1 //===--- BitTracker.cpp ---------------------------------------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9
10 // SSA-based bit propagation.
11 //
12 // The purpose of this code is, for a given virtual register, to provide
13 // information about the value of each bit in the register. The values
14 // of bits are represented by the class BitValue, and take one of four
15 // cases: 0, 1, "ref" and "bottom". The 0 and 1 are rather clear, the
16 // "ref" value means that the bit is a copy of another bit (which itself
17 // cannot be a copy of yet another bit---such chains are not allowed).
18 // A "ref" value is associated with a BitRef structure, which indicates
19 // which virtual register, and which bit in that register is the origin
20 // of the value. For example, given an instruction
21 // vreg2 = ASL vreg1, 1
22 // assuming that nothing is known about bits of vreg1, bit 1 of vreg2
23 // will be a "ref" to (vreg1, 0). If there is a subsequent instruction
24 // vreg3 = ASL vreg2, 2
25 // then bit 3 of vreg3 will be a "ref" to (vreg1, 0) as well.
26 // The "bottom" case means that the bit's value cannot be determined,
27 // and that this virtual register actually defines it. The "bottom" case
28 // is discussed in detail in BitTracker.h. In fact, "bottom" is a "ref
29 // to self", so for the vreg1 above, the bit 0 of it will be a "ref" to
30 // (vreg1, 0), bit 1 will be a "ref" to (vreg1, 1), etc.
31 //
32 // The tracker implements the Wegman-Zadeck algorithm, originally developed
33 // for SSA-based constant propagation. Each register is represented as
34 // a sequence of bits, with the convention that bit 0 is the least signi-
35 // ficant bit. Each bit is propagated individually. The class RegisterCell
36 // implements the register's representation, and is also the subject of
37 // the lattice operations in the tracker.
38 //
39 // The intended usage of the bit tracker is to create a target-specific
40 // machine instruction evaluator, pass the evaluator to the BitTracker
41 // object, and run the tracker. The tracker will then collect the bit
42 // value information for a given machine function. After that, it can be
43 // queried for the cells for each virtual register.
44 // Sample code:
45 // const TargetSpecificEvaluator TSE(TRI, MRI);
46 // BitTracker BT(TSE, MF);
47 // BT.run();
48 // ...
49 // unsigned Reg = interestingRegister();
50 // RegisterCell RC = BT.get(Reg);
51 // if (RC[3].is(1))
52 // Reg0bit3 = 1;
53 //
54 // The code below is intended to be fully target-independent.
55
56 #include "llvm/CodeGen/MachineBasicBlock.h"
57 #include "llvm/CodeGen/MachineFunction.h"
58 #include "llvm/CodeGen/MachineInstr.h"
59 #include "llvm/CodeGen/MachineRegisterInfo.h"
60 #include "llvm/IR/Constants.h"
61 #include "llvm/Support/Debug.h"
62 #include "llvm/Support/raw_ostream.h"
63 #include "llvm/Target/TargetRegisterInfo.h"
64
65 #include "BitTracker.h"
66
67 using namespace llvm;
68
69 typedef BitTracker BT;
70
71 namespace {
72 // Local trickery to pretty print a register (without the whole "%vreg"
73 // business).
74 struct printv {
printv__anonec0588a90111::printv75 printv(unsigned r) : R(r) {}
76 unsigned R;
77 };
operator <<(raw_ostream & OS,const printv & PV)78 raw_ostream &operator<< (raw_ostream &OS, const printv &PV) {
79 if (PV.R)
80 OS << 'v' << TargetRegisterInfo::virtReg2Index(PV.R);
81 else
82 OS << 's';
83 return OS;
84 }
85 }
86
87 namespace llvm {
operator <<(raw_ostream & OS,const BT::BitValue & BV)88 raw_ostream &operator<<(raw_ostream &OS, const BT::BitValue &BV) {
89 switch (BV.Type) {
90 case BT::BitValue::Top:
91 OS << 'T';
92 break;
93 case BT::BitValue::Zero:
94 OS << '0';
95 break;
96 case BT::BitValue::One:
97 OS << '1';
98 break;
99 case BT::BitValue::Ref:
100 OS << printv(BV.RefI.Reg) << '[' << BV.RefI.Pos << ']';
101 break;
102 }
103 return OS;
104 }
105
operator <<(raw_ostream & OS,const BT::RegisterCell & RC)106 raw_ostream &operator<<(raw_ostream &OS, const BT::RegisterCell &RC) {
107 unsigned n = RC.Bits.size();
108 OS << "{ w:" << n;
109 // Instead of printing each bit value individually, try to group them
110 // into logical segments, such as sequences of 0 or 1 bits or references
111 // to consecutive bits (e.g. "bits 3-5 are same as bits 7-9 of reg xyz").
112 // "Start" will be the index of the beginning of the most recent segment.
113 unsigned Start = 0;
114 bool SeqRef = false; // A sequence of refs to consecutive bits.
115 bool ConstRef = false; // A sequence of refs to the same bit.
116
117 for (unsigned i = 1, n = RC.Bits.size(); i < n; ++i) {
118 const BT::BitValue &V = RC[i];
119 const BT::BitValue &SV = RC[Start];
120 bool IsRef = (V.Type == BT::BitValue::Ref);
121 // If the current value is the same as Start, skip to the next one.
122 if (!IsRef && V == SV)
123 continue;
124 if (IsRef && SV.Type == BT::BitValue::Ref && V.RefI.Reg == SV.RefI.Reg) {
125 if (Start+1 == i) {
126 SeqRef = (V.RefI.Pos == SV.RefI.Pos+1);
127 ConstRef = (V.RefI.Pos == SV.RefI.Pos);
128 }
129 if (SeqRef && V.RefI.Pos == SV.RefI.Pos+(i-Start))
130 continue;
131 if (ConstRef && V.RefI.Pos == SV.RefI.Pos)
132 continue;
133 }
134
135 // The current value is different. Print the previous one and reset
136 // the Start.
