[ARM] Modify codegen for memcpy intrinsic to prefer LDM/STM.

We were previously codegen'ing memcpy as regular load/store operations and hoping that the register allocator would allocate registers in ascending order so that we could apply an LDM/STM combine after register allocation. According to the commit that first introduced this code (r37179), we planned to teach the register allocator to allocate the registers in ascending order. This never got implemented, and up to now we've been stuck with very poor codegen. A much simpler approach for achieving better codegen is to create MEMCPY pseudo instructions, attach scratch virtual registers to them and then, post register allocation, expand the MEMCPYs into LDM/STM pairs using the scratch registers. The register allocator will have picked arbitrary registers which we sort when expanding the MEMCPY. This approach also avoids the need to repeatedly calculate offsets which ultimately ought to be eliminated pre-RA in order to decrease register pressure. Fixes PR9199 and PR23768. [This is based on Peter Collingbourne's r238473 which was reverted.] Differential Revision: http://reviews.llvm.org/D13239 Change-Id: I727543c2e94136e0f80b8e22d5642d7b9ee5b458 Author: Peter Collingbourne <[email protected]> llvm-svn: 249322
llvm · Oct 5, 2015 · 953f908 · 953f908
1 parent 3fd6c11
commit 953f908
Show file tree

Hide file tree

Showing 14 changed files with 520 additions and 73 deletions.
diff --git a/llvm/lib/Target/ARM/ARMBaseInstrInfo.cpp b/llvm/lib/Target/ARM/ARMBaseInstrInfo.cpp
@@ -1219,6 +1219,60 @@ unsigned ARMBaseInstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI,
   return MI->mayLoad() && hasLoadFromStackSlot(MI, Dummy, FrameIndex);
 }
 
+/// \brief Expands MEMCPY to either LDMIA/STMIA or LDMIA_UPD/STMID_UPD
+/// depending on whether the result is used.
+void ARMBaseInstrInfo::expandMEMCPY(MachineBasicBlock::iterator MBBI) const {
+  bool isThumb1 = Subtarget.isThumb1Only();
+  bool isThumb2 = Subtarget.isThumb2();
+  const ARMBaseInstrInfo *TII = Subtarget.getInstrInfo();
+
+  MachineInstr *MI = MBBI;
+  DebugLoc dl = MI->getDebugLoc();
+  MachineBasicBlock *BB = MI->getParent();
+
+  MachineInstrBuilder LDM, STM;
+  if (isThumb1 || !MI->getOperand(1).isDead()) {
+    LDM = BuildMI(*BB, MI, dl, TII->get(isThumb2 ? ARM::t2LDMIA_UPD
+                                                 : isThumb1 ? ARM::tLDMIA_UPD
+                                                            : ARM::LDMIA_UPD))
+             .addOperand(MI->getOperand(1));
+  } else {
+    LDM = BuildMI(*BB, MI, dl, TII->get(isThumb2 ? ARM::t2LDMIA : ARM::LDMIA));
+  }
+
+  if (isThumb1 || !MI->getOperand(0).isDead()) {
+    STM = BuildMI(*BB, MI, dl, TII->get(isThumb2 ? ARM::t2STMIA_UPD
+                                                 : isThumb1 ? ARM::tSTMIA_UPD
+                                                            : ARM::STMIA_UPD))
+             .addOperand(MI->getOperand(0));
+  } else {
+    STM = BuildMI(*BB, MI, dl, TII->get(isThumb2 ? ARM::t2STMIA : ARM::STMIA));
+  }
+
+  AddDefaultPred(LDM.addOperand(MI->getOperand(3)));
+  AddDefaultPred(STM.addOperand(MI->getOperand(2)));
+
+  // Sort the scratch registers into ascending order.
+  const TargetRegisterInfo &TRI = getRegisterInfo();
+  llvm::SmallVector<unsigned, 6> ScratchRegs;
+  for(unsigned I = 5; I < MI->getNumOperands(); ++I)
+    ScratchRegs.push_back(MI->getOperand(I).getReg());
+  std::sort(ScratchRegs.begin(), ScratchRegs.end(),
+            [&TRI](const unsigned &Reg1,
+                   const unsigned &Reg2) -> bool {
+              return TRI.getEncodingValue(Reg1) <
+                     TRI.getEncodingValue(Reg2);
+            });
+
+  for (const auto &Reg : ScratchRegs) {
+    LDM.addReg(Reg, RegState::Define);
+    STM.addReg(Reg, RegState::Kill);
+  }
+
+  BB->erase(MBBI);
+}
+
+
 bool
 ARMBaseInstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
   MachineFunction &MF = *MI->getParent()->getParent();
@@ -1232,6 +1286,11 @@ ARMBaseInstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
     return true;
   }
 
