File: | lib/Target/PowerPC/PPCISelDAGToDAG.cpp |
Location: | line 567, column 12 |
Description: | Value stored to 'Hi' during its initialization is never read |
1 | //===-- PPCISelDAGToDAG.cpp - PPC --pattern matching inst selector --------===// |
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 defines a pattern matching instruction selector for PowerPC, |
11 | // converting from a legalized dag to a PPC dag. |
12 | // |
13 | //===----------------------------------------------------------------------===// |
14 | |
15 | #include "PPC.h" |
16 | #include "MCTargetDesc/PPCPredicates.h" |
17 | #include "PPCMachineFunctionInfo.h" |
18 | #include "PPCTargetMachine.h" |
19 | #include "llvm/CodeGen/MachineFunction.h" |
20 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
21 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
22 | #include "llvm/CodeGen/SelectionDAG.h" |
23 | #include "llvm/CodeGen/SelectionDAGISel.h" |
24 | #include "llvm/IR/Constants.h" |
25 | #include "llvm/IR/Function.h" |
26 | #include "llvm/IR/GlobalAlias.h" |
27 | #include "llvm/IR/GlobalValue.h" |
28 | #include "llvm/IR/GlobalVariable.h" |
29 | #include "llvm/IR/Intrinsics.h" |
30 | #include "llvm/IR/Module.h" |
31 | #include "llvm/Support/CommandLine.h" |
32 | #include "llvm/Support/Debug.h" |
33 | #include "llvm/Support/ErrorHandling.h" |
34 | #include "llvm/Support/MathExtras.h" |
35 | #include "llvm/Support/raw_ostream.h" |
36 | #include "llvm/Target/TargetOptions.h" |
37 | using namespace llvm; |
38 | |
39 | #define DEBUG_TYPE"ppc-codegen" "ppc-codegen" |
40 | |
41 | // FIXME: Remove this once the bug has been fixed! |
42 | cl::opt<bool> ANDIGlueBug("expose-ppc-andi-glue-bug", |
43 | cl::desc("expose the ANDI glue bug on PPC"), cl::Hidden); |
44 | |
45 | static cl::opt<bool> |
46 | UseBitPermRewriter("ppc-use-bit-perm-rewriter", cl::init(true), |
47 | cl::desc("use aggressive ppc isel for bit permutations"), |
48 | cl::Hidden); |
49 | static cl::opt<bool> BPermRewriterNoMasking( |
50 | "ppc-bit-perm-rewriter-stress-rotates", |
51 | cl::desc("stress rotate selection in aggressive ppc isel for " |
52 | "bit permutations"), |
53 | cl::Hidden); |
54 | |
55 | namespace llvm { |
56 | void initializePPCDAGToDAGISelPass(PassRegistry&); |
57 | } |
58 | |
59 | namespace { |
60 | //===--------------------------------------------------------------------===// |
61 | /// PPCDAGToDAGISel - PPC specific code to select PPC machine |
62 | /// instructions for SelectionDAG operations. |
63 | /// |
64 | class PPCDAGToDAGISel : public SelectionDAGISel { |
65 | const PPCTargetMachine &TM; |
66 | const PPCSubtarget *PPCSubTarget; |
67 | const PPCTargetLowering *PPCLowering; |
68 | unsigned GlobalBaseReg; |
69 | public: |
70 | explicit PPCDAGToDAGISel(PPCTargetMachine &tm) |
71 | : SelectionDAGISel(tm), TM(tm) { |
72 | initializePPCDAGToDAGISelPass(*PassRegistry::getPassRegistry()); |
73 | } |
74 | |
75 | bool runOnMachineFunction(MachineFunction &MF) override { |
76 | // Make sure we re-emit a set of the global base reg if necessary |
77 | GlobalBaseReg = 0; |
78 | PPCSubTarget = &MF.getSubtarget<PPCSubtarget>(); |
79 | PPCLowering = PPCSubTarget->getTargetLowering(); |
80 | SelectionDAGISel::runOnMachineFunction(MF); |
81 | |
82 | if (!PPCSubTarget->isSVR4ABI()) |
83 | InsertVRSaveCode(MF); |
84 | |
85 | return true; |
86 | } |
87 | |
88 | void PreprocessISelDAG() override; |
89 | void PostprocessISelDAG() override; |
90 | |
91 | /// getI32Imm - Return a target constant with the specified value, of type |
92 | /// i32. |
93 | inline SDValue getI32Imm(unsigned Imm, SDLoc dl) { |
94 | return CurDAG->getTargetConstant(Imm, dl, MVT::i32); |
95 | } |
96 | |
97 | /// getI64Imm - Return a target constant with the specified value, of type |
98 | /// i64. |
99 | inline SDValue getI64Imm(uint64_t Imm, SDLoc dl) { |
100 | return CurDAG->getTargetConstant(Imm, dl, MVT::i64); |
101 | } |
102 | |
103 | /// getSmallIPtrImm - Return a target constant of pointer type. |
104 | inline SDValue getSmallIPtrImm(unsigned Imm, SDLoc dl) { |
105 | return CurDAG->getTargetConstant( |
106 | Imm, dl, PPCLowering->getPointerTy(CurDAG->getDataLayout())); |
107 | } |
108 | |
109 | /// isRotateAndMask - Returns true if Mask and Shift can be folded into a |
110 | /// rotate and mask opcode and mask operation. |
111 | static bool isRotateAndMask(SDNode *N, unsigned Mask, bool isShiftMask, |
112 | unsigned &SH, unsigned &MB, unsigned &ME); |
113 | |
114 | /// getGlobalBaseReg - insert code into the entry mbb to materialize the PIC |
115 | /// base register. Return the virtual register that holds this value. |
116 | SDNode *getGlobalBaseReg(); |
117 | |
118 | SDNode *getFrameIndex(SDNode *SN, SDNode *N, unsigned Offset = 0); |
119 | |
120 | // Select - Convert the specified operand from a target-independent to a |
121 | // target-specific node if it hasn't already been changed. |
122 | SDNode *Select(SDNode *N) override; |
123 | |
124 | SDNode *SelectBitfieldInsert(SDNode *N); |
125 | SDNode *SelectBitPermutation(SDNode *N); |
126 | |
127 | /// SelectCC - Select a comparison of the specified values with the |
128 | /// specified condition code, returning the CR# of the expression. |
129 | SDValue SelectCC(SDValue LHS, SDValue RHS, ISD::CondCode CC, SDLoc dl); |
130 | |
131 | /// SelectAddrImm - Returns true if the address N can be represented by |
132 | /// a base register plus a signed 16-bit displacement [r+imm]. |
133 | bool SelectAddrImm(SDValue N, SDValue &Disp, |
134 | SDValue &Base) { |
135 | return PPCLowering->SelectAddressRegImm(N, Disp, Base, *CurDAG, false); |
136 | } |
137 | |
138 | /// SelectAddrImmOffs - Return true if the operand is valid for a preinc |
139 | /// immediate field. Note that the operand at this point is already the |
140 | /// result of a prior SelectAddressRegImm call. |
141 | bool SelectAddrImmOffs(SDValue N, SDValue &Out) const { |
142 | if (N.getOpcode() == ISD::TargetConstant || |
143 | N.getOpcode() == ISD::TargetGlobalAddress) { |
144 | Out = N; |
145 | return true; |
146 | } |
147 | |
148 | return false; |
149 | } |
150 | |
151 | /// SelectAddrIdx - Given the specified addressed, check to see if it can be |
152 | /// represented as an indexed [r+r] operation. Returns false if it can |
153 | /// be represented by [r+imm], which are preferred. |
154 | bool SelectAddrIdx(SDValue N, SDValue &Base, SDValue &Index) { |
155 | return PPCLowering->SelectAddressRegReg(N, Base, Index, *CurDAG); |
156 | } |
157 | |
158 | /// SelectAddrIdxOnly - Given the specified addressed, force it to be |
159 | /// represented as an indexed [r+r] operation. |
160 | bool SelectAddrIdxOnly(SDValue N, SDValue &Base, SDValue &Index) { |
161 | return PPCLowering->SelectAddressRegRegOnly(N, Base, Index, *CurDAG); |
162 | } |
163 | |
164 | /// SelectAddrImmX4 - Returns true if the address N can be represented by |
165 | /// a base register plus a signed 16-bit displacement that is a multiple of 4. |
166 | /// Suitable for use by STD and friends. |
167 | bool SelectAddrImmX4(SDValue N, SDValue &Disp, SDValue &Base) { |
168 | return PPCLowering->SelectAddressRegImm(N, Disp, Base, *CurDAG, true); |
169 | } |
170 | |
171 | // Select an address into a single register. |
172 | bool SelectAddr(SDValue N, SDValue &Base) { |
173 | Base = N; |
174 | return true; |
175 | } |
176 | |
177 | /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for |
178 | /// inline asm expressions. It is always correct to compute the value into |
179 | /// a register. The case of adding a (possibly relocatable) constant to a |
180 | /// register can be improved, but it is wrong to substitute Reg+Reg for |
181 | /// Reg in an asm, because the load or store opcode would have to change. |
182 | bool SelectInlineAsmMemoryOperand(const SDValue &Op, |
183 | unsigned ConstraintID, |
184 | std::vector<SDValue> &OutOps) override { |
185 | |
186 | switch(ConstraintID) { |
187 | default: |
188 | errs() << "ConstraintID: " << ConstraintID << "\n"; |
189 | llvm_unreachable("Unexpected asm memory constraint")::llvm::llvm_unreachable_internal("Unexpected asm memory constraint" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 189); |
190 | case InlineAsm::Constraint_es: |
191 | case InlineAsm::Constraint_i: |
192 | case InlineAsm::Constraint_m: |
193 | case InlineAsm::Constraint_o: |
194 | case InlineAsm::Constraint_Q: |
195 | case InlineAsm::Constraint_Z: |
196 | case InlineAsm::Constraint_Zy: |
197 | // We need to make sure that this one operand does not end up in r0 |
198 | // (because we might end up lowering this as 0(%op)). |
199 | const TargetRegisterInfo *TRI = PPCSubTarget->getRegisterInfo(); |
200 | const TargetRegisterClass *TRC = TRI->getPointerRegClass(*MF, /*Kind=*/1); |
201 | SDLoc dl(Op); |
202 | SDValue RC = CurDAG->getTargetConstant(TRC->getID(), dl, MVT::i32); |
203 | SDValue NewOp = |
204 | SDValue(CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS, |
205 | dl, Op.getValueType(), |
206 | Op, RC), 0); |
207 | |
208 | OutOps.push_back(NewOp); |
209 | return false; |
210 | } |
211 | return true; |
212 | } |
213 | |
214 | void InsertVRSaveCode(MachineFunction &MF); |
215 | |
216 | const char *getPassName() const override { |
217 | return "PowerPC DAG->DAG Pattern Instruction Selection"; |
218 | } |
219 | |
220 | // Include the pieces autogenerated from the target description. |
221 | #include "PPCGenDAGISel.inc" |
222 | |
223 | private: |
224 | SDNode *SelectSETCC(SDNode *N); |
225 | |
226 | void PeepholePPC64(); |
227 | void PeepholePPC64ZExt(); |
228 | void PeepholeCROps(); |
229 | |
230 | SDValue combineToCMPB(SDNode *N); |
231 | void foldBoolExts(SDValue &Res, SDNode *&N); |
232 | |
233 | bool AllUsersSelectZero(SDNode *N); |
234 | void SwapAllSelectUsers(SDNode *N); |
235 | |
236 | SDNode *transferMemOperands(SDNode *N, SDNode *Result); |
237 | }; |
238 | } |
239 | |
240 | /// InsertVRSaveCode - Once the entire function has been instruction selected, |
241 | /// all virtual registers are created and all machine instructions are built, |
242 | /// check to see if we need to save/restore VRSAVE. If so, do it. |
243 | void PPCDAGToDAGISel::InsertVRSaveCode(MachineFunction &Fn) { |
244 | // Check to see if this function uses vector registers, which means we have to |
245 | // save and restore the VRSAVE register and update it with the regs we use. |
246 | // |
247 | // In this case, there will be virtual registers of vector type created |
248 | // by the scheduler. Detect them now. |
249 | bool HasVectorVReg = false; |
250 | for (unsigned i = 0, e = RegInfo->getNumVirtRegs(); i != e; ++i) { |
251 | unsigned Reg = TargetRegisterInfo::index2VirtReg(i); |
252 | if (RegInfo->getRegClass(Reg) == &PPC::VRRCRegClass) { |
253 | HasVectorVReg = true; |
254 | break; |
255 | } |
256 | } |
257 | if (!HasVectorVReg) return; // nothing to do. |
258 | |
259 | // If we have a vector register, we want to emit code into the entry and exit |
260 | // blocks to save and restore the VRSAVE register. We do this here (instead |
261 | // of marking all vector instructions as clobbering VRSAVE) for two reasons: |
262 | // |
263 | // 1. This (trivially) reduces the load on the register allocator, by not |
264 | // having to represent the live range of the VRSAVE register. |
265 | // 2. This (more significantly) allows us to create a temporary virtual |
266 | // register to hold the saved VRSAVE value, allowing this temporary to be |
267 | // register allocated, instead of forcing it to be spilled to the stack. |
268 | |
269 | // Create two vregs - one to hold the VRSAVE register that is live-in to the |
270 | // function and one for the value after having bits or'd into it. |
271 | unsigned InVRSAVE = RegInfo->createVirtualRegister(&PPC::GPRCRegClass); |
272 | unsigned UpdatedVRSAVE = RegInfo->createVirtualRegister(&PPC::GPRCRegClass); |
273 | |
274 | const TargetInstrInfo &TII = *PPCSubTarget->getInstrInfo(); |
275 | MachineBasicBlock &EntryBB = *Fn.begin(); |
276 | DebugLoc dl; |
277 | // Emit the following code into the entry block: |
278 | // InVRSAVE = MFVRSAVE |
279 | // UpdatedVRSAVE = UPDATE_VRSAVE InVRSAVE |
280 | // MTVRSAVE UpdatedVRSAVE |
281 | MachineBasicBlock::iterator IP = EntryBB.begin(); // Insert Point |
282 | BuildMI(EntryBB, IP, dl, TII.get(PPC::MFVRSAVE), InVRSAVE); |
283 | BuildMI(EntryBB, IP, dl, TII.get(PPC::UPDATE_VRSAVE), |
284 | UpdatedVRSAVE).addReg(InVRSAVE); |
285 | BuildMI(EntryBB, IP, dl, TII.get(PPC::MTVRSAVE)).addReg(UpdatedVRSAVE); |
286 | |
287 | // Find all return blocks, outputting a restore in each epilog. |
288 | for (MachineFunction::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) { |
289 | if (BB->isReturnBlock()) { |
290 | IP = BB->end(); --IP; |
291 | |
292 | // Skip over all terminator instructions, which are part of the return |
293 | // sequence. |
294 | MachineBasicBlock::iterator I2 = IP; |
295 | while (I2 != BB->begin() && (--I2)->isTerminator()) |
296 | IP = I2; |
297 | |
298 | // Emit: MTVRSAVE InVRSave |
299 | BuildMI(*BB, IP, dl, TII.get(PPC::MTVRSAVE)).addReg(InVRSAVE); |
300 | } |
301 | } |
302 | } |
303 | |
304 | |
305 | /// getGlobalBaseReg - Output the instructions required to put the |
306 | /// base address to use for accessing globals into a register. |
307 | /// |
308 | SDNode *PPCDAGToDAGISel::getGlobalBaseReg() { |
309 | if (!GlobalBaseReg) { |
310 | const TargetInstrInfo &TII = *PPCSubTarget->getInstrInfo(); |
311 | // Insert the set of GlobalBaseReg into the first MBB of the function |
312 | MachineBasicBlock &FirstMBB = MF->front(); |
313 | MachineBasicBlock::iterator MBBI = FirstMBB.begin(); |
314 | const Module *M = MF->getFunction()->getParent(); |
315 | DebugLoc dl; |
316 | |
317 | if (PPCLowering->getPointerTy(CurDAG->getDataLayout()) == MVT::i32) { |
318 | if (PPCSubTarget->isTargetELF()) { |
319 | GlobalBaseReg = PPC::R30; |
320 | if (M->getPICLevel() == PICLevel::Small) { |
321 | BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MoveGOTtoLR)); |
322 | BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR), GlobalBaseReg); |
323 | MF->getInfo<PPCFunctionInfo>()->setUsesPICBase(true); |
324 | } else { |
325 | BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR)); |
326 | BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR), GlobalBaseReg); |
327 | unsigned TempReg = RegInfo->createVirtualRegister(&PPC::GPRCRegClass); |
328 | BuildMI(FirstMBB, MBBI, dl, |
329 | TII.get(PPC::UpdateGBR), GlobalBaseReg) |
330 | .addReg(TempReg, RegState::Define).addReg(GlobalBaseReg); |
331 | MF->getInfo<PPCFunctionInfo>()->setUsesPICBase(true); |
332 | } |
333 | } else { |
334 | GlobalBaseReg = |
335 | RegInfo->createVirtualRegister(&PPC::GPRC_NOR0RegClass); |
336 | BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR)); |
337 | BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR), GlobalBaseReg); |
338 | } |
339 | } else { |
340 | GlobalBaseReg = RegInfo->createVirtualRegister(&PPC::G8RC_NOX0RegClass); |
341 | BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR8)); |
342 | BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR8), GlobalBaseReg); |
343 | } |
344 | } |
345 | return CurDAG->getRegister(GlobalBaseReg, |
346 | PPCLowering->getPointerTy(CurDAG->getDataLayout())) |
347 | .getNode(); |
348 | } |
349 | |
350 | /// isIntS16Immediate - This method tests to see if the node is either a 32-bit |
351 | /// or 64-bit immediate, and if the value can be accurately represented as a |
352 | /// sign extension from a 16-bit value. If so, this returns true and the |
353 | /// immediate. |
354 | static bool isIntS16Immediate(SDNode *N, short &Imm) { |
355 | if (N->getOpcode() != ISD::Constant) |
356 | return false; |
357 | |
358 | Imm = (short)cast<ConstantSDNode>(N)->getZExtValue(); |
359 | if (N->getValueType(0) == MVT::i32) |
360 | return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue(); |
361 | else |
362 | return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue(); |
363 | } |
364 | |
365 | static bool isIntS16Immediate(SDValue Op, short &Imm) { |
366 | return isIntS16Immediate(Op.getNode(), Imm); |
367 | } |
368 | |
369 | |
370 | /// isInt32Immediate - This method tests to see if the node is a 32-bit constant |
371 | /// operand. If so Imm will receive the 32-bit value. |
372 | static bool isInt32Immediate(SDNode *N, unsigned &Imm) { |
373 | if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i32) { |
374 | Imm = cast<ConstantSDNode>(N)->getZExtValue(); |
375 | return true; |
376 | } |
377 | return false; |
378 | } |
379 | |
380 | /// isInt64Immediate - This method tests to see if the node is a 64-bit constant |
381 | /// operand. If so Imm will receive the 64-bit value. |
382 | static bool isInt64Immediate(SDNode *N, uint64_t &Imm) { |
383 | if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i64) { |
384 | Imm = cast<ConstantSDNode>(N)->getZExtValue(); |
385 | return true; |
386 | } |
387 | return false; |
388 | } |
389 | |
390 | // isInt32Immediate - This method tests to see if a constant operand. |
391 | // If so Imm will receive the 32 bit value. |
392 | static bool isInt32Immediate(SDValue N, unsigned &Imm) { |
393 | return isInt32Immediate(N.getNode(), Imm); |
394 | } |
395 | |
396 | |
397 | // isOpcWithIntImmediate - This method tests to see if the node is a specific |
398 | // opcode and that it has a immediate integer right operand. |
399 | // If so Imm will receive the 32 bit value. |
400 | static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) { |
401 | return N->getOpcode() == Opc |
402 | && isInt32Immediate(N->getOperand(1).getNode(), Imm); |
403 | } |
404 | |
405 | SDNode *PPCDAGToDAGISel::getFrameIndex(SDNode *SN, SDNode *N, unsigned Offset) { |
406 | SDLoc dl(SN); |
407 | int FI = cast<FrameIndexSDNode>(N)->getIndex(); |
408 | SDValue TFI = CurDAG->getTargetFrameIndex(FI, N->getValueType(0)); |
409 | unsigned Opc = N->getValueType(0) == MVT::i32 ? PPC::ADDI : PPC::ADDI8; |
410 | if (SN->hasOneUse()) |
411 | return CurDAG->SelectNodeTo(SN, Opc, N->getValueType(0), TFI, |
412 | getSmallIPtrImm(Offset, dl)); |
413 | return CurDAG->getMachineNode(Opc, dl, N->getValueType(0), TFI, |
414 | getSmallIPtrImm(Offset, dl)); |
415 | } |
416 | |
417 | bool PPCDAGToDAGISel::isRotateAndMask(SDNode *N, unsigned Mask, |
418 | bool isShiftMask, unsigned &SH, |
419 | unsigned &MB, unsigned &ME) { |
420 | // Don't even go down this path for i64, since different logic will be |
421 | // necessary for rldicl/rldicr/rldimi. |
422 | if (N->getValueType(0) != MVT::i32) |
423 | return false; |
424 | |
425 | unsigned Shift = 32; |
426 | unsigned Indeterminant = ~0; // bit mask marking indeterminant results |
427 | unsigned Opcode = N->getOpcode(); |
428 | if (N->getNumOperands() != 2 || |
429 | !isInt32Immediate(N->getOperand(1).getNode(), Shift) || (Shift > 31)) |
430 | return false; |
431 | |
432 | if (Opcode == ISD::SHL) { |
433 | // apply shift left to mask if it comes first |
434 | if (isShiftMask) Mask = Mask << Shift; |
435 | // determine which bits are made indeterminant by shift |
436 | Indeterminant = ~(0xFFFFFFFFu << Shift); |
437 | } else if (Opcode == ISD::SRL) { |
438 | // apply shift right to mask if it comes first |
439 | if (isShiftMask) Mask = Mask >> Shift; |
440 | // determine which bits are made indeterminant by shift |
441 | Indeterminant = ~(0xFFFFFFFFu >> Shift); |
442 | // adjust for the left rotate |
443 | Shift = 32 - Shift; |
444 | } else if (Opcode == ISD::ROTL) { |
445 | Indeterminant = 0; |
446 | } else { |
447 | return false; |
448 | } |
449 | |
450 | // if the mask doesn't intersect any Indeterminant bits |
451 | if (Mask && !(Mask & Indeterminant)) { |
452 | SH = Shift & 31; |
453 | // make sure the mask is still a mask (wrap arounds may not be) |
454 | return isRunOfOnes(Mask, MB, ME); |
455 | } |
456 | return false; |
457 | } |
458 | |
459 | /// SelectBitfieldInsert - turn an or of two masked values into |
460 | /// the rotate left word immediate then mask insert (rlwimi) instruction. |
461 | SDNode *PPCDAGToDAGISel::SelectBitfieldInsert(SDNode *N) { |
462 | SDValue Op0 = N->getOperand(0); |
463 | SDValue Op1 = N->getOperand(1); |
464 | SDLoc dl(N); |
465 | |
466 | APInt LKZ, LKO, RKZ, RKO; |
467 | CurDAG->computeKnownBits(Op0, LKZ, LKO); |
468 | CurDAG->computeKnownBits(Op1, RKZ, RKO); |
469 | |
470 | unsigned TargetMask = LKZ.getZExtValue(); |
471 | unsigned InsertMask = RKZ.getZExtValue(); |
472 | |
473 | if ((TargetMask | InsertMask) == 0xFFFFFFFF) { |
474 | unsigned Op0Opc = Op0.getOpcode(); |
475 | unsigned Op1Opc = Op1.getOpcode(); |
476 | unsigned Value, SH = 0; |
477 | TargetMask = ~TargetMask; |
478 | InsertMask = ~InsertMask; |
479 | |
480 | // If the LHS has a foldable shift and the RHS does not, then swap it to the |
481 | // RHS so that we can fold the shift into the insert. |
482 | if (Op0Opc == ISD::AND && Op1Opc == ISD::AND) { |
483 | if (Op0.getOperand(0).getOpcode() == ISD::SHL || |
484 | Op0.getOperand(0).getOpcode() == ISD::SRL) { |
485 | if (Op1.getOperand(0).getOpcode() != ISD::SHL && |
486 | Op1.getOperand(0).getOpcode() != ISD::SRL) { |
487 | std::swap(Op0, Op1); |
488 | std::swap(Op0Opc, Op1Opc); |
489 | std::swap(TargetMask, InsertMask); |
490 | } |
491 | } |
492 | } else if (Op0Opc == ISD::SHL || Op0Opc == ISD::SRL) { |
493 | if (Op1Opc == ISD::AND && Op1.getOperand(0).getOpcode() != ISD::SHL && |
494 | Op1.getOperand(0).getOpcode() != ISD::SRL) { |
495 | std::swap(Op0, Op1); |
496 | std::swap(Op0Opc, Op1Opc); |
497 | std::swap(TargetMask, InsertMask); |
498 | } |
499 | } |
500 | |
501 | unsigned MB, ME; |
502 | if (isRunOfOnes(InsertMask, MB, ME)) { |
503 | SDValue Tmp1, Tmp2; |
504 | |
505 | if ((Op1Opc == ISD::SHL || Op1Opc == ISD::SRL) && |
506 | isInt32Immediate(Op1.getOperand(1), Value)) { |
507 | Op1 = Op1.getOperand(0); |
508 | SH = (Op1Opc == ISD::SHL) ? Value : 32 - Value; |
509 | } |
510 | if (Op1Opc == ISD::AND) { |
511 | // The AND mask might not be a constant, and we need to make sure that |
512 | // if we're going to fold the masking with the insert, all bits not |
513 | // know to be zero in the mask are known to be one. |
514 | APInt MKZ, MKO; |
515 | CurDAG->computeKnownBits(Op1.getOperand(1), MKZ, MKO); |
516 | bool CanFoldMask = InsertMask == MKO.getZExtValue(); |
517 | |
518 | unsigned SHOpc = Op1.getOperand(0).getOpcode(); |
519 | if ((SHOpc == ISD::SHL || SHOpc == ISD::SRL) && CanFoldMask && |
520 | isInt32Immediate(Op1.getOperand(0).getOperand(1), Value)) { |
521 | // Note that Value must be in range here (less than 32) because |
522 | // otherwise there would not be any bits set in InsertMask. |
523 | Op1 = Op1.getOperand(0).getOperand(0); |
524 | SH = (SHOpc == ISD::SHL) ? Value : 32 - Value; |
525 | } |
526 | } |
527 | |
528 | SH &= 31; |
529 | SDValue Ops[] = { Op0, Op1, getI32Imm(SH, dl), getI32Imm(MB, dl), |
530 | getI32Imm(ME, dl) }; |
531 | return CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops); |
532 | } |
533 | } |
534 | return nullptr; |
535 | } |
536 | |
537 | // Predict the number of instructions that would be generated by calling |
538 | // SelectInt64(N). |
539 | static unsigned SelectInt64CountDirect(int64_t Imm) { |
540 | // Assume no remaining bits. |
541 | unsigned Remainder = 0; |
542 | // Assume no shift required. |
543 | unsigned Shift = 0; |
544 | |
545 | // If it can't be represented as a 32 bit value. |
546 | if (!isInt<32>(Imm)) { |
547 | Shift = countTrailingZeros<uint64_t>(Imm); |
548 | int64_t ImmSh = static_cast<uint64_t>(Imm) >> Shift; |
549 | |
550 | // If the shifted value fits 32 bits. |
551 | if (isInt<32>(ImmSh)) { |
552 | // Go with the shifted value. |
553 | Imm = ImmSh; |
554 | } else { |
555 | // Still stuck with a 64 bit value. |
556 | Remainder = Imm; |
557 | Shift = 32; |
558 | Imm >>= 32; |
559 | } |
560 | } |
561 | |
562 | // Intermediate operand. |
563 | unsigned Result = 0; |
564 | |
565 | // Handle first 32 bits. |
566 | unsigned Lo = Imm & 0xFFFF; |
567 | unsigned Hi = (Imm >> 16) & 0xFFFF; |
Value stored to 'Hi' during its initialization is never read | |
568 | |
569 | // Simple value. |
570 | if (isInt<16>(Imm)) { |
571 | // Just the Lo bits. |
572 | ++Result; |
573 | } else if (Lo) { |
