File: | lib/Target/PowerPC/PPCInstrInfo.cpp |
Location: | line 434, column 7 |
Description: | Called C++ object pointer is null |
1 | //===-- PPCInstrInfo.cpp - PowerPC Instruction Information ----------------===// | |||
2 | // | |||
3 | // The LLVM Compiler Infrastructure | |||
4 | // | |||
5 | // This file is distributed under the University of Illinois Open Source | |||
6 | // License. See LICENSE.TXT for details. | |||
7 | // | |||
8 | //===----------------------------------------------------------------------===// | |||
9 | // | |||
10 | // This file contains the PowerPC implementation of the TargetInstrInfo class. | |||
11 | // | |||
12 | //===----------------------------------------------------------------------===// | |||
13 | ||||
14 | #include "PPCInstrInfo.h" | |||
15 | #include "MCTargetDesc/PPCPredicates.h" | |||
16 | #include "PPC.h" | |||
17 | #include "PPCHazardRecognizers.h" | |||
18 | #include "PPCInstrBuilder.h" | |||
19 | #include "PPCMachineFunctionInfo.h" | |||
20 | #include "PPCTargetMachine.h" | |||
21 | #include "llvm/ADT/STLExtras.h" | |||
22 | #include "llvm/ADT/Statistic.h" | |||
23 | #include "llvm/CodeGen/LiveIntervalAnalysis.h" | |||
24 | #include "llvm/CodeGen/MachineFrameInfo.h" | |||
25 | #include "llvm/CodeGen/MachineFunctionPass.h" | |||
26 | #include "llvm/CodeGen/MachineInstrBuilder.h" | |||
27 | #include "llvm/CodeGen/MachineMemOperand.h" | |||
28 | #include "llvm/CodeGen/MachineRegisterInfo.h" | |||
29 | #include "llvm/CodeGen/PseudoSourceValue.h" | |||
30 | #include "llvm/CodeGen/ScheduleDAG.h" | |||
31 | #include "llvm/CodeGen/SlotIndexes.h" | |||
32 | #include "llvm/MC/MCAsmInfo.h" | |||
33 | #include "llvm/Support/CommandLine.h" | |||
34 | #include "llvm/Support/Debug.h" | |||
35 | #include "llvm/Support/ErrorHandling.h" | |||
36 | #include "llvm/Support/TargetRegistry.h" | |||
37 | #include "llvm/Support/raw_ostream.h" | |||
38 | ||||
39 | using namespace llvm; | |||
40 | ||||
41 | #define DEBUG_TYPE"ppc-early-ret" "ppc-instr-info" | |||
42 | ||||
43 | #define GET_INSTRMAP_INFO | |||
44 | #define GET_INSTRINFO_CTOR_DTOR | |||
45 | #include "PPCGenInstrInfo.inc" | |||
46 | ||||
47 | static cl:: | |||
48 | opt<bool> DisableCTRLoopAnal("disable-ppc-ctrloop-analysis", cl::Hidden, | |||
49 | cl::desc("Disable analysis for CTR loops")); | |||
50 | ||||
51 | static cl::opt<bool> DisableCmpOpt("disable-ppc-cmp-opt", | |||
52 | cl::desc("Disable compare instruction optimization"), cl::Hidden); | |||
53 | ||||
54 | static cl::opt<bool> DisableVSXFMAMutate("disable-ppc-vsx-fma-mutation", | |||
55 | cl::desc("Disable VSX FMA instruction mutation"), cl::Hidden); | |||
56 | ||||
57 | static cl::opt<bool> VSXSelfCopyCrash("crash-on-ppc-vsx-self-copy", | |||
58 | cl::desc("Causes the backend to crash instead of generating a nop VSX copy"), | |||
59 | cl::Hidden); | |||
60 | ||||
61 | // Pin the vtable to this file. | |||
62 | void PPCInstrInfo::anchor() {} | |||
63 | ||||
64 | PPCInstrInfo::PPCInstrInfo(PPCSubtarget &STI) | |||
65 | : PPCGenInstrInfo(PPC::ADJCALLSTACKDOWN, PPC::ADJCALLSTACKUP), | |||
66 | Subtarget(STI), RI(STI) {} | |||
67 | ||||
68 | /// CreateTargetHazardRecognizer - Return the hazard recognizer to use for | |||
69 | /// this target when scheduling the DAG. | |||
70 | ScheduleHazardRecognizer * | |||
71 | PPCInstrInfo::CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI, | |||
72 | const ScheduleDAG *DAG) const { | |||
73 | unsigned Directive = | |||
74 | static_cast<const PPCSubtarget *>(STI)->getDarwinDirective(); | |||
75 | if (Directive == PPC::DIR_440 || Directive == PPC::DIR_A2 || | |||
76 | Directive == PPC::DIR_E500mc || Directive == PPC::DIR_E5500) { | |||
77 | const InstrItineraryData *II = | |||
78 | static_cast<const PPCSubtarget *>(STI)->getInstrItineraryData(); | |||
79 | return new ScoreboardHazardRecognizer(II, DAG); | |||
80 | } | |||
81 | ||||
82 | return TargetInstrInfo::CreateTargetHazardRecognizer(STI, DAG); | |||
83 | } | |||
84 | ||||
85 | /// CreateTargetPostRAHazardRecognizer - Return the postRA hazard recognizer | |||
86 | /// to use for this target when scheduling the DAG. | |||
87 | ScheduleHazardRecognizer *PPCInstrInfo::CreateTargetPostRAHazardRecognizer( | |||
88 | const InstrItineraryData *II, | |||
89 | const ScheduleDAG *DAG) const { | |||
90 | unsigned Directive = | |||
91 | DAG->TM.getSubtarget<PPCSubtarget>().getDarwinDirective(); | |||
92 | ||||
93 | if (Directive == PPC::DIR_PWR7 || Directive == PPC::DIR_PWR8) | |||
94 | return new PPCDispatchGroupSBHazardRecognizer(II, DAG); | |||
95 | ||||
96 | // Most subtargets use a PPC970 recognizer. | |||
97 | if (Directive != PPC::DIR_440 && Directive != PPC::DIR_A2 && | |||
98 | Directive != PPC::DIR_E500mc && Directive != PPC::DIR_E5500) { | |||
99 | assert(DAG->TII && "No InstrInfo?")((DAG->TII && "No InstrInfo?") ? static_cast<void > (0) : __assert_fail ("DAG->TII && \"No InstrInfo?\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 99, __PRETTY_FUNCTION__)); | |||
100 | ||||
101 | return new PPCHazardRecognizer970(*DAG); | |||
102 | } | |||
103 | ||||
104 | return new ScoreboardHazardRecognizer(II, DAG); | |||
105 | } | |||
106 | ||||
107 | ||||
108 | int PPCInstrInfo::getOperandLatency(const InstrItineraryData *ItinData, | |||
109 | const MachineInstr *DefMI, unsigned DefIdx, | |||
110 | const MachineInstr *UseMI, | |||
111 | unsigned UseIdx) const { | |||
112 | int Latency = PPCGenInstrInfo::getOperandLatency(ItinData, DefMI, DefIdx, | |||
113 | UseMI, UseIdx); | |||
114 | ||||
115 | const MachineOperand &DefMO = DefMI->getOperand(DefIdx); | |||
116 | unsigned Reg = DefMO.getReg(); | |||
117 | ||||
118 | const TargetRegisterInfo *TRI = &getRegisterInfo(); | |||
119 | bool IsRegCR; | |||
120 | if (TRI->isVirtualRegister(Reg)) { | |||
121 | const MachineRegisterInfo *MRI = | |||
122 | &DefMI->getParent()->getParent()->getRegInfo(); | |||
123 | IsRegCR = MRI->getRegClass(Reg)->hasSuperClassEq(&PPC::CRRCRegClass) || | |||
124 | MRI->getRegClass(Reg)->hasSuperClassEq(&PPC::CRBITRCRegClass); | |||
125 | } else { | |||
126 | IsRegCR = PPC::CRRCRegClass.contains(Reg) || | |||
127 | PPC::CRBITRCRegClass.contains(Reg); | |||
128 | } | |||
129 | ||||
130 | if (UseMI->isBranch() && IsRegCR) { | |||
131 | if (Latency < 0) | |||
132 | Latency = getInstrLatency(ItinData, DefMI); | |||
133 | ||||
134 | // On some cores, there is an additional delay between writing to a condition | |||
135 | // register, and using it from a branch. | |||
136 | unsigned Directive = Subtarget.getDarwinDirective(); | |||
137 | switch (Directive) { | |||
138 | default: break; | |||
139 | case PPC::DIR_7400: | |||
140 | case PPC::DIR_750: | |||
141 | case PPC::DIR_970: | |||
142 | case PPC::DIR_E5500: | |||
143 | case PPC::DIR_PWR4: | |||
144 | case PPC::DIR_PWR5: | |||
145 | case PPC::DIR_PWR5X: | |||
146 | case PPC::DIR_PWR6: | |||
147 | case PPC::DIR_PWR6X: | |||
148 | case PPC::DIR_PWR7: | |||
149 | case PPC::DIR_PWR8: | |||
150 | Latency += 2; | |||
151 | break; | |||
152 | } | |||
153 | } | |||
154 | ||||
155 | return Latency; | |||
156 | } | |||
157 | ||||
158 | // Detect 32 -> 64-bit extensions where we may reuse the low sub-register. | |||
159 | bool PPCInstrInfo::isCoalescableExtInstr(const MachineInstr &MI, | |||
160 | unsigned &SrcReg, unsigned &DstReg, | |||
161 | unsigned &SubIdx) const { | |||
162 | switch (MI.getOpcode()) { | |||
163 | default: return false; | |||
164 | case PPC::EXTSW: | |||
165 | case PPC::EXTSW_32_64: | |||
166 | SrcReg = MI.getOperand(1).getReg(); | |||
167 | DstReg = MI.getOperand(0).getReg(); | |||
168 | SubIdx = PPC::sub_32; | |||
169 | return true; | |||
170 | } | |||
171 | } | |||
172 | ||||
173 | unsigned PPCInstrInfo::isLoadFromStackSlot(const MachineInstr *MI, | |||
174 | int &FrameIndex) const { | |||
175 | // Note: This list must be kept consistent with LoadRegFromStackSlot. | |||
176 | switch (MI->getOpcode()) { | |||
177 | default: break; | |||
178 | case PPC::LD: | |||
179 | case PPC::LWZ: | |||
180 | case PPC::LFS: | |||
181 | case PPC::LFD: | |||
182 | case PPC::RESTORE_CR: | |||
183 | case PPC::RESTORE_CRBIT: | |||
184 | case PPC::LVX: | |||
185 | case PPC::LXVD2X: | |||
186 | case PPC::RESTORE_VRSAVE: | |||
187 | // Check for the operands added by addFrameReference (the immediate is the | |||
188 | // offset which defaults to 0). | |||
189 | if (MI->getOperand(1).isImm() && !MI->getOperand(1).getImm() && | |||
190 | MI->getOperand(2).isFI()) { | |||
191 | FrameIndex = MI->getOperand(2).getIndex(); | |||
192 | return MI->getOperand(0).getReg(); | |||
193 | } | |||
194 | break; | |||
195 | } | |||
196 | return 0; | |||
197 | } | |||
198 | ||||
199 | unsigned PPCInstrInfo::isStoreToStackSlot(const MachineInstr *MI, | |||
200 | int &FrameIndex) const { | |||
201 | // Note: This list must be kept consistent with StoreRegToStackSlot. | |||
202 | switch (MI->getOpcode()) { | |||
203 | default: break; | |||
204 | case PPC::STD: | |||
205 | case PPC::STW: | |||
206 | case PPC::STFS: | |||
207 | case PPC::STFD: | |||
208 | case PPC::SPILL_CR: | |||
209 | case PPC::SPILL_CRBIT: | |||
210 | case PPC::STVX: | |||
211 | case PPC::STXVD2X: | |||
212 | case PPC::SPILL_VRSAVE: | |||
213 | // Check for the operands added by addFrameReference (the immediate is the | |||
214 | // offset which defaults to 0). | |||
215 | if (MI->getOperand(1).isImm() && !MI->getOperand(1).getImm() && | |||
216 | MI->getOperand(2).isFI()) { | |||
217 | FrameIndex = MI->getOperand(2).getIndex(); | |||
218 | return MI->getOperand(0).getReg(); | |||
219 | } | |||
220 | break; | |||
221 | } | |||
222 | return 0; | |||
223 | } | |||
224 | ||||
225 | // commuteInstruction - We can commute rlwimi instructions, but only if the | |||
226 | // rotate amt is zero. We also have to munge the immediates a bit. | |||
227 | MachineInstr * | |||
228 | PPCInstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const { | |||
229 | MachineFunction &MF = *MI->getParent()->getParent(); | |||
230 | ||||
231 | // Normal instructions can be commuted the obvious way. | |||
232 | if (MI->getOpcode() != PPC::RLWIMI && | |||
233 | MI->getOpcode() != PPC::RLWIMIo) | |||
234 | return TargetInstrInfo::commuteInstruction(MI, NewMI); | |||
235 | // Note that RLWIMI can be commuted as a 32-bit instruction, but not as a | |||
236 | // 64-bit instruction (so we don't handle PPC::RLWIMI8 here), because | |||
237 | // changing the relative order of the mask operands might change what happens | |||
238 | // to the high-bits of the mask (and, thus, the result). | |||
239 | ||||
240 | // Cannot commute if it has a non-zero rotate count. | |||
241 | if (MI->getOperand(3).getImm() != 0) | |||
242 | return nullptr; | |||
243 | ||||
244 | // If we have a zero rotate count, we have: | |||
245 | // M = mask(MB,ME) | |||
246 | // Op0 = (Op1 & ~M) | (Op2 & M) | |||
247 | // Change this to: | |||
248 | // M = mask((ME+1)&31, (MB-1)&31) | |||
249 | // Op0 = (Op2 & ~M) | (Op1 & M) | |||
250 | ||||
251 | // Swap op1/op2 | |||
252 | unsigned Reg0 = MI->getOperand(0).getReg(); | |||
253 | unsigned Reg1 = MI->getOperand(1).getReg(); | |||
254 | unsigned Reg2 = MI->getOperand(2).getReg(); | |||
255 | unsigned SubReg1 = MI->getOperand(1).getSubReg(); | |||
256 | unsigned SubReg2 = MI->getOperand(2).getSubReg(); | |||
257 | bool Reg1IsKill = MI->getOperand(1).isKill(); | |||
258 | bool Reg2IsKill = MI->getOperand(2).isKill(); | |||
259 | bool ChangeReg0 = false; | |||
260 | // If machine instrs are no longer in two-address forms, update | |||
261 | // destination register as well. | |||
262 | if (Reg0 == Reg1) { | |||
263 | // Must be two address instruction! | |||
264 | assert(MI->getDesc().getOperandConstraint(0, MCOI::TIED_TO) &&((MI->getDesc().getOperandConstraint(0, MCOI::TIED_TO) && "Expecting a two-address instruction!") ? static_cast<void > (0) : __assert_fail ("MI->getDesc().getOperandConstraint(0, MCOI::TIED_TO) && \"Expecting a two-address instruction!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 265, __PRETTY_FUNCTION__)) | |||
265 | "Expecting a two-address instruction!")((MI->getDesc().getOperandConstraint(0, MCOI::TIED_TO) && "Expecting a two-address instruction!") ? static_cast<void > (0) : __assert_fail ("MI->getDesc().getOperandConstraint(0, MCOI::TIED_TO) && \"Expecting a two-address instruction!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 265, __PRETTY_FUNCTION__)); | |||
266 | assert(MI->getOperand(0).getSubReg() == SubReg1 && "Tied subreg mismatch")((MI->getOperand(0).getSubReg() == SubReg1 && "Tied subreg mismatch" ) ? static_cast<void> (0) : __assert_fail ("MI->getOperand(0).getSubReg() == SubReg1 && \"Tied subreg mismatch\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 266, __PRETTY_FUNCTION__)); | |||
267 | Reg2IsKill = false; | |||
268 | ChangeReg0 = true; | |||
269 | } | |||
270 | ||||
271 | // Masks. | |||
272 | unsigned MB = MI->getOperand(4).getImm(); | |||
273 | unsigned ME = MI->getOperand(5).getImm(); | |||
274 | ||||
275 | if (NewMI) { | |||
276 | // Create a new instruction. | |||
277 | unsigned Reg0 = ChangeReg0 ? Reg2 : MI->getOperand(0).getReg(); | |||
278 | bool Reg0IsDead = MI->getOperand(0).isDead(); | |||
279 | return BuildMI(MF, MI->getDebugLoc(), MI->getDesc()) | |||
280 | .addReg(Reg0, RegState::Define | getDeadRegState(Reg0IsDead)) | |||
281 | .addReg(Reg2, getKillRegState(Reg2IsKill)) | |||
282 | .addReg(Reg1, getKillRegState(Reg1IsKill)) | |||
283 | .addImm((ME+1) & 31) | |||
284 | .addImm((MB-1) & 31); | |||
285 | } | |||
286 | ||||
287 | if (ChangeReg0) { | |||
288 | MI->getOperand(0).setReg(Reg2); | |||
289 | MI->getOperand(0).setSubReg(SubReg2); | |||
290 | } | |||
291 | MI->getOperand(2).setReg(Reg1); | |||
292 | MI->getOperand(1).setReg(Reg2); | |||
293 | MI->getOperand(2).setSubReg(SubReg1); | |||
294 | MI->getOperand(1).setSubReg(SubReg2); | |||
295 | MI->getOperand(2).setIsKill(Reg1IsKill); | |||
296 | MI->getOperand(1).setIsKill(Reg2IsKill); | |||
297 | ||||
298 | // Swap the mask around. | |||
299 | MI->getOperand(4).setImm((ME+1) & 31); | |||
300 | MI->getOperand(5).setImm((MB-1) & 31); | |||
301 | return MI; | |||
302 | } | |||
303 | ||||
304 | bool PPCInstrInfo::findCommutedOpIndices(MachineInstr *MI, unsigned &SrcOpIdx1, | |||
305 | unsigned &SrcOpIdx2) const { | |||
306 | // For VSX A-Type FMA instructions, it is the first two operands that can be | |||
307 | // commuted, however, because the non-encoded tied input operand is listed | |||
308 | // first, the operands to swap are actually the second and third. | |||
309 | ||||
310 | int AltOpc = PPC::getAltVSXFMAOpcode(MI->getOpcode()); | |||
311 | if (AltOpc == -1) | |||
312 | return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2); | |||
313 | ||||
314 | SrcOpIdx1 = 2; | |||
315 | SrcOpIdx2 = 3; | |||
316 | return true; | |||
317 | } | |||
318 | ||||
319 | void PPCInstrInfo::insertNoop(MachineBasicBlock &MBB, | |||
320 | MachineBasicBlock::iterator MI) const { | |||
321 | // This function is used for scheduling, and the nop wanted here is the type | |||
322 | // that terminates dispatch groups on the POWER cores. | |||
323 | unsigned Directive = Subtarget.getDarwinDirective(); | |||
324 | unsigned Opcode; | |||
325 | switch (Directive) { | |||
326 | default: Opcode = PPC::NOP; break; | |||
