1 //===- X86ISelDAGToDAG.cpp - A DAG pattern matching inst selector for X86 -===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines a DAG pattern matching instruction selector for X86,
11 // converting from a legalized dag to a X86 dag.
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "x86-isel"
17 #include "X86InstrBuilder.h"
18 #include "X86MachineFunctionInfo.h"
19 #include "X86RegisterInfo.h"
20 #include "X86Subtarget.h"
21 #include "X86TargetMachine.h"
22 #include "llvm/Instructions.h"
23 #include "llvm/Intrinsics.h"
24 #include "llvm/Support/CFG.h"
25 #include "llvm/Type.h"
26 #include "llvm/CodeGen/MachineConstantPool.h"
27 #include "llvm/CodeGen/MachineFunction.h"
28 #include "llvm/CodeGen/MachineFrameInfo.h"
29 #include "llvm/CodeGen/MachineInstrBuilder.h"
30 #include "llvm/CodeGen/MachineRegisterInfo.h"
31 #include "llvm/CodeGen/SelectionDAGISel.h"
32 #include "llvm/Target/TargetMachine.h"
33 #include "llvm/Target/TargetOptions.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/ErrorHandling.h"
36 #include "llvm/Support/MathExtras.h"
37 #include "llvm/Support/raw_ostream.h"
38 #include "llvm/ADT/SmallPtrSet.h"
39 #include "llvm/ADT/Statistic.h"
42 STATISTIC(NumLoadMoved, "Number of loads moved below TokenFactor");
44 //===----------------------------------------------------------------------===//
45 // Pattern Matcher Implementation
46 //===----------------------------------------------------------------------===//
49 /// X86ISelAddressMode - This corresponds to X86AddressMode, but uses
50 /// SDValue's instead of register numbers for the leaves of the matched
52 struct X86ISelAddressMode {
58 // This is really a union, discriminated by BaseType!
66 const GlobalValue *GV;
68 const BlockAddress *BlockAddr;
71 unsigned Align; // CP alignment.
72 unsigned char SymbolFlags; // X86II::MO_*
75 : BaseType(RegBase), Base_FrameIndex(0), Scale(1), IndexReg(), Disp(0),
76 Segment(), GV(0), CP(0), BlockAddr(0), ES(0), JT(-1), Align(0),
77 SymbolFlags(X86II::MO_NO_FLAG) {
80 bool hasSymbolicDisplacement() const {
81 return GV != 0 || CP != 0 || ES != 0 || JT != -1 || BlockAddr != 0;
84 bool hasBaseOrIndexReg() const {
85 return IndexReg.getNode() != 0 || Base_Reg.getNode() != 0;
88 /// isRIPRelative - Return true if this addressing mode is already RIP
90 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;
98 void setBaseReg(SDValue Reg) {
104 dbgs() << "X86ISelAddressMode " << this << '\n';
105 dbgs() << "Base_Reg ";
106 if (Base_Reg.getNode() != 0)
107 Base_Reg.getNode()->dump();
110 dbgs() << " Base.FrameIndex " << Base_FrameIndex << '\n'
111 << " Scale" << Scale << '\n'
113 if (IndexReg.getNode() != 0)
114 IndexReg.getNode()->dump();
117 dbgs() << " Disp " << Disp << '\n'
134 dbgs() << " JT" << JT << " Align" << Align << '\n';
140 //===--------------------------------------------------------------------===//
141 /// ISel - X86 specific code to select X86 machine instructions for
142 /// SelectionDAG operations.
144 class X86DAGToDAGISel : public SelectionDAGISel {
145 /// X86Lowering - This object fully describes how to lower LLVM code to an
146 /// X86-specific SelectionDAG.
147 const X86TargetLowering &X86Lowering;
149 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
150 /// make the right decision when generating code for different targets.
151 const X86Subtarget *Subtarget;
153 /// OptForSize - If true, selector should try to optimize for code size
154 /// instead of performance.
158 explicit X86DAGToDAGISel(X86TargetMachine &tm, CodeGenOpt::Level OptLevel)
159 : SelectionDAGISel(tm, OptLevel),
160 X86Lowering(*tm.getTargetLowering()),
161 Subtarget(&tm.getSubtarget<X86Subtarget>()),
164 virtual const char *getPassName() const {
165 return "X86 DAG->DAG Instruction Selection";
168 virtual void EmitFunctionEntryCode();
170 virtual bool IsProfitableToFold(SDValue N, SDNode *U, SDNode *Root) const;
172 virtual void PreprocessISelDAG();
174 inline bool immSext8(SDNode *N) const {
175 return isInt<8>(cast<ConstantSDNode>(N)->getSExtValue());
178 // i64immSExt32 predicate - True if the 64-bit immediate fits in a 32-bit
179 // sign extended field.
180 inline bool i64immSExt32(SDNode *N) const {
181 uint64_t v = cast<ConstantSDNode>(N)->getZExtValue();
182 return (int64_t)v == (int32_t)v;
185 // Include the pieces autogenerated from the target description.