137 OS << " [" << Start;
138 unsigned Count = i - Start;
139 if (Count == 1) {
140 OS << "]:" << SV;
141 } else {
142 OS << '-' << i-1 << "]:";
143 if (SV.Type == BT::BitValue::Ref && SeqRef)
144 OS << printv(SV.RefI.Reg) << '[' << SV.RefI.Pos << '-'
145 << SV.RefI.Pos+(Count-1) << ']';
146 else
147 OS << SV;
148 }
149 Start = i;
150 SeqRef = ConstRef = false;
151 }
152
153 OS << " [" << Start;
154 unsigned Count = n - Start;
155 if (n-Start == 1) {
156 OS << "]:" << RC[Start];
157 } else {
158 OS << '-' << n-1 << "]:";
159 const BT::BitValue &SV = RC[Start];
160 if (SV.Type == BT::BitValue::Ref && SeqRef)
161 OS << printv(SV.RefI.Reg) << '[' << SV.RefI.Pos << '-'
162 << SV.RefI.Pos+(Count-1) << ']';
163 else
164 OS << SV;
165 }
166 OS << " }";
167
168 return OS;
169 }
170 }
171
BitTracker(const MachineEvaluator & E,MachineFunction & F)172 BitTracker::BitTracker(const MachineEvaluator &E, MachineFunction &F)
173 : Trace(false), ME(E), MF(F), MRI(F.getRegInfo()), Map(*new CellMapType) {}
174
~BitTracker()175 BitTracker::~BitTracker() {
176 delete ⤅
177 }
178
179
180 // If we were allowed to update a cell for a part of a register, the meet
181 // operation would need to be parametrized by the register number and the
182 // exact part of the register, so that the computer BitRefs correspond to
183 // the actual bits of the "self" register.
184 // While this cannot happen in the current implementation, I'm not sure
185 // if this should be ruled out in the future.
meet(const RegisterCell & RC,unsigned SelfR)186 bool BT::RegisterCell::meet(const RegisterCell &RC, unsigned SelfR) {
187 // An example when "meet" can be invoked with SelfR == 0 is a phi node
188 // with a physical register as an operand.
189 assert(SelfR == 0 || TargetRegisterInfo::isVirtualRegister(SelfR));
190 bool Changed = false;
191 for (uint16_t i = 0, n = Bits.size(); i < n; ++i) {
192 const BitValue &RCV = RC[i];
193 Changed |= Bits[i].meet(RCV, BitRef(SelfR, i));
194 }
195 return Changed;
196 }
197
198
199 // Insert the entire cell RC into the current cell at position given by M.
insert(const BT::RegisterCell & RC,const BitMask & M)200 BT::RegisterCell &BT::RegisterCell::insert(const BT::RegisterCell &RC,
201 const BitMask &M) {
202 uint16_t B = M.first(), E = M.last(), W = width();
203 // Sanity: M must be a valid mask for *this.
204 assert(B < W && E < W);
205 // Sanity: the masked part of *this must have the same number of bits
206 // as the source.
207 assert(B > E || E-B+1 == RC.width()); // B <= E => E-B+1 = |RC|.
208 assert(B <= E || E+(W-B)+1 == RC.width()); // E < B => E+(W-B)+1 = |RC|.
209 if (B <= E) {
210 for (uint16_t i = 0; i <= E-B; ++i)
211 Bits[i+B] = RC[i];
212 } else {
213 for (uint16_t i = 0; i < W-B; ++i)
214 Bits[i+B] = RC[i];
215 for (uint16_t i = 0; i <= E; ++i)
216 Bits[i] = RC[i+(W-B)];
217 }
218 return *this;
219 }
220
221
extract(const BitMask & M) const222 BT::RegisterCell BT::RegisterCell::extract(const BitMask &M) const {
223 uint16_t B = M.first(), E = M.last(), W = width();
224 assert(B < W && E < W);
225 if (B <= E) {
226 RegisterCell RC(E-B+1);
227 for (uint16_t i = B; i <= E; ++i)
228 RC.Bits[i-B] = Bits[i];
229 return RC;
230 }
231
232 RegisterCell RC(E+(W-B)+1);
233 for (uint16_t i = 0; i < W-B; ++i)
234 RC.Bits[i] = Bits[i+B];
235 for (uint16_t i = 0; i <= E; ++i)
236 RC.Bits[i+(W-B)] = Bits[i];
237 return RC;
238 }
239
240
rol(uint16_t Sh)241 BT::RegisterCell &BT::RegisterCell::rol(uint16_t Sh) {
242 // Rotate left (i.e. towards increasing bit indices).
243 // Swap the two parts: [0..W-Sh-1] [W-Sh..W-1]
244 uint16_t W = width();
245 Sh = Sh % W;
246 if (Sh == 0)
247 return *this;
248
249 RegisterCell Tmp(W-Sh);
250 // Tmp = [0..W-Sh-1].
251 for (uint16_t i = 0; i < W-Sh; ++i)
252 Tmp[i] = Bits[i];
253 // Shift [W-Sh..W-1] to [0..Sh-1].
254 for (uint16_t i = 0; i < Sh; ++i)
255 Bits[i] = Bits[W-Sh+i];
256 // Copy Tmp to [Sh..W-1].
257 for (uint16_t i = 0; i < W-Sh; ++i)
258 Bits[i+Sh] = Tmp.Bits[i];
259 return *this;
260 }
261
262
fill(uint16_t B,uint16_t E,const BitValue & V)263 BT::RegisterCell &BT::RegisterCell::fill(uint16_t B, uint16_t E,
264 const BitValue &V) {
265 assert(B <= E);
266 while (B < E)
267 Bits[B++] = V;
268 return *this;
269 }
270
271
cat(const RegisterCell & RC)272 BT::RegisterCell &BT::RegisterCell::cat(const RegisterCell &RC) {
273 // Append the cell given as the argument to the "this" cell.
274 // Bit 0 of RC becomes bit W of the result, where W is this->width().