327 | case PPC::DIR_PWR6: Opcode = PPC::NOP_GT_PWR6; break; | |||
328 | case PPC::DIR_PWR7: Opcode = PPC::NOP_GT_PWR7; break; | |||
329 | case PPC::DIR_PWR8: Opcode = PPC::NOP_GT_PWR7; break; /* FIXME: Update when P8 InstrScheduling model is ready */ | |||
330 | } | |||
331 | ||||
332 | DebugLoc DL; | |||
333 | BuildMI(MBB, MI, DL, get(Opcode)); | |||
334 | } | |||
335 | ||||
336 | /// getNoopForMachoTarget - Return the noop instruction to use for a noop. | |||
337 | void PPCInstrInfo::getNoopForMachoTarget(MCInst &NopInst) const { | |||
338 | NopInst.setOpcode(PPC::NOP); | |||
339 | } | |||
340 | ||||
341 | // Branch analysis. | |||
342 | // Note: If the condition register is set to CTR or CTR8 then this is a | |||
343 | // BDNZ (imm == 1) or BDZ (imm == 0) branch. | |||
344 | bool PPCInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,MachineBasicBlock *&TBB, | |||
345 | MachineBasicBlock *&FBB, | |||
346 | SmallVectorImpl<MachineOperand> &Cond, | |||
347 | bool AllowModify) const { | |||
348 | bool isPPC64 = Subtarget.isPPC64(); | |||
349 | ||||
350 | // If the block has no terminators, it just falls into the block after it. | |||
351 | MachineBasicBlock::iterator I = MBB.end(); | |||
352 | if (I == MBB.begin()) | |||
| ||||
353 | return false; | |||
354 | --I; | |||
355 | while (I->isDebugValue()) { | |||
356 | if (I == MBB.begin()) | |||
357 | return false; | |||
358 | --I; | |||
359 | } | |||
360 | if (!isUnpredicatedTerminator(I)) | |||
361 | return false; | |||
362 | ||||
363 | // Get the last instruction in the block. | |||
364 | MachineInstr *LastInst = I; | |||
365 | ||||
366 | // If there is only one terminator instruction, process it. | |||
367 | if (I == MBB.begin() || !isUnpredicatedTerminator(--I)) { | |||
368 | if (LastInst->getOpcode() == PPC::B) { | |||
369 | if (!LastInst->getOperand(0).isMBB()) | |||
370 | return true; | |||
371 | TBB = LastInst->getOperand(0).getMBB(); | |||
372 | return false; | |||
373 | } else if (LastInst->getOpcode() == PPC::BCC) { | |||
374 | if (!LastInst->getOperand(2).isMBB()) | |||
375 | return true; | |||
376 | // Block ends with fall-through condbranch. | |||
377 | TBB = LastInst->getOperand(2).getMBB(); | |||
378 | Cond.push_back(LastInst->getOperand(0)); | |||
379 | Cond.push_back(LastInst->getOperand(1)); | |||
380 | return false; | |||
381 | } else if (LastInst->getOpcode() == PPC::BC) { | |||
382 | if (!LastInst->getOperand(1).isMBB()) | |||
383 | return true; | |||
384 | // Block ends with fall-through condbranch. | |||
385 | TBB = LastInst->getOperand(1).getMBB(); | |||
386 | Cond.push_back(MachineOperand::CreateImm(PPC::PRED_BIT_SET)); | |||
387 | Cond.push_back(LastInst->getOperand(0)); | |||
388 | return false; | |||
389 | } else if (LastInst->getOpcode() == PPC::BCn) { | |||
390 | if (!LastInst->getOperand(1).isMBB()) | |||
391 | return true; | |||
392 | // Block ends with fall-through condbranch. | |||
393 | TBB = LastInst->getOperand(1).getMBB(); | |||
394 | Cond.push_back(MachineOperand::CreateImm(PPC::PRED_BIT_UNSET)); | |||
395 | Cond.push_back(LastInst->getOperand(0)); | |||
396 | return false; | |||
397 | } else if (LastInst->getOpcode() == PPC::BDNZ8 || | |||
398 | LastInst->getOpcode() == PPC::BDNZ) { | |||
399 | if (!LastInst->getOperand(0).isMBB()) | |||
400 | return true; | |||
401 | if (DisableCTRLoopAnal) | |||
402 | return true; | |||
403 | TBB = LastInst->getOperand(0).getMBB(); | |||
404 | Cond.push_back(MachineOperand::CreateImm(1)); | |||
405 | Cond.push_back(MachineOperand::CreateReg(isPPC64 ? PPC::CTR8 : PPC::CTR, | |||
406 | true)); | |||
407 | return false; | |||
408 | } else if (LastInst->getOpcode() == PPC::BDZ8 || | |||
409 | LastInst->getOpcode() == PPC::BDZ) { | |||
410 | if (!LastInst->getOperand(0).isMBB()) | |||
411 | return true; | |||
412 | if (DisableCTRLoopAnal) | |||
413 | return true; | |||
414 | TBB = LastInst->getOperand(0).getMBB(); | |||
415 | Cond.push_back(MachineOperand::CreateImm(0)); | |||
416 | Cond.push_back(MachineOperand::CreateReg(isPPC64 ? PPC::CTR8 : PPC::CTR, | |||
417 | true)); | |||
418 | return false; | |||
419 | } | |||
420 | ||||
421 | // Otherwise, don't know what this is. | |||
422 | return true; | |||
423 | } | |||
424 | ||||
425 | // Get the instruction before it if it's a terminator. | |||
426 | MachineInstr *SecondLastInst = I; | |||
427 | ||||
428 | // If there are three terminators, we don't know what sort of block this is. | |||
429 | if (SecondLastInst && I != MBB.begin() && | |||
430 | isUnpredicatedTerminator(--I)) | |||
431 | return true; | |||
432 | ||||
433 | // If the block ends with PPC::B and PPC:BCC, handle it. | |||
434 | if (SecondLastInst->getOpcode() == PPC::BCC && | |||
| ||||
435 | LastInst->getOpcode() == PPC::B) { | |||
436 | if (!SecondLastInst->getOperand(2).isMBB() || | |||
437 | !LastInst->getOperand(0).isMBB()) | |||
438 | return true; | |||
439 | TBB = SecondLastInst->getOperand(2).getMBB(); | |||
440 | Cond.push_back(SecondLastInst->getOperand(0)); | |||
441 | Cond.push_back(SecondLastInst->getOperand(1)); | |||
442 | FBB = LastInst->getOperand(0).getMBB(); | |||
443 | return false; | |||
444 | } else if (SecondLastInst->getOpcode() == PPC::BC && | |||
445 | LastInst->getOpcode() == PPC::B) { | |||
446 | if (!SecondLastInst->getOperand(1).isMBB() || | |||
447 | !LastInst->getOperand(0).isMBB()) | |||
448 | return true; | |||
449 | TBB = SecondLastInst->getOperand(1).getMBB(); | |||
450 | Cond.push_back(MachineOperand::CreateImm(PPC::PRED_BIT_SET)); | |||
451 | Cond.push_back(SecondLastInst->getOperand(0)); | |||
452 | FBB = LastInst->getOperand(0).getMBB(); | |||
453 | return false; | |||
454 | } else if (SecondLastInst->getOpcode() == PPC::BCn && | |||
455 | LastInst->getOpcode() == PPC::B) { | |||
456 | if (!SecondLastInst->getOperand(1).isMBB() || | |||
457 | !LastInst->getOperand(0).isMBB()) | |||
458 | return true; | |||
459 | TBB = SecondLastInst->getOperand(1).getMBB(); | |||
460 | Cond.push_back(MachineOperand::CreateImm(PPC::PRED_BIT_UNSET)); | |||
461 | Cond.push_back(SecondLastInst->getOperand(0)); | |||
462 | FBB = LastInst->getOperand(0).getMBB(); | |||
463 | return false; | |||
464 | } else if ((SecondLastInst->getOpcode() == PPC::BDNZ8 || | |||
465 | SecondLastInst->getOpcode() == PPC::BDNZ) && | |||
466 | LastInst->getOpcode() == PPC::B) { | |||
467 | if (!SecondLastInst->getOperand(0).isMBB() || | |||
468 | !LastInst->getOperand(0).isMBB()) | |||
469 | return true; | |||
470 | if (DisableCTRLoopAnal) | |||
471 | return true; | |||
472 | TBB = SecondLastInst->getOperand(0).getMBB(); | |||
473 | Cond.push_back(MachineOperand::CreateImm(1)); | |||
474 | Cond.push_back(MachineOperand::CreateReg(isPPC64 ? PPC::CTR8 : PPC::CTR, | |||
475 | true)); | |||
476 | FBB = LastInst->getOperand(0).getMBB(); | |||
477 | return false; | |||
478 | } else if ((SecondLastInst->getOpcode() == PPC::BDZ8 || | |||
479 | SecondLastInst->getOpcode() == PPC::BDZ) && | |||
480 | LastInst->getOpcode() == PPC::B) { | |||
481 | if (!SecondLastInst->getOperand(0).isMBB() || | |||
482 | !LastInst->getOperand(0).isMBB()) | |||
483 | return true; | |||
484 | if (DisableCTRLoopAnal) | |||
485 | return true; | |||
486 | TBB = SecondLastInst->getOperand(0).getMBB(); | |||
487 | Cond.push_back(MachineOperand::CreateImm(0)); | |||
488 | Cond.push_back(MachineOperand::CreateReg(isPPC64 ? PPC::CTR8 : PPC::CTR, | |||
489 | true)); | |||
490 | FBB = LastInst->getOperand(0).getMBB(); | |||
491 | return false; | |||
492 | } | |||
493 | ||||
494 | // If the block ends with two PPC:Bs, handle it. The second one is not | |||
495 | // executed, so remove it. | |||
496 | if (SecondLastInst->getOpcode() == PPC::B && | |||
497 | LastInst->getOpcode() == PPC::B) { | |||
498 | if (!SecondLastInst->getOperand(0).isMBB()) | |||
499 | return true; | |||
500 | TBB = SecondLastInst->getOperand(0).getMBB(); | |||
501 | I = LastInst; | |||
502 | if (AllowModify) | |||
503 | I->eraseFromParent(); | |||
504 | return false; | |||
505 | } | |||
506 | ||||
507 | // Otherwise, can't handle this. | |||
508 | return true; | |||
509 | } | |||
510 | ||||
511 | unsigned PPCInstrInfo::RemoveBranch(MachineBasicBlock &MBB) const { | |||
512 | MachineBasicBlock::iterator I = MBB.end(); | |||
513 | if (I == MBB.begin()) return 0; | |||
514 | --I; | |||
515 | while (I->isDebugValue()) { | |||
516 | if (I == MBB.begin()) | |||
517 | return 0; | |||
518 | --I; | |||
519 | } | |||
520 | if (I->getOpcode() != PPC::B && I->getOpcode() != PPC::BCC && | |||
521 | I->getOpcode() != PPC::BC && I->getOpcode() != PPC::BCn && | |||
522 | I->getOpcode() != PPC::BDNZ8 && I->getOpcode() != PPC::BDNZ && | |||
523 | I->getOpcode() != PPC::BDZ8 && I->getOpcode() != PPC::BDZ) | |||
524 | return 0; | |||
525 | ||||
526 | // Remove the branch. | |||
527 | I->eraseFromParent(); | |||
528 | ||||
529 | I = MBB.end(); | |||
530 | ||||
531 | if (I == MBB.begin()) return 1; | |||
532 | --I; | |||
533 | if (I->getOpcode() != PPC::BCC && | |||
534 | I->getOpcode() != PPC::BC && I->getOpcode() != PPC::BCn && | |||
535 | I->getOpcode() != PPC::BDNZ8 && I->getOpcode() != PPC::BDNZ && | |||
536 | I->getOpcode() != PPC::BDZ8 && I->getOpcode() != PPC::BDZ) | |||
537 | return 1; | |||
538 | ||||
539 | // Remove the branch. | |||
540 | I->eraseFromParent(); | |||
541 | return 2; | |||
542 | } | |||
543 | ||||
544 | unsigned | |||
545 | PPCInstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB, | |||
546 | MachineBasicBlock *FBB, | |||
547 | const SmallVectorImpl<MachineOperand> &Cond, | |||
548 | DebugLoc DL) const { | |||
549 | // Shouldn't be a fall through. | |||
550 | assert(TBB && "InsertBranch must not be told to insert a fallthrough")((TBB && "InsertBranch must not be told to insert a fallthrough" ) ? static_cast<void> (0) : __assert_fail ("TBB && \"InsertBranch must not be told to insert a fallthrough\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 550, __PRETTY_FUNCTION__)); | |||
551 | assert((Cond.size() == 2 || Cond.size() == 0) &&(((Cond.size() == 2 || Cond.size() == 0) && "PPC branch conditions have two components!" ) ? static_cast<void> (0) : __assert_fail ("(Cond.size() == 2 || Cond.size() == 0) && \"PPC branch conditions have two components!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 552, __PRETTY_FUNCTION__)) | |||
552 | "PPC branch conditions have two components!")(((Cond.size() == 2 || Cond.size() == 0) && "PPC branch conditions have two components!" ) ? static_cast<void> (0) : __assert_fail ("(Cond.size() == 2 || Cond.size() == 0) && \"PPC branch conditions have two components!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 552, __PRETTY_FUNCTION__)); | |||
553 | ||||
554 | bool isPPC64 = Subtarget.isPPC64(); | |||
555 | ||||
556 | // One-way branch. | |||
557 | if (!FBB) { | |||
558 | if (Cond.empty()) // Unconditional branch | |||
559 | BuildMI(&MBB, DL, get(PPC::B)).addMBB(TBB); | |||
560 | else if (Cond[1].getReg() == PPC::CTR || Cond[1].getReg() == PPC::CTR8) | |||
561 | BuildMI(&MBB, DL, get(Cond[0].getImm() ? | |||
562 | (isPPC64 ? PPC::BDNZ8 : PPC::BDNZ) : | |||
563 | (isPPC64 ? PPC::BDZ8 : PPC::BDZ))).addMBB(TBB); | |||
564 | else if (Cond[0].getImm() == PPC::PRED_BIT_SET) | |||
565 | BuildMI(&MBB, DL, get(PPC::BC)).addOperand(Cond[1]).addMBB(TBB); | |||
566 | else if (Cond[0].getImm() == PPC::PRED_BIT_UNSET) | |||
567 | BuildMI(&MBB, DL, get(PPC::BCn)).addOperand(Cond[1]).addMBB(TBB); | |||
568 | else // Conditional branch | |||
569 | BuildMI(&MBB, DL, get(PPC::BCC)) | |||
570 | .addImm(Cond[0].getImm()).addOperand(Cond[1]).addMBB(TBB); | |||
571 | return 1; | |||
572 | } | |||
573 | ||||
574 | // Two-way Conditional Branch. | |||
575 | if (Cond[1].getReg() == PPC::CTR || Cond[1].getReg() == PPC::CTR8) | |||
576 | BuildMI(&MBB, DL, get(Cond[0].getImm() ? | |||
577 | (isPPC64 ? PPC::BDNZ8 : PPC::BDNZ) : | |||
578 | (isPPC64 ? PPC::BDZ8 : PPC::BDZ))).addMBB(TBB); | |||
579 | else if (Cond[0].getImm() == PPC::PRED_BIT_SET) | |||
580 | BuildMI(&MBB, DL, get(PPC::BC)).addOperand(Cond[1]).addMBB(TBB); | |||
581 | else if (Cond[0].getImm() == PPC::PRED_BIT_UNSET) | |||
582 | BuildMI(&MBB, DL, get(PPC::BCn)).addOperand(Cond[1]).addMBB(TBB); | |||
583 | else | |||
584 | BuildMI(&MBB, DL, get(PPC::BCC)) | |||
585 | .addImm(Cond[0].getImm()).addOperand(Cond[1]).addMBB(TBB); | |||
586 | BuildMI(&MBB, DL, get(PPC::B)).addMBB(FBB); | |||
587 | return 2; | |||
588 | } | |||
589 | ||||
590 | // Select analysis. | |||
591 | bool PPCInstrInfo::canInsertSelect(const MachineBasicBlock &MBB, | |||
592 | const SmallVectorImpl<MachineOperand> &Cond, | |||
593 | unsigned TrueReg, unsigned FalseReg, | |||
594 | int &CondCycles, int &TrueCycles, int &FalseCycles) const { | |||
595 | if (!Subtarget.hasISEL()) | |||
596 | return false; | |||
597 | ||||
598 | if (Cond.size() != 2) | |||
599 | return false; | |||
600 | ||||
601 | // If this is really a bdnz-like condition, then it cannot be turned into a | |||
602 | // select. | |||
603 | if (Cond[1].getReg() == PPC::CTR || Cond[1].getReg() == PPC::CTR8) | |||
604 | return false; | |||
605 | ||||
606 | // Check register classes. | |||
607 | const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); | |||
608 | const TargetRegisterClass *RC = | |||
609 | RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg)); | |||
610 | if (!RC) | |||
611 | return false; | |||
612 | ||||
613 | // isel is for regular integer GPRs only. | |||
614 | if (!PPC::GPRCRegClass.hasSubClassEq(RC) && | |||
615 | !PPC::GPRC_NOR0RegClass.hasSubClassEq(RC) && | |||
616 | !PPC::G8RCRegClass.hasSubClassEq(RC) && | |||
617 | !PPC::G8RC_NOX0RegClass.hasSubClassEq(RC)) | |||
618 | return false; | |||
619 | ||||
620 | // FIXME: These numbers are for the A2, how well they work for other cores is | |||
621 | // an open question. On the A2, the isel instruction has a 2-cycle latency | |||
622 | // but single-cycle throughput. These numbers are used in combination with | |||
623 | // the MispredictPenalty setting from the active SchedMachineModel. | |||
624 | CondCycles = 1; | |||
625 | TrueCycles = 1; | |||
626 | FalseCycles = 1; | |||
627 | ||||
628 | return true; | |||
629 | } | |||
630 | ||||
631 | void PPCInstrInfo::insertSelect(MachineBasicBlock &MBB, | |||
632 | MachineBasicBlock::iterator MI, DebugLoc dl, | |||
633 | unsigned DestReg, | |||
634 | const SmallVectorImpl<MachineOperand> &Cond, | |||
635 | unsigned TrueReg, unsigned FalseReg) const { | |||
636 | assert(Cond.size() == 2 &&((Cond.size() == 2 && "PPC branch conditions have two components!" ) ? static_cast<void> (0) : __assert_fail ("Cond.size() == 2 && \"PPC branch conditions have two components!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 637, __PRETTY_FUNCTION__)) | |||
637 | "PPC branch conditions have two components!")((Cond.size() == 2 && "PPC branch conditions have two components!" ) ? static_cast<void> (0) : __assert_fail ("Cond.size() == 2 && \"PPC branch conditions have two components!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 637, __PRETTY_FUNCTION__)); | |||
638 | ||||
639 | assert(Subtarget.hasISEL() &&((Subtarget.hasISEL() && "Cannot insert select on target without ISEL support" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasISEL() && \"Cannot insert select on target without ISEL support\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 640, __PRETTY_FUNCTION__)) | |||