+  if (MI->getOpcode() == ARM::MEMCPY) {
+    expandMEMCPY(MI);
+    return true;
+  }
+
   // This hook gets to expand COPY instructions before they become
   // copyPhysReg() calls.  Look for VMOVS instructions that can legally be
   // widened to VMOVD.  We prefer the VMOVD when possible because it may be

diff --git a/llvm/lib/Target/ARM/ARMBaseInstrInfo.h b/llvm/lib/Target/ARM/ARMBaseInstrInfo.h
@@ -352,6 +352,8 @@ class ARMBaseInstrInfo : public ARMGenInstrInfo {
   virtual void expandLoadStackGuard(MachineBasicBlock::iterator MI,
                                     Reloc::Model RM) const = 0;
 
+  void expandMEMCPY(MachineBasicBlock::iterator) const;
+
 private:
   /// Modeling special VFP / NEON fp MLA / MLS hazards.
 

diff --git a/llvm/lib/Target/ARM/ARMISelLowering.cpp b/llvm/lib/Target/ARM/ARMISelLowering.cpp
@@ -1178,6 +1178,7 @@ const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
   case ARMISD::VORRIMM:       return "ARMISD::VORRIMM";
   case ARMISD::VBICIMM:       return "ARMISD::VBICIMM";
   case ARMISD::VBSL:          return "ARMISD::VBSL";
+  case ARMISD::MEMCPY:        return "ARMISD::MEMCPY";
   case ARMISD::VLD2DUP:       return "ARMISD::VLD2DUP";
   case ARMISD::VLD3DUP:       return "ARMISD::VLD3DUP";
   case ARMISD::VLD4DUP:       return "ARMISD::VLD4DUP";
@@ -8072,8 +8073,41 @@ ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
   }
 }
 
+/// \brief Attaches vregs to MEMCPY that it will use as scratch registers
+/// when it is expanded into LDM/STM. This is done as a post-isel lowering
+/// instead of as a custom inserter because we need the use list from the SDNode.
+static void attachMEMCPYScratchRegs(const ARMSubtarget *Subtarget,
+                                   MachineInstr *MI, const SDNode *Node) {
+  bool isThumb1 = Subtarget->isThumb1Only();
+
+  DebugLoc DL = MI->getDebugLoc();
+  MachineFunction *MF = MI->getParent()->getParent();
+  MachineRegisterInfo &MRI = MF->getRegInfo();
+  MachineInstrBuilder MIB(*MF, MI);
+
+  // If the new dst/src is unused mark it as dead.
+  if (!Node->hasAnyUseOfValue(0)) {
+    MI->getOperand(0).setIsDead(true);
+  }
+  if (!Node->hasAnyUseOfValue(1)) {
+    MI->getOperand(1).setIsDead(true);
+  }
+
+  // The MEMCPY both defines and kills the scratch registers.
+  for (unsigned I = 0; I != MI->getOperand(4).getImm(); ++I) {
+    unsigned TmpReg = MRI.createVirtualRegister(isThumb1 ? &ARM::tGPRRegClass
+                                                         : &ARM::GPRRegClass);
+    MIB.addReg(TmpReg, RegState::Define|RegState::Dead);
+  }
+}
+
 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
                                                       SDNode *Node) const {
+  if (MI->getOpcode() == ARM::MEMCPY) {
+    attachMEMCPYScratchRegs(Subtarget, MI, Node);
+    return;
+  }
+
   const MCInstrDesc *MCID = &MI->getDesc();
   // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
   // RSC. Coming out of isel, they have an implicit CPSR def, but the optional

diff --git a/llvm/lib/Target/ARM/ARMISelLowering.h b/llvm/lib/Target/ARM/ARMISelLowering.h
@@ -185,6 +185,10 @@ namespace llvm {
       // Vector bitwise select
       VBSL,
 
