Add support for ARM ldrexd/strexd intrinsics. They both use i32 register pairs

to load/store i64 values. Since there's no current support to explicitly
declare such restrictions, implement it by using specific hardcoded register
pairs during isel.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@132248 91177308-0d34-0410-b5e6-96231b3b80d8

6 files changed, 182 insertions, 5 deletions

diff --git a/include/llvm/IntrinsicsARM.td b/include/llvm/IntrinsicsARM.td
index 3954412a76..fa8034e0c2 100644
--- a/include/llvm/IntrinsicsARM.td
+++ b/include/llvm/IntrinsicsARM.td
@@ -36,6 +36,16 @@ let TargetPrefix = "arm" in {  // All intrinsics start with "llvm.arm.".
 }
 
 //===----------------------------------------------------------------------===//
+// Load and Store exclusive doubleword
+
+let TargetPrefix = "arm" in {  // All intrinsics start with "llvm.arm.".
+  def int_arm_strexd : Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty,
+                                  llvm_ptr_ty], [IntrReadWriteArgMem]>;
+  def int_arm_ldrexd : Intrinsic<[llvm_i32_ty, llvm_i32_ty], [llvm_ptr_ty],
+                                 [IntrReadArgMem]>;
+}
+
+//===----------------------------------------------------------------------===//
 // VFP
 
 let TargetPrefix = "arm" in {  // All intrinsics start with "llvm.arm.".
diff --git a/lib/Target/ARM/ARMISelDAGToDAG.cpp b/lib/Target/ARM/ARMISelDAGToDAG.cpp
index abe5a316a4..9ad516dafb 100644
--- a/lib/Target/ARM/ARMISelDAGToDAG.cpp
+++ b/lib/Target/ARM/ARMISelDAGToDAG.cpp
@@ -2691,6 +2691,111 @@ SDNode *ARMDAGToDAGISel::Select(SDNode *N) {
     default:
       break;
 