640 | "Cannot insert select on target without ISEL support")((Subtarget.hasISEL() && "Cannot insert select on target without ISEL support" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasISEL() && \"Cannot insert select on target without ISEL support\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 640, __PRETTY_FUNCTION__)); | |||
641 | ||||
642 | // Get the register classes. | |||
643 | MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); | |||
644 | const TargetRegisterClass *RC = | |||
645 | RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg)); | |||
646 | assert(RC && "TrueReg and FalseReg must have overlapping register classes")((RC && "TrueReg and FalseReg must have overlapping register classes" ) ? static_cast<void> (0) : __assert_fail ("RC && \"TrueReg and FalseReg must have overlapping register classes\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 646, __PRETTY_FUNCTION__)); | |||
647 | ||||
648 | bool Is64Bit = PPC::G8RCRegClass.hasSubClassEq(RC) || | |||
649 | PPC::G8RC_NOX0RegClass.hasSubClassEq(RC); | |||
650 | assert((Is64Bit ||(((Is64Bit || PPC::GPRCRegClass.hasSubClassEq(RC) || PPC::GPRC_NOR0RegClass .hasSubClassEq(RC)) && "isel is for regular integer GPRs only" ) ? static_cast<void> (0) : __assert_fail ("(Is64Bit || PPC::GPRCRegClass.hasSubClassEq(RC) || PPC::GPRC_NOR0RegClass.hasSubClassEq(RC)) && \"isel is for regular integer GPRs only\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 653, __PRETTY_FUNCTION__)) | |||
651 | PPC::GPRCRegClass.hasSubClassEq(RC) ||(((Is64Bit || PPC::GPRCRegClass.hasSubClassEq(RC) || PPC::GPRC_NOR0RegClass .hasSubClassEq(RC)) && "isel is for regular integer GPRs only" ) ? static_cast<void> (0) : __assert_fail ("(Is64Bit || PPC::GPRCRegClass.hasSubClassEq(RC) || PPC::GPRC_NOR0RegClass.hasSubClassEq(RC)) && \"isel is for regular integer GPRs only\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 653, __PRETTY_FUNCTION__)) | |||
652 | PPC::GPRC_NOR0RegClass.hasSubClassEq(RC)) &&(((Is64Bit || PPC::GPRCRegClass.hasSubClassEq(RC) || PPC::GPRC_NOR0RegClass .hasSubClassEq(RC)) && "isel is for regular integer GPRs only" ) ? static_cast<void> (0) : __assert_fail ("(Is64Bit || PPC::GPRCRegClass.hasSubClassEq(RC) || PPC::GPRC_NOR0RegClass.hasSubClassEq(RC)) && \"isel is for regular integer GPRs only\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 653, __PRETTY_FUNCTION__)) | |||
653 | "isel is for regular integer GPRs only")(((Is64Bit || PPC::GPRCRegClass.hasSubClassEq(RC) || PPC::GPRC_NOR0RegClass .hasSubClassEq(RC)) && "isel is for regular integer GPRs only" ) ? static_cast<void> (0) : __assert_fail ("(Is64Bit || PPC::GPRCRegClass.hasSubClassEq(RC) || PPC::GPRC_NOR0RegClass.hasSubClassEq(RC)) && \"isel is for regular integer GPRs only\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 653, __PRETTY_FUNCTION__)); | |||
654 | ||||
655 | unsigned OpCode = Is64Bit ? PPC::ISEL8 : PPC::ISEL; | |||
656 | unsigned SelectPred = Cond[0].getImm(); | |||
657 | ||||
658 | unsigned SubIdx; | |||
659 | bool SwapOps; | |||
660 | switch (SelectPred) { | |||
661 | default: llvm_unreachable("invalid predicate for isel")::llvm::llvm_unreachable_internal("invalid predicate for isel" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 661); | |||
662 | case PPC::PRED_EQ: SubIdx = PPC::sub_eq; SwapOps = false; break; | |||
663 | case PPC::PRED_NE: SubIdx = PPC::sub_eq; SwapOps = true; break; | |||
664 | case PPC::PRED_LT: SubIdx = PPC::sub_lt; SwapOps = false; break; | |||
665 | case PPC::PRED_GE: SubIdx = PPC::sub_lt; SwapOps = true; break; | |||
666 | case PPC::PRED_GT: SubIdx = PPC::sub_gt; SwapOps = false; break; | |||
667 | case PPC::PRED_LE: SubIdx = PPC::sub_gt; SwapOps = true; break; | |||
668 | case PPC::PRED_UN: SubIdx = PPC::sub_un; SwapOps = false; break; | |||
669 | case PPC::PRED_NU: SubIdx = PPC::sub_un; SwapOps = true; break; | |||
670 | case PPC::PRED_BIT_SET: SubIdx = 0; SwapOps = false; break; | |||
671 | case PPC::PRED_BIT_UNSET: SubIdx = 0; SwapOps = true; break; | |||
672 | } | |||
673 | ||||
674 | unsigned FirstReg = SwapOps ? FalseReg : TrueReg, | |||
675 | SecondReg = SwapOps ? TrueReg : FalseReg; | |||
676 | ||||
677 | // The first input register of isel cannot be r0. If it is a member | |||
678 | // of a register class that can be r0, then copy it first (the | |||
679 | // register allocator should eliminate the copy). | |||
680 | if (MRI.getRegClass(FirstReg)->contains(PPC::R0) || | |||
681 | MRI.getRegClass(FirstReg)->contains(PPC::X0)) { | |||
682 | const TargetRegisterClass *FirstRC = | |||
683 | MRI.getRegClass(FirstReg)->contains(PPC::X0) ? | |||
684 | &PPC::G8RC_NOX0RegClass : &PPC::GPRC_NOR0RegClass; | |||
685 | unsigned OldFirstReg = FirstReg; | |||
686 | FirstReg = MRI.createVirtualRegister(FirstRC); | |||
687 | BuildMI(MBB, MI, dl, get(TargetOpcode::COPY), FirstReg) | |||
688 | .addReg(OldFirstReg); | |||
689 | } | |||
690 | ||||
691 | BuildMI(MBB, MI, dl, get(OpCode), DestReg) | |||
692 | .addReg(FirstReg).addReg(SecondReg) | |||
693 | .addReg(Cond[1].getReg(), 0, SubIdx); | |||
694 | } | |||
695 | ||||
696 | void PPCInstrInfo::copyPhysReg(MachineBasicBlock &MBB, | |||
697 | MachineBasicBlock::iterator I, DebugLoc DL, | |||
698 | unsigned DestReg, unsigned SrcReg, | |||
699 | bool KillSrc) const { | |||
700 | // We can end up with self copies and similar things as a result of VSX copy | |||
701 | // legalization. Promote them here. | |||
702 | const TargetRegisterInfo *TRI = &getRegisterInfo(); | |||
703 | if (PPC::F8RCRegClass.contains(DestReg) && | |||
704 | PPC::VSLRCRegClass.contains(SrcReg)) { | |||
705 | unsigned SuperReg = | |||
706 | TRI->getMatchingSuperReg(DestReg, PPC::sub_64, &PPC::VSRCRegClass); | |||
707 | ||||
708 | if (VSXSelfCopyCrash && SrcReg == SuperReg) | |||
709 | llvm_unreachable("nop VSX copy")::llvm::llvm_unreachable_internal("nop VSX copy", "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 709); | |||
710 | ||||
711 | DestReg = SuperReg; | |||
712 | } else if (PPC::VRRCRegClass.contains(DestReg) && | |||
713 | PPC::VSHRCRegClass.contains(SrcReg)) { | |||
714 | unsigned SuperReg = | |||
715 | TRI->getMatchingSuperReg(DestReg, PPC::sub_128, &PPC::VSRCRegClass); | |||
716 | ||||
717 | if (VSXSelfCopyCrash && SrcReg == SuperReg) | |||
718 | llvm_unreachable("nop VSX copy")::llvm::llvm_unreachable_internal("nop VSX copy", "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 718); | |||
719 | ||||
720 | DestReg = SuperReg; | |||
721 | } else if (PPC::F8RCRegClass.contains(SrcReg) && | |||
722 | PPC::VSLRCRegClass.contains(DestReg)) { | |||
723 | unsigned SuperReg = | |||
724 | TRI->getMatchingSuperReg(SrcReg, PPC::sub_64, &PPC::VSRCRegClass); | |||
725 | ||||
726 | if (VSXSelfCopyCrash && DestReg == SuperReg) | |||
727 | llvm_unreachable("nop VSX copy")::llvm::llvm_unreachable_internal("nop VSX copy", "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 727); | |||
728 | ||||
729 | SrcReg = SuperReg; | |||
730 | } else if (PPC::VRRCRegClass.contains(SrcReg) && | |||
731 | PPC::VSHRCRegClass.contains(DestReg)) { | |||
732 | unsigned SuperReg = | |||
733 | TRI->getMatchingSuperReg(SrcReg, PPC::sub_128, &PPC::VSRCRegClass); | |||
734 | ||||
735 | if (VSXSelfCopyCrash && DestReg == SuperReg) | |||
736 | llvm_unreachable("nop VSX copy")::llvm::llvm_unreachable_internal("nop VSX copy", "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 736); | |||
737 | ||||
738 | SrcReg = SuperReg; | |||
739 | } | |||
740 | ||||
741 | unsigned Opc; | |||
742 | if (PPC::GPRCRegClass.contains(DestReg, SrcReg)) | |||
743 | Opc = PPC::OR; | |||
744 | else if (PPC::G8RCRegClass.contains(DestReg, SrcReg)) | |||
745 | Opc = PPC::OR8; | |||
746 | else if (PPC::F4RCRegClass.contains(DestReg, SrcReg)) | |||
747 | Opc = PPC::FMR; | |||
748 | else if (PPC::CRRCRegClass.contains(DestReg, SrcReg)) | |||
749 | Opc = PPC::MCRF; | |||
750 | else if (PPC::VRRCRegClass.contains(DestReg, SrcReg)) | |||
751 | Opc = PPC::VOR; | |||
752 | else if (PPC::VSRCRegClass.contains(DestReg, SrcReg)) | |||
753 | // There are two different ways this can be done: | |||
754 | // 1. xxlor : This has lower latency (on the P7), 2 cycles, but can only | |||
755 | // issue in VSU pipeline 0. | |||
756 | // 2. xmovdp/xmovsp: This has higher latency (on the P7), 6 cycles, but | |||
757 | // can go to either pipeline. | |||
758 | // We'll always use xxlor here, because in practically all cases where | |||
759 | // copies are generated, they are close enough to some use that the | |||
760 | // lower-latency form is preferable. | |||
761 | Opc = PPC::XXLOR; | |||
762 | else if (PPC::VSFRCRegClass.contains(DestReg, SrcReg)) | |||
763 | Opc = PPC::XXLORf; | |||
764 | else if (PPC::CRBITRCRegClass.contains(DestReg, SrcReg)) | |||
765 | Opc = PPC::CROR; | |||
766 | else | |||
767 | llvm_unreachable("Impossible reg-to-reg copy")::llvm::llvm_unreachable_internal("Impossible reg-to-reg copy" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 767); | |||
768 | ||||
769 | const MCInstrDesc &MCID = get(Opc); | |||
770 | if (MCID.getNumOperands() == 3) | |||
771 | BuildMI(MBB, I, DL, MCID, DestReg) | |||
772 | .addReg(SrcReg).addReg(SrcReg, getKillRegState(KillSrc)); | |||
773 | else | |||
774 | BuildMI(MBB, I, DL, MCID, DestReg).addReg(SrcReg, getKillRegState(KillSrc)); | |||
775 | } | |||
776 | ||||
777 | // This function returns true if a CR spill is necessary and false otherwise. | |||
778 | bool | |||
779 | PPCInstrInfo::StoreRegToStackSlot(MachineFunction &MF, | |||
780 | unsigned SrcReg, bool isKill, | |||
781 | int FrameIdx, | |||
782 | const TargetRegisterClass *RC, | |||
783 | SmallVectorImpl<MachineInstr*> &NewMIs, | |||
784 | bool &NonRI, bool &SpillsVRS) const{ | |||
785 | // Note: If additional store instructions are added here, | |||
786 | // update isStoreToStackSlot. | |||
787 | ||||
788 | DebugLoc DL; | |||
789 | if (PPC::GPRCRegClass.hasSubClassEq(RC) || | |||
790 | PPC::GPRC_NOR0RegClass.hasSubClassEq(RC)) { | |||
791 | NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STW)) | |||
792 | .addReg(SrcReg, | |||
793 | getKillRegState(isKill)), | |||
794 | FrameIdx)); | |||
795 | } else if (PPC::G8RCRegClass.hasSubClassEq(RC) || | |||
796 | PPC::G8RC_NOX0RegClass.hasSubClassEq(RC)) { | |||
797 | NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STD)) | |||
798 | .addReg(SrcReg, | |||
799 | getKillRegState(isKill)), | |||
800 | FrameIdx)); | |||
801 | } else if (PPC::F8RCRegClass.hasSubClassEq(RC)) { | |||
802 | NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STFD)) | |||
803 | .addReg(SrcReg, | |||
804 | getKillRegState(isKill)), | |||
805 | FrameIdx)); | |||
806 | } else if (PPC::F4RCRegClass.hasSubClassEq(RC)) { | |||
807 | NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STFS)) | |||
808 | .addReg(SrcReg, | |||
809 | getKillRegState(isKill)), | |||
810 | FrameIdx)); | |||
811 | } else if (PPC::CRRCRegClass.hasSubClassEq(RC)) { | |||
812 | NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::SPILL_CR)) | |||
813 | .addReg(SrcReg, | |||
814 | getKillRegState(isKill)), | |||
815 | FrameIdx)); | |||
816 | return true; | |||
817 | } else if (PPC::CRBITRCRegClass.hasSubClassEq(RC)) { | |||
818 | NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::SPILL_CRBIT)) | |||
819 | .addReg(SrcReg, | |||
820 | getKillRegState(isKill)), | |||
821 | FrameIdx)); | |||
822 | return true; | |||
823 | } else if (PPC::VRRCRegClass.hasSubClassEq(RC)) { | |||
824 | NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STVX)) | |||
825 | .addReg(SrcReg, | |||
826 | getKillRegState(isKill)), | |||
827 | FrameIdx)); | |||
828 | NonRI = true; | |||
829 | } else if (PPC::VSRCRegClass.hasSubClassEq(RC)) { | |||
830 | NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STXVD2X)) | |||
831 | .addReg(SrcReg, | |||
832 | getKillRegState(isKill)), | |||
833 | FrameIdx)); | |||
834 | NonRI = true; | |||
835 | } else if (PPC::VSFRCRegClass.hasSubClassEq(RC)) { | |||
836 | NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STXSDX)) | |||
837 | .addReg(SrcReg, | |||
838 | getKillRegState(isKill)), | |||
839 | FrameIdx)); | |||
840 | NonRI = true; | |||
841 | } else if (PPC::VRSAVERCRegClass.hasSubClassEq(RC)) { | |||
842 | assert(Subtarget.isDarwin() &&((Subtarget.isDarwin() && "VRSAVE only needs spill/restore on Darwin" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.isDarwin() && \"VRSAVE only needs spill/restore on Darwin\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 843, __PRETTY_FUNCTION__)) | |||
843 | "VRSAVE only needs spill/restore on Darwin")((Subtarget.isDarwin() && "VRSAVE only needs spill/restore on Darwin" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.isDarwin() && \"VRSAVE only needs spill/restore on Darwin\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 843, __PRETTY_FUNCTION__)); | |||
844 | NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::SPILL_VRSAVE)) | |||
845 | .addReg(SrcReg, | |||
846 | getKillRegState(isKill)), | |||
847 | FrameIdx)); | |||
848 | SpillsVRS = true; | |||
849 | } else { | |||
850 | llvm_unreachable("Unknown regclass!")::llvm::llvm_unreachable_internal("Unknown regclass!", "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 850); | |||
851 | } | |||
852 | ||||
853 | return false; | |||
854 | } | |||
855 | ||||
856 | void | |||
857 | PPCInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB, | |||
858 | MachineBasicBlock::iterator MI, | |||
859 | unsigned SrcReg, bool isKill, int FrameIdx, | |||
860 | const TargetRegisterClass *RC, | |||
861 | const TargetRegisterInfo *TRI) const { | |||
862 | MachineFunction &MF = *MBB.getParent(); | |||
863 | SmallVector<MachineInstr*, 4> NewMIs; | |||
864 | ||||
865 | PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>(); | |||
866 | FuncInfo->setHasSpills(); | |||
867 | ||||
868 | bool NonRI = false, SpillsVRS = false; | |||
869 | if (StoreRegToStackSlot(MF, SrcReg, isKill, FrameIdx, RC, NewMIs, | |||
870 | NonRI, SpillsVRS)) | |||
871 | FuncInfo->setSpillsCR(); | |||
872 | ||||
873 | if (SpillsVRS) | |||
874 | FuncInfo->setSpillsVRSAVE(); | |||
875 | ||||
876 | if (NonRI) | |||
877 | FuncInfo->setHasNonRISpills(); | |||
878 | ||||
879 | for (unsigned i = 0, e = NewMIs.size(); i != e; ++i) | |||
880 | MBB.insert(MI, NewMIs[i]); | |||
881 | ||||
882 | const MachineFrameInfo &MFI = *MF.getFrameInfo(); | |||
883 | MachineMemOperand *MMO = | |||
884 | MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FrameIdx), | |||
885 | MachineMemOperand::MOStore, | |||
886 | MFI.getObjectSize(FrameIdx), | |||
887 | MFI.getObjectAlignment(FrameIdx)); | |||
888 | NewMIs.back()->addMemOperand(MF, MMO); | |||
889 | } | |||
890 | ||||
891 | bool | |||
892 | PPCInstrInfo::LoadRegFromStackSlot(MachineFunction &MF, DebugLoc DL, | |||
893 | unsigned DestReg, int FrameIdx, | |||
894 | const TargetRegisterClass *RC, | |||
895 | SmallVectorImpl<MachineInstr*> &NewMIs, | |||
896 | bool &NonRI, bool &SpillsVRS) const{ | |||
897 | // Note: If additional load instructions are added here, | |||
898 | // update isLoadFromStackSlot. | |||
899 | ||||
900 | if (PPC::GPRCRegClass.hasSubClassEq(RC) || | |||
901 | PPC::GPRC_NOR0RegClass.hasSubClassEq(RC)) { | |||