275 uint16_t W = width(), WRC = RC.width();
276 Bits.resize(W+WRC);
277 for (uint16_t i = 0; i < WRC; ++i)
278 Bits[i+W] = RC.Bits[i];
279 return *this;
280 }
281
282
ct(bool B) const283 uint16_t BT::RegisterCell::ct(bool B) const {
284 uint16_t W = width();
285 uint16_t C = 0;
286 BitValue V = B;
287 while (C < W && Bits[C] == V)
288 C++;
289 return C;
290 }
291
292
cl(bool B) const293 uint16_t BT::RegisterCell::cl(bool B) const {
294 uint16_t W = width();
295 uint16_t C = 0;
296 BitValue V = B;
297 while (C < W && Bits[W-(C+1)] == V)
298 C++;
299 return C;
300 }
301
302
operator ==(const RegisterCell & RC) const303 bool BT::RegisterCell::operator== (const RegisterCell &RC) const {
304 uint16_t W = Bits.size();
305 if (RC.Bits.size() != W)
306 return false;
307 for (uint16_t i = 0; i < W; ++i)
308 if (Bits[i] != RC[i])
309 return false;
310 return true;
311 }
312
313
getRegBitWidth(const RegisterRef & RR) const314 uint16_t BT::MachineEvaluator::getRegBitWidth(const RegisterRef &RR) const {
315 // The general problem is with finding a register class that corresponds
316 // to a given reference reg:sub. There can be several such classes, and
317 // since we only care about the register size, it does not matter which
318 // such class we would find.
319 // The easiest way to accomplish what we want is to
320 // 1. find a physical register PhysR from the same class as RR.Reg,
321 // 2. find a physical register PhysS that corresponds to PhysR:RR.Sub,
322 // 3. find a register class that contains PhysS.
323 unsigned PhysR;
324 if (TargetRegisterInfo::isVirtualRegister(RR.Reg)) {
325 const TargetRegisterClass *VC = MRI.getRegClass(RR.Reg);
326 assert(VC->begin() != VC->end() && "Empty register class");
327 PhysR = *VC->begin();
328 } else {
329 assert(TargetRegisterInfo::isPhysicalRegister(RR.Reg));
330 PhysR = RR.Reg;
331 }
332
333 unsigned PhysS = (RR.Sub == 0) ? PhysR : TRI.getSubReg(PhysR, RR.Sub);
334 const TargetRegisterClass *RC = TRI.getMinimalPhysRegClass(PhysS);
335 uint16_t BW = RC->getSize()*8;
336 return BW;
337 }
338
339
getCell(const RegisterRef & RR,const CellMapType & M) const340 BT::RegisterCell BT::MachineEvaluator::getCell(const RegisterRef &RR,
341 const CellMapType &M) const {
342 uint16_t BW = getRegBitWidth(RR);
343
344 // Physical registers are assumed to be present in the map with an unknown
345 // value. Don't actually insert anything in the map, just return the cell.
346 if (TargetRegisterInfo::isPhysicalRegister(RR.Reg))
347 return RegisterCell::self(0, BW);
348
349 assert(TargetRegisterInfo::isVirtualRegister(RR.Reg));
350 // For virtual registers that belong to a class that is not tracked,
351 // generate an "unknown" value as well.
352 const TargetRegisterClass *C = MRI.getRegClass(RR.Reg);
353 if (!track(C))
354 return RegisterCell::self(0, BW);
355
356 CellMapType::const_iterator F = M.find(RR.Reg);
357 if (F != M.end()) {
358 if (!RR.Sub)
359 return F->second;
360 BitMask M = mask(RR.Reg, RR.Sub);
361 return F->second.extract(M);
362 }
363 // If not found, create a "top" entry, but do not insert it in the map.
364 return RegisterCell::top(BW);
365 }
366
367
putCell(const RegisterRef & RR,RegisterCell RC,CellMapType & M) const368 void BT::MachineEvaluator::putCell(const RegisterRef &RR, RegisterCell RC,
369 CellMapType &M) const {
370 // While updating the cell map can be done in a meaningful way for
371 // a part of a register, it makes little sense to implement it as the
372 // SSA representation would never contain such "partial definitions".
373 if (!TargetRegisterInfo::isVirtualRegister(RR.Reg))
374 return;
375 assert(RR.Sub == 0 && "Unexpected sub-register in definition");
376 // Eliminate all ref-to-reg-0 bit values: replace them with "self".
377 for (unsigned i = 0, n = RC.width(); i < n; ++i) {
378 const BitValue &V = RC[i];
379 if (V.Type == BitValue::Ref && V.RefI.Reg == 0)
380 RC[i].RefI = BitRef(RR.Reg, i);
381 }
382 M[RR.Reg] = RC;
383 }
384
385
386 // Check if the cell represents a compile-time integer value.
isInt(const RegisterCell & A) const387 bool BT::MachineEvaluator::isInt(const RegisterCell &A) const {
388 uint16_t W = A.width();
389 for (uint16_t i = 0; i < W; ++i)
390 if (!A[i].is(0) && !A[i].is(1))
391 return false;
392 return true;
393 }
394
395
396 // Convert a cell to the integer value. The result must fit in uint64_t.
toInt(const RegisterCell & A) const397 uint64_t BT::MachineEvaluator::toInt(const RegisterCell &A) const {
398 assert(isInt(A));
399 uint64_t Val = 0;
400 uint16_t W = A.width();
401 for (uint16_t i = 0; i < W; ++i) {
402 Val <<= 1;
403 Val |= A[i].is(1);
404 }
405 return Val;
406 }
407
408
409 // Evaluator helper functions. These implement some common operation on
410 // register cells that can be used to implement target-specific instructions
411 // in a target-specific evaluator.
412
eIMM(int64_t V,uint16_t W) const413 BT::RegisterCell BT::MachineEvaluator::eIMM(int64_t V, uint16_t W) const {
414 RegisterCell Res(W);
415 // For bits beyond the 63rd, this will generate the sign bit of V.