574 | // Handle the Hi bits and Lo bits. |
575 | Result += 2; |
576 | } else { |
577 | // Just the Hi bits. |
578 | ++Result; |
579 | } |
580 | |
581 | // If no shift, we're done. |
582 | if (!Shift) return Result; |
583 | |
584 | // Shift for next step if the upper 32-bits were not zero. |
585 | if (Imm) |
586 | ++Result; |
587 | |
588 | // Add in the last bits as required. |
589 | if ((Hi = (Remainder >> 16) & 0xFFFF)) |
590 | ++Result; |
591 | if ((Lo = Remainder & 0xFFFF)) |
592 | ++Result; |
593 | |
594 | return Result; |
595 | } |
596 | |
597 | static uint64_t Rot64(uint64_t Imm, unsigned R) { |
598 | return (Imm << R) | (Imm >> (64 - R)); |
599 | } |
600 | |
601 | static unsigned SelectInt64Count(int64_t Imm) { |
602 | unsigned Count = SelectInt64CountDirect(Imm); |
603 | if (Count == 1) |
604 | return Count; |
605 | |
606 | for (unsigned r = 1; r < 63; ++r) { |
607 | uint64_t RImm = Rot64(Imm, r); |
608 | unsigned RCount = SelectInt64CountDirect(RImm) + 1; |
609 | Count = std::min(Count, RCount); |
610 | |
611 | // See comments in SelectInt64 for an explanation of the logic below. |
612 | unsigned LS = findLastSet(RImm); |
613 | if (LS != r-1) |
614 | continue; |
615 | |
616 | uint64_t OnesMask = -(int64_t) (UINT64_C(1)1UL << (LS+1)); |
617 | uint64_t RImmWithOnes = RImm | OnesMask; |
618 | |
619 | RCount = SelectInt64CountDirect(RImmWithOnes) + 1; |
620 | Count = std::min(Count, RCount); |
621 | } |
622 | |
623 | return Count; |
624 | } |
625 | |
626 | // Select a 64-bit constant. For cost-modeling purposes, SelectInt64Count |
627 | // (above) needs to be kept in sync with this function. |
628 | static SDNode *SelectInt64Direct(SelectionDAG *CurDAG, SDLoc dl, int64_t Imm) { |
629 | // Assume no remaining bits. |
630 | unsigned Remainder = 0; |
631 | // Assume no shift required. |
632 | unsigned Shift = 0; |
633 | |
634 | // If it can't be represented as a 32 bit value. |
635 | if (!isInt<32>(Imm)) { |
636 | Shift = countTrailingZeros<uint64_t>(Imm); |
637 | int64_t ImmSh = static_cast<uint64_t>(Imm) >> Shift; |
638 | |
639 | // If the shifted value fits 32 bits. |
640 | if (isInt<32>(ImmSh)) { |
641 | // Go with the shifted value. |
642 | Imm = ImmSh; |
643 | } else { |
644 | // Still stuck with a 64 bit value. |
645 | Remainder = Imm; |
646 | Shift = 32; |
647 | Imm >>= 32; |
648 | } |
649 | } |
650 | |
651 | // Intermediate operand. |
652 | SDNode *Result; |
653 | |
654 | // Handle first 32 bits. |
655 | unsigned Lo = Imm & 0xFFFF; |
656 | unsigned Hi = (Imm >> 16) & 0xFFFF; |
657 | |
658 | auto getI32Imm = [CurDAG, dl](unsigned Imm) { |
659 | return CurDAG->getTargetConstant(Imm, dl, MVT::i32); |
660 | }; |
661 | |
662 | // Simple value. |
663 | if (isInt<16>(Imm)) { |
664 | // Just the Lo bits. |
665 | Result = CurDAG->getMachineNode(PPC::LI8, dl, MVT::i64, getI32Imm(Lo)); |
666 | } else if (Lo) { |
667 | // Handle the Hi bits. |
668 | unsigned OpC = Hi ? PPC::LIS8 : PPC::LI8; |
669 | Result = CurDAG->getMachineNode(OpC, dl, MVT::i64, getI32Imm(Hi)); |
670 | // And Lo bits. |
671 | Result = CurDAG->getMachineNode(PPC::ORI8, dl, MVT::i64, |
672 | SDValue(Result, 0), getI32Imm(Lo)); |
673 | } else { |
674 | // Just the Hi bits. |
675 | Result = CurDAG->getMachineNode(PPC::LIS8, dl, MVT::i64, getI32Imm(Hi)); |
676 | } |
677 | |
678 | // If no shift, we're done. |
679 | if (!Shift) return Result; |
680 | |
681 | // Shift for next step if the upper 32-bits were not zero. |
682 | if (Imm) { |
683 | Result = CurDAG->getMachineNode(PPC::RLDICR, dl, MVT::i64, |
684 | SDValue(Result, 0), |
685 | getI32Imm(Shift), |
686 | getI32Imm(63 - Shift)); |
687 | } |
688 | |
689 | // Add in the last bits as required. |
690 | if ((Hi = (Remainder >> 16) & 0xFFFF)) { |
691 | Result = CurDAG->getMachineNode(PPC::ORIS8, dl, MVT::i64, |
692 | SDValue(Result, 0), getI32Imm(Hi)); |
693 | } |
694 | if ((Lo = Remainder & 0xFFFF)) { |
695 | Result = CurDAG->getMachineNode(PPC::ORI8, dl, MVT::i64, |
696 | SDValue(Result, 0), getI32Imm(Lo)); |
697 | } |
698 | |
699 | return Result; |
700 | } |
701 | |
702 | static SDNode *SelectInt64(SelectionDAG *CurDAG, SDLoc dl, int64_t Imm) { |
703 | unsigned Count = SelectInt64CountDirect(Imm); |
704 | if (Count == 1) |
705 | return SelectInt64Direct(CurDAG, dl, Imm); |
706 | |
707 | unsigned RMin = 0; |
708 | |
709 | int64_t MatImm; |
710 | unsigned MaskEnd; |
711 | |
712 | for (unsigned r = 1; r < 63; ++r) { |
713 | uint64_t RImm = Rot64(Imm, r); |
714 | unsigned RCount = SelectInt64CountDirect(RImm) + 1; |
715 | if (RCount < Count) { |
716 | Count = RCount; |
717 | RMin = r; |
718 | MatImm = RImm; |
719 | MaskEnd = 63; |
720 | } |
721 | |
722 | // If the immediate to generate has many trailing zeros, it might be |
723 | // worthwhile to generate a rotated value with too many leading ones |
724 | // (because that's free with li/lis's sign-extension semantics), and then |
725 | // mask them off after rotation. |
726 | |
727 | unsigned LS = findLastSet(RImm); |
728 | // We're adding (63-LS) higher-order ones, and we expect to mask them off |
729 | // after performing the inverse rotation by (64-r). So we need that: |
730 | // 63-LS == 64-r => LS == r-1 |
731 | if (LS != r-1) |
732 | continue; |
733 | |
734 | uint64_t OnesMask = -(int64_t) (UINT64_C(1)1UL << (LS+1)); |
735 | uint64_t RImmWithOnes = RImm | OnesMask; |
736 | |
737 | RCount = SelectInt64CountDirect(RImmWithOnes) + 1; |
738 | if (RCount < Count) { |
739 | Count = RCount; |
740 | RMin = r; |
741 | MatImm = RImmWithOnes; |
742 | MaskEnd = LS; |
743 | } |
744 | } |
745 | |
746 | if (!RMin) |
747 | return SelectInt64Direct(CurDAG, dl, Imm); |
748 | |
749 | auto getI32Imm = [CurDAG, dl](unsigned Imm) { |
750 | return CurDAG->getTargetConstant(Imm, dl, MVT::i32); |
751 | }; |
752 | |
753 | SDValue Val = SDValue(SelectInt64Direct(CurDAG, dl, MatImm), 0); |
754 | return CurDAG->getMachineNode(PPC::RLDICR, dl, MVT::i64, Val, |
755 | getI32Imm(64 - RMin), getI32Imm(MaskEnd)); |
756 | } |
757 | |
758 | // Select a 64-bit constant. |
759 | static SDNode *SelectInt64(SelectionDAG *CurDAG, SDNode *N) { |
760 | SDLoc dl(N); |
761 | |
762 | // Get 64 bit value. |
763 | int64_t Imm = cast<ConstantSDNode>(N)->getZExtValue(); |
764 | return SelectInt64(CurDAG, dl, Imm); |
765 | } |
766 | |
767 | namespace { |
768 | class BitPermutationSelector { |
769 | struct ValueBit { |
770 | SDValue V; |
771 | |
772 | // The bit number in the value, using a convention where bit 0 is the |
773 | // lowest-order bit. |
774 | unsigned Idx; |
775 | |
776 | enum Kind { |
777 | ConstZero, |
778 | Variable |
779 | } K; |
780 | |
781 | ValueBit(SDValue V, unsigned I, Kind K = Variable) |
782 | : V(V), Idx(I), K(K) {} |
783 | ValueBit(Kind K = Variable) |
784 | : V(SDValue(nullptr, 0)), Idx(UINT32_MAX(4294967295U)), K(K) {} |
785 | |
786 | bool isZero() const { |
787 | return K == ConstZero; |
788 | } |
789 | |
790 | bool hasValue() const { |
791 | return K == Variable; |
792 | } |
793 | |
794 | SDValue getValue() const { |
795 | assert(hasValue() && "Cannot get the value of a constant bit")((hasValue() && "Cannot get the value of a constant bit" ) ? static_cast<void> (0) : __assert_fail ("hasValue() && \"Cannot get the value of a constant bit\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 795, __PRETTY_FUNCTION__)); |
796 | return V; |
797 | } |
798 | |
799 | unsigned getValueBitIndex() const { |
800 | assert(hasValue() && "Cannot get the value bit index of a constant bit")((hasValue() && "Cannot get the value bit index of a constant bit" ) ? static_cast<void> (0) : __assert_fail ("hasValue() && \"Cannot get the value bit index of a constant bit\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 800, __PRETTY_FUNCTION__)); |
801 | return Idx; |
802 | } |
803 | }; |
804 | |
805 | // A bit group has the same underlying value and the same rotate factor. |
806 | struct BitGroup { |
807 | SDValue V; |
808 | unsigned RLAmt; |
809 | unsigned StartIdx, EndIdx; |
810 | |
811 | // This rotation amount assumes that the lower 32 bits of the quantity are |
812 | // replicated in the high 32 bits by the rotation operator (which is done |
813 | // by rlwinm and friends in 64-bit mode). |
814 | bool Repl32; |
815 | // Did converting to Repl32 == true change the rotation factor? If it did, |
816 | // it decreased it by 32. |
817 | bool Repl32CR; |
818 | // Was this group coalesced after setting Repl32 to true? |
819 | bool Repl32Coalesced; |
820 | |
821 | BitGroup(SDValue V, unsigned R, unsigned S, unsigned E) |
822 | : V(V), RLAmt(R), StartIdx(S), EndIdx(E), Repl32(false), Repl32CR(false), |
823 | Repl32Coalesced(false) { |
824 | DEBUG(dbgs() << "\tbit group for " << V.getNode() << " RLAmt = " << R <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\tbit group for " << V.getNode() << " RLAmt = " << R << " [" << S << ", " << E << "]\n"; } } while (0) |
825 | " [" << S << ", " << E << "]\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\tbit group for " << V.getNode() << " RLAmt = " << R << " [" << S << ", " << E << "]\n"; } } while (0); |
826 | } |
827 | }; |
828 | |
829 | // Information on each (Value, RLAmt) pair (like the number of groups |
830 | // associated with each) used to choose the lowering method. |
831 | struct ValueRotInfo { |
832 | SDValue V; |
833 | unsigned RLAmt; |
834 | unsigned NumGroups; |
835 | unsigned FirstGroupStartIdx; |
836 | bool Repl32; |
837 | |
838 | ValueRotInfo() |
839 | : RLAmt(UINT32_MAX(4294967295U)), NumGroups(0), FirstGroupStartIdx(UINT32_MAX(4294967295U)), |
840 | Repl32(false) {} |
841 | |
842 | // For sorting (in reverse order) by NumGroups, and then by |
843 | // FirstGroupStartIdx. |
844 | bool operator < (const ValueRotInfo &Other) const { |
845 | // We need to sort so that the non-Repl32 come first because, when we're |
846 | // doing masking, the Repl32 bit groups might be subsumed into the 64-bit |
847 | // masking operation. |
848 | if (Repl32 < Other.Repl32) |
849 | return true; |
850 | else if (Repl32 > Other.Repl32) |
851 | return false; |
852 | else if (NumGroups > Other.NumGroups) |
853 | return true; |
854 | else if (NumGroups < Other.NumGroups) |
855 | return false; |
856 | else if (FirstGroupStartIdx < Other.FirstGroupStartIdx) |
857 | return true; |
858 | return false; |
859 | } |
860 | }; |
861 | |
862 | // Return true if something interesting was deduced, return false if we're |
863 | // providing only a generic representation of V (or something else likewise |
864 | // uninteresting for instruction selection). |
865 | bool getValueBits(SDValue V, SmallVector<ValueBit, 64> &Bits) { |
866 | switch (V.getOpcode()) { |
867 | default: break; |
868 | case ISD::ROTL: |
869 | if (isa<ConstantSDNode>(V.getOperand(1))) { |
870 | unsigned RotAmt = V.getConstantOperandVal(1); |
871 | |
872 | SmallVector<ValueBit, 64> LHSBits(Bits.size()); |
873 | getValueBits(V.getOperand(0), LHSBits); |
874 | |
875 | for (unsigned i = 0; i < Bits.size(); ++i) |
876 | Bits[i] = LHSBits[i < RotAmt ? i + (Bits.size() - RotAmt) : i - RotAmt]; |
877 | |
878 | return true; |
879 | } |
880 | break; |
881 | case ISD::SHL: |
882 | if (isa<ConstantSDNode>(V.getOperand(1))) { |
883 | unsigned ShiftAmt = V.getConstantOperandVal(1); |
884 | |
885 | SmallVector<ValueBit, 64> LHSBits(Bits.size()); |
886 | getValueBits(V.getOperand(0), LHSBits); |
887 | |
888 | for (unsigned i = ShiftAmt; i < Bits.size(); ++i) |
889 | Bits[i] = LHSBits[i - ShiftAmt]; |
890 | |
891 | for (unsigned i = 0; i < ShiftAmt; ++i) |
892 | Bits[i] = ValueBit(ValueBit::ConstZero); |
893 | |
894 | return true; |
895 | } |
896 | break; |
897 | case ISD::SRL: |
898 | if (isa<ConstantSDNode>(V.getOperand(1))) { |
899 | unsigned ShiftAmt = V.getConstantOperandVal(1); |
900 | |
901 | SmallVector<ValueBit, 64> LHSBits(Bits.size()); |
902 | getValueBits(V.getOperand(0), LHSBits); |
903 | |
904 | for (unsigned i = 0; i < Bits.size() - ShiftAmt; ++i) |
905 | Bits[i] = LHSBits[i + ShiftAmt]; |
906 | |
907 | for (unsigned i = Bits.size() - ShiftAmt; i < Bits.size(); ++i) |
908 | Bits[i] = ValueBit(ValueBit::ConstZero); |
909 | |
910 | return true; |
911 | } |
912 | break; |
913 | case ISD::AND: |
914 | if (isa<ConstantSDNode>(V.getOperand(1))) { |
915 | uint64_t Mask = V.getConstantOperandVal(1); |
916 | |
917 | SmallVector<ValueBit, 64> LHSBits(Bits.size()); |
918 | bool LHSTrivial = getValueBits(V.getOperand(0), LHSBits); |
919 | |
920 | for (unsigned i = 0; i < Bits.size(); ++i) |
921 | if (((Mask >> i) & 1) == 1) |
922 | Bits[i] = LHSBits[i]; |
923 | else |
924 | Bits[i] = ValueBit(ValueBit::ConstZero); |
925 | |
926 | // Mark this as interesting, only if the LHS was also interesting. This |
927 | // prevents the overall procedure from matching a single immediate 'and' |
928 | // (which is non-optimal because such an and might be folded with other |
929 | // things if we don't select it here). |
930 | return LHSTrivial; |
931 | } |
932 | break; |
933 | case ISD::OR: { |
934 | SmallVector<ValueBit, 64> LHSBits(Bits.size()), RHSBits(Bits.size()); |
935 | getValueBits(V.getOperand(0), LHSBits); |
936 | getValueBits(V.getOperand(1), RHSBits); |
937 | |
938 | bool AllDisjoint = true; |
939 | for (unsigned i = 0; i < Bits.size(); ++i) |
940 | if (LHSBits[i].isZero()) |
941 | Bits[i] = RHSBits[i]; |
942 | else if (RHSBits[i].isZero()) |
943 | Bits[i] = LHSBits[i]; |
944 | else { |
945 | AllDisjoint = false; |
946 | break; |
947 | } |
948 | |
949 | if (!AllDisjoint) |
950 | break; |
951 | |
952 | return true; |
953 | } |
954 | } |
955 | |
956 | for (unsigned i = 0; i < Bits.size(); ++i) |
957 | Bits[i] = ValueBit(V, i); |
958 | |
959 | return false; |
960 | } |
961 | |
962 | // For each value (except the constant ones), compute the left-rotate amount |
963 | // to get it from its original to final position. |
964 | void computeRotationAmounts() { |
965 | HasZeros = false; |
966 | RLAmt.resize(Bits.size()); |
967 | for (unsigned i = 0; i < Bits.size(); ++i) |
968 | if (Bits[i].hasValue()) { |
969 | unsigned VBI = Bits[i].getValueBitIndex(); |
970 | if (i >= VBI) |
971 | RLAmt[i] = i - VBI; |
972 | else |
973 | RLAmt[i] = Bits.size() - (VBI - i); |
974 | } else if (Bits[i].isZero()) { |
975 | HasZeros = true; |
976 | RLAmt[i] = UINT32_MAX(4294967295U); |
977 | } else { |
978 | llvm_unreachable("Unknown value bit type")::llvm::llvm_unreachable_internal("Unknown value bit type", "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 978); |
979 | } |
980 | } |
981 | |
982 | // Collect groups of consecutive bits with the same underlying value and |
983 | // rotation factor. If we're doing late masking, we ignore zeros, otherwise |
984 | // they break up groups. |
985 | void collectBitGroups(bool LateMask) { |
986 | BitGroups.clear(); |
987 | |
988 | unsigned LastRLAmt = RLAmt[0]; |
989 | SDValue LastValue = Bits[0].hasValue() ? Bits[0].getValue() : SDValue(); |
990 | unsigned LastGroupStartIdx = 0; |
991 | for (unsigned i = 1; i < Bits.size(); ++i) { |
992 | unsigned ThisRLAmt = RLAmt[i]; |
993 | SDValue ThisValue = Bits[i].hasValue() ? Bits[i].getValue() : SDValue(); |
994 | if (LateMask && !ThisValue) { |
995 | ThisValue = LastValue; |
996 | ThisRLAmt = LastRLAmt; |
997 | // If we're doing late masking, then the first bit group always starts |
998 | // at zero (even if the first bits were zero). |
999 | if (BitGroups.empty()) |
1000 | LastGroupStartIdx = 0; |
1001 | } |
1002 | |
1003 | // If this bit has the same underlying value and the same rotate factor as |
1004 | // the last one, then they're part of the same group. |
1005 | if (ThisRLAmt == LastRLAmt && ThisValue == LastValue) |
1006 | continue; |
1007 | |
1008 | if (LastValue.getNode()) |
1009 | BitGroups.push_back(BitGroup(LastValue, LastRLAmt, LastGroupStartIdx, |
1010 | i-1)); |
1011 | LastRLAmt = ThisRLAmt; |
1012 | LastValue = ThisValue; |
1013 | LastGroupStartIdx = i; |
1014 | } |
1015 | if (LastValue.getNode()) |
1016 | BitGroups.push_back(BitGroup(LastValue, LastRLAmt, LastGroupStartIdx, |
1017 | Bits.size()-1)); |
1018 | |
1019 | if (BitGroups.empty()) |
1020 | return; |
1021 | |
1022 | // We might be able to combine the first and last groups. |
1023 | if (BitGroups.size() > 1) { |
1024 | // If the first and last groups are the same, then remove the first group |
1025 | // in favor of the last group, making the ending index of the last group |
1026 | // equal to the ending index of the to-be-removed first group. |
1027 | if (BitGroups[0].StartIdx == 0 && |
1028 | BitGroups[BitGroups.size()-1].EndIdx == Bits.size()-1 && |
1029 | BitGroups[0].V == BitGroups[BitGroups.size()-1].V && |
1030 | BitGroups[0].RLAmt == BitGroups[BitGroups.size()-1].RLAmt) { |
1031 | DEBUG(dbgs() << "\tcombining final bit group with initial one\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\tcombining final bit group with initial one\n" ; } } while (0); |
1032 | BitGroups[BitGroups.size()-1].EndIdx = BitGroups[0].EndIdx; |
1033 | BitGroups.erase(BitGroups.begin()); |
1034 | } |
1035 | } |
1036 | } |
1037 | |
1038 | // Take all (SDValue, RLAmt) pairs and sort them by the number of groups |
1039 | // associated with each. If there is a degeneracy, pick the one that occurs |
1040 | // first (in the final value). |
1041 | void collectValueRotInfo() { |
1042 | ValueRots.clear(); |
1043 | |
1044 | for (auto &BG : BitGroups) { |
1045 | unsigned RLAmtKey = BG.RLAmt + (BG.Repl32 ? 64 : 0); |
1046 | ValueRotInfo &VRI = ValueRots[std::make_pair(BG.V, RLAmtKey)]; |
1047 | VRI.V = BG.V; |
1048 | VRI.RLAmt = BG.RLAmt; |
1049 | VRI.Repl32 = BG.Repl32; |
1050 | VRI.NumGroups += 1; |
1051 | VRI.FirstGroupStartIdx = std::min(VRI.FirstGroupStartIdx, BG.StartIdx); |
1052 | } |
1053 | |
1054 | // Now that we've collected the various ValueRotInfo instances, we need to |
1055 | // sort them. |
1056 | ValueRotsVec.clear(); |
1057 | for (auto &I : ValueRots) { |
1058 | ValueRotsVec.push_back(I.second); |
1059 | } |
1060 | std::sort(ValueRotsVec.begin(), ValueRotsVec.end()); |
1061 | } |
1062 | |
1063 | // In 64-bit mode, rlwinm and friends have a rotation operator that |
1064 | // replicates the low-order 32 bits into the high-order 32-bits. The mask |
1065 | // indices of these instructions can only be in the lower 32 bits, so they |
1066 | // can only represent some 64-bit bit groups. However, when they can be used, |
1067 | // the 32-bit replication can be used to represent, as a single bit group, |
1068 | // otherwise separate bit groups. We'll convert to replicated-32-bit bit |
1069 | // groups when possible. Returns true if any of the bit groups were |
1070 | // converted. |
1071 | void assignRepl32BitGroups() { |
1072 | // If we have bits like this: |
1073 | // |
1074 | // Indices: 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0 |
1075 | // V bits: ... 7 6 5 4 3 2 1 0 31 30 29 28 27 26 25 24 |
1076 | // Groups: | RLAmt = 8 | RLAmt = 40 | |
1077 | // |
1078 | // But, making use of a 32-bit operation that replicates the low-order 32 |
1079 | // bits into the high-order 32 bits, this can be one bit group with a RLAmt |
1080 | // of 8. |
1081 | |
1082 | auto IsAllLow32 = [this](BitGroup & BG) { |
1083 | if (BG.StartIdx <= BG.EndIdx) { |
1084 | for (unsigned i = BG.StartIdx; i <= BG.EndIdx; ++i) { |
1085 | if (!Bits[i].hasValue()) |
1086 | continue; |
1087 | if (Bits[i].getValueBitIndex() >= 32) |
1088 | return false; |
1089 | } |
1090 | } else { |
1091 | for (unsigned i = BG.StartIdx; i < Bits.size(); ++i) { |
1092 | if (!Bits[i].hasValue()) |
1093 | continue; |
1094 | if (Bits[i].getValueBitIndex() >= 32) |
1095 | return false; |
1096 | } |
1097 | for (unsigned i = 0; i <= BG.EndIdx; ++i) { |
1098 | if (!Bits[i].hasValue()) |
1099 | continue; |
1100 | if (Bits[i].getValueBitIndex() >= 32) |
1101 | return false; |
1102 | } |
1103 | } |
1104 | |
1105 | return true; |
1106 | }; |
1107 | |
1108 | for (auto &BG : BitGroups) { |
1109 | if (BG.StartIdx < 32 && BG.EndIdx < 32) { |
1110 | if (IsAllLow32(BG)) { |
1111 | if (BG.RLAmt >= 32) { |
1112 | BG.RLAmt -= 32; |
1113 | BG.Repl32CR = true; |
1114 | } |
1115 | |
1116 | BG.Repl32 = true; |
1117 | |
1118 | DEBUG(dbgs() << "\t32-bit replicated bit group for " <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\t32-bit replicated bit group for " << BG.V.getNode() << " RLAmt = " << BG.RLAmt << " [" << BG.StartIdx << ", " << BG .EndIdx << "]\n"; } } while (0) |
1119 | BG.V.getNode() << " RLAmt = " << BG.RLAmt <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\t32-bit replicated bit group for " << BG.V.getNode() << " RLAmt = " << BG.RLAmt << " [" << BG.StartIdx << ", " << BG .EndIdx << "]\n"; } } while (0) |
1120 | " [" << BG.StartIdx << ", " << BG.EndIdx << "]\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\t32-bit replicated bit group for " << BG.V.getNode() << " RLAmt = " << BG.RLAmt << " [" << BG.StartIdx << ", " << BG .EndIdx << "]\n"; } } while (0); |
1121 | } |
1122 | } |
1123 | } |
1124 | |
1125 | // Now walk through the bit groups, consolidating where possible. |
1126 | for (auto I = BitGroups.begin(); I != BitGroups.end();) { |
1127 | // We might want to remove this bit group by merging it with the previous |
1128 | // group (which might be the ending group). |
1129 | auto IP = (I == BitGroups.begin()) ? |
1130 | std::prev(BitGroups.end()) : std::prev(I); |
1131 | if (I->Repl32 && IP->Repl32 && I->V == IP->V && I->RLAmt == IP->RLAmt && |
1132 | I->StartIdx == (IP->EndIdx + 1) % 64 && I != IP) { |
1133 | |
1134 | DEBUG(dbgs() << "\tcombining 32-bit replicated bit group for " <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\tcombining 32-bit replicated bit group for " << I->V.getNode() << " RLAmt = " << I-> RLAmt << " [" << I->StartIdx << ", " << I->EndIdx << "] with group with range [" << IP ->StartIdx << ", " << IP->EndIdx << "]\n" ; } } while (0) |
1135 | I->V.getNode() << " RLAmt = " << I->RLAmt <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\tcombining 32-bit replicated bit group for " << I->V.getNode() << " RLAmt = " << I-> RLAmt << " [" << I->StartIdx << ", " << I->EndIdx << "] with group with range [" << IP ->StartIdx << ", " << IP->EndIdx << "]\n" ; } } while (0) |
1136 | " [" << I->StartIdx << ", " << I->EndIdx <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\tcombining 32-bit replicated bit group for " << I->V.getNode() << " RLAmt = " << I-> RLAmt << " [" << I->StartIdx << ", " << I->EndIdx << "] with group with range [" << IP ->StartIdx << ", " << IP->EndIdx << "]\n" ; } } while (0) |
1137 | "] with group with range [" <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\tcombining 32-bit replicated bit group for " << I->V.getNode() << " RLAmt = " << I-> RLAmt << " [" << I->StartIdx << ", " << I->EndIdx << "] with group with range [" << IP ->StartIdx << ", " << IP->EndIdx << "]\n" ; } } while (0) |
1138 | IP->StartIdx << ", " << IP->EndIdx << "]\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\tcombining 32-bit replicated bit group for " << I->V.getNode() << " RLAmt = " << I-> RLAmt << " [" << I->StartIdx << ", " << I->EndIdx << "] with group with range [" << IP ->StartIdx << ", " << IP->EndIdx << "]\n" ; } } while (0); |
1139 | |
1140 | IP->EndIdx = I->EndIdx; |
1141 | IP->Repl32CR = IP->Repl32CR || I->Repl32CR; |
1142 | IP->Repl32Coalesced = true; |
1143 | I = BitGroups.erase(I); |
1144 | continue; |
1145 | } else { |
1146 | // There is a special case worth handling: If there is a single group |
1147 | // covering the entire upper 32 bits, and it can be merged with both |
1148 | // the next and previous groups (which might be the same group), then |
1149 | // do so. If it is the same group (so there will be only one group in |
1150 | // total), then we need to reverse the order of the range so that it |
1151 | // covers the entire 64 bits. |
1152 | if (I->StartIdx == 32 && I->EndIdx == 63) { |
1153 | assert(std::next(I) == BitGroups.end() &&((std::next(I) == BitGroups.end() && "bit group ends at index 63 but there is another?" ) ? static_cast<void> (0) : __assert_fail ("std::next(I) == BitGroups.end() && \"bit group ends at index 63 but there is another?\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 1154, __PRETTY_FUNCTION__)) |