902 | NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LWZ), | |||
903 | DestReg), FrameIdx)); | |||
904 | } else if (PPC::G8RCRegClass.hasSubClassEq(RC) || | |||
905 | PPC::G8RC_NOX0RegClass.hasSubClassEq(RC)) { | |||
906 | NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LD), DestReg), | |||
907 | FrameIdx)); | |||
908 | } else if (PPC::F8RCRegClass.hasSubClassEq(RC)) { | |||
909 | NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LFD), DestReg), | |||
910 | FrameIdx)); | |||
911 | } else if (PPC::F4RCRegClass.hasSubClassEq(RC)) { | |||
912 | NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LFS), DestReg), | |||
913 | FrameIdx)); | |||
914 | } else if (PPC::CRRCRegClass.hasSubClassEq(RC)) { | |||
915 | NewMIs.push_back(addFrameReference(BuildMI(MF, DL, | |||
916 | get(PPC::RESTORE_CR), DestReg), | |||
917 | FrameIdx)); | |||
918 | return true; | |||
919 | } else if (PPC::CRBITRCRegClass.hasSubClassEq(RC)) { | |||
920 | NewMIs.push_back(addFrameReference(BuildMI(MF, DL, | |||
921 | get(PPC::RESTORE_CRBIT), DestReg), | |||
922 | FrameIdx)); | |||
923 | return true; | |||
924 | } else if (PPC::VRRCRegClass.hasSubClassEq(RC)) { | |||
925 | NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LVX), DestReg), | |||
926 | FrameIdx)); | |||
927 | NonRI = true; | |||
928 | } else if (PPC::VSRCRegClass.hasSubClassEq(RC)) { | |||
929 | NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LXVD2X), DestReg), | |||
930 | FrameIdx)); | |||
931 | NonRI = true; | |||
932 | } else if (PPC::VSFRCRegClass.hasSubClassEq(RC)) { | |||
933 | NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LXSDX), DestReg), | |||
934 | FrameIdx)); | |||
935 | NonRI = true; | |||
936 | } else if (PPC::VRSAVERCRegClass.hasSubClassEq(RC)) { | |||
937 | assert(Subtarget.isDarwin() &&((Subtarget.isDarwin() && "VRSAVE only needs spill/restore on Darwin" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.isDarwin() && \"VRSAVE only needs spill/restore on Darwin\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 938, __PRETTY_FUNCTION__)) | |||
938 | "VRSAVE only needs spill/restore on Darwin")((Subtarget.isDarwin() && "VRSAVE only needs spill/restore on Darwin" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.isDarwin() && \"VRSAVE only needs spill/restore on Darwin\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 938, __PRETTY_FUNCTION__)); | |||
939 | NewMIs.push_back(addFrameReference(BuildMI(MF, DL, | |||
940 | get(PPC::RESTORE_VRSAVE), | |||
941 | DestReg), | |||
942 | FrameIdx)); | |||
943 | SpillsVRS = true; | |||
944 | } else { | |||
945 | llvm_unreachable("Unknown regclass!")::llvm::llvm_unreachable_internal("Unknown regclass!", "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 945); | |||
946 | } | |||
947 | ||||
948 | return false; | |||
949 | } | |||
950 | ||||
951 | void | |||
952 | PPCInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB, | |||
953 | MachineBasicBlock::iterator MI, | |||
954 | unsigned DestReg, int FrameIdx, | |||
955 | const TargetRegisterClass *RC, | |||
956 | const TargetRegisterInfo *TRI) const { | |||
957 | MachineFunction &MF = *MBB.getParent(); | |||
958 | SmallVector<MachineInstr*, 4> NewMIs; | |||
959 | DebugLoc DL; | |||
960 | if (MI != MBB.end()) DL = MI->getDebugLoc(); | |||
961 | ||||
962 | PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>(); | |||
963 | FuncInfo->setHasSpills(); | |||
964 | ||||
965 | bool NonRI = false, SpillsVRS = false; | |||
966 | if (LoadRegFromStackSlot(MF, DL, DestReg, FrameIdx, RC, NewMIs, | |||
967 | NonRI, SpillsVRS)) | |||
968 | FuncInfo->setSpillsCR(); | |||
969 | ||||
970 | if (SpillsVRS) | |||
971 | FuncInfo->setSpillsVRSAVE(); | |||
972 | ||||
973 | if (NonRI) | |||
974 | FuncInfo->setHasNonRISpills(); | |||
975 | ||||
976 | for (unsigned i = 0, e = NewMIs.size(); i != e; ++i) | |||
977 | MBB.insert(MI, NewMIs[i]); | |||
978 | ||||
979 | const MachineFrameInfo &MFI = *MF.getFrameInfo(); | |||
980 | MachineMemOperand *MMO = | |||
981 | MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FrameIdx), | |||
982 | MachineMemOperand::MOLoad, | |||
983 | MFI.getObjectSize(FrameIdx), | |||
984 | MFI.getObjectAlignment(FrameIdx)); | |||
985 | NewMIs.back()->addMemOperand(MF, MMO); | |||
986 | } | |||
987 | ||||
988 | bool PPCInstrInfo:: | |||
989 | ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const { | |||
990 | assert(Cond.size() == 2 && "Invalid PPC branch opcode!")((Cond.size() == 2 && "Invalid PPC branch opcode!") ? static_cast<void> (0) : __assert_fail ("Cond.size() == 2 && \"Invalid PPC branch opcode!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 990, __PRETTY_FUNCTION__)); | |||
991 | if (Cond[1].getReg() == PPC::CTR8 || Cond[1].getReg() == PPC::CTR) | |||
992 | Cond[0].setImm(Cond[0].getImm() == 0 ? 1 : 0); | |||
993 | else | |||
994 | // Leave the CR# the same, but invert the condition. | |||
995 | Cond[0].setImm(PPC::InvertPredicate((PPC::Predicate)Cond[0].getImm())); | |||
996 | return false; | |||
997 | } | |||
998 | ||||
999 | bool PPCInstrInfo::FoldImmediate(MachineInstr *UseMI, MachineInstr *DefMI, | |||
1000 | unsigned Reg, MachineRegisterInfo *MRI) const { | |||
1001 | // For some instructions, it is legal to fold ZERO into the RA register field. | |||
1002 | // A zero immediate should always be loaded with a single li. | |||
1003 | unsigned DefOpc = DefMI->getOpcode(); | |||
1004 | if (DefOpc != PPC::LI && DefOpc != PPC::LI8) | |||
1005 | return false; | |||
1006 | if (!DefMI->getOperand(1).isImm()) | |||
1007 | return false; | |||
1008 | if (DefMI->getOperand(1).getImm() != 0) | |||
1009 | return false; | |||
1010 | ||||
1011 | // Note that we cannot here invert the arguments of an isel in order to fold | |||
1012 | // a ZERO into what is presented as the second argument. All we have here | |||
1013 | // is the condition bit, and that might come from a CR-logical bit operation. | |||
1014 | ||||
1015 | const MCInstrDesc &UseMCID = UseMI->getDesc(); | |||
1016 | ||||
1017 | // Only fold into real machine instructions. | |||
1018 | if (UseMCID.isPseudo()) | |||
1019 | return false; | |||
1020 | ||||
1021 | unsigned UseIdx; | |||
1022 | for (UseIdx = 0; UseIdx < UseMI->getNumOperands(); ++UseIdx) | |||
1023 | if (UseMI->getOperand(UseIdx).isReg() && | |||
1024 | UseMI->getOperand(UseIdx).getReg() == Reg) | |||
1025 | break; | |||
1026 | ||||
1027 | assert(UseIdx < UseMI->getNumOperands() && "Cannot find Reg in UseMI")((UseIdx < UseMI->getNumOperands() && "Cannot find Reg in UseMI" ) ? static_cast<void> (0) : __assert_fail ("UseIdx < UseMI->getNumOperands() && \"Cannot find Reg in UseMI\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 1027, __PRETTY_FUNCTION__)); | |||
1028 | assert(UseIdx < UseMCID.getNumOperands() && "No operand description for Reg")((UseIdx < UseMCID.getNumOperands() && "No operand description for Reg" ) ? static_cast<void> (0) : __assert_fail ("UseIdx < UseMCID.getNumOperands() && \"No operand description for Reg\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 1028, __PRETTY_FUNCTION__)); | |||
1029 | ||||
1030 | const MCOperandInfo *UseInfo = &UseMCID.OpInfo[UseIdx]; | |||
1031 | ||||
1032 | // We can fold the zero if this register requires a GPRC_NOR0/G8RC_NOX0 | |||
1033 | // register (which might also be specified as a pointer class kind). | |||
1034 | if (UseInfo->isLookupPtrRegClass()) { | |||
1035 | if (UseInfo->RegClass /* Kind */ != 1) | |||
1036 | return false; | |||
1037 | } else { | |||
1038 | if (UseInfo->RegClass != PPC::GPRC_NOR0RegClassID && | |||
1039 | UseInfo->RegClass != PPC::G8RC_NOX0RegClassID) | |||
1040 | return false; | |||
1041 | } | |||
1042 | ||||
1043 | // Make sure this is not tied to an output register (or otherwise | |||
1044 | // constrained). This is true for ST?UX registers, for example, which | |||
1045 | // are tied to their output registers. | |||
1046 | if (UseInfo->Constraints != 0) | |||
1047 | return false; | |||
1048 | ||||
1049 | unsigned ZeroReg; | |||
1050 | if (UseInfo->isLookupPtrRegClass()) { | |||
1051 | bool isPPC64 = Subtarget.isPPC64(); | |||
1052 | ZeroReg = isPPC64 ? PPC::ZERO8 : PPC::ZERO; | |||
1053 | } else { | |||
1054 | ZeroReg = UseInfo->RegClass == PPC::G8RC_NOX0RegClassID ? | |||
1055 | PPC::ZERO8 : PPC::ZERO; | |||
1056 | } | |||
1057 | ||||
1058 | bool DeleteDef = MRI->hasOneNonDBGUse(Reg); | |||
1059 | UseMI->getOperand(UseIdx).setReg(ZeroReg); | |||
1060 | ||||
1061 | if (DeleteDef) | |||
1062 | DefMI->eraseFromParent(); | |||
1063 | ||||
1064 | return true; | |||
1065 | } | |||
1066 | ||||
1067 | static bool MBBDefinesCTR(MachineBasicBlock &MBB) { | |||
1068 | for (MachineBasicBlock::iterator I = MBB.begin(), IE = MBB.end(); | |||
1069 | I != IE; ++I) | |||
1070 | if (I->definesRegister(PPC::CTR) || I->definesRegister(PPC::CTR8)) | |||
1071 | return true; | |||
1072 | return false; | |||
1073 | } | |||
1074 | ||||
1075 | // We should make sure that, if we're going to predicate both sides of a | |||
1076 | // condition (a diamond), that both sides don't define the counter register. We | |||
1077 | // can predicate counter-decrement-based branches, but while that predicates | |||
1078 | // the branching, it does not predicate the counter decrement. If we tried to | |||
1079 | // merge the triangle into one predicated block, we'd decrement the counter | |||
1080 | // twice. | |||
1081 | bool PPCInstrInfo::isProfitableToIfCvt(MachineBasicBlock &TMBB, | |||
1082 | unsigned NumT, unsigned ExtraT, | |||
1083 | MachineBasicBlock &FMBB, | |||
1084 | unsigned NumF, unsigned ExtraF, | |||
1085 | const BranchProbability &Probability) const { | |||
1086 | return !(MBBDefinesCTR(TMBB) && MBBDefinesCTR(FMBB)); | |||
1087 | } | |||
1088 | ||||
1089 | ||||
1090 | bool PPCInstrInfo::isPredicated(const MachineInstr *MI) const { | |||
1091 | // The predicated branches are identified by their type, not really by the | |||
1092 | // explicit presence of a predicate. Furthermore, some of them can be | |||
1093 | // predicated more than once. Because if conversion won't try to predicate | |||
1094 | // any instruction which already claims to be predicated (by returning true | |||
1095 | // here), always return false. In doing so, we let isPredicable() be the | |||
1096 | // final word on whether not the instruction can be (further) predicated. | |||
1097 | ||||
1098 | return false; | |||
1099 | } | |||
1100 | ||||
1101 | bool PPCInstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const { | |||
1102 | if (!MI->isTerminator()) | |||
1103 | return false; | |||
1104 | ||||
1105 | // Conditional branch is a special case. | |||
1106 | if (MI->isBranch() && !MI->isBarrier()) | |||
1107 | return true; | |||
1108 | ||||
1109 | return !isPredicated(MI); | |||
1110 | } | |||
1111 | ||||
1112 | bool PPCInstrInfo::PredicateInstruction( | |||
1113 | MachineInstr *MI, | |||
1114 | const SmallVectorImpl<MachineOperand> &Pred) const { | |||
1115 | unsigned OpC = MI->getOpcode(); | |||
1116 | if (OpC == PPC::BLR) { | |||
1117 | if (Pred[1].getReg() == PPC::CTR8 || Pred[1].getReg() == PPC::CTR) { | |||
1118 | bool isPPC64 = Subtarget.isPPC64(); | |||
1119 | MI->setDesc(get(Pred[0].getImm() ? | |||
1120 | (isPPC64 ? PPC::BDNZLR8 : PPC::BDNZLR) : | |||
1121 | (isPPC64 ? PPC::BDZLR8 : PPC::BDZLR))); | |||
1122 | } else if (Pred[0].getImm() == PPC::PRED_BIT_SET) { | |||
1123 | MI->setDesc(get(PPC::BCLR)); | |||
1124 | MachineInstrBuilder(*MI->getParent()->getParent(), MI) | |||
1125 | .addReg(Pred[1].getReg()); | |||
1126 | } else if (Pred[0].getImm() == PPC::PRED_BIT_UNSET) { | |||
1127 | MI->setDesc(get(PPC::BCLRn)); | |||
1128 | MachineInstrBuilder(*MI->getParent()->getParent(), MI) | |||
1129 | .addReg(Pred[1].getReg()); | |||
1130 | } else { | |||
1131 | MI->setDesc(get(PPC::BCCLR)); | |||
1132 | MachineInstrBuilder(*MI->getParent()->getParent(), MI) | |||
1133 | .addImm(Pred[0].getImm()) | |||
1134 | .addReg(Pred[1].getReg()); | |||
1135 | } | |||
1136 | ||||
1137 | return true; | |||
1138 | } else if (OpC == PPC::B) { | |||
1139 | if (Pred[1].getReg() == PPC::CTR8 || Pred[1].getReg() == PPC::CTR) { | |||
1140 | bool isPPC64 = Subtarget.isPPC64(); | |||
1141 | MI->setDesc(get(Pred[0].getImm() ? | |||
1142 | (isPPC64 ? PPC::BDNZ8 : PPC::BDNZ) : | |||
1143 | (isPPC64 ? PPC::BDZ8 : PPC::BDZ))); | |||
1144 | } else if (Pred[0].getImm() == PPC::PRED_BIT_SET) { | |||
1145 | MachineBasicBlock *MBB = MI->getOperand(0).getMBB(); | |||
1146 | MI->RemoveOperand(0); | |||
1147 | ||||
1148 | MI->setDesc(get(PPC::BC)); | |||
1149 | MachineInstrBuilder(*MI->getParent()->getParent(), MI) | |||
1150 | .addReg(Pred[1].getReg()) | |||
1151 | .addMBB(MBB); | |||
1152 | } else if (Pred[0].getImm() == PPC::PRED_BIT_UNSET) { | |||
1153 | MachineBasicBlock *MBB = MI->getOperand(0).getMBB(); | |||
1154 | MI->RemoveOperand(0); | |||
1155 | ||||
1156 | MI->setDesc(get(PPC::BCn)); | |||
1157 | MachineInstrBuilder(*MI->getParent()->getParent(), MI) | |||
1158 | .addReg(Pred[1].getReg()) | |||
1159 | .addMBB(MBB); | |||
1160 | } else { | |||
1161 | MachineBasicBlock *MBB = MI->getOperand(0).getMBB(); | |||
1162 | MI->RemoveOperand(0); | |||
1163 | ||||
1164 | MI->setDesc(get(PPC::BCC)); | |||
1165 | MachineInstrBuilder(*MI->getParent()->getParent(), MI) | |||
1166 | .addImm(Pred[0].getImm()) | |||
1167 | .addReg(Pred[1].getReg()) | |||
1168 | .addMBB(MBB); | |||
1169 | } | |||
1170 | ||||
1171 | return true; | |||
1172 | } else if (OpC == PPC::BCTR || OpC == PPC::BCTR8 || | |||
1173 | OpC == PPC::BCTRL || OpC == PPC::BCTRL8) { | |||
1174 | if (Pred[1].getReg() == PPC::CTR8 || Pred[1].getReg() == PPC::CTR) | |||
1175 | llvm_unreachable("Cannot predicate bctr[l] on the ctr register")::llvm::llvm_unreachable_internal("Cannot predicate bctr[l] on the ctr register" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 1175); | |||
1176 | ||||
1177 | bool setLR = OpC == PPC::BCTRL || OpC == PPC::BCTRL8; | |||
1178 | bool isPPC64 = Subtarget.isPPC64(); | |||
1179 | ||||
1180 | if (Pred[0].getImm() == PPC::PRED_BIT_SET) { | |||
1181 | MI->setDesc(get(isPPC64 ? (setLR ? PPC::BCCTRL8 : PPC::BCCTR8) : | |||
1182 | (setLR ? PPC::BCCTRL : PPC::BCCTR))); | |||
1183 | MachineInstrBuilder(*MI->getParent()->getParent(), MI) | |||
1184 | .addReg(Pred[1].getReg()); | |||
1185 | return true; | |||
1186 | } else if (Pred[0].getImm() == PPC::PRED_BIT_UNSET) { | |||
1187 | MI->setDesc(get(isPPC64 ? (setLR ? PPC::BCCTRL8n : PPC::BCCTR8n) : | |||
1188 | (setLR ? PPC::BCCTRLn : PPC::BCCTRn))); | |||
1189 | MachineInstrBuilder(*MI->getParent()->getParent(), MI) | |||
1190 | .addReg(Pred[1].getReg()); | |||
1191 | return true; | |||
1192 | } | |||
1193 | ||||
1194 | MI->setDesc(get(isPPC64 ? (setLR ? PPC::BCCCTRL8 : PPC::BCCCTR8) : | |||
1195 | (setLR ? PPC::BCCCTRL : PPC::BCCCTR))); | |||
1196 | MachineInstrBuilder(*MI->getParent()->getParent(), MI) | |||
1197 | .addImm(Pred[0].getImm()) | |||
1198 | .addReg(Pred[1].getReg()); | |||
1199 | return true; | |||
1200 | } | |||
1201 | ||||
1202 | return false; | |||
1203 | } | |||
1204 | ||||
1205 | bool PPCInstrInfo::SubsumesPredicate( | |||
1206 | const SmallVectorImpl<MachineOperand> &Pred1, | |||
1207 | const SmallVectorImpl<MachineOperand> &Pred2) const { | |||
