1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
|
//===--- ScheduleDAGSDNodes.cpp - Implement the ScheduleDAGSDNodes class --===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This implements the ScheduleDAG class, which is a base class used by
// scheduling implementation classes.
//
//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "pre-RA-sched"
#include "ScheduleDAGSDNodes.h"
#include "llvm/CodeGen/SelectionDAG.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetRegisterInfo.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
using namespace llvm;
ScheduleDAGSDNodes::ScheduleDAGSDNodes(MachineFunction &mf)
: ScheduleDAG(mf) {
}
/// Run - perform scheduling.
///
void ScheduleDAGSDNodes::Run(SelectionDAG *dag, MachineBasicBlock *bb,
MachineBasicBlock::iterator insertPos) {
DAG = dag;
ScheduleDAG::Run(bb, insertPos);
}
SUnit *ScheduleDAGSDNodes::Clone(SUnit *Old) {
SUnit *SU = NewSUnit(Old->getNode());
SU->OrigNode = Old->OrigNode;
SU->Latency = Old->Latency;
SU->isTwoAddress = Old->isTwoAddress;
SU->isCommutable = Old->isCommutable;
SU->hasPhysRegDefs = Old->hasPhysRegDefs;
Old->isCloned = true;
return SU;
}
/// CheckForPhysRegDependency - Check if the dependency between def and use of
/// a specified operand is a physical register dependency. If so, returns the
/// register and the cost of copying the register.
static void CheckForPhysRegDependency(SDNode *Def, SDNode *User, unsigned Op,
const TargetRegisterInfo *TRI,
const TargetInstrInfo *TII,
unsigned &PhysReg, int &Cost) {
if (Op != 2 || User->getOpcode() != ISD::CopyToReg)
return;
unsigned Reg = cast<RegisterSDNode>(User->getOperand(1))->getReg();
if (TargetRegisterInfo::isVirtualRegister(Reg))
return;
unsigned ResNo = User->getOperand(2).getResNo();
if (Def->isMachineOpcode()) {
const TargetInstrDesc &II = TII->get(Def->getMachineOpcode());
if (ResNo >= II.getNumDefs() &&
II.ImplicitDefs[ResNo - II.getNumDefs()] == Reg) {
PhysReg = Reg;
const TargetRegisterClass *RC =
TRI->getPhysicalRegisterRegClass(Reg, Def->getValueType(ResNo));
Cost = RC->getCopyCost();
}
}
}
void ScheduleDAGSDNodes::BuildSchedUnits() {
// During scheduling, the NodeId field of SDNode is used to map SDNodes
// to their associated SUnits by holding SUnits table indices. A value
// of -1 means the SDNode does not yet have an associated SUnit.
unsigned NumNodes = 0;
for (SelectionDAG::allnodes_iterator NI = DAG->allnodes_begin(),
E = DAG->allnodes_end(); NI != E; ++NI) {
NI->setNodeId(-1);
++NumNodes;
}
// Reserve entries in the vector for each of the SUnits we are creating. This
// ensure that reallocation of the vector won't happen, so SUnit*'s won't get
// invalidated.
// FIXME: Multiply by 2 because we may clone nodes during scheduling.
// This is a temporary workaround.
SUnits.reserve(NumNodes * 2);
// Check to see if the scheduler cares about latencies.
bool UnitLatencies = ForceUnitLatencies();
for (SelectionDAG::allnodes_iterator NI = DAG->allnodes_begin(),
E = DAG->allnodes_end(); NI != E; ++NI) {
if (isPassiveNode(NI)) // Leaf node, e.g. a TargetImmediate.
continue;
// If this node has already been processed, stop now.
if (NI->getNodeId() != -1) continue;
SUnit *NodeSUnit = NewSUnit(NI);
// See if anything is flagged to this node, if so, add them to flagged
// nodes. Nodes can have at most one flag input and one flag output. Flags
// are required the be the last operand and result of a node.
// Scan up to find flagged preds.
SDNode *N = NI;
if (N->getNumOperands() &&
N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Flag) {
do {
N = N->getOperand(N->getNumOperands()-1).getNode();
assert(N->getNodeId() == -1 && "Node already inserted!");
N->setNodeId(NodeSUnit->NodeNum);
} while (N->getNumOperands() &&
N->getOperand(N->getNumOperands()-1).getValueType()== MVT::Flag);
}
// Scan down to find any flagged succs.
N = NI;
while (N->getValueType(N->getNumValues()-1) == MVT::Flag) {
SDValue FlagVal(N, N->getNumValues()-1);
// There are either zero or one users of the Flag result.
bool HasFlagUse = false;
for (SDNode::use_iterator UI = N->use_begin(), E = N->use_end();
UI != E; ++UI)
if (FlagVal.isOperandOf(*UI)) {
HasFlagUse = true;
assert(N->getNodeId() == -1 && "Node already inserted!");
N->setNodeId(NodeSUnit->NodeNum);
N = *UI;
break;
}
if (!HasFlagUse) break;
}
// If there are flag operands involved, N is now the bottom-most node
// of the sequence of nodes that are flagged together.
// Update the SUnit.
