1 //===-- SIModeRegister.cpp - Mode Register --------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 /// This pass inserts changes to the Mode register settings as required.
11 /// Note that currently it only deals with the Double Precision Floating Point
12 /// rounding mode setting, but is intended to be generic enough to be easily
15 //===----------------------------------------------------------------------===//
18 #include "AMDGPUInstrInfo.h"
19 #include "AMDGPUSubtarget.h"
20 #include "SIInstrInfo.h"
21 #include "SIMachineFunctionInfo.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/CodeGen/MachineFunctionPass.h"
24 #include "llvm/CodeGen/MachineInstrBuilder.h"
25 #include "llvm/CodeGen/MachineRegisterInfo.h"
26 #include "llvm/IR/Constants.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/LLVMContext.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/Support/raw_ostream.h"
31 #include "llvm/Target/TargetMachine.h"
34 #define DEBUG_TYPE "si-mode-register"
36 STATISTIC(NumSetregInserted, "Number of setreg of mode register inserted.");
41 // Mask is a bitmask where a '1' indicates the corresponding Mode bit has a
46 Status() : Mask(0), Mode(0){};
48 Status(unsigned Mask, unsigned Mode) : Mask(Mask), Mode(Mode) {
52 // merge two status values such that only values that don't conflict are
54 Status merge(const Status &S) const {
55 return Status((Mask | S.Mask), ((Mode & ~S.Mask) | (S.Mode & S.Mask)));
58 // merge an unknown value by using the unknown value's mask to remove bits
60 Status mergeUnknown(unsigned newMask) {
61 return Status(Mask & ~newMask, Mode & ~newMask);
64 // intersect two Status values to produce a mode and mask that is a subset
66 Status intersect(const Status &S) const {
67 unsigned NewMask = (Mask & S.Mask) & (Mode ^ ~S.Mode);
68 unsigned NewMode = (Mode & NewMask);
69 return Status(NewMask, NewMode);
72 // produce the delta required to change the Mode to the required Mode
73 Status delta(const Status &S) const {
74 return Status((S.Mask & (Mode ^ S.Mode)) | (~Mask & S.Mask), S.Mode);
77 bool operator==(const Status &S) const {
78 return (Mask == S.Mask) && (Mode == S.Mode);
81 bool operator!=(const Status &S) const { return !(*this == S); }
83 bool isCompatible(Status &S) {
84 return ((Mask & S.Mask) == S.Mask) && ((Mode & S.Mask) == S.Mode);
87 bool isCombinable(Status &S) {
88 return !(Mask & S.Mask) || isCompatible(S);
94 // The Status that represents the mode register settings required by the
95 // FirstInsertionPoint (if any) in this block. Calculated in Phase 1.
98 // The Status that represents the net changes to the Mode register made by
99 // this block, Calculated in Phase 1.
102 // The Status that represents the mode register settings on exit from this
103 // block. Calculated in Phase 2.
106 // The Status that represents the intersection of exit Mode register settings
107 // from all predecessor blocks. Calculated in Phase 2, and used by Phase 3.
110 // In Phase 1 we record the first instruction that has a mode requirement,
111 // which is used in Phase 3 if we need to insert a mode change.
112 MachineInstr *FirstInsertionPoint;
114 BlockData() : FirstInsertionPoint(nullptr) {};
119 class SIModeRegister : public MachineFunctionPass {
123 std::vector<std::unique_ptr<BlockData>> BlockInfo;
124 std::queue<MachineBasicBlock *> Phase2List;
126 // The default mode register setting currently only caters for the floating
127 // point double precision rounding mode.