1154 | "bit group ends at index 63 but there is another?")((std::next(I) == BitGroups.end() && "bit group ends at index 63 but there is another?" ) ? static_cast<void> (0) : __assert_fail ("std::next(I) == BitGroups.end() && \"bit group ends at index 63 but there is another?\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 1154, __PRETTY_FUNCTION__)); |
1155 | auto IN = BitGroups.begin(); |
1156 | |
1157 | if (IP->Repl32 && IN->Repl32 && I->V == IP->V && I->V == IN->V && |
1158 | (I->RLAmt % 32) == IP->RLAmt && (I->RLAmt % 32) == IN->RLAmt && |
1159 | IP->EndIdx == 31 && IN->StartIdx == 0 && I != IP && |
1160 | IsAllLow32(*I)) { |
1161 | |
1162 | DEBUG(dbgs() << "\tcombining bit group for " <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\tcombining bit group for " << I->V.getNode() << " RLAmt = " << I-> RLAmt << " [" << I->StartIdx << ", " << I->EndIdx << "] with 32-bit replicated groups with ranges [" << IP->StartIdx << ", " << IP->EndIdx << "] and [" << IN->StartIdx << ", " << IN->EndIdx << "]\n"; } } while (0) |
1163 | I->V.getNode() << " RLAmt = " << I->RLAmt <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\tcombining bit group for " << I->V.getNode() << " RLAmt = " << I-> RLAmt << " [" << I->StartIdx << ", " << I->EndIdx << "] with 32-bit replicated groups with ranges [" << IP->StartIdx << ", " << IP->EndIdx << "] and [" << IN->StartIdx << ", " << IN->EndIdx << "]\n"; } } while (0) |
1164 | " [" << I->StartIdx << ", " << I->EndIdx <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\tcombining bit group for " << I->V.getNode() << " RLAmt = " << I-> RLAmt << " [" << I->StartIdx << ", " << I->EndIdx << "] with 32-bit replicated groups with ranges [" << IP->StartIdx << ", " << IP->EndIdx << "] and [" << IN->StartIdx << ", " << IN->EndIdx << "]\n"; } } while (0) |
1165 | "] with 32-bit replicated groups with ranges [" <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\tcombining bit group for " << I->V.getNode() << " RLAmt = " << I-> RLAmt << " [" << I->StartIdx << ", " << I->EndIdx << "] with 32-bit replicated groups with ranges [" << IP->StartIdx << ", " << IP->EndIdx << "] and [" << IN->StartIdx << ", " << IN->EndIdx << "]\n"; } } while (0) |
1166 | IP->StartIdx << ", " << IP->EndIdx << "] and [" <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\tcombining bit group for " << I->V.getNode() << " RLAmt = " << I-> RLAmt << " [" << I->StartIdx << ", " << I->EndIdx << "] with 32-bit replicated groups with ranges [" << IP->StartIdx << ", " << IP->EndIdx << "] and [" << IN->StartIdx << ", " << IN->EndIdx << "]\n"; } } while (0) |
1167 | IN->StartIdx << ", " << IN->EndIdx << "]\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\tcombining bit group for " << I->V.getNode() << " RLAmt = " << I-> RLAmt << " [" << I->StartIdx << ", " << I->EndIdx << "] with 32-bit replicated groups with ranges [" << IP->StartIdx << ", " << IP->EndIdx << "] and [" << IN->StartIdx << ", " << IN->EndIdx << "]\n"; } } while (0); |
1168 | |
1169 | if (IP == IN) { |
1170 | // There is only one other group; change it to cover the whole |
1171 | // range (backward, so that it can still be Repl32 but cover the |
1172 | // whole 64-bit range). |
1173 | IP->StartIdx = 31; |
1174 | IP->EndIdx = 30; |
1175 | IP->Repl32CR = IP->Repl32CR || I->RLAmt >= 32; |
1176 | IP->Repl32Coalesced = true; |
1177 | I = BitGroups.erase(I); |
1178 | } else { |
1179 | // There are two separate groups, one before this group and one |
1180 | // after us (at the beginning). We're going to remove this group, |
1181 | // but also the group at the very beginning. |
1182 | IP->EndIdx = IN->EndIdx; |
1183 | IP->Repl32CR = IP->Repl32CR || IN->Repl32CR || I->RLAmt >= 32; |
1184 | IP->Repl32Coalesced = true; |
1185 | I = BitGroups.erase(I); |
1186 | BitGroups.erase(BitGroups.begin()); |
1187 | } |
1188 | |
1189 | // This must be the last group in the vector (and we might have |
1190 | // just invalidated the iterator above), so break here. |
1191 | break; |
1192 | } |
1193 | } |
1194 | } |
1195 | |
1196 | ++I; |
1197 | } |
1198 | } |
1199 | |
1200 | SDValue getI32Imm(unsigned Imm, SDLoc dl) { |
1201 | return CurDAG->getTargetConstant(Imm, dl, MVT::i32); |
1202 | } |
1203 | |
1204 | uint64_t getZerosMask() { |
1205 | uint64_t Mask = 0; |
1206 | for (unsigned i = 0; i < Bits.size(); ++i) { |
1207 | if (Bits[i].hasValue()) |
1208 | continue; |
1209 | Mask |= (UINT64_C(1)1UL << i); |
1210 | } |
1211 | |
1212 | return ~Mask; |
1213 | } |
1214 | |
1215 | // Depending on the number of groups for a particular value, it might be |
1216 | // better to rotate, mask explicitly (using andi/andis), and then or the |
1217 | // result. Select this part of the result first. |
1218 | void SelectAndParts32(SDLoc dl, SDValue &Res, unsigned *InstCnt) { |
1219 | if (BPermRewriterNoMasking) |
1220 | return; |
1221 | |
1222 | for (ValueRotInfo &VRI : ValueRotsVec) { |
1223 | unsigned Mask = 0; |
1224 | for (unsigned i = 0; i < Bits.size(); ++i) { |
1225 | if (!Bits[i].hasValue() || Bits[i].getValue() != VRI.V) |
1226 | continue; |
1227 | if (RLAmt[i] != VRI.RLAmt) |
1228 | continue; |
1229 | Mask |= (1u << i); |
1230 | } |
1231 | |
1232 | // Compute the masks for andi/andis that would be necessary. |
1233 | unsigned ANDIMask = (Mask & UINT16_MAX(65535)), ANDISMask = Mask >> 16; |
1234 | assert((ANDIMask != 0 || ANDISMask != 0) &&(((ANDIMask != 0 || ANDISMask != 0) && "No set bits in mask for value bit groups" ) ? static_cast<void> (0) : __assert_fail ("(ANDIMask != 0 || ANDISMask != 0) && \"No set bits in mask for value bit groups\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 1235, __PRETTY_FUNCTION__)) |
1235 | "No set bits in mask for value bit groups")(((ANDIMask != 0 || ANDISMask != 0) && "No set bits in mask for value bit groups" ) ? static_cast<void> (0) : __assert_fail ("(ANDIMask != 0 || ANDISMask != 0) && \"No set bits in mask for value bit groups\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 1235, __PRETTY_FUNCTION__)); |
1236 | bool NeedsRotate = VRI.RLAmt != 0; |
1237 | |
1238 | // We're trying to minimize the number of instructions. If we have one |
1239 | // group, using one of andi/andis can break even. If we have three |
1240 | // groups, we can use both andi and andis and break even (to use both |
1241 | // andi and andis we also need to or the results together). We need four |
1242 | // groups if we also need to rotate. To use andi/andis we need to do more |
1243 | // than break even because rotate-and-mask instructions tend to be easier |
1244 | // to schedule. |
1245 | |
1246 | // FIXME: We've biased here against using andi/andis, which is right for |
1247 | // POWER cores, but not optimal everywhere. For example, on the A2, |
1248 | // andi/andis have single-cycle latency whereas the rotate-and-mask |
1249 | // instructions take two cycles, and it would be better to bias toward |
1250 | // andi/andis in break-even cases. |
1251 | |
1252 | unsigned NumAndInsts = (unsigned) NeedsRotate + |
1253 | (unsigned) (ANDIMask != 0) + |
1254 | (unsigned) (ANDISMask != 0) + |
1255 | (unsigned) (ANDIMask != 0 && ANDISMask != 0) + |
1256 | (unsigned) (bool) Res; |
1257 | |
1258 | DEBUG(dbgs() << "\t\trotation groups for " << VRI.V.getNode() <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\t\trotation groups for " << VRI.V.getNode() << " RL: " << VRI.RLAmt << ":" << "\n\t\t\tisel using masking: " << NumAndInsts << " using rotates: " << VRI.NumGroups << "\n"; } } while (0) |
1259 | " RL: " << VRI.RLAmt << ":" <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\t\trotation groups for " << VRI.V.getNode() << " RL: " << VRI.RLAmt << ":" << "\n\t\t\tisel using masking: " << NumAndInsts << " using rotates: " << VRI.NumGroups << "\n"; } } while (0) |
1260 | "\n\t\t\tisel using masking: " << NumAndInsts <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\t\trotation groups for " << VRI.V.getNode() << " RL: " << VRI.RLAmt << ":" << "\n\t\t\tisel using masking: " << NumAndInsts << " using rotates: " << VRI.NumGroups << "\n"; } } while (0) |
1261 | " using rotates: " << VRI.NumGroups << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\t\trotation groups for " << VRI.V.getNode() << " RL: " << VRI.RLAmt << ":" << "\n\t\t\tisel using masking: " << NumAndInsts << " using rotates: " << VRI.NumGroups << "\n"; } } while (0); |
1262 | |
1263 | if (NumAndInsts >= VRI.NumGroups) |
1264 | continue; |
1265 | |
1266 | DEBUG(dbgs() << "\t\t\t\tusing masking\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\t\t\t\tusing masking\n"; } } while (0); |
1267 | |
1268 | if (InstCnt) *InstCnt += NumAndInsts; |
1269 | |
1270 | SDValue VRot; |
1271 | if (VRI.RLAmt) { |
1272 | SDValue Ops[] = |
1273 | { VRI.V, getI32Imm(VRI.RLAmt, dl), getI32Imm(0, dl), |
1274 | getI32Imm(31, dl) }; |
1275 | VRot = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, |
1276 | Ops), 0); |
1277 | } else { |
1278 | VRot = VRI.V; |
1279 | } |
1280 | |
1281 | SDValue ANDIVal, ANDISVal; |
1282 | if (ANDIMask != 0) |
1283 | ANDIVal = SDValue(CurDAG->getMachineNode(PPC::ANDIo, dl, MVT::i32, |
1284 | VRot, getI32Imm(ANDIMask, dl)), 0); |
1285 | if (ANDISMask != 0) |
1286 | ANDISVal = SDValue(CurDAG->getMachineNode(PPC::ANDISo, dl, MVT::i32, |
1287 | VRot, getI32Imm(ANDISMask, dl)), 0); |
1288 | |
1289 | SDValue TotalVal; |
1290 | if (!ANDIVal) |
1291 | TotalVal = ANDISVal; |
1292 | else if (!ANDISVal) |
1293 | TotalVal = ANDIVal; |
1294 | else |
1295 | TotalVal = SDValue(CurDAG->getMachineNode(PPC::OR, dl, MVT::i32, |
1296 | ANDIVal, ANDISVal), 0); |
1297 | |
1298 | if (!Res) |
1299 | Res = TotalVal; |
1300 | else |
1301 | Res = SDValue(CurDAG->getMachineNode(PPC::OR, dl, MVT::i32, |
1302 | Res, TotalVal), 0); |
1303 | |
1304 | // Now, remove all groups with this underlying value and rotation |
1305 | // factor. |
1306 | eraseMatchingBitGroups([VRI](const BitGroup &BG) { |
1307 | return BG.V == VRI.V && BG.RLAmt == VRI.RLAmt; |
1308 | }); |
1309 | } |
1310 | } |
1311 | |
1312 | // Instruction selection for the 32-bit case. |
1313 | SDNode *Select32(SDNode *N, bool LateMask, unsigned *InstCnt) { |
1314 | SDLoc dl(N); |
1315 | SDValue Res; |
1316 | |
1317 | if (InstCnt) *InstCnt = 0; |
1318 | |
1319 | // Take care of cases that should use andi/andis first. |
1320 | SelectAndParts32(dl, Res, InstCnt); |
1321 | |
1322 | // If we've not yet selected a 'starting' instruction, and we have no zeros |
1323 | // to fill in, select the (Value, RLAmt) with the highest priority (largest |
1324 | // number of groups), and start with this rotated value. |
1325 | if ((!HasZeros || LateMask) && !Res) { |
1326 | ValueRotInfo &VRI = ValueRotsVec[0]; |
1327 | if (VRI.RLAmt) { |
1328 | if (InstCnt) *InstCnt += 1; |
1329 | SDValue Ops[] = |
1330 | { VRI.V, getI32Imm(VRI.RLAmt, dl), getI32Imm(0, dl), |
1331 | getI32Imm(31, dl) }; |
1332 | Res = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), |
1333 | 0); |
1334 | } else { |
1335 | Res = VRI.V; |
1336 | } |
1337 | |
1338 | // Now, remove all groups with this underlying value and rotation factor. |
1339 | eraseMatchingBitGroups([VRI](const BitGroup &BG) { |
1340 | return BG.V == VRI.V && BG.RLAmt == VRI.RLAmt; |
1341 | }); |
1342 | } |
1343 | |
1344 | if (InstCnt) *InstCnt += BitGroups.size(); |
1345 | |
1346 | // Insert the other groups (one at a time). |
1347 | for (auto &BG : BitGroups) { |
1348 | if (!Res) { |
1349 | SDValue Ops[] = |
1350 | { BG.V, getI32Imm(BG.RLAmt, dl), |
1351 | getI32Imm(Bits.size() - BG.EndIdx - 1, dl), |
1352 | getI32Imm(Bits.size() - BG.StartIdx - 1, dl) }; |
1353 | Res = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 0); |
1354 | } else { |
1355 | SDValue Ops[] = |
1356 | { Res, BG.V, getI32Imm(BG.RLAmt, dl), |
1357 | getI32Imm(Bits.size() - BG.EndIdx - 1, dl), |
1358 | getI32Imm(Bits.size() - BG.StartIdx - 1, dl) }; |
1359 | Res = SDValue(CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops), 0); |
1360 | } |
1361 | } |
1362 | |
1363 | if (LateMask) { |
1364 | unsigned Mask = (unsigned) getZerosMask(); |
1365 | |
1366 | unsigned ANDIMask = (Mask & UINT16_MAX(65535)), ANDISMask = Mask >> 16; |
1367 | assert((ANDIMask != 0 || ANDISMask != 0) &&(((ANDIMask != 0 || ANDISMask != 0) && "No set bits in zeros mask?" ) ? static_cast<void> (0) : __assert_fail ("(ANDIMask != 0 || ANDISMask != 0) && \"No set bits in zeros mask?\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 1368, __PRETTY_FUNCTION__)) |
1368 | "No set bits in zeros mask?")(((ANDIMask != 0 || ANDISMask != 0) && "No set bits in zeros mask?" ) ? static_cast<void> (0) : __assert_fail ("(ANDIMask != 0 || ANDISMask != 0) && \"No set bits in zeros mask?\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 1368, __PRETTY_FUNCTION__)); |
1369 | |
1370 | if (InstCnt) *InstCnt += (unsigned) (ANDIMask != 0) + |
1371 | (unsigned) (ANDISMask != 0) + |
1372 | (unsigned) (ANDIMask != 0 && ANDISMask != 0); |
1373 | |
1374 | SDValue ANDIVal, ANDISVal; |
1375 | if (ANDIMask != 0) |
1376 | ANDIVal = SDValue(CurDAG->getMachineNode(PPC::ANDIo, dl, MVT::i32, |
1377 | Res, getI32Imm(ANDIMask, dl)), 0); |
1378 | if (ANDISMask != 0) |
1379 | ANDISVal = SDValue(CurDAG->getMachineNode(PPC::ANDISo, dl, MVT::i32, |
1380 | Res, getI32Imm(ANDISMask, dl)), 0); |
1381 | |
1382 | if (!ANDIVal) |
1383 | Res = ANDISVal; |
1384 | else if (!ANDISVal) |
1385 | Res = ANDIVal; |
1386 | else |
1387 | Res = SDValue(CurDAG->getMachineNode(PPC::OR, dl, MVT::i32, |
1388 | ANDIVal, ANDISVal), 0); |
1389 | } |
1390 | |
1391 | return Res.getNode(); |
1392 | } |
1393 | |
1394 | unsigned SelectRotMask64Count(unsigned RLAmt, bool Repl32, |
1395 | unsigned MaskStart, unsigned MaskEnd, |
1396 | bool IsIns) { |
1397 | // In the notation used by the instructions, 'start' and 'end' are reversed |
1398 | // because bits are counted from high to low order. |
1399 | unsigned InstMaskStart = 64 - MaskEnd - 1, |
1400 | InstMaskEnd = 64 - MaskStart - 1; |
1401 | |
1402 | if (Repl32) |
1403 | return 1; |
1404 | |
1405 | if ((!IsIns && (InstMaskEnd == 63 || InstMaskStart == 0)) || |
1406 | InstMaskEnd == 63 - RLAmt) |
1407 | return 1; |
1408 | |
1409 | return 2; |
1410 | } |
1411 | |
1412 | // For 64-bit values, not all combinations of rotates and masks are |
1413 | // available. Produce one if it is available. |
1414 | SDValue SelectRotMask64(SDValue V, SDLoc dl, unsigned RLAmt, bool Repl32, |
1415 | unsigned MaskStart, unsigned MaskEnd, |
1416 | unsigned *InstCnt = nullptr) { |
1417 | // In the notation used by the instructions, 'start' and 'end' are reversed |
1418 | // because bits are counted from high to low order. |
1419 | unsigned InstMaskStart = 64 - MaskEnd - 1, |
1420 | InstMaskEnd = 64 - MaskStart - 1; |
1421 | |
1422 | if (InstCnt) *InstCnt += 1; |
1423 | |
1424 | if (Repl32) { |
1425 | // This rotation amount assumes that the lower 32 bits of the quantity |
1426 | // are replicated in the high 32 bits by the rotation operator (which is |
1427 | // done by rlwinm and friends). |
1428 | assert(InstMaskStart >= 32 && "Mask cannot start out of range")((InstMaskStart >= 32 && "Mask cannot start out of range" ) ? static_cast<void> (0) : __assert_fail ("InstMaskStart >= 32 && \"Mask cannot start out of range\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 1428, __PRETTY_FUNCTION__)); |
1429 | assert(InstMaskEnd >= 32 && "Mask cannot end out of range")((InstMaskEnd >= 32 && "Mask cannot end out of range" ) ? static_cast<void> (0) : __assert_fail ("InstMaskEnd >= 32 && \"Mask cannot end out of range\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 1429, __PRETTY_FUNCTION__)); |
1430 | SDValue Ops[] = |
1431 | { V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskStart - 32, dl), |
1432 | getI32Imm(InstMaskEnd - 32, dl) }; |
1433 | return SDValue(CurDAG->getMachineNode(PPC::RLWINM8, dl, MVT::i64, |
1434 | Ops), 0); |
1435 | } |
1436 | |
1437 | if (InstMaskEnd == 63) { |
1438 | SDValue Ops[] = |
1439 | { V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskStart, dl) }; |
1440 | return SDValue(CurDAG->getMachineNode(PPC::RLDICL, dl, MVT::i64, Ops), 0); |
1441 | } |
1442 | |
1443 | if (InstMaskStart == 0) { |
1444 | SDValue Ops[] = |
1445 | { V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskEnd, dl) }; |
1446 | return SDValue(CurDAG->getMachineNode(PPC::RLDICR, dl, MVT::i64, Ops), 0); |
1447 | } |
1448 | |
1449 | if (InstMaskEnd == 63 - RLAmt) { |
1450 | SDValue Ops[] = |
1451 | { V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskStart, dl) }; |
1452 | return SDValue(CurDAG->getMachineNode(PPC::RLDIC, dl, MVT::i64, Ops), 0); |
1453 | } |
1454 | |
1455 | // We cannot do this with a single instruction, so we'll use two. The |
1456 | // problem is that we're not free to choose both a rotation amount and mask |
1457 | // start and end independently. We can choose an arbitrary mask start and |
1458 | // end, but then the rotation amount is fixed. Rotation, however, can be |
1459 | // inverted, and so by applying an "inverse" rotation first, we can get the |
1460 | // desired result. |
1461 | if (InstCnt) *InstCnt += 1; |
1462 | |
1463 | // The rotation mask for the second instruction must be MaskStart. |
1464 | unsigned RLAmt2 = MaskStart; |
1465 | // The first instruction must rotate V so that the overall rotation amount |
1466 | // is RLAmt. |
1467 | unsigned RLAmt1 = (64 + RLAmt - RLAmt2) % 64; |
1468 | if (RLAmt1) |
1469 | V = SelectRotMask64(V, dl, RLAmt1, false, 0, 63); |
1470 | return SelectRotMask64(V, dl, RLAmt2, false, MaskStart, MaskEnd); |
1471 | } |
1472 | |
1473 | // For 64-bit values, not all combinations of rotates and masks are |
1474 | // available. Produce a rotate-mask-and-insert if one is available. |
1475 | SDValue SelectRotMaskIns64(SDValue Base, SDValue V, SDLoc dl, unsigned RLAmt, |
1476 | bool Repl32, unsigned MaskStart, |
1477 | unsigned MaskEnd, unsigned *InstCnt = nullptr) { |
1478 | // In the notation used by the instructions, 'start' and 'end' are reversed |
1479 | // because bits are counted from high to low order. |
1480 | unsigned InstMaskStart = 64 - MaskEnd - 1, |
1481 | InstMaskEnd = 64 - MaskStart - 1; |
1482 | |
1483 | if (InstCnt) *InstCnt += 1; |
1484 | |
1485 | if (Repl32) { |
1486 | // This rotation amount assumes that the lower 32 bits of the quantity |
1487 | // are replicated in the high 32 bits by the rotation operator (which is |
1488 | // done by rlwinm and friends). |
1489 | assert(InstMaskStart >= 32 && "Mask cannot start out of range")((InstMaskStart >= 32 && "Mask cannot start out of range" ) ? static_cast<void> (0) : __assert_fail ("InstMaskStart >= 32 && \"Mask cannot start out of range\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 1489, __PRETTY_FUNCTION__)); |
1490 | assert(InstMaskEnd >= 32 && "Mask cannot end out of range")((InstMaskEnd >= 32 && "Mask cannot end out of range" ) ? static_cast<void> (0) : __assert_fail ("InstMaskEnd >= 32 && \"Mask cannot end out of range\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 1490, __PRETTY_FUNCTION__)); |
1491 | SDValue Ops[] = |
1492 | { Base, V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskStart - 32, dl), |
1493 | getI32Imm(InstMaskEnd - 32, dl) }; |
1494 | return SDValue(CurDAG->getMachineNode(PPC::RLWIMI8, dl, MVT::i64, |
1495 | Ops), 0); |
1496 | } |
1497 | |
1498 | if (InstMaskEnd == 63 - RLAmt) { |
1499 | SDValue Ops[] = |
1500 | { Base, V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskStart, dl) }; |
1501 | return SDValue(CurDAG->getMachineNode(PPC::RLDIMI, dl, MVT::i64, Ops), 0); |
1502 | } |
1503 | |
1504 | // We cannot do this with a single instruction, so we'll use two. The |
1505 | // problem is that we're not free to choose both a rotation amount and mask |
1506 | // start and end independently. We can choose an arbitrary mask start and |
1507 | // end, but then the rotation amount is fixed. Rotation, however, can be |
1508 | // inverted, and so by applying an "inverse" rotation first, we can get the |
1509 | // desired result. |
1510 | if (InstCnt) *InstCnt += 1; |
1511 | |
1512 | // The rotation mask for the second instruction must be MaskStart. |
1513 | unsigned RLAmt2 = MaskStart; |
1514 | // The first instruction must rotate V so that the overall rotation amount |
1515 | // is RLAmt. |
1516 | unsigned RLAmt1 = (64 + RLAmt - RLAmt2) % 64; |
1517 | if (RLAmt1) |
1518 | V = SelectRotMask64(V, dl, RLAmt1, false, 0, 63); |
1519 | return SelectRotMaskIns64(Base, V, dl, RLAmt2, false, MaskStart, MaskEnd); |
1520 | } |
1521 | |
1522 | void SelectAndParts64(SDLoc dl, SDValue &Res, unsigned *InstCnt) { |
1523 | if (BPermRewriterNoMasking) |
1524 | return; |
1525 | |
1526 | // The idea here is the same as in the 32-bit version, but with additional |
1527 | // complications from the fact that Repl32 might be true. Because we |
1528 | // aggressively convert bit groups to Repl32 form (which, for small |
1529 | // rotation factors, involves no other change), and then coalesce, it might |
1530 | // be the case that a single 64-bit masking operation could handle both |
1531 | // some Repl32 groups and some non-Repl32 groups. If converting to Repl32 |
1532 | // form allowed coalescing, then we must use a 32-bit rotaton in order to |
1533 | // completely capture the new combined bit group. |
1534 | |
1535 | for (ValueRotInfo &VRI : ValueRotsVec) { |
1536 | uint64_t Mask = 0; |
1537 | |
1538 | // We need to add to the mask all bits from the associated bit groups. |
1539 | // If Repl32 is false, we need to add bits from bit groups that have |
1540 | // Repl32 true, but are trivially convertable to Repl32 false. Such a |
1541 | // group is trivially convertable if it overlaps only with the lower 32 |
1542 | // bits, and the group has not been coalesced. |
1543 | auto MatchingBG = [VRI](const BitGroup &BG) { |
1544 | if (VRI.V != BG.V) |
1545 | return false; |
1546 | |
1547 | unsigned EffRLAmt = BG.RLAmt; |
1548 | if (!VRI.Repl32 && BG.Repl32) { |
1549 | if (BG.StartIdx < 32 && BG.EndIdx < 32 && BG.StartIdx <= BG.EndIdx && |
1550 | !BG.Repl32Coalesced) { |
1551 | if (BG.Repl32CR) |
1552 | EffRLAmt += 32; |
1553 | } else { |
1554 | return false; |
1555 | } |
1556 | } else if (VRI.Repl32 != BG.Repl32) { |
1557 | return false; |
1558 | } |
1559 | |
1560 | if (VRI.RLAmt != EffRLAmt) |
1561 | return false; |
1562 | |
1563 | return true; |
1564 | }; |
1565 | |
1566 | for (auto &BG : BitGroups) { |
1567 | if (!MatchingBG(BG)) |
1568 | continue; |
1569 | |
1570 | if (BG.StartIdx <= BG.EndIdx) { |
1571 | for (unsigned i = BG.StartIdx; i <= BG.EndIdx; ++i) |
1572 | Mask |= (UINT64_C(1)1UL << i); |
1573 | } else { |
1574 | for (unsigned i = BG.StartIdx; i < Bits.size(); ++i) |
1575 | Mask |= (UINT64_C(1)1UL << i); |
1576 | for (unsigned i = 0; i <= BG.EndIdx; ++i) |
1577 | Mask |= (UINT64_C(1)1UL << i); |
1578 | } |
1579 | } |
1580 | |
1581 | // We can use the 32-bit andi/andis technique if the mask does not |
1582 | // require any higher-order bits. This can save an instruction compared |
1583 | // to always using the general 64-bit technique. |
1584 | bool Use32BitInsts = isUInt<32>(Mask); |
1585 | // Compute the masks for andi/andis that would be necessary. |
1586 | unsigned ANDIMask = (Mask & UINT16_MAX(65535)), |
1587 | ANDISMask = (Mask >> 16) & UINT16_MAX(65535); |
1588 | |
1589 | bool NeedsRotate = VRI.RLAmt || (VRI.Repl32 && !isUInt<32>(Mask)); |
1590 | |
1591 | unsigned NumAndInsts = (unsigned) NeedsRotate + |
1592 | (unsigned) (bool) Res; |
1593 | if (Use32BitInsts) |
1594 | NumAndInsts += (unsigned) (ANDIMask != 0) + (unsigned) (ANDISMask != 0) + |
1595 | (unsigned) (ANDIMask != 0 && ANDISMask != 0); |
1596 | else |
1597 | NumAndInsts += SelectInt64Count(Mask) + /* and */ 1; |
1598 | |
1599 | unsigned NumRLInsts = 0; |
1600 | bool FirstBG = true; |
1601 | for (auto &BG : BitGroups) { |
1602 | if (!MatchingBG(BG)) |
1603 | continue; |
1604 | NumRLInsts += |
1605 | SelectRotMask64Count(BG.RLAmt, BG.Repl32, BG.StartIdx, BG.EndIdx, |
1606 | !FirstBG); |
1607 | FirstBG = false; |
1608 | } |
1609 | |
1610 | DEBUG(dbgs() << "\t\trotation groups for " << VRI.V.getNode() <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\t\trotation groups for " << VRI.V.getNode() << " RL: " << VRI.RLAmt << (VRI.Repl32 ? " (32):" : ":") << "\n\t\t\tisel using masking: " << NumAndInsts << " using rotates: " << NumRLInsts << "\n"; } } while (0) |
