$val),
+             (tSTRHi tGPR:$val, t_addrmode_is2:$ptr)>;
+def : T1Pat<(atomic_store_16 t_addrmode_rrs2:$ptr, tGPR:$val),
+             (tSTRHr tGPR:$val, t_addrmode_rrs2:$ptr)>;
+def : T1Pat<(atomic_store_32 t_addrmode_is4:$ptr, tGPR:$val),
+             (tSTRi tGPR:$val, t_addrmode_is4:$ptr)>;
+def : T1Pat<(atomic_store_32 t_addrmode_rrs4:$ptr, tGPR:$val),
+             (tSTRr tGPR:$val, t_addrmode_rrs4:$ptr)>;
+
 // Large immediate handling.
 
 // Two piece imms.