?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
[ARM] Remove declaration of unimplemented function. NFC.
[llvm-complete.git] / lib / Target / Hexagon / HexagonBitSimplify.cpp
blob3068fb6f96296d86bedb6155729aec7f1a1b28a2
1 //===- HexagonBitSimplify.cpp ---------------------------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8
9 #include "BitTracker.h"
10 #include "HexagonBitTracker.h"
11 #include "HexagonInstrInfo.h"
12 #include "HexagonRegisterInfo.h"
13 #include "HexagonSubtarget.h"
14 #include "llvm/ADT/BitVector.h"
15 #include "llvm/ADT/DenseMap.h"
16 #include "llvm/ADT/GraphTraits.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/SmallVector.h"
19 #include "llvm/ADT/StringRef.h"
20 #include "llvm/CodeGen/MachineBasicBlock.h"
21 #include "llvm/CodeGen/MachineDominators.h"
22 #include "llvm/CodeGen/MachineFunction.h"
23 #include "llvm/CodeGen/MachineFunctionPass.h"
24 #include "llvm/CodeGen/MachineInstr.h"
25 #include "llvm/CodeGen/MachineInstrBuilder.h"
26 #include "llvm/CodeGen/MachineOperand.h"
27 #include "llvm/CodeGen/MachineRegisterInfo.h"
28 #include "llvm/CodeGen/TargetRegisterInfo.h"
29 #include "llvm/IR/DebugLoc.h"
30 #include "llvm/MC/MCInstrDesc.h"
31 #include "llvm/Pass.h"
32 #include "llvm/Support/CommandLine.h"
33 #include "llvm/Support/Compiler.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/ErrorHandling.h"
36 #include "llvm/Support/MathExtras.h"
37 #include "llvm/Support/raw_ostream.h"
38 #include <algorithm>
39 #include <cassert>
40 #include <cstdint>
41 #include <iterator>
42 #include <limits>
43 #include <utility>
44 #include <vector>
45
46 #define DEBUG_TYPE "hexbit"
47
48 using namespace llvm;
49
50 static cl::opt<bool> PreserveTiedOps("hexbit-keep-tied", cl::Hidden,
51 cl::init(true), cl::desc("Preserve subregisters in tied operands"));
52 static cl::opt<bool> GenExtract("hexbit-extract", cl::Hidden,
53 cl::init(true), cl::desc("Generate extract instructions"));
54 static cl::opt<bool> GenBitSplit("hexbit-bitsplit", cl::Hidden,
55 cl::init(true), cl::desc("Generate bitsplit instructions"));
56
57 static cl::opt<unsigned> MaxExtract("hexbit-max-extract", cl::Hidden,
58 cl::init(std::numeric_limits<unsigned>::max()));
59 static unsigned CountExtract = 0;
60 static cl::opt<unsigned> MaxBitSplit("hexbit-max-bitsplit", cl::Hidden,
61 cl::init(std::numeric_limits<unsigned>::max()));
62 static unsigned CountBitSplit = 0;
63
64 namespace llvm {
65
66 void initializeHexagonBitSimplifyPass(PassRegistry& Registry);
67 FunctionPass *createHexagonBitSimplify();
68
69 } // end namespace llvm
70
71 namespace {
72
73 // Set of virtual registers, based on BitVector.
74 struct RegisterSet : private BitVector {
75 RegisterSet() = default;
76 explicit RegisterSet(unsigned s, bool t = false) : BitVector(s, t) {}
77 RegisterSet(const RegisterSet &RS) = default;
78
79 using BitVector::clear;
80 using BitVector::count;
81
82 unsigned find_first() const {
83 int First = BitVector::find_first();
84 if (First < 0)
85 return 0;
86 return x2v(First);
87 }
88
89 unsigned find_next(unsigned Prev) const {
90 int Next = BitVector::find_next(v2x(Prev));
91 if (Next < 0)
92 return 0;
93 return x2v(Next);
94 }
95
96 RegisterSet &insert(unsigned R) {
97 unsigned Idx = v2x(R);
98 ensure(Idx);
99 return static_cast<RegisterSet&>(BitVector::set(Idx));
100 }
101 RegisterSet &remove(unsigned R) {
102 unsigned Idx = v2x(R);
103 if (Idx >= size())
104 return *this;
105 return static_cast<RegisterSet&>(BitVector::reset(Idx));
106 }
107
108 RegisterSet &insert(const RegisterSet &Rs) {
109 return static_cast<RegisterSet&>(BitVector::operator|=(Rs));
110 }
111 RegisterSet &remove(const RegisterSet &Rs) {
112 return static_cast<RegisterSet&>(BitVector::reset(Rs));
113 }
114
115 reference operator[](unsigned R) {
116 unsigned Idx = v2x(R);
117 ensure(Idx);
118 return BitVector::operator[](Idx);
119 }
120 bool operator[](unsigned R) const {
121 unsigned Idx = v2x(R);
122 assert(Idx < size());
123 return BitVector::operator[](Idx);
124 }
125 bool has(unsigned R) const {
126 unsigned Idx = v2x(R);
127 if (Idx >= size())
128 return false;
129 return BitVector::test(Idx);
130 }
131
132 bool empty() const {
133 return !BitVector::any();
134 }
135 bool includes(const RegisterSet &Rs) const {
136 // A.BitVector::test(B) <=> A-B != {}
137 return !Rs.BitVector::test(*this);
138 }
139 bool intersects(const RegisterSet &Rs) const {
140 return BitVector::anyCommon(Rs);
141 }
142
143 private:
144 void ensure(unsigned Idx) {
145 if (size() <= Idx)
146 resize(std::max(Idx+1, 32U));
147 }
148
149 static inline unsigned v2x(unsigned v) {
150 return Register::virtReg2Index(v);
151 }
152
153 static inline unsigned x2v(unsigned x) {
154 return Register::index2VirtReg(x);
155 }
156 };
157
158 struct PrintRegSet {
159 PrintRegSet(const RegisterSet &S, const TargetRegisterInfo *RI)
160 : RS(S), TRI(RI) {}
161
162 friend raw_ostream &operator<< (raw_ostream &OS,
163 const PrintRegSet &P);
164
165 private:
166 const RegisterSet &RS;
167 const TargetRegisterInfo *TRI;
168 };
169
170 raw_ostream &operator<< (raw_ostream &OS, const PrintRegSet &P)
171 LLVM_ATTRIBUTE_UNUSED;
172 raw_ostream &operator<< (raw_ostream &OS, const PrintRegSet &P) {
173 OS << '{';
174 for (unsigned R = P.RS.find_first(); R; R = P.RS.find_next(R))
175 OS << ' ' << printReg(R, P.TRI);
176 OS << " }";
177 return OS;
178 }
179
180 class Transformation;
181
182 class HexagonBitSimplify : public MachineFunctionPass {
183 public:
184 static char ID;
185
186 HexagonBitSimplify() : MachineFunctionPass(ID) {}
187
188 StringRef getPassName() const override {
189 return "Hexagon bit simplification";
190 }
191
192 void getAnalysisUsage(AnalysisUsage &AU) const override {
193 AU.addRequired<MachineDominatorTree>();
194 AU.addPreserved<MachineDominatorTree>();
195 MachineFunctionPass::getAnalysisUsage(AU);
196 }
197
198 bool runOnMachineFunction(MachineFunction &MF) override;
199
200 static void getInstrDefs(const MachineInstr &MI, RegisterSet &Defs);
201 static void getInstrUses(const MachineInstr &MI, RegisterSet &Uses);
202 static bool isEqual(const BitTracker::RegisterCell &RC1, uint16_t B1,
203 const BitTracker::RegisterCell &RC2, uint16_t B2, uint16_t W);
204 static bool isZero(const BitTracker::RegisterCell &RC, uint16_t B,
205 uint16_t W);
206 static bool getConst(const BitTracker::RegisterCell &RC, uint16_t B,
207 uint16_t W, uint64_t &U);
208 static bool replaceReg(unsigned OldR, unsigned NewR,
209 MachineRegisterInfo &MRI);
210 static bool getSubregMask(const BitTracker::RegisterRef &RR,
211 unsigned &Begin, unsigned &Width, MachineRegisterInfo &MRI);
212 static bool replaceRegWithSub(unsigned OldR, unsigned NewR,
213 unsigned NewSR, MachineRegisterInfo &MRI);
214 static bool replaceSubWithSub(unsigned OldR, unsigned OldSR,
215 unsigned NewR, unsigned NewSR, MachineRegisterInfo &MRI);
216 static bool parseRegSequence(const MachineInstr &I,
217 BitTracker::RegisterRef &SL, BitTracker::RegisterRef &SH,
218 const MachineRegisterInfo &MRI);
219
220 static bool getUsedBitsInStore(unsigned Opc, BitVector &Bits,
221 uint16_t Begin);
222 static bool getUsedBits(unsigned Opc, unsigned OpN, BitVector &Bits,
223 uint16_t Begin, const HexagonInstrInfo &HII);
224
225 static const TargetRegisterClass *getFinalVRegClass(
226 const BitTracker::RegisterRef &RR, MachineRegisterInfo &MRI);
227 static bool isTransparentCopy(const BitTracker::RegisterRef &RD,
228 const BitTracker::RegisterRef &RS, MachineRegisterInfo &MRI);
229
230 private:
231 MachineDominatorTree *MDT = nullptr;
232
233 bool visitBlock(MachineBasicBlock &B, Transformation &T, RegisterSet &AVs);
234 static bool hasTiedUse(unsigned Reg, MachineRegisterInfo &MRI,
235 unsigned NewSub = Hexagon::NoSubRegister);
236 };
237
238 using HBS = HexagonBitSimplify;
239
240 // The purpose of this class is to provide a common facility to traverse
241 // the function top-down or bottom-up via the dominator tree, and keep
242 // track of the available registers.
243 class Transformation {
244 public:
245 bool TopDown;
246
247 Transformation(bool TD) : TopDown(TD) {}
248 virtual ~Transformation() = default;
249
250 virtual bool processBlock(MachineBasicBlock &B, const RegisterSet &AVs) = 0;
251 };
252
253 } // end anonymous namespace
254
255 char HexagonBitSimplify::ID = 0;
256
257 INITIALIZE_PASS_BEGIN(HexagonBitSimplify, "hexagon-bit-simplify",
258 "Hexagon bit simplification", false, false)
259 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
260 INITIALIZE_PASS_END(HexagonBitSimplify, "hexagon-bit-simplify",
261 "Hexagon bit simplification", false, false)
262
263 bool HexagonBitSimplify::visitBlock(MachineBasicBlock &B, Transformation &T,
264 RegisterSet &AVs) {
265 bool Changed = false;
266
267 if (T.TopDown)
268 Changed = T.processBlock(B, AVs);
269
270 RegisterSet Defs;
271 for (auto &I : B)
272 getInstrDefs(I, Defs);
273 RegisterSet NewAVs = AVs;
274 NewAVs.insert(Defs);
275
276 for (auto *DTN : children<MachineDomTreeNode*>(MDT->getNode(&B)))
277 Changed |= visitBlock(*(DTN->getBlock()), T, NewAVs);
278
279 if (!T.TopDown)
280 Changed |= T.processBlock(B, AVs);
281
282 return Changed;
283 }
284
285 //
286 // Utility functions:
287 //
288 void HexagonBitSimplify::getInstrDefs(const MachineInstr &MI,
289 RegisterSet &Defs) {
290 for (auto &Op : MI.operands()) {
291 if (!Op.isReg() || !Op.isDef())
292 continue;
293 Register R = Op.getReg();
294 if (!Register::isVirtualRegister(R))
295 continue;
296 Defs.insert(R);
297 }
298 }
299
300 void HexagonBitSimplify::getInstrUses(const MachineInstr &MI,
301 RegisterSet &Uses) {
302 for (auto &Op : MI.operands()) {
303 if (!Op.isReg() || !Op.isUse())
304 continue;
305 Register R = Op.getReg();
306 if (!Register::isVirtualRegister(R))
307 continue;
308 Uses.insert(R);
309 }
310 }
311
312 // Check if all the bits in range [B, E) in both cells are equal.
313 bool HexagonBitSimplify::isEqual(const BitTracker::RegisterCell &RC1,
314 uint16_t B1, const BitTracker::RegisterCell &RC2, uint16_t B2,
315 uint16_t W) {
316 for (uint16_t i = 0; i < W; ++i) {
317 // If RC1[i] is "bottom", it cannot be proven equal to RC2[i].
318 if (RC1[B1+i].Type == BitTracker::BitValue::Ref && RC1[B1+i].RefI.Reg == 0)
319 return false;
320 // Same for RC2[i].
321 if (RC2[B2+i].Type == BitTracker::BitValue::Ref && RC2[B2+i].RefI.Reg == 0)
322 return false;
323 if (RC1[B1+i] != RC2[B2+i])
324 return false;
325 }
326 return true;
327 }
328
329 bool HexagonBitSimplify::isZero(const BitTracker::RegisterCell &RC,
330 uint16_t B, uint16_t W) {
331 assert(B < RC.width() && B+W <= RC.width());
332 for (uint16_t i = B; i < B+W; ++i)
333 if (!RC[i].is(0))
334 return false;
335 return true;
336 }
337
338 bool HexagonBitSimplify::getConst(const BitTracker::RegisterCell &RC,
339 uint16_t B, uint16_t W, uint64_t &U) {
340 assert(B < RC.width() && B+W <= RC.width());
341 int64_t T = 0;
342 for (uint16_t i = B+W; i > B; --i) {
343 const BitTracker::BitValue &BV = RC[i-1];
344 T <<= 1;
345 if (BV.is(1))
346 T |= 1;
347 else if (!BV.is(0))
348 return false;
349 }
350 U = T;
351 return true;
352 }
353
354 bool HexagonBitSimplify::replaceReg(unsigned OldR, unsigned NewR,
355 MachineRegisterInfo &MRI) {
356 if (!Register::isVirtualRegister(OldR) || !Register::isVirtualRegister(NewR))
357 return false;
358 auto Begin = MRI.use_begin(OldR), End = MRI.use_end();
359 decltype(End) NextI;
360 for (auto I = Begin; I != End; I = NextI) {
361 NextI = std::next(I);
362 I->setReg(NewR);
363 }
364 return Begin != End;
365 }
366
367 bool HexagonBitSimplify::replaceRegWithSub(unsigned OldR, unsigned NewR,
368 unsigned NewSR, MachineRegisterInfo &MRI) {
369 if (!Register::isVirtualRegister(OldR) || !Register::isVirtualRegister(NewR))
370 return false;
371 if (hasTiedUse(OldR, MRI, NewSR))
372 return false;
373 auto Begin = MRI.use_begin(OldR), End = MRI.use_end();
374 decltype(End) NextI;
375 for (auto I = Begin; I != End; I = NextI) {
376 NextI = std::next(I);
377 I->setReg(NewR);
378 I->setSubReg(NewSR);
379 }
380 return Begin != End;
381 }
382
383 bool HexagonBitSimplify::replaceSubWithSub(unsigned OldR, unsigned OldSR,
384 unsigned NewR, unsigned NewSR, MachineRegisterInfo &MRI) {
385 if (!Register::isVirtualRegister(OldR) || !Register::isVirtualRegister(NewR))
386 return false;
387 if (OldSR != NewSR && hasTiedUse(OldR, MRI, NewSR))
388 return false;
389 auto Begin = MRI.use_begin(OldR), End = MRI.use_end();
390 decltype(End) NextI;
391 for (auto I = Begin; I != End; I = NextI) {
392 NextI = std::next(I);
393 if (I->getSubReg() != OldSR)
394 continue;
395 I->setReg(NewR);
396 I->setSubReg(NewSR);
397 }
398 return Begin != End;
399 }
400
401 // For a register ref (pair Reg:Sub), set Begin to the position of the LSB
402 // of Sub in Reg, and set Width to the size of Sub in bits. Return true,
403 // if this succeeded, otherwise return false.
404 bool HexagonBitSimplify::getSubregMask(const BitTracker::RegisterRef &RR,
405 unsigned &Begin, unsigned &Width, MachineRegisterInfo &MRI) {
406 const TargetRegisterClass *RC = MRI.getRegClass(RR.Reg);
407 if (RR.Sub == 0) {
408 Begin = 0;
409 Width = MRI.getTargetRegisterInfo()->getRegSizeInBits(*RC);
410 return true;
411 }
412
413 Begin = 0;
414
415 switch (RC->getID()) {
416 case Hexagon::DoubleRegsRegClassID:
417 case Hexagon::HvxWRRegClassID:
418 Width = MRI.getTargetRegisterInfo()->getRegSizeInBits(*RC) / 2;
419 if (RR.Sub == Hexagon::isub_hi || RR.Sub == Hexagon::vsub_hi)
420 Begin = Width;
421 break;
422 default:
423 return false;
424 }
425 return true;
426 }
427
428
429 // For a REG_SEQUENCE, set SL to the low subregister and SH to the high
430 // subregister.
431 bool HexagonBitSimplify::parseRegSequence(const MachineInstr &I,
432 BitTracker::RegisterRef &SL, BitTracker::RegisterRef &SH,
433 const MachineRegisterInfo &MRI) {
434 assert(I.getOpcode() == TargetOpcode::REG_SEQUENCE);
435 unsigned Sub1 = I.getOperand(2).getImm(), Sub2 = I.getOperand(4).getImm();
436 auto &DstRC = *MRI.getRegClass(I.getOperand(0).getReg());
437 auto &HRI = static_cast<const HexagonRegisterInfo&>(
438 *MRI.getTargetRegisterInfo());
439 unsigned SubLo = HRI.getHexagonSubRegIndex(DstRC, Hexagon::ps_sub_lo);
440 unsigned SubHi = HRI.getHexagonSubRegIndex(DstRC, Hexagon::ps_sub_hi);
441 assert((Sub1 == SubLo && Sub2 == SubHi) || (Sub1 == SubHi && Sub2 == SubLo));
442 if (Sub1 == SubLo && Sub2 == SubHi) {
443 SL = I.getOperand(1);
444 SH = I.getOperand(3);
445 return true;
446 }
447 if (Sub1 == SubHi && Sub2 == SubLo) {
448 SH = I.getOperand(1);
449 SL = I.getOperand(3);
450 return true;
451 }
452 return false;
453 }
454
455 // All stores (except 64-bit stores) take a 32-bit register as the source
456 // of the value to be stored. If the instruction stores into a location
457 // that is shorter than 32 bits, some bits of the source register are not
458 // used. For each store instruction, calculate the set of used bits in
459 // the source register, and set appropriate bits in Bits. Return true if
460 // the bits are calculated, false otherwise.