186 #include "X86GenDAGISel.inc"
189 SDNode *Select(SDNode *N);
190 SDNode *SelectAtomic64(SDNode *Node, unsigned Opc);
191 SDNode *SelectAtomicLoadAdd(SDNode *Node, EVT NVT);
193 bool MatchSegmentBaseAddress(SDValue N, X86ISelAddressMode &AM);
194 bool MatchLoad(SDValue N, X86ISelAddressMode &AM);
195 bool MatchWrapper(SDValue N, X86ISelAddressMode &AM);
196 bool MatchAddress(SDValue N, X86ISelAddressMode &AM);
197 bool MatchAddressRecursively(SDValue N, X86ISelAddressMode &AM,
199 bool MatchAddressBase(SDValue N, X86ISelAddressMode &AM);
200 bool SelectAddr(SDNode *Parent, SDValue N, SDValue &Base,
201 SDValue &Scale, SDValue &Index, SDValue &Disp,
203 bool SelectLEAAddr(SDValue N, SDValue &Base,
204 SDValue &Scale, SDValue &Index, SDValue &Disp,
206 bool SelectTLSADDRAddr(SDValue N, SDValue &Base,
207 SDValue &Scale, SDValue &Index, SDValue &Disp,
209 bool SelectScalarSSELoad(SDNode *Root, SDValue N,
210 SDValue &Base, SDValue &Scale,
211 SDValue &Index, SDValue &Disp,
213 SDValue &NodeWithChain);
215 bool TryFoldLoad(SDNode *P, SDValue N,
216 SDValue &Base, SDValue &Scale,
217 SDValue &Index, SDValue &Disp,
220 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
221 /// inline asm expressions.
222 virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
224 std::vector<SDValue> &OutOps);
226 void EmitSpecialCodeForMain(MachineBasicBlock *BB, MachineFrameInfo *MFI);
228 inline void getAddressOperands(X86ISelAddressMode &AM, SDValue &Base,
229 SDValue &Scale, SDValue &Index,
230 SDValue &Disp, SDValue &Segment) {
231 Base = (AM.BaseType == X86ISelAddressMode::FrameIndexBase) ?
232 CurDAG->getTargetFrameIndex(AM.Base_FrameIndex, TLI.getPointerTy()) :
234 Scale = getI8Imm(AM.Scale);
236 // These are 32-bit even in 64-bit mode since RIP relative offset
239 Disp = CurDAG->getTargetGlobalAddress(AM.GV, DebugLoc(),
243 Disp = CurDAG->getTargetConstantPool(AM.CP, MVT::i32,
244 AM.Align, AM.Disp, AM.SymbolFlags);
246 Disp = CurDAG->getTargetExternalSymbol(AM.ES, MVT::i32, AM.SymbolFlags);
247 else if (AM.JT != -1)
248 Disp = CurDAG->getTargetJumpTable(AM.JT, MVT::i32, AM.SymbolFlags);
249 else if (AM.BlockAddr)
250 Disp = CurDAG->getBlockAddress(AM.BlockAddr, MVT::i32,
251 true, AM.SymbolFlags);
253 Disp = CurDAG->getTargetConstant(AM.Disp, MVT::i32);
255 if (AM.Segment.getNode())
256 Segment = AM.Segment;
258 Segment = CurDAG->getRegister(0, MVT::i32);
261 /// getI8Imm - Return a target constant with the specified value, of type
263 inline SDValue getI8Imm(unsigned Imm) {
264 return CurDAG->getTargetConstant(Imm, MVT::i8);
267 /// getI32Imm - Return a target constant with the specified value, of type
269 inline SDValue getI32Imm(unsigned Imm) {
270 return CurDAG->getTargetConstant(Imm, MVT::i32);
273 /// getGlobalBaseReg - Return an SDNode that returns the value of
274 /// the global base register. Output instructions required to
275 /// initialize the global base register, if necessary.
277 SDNode *getGlobalBaseReg();
279 /// getTargetMachine - Return a reference to the TargetMachine, casted
280 /// to the target-specific type.
281 const X86TargetMachine &getTargetMachine() {
282 return static_cast<const X86TargetMachine &>(TM);
285 /// getInstrInfo - Return a reference to the TargetInstrInfo, casted
286 /// to the target-specific type.
287 const X86InstrInfo *getInstrInfo() {
288 return getTargetMachine().getInstrInfo();
295 X86DAGToDAGISel::IsProfitableToFold(SDValue N, SDNode *U, SDNode *Root) const {
296 if (OptLevel == CodeGenOpt::None) return false;
301 if (N.getOpcode() != ISD::LOAD)
304 // If N is a load, do additional profitability checks.
306 switch (U->getOpcode()) {
319 SDValue Op1 = U->getOperand(1);
321 // If the other operand is a 8-bit immediate we should fold the immediate
322 // instead. This reduces code size.
324 // movl 4(%esp), %eax
328 // addl 4(%esp), %eax
329 // The former is 2 bytes shorter. In case where the increment is 1, then
330 // the saving can be 4 bytes (by using incl %eax).
331 if (ConstantSDNode *Imm = dyn_cast<ConstantSDNode>(Op1))
332 if (Imm->getAPIntValue().isSignedIntN(8))
335 // If the other operand is a TLS address, we should fold it instead.
338 // leal i@NTPOFF(%eax), %eax
340 // movl $i@NTPOFF, %eax
342 // if the block also has an access to a second TLS address this will save
344 // FIXME: This is probably also true for non TLS addresses.
345 if (Op1.getOpcode() == X86ISD::Wrapper) {
346 SDValue Val = Op1.getOperand(0);
347 if (Val.getOpcode() == ISD::TargetGlobalTLSAddress)
357 /// MoveBelowCallOrigChain - Replace the original chain operand of the call with
358 /// load's chain operand and move load below the call's chain operand.
