1 | //===-- PPCInstrInfo.cpp - PowerPC Instruction Information ----------------===// |
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 | // This file contains the PowerPC implementation of the TargetInstrInfo class. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "PPCInstrInfo.h" |
14 | #include "MCTargetDesc/PPCPredicates.h" |
15 | #include "PPC.h" |
16 | #include "PPCHazardRecognizers.h" |
17 | #include "PPCInstrBuilder.h" |
18 | #include "PPCMachineFunctionInfo.h" |
19 | #include "PPCTargetMachine.h" |
20 | #include "llvm/ADT/STLExtras.h" |
21 | #include "llvm/ADT/Statistic.h" |
22 | #include "llvm/Analysis/AliasAnalysis.h" |
23 | #include "llvm/CodeGen/LiveIntervals.h" |
24 | #include "llvm/CodeGen/LivePhysRegs.h" |
25 | #include "llvm/CodeGen/MachineCombinerPattern.h" |
26 | #include "llvm/CodeGen/MachineConstantPool.h" |
27 | #include "llvm/CodeGen/MachineFrameInfo.h" |
28 | #include "llvm/CodeGen/MachineFunctionPass.h" |
29 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
30 | #include "llvm/CodeGen/MachineMemOperand.h" |
31 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
32 | #include "llvm/CodeGen/PseudoSourceValue.h" |
33 | #include "llvm/CodeGen/RegisterClassInfo.h" |
34 | #include "llvm/CodeGen/RegisterPressure.h" |
35 | #include "llvm/CodeGen/ScheduleDAG.h" |
36 | #include "llvm/CodeGen/SlotIndexes.h" |
37 | #include "llvm/CodeGen/StackMaps.h" |
38 | #include "llvm/MC/MCAsmInfo.h" |
39 | #include "llvm/MC/MCInst.h" |
40 | #include "llvm/MC/TargetRegistry.h" |
41 | #include "llvm/Support/CommandLine.h" |
42 | #include "llvm/Support/Debug.h" |
43 | #include "llvm/Support/ErrorHandling.h" |
44 | #include "llvm/Support/raw_ostream.h" |
45 | |
46 | using namespace llvm; |
47 | |
48 | #define DEBUG_TYPE "ppc-instr-info" |
49 | |
50 | #define GET_INSTRMAP_INFO |
51 | #define GET_INSTRINFO_CTOR_DTOR |
52 | #include "PPCGenInstrInfo.inc" |
53 | |
54 | STATISTIC(NumStoreSPILLVSRRCAsVec, |
55 | "Number of spillvsrrc spilled to stack as vec" ); |
56 | STATISTIC(NumStoreSPILLVSRRCAsGpr, |
57 | "Number of spillvsrrc spilled to stack as gpr" ); |
58 | STATISTIC(NumGPRtoVSRSpill, "Number of gpr spills to spillvsrrc" ); |
59 | STATISTIC(CmpIselsConverted, |
60 | "Number of ISELs that depend on comparison of constants converted" ); |
61 | STATISTIC(MissedConvertibleImmediateInstrs, |
62 | "Number of compare-immediate instructions fed by constants" ); |
63 | STATISTIC(NumRcRotatesConvertedToRcAnd, |
64 | "Number of record-form rotates converted to record-form andi" ); |
65 | |
66 | static cl:: |
67 | opt<bool> DisableCTRLoopAnal("disable-ppc-ctrloop-analysis" , cl::Hidden, |
68 | cl::desc("Disable analysis for CTR loops" )); |
69 | |
70 | static cl::opt<bool> DisableCmpOpt("disable-ppc-cmp-opt" , |
71 | cl::desc("Disable compare instruction optimization" ), cl::Hidden); |
72 | |
73 | static cl::opt<bool> VSXSelfCopyCrash("crash-on-ppc-vsx-self-copy" , |
74 | cl::desc("Causes the backend to crash instead of generating a nop VSX copy" ), |
75 | cl::Hidden); |
76 | |
77 | static cl::opt<bool> |
78 | UseOldLatencyCalc("ppc-old-latency-calc" , cl::Hidden, |
79 | cl::desc("Use the old (incorrect) instruction latency calculation" )); |
80 | |
81 | static cl::opt<float> |
82 | FMARPFactor("ppc-fma-rp-factor" , cl::Hidden, cl::init(Val: 1.5), |
83 | cl::desc("register pressure factor for the transformations." )); |
84 | |
85 | static cl::opt<bool> EnableFMARegPressureReduction( |
86 | "ppc-fma-rp-reduction" , cl::Hidden, cl::init(Val: true), |
87 | cl::desc("enable register pressure reduce in machine combiner pass." )); |
88 | |
89 | // Pin the vtable to this file. |
90 | void PPCInstrInfo::anchor() {} |
91 | |
92 | PPCInstrInfo::PPCInstrInfo(PPCSubtarget &STI) |
93 | : PPCGenInstrInfo(PPC::ADJCALLSTACKDOWN, PPC::ADJCALLSTACKUP, |
94 | /* CatchRetOpcode */ -1, |
95 | STI.isPPC64() ? PPC::BLR8 : PPC::BLR), |
96 | Subtarget(STI), RI(STI.getTargetMachine()) {} |
97 | |
98 | /// CreateTargetHazardRecognizer - Return the hazard recognizer to use for |
99 | /// this target when scheduling the DAG. |
100 | ScheduleHazardRecognizer * |
101 | PPCInstrInfo::CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI, |
102 | const ScheduleDAG *DAG) const { |
103 | unsigned Directive = |
104 | static_cast<const PPCSubtarget *>(STI)->getCPUDirective(); |
105 | if (Directive == PPC::DIR_440 || Directive == PPC::DIR_A2 || |
106 | Directive == PPC::DIR_E500mc || Directive == PPC::DIR_E5500) { |
107 | const InstrItineraryData *II = |
108 | static_cast<const PPCSubtarget *>(STI)->getInstrItineraryData(); |
109 | return new ScoreboardHazardRecognizer(II, DAG); |
110 | } |
111 | |
112 | return TargetInstrInfo::CreateTargetHazardRecognizer(STI, DAG); |
113 | } |
114 | |
115 | /// CreateTargetPostRAHazardRecognizer - Return the postRA hazard recognizer |
116 | /// to use for this target when scheduling the DAG. |
117 | ScheduleHazardRecognizer * |
118 | PPCInstrInfo::CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II, |
119 | const ScheduleDAG *DAG) const { |
120 | unsigned Directive = |
121 | DAG->MF.getSubtarget<PPCSubtarget>().getCPUDirective(); |
122 | |
123 | // FIXME: Leaving this as-is until we have POWER9 scheduling info |
124 | if (Directive == PPC::DIR_PWR7 || Directive == PPC::DIR_PWR8) |
125 | return new PPCDispatchGroupSBHazardRecognizer(II, DAG); |
126 | |
127 | // Most subtargets use a PPC970 recognizer. |
128 | if (Directive != PPC::DIR_440 && Directive != PPC::DIR_A2 && |
129 | Directive != PPC::DIR_E500mc && Directive != PPC::DIR_E5500) { |
130 | assert(DAG->TII && "No InstrInfo?" ); |
131 | |
132 | return new PPCHazardRecognizer970(*DAG); |
133 | } |
134 | |
135 | return new ScoreboardHazardRecognizer(II, DAG); |
136 | } |
137 | |
138 | unsigned PPCInstrInfo::getInstrLatency(const InstrItineraryData *ItinData, |
139 | const MachineInstr &MI, |
140 | unsigned *PredCost) const { |
141 | if (!ItinData || UseOldLatencyCalc) |
142 | return PPCGenInstrInfo::getInstrLatency(ItinData, MI, PredCost); |
143 | |
144 | // The default implementation of getInstrLatency calls getStageLatency, but |
145 | // getStageLatency does not do the right thing for us. While we have |
146 | // itinerary, most cores are fully pipelined, and so the itineraries only |
147 | // express the first part of the pipeline, not every stage. Instead, we need |
148 | // to use the listed output operand cycle number (using operand 0 here, which |
149 | // is an output). |
150 | |
151 | unsigned Latency = 1; |
152 | unsigned DefClass = MI.getDesc().getSchedClass(); |
153 | for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) { |
154 | const MachineOperand &MO = MI.getOperand(i); |
155 | if (!MO.isReg() || !MO.isDef() || MO.isImplicit()) |
156 | continue; |
157 | |
158 | std::optional<unsigned> Cycle = ItinData->getOperandCycle(ItinClassIndx: DefClass, OperandIdx: i); |
159 | if (!Cycle) |
160 | continue; |
161 | |
162 | Latency = std::max(a: Latency, b: *Cycle); |
163 | } |
164 | |
165 | return Latency; |
166 | } |
167 | |
168 | std::optional<unsigned> PPCInstrInfo::getOperandLatency( |
169 | const InstrItineraryData *ItinData, const MachineInstr &DefMI, |
170 | unsigned DefIdx, const MachineInstr &UseMI, unsigned UseIdx) const { |
171 | std::optional<unsigned> Latency = PPCGenInstrInfo::getOperandLatency( |
172 | ItinData, DefMI, DefIdx, UseMI, UseIdx); |
173 | |
174 | if (!DefMI.getParent()) |
175 | return Latency; |
176 | |
177 | const MachineOperand &DefMO = DefMI.getOperand(i: DefIdx); |
178 | Register Reg = DefMO.getReg(); |
179 | |
180 | bool IsRegCR; |
181 | if (Reg.isVirtual()) { |
182 | const MachineRegisterInfo *MRI = |
183 | &DefMI.getParent()->getParent()->getRegInfo(); |
184 | IsRegCR = MRI->getRegClass(Reg)->hasSuperClassEq(RC: &PPC::CRRCRegClass) || |
185 | MRI->getRegClass(Reg)->hasSuperClassEq(RC: &PPC::CRBITRCRegClass); |
186 | } else { |
187 | IsRegCR = PPC::CRRCRegClass.contains(Reg) || |
188 | PPC::CRBITRCRegClass.contains(Reg); |
189 | } |
190 | |
191 | if (UseMI.isBranch() && IsRegCR) { |
192 | if (!Latency) |
193 | Latency = getInstrLatency(ItinData, MI: DefMI); |
194 | |
195 | // On some cores, there is an additional delay between writing to a condition |
196 | // register, and using it from a branch. |
197 | unsigned Directive = Subtarget.getCPUDirective(); |
198 | switch (Directive) { |
199 | default: break; |
200 | case PPC::DIR_7400: |
201 | case PPC::DIR_750: |
202 | case PPC::DIR_970: |
203 | case PPC::DIR_E5500: |
204 | case PPC::DIR_PWR4: |
205 | case PPC::DIR_PWR5: |
206 | case PPC::DIR_PWR5X: |
207 | case PPC::DIR_PWR6: |
208 | case PPC::DIR_PWR6X: |
209 | case PPC::DIR_PWR7: |
210 | case PPC::DIR_PWR8: |
211 | // FIXME: Is this needed for POWER9? |
212 | Latency = *Latency + 2; |
213 | break; |
214 | } |
215 | } |
216 | |
217 | return Latency; |
218 | } |
219 | |
220 | void PPCInstrInfo::setSpecialOperandAttr(MachineInstr &MI, |
221 | uint32_t Flags) const { |
222 | MI.setFlags(Flags); |
223 | MI.clearFlag(Flag: MachineInstr::MIFlag::NoSWrap); |
224 | MI.clearFlag(Flag: MachineInstr::MIFlag::NoUWrap); |
225 | MI.clearFlag(Flag: MachineInstr::MIFlag::IsExact); |
226 | } |
227 | |
228 | // This function does not list all associative and commutative operations, but |
229 | // only those worth feeding through the machine combiner in an attempt to |
230 | // reduce the critical path. Mostly, this means floating-point operations, |
231 | // because they have high latencies(>=5) (compared to other operations, such as |
232 | // and/or, which are also associative and commutative, but have low latencies). |
233 | bool PPCInstrInfo::isAssociativeAndCommutative(const MachineInstr &Inst, |
234 | bool Invert) const { |
235 | if (Invert) |
236 | return false; |
237 | switch (Inst.getOpcode()) { |
238 | // Floating point: |
239 | // FP Add: |
240 | case PPC::FADD: |
241 | case PPC::FADDS: |
242 | // FP Multiply: |
243 | case PPC::FMUL: |
244 | case PPC::FMULS: |
245 | // Altivec Add: |
246 | case PPC::VADDFP: |
247 | // VSX Add: |
248 | case PPC::XSADDDP: |
249 | case PPC::XVADDDP: |
250 | case PPC::XVADDSP: |
251 | case PPC::XSADDSP: |
252 | // VSX Multiply: |
253 | case PPC::XSMULDP: |
254 | case PPC::XVMULDP: |
255 | case PPC::XVMULSP: |
256 | case PPC::XSMULSP: |
257 | return Inst.getFlag(Flag: MachineInstr::MIFlag::FmReassoc) && |
258 | Inst.getFlag(Flag: MachineInstr::MIFlag::FmNsz); |
259 | // Fixed point: |
260 | // Multiply: |
261 | case PPC::MULHD: |
262 | case PPC::MULLD: |
263 | case PPC::MULHW: |
264 | case PPC::MULLW: |
265 | return true; |
266 | default: |
267 | return false; |
268 | } |
269 | } |
270 | |
271 | #define InfoArrayIdxFMAInst 0 |
272 | #define InfoArrayIdxFAddInst 1 |
273 | #define InfoArrayIdxFMULInst 2 |
274 | #define InfoArrayIdxAddOpIdx 3 |
275 | #define InfoArrayIdxMULOpIdx 4 |
276 | #define InfoArrayIdxFSubInst 5 |
277 | // Array keeps info for FMA instructions: |
278 | // Index 0(InfoArrayIdxFMAInst): FMA instruction; |
279 | // Index 1(InfoArrayIdxFAddInst): ADD instruction associated with FMA; |
280 | // Index 2(InfoArrayIdxFMULInst): MUL instruction associated with FMA; |
281 | // Index 3(InfoArrayIdxAddOpIdx): ADD operand index in FMA operands; |
282 | // Index 4(InfoArrayIdxMULOpIdx): first MUL operand index in FMA operands; |
283 | // second MUL operand index is plus 1; |
284 | // Index 5(InfoArrayIdxFSubInst): SUB instruction associated with FMA. |
285 | static const uint16_t FMAOpIdxInfo[][6] = { |
286 | // FIXME: Add more FMA instructions like XSNMADDADP and so on. |
287 | {PPC::XSMADDADP, PPC::XSADDDP, PPC::XSMULDP, 1, 2, PPC::XSSUBDP}, |
288 | {PPC::XSMADDASP, PPC::XSADDSP, PPC::XSMULSP, 1, 2, PPC::XSSUBSP}, |
289 | {PPC::XVMADDADP, PPC::XVADDDP, PPC::XVMULDP, 1, 2, PPC::XVSUBDP}, |
290 | {PPC::XVMADDASP, PPC::XVADDSP, PPC::XVMULSP, 1, 2, PPC::XVSUBSP}, |
291 | {PPC::FMADD, PPC::FADD, PPC::FMUL, 3, 1, PPC::FSUB}, |
292 | {PPC::FMADDS, PPC::FADDS, PPC::FMULS, 3, 1, PPC::FSUBS}}; |
293 | |
294 | // Check if an opcode is a FMA instruction. If it is, return the index in array |
295 | // FMAOpIdxInfo. Otherwise, return -1. |
296 | int16_t PPCInstrInfo::getFMAOpIdxInfo(unsigned Opcode) const { |
297 | for (unsigned I = 0; I < std::size(FMAOpIdxInfo); I++) |
298 | if (FMAOpIdxInfo[I][InfoArrayIdxFMAInst] == Opcode) |
299 | return I; |
300 | return -1; |
301 | } |
302 | |
303 | // On PowerPC target, we have two kinds of patterns related to FMA: |
304 | // 1: Improve ILP. |
305 | // Try to reassociate FMA chains like below: |
306 | // |
307 | // Pattern 1: |
308 | // A = FADD X, Y (Leaf) |
309 | // B = FMA A, M21, M22 (Prev) |
310 | // C = FMA B, M31, M32 (Root) |
311 | // --> |
312 | // A = FMA X, M21, M22 |
313 | // B = FMA Y, M31, M32 |
314 | // C = FADD A, B |
315 | // |
316 | // Pattern 2: |
317 | // A = FMA X, M11, M12 (Leaf) |
318 | // B = FMA A, M21, M22 (Prev) |
319 | // C = FMA B, M31, M32 (Root) |
320 | // --> |
321 | // A = FMUL M11, M12 |
322 | // B = FMA X, M21, M22 |
323 | // D = FMA A, M31, M32 |
324 | // C = FADD B, D |
325 | // |
326 | // breaking the dependency between A and B, allowing FMA to be executed in |
327 | // parallel (or back-to-back in a pipeline) instead of depending on each other. |
328 | // |
329 | // 2: Reduce register pressure. |
330 | // Try to reassociate FMA with FSUB and a constant like below: |
331 | // C is a floating point const. |
332 | // |
333 | // Pattern 1: |
334 | // A = FSUB X, Y (Leaf) |
335 | // D = FMA B, C, A (Root) |
336 | // --> |
337 | // A = FMA B, Y, -C |
338 | // D = FMA A, X, C |
339 | // |
340 | // Pattern 2: |
341 | // A = FSUB X, Y (Leaf) |
342 | // D = FMA B, A, C (Root) |
343 | // --> |
344 | // A = FMA B, Y, -C |
345 | // D = FMA A, X, C |
346 | // |
347 | // Before the transformation, A must be assigned with different hardware |
348 | // register with D. After the transformation, A and D must be assigned with |
349 | // same hardware register due to TIE attribute of FMA instructions. |
350 | // |
351 | bool PPCInstrInfo::getFMAPatterns(MachineInstr &Root, |
352 | SmallVectorImpl<unsigned> &Patterns, |
353 | bool DoRegPressureReduce) const { |
354 | MachineBasicBlock *MBB = Root.getParent(); |
355 | const MachineRegisterInfo *MRI = &MBB->getParent()->getRegInfo(); |
356 | const TargetRegisterInfo *TRI = &getRegisterInfo(); |
357 | |
358 | auto IsAllOpsVirtualReg = [](const MachineInstr &Instr) { |
359 | for (const auto &MO : Instr.explicit_operands()) |
360 | if (!(MO.isReg() && MO.getReg().isVirtual())) |
361 | return false; |
362 | return true; |
363 | }; |
364 | |
365 | auto IsReassociableAddOrSub = [&](const MachineInstr &Instr, |
366 | unsigned OpType) { |
367 | if (Instr.getOpcode() != |
368 | FMAOpIdxInfo[getFMAOpIdxInfo(Opcode: Root.getOpcode())][OpType]) |
369 | return false; |
370 | |
371 | // Instruction can be reassociated. |
372 | // fast math flags may prohibit reassociation. |
373 | if (!(Instr.getFlag(Flag: MachineInstr::MIFlag::FmReassoc) && |
374 | Instr.getFlag(Flag: MachineInstr::MIFlag::FmNsz))) |
375 | return false; |
376 | |
377 | // Instruction operands are virtual registers for reassociation. |
378 | if (!IsAllOpsVirtualReg(Instr)) |
379 | return false; |
380 | |
381 | // For register pressure reassociation, the FSub must have only one use as |
382 | // we want to delete the sub to save its def. |
383 | if (OpType == InfoArrayIdxFSubInst && |
384 | !MRI->hasOneNonDBGUse(RegNo: Instr.getOperand(i: 0).getReg())) |
385 | return false; |
386 | |
387 | return true; |
388 | }; |
389 | |
390 | auto IsReassociableFMA = [&](const MachineInstr &Instr, int16_t &AddOpIdx, |
391 | int16_t &MulOpIdx, bool IsLeaf) { |
392 | int16_t Idx = getFMAOpIdxInfo(Opcode: Instr.getOpcode()); |
393 | if (Idx < 0) |
394 | return false; |
395 | |
396 | // Instruction can be reassociated. |
397 | // fast math flags may prohibit reassociation. |
398 | if (!(Instr.getFlag(Flag: MachineInstr::MIFlag::FmReassoc) && |
399 | Instr.getFlag(Flag: MachineInstr::MIFlag::FmNsz))) |
400 | return false; |
401 | |
402 | // Instruction operands are virtual registers for reassociation. |
403 | if (!IsAllOpsVirtualReg(Instr)) |
404 | return false; |
405 | |
406 | MulOpIdx = FMAOpIdxInfo[Idx][InfoArrayIdxMULOpIdx]; |
407 | if (IsLeaf) |
408 | return true; |
409 | |
410 | AddOpIdx = FMAOpIdxInfo[Idx][InfoArrayIdxAddOpIdx]; |
411 | |
412 | const MachineOperand &OpAdd = Instr.getOperand(i: AddOpIdx); |
413 | MachineInstr *MIAdd = MRI->getUniqueVRegDef(Reg: OpAdd.getReg()); |
414 | // If 'add' operand's def is not in current block, don't do ILP related opt. |
415 | if (!MIAdd || MIAdd->getParent() != MBB) |
416 | return false; |
417 | |
418 | // If this is not Leaf FMA Instr, its 'add' operand should only have one use |
419 | // as this fma will be changed later. |
420 | return IsLeaf ? true : MRI->hasOneNonDBGUse(RegNo: OpAdd.getReg()); |
421 | }; |
422 | |
423 | int16_t AddOpIdx = -1; |
424 | int16_t MulOpIdx = -1; |
425 | |
426 | bool IsUsedOnceL = false; |
427 | bool IsUsedOnceR = false; |
428 | MachineInstr *MULInstrL = nullptr; |
429 | MachineInstr *MULInstrR = nullptr; |
430 | |
431 | auto IsRPReductionCandidate = [&]() { |
432 | // Currently, we only support float and double. |
433 | // FIXME: add support for other types. |
434 | unsigned Opcode = Root.getOpcode(); |
435 | if (Opcode != PPC::XSMADDASP && Opcode != PPC::XSMADDADP) |
436 | return false; |
437 | |
438 | // Root must be a valid FMA like instruction. |
439 | // Treat it as leaf as we don't care its add operand. |
440 | if (IsReassociableFMA(Root, AddOpIdx, MulOpIdx, true)) { |
441 | assert((MulOpIdx >= 0) && "mul operand index not right!" ); |
442 | Register MULRegL = TRI->lookThruSingleUseCopyChain( |
443 | SrcReg: Root.getOperand(i: MulOpIdx).getReg(), MRI); |
444 | Register MULRegR = TRI->lookThruSingleUseCopyChain( |
445 | SrcReg: Root.getOperand(i: MulOpIdx + 1).getReg(), MRI); |
446 | if (!MULRegL && !MULRegR) |
447 | return false; |
448 | |
449 | if (MULRegL && !MULRegR) { |
450 | MULRegR = |
451 | TRI->lookThruCopyLike(SrcReg: Root.getOperand(i: MulOpIdx + 1).getReg(), MRI); |
452 | IsUsedOnceL = true; |
453 | } else if (!MULRegL && MULRegR) { |
454 | MULRegL = |
455 | TRI->lookThruCopyLike(SrcReg: Root.getOperand(i: MulOpIdx).getReg(), MRI); |
456 | IsUsedOnceR = true; |
457 | } else { |
458 | IsUsedOnceL = true; |
459 | IsUsedOnceR = true; |
460 | } |
461 | |
462 | if (!MULRegL.isVirtual() || !MULRegR.isVirtual()) |
463 | return false; |
464 | |
465 | MULInstrL = MRI->getVRegDef(Reg: MULRegL); |
466 | MULInstrR = MRI->getVRegDef(Reg: MULRegR); |
467 | return true; |
468 | } |
469 | return false; |
470 | }; |
471 | |
472 | // Register pressure fma reassociation patterns. |
473 | if (DoRegPressureReduce && IsRPReductionCandidate()) { |
474 | assert((MULInstrL && MULInstrR) && "wrong register preduction candidate!" ); |
475 | // Register pressure pattern 1 |
476 | if (isLoadFromConstantPool(I: MULInstrL) && IsUsedOnceR && |
477 | IsReassociableAddOrSub(*MULInstrR, InfoArrayIdxFSubInst)) { |
478 | LLVM_DEBUG(dbgs() << "add pattern REASSOC_XY_BCA\n" ); |
479 | Patterns.push_back(Elt: PPCMachineCombinerPattern::REASSOC_XY_BCA); |
480 | return true; |
481 | } |
482 | |
483 | // Register pressure pattern 2 |
484 | if ((isLoadFromConstantPool(I: MULInstrR) && IsUsedOnceL && |
485 | IsReassociableAddOrSub(*MULInstrL, InfoArrayIdxFSubInst))) { |
486 | LLVM_DEBUG(dbgs() << "add pattern REASSOC_XY_BAC\n" ); |
487 | Patterns.push_back(Elt: PPCMachineCombinerPattern::REASSOC_XY_BAC); |
488 | return true; |
489 | } |
490 | } |
491 | |
492 | // ILP fma reassociation patterns. |
493 | // Root must be a valid FMA like instruction. |
494 | AddOpIdx = -1; |
495 | if (!IsReassociableFMA(Root, AddOpIdx, MulOpIdx, false)) |
496 | return false; |
497 | |
498 | assert((AddOpIdx >= 0) && "add operand index not right!" ); |
499 | |
500 | Register RegB = Root.getOperand(i: AddOpIdx).getReg(); |
501 | MachineInstr *Prev = MRI->getUniqueVRegDef(Reg: RegB); |
502 | |
503 | // Prev must be a valid FMA like instruction. |
504 | AddOpIdx = -1; |
505 | if (!IsReassociableFMA(*Prev, AddOpIdx, MulOpIdx, false)) |
506 | return false; |
507 | |
508 | assert((AddOpIdx >= 0) && "add operand index not right!" ); |
509 | |
510 | Register RegA = Prev->getOperand(i: AddOpIdx).getReg(); |
511 | MachineInstr *Leaf = MRI->getUniqueVRegDef(Reg: RegA); |
512 | AddOpIdx = -1; |
513 | if (IsReassociableFMA(*Leaf, AddOpIdx, MulOpIdx, true)) { |
514 | Patterns.push_back(Elt: PPCMachineCombinerPattern::REASSOC_XMM_AMM_BMM); |
515 | LLVM_DEBUG(dbgs() << "add pattern REASSOC_XMM_AMM_BMM\n" ); |
516 | return true; |
517 | } |
518 | if (IsReassociableAddOrSub(*Leaf, InfoArrayIdxFAddInst)) { |
519 | Patterns.push_back(Elt: PPCMachineCombinerPattern::REASSOC_XY_AMM_BMM); |
520 | LLVM_DEBUG(dbgs() << "add pattern REASSOC_XY_AMM_BMM\n" ); |
521 | return true; |
522 | } |
523 | return false; |
524 | } |
525 | |
526 | void PPCInstrInfo::finalizeInsInstrs( |
527 | MachineInstr &Root, unsigned &Pattern, |
528 | SmallVectorImpl<MachineInstr *> &InsInstrs) const { |
529 | assert(!InsInstrs.empty() && "Instructions set to be inserted is empty!" ); |
530 | |
531 | MachineFunction *MF = Root.getMF(); |
532 | MachineRegisterInfo *MRI = &MF->getRegInfo(); |
533 | const TargetRegisterInfo *TRI = &getRegisterInfo(); |
534 | MachineConstantPool *MCP = MF->getConstantPool(); |
535 | |
536 | int16_t Idx = getFMAOpIdxInfo(Opcode: Root.getOpcode()); |
537 | if (Idx < 0) |
538 | return; |
539 | |
540 | uint16_t FirstMulOpIdx = FMAOpIdxInfo[Idx][InfoArrayIdxMULOpIdx]; |
541 | |
542 | // For now we only need to fix up placeholder for register pressure reduce |
543 | // patterns. |
544 | Register ConstReg = 0; |
545 | switch (Pattern) { |
546 | case PPCMachineCombinerPattern::REASSOC_XY_BCA: |
547 | ConstReg = |
548 | TRI->lookThruCopyLike(SrcReg: Root.getOperand(i: FirstMulOpIdx).getReg(), MRI); |
549 | break; |
550 | case PPCMachineCombinerPattern::REASSOC_XY_BAC: |
551 | ConstReg = |
552 | TRI->lookThruCopyLike(SrcReg: Root.getOperand(i: FirstMulOpIdx + 1).getReg(), MRI); |
553 | break; |
554 | default: |
555 | // Not register pressure reduce patterns. |
556 | return; |
557 | } |
558 | |
559 | MachineInstr *ConstDefInstr = MRI->getVRegDef(Reg: ConstReg); |
560 | // Get const value from const pool. |
561 | const Constant *C = getConstantFromConstantPool(I: ConstDefInstr); |
562 | assert(isa<llvm::ConstantFP>(C) && "not a valid constant!" ); |
563 | |
564 | // Get negative fp const. |
565 | APFloat F1((dyn_cast<ConstantFP>(Val: C))->getValueAPF()); |
566 | F1.changeSign(); |
567 | Constant *NegC = ConstantFP::get(Context&: dyn_cast<ConstantFP>(Val: C)->getContext(), V: F1); |
568 | Align Alignment = MF->getDataLayout().getPrefTypeAlign(Ty: C->getType()); |
569 | |
570 | // Put negative fp const into constant pool. |
571 | unsigned ConstPoolIdx = MCP->getConstantPoolIndex(C: NegC, Alignment); |
572 | |
573 | MachineOperand *Placeholder = nullptr; |
574 | // Record the placeholder PPC::ZERO8 we add in reassociateFMA. |
575 | for (auto *Inst : InsInstrs) { |
576 | for (MachineOperand &Operand : Inst->explicit_operands()) { |
577 | assert(Operand.isReg() && "Invalid instruction in InsInstrs!" ); |
578 | if (Operand.getReg() == PPC::ZERO8) { |
579 | Placeholder = &Operand; |
580 | break; |
581 | } |
582 | } |
583 | } |
584 | |
585 | assert(Placeholder && "Placeholder does not exist!" ); |
586 | |
587 | // Generate instructions to load the const fp from constant pool. |
588 | // We only support PPC64 and medium code model. |
589 | Register LoadNewConst = |
590 | generateLoadForNewConst(Idx: ConstPoolIdx, MI: &Root, Ty: C->getType(), InsInstrs); |
591 | |
592 | // Fill the placeholder with the new load from constant pool. |
593 | Placeholder->setReg(LoadNewConst); |
594 | } |
595 | |
596 | bool PPCInstrInfo::shouldReduceRegisterPressure( |
597 | const MachineBasicBlock *MBB, const RegisterClassInfo *RegClassInfo) const { |
598 | |
599 | if (!EnableFMARegPressureReduction) |
600 | return false; |
601 | |
602 | // Currently, we only enable register pressure reducing in machine combiner |
603 | // for: 1: PPC64; 2: Code Model is Medium; 3: Power9 which also has vector |
604 | // support. |
605 | // |
606 | // So we need following instructions to access a TOC entry: |
607 | // |
608 | // %6:g8rc_and_g8rc_nox0 = ADDIStocHA8 $x2, %const.0 |
609 | // %7:vssrc = DFLOADf32 target-flags(ppc-toc-lo) %const.0, |
610 | // killed %6:g8rc_and_g8rc_nox0, implicit $x2 :: (load 4 from constant-pool) |
611 | // |
612 | // FIXME: add more supported targets, like Small and Large code model, PPC32, |
613 | // AIX. |
614 | if (!(Subtarget.isPPC64() && Subtarget.hasP9Vector() && |
615 | Subtarget.getTargetMachine().getCodeModel() == CodeModel::Medium)) |
616 | return false; |
617 | |
618 | const TargetRegisterInfo *TRI = &getRegisterInfo(); |
619 | const MachineFunction *MF = MBB->getParent(); |
620 | const MachineRegisterInfo *MRI = &MF->getRegInfo(); |
621 | |
622 | auto GetMBBPressure = |
623 | [&](const MachineBasicBlock *MBB) -> std::vector<unsigned> { |
624 | RegionPressure Pressure; |
625 | RegPressureTracker RPTracker(Pressure); |
626 | |
627 | // Initialize the register pressure tracker. |
628 | RPTracker.init(mf: MBB->getParent(), rci: RegClassInfo, lis: nullptr, mbb: MBB, pos: MBB->end(), |
629 | /*TrackLaneMasks*/ false, /*TrackUntiedDefs=*/true); |
630 | |
631 | for (const auto &MI : reverse(C: *MBB)) { |
632 | if (MI.isDebugValue() || MI.isDebugLabel()) |
633 | continue; |
634 | RegisterOperands RegOpers; |
635 | RegOpers.collect(MI, TRI: *TRI, MRI: *MRI, TrackLaneMasks: false, IgnoreDead: false); |
636 | RPTracker.recedeSkipDebugValues(); |
637 | assert(&*RPTracker.getPos() == &MI && "RPTracker sync error!" ); |
638 | RPTracker.recede(RegOpers); |
639 | } |
640 | |
641 | // Close the RPTracker to finalize live ins. |
642 | RPTracker.closeRegion(); |
643 | |
644 | return RPTracker.getPressure().MaxSetPressure; |
645 | }; |
646 | |
647 | // For now we only care about float and double type fma. |
648 | unsigned VSSRCLimit = TRI->getRegPressureSetLimit( |
649 | MF: *MBB->getParent(), Idx: PPC::RegisterPressureSets::VSSRC); |
650 | |
651 | // Only reduce register pressure when pressure is high. |
652 | return GetMBBPressure(MBB)[PPC::RegisterPressureSets::VSSRC] > |
653 | (float)VSSRCLimit * FMARPFactor; |
654 | } |
655 | |
656 | bool PPCInstrInfo::isLoadFromConstantPool(MachineInstr *I) const { |
657 | // I has only one memory operand which is load from constant pool. |
658 | if (!I->hasOneMemOperand()) |
659 | return false; |
660 | |
661 | MachineMemOperand *Op = I->memoperands()[0]; |
662 | return Op->isLoad() && Op->getPseudoValue() && |
663 | Op->getPseudoValue()->kind() == PseudoSourceValue::ConstantPool; |
664 | } |
665 | |
666 | Register PPCInstrInfo::generateLoadForNewConst( |
667 | unsigned Idx, MachineInstr *MI, Type *Ty, |
668 | SmallVectorImpl<MachineInstr *> &InsInstrs) const { |
669 | // Now we only support PPC64, Medium code model and P9 with vector. |
670 | // We have immutable pattern to access const pool. See function |
671 | // shouldReduceRegisterPressure. |
672 | assert((Subtarget.isPPC64() && Subtarget.hasP9Vector() && |
673 | Subtarget.getTargetMachine().getCodeModel() == CodeModel::Medium) && |
674 | "Target not supported!\n" ); |
675 | |
676 | MachineFunction *MF = MI->getMF(); |
677 | MachineRegisterInfo *MRI = &MF->getRegInfo(); |
678 | |
679 | // Generate ADDIStocHA8 |
680 | Register VReg1 = MRI->createVirtualRegister(RegClass: &PPC::G8RC_and_G8RC_NOX0RegClass); |
681 | MachineInstrBuilder TOCOffset = |
682 | BuildMI(MF&: *MF, MIMD: MI->getDebugLoc(), MCID: get(Opcode: PPC::ADDIStocHA8), DestReg: VReg1) |
683 | .addReg(RegNo: PPC::X2) |
684 | .addConstantPoolIndex(Idx); |
685 | |
686 | assert((Ty->isFloatTy() || Ty->isDoubleTy()) && |
687 | "Only float and double are supported!" ); |
688 | |
689 | unsigned LoadOpcode; |
690 | // Should be float type or double type. |
691 | if (Ty->isFloatTy()) |
692 | LoadOpcode = PPC::DFLOADf32; |
693 | else |
694 | LoadOpcode = PPC::DFLOADf64; |
695 | |
696 | const TargetRegisterClass *RC = MRI->getRegClass(Reg: MI->getOperand(i: 0).getReg()); |
697 | Register VReg2 = MRI->createVirtualRegister(RegClass: RC); |
698 | MachineMemOperand *MMO = MF->getMachineMemOperand( |
699 | PtrInfo: MachinePointerInfo::getConstantPool(MF&: *MF), F: MachineMemOperand::MOLoad, |
700 | Size: Ty->getScalarSizeInBits() / 8, BaseAlignment: MF->getDataLayout().getPrefTypeAlign(Ty)); |
701 | |
702 | // Generate Load from constant pool. |
703 | MachineInstrBuilder Load = |
704 | BuildMI(MF&: *MF, MIMD: MI->getDebugLoc(), MCID: get(Opcode: LoadOpcode), DestReg: VReg2) |
705 | .addConstantPoolIndex(Idx) |
706 | .addReg(RegNo: VReg1, flags: getKillRegState(B: true)) |
707 | .addMemOperand(MMO); |
708 | |
709 | Load->getOperand(i: 1).setTargetFlags(PPCII::MO_TOC_LO); |
710 | |
711 | // Insert the toc load instructions into InsInstrs. |
712 | InsInstrs.insert(I: InsInstrs.begin(), Elt: Load); |
713 | InsInstrs.insert(I: InsInstrs.begin(), Elt: TOCOffset); |
714 | return VReg2; |
715 | } |
716 | |
717 | // This function returns the const value in constant pool if the \p I is a load |
718 | // from constant pool. |
719 | const Constant * |
720 | PPCInstrInfo::getConstantFromConstantPool(MachineInstr *I) const { |
721 | MachineFunction *MF = I->getMF(); |
722 | MachineRegisterInfo *MRI = &MF->getRegInfo(); |
723 | MachineConstantPool *MCP = MF->getConstantPool(); |
724 | assert(I->mayLoad() && "Should be a load instruction.\n" ); |
725 | for (auto MO : I->uses()) { |
726 | if (!MO.isReg()) |
727 | continue; |
728 | Register Reg = MO.getReg(); |
729 | if (Reg == 0 || !Reg.isVirtual()) |
730 | continue; |
731 | // Find the toc address. |
732 | MachineInstr *DefMI = MRI->getVRegDef(Reg); |
733 | for (auto MO2 : DefMI->uses()) |
734 | if (MO2.isCPI()) |
735 | return (MCP->getConstants())[MO2.getIndex()].Val.ConstVal; |
736 | } |
737 | return nullptr; |
738 | } |
739 | |
740 | CombinerObjective PPCInstrInfo::getCombinerObjective(unsigned Pattern) const { |