461 bool HexagonBitSimplify::getUsedBitsInStore(unsigned Opc, BitVector &Bits,
462 uint16_t Begin) {
463 using namespace Hexagon;
464
465 switch (Opc) {
466 // Store byte
467 case S2_storerb_io: // memb(Rs32+#s11:0)=Rt32
468 case S2_storerbnew_io: // memb(Rs32+#s11:0)=Nt8.new
469 case S2_pstorerbt_io: // if (Pv4) memb(Rs32+#u6:0)=Rt32
470 case S2_pstorerbf_io: // if (!Pv4) memb(Rs32+#u6:0)=Rt32
471 case S4_pstorerbtnew_io: // if (Pv4.new) memb(Rs32+#u6:0)=Rt32
472 case S4_pstorerbfnew_io: // if (!Pv4.new) memb(Rs32+#u6:0)=Rt32
473 case S2_pstorerbnewt_io: // if (Pv4) memb(Rs32+#u6:0)=Nt8.new
474 case S2_pstorerbnewf_io: // if (!Pv4) memb(Rs32+#u6:0)=Nt8.new
475 case S4_pstorerbnewtnew_io: // if (Pv4.new) memb(Rs32+#u6:0)=Nt8.new
476 case S4_pstorerbnewfnew_io: // if (!Pv4.new) memb(Rs32+#u6:0)=Nt8.new
477 case S2_storerb_pi: // memb(Rx32++#s4:0)=Rt32
478 case S2_storerbnew_pi: // memb(Rx32++#s4:0)=Nt8.new
479 case S2_pstorerbt_pi: // if (Pv4) memb(Rx32++#s4:0)=Rt32
480 case S2_pstorerbf_pi: // if (!Pv4) memb(Rx32++#s4:0)=Rt32
481 case S2_pstorerbtnew_pi: // if (Pv4.new) memb(Rx32++#s4:0)=Rt32
482 case S2_pstorerbfnew_pi: // if (!Pv4.new) memb(Rx32++#s4:0)=Rt32
483 case S2_pstorerbnewt_pi: // if (Pv4) memb(Rx32++#s4:0)=Nt8.new
484 case S2_pstorerbnewf_pi: // if (!Pv4) memb(Rx32++#s4:0)=Nt8.new
485 case S2_pstorerbnewtnew_pi: // if (Pv4.new) memb(Rx32++#s4:0)=Nt8.new
486 case S2_pstorerbnewfnew_pi: // if (!Pv4.new) memb(Rx32++#s4:0)=Nt8.new
487 case S4_storerb_ap: // memb(Re32=#U6)=Rt32
488 case S4_storerbnew_ap: // memb(Re32=#U6)=Nt8.new
489 case S2_storerb_pr: // memb(Rx32++Mu2)=Rt32
490 case S2_storerbnew_pr: // memb(Rx32++Mu2)=Nt8.new
491 case S4_storerb_ur: // memb(Ru32<<#u2+#U6)=Rt32
492 case S4_storerbnew_ur: // memb(Ru32<<#u2+#U6)=Nt8.new
493 case S2_storerb_pbr: // memb(Rx32++Mu2:brev)=Rt32
494 case S2_storerbnew_pbr: // memb(Rx32++Mu2:brev)=Nt8.new
495 case S2_storerb_pci: // memb(Rx32++#s4:0:circ(Mu2))=Rt32
496 case S2_storerbnew_pci: // memb(Rx32++#s4:0:circ(Mu2))=Nt8.new
497 case S2_storerb_pcr: // memb(Rx32++I:circ(Mu2))=Rt32
498 case S2_storerbnew_pcr: // memb(Rx32++I:circ(Mu2))=Nt8.new
499 case S4_storerb_rr: // memb(Rs32+Ru32<<#u2)=Rt32
500 case S4_storerbnew_rr: // memb(Rs32+Ru32<<#u2)=Nt8.new
501 case S4_pstorerbt_rr: // if (Pv4) memb(Rs32+Ru32<<#u2)=Rt32
502 case S4_pstorerbf_rr: // if (!Pv4) memb(Rs32+Ru32<<#u2)=Rt32
503 case S4_pstorerbtnew_rr: // if (Pv4.new) memb(Rs32+Ru32<<#u2)=Rt32
504 case S4_pstorerbfnew_rr: // if (!Pv4.new) memb(Rs32+Ru32<<#u2)=Rt32
505 case S4_pstorerbnewt_rr: // if (Pv4) memb(Rs32+Ru32<<#u2)=Nt8.new
506 case S4_pstorerbnewf_rr: // if (!Pv4) memb(Rs32+Ru32<<#u2)=Nt8.new
507 case S4_pstorerbnewtnew_rr: // if (Pv4.new) memb(Rs32+Ru32<<#u2)=Nt8.new
508 case S4_pstorerbnewfnew_rr: // if (!Pv4.new) memb(Rs32+Ru32<<#u2)=Nt8.new
509 case S2_storerbgp: // memb(gp+#u16:0)=Rt32
510 case S2_storerbnewgp: // memb(gp+#u16:0)=Nt8.new
511 case S4_pstorerbt_abs: // if (Pv4) memb(#u6)=Rt32
512 case S4_pstorerbf_abs: // if (!Pv4) memb(#u6)=Rt32
513 case S4_pstorerbtnew_abs: // if (Pv4.new) memb(#u6)=Rt32
514 case S4_pstorerbfnew_abs: // if (!Pv4.new) memb(#u6)=Rt32
515 case S4_pstorerbnewt_abs: // if (Pv4) memb(#u6)=Nt8.new
516 case S4_pstorerbnewf_abs: // if (!Pv4) memb(#u6)=Nt8.new
517 case S4_pstorerbnewtnew_abs: // if (Pv4.new) memb(#u6)=Nt8.new
518 case S4_pstorerbnewfnew_abs: // if (!Pv4.new) memb(#u6)=Nt8.new
519 Bits.set(Begin, Begin+8);
520 return true;
521
522 // Store low half
523 case S2_storerh_io: // memh(Rs32+#s11:1)=Rt32
524 case S2_storerhnew_io: // memh(Rs32+#s11:1)=Nt8.new
525 case S2_pstorerht_io: // if (Pv4) memh(Rs32+#u6:1)=Rt32
526 case S2_pstorerhf_io: // if (!Pv4) memh(Rs32+#u6:1)=Rt32
527 case S4_pstorerhtnew_io: // if (Pv4.new) memh(Rs32+#u6:1)=Rt32
528 case S4_pstorerhfnew_io: // if (!Pv4.new) memh(Rs32+#u6:1)=Rt32
529 case S2_pstorerhnewt_io: // if (Pv4) memh(Rs32+#u6:1)=Nt8.new
530 case S2_pstorerhnewf_io: // if (!Pv4) memh(Rs32+#u6:1)=Nt8.new
531 case S4_pstorerhnewtnew_io: // if (Pv4.new) memh(Rs32+#u6:1)=Nt8.new
532 case S4_pstorerhnewfnew_io: // if (!Pv4.new) memh(Rs32+#u6:1)=Nt8.new
533 case S2_storerh_pi: // memh(Rx32++#s4:1)=Rt32
534 case S2_storerhnew_pi: // memh(Rx32++#s4:1)=Nt8.new
535 case S2_pstorerht_pi: // if (Pv4) memh(Rx32++#s4:1)=Rt32
536 case S2_pstorerhf_pi: // if (!Pv4) memh(Rx32++#s4:1)=Rt32
537 case S2_pstorerhtnew_pi: // if (Pv4.new) memh(Rx32++#s4:1)=Rt32
538 case S2_pstorerhfnew_pi: // if (!Pv4.new) memh(Rx32++#s4:1)=Rt32
539 case S2_pstorerhnewt_pi: // if (Pv4) memh(Rx32++#s4:1)=Nt8.new
540 case S2_pstorerhnewf_pi: // if (!Pv4) memh(Rx32++#s4:1)=Nt8.new
541 case S2_pstorerhnewtnew_pi: // if (Pv4.new) memh(Rx32++#s4:1)=Nt8.new
542 case S2_pstorerhnewfnew_pi: // if (!Pv4.new) memh(Rx32++#s4:1)=Nt8.new
543 case S4_storerh_ap: // memh(Re32=#U6)=Rt32
544 case S4_storerhnew_ap: // memh(Re32=#U6)=Nt8.new
545 case S2_storerh_pr: // memh(Rx32++Mu2)=Rt32
546 case S2_storerhnew_pr: // memh(Rx32++Mu2)=Nt8.new
547 case S4_storerh_ur: // memh(Ru32<<#u2+#U6)=Rt32
548 case S4_storerhnew_ur: // memh(Ru32<<#u2+#U6)=Nt8.new
549 case S2_storerh_pbr: // memh(Rx32++Mu2:brev)=Rt32
550 case S2_storerhnew_pbr: // memh(Rx32++Mu2:brev)=Nt8.new
551 case S2_storerh_pci: // memh(Rx32++#s4:1:circ(Mu2))=Rt32
552 case S2_storerhnew_pci: // memh(Rx32++#s4:1:circ(Mu2))=Nt8.new
553 case S2_storerh_pcr: // memh(Rx32++I:circ(Mu2))=Rt32
554 case S2_storerhnew_pcr: // memh(Rx32++I:circ(Mu2))=Nt8.new
555 case S4_storerh_rr: // memh(Rs32+Ru32<<#u2)=Rt32
556 case S4_pstorerht_rr: // if (Pv4) memh(Rs32+Ru32<<#u2)=Rt32
557 case S4_pstorerhf_rr: // if (!Pv4) memh(Rs32+Ru32<<#u2)=Rt32
558 case S4_pstorerhtnew_rr: // if (Pv4.new) memh(Rs32+Ru32<<#u2)=Rt32
559 case S4_pstorerhfnew_rr: // if (!Pv4.new) memh(Rs32+Ru32<<#u2)=Rt32
560 case S4_storerhnew_rr: // memh(Rs32+Ru32<<#u2)=Nt8.new
561 case S4_pstorerhnewt_rr: // if (Pv4) memh(Rs32+Ru32<<#u2)=Nt8.new
562 case S4_pstorerhnewf_rr: // if (!Pv4) memh(Rs32+Ru32<<#u2)=Nt8.new
563 case S4_pstorerhnewtnew_rr: // if (Pv4.new) memh(Rs32+Ru32<<#u2)=Nt8.new
564 case S4_pstorerhnewfnew_rr: // if (!Pv4.new) memh(Rs32+Ru32<<#u2)=Nt8.new
565 case S2_storerhgp: // memh(gp+#u16:1)=Rt32
566 case S2_storerhnewgp: // memh(gp+#u16:1)=Nt8.new
567 case S4_pstorerht_abs: // if (Pv4) memh(#u6)=Rt32
568 case S4_pstorerhf_abs: // if (!Pv4) memh(#u6)=Rt32
569 case S4_pstorerhtnew_abs: // if (Pv4.new) memh(#u6)=Rt32
570 case S4_pstorerhfnew_abs: // if (!Pv4.new) memh(#u6)=Rt32
571 case S4_pstorerhnewt_abs: // if (Pv4) memh(#u6)=Nt8.new
572 case S4_pstorerhnewf_abs: // if (!Pv4) memh(#u6)=Nt8.new
573 case S4_pstorerhnewtnew_abs: // if (Pv4.new) memh(#u6)=Nt8.new
574 case S4_pstorerhnewfnew_abs: // if (!Pv4.new) memh(#u6)=Nt8.new
575 Bits.set(Begin, Begin+16);
576 return true;
577
578 // Store high half
579 case S2_storerf_io: // memh(Rs32+#s11:1)=Rt.H32
580 case S2_pstorerft_io: // if (Pv4) memh(Rs32+#u6:1)=Rt.H32
581 case S2_pstorerff_io: // if (!Pv4) memh(Rs32+#u6:1)=Rt.H32
582 case S4_pstorerftnew_io: // if (Pv4.new) memh(Rs32+#u6:1)=Rt.H32
583 case S4_pstorerffnew_io: // if (!Pv4.new) memh(Rs32+#u6:1)=Rt.H32
584 case S2_storerf_pi: // memh(Rx32++#s4:1)=Rt.H32
585 case S2_pstorerft_pi: // if (Pv4) memh(Rx32++#s4:1)=Rt.H32
586 case S2_pstorerff_pi: // if (!Pv4) memh(Rx32++#s4:1)=Rt.H32
587 case S2_pstorerftnew_pi: // if (Pv4.new) memh(Rx32++#s4:1)=Rt.H32
588 case S2_pstorerffnew_pi: // if (!Pv4.new) memh(Rx32++#s4:1)=Rt.H32
589 case S4_storerf_ap: // memh(Re32=#U6)=Rt.H32
590 case S2_storerf_pr: // memh(Rx32++Mu2)=Rt.H32
591 case S4_storerf_ur: // memh(Ru32<<#u2+#U6)=Rt.H32
592 case S2_storerf_pbr: // memh(Rx32++Mu2:brev)=Rt.H32
593 case S2_storerf_pci: // memh(Rx32++#s4:1:circ(Mu2))=Rt.H32
594 case S2_storerf_pcr: // memh(Rx32++I:circ(Mu2))=Rt.H32
595 case S4_storerf_rr: // memh(Rs32+Ru32<<#u2)=Rt.H32
596 case S4_pstorerft_rr: // if (Pv4) memh(Rs32+Ru32<<#u2)=Rt.H32
597 case S4_pstorerff_rr: // if (!Pv4) memh(Rs32+Ru32<<#u2)=Rt.H32
598 case S4_pstorerftnew_rr: // if (Pv4.new) memh(Rs32+Ru32<<#u2)=Rt.H32
599 case S4_pstorerffnew_rr: // if (!Pv4.new) memh(Rs32+Ru32<<#u2)=Rt.H32
600 case S2_storerfgp: // memh(gp+#u16:1)=Rt.H32
601 case S4_pstorerft_abs: // if (Pv4) memh(#u6)=Rt.H32
602 case S4_pstorerff_abs: // if (!Pv4) memh(#u6)=Rt.H32
603 case S4_pstorerftnew_abs: // if (Pv4.new) memh(#u6)=Rt.H32
604 case S4_pstorerffnew_abs: // if (!Pv4.new) memh(#u6)=Rt.H32
605 Bits.set(Begin+16, Begin+32);
606 return true;
607 }
608
609 return false;
610 }
611
612 // For an instruction with opcode Opc, calculate the set of bits that it
613 // uses in a register in operand OpN. This only calculates the set of used
614 // bits for cases where it does not depend on any operands (as is the case
615 // in shifts, for example). For concrete instructions from a program, the
616 // operand may be a subregister of a larger register, while Bits would
617 // correspond to the larger register in its entirety. Because of that,
618 // the parameter Begin can be used to indicate which bit of Bits should be
619 // considered the LSB of the operand.
620 bool HexagonBitSimplify::getUsedBits(unsigned Opc, unsigned OpN,
621 BitVector &Bits, uint16_t Begin, const HexagonInstrInfo &HII) {
622 using namespace Hexagon;
623
624 const MCInstrDesc &D = HII.get(Opc);
625 if (D.mayStore()) {
626 if (OpN == D.getNumOperands()-1)
627 return getUsedBitsInStore(Opc, Bits, Begin);
628 return false;
629 }
630
631 switch (Opc) {
632 // One register source. Used bits: R1[0-7].
633 case A2_sxtb:
634 case A2_zxtb:
635 case A4_cmpbeqi:
636 case A4_cmpbgti:
637 case A4_cmpbgtui:
638 if (OpN == 1) {
639 Bits.set(Begin, Begin+8);
640 return true;
641 }
642 break;
643
644 // One register source. Used bits: R1[0-15].
645 case A2_aslh:
646 case A2_sxth:
647 case A2_zxth:
648 case A4_cmpheqi:
649 case A4_cmphgti:
650 case A4_cmphgtui:
651 if (OpN == 1) {
652 Bits.set(Begin, Begin+16);
653 return true;
654 }
655 break;
656
657 // One register source. Used bits: R1[16-31].
658 case A2_asrh:
659 if (OpN == 1) {
660 Bits.set(Begin+16, Begin+32);
661 return true;
662 }
663 break;
664
665 // Two register sources. Used bits: R1[0-7], R2[0-7].
666 case A4_cmpbeq:
667 case A4_cmpbgt:
668 case A4_cmpbgtu:
669 if (OpN == 1) {
670 Bits.set(Begin, Begin+8);
671 return true;
672 }
673 break;
674
675 // Two register sources. Used bits: R1[0-15], R2[0-15].
676 case A4_cmpheq:
677 case A4_cmphgt:
678 case A4_cmphgtu:
679 case A2_addh_h16_ll:
680 case A2_addh_h16_sat_ll:
681 case A2_addh_l16_ll:
682 case A2_addh_l16_sat_ll:
683 case A2_combine_ll:
684 case A2_subh_h16_ll:
685 case A2_subh_h16_sat_ll:
686 case A2_subh_l16_ll:
687 case A2_subh_l16_sat_ll:
688 case M2_mpy_acc_ll_s0:
689 case M2_mpy_acc_ll_s1:
690 case M2_mpy_acc_sat_ll_s0:
691 case M2_mpy_acc_sat_ll_s1:
692 case M2_mpy_ll_s0:
693 case M2_mpy_ll_s1:
694 case M2_mpy_nac_ll_s0:
695 case M2_mpy_nac_ll_s1:
696 case M2_mpy_nac_sat_ll_s0:
697 case M2_mpy_nac_sat_ll_s1:
698 case M2_mpy_rnd_ll_s0:
699 case M2_mpy_rnd_ll_s1:
700 case M2_mpy_sat_ll_s0:
701 case M2_mpy_sat_ll_s1:
702 case M2_mpy_sat_rnd_ll_s0:
703 case M2_mpy_sat_rnd_ll_s1:
704 case M2_mpyd_acc_ll_s0:
705 case M2_mpyd_acc_ll_s1:
706 case M2_mpyd_ll_s0:
707 case M2_mpyd_ll_s1:
708 case M2_mpyd_nac_ll_s0:
709 case M2_mpyd_nac_ll_s1:
710 case M2_mpyd_rnd_ll_s0:
711 case M2_mpyd_rnd_ll_s1:
712 case M2_mpyu_acc_ll_s0:
713 case M2_mpyu_acc_ll_s1:
714 case M2_mpyu_ll_s0:
715 case M2_mpyu_ll_s1:
716 case M2_mpyu_nac_ll_s0:
717 case M2_mpyu_nac_ll_s1:
718 case M2_mpyud_acc_ll_s0:
719 case M2_mpyud_acc_ll_s1:
720 case M2_mpyud_ll_s0:
721 case M2_mpyud_ll_s1:
722 case M2_mpyud_nac_ll_s0:
723 case M2_mpyud_nac_ll_s1:
724 if (OpN == 1 || OpN == 2) {
725 Bits.set(Begin, Begin+16);
726 return true;
727 }
728 break;
729
730 // Two register sources. Used bits: R1[0-15], R2[16-31].