416 for (uint16_t i = 0; i < W; ++i) {
417 Res[i] = BitValue(V & 1);
418 V >>= 1;
419 }
420 return Res;
421 }
422
423
eIMM(const ConstantInt * CI) const424 BT::RegisterCell BT::MachineEvaluator::eIMM(const ConstantInt *CI) const {
425 const APInt &A = CI->getValue();
426 uint16_t BW = A.getBitWidth();
427 assert((unsigned)BW == A.getBitWidth() && "BitWidth overflow");
428 RegisterCell Res(BW);
429 for (uint16_t i = 0; i < BW; ++i)
430 Res[i] = A[i];
431 return Res;
432 }
433
434
eADD(const RegisterCell & A1,const RegisterCell & A2) const435 BT::RegisterCell BT::MachineEvaluator::eADD(const RegisterCell &A1,
436 const RegisterCell &A2) const {
437 uint16_t W = A1.width();
438 assert(W == A2.width());
439 RegisterCell Res(W);
440 bool Carry = false;
441 uint16_t I;
442 for (I = 0; I < W; ++I) {
443 const BitValue &V1 = A1[I];
444 const BitValue &V2 = A2[I];
445 if (!V1.num() || !V2.num())
446 break;
447 unsigned S = bool(V1) + bool(V2) + Carry;
448 Res[I] = BitValue(S & 1);
449 Carry = (S > 1);
450 }
451 for (; I < W; ++I) {
452 const BitValue &V1 = A1[I];
453 const BitValue &V2 = A2[I];
454 // If the next bit is same as Carry, the result will be 0 plus the
455 // other bit. The Carry bit will remain unchanged.
456 if (V1.is(Carry))
457 Res[I] = BitValue::ref(V2);
458 else if (V2.is(Carry))
459 Res[I] = BitValue::ref(V1);
460 else
461 break;
462 }
463 for (; I < W; ++I)
464 Res[I] = BitValue::self();
465 return Res;
466 }
467
468
eSUB(const RegisterCell & A1,const RegisterCell & A2) const469 BT::RegisterCell BT::MachineEvaluator::eSUB(const RegisterCell &A1,
470 const RegisterCell &A2) const {
471 uint16_t W = A1.width();
472 assert(W == A2.width());
473 RegisterCell Res(W);
474 bool Borrow = false;
475 uint16_t I;
476 for (I = 0; I < W; ++I) {
477 const BitValue &V1 = A1[I];
478 const BitValue &V2 = A2[I];
479 if (!V1.num() || !V2.num())
480 break;
481 unsigned S = bool(V1) - bool(V2) - Borrow;
482 Res[I] = BitValue(S & 1);
483 Borrow = (S > 1);
484 }
485 for (; I < W; ++I) {
486 const BitValue &V1 = A1[I];
487 const BitValue &V2 = A2[I];
488 if (V1.is(Borrow)) {
489 Res[I] = BitValue::ref(V2);
490 break;
491 }
492 if (V2.is(Borrow))
493 Res[I] = BitValue::ref(V1);
494 else
495 break;
496 }
497 for (; I < W; ++I)
498 Res[I] = BitValue::self();
499 return Res;
500 }
501
502
eMLS(const RegisterCell & A1,const RegisterCell & A2) const503 BT::RegisterCell BT::MachineEvaluator::eMLS(const RegisterCell &A1,
504 const RegisterCell &A2) const {
505 uint16_t W = A1.width() + A2.width();
506 uint16_t Z = A1.ct(0) + A2.ct(0);
507 RegisterCell Res(W);
508 Res.fill(0, Z, BitValue::Zero);
509 Res.fill(Z, W, BitValue::self());
510 return Res;
511 }
512
513
eMLU(const RegisterCell & A1,const RegisterCell & A2) const514 BT::RegisterCell BT::MachineEvaluator::eMLU(const RegisterCell &A1,
515 const RegisterCell &A2) const {
516 uint16_t W = A1.width() + A2.width();
517 uint16_t Z = A1.ct(0) + A2.ct(0);
518 RegisterCell Res(W);
519 Res.fill(0, Z, BitValue::Zero);
520 Res.fill(Z, W, BitValue::self());
521 return Res;
522 }
523
524
eASL(const RegisterCell & A1,uint16_t Sh) const525 BT::RegisterCell BT::MachineEvaluator::eASL(const RegisterCell &A1,
526 uint16_t Sh) const {
527 assert(Sh <= A1.width());
528 RegisterCell Res = RegisterCell::ref(A1);
529 Res.rol(Sh);
530 Res.fill(0, Sh, BitValue::Zero);
531 return Res;
532 }
533
534
eLSR(const RegisterCell & A1,uint16_t Sh) const535 BT::RegisterCell BT::MachineEvaluator::eLSR(const RegisterCell &A1,
536 uint16_t Sh) const {
537 uint16_t W = A1.width();
538 assert(Sh <= W);
539 RegisterCell Res = RegisterCell::ref(A1);
540 Res.rol(W-Sh);
541 Res.fill(W-Sh, W, BitValue::Zero);
542 return Res;
543 }
544
545
eASR(const RegisterCell & A1,uint16_t Sh) const546 BT::RegisterCell BT::MachineEvaluator::eASR(const RegisterCell &A1,
547 uint16_t Sh) const {
548 uint16_t W = A1.width();
549 assert(Sh <= W);
550 RegisterCell Res = RegisterCell::ref(A1);
551 BitValue Sign = Res[W-1];
552 Res.rol(W-Sh);
553 Res.fill(W-Sh, W, Sign);
554 return Res;
555 }
556
557
eAND(const RegisterCell & A1,const RegisterCell & A2) const558 BT::RegisterCell BT::MachineEvaluator::eAND(const RegisterCell &A1,
559 const RegisterCell &A2) const {
560 uint16_t W = A1.width();
561 assert(W == A2.width());
562 RegisterCell Res(W);
563 for (uint16_t i = 0; i < W; ++i) {
564 const BitValue &V1 = A1[i];
565 const BitValue &V2 = A2[i];
566 if (V1.is(1))
