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
|
//===-- SIISelLowering.cpp - SI DAG Lowering Implementation ---------------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// Most of the DAG lowering is handled in AMDGPUISelLowering.cpp. This file is
// mostly EmitInstrWithCustomInserter().
//
//===----------------------------------------------------------------------===//
#include "SIISelLowering.h"
#include "SIInstrInfo.h"
#include "SIRegisterInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
using namespace llvm;
SITargetLowering::SITargetLowering(TargetMachine &TM) :
AMDGPUTargetLowering(TM),
TII(static_cast<const SIInstrInfo*>(TM.getInstrInfo()))
{
addRegisterClass(MVT::v4f32, &AMDGPU::VReg_128RegClass);
addRegisterClass(MVT::f32, &AMDGPU::VReg_32RegClass);
addRegisterClass(MVT::i32, &AMDGPU::VReg_32RegClass);
addRegisterClass(MVT::i64, &AMDGPU::VReg_64RegClass);
addRegisterClass(MVT::v4i32, &AMDGPU::SReg_128RegClass);
addRegisterClass(MVT::v8i32, &AMDGPU::SReg_256RegClass);
computeRegisterProperties();
setOperationAction(ISD::ADD, MVT::i64, Legal);
setOperationAction(ISD::ADD, MVT::i32, Legal);
setOperationAction(ISD::BR_CC, MVT::i32, Custom);
}
MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter(
MachineInstr * MI, MachineBasicBlock * BB) const
{
const TargetInstrInfo * TII = getTargetMachine().getInstrInfo();
MachineRegisterInfo & MRI = BB->getParent()->getRegInfo();
MachineBasicBlock::iterator I = MI;
if (TII->get(MI->getOpcode()).TSFlags & SIInstrFlags::NEED_WAIT) {
AppendS_WAITCNT(MI, *BB, llvm::next(I));
return BB;
}
switch (MI->getOpcode()) {
default:
return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
case AMDGPU::CLAMP_SI:
BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::V_MOV_B32_e64))
.addOperand(MI->getOperand(0))
.addOperand(MI->getOperand(1))
// VSRC1-2 are unused, but we still need to fill all the
// operand slots, so we just reuse the VSRC0 operand
.addOperand(MI->getOperand(1))
.addOperand(MI->getOperand(1))
.addImm(0) // ABS
.addImm(1) // CLAMP
.addImm(0) // OMOD
.addImm(0); // NEG
MI->eraseFromParent();
break;
case AMDGPU::FABS_SI:
BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::V_MOV_B32_e64))
.addOperand(MI->getOperand(0))
.addOperand(MI->getOperand(1))
// VSRC1-2 are unused, but we still need to fill all the
// operand slots, so we just reuse the VSRC0 operand
.addOperand(MI->getOperand(1))
.addOperand(MI->getOperand(1))
.addImm(1) // ABS
.addImm(0) // CLAMP
.addImm(0) // OMOD
.addImm(0); // NEG
MI->eraseFromParent();
break;
case AMDGPU::SI_INTERP:
LowerSI_INTERP(MI, *BB, I, MRI);
break;
case AMDGPU::SI_INTERP_CONST:
LowerSI_INTERP_CONST(MI, *BB, I);
break;
case AMDGPU::SI_V_CNDLT:
LowerSI_V_CNDLT(MI, *BB, I, MRI);
break;
case AMDGPU::USE_SGPR_32:
case AMDGPU::USE_SGPR_64:
lowerUSE_SGPR(MI, BB->getParent(), MRI);
MI->eraseFromParent();
break;
case AMDGPU::VS_LOAD_BUFFER_INDEX:
addLiveIn(MI, BB->getParent(), MRI, TII, AMDGPU::VGPR0);
MI->eraseFromParent();
break;
}
return BB;
}
void SITargetLowering::AppendS_WAITCNT(MachineInstr *MI, MachineBasicBlock &BB,
MachineBasicBlock::iterator I) const
{
BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::S_WAITCNT))
.addImm(0);
}
void SITargetLowering::LowerSI_INTERP(MachineInstr *MI, MachineBasicBlock &BB,
MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const
{
unsigned tmp = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
MachineOperand dst = MI->getOperand(0);
MachineOperand iReg = MI->getOperand(1);
MachineOperand jReg = MI->getOperand(2);
MachineOperand attr_chan = MI->getOperand(3);
MachineOperand attr = MI->getOperand(4);
MachineOperand params = MI->getOperand(5);
BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::S_MOV_B32))
.addReg(AMDGPU::M0)
.addOperand(params);
BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::V_INTERP_P1_F32), tmp)
.addOperand(iReg)
.addOperand(attr_chan)
.addOperand(attr);
BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::V_INTERP_P2_F32))
.addOperand(dst)
.addReg(tmp)
.addOperand(jReg)
.addOperand(attr_chan)
.addOperand(attr);
MI->eraseFromParent();
}
void SITargetLowering::LowerSI_INTERP_CONST(MachineInstr *MI,
MachineBasicBlock &BB, MachineBasicBlock::iterator I) const
{
MachineOperand dst = MI->getOperand(0);
MachineOperand attr_chan = MI->getOperand(1);
MachineOperand attr = MI->getOperand(2);
MachineOperand params = MI->getOperand(3);
BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::S_MOV_B32))
.addReg(AMDGPU::M0)
.addOperand(params);
BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::V_INTERP_MOV_F32))
.addOperand(dst)
.addOperand(attr_chan)
.addOperand(attr);
MI->eraseFromParent();
}
void SITargetLowering::LowerSI_V_CNDLT(MachineInstr *MI, MachineBasicBlock &BB,
MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const
{
BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::V_CMP_LT_F32_e32))
.addOperand(MI->getOperand(1))
.addReg(AMDGPU::SREG_LIT_0);
BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::V_CNDMASK_B32))
.addOperand(MI->getOperand(0))
.addOperand(MI->getOperand(2))
.addOperand(MI->getOperand(3));
MI->eraseFromParent();
}
void SITargetLowering::lowerUSE_SGPR(MachineInstr *MI,
MachineFunction * MF, MachineRegisterInfo & MRI) const
{
const TargetInstrInfo * TII = getTargetMachine().getInstrInfo();
unsigned dstReg = MI->getOperand(0).getReg();
int64_t newIndex = MI->getOperand(1).getImm();
const TargetRegisterClass * dstClass = MRI.getRegClass(dstReg);
unsigned DwordWidth = dstClass->getSize() / 4;
assert(newIndex % DwordWidth == 0 && "USER_SGPR not properly aligned");
newIndex = newIndex / DwordWidth;
unsigned newReg = dstClass->getRegister(newIndex);
addLiveIn(MI, MF, MRI, TII, newReg);
}
//===----------------------------------------------------------------------===//
// Custom DAG Lowering Operations
//===----------------------------------------------------------------------===//
SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const
{
switch (Op.getOpcode()) {
default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
case ISD::BR_CC: return LowerBR_CC(Op, DAG);
}
}
SDValue SITargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const
{
SDValue Chain = Op.getOperand(0);
SDValue CC = Op.getOperand(1);
SDValue LHS = Op.getOperand(2);
SDValue RHS = Op.getOperand(3);
SDValue JumpT = Op.getOperand(4);
SDValue CmpValue;
SDValue Result;
CmpValue = DAG.getNode(
ISD::SETCC,
Op.getDebugLoc(),
MVT::i1,
LHS, RHS,
CC);
Result = DAG.getNode(
AMDILISD::BRANCH_COND,
CmpValue.getDebugLoc(),
MVT::Other, Chain,
JumpT, CmpValue);
return Result;
}
|