1208 | assert(Pred1.size() == 2 && "Invalid PPC first predicate")((Pred1.size() == 2 && "Invalid PPC first predicate") ? static_cast<void> (0) : __assert_fail ("Pred1.size() == 2 && \"Invalid PPC first predicate\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 1208, __PRETTY_FUNCTION__)); | |||
1209 | assert(Pred2.size() == 2 && "Invalid PPC second predicate")((Pred2.size() == 2 && "Invalid PPC second predicate" ) ? static_cast<void> (0) : __assert_fail ("Pred2.size() == 2 && \"Invalid PPC second predicate\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 1209, __PRETTY_FUNCTION__)); | |||
1210 | ||||
1211 | if (Pred1[1].getReg() == PPC::CTR8 || Pred1[1].getReg() == PPC::CTR) | |||
1212 | return false; | |||
1213 | if (Pred2[1].getReg() == PPC::CTR8 || Pred2[1].getReg() == PPC::CTR) | |||
1214 | return false; | |||
1215 | ||||
1216 | // P1 can only subsume P2 if they test the same condition register. | |||
1217 | if (Pred1[1].getReg() != Pred2[1].getReg()) | |||
1218 | return false; | |||
1219 | ||||
1220 | PPC::Predicate P1 = (PPC::Predicate) Pred1[0].getImm(); | |||
1221 | PPC::Predicate P2 = (PPC::Predicate) Pred2[0].getImm(); | |||
1222 | ||||
1223 | if (P1 == P2) | |||
1224 | return true; | |||
1225 | ||||
1226 | // Does P1 subsume P2, e.g. GE subsumes GT. | |||
1227 | if (P1 == PPC::PRED_LE && | |||
1228 | (P2 == PPC::PRED_LT || P2 == PPC::PRED_EQ)) | |||
1229 | return true; | |||
1230 | if (P1 == PPC::PRED_GE && | |||
1231 | (P2 == PPC::PRED_GT || P2 == PPC::PRED_EQ)) | |||
1232 | return true; | |||
1233 | ||||
1234 | return false; | |||
1235 | } | |||
1236 | ||||
1237 | bool PPCInstrInfo::DefinesPredicate(MachineInstr *MI, | |||
1238 | std::vector<MachineOperand> &Pred) const { | |||
1239 | // Note: At the present time, the contents of Pred from this function is | |||
1240 | // unused by IfConversion. This implementation follows ARM by pushing the | |||
1241 | // CR-defining operand. Because the 'DZ' and 'DNZ' count as types of | |||
1242 | // predicate, instructions defining CTR or CTR8 are also included as | |||
1243 | // predicate-defining instructions. | |||
1244 | ||||
1245 | const TargetRegisterClass *RCs[] = | |||
1246 | { &PPC::CRRCRegClass, &PPC::CRBITRCRegClass, | |||
1247 | &PPC::CTRRCRegClass, &PPC::CTRRC8RegClass }; | |||
1248 | ||||
1249 | bool Found = false; | |||
1250 | for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { | |||
1251 | const MachineOperand &MO = MI->getOperand(i); | |||
1252 | for (unsigned c = 0; c < array_lengthof(RCs) && !Found; ++c) { | |||
1253 | const TargetRegisterClass *RC = RCs[c]; | |||
1254 | if (MO.isReg()) { | |||
1255 | if (MO.isDef() && RC->contains(MO.getReg())) { | |||
1256 | Pred.push_back(MO); | |||
1257 | Found = true; | |||
1258 | } | |||
1259 | } else if (MO.isRegMask()) { | |||
1260 | for (TargetRegisterClass::iterator I = RC->begin(), | |||
1261 | IE = RC->end(); I != IE; ++I) | |||
1262 | if (MO.clobbersPhysReg(*I)) { | |||
1263 | Pred.push_back(MO); | |||
1264 | Found = true; | |||
1265 | } | |||
1266 | } | |||
1267 | } | |||
1268 | } | |||
1269 | ||||
1270 | return Found; | |||
1271 | } | |||
1272 | ||||
1273 | bool PPCInstrInfo::isPredicable(MachineInstr *MI) const { | |||
1274 | unsigned OpC = MI->getOpcode(); | |||
1275 | switch (OpC) { | |||
1276 | default: | |||
1277 | return false; | |||
1278 | case PPC::B: | |||
1279 | case PPC::BLR: | |||
1280 | case PPC::BCTR: | |||
1281 | case PPC::BCTR8: | |||
1282 | case PPC::BCTRL: | |||
1283 | case PPC::BCTRL8: | |||
1284 | return true; | |||
1285 | } | |||
1286 | } | |||
1287 | ||||
1288 | bool PPCInstrInfo::analyzeCompare(const MachineInstr *MI, | |||
1289 | unsigned &SrcReg, unsigned &SrcReg2, | |||
1290 | int &Mask, int &Value) const { | |||
1291 | unsigned Opc = MI->getOpcode(); | |||
1292 | ||||
1293 | switch (Opc) { | |||
1294 | default: return false; | |||
1295 | case PPC::CMPWI: | |||
1296 | case PPC::CMPLWI: | |||
1297 | case PPC::CMPDI: | |||
1298 | case PPC::CMPLDI: | |||
1299 | SrcReg = MI->getOperand(1).getReg(); | |||
1300 | SrcReg2 = 0; | |||
1301 | Value = MI->getOperand(2).getImm(); | |||
1302 | Mask = 0xFFFF; | |||
1303 | return true; | |||
1304 | case PPC::CMPW: | |||
1305 | case PPC::CMPLW: | |||
1306 | case PPC::CMPD: | |||
1307 | case PPC::CMPLD: | |||
1308 | case PPC::FCMPUS: | |||
1309 | case PPC::FCMPUD: | |||
1310 | SrcReg = MI->getOperand(1).getReg(); | |||
1311 | SrcReg2 = MI->getOperand(2).getReg(); | |||
1312 | return true; | |||
1313 | } | |||
1314 | } | |||
1315 | ||||
1316 | bool PPCInstrInfo::optimizeCompareInstr(MachineInstr *CmpInstr, | |||
1317 | unsigned SrcReg, unsigned SrcReg2, | |||
1318 | int Mask, int Value, | |||
1319 | const MachineRegisterInfo *MRI) const { | |||
1320 | if (DisableCmpOpt) | |||
1321 | return false; | |||
1322 | ||||
1323 | int OpC = CmpInstr->getOpcode(); | |||
1324 | unsigned CRReg = CmpInstr->getOperand(0).getReg(); | |||
1325 | ||||
1326 | // FP record forms set CR1 based on the execption status bits, not a | |||
1327 | // comparison with zero. | |||
1328 | if (OpC == PPC::FCMPUS || OpC == PPC::FCMPUD) | |||
1329 | return false; | |||
1330 | ||||
1331 | // The record forms set the condition register based on a signed comparison | |||
1332 | // with zero (so says the ISA manual). This is not as straightforward as it | |||
1333 | // seems, however, because this is always a 64-bit comparison on PPC64, even | |||
1334 | // for instructions that are 32-bit in nature (like slw for example). | |||
1335 | // So, on PPC32, for unsigned comparisons, we can use the record forms only | |||
1336 | // for equality checks (as those don't depend on the sign). On PPC64, | |||
1337 | // we are restricted to equality for unsigned 64-bit comparisons and for | |||
1338 | // signed 32-bit comparisons the applicability is more restricted. | |||
1339 | bool isPPC64 = Subtarget.isPPC64(); | |||
1340 | bool is32BitSignedCompare = OpC == PPC::CMPWI || OpC == PPC::CMPW; | |||
1341 | bool is32BitUnsignedCompare = OpC == PPC::CMPLWI || OpC == PPC::CMPLW; | |||
1342 | bool is64BitUnsignedCompare = OpC == PPC::CMPLDI || OpC == PPC::CMPLD; | |||
1343 | ||||
1344 | // Get the unique definition of SrcReg. | |||
1345 | MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg); | |||
1346 | if (!MI) return false; | |||
1347 | int MIOpC = MI->getOpcode(); | |||
1348 | ||||
1349 | bool equalityOnly = false; | |||
1350 | bool noSub = false; | |||
1351 | if (isPPC64) { | |||
1352 | if (is32BitSignedCompare) { | |||
1353 | // We can perform this optimization only if MI is sign-extending. | |||
1354 | if (MIOpC == PPC::SRAW || MIOpC == PPC::SRAWo || | |||
1355 | MIOpC == PPC::SRAWI || MIOpC == PPC::SRAWIo || | |||
1356 | MIOpC == PPC::EXTSB || MIOpC == PPC::EXTSBo || | |||
1357 | MIOpC == PPC::EXTSH || MIOpC == PPC::EXTSHo || | |||
1358 | MIOpC == PPC::EXTSW || MIOpC == PPC::EXTSWo) { | |||
1359 | noSub = true; | |||
1360 | } else | |||
1361 | return false; | |||
1362 | } else if (is32BitUnsignedCompare) { | |||
1363 | // We can perform this optimization, equality only, if MI is | |||
1364 | // zero-extending. | |||
1365 | if (MIOpC == PPC::CNTLZW || MIOpC == PPC::CNTLZWo || | |||
1366 | MIOpC == PPC::SLW || MIOpC == PPC::SLWo || | |||
1367 | MIOpC == PPC::SRW || MIOpC == PPC::SRWo) { | |||
1368 | noSub = true; | |||
1369 | equalityOnly = true; | |||
1370 | } else | |||
1371 | return false; | |||
1372 | } else | |||
1373 | equalityOnly = is64BitUnsignedCompare; | |||
1374 | } else | |||
1375 | equalityOnly = is32BitUnsignedCompare; | |||
1376 | ||||
1377 | if (equalityOnly) { | |||
1378 | // We need to check the uses of the condition register in order to reject | |||
1379 | // non-equality comparisons. | |||
1380 | for (MachineRegisterInfo::use_instr_iterator I =MRI->use_instr_begin(CRReg), | |||
1381 | IE = MRI->use_instr_end(); I != IE; ++I) { | |||
1382 | MachineInstr *UseMI = &*I; | |||
1383 | if (UseMI->getOpcode() == PPC::BCC) { | |||
1384 | unsigned Pred = UseMI->getOperand(0).getImm(); | |||
1385 | if (Pred != PPC::PRED_EQ && Pred != PPC::PRED_NE) | |||
1386 | return false; | |||
1387 | } else if (UseMI->getOpcode() == PPC::ISEL || | |||
1388 | UseMI->getOpcode() == PPC::ISEL8) { | |||
1389 | unsigned SubIdx = UseMI->getOperand(3).getSubReg(); | |||
1390 | if (SubIdx != PPC::sub_eq) | |||
1391 | return false; | |||
1392 | } else | |||
1393 | return false; | |||
1394 | } | |||
1395 | } | |||
1396 | ||||
1397 | MachineBasicBlock::iterator I = CmpInstr; | |||
1398 | ||||
1399 | // Scan forward to find the first use of the compare. | |||
1400 | for (MachineBasicBlock::iterator EL = CmpInstr->getParent()->end(); | |||
1401 | I != EL; ++I) { | |||
1402 | bool FoundUse = false; | |||
1403 | for (MachineRegisterInfo::use_instr_iterator J =MRI->use_instr_begin(CRReg), | |||
1404 | JE = MRI->use_instr_end(); J != JE; ++J) | |||
1405 | if (&*J == &*I) { | |||
1406 | FoundUse = true; | |||
1407 | break; | |||
1408 | } | |||
1409 | ||||
1410 | if (FoundUse) | |||
1411 | break; | |||
1412 | } | |||
1413 | ||||
1414 | // There are two possible candidates which can be changed to set CR[01]. | |||
1415 | // One is MI, the other is a SUB instruction. | |||
1416 | // For CMPrr(r1,r2), we are looking for SUB(r1,r2) or SUB(r2,r1). | |||
1417 | MachineInstr *Sub = nullptr; | |||
1418 | if (SrcReg2 != 0) | |||
1419 | // MI is not a candidate for CMPrr. | |||
1420 | MI = nullptr; | |||
1421 | // FIXME: Conservatively refuse to convert an instruction which isn't in the | |||
1422 | // same BB as the comparison. This is to allow the check below to avoid calls | |||
1423 | // (and other explicit clobbers); instead we should really check for these | |||
1424 | // more explicitly (in at least a few predecessors). | |||
1425 | else if (MI->getParent() != CmpInstr->getParent() || Value != 0) { | |||
1426 | // PPC does not have a record-form SUBri. | |||
1427 | return false; | |||
1428 | } | |||
1429 | ||||
1430 | // Search for Sub. | |||
1431 | const TargetRegisterInfo *TRI = &getRegisterInfo(); | |||
1432 | --I; | |||
1433 | ||||
1434 | // Get ready to iterate backward from CmpInstr. | |||
1435 | MachineBasicBlock::iterator E = MI, | |||
1436 | B = CmpInstr->getParent()->begin(); | |||
1437 | ||||
1438 | for (; I != E && !noSub; --I) { | |||
1439 | const MachineInstr &Instr = *I; | |||
1440 | unsigned IOpC = Instr.getOpcode(); | |||
1441 | ||||
1442 | if (&*I != CmpInstr && ( | |||
1443 | Instr.modifiesRegister(PPC::CR0, TRI) || | |||
1444 | Instr.readsRegister(PPC::CR0, TRI))) | |||
1445 | // This instruction modifies or uses the record condition register after | |||
1446 | // the one we want to change. While we could do this transformation, it | |||
1447 | // would likely not be profitable. This transformation removes one | |||
1448 | // instruction, and so even forcing RA to generate one move probably | |||
1449 | // makes it unprofitable. | |||
1450 | return false; | |||
1451 | ||||
1452 | // Check whether CmpInstr can be made redundant by the current instruction. | |||
1453 | if ((OpC == PPC::CMPW || OpC == PPC::CMPLW || | |||
1454 | OpC == PPC::CMPD || OpC == PPC::CMPLD) && | |||
1455 | (IOpC == PPC::SUBF || IOpC == PPC::SUBF8) && | |||
1456 | ((Instr.getOperand(1).getReg() == SrcReg && | |||
1457 | Instr.getOperand(2).getReg() == SrcReg2) || | |||
1458 | (Instr.getOperand(1).getReg() == SrcReg2 && | |||
1459 | Instr.getOperand(2).getReg() == SrcReg))) { | |||
1460 | Sub = &*I; | |||
1461 | break; | |||
1462 | } | |||
1463 | ||||
1464 | if (I == B) | |||
1465 | // The 'and' is below the comparison instruction. | |||
1466 | return false; | |||
1467 | } | |||
1468 | ||||
1469 | // Return false if no candidates exist. | |||
1470 | if (!MI && !Sub) | |||
1471 | return false; | |||
1472 | ||||
1473 | // The single candidate is called MI. | |||
1474 | if (!MI) MI = Sub; | |||
1475 | ||||
1476 | int NewOpC = -1; | |||
1477 | MIOpC = MI->getOpcode(); | |||
1478 | if (MIOpC == PPC::ANDIo || MIOpC == PPC::ANDIo8) | |||
1479 | NewOpC = MIOpC; | |||
1480 | else { | |||
1481 | NewOpC = PPC::getRecordFormOpcode(MIOpC); | |||
1482 | if (NewOpC == -1 && PPC::getNonRecordFormOpcode(MIOpC) != -1) | |||
1483 | NewOpC = MIOpC; | |||
1484 | } | |||
1485 | ||||
1486 | // FIXME: On the non-embedded POWER architectures, only some of the record | |||
1487 | // forms are fast, and we should use only the fast ones. | |||
1488 | ||||
1489 | // The defining instruction has a record form (or is already a record | |||
1490 | // form). It is possible, however, that we'll need to reverse the condition | |||
1491 | // code of the users. | |||
1492 | if (NewOpC == -1) | |||
1493 | return false; | |||
1494 | ||||
1495 | SmallVector<std::pair<MachineOperand*, PPC::Predicate>, 4> PredsToUpdate; | |||
1496 | SmallVector<std::pair<MachineOperand*, unsigned>, 4> SubRegsToUpdate; | |||
1497 | ||||
1498 | // If we have SUB(r1, r2) and CMP(r2, r1), the condition code based on CMP | |||
1499 | // needs to be updated to be based on SUB. Push the condition code | |||
1500 | // operands to OperandsToUpdate. If it is safe to remove CmpInstr, the | |||
1501 | // condition code of these operands will be modified. | |||
1502 | bool ShouldSwap = false; | |||
1503 | if (Sub) { | |||
1504 | ShouldSwap = SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 && | |||
1505 | Sub->getOperand(2).getReg() == SrcReg; | |||
1506 | ||||
1507 | // The operands to subf are the opposite of sub, so only in the fixed-point | |||
1508 | // case, invert the order. | |||
1509 | ShouldSwap = !ShouldSwap; | |||
1510 | } | |||
1511 | ||||
1512 | if (ShouldSwap) | |||
1513 | for (MachineRegisterInfo::use_instr_iterator | |||
1514 | I = MRI->use_instr_begin(CRReg), IE = MRI->use_instr_end(); | |||
1515 | I != IE; ++I) { | |||
1516 | MachineInstr *UseMI = &*I; | |||
1517 | if (UseMI->getOpcode() == PPC::BCC) { | |||
1518 | PPC::Predicate Pred = (PPC::Predicate) UseMI->getOperand(0).getImm(); | |||
1519 | assert((!equalityOnly ||(((!equalityOnly || Pred == PPC::PRED_EQ || Pred == PPC::PRED_NE ) && "Invalid predicate for equality-only optimization" ) ? static_cast<void> (0) : __assert_fail ("(!equalityOnly || Pred == PPC::PRED_EQ || Pred == PPC::PRED_NE) && \"Invalid predicate for equality-only optimization\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 1521, __PRETTY_FUNCTION__)) | |||
1520 | Pred == PPC::PRED_EQ || Pred == PPC::PRED_NE) &&(((!equalityOnly || Pred == PPC::PRED_EQ || Pred == PPC::PRED_NE ) && "Invalid predicate for equality-only optimization" ) ? static_cast<void> (0) : __assert_fail ("(!equalityOnly || Pred == PPC::PRED_EQ || Pred == PPC::PRED_NE) && \"Invalid predicate for equality-only optimization\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 1521, __PRETTY_FUNCTION__)) | |||
1521 | "Invalid predicate for equality-only optimization")(((!equalityOnly || Pred == PPC::PRED_EQ || Pred == PPC::PRED_NE ) && "Invalid predicate for equality-only optimization" ) ? static_cast<void> (0) : __assert_fail ("(!equalityOnly || Pred == PPC::PRED_EQ || Pred == PPC::PRED_NE) && \"Invalid predicate for equality-only optimization\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 1521, __PRETTY_FUNCTION__)); | |||
1522 | PredsToUpdate.push_back(std::make_pair(&(UseMI->getOperand(0)), | |||
1523 | PPC::getSwappedPredicate(Pred))); | |||
1524 | } else if (UseMI->getOpcode() == PPC::ISEL || | |||