741 | switch (Pattern) { |
742 | case PPCMachineCombinerPattern::REASSOC_XY_AMM_BMM: |
743 | case PPCMachineCombinerPattern::REASSOC_XMM_AMM_BMM: |
744 | return CombinerObjective::MustReduceDepth; |
745 | case PPCMachineCombinerPattern::REASSOC_XY_BCA: |
746 | case PPCMachineCombinerPattern::REASSOC_XY_BAC: |
747 | return CombinerObjective::MustReduceRegisterPressure; |
748 | default: |
749 | return TargetInstrInfo::getCombinerObjective(Pattern); |
750 | } |
751 | } |
752 | |
753 | bool PPCInstrInfo::getMachineCombinerPatterns( |
754 | MachineInstr &Root, SmallVectorImpl<unsigned> &Patterns, |
755 | bool DoRegPressureReduce) const { |
756 | // Using the machine combiner in this way is potentially expensive, so |
757 | // restrict to when aggressive optimizations are desired. |
758 | if (Subtarget.getTargetMachine().getOptLevel() != CodeGenOptLevel::Aggressive) |
759 | return false; |
760 | |
761 | if (getFMAPatterns(Root, Patterns, DoRegPressureReduce)) |
762 | return true; |
763 | |
764 | return TargetInstrInfo::getMachineCombinerPatterns(Root, Patterns, |
765 | DoRegPressureReduce); |
766 | } |
767 | |
768 | void PPCInstrInfo::genAlternativeCodeSequence( |
769 | MachineInstr &Root, unsigned Pattern, |
770 | SmallVectorImpl<MachineInstr *> &InsInstrs, |
771 | SmallVectorImpl<MachineInstr *> &DelInstrs, |
772 | DenseMap<unsigned, unsigned> &InstrIdxForVirtReg) const { |
773 | switch (Pattern) { |
774 | case PPCMachineCombinerPattern::REASSOC_XY_AMM_BMM: |
775 | case PPCMachineCombinerPattern::REASSOC_XMM_AMM_BMM: |
776 | case PPCMachineCombinerPattern::REASSOC_XY_BCA: |
777 | case PPCMachineCombinerPattern::REASSOC_XY_BAC: |
778 | reassociateFMA(Root, Pattern, InsInstrs, DelInstrs, InstrIdxForVirtReg); |
779 | break; |
780 | default: |
781 | // Reassociate default patterns. |
782 | TargetInstrInfo::genAlternativeCodeSequence(Root, Pattern, InsInstrs, |
783 | DelInstrs, InstIdxForVirtReg&: InstrIdxForVirtReg); |
784 | break; |
785 | } |
786 | } |
787 | |
788 | void PPCInstrInfo::reassociateFMA( |
789 | MachineInstr &Root, unsigned Pattern, |
790 | SmallVectorImpl<MachineInstr *> &InsInstrs, |
791 | SmallVectorImpl<MachineInstr *> &DelInstrs, |
792 | DenseMap<unsigned, unsigned> &InstrIdxForVirtReg) const { |
793 | MachineFunction *MF = Root.getMF(); |
794 | MachineRegisterInfo &MRI = MF->getRegInfo(); |
795 | const TargetRegisterInfo *TRI = &getRegisterInfo(); |
796 | MachineOperand &OpC = Root.getOperand(i: 0); |
797 | Register RegC = OpC.getReg(); |
798 | const TargetRegisterClass *RC = MRI.getRegClass(Reg: RegC); |
799 | MRI.constrainRegClass(Reg: RegC, RC); |
800 | |
801 | unsigned FmaOp = Root.getOpcode(); |
802 | int16_t Idx = getFMAOpIdxInfo(Opcode: FmaOp); |
803 | assert(Idx >= 0 && "Root must be a FMA instruction" ); |
804 | |
805 | bool IsILPReassociate = |
806 | (Pattern == PPCMachineCombinerPattern::REASSOC_XY_AMM_BMM) || |
807 | (Pattern == PPCMachineCombinerPattern::REASSOC_XMM_AMM_BMM); |
808 | |
809 | uint16_t AddOpIdx = FMAOpIdxInfo[Idx][InfoArrayIdxAddOpIdx]; |
810 | uint16_t FirstMulOpIdx = FMAOpIdxInfo[Idx][InfoArrayIdxMULOpIdx]; |
811 | |
812 | MachineInstr *Prev = nullptr; |
813 | MachineInstr *Leaf = nullptr; |
814 | switch (Pattern) { |
815 | default: |
816 | llvm_unreachable("not recognized pattern!" ); |
817 | case PPCMachineCombinerPattern::REASSOC_XY_AMM_BMM: |
818 | case PPCMachineCombinerPattern::REASSOC_XMM_AMM_BMM: |
819 | Prev = MRI.getUniqueVRegDef(Reg: Root.getOperand(i: AddOpIdx).getReg()); |
820 | Leaf = MRI.getUniqueVRegDef(Reg: Prev->getOperand(i: AddOpIdx).getReg()); |
821 | break; |
822 | case PPCMachineCombinerPattern::REASSOC_XY_BAC: { |
823 | Register MULReg = |
824 | TRI->lookThruCopyLike(SrcReg: Root.getOperand(i: FirstMulOpIdx).getReg(), MRI: &MRI); |
825 | Leaf = MRI.getVRegDef(Reg: MULReg); |
826 | break; |
827 | } |
828 | case PPCMachineCombinerPattern::REASSOC_XY_BCA: { |
829 | Register MULReg = TRI->lookThruCopyLike( |
830 | SrcReg: Root.getOperand(i: FirstMulOpIdx + 1).getReg(), MRI: &MRI); |
831 | Leaf = MRI.getVRegDef(Reg: MULReg); |
832 | break; |
833 | } |
834 | } |
835 | |
836 | uint32_t IntersectedFlags = 0; |
837 | if (IsILPReassociate) |
838 | IntersectedFlags = Root.getFlags() & Prev->getFlags() & Leaf->getFlags(); |
839 | else |
840 | IntersectedFlags = Root.getFlags() & Leaf->getFlags(); |
841 | |
842 | auto GetOperandInfo = [&](const MachineOperand &Operand, Register &Reg, |
843 | bool &KillFlag) { |
844 | Reg = Operand.getReg(); |
845 | MRI.constrainRegClass(Reg, RC); |
846 | KillFlag = Operand.isKill(); |
847 | }; |
848 | |
849 | auto GetFMAInstrInfo = [&](const MachineInstr &Instr, Register &MulOp1, |
850 | Register &MulOp2, Register &AddOp, |
851 | bool &MulOp1KillFlag, bool &MulOp2KillFlag, |
852 | bool &AddOpKillFlag) { |
853 | GetOperandInfo(Instr.getOperand(i: FirstMulOpIdx), MulOp1, MulOp1KillFlag); |
854 | GetOperandInfo(Instr.getOperand(i: FirstMulOpIdx + 1), MulOp2, MulOp2KillFlag); |
855 | GetOperandInfo(Instr.getOperand(i: AddOpIdx), AddOp, AddOpKillFlag); |
856 | }; |
857 | |
858 | Register RegM11, RegM12, RegX, RegY, RegM21, RegM22, RegM31, RegM32, RegA11, |
859 | RegA21, RegB; |
860 | bool KillX = false, KillY = false, KillM11 = false, KillM12 = false, |
861 | KillM21 = false, KillM22 = false, KillM31 = false, KillM32 = false, |
862 | KillA11 = false, KillA21 = false, KillB = false; |
863 | |
864 | GetFMAInstrInfo(Root, RegM31, RegM32, RegB, KillM31, KillM32, KillB); |
865 | |
866 | if (IsILPReassociate) |
867 | GetFMAInstrInfo(*Prev, RegM21, RegM22, RegA21, KillM21, KillM22, KillA21); |
868 | |
869 | if (Pattern == PPCMachineCombinerPattern::REASSOC_XMM_AMM_BMM) { |
870 | GetFMAInstrInfo(*Leaf, RegM11, RegM12, RegA11, KillM11, KillM12, KillA11); |
871 | GetOperandInfo(Leaf->getOperand(i: AddOpIdx), RegX, KillX); |
872 | } else if (Pattern == PPCMachineCombinerPattern::REASSOC_XY_AMM_BMM) { |
873 | GetOperandInfo(Leaf->getOperand(i: 1), RegX, KillX); |
874 | GetOperandInfo(Leaf->getOperand(i: 2), RegY, KillY); |
875 | } else { |
876 | // Get FSUB instruction info. |
877 | GetOperandInfo(Leaf->getOperand(i: 1), RegX, KillX); |
878 | GetOperandInfo(Leaf->getOperand(i: 2), RegY, KillY); |
879 | } |
880 | |
881 | // Create new virtual registers for the new results instead of |
882 | // recycling legacy ones because the MachineCombiner's computation of the |
883 | // critical path requires a new register definition rather than an existing |
884 | // one. |
885 | // For register pressure reassociation, we only need create one virtual |
886 | // register for the new fma. |
887 | Register NewVRA = MRI.createVirtualRegister(RegClass: RC); |
888 | InstrIdxForVirtReg.insert(KV: std::make_pair(x&: NewVRA, y: 0)); |
889 | |
890 | Register NewVRB = 0; |
891 | if (IsILPReassociate) { |
892 | NewVRB = MRI.createVirtualRegister(RegClass: RC); |
893 | InstrIdxForVirtReg.insert(KV: std::make_pair(x&: NewVRB, y: 1)); |
894 | } |
895 | |
896 | Register NewVRD = 0; |
897 | if (Pattern == PPCMachineCombinerPattern::REASSOC_XMM_AMM_BMM) { |
898 | NewVRD = MRI.createVirtualRegister(RegClass: RC); |
899 | InstrIdxForVirtReg.insert(KV: std::make_pair(x&: NewVRD, y: 2)); |
900 | } |
901 | |
902 | auto AdjustOperandOrder = [&](MachineInstr *MI, Register RegAdd, bool KillAdd, |
903 | Register RegMul1, bool KillRegMul1, |
904 | Register RegMul2, bool KillRegMul2) { |
905 | MI->getOperand(i: AddOpIdx).setReg(RegAdd); |
906 | MI->getOperand(i: AddOpIdx).setIsKill(KillAdd); |
907 | MI->getOperand(i: FirstMulOpIdx).setReg(RegMul1); |
908 | MI->getOperand(i: FirstMulOpIdx).setIsKill(KillRegMul1); |
909 | MI->getOperand(i: FirstMulOpIdx + 1).setReg(RegMul2); |
910 | MI->getOperand(i: FirstMulOpIdx + 1).setIsKill(KillRegMul2); |
911 | }; |
912 | |
913 | MachineInstrBuilder NewARegPressure, NewCRegPressure; |
914 | switch (Pattern) { |
915 | default: |
916 | llvm_unreachable("not recognized pattern!" ); |
917 | case PPCMachineCombinerPattern::REASSOC_XY_AMM_BMM: { |
918 | // Create new instructions for insertion. |
919 | MachineInstrBuilder MINewB = |
920 | BuildMI(MF&: *MF, MIMD: Prev->getDebugLoc(), MCID: get(Opcode: FmaOp), DestReg: NewVRB) |
921 | .addReg(RegNo: RegX, flags: getKillRegState(B: KillX)) |
922 | .addReg(RegNo: RegM21, flags: getKillRegState(B: KillM21)) |
923 | .addReg(RegNo: RegM22, flags: getKillRegState(B: KillM22)); |
924 | MachineInstrBuilder MINewA = |
925 | BuildMI(MF&: *MF, MIMD: Root.getDebugLoc(), MCID: get(Opcode: FmaOp), DestReg: NewVRA) |
926 | .addReg(RegNo: RegY, flags: getKillRegState(B: KillY)) |
927 | .addReg(RegNo: RegM31, flags: getKillRegState(B: KillM31)) |
928 | .addReg(RegNo: RegM32, flags: getKillRegState(B: KillM32)); |
929 | // If AddOpIdx is not 1, adjust the order. |
930 | if (AddOpIdx != 1) { |
931 | AdjustOperandOrder(MINewB, RegX, KillX, RegM21, KillM21, RegM22, KillM22); |
932 | AdjustOperandOrder(MINewA, RegY, KillY, RegM31, KillM31, RegM32, KillM32); |
933 | } |
934 | |
935 | MachineInstrBuilder MINewC = |
936 | BuildMI(MF&: *MF, MIMD: Root.getDebugLoc(), |
937 | MCID: get(Opcode: FMAOpIdxInfo[Idx][InfoArrayIdxFAddInst]), DestReg: RegC) |
938 | .addReg(RegNo: NewVRB, flags: getKillRegState(B: true)) |
939 | .addReg(RegNo: NewVRA, flags: getKillRegState(B: true)); |
940 | |
941 | // Update flags for newly created instructions. |
942 | setSpecialOperandAttr(MI&: *MINewA, Flags: IntersectedFlags); |
943 | setSpecialOperandAttr(MI&: *MINewB, Flags: IntersectedFlags); |
944 | setSpecialOperandAttr(MI&: *MINewC, Flags: IntersectedFlags); |
945 | |
946 | // Record new instructions for insertion. |
947 | InsInstrs.push_back(Elt: MINewA); |
948 | InsInstrs.push_back(Elt: MINewB); |
949 | InsInstrs.push_back(Elt: MINewC); |
950 | break; |
951 | } |
952 | case PPCMachineCombinerPattern::REASSOC_XMM_AMM_BMM: { |
953 | assert(NewVRD && "new FMA register not created!" ); |
954 | // Create new instructions for insertion. |
955 | MachineInstrBuilder MINewA = |
956 | BuildMI(MF&: *MF, MIMD: Leaf->getDebugLoc(), |
957 | MCID: get(Opcode: FMAOpIdxInfo[Idx][InfoArrayIdxFMULInst]), DestReg: NewVRA) |
958 | .addReg(RegNo: RegM11, flags: getKillRegState(B: KillM11)) |
959 | .addReg(RegNo: RegM12, flags: getKillRegState(B: KillM12)); |
960 | MachineInstrBuilder MINewB = |
961 | BuildMI(MF&: *MF, MIMD: Prev->getDebugLoc(), MCID: get(Opcode: FmaOp), DestReg: NewVRB) |
962 | .addReg(RegNo: RegX, flags: getKillRegState(B: KillX)) |
963 | .addReg(RegNo: RegM21, flags: getKillRegState(B: KillM21)) |
964 | .addReg(RegNo: RegM22, flags: getKillRegState(B: KillM22)); |
965 | MachineInstrBuilder MINewD = |
966 | BuildMI(MF&: *MF, MIMD: Root.getDebugLoc(), MCID: get(Opcode: FmaOp), DestReg: NewVRD) |
967 | .addReg(RegNo: NewVRA, flags: getKillRegState(B: true)) |
968 | .addReg(RegNo: RegM31, flags: getKillRegState(B: KillM31)) |
969 | .addReg(RegNo: RegM32, flags: getKillRegState(B: KillM32)); |
970 | // If AddOpIdx is not 1, adjust the order. |
971 | if (AddOpIdx != 1) { |
972 | AdjustOperandOrder(MINewB, RegX, KillX, RegM21, KillM21, RegM22, KillM22); |
973 | AdjustOperandOrder(MINewD, NewVRA, true, RegM31, KillM31, RegM32, |
974 | KillM32); |
975 | } |
976 | |
977 | MachineInstrBuilder MINewC = |
978 | BuildMI(MF&: *MF, MIMD: Root.getDebugLoc(), |
979 | MCID: get(Opcode: FMAOpIdxInfo[Idx][InfoArrayIdxFAddInst]), DestReg: RegC) |
980 | .addReg(RegNo: NewVRB, flags: getKillRegState(B: true)) |
981 | .addReg(RegNo: NewVRD, flags: getKillRegState(B: true)); |
982 | |
983 | // Update flags for newly created instructions. |
984 | setSpecialOperandAttr(MI&: *MINewA, Flags: IntersectedFlags); |
985 | setSpecialOperandAttr(MI&: *MINewB, Flags: IntersectedFlags); |
986 | setSpecialOperandAttr(MI&: *MINewD, Flags: IntersectedFlags); |
987 | setSpecialOperandAttr(MI&: *MINewC, Flags: IntersectedFlags); |
988 | |
989 | // Record new instructions for insertion. |
990 | InsInstrs.push_back(Elt: MINewA); |
991 | InsInstrs.push_back(Elt: MINewB); |
992 | InsInstrs.push_back(Elt: MINewD); |
993 | InsInstrs.push_back(Elt: MINewC); |
994 | break; |
995 | } |
996 | case PPCMachineCombinerPattern::REASSOC_XY_BAC: |
997 | case PPCMachineCombinerPattern::REASSOC_XY_BCA: { |
998 | Register VarReg; |
999 | bool KillVarReg = false; |
1000 | if (Pattern == PPCMachineCombinerPattern::REASSOC_XY_BCA) { |
1001 | VarReg = RegM31; |
1002 | KillVarReg = KillM31; |
1003 | } else { |
1004 | VarReg = RegM32; |
1005 | KillVarReg = KillM32; |
1006 | } |
1007 | // We don't want to get negative const from memory pool too early, as the |
1008 | // created entry will not be deleted even if it has no users. Since all |
1009 | // operand of Leaf and Root are virtual register, we use zero register |
1010 | // here as a placeholder. When the InsInstrs is selected in |
1011 | // MachineCombiner, we call finalizeInsInstrs to replace the zero register |
1012 | // with a virtual register which is a load from constant pool. |
1013 | NewARegPressure = BuildMI(MF&: *MF, MIMD: Root.getDebugLoc(), MCID: get(Opcode: FmaOp), DestReg: NewVRA) |
1014 | .addReg(RegNo: RegB, flags: getKillRegState(B: RegB)) |
1015 | .addReg(RegNo: RegY, flags: getKillRegState(B: KillY)) |
1016 | .addReg(RegNo: PPC::ZERO8); |
1017 | NewCRegPressure = BuildMI(MF&: *MF, MIMD: Root.getDebugLoc(), MCID: get(Opcode: FmaOp), DestReg: RegC) |
1018 | .addReg(RegNo: NewVRA, flags: getKillRegState(B: true)) |
1019 | .addReg(RegNo: RegX, flags: getKillRegState(B: KillX)) |
1020 | .addReg(RegNo: VarReg, flags: getKillRegState(B: KillVarReg)); |
1021 | // For now, we only support xsmaddadp/xsmaddasp, their add operand are |
1022 | // both at index 1, no need to adjust. |
1023 | // FIXME: when add more fma instructions support, like fma/fmas, adjust |
1024 | // the operand index here. |
1025 | break; |
1026 | } |
1027 | } |
1028 | |
1029 | if (!IsILPReassociate) { |
1030 | setSpecialOperandAttr(MI&: *NewARegPressure, Flags: IntersectedFlags); |
1031 | setSpecialOperandAttr(MI&: *NewCRegPressure, Flags: IntersectedFlags); |
1032 | |
1033 | InsInstrs.push_back(Elt: NewARegPressure); |
1034 | InsInstrs.push_back(Elt: NewCRegPressure); |
1035 | } |
1036 | |
1037 | assert(!InsInstrs.empty() && |
1038 | "Insertion instructions set should not be empty!" ); |
1039 | |
1040 | // Record old instructions for deletion. |
1041 | DelInstrs.push_back(Elt: Leaf); |
1042 | if (IsILPReassociate) |
1043 | DelInstrs.push_back(Elt: Prev); |
1044 | DelInstrs.push_back(Elt: &Root); |
1045 | } |
1046 | |
1047 | // Detect 32 -> 64-bit extensions where we may reuse the low sub-register. |
1048 | bool PPCInstrInfo::isCoalescableExtInstr(const MachineInstr &MI, |
1049 | Register &SrcReg, Register &DstReg, |
1050 | unsigned &SubIdx) const { |
1051 | switch (MI.getOpcode()) { |
1052 | default: return false; |
1053 | case PPC::EXTSW: |
1054 | case PPC::EXTSW_32: |
1055 | case PPC::EXTSW_32_64: |
1056 | SrcReg = MI.getOperand(i: 1).getReg(); |
1057 | DstReg = MI.getOperand(i: 0).getReg(); |
1058 | SubIdx = PPC::sub_32; |
1059 | return true; |
1060 | } |
1061 | } |
1062 | |
1063 | Register PPCInstrInfo::isLoadFromStackSlot(const MachineInstr &MI, |
1064 | int &FrameIndex) const { |
1065 | if (llvm::is_contained(Range: getLoadOpcodesForSpillArray(), Element: MI.getOpcode())) { |
1066 | // Check for the operands added by addFrameReference (the immediate is the |
1067 | // offset which defaults to 0). |
1068 | if (MI.getOperand(i: 1).isImm() && !MI.getOperand(i: 1).getImm() && |
1069 | MI.getOperand(i: 2).isFI()) { |
1070 | FrameIndex = MI.getOperand(i: 2).getIndex(); |
1071 | return MI.getOperand(i: 0).getReg(); |
1072 | } |
1073 | } |
1074 | return 0; |
1075 | } |
1076 | |
1077 | // For opcodes with the ReMaterializable flag set, this function is called to |
1078 | // verify the instruction is really rematable. |
1079 | bool PPCInstrInfo::isReallyTriviallyReMaterializable( |
1080 | const MachineInstr &MI) const { |
1081 | switch (MI.getOpcode()) { |
1082 | default: |
1083 | // Let base implementaion decide. |
1084 | break; |
1085 | case PPC::LI: |
1086 | case PPC::LI8: |
1087 | case PPC::PLI: |
1088 | case PPC::PLI8: |
1089 | case PPC::LIS: |
1090 | case PPC::LIS8: |
1091 | case PPC::ADDIStocHA: |
1092 | case PPC::ADDIStocHA8: |
1093 | case PPC::ADDItocL: |
1094 | case PPC::ADDItocL8: |
1095 | case PPC::LOAD_STACK_GUARD: |
1096 | case PPC::PPCLdFixedAddr: |
1097 | case PPC::XXLXORz: |
1098 | case PPC::XXLXORspz: |
1099 | case PPC::XXLXORdpz: |
1100 | case PPC::XXLEQVOnes: |
1101 | case PPC::XXSPLTI32DX: |
1102 | case PPC::XXSPLTIW: |
1103 | case PPC::XXSPLTIDP: |
1104 | case PPC::V_SET0B: |
1105 | case PPC::V_SET0H: |
1106 | case PPC::V_SET0: |
1107 | case PPC::V_SETALLONESB: |
1108 | case PPC::V_SETALLONESH: |
1109 | case PPC::V_SETALLONES: |
1110 | case PPC::CRSET: |
1111 | case PPC::CRUNSET: |
1112 | case PPC::XXSETACCZ: |
1113 | case PPC::XXSETACCZW: |
1114 | return true; |
1115 | } |
1116 | return TargetInstrInfo::isReallyTriviallyReMaterializable(MI); |
1117 | } |
1118 | |
1119 | Register PPCInstrInfo::isStoreToStackSlot(const MachineInstr &MI, |
1120 | int &FrameIndex) const { |
1121 | if (llvm::is_contained(Range: getStoreOpcodesForSpillArray(), Element: MI.getOpcode())) { |
1122 | if (MI.getOperand(i: 1).isImm() && !MI.getOperand(i: 1).getImm() && |
1123 | MI.getOperand(i: 2).isFI()) { |
1124 | FrameIndex = MI.getOperand(i: 2).getIndex(); |
1125 | return MI.getOperand(i: 0).getReg(); |
1126 | } |
1127 | } |
1128 | return 0; |
1129 | } |
1130 | |
1131 | MachineInstr *PPCInstrInfo::commuteInstructionImpl(MachineInstr &MI, bool NewMI, |
1132 | unsigned OpIdx1, |
1133 | unsigned OpIdx2) const { |
1134 | MachineFunction &MF = *MI.getParent()->getParent(); |
1135 | |
1136 | // Normal instructions can be commuted the obvious way. |
1137 | if (MI.getOpcode() != PPC::RLWIMI && MI.getOpcode() != PPC::RLWIMI_rec) |
1138 | return TargetInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2); |
1139 | // Note that RLWIMI can be commuted as a 32-bit instruction, but not as a |
1140 | // 64-bit instruction (so we don't handle PPC::RLWIMI8 here), because |
1141 | // changing the relative order of the mask operands might change what happens |
1142 | // to the high-bits of the mask (and, thus, the result). |
1143 | |
1144 | // Cannot commute if it has a non-zero rotate count. |
1145 | if (MI.getOperand(i: 3).getImm() != 0) |
1146 | return nullptr; |
1147 | |
1148 | // If we have a zero rotate count, we have: |
1149 | // M = mask(MB,ME) |
1150 | // Op0 = (Op1 & ~M) | (Op2 & M) |
1151 | // Change this to: |
1152 | // M = mask((ME+1)&31, (MB-1)&31) |
1153 | // Op0 = (Op2 & ~M) | (Op1 & M) |
1154 | |
1155 | // Swap op1/op2 |
1156 | assert(((OpIdx1 == 1 && OpIdx2 == 2) || (OpIdx1 == 2 && OpIdx2 == 1)) && |
1157 | "Only the operands 1 and 2 can be swapped in RLSIMI/RLWIMI_rec." ); |
1158 | Register Reg0 = MI.getOperand(i: 0).getReg(); |
1159 | Register Reg1 = MI.getOperand(i: 1).getReg(); |
1160 | Register Reg2 = MI.getOperand(i: 2).getReg(); |
1161 | unsigned SubReg1 = MI.getOperand(i: 1).getSubReg(); |
1162 | unsigned SubReg2 = MI.getOperand(i: 2).getSubReg(); |
1163 | bool Reg1IsKill = MI.getOperand(i: 1).isKill(); |
1164 | bool Reg2IsKill = MI.getOperand(i: 2).isKill(); |
1165 | bool ChangeReg0 = false; |
1166 | // If machine instrs are no longer in two-address forms, update |
1167 | // destination register as well. |
1168 | if (Reg0 == Reg1) { |
1169 | // Must be two address instruction (i.e. op1 is tied to op0). |
1170 | assert(MI.getDesc().getOperandConstraint(1, MCOI::TIED_TO) == 0 && |
1171 | "Expecting a two-address instruction!" ); |
1172 | assert(MI.getOperand(0).getSubReg() == SubReg1 && "Tied subreg mismatch" ); |
1173 | Reg2IsKill = false; |
1174 | ChangeReg0 = true; |
1175 | } |
1176 | |
1177 | // Masks. |
1178 | unsigned MB = MI.getOperand(i: 4).getImm(); |
1179 | unsigned ME = MI.getOperand(i: 5).getImm(); |
1180 | |
1181 | // We can't commute a trivial mask (there is no way to represent an all-zero |
1182 | // mask). |
1183 | if (MB == 0 && ME == 31) |
1184 | return nullptr; |
1185 | |
1186 | if (NewMI) { |
1187 | // Create a new instruction. |
1188 | Register Reg0 = ChangeReg0 ? Reg2 : MI.getOperand(i: 0).getReg(); |
1189 | bool Reg0IsDead = MI.getOperand(i: 0).isDead(); |
1190 | return BuildMI(MF, MIMD: MI.getDebugLoc(), MCID: MI.getDesc()) |
1191 | .addReg(RegNo: Reg0, flags: RegState::Define | getDeadRegState(B: Reg0IsDead)) |
1192 | .addReg(RegNo: Reg2, flags: getKillRegState(B: Reg2IsKill)) |
1193 | .addReg(RegNo: Reg1, flags: getKillRegState(B: Reg1IsKill)) |
1194 | .addImm(Val: (ME + 1) & 31) |
1195 | .addImm(Val: (MB - 1) & 31); |
1196 | } |
1197 | |
1198 | if (ChangeReg0) { |
1199 | MI.getOperand(i: 0).setReg(Reg2); |
1200 | MI.getOperand(i: 0).setSubReg(SubReg2); |
1201 | } |
1202 | MI.getOperand(i: 2).setReg(Reg1); |
1203 | MI.getOperand(i: 1).setReg(Reg2); |
1204 | MI.getOperand(i: 2).setSubReg(SubReg1); |
1205 | MI.getOperand(i: 1).setSubReg(SubReg2); |
1206 | MI.getOperand(i: 2).setIsKill(Reg1IsKill); |
1207 | MI.getOperand(i: 1).setIsKill(Reg2IsKill); |
1208 | |
1209 | // Swap the mask around. |
1210 | MI.getOperand(i: 4).setImm((ME + 1) & 31); |
1211 | MI.getOperand(i: 5).setImm((MB - 1) & 31); |
1212 | return &MI; |
1213 | } |
1214 | |
1215 | bool PPCInstrInfo::findCommutedOpIndices(const MachineInstr &MI, |
1216 | unsigned &SrcOpIdx1, |
1217 | unsigned &SrcOpIdx2) const { |
1218 | // For VSX A-Type FMA instructions, it is the first two operands that can be |
1219 | // commuted, however, because the non-encoded tied input operand is listed |
1220 | // first, the operands to swap are actually the second and third. |
1221 | |
1222 | int AltOpc = PPC::getAltVSXFMAOpcode(Opcode: MI.getOpcode()); |
1223 | if (AltOpc == -1) |
1224 | return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2); |
1225 | |
1226 | // The commutable operand indices are 2 and 3. Return them in SrcOpIdx1 |
1227 | // and SrcOpIdx2. |
1228 | return fixCommutedOpIndices(ResultIdx1&: SrcOpIdx1, ResultIdx2&: SrcOpIdx2, CommutableOpIdx1: 2, CommutableOpIdx2: 3); |
1229 | } |
1230 | |
1231 | void PPCInstrInfo::insertNoop(MachineBasicBlock &MBB, |
1232 | MachineBasicBlock::iterator MI) const { |
1233 | // This function is used for scheduling, and the nop wanted here is the type |
1234 | // that terminates dispatch groups on the POWER cores. |
1235 | unsigned Directive = Subtarget.getCPUDirective(); |
1236 | unsigned Opcode; |
1237 | switch (Directive) { |
1238 | default: Opcode = PPC::NOP; break; |
1239 | case PPC::DIR_PWR6: Opcode = PPC::NOP_GT_PWR6; break; |
1240 | case PPC::DIR_PWR7: Opcode = PPC::NOP_GT_PWR7; break; |
1241 | case PPC::DIR_PWR8: Opcode = PPC::NOP_GT_PWR7; break; /* FIXME: Update when P8 InstrScheduling model is ready */ |
1242 | // FIXME: Update when POWER9 scheduling model is ready. |
1243 | case PPC::DIR_PWR9: Opcode = PPC::NOP_GT_PWR7; break; |
1244 | } |
1245 | |
1246 | DebugLoc DL; |
1247 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: get(Opcode)); |
1248 | } |
1249 | |
1250 | /// Return the noop instruction to use for a noop. |
1251 | MCInst PPCInstrInfo::getNop() const { |
1252 | MCInst Nop; |
1253 | Nop.setOpcode(PPC::NOP); |
1254 | return Nop; |
1255 | } |
1256 | |
1257 | // Branch analysis. |
1258 | // Note: If the condition register is set to CTR or CTR8 then this is a |
1259 | // BDNZ (imm == 1) or BDZ (imm == 0) branch. |
1260 | bool PPCInstrInfo::analyzeBranch(MachineBasicBlock &MBB, |
1261 | MachineBasicBlock *&TBB, |
1262 | MachineBasicBlock *&FBB, |
1263 | SmallVectorImpl<MachineOperand> &Cond, |
1264 | bool AllowModify) const { |
1265 | bool isPPC64 = Subtarget.isPPC64(); |
1266 | |
1267 | // If the block has no terminators, it just falls into the block after it. |
1268 | MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr(); |
1269 | if (I == MBB.end()) |
1270 | return false; |
1271 | |
1272 | if (!isUnpredicatedTerminator(MI: *I)) |
1273 | return false; |
1274 | |
1275 | if (AllowModify) { |
1276 | // If the BB ends with an unconditional branch to the fallthrough BB, |
1277 | // we eliminate the branch instruction. |
1278 | if (I->getOpcode() == PPC::B && |
1279 | MBB.isLayoutSuccessor(MBB: I->getOperand(i: 0).getMBB())) { |
1280 | I->eraseFromParent(); |
1281 | |
1282 | // We update iterator after deleting the last branch. |
1283 | I = MBB.getLastNonDebugInstr(); |
1284 | if (I == MBB.end() || !isUnpredicatedTerminator(MI: *I)) |
1285 | return false; |
1286 | } |
1287 | } |
1288 | |
1289 | // Get the last instruction in the block. |
1290 | MachineInstr &LastInst = *I; |
1291 | |
1292 | // If there is only one terminator instruction, process it. |
1293 | if (I == MBB.begin() || !isUnpredicatedTerminator(MI: *--I)) { |
1294 | if (LastInst.getOpcode() == PPC::B) { |
1295 | if (!LastInst.getOperand(i: 0).isMBB()) |
1296 | return true; |
1297 | TBB = LastInst.getOperand(i: 0).getMBB(); |
1298 | return false; |
1299 | } else if (LastInst.getOpcode() == PPC::BCC) { |
1300 | if (!LastInst.getOperand(i: 2).isMBB()) |
1301 | return true; |
1302 | // Block ends with fall-through condbranch. |
1303 | TBB = LastInst.getOperand(i: 2).getMBB(); |
1304 | Cond.push_back(Elt: LastInst.getOperand(i: 0)); |
1305 | Cond.push_back(Elt: LastInst.getOperand(i: 1)); |
1306 | return false; |
1307 | } else if (LastInst.getOpcode() == PPC::BC) { |
1308 | if (!LastInst.getOperand(i: 1).isMBB()) |
1309 | return true; |
1310 | // Block ends with fall-through condbranch. |
1311 | TBB = LastInst.getOperand(i: 1).getMBB(); |
1312 | Cond.push_back(Elt: MachineOperand::CreateImm(Val: PPC::PRED_BIT_SET)); |
1313 | Cond.push_back(Elt: LastInst.getOperand(i: 0)); |
1314 | return false; |
1315 | } else if (LastInst.getOpcode() == PPC::BCn) { |
1316 | if (!LastInst.getOperand(i: 1).isMBB()) |
1317 | return true; |
1318 | // Block ends with fall-through condbranch. |
1319 | TBB = LastInst.getOperand(i: 1).getMBB(); |
1320 | Cond.push_back(Elt: MachineOperand::CreateImm(Val: PPC::PRED_BIT_UNSET)); |
1321 | Cond.push_back(Elt: LastInst.getOperand(i: 0)); |
1322 | return false; |
1323 | } else if (LastInst.getOpcode() == PPC::BDNZ8 || |
1324 | LastInst.getOpcode() == PPC::BDNZ) { |
1325 | if (!LastInst.getOperand(i: 0).isMBB()) |
1326 | return true; |
1327 | if (DisableCTRLoopAnal) |
1328 | return true; |
1329 | TBB = LastInst.getOperand(i: 0).getMBB(); |
1330 | Cond.push_back(Elt: MachineOperand::CreateImm(Val: 1)); |
1331 | Cond.push_back(Elt: MachineOperand::CreateReg(Reg: isPPC64 ? PPC::CTR8 : PPC::CTR, |
1332 | isDef: true)); |
1333 | return false; |
1334 | } else if (LastInst.getOpcode() == PPC::BDZ8 || |
1335 | LastInst.getOpcode() == PPC::BDZ) { |
1336 | if (!LastInst.getOperand(i: 0).isMBB()) |
1337 | return true; |
1338 | if (DisableCTRLoopAnal) |
1339 | return true; |
1340 | TBB = LastInst.getOperand(i: 0).getMBB(); |
1341 | Cond.push_back(Elt: MachineOperand::CreateImm(Val: 0)); |
1342 | Cond.push_back(Elt: MachineOperand::CreateReg(Reg: isPPC64 ? PPC::CTR8 : PPC::CTR, |
1343 | isDef: true)); |
1344 | return false; |
1345 | } |
1346 | |
1347 | // Otherwise, don't know what this is. |
1348 | return true; |
1349 | } |
1350 | |
1351 | // Get the instruction before it if it's a terminator. |
1352 | MachineInstr &SecondLastInst = *I; |
1353 | |
1354 | // If there are three terminators, we don't know what sort of block this is. |
1355 | if (I != MBB.begin() && isUnpredicatedTerminator(MI: *--I)) |
1356 | return true; |
1357 | |
1358 | // If the block ends with PPC::B and PPC:BCC, handle it. |
1359 | if (SecondLastInst.getOpcode() == PPC::BCC && |
1360 | LastInst.getOpcode() == PPC::B) { |
1361 | if (!SecondLastInst.getOperand(i: 2).isMBB() || |
1362 | !LastInst.getOperand(i: 0).isMBB()) |
1363 | return true; |
1364 | TBB = SecondLastInst.getOperand(i: 2).getMBB(); |
1365 | Cond.push_back(Elt: SecondLastInst.getOperand(i: 0)); |
1366 | Cond.push_back(Elt: SecondLastInst.getOperand(i: 1)); |
1367 | FBB = LastInst.getOperand(i: 0).getMBB(); |
1368 | return false; |
1369 | } else if (SecondLastInst.getOpcode() == PPC::BC && |
1370 | LastInst.getOpcode() == PPC::B) { |
1371 | if (!SecondLastInst.getOperand(i: 1).isMBB() || |
1372 | !LastInst.getOperand(i: 0).isMBB()) |
1373 | return true; |
1374 | TBB = SecondLastInst.getOperand(i: 1).getMBB(); |
1375 | Cond.push_back(Elt: MachineOperand::CreateImm(Val: PPC::PRED_BIT_SET)); |
1376 | Cond.push_back(Elt: SecondLastInst.getOperand(i: 0)); |
1377 | FBB = LastInst.getOperand(i: 0).getMBB(); |
1378 | return false; |
1379 | } else if (SecondLastInst.getOpcode() == PPC::BCn && |
1380 | LastInst.getOpcode() == PPC::B) { |
1381 | if (!SecondLastInst.getOperand(i: 1).isMBB() || |
1382 | !LastInst.getOperand(i: 0).isMBB()) |
1383 | return true; |
1384 | TBB = SecondLastInst.getOperand(i: 1).getMBB(); |
1385 | Cond.push_back(Elt: MachineOperand::CreateImm(Val: PPC::PRED_BIT_UNSET)); |
1386 | Cond.push_back(Elt: SecondLastInst.getOperand(i: 0)); |
1387 | FBB = LastInst.getOperand(i: 0).getMBB(); |
1388 | return false; |
1389 | } else if ((SecondLastInst.getOpcode() == PPC::BDNZ8 || |
1390 | SecondLastInst.getOpcode() == PPC::BDNZ) && |
1391 | LastInst.getOpcode() == PPC::B) { |
1392 | if (!SecondLastInst.getOperand(i: 0).isMBB() || |
1393 | !LastInst.getOperand(i: 0).isMBB()) |
1394 | return true; |
1395 | if (DisableCTRLoopAnal) |
1396 | return true; |
1397 | TBB = SecondLastInst.getOperand(i: 0).getMBB(); |
1398 | Cond.push_back(Elt: MachineOperand::CreateImm(Val: 1)); |
1399 | Cond.push_back(Elt: MachineOperand::CreateReg(Reg: isPPC64 ? PPC::CTR8 : PPC::CTR, |
1400 | isDef: true)); |
1401 | FBB = LastInst.getOperand(i: 0).getMBB(); |
1402 | return false; |
1403 | } else if ((SecondLastInst.getOpcode() == PPC::BDZ8 || |
1404 | SecondLastInst.getOpcode() == PPC::BDZ) && |
1405 | LastInst.getOpcode() == PPC::B) { |
1406 | if (!SecondLastInst.getOperand(i: 0).isMBB() || |
1407 | !LastInst.getOperand(i: 0).isMBB()) |
1408 | return true; |
1409 | if (DisableCTRLoopAnal) |
1410 | return true; |
1411 | TBB = SecondLastInst.getOperand(i: 0).getMBB(); |
1412 | Cond.push_back(Elt: MachineOperand::CreateImm(Val: 0)); |