128 // We currently assume the default rounding mode is Round to Nearest
129 // NOTE: this should come from a per function rounding mode setting once such
131 unsigned DefaultMode = FP_ROUND_ROUND_TO_NEAREST;
132 Status DefaultStatus =
133 Status(FP_ROUND_MODE_DP(0x3), FP_ROUND_MODE_DP(DefaultMode));
136 SIModeRegister() : MachineFunctionPass(ID) {}
138 bool runOnMachineFunction(MachineFunction &MF) override;
140 void getAnalysisUsage(AnalysisUsage &AU) const override {
141 AU.setPreservesCFG();
142 MachineFunctionPass::getAnalysisUsage(AU);
145 void processBlockPhase1(MachineBasicBlock &MBB, const SIInstrInfo *TII);
147 void processBlockPhase2(MachineBasicBlock &MBB, const SIInstrInfo *TII);
149 void processBlockPhase3(MachineBasicBlock &MBB, const SIInstrInfo *TII);
151 Status getInstructionMode(MachineInstr &MI, const SIInstrInfo *TII);
153 void insertSetreg(MachineBasicBlock &MBB, MachineInstr *I,
154 const SIInstrInfo *TII, Status InstrMode);
156 } // End anonymous namespace.
158 INITIALIZE_PASS(SIModeRegister, DEBUG_TYPE,
159 "Insert required mode register values", false, false)
161 char SIModeRegister::ID = 0;
163 char &llvm::SIModeRegisterID = SIModeRegister::ID;
165 FunctionPass *llvm::createSIModeRegisterPass() { return new SIModeRegister(); }
167 // Determine the Mode register setting required for this instruction.
168 // Instructions which don't use the Mode register return a null Status.
169 // Note this currently only deals with instructions that use the floating point
170 // double precision setting.
171 Status SIModeRegister::getInstructionMode(MachineInstr &MI,
172 const SIInstrInfo *TII) {
173 if (TII->usesFPDPRounding(MI)) {
174 switch (MI.getOpcode()) {
175 case AMDGPU::V_INTERP_P1LL_F16:
176 case AMDGPU::V_INTERP_P1LV_F16:
177 case AMDGPU::V_INTERP_P2_F16:
178 // f16 interpolation instructions need double precision round to zero
179 return Status(FP_ROUND_MODE_DP(3),
180 FP_ROUND_MODE_DP(FP_ROUND_ROUND_TO_ZERO));
182 return DefaultStatus;
188 // Insert a setreg instruction to update the Mode register.
189 // It is possible (though unlikely) for an instruction to require a change to
190 // the value of disjoint parts of the Mode register when we don't know the
191 // value of the intervening bits. In that case we need to use more than one
192 // setreg instruction.
193 void SIModeRegister::insertSetreg(MachineBasicBlock &MBB, MachineInstr *MI,
194 const SIInstrInfo *TII, Status InstrMode) {
195 while (InstrMode.Mask) {
196 unsigned Offset = countTrailingZeros<unsigned>(InstrMode.Mask);
197 unsigned Width = countTrailingOnes<unsigned>(InstrMode.Mask >> Offset);
198 unsigned Value = (InstrMode.Mode >> Offset) & ((1 << Width) - 1);
199 BuildMI(MBB, MI, 0, TII->get(AMDGPU::S_SETREG_IMM32_B32))
201 .addImm(((Width - 1) << AMDGPU::Hwreg::WIDTH_M1_SHIFT_) |
202 (Offset << AMDGPU::Hwreg::OFFSET_SHIFT_) |
203 (AMDGPU::Hwreg::ID_MODE << AMDGPU::Hwreg::ID_SHIFT_));
205 InstrMode.Mask &= ~(((1 << Width) - 1) << Offset);
209 // In Phase 1 we iterate through the instructions of the block and for each
210 // instruction we get its mode usage. If the instruction uses the Mode register
212 // - update the Change status, which tracks the changes to the Mode register
213 // made by this block
214 // - if this instruction's requirements are compatible with the current setting
215 // of the Mode register we merge the modes
216 // - if it isn't compatible and an InsertionPoint isn't set, then we set the
217 // InsertionPoint to the current instruction, and we remember the current
219 // - if it isn't compatible and InsertionPoint is set we insert a seteg before
220 // that instruction (unless this instruction forms part of the block's
221 // entry requirements in which case the insertion is deferred until Phase 3
222 // when predecessor exit values are known), and move the insertion point to
224 // - if this is a setreg instruction we treat it as an incompatible instruction.
225 // This is sub-optimal but avoids some nasty corner cases, and is expected to
226 // occur very rarely.
227 // - on exit we have set the Require, Change, and initial Exit modes.