359 static void MoveBelowOrigChain(SelectionDAG *CurDAG, SDValue Load,
360 SDValue Call, SDValue OrigChain) {
361 SmallVector<SDValue, 8> Ops;
362 SDValue Chain = OrigChain.getOperand(0);
363 if (Chain.getNode() == Load.getNode())
364 Ops.push_back(Load.getOperand(0));
366 assert(Chain.getOpcode() == ISD::TokenFactor &&
367 "Unexpected chain operand");
368 for (unsigned i = 0, e = Chain.getNumOperands(); i != e; ++i)
369 if (Chain.getOperand(i).getNode() == Load.getNode())
370 Ops.push_back(Load.getOperand(0));
372 Ops.push_back(Chain.getOperand(i));
374 CurDAG->getNode(ISD::TokenFactor, Load.getDebugLoc(),
375 MVT::Other, &Ops[0], Ops.size());
377 Ops.push_back(NewChain);
379 for (unsigned i = 1, e = OrigChain.getNumOperands(); i != e; ++i)
380 Ops.push_back(OrigChain.getOperand(i));
381 CurDAG->UpdateNodeOperands(OrigChain.getNode(), &Ops[0], Ops.size());
382 CurDAG->UpdateNodeOperands(Load.getNode(), Call.getOperand(0),
383 Load.getOperand(1), Load.getOperand(2));
385 Ops.push_back(SDValue(Load.getNode(), 1));
386 for (unsigned i = 1, e = Call.getNode()->getNumOperands(); i != e; ++i)
387 Ops.push_back(Call.getOperand(i));
388 CurDAG->UpdateNodeOperands(Call.getNode(), &Ops[0], Ops.size());
391 /// isCalleeLoad - Return true if call address is a load and it can be
392 /// moved below CALLSEQ_START and the chains leading up to the call.
393 /// Return the CALLSEQ_START by reference as a second output.
394 /// In the case of a tail call, there isn't a callseq node between the call
395 /// chain and the load.
396 static bool isCalleeLoad(SDValue Callee, SDValue &Chain, bool HasCallSeq) {
397 if (Callee.getNode() == Chain.getNode() || !Callee.hasOneUse())
399 LoadSDNode *LD = dyn_cast<LoadSDNode>(Callee.getNode());
402 LD->getAddressingMode() != ISD::UNINDEXED ||
403 LD->getExtensionType() != ISD::NON_EXTLOAD)
406 // Now let's find the callseq_start.
407 while (HasCallSeq && Chain.getOpcode() != ISD::CALLSEQ_START) {
408 if (!Chain.hasOneUse())
410 Chain = Chain.getOperand(0);
413 if (!Chain.getNumOperands())
415 if (Chain.getOperand(0).getNode() == Callee.getNode())
417 if (Chain.getOperand(0).getOpcode() == ISD::TokenFactor &&
418 Callee.getValue(1).isOperandOf(Chain.getOperand(0).getNode()) &&
419 Callee.getValue(1).hasOneUse())
424 void X86DAGToDAGISel::PreprocessISelDAG() {
425 // OptForSize is used in pattern predicates that isel is matching.
426 OptForSize = MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize);
428 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
429 E = CurDAG->allnodes_end(); I != E; ) {
430 SDNode *N = I++; // Preincrement iterator to avoid invalidation issues.
432 if (OptLevel != CodeGenOpt::None &&
433 (N->getOpcode() == X86ISD::CALL ||
434 N->getOpcode() == X86ISD::TC_RETURN)) {
435 /// Also try moving call address load from outside callseq_start to just
436 /// before the call to allow it to be folded.
454 bool HasCallSeq = N->getOpcode() == X86ISD::CALL;
455 SDValue Chain = N->getOperand(0);
456 SDValue Load = N->getOperand(1);
457 if (!isCalleeLoad(Load, Chain, HasCallSeq))
459 MoveBelowOrigChain(CurDAG, Load, SDValue(N, 0), Chain);
464 // Lower fpround and fpextend nodes that target the FP stack to be store and
465 // load to the stack. This is a gross hack. We would like to simply mark
466 // these as being illegal, but when we do that, legalize produces these when
467 // it expands calls, then expands these in the same legalize pass. We would
468 // like dag combine to be able to hack on these between the call expansion
469 // and the node legalization. As such this pass basically does "really
470 // late" legalization of these inline with the X86 isel pass.
471 // FIXME: This should only happen when not compiled with -O0.
472 if (N->getOpcode() != ISD::FP_ROUND && N->getOpcode() != ISD::FP_EXTEND)
475 // If the source and destination are SSE registers, then this is a legal
476 // conversion that should not be lowered.
477 EVT SrcVT = N->getOperand(0).getValueType();
478 EVT DstVT = N->getValueType(0);
479 bool SrcIsSSE = X86Lowering.isScalarFPTypeInSSEReg(SrcVT);
480 bool DstIsSSE = X86Lowering.isScalarFPTypeInSSEReg(DstVT);
481 if (SrcIsSSE && DstIsSSE)
484 if (!SrcIsSSE && !DstIsSSE) {
485 // If this is an FPStack extension, it is a noop.
486 if (N->getOpcode() == ISD::FP_EXTEND)
488 // If this is a value-preserving FPStack truncation, it is a noop.
