1//===-- SystemZSelectionDAGInfo.cpp - SystemZ SelectionDAG Info -----------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file implements the SystemZSelectionDAGInfo class.
10//
11//===----------------------------------------------------------------------===//
12
13#include "SystemZTargetMachine.h"
14#include "llvm/CodeGen/SelectionDAG.h"
15
16using namespace llvm;
17
18#define DEBUG_TYPE "systemz-selectiondag-info"
19
20// Decide whether it is best to use a loop or straight-line code for
21// a block operation of Size bytes with source address Src and destination
22// address Dest.  Sequence is the opcode to use for straight-line code
23// (such as MVC) and Loop is the opcode to use for loops (such as MVC_LOOP).
24// Return the chain for the completed operation.
25static SDValue emitMemMem(SelectionDAG &DAG, const SDLoc &DL, unsigned Sequence,
26                          unsigned Loop, SDValue Chain, SDValue Dst,
27                          SDValue Src, uint64_t Size) {
28  EVT PtrVT = Src.getValueType();
29  // The heuristic we use is to prefer loops for anything that would
30  // require 7 or more MVCs.  With these kinds of sizes there isn't
31  // much to choose between straight-line code and looping code,
32  // since the time will be dominated by the MVCs themselves.
33  // However, the loop has 4 or 5 instructions (depending on whether
34  // the base addresses can be proved equal), so there doesn't seem
35  // much point using a loop for 5 * 256 bytes or fewer.  Anything in
36  // the range (5 * 256, 6 * 256) will need another instruction after
37  // the loop, so it doesn't seem worth using a loop then either.
38  // The next value up, 6 * 256, can be implemented in the same
39  // number of straight-line MVCs as 6 * 256 - 1.
40  if (Size > 6 * 256)
41    return DAG.getNode(Loop, DL, MVT::Other, Chain, Dst, Src,
42                       DAG.getConstant(Size, DL, PtrVT),
43                       DAG.getConstant(Size / 256, DL, PtrVT));
44  return DAG.getNode(Sequence, DL, MVT::Other, Chain, Dst, Src,
45                     DAG.getConstant(Size, DL, PtrVT));
46}
47
48SDValue SystemZSelectionDAGInfo::EmitTargetCodeForMemcpy(
49    SelectionDAG &DAG, const SDLoc &DL, SDValue Chain, SDValue Dst, SDValue Src,
50    SDValue Size, unsigned Align, bool IsVolatile, bool AlwaysInline,
51    MachinePointerInfo DstPtrInfo, MachinePointerInfo SrcPtrInfo) const {
52  if (IsVolatile)
53    return SDValue();
54
55  if (auto *CSize = dyn_cast<ConstantSDNode>(Size))
56    return emitMemMem(DAG, DL, SystemZISD::MVC, SystemZISD::MVC_LOOP,
57                      Chain, Dst, Src, CSize->getZExtValue());
58  return SDValue();
59}
60
61// Handle a memset of 1, 2, 4 or 8 bytes with the operands given by
62// Chain, Dst, ByteVal and Size.  These cases are expected to use
63// MVI, MVHHI, MVHI and MVGHI respectively.
64static SDValue memsetStore(SelectionDAG &DAG, const SDLoc &DL, SDValue Chain,
65                           SDValue Dst, uint64_t ByteVal, uint64_t Size,
66                           unsigned Align, MachinePointerInfo DstPtrInfo) {
67  uint64_t StoreVal = ByteVal;
68  for (unsigned I = 1; I < Size; ++I)
69    StoreVal |= ByteVal << (I * 8);
70  return DAG.getStore(
71      Chain, DL, DAG.getConstant(StoreVal, DL, MVT::getIntegerVT(Size * 8)),
72      Dst, DstPtrInfo, Align);
73}
74
75SDValue SystemZSelectionDAGInfo::EmitTargetCodeForMemset(
76    SelectionDAG &DAG, const SDLoc &DL, SDValue Chain, SDValue Dst,
77    SDValue Byte, SDValue Size, unsigned Align, bool IsVolatile,
78    MachinePointerInfo DstPtrInfo) const {
79  EVT PtrVT = Dst.getValueType();
80
81  if (IsVolatile)
82    return SDValue();
83
84  if (auto *CSize = dyn_cast<ConstantSDNode>(Size)) {
85    uint64_t Bytes = CSize->getZExtValue();
86    if (Bytes == 0)
87      return SDValue();
88    if (auto *CByte = dyn_cast<ConstantSDNode>(Byte)) {
89      // Handle cases that can be done using at most two of
90      // MVI, MVHI, MVHHI and MVGHI.  The latter two can only be
91      // used if ByteVal is all zeros or all ones; in other casees,
92      // we can move at most 2 halfwords.
