1 //===- X86ISelDAGToDAG.cpp - A DAG pattern matching inst selector for X86 -===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file defines a DAG pattern matching instruction selector for X86,
11 // converting from a legalized dag to a X86 dag.
12 //
13 //===----------------------------------------------------------------------===//
14
15 #include "X86.h"
16 #include "X86InstrBuilder.h"
17 #include "X86MachineFunctionInfo.h"
18 #include "X86RegisterInfo.h"
19 #include "X86Subtarget.h"
20 #include "X86TargetMachine.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/CodeGen/MachineFrameInfo.h"
23 #include "llvm/CodeGen/MachineFunction.h"
24 #include "llvm/CodeGen/MachineInstrBuilder.h"
25 #include "llvm/CodeGen/MachineRegisterInfo.h"
26 #include "llvm/CodeGen/SelectionDAGISel.h"
27 #include "llvm/IR/Instructions.h"
28 #include "llvm/IR/Intrinsics.h"
29 #include "llvm/IR/Type.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/ErrorHandling.h"
32 #include "llvm/Support/MathExtras.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/Target/TargetMachine.h"
35 #include "llvm/Target/TargetOptions.h"
36 using namespace llvm;
37
38 #define DEBUG_TYPE "x86-isel"
39
40 STATISTIC(NumLoadMoved, "Number of loads moved below TokenFactor");
41
42 //===----------------------------------------------------------------------===//
43 // Pattern Matcher Implementation
44 //===----------------------------------------------------------------------===//
45
46 namespace {
47 /// X86ISelAddressMode - This corresponds to X86AddressMode, but uses
48 /// SDValue's instead of register numbers for the leaves of the matched
49 /// tree.
50 struct X86ISelAddressMode {
51 enum {
52 RegBase,
53 FrameIndexBase
54 } BaseType;
55
56 // This is really a union, discriminated by BaseType!
57 SDValue Base_Reg;
58 int Base_FrameIndex;
59
60 unsigned Scale;
61 SDValue IndexReg;
62 int32_t Disp;
63 SDValue Segment;
64 const GlobalValue *GV;
65 const Constant *CP;
66 const BlockAddress *BlockAddr;
67 const char *ES;
68 int JT;
69 unsigned Align; // CP alignment.
70 unsigned char SymbolFlags; // X86II::MO_*
71
X86ISelAddressMode__anon4d463def0111::X86ISelAddressMode72 X86ISelAddressMode()
73 : BaseType(RegBase), Base_FrameIndex(0), Scale(1), IndexReg(), Disp(0),
74 Segment(), GV(nullptr), CP(nullptr), BlockAddr(nullptr), ES(nullptr),
75 JT(-1), Align(0), SymbolFlags(X86II::MO_NO_FLAG) {
76 }
77
hasSymbolicDisplacement__anon4d463def0111::X86ISelAddressMode78 bool hasSymbolicDisplacement() const {
79 return GV != nullptr || CP != nullptr || ES != nullptr ||
80 JT != -1 || BlockAddr != nullptr;
81 }
82
hasBaseOrIndexReg__anon4d463def0111::X86ISelAddressMode83 bool hasBaseOrIndexReg() const {
84 return BaseType == FrameIndexBase ||
85 IndexReg.getNode() != nullptr || Base_Reg.getNode() != nullptr;
86 }
87
88 /// isRIPRelative - Return true if this addressing mode is already RIP
89 /// relative.
isRIPRelative__anon4d463def0111::X86ISelAddressMode90 bool isRIPRelative() const {
91 if (BaseType != RegBase) return false;
92 if (RegisterSDNode *RegNode =
93 dyn_cast_or_null<RegisterSDNode>(Base_Reg.getNode()))
94 return RegNode->getReg() == X86::RIP;
95 return false;
96 }
97
setBaseReg__anon4d463def0111::X86ISelAddressMode98 void setBaseReg(SDValue Reg) {
99 BaseType = RegBase;
100 Base_Reg = Reg;
101 }
102
103 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
dump__anon4d463def0111::X86ISelAddressMode104 void dump() {
105 dbgs() << "X86ISelAddressMode " << this << '\n';
106 dbgs() << "Base_Reg ";
107 if (Base_Reg.getNode())
108 Base_Reg.getNode()->dump();
109 else
110 dbgs() << "nul";
111 dbgs() << " Base.FrameIndex " << Base_FrameIndex << '\n'
112 << " Scale" << Scale << '\n'
113 << "IndexReg ";
114 if (IndexReg.getNode())
115 IndexReg.getNode()->dump();
116 else
117 dbgs() << "nul";
118 dbgs() << " Disp " << Disp << '\n'
119 << "GV ";
120 if (GV)
121 GV->dump();
122 else
123 dbgs() << "nul";
124 dbgs() << " CP ";
125 if (CP)
126 CP->dump();
127 else
128 dbgs() << "nul";
129 dbgs() << '\n'
130 << "ES ";
131 if (ES)
132 dbgs() << ES;
133 else
134 dbgs() << "nul";
135 dbgs() << " JT" << JT << " Align" << Align << '\n';
136 }
137 #endif
138 };
139 }
140
141 namespace {
142 //===--------------------------------------------------------------------===//
143 /// ISel - X86 specific code to select X86 machine instructions for
144 /// SelectionDAG operations.
145 ///
146 class X86DAGToDAGISel final : public SelectionDAGISel {
147 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
148 /// make the right decision when generating code for different targets.
149 const X86Subtarget *Subtarget;
150
151 /// OptForSize - If true, selector should try to optimize for code size
152 /// instead of performance.
153 bool OptForSize;
154
155 public:
X86DAGToDAGISel(X86TargetMachine & tm,CodeGenOpt::Level OptLevel)156 explicit X86DAGToDAGISel(X86TargetMachine &tm, CodeGenOpt::Level OptLevel)
157 : SelectionDAGISel(tm, OptLevel),
158 Subtarget(&tm.getSubtarget<X86Subtarget>()),
159 OptForSize(false) {}
160
getPassName() const161 const char *getPassName() const override {
162 return "X86 DAG->DAG Instruction Selection";
163 }
164
runOnMachineFunction(MachineFunction & MF)165 bool runOnMachineFunction(MachineFunction &MF) override {
166 // Reset the subtarget each time through.
167 Subtarget = &TM.getSubtarget<X86Subtarget>();
168 SelectionDAGISel::runOnMachineFunction(MF);
169 return true;
170 }
171
172 void EmitFunctionEntryCode() override;
173
174 bool IsProfitableToFold(SDValue N, SDNode *U, SDNode *Root) const override;
175
176 void PreprocessISelDAG() override;
177
immSext8(SDNode * N) const178 inline bool immSext8(SDNode *N) const {
179 return isInt<8>(cast<ConstantSDNode>(N)->getSExtValue());
180 }
181
182 // i64immSExt32 predicate - True if the 64-bit immediate fits in a 32-bit
183 // sign extended field.
i64immSExt32(SDNode * N) const184 inline bool i64immSExt32(SDNode *N) const {
185 uint64_t v = cast<ConstantSDNode>(N)->getZExtValue();
186 return (int64_t)v == (int32_t)v;
187 }
188
189 // Include the pieces autogenerated from the target description.
190 #include "X86GenDAGISel.inc"
191
192 private:
193 SDNode *Select(SDNode *N) override;
194 SDNode *SelectGather(SDNode *N, unsigned Opc);
195 SDNode *SelectAtomic64(SDNode *Node, unsigned Opc);
196 SDNode *SelectAtomicLoadArith(SDNode *Node, MVT NVT);
197
198 bool FoldOffsetIntoAddress(uint64_t Offset, X86ISelAddressMode &AM);
199 bool MatchLoadInAddress(LoadSDNode *N, X86ISelAddressMode &AM);
200 bool MatchWrapper(SDValue N, X86ISelAddressMode &AM);
201 bool MatchAddress(SDValue N, X86ISelAddressMode &AM);
202 bool MatchAddressRecursively(SDValue N, X86ISelAddressMode &AM,
203 unsigned Depth);
204 bool MatchAddressBase(SDValue N, X86ISelAddressMode &AM);
205 bool SelectAddr(SDNode *Parent, SDValue N, SDValue &Base,
206 SDValue &Scale, SDValue &Index, SDValue &Disp,
207 SDValue &Segment);
208 bool SelectMOV64Imm32(SDValue N, SDValue &Imm);
209 bool SelectLEAAddr(SDValue N, SDValue &Base,
210 SDValue &Scale, SDValue &Index, SDValue &Disp,
211 SDValue &Segment);
212 bool SelectLEA64_32Addr(SDValue N, SDValue &Base,
213 SDValue &Scale, SDValue &Index, SDValue &Disp,
214 SDValue &Segment);
215 bool SelectTLSADDRAddr(SDValue N, SDValue &Base,
216 SDValue &Scale, SDValue &Index, SDValue &Disp,
217 SDValue &Segment);
218 bool SelectScalarSSELoad(SDNode *Root, SDValue N,
219 SDValue &Base, SDValue &Scale,
220 SDValue &Index, SDValue &Disp,
221 SDValue &Segment,
222 SDValue &NodeWithChain);
223
224 bool TryFoldLoad(SDNode *P, SDValue N,
225 SDValue &Base, SDValue &Scale,
226 SDValue &Index, SDValue &Disp,
227 SDValue &Segment);
228
229 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
230 /// inline asm expressions.
231 bool SelectInlineAsmMemoryOperand(const SDValue &Op,
232 char ConstraintCode,
233 std::vector<SDValue> &OutOps) override;
234
235 void EmitSpecialCodeForMain(MachineBasicBlock *BB, MachineFrameInfo *MFI);
236
getAddressOperands(X86ISelAddressMode & AM,SDValue & Base,SDValue & Scale,SDValue & Index,SDValue & Disp,SDValue & Segment)237 inline void getAddressOperands(X86ISelAddressMode &AM, SDValue &Base,
238 SDValue &Scale, SDValue &Index,
239 SDValue &Disp, SDValue &Segment) {
240 Base = (AM.BaseType == X86ISelAddressMode::FrameIndexBase) ?
241 CurDAG->getTargetFrameIndex(AM.Base_FrameIndex,
242 getTargetLowering()->getPointerTy()) :
243 AM.Base_Reg;
244 Scale = getI8Imm(AM.Scale);
245 Index = AM.IndexReg;
246 // These are 32-bit even in 64-bit mode since RIP relative offset
247 // is 32-bit.
248 if (AM.GV)
249 Disp = CurDAG->getTargetGlobalAddress(AM.GV, SDLoc(),
250 MVT::i32, AM.Disp,
251 AM.SymbolFlags);
252 else if (AM.CP)
253 Disp = CurDAG->getTargetConstantPool(AM.CP, MVT::i32,
254 AM.Align, AM.Disp, AM.SymbolFlags);
255 else if (AM.ES) {
256 assert(!AM.Disp && "Non-zero displacement is ignored with ES.");
257 Disp = CurDAG->getTargetExternalSymbol(AM.ES, MVT::i32, AM.SymbolFlags);
258 } else if (AM.JT != -1) {
259 assert(!AM.Disp && "Non-zero displacement is ignored with JT.");
260 Disp = CurDAG->getTargetJumpTable(AM.JT, MVT::i32, AM.SymbolFlags);
261 } else if (AM.BlockAddr)
262 Disp = CurDAG->getTargetBlockAddress(AM.BlockAddr, MVT::i32, AM.Disp,
263 AM.SymbolFlags);
264 else
265 Disp = CurDAG->getTargetConstant(AM.Disp, MVT::i32);
266
267 if (AM.Segment.getNode())
268 Segment = AM.Segment;
269 else
270 Segment = CurDAG->getRegister(0, MVT::i32);
271 }
272
273 /// getI8Imm - Return a target constant with the specified value, of type
274 /// i8.
getI8Imm(unsigned Imm)275 inline SDValue getI8Imm(unsigned Imm) {
276 return CurDAG->getTargetConstant(Imm, MVT::i8);
277 }
278
279 /// getI32Imm - Return a target constant with the specified value, of type
280 /// i32.
getI32Imm(unsigned Imm)281 inline SDValue getI32Imm(unsigned Imm) {
282 return CurDAG->getTargetConstant(Imm, MVT::i32);
283 }
284
285 /// getGlobalBaseReg - Return an SDNode that returns the value of
286 /// the global base register. Output instructions required to
287 /// initialize the global base register, if necessary.
288 ///
289 SDNode *getGlobalBaseReg();
290
291 /// getTargetMachine - Return a reference to the TargetMachine, casted
292 /// to the target-specific type.
getTargetMachine() const293 const X86TargetMachine &getTargetMachine() const {
294 return static_cast<const X86TargetMachine &>(TM);
295 }
296
297 /// getInstrInfo - Return a reference to the TargetInstrInfo, casted
298 /// to the target-specific type.
getInstrInfo() const299 const X86InstrInfo *getInstrInfo() const {
300 return getTargetMachine().getInstrInfo();
301 }
302 };
303 }
304
305
306 bool
IsProfitableToFold(SDValue N,SDNode * U,SDNode * Root) const307 X86DAGToDAGISel::IsProfitableToFold(SDValue N, SDNode *U, SDNode *Root) const {
308 if (OptLevel == CodeGenOpt::None) return false;
309
310 if (!N.hasOneUse())
311 return false;
312
313 if (N.getOpcode() != ISD::LOAD)
314 return true;
315
316 // If N is a load, do additional profitability checks.
317 if (U == Root) {
318 switch (U->getOpcode()) {
319 default: break;
320 case X86ISD::ADD:
321 case X86ISD::SUB:
322 case X86ISD::AND:
323 case X86ISD::XOR:
324 case X86ISD::OR:
325 case ISD::ADD:
326 case ISD::ADDC:
327 case ISD::ADDE:
328 case ISD::AND:
329 case ISD::OR:
330 case ISD::XOR: {
331 SDValue Op1 = U->getOperand(1);
332
333 // If the other operand is a 8-bit immediate we should fold the immediate
334 // instead. This reduces code size.
335 // e.g.
336 // movl 4(%esp), %eax
337 // addl $4, %eax
338 // vs.
339 // movl $4, %eax
340 // addl 4(%esp), %eax
341 // The former is 2 bytes shorter. In case where the increment is 1, then
342 // the saving can be 4 bytes (by using incl %eax).
343 if (ConstantSDNode *Imm = dyn_cast<ConstantSDNode>(Op1))
344 if (Imm->getAPIntValue().isSignedIntN(8))
345 return false;
346
347 // If the other operand is a TLS address, we should fold it instead.
348 // This produces
349 // movl %gs:0, %eax
350 // leal i@NTPOFF(%eax), %eax
351 // instead of
352 // movl $i@NTPOFF, %eax
353 // addl %gs:0, %eax
354 // if the block also has an access to a second TLS address this will save
355 // a load.
356 // FIXME: This is probably also true for non-TLS addresses.
357 if (Op1.getOpcode() == X86ISD::Wrapper) {
358 SDValue Val = Op1.getOperand(0);
359 if (Val.getOpcode() == ISD::TargetGlobalTLSAddress)
360 return false;
361 }
362 }
363 }
364 }
365
366 return true;
367 }
368
369 /// MoveBelowCallOrigChain - Replace the original chain operand of the call with
370 /// load's chain operand and move load below the call's chain operand.
MoveBelowOrigChain(SelectionDAG * CurDAG,SDValue Load,SDValue Call,SDValue OrigChain)371 static void MoveBelowOrigChain(SelectionDAG *CurDAG, SDValue Load,
372 SDValue Call, SDValue OrigChain) {
373 SmallVector<SDValue, 8> Ops;
374 SDValue Chain = OrigChain.getOperand(0);
375 if (Chain.getNode() == Load.getNode())
376 Ops.push_back(Load.getOperand(0));
377 else {
378 assert(Chain.getOpcode() == ISD::TokenFactor &&
379 "Unexpected chain operand");
380 for (unsigned i = 0, e = Chain.getNumOperands(); i != e; ++i)
381 if (Chain.getOperand(i).getNode() == Load.getNode())
382 Ops.push_back(Load.getOperand(0));
383 else
384 Ops.push_back(Chain.getOperand(i));
385 SDValue NewChain =
386 CurDAG->getNode(ISD::TokenFactor, SDLoc(Load), MVT::Other, Ops);
387 Ops.clear();
388 Ops.push_back(NewChain);
389 }
390 for (unsigned i = 1, e = OrigChain.getNumOperands(); i != e; ++i)
391 Ops.push_back(OrigChain.getOperand(i));
392 CurDAG->UpdateNodeOperands(OrigChain.getNode(), Ops);
393 CurDAG->UpdateNodeOperands(Load.getNode(), Call.getOperand(0),
394 Load.getOperand(1), Load.getOperand(2));
395
396 unsigned NumOps = Call.getNode()->getNumOperands();
397 Ops.clear();
398 Ops.push_back(SDValue(Load.getNode(), 1));
399 for (unsigned i = 1, e = NumOps; i != e; ++i)
400 Ops.push_back(Call.getOperand(i));
401 CurDAG->UpdateNodeOperands(Call.getNode(), Ops);
402 }
403
404 /// isCalleeLoad - Return true if call address is a load and it can be
405 /// moved below CALLSEQ_START and the chains leading up to the call.
