[ARM] VQADD instructions
[llvm-complete.git] / lib / Target / AMDGPU / SILowerSGPRSpills.cpp
blob714d403a3e8fdbe902879c9e34f38305662fa4a6
1 //===-- SILowerSGPRSPills.cpp ---------------------------------------------===//
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 // Handle SGPR spills. This pass takes the place of PrologEpilogInserter for all
10 // SGPR spills, so must insert CSR SGPR spills as well as expand them.
12 // This pass must never create new SGPR virtual registers.
14 // FIXME: Must stop RegScavenger spills in later passes.
16 //===----------------------------------------------------------------------===//
18 #include "AMDGPU.h"
19 #include "AMDGPUSubtarget.h"
20 #include "SIInstrInfo.h"
21 #include "SIMachineFunctionInfo.h"
22 #include "llvm/CodeGen/LiveIntervals.h"
23 #include "llvm/CodeGen/MachineBasicBlock.h"
24 #include "llvm/CodeGen/MachineFunction.h"
25 #include "llvm/CodeGen/MachineFunctionPass.h"
26 #include "llvm/CodeGen/MachineInstr.h"
27 #include "llvm/CodeGen/MachineInstrBuilder.h"
28 #include "llvm/CodeGen/MachineOperand.h"
29 #include "llvm/CodeGen/VirtRegMap.h"
30 #include "llvm/Target/TargetMachine.h"
32 using namespace llvm;
34 #define DEBUG_TYPE "si-lower-sgpr-spills"
36 using MBBVector = SmallVector<MachineBasicBlock *, 4>;
38 namespace {
40 static cl::opt<bool> EnableSpillVGPRToAGPR(
41 "amdgpu-spill-vgpr-to-agpr",
42 cl::desc("Enable spilling VGPRs to AGPRs"),
43 cl::ReallyHidden,
44 cl::init(true));
46 class SILowerSGPRSpills : public MachineFunctionPass {
47 private:
48 const SIRegisterInfo *TRI = nullptr;
49 const SIInstrInfo *TII = nullptr;
50 VirtRegMap *VRM = nullptr;
51 LiveIntervals *LIS = nullptr;
53 // Save and Restore blocks of the current function. Typically there is a
54 // single save block, unless Windows EH funclets are involved.
55 MBBVector SaveBlocks;
56 MBBVector RestoreBlocks;
58 public:
59 static char ID;
61 SILowerSGPRSpills() : MachineFunctionPass(ID) {}
63 void calculateSaveRestoreBlocks(MachineFunction &MF);
64 bool spillCalleeSavedRegs(MachineFunction &MF);
66 bool runOnMachineFunction(MachineFunction &MF) override;
68 void getAnalysisUsage(AnalysisUsage &AU) const override {
69 AU.setPreservesAll();
70 MachineFunctionPass::getAnalysisUsage(AU);
74 } // end anonymous namespace
76 char SILowerSGPRSpills::ID = 0;
78 INITIALIZE_PASS_BEGIN(SILowerSGPRSpills, DEBUG_TYPE,
79 "SI lower SGPR spill instructions", false, false)
80 INITIALIZE_PASS_DEPENDENCY(VirtRegMap)
81 INITIALIZE_PASS_END(SILowerSGPRSpills, DEBUG_TYPE,
82 "SI lower SGPR spill instructions", false, false)
84 char &llvm::SILowerSGPRSpillsID = SILowerSGPRSpills::ID;
86 /// Insert restore code for the callee-saved registers used in the function.
87 static void insertCSRSaves(MachineBasicBlock &SaveBlock,
88 ArrayRef<CalleeSavedInfo> CSI,
89 LiveIntervals *LIS) {
90 MachineFunction &MF = *SaveBlock.getParent();
91 const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
92 const TargetFrameLowering *TFI = MF.getSubtarget().getFrameLowering();
93 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
95 MachineBasicBlock::iterator I = SaveBlock.begin();
96 if (!TFI->spillCalleeSavedRegisters(SaveBlock, I, CSI, TRI)) {
97 for (const CalleeSavedInfo &CS : CSI) {
98 // Insert the spill to the stack frame.
99 unsigned Reg = CS.getReg();
101 MachineInstrSpan MIS(I, &SaveBlock);
102 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg);
104 TII.storeRegToStackSlot(SaveBlock, I, Reg, true, CS.getFrameIdx(), RC,
105 TRI);
107 if (LIS) {
108 assert(std::distance(MIS.begin(), I) == 1);
109 MachineInstr &Inst = *std::prev(I);
111 LIS->InsertMachineInstrInMaps(Inst);
112 LIS->removeAllRegUnitsForPhysReg(Reg);
118 /// Insert restore code for the callee-saved registers used in the function.