93      uint64_t ByteVal = CByte->getZExtValue();
94      if (ByteVal == 0 || ByteVal == 255 ?
95          Bytes <= 16 && countPopulation(Bytes) <= 2 :
96          Bytes <= 4) {
97        unsigned Size1 = Bytes == 16 ? 8 : 1 << findLastSet(Bytes);
98        unsigned Size2 = Bytes - Size1;
99        SDValue Chain1 = memsetStore(DAG, DL, Chain, Dst, ByteVal, Size1,
100                                     Align, DstPtrInfo);
101        if (Size2 == 0)
102          return Chain1;
103        Dst = DAG.getNode(ISD::ADD, DL, PtrVT, Dst,
104                          DAG.getConstant(Size1, DL, PtrVT));
105        DstPtrInfo = DstPtrInfo.getWithOffset(Size1);
106        SDValue Chain2 = memsetStore(DAG, DL, Chain, Dst, ByteVal, Size2,
107                                     std::min(Align, Size1), DstPtrInfo);
108        return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chain1, Chain2);
109      }
110    } else {
111      // Handle one and two bytes using STC.
112      if (Bytes <= 2) {
113        SDValue Chain1 = DAG.getStore(Chain, DL, Byte, Dst, DstPtrInfo, Align);
114        if (Bytes == 1)
115          return Chain1;
116        SDValue Dst2 = DAG.getNode(ISD::ADD, DL, PtrVT, Dst,
117                                   DAG.getConstant(1, DL, PtrVT));
118        SDValue Chain2 =
119            DAG.getStore(Chain, DL, Byte, Dst2, DstPtrInfo.getWithOffset(1),
120                         /* Alignment = */ 1);
121        return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chain1, Chain2);
122      }
123    }
124    assert(Bytes >= 2 && "Should have dealt with 0- and 1-byte cases already");
125
126    // Handle the special case of a memset of 0, which can use XC.
127    auto *CByte = dyn_cast<ConstantSDNode>(Byte);
128    if (CByte && CByte->getZExtValue() == 0)
129      return emitMemMem(DAG, DL, SystemZISD::XC, SystemZISD::XC_LOOP,
130                        Chain, Dst, Dst, Bytes);
131
132    // Copy the byte to the first location and then use MVC to copy
133    // it to the rest.
134    Chain = DAG.getStore(Chain, DL, Byte, Dst, DstPtrInfo, Align);
135    SDValue DstPlus1 = DAG.getNode(ISD::ADD, DL, PtrVT, Dst,
136                                   DAG.getConstant(1, DL, PtrVT));
137    return emitMemMem(DAG, DL, SystemZISD::MVC, SystemZISD::MVC_LOOP,
138                      Chain, DstPlus1, Dst, Bytes - 1);
139  }
140  return SDValue();
141}
142
143// Use CLC to compare [Src1, Src1 + Size) with [Src2, Src2 + Size),
144// deciding whether to use a loop or straight-line code.
145static SDValue emitCLC(SelectionDAG &DAG, const SDLoc &DL, SDValue Chain,
146                       SDValue Src1, SDValue Src2, uint64_t Size) {
147  SDVTList VTs = DAG.getVTList(MVT::i32, MVT::Other);
148  EVT PtrVT = Src1.getValueType();
149  // A two-CLC sequence is a clear win over a loop, not least because it
150  // needs only one branch.  A three-CLC sequence needs the same number
151  // of branches as a loop (i.e. 2), but is shorter.  That brings us to
152  // lengths greater than 768 bytes.  It seems relatively likely that
153  // a difference will be found within the first 768 bytes, so we just
154  // optimize for the smallest number of branch instructions, in order
155  // to avoid polluting the prediction buffer too much.  A loop only ever
156  // needs 2 branches, whereas a straight-line sequence would need 3 or more.