1413 | Cond.push_back(Elt: MachineOperand::CreateReg(Reg: isPPC64 ? PPC::CTR8 : PPC::CTR, |
1414 | isDef: true)); |
1415 | FBB = LastInst.getOperand(i: 0).getMBB(); |
1416 | return false; |
1417 | } |
1418 | |
1419 | // If the block ends with two PPC:Bs, handle it. The second one is not |
1420 | // executed, so remove it. |
1421 | if (SecondLastInst.getOpcode() == PPC::B && LastInst.getOpcode() == PPC::B) { |
1422 | if (!SecondLastInst.getOperand(i: 0).isMBB()) |
1423 | return true; |
1424 | TBB = SecondLastInst.getOperand(i: 0).getMBB(); |
1425 | I = LastInst; |
1426 | if (AllowModify) |
1427 | I->eraseFromParent(); |
1428 | return false; |
1429 | } |
1430 | |
1431 | // Otherwise, can't handle this. |
1432 | return true; |
1433 | } |
1434 | |
1435 | unsigned PPCInstrInfo::removeBranch(MachineBasicBlock &MBB, |
1436 | int *BytesRemoved) const { |
1437 | assert(!BytesRemoved && "code size not handled" ); |
1438 | |
1439 | MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr(); |
1440 | if (I == MBB.end()) |
1441 | return 0; |
1442 | |
1443 | if (I->getOpcode() != PPC::B && I->getOpcode() != PPC::BCC && |
1444 | I->getOpcode() != PPC::BC && I->getOpcode() != PPC::BCn && |
1445 | I->getOpcode() != PPC::BDNZ8 && I->getOpcode() != PPC::BDNZ && |
1446 | I->getOpcode() != PPC::BDZ8 && I->getOpcode() != PPC::BDZ) |
1447 | return 0; |
1448 | |
1449 | // Remove the branch. |
1450 | I->eraseFromParent(); |
1451 | |
1452 | I = MBB.end(); |
1453 | |
1454 | if (I == MBB.begin()) return 1; |
1455 | --I; |
1456 | if (I->getOpcode() != PPC::BCC && |
1457 | I->getOpcode() != PPC::BC && I->getOpcode() != PPC::BCn && |
1458 | I->getOpcode() != PPC::BDNZ8 && I->getOpcode() != PPC::BDNZ && |
1459 | I->getOpcode() != PPC::BDZ8 && I->getOpcode() != PPC::BDZ) |
1460 | return 1; |
1461 | |
1462 | // Remove the branch. |
1463 | I->eraseFromParent(); |
1464 | return 2; |
1465 | } |
1466 | |
1467 | unsigned PPCInstrInfo::insertBranch(MachineBasicBlock &MBB, |
1468 | MachineBasicBlock *TBB, |
1469 | MachineBasicBlock *FBB, |
1470 | ArrayRef<MachineOperand> Cond, |
1471 | const DebugLoc &DL, |
1472 | int *BytesAdded) const { |
1473 | // Shouldn't be a fall through. |
1474 | assert(TBB && "insertBranch must not be told to insert a fallthrough" ); |
1475 | assert((Cond.size() == 2 || Cond.size() == 0) && |
1476 | "PPC branch conditions have two components!" ); |
1477 | assert(!BytesAdded && "code size not handled" ); |
1478 | |
1479 | bool isPPC64 = Subtarget.isPPC64(); |
1480 | |
1481 | // One-way branch. |
1482 | if (!FBB) { |
1483 | if (Cond.empty()) // Unconditional branch |
1484 | BuildMI(BB: &MBB, MIMD: DL, MCID: get(Opcode: PPC::B)).addMBB(MBB: TBB); |
1485 | else if (Cond[1].getReg() == PPC::CTR || Cond[1].getReg() == PPC::CTR8) |
1486 | BuildMI(BB: &MBB, MIMD: DL, MCID: get(Opcode: Cond[0].getImm() ? |
1487 | (isPPC64 ? PPC::BDNZ8 : PPC::BDNZ) : |
1488 | (isPPC64 ? PPC::BDZ8 : PPC::BDZ))).addMBB(MBB: TBB); |
1489 | else if (Cond[0].getImm() == PPC::PRED_BIT_SET) |
1490 | BuildMI(BB: &MBB, MIMD: DL, MCID: get(Opcode: PPC::BC)).add(MO: Cond[1]).addMBB(MBB: TBB); |
1491 | else if (Cond[0].getImm() == PPC::PRED_BIT_UNSET) |
1492 | BuildMI(BB: &MBB, MIMD: DL, MCID: get(Opcode: PPC::BCn)).add(MO: Cond[1]).addMBB(MBB: TBB); |
1493 | else // Conditional branch |
1494 | BuildMI(BB: &MBB, MIMD: DL, MCID: get(Opcode: PPC::BCC)) |
1495 | .addImm(Val: Cond[0].getImm()) |
1496 | .add(MO: Cond[1]) |
1497 | .addMBB(MBB: TBB); |
1498 | return 1; |
1499 | } |
1500 | |
1501 | // Two-way Conditional Branch. |
1502 | if (Cond[1].getReg() == PPC::CTR || Cond[1].getReg() == PPC::CTR8) |
1503 | BuildMI(BB: &MBB, MIMD: DL, MCID: get(Opcode: Cond[0].getImm() ? |
1504 | (isPPC64 ? PPC::BDNZ8 : PPC::BDNZ) : |
1505 | (isPPC64 ? PPC::BDZ8 : PPC::BDZ))).addMBB(MBB: TBB); |
1506 | else if (Cond[0].getImm() == PPC::PRED_BIT_SET) |
1507 | BuildMI(BB: &MBB, MIMD: DL, MCID: get(Opcode: PPC::BC)).add(MO: Cond[1]).addMBB(MBB: TBB); |
1508 | else if (Cond[0].getImm() == PPC::PRED_BIT_UNSET) |
1509 | BuildMI(BB: &MBB, MIMD: DL, MCID: get(Opcode: PPC::BCn)).add(MO: Cond[1]).addMBB(MBB: TBB); |
1510 | else |
1511 | BuildMI(BB: &MBB, MIMD: DL, MCID: get(Opcode: PPC::BCC)) |
1512 | .addImm(Val: Cond[0].getImm()) |
1513 | .add(MO: Cond[1]) |
1514 | .addMBB(MBB: TBB); |
1515 | BuildMI(BB: &MBB, MIMD: DL, MCID: get(Opcode: PPC::B)).addMBB(MBB: FBB); |
1516 | return 2; |
1517 | } |
1518 | |
1519 | // Select analysis. |
1520 | bool PPCInstrInfo::canInsertSelect(const MachineBasicBlock &MBB, |
1521 | ArrayRef<MachineOperand> Cond, |
1522 | Register DstReg, Register TrueReg, |
1523 | Register FalseReg, int &CondCycles, |
1524 | int &TrueCycles, int &FalseCycles) const { |
1525 | if (!Subtarget.hasISEL()) |
1526 | return false; |
1527 | |
1528 | if (Cond.size() != 2) |
1529 | return false; |
1530 | |
1531 | // If this is really a bdnz-like condition, then it cannot be turned into a |
1532 | // select. |
1533 | if (Cond[1].getReg() == PPC::CTR || Cond[1].getReg() == PPC::CTR8) |
1534 | return false; |
1535 | |
1536 | // If the conditional branch uses a physical register, then it cannot be |
1537 | // turned into a select. |
1538 | if (Cond[1].getReg().isPhysical()) |
1539 | return false; |
1540 | |
1541 | // Check register classes. |
1542 | const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); |
1543 | const TargetRegisterClass *RC = |
1544 | RI.getCommonSubClass(A: MRI.getRegClass(Reg: TrueReg), B: MRI.getRegClass(Reg: FalseReg)); |
1545 | if (!RC) |
1546 | return false; |
1547 | |
1548 | // isel is for regular integer GPRs only. |
1549 | if (!PPC::GPRCRegClass.hasSubClassEq(RC) && |
1550 | !PPC::GPRC_NOR0RegClass.hasSubClassEq(RC) && |
1551 | !PPC::G8RCRegClass.hasSubClassEq(RC) && |
1552 | !PPC::G8RC_NOX0RegClass.hasSubClassEq(RC)) |
1553 | return false; |
1554 | |
1555 | // FIXME: These numbers are for the A2, how well they work for other cores is |
1556 | // an open question. On the A2, the isel instruction has a 2-cycle latency |
1557 | // but single-cycle throughput. These numbers are used in combination with |
1558 | // the MispredictPenalty setting from the active SchedMachineModel. |
1559 | CondCycles = 1; |
1560 | TrueCycles = 1; |
1561 | FalseCycles = 1; |
1562 | |
1563 | return true; |
1564 | } |
1565 | |
1566 | void PPCInstrInfo::insertSelect(MachineBasicBlock &MBB, |
1567 | MachineBasicBlock::iterator MI, |
1568 | const DebugLoc &dl, Register DestReg, |
1569 | ArrayRef<MachineOperand> Cond, Register TrueReg, |
1570 | Register FalseReg) const { |
1571 | assert(Cond.size() == 2 && |
1572 | "PPC branch conditions have two components!" ); |
1573 | |
1574 | // Get the register classes. |
1575 | MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); |
1576 | const TargetRegisterClass *RC = |
1577 | RI.getCommonSubClass(A: MRI.getRegClass(Reg: TrueReg), B: MRI.getRegClass(Reg: FalseReg)); |
1578 | assert(RC && "TrueReg and FalseReg must have overlapping register classes" ); |
1579 | |
1580 | bool Is64Bit = PPC::G8RCRegClass.hasSubClassEq(RC) || |
1581 | PPC::G8RC_NOX0RegClass.hasSubClassEq(RC); |
1582 | assert((Is64Bit || |
1583 | PPC::GPRCRegClass.hasSubClassEq(RC) || |
1584 | PPC::GPRC_NOR0RegClass.hasSubClassEq(RC)) && |
1585 | "isel is for regular integer GPRs only" ); |
1586 | |
1587 | unsigned OpCode = Is64Bit ? PPC::ISEL8 : PPC::ISEL; |
1588 | auto SelectPred = static_cast<PPC::Predicate>(Cond[0].getImm()); |
1589 | |
1590 | unsigned SubIdx = 0; |
1591 | bool SwapOps = false; |
1592 | switch (SelectPred) { |
1593 | case PPC::PRED_EQ: |
1594 | case PPC::PRED_EQ_MINUS: |
1595 | case PPC::PRED_EQ_PLUS: |
1596 | SubIdx = PPC::sub_eq; SwapOps = false; break; |
1597 | case PPC::PRED_NE: |
1598 | case PPC::PRED_NE_MINUS: |
1599 | case PPC::PRED_NE_PLUS: |
1600 | SubIdx = PPC::sub_eq; SwapOps = true; break; |
1601 | case PPC::PRED_LT: |
1602 | case PPC::PRED_LT_MINUS: |
1603 | case PPC::PRED_LT_PLUS: |
1604 | SubIdx = PPC::sub_lt; SwapOps = false; break; |
1605 | case PPC::PRED_GE: |
1606 | case PPC::PRED_GE_MINUS: |
1607 | case PPC::PRED_GE_PLUS: |
1608 | SubIdx = PPC::sub_lt; SwapOps = true; break; |
1609 | case PPC::PRED_GT: |
1610 | case PPC::PRED_GT_MINUS: |
1611 | case PPC::PRED_GT_PLUS: |
1612 | SubIdx = PPC::sub_gt; SwapOps = false; break; |
1613 | case PPC::PRED_LE: |
1614 | case PPC::PRED_LE_MINUS: |
1615 | case PPC::PRED_LE_PLUS: |
1616 | SubIdx = PPC::sub_gt; SwapOps = true; break; |
1617 | case PPC::PRED_UN: |
1618 | case PPC::PRED_UN_MINUS: |
1619 | case PPC::PRED_UN_PLUS: |
1620 | SubIdx = PPC::sub_un; SwapOps = false; break; |
1621 | case PPC::PRED_NU: |
1622 | case PPC::PRED_NU_MINUS: |
1623 | case PPC::PRED_NU_PLUS: |
1624 | SubIdx = PPC::sub_un; SwapOps = true; break; |
1625 | case PPC::PRED_BIT_SET: SubIdx = 0; SwapOps = false; break; |
1626 | case PPC::PRED_BIT_UNSET: SubIdx = 0; SwapOps = true; break; |
1627 | } |
1628 | |
1629 | Register FirstReg = SwapOps ? FalseReg : TrueReg, |
1630 | SecondReg = SwapOps ? TrueReg : FalseReg; |
1631 | |
1632 | // The first input register of isel cannot be r0. If it is a member |
1633 | // of a register class that can be r0, then copy it first (the |
1634 | // register allocator should eliminate the copy). |
1635 | if (MRI.getRegClass(Reg: FirstReg)->contains(Reg: PPC::R0) || |
1636 | MRI.getRegClass(Reg: FirstReg)->contains(Reg: PPC::X0)) { |
1637 | const TargetRegisterClass *FirstRC = |
1638 | MRI.getRegClass(Reg: FirstReg)->contains(Reg: PPC::X0) ? |
1639 | &PPC::G8RC_NOX0RegClass : &PPC::GPRC_NOR0RegClass; |
1640 | Register OldFirstReg = FirstReg; |
1641 | FirstReg = MRI.createVirtualRegister(RegClass: FirstRC); |
1642 | BuildMI(BB&: MBB, I: MI, MIMD: dl, MCID: get(Opcode: TargetOpcode::COPY), DestReg: FirstReg) |
1643 | .addReg(RegNo: OldFirstReg); |
1644 | } |
1645 | |
1646 | BuildMI(BB&: MBB, I: MI, MIMD: dl, MCID: get(Opcode: OpCode), DestReg) |
1647 | .addReg(RegNo: FirstReg).addReg(RegNo: SecondReg) |
1648 | .addReg(RegNo: Cond[1].getReg(), flags: 0, SubReg: SubIdx); |
1649 | } |
1650 | |
1651 | static unsigned getCRBitValue(unsigned CRBit) { |
1652 | unsigned Ret = 4; |
1653 | if (CRBit == PPC::CR0LT || CRBit == PPC::CR1LT || |
1654 | CRBit == PPC::CR2LT || CRBit == PPC::CR3LT || |
1655 | CRBit == PPC::CR4LT || CRBit == PPC::CR5LT || |
1656 | CRBit == PPC::CR6LT || CRBit == PPC::CR7LT) |
1657 | Ret = 3; |
1658 | if (CRBit == PPC::CR0GT || CRBit == PPC::CR1GT || |
1659 | CRBit == PPC::CR2GT || CRBit == PPC::CR3GT || |
1660 | CRBit == PPC::CR4GT || CRBit == PPC::CR5GT || |
1661 | CRBit == PPC::CR6GT || CRBit == PPC::CR7GT) |
1662 | Ret = 2; |
1663 | if (CRBit == PPC::CR0EQ || CRBit == PPC::CR1EQ || |
1664 | CRBit == PPC::CR2EQ || CRBit == PPC::CR3EQ || |
1665 | CRBit == PPC::CR4EQ || CRBit == PPC::CR5EQ || |
1666 | CRBit == PPC::CR6EQ || CRBit == PPC::CR7EQ) |
1667 | Ret = 1; |
1668 | if (CRBit == PPC::CR0UN || CRBit == PPC::CR1UN || |
1669 | CRBit == PPC::CR2UN || CRBit == PPC::CR3UN || |
1670 | CRBit == PPC::CR4UN || CRBit == PPC::CR5UN || |
1671 | CRBit == PPC::CR6UN || CRBit == PPC::CR7UN) |
1672 | Ret = 0; |
1673 | |
1674 | assert(Ret != 4 && "Invalid CR bit register" ); |
1675 | return Ret; |
1676 | } |
1677 | |
1678 | void PPCInstrInfo::copyPhysReg(MachineBasicBlock &MBB, |
1679 | MachineBasicBlock::iterator I, |
1680 | const DebugLoc &DL, MCRegister DestReg, |
1681 | MCRegister SrcReg, bool KillSrc) const { |
1682 | // We can end up with self copies and similar things as a result of VSX copy |
1683 | // legalization. Promote them here. |
1684 | const TargetRegisterInfo *TRI = &getRegisterInfo(); |
1685 | if (PPC::F8RCRegClass.contains(Reg: DestReg) && |
1686 | PPC::VSRCRegClass.contains(Reg: SrcReg)) { |
1687 | MCRegister SuperReg = |
1688 | TRI->getMatchingSuperReg(Reg: DestReg, SubIdx: PPC::sub_64, RC: &PPC::VSRCRegClass); |
1689 | |
1690 | if (VSXSelfCopyCrash && SrcReg == SuperReg) |
1691 | llvm_unreachable("nop VSX copy" ); |
1692 | |
1693 | DestReg = SuperReg; |
1694 | } else if (PPC::F8RCRegClass.contains(Reg: SrcReg) && |
1695 | PPC::VSRCRegClass.contains(Reg: DestReg)) { |
1696 | MCRegister SuperReg = |
1697 | TRI->getMatchingSuperReg(Reg: SrcReg, SubIdx: PPC::sub_64, RC: &PPC::VSRCRegClass); |
1698 | |
1699 | if (VSXSelfCopyCrash && DestReg == SuperReg) |
1700 | llvm_unreachable("nop VSX copy" ); |
1701 | |
1702 | SrcReg = SuperReg; |
1703 | } |
1704 | |
1705 | // Different class register copy |
1706 | if (PPC::CRBITRCRegClass.contains(Reg: SrcReg) && |
1707 | PPC::GPRCRegClass.contains(Reg: DestReg)) { |
1708 | MCRegister CRReg = getCRFromCRBit(SrcReg); |
1709 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: PPC::MFOCRF), DestReg).addReg(RegNo: CRReg); |
1710 | getKillRegState(B: KillSrc); |
1711 | // Rotate the CR bit in the CR fields to be the least significant bit and |
1712 | // then mask with 0x1 (MB = ME = 31). |
1713 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: PPC::RLWINM), DestReg) |
1714 | .addReg(RegNo: DestReg, flags: RegState::Kill) |
1715 | .addImm(Val: TRI->getEncodingValue(RegNo: CRReg) * 4 + (4 - getCRBitValue(CRBit: SrcReg))) |
1716 | .addImm(Val: 31) |
1717 | .addImm(Val: 31); |
1718 | return; |
1719 | } else if (PPC::CRRCRegClass.contains(Reg: SrcReg) && |
1720 | (PPC::G8RCRegClass.contains(Reg: DestReg) || |
1721 | PPC::GPRCRegClass.contains(Reg: DestReg))) { |
1722 | bool Is64Bit = PPC::G8RCRegClass.contains(Reg: DestReg); |
1723 | unsigned MvCode = Is64Bit ? PPC::MFOCRF8 : PPC::MFOCRF; |
1724 | unsigned ShCode = Is64Bit ? PPC::RLWINM8 : PPC::RLWINM; |
1725 | unsigned CRNum = TRI->getEncodingValue(RegNo: SrcReg); |
1726 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: MvCode), DestReg).addReg(RegNo: SrcReg); |
1727 | getKillRegState(B: KillSrc); |
1728 | if (CRNum == 7) |
1729 | return; |
1730 | // Shift the CR bits to make the CR field in the lowest 4 bits of GRC. |
1731 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: ShCode), DestReg) |
1732 | .addReg(RegNo: DestReg, flags: RegState::Kill) |
1733 | .addImm(Val: CRNum * 4 + 4) |
1734 | .addImm(Val: 28) |
1735 | .addImm(Val: 31); |
1736 | return; |
1737 | } else if (PPC::G8RCRegClass.contains(Reg: SrcReg) && |
1738 | PPC::VSFRCRegClass.contains(Reg: DestReg)) { |
1739 | assert(Subtarget.hasDirectMove() && |
1740 | "Subtarget doesn't support directmove, don't know how to copy." ); |
1741 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: PPC::MTVSRD), DestReg).addReg(RegNo: SrcReg); |
1742 | NumGPRtoVSRSpill++; |
1743 | getKillRegState(B: KillSrc); |
1744 | return; |
1745 | } else if (PPC::VSFRCRegClass.contains(Reg: SrcReg) && |
1746 | PPC::G8RCRegClass.contains(Reg: DestReg)) { |
1747 | assert(Subtarget.hasDirectMove() && |
1748 | "Subtarget doesn't support directmove, don't know how to copy." ); |
1749 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: PPC::MFVSRD), DestReg).addReg(RegNo: SrcReg); |
1750 | getKillRegState(B: KillSrc); |
1751 | return; |
1752 | } else if (PPC::SPERCRegClass.contains(Reg: SrcReg) && |
1753 | PPC::GPRCRegClass.contains(Reg: DestReg)) { |
1754 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: PPC::EFSCFD), DestReg).addReg(RegNo: SrcReg); |
1755 | getKillRegState(B: KillSrc); |
1756 | return; |
1757 | } else if (PPC::GPRCRegClass.contains(Reg: SrcReg) && |
1758 | PPC::SPERCRegClass.contains(Reg: DestReg)) { |
1759 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: PPC::EFDCFS), DestReg).addReg(RegNo: SrcReg); |
1760 | getKillRegState(B: KillSrc); |
1761 | return; |
1762 | } |
1763 | |
1764 | unsigned Opc; |
1765 | if (PPC::GPRCRegClass.contains(Reg1: DestReg, Reg2: SrcReg)) |
1766 | Opc = PPC::OR; |
1767 | else if (PPC::G8RCRegClass.contains(Reg1: DestReg, Reg2: SrcReg)) |
1768 | Opc = PPC::OR8; |
1769 | else if (PPC::F4RCRegClass.contains(Reg1: DestReg, Reg2: SrcReg)) |
1770 | Opc = PPC::FMR; |
1771 | else if (PPC::CRRCRegClass.contains(Reg1: DestReg, Reg2: SrcReg)) |
1772 | Opc = PPC::MCRF; |
1773 | else if (PPC::VRRCRegClass.contains(Reg1: DestReg, Reg2: SrcReg)) |
1774 | Opc = PPC::VOR; |
1775 | else if (PPC::VSRCRegClass.contains(Reg1: DestReg, Reg2: SrcReg)) |
1776 | // There are two different ways this can be done: |
1777 | // 1. xxlor : This has lower latency (on the P7), 2 cycles, but can only |
1778 | // issue in VSU pipeline 0. |
1779 | // 2. xmovdp/xmovsp: This has higher latency (on the P7), 6 cycles, but |
1780 | // can go to either pipeline. |
1781 | // We'll always use xxlor here, because in practically all cases where |
1782 | // copies are generated, they are close enough to some use that the |
1783 | // lower-latency form is preferable. |
1784 | Opc = PPC::XXLOR; |
1785 | else if (PPC::VSFRCRegClass.contains(Reg1: DestReg, Reg2: SrcReg) || |
1786 | PPC::VSSRCRegClass.contains(Reg1: DestReg, Reg2: SrcReg)) |
1787 | Opc = (Subtarget.hasP9Vector()) ? PPC::XSCPSGNDP : PPC::XXLORf; |
1788 | else if (Subtarget.pairedVectorMemops() && |
1789 | PPC::VSRpRCRegClass.contains(Reg1: DestReg, Reg2: SrcReg)) { |
1790 | if (SrcReg > PPC::VSRp15) |
1791 | SrcReg = PPC::V0 + (SrcReg - PPC::VSRp16) * 2; |
1792 | else |
1793 | SrcReg = PPC::VSL0 + (SrcReg - PPC::VSRp0) * 2; |
1794 | if (DestReg > PPC::VSRp15) |
1795 | DestReg = PPC::V0 + (DestReg - PPC::VSRp16) * 2; |
1796 | else |
1797 | DestReg = PPC::VSL0 + (DestReg - PPC::VSRp0) * 2; |
1798 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: PPC::XXLOR), DestReg). |
1799 | addReg(RegNo: SrcReg).addReg(RegNo: SrcReg, flags: getKillRegState(B: KillSrc)); |
1800 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: PPC::XXLOR), DestReg: DestReg + 1). |
1801 | addReg(RegNo: SrcReg + 1).addReg(RegNo: SrcReg + 1, flags: getKillRegState(B: KillSrc)); |
1802 | return; |
1803 | } |
1804 | else if (PPC::CRBITRCRegClass.contains(Reg1: DestReg, Reg2: SrcReg)) |
1805 | Opc = PPC::CROR; |
1806 | else if (PPC::SPERCRegClass.contains(Reg1: DestReg, Reg2: SrcReg)) |
1807 | Opc = PPC::EVOR; |
1808 | else if ((PPC::ACCRCRegClass.contains(Reg: DestReg) || |
1809 | PPC::UACCRCRegClass.contains(Reg: DestReg)) && |
1810 | (PPC::ACCRCRegClass.contains(Reg: SrcReg) || |
1811 | PPC::UACCRCRegClass.contains(Reg: SrcReg))) { |
1812 | // If primed, de-prime the source register, copy the individual registers |
1813 | // and prime the destination if needed. The vector subregisters are |
1814 | // vs[(u)acc * 4] - vs[(u)acc * 4 + 3]. If the copy is not a kill and the |
1815 | // source is primed, we need to re-prime it after the copy as well. |
1816 | PPCRegisterInfo::emitAccCopyInfo(MBB, DestReg, SrcReg); |
1817 | bool DestPrimed = PPC::ACCRCRegClass.contains(Reg: DestReg); |
1818 | bool SrcPrimed = PPC::ACCRCRegClass.contains(Reg: SrcReg); |
1819 | MCRegister VSLSrcReg = |
1820 | PPC::VSL0 + (SrcReg - (SrcPrimed ? PPC::ACC0 : PPC::UACC0)) * 4; |
1821 | MCRegister VSLDestReg = |
1822 | PPC::VSL0 + (DestReg - (DestPrimed ? PPC::ACC0 : PPC::UACC0)) * 4; |
1823 | if (SrcPrimed) |
1824 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: PPC::XXMFACC), DestReg: SrcReg).addReg(RegNo: SrcReg); |
1825 | for (unsigned Idx = 0; Idx < 4; Idx++) |
1826 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: PPC::XXLOR), DestReg: VSLDestReg + Idx) |
1827 | .addReg(RegNo: VSLSrcReg + Idx) |
1828 | .addReg(RegNo: VSLSrcReg + Idx, flags: getKillRegState(B: KillSrc)); |
1829 | if (DestPrimed) |
1830 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: PPC::XXMTACC), DestReg).addReg(RegNo: DestReg); |
1831 | if (SrcPrimed && !KillSrc) |
1832 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: PPC::XXMTACC), DestReg: SrcReg).addReg(RegNo: SrcReg); |
1833 | return; |
1834 | } else if (PPC::G8pRCRegClass.contains(Reg: DestReg) && |
1835 | PPC::G8pRCRegClass.contains(Reg: SrcReg)) { |
1836 | // TODO: Handle G8RC to G8pRC (and vice versa) copy. |
1837 | unsigned DestRegIdx = DestReg - PPC::G8p0; |
1838 | MCRegister DestRegSub0 = PPC::X0 + 2 * DestRegIdx; |
1839 | MCRegister DestRegSub1 = PPC::X0 + 2 * DestRegIdx + 1; |
1840 | unsigned SrcRegIdx = SrcReg - PPC::G8p0; |
1841 | MCRegister SrcRegSub0 = PPC::X0 + 2 * SrcRegIdx; |
1842 | MCRegister SrcRegSub1 = PPC::X0 + 2 * SrcRegIdx + 1; |
1843 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: PPC::OR8), DestReg: DestRegSub0) |
1844 | .addReg(RegNo: SrcRegSub0) |
1845 | .addReg(RegNo: SrcRegSub0, flags: getKillRegState(B: KillSrc)); |
1846 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: PPC::OR8), DestReg: DestRegSub1) |
1847 | .addReg(RegNo: SrcRegSub1) |
1848 | .addReg(RegNo: SrcRegSub1, flags: getKillRegState(B: KillSrc)); |
1849 | return; |
1850 | } else |
1851 | llvm_unreachable("Impossible reg-to-reg copy" ); |
1852 | |
1853 | const MCInstrDesc &MCID = get(Opcode: Opc); |
1854 | if (MCID.getNumOperands() == 3) |
1855 | BuildMI(BB&: MBB, I, MIMD: DL, MCID, DestReg) |
1856 | .addReg(RegNo: SrcReg).addReg(RegNo: SrcReg, flags: getKillRegState(B: KillSrc)); |
1857 | else |
1858 | BuildMI(BB&: MBB, I, MIMD: DL, MCID, DestReg).addReg(RegNo: SrcReg, flags: getKillRegState(B: KillSrc)); |
1859 | } |
1860 | |
1861 | unsigned PPCInstrInfo::getSpillIndex(const TargetRegisterClass *RC) const { |
1862 | int OpcodeIndex = 0; |
1863 | |
1864 | if (PPC::GPRCRegClass.hasSubClassEq(RC) || |
1865 | PPC::GPRC_NOR0RegClass.hasSubClassEq(RC)) { |
1866 | OpcodeIndex = SOK_Int4Spill; |
1867 | } else if (PPC::G8RCRegClass.hasSubClassEq(RC) || |
1868 | PPC::G8RC_NOX0RegClass.hasSubClassEq(RC)) { |
1869 | OpcodeIndex = SOK_Int8Spill; |
1870 | } else if (PPC::F8RCRegClass.hasSubClassEq(RC)) { |
1871 | OpcodeIndex = SOK_Float8Spill; |
1872 | } else if (PPC::F4RCRegClass.hasSubClassEq(RC)) { |
1873 | OpcodeIndex = SOK_Float4Spill; |
1874 | } else if (PPC::SPERCRegClass.hasSubClassEq(RC)) { |
1875 | OpcodeIndex = SOK_SPESpill; |
1876 | } else if (PPC::CRRCRegClass.hasSubClassEq(RC)) { |
1877 | OpcodeIndex = SOK_CRSpill; |
1878 | } else if (PPC::CRBITRCRegClass.hasSubClassEq(RC)) { |
1879 | OpcodeIndex = SOK_CRBitSpill; |
1880 | } else if (PPC::VRRCRegClass.hasSubClassEq(RC)) { |
1881 | OpcodeIndex = SOK_VRVectorSpill; |
1882 | } else if (PPC::VSRCRegClass.hasSubClassEq(RC)) { |
1883 | OpcodeIndex = SOK_VSXVectorSpill; |
1884 | } else if (PPC::VSFRCRegClass.hasSubClassEq(RC)) { |
1885 | OpcodeIndex = SOK_VectorFloat8Spill; |
1886 | } else if (PPC::VSSRCRegClass.hasSubClassEq(RC)) { |
1887 | OpcodeIndex = SOK_VectorFloat4Spill; |
1888 | } else if (PPC::SPILLTOVSRRCRegClass.hasSubClassEq(RC)) { |
1889 | OpcodeIndex = SOK_SpillToVSR; |
1890 | } else if (PPC::ACCRCRegClass.hasSubClassEq(RC)) { |
1891 | assert(Subtarget.pairedVectorMemops() && |
1892 | "Register unexpected when paired memops are disabled." ); |
1893 | OpcodeIndex = SOK_AccumulatorSpill; |
1894 | } else if (PPC::UACCRCRegClass.hasSubClassEq(RC)) { |
1895 | assert(Subtarget.pairedVectorMemops() && |
1896 | "Register unexpected when paired memops are disabled." ); |
1897 | OpcodeIndex = SOK_UAccumulatorSpill; |
1898 | } else if (PPC::WACCRCRegClass.hasSubClassEq(RC)) { |
1899 | assert(Subtarget.pairedVectorMemops() && |
1900 | "Register unexpected when paired memops are disabled." ); |
1901 | OpcodeIndex = SOK_WAccumulatorSpill; |
1902 | } else if (PPC::VSRpRCRegClass.hasSubClassEq(RC)) { |
1903 | assert(Subtarget.pairedVectorMemops() && |
1904 | "Register unexpected when paired memops are disabled." ); |
1905 | OpcodeIndex = SOK_PairedVecSpill; |
1906 | } else if (PPC::G8pRCRegClass.hasSubClassEq(RC)) { |
1907 | OpcodeIndex = SOK_PairedG8Spill; |
1908 | } else { |
1909 | llvm_unreachable("Unknown regclass!" ); |
1910 | } |
1911 | return OpcodeIndex; |
1912 | } |
1913 | |
1914 | unsigned |
1915 | PPCInstrInfo::getStoreOpcodeForSpill(const TargetRegisterClass *RC) const { |
1916 | ArrayRef<unsigned> OpcodesForSpill = getStoreOpcodesForSpillArray(); |
1917 | return OpcodesForSpill[getSpillIndex(RC)]; |
1918 | } |
1919 | |
1920 | unsigned |
1921 | PPCInstrInfo::getLoadOpcodeForSpill(const TargetRegisterClass *RC) const { |
1922 | ArrayRef<unsigned> OpcodesForSpill = getLoadOpcodesForSpillArray(); |
1923 | return OpcodesForSpill[getSpillIndex(RC)]; |
1924 | } |
1925 | |
1926 | void PPCInstrInfo::StoreRegToStackSlot( |
1927 | MachineFunction &MF, unsigned SrcReg, bool isKill, int FrameIdx, |
1928 | const TargetRegisterClass *RC, |
1929 | SmallVectorImpl<MachineInstr *> &NewMIs) const { |
1930 | unsigned Opcode = getStoreOpcodeForSpill(RC); |
1931 | DebugLoc DL; |
1932 | |
1933 | PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>(); |
1934 | FuncInfo->setHasSpills(); |
1935 | |
1936 | NewMIs.push_back(Elt: addFrameReference( |
1937 | MIB: BuildMI(MF, MIMD: DL, MCID: get(Opcode)).addReg(RegNo: SrcReg, flags: getKillRegState(B: isKill)), |
1938 | FI: FrameIdx)); |
1939 | |
1940 | if (PPC::CRRCRegClass.hasSubClassEq(RC) || |
1941 | PPC::CRBITRCRegClass.hasSubClassEq(RC)) |
1942 | FuncInfo->setSpillsCR(); |
1943 | |
1944 | if (isXFormMemOp(Opcode)) |
1945 | FuncInfo->setHasNonRISpills(); |
1946 | } |
1947 | |
1948 | void PPCInstrInfo::storeRegToStackSlotNoUpd( |
1949 | MachineBasicBlock &MBB, MachineBasicBlock::iterator MI, unsigned SrcReg, |
1950 | bool isKill, int FrameIdx, const TargetRegisterClass *RC, |
1951 | const TargetRegisterInfo *TRI) const { |
1952 | MachineFunction &MF = *MBB.getParent(); |
1953 | SmallVector<MachineInstr *, 4> NewMIs; |
1954 | |
1955 | StoreRegToStackSlot(MF, SrcReg, isKill, FrameIdx, RC, NewMIs); |
1956 | |
1957 | for (MachineInstr *NewMI : NewMIs) |
1958 | MBB.insert(I: MI, MI: NewMI); |
1959 | |
1960 | const MachineFrameInfo &MFI = MF.getFrameInfo(); |
1961 | MachineMemOperand *MMO = MF.getMachineMemOperand( |
1962 | PtrInfo: MachinePointerInfo::getFixedStack(MF, FI: FrameIdx), |
1963 | F: MachineMemOperand::MOStore, Size: MFI.getObjectSize(ObjectIdx: FrameIdx), |
1964 | BaseAlignment: MFI.getObjectAlign(ObjectIdx: FrameIdx)); |
1965 | NewMIs.back()->addMemOperand(MF, MO: MMO); |
1966 | } |
1967 | |
1968 | void PPCInstrInfo::storeRegToStackSlot( |
1969 | MachineBasicBlock &MBB, MachineBasicBlock::iterator MI, Register SrcReg, |
1970 | bool isKill, int FrameIdx, const TargetRegisterClass *RC, |
1971 | const TargetRegisterInfo *TRI, Register VReg) const { |
1972 | // We need to avoid a situation in which the value from a VRRC register is |
1973 | // spilled using an Altivec instruction and reloaded into a VSRC register |
1974 | // using a VSX instruction. The issue with this is that the VSX |
1975 | // load/store instructions swap the doublewords in the vector and the Altivec |
1976 | // ones don't. The register classes on the spill/reload may be different if |
1977 | // the register is defined using an Altivec instruction and is then used by a |
1978 | // VSX instruction. |
1979 | RC = updatedRC(RC); |
1980 | storeRegToStackSlotNoUpd(MBB, MI, SrcReg, isKill, FrameIdx, RC, TRI); |
1981 | } |
1982 | |
1983 | void PPCInstrInfo::LoadRegFromStackSlot(MachineFunction &MF, const DebugLoc &DL, |
1984 | unsigned DestReg, int FrameIdx, |
1985 | const TargetRegisterClass *RC, |
1986 | SmallVectorImpl<MachineInstr *> &NewMIs) |
1987 | const { |
1988 | unsigned Opcode = getLoadOpcodeForSpill(RC); |
1989 | NewMIs.push_back(Elt: addFrameReference(MIB: BuildMI(MF, MIMD: DL, MCID: get(Opcode), DestReg), |
1990 | FI: FrameIdx)); |
1991 | } |
1992 | |
1993 | void PPCInstrInfo::loadRegFromStackSlotNoUpd( |
1994 | MachineBasicBlock &MBB, MachineBasicBlock::iterator MI, unsigned DestReg, |
1995 | int FrameIdx, const TargetRegisterClass *RC, |
1996 | const TargetRegisterInfo *TRI) const { |
1997 | MachineFunction &MF = *MBB.getParent(); |
1998 | SmallVector<MachineInstr*, 4> NewMIs; |
1999 | DebugLoc DL; |
2000 | if (MI != MBB.end()) DL = MI->getDebugLoc(); |
2001 | |
2002 | LoadRegFromStackSlot(MF, DL, DestReg, FrameIdx, RC, NewMIs); |
2003 | |
2004 | for (MachineInstr *NewMI : NewMIs) |
2005 | MBB.insert(I: MI, MI: NewMI); |
2006 | |
2007 | const MachineFrameInfo &MFI = MF.getFrameInfo(); |
2008 | MachineMemOperand *MMO = MF.getMachineMemOperand( |
2009 | PtrInfo: MachinePointerInfo::getFixedStack(MF, FI: FrameIdx), |
2010 | F: MachineMemOperand::MOLoad, Size: MFI.getObjectSize(ObjectIdx: FrameIdx), |
2011 | BaseAlignment: MFI.getObjectAlign(ObjectIdx: FrameIdx)); |
2012 | NewMIs.back()->addMemOperand(MF, MO: MMO); |
2013 | } |
2014 | |
2015 | void PPCInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB, |
2016 | MachineBasicBlock::iterator MI, |
2017 | Register DestReg, int FrameIdx, |
2018 | const TargetRegisterClass *RC, |
2019 | const TargetRegisterInfo *TRI, |
2020 | Register VReg) const { |
2021 | // We need to avoid a situation in which the value from a VRRC register is |
2022 | // spilled using an Altivec instruction and reloaded into a VSRC register |
2023 | // using a VSX instruction. The issue with this is that the VSX |
2024 | // load/store instructions swap the doublewords in the vector and the Altivec |
2025 | // ones don't. The register classes on the spill/reload may be different if |
2026 | // the register is defined using an Altivec instruction and is then used by a |
2027 | // VSX instruction. |
2028 | RC = updatedRC(RC); |
2029 | |
2030 | loadRegFromStackSlotNoUpd(MBB, MI, DestReg, FrameIdx, RC, TRI); |
2031 | } |
2032 | |
2033 | bool PPCInstrInfo:: |
2034 | reverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const { |
2035 | assert(Cond.size() == 2 && "Invalid PPC branch opcode!" ); |
2036 | if (Cond[1].getReg() == PPC::CTR8 || Cond[1].getReg() == PPC::CTR) |
2037 | Cond[0].setImm(Cond[0].getImm() == 0 ? 1 : 0); |
2038 | else |
2039 | // Leave the CR# the same, but invert the condition. |
2040 | Cond[0].setImm(PPC::InvertPredicate(Opcode: (PPC::Predicate)Cond[0].getImm())); |
2041 | return false; |
2042 | } |
2043 | |
2044 | // For some instructions, it is legal to fold ZERO into the RA register field. |
2045 | // This function performs that fold by replacing the operand with PPC::ZERO, |
2046 | // it does not consider whether the load immediate zero is no longer in use. |
2047 | bool PPCInstrInfo::onlyFoldImmediate(MachineInstr &UseMI, MachineInstr &DefMI, |
2048 | Register Reg) const { |
2049 | // A zero immediate should always be loaded with a single li. |
2050 | unsigned DefOpc = DefMI.getOpcode(); |