406 /// Return the CALLSEQ_START by reference as a second output.
407 /// In the case of a tail call, there isn't a callseq node between the call
408 /// chain and the load.
isCalleeLoad(SDValue Callee,SDValue & Chain,bool HasCallSeq)409 static bool isCalleeLoad(SDValue Callee, SDValue &Chain, bool HasCallSeq) {
410 // The transformation is somewhat dangerous if the call's chain was glued to
411 // the call. After MoveBelowOrigChain the load is moved between the call and
412 // the chain, this can create a cycle if the load is not folded. So it is
413 // *really* important that we are sure the load will be folded.
414 if (Callee.getNode() == Chain.getNode() || !Callee.hasOneUse())
415 return false;
416 LoadSDNode *LD = dyn_cast<LoadSDNode>(Callee.getNode());
417 if (!LD ||
418 LD->isVolatile() ||
419 LD->getAddressingMode() != ISD::UNINDEXED ||
420 LD->getExtensionType() != ISD::NON_EXTLOAD)
421 return false;
422
423 // Now let's find the callseq_start.
424 while (HasCallSeq && Chain.getOpcode() != ISD::CALLSEQ_START) {
425 if (!Chain.hasOneUse())
426 return false;
427 Chain = Chain.getOperand(0);
428 }
429
430 if (!Chain.getNumOperands())
431 return false;
432 // Since we are not checking for AA here, conservatively abort if the chain
433 // writes to memory. It's not safe to move the callee (a load) across a store.
434 if (isa<MemSDNode>(Chain.getNode()) &&
435 cast<MemSDNode>(Chain.getNode())->writeMem())
436 return false;
437 if (Chain.getOperand(0).getNode() == Callee.getNode())
438 return true;
439 if (Chain.getOperand(0).getOpcode() == ISD::TokenFactor &&
440 Callee.getValue(1).isOperandOf(Chain.getOperand(0).getNode()) &&
441 Callee.getValue(1).hasOneUse())
442 return true;
443 return false;
444 }
445
PreprocessISelDAG()446 void X86DAGToDAGISel::PreprocessISelDAG() {
447 // OptForSize is used in pattern predicates that isel is matching.
448 OptForSize = MF->getFunction()->getAttributes().
449 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize);
450
451 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
452 E = CurDAG->allnodes_end(); I != E; ) {
453 SDNode *N = I++; // Preincrement iterator to avoid invalidation issues.
454
455 if (OptLevel != CodeGenOpt::None &&
456 // Only does this when target favors doesn't favor register indirect
457 // call.
458 ((N->getOpcode() == X86ISD::CALL && !Subtarget->callRegIndirect()) ||
459 (N->getOpcode() == X86ISD::TC_RETURN &&
460 // Only does this if load can be folded into TC_RETURN.
461 (Subtarget->is64Bit() ||
462 getTargetMachine().getRelocationModel() != Reloc::PIC_)))) {
463 /// Also try moving call address load from outside callseq_start to just
464 /// before the call to allow it to be folded.
465 ///
466 /// [Load chain]
467 /// ^
468 /// |
469 /// [Load]
470 /// ^ ^
471 /// | |
472 /// / \--
473 /// / |
474 ///[CALLSEQ_START] |
475 /// ^ |
476 /// | |
477 /// [LOAD/C2Reg] |
478 /// | |
479 /// \ /
480 /// \ /
481 /// [CALL]
482 bool HasCallSeq = N->getOpcode() == X86ISD::CALL;
483 SDValue Chain = N->getOperand(0);
484 SDValue Load = N->getOperand(1);
485 if (!isCalleeLoad(Load, Chain, HasCallSeq))
486 continue;
487 MoveBelowOrigChain(CurDAG, Load, SDValue(N, 0), Chain);
488 ++NumLoadMoved;
489 continue;
490 }
491
492 // Lower fpround and fpextend nodes that target the FP stack to be store and
493 // load to the stack. This is a gross hack. We would like to simply mark
494 // these as being illegal, but when we do that, legalize produces these when
495 // it expands calls, then expands these in the same legalize pass. We would
496 // like dag combine to be able to hack on these between the call expansion
497 // and the node legalization. As such this pass basically does "really
498 // late" legalization of these inline with the X86 isel pass.
499 // FIXME: This should only happen when not compiled with -O0.
500 if (N->getOpcode() != ISD::FP_ROUND && N->getOpcode() != ISD::FP_EXTEND)
501 continue;
502
503 MVT SrcVT = N->getOperand(0).getSimpleValueType();
504 MVT DstVT = N->getSimpleValueType(0);
505
506 // If any of the sources are vectors, no fp stack involved.
507 if (SrcVT.isVector() || DstVT.isVector())
508 continue;
509
510 // If the source and destination are SSE registers, then this is a legal
511 // conversion that should not be lowered.
512 const X86TargetLowering *X86Lowering =
513 static_cast<const X86TargetLowering *>(getTargetLowering());
514 bool SrcIsSSE = X86Lowering->isScalarFPTypeInSSEReg(SrcVT);
515 bool DstIsSSE = X86Lowering->isScalarFPTypeInSSEReg(DstVT);
516 if (SrcIsSSE && DstIsSSE)
517 continue;
518
519 if (!SrcIsSSE && !DstIsSSE) {
520 // If this is an FPStack extension, it is a noop.
521 if (N->getOpcode() == ISD::FP_EXTEND)
522 continue;
523 // If this is a value-preserving FPStack truncation, it is a noop.
524 if (N->getConstantOperandVal(1))
525 continue;
526 }
527
528 // Here we could have an FP stack truncation or an FPStack <-> SSE convert.
529 // FPStack has extload and truncstore. SSE can fold direct loads into other
530 // operations. Based on this, decide what we want to do.
531 MVT MemVT;
532 if (N->getOpcode() == ISD::FP_ROUND)
533 MemVT = DstVT; // FP_ROUND must use DstVT, we can't do a 'trunc load'.
534 else
535 MemVT = SrcIsSSE ? SrcVT : DstVT;
536
537 SDValue MemTmp = CurDAG->CreateStackTemporary(MemVT);
538 SDLoc dl(N);
539
540 // FIXME: optimize the case where the src/dest is a load or store?
541 SDValue Store = CurDAG->getTruncStore(CurDAG->getEntryNode(), dl,
542 N->getOperand(0),
543 MemTmp, MachinePointerInfo(), MemVT,
544 false, false, 0);
545 SDValue Result = CurDAG->getExtLoad(ISD::EXTLOAD, dl, DstVT, Store, MemTmp,
546 MachinePointerInfo(),
547 MemVT, false, false, 0);
548
549 // We're about to replace all uses of the FP_ROUND/FP_EXTEND with the
550 // extload we created. This will cause general havok on the dag because
551 // anything below the conversion could be folded into other existing nodes.
552 // To avoid invalidating 'I', back it up to the convert node.
553 --I;
554 CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
555
556 // Now that we did that, the node is dead. Increment the iterator to the
557 // next node to process, then delete N.
558 ++I;
559 CurDAG->DeleteNode(N);
560 }
561 }
562
563
564 /// EmitSpecialCodeForMain - Emit any code that needs to be executed only in
565 /// the main function.
EmitSpecialCodeForMain(MachineBasicBlock * BB,MachineFrameInfo * MFI)566 void X86DAGToDAGISel::EmitSpecialCodeForMain(MachineBasicBlock *BB,
567 MachineFrameInfo *MFI) {
568 const TargetInstrInfo *TII = TM.getInstrInfo();
569 if (Subtarget->isTargetCygMing()) {
570 unsigned CallOp =
571 Subtarget->is64Bit() ? X86::CALL64pcrel32 : X86::CALLpcrel32;
572 BuildMI(BB, DebugLoc(),
573 TII->get(CallOp)).addExternalSymbol("__main");
574 }
575 }
576
EmitFunctionEntryCode()577 void X86DAGToDAGISel::EmitFunctionEntryCode() {
578 // If this is main, emit special code for main.
579 if (const Function *Fn = MF->getFunction())
580 if (Fn->hasExternalLinkage() && Fn->getName() == "main")
581 EmitSpecialCodeForMain(MF->begin(), MF->getFrameInfo());
582 }
583
isDispSafeForFrameIndex(int64_t Val)584 static bool isDispSafeForFrameIndex(int64_t Val) {
585 // On 64-bit platforms, we can run into an issue where a frame index
586 // includes a displacement that, when added to the explicit displacement,
587 // will overflow the displacement field. Assuming that the frame index
588 // displacement fits into a 31-bit integer (which is only slightly more
589 // aggressive than the current fundamental assumption that it fits into
590 // a 32-bit integer), a 31-bit disp should always be safe.
591 return isInt<31>(Val);
592 }
593
FoldOffsetIntoAddress(uint64_t Offset,X86ISelAddressMode & AM)594 bool X86DAGToDAGISel::FoldOffsetIntoAddress(uint64_t Offset,
595 X86ISelAddressMode &AM) {
596 int64_t Val = AM.Disp + Offset;
597 CodeModel::Model M = TM.getCodeModel();
598 if (Subtarget->is64Bit()) {
599 if (!X86::isOffsetSuitableForCodeModel(Val, M,
600 AM.hasSymbolicDisplacement()))
601 return true;
602 // In addition to the checks required for a register base, check that
603 // we do not try to use an unsafe Disp with a frame index.
604 if (AM.BaseType == X86ISelAddressMode::FrameIndexBase &&
605 !isDispSafeForFrameIndex(Val))
606 return true;
607 }
608 AM.Disp = Val;
609 return false;
610
611 }
612
MatchLoadInAddress(LoadSDNode * N,X86ISelAddressMode & AM)613 bool X86DAGToDAGISel::MatchLoadInAddress(LoadSDNode *N, X86ISelAddressMode &AM){
614 SDValue Address = N->getOperand(1);
615
616 // load gs:0 -> GS segment register.
617 // load fs:0 -> FS segment register.
618 //
619 // This optimization is valid because the GNU TLS model defines that
620 // gs:0 (or fs:0 on X86-64) contains its own address.
621 // For more information see http://people.redhat.com/drepper/tls.pdf
622 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Address))
623 if (C->getSExtValue() == 0 && AM.Segment.getNode() == nullptr &&
624 Subtarget->isTargetLinux())
625 switch (N->getPointerInfo().getAddrSpace()) {
626 case 256:
627 AM.Segment = CurDAG->getRegister(X86::GS, MVT::i16);
628 return false;
629 case 257:
630 AM.Segment = CurDAG->getRegister(X86::FS, MVT::i16);
631 return false;
632 }
633
634 return true;
635 }
636
637 /// MatchWrapper - Try to match X86ISD::Wrapper and X86ISD::WrapperRIP nodes
638 /// into an addressing mode. These wrap things that will resolve down into a
639 /// symbol reference. If no match is possible, this returns true, otherwise it
640 /// returns false.
MatchWrapper(SDValue N,X86ISelAddressMode & AM)641 bool X86DAGToDAGISel::MatchWrapper(SDValue N, X86ISelAddressMode &AM) {
642 // If the addressing mode already has a symbol as the displacement, we can
643 // never match another symbol.
644 if (AM.hasSymbolicDisplacement())
645 return true;
646
647 SDValue N0 = N.getOperand(0);
648 CodeModel::Model M = TM.getCodeModel();
649
650 // Handle X86-64 rip-relative addresses. We check this before checking direct
651 // folding because RIP is preferable to non-RIP accesses.
652 if (Subtarget->is64Bit() && N.getOpcode() == X86ISD::WrapperRIP &&
653 // Under X86-64 non-small code model, GV (and friends) are 64-bits, so
654 // they cannot be folded into immediate fields.
655 // FIXME: This can be improved for kernel and other models?
656 (M == CodeModel::Small || M == CodeModel::Kernel)) {
657 // Base and index reg must be 0 in order to use %rip as base.
658 if (AM.hasBaseOrIndexReg())
659 return true;
660 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(N0)) {
661 X86ISelAddressMode Backup = AM;
662 AM.GV = G->getGlobal();
663 AM.SymbolFlags = G->getTargetFlags();
664 if (FoldOffsetIntoAddress(G->getOffset(), AM)) {
665 AM = Backup;
666 return true;
667 }
668 } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N0)) {
669 X86ISelAddressMode Backup = AM;
670 AM.CP = CP->getConstVal();
671 AM.Align = CP->getAlignment();
672 AM.SymbolFlags = CP->getTargetFlags();
673 if (FoldOffsetIntoAddress(CP->getOffset(), AM)) {
674 AM = Backup;
675 return true;
676 }
677 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(N0)) {
678 AM.ES = S->getSymbol();
679 AM.SymbolFlags = S->getTargetFlags();
680 } else if (JumpTableSDNode *J = dyn_cast<JumpTableSDNode>(N0)) {
681 AM.JT = J->getIndex();
682 AM.SymbolFlags = J->getTargetFlags();
683 } else if (BlockAddressSDNode *BA = dyn_cast<BlockAddressSDNode>(N0)) {
684 X86ISelAddressMode Backup = AM;
685 AM.BlockAddr = BA->getBlockAddress();
686 AM.SymbolFlags = BA->getTargetFlags();
687 if (FoldOffsetIntoAddress(BA->getOffset(), AM)) {
688 AM = Backup;
689 return true;
690 }
691 } else
692 llvm_unreachable("Unhandled symbol reference node.");
693
694 if (N.getOpcode() == X86ISD::WrapperRIP)
695 AM.setBaseReg(CurDAG->getRegister(X86::RIP, MVT::i64));
696 return false;
697 }
698
699 // Handle the case when globals fit in our immediate field: This is true for
700 // X86-32 always and X86-64 when in -mcmodel=small mode. In 64-bit
701 // mode, this only applies to a non-RIP-relative computation.
702 if (!Subtarget->is64Bit() ||
703 M == CodeModel::Small || M == CodeModel::Kernel) {
704 assert(N.getOpcode() != X86ISD::WrapperRIP &&
705 "RIP-relative addressing already handled");
706 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(N0)) {
707 AM.GV = G->getGlobal();
708 AM.Disp += G->getOffset();
709 AM.SymbolFlags = G->getTargetFlags();
710 } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N0)) {
711 AM.CP = CP->getConstVal();
712 AM.Align = CP->getAlignment();
713 AM.Disp += CP->getOffset();
714 AM.SymbolFlags = CP->getTargetFlags();
715 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(N0)) {
716 AM.ES = S->getSymbol();
717 AM.SymbolFlags = S->getTargetFlags();
718 } else if (JumpTableSDNode *J = dyn_cast<JumpTableSDNode>(N0)) {
719 AM.JT = J->getIndex();
720 AM.SymbolFlags = J->getTargetFlags();
721 } else if (BlockAddressSDNode *BA = dyn_cast<BlockAddressSDNode>(N0)) {
722 AM.BlockAddr = BA->getBlockAddress();
723 AM.Disp += BA->getOffset();
724 AM.SymbolFlags = BA->getTargetFlags();
725 } else
726 llvm_unreachable("Unhandled symbol reference node.");
727 return false;
728 }
729
730 return true;
731 }
732
733 /// MatchAddress - Add the specified node to the specified addressing mode,
734 /// returning true if it cannot be done. This just pattern matches for the
735 /// addressing mode.
MatchAddress(SDValue N,X86ISelAddressMode & AM)736 bool X86DAGToDAGISel::MatchAddress(SDValue N, X86ISelAddressMode &AM) {
737 if (MatchAddressRecursively(N, AM, 0))
738 return true;
739
740 // Post-processing: Convert lea(,%reg,2) to lea(%reg,%reg), which has
741 // a smaller encoding and avoids a scaled-index.