NodeSUnit->setNode(N);
assert(N->getNodeId() == -1 && "Node already inserted!");
N->setNodeId(NodeSUnit->NodeNum);
// Assign the Latency field of NodeSUnit using target-provided information.
if (UnitLatencies)
NodeSUnit->Latency = 1;
else
ComputeLatency(NodeSUnit);
}
}
void ScheduleDAGSDNodes::AddSchedEdges() {
// Pass 2: add the preds, succs, etc.
for (unsigned su = 0, e = SUnits.size(); su != e; ++su) {
SUnit *SU = &SUnits[su];
SDNode *MainNode = SU->getNode();
if (MainNode->isMachineOpcode()) {
unsigned Opc = MainNode->getMachineOpcode();
const TargetInstrDesc &TID = TII->get(Opc);
for (unsigned i = 0; i != TID.getNumOperands(); ++i) {
if (TID.getOperandConstraint(i, TOI::TIED_TO) != -1) {
SU->isTwoAddress = true;
break;
}
}
if (TID.isCommutable())
SU->isCommutable = true;
}
// Find all predecessors and successors of the group.
for (SDNode *N = SU->getNode(); N; N = N->getFlaggedNode()) {
if (N->isMachineOpcode() &&
TII->get(N->getMachineOpcode()).getImplicitDefs() &&
CountResults(N) > TII->get(N->getMachineOpcode()).getNumDefs())
SU->hasPhysRegDefs = true;
for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
SDNode *OpN = N->getOperand(i).getNode();
if (isPassiveNode(OpN)) continue; // Not scheduled.
SUnit *OpSU = &SUnits[OpN->getNodeId()];
assert(OpSU && "Node has no SUnit!");
if (OpSU == SU) continue; // In the same group.
MVT OpVT = N->getOperand(i).getValueType();
assert(OpVT != MVT::Flag && "Flagged nodes should be in same sunit!");
bool isChain = OpVT == MVT::Other;
unsigned PhysReg = 0;
int Cost = 1;
// Determine if this is a physical register dependency.
CheckForPhysRegDependency(OpN, N, i, TRI, TII, PhysReg, Cost);
assert((PhysReg == 0 || !isChain) &&
"Chain dependence via physreg data?");
// FIXME: See ScheduleDAGSDNodes::EmitCopyFromReg. For now, scheduler
// emits a copy from the physical register to a virtual register unless
// it requires a cross class copy (cost < 0). That means we are only
// treating "expensive to copy" register dependency as physical register
// dependency. This may change in the future though.
if (Cost >= 0)
PhysReg = 0;
SU->addPred(SDep(OpSU, isChain ? SDep::Order : SDep::Data,
OpSU->Latency, PhysReg));
}
}
}
}
/// BuildSchedGraph - Build the SUnit graph from the selection dag that we
/// are input. This SUnit graph is similar to the SelectionDAG, but
/// excludes nodes that aren't interesting to scheduling, and represents
/// flagged together nodes with a single SUnit.
void ScheduleDAGSDNodes::BuildSchedGraph() {
// Populate the SUnits array.
BuildSchedUnits();
// Compute all the scheduling dependencies between nodes.
AddSchedEdges();
}
void ScheduleDAGSDNodes::ComputeLatency(SUnit *SU) {
const InstrItineraryData &InstrItins = TM.getInstrItineraryData();
// Compute the latency for the node. We use the sum of the latencies for
// all nodes flagged together into this SUnit.
SU->Latency = 0;
bool SawMachineOpcode = false;
for (SDNode *N = SU->getNode(); N; N = N->getFlaggedNode())
if (N->isMachineOpcode()) {
SawMachineOpcode = true;
SU->Latency +=
InstrItins.getLatency(TII->get(N->getMachineOpcode()).getSchedClass());
}
}
/// CountResults - The results of target nodes have register or immediate
/// operands first, then an optional chain, and optional flag operands (which do
/// not go into the resulting MachineInstr).
unsigned ScheduleDAGSDNodes::CountResults(SDNode *Node) {
unsigned N = Node->getNumValues();
while (N && Node->getValueType(N - 1) == MVT::Flag)
--N;
if (N && Node->getValueType(N - 1) == MVT::Other)
--N; // Skip over chain result.
return N;
}
/// CountOperands - The inputs to target nodes have any actual inputs first,
/// followed by special operands that describe memory references, then an
/// optional chain operand, then an optional flag operand. Compute the number
/// of actual operands that will go into the resulting MachineInstr.
unsigned ScheduleDAGSDNodes::CountOperands(SDNode *Node) {
unsigned N = ComputeMemOperandsEnd(Node);
while (N && isa<MemOperandSDNode>(Node->getOperand(N - 1).getNode()))
--N; // Ignore MEMOPERAND nodes
return N;
}
/// ComputeMemOperandsEnd - Find the index one past the last MemOperandSDNode
/// operand
unsigned ScheduleDAGSDNodes::ComputeMemOperandsEnd(SDNode *Node) {
unsigned N = Node->getNumOperands();
while (N && Node->getOperand(N - 1).getValueType() == MVT::Flag)
--N;
if (N && Node->getOperand(N - 1).getValueType() == MVT::Other)
--N; // Ignore chain if it exists.
return N;
}
void ScheduleDAGSDNodes::dumpNode(const SUnit *SU) const {
if (!SU->getNode()) {
cerr << "PHYS REG COPY\n";
return;
}
SU->getNode()->dump(DAG);
cerr << "\n";
SmallVector<SDNode *, 4> FlaggedNodes;
for (SDNode *N = SU->getNode()->getFlaggedNode(); N; N = N->getFlaggedNode())
FlaggedNodes.push_back(N);
while (!FlaggedNodes.empty()) {
cerr << " ";
FlaggedNodes.back()->dump(DAG);
cerr << "\n";
FlaggedNodes.pop_back();
}
}
|