2051 | if (DefOpc != PPC::LI && DefOpc != PPC::LI8) |
2052 | return false; |
2053 | if (!DefMI.getOperand(i: 1).isImm()) |
2054 | return false; |
2055 | if (DefMI.getOperand(i: 1).getImm() != 0) |
2056 | return false; |
2057 | |
2058 | // Note that we cannot here invert the arguments of an isel in order to fold |
2059 | // a ZERO into what is presented as the second argument. All we have here |
2060 | // is the condition bit, and that might come from a CR-logical bit operation. |
2061 | |
2062 | const MCInstrDesc &UseMCID = UseMI.getDesc(); |
2063 | |
2064 | // Only fold into real machine instructions. |
2065 | if (UseMCID.isPseudo()) |
2066 | return false; |
2067 | |
2068 | // We need to find which of the User's operands is to be folded, that will be |
2069 | // the operand that matches the given register ID. |
2070 | unsigned UseIdx; |
2071 | for (UseIdx = 0; UseIdx < UseMI.getNumOperands(); ++UseIdx) |
2072 | if (UseMI.getOperand(i: UseIdx).isReg() && |
2073 | UseMI.getOperand(i: UseIdx).getReg() == Reg) |
2074 | break; |
2075 | |
2076 | assert(UseIdx < UseMI.getNumOperands() && "Cannot find Reg in UseMI" ); |
2077 | assert(UseIdx < UseMCID.getNumOperands() && "No operand description for Reg" ); |
2078 | |
2079 | const MCOperandInfo *UseInfo = &UseMCID.operands()[UseIdx]; |
2080 | |
2081 | // We can fold the zero if this register requires a GPRC_NOR0/G8RC_NOX0 |
2082 | // register (which might also be specified as a pointer class kind). |
2083 | if (UseInfo->isLookupPtrRegClass()) { |
2084 | if (UseInfo->RegClass /* Kind */ != 1) |
2085 | return false; |
2086 | } else { |
2087 | if (UseInfo->RegClass != PPC::GPRC_NOR0RegClassID && |
2088 | UseInfo->RegClass != PPC::G8RC_NOX0RegClassID) |
2089 | return false; |
2090 | } |
2091 | |
2092 | // Make sure this is not tied to an output register (or otherwise |
2093 | // constrained). This is true for ST?UX registers, for example, which |
2094 | // are tied to their output registers. |
2095 | if (UseInfo->Constraints != 0) |
2096 | return false; |
2097 | |
2098 | MCRegister ZeroReg; |
2099 | if (UseInfo->isLookupPtrRegClass()) { |
2100 | bool isPPC64 = Subtarget.isPPC64(); |
2101 | ZeroReg = isPPC64 ? PPC::ZERO8 : PPC::ZERO; |
2102 | } else { |
2103 | ZeroReg = UseInfo->RegClass == PPC::G8RC_NOX0RegClassID ? |
2104 | PPC::ZERO8 : PPC::ZERO; |
2105 | } |
2106 | |
2107 | LLVM_DEBUG(dbgs() << "Folded immediate zero for: " ); |
2108 | LLVM_DEBUG(UseMI.dump()); |
2109 | UseMI.getOperand(i: UseIdx).setReg(ZeroReg); |
2110 | LLVM_DEBUG(dbgs() << "Into: " ); |
2111 | LLVM_DEBUG(UseMI.dump()); |
2112 | return true; |
2113 | } |
2114 | |
2115 | // Folds zero into instructions which have a load immediate zero as an operand |
2116 | // but also recognize zero as immediate zero. If the definition of the load |
2117 | // has no more users it is deleted. |
2118 | bool PPCInstrInfo::foldImmediate(MachineInstr &UseMI, MachineInstr &DefMI, |
2119 | Register Reg, MachineRegisterInfo *MRI) const { |
2120 | bool Changed = onlyFoldImmediate(UseMI, DefMI, Reg); |
2121 | if (MRI->use_nodbg_empty(RegNo: Reg)) |
2122 | DefMI.eraseFromParent(); |
2123 | return Changed; |
2124 | } |
2125 | |
2126 | static bool MBBDefinesCTR(MachineBasicBlock &MBB) { |
2127 | for (MachineInstr &MI : MBB) |
2128 | if (MI.definesRegister(Reg: PPC::CTR, /*TRI=*/nullptr) || |
2129 | MI.definesRegister(Reg: PPC::CTR8, /*TRI=*/nullptr)) |
2130 | return true; |
2131 | return false; |
2132 | } |
2133 | |
2134 | // We should make sure that, if we're going to predicate both sides of a |
2135 | // condition (a diamond), that both sides don't define the counter register. We |
2136 | // can predicate counter-decrement-based branches, but while that predicates |
2137 | // the branching, it does not predicate the counter decrement. If we tried to |
2138 | // merge the triangle into one predicated block, we'd decrement the counter |
2139 | // twice. |
2140 | bool PPCInstrInfo::isProfitableToIfCvt(MachineBasicBlock &TMBB, |
2141 | unsigned NumT, unsigned , |
2142 | MachineBasicBlock &FMBB, |
2143 | unsigned NumF, unsigned , |
2144 | BranchProbability Probability) const { |
2145 | return !(MBBDefinesCTR(MBB&: TMBB) && MBBDefinesCTR(MBB&: FMBB)); |
2146 | } |
2147 | |
2148 | |
2149 | bool PPCInstrInfo::isPredicated(const MachineInstr &MI) const { |
2150 | // The predicated branches are identified by their type, not really by the |
2151 | // explicit presence of a predicate. Furthermore, some of them can be |
2152 | // predicated more than once. Because if conversion won't try to predicate |
2153 | // any instruction which already claims to be predicated (by returning true |
2154 | // here), always return false. In doing so, we let isPredicable() be the |
2155 | // final word on whether not the instruction can be (further) predicated. |
2156 | |
2157 | return false; |
2158 | } |
2159 | |
2160 | bool PPCInstrInfo::isSchedulingBoundary(const MachineInstr &MI, |
2161 | const MachineBasicBlock *MBB, |
2162 | const MachineFunction &MF) const { |
2163 | switch (MI.getOpcode()) { |
2164 | default: |
2165 | break; |
2166 | // Set MFFS and MTFSF as scheduling boundary to avoid unexpected code motion |
2167 | // across them, since some FP operations may change content of FPSCR. |
2168 | // TODO: Model FPSCR in PPC instruction definitions and remove the workaround |
2169 | case PPC::MFFS: |
2170 | case PPC::MTFSF: |
2171 | case PPC::FENCE: |
2172 | return true; |
2173 | } |
2174 | return TargetInstrInfo::isSchedulingBoundary(MI, MBB, MF); |
2175 | } |
2176 | |
2177 | bool PPCInstrInfo::PredicateInstruction(MachineInstr &MI, |
2178 | ArrayRef<MachineOperand> Pred) const { |
2179 | unsigned OpC = MI.getOpcode(); |
2180 | if (OpC == PPC::BLR || OpC == PPC::BLR8) { |
2181 | if (Pred[1].getReg() == PPC::CTR8 || Pred[1].getReg() == PPC::CTR) { |
2182 | bool isPPC64 = Subtarget.isPPC64(); |
2183 | MI.setDesc(get(Opcode: Pred[0].getImm() ? (isPPC64 ? PPC::BDNZLR8 : PPC::BDNZLR) |
2184 | : (isPPC64 ? PPC::BDZLR8 : PPC::BDZLR))); |
2185 | // Need add Def and Use for CTR implicit operand. |
2186 | MachineInstrBuilder(*MI.getParent()->getParent(), MI) |
2187 | .addReg(RegNo: Pred[1].getReg(), flags: RegState::Implicit) |
2188 | .addReg(RegNo: Pred[1].getReg(), flags: RegState::ImplicitDefine); |
2189 | } else if (Pred[0].getImm() == PPC::PRED_BIT_SET) { |
2190 | MI.setDesc(get(Opcode: PPC::BCLR)); |
2191 | MachineInstrBuilder(*MI.getParent()->getParent(), MI).add(MO: Pred[1]); |
2192 | } else if (Pred[0].getImm() == PPC::PRED_BIT_UNSET) { |
2193 | MI.setDesc(get(Opcode: PPC::BCLRn)); |
2194 | MachineInstrBuilder(*MI.getParent()->getParent(), MI).add(MO: Pred[1]); |
2195 | } else { |
2196 | MI.setDesc(get(Opcode: PPC::BCCLR)); |
2197 | MachineInstrBuilder(*MI.getParent()->getParent(), MI) |
2198 | .addImm(Val: Pred[0].getImm()) |
2199 | .add(MO: Pred[1]); |
2200 | } |
2201 | |
2202 | return true; |
2203 | } else if (OpC == PPC::B) { |
2204 | if (Pred[1].getReg() == PPC::CTR8 || Pred[1].getReg() == PPC::CTR) { |
2205 | bool isPPC64 = Subtarget.isPPC64(); |
2206 | MI.setDesc(get(Opcode: Pred[0].getImm() ? (isPPC64 ? PPC::BDNZ8 : PPC::BDNZ) |
2207 | : (isPPC64 ? PPC::BDZ8 : PPC::BDZ))); |
2208 | // Need add Def and Use for CTR implicit operand. |
2209 | MachineInstrBuilder(*MI.getParent()->getParent(), MI) |
2210 | .addReg(RegNo: Pred[1].getReg(), flags: RegState::Implicit) |
2211 | .addReg(RegNo: Pred[1].getReg(), flags: RegState::ImplicitDefine); |
2212 | } else if (Pred[0].getImm() == PPC::PRED_BIT_SET) { |
2213 | MachineBasicBlock *MBB = MI.getOperand(i: 0).getMBB(); |
2214 | MI.removeOperand(OpNo: 0); |
2215 | |
2216 | MI.setDesc(get(Opcode: PPC::BC)); |
2217 | MachineInstrBuilder(*MI.getParent()->getParent(), MI) |
2218 | .add(MO: Pred[1]) |
2219 | .addMBB(MBB); |
2220 | } else if (Pred[0].getImm() == PPC::PRED_BIT_UNSET) { |
2221 | MachineBasicBlock *MBB = MI.getOperand(i: 0).getMBB(); |
2222 | MI.removeOperand(OpNo: 0); |
2223 | |
2224 | MI.setDesc(get(Opcode: PPC::BCn)); |
2225 | MachineInstrBuilder(*MI.getParent()->getParent(), MI) |
2226 | .add(MO: Pred[1]) |
2227 | .addMBB(MBB); |
2228 | } else { |
2229 | MachineBasicBlock *MBB = MI.getOperand(i: 0).getMBB(); |
2230 | MI.removeOperand(OpNo: 0); |
2231 | |
2232 | MI.setDesc(get(Opcode: PPC::BCC)); |
2233 | MachineInstrBuilder(*MI.getParent()->getParent(), MI) |
2234 | .addImm(Val: Pred[0].getImm()) |
2235 | .add(MO: Pred[1]) |
2236 | .addMBB(MBB); |
2237 | } |
2238 | |
2239 | return true; |
2240 | } else if (OpC == PPC::BCTR || OpC == PPC::BCTR8 || OpC == PPC::BCTRL || |
2241 | OpC == PPC::BCTRL8 || OpC == PPC::BCTRL_RM || |
2242 | OpC == PPC::BCTRL8_RM) { |
2243 | if (Pred[1].getReg() == PPC::CTR8 || Pred[1].getReg() == PPC::CTR) |
2244 | llvm_unreachable("Cannot predicate bctr[l] on the ctr register" ); |
2245 | |
2246 | bool setLR = OpC == PPC::BCTRL || OpC == PPC::BCTRL8 || |
2247 | OpC == PPC::BCTRL_RM || OpC == PPC::BCTRL8_RM; |
2248 | bool isPPC64 = Subtarget.isPPC64(); |
2249 | |
2250 | if (Pred[0].getImm() == PPC::PRED_BIT_SET) { |
2251 | MI.setDesc(get(Opcode: isPPC64 ? (setLR ? PPC::BCCTRL8 : PPC::BCCTR8) |
2252 | : (setLR ? PPC::BCCTRL : PPC::BCCTR))); |
2253 | MachineInstrBuilder(*MI.getParent()->getParent(), MI).add(MO: Pred[1]); |
2254 | } else if (Pred[0].getImm() == PPC::PRED_BIT_UNSET) { |
2255 | MI.setDesc(get(Opcode: isPPC64 ? (setLR ? PPC::BCCTRL8n : PPC::BCCTR8n) |
2256 | : (setLR ? PPC::BCCTRLn : PPC::BCCTRn))); |
2257 | MachineInstrBuilder(*MI.getParent()->getParent(), MI).add(MO: Pred[1]); |
2258 | } else { |
2259 | MI.setDesc(get(Opcode: isPPC64 ? (setLR ? PPC::BCCCTRL8 : PPC::BCCCTR8) |
2260 | : (setLR ? PPC::BCCCTRL : PPC::BCCCTR))); |
2261 | MachineInstrBuilder(*MI.getParent()->getParent(), MI) |
2262 | .addImm(Val: Pred[0].getImm()) |
2263 | .add(MO: Pred[1]); |
2264 | } |
2265 | |
2266 | // Need add Def and Use for LR implicit operand. |
2267 | if (setLR) |
2268 | MachineInstrBuilder(*MI.getParent()->getParent(), MI) |
2269 | .addReg(RegNo: isPPC64 ? PPC::LR8 : PPC::LR, flags: RegState::Implicit) |
2270 | .addReg(RegNo: isPPC64 ? PPC::LR8 : PPC::LR, flags: RegState::ImplicitDefine); |
2271 | if (OpC == PPC::BCTRL_RM || OpC == PPC::BCTRL8_RM) |
2272 | MachineInstrBuilder(*MI.getParent()->getParent(), MI) |
2273 | .addReg(RegNo: PPC::RM, flags: RegState::ImplicitDefine); |
2274 | |
2275 | return true; |
2276 | } |
2277 | |
2278 | return false; |
2279 | } |
2280 | |
2281 | bool PPCInstrInfo::SubsumesPredicate(ArrayRef<MachineOperand> Pred1, |
2282 | ArrayRef<MachineOperand> Pred2) const { |
2283 | assert(Pred1.size() == 2 && "Invalid PPC first predicate" ); |
2284 | assert(Pred2.size() == 2 && "Invalid PPC second predicate" ); |
2285 | |
2286 | if (Pred1[1].getReg() == PPC::CTR8 || Pred1[1].getReg() == PPC::CTR) |
2287 | return false; |
2288 | if (Pred2[1].getReg() == PPC::CTR8 || Pred2[1].getReg() == PPC::CTR) |
2289 | return false; |
2290 | |
2291 | // P1 can only subsume P2 if they test the same condition register. |
2292 | if (Pred1[1].getReg() != Pred2[1].getReg()) |
2293 | return false; |
2294 | |
2295 | PPC::Predicate P1 = (PPC::Predicate) Pred1[0].getImm(); |
2296 | PPC::Predicate P2 = (PPC::Predicate) Pred2[0].getImm(); |
2297 | |
2298 | if (P1 == P2) |
2299 | return true; |
2300 | |
2301 | // Does P1 subsume P2, e.g. GE subsumes GT. |
2302 | if (P1 == PPC::PRED_LE && |
2303 | (P2 == PPC::PRED_LT || P2 == PPC::PRED_EQ)) |
2304 | return true; |
2305 | if (P1 == PPC::PRED_GE && |
2306 | (P2 == PPC::PRED_GT || P2 == PPC::PRED_EQ)) |
2307 | return true; |
2308 | |
2309 | return false; |
2310 | } |
2311 | |
2312 | bool PPCInstrInfo::ClobbersPredicate(MachineInstr &MI, |
2313 | std::vector<MachineOperand> &Pred, |
2314 | bool SkipDead) const { |
2315 | // Note: At the present time, the contents of Pred from this function is |
2316 | // unused by IfConversion. This implementation follows ARM by pushing the |
2317 | // CR-defining operand. Because the 'DZ' and 'DNZ' count as types of |
2318 | // predicate, instructions defining CTR or CTR8 are also included as |
2319 | // predicate-defining instructions. |
2320 | |
2321 | const TargetRegisterClass *RCs[] = |
2322 | { &PPC::CRRCRegClass, &PPC::CRBITRCRegClass, |
2323 | &PPC::CTRRCRegClass, &PPC::CTRRC8RegClass }; |
2324 | |
2325 | bool Found = false; |
2326 | for (const MachineOperand &MO : MI.operands()) { |
2327 | for (unsigned c = 0; c < std::size(RCs) && !Found; ++c) { |
2328 | const TargetRegisterClass *RC = RCs[c]; |
2329 | if (MO.isReg()) { |
2330 | if (MO.isDef() && RC->contains(Reg: MO.getReg())) { |
2331 | Pred.push_back(x: MO); |
2332 | Found = true; |
2333 | } |
2334 | } else if (MO.isRegMask()) { |
2335 | for (MCPhysReg R : *RC) |
2336 | if (MO.clobbersPhysReg(PhysReg: R)) { |
2337 | Pred.push_back(x: MO); |
2338 | Found = true; |
2339 | } |
2340 | } |
2341 | } |
2342 | } |
2343 | |
2344 | return Found; |
2345 | } |
2346 | |
2347 | bool PPCInstrInfo::analyzeCompare(const MachineInstr &MI, Register &SrcReg, |
2348 | Register &SrcReg2, int64_t &Mask, |
2349 | int64_t &Value) const { |
2350 | unsigned Opc = MI.getOpcode(); |
2351 | |
2352 | switch (Opc) { |
2353 | default: return false; |
2354 | case PPC::CMPWI: |
2355 | case PPC::CMPLWI: |
2356 | case PPC::CMPDI: |
2357 | case PPC::CMPLDI: |
2358 | SrcReg = MI.getOperand(i: 1).getReg(); |
2359 | SrcReg2 = 0; |
2360 | Value = MI.getOperand(i: 2).getImm(); |
2361 | Mask = 0xFFFF; |
2362 | return true; |
2363 | case PPC::CMPW: |
2364 | case PPC::CMPLW: |
2365 | case PPC::CMPD: |
2366 | case PPC::CMPLD: |
2367 | case PPC::FCMPUS: |
2368 | case PPC::FCMPUD: |
2369 | SrcReg = MI.getOperand(i: 1).getReg(); |
2370 | SrcReg2 = MI.getOperand(i: 2).getReg(); |
2371 | Value = 0; |
2372 | Mask = 0; |
2373 | return true; |
2374 | } |
2375 | } |
2376 | |
2377 | bool PPCInstrInfo::optimizeCompareInstr(MachineInstr &CmpInstr, Register SrcReg, |
2378 | Register SrcReg2, int64_t Mask, |
2379 | int64_t Value, |
2380 | const MachineRegisterInfo *MRI) const { |
2381 | if (DisableCmpOpt) |
2382 | return false; |
2383 | |
2384 | int OpC = CmpInstr.getOpcode(); |
2385 | Register CRReg = CmpInstr.getOperand(i: 0).getReg(); |
2386 | |
2387 | // FP record forms set CR1 based on the exception status bits, not a |
2388 | // comparison with zero. |
2389 | if (OpC == PPC::FCMPUS || OpC == PPC::FCMPUD) |
2390 | return false; |
2391 | |
2392 | const TargetRegisterInfo *TRI = &getRegisterInfo(); |
2393 | // The record forms set the condition register based on a signed comparison |
2394 | // with zero (so says the ISA manual). This is not as straightforward as it |
2395 | // seems, however, because this is always a 64-bit comparison on PPC64, even |
2396 | // for instructions that are 32-bit in nature (like slw for example). |
2397 | // So, on PPC32, for unsigned comparisons, we can use the record forms only |
2398 | // for equality checks (as those don't depend on the sign). On PPC64, |
2399 | // we are restricted to equality for unsigned 64-bit comparisons and for |
2400 | // signed 32-bit comparisons the applicability is more restricted. |
2401 | bool isPPC64 = Subtarget.isPPC64(); |
2402 | bool is32BitSignedCompare = OpC == PPC::CMPWI || OpC == PPC::CMPW; |
2403 | bool is32BitUnsignedCompare = OpC == PPC::CMPLWI || OpC == PPC::CMPLW; |
2404 | bool is64BitUnsignedCompare = OpC == PPC::CMPLDI || OpC == PPC::CMPLD; |
2405 | |
2406 | // Look through copies unless that gets us to a physical register. |
2407 | Register ActualSrc = TRI->lookThruCopyLike(SrcReg, MRI); |
2408 | if (ActualSrc.isVirtual()) |
2409 | SrcReg = ActualSrc; |
2410 | |
2411 | // Get the unique definition of SrcReg. |
2412 | MachineInstr *MI = MRI->getUniqueVRegDef(Reg: SrcReg); |
2413 | if (!MI) return false; |
2414 | |
2415 | bool equalityOnly = false; |
2416 | bool noSub = false; |
2417 | if (isPPC64) { |
2418 | if (is32BitSignedCompare) { |
2419 | // We can perform this optimization only if SrcReg is sign-extending. |
2420 | if (isSignExtended(Reg: SrcReg, MRI)) |
2421 | noSub = true; |
2422 | else |
2423 | return false; |
2424 | } else if (is32BitUnsignedCompare) { |
2425 | // We can perform this optimization, equality only, if SrcReg is |
2426 | // zero-extending. |
2427 | if (isZeroExtended(Reg: SrcReg, MRI)) { |
2428 | noSub = true; |
2429 | equalityOnly = true; |
2430 | } else |
2431 | return false; |
2432 | } else |
2433 | equalityOnly = is64BitUnsignedCompare; |
2434 | } else |
2435 | equalityOnly = is32BitUnsignedCompare; |
2436 | |
2437 | if (equalityOnly) { |
2438 | // We need to check the uses of the condition register in order to reject |
2439 | // non-equality comparisons. |
2440 | for (MachineRegisterInfo::use_instr_iterator |
2441 | I = MRI->use_instr_begin(RegNo: CRReg), IE = MRI->use_instr_end(); |
2442 | I != IE; ++I) { |
2443 | MachineInstr *UseMI = &*I; |
2444 | if (UseMI->getOpcode() == PPC::BCC) { |
2445 | PPC::Predicate Pred = (PPC::Predicate)UseMI->getOperand(i: 0).getImm(); |
2446 | unsigned PredCond = PPC::getPredicateCondition(Opcode: Pred); |
2447 | // We ignore hint bits when checking for non-equality comparisons. |
2448 | if (PredCond != PPC::PRED_EQ && PredCond != PPC::PRED_NE) |
2449 | return false; |
2450 | } else if (UseMI->getOpcode() == PPC::ISEL || |
2451 | UseMI->getOpcode() == PPC::ISEL8) { |
2452 | unsigned SubIdx = UseMI->getOperand(i: 3).getSubReg(); |
2453 | if (SubIdx != PPC::sub_eq) |
2454 | return false; |
2455 | } else |
2456 | return false; |
2457 | } |
2458 | } |
2459 | |
2460 | MachineBasicBlock::iterator I = CmpInstr; |
2461 | |
2462 | // Scan forward to find the first use of the compare. |
2463 | for (MachineBasicBlock::iterator EL = CmpInstr.getParent()->end(); I != EL; |
2464 | ++I) { |
2465 | bool FoundUse = false; |
2466 | for (MachineRegisterInfo::use_instr_iterator |
2467 | J = MRI->use_instr_begin(RegNo: CRReg), JE = MRI->use_instr_end(); |
2468 | J != JE; ++J) |
2469 | if (&*J == &*I) { |
2470 | FoundUse = true; |
2471 | break; |
2472 | } |
2473 | |
2474 | if (FoundUse) |
2475 | break; |
2476 | } |
2477 | |
2478 | SmallVector<std::pair<MachineOperand*, PPC::Predicate>, 4> PredsToUpdate; |
2479 | SmallVector<std::pair<MachineOperand*, unsigned>, 4> SubRegsToUpdate; |
2480 | |
2481 | // There are two possible candidates which can be changed to set CR[01]. |
2482 | // One is MI, the other is a SUB instruction. |
2483 | // For CMPrr(r1,r2), we are looking for SUB(r1,r2) or SUB(r2,r1). |
2484 | MachineInstr *Sub = nullptr; |
2485 | if (SrcReg2 != 0) |
2486 | // MI is not a candidate for CMPrr. |
2487 | MI = nullptr; |
2488 | // FIXME: Conservatively refuse to convert an instruction which isn't in the |
2489 | // same BB as the comparison. This is to allow the check below to avoid calls |
2490 | // (and other explicit clobbers); instead we should really check for these |
2491 | // more explicitly (in at least a few predecessors). |
2492 | else if (MI->getParent() != CmpInstr.getParent()) |
2493 | return false; |
2494 | else if (Value != 0) { |
2495 | // The record-form instructions set CR bit based on signed comparison |
2496 | // against 0. We try to convert a compare against 1 or -1 into a compare |
2497 | // against 0 to exploit record-form instructions. For example, we change |
2498 | // the condition "greater than -1" into "greater than or equal to 0" |
2499 | // and "less than 1" into "less than or equal to 0". |
2500 | |
2501 | // Since we optimize comparison based on a specific branch condition, |
2502 | // we don't optimize if condition code is used by more than once. |
2503 | if (equalityOnly || !MRI->hasOneUse(RegNo: CRReg)) |
2504 | return false; |
2505 | |
2506 | MachineInstr *UseMI = &*MRI->use_instr_begin(RegNo: CRReg); |
2507 | if (UseMI->getOpcode() != PPC::BCC) |
2508 | return false; |
2509 | |
2510 | PPC::Predicate Pred = (PPC::Predicate)UseMI->getOperand(i: 0).getImm(); |
2511 | unsigned PredCond = PPC::getPredicateCondition(Opcode: Pred); |
2512 | unsigned PredHint = PPC::getPredicateHint(Opcode: Pred); |
2513 | int16_t Immed = (int16_t)Value; |
2514 | |
2515 | // When modifying the condition in the predicate, we propagate hint bits |
2516 | // from the original predicate to the new one. |
2517 | if (Immed == -1 && PredCond == PPC::PRED_GT) |
2518 | // We convert "greater than -1" into "greater than or equal to 0", |
2519 | // since we are assuming signed comparison by !equalityOnly |
2520 | Pred = PPC::getPredicate(Condition: PPC::PRED_GE, Hint: PredHint); |
2521 | else if (Immed == -1 && PredCond == PPC::PRED_LE) |
2522 | // We convert "less than or equal to -1" into "less than 0". |
2523 | Pred = PPC::getPredicate(Condition: PPC::PRED_LT, Hint: PredHint); |
2524 | else if (Immed == 1 && PredCond == PPC::PRED_LT) |
2525 | // We convert "less than 1" into "less than or equal to 0". |
2526 | Pred = PPC::getPredicate(Condition: PPC::PRED_LE, Hint: PredHint); |
2527 | else if (Immed == 1 && PredCond == PPC::PRED_GE) |
2528 | // We convert "greater than or equal to 1" into "greater than 0". |
2529 | Pred = PPC::getPredicate(Condition: PPC::PRED_GT, Hint: PredHint); |
2530 | else |
2531 | return false; |
2532 | |
2533 | // Convert the comparison and its user to a compare against zero with the |
2534 | // appropriate predicate on the branch. Zero comparison might provide |
2535 | // optimization opportunities post-RA (see optimization in |
2536 | // PPCPreEmitPeephole.cpp). |
2537 | UseMI->getOperand(i: 0).setImm(Pred); |
2538 | CmpInstr.getOperand(i: 2).setImm(0); |
2539 | } |
2540 | |
2541 | // Search for Sub. |
2542 | --I; |
2543 | |
2544 | // Get ready to iterate backward from CmpInstr. |
2545 | MachineBasicBlock::iterator E = MI, B = CmpInstr.getParent()->begin(); |
2546 | |
2547 | for (; I != E && !noSub; --I) { |
2548 | const MachineInstr &Instr = *I; |
2549 | unsigned IOpC = Instr.getOpcode(); |
2550 | |
2551 | if (&*I != &CmpInstr && (Instr.modifiesRegister(Reg: PPC::CR0, TRI) || |
2552 | Instr.readsRegister(Reg: PPC::CR0, TRI))) |
2553 | // This instruction modifies or uses the record condition register after |
2554 | // the one we want to change. While we could do this transformation, it |
2555 | // would likely not be profitable. This transformation removes one |
2556 | // instruction, and so even forcing RA to generate one move probably |
2557 | // makes it unprofitable. |
2558 | return false; |
2559 | |
2560 | // Check whether CmpInstr can be made redundant by the current instruction. |
2561 | if ((OpC == PPC::CMPW || OpC == PPC::CMPLW || |
2562 | OpC == PPC::CMPD || OpC == PPC::CMPLD) && |
2563 | (IOpC == PPC::SUBF || IOpC == PPC::SUBF8) && |
2564 | ((Instr.getOperand(i: 1).getReg() == SrcReg && |
2565 | Instr.getOperand(i: 2).getReg() == SrcReg2) || |
2566 | (Instr.getOperand(i: 1).getReg() == SrcReg2 && |
2567 | Instr.getOperand(i: 2).getReg() == SrcReg))) { |
2568 | Sub = &*I; |
2569 | break; |
2570 | } |
2571 | |
2572 | if (I == B) |
2573 | // The 'and' is below the comparison instruction. |
2574 | return false; |
2575 | } |
2576 | |
2577 | // Return false if no candidates exist. |
2578 | if (!MI && !Sub) |
2579 | return false; |
2580 | |
2581 | // The single candidate is called MI. |
2582 | if (!MI) MI = Sub; |
2583 | |
2584 | int NewOpC = -1; |
2585 | int MIOpC = MI->getOpcode(); |
2586 | if (MIOpC == PPC::ANDI_rec || MIOpC == PPC::ANDI8_rec || |
2587 | MIOpC == PPC::ANDIS_rec || MIOpC == PPC::ANDIS8_rec) |
2588 | NewOpC = MIOpC; |
2589 | else { |
2590 | NewOpC = PPC::getRecordFormOpcode(Opcode: MIOpC); |
2591 | if (NewOpC == -1 && PPC::getNonRecordFormOpcode(Opcode: MIOpC) != -1) |
2592 | NewOpC = MIOpC; |
2593 | } |
2594 | |
2595 | // FIXME: On the non-embedded POWER architectures, only some of the record |
2596 | // forms are fast, and we should use only the fast ones. |
2597 | |
2598 | // The defining instruction has a record form (or is already a record |
2599 | // form). It is possible, however, that we'll need to reverse the condition |
2600 | // code of the users. |
2601 | if (NewOpC == -1) |
2602 | return false; |
2603 | |
2604 | // This transformation should not be performed if `nsw` is missing and is not |
2605 | // `equalityOnly` comparison. Since if there is overflow, sub_lt, sub_gt in |
2606 | // CRReg do not reflect correct order. If `equalityOnly` is true, sub_eq in |
2607 | // CRReg can reflect if compared values are equal, this optz is still valid. |
2608 | if (!equalityOnly && (NewOpC == PPC::SUBF_rec || NewOpC == PPC::SUBF8_rec) && |
2609 | Sub && !Sub->getFlag(Flag: MachineInstr::NoSWrap)) |
2610 | return false; |
2611 | |
2612 | // If we have SUB(r1, r2) and CMP(r2, r1), the condition code based on CMP |
2613 | // needs to be updated to be based on SUB. Push the condition code |
2614 | // operands to OperandsToUpdate. If it is safe to remove CmpInstr, the |
2615 | // condition code of these operands will be modified. |
2616 | // Here, Value == 0 means we haven't converted comparison against 1 or -1 to |
2617 | // comparison against 0, which may modify predicate. |
2618 | bool ShouldSwap = false; |
2619 | if (Sub && Value == 0) { |
2620 | ShouldSwap = SrcReg2 != 0 && Sub->getOperand(i: 1).getReg() == SrcReg2 && |
2621 | Sub->getOperand(i: 2).getReg() == SrcReg; |
2622 | |
2623 | // The operands to subf are the opposite of sub, so only in the fixed-point |
2624 | // case, invert the order. |
2625 | ShouldSwap = !ShouldSwap; |
2626 | } |
2627 | |
2628 | if (ShouldSwap) |
2629 | for (MachineRegisterInfo::use_instr_iterator |
2630 | I = MRI->use_instr_begin(RegNo: CRReg), IE = MRI->use_instr_end(); |
2631 | I != IE; ++I) { |
2632 | MachineInstr *UseMI = &*I; |
2633 | if (UseMI->getOpcode() == PPC::BCC) { |
2634 | PPC::Predicate Pred = (PPC::Predicate) UseMI->getOperand(i: 0).getImm(); |
2635 | unsigned PredCond = PPC::getPredicateCondition(Opcode: Pred); |
2636 | assert((!equalityOnly || |
2637 | PredCond == PPC::PRED_EQ || PredCond == PPC::PRED_NE) && |
2638 | "Invalid predicate for equality-only optimization" ); |
2639 | (void)PredCond; // To suppress warning in release build. |
2640 | PredsToUpdate.push_back(Elt: std::make_pair(x: &(UseMI->getOperand(i: 0)), |
2641 | y: PPC::getSwappedPredicate(Opcode: Pred))); |
2642 | } else if (UseMI->getOpcode() == PPC::ISEL || |
2643 | UseMI->getOpcode() == PPC::ISEL8) { |
2644 | unsigned NewSubReg = UseMI->getOperand(i: 3).getSubReg(); |
2645 | assert((!equalityOnly || NewSubReg == PPC::sub_eq) && |
2646 | "Invalid CR bit for equality-only optimization" ); |
2647 | |
2648 | if (NewSubReg == PPC::sub_lt) |
2649 | NewSubReg = PPC::sub_gt; |
2650 | else if (NewSubReg == PPC::sub_gt) |
2651 | NewSubReg = PPC::sub_lt; |
2652 | |
2653 | SubRegsToUpdate.push_back(Elt: std::make_pair(x: &(UseMI->getOperand(i: 3)), |
2654 | y&: NewSubReg)); |
2655 | } else // We need to abort on a user we don't understand. |
2656 | return false; |
2657 | } |
2658 | assert(!(Value != 0 && ShouldSwap) && |
2659 | "Non-zero immediate support and ShouldSwap" |
2660 | "may conflict in updating predicate" ); |
2661 | |
2662 | // Create a new virtual register to hold the value of the CR set by the |
2663 | // record-form instruction. If the instruction was not previously in |
2664 | // record form, then set the kill flag on the CR. |
2665 | CmpInstr.eraseFromParent(); |
2666 | |
2667 | MachineBasicBlock::iterator MII = MI; |
2668 | BuildMI(BB&: *MI->getParent(), I: std::next(x: MII), MIMD: MI->getDebugLoc(), |
2669 | MCID: get(Opcode: TargetOpcode::COPY), DestReg: CRReg) |
2670 | .addReg(RegNo: PPC::CR0, flags: MIOpC != NewOpC ? RegState::Kill : 0); |
2671 | |
2672 | // Even if CR0 register were dead before, it is alive now since the |
2673 | // instruction we just built uses it. |
2674 | MI->clearRegisterDeads(Reg: PPC::CR0); |
2675 | |
2676 | if (MIOpC != NewOpC) { |
2677 | // We need to be careful here: we're replacing one instruction with |
2678 | // another, and we need to make sure that we get all of the right |
2679 | // implicit uses and defs. On the other hand, the caller may be holding |
2680 | // an iterator to this instruction, and so we can't delete it (this is |
2681 | // specifically the case if this is the instruction directly after the |
2682 | // compare). |
2683 | |
2684 | // Rotates are expensive instructions. If we're emitting a record-form |
2685 | // rotate that can just be an andi/andis, we should just emit that. |
2686 | if (MIOpC == PPC::RLWINM || MIOpC == PPC::RLWINM8) { |
2687 | Register GPRRes = MI->getOperand(i: 0).getReg(); |
2688 | int64_t SH = MI->getOperand(i: 2).getImm(); |
2689 | int64_t MB = MI->getOperand(i: 3).getImm(); |
2690 | int64_t ME = MI->getOperand(i: 4).getImm(); |
2691 | // We can only do this if both the start and end of the mask are in the |
2692 | // same halfword. |
2693 | bool MBInLoHWord = MB >= 16; |
2694 | bool MEInLoHWord = ME >= 16; |
2695 | uint64_t Mask = ~0LLU; |
2696 | |
2697 | if (MB <= ME && MBInLoHWord == MEInLoHWord && SH == 0) { |
2698 | Mask = ((1LLU << (32 - MB)) - 1) & ~((1LLU << (31 - ME)) - 1); |
2699 | // The mask value needs to shift right 16 if we're emitting andis. |
2700 | Mask >>= MBInLoHWord ? 0 : 16; |
2701 | NewOpC = MIOpC == PPC::RLWINM |
2702 | ? (MBInLoHWord ? PPC::ANDI_rec : PPC::ANDIS_rec) |
2703 | : (MBInLoHWord ? PPC::ANDI8_rec : PPC::ANDIS8_rec); |
2704 | } else if (MRI->use_empty(RegNo: GPRRes) && (ME == 31) && |
2705 | (ME - MB + 1 == SH) && (MB >= 16)) { |
2706 | // If we are rotating by the exact number of bits as are in the mask |
2707 | // and the mask is in the least significant bits of the register, |
2708 | // that's just an andis. (as long as the GPR result has no uses). |
2709 | Mask = ((1LLU << 32) - 1) & ~((1LLU << (32 - SH)) - 1); |
2710 | Mask >>= 16; |
2711 | NewOpC = MIOpC == PPC::RLWINM ? PPC::ANDIS_rec : PPC::ANDIS8_rec; |
2712 | } |
2713 | // If we've set the mask, we can transform. |
2714 | if (Mask != ~0LLU) { |
2715 | MI->removeOperand(OpNo: 4); |
2716 | MI->removeOperand(OpNo: 3); |
2717 | MI->getOperand(i: 2).setImm(Mask); |
2718 | NumRcRotatesConvertedToRcAnd++; |
2719 | } |
2720 | } else if (MIOpC == PPC::RLDICL && MI->getOperand(i: 2).getImm() == 0) { |