731 case A2_addh_h16_lh:
732 case A2_addh_h16_sat_lh:
733 case A2_combine_lh:
734 case A2_subh_h16_lh:
735 case A2_subh_h16_sat_lh:
736 case M2_mpy_acc_lh_s0:
737 case M2_mpy_acc_lh_s1:
738 case M2_mpy_acc_sat_lh_s0:
739 case M2_mpy_acc_sat_lh_s1:
740 case M2_mpy_lh_s0:
741 case M2_mpy_lh_s1:
742 case M2_mpy_nac_lh_s0:
743 case M2_mpy_nac_lh_s1:
744 case M2_mpy_nac_sat_lh_s0:
745 case M2_mpy_nac_sat_lh_s1:
746 case M2_mpy_rnd_lh_s0:
747 case M2_mpy_rnd_lh_s1:
748 case M2_mpy_sat_lh_s0:
749 case M2_mpy_sat_lh_s1:
750 case M2_mpy_sat_rnd_lh_s0:
751 case M2_mpy_sat_rnd_lh_s1:
752 case M2_mpyd_acc_lh_s0:
753 case M2_mpyd_acc_lh_s1:
754 case M2_mpyd_lh_s0:
755 case M2_mpyd_lh_s1:
756 case M2_mpyd_nac_lh_s0:
757 case M2_mpyd_nac_lh_s1:
758 case M2_mpyd_rnd_lh_s0:
759 case M2_mpyd_rnd_lh_s1:
760 case M2_mpyu_acc_lh_s0:
761 case M2_mpyu_acc_lh_s1:
762 case M2_mpyu_lh_s0:
763 case M2_mpyu_lh_s1:
764 case M2_mpyu_nac_lh_s0:
765 case M2_mpyu_nac_lh_s1:
766 case M2_mpyud_acc_lh_s0:
767 case M2_mpyud_acc_lh_s1:
768 case M2_mpyud_lh_s0:
769 case M2_mpyud_lh_s1:
770 case M2_mpyud_nac_lh_s0:
771 case M2_mpyud_nac_lh_s1:
772 // These four are actually LH.
773 case A2_addh_l16_hl:
774 case A2_addh_l16_sat_hl:
775 case A2_subh_l16_hl:
776 case A2_subh_l16_sat_hl:
777 if (OpN == 1) {
778 Bits.set(Begin, Begin+16);
779 return true;
780 }
781 if (OpN == 2) {
782 Bits.set(Begin+16, Begin+32);
783 return true;
784 }
785 break;
786
787 // Two register sources, used bits: R1[16-31], R2[0-15].
788 case A2_addh_h16_hl:
789 case A2_addh_h16_sat_hl:
790 case A2_combine_hl:
791 case A2_subh_h16_hl:
792 case A2_subh_h16_sat_hl:
793 case M2_mpy_acc_hl_s0:
794 case M2_mpy_acc_hl_s1:
795 case M2_mpy_acc_sat_hl_s0:
796 case M2_mpy_acc_sat_hl_s1:
797 case M2_mpy_hl_s0:
798 case M2_mpy_hl_s1:
799 case M2_mpy_nac_hl_s0:
800 case M2_mpy_nac_hl_s1:
801 case M2_mpy_nac_sat_hl_s0:
802 case M2_mpy_nac_sat_hl_s1:
803 case M2_mpy_rnd_hl_s0:
804 case M2_mpy_rnd_hl_s1:
805 case M2_mpy_sat_hl_s0:
806 case M2_mpy_sat_hl_s1:
807 case M2_mpy_sat_rnd_hl_s0:
808 case M2_mpy_sat_rnd_hl_s1:
809 case M2_mpyd_acc_hl_s0:
810 case M2_mpyd_acc_hl_s1:
811 case M2_mpyd_hl_s0:
812 case M2_mpyd_hl_s1:
813 case M2_mpyd_nac_hl_s0:
814 case M2_mpyd_nac_hl_s1:
815 case M2_mpyd_rnd_hl_s0:
816 case M2_mpyd_rnd_hl_s1:
817 case M2_mpyu_acc_hl_s0:
818 case M2_mpyu_acc_hl_s1:
819 case M2_mpyu_hl_s0:
820 case M2_mpyu_hl_s1:
821 case M2_mpyu_nac_hl_s0:
822 case M2_mpyu_nac_hl_s1:
823 case M2_mpyud_acc_hl_s0:
824 case M2_mpyud_acc_hl_s1:
825 case M2_mpyud_hl_s0:
826 case M2_mpyud_hl_s1:
827 case M2_mpyud_nac_hl_s0:
828 case M2_mpyud_nac_hl_s1:
829 if (OpN == 1) {
830 Bits.set(Begin+16, Begin+32);
831 return true;
832 }
833 if (OpN == 2) {
834 Bits.set(Begin, Begin+16);
835 return true;
836 }
837 break;
838
839 // Two register sources, used bits: R1[16-31], R2[16-31].
840 case A2_addh_h16_hh:
841 case A2_addh_h16_sat_hh:
842 case A2_combine_hh:
843 case A2_subh_h16_hh:
844 case A2_subh_h16_sat_hh:
845 case M2_mpy_acc_hh_s0:
846 case M2_mpy_acc_hh_s1:
847 case M2_mpy_acc_sat_hh_s0:
848 case M2_mpy_acc_sat_hh_s1:
849 case M2_mpy_hh_s0:
850 case M2_mpy_hh_s1:
851 case M2_mpy_nac_hh_s0:
852 case M2_mpy_nac_hh_s1:
853 case M2_mpy_nac_sat_hh_s0:
854 case M2_mpy_nac_sat_hh_s1:
855 case M2_mpy_rnd_hh_s0:
856 case M2_mpy_rnd_hh_s1:
857 case M2_mpy_sat_hh_s0:
858 case M2_mpy_sat_hh_s1:
859 case M2_mpy_sat_rnd_hh_s0:
860 case M2_mpy_sat_rnd_hh_s1:
861 case M2_mpyd_acc_hh_s0:
862 case M2_mpyd_acc_hh_s1:
863 case M2_mpyd_hh_s0:
864 case M2_mpyd_hh_s1:
865 case M2_mpyd_nac_hh_s0:
866 case M2_mpyd_nac_hh_s1:
867 case M2_mpyd_rnd_hh_s0:
868 case M2_mpyd_rnd_hh_s1:
869 case M2_mpyu_acc_hh_s0:
870 case M2_mpyu_acc_hh_s1:
871 case M2_mpyu_hh_s0:
872 case M2_mpyu_hh_s1:
873 case M2_mpyu_nac_hh_s0:
874 case M2_mpyu_nac_hh_s1:
875 case M2_mpyud_acc_hh_s0:
876 case M2_mpyud_acc_hh_s1:
877 case M2_mpyud_hh_s0:
878 case M2_mpyud_hh_s1:
879 case M2_mpyud_nac_hh_s0:
880 case M2_mpyud_nac_hh_s1:
881 if (OpN == 1 || OpN == 2) {
882 Bits.set(Begin+16, Begin+32);
883 return true;
884 }
885 break;
886 }
887
888 return false;
889 }
890
891 // Calculate the register class that matches Reg:Sub. For example, if
892 // %1 is a double register, then %1:isub_hi would match the "int"
893 // register class.
894 const TargetRegisterClass *HexagonBitSimplify::getFinalVRegClass(
895 const BitTracker::RegisterRef &RR, MachineRegisterInfo &MRI) {
896 if (!Register::isVirtualRegister(RR.Reg))
897 return nullptr;
898 auto *RC = MRI.getRegClass(RR.Reg);
899 if (RR.Sub == 0)
900 return RC;
901 auto &HRI = static_cast<const HexagonRegisterInfo&>(
902 *MRI.getTargetRegisterInfo());
903
904 auto VerifySR = [&HRI] (const TargetRegisterClass *RC, unsigned Sub) -> void {
905 (void)HRI;
906 assert(Sub == HRI.getHexagonSubRegIndex(*RC, Hexagon::ps_sub_lo) ||
907 Sub == HRI.getHexagonSubRegIndex(*RC, Hexagon::ps_sub_hi));
908 };
909
910 switch (RC->getID()) {
911 case Hexagon::DoubleRegsRegClassID:
912 VerifySR(RC, RR.Sub);
913 return &Hexagon::IntRegsRegClass;
914 case Hexagon::HvxWRRegClassID:
915 VerifySR(RC, RR.Sub);
916 return &Hexagon::HvxVRRegClass;
917 }
918 return nullptr;
919 }
920
921 // Check if RD could be replaced with RS at any possible use of RD.
922 // For example a predicate register cannot be replaced with a integer
923 // register, but a 64-bit register with a subregister can be replaced
924 // with a 32-bit register.
925 bool HexagonBitSimplify::isTransparentCopy(const BitTracker::RegisterRef &RD,
926 const BitTracker::RegisterRef &RS, MachineRegisterInfo &MRI) {
927 if (!Register::isVirtualRegister(RD.Reg) ||
928 !Register::isVirtualRegister(RS.Reg))
929 return false;
930 // Return false if one (or both) classes are nullptr.
931 auto *DRC = getFinalVRegClass(RD, MRI);
932 if (!DRC)
933 return false;
934
935 return DRC == getFinalVRegClass(RS, MRI);
936 }
937
938 bool HexagonBitSimplify::hasTiedUse(unsigned Reg, MachineRegisterInfo &MRI,
939 unsigned NewSub) {
940 if (!PreserveTiedOps)
941 return false;
942 return llvm::any_of(MRI.use_operands(Reg),
943 [NewSub] (const MachineOperand &Op) -> bool {
944 return Op.getSubReg() != NewSub && Op.isTied();
945 });
946 }
947
948 namespace {
949
950 class DeadCodeElimination {
951 public:
952 DeadCodeElimination(MachineFunction &mf, MachineDominatorTree &mdt)
953 : MF(mf), HII(*MF.getSubtarget<HexagonSubtarget>().getInstrInfo()),
954 MDT(mdt), MRI(mf.getRegInfo()) {}
955
956 bool run() {
957 return runOnNode(MDT.getRootNode());
958 }
959
960 private:
961 bool isDead(unsigned R) const;
962 bool runOnNode(MachineDomTreeNode *N);
963
964 MachineFunction &MF;
965 const HexagonInstrInfo &HII;
966 MachineDominatorTree &MDT;
967 MachineRegisterInfo &MRI;
968 };
969
970 } // end anonymous namespace
971
972 bool DeadCodeElimination::isDead(unsigned R) const {
973 for (auto I = MRI.use_begin(R), E = MRI.use_end(); I != E; ++I) {
974 MachineInstr *UseI = I->getParent();
975 if (UseI->isDebugValue())
976 continue;
977 if (UseI->isPHI()) {
978 assert(!UseI->getOperand(0).getSubReg());
979 Register DR = UseI->getOperand(0).getReg();
980 if (DR == R)
981 continue;
982 }
983 return false;
984 }
985 return true;
986 }
987
988 bool DeadCodeElimination::runOnNode(MachineDomTreeNode *N) {
989 bool Changed = false;
990
991 for (auto *DTN : children<MachineDomTreeNode*>(N))
992 Changed |= runOnNode(DTN);
993
994 MachineBasicBlock *B = N->getBlock();
995 std::vector<MachineInstr*> Instrs;
996 for (auto I = B->rbegin(), E = B->rend(); I != E; ++I)
997 Instrs.push_back(&*I);
998
999 for (auto MI : Instrs) {
1000 unsigned Opc = MI->getOpcode();
1001 // Do not touch lifetime markers. This is why the target-independent DCE
1002 // cannot be used.
1003 if (Opc == TargetOpcode::LIFETIME_START ||
1004 Opc == TargetOpcode::LIFETIME_END)
1005 continue;
1006 bool Store = false;
1007 if (MI->isInlineAsm())
1008 continue;
1009 // Delete PHIs if possible.
1010 if (!MI->isPHI() && !MI->isSafeToMove(nullptr, Store))
1011 continue;
1012
1013 bool AllDead = true;
1014 SmallVector<unsigned,2> Regs;
1015 for (auto &Op : MI->operands()) {
1016 if (!Op.isReg() || !Op.isDef())
1017 continue;
1018 Register R = Op.getReg();
1019 if (!Register::isVirtualRegister(R) || !isDead(R)) {
1020 AllDead = false;
1021 break;
1022 }
1023 Regs.push_back(R);
1024 }
1025 if (!AllDead)
1026 continue;
1027
1028 B->erase(MI);
1029 for (unsigned i = 0, n = Regs.size(); i != n; ++i)
1030 MRI.markUsesInDebugValueAsUndef(Regs[i]);
1031 Changed = true;
1032 }
1033
1034 return Changed;
1035 }
1036
1037 namespace {
1038
1039 // Eliminate redundant instructions
1040 //
1041 // This transformation will identify instructions where the output register
1042 // is the same as one of its input registers. This only works on instructions
1043 // that define a single register (unlike post-increment loads, for example).
1044 // The equality check is actually more detailed: the code calculates which
1045 // bits of the output are used, and only compares these bits with the input
1046 // registers.
1047 // If the output matches an input, the instruction is replaced with COPY.
1048 // The copies will be removed by another transformation.
1049 class RedundantInstrElimination : public Transformation {
1050 public:
1051 RedundantInstrElimination(BitTracker &bt, const HexagonInstrInfo &hii,
1052 const HexagonRegisterInfo &hri, MachineRegisterInfo &mri)
1053 : Transformation(true), HII(hii), HRI(hri), MRI(mri), BT(bt) {}
1054
1055 bool processBlock(MachineBasicBlock &B, const RegisterSet &AVs) override;
1056
1057 private:
1058 bool isLossyShiftLeft(const MachineInstr &MI, unsigned OpN,
1059 unsigned &LostB, unsigned &LostE);
1060 bool isLossyShiftRight(const MachineInstr &MI, unsigned OpN,
1061 unsigned &LostB, unsigned &LostE);
1062 bool computeUsedBits(unsigned Reg, BitVector &Bits);
1063 bool computeUsedBits(const MachineInstr &MI, unsigned OpN, BitVector &Bits,
1064 uint16_t Begin);
1065 bool usedBitsEqual(BitTracker::RegisterRef RD, BitTracker::RegisterRef RS);
1066
1067 const HexagonInstrInfo &HII;
1068 const HexagonRegisterInfo &HRI;
1069 MachineRegisterInfo &MRI;
1070 BitTracker &BT;
1071 };
1072
1073 } // end anonymous namespace
1074
1075 // Check if the instruction is a lossy shift left, where the input being
1076 // shifted is the operand OpN of MI. If true, [LostB, LostE) is the range
1077 // of bit indices that are lost.
1078 bool RedundantInstrElimination::isLossyShiftLeft(const MachineInstr &MI,
1079 unsigned OpN, unsigned &LostB, unsigned &LostE) {
1080 using namespace Hexagon;
1081
1082 unsigned Opc = MI.getOpcode();
1083 unsigned ImN, RegN, Width;
1084 switch (Opc) {
1085 case S2_asl_i_p:
1086 ImN = 2;
1087 RegN = 1;
1088 Width = 64;
1089 break;
1090 case S2_asl_i_p_acc:
1091 case S2_asl_i_p_and:
1092 case S2_asl_i_p_nac:
1093 case S2_asl_i_p_or:
1094 case S2_asl_i_p_xacc:
1095 ImN = 3;
1096 RegN = 2;
1097 Width = 64;
1098 break;
1099 case S2_asl_i_r:
1100 ImN = 2;
1101 RegN = 1;
1102 Width = 32;
1103 break;
1104 case S2_addasl_rrri:
1105 case S4_andi_asl_ri:
1106 case S4_ori_asl_ri:
1107 case S4_addi_asl_ri:
1108 case S4_subi_asl_ri:
1109 case S2_asl_i_r_acc:
1110 case S2_asl_i_r_and:
1111 case S2_asl_i_r_nac:
1112 case S2_asl_i_r_or:
1113 case S2_asl_i_r_sat:
1114 case S2_asl_i_r_xacc:
1115 ImN = 3;
1116 RegN = 2;
1117 Width = 32;
1118 break;
1119 default:
1120 return false;
1121 }
1122
1123 if (RegN != OpN)
1124 return false;
1125
1126 assert(MI.getOperand(ImN).isImm());
1127 unsigned S = MI.getOperand(ImN).getImm();
1128 if (S == 0)
1129 return false;
1130 LostB = Width-S;
1131 LostE = Width;
1132 return true;
1133 }
1134
1135 // Check if the instruction is a lossy shift right, where the input being
1136 // shifted is the operand OpN of MI. If true, [LostB, LostE) is the range
1137 // of bit indices that are lost.
1138 bool RedundantInstrElimination::isLossyShiftRight(const MachineInstr &MI,
1139 unsigned OpN, unsigned &LostB, unsigned &LostE) {
1140 using namespace Hexagon;
1141
1142 unsigned Opc = MI.getOpcode();
1143 unsigned ImN, RegN;
1144 switch (Opc) {
1145 case S2_asr_i_p:
1146 case S2_lsr_i_p:
1147 ImN = 2;
1148 RegN = 1;
1149 break;
1150 case S2_asr_i_p_acc:
1151 case S2_asr_i_p_and:
1152 case S2_asr_i_p_nac:
1153 case S2_asr_i_p_or:
1154 case S2_lsr_i_p_acc:
1155 case S2_lsr_i_p_and:
1156 case S2_lsr_i_p_nac:
1157 case S2_lsr_i_p_or:
1158 case S2_lsr_i_p_xacc:
1159 ImN = 3;
1160 RegN = 2;
1161 break;
1162 case S2_asr_i_r:
1163 case S2_lsr_i_r:
1164 ImN = 2;
1165 RegN = 1;
1166 break;
1167 case S4_andi_lsr_ri:
1168 case S4_ori_lsr_ri:
1169 case S4_addi_lsr_ri:
1170 case S4_subi_lsr_ri:
1171 case S2_asr_i_r_acc:
1172 case S2_asr_i_r_and:
1173 case S2_asr_i_r_nac:
1174 case S2_asr_i_r_or:
1175 case S2_lsr_i_r_acc:
1176 case S2_lsr_i_r_and:
1177 case S2_lsr_i_r_nac:
1178 case S2_lsr_i_r_or:
1179 case S2_lsr_i_r_xacc:
1180 ImN = 3;
1181 RegN = 2;
1182 break;
1183
1184 default:
1185 return false;
1186 }
1187
1188 if (RegN != OpN)
1189 return false;
1190
1191 assert(MI.getOperand(ImN).isImm());
1192 unsigned S = MI.getOperand(ImN).getImm();
1193 LostB = 0;
1194 LostE = S;
1195 return true;
1196 }
1197
1198 // Calculate the bit vector that corresponds to the used bits of register Reg.
1199 // The vector Bits has the same size, as the size of Reg in bits. If the cal-
1200 // culation fails (i.e. the used bits are unknown), it returns false. Other-
1201 // wise, it returns true and sets the corresponding bits in Bits.