+    case Intrinsic::arm_ldrexd: {
+      SDValue MemAddr = N->getOperand(2);
+      DebugLoc dl = N->getDebugLoc();
+      SDValue Chain = N->getOperand(0);
+
+      unsigned NewOpc = ARM::LDREXD;
+      if (Subtarget->isThumb() && Subtarget->hasThumb2())
+        NewOpc = ARM::t2LDREXD;
+
+      // arm_ldrexd returns a i64 value in {i32, i32}
+      std::vector<EVT> ResTys;
+      ResTys.push_back(MVT::i32);
+      ResTys.push_back(MVT::i32);
+      ResTys.push_back(MVT::Other);
+
+      // place arguments in the right order
+      SmallVector<SDValue, 7> Ops;
+      Ops.push_back(MemAddr);
+      Ops.push_back(getAL(CurDAG));
+      Ops.push_back(CurDAG->getRegister(0, MVT::i32));
+      Ops.push_back(Chain);
+      SDNode *Ld = CurDAG->getMachineNode(NewOpc, dl, ResTys, Ops.data(),
+                                          Ops.size());
+      // Transfer memoperands.
+      MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
+      MemOp[0] = cast<MemIntrinsicSDNode>(N)->getMemOperand();
+      cast<MachineSDNode>(Ld)->setMemRefs(MemOp, MemOp + 1);
+
+      // Until there's support for specifing explicit register constraints
+      // like the use of even/odd register pair, hardcode ldrexd to always
+      // use the pair [R0, R1] to hold the load result.
+      Chain = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, ARM::R0,
+                                   SDValue(Ld, 0), SDValue(0,0));
+      Chain = CurDAG->getCopyToReg(Chain, dl, ARM::R1,
+                                   SDValue(Ld, 1), Chain.getValue(1));
+
+      // Remap uses.
+      SDValue Glue = Chain.getValue(1);
+      if (!SDValue(N, 0).use_empty()) {
+        SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
+                                                ARM::R0, MVT::i32, Glue);
+        Glue = Result.getValue(2);
+        ReplaceUses(SDValue(N, 0), Result);
+      }
+      if (!SDValue(N, 1).use_empty()) {
+        SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
+                                                ARM::R1, MVT::i32, Glue);
+        Glue = Result.getValue(2);
+        ReplaceUses(SDValue(N, 1), Result);
+      }
+
+      ReplaceUses(SDValue(N, 2), SDValue(Ld, 2));
+      return NULL;
+    }
+
+    case Intrinsic::arm_strexd: {
+      DebugLoc dl = N->getDebugLoc();
+      SDValue Chain = N->getOperand(0);
+      SDValue Val0 = N->getOperand(2);
+      SDValue Val1 = N->getOperand(3);
+      SDValue MemAddr = N->getOperand(4);
+
+      // Until there's support for specifing explicit register constraints
+      // like the use of even/odd register pair, hardcode strexd to always
+      // use the pair [R2, R3] to hold the i64 (i32, i32) value to be stored.
+      Chain = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, ARM::R2, Val0,
+                                   SDValue(0, 0));
+      Chain = CurDAG->getCopyToReg(Chain, dl, ARM::R3, Val1, Chain.getValue(1));
+
+      SDValue Glue = Chain.getValue(1);
+      Val0 = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
+                                    ARM::R2, MVT::i32, Glue);
+      Glue = Val0.getValue(1);
+      Val1 = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
+                                    ARM::R3, MVT::i32, Glue);
+
+      // Store exclusive double return a i32 value which is the return status
+      // of the issued store.
+      std::vector<EVT> ResTys;
+      ResTys.push_back(MVT::i32);
+      ResTys.push_back(MVT::Other);
+
+      // place arguments in the right order
+      SmallVector<SDValue, 7> Ops;
+      Ops.push_back(Val0);
+      Ops.push_back(Val1);
+      Ops.push_back(MemAddr);
+      Ops.push_back(getAL(CurDAG));
+      Ops.push_back(CurDAG->getRegister(0, MVT::i32));
+      Ops.push_back(Chain);
+
+      unsigned NewOpc = ARM::STREXD;
+      if (Subtarget->isThumb() && Subtarget->hasThumb2())
+        NewOpc = ARM::t2STREXD;
+
+      SDNode *St = CurDAG->getMachineNode(NewOpc, dl, ResTys, Ops.data(),
+                                          Ops.size());
+      // Transfer memoperands.
+      MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
+      MemOp[0] = cast<MemIntrinsicSDNode>(N)->getMemOperand();
+      cast<MachineSDNode>(St)->setMemRefs(MemOp, MemOp + 1);
+
+      return St;
+    }
+
     case Intrinsic::arm_neon_vld1: {
       unsigned DOpcodes[] = { ARM::VLD1d8, ARM::VLD1d16,
                               ARM::VLD1d32, ARM::VLD1d64 };
diff --git a/lib/Target/ARM/ARMISelLowering.cpp b/lib/Target/ARM/ARMISelLowering.cpp
index f23d890d51..a78c3ed2a4 100644
--- a/lib/Target/ARM/ARMISelLowering.cpp
+++ b/lib/Target/ARM/ARMISelLowering.cpp
@@ -7658,6 +7658,28 @@ bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
     Info.writeMem = true;
     return true;
   }
+  case Intrinsic::arm_strexd: {
+    Info.opc = ISD::INTRINSIC_W_CHAIN;
+    Info.memVT = MVT::i64;
+    Info.ptrVal = I.getArgOperand(2);
+    Info.offset = 0;
+    Info.align = 8;
+    Info.vol = false;
+    Info.readMem = false;
+    Info.writeMem = true;
+    return true;
+  }
+  case Intrinsic::arm_ldrexd: {
+    Info.opc = ISD::INTRINSIC_W_CHAIN;
+    Info.memVT = MVT::i64;
+    Info.ptrVal = I.getArgOperand(0);
+    Info.offset = 0;
+    Info.align = 8;
+    Info.vol = false;
+    Info.readMem = true;
+    Info.writeMem = false;
+    return true;
+  }
   default:
     break;
   }
diff --git a/lib/Target/ARM/ARMInstrInfo.td b/lib/Target/ARM/ARMInstrInfo.td
index e9e11b0465..fd550219fe 100644
--- a/lib/Target/ARM/ARMInstrInfo.td
+++ b/lib/Target/ARM/ARMInstrInfo.td
@@ -3366,8 +3366,9 @@ def LDREXH : AIldrex<0b11, (outs GPR:$Rt), (ins addrmode7:$addr), NoItinerary,
                     "ldrexh", "\t$Rt, $addr", []>;
 def LDREX  : AIldrex<0b00, (outs GPR:$Rt), (ins addrmode7:$addr), NoItinerary,
                     "ldrex", "\t$Rt, $addr", []>;
-def LDREXD : AIldrex<0b01, (outs GPR:$Rt, GPR:$Rt2), (ins addrmode7:$addr),
-                    NoItinerary, "ldrexd", "\t$Rt, $Rt2, $addr", []>;
+let hasExtraDefRegAllocReq = 1 in
+  def LDREXD : AIldrex<0b01, (outs GPR:$Rt, GPR:$Rt2), (ins addrmode7:$addr),
+                      NoItinerary, "ldrexd", "\t$Rt, $Rt2, $addr", []>;
 }
 
 let mayStore = 1, Constraints = "@earlyclobber $Rd" in {
@@ -3377,10 +3378,12 @@ def STREXH : AIstrex<0b11, (outs GPR:$Rd), (ins GPR:$Rt, addrmode7:$addr),
                     NoItinerary, "strexh", "\t$Rd, $Rt, $addr", []>;
 def STREX  : AIstrex<0b00, (outs GPR:$Rd), (ins GPR:$Rt, addrmode7:$addr),
                     NoItinerary, "strex", "\t$Rd, $Rt, $addr", []>;
+}
+
+let hasExtraSrcRegAllocReq = 1, Constraints = "@earlyclobber $Rd" in
 def STREXD : AIstrex<0b01, (outs GPR:$Rd),
                     (ins GPR:$Rt, GPR:$Rt2, addrmode7:$addr),
                     NoItinerary, "strexd", "\t$Rd, $Rt, $Rt2, $addr", []>;
-}
 
