1 //===-- SystemZSelectionDAGInfo.cpp - SystemZ 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 SystemZSelectionDAGInfo class.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "systemz-selectiondag-info"
15 #include "SystemZTargetMachine.h"
16 #include "llvm/CodeGen/SelectionDAG.h"
20 SystemZSelectionDAGInfo::
21 SystemZSelectionDAGInfo(const SystemZTargetMachine &TM)
22 : TargetSelectionDAGInfo(TM) {
25 SystemZSelectionDAGInfo::~SystemZSelectionDAGInfo() {
28 // Use MVC to copy Size bytes from Src to Dest, deciding whether to use
29 // a loop or straight-line code.
30 static SDValue emitMVC(SelectionDAG &DAG, SDLoc DL, SDValue Chain,
31 SDValue Dst, SDValue Src, uint64_t Size) {
32 EVT PtrVT = Src.getValueType();
33 // The heuristic we use is to prefer loops for anything that would
34 // require 7 or more MVCs. With these kinds of sizes there isn't
35 // much to choose between straight-line code and looping code,
36 // since the time will be dominated by the MVCs themselves.
37 // However, the loop has 4 or 5 instructions (depending on whether
38 // the base addresses can be proved equal), so there doesn't seem
39 // much point using a loop for 5 * 256 bytes or fewer. Anything in
40 // the range (5 * 256, 6 * 256) will need another instruction after
41 // the loop, so it doesn't seem worth using a loop then either.
42 // The next value up, 6 * 256, can be implemented in the same
43 // number of straight-line MVCs as 6 * 256 - 1.
45 return DAG.getNode(SystemZISD::MVC_LOOP, DL, MVT::Other, Chain, Dst, Src,
46 DAG.getConstant(Size, PtrVT),
47 DAG.getConstant(Size / 256, PtrVT));
48 return DAG.getNode(SystemZISD::MVC, DL, MVT::Other, Chain, Dst, Src,
49 DAG.getConstant(Size, PtrVT));
52 SDValue SystemZSelectionDAGInfo::
53 EmitTargetCodeForMemcpy(SelectionDAG &DAG, SDLoc DL, SDValue Chain,
54 SDValue Dst, SDValue Src, SDValue Size, unsigned Align,
55 bool IsVolatile, bool AlwaysInline,
56 MachinePointerInfo DstPtrInfo,
57 MachinePointerInfo SrcPtrInfo) const {
61 if (ConstantSDNode *CSize = dyn_cast<ConstantSDNode>(Size))
62 return emitMVC(DAG, DL, Chain, Dst, Src, CSize->getZExtValue());
66 // Handle a memset of 1, 2, 4 or 8 bytes with the operands given by
67 // Chain, Dst, ByteVal and Size. These cases are expected to use
68 // MVI, MVHHI, MVHI and MVGHI respectively.
69 static SDValue memsetStore(SelectionDAG &DAG, SDLoc DL, SDValue Chain,
70 SDValue Dst, uint64_t ByteVal, uint64_t Size,
72 MachinePointerInfo DstPtrInfo) {
73 uint64_t StoreVal = ByteVal;
74 for (unsigned I = 1; I < Size; ++I)
75 StoreVal |= ByteVal << (I * 8);
76 return DAG.getStore(Chain, DL,
77 DAG.getConstant(StoreVal, MVT::getIntegerVT(Size * 8)),
78 Dst, DstPtrInfo, false, false, Align);
81 SDValue SystemZSelectionDAGInfo::
82 EmitTargetCodeForMemset(SelectionDAG &DAG, SDLoc DL, SDValue Chain,
83 SDValue Dst, SDValue Byte, SDValue Size,
84 unsigned Align, bool IsVolatile,
85 MachinePointerInfo DstPtrInfo) const {
86 EVT PtrVT = Dst.getValueType();
91 if (ConstantSDNode *CSize = dyn_cast<ConstantSDNode>(Size)) {
92 uint64_t Bytes = CSize->getZExtValue();
95 if (ConstantSDNode *CByte = dyn_cast<ConstantSDNode>(Byte)) {
96 // Handle cases that can be done using at most two of
97 // MVI, MVHI, MVHHI and MVGHI. The latter two can only be
98 // used if ByteVal is all zeros or all ones; in other casees,
99 // we can move at most 2 halfwords.
100 uint64_t ByteVal = CByte->getZExtValue();
101 if (ByteVal == 0 || ByteVal == 255 ?