742 if (AM.Scale == 2 &&
743 AM.BaseType == X86ISelAddressMode::RegBase &&
744 AM.Base_Reg.getNode() == nullptr) {
745 AM.Base_Reg = AM.IndexReg;
746 AM.Scale = 1;
747 }
748
749 // Post-processing: Convert foo to foo(%rip), even in non-PIC mode,
750 // because it has a smaller encoding.
751 // TODO: Which other code models can use this?
752 if (TM.getCodeModel() == CodeModel::Small &&
753 Subtarget->is64Bit() &&
754 AM.Scale == 1 &&
755 AM.BaseType == X86ISelAddressMode::RegBase &&
756 AM.Base_Reg.getNode() == nullptr &&
757 AM.IndexReg.getNode() == nullptr &&
758 AM.SymbolFlags == X86II::MO_NO_FLAG &&
759 AM.hasSymbolicDisplacement())
760 AM.Base_Reg = CurDAG->getRegister(X86::RIP, MVT::i64);
761
762 return false;
763 }
764
765 // Insert a node into the DAG at least before the Pos node's position. This
766 // will reposition the node as needed, and will assign it a node ID that is <=
767 // the Pos node's ID. Note that this does *not* preserve the uniqueness of node
768 // IDs! The selection DAG must no longer depend on their uniqueness when this
769 // is used.
InsertDAGNode(SelectionDAG & DAG,SDValue Pos,SDValue N)770 static void InsertDAGNode(SelectionDAG &DAG, SDValue Pos, SDValue N) {
771 if (N.getNode()->getNodeId() == -1 ||
772 N.getNode()->getNodeId() > Pos.getNode()->getNodeId()) {
773 DAG.RepositionNode(Pos.getNode(), N.getNode());
774 N.getNode()->setNodeId(Pos.getNode()->getNodeId());
775 }
776 }
777
778 // Transform "(X >> (8-C1)) & C2" to "(X >> 8) & 0xff)" if safe. This
779 // allows us to convert the shift and and into an h-register extract and
780 // a scaled index. Returns false if the simplification is performed.
FoldMaskAndShiftToExtract(SelectionDAG & DAG,SDValue N,uint64_t Mask,SDValue Shift,SDValue X,X86ISelAddressMode & AM)781 static bool FoldMaskAndShiftToExtract(SelectionDAG &DAG, SDValue N,
782 uint64_t Mask,
783 SDValue Shift, SDValue X,
784 X86ISelAddressMode &AM) {
785 if (Shift.getOpcode() != ISD::SRL ||
786 !isa<ConstantSDNode>(Shift.getOperand(1)) ||
787 !Shift.hasOneUse())
788 return true;
789
790 int ScaleLog = 8 - Shift.getConstantOperandVal(1);
791 if (ScaleLog <= 0 || ScaleLog >= 4 ||
792 Mask != (0xffu << ScaleLog))
793 return true;
794
795 MVT VT = N.getSimpleValueType();
796 SDLoc DL(N);
797 SDValue Eight = DAG.getConstant(8, MVT::i8);
798 SDValue NewMask = DAG.getConstant(0xff, VT);
799 SDValue Srl = DAG.getNode(ISD::SRL, DL, VT, X, Eight);
800 SDValue And = DAG.getNode(ISD::AND, DL, VT, Srl, NewMask);
801 SDValue ShlCount = DAG.getConstant(ScaleLog, MVT::i8);
802 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, And, ShlCount);
803
804 // Insert the new nodes into the topological ordering. We must do this in
805 // a valid topological ordering as nothing is going to go back and re-sort
806 // these nodes. We continually insert before 'N' in sequence as this is
807 // essentially a pre-flattened and pre-sorted sequence of nodes. There is no
808 // hierarchy left to express.
809 InsertDAGNode(DAG, N, Eight);
810 InsertDAGNode(DAG, N, Srl);
811 InsertDAGNode(DAG, N, NewMask);
812 InsertDAGNode(DAG, N, And);
813 InsertDAGNode(DAG, N, ShlCount);
814 InsertDAGNode(DAG, N, Shl);
815 DAG.ReplaceAllUsesWith(N, Shl);
816 AM.IndexReg = And;
817 AM.Scale = (1 << ScaleLog);
818 return false;
819 }
820
821 // Transforms "(X << C1) & C2" to "(X & (C2>>C1)) << C1" if safe and if this
822 // allows us to fold the shift into this addressing mode. Returns false if the
823 // transform succeeded.
FoldMaskedShiftToScaledMask(SelectionDAG & DAG,SDValue N,uint64_t Mask,SDValue Shift,SDValue X,X86ISelAddressMode & AM)824 static bool FoldMaskedShiftToScaledMask(SelectionDAG &DAG, SDValue N,
825 uint64_t Mask,
826 SDValue Shift, SDValue X,
827 X86ISelAddressMode &AM) {
828 if (Shift.getOpcode() != ISD::SHL ||
829 !isa<ConstantSDNode>(Shift.getOperand(1)))
830 return true;
831
832 // Not likely to be profitable if either the AND or SHIFT node has more
833 // than one use (unless all uses are for address computation). Besides,
834 // isel mechanism requires their node ids to be reused.
835 if (!N.hasOneUse() || !Shift.hasOneUse())
836 return true;
837
838 // Verify that the shift amount is something we can fold.
839 unsigned ShiftAmt = Shift.getConstantOperandVal(1);
840 if (ShiftAmt != 1 && ShiftAmt != 2 && ShiftAmt != 3)
841 return true;
842
843 MVT VT = N.getSimpleValueType();
844 SDLoc DL(N);
845 SDValue NewMask = DAG.getConstant(Mask >> ShiftAmt, VT);
846 SDValue NewAnd = DAG.getNode(ISD::AND, DL, VT, X, NewMask);
847 SDValue NewShift = DAG.getNode(ISD::SHL, DL, VT, NewAnd, Shift.getOperand(1));
848
849 // Insert the new nodes into the topological ordering. We must do this in
850 // a valid topological ordering as nothing is going to go back and re-sort
851 // these nodes. We continually insert before 'N' in sequence as this is
852 // essentially a pre-flattened and pre-sorted sequence of nodes. There is no
853 // hierarchy left to express.
854 InsertDAGNode(DAG, N, NewMask);
855 InsertDAGNode(DAG, N, NewAnd);
856 InsertDAGNode(DAG, N, NewShift);
857 DAG.ReplaceAllUsesWith(N, NewShift);
858
859 AM.Scale = 1 << ShiftAmt;
860 AM.IndexReg = NewAnd;
861 return false;
862 }
863
864 // Implement some heroics to detect shifts of masked values where the mask can
865 // be replaced by extending the shift and undoing that in the addressing mode
866 // scale. Patterns such as (shl (srl x, c1), c2) are canonicalized into (and
867 // (srl x, SHIFT), MASK) by DAGCombines that don't know the shl can be done in
868 // the addressing mode. This results in code such as:
869 //
870 // int f(short *y, int *lookup_table) {
871 // ...
872 // return *y + lookup_table[*y >> 11];
873 // }
874 //
875 // Turning into:
876 // movzwl (%rdi), %eax
877 // movl %eax, %ecx
878 // shrl $11, %ecx
879 // addl (%rsi,%rcx,4), %eax
880 //
881 // Instead of:
882 // movzwl (%rdi), %eax
883 // movl %eax, %ecx
884 // shrl $9, %ecx
885 // andl $124, %rcx
886 // addl (%rsi,%rcx), %eax
887 //
888 // Note that this function assumes the mask is provided as a mask *after* the
889 // value is shifted. The input chain may or may not match that, but computing
890 // such a mask is trivial.
FoldMaskAndShiftToScale(SelectionDAG & DAG,SDValue N,uint64_t Mask,SDValue Shift,SDValue X,X86ISelAddressMode & AM)891 static bool FoldMaskAndShiftToScale(SelectionDAG &DAG, SDValue N,
892 uint64_t Mask,
893 SDValue Shift, SDValue X,
894 X86ISelAddressMode &AM) {
895 if (Shift.getOpcode() != ISD::SRL || !Shift.hasOneUse() ||
896 !isa<ConstantSDNode>(Shift.getOperand(1)))
897 return true;
898
899 unsigned ShiftAmt = Shift.getConstantOperandVal(1);
900 unsigned MaskLZ = countLeadingZeros(Mask);
901 unsigned MaskTZ = countTrailingZeros(Mask);
902
903 // The amount of shift we're trying to fit into the addressing mode is taken
904 // from the trailing zeros of the mask.
905 unsigned AMShiftAmt = MaskTZ;
906
907 // There is nothing we can do here unless the mask is removing some bits.
908 // Also, the addressing mode can only represent shifts of 1, 2, or 3 bits.
909 if (AMShiftAmt <= 0 || AMShiftAmt > 3) return true;
910
911 // We also need to ensure that mask is a continuous run of bits.
912 if (CountTrailingOnes_64(Mask >> MaskTZ) + MaskTZ + MaskLZ != 64) return true;
913
914 // Scale the leading zero count down based on the actual size of the value.
915 // Also scale it down based on the size of the shift.
916 MaskLZ -= (64 - X.getSimpleValueType().getSizeInBits()) + ShiftAmt;
917
918 // The final check is to ensure that any masked out high bits of X are
919 // already known to be zero. Otherwise, the mask has a semantic impact
920 // other than masking out a couple of low bits. Unfortunately, because of
921 // the mask, zero extensions will be removed from operands in some cases.
922 // This code works extra hard to look through extensions because we can
923 // replace them with zero extensions cheaply if necessary.
924 bool ReplacingAnyExtend = false;
925 if (X.getOpcode() == ISD::ANY_EXTEND) {
926 unsigned ExtendBits = X.getSimpleValueType().getSizeInBits() -
927 X.getOperand(0).getSimpleValueType().getSizeInBits();
928 // Assume that we'll replace the any-extend with a zero-extend, and
929 // narrow the search to the extended value.
930 X = X.getOperand(0);
931 MaskLZ = ExtendBits > MaskLZ ? 0 : MaskLZ - ExtendBits;
932 ReplacingAnyExtend = true;
933 }
934 APInt MaskedHighBits =
935 APInt::getHighBitsSet(X.getSimpleValueType().getSizeInBits(), MaskLZ);
936 APInt KnownZero, KnownOne;
937 DAG.computeKnownBits(X, KnownZero, KnownOne);
938 if (MaskedHighBits != KnownZero) return true;
939
940 // We've identified a pattern that can be transformed into a single shift
941 // and an addressing mode. Make it so.
942 MVT VT = N.getSimpleValueType();
943 if (ReplacingAnyExtend) {
944 assert(X.getValueType() != VT);
945 // We looked through an ANY_EXTEND node, insert a ZERO_EXTEND.
946 SDValue NewX = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(X), VT, X);
947 InsertDAGNode(DAG, N, NewX);
948 X = NewX;
949 }
950 SDLoc DL(N);
951 SDValue NewSRLAmt = DAG.getConstant(ShiftAmt + AMShiftAmt, MVT::i8);
952 SDValue NewSRL = DAG.getNode(ISD::SRL, DL, VT, X, NewSRLAmt);
953 SDValue NewSHLAmt = DAG.getConstant(AMShiftAmt, MVT::i8);
954 SDValue NewSHL = DAG.getNode(ISD::SHL, DL, VT, NewSRL, NewSHLAmt);
955
956 // Insert the new nodes into the topological ordering. We must do this in
957 // a valid topological ordering as nothing is going to go back and re-sort
958 // these nodes. We continually insert before 'N' in sequence as this is
959 // essentially a pre-flattened and pre-sorted sequence of nodes. There is no
960 // hierarchy left to express.
961 InsertDAGNode(DAG, N, NewSRLAmt);
962 InsertDAGNode(DAG, N, NewSRL);
963 InsertDAGNode(DAG, N, NewSHLAmt);
964 InsertDAGNode(DAG, N, NewSHL);
965 DAG.ReplaceAllUsesWith(N, NewSHL);
966
967 AM.Scale = 1 << AMShiftAmt;
968 AM.IndexReg = NewSRL;
969 return false;
970 }
971
MatchAddressRecursively(SDValue N,X86ISelAddressMode & AM,unsigned Depth)972 bool X86DAGToDAGISel::MatchAddressRecursively(SDValue N, X86ISelAddressMode &AM,
973 unsigned Depth) {
974 SDLoc dl(N);
975 DEBUG({
976 dbgs() << "MatchAddress: ";
977 AM.dump();
978 });
979 // Limit recursion.
980 if (Depth > 5)
981 return MatchAddressBase(N, AM);
982
983 // If this is already a %rip relative address, we can only merge immediates
984 // into it. Instead of handling this in every case, we handle it here.
985 // RIP relative addressing: %rip + 32-bit displacement!
986 if (AM.isRIPRelative()) {
987 // FIXME: JumpTable and ExternalSymbol address currently don't like
988 // displacements. It isn't very important, but this should be fixed for
989 // consistency.
990 if (!AM.ES && AM.JT != -1) return true;
991
992 if (ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(N))
993 if (!FoldOffsetIntoAddress(Cst->getSExtValue(), AM))
994 return false;
995 return true;
996 }
997
998 switch (N.getOpcode()) {
999 default: break;
1000 case ISD::Constant: {
1001 uint64_t Val = cast<ConstantSDNode>(N)->getSExtValue();
1002 if (!FoldOffsetIntoAddress(Val, AM))
1003 return false;
1004 break;
1005 }
1006
1007 case X86ISD::Wrapper:
1008 case X86ISD::WrapperRIP:
1009 if (!MatchWrapper(N, AM))
1010 return false;
1011 break;
1012
1013 case ISD::LOAD:
1014 if (!MatchLoadInAddress(cast<LoadSDNode>(N), AM))
1015 return false;
1016 break;
1017
1018 case ISD::FrameIndex:
1019 if (AM.BaseType == X86ISelAddressMode::RegBase &&
1020 AM.Base_Reg.getNode() == nullptr &&
1021 (!Subtarget->is64Bit() || isDispSafeForFrameIndex(AM.Disp))) {
1022 AM.BaseType = X86ISelAddressMode::FrameIndexBase;
1023 AM.Base_FrameIndex = cast<FrameIndexSDNode>(N)->getIndex();
1024 return false;
1025 }
1026 break;
1027
1028 case ISD::SHL:
1029 if (AM.IndexReg.getNode() != nullptr || AM.Scale != 1)
1030 break;
1031
1032 if (ConstantSDNode
1033 *CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1))) {
1034 unsigned Val = CN->getZExtValue();
1035 // Note that we handle x<<1 as (,x,2) rather than (x,x) here so
1036 // that the base operand remains free for further matching. If
1037 // the base doesn't end up getting used, a post-processing step
1038 // in MatchAddress turns (,x,2) into (x,x), which is cheaper.
1039 if (Val == 1 || Val == 2 || Val == 3) {
1040 AM.Scale = 1 << Val;
1041 SDValue ShVal = N.getNode()->getOperand(0);
1042
1043 // Okay, we know that we have a scale by now. However, if the scaled
1044 // value is an add of something and a constant, we can fold the
1045 // constant into the disp field here.
1046 if (CurDAG->isBaseWithConstantOffset(ShVal)) {
1047 AM.IndexReg = ShVal.getNode()->getOperand(0);
1048 ConstantSDNode *AddVal =
1049 cast<ConstantSDNode>(ShVal.getNode()->getOperand(1));
1050 uint64_t Disp = (uint64_t)AddVal->getSExtValue() << Val;
1051 if (!FoldOffsetIntoAddress(Disp, AM))
1052 return false;
1053 }
1054
1055 AM.IndexReg = ShVal;
1056 return false;
1057 }
1058 }
1059 break;
1060
1061 case ISD::SRL: {
1062 // Scale must not be used already.
1063 if (AM.IndexReg.getNode() != nullptr || AM.Scale != 1) break;
1064
1065 SDValue And = N.getOperand(0);
1066 if (And.getOpcode() != ISD::AND) break;
1067 SDValue X = And.getOperand(0);
1068
1069 // We only handle up to 64-bit values here as those are what matter for
1070 // addressing mode optimizations.
1071 if (X.getSimpleValueType().getSizeInBits() > 64) break;
1072
1073 // The mask used for the transform is expected to be post-shift, but we
1074 // found the shift first so just apply the shift to the mask before passing
1075 // it down.