119 static void insertCSRRestores(MachineBasicBlock &RestoreBlock,
120 std::vector<CalleeSavedInfo> &CSI,
121 LiveIntervals *LIS) {
122 MachineFunction &MF = *RestoreBlock.getParent();
123 const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
124 const TargetFrameLowering *TFI = MF.getSubtarget().getFrameLowering();
125 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
127 // Restore all registers immediately before the return and any
128 // terminators that precede it.
129 MachineBasicBlock::iterator I = RestoreBlock.getFirstTerminator();
131 // FIXME: Just emit the readlane/writelane directly
132 if (!TFI->restoreCalleeSavedRegisters(RestoreBlock, I, CSI, TRI)) {
133 for (const CalleeSavedInfo &CI : reverse(CSI)) {
134 unsigned Reg = CI.getReg();
135 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg);
137 TII.loadRegFromStackSlot(RestoreBlock, I, Reg, CI.getFrameIdx(), RC, TRI);
138 assert(I != RestoreBlock.begin() &&
139 "loadRegFromStackSlot didn't insert any code!");
140 // Insert in reverse order. loadRegFromStackSlot can insert
141 // multiple instructions.
143 if (LIS) {
144 MachineInstr &Inst = *std::prev(I);
145 LIS->InsertMachineInstrInMaps(Inst);
146 LIS->removeAllRegUnitsForPhysReg(Reg);
152 /// Compute the sets of entry and return blocks for saving and restoring
153 /// callee-saved registers, and placing prolog and epilog code.
154 void SILowerSGPRSpills::calculateSaveRestoreBlocks(MachineFunction &MF) {
155 const MachineFrameInfo &MFI = MF.getFrameInfo();
157 // Even when we do not change any CSR, we still want to insert the
158 // prologue and epilogue of the function.
159 // So set the save points for those.
161 // Use the points found by shrink-wrapping, if any.
162 if (MFI.getSavePoint()) {
163 SaveBlocks.push_back(MFI.getSavePoint());
164 assert(MFI.getRestorePoint() && "Both restore and save must be set");
165 MachineBasicBlock *RestoreBlock = MFI.getRestorePoint();
166 // If RestoreBlock does not have any successor and is not a return block
167 // then the end point is unreachable and we do not need to insert any
168 // epilogue.
169 if (!RestoreBlock->succ_empty() || RestoreBlock->isReturnBlock())
170 RestoreBlocks.push_back(RestoreBlock);
171 return;
174 // Save refs to entry and return blocks.
175 SaveBlocks.push_back(&MF.front());
176 for (MachineBasicBlock &MBB : MF) {
177 if (MBB.isEHFuncletEntry())
178 SaveBlocks.push_back(&MBB);
179 if (MBB.isReturnBlock())
180 RestoreBlocks.push_back(&MBB);
184 bool SILowerSGPRSpills::spillCalleeSavedRegs(MachineFunction &MF) {
185 MachineRegisterInfo &MRI = MF.getRegInfo();
186 const Function &F = MF.getFunction();
187 const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
188 const SIFrameLowering *TFI = ST.getFrameLowering();
189 MachineFrameInfo &MFI = MF.getFrameInfo();
190 RegScavenger *RS = nullptr;
192 // Determine which of the registers in the callee save list should be saved.
193 BitVector SavedRegs;
194 TFI->determineCalleeSavesSGPR(MF, SavedRegs, RS);
196 // Add the code to save and restore the callee saved registers.
197 if (!F.hasFnAttribute(Attribute::Naked)) {
198 // FIXME: This is a lie. The CalleeSavedInfo is incomplete, but this is
199 // necessary for verifier liveness checks.