228 void SIModeRegister::processBlockPhase1(MachineBasicBlock &MBB,
229 const SIInstrInfo *TII) {
230 auto NewInfo = llvm::make_unique<BlockData>();
231 MachineInstr *InsertionPoint = nullptr;
232 // RequirePending is used to indicate whether we are collecting the initial
233 // requirements for the block, and need to defer the first InsertionPoint to
234 // Phase 3. It is set to false once we have set FirstInsertionPoint, or when
235 // we discover an explict setreg that means this block doesn't have any
236 // initial requirements.
237 bool RequirePending = true;
239 for (MachineInstr &MI : MBB) {
240 Status InstrMode = getInstructionMode(MI, TII);
241 if ((MI.getOpcode() == AMDGPU::S_SETREG_B32) ||
242 (MI.getOpcode() == AMDGPU::S_SETREG_IMM32_B32)) {
243 // We preserve any explicit mode register setreg instruction we encounter,
244 // as we assume it has been inserted by a higher authority (this is
245 // likely to be a very rare occurrence).
246 unsigned Dst = TII->getNamedOperand(MI, AMDGPU::OpName::simm16)->getImm();
247 if (((Dst & AMDGPU::Hwreg::ID_MASK_) >> AMDGPU::Hwreg::ID_SHIFT_) !=
248 AMDGPU::Hwreg::ID_MODE)
251 unsigned Width = ((Dst & AMDGPU::Hwreg::WIDTH_M1_MASK_) >>
252 AMDGPU::Hwreg::WIDTH_M1_SHIFT_) +
255 (Dst & AMDGPU::Hwreg::OFFSET_MASK_) >> AMDGPU::Hwreg::OFFSET_SHIFT_;
256 unsigned Mask = ((1 << Width) - 1) << Offset;
258 // If an InsertionPoint is set we will insert a setreg there.
259 if (InsertionPoint) {
260 insertSetreg(MBB, InsertionPoint, TII, IPChange.delta(NewInfo->Change));
261 InsertionPoint = nullptr;
263 // If this is an immediate then we know the value being set, but if it is
264 // not an immediate then we treat the modified bits of the mode register
266 if (MI.getOpcode() == AMDGPU::S_SETREG_IMM32_B32) {
267 unsigned Val = TII->getNamedOperand(MI, AMDGPU::OpName::imm)->getImm();
268 unsigned Mode = (Val << Offset) & Mask;
269 Status Setreg = Status(Mask, Mode);
270 // If we haven't already set the initial requirements for the block we
271 // don't need to as the requirements start from this explicit setreg.
272 RequirePending = false;
273 NewInfo->Change = NewInfo->Change.merge(Setreg);
275 NewInfo->Change = NewInfo->Change.mergeUnknown(Mask);
277 } else if (!NewInfo->Change.isCompatible(InstrMode)) {
278 // This instruction uses the Mode register and its requirements aren't
279 // compatible with the current mode.
280 if (InsertionPoint) {
281 // If the required mode change cannot be included in the current
282 // InsertionPoint changes, we need a setreg and start a new
284 if (!IPChange.delta(NewInfo->Change).isCombinable(InstrMode)) {
285 if (RequirePending) {
286 // This is the first insertionPoint in the block so we will defer
287 // the insertion of the setreg to Phase 3 where we know whether or
288 // not it is actually needed.
289 NewInfo->FirstInsertionPoint = InsertionPoint;
290 NewInfo->Require = NewInfo->Change;
291 RequirePending = false;
293 insertSetreg(MBB, InsertionPoint, TII,
294 IPChange.delta(NewInfo->Change));
295 IPChange = NewInfo->Change;
297 // Set the new InsertionPoint
298 InsertionPoint = &MI;
300 NewInfo->Change = NewInfo->Change.merge(InstrMode);
302 // No InsertionPoint is currently set - this is either the first in
303 // the block or we have previously seen an explicit setreg.