1076 if (!isa<ConstantSDNode>(N.getOperand(1)) ||
1077 !isa<ConstantSDNode>(And.getOperand(1)))
1078 break;
1079 uint64_t Mask = And.getConstantOperandVal(1) >> N.getConstantOperandVal(1);
1080
1081 // Try to fold the mask and shift into the scale, and return false if we
1082 // succeed.
1083 if (!FoldMaskAndShiftToScale(*CurDAG, N, Mask, N, X, AM))
1084 return false;
1085 break;
1086 }
1087
1088 case ISD::SMUL_LOHI:
1089 case ISD::UMUL_LOHI:
1090 // A mul_lohi where we need the low part can be folded as a plain multiply.
1091 if (N.getResNo() != 0) break;
1092 // FALL THROUGH
1093 case ISD::MUL:
1094 case X86ISD::MUL_IMM:
1095 // X*[3,5,9] -> X+X*[2,4,8]
1096 if (AM.BaseType == X86ISelAddressMode::RegBase &&
1097 AM.Base_Reg.getNode() == nullptr &&
1098 AM.IndexReg.getNode() == nullptr) {
1099 if (ConstantSDNode
1100 *CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1)))
1101 if (CN->getZExtValue() == 3 || CN->getZExtValue() == 5 ||
1102 CN->getZExtValue() == 9) {
1103 AM.Scale = unsigned(CN->getZExtValue())-1;
1104
1105 SDValue MulVal = N.getNode()->getOperand(0);
1106 SDValue Reg;
1107
1108 // Okay, we know that we have a scale by now. However, if the scaled
1109 // value is an add of something and a constant, we can fold the
1110 // constant into the disp field here.
1111 if (MulVal.getNode()->getOpcode() == ISD::ADD && MulVal.hasOneUse() &&
1112 isa<ConstantSDNode>(MulVal.getNode()->getOperand(1))) {
1113 Reg = MulVal.getNode()->getOperand(0);
1114 ConstantSDNode *AddVal =
1115 cast<ConstantSDNode>(MulVal.getNode()->getOperand(1));
1116 uint64_t Disp = AddVal->getSExtValue() * CN->getZExtValue();
1117 if (FoldOffsetIntoAddress(Disp, AM))
1118 Reg = N.getNode()->getOperand(0);
1119 } else {
1120 Reg = N.getNode()->getOperand(0);
1121 }
1122
1123 AM.IndexReg = AM.Base_Reg = Reg;
1124 return false;
1125 }
1126 }
1127 break;
1128
1129 case ISD::SUB: {
1130 // Given A-B, if A can be completely folded into the address and
1131 // the index field with the index field unused, use -B as the index.
1132 // This is a win if a has multiple parts that can be folded into
1133 // the address. Also, this saves a mov if the base register has
1134 // other uses, since it avoids a two-address sub instruction, however
1135 // it costs an additional mov if the index register has other uses.
1136
1137 // Add an artificial use to this node so that we can keep track of
1138 // it if it gets CSE'd with a different node.
1139 HandleSDNode Handle(N);
1140
1141 // Test if the LHS of the sub can be folded.
1142 X86ISelAddressMode Backup = AM;
1143 if (MatchAddressRecursively(N.getNode()->getOperand(0), AM, Depth+1)) {
1144 AM = Backup;
1145 break;
1146 }
1147 // Test if the index field is free for use.
1148 if (AM.IndexReg.getNode() || AM.isRIPRelative()) {
1149 AM = Backup;
1150 break;
1151 }
1152
1153 int Cost = 0;
1154 SDValue RHS = Handle.getValue().getNode()->getOperand(1);
1155 // If the RHS involves a register with multiple uses, this
1156 // transformation incurs an extra mov, due to the neg instruction
1157 // clobbering its operand.
1158 if (!RHS.getNode()->hasOneUse() ||
1159 RHS.getNode()->getOpcode() == ISD::CopyFromReg ||
1160 RHS.getNode()->getOpcode() == ISD::TRUNCATE ||
1161 RHS.getNode()->getOpcode() == ISD::ANY_EXTEND ||
1162 (RHS.getNode()->getOpcode() == ISD::ZERO_EXTEND &&
1163 RHS.getNode()->getOperand(0).getValueType() == MVT::i32))
1164 ++Cost;
1165 // If the base is a register with multiple uses, this
1166 // transformation may save a mov.
1167 if ((AM.BaseType == X86ISelAddressMode::RegBase &&
1168 AM.Base_Reg.getNode() &&
1169 !AM.Base_Reg.getNode()->hasOneUse()) ||
1170 AM.BaseType == X86ISelAddressMode::FrameIndexBase)
1171 --Cost;
1172 // If the folded LHS was interesting, this transformation saves
1173 // address arithmetic.
1174 if ((AM.hasSymbolicDisplacement() && !Backup.hasSymbolicDisplacement()) +
1175 ((AM.Disp != 0) && (Backup.Disp == 0)) +
1176 (AM.Segment.getNode() && !Backup.Segment.getNode()) >= 2)
1177 --Cost;
1178 // If it doesn't look like it may be an overall win, don't do it.
1179 if (Cost >= 0) {
1180 AM = Backup;
1181 break;
1182 }
1183
1184 // Ok, the transformation is legal and appears profitable. Go for it.
1185 SDValue Zero = CurDAG->getConstant(0, N.getValueType());
1186 SDValue Neg = CurDAG->getNode(ISD::SUB, dl, N.getValueType(), Zero, RHS);
1187 AM.IndexReg = Neg;
1188 AM.Scale = 1;
1189
1190 // Insert the new nodes into the topological ordering.
1191 InsertDAGNode(*CurDAG, N, Zero);
1192 InsertDAGNode(*CurDAG, N, Neg);
1193 return false;
1194 }
1195
1196 case ISD::ADD: {
1197 // Add an artificial use to this node so that we can keep track of
1198 // it if it gets CSE'd with a different node.
1199 HandleSDNode Handle(N);
1200
1201 X86ISelAddressMode Backup = AM;
1202 if (!MatchAddressRecursively(N.getOperand(0), AM, Depth+1) &&
1203 !MatchAddressRecursively(Handle.getValue().getOperand(1), AM, Depth+1))
1204 return false;
1205 AM = Backup;
1206
1207 // Try again after commuting the operands.
1208 if (!MatchAddressRecursively(Handle.getValue().getOperand(1), AM, Depth+1)&&
1209 !MatchAddressRecursively(Handle.getValue().getOperand(0), AM, Depth+1))
1210 return false;
1211 AM = Backup;
1212
1213 // If we couldn't fold both operands into the address at the same time,
1214 // see if we can just put each operand into a register and fold at least
1215 // the add.
1216 if (AM.BaseType == X86ISelAddressMode::RegBase &&
1217 !AM.Base_Reg.getNode() &&
1218 !AM.IndexReg.getNode()) {
1219 N = Handle.getValue();
1220 AM.Base_Reg = N.getOperand(0);
1221 AM.IndexReg = N.getOperand(1);
1222 AM.Scale = 1;
1223 return false;
1224 }
1225 N = Handle.getValue();
1226 break;
1227 }
1228
1229 case ISD::OR:
1230 // Handle "X | C" as "X + C" iff X is known to have C bits clear.
1231 if (CurDAG->isBaseWithConstantOffset(N)) {
1232 X86ISelAddressMode Backup = AM;
1233 ConstantSDNode *CN = cast<ConstantSDNode>(N.getOperand(1));
1234
1235 // Start with the LHS as an addr mode.
1236 if (!MatchAddressRecursively(N.getOperand(0), AM, Depth+1) &&
1237 !FoldOffsetIntoAddress(CN->getSExtValue(), AM))
1238 return false;
1239 AM = Backup;
1240 }
1241 break;
1242
1243 case ISD::AND: {
1244 // Perform some heroic transforms on an and of a constant-count shift
1245 // with a constant to enable use of the scaled offset field.
1246
1247 // Scale must not be used already.
1248 if (AM.IndexReg.getNode() != nullptr || AM.Scale != 1) break;
1249
1250 SDValue Shift = N.getOperand(0);
1251 if (Shift.getOpcode() != ISD::SRL && Shift.getOpcode() != ISD::SHL) break;
1252 SDValue X = Shift.getOperand(0);
1253
1254 // We only handle up to 64-bit values here as those are what matter for
1255 // addressing mode optimizations.
1256 if (X.getSimpleValueType().getSizeInBits() > 64) break;
1257
1258 if (!isa<ConstantSDNode>(N.getOperand(1)))
1259 break;
1260 uint64_t Mask = N.getConstantOperandVal(1);
1261
1262 // Try to fold the mask and shift into an extract and scale.
1263 if (!FoldMaskAndShiftToExtract(*CurDAG, N, Mask, Shift, X, AM))
1264 return false;
1265
1266 // Try to fold the mask and shift directly into the scale.
1267 if (!FoldMaskAndShiftToScale(*CurDAG, N, Mask, Shift, X, AM))
1268 return false;
1269
1270 // Try to swap the mask and shift to place shifts which can be done as
1271 // a scale on the outside of the mask.
1272 if (!FoldMaskedShiftToScaledMask(*CurDAG, N, Mask, Shift, X, AM))
1273 return false;
1274 break;
1275 }
1276 }
1277
1278 return MatchAddressBase(N, AM);
1279 }
1280
1281 /// MatchAddressBase - Helper for MatchAddress. Add the specified node to the
1282 /// specified addressing mode without any further recursion.
MatchAddressBase(SDValue N,X86ISelAddressMode & AM)1283 bool X86DAGToDAGISel::MatchAddressBase(SDValue N, X86ISelAddressMode &AM) {
1284 // Is the base register already occupied?
1285 if (AM.BaseType != X86ISelAddressMode::RegBase || AM.Base_Reg.getNode()) {
1286 // If so, check to see if the scale index register is set.
1287 if (!AM.IndexReg.getNode()) {
1288 AM.IndexReg = N;
1289 AM.Scale = 1;
1290 return false;
1291 }
1292
1293 // Otherwise, we cannot select it.
1294 return true;
1295 }
1296
1297 // Default, generate it as a register.
1298 AM.BaseType = X86ISelAddressMode::RegBase;
1299 AM.Base_Reg = N;
1300 return false;
1301 }
1302
1303 /// SelectAddr - returns true if it is able pattern match an addressing mode.
1304 /// It returns the operands which make up the maximal addressing mode it can
1305 /// match by reference.
1306 ///
1307 /// Parent is the parent node of the addr operand that is being matched. It
1308 /// is always a load, store, atomic node, or null. It is only null when
1309 /// checking memory operands for inline asm nodes.
SelectAddr(SDNode * Parent,SDValue N,SDValue & Base,SDValue & Scale,SDValue & Index,SDValue & Disp,SDValue & Segment)1310 bool X86DAGToDAGISel::SelectAddr(SDNode *Parent, SDValue N, SDValue &Base,
1311 SDValue &Scale, SDValue &Index,
1312 SDValue &Disp, SDValue &Segment) {
1313 X86ISelAddressMode AM;
1314
1315 if (Parent &&
1316 // This list of opcodes are all the nodes that have an "addr:$ptr" operand
1317 // that are not a MemSDNode, and thus don't have proper addrspace info.
1318 Parent->getOpcode() != ISD::INTRINSIC_W_CHAIN && // unaligned loads, fixme
1319 Parent->getOpcode() != ISD::INTRINSIC_VOID && // nontemporal stores
1320 Parent->getOpcode() != X86ISD::TLSCALL && // Fixme
1321 Parent->getOpcode() != X86ISD::EH_SJLJ_SETJMP && // setjmp
1322 Parent->getOpcode() != X86ISD::EH_SJLJ_LONGJMP) { // longjmp
1323 unsigned AddrSpace =
1324 cast<MemSDNode>(Parent)->getPointerInfo().getAddrSpace();
1325 // AddrSpace 256 -> GS, 257 -> FS.
1326 if (AddrSpace == 256)
1327 AM.Segment = CurDAG->getRegister(X86::GS, MVT::i16);
1328 if (AddrSpace == 257)
1329 AM.Segment = CurDAG->getRegister(X86::FS, MVT::i16);
1330 }
1331
1332 if (MatchAddress(N, AM))
1333 return false;
1334
1335 MVT VT = N.getSimpleValueType();
1336 if (AM.BaseType == X86ISelAddressMode::RegBase) {
1337 if (!AM.Base_Reg.getNode())
1338 AM.Base_Reg = CurDAG->getRegister(0, VT);
1339 }
1340
1341 if (!AM.IndexReg.getNode())
1342 AM.IndexReg = CurDAG->getRegister(0, VT);
1343
1344 getAddressOperands(AM, Base, Scale, Index, Disp, Segment);
1345 return true;
1346 }
1347
1348 /// SelectScalarSSELoad - Match a scalar SSE load. In particular, we want to
1349 /// match a load whose top elements are either undef or zeros. The load flavor
1350 /// is derived from the type of N, which is either v4f32 or v2f64.
1351 ///
1352 /// We also return:
1353 /// PatternChainNode: this is the matched node that has a chain input and
1354 /// output.
SelectScalarSSELoad(SDNode * Root,SDValue N,SDValue & Base,SDValue & Scale,SDValue & Index,SDValue & Disp,SDValue & Segment,SDValue & PatternNodeWithChain)1355 bool X86DAGToDAGISel::SelectScalarSSELoad(SDNode *Root,
1356 SDValue N, SDValue &Base,
1357 SDValue &Scale, SDValue &Index,
1358 SDValue &Disp, SDValue &Segment,
1359 SDValue &PatternNodeWithChain) {
1360 if (N.getOpcode() == ISD::SCALAR_TO_VECTOR) {
1361 PatternNodeWithChain = N.getOperand(0);
1362 if (ISD::isNON_EXTLoad(PatternNodeWithChain.getNode()) &&
1363 PatternNodeWithChain.hasOneUse() &&
1364 IsProfitableToFold(N.getOperand(0), N.getNode(), Root) &&
1365 IsLegalToFold(N.getOperand(0), N.getNode(), Root, OptLevel)) {
1366 LoadSDNode *LD = cast<LoadSDNode>(PatternNodeWithChain);
1367 if (!SelectAddr(LD, LD->getBasePtr(), Base, Scale, Index, Disp, Segment))
1368 return false;
1369 return true;
1370 }
1371 }
1372
1373 // Also handle the case where we explicitly require zeros in the top
1374 // elements. This is a vector shuffle from the zero vector.
1375 if (N.getOpcode() == X86ISD::VZEXT_MOVL && N.getNode()->hasOneUse() &&
1376 // Check to see if the top elements are all zeros (or bitcast of zeros).
1377 N.getOperand(0).getOpcode() == ISD::SCALAR_TO_VECTOR &&
1378 N.getOperand(0).getNode()->hasOneUse() &&
1379 ISD::isNON_EXTLoad(N.getOperand(0).getOperand(0).getNode()) &&
1380 N.getOperand(0).getOperand(0).hasOneUse() &&
1381 IsProfitableToFold(N.getOperand(0), N.getNode(), Root) &&
1382 IsLegalToFold(N.getOperand(0), N.getNode(), Root, OptLevel)) {
1383 // Okay, this is a zero extending load. Fold it.
1384 LoadSDNode *LD = cast<LoadSDNode>(N.getOperand(0).getOperand(0));
1385 if (!SelectAddr(LD, LD->getBasePtr(), Base, Scale, Index, Disp, Segment))
1386 return false;
1387 PatternNodeWithChain = SDValue(LD, 0);
1388 return true;
1389 }
1390 return false;
1391 }
1392
1393
SelectMOV64Imm32(SDValue N,SDValue & Imm)1394 bool X86DAGToDAGISel::SelectMOV64Imm32(SDValue N, SDValue &Imm) {
1395 if (const ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
1396 uint64_t ImmVal = CN->getZExtValue();
1397 if ((uint32_t)ImmVal != (uint64_t)ImmVal)
1398 return false;
1399
1400 Imm = CurDAG->getTargetConstant(ImmVal, MVT::i64);
1401 return true;
1402 }
1403
1404 // In static codegen with small code model, we can get the address of a label
1405 // into a register with 'movl'. TableGen has already made sure we're looking
1406 // at a label of some kind.