1611 | " RL: " << VRI.RLAmt << (VRI.Repl32 ? " (32):" : ":") <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\t\trotation groups for " << VRI.V.getNode() << " RL: " << VRI.RLAmt << (VRI.Repl32 ? " (32):" : ":") << "\n\t\t\tisel using masking: " << NumAndInsts << " using rotates: " << NumRLInsts << "\n"; } } while (0) |
1612 | "\n\t\t\tisel using masking: " << NumAndInsts <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\t\trotation groups for " << VRI.V.getNode() << " RL: " << VRI.RLAmt << (VRI.Repl32 ? " (32):" : ":") << "\n\t\t\tisel using masking: " << NumAndInsts << " using rotates: " << NumRLInsts << "\n"; } } while (0) |
1613 | " using rotates: " << NumRLInsts << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\t\trotation groups for " << VRI.V.getNode() << " RL: " << VRI.RLAmt << (VRI.Repl32 ? " (32):" : ":") << "\n\t\t\tisel using masking: " << NumAndInsts << " using rotates: " << NumRLInsts << "\n"; } } while (0); |
1614 | |
1615 | // When we'd use andi/andis, we bias toward using the rotates (andi only |
1616 | // has a record form, and is cracked on POWER cores). However, when using |
1617 | // general 64-bit constant formation, bias toward the constant form, |
1618 | // because that exposes more opportunities for CSE. |
1619 | if (NumAndInsts > NumRLInsts) |
1620 | continue; |
1621 | if (Use32BitInsts && NumAndInsts == NumRLInsts) |
1622 | continue; |
1623 | |
1624 | DEBUG(dbgs() << "\t\t\t\tusing masking\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\t\t\t\tusing masking\n"; } } while (0); |
1625 | |
1626 | if (InstCnt) *InstCnt += NumAndInsts; |
1627 | |
1628 | SDValue VRot; |
1629 | // We actually need to generate a rotation if we have a non-zero rotation |
1630 | // factor or, in the Repl32 case, if we care about any of the |
1631 | // higher-order replicated bits. In the latter case, we generate a mask |
1632 | // backward so that it actually includes the entire 64 bits. |
1633 | if (VRI.RLAmt || (VRI.Repl32 && !isUInt<32>(Mask))) |
1634 | VRot = SelectRotMask64(VRI.V, dl, VRI.RLAmt, VRI.Repl32, |
1635 | VRI.Repl32 ? 31 : 0, VRI.Repl32 ? 30 : 63); |
1636 | else |
1637 | VRot = VRI.V; |
1638 | |
1639 | SDValue TotalVal; |
1640 | if (Use32BitInsts) { |
1641 | assert((ANDIMask != 0 || ANDISMask != 0) &&(((ANDIMask != 0 || ANDISMask != 0) && "No set bits in mask when using 32-bit ands for 64-bit value" ) ? static_cast<void> (0) : __assert_fail ("(ANDIMask != 0 || ANDISMask != 0) && \"No set bits in mask when using 32-bit ands for 64-bit value\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 1642, __PRETTY_FUNCTION__)) |
1642 | "No set bits in mask when using 32-bit ands for 64-bit value")(((ANDIMask != 0 || ANDISMask != 0) && "No set bits in mask when using 32-bit ands for 64-bit value" ) ? static_cast<void> (0) : __assert_fail ("(ANDIMask != 0 || ANDISMask != 0) && \"No set bits in mask when using 32-bit ands for 64-bit value\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 1642, __PRETTY_FUNCTION__)); |
1643 | |
1644 | SDValue ANDIVal, ANDISVal; |
1645 | if (ANDIMask != 0) |
1646 | ANDIVal = SDValue(CurDAG->getMachineNode(PPC::ANDIo8, dl, MVT::i64, |
1647 | VRot, getI32Imm(ANDIMask, dl)), 0); |
1648 | if (ANDISMask != 0) |
1649 | ANDISVal = SDValue(CurDAG->getMachineNode(PPC::ANDISo8, dl, MVT::i64, |
1650 | VRot, getI32Imm(ANDISMask, dl)), 0); |
1651 | |
1652 | if (!ANDIVal) |
1653 | TotalVal = ANDISVal; |
1654 | else if (!ANDISVal) |
1655 | TotalVal = ANDIVal; |
1656 | else |
1657 | TotalVal = SDValue(CurDAG->getMachineNode(PPC::OR8, dl, MVT::i64, |
1658 | ANDIVal, ANDISVal), 0); |
1659 | } else { |
1660 | TotalVal = SDValue(SelectInt64(CurDAG, dl, Mask), 0); |
1661 | TotalVal = |
1662 | SDValue(CurDAG->getMachineNode(PPC::AND8, dl, MVT::i64, |
1663 | VRot, TotalVal), 0); |
1664 | } |
1665 | |
1666 | if (!Res) |
1667 | Res = TotalVal; |
1668 | else |
1669 | Res = SDValue(CurDAG->getMachineNode(PPC::OR8, dl, MVT::i64, |
1670 | Res, TotalVal), 0); |
1671 | |
1672 | // Now, remove all groups with this underlying value and rotation |
1673 | // factor. |
1674 | eraseMatchingBitGroups(MatchingBG); |
1675 | } |
1676 | } |
1677 | |
1678 | // Instruction selection for the 64-bit case. |
1679 | SDNode *Select64(SDNode *N, bool LateMask, unsigned *InstCnt) { |
1680 | SDLoc dl(N); |
1681 | SDValue Res; |
1682 | |
1683 | if (InstCnt) *InstCnt = 0; |
1684 | |
1685 | // Take care of cases that should use andi/andis first. |
1686 | SelectAndParts64(dl, Res, InstCnt); |
1687 | |
1688 | // If we've not yet selected a 'starting' instruction, and we have no zeros |
1689 | // to fill in, select the (Value, RLAmt) with the highest priority (largest |
1690 | // number of groups), and start with this rotated value. |
1691 | if ((!HasZeros || LateMask) && !Res) { |
1692 | // If we have both Repl32 groups and non-Repl32 groups, the non-Repl32 |
1693 | // groups will come first, and so the VRI representing the largest number |
1694 | // of groups might not be first (it might be the first Repl32 groups). |
1695 | unsigned MaxGroupsIdx = 0; |
1696 | if (!ValueRotsVec[0].Repl32) { |
1697 | for (unsigned i = 0, ie = ValueRotsVec.size(); i < ie; ++i) |
1698 | if (ValueRotsVec[i].Repl32) { |
1699 | if (ValueRotsVec[i].NumGroups > ValueRotsVec[0].NumGroups) |
1700 | MaxGroupsIdx = i; |
1701 | break; |
1702 | } |
1703 | } |
1704 | |
1705 | ValueRotInfo &VRI = ValueRotsVec[MaxGroupsIdx]; |
1706 | bool NeedsRotate = false; |
1707 | if (VRI.RLAmt) { |
1708 | NeedsRotate = true; |
1709 | } else if (VRI.Repl32) { |
1710 | for (auto &BG : BitGroups) { |
1711 | if (BG.V != VRI.V || BG.RLAmt != VRI.RLAmt || |
1712 | BG.Repl32 != VRI.Repl32) |
1713 | continue; |
1714 | |
1715 | // We don't need a rotate if the bit group is confined to the lower |
1716 | // 32 bits. |
1717 | if (BG.StartIdx < 32 && BG.EndIdx < 32 && BG.StartIdx < BG.EndIdx) |
1718 | continue; |
1719 | |
1720 | NeedsRotate = true; |
1721 | break; |
1722 | } |
1723 | } |
1724 | |
1725 | if (NeedsRotate) |
1726 | Res = SelectRotMask64(VRI.V, dl, VRI.RLAmt, VRI.Repl32, |
1727 | VRI.Repl32 ? 31 : 0, VRI.Repl32 ? 30 : 63, |
1728 | InstCnt); |
1729 | else |
1730 | Res = VRI.V; |
1731 | |
1732 | // Now, remove all groups with this underlying value and rotation factor. |
1733 | if (Res) |
1734 | eraseMatchingBitGroups([VRI](const BitGroup &BG) { |
1735 | return BG.V == VRI.V && BG.RLAmt == VRI.RLAmt && |
1736 | BG.Repl32 == VRI.Repl32; |
1737 | }); |
1738 | } |
1739 | |
1740 | // Because 64-bit rotates are more flexible than inserts, we might have a |
1741 | // preference regarding which one we do first (to save one instruction). |
1742 | if (!Res) |
1743 | for (auto I = BitGroups.begin(), IE = BitGroups.end(); I != IE; ++I) { |
1744 | if (SelectRotMask64Count(I->RLAmt, I->Repl32, I->StartIdx, I->EndIdx, |
1745 | false) < |
1746 | SelectRotMask64Count(I->RLAmt, I->Repl32, I->StartIdx, I->EndIdx, |
1747 | true)) { |
1748 | if (I != BitGroups.begin()) { |
1749 | BitGroup BG = *I; |
1750 | BitGroups.erase(I); |
1751 | BitGroups.insert(BitGroups.begin(), BG); |
1752 | } |
1753 | |
1754 | break; |
1755 | } |
1756 | } |
1757 | |
1758 | // Insert the other groups (one at a time). |
1759 | for (auto &BG : BitGroups) { |
1760 | if (!Res) |
1761 | Res = SelectRotMask64(BG.V, dl, BG.RLAmt, BG.Repl32, BG.StartIdx, |
1762 | BG.EndIdx, InstCnt); |
1763 | else |
1764 | Res = SelectRotMaskIns64(Res, BG.V, dl, BG.RLAmt, BG.Repl32, |
1765 | BG.StartIdx, BG.EndIdx, InstCnt); |
1766 | } |
1767 | |
1768 | if (LateMask) { |
1769 | uint64_t Mask = getZerosMask(); |
1770 | |
1771 | // We can use the 32-bit andi/andis technique if the mask does not |
1772 | // require any higher-order bits. This can save an instruction compared |
1773 | // to always using the general 64-bit technique. |
1774 | bool Use32BitInsts = isUInt<32>(Mask); |
1775 | // Compute the masks for andi/andis that would be necessary. |
1776 | unsigned ANDIMask = (Mask & UINT16_MAX(65535)), |
1777 | ANDISMask = (Mask >> 16) & UINT16_MAX(65535); |
1778 | |
1779 | if (Use32BitInsts) { |
1780 | assert((ANDIMask != 0 || ANDISMask != 0) &&(((ANDIMask != 0 || ANDISMask != 0) && "No set bits in mask when using 32-bit ands for 64-bit value" ) ? static_cast<void> (0) : __assert_fail ("(ANDIMask != 0 || ANDISMask != 0) && \"No set bits in mask when using 32-bit ands for 64-bit value\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 1781, __PRETTY_FUNCTION__)) |
1781 | "No set bits in mask when using 32-bit ands for 64-bit value")(((ANDIMask != 0 || ANDISMask != 0) && "No set bits in mask when using 32-bit ands for 64-bit value" ) ? static_cast<void> (0) : __assert_fail ("(ANDIMask != 0 || ANDISMask != 0) && \"No set bits in mask when using 32-bit ands for 64-bit value\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 1781, __PRETTY_FUNCTION__)); |
1782 | |
1783 | if (InstCnt) *InstCnt += (unsigned) (ANDIMask != 0) + |
1784 | (unsigned) (ANDISMask != 0) + |
1785 | (unsigned) (ANDIMask != 0 && ANDISMask != 0); |
1786 | |
1787 | SDValue ANDIVal, ANDISVal; |
1788 | if (ANDIMask != 0) |
1789 | ANDIVal = SDValue(CurDAG->getMachineNode(PPC::ANDIo8, dl, MVT::i64, |
1790 | Res, getI32Imm(ANDIMask, dl)), 0); |
1791 | if (ANDISMask != 0) |
1792 | ANDISVal = SDValue(CurDAG->getMachineNode(PPC::ANDISo8, dl, MVT::i64, |
1793 | Res, getI32Imm(ANDISMask, dl)), 0); |
1794 | |
1795 | if (!ANDIVal) |
1796 | Res = ANDISVal; |
1797 | else if (!ANDISVal) |
1798 | Res = ANDIVal; |
1799 | else |
1800 | Res = SDValue(CurDAG->getMachineNode(PPC::OR8, dl, MVT::i64, |
1801 | ANDIVal, ANDISVal), 0); |
1802 | } else { |
1803 | if (InstCnt) *InstCnt += SelectInt64Count(Mask) + /* and */ 1; |
1804 | |
1805 | SDValue MaskVal = SDValue(SelectInt64(CurDAG, dl, Mask), 0); |
1806 | Res = |
1807 | SDValue(CurDAG->getMachineNode(PPC::AND8, dl, MVT::i64, |
1808 | Res, MaskVal), 0); |
1809 | } |
1810 | } |
1811 | |
1812 | return Res.getNode(); |
1813 | } |
1814 | |
1815 | SDNode *Select(SDNode *N, bool LateMask, unsigned *InstCnt = nullptr) { |
1816 | // Fill in BitGroups. |
1817 | collectBitGroups(LateMask); |
1818 | if (BitGroups.empty()) |
1819 | return nullptr; |
1820 | |
1821 | // For 64-bit values, figure out when we can use 32-bit instructions. |
1822 | if (Bits.size() == 64) |
1823 | assignRepl32BitGroups(); |
1824 | |
1825 | // Fill in ValueRotsVec. |
1826 | collectValueRotInfo(); |
1827 | |
1828 | if (Bits.size() == 32) { |
1829 | return Select32(N, LateMask, InstCnt); |
1830 | } else { |
1831 | assert(Bits.size() == 64 && "Not 64 bits here?")((Bits.size() == 64 && "Not 64 bits here?") ? static_cast <void> (0) : __assert_fail ("Bits.size() == 64 && \"Not 64 bits here?\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 1831, __PRETTY_FUNCTION__)); |
1832 | return Select64(N, LateMask, InstCnt); |
1833 | } |
1834 | |
1835 | return nullptr; |
1836 | } |
1837 | |
1838 | void eraseMatchingBitGroups(function_ref<bool(const BitGroup &)> F) { |
1839 | BitGroups.erase(std::remove_if(BitGroups.begin(), BitGroups.end(), F), |
1840 | BitGroups.end()); |
1841 | } |
1842 | |
1843 | SmallVector<ValueBit, 64> Bits; |
1844 | |
1845 | bool HasZeros; |
1846 | SmallVector<unsigned, 64> RLAmt; |
1847 | |
1848 | SmallVector<BitGroup, 16> BitGroups; |
1849 | |
1850 | DenseMap<std::pair<SDValue, unsigned>, ValueRotInfo> ValueRots; |
1851 | SmallVector<ValueRotInfo, 16> ValueRotsVec; |
1852 | |
1853 | SelectionDAG *CurDAG; |
1854 | |
1855 | public: |
1856 | BitPermutationSelector(SelectionDAG *DAG) |
1857 | : CurDAG(DAG) {} |
1858 | |
1859 | // Here we try to match complex bit permutations into a set of |
1860 | // rotate-and-shift/shift/and/or instructions, using a set of heuristics |
1861 | // known to produce optimial code for common cases (like i32 byte swapping). |
1862 | SDNode *Select(SDNode *N) { |
1863 | Bits.resize(N->getValueType(0).getSizeInBits()); |
1864 | if (!getValueBits(SDValue(N, 0), Bits)) |
1865 | return nullptr; |
1866 | |
1867 | DEBUG(dbgs() << "Considering bit-permutation-based instruction"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "Considering bit-permutation-based instruction" " selection for: "; } } while (0) |
1868 | " selection for: ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "Considering bit-permutation-based instruction" " selection for: "; } } while (0); |
1869 | DEBUG(N->dump(CurDAG))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { N->dump(CurDAG); } } while (0); |
1870 | |
1871 | // Fill it RLAmt and set HasZeros. |
1872 | computeRotationAmounts(); |
1873 | |
1874 | if (!HasZeros) |
1875 | return Select(N, false); |
1876 | |
1877 | // We currently have two techniques for handling results with zeros: early |
1878 | // masking (the default) and late masking. Late masking is sometimes more |
1879 | // efficient, but because the structure of the bit groups is different, it |
1880 | // is hard to tell without generating both and comparing the results. With |
1881 | // late masking, we ignore zeros in the resulting value when inserting each |
1882 | // set of bit groups, and then mask in the zeros at the end. With early |
1883 | // masking, we only insert the non-zero parts of the result at every step. |
1884 | |
1885 | unsigned InstCnt, InstCntLateMask; |
1886 | DEBUG(dbgs() << "\tEarly masking:\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\tEarly masking:\n"; } } while (0); |
1887 | SDNode *RN = Select(N, false, &InstCnt); |
1888 | DEBUG(dbgs() << "\t\tisel would use " << InstCnt << " instructions\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\t\tisel would use " << InstCnt << " instructions\n"; } } while (0); |
1889 | |
1890 | DEBUG(dbgs() << "\tLate masking:\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\tLate masking:\n"; } } while (0); |
1891 | SDNode *RNLM = Select(N, true, &InstCntLateMask); |
1892 | DEBUG(dbgs() << "\t\tisel would use " << InstCntLateMask <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\t\tisel would use " << InstCntLateMask << " instructions\n"; } } while (0) |
1893 | " instructions\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\t\tisel would use " << InstCntLateMask << " instructions\n"; } } while (0); |
1894 | |
1895 | if (InstCnt <= InstCntLateMask) { |
1896 | DEBUG(dbgs() << "\tUsing early-masking for isel\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\tUsing early-masking for isel\n" ; } } while (0); |
1897 | return RN; |
1898 | } |
1899 | |
1900 | DEBUG(dbgs() << "\tUsing late-masking for isel\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\tUsing late-masking for isel\n" ; } } while (0); |
1901 | return RNLM; |
1902 | } |
1903 | }; |
1904 | } // anonymous namespace |
1905 | |
1906 | SDNode *PPCDAGToDAGISel::SelectBitPermutation(SDNode *N) { |
1907 | if (N->getValueType(0) != MVT::i32 && |
1908 | N->getValueType(0) != MVT::i64) |
1909 | return nullptr; |
1910 | |
1911 | if (!UseBitPermRewriter) |
1912 | return nullptr; |
1913 | |
1914 | switch (N->getOpcode()) { |
1915 | default: break; |
1916 | case ISD::ROTL: |
1917 | case ISD::SHL: |
1918 | case ISD::SRL: |
1919 | case ISD::AND: |
1920 | case ISD::OR: { |
1921 | BitPermutationSelector BPS(CurDAG); |
1922 | return BPS.Select(N); |
1923 | } |
1924 | } |
1925 | |
1926 | return nullptr; |
1927 | } |
1928 | |
1929 | /// SelectCC - Select a comparison of the specified values with the specified |
1930 | /// condition code, returning the CR# of the expression. |
1931 | SDValue PPCDAGToDAGISel::SelectCC(SDValue LHS, SDValue RHS, |
1932 | ISD::CondCode CC, SDLoc dl) { |
1933 | // Always select the LHS. |
1934 | unsigned Opc; |
1935 | |
1936 | if (LHS.getValueType() == MVT::i32) { |
1937 | unsigned Imm; |
1938 | if (CC == ISD::SETEQ || CC == ISD::SETNE) { |
1939 | if (isInt32Immediate(RHS, Imm)) { |
1940 | // SETEQ/SETNE comparison with 16-bit immediate, fold it. |
1941 | if (isUInt<16>(Imm)) |
1942 | return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, LHS, |
1943 | getI32Imm(Imm & 0xFFFF, dl)), |
1944 | 0); |
1945 | // If this is a 16-bit signed immediate, fold it. |
1946 | if (isInt<16>((int)Imm)) |
1947 | return SDValue(CurDAG->getMachineNode(PPC::CMPWI, dl, MVT::i32, LHS, |
1948 | getI32Imm(Imm & 0xFFFF, dl)), |
1949 | 0); |
1950 | |
1951 | // For non-equality comparisons, the default code would materialize the |
1952 | // constant, then compare against it, like this: |
1953 | // lis r2, 4660 |
1954 | // ori r2, r2, 22136 |
1955 | // cmpw cr0, r3, r2 |
1956 | // Since we are just comparing for equality, we can emit this instead: |
1957 | // xoris r0,r3,0x1234 |
1958 | // cmplwi cr0,r0,0x5678 |
1959 | // beq cr0,L6 |
1960 | SDValue Xor(CurDAG->getMachineNode(PPC::XORIS, dl, MVT::i32, LHS, |
1961 | getI32Imm(Imm >> 16, dl)), 0); |
1962 | return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, Xor, |
1963 | getI32Imm(Imm & 0xFFFF, dl)), 0); |
1964 | } |
1965 | Opc = PPC::CMPLW; |
1966 | } else if (ISD::isUnsignedIntSetCC(CC)) { |
1967 | if (isInt32Immediate(RHS, Imm) && isUInt<16>(Imm)) |
1968 | return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, LHS, |
1969 | getI32Imm(Imm & 0xFFFF, dl)), 0); |
1970 | Opc = PPC::CMPLW; |
1971 | } else { |
1972 | short SImm; |
1973 | if (isIntS16Immediate(RHS, SImm)) |
1974 | return SDValue(CurDAG->getMachineNode(PPC::CMPWI, dl, MVT::i32, LHS, |
1975 | getI32Imm((int)SImm & 0xFFFF, |
1976 | dl)), |
1977 | 0); |
1978 | Opc = PPC::CMPW; |
1979 | } |
1980 | } else if (LHS.getValueType() == MVT::i64) { |
1981 | uint64_t Imm; |
1982 | if (CC == ISD::SETEQ || CC == ISD::SETNE) { |
1983 | if (isInt64Immediate(RHS.getNode(), Imm)) { |
1984 | // SETEQ/SETNE comparison with 16-bit immediate, fold it. |
1985 | if (isUInt<16>(Imm)) |
1986 | return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, LHS, |
1987 | getI32Imm(Imm & 0xFFFF, dl)), |
1988 | 0); |
1989 | // If this is a 16-bit signed immediate, fold it. |
1990 | if (isInt<16>(Imm)) |
1991 | return SDValue(CurDAG->getMachineNode(PPC::CMPDI, dl, MVT::i64, LHS, |
1992 | getI32Imm(Imm & 0xFFFF, dl)), |
1993 | 0); |
1994 | |
1995 | // For non-equality comparisons, the default code would materialize the |
1996 | // constant, then compare against it, like this: |
1997 | // lis r2, 4660 |
1998 | // ori r2, r2, 22136 |
1999 | // cmpd cr0, r3, r2 |
2000 | // Since we are just comparing for equality, we can emit this instead: |
2001 | // xoris r0,r3,0x1234 |
2002 | // cmpldi cr0,r0,0x5678 |
2003 | // beq cr0,L6 |
2004 | if (isUInt<32>(Imm)) { |
2005 | SDValue Xor(CurDAG->getMachineNode(PPC::XORIS8, dl, MVT::i64, LHS, |
2006 | getI64Imm(Imm >> 16, dl)), 0); |
2007 | return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, Xor, |
2008 | getI64Imm(Imm & 0xFFFF, dl)), |
2009 | 0); |
2010 | } |
2011 | } |
2012 | Opc = PPC::CMPLD; |
2013 | } else if (ISD::isUnsignedIntSetCC(CC)) { |
2014 | if (isInt64Immediate(RHS.getNode(), Imm) && isUInt<16>(Imm)) |
2015 | return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, LHS, |
2016 | getI64Imm(Imm & 0xFFFF, dl)), 0); |
2017 | Opc = PPC::CMPLD; |
2018 | } else { |
2019 | short SImm; |
2020 | if (isIntS16Immediate(RHS, SImm)) |
2021 | return SDValue(CurDAG->getMachineNode(PPC::CMPDI, dl, MVT::i64, LHS, |
2022 | getI64Imm(SImm & 0xFFFF, dl)), |
2023 | 0); |
2024 | Opc = PPC::CMPD; |
2025 | } |
2026 | } else if (LHS.getValueType() == MVT::f32) { |
2027 | Opc = PPC::FCMPUS; |
2028 | } else { |
2029 | assert(LHS.getValueType() == MVT::f64 && "Unknown vt!")((LHS.getValueType() == MVT::f64 && "Unknown vt!") ? static_cast <void> (0) : __assert_fail ("LHS.getValueType() == MVT::f64 && \"Unknown vt!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2029, __PRETTY_FUNCTION__)); |
2030 | Opc = PPCSubTarget->hasVSX() ? PPC::XSCMPUDP : PPC::FCMPUD; |
2031 | } |
2032 | return SDValue(CurDAG->getMachineNode(Opc, dl, MVT::i32, LHS, RHS), 0); |
2033 | } |
2034 | |
2035 | static PPC::Predicate getPredicateForSetCC(ISD::CondCode CC) { |
2036 | switch (CC) { |
2037 | case ISD::SETUEQ: |
2038 | case ISD::SETONE: |
2039 | case ISD::SETOLE: |
2040 | case ISD::SETOGE: |
2041 | llvm_unreachable("Should be lowered by legalize!")::llvm::llvm_unreachable_internal("Should be lowered by legalize!" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2041); |
2042 | default: llvm_unreachable("Unknown condition!")::llvm::llvm_unreachable_internal("Unknown condition!", "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2042); |
2043 | case ISD::SETOEQ: |
2044 | case ISD::SETEQ: return PPC::PRED_EQ; |
2045 | case ISD::SETUNE: |
2046 | case ISD::SETNE: return PPC::PRED_NE; |
2047 | case ISD::SETOLT: |
2048 | case ISD::SETLT: return PPC::PRED_LT; |
2049 | case ISD::SETULE: |
2050 | case ISD::SETLE: return PPC::PRED_LE; |
2051 | case ISD::SETOGT: |
2052 | case ISD::SETGT: return PPC::PRED_GT; |
2053 | case ISD::SETUGE: |
2054 | case ISD::SETGE: return PPC::PRED_GE; |
2055 | case ISD::SETO: return PPC::PRED_NU; |
2056 | case ISD::SETUO: return PPC::PRED_UN; |
2057 | // These two are invalid for floating point. Assume we have int. |
2058 | case ISD::SETULT: return PPC::PRED_LT; |
2059 | case ISD::SETUGT: return PPC::PRED_GT; |
2060 | } |
2061 | } |
2062 | |
2063 | /// getCRIdxForSetCC - Return the index of the condition register field |
2064 | /// associated with the SetCC condition, and whether or not the field is |
2065 | /// treated as inverted. That is, lt = 0; ge = 0 inverted. |
2066 | static unsigned getCRIdxForSetCC(ISD::CondCode CC, bool &Invert) { |
2067 | Invert = false; |
2068 | switch (CC) { |
2069 | default: llvm_unreachable("Unknown condition!")::llvm::llvm_unreachable_internal("Unknown condition!", "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2069); |
2070 | case ISD::SETOLT: |
2071 | case ISD::SETLT: return 0; // Bit #0 = SETOLT |
2072 | case ISD::SETOGT: |
2073 | case ISD::SETGT: return 1; // Bit #1 = SETOGT |
2074 | case ISD::SETOEQ: |
2075 | case ISD::SETEQ: return 2; // Bit #2 = SETOEQ |
2076 | case ISD::SETUO: return 3; // Bit #3 = SETUO |
2077 | case ISD::SETUGE: |
2078 | case ISD::SETGE: Invert = true; return 0; // !Bit #0 = SETUGE |
2079 | case ISD::SETULE: |
2080 | case ISD::SETLE: Invert = true; return 1; // !Bit #1 = SETULE |
2081 | case ISD::SETUNE: |
2082 | case ISD::SETNE: Invert = true; return 2; // !Bit #2 = SETUNE |
2083 | case ISD::SETO: Invert = true; return 3; // !Bit #3 = SETO |
2084 | case ISD::SETUEQ: |
2085 | case ISD::SETOGE: |
2086 | case ISD::SETOLE: |
2087 | case ISD::SETONE: |
2088 | llvm_unreachable("Invalid branch code: should be expanded by legalize")::llvm::llvm_unreachable_internal("Invalid branch code: should be expanded by legalize" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2088); |
2089 | // These are invalid for floating point. Assume integer. |
2090 | case ISD::SETULT: return 0; |
2091 | case ISD::SETUGT: return 1; |
2092 | } |
2093 | } |
2094 | |
2095 | // getVCmpInst: return the vector compare instruction for the specified |
2096 | // vector type and condition code. Since this is for altivec specific code, |
2097 | // only support the altivec types (v16i8, v8i16, v4i32, v2i64, and v4f32). |
2098 | static unsigned int getVCmpInst(MVT VecVT, ISD::CondCode CC, |
2099 | bool HasVSX, bool &Swap, bool &Negate) { |
2100 | Swap = false; |
2101 | Negate = false; |
2102 | |
2103 | if (VecVT.isFloatingPoint()) { |
2104 | /* Handle some cases by swapping input operands. */ |
2105 | switch (CC) { |
2106 | case ISD::SETLE: CC = ISD::SETGE; Swap = true; break; |
2107 | case ISD::SETLT: CC = ISD::SETGT; Swap = true; break; |
2108 | case ISD::SETOLE: CC = ISD::SETOGE; Swap = true; break; |
2109 | case ISD::SETOLT: CC = ISD::SETOGT; Swap = true; break; |
2110 | case ISD::SETUGE: CC = ISD::SETULE; Swap = true; break; |
2111 | case ISD::SETUGT: CC = ISD::SETULT; Swap = true; break; |
2112 | default: break; |
2113 | } |
2114 | /* Handle some cases by negating the result. */ |
2115 | switch (CC) { |
2116 | case ISD::SETNE: CC = ISD::SETEQ; Negate = true; break; |
2117 | case ISD::SETUNE: CC = ISD::SETOEQ; Negate = true; break; |
2118 | case ISD::SETULE: CC = ISD::SETOGT; Negate = true; break; |
