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
|
//===-- SIPrepareScratchRegs.cpp - Use predicates for control flow --------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
/// \file
///
/// This pass loads scratch pointer and scratch offset into a register or a
/// frame index which can be used anywhere in the program. These values will
/// be used for spilling VGPRs.
///
//===----------------------------------------------------------------------===//
#include "AMDGPU.h"
#include "AMDGPUSubtarget.h"
#include "SIDefines.h"
#include "SIInstrInfo.h"
#include "SIMachineFunctionInfo.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/RegisterScavenging.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/LLVMContext.h"
using namespace llvm;
namespace {
class SIPrepareScratchRegs : public MachineFunctionPass {
private:
static char ID;
public:
SIPrepareScratchRegs() : MachineFunctionPass(ID) { }
bool runOnMachineFunction(MachineFunction &MF) override;
const char *getPassName() const override {
return "SI prepare scratch registers";
}
};
} // End anonymous namespace
char SIPrepareScratchRegs::ID = 0;
FunctionPass *llvm::createSIPrepareScratchRegs() {
return new SIPrepareScratchRegs();
}
bool SIPrepareScratchRegs::runOnMachineFunction(MachineFunction &MF) {
SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
const SIInstrInfo *TII =
static_cast<const SIInstrInfo *>(MF.getSubtarget().getInstrInfo());
const SIRegisterInfo *TRI = &TII->getRegisterInfo();
MachineRegisterInfo &MRI = MF.getRegInfo();
MachineFrameInfo *FrameInfo = MF.getFrameInfo();
MachineBasicBlock *Entry = MF.begin();
MachineBasicBlock::iterator I = Entry->begin();
DebugLoc DL = I->getDebugLoc();
// FIXME: If we don't have enough VGPRs for SGPR spilling we will need to
// run this pass.
if (!MFI->hasSpilledVGPRs())
return false;
unsigned ScratchPtrPreloadReg =
TRI->getPreloadedValue(MF, SIRegisterInfo::SCRATCH_PTR);
unsigned ScratchOffsetPreloadReg =
TRI->getPreloadedValue(MF, SIRegisterInfo::SCRATCH_WAVE_OFFSET);
if (!Entry->isLiveIn(ScratchPtrPreloadReg))
Entry->addLiveIn(ScratchPtrPreloadReg);
if (!Entry->isLiveIn(ScratchOffsetPreloadReg))
Entry->addLiveIn(ScratchOffsetPreloadReg);
// Load the scratch offset.
unsigned ScratchOffsetReg =
TRI->findUnusedRegister(MRI, &AMDGPU::SGPR_32RegClass);
int ScratchOffsetFI = -1;
if (ScratchOffsetReg != AMDGPU::NoRegister) {
// Found an SGPR to use
MRI.setPhysRegUsed(ScratchOffsetReg);
BuildMI(*Entry, I, DL, TII->get(AMDGPU::S_MOV_B32), ScratchOffsetReg)
.addReg(ScratchOffsetPreloadReg);
} else {
// No SGPR is available, we must spill.
ScratchOffsetFI = FrameInfo->CreateSpillStackObject(4,4);
BuildMI(*Entry, I, DL, TII->get(AMDGPU::SI_SPILL_S32_SAVE))
.addReg(ScratchOffsetPreloadReg)
.addFrameIndex(ScratchOffsetFI)
.addReg(AMDGPU::SGPR0_SGPR1_SGPR2_SGPR3, RegState::Undef)
.addReg(AMDGPU::SGPR0, RegState::Undef);
}
// Now that we have the scratch pointer and offset values, we need to
// add them to all the SI_SPILL_V* instructions.
RegScavenger RS;
unsigned ScratchRsrcFI = FrameInfo->CreateSpillStackObject(16, 4);
RS.addScavengingFrameIndex(ScratchRsrcFI);
for (MachineFunction::iterator BI = MF.begin(), BE = MF.end();
BI != BE; ++BI) {
MachineBasicBlock &MBB = *BI;
// Add the scratch offset reg as a live-in so that the register scavenger
// doesn't re-use it.