1407 assert(N->getOpcode() == X86ISD::Wrapper &&
1408 "Unexpected node type for MOV32ri64");
1409 N = N.getOperand(0);
1410
1411 if (N->getOpcode() != ISD::TargetConstantPool &&
1412 N->getOpcode() != ISD::TargetJumpTable &&
1413 N->getOpcode() != ISD::TargetGlobalAddress &&
1414 N->getOpcode() != ISD::TargetExternalSymbol &&
1415 N->getOpcode() != ISD::TargetBlockAddress)
1416 return false;
1417
1418 Imm = N;
1419 return TM.getCodeModel() == CodeModel::Small;
1420 }
1421
SelectLEA64_32Addr(SDValue N,SDValue & Base,SDValue & Scale,SDValue & Index,SDValue & Disp,SDValue & Segment)1422 bool X86DAGToDAGISel::SelectLEA64_32Addr(SDValue N, SDValue &Base,
1423 SDValue &Scale, SDValue &Index,
1424 SDValue &Disp, SDValue &Segment) {
1425 if (!SelectLEAAddr(N, Base, Scale, Index, Disp, Segment))
1426 return false;
1427
1428 SDLoc DL(N);
1429 RegisterSDNode *RN = dyn_cast<RegisterSDNode>(Base);
1430 if (RN && RN->getReg() == 0)
1431 Base = CurDAG->getRegister(0, MVT::i64);
1432 else if (Base.getValueType() == MVT::i32 && !dyn_cast<FrameIndexSDNode>(N)) {
1433 // Base could already be %rip, particularly in the x32 ABI.
1434 Base = SDValue(CurDAG->getMachineNode(
1435 TargetOpcode::SUBREG_TO_REG, DL, MVT::i64,
1436 CurDAG->getTargetConstant(0, MVT::i64),
1437 Base,
1438 CurDAG->getTargetConstant(X86::sub_32bit, MVT::i32)),
1439 0);
1440 }
1441
1442 RN = dyn_cast<RegisterSDNode>(Index);
1443 if (RN && RN->getReg() == 0)
1444 Index = CurDAG->getRegister(0, MVT::i64);
1445 else {
1446 assert(Index.getValueType() == MVT::i32 &&
1447 "Expect to be extending 32-bit registers for use in LEA");
1448 Index = SDValue(CurDAG->getMachineNode(
1449 TargetOpcode::SUBREG_TO_REG, DL, MVT::i64,
1450 CurDAG->getTargetConstant(0, MVT::i64),
1451 Index,
1452 CurDAG->getTargetConstant(X86::sub_32bit, MVT::i32)),
1453 0);
1454 }
1455
1456 return true;
1457 }
1458
1459 /// SelectLEAAddr - it calls SelectAddr and determines if the maximal addressing
1460 /// mode it matches can be cost effectively emitted as an LEA instruction.
SelectLEAAddr(SDValue N,SDValue & Base,SDValue & Scale,SDValue & Index,SDValue & Disp,SDValue & Segment)1461 bool X86DAGToDAGISel::SelectLEAAddr(SDValue N,
1462 SDValue &Base, SDValue &Scale,
1463 SDValue &Index, SDValue &Disp,
1464 SDValue &Segment) {
1465 X86ISelAddressMode AM;
1466
1467 // Set AM.Segment to prevent MatchAddress from using one. LEA doesn't support
1468 // segments.
1469 SDValue Copy = AM.Segment;
1470 SDValue T = CurDAG->getRegister(0, MVT::i32);
1471 AM.Segment = T;
1472 if (MatchAddress(N, AM))
1473 return false;
1474 assert (T == AM.Segment);
1475 AM.Segment = Copy;
1476
1477 MVT VT = N.getSimpleValueType();
1478 unsigned Complexity = 0;
1479 if (AM.BaseType == X86ISelAddressMode::RegBase)
1480 if (AM.Base_Reg.getNode())
1481 Complexity = 1;
1482 else
1483 AM.Base_Reg = CurDAG->getRegister(0, VT);
1484 else if (AM.BaseType == X86ISelAddressMode::FrameIndexBase)
1485 Complexity = 4;
1486
1487 if (AM.IndexReg.getNode())
1488 Complexity++;
1489 else
1490 AM.IndexReg = CurDAG->getRegister(0, VT);
1491
1492 // Don't match just leal(,%reg,2). It's cheaper to do addl %reg, %reg, or with
1493 // a simple shift.
1494 if (AM.Scale > 1)
1495 Complexity++;
1496
1497 // FIXME: We are artificially lowering the criteria to turn ADD %reg, $GA
1498 // to a LEA. This is determined with some expermentation but is by no means
1499 // optimal (especially for code size consideration). LEA is nice because of
1500 // its three-address nature. Tweak the cost function again when we can run
1501 // convertToThreeAddress() at register allocation time.
1502 if (AM.hasSymbolicDisplacement()) {
1503 // For X86-64, we should always use lea to materialize RIP relative
1504 // addresses.
1505 if (Subtarget->is64Bit())
1506 Complexity = 4;
1507 else
1508 Complexity += 2;
1509 }
1510
1511 if (AM.Disp && (AM.Base_Reg.getNode() || AM.IndexReg.getNode()))
1512 Complexity++;
1513
1514 // If it isn't worth using an LEA, reject it.
1515 if (Complexity <= 2)
1516 return false;
1517
1518 getAddressOperands(AM, Base, Scale, Index, Disp, Segment);
1519 return true;
1520 }
1521
1522 /// SelectTLSADDRAddr - This is only run on TargetGlobalTLSAddress nodes.
SelectTLSADDRAddr(SDValue N,SDValue & Base,SDValue & Scale,SDValue & Index,SDValue & Disp,SDValue & Segment)1523 bool X86DAGToDAGISel::SelectTLSADDRAddr(SDValue N, SDValue &Base,
1524 SDValue &Scale, SDValue &Index,
1525 SDValue &Disp, SDValue &Segment) {
1526 assert(N.getOpcode() == ISD::TargetGlobalTLSAddress);
1527 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(N);
1528
1529 X86ISelAddressMode AM;
1530 AM.GV = GA->getGlobal();
1531 AM.Disp += GA->getOffset();
1532 AM.Base_Reg = CurDAG->getRegister(0, N.getValueType());
1533 AM.SymbolFlags = GA->getTargetFlags();
1534
1535 if (N.getValueType() == MVT::i32) {
1536 AM.Scale = 1;
1537 AM.IndexReg = CurDAG->getRegister(X86::EBX, MVT::i32);
1538 } else {
1539 AM.IndexReg = CurDAG->getRegister(0, MVT::i64);
1540 }
1541
1542 getAddressOperands(AM, Base, Scale, Index, Disp, Segment);
1543 return true;
1544 }
1545
1546
TryFoldLoad(SDNode * P,SDValue N,SDValue & Base,SDValue & Scale,SDValue & Index,SDValue & Disp,SDValue & Segment)1547 bool X86DAGToDAGISel::TryFoldLoad(SDNode *P, SDValue N,
1548 SDValue &Base, SDValue &Scale,
1549 SDValue &Index, SDValue &Disp,
1550 SDValue &Segment) {
1551 if (!ISD::isNON_EXTLoad(N.getNode()) ||
1552 !IsProfitableToFold(N, P, P) ||
1553 !IsLegalToFold(N, P, P, OptLevel))
1554 return false;
1555
1556 return SelectAddr(N.getNode(),
1557 N.getOperand(1), Base, Scale, Index, Disp, Segment);
1558 }
1559
1560 /// getGlobalBaseReg - Return an SDNode that returns the value of
1561 /// the global base register. Output instructions required to
1562 /// initialize the global base register, if necessary.
1563 ///
getGlobalBaseReg()1564 SDNode *X86DAGToDAGISel::getGlobalBaseReg() {
1565 unsigned GlobalBaseReg = getInstrInfo()->getGlobalBaseReg(MF);
1566 return CurDAG->getRegister(GlobalBaseReg,
1567 getTargetLowering()->getPointerTy()).getNode();
1568 }
1569
SelectAtomic64(SDNode * Node,unsigned Opc)1570 SDNode *X86DAGToDAGISel::SelectAtomic64(SDNode *Node, unsigned Opc) {
1571 SDValue Chain = Node->getOperand(0);
1572 SDValue In1 = Node->getOperand(1);
1573 SDValue In2L = Node->getOperand(2);
1574 SDValue In2H = Node->getOperand(3);
1575
1576 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1577 if (!SelectAddr(Node, In1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4))
1578 return nullptr;
1579 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
1580 MemOp[0] = cast<MemSDNode>(Node)->getMemOperand();
1581 const SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, In2L, In2H, Chain};
1582 SDNode *ResNode = CurDAG->getMachineNode(Opc, SDLoc(Node),
1583 MVT::i32, MVT::i32, MVT::Other, Ops);
1584 cast<MachineSDNode>(ResNode)->setMemRefs(MemOp, MemOp + 1);
1585 return ResNode;
1586 }
1587
1588 /// Atomic opcode table
1589 ///
1590 enum AtomicOpc {
1591 ADD,
1592 SUB,
1593 INC,
1594 DEC,
1595 OR,
1596 AND,
1597 XOR,
1598 AtomicOpcEnd
1599 };
1600
1601 enum AtomicSz {
1602 ConstantI8,
1603 I8,
1604 SextConstantI16,
1605 ConstantI16,
1606 I16,
1607 SextConstantI32,
1608 ConstantI32,
1609 I32,
1610 SextConstantI64,
1611 ConstantI64,
1612 I64,
1613 AtomicSzEnd
1614 };
1615
1616 static const uint16_t AtomicOpcTbl[AtomicOpcEnd][AtomicSzEnd] = {
1617 {
1618 X86::LOCK_ADD8mi,
1619 X86::LOCK_ADD8mr,
1620 X86::LOCK_ADD16mi8,
1621 X86::LOCK_ADD16mi,
1622 X86::LOCK_ADD16mr,
1623 X86::LOCK_ADD32mi8,
1624 X86::LOCK_ADD32mi,
1625 X86::LOCK_ADD32mr,
1626 X86::LOCK_ADD64mi8,
1627 X86::LOCK_ADD64mi32,
1628 X86::LOCK_ADD64mr,
1629 },
1630 {
1631 X86::LOCK_SUB8mi,
1632 X86::LOCK_SUB8mr,
1633 X86::LOCK_SUB16mi8,
1634 X86::LOCK_SUB16mi,
1635 X86::LOCK_SUB16mr,
1636 X86::LOCK_SUB32mi8,
1637 X86::LOCK_SUB32mi,
1638 X86::LOCK_SUB32mr,
1639 X86::LOCK_SUB64mi8,
1640 X86::LOCK_SUB64mi32,
1641 X86::LOCK_SUB64mr,
1642 },
1643 {
1644 0,
1645 X86::LOCK_INC8m,
1646 0,
1647 0,
1648 X86::LOCK_INC16m,
1649 0,
1650 0,
1651 X86::LOCK_INC32m,
1652 0,
1653 0,
1654 X86::LOCK_INC64m,
1655 },
1656 {
1657 0,
1658 X86::LOCK_DEC8m,
1659 0,
1660 0,
1661 X86::LOCK_DEC16m,
1662 0,
1663 0,
1664 X86::LOCK_DEC32m,
1665 0,
1666 0,
1667 X86::LOCK_DEC64m,
1668 },
1669 {
1670 X86::LOCK_OR8mi,
1671 X86::LOCK_OR8mr,
1672 X86::LOCK_OR16mi8,
1673 X86::LOCK_OR16mi,
1674 X86::LOCK_OR16mr,
1675 X86::LOCK_OR32mi8,
1676 X86::LOCK_OR32mi,
1677 X86::LOCK_OR32mr,
1678 X86::LOCK_OR64mi8,
1679 X86::LOCK_OR64mi32,
1680 X86::LOCK_OR64mr,
1681 },
1682 {
1683 X86::LOCK_AND8mi,
1684 X86::LOCK_AND8mr,
1685 X86::LOCK_AND16mi8,
1686 X86::LOCK_AND16mi,
1687 X86::LOCK_AND16mr,
1688 X86::LOCK_AND32mi8,
1689 X86::LOCK_AND32mi,
1690 X86::LOCK_AND32mr,
1691 X86::LOCK_AND64mi8,
1692 X86::LOCK_AND64mi32,
1693 X86::LOCK_AND64mr,
1694 },
1695 {
1696 X86::LOCK_XOR8mi,
1697 X86::LOCK_XOR8mr,
1698 X86::LOCK_XOR16mi8,
1699 X86::LOCK_XOR16mi,
1700 X86::LOCK_XOR16mr,
1701 X86::LOCK_XOR32mi8,
1702 X86::LOCK_XOR32mi,
1703 X86::LOCK_XOR32mr,
1704 X86::LOCK_XOR64mi8,
1705 X86::LOCK_XOR64mi32,
1706 X86::LOCK_XOR64mr,
1707 }
1708 };
1709
1710 // Return the target constant operand for atomic-load-op and do simple
1711 // translations, such as from atomic-load-add to lock-sub. The return value is
1712 // one of the following 3 cases:
1713 // + target-constant, the operand could be supported as a target constant.
1714 // + empty, the operand is not needed any more with the new op selected.
1715 // + non-empty, otherwise.
getAtomicLoadArithTargetConstant(SelectionDAG * CurDAG,SDLoc dl,enum AtomicOpc & Op,MVT NVT,SDValue Val)1716 static SDValue getAtomicLoadArithTargetConstant(SelectionDAG *CurDAG,
1717 SDLoc dl,
1718 enum AtomicOpc &Op, MVT NVT,
1719 SDValue Val) {
1720 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Val)) {
1721 int64_t CNVal = CN->getSExtValue();
1722 // Quit if not 32-bit imm.
1723 if ((int32_t)CNVal != CNVal)
1724 return Val;
1725 // For atomic-load-add, we could do some optimizations.
1726 if (Op == ADD) {
1727 // Translate to INC/DEC if ADD by 1 or -1.
1728 if ((CNVal == 1) || (CNVal == -1)) {
1729 Op = (CNVal == 1) ? INC : DEC;
1730 // No more constant operand after being translated into INC/DEC.
1731 return SDValue();
1732 }
1733 // Translate to SUB if ADD by negative value.
1734 if (CNVal < 0) {
1735 Op = SUB;
1736 CNVal = -CNVal;
1737 }
1738 }
1739 return CurDAG->getTargetConstant(CNVal, NVT);
1740 }
1741
1742 // If the value operand is single-used, try to optimize it.
1743 if (Op == ADD && Val.hasOneUse()) {
1744 // Translate (atomic-load-add ptr (sub 0 x)) back to (lock-sub x).
1745 if (Val.getOpcode() == ISD::SUB && X86::isZeroNode(Val.getOperand(0))) {
1746 Op = SUB;
1747 return Val.getOperand(1);
1748 }
1749 // A special case for i16, which needs truncating as, in most cases, it's
1750 // promoted to i32. We will translate
1751 // (atomic-load-add (truncate (sub 0 x))) to (lock-sub (EXTRACT_SUBREG x))
1752 if (Val.getOpcode() == ISD::TRUNCATE && NVT == MVT::i16 &&
1753 Val.getOperand(0).getOpcode() == ISD::SUB &&
1754 X86::isZeroNode(Val.getOperand(0).getOperand(0))) {
1755 Op = SUB;
1756 Val = Val.getOperand(0);
1757 return CurDAG->getTargetExtractSubreg(X86::sub_16bit, dl, NVT,
1758 Val.getOperand(1));
1759 }
1760 }
1761
1762 return Val;
1763 }
1764
SelectAtomicLoadArith(SDNode * Node,MVT NVT)1765 SDNode *X86DAGToDAGISel::SelectAtomicLoadArith(SDNode *Node, MVT NVT) {
1766 if (Node->hasAnyUseOfValue(0))
1767 return nullptr;
1768
1769 SDLoc dl(Node);
1770
1771 // Optimize common patterns for __sync_or_and_fetch and similar arith
1772 // operations where the result is not used. This allows us to use the "lock"
1773 // version of the arithmetic instruction.
1774 SDValue Chain = Node->getOperand(0);
1775 SDValue Ptr = Node->getOperand(1);
1776 SDValue Val = Node->getOperand(2);
1777 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1778 if (!SelectAddr(Node, Ptr, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4))
1779 return nullptr;
1780
1781 // Which index into the table.