+      // Pseudo-instruction representing a memory copy using ldm/stm
+      // instructions.
+      MEMCPY,
+
       // Vector load N-element structure to all lanes:
       VLD2DUP = ISD::FIRST_TARGET_MEMORY_OPCODE,
       VLD3DUP,

diff --git a/llvm/lib/Target/ARM/ARMInstrInfo.td b/llvm/lib/Target/ARM/ARMInstrInfo.td
@@ -73,6 +73,10 @@ def SDT_ARMBFI : SDTypeProfile<1, 3, [SDTCisVT<0, i32>, SDTCisVT<1, i32>,
 
 def SDT_WIN__DBZCHK : SDTypeProfile<0, 1, [SDTCisVT<0, i32>]>;
 
+def SDT_ARMMEMCPY  : SDTypeProfile<2, 3, [SDTCisVT<0, i32>, SDTCisVT<1, i32>,
+                                          SDTCisVT<2, i32>, SDTCisVT<3, i32>,
+                                          SDTCisVT<4, i32>]>;
+
 def SDTBinaryArithWithFlags : SDTypeProfile<2, 2,
                                             [SDTCisSameAs<0, 2>,
                                              SDTCisSameAs<0, 3>,
@@ -179,6 +183,10 @@ def ARMtcret         : SDNode<"ARMISD::TC_RETURN", SDT_ARMTCRET,
 
 def ARMbfi           : SDNode<"ARMISD::BFI", SDT_ARMBFI>;
 
+def ARMmemcopy : SDNode<"ARMISD::MEMCPY", SDT_ARMMEMCPY,
+                        [SDNPHasChain, SDNPInGlue, SDNPOutGlue,
+                         SDNPMayStore, SDNPMayLoad]>;
+
 //===----------------------------------------------------------------------===//
 // ARM Instruction Predicate Definitions.
 //
@@ -4577,6 +4585,19 @@ let usesCustomInserter = 1 in {
       [(ARMcopystructbyval GPR:$dst, GPR:$src, imm:$size, imm:$alignment)]>;
 }
 
+let hasPostISelHook = 1, Constraints = "$newdst = $dst, $newsrc = $src" in {
+    // %newsrc, %newdst = MEMCPY %dst, %src, N, ...N scratch regs...
+    // Copies N registers worth of memory from address %src to address %dst
+    // and returns the incremented addresses.  N scratch register will
+    // be attached for the copy to use.
+    def MEMCPY : PseudoInst<
+      (outs GPR:$newdst, GPR:$newsrc),
+      (ins GPR:$dst, GPR:$src, i32imm:$nreg, variable_ops),
+      NoItinerary,
+      [(set GPR:$newdst, GPR:$newsrc,
+            (ARMmemcopy GPR:$dst, GPR:$src, imm:$nreg))]>;
+}
+
 def ldrex_1 : PatFrag<(ops node:$ptr), (int_arm_ldrex node:$ptr), [{
   return cast<MemIntrinsicSDNode>(N)->getMemoryVT() == MVT::i8;
 }]>;

diff --git a/llvm/lib/Target/ARM/ARMInstrThumb.td b/llvm/lib/Target/ARM/ARMInstrThumb.td
@@ -740,6 +740,7 @@ def tLDMIA : T1I<(outs), (ins tGPR:$Rn, pred:$p, reglist:$regs, variable_ops),
 // Writeback version is just a pseudo, as there's no encoding difference.
 // Writeback happens iff the base register is not in the destination register
 // list.
+let mayLoad = 1, hasExtraDefRegAllocReq = 1 in
 def tLDMIA_UPD :
     InstTemplate<AddrModeNone, 0, IndexModeNone, Pseudo, GenericDomain,
                  "$Rn = $wb", IIC_iLoad_mu>,

diff --git a/llvm/lib/Target/ARM/ARMSelectionDAGInfo.cpp b/llvm/lib/Target/ARM/ARMSelectionDAGInfo.cpp
@@ -160,41 +160,39 @@ ARMSelectionDAGInfo::EmitTargetCodeForMemcpy(SelectionDAG &DAG, SDLoc dl,
   unsigned VTSize = 4;
   unsigned i = 0;
   // Emit a maximum of 4 loads in Thumb1 since we have fewer registers
-  const unsigned MAX_LOADS_IN_LDM = Subtarget.isThumb1Only() ? 4 : 6;
+  const unsigned MaxLoadsInLDM = Subtarget.isThumb1Only() ? 4 : 6;
   SDValue TFOps[6];
   SDValue Loads[6];
   uint64_t SrcOff = 0, DstOff = 0;
 