2721 | int64_t MB = MI->getOperand(i: 3).getImm(); |
2722 | if (MB >= 48) { |
2723 | uint64_t Mask = (1LLU << (63 - MB + 1)) - 1; |
2724 | NewOpC = PPC::ANDI8_rec; |
2725 | MI->removeOperand(OpNo: 3); |
2726 | MI->getOperand(i: 2).setImm(Mask); |
2727 | NumRcRotatesConvertedToRcAnd++; |
2728 | } |
2729 | } |
2730 | |
2731 | const MCInstrDesc &NewDesc = get(Opcode: NewOpC); |
2732 | MI->setDesc(NewDesc); |
2733 | |
2734 | for (MCPhysReg ImpDef : NewDesc.implicit_defs()) { |
2735 | if (!MI->definesRegister(Reg: ImpDef, /*TRI=*/nullptr)) { |
2736 | MI->addOperand(MF&: *MI->getParent()->getParent(), |
2737 | Op: MachineOperand::CreateReg(Reg: ImpDef, isDef: true, isImp: true)); |
2738 | } |
2739 | } |
2740 | for (MCPhysReg ImpUse : NewDesc.implicit_uses()) { |
2741 | if (!MI->readsRegister(Reg: ImpUse, /*TRI=*/nullptr)) { |
2742 | MI->addOperand(MF&: *MI->getParent()->getParent(), |
2743 | Op: MachineOperand::CreateReg(Reg: ImpUse, isDef: false, isImp: true)); |
2744 | } |
2745 | } |
2746 | } |
2747 | assert(MI->definesRegister(PPC::CR0, /*TRI=*/nullptr) && |
2748 | "Record-form instruction does not define cr0?" ); |
2749 | |
2750 | // Modify the condition code of operands in OperandsToUpdate. |
2751 | // Since we have SUB(r1, r2) and CMP(r2, r1), the condition code needs to |
2752 | // be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc. |
2753 | for (unsigned i = 0, e = PredsToUpdate.size(); i < e; i++) |
2754 | PredsToUpdate[i].first->setImm(PredsToUpdate[i].second); |
2755 | |
2756 | for (unsigned i = 0, e = SubRegsToUpdate.size(); i < e; i++) |
2757 | SubRegsToUpdate[i].first->setSubReg(SubRegsToUpdate[i].second); |
2758 | |
2759 | return true; |
2760 | } |
2761 | |
2762 | bool PPCInstrInfo::optimizeCmpPostRA(MachineInstr &CmpMI) const { |
2763 | MachineRegisterInfo *MRI = &CmpMI.getParent()->getParent()->getRegInfo(); |
2764 | if (MRI->isSSA()) |
2765 | return false; |
2766 | |
2767 | Register SrcReg, SrcReg2; |
2768 | int64_t CmpMask, CmpValue; |
2769 | if (!analyzeCompare(MI: CmpMI, SrcReg, SrcReg2, Mask&: CmpMask, Value&: CmpValue)) |
2770 | return false; |
2771 | |
2772 | // Try to optimize the comparison against 0. |
2773 | if (CmpValue || !CmpMask || SrcReg2) |
2774 | return false; |
2775 | |
2776 | // The record forms set the condition register based on a signed comparison |
2777 | // with zero (see comments in optimizeCompareInstr). Since we can't do the |
2778 | // equality checks in post-RA, we are more restricted on a unsigned |
2779 | // comparison. |
2780 | unsigned Opc = CmpMI.getOpcode(); |
2781 | if (Opc == PPC::CMPLWI || Opc == PPC::CMPLDI) |
2782 | return false; |
2783 | |
2784 | // The record forms are always based on a 64-bit comparison on PPC64 |
2785 | // (similary, a 32-bit comparison on PPC32), while the CMPWI is a 32-bit |
2786 | // comparison. Since we can't do the equality checks in post-RA, we bail out |
2787 | // the case. |
2788 | if (Subtarget.isPPC64() && Opc == PPC::CMPWI) |
2789 | return false; |
2790 | |
2791 | // CmpMI can't be deleted if it has implicit def. |
2792 | if (CmpMI.hasImplicitDef()) |
2793 | return false; |
2794 | |
2795 | bool SrcRegHasOtherUse = false; |
2796 | MachineInstr *SrcMI = getDefMIPostRA(Reg: SrcReg, MI&: CmpMI, SeenIntermediateUse&: SrcRegHasOtherUse); |
2797 | if (!SrcMI || !SrcMI->definesRegister(Reg: SrcReg, /*TRI=*/nullptr)) |
2798 | return false; |
2799 | |
2800 | MachineOperand RegMO = CmpMI.getOperand(i: 0); |
2801 | Register CRReg = RegMO.getReg(); |
2802 | if (CRReg != PPC::CR0) |
2803 | return false; |
2804 | |
2805 | // Make sure there is no def/use of CRReg between SrcMI and CmpMI. |
2806 | bool SeenUseOfCRReg = false; |
2807 | bool IsCRRegKilled = false; |
2808 | if (!isRegElgibleForForwarding(RegMO, DefMI: *SrcMI, MI: CmpMI, KillDefMI: false, IsFwdFeederRegKilled&: IsCRRegKilled, |
2809 | SeenIntermediateUse&: SeenUseOfCRReg) || |
2810 | SrcMI->definesRegister(Reg: CRReg, /*TRI=*/nullptr) || SeenUseOfCRReg) |
2811 | return false; |
2812 | |
2813 | int SrcMIOpc = SrcMI->getOpcode(); |
2814 | int NewOpC = PPC::getRecordFormOpcode(Opcode: SrcMIOpc); |
2815 | if (NewOpC == -1) |
2816 | return false; |
2817 | |
2818 | LLVM_DEBUG(dbgs() << "Replace Instr: " ); |
2819 | LLVM_DEBUG(SrcMI->dump()); |
2820 | |
2821 | const MCInstrDesc &NewDesc = get(Opcode: NewOpC); |
2822 | SrcMI->setDesc(NewDesc); |
2823 | MachineInstrBuilder(*SrcMI->getParent()->getParent(), SrcMI) |
2824 | .addReg(RegNo: CRReg, flags: RegState::ImplicitDefine); |
2825 | SrcMI->clearRegisterDeads(Reg: CRReg); |
2826 | |
2827 | assert(SrcMI->definesRegister(PPC::CR0, /*TRI=*/nullptr) && |
2828 | "Record-form instruction does not define cr0?" ); |
2829 | |
2830 | LLVM_DEBUG(dbgs() << "with: " ); |
2831 | LLVM_DEBUG(SrcMI->dump()); |
2832 | LLVM_DEBUG(dbgs() << "Delete dead instruction: " ); |
2833 | LLVM_DEBUG(CmpMI.dump()); |
2834 | return true; |
2835 | } |
2836 | |
2837 | bool PPCInstrInfo::getMemOperandsWithOffsetWidth( |
2838 | const MachineInstr &LdSt, SmallVectorImpl<const MachineOperand *> &BaseOps, |
2839 | int64_t &Offset, bool &OffsetIsScalable, LocationSize &Width, |
2840 | const TargetRegisterInfo *TRI) const { |
2841 | const MachineOperand *BaseOp; |
2842 | OffsetIsScalable = false; |
2843 | if (!getMemOperandWithOffsetWidth(LdSt, BaseOp, Offset, Width, TRI)) |
2844 | return false; |
2845 | BaseOps.push_back(Elt: BaseOp); |
2846 | return true; |
2847 | } |
2848 | |
2849 | static bool isLdStSafeToCluster(const MachineInstr &LdSt, |
2850 | const TargetRegisterInfo *TRI) { |
2851 | // If this is a volatile load/store, don't mess with it. |
2852 | if (LdSt.hasOrderedMemoryRef() || LdSt.getNumExplicitOperands() != 3) |
2853 | return false; |
2854 | |
2855 | if (LdSt.getOperand(i: 2).isFI()) |
2856 | return true; |
2857 | |
2858 | assert(LdSt.getOperand(2).isReg() && "Expected a reg operand." ); |
2859 | // Can't cluster if the instruction modifies the base register |
2860 | // or it is update form. e.g. ld r2,3(r2) |
2861 | if (LdSt.modifiesRegister(Reg: LdSt.getOperand(i: 2).getReg(), TRI)) |
2862 | return false; |
2863 | |
2864 | return true; |
2865 | } |
2866 | |
2867 | // Only cluster instruction pair that have the same opcode, and they are |
2868 | // clusterable according to PowerPC specification. |
2869 | static bool isClusterableLdStOpcPair(unsigned FirstOpc, unsigned SecondOpc, |
2870 | const PPCSubtarget &Subtarget) { |
2871 | switch (FirstOpc) { |
2872 | default: |
2873 | return false; |
2874 | case PPC::STD: |
2875 | case PPC::STFD: |
2876 | case PPC::STXSD: |
2877 | case PPC::DFSTOREf64: |
2878 | return FirstOpc == SecondOpc; |
2879 | // PowerPC backend has opcode STW/STW8 for instruction "stw" to deal with |
2880 | // 32bit and 64bit instruction selection. They are clusterable pair though |
2881 | // they are different opcode. |
2882 | case PPC::STW: |
2883 | case PPC::STW8: |
2884 | return SecondOpc == PPC::STW || SecondOpc == PPC::STW8; |
2885 | } |
2886 | } |
2887 | |
2888 | bool PPCInstrInfo::shouldClusterMemOps( |
2889 | ArrayRef<const MachineOperand *> BaseOps1, int64_t OpOffset1, |
2890 | bool OffsetIsScalable1, ArrayRef<const MachineOperand *> BaseOps2, |
2891 | int64_t OpOffset2, bool OffsetIsScalable2, unsigned ClusterSize, |
2892 | unsigned NumBytes) const { |
2893 | |
2894 | assert(BaseOps1.size() == 1 && BaseOps2.size() == 1); |
2895 | const MachineOperand &BaseOp1 = *BaseOps1.front(); |
2896 | const MachineOperand &BaseOp2 = *BaseOps2.front(); |
2897 | assert((BaseOp1.isReg() || BaseOp1.isFI()) && |
2898 | "Only base registers and frame indices are supported." ); |
2899 | |
2900 | // ClusterSize means the number of memory operations that will have been |
2901 | // clustered if this hook returns true. |
2902 | // Don't cluster memory op if there are already two ops clustered at least. |
2903 | if (ClusterSize > 2) |
2904 | return false; |
2905 | |
2906 | // Cluster the load/store only when they have the same base |
2907 | // register or FI. |
2908 | if ((BaseOp1.isReg() != BaseOp2.isReg()) || |
2909 | (BaseOp1.isReg() && BaseOp1.getReg() != BaseOp2.getReg()) || |
2910 | (BaseOp1.isFI() && BaseOp1.getIndex() != BaseOp2.getIndex())) |
2911 | return false; |
2912 | |
2913 | // Check if the load/store are clusterable according to the PowerPC |
2914 | // specification. |
2915 | const MachineInstr &FirstLdSt = *BaseOp1.getParent(); |
2916 | const MachineInstr &SecondLdSt = *BaseOp2.getParent(); |
2917 | unsigned FirstOpc = FirstLdSt.getOpcode(); |
2918 | unsigned SecondOpc = SecondLdSt.getOpcode(); |
2919 | const TargetRegisterInfo *TRI = &getRegisterInfo(); |
2920 | // Cluster the load/store only when they have the same opcode, and they are |
2921 | // clusterable opcode according to PowerPC specification. |
2922 | if (!isClusterableLdStOpcPair(FirstOpc, SecondOpc, Subtarget)) |
2923 | return false; |
2924 | |
2925 | // Can't cluster load/store that have ordered or volatile memory reference. |
2926 | if (!isLdStSafeToCluster(LdSt: FirstLdSt, TRI) || |
2927 | !isLdStSafeToCluster(LdSt: SecondLdSt, TRI)) |
2928 | return false; |
2929 | |
2930 | int64_t Offset1 = 0, Offset2 = 0; |
2931 | LocationSize Width1 = 0, Width2 = 0; |
2932 | const MachineOperand *Base1 = nullptr, *Base2 = nullptr; |
2933 | if (!getMemOperandWithOffsetWidth(LdSt: FirstLdSt, BaseOp&: Base1, Offset&: Offset1, Width&: Width1, TRI) || |
2934 | !getMemOperandWithOffsetWidth(LdSt: SecondLdSt, BaseOp&: Base2, Offset&: Offset2, Width&: Width2, TRI) || |
2935 | Width1 != Width2) |
2936 | return false; |
2937 | |
2938 | assert(Base1 == &BaseOp1 && Base2 == &BaseOp2 && |
2939 | "getMemOperandWithOffsetWidth return incorrect base op" ); |
2940 | // The caller should already have ordered FirstMemOp/SecondMemOp by offset. |
2941 | assert(Offset1 <= Offset2 && "Caller should have ordered offsets." ); |
2942 | return Offset1 + (int64_t)Width1.getValue() == Offset2; |
2943 | } |
2944 | |
2945 | /// GetInstSize - Return the number of bytes of code the specified |
2946 | /// instruction may be. This returns the maximum number of bytes. |
2947 | /// |
2948 | unsigned PPCInstrInfo::getInstSizeInBytes(const MachineInstr &MI) const { |
2949 | unsigned Opcode = MI.getOpcode(); |
2950 | |
2951 | if (Opcode == PPC::INLINEASM || Opcode == PPC::INLINEASM_BR) { |
2952 | const MachineFunction *MF = MI.getParent()->getParent(); |
2953 | const char *AsmStr = MI.getOperand(i: 0).getSymbolName(); |
2954 | return getInlineAsmLength(Str: AsmStr, MAI: *MF->getTarget().getMCAsmInfo()); |
2955 | } else if (Opcode == TargetOpcode::STACKMAP) { |
2956 | StackMapOpers Opers(&MI); |
2957 | return Opers.getNumPatchBytes(); |
2958 | } else if (Opcode == TargetOpcode::PATCHPOINT) { |
2959 | PatchPointOpers Opers(&MI); |
2960 | return Opers.getNumPatchBytes(); |
2961 | } else { |
2962 | return get(Opcode).getSize(); |
2963 | } |
2964 | } |
2965 | |
2966 | std::pair<unsigned, unsigned> |
2967 | PPCInstrInfo::decomposeMachineOperandsTargetFlags(unsigned TF) const { |
2968 | // PPC always uses a direct mask. |
2969 | return std::make_pair(x&: TF, y: 0u); |
2970 | } |
2971 | |
2972 | ArrayRef<std::pair<unsigned, const char *>> |
2973 | PPCInstrInfo::getSerializableDirectMachineOperandTargetFlags() const { |
2974 | using namespace PPCII; |
2975 | static const std::pair<unsigned, const char *> TargetFlags[] = { |
2976 | {MO_PLT, "ppc-plt" }, |
2977 | {MO_PIC_FLAG, "ppc-pic" }, |
2978 | {MO_PCREL_FLAG, "ppc-pcrel" }, |
2979 | {MO_GOT_FLAG, "ppc-got" }, |
2980 | {MO_PCREL_OPT_FLAG, "ppc-opt-pcrel" }, |
2981 | {MO_TLSGD_FLAG, "ppc-tlsgd" }, |
2982 | {MO_TPREL_FLAG, "ppc-tprel" }, |
2983 | {MO_TLSLDM_FLAG, "ppc-tlsldm" }, |
2984 | {MO_TLSLD_FLAG, "ppc-tlsld" }, |
2985 | {MO_TLSGDM_FLAG, "ppc-tlsgdm" }, |
2986 | {MO_GOT_TLSGD_PCREL_FLAG, "ppc-got-tlsgd-pcrel" }, |
2987 | {MO_GOT_TLSLD_PCREL_FLAG, "ppc-got-tlsld-pcrel" }, |
2988 | {MO_GOT_TPREL_PCREL_FLAG, "ppc-got-tprel-pcrel" }, |
2989 | {MO_LO, "ppc-lo" }, |
2990 | {MO_HA, "ppc-ha" }, |
2991 | {MO_TPREL_LO, "ppc-tprel-lo" }, |
2992 | {MO_TPREL_HA, "ppc-tprel-ha" }, |
2993 | {MO_DTPREL_LO, "ppc-dtprel-lo" }, |
2994 | {MO_TLSLD_LO, "ppc-tlsld-lo" }, |
2995 | {MO_TOC_LO, "ppc-toc-lo" }, |
2996 | {MO_TLS, "ppc-tls" }, |
2997 | {MO_PIC_HA_FLAG, "ppc-ha-pic" }, |
2998 | {MO_PIC_LO_FLAG, "ppc-lo-pic" }, |
2999 | {MO_TPREL_PCREL_FLAG, "ppc-tprel-pcrel" }, |
3000 | {MO_TLS_PCREL_FLAG, "ppc-tls-pcrel" }, |
3001 | {MO_GOT_PCREL_FLAG, "ppc-got-pcrel" }, |
3002 | }; |
3003 | return ArrayRef(TargetFlags); |
3004 | } |
3005 | |
3006 | // Expand VSX Memory Pseudo instruction to either a VSX or a FP instruction. |
3007 | // The VSX versions have the advantage of a full 64-register target whereas |
3008 | // the FP ones have the advantage of lower latency and higher throughput. So |
3009 | // what we are after is using the faster instructions in low register pressure |
3010 | // situations and using the larger register file in high register pressure |
3011 | // situations. |
3012 | bool PPCInstrInfo::expandVSXMemPseudo(MachineInstr &MI) const { |
3013 | unsigned UpperOpcode, LowerOpcode; |
3014 | switch (MI.getOpcode()) { |
3015 | case PPC::DFLOADf32: |
3016 | UpperOpcode = PPC::LXSSP; |
3017 | LowerOpcode = PPC::LFS; |
3018 | break; |
3019 | case PPC::DFLOADf64: |
3020 | UpperOpcode = PPC::LXSD; |
3021 | LowerOpcode = PPC::LFD; |
3022 | break; |
3023 | case PPC::DFSTOREf32: |
3024 | UpperOpcode = PPC::STXSSP; |
3025 | LowerOpcode = PPC::STFS; |
3026 | break; |
3027 | case PPC::DFSTOREf64: |
3028 | UpperOpcode = PPC::STXSD; |
3029 | LowerOpcode = PPC::STFD; |
3030 | break; |
3031 | case PPC::XFLOADf32: |
3032 | UpperOpcode = PPC::LXSSPX; |
3033 | LowerOpcode = PPC::LFSX; |
3034 | break; |
3035 | case PPC::XFLOADf64: |
3036 | UpperOpcode = PPC::LXSDX; |
3037 | LowerOpcode = PPC::LFDX; |
3038 | break; |
3039 | case PPC::XFSTOREf32: |
3040 | UpperOpcode = PPC::STXSSPX; |
3041 | LowerOpcode = PPC::STFSX; |
3042 | break; |
3043 | case PPC::XFSTOREf64: |
3044 | UpperOpcode = PPC::STXSDX; |
3045 | LowerOpcode = PPC::STFDX; |
3046 | break; |
3047 | case PPC::LIWAX: |
3048 | UpperOpcode = PPC::LXSIWAX; |
3049 | LowerOpcode = PPC::LFIWAX; |
3050 | break; |
3051 | case PPC::LIWZX: |
3052 | UpperOpcode = PPC::LXSIWZX; |
3053 | LowerOpcode = PPC::LFIWZX; |
3054 | break; |
3055 | case PPC::STIWX: |
3056 | UpperOpcode = PPC::STXSIWX; |
3057 | LowerOpcode = PPC::STFIWX; |
3058 | break; |
3059 | default: |
3060 | llvm_unreachable("Unknown Operation!" ); |
3061 | } |
3062 | |
3063 | Register TargetReg = MI.getOperand(i: 0).getReg(); |
3064 | unsigned Opcode; |
3065 | if ((TargetReg >= PPC::F0 && TargetReg <= PPC::F31) || |
3066 | (TargetReg >= PPC::VSL0 && TargetReg <= PPC::VSL31)) |
3067 | Opcode = LowerOpcode; |
3068 | else |
3069 | Opcode = UpperOpcode; |
3070 | MI.setDesc(get(Opcode)); |
3071 | return true; |
3072 | } |
3073 | |
3074 | static bool isAnImmediateOperand(const MachineOperand &MO) { |
3075 | return MO.isCPI() || MO.isGlobal() || MO.isImm(); |
3076 | } |
3077 | |
3078 | bool PPCInstrInfo::expandPostRAPseudo(MachineInstr &MI) const { |
3079 | auto &MBB = *MI.getParent(); |
3080 | auto DL = MI.getDebugLoc(); |
3081 | |
3082 | switch (MI.getOpcode()) { |
3083 | case PPC::BUILD_UACC: { |
3084 | MCRegister ACC = MI.getOperand(i: 0).getReg(); |
3085 | MCRegister UACC = MI.getOperand(i: 1).getReg(); |
3086 | if (ACC - PPC::ACC0 != UACC - PPC::UACC0) { |
3087 | MCRegister SrcVSR = PPC::VSL0 + (UACC - PPC::UACC0) * 4; |
3088 | MCRegister DstVSR = PPC::VSL0 + (ACC - PPC::ACC0) * 4; |
3089 | // FIXME: This can easily be improved to look up to the top of the MBB |
3090 | // to see if the inputs are XXLOR's. If they are and SrcReg is killed, |
3091 | // we can just re-target any such XXLOR's to DstVSR + offset. |
3092 | for (int VecNo = 0; VecNo < 4; VecNo++) |
3093 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: get(Opcode: PPC::XXLOR), DestReg: DstVSR + VecNo) |
3094 | .addReg(RegNo: SrcVSR + VecNo) |
3095 | .addReg(RegNo: SrcVSR + VecNo); |
3096 | } |
3097 | // BUILD_UACC is expanded to 4 copies of the underlying vsx registers. |
3098 | // So after building the 4 copies, we can replace the BUILD_UACC instruction |
3099 | // with a NOP. |
3100 | [[fallthrough]]; |
3101 | } |
3102 | case PPC::KILL_PAIR: { |
3103 | MI.setDesc(get(Opcode: PPC::UNENCODED_NOP)); |
3104 | MI.removeOperand(OpNo: 1); |
3105 | MI.removeOperand(OpNo: 0); |
3106 | return true; |
3107 | } |
3108 | case TargetOpcode::LOAD_STACK_GUARD: { |
3109 | assert(Subtarget.isTargetLinux() && |
3110 | "Only Linux target is expected to contain LOAD_STACK_GUARD" ); |
3111 | const int64_t Offset = Subtarget.isPPC64() ? -0x7010 : -0x7008; |
3112 | const unsigned Reg = Subtarget.isPPC64() ? PPC::X13 : PPC::R2; |
3113 | MI.setDesc(get(Opcode: Subtarget.isPPC64() ? PPC::LD : PPC::LWZ)); |
3114 | MachineInstrBuilder(*MI.getParent()->getParent(), MI) |
3115 | .addImm(Val: Offset) |
3116 | .addReg(RegNo: Reg); |
3117 | return true; |
3118 | } |
3119 | case PPC::PPCLdFixedAddr: { |
3120 | assert(Subtarget.getTargetTriple().isOSGlibc() && |
3121 | "Only targets with Glibc expected to contain PPCLdFixedAddr" ); |
3122 | int64_t Offset = 0; |
3123 | const unsigned Reg = Subtarget.isPPC64() ? PPC::X13 : PPC::R2; |
3124 | MI.setDesc(get(Opcode: PPC::LWZ)); |
3125 | uint64_t FAType = MI.getOperand(i: 1).getImm(); |
3126 | #undef PPC_LNX_FEATURE |
3127 | #undef PPC_CPU |
3128 | #define PPC_LNX_DEFINE_OFFSETS |
3129 | #include "llvm/TargetParser/PPCTargetParser.def" |
3130 | bool IsLE = Subtarget.isLittleEndian(); |
3131 | bool Is64 = Subtarget.isPPC64(); |
3132 | if (FAType == PPC_FAWORD_HWCAP) { |
3133 | if (IsLE) |
3134 | Offset = Is64 ? PPC_HWCAP_OFFSET_LE64 : PPC_HWCAP_OFFSET_LE32; |
3135 | else |
3136 | Offset = Is64 ? PPC_HWCAP_OFFSET_BE64 : PPC_HWCAP_OFFSET_BE32; |
3137 | } else if (FAType == PPC_FAWORD_HWCAP2) { |
3138 | if (IsLE) |
3139 | Offset = Is64 ? PPC_HWCAP2_OFFSET_LE64 : PPC_HWCAP2_OFFSET_LE32; |
3140 | else |
3141 | Offset = Is64 ? PPC_HWCAP2_OFFSET_BE64 : PPC_HWCAP2_OFFSET_BE32; |
3142 | } else if (FAType == PPC_FAWORD_CPUID) { |
3143 | if (IsLE) |
3144 | Offset = Is64 ? PPC_CPUID_OFFSET_LE64 : PPC_CPUID_OFFSET_LE32; |
3145 | else |
3146 | Offset = Is64 ? PPC_CPUID_OFFSET_BE64 : PPC_CPUID_OFFSET_BE32; |
3147 | } |
3148 | assert(Offset && "Do not know the offset for this fixed addr load" ); |
3149 | MI.removeOperand(OpNo: 1); |
3150 | Subtarget.getTargetMachine().setGlibcHWCAPAccess(); |
3151 | MachineInstrBuilder(*MI.getParent()->getParent(), MI) |
3152 | .addImm(Val: Offset) |
3153 | .addReg(RegNo: Reg); |
3154 | return true; |
3155 | #define PPC_TGT_PARSER_UNDEF_MACROS |
3156 | #include "llvm/TargetParser/PPCTargetParser.def" |
3157 | #undef PPC_TGT_PARSER_UNDEF_MACROS |
3158 | } |
3159 | case PPC::DFLOADf32: |
3160 | case PPC::DFLOADf64: |
3161 | case PPC::DFSTOREf32: |
3162 | case PPC::DFSTOREf64: { |
3163 | assert(Subtarget.hasP9Vector() && |
3164 | "Invalid D-Form Pseudo-ops on Pre-P9 target." ); |
3165 | assert(MI.getOperand(2).isReg() && |
3166 | isAnImmediateOperand(MI.getOperand(1)) && |
3167 | "D-form op must have register and immediate operands" ); |
3168 | return expandVSXMemPseudo(MI); |
3169 | } |
3170 | case PPC::XFLOADf32: |
3171 | case PPC::XFSTOREf32: |
3172 | case PPC::LIWAX: |
3173 | case PPC::LIWZX: |
3174 | case PPC::STIWX: { |
3175 | assert(Subtarget.hasP8Vector() && |
3176 | "Invalid X-Form Pseudo-ops on Pre-P8 target." ); |
3177 | assert(MI.getOperand(2).isReg() && MI.getOperand(1).isReg() && |
3178 | "X-form op must have register and register operands" ); |
3179 | return expandVSXMemPseudo(MI); |
3180 | } |
3181 | case PPC::XFLOADf64: |
3182 | case PPC::XFSTOREf64: { |
3183 | assert(Subtarget.hasVSX() && |
3184 | "Invalid X-Form Pseudo-ops on target that has no VSX." ); |
3185 | assert(MI.getOperand(2).isReg() && MI.getOperand(1).isReg() && |
3186 | "X-form op must have register and register operands" ); |
3187 | return expandVSXMemPseudo(MI); |
3188 | } |
3189 | case PPC::SPILLTOVSR_LD: { |
3190 | Register TargetReg = MI.getOperand(i: 0).getReg(); |
3191 | if (PPC::VSFRCRegClass.contains(Reg: TargetReg)) { |
3192 | MI.setDesc(get(Opcode: PPC::DFLOADf64)); |
3193 | return expandPostRAPseudo(MI); |
3194 | } |
3195 | else |
3196 | MI.setDesc(get(Opcode: PPC::LD)); |
3197 | return true; |
3198 | } |
3199 | case PPC::SPILLTOVSR_ST: { |
3200 | Register SrcReg = MI.getOperand(i: 0).getReg(); |
3201 | if (PPC::VSFRCRegClass.contains(Reg: SrcReg)) { |
3202 | NumStoreSPILLVSRRCAsVec++; |
3203 | MI.setDesc(get(Opcode: PPC::DFSTOREf64)); |
3204 | return expandPostRAPseudo(MI); |
3205 | } else { |
3206 | NumStoreSPILLVSRRCAsGpr++; |
3207 | MI.setDesc(get(Opcode: PPC::STD)); |
3208 | } |
3209 | return true; |
3210 | } |
3211 | case PPC::SPILLTOVSR_LDX: { |
3212 | Register TargetReg = MI.getOperand(i: 0).getReg(); |
3213 | if (PPC::VSFRCRegClass.contains(Reg: TargetReg)) |
3214 | MI.setDesc(get(Opcode: PPC::LXSDX)); |
3215 | else |
3216 | MI.setDesc(get(Opcode: PPC::LDX)); |
3217 | return true; |
3218 | } |
3219 | case PPC::SPILLTOVSR_STX: { |
3220 | Register SrcReg = MI.getOperand(i: 0).getReg(); |
3221 | if (PPC::VSFRCRegClass.contains(Reg: SrcReg)) { |
3222 | NumStoreSPILLVSRRCAsVec++; |
3223 | MI.setDesc(get(Opcode: PPC::STXSDX)); |
3224 | } else { |
3225 | NumStoreSPILLVSRRCAsGpr++; |
3226 | MI.setDesc(get(Opcode: PPC::STDX)); |
3227 | } |
3228 | return true; |
3229 | } |
3230 | |
3231 | // FIXME: Maybe we can expand it in 'PowerPC Expand Atomic' pass. |
3232 | case PPC::CFENCE: |
3233 | case PPC::CFENCE8: { |
3234 | auto Val = MI.getOperand(i: 0).getReg(); |
3235 | unsigned CmpOp = Subtarget.isPPC64() ? PPC::CMPD : PPC::CMPW; |
3236 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: get(Opcode: CmpOp), DestReg: PPC::CR7).addReg(RegNo: Val).addReg(RegNo: Val); |
3237 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: get(Opcode: PPC::CTRL_DEP)) |
3238 | .addImm(Val: PPC::PRED_NE_MINUS) |
3239 | .addReg(RegNo: PPC::CR7) |
3240 | .addImm(Val: 1); |
3241 | MI.setDesc(get(Opcode: PPC::ISYNC)); |
3242 | MI.removeOperand(OpNo: 0); |
3243 | return true; |
3244 | } |
3245 | } |
3246 | return false; |
3247 | } |
3248 | |
3249 | // Essentially a compile-time implementation of a compare->isel sequence. |
3250 | // It takes two constants to compare, along with the true/false registers |
3251 | // and the comparison type (as a subreg to a CR field) and returns one |
3252 | // of the true/false registers, depending on the comparison results. |
3253 | static unsigned selectReg(int64_t Imm1, int64_t Imm2, unsigned CompareOpc, |
3254 | unsigned TrueReg, unsigned FalseReg, |
3255 | unsigned CRSubReg) { |
3256 | // Signed comparisons. The immediates are assumed to be sign-extended. |
3257 | if (CompareOpc == PPC::CMPWI || CompareOpc == PPC::CMPDI) { |
3258 | switch (CRSubReg) { |
3259 | default: llvm_unreachable("Unknown integer comparison type." ); |
3260 | case PPC::sub_lt: |
3261 | return Imm1 < Imm2 ? TrueReg : FalseReg; |
3262 | case PPC::sub_gt: |
3263 | return Imm1 > Imm2 ? TrueReg : FalseReg; |
3264 | case PPC::sub_eq: |
3265 | return Imm1 == Imm2 ? TrueReg : FalseReg; |
3266 | } |
3267 | } |
3268 | // Unsigned comparisons. |
3269 | else if (CompareOpc == PPC::CMPLWI || CompareOpc == PPC::CMPLDI) { |
3270 | switch (CRSubReg) { |
3271 | default: llvm_unreachable("Unknown integer comparison type." ); |
3272 | case PPC::sub_lt: |
3273 | return (uint64_t)Imm1 < (uint64_t)Imm2 ? TrueReg : FalseReg; |
3274 | case PPC::sub_gt: |
3275 | return (uint64_t)Imm1 > (uint64_t)Imm2 ? TrueReg : FalseReg; |
3276 | case PPC::sub_eq: |
3277 | return Imm1 == Imm2 ? TrueReg : FalseReg; |
3278 | } |
3279 | } |
3280 | return PPC::NoRegister; |
3281 | } |
3282 | |
3283 | void PPCInstrInfo::replaceInstrOperandWithImm(MachineInstr &MI, |
3284 | unsigned OpNo, |
3285 | int64_t Imm) const { |
3286 | assert(MI.getOperand(OpNo).isReg() && "Operand must be a REG" ); |
3287 | // Replace the REG with the Immediate. |
3288 | Register InUseReg = MI.getOperand(i: OpNo).getReg(); |
3289 | MI.getOperand(i: OpNo).ChangeToImmediate(ImmVal: Imm); |
3290 | |
3291 | // We need to make sure that the MI didn't have any implicit use |
3292 | // of this REG any more. We don't call MI.implicit_operands().empty() to |
3293 | // return early, since MI's MCID might be changed in calling context, as a |
3294 | // result its number of explicit operands may be changed, thus the begin of |
3295 | // implicit operand is changed. |
3296 | const TargetRegisterInfo *TRI = &getRegisterInfo(); |
3297 | int UseOpIdx = MI.findRegisterUseOperandIdx(Reg: InUseReg, TRI, isKill: false); |
3298 | if (UseOpIdx >= 0) { |
3299 | MachineOperand &MO = MI.getOperand(i: UseOpIdx); |
3300 | if (MO.isImplicit()) |
3301 | // The operands must always be in the following order: |
3302 | // - explicit reg defs, |
3303 | // - other explicit operands (reg uses, immediates, etc.), |
3304 | // - implicit reg defs |
3305 | // - implicit reg uses |
3306 | // Therefore, removing the implicit operand won't change the explicit |
3307 | // operands layout. |
3308 | MI.removeOperand(OpNo: UseOpIdx); |
3309 | } |
3310 | } |
3311 | |
3312 | // Replace an instruction with one that materializes a constant (and sets |
3313 | // CR0 if the original instruction was a record-form instruction). |
3314 | void PPCInstrInfo::replaceInstrWithLI(MachineInstr &MI, |
3315 | const LoadImmediateInfo &LII) const { |
3316 | // Remove existing operands. |
3317 | int OperandToKeep = LII.SetCR ? 1 : 0; |
3318 | for (int i = MI.getNumOperands() - 1; i > OperandToKeep; i--) |
3319 | MI.removeOperand(OpNo: i); |
3320 | |
3321 | // Replace the instruction. |
3322 | if (LII.SetCR) { |
3323 | MI.setDesc(get(Opcode: LII.Is64Bit ? PPC::ANDI8_rec : PPC::ANDI_rec)); |
3324 | // Set the immediate. |
3325 | MachineInstrBuilder(*MI.getParent()->getParent(), MI) |
3326 | .addImm(Val: LII.Imm).addReg(RegNo: PPC::CR0, flags: RegState::ImplicitDefine); |
3327 | return; |
3328 | } |
3329 | else |
3330 | MI.setDesc(get(Opcode: LII.Is64Bit ? PPC::LI8 : PPC::LI)); |
3331 | |
3332 | // Set the immediate. |
3333 | MachineInstrBuilder(*MI.getParent()->getParent(), MI) |
3334 | .addImm(Val: LII.Imm); |
3335 | } |
3336 | |
3337 | MachineInstr *PPCInstrInfo::getDefMIPostRA(unsigned Reg, MachineInstr &MI, |
3338 | bool &SeenIntermediateUse) const { |
3339 | assert(!MI.getParent()->getParent()->getRegInfo().isSSA() && |
3340 | "Should be called after register allocation." ); |
3341 | const TargetRegisterInfo *TRI = &getRegisterInfo(); |
3342 | MachineBasicBlock::reverse_iterator E = MI.getParent()->rend(), It = MI; |
3343 | It++; |
3344 | SeenIntermediateUse = false; |
3345 | for (; It != E; ++It) { |
3346 | if (It->modifiesRegister(Reg, TRI)) |
3347 | return &*It; |
3348 | if (It->readsRegister(Reg, TRI)) |
3349 | SeenIntermediateUse = true; |
3350 | } |
3351 | return nullptr; |
3352 | } |
3353 | |
3354 | void PPCInstrInfo::materializeImmPostRA(MachineBasicBlock &MBB, |
3355 | MachineBasicBlock::iterator MBBI, |
3356 | const DebugLoc &DL, Register Reg, |
3357 | int64_t Imm) const { |
3358 | assert(!MBB.getParent()->getRegInfo().isSSA() && |
3359 | "Register should be in non-SSA form after RA" ); |
3360 | bool isPPC64 = Subtarget.isPPC64(); |
3361 | // FIXME: Materialization here is not optimal. |
3362 | // For some special bit patterns we can use less instructions. |
3363 | // See `selectI64ImmDirect` in PPCISelDAGToDAG.cpp. |
3364 | if (isInt<16>(x: Imm)) { |
3365 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: get(Opcode: isPPC64 ? PPC::LI8 : PPC::LI), DestReg: Reg).addImm(Val: Imm); |
3366 | } else if (isInt<32>(x: Imm)) { |
3367 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: get(Opcode: isPPC64 ? PPC::LIS8 : PPC::LIS), DestReg: Reg) |
3368 | .addImm(Val: Imm >> 16); |
3369 | if (Imm & 0xFFFF) |
3370 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: get(Opcode: isPPC64 ? PPC::ORI8 : PPC::ORI), DestReg: Reg) |
3371 | .addReg(RegNo: Reg, flags: RegState::Kill) |
3372 | .addImm(Val: Imm & 0xFFFF); |
3373 | } else { |
3374 | assert(isPPC64 && "Materializing 64-bit immediate to single register is " |
3375 | "only supported in PPC64" ); |
3376 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: get(Opcode: PPC::LIS8), DestReg: Reg).addImm(Val: Imm >> 48); |
3377 | if ((Imm >> 32) & 0xFFFF) |
3378 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: get(Opcode: PPC::ORI8), DestReg: Reg) |
3379 | .addReg(RegNo: Reg, flags: RegState::Kill) |
3380 | .addImm(Val: (Imm >> 32) & 0xFFFF); |
3381 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: get(Opcode: PPC::RLDICR), DestReg: Reg) |
3382 | .addReg(RegNo: Reg, flags: RegState::Kill) |
3383 | .addImm(Val: 32) |
3384 | .addImm(Val: 31); |
3385 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: get(Opcode: PPC::ORIS8), DestReg: Reg) |
3386 | .addReg(RegNo: Reg, flags: RegState::Kill) |
3387 | .addImm(Val: (Imm >> 16) & 0xFFFF); |
3388 | if (Imm & 0xFFFF) |
3389 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: get(Opcode: PPC::ORI8), DestReg: Reg) |
3390 | .addReg(RegNo: Reg, flags: RegState::Kill) |
3391 | .addImm(Val: Imm & 0xFFFF); |
3392 | } |
3393 | } |
3394 | |
3395 | MachineInstr *PPCInstrInfo::getForwardingDefMI( |
3396 | MachineInstr &MI, |
3397 | unsigned &OpNoForForwarding, |
3398 | bool &SeenIntermediateUse) const { |
3399 | OpNoForForwarding = ~0U; |
3400 | MachineInstr *DefMI = nullptr; |
3401 | MachineRegisterInfo *MRI = &MI.getParent()->getParent()->getRegInfo(); |
3402 | const TargetRegisterInfo *TRI = &getRegisterInfo(); |
3403 | // If we're in SSA, get the defs through the MRI. Otherwise, only look |
3404 | // within the basic block to see if the register is defined using an |
3405 | // LI/LI8/ADDI/ADDI8. |
3406 | if (MRI->isSSA()) { |
3407 | for (int i = 1, e = MI.getNumOperands(); i < e; i++) { |
3408 | if (!MI.getOperand(i).isReg()) |
3409 | continue; |
3410 | Register Reg = MI.getOperand(i).getReg(); |
3411 | if (!Reg.isVirtual()) |
3412 | continue; |
3413 | Register TrueReg = TRI->lookThruCopyLike(SrcReg: Reg, MRI); |
3414 | if (TrueReg.isVirtual()) { |