1782 enum AtomicOpc Op;
1783 switch (Node->getOpcode()) {
1784 default:
1785 return nullptr;
1786 case ISD::ATOMIC_LOAD_OR:
1787 Op = OR;
1788 break;
1789 case ISD::ATOMIC_LOAD_AND:
1790 Op = AND;
1791 break;
1792 case ISD::ATOMIC_LOAD_XOR:
1793 Op = XOR;
1794 break;
1795 case ISD::ATOMIC_LOAD_ADD:
1796 Op = ADD;
1797 break;
1798 }
1799
1800 Val = getAtomicLoadArithTargetConstant(CurDAG, dl, Op, NVT, Val);
1801 bool isUnOp = !Val.getNode();
1802 bool isCN = Val.getNode() && (Val.getOpcode() == ISD::TargetConstant);
1803
1804 unsigned Opc = 0;
1805 switch (NVT.SimpleTy) {
1806 default: return nullptr;
1807 case MVT::i8:
1808 if (isCN)
1809 Opc = AtomicOpcTbl[Op][ConstantI8];
1810 else
1811 Opc = AtomicOpcTbl[Op][I8];
1812 break;
1813 case MVT::i16:
1814 if (isCN) {
1815 if (immSext8(Val.getNode()))
1816 Opc = AtomicOpcTbl[Op][SextConstantI16];
1817 else
1818 Opc = AtomicOpcTbl[Op][ConstantI16];
1819 } else
1820 Opc = AtomicOpcTbl[Op][I16];
1821 break;
1822 case MVT::i32:
1823 if (isCN) {
1824 if (immSext8(Val.getNode()))
1825 Opc = AtomicOpcTbl[Op][SextConstantI32];
1826 else
1827 Opc = AtomicOpcTbl[Op][ConstantI32];
1828 } else
1829 Opc = AtomicOpcTbl[Op][I32];
1830 break;
1831 case MVT::i64:
1832 Opc = AtomicOpcTbl[Op][I64];
1833 if (isCN) {
1834 if (immSext8(Val.getNode()))
1835 Opc = AtomicOpcTbl[Op][SextConstantI64];
1836 else if (i64immSExt32(Val.getNode()))
1837 Opc = AtomicOpcTbl[Op][ConstantI64];
1838 }
1839 break;
1840 }
1841
1842 assert(Opc != 0 && "Invalid arith lock transform!");
1843
1844 SDValue Ret;
1845 SDValue Undef = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1846 dl, NVT), 0);
1847 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
1848 MemOp[0] = cast<MemSDNode>(Node)->getMemOperand();
1849 if (isUnOp) {
1850 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Chain };
1851 Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops), 0);
1852 } else {
1853 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Val, Chain };
1854 Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops), 0);
1855 }
1856 cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
1857 SDValue RetVals[] = { Undef, Ret };
1858 return CurDAG->getMergeValues(RetVals, dl).getNode();
1859 }
1860
1861 /// HasNoSignedComparisonUses - Test whether the given X86ISD::CMP node has
1862 /// any uses which require the SF or OF bits to be accurate.
HasNoSignedComparisonUses(SDNode * N)1863 static bool HasNoSignedComparisonUses(SDNode *N) {
1864 // Examine each user of the node.
1865 for (SDNode::use_iterator UI = N->use_begin(),
1866 UE = N->use_end(); UI != UE; ++UI) {
1867 // Only examine CopyToReg uses.
1868 if (UI->getOpcode() != ISD::CopyToReg)
1869 return false;
1870 // Only examine CopyToReg uses that copy to EFLAGS.
1871 if (cast<RegisterSDNode>(UI->getOperand(1))->getReg() !=
1872 X86::EFLAGS)
1873 return false;
1874 // Examine each user of the CopyToReg use.
1875 for (SDNode::use_iterator FlagUI = UI->use_begin(),
1876 FlagUE = UI->use_end(); FlagUI != FlagUE; ++FlagUI) {
1877 // Only examine the Flag result.
1878 if (FlagUI.getUse().getResNo() != 1) continue;
1879 // Anything unusual: assume conservatively.
1880 if (!FlagUI->isMachineOpcode()) return false;
1881 // Examine the opcode of the user.
1882 switch (FlagUI->getMachineOpcode()) {
1883 // These comparisons don't treat the most significant bit specially.
1884 case X86::SETAr: case X86::SETAEr: case X86::SETBr: case X86::SETBEr:
1885 case X86::SETEr: case X86::SETNEr: case X86::SETPr: case X86::SETNPr:
1886 case X86::SETAm: case X86::SETAEm: case X86::SETBm: case X86::SETBEm:
1887 case X86::SETEm: case X86::SETNEm: case X86::SETPm: case X86::SETNPm:
1888 case X86::JA_4: case X86::JAE_4: case X86::JB_4: case X86::JBE_4:
1889 case X86::JE_4: case X86::JNE_4: case X86::JP_4: case X86::JNP_4:
1890 case X86::CMOVA16rr: case X86::CMOVA16rm:
1891 case X86::CMOVA32rr: case X86::CMOVA32rm:
1892 case X86::CMOVA64rr: case X86::CMOVA64rm:
1893 case X86::CMOVAE16rr: case X86::CMOVAE16rm:
1894 case X86::CMOVAE32rr: case X86::CMOVAE32rm:
1895 case X86::CMOVAE64rr: case X86::CMOVAE64rm:
1896 case X86::CMOVB16rr: case X86::CMOVB16rm:
1897 case X86::CMOVB32rr: case X86::CMOVB32rm:
1898 case X86::CMOVB64rr: case X86::CMOVB64rm:
1899 case X86::CMOVBE16rr: case X86::CMOVBE16rm:
1900 case X86::CMOVBE32rr: case X86::CMOVBE32rm:
1901 case X86::CMOVBE64rr: case X86::CMOVBE64rm:
1902 case X86::CMOVE16rr: case X86::CMOVE16rm:
1903 case X86::CMOVE32rr: case X86::CMOVE32rm:
1904 case X86::CMOVE64rr: case X86::CMOVE64rm:
1905 case X86::CMOVNE16rr: case X86::CMOVNE16rm:
1906 case X86::CMOVNE32rr: case X86::CMOVNE32rm:
1907 case X86::CMOVNE64rr: case X86::CMOVNE64rm:
1908 case X86::CMOVNP16rr: case X86::CMOVNP16rm:
1909 case X86::CMOVNP32rr: case X86::CMOVNP32rm:
1910 case X86::CMOVNP64rr: case X86::CMOVNP64rm:
1911 case X86::CMOVP16rr: case X86::CMOVP16rm:
1912 case X86::CMOVP32rr: case X86::CMOVP32rm:
1913 case X86::CMOVP64rr: case X86::CMOVP64rm:
1914 continue;
1915 // Anything else: assume conservatively.
1916 default: return false;
1917 }
1918 }
1919 }
1920 return true;
1921 }
1922
1923 /// isLoadIncOrDecStore - Check whether or not the chain ending in StoreNode
1924 /// is suitable for doing the {load; increment or decrement; store} to modify
1925 /// transformation.
isLoadIncOrDecStore(StoreSDNode * StoreNode,unsigned Opc,SDValue StoredVal,SelectionDAG * CurDAG,LoadSDNode * & LoadNode,SDValue & InputChain)1926 static bool isLoadIncOrDecStore(StoreSDNode *StoreNode, unsigned Opc,
1927 SDValue StoredVal, SelectionDAG *CurDAG,
1928 LoadSDNode* &LoadNode, SDValue &InputChain) {
1929
1930 // is the value stored the result of a DEC or INC?
1931 if (!(Opc == X86ISD::DEC || Opc == X86ISD::INC)) return false;
1932
1933 // is the stored value result 0 of the load?
1934 if (StoredVal.getResNo() != 0) return false;
1935
1936 // are there other uses of the loaded value than the inc or dec?
1937 if (!StoredVal.getNode()->hasNUsesOfValue(1, 0)) return false;
1938
1939 // is the store non-extending and non-indexed?
1940 if (!ISD::isNormalStore(StoreNode) || StoreNode->isNonTemporal())
1941 return false;
1942
1943 SDValue Load = StoredVal->getOperand(0);
1944 // Is the stored value a non-extending and non-indexed load?
1945 if (!ISD::isNormalLoad(Load.getNode())) return false;
1946
1947 // Return LoadNode by reference.
1948 LoadNode = cast<LoadSDNode>(Load);
1949 // is the size of the value one that we can handle? (i.e. 64, 32, 16, or 8)
1950 EVT LdVT = LoadNode->getMemoryVT();
1951 if (LdVT != MVT::i64 && LdVT != MVT::i32 && LdVT != MVT::i16 &&
1952 LdVT != MVT::i8)
1953 return false;
1954
1955 // Is store the only read of the loaded value?
1956 if (!Load.hasOneUse())
1957 return false;
1958
1959 // Is the address of the store the same as the load?
1960 if (LoadNode->getBasePtr() != StoreNode->getBasePtr() ||
1961 LoadNode->getOffset() != StoreNode->getOffset())
1962 return false;
1963
1964 // Check if the chain is produced by the load or is a TokenFactor with
1965 // the load output chain as an operand. Return InputChain by reference.
1966 SDValue Chain = StoreNode->getChain();
1967
1968 bool ChainCheck = false;
1969 if (Chain == Load.getValue(1)) {
1970 ChainCheck = true;
1971 InputChain = LoadNode->getChain();
1972 } else if (Chain.getOpcode() == ISD::TokenFactor) {
1973 SmallVector<SDValue, 4> ChainOps;
1974 for (unsigned i = 0, e = Chain.getNumOperands(); i != e; ++i) {
1975 SDValue Op = Chain.getOperand(i);
1976 if (Op == Load.getValue(1)) {
1977 ChainCheck = true;
1978 continue;
1979 }
1980
1981 // Make sure using Op as part of the chain would not cause a cycle here.
1982 // In theory, we could check whether the chain node is a predecessor of
1983 // the load. But that can be very expensive. Instead visit the uses and
1984 // make sure they all have smaller node id than the load.
1985 int LoadId = LoadNode->getNodeId();
1986 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
1987 UE = UI->use_end(); UI != UE; ++UI) {
1988 if (UI.getUse().getResNo() != 0)
1989 continue;
1990 if (UI->getNodeId() > LoadId)
1991 return false;
1992 }
1993
1994 ChainOps.push_back(Op);
1995 }
1996
1997 if (ChainCheck)
1998 // Make a new TokenFactor with all the other input chains except
1999 // for the load.
2000 InputChain = CurDAG->getNode(ISD::TokenFactor, SDLoc(Chain),
2001 MVT::Other, ChainOps);
2002 }
2003 if (!ChainCheck)
2004 return false;
2005
2006 return true;
2007 }
2008
2009 /// getFusedLdStOpcode - Get the appropriate X86 opcode for an in memory
2010 /// increment or decrement. Opc should be X86ISD::DEC or X86ISD::INC.
getFusedLdStOpcode(EVT & LdVT,unsigned Opc)2011 static unsigned getFusedLdStOpcode(EVT &LdVT, unsigned Opc) {
2012 if (Opc == X86ISD::DEC) {
2013 if (LdVT == MVT::i64) return X86::DEC64m;
2014 if (LdVT == MVT::i32) return X86::DEC32m;
2015 if (LdVT == MVT::i16) return X86::DEC16m;
2016 if (LdVT == MVT::i8) return X86::DEC8m;
2017 } else {
2018 assert(Opc == X86ISD::INC && "unrecognized opcode");
2019 if (LdVT == MVT::i64) return X86::INC64m;
2020 if (LdVT == MVT::i32) return X86::INC32m;
2021 if (LdVT == MVT::i16) return X86::INC16m;
2022 if (LdVT == MVT::i8) return X86::INC8m;
2023 }
2024 llvm_unreachable("unrecognized size for LdVT");
2025 }
2026
2027 /// SelectGather - Customized ISel for GATHER operations.
2028 ///
SelectGather(SDNode * Node,unsigned Opc)2029 SDNode *X86DAGToDAGISel::SelectGather(SDNode *Node, unsigned Opc) {
2030 // Operands of Gather: VSrc, Base, VIdx, VMask, Scale
2031 SDValue Chain = Node->getOperand(0);
2032 SDValue VSrc = Node->getOperand(2);
2033 SDValue Base = Node->getOperand(3);
2034 SDValue VIdx = Node->getOperand(4);
2035 SDValue VMask = Node->getOperand(5);
2036 ConstantSDNode *Scale = dyn_cast<ConstantSDNode>(Node->getOperand(6));
2037 if (!Scale)
2038 return nullptr;
2039
2040 SDVTList VTs = CurDAG->getVTList(VSrc.getValueType(), VSrc.getValueType(),
2041 MVT::Other);
2042
2043 // Memory Operands: Base, Scale, Index, Disp, Segment
2044 SDValue Disp = CurDAG->getTargetConstant(0, MVT::i32);
2045 SDValue Segment = CurDAG->getRegister(0, MVT::i32);
2046 const SDValue Ops[] = { VSrc, Base, getI8Imm(Scale->getSExtValue()), VIdx,
2047 Disp, Segment, VMask, Chain};
2048 SDNode *ResNode = CurDAG->getMachineNode(Opc, SDLoc(Node), VTs, Ops);
2049 // Node has 2 outputs: VDst and MVT::Other.
2050 // ResNode has 3 outputs: VDst, VMask_wb, and MVT::Other.
2051 // We replace VDst of Node with VDst of ResNode, and Other of Node with Other
2052 // of ResNode.
2053 ReplaceUses(SDValue(Node, 0), SDValue(ResNode, 0));
2054 ReplaceUses(SDValue(Node, 1), SDValue(ResNode, 2));
2055 return ResNode;
2056 }
2057
Select(SDNode * Node)2058 SDNode *X86DAGToDAGISel::Select(SDNode *Node) {
2059 MVT NVT = Node->getSimpleValueType(0);
2060 unsigned Opc, MOpc;
2061 unsigned Opcode = Node->getOpcode();
2062 SDLoc dl(Node);
2063
2064 DEBUG(dbgs() << "Selecting: "; Node->dump(CurDAG); dbgs() << '\n');
2065
2066 if (Node->isMachineOpcode()) {
2067 DEBUG(dbgs() << "== "; Node->dump(CurDAG); dbgs() << '\n');
2068 Node->setNodeId(-1);
2069 return nullptr; // Already selected.
2070 }
2071
2072 switch (Opcode) {
2073 default: break;
2074 case ISD::INTRINSIC_W_CHAIN: {
2075 unsigned IntNo = cast<ConstantSDNode>(Node->getOperand(1))->getZExtValue();
2076 switch (IntNo) {
2077 default: break;
2078 case Intrinsic::x86_avx2_gather_d_pd:
2079 case Intrinsic::x86_avx2_gather_d_pd_256:
2080 case Intrinsic::x86_avx2_gather_q_pd:
2081 case Intrinsic::x86_avx2_gather_q_pd_256:
2082 case Intrinsic::x86_avx2_gather_d_ps:
2083 case Intrinsic::x86_avx2_gather_d_ps_256:
2084 case Intrinsic::x86_avx2_gather_q_ps:
2085 case Intrinsic::x86_avx2_gather_q_ps_256:
2086 case Intrinsic::x86_avx2_gather_d_q:
2087 case Intrinsic::x86_avx2_gather_d_q_256:
2088 case Intrinsic::x86_avx2_gather_q_q:
2089 case Intrinsic::x86_avx2_gather_q_q_256:
2090 case Intrinsic::x86_avx2_gather_d_d:
2091 case Intrinsic::x86_avx2_gather_d_d_256:
2092 case Intrinsic::x86_avx2_gather_q_d:
2093 case Intrinsic::x86_avx2_gather_q_d_256: {
2094 if (!Subtarget->hasAVX2())
2095 break;
2096 unsigned Opc;
2097 switch (IntNo) {
2098 default: llvm_unreachable("Impossible intrinsic");
2099 case Intrinsic::x86_avx2_gather_d_pd: Opc = X86::VGATHERDPDrm; break;
2100 case Intrinsic::x86_avx2_gather_d_pd_256: Opc = X86::VGATHERDPDYrm; break;
2101 case Intrinsic::x86_avx2_gather_q_pd: Opc = X86::VGATHERQPDrm; break;
2102 case Intrinsic::x86_avx2_gather_q_pd_256: Opc = X86::VGATHERQPDYrm; break;
2103 case Intrinsic::x86_avx2_gather_d_ps: Opc = X86::VGATHERDPSrm; break;
2104 case Intrinsic::x86_avx2_gather_d_ps_256: Opc = X86::VGATHERDPSYrm; break;
2105 case Intrinsic::x86_avx2_gather_q_ps: Opc = X86::VGATHERQPSrm; break;
2106 case Intrinsic::x86_avx2_gather_q_ps_256: Opc = X86::VGATHERQPSYrm; break;
2107 case Intrinsic::x86_avx2_gather_d_q: Opc = X86::VPGATHERDQrm; break;
2108 case Intrinsic::x86_avx2_gather_d_q_256: Opc = X86::VPGATHERDQYrm; break;
2109 case Intrinsic::x86_avx2_gather_q_q: Opc = X86::VPGATHERQQrm; break;
2110 case Intrinsic::x86_avx2_gather_q_q_256: Opc = X86::VPGATHERQQYrm; break;
2111 case Intrinsic::x86_avx2_gather_d_d: Opc = X86::VPGATHERDDrm; break;
2112 case Intrinsic::x86_avx2_gather_d_d_256: Opc = X86::VPGATHERDDYrm; break;
2113 case Intrinsic::x86_avx2_gather_q_d: Opc = X86::VPGATHERQDrm; break;
2114 case Intrinsic::x86_avx2_gather_q_d_256: Opc = X86::VPGATHERQDYrm; break;
2115 }
2116 SDNode *RetVal = SelectGather(Node, Opc);
2117 if (RetVal)
2118 // We already called ReplaceUses inside SelectGather.