489 if (N->getConstantOperandVal(1))
493 // Here we could have an FP stack truncation or an FPStack <-> SSE convert.
494 // FPStack has extload and truncstore. SSE can fold direct loads into other
495 // operations. Based on this, decide what we want to do.
497 if (N->getOpcode() == ISD::FP_ROUND)
498 MemVT = DstVT; // FP_ROUND must use DstVT, we can't do a 'trunc load'.
500 MemVT = SrcIsSSE ? SrcVT : DstVT;
502 SDValue MemTmp = CurDAG->CreateStackTemporary(MemVT);
503 DebugLoc dl = N->getDebugLoc();
505 // FIXME: optimize the case where the src/dest is a load or store?
506 SDValue Store = CurDAG->getTruncStore(CurDAG->getEntryNode(), dl,
508 MemTmp, MachinePointerInfo(), MemVT,
510 SDValue Result = CurDAG->getExtLoad(ISD::EXTLOAD, DstVT, dl, Store, MemTmp,
511 MachinePointerInfo(),
512 MemVT, false, false, 0);
514 // We're about to replace all uses of the FP_ROUND/FP_EXTEND with the
515 // extload we created. This will cause general havok on the dag because
516 // anything below the conversion could be folded into other existing nodes.
517 // To avoid invalidating 'I', back it up to the convert node.
519 CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
521 // Now that we did that, the node is dead. Increment the iterator to the
522 // next node to process, then delete N.
524 CurDAG->DeleteNode(N);
529 /// EmitSpecialCodeForMain - Emit any code that needs to be executed only in
530 /// the main function.
531 void X86DAGToDAGISel::EmitSpecialCodeForMain(MachineBasicBlock *BB,
532 MachineFrameInfo *MFI) {
533 const TargetInstrInfo *TII = TM.getInstrInfo();
534 if (Subtarget->isTargetCygMing())
535 BuildMI(BB, DebugLoc(),
536 TII->get(X86::CALLpcrel32)).addExternalSymbol("__main");
539 void X86DAGToDAGISel::EmitFunctionEntryCode() {
540 // If this is main, emit special code for main.
541 if (const Function *Fn = MF->getFunction())
542 if (Fn->hasExternalLinkage() && Fn->getName() == "main")
543 EmitSpecialCodeForMain(MF->begin(), MF->getFrameInfo());
547 bool X86DAGToDAGISel::MatchSegmentBaseAddress(SDValue N,
548 X86ISelAddressMode &AM) {
549 assert(N.getOpcode() == X86ISD::SegmentBaseAddress);
550 SDValue Segment = N.getOperand(0);
552 if (AM.Segment.getNode() == 0) {
553 AM.Segment = Segment;
560 bool X86DAGToDAGISel::MatchLoad(SDValue N, X86ISelAddressMode &AM) {
561 // This optimization is valid because the GNU TLS model defines that
562 // gs:0 (or fs:0 on X86-64) contains its own address.
563 // For more information see http://people.redhat.com/drepper/tls.pdf
565 SDValue Address = N.getOperand(1);
566 if (Address.getOpcode() == X86ISD::SegmentBaseAddress &&
567 !MatchSegmentBaseAddress(Address, AM))
573 /// MatchWrapper - Try to match X86ISD::Wrapper and X86ISD::WrapperRIP nodes
574 /// into an addressing mode. These wrap things that will resolve down into a
575 /// symbol reference. If no match is possible, this returns true, otherwise it
577 bool X86DAGToDAGISel::MatchWrapper(SDValue N, X86ISelAddressMode &AM) {
578 // If the addressing mode already has a symbol as the displacement, we can
579 // never match another symbol.
580 if (AM.hasSymbolicDisplacement())
583 SDValue N0 = N.getOperand(0);
584 CodeModel::Model M = TM.getCodeModel();
586 // Handle X86-64 rip-relative addresses. We check this before checking direct
587 // folding because RIP is preferable to non-RIP accesses.
588 if (Subtarget->is64Bit() &&
589 // Under X86-64 non-small code model, GV (and friends) are 64-bits, so
590 // they cannot be folded into immediate fields.
591 // FIXME: This can be improved for kernel and other models?
592 (M == CodeModel::Small || M == CodeModel::Kernel) &&
593 // Base and index reg must be 0 in order to use %rip as base and lowering
595 !AM.hasBaseOrIndexReg() && N.getOpcode() == X86ISD::WrapperRIP) {
596 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(N0)) {
597 int64_t Offset = AM.Disp + G->getOffset();
598 if (!X86::isOffsetSuitableForCodeModel(Offset, M)) return true;
599 AM.GV = G->getGlobal();
601 AM.SymbolFlags = G->getTargetFlags();
602 } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N0)) {
603 int64_t Offset = AM.Disp + CP->getOffset();
604 if (!X86::isOffsetSuitableForCodeModel(Offset, M)) return true;
605 AM.CP = CP->getConstVal();
606 AM.Align = CP->getAlignment();
608 AM.SymbolFlags = CP->getTargetFlags();
609 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(N0)) {
610 AM.ES = S->getSymbol();
611 AM.SymbolFlags = S->getTargetFlags();
612 } else if (JumpTableSDNode *J = dyn_cast<JumpTableSDNode>(N0)) {
613 AM.JT = J->getIndex();
614 AM.SymbolFlags = J->getTargetFlags();
616 AM.BlockAddr = cast<BlockAddressSDNode>(N0)->getBlockAddress();
617 AM.SymbolFlags = cast<BlockAddressSDNode>(N0)->getTargetFlags();
620 if (N.getOpcode() == X86ISD::WrapperRIP)
621 AM.setBaseReg(CurDAG->getRegister(X86::RIP, MVT::i64));
625 // Handle the case when globals fit in our immediate field: This is true for
626 // X86-32 always and X86-64 when in -static -mcmodel=small mode. In 64-bit
627 // mode, this results in a non-RIP-relative computation.