1202 bool RedundantInstrElimination::computeUsedBits(unsigned Reg, BitVector &Bits) {
1203 BitVector Used(Bits.size());
1204 RegisterSet Visited;
1205 std::vector<unsigned> Pending;
1206 Pending.push_back(Reg);
1207
1208 for (unsigned i = 0; i < Pending.size(); ++i) {
1209 unsigned R = Pending[i];
1210 if (Visited.has(R))
1211 continue;
1212 Visited.insert(R);
1213 for (auto I = MRI.use_begin(R), E = MRI.use_end(); I != E; ++I) {
1214 BitTracker::RegisterRef UR = *I;
1215 unsigned B, W;
1216 if (!HBS::getSubregMask(UR, B, W, MRI))
1217 return false;
1218 MachineInstr &UseI = *I->getParent();
1219 if (UseI.isPHI() || UseI.isCopy()) {
1220 Register DefR = UseI.getOperand(0).getReg();
1221 if (!Register::isVirtualRegister(DefR))
1222 return false;
1223 Pending.push_back(DefR);
1224 } else {
1225 if (!computeUsedBits(UseI, I.getOperandNo(), Used, B))
1226 return false;
1227 }
1228 }
1229 }
1230 Bits |= Used;
1231 return true;
1232 }
1233
1234 // Calculate the bits used by instruction MI in a register in operand OpN.
1235 // Return true/false if the calculation succeeds/fails. If is succeeds, set
1236 // used bits in Bits. This function does not reset any bits in Bits, so
1237 // subsequent calls over different instructions will result in the union
1238 // of the used bits in all these instructions.
1239 // The register in question may be used with a sub-register, whereas Bits
1240 // holds the bits for the entire register. To keep track of that, the
1241 // argument Begin indicates where in Bits is the lowest-significant bit
1242 // of the register used in operand OpN. For example, in instruction:
1243 // %1 = S2_lsr_i_r %2:isub_hi, 10
1244 // the operand 1 is a 32-bit register, which happens to be a subregister
1245 // of the 64-bit register %2, and that subregister starts at position 32.
1246 // In this case Begin=32, since Bits[32] would be the lowest-significant bit
1247 // of %2:isub_hi.
1248 bool RedundantInstrElimination::computeUsedBits(const MachineInstr &MI,
1249 unsigned OpN, BitVector &Bits, uint16_t Begin) {
1250 unsigned Opc = MI.getOpcode();
1251 BitVector T(Bits.size());
1252 bool GotBits = HBS::getUsedBits(Opc, OpN, T, Begin, HII);
1253 // Even if we don't have bits yet, we could still provide some information
1254 // if the instruction is a lossy shift: the lost bits will be marked as
1255 // not used.
1256 unsigned LB, LE;
1257 if (isLossyShiftLeft(MI, OpN, LB, LE) || isLossyShiftRight(MI, OpN, LB, LE)) {
1258 assert(MI.getOperand(OpN).isReg());
1259 BitTracker::RegisterRef RR = MI.getOperand(OpN);
1260 const TargetRegisterClass *RC = HBS::getFinalVRegClass(RR, MRI);
1261 uint16_t Width = HRI.getRegSizeInBits(*RC);
1262
1263 if (!GotBits)
1264 T.set(Begin, Begin+Width);
1265 assert(LB <= LE && LB < Width && LE <= Width);
1266 T.reset(Begin+LB, Begin+LE);
1267 GotBits = true;
1268 }
1269 if (GotBits)
1270 Bits |= T;
1271 return GotBits;
1272 }
1273
1274 // Calculates the used bits in RD ("defined register"), and checks if these
1275 // bits in RS ("used register") and RD are identical.
1276 bool RedundantInstrElimination::usedBitsEqual(BitTracker::RegisterRef RD,
1277 BitTracker::RegisterRef RS) {
1278 const BitTracker::RegisterCell &DC = BT.lookup(RD.Reg);
1279 const BitTracker::RegisterCell &SC = BT.lookup(RS.Reg);
1280
1281 unsigned DB, DW;
1282 if (!HBS::getSubregMask(RD, DB, DW, MRI))
1283 return false;
1284 unsigned SB, SW;
1285 if (!HBS::getSubregMask(RS, SB, SW, MRI))
1286 return false;
1287 if (SW != DW)
1288 return false;
1289
1290 BitVector Used(DC.width());
1291 if (!computeUsedBits(RD.Reg, Used))
1292 return false;
1293
1294 for (unsigned i = 0; i != DW; ++i)
1295 if (Used[i+DB] && DC[DB+i] != SC[SB+i])
1296 return false;
1297 return true;
1298 }
1299
1300 bool RedundantInstrElimination::processBlock(MachineBasicBlock &B,
1301 const RegisterSet&) {
1302 if (!BT.reached(&B))
1303 return false;
1304 bool Changed = false;
1305
1306 for (auto I = B.begin(), E = B.end(), NextI = I; I != E; ++I) {
1307 NextI = std::next(I);
1308 MachineInstr *MI = &*I;
1309
1310 if (MI->getOpcode() == TargetOpcode::COPY)
1311 continue;
1312 if (MI->isPHI() || MI->hasUnmodeledSideEffects() || MI->isInlineAsm())
1313 continue;
1314 unsigned NumD = MI->getDesc().getNumDefs();
1315 if (NumD != 1)
1316 continue;
1317
1318 BitTracker::RegisterRef RD = MI->getOperand(0);
1319 if (!BT.has(RD.Reg))
1320 continue;
1321 const BitTracker::RegisterCell &DC = BT.lookup(RD.Reg);
1322 auto At = MachineBasicBlock::iterator(MI);
1323
1324 // Find a source operand that is equal to the result.
1325 for (auto &Op : MI->uses()) {
1326 if (!Op.isReg())
1327 continue;
1328 BitTracker::RegisterRef RS = Op;
1329 if (!BT.has(RS.Reg))
1330 continue;
1331 if (!HBS::isTransparentCopy(RD, RS, MRI))
1332 continue;
1333
1334 unsigned BN, BW;
1335 if (!HBS::getSubregMask(RS, BN, BW, MRI))
1336 continue;
1337
1338 const BitTracker::RegisterCell &SC = BT.lookup(RS.Reg);
1339 if (!usedBitsEqual(RD, RS) && !HBS::isEqual(DC, 0, SC, BN, BW))
1340 continue;
1341
1342 // If found, replace the instruction with a COPY.
1343 const DebugLoc &DL = MI->getDebugLoc();
1344 const TargetRegisterClass *FRC = HBS::getFinalVRegClass(RD, MRI);
1345 Register NewR = MRI.createVirtualRegister(FRC);
1346 MachineInstr *CopyI =
1347 BuildMI(B, At, DL, HII.get(TargetOpcode::COPY), NewR)
1348 .addReg(RS.Reg, 0, RS.Sub);
1349 HBS::replaceSubWithSub(RD.Reg, RD.Sub, NewR, 0, MRI);
1350 // This pass can create copies between registers that don't have the
1351 // exact same values. Updating the tracker has to involve updating
1352 // all dependent cells. Example:
1353 // %1 = inst %2 ; %1 != %2, but used bits are equal
1354 //
1355 // %3 = copy %2 ; <- inserted
1356 // ... = %3 ; <- replaced from %2
1357 // Indirectly, we can create a "copy" between %1 and %2 even
1358 // though their exact values do not match.
1359 BT.visit(*CopyI);
1360 Changed = true;
1361 break;
1362 }
1363 }
1364
1365 return Changed;
1366 }
1367
1368 namespace {
1369
1370 // Recognize instructions that produce constant values known at compile-time.
1371 // Replace them with register definitions that load these constants directly.
1372 class ConstGeneration : public Transformation {
1373 public:
1374 ConstGeneration(BitTracker &bt, const HexagonInstrInfo &hii,
1375 MachineRegisterInfo &mri)
1376 : Transformation(true), HII(hii), MRI(mri), BT(bt) {}
1377
1378 bool processBlock(MachineBasicBlock &B, const RegisterSet &AVs) override;
1379 static bool isTfrConst(const MachineInstr &MI);
1380
1381 private:
1382 unsigned genTfrConst(const TargetRegisterClass *RC, int64_t C,
1383 MachineBasicBlock &B, MachineBasicBlock::iterator At, DebugLoc &DL);
1384
1385 const HexagonInstrInfo &HII;
1386 MachineRegisterInfo &MRI;
1387 BitTracker &BT;
1388 };
1389
1390 } // end anonymous namespace
1391
1392 bool ConstGeneration::isTfrConst(const MachineInstr &MI) {
1393 unsigned Opc = MI.getOpcode();
1394 switch (Opc) {
1395 case Hexagon::A2_combineii:
1396 case Hexagon::A4_combineii:
1397 case Hexagon::A2_tfrsi:
1398 case Hexagon::A2_tfrpi:
1399 case Hexagon::PS_true:
1400 case Hexagon::PS_false:
1401 case Hexagon::CONST32:
1402 case Hexagon::CONST64:
1403 return true;
1404 }
1405 return false;
1406 }
1407
1408 // Generate a transfer-immediate instruction that is appropriate for the
1409 // register class and the actual value being transferred.
1410 unsigned ConstGeneration::genTfrConst(const TargetRegisterClass *RC, int64_t C,
1411 MachineBasicBlock &B, MachineBasicBlock::iterator At, DebugLoc &DL) {
1412 Register Reg = MRI.createVirtualRegister(RC);
1413 if (RC == &Hexagon::IntRegsRegClass) {
1414 BuildMI(B, At, DL, HII.get(Hexagon::A2_tfrsi), Reg)
1415 .addImm(int32_t(C));
1416 return Reg;
1417 }
1418
1419 if (RC == &Hexagon::DoubleRegsRegClass) {
1420 if (isInt<8>(C)) {
1421 BuildMI(B, At, DL, HII.get(Hexagon::A2_tfrpi), Reg)
1422 .addImm(C);
1423 return Reg;
1424 }
1425
1426 unsigned Lo = Lo_32(C), Hi = Hi_32(C);
1427 if (isInt<8>(Lo) || isInt<8>(Hi)) {
1428 unsigned Opc = isInt<8>(Lo) ? Hexagon::A2_combineii
1429 : Hexagon::A4_combineii;
1430 BuildMI(B, At, DL, HII.get(Opc), Reg)
1431 .addImm(int32_t(Hi))
1432 .addImm(int32_t(Lo));
1433 return Reg;
1434 }
1435
1436 BuildMI(B, At, DL, HII.get(Hexagon::CONST64), Reg)
1437 .addImm(C);
1438 return Reg;
1439 }
1440
1441 if (RC == &Hexagon::PredRegsRegClass) {
1442 unsigned Opc;
1443 if (C == 0)
1444 Opc = Hexagon::PS_false;
1445 else if ((C & 0xFF) == 0xFF)
1446 Opc = Hexagon::PS_true;
1447 else
1448 return 0;
1449 BuildMI(B, At, DL, HII.get(Opc), Reg);
1450 return Reg;
1451 }
1452
1453 return 0;
1454 }
1455
1456 bool ConstGeneration::processBlock(MachineBasicBlock &B, const RegisterSet&) {
1457 if (!BT.reached(&B))
1458 return false;
1459 bool Changed = false;
1460 RegisterSet Defs;
1461
1462 for (auto I = B.begin(), E = B.end(); I != E; ++I) {
1463 if (isTfrConst(*I))
1464 continue;
1465 Defs.clear();
1466 HBS::getInstrDefs(*I, Defs);
1467 if (Defs.count() != 1)
1468 continue;
1469 unsigned DR = Defs.find_first();
1470 if (!Register::isVirtualRegister(DR))
1471 continue;
1472 uint64_t U;
1473 const BitTracker::RegisterCell &DRC = BT.lookup(DR);
1474 if (HBS::getConst(DRC, 0, DRC.width(), U)) {
1475 int64_t C = U;
1476 DebugLoc DL = I->getDebugLoc();
1477 auto At = I->isPHI() ? B.getFirstNonPHI() : I;
1478 unsigned ImmReg = genTfrConst(MRI.getRegClass(DR), C, B, At, DL);
1479 if (ImmReg) {
1480 HBS::replaceReg(DR, ImmReg, MRI);
1481 BT.put(ImmReg, DRC);
1482 Changed = true;
1483 }
1484 }
1485 }
1486 return Changed;
1487 }
1488
1489 namespace {
1490
1491 // Identify pairs of available registers which hold identical values.
1492 // In such cases, only one of them needs to be calculated, the other one
1493 // will be defined as a copy of the first.
1494 class CopyGeneration : public Transformation {
1495 public:
1496 CopyGeneration(BitTracker &bt, const HexagonInstrInfo &hii,
1497 const HexagonRegisterInfo &hri, MachineRegisterInfo &mri)
1498 : Transformation(true), HII(hii), HRI(hri), MRI(mri), BT(bt) {}
1499
1500 bool processBlock(MachineBasicBlock &B, const RegisterSet &AVs) override;
1501
1502 private:
1503 bool findMatch(const BitTracker::RegisterRef &Inp,
1504 BitTracker::RegisterRef &Out, const RegisterSet &AVs);
1505
1506 const HexagonInstrInfo &HII;
1507 const HexagonRegisterInfo &HRI;
1508 MachineRegisterInfo &MRI;
1509 BitTracker &BT;
1510 RegisterSet Forbidden;
1511 };
1512
1513 // Eliminate register copies RD = RS, by replacing the uses of RD with
1514 // with uses of RS.
1515 class CopyPropagation : public Transformation {
1516 public:
1517 CopyPropagation(const HexagonRegisterInfo &hri, MachineRegisterInfo &mri)
1518 : Transformation(false), HRI(hri), MRI(mri) {}
1519
1520 bool processBlock(MachineBasicBlock &B, const RegisterSet &AVs) override;
1521
1522 static bool isCopyReg(unsigned Opc, bool NoConv);
1523
1524 private:
1525 bool propagateRegCopy(MachineInstr &MI);
1526
1527 const HexagonRegisterInfo &HRI;
1528 MachineRegisterInfo &MRI;
1529 };
1530
1531 } // end anonymous namespace
1532
1533 /// Check if there is a register in AVs that is identical to Inp. If so,
1534 /// set Out to the found register. The output may be a pair Reg:Sub.
1535 bool CopyGeneration::findMatch(const BitTracker::RegisterRef &Inp,
1536 BitTracker::RegisterRef &Out, const RegisterSet &AVs) {
1537 if (!BT.has(Inp.Reg))
1538 return false;
1539 const BitTracker::RegisterCell &InpRC = BT.lookup(Inp.Reg);
1540 auto *FRC = HBS::getFinalVRegClass(Inp, MRI);
1541 unsigned B, W;
1542 if (!HBS::getSubregMask(Inp, B, W, MRI))
1543 return false;
1544
1545 for (unsigned R = AVs.find_first(); R; R = AVs.find_next(R)) {
1546 if (!BT.has(R) || Forbidden[R])
1547 continue;
1548 const BitTracker::RegisterCell &RC = BT.lookup(R);
1549 unsigned RW = RC.width();
1550 if (W == RW) {
1551 if (FRC != MRI.getRegClass(R))
1552 continue;
1553 if (!HBS::isTransparentCopy(R, Inp, MRI))
1554 continue;
1555 if (!HBS::isEqual(InpRC, B, RC, 0, W))
1556 continue;
1557 Out.Reg = R;
1558 Out.Sub = 0;
1559 return true;
1560 }
1561 // Check if there is a super-register, whose part (with a subregister)
1562 // is equal to the input.
1563 // Only do double registers for now.
1564 if (W*2 != RW)
1565 continue;
1566 if (MRI.getRegClass(R) != &Hexagon::DoubleRegsRegClass)
1567 continue;
1568
1569 if (HBS::isEqual(InpRC, B, RC, 0, W))
1570 Out.Sub = Hexagon::isub_lo;
1571 else if (HBS::isEqual(InpRC, B, RC, W, W))
1572 Out.Sub = Hexagon::isub_hi;
1573 else
1574 continue;
1575 Out.Reg = R;
1576 if (HBS::isTransparentCopy(Out, Inp, MRI))
1577 return true;
1578 }
1579 return false;
1580 }
1581
1582 bool CopyGeneration::processBlock(MachineBasicBlock &B,
1583 const RegisterSet &AVs) {
1584 if (!BT.reached(&B))
1585 return false;
1586 RegisterSet AVB(AVs);
1587 bool Changed = false;
1588 RegisterSet Defs;
1589
1590 for (auto I = B.begin(), E = B.end(), NextI = I; I != E;
1591 ++I, AVB.insert(Defs)) {
1592 NextI = std::next(I);
1593 Defs.clear();
1594 HBS::getInstrDefs(*I, Defs);
1595
1596 unsigned Opc = I->getOpcode();
1597 if (CopyPropagation::isCopyReg(Opc, false) ||
1598 ConstGeneration::isTfrConst(*I))
1599 continue;
1600
1601 DebugLoc DL = I->getDebugLoc();
1602 auto At = I->isPHI() ? B.getFirstNonPHI() : I;
1603
1604 for (unsigned R = Defs.find_first(); R; R = Defs.find_next(R)) {
1605 BitTracker::RegisterRef MR;
1606 auto *FRC = HBS::getFinalVRegClass(R, MRI);
1607
1608 if (findMatch(R, MR, AVB)) {
1609 Register NewR = MRI.createVirtualRegister(FRC);
1610 BuildMI(B, At, DL, HII.get(TargetOpcode::COPY), NewR)
1611 .addReg(MR.Reg, 0, MR.Sub);
1612 BT.put(BitTracker::RegisterRef(NewR), BT.get(MR));
1613 HBS::replaceReg(R, NewR, MRI);
1614 Forbidden.insert(R);
1615 continue;
1616 }
1617
1618 if (FRC == &Hexagon::DoubleRegsRegClass ||
1619 FRC == &Hexagon::HvxWRRegClass) {
1620 // Try to generate REG_SEQUENCE.