2119 return nullptr;
2120 break;
2121 }
2122 }
2123 break;
2124 }
2125 case X86ISD::GlobalBaseReg:
2126 return getGlobalBaseReg();
2127
2128
2129 case ISD::ATOMIC_LOAD_XOR:
2130 case ISD::ATOMIC_LOAD_AND:
2131 case ISD::ATOMIC_LOAD_OR:
2132 case ISD::ATOMIC_LOAD_ADD: {
2133 SDNode *RetVal = SelectAtomicLoadArith(Node, NVT);
2134 if (RetVal)
2135 return RetVal;
2136 break;
2137 }
2138 case ISD::AND:
2139 case ISD::OR:
2140 case ISD::XOR: {
2141 // For operations of the form (x << C1) op C2, check if we can use a smaller
2142 // encoding for C2 by transforming it into (x op (C2>>C1)) << C1.
2143 SDValue N0 = Node->getOperand(0);
2144 SDValue N1 = Node->getOperand(1);
2145
2146 if (N0->getOpcode() != ISD::SHL || !N0->hasOneUse())
2147 break;
2148
2149 // i8 is unshrinkable, i16 should be promoted to i32.
2150 if (NVT != MVT::i32 && NVT != MVT::i64)
2151 break;
2152
2153 ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(N1);
2154 ConstantSDNode *ShlCst = dyn_cast<ConstantSDNode>(N0->getOperand(1));
2155 if (!Cst || !ShlCst)
2156 break;
2157
2158 int64_t Val = Cst->getSExtValue();
2159 uint64_t ShlVal = ShlCst->getZExtValue();
2160
2161 // Make sure that we don't change the operation by removing bits.
2162 // This only matters for OR and XOR, AND is unaffected.
2163 uint64_t RemovedBitsMask = (1ULL << ShlVal) - 1;
2164 if (Opcode != ISD::AND && (Val & RemovedBitsMask) != 0)
2165 break;
2166
2167 unsigned ShlOp, Op;
2168 MVT CstVT = NVT;
2169
2170 // Check the minimum bitwidth for the new constant.
2171 // TODO: AND32ri is the same as AND64ri32 with zext imm.
2172 // TODO: MOV32ri+OR64r is cheaper than MOV64ri64+OR64rr
2173 // TODO: Using 16 and 8 bit operations is also possible for or32 & xor32.
2174 if (!isInt<8>(Val) && isInt<8>(Val >> ShlVal))
2175 CstVT = MVT::i8;
2176 else if (!isInt<32>(Val) && isInt<32>(Val >> ShlVal))
2177 CstVT = MVT::i32;
2178
2179 // Bail if there is no smaller encoding.
2180 if (NVT == CstVT)
2181 break;
2182
2183 switch (NVT.SimpleTy) {
2184 default: llvm_unreachable("Unsupported VT!");
2185 case MVT::i32:
2186 assert(CstVT == MVT::i8);
2187 ShlOp = X86::SHL32ri;
2188
2189 switch (Opcode) {
2190 default: llvm_unreachable("Impossible opcode");
2191 case ISD::AND: Op = X86::AND32ri8; break;
2192 case ISD::OR: Op = X86::OR32ri8; break;
2193 case ISD::XOR: Op = X86::XOR32ri8; break;
2194 }
2195 break;
2196 case MVT::i64:
2197 assert(CstVT == MVT::i8 || CstVT == MVT::i32);
2198 ShlOp = X86::SHL64ri;
2199
2200 switch (Opcode) {
2201 default: llvm_unreachable("Impossible opcode");
2202 case ISD::AND: Op = CstVT==MVT::i8? X86::AND64ri8 : X86::AND64ri32; break;
2203 case ISD::OR: Op = CstVT==MVT::i8? X86::OR64ri8 : X86::OR64ri32; break;
2204 case ISD::XOR: Op = CstVT==MVT::i8? X86::XOR64ri8 : X86::XOR64ri32; break;
2205 }
2206 break;
2207 }
2208
2209 // Emit the smaller op and the shift.
2210 SDValue NewCst = CurDAG->getTargetConstant(Val >> ShlVal, CstVT);
2211 SDNode *New = CurDAG->getMachineNode(Op, dl, NVT, N0->getOperand(0),NewCst);
2212 return CurDAG->SelectNodeTo(Node, ShlOp, NVT, SDValue(New, 0),
2213 getI8Imm(ShlVal));
2214 }
2215 case X86ISD::UMUL: {
2216 SDValue N0 = Node->getOperand(0);
2217 SDValue N1 = Node->getOperand(1);
2218
2219 unsigned LoReg;
2220 switch (NVT.SimpleTy) {
2221 default: llvm_unreachable("Unsupported VT!");
2222 case MVT::i8: LoReg = X86::AL; Opc = X86::MUL8r; break;
2223 case MVT::i16: LoReg = X86::AX; Opc = X86::MUL16r; break;
2224 case MVT::i32: LoReg = X86::EAX; Opc = X86::MUL32r; break;
2225 case MVT::i64: LoReg = X86::RAX; Opc = X86::MUL64r; break;
2226 }
2227
2228 SDValue InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, LoReg,
2229 N0, SDValue()).getValue(1);
2230
2231 SDVTList VTs = CurDAG->getVTList(NVT, NVT, MVT::i32);
2232 SDValue Ops[] = {N1, InFlag};
2233 SDNode *CNode = CurDAG->getMachineNode(Opc, dl, VTs, Ops);
2234
2235 ReplaceUses(SDValue(Node, 0), SDValue(CNode, 0));
2236 ReplaceUses(SDValue(Node, 1), SDValue(CNode, 1));
2237 ReplaceUses(SDValue(Node, 2), SDValue(CNode, 2));
2238 return nullptr;
2239 }
2240
2241 case ISD::SMUL_LOHI:
2242 case ISD::UMUL_LOHI: {
2243 SDValue N0 = Node->getOperand(0);
2244 SDValue N1 = Node->getOperand(1);
2245
2246 bool isSigned = Opcode == ISD::SMUL_LOHI;
2247 bool hasBMI2 = Subtarget->hasBMI2();
2248 if (!isSigned) {
2249 switch (NVT.SimpleTy) {
2250 default: llvm_unreachable("Unsupported VT!");
2251 case MVT::i8: Opc = X86::MUL8r; MOpc = X86::MUL8m; break;
2252 case MVT::i16: Opc = X86::MUL16r; MOpc = X86::MUL16m; break;
2253 case MVT::i32: Opc = hasBMI2 ? X86::MULX32rr : X86::MUL32r;
2254 MOpc = hasBMI2 ? X86::MULX32rm : X86::MUL32m; break;
2255 case MVT::i64: Opc = hasBMI2 ? X86::MULX64rr : X86::MUL64r;
2256 MOpc = hasBMI2 ? X86::MULX64rm : X86::MUL64m; break;
2257 }
2258 } else {
2259 switch (NVT.SimpleTy) {
2260 default: llvm_unreachable("Unsupported VT!");
2261 case MVT::i8: Opc = X86::IMUL8r; MOpc = X86::IMUL8m; break;
2262 case MVT::i16: Opc = X86::IMUL16r; MOpc = X86::IMUL16m; break;
2263 case MVT::i32: Opc = X86::IMUL32r; MOpc = X86::IMUL32m; break;
2264 case MVT::i64: Opc = X86::IMUL64r; MOpc = X86::IMUL64m; break;
2265 }
2266 }
2267
2268 unsigned SrcReg, LoReg, HiReg;
2269 switch (Opc) {
2270 default: llvm_unreachable("Unknown MUL opcode!");
2271 case X86::IMUL8r:
2272 case X86::MUL8r:
2273 SrcReg = LoReg = X86::AL; HiReg = X86::AH;
2274 break;
2275 case X86::IMUL16r:
2276 case X86::MUL16r:
2277 SrcReg = LoReg = X86::AX; HiReg = X86::DX;
2278 break;
2279 case X86::IMUL32r:
2280 case X86::MUL32r:
2281 SrcReg = LoReg = X86::EAX; HiReg = X86::EDX;
2282 break;
2283 case X86::IMUL64r:
2284 case X86::MUL64r:
2285 SrcReg = LoReg = X86::RAX; HiReg = X86::RDX;
2286 break;
2287 case X86::MULX32rr:
2288 SrcReg = X86::EDX; LoReg = HiReg = 0;
2289 break;
2290 case X86::MULX64rr:
2291 SrcReg = X86::RDX; LoReg = HiReg = 0;
2292 break;
2293 }
2294
2295 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
2296 bool foldedLoad = TryFoldLoad(Node, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
2297 // Multiply is commmutative.
2298 if (!foldedLoad) {
2299 foldedLoad = TryFoldLoad(Node, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
2300 if (foldedLoad)
2301 std::swap(N0, N1);
2302 }
2303
2304 SDValue InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, SrcReg,
2305 N0, SDValue()).getValue(1);
2306 SDValue ResHi, ResLo;
2307
2308 if (foldedLoad) {
2309 SDValue Chain;
2310 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
2311 InFlag };
2312 if (MOpc == X86::MULX32rm || MOpc == X86::MULX64rm) {
2313 SDVTList VTs = CurDAG->getVTList(NVT, NVT, MVT::Other, MVT::Glue);
2314 SDNode *CNode = CurDAG->getMachineNode(MOpc, dl, VTs, Ops);
2315 ResHi = SDValue(CNode, 0);
2316 ResLo = SDValue(CNode, 1);
2317 Chain = SDValue(CNode, 2);
2318 InFlag = SDValue(CNode, 3);
2319 } else {
2320 SDVTList VTs = CurDAG->getVTList(MVT::Other, MVT::Glue);
2321 SDNode *CNode = CurDAG->getMachineNode(MOpc, dl, VTs, Ops);
2322 Chain = SDValue(CNode, 0);
2323 InFlag = SDValue(CNode, 1);
2324 }
2325
2326 // Update the chain.
2327 ReplaceUses(N1.getValue(1), Chain);
2328 } else {
2329 SDValue Ops[] = { N1, InFlag };
2330 if (Opc == X86::MULX32rr || Opc == X86::MULX64rr) {
2331 SDVTList VTs = CurDAG->getVTList(NVT, NVT, MVT::Glue);
2332 SDNode *CNode = CurDAG->getMachineNode(Opc, dl, VTs, Ops);
2333 ResHi = SDValue(CNode, 0);
2334 ResLo = SDValue(CNode, 1);
2335 InFlag = SDValue(CNode, 2);
2336 } else {
2337 SDVTList VTs = CurDAG->getVTList(MVT::Glue);
2338 SDNode *CNode = CurDAG->getMachineNode(Opc, dl, VTs, Ops);
2339 InFlag = SDValue(CNode, 0);
2340 }
2341 }
2342
2343 // Prevent use of AH in a REX instruction by referencing AX instead.
2344 if (HiReg == X86::AH && Subtarget->is64Bit() &&
2345 !SDValue(Node, 1).use_empty()) {
2346 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2347 X86::AX, MVT::i16, InFlag);
2348 InFlag = Result.getValue(2);
2349 // Get the low part if needed. Don't use getCopyFromReg for aliasing
2350 // registers.
2351 if (!SDValue(Node, 0).use_empty())
2352 ReplaceUses(SDValue(Node, 1),
2353 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
2354
2355 // Shift AX down 8 bits.
2356 Result = SDValue(CurDAG->getMachineNode(X86::SHR16ri, dl, MVT::i16,
2357 Result,
2358 CurDAG->getTargetConstant(8, MVT::i8)), 0);
2359 // Then truncate it down to i8.
2360 ReplaceUses(SDValue(Node, 1),
2361 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
2362 }
2363 // Copy the low half of the result, if it is needed.
2364 if (!SDValue(Node, 0).use_empty()) {
2365 if (!ResLo.getNode()) {
2366 assert(LoReg && "Register for low half is not defined!");
2367 ResLo = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl, LoReg, NVT,
2368 InFlag);
2369 InFlag = ResLo.getValue(2);
2370 }
2371 ReplaceUses(SDValue(Node, 0), ResLo);
2372 DEBUG(dbgs() << "=> "; ResLo.getNode()->dump(CurDAG); dbgs() << '\n');
2373 }
2374 // Copy the high half of the result, if it is needed.
2375 if (!SDValue(Node, 1).use_empty()) {
2376 if (!ResHi.getNode()) {
2377 assert(HiReg && "Register for high half is not defined!");
2378 ResHi = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl, HiReg, NVT,
2379 InFlag);
2380 InFlag = ResHi.getValue(2);
2381 }
2382 ReplaceUses(SDValue(Node, 1), ResHi);
2383 DEBUG(dbgs() << "=> "; ResHi.getNode()->dump(CurDAG); dbgs() << '\n');
2384 }
2385
2386 return nullptr;
2387 }
2388
2389 case ISD::SDIVREM:
2390 case ISD::UDIVREM: {
2391 SDValue N0 = Node->getOperand(0);
2392 SDValue N1 = Node->getOperand(1);
2393
2394 bool isSigned = Opcode == ISD::SDIVREM;
2395 if (!isSigned) {
2396 switch (NVT.SimpleTy) {
2397 default: llvm_unreachable("Unsupported VT!");
2398 case MVT::i8: Opc = X86::DIV8r; MOpc = X86::DIV8m; break;
2399 case MVT::i16: Opc = X86::DIV16r; MOpc = X86::DIV16m; break;
2400 case MVT::i32: Opc = X86::DIV32r; MOpc = X86::DIV32m; break;
2401 case MVT::i64: Opc = X86::DIV64r; MOpc = X86::DIV64m; break;
2402 }
2403 } else {
2404 switch (NVT.SimpleTy) {
2405 default: llvm_unreachable("Unsupported VT!");
2406 case MVT::i8: Opc = X86::IDIV8r; MOpc = X86::IDIV8m; break;
2407 case MVT::i16: Opc = X86::IDIV16r; MOpc = X86::IDIV16m; break;
2408 case MVT::i32: Opc = X86::IDIV32r; MOpc = X86::IDIV32m; break;
2409 case MVT::i64: Opc = X86::IDIV64r; MOpc = X86::IDIV64m; break;
2410 }
2411 }
2412
2413 unsigned LoReg, HiReg, ClrReg;
2414 unsigned SExtOpcode;
2415 switch (NVT.SimpleTy) {
2416 default: llvm_unreachable("Unsupported VT!");
2417 case MVT::i8:
2418 LoReg = X86::AL; ClrReg = HiReg = X86::AH;
2419 SExtOpcode = X86::CBW;
2420 break;
2421 case MVT::i16:
2422 LoReg = X86::AX; HiReg = X86::DX;
2423 ClrReg = X86::DX;
2424 SExtOpcode = X86::CWD;
2425 break;
2426 case MVT::i32:
2427 LoReg = X86::EAX; ClrReg = HiReg = X86::EDX;
2428 SExtOpcode = X86::CDQ;
2429 break;
2430 case MVT::i64:
2431 LoReg = X86::RAX; ClrReg = HiReg = X86::RDX;
2432 SExtOpcode = X86::CQO;
2433 break;
2434 }
2435
2436 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
2437 bool foldedLoad = TryFoldLoad(Node, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
2438 bool signBitIsZero = CurDAG->SignBitIsZero(N0);
2439
2440 SDValue InFlag;
2441 if (NVT == MVT::i8 && (!isSigned || signBitIsZero)) {
2442 // Special case for div8, just use a move with zero extension to AX to
2443 // clear the upper 8 bits (AH).