1525 | UseMI->getOpcode() == PPC::ISEL8) { | |||
1526 | unsigned NewSubReg = UseMI->getOperand(3).getSubReg(); | |||
1527 | assert((!equalityOnly || NewSubReg == PPC::sub_eq) &&(((!equalityOnly || NewSubReg == PPC::sub_eq) && "Invalid CR bit for equality-only optimization" ) ? static_cast<void> (0) : __assert_fail ("(!equalityOnly || NewSubReg == PPC::sub_eq) && \"Invalid CR bit for equality-only optimization\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 1528, __PRETTY_FUNCTION__)) | |||
1528 | "Invalid CR bit for equality-only optimization")(((!equalityOnly || NewSubReg == PPC::sub_eq) && "Invalid CR bit for equality-only optimization" ) ? static_cast<void> (0) : __assert_fail ("(!equalityOnly || NewSubReg == PPC::sub_eq) && \"Invalid CR bit for equality-only optimization\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 1528, __PRETTY_FUNCTION__)); | |||
1529 | ||||
1530 | if (NewSubReg == PPC::sub_lt) | |||
1531 | NewSubReg = PPC::sub_gt; | |||
1532 | else if (NewSubReg == PPC::sub_gt) | |||
1533 | NewSubReg = PPC::sub_lt; | |||
1534 | ||||
1535 | SubRegsToUpdate.push_back(std::make_pair(&(UseMI->getOperand(3)), | |||
1536 | NewSubReg)); | |||
1537 | } else // We need to abort on a user we don't understand. | |||
1538 | return false; | |||
1539 | } | |||
1540 | ||||
1541 | // Create a new virtual register to hold the value of the CR set by the | |||
1542 | // record-form instruction. If the instruction was not previously in | |||
1543 | // record form, then set the kill flag on the CR. | |||
1544 | CmpInstr->eraseFromParent(); | |||
1545 | ||||
1546 | MachineBasicBlock::iterator MII = MI; | |||
1547 | BuildMI(*MI->getParent(), std::next(MII), MI->getDebugLoc(), | |||
1548 | get(TargetOpcode::COPY), CRReg) | |||
1549 | .addReg(PPC::CR0, MIOpC != NewOpC ? RegState::Kill : 0); | |||
1550 | ||||
1551 | if (MIOpC != NewOpC) { | |||
1552 | // We need to be careful here: we're replacing one instruction with | |||
1553 | // another, and we need to make sure that we get all of the right | |||
1554 | // implicit uses and defs. On the other hand, the caller may be holding | |||
1555 | // an iterator to this instruction, and so we can't delete it (this is | |||
1556 | // specifically the case if this is the instruction directly after the | |||
1557 | // compare). | |||
1558 | ||||
1559 | const MCInstrDesc &NewDesc = get(NewOpC); | |||
1560 | MI->setDesc(NewDesc); | |||
1561 | ||||
1562 | if (NewDesc.ImplicitDefs) | |||
1563 | for (const uint16_t *ImpDefs = NewDesc.getImplicitDefs(); | |||
1564 | *ImpDefs; ++ImpDefs) | |||
1565 | if (!MI->definesRegister(*ImpDefs)) | |||
1566 | MI->addOperand(*MI->getParent()->getParent(), | |||
1567 | MachineOperand::CreateReg(*ImpDefs, true, true)); | |||
1568 | if (NewDesc.ImplicitUses) | |||
1569 | for (const uint16_t *ImpUses = NewDesc.getImplicitUses(); | |||
1570 | *ImpUses; ++ImpUses) | |||
1571 | if (!MI->readsRegister(*ImpUses)) | |||
1572 | MI->addOperand(*MI->getParent()->getParent(), | |||
1573 | MachineOperand::CreateReg(*ImpUses, false, true)); | |||
1574 | } | |||
1575 | ||||
1576 | // Modify the condition code of operands in OperandsToUpdate. | |||
1577 | // Since we have SUB(r1, r2) and CMP(r2, r1), the condition code needs to | |||
1578 | // be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc. | |||
1579 | for (unsigned i = 0, e = PredsToUpdate.size(); i < e; i++) | |||
1580 | PredsToUpdate[i].first->setImm(PredsToUpdate[i].second); | |||
1581 | ||||
1582 | for (unsigned i = 0, e = SubRegsToUpdate.size(); i < e; i++) | |||
1583 | SubRegsToUpdate[i].first->setSubReg(SubRegsToUpdate[i].second); | |||
1584 | ||||
1585 | return true; | |||
1586 | } | |||
1587 | ||||
1588 | /// GetInstSize - Return the number of bytes of code the specified | |||
1589 | /// instruction may be. This returns the maximum number of bytes. | |||
1590 | /// | |||
1591 | unsigned PPCInstrInfo::GetInstSizeInBytes(const MachineInstr *MI) const { | |||
1592 | unsigned Opcode = MI->getOpcode(); | |||
1593 | ||||
1594 | if (Opcode == PPC::INLINEASM) { | |||
1595 | const MachineFunction *MF = MI->getParent()->getParent(); | |||
1596 | const char *AsmStr = MI->getOperand(0).getSymbolName(); | |||
1597 | return getInlineAsmLength(AsmStr, *MF->getTarget().getMCAsmInfo()); | |||
1598 | } else { | |||
1599 | const MCInstrDesc &Desc = get(Opcode); | |||
1600 | return Desc.getSize(); | |||
1601 | } | |||
1602 | } | |||
1603 | ||||
1604 | #undef DEBUG_TYPE"ppc-early-ret" | |||
1605 | #define DEBUG_TYPE"ppc-early-ret" "ppc-vsx-fma-mutate" | |||
1606 | ||||
1607 | namespace { | |||
1608 | // PPCVSXFMAMutate pass - For copies between VSX registers and non-VSX registers | |||
1609 | // (Altivec and scalar floating-point registers), we need to transform the | |||
1610 | // copies into subregister copies with other restrictions. | |||
1611 | struct PPCVSXFMAMutate : public MachineFunctionPass { | |||
1612 | static char ID; | |||
1613 | PPCVSXFMAMutate() : MachineFunctionPass(ID) { | |||
1614 | initializePPCVSXFMAMutatePass(*PassRegistry::getPassRegistry()); | |||
1615 | } | |||
1616 | ||||
1617 | LiveIntervals *LIS; | |||
1618 | ||||
1619 | const PPCTargetMachine *TM; | |||
1620 | const PPCInstrInfo *TII; | |||
1621 | ||||
1622 | protected: | |||
1623 | bool processBlock(MachineBasicBlock &MBB) { | |||
1624 | bool Changed = false; | |||
1625 | ||||
1626 | MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); | |||
1627 | const TargetRegisterInfo *TRI = &TII->getRegisterInfo(); | |||
1628 | for (MachineBasicBlock::iterator I = MBB.begin(), IE = MBB.end(); | |||
1629 | I != IE; ++I) { | |||
1630 | MachineInstr *MI = I; | |||
1631 | ||||
1632 | // The default (A-type) VSX FMA form kills the addend (it is taken from | |||
1633 | // the target register, which is then updated to reflect the result of | |||
1634 | // the FMA). If the instruction, however, kills one of the registers | |||
1635 | // used for the product, then we can use the M-form instruction (which | |||
1636 | // will take that value from the to-be-defined register). | |||
1637 | ||||
1638 | int AltOpc = PPC::getAltVSXFMAOpcode(MI->getOpcode()); | |||
1639 | if (AltOpc == -1) | |||
1640 | continue; | |||
1641 | ||||
1642 | // This pass is run after register coalescing, and so we're looking for | |||
1643 | // a situation like this: | |||
1644 | // ... | |||
1645 | // %vreg5<def> = COPY %vreg9; VSLRC:%vreg5,%vreg9 | |||
1646 | // %vreg5<def,tied1> = XSMADDADP %vreg5<tied0>, %vreg17, %vreg16, | |||
1647 | // %RM<imp-use>; VSLRC:%vreg5,%vreg17,%vreg16 | |||
1648 | // ... | |||
1649 | // %vreg9<def,tied1> = XSMADDADP %vreg9<tied0>, %vreg17, %vreg19, | |||
1650 | // %RM<imp-use>; VSLRC:%vreg9,%vreg17,%vreg19 | |||
1651 | // ... | |||
1652 | // Where we can eliminate the copy by changing from the A-type to the | |||
1653 | // M-type instruction. Specifically, for this example, this means: | |||
1654 | // %vreg5<def,tied1> = XSMADDADP %vreg5<tied0>, %vreg17, %vreg16, | |||
1655 | // %RM<imp-use>; VSLRC:%vreg5,%vreg17,%vreg16 | |||
1656 | // is replaced by: | |||
1657 | // %vreg16<def,tied1> = XSMADDMDP %vreg16<tied0>, %vreg18, %vreg9, | |||
1658 | // %RM<imp-use>; VSLRC:%vreg16,%vreg18,%vreg9 | |||
1659 | // and we remove: %vreg5<def> = COPY %vreg9; VSLRC:%vreg5,%vreg9 | |||
1660 | ||||
1661 | SlotIndex FMAIdx = LIS->getInstructionIndex(MI); | |||
1662 | ||||
1663 | VNInfo *AddendValNo = | |||
1664 | LIS->getInterval(MI->getOperand(1).getReg()).Query(FMAIdx).valueIn(); | |||
1665 | MachineInstr *AddendMI = LIS->getInstructionFromIndex(AddendValNo->def); | |||
1666 | ||||
1667 | // The addend and this instruction must be in the same block. | |||
1668 | ||||
1669 | if (!AddendMI || AddendMI->getParent() != MI->getParent()) | |||
1670 | continue; | |||
1671 | ||||
1672 | // The addend must be a full copy within the same register class. | |||
1673 | ||||
1674 | if (!AddendMI->isFullCopy()) | |||
1675 | continue; | |||
1676 | ||||
1677 | unsigned AddendSrcReg = AddendMI->getOperand(1).getReg(); | |||
1678 | if (TargetRegisterInfo::isVirtualRegister(AddendSrcReg)) { | |||
1679 | if (MRI.getRegClass(AddendMI->getOperand(0).getReg()) != | |||
1680 | MRI.getRegClass(AddendSrcReg)) | |||
1681 | continue; | |||
1682 | } else { | |||
1683 | // If AddendSrcReg is a physical register, make sure the destination | |||
1684 | // register class contains it. | |||
1685 | if (!MRI.getRegClass(AddendMI->getOperand(0).getReg()) | |||
1686 | ->contains(AddendSrcReg)) | |||
1687 | continue; | |||
1688 | } | |||
1689 | ||||
1690 | // In theory, there could be other uses of the addend copy before this | |||
1691 | // fma. We could deal with this, but that would require additional | |||
1692 | // logic below and I suspect it will not occur in any relevant | |||
1693 | // situations. Additionally, check whether the copy source is killed | |||
1694 | // prior to the fma. In order to replace the addend here with the | |||
1695 | // source of the copy, it must still be live here. We can't use | |||
1696 | // interval testing for a physical register, so as long as we're | |||
1697 | // walking the MIs we may as well test liveness here. | |||
1698 | bool OtherUsers = false, KillsAddendSrc = false; | |||
1699 | for (auto J = std::prev(I), JE = MachineBasicBlock::iterator(AddendMI); | |||
1700 | J != JE; --J) { | |||
1701 | if (J->readsVirtualRegister(AddendMI->getOperand(0).getReg())) { | |||
1702 | OtherUsers = true; | |||
1703 | break; | |||
1704 | } | |||
1705 | if (J->modifiesRegister(AddendSrcReg, TRI) || | |||
1706 | J->killsRegister(AddendSrcReg, TRI)) { | |||
1707 | KillsAddendSrc = true; | |||
1708 | break; | |||
1709 | } | |||
1710 | } | |||
1711 | ||||
1712 | if (OtherUsers || KillsAddendSrc) | |||
1713 | continue; | |||
1714 | ||||
1715 | // Find one of the product operands that is killed by this instruction. | |||
1716 | ||||
1717 | unsigned KilledProdOp = 0, OtherProdOp = 0; | |||
1718 | if (LIS->getInterval(MI->getOperand(2).getReg()) | |||
1719 | .Query(FMAIdx).isKill()) { | |||
1720 | KilledProdOp = 2; | |||
1721 | OtherProdOp = 3; | |||
1722 | } else if (LIS->getInterval(MI->getOperand(3).getReg()) | |||
1723 | .Query(FMAIdx).isKill()) { | |||
1724 | KilledProdOp = 3; | |||
1725 | OtherProdOp = 2; | |||
1726 | } | |||
1727 | ||||
1728 | // If there are no killed product operands, then this transformation is | |||
1729 | // likely not profitable. | |||
1730 | if (!KilledProdOp) | |||
1731 | continue; | |||
1732 | ||||
1733 | // For virtual registers, verify that the addend source register | |||
1734 | // is live here (as should have been assured above). | |||
1735 | assert((!TargetRegisterInfo::isVirtualRegister(AddendSrcReg) ||(((!TargetRegisterInfo::isVirtualRegister(AddendSrcReg) || LIS ->getInterval(AddendSrcReg).liveAt(FMAIdx)) && "Addend source register is not live!" ) ? static_cast<void> (0) : __assert_fail ("(!TargetRegisterInfo::isVirtualRegister(AddendSrcReg) || LIS->getInterval(AddendSrcReg).liveAt(FMAIdx)) && \"Addend source register is not live!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 1737, __PRETTY_FUNCTION__)) | |||
1736 | LIS->getInterval(AddendSrcReg).liveAt(FMAIdx)) &&(((!TargetRegisterInfo::isVirtualRegister(AddendSrcReg) || LIS ->getInterval(AddendSrcReg).liveAt(FMAIdx)) && "Addend source register is not live!" ) ? static_cast<void> (0) : __assert_fail ("(!TargetRegisterInfo::isVirtualRegister(AddendSrcReg) || LIS->getInterval(AddendSrcReg).liveAt(FMAIdx)) && \"Addend source register is not live!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 1737, __PRETTY_FUNCTION__)) | |||
1737 | "Addend source register is not live!")(((!TargetRegisterInfo::isVirtualRegister(AddendSrcReg) || LIS ->getInterval(AddendSrcReg).liveAt(FMAIdx)) && "Addend source register is not live!" ) ? static_cast<void> (0) : __assert_fail ("(!TargetRegisterInfo::isVirtualRegister(AddendSrcReg) || LIS->getInterval(AddendSrcReg).liveAt(FMAIdx)) && \"Addend source register is not live!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 1737, __PRETTY_FUNCTION__)); | |||
1738 | ||||
1739 | // Transform: (O2 * O3) + O1 -> (O2 * O1) + O3. | |||
1740 | ||||
1741 | unsigned AddReg = AddendMI->getOperand(1).getReg(); | |||
1742 | unsigned KilledProdReg = MI->getOperand(KilledProdOp).getReg(); | |||
1743 | unsigned OtherProdReg = MI->getOperand(OtherProdOp).getReg(); | |||
1744 | ||||
1745 | unsigned AddSubReg = AddendMI->getOperand(1).getSubReg(); | |||
1746 | unsigned KilledProdSubReg = MI->getOperand(KilledProdOp).getSubReg(); | |||
1747 | unsigned OtherProdSubReg = MI->getOperand(OtherProdOp).getSubReg(); | |||
1748 | ||||
1749 | bool AddRegKill = AddendMI->getOperand(1).isKill(); | |||
1750 | bool KilledProdRegKill = MI->getOperand(KilledProdOp).isKill(); | |||
1751 | bool OtherProdRegKill = MI->getOperand(OtherProdOp).isKill(); | |||
1752 | ||||
1753 | bool AddRegUndef = AddendMI->getOperand(1).isUndef(); | |||
1754 | bool KilledProdRegUndef = MI->getOperand(KilledProdOp).isUndef(); | |||
1755 | bool OtherProdRegUndef = MI->getOperand(OtherProdOp).isUndef(); | |||
1756 | ||||
1757 | unsigned OldFMAReg = MI->getOperand(0).getReg(); | |||
1758 | ||||
1759 | // The transformation doesn't work well with things like: | |||
1760 | // %vreg5 = A-form-op %vreg5, %vreg11, %vreg5; | |||
1761 | // so leave such things alone. | |||
1762 | if (OldFMAReg == KilledProdReg) | |||
1763 | continue; | |||
1764 | ||||
1765 | assert(OldFMAReg == AddendMI->getOperand(0).getReg() &&((OldFMAReg == AddendMI->getOperand(0).getReg() && "Addend copy not tied to old FMA output!") ? static_cast< void> (0) : __assert_fail ("OldFMAReg == AddendMI->getOperand(0).getReg() && \"Addend copy not tied to old FMA output!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 1766, __PRETTY_FUNCTION__)) | |||
1766 | "Addend copy not tied to old FMA output!")((OldFMAReg == AddendMI->getOperand(0).getReg() && "Addend copy not tied to old FMA output!") ? static_cast< void> (0) : __assert_fail ("OldFMAReg == AddendMI->getOperand(0).getReg() && \"Addend copy not tied to old FMA output!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 1766, __PRETTY_FUNCTION__)); | |||
1767 | ||||
1768 | DEBUG(dbgs() << "VSX FMA Mutation:\n " << *MI;)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-early-ret")) { dbgs() << "VSX FMA Mutation:\n " << *MI;; } } while (0); | |||
1769 | ||||
1770 | MI->getOperand(0).setReg(KilledProdReg); | |||
1771 | MI->getOperand(1).setReg(KilledProdReg); | |||
1772 | MI->getOperand(3).setReg(AddReg); | |||
1773 | MI->getOperand(2).setReg(OtherProdReg); | |||
1774 | ||||
1775 | MI->getOperand(0).setSubReg(KilledProdSubReg); | |||
1776 | MI->getOperand(1).setSubReg(KilledProdSubReg); | |||
1777 | MI->getOperand(3).setSubReg(AddSubReg); | |||
1778 | MI->getOperand(2).setSubReg(OtherProdSubReg); | |||
1779 | ||||
1780 | MI->getOperand(1).setIsKill(KilledProdRegKill); | |||
1781 | MI->getOperand(3).setIsKill(AddRegKill); | |||
1782 | MI->getOperand(2).setIsKill(OtherProdRegKill); | |||
1783 | ||||
1784 | MI->getOperand(1).setIsUndef(KilledProdRegUndef); | |||
1785 | MI->getOperand(3).setIsUndef(AddRegUndef); | |||
1786 | MI->getOperand(2).setIsUndef(OtherProdRegUndef); | |||
1787 | ||||
1788 | MI->setDesc(TII->get(AltOpc)); | |||
1789 | ||||