567 Res[i] = BitValue::ref(V2);
568 else if (V2.is(1))
569 Res[i] = BitValue::ref(V1);
570 else if (V1.is(0) || V2.is(0))
571 Res[i] = BitValue::Zero;
572 else if (V1 == V2)
573 Res[i] = V1;
574 else
575 Res[i] = BitValue::self();
576 }
577 return Res;
578 }
579
580
eORL(const RegisterCell & A1,const RegisterCell & A2) const581 BT::RegisterCell BT::MachineEvaluator::eORL(const RegisterCell &A1,
582 const RegisterCell &A2) const {
583 uint16_t W = A1.width();
584 assert(W == A2.width());
585 RegisterCell Res(W);
586 for (uint16_t i = 0; i < W; ++i) {
587 const BitValue &V1 = A1[i];
588 const BitValue &V2 = A2[i];
589 if (V1.is(1) || V2.is(1))
590 Res[i] = BitValue::One;
591 else if (V1.is(0))
592 Res[i] = BitValue::ref(V2);
593 else if (V2.is(0))
594 Res[i] = BitValue::ref(V1);
595 else if (V1 == V2)
596 Res[i] = V1;
597 else
598 Res[i] = BitValue::self();
599 }
600 return Res;
601 }
602
603
eXOR(const RegisterCell & A1,const RegisterCell & A2) const604 BT::RegisterCell BT::MachineEvaluator::eXOR(const RegisterCell &A1,
605 const RegisterCell &A2) const {
606 uint16_t W = A1.width();
607 assert(W == A2.width());
608 RegisterCell Res(W);
609 for (uint16_t i = 0; i < W; ++i) {
610 const BitValue &V1 = A1[i];
611 const BitValue &V2 = A2[i];
612 if (V1.is(0))
613 Res[i] = BitValue::ref(V2);
614 else if (V2.is(0))
615 Res[i] = BitValue::ref(V1);
616 else if (V1 == V2)
617 Res[i] = BitValue::Zero;
618 else
619 Res[i] = BitValue::self();
620 }
621 return Res;
622 }
623
624
eNOT(const RegisterCell & A1) const625 BT::RegisterCell BT::MachineEvaluator::eNOT(const RegisterCell &A1) const {
626 uint16_t W = A1.width();
627 RegisterCell Res(W);
628 for (uint16_t i = 0; i < W; ++i) {
629 const BitValue &V = A1[i];
630 if (V.is(0))
631 Res[i] = BitValue::One;
632 else if (V.is(1))
633 Res[i] = BitValue::Zero;
634 else
635 Res[i] = BitValue::self();
636 }
637 return Res;
638 }
639
640
eSET(const RegisterCell & A1,uint16_t BitN) const641 BT::RegisterCell BT::MachineEvaluator::eSET(const RegisterCell &A1,
642 uint16_t BitN) const {
643 assert(BitN < A1.width());
644 RegisterCell Res = RegisterCell::ref(A1);
645 Res[BitN] = BitValue::One;
646 return Res;
647 }
648
649
eCLR(const RegisterCell & A1,uint16_t BitN) const650 BT::RegisterCell BT::MachineEvaluator::eCLR(const RegisterCell &A1,
651 uint16_t BitN) const {
652 assert(BitN < A1.width());
653 RegisterCell Res = RegisterCell::ref(A1);
654 Res[BitN] = BitValue::Zero;
655 return Res;
656 }
657
658
eCLB(const RegisterCell & A1,bool B,uint16_t W) const659 BT::RegisterCell BT::MachineEvaluator::eCLB(const RegisterCell &A1, bool B,
660 uint16_t W) const {
661 uint16_t C = A1.cl(B), AW = A1.width();
662 // If the last leading non-B bit is not a constant, then we don't know
663 // the real count.
664 if ((C < AW && A1[AW-1-C].num()) || C == AW)
665 return eIMM(C, W);
666 return RegisterCell::self(0, W);
667 }
668
669
eCTB(const RegisterCell & A1,bool B,uint16_t W) const670 BT::RegisterCell BT::MachineEvaluator::eCTB(const RegisterCell &A1, bool B,
671 uint16_t W) const {
672 uint16_t C = A1.ct(B), AW = A1.width();
673 // If the last trailing non-B bit is not a constant, then we don't know
674 // the real count.
675 if ((C < AW && A1[C].num()) || C == AW)
676 return eIMM(C, W);
677 return RegisterCell::self(0, W);
678 }
679
680
eSXT(const RegisterCell & A1,uint16_t FromN) const681 BT::RegisterCell BT::MachineEvaluator::eSXT(const RegisterCell &A1,
682 uint16_t FromN) const {
683 uint16_t W = A1.width();
684 assert(FromN <= W);
685 RegisterCell Res = RegisterCell::ref(A1);
686 BitValue Sign = Res[FromN-1];
687 // Sign-extend "inreg".
688 Res.fill(FromN, W, Sign);
689 return Res;
690 }
691
692
eZXT(const RegisterCell & A1,uint16_t FromN) const693 BT::RegisterCell BT::MachineEvaluator::eZXT(const RegisterCell &A1,
694 uint16_t FromN) const {
695 uint16_t W = A1.width();
696 assert(FromN <= W);
697 RegisterCell Res = RegisterCell::ref(A1);
698 Res.fill(FromN, W, BitValue::Zero);
699 return Res;
700 }
701
702
eXTR(const RegisterCell & A1,uint16_t B,uint16_t E) const703 BT::RegisterCell BT::MachineEvaluator::eXTR(const RegisterCell &A1,
704 uint16_t B, uint16_t E) const {
705 uint16_t W = A1.width();
706 assert(B < W && E <= W);
707 if (B == E)
708 return RegisterCell(0);
709 uint16_t Last = (E > 0) ? E-1 : W-1;
710 RegisterCell Res = RegisterCell::ref(A1).extract(BT::BitMask(B, Last));
711 // Return shorter cell.