3415 | MachineInstr *DefMIForTrueReg = MRI->getVRegDef(Reg: TrueReg); |
3416 | if (DefMIForTrueReg->getOpcode() == PPC::LI || |
3417 | DefMIForTrueReg->getOpcode() == PPC::LI8 || |
3418 | DefMIForTrueReg->getOpcode() == PPC::ADDI || |
3419 | DefMIForTrueReg->getOpcode() == PPC::ADDI8) { |
3420 | OpNoForForwarding = i; |
3421 | DefMI = DefMIForTrueReg; |
3422 | // The ADDI and LI operand maybe exist in one instruction at same |
3423 | // time. we prefer to fold LI operand as LI only has one Imm operand |
3424 | // and is more possible to be converted. So if current DefMI is |
3425 | // ADDI/ADDI8, we continue to find possible LI/LI8. |
3426 | if (DefMI->getOpcode() == PPC::LI || DefMI->getOpcode() == PPC::LI8) |
3427 | break; |
3428 | } |
3429 | } |
3430 | } |
3431 | } else { |
3432 | // Looking back through the definition for each operand could be expensive, |
3433 | // so exit early if this isn't an instruction that either has an immediate |
3434 | // form or is already an immediate form that we can handle. |
3435 | ImmInstrInfo III; |
3436 | unsigned Opc = MI.getOpcode(); |
3437 | bool ConvertibleImmForm = |
3438 | Opc == PPC::CMPWI || Opc == PPC::CMPLWI || Opc == PPC::CMPDI || |
3439 | Opc == PPC::CMPLDI || Opc == PPC::ADDI || Opc == PPC::ADDI8 || |
3440 | Opc == PPC::ORI || Opc == PPC::ORI8 || Opc == PPC::XORI || |
3441 | Opc == PPC::XORI8 || Opc == PPC::RLDICL || Opc == PPC::RLDICL_rec || |
3442 | Opc == PPC::RLDICL_32 || Opc == PPC::RLDICL_32_64 || |
3443 | Opc == PPC::RLWINM || Opc == PPC::RLWINM_rec || Opc == PPC::RLWINM8 || |
3444 | Opc == PPC::RLWINM8_rec; |
3445 | bool IsVFReg = (MI.getNumOperands() && MI.getOperand(i: 0).isReg()) |
3446 | ? PPC::isVFRegister(Reg: MI.getOperand(i: 0).getReg()) |
3447 | : false; |
3448 | if (!ConvertibleImmForm && !instrHasImmForm(Opc, IsVFReg, III, PostRA: true)) |
3449 | return nullptr; |
3450 | |
3451 | // Don't convert or %X, %Y, %Y since that's just a register move. |
3452 | if ((Opc == PPC::OR || Opc == PPC::OR8) && |
3453 | MI.getOperand(i: 1).getReg() == MI.getOperand(i: 2).getReg()) |
3454 | return nullptr; |
3455 | for (int i = 1, e = MI.getNumOperands(); i < e; i++) { |
3456 | MachineOperand &MO = MI.getOperand(i); |
3457 | SeenIntermediateUse = false; |
3458 | if (MO.isReg() && MO.isUse() && !MO.isImplicit()) { |
3459 | Register Reg = MI.getOperand(i).getReg(); |
3460 | // If we see another use of this reg between the def and the MI, |
3461 | // we want to flag it so the def isn't deleted. |
3462 | MachineInstr *DefMI = getDefMIPostRA(Reg, MI, SeenIntermediateUse); |
3463 | if (DefMI) { |
3464 | // Is this register defined by some form of add-immediate (including |
3465 | // load-immediate) within this basic block? |
3466 | switch (DefMI->getOpcode()) { |
3467 | default: |
3468 | break; |
3469 | case PPC::LI: |
3470 | case PPC::LI8: |
3471 | case PPC::ADDItocL8: |
3472 | case PPC::ADDI: |
3473 | case PPC::ADDI8: |
3474 | OpNoForForwarding = i; |
3475 | return DefMI; |
3476 | } |
3477 | } |
3478 | } |
3479 | } |
3480 | } |
3481 | return OpNoForForwarding == ~0U ? nullptr : DefMI; |
3482 | } |
3483 | |
3484 | unsigned PPCInstrInfo::getSpillTarget() const { |
3485 | // With P10, we may need to spill paired vector registers or accumulator |
3486 | // registers. MMA implies paired vectors, so we can just check that. |
3487 | bool IsP10Variant = Subtarget.isISA3_1() || Subtarget.pairedVectorMemops(); |
3488 | // P11 uses the P10 target. |
3489 | return Subtarget.isISAFuture() ? 3 : IsP10Variant ? |
3490 | 2 : Subtarget.hasP9Vector() ? |
3491 | 1 : 0; |
3492 | } |
3493 | |
3494 | ArrayRef<unsigned> PPCInstrInfo::getStoreOpcodesForSpillArray() const { |
3495 | return {StoreSpillOpcodesArray[getSpillTarget()], SOK_LastOpcodeSpill}; |
3496 | } |
3497 | |
3498 | ArrayRef<unsigned> PPCInstrInfo::getLoadOpcodesForSpillArray() const { |
3499 | return {LoadSpillOpcodesArray[getSpillTarget()], SOK_LastOpcodeSpill}; |
3500 | } |
3501 | |
3502 | // This opt tries to convert the following imm form to an index form to save an |
3503 | // add for stack variables. |
3504 | // Return false if no such pattern found. |
3505 | // |
3506 | // ADDI instr: ToBeChangedReg = ADDI FrameBaseReg, OffsetAddi |
3507 | // ADD instr: ToBeDeletedReg = ADD ToBeChangedReg(killed), ScaleReg |
3508 | // Imm instr: Reg = op OffsetImm, ToBeDeletedReg(killed) |
3509 | // |
3510 | // can be converted to: |
3511 | // |
3512 | // new ADDI instr: ToBeChangedReg = ADDI FrameBaseReg, (OffsetAddi + OffsetImm) |
3513 | // Index instr: Reg = opx ScaleReg, ToBeChangedReg(killed) |
3514 | // |
3515 | // In order to eliminate ADD instr, make sure that: |
3516 | // 1: (OffsetAddi + OffsetImm) must be int16 since this offset will be used in |
3517 | // new ADDI instr and ADDI can only take int16 Imm. |
3518 | // 2: ToBeChangedReg must be killed in ADD instr and there is no other use |
3519 | // between ADDI and ADD instr since its original def in ADDI will be changed |
3520 | // in new ADDI instr. And also there should be no new def for it between |
3521 | // ADD and Imm instr as ToBeChangedReg will be used in Index instr. |
3522 | // 3: ToBeDeletedReg must be killed in Imm instr and there is no other use |
3523 | // between ADD and Imm instr since ADD instr will be eliminated. |
3524 | // 4: ScaleReg must not be redefined between ADD and Imm instr since it will be |
3525 | // moved to Index instr. |
3526 | bool PPCInstrInfo::foldFrameOffset(MachineInstr &MI) const { |
3527 | MachineFunction *MF = MI.getParent()->getParent(); |
3528 | MachineRegisterInfo *MRI = &MF->getRegInfo(); |
3529 | bool PostRA = !MRI->isSSA(); |
3530 | // Do this opt after PEI which is after RA. The reason is stack slot expansion |
3531 | // in PEI may expose such opportunities since in PEI, stack slot offsets to |
3532 | // frame base(OffsetAddi) are determined. |
3533 | if (!PostRA) |
3534 | return false; |
3535 | unsigned ToBeDeletedReg = 0; |
3536 | int64_t OffsetImm = 0; |
3537 | unsigned XFormOpcode = 0; |
3538 | ImmInstrInfo III; |
3539 | |
3540 | // Check if Imm instr meets requirement. |
3541 | if (!isImmInstrEligibleForFolding(MI, BaseReg&: ToBeDeletedReg, XFormOpcode, OffsetOfImmInstr&: OffsetImm, |
3542 | III)) |
3543 | return false; |
3544 | |
3545 | bool OtherIntermediateUse = false; |
3546 | MachineInstr *ADDMI = getDefMIPostRA(Reg: ToBeDeletedReg, MI, SeenIntermediateUse&: OtherIntermediateUse); |
3547 | |
3548 | // Exit if there is other use between ADD and Imm instr or no def found. |
3549 | if (OtherIntermediateUse || !ADDMI) |
3550 | return false; |
3551 | |
3552 | // Check if ADD instr meets requirement. |
3553 | if (!isADDInstrEligibleForFolding(ADDMI&: *ADDMI)) |
3554 | return false; |
3555 | |
3556 | unsigned ScaleRegIdx = 0; |
3557 | int64_t OffsetAddi = 0; |
3558 | MachineInstr *ADDIMI = nullptr; |
3559 | |
3560 | // Check if there is a valid ToBeChangedReg in ADDMI. |
3561 | // 1: It must be killed. |
3562 | // 2: Its definition must be a valid ADDIMI. |
3563 | // 3: It must satify int16 offset requirement. |
3564 | if (isValidToBeChangedReg(ADDMI, Index: 1, ADDIMI, OffsetAddi, OffsetImm)) |
3565 | ScaleRegIdx = 2; |
3566 | else if (isValidToBeChangedReg(ADDMI, Index: 2, ADDIMI, OffsetAddi, OffsetImm)) |
3567 | ScaleRegIdx = 1; |
3568 | else |
3569 | return false; |
3570 | |
3571 | assert(ADDIMI && "There should be ADDIMI for valid ToBeChangedReg." ); |
3572 | Register ToBeChangedReg = ADDIMI->getOperand(i: 0).getReg(); |
3573 | Register ScaleReg = ADDMI->getOperand(i: ScaleRegIdx).getReg(); |
3574 | auto NewDefFor = [&](unsigned Reg, MachineBasicBlock::iterator Start, |
3575 | MachineBasicBlock::iterator End) { |
3576 | for (auto It = ++Start; It != End; It++) |
3577 | if (It->modifiesRegister(Reg, TRI: &getRegisterInfo())) |
3578 | return true; |
3579 | return false; |
3580 | }; |
3581 | |
3582 | // We are trying to replace the ImmOpNo with ScaleReg. Give up if it is |
3583 | // treated as special zero when ScaleReg is R0/X0 register. |
3584 | if (III.ZeroIsSpecialOrig == III.ImmOpNo && |
3585 | (ScaleReg == PPC::R0 || ScaleReg == PPC::X0)) |
3586 | return false; |
3587 | |
3588 | // Make sure no other def for ToBeChangedReg and ScaleReg between ADD Instr |
3589 | // and Imm Instr. |
3590 | if (NewDefFor(ToBeChangedReg, *ADDMI, MI) || NewDefFor(ScaleReg, *ADDMI, MI)) |
3591 | return false; |
3592 | |
3593 | // Now start to do the transformation. |
3594 | LLVM_DEBUG(dbgs() << "Replace instruction: " |
3595 | << "\n" ); |
3596 | LLVM_DEBUG(ADDIMI->dump()); |
3597 | LLVM_DEBUG(ADDMI->dump()); |
3598 | LLVM_DEBUG(MI.dump()); |
3599 | LLVM_DEBUG(dbgs() << "with: " |
3600 | << "\n" ); |
3601 | |
3602 | // Update ADDI instr. |
3603 | ADDIMI->getOperand(i: 2).setImm(OffsetAddi + OffsetImm); |
3604 | |
3605 | // Update Imm instr. |
3606 | MI.setDesc(get(Opcode: XFormOpcode)); |
3607 | MI.getOperand(i: III.ImmOpNo) |
3608 | .ChangeToRegister(Reg: ScaleReg, isDef: false, isImp: false, |
3609 | isKill: ADDMI->getOperand(i: ScaleRegIdx).isKill()); |
3610 | |
3611 | MI.getOperand(i: III.OpNoForForwarding) |
3612 | .ChangeToRegister(Reg: ToBeChangedReg, isDef: false, isImp: false, isKill: true); |
3613 | |
3614 | // Eliminate ADD instr. |
3615 | ADDMI->eraseFromParent(); |
3616 | |
3617 | LLVM_DEBUG(ADDIMI->dump()); |
3618 | LLVM_DEBUG(MI.dump()); |
3619 | |
3620 | return true; |
3621 | } |
3622 | |
3623 | bool PPCInstrInfo::isADDIInstrEligibleForFolding(MachineInstr &ADDIMI, |
3624 | int64_t &Imm) const { |
3625 | unsigned Opc = ADDIMI.getOpcode(); |
3626 | |
3627 | // Exit if the instruction is not ADDI. |
3628 | if (Opc != PPC::ADDI && Opc != PPC::ADDI8) |
3629 | return false; |
3630 | |
3631 | // The operand may not necessarily be an immediate - it could be a relocation. |
3632 | if (!ADDIMI.getOperand(i: 2).isImm()) |
3633 | return false; |
3634 | |
3635 | Imm = ADDIMI.getOperand(i: 2).getImm(); |
3636 | |
3637 | return true; |
3638 | } |
3639 | |
3640 | bool PPCInstrInfo::isADDInstrEligibleForFolding(MachineInstr &ADDMI) const { |
3641 | unsigned Opc = ADDMI.getOpcode(); |
3642 | |
3643 | // Exit if the instruction is not ADD. |
3644 | return Opc == PPC::ADD4 || Opc == PPC::ADD8; |
3645 | } |
3646 | |
3647 | bool PPCInstrInfo::isImmInstrEligibleForFolding(MachineInstr &MI, |
3648 | unsigned &ToBeDeletedReg, |
3649 | unsigned &XFormOpcode, |
3650 | int64_t &OffsetImm, |
3651 | ImmInstrInfo &III) const { |
3652 | // Only handle load/store. |
3653 | if (!MI.mayLoadOrStore()) |
3654 | return false; |
3655 | |
3656 | unsigned Opc = MI.getOpcode(); |
3657 | |
3658 | XFormOpcode = RI.getMappedIdxOpcForImmOpc(ImmOpcode: Opc); |
3659 | |
3660 | // Exit if instruction has no index form. |
3661 | if (XFormOpcode == PPC::INSTRUCTION_LIST_END) |
3662 | return false; |
3663 | |
3664 | // TODO: sync the logic between instrHasImmForm() and ImmToIdxMap. |
3665 | if (!instrHasImmForm(Opc: XFormOpcode, |
3666 | IsVFReg: PPC::isVFRegister(Reg: MI.getOperand(i: 0).getReg()), III, PostRA: true)) |
3667 | return false; |
3668 | |
3669 | if (!III.IsSummingOperands) |
3670 | return false; |
3671 | |
3672 | MachineOperand ImmOperand = MI.getOperand(i: III.ImmOpNo); |
3673 | MachineOperand RegOperand = MI.getOperand(i: III.OpNoForForwarding); |
3674 | // Only support imm operands, not relocation slots or others. |
3675 | if (!ImmOperand.isImm()) |
3676 | return false; |
3677 | |
3678 | assert(RegOperand.isReg() && "Instruction format is not right" ); |
3679 | |
3680 | // There are other use for ToBeDeletedReg after Imm instr, can not delete it. |
3681 | if (!RegOperand.isKill()) |
3682 | return false; |
3683 | |
3684 | ToBeDeletedReg = RegOperand.getReg(); |
3685 | OffsetImm = ImmOperand.getImm(); |
3686 | |
3687 | return true; |
3688 | } |
3689 | |
3690 | bool PPCInstrInfo::isValidToBeChangedReg(MachineInstr *ADDMI, unsigned Index, |
3691 | MachineInstr *&ADDIMI, |
3692 | int64_t &OffsetAddi, |
3693 | int64_t OffsetImm) const { |
3694 | assert((Index == 1 || Index == 2) && "Invalid operand index for add." ); |
3695 | MachineOperand &MO = ADDMI->getOperand(i: Index); |
3696 | |
3697 | if (!MO.isKill()) |
3698 | return false; |
3699 | |
3700 | bool OtherIntermediateUse = false; |
3701 | |
3702 | ADDIMI = getDefMIPostRA(Reg: MO.getReg(), MI&: *ADDMI, SeenIntermediateUse&: OtherIntermediateUse); |
3703 | // Currently handle only one "add + Imminstr" pair case, exit if other |
3704 | // intermediate use for ToBeChangedReg found. |
3705 | // TODO: handle the cases where there are other "add + Imminstr" pairs |
3706 | // with same offset in Imminstr which is like: |
3707 | // |
3708 | // ADDI instr: ToBeChangedReg = ADDI FrameBaseReg, OffsetAddi |
3709 | // ADD instr1: ToBeDeletedReg1 = ADD ToBeChangedReg, ScaleReg1 |
3710 | // Imm instr1: Reg1 = op1 OffsetImm, ToBeDeletedReg1(killed) |
3711 | // ADD instr2: ToBeDeletedReg2 = ADD ToBeChangedReg(killed), ScaleReg2 |
3712 | // Imm instr2: Reg2 = op2 OffsetImm, ToBeDeletedReg2(killed) |
3713 | // |
3714 | // can be converted to: |
3715 | // |
3716 | // new ADDI instr: ToBeChangedReg = ADDI FrameBaseReg, |
3717 | // (OffsetAddi + OffsetImm) |
3718 | // Index instr1: Reg1 = opx1 ScaleReg1, ToBeChangedReg |
3719 | // Index instr2: Reg2 = opx2 ScaleReg2, ToBeChangedReg(killed) |
3720 | |
3721 | if (OtherIntermediateUse || !ADDIMI) |
3722 | return false; |
3723 | // Check if ADDI instr meets requirement. |
3724 | if (!isADDIInstrEligibleForFolding(ADDIMI&: *ADDIMI, Imm&: OffsetAddi)) |
3725 | return false; |
3726 | |
3727 | if (isInt<16>(x: OffsetAddi + OffsetImm)) |
3728 | return true; |
3729 | return false; |
3730 | } |
3731 | |
3732 | // If this instruction has an immediate form and one of its operands is a |
3733 | // result of a load-immediate or an add-immediate, convert it to |
3734 | // the immediate form if the constant is in range. |
3735 | bool PPCInstrInfo::convertToImmediateForm(MachineInstr &MI, |
3736 | SmallSet<Register, 4> &RegsToUpdate, |
3737 | MachineInstr **KilledDef) const { |
3738 | MachineFunction *MF = MI.getParent()->getParent(); |
3739 | MachineRegisterInfo *MRI = &MF->getRegInfo(); |
3740 | bool PostRA = !MRI->isSSA(); |
3741 | bool SeenIntermediateUse = true; |
3742 | unsigned ForwardingOperand = ~0U; |
3743 | MachineInstr *DefMI = getForwardingDefMI(MI, OpNoForForwarding&: ForwardingOperand, |
3744 | SeenIntermediateUse); |
3745 | if (!DefMI) |
3746 | return false; |
3747 | assert(ForwardingOperand < MI.getNumOperands() && |
3748 | "The forwarding operand needs to be valid at this point" ); |
3749 | bool IsForwardingOperandKilled = MI.getOperand(i: ForwardingOperand).isKill(); |
3750 | bool KillFwdDefMI = !SeenIntermediateUse && IsForwardingOperandKilled; |
3751 | if (KilledDef && KillFwdDefMI) |
3752 | *KilledDef = DefMI; |
3753 | |
3754 | // Conservatively add defs from DefMI and defs/uses from MI to the set of |
3755 | // registers that need their kill flags updated. |
3756 | for (const MachineOperand &MO : DefMI->operands()) |
3757 | if (MO.isReg() && MO.isDef()) |
3758 | RegsToUpdate.insert(V: MO.getReg()); |
3759 | for (const MachineOperand &MO : MI.operands()) |
3760 | if (MO.isReg()) |
3761 | RegsToUpdate.insert(V: MO.getReg()); |
3762 | |
3763 | // If this is a imm instruction and its register operands is produced by ADDI, |
3764 | // put the imm into imm inst directly. |
3765 | if (RI.getMappedIdxOpcForImmOpc(ImmOpcode: MI.getOpcode()) != |
3766 | PPC::INSTRUCTION_LIST_END && |
3767 | transformToNewImmFormFedByAdd(MI, DefMI&: *DefMI, OpNoForForwarding: ForwardingOperand)) |
3768 | return true; |
3769 | |
3770 | ImmInstrInfo III; |
3771 | bool IsVFReg = MI.getOperand(i: 0).isReg() |
3772 | ? PPC::isVFRegister(Reg: MI.getOperand(i: 0).getReg()) |
3773 | : false; |
3774 | bool HasImmForm = instrHasImmForm(Opc: MI.getOpcode(), IsVFReg, III, PostRA); |
3775 | // If this is a reg+reg instruction that has a reg+imm form, |
3776 | // and one of the operands is produced by an add-immediate, |
3777 | // try to convert it. |
3778 | if (HasImmForm && |
3779 | transformToImmFormFedByAdd(MI, III, ConstantOpNo: ForwardingOperand, DefMI&: *DefMI, |
3780 | KillDefMI: KillFwdDefMI)) |
3781 | return true; |
3782 | |
3783 | // If this is a reg+reg instruction that has a reg+imm form, |
3784 | // and one of the operands is produced by LI, convert it now. |
3785 | if (HasImmForm && |
3786 | transformToImmFormFedByLI(MI, III, ConstantOpNo: ForwardingOperand, DefMI&: *DefMI)) |
3787 | return true; |
3788 | |
3789 | // If this is not a reg+reg, but the DefMI is LI/LI8, check if its user MI |
3790 | // can be simpified to LI. |
3791 | if (!HasImmForm && simplifyToLI(MI, DefMI&: *DefMI, OpNoForForwarding: ForwardingOperand, KilledDef)) |
3792 | return true; |
3793 | |
3794 | return false; |
3795 | } |
3796 | |
3797 | bool PPCInstrInfo::combineRLWINM(MachineInstr &MI, |
3798 | MachineInstr **ToErase) const { |
3799 | MachineRegisterInfo *MRI = &MI.getParent()->getParent()->getRegInfo(); |
3800 | Register FoldingReg = MI.getOperand(i: 1).getReg(); |
3801 | if (!FoldingReg.isVirtual()) |
3802 | return false; |
3803 | MachineInstr *SrcMI = MRI->getVRegDef(Reg: FoldingReg); |
3804 | if (SrcMI->getOpcode() != PPC::RLWINM && |
3805 | SrcMI->getOpcode() != PPC::RLWINM_rec && |
3806 | SrcMI->getOpcode() != PPC::RLWINM8 && |
3807 | SrcMI->getOpcode() != PPC::RLWINM8_rec) |
3808 | return false; |
3809 | assert((MI.getOperand(2).isImm() && MI.getOperand(3).isImm() && |
3810 | MI.getOperand(4).isImm() && SrcMI->getOperand(2).isImm() && |
3811 | SrcMI->getOperand(3).isImm() && SrcMI->getOperand(4).isImm()) && |
3812 | "Invalid PPC::RLWINM Instruction!" ); |
3813 | uint64_t SHSrc = SrcMI->getOperand(i: 2).getImm(); |
3814 | uint64_t SHMI = MI.getOperand(i: 2).getImm(); |
3815 | uint64_t MBSrc = SrcMI->getOperand(i: 3).getImm(); |
3816 | uint64_t MBMI = MI.getOperand(i: 3).getImm(); |
3817 | uint64_t MESrc = SrcMI->getOperand(i: 4).getImm(); |
3818 | uint64_t MEMI = MI.getOperand(i: 4).getImm(); |
3819 | |
3820 | assert((MEMI < 32 && MESrc < 32 && MBMI < 32 && MBSrc < 32) && |
3821 | "Invalid PPC::RLWINM Instruction!" ); |
3822 | // If MBMI is bigger than MEMI, we always can not get run of ones. |
3823 | // RotatedSrcMask non-wrap: |
3824 | // 0........31|32........63 |
3825 | // RotatedSrcMask: B---E B---E |
3826 | // MaskMI: -----------|--E B------ |
3827 | // Result: ----- --- (Bad candidate) |
3828 | // |
3829 | // RotatedSrcMask wrap: |
3830 | // 0........31|32........63 |
3831 | // RotatedSrcMask: --E B----|--E B---- |
3832 | // MaskMI: -----------|--E B------ |
3833 | // Result: --- -----|--- ----- (Bad candidate) |
3834 | // |
3835 | // One special case is RotatedSrcMask is a full set mask. |
3836 | // RotatedSrcMask full: |
3837 | // 0........31|32........63 |
3838 | // RotatedSrcMask: ------EB---|-------EB--- |
3839 | // MaskMI: -----------|--E B------ |
3840 | // Result: -----------|--- ------- (Good candidate) |
3841 | |
3842 | // Mark special case. |
3843 | bool SrcMaskFull = (MBSrc - MESrc == 1) || (MBSrc == 0 && MESrc == 31); |
3844 | |
3845 | // For other MBMI > MEMI cases, just return. |
3846 | if ((MBMI > MEMI) && !SrcMaskFull) |
3847 | return false; |
3848 | |
3849 | // Handle MBMI <= MEMI cases. |
3850 | APInt MaskMI = APInt::getBitsSetWithWrap(numBits: 32, loBit: 32 - MEMI - 1, hiBit: 32 - MBMI); |
3851 | // In MI, we only need low 32 bits of SrcMI, just consider about low 32 |
3852 | // bit of SrcMI mask. Note that in APInt, lowerest bit is at index 0, |
3853 | // while in PowerPC ISA, lowerest bit is at index 63. |
3854 | APInt MaskSrc = APInt::getBitsSetWithWrap(numBits: 32, loBit: 32 - MESrc - 1, hiBit: 32 - MBSrc); |
3855 | |
3856 | APInt RotatedSrcMask = MaskSrc.rotl(rotateAmt: SHMI); |
3857 | APInt FinalMask = RotatedSrcMask & MaskMI; |
3858 | uint32_t NewMB, NewME; |
3859 | bool Simplified = false; |
3860 | |
3861 | // If final mask is 0, MI result should be 0 too. |
3862 | if (FinalMask.isZero()) { |
3863 | bool Is64Bit = |
3864 | (MI.getOpcode() == PPC::RLWINM8 || MI.getOpcode() == PPC::RLWINM8_rec); |
3865 | Simplified = true; |
3866 | LLVM_DEBUG(dbgs() << "Replace Instr: " ); |
3867 | LLVM_DEBUG(MI.dump()); |
3868 | |
3869 | if (MI.getOpcode() == PPC::RLWINM || MI.getOpcode() == PPC::RLWINM8) { |
3870 | // Replace MI with "LI 0" |
3871 | MI.removeOperand(OpNo: 4); |
3872 | MI.removeOperand(OpNo: 3); |
3873 | MI.removeOperand(OpNo: 2); |
3874 | MI.getOperand(i: 1).ChangeToImmediate(ImmVal: 0); |
3875 | MI.setDesc(get(Opcode: Is64Bit ? PPC::LI8 : PPC::LI)); |
3876 | } else { |
3877 | // Replace MI with "ANDI_rec reg, 0" |
3878 | MI.removeOperand(OpNo: 4); |
3879 | MI.removeOperand(OpNo: 3); |
3880 | MI.getOperand(i: 2).setImm(0); |
3881 | MI.setDesc(get(Opcode: Is64Bit ? PPC::ANDI8_rec : PPC::ANDI_rec)); |
3882 | MI.getOperand(i: 1).setReg(SrcMI->getOperand(i: 1).getReg()); |
3883 | if (SrcMI->getOperand(i: 1).isKill()) { |
3884 | MI.getOperand(i: 1).setIsKill(true); |
3885 | SrcMI->getOperand(i: 1).setIsKill(false); |
3886 | } else |
3887 | // About to replace MI.getOperand(1), clear its kill flag. |
3888 | MI.getOperand(i: 1).setIsKill(false); |
3889 | } |
3890 | |
3891 | LLVM_DEBUG(dbgs() << "With: " ); |
3892 | LLVM_DEBUG(MI.dump()); |
3893 | |
3894 | } else if ((isRunOfOnes(Val: (unsigned)(FinalMask.getZExtValue()), MB&: NewMB, ME&: NewME) && |
3895 | NewMB <= NewME) || |
3896 | SrcMaskFull) { |
3897 | // Here we only handle MBMI <= MEMI case, so NewMB must be no bigger |
3898 | // than NewME. Otherwise we get a 64 bit value after folding, but MI |
3899 | // return a 32 bit value. |
3900 | Simplified = true; |
3901 | LLVM_DEBUG(dbgs() << "Converting Instr: " ); |
3902 | LLVM_DEBUG(MI.dump()); |
3903 | |
3904 | uint16_t NewSH = (SHSrc + SHMI) % 32; |
3905 | MI.getOperand(i: 2).setImm(NewSH); |
3906 | // If SrcMI mask is full, no need to update MBMI and MEMI. |
3907 | if (!SrcMaskFull) { |
3908 | MI.getOperand(i: 3).setImm(NewMB); |
3909 | MI.getOperand(i: 4).setImm(NewME); |
3910 | } |
3911 | MI.getOperand(i: 1).setReg(SrcMI->getOperand(i: 1).getReg()); |
3912 | if (SrcMI->getOperand(i: 1).isKill()) { |
3913 | MI.getOperand(i: 1).setIsKill(true); |
3914 | SrcMI->getOperand(i: 1).setIsKill(false); |
3915 | } else |
3916 | // About to replace MI.getOperand(1), clear its kill flag. |
3917 | MI.getOperand(i: 1).setIsKill(false); |
3918 | |
3919 | LLVM_DEBUG(dbgs() << "To: " ); |
3920 | LLVM_DEBUG(MI.dump()); |
3921 | } |
3922 | if (Simplified & MRI->use_nodbg_empty(RegNo: FoldingReg) && |
3923 | !SrcMI->hasImplicitDef()) { |
3924 | // If FoldingReg has no non-debug use and it has no implicit def (it |
3925 | // is not RLWINMO or RLWINM8o), it's safe to delete its def SrcMI. |
3926 | // Otherwise keep it. |
3927 | *ToErase = SrcMI; |
3928 | LLVM_DEBUG(dbgs() << "Delete dead instruction: " ); |
3929 | LLVM_DEBUG(SrcMI->dump()); |
3930 | } |
3931 | return Simplified; |
3932 | } |
3933 | |
3934 | bool PPCInstrInfo::instrHasImmForm(unsigned Opc, bool IsVFReg, |
3935 | ImmInstrInfo &III, bool PostRA) const { |
3936 | // The vast majority of the instructions would need their operand 2 replaced |
3937 | // with an immediate when switching to the reg+imm form. A marked exception |
3938 | // are the update form loads/stores for which a constant operand 2 would need |
3939 | // to turn into a displacement and move operand 1 to the operand 2 position. |
3940 | III.ImmOpNo = 2; |
3941 | III.OpNoForForwarding = 2; |
3942 | III.ImmWidth = 16; |
3943 | III.ImmMustBeMultipleOf = 1; |
3944 | III.TruncateImmTo = 0; |
3945 | III.IsSummingOperands = false; |
3946 | switch (Opc) { |
3947 | default: return false; |
3948 | case PPC::ADD4: |
3949 | case PPC::ADD8: |
3950 | III.SignedImm = true; |
3951 | III.ZeroIsSpecialOrig = 0; |
3952 | III.ZeroIsSpecialNew = 1; |
3953 | III.IsCommutative = true; |
3954 | III.IsSummingOperands = true; |
3955 | III.ImmOpcode = Opc == PPC::ADD4 ? PPC::ADDI : PPC::ADDI8; |
3956 | break; |
3957 | case PPC::ADDC: |
3958 | case PPC::ADDC8: |
3959 | III.SignedImm = true; |
3960 | III.ZeroIsSpecialOrig = 0; |
3961 | III.ZeroIsSpecialNew = 0; |
3962 | III.IsCommutative = true; |
3963 | III.IsSummingOperands = true; |
3964 | III.ImmOpcode = Opc == PPC::ADDC ? PPC::ADDIC : PPC::ADDIC8; |
3965 | break; |
3966 | case PPC::ADDC_rec: |
3967 | III.SignedImm = true; |
3968 | III.ZeroIsSpecialOrig = 0; |
3969 | III.ZeroIsSpecialNew = 0; |
3970 | III.IsCommutative = true; |
3971 | III.IsSummingOperands = true; |
3972 | III.ImmOpcode = PPC::ADDIC_rec; |
3973 | break; |
3974 | case PPC::SUBFC: |
3975 | case PPC::SUBFC8: |
3976 | III.SignedImm = true; |
3977 | III.ZeroIsSpecialOrig = 0; |
3978 | III.ZeroIsSpecialNew = 0; |
3979 | III.IsCommutative = false; |
3980 | III.ImmOpcode = Opc == PPC::SUBFC ? PPC::SUBFIC : PPC::SUBFIC8; |
3981 | break; |
3982 | case PPC::CMPW: |
3983 | case PPC::CMPD: |
3984 | III.SignedImm = true; |
3985 | III.ZeroIsSpecialOrig = 0; |
3986 | III.ZeroIsSpecialNew = 0; |
3987 | III.IsCommutative = false; |
3988 | III.ImmOpcode = Opc == PPC::CMPW ? PPC::CMPWI : PPC::CMPDI; |
3989 | break; |
3990 | case PPC::CMPLW: |
3991 | case PPC::CMPLD: |
3992 | III.SignedImm = false; |
3993 | III.ZeroIsSpecialOrig = 0; |
3994 | III.ZeroIsSpecialNew = 0; |
3995 | III.IsCommutative = false; |
3996 | III.ImmOpcode = Opc == PPC::CMPLW ? PPC::CMPLWI : PPC::CMPLDI; |
3997 | break; |
3998 | case PPC::AND_rec: |
3999 | case PPC::AND8_rec: |
4000 | case PPC::OR: |
4001 | case PPC::OR8: |
4002 | case PPC::XOR: |
4003 | case PPC::XOR8: |
4004 | III.SignedImm = false; |
4005 | III.ZeroIsSpecialOrig = 0; |
4006 | III.ZeroIsSpecialNew = 0; |
4007 | III.IsCommutative = true; |
4008 | switch(Opc) { |
4009 | default: llvm_unreachable("Unknown opcode" ); |
4010 | case PPC::AND_rec: |
4011 | III.ImmOpcode = PPC::ANDI_rec; |
4012 | break; |
4013 | case PPC::AND8_rec: |
4014 | III.ImmOpcode = PPC::ANDI8_rec; |
4015 | break; |
4016 | case PPC::OR: III.ImmOpcode = PPC::ORI; break; |
4017 | case PPC::OR8: III.ImmOpcode = PPC::ORI8; break; |
4018 | case PPC::XOR: III.ImmOpcode = PPC::XORI; break; |
4019 | case PPC::XOR8: III.ImmOpcode = PPC::XORI8; break; |
4020 | } |
4021 | break; |
4022 | case PPC::RLWNM: |
4023 | case PPC::RLWNM8: |
4024 | case PPC::RLWNM_rec: |
4025 | case PPC::RLWNM8_rec: |
4026 | case PPC::SLW: |
4027 | case PPC::SLW8: |
4028 | case PPC::SLW_rec: |
4029 | case PPC::SLW8_rec: |
4030 | case PPC::SRW: |
4031 | case PPC::SRW8: |
4032 | case PPC::SRW_rec: |
4033 | case PPC::SRW8_rec: |
4034 | case PPC::SRAW: |
4035 | case PPC::SRAW_rec: |
4036 | III.SignedImm = false; |
4037 | III.ZeroIsSpecialOrig = 0; |
4038 | III.ZeroIsSpecialNew = 0; |
4039 | III.IsCommutative = false; |
4040 | // This isn't actually true, but the instructions ignore any of the |
4041 | // upper bits, so any immediate loaded with an LI is acceptable. |
4042 | // This does not apply to shift right algebraic because a value |
4043 | // out of range will produce a -1/0. |
4044 | III.ImmWidth = 16; |
4045 | if (Opc == PPC::RLWNM || Opc == PPC::RLWNM8 || Opc == PPC::RLWNM_rec || |
4046 | Opc == PPC::RLWNM8_rec) |
4047 | III.TruncateImmTo = 5; |
4048 | else |
4049 | III.TruncateImmTo = 6; |
4050 | switch(Opc) { |
4051 | default: llvm_unreachable("Unknown opcode" ); |
4052 | case PPC::RLWNM: III.ImmOpcode = PPC::RLWINM; break; |
4053 | case PPC::RLWNM8: III.ImmOpcode = PPC::RLWINM8; break; |
4054 | case PPC::RLWNM_rec: |
4055 | III.ImmOpcode = PPC::RLWINM_rec; |
4056 | break; |
4057 | case PPC::RLWNM8_rec: |
4058 | III.ImmOpcode = PPC::RLWINM8_rec; |
4059 | break; |
4060 | case PPC::SLW: III.ImmOpcode = PPC::RLWINM; break; |
4061 | case PPC::SLW8: III.ImmOpcode = PPC::RLWINM8; break; |
4062 | case PPC::SLW_rec: |
4063 | III.ImmOpcode = PPC::RLWINM_rec; |
4064 | break; |
4065 | case PPC::SLW8_rec: |
4066 | III.ImmOpcode = PPC::RLWINM8_rec; |
4067 | break; |
4068 | case PPC::SRW: III.ImmOpcode = PPC::RLWINM; break; |
4069 | case PPC::SRW8: III.ImmOpcode = PPC::RLWINM8; break; |
4070 | case PPC::SRW_rec: |
4071 | III.ImmOpcode = PPC::RLWINM_rec; |
4072 | break; |
4073 | case PPC::SRW8_rec: |
4074 | III.ImmOpcode = PPC::RLWINM8_rec; |
4075 | break; |
4076 | case PPC::SRAW: |
4077 | III.ImmWidth = 5; |
4078 | III.TruncateImmTo = 0; |
4079 | III.ImmOpcode = PPC::SRAWI; |
4080 | break; |
4081 | case PPC::SRAW_rec: |
4082 | III.ImmWidth = 5; |
4083 | III.TruncateImmTo = 0; |
4084 | III.ImmOpcode = PPC::SRAWI_rec; |
4085 | break; |
4086 | } |
4087 | break; |
4088 | case PPC::RLDCL: |
4089 | case PPC::RLDCL_rec: |
4090 | case PPC::RLDCR: |
4091 | case PPC::RLDCR_rec: |
4092 | case PPC::SLD: |
4093 | case PPC::SLD_rec: |
4094 | case PPC::SRD: |
4095 | case PPC::SRD_rec: |
4096 | case PPC::SRAD: |
4097 | case PPC::SRAD_rec: |
4098 | III.SignedImm = false; |
4099 | III.ZeroIsSpecialOrig = 0; |
4100 | III.ZeroIsSpecialNew = 0; |
4101 | III.IsCommutative = false; |
4102 | // This isn't actually true, but the instructions ignore any of the |
4103 | // upper bits, so any immediate loaded with an LI is acceptable. |
4104 | // This does not apply to shift right algebraic because a value |
4105 | // out of range will produce a -1/0. |
4106 | III.ImmWidth = 16; |
4107 | if (Opc == PPC::RLDCL || Opc == PPC::RLDCL_rec || Opc == PPC::RLDCR || |
4108 | Opc == PPC::RLDCR_rec) |
4109 | III.TruncateImmTo = 6; |
4110 | else |
4111 | III.TruncateImmTo = 7; |
4112 | switch(Opc) { |
4113 | default: llvm_unreachable("Unknown opcode" ); |
4114 | case PPC::RLDCL: III.ImmOpcode = PPC::RLDICL; break; |
4115 | case PPC::RLDCL_rec: |
4116 | III.ImmOpcode = PPC::RLDICL_rec; |
4117 | break; |
4118 | case PPC::RLDCR: III.ImmOpcode = PPC::RLDICR; break; |
4119 | case PPC::RLDCR_rec: |
4120 | III.ImmOpcode = PPC::RLDICR_rec; |
4121 | break; |
4122 | case PPC::SLD: III.ImmOpcode = PPC::RLDICR; break; |
4123 | case PPC::SLD_rec: |
4124 | III.ImmOpcode = PPC::RLDICR_rec; |
4125 | break; |
4126 | case PPC::SRD: III.ImmOpcode = PPC::RLDICL; break; |
4127 | case PPC::SRD_rec: |
4128 | III.ImmOpcode = PPC::RLDICL_rec; |
4129 | break; |
4130 | case PPC::SRAD: |
4131 | III.ImmWidth = 6; |
4132 | III.TruncateImmTo = 0; |
4133 | III.ImmOpcode = PPC::SRADI; |
4134 | break; |
4135 | case PPC::SRAD_rec: |