200 MFI.setCalleeSavedInfoValid(true);
202 std::vector<CalleeSavedInfo> CSI;
203 const MCPhysReg *CSRegs = MRI.getCalleeSavedRegs();
205 for (unsigned I = 0; CSRegs[I]; ++I) {
206 unsigned Reg = CSRegs[I];
207 if (SavedRegs.test(Reg)) {
208 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg);
209 int JunkFI = MFI.CreateStackObject(TRI->getSpillSize(*RC),
210 TRI->getSpillAlignment(*RC),
211 true);
213 CSI.push_back(CalleeSavedInfo(Reg, JunkFI));
217 if (!CSI.empty()) {
218 for (MachineBasicBlock *SaveBlock : SaveBlocks)
219 insertCSRSaves(*SaveBlock, CSI, LIS);
221 for (MachineBasicBlock *RestoreBlock : RestoreBlocks)
222 insertCSRRestores(*RestoreBlock, CSI, LIS);
223 return true;
227 return false;
230 bool SILowerSGPRSpills::runOnMachineFunction(MachineFunction &MF) {
231 const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
232 TII = ST.getInstrInfo();
233 TRI = &TII->getRegisterInfo();
235 VRM = getAnalysisIfAvailable<VirtRegMap>();
237 assert(SaveBlocks.empty() && RestoreBlocks.empty());
239 // First, expose any CSR SGPR spills. This is mostly the same as what PEI
240 // does, but somewhat simpler.
241 calculateSaveRestoreBlocks(MF);
242 bool HasCSRs = spillCalleeSavedRegs(MF);
244 MachineFrameInfo &MFI = MF.getFrameInfo();
245 if (!MFI.hasStackObjects() && !HasCSRs) {
246 SaveBlocks.clear();
247 RestoreBlocks.clear();
248 return false;
251 MachineRegisterInfo &MRI = MF.getRegInfo();
252 SIMachineFunctionInfo *FuncInfo = MF.getInfo<SIMachineFunctionInfo>();
253 const bool SpillVGPRToAGPR = ST.hasMAIInsts() && FuncInfo->hasSpilledVGPRs()
254 && EnableSpillVGPRToAGPR;
256 bool MadeChange = false;
258 const bool SpillToAGPR = EnableSpillVGPRToAGPR && ST.hasMAIInsts();
260 // TODO: CSR VGPRs will never be spilled to AGPRs. These can probably be
261 // handled as SpilledToReg in regular PrologEpilogInserter.
262 if ((TRI->spillSGPRToVGPR() && (HasCSRs || FuncInfo->hasSpilledSGPRs())) ||
263 SpillVGPRToAGPR) {
264 // Process all SGPR spills before frame offsets are finalized. Ideally SGPRs
265 // are spilled to VGPRs, in which case we can eliminate the stack usage.
267 // This operates under the assumption that only other SGPR spills are users
268 // of the frame index.
269 for (MachineBasicBlock &MBB : MF) {
270 MachineBasicBlock::iterator Next;
271 for (auto I = MBB.begin(), E = MBB.end(); I != E; I = Next) {
272 MachineInstr &MI = *I;
273 Next = std::next(I);
275 if (SpillToAGPR && TII->isVGPRSpill(MI)) {
276 // Try to eliminate stack used by VGPR spills before frame
277 // finalization.
278 unsigned FIOp = AMDGPU::getNamedOperandIdx(MI.getOpcode(),
279 AMDGPU::OpName::vaddr);
280 int FI = MI.getOperand(FIOp).getIndex();
281 Register VReg =
282 TII->getNamedOperand(MI, AMDGPU::OpName::vdata)->getReg();
283 if (FuncInfo->allocateVGPRSpillToAGPR(MF, FI,
284 TRI->isAGPR(MRI, VReg))) {
285 TRI->eliminateFrameIndex(MI, 0, FIOp, nullptr);
286 continue;
290 if (!TII->isSGPRSpill(MI))
291 continue;
293 int FI = TII->getNamedOperand(MI, AMDGPU::OpName::addr)->getIndex();
294 assert(MFI.getStackID(FI) == TargetStackID::SGPRSpill);
295 if (FuncInfo->allocateSGPRSpillToVGPR(MF, FI)) {
296 bool Spilled = TRI->eliminateSGPRToVGPRSpillFrameIndex(MI, FI, nullptr);
297 (void)Spilled;
298 assert(Spilled && "failed to spill SGPR to VGPR when allocated");
303 for (MachineBasicBlock &MBB : MF) {
304 for (auto SSpill : FuncInfo->getSGPRSpillVGPRs())
305 MBB.addLiveIn(SSpill.VGPR);
307 for (MCPhysReg Reg : FuncInfo->getVGPRSpillAGPRs())
308 MBB.addLiveIn(Reg);
310 for (MCPhysReg Reg : FuncInfo->getAGPRSpillVGPRs())
311 MBB.addLiveIn(Reg);
313 MBB.sortUniqueLiveIns();
316 MadeChange = true;
319 SaveBlocks.clear();
320 RestoreBlocks.clear();
322 return MadeChange;