157  if (Size > 3 * 256)
158    return DAG.getNode(SystemZISD::CLC_LOOP, DL, VTs, Chain, Src1, Src2,
159                       DAG.getConstant(Size, DL, PtrVT),
160                       DAG.getConstant(Size / 256, DL, PtrVT));
161  return DAG.getNode(SystemZISD::CLC, DL, VTs, Chain, Src1, Src2,
162                     DAG.getConstant(Size, DL, PtrVT));
163}
164
165// Convert the current CC value into an integer that is 0 if CC == 0,
166// greater than zero if CC == 1 and less than zero if CC >= 2.
167// The sequence starts with IPM, which puts CC into bits 29 and 28
168// of an integer and clears bits 30 and 31.
169static SDValue addIPMSequence(const SDLoc &DL, SDValue CCReg,
170                              SelectionDAG &DAG) {
171  SDValue IPM = DAG.getNode(SystemZISD::IPM, DL, MVT::i32, CCReg);
172  SDValue SHL = DAG.getNode(ISD::SHL, DL, MVT::i32, IPM,
173                            DAG.getConstant(30 - SystemZ::IPM_CC, DL, MVT::i32));
174  SDValue SRA = DAG.getNode(ISD::SRA, DL, MVT::i32, SHL,
175                            DAG.getConstant(30, DL, MVT::i32));
176  return SRA;
177}
178
179std::pair<SDValue, SDValue> SystemZSelectionDAGInfo::EmitTargetCodeForMemcmp(
180    SelectionDAG &DAG, const SDLoc &DL, SDValue Chain, SDValue Src1,
181    SDValue Src2, SDValue Size, MachinePointerInfo Op1PtrInfo,
182    MachinePointerInfo Op2PtrInfo) const {
183  if (auto *CSize = dyn_cast<ConstantSDNode>(Size)) {
184    uint64_t Bytes = CSize->getZExtValue();
185    assert(Bytes > 0 && "Caller should have handled 0-size case");
186    // Swap operands to invert CC == 1 vs. CC == 2 cases.
187    SDValue CCReg = emitCLC(DAG, DL, Chain, Src2, Src1, Bytes);
188    Chain = CCReg.getValue(1);
189    return std::make_pair(addIPMSequence(DL, CCReg, DAG), Chain);
190  }
191  return std::make_pair(SDValue(), SDValue());
192}
193
194std::pair<SDValue, SDValue> SystemZSelectionDAGInfo::EmitTargetCodeForMemchr(
195    SelectionDAG &DAG, const SDLoc &DL, SDValue Chain, SDValue Src,
196    SDValue Char, SDValue Length, MachinePointerInfo SrcPtrInfo) const {
197  // Use SRST to find the character.  End is its address on success.
198  EVT PtrVT = Src.getValueType();
199  SDVTList VTs = DAG.getVTList(PtrVT, MVT::i32, MVT::Other);
200  Length = DAG.getZExtOrTrunc(Length, DL, PtrVT);
201  Char = DAG.getZExtOrTrunc(Char, DL, MVT::i32);
202  Char = DAG.getNode(ISD::AND, DL, MVT::i32, Char,
203                     DAG.getConstant(255, DL, MVT::i32));
204  SDValue Limit = DAG.getNode(ISD::ADD, DL, PtrVT, Src, Length);
205  SDValue End = DAG.getNode(SystemZISD::SEARCH_STRING, DL, VTs, Chain,
206                            Limit, Src, Char);
207  SDValue CCReg = End.getValue(1);
208  Chain = End.getValue(2);
209
210  // Now select between End and null, depending on whether the character
211  // was found.