4136 | III.ImmWidth = 6; |
4137 | III.TruncateImmTo = 0; |
4138 | III.ImmOpcode = PPC::SRADI_rec; |
4139 | break; |
4140 | } |
4141 | break; |
4142 | // Loads and stores: |
4143 | case PPC::LBZX: |
4144 | case PPC::LBZX8: |
4145 | case PPC::LHZX: |
4146 | case PPC::LHZX8: |
4147 | case PPC::LHAX: |
4148 | case PPC::LHAX8: |
4149 | case PPC::LWZX: |
4150 | case PPC::LWZX8: |
4151 | case PPC::LWAX: |
4152 | case PPC::LDX: |
4153 | case PPC::LFSX: |
4154 | case PPC::LFDX: |
4155 | case PPC::STBX: |
4156 | case PPC::STBX8: |
4157 | case PPC::STHX: |
4158 | case PPC::STHX8: |
4159 | case PPC::STWX: |
4160 | case PPC::STWX8: |
4161 | case PPC::STDX: |
4162 | case PPC::STFSX: |
4163 | case PPC::STFDX: |
4164 | III.SignedImm = true; |
4165 | III.ZeroIsSpecialOrig = 1; |
4166 | III.ZeroIsSpecialNew = 2; |
4167 | III.IsCommutative = true; |
4168 | III.IsSummingOperands = true; |
4169 | III.ImmOpNo = 1; |
4170 | III.OpNoForForwarding = 2; |
4171 | switch(Opc) { |
4172 | default: llvm_unreachable("Unknown opcode" ); |
4173 | case PPC::LBZX: III.ImmOpcode = PPC::LBZ; break; |
4174 | case PPC::LBZX8: III.ImmOpcode = PPC::LBZ8; break; |
4175 | case PPC::LHZX: III.ImmOpcode = PPC::LHZ; break; |
4176 | case PPC::LHZX8: III.ImmOpcode = PPC::LHZ8; break; |
4177 | case PPC::LHAX: III.ImmOpcode = PPC::LHA; break; |
4178 | case PPC::LHAX8: III.ImmOpcode = PPC::LHA8; break; |
4179 | case PPC::LWZX: III.ImmOpcode = PPC::LWZ; break; |
4180 | case PPC::LWZX8: III.ImmOpcode = PPC::LWZ8; break; |
4181 | case PPC::LWAX: |
4182 | III.ImmOpcode = PPC::LWA; |
4183 | III.ImmMustBeMultipleOf = 4; |
4184 | break; |
4185 | case PPC::LDX: III.ImmOpcode = PPC::LD; III.ImmMustBeMultipleOf = 4; break; |
4186 | case PPC::LFSX: III.ImmOpcode = PPC::LFS; break; |
4187 | case PPC::LFDX: III.ImmOpcode = PPC::LFD; break; |
4188 | case PPC::STBX: III.ImmOpcode = PPC::STB; break; |
4189 | case PPC::STBX8: III.ImmOpcode = PPC::STB8; break; |
4190 | case PPC::STHX: III.ImmOpcode = PPC::STH; break; |
4191 | case PPC::STHX8: III.ImmOpcode = PPC::STH8; break; |
4192 | case PPC::STWX: III.ImmOpcode = PPC::STW; break; |
4193 | case PPC::STWX8: III.ImmOpcode = PPC::STW8; break; |
4194 | case PPC::STDX: |
4195 | III.ImmOpcode = PPC::STD; |
4196 | III.ImmMustBeMultipleOf = 4; |
4197 | break; |
4198 | case PPC::STFSX: III.ImmOpcode = PPC::STFS; break; |
4199 | case PPC::STFDX: III.ImmOpcode = PPC::STFD; break; |
4200 | } |
4201 | break; |
4202 | case PPC::LBZUX: |
4203 | case PPC::LBZUX8: |
4204 | case PPC::LHZUX: |
4205 | case PPC::LHZUX8: |
4206 | case PPC::LHAUX: |
4207 | case PPC::LHAUX8: |
4208 | case PPC::LWZUX: |
4209 | case PPC::LWZUX8: |
4210 | case PPC::LDUX: |
4211 | case PPC::LFSUX: |
4212 | case PPC::LFDUX: |
4213 | case PPC::STBUX: |
4214 | case PPC::STBUX8: |
4215 | case PPC::STHUX: |
4216 | case PPC::STHUX8: |
4217 | case PPC::STWUX: |
4218 | case PPC::STWUX8: |
4219 | case PPC::STDUX: |
4220 | case PPC::STFSUX: |
4221 | case PPC::STFDUX: |
4222 | III.SignedImm = true; |
4223 | III.ZeroIsSpecialOrig = 2; |
4224 | III.ZeroIsSpecialNew = 3; |
4225 | III.IsCommutative = false; |
4226 | III.IsSummingOperands = true; |
4227 | III.ImmOpNo = 2; |
4228 | III.OpNoForForwarding = 3; |
4229 | switch(Opc) { |
4230 | default: llvm_unreachable("Unknown opcode" ); |
4231 | case PPC::LBZUX: III.ImmOpcode = PPC::LBZU; break; |
4232 | case PPC::LBZUX8: III.ImmOpcode = PPC::LBZU8; break; |
4233 | case PPC::LHZUX: III.ImmOpcode = PPC::LHZU; break; |
4234 | case PPC::LHZUX8: III.ImmOpcode = PPC::LHZU8; break; |
4235 | case PPC::LHAUX: III.ImmOpcode = PPC::LHAU; break; |
4236 | case PPC::LHAUX8: III.ImmOpcode = PPC::LHAU8; break; |
4237 | case PPC::LWZUX: III.ImmOpcode = PPC::LWZU; break; |
4238 | case PPC::LWZUX8: III.ImmOpcode = PPC::LWZU8; break; |
4239 | case PPC::LDUX: |
4240 | III.ImmOpcode = PPC::LDU; |
4241 | III.ImmMustBeMultipleOf = 4; |
4242 | break; |
4243 | case PPC::LFSUX: III.ImmOpcode = PPC::LFSU; break; |
4244 | case PPC::LFDUX: III.ImmOpcode = PPC::LFDU; break; |
4245 | case PPC::STBUX: III.ImmOpcode = PPC::STBU; break; |
4246 | case PPC::STBUX8: III.ImmOpcode = PPC::STBU8; break; |
4247 | case PPC::STHUX: III.ImmOpcode = PPC::STHU; break; |
4248 | case PPC::STHUX8: III.ImmOpcode = PPC::STHU8; break; |
4249 | case PPC::STWUX: III.ImmOpcode = PPC::STWU; break; |
4250 | case PPC::STWUX8: III.ImmOpcode = PPC::STWU8; break; |
4251 | case PPC::STDUX: |
4252 | III.ImmOpcode = PPC::STDU; |
4253 | III.ImmMustBeMultipleOf = 4; |
4254 | break; |
4255 | case PPC::STFSUX: III.ImmOpcode = PPC::STFSU; break; |
4256 | case PPC::STFDUX: III.ImmOpcode = PPC::STFDU; break; |
4257 | } |
4258 | break; |
4259 | // Power9 and up only. For some of these, the X-Form version has access to all |
4260 | // 64 VSR's whereas the D-Form only has access to the VR's. We replace those |
4261 | // with pseudo-ops pre-ra and for post-ra, we check that the register loaded |
4262 | // into or stored from is one of the VR registers. |
4263 | case PPC::LXVX: |
4264 | case PPC::LXSSPX: |
4265 | case PPC::LXSDX: |
4266 | case PPC::STXVX: |
4267 | case PPC::STXSSPX: |
4268 | case PPC::STXSDX: |
4269 | case PPC::XFLOADf32: |
4270 | case PPC::XFLOADf64: |
4271 | case PPC::XFSTOREf32: |
4272 | case PPC::XFSTOREf64: |
4273 | if (!Subtarget.hasP9Vector()) |
4274 | return false; |
4275 | III.SignedImm = true; |
4276 | III.ZeroIsSpecialOrig = 1; |
4277 | III.ZeroIsSpecialNew = 2; |
4278 | III.IsCommutative = true; |
4279 | III.IsSummingOperands = true; |
4280 | III.ImmOpNo = 1; |
4281 | III.OpNoForForwarding = 2; |
4282 | III.ImmMustBeMultipleOf = 4; |
4283 | switch(Opc) { |
4284 | default: llvm_unreachable("Unknown opcode" ); |
4285 | case PPC::LXVX: |
4286 | III.ImmOpcode = PPC::LXV; |
4287 | III.ImmMustBeMultipleOf = 16; |
4288 | break; |
4289 | case PPC::LXSSPX: |
4290 | if (PostRA) { |
4291 | if (IsVFReg) |
4292 | III.ImmOpcode = PPC::LXSSP; |
4293 | else { |
4294 | III.ImmOpcode = PPC::LFS; |
4295 | III.ImmMustBeMultipleOf = 1; |
4296 | } |
4297 | break; |
4298 | } |
4299 | [[fallthrough]]; |
4300 | case PPC::XFLOADf32: |
4301 | III.ImmOpcode = PPC::DFLOADf32; |
4302 | break; |
4303 | case PPC::LXSDX: |
4304 | if (PostRA) { |
4305 | if (IsVFReg) |
4306 | III.ImmOpcode = PPC::LXSD; |
4307 | else { |
4308 | III.ImmOpcode = PPC::LFD; |
4309 | III.ImmMustBeMultipleOf = 1; |
4310 | } |
4311 | break; |
4312 | } |
4313 | [[fallthrough]]; |
4314 | case PPC::XFLOADf64: |
4315 | III.ImmOpcode = PPC::DFLOADf64; |
4316 | break; |
4317 | case PPC::STXVX: |
4318 | III.ImmOpcode = PPC::STXV; |
4319 | III.ImmMustBeMultipleOf = 16; |
4320 | break; |
4321 | case PPC::STXSSPX: |
4322 | if (PostRA) { |
4323 | if (IsVFReg) |
4324 | III.ImmOpcode = PPC::STXSSP; |
4325 | else { |
4326 | III.ImmOpcode = PPC::STFS; |
4327 | III.ImmMustBeMultipleOf = 1; |
4328 | } |
4329 | break; |
4330 | } |
4331 | [[fallthrough]]; |
4332 | case PPC::XFSTOREf32: |
4333 | III.ImmOpcode = PPC::DFSTOREf32; |
4334 | break; |
4335 | case PPC::STXSDX: |
4336 | if (PostRA) { |
4337 | if (IsVFReg) |
4338 | III.ImmOpcode = PPC::STXSD; |
4339 | else { |
4340 | III.ImmOpcode = PPC::STFD; |
4341 | III.ImmMustBeMultipleOf = 1; |
4342 | } |
4343 | break; |
4344 | } |
4345 | [[fallthrough]]; |
4346 | case PPC::XFSTOREf64: |
4347 | III.ImmOpcode = PPC::DFSTOREf64; |
4348 | break; |
4349 | } |
4350 | break; |
4351 | } |
4352 | return true; |
4353 | } |
4354 | |
4355 | // Utility function for swaping two arbitrary operands of an instruction. |
4356 | static void swapMIOperands(MachineInstr &MI, unsigned Op1, unsigned Op2) { |
4357 | assert(Op1 != Op2 && "Cannot swap operand with itself." ); |
4358 | |
4359 | unsigned MaxOp = std::max(a: Op1, b: Op2); |
4360 | unsigned MinOp = std::min(a: Op1, b: Op2); |
4361 | MachineOperand MOp1 = MI.getOperand(i: MinOp); |
4362 | MachineOperand MOp2 = MI.getOperand(i: MaxOp); |
4363 | MI.removeOperand(OpNo: std::max(a: Op1, b: Op2)); |
4364 | MI.removeOperand(OpNo: std::min(a: Op1, b: Op2)); |
4365 | |
4366 | // If the operands we are swapping are the two at the end (the common case) |
4367 | // we can just remove both and add them in the opposite order. |
4368 | if (MaxOp - MinOp == 1 && MI.getNumOperands() == MinOp) { |
4369 | MI.addOperand(Op: MOp2); |
4370 | MI.addOperand(Op: MOp1); |
4371 | } else { |
4372 | // Store all operands in a temporary vector, remove them and re-add in the |
4373 | // right order. |
4374 | SmallVector<MachineOperand, 2> MOps; |
4375 | unsigned TotalOps = MI.getNumOperands() + 2; // We've already removed 2 ops. |
4376 | for (unsigned i = MI.getNumOperands() - 1; i >= MinOp; i--) { |
4377 | MOps.push_back(Elt: MI.getOperand(i)); |
4378 | MI.removeOperand(OpNo: i); |
4379 | } |
4380 | // MOp2 needs to be added next. |
4381 | MI.addOperand(Op: MOp2); |
4382 | // Now add the rest. |
4383 | for (unsigned i = MI.getNumOperands(); i < TotalOps; i++) { |
4384 | if (i == MaxOp) |
4385 | MI.addOperand(Op: MOp1); |
4386 | else { |
4387 | MI.addOperand(Op: MOps.back()); |
4388 | MOps.pop_back(); |
4389 | } |
4390 | } |
4391 | } |
4392 | } |
4393 | |
4394 | // Check if the 'MI' that has the index OpNoForForwarding |
4395 | // meets the requirement described in the ImmInstrInfo. |
4396 | bool PPCInstrInfo::isUseMIElgibleForForwarding(MachineInstr &MI, |
4397 | const ImmInstrInfo &III, |
4398 | unsigned OpNoForForwarding |
4399 | ) const { |
4400 | // As the algorithm of checking for PPC::ZERO/PPC::ZERO8 |
4401 | // would not work pre-RA, we can only do the check post RA. |
4402 | MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo(); |
4403 | if (MRI.isSSA()) |
4404 | return false; |
4405 | |
4406 | // Cannot do the transform if MI isn't summing the operands. |
4407 | if (!III.IsSummingOperands) |
4408 | return false; |
4409 | |
4410 | // The instruction we are trying to replace must have the ZeroIsSpecialOrig set. |
4411 | if (!III.ZeroIsSpecialOrig) |
4412 | return false; |
4413 | |
4414 | // We cannot do the transform if the operand we are trying to replace |
4415 | // isn't the same as the operand the instruction allows. |
4416 | if (OpNoForForwarding != III.OpNoForForwarding) |
4417 | return false; |
4418 | |
4419 | // Check if the instruction we are trying to transform really has |
4420 | // the special zero register as its operand. |
4421 | if (MI.getOperand(i: III.ZeroIsSpecialOrig).getReg() != PPC::ZERO && |
4422 | MI.getOperand(i: III.ZeroIsSpecialOrig).getReg() != PPC::ZERO8) |
4423 | return false; |
4424 | |
4425 | // This machine instruction is convertible if it is, |
4426 | // 1. summing the operands. |
4427 | // 2. one of the operands is special zero register. |
4428 | // 3. the operand we are trying to replace is allowed by the MI. |
4429 | return true; |
4430 | } |
4431 | |
4432 | // Check if the DefMI is the add inst and set the ImmMO and RegMO |
4433 | // accordingly. |
4434 | bool PPCInstrInfo::isDefMIElgibleForForwarding(MachineInstr &DefMI, |
4435 | const ImmInstrInfo &III, |
4436 | MachineOperand *&ImmMO, |
4437 | MachineOperand *&RegMO) const { |
4438 | unsigned Opc = DefMI.getOpcode(); |
4439 | if (Opc != PPC::ADDItocL8 && Opc != PPC::ADDI && Opc != PPC::ADDI8) |
4440 | return false; |
4441 | |
4442 | // Skip the optimization of transformTo[NewImm|Imm]FormFedByAdd for ADDItocL8 |
4443 | // on AIX which is used for toc-data access. TODO: Follow up to see if it can |
4444 | // apply for AIX toc-data as well. |
4445 | if (Opc == PPC::ADDItocL8 && Subtarget.isAIX()) |
4446 | return false; |
4447 | |
4448 | assert(DefMI.getNumOperands() >= 3 && |
4449 | "Add inst must have at least three operands" ); |
4450 | RegMO = &DefMI.getOperand(i: 1); |
4451 | ImmMO = &DefMI.getOperand(i: 2); |
4452 | |
4453 | // Before RA, ADDI first operand could be a frame index. |
4454 | if (!RegMO->isReg()) |
4455 | return false; |
4456 | |
4457 | // This DefMI is elgible for forwarding if it is: |
4458 | // 1. add inst |
4459 | // 2. one of the operands is Imm/CPI/Global. |
4460 | return isAnImmediateOperand(MO: *ImmMO); |
4461 | } |
4462 | |
4463 | bool PPCInstrInfo::isRegElgibleForForwarding( |
4464 | const MachineOperand &RegMO, const MachineInstr &DefMI, |
4465 | const MachineInstr &MI, bool KillDefMI, |
4466 | bool &IsFwdFeederRegKilled, bool &SeenIntermediateUse) const { |
4467 | // x = addi y, imm |
4468 | // ... |
4469 | // z = lfdx 0, x -> z = lfd imm(y) |
4470 | // The Reg "y" can be forwarded to the MI(z) only when there is no DEF |
4471 | // of "y" between the DEF of "x" and "z". |
4472 | // The query is only valid post RA. |
4473 | const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo(); |
4474 | if (MRI.isSSA()) |
4475 | return false; |
4476 | |
4477 | Register Reg = RegMO.getReg(); |
4478 | |
4479 | // Walking the inst in reverse(MI-->DefMI) to get the last DEF of the Reg. |
4480 | MachineBasicBlock::const_reverse_iterator It = MI; |
4481 | MachineBasicBlock::const_reverse_iterator E = MI.getParent()->rend(); |
4482 | It++; |
4483 | for (; It != E; ++It) { |
4484 | if (It->modifiesRegister(Reg, TRI: &getRegisterInfo()) && (&*It) != &DefMI) |
4485 | return false; |
4486 | else if (It->killsRegister(Reg, TRI: &getRegisterInfo()) && (&*It) != &DefMI) |
4487 | IsFwdFeederRegKilled = true; |
4488 | if (It->readsRegister(Reg, TRI: &getRegisterInfo()) && (&*It) != &DefMI) |
4489 | SeenIntermediateUse = true; |
4490 | // Made it to DefMI without encountering a clobber. |
4491 | if ((&*It) == &DefMI) |
4492 | break; |
4493 | } |
4494 | assert((&*It) == &DefMI && "DefMI is missing" ); |
4495 | |
4496 | // If DefMI also defines the register to be forwarded, we can only forward it |
4497 | // if DefMI is being erased. |
4498 | if (DefMI.modifiesRegister(Reg, TRI: &getRegisterInfo())) |
4499 | return KillDefMI; |
4500 | |
4501 | return true; |
4502 | } |
4503 | |
4504 | bool PPCInstrInfo::isImmElgibleForForwarding(const MachineOperand &ImmMO, |
4505 | const MachineInstr &DefMI, |
4506 | const ImmInstrInfo &III, |
4507 | int64_t &Imm, |
4508 | int64_t BaseImm) const { |
4509 | assert(isAnImmediateOperand(ImmMO) && "ImmMO is NOT an immediate" ); |
4510 | if (DefMI.getOpcode() == PPC::ADDItocL8) { |
4511 | // The operand for ADDItocL8 is CPI, which isn't imm at compiling time, |
4512 | // However, we know that, it is 16-bit width, and has the alignment of 4. |
4513 | // Check if the instruction met the requirement. |
4514 | if (III.ImmMustBeMultipleOf > 4 || |
4515 | III.TruncateImmTo || III.ImmWidth != 16) |
4516 | return false; |
4517 | |
4518 | // Going from XForm to DForm loads means that the displacement needs to be |
4519 | // not just an immediate but also a multiple of 4, or 16 depending on the |
4520 | // load. A DForm load cannot be represented if it is a multiple of say 2. |
4521 | // XForm loads do not have this restriction. |
4522 | if (ImmMO.isGlobal()) { |
4523 | const DataLayout &DL = ImmMO.getGlobal()->getDataLayout(); |
4524 | if (ImmMO.getGlobal()->getPointerAlignment(DL) < III.ImmMustBeMultipleOf) |
4525 | return false; |
4526 | } |
4527 | |
4528 | return true; |
4529 | } |
4530 | |
4531 | if (ImmMO.isImm()) { |
4532 | // It is Imm, we need to check if the Imm fit the range. |
4533 | // Sign-extend to 64-bits. |
4534 | // DefMI may be folded with another imm form instruction, the result Imm is |
4535 | // the sum of Imm of DefMI and BaseImm which is from imm form instruction. |
4536 | APInt ActualValue(64, ImmMO.getImm() + BaseImm, true); |
4537 | if (III.SignedImm && !ActualValue.isSignedIntN(N: III.ImmWidth)) |
4538 | return false; |
4539 | if (!III.SignedImm && !ActualValue.isIntN(N: III.ImmWidth)) |
4540 | return false; |
4541 | Imm = SignExtend64<16>(x: ImmMO.getImm() + BaseImm); |
4542 | |
4543 | if (Imm % III.ImmMustBeMultipleOf) |
4544 | return false; |
4545 | if (III.TruncateImmTo) |
4546 | Imm &= ((1 << III.TruncateImmTo) - 1); |
4547 | } |
4548 | else |
4549 | return false; |
4550 | |
4551 | // This ImmMO is forwarded if it meets the requriement describle |
4552 | // in ImmInstrInfo |
4553 | return true; |
4554 | } |
4555 | |
4556 | bool PPCInstrInfo::simplifyToLI(MachineInstr &MI, MachineInstr &DefMI, |
4557 | unsigned OpNoForForwarding, |
4558 | MachineInstr **KilledDef) const { |
4559 | if ((DefMI.getOpcode() != PPC::LI && DefMI.getOpcode() != PPC::LI8) || |
4560 | !DefMI.getOperand(i: 1).isImm()) |
4561 | return false; |
4562 | |
4563 | MachineFunction *MF = MI.getParent()->getParent(); |
4564 | MachineRegisterInfo *MRI = &MF->getRegInfo(); |
4565 | bool PostRA = !MRI->isSSA(); |
4566 | |
4567 | int64_t Immediate = DefMI.getOperand(i: 1).getImm(); |
4568 | // Sign-extend to 64-bits. |
4569 | int64_t SExtImm = SignExtend64<16>(x: Immediate); |
4570 | |
4571 | bool ReplaceWithLI = false; |
4572 | bool Is64BitLI = false; |
4573 | int64_t NewImm = 0; |
4574 | bool SetCR = false; |
4575 | unsigned Opc = MI.getOpcode(); |
4576 | switch (Opc) { |
4577 | default: |
4578 | return false; |
4579 | |
4580 | // FIXME: Any branches conditional on such a comparison can be made |
4581 | // unconditional. At this time, this happens too infrequently to be worth |
4582 | // the implementation effort, but if that ever changes, we could convert |
4583 | // such a pattern here. |
4584 | case PPC::CMPWI: |
4585 | case PPC::CMPLWI: |
4586 | case PPC::CMPDI: |
4587 | case PPC::CMPLDI: { |
4588 | // Doing this post-RA would require dataflow analysis to reliably find uses |
4589 | // of the CR register set by the compare. |
4590 | // No need to fixup killed/dead flag since this transformation is only valid |
4591 | // before RA. |
4592 | if (PostRA) |
4593 | return false; |
4594 | // If a compare-immediate is fed by an immediate and is itself an input of |
4595 | // an ISEL (the most common case) into a COPY of the correct register. |
4596 | bool Changed = false; |
4597 | Register DefReg = MI.getOperand(i: 0).getReg(); |
4598 | int64_t Comparand = MI.getOperand(i: 2).getImm(); |
4599 | int64_t SExtComparand = ((uint64_t)Comparand & ~0x7FFFuLL) != 0 |
4600 | ? (Comparand | 0xFFFFFFFFFFFF0000) |
4601 | : Comparand; |
4602 | |
4603 | for (auto &CompareUseMI : MRI->use_instructions(Reg: DefReg)) { |
4604 | unsigned UseOpc = CompareUseMI.getOpcode(); |
4605 | if (UseOpc != PPC::ISEL && UseOpc != PPC::ISEL8) |
4606 | continue; |
4607 | unsigned CRSubReg = CompareUseMI.getOperand(i: 3).getSubReg(); |
4608 | Register TrueReg = CompareUseMI.getOperand(i: 1).getReg(); |
4609 | Register FalseReg = CompareUseMI.getOperand(i: 2).getReg(); |
4610 | unsigned RegToCopy = |
4611 | selectReg(Imm1: SExtImm, Imm2: SExtComparand, CompareOpc: Opc, TrueReg, FalseReg, CRSubReg); |
4612 | if (RegToCopy == PPC::NoRegister) |
4613 | continue; |
4614 | // Can't use PPC::COPY to copy PPC::ZERO[8]. Convert it to LI[8] 0. |
4615 | if (RegToCopy == PPC::ZERO || RegToCopy == PPC::ZERO8) { |
4616 | CompareUseMI.setDesc(get(Opcode: UseOpc == PPC::ISEL8 ? PPC::LI8 : PPC::LI)); |
4617 | replaceInstrOperandWithImm(MI&: CompareUseMI, OpNo: 1, Imm: 0); |
4618 | CompareUseMI.removeOperand(OpNo: 3); |
4619 | CompareUseMI.removeOperand(OpNo: 2); |
4620 | continue; |
4621 | } |
4622 | LLVM_DEBUG( |
4623 | dbgs() << "Found LI -> CMPI -> ISEL, replacing with a copy.\n" ); |
4624 | LLVM_DEBUG(DefMI.dump(); MI.dump(); CompareUseMI.dump()); |
4625 | LLVM_DEBUG(dbgs() << "Is converted to:\n" ); |
4626 | // Convert to copy and remove unneeded operands. |
4627 | CompareUseMI.setDesc(get(Opcode: PPC::COPY)); |
4628 | CompareUseMI.removeOperand(OpNo: 3); |
4629 | CompareUseMI.removeOperand(OpNo: RegToCopy == TrueReg ? 2 : 1); |
4630 | CmpIselsConverted++; |
4631 | Changed = true; |
4632 | LLVM_DEBUG(CompareUseMI.dump()); |
4633 | } |
4634 | if (Changed) |
4635 | return true; |
4636 | // This may end up incremented multiple times since this function is called |
4637 | // during a fixed-point transformation, but it is only meant to indicate the |
4638 | // presence of this opportunity. |
4639 | MissedConvertibleImmediateInstrs++; |
4640 | return false; |
4641 | } |
4642 | |
4643 | // Immediate forms - may simply be convertable to an LI. |
4644 | case PPC::ADDI: |
4645 | case PPC::ADDI8: { |
4646 | // Does the sum fit in a 16-bit signed field? |
4647 | int64_t Addend = MI.getOperand(i: 2).getImm(); |
4648 | if (isInt<16>(x: Addend + SExtImm)) { |
4649 | ReplaceWithLI = true; |
4650 | Is64BitLI = Opc == PPC::ADDI8; |
4651 | NewImm = Addend + SExtImm; |
4652 | break; |
4653 | } |
4654 | return false; |
4655 | } |
4656 | case PPC::SUBFIC: |
4657 | case PPC::SUBFIC8: { |
4658 | // Only transform this if the CARRY implicit operand is dead. |
4659 | if (MI.getNumOperands() > 3 && !MI.getOperand(i: 3).isDead()) |
4660 | return false; |
4661 | int64_t Minuend = MI.getOperand(i: 2).getImm(); |
4662 | if (isInt<16>(x: Minuend - SExtImm)) { |
4663 | ReplaceWithLI = true; |
4664 | Is64BitLI = Opc == PPC::SUBFIC8; |
4665 | NewImm = Minuend - SExtImm; |
4666 | break; |
4667 | } |
4668 | return false; |
4669 | } |
4670 | case PPC::RLDICL: |
4671 | case PPC::RLDICL_rec: |
4672 | case PPC::RLDICL_32: |
4673 | case PPC::RLDICL_32_64: { |
4674 | // Use APInt's rotate function. |
4675 | int64_t SH = MI.getOperand(i: 2).getImm(); |
4676 | int64_t MB = MI.getOperand(i: 3).getImm(); |
4677 | APInt InVal((Opc == PPC::RLDICL || Opc == PPC::RLDICL_rec) ? 64 : 32, |
4678 | SExtImm, true); |
4679 | InVal = InVal.rotl(rotateAmt: SH); |
4680 | uint64_t Mask = MB == 0 ? -1LLU : (1LLU << (63 - MB + 1)) - 1; |
4681 | InVal &= Mask; |
4682 | // Can't replace negative values with an LI as that will sign-extend |
4683 | // and not clear the left bits. If we're setting the CR bit, we will use |
4684 | // ANDI_rec which won't sign extend, so that's safe. |
4685 | if (isUInt<15>(x: InVal.getSExtValue()) || |
4686 | (Opc == PPC::RLDICL_rec && isUInt<16>(x: InVal.getSExtValue()))) { |
4687 | ReplaceWithLI = true; |
4688 | Is64BitLI = Opc != PPC::RLDICL_32; |
4689 | NewImm = InVal.getSExtValue(); |
4690 | SetCR = Opc == PPC::RLDICL_rec; |
4691 | break; |
4692 | } |
4693 | return false; |
4694 | } |
4695 | case PPC::RLWINM: |
4696 | case PPC::RLWINM8: |
4697 | case PPC::RLWINM_rec: |
4698 | case PPC::RLWINM8_rec: { |
4699 | int64_t SH = MI.getOperand(i: 2).getImm(); |
4700 | int64_t MB = MI.getOperand(i: 3).getImm(); |
4701 | int64_t ME = MI.getOperand(i: 4).getImm(); |
4702 | APInt InVal(32, SExtImm, true); |
4703 | InVal = InVal.rotl(rotateAmt: SH); |
4704 | APInt Mask = APInt::getBitsSetWithWrap(numBits: 32, loBit: 32 - ME - 1, hiBit: 32 - MB); |
4705 | InVal &= Mask; |
4706 | // Can't replace negative values with an LI as that will sign-extend |
4707 | // and not clear the left bits. If we're setting the CR bit, we will use |
4708 | // ANDI_rec which won't sign extend, so that's safe. |
4709 | bool ValueFits = isUInt<15>(x: InVal.getSExtValue()); |
4710 | ValueFits |= ((Opc == PPC::RLWINM_rec || Opc == PPC::RLWINM8_rec) && |
4711 | isUInt<16>(x: InVal.getSExtValue())); |
4712 | if (ValueFits) { |
4713 | ReplaceWithLI = true; |
4714 | Is64BitLI = Opc == PPC::RLWINM8 || Opc == PPC::RLWINM8_rec; |
4715 | NewImm = InVal.getSExtValue(); |
4716 | SetCR = Opc == PPC::RLWINM_rec || Opc == PPC::RLWINM8_rec; |
4717 | break; |
4718 | } |
4719 | return false; |
4720 | } |
4721 | case PPC::ORI: |
4722 | case PPC::ORI8: |
4723 | case PPC::XORI: |
4724 | case PPC::XORI8: { |
4725 | int64_t LogicalImm = MI.getOperand(i: 2).getImm(); |
4726 | int64_t Result = 0; |
4727 | if (Opc == PPC::ORI || Opc == PPC::ORI8) |
4728 | Result = LogicalImm | SExtImm; |
4729 | else |
4730 | Result = LogicalImm ^ SExtImm; |
4731 | if (isInt<16>(x: Result)) { |
4732 | ReplaceWithLI = true; |
4733 | Is64BitLI = Opc == PPC::ORI8 || Opc == PPC::XORI8; |
4734 | NewImm = Result; |
4735 | break; |
4736 | } |
4737 | return false; |
4738 | } |
4739 | } |
4740 | |
4741 | if (ReplaceWithLI) { |
4742 | // We need to be careful with CR-setting instructions we're replacing. |
4743 | if (SetCR) { |
4744 | // We don't know anything about uses when we're out of SSA, so only |
4745 | // replace if the new immediate will be reproduced. |
4746 | bool ImmChanged = (SExtImm & NewImm) != NewImm; |
4747 | if (PostRA && ImmChanged) |
4748 | return false; |
4749 | |
4750 | if (!PostRA) { |
4751 | // If the defining load-immediate has no other uses, we can just replace |
4752 | // the immediate with the new immediate. |
4753 | if (MRI->hasOneUse(RegNo: DefMI.getOperand(i: 0).getReg())) |
4754 | DefMI.getOperand(i: 1).setImm(NewImm); |
4755 | |
4756 | // If we're not using the GPR result of the CR-setting instruction, we |
4757 | // just need to and with zero/non-zero depending on the new immediate. |
4758 | else if (MRI->use_empty(RegNo: MI.getOperand(i: 0).getReg())) { |
4759 | if (NewImm) { |
4760 | assert(Immediate && "Transformation converted zero to non-zero?" ); |
4761 | NewImm = Immediate; |
4762 | } |
4763 | } else if (ImmChanged) |
4764 | return false; |
4765 | } |
4766 | } |
4767 | |
4768 | LLVM_DEBUG(dbgs() << "Replacing constant instruction:\n" ); |
4769 | LLVM_DEBUG(MI.dump()); |
4770 | LLVM_DEBUG(dbgs() << "Fed by:\n" ); |
4771 | LLVM_DEBUG(DefMI.dump()); |
4772 | LoadImmediateInfo LII; |
4773 | LII.Imm = NewImm; |
4774 | LII.Is64Bit = Is64BitLI; |
4775 | LII.SetCR = SetCR; |
4776 | // If we're setting the CR, the original load-immediate must be kept (as an |
4777 | // operand to ANDI_rec/ANDI8_rec). |
4778 | if (KilledDef && SetCR) |
4779 | *KilledDef = nullptr; |
4780 | replaceInstrWithLI(MI, LII); |
4781 | |
4782 | if (PostRA) |
4783 | recomputeLivenessFlags(MBB&: *MI.getParent()); |
4784 | |
4785 | LLVM_DEBUG(dbgs() << "With:\n" ); |
4786 | LLVM_DEBUG(MI.dump()); |
4787 | return true; |
4788 | } |
4789 | return false; |
4790 | } |
4791 | |
4792 | bool PPCInstrInfo::transformToNewImmFormFedByAdd( |
4793 | MachineInstr &MI, MachineInstr &DefMI, unsigned OpNoForForwarding) const { |
4794 | MachineRegisterInfo *MRI = &MI.getParent()->getParent()->getRegInfo(); |
4795 | bool PostRA = !MRI->isSSA(); |
4796 | // FIXME: extend this to post-ra. Need to do some change in getForwardingDefMI |
4797 | // for post-ra. |
4798 | if (PostRA) |
4799 | return false; |
4800 | |
4801 | // Only handle load/store. |
4802 | if (!MI.mayLoadOrStore()) |
4803 | return false; |
4804 | |
4805 | unsigned XFormOpcode = RI.getMappedIdxOpcForImmOpc(ImmOpcode: MI.getOpcode()); |
4806 | |
4807 | assert((XFormOpcode != PPC::INSTRUCTION_LIST_END) && |
4808 | "MI must have x-form opcode" ); |
4809 | |
4810 | // get Imm Form info. |
4811 | ImmInstrInfo III; |
4812 | bool IsVFReg = MI.getOperand(i: 0).isReg() |
4813 | ? PPC::isVFRegister(Reg: MI.getOperand(i: 0).getReg()) |
4814 | : false; |
4815 | |
4816 | if (!instrHasImmForm(Opc: XFormOpcode, IsVFReg, III, PostRA)) |
4817 | return false; |
4818 | |
4819 | if (!III.IsSummingOperands) |
4820 | return false; |
4821 | |
4822 | if (OpNoForForwarding != III.OpNoForForwarding) |
4823 | return false; |
4824 | |
4825 | MachineOperand ImmOperandMI = MI.getOperand(i: III.ImmOpNo); |
4826 | if (!ImmOperandMI.isImm()) |
4827 | return false; |
4828 | |
4829 | // Check DefMI. |
4830 | MachineOperand *ImmMO = nullptr; |
4831 | MachineOperand *RegMO = nullptr; |
4832 | if (!isDefMIElgibleForForwarding(DefMI, III, ImmMO, RegMO)) |
4833 | return false; |
4834 | assert(ImmMO && RegMO && "Imm and Reg operand must have been set" ); |
4835 | |
4836 | // Check Imm. |
4837 | // Set ImmBase from imm instruction as base and get new Imm inside |
4838 | // isImmElgibleForForwarding. |
4839 | int64_t ImmBase = ImmOperandMI.getImm(); |
4840 | int64_t Imm = 0; |
4841 | if (!isImmElgibleForForwarding(ImmMO: *ImmMO, DefMI, III, Imm, BaseImm: ImmBase)) |
4842 | return false; |
4843 | |
4844 | // Do the transform |
4845 | LLVM_DEBUG(dbgs() << "Replacing existing reg+imm instruction:\n" ); |
4846 | LLVM_DEBUG(MI.dump()); |
4847 | LLVM_DEBUG(dbgs() << "Fed by:\n" ); |
4848 | LLVM_DEBUG(DefMI.dump()); |
4849 | |
4850 | MI.getOperand(i: III.OpNoForForwarding).setReg(RegMO->getReg()); |
4851 | MI.getOperand(i: III.ImmOpNo).setImm(Imm); |
4852 | |
4853 | LLVM_DEBUG(dbgs() << "With:\n" ); |
4854 | LLVM_DEBUG(MI.dump()); |
4855 | return true; |
4856 | } |
4857 | |
4858 | // If an X-Form instruction is fed by an add-immediate and one of its operands |
4859 | // is the literal zero, attempt to forward the source of the add-immediate to |
4860 | // the corresponding D-Form instruction with the displacement coming from |
4861 | // the immediate being added. |
4862 | bool PPCInstrInfo::transformToImmFormFedByAdd( |
4863 | MachineInstr &MI, const ImmInstrInfo &III, unsigned OpNoForForwarding, |
4864 | MachineInstr &DefMI, bool KillDefMI) const { |
4865 | // RegMO ImmMO |
4866 | // | | |
4867 | // x = addi reg, imm <----- DefMI |
4868 | // y = op 0 , x <----- MI |
4869 | // | |
4870 | // OpNoForForwarding |
4871 | // Check if the MI meet the requirement described in the III. |
4872 | if (!isUseMIElgibleForForwarding(MI, III, OpNoForForwarding)) |
4873 | return false; |
4874 | |
4875 | // Check if the DefMI meet the requirement |
4876 | // described in the III. If yes, set the ImmMO and RegMO accordingly. |
4877 | MachineOperand *ImmMO = nullptr; |
4878 | MachineOperand *RegMO = nullptr; |
4879 | if (!isDefMIElgibleForForwarding(DefMI, III, ImmMO, RegMO)) |
4880 | return false; |
4881 | assert(ImmMO && RegMO && "Imm and Reg operand must have been set" ); |
4882 | |
4883 | // As we get the Imm operand now, we need to check if the ImmMO meet |
4884 | // the requirement described in the III. If yes set the Imm. |
4885 | int64_t Imm = 0; |
4886 | if (!isImmElgibleForForwarding(ImmMO: *ImmMO, DefMI, III, Imm)) |
4887 | return false; |
4888 | |
4889 | bool IsFwdFeederRegKilled = false; |
4890 | bool SeenIntermediateUse = false; |
4891 | // Check if the RegMO can be forwarded to MI. |
4892 | if (!isRegElgibleForForwarding(RegMO: *RegMO, DefMI, MI, KillDefMI, |
4893 | IsFwdFeederRegKilled, SeenIntermediateUse)) |
4894 | return false; |
4895 | |
4896 | MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo(); |
4897 | bool PostRA = !MRI.isSSA(); |
4898 | |
4899 | // We know that, the MI and DefMI both meet the pattern, and |
4900 | // the Imm also meet the requirement with the new Imm-form. |