1621 unsigned SubLo = HRI.getHexagonSubRegIndex(*FRC, Hexagon::ps_sub_lo);
1622 unsigned SubHi = HRI.getHexagonSubRegIndex(*FRC, Hexagon::ps_sub_hi);
1623 BitTracker::RegisterRef TL = { R, SubLo };
1624 BitTracker::RegisterRef TH = { R, SubHi };
1625 BitTracker::RegisterRef ML, MH;
1626 if (findMatch(TL, ML, AVB) && findMatch(TH, MH, AVB)) {
1627 auto *FRC = HBS::getFinalVRegClass(R, MRI);
1628 Register NewR = MRI.createVirtualRegister(FRC);
1629 BuildMI(B, At, DL, HII.get(TargetOpcode::REG_SEQUENCE), NewR)
1630 .addReg(ML.Reg, 0, ML.Sub)
1631 .addImm(SubLo)
1632 .addReg(MH.Reg, 0, MH.Sub)
1633 .addImm(SubHi);
1634 BT.put(BitTracker::RegisterRef(NewR), BT.get(R));
1635 HBS::replaceReg(R, NewR, MRI);
1636 Forbidden.insert(R);
1637 }
1638 }
1639 }
1640 }
1641
1642 return Changed;
1643 }
1644
1645 bool CopyPropagation::isCopyReg(unsigned Opc, bool NoConv) {
1646 switch (Opc) {
1647 case TargetOpcode::COPY:
1648 case TargetOpcode::REG_SEQUENCE:
1649 case Hexagon::A4_combineir:
1650 case Hexagon::A4_combineri:
1651 return true;
1652 case Hexagon::A2_tfr:
1653 case Hexagon::A2_tfrp:
1654 case Hexagon::A2_combinew:
1655 case Hexagon::V6_vcombine:
1656 return NoConv;
1657 default:
1658 break;
1659 }
1660 return false;
1661 }
1662
1663 bool CopyPropagation::propagateRegCopy(MachineInstr &MI) {
1664 bool Changed = false;
1665 unsigned Opc = MI.getOpcode();
1666 BitTracker::RegisterRef RD = MI.getOperand(0);
1667 assert(MI.getOperand(0).getSubReg() == 0);
1668
1669 switch (Opc) {
1670 case TargetOpcode::COPY:
1671 case Hexagon::A2_tfr:
1672 case Hexagon::A2_tfrp: {
1673 BitTracker::RegisterRef RS = MI.getOperand(1);
1674 if (!HBS::isTransparentCopy(RD, RS, MRI))
1675 break;
1676 if (RS.Sub != 0)
1677 Changed = HBS::replaceRegWithSub(RD.Reg, RS.Reg, RS.Sub, MRI);
1678 else
1679 Changed = HBS::replaceReg(RD.Reg, RS.Reg, MRI);
1680 break;
1681 }
1682 case TargetOpcode::REG_SEQUENCE: {
1683 BitTracker::RegisterRef SL, SH;
1684 if (HBS::parseRegSequence(MI, SL, SH, MRI)) {
1685 const TargetRegisterClass &RC = *MRI.getRegClass(RD.Reg);
1686 unsigned SubLo = HRI.getHexagonSubRegIndex(RC, Hexagon::ps_sub_lo);
1687 unsigned SubHi = HRI.getHexagonSubRegIndex(RC, Hexagon::ps_sub_hi);
1688 Changed = HBS::replaceSubWithSub(RD.Reg, SubLo, SL.Reg, SL.Sub, MRI);
1689 Changed |= HBS::replaceSubWithSub(RD.Reg, SubHi, SH.Reg, SH.Sub, MRI);
1690 }
1691 break;
1692 }
1693 case Hexagon::A2_combinew:
1694 case Hexagon::V6_vcombine: {
1695 const TargetRegisterClass &RC = *MRI.getRegClass(RD.Reg);
1696 unsigned SubLo = HRI.getHexagonSubRegIndex(RC, Hexagon::ps_sub_lo);
1697 unsigned SubHi = HRI.getHexagonSubRegIndex(RC, Hexagon::ps_sub_hi);
1698 BitTracker::RegisterRef RH = MI.getOperand(1), RL = MI.getOperand(2);
1699 Changed = HBS::replaceSubWithSub(RD.Reg, SubLo, RL.Reg, RL.Sub, MRI);
1700 Changed |= HBS::replaceSubWithSub(RD.Reg, SubHi, RH.Reg, RH.Sub, MRI);
1701 break;
1702 }
1703 case Hexagon::A4_combineir:
1704 case Hexagon::A4_combineri: {
1705 unsigned SrcX = (Opc == Hexagon::A4_combineir) ? 2 : 1;
1706 unsigned Sub = (Opc == Hexagon::A4_combineir) ? Hexagon::isub_lo
1707 : Hexagon::isub_hi;
1708 BitTracker::RegisterRef RS = MI.getOperand(SrcX);
1709 Changed = HBS::replaceSubWithSub(RD.Reg, Sub, RS.Reg, RS.Sub, MRI);
1710 break;
1711 }
1712 }
1713 return Changed;
1714 }
1715
1716 bool CopyPropagation::processBlock(MachineBasicBlock &B, const RegisterSet&) {
1717 std::vector<MachineInstr*> Instrs;
1718 for (auto I = B.rbegin(), E = B.rend(); I != E; ++I)
1719 Instrs.push_back(&*I);
1720
1721 bool Changed = false;
1722 for (auto I : Instrs) {
1723 unsigned Opc = I->getOpcode();
1724 if (!CopyPropagation::isCopyReg(Opc, true))
1725 continue;
1726 Changed |= propagateRegCopy(*I);
1727 }
1728
1729 return Changed;
1730 }
1731
1732 namespace {
1733
1734 // Recognize patterns that can be simplified and replace them with the
1735 // simpler forms.
1736 // This is by no means complete
1737 class BitSimplification : public Transformation {
1738 public:
1739 BitSimplification(BitTracker &bt, const MachineDominatorTree &mdt,
1740 const HexagonInstrInfo &hii, const HexagonRegisterInfo &hri,
1741 MachineRegisterInfo &mri, MachineFunction &mf)
1742 : Transformation(true), MDT(mdt), HII(hii), HRI(hri), MRI(mri),
1743 MF(mf), BT(bt) {}
1744
1745 bool processBlock(MachineBasicBlock &B, const RegisterSet &AVs) override;
1746
1747 private:
1748 struct RegHalf : public BitTracker::RegisterRef {
1749 bool Low; // Low/High halfword.
1750 };
1751
1752 bool matchHalf(unsigned SelfR, const BitTracker::RegisterCell &RC,
1753 unsigned B, RegHalf &RH);
1754 bool validateReg(BitTracker::RegisterRef R, unsigned Opc, unsigned OpNum);
1755
1756 bool matchPackhl(unsigned SelfR, const BitTracker::RegisterCell &RC,
1757 BitTracker::RegisterRef &Rs, BitTracker::RegisterRef &Rt);
1758 unsigned getCombineOpcode(bool HLow, bool LLow);
1759
1760 bool genStoreUpperHalf(MachineInstr *MI);
1761 bool genStoreImmediate(MachineInstr *MI);
1762 bool genPackhl(MachineInstr *MI, BitTracker::RegisterRef RD,
1763 const BitTracker::RegisterCell &RC);
1764 bool genExtractHalf(MachineInstr *MI, BitTracker::RegisterRef RD,
1765 const BitTracker::RegisterCell &RC);
1766 bool genCombineHalf(MachineInstr *MI, BitTracker::RegisterRef RD,
1767 const BitTracker::RegisterCell &RC);
1768 bool genExtractLow(MachineInstr *MI, BitTracker::RegisterRef RD,
1769 const BitTracker::RegisterCell &RC);
1770 bool genBitSplit(MachineInstr *MI, BitTracker::RegisterRef RD,
1771 const BitTracker::RegisterCell &RC, const RegisterSet &AVs);
1772 bool simplifyTstbit(MachineInstr *MI, BitTracker::RegisterRef RD,
1773 const BitTracker::RegisterCell &RC);
1774 bool simplifyExtractLow(MachineInstr *MI, BitTracker::RegisterRef RD,
1775 const BitTracker::RegisterCell &RC, const RegisterSet &AVs);
1776 bool simplifyRCmp0(MachineInstr *MI, BitTracker::RegisterRef RD);
1777
1778 // Cache of created instructions to avoid creating duplicates.
1779 // XXX Currently only used by genBitSplit.
1780 std::vector<MachineInstr*> NewMIs;
1781
1782 const MachineDominatorTree &MDT;
1783 const HexagonInstrInfo &HII;
1784 const HexagonRegisterInfo &HRI;
1785 MachineRegisterInfo &MRI;
1786 MachineFunction &MF;
1787 BitTracker &BT;
1788 };
1789
1790 } // end anonymous namespace
1791
1792 // Check if the bits [B..B+16) in register cell RC form a valid halfword,
1793 // i.e. [0..16), [16..32), etc. of some register. If so, return true and
1794 // set the information about the found register in RH.
1795 bool BitSimplification::matchHalf(unsigned SelfR,
1796 const BitTracker::RegisterCell &RC, unsigned B, RegHalf &RH) {
1797 // XXX This could be searching in the set of available registers, in case
1798 // the match is not exact.
1799
1800 // Match 16-bit chunks, where the RC[B..B+15] references exactly one
1801 // register and all the bits B..B+15 match between RC and the register.
1802 // This is meant to match "v1[0-15]", where v1 = { [0]:0 [1-15]:v1... },
1803 // and RC = { [0]:0 [1-15]:v1[1-15]... }.
1804 bool Low = false;
1805 unsigned I = B;
1806 while (I < B+16 && RC[I].num())
1807 I++;
1808 if (I == B+16)
1809 return false;
1810
1811 unsigned Reg = RC[I].RefI.Reg;
1812 unsigned P = RC[I].RefI.Pos; // The RefI.Pos will be advanced by I-B.
1813 if (P < I-B)
1814 return false;
1815 unsigned Pos = P - (I-B);
1816
1817 if (Reg == 0 || Reg == SelfR) // Don't match "self".
1818 return false;
1819 if (!Register::isVirtualRegister(Reg))
1820 return false;
1821 if (!BT.has(Reg))
1822 return false;
1823
1824 const BitTracker::RegisterCell &SC = BT.lookup(Reg);
1825 if (Pos+16 > SC.width())
1826 return false;
1827
1828 for (unsigned i = 0; i < 16; ++i) {
1829 const BitTracker::BitValue &RV = RC[i+B];
1830 if (RV.Type == BitTracker::BitValue::Ref) {
1831 if (RV.RefI.Reg != Reg)
1832 return false;
1833 if (RV.RefI.Pos != i+Pos)
1834 return false;
1835 continue;
1836 }
1837 if (RC[i+B] != SC[i+Pos])
1838 return false;
1839 }
1840
1841 unsigned Sub = 0;
1842 switch (Pos) {
1843 case 0:
1844 Sub = Hexagon::isub_lo;
1845 Low = true;
1846 break;
1847 case 16:
1848 Sub = Hexagon::isub_lo;
1849 Low = false;
1850 break;
1851 case 32:
1852 Sub = Hexagon::isub_hi;
1853 Low = true;
1854 break;
1855 case 48:
1856 Sub = Hexagon::isub_hi;
1857 Low = false;
1858 break;
1859 default:
1860 return false;
1861 }
1862
1863 RH.Reg = Reg;
1864 RH.Sub = Sub;
1865 RH.Low = Low;
1866 // If the subregister is not valid with the register, set it to 0.
1867 if (!HBS::getFinalVRegClass(RH, MRI))
1868 RH.Sub = 0;
1869
1870 return true;
1871 }
1872
1873 bool BitSimplification::validateReg(BitTracker::RegisterRef R, unsigned Opc,
1874 unsigned OpNum) {
1875 auto *OpRC = HII.getRegClass(HII.get(Opc), OpNum, &HRI, MF);
1876 auto *RRC = HBS::getFinalVRegClass(R, MRI);
1877 return OpRC->hasSubClassEq(RRC);
1878 }
1879
1880 // Check if RC matches the pattern of a S2_packhl. If so, return true and
1881 // set the inputs Rs and Rt.
1882 bool BitSimplification::matchPackhl(unsigned SelfR,
1883 const BitTracker::RegisterCell &RC, BitTracker::RegisterRef &Rs,
1884 BitTracker::RegisterRef &Rt) {
1885 RegHalf L1, H1, L2, H2;
1886
1887 if (!matchHalf(SelfR, RC, 0, L2) || !matchHalf(SelfR, RC, 16, L1))
1888 return false;
1889 if (!matchHalf(SelfR, RC, 32, H2) || !matchHalf(SelfR, RC, 48, H1))
1890 return false;
1891
1892 // Rs = H1.L1, Rt = H2.L2
1893 if (H1.Reg != L1.Reg || H1.Sub != L1.Sub || H1.Low || !L1.Low)
1894 return false;
1895 if (H2.Reg != L2.Reg || H2.Sub != L2.Sub || H2.Low || !L2.Low)
1896 return false;
1897
1898 Rs = H1;
1899 Rt = H2;
1900 return true;
1901 }
1902
1903 unsigned BitSimplification::getCombineOpcode(bool HLow, bool LLow) {
1904 return HLow ? LLow ? Hexagon::A2_combine_ll
1905 : Hexagon::A2_combine_lh
1906 : LLow ? Hexagon::A2_combine_hl
1907 : Hexagon::A2_combine_hh;
1908 }
1909
1910 // If MI stores the upper halfword of a register (potentially obtained via
1911 // shifts or extracts), replace it with a storerf instruction. This could
1912 // cause the "extraction" code to become dead.
1913 bool BitSimplification::genStoreUpperHalf(MachineInstr *MI) {
1914 unsigned Opc = MI->getOpcode();
1915 if (Opc != Hexagon::S2_storerh_io)
1916 return false;
1917
1918 MachineOperand &ValOp = MI->getOperand(2);
1919 BitTracker::RegisterRef RS = ValOp;
1920 if (!BT.has(RS.Reg))
1921 return false;
1922 const BitTracker::RegisterCell &RC = BT.lookup(RS.Reg);
1923 RegHalf H;
1924 if (!matchHalf(0, RC, 0, H))
1925 return false;
1926 if (H.Low)
1927 return false;
1928 MI->setDesc(HII.get(Hexagon::S2_storerf_io));
1929 ValOp.setReg(H.Reg);
1930 ValOp.setSubReg(H.Sub);
1931 return true;
1932 }
1933
1934 // If MI stores a value known at compile-time, and the value is within a range
1935 // that avoids using constant-extenders, replace it with a store-immediate.
1936 bool BitSimplification::genStoreImmediate(MachineInstr *MI) {
1937 unsigned Opc = MI->getOpcode();
1938 unsigned Align = 0;
1939 switch (Opc) {
1940 case Hexagon::S2_storeri_io:
1941 Align++;
1942 LLVM_FALLTHROUGH;
1943 case Hexagon::S2_storerh_io:
1944 Align++;
1945 LLVM_FALLTHROUGH;
1946 case Hexagon::S2_storerb_io:
1947 break;
1948 default:
1949 return false;
1950 }
1951
1952 // Avoid stores to frame-indices (due to an unknown offset).
1953 if (!MI->getOperand(0).isReg())
1954 return false;
1955 MachineOperand &OffOp = MI->getOperand(1);
1956 if (!OffOp.isImm())
1957 return false;
1958
1959 int64_t Off = OffOp.getImm();
1960 // Offset is u6:a. Sadly, there is no isShiftedUInt(n,x).
1961 if (!isUIntN(6+Align, Off) || (Off & ((1<<Align)-1)))
1962 return false;
1963 // Source register:
1964 BitTracker::RegisterRef RS = MI->getOperand(2);
1965 if (!BT.has(RS.Reg))
1966 return false;
1967 const BitTracker::RegisterCell &RC = BT.lookup(RS.Reg);
1968 uint64_t U;
1969 if (!HBS::getConst(RC, 0, RC.width(), U))
1970 return false;
1971
1972 // Only consider 8-bit values to avoid constant-extenders.
1973 int V;
1974 switch (Opc) {
1975 case Hexagon::S2_storerb_io:
1976 V = int8_t(U);
1977 break;
1978 case Hexagon::S2_storerh_io:
1979 V = int16_t(U);
1980 break;
1981 case Hexagon::S2_storeri_io:
1982 V = int32_t(U);
1983 break;
1984 default:
1985 // Opc is already checked above to be one of the three store instructions.
1986 // This silences a -Wuninitialized false positive on GCC 5.4.
1987 llvm_unreachable("Unexpected store opcode");
1988 }
1989 if (!isInt<8>(V))
1990 return false;
1991
1992 MI->RemoveOperand(2);
1993 switch (Opc) {
1994 case Hexagon::S2_storerb_io:
1995 MI->setDesc(HII.get(Hexagon::S4_storeirb_io));
1996 break;
1997 case Hexagon::S2_storerh_io:
1998 MI->setDesc(HII.get(Hexagon::S4_storeirh_io));
1999 break;
2000 case Hexagon::S2_storeri_io:
2001 MI->setDesc(HII.get(Hexagon::S4_storeiri_io));
2002 break;
2003 }
2004 MI->addOperand(MachineOperand::CreateImm(V));
2005 return true;
2006 }
2007
2008 // If MI is equivalent o S2_packhl, generate the S2_packhl. MI could be the
2009 // last instruction in a sequence that results in something equivalent to
2010 // the pack-halfwords. The intent is to cause the entire sequence to become
2011 // dead.
2012 bool BitSimplification::genPackhl(MachineInstr *MI,
2013 BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC) {
2014 unsigned Opc = MI->getOpcode();
2015 if (Opc == Hexagon::S2_packhl)
2016 return false;
2017 BitTracker::RegisterRef Rs, Rt;
2018 if (!matchPackhl(RD.Reg, RC, Rs, Rt))
2019 return false;
2020 if (!validateReg(Rs, Hexagon::S2_packhl, 1) ||
2021 !validateReg(Rt, Hexagon::S2_packhl, 2))
2022 return false;
2023
2024 MachineBasicBlock &B = *MI->getParent();
2025 Register NewR = MRI.createVirtualRegister(&Hexagon::DoubleRegsRegClass);
2026 DebugLoc DL = MI->getDebugLoc();
2027 auto At = MI->isPHI() ? B.getFirstNonPHI()
2028 : MachineBasicBlock::iterator(MI);
2029 BuildMI(B, At, DL, HII.get(Hexagon::S2_packhl), NewR)
2030 .addReg(Rs.Reg, 0, Rs.Sub)
2031 .addReg(Rt.Reg, 0, Rt.Sub);
2032 HBS::replaceSubWithSub(RD.Reg, RD.Sub, NewR, 0, MRI);
2033 BT.put(BitTracker::RegisterRef(NewR), RC);
2034 return true;
2035 }
2036
2037 // If MI produces halfword of the input in the low half of the output,
2038 // replace it with zero-extend or extractu.
2039 bool BitSimplification::genExtractHalf(MachineInstr *MI,
2040 BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC) {
2041 RegHalf L;
2042 // Check for halfword in low 16 bits, zeros elsewhere.
2043 if (!matchHalf(RD.Reg, RC, 0, L) || !HBS::isZero(RC, 16, 16))
2044 return false;
2045
2046 unsigned Opc = MI->getOpcode();
2047 MachineBasicBlock &B = *MI->getParent();
2048 DebugLoc DL = MI->getDebugLoc();
2049
2050 // Prefer zxth, since zxth can go in any slot, while extractu only in
2051 // slots 2 and 3.