628 if (!Subtarget->is64Bit() ||
629 ((M == CodeModel::Small || M == CodeModel::Kernel) &&
630 TM.getRelocationModel() == Reloc::Static)) {
631 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(N0)) {
632 AM.GV = G->getGlobal();
633 AM.Disp += G->getOffset();
634 AM.SymbolFlags = G->getTargetFlags();
635 } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N0)) {
636 AM.CP = CP->getConstVal();
637 AM.Align = CP->getAlignment();
638 AM.Disp += CP->getOffset();
639 AM.SymbolFlags = CP->getTargetFlags();
640 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(N0)) {
641 AM.ES = S->getSymbol();
642 AM.SymbolFlags = S->getTargetFlags();
643 } else if (JumpTableSDNode *J = dyn_cast<JumpTableSDNode>(N0)) {
644 AM.JT = J->getIndex();
645 AM.SymbolFlags = J->getTargetFlags();
647 AM.BlockAddr = cast<BlockAddressSDNode>(N0)->getBlockAddress();
648 AM.SymbolFlags = cast<BlockAddressSDNode>(N0)->getTargetFlags();
656 /// MatchAddress - Add the specified node to the specified addressing mode,
657 /// returning true if it cannot be done. This just pattern matches for the
659 bool X86DAGToDAGISel::MatchAddress(SDValue N, X86ISelAddressMode &AM) {
660 if (MatchAddressRecursively(N, AM, 0))
663 // Post-processing: Convert lea(,%reg,2) to lea(%reg,%reg), which has
664 // a smaller encoding and avoids a scaled-index.
666 AM.BaseType == X86ISelAddressMode::RegBase &&
667 AM.Base_Reg.getNode() == 0) {
668 AM.Base_Reg = AM.IndexReg;
672 // Post-processing: Convert foo to foo(%rip), even in non-PIC mode,
673 // because it has a smaller encoding.
674 // TODO: Which other code models can use this?
675 if (TM.getCodeModel() == CodeModel::Small &&
676 Subtarget->is64Bit() &&
678 AM.BaseType == X86ISelAddressMode::RegBase &&
679 AM.Base_Reg.getNode() == 0 &&
680 AM.IndexReg.getNode() == 0 &&
681 AM.SymbolFlags == X86II::MO_NO_FLAG &&
682 AM.hasSymbolicDisplacement())
683 AM.Base_Reg = CurDAG->getRegister(X86::RIP, MVT::i64);
688 /// isLogicallyAddWithConstant - Return true if this node is semantically an
689 /// add of a value with a constantint.
690 static bool isLogicallyAddWithConstant(SDValue V, SelectionDAG *CurDAG) {
691 // Check for (add x, Cst)
692 if (V->getOpcode() == ISD::ADD)
693 return isa<ConstantSDNode>(V->getOperand(1));
695 // Check for (or x, Cst), where Cst & x == 0.
696 if (V->getOpcode() != ISD::OR ||
697 !isa<ConstantSDNode>(V->getOperand(1)))
700 // Handle "X | C" as "X + C" iff X is known to have C bits clear.
701 ConstantSDNode *CN = cast<ConstantSDNode>(V->getOperand(1));
703 // Check to see if the LHS & C is zero.
704 return CurDAG->MaskedValueIsZero(V->getOperand(0), CN->getAPIntValue());
707 bool X86DAGToDAGISel::MatchAddressRecursively(SDValue N, X86ISelAddressMode &AM,
709 bool is64Bit = Subtarget->is64Bit();
710 DebugLoc dl = N.getDebugLoc();
712 dbgs() << "MatchAddress: ";
717 return MatchAddressBase(N, AM);
719 CodeModel::Model M = TM.getCodeModel();
721 // If this is already a %rip relative address, we can only merge immediates
722 // into it. Instead of handling this in every case, we handle it here.
723 // RIP relative addressing: %rip + 32-bit displacement!
724 if (AM.isRIPRelative()) {
725 // FIXME: JumpTable and ExternalSymbol address currently don't like
726 // displacements. It isn't very important, but this should be fixed for
728 if (!AM.ES && AM.JT != -1) return true;
730 if (ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(N)) {
731 int64_t Val = AM.Disp + Cst->getSExtValue();
732 if (X86::isOffsetSuitableForCodeModel(Val, M,
733 AM.hasSymbolicDisplacement())) {
741 switch (N.getOpcode()) {
743 case ISD::Constant: {
744 uint64_t Val = cast<ConstantSDNode>(N)->getSExtValue();
746 X86::isOffsetSuitableForCodeModel(AM.Disp + Val, M,
747 AM.hasSymbolicDisplacement())) {
754 case X86ISD::SegmentBaseAddress:
755 if (!MatchSegmentBaseAddress(N, AM))
759 case X86ISD::Wrapper:
760 case X86ISD::WrapperRIP:
761 if (!MatchWrapper(N, AM))
766 if (!MatchLoad(N, AM))
770 case ISD::FrameIndex:
771 if (AM.BaseType == X86ISelAddressMode::RegBase
772 && AM.Base_Reg.getNode() == 0) {
773 AM.BaseType = X86ISelAddressMode::FrameIndexBase;
774 AM.Base_FrameIndex = cast<FrameIndexSDNode>(N)->getIndex();
780 if (AM.IndexReg.getNode() != 0 || AM.Scale != 1)
784 *CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1))) {
785 unsigned Val = CN->getZExtValue();
786 // Note that we handle x<<1 as (,x,2) rather than (x,x) here so
787 // that the base operand remains free for further matching. If
788 // the base doesn't end up getting used, a post-processing step
789 // in MatchAddress turns (,x,2) into (x,x), which is cheaper.