 // Clear-Exclusive is for disassembly only.
 def CLREX : AXI<(outs), (ins), MiscFrm, NoItinerary, "clrex",
diff --git a/lib/Target/ARM/ARMInstrThumb2.td b/lib/Target/ARM/ARMInstrThumb2.td
index b8159f13a3..e86f2767ca 100644
--- a/lib/Target/ARM/ARMInstrThumb2.td
+++ b/lib/Target/ARM/ARMInstrThumb2.td
@@ -2881,7 +2881,9 @@ def t2LDREX  : Thumb2I<(outs rGPR:$Rt), (ins t2addrmode_reg:$addr), AddrModeNone
   let Inst{19-16} = addr;
   let Inst{15-12} = Rt;
 }
-def t2LDREXD : T2I_ldrex<0b11, (outs rGPR:$Rt, rGPR:$Rt2), (ins t2addrmode_reg:$addr),
+let hasExtraDefRegAllocReq = 1 in
+def t2LDREXD : T2I_ldrex<0b11, (outs rGPR:$Rt, rGPR:$Rt2),
+                         (ins t2addrmode_reg:$addr),
                          AddrModeNone, Size4Bytes, NoItinerary,
                          "ldrexd", "\t$Rt, $Rt2, $addr", "",
                          [], {?, ?, ?, ?}> {
@@ -2912,6 +2914,9 @@ def t2STREX  : Thumb2I<(outs rGPR:$Rd), (ins rGPR:$Rt, t2addrmode_reg:$addr),
   let Inst{19-16} = addr;
   let Inst{15-12} = Rt;
 }
+}
+
+let hasExtraSrcRegAllocReq = 1, Constraints = "@earlyclobber $Rd" in
 def t2STREXD : T2I_strex<0b11, (outs rGPR:$Rd),
                          (ins rGPR:$Rt, rGPR:$Rt2, t2addrmode_reg:$addr),
                          AddrModeNone, Size4Bytes, NoItinerary,
@@ -2920,7 +2925,6 @@ def t2STREXD : T2I_strex<0b11, (outs rGPR:$Rd),
   bits<4> Rt2;
   let Inst{11-8} = Rt2;
 }
-}
 
 // Clear-Exclusive is for disassembly only.
 def t2CLREX : T2XI<(outs), (ins), NoItinerary, "clrex",
diff --git a/test/CodeGen/ARM/ldstrexd.ll b/test/CodeGen/ARM/ldstrexd.ll
new file mode 100644
index 0000000000..0c0911a86e
--- /dev/null
+++ b/test/CodeGen/ARM/ldstrexd.ll
@@ -0,0 +1,33 @@
+; RUN: llc < %s -mtriple=armv7-apple-darwin   | FileCheck %s
+; RUN: llc < %s -mtriple=thumbv7-apple-darwin | FileCheck %s
+
+%0 = type { i32, i32 }
+
+; CHECK: f0:
+; CHECK: ldrexd
+define i64 @f0(i8* %p) nounwind readonly {
+entry:
+  %ldrexd = tail call %0 @llvm.arm.ldrexd(i8* %p)
+  %0 = extractvalue %0 %ldrexd, 1
+  %1 = extractvalue %0 %ldrexd, 0
+  %2 = zext i32 %0 to i64
+  %3 = zext i32 %1 to i64
+  %shl = shl nuw i64 %2, 32
+  %4 = or i64 %shl, %3
+  ret i64 %4
+}
+
+; CHECK: f1:
+; CHECK: strexd
+define i32 @f1(i8* %ptr, i64 %val) nounwind {
+entry:
+  %tmp4 = trunc i64 %val to i32
+  %tmp6 = lshr i64 %val, 32
+  %tmp7 = trunc i64 %tmp6 to i32
+  %strexd = tail call i32 @llvm.arm.strexd(i32 %tmp4, i32 %tmp7, i8* %ptr)
+  ret i32 %strexd
+}
+
+declare %0 @llvm.arm.ldrexd(i8*) nounwind readonly
+declare i32 @llvm.arm.strexd(i32, i32, i8*) nounwind
+