1 //===-- ARMSelectionDAGInfo.cpp - ARM SelectionDAG Info -------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the ARMSelectionDAGInfo class.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "arm-selectiondag-info"
15 #include "ARMTargetMachine.h"
16 #include "llvm/DerivedTypes.h"
17 #include "llvm/CodeGen/SelectionDAG.h"
20 ARMSelectionDAGInfo::ARMSelectionDAGInfo(const TargetMachine &TM)
21 : TargetSelectionDAGInfo(TM),
22 Subtarget(&TM.getSubtarget<ARMSubtarget>()) {
25 ARMSelectionDAGInfo::~ARMSelectionDAGInfo() {
29 ARMSelectionDAGInfo::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
31 SDValue Dst, SDValue Src,
32 SDValue Size, unsigned Align,
33 bool isVolatile, bool AlwaysInline,
34 MachinePointerInfo DstPtrInfo,
35 MachinePointerInfo SrcPtrInfo) const {
36 // Do repeated 4-byte loads and stores. To be improved.
37 // This requires 4-byte alignment.
40 // This requires the copy size to be a constant, preferably
41 // within a subtarget-specific limit.
42 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
45 uint64_t SizeVal = ConstantSize->getZExtValue();
46 if (!AlwaysInline && SizeVal > Subtarget->getMaxInlineSizeThreshold())
49 unsigned BytesLeft = SizeVal & 3;
50 unsigned NumMemOps = SizeVal >> 2;
51 unsigned EmittedNumMemOps = 0;
55 const unsigned MAX_LOADS_IN_LDM = 6;
56 SDValue TFOps[MAX_LOADS_IN_LDM];
57 SDValue Loads[MAX_LOADS_IN_LDM];
58 uint64_t SrcOff = 0, DstOff = 0;
60 // Emit up to MAX_LOADS_IN_LDM loads, then a TokenFactor barrier, then the
61 // same number of stores. The loads and stores will get combined into
63 while (EmittedNumMemOps < NumMemOps) {
65 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
66 Loads[i] = DAG.getLoad(VT, dl, Chain,
67 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
68 DAG.getConstant(SrcOff, MVT::i32)),
69 SrcPtrInfo.getWithOffset(SrcOff), isVolatile,
71 TFOps[i] = Loads[i].getValue(1);
74 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
77 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
78 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
79 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
80 DAG.getConstant(DstOff, MVT::i32)),
81 DstPtrInfo.getWithOffset(DstOff),
82 isVolatile, false, 0);
85 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
87 EmittedNumMemOps += i;
93 // Issue loads / stores for the trailing (1 - 3) bytes.
94 unsigned BytesLeftSave = BytesLeft;
105 Loads[i] = DAG.getLoad(VT, dl, Chain,
106 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
107 DAG.getConstant(SrcOff, MVT::i32)),
108 SrcPtrInfo.getWithOffset(SrcOff), false, false, 0);
109 TFOps[i] = Loads[i].getValue(1);
114 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
117 BytesLeft = BytesLeftSave;
119 if (BytesLeft >= 2) {
127 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
128 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
129 DAG.getConstant(DstOff, MVT::i32)),
130 DstPtrInfo.getWithOffset(DstOff), false, false, 0);
135 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
138 // Adjust parameters for memset, EABI uses format (ptr, size, value),
139 // GNU library uses (ptr, value, size)
140 // See RTABI section 4.3.4
142 ARMSelectionDAGInfo::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
143 SDValue Chain, SDValue Dst,
144 SDValue Src, SDValue Size,
145 unsigned Align, bool isVolatile,
146 MachinePointerInfo DstPtrInfo) const
148 // Use default for non AAPCS subtargets
149 if (!Subtarget->isAAPCS_ABI())
152 const ARMTargetLowering &TLI =
153 *static_cast<const ARMTargetLowering*>(DAG.getTarget().getTargetLowering());
154 TargetLowering::ArgListTy Args;
155 TargetLowering::ArgListEntry Entry;
157 // First argument: data pointer
158 Type *IntPtrTy = TLI.getTargetData()->getIntPtrType(*DAG.getContext());
161 Args.push_back(Entry);
163 // Second argument: buffer size
166 Entry.isSExt = false;
167 Args.push_back(Entry);
169 // Extend or truncate the argument to be an i32 value for the call.
170 if (Src.getValueType().bitsGT(MVT::i32))
171 Src = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Src);
173 Src = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Src);
175 // Third argument: value to fill
177 Entry.Ty = Type::getInt32Ty(*DAG.getContext());
179 Args.push_back(Entry);
181 // Emit __eabi_memset call
182 std::pair<SDValue,SDValue> CallResult =
183 TLI.LowerCallTo(Chain,
184 Type::getVoidTy(*DAG.getContext()), // return type
185 false, // return sign ext
186 false, // return zero ext
189 0, // number of fixed arguments
190 TLI.getLibcallCallingConv(RTLIB::MEMSET), // call conv
191 false, // is tail call
192 false, // is return val used
193 DAG.getExternalSymbol(TLI.getLibcallName(RTLIB::MEMSET),
194 TLI.getPointerTy()), // callee
195 Args, DAG, dl); // arg list, DAG and debug
197 return CallResult.second;