if (!MBB.isLiveIn(ScratchOffsetReg) &&
ScratchOffsetReg != AMDGPU::NoRegister)
MBB.addLiveIn(ScratchOffsetReg);
RS.enterBasicBlock(&MBB);
for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end();
I != E; ++I) {
MachineInstr &MI = *I;
RS.forward(I);
DebugLoc DL = MI.getDebugLoc();
switch(MI.getOpcode()) {
default: break;
case AMDGPU::SI_SPILL_V512_SAVE:
case AMDGPU::SI_SPILL_V256_SAVE:
case AMDGPU::SI_SPILL_V128_SAVE:
case AMDGPU::SI_SPILL_V96_SAVE:
case AMDGPU::SI_SPILL_V64_SAVE:
case AMDGPU::SI_SPILL_V32_SAVE:
case AMDGPU::SI_SPILL_V32_RESTORE:
case AMDGPU::SI_SPILL_V64_RESTORE:
case AMDGPU::SI_SPILL_V128_RESTORE:
case AMDGPU::SI_SPILL_V256_RESTORE:
case AMDGPU::SI_SPILL_V512_RESTORE:
// Scratch resource
unsigned ScratchRsrcReg =
RS.scavengeRegister(&AMDGPU::SReg_128RegClass, 0);
uint64_t Rsrc = AMDGPU::RSRC_DATA_FORMAT | AMDGPU::RSRC_TID_ENABLE |
0xffffffff; // Size
unsigned Rsrc0 = TRI->getSubReg(ScratchRsrcReg, AMDGPU::sub0);
unsigned Rsrc1 = TRI->getSubReg(ScratchRsrcReg, AMDGPU::sub1);
unsigned Rsrc2 = TRI->getSubReg(ScratchRsrcReg, AMDGPU::sub2);
unsigned Rsrc3 = TRI->getSubReg(ScratchRsrcReg, AMDGPU::sub3);
BuildMI(MBB, I, DL, TII->get(AMDGPU::S_MOV_B32), Rsrc0)
.addExternalSymbol("SCRATCH_RSRC_DWORD0")
.addReg(ScratchRsrcReg, RegState::ImplicitDefine);
BuildMI(MBB, I, DL, TII->get(AMDGPU::S_MOV_B32), Rsrc1)
.addExternalSymbol("SCRATCH_RSRC_DWORD1")
.addReg(ScratchRsrcReg, RegState::ImplicitDefine);
BuildMI(MBB, I, DL, TII->get(AMDGPU::S_MOV_B32), Rsrc2)
.addImm(Rsrc & 0xffffffff)
.addReg(ScratchRsrcReg, RegState::ImplicitDefine);
BuildMI(MBB, I, DL, TII->get(AMDGPU::S_MOV_B32), Rsrc3)
.addImm(Rsrc >> 32)
.addReg(ScratchRsrcReg, RegState::ImplicitDefine);
// Scratch Offset
if (ScratchOffsetReg == AMDGPU::NoRegister) {
ScratchOffsetReg = RS.scavengeRegister(&AMDGPU::SGPR_32RegClass, 0);
BuildMI(MBB, I, DL, TII->get(AMDGPU::SI_SPILL_S32_RESTORE),
ScratchOffsetReg)
.addFrameIndex(ScratchOffsetFI)
.addReg(AMDGPU::SGPR0_SGPR1_SGPR2_SGPR3, RegState::Undef)
.addReg(AMDGPU::SGPR0, RegState::Undef);
} else if (!MBB.isLiveIn(ScratchOffsetReg)) {
MBB.addLiveIn(ScratchOffsetReg);
}
if (ScratchRsrcReg == AMDGPU::NoRegister ||
ScratchOffsetReg == AMDGPU::NoRegister) {
LLVMContext &Ctx = MF.getFunction()->getContext();
Ctx.emitError("ran out of SGPRs for spilling VGPRs");
ScratchRsrcReg = AMDGPU::SGPR0;
ScratchOffsetReg = AMDGPU::SGPR0;
}
MI.getOperand(2).setReg(ScratchRsrcReg);
MI.getOperand(2).setIsKill(true);
MI.getOperand(2).setIsUndef(false);
MI.getOperand(3).setReg(ScratchOffsetReg);
MI.getOperand(3).setIsUndef(false);
MI.getOperand(3).setIsKill(false);
MI.addOperand(MachineOperand::CreateReg(Rsrc0, false, true, true));
MI.addOperand(MachineOperand::CreateReg(Rsrc1, false, true, true));
MI.addOperand(MachineOperand::CreateReg(Rsrc2, false, true, true));
MI.addOperand(MachineOperand::CreateReg(Rsrc3, false, true, true));
break;
}
}
}
return true;
}
|