-  // Emit up to MAX_LOADS_IN_LDM loads, then a TokenFactor barrier, then the
-  // same number of stores.  The loads and stores will get combined into
-  // ldm/stm later on.
-  while (EmittedNumMemOps < NumMemOps) {
-    for (i = 0;
-         i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
-      Loads[i] = DAG.getLoad(VT, dl, Chain,
-                             DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
-                                         DAG.getConstant(SrcOff, dl, MVT::i32)),
-                             SrcPtrInfo.getWithOffset(SrcOff), isVolatile,
-                             false, false, 0);
-      TFOps[i] = Loads[i].getValue(1);
-      SrcOff += VTSize;
-    }
-    Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
-                        makeArrayRef(TFOps, i));
+  // FIXME: We should invent a VMEMCPY pseudo-instruction that lowers to
+  // VLDM/VSTM and make this code emit it when appropriate. This would reduce
+  // pressure on the general purpose registers. However this seems harder to map
+  // onto the register allocator's view of the world.
 
-    for (i = 0;
-         i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
-      TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
-                              DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
-                                          DAG.getConstant(DstOff, dl, MVT::i32)),
-                              DstPtrInfo.getWithOffset(DstOff),
-                              isVolatile, false, 0);
-      DstOff += VTSize;
-    }
-    Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
-                        makeArrayRef(TFOps, i));
+  // The number of MEMCPY pseudo-instructions to emit. We use up to
+  // MaxLoadsInLDM registers per mcopy, which will get lowered into ldm/stm
+  // later on. This is a lower bound on the number of MEMCPY operations we must
+  // emit.
+  unsigned NumMEMCPYs = (NumMemOps + MaxLoadsInLDM - 1) / MaxLoadsInLDM;
+
+  SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other, MVT::Glue);
+
+  for (unsigned I = 0; I != NumMEMCPYs; ++I) {
+    // Evenly distribute registers among MEMCPY operations to reduce register
+    // pressure.
+    unsigned NextEmittedNumMemOps = NumMemOps * (I + 1) / NumMEMCPYs;
+    unsigned NumRegs = NextEmittedNumMemOps - EmittedNumMemOps;
+
+    Dst = DAG.getNode(ARMISD::MEMCPY, dl, VTs, Chain, Dst, Src,
+                      DAG.getConstant(NumRegs, dl, MVT::i32));
+    Src = Dst.getValue(1);
+    Chain = Dst.getValue(2);
+
+    DstPtrInfo = DstPtrInfo.getWithOffset(NumRegs * VTSize);
+    SrcPtrInfo = SrcPtrInfo.getWithOffset(NumRegs * VTSize);
 
-    EmittedNumMemOps += i;
+    EmittedNumMemOps = NextEmittedNumMemOps;
   }
 
   if (BytesLeft == 0)