102 Bytes <= 16 && CountPopulation_64(Bytes) <= 2 :
104 unsigned Size1 = Bytes == 16 ? 8 : 1 << findLastSet(Bytes);
105 unsigned Size2 = Bytes - Size1;
106 SDValue Chain1 = memsetStore(DAG, DL, Chain, Dst, ByteVal, Size1,
110 Dst = DAG.getNode(ISD::ADD, DL, PtrVT, Dst,
111 DAG.getConstant(Size1, PtrVT));
112 DstPtrInfo = DstPtrInfo.getWithOffset(Size1);
113 SDValue Chain2 = memsetStore(DAG, DL, Chain, Dst, ByteVal, Size2,
114 std::min(Align, Size1), DstPtrInfo);
115 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chain1, Chain2);
118 // Handle one and two bytes using STC.
120 SDValue Chain1 = DAG.getStore(Chain, DL, Byte, Dst, DstPtrInfo,
121 false, false, Align);
124 SDValue Dst2 = DAG.getNode(ISD::ADD, DL, PtrVT, Dst,
125 DAG.getConstant(1, PtrVT));
126 SDValue Chain2 = DAG.getStore(Chain, DL, Byte, Dst2,
127 DstPtrInfo.getWithOffset(1),
129 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chain1, Chain2);
132 assert(Bytes >= 2 && "Should have dealt with 0- and 1-byte cases already");
133 // Copy the byte to the first location and then use MVC to copy
135 Chain = DAG.getStore(Chain, DL, Byte, Dst, DstPtrInfo,
136 false, false, Align);
137 SDValue DstPlus1 = DAG.getNode(ISD::ADD, DL, PtrVT, Dst,
138 DAG.getConstant(1, PtrVT));
139 return emitMVC(DAG, DL, Chain, DstPlus1, Dst, Bytes - 1);
144 // Use CLC to compare [Src1, Src1 + Size) with [Src2, Src2 + Size),
145 // deciding whether to use a loop or straight-line code.
146 static SDValue emitCLC(SelectionDAG &DAG, SDLoc DL, SDValue Chain,
147 SDValue Src1, SDValue Src2, uint64_t Size) {
148 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
149 EVT PtrVT = Src1.getValueType();
150 // A two-CLC sequence is a clear win over a loop, not least because it
151 // needs only one branch. A three-CLC sequence needs the same number
152 // of branches as a loop (i.e. 2), but is shorter. That brings us to
153 // lengths greater than 768 bytes. It seems relatively likely that
154 // a difference will be found within the first 768 bytes, so we just
155 // optimize for the smallest number of branch instructions, in order
156 // to avoid polluting the prediction buffer too much. A loop only ever
157 // needs 2 branches, whereas a straight-line sequence would need 3 or more.
159 return DAG.getNode(SystemZISD::CLC_LOOP, DL, VTs, Chain, Src1, Src2,
160 DAG.getConstant(Size, PtrVT),
161 DAG.getConstant(Size / 256, PtrVT));
162 return DAG.getNode(SystemZISD::CLC, DL, VTs, Chain, Src1, Src2,
163 DAG.getConstant(Size, PtrVT));
166 // Convert the current CC value into an integer that is 0 if CC == 0,
167 // less than zero if CC == 1 and greater than zero if CC >= 2.
168 // The sequence starts with IPM, which puts CC into bits 29 and 28
169 // of an integer and clears bits 30 and 31.
170 static SDValue addIPMSequence(SDLoc DL, SDValue Glue, SelectionDAG &DAG) {
171 SDValue IPM = DAG.getNode(SystemZISD::IPM, DL, MVT::i32, Glue);
172 SDValue SRL = DAG.getNode(ISD::SRL, DL, MVT::i32, IPM,
173 DAG.getConstant(28, MVT::i32));
174 SDValue ROTL = DAG.getNode(ISD::ROTL, DL, MVT::i32, SRL,
175 DAG.getConstant(31, MVT::i32));
179 std::pair<SDValue, SDValue> SystemZSelectionDAGInfo::
180 EmitTargetCodeForMemcmp(SelectionDAG &DAG, SDLoc DL, SDValue Chain,
181 SDValue Src1, SDValue Src2, SDValue Size,
182 MachinePointerInfo Op1PtrInfo,
183 MachinePointerInfo Op2PtrInfo) const {
184 if (ConstantSDNode *CSize = dyn_cast<ConstantSDNode>(Size)) {
185 uint64_t Bytes = CSize->getZExtValue();
186 assert(Bytes > 0 && "Caller should have handled 0-size case");
187 Chain = emitCLC(DAG, DL, Chain, Src1, Src2, Bytes);
188 SDValue Glue = Chain.getValue(1);
189 return std::make_pair(addIPMSequence(DL, Glue, DAG), Chain);
191 return std::make_pair(SDValue(), SDValue());
194 std::pair<SDValue, SDValue> SystemZSelectionDAGInfo::
195 EmitTargetCodeForMemchr(SelectionDAG &DAG, SDLoc DL, SDValue Chain,
196 SDValue Src, SDValue Char, SDValue Length,
197 MachinePointerInfo SrcPtrInfo) const {
198 // Use SRST to find the character. End is its address on success.