212  SDValue Ops[] = {
213      End, DAG.getConstant(0, DL, PtrVT),
214      DAG.getTargetConstant(SystemZ::CCMASK_SRST, DL, MVT::i32),
215      DAG.getTargetConstant(SystemZ::CCMASK_SRST_FOUND, DL, MVT::i32), CCReg};
216  End = DAG.getNode(SystemZISD::SELECT_CCMASK, DL, PtrVT, Ops);
217  return std::make_pair(End, Chain);
218}
219
220std::pair<SDValue, SDValue> SystemZSelectionDAGInfo::EmitTargetCodeForStrcpy(
221    SelectionDAG &DAG, const SDLoc &DL, SDValue Chain, SDValue Dest,
222    SDValue Src, MachinePointerInfo DestPtrInfo, MachinePointerInfo SrcPtrInfo,
223    bool isStpcpy) const {
224  SDVTList VTs = DAG.getVTList(Dest.getValueType(), MVT::Other);
225  SDValue EndDest = DAG.getNode(SystemZISD::STPCPY, DL, VTs, Chain, Dest, Src,
226                                DAG.getConstant(0, DL, MVT::i32));
227  return std::make_pair(isStpcpy ? EndDest : Dest, EndDest.getValue(1));
228}
229
230std::pair<SDValue, SDValue> SystemZSelectionDAGInfo::EmitTargetCodeForStrcmp(
231    SelectionDAG &DAG, const SDLoc &DL, SDValue Chain, SDValue Src1,
232    SDValue Src2, MachinePointerInfo Op1PtrInfo,
233    MachinePointerInfo Op2PtrInfo) const {
234  SDVTList VTs = DAG.getVTList(Src1.getValueType(), MVT::i32, MVT::Other);
235  // Swap operands to invert CC == 1 vs. CC == 2 cases.
236  SDValue Unused = DAG.getNode(SystemZISD::STRCMP, DL, VTs, Chain, Src2, Src1,
237                               DAG.getConstant(0, DL, MVT::i32));
238  SDValue CCReg = Unused.getValue(1);
239  Chain = Unused.getValue(2);
240  return std::make_pair(addIPMSequence(DL, CCReg, DAG), Chain);
241}
242
243// Search from Src for a null character, stopping once Src reaches Limit.
244// Return a pair of values, the first being the number of nonnull characters
245// and the second being the out chain.
246//
247// This can be used for strlen by setting Limit to 0.
248static std::pair<SDValue, SDValue> getBoundedStrlen(SelectionDAG &DAG,
249                                                    const SDLoc &DL,
250                                                    SDValue Chain, SDValue Src,
251                                                    SDValue Limit) {
252  EVT PtrVT = Src.getValueType();
253  SDVTList VTs = DAG.getVTList(PtrVT, MVT::i32, MVT::Other);
254  SDValue End = DAG.getNode(SystemZISD::SEARCH_STRING, DL, VTs, Chain,
255                            Limit, Src, DAG.getConstant(0, DL, MVT::i32));
256  Chain = End.getValue(2);
257  SDValue Len = DAG.getNode(ISD::SUB, DL, PtrVT, End, Src);
258  return std::make_pair(Len, Chain);
259}
260
261std::pair<SDValue, SDValue> SystemZSelectionDAGInfo::EmitTargetCodeForStrlen(
262    SelectionDAG &DAG, const SDLoc &DL, SDValue Chain, SDValue Src,
263    MachinePointerInfo SrcPtrInfo) const {
264  EVT PtrVT = Src.getValueType();
265  return getBoundedStrlen(DAG, DL, Chain, Src, DAG.getConstant(0, DL, PtrVT));
266}
267
268std::pair<SDValue, SDValue> SystemZSelectionDAGInfo::EmitTargetCodeForStrnlen(
269    SelectionDAG &DAG, const SDLoc &DL, SDValue Chain, SDValue Src,
270    SDValue MaxLength, MachinePointerInfo SrcPtrInfo) const {
271  EVT PtrVT = Src.getValueType();
272  MaxLength = DAG.getZExtOrTrunc(MaxLength, DL, PtrVT);
273  SDValue Limit = DAG.getNode(ISD::ADD, DL, PtrVT, Src, MaxLength);
274  return getBoundedStrlen(DAG, DL, Chain, Src, Limit);
275}
276