1790 | DEBUG(dbgs() << " -> " << *MI)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-early-ret")) { dbgs() << " -> " << *MI; } } while (0); | |||
1791 | ||||
1792 | // The killed product operand was killed here, so we can reuse it now | |||
1793 | // for the result of the fma. | |||
1794 | ||||
1795 | LiveInterval &FMAInt = LIS->getInterval(OldFMAReg); | |||
1796 | VNInfo *FMAValNo = FMAInt.getVNInfoAt(FMAIdx.getRegSlot()); | |||
1797 | for (auto UI = MRI.reg_nodbg_begin(OldFMAReg), UE = MRI.reg_nodbg_end(); | |||
1798 | UI != UE;) { | |||
1799 | MachineOperand &UseMO = *UI; | |||
1800 | MachineInstr *UseMI = UseMO.getParent(); | |||
1801 | ++UI; | |||
1802 | ||||
1803 | // Don't replace the result register of the copy we're about to erase. | |||
1804 | if (UseMI == AddendMI) | |||
1805 | continue; | |||
1806 | ||||
1807 | UseMO.setReg(KilledProdReg); | |||
1808 | UseMO.setSubReg(KilledProdSubReg); | |||
1809 | } | |||
1810 | ||||
1811 | // Extend the live intervals of the killed product operand to hold the | |||
1812 | // fma result. | |||
1813 | ||||
1814 | LiveInterval &NewFMAInt = LIS->getInterval(KilledProdReg); | |||
1815 | for (LiveInterval::iterator AI = FMAInt.begin(), AE = FMAInt.end(); | |||
1816 | AI != AE; ++AI) { | |||
1817 | // Don't add the segment that corresponds to the original copy. | |||
1818 | if (AI->valno == AddendValNo) | |||
1819 | continue; | |||
1820 | ||||
1821 | VNInfo *NewFMAValNo = | |||
1822 | NewFMAInt.getNextValue(AI->start, | |||
1823 | LIS->getVNInfoAllocator()); | |||
1824 | ||||
1825 | NewFMAInt.addSegment(LiveInterval::Segment(AI->start, AI->end, | |||
1826 | NewFMAValNo)); | |||
1827 | } | |||
1828 | DEBUG(dbgs() << " extended: " << NewFMAInt << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-early-ret")) { dbgs() << " extended: " << NewFMAInt << '\n'; } } while (0); | |||
1829 | ||||
1830 | FMAInt.removeValNo(FMAValNo); | |||
1831 | DEBUG(dbgs() << " trimmed: " << FMAInt << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-early-ret")) { dbgs() << " trimmed: " << FMAInt << '\n'; } } while (0); | |||
1832 | ||||
1833 | // Remove the (now unused) copy. | |||
1834 | ||||
1835 | DEBUG(dbgs() << " removing: " << *AddendMI << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-early-ret")) { dbgs() << " removing: " << * AddendMI << '\n'; } } while (0); | |||
1836 | LIS->RemoveMachineInstrFromMaps(AddendMI); | |||
1837 | AddendMI->eraseFromParent(); | |||
1838 | ||||
1839 | Changed = true; | |||
1840 | } | |||
1841 | ||||
1842 | return Changed; | |||
1843 | } | |||
1844 | ||||
1845 | public: | |||
1846 | bool runOnMachineFunction(MachineFunction &MF) override { | |||
1847 | TM = static_cast<const PPCTargetMachine *>(&MF.getTarget()); | |||
1848 | // If we don't have VSX then go ahead and return without doing | |||
1849 | // anything. | |||
1850 | if (!TM->getSubtargetImpl()->hasVSX()) | |||
1851 | return false; | |||
1852 | ||||
1853 | LIS = &getAnalysis<LiveIntervals>(); | |||
1854 | ||||
1855 | TII = TM->getSubtargetImpl()->getInstrInfo(); | |||
1856 | ||||
1857 | bool Changed = false; | |||
1858 | ||||
1859 | if (DisableVSXFMAMutate) | |||
1860 | return Changed; | |||
1861 | ||||
1862 | for (MachineFunction::iterator I = MF.begin(); I != MF.end();) { | |||
1863 | MachineBasicBlock &B = *I++; | |||
1864 | if (processBlock(B)) | |||
1865 | Changed = true; | |||
1866 | } | |||
1867 | ||||
1868 | return Changed; | |||
1869 | } | |||
1870 | ||||
1871 | void getAnalysisUsage(AnalysisUsage &AU) const override { | |||
1872 | AU.addRequired<LiveIntervals>(); | |||
1873 | AU.addPreserved<LiveIntervals>(); | |||
1874 | AU.addRequired<SlotIndexes>(); | |||
1875 | AU.addPreserved<SlotIndexes>(); | |||
1876 | MachineFunctionPass::getAnalysisUsage(AU); | |||
1877 | } | |||
1878 | }; | |||
1879 | } | |||
1880 | ||||
1881 | INITIALIZE_PASS_BEGIN(PPCVSXFMAMutate, DEBUG_TYPE,static void* initializePPCVSXFMAMutatePassOnce(PassRegistry & Registry) { | |||
1882 | "PowerPC VSX FMA Mutation", false, false)static void* initializePPCVSXFMAMutatePassOnce(PassRegistry & Registry) { | |||
1883 | INITIALIZE_PASS_DEPENDENCY(LiveIntervals)initializeLiveIntervalsPass(Registry); | |||
1884 | INITIALIZE_PASS_DEPENDENCY(SlotIndexes)initializeSlotIndexesPass(Registry); | |||
1885 | INITIALIZE_PASS_END(PPCVSXFMAMutate, DEBUG_TYPE,PassInfo *PI = new PassInfo("PowerPC VSX FMA Mutation", "ppc-early-ret" , & PPCVSXFMAMutate ::ID, PassInfo::NormalCtor_t(callDefaultCtor < PPCVSXFMAMutate >), false, false); Registry.registerPass (*PI, true); return PI; } void llvm::initializePPCVSXFMAMutatePass (PassRegistry &Registry) { static volatile sys::cas_flag initialized = 0; sys::cas_flag old_val = sys::CompareAndSwap(&initialized , 1, 0); if (old_val == 0) { initializePPCVSXFMAMutatePassOnce (Registry); sys::MemoryFence(); AnnotateIgnoreWritesBegin("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 1886); AnnotateHappensBefore("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 1886, &initialized); initialized = 2; AnnotateIgnoreWritesEnd ("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 1886); } else { sys::cas_flag tmp = initialized; sys::MemoryFence (); while (tmp != 2) { tmp = initialized; sys::MemoryFence(); } } AnnotateHappensAfter("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 1886, &initialized); } | |||
1886 | "PowerPC VSX FMA Mutation", false, false)PassInfo *PI = new PassInfo("PowerPC VSX FMA Mutation", "ppc-early-ret" , & PPCVSXFMAMutate ::ID, PassInfo::NormalCtor_t(callDefaultCtor < PPCVSXFMAMutate >), false, false); Registry.registerPass (*PI, true); return PI; } void llvm::initializePPCVSXFMAMutatePass (PassRegistry &Registry) { static volatile sys::cas_flag initialized = 0; sys::cas_flag old_val = sys::CompareAndSwap(&initialized , 1, 0); if (old_val == 0) { initializePPCVSXFMAMutatePassOnce (Registry); sys::MemoryFence(); AnnotateIgnoreWritesBegin("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 1886); AnnotateHappensBefore("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 1886, &initialized); initialized = 2; AnnotateIgnoreWritesEnd ("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 1886); } else { sys::cas_flag tmp = initialized; sys::MemoryFence (); while (tmp != 2) { tmp = initialized; sys::MemoryFence(); } } AnnotateHappensAfter("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 1886, &initialized); } | |||
1887 | ||||
1888 | char &llvm::PPCVSXFMAMutateID = PPCVSXFMAMutate::ID; | |||
1889 | ||||
1890 | char PPCVSXFMAMutate::ID = 0; | |||
1891 | FunctionPass* | |||
1892 | llvm::createPPCVSXFMAMutatePass() { return new PPCVSXFMAMutate(); } | |||
1893 | ||||
1894 | #undef DEBUG_TYPE"ppc-early-ret" | |||
1895 | #define DEBUG_TYPE"ppc-early-ret" "ppc-vsx-copy" | |||
1896 | ||||
1897 | namespace llvm { | |||
1898 | void initializePPCVSXCopyPass(PassRegistry&); | |||
1899 | } | |||
1900 | ||||
1901 | namespace { | |||
1902 | // PPCVSXCopy pass - For copies between VSX registers and non-VSX registers | |||
1903 | // (Altivec and scalar floating-point registers), we need to transform the | |||
1904 | // copies into subregister copies with other restrictions. | |||
1905 | struct PPCVSXCopy : public MachineFunctionPass { | |||
1906 | static char ID; | |||
1907 | PPCVSXCopy() : MachineFunctionPass(ID) { | |||
1908 | initializePPCVSXCopyPass(*PassRegistry::getPassRegistry()); | |||
1909 | } | |||
1910 | ||||
1911 | const PPCTargetMachine *TM; | |||
1912 | const PPCInstrInfo *TII; | |||
1913 | ||||
1914 | bool IsRegInClass(unsigned Reg, const TargetRegisterClass *RC, | |||
1915 | MachineRegisterInfo &MRI) { | |||
1916 | if (TargetRegisterInfo::isVirtualRegister(Reg)) { | |||
1917 | return RC->hasSubClassEq(MRI.getRegClass(Reg)); | |||
1918 | } else if (RC->contains(Reg)) { | |||
1919 | return true; | |||
1920 | } | |||
1921 | ||||
1922 | return false; | |||
1923 | } | |||
1924 | ||||
1925 | bool IsVSReg(unsigned Reg, MachineRegisterInfo &MRI) { | |||
1926 | return IsRegInClass(Reg, &PPC::VSRCRegClass, MRI); | |||
1927 | } | |||
1928 | ||||
1929 | bool IsVRReg(unsigned Reg, MachineRegisterInfo &MRI) { | |||
1930 | return IsRegInClass(Reg, &PPC::VRRCRegClass, MRI); | |||
1931 | } | |||
1932 | ||||
1933 | bool IsF8Reg(unsigned Reg, MachineRegisterInfo &MRI) { | |||
1934 | return IsRegInClass(Reg, &PPC::F8RCRegClass, MRI); | |||
1935 | } | |||
1936 | ||||
1937 | protected: | |||
1938 | bool processBlock(MachineBasicBlock &MBB) { | |||
1939 | bool Changed = false; | |||
1940 | ||||
1941 | MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); | |||
1942 | for (MachineBasicBlock::iterator I = MBB.begin(), IE = MBB.end(); | |||
1943 | I != IE; ++I) { | |||
1944 | MachineInstr *MI = I; | |||
1945 | if (!MI->isFullCopy()) | |||
1946 | continue; | |||
1947 | ||||
1948 | MachineOperand &DstMO = MI->getOperand(0); | |||
1949 | MachineOperand &SrcMO = MI->getOperand(1); | |||
1950 | ||||
1951 | if ( IsVSReg(DstMO.getReg(), MRI) && | |||
1952 | !IsVSReg(SrcMO.getReg(), MRI)) { | |||
1953 | // This is a copy *to* a VSX register from a non-VSX register. | |||
1954 | Changed = true; | |||
1955 | ||||
1956 | const TargetRegisterClass *SrcRC = | |||
1957 | IsVRReg(SrcMO.getReg(), MRI) ? &PPC::VSHRCRegClass : | |||
1958 | &PPC::VSLRCRegClass; | |||
1959 | assert((IsF8Reg(SrcMO.getReg(), MRI) ||(((IsF8Reg(SrcMO.getReg(), MRI) || IsVRReg(SrcMO.getReg(), MRI )) && "Unknown source for a VSX copy") ? static_cast< void> (0) : __assert_fail ("(IsF8Reg(SrcMO.getReg(), MRI) || IsVRReg(SrcMO.getReg(), MRI)) && \"Unknown source for a VSX copy\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 1961, __PRETTY_FUNCTION__)) | |||
1960 | IsVRReg(SrcMO.getReg(), MRI)) &&(((IsF8Reg(SrcMO.getReg(), MRI) || IsVRReg(SrcMO.getReg(), MRI )) && "Unknown source for a VSX copy") ? static_cast< void> (0) : __assert_fail ("(IsF8Reg(SrcMO.getReg(), MRI) || IsVRReg(SrcMO.getReg(), MRI)) && \"Unknown source for a VSX copy\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 1961, __PRETTY_FUNCTION__)) | |||
1961 | "Unknown source for a VSX copy")(((IsF8Reg(SrcMO.getReg(), MRI) || IsVRReg(SrcMO.getReg(), MRI )) && "Unknown source for a VSX copy") ? static_cast< void> (0) : __assert_fail ("(IsF8Reg(SrcMO.getReg(), MRI) || IsVRReg(SrcMO.getReg(), MRI)) && \"Unknown source for a VSX copy\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 1961, __PRETTY_FUNCTION__)); | |||
1962 | ||||
1963 | unsigned NewVReg = MRI.createVirtualRegister(SrcRC); | |||
1964 | BuildMI(MBB, MI, MI->getDebugLoc(), | |||
1965 | TII->get(TargetOpcode::SUBREG_TO_REG), NewVReg) | |||
1966 | .addImm(1) // add 1, not 0, because there is no implicit clearing | |||
1967 | // of the high bits. | |||
1968 | .addOperand(SrcMO) | |||
1969 | .addImm(IsVRReg(SrcMO.getReg(), MRI) ? PPC::sub_128 : | |||
1970 | PPC::sub_64); | |||
1971 | ||||
1972 | // The source of the original copy is now the new virtual register. | |||
1973 | SrcMO.setReg(NewVReg); | |||
1974 | } else if (!IsVSReg(DstMO.getReg(), MRI) && | |||
1975 | IsVSReg(SrcMO.getReg(), MRI)) { | |||
1976 | // This is a copy *from* a VSX register to a non-VSX register. | |||
1977 | Changed = true; | |||
1978 | ||||
1979 | const TargetRegisterClass *DstRC = | |||
1980 | IsVRReg(DstMO.getReg(), MRI) ? &PPC::VSHRCRegClass : | |||
1981 | &PPC::VSLRCRegClass; | |||
1982 | assert((IsF8Reg(DstMO.getReg(), MRI) ||(((IsF8Reg(DstMO.getReg(), MRI) || IsVRReg(DstMO.getReg(), MRI )) && "Unknown destination for a VSX copy") ? static_cast <void> (0) : __assert_fail ("(IsF8Reg(DstMO.getReg(), MRI) || IsVRReg(DstMO.getReg(), MRI)) && \"Unknown destination for a VSX copy\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 1984, __PRETTY_FUNCTION__)) | |||
1983 | IsVRReg(DstMO.getReg(), MRI)) &&(((IsF8Reg(DstMO.getReg(), MRI) || IsVRReg(DstMO.getReg(), MRI )) && "Unknown destination for a VSX copy") ? static_cast <void> (0) : __assert_fail ("(IsF8Reg(DstMO.getReg(), MRI) || IsVRReg(DstMO.getReg(), MRI)) && \"Unknown destination for a VSX copy\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 1984, __PRETTY_FUNCTION__)) | |||
1984 | "Unknown destination for a VSX copy")(((IsF8Reg(DstMO.getReg(), MRI) || IsVRReg(DstMO.getReg(), MRI )) && "Unknown destination for a VSX copy") ? static_cast <void> (0) : __assert_fail ("(IsF8Reg(DstMO.getReg(), MRI) || IsVRReg(DstMO.getReg(), MRI)) && \"Unknown destination for a VSX copy\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 1984, __PRETTY_FUNCTION__)); | |||
1985 | ||||
1986 | // Copy the VSX value into a new VSX register of the correct subclass. | |||
1987 | unsigned NewVReg = MRI.createVirtualRegister(DstRC); | |||
1988 | BuildMI(MBB, MI, MI->getDebugLoc(), | |||
1989 | TII->get(TargetOpcode::COPY), NewVReg) | |||
1990 | .addOperand(SrcMO); | |||
1991 | ||||
1992 | // Transform the original copy into a subregister extraction copy. | |||
1993 | SrcMO.setReg(NewVReg); | |||
1994 | SrcMO.setSubReg(IsVRReg(DstMO.getReg(), MRI) ? PPC::sub_128 : | |||
1995 | PPC::sub_64); | |||
1996 | } | |||
1997 | } | |||
1998 | ||||
1999 | return Changed; | |||
2000 | } | |||
2001 | ||||
2002 | public: | |||
2003 | bool runOnMachineFunction(MachineFunction &MF) override { | |||
2004 | TM = static_cast<const PPCTargetMachine *>(&MF.getTarget()); | |||
2005 | // If we don't have VSX on the subtarget, don't do anything. | |||
2006 | if (!TM->getSubtargetImpl()->hasVSX()) | |||
2007 | return false; | |||
2008 | TII = TM->getSubtargetImpl()->getInstrInfo(); | |||
2009 | ||||
2010 | bool Changed = false; | |||
2011 | ||||
2012 | for (MachineFunction::iterator I = MF.begin(); I != MF.end();) { | |||
2013 | MachineBasicBlock &B = *I++; | |||
2014 | if (processBlock(B)) | |||
2015 | Changed = true; | |||
2016 | } | |||
2017 | ||||
2018 | return Changed; | |||
2019 | } | |||
2020 | ||||
2021 | void getAnalysisUsage(AnalysisUsage &AU) const override { | |||
2022 | MachineFunctionPass::getAnalysisUsage(AU); | |||
2023 | } | |||
2024 | }; | |||
2025 | } | |||
2026 | ||||
2027 | INITIALIZE_PASS(PPCVSXCopy, DEBUG_TYPE,static void* initializePPCVSXCopyPassOnce(PassRegistry &Registry ) { PassInfo *PI = new PassInfo("PowerPC VSX Copy Legalization" , "ppc-early-ret", & PPCVSXCopy ::ID, PassInfo::NormalCtor_t (callDefaultCtor< PPCVSXCopy >), false, false); Registry .registerPass(*PI, true); return PI; } void llvm::initializePPCVSXCopyPass (PassRegistry &Registry) { static volatile sys::cas_flag initialized = 0; sys::cas_flag old_val = sys::CompareAndSwap(&initialized , 1, 0); if (old_val == 0) { initializePPCVSXCopyPassOnce(Registry ); sys::MemoryFence(); AnnotateIgnoreWritesBegin("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 2028); AnnotateHappensBefore("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 2028, &initialized); initialized = 2; AnnotateIgnoreWritesEnd ("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 2028); } else { sys::cas_flag tmp = initialized; sys::MemoryFence (); while (tmp != 2) { tmp = initialized; sys::MemoryFence(); } } AnnotateHappensAfter("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 2028, &initialized); } | |||