790 if (Val == 1 || Val == 2 || Val == 3) {
792 SDValue ShVal = N.getNode()->getOperand(0);
794 // Okay, we know that we have a scale by now. However, if the scaled
795 // value is an add of something and a constant, we can fold the
796 // constant into the disp field here.
797 if (isLogicallyAddWithConstant(ShVal, CurDAG)) {
798 AM.IndexReg = ShVal.getNode()->getOperand(0);
799 ConstantSDNode *AddVal =
800 cast<ConstantSDNode>(ShVal.getNode()->getOperand(1));
801 uint64_t Disp = AM.Disp + (AddVal->getSExtValue() << Val);
803 X86::isOffsetSuitableForCodeModel(Disp, M,
804 AM.hasSymbolicDisplacement()))
818 // A mul_lohi where we need the low part can be folded as a plain multiply.
819 if (N.getResNo() != 0) break;
822 case X86ISD::MUL_IMM:
823 // X*[3,5,9] -> X+X*[2,4,8]
824 if (AM.BaseType == X86ISelAddressMode::RegBase &&
825 AM.Base_Reg.getNode() == 0 &&
826 AM.IndexReg.getNode() == 0) {
828 *CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1)))
829 if (CN->getZExtValue() == 3 || CN->getZExtValue() == 5 ||
830 CN->getZExtValue() == 9) {
831 AM.Scale = unsigned(CN->getZExtValue())-1;
833 SDValue MulVal = N.getNode()->getOperand(0);
836 // Okay, we know that we have a scale by now. However, if the scaled
837 // value is an add of something and a constant, we can fold the
838 // constant into the disp field here.
839 if (MulVal.getNode()->getOpcode() == ISD::ADD && MulVal.hasOneUse() &&
840 isa<ConstantSDNode>(MulVal.getNode()->getOperand(1))) {
841 Reg = MulVal.getNode()->getOperand(0);
842 ConstantSDNode *AddVal =
843 cast<ConstantSDNode>(MulVal.getNode()->getOperand(1));
844 uint64_t Disp = AM.Disp + AddVal->getSExtValue() *
847 X86::isOffsetSuitableForCodeModel(Disp, M,
848 AM.hasSymbolicDisplacement()))
851 Reg = N.getNode()->getOperand(0);
853 Reg = N.getNode()->getOperand(0);
856 AM.IndexReg = AM.Base_Reg = Reg;
863 // Given A-B, if A can be completely folded into the address and
864 // the index field with the index field unused, use -B as the index.
865 // This is a win if a has multiple parts that can be folded into
866 // the address. Also, this saves a mov if the base register has
867 // other uses, since it avoids a two-address sub instruction, however
868 // it costs an additional mov if the index register has other uses.
870 // Add an artificial use to this node so that we can keep track of
871 // it if it gets CSE'd with a different node.
872 HandleSDNode Handle(N);
874 // Test if the LHS of the sub can be folded.
875 X86ISelAddressMode Backup = AM;
876 if (MatchAddressRecursively(N.getNode()->getOperand(0), AM, Depth+1)) {
880 // Test if the index field is free for use.
881 if (AM.IndexReg.getNode() || AM.isRIPRelative()) {
887 SDValue RHS = Handle.getValue().getNode()->getOperand(1);
888 // If the RHS involves a register with multiple uses, this
889 // transformation incurs an extra mov, due to the neg instruction
890 // clobbering its operand.
891 if (!RHS.getNode()->hasOneUse() ||
892 RHS.getNode()->getOpcode() == ISD::CopyFromReg ||
893 RHS.getNode()->getOpcode() == ISD::TRUNCATE ||
894 RHS.getNode()->getOpcode() == ISD::ANY_EXTEND ||
895 (RHS.getNode()->getOpcode() == ISD::ZERO_EXTEND &&
896 RHS.getNode()->getOperand(0).getValueType() == MVT::i32))
898 // If the base is a register with multiple uses, this
899 // transformation may save a mov.
900 if ((AM.BaseType == X86ISelAddressMode::RegBase &&
901 AM.Base_Reg.getNode() &&
902 !AM.Base_Reg.getNode()->hasOneUse()) ||
903 AM.BaseType == X86ISelAddressMode::FrameIndexBase)
905 // If the folded LHS was interesting, this transformation saves
906 // address arithmetic.
907 if ((AM.hasSymbolicDisplacement() && !Backup.hasSymbolicDisplacement()) +
908 ((AM.Disp != 0) && (Backup.Disp == 0)) +
909 (AM.Segment.getNode() && !Backup.Segment.getNode()) >= 2)
911 // If it doesn't look like it may be an overall win, don't do it.