712 return Res;
713 }
714
715
eINS(const RegisterCell & A1,const RegisterCell & A2,uint16_t AtN) const716 BT::RegisterCell BT::MachineEvaluator::eINS(const RegisterCell &A1,
717 const RegisterCell &A2, uint16_t AtN) const {
718 uint16_t W1 = A1.width(), W2 = A2.width();
719 (void)W1;
720 assert(AtN < W1 && AtN+W2 <= W1);
721 // Copy bits from A1, insert A2 at position AtN.
722 RegisterCell Res = RegisterCell::ref(A1);
723 if (W2 > 0)
724 Res.insert(RegisterCell::ref(A2), BT::BitMask(AtN, AtN+W2-1));
725 return Res;
726 }
727
728
mask(unsigned Reg,unsigned Sub) const729 BT::BitMask BT::MachineEvaluator::mask(unsigned Reg, unsigned Sub) const {
730 assert(Sub == 0 && "Generic BitTracker::mask called for Sub != 0");
731 uint16_t W = getRegBitWidth(Reg);
732 assert(W > 0 && "Cannot generate mask for empty register");
733 return BitMask(0, W-1);
734 }
735
evaluate(const MachineInstr & MI,const CellMapType & Inputs,CellMapType & Outputs) const736 bool BT::MachineEvaluator::evaluate(const MachineInstr &MI,
737 const CellMapType &Inputs,
738 CellMapType &Outputs) const {
739 unsigned Opc = MI.getOpcode();
740 switch (Opc) {
741 case TargetOpcode::REG_SEQUENCE: {
742 RegisterRef RD = MI.getOperand(0);
743 assert(RD.Sub == 0);
744 RegisterRef RS = MI.getOperand(1);
745 unsigned SS = MI.getOperand(2).getImm();
746 RegisterRef RT = MI.getOperand(3);
747 unsigned ST = MI.getOperand(4).getImm();
748 assert(SS != ST);
749
750 uint16_t W = getRegBitWidth(RD);
751 RegisterCell Res(W);
752 Res.insert(RegisterCell::ref(getCell(RS, Inputs)), mask(RD.Reg, SS));
753 Res.insert(RegisterCell::ref(getCell(RT, Inputs)), mask(RD.Reg, ST));
754 putCell(RD, Res, Outputs);
755 break;
756 }
757
758 case TargetOpcode::COPY: {
759 // COPY can transfer a smaller register into a wider one.
760 // If that is the case, fill the remaining high bits with 0.
761 RegisterRef RD = MI.getOperand(0);
762 RegisterRef RS = MI.getOperand(1);
763 assert(RD.Sub == 0);
764 uint16_t WD = getRegBitWidth(RD);
765 uint16_t WS = getRegBitWidth(RS);
766 assert(WD >= WS);
767 RegisterCell Src = getCell(RS, Inputs);
768 RegisterCell Res(WD);
769 Res.insert(Src, BitMask(0, WS-1));
770 Res.fill(WS, WD, BitValue::Zero);
771 putCell(RD, Res, Outputs);
772 break;
773 }
774
775 default:
776 return false;
777 }
778
779 return true;
780 }
781
782
783 // Main W-Z implementation.
784
visitPHI(const MachineInstr & PI)785 void BT::visitPHI(const MachineInstr &PI) {
786 int ThisN = PI.getParent()->getNumber();
787 if (Trace)
788 dbgs() << "Visit FI(BB#" << ThisN << "): " << PI;
789
790 const MachineOperand &MD = PI.getOperand(0);
791 assert(MD.getSubReg() == 0 && "Unexpected sub-register in definition");
792 RegisterRef DefRR(MD);
793 uint16_t DefBW = ME.getRegBitWidth(DefRR);
794
795 RegisterCell DefC = ME.getCell(DefRR, Map);
796 if (DefC == RegisterCell::self(DefRR.Reg, DefBW)) // XXX slow
797 return;
798
799 bool Changed = false;
800
801 for (unsigned i = 1, n = PI.getNumOperands(); i < n; i += 2) {
802 const MachineBasicBlock *PB = PI.getOperand(i + 1).getMBB();
803 int PredN = PB->getNumber();
804 if (Trace)
805 dbgs() << " edge BB#" << PredN << "->BB#" << ThisN;
806 if (!EdgeExec.count(CFGEdge(PredN, ThisN))) {
807 if (Trace)
808 dbgs() << " not executable\n";
809 continue;
810 }
811
812 RegisterRef RU = PI.getOperand(i);
813 RegisterCell ResC = ME.getCell(RU, Map);
814 if (Trace)
815 dbgs() << " input reg: " << PrintReg(RU.Reg, &ME.TRI, RU.Sub)
816 << " cell: " << ResC << "\n";
817 Changed |= DefC.meet(ResC, DefRR.Reg);
818 }
819
820 if (Changed) {
821 if (Trace)
822 dbgs() << "Output: " << PrintReg(DefRR.Reg, &ME.TRI, DefRR.Sub)
823 << " cell: " << DefC << "\n";
824 ME.putCell(DefRR, DefC, Map);
825 visitUsesOf(DefRR.Reg);
826 }
827 }
828
visitNonBranch(const MachineInstr & MI)829 void BT::visitNonBranch(const MachineInstr &MI) {
830 if (Trace) {
831 int ThisN = MI.getParent()->getNumber();
832 dbgs() << "Visit MI(BB#" << ThisN << "): " << MI;
833 }
834 if (MI.isDebugValue())
835 return;
836 assert(!MI.isBranch() && "Unexpected branch instruction");
837
838 CellMapType ResMap;
839 bool Eval = ME.evaluate(MI, Map, ResMap);
840
841 if (Trace && Eval) {
842 for (unsigned i = 0, n = MI.getNumOperands(); i < n; ++i) {
843 const MachineOperand &MO = MI.getOperand(i);
844 if (!MO.isReg() || !MO.isUse())
845 continue;
846 RegisterRef RU(MO);
847 dbgs() << " input reg: " << PrintReg(RU.Reg, &ME.TRI, RU.Sub)
848 << " cell: " << ME.getCell(RU, Map) << "\n";
849 }
850 dbgs() << "Outputs:\n";
851 for (CellMapType::iterator I = ResMap.begin(), E = ResMap.end();
852 I != E; ++I) {
853 RegisterRef RD(I->first);
854 dbgs() << " " << PrintReg(I->first, &ME.TRI) << " cell: "
855 << ME.getCell(RD, ResMap) << "\n";
856 }
857 }
858
859 // Iterate over all definitions of the instruction, and update the
860 // cells accordingly.
