Implement SPARCv9 atomic_swap_64 with a pseudo.

The SWAP instruction only exists in a 32-bit variant, but the 64-bit
atomic swap can be implemented in terms of CASX, like the other atomic
rmw primitives.

llvm-svn: 200453

3 files changed, 24 insertions, 3 deletions

diff --git a/llvm/lib/Target/Sparc/SparcISelLowering.cpp b/llvm/lib/Target/Sparc/SparcISelLowering.cpp
index 75b0167acd5..da5788d8801 100644
--- a/llvm/lib/Target/Sparc/SparcISelLowering.cpp
+++ b/llvm/lib/Target/Sparc/SparcISelLowering.cpp
@@ -1495,7 +1495,7 @@ SparcTargetLowering::SparcTargetLowering(TargetMachine &TM)
 
   if (Subtarget->is64Bit()) {
     setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Legal);
-    setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Expand);
+    setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Legal);
     setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Custom);
     setOperationAction(ISD::ATOMIC_STORE, MVT::i64, Custom);
   }
@@ -2874,6 +2874,9 @@ SparcTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
   case SP::ATOMIC_LOAD_NAND_64:
     return expandAtomicRMW(MI, BB, SP::ANDXrr);
 
+  case SP::ATOMIC_SWAP_64:
+    return expandAtomicRMW(MI, BB, 0);
+
   case SP::ATOMIC_LOAD_MAX_32:
     return expandAtomicRMW(MI, BB, SP::MOVICCrr, SPCC::ICC_G);
   case SP::ATOMIC_LOAD_MAX_64:
@@ -3012,7 +3015,8 @@ SparcTargetLowering::expandAtomicRMW(MachineInstr *MI,
 
   // Build the loop block.
   unsigned ValReg = MRI.createVirtualRegister(ValueRC);
-  unsigned UpdReg = MRI.createVirtualRegister(ValueRC);
+  // Opcode == 0 means try to write Rs2Reg directly (ATOMIC_SWAP).
+  unsigned UpdReg = (Opcode ? MRI.createVirtualRegister(ValueRC) : Rs2Reg);
 
   BuildMI(LoopMBB, DL, TII.get(SP::PHI), ValReg)
     .addReg(Val0Reg).addMBB(MBB)
@@ -3024,7 +3028,7 @@ SparcTargetLowering::expandAtomicRMW(MachineInstr *MI,
     BuildMI(LoopMBB, DL, TII.get(SP::CMPrr)).addReg(ValReg).addReg(Rs2Reg);
     BuildMI(LoopMBB, DL, TII.get(Opcode), UpdReg)
       .addReg(ValReg).addReg(Rs2Reg).addImm(CondCode);
-  } else {
+  } else if (Opcode) {
     BuildMI(LoopMBB, DL, TII.get(Opcode), UpdReg)
       .addReg(ValReg).addReg(Rs2Reg);
   }
diff --git a/llvm/lib/Target/Sparc/SparcInstr64Bit.td b/llvm/lib/Target/Sparc/SparcInstr64Bit.td
index 5f213342914..9146098a230 100644
--- a/llvm/lib/Target/Sparc/SparcInstr64Bit.td
+++ b/llvm/lib/Target/Sparc/SparcInstr64Bit.td
@@ -463,6 +463,14 @@ defm ATOMIC_LOAD_MAX  : AtomicRMW<atomic_load_max_32,  atomic_load_max_64>;
 defm ATOMIC_LOAD_UMIN : AtomicRMW<atomic_load_umin_32, atomic_load_umin_64>;
 defm ATOMIC_LOAD_UMAX : AtomicRMW<atomic_load_umax_32, atomic_load_umax_64>;
 
+// There is no 64-bit variant of SWAP, so use a pseudo.
+let usesCustomInserter = 1, hasCtrlDep = 1, mayLoad = 1, mayStore = 1,
+    Defs = [ICC], Predicates = [Is64Bit] in
+def ATOMIC_SWAP_64 : Pseudo<(outs I64Regs:$rd),
+                            (ins ptr_rc:$addr, I64Regs:$rs2), "",
+                            [(set i64:$rd,
+                                  (atomic_swap_64 iPTR:$addr, i64:$rs2))]>;
+
 // Global addresses, constant pool entries
 let Predicates = [Is64Bit] in {
 
diff --git a/llvm/test/CodeGen/SPARC/atomics.ll b/llvm/test/CodeGen/SPARC/atomics.ll
index 6d2b48287ca..b10336c9808 100644
--- a/llvm/test/CodeGen/SPARC/atomics.ll
+++ b/llvm/test/CodeGen/SPARC/atomics.ll
@@ -62,6 +62,15 @@ entry:
   ret i32 %b
 }
 
+; CHECK-LABEL: test_swap_i64
+; CHECK:       casx [%o1],
+
+define i64 @test_swap_i64(i64 %a, i64* %ptr) {
+entry:
+  %b = atomicrmw xchg i64* %ptr, i64 42 monotonic
+  ret i64 %b
+}
+
 ; CHECK-LABEL: test_load_add_32
 ; CHECK: membar
 ; CHECK: add [[V:%[gilo][0-7]]], %o1, [[U:%[gilo][0-7]]]