2119 | case ISD::SETULT: CC = ISD::SETOGE; Negate = true; break; |
2120 | default: break; |
2121 | } |
2122 | /* We have instructions implementing the remaining cases. */ |
2123 | switch (CC) { |
2124 | case ISD::SETEQ: |
2125 | case ISD::SETOEQ: |
2126 | if (VecVT == MVT::v4f32) |
2127 | return HasVSX ? PPC::XVCMPEQSP : PPC::VCMPEQFP; |
2128 | else if (VecVT == MVT::v2f64) |
2129 | return PPC::XVCMPEQDP; |
2130 | break; |
2131 | case ISD::SETGT: |
2132 | case ISD::SETOGT: |
2133 | if (VecVT == MVT::v4f32) |
2134 | return HasVSX ? PPC::XVCMPGTSP : PPC::VCMPGTFP; |
2135 | else if (VecVT == MVT::v2f64) |
2136 | return PPC::XVCMPGTDP; |
2137 | break; |
2138 | case ISD::SETGE: |
2139 | case ISD::SETOGE: |
2140 | if (VecVT == MVT::v4f32) |
2141 | return HasVSX ? PPC::XVCMPGESP : PPC::VCMPGEFP; |
2142 | else if (VecVT == MVT::v2f64) |
2143 | return PPC::XVCMPGEDP; |
2144 | break; |
2145 | default: |
2146 | break; |
2147 | } |
2148 | llvm_unreachable("Invalid floating-point vector compare condition")::llvm::llvm_unreachable_internal("Invalid floating-point vector compare condition" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2148); |
2149 | } else { |
2150 | /* Handle some cases by swapping input operands. */ |
2151 | switch (CC) { |
2152 | case ISD::SETGE: CC = ISD::SETLE; Swap = true; break; |
2153 | case ISD::SETLT: CC = ISD::SETGT; Swap = true; break; |
2154 | case ISD::SETUGE: CC = ISD::SETULE; Swap = true; break; |
2155 | case ISD::SETULT: CC = ISD::SETUGT; Swap = true; break; |
2156 | default: break; |
2157 | } |
2158 | /* Handle some cases by negating the result. */ |
2159 | switch (CC) { |
2160 | case ISD::SETNE: CC = ISD::SETEQ; Negate = true; break; |
2161 | case ISD::SETUNE: CC = ISD::SETUEQ; Negate = true; break; |
2162 | case ISD::SETLE: CC = ISD::SETGT; Negate = true; break; |
2163 | case ISD::SETULE: CC = ISD::SETUGT; Negate = true; break; |
2164 | default: break; |
2165 | } |
2166 | /* We have instructions implementing the remaining cases. */ |
2167 | switch (CC) { |
2168 | case ISD::SETEQ: |
2169 | case ISD::SETUEQ: |
2170 | if (VecVT == MVT::v16i8) |
2171 | return PPC::VCMPEQUB; |
2172 | else if (VecVT == MVT::v8i16) |
2173 | return PPC::VCMPEQUH; |
2174 | else if (VecVT == MVT::v4i32) |
2175 | return PPC::VCMPEQUW; |
2176 | else if (VecVT == MVT::v2i64) |
2177 | return PPC::VCMPEQUD; |
2178 | break; |
2179 | case ISD::SETGT: |
2180 | if (VecVT == MVT::v16i8) |
2181 | return PPC::VCMPGTSB; |
2182 | else if (VecVT == MVT::v8i16) |
2183 | return PPC::VCMPGTSH; |
2184 | else if (VecVT == MVT::v4i32) |
2185 | return PPC::VCMPGTSW; |
2186 | else if (VecVT == MVT::v2i64) |
2187 | return PPC::VCMPGTSD; |
2188 | break; |
2189 | case ISD::SETUGT: |
2190 | if (VecVT == MVT::v16i8) |
2191 | return PPC::VCMPGTUB; |
2192 | else if (VecVT == MVT::v8i16) |
2193 | return PPC::VCMPGTUH; |
2194 | else if (VecVT == MVT::v4i32) |
2195 | return PPC::VCMPGTUW; |
2196 | else if (VecVT == MVT::v2i64) |
2197 | return PPC::VCMPGTUD; |
2198 | break; |
2199 | default: |
2200 | break; |
2201 | } |
2202 | llvm_unreachable("Invalid integer vector compare condition")::llvm::llvm_unreachable_internal("Invalid integer vector compare condition" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2202); |
2203 | } |
2204 | } |
2205 | |
2206 | SDNode *PPCDAGToDAGISel::SelectSETCC(SDNode *N) { |
2207 | SDLoc dl(N); |
2208 | unsigned Imm; |
2209 | ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get(); |
2210 | EVT PtrVT = |
2211 | CurDAG->getTargetLoweringInfo().getPointerTy(CurDAG->getDataLayout()); |
2212 | bool isPPC64 = (PtrVT == MVT::i64); |
2213 | |
2214 | if (!PPCSubTarget->useCRBits() && |
2215 | isInt32Immediate(N->getOperand(1), Imm)) { |
2216 | // We can codegen setcc op, imm very efficiently compared to a brcond. |
2217 | // Check for those cases here. |
2218 | // setcc op, 0 |
2219 | if (Imm == 0) { |
2220 | SDValue Op = N->getOperand(0); |
2221 | switch (CC) { |
2222 | default: break; |
2223 | case ISD::SETEQ: { |
2224 | Op = SDValue(CurDAG->getMachineNode(PPC::CNTLZW, dl, MVT::i32, Op), 0); |
2225 | SDValue Ops[] = { Op, getI32Imm(27, dl), getI32Imm(5, dl), |
2226 | getI32Imm(31, dl) }; |
2227 | return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); |
2228 | } |
2229 | case ISD::SETNE: { |
2230 | if (isPPC64) break; |
2231 | SDValue AD = |
2232 | SDValue(CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue, |
2233 | Op, getI32Imm(~0U, dl)), 0); |
2234 | return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, AD, Op, |
2235 | AD.getValue(1)); |
2236 | } |
2237 | case ISD::SETLT: { |
2238 | SDValue Ops[] = { Op, getI32Imm(1, dl), getI32Imm(31, dl), |
2239 | getI32Imm(31, dl) }; |
2240 | return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); |
2241 | } |
2242 | case ISD::SETGT: { |
2243 | SDValue T = |
2244 | SDValue(CurDAG->getMachineNode(PPC::NEG, dl, MVT::i32, Op), 0); |
2245 | T = SDValue(CurDAG->getMachineNode(PPC::ANDC, dl, MVT::i32, T, Op), 0); |
2246 | SDValue Ops[] = { T, getI32Imm(1, dl), getI32Imm(31, dl), |
2247 | getI32Imm(31, dl) }; |
2248 | return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); |
2249 | } |
2250 | } |
2251 | } else if (Imm == ~0U) { // setcc op, -1 |
2252 | SDValue Op = N->getOperand(0); |
2253 | switch (CC) { |
2254 | default: break; |
2255 | case ISD::SETEQ: |
2256 | if (isPPC64) break; |
2257 | Op = SDValue(CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue, |
2258 | Op, getI32Imm(1, dl)), 0); |
2259 | return CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32, |
2260 | SDValue(CurDAG->getMachineNode(PPC::LI, dl, |
2261 | MVT::i32, |
2262 | getI32Imm(0, dl)), |
2263 | 0), Op.getValue(1)); |
2264 | case ISD::SETNE: { |
2265 | if (isPPC64) break; |
2266 | Op = SDValue(CurDAG->getMachineNode(PPC::NOR, dl, MVT::i32, Op, Op), 0); |
2267 | SDNode *AD = CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue, |
2268 | Op, getI32Imm(~0U, dl)); |
2269 | return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, SDValue(AD, 0), |
2270 | Op, SDValue(AD, 1)); |
2271 | } |
2272 | case ISD::SETLT: { |
2273 | SDValue AD = SDValue(CurDAG->getMachineNode(PPC::ADDI, dl, MVT::i32, Op, |
2274 | getI32Imm(1, dl)), 0); |
2275 | SDValue AN = SDValue(CurDAG->getMachineNode(PPC::AND, dl, MVT::i32, AD, |
2276 | Op), 0); |
2277 | SDValue Ops[] = { AN, getI32Imm(1, dl), getI32Imm(31, dl), |
2278 | getI32Imm(31, dl) }; |
2279 | return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); |
2280 | } |
2281 | case ISD::SETGT: { |
2282 | SDValue Ops[] = { Op, getI32Imm(1, dl), getI32Imm(31, dl), |
2283 | getI32Imm(31, dl) }; |
2284 | Op = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 0); |
2285 | return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Op, |
2286 | getI32Imm(1, dl)); |
2287 | } |
2288 | } |
2289 | } |
2290 | } |
2291 | |
2292 | SDValue LHS = N->getOperand(0); |
2293 | SDValue RHS = N->getOperand(1); |
2294 | |
2295 | // Altivec Vector compare instructions do not set any CR register by default and |
2296 | // vector compare operations return the same type as the operands. |
2297 | if (LHS.getValueType().isVector()) { |
2298 | if (PPCSubTarget->hasQPX()) |
2299 | return nullptr; |
2300 | |
2301 | EVT VecVT = LHS.getValueType(); |
2302 | bool Swap, Negate; |
2303 | unsigned int VCmpInst = getVCmpInst(VecVT.getSimpleVT(), CC, |
2304 | PPCSubTarget->hasVSX(), Swap, Negate); |
2305 | if (Swap) |
2306 | std::swap(LHS, RHS); |
2307 | |
2308 | EVT ResVT = VecVT.changeVectorElementTypeToInteger(); |
2309 | if (Negate) { |
2310 | SDValue VCmp(CurDAG->getMachineNode(VCmpInst, dl, ResVT, LHS, RHS), 0); |
2311 | return CurDAG->SelectNodeTo(N, PPCSubTarget->hasVSX() ? PPC::XXLNOR : |
2312 | PPC::VNOR, |
2313 | ResVT, VCmp, VCmp); |
2314 | } |
2315 | |
2316 | return CurDAG->SelectNodeTo(N, VCmpInst, ResVT, LHS, RHS); |
2317 | } |
2318 | |
2319 | if (PPCSubTarget->useCRBits()) |
2320 | return nullptr; |
2321 | |
2322 | bool Inv; |
2323 | unsigned Idx = getCRIdxForSetCC(CC, Inv); |
2324 | SDValue CCReg = SelectCC(LHS, RHS, CC, dl); |
2325 | SDValue IntCR; |
2326 | |
2327 | // Force the ccreg into CR7. |
2328 | SDValue CR7Reg = CurDAG->getRegister(PPC::CR7, MVT::i32); |
2329 | |
2330 | SDValue InFlag(nullptr, 0); // Null incoming flag value. |
2331 | CCReg = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, CR7Reg, CCReg, |
2332 | InFlag).getValue(1); |
2333 | |
2334 | IntCR = SDValue(CurDAG->getMachineNode(PPC::MFOCRF, dl, MVT::i32, CR7Reg, |
2335 | CCReg), 0); |
2336 | |
2337 | SDValue Ops[] = { IntCR, getI32Imm((32 - (3 - Idx)) & 31, dl), |
2338 | getI32Imm(31, dl), getI32Imm(31, dl) }; |
2339 | if (!Inv) |
2340 | return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); |
2341 | |
2342 | // Get the specified bit. |
2343 | SDValue Tmp = |
2344 | SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 0); |
2345 | return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Tmp, getI32Imm(1, dl)); |
2346 | } |
2347 | |
2348 | SDNode *PPCDAGToDAGISel::transferMemOperands(SDNode *N, SDNode *Result) { |
2349 | // Transfer memoperands. |
2350 | MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1); |
2351 | MemOp[0] = cast<MemSDNode>(N)->getMemOperand(); |
2352 | cast<MachineSDNode>(Result)->setMemRefs(MemOp, MemOp + 1); |
2353 | return Result; |
2354 | } |
2355 | |
2356 | |
2357 | // Select - Convert the specified operand from a target-independent to a |
2358 | // target-specific node if it hasn't already been changed. |
2359 | SDNode *PPCDAGToDAGISel::Select(SDNode *N) { |
2360 | SDLoc dl(N); |
2361 | if (N->isMachineOpcode()) { |
2362 | N->setNodeId(-1); |
2363 | return nullptr; // Already selected. |
2364 | } |
2365 | |
2366 | // In case any misguided DAG-level optimizations form an ADD with a |
2367 | // TargetConstant operand, crash here instead of miscompiling (by selecting |
2368 | // an r+r add instead of some kind of r+i add). |
2369 | if (N->getOpcode() == ISD::ADD && |
2370 | N->getOperand(1).getOpcode() == ISD::TargetConstant) |
2371 | llvm_unreachable("Invalid ADD with TargetConstant operand")::llvm::llvm_unreachable_internal("Invalid ADD with TargetConstant operand" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2371); |
2372 | |
2373 | // Try matching complex bit permutations before doing anything else. |
2374 | if (SDNode *NN = SelectBitPermutation(N)) |
2375 | return NN; |
2376 | |
2377 | switch (N->getOpcode()) { |
2378 | default: break; |
2379 | |
2380 | case ISD::Constant: { |
2381 | if (N->getValueType(0) == MVT::i64) |
2382 | return SelectInt64(CurDAG, N); |
2383 | break; |
2384 | } |
2385 | |
2386 | case ISD::SETCC: { |
2387 | SDNode *SN = SelectSETCC(N); |
2388 | if (SN) |
2389 | return SN; |
2390 | break; |
2391 | } |
2392 | case PPCISD::GlobalBaseReg: |
2393 | return getGlobalBaseReg(); |
2394 | |
2395 | case ISD::FrameIndex: |
2396 | return getFrameIndex(N, N); |
2397 | |
2398 | case PPCISD::MFOCRF: { |
2399 | SDValue InFlag = N->getOperand(1); |
2400 | return CurDAG->getMachineNode(PPC::MFOCRF, dl, MVT::i32, |
2401 | N->getOperand(0), InFlag); |
2402 | } |
2403 | |
2404 | case PPCISD::READ_TIME_BASE: { |
2405 | return CurDAG->getMachineNode(PPC::ReadTB, dl, MVT::i32, MVT::i32, |
2406 | MVT::Other, N->getOperand(0)); |
2407 | } |
2408 | |
2409 | case PPCISD::SRA_ADDZE: { |
2410 | SDValue N0 = N->getOperand(0); |
2411 | SDValue ShiftAmt = |
2412 | CurDAG->getTargetConstant(*cast<ConstantSDNode>(N->getOperand(1))-> |
2413 | getConstantIntValue(), dl, |
2414 | N->getValueType(0)); |
2415 | if (N->getValueType(0) == MVT::i64) { |
2416 | SDNode *Op = |
2417 | CurDAG->getMachineNode(PPC::SRADI, dl, MVT::i64, MVT::Glue, |
2418 | N0, ShiftAmt); |
2419 | return CurDAG->SelectNodeTo(N, PPC::ADDZE8, MVT::i64, |
2420 | SDValue(Op, 0), SDValue(Op, 1)); |
2421 | } else { |
2422 | assert(N->getValueType(0) == MVT::i32 &&((N->getValueType(0) == MVT::i32 && "Expecting i64 or i32 in PPCISD::SRA_ADDZE" ) ? static_cast<void> (0) : __assert_fail ("N->getValueType(0) == MVT::i32 && \"Expecting i64 or i32 in PPCISD::SRA_ADDZE\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2423, __PRETTY_FUNCTION__)) |
2423 | "Expecting i64 or i32 in PPCISD::SRA_ADDZE")((N->getValueType(0) == MVT::i32 && "Expecting i64 or i32 in PPCISD::SRA_ADDZE" ) ? static_cast<void> (0) : __assert_fail ("N->getValueType(0) == MVT::i32 && \"Expecting i64 or i32 in PPCISD::SRA_ADDZE\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2423, __PRETTY_FUNCTION__)); |
2424 | SDNode *Op = |
2425 | CurDAG->getMachineNode(PPC::SRAWI, dl, MVT::i32, MVT::Glue, |
2426 | N0, ShiftAmt); |
2427 | return CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32, |
2428 | SDValue(Op, 0), SDValue(Op, 1)); |
2429 | } |
2430 | } |
2431 | |
2432 | case ISD::LOAD: { |
2433 | // Handle preincrement loads. |
2434 | LoadSDNode *LD = cast<LoadSDNode>(N); |
2435 | EVT LoadedVT = LD->getMemoryVT(); |
2436 | |
2437 | // Normal loads are handled by code generated from the .td file. |
2438 | if (LD->getAddressingMode() != ISD::PRE_INC) |
2439 | break; |
2440 | |
2441 | SDValue Offset = LD->getOffset(); |
2442 | if (Offset.getOpcode() == ISD::TargetConstant || |
2443 | Offset.getOpcode() == ISD::TargetGlobalAddress) { |
2444 | |
2445 | unsigned Opcode; |
2446 | bool isSExt = LD->getExtensionType() == ISD::SEXTLOAD; |
2447 | if (LD->getValueType(0) != MVT::i64) { |
2448 | // Handle PPC32 integer and normal FP loads. |
2449 | assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load")(((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load" ) ? static_cast<void> (0) : __assert_fail ("(!isSExt || LoadedVT == MVT::i16) && \"Invalid sext update load\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2449, __PRETTY_FUNCTION__)); |
2450 | switch (LoadedVT.getSimpleVT().SimpleTy) { |
2451 | default: llvm_unreachable("Invalid PPC load type!")::llvm::llvm_unreachable_internal("Invalid PPC load type!", "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2451); |
2452 | case MVT::f64: Opcode = PPC::LFDU; break; |
2453 | case MVT::f32: Opcode = PPC::LFSU; break; |
2454 | case MVT::i32: Opcode = PPC::LWZU; break; |
2455 | case MVT::i16: Opcode = isSExt ? PPC::LHAU : PPC::LHZU; break; |
2456 | case MVT::i1: |
2457 | case MVT::i8: Opcode = PPC::LBZU; break; |
2458 | } |
2459 | } else { |
2460 | assert(LD->getValueType(0) == MVT::i64 && "Unknown load result type!")((LD->getValueType(0) == MVT::i64 && "Unknown load result type!" ) ? static_cast<void> (0) : __assert_fail ("LD->getValueType(0) == MVT::i64 && \"Unknown load result type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2460, __PRETTY_FUNCTION__)); |
2461 | assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load")(((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load" ) ? static_cast<void> (0) : __assert_fail ("(!isSExt || LoadedVT == MVT::i16) && \"Invalid sext update load\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2461, __PRETTY_FUNCTION__)); |
2462 | switch (LoadedVT.getSimpleVT().SimpleTy) { |
2463 | default: llvm_unreachable("Invalid PPC load type!")::llvm::llvm_unreachable_internal("Invalid PPC load type!", "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2463); |
2464 | case MVT::i64: Opcode = PPC::LDU; break; |
2465 | case MVT::i32: Opcode = PPC::LWZU8; break; |
2466 | case MVT::i16: Opcode = isSExt ? PPC::LHAU8 : PPC::LHZU8; break; |
2467 | case MVT::i1: |
2468 | case MVT::i8: Opcode = PPC::LBZU8; break; |
2469 | } |
2470 | } |
2471 | |
2472 | SDValue Chain = LD->getChain(); |
2473 | SDValue Base = LD->getBasePtr(); |
2474 | SDValue Ops[] = { Offset, Base, Chain }; |
2475 | return transferMemOperands( |
2476 | N, CurDAG->getMachineNode( |
2477 | Opcode, dl, LD->getValueType(0), |
2478 | PPCLowering->getPointerTy(CurDAG->getDataLayout()), MVT::Other, |
2479 | Ops)); |
2480 | } else { |
2481 | unsigned Opcode; |
2482 | bool isSExt = LD->getExtensionType() == ISD::SEXTLOAD; |
2483 | if (LD->getValueType(0) != MVT::i64) { |
2484 | // Handle PPC32 integer and normal FP loads. |
2485 | assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load")(((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load" ) ? static_cast<void> (0) : __assert_fail ("(!isSExt || LoadedVT == MVT::i16) && \"Invalid sext update load\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2485, __PRETTY_FUNCTION__)); |
2486 | switch (LoadedVT.getSimpleVT().SimpleTy) { |
2487 | default: llvm_unreachable("Invalid PPC load type!")::llvm::llvm_unreachable_internal("Invalid PPC load type!", "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2487); |
2488 | case MVT::v4f64: Opcode = PPC::QVLFDUX; break; // QPX |
2489 | case MVT::v4f32: Opcode = PPC::QVLFSUX; break; // QPX |
2490 | case MVT::f64: Opcode = PPC::LFDUX; break; |
2491 | case MVT::f32: Opcode = PPC::LFSUX; break; |
2492 | case MVT::i32: Opcode = PPC::LWZUX; break; |
2493 | case MVT::i16: Opcode = isSExt ? PPC::LHAUX : PPC::LHZUX; break; |
2494 | case MVT::i1: |
2495 | case MVT::i8: Opcode = PPC::LBZUX; break; |
2496 | } |
2497 | } else { |
2498 | assert(LD->getValueType(0) == MVT::i64 && "Unknown load result type!")((LD->getValueType(0) == MVT::i64 && "Unknown load result type!" ) ? static_cast<void> (0) : __assert_fail ("LD->getValueType(0) == MVT::i64 && \"Unknown load result type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2498, __PRETTY_FUNCTION__)); |
2499 | assert((!isSExt || LoadedVT == MVT::i16 || LoadedVT == MVT::i32) &&(((!isSExt || LoadedVT == MVT::i16 || LoadedVT == MVT::i32) && "Invalid sext update load") ? static_cast<void> (0) : __assert_fail ("(!isSExt || LoadedVT == MVT::i16 || LoadedVT == MVT::i32) && \"Invalid sext update load\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2500, __PRETTY_FUNCTION__)) |
2500 | "Invalid sext update load")(((!isSExt || LoadedVT == MVT::i16 || LoadedVT == MVT::i32) && "Invalid sext update load") ? static_cast<void> (0) : __assert_fail ("(!isSExt || LoadedVT == MVT::i16 || LoadedVT == MVT::i32) && \"Invalid sext update load\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2500, __PRETTY_FUNCTION__)); |
2501 | switch (LoadedVT.getSimpleVT().SimpleTy) { |
2502 | default: llvm_unreachable("Invalid PPC load type!")::llvm::llvm_unreachable_internal("Invalid PPC load type!", "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2502); |
2503 | case MVT::i64: Opcode = PPC::LDUX; break; |
2504 | case MVT::i32: Opcode = isSExt ? PPC::LWAUX : PPC::LWZUX8; break; |
2505 | case MVT::i16: Opcode = isSExt ? PPC::LHAUX8 : PPC::LHZUX8; break; |
2506 | case MVT::i1: |
2507 | case MVT::i8: Opcode = PPC::LBZUX8; break; |
2508 | } |
2509 | } |
2510 | |
2511 | SDValue Chain = LD->getChain(); |
2512 | SDValue Base = LD->getBasePtr(); |
2513 | SDValue Ops[] = { Base, Offset, Chain }; |
2514 | return transferMemOperands( |
2515 | N, CurDAG->getMachineNode( |
2516 | Opcode, dl, LD->getValueType(0), |
2517 | PPCLowering->getPointerTy(CurDAG->getDataLayout()), MVT::Other, |
2518 | Ops)); |
2519 | } |
2520 | } |
2521 | |
2522 | case ISD::AND: { |
2523 | unsigned Imm, Imm2, SH, MB, ME; |
2524 | uint64_t Imm64; |
2525 | |
2526 | // If this is an and of a value rotated between 0 and 31 bits and then and'd |
2527 | // with a mask, emit rlwinm |
2528 | if (isInt32Immediate(N->getOperand(1), Imm) && |
2529 | isRotateAndMask(N->getOperand(0).getNode(), Imm, false, SH, MB, ME)) { |
2530 | SDValue Val = N->getOperand(0).getOperand(0); |
2531 | SDValue Ops[] = { Val, getI32Imm(SH, dl), getI32Imm(MB, dl), |
2532 | getI32Imm(ME, dl) }; |
2533 | return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); |
2534 | } |
2535 | // If this is just a masked value where the input is not handled above, and |
2536 | // is not a rotate-left (handled by a pattern in the .td file), emit rlwinm |
2537 | if (isInt32Immediate(N->getOperand(1), Imm) && |
2538 | isRunOfOnes(Imm, MB, ME) && |
2539 | N->getOperand(0).getOpcode() != ISD::ROTL) { |
2540 | SDValue Val = N->getOperand(0); |
2541 | SDValue Ops[] = { Val, getI32Imm(0, dl), getI32Imm(MB, dl), |
2542 | getI32Imm(ME, dl) }; |
2543 | return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); |
2544 | } |
2545 | // If this is a 64-bit zero-extension mask, emit rldicl. |
2546 | if (isInt64Immediate(N->getOperand(1).getNode(), Imm64) && |
2547 | isMask_64(Imm64)) { |
2548 | SDValue Val = N->getOperand(0); |
2549 | MB = 64 - countTrailingOnes(Imm64); |
2550 | SH = 0; |
2551 | |
2552 | // If the operand is a logical right shift, we can fold it into this |
2553 | // instruction: rldicl(rldicl(x, 64-n, n), 0, mb) -> rldicl(x, 64-n, mb) |
2554 | // for n <= mb. The right shift is really a left rotate followed by a |
2555 | // mask, and this mask is a more-restrictive sub-mask of the mask implied |
2556 | // by the shift. |
2557 | if (Val.getOpcode() == ISD::SRL && |
2558 | isInt32Immediate(Val.getOperand(1).getNode(), Imm) && Imm <= MB) { |
2559 | assert(Imm < 64 && "Illegal shift amount")((Imm < 64 && "Illegal shift amount") ? static_cast <void> (0) : __assert_fail ("Imm < 64 && \"Illegal shift amount\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2559, __PRETTY_FUNCTION__)); |
2560 | Val = Val.getOperand(0); |
2561 | SH = 64 - Imm; |
2562 | } |
2563 | |
2564 | SDValue Ops[] = { Val, getI32Imm(SH, dl), getI32Imm(MB, dl) }; |
2565 | return CurDAG->SelectNodeTo(N, PPC::RLDICL, MVT::i64, Ops); |
2566 | } |
2567 | // AND X, 0 -> 0, not "rlwinm 32". |
2568 | if (isInt32Immediate(N->getOperand(1), Imm) && (Imm == 0)) { |
2569 | ReplaceUses(SDValue(N, 0), N->getOperand(1)); |
2570 | return nullptr; |
2571 | } |
2572 | // ISD::OR doesn't get all the bitfield insertion fun. |
2573 | // (and (or x, c1), c2) where isRunOfOnes(~(c1^c2)) might be a |
2574 | // bitfield insert. |
2575 | if (isInt32Immediate(N->getOperand(1), Imm) && |
2576 | N->getOperand(0).getOpcode() == ISD::OR && |
2577 | isInt32Immediate(N->getOperand(0).getOperand(1), Imm2)) { |
2578 | // The idea here is to check whether this is equivalent to: |
2579 | // (c1 & m) | (x & ~m) |
2580 | // where m is a run-of-ones mask. The logic here is that, for each bit in |
2581 | // c1 and c2: |
2582 | // - if both are 1, then the output will be 1. |
2583 | // - if both are 0, then the output will be 0. |
2584 | // - if the bit in c1 is 0, and the bit in c2 is 1, then the output will |
2585 | // come from x. |
2586 | // - if the bit in c1 is 1, and the bit in c2 is 0, then the output will |
2587 | // be 0. |
2588 | // If that last condition is never the case, then we can form m from the |
2589 | // bits that are the same between c1 and c2. |
2590 | unsigned MB, ME; |
2591 | if (isRunOfOnes(~(Imm^Imm2), MB, ME) && !(~Imm & Imm2)) { |
2592 | SDValue Ops[] = { N->getOperand(0).getOperand(0), |
2593 | N->getOperand(0).getOperand(1), |
2594 | getI32Imm(0, dl), getI32Imm(MB, dl), |
2595 | getI32Imm(ME, dl) }; |
2596 | return CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops); |
2597 | } |
2598 | } |
2599 | |
2600 | // Other cases are autogenerated. |
2601 | break; |
2602 | } |
2603 | case ISD::OR: { |
2604 | if (N->getValueType(0) == MVT::i32) |
2605 | if (SDNode *I = SelectBitfieldInsert(N)) |
2606 | return I; |
2607 | |
2608 | short Imm; |
2609 | if (N->getOperand(0)->getOpcode() == ISD::FrameIndex && |
2610 | isIntS16Immediate(N->getOperand(1), Imm)) { |
2611 | APInt LHSKnownZero, LHSKnownOne; |
2612 | CurDAG->computeKnownBits(N->getOperand(0), LHSKnownZero, LHSKnownOne); |
2613 | |
2614 | // If this is equivalent to an add, then we can fold it with the |
2615 | // FrameIndex calculation. |
2616 | if ((LHSKnownZero.getZExtValue()|~(uint64_t)Imm) == ~0ULL) |
2617 | return getFrameIndex(N, N->getOperand(0).getNode(), (int)Imm); |
2618 | } |
2619 | |
2620 | // Other cases are autogenerated. |
2621 | break; |
2622 | } |
2623 | case ISD::ADD: { |
2624 | short Imm; |
2625 | if (N->getOperand(0)->getOpcode() == ISD::FrameIndex && |
2626 | isIntS16Immediate(N->getOperand(1), Imm)) |
2627 | return getFrameIndex(N, N->getOperand(0).getNode(), (int)Imm); |
2628 | |
2629 | break; |
2630 | } |
2631 | case ISD::SHL: { |
2632 | unsigned Imm, SH, MB, ME; |
2633 | if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::AND, Imm) && |
2634 | isRotateAndMask(N, Imm, true, SH, MB, ME)) { |