diff --git a/llvm/lib/Target/ARM/Thumb2SizeReduction.cpp b/llvm/lib/Target/ARM/Thumb2SizeReduction.cpp
@@ -125,7 +125,10 @@ namespace {
   { ARM::t2LDMIA, ARM::tLDMIA,  0,             0,   0,   1,   1,  1,1, 0,1,0 },
   { ARM::t2LDMIA_RET,0,         ARM::tPOP_RET, 0,   0,   1,   1,  1,1, 0,1,0 },
   { ARM::t2LDMIA_UPD,ARM::tLDMIA_UPD,ARM::tPOP,0,   0,   1,   1,  1,1, 0,1,0 },
-  // ARM::t2STM (with no basereg writeback) has no Thumb1 equivalent
+  // ARM::t2STMIA (with no basereg writeback) has no Thumb1 equivalent.
+  // tSTMIA_UPD is a change in semantics which can only be used if the base
+  // register is killed. This difference is correctly handled elsewhere.
+  { ARM::t2STMIA, ARM::tSTMIA_UPD, 0,          0,   0,   1,   1,  1,1, 0,1,0 },
   { ARM::t2STMIA_UPD,ARM::tSTMIA_UPD, 0,       0,   0,   1,   1,  1,1, 0,1,0 },
   { ARM::t2STMDB_UPD, 0,        ARM::tPUSH,    0,   0,   1,   1,  1,1, 0,1,0 }
   };
@@ -435,6 +438,14 @@ Thumb2SizeReduce::ReduceLoadStore(MachineBasicBlock &MBB, MachineInstr *MI,
     isLdStMul = true;
     break;
   }
+  case ARM::t2STMIA: {
+    // If the base register is killed, we don't care what its value is after the
+    // instruction, so we can use an updating STMIA.
+    if (!MI->getOperand(0).isKill())
+      return false;
+
+    break;
+  }
   case ARM::t2LDMIA_RET: {
     unsigned BaseReg = MI->getOperand(1).getReg();
     if (BaseReg != ARM::SP)
@@ -492,6 +503,12 @@ Thumb2SizeReduce::ReduceLoadStore(MachineBasicBlock &MBB, MachineInstr *MI,
   // Add the 16-bit load / store instruction.
   DebugLoc dl = MI->getDebugLoc();
   MachineInstrBuilder MIB = BuildMI(MBB, MI, dl, TII->get(Opc));
+
+  // tSTMIA_UPD takes a defining register operand. We've already checked that
+  // the register is killed, so mark it as dead here.
+  if (Entry.WideOpc == ARM::t2STMIA)
+    MIB.addReg(MI->getOperand(0).getReg(), RegState::Define | RegState::Dead);
+
   if (!isLdStMul) {
     MIB.addOperand(MI->getOperand(0));
     MIB.addOperand(MI->getOperand(1));