2052 unsigned NewR = 0;
2053 auto At = MI->isPHI() ? B.getFirstNonPHI()
2054 : MachineBasicBlock::iterator(MI);
2055 if (L.Low && Opc != Hexagon::A2_zxth) {
2056 if (validateReg(L, Hexagon::A2_zxth, 1)) {
2057 NewR = MRI.createVirtualRegister(&Hexagon::IntRegsRegClass);
2058 BuildMI(B, At, DL, HII.get(Hexagon::A2_zxth), NewR)
2059 .addReg(L.Reg, 0, L.Sub);
2060 }
2061 } else if (!L.Low && Opc != Hexagon::S2_lsr_i_r) {
2062 if (validateReg(L, Hexagon::S2_lsr_i_r, 1)) {
2063 NewR = MRI.createVirtualRegister(&Hexagon::IntRegsRegClass);
2064 BuildMI(B, MI, DL, HII.get(Hexagon::S2_lsr_i_r), NewR)
2065 .addReg(L.Reg, 0, L.Sub)
2066 .addImm(16);
2067 }
2068 }
2069 if (NewR == 0)
2070 return false;
2071 HBS::replaceSubWithSub(RD.Reg, RD.Sub, NewR, 0, MRI);
2072 BT.put(BitTracker::RegisterRef(NewR), RC);
2073 return true;
2074 }
2075
2076 // If MI is equivalent to a combine(.L/.H, .L/.H) replace with with the
2077 // combine.
2078 bool BitSimplification::genCombineHalf(MachineInstr *MI,
2079 BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC) {
2080 RegHalf L, H;
2081 // Check for combine h/l
2082 if (!matchHalf(RD.Reg, RC, 0, L) || !matchHalf(RD.Reg, RC, 16, H))
2083 return false;
2084 // Do nothing if this is just a reg copy.
2085 if (L.Reg == H.Reg && L.Sub == H.Sub && !H.Low && L.Low)
2086 return false;
2087
2088 unsigned Opc = MI->getOpcode();
2089 unsigned COpc = getCombineOpcode(H.Low, L.Low);
2090 if (COpc == Opc)
2091 return false;
2092 if (!validateReg(H, COpc, 1) || !validateReg(L, COpc, 2))
2093 return false;
2094
2095 MachineBasicBlock &B = *MI->getParent();
2096 DebugLoc DL = MI->getDebugLoc();
2097 Register NewR = MRI.createVirtualRegister(&Hexagon::IntRegsRegClass);
2098 auto At = MI->isPHI() ? B.getFirstNonPHI()
2099 : MachineBasicBlock::iterator(MI);
2100 BuildMI(B, At, DL, HII.get(COpc), NewR)
2101 .addReg(H.Reg, 0, H.Sub)
2102 .addReg(L.Reg, 0, L.Sub);
2103 HBS::replaceSubWithSub(RD.Reg, RD.Sub, NewR, 0, MRI);
2104 BT.put(BitTracker::RegisterRef(NewR), RC);
2105 return true;
2106 }
2107
2108 // If MI resets high bits of a register and keeps the lower ones, replace it
2109 // with zero-extend byte/half, and-immediate, or extractu, as appropriate.
2110 bool BitSimplification::genExtractLow(MachineInstr *MI,
2111 BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC) {
2112 unsigned Opc = MI->getOpcode();
2113 switch (Opc) {
2114 case Hexagon::A2_zxtb:
2115 case Hexagon::A2_zxth:
2116 case Hexagon::S2_extractu:
2117 return false;
2118 }
2119 if (Opc == Hexagon::A2_andir && MI->getOperand(2).isImm()) {
2120 int32_t Imm = MI->getOperand(2).getImm();
2121 if (isInt<10>(Imm))
2122 return false;
2123 }
2124
2125 if (MI->hasUnmodeledSideEffects() || MI->isInlineAsm())
2126 return false;
2127 unsigned W = RC.width();
2128 while (W > 0 && RC[W-1].is(0))
2129 W--;
2130 if (W == 0 || W == RC.width())
2131 return false;
2132 unsigned NewOpc = (W == 8) ? Hexagon::A2_zxtb
2133 : (W == 16) ? Hexagon::A2_zxth
2134 : (W < 10) ? Hexagon::A2_andir
2135 : Hexagon::S2_extractu;
2136 MachineBasicBlock &B = *MI->getParent();
2137 DebugLoc DL = MI->getDebugLoc();
2138
2139 for (auto &Op : MI->uses()) {
2140 if (!Op.isReg())
2141 continue;
2142 BitTracker::RegisterRef RS = Op;
2143 if (!BT.has(RS.Reg))
2144 continue;
2145 const BitTracker::RegisterCell &SC = BT.lookup(RS.Reg);
2146 unsigned BN, BW;
2147 if (!HBS::getSubregMask(RS, BN, BW, MRI))
2148 continue;
2149 if (BW < W || !HBS::isEqual(RC, 0, SC, BN, W))
2150 continue;
2151 if (!validateReg(RS, NewOpc, 1))
2152 continue;
2153
2154 Register NewR = MRI.createVirtualRegister(&Hexagon::IntRegsRegClass);
2155 auto At = MI->isPHI() ? B.getFirstNonPHI()
2156 : MachineBasicBlock::iterator(MI);
2157 auto MIB = BuildMI(B, At, DL, HII.get(NewOpc), NewR)
2158 .addReg(RS.Reg, 0, RS.Sub);
2159 if (NewOpc == Hexagon::A2_andir)
2160 MIB.addImm((1 << W) - 1);
2161 else if (NewOpc == Hexagon::S2_extractu)
2162 MIB.addImm(W).addImm(0);
2163 HBS::replaceSubWithSub(RD.Reg, RD.Sub, NewR, 0, MRI);
2164 BT.put(BitTracker::RegisterRef(NewR), RC);
2165 return true;
2166 }
2167 return false;
2168 }
2169
2170 bool BitSimplification::genBitSplit(MachineInstr *MI,
2171 BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC,
2172 const RegisterSet &AVs) {
2173 if (!GenBitSplit)
2174 return false;
2175 if (MaxBitSplit.getNumOccurrences()) {
2176 if (CountBitSplit >= MaxBitSplit)
2177 return false;
2178 }
2179
2180 unsigned Opc = MI->getOpcode();
2181 switch (Opc) {
2182 case Hexagon::A4_bitsplit:
2183 case Hexagon::A4_bitspliti:
2184 return false;
2185 }
2186
2187 unsigned W = RC.width();
2188 if (W != 32)
2189 return false;
2190
2191 auto ctlz = [] (const BitTracker::RegisterCell &C) -> unsigned {
2192 unsigned Z = C.width();
2193 while (Z > 0 && C[Z-1].is(0))
2194 --Z;
2195 return C.width() - Z;
2196 };
2197
2198 // Count the number of leading zeros in the target RC.
2199 unsigned Z = ctlz(RC);
2200 if (Z == 0 || Z == W)
2201 return false;
2202
2203 // A simplistic analysis: assume the source register (the one being split)
2204 // is fully unknown, and that all its bits are self-references.
2205 const BitTracker::BitValue &B0 = RC[0];
2206 if (B0.Type != BitTracker::BitValue::Ref)
2207 return false;
2208
2209 unsigned SrcR = B0.RefI.Reg;
2210 unsigned SrcSR = 0;
2211 unsigned Pos = B0.RefI.Pos;
2212
2213 // All the non-zero bits should be consecutive bits from the same register.
2214 for (unsigned i = 1; i < W-Z; ++i) {
2215 const BitTracker::BitValue &V = RC[i];
2216 if (V.Type != BitTracker::BitValue::Ref)
2217 return false;
2218 if (V.RefI.Reg != SrcR || V.RefI.Pos != Pos+i)
2219 return false;
2220 }
2221
2222 // Now, find the other bitfield among AVs.
2223 for (unsigned S = AVs.find_first(); S; S = AVs.find_next(S)) {
2224 // The number of leading zeros here should be the number of trailing
2225 // non-zeros in RC.
2226 unsigned SRC = MRI.getRegClass(S)->getID();
2227 if (SRC != Hexagon::IntRegsRegClassID &&
2228 SRC != Hexagon::DoubleRegsRegClassID)
2229 continue;
2230 if (!BT.has(S))
2231 continue;
2232 const BitTracker::RegisterCell &SC = BT.lookup(S);
2233 if (SC.width() != W || ctlz(SC) != W-Z)
2234 continue;
2235 // The Z lower bits should now match SrcR.
2236 const BitTracker::BitValue &S0 = SC[0];
2237 if (S0.Type != BitTracker::BitValue::Ref || S0.RefI.Reg != SrcR)
2238 continue;
2239 unsigned P = S0.RefI.Pos;
2240
2241 if (Pos <= P && (Pos + W-Z) != P)
2242 continue;
2243 if (P < Pos && (P + Z) != Pos)
2244 continue;
2245 // The starting bitfield position must be at a subregister boundary.
2246 if (std::min(P, Pos) != 0 && std::min(P, Pos) != 32)
2247 continue;
2248
2249 unsigned I;
2250 for (I = 1; I < Z; ++I) {
2251 const BitTracker::BitValue &V = SC[I];
2252 if (V.Type != BitTracker::BitValue::Ref)
2253 break;
2254 if (V.RefI.Reg != SrcR || V.RefI.Pos != P+I)
2255 break;
2256 }
2257 if (I != Z)
2258 continue;
2259
2260 // Generate bitsplit where S is defined.
2261 if (MaxBitSplit.getNumOccurrences())
2262 CountBitSplit++;
2263 MachineInstr *DefS = MRI.getVRegDef(S);
2264 assert(DefS != nullptr);
2265 DebugLoc DL = DefS->getDebugLoc();
2266 MachineBasicBlock &B = *DefS->getParent();
2267 auto At = DefS->isPHI() ? B.getFirstNonPHI()
2268 : MachineBasicBlock::iterator(DefS);
2269 if (MRI.getRegClass(SrcR)->getID() == Hexagon::DoubleRegsRegClassID)
2270 SrcSR = (std::min(Pos, P) == 32) ? Hexagon::isub_hi : Hexagon::isub_lo;
2271 if (!validateReg({SrcR,SrcSR}, Hexagon::A4_bitspliti, 1))
2272 continue;
2273 unsigned ImmOp = Pos <= P ? W-Z : Z;
2274
2275 // Find an existing bitsplit instruction if one already exists.
2276 unsigned NewR = 0;
2277 for (MachineInstr *In : NewMIs) {
2278 if (In->getOpcode() != Hexagon::A4_bitspliti)
2279 continue;
2280 MachineOperand &Op1 = In->getOperand(1);
2281 if (Op1.getReg() != SrcR || Op1.getSubReg() != SrcSR)
2282 continue;
2283 if (In->getOperand(2).getImm() != ImmOp)
2284 continue;
2285 // Check if the target register is available here.
2286 MachineOperand &Op0 = In->getOperand(0);
2287 MachineInstr *DefI = MRI.getVRegDef(Op0.getReg());
2288 assert(DefI != nullptr);
2289 if (!MDT.dominates(DefI, &*At))
2290 continue;
2291
2292 // Found one that can be reused.
2293 assert(Op0.getSubReg() == 0);
2294 NewR = Op0.getReg();
2295 break;
2296 }
2297 if (!NewR) {
2298 NewR = MRI.createVirtualRegister(&Hexagon::DoubleRegsRegClass);
2299 auto NewBS = BuildMI(B, At, DL, HII.get(Hexagon::A4_bitspliti), NewR)
2300 .addReg(SrcR, 0, SrcSR)
2301 .addImm(ImmOp);
2302 NewMIs.push_back(NewBS);
2303 }
2304 if (Pos <= P) {
2305 HBS::replaceRegWithSub(RD.Reg, NewR, Hexagon::isub_lo, MRI);
2306 HBS::replaceRegWithSub(S, NewR, Hexagon::isub_hi, MRI);
2307 } else {
2308 HBS::replaceRegWithSub(S, NewR, Hexagon::isub_lo, MRI);
2309 HBS::replaceRegWithSub(RD.Reg, NewR, Hexagon::isub_hi, MRI);
2310 }
2311 return true;
2312 }
2313
2314 return false;
2315 }
2316
2317 // Check for tstbit simplification opportunity, where the bit being checked
2318 // can be tracked back to another register. For example:
2319 // %2 = S2_lsr_i_r %1, 5
2320 // %3 = S2_tstbit_i %2, 0
2321 // =>
2322 // %3 = S2_tstbit_i %1, 5
2323 bool BitSimplification::simplifyTstbit(MachineInstr *MI,
2324 BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC) {
2325 unsigned Opc = MI->getOpcode();
2326 if (Opc != Hexagon::S2_tstbit_i)
2327 return false;
2328
2329 unsigned BN = MI->getOperand(2).getImm();
2330 BitTracker::RegisterRef RS = MI->getOperand(1);
2331 unsigned F, W;
2332 DebugLoc DL = MI->getDebugLoc();
2333 if (!BT.has(RS.Reg) || !HBS::getSubregMask(RS, F, W, MRI))
2334 return false;
2335 MachineBasicBlock &B = *MI->getParent();
2336 auto At = MI->isPHI() ? B.getFirstNonPHI()
2337 : MachineBasicBlock::iterator(MI);
2338
2339 const BitTracker::RegisterCell &SC = BT.lookup(RS.Reg);
2340 const BitTracker::BitValue &V = SC[F+BN];
2341 if (V.Type == BitTracker::BitValue::Ref && V.RefI.Reg != RS.Reg) {
2342 const TargetRegisterClass *TC = MRI.getRegClass(V.RefI.Reg);
2343 // Need to map V.RefI.Reg to a 32-bit register, i.e. if it is
2344 // a double register, need to use a subregister and adjust bit
2345 // number.
2346 unsigned P = std::numeric_limits<unsigned>::max();
2347 BitTracker::RegisterRef RR(V.RefI.Reg, 0);
2348 if (TC == &Hexagon::DoubleRegsRegClass) {
2349 P = V.RefI.Pos;
2350 RR.Sub = Hexagon::isub_lo;
2351 if (P >= 32) {
2352 P -= 32;
2353 RR.Sub = Hexagon::isub_hi;
2354 }
2355 } else if (TC == &Hexagon::IntRegsRegClass) {
2356 P = V.RefI.Pos;
2357 }
2358 if (P != std::numeric_limits<unsigned>::max()) {
2359 unsigned NewR = MRI.createVirtualRegister(&Hexagon::PredRegsRegClass);
2360 BuildMI(B, At, DL, HII.get(Hexagon::S2_tstbit_i), NewR)
2361 .addReg(RR.Reg, 0, RR.Sub)
2362 .addImm(P);
2363 HBS::replaceReg(RD.Reg, NewR, MRI);
2364 BT.put(NewR, RC);
2365 return true;
2366 }
2367 } else if (V.is(0) || V.is(1)) {
2368 Register NewR = MRI.createVirtualRegister(&Hexagon::PredRegsRegClass);
2369 unsigned NewOpc = V.is(0) ? Hexagon::PS_false : Hexagon::PS_true;
2370 BuildMI(B, At, DL, HII.get(NewOpc), NewR);
2371 HBS::replaceReg(RD.Reg, NewR, MRI);
2372 return true;
2373 }
2374
2375 return false;
2376 }
2377
2378 // Detect whether RD is a bitfield extract (sign- or zero-extended) of
2379 // some register from the AVs set. Create a new corresponding instruction
2380 // at the location of MI. The intent is to recognize situations where
2381 // a sequence of instructions performs an operation that is equivalent to
2382 // an extract operation, such as a shift left followed by a shift right.
2383 bool BitSimplification::simplifyExtractLow(MachineInstr *MI,
2384 BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC,
2385 const RegisterSet &AVs) {
2386 if (!GenExtract)
2387 return false;
2388 if (MaxExtract.getNumOccurrences()) {
2389 if (CountExtract >= MaxExtract)
2390 return false;
2391 CountExtract++;
2392 }
2393
2394 unsigned W = RC.width();
2395 unsigned RW = W;
2396 unsigned Len;
2397 bool Signed;
2398
2399 // The code is mostly class-independent, except for the part that generates
2400 // the extract instruction, and establishes the source register (in case it
2401 // needs to use a subregister).
2402 const TargetRegisterClass *FRC = HBS::getFinalVRegClass(RD, MRI);
2403 if (FRC != &Hexagon::IntRegsRegClass && FRC != &Hexagon::DoubleRegsRegClass)
2404 return false;
2405 assert(RD.Sub == 0);
2406
2407 // Observation:
2408 // If the cell has a form of 00..0xx..x with k zeros and n remaining
2409 // bits, this could be an extractu of the n bits, but it could also be
2410 // an extractu of a longer field which happens to have 0s in the top
2411 // bit positions.
2412 // The same logic applies to sign-extended fields.
2413 //
2414 // Do not check for the extended extracts, since it would expand the
2415 // search space quite a bit. The search may be expensive as it is.
2416
2417 const BitTracker::BitValue &TopV = RC[W-1];
2418
2419 // Eliminate candidates that have self-referential bits, since they
2420 // cannot be extracts from other registers. Also, skip registers that
2421 // have compile-time constant values.
2422 bool IsConst = true;
2423 for (unsigned I = 0; I != W; ++I) {
2424 const BitTracker::BitValue &V = RC[I];
2425 if (V.Type == BitTracker::BitValue::Ref && V.RefI.Reg == RD.Reg)
2426 return false;
2427 IsConst = IsConst && (V.is(0) || V.is(1));
2428 }
2429 if (IsConst)
2430 return false;
2431
2432 if (TopV.is(0) || TopV.is(1)) {
2433 bool S = TopV.is(1);
2434 for (--W; W > 0 && RC[W-1].is(S); --W)
2435 ;
2436 Len = W;
2437 Signed = S;
2438 // The sign bit must be a part of the field being extended.
2439 if (Signed)
2440 ++Len;
2441 } else {
2442 // This could still be a sign-extended extract.
2443 assert(TopV.Type == BitTracker::BitValue::Ref);
2444 if (TopV.RefI.Reg == RD.Reg || TopV.RefI.Pos == W-1)
2445 return false;
2446 for (--W; W > 0 && RC[W-1] == TopV; --W)
2447 ;
2448 // The top bits of RC are copies of TopV. One occurrence of TopV will
2449 // be a part of the field.
2450 Len = W + 1;
2451 Signed = true;
2452 }
2453
2454 // This would be just a copy. It should be handled elsewhere.
2455 if (Len == RW)
2456 return false;
2457
2458 LLVM_DEBUG({
2459 dbgs() << __func__ << " on reg: " << printReg(RD.Reg, &HRI, RD.Sub)
2460 << ", MI: " << *MI;
2461 dbgs() << "Cell: " << RC << '\n';
2462 dbgs() << "Expected bitfield size: " << Len << " bits, "
2463 << (Signed ? "sign" : "zero") << "-extended\n";
2464 });
2465
2466 bool Changed = false;
2467
2468 for (unsigned R = AVs.find_first(); R != 0; R = AVs.find_next(R)) {
2469 if (!BT.has(R))
2470 continue;
2471 const BitTracker::RegisterCell &SC = BT.lookup(R);
2472 unsigned SW = SC.width();
2473
2474 // The source can be longer than the destination, as long as its size is
2475 // a multiple of the size of the destination. Also, we would need to be
2476 // able to refer to the subregister in the source that would be of the
2477 // same size as the destination, but only check the sizes here.