917 // Ok, the transformation is legal and appears profitable. Go for it.
918 SDValue Zero = CurDAG->getConstant(0, N.getValueType());
919 SDValue Neg = CurDAG->getNode(ISD::SUB, dl, N.getValueType(), Zero, RHS);
923 // Insert the new nodes into the topological ordering.
924 if (Zero.getNode()->getNodeId() == -1 ||
925 Zero.getNode()->getNodeId() > N.getNode()->getNodeId()) {
926 CurDAG->RepositionNode(N.getNode(), Zero.getNode());
927 Zero.getNode()->setNodeId(N.getNode()->getNodeId());
929 if (Neg.getNode()->getNodeId() == -1 ||
930 Neg.getNode()->getNodeId() > N.getNode()->getNodeId()) {
931 CurDAG->RepositionNode(N.getNode(), Neg.getNode());
932 Neg.getNode()->setNodeId(N.getNode()->getNodeId());
938 // Add an artificial use to this node so that we can keep track of
939 // it if it gets CSE'd with a different node.
940 HandleSDNode Handle(N);
941 SDValue LHS = Handle.getValue().getNode()->getOperand(0);
942 SDValue RHS = Handle.getValue().getNode()->getOperand(1);
944 X86ISelAddressMode Backup = AM;
945 if (!MatchAddressRecursively(LHS, AM, Depth+1) &&
946 !MatchAddressRecursively(RHS, AM, Depth+1))
949 LHS = Handle.getValue().getNode()->getOperand(0);
950 RHS = Handle.getValue().getNode()->getOperand(1);
952 // Try again after commuting the operands.
953 if (!MatchAddressRecursively(RHS, AM, Depth+1) &&
954 !MatchAddressRecursively(LHS, AM, Depth+1))
957 LHS = Handle.getValue().getNode()->getOperand(0);
958 RHS = Handle.getValue().getNode()->getOperand(1);
960 // If we couldn't fold both operands into the address at the same time,
961 // see if we can just put each operand into a register and fold at least
963 if (AM.BaseType == X86ISelAddressMode::RegBase &&
964 !AM.Base_Reg.getNode() &&
965 !AM.IndexReg.getNode()) {
975 // Handle "X | C" as "X + C" iff X is known to have C bits clear.
976 if (isLogicallyAddWithConstant(N, CurDAG)) {
977 X86ISelAddressMode Backup = AM;
978 ConstantSDNode *CN = cast<ConstantSDNode>(N.getOperand(1));
979 uint64_t Offset = CN->getSExtValue();
981 // Start with the LHS as an addr mode.
982 if (!MatchAddressRecursively(N.getOperand(0), AM, Depth+1) &&
983 // Address could not have picked a GV address for the displacement.
985 // On x86-64, the resultant disp must fit in 32-bits.
987 X86::isOffsetSuitableForCodeModel(AM.Disp + Offset, M,
988 AM.hasSymbolicDisplacement()))) {
997 // Perform some heroic transforms on an and of a constant-count shift
998 // with a constant to enable use of the scaled offset field.
1000 SDValue Shift = N.getOperand(0);
1001 if (Shift.getNumOperands() != 2) break;
1003 // Scale must not be used already.
1004 if (AM.IndexReg.getNode() != 0 || AM.Scale != 1) break;
1006 SDValue X = Shift.getOperand(0);
1007 ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N.getOperand(1));
1008 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(Shift.getOperand(1));
1009 if (!C1 || !C2) break;
1011 // Handle "(X >> (8-C1)) & C2" as "(X >> 8) & 0xff)" if safe. This
1012 // allows us to convert the shift and and into an h-register extract and
1014 if (Shift.getOpcode() == ISD::SRL && Shift.hasOneUse()) {
1015 unsigned ScaleLog = 8 - C1->getZExtValue();
1016 if (ScaleLog > 0 && ScaleLog < 4 &&
1017 C2->getZExtValue() == (UINT64_C(0xff) << ScaleLog)) {
1018 SDValue Eight = CurDAG->getConstant(8, MVT::i8);
1019 SDValue Mask = CurDAG->getConstant(0xff, N.getValueType());
1020 SDValue Srl = CurDAG->getNode(ISD::SRL, dl, N.getValueType(),
1022 SDValue And = CurDAG->getNode(ISD::AND, dl, N.getValueType(),
1024 SDValue ShlCount = CurDAG->getConstant(ScaleLog, MVT::i8);
1025 SDValue Shl = CurDAG->getNode(ISD::SHL, dl, N.getValueType(),
1028 // Insert the new nodes into the topological ordering.