861 for (unsigned i = 0, n = MI.getNumOperands(); i < n; ++i) {
862 const MachineOperand &MO = MI.getOperand(i);
863 // Visit register defs only.
864 if (!MO.isReg() || !MO.isDef())
865 continue;
866 RegisterRef RD(MO);
867 assert(RD.Sub == 0 && "Unexpected sub-register in definition");
868 if (!TargetRegisterInfo::isVirtualRegister(RD.Reg))
869 continue;
870
871 bool Changed = false;
872 if (!Eval || ResMap.count(RD.Reg) == 0) {
873 // Set to "ref" (aka "bottom").
874 uint16_t DefBW = ME.getRegBitWidth(RD);
875 RegisterCell RefC = RegisterCell::self(RD.Reg, DefBW);
876 if (RefC != ME.getCell(RD, Map)) {
877 ME.putCell(RD, RefC, Map);
878 Changed = true;
879 }
880 } else {
881 RegisterCell DefC = ME.getCell(RD, Map);
882 RegisterCell ResC = ME.getCell(RD, ResMap);
883 // This is a non-phi instruction, so the values of the inputs come
884 // from the same registers each time this instruction is evaluated.
885 // During the propagation, the values of the inputs can become lowered
886 // in the sense of the lattice operation, which may cause different
887 // results to be calculated in subsequent evaluations. This should
888 // not cause the bottoming of the result in the map, since the new
889 // result is already reflecting the lowered inputs.
890 for (uint16_t i = 0, w = DefC.width(); i < w; ++i) {
891 BitValue &V = DefC[i];
892 // Bits that are already "bottom" should not be updated.
893 if (V.Type == BitValue::Ref && V.RefI.Reg == RD.Reg)
894 continue;
895 // Same for those that are identical in DefC and ResC.
896 if (V == ResC[i])
897 continue;
898 V = ResC[i];
899 Changed = true;
900 }
901 if (Changed)
902 ME.putCell(RD, DefC, Map);
903 }
904 if (Changed)
905 visitUsesOf(RD.Reg);
906 }
907 }
908
visitBranchesFrom(const MachineInstr & BI)909 void BT::visitBranchesFrom(const MachineInstr &BI) {
910 const MachineBasicBlock &B = *BI.getParent();
911 MachineBasicBlock::const_iterator It = BI, End = B.end();
912 BranchTargetList Targets, BTs;
913 bool FallsThrough = true, DefaultToAll = false;
914 int ThisN = B.getNumber();
915
916 do {
917 BTs.clear();
918 const MachineInstr &MI = *It;
919 if (Trace)
920 dbgs() << "Visit BR(BB#" << ThisN << "): " << MI;
921 assert(MI.isBranch() && "Expecting branch instruction");
922 InstrExec.insert(&MI);
923 bool Eval = ME.evaluate(MI, Map, BTs, FallsThrough);
924 if (!Eval) {
925 // If the evaluation failed, we will add all targets. Keep going in
926 // the loop to mark all executable branches as such.
927 DefaultToAll = true;
928 FallsThrough = true;
929 if (Trace)
930 dbgs() << " failed to evaluate: will add all CFG successors\n";
931 } else if (!DefaultToAll) {
932 // If evaluated successfully add the targets to the cumulative list.
933 if (Trace) {
934 dbgs() << " adding targets:";
935 for (unsigned i = 0, n = BTs.size(); i < n; ++i)
936 dbgs() << " BB#" << BTs[i]->getNumber();
937 if (FallsThrough)
938 dbgs() << "\n falls through\n";
939 else
940 dbgs() << "\n does not fall through\n";
941 }
942 Targets.insert(BTs.begin(), BTs.end());
943 }
944 ++It;
945 } while (FallsThrough && It != End);
946
947 typedef MachineBasicBlock::const_succ_iterator succ_iterator;
948 if (!DefaultToAll) {
949 // Need to add all CFG successors that lead to EH landing pads.
950 // There won't be explicit branches to these blocks, but they must
951 // be processed.
952 for (succ_iterator I = B.succ_begin(), E = B.succ_end(); I != E; ++I) {
953 const MachineBasicBlock *SB = *I;
954 if (SB->isEHPad())
955 Targets.insert(SB);
956 }
957 if (FallsThrough) {
958 MachineFunction::const_iterator BIt = B.getIterator();
959 MachineFunction::const_iterator Next = std::next(BIt);
960 if (Next != MF.end())
961 Targets.insert(&*Next);
962 }
963 } else {
964 for (succ_iterator I = B.succ_begin(), E = B.succ_end(); I != E; ++I)
965 Targets.insert(*I);
966 }
967
968 for (unsigned i = 0, n = Targets.size(); i < n; ++i) {
969 int TargetN = Targets[i]->getNumber();
970 FlowQ.push(CFGEdge(ThisN, TargetN));
971 }
972 }
973
974
visitUsesOf(unsigned Reg)975 void BT::visitUsesOf(unsigned Reg) {
976 if (Trace)
977 dbgs() << "visiting uses of " << PrintReg(Reg, &ME.TRI) << "\n";
978
979 typedef MachineRegisterInfo::use_nodbg_iterator use_iterator;
980 use_iterator End = MRI.use_nodbg_end();
981 for (use_iterator I = MRI.use_nodbg_begin(Reg); I != End; ++I) {
982 MachineInstr *UseI = I->getParent();
983 if (!InstrExec.count(UseI))
984 continue;
985 if (UseI->isPHI())
986 visitPHI(*UseI);
987 else if (!UseI->isBranch())
988 visitNonBranch(*UseI);
989 else
990 visitBranchesFrom(*UseI);
991 }
992 }
993
994
get(RegisterRef RR) const995 BT::RegisterCell BT::get(RegisterRef RR) const {
996 return ME.getCell(RR, Map);
997 }
998
999
put(RegisterRef RR,const RegisterCell & RC)1000 void BT::put(RegisterRef RR, const RegisterCell &RC) {
1001 ME.putCell(RR, RC, Map);
1002 }
1003
1004
1005 // Replace all references to bits from OldRR with the corresponding bits
1006 // in NewRR.