2478 if (SW < RW || (SW % RW) != 0)
2479 continue;
2480
2481 // The field can start at any offset in SC as long as it contains Len
2482 // bits and does not cross subregister boundary (if the source register
2483 // is longer than the destination).
2484 unsigned Off = 0;
2485 while (Off <= SW-Len) {
2486 unsigned OE = (Off+Len)/RW;
2487 if (OE != Off/RW) {
2488 // The assumption here is that if the source (R) is longer than the
2489 // destination, then the destination is a sequence of words of
2490 // size RW, and each such word in R can be accessed via a subregister.
2491 //
2492 // If the beginning and the end of the field cross the subregister
2493 // boundary, advance to the next subregister.
2494 Off = OE*RW;
2495 continue;
2496 }
2497 if (HBS::isEqual(RC, 0, SC, Off, Len))
2498 break;
2499 ++Off;
2500 }
2501
2502 if (Off > SW-Len)
2503 continue;
2504
2505 // Found match.
2506 unsigned ExtOpc = 0;
2507 if (Off == 0) {
2508 if (Len == 8)
2509 ExtOpc = Signed ? Hexagon::A2_sxtb : Hexagon::A2_zxtb;
2510 else if (Len == 16)
2511 ExtOpc = Signed ? Hexagon::A2_sxth : Hexagon::A2_zxth;
2512 else if (Len < 10 && !Signed)
2513 ExtOpc = Hexagon::A2_andir;
2514 }
2515 if (ExtOpc == 0) {
2516 ExtOpc =
2517 Signed ? (RW == 32 ? Hexagon::S4_extract : Hexagon::S4_extractp)
2518 : (RW == 32 ? Hexagon::S2_extractu : Hexagon::S2_extractup);
2519 }
2520 unsigned SR = 0;
2521 // This only recognizes isub_lo and isub_hi.
2522 if (RW != SW && RW*2 != SW)
2523 continue;
2524 if (RW != SW)
2525 SR = (Off/RW == 0) ? Hexagon::isub_lo : Hexagon::isub_hi;
2526 Off = Off % RW;
2527
2528 if (!validateReg({R,SR}, ExtOpc, 1))
2529 continue;
2530
2531 // Don't generate the same instruction as the one being optimized.
2532 if (MI->getOpcode() == ExtOpc) {
2533 // All possible ExtOpc's have the source in operand(1).
2534 const MachineOperand &SrcOp = MI->getOperand(1);
2535 if (SrcOp.getReg() == R)
2536 continue;
2537 }
2538
2539 DebugLoc DL = MI->getDebugLoc();
2540 MachineBasicBlock &B = *MI->getParent();
2541 Register NewR = MRI.createVirtualRegister(FRC);
2542 auto At = MI->isPHI() ? B.getFirstNonPHI()
2543 : MachineBasicBlock::iterator(MI);
2544 auto MIB = BuildMI(B, At, DL, HII.get(ExtOpc), NewR)
2545 .addReg(R, 0, SR);
2546 switch (ExtOpc) {
2547 case Hexagon::A2_sxtb:
2548 case Hexagon::A2_zxtb:
2549 case Hexagon::A2_sxth:
2550 case Hexagon::A2_zxth:
2551 break;
2552 case Hexagon::A2_andir:
2553 MIB.addImm((1u << Len) - 1);
2554 break;
2555 case Hexagon::S4_extract:
2556 case Hexagon::S2_extractu:
2557 case Hexagon::S4_extractp:
2558 case Hexagon::S2_extractup:
2559 MIB.addImm(Len)
2560 .addImm(Off);
2561 break;
2562 default:
2563 llvm_unreachable("Unexpected opcode");
2564 }
2565
2566 HBS::replaceReg(RD.Reg, NewR, MRI);
2567 BT.put(BitTracker::RegisterRef(NewR), RC);
2568 Changed = true;
2569 break;
2570 }
2571
2572 return Changed;
2573 }
2574
2575 bool BitSimplification::simplifyRCmp0(MachineInstr *MI,
2576 BitTracker::RegisterRef RD) {
2577 unsigned Opc = MI->getOpcode();
2578 if (Opc != Hexagon::A4_rcmpeqi && Opc != Hexagon::A4_rcmpneqi)
2579 return false;
2580 MachineOperand &CmpOp = MI->getOperand(2);
2581 if (!CmpOp.isImm() || CmpOp.getImm() != 0)
2582 return false;
2583
2584 const TargetRegisterClass *FRC = HBS::getFinalVRegClass(RD, MRI);
2585 if (FRC != &Hexagon::IntRegsRegClass && FRC != &Hexagon::DoubleRegsRegClass)
2586 return false;
2587 assert(RD.Sub == 0);
2588
2589 MachineBasicBlock &B = *MI->getParent();
2590 const DebugLoc &DL = MI->getDebugLoc();
2591 auto At = MI->isPHI() ? B.getFirstNonPHI()
2592 : MachineBasicBlock::iterator(MI);
2593 bool KnownZ = true;
2594 bool KnownNZ = false;
2595
2596 BitTracker::RegisterRef SR = MI->getOperand(1);
2597 if (!BT.has(SR.Reg))
2598 return false;
2599 const BitTracker::RegisterCell &SC = BT.lookup(SR.Reg);
2600 unsigned F, W;
2601 if (!HBS::getSubregMask(SR, F, W, MRI))
2602 return false;
2603
2604 for (uint16_t I = F; I != F+W; ++I) {
2605 const BitTracker::BitValue &V = SC[I];
2606 if (!V.is(0))
2607 KnownZ = false;
2608 if (V.is(1))
2609 KnownNZ = true;
2610 }
2611
2612 auto ReplaceWithConst = [&](int C) {
2613 Register NewR = MRI.createVirtualRegister(FRC);
2614 BuildMI(B, At, DL, HII.get(Hexagon::A2_tfrsi), NewR)
2615 .addImm(C);
2616 HBS::replaceReg(RD.Reg, NewR, MRI);
2617 BitTracker::RegisterCell NewRC(W);
2618 for (uint16_t I = 0; I != W; ++I) {
2619 NewRC[I] = BitTracker::BitValue(C & 1);
2620 C = unsigned(C) >> 1;
2621 }
2622 BT.put(BitTracker::RegisterRef(NewR), NewRC);
2623 return true;
2624 };
2625
2626 auto IsNonZero = [] (const MachineOperand &Op) {
2627 if (Op.isGlobal() || Op.isBlockAddress())
2628 return true;
2629 if (Op.isImm())
2630 return Op.getImm() != 0;
2631 if (Op.isCImm())
2632 return !Op.getCImm()->isZero();
2633 if (Op.isFPImm())
2634 return !Op.getFPImm()->isZero();
2635 return false;
2636 };
2637
2638 auto IsZero = [] (const MachineOperand &Op) {
2639 if (Op.isGlobal() || Op.isBlockAddress())
2640 return false;
2641 if (Op.isImm())
2642 return Op.getImm() == 0;
2643 if (Op.isCImm())
2644 return Op.getCImm()->isZero();
2645 if (Op.isFPImm())
2646 return Op.getFPImm()->isZero();
2647 return false;
2648 };
2649
2650 // If the source register is known to be 0 or non-0, the comparison can
2651 // be folded to a load of a constant.
2652 if (KnownZ || KnownNZ) {
2653 assert(KnownZ != KnownNZ && "Register cannot be both 0 and non-0");
2654 return ReplaceWithConst(KnownZ == (Opc == Hexagon::A4_rcmpeqi));
2655 }
2656
2657 // Special case: if the compare comes from a C2_muxii, then we know the
2658 // two possible constants that can be the source value.
2659 MachineInstr *InpDef = MRI.getVRegDef(SR.Reg);
2660 if (!InpDef)
2661 return false;
2662 if (SR.Sub == 0 && InpDef->getOpcode() == Hexagon::C2_muxii) {
2663 MachineOperand &Src1 = InpDef->getOperand(2);
2664 MachineOperand &Src2 = InpDef->getOperand(3);
2665 // Check if both are non-zero.
2666 bool KnownNZ1 = IsNonZero(Src1), KnownNZ2 = IsNonZero(Src2);
2667 if (KnownNZ1 && KnownNZ2)
2668 return ReplaceWithConst(Opc == Hexagon::A4_rcmpneqi);
2669 // Check if both are zero.
2670 bool KnownZ1 = IsZero(Src1), KnownZ2 = IsZero(Src2);
2671 if (KnownZ1 && KnownZ2)
2672 return ReplaceWithConst(Opc == Hexagon::A4_rcmpeqi);
2673
2674 // If for both operands we know that they are either 0 or non-0,
2675 // replace the comparison with a C2_muxii, using the same predicate
2676 // register, but with operands substituted with 0/1 accordingly.
2677 if ((KnownZ1 || KnownNZ1) && (KnownZ2 || KnownNZ2)) {
2678 Register NewR = MRI.createVirtualRegister(FRC);
2679 BuildMI(B, At, DL, HII.get(Hexagon::C2_muxii), NewR)
2680 .addReg(InpDef->getOperand(1).getReg())
2681 .addImm(KnownZ1 == (Opc == Hexagon::A4_rcmpeqi))
2682 .addImm(KnownZ2 == (Opc == Hexagon::A4_rcmpeqi));
2683 HBS::replaceReg(RD.Reg, NewR, MRI);
2684 // Create a new cell with only the least significant bit unknown.
2685 BitTracker::RegisterCell NewRC(W);
2686 NewRC[0] = BitTracker::BitValue::self();
2687 NewRC.fill(1, W, BitTracker::BitValue::Zero);
2688 BT.put(BitTracker::RegisterRef(NewR), NewRC);
2689 return true;
2690 }
2691 }
2692
2693 return false;
2694 }
2695
2696 bool BitSimplification::processBlock(MachineBasicBlock &B,
2697 const RegisterSet &AVs) {
2698 if (!BT.reached(&B))
2699 return false;
2700 bool Changed = false;
2701 RegisterSet AVB = AVs;
2702 RegisterSet Defs;
2703
2704 for (auto I = B.begin(), E = B.end(); I != E; ++I, AVB.insert(Defs)) {
2705 MachineInstr *MI = &*I;
2706 Defs.clear();
2707 HBS::getInstrDefs(*MI, Defs);
2708
2709 unsigned Opc = MI->getOpcode();
2710 if (Opc == TargetOpcode::COPY || Opc == TargetOpcode::REG_SEQUENCE)
2711 continue;
2712
2713 if (MI->mayStore()) {
2714 bool T = genStoreUpperHalf(MI);
2715 T = T || genStoreImmediate(MI);
2716 Changed |= T;
2717 continue;
2718 }
2719
2720 if (Defs.count() != 1)
2721 continue;
2722 const MachineOperand &Op0 = MI->getOperand(0);
2723 if (!Op0.isReg() || !Op0.isDef())
2724 continue;
2725 BitTracker::RegisterRef RD = Op0;
2726 if (!BT.has(RD.Reg))
2727 continue;
2728 const TargetRegisterClass *FRC = HBS::getFinalVRegClass(RD, MRI);
2729 const BitTracker::RegisterCell &RC = BT.lookup(RD.Reg);
2730
2731 if (FRC->getID() == Hexagon::DoubleRegsRegClassID) {
2732 bool T = genPackhl(MI, RD, RC);
2733 T = T || simplifyExtractLow(MI, RD, RC, AVB);
2734 Changed |= T;
2735 continue;
2736 }
2737
2738 if (FRC->getID() == Hexagon::IntRegsRegClassID) {
2739 bool T = genBitSplit(MI, RD, RC, AVB);
2740 T = T || simplifyExtractLow(MI, RD, RC, AVB);
2741 T = T || genExtractHalf(MI, RD, RC);
2742 T = T || genCombineHalf(MI, RD, RC);
2743 T = T || genExtractLow(MI, RD, RC);
2744 T = T || simplifyRCmp0(MI, RD);
2745 Changed |= T;
2746 continue;
2747 }
2748
2749 if (FRC->getID() == Hexagon::PredRegsRegClassID) {
2750 bool T = simplifyTstbit(MI, RD, RC);
2751 Changed |= T;
2752 continue;
2753 }
2754 }
2755 return Changed;
2756 }
2757
2758 bool HexagonBitSimplify::runOnMachineFunction(MachineFunction &MF) {
2759 if (skipFunction(MF.getFunction()))
2760 return false;
2761
2762 auto &HST = MF.getSubtarget<HexagonSubtarget>();
2763 auto &HRI = *HST.getRegisterInfo();
2764 auto &HII = *HST.getInstrInfo();
2765
2766 MDT = &getAnalysis<MachineDominatorTree>();
2767 MachineRegisterInfo &MRI = MF.getRegInfo();
2768 bool Changed;
2769
2770 Changed = DeadCodeElimination(MF, *MDT).run();
2771
2772 const HexagonEvaluator HE(HRI, MRI, HII, MF);
2773 BitTracker BT(HE, MF);
2774 LLVM_DEBUG(BT.trace(true));
2775 BT.run();
2776
2777 MachineBasicBlock &Entry = MF.front();
2778
2779 RegisterSet AIG; // Available registers for IG.
2780 ConstGeneration ImmG(BT, HII, MRI);
2781 Changed |= visitBlock(Entry, ImmG, AIG);
2782
2783 RegisterSet ARE; // Available registers for RIE.
2784 RedundantInstrElimination RIE(BT, HII, HRI, MRI);
2785 bool Ried = visitBlock(Entry, RIE, ARE);
2786 if (Ried) {
2787 Changed = true;
2788 BT.run();
2789 }
2790
2791 RegisterSet ACG; // Available registers for CG.
2792 CopyGeneration CopyG(BT, HII, HRI, MRI);
2793 Changed |= visitBlock(Entry, CopyG, ACG);
2794
2795 RegisterSet ACP; // Available registers for CP.
2796 CopyPropagation CopyP(HRI, MRI);
2797 Changed |= visitBlock(Entry, CopyP, ACP);
2798
2799 Changed = DeadCodeElimination(MF, *MDT).run() || Changed;
2800
2801 BT.run();
2802 RegisterSet ABS; // Available registers for BS.
2803 BitSimplification BitS(BT, *MDT, HII, HRI, MRI, MF);
2804 Changed |= visitBlock(Entry, BitS, ABS);
2805
2806 Changed = DeadCodeElimination(MF, *MDT).run() || Changed;
2807
2808 if (Changed) {
2809 for (auto &B : MF)
2810 for (auto &I : B)
2811 I.clearKillInfo();
2812 DeadCodeElimination(MF, *MDT).run();
2813 }
2814 return Changed;
2815 }
2816
2817 // Recognize loops where the code at the end of the loop matches the code
2818 // before the entry of the loop, and the matching code is such that is can
2819 // be simplified. This pass relies on the bit simplification above and only
2820 // prepares code in a way that can be handled by the bit simplifcation.
2821 //
2822 // This is the motivating testcase (and explanation):
2823 //
2824 // {
2825 // loop0(.LBB0_2, r1) // %for.body.preheader
2826 // r5:4 = memd(r0++#8)
2827 // }
2828 // {
2829 // r3 = lsr(r4, #16)
2830 // r7:6 = combine(r5, r5)
2831 // }
2832 // {
2833 // r3 = insert(r5, #16, #16)
2834 // r7:6 = vlsrw(r7:6, #16)
2835 // }
2836 // .LBB0_2:
2837 // {
2838 // memh(r2+#4) = r5
2839 // memh(r2+#6) = r6 # R6 is really R5.H
2840 // }
2841 // {
2842 // r2 = add(r2, #8)
2843 // memh(r2+#0) = r4
2844 // memh(r2+#2) = r3 # R3 is really R4.H
2845 // }
2846 // {
2847 // r5:4 = memd(r0++#8)
2848 // }
2849 // { # "Shuffling" code that sets up R3 and R6
2850 // r3 = lsr(r4, #16) # so that their halves can be stored in the
2851 // r7:6 = combine(r5, r5) # next iteration. This could be folded into
2852 // } # the stores if the code was at the beginning
2853 // { # of the loop iteration. Since the same code
2854 // r3 = insert(r5, #16, #16) # precedes the loop, it can actually be moved
2855 // r7:6 = vlsrw(r7:6, #16) # there.