2028 | "PowerPC VSX Copy Legalization", false, false)static void* initializePPCVSXCopyPassOnce(PassRegistry &Registry ) { PassInfo *PI = new PassInfo("PowerPC VSX Copy Legalization" , "ppc-early-ret", & PPCVSXCopy ::ID, PassInfo::NormalCtor_t (callDefaultCtor< PPCVSXCopy >), false, false); Registry .registerPass(*PI, true); return PI; } void llvm::initializePPCVSXCopyPass (PassRegistry &Registry) { static volatile sys::cas_flag initialized = 0; sys::cas_flag old_val = sys::CompareAndSwap(&initialized , 1, 0); if (old_val == 0) { initializePPCVSXCopyPassOnce(Registry ); sys::MemoryFence(); AnnotateIgnoreWritesBegin("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 2028); AnnotateHappensBefore("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 2028, &initialized); initialized = 2; AnnotateIgnoreWritesEnd ("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 2028); } else { sys::cas_flag tmp = initialized; sys::MemoryFence (); while (tmp != 2) { tmp = initialized; sys::MemoryFence(); } } AnnotateHappensAfter("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 2028, &initialized); } | |||
2029 | ||||
2030 | char PPCVSXCopy::ID = 0; | |||
2031 | FunctionPass* | |||
2032 | llvm::createPPCVSXCopyPass() { return new PPCVSXCopy(); } | |||
2033 | ||||
2034 | #undef DEBUG_TYPE"ppc-early-ret" | |||
2035 | #define DEBUG_TYPE"ppc-early-ret" "ppc-vsx-copy-cleanup" | |||
2036 | ||||
2037 | namespace llvm { | |||
2038 | void initializePPCVSXCopyCleanupPass(PassRegistry&); | |||
2039 | } | |||
2040 | ||||
2041 | namespace { | |||
2042 | // PPCVSXCopyCleanup pass - We sometimes end up generating self copies of VSX | |||
2043 | // registers (mostly because the ABI code still places all values into the | |||
2044 | // "traditional" floating-point and vector registers). Remove them here. | |||
2045 | struct PPCVSXCopyCleanup : public MachineFunctionPass { | |||
2046 | static char ID; | |||
2047 | PPCVSXCopyCleanup() : MachineFunctionPass(ID) { | |||
2048 | initializePPCVSXCopyCleanupPass(*PassRegistry::getPassRegistry()); | |||
2049 | } | |||
2050 | ||||
2051 | const PPCTargetMachine *TM; | |||
2052 | const PPCInstrInfo *TII; | |||
2053 | ||||
2054 | protected: | |||
2055 | bool processBlock(MachineBasicBlock &MBB) { | |||
2056 | bool Changed = false; | |||
2057 | ||||
2058 | SmallVector<MachineInstr *, 4> ToDelete; | |||
2059 | for (MachineBasicBlock::iterator I = MBB.begin(), IE = MBB.end(); | |||
2060 | I != IE; ++I) { | |||
2061 | MachineInstr *MI = I; | |||
2062 | if (MI->getOpcode() == PPC::XXLOR && | |||
2063 | MI->getOperand(0).getReg() == MI->getOperand(1).getReg() && | |||
2064 | MI->getOperand(0).getReg() == MI->getOperand(2).getReg()) | |||
2065 | ToDelete.push_back(MI); | |||
2066 | } | |||
2067 | ||||
2068 | if (!ToDelete.empty()) | |||
2069 | Changed = true; | |||
2070 | ||||
2071 | for (unsigned i = 0, ie = ToDelete.size(); i != ie; ++i) { | |||
2072 | DEBUG(dbgs() << "Removing VSX self-copy: " << *ToDelete[i])do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-early-ret")) { dbgs() << "Removing VSX self-copy: " << *ToDelete[i]; } } while (0); | |||
2073 | ToDelete[i]->eraseFromParent(); | |||
2074 | } | |||
2075 | ||||
2076 | return Changed; | |||
2077 | } | |||
2078 | ||||
2079 | public: | |||
2080 | bool runOnMachineFunction(MachineFunction &MF) override { | |||
2081 | TM = static_cast<const PPCTargetMachine *>(&MF.getTarget()); | |||
2082 | // If we don't have VSX don't bother doing anything here. | |||
2083 | if (!TM->getSubtargetImpl()->hasVSX()) | |||
2084 | return false; | |||
2085 | TII = TM->getSubtargetImpl()->getInstrInfo(); | |||
2086 | ||||
2087 | bool Changed = false; | |||
2088 | ||||
2089 | for (MachineFunction::iterator I = MF.begin(); I != MF.end();) { | |||
2090 | MachineBasicBlock &B = *I++; | |||
2091 | if (processBlock(B)) | |||
2092 | Changed = true; | |||
2093 | } | |||
2094 | ||||
2095 | return Changed; | |||
2096 | } | |||
2097 | ||||
2098 | void getAnalysisUsage(AnalysisUsage &AU) const override { | |||
2099 | MachineFunctionPass::getAnalysisUsage(AU); | |||
2100 | } | |||
2101 | }; | |||
2102 | } | |||
2103 | ||||
2104 | INITIALIZE_PASS(PPCVSXCopyCleanup, DEBUG_TYPE,static void* initializePPCVSXCopyCleanupPassOnce(PassRegistry &Registry) { PassInfo *PI = new PassInfo("PowerPC VSX Copy Cleanup" , "ppc-early-ret", & PPCVSXCopyCleanup ::ID, PassInfo::NormalCtor_t (callDefaultCtor< PPCVSXCopyCleanup >), false, false); Registry .registerPass(*PI, true); return PI; } void llvm::initializePPCVSXCopyCleanupPass (PassRegistry &Registry) { static volatile sys::cas_flag initialized = 0; sys::cas_flag old_val = sys::CompareAndSwap(&initialized , 1, 0); if (old_val == 0) { initializePPCVSXCopyCleanupPassOnce (Registry); sys::MemoryFence(); AnnotateIgnoreWritesBegin("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 2105); AnnotateHappensBefore("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 2105, &initialized); initialized = 2; AnnotateIgnoreWritesEnd ("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 2105); } else { sys::cas_flag tmp = initialized; sys::MemoryFence (); while (tmp != 2) { tmp = initialized; sys::MemoryFence(); } } AnnotateHappensAfter("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 2105, &initialized); } | |||
2105 | "PowerPC VSX Copy Cleanup", false, false)static void* initializePPCVSXCopyCleanupPassOnce(PassRegistry &Registry) { PassInfo *PI = new PassInfo("PowerPC VSX Copy Cleanup" , "ppc-early-ret", & PPCVSXCopyCleanup ::ID, PassInfo::NormalCtor_t (callDefaultCtor< PPCVSXCopyCleanup >), false, false); Registry .registerPass(*PI, true); return PI; } void llvm::initializePPCVSXCopyCleanupPass (PassRegistry &Registry) { static volatile sys::cas_flag initialized = 0; sys::cas_flag old_val = sys::CompareAndSwap(&initialized , 1, 0); if (old_val == 0) { initializePPCVSXCopyCleanupPassOnce (Registry); sys::MemoryFence(); AnnotateIgnoreWritesBegin("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 2105); AnnotateHappensBefore("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 2105, &initialized); initialized = 2; AnnotateIgnoreWritesEnd ("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 2105); } else { sys::cas_flag tmp = initialized; sys::MemoryFence (); while (tmp != 2) { tmp = initialized; sys::MemoryFence(); } } AnnotateHappensAfter("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 2105, &initialized); } | |||
2106 | ||||
2107 | char PPCVSXCopyCleanup::ID = 0; | |||
2108 | FunctionPass* | |||
2109 | llvm::createPPCVSXCopyCleanupPass() { return new PPCVSXCopyCleanup(); } | |||
2110 | ||||
2111 | #undef DEBUG_TYPE"ppc-early-ret" | |||
2112 | #define DEBUG_TYPE"ppc-early-ret" "ppc-early-ret" | |||
2113 | STATISTIC(NumBCLR, "Number of early conditional returns")static llvm::Statistic NumBCLR = { "ppc-early-ret", "Number of early conditional returns" , 0, 0 }; | |||
2114 | STATISTIC(NumBLR, "Number of early returns")static llvm::Statistic NumBLR = { "ppc-early-ret", "Number of early returns" , 0, 0 }; | |||
2115 | ||||
2116 | namespace llvm { | |||
2117 | void initializePPCEarlyReturnPass(PassRegistry&); | |||
2118 | } | |||
2119 | ||||
2120 | namespace { | |||
2121 | // PPCEarlyReturn pass - For simple functions without epilogue code, move | |||
2122 | // returns up, and create conditional returns, to avoid unnecessary | |||
2123 | // branch-to-blr sequences. | |||
2124 | struct PPCEarlyReturn : public MachineFunctionPass { | |||
2125 | static char ID; | |||
2126 | PPCEarlyReturn() : MachineFunctionPass(ID) { | |||
2127 | initializePPCEarlyReturnPass(*PassRegistry::getPassRegistry()); | |||
2128 | } | |||
2129 | ||||
2130 | const PPCTargetMachine *TM; | |||
2131 | const PPCInstrInfo *TII; | |||
2132 | ||||
2133 | protected: | |||
2134 | bool processBlock(MachineBasicBlock &ReturnMBB) { | |||
2135 | bool Changed = false; | |||
2136 | ||||
2137 | MachineBasicBlock::iterator I = ReturnMBB.begin(); | |||
2138 | I = ReturnMBB.SkipPHIsAndLabels(I); | |||
2139 | ||||
2140 | // The block must be essentially empty except for the blr. | |||
2141 | if (I == ReturnMBB.end() || I->getOpcode() != PPC::BLR || | |||
2142 | I != ReturnMBB.getLastNonDebugInstr()) | |||
2143 | return Changed; | |||
2144 | ||||
2145 | SmallVector<MachineBasicBlock*, 8> PredToRemove; | |||
2146 | for (MachineBasicBlock::pred_iterator PI = ReturnMBB.pred_begin(), | |||
2147 | PIE = ReturnMBB.pred_end(); PI != PIE; ++PI) { | |||
2148 | bool OtherReference = false, BlockChanged = false; | |||
2149 | for (MachineBasicBlock::iterator J = (*PI)->getLastNonDebugInstr();;) { | |||
2150 | if (J->getOpcode() == PPC::B) { | |||
2151 | if (J->getOperand(0).getMBB() == &ReturnMBB) { | |||
2152 | // This is an unconditional branch to the return. Replace the | |||
2153 | // branch with a blr. | |||
2154 | BuildMI(**PI, J, J->getDebugLoc(), TII->get(PPC::BLR)); | |||
2155 | MachineBasicBlock::iterator K = J--; | |||
2156 | K->eraseFromParent(); | |||
2157 | BlockChanged = true; | |||
2158 | ++NumBLR; | |||
2159 | continue; | |||
2160 | } | |||
2161 | } else if (J->getOpcode() == PPC::BCC) { | |||
2162 | if (J->getOperand(2).getMBB() == &ReturnMBB) { | |||
2163 | // This is a conditional branch to the return. Replace the branch | |||
2164 | // with a bclr. | |||
2165 | BuildMI(**PI, J, J->getDebugLoc(), TII->get(PPC::BCCLR)) | |||
2166 | .addImm(J->getOperand(0).getImm()) | |||
2167 | .addReg(J->getOperand(1).getReg()); | |||
2168 | MachineBasicBlock::iterator K = J--; | |||
2169 | K->eraseFromParent(); | |||
2170 | BlockChanged = true; | |||
2171 | ++NumBCLR; | |||
2172 | continue; | |||
2173 | } | |||
2174 | } else if (J->getOpcode() == PPC::BC || J->getOpcode() == PPC::BCn) { | |||
2175 | if (J->getOperand(1).getMBB() == &ReturnMBB) { | |||
2176 | // This is a conditional branch to the return. Replace the branch | |||
2177 | // with a bclr. | |||
2178 | BuildMI(**PI, J, J->getDebugLoc(), | |||
2179 | TII->get(J->getOpcode() == PPC::BC ? | |||
2180 | PPC::BCLR : PPC::BCLRn)) | |||
2181 | .addReg(J->getOperand(0).getReg()); | |||
2182 | MachineBasicBlock::iterator K = J--; | |||
2183 | K->eraseFromParent(); | |||
2184 | BlockChanged = true; | |||
2185 | ++NumBCLR; | |||
2186 | continue; | |||
2187 | } | |||
2188 | } else if (J->isBranch()) { | |||
2189 | if (J->isIndirectBranch()) { | |||
2190 | if (ReturnMBB.hasAddressTaken()) | |||
2191 | OtherReference = true; | |||
2192 | } else | |||
2193 | for (unsigned i = 0; i < J->getNumOperands(); ++i) | |||
2194 | if (J->getOperand(i).isMBB() && | |||
2195 | J->getOperand(i).getMBB() == &ReturnMBB) | |||
2196 | OtherReference = true; | |||
2197 | } else if (!J->isTerminator() && !J->isDebugValue()) | |||
2198 | break; | |||
2199 | ||||
2200 | if (J == (*PI)->begin()) | |||
2201 | break; | |||
2202 | ||||
2203 | --J; | |||
2204 | } | |||
2205 | ||||
2206 | if ((*PI)->canFallThrough() && (*PI)->isLayoutSuccessor(&ReturnMBB)) | |||
2207 | OtherReference = true; | |||
2208 | ||||
2209 | // Predecessors are stored in a vector and can't be removed here. | |||
2210 | if (!OtherReference && BlockChanged) { | |||
2211 | PredToRemove.push_back(*PI); | |||
2212 | } | |||
2213 | ||||
2214 | if (BlockChanged) | |||
2215 | Changed = true; | |||
2216 | } | |||
2217 | ||||
2218 | for (unsigned i = 0, ie = PredToRemove.size(); i != ie; ++i) | |||
2219 | PredToRemove[i]->removeSuccessor(&ReturnMBB); | |||
2220 | ||||
2221 | if (Changed && !ReturnMBB.hasAddressTaken()) { | |||
2222 | // We now might be able to merge this blr-only block into its | |||
2223 | // by-layout predecessor. | |||
2224 | if (ReturnMBB.pred_size() == 1 && | |||
2225 | (*ReturnMBB.pred_begin())->isLayoutSuccessor(&ReturnMBB)) { | |||
2226 | // Move the blr into the preceding block. | |||
2227 | MachineBasicBlock &PrevMBB = **ReturnMBB.pred_begin(); | |||
2228 | PrevMBB.splice(PrevMBB.end(), &ReturnMBB, I); | |||
2229 | PrevMBB.removeSuccessor(&ReturnMBB); | |||
2230 | } | |||
2231 | ||||
2232 | if (ReturnMBB.pred_empty()) | |||
2233 | ReturnMBB.eraseFromParent(); | |||
2234 | } | |||
2235 | ||||
2236 | return Changed; | |||
2237 | } | |||
2238 | ||||
2239 | public: | |||
2240 | bool runOnMachineFunction(MachineFunction &MF) override { | |||
2241 | TM = static_cast<const PPCTargetMachine *>(&MF.getTarget()); | |||
2242 | TII = TM->getSubtargetImpl()->getInstrInfo(); | |||
2243 | ||||
2244 | bool Changed = false; | |||
2245 | ||||
2246 | // If the function does not have at least two blocks, then there is | |||
2247 | // nothing to do. | |||
2248 | if (MF.size() < 2) | |||
2249 | return Changed; | |||
2250 | ||||
2251 | for (MachineFunction::iterator I = MF.begin(); I != MF.end();) { | |||
2252 | MachineBasicBlock &B = *I++; | |||
2253 | if (processBlock(B)) | |||
2254 | Changed = true; | |||
2255 | } | |||
2256 | ||||
2257 | return Changed; | |||
2258 | } | |||
2259 | ||||
2260 | void getAnalysisUsage(AnalysisUsage &AU) const override { | |||
2261 | MachineFunctionPass::getAnalysisUsage(AU); | |||
2262 | } | |||
2263 | }; | |||
2264 | } | |||
2265 | ||||
2266 | INITIALIZE_PASS(PPCEarlyReturn, DEBUG_TYPE,static void* initializePPCEarlyReturnPassOnce(PassRegistry & Registry) { PassInfo *PI = new PassInfo("PowerPC Early-Return Creation" , "ppc-early-ret", & PPCEarlyReturn ::ID, PassInfo::NormalCtor_t (callDefaultCtor< PPCEarlyReturn >), false, false); Registry .registerPass(*PI, true); return PI; } void llvm::initializePPCEarlyReturnPass (PassRegistry &Registry) { static volatile sys::cas_flag initialized = 0; sys::cas_flag old_val = sys::CompareAndSwap(&initialized , 1, 0); if (old_val == 0) { initializePPCEarlyReturnPassOnce (Registry); sys::MemoryFence(); AnnotateIgnoreWritesBegin("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 2267); AnnotateHappensBefore("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 2267, &initialized); initialized = 2; AnnotateIgnoreWritesEnd ("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 2267); } else { sys::cas_flag tmp = initialized; sys::MemoryFence (); while (tmp != 2) { tmp = initialized; sys::MemoryFence(); } } AnnotateHappensAfter("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 2267, &initialized); } | |||
2267 | "PowerPC Early-Return Creation", false, false)static void* initializePPCEarlyReturnPassOnce(PassRegistry & Registry) { PassInfo *PI = new PassInfo("PowerPC Early-Return Creation" , "ppc-early-ret", & PPCEarlyReturn ::ID, PassInfo::NormalCtor_t (callDefaultCtor< PPCEarlyReturn >), false, false); Registry .registerPass(*PI, true); return PI; } void llvm::initializePPCEarlyReturnPass (PassRegistry &Registry) { static volatile sys::cas_flag initialized = 0; sys::cas_flag old_val = sys::CompareAndSwap(&initialized , 1, 0); if (old_val == 0) { initializePPCEarlyReturnPassOnce (Registry); sys::MemoryFence(); AnnotateIgnoreWritesBegin("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 2267); AnnotateHappensBefore("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 2267, &initialized); initialized = 2; AnnotateIgnoreWritesEnd ("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 2267); } else { sys::cas_flag tmp = initialized; sys::MemoryFence (); while (tmp != 2) { tmp = initialized; sys::MemoryFence(); } } AnnotateHappensAfter("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn224369/lib/Target/PowerPC/PPCInstrInfo.cpp" , 2267, &initialized); } | |||
2268 | ||||
2269 | char PPCEarlyReturn::ID = 0; | |||
2270 | FunctionPass* | |||
2271 | llvm::createPPCEarlyReturnPass() { return new PPCEarlyReturn(); } |