2635 | SDValue Ops[] = { N->getOperand(0).getOperand(0), |
2636 | getI32Imm(SH, dl), getI32Imm(MB, dl), |
2637 | getI32Imm(ME, dl) }; |
2638 | return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); |
2639 | } |
2640 | |
2641 | // Other cases are autogenerated. |
2642 | break; |
2643 | } |
2644 | case ISD::SRL: { |
2645 | unsigned Imm, SH, MB, ME; |
2646 | if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::AND, Imm) && |
2647 | isRotateAndMask(N, Imm, true, SH, MB, ME)) { |
2648 | SDValue Ops[] = { N->getOperand(0).getOperand(0), |
2649 | getI32Imm(SH, dl), getI32Imm(MB, dl), |
2650 | getI32Imm(ME, dl) }; |
2651 | return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); |
2652 | } |
2653 | |
2654 | // Other cases are autogenerated. |
2655 | break; |
2656 | } |
2657 | // FIXME: Remove this once the ANDI glue bug is fixed: |
2658 | case PPCISD::ANDIo_1_EQ_BIT: |
2659 | case PPCISD::ANDIo_1_GT_BIT: { |
2660 | if (!ANDIGlueBug) |
2661 | break; |
2662 | |
2663 | EVT InVT = N->getOperand(0).getValueType(); |
2664 | assert((InVT == MVT::i64 || InVT == MVT::i32) &&(((InVT == MVT::i64 || InVT == MVT::i32) && "Invalid input type for ANDIo_1_EQ_BIT" ) ? static_cast<void> (0) : __assert_fail ("(InVT == MVT::i64 || InVT == MVT::i32) && \"Invalid input type for ANDIo_1_EQ_BIT\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2665, __PRETTY_FUNCTION__)) |
2665 | "Invalid input type for ANDIo_1_EQ_BIT")(((InVT == MVT::i64 || InVT == MVT::i32) && "Invalid input type for ANDIo_1_EQ_BIT" ) ? static_cast<void> (0) : __assert_fail ("(InVT == MVT::i64 || InVT == MVT::i32) && \"Invalid input type for ANDIo_1_EQ_BIT\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2665, __PRETTY_FUNCTION__)); |
2666 | |
2667 | unsigned Opcode = (InVT == MVT::i64) ? PPC::ANDIo8 : PPC::ANDIo; |
2668 | SDValue AndI(CurDAG->getMachineNode(Opcode, dl, InVT, MVT::Glue, |
2669 | N->getOperand(0), |
2670 | CurDAG->getTargetConstant(1, dl, InVT)), |
2671 | 0); |
2672 | SDValue CR0Reg = CurDAG->getRegister(PPC::CR0, MVT::i32); |
2673 | SDValue SRIdxVal = |
2674 | CurDAG->getTargetConstant(N->getOpcode() == PPCISD::ANDIo_1_EQ_BIT ? |
2675 | PPC::sub_eq : PPC::sub_gt, dl, MVT::i32); |
2676 | |
2677 | return CurDAG->SelectNodeTo(N, TargetOpcode::EXTRACT_SUBREG, MVT::i1, |
2678 | CR0Reg, SRIdxVal, |
2679 | SDValue(AndI.getNode(), 1) /* glue */); |
2680 | } |
2681 | case ISD::SELECT_CC: { |
2682 | ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get(); |
2683 | EVT PtrVT = |
2684 | CurDAG->getTargetLoweringInfo().getPointerTy(CurDAG->getDataLayout()); |
2685 | bool isPPC64 = (PtrVT == MVT::i64); |
2686 | |
2687 | // If this is a select of i1 operands, we'll pattern match it. |
2688 | if (PPCSubTarget->useCRBits() && |
2689 | N->getOperand(0).getValueType() == MVT::i1) |
2690 | break; |
2691 | |
2692 | // Handle the setcc cases here. select_cc lhs, 0, 1, 0, cc |
2693 | if (!isPPC64) |
2694 | if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N->getOperand(1))) |
2695 | if (ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N->getOperand(2))) |
2696 | if (ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N->getOperand(3))) |
2697 | if (N1C->isNullValue() && N3C->isNullValue() && |
2698 | N2C->getZExtValue() == 1ULL && CC == ISD::SETNE && |
2699 | // FIXME: Implement this optzn for PPC64. |
2700 | N->getValueType(0) == MVT::i32) { |
2701 | SDNode *Tmp = |
2702 | CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue, |
2703 | N->getOperand(0), getI32Imm(~0U, dl)); |
2704 | return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, |
2705 | SDValue(Tmp, 0), N->getOperand(0), |
2706 | SDValue(Tmp, 1)); |
2707 | } |
2708 | |
2709 | SDValue CCReg = SelectCC(N->getOperand(0), N->getOperand(1), CC, dl); |
2710 | |
2711 | if (N->getValueType(0) == MVT::i1) { |
2712 | // An i1 select is: (c & t) | (!c & f). |
2713 | bool Inv; |
2714 | unsigned Idx = getCRIdxForSetCC(CC, Inv); |
2715 | |
2716 | unsigned SRI; |
2717 | switch (Idx) { |
2718 | default: llvm_unreachable("Invalid CC index")::llvm::llvm_unreachable_internal("Invalid CC index", "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2718); |
2719 | case 0: SRI = PPC::sub_lt; break; |
2720 | case 1: SRI = PPC::sub_gt; break; |
2721 | case 2: SRI = PPC::sub_eq; break; |
2722 | case 3: SRI = PPC::sub_un; break; |
2723 | } |
2724 | |
2725 | SDValue CCBit = CurDAG->getTargetExtractSubreg(SRI, dl, MVT::i1, CCReg); |
2726 | |
2727 | SDValue NotCCBit(CurDAG->getMachineNode(PPC::CRNOR, dl, MVT::i1, |
2728 | CCBit, CCBit), 0); |
2729 | SDValue C = Inv ? NotCCBit : CCBit, |
2730 | NotC = Inv ? CCBit : NotCCBit; |
2731 | |
2732 | SDValue CAndT(CurDAG->getMachineNode(PPC::CRAND, dl, MVT::i1, |
2733 | C, N->getOperand(2)), 0); |
2734 | SDValue NotCAndF(CurDAG->getMachineNode(PPC::CRAND, dl, MVT::i1, |
2735 | NotC, N->getOperand(3)), 0); |
2736 | |
2737 | return CurDAG->SelectNodeTo(N, PPC::CROR, MVT::i1, CAndT, NotCAndF); |
2738 | } |
2739 | |
2740 | unsigned BROpc = getPredicateForSetCC(CC); |
2741 | |
2742 | unsigned SelectCCOp; |
2743 | if (N->getValueType(0) == MVT::i32) |
2744 | SelectCCOp = PPC::SELECT_CC_I4; |
2745 | else if (N->getValueType(0) == MVT::i64) |
2746 | SelectCCOp = PPC::SELECT_CC_I8; |
2747 | else if (N->getValueType(0) == MVT::f32) |
2748 | if (PPCSubTarget->hasP8Vector()) |
2749 | SelectCCOp = PPC::SELECT_CC_VSSRC; |
2750 | else |
2751 | SelectCCOp = PPC::SELECT_CC_F4; |
2752 | else if (N->getValueType(0) == MVT::f64) |
2753 | if (PPCSubTarget->hasVSX()) |
2754 | SelectCCOp = PPC::SELECT_CC_VSFRC; |
2755 | else |
2756 | SelectCCOp = PPC::SELECT_CC_F8; |
2757 | else if (PPCSubTarget->hasQPX() && N->getValueType(0) == MVT::v4f64) |
2758 | SelectCCOp = PPC::SELECT_CC_QFRC; |
2759 | else if (PPCSubTarget->hasQPX() && N->getValueType(0) == MVT::v4f32) |
2760 | SelectCCOp = PPC::SELECT_CC_QSRC; |
2761 | else if (PPCSubTarget->hasQPX() && N->getValueType(0) == MVT::v4i1) |
2762 | SelectCCOp = PPC::SELECT_CC_QBRC; |
2763 | else if (N->getValueType(0) == MVT::v2f64 || |
2764 | N->getValueType(0) == MVT::v2i64) |
2765 | SelectCCOp = PPC::SELECT_CC_VSRC; |
2766 | else |
2767 | SelectCCOp = PPC::SELECT_CC_VRRC; |
2768 | |
2769 | SDValue Ops[] = { CCReg, N->getOperand(2), N->getOperand(3), |
2770 | getI32Imm(BROpc, dl) }; |
2771 | return CurDAG->SelectNodeTo(N, SelectCCOp, N->getValueType(0), Ops); |
2772 | } |
2773 | case ISD::VSELECT: |
2774 | if (PPCSubTarget->hasVSX()) { |
2775 | SDValue Ops[] = { N->getOperand(2), N->getOperand(1), N->getOperand(0) }; |
2776 | return CurDAG->SelectNodeTo(N, PPC::XXSEL, N->getValueType(0), Ops); |
2777 | } |
2778 | |
2779 | break; |
2780 | case ISD::VECTOR_SHUFFLE: |
2781 | if (PPCSubTarget->hasVSX() && (N->getValueType(0) == MVT::v2f64 || |
2782 | N->getValueType(0) == MVT::v2i64)) { |
2783 | ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N); |
2784 | |
2785 | SDValue Op1 = N->getOperand(SVN->getMaskElt(0) < 2 ? 0 : 1), |
2786 | Op2 = N->getOperand(SVN->getMaskElt(1) < 2 ? 0 : 1); |
2787 | unsigned DM[2]; |
2788 | |
2789 | for (int i = 0; i < 2; ++i) |
2790 | if (SVN->getMaskElt(i) <= 0 || SVN->getMaskElt(i) == 2) |
2791 | DM[i] = 0; |
2792 | else |
2793 | DM[i] = 1; |
2794 | |
2795 | if (Op1 == Op2 && DM[0] == 0 && DM[1] == 0 && |
2796 | Op1.getOpcode() == ISD::SCALAR_TO_VECTOR && |
2797 | isa<LoadSDNode>(Op1.getOperand(0))) { |
2798 | LoadSDNode *LD = cast<LoadSDNode>(Op1.getOperand(0)); |
2799 | SDValue Base, Offset; |
2800 | |
2801 | if (LD->isUnindexed() && |
2802 | (LD->getMemoryVT() == MVT::f64 || |
2803 | LD->getMemoryVT() == MVT::i64) && |
2804 | SelectAddrIdxOnly(LD->getBasePtr(), Base, Offset)) { |
2805 | SDValue Chain = LD->getChain(); |
2806 | SDValue Ops[] = { Base, Offset, Chain }; |
2807 | return CurDAG->SelectNodeTo(N, PPC::LXVDSX, |
2808 | N->getValueType(0), Ops); |
2809 | } |
2810 | } |
2811 | |
2812 | // For little endian, we must swap the input operands and adjust |
2813 | // the mask elements (reverse and invert them). |
2814 | if (PPCSubTarget->isLittleEndian()) { |
2815 | std::swap(Op1, Op2); |
2816 | unsigned tmp = DM[0]; |
2817 | DM[0] = 1 - DM[1]; |
2818 | DM[1] = 1 - tmp; |
2819 | } |
2820 | |
2821 | SDValue DMV = CurDAG->getTargetConstant(DM[1] | (DM[0] << 1), dl, |
2822 | MVT::i32); |
2823 | SDValue Ops[] = { Op1, Op2, DMV }; |
2824 | return CurDAG->SelectNodeTo(N, PPC::XXPERMDI, N->getValueType(0), Ops); |
2825 | } |
2826 | |
2827 | break; |
2828 | case PPCISD::BDNZ: |
2829 | case PPCISD::BDZ: { |
2830 | bool IsPPC64 = PPCSubTarget->isPPC64(); |
2831 | SDValue Ops[] = { N->getOperand(1), N->getOperand(0) }; |
2832 | return CurDAG->SelectNodeTo(N, N->getOpcode() == PPCISD::BDNZ ? |
2833 | (IsPPC64 ? PPC::BDNZ8 : PPC::BDNZ) : |
2834 | (IsPPC64 ? PPC::BDZ8 : PPC::BDZ), |
2835 | MVT::Other, Ops); |
2836 | } |
2837 | case PPCISD::COND_BRANCH: { |
2838 | // Op #0 is the Chain. |
2839 | // Op #1 is the PPC::PRED_* number. |
2840 | // Op #2 is the CR# |
2841 | // Op #3 is the Dest MBB |
2842 | // Op #4 is the Flag. |
2843 | // Prevent PPC::PRED_* from being selected into LI. |
2844 | SDValue Pred = |
2845 | getI32Imm(cast<ConstantSDNode>(N->getOperand(1))->getZExtValue(), dl); |
2846 | SDValue Ops[] = { Pred, N->getOperand(2), N->getOperand(3), |
2847 | N->getOperand(0), N->getOperand(4) }; |
2848 | return CurDAG->SelectNodeTo(N, PPC::BCC, MVT::Other, Ops); |
2849 | } |
2850 | case ISD::BR_CC: { |
2851 | ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get(); |
2852 | unsigned PCC = getPredicateForSetCC(CC); |
2853 | |
2854 | if (N->getOperand(2).getValueType() == MVT::i1) { |
2855 | unsigned Opc; |
2856 | bool Swap; |
2857 | switch (PCC) { |
2858 | default: llvm_unreachable("Unexpected Boolean-operand predicate")::llvm::llvm_unreachable_internal("Unexpected Boolean-operand predicate" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2858); |
2859 | case PPC::PRED_LT: Opc = PPC::CRANDC; Swap = true; break; |
2860 | case PPC::PRED_LE: Opc = PPC::CRORC; Swap = true; break; |
2861 | case PPC::PRED_EQ: Opc = PPC::CREQV; Swap = false; break; |
2862 | case PPC::PRED_GE: Opc = PPC::CRORC; Swap = false; break; |
2863 | case PPC::PRED_GT: Opc = PPC::CRANDC; Swap = false; break; |
2864 | case PPC::PRED_NE: Opc = PPC::CRXOR; Swap = false; break; |
2865 | } |
2866 | |
2867 | SDValue BitComp(CurDAG->getMachineNode(Opc, dl, MVT::i1, |
2868 | N->getOperand(Swap ? 3 : 2), |
2869 | N->getOperand(Swap ? 2 : 3)), 0); |
2870 | return CurDAG->SelectNodeTo(N, PPC::BC, MVT::Other, |
2871 | BitComp, N->getOperand(4), N->getOperand(0)); |
2872 | } |
2873 | |
2874 | SDValue CondCode = SelectCC(N->getOperand(2), N->getOperand(3), CC, dl); |
2875 | SDValue Ops[] = { getI32Imm(PCC, dl), CondCode, |
2876 | N->getOperand(4), N->getOperand(0) }; |
2877 | return CurDAG->SelectNodeTo(N, PPC::BCC, MVT::Other, Ops); |
2878 | } |
2879 | case ISD::BRIND: { |
2880 | // FIXME: Should custom lower this. |
2881 | SDValue Chain = N->getOperand(0); |
2882 | SDValue Target = N->getOperand(1); |
2883 | unsigned Opc = Target.getValueType() == MVT::i32 ? PPC::MTCTR : PPC::MTCTR8; |
2884 | unsigned Reg = Target.getValueType() == MVT::i32 ? PPC::BCTR : PPC::BCTR8; |
2885 | Chain = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Glue, Target, |
2886 | Chain), 0); |
2887 | return CurDAG->SelectNodeTo(N, Reg, MVT::Other, Chain); |
2888 | } |
2889 | case PPCISD::TOC_ENTRY: { |
2890 | assert ((PPCSubTarget->isPPC64() || PPCSubTarget->isSVR4ABI()) &&(((PPCSubTarget->isPPC64() || PPCSubTarget->isSVR4ABI() ) && "Only supported for 64-bit ABI and 32-bit SVR4") ? static_cast<void> (0) : __assert_fail ("(PPCSubTarget->isPPC64() || PPCSubTarget->isSVR4ABI()) && \"Only supported for 64-bit ABI and 32-bit SVR4\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2891, __PRETTY_FUNCTION__)) |
2891 | "Only supported for 64-bit ABI and 32-bit SVR4")(((PPCSubTarget->isPPC64() || PPCSubTarget->isSVR4ABI() ) && "Only supported for 64-bit ABI and 32-bit SVR4") ? static_cast<void> (0) : __assert_fail ("(PPCSubTarget->isPPC64() || PPCSubTarget->isSVR4ABI()) && \"Only supported for 64-bit ABI and 32-bit SVR4\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2891, __PRETTY_FUNCTION__)); |
2892 | if (PPCSubTarget->isSVR4ABI() && !PPCSubTarget->isPPC64()) { |
2893 | SDValue GA = N->getOperand(0); |
2894 | return transferMemOperands(N, CurDAG->getMachineNode(PPC::LWZtoc, dl, |
2895 | MVT::i32, GA, N->getOperand(1))); |
2896 | } |
2897 | |
2898 | // For medium and large code model, we generate two instructions as |
2899 | // described below. Otherwise we allow SelectCodeCommon to handle this, |
2900 | // selecting one of LDtoc, LDtocJTI, LDtocCPT, and LDtocBA. |
2901 | CodeModel::Model CModel = TM.getCodeModel(); |
2902 | if (CModel != CodeModel::Medium && CModel != CodeModel::Large) |
2903 | break; |
2904 | |
2905 | // The first source operand is a TargetGlobalAddress or a TargetJumpTable. |
2906 | // If it is an externally defined symbol, a symbol with common linkage, |
2907 | // a non-local function address, or a jump table address, or if we are |
2908 | // generating code for large code model, we generate: |
2909 | // LDtocL(<ga:@sym>, ADDIStocHA(%X2, <ga:@sym>)) |
2910 | // Otherwise we generate: |
2911 | // ADDItocL(ADDIStocHA(%X2, <ga:@sym>), <ga:@sym>) |
2912 | SDValue GA = N->getOperand(0); |
2913 | SDValue TOCbase = N->getOperand(1); |
2914 | SDNode *Tmp = CurDAG->getMachineNode(PPC::ADDIStocHA, dl, MVT::i64, |
2915 | TOCbase, GA); |
2916 | |
2917 | if (isa<JumpTableSDNode>(GA) || isa<BlockAddressSDNode>(GA) || |
2918 | CModel == CodeModel::Large) |
2919 | return transferMemOperands(N, CurDAG->getMachineNode(PPC::LDtocL, dl, |
2920 | MVT::i64, GA, SDValue(Tmp, 0))); |
2921 | |
2922 | if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(GA)) { |
2923 | const GlobalValue *GValue = G->getGlobal(); |
2924 | if ((GValue->getType()->getElementType()->isFunctionTy() && |
2925 | !GValue->isStrongDefinitionForLinker()) || |
2926 | GValue->isDeclaration() || GValue->hasCommonLinkage() || |
2927 | GValue->hasAvailableExternallyLinkage()) |
2928 | return transferMemOperands(N, CurDAG->getMachineNode(PPC::LDtocL, dl, |
2929 | MVT::i64, GA, SDValue(Tmp, 0))); |
2930 | } |
2931 | |
2932 | return CurDAG->getMachineNode(PPC::ADDItocL, dl, MVT::i64, |
2933 | SDValue(Tmp, 0), GA); |
2934 | } |
2935 | case PPCISD::PPC32_PICGOT: { |
2936 | // Generate a PIC-safe GOT reference. |
2937 | assert(!PPCSubTarget->isPPC64() && PPCSubTarget->isSVR4ABI() &&((!PPCSubTarget->isPPC64() && PPCSubTarget->isSVR4ABI () && "PPCISD::PPC32_PICGOT is only supported for 32-bit SVR4" ) ? static_cast<void> (0) : __assert_fail ("!PPCSubTarget->isPPC64() && PPCSubTarget->isSVR4ABI() && \"PPCISD::PPC32_PICGOT is only supported for 32-bit SVR4\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2938, __PRETTY_FUNCTION__)) |
2938 | "PPCISD::PPC32_PICGOT is only supported for 32-bit SVR4")((!PPCSubTarget->isPPC64() && PPCSubTarget->isSVR4ABI () && "PPCISD::PPC32_PICGOT is only supported for 32-bit SVR4" ) ? static_cast<void> (0) : __assert_fail ("!PPCSubTarget->isPPC64() && PPCSubTarget->isSVR4ABI() && \"PPCISD::PPC32_PICGOT is only supported for 32-bit SVR4\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2938, __PRETTY_FUNCTION__)); |
2939 | return CurDAG->SelectNodeTo( |
2940 | N, PPC::PPC32PICGOT, PPCLowering->getPointerTy(CurDAG->getDataLayout()), |
2941 | MVT::i32); |
2942 | } |
2943 | case PPCISD::VADD_SPLAT: { |
2944 | // This expands into one of three sequences, depending on whether |
2945 | // the first operand is odd or even, positive or negative. |
2946 | assert(isa<ConstantSDNode>(N->getOperand(0)) &&((isa<ConstantSDNode>(N->getOperand(0)) && isa <ConstantSDNode>(N->getOperand(1)) && "Invalid operand on VADD_SPLAT!" ) ? static_cast<void> (0) : __assert_fail ("isa<ConstantSDNode>(N->getOperand(0)) && isa<ConstantSDNode>(N->getOperand(1)) && \"Invalid operand on VADD_SPLAT!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2948, __PRETTY_FUNCTION__)) |
2947 | isa<ConstantSDNode>(N->getOperand(1)) &&((isa<ConstantSDNode>(N->getOperand(0)) && isa <ConstantSDNode>(N->getOperand(1)) && "Invalid operand on VADD_SPLAT!" ) ? static_cast<void> (0) : __assert_fail ("isa<ConstantSDNode>(N->getOperand(0)) && isa<ConstantSDNode>(N->getOperand(1)) && \"Invalid operand on VADD_SPLAT!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2948, __PRETTY_FUNCTION__)) |
2948 | "Invalid operand on VADD_SPLAT!")((isa<ConstantSDNode>(N->getOperand(0)) && isa <ConstantSDNode>(N->getOperand(1)) && "Invalid operand on VADD_SPLAT!" ) ? static_cast<void> (0) : __assert_fail ("isa<ConstantSDNode>(N->getOperand(0)) && isa<ConstantSDNode>(N->getOperand(1)) && \"Invalid operand on VADD_SPLAT!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2948, __PRETTY_FUNCTION__)); |
2949 | |
2950 | int Elt = N->getConstantOperandVal(0); |
2951 | int EltSize = N->getConstantOperandVal(1); |
2952 | unsigned Opc1, Opc2, Opc3; |
2953 | EVT VT; |
2954 | |
2955 | if (EltSize == 1) { |
2956 | Opc1 = PPC::VSPLTISB; |
2957 | Opc2 = PPC::VADDUBM; |
2958 | Opc3 = PPC::VSUBUBM; |
2959 | VT = MVT::v16i8; |
2960 | } else if (EltSize == 2) { |
2961 | Opc1 = PPC::VSPLTISH; |
2962 | Opc2 = PPC::VADDUHM; |
2963 | Opc3 = PPC::VSUBUHM; |
2964 | VT = MVT::v8i16; |
2965 | } else { |
2966 | assert(EltSize == 4 && "Invalid element size on VADD_SPLAT!")((EltSize == 4 && "Invalid element size on VADD_SPLAT!" ) ? static_cast<void> (0) : __assert_fail ("EltSize == 4 && \"Invalid element size on VADD_SPLAT!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 2966, __PRETTY_FUNCTION__)); |
2967 | Opc1 = PPC::VSPLTISW; |
2968 | Opc2 = PPC::VADDUWM; |
2969 | Opc3 = PPC::VSUBUWM; |
2970 | VT = MVT::v4i32; |
2971 | } |
2972 | |
2973 | if ((Elt & 1) == 0) { |
2974 | // Elt is even, in the range [-32,-18] + [16,30]. |
2975 | // |
2976 | // Convert: VADD_SPLAT elt, size |
2977 | // Into: tmp = VSPLTIS[BHW] elt |
2978 | // VADDU[BHW]M tmp, tmp |
2979 | // Where: [BHW] = B for size = 1, H for size = 2, W for size = 4 |
2980 | SDValue EltVal = getI32Imm(Elt >> 1, dl); |
2981 | SDNode *Tmp = CurDAG->getMachineNode(Opc1, dl, VT, EltVal); |
2982 | SDValue TmpVal = SDValue(Tmp, 0); |
2983 | return CurDAG->getMachineNode(Opc2, dl, VT, TmpVal, TmpVal); |
2984 | |
2985 | } else if (Elt > 0) { |
2986 | // Elt is odd and positive, in the range [17,31]. |
2987 | // |
2988 | // Convert: VADD_SPLAT elt, size |
2989 | // Into: tmp1 = VSPLTIS[BHW] elt-16 |
2990 | // tmp2 = VSPLTIS[BHW] -16 |
2991 | // VSUBU[BHW]M tmp1, tmp2 |
2992 | SDValue EltVal = getI32Imm(Elt - 16, dl); |
2993 | SDNode *Tmp1 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal); |
2994 | EltVal = getI32Imm(-16, dl); |
2995 | SDNode *Tmp2 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal); |
2996 | return CurDAG->getMachineNode(Opc3, dl, VT, SDValue(Tmp1, 0), |
2997 | SDValue(Tmp2, 0)); |
2998 | |
2999 | } else { |
3000 | // Elt is odd and negative, in the range [-31,-17]. |
3001 | // |
3002 | // Convert: VADD_SPLAT elt, size |
3003 | // Into: tmp1 = VSPLTIS[BHW] elt+16 |
3004 | // tmp2 = VSPLTIS[BHW] -16 |
3005 | // VADDU[BHW]M tmp1, tmp2 |
3006 | SDValue EltVal = getI32Imm(Elt + 16, dl); |
3007 | SDNode *Tmp1 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal); |
3008 | EltVal = getI32Imm(-16, dl); |
3009 | SDNode *Tmp2 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal); |
3010 | return CurDAG->getMachineNode(Opc2, dl, VT, SDValue(Tmp1, 0), |
3011 | SDValue(Tmp2, 0)); |
3012 | } |
3013 | } |
3014 | } |
3015 | |
3016 | return SelectCode(N); |
3017 | } |
3018 | |
3019 | // If the target supports the cmpb instruction, do the idiom recognition here. |
3020 | // We don't do this as a DAG combine because we don't want to do it as nodes |
3021 | // are being combined (because we might miss part of the eventual idiom). We |
3022 | // don't want to do it during instruction selection because we want to reuse |
3023 | // the logic for lowering the masking operations already part of the |
3024 | // instruction selector. |
3025 | SDValue PPCDAGToDAGISel::combineToCMPB(SDNode *N) { |
3026 | SDLoc dl(N); |
3027 | |
3028 | assert(N->getOpcode() == ISD::OR &&((N->getOpcode() == ISD::OR && "Only OR nodes are supported for CMPB" ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() == ISD::OR && \"Only OR nodes are supported for CMPB\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 3029, __PRETTY_FUNCTION__)) |
3029 | "Only OR nodes are supported for CMPB")((N->getOpcode() == ISD::OR && "Only OR nodes are supported for CMPB" ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() == ISD::OR && \"Only OR nodes are supported for CMPB\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 3029, __PRETTY_FUNCTION__)); |
3030 | |
3031 | SDValue Res; |
3032 | if (!PPCSubTarget->hasCMPB()) |
3033 | return Res; |
3034 | |
3035 | if (N->getValueType(0) != MVT::i32 && |
3036 | N->getValueType(0) != MVT::i64) |
3037 | return Res; |
3038 | |
3039 | EVT VT = N->getValueType(0); |
3040 | |
3041 | SDValue RHS, LHS; |
3042 | bool BytesFound[8] = { 0, 0, 0, 0, 0, 0, 0, 0 }; |
3043 | uint64_t Mask = 0, Alt = 0; |
3044 | |
3045 | auto IsByteSelectCC = [this](SDValue O, unsigned &b, |
3046 | uint64_t &Mask, uint64_t &Alt, |
3047 | SDValue &LHS, SDValue &RHS) { |
3048 | if (O.getOpcode() != ISD::SELECT_CC) |
3049 | return false; |
3050 | ISD::CondCode CC = cast<CondCodeSDNode>(O.getOperand(4))->get(); |
3051 | |
3052 | if (!isa<ConstantSDNode>(O.getOperand(2)) || |
3053 | !isa<ConstantSDNode>(O.getOperand(3))) |
3054 | return false; |
3055 | |
3056 | uint64_t PM = O.getConstantOperandVal(2); |
3057 | uint64_t PAlt = O.getConstantOperandVal(3); |
3058 | for (b = 0; b < 8; ++b) { |
3059 | uint64_t Mask = UINT64_C(0xFF)0xFFUL << (8*b); |
3060 | if (PM && (PM & Mask) == PM && (PAlt & Mask) == PAlt) |
3061 | break; |
3062 | } |
3063 | |
3064 | if (b == 8) |
3065 | return false; |
3066 | Mask |= PM; |
3067 | Alt |= PAlt; |
3068 | |
3069 | if (!isa<ConstantSDNode>(O.getOperand(1)) || |
3070 | O.getConstantOperandVal(1) != 0) { |
3071 | SDValue Op0 = O.getOperand(0), Op1 = O.getOperand(1); |
3072 | if (Op0.getOpcode() == ISD::TRUNCATE) |
3073 | Op0 = Op0.getOperand(0); |
3074 | if (Op1.getOpcode() == ISD::TRUNCATE) |
3075 | Op1 = Op1.getOperand(0); |
3076 | |
3077 | if (Op0.getOpcode() == ISD::SRL && Op1.getOpcode() == ISD::SRL && |
3078 | Op0.getOperand(1) == Op1.getOperand(1) && CC == ISD::SETEQ && |
3079 | isa<ConstantSDNode>(Op0.getOperand(1))) { |
3080 | |
3081 | unsigned Bits = Op0.getValueType().getSizeInBits(); |
3082 | if (b != Bits/8-1) |
3083 | return false; |
3084 | if (Op0.getConstantOperandVal(1) != Bits-8) |
3085 | return false; |
3086 | |
3087 | LHS = Op0.getOperand(0); |
3088 | RHS = Op1.getOperand(0); |
3089 | return true; |
3090 | } |
3091 | |
3092 | // When we have small integers (i16 to be specific), the form present |
3093 | // post-legalization uses SETULT in the SELECT_CC for the |
3094 | // higher-order byte, depending on the fact that the |
3095 | // even-higher-order bytes are known to all be zero, for example: |
3096 | // select_cc (xor $lhs, $rhs), 256, 65280, 0, setult |
3097 | // (so when the second byte is the same, because all higher-order |
3098 | // bits from bytes 3 and 4 are known to be zero, the result of the |
3099 | // xor can be at most 255) |
3100 | if (Op0.getOpcode() == ISD::XOR && CC == ISD::SETULT && |
3101 | isa<ConstantSDNode>(O.getOperand(1))) { |
3102 | |
3103 | uint64_t ULim = O.getConstantOperandVal(1); |
3104 | if (ULim != (UINT64_C(1)1UL << b*8)) |
3105 | return false; |
3106 | |
3107 | // Now we need to make sure that the upper bytes are known to be |
3108 | // zero. |
3109 | unsigned Bits = Op0.getValueType().getSizeInBits(); |
3110 | if (!CurDAG->MaskedValueIsZero(Op0, |
3111 | APInt::getHighBitsSet(Bits, Bits - (b+1)*8))) |
3112 | return false; |
3113 | |
3114 | LHS = Op0.getOperand(0); |
3115 | RHS = Op0.getOperand(1); |
3116 | return true; |
3117 | } |
3118 | |
3119 | return false; |
3120 | } |
3121 | |
3122 | if (CC != ISD::SETEQ) |
3123 | return false; |
3124 | |
3125 | SDValue Op = O.getOperand(0); |
3126 | if (Op.getOpcode() == ISD::AND) { |
3127 | if (!isa<ConstantSDNode>(Op.getOperand(1))) |