2444 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Move, Chain;
2445 if (TryFoldLoad(Node, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4)) {
2446 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N0.getOperand(0) };
2447 Move =
2448 SDValue(CurDAG->getMachineNode(X86::MOVZX32rm8, dl, MVT::i32,
2449 MVT::Other, Ops), 0);
2450 Chain = Move.getValue(1);
2451 ReplaceUses(N0.getValue(1), Chain);
2452 } else {
2453 Move =
2454 SDValue(CurDAG->getMachineNode(X86::MOVZX32rr8, dl, MVT::i32, N0),0);
2455 Chain = CurDAG->getEntryNode();
2456 }
2457 Chain = CurDAG->getCopyToReg(Chain, dl, X86::EAX, Move, SDValue());
2458 InFlag = Chain.getValue(1);
2459 } else {
2460 InFlag =
2461 CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl,
2462 LoReg, N0, SDValue()).getValue(1);
2463 if (isSigned && !signBitIsZero) {
2464 // Sign extend the low part into the high part.
2465 InFlag =
2466 SDValue(CurDAG->getMachineNode(SExtOpcode, dl, MVT::Glue, InFlag),0);
2467 } else {
2468 // Zero out the high part, effectively zero extending the input.
2469 SDValue ClrNode = SDValue(CurDAG->getMachineNode(X86::MOV32r0, dl, NVT), 0);
2470 switch (NVT.SimpleTy) {
2471 case MVT::i16:
2472 ClrNode =
2473 SDValue(CurDAG->getMachineNode(
2474 TargetOpcode::EXTRACT_SUBREG, dl, MVT::i16, ClrNode,
2475 CurDAG->getTargetConstant(X86::sub_16bit, MVT::i32)),
2476 0);
2477 break;
2478 case MVT::i32:
2479 break;
2480 case MVT::i64:
2481 ClrNode =
2482 SDValue(CurDAG->getMachineNode(
2483 TargetOpcode::SUBREG_TO_REG, dl, MVT::i64,
2484 CurDAG->getTargetConstant(0, MVT::i64), ClrNode,
2485 CurDAG->getTargetConstant(X86::sub_32bit, MVT::i32)),
2486 0);
2487 break;
2488 default:
2489 llvm_unreachable("Unexpected division source");
2490 }
2491
2492 InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, ClrReg,
2493 ClrNode, InFlag).getValue(1);
2494 }
2495 }
2496
2497 if (foldedLoad) {
2498 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
2499 InFlag };
2500 SDNode *CNode =
2501 CurDAG->getMachineNode(MOpc, dl, MVT::Other, MVT::Glue, Ops);
2502 InFlag = SDValue(CNode, 1);
2503 // Update the chain.
2504 ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
2505 } else {
2506 InFlag =
2507 SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Glue, N1, InFlag), 0);
2508 }
2509
2510 // Prevent use of AH in a REX instruction by referencing AX instead.
2511 // Shift it down 8 bits.
2512 //
2513 // The current assumption of the register allocator is that isel
2514 // won't generate explicit references to the GPR8_NOREX registers. If
2515 // the allocator and/or the backend get enhanced to be more robust in
2516 // that regard, this can be, and should be, removed.
2517 if (HiReg == X86::AH && Subtarget->is64Bit() &&
2518 !SDValue(Node, 1).use_empty()) {
2519 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2520 X86::AX, MVT::i16, InFlag);
2521 InFlag = Result.getValue(2);
2522
2523 // If we also need AL (the quotient), get it by extracting a subreg from
2524 // Result. The fast register allocator does not like multiple CopyFromReg
2525 // nodes using aliasing registers.
2526 if (!SDValue(Node, 0).use_empty())
2527 ReplaceUses(SDValue(Node, 0),
2528 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
2529
2530 // Shift AX right by 8 bits instead of using AH.
2531 Result = SDValue(CurDAG->getMachineNode(X86::SHR16ri, dl, MVT::i16,
2532 Result,
2533 CurDAG->getTargetConstant(8, MVT::i8)),
2534 0);
2535 ReplaceUses(SDValue(Node, 1),
2536 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
2537 }
2538 // Copy the division (low) result, if it is needed.
2539 if (!SDValue(Node, 0).use_empty()) {
2540 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2541 LoReg, NVT, InFlag);
2542 InFlag = Result.getValue(2);
2543 ReplaceUses(SDValue(Node, 0), Result);
2544 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
2545 }
2546 // Copy the remainder (high) result, if it is needed.
2547 if (!SDValue(Node, 1).use_empty()) {
2548 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
2549 HiReg, NVT, InFlag);
2550 InFlag = Result.getValue(2);
2551 ReplaceUses(SDValue(Node, 1), Result);
2552 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
2553 }
2554 return nullptr;
2555 }
2556
2557 case X86ISD::CMP:
2558 case X86ISD::SUB: {
2559 // Sometimes a SUB is used to perform comparison.
2560 if (Opcode == X86ISD::SUB && Node->hasAnyUseOfValue(0))
2561 // This node is not a CMP.
2562 break;
2563 SDValue N0 = Node->getOperand(0);
2564 SDValue N1 = Node->getOperand(1);
2565
2566 // Look for (X86cmp (and $op, $imm), 0) and see if we can convert it to
2567 // use a smaller encoding.
2568 if (N0.getOpcode() == ISD::TRUNCATE && N0.hasOneUse() &&
2569 HasNoSignedComparisonUses(Node))
2570 // Look past the truncate if CMP is the only use of it.
2571 N0 = N0.getOperand(0);
2572 if ((N0.getNode()->getOpcode() == ISD::AND ||
2573 (N0.getResNo() == 0 && N0.getNode()->getOpcode() == X86ISD::AND)) &&
2574 N0.getNode()->hasOneUse() &&
2575 N0.getValueType() != MVT::i8 &&
2576 X86::isZeroNode(N1)) {
2577 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getNode()->getOperand(1));
2578 if (!C) break;
2579
2580 // For example, convert "testl %eax, $8" to "testb %al, $8"
2581 if ((C->getZExtValue() & ~UINT64_C(0xff)) == 0 &&
2582 (!(C->getZExtValue() & 0x80) ||
2583 HasNoSignedComparisonUses(Node))) {
2584 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), MVT::i8);
2585 SDValue Reg = N0.getNode()->getOperand(0);
2586
2587 // On x86-32, only the ABCD registers have 8-bit subregisters.
2588 if (!Subtarget->is64Bit()) {
2589 const TargetRegisterClass *TRC;
2590 switch (N0.getSimpleValueType().SimpleTy) {
2591 case MVT::i32: TRC = &X86::GR32_ABCDRegClass; break;
2592 case MVT::i16: TRC = &X86::GR16_ABCDRegClass; break;
2593 default: llvm_unreachable("Unsupported TEST operand type!");
2594 }
2595 SDValue RC = CurDAG->getTargetConstant(TRC->getID(), MVT::i32);
2596 Reg = SDValue(CurDAG->getMachineNode(X86::COPY_TO_REGCLASS, dl,
2597 Reg.getValueType(), Reg, RC), 0);
2598 }
2599
2600 // Extract the l-register.
2601 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl,
2602 MVT::i8, Reg);
2603
2604 // Emit a testb.
2605 SDNode *NewNode = CurDAG->getMachineNode(X86::TEST8ri, dl, MVT::i32,
2606 Subreg, Imm);
2607 // Replace SUB|CMP with TEST, since SUB has two outputs while TEST has
2608 // one, do not call ReplaceAllUsesWith.
2609 ReplaceUses(SDValue(Node, (Opcode == X86ISD::SUB ? 1 : 0)),
2610 SDValue(NewNode, 0));
2611 return nullptr;
2612 }
2613
2614 // For example, "testl %eax, $2048" to "testb %ah, $8".
2615 if ((C->getZExtValue() & ~UINT64_C(0xff00)) == 0 &&
2616 (!(C->getZExtValue() & 0x8000) ||
2617 HasNoSignedComparisonUses(Node))) {
2618 // Shift the immediate right by 8 bits.
2619 SDValue ShiftedImm = CurDAG->getTargetConstant(C->getZExtValue() >> 8,
2620 MVT::i8);
2621 SDValue Reg = N0.getNode()->getOperand(0);
2622
2623 // Put the value in an ABCD register.
2624 const TargetRegisterClass *TRC;
2625 switch (N0.getSimpleValueType().SimpleTy) {
2626 case MVT::i64: TRC = &X86::GR64_ABCDRegClass; break;
2627 case MVT::i32: TRC = &X86::GR32_ABCDRegClass; break;
2628 case MVT::i16: TRC = &X86::GR16_ABCDRegClass; break;
2629 default: llvm_unreachable("Unsupported TEST operand type!");
2630 }
2631 SDValue RC = CurDAG->getTargetConstant(TRC->getID(), MVT::i32);
2632 Reg = SDValue(CurDAG->getMachineNode(X86::COPY_TO_REGCLASS, dl,
2633 Reg.getValueType(), Reg, RC), 0);
2634
2635 // Extract the h-register.
2636 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_8bit_hi, dl,
2637 MVT::i8, Reg);
2638
2639 // Emit a testb. The EXTRACT_SUBREG becomes a COPY that can only
2640 // target GR8_NOREX registers, so make sure the register class is
2641 // forced.
2642 SDNode *NewNode = CurDAG->getMachineNode(X86::TEST8ri_NOREX, dl,
2643 MVT::i32, Subreg, ShiftedImm);
2644 // Replace SUB|CMP with TEST, since SUB has two outputs while TEST has
2645 // one, do not call ReplaceAllUsesWith.
2646 ReplaceUses(SDValue(Node, (Opcode == X86ISD::SUB ? 1 : 0)),
2647 SDValue(NewNode, 0));
2648 return nullptr;
2649 }
2650
2651 // For example, "testl %eax, $32776" to "testw %ax, $32776".
2652 if ((C->getZExtValue() & ~UINT64_C(0xffff)) == 0 &&
2653 N0.getValueType() != MVT::i16 &&
2654 (!(C->getZExtValue() & 0x8000) ||
2655 HasNoSignedComparisonUses(Node))) {
2656 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), MVT::i16);
2657 SDValue Reg = N0.getNode()->getOperand(0);
2658
2659 // Extract the 16-bit subregister.
2660 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_16bit, dl,
2661 MVT::i16, Reg);
2662
2663 // Emit a testw.
2664 SDNode *NewNode = CurDAG->getMachineNode(X86::TEST16ri, dl, MVT::i32,
2665 Subreg, Imm);
2666 // Replace SUB|CMP with TEST, since SUB has two outputs while TEST has
2667 // one, do not call ReplaceAllUsesWith.
2668 ReplaceUses(SDValue(Node, (Opcode == X86ISD::SUB ? 1 : 0)),
2669 SDValue(NewNode, 0));
2670 return nullptr;
2671 }
2672
2673 // For example, "testq %rax, $268468232" to "testl %eax, $268468232".
2674 if ((C->getZExtValue() & ~UINT64_C(0xffffffff)) == 0 &&
2675 N0.getValueType() == MVT::i64 &&
2676 (!(C->getZExtValue() & 0x80000000) ||
2677 HasNoSignedComparisonUses(Node))) {
2678 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), MVT::i32);
2679 SDValue Reg = N0.getNode()->getOperand(0);
2680
2681 // Extract the 32-bit subregister.
2682 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_32bit, dl,
2683 MVT::i32, Reg);
2684
2685 // Emit a testl.
2686 SDNode *NewNode = CurDAG->getMachineNode(X86::TEST32ri, dl, MVT::i32,
2687 Subreg, Imm);
2688 // Replace SUB|CMP with TEST, since SUB has two outputs while TEST has
2689 // one, do not call ReplaceAllUsesWith.
2690 ReplaceUses(SDValue(Node, (Opcode == X86ISD::SUB ? 1 : 0)),
2691 SDValue(NewNode, 0));
2692 return nullptr;
2693 }
2694 }
2695 break;
2696 }
2697 case ISD::STORE: {
2698 // Change a chain of {load; incr or dec; store} of the same value into
2699 // a simple increment or decrement through memory of that value, if the
2700 // uses of the modified value and its address are suitable.
2701 // The DEC64m tablegen pattern is currently not able to match the case where
2702 // the EFLAGS on the original DEC are used. (This also applies to
2703 // {INC,DEC}X{64,32,16,8}.)
2704 // We'll need to improve tablegen to allow flags to be transferred from a
2705 // node in the pattern to the result node. probably with a new keyword
2706 // for example, we have this
2707 // def DEC64m : RI<0xFF, MRM1m, (outs), (ins i64mem:$dst), "dec{q}\t$dst",
2708 // [(store (add (loadi64 addr:$dst), -1), addr:$dst),
2709 // (implicit EFLAGS)]>;
2710 // but maybe need something like this
2711 // def DEC64m : RI<0xFF, MRM1m, (outs), (ins i64mem:$dst), "dec{q}\t$dst",
2712 // [(store (add (loadi64 addr:$dst), -1), addr:$dst),
2713 // (transferrable EFLAGS)]>;
2714
2715 StoreSDNode *StoreNode = cast<StoreSDNode>(Node);
2716 SDValue StoredVal = StoreNode->getOperand(1);
2717 unsigned Opc = StoredVal->getOpcode();
2718
2719 LoadSDNode *LoadNode = nullptr;
2720 SDValue InputChain;
2721 if (!isLoadIncOrDecStore(StoreNode, Opc, StoredVal, CurDAG,
2722 LoadNode, InputChain))
2723 break;
2724
2725 SDValue Base, Scale, Index, Disp, Segment;
2726 if (!SelectAddr(LoadNode, LoadNode->getBasePtr(),
2727 Base, Scale, Index, Disp, Segment))
2728 break;
2729
2730 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(2);
2731 MemOp[0] = StoreNode->getMemOperand();
2732 MemOp[1] = LoadNode->getMemOperand();
2733 const SDValue Ops[] = { Base, Scale, Index, Disp, Segment, InputChain };
2734 EVT LdVT = LoadNode->getMemoryVT();
2735 unsigned newOpc = getFusedLdStOpcode(LdVT, Opc);
2736 MachineSDNode *Result = CurDAG->getMachineNode(newOpc,
2737 SDLoc(Node),
2738 MVT::i32, MVT::Other, Ops);
2739 Result->setMemRefs(MemOp, MemOp + 2);
2740
2741 ReplaceUses(SDValue(StoreNode, 0), SDValue(Result, 1));
2742 ReplaceUses(SDValue(StoredVal.getNode(), 1), SDValue(Result, 0));
2743
2744 return Result;
2745 }
2746 }
2747
2748 SDNode *ResNode = SelectCode(Node);
2749
2750 DEBUG(dbgs() << "=> ";
2751 if (ResNode == nullptr || ResNode == Node)
2752 Node->dump(CurDAG);
2753 else
2754 ResNode->dump(CurDAG);
2755 dbgs() << '\n');
2756
2757 return ResNode;
2758 }
2759
2760 bool X86DAGToDAGISel::
SelectInlineAsmMemoryOperand(const SDValue & Op,char ConstraintCode,std::vector<SDValue> & OutOps)2761 SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
2762 std::vector<SDValue> &OutOps) {
2763 SDValue Op0, Op1, Op2, Op3, Op4;
2764 switch (ConstraintCode) {
2765 case 'o': // offsetable ??
2766 case 'v': // not offsetable ??
2767 default: return true;
2768 case 'm': // memory
2769 if (!SelectAddr(nullptr, Op, Op0, Op1, Op2, Op3, Op4))
2770 return true;
2771 break;
2772 }
2773
2774 OutOps.push_back(Op0);
2775 OutOps.push_back(Op1);
2776 OutOps.push_back(Op2);
2777 OutOps.push_back(Op3);
2778 OutOps.push_back(Op4);
2779 return false;
2780 }
2781
2782 /// createX86ISelDag - This pass converts a legalized DAG into a
2783 /// X86-specific DAG, ready for instruction scheduling.
2784 ///
createX86ISelDag(X86TargetMachine & TM,CodeGenOpt::Level OptLevel)2785 FunctionPass *llvm::createX86ISelDag(X86TargetMachine &TM,
2786 CodeGenOpt::Level OptLevel) {
2787 return new X86DAGToDAGISel(TM, OptLevel);
2788 }
2789