304 InsertionPoint = &MI;
305 IPChange = NewInfo->Change;
306 NewInfo->Change = NewInfo->Change.merge(InstrMode);
310 if (RequirePending) {
311 // If we haven't yet set the initial requirements for the block we set them
313 NewInfo->FirstInsertionPoint = InsertionPoint;
314 NewInfo->Require = NewInfo->Change;
315 } else if (InsertionPoint) {
316 // We need to insert a setreg at the InsertionPoint
317 insertSetreg(MBB, InsertionPoint, TII, IPChange.delta(NewInfo->Change));
319 NewInfo->Exit = NewInfo->Change;
320 BlockInfo[MBB.getNumber()] = std::move(NewInfo);
323 // In Phase 2 we revisit each block and calculate the common Mode register
324 // value provided by all predecessor blocks. If the Exit value for the block
325 // is changed, then we add the successor blocks to the worklist so that the
326 // exit value is propagated.
327 void SIModeRegister::processBlockPhase2(MachineBasicBlock &MBB,
328 const SIInstrInfo *TII) {
329 // BlockData *BI = BlockInfo[MBB.getNumber()];
330 unsigned ThisBlock = MBB.getNumber();
331 if (MBB.pred_empty()) {
332 // There are no predecessors, so use the default starting status.
333 BlockInfo[ThisBlock]->Pred = DefaultStatus;
335 // Build a status that is common to all the predecessors by intersecting
336 // all the predecessor exit status values.
337 MachineBasicBlock::pred_iterator P = MBB.pred_begin(), E = MBB.pred_end();
338 MachineBasicBlock &PB = *(*P);
339 BlockInfo[ThisBlock]->Pred = BlockInfo[PB.getNumber()]->Exit;
341 for (P = std::next(P); P != E; P = std::next(P)) {
342 MachineBasicBlock *Pred = *P;
343 BlockInfo[ThisBlock]->Pred = BlockInfo[ThisBlock]->Pred.intersect(BlockInfo[Pred->getNumber()]->Exit);
346 Status TmpStatus = BlockInfo[ThisBlock]->Pred.merge(BlockInfo[ThisBlock]->Change);
347 if (BlockInfo[ThisBlock]->Exit != TmpStatus) {
348 BlockInfo[ThisBlock]->Exit = TmpStatus;
349 // Add the successors to the work list so we can propagate the changed exit
351 for (MachineBasicBlock::succ_iterator S = MBB.succ_begin(),
353 S != E; S = std::next(S)) {
354 MachineBasicBlock &B = *(*S);
360 // In Phase 3 we revisit each block and if it has an insertion point defined we
361 // check whether the predecessor mode meets the block's entry requirements. If
362 // not we insert an appropriate setreg instruction to modify the Mode register.
363 void SIModeRegister::processBlockPhase3(MachineBasicBlock &MBB,
364 const SIInstrInfo *TII) {
365 // BlockData *BI = BlockInfo[MBB.getNumber()];
366 unsigned ThisBlock = MBB.getNumber();
367 if (!BlockInfo[ThisBlock]->Pred.isCompatible(BlockInfo[ThisBlock]->Require)) {
368 Status Delta = BlockInfo[ThisBlock]->Pred.delta(BlockInfo[ThisBlock]->Require);
369 if (BlockInfo[ThisBlock]->FirstInsertionPoint)
370 insertSetreg(MBB, BlockInfo[ThisBlock]->FirstInsertionPoint, TII, Delta);
372 insertSetreg(MBB, &MBB.instr_front(), TII, Delta);
376 bool SIModeRegister::runOnMachineFunction(MachineFunction &MF) {
377 BlockInfo.resize(MF.getNumBlockIDs());
378 const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
379 const SIInstrInfo *TII = ST.getInstrInfo();
381 // Processing is performed in a number of phases
383 // Phase 1 - determine the initial mode required by each block, and add setreg
384 // instructions for intra block requirements.
385 for (MachineBasicBlock &BB : MF)
386 processBlockPhase1(BB, TII);
388 // Phase 2 - determine the exit mode from each block. We add all blocks to the
389 // list here, but will also add any that need to be revisited during Phase 2
391 for (MachineBasicBlock &BB : MF)
392 Phase2List.push(&BB);
393 while (!Phase2List.empty()) {
394 processBlockPhase2(*Phase2List.front(), TII);
398 // Phase 3 - add an initial setreg to each block where the required entry mode
399 // is not satisfied by the exit mode of all its predecessors.
400 for (MachineBasicBlock &BB : MF)
401 processBlockPhase3(BB, TII);
405 return NumSetregInserted > 0;