1029 if (Eight.getNode()->getNodeId() == -1 ||
1030 Eight.getNode()->getNodeId() > X.getNode()->getNodeId()) {
1031 CurDAG->RepositionNode(X.getNode(), Eight.getNode());
1032 Eight.getNode()->setNodeId(X.getNode()->getNodeId());
1034 if (Mask.getNode()->getNodeId() == -1 ||
1035 Mask.getNode()->getNodeId() > X.getNode()->getNodeId()) {
1036 CurDAG->RepositionNode(X.getNode(), Mask.getNode());
1037 Mask.getNode()->setNodeId(X.getNode()->getNodeId());
1039 if (Srl.getNode()->getNodeId() == -1 ||
1040 Srl.getNode()->getNodeId() > Shift.getNode()->getNodeId()) {
1041 CurDAG->RepositionNode(Shift.getNode(), Srl.getNode());
1042 Srl.getNode()->setNodeId(Shift.getNode()->getNodeId());
1044 if (And.getNode()->getNodeId() == -1 ||
1045 And.getNode()->getNodeId() > N.getNode()->getNodeId()) {
1046 CurDAG->RepositionNode(N.getNode(), And.getNode());
1047 And.getNode()->setNodeId(N.getNode()->getNodeId());
1049 if (ShlCount.getNode()->getNodeId() == -1 ||
1050 ShlCount.getNode()->getNodeId() > X.getNode()->getNodeId()) {
1051 CurDAG->RepositionNode(X.getNode(), ShlCount.getNode());
1052 ShlCount.getNode()->setNodeId(N.getNode()->getNodeId());
1054 if (Shl.getNode()->getNodeId() == -1 ||
1055 Shl.getNode()->getNodeId() > N.getNode()->getNodeId()) {
1056 CurDAG->RepositionNode(N.getNode(), Shl.getNode());
1057 Shl.getNode()->setNodeId(N.getNode()->getNodeId());
1059 CurDAG->ReplaceAllUsesWith(N, Shl);
1061 AM.Scale = (1 << ScaleLog);
1066 // Handle "(X << C1) & C2" as "(X & (C2>>C1)) << C1" if safe and if this
1067 // allows us to fold the shift into this addressing mode.
1068 if (Shift.getOpcode() != ISD::SHL) break;
1070 // Not likely to be profitable if either the AND or SHIFT node has more
1071 // than one use (unless all uses are for address computation). Besides,
1072 // isel mechanism requires their node ids to be reused.
1073 if (!N.hasOneUse() || !Shift.hasOneUse())
1076 // Verify that the shift amount is something we can fold.
1077 unsigned ShiftCst = C1->getZExtValue();
1078 if (ShiftCst != 1 && ShiftCst != 2 && ShiftCst != 3)
1081 // Get the new AND mask, this folds to a constant.
1082 SDValue NewANDMask = CurDAG->getNode(ISD::SRL, dl, N.getValueType(),
1083 SDValue(C2, 0), SDValue(C1, 0));
1084 SDValue NewAND = CurDAG->getNode(ISD::AND, dl, N.getValueType(), X,
1086 SDValue NewSHIFT = CurDAG->getNode(ISD::SHL, dl, N.getValueType(),
1087 NewAND, SDValue(C1, 0));
1089 // Insert the new nodes into the topological ordering.
1090 if (C1->getNodeId() > X.getNode()->getNodeId()) {
1091 CurDAG->RepositionNode(X.getNode(), C1);
1092 C1->setNodeId(X.getNode()->getNodeId());
1094 if (NewANDMask.getNode()->getNodeId() == -1 ||
1095 NewANDMask.getNode()->getNodeId() > X.getNode()->getNodeId()) {
1096 CurDAG->RepositionNode(X.getNode(), NewANDMask.getNode());
1097 NewANDMask.getNode()->setNodeId(X.getNode()->getNodeId());
1099 if (NewAND.getNode()->getNodeId() == -1 ||
1100 NewAND.getNode()->getNodeId() > Shift.getNode()->getNodeId()) {
1101 CurDAG->RepositionNode(Shift.getNode(), NewAND.getNode());
1102 NewAND.getNode()->setNodeId(Shift.getNode()->getNodeId());
1104 if (NewSHIFT.getNode()->getNodeId() == -1 ||
1105 NewSHIFT.getNode()->getNodeId() > N.getNode()->getNodeId()) {
1106 CurDAG->RepositionNode(N.getNode(), NewSHIFT.getNode());
1107 NewSHIFT.getNode()->setNodeId(N.getNode()->getNodeId());
1110 CurDAG->ReplaceAllUsesWith(N, NewSHIFT);
1112 AM.Scale = 1 << ShiftCst;
1113 AM.IndexReg = NewAND;
1118 return MatchAddressBase(N, AM);
1121 /// MatchAddressBase - Helper for MatchAddress. Add the specified node to the
1122 /// specified addressing mode without any further recursion.
1123 bool X86DAGToDAGISel::MatchAddressBase(SDValue N, X86ISelAddressMode &AM) {
1124 // Is the base register already occupied?
1125 if (AM.BaseType != X86ISelAddressMode::RegBase || AM.Base_Reg.getNode()) {
1126 // If so, check to see if the scale index register is set.
1127 if (AM.IndexReg.getNode() == 0) {
1133 // Otherwise, we cannot select it.
1137 // Default, generate it as a register.
1138 AM.BaseType = X86ISelAddressMode::RegBase;
1143 /// SelectAddr - returns true if it is able pattern match an addressing mode.
1144 /// It returns the operands which make up the maximal addressing mode it can
1145 /// match by reference.
1147 /// Parent is the parent node of the addr operand that is being matched. It
1148 /// is always a load, store, atomic node, or null. It is only null when
1149 /// checking memory operands for inline asm nodes.
1150 bool X86DAGToDAGISel::SelectAddr(SDNode *Parent, SDValue N, SDValue &Base,
1151 SDValue &Scale, SDValue &Index,
1152 SDValue &Disp, SDValue &Segment) {
1153 X86ISