1 //===-- FastISel.cpp - Implementation of the FastISel class ---------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file contains the implementation of the FastISel class.
11 //
12 // "Fast" instruction selection is designed to emit very poor code quickly.
13 // Also, it is not designed to be able to do much lowering, so most illegal
14 // types (e.g. i64 on 32-bit targets) and operations are not supported. It is
15 // also not intended to be able to do much optimization, except in a few cases
16 // where doing optimizations reduces overall compile time. For example, folding
17 // constants into immediate fields is often done, because it's cheap and it
18 // reduces the number of instructions later phases have to examine.
19 //
20 // "Fast" instruction selection is able to fail gracefully and transfer
21 // control to the SelectionDAG selector for operations that it doesn't
22 // support. In many cases, this allows us to avoid duplicating a lot of
23 // the complicated lowering logic that SelectionDAG currently has.
24 //
25 // The intended use for "fast" instruction selection is "-O0" mode
26 // compilation, where the quality of the generated code is irrelevant when
27 // weighed against the speed at which the code can be generated. Also,
28 // at -O0, the LLVM optimizers are not running, and this makes the
29 // compile time of codegen a much higher portion of the overall compile
30 // time. Despite its limitations, "fast" instruction selection is able to
31 // handle enough code on its own to provide noticeable overall speedups
32 // in -O0 compiles.
33 //
34 // Basic operations are supported in a target-independent way, by reading
35 // the same instruction descriptions that the SelectionDAG selector reads,
36 // and identifying simple arithmetic operations that can be directly selected
37 // from simple operators. More complicated operations currently require
38 // target-specific code.
39 //
40 //===----------------------------------------------------------------------===//
41
42 #define DEBUG_TYPE "isel"
43 #include "llvm/CodeGen/FastISel.h"
44 #include "llvm/ADT/Statistic.h"
45 #include "llvm/Analysis/Loads.h"
46 #include "llvm/CodeGen/Analysis.h"
47 #include "llvm/CodeGen/FunctionLoweringInfo.h"
48 #include "llvm/CodeGen/MachineInstrBuilder.h"
49 #include "llvm/CodeGen/MachineModuleInfo.h"
50 #include "llvm/CodeGen/MachineRegisterInfo.h"
51 #include "llvm/DebugInfo.h"
52 #include "llvm/IR/DataLayout.h"
53 #include "llvm/IR/Function.h"
54 #include "llvm/IR/GlobalVariable.h"
55 #include "llvm/IR/Instructions.h"
56 #include "llvm/IR/IntrinsicInst.h"
57 #include "llvm/IR/Operator.h"
58 #include "llvm/Support/Debug.h"
59 #include "llvm/Support/ErrorHandling.h"
60 #include "llvm/Target/TargetInstrInfo.h"
61 #include "llvm/Target/TargetLibraryInfo.h"
62 #include "llvm/Target/TargetLowering.h"
63 #include "llvm/Target/TargetMachine.h"
64 using namespace llvm;
65
66 STATISTIC(NumFastIselSuccessIndependent, "Number of insts selected by "
67 "target-independent selector");
68 STATISTIC(NumFastIselSuccessTarget, "Number of insts selected by "
69 "target-specific selector");
70 STATISTIC(NumFastIselDead, "Number of dead insts removed on failure");
71
72 /// startNewBlock - Set the current block to which generated machine
73 /// instructions will be appended, and clear the local CSE map.
74 ///
startNewBlock()75 void FastISel::startNewBlock() {
76 LocalValueMap.clear();
77
78 EmitStartPt = 0;
79
80 // Advance the emit start point past any EH_LABEL instructions.
81 MachineBasicBlock::iterator
82 I = FuncInfo.MBB->begin(), E = FuncInfo.MBB->end();
83 while (I != E && I->getOpcode() == TargetOpcode::EH_LABEL) {
84 EmitStartPt = I;
85 ++I;
86 }
87 LastLocalValue = EmitStartPt;
88 }
89
LowerArguments()90 bool FastISel::LowerArguments() {
91 if (!FuncInfo.CanLowerReturn)
92 // Fallback to SDISel argument lowering code to deal with sret pointer
93 // parameter.
94 return false;
95
96 if (!FastLowerArguments())
97 return false;
98
99 // Enter non-dead arguments into ValueMap for uses in non-entry BBs.
100 for (Function::const_arg_iterator I = FuncInfo.Fn->arg_begin(),
101 E = FuncInfo.Fn->arg_end(); I != E; ++I) {
102 if (!I->use_empty()) {
103 DenseMap<const Value *, unsigned>::iterator VI = LocalValueMap.find(I);
104 assert(VI != LocalValueMap.end() && "Missed an argument?");
105 FuncInfo.ValueMap[I] = VI->second;
106 }
107 }
108 return true;
109 }
110
flushLocalValueMap()111 void FastISel::flushLocalValueMap() {
112 LocalValueMap.clear();
113 LastLocalValue = EmitStartPt;
114 recomputeInsertPt();
115 }
116
hasTrivialKill(const Value * V) const117 bool FastISel::hasTrivialKill(const Value *V) const {
118 // Don't consider constants or arguments to have trivial kills.
119 const Instruction *I = dyn_cast<Instruction>(V);
120 if (!I)
121 return false;
122
123 // No-op casts are trivially coalesced by fast-isel.
124 if (const CastInst *Cast = dyn_cast<CastInst>(I))
125 if (Cast->isNoopCast(TD.getIntPtrType(Cast->getContext())) &&
126 !hasTrivialKill(Cast->getOperand(0)))
127 return false;
128
129 // GEPs with all zero indices are trivially coalesced by fast-isel.
130 if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I))
131 if (GEP->hasAllZeroIndices() && !hasTrivialKill(GEP->getOperand(0)))
132 return false;
133
134 // Only instructions with a single use in the same basic block are considered
135 // to have trivial kills.
136 return I->hasOneUse() &&
137 !(I->getOpcode() == Instruction::BitCast ||
138 I->getOpcode() == Instruction::PtrToInt ||
139 I->getOpcode() == Instruction::IntToPtr) &&
140 cast<Instruction>(*I->use_begin())->getParent() == I->getParent();
141 }
142
getRegForValue(const Value * V)143 unsigned FastISel::getRegForValue(const Value *V) {
144 EVT RealVT = TLI.getValueType(V->getType(), /*AllowUnknown=*/true);
145 // Don't handle non-simple values in FastISel.
146 if (!RealVT.isSimple())
147 return 0;
148
149 // Ignore illegal types. We must do this before looking up the value
150 // in ValueMap because Arguments are given virtual registers regardless
151 // of whether FastISel can handle them.
152 MVT VT = RealVT.getSimpleVT();
153 if (!TLI.isTypeLegal(VT)) {
154 // Handle integer promotions, though, because they're common and easy.
155 if (VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16)
156 VT = TLI.getTypeToTransformTo(V->getContext(), VT).getSimpleVT();
157 else
158 return 0;
159 }
160
161 // Look up the value to see if we already have a register for it.
162 unsigned Reg = lookUpRegForValue(V);
163 if (Reg != 0)
164 return Reg;
165
166 // In bottom-up mode, just create the virtual register which will be used
167 // to hold the value. It will be materialized later.
168 if (isa<Instruction>(V) &&
169 (!isa<AllocaInst>(V) ||
170 !FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(V))))
171 return FuncInfo.InitializeRegForValue(V);
172
173 SavePoint SaveInsertPt = enterLocalValueArea();
174
175 // Materialize the value in a register. Emit any instructions in the
176 // local value area.
177 Reg = materializeRegForValue(V, VT);
178
179 leaveLocalValueArea(SaveInsertPt);
180
181 return Reg;
182 }
183
184 /// materializeRegForValue - Helper for getRegForValue. This function is
185 /// called when the value isn't already available in a register and must
186 /// be materialized with new instructions.
materializeRegForValue(const Value * V,MVT VT)187 unsigned FastISel::materializeRegForValue(const Value *V, MVT VT) {
188 unsigned Reg = 0;
189
190 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
191 if (CI->getValue().getActiveBits() <= 64)
192 Reg = FastEmit_i(VT, VT, ISD::Constant, CI->getZExtValue());
193 } else if (isa<AllocaInst>(V)) {
194 Reg = TargetMaterializeAlloca(cast<AllocaInst>(V));
195 } else if (isa<ConstantPointerNull>(V)) {
196 // Translate this as an integer zero so that it can be
197 // local-CSE'd with actual integer zeros.
198 Reg =
199 getRegForValue(Constant::getNullValue(TD.getIntPtrType(V->getContext())));
200 } else if (const ConstantFP *CF = dyn_cast<ConstantFP>(V)) {
201 if (CF->isNullValue()) {
202 Reg = TargetMaterializeFloatZero(CF);
203 } else {
204 // Try to emit the constant directly.
205 Reg = FastEmit_f(VT, VT, ISD::ConstantFP, CF);
206 }
207
208 if (!Reg) {
209 // Try to emit the constant by using an integer constant with a cast.
210 const APFloat &Flt = CF->getValueAPF();
211 EVT IntVT = TLI.getPointerTy();
212
213 uint64_t x[2];
214 uint32_t IntBitWidth = IntVT.getSizeInBits();
215 bool isExact;
216 (void) Flt.convertToInteger(x, IntBitWidth, /*isSigned=*/true,
217 APFloat::rmTowardZero, &isExact);
218 if (isExact) {
219 APInt IntVal(IntBitWidth, x);
220
221 unsigned IntegerReg =
222 getRegForValue(ConstantInt::get(V->getContext(), IntVal));
223 if (IntegerReg != 0)
224 Reg = FastEmit_r(IntVT.getSimpleVT(), VT, ISD::SINT_TO_FP,
225 IntegerReg, /*Kill=*/false);
226 }
227 }
228 } else if (const Operator *Op = dyn_cast<Operator>(V)) {
229 if (!SelectOperator(Op, Op->getOpcode()))
230 if (!isa<Instruction>(Op) ||
231 !TargetSelectInstruction(cast<Instruction>(Op)))
232 return 0;
233 Reg = lookUpRegForValue(Op);
234 } else if (isa<UndefValue>(V)) {
235 Reg = createResultReg(TLI.getRegClassFor(VT));
236 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
237 TII.get(TargetOpcode::IMPLICIT_DEF), Reg);
238 }
239
240 // If target-independent code couldn't handle the value, give target-specific
241 // code a try.
242 if (!Reg && isa<Constant>(V))
243 Reg = TargetMaterializeConstant(cast<Constant>(V));
244
245 // Don't cache constant materializations in the general ValueMap.
246 // To do so would require tracking what uses they dominate.
247 if (Reg != 0) {
248 LocalValueMap[V] = Reg;
249 LastLocalValue = MRI.getVRegDef(Reg);
250 }
251 return Reg;
252 }
253
lookUpRegForValue(const Value * V)254 unsigned FastISel::lookUpRegForValue(const Value *V) {
255 // Look up the value to see if we already have a register for it. We
256 // cache values defined by Instructions across blocks, and other values
257 // only locally. This is because Instructions already have the SSA
258 // def-dominates-use requirement enforced.
259 DenseMap<const Value *, unsigned>::iterator I = FuncInfo.ValueMap.find(V);
260 if (I != FuncInfo.ValueMap.end())
261 return I->second;
262 return LocalValueMap[V];
263 }
264
265 /// UpdateValueMap - Update the value map to include the new mapping for this
266 /// instruction, or insert an extra copy to get the result in a previous
267 /// determined register.
268 /// NOTE: This is only necessary because we might select a block that uses
269 /// a value before we select the block that defines the value. It might be
270 /// possible to fix this by selecting blocks in reverse postorder.
UpdateValueMap(const Value * I,unsigned Reg,unsigned NumRegs)271 void FastISel::UpdateValueMap(const Value *I, unsigned Reg, unsigned NumRegs) {
272 if (!isa<Instruction>(I)) {
273 LocalValueMap[I] = Reg;
274 return;
275 }
276
277 unsigned &AssignedReg = FuncInfo.ValueMap[I];
278 if (AssignedReg == 0)
279 // Use the new register.
280 AssignedReg = Reg;
281 else if (Reg != AssignedReg) {
282 // Arrange for uses of AssignedReg to be replaced by uses of Reg.
283 for (unsigned i = 0; i < NumRegs; i++)
284 FuncInfo.RegFixups[AssignedReg+i] = Reg+i;
285
286 AssignedReg = Reg;
287 }
288 }
289
getRegForGEPIndex(const Value * Idx)290 std::pair<unsigned, bool> FastISel::getRegForGEPIndex(const Value *Idx) {
291 unsigned IdxN = getRegForValue(Idx);
292 if (IdxN == 0)
293 // Unhandled operand. Halt "fast" selection and bail.
294 return std::pair<unsigned, bool>(0, false);
295
296 bool IdxNIsKill = hasTrivialKill(Idx);
297
298 // If the index is smaller or larger than intptr_t, truncate or extend it.
299 MVT PtrVT = TLI.getPointerTy();
300 EVT IdxVT = EVT::getEVT(Idx->getType(), /*HandleUnknown=*/false);
301 if (IdxVT.bitsLT(PtrVT)) {
302 IdxN = FastEmit_r(IdxVT.getSimpleVT(), PtrVT, ISD::SIGN_EXTEND,
303 IdxN, IdxNIsKill);
304 IdxNIsKill = true;
305 }
306 else if (IdxVT.bitsGT(PtrVT)) {
307 IdxN = FastEmit_r(IdxVT.getSimpleVT(), PtrVT, ISD::TRUNCATE,
308 IdxN, IdxNIsKill);
309 IdxNIsKill = true;
310 }
311 return std::pair<unsigned, bool>(IdxN, IdxNIsKill);
312 }
313
recomputeInsertPt()314 void FastISel::recomputeInsertPt() {
315 if (getLastLocalValue()) {
316 FuncInfo.InsertPt = getLastLocalValue();
317 FuncInfo.MBB = FuncInfo.InsertPt->getParent();
318 ++FuncInfo.InsertPt;
319 } else
320 FuncInfo.InsertPt = FuncInfo.MBB->getFirstNonPHI();
321
322 // Now skip past any EH_LABELs, which must remain at the beginning.
323 while (FuncInfo.InsertPt != FuncInfo.MBB->end() &&
324 FuncInfo.InsertPt->getOpcode() == TargetOpcode::EH_LABEL)
325 ++FuncInfo.InsertPt;
326 }
327
removeDeadCode(MachineBasicBlock::iterator I,MachineBasicBlock::iterator E)328 void FastISel::removeDeadCode(MachineBasicBlock::iterator I,
329 MachineBasicBlock::iterator E) {
330 assert (I && E && std::distance(I, E) > 0 && "Invalid iterator!");
331 while (I != E) {
332 MachineInstr *Dead = &*I;
333 ++I;
334 Dead->eraseFromParent();
335 ++NumFastIselDead;
336 }
337 recomputeInsertPt();
338 }
339
enterLocalValueArea()340 FastISel::SavePoint FastISel::enterLocalValueArea() {
341 MachineBasicBlock::iterator OldInsertPt = FuncInfo.InsertPt;
342 DebugLoc OldDL = DL;
343 recomputeInsertPt();
344 DL = DebugLoc();
345 SavePoint SP = { OldInsertPt, OldDL };
346 return SP;
347 }
348
leaveLocalValueArea(SavePoint OldInsertPt)349 void FastISel::leaveLocalValueArea(SavePoint OldInsertPt) {
350 if (FuncInfo.InsertPt != FuncInfo.MBB->begin())
351 LastLocalValue = llvm::prior(FuncInfo.InsertPt);
352
353 // Restore the previous insert position.
354 FuncInfo.InsertPt = OldInsertPt.InsertPt;
355 DL = OldInsertPt.DL;
356 }
357
358 /// SelectBinaryOp - Select and emit code for a binary operator instruction,
359 /// which has an opcode which directly corresponds to the given ISD opcode.
360 ///
SelectBinaryOp(const User * I,unsigned ISDOpcode)361 bool FastISel::SelectBinaryOp(const User *I, unsigned ISDOpcode) {
362 EVT VT = EVT::getEVT(I->getType(), /*HandleUnknown=*/true);
363 if (VT == MVT::Other || !VT.isSimple())
364 // Unhandled type. Halt "fast" selection and bail.
365 return false;
366
367 // We only handle legal types. For example, on x86-32 the instruction
368 // selector contains all of the 64-bit instructions from x86-64,
369 // under the assumption that i64 won't be used if the target doesn't
370 // support it.
371 if (!TLI.isTypeLegal(VT)) {
372 // MVT::i1 is special. Allow AND, OR, or XOR because they
373 // don't require additional zeroing, which makes them easy.
374 if (VT == MVT::i1 &&
375 (ISDOpcode == ISD::AND || ISDOpcode == ISD::OR ||
376 ISDOpcode == ISD::XOR))
377 VT = TLI.getTypeToTransformTo(I->getContext(), VT);
378 else
379 return false;
380 }
381
382 // Check if the first operand is a constant, and handle it as "ri". At -O0,
383 // we don't have anything that canonicalizes operand order.
384 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(0)))
385 if (isa<Instruction>(I) && cast<Instruction>(I)->isCommutative()) {
386 unsigned Op1 = getRegForValue(I->getOperand(1));
387 if (Op1 == 0) return false;
388
389 bool Op1IsKill = hasTrivialKill(I->getOperand(1));
390
391 unsigned ResultReg = FastEmit_ri_(VT.getSimpleVT(), ISDOpcode, Op1,
392 Op1IsKill, CI->getZExtValue(),
393 VT.getSimpleVT());
394 if (ResultReg == 0) return false;
395
396 // We successfully emitted code for the given LLVM Instruction.
397 UpdateValueMap(I, ResultReg);
398 return true;
399 }
400
401
402 unsigned Op0 = getRegForValue(I->getOperand(0));
403 if (Op0 == 0) // Unhandled operand. Halt "fast" selection and bail.
404 return false;
405
406 bool Op0IsKill = hasTrivialKill(I->getOperand(0));
407
408 // Check if the second operand is a constant and handle it appropriately.
409 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
410 uint64_t Imm = CI->getZExtValue();
411
412 // Transform "sdiv exact X, 8" -> "sra X, 3".
413 if (ISDOpcode == ISD::SDIV && isa<BinaryOperator>(I) &&
414 cast<BinaryOperator>(I)->isExact() &&
415 isPowerOf2_64(Imm)) {
416 Imm = Log2_64(Imm);
417 ISDOpcode = ISD::SRA;
418 }
419
420 // Transform "urem x, pow2" -> "and x, pow2-1".
421 if (ISDOpcode == ISD::UREM && isa<BinaryOperator>(I) &&
422 isPowerOf2_64(Imm)) {
423 --Imm;
424 ISDOpcode = ISD::AND;
425 }
426
427 unsigned ResultReg = FastEmit_ri_(VT.getSimpleVT(), ISDOpcode, Op0,
428 Op0IsKill, Imm, VT.getSimpleVT());
429 if (ResultReg == 0) return false;
430
431 // We successfully emitted code for the given LLVM Instruction.
432 UpdateValueMap(I, ResultReg);
433 return true;
434 }
435
436 // Check if the second operand is a constant float.
437 if (ConstantFP *CF = dyn_cast<ConstantFP>(I->getOperand(1))) {
438 unsigned ResultReg = FastEmit_rf(VT.getSimpleVT(), VT.getSimpleVT(),
439 ISDOpcode, Op0, Op0IsKill, CF);
440 if (ResultReg != 0) {
441 // We successfully emitted code for the given LLVM Instruction.
442 UpdateValueMap(I, ResultReg);
443 return true;
444 }
445 }
446
447 unsigned Op1 = getRegForValue(I->getOperand(1));
448 if (Op1 == 0)
449 // Unhandled operand. Halt "fast" selection and bail.
450 return false;
451
452 bool Op1IsKill = hasTrivialKill(I->getOperand(1));
453
454 // Now we have both operands in registers. Emit the instruction.
455 unsigned ResultReg = FastEmit_rr(VT.getSimpleVT(), VT.getSimpleVT(),
456 ISDOpcode,
457 Op0, Op0IsKill,
458 Op1, Op1IsKill);
459 if (ResultReg == 0)
460 // Target-specific code wasn't able to find a machine opcode for
461 // the given ISD opcode and type. Halt "fast" selection and bail.
462 return false;
463
464 // We successfully emitted code for the given LLVM Instruction.
465 UpdateValueMap(I, ResultReg);
466 return true;
467 }
468
SelectGetElementPtr(const User * I)469 bool FastISel::SelectGetElementPtr(const User *I) {
470 unsigned N = getRegForValue(I->getOperand(0));
471 if (N == 0)
472 // Unhandled operand. Halt "fast" selection and bail.
473 return false;
474
475 bool NIsKill = hasTrivialKill(I->getOperand(0));
476
477 // Keep a running tab of the total offset to coalesce multiple N = N + Offset
478 // into a single N = N + TotalOffset.
479 uint64_t TotalOffs = 0;
480 // FIXME: What's a good SWAG number for MaxOffs?
481 uint64_t MaxOffs = 2048;
482 Type *Ty = I->getOperand(0)->getType();
483 MVT VT = TLI.getPointerTy();
484 for (GetElementPtrInst::const_op_iterator OI = I->op_begin()+1,
485 E = I->op_end(); OI != E; ++OI) {
486 const Value *Idx = *OI;
487 if (StructType *StTy = dyn_cast<StructType>(Ty)) {
488 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
489 if (Field) {
490 // N = N + Offset
491 TotalOffs += TD.getStructLayout(StTy)->getElementOffset(Field);
492 if (TotalOffs >= MaxOffs) {
493 N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, TotalOffs, VT);
494 if (N == 0)
495 // Unhandled operand. Halt "fast" selection and bail.
496 return false;
497 NIsKill = true;
498 TotalOffs = 0;
499 }
500 }
501 Ty = StTy->getElementType(Field);
502 } else {
503 Ty = cast<SequentialType>(Ty)->getElementType();
504
505 // If this is a constant subscript, handle it quickly.
506 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
507 if (CI->isZero()) continue;
508 // N = N + Offset
509 TotalOffs +=
510 TD.getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
511 if (TotalOffs >= MaxOffs) {
512 N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, TotalOffs, VT);
513 if (N == 0)
514 // Unhandled operand. Halt "fast" selection and bail.
515 return false;
516 NIsKill = true;
517 TotalOffs = 0;
518 }
519 continue;
520 }
521 if (TotalOffs) {
522 N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, TotalOffs, VT);
523 if (N == 0)
524 // Unhandled operand. Halt "fast" selection and bail.
525 return false;
526 NIsKill = true;
527 TotalOffs = 0;
528 }
529
530 // N = N + Idx * ElementSize;
531 uint64_t ElementSize = TD.getTypeAllocSize(Ty);
532 std::pair<unsigned, bool> Pair = getRegForGEPIndex(Idx);
533 unsigned IdxN = Pair.first;
534 bool IdxNIsKill = Pair.second;
535 if (IdxN == 0)
536 // Unhandled operand. Halt "fast" selection and bail.
537 return false;
538
539 if (ElementSize != 1) {
540 IdxN = FastEmit_ri_(VT, ISD::MUL, IdxN, IdxNIsKill, ElementSize, VT);
541 if (IdxN == 0)
542 // Unhandled operand. Halt "fast" selection and bail.
543 return false;
544 IdxNIsKill = true;
545 }
546 N = FastEmit_rr(VT, VT, ISD::ADD, N, NIsKill, IdxN, IdxNIsKill);
547 if (N == 0)
548 // Unhandled operand. Halt "fast" selection and bail.
549 return false;
550 }
551 }
552 if (TotalOffs) {
553 N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, TotalOffs, VT);
554 if (N == 0)
555 // Unhandled operand. Halt "fast" selection and bail.
556 return false;
557 }
558
559 // We successfully emitted code for the given LLVM Instruction.
560 UpdateValueMap(I, N);
561 return true;
562 }
563
SelectCall(const User * I)564 bool FastISel::SelectCall(const User *I) {
565 const CallInst *Call = cast<CallInst>(I);
566
567 // Handle simple inline asms.
568 if (const InlineAsm *IA = dyn_cast<InlineAsm>(Call->getCalledValue())) {
569 // Don't attempt to handle constraints.
570 if (!IA->getConstraintString().empty())
571 return false;
572
573 unsigned ExtraInfo = 0;
574 if (IA->hasSideEffects())
575 ExtraInfo |= InlineAsm::Extra_HasSideEffects;
576 if (IA->isAlignStack())
577 ExtraInfo |= InlineAsm::Extra_IsAlignStack;
578
579 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
580 TII.get(TargetOpcode::INLINEASM))
581 .addExternalSymbol(IA->getAsmString().c_str())
582 .addImm(ExtraInfo);
583 return true;
584 }
585
586 MachineModuleInfo &MMI = FuncInfo.MF->getMMI();
587 ComputeUsesVAFloatArgument(*Call, &MMI);
588
589 const Function *F = Call->getCalledFunction();
590 if (!F) return false;
591
592 // Handle selected intrinsic function calls.
593 switch (F->getIntrinsicID()) {
594 default: break;
595 // At -O0 we don't care about the lifetime intrinsics.
596 case Intrinsic::lifetime_start:
597 case Intrinsic::lifetime_end:
598 // The donothing intrinsic does, well, nothing.
599 case Intrinsic::donothing:
600 return true;
601
602 case Intrinsic::dbg_declare: {
603 const DbgDeclareInst *DI = cast<DbgDeclareInst>(Call);
604 if (!DIVariable(DI->getVariable()).Verify() ||
605 !FuncInfo.MF->getMMI().hasDebugInfo()) {
606 DEBUG(dbgs() << "Dropping debug info for " << *DI << "\n");
607 return true;
608 }
609
610 const Value *Address = DI->getAddress();
611 if (!Address || isa<UndefValue>(Address)) {
612 DEBUG(dbgs() << "Dropping debug info for " << *DI << "\n");
613 return true;
614 }
615
616 unsigned Reg = 0;
617 unsigned Offset = 0;
618 if (const Argument *Arg = dyn_cast<Argument>(Address)) {
619 // Some arguments' frame index is recorded during argument lowering.
620 Offset = FuncInfo.getArgumentFrameIndex(Arg);
621 if (Offset)
622 Reg = TRI.getFrameRegister(*FuncInfo.MF);
623 }
624 if (!Reg)
625 Reg = lookUpRegForValue(Address);
626
627 // If we have a VLA that has a "use" in a metadata node that's then used
628 // here but it has no other uses, then we have a problem. E.g.,
629 //
630 // int foo (const int *x) {
631 // char a[*x];
632 // return 0;
633 // }
634 //
635 // If we assign 'a' a vreg and fast isel later on has to use the selection
636 // DAG isel, it will want to copy the value to the vreg. However, there are
637 // no uses, which goes counter to what selection DAG isel expects.
638 if (!Reg && !Address->use_empty() && isa<Instruction>(Address) &&
639 (!isa<AllocaInst>(Address) ||
640 !FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(Address))))
641 Reg = FuncInfo.InitializeRegForValue(Address);
642
643 if (Reg)
644 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
645 TII.get(TargetOpcode::DBG_VALUE))
646 .addReg(Reg, RegState::Debug).addImm(Offset)
647 .addMetadata(DI->getVariable());
648 else
649 // We can't yet handle anything else here because it would require
650 // generating code, thus altering codegen because of debug info.
651 DEBUG(dbgs() << "Dropping debug info for " << DI);
652 return true;
653 }
654 case Intrinsic::dbg_value: {
655 // This form of DBG_VALUE is target-independent.
656 const DbgValueInst *DI = cast<DbgValueInst>(Call);
657 const MCInstrDesc &II = TII.get(TargetOpcode::DBG_VALUE);
658 const Value *V = DI->getValue();
659 if (!V) {
660 // Currently the optimizer can produce this; insert an undef to
661 // help debugging. Probably the optimizer should not do this.
662 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
663 .addReg(0U).addImm(DI->getOffset())
664 .addMetadata(DI->getVariable());
665 } else if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
666 if (CI->getBitWidth() > 64)
667 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
668 .addCImm(CI).addImm(DI->getOffset())
669 .addMetadata(DI->getVariable());
670 else
671 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
672 .addImm(CI->getZExtValue()).addImm(DI->getOffset())
673 .addMetadata(DI->getVariable());
674 } else if (const ConstantFP *CF = dyn_cast<ConstantFP>(V)) {
675 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
676 .addFPImm(CF).addImm(DI->getOffset())
677 .addMetadata(DI->getVariable());
678 } else if (unsigned Reg = lookUpRegForValue(V)) {
679 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
680 .addReg(Reg, RegState::Debug).addImm(DI->getOffset())
681 .addMetadata(DI->getVariable());
682 } else {
683 // We can't yet handle anything else here because it would require
684 // generating code, thus altering codegen because of debug info.
685 DEBUG(dbgs() << "Dropping debug info for " << DI);
686 }
687 return true;
688 }
689 case Intrinsic::objectsize: {
690 ConstantInt *CI = cast<ConstantInt>(Call->getArgOperand(1));
691 unsigned long long Res = CI->isZero() ? -1ULL : 0;
692 Constant *ResCI = ConstantInt::get(Call->getType(), Res);
693 unsigned ResultReg = getRegForValue(ResCI);
694 if (ResultReg == 0)
695 return false;
696 UpdateValueMap(Call, ResultReg);
697 return true;
698 }
699 case Intrinsic::expect: {
700 unsigned ResultReg = getRegForValue(Call->getArgOperand(0));
701 if (ResultReg == 0)
702 return false;
703 UpdateValueMap(Call, ResultReg);
704 return true;
705 }
706 }
707
708 // Usually, it does not make sense to initialize a value,
709 // make an unrelated function call and use the value, because
710 // it tends to be spilled on the stack. So, we move the pointer
711 // to the last local value to the beginning of the block, so that
712 // all the values which have already been materialized,
713 // appear after the call. It also makes sense to skip intrinsics
714 // since they tend to be inlined.
715 if (!isa<IntrinsicInst>(Call))
716 flushLocalValueMap();
717
718 // An arbitrary call. Bail.
719 return false;
720 }
721
SelectCast(const User * I,unsigned Opcode)722 bool FastISel::SelectCast(const User *I, unsigned Opcode) {
723 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
724 EVT DstVT = TLI.getValueType(I->getType());
725
726 if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
727 DstVT == MVT::Other || !DstVT.isSimple())
728 // Unhandled type. Halt "fast" selection and bail.
729 return false;
730
731 // Check if the destination type is legal.
732 if (!TLI.isTypeLegal(DstVT))
733 return false;
734
735 // Check if the source operand is legal.
736 if (!TLI.isTypeLegal(SrcVT))
737 return false;
738
739 unsigned InputReg = getRegForValue(I->getOperand(0));
740 if (!InputReg)
741 // Unhandled operand. Halt "fast" selection and bail.
742 return false;
743
744 bool InputRegIsKill = hasTrivialKill(I->getOperand(0));
745
746 unsigned ResultReg = FastEmit_r(SrcVT.getSimpleVT(),
747 DstVT.getSimpleVT(),
748 Opcode,
749 InputReg, InputRegIsKill);
750 if (!ResultReg)
751 return false;
752
753 UpdateValueMap(I, ResultReg);
754 return true;
755 }
756
SelectBitCast(const User * I)757 bool FastISel::SelectBitCast(const User *I) {
758 // If the bitcast doesn't change the type, just use the operand value.
759 if (I->getType() == I->getOperand(0)->getType()) {
760 unsigned Reg = getRegForValue(I->getOperand(0));
761 if (Reg == 0)
762 return false;
763 UpdateValueMap(I, Reg);
764 return true;
765 }
766
767 // Bitcasts of other values become reg-reg copies or BITCAST operators.
768 EVT SrcEVT = TLI.getValueType(I->getOperand(0)->getType());
769 EVT DstEVT = TLI.getValueType(I->getType());
770 if (SrcEVT == MVT::Other || DstEVT == MVT::Other ||
771 !TLI.isTypeLegal(SrcEVT) || !TLI.isTypeLegal(DstEVT))
772 // Unhandled type. Halt "fast" selection and bail.
773 return false;
774
775 MVT SrcVT = SrcEVT.getSimpleVT();
776 MVT DstVT = DstEVT.getSimpleVT();
777 unsigned Op0 = getRegForValue(I->getOperand(0));
778 if (Op0 == 0)
779 // Unhandled operand. Halt "fast" selection and bail.
780 return false;
781
782 bool Op0IsKill = hasTrivialKill(I->getOperand(0));
783
784 // First, try to perform the bitcast by inserting a reg-reg copy.
785 unsigned ResultReg = 0;
786 if (SrcVT == DstVT) {
787 const TargetRegisterClass* SrcClass = TLI.getRegClassFor(SrcVT);
788 const TargetRegisterClass* DstClass = TLI.getRegClassFor(DstVT);
789 // Don't attempt a cross-class copy. It will likely fail.
790 if (SrcClass == DstClass) {
791 ResultReg = createResultReg(DstClass);
792 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
793 ResultReg).addReg(Op0);
794 }
795 }
796
797 // If the reg-reg copy failed, select a BITCAST opcode.
798 if (!ResultReg)
799 ResultReg = FastEmit_r(SrcVT, DstVT, ISD::BITCAST, Op0, Op0IsKill);
800
801 if (!ResultReg)
802 return false;
803
804 UpdateValueMap(I, ResultReg);
805 return true;
806 }
807
808 bool
SelectInstruction(const Instruction * I)809 FastISel::SelectInstruction(const Instruction *I) {
810 // Just before the terminator instruction, insert instructions to
811 // feed PHI nodes in successor blocks.
812 if (isa<TerminatorInst>(I))
813 if (!HandlePHINodesInSuccessorBlocks(I->getParent()))
814 return false;
815
816 DL = I->getDebugLoc();
817
818 MachineBasicBlock::iterator SavedInsertPt = FuncInfo.InsertPt;
819
820 // As a special case, don't handle calls to builtin library functions that
821 // may be translated directly to target instructions.
822 if (const CallInst *Call = dyn_cast<CallInst>(I)) {
823 const Function *F = Call->getCalledFunction();
824 LibFunc::Func Func;
825 if (F && !F->hasLocalLinkage() && F->hasName() &&
826 LibInfo->getLibFunc(F->getName(), Func) &&
827 LibInfo->hasOptimizedCodeGen(Func))
828 return false;
829 }
830
831 // First, try doing target-independent selection.
832 if (SelectOperator(I, I->getOpcode())) {
833 ++NumFastIselSuccessIndependent;
834 DL = DebugLoc();
835 return true;
836 }
837 // Remove dead code. However, ignore call instructions since we've flushed
838 // the local value map and recomputed the insert point.
839 if (!isa<CallInst>(I)) {
840 recomputeInsertPt();
841 if (SavedInsertPt != FuncInfo.InsertPt)
842 removeDeadCode(FuncInfo.InsertPt, SavedInsertPt);
843 }
844
845 // Next, try calling the target to attempt to handle the instruction.
846 SavedInsertPt = FuncInfo.InsertPt;
847 if (TargetSelectInstruction(I)) {
848 ++NumFastIselSuccessTarget;
849 DL = DebugLoc();
850 return true;
851 }
852 // Check for dead code and remove as necessary.
853 recomputeInsertPt();
854 if (SavedInsertPt != FuncInfo.InsertPt)
855 removeDeadCode(FuncInfo.InsertPt, SavedInsertPt);
856
857 DL = DebugLoc();
858 return false;
859 }
860
861 /// FastEmitBranch - Emit an unconditional branch to the given block,
862 /// unless it is the immediate (fall-through) successor, and update
863 /// the CFG.
864 void
FastEmitBranch(MachineBasicBlock * MSucc,DebugLoc DL)865 FastISel::FastEmitBranch(MachineBasicBlock *MSucc, DebugLoc DL) {
866
867 if (FuncInfo.MBB->getBasicBlock()->size() > 1 &&
868 FuncInfo.MBB->isLayoutSuccessor(MSucc)) {
869 // For more accurate line information if this is the only instruction
870 // in the block then emit it, otherwise we have the unconditional
871 // fall-through case, which needs no instructions.
872 } else {
873 // The unconditional branch case.
874 TII.InsertBranch(*FuncInfo.MBB, MSucc, NULL,
875 SmallVector<MachineOperand, 0>(), DL);
876 }
877 FuncInfo.MBB->addSuccessor(MSucc);
878 }
879
880 /// SelectFNeg - Emit an FNeg operation.
881 ///
882 bool
SelectFNeg(const User * I)883 FastISel::SelectFNeg(const User *I) {
884 unsigned OpReg = getRegForValue(BinaryOperator::getFNegArgument(I));
885 if (OpReg == 0) return false;
886
887 bool OpRegIsKill = hasTrivialKill(I);
888
889 // If the target has ISD::FNEG, use it.
890 EVT VT = TLI.getValueType(I->getType());
891 unsigned ResultReg = FastEmit_r(VT.getSimpleVT(), VT.getSimpleVT(),
892 ISD::FNEG, OpReg, OpRegIsKill);
893 if (ResultReg != 0) {
894 UpdateValueMap(I, ResultReg);
895 return true;
896 }
897
898 // Bitcast the value to integer, twiddle the sign bit with xor,
899 // and then bitcast it back to floating-point.
900 if (VT.getSizeInBits() > 64) return false;
901 EVT IntVT = EVT::getIntegerVT(I->getContext(), VT.getSizeInBits());
902 if (!TLI.isTypeLegal(IntVT))
903 return false;
904
905 unsigned IntReg = FastEmit_r(VT.getSimpleVT(), IntVT.getSimpleVT(),
906 ISD::BITCAST, OpReg, OpRegIsKill);
907 if (IntReg == 0)
908 return false;
909
910 unsigned IntResultReg = FastEmit_ri_(IntVT.getSimpleVT(), ISD::XOR,
911 IntReg, /*Kill=*/true,
912 UINT64_C(1) << (VT.getSizeInBits()-1),
913 IntVT.getSimpleVT());
914 if (IntResultReg == 0)
915 return false;
916
917 ResultReg = FastEmit_r(IntVT.getSimpleVT(), VT.getSimpleVT(),
918 ISD::BITCAST, IntResultReg, /*Kill=*/true);
919 if (ResultReg == 0)
920 return false;
921
922 UpdateValueMap(I, ResultReg);
923 return true;
924 }
925
926 bool
SelectExtractValue(const User * U)927 FastISel::SelectExtractValue(const User *U) {
928 const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(U);
929 if (!EVI)
930 return false;
931
932 // Make sure we only try to handle extracts with a legal result. But also
933 // allow i1 because it's easy.
934 EVT RealVT = TLI.getValueType(EVI->getType(), /*AllowUnknown=*/true);
935 if (!RealVT.isSimple())
936 return false;
937 MVT VT = RealVT.getSimpleVT();
938 if (!TLI.isTypeLegal(VT) && VT != MVT::i1)
939 return false;
940
941 const Value *Op0 = EVI->getOperand(0);
942 Type *AggTy = Op0->getType();
943
944 // Get the base result register.
945 unsigned ResultReg;
946 DenseMap<const Value *, unsigned>::iterator I = FuncInfo.ValueMap.find(Op0);
947 if (I != FuncInfo.ValueMap.end())
948 ResultReg = I->second;
949 else if (isa<Instruction>(Op0))
950 ResultReg = FuncInfo.InitializeRegForValue(Op0);
951 else
952 return false; // fast-isel can't handle aggregate constants at the moment
953
954 // Get the actual result register, which is an offset from the base register.
955 unsigned VTIndex = ComputeLinearIndex(AggTy, EVI->getIndices());
956
957 SmallVector<EVT, 4> AggValueVTs;
958 ComputeValueVTs(TLI, AggTy, AggValueVTs);
959
960 for (unsigned i = 0; i < VTIndex; i++)
961 ResultReg += TLI.getNumRegisters(FuncInfo.Fn->getContext(), AggValueVTs[i]);
962
963 UpdateValueMap(EVI, ResultReg);
964 return true;
965 }
966
967 bool
SelectOperator(const User * I,unsigned Opcode)968 FastISel::SelectOperator(const User *I, unsigned Opcode) {
969 switch (Opcode) {
970 case Instruction::Add:
971 return SelectBinaryOp(I, ISD::ADD);
972 case Instruction::FAdd:
973 return SelectBinaryOp(I, ISD::FADD);
974 case Instruction::Sub:
975 return SelectBinaryOp(I, ISD::SUB);
976 case Instruction::FSub:
977 // FNeg is currently represented in LLVM IR as a special case of FSub.
978 if (BinaryOperator::isFNeg(I))
979 return SelectFNeg(I);
980 return SelectBinaryOp(I, ISD::FSUB);
981 case Instruction::Mul:
982 return SelectBinaryOp(I, ISD::MUL);
983 case Instruction::FMul:
984 return SelectBinaryOp(I, ISD::FMUL);
985 case Instruction::SDiv:
986 return SelectBinaryOp(I, ISD::SDIV);
987 case Instruction::UDiv:
988 return SelectBinaryOp(I, ISD::UDIV);
989 case Instruction::FDiv:
990 return SelectBinaryOp(I, ISD::FDIV);
991 case Instruction::SRem:
992 return SelectBinaryOp(I, ISD::SREM);
993 case Instruction::URem:
994 return SelectBinaryOp(I, ISD::UREM);
995 case Instruction::FRem:
996 return SelectBinaryOp(I, ISD::FREM);
997 case Instruction::Shl:
998 return SelectBinaryOp(I, ISD::SHL);
999 case Instruction::LShr:
1000 return SelectBinaryOp(I, ISD::SRL);
1001 case Instruction::AShr:
1002 return SelectBinaryOp(I, ISD::SRA);
1003 case Instruction::And:
1004 return SelectBinaryOp(I, ISD::AND);
1005 case Instruction::Or:
1006 return SelectBinaryOp(I, ISD::OR);
1007 case Instruction::Xor:
1008 return SelectBinaryOp(I, ISD::XOR);
1009
1010 case Instruction::GetElementPtr:
1011 return SelectGetElementPtr(I);
1012
1013 case Instruction::Br: {
1014 const BranchInst *BI = cast<BranchInst>(I);
1015
1016 if (BI->isUnconditional()) {
1017 const BasicBlock *LLVMSucc = BI->getSuccessor(0);
1018 MachineBasicBlock *MSucc = FuncInfo.MBBMap[LLVMSucc];
1019 FastEmitBranch(MSucc, BI->getDebugLoc());
1020 return true;
1021 }
1022
1023 // Conditional branches are not handed yet.
1024 // Halt "fast" selection and bail.
1025 return false;
1026 }
1027
1028 case Instruction::Unreachable:
1029 // Nothing to emit.
1030 return true;
1031
1032 case Instruction::Alloca:
1033 // FunctionLowering has the static-sized case covered.
1034 if (FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(I)))
1035 return true;
1036
1037 // Dynamic-sized alloca is not handled yet.
1038 return false;
1039
1040 case Instruction::Call:
1041 return SelectCall(I);
1042
1043 case Instruction::BitCast:
1044 return SelectBitCast(I);
1045
1046 case Instruction::FPToSI:
1047 return SelectCast(I, ISD::FP_TO_SINT);
1048 case Instruction::ZExt:
1049 return SelectCast(I, ISD::ZERO_EXTEND);
1050 case Instruction::SExt:
1051 return SelectCast(I, ISD::SIGN_EXTEND);
1052 case Instruction::Trunc:
1053 return SelectCast(I, ISD::TRUNCATE);
1054 case Instruction::SIToFP:
1055 return SelectCast(I, ISD::SINT_TO_FP);
1056
1057 case Instruction::IntToPtr: // Deliberate fall-through.
1058 case Instruction::PtrToInt: {
1059 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
1060 EVT DstVT = TLI.getValueType(I->getType());
1061 if (DstVT.bitsGT(SrcVT))
1062 return SelectCast(I, ISD::ZERO_EXTEND);
1063 if (DstVT.bitsLT(SrcVT))
1064 return SelectCast(I, ISD::TRUNCATE);
1065 unsigned Reg = getRegForValue(I->getOperand(0));
1066 if (Reg == 0) return false;
1067 UpdateValueMap(I, Reg);
1068 return true;
1069 }
1070
1071 case Instruction::ExtractValue:
1072 return SelectExtractValue(I);
1073
1074 case Instruction::PHI:
1075 llvm_unreachable("FastISel shouldn't visit PHI nodes!");
1076
1077 default:
1078 // Unhandled instruction. Halt "fast" selection and bail.
1079 return false;
1080 }
1081 }
1082
FastISel(FunctionLoweringInfo & funcInfo,const TargetLibraryInfo * libInfo)1083 FastISel::FastISel(FunctionLoweringInfo &funcInfo,
1084 const TargetLibraryInfo *libInfo)
1085 : FuncInfo(funcInfo),
1086 MRI(FuncInfo.MF->getRegInfo()),
1087 MFI(*FuncInfo.MF->getFrameInfo()),
1088 MCP(*FuncInfo.MF->getConstantPool()),
1089 TM(FuncInfo.MF->getTarget()),
1090 TD(*TM.getDataLayout()),
1091 TII(*TM.getInstrInfo()),
1092 TLI(*TM.getTargetLowering()),
1093 TRI(*TM.getRegisterInfo()),
1094 LibInfo(libInfo) {
1095 }
1096
~FastISel()1097 FastISel::~FastISel() {}
1098
FastLowerArguments()1099 bool FastISel::FastLowerArguments() {
1100 return false;
1101 }
1102
FastEmit_(MVT,MVT,unsigned)1103 unsigned FastISel::FastEmit_(MVT, MVT,
1104 unsigned) {
1105 return 0;
1106 }
1107
FastEmit_r(MVT,MVT,unsigned,unsigned,bool)1108 unsigned FastISel::FastEmit_r(MVT, MVT,
1109 unsigned,
1110 unsigned /*Op0*/, bool /*Op0IsKill*/) {
1111 return 0;
1112 }
1113
FastEmit_rr(MVT,MVT,unsigned,unsigned,bool,unsigned,bool)1114 unsigned FastISel::FastEmit_rr(MVT, MVT,
1115 unsigned,
1116 unsigned /*Op0*/, bool /*Op0IsKill*/,
1117 unsigned /*Op1*/, bool /*Op1IsKill*/) {
1118 return 0;
1119 }
1120
FastEmit_i(MVT,MVT,unsigned,uint64_t)1121 unsigned FastISel::FastEmit_i(MVT, MVT, unsigned, uint64_t /*Imm*/) {
1122 return 0;
1123 }
1124
FastEmit_f(MVT,MVT,unsigned,const ConstantFP *)1125 unsigned FastISel::FastEmit_f(MVT, MVT,
1126 unsigned, const ConstantFP * /*FPImm*/) {
1127 return 0;
1128 }
1129
FastEmit_ri(MVT,MVT,unsigned,unsigned,bool,uint64_t)1130 unsigned FastISel::FastEmit_ri(MVT, MVT,
1131 unsigned,
1132 unsigned /*Op0*/, bool /*Op0IsKill*/,
1133 uint64_t /*Imm*/) {
1134 return 0;
1135 }
1136
FastEmit_rf(MVT,MVT,unsigned,unsigned,bool,const ConstantFP *)1137 unsigned FastISel::FastEmit_rf(MVT, MVT,
1138 unsigned,
1139 unsigned /*Op0*/, bool /*Op0IsKill*/,
1140 const ConstantFP * /*FPImm*/) {
1141 return 0;
1142 }
1143
FastEmit_rri(MVT,MVT,unsigned,unsigned,bool,unsigned,bool,uint64_t)1144 unsigned FastISel::FastEmit_rri(MVT, MVT,
1145 unsigned,
1146 unsigned /*Op0*/, bool /*Op0IsKill*/,
1147 unsigned /*Op1*/, bool /*Op1IsKill*/,
1148 uint64_t /*Imm*/) {
1149 return 0;
1150 }
1151
1152 /// FastEmit_ri_ - This method is a wrapper of FastEmit_ri. It first tries
1153 /// to emit an instruction with an immediate operand using FastEmit_ri.
1154 /// If that fails, it materializes the immediate into a register and try
1155 /// FastEmit_rr instead.
FastEmit_ri_(MVT VT,unsigned Opcode,unsigned Op0,bool Op0IsKill,uint64_t Imm,MVT ImmType)1156 unsigned FastISel::FastEmit_ri_(MVT VT, unsigned Opcode,
1157 unsigned Op0, bool Op0IsKill,
1158 uint64_t Imm, MVT ImmType) {
1159 // If this is a multiply by a power of two, emit this as a shift left.
1160 if (Opcode == ISD::MUL && isPowerOf2_64(Imm)) {
1161 Opcode = ISD::SHL;
1162 Imm = Log2_64(Imm);
1163 } else if (Opcode == ISD::UDIV && isPowerOf2_64(Imm)) {
1164 // div x, 8 -> srl x, 3
1165 Opcode = ISD::SRL;
1166 Imm = Log2_64(Imm);
1167 }
1168
1169 // Horrible hack (to be removed), check to make sure shift amounts are
1170 // in-range.
1171 if ((Opcode == ISD::SHL || Opcode == ISD::SRA || Opcode == ISD::SRL) &&
1172 Imm >= VT.getSizeInBits())
1173 return 0;
1174
1175 // First check if immediate type is legal. If not, we can't use the ri form.
1176 unsigned ResultReg = FastEmit_ri(VT, VT, Opcode, Op0, Op0IsKill, Imm);
1177 if (ResultReg != 0)
1178 return ResultReg;
1179 unsigned MaterialReg = FastEmit_i(ImmType, ImmType, ISD::Constant, Imm);
1180 if (MaterialReg == 0) {
1181 // This is a bit ugly/slow, but failing here means falling out of
1182 // fast-isel, which would be very slow.
1183 IntegerType *ITy = IntegerType::get(FuncInfo.Fn->getContext(),
1184 VT.getSizeInBits());
1185 MaterialReg = getRegForValue(ConstantInt::get(ITy, Imm));
1186 }
1187 return FastEmit_rr(VT, VT, Opcode,
1188 Op0, Op0IsKill,
1189 MaterialReg, /*Kill=*/true);
1190 }
1191
createResultReg(const TargetRegisterClass * RC)1192 unsigned FastISel::createResultReg(const TargetRegisterClass* RC) {
1193 return MRI.createVirtualRegister(RC);
1194 }
1195
FastEmitInst_(unsigned MachineInstOpcode,const TargetRegisterClass * RC)1196 unsigned FastISel::FastEmitInst_(unsigned MachineInstOpcode,
1197 const TargetRegisterClass* RC) {
1198 unsigned ResultReg = createResultReg(RC);
1199 const MCInstrDesc &II = TII.get(MachineInstOpcode);
1200
1201 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg);
1202 return ResultReg;
1203 }
1204
FastEmitInst_r(unsigned MachineInstOpcode,const TargetRegisterClass * RC,unsigned Op0,bool Op0IsKill)1205 unsigned FastISel::FastEmitInst_r(unsigned MachineInstOpcode,
1206 const TargetRegisterClass *RC,
1207 unsigned Op0, bool Op0IsKill) {
1208 unsigned ResultReg = createResultReg(RC);
1209 const MCInstrDesc &II = TII.get(MachineInstOpcode);
1210
1211 if (II.getNumDefs() >= 1)
1212 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1213 .addReg(Op0, Op0IsKill * RegState::Kill);
1214 else {
1215 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1216 .addReg(Op0, Op0IsKill * RegState::Kill);
1217 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1218 ResultReg).addReg(II.ImplicitDefs[0]);
1219 }
1220
1221 return ResultReg;
1222 }
1223
FastEmitInst_rr(unsigned MachineInstOpcode,const TargetRegisterClass * RC,unsigned Op0,bool Op0IsKill,unsigned Op1,bool Op1IsKill)1224 unsigned FastISel::FastEmitInst_rr(unsigned MachineInstOpcode,
1225 const TargetRegisterClass *RC,
1226 unsigned Op0, bool Op0IsKill,
1227 unsigned Op1, bool Op1IsKill) {
1228 unsigned ResultReg = createResultReg(RC);
1229 const MCInstrDesc &II = TII.get(MachineInstOpcode);
1230
1231 if (II.getNumDefs() >= 1)
1232 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1233 .addReg(Op0, Op0IsKill * RegState::Kill)
1234 .addReg(Op1, Op1IsKill * RegState::Kill);
1235 else {
1236 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1237 .addReg(Op0, Op0IsKill * RegState::Kill)
1238 .addReg(Op1, Op1IsKill * RegState::Kill);
1239 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1240 ResultReg).addReg(II.ImplicitDefs[0]);
1241 }
1242 return ResultReg;
1243 }
1244
FastEmitInst_rrr(unsigned MachineInstOpcode,const TargetRegisterClass * RC,unsigned Op0,bool Op0IsKill,unsigned Op1,bool Op1IsKill,unsigned Op2,bool Op2IsKill)1245 unsigned FastISel::FastEmitInst_rrr(unsigned MachineInstOpcode,
1246 const TargetRegisterClass *RC,
1247 unsigned Op0, bool Op0IsKill,
1248 unsigned Op1, bool Op1IsKill,
1249 unsigned Op2, bool Op2IsKill) {
1250 unsigned ResultReg = createResultReg(RC);
1251 const MCInstrDesc &II = TII.get(MachineInstOpcode);
1252
1253 if (II.getNumDefs() >= 1)
1254 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1255 .addReg(Op0, Op0IsKill * RegState::Kill)
1256 .addReg(Op1, Op1IsKill * RegState::Kill)
1257 .addReg(Op2, Op2IsKill * RegState::Kill);
1258 else {
1259 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1260 .addReg(Op0, Op0IsKill * RegState::Kill)
1261 .addReg(Op1, Op1IsKill * RegState::Kill)
1262 .addReg(Op2, Op2IsKill * RegState::Kill);
1263 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1264 ResultReg).addReg(II.ImplicitDefs[0]);
1265 }
1266 return ResultReg;
1267 }
1268
FastEmitInst_ri(unsigned MachineInstOpcode,const TargetRegisterClass * RC,unsigned Op0,bool Op0IsKill,uint64_t Imm)1269 unsigned FastISel::FastEmitInst_ri(unsigned MachineInstOpcode,
1270 const TargetRegisterClass *RC,
1271 unsigned Op0, bool Op0IsKill,
1272 uint64_t Imm) {
1273 unsigned ResultReg = createResultReg(RC);
1274 const MCInstrDesc &II = TII.get(MachineInstOpcode);
1275
1276 if (II.getNumDefs() >= 1)
1277 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1278 .addReg(Op0, Op0IsKill * RegState::Kill)
1279 .addImm(Imm);
1280 else {
1281 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1282 .addReg(Op0, Op0IsKill * RegState::Kill)
1283 .addImm(Imm);
1284 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1285 ResultReg).addReg(II.ImplicitDefs[0]);
1286 }
1287 return ResultReg;
1288 }
1289
FastEmitInst_rii(unsigned MachineInstOpcode,const TargetRegisterClass * RC,unsigned Op0,bool Op0IsKill,uint64_t Imm1,uint64_t Imm2)1290 unsigned FastISel::FastEmitInst_rii(unsigned MachineInstOpcode,
1291 const TargetRegisterClass *RC,
1292 unsigned Op0, bool Op0IsKill,
1293 uint64_t Imm1, uint64_t Imm2) {
1294 unsigned ResultReg = createResultReg(RC);
1295 const MCInstrDesc &II = TII.get(MachineInstOpcode);
1296
1297 if (II.getNumDefs() >= 1)
1298 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1299 .addReg(Op0, Op0IsKill * RegState::Kill)
1300 .addImm(Imm1)
1301 .addImm(Imm2);
1302 else {
1303 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1304 .addReg(Op0, Op0IsKill * RegState::Kill)
1305 .addImm(Imm1)
1306 .addImm(Imm2);
1307 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1308 ResultReg).addReg(II.ImplicitDefs[0]);
1309 }
1310 return ResultReg;
1311 }
1312
FastEmitInst_rf(unsigned MachineInstOpcode,const TargetRegisterClass * RC,unsigned Op0,bool Op0IsKill,const ConstantFP * FPImm)1313 unsigned FastISel::FastEmitInst_rf(unsigned MachineInstOpcode,
1314 const TargetRegisterClass *RC,
1315 unsigned Op0, bool Op0IsKill,
1316 const ConstantFP *FPImm) {
1317 unsigned ResultReg = createResultReg(RC);
1318 const MCInstrDesc &II = TII.get(MachineInstOpcode);
1319
1320 if (II.getNumDefs() >= 1)
1321 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1322 .addReg(Op0, Op0IsKill * RegState::Kill)
1323 .addFPImm(FPImm);
1324 else {
1325 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1326 .addReg(Op0, Op0IsKill * RegState::Kill)
1327 .addFPImm(FPImm);
1328 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1329 ResultReg).addReg(II.ImplicitDefs[0]);
1330 }
1331 return ResultReg;
1332 }
1333
FastEmitInst_rri(unsigned MachineInstOpcode,const TargetRegisterClass * RC,unsigned Op0,bool Op0IsKill,unsigned Op1,bool Op1IsKill,uint64_t Imm)1334 unsigned FastISel::FastEmitInst_rri(unsigned MachineInstOpcode,
1335 const TargetRegisterClass *RC,
1336 unsigned Op0, bool Op0IsKill,
1337 unsigned Op1, bool Op1IsKill,
1338 uint64_t Imm) {
1339 unsigned ResultReg = createResultReg(RC);
1340 const MCInstrDesc &II = TII.get(MachineInstOpcode);
1341
1342 if (II.getNumDefs() >= 1)
1343 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1344 .addReg(Op0, Op0IsKill * RegState::Kill)
1345 .addReg(Op1, Op1IsKill * RegState::Kill)
1346 .addImm(Imm);
1347 else {
1348 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1349 .addReg(Op0, Op0IsKill * RegState::Kill)
1350 .addReg(Op1, Op1IsKill * RegState::Kill)
1351 .addImm(Imm);
1352 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1353 ResultReg).addReg(II.ImplicitDefs[0]);
1354 }
1355 return ResultReg;
1356 }
1357
FastEmitInst_rrii(unsigned MachineInstOpcode,const TargetRegisterClass * RC,unsigned Op0,bool Op0IsKill,unsigned Op1,bool Op1IsKill,uint64_t Imm1,uint64_t Imm2)1358 unsigned FastISel::FastEmitInst_rrii(unsigned MachineInstOpcode,
1359 const TargetRegisterClass *RC,
1360 unsigned Op0, bool Op0IsKill,
1361 unsigned Op1, bool Op1IsKill,
1362 uint64_t Imm1, uint64_t Imm2) {
1363 unsigned ResultReg = createResultReg(RC);
1364 const MCInstrDesc &II = TII.get(MachineInstOpcode);
1365
1366 if (II.getNumDefs() >= 1)
1367 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1368 .addReg(Op0, Op0IsKill * RegState::Kill)
1369 .addReg(Op1, Op1IsKill * RegState::Kill)
1370 .addImm(Imm1).addImm(Imm2);
1371 else {
1372 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1373 .addReg(Op0, Op0IsKill * RegState::Kill)
1374 .addReg(Op1, Op1IsKill * RegState::Kill)
1375 .addImm(Imm1).addImm(Imm2);
1376 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1377 ResultReg).addReg(II.ImplicitDefs[0]);
1378 }
1379 return ResultReg;
1380 }
1381
FastEmitInst_i(unsigned MachineInstOpcode,const TargetRegisterClass * RC,uint64_t Imm)1382 unsigned FastISel::FastEmitInst_i(unsigned MachineInstOpcode,
1383 const TargetRegisterClass *RC,
1384 uint64_t Imm) {
1385 unsigned ResultReg = createResultReg(RC);
1386 const MCInstrDesc &II = TII.get(MachineInstOpcode);
1387
1388 if (II.getNumDefs() >= 1)
1389 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg).addImm(Imm);
1390 else {
1391 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II).addImm(Imm);
1392 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1393 ResultReg).addReg(II.ImplicitDefs[0]);
1394 }
1395 return ResultReg;
1396 }
1397
FastEmitInst_ii(unsigned MachineInstOpcode,const TargetRegisterClass * RC,uint64_t Imm1,uint64_t Imm2)1398 unsigned FastISel::FastEmitInst_ii(unsigned MachineInstOpcode,
1399 const TargetRegisterClass *RC,
1400 uint64_t Imm1, uint64_t Imm2) {
1401 unsigned ResultReg = createResultReg(RC);
1402 const MCInstrDesc &II = TII.get(MachineInstOpcode);
1403
1404 if (II.getNumDefs() >= 1)
1405 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1406 .addImm(Imm1).addImm(Imm2);
1407 else {
1408 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II).addImm(Imm1).addImm(Imm2);
1409 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1410 ResultReg).addReg(II.ImplicitDefs[0]);
1411 }
1412 return ResultReg;
1413 }
1414
FastEmitInst_extractsubreg(MVT RetVT,unsigned Op0,bool Op0IsKill,uint32_t Idx)1415 unsigned FastISel::FastEmitInst_extractsubreg(MVT RetVT,
1416 unsigned Op0, bool Op0IsKill,
1417 uint32_t Idx) {
1418 unsigned ResultReg = createResultReg(TLI.getRegClassFor(RetVT));
1419 assert(TargetRegisterInfo::isVirtualRegister(Op0) &&
1420 "Cannot yet extract from physregs");
1421 const TargetRegisterClass *RC = MRI.getRegClass(Op0);
1422 MRI.constrainRegClass(Op0, TRI.getSubClassWithSubReg(RC, Idx));
1423 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
1424 DL, TII.get(TargetOpcode::COPY), ResultReg)
1425 .addReg(Op0, getKillRegState(Op0IsKill), Idx);
1426 return ResultReg;
1427 }
1428
1429 /// FastEmitZExtFromI1 - Emit MachineInstrs to compute the value of Op
1430 /// with all but the least significant bit set to zero.
FastEmitZExtFromI1(MVT VT,unsigned Op0,bool Op0IsKill)1431 unsigned FastISel::FastEmitZExtFromI1(MVT VT, unsigned Op0, bool Op0IsKill) {
1432 return FastEmit_ri(VT, VT, ISD::AND, Op0, Op0IsKill, 1);
1433 }
1434
1435 /// HandlePHINodesInSuccessorBlocks - Handle PHI nodes in successor blocks.
1436 /// Emit code to ensure constants are copied into registers when needed.
1437 /// Remember the virtual registers that need to be added to the Machine PHI
1438 /// nodes as input. We cannot just directly add them, because expansion
1439 /// might result in multiple MBB's for one BB. As such, the start of the
1440 /// BB might correspond to a different MBB than the end.
HandlePHINodesInSuccessorBlocks(const BasicBlock * LLVMBB)1441 bool FastISel::HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB) {
1442 const TerminatorInst *TI = LLVMBB->getTerminator();
1443
1444 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled;
1445 unsigned OrigNumPHINodesToUpdate = FuncInfo.PHINodesToUpdate.size();
1446
1447 // Check successor nodes' PHI nodes that expect a constant to be available
1448 // from this block.
1449 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
1450 const BasicBlock *SuccBB = TI->getSuccessor(succ);
1451 if (!isa<PHINode>(SuccBB->begin())) continue;
1452 MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB];
1453
1454 // If this terminator has multiple identical successors (common for
1455 // switches), only handle each succ once.
1456 if (!SuccsHandled.insert(SuccMBB)) continue;
1457
1458 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
1459
1460 // At this point we know that there is a 1-1 correspondence between LLVM PHI
1461 // nodes and Machine PHI nodes, but the incoming operands have not been
1462 // emitted yet.
1463 for (BasicBlock::const_iterator I = SuccBB->begin();
1464 const PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1465
1466 // Ignore dead phi's.
1467 if (PN->use_empty()) continue;
1468
1469 // Only handle legal types. Two interesting things to note here. First,
1470 // by bailing out early, we may leave behind some dead instructions,
1471 // since SelectionDAG's HandlePHINodesInSuccessorBlocks will insert its
1472 // own moves. Second, this check is necessary because FastISel doesn't
1473 // use CreateRegs to create registers, so it always creates
1474 // exactly one register for each non-void instruction.
1475 EVT VT = TLI.getValueType(PN->getType(), /*AllowUnknown=*/true);
1476 if (VT == MVT::Other || !TLI.isTypeLegal(VT)) {
1477 // Handle integer promotions, though, because they're common and easy.
1478 if (VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16)
1479 VT = TLI.getTypeToTransformTo(LLVMBB->getContext(), VT);
1480 else {
1481 FuncInfo.PHINodesToUpdate.resize(OrigNumPHINodesToUpdate);
1482 return false;
1483 }
1484 }
1485
1486 const Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
1487
1488 // Set the DebugLoc for the copy. Prefer the location of the operand
1489 // if there is one; use the location of the PHI otherwise.
1490 DL = PN->getDebugLoc();
1491 if (const Instruction *Inst = dyn_cast<Instruction>(PHIOp))
1492 DL = Inst->getDebugLoc();
1493
1494 unsigned Reg = getRegForValue(PHIOp);
1495 if (Reg == 0) {
1496 FuncInfo.PHINodesToUpdate.resize(OrigNumPHINodesToUpdate);
1497 return false;
1498 }
1499 FuncInfo.PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg));
1500 DL = DebugLoc();
1501 }
1502 }
1503
1504 return true;
1505 }
1506