2856 // }:endloop0
2857 //
2858 //
2859 // The outcome:
2860 //
2861 // {
2862 // loop0(.LBB0_2, r1)
2863 // r5:4 = memd(r0++#8)
2864 // }
2865 // .LBB0_2:
2866 // {
2867 // memh(r2+#4) = r5
2868 // memh(r2+#6) = r5.h
2869 // }
2870 // {
2871 // r2 = add(r2, #8)
2872 // memh(r2+#0) = r4
2873 // memh(r2+#2) = r4.h
2874 // }
2875 // {
2876 // r5:4 = memd(r0++#8)
2877 // }:endloop0
2878
2879 namespace llvm {
2880
2881 FunctionPass *createHexagonLoopRescheduling();
2882 void initializeHexagonLoopReschedulingPass(PassRegistry&);
2883
2884 } // end namespace llvm
2885
2886 namespace {
2887
2888 class HexagonLoopRescheduling : public MachineFunctionPass {
2889 public:
2890 static char ID;
2891
2892 HexagonLoopRescheduling() : MachineFunctionPass(ID) {
2893 initializeHexagonLoopReschedulingPass(*PassRegistry::getPassRegistry());
2894 }
2895
2896 bool runOnMachineFunction(MachineFunction &MF) override;
2897
2898 private:
2899 const HexagonInstrInfo *HII = nullptr;
2900 const HexagonRegisterInfo *HRI = nullptr;
2901 MachineRegisterInfo *MRI = nullptr;
2902 BitTracker *BTP = nullptr;
2903
2904 struct LoopCand {
2905 LoopCand(MachineBasicBlock *lb, MachineBasicBlock *pb,
2906 MachineBasicBlock *eb) : LB(lb), PB(pb), EB(eb) {}
2907
2908 MachineBasicBlock *LB, *PB, *EB;
2909 };
2910 using InstrList = std::vector<MachineInstr *>;
2911 struct InstrGroup {
2912 BitTracker::RegisterRef Inp, Out;
2913 InstrList Ins;
2914 };
2915 struct PhiInfo {
2916 PhiInfo(MachineInstr &P, MachineBasicBlock &B);
2917
2918 unsigned DefR;
2919 BitTracker::RegisterRef LR, PR; // Loop Register, Preheader Register
2920 MachineBasicBlock *LB, *PB; // Loop Block, Preheader Block
2921 };
2922
2923 static unsigned getDefReg(const MachineInstr *MI);
2924 bool isConst(unsigned Reg) const;
2925 bool isBitShuffle(const MachineInstr *MI, unsigned DefR) const;
2926 bool isStoreInput(const MachineInstr *MI, unsigned DefR) const;
2927 bool isShuffleOf(unsigned OutR, unsigned InpR) const;
2928 bool isSameShuffle(unsigned OutR1, unsigned InpR1, unsigned OutR2,
2929 unsigned &InpR2) const;
2930 void moveGroup(InstrGroup &G, MachineBasicBlock &LB, MachineBasicBlock &PB,
2931 MachineBasicBlock::iterator At, unsigned OldPhiR, unsigned NewPredR);
2932 bool processLoop(LoopCand &C);
2933 };
2934
2935 } // end anonymous namespace
2936
2937 char HexagonLoopRescheduling::ID = 0;
2938
2939 INITIALIZE_PASS(HexagonLoopRescheduling, "hexagon-loop-resched",
2940 "Hexagon Loop Rescheduling", false, false)
2941
2942 HexagonLoopRescheduling::PhiInfo::PhiInfo(MachineInstr &P,
2943 MachineBasicBlock &B) {
2944 DefR = HexagonLoopRescheduling::getDefReg(&P);
2945 LB = &B;
2946 PB = nullptr;
2947 for (unsigned i = 1, n = P.getNumOperands(); i < n; i += 2) {
2948 const MachineOperand &OpB = P.getOperand(i+1);
2949 if (OpB.getMBB() == &B) {
2950 LR = P.getOperand(i);
2951 continue;
2952 }
2953 PB = OpB.getMBB();
2954 PR = P.getOperand(i);
2955 }
2956 }
2957
2958 unsigned HexagonLoopRescheduling::getDefReg(const MachineInstr *MI) {
2959 RegisterSet Defs;
2960 HBS::getInstrDefs(*MI, Defs);
2961 if (Defs.count() != 1)
2962 return 0;
2963 return Defs.find_first();
2964 }
2965
2966 bool HexagonLoopRescheduling::isConst(unsigned Reg) const {
2967 if (!BTP->has(Reg))
2968 return false;
2969 const BitTracker::RegisterCell &RC = BTP->lookup(Reg);
2970 for (unsigned i = 0, w = RC.width(); i < w; ++i) {
2971 const BitTracker::BitValue &V = RC[i];
2972 if (!V.is(0) && !V.is(1))
2973 return false;
2974 }
2975 return true;
2976 }
2977
2978 bool HexagonLoopRescheduling::isBitShuffle(const MachineInstr *MI,
2979 unsigned DefR) const {
2980 unsigned Opc = MI->getOpcode();
2981 switch (Opc) {
2982 case TargetOpcode::COPY:
2983 case Hexagon::S2_lsr_i_r:
2984 case Hexagon::S2_asr_i_r:
2985 case Hexagon::S2_asl_i_r:
2986 case Hexagon::S2_lsr_i_p:
2987 case Hexagon::S2_asr_i_p:
2988 case Hexagon::S2_asl_i_p:
2989 case Hexagon::S2_insert:
2990 case Hexagon::A2_or:
2991 case Hexagon::A2_orp:
2992 case Hexagon::A2_and:
2993 case Hexagon::A2_andp:
2994 case Hexagon::A2_combinew:
2995 case Hexagon::A4_combineri:
2996 case Hexagon::A4_combineir:
2997 case Hexagon::A2_combineii:
2998 case Hexagon::A4_combineii:
2999 case Hexagon::A2_combine_ll:
3000 case Hexagon::A2_combine_lh:
3001 case Hexagon::A2_combine_hl:
3002 case Hexagon::A2_combine_hh:
3003 return true;
3004 }
3005 return false;
3006 }
3007
3008 bool HexagonLoopRescheduling::isStoreInput(const MachineInstr *MI,
3009 unsigned InpR) const {
3010 for (unsigned i = 0, n = MI->getNumOperands(); i < n; ++i) {
3011 const MachineOperand &Op = MI->getOperand(i);
3012 if (!Op.isReg())
3013 continue;
3014 if (Op.getReg() == InpR)
3015 return i == n-1;
3016 }
3017 return false;
3018 }
3019
3020 bool HexagonLoopRescheduling::isShuffleOf(unsigned OutR, unsigned InpR) const {
3021 if (!BTP->has(OutR) || !BTP->has(InpR))
3022 return false;
3023 const BitTracker::RegisterCell &OutC = BTP->lookup(OutR);
3024 for (unsigned i = 0, w = OutC.width(); i < w; ++i) {
3025 const BitTracker::BitValue &V = OutC[i];
3026 if (V.Type != BitTracker::BitValue::Ref)
3027 continue;
3028 if (V.RefI.Reg != InpR)
3029 return false;
3030 }
3031 return true;
3032 }
3033
3034 bool HexagonLoopRescheduling::isSameShuffle(unsigned OutR1, unsigned InpR1,
3035 unsigned OutR2, unsigned &InpR2) const {
3036 if (!BTP->has(OutR1) || !BTP->has(InpR1) || !BTP->has(OutR2))
3037 return false;
3038 const BitTracker::RegisterCell &OutC1 = BTP->lookup(OutR1);
3039 const BitTracker::RegisterCell &OutC2 = BTP->lookup(OutR2);
3040 unsigned W = OutC1.width();
3041 unsigned MatchR = 0;
3042 if (W != OutC2.width())
3043 return false;
3044 for (unsigned i = 0; i < W; ++i) {
3045 const BitTracker::BitValue &V1 = OutC1[i], &V2 = OutC2[i];
3046 if (V1.Type != V2.Type || V1.Type == BitTracker::BitValue::One)
3047 return false;
3048 if (V1.Type != BitTracker::BitValue::Ref)
3049 continue;
3050 if (V1.RefI.Pos != V2.RefI.Pos)
3051 return false;
3052 if (V1.RefI.Reg != InpR1)
3053 return false;
3054 if (V2.RefI.Reg == 0 || V2.RefI.Reg == OutR2)
3055 return false;
3056 if (!MatchR)
3057 MatchR = V2.RefI.Reg;
3058 else if (V2.RefI.Reg != MatchR)
3059 return false;
3060 }
3061 InpR2 = MatchR;
3062 return true;
3063 }
3064
3065 void HexagonLoopRescheduling::moveGroup(InstrGroup &G, MachineBasicBlock &LB,
3066 MachineBasicBlock &PB, MachineBasicBlock::iterator At, unsigned OldPhiR,
3067 unsigned NewPredR) {
3068 DenseMap<unsigned,unsigned> RegMap;
3069
3070 const TargetRegisterClass *PhiRC = MRI->getRegClass(NewPredR);
3071 Register PhiR = MRI->createVirtualRegister(PhiRC);
3072 BuildMI(LB, At, At->getDebugLoc(), HII->get(TargetOpcode::PHI), PhiR)
3073 .addReg(NewPredR)
3074 .addMBB(&PB)
3075 .addReg(G.Inp.Reg)
3076 .addMBB(&LB);
3077 RegMap.insert(std::make_pair(G.Inp.Reg, PhiR));
3078
3079 for (unsigned i = G.Ins.size(); i > 0; --i) {
3080 const MachineInstr *SI = G.Ins[i-1];
3081 unsigned DR = getDefReg(SI);
3082 const TargetRegisterClass *RC = MRI->getRegClass(DR);
3083 Register NewDR = MRI->createVirtualRegister(RC);
3084 DebugLoc DL = SI->getDebugLoc();
3085
3086 auto MIB = BuildMI(LB, At, DL, HII->get(SI->getOpcode()), NewDR);
3087 for (unsigned j = 0, m = SI->getNumOperands(); j < m; ++j) {
3088 const MachineOperand &Op = SI->getOperand(j);
3089 if (!Op.isReg()) {
3090 MIB.add(Op);
3091 continue;
3092 }
3093 if (!Op.isUse())
3094 continue;
3095 unsigned UseR = RegMap[Op.getReg()];
3096 MIB.addReg(UseR, 0, Op.getSubReg());
3097 }
3098 RegMap.insert(std::make_pair(DR, NewDR));
3099 }
3100
3101 HBS::replaceReg(OldPhiR, RegMap[G.Out.Reg], *MRI);
3102 }
3103
3104 bool HexagonLoopRescheduling::processLoop(LoopCand &C) {
3105 LLVM_DEBUG(dbgs() << "Processing loop in " << printMBBReference(*C.LB)
3106 << "\n");
3107 std::vector<PhiInfo> Phis;
3108 for (auto &I : *C.LB) {
3109 if (!I.isPHI())
3110 break;
3111 unsigned PR = getDefReg(&I);
3112 if (isConst(PR))
3113 continue;
3114 bool BadUse = false, GoodUse = false;
3115 for (auto UI = MRI->use_begin(PR), UE = MRI->use_end(); UI != UE; ++UI) {
3116 MachineInstr *UseI = UI->getParent();
3117 if (UseI->getParent() != C.LB) {
3118 BadUse = true;
3119 break;
3120 }
3121 if (isBitShuffle(UseI, PR) || isStoreInput(UseI, PR))
3122 GoodUse = true;
3123 }
3124 if (BadUse || !GoodUse)
3125 continue;
3126
3127 Phis.push_back(PhiInfo(I, *C.LB));
3128 }
3129
3130 LLVM_DEBUG({
3131 dbgs() << "Phis: {";
3132 for (auto &I : Phis) {
3133 dbgs() << ' ' << printReg(I.DefR, HRI) << "=phi("
3134 << printReg(I.PR.Reg, HRI, I.PR.Sub) << ":b" << I.PB->getNumber()
3135 << ',' << printReg(I.LR.Reg, HRI, I.LR.Sub) << ":b"
3136 << I.LB->getNumber() << ')';
3137 }
3138 dbgs() << " }\n";
3139 });
3140
3141 if (Phis.empty())
3142 return false;
3143
3144 bool Changed = false;
3145 InstrList ShufIns;
3146
3147 // Go backwards in the block: for each bit shuffling instruction, check
3148 // if that instruction could potentially be moved to the front of the loop:
3149 // the output of the loop cannot be used in a non-shuffling instruction
3150 // in this loop.
3151 for (auto I = C.LB->rbegin(), E = C.LB->rend(); I != E; ++I) {
3152 if (I->isTerminator())
3153 continue;
3154 if (I->isPHI())
3155 break;
3156
3157 RegisterSet Defs;
3158 HBS::getInstrDefs(*I, Defs);
3159 if (Defs.count() != 1)
3160 continue;
3161 unsigned DefR = Defs.find_first();
3162 if (!Register::isVirtualRegister(DefR))
3163 continue;
3164 if (!isBitShuffle(&*I, DefR))
3165 continue;
3166
3167 bool BadUse = false;
3168 for (auto UI = MRI->use_begin(DefR), UE = MRI->use_end(); UI != UE; ++UI) {
3169 MachineInstr *UseI = UI->getParent();
3170 if (UseI->getParent() == C.LB) {
3171 if (UseI->isPHI()) {
3172 // If the use is in a phi node in this loop, then it should be
3173 // the value corresponding to the back edge.
3174 unsigned Idx = UI.getOperandNo();
3175 if (UseI->getOperand(Idx+1).getMBB() != C.LB)
3176 BadUse = true;
3177 } else {
3178 auto F = find(ShufIns, UseI);
3179 if (F == ShufIns.end())
3180 BadUse = true;
3181 }
3182 } else {
3183 // There is a use outside of the loop, but there is no epilog block
3184 // suitable for a copy-out.
3185 if (C.EB == nullptr)
3186 BadUse = true;
3187 }
3188 if (BadUse)
3189 break;
3190 }
3191
3192 if (BadUse)
3193 continue;
3194 ShufIns.push_back(&*I);
3195 }
3196
3197 // Partition the list of shuffling instructions into instruction groups,
3198 // where each group has to be moved as a whole (i.e. a group is a chain of
3199 // dependent instructions). A group produces a single live output register,
3200 // which is meant to be the input of the loop phi node (although this is
3201 // not checked here yet). It also uses a single register as its input,
3202 // which is some value produced in the loop body. After moving the group
3203 // to the beginning of the loop, that input register would need to be
3204 // the loop-carried register (through a phi node) instead of the (currently
3205 // loop-carried) output register.
3206 using InstrGroupList = std::vector<InstrGroup>;
3207 InstrGroupList Groups;
3208
3209 for (unsigned i = 0, n = ShufIns.size(); i < n; ++i) {
3210 MachineInstr *SI = ShufIns[i];
3211 if (SI == nullptr)
3212 continue;
3213
3214 InstrGroup G;
3215 G.Ins.push_back(SI);
3216 G.Out.Reg = getDefReg(SI);
3217 RegisterSet Inputs;
3218 HBS::getInstrUses(*SI, Inputs);
3219
3220 for (unsigned j = i+1; j < n; ++j) {
3221 MachineInstr *MI = ShufIns[j];
3222 if (MI == nullptr)
3223 continue;
3224 RegisterSet Defs;
3225 HBS::getInstrDefs(*MI, Defs);
3226 // If this instruction does not define any pending inputs, skip it.
3227 if (!Defs.intersects(Inputs))
3228 continue;
3229 // Otherwise, add it to the current group and remove the inputs that
3230 // are defined by MI.
3231 G.Ins.push_back(MI);
3232 Inputs.remove(Defs);
3233 // Then add all registers used by MI.
3234 HBS::getInstrUses(*MI, Inputs);
3235 ShufIns[j] = nullptr;
3236 }
3237
3238 // Only add a group if it requires at most one register.
3239 if (Inputs.count() > 1)
3240 continue;
3241 auto LoopInpEq = [G] (const PhiInfo &P) -> bool {
3242 return G.Out.Reg == P.LR.Reg;
3243 };
3244 if (llvm::find_if(Phis, LoopInpEq) == Phis.end())
3245 continue;
3246
3247 G.Inp.Reg = Inputs.find_first();
3248 Groups.push_back(G);
3249 }
3250
3251 LLVM_DEBUG({
3252 for (unsigned i = 0, n = Groups.size(); i < n; ++i) {
3253 InstrGroup &G = Groups[i];
3254 dbgs() << "Group[" << i << "] inp: "
3255 << printReg(G.Inp.Reg, HRI, G.Inp.Sub)
3256 << " out: " << printReg(G.Out.Reg, HRI, G.Out.Sub) << "\n";
3257 for (unsigned j = 0, m = G.Ins.size(); j < m; ++j)
3258 dbgs() << " " << *G.Ins[j];
3259 }
3260 });
3261
3262 for (unsigned i = 0, n = Groups.size(); i < n; ++i) {
3263 InstrGroup &G = Groups[i];
3264 if (!isShuffleOf(G.Out.Reg, G.Inp.Reg))
3265 continue;
3266 auto LoopInpEq = [G] (const PhiInfo &P) -> bool {
3267 return G.Out.Reg == P.LR.Reg;
3268 };
3269 auto F = llvm::find_if(Phis, LoopInpEq);
3270 if (F == Phis.end())
3271 continue;
3272 unsigned PrehR = 0;
3273 if (!isSameShuffle(G.Out.Reg, G.Inp.Reg, F->PR.Reg, PrehR)) {
3274 const MachineInstr *DefPrehR = MRI->getVRegDef(F->PR.Reg);
3275 unsigned Opc = DefPrehR->getOpcode();
3276 if (Opc != Hexagon::A2_tfrsi && Opc != Hexagon::A2_tfrpi)
3277 continue;
3278 if (!DefPrehR->getOperand(1).isImm())
3279 continue;
3280 if (DefPrehR->getOperand(1).getImm() != 0)
3281 continue;
3282 const TargetRegisterClass *RC = MRI->getRegClass(G.Inp.Reg);
3283 if (RC != MRI->getRegClass(F->PR.Reg)) {
3284 PrehR = MRI->createVirtualRegister(RC);
3285 unsigned TfrI = (RC == &Hexagon::IntRegsRegClass) ? Hexagon::A2_tfrsi
3286 : Hexagon::A2_tfrpi;
3287 auto T = C.PB->getFirstTerminator();
3288 DebugLoc DL = (T != C.PB->end()) ? T->getDebugLoc() : DebugLoc();
3289 BuildMI(*C.PB, T, DL, HII->get(TfrI), PrehR)
3290 .addImm(0);
3291 } else {
3292 PrehR = F->PR.Reg;
3293 }
3294 }
3295 // isSameShuffle could match with PrehR being of a wider class than
3296 // G.Inp.Reg, for example if G shuffles the low 32 bits of its input,
3297 // it would match for the input being a 32-bit register, and PrehR
3298 // being a 64-bit register (where the low 32 bits match). This could
3299 // be handled, but for now skip these cases.
3300 if (MRI->getRegClass(PrehR) != MRI->getRegClass(G.Inp.Reg))
3301 continue;
3302 moveGroup(G, *F->LB, *F->PB, F->LB->getFirstNonPHI(), F->DefR, PrehR);
3303 Changed = true;
3304 }
3305
3306 return Changed;
3307 }
3308
3309 bool HexagonLoopRescheduling::runOnMachineFunction(MachineFunction &MF) {
3310 if (skipFunction(MF.getFunction()))
3311 return false;
3312
3313 auto &HST = MF.getSubtarget<HexagonSubtarget>();
3314 HII = HST.getInstrInfo();
3315 HRI = HST.getRegisterInfo();
3316 MRI = &MF.getRegInfo();
3317 const HexagonEvaluator HE(*HRI, *MRI, *HII, MF);
3318 BitTracker BT(HE, MF);
3319 LLVM_DEBUG(BT.trace(true));
3320 BT.run();
3321 BTP = &BT;
3322
3323 std::vector<LoopCand> Cand;
3324
3325 for (auto &B : MF) {
3326 if (B.pred_size() != 2 || B.succ_size() != 2)
3327 continue;
3328 MachineBasicBlock *PB = nullptr;
3329 bool IsLoop = false;
3330 for (auto PI = B.pred_begin(), PE = B.pred_end(); PI != PE; ++PI) {
3331 if (*PI != &B)
3332 PB = *PI;
3333 else
3334 IsLoop = true;
3335 }
3336 if (!IsLoop)
3337 continue;
3338
3339 MachineBasicBlock *EB = nullptr;
3340 for (auto SI = B.succ_begin(), SE = B.succ_end(); SI != SE; ++SI) {
3341 if (*SI == &B)
3342 continue;
3343 // Set EP to the epilog block, if it has only 1 predecessor (i.e. the
3344 // edge from B to EP is non-critical.
3345 if ((*SI)->pred_size() == 1)
3346 EB = *SI;
3347 break;
3348 }
3349
3350 Cand.push_back(LoopCand(&B, PB, EB));
3351 }
3352
3353 bool Changed = false;
3354 for (auto &C : Cand)
3355 Changed |= processLoop(C);
3356
3357 return Changed;
3358 }
3359
3360 //===----------------------------------------------------------------------===//
3361 // Public Constructor Functions
3362 //===----------------------------------------------------------------------===//
3363
3364 FunctionPass *llvm::createHexagonLoopRescheduling() {
3365 return new HexagonLoopRescheduling();
3366 }
3367
3368 FunctionPass *llvm::createHexagonBitSimplify() {
3369 return new HexagonBitSimplify();
3370 }
The low level virtual machine