diff --git a/llvm/test/CodeGen/ARM/ldm-stm-base-materialization.ll b/llvm/test/CodeGen/ARM/ldm-stm-base-materialization.ll
@@ -0,0 +1,93 @@
+; RUN: llc -mtriple armv7a-none-eabi -mattr=-neon < %s -verify-machineinstrs -o - | FileCheck %s
+
+; Thumb1 (thumbv6m) is tested in tests/Thumb
+
+@a = external global i32*
+@b = external global i32*
+
+; Function Attrs: nounwind
+define void @foo24() #0 {
+entry:
+; CHECK-LABEL: foo24:
+; We use '[rl0-9]*' to allow 'r0'..'r12', 'lr'
+; CHECK: movt [[LB:[rl0-9]+]], :upper16:b
+; CHECK: movt [[SB:[rl0-9]+]], :upper16:a
+; CHECK: add [[NLB:[rl0-9]+]], [[LB]], #4
+; CHECK: add [[NSB:[rl0-9]+]], [[SB]], #4
+; CHECK-NEXT: ldm [[NLB]], {[[R1:[rl0-9]+]], [[R2:[rl0-9]+]], [[R3:[rl0-9]+]], [[R4:[rl0-9]+]], [[R5:[rl0-9]+]], [[R6:[rl0-9]+]]}
+; CHECK-NEXT: stm [[NSB]], {[[R1]], [[R2]], [[R3]], [[R4]], [[R5]], [[R6]]}
+  %0 = load i32*, i32** @a, align 4
+  %arrayidx = getelementptr inbounds i32, i32* %0, i32 1
+  %1 = bitcast i32* %arrayidx to i8*
+  %2 = load i32*, i32** @b, align 4
+  %arrayidx1 = getelementptr inbounds i32, i32* %2, i32 1
+  %3 = bitcast i32* %arrayidx1 to i8*
+  tail call void @llvm.memcpy.p0i8.p0i8.i32(i8* %1, i8* %3, i32 24, i32 4, i1 false)
+  ret void
+}
+
+define void @foo28() #0 {
+entry:
+; CHECK-LABEL: foo28:
+; CHECK: movt [[LB:[rl0-9]+]], :upper16:b
+; CHECK: movt [[SB:[rl0-9]+]], :upper16:a
+; CHECK: add [[NLB:[rl0-9]+]], [[LB]], #4
+; CHECK: add [[NSB:[rl0-9]+]], [[SB]], #4
+; CHECK-NEXT: ldm [[NLB]]!, {[[R1:[rl0-9]+]], [[R2:[rl0-9]+]], [[R3:[rl0-9]+]]}
+; CHECK-NEXT: stm [[NSB]]!, {[[R1]], [[R2]], [[R3]]}
+; CHECK-NEXT: ldm [[NLB]], {[[R1:[rl0-9]+]], [[R2:[rl0-9]+]], [[R3:[rl0-9]+]], [[R4:[rl0-9]+]]}
+; CHECK-NEXT: stm [[NSB]], {[[R1]], [[R2]], [[R3]], [[R4]]}
+  %0 = load i32*, i32** @a, align 4
+  %arrayidx = getelementptr inbounds i32, i32* %0, i32 1
+  %1 = bitcast i32* %arrayidx to i8*
+  %2 = load i32*, i32** @b, align 4
+  %arrayidx1 = getelementptr inbounds i32, i32* %2, i32 1
+  %3 = bitcast i32* %arrayidx1 to i8*
+  tail call void @llvm.memcpy.p0i8.p0i8.i32(i8* %1, i8* %3, i32 28, i32 4, i1 false)
+  ret void
+}
+
+define void @foo32() #0 {
+entry:
+; CHECK-LABEL: foo32:
+; CHECK: movt [[LB:[rl0-9]+]], :upper16:b
+; CHECK: movt [[SB:[rl0-9]+]], :upper16:a
+; CHECK: add [[NLB:[rl0-9]+]], [[LB]], #4
+; CHECK: add [[NSB:[rl0-9]+]], [[SB]], #4
+; CHECK-NEXT: ldm [[NLB]]!, {[[R1:[rl0-9]+]], [[R2:[rl0-9]+]], [[R3:[rl0-9]+]], [[R4:[rl0-9]+]]}
+; CHECK-NEXT: stm [[NSB]]!, {[[R1]], [[R2]], [[R3]], [[R4]]}
+; CHECK-NEXT: ldm [[NLB]], {[[R1:[rl0-9]+]], [[R2:[rl0-9]+]], [[R3:[rl0-9]+]], [[R4:[rl0-9]+]]}
+; CHECK-NEXT: stm [[NSB]], {[[R1]], [[R2]], [[R3]], [[R4]]}
+  %0 = load i32*, i32** @a, align 4
+  %arrayidx = getelementptr inbounds i32, i32* %0, i32 1
+  %1 = bitcast i32* %arrayidx to i8*
+  %2 = load i32*, i32** @b, align 4
+  %arrayidx1 = getelementptr inbounds i32, i32* %2, i32 1
+  %3 = bitcast i32* %arrayidx1 to i8*
+  tail call void @llvm.memcpy.p0i8.p0i8.i32(i8* %1, i8* %3, i32 32, i32 4, i1 false)
+  ret void
+}
+
+define void @foo36() #0 {
+entry:
+; CHECK-LABEL: foo36:
+; CHECK: movt [[LB:[rl0-9]+]], :upper16:b
+; CHECK: movt [[SB:[rl0-9]+]], :upper16:a
+; CHECK: add [[NLB:[rl0-9]+]], [[LB]], #4
+; CHECK: add [[NSB:[rl0-9]+]], [[SB]], #4
+; CHECK-NEXT: ldm [[NLB]]!, {[[R1:[rl0-9]+]], [[R2:[rl0-9]+]], [[R3:[rl0-9]+]], [[R4:[rl0-9]+]]}
+; CHECK-NEXT: stm [[NSB]]!, {[[R1]], [[R2]], [[R3]], [[R4]]}
+; CHECK-NEXT: ldm [[NLB]], {[[R1:[rl0-9]+]], [[R2:[rl0-9]+]], [[R3:[rl0-9]+]], [[R4:[rl0-9]+]], [[R5:[rl0-9]+]]}
+; CHECK-NEXT: stm [[NSB]], {[[R1]], [[R2]], [[R3]], [[R4]], [[R5]]}
+  %0 = load i32*, i32** @a, align 4
+  %arrayidx = getelementptr inbounds i32, i32* %0, i32 1
+  %1 = bitcast i32* %arrayidx to i8*
+  %2 = load i32*, i32** @b, align 4
+  %arrayidx1 = getelementptr inbounds i32, i32* %2, i32 1
+  %3 = bitcast i32* %arrayidx1 to i8*
+  tail call void @llvm.memcpy.p0i8.p0i8.i32(i8* %1, i8* %3, i32 36, i32 4, i1 false)
+  ret void
+}
+
+; Function Attrs: nounwind
+declare void @llvm.memcpy.p0i8.p0i8.i32(i8* nocapture, i8* nocapture readonly, i32, i32, i1) #1