3128 | return false; |
3129 | if (Op.getConstantOperandVal(1) != (UINT64_C(0xFF)0xFFUL << (8*b))) |
3130 | return false; |
3131 | |
3132 | SDValue XOR = Op.getOperand(0); |
3133 | if (XOR.getOpcode() == ISD::TRUNCATE) |
3134 | XOR = XOR.getOperand(0); |
3135 | if (XOR.getOpcode() != ISD::XOR) |
3136 | return false; |
3137 | |
3138 | LHS = XOR.getOperand(0); |
3139 | RHS = XOR.getOperand(1); |
3140 | return true; |
3141 | } else if (Op.getOpcode() == ISD::SRL) { |
3142 | if (!isa<ConstantSDNode>(Op.getOperand(1))) |
3143 | return false; |
3144 | unsigned Bits = Op.getValueType().getSizeInBits(); |
3145 | if (b != Bits/8-1) |
3146 | return false; |
3147 | if (Op.getConstantOperandVal(1) != Bits-8) |
3148 | return false; |
3149 | |
3150 | SDValue XOR = Op.getOperand(0); |
3151 | if (XOR.getOpcode() == ISD::TRUNCATE) |
3152 | XOR = XOR.getOperand(0); |
3153 | if (XOR.getOpcode() != ISD::XOR) |
3154 | return false; |
3155 | |
3156 | LHS = XOR.getOperand(0); |
3157 | RHS = XOR.getOperand(1); |
3158 | return true; |
3159 | } |
3160 | |
3161 | return false; |
3162 | }; |
3163 | |
3164 | SmallVector<SDValue, 8> Queue(1, SDValue(N, 0)); |
3165 | while (!Queue.empty()) { |
3166 | SDValue V = Queue.pop_back_val(); |
3167 | |
3168 | for (const SDValue &O : V.getNode()->ops()) { |
3169 | unsigned b; |
3170 | uint64_t M = 0, A = 0; |
3171 | SDValue OLHS, ORHS; |
3172 | if (O.getOpcode() == ISD::OR) { |
3173 | Queue.push_back(O); |
3174 | } else if (IsByteSelectCC(O, b, M, A, OLHS, ORHS)) { |
3175 | if (!LHS) { |
3176 | LHS = OLHS; |
3177 | RHS = ORHS; |
3178 | BytesFound[b] = true; |
3179 | Mask |= M; |
3180 | Alt |= A; |
3181 | } else if ((LHS == ORHS && RHS == OLHS) || |
3182 | (RHS == ORHS && LHS == OLHS)) { |
3183 | BytesFound[b] = true; |
3184 | Mask |= M; |
3185 | Alt |= A; |
3186 | } else { |
3187 | return Res; |
3188 | } |
3189 | } else { |
3190 | return Res; |
3191 | } |
3192 | } |
3193 | } |
3194 | |
3195 | unsigned LastB = 0, BCnt = 0; |
3196 | for (unsigned i = 0; i < 8; ++i) |
3197 | if (BytesFound[LastB]) { |
3198 | ++BCnt; |
3199 | LastB = i; |
3200 | } |
3201 | |
3202 | if (!LastB || BCnt < 2) |
3203 | return Res; |
3204 | |
3205 | // Because we'll be zero-extending the output anyway if don't have a specific |
3206 | // value for each input byte (via the Mask), we can 'anyext' the inputs. |
3207 | if (LHS.getValueType() != VT) { |
3208 | LHS = CurDAG->getAnyExtOrTrunc(LHS, dl, VT); |
3209 | RHS = CurDAG->getAnyExtOrTrunc(RHS, dl, VT); |
3210 | } |
3211 | |
3212 | Res = CurDAG->getNode(PPCISD::CMPB, dl, VT, LHS, RHS); |
3213 | |
3214 | bool NonTrivialMask = ((int64_t) Mask) != INT64_C(-1)-1L; |
3215 | if (NonTrivialMask && !Alt) { |
3216 | // Res = Mask & CMPB |
3217 | Res = CurDAG->getNode(ISD::AND, dl, VT, Res, |
3218 | CurDAG->getConstant(Mask, dl, VT)); |
3219 | } else if (Alt) { |
3220 | // Res = (CMPB & Mask) | (~CMPB & Alt) |
3221 | // Which, as suggested here: |
3222 | // https://graphics.stanford.edu/~seander/bithacks.html#MaskedMerge |
3223 | // can be written as: |
3224 | // Res = Alt ^ ((Alt ^ Mask) & CMPB) |
3225 | // useful because the (Alt ^ Mask) can be pre-computed. |
3226 | Res = CurDAG->getNode(ISD::AND, dl, VT, Res, |
3227 | CurDAG->getConstant(Mask ^ Alt, dl, VT)); |
3228 | Res = CurDAG->getNode(ISD::XOR, dl, VT, Res, |
3229 | CurDAG->getConstant(Alt, dl, VT)); |
3230 | } |
3231 | |
3232 | return Res; |
3233 | } |
3234 | |
3235 | // When CR bit registers are enabled, an extension of an i1 variable to a i32 |
3236 | // or i64 value is lowered in terms of a SELECT_I[48] operation, and thus |
3237 | // involves constant materialization of a 0 or a 1 or both. If the result of |
3238 | // the extension is then operated upon by some operator that can be constant |
3239 | // folded with a constant 0 or 1, and that constant can be materialized using |
3240 | // only one instruction (like a zero or one), then we should fold in those |
3241 | // operations with the select. |
3242 | void PPCDAGToDAGISel::foldBoolExts(SDValue &Res, SDNode *&N) { |
3243 | if (!PPCSubTarget->useCRBits()) |
3244 | return; |
3245 | |
3246 | if (N->getOpcode() != ISD::ZERO_EXTEND && |
3247 | N->getOpcode() != ISD::SIGN_EXTEND && |
3248 | N->getOpcode() != ISD::ANY_EXTEND) |
3249 | return; |
3250 | |
3251 | if (N->getOperand(0).getValueType() != MVT::i1) |
3252 | return; |
3253 | |
3254 | if (!N->hasOneUse()) |
3255 | return; |
3256 | |
3257 | SDLoc dl(N); |
3258 | EVT VT = N->getValueType(0); |
3259 | SDValue Cond = N->getOperand(0); |
3260 | SDValue ConstTrue = |
3261 | CurDAG->getConstant(N->getOpcode() == ISD::SIGN_EXTEND ? -1 : 1, dl, VT); |
3262 | SDValue ConstFalse = CurDAG->getConstant(0, dl, VT); |
3263 | |
3264 | do { |
3265 | SDNode *User = *N->use_begin(); |
3266 | if (User->getNumOperands() != 2) |
3267 | break; |
3268 | |
3269 | auto TryFold = [this, N, User, dl](SDValue Val) { |
3270 | SDValue UserO0 = User->getOperand(0), UserO1 = User->getOperand(1); |
3271 | SDValue O0 = UserO0.getNode() == N ? Val : UserO0; |
3272 | SDValue O1 = UserO1.getNode() == N ? Val : UserO1; |
3273 | |
3274 | return CurDAG->FoldConstantArithmetic(User->getOpcode(), dl, |
3275 | User->getValueType(0), |
3276 | O0.getNode(), O1.getNode()); |
3277 | }; |
3278 | |
3279 | SDValue TrueRes = TryFold(ConstTrue); |
3280 | if (!TrueRes) |
3281 | break; |
3282 | SDValue FalseRes = TryFold(ConstFalse); |
3283 | if (!FalseRes) |
3284 | break; |
3285 | |
3286 | // For us to materialize these using one instruction, we must be able to |
3287 | // represent them as signed 16-bit integers. |
3288 | uint64_t True = cast<ConstantSDNode>(TrueRes)->getZExtValue(), |
3289 | False = cast<ConstantSDNode>(FalseRes)->getZExtValue(); |
3290 | if (!isInt<16>(True) || !isInt<16>(False)) |
3291 | break; |
3292 | |
3293 | // We can replace User with a new SELECT node, and try again to see if we |
3294 | // can fold the select with its user. |
3295 | Res = CurDAG->getSelect(dl, User->getValueType(0), Cond, TrueRes, FalseRes); |
3296 | N = User; |
3297 | ConstTrue = TrueRes; |
3298 | ConstFalse = FalseRes; |
3299 | } while (N->hasOneUse()); |
3300 | } |
3301 | |
3302 | void PPCDAGToDAGISel::PreprocessISelDAG() { |
3303 | SelectionDAG::allnodes_iterator Position(CurDAG->getRoot().getNode()); |
3304 | ++Position; |
3305 | |
3306 | bool MadeChange = false; |
3307 | while (Position != CurDAG->allnodes_begin()) { |
3308 | SDNode *N = &*--Position; |
3309 | if (N->use_empty()) |
3310 | continue; |
3311 | |
3312 | SDValue Res; |
3313 | switch (N->getOpcode()) { |
3314 | default: break; |
3315 | case ISD::OR: |
3316 | Res = combineToCMPB(N); |
3317 | break; |
3318 | } |
3319 | |
3320 | if (!Res) |
3321 | foldBoolExts(Res, N); |
3322 | |
3323 | if (Res) { |
3324 | DEBUG(dbgs() << "PPC DAG preprocessing replacing:\nOld: ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "PPC DAG preprocessing replacing:\nOld: " ; } } while (0); |
3325 | DEBUG(N->dump(CurDAG))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { N->dump(CurDAG); } } while (0); |
3326 | DEBUG(dbgs() << "\nNew: ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\nNew: "; } } while (0); |
3327 | DEBUG(Res.getNode()->dump(CurDAG))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { Res.getNode()->dump(CurDAG); } } while ( 0); |
3328 | DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\n"; } } while (0); |
3329 | |
3330 | CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), Res); |
3331 | MadeChange = true; |
3332 | } |
3333 | } |
3334 | |
3335 | if (MadeChange) |
3336 | CurDAG->RemoveDeadNodes(); |
3337 | } |
3338 | |
3339 | /// PostprocessISelDAG - Perform some late peephole optimizations |
3340 | /// on the DAG representation. |
3341 | void PPCDAGToDAGISel::PostprocessISelDAG() { |
3342 | |
3343 | // Skip peepholes at -O0. |
3344 | if (TM.getOptLevel() == CodeGenOpt::None) |
3345 | return; |
3346 | |
3347 | PeepholePPC64(); |
3348 | PeepholeCROps(); |
3349 | PeepholePPC64ZExt(); |
3350 | } |
3351 | |
3352 | // Check if all users of this node will become isel where the second operand |
3353 | // is the constant zero. If this is so, and if we can negate the condition, |
3354 | // then we can flip the true and false operands. This will allow the zero to |
3355 | // be folded with the isel so that we don't need to materialize a register |
3356 | // containing zero. |
3357 | bool PPCDAGToDAGISel::AllUsersSelectZero(SDNode *N) { |
3358 | // If we're not using isel, then this does not matter. |
3359 | if (!PPCSubTarget->hasISEL()) |
3360 | return false; |
3361 | |
3362 | for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end(); |
3363 | UI != UE; ++UI) { |
3364 | SDNode *User = *UI; |
3365 | if (!User->isMachineOpcode()) |
3366 | return false; |
3367 | if (User->getMachineOpcode() != PPC::SELECT_I4 && |
3368 | User->getMachineOpcode() != PPC::SELECT_I8) |
3369 | return false; |
3370 | |
3371 | SDNode *Op2 = User->getOperand(2).getNode(); |
3372 | if (!Op2->isMachineOpcode()) |
3373 | return false; |
3374 | |
3375 | if (Op2->getMachineOpcode() != PPC::LI && |
3376 | Op2->getMachineOpcode() != PPC::LI8) |
3377 | return false; |
3378 | |
3379 | ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op2->getOperand(0)); |
3380 | if (!C) |
3381 | return false; |
3382 | |
3383 | if (!C->isNullValue()) |
3384 | return false; |
3385 | } |
3386 | |
3387 | return true; |
3388 | } |
3389 | |
3390 | void PPCDAGToDAGISel::SwapAllSelectUsers(SDNode *N) { |
3391 | SmallVector<SDNode *, 4> ToReplace; |
3392 | for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end(); |
3393 | UI != UE; ++UI) { |
3394 | SDNode *User = *UI; |
3395 | assert((User->getMachineOpcode() == PPC::SELECT_I4 ||(((User->getMachineOpcode() == PPC::SELECT_I4 || User-> getMachineOpcode() == PPC::SELECT_I8) && "Must have all select users" ) ? static_cast<void> (0) : __assert_fail ("(User->getMachineOpcode() == PPC::SELECT_I4 || User->getMachineOpcode() == PPC::SELECT_I8) && \"Must have all select users\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 3397, __PRETTY_FUNCTION__)) |
3396 | User->getMachineOpcode() == PPC::SELECT_I8) &&(((User->getMachineOpcode() == PPC::SELECT_I4 || User-> getMachineOpcode() == PPC::SELECT_I8) && "Must have all select users" ) ? static_cast<void> (0) : __assert_fail ("(User->getMachineOpcode() == PPC::SELECT_I4 || User->getMachineOpcode() == PPC::SELECT_I8) && \"Must have all select users\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 3397, __PRETTY_FUNCTION__)) |
3397 | "Must have all select users")(((User->getMachineOpcode() == PPC::SELECT_I4 || User-> getMachineOpcode() == PPC::SELECT_I8) && "Must have all select users" ) ? static_cast<void> (0) : __assert_fail ("(User->getMachineOpcode() == PPC::SELECT_I4 || User->getMachineOpcode() == PPC::SELECT_I8) && \"Must have all select users\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 3397, __PRETTY_FUNCTION__)); |
3398 | ToReplace.push_back(User); |
3399 | } |
3400 | |
3401 | for (SmallVector<SDNode *, 4>::iterator UI = ToReplace.begin(), |
3402 | UE = ToReplace.end(); UI != UE; ++UI) { |
3403 | SDNode *User = *UI; |
3404 | SDNode *ResNode = |
3405 | CurDAG->getMachineNode(User->getMachineOpcode(), SDLoc(User), |
3406 | User->getValueType(0), User->getOperand(0), |
3407 | User->getOperand(2), |
3408 | User->getOperand(1)); |
3409 | |
3410 | DEBUG(dbgs() << "CR Peephole replacing:\nOld: ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "CR Peephole replacing:\nOld: " ; } } while (0); |
3411 | DEBUG(User->dump(CurDAG))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { User->dump(CurDAG); } } while (0); |
3412 | DEBUG(dbgs() << "\nNew: ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\nNew: "; } } while (0); |
3413 | DEBUG(ResNode->dump(CurDAG))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { ResNode->dump(CurDAG); } } while (0); |
3414 | DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\n"; } } while (0); |
3415 | |
3416 | ReplaceUses(User, ResNode); |
3417 | } |
3418 | } |
3419 | |
3420 | void PPCDAGToDAGISel::PeepholeCROps() { |
3421 | bool IsModified; |
3422 | do { |
3423 | IsModified = false; |
3424 | for (SDNode &Node : CurDAG->allnodes()) { |
3425 | MachineSDNode *MachineNode = dyn_cast<MachineSDNode>(&Node); |
3426 | if (!MachineNode || MachineNode->use_empty()) |
3427 | continue; |
3428 | SDNode *ResNode = MachineNode; |
3429 | |
3430 | bool Op1Set = false, Op1Unset = false, |
3431 | Op1Not = false, |
3432 | Op2Set = false, Op2Unset = false, |
3433 | Op2Not = false; |
3434 | |
3435 | unsigned Opcode = MachineNode->getMachineOpcode(); |
3436 | switch (Opcode) { |
3437 | default: break; |
3438 | case PPC::CRAND: |
3439 | case PPC::CRNAND: |
3440 | case PPC::CROR: |
3441 | case PPC::CRXOR: |
3442 | case PPC::CRNOR: |
3443 | case PPC::CREQV: |
3444 | case PPC::CRANDC: |
3445 | case PPC::CRORC: { |
3446 | SDValue Op = MachineNode->getOperand(1); |
3447 | if (Op.isMachineOpcode()) { |
3448 | if (Op.getMachineOpcode() == PPC::CRSET) |
3449 | Op2Set = true; |
3450 | else if (Op.getMachineOpcode() == PPC::CRUNSET) |
3451 | Op2Unset = true; |
3452 | else if (Op.getMachineOpcode() == PPC::CRNOR && |
3453 | Op.getOperand(0) == Op.getOperand(1)) |
3454 | Op2Not = true; |
3455 | } |
3456 | } // fallthrough |
3457 | case PPC::BC: |
3458 | case PPC::BCn: |
3459 | case PPC::SELECT_I4: |
3460 | case PPC::SELECT_I8: |
3461 | case PPC::SELECT_F4: |
3462 | case PPC::SELECT_F8: |
3463 | case PPC::SELECT_QFRC: |
3464 | case PPC::SELECT_QSRC: |
3465 | case PPC::SELECT_QBRC: |
3466 | case PPC::SELECT_VRRC: |
3467 | case PPC::SELECT_VSFRC: |
3468 | case PPC::SELECT_VSSRC: |
3469 | case PPC::SELECT_VSRC: { |
3470 | SDValue Op = MachineNode->getOperand(0); |
3471 | if (Op.isMachineOpcode()) { |
3472 | if (Op.getMachineOpcode() == PPC::CRSET) |
3473 | Op1Set = true; |
3474 | else if (Op.getMachineOpcode() == PPC::CRUNSET) |
3475 | Op1Unset = true; |
3476 | else if (Op.getMachineOpcode() == PPC::CRNOR && |
3477 | Op.getOperand(0) == Op.getOperand(1)) |
3478 | Op1Not = true; |
3479 | } |
3480 | } |
3481 | break; |
3482 | } |
3483 | |
3484 | bool SelectSwap = false; |
3485 | switch (Opcode) { |
3486 | default: break; |
3487 | case PPC::CRAND: |
3488 | if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) |
3489 | // x & x = x |
3490 | ResNode = MachineNode->getOperand(0).getNode(); |
3491 | else if (Op1Set) |
3492 | // 1 & y = y |
3493 | ResNode = MachineNode->getOperand(1).getNode(); |
3494 | else if (Op2Set) |
3495 | // x & 1 = x |
3496 | ResNode = MachineNode->getOperand(0).getNode(); |
3497 | else if (Op1Unset || Op2Unset) |
3498 | // x & 0 = 0 & y = 0 |
3499 | ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode), |
3500 | MVT::i1); |
3501 | else if (Op1Not) |
3502 | // ~x & y = andc(y, x) |
3503 | ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode), |
3504 | MVT::i1, MachineNode->getOperand(1), |
3505 | MachineNode->getOperand(0). |
3506 | getOperand(0)); |
3507 | else if (Op2Not) |
3508 | // x & ~y = andc(x, y) |
3509 | ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode), |
3510 | MVT::i1, MachineNode->getOperand(0), |
3511 | MachineNode->getOperand(1). |
3512 | getOperand(0)); |
3513 | else if (AllUsersSelectZero(MachineNode)) |
3514 | ResNode = CurDAG->getMachineNode(PPC::CRNAND, SDLoc(MachineNode), |
3515 | MVT::i1, MachineNode->getOperand(0), |
3516 | MachineNode->getOperand(1)), |
3517 | SelectSwap = true; |
3518 | break; |
3519 | case PPC::CRNAND: |
3520 | if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) |
3521 | // nand(x, x) -> nor(x, x) |
3522 | ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), |
3523 | MVT::i1, MachineNode->getOperand(0), |
3524 | MachineNode->getOperand(0)); |
3525 | else if (Op1Set) |
3526 | // nand(1, y) -> nor(y, y) |
3527 | ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), |
3528 | MVT::i1, MachineNode->getOperand(1), |
3529 | MachineNode->getOperand(1)); |
3530 | else if (Op2Set) |
3531 | // nand(x, 1) -> nor(x, x) |
3532 | ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), |
3533 | MVT::i1, MachineNode->getOperand(0), |
3534 | MachineNode->getOperand(0)); |
3535 | else if (Op1Unset || Op2Unset) |
3536 | // nand(x, 0) = nand(0, y) = 1 |
3537 | ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode), |
3538 | MVT::i1); |
3539 | else if (Op1Not) |
3540 | // nand(~x, y) = ~(~x & y) = x | ~y = orc(x, y) |
3541 | ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode), |
3542 | MVT::i1, MachineNode->getOperand(0). |
3543 | getOperand(0), |
3544 | MachineNode->getOperand(1)); |
3545 | else if (Op2Not) |
3546 | // nand(x, ~y) = ~x | y = orc(y, x) |
3547 | ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode), |
3548 | MVT::i1, MachineNode->getOperand(1). |
3549 | getOperand(0), |
3550 | MachineNode->getOperand(0)); |
3551 | else if (AllUsersSelectZero(MachineNode)) |
3552 | ResNode = CurDAG->getMachineNode(PPC::CRAND, SDLoc(MachineNode), |
3553 | MVT::i1, MachineNode->getOperand(0), |
3554 | MachineNode->getOperand(1)), |
3555 | SelectSwap = true; |
3556 | break; |
3557 | case PPC::CROR: |
3558 | if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) |
3559 | // x | x = x |
3560 | ResNode = MachineNode->getOperand(0).getNode(); |
3561 | else if (Op1Set || Op2Set) |
3562 | // x | 1 = 1 | y = 1 |
3563 | ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode), |
3564 | MVT::i1); |
3565 | else if (Op1Unset) |
3566 | // 0 | y = y |
3567 | ResNode = MachineNode->getOperand(1).getNode(); |
3568 | else if (Op2Unset) |
3569 | // x | 0 = x |
3570 | ResNode = MachineNode->getOperand(0).getNode(); |
3571 | else if (Op1Not) |
3572 | // ~x | y = orc(y, x) |
3573 | ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode), |
3574 | MVT::i1, MachineNode->getOperand(1), |
3575 | MachineNode->getOperand(0). |
3576 | getOperand(0)); |
3577 | else if (Op2Not) |
3578 | // x | ~y = orc(x, y) |
3579 | ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode), |
3580 | MVT::i1, MachineNode->getOperand(0), |
3581 | MachineNode->getOperand(1). |
3582 | getOperand(0)); |
3583 | else if (AllUsersSelectZero(MachineNode)) |
3584 | ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), |
3585 | MVT::i1, MachineNode->getOperand(0), |
3586 | MachineNode->getOperand(1)), |
3587 | SelectSwap = true; |
3588 | break; |
3589 | case PPC::CRXOR: |
3590 | if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) |
3591 | // xor(x, x) = 0 |
3592 | ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode), |
3593 | MVT::i1); |
3594 | else if (Op1Set) |
3595 | // xor(1, y) -> nor(y, y) |
3596 | ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), |
3597 | MVT::i1, MachineNode->getOperand(1), |
3598 | MachineNode->getOperand(1)); |
3599 | else if (Op2Set) |
3600 | // xor(x, 1) -> nor(x, x) |
3601 | ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), |
3602 | MVT::i1, MachineNode->getOperand(0), |
3603 | MachineNode->getOperand(0)); |
3604 | else if (Op1Unset) |
3605 | // xor(0, y) = y |
3606 | ResNode = MachineNode->getOperand(1).getNode(); |
3607 | else if (Op2Unset) |
3608 | // xor(x, 0) = x |
3609 | ResNode = MachineNode->getOperand(0).getNode(); |
3610 | else if (Op1Not) |
3611 | // xor(~x, y) = eqv(x, y) |
3612 | ResNode = CurDAG->getMachineNode(PPC::CREQV, SDLoc(MachineNode), |
3613 | MVT::i1, MachineNode->getOperand(0). |
3614 | getOperand(0), |
3615 | MachineNode->getOperand(1)); |
3616 | else if (Op2Not) |
3617 | // xor(x, ~y) = eqv(x, y) |
3618 | ResNode = CurDAG->getMachineNode(PPC::CREQV, SDLoc(MachineNode), |
3619 | MVT::i1, MachineNode->getOperand(0), |
3620 | MachineNode->getOperand(1). |
3621 | getOperand(0)); |
3622 | else if (AllUsersSelectZero(MachineNode)) |
3623 | ResNode = CurDAG->getMachineNode(PPC::CREQV, SDLoc(MachineNode), |
3624 | MVT::i1, MachineNode->getOperand(0), |
3625 | MachineNode->getOperand(1)), |
3626 | SelectSwap = true; |
3627 | break; |
3628 | case PPC::CRNOR: |
3629 | if (Op1Set || Op2Set) |
3630 | // nor(1, y) -> 0 |
3631 | ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode), |
3632 | MVT::i1); |
3633 | else if (Op1Unset) |
3634 | // nor(0, y) = ~y -> nor(y, y) |
3635 | ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), |
3636 | MVT::i1, MachineNode->getOperand(1), |
3637 | MachineNode->getOperand(1)); |
3638 | else if (Op2Unset) |
3639 | // nor(x, 0) = ~x |
3640 | ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), |
3641 | MVT::i1, MachineNode->getOperand(0), |
3642 | MachineNode->getOperand(0)); |
3643 | else if (Op1Not) |
3644 | // nor(~x, y) = andc(x, y) |
3645 | ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode), |
3646 | MVT::i1, MachineNode->getOperand(0). |
3647 | getOperand(0), |
3648 | MachineNode->getOperand(1)); |
3649 | else if (Op2Not) |
3650 | // nor(x, ~y) = andc(y, x) |
3651 | ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode), |
3652 | MVT::i1, MachineNode->getOperand(1). |
3653 | getOperand(0), |
3654 | MachineNode->getOperand(0)); |
3655 | else if (AllUsersSelectZero(MachineNode)) |
3656 | ResNode = CurDAG->getMachineNode(PPC::CROR, SDLoc(MachineNode), |
3657 | MVT::i1, MachineNode->getOperand(0), |
3658 | MachineNode->getOperand(1)), |
3659 | SelectSwap = true; |
3660 | break; |
3661 | case PPC::CREQV: |
3662 | if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) |
3663 | // eqv(x, x) = 1 |
3664 | ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode), |
3665 | MVT::i1); |
3666 | else if (Op1Set) |
3667 | // eqv(1, y) = y |
3668 | ResNode = MachineNode->getOperand(1).getNode(); |
3669 | else if (Op2Set) |
3670 | // eqv(x, 1) = x |
3671 | ResNode = MachineNode->getOperand(0).getNode(); |
3672 | else if (Op1Unset) |
3673 | // eqv(0, y) = ~y -> nor(y, y) |
3674 | ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), |
3675 | MVT::i1, MachineNode->getOperand(1), |
3676 | MachineNode->getOperand(1)); |
3677 | else if (Op2Unset) |
3678 | // eqv(x, 0) = ~x |
3679 | ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), |
3680 | MVT::i1, MachineNode->getOperand(0), |
3681 | MachineNode->getOperand(0)); |
3682 | else if (Op1Not) |
3683 | // eqv(~x, y) = xor(x, y) |
3684 | ResNode = CurDAG->getMachineNode(PPC::CRXOR, SDLoc(MachineNode), |
3685 | MVT::i1, MachineNode->getOperand(0). |
3686 | getOperand(0), |
3687 | MachineNode->getOperand(1)); |
3688 | else if (Op2Not) |
3689 | // eqv(x, ~y) = xor(x, y) |
3690 | ResNode = CurDAG->getMachineNode(PPC::CRXOR, SDLoc(MachineNode), |
3691 | MVT::i1, MachineNode->getOperand(0), |
3692 | MachineNode->getOperand(1). |
3693 | getOperand(0)); |
3694 | else if (AllUsersSelectZero(MachineNode)) |
3695 | ResNode = CurDAG->getMachineNode(PPC::CRXOR, SDLoc(MachineNode), |
3696 | MVT::i1, MachineNode->getOperand(0), |
3697 | MachineNode->getOperand(1)), |
3698 | SelectSwap = true; |
3699 | break; |
3700 | case PPC::CRANDC: |
3701 | if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) |
3702 | // andc(x, x) = 0 |
3703 | ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode), |
3704 | MVT::i1); |
3705 | else if (Op1Set) |
3706 | // andc(1, y) = ~y |
3707 | ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), |
3708 | MVT::i1, MachineNode->getOperand(1), |
3709 | MachineNode->getOperand(1)); |
3710 | else if (Op1Unset || Op2Set) |
3711 | // andc(0, y) = andc(x, 1) = 0 |
3712 | ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode), |
3713 | MVT::i1); |
3714 | else if (Op2Unset) |
3715 | // andc(x, 0) = x |
3716 | ResNode = MachineNode->getOperand(0).getNode(); |
3717 | else if (Op1Not) |
3718 | // andc(~x, y) = ~(x | y) = nor(x, y) |
3719 | ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), |
3720 | MVT::i1, MachineNode->getOperand(0). |
3721 | getOperand(0), |
3722 | MachineNode->getOperand(1)); |
3723 | else if (Op2Not) |
3724 | // andc(x, ~y) = x & y |
3725 | ResNode = CurDAG->getMachineNode(PPC::CRAND, SDLoc(MachineNode), |
3726 | MVT::i1, MachineNode->getOperand(0), |
3727 | MachineNode->getOperand(1). |
3728 | getOperand(0)); |
3729 | else if (AllUsersSelectZero(MachineNode)) |
3730 | ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode), |
3731 | MVT::i1, MachineNode->getOperand(1), |
3732 | MachineNode->getOperand(0)), |