199 EVT PtrVT = Src.getValueType();
200 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other, MVT::Glue);
201 Length = DAG.getZExtOrTrunc(Length, DL, PtrVT);
202 Char = DAG.getZExtOrTrunc(Char, DL, MVT::i32);
203 Char = DAG.getNode(ISD::AND, DL, MVT::i32, Char,
204 DAG.getConstant(255, MVT::i32));
205 SDValue Limit = DAG.getNode(ISD::ADD, DL, PtrVT, Src, Length);
206 SDValue End = DAG.getNode(SystemZISD::SEARCH_STRING, DL, VTs, Chain,
208 Chain = End.getValue(1);
209 SDValue Glue = End.getValue(2);
211 // Now select between End and null, depending on whether the character
213 SmallVector<SDValue, 5> Ops;
215 Ops.push_back(DAG.getConstant(0, PtrVT));
216 Ops.push_back(DAG.getConstant(SystemZ::CCMASK_SRST, MVT::i32));
217 Ops.push_back(DAG.getConstant(SystemZ::CCMASK_SRST_FOUND, MVT::i32));
219 VTs = DAG.getVTList(PtrVT, MVT::Glue);
220 End = DAG.getNode(SystemZISD::SELECT_CCMASK, DL, VTs, &Ops[0], Ops.size());
221 return std::make_pair(End, Chain);
224 std::pair<SDValue, SDValue> SystemZSelectionDAGInfo::
225 EmitTargetCodeForStrcpy(SelectionDAG &DAG, SDLoc DL, SDValue Chain,
226 SDValue Dest, SDValue Src,
227 MachinePointerInfo DestPtrInfo,
228 MachinePointerInfo SrcPtrInfo, bool isStpcpy) const {
229 SDVTList VTs = DAG.getVTList(Dest.getValueType(), MVT::Other);
230 SDValue EndDest = DAG.getNode(SystemZISD::STPCPY, DL, VTs, Chain, Dest, Src,
231 DAG.getConstant(0, MVT::i32));
232 return std::make_pair(isStpcpy ? EndDest : Dest, EndDest.getValue(1));
235 std::pair<SDValue, SDValue> SystemZSelectionDAGInfo::
236 EmitTargetCodeForStrcmp(SelectionDAG &DAG, SDLoc DL, SDValue Chain,
237 SDValue Src1, SDValue Src2,
238 MachinePointerInfo Op1PtrInfo,
239 MachinePointerInfo Op2PtrInfo) const {
240 SDVTList VTs = DAG.getVTList(Src1.getValueType(), MVT::Other, MVT::Glue);
241 SDValue Unused = DAG.getNode(SystemZISD::STRCMP, DL, VTs, Chain, Src1, Src2,
242 DAG.getConstant(0, MVT::i32));
243 Chain = Unused.getValue(1);
244 SDValue Glue = Chain.getValue(2);
245 return std::make_pair(addIPMSequence(DL, Glue, DAG), Chain);
248 // Search from Src for a null character, stopping once Src reaches Limit.
249 // Return a pair of values, the first being the number of nonnull characters
250 // and the second being the out chain.
252 // This can be used for strlen by setting Limit to 0.
253 static std::pair<SDValue, SDValue> getBoundedStrlen(SelectionDAG &DAG, SDLoc DL,
254 SDValue Chain, SDValue Src,
256 EVT PtrVT = Src.getValueType();
257 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other, MVT::Glue);
258 SDValue End = DAG.getNode(SystemZISD::SEARCH_STRING, DL, VTs, Chain,
259 Limit, Src, DAG.getConstant(0, MVT::i32));
260 Chain = End.getValue(1);
261 SDValue Len = DAG.getNode(ISD::SUB, DL, PtrVT, End, Src);
262 return std::make_pair(Len, Chain);
265 std::pair<SDValue, SDValue> SystemZSelectionDAGInfo::
266 EmitTargetCodeForStrlen(SelectionDAG &DAG, SDLoc DL, SDValue Chain,
267 SDValue Src, MachinePointerInfo SrcPtrInfo) const {
268 EVT PtrVT = Src.getValueType();
269 return getBoundedStrlen(DAG, DL, Chain, Src, DAG.getConstant(0, PtrVT));
272 std::pair<SDValue, SDValue> SystemZSelectionDAGInfo::
273 EmitTargetCodeForStrnlen(SelectionDAG &DAG, SDLoc DL, SDValue Chain,
274 SDValue Src, SDValue MaxLength,
275 MachinePointerInfo SrcPtrInfo) const {
276 EVT PtrVT = Src.getValueType();
277 MaxLength = DAG.getZExtOrTrunc(MaxLength, DL, PtrVT);
278 SDValue Limit = DAG.getNode(ISD::ADD, DL, PtrVT, Src, MaxLength);
279 return getBoundedStrlen(DAG, DL, Chain, Src, Limit);