subst(RegisterRef OldRR,RegisterRef NewRR)1007 void BT::subst(RegisterRef OldRR, RegisterRef NewRR) {
1008 assert(Map.count(OldRR.Reg) > 0 && "OldRR not present in map");
1009 BitMask OM = ME.mask(OldRR.Reg, OldRR.Sub);
1010 BitMask NM = ME.mask(NewRR.Reg, NewRR.Sub);
1011 uint16_t OMB = OM.first(), OME = OM.last();
1012 uint16_t NMB = NM.first(), NME = NM.last();
1013 (void)NME;
1014 assert((OME-OMB == NME-NMB) &&
1015 "Substituting registers of different lengths");
1016 for (CellMapType::iterator I = Map.begin(), E = Map.end(); I != E; ++I) {
1017 RegisterCell &RC = I->second;
1018 for (uint16_t i = 0, w = RC.width(); i < w; ++i) {
1019 BitValue &V = RC[i];
1020 if (V.Type != BitValue::Ref || V.RefI.Reg != OldRR.Reg)
1021 continue;
1022 if (V.RefI.Pos < OMB || V.RefI.Pos > OME)
1023 continue;
1024 V.RefI.Reg = NewRR.Reg;
1025 V.RefI.Pos += NMB-OMB;
1026 }
1027 }
1028 }
1029
1030
1031 // Check if the block has been "executed" during propagation. (If not, the
1032 // block is dead, but it may still appear to be reachable.)
reached(const MachineBasicBlock * B) const1033 bool BT::reached(const MachineBasicBlock *B) const {
1034 int BN = B->getNumber();
1035 assert(BN >= 0);
1036 for (EdgeSetType::iterator I = EdgeExec.begin(), E = EdgeExec.end();
1037 I != E; ++I) {
1038 if (I->second == BN)
1039 return true;
1040 }
1041 return false;
1042 }
1043
1044
reset()1045 void BT::reset() {
1046 EdgeExec.clear();
1047 InstrExec.clear();
1048 Map.clear();
1049 }
1050
1051
run()1052 void BT::run() {
1053 reset();
1054 assert(FlowQ.empty());
1055
1056 typedef GraphTraits<const MachineFunction*> MachineFlowGraphTraits;
1057 const MachineBasicBlock *Entry = MachineFlowGraphTraits::getEntryNode(&MF);
1058
1059 unsigned MaxBN = 0;
1060 for (MachineFunction::const_iterator I = MF.begin(), E = MF.end();
1061 I != E; ++I) {
1062 assert(I->getNumber() >= 0 && "Disconnected block");
1063 unsigned BN = I->getNumber();
1064 if (BN > MaxBN)
1065 MaxBN = BN;
1066 }
1067
1068 // Keep track of visited blocks.
1069 BitVector BlockScanned(MaxBN+1);
1070
1071 int EntryN = Entry->getNumber();
1072 // Generate a fake edge to get something to start with.
1073 FlowQ.push(CFGEdge(-1, EntryN));
1074
1075 while (!FlowQ.empty()) {
1076 CFGEdge Edge = FlowQ.front();
1077 FlowQ.pop();
1078
1079 if (EdgeExec.count(Edge))
1080 continue;
1081 EdgeExec.insert(Edge);
1082
1083 const MachineBasicBlock &B = *MF.getBlockNumbered(Edge.second);
1084 MachineBasicBlock::const_iterator It = B.begin(), End = B.end();
1085 // Visit PHI nodes first.
1086 while (It != End && It->isPHI()) {
1087 const MachineInstr &PI = *It++;
1088 InstrExec.insert(&PI);
1089 visitPHI(PI);
1090 }
1091
1092 // If this block has already been visited through a flow graph edge,
1093 // then the instructions have already been processed. Any updates to
1094 // the cells would now only happen through visitUsesOf...
1095 if (BlockScanned[Edge.second])
1096 continue;
1097 BlockScanned[Edge.second] = true;
1098
1099 // Visit non-branch instructions.
1100 while (It != End && !It->isBranch()) {
1101 const MachineInstr &MI = *It++;
1102 InstrExec.insert(&MI);
1103 visitNonBranch(MI);
1104 }
1105 // If block end has been reached, add the fall-through edge to the queue.
1106 if (It == End) {
1107 MachineFunction::const_iterator BIt = B.getIterator();
1108 MachineFunction::const_iterator Next = std::next(BIt);
1109 if (Next != MF.end() && B.isSuccessor(&*Next)) {
1110 int ThisN = B.getNumber();
1111 int NextN = Next->getNumber();
1112 FlowQ.push(CFGEdge(ThisN, NextN));
1113 }
1114 } else {
1115 // Handle the remaining sequence of branches. This function will update
1116 // the work queue.
1117 visitBranchesFrom(*It);
1118 }
1119 } // while (!FlowQ->empty())
1120
1121 if (Trace) {
1122 dbgs() << "Cells after propagation:\n";
1123 for (CellMapType::iterator I = Map.begin(), E = Map.end(); I != E; ++I)
1124 dbgs() << PrintReg(I->first, &ME.TRI) << " -> " << I->second << "\n";
1125 }
1126 }
1127
1128