4901 | // It is safe to do the transformation now. |
4902 | LLVM_DEBUG(dbgs() << "Replacing indexed instruction:\n" ); |
4903 | LLVM_DEBUG(MI.dump()); |
4904 | LLVM_DEBUG(dbgs() << "Fed by:\n" ); |
4905 | LLVM_DEBUG(DefMI.dump()); |
4906 | |
4907 | // Update the base reg first. |
4908 | MI.getOperand(i: III.OpNoForForwarding).ChangeToRegister(Reg: RegMO->getReg(), |
4909 | isDef: false, isImp: false, |
4910 | isKill: RegMO->isKill()); |
4911 | |
4912 | // Then, update the imm. |
4913 | if (ImmMO->isImm()) { |
4914 | // If the ImmMO is Imm, change the operand that has ZERO to that Imm |
4915 | // directly. |
4916 | replaceInstrOperandWithImm(MI, OpNo: III.ZeroIsSpecialOrig, Imm); |
4917 | } |
4918 | else { |
4919 | // Otherwise, it is Constant Pool Index(CPI) or Global, |
4920 | // which is relocation in fact. We need to replace the special zero |
4921 | // register with ImmMO. |
4922 | // Before that, we need to fixup the target flags for imm. |
4923 | // For some reason, we miss to set the flag for the ImmMO if it is CPI. |
4924 | if (DefMI.getOpcode() == PPC::ADDItocL8) |
4925 | ImmMO->setTargetFlags(PPCII::MO_TOC_LO); |
4926 | |
4927 | // MI didn't have the interface such as MI.setOperand(i) though |
4928 | // it has MI.getOperand(i). To repalce the ZERO MachineOperand with |
4929 | // ImmMO, we need to remove ZERO operand and all the operands behind it, |
4930 | // and, add the ImmMO, then, move back all the operands behind ZERO. |
4931 | SmallVector<MachineOperand, 2> MOps; |
4932 | for (unsigned i = MI.getNumOperands() - 1; i >= III.ZeroIsSpecialOrig; i--) { |
4933 | MOps.push_back(Elt: MI.getOperand(i)); |
4934 | MI.removeOperand(OpNo: i); |
4935 | } |
4936 | |
4937 | // Remove the last MO in the list, which is ZERO operand in fact. |
4938 | MOps.pop_back(); |
4939 | // Add the imm operand. |
4940 | MI.addOperand(Op: *ImmMO); |
4941 | // Now add the rest back. |
4942 | for (auto &MO : MOps) |
4943 | MI.addOperand(Op: MO); |
4944 | } |
4945 | |
4946 | // Update the opcode. |
4947 | MI.setDesc(get(Opcode: III.ImmOpcode)); |
4948 | |
4949 | if (PostRA) |
4950 | recomputeLivenessFlags(MBB&: *MI.getParent()); |
4951 | LLVM_DEBUG(dbgs() << "With:\n" ); |
4952 | LLVM_DEBUG(MI.dump()); |
4953 | |
4954 | return true; |
4955 | } |
4956 | |
4957 | bool PPCInstrInfo::transformToImmFormFedByLI(MachineInstr &MI, |
4958 | const ImmInstrInfo &III, |
4959 | unsigned ConstantOpNo, |
4960 | MachineInstr &DefMI) const { |
4961 | // DefMI must be LI or LI8. |
4962 | if ((DefMI.getOpcode() != PPC::LI && DefMI.getOpcode() != PPC::LI8) || |
4963 | !DefMI.getOperand(i: 1).isImm()) |
4964 | return false; |
4965 | |
4966 | // Get Imm operand and Sign-extend to 64-bits. |
4967 | int64_t Imm = SignExtend64<16>(x: DefMI.getOperand(i: 1).getImm()); |
4968 | |
4969 | MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo(); |
4970 | bool PostRA = !MRI.isSSA(); |
4971 | // Exit early if we can't convert this. |
4972 | if ((ConstantOpNo != III.OpNoForForwarding) && !III.IsCommutative) |
4973 | return false; |
4974 | if (Imm % III.ImmMustBeMultipleOf) |
4975 | return false; |
4976 | if (III.TruncateImmTo) |
4977 | Imm &= ((1 << III.TruncateImmTo) - 1); |
4978 | if (III.SignedImm) { |
4979 | APInt ActualValue(64, Imm, true); |
4980 | if (!ActualValue.isSignedIntN(N: III.ImmWidth)) |
4981 | return false; |
4982 | } else { |
4983 | uint64_t UnsignedMax = (1 << III.ImmWidth) - 1; |
4984 | if ((uint64_t)Imm > UnsignedMax) |
4985 | return false; |
4986 | } |
4987 | |
4988 | // If we're post-RA, the instructions don't agree on whether register zero is |
4989 | // special, we can transform this as long as the register operand that will |
4990 | // end up in the location where zero is special isn't R0. |
4991 | if (PostRA && III.ZeroIsSpecialOrig != III.ZeroIsSpecialNew) { |
4992 | unsigned PosForOrigZero = III.ZeroIsSpecialOrig ? III.ZeroIsSpecialOrig : |
4993 | III.ZeroIsSpecialNew + 1; |
4994 | Register OrigZeroReg = MI.getOperand(i: PosForOrigZero).getReg(); |
4995 | Register NewZeroReg = MI.getOperand(i: III.ZeroIsSpecialNew).getReg(); |
4996 | // If R0 is in the operand where zero is special for the new instruction, |
4997 | // it is unsafe to transform if the constant operand isn't that operand. |
4998 | if ((NewZeroReg == PPC::R0 || NewZeroReg == PPC::X0) && |
4999 | ConstantOpNo != III.ZeroIsSpecialNew) |
5000 | return false; |
5001 | if ((OrigZeroReg == PPC::R0 || OrigZeroReg == PPC::X0) && |
5002 | ConstantOpNo != PosForOrigZero) |
5003 | return false; |
5004 | } |
5005 | |
5006 | unsigned Opc = MI.getOpcode(); |
5007 | bool SpecialShift32 = Opc == PPC::SLW || Opc == PPC::SLW_rec || |
5008 | Opc == PPC::SRW || Opc == PPC::SRW_rec || |
5009 | Opc == PPC::SLW8 || Opc == PPC::SLW8_rec || |
5010 | Opc == PPC::SRW8 || Opc == PPC::SRW8_rec; |
5011 | bool SpecialShift64 = Opc == PPC::SLD || Opc == PPC::SLD_rec || |
5012 | Opc == PPC::SRD || Opc == PPC::SRD_rec; |
5013 | bool SetCR = Opc == PPC::SLW_rec || Opc == PPC::SRW_rec || |
5014 | Opc == PPC::SLD_rec || Opc == PPC::SRD_rec; |
5015 | bool RightShift = Opc == PPC::SRW || Opc == PPC::SRW_rec || Opc == PPC::SRD || |
5016 | Opc == PPC::SRD_rec; |
5017 | |
5018 | LLVM_DEBUG(dbgs() << "Replacing reg+reg instruction: " ); |
5019 | LLVM_DEBUG(MI.dump()); |
5020 | LLVM_DEBUG(dbgs() << "Fed by load-immediate: " ); |
5021 | LLVM_DEBUG(DefMI.dump()); |
5022 | MI.setDesc(get(Opcode: III.ImmOpcode)); |
5023 | if (ConstantOpNo == III.OpNoForForwarding) { |
5024 | // Converting shifts to immediate form is a bit tricky since they may do |
5025 | // one of three things: |
5026 | // 1. If the shift amount is between OpSize and 2*OpSize, the result is zero |
5027 | // 2. If the shift amount is zero, the result is unchanged (save for maybe |
5028 | // setting CR0) |
5029 | // 3. If the shift amount is in [1, OpSize), it's just a shift |
5030 | if (SpecialShift32 || SpecialShift64) { |
5031 | LoadImmediateInfo LII; |
5032 | LII.Imm = 0; |
5033 | LII.SetCR = SetCR; |
5034 | LII.Is64Bit = SpecialShift64; |
5035 | uint64_t ShAmt = Imm & (SpecialShift32 ? 0x1F : 0x3F); |
5036 | if (Imm & (SpecialShift32 ? 0x20 : 0x40)) |
5037 | replaceInstrWithLI(MI, LII); |
5038 | // Shifts by zero don't change the value. If we don't need to set CR0, |
5039 | // just convert this to a COPY. Can't do this post-RA since we've already |
5040 | // cleaned up the copies. |
5041 | else if (!SetCR && ShAmt == 0 && !PostRA) { |
5042 | MI.removeOperand(OpNo: 2); |
5043 | MI.setDesc(get(Opcode: PPC::COPY)); |
5044 | } else { |
5045 | // The 32 bit and 64 bit instructions are quite different. |
5046 | if (SpecialShift32) { |
5047 | // Left shifts use (N, 0, 31-N). |
5048 | // Right shifts use (32-N, N, 31) if 0 < N < 32. |
5049 | // use (0, 0, 31) if N == 0. |
5050 | uint64_t SH = ShAmt == 0 ? 0 : RightShift ? 32 - ShAmt : ShAmt; |
5051 | uint64_t MB = RightShift ? ShAmt : 0; |
5052 | uint64_t ME = RightShift ? 31 : 31 - ShAmt; |
5053 | replaceInstrOperandWithImm(MI, OpNo: III.OpNoForForwarding, Imm: SH); |
5054 | MachineInstrBuilder(*MI.getParent()->getParent(), MI).addImm(Val: MB) |
5055 | .addImm(Val: ME); |
5056 | } else { |
5057 | // Left shifts use (N, 63-N). |
5058 | // Right shifts use (64-N, N) if 0 < N < 64. |
5059 | // use (0, 0) if N == 0. |
5060 | uint64_t SH = ShAmt == 0 ? 0 : RightShift ? 64 - ShAmt : ShAmt; |
5061 | uint64_t ME = RightShift ? ShAmt : 63 - ShAmt; |
5062 | replaceInstrOperandWithImm(MI, OpNo: III.OpNoForForwarding, Imm: SH); |
5063 | MachineInstrBuilder(*MI.getParent()->getParent(), MI).addImm(Val: ME); |
5064 | } |
5065 | } |
5066 | } else |
5067 | replaceInstrOperandWithImm(MI, OpNo: ConstantOpNo, Imm); |
5068 | } |
5069 | // Convert commutative instructions (switch the operands and convert the |
5070 | // desired one to an immediate. |
5071 | else if (III.IsCommutative) { |
5072 | replaceInstrOperandWithImm(MI, OpNo: ConstantOpNo, Imm); |
5073 | swapMIOperands(MI, Op1: ConstantOpNo, Op2: III.OpNoForForwarding); |
5074 | } else |
5075 | llvm_unreachable("Should have exited early!" ); |
5076 | |
5077 | // For instructions for which the constant register replaces a different |
5078 | // operand than where the immediate goes, we need to swap them. |
5079 | if (III.OpNoForForwarding != III.ImmOpNo) |
5080 | swapMIOperands(MI, Op1: III.OpNoForForwarding, Op2: III.ImmOpNo); |
5081 | |
5082 | // If the special R0/X0 register index are different for original instruction |
5083 | // and new instruction, we need to fix up the register class in new |
5084 | // instruction. |
5085 | if (!PostRA && III.ZeroIsSpecialOrig != III.ZeroIsSpecialNew) { |
5086 | if (III.ZeroIsSpecialNew) { |
5087 | // If operand at III.ZeroIsSpecialNew is physical reg(eg: ZERO/ZERO8), no |
5088 | // need to fix up register class. |
5089 | Register RegToModify = MI.getOperand(i: III.ZeroIsSpecialNew).getReg(); |
5090 | if (RegToModify.isVirtual()) { |
5091 | const TargetRegisterClass *NewRC = |
5092 | MRI.getRegClass(Reg: RegToModify)->hasSuperClassEq(RC: &PPC::GPRCRegClass) ? |
5093 | &PPC::GPRC_and_GPRC_NOR0RegClass : &PPC::G8RC_and_G8RC_NOX0RegClass; |
5094 | MRI.setRegClass(Reg: RegToModify, RC: NewRC); |
5095 | } |
5096 | } |
5097 | } |
5098 | |
5099 | if (PostRA) |
5100 | recomputeLivenessFlags(MBB&: *MI.getParent()); |
5101 | |
5102 | LLVM_DEBUG(dbgs() << "With: " ); |
5103 | LLVM_DEBUG(MI.dump()); |
5104 | LLVM_DEBUG(dbgs() << "\n" ); |
5105 | return true; |
5106 | } |
5107 | |
5108 | const TargetRegisterClass * |
5109 | PPCInstrInfo::updatedRC(const TargetRegisterClass *RC) const { |
5110 | if (Subtarget.hasVSX() && RC == &PPC::VRRCRegClass) |
5111 | return &PPC::VSRCRegClass; |
5112 | return RC; |
5113 | } |
5114 | |
5115 | int PPCInstrInfo::getRecordFormOpcode(unsigned Opcode) { |
5116 | return PPC::getRecordFormOpcode(Opcode); |
5117 | } |
5118 | |
5119 | static bool isOpZeroOfSubwordPreincLoad(int Opcode) { |
5120 | return (Opcode == PPC::LBZU || Opcode == PPC::LBZUX || Opcode == PPC::LBZU8 || |
5121 | Opcode == PPC::LBZUX8 || Opcode == PPC::LHZU || |
5122 | Opcode == PPC::LHZUX || Opcode == PPC::LHZU8 || |
5123 | Opcode == PPC::LHZUX8); |
5124 | } |
5125 | |
5126 | // This function checks for sign extension from 32 bits to 64 bits. |
5127 | static bool definedBySignExtendingOp(const unsigned Reg, |
5128 | const MachineRegisterInfo *MRI) { |
5129 | if (!Register::isVirtualRegister(Reg)) |
5130 | return false; |
5131 | |
5132 | MachineInstr *MI = MRI->getVRegDef(Reg); |
5133 | if (!MI) |
5134 | return false; |
5135 | |
5136 | int Opcode = MI->getOpcode(); |
5137 | const PPCInstrInfo *TII = |
5138 | MI->getMF()->getSubtarget<PPCSubtarget>().getInstrInfo(); |
5139 | if (TII->isSExt32To64(Opcode)) |
5140 | return true; |
5141 | |
5142 | // The first def of LBZU/LHZU is sign extended. |
5143 | if (isOpZeroOfSubwordPreincLoad(Opcode) && MI->getOperand(i: 0).getReg() == Reg) |
5144 | return true; |
5145 | |
5146 | // RLDICL generates sign-extended output if it clears at least |
5147 | // 33 bits from the left (MSB). |
5148 | if (Opcode == PPC::RLDICL && MI->getOperand(i: 3).getImm() >= 33) |
5149 | return true; |
5150 | |
5151 | // If at least one bit from left in a lower word is masked out, |
5152 | // all of 0 to 32-th bits of the output are cleared. |
5153 | // Hence the output is already sign extended. |
5154 | if ((Opcode == PPC::RLWINM || Opcode == PPC::RLWINM_rec || |
5155 | Opcode == PPC::RLWNM || Opcode == PPC::RLWNM_rec) && |
5156 | MI->getOperand(i: 3).getImm() > 0 && |
5157 | MI->getOperand(i: 3).getImm() <= MI->getOperand(i: 4).getImm()) |
5158 | return true; |
5159 | |
5160 | // If the most significant bit of immediate in ANDIS is zero, |
5161 | // all of 0 to 32-th bits are cleared. |
5162 | if (Opcode == PPC::ANDIS_rec || Opcode == PPC::ANDIS8_rec) { |
5163 | uint16_t Imm = MI->getOperand(i: 2).getImm(); |
5164 | if ((Imm & 0x8000) == 0) |
5165 | return true; |
5166 | } |
5167 | |
5168 | return false; |
5169 | } |
5170 | |
5171 | // This function checks the machine instruction that defines the input register |
5172 | // Reg. If that machine instruction always outputs a value that has only zeros |
5173 | // in the higher 32 bits then this function will return true. |
5174 | static bool definedByZeroExtendingOp(const unsigned Reg, |
5175 | const MachineRegisterInfo *MRI) { |
5176 | if (!Register::isVirtualRegister(Reg)) |
5177 | return false; |
5178 | |
5179 | MachineInstr *MI = MRI->getVRegDef(Reg); |
5180 | if (!MI) |
5181 | return false; |
5182 | |
5183 | int Opcode = MI->getOpcode(); |
5184 | const PPCInstrInfo *TII = |
5185 | MI->getMF()->getSubtarget<PPCSubtarget>().getInstrInfo(); |
5186 | if (TII->isZExt32To64(Opcode)) |
5187 | return true; |
5188 | |
5189 | // The first def of LBZU/LHZU/LWZU are zero extended. |
5190 | if ((isOpZeroOfSubwordPreincLoad(Opcode) || Opcode == PPC::LWZU || |
5191 | Opcode == PPC::LWZUX || Opcode == PPC::LWZU8 || Opcode == PPC::LWZUX8) && |
5192 | MI->getOperand(i: 0).getReg() == Reg) |
5193 | return true; |
5194 | |
5195 | // The 16-bit immediate is sign-extended in li/lis. |
5196 | // If the most significant bit is zero, all higher bits are zero. |
5197 | if (Opcode == PPC::LI || Opcode == PPC::LI8 || |
5198 | Opcode == PPC::LIS || Opcode == PPC::LIS8) { |
5199 | int64_t Imm = MI->getOperand(i: 1).getImm(); |
5200 | if (((uint64_t)Imm & ~0x7FFFuLL) == 0) |
5201 | return true; |
5202 | } |
5203 | |
5204 | // We have some variations of rotate-and-mask instructions |
5205 | // that clear higher 32-bits. |
5206 | if ((Opcode == PPC::RLDICL || Opcode == PPC::RLDICL_rec || |
5207 | Opcode == PPC::RLDCL || Opcode == PPC::RLDCL_rec || |
5208 | Opcode == PPC::RLDICL_32_64) && |
5209 | MI->getOperand(i: 3).getImm() >= 32) |
5210 | return true; |
5211 | |
5212 | if ((Opcode == PPC::RLDIC || Opcode == PPC::RLDIC_rec) && |
5213 | MI->getOperand(i: 3).getImm() >= 32 && |
5214 | MI->getOperand(i: 3).getImm() <= 63 - MI->getOperand(i: 2).getImm()) |
5215 | return true; |
5216 | |
5217 | if ((Opcode == PPC::RLWINM || Opcode == PPC::RLWINM_rec || |
5218 | Opcode == PPC::RLWNM || Opcode == PPC::RLWNM_rec || |
5219 | Opcode == PPC::RLWINM8 || Opcode == PPC::RLWNM8) && |
5220 | MI->getOperand(i: 3).getImm() <= MI->getOperand(i: 4).getImm()) |
5221 | return true; |
5222 | |
5223 | return false; |
5224 | } |
5225 | |
5226 | // This function returns true if the input MachineInstr is a TOC save |
5227 | // instruction. |
5228 | bool PPCInstrInfo::isTOCSaveMI(const MachineInstr &MI) const { |
5229 | if (!MI.getOperand(i: 1).isImm() || !MI.getOperand(i: 2).isReg()) |
5230 | return false; |
5231 | unsigned TOCSaveOffset = Subtarget.getFrameLowering()->getTOCSaveOffset(); |
5232 | unsigned StackOffset = MI.getOperand(i: 1).getImm(); |
5233 | Register StackReg = MI.getOperand(i: 2).getReg(); |
5234 | Register SPReg = Subtarget.isPPC64() ? PPC::X1 : PPC::R1; |
5235 | if (StackReg == SPReg && StackOffset == TOCSaveOffset) |
5236 | return true; |
5237 | |
5238 | return false; |
5239 | } |
5240 | |
5241 | // We limit the max depth to track incoming values of PHIs or binary ops |
5242 | // (e.g. AND) to avoid excessive cost. |
5243 | const unsigned MAX_BINOP_DEPTH = 1; |
5244 | // The isSignOrZeroExtended function is recursive. The parameter BinOpDepth |
5245 | // does not count all of the recursions. The parameter BinOpDepth is incremented |
5246 | // only when isSignOrZeroExtended calls itself more than once. This is done to |
5247 | // prevent expontential recursion. There is no parameter to track linear |
5248 | // recursion. |
5249 | std::pair<bool, bool> |
5250 | PPCInstrInfo::isSignOrZeroExtended(const unsigned Reg, |
5251 | const unsigned BinOpDepth, |
5252 | const MachineRegisterInfo *MRI) const { |
5253 | if (!Register::isVirtualRegister(Reg)) |
5254 | return std::pair<bool, bool>(false, false); |
5255 | |
5256 | MachineInstr *MI = MRI->getVRegDef(Reg); |
5257 | if (!MI) |
5258 | return std::pair<bool, bool>(false, false); |
5259 | |
5260 | bool IsSExt = definedBySignExtendingOp(Reg, MRI); |
5261 | bool IsZExt = definedByZeroExtendingOp(Reg, MRI); |
5262 | |
5263 | // If we know the instruction always returns sign- and zero-extended result, |
5264 | // return here. |
5265 | if (IsSExt && IsZExt) |
5266 | return std::pair<bool, bool>(IsSExt, IsZExt); |
5267 | |
5268 | switch (MI->getOpcode()) { |
5269 | case PPC::COPY: { |
5270 | Register SrcReg = MI->getOperand(i: 1).getReg(); |
5271 | |
5272 | // In both ELFv1 and v2 ABI, method parameters and the return value |
5273 | // are sign- or zero-extended. |
5274 | const MachineFunction *MF = MI->getMF(); |
5275 | |
5276 | if (!MF->getSubtarget<PPCSubtarget>().isSVR4ABI()) { |
5277 | // If this is a copy from another register, we recursively check source. |
5278 | auto SrcExt = isSignOrZeroExtended(Reg: SrcReg, BinOpDepth, MRI); |
5279 | return std::pair<bool, bool>(SrcExt.first || IsSExt, |
5280 | SrcExt.second || IsZExt); |
5281 | } |
5282 | |
5283 | // From here on everything is SVR4ABI |
5284 | const PPCFunctionInfo *FuncInfo = MF->getInfo<PPCFunctionInfo>(); |
5285 | // We check the ZExt/SExt flags for a method parameter. |
5286 | if (MI->getParent()->getBasicBlock() == |
5287 | &MF->getFunction().getEntryBlock()) { |
5288 | Register VReg = MI->getOperand(i: 0).getReg(); |
5289 | if (MF->getRegInfo().isLiveIn(Reg: VReg)) { |
5290 | IsSExt |= FuncInfo->isLiveInSExt(VReg); |
5291 | IsZExt |= FuncInfo->isLiveInZExt(VReg); |
5292 | return std::pair<bool, bool>(IsSExt, IsZExt); |
5293 | } |
5294 | } |
5295 | |
5296 | if (SrcReg != PPC::X3) { |
5297 | // If this is a copy from another register, we recursively check source. |
5298 | auto SrcExt = isSignOrZeroExtended(Reg: SrcReg, BinOpDepth, MRI); |
5299 | return std::pair<bool, bool>(SrcExt.first || IsSExt, |
5300 | SrcExt.second || IsZExt); |
5301 | } |
5302 | |
5303 | // For a method return value, we check the ZExt/SExt flags in attribute. |
5304 | // We assume the following code sequence for method call. |
5305 | // ADJCALLSTACKDOWN 32, implicit dead %r1, implicit %r1 |
5306 | // BL8_NOP @func,... |
5307 | // ADJCALLSTACKUP 32, 0, implicit dead %r1, implicit %r1 |
5308 | // %5 = COPY %x3; G8RC:%5 |
5309 | const MachineBasicBlock *MBB = MI->getParent(); |
5310 | std::pair<bool, bool> IsExtendPair = std::pair<bool, bool>(IsSExt, IsZExt); |
5311 | MachineBasicBlock::const_instr_iterator II = |
5312 | MachineBasicBlock::const_instr_iterator(MI); |
5313 | if (II == MBB->instr_begin() || (--II)->getOpcode() != PPC::ADJCALLSTACKUP) |
5314 | return IsExtendPair; |
5315 | |
5316 | const MachineInstr &CallMI = *(--II); |
5317 | if (!CallMI.isCall() || !CallMI.getOperand(i: 0).isGlobal()) |
5318 | return IsExtendPair; |
5319 | |
5320 | const Function *CalleeFn = |
5321 | dyn_cast_if_present<Function>(Val: CallMI.getOperand(i: 0).getGlobal()); |
5322 | if (!CalleeFn) |
5323 | return IsExtendPair; |
5324 | const IntegerType *IntTy = dyn_cast<IntegerType>(Val: CalleeFn->getReturnType()); |
5325 | if (IntTy && IntTy->getBitWidth() <= 32) { |
5326 | const AttributeSet &Attrs = CalleeFn->getAttributes().getRetAttrs(); |
5327 | IsSExt |= Attrs.hasAttribute(Kind: Attribute::SExt); |
5328 | IsZExt |= Attrs.hasAttribute(Kind: Attribute::ZExt); |
5329 | return std::pair<bool, bool>(IsSExt, IsZExt); |
5330 | } |
5331 | |
5332 | return IsExtendPair; |
5333 | } |
5334 | |
5335 | // OR, XOR with 16-bit immediate does not change the upper 48 bits. |
5336 | // So, we track the operand register as we do for register copy. |
5337 | case PPC::ORI: |
5338 | case PPC::XORI: |
5339 | case PPC::ORI8: |
5340 | case PPC::XORI8: { |
5341 | Register SrcReg = MI->getOperand(i: 1).getReg(); |
5342 | auto SrcExt = isSignOrZeroExtended(Reg: SrcReg, BinOpDepth, MRI); |
5343 | return std::pair<bool, bool>(SrcExt.first || IsSExt, |
5344 | SrcExt.second || IsZExt); |
5345 | } |
5346 | |
5347 | // OR, XOR with shifted 16-bit immediate does not change the upper |
5348 | // 32 bits. So, we track the operand register for zero extension. |
5349 | // For sign extension when the MSB of the immediate is zero, we also |
5350 | // track the operand register since the upper 33 bits are unchanged. |
5351 | case PPC::ORIS: |
5352 | case PPC::XORIS: |
5353 | case PPC::ORIS8: |
5354 | case PPC::XORIS8: { |
5355 | Register SrcReg = MI->getOperand(i: 1).getReg(); |
5356 | auto SrcExt = isSignOrZeroExtended(Reg: SrcReg, BinOpDepth, MRI); |
5357 | uint16_t Imm = MI->getOperand(i: 2).getImm(); |
5358 | if (Imm & 0x8000) |
5359 | return std::pair<bool, bool>(false, SrcExt.second || IsZExt); |
5360 | else |
5361 | return std::pair<bool, bool>(SrcExt.first || IsSExt, |
5362 | SrcExt.second || IsZExt); |
5363 | } |
5364 | |
5365 | // If all incoming values are sign-/zero-extended, |
5366 | // the output of OR, ISEL or PHI is also sign-/zero-extended. |
5367 | case PPC::OR: |
5368 | case PPC::OR8: |
5369 | case PPC::ISEL: |
5370 | case PPC::PHI: { |
5371 | if (BinOpDepth >= MAX_BINOP_DEPTH) |
5372 | return std::pair<bool, bool>(false, false); |
5373 | |
5374 | // The input registers for PHI are operand 1, 3, ... |
5375 | // The input registers for others are operand 1 and 2. |
5376 | unsigned OperandEnd = 3, OperandStride = 1; |
5377 | if (MI->getOpcode() == PPC::PHI) { |
5378 | OperandEnd = MI->getNumOperands(); |
5379 | OperandStride = 2; |
5380 | } |
5381 | |
5382 | IsSExt = true; |
5383 | IsZExt = true; |
5384 | for (unsigned I = 1; I != OperandEnd; I += OperandStride) { |
5385 | if (!MI->getOperand(i: I).isReg()) |
5386 | return std::pair<bool, bool>(false, false); |
5387 | |
5388 | Register SrcReg = MI->getOperand(i: I).getReg(); |
5389 | auto SrcExt = isSignOrZeroExtended(Reg: SrcReg, BinOpDepth: BinOpDepth + 1, MRI); |
5390 | IsSExt &= SrcExt.first; |
5391 | IsZExt &= SrcExt.second; |
5392 | } |
5393 | return std::pair<bool, bool>(IsSExt, IsZExt); |
5394 | } |
5395 | |
5396 | // If at least one of the incoming values of an AND is zero extended |
5397 | // then the output is also zero-extended. If both of the incoming values |
5398 | // are sign-extended then the output is also sign extended. |
5399 | case PPC::AND: |
5400 | case PPC::AND8: { |
5401 | if (BinOpDepth >= MAX_BINOP_DEPTH) |
5402 | return std::pair<bool, bool>(false, false); |
5403 | |
5404 | Register SrcReg1 = MI->getOperand(i: 1).getReg(); |
5405 | Register SrcReg2 = MI->getOperand(i: 2).getReg(); |
5406 | auto Src1Ext = isSignOrZeroExtended(Reg: SrcReg1, BinOpDepth: BinOpDepth + 1, MRI); |
5407 | auto Src2Ext = isSignOrZeroExtended(Reg: SrcReg2, BinOpDepth: BinOpDepth + 1, MRI); |
5408 | return std::pair<bool, bool>(Src1Ext.first && Src2Ext.first, |
5409 | Src1Ext.second || Src2Ext.second); |
5410 | } |
5411 | |
5412 | default: |
5413 | break; |
5414 | } |
5415 | return std::pair<bool, bool>(IsSExt, IsZExt); |
5416 | } |
5417 | |
5418 | bool PPCInstrInfo::isBDNZ(unsigned Opcode) const { |
5419 | return (Opcode == (Subtarget.isPPC64() ? PPC::BDNZ8 : PPC::BDNZ)); |
5420 | } |
5421 | |
5422 | namespace { |
5423 | class PPCPipelinerLoopInfo : public TargetInstrInfo::PipelinerLoopInfo { |
5424 | MachineInstr *Loop, *EndLoop, *LoopCount; |
5425 | MachineFunction *MF; |
5426 | const TargetInstrInfo *TII; |
5427 | int64_t TripCount; |
5428 | |
5429 | public: |
5430 | PPCPipelinerLoopInfo(MachineInstr *Loop, MachineInstr *EndLoop, |
5431 | MachineInstr *LoopCount) |
5432 | : Loop(Loop), EndLoop(EndLoop), LoopCount(LoopCount), |
5433 | MF(Loop->getParent()->getParent()), |
5434 | TII(MF->getSubtarget().getInstrInfo()) { |
5435 | // Inspect the Loop instruction up-front, as it may be deleted when we call |
5436 | // createTripCountGreaterCondition. |
5437 | if (LoopCount->getOpcode() == PPC::LI8 || LoopCount->getOpcode() == PPC::LI) |
5438 | TripCount = LoopCount->getOperand(i: 1).getImm(); |
5439 | else |
5440 | TripCount = -1; |
5441 | } |
5442 | |
5443 | bool shouldIgnoreForPipelining(const MachineInstr *MI) const override { |
5444 | // Only ignore the terminator. |
5445 | return MI == EndLoop; |
5446 | } |
5447 | |
5448 | std::optional<bool> createTripCountGreaterCondition( |
5449 | int TC, MachineBasicBlock &MBB, |
5450 | SmallVectorImpl<MachineOperand> &Cond) override { |
5451 | if (TripCount == -1) { |
5452 | // Since BDZ/BDZ8 that we will insert will also decrease the ctr by 1, |
5453 | // so we don't need to generate any thing here. |
5454 | Cond.push_back(Elt: MachineOperand::CreateImm(Val: 0)); |
5455 | Cond.push_back(Elt: MachineOperand::CreateReg( |
5456 | Reg: MF->getSubtarget<PPCSubtarget>().isPPC64() ? PPC::CTR8 : PPC::CTR, |
5457 | isDef: true)); |
5458 | return {}; |
5459 | } |
5460 | |
5461 | return TripCount > TC; |
5462 | } |
5463 | |
5464 | void (MachineBasicBlock *) override { |
5465 | // Do nothing. We want the LOOP setup instruction to stay in the *old* |
5466 | // preheader, so we can use BDZ in the prologs to adapt the loop trip count. |
5467 | } |
5468 | |
5469 | void adjustTripCount(int TripCountAdjust) override { |
5470 | // If the loop trip count is a compile-time value, then just change the |
5471 | // value. |
5472 | if (LoopCount->getOpcode() == PPC::LI8 || |
5473 | LoopCount->getOpcode() == PPC::LI) { |
5474 | int64_t TripCount = LoopCount->getOperand(i: 1).getImm() + TripCountAdjust; |
5475 | LoopCount->getOperand(i: 1).setImm(TripCount); |
5476 | return; |
5477 | } |
5478 | |
5479 | // Since BDZ/BDZ8 that we will insert will also decrease the ctr by 1, |
5480 | // so we don't need to generate any thing here. |
5481 | } |
5482 | |
5483 | void disposed() override { |
5484 | Loop->eraseFromParent(); |
5485 | // Ensure the loop setup instruction is deleted too. |
5486 | LoopCount->eraseFromParent(); |
5487 | } |
5488 | }; |
5489 | } // namespace |
5490 | |
5491 | std::unique_ptr<TargetInstrInfo::PipelinerLoopInfo> |
5492 | PPCInstrInfo::analyzeLoopForPipelining(MachineBasicBlock *LoopBB) const { |
5493 | // We really "analyze" only hardware loops right now. |
5494 | MachineBasicBlock::iterator I = LoopBB->getFirstTerminator(); |
5495 | MachineBasicBlock * = *LoopBB->pred_begin(); |
5496 | if (Preheader == LoopBB) |
5497 | Preheader = *std::next(x: LoopBB->pred_begin()); |
5498 | MachineFunction *MF = Preheader->getParent(); |
5499 | |
5500 | if (I != LoopBB->end() && isBDNZ(Opcode: I->getOpcode())) { |
5501 | SmallPtrSet<MachineBasicBlock *, 8> Visited; |
5502 | if (MachineInstr *LoopInst = findLoopInstr(PreHeader&: *Preheader, Visited)) { |
5503 | Register LoopCountReg = LoopInst->getOperand(i: 0).getReg(); |
5504 | MachineRegisterInfo &MRI = MF->getRegInfo(); |
5505 | MachineInstr *LoopCount = MRI.getUniqueVRegDef(Reg: LoopCountReg); |
5506 | return std::make_unique<PPCPipelinerLoopInfo>(args&: LoopInst, args: &*I, args&: LoopCount); |
5507 | } |
5508 | } |
5509 | return nullptr; |
5510 | } |
5511 | |
5512 | MachineInstr *PPCInstrInfo::findLoopInstr( |
5513 | MachineBasicBlock &, |
5514 | SmallPtrSet<MachineBasicBlock *, 8> &Visited) const { |
5515 | |
5516 | unsigned LOOPi = (Subtarget.isPPC64() ? PPC::MTCTR8loop : PPC::MTCTRloop); |
5517 | |
5518 | // The loop set-up instruction should be in preheader |
5519 | for (auto &I : PreHeader.instrs()) |
5520 | if (I.getOpcode() == LOOPi) |
5521 | return &I; |
5522 | return nullptr; |
5523 | } |
5524 | |
5525 | // Return true if get the base operand, byte offset of an instruction and the |
5526 | // memory width. Width is the size of memory that is being loaded/stored. |
5527 | bool PPCInstrInfo::getMemOperandWithOffsetWidth( |
5528 | const MachineInstr &LdSt, const MachineOperand *&BaseReg, int64_t &Offset, |
5529 | LocationSize &Width, const TargetRegisterInfo *TRI) const { |
5530 | if (!LdSt.mayLoadOrStore() || LdSt.getNumExplicitOperands() != 3) |
5531 | return false; |
5532 | |
5533 | // Handle only loads/stores with base register followed by immediate offset. |
5534 | if (!LdSt.getOperand(i: 1).isImm() || |
5535 | (!LdSt.getOperand(i: 2).isReg() && !LdSt.getOperand(i: 2).isFI())) |
5536 | return false; |
5537 | if (!LdSt.getOperand(i: 1).isImm() || |
5538 | (!LdSt.getOperand(i: 2).isReg() && !LdSt.getOperand(i: 2).isFI())) |
5539 | return false; |
5540 | |
5541 | if (!LdSt.hasOneMemOperand()) |
5542 | return false; |
5543 | |
5544 | Width = (*LdSt.memoperands_begin())->getSize(); |
5545 | Offset = LdSt.getOperand(i: 1).getImm(); |
5546 | BaseReg = &LdSt.getOperand(i: 2); |
5547 | return true; |
5548 | } |
5549 | |
5550 | bool PPCInstrInfo::areMemAccessesTriviallyDisjoint( |
5551 | const MachineInstr &MIa, const MachineInstr &MIb) const { |
5552 | assert(MIa.mayLoadOrStore() && "MIa must be a load or store." ); |
5553 | assert(MIb.mayLoadOrStore() && "MIb must be a load or store." ); |
5554 | |
5555 | if (MIa.hasUnmodeledSideEffects() || MIb.hasUnmodeledSideEffects() || |
5556 | MIa.hasOrderedMemoryRef() || MIb.hasOrderedMemoryRef()) |
5557 | return false; |
5558 | |
5559 | // Retrieve the base register, offset from the base register and width. Width |
5560 | // is the size of memory that is being loaded/stored (e.g. 1, 2, 4). If |
5561 | // base registers are identical, and the offset of a lower memory access + |
5562 | // the width doesn't overlap the offset of a higher memory access, |
5563 | // then the memory accesses are different. |
5564 | const TargetRegisterInfo *TRI = &getRegisterInfo(); |
5565 | const MachineOperand *BaseOpA = nullptr, *BaseOpB = nullptr; |
5566 | int64_t OffsetA = 0, OffsetB = 0; |
5567 | LocationSize WidthA = 0, WidthB = 0; |
5568 | if (getMemOperandWithOffsetWidth(LdSt: MIa, BaseReg&: BaseOpA, Offset&: OffsetA, Width&: WidthA, TRI) && |
5569 | getMemOperandWithOffsetWidth(LdSt: MIb, BaseReg&: BaseOpB, Offset&: OffsetB, Width&: WidthB, TRI)) { |
5570 | if (BaseOpA->isIdenticalTo(Other: *BaseOpB)) { |
5571 | int LowOffset = std::min(a: OffsetA, b: OffsetB); |
5572 | int HighOffset = std::max(a: OffsetA, b: OffsetB); |
5573 | LocationSize LowWidth = (LowOffset == OffsetA) ? WidthA : WidthB; |
5574 | if (LowWidth.hasValue() && |
5575 | LowOffset + (int)LowWidth.getValue() <= HighOffset) |
5576 | return true; |
5577 | } |
5578 | } |
5579 | return false; |
5580 | } |
5581 | |