3733 | SelectSwap = true; |
3734 | break; |
3735 | case PPC::CRORC: |
3736 | if (MachineNode->getOperand(0) == MachineNode->getOperand(1)) |
3737 | // orc(x, x) = 1 |
3738 | ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode), |
3739 | MVT::i1); |
3740 | else if (Op1Set || Op2Unset) |
3741 | // orc(1, y) = orc(x, 0) = 1 |
3742 | ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode), |
3743 | MVT::i1); |
3744 | else if (Op2Set) |
3745 | // orc(x, 1) = x |
3746 | ResNode = MachineNode->getOperand(0).getNode(); |
3747 | else if (Op1Unset) |
3748 | // orc(0, y) = ~y |
3749 | ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode), |
3750 | MVT::i1, MachineNode->getOperand(1), |
3751 | MachineNode->getOperand(1)); |
3752 | else if (Op1Not) |
3753 | // orc(~x, y) = ~(x & y) = nand(x, y) |
3754 | ResNode = CurDAG->getMachineNode(PPC::CRNAND, SDLoc(MachineNode), |
3755 | MVT::i1, MachineNode->getOperand(0). |
3756 | getOperand(0), |
3757 | MachineNode->getOperand(1)); |
3758 | else if (Op2Not) |
3759 | // orc(x, ~y) = x | y |
3760 | ResNode = CurDAG->getMachineNode(PPC::CROR, SDLoc(MachineNode), |
3761 | MVT::i1, MachineNode->getOperand(0), |
3762 | MachineNode->getOperand(1). |
3763 | getOperand(0)); |
3764 | else if (AllUsersSelectZero(MachineNode)) |
3765 | ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode), |
3766 | MVT::i1, MachineNode->getOperand(1), |
3767 | MachineNode->getOperand(0)), |
3768 | SelectSwap = true; |
3769 | break; |
3770 | case PPC::SELECT_I4: |
3771 | case PPC::SELECT_I8: |
3772 | case PPC::SELECT_F4: |
3773 | case PPC::SELECT_F8: |
3774 | case PPC::SELECT_QFRC: |
3775 | case PPC::SELECT_QSRC: |
3776 | case PPC::SELECT_QBRC: |
3777 | case PPC::SELECT_VRRC: |
3778 | case PPC::SELECT_VSFRC: |
3779 | case PPC::SELECT_VSSRC: |
3780 | case PPC::SELECT_VSRC: |
3781 | if (Op1Set) |
3782 | ResNode = MachineNode->getOperand(1).getNode(); |
3783 | else if (Op1Unset) |
3784 | ResNode = MachineNode->getOperand(2).getNode(); |
3785 | else if (Op1Not) |
3786 | ResNode = CurDAG->getMachineNode(MachineNode->getMachineOpcode(), |
3787 | SDLoc(MachineNode), |
3788 | MachineNode->getValueType(0), |
3789 | MachineNode->getOperand(0). |
3790 | getOperand(0), |
3791 | MachineNode->getOperand(2), |
3792 | MachineNode->getOperand(1)); |
3793 | break; |
3794 | case PPC::BC: |
3795 | case PPC::BCn: |
3796 | if (Op1Not) |
3797 | ResNode = CurDAG->getMachineNode(Opcode == PPC::BC ? PPC::BCn : |
3798 | PPC::BC, |
3799 | SDLoc(MachineNode), |
3800 | MVT::Other, |
3801 | MachineNode->getOperand(0). |
3802 | getOperand(0), |
3803 | MachineNode->getOperand(1), |
3804 | MachineNode->getOperand(2)); |
3805 | // FIXME: Handle Op1Set, Op1Unset here too. |
3806 | break; |
3807 | } |
3808 | |
3809 | // If we're inverting this node because it is used only by selects that |
3810 | // we'd like to swap, then swap the selects before the node replacement. |
3811 | if (SelectSwap) |
3812 | SwapAllSelectUsers(MachineNode); |
3813 | |
3814 | if (ResNode != MachineNode) { |
3815 | DEBUG(dbgs() << "CR Peephole replacing:\nOld: ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "CR Peephole replacing:\nOld: " ; } } while (0); |
3816 | DEBUG(MachineNode->dump(CurDAG))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { MachineNode->dump(CurDAG); } } while (0 ); |
3817 | DEBUG(dbgs() << "\nNew: ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\nNew: "; } } while (0); |
3818 | DEBUG(ResNode->dump(CurDAG))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { ResNode->dump(CurDAG); } } while (0); |
3819 | DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\n"; } } while (0); |
3820 | |
3821 | ReplaceUses(MachineNode, ResNode); |
3822 | IsModified = true; |
3823 | } |
3824 | } |
3825 | if (IsModified) |
3826 | CurDAG->RemoveDeadNodes(); |
3827 | } while (IsModified); |
3828 | } |
3829 | |
3830 | // Gather the set of 32-bit operations that are known to have their |
3831 | // higher-order 32 bits zero, where ToPromote contains all such operations. |
3832 | static bool PeepholePPC64ZExtGather(SDValue Op32, |
3833 | SmallPtrSetImpl<SDNode *> &ToPromote) { |
3834 | if (!Op32.isMachineOpcode()) |
3835 | return false; |
3836 | |
3837 | // First, check for the "frontier" instructions (those that will clear the |
3838 | // higher-order 32 bits. |
3839 | |
3840 | // For RLWINM and RLWNM, we need to make sure that the mask does not wrap |
3841 | // around. If it does not, then these instructions will clear the |
3842 | // higher-order bits. |
3843 | if ((Op32.getMachineOpcode() == PPC::RLWINM || |
3844 | Op32.getMachineOpcode() == PPC::RLWNM) && |
3845 | Op32.getConstantOperandVal(2) <= Op32.getConstantOperandVal(3)) { |
3846 | ToPromote.insert(Op32.getNode()); |
3847 | return true; |
3848 | } |
3849 | |
3850 | // SLW and SRW always clear the higher-order bits. |
3851 | if (Op32.getMachineOpcode() == PPC::SLW || |
3852 | Op32.getMachineOpcode() == PPC::SRW) { |
3853 | ToPromote.insert(Op32.getNode()); |
3854 | return true; |
3855 | } |
3856 | |
3857 | // For LI and LIS, we need the immediate to be positive (so that it is not |
3858 | // sign extended). |
3859 | if (Op32.getMachineOpcode() == PPC::LI || |
3860 | Op32.getMachineOpcode() == PPC::LIS) { |
3861 | if (!isUInt<15>(Op32.getConstantOperandVal(0))) |
3862 | return false; |
3863 | |
3864 | ToPromote.insert(Op32.getNode()); |
3865 | return true; |
3866 | } |
3867 | |
3868 | // LHBRX and LWBRX always clear the higher-order bits. |
3869 | if (Op32.getMachineOpcode() == PPC::LHBRX || |
3870 | Op32.getMachineOpcode() == PPC::LWBRX) { |
3871 | ToPromote.insert(Op32.getNode()); |
3872 | return true; |
3873 | } |
3874 | |
3875 | // CNTLZW always produces a 64-bit value in [0,32], and so is zero extended. |
3876 | if (Op32.getMachineOpcode() == PPC::CNTLZW) { |
3877 | ToPromote.insert(Op32.getNode()); |
3878 | return true; |
3879 | } |
3880 | |
3881 | // Next, check for those instructions we can look through. |
3882 | |
3883 | // Assuming the mask does not wrap around, then the higher-order bits are |
3884 | // taken directly from the first operand. |
3885 | if (Op32.getMachineOpcode() == PPC::RLWIMI && |
3886 | Op32.getConstantOperandVal(3) <= Op32.getConstantOperandVal(4)) { |
3887 | SmallPtrSet<SDNode *, 16> ToPromote1; |
3888 | if (!PeepholePPC64ZExtGather(Op32.getOperand(0), ToPromote1)) |
3889 | return false; |
3890 | |
3891 | ToPromote.insert(Op32.getNode()); |
3892 | ToPromote.insert(ToPromote1.begin(), ToPromote1.end()); |
3893 | return true; |
3894 | } |
3895 | |
3896 | // For OR, the higher-order bits are zero if that is true for both operands. |
3897 | // For SELECT_I4, the same is true (but the relevant operand numbers are |
3898 | // shifted by 1). |
3899 | if (Op32.getMachineOpcode() == PPC::OR || |
3900 | Op32.getMachineOpcode() == PPC::SELECT_I4) { |
3901 | unsigned B = Op32.getMachineOpcode() == PPC::SELECT_I4 ? 1 : 0; |
3902 | SmallPtrSet<SDNode *, 16> ToPromote1; |
3903 | if (!PeepholePPC64ZExtGather(Op32.getOperand(B+0), ToPromote1)) |
3904 | return false; |
3905 | if (!PeepholePPC64ZExtGather(Op32.getOperand(B+1), ToPromote1)) |
3906 | return false; |
3907 | |
3908 | ToPromote.insert(Op32.getNode()); |
3909 | ToPromote.insert(ToPromote1.begin(), ToPromote1.end()); |
3910 | return true; |
3911 | } |
3912 | |
3913 | // For ORI and ORIS, we need the higher-order bits of the first operand to be |
3914 | // zero, and also for the constant to be positive (so that it is not sign |
3915 | // extended). |
3916 | if (Op32.getMachineOpcode() == PPC::ORI || |
3917 | Op32.getMachineOpcode() == PPC::ORIS) { |
3918 | SmallPtrSet<SDNode *, 16> ToPromote1; |
3919 | if (!PeepholePPC64ZExtGather(Op32.getOperand(0), ToPromote1)) |
3920 | return false; |
3921 | if (!isUInt<15>(Op32.getConstantOperandVal(1))) |
3922 | return false; |
3923 | |
3924 | ToPromote.insert(Op32.getNode()); |
3925 | ToPromote.insert(ToPromote1.begin(), ToPromote1.end()); |
3926 | return true; |
3927 | } |
3928 | |
3929 | // The higher-order bits of AND are zero if that is true for at least one of |
3930 | // the operands. |
3931 | if (Op32.getMachineOpcode() == PPC::AND) { |
3932 | SmallPtrSet<SDNode *, 16> ToPromote1, ToPromote2; |
3933 | bool Op0OK = |
3934 | PeepholePPC64ZExtGather(Op32.getOperand(0), ToPromote1); |
3935 | bool Op1OK = |
3936 | PeepholePPC64ZExtGather(Op32.getOperand(1), ToPromote2); |
3937 | if (!Op0OK && !Op1OK) |
3938 | return false; |
3939 | |
3940 | ToPromote.insert(Op32.getNode()); |
3941 | |
3942 | if (Op0OK) |
3943 | ToPromote.insert(ToPromote1.begin(), ToPromote1.end()); |
3944 | |
3945 | if (Op1OK) |
3946 | ToPromote.insert(ToPromote2.begin(), ToPromote2.end()); |
3947 | |
3948 | return true; |
3949 | } |
3950 | |
3951 | // For ANDI and ANDIS, the higher-order bits are zero if either that is true |
3952 | // of the first operand, or if the second operand is positive (so that it is |
3953 | // not sign extended). |
3954 | if (Op32.getMachineOpcode() == PPC::ANDIo || |
3955 | Op32.getMachineOpcode() == PPC::ANDISo) { |
3956 | SmallPtrSet<SDNode *, 16> ToPromote1; |
3957 | bool Op0OK = |
3958 | PeepholePPC64ZExtGather(Op32.getOperand(0), ToPromote1); |
3959 | bool Op1OK = isUInt<15>(Op32.getConstantOperandVal(1)); |
3960 | if (!Op0OK && !Op1OK) |
3961 | return false; |
3962 | |
3963 | ToPromote.insert(Op32.getNode()); |
3964 | |
3965 | if (Op0OK) |
3966 | ToPromote.insert(ToPromote1.begin(), ToPromote1.end()); |
3967 | |
3968 | return true; |
3969 | } |
3970 | |
3971 | return false; |
3972 | } |
3973 | |
3974 | void PPCDAGToDAGISel::PeepholePPC64ZExt() { |
3975 | if (!PPCSubTarget->isPPC64()) |
3976 | return; |
3977 | |
3978 | // When we zero-extend from i32 to i64, we use a pattern like this: |
3979 | // def : Pat<(i64 (zext i32:$in)), |
3980 | // (RLDICL (INSERT_SUBREG (i64 (IMPLICIT_DEF)), $in, sub_32), |
3981 | // 0, 32)>; |
3982 | // There are several 32-bit shift/rotate instructions, however, that will |
3983 | // clear the higher-order bits of their output, rendering the RLDICL |
3984 | // unnecessary. When that happens, we remove it here, and redefine the |
3985 | // relevant 32-bit operation to be a 64-bit operation. |
3986 | |
3987 | SelectionDAG::allnodes_iterator Position(CurDAG->getRoot().getNode()); |
3988 | ++Position; |
3989 | |
3990 | bool MadeChange = false; |
3991 | while (Position != CurDAG->allnodes_begin()) { |
3992 | SDNode *N = &*--Position; |
3993 | // Skip dead nodes and any non-machine opcodes. |
3994 | if (N->use_empty() || !N->isMachineOpcode()) |
3995 | continue; |
3996 | |
3997 | if (N->getMachineOpcode() != PPC::RLDICL) |
3998 | continue; |
3999 | |
4000 | if (N->getConstantOperandVal(1) != 0 || |
4001 | N->getConstantOperandVal(2) != 32) |
4002 | continue; |
4003 | |
4004 | SDValue ISR = N->getOperand(0); |
4005 | if (!ISR.isMachineOpcode() || |
4006 | ISR.getMachineOpcode() != TargetOpcode::INSERT_SUBREG) |
4007 | continue; |
4008 | |
4009 | if (!ISR.hasOneUse()) |
4010 | continue; |
4011 | |
4012 | if (ISR.getConstantOperandVal(2) != PPC::sub_32) |
4013 | continue; |
4014 | |
4015 | SDValue IDef = ISR.getOperand(0); |
4016 | if (!IDef.isMachineOpcode() || |
4017 | IDef.getMachineOpcode() != TargetOpcode::IMPLICIT_DEF) |
4018 | continue; |
4019 | |
4020 | // We now know that we're looking at a canonical i32 -> i64 zext. See if we |
4021 | // can get rid of it. |
4022 | |
4023 | SDValue Op32 = ISR->getOperand(1); |
4024 | if (!Op32.isMachineOpcode()) |
4025 | continue; |
4026 | |
4027 | // There are some 32-bit instructions that always clear the high-order 32 |
4028 | // bits, there are also some instructions (like AND) that we can look |
4029 | // through. |
4030 | SmallPtrSet<SDNode *, 16> ToPromote; |
4031 | if (!PeepholePPC64ZExtGather(Op32, ToPromote)) |
4032 | continue; |
4033 | |
4034 | // If the ToPromote set contains nodes that have uses outside of the set |
4035 | // (except for the original INSERT_SUBREG), then abort the transformation. |
4036 | bool OutsideUse = false; |
4037 | for (SDNode *PN : ToPromote) { |
4038 | for (SDNode *UN : PN->uses()) { |
4039 | if (!ToPromote.count(UN) && UN != ISR.getNode()) { |
4040 | OutsideUse = true; |
4041 | break; |
4042 | } |
4043 | } |
4044 | |
4045 | if (OutsideUse) |
4046 | break; |
4047 | } |
4048 | if (OutsideUse) |
4049 | continue; |
4050 | |
4051 | MadeChange = true; |
4052 | |
4053 | // We now know that this zero extension can be removed by promoting to |
4054 | // nodes in ToPromote to 64-bit operations, where for operations in the |
4055 | // frontier of the set, we need to insert INSERT_SUBREGs for their |
4056 | // operands. |
4057 | for (SDNode *PN : ToPromote) { |
4058 | unsigned NewOpcode; |
4059 | switch (PN->getMachineOpcode()) { |
4060 | default: |
4061 | llvm_unreachable("Don't know the 64-bit variant of this instruction")::llvm::llvm_unreachable_internal("Don't know the 64-bit variant of this instruction" , "/tmp/buildd/llvm-toolchain-snapshot-3.8~svn251506/lib/Target/PowerPC/PPCISelDAGToDAG.cpp" , 4061); |
4062 | case PPC::RLWINM: NewOpcode = PPC::RLWINM8; break; |
4063 | case PPC::RLWNM: NewOpcode = PPC::RLWNM8; break; |
4064 | case PPC::SLW: NewOpcode = PPC::SLW8; break; |
4065 | case PPC::SRW: NewOpcode = PPC::SRW8; break; |
4066 | case PPC::LI: NewOpcode = PPC::LI8; break; |
4067 | case PPC::LIS: NewOpcode = PPC::LIS8; break; |
4068 | case PPC::LHBRX: NewOpcode = PPC::LHBRX8; break; |
4069 | case PPC::LWBRX: NewOpcode = PPC::LWBRX8; break; |
4070 | case PPC::CNTLZW: NewOpcode = PPC::CNTLZW8; break; |
4071 | case PPC::RLWIMI: NewOpcode = PPC::RLWIMI8; break; |
4072 | case PPC::OR: NewOpcode = PPC::OR8; break; |
4073 | case PPC::SELECT_I4: NewOpcode = PPC::SELECT_I8; break; |
4074 | case PPC::ORI: NewOpcode = PPC::ORI8; break; |
4075 | case PPC::ORIS: NewOpcode = PPC::ORIS8; break; |
4076 | case PPC::AND: NewOpcode = PPC::AND8; break; |
4077 | case PPC::ANDIo: NewOpcode = PPC::ANDIo8; break; |
4078 | case PPC::ANDISo: NewOpcode = PPC::ANDISo8; break; |
4079 | } |
4080 | |
4081 | // Note: During the replacement process, the nodes will be in an |
4082 | // inconsistent state (some instructions will have operands with values |
4083 | // of the wrong type). Once done, however, everything should be right |
4084 | // again. |
4085 | |
4086 | SmallVector<SDValue, 4> Ops; |
4087 | for (const SDValue &V : PN->ops()) { |
4088 | if (!ToPromote.count(V.getNode()) && V.getValueType() == MVT::i32 && |
4089 | !isa<ConstantSDNode>(V)) { |
4090 | SDValue ReplOpOps[] = { ISR.getOperand(0), V, ISR.getOperand(2) }; |
4091 | SDNode *ReplOp = |
4092 | CurDAG->getMachineNode(TargetOpcode::INSERT_SUBREG, SDLoc(V), |
4093 | ISR.getNode()->getVTList(), ReplOpOps); |
4094 | Ops.push_back(SDValue(ReplOp, 0)); |
4095 | } else { |
4096 | Ops.push_back(V); |
4097 | } |
4098 | } |
4099 | |
4100 | // Because all to-be-promoted nodes only have users that are other |
4101 | // promoted nodes (or the original INSERT_SUBREG), we can safely replace |
4102 | // the i32 result value type with i64. |
4103 | |
4104 | SmallVector<EVT, 2> NewVTs; |
4105 | SDVTList VTs = PN->getVTList(); |
4106 | for (unsigned i = 0, ie = VTs.NumVTs; i != ie; ++i) |
4107 | if (VTs.VTs[i] == MVT::i32) |
4108 | NewVTs.push_back(MVT::i64); |
4109 | else |
4110 | NewVTs.push_back(VTs.VTs[i]); |
4111 | |
4112 | DEBUG(dbgs() << "PPC64 ZExt Peephole morphing:\nOld: ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "PPC64 ZExt Peephole morphing:\nOld: " ; } } while (0); |
4113 | DEBUG(PN->dump(CurDAG))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { PN->dump(CurDAG); } } while (0); |
4114 | |
4115 | CurDAG->SelectNodeTo(PN, NewOpcode, CurDAG->getVTList(NewVTs), Ops); |
4116 | |
4117 | DEBUG(dbgs() << "\nNew: ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\nNew: "; } } while (0); |
4118 | DEBUG(PN->dump(CurDAG))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { PN->dump(CurDAG); } } while (0); |
4119 | DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\n"; } } while (0); |
4120 | } |
4121 | |
4122 | // Now we replace the original zero extend and its associated INSERT_SUBREG |
4123 | // with the value feeding the INSERT_SUBREG (which has now been promoted to |
4124 | // return an i64). |
4125 | |
4126 | DEBUG(dbgs() << "PPC64 ZExt Peephole replacing:\nOld: ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "PPC64 ZExt Peephole replacing:\nOld: " ; } } while (0); |
4127 | DEBUG(N->dump(CurDAG))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { N->dump(CurDAG); } } while (0); |
4128 | DEBUG(dbgs() << "\nNew: ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\nNew: "; } } while (0); |
4129 | DEBUG(Op32.getNode()->dump(CurDAG))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { Op32.getNode()->dump(CurDAG); } } while (0); |
4130 | DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\n"; } } while (0); |
4131 | |
4132 | ReplaceUses(N, Op32.getNode()); |
4133 | } |
4134 | |
4135 | if (MadeChange) |
4136 | CurDAG->RemoveDeadNodes(); |
4137 | } |
4138 | |
4139 | void PPCDAGToDAGISel::PeepholePPC64() { |
4140 | // These optimizations are currently supported only for 64-bit SVR4. |
4141 | if (PPCSubTarget->isDarwin() || !PPCSubTarget->isPPC64()) |
4142 | return; |
4143 | |
4144 | SelectionDAG::allnodes_iterator Position(CurDAG->getRoot().getNode()); |
4145 | ++Position; |
4146 | |
4147 | while (Position != CurDAG->allnodes_begin()) { |
4148 | SDNode *N = &*--Position; |
4149 | // Skip dead nodes and any non-machine opcodes. |
4150 | if (N->use_empty() || !N->isMachineOpcode()) |
4151 | continue; |
4152 | |
4153 | unsigned FirstOp; |
4154 | unsigned StorageOpcode = N->getMachineOpcode(); |
4155 | |
4156 | switch (StorageOpcode) { |
4157 | default: continue; |
4158 | |
4159 | case PPC::LBZ: |
4160 | case PPC::LBZ8: |
4161 | case PPC::LD: |
4162 | case PPC::LFD: |
4163 | case PPC::LFS: |
4164 | case PPC::LHA: |
4165 | case PPC::LHA8: |
4166 | case PPC::LHZ: |
4167 | case PPC::LHZ8: |
4168 | case PPC::LWA: |
4169 | case PPC::LWZ: |
4170 | case PPC::LWZ8: |
4171 | FirstOp = 0; |
4172 | break; |
4173 | |
4174 | case PPC::STB: |
4175 | case PPC::STB8: |
4176 | case PPC::STD: |
4177 | case PPC::STFD: |
4178 | case PPC::STFS: |
4179 | case PPC::STH: |
4180 | case PPC::STH8: |
4181 | case PPC::STW: |
4182 | case PPC::STW8: |
4183 | FirstOp = 1; |
4184 | break; |
4185 | } |
4186 | |
4187 | // If this is a load or store with a zero offset, we may be able to |
4188 | // fold an add-immediate into the memory operation. |
4189 | if (!isa<ConstantSDNode>(N->getOperand(FirstOp)) || |
4190 | N->getConstantOperandVal(FirstOp) != 0) |
4191 | continue; |
4192 | |
4193 | SDValue Base = N->getOperand(FirstOp + 1); |
4194 | if (!Base.isMachineOpcode()) |
4195 | continue; |
4196 | |
4197 | unsigned Flags = 0; |
4198 | bool ReplaceFlags = true; |
4199 | |
4200 | // When the feeding operation is an add-immediate of some sort, |
4201 | // determine whether we need to add relocation information to the |
4202 | // target flags on the immediate operand when we fold it into the |
4203 | // load instruction. |
4204 | // |
4205 | // For something like ADDItocL, the relocation information is |
4206 | // inferred from the opcode; when we process it in the AsmPrinter, |
4207 | // we add the necessary relocation there. A load, though, can receive |
4208 | // relocation from various flavors of ADDIxxx, so we need to carry |
4209 | // the relocation information in the target flags. |
4210 | switch (Base.getMachineOpcode()) { |
4211 | default: continue; |
4212 | |
4213 | case PPC::ADDI8: |
4214 | case PPC::ADDI: |
4215 | // In some cases (such as TLS) the relocation information |
4216 | // is already in place on the operand, so copying the operand |
4217 | // is sufficient. |
4218 | ReplaceFlags = false; |
4219 | // For these cases, the immediate may not be divisible by 4, in |
4220 | // which case the fold is illegal for DS-form instructions. (The |
4221 | // other cases provide aligned addresses and are always safe.) |
4222 | if ((StorageOpcode == PPC::LWA || |
4223 | StorageOpcode == PPC::LD || |
4224 | StorageOpcode == PPC::STD) && |
4225 | (!isa<ConstantSDNode>(Base.getOperand(1)) || |
4226 | Base.getConstantOperandVal(1) % 4 != 0)) |
4227 | continue; |
4228 | break; |
4229 | case PPC::ADDIdtprelL: |
4230 | Flags = PPCII::MO_DTPREL_LO; |
4231 | break; |
4232 | case PPC::ADDItlsldL: |
4233 | Flags = PPCII::MO_TLSLD_LO; |
4234 | break; |
4235 | case PPC::ADDItocL: |
4236 | Flags = PPCII::MO_TOC_LO; |
4237 | break; |
4238 | } |
4239 | |
4240 | // We found an opportunity. Reverse the operands from the add |
4241 | // immediate and substitute them into the load or store. If |
4242 | // needed, update the target flags for the immediate operand to |
4243 | // reflect the necessary relocation information. |
4244 | DEBUG(dbgs() << "Folding add-immediate into mem-op:\nBase: ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "Folding add-immediate into mem-op:\nBase: " ; } } while (0); |
4245 | DEBUG(Base->dump(CurDAG))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { Base->dump(CurDAG); } } while (0); |
4246 | DEBUG(dbgs() << "\nN: ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\nN: "; } } while (0); |
4247 | DEBUG(N->dump(CurDAG))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { N->dump(CurDAG); } } while (0); |
4248 | DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "\n"; } } while (0); |
4249 | |
4250 | SDValue ImmOpnd = Base.getOperand(1); |
4251 | |
4252 | // If the relocation information isn't already present on the |
4253 | // immediate operand, add it now. |
4254 | if (ReplaceFlags) { |
4255 | if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(ImmOpnd)) { |
4256 | SDLoc dl(GA); |
4257 | const GlobalValue *GV = GA->getGlobal(); |
4258 | // We can't perform this optimization for data whose alignment |
4259 | // is insufficient for the instruction encoding. |
4260 | if (GV->getAlignment() < 4 && |
4261 | (StorageOpcode == PPC::LD || StorageOpcode == PPC::STD || |
4262 | StorageOpcode == PPC::LWA)) { |
4263 | DEBUG(dbgs() << "Rejected this candidate for alignment.\n\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-codegen")) { dbgs() << "Rejected this candidate for alignment.\n\n" ; } } while (0); |
4264 | continue; |
4265 | } |
4266 | ImmOpnd = CurDAG->getTargetGlobalAddress(GV, dl, MVT::i64, 0, Flags); |
4267 | } else if (ConstantPoolSDNode *CP = |
4268 | dyn_cast<ConstantPoolSDNode>(ImmOpnd)) { |
4269 | const Constant *C = CP->getConstVal(); |
4270 | ImmOpnd = CurDAG->getTargetConstantPool(C, MVT::i64, |
4271 | CP->getAlignment(), |
4272 | 0, Flags); |
4273 | } |
4274 | } |
4275 | |
4276 | if (FirstOp == 1) // Store |
4277 | (void)CurDAG->UpdateNodeOperands(N, N->getOperand(0), ImmOpnd, |
4278 | Base.getOperand(0), N->getOperand(3)); |
4279 | else // Load |
4280 | (void)CurDAG->UpdateNodeOperands(N, ImmOpnd, Base.getOperand(0), |
4281 | N->getOperand(2)); |
4282 | |
4283 | // The add-immediate may now be dead, in which case remove it. |
4284 | if (Base.getNode()->use_empty()) |
4285 | CurDAG->RemoveDeadNode(Base.getNode()); |
4286 | } |
4287 | } |
4288 | |
4289 | |
4290 | /// createPPCISelDag - This pass converts a legalized DAG into a |
4291 | /// PowerPC-specific DAG, ready for instruction scheduling. |
4292 | /// |
4293 | FunctionPass *llvm::createPPCISelDag(PPCTargetMachine &TM) { |
4294 | return new PPCDAGToDAGISel(TM); |
4295 | } |
4296 | |
4297 | static void initializePassOnce(PassRegistry &Registry) { |
4298 | const char *Name = "PowerPC DAG->DAG Pattern Instruction Selection"; |
4299 | PassInfo *PI = new PassInfo(Name, "ppc-codegen", &SelectionDAGISel::ID, |
4300 | nullptr, false, false); |
4301 | Registry.registerPass(*PI, true); |
4302 | } |
4303 | |
4304 | void llvm::initializePPCDAGToDAGISelPass(PassRegistry &Registry) { |
4305 | CALL_ONCE_INITIALIZATION(initializePassOnce)static volatile sys::cas_flag initialized = 0; sys::cas_flag old_val = sys::CompareAndSwap(&initialized, 1, 0); if (old_val == 0) { initializePassOnce(Registry); sys::MemoryFence(); ; ; initialized = 2; ; } else { sys::cas_flag tmp = initialized; sys::MemoryFence (); while (tmp != 2) { tmp = initialized; sys::MemoryFence(); } } ;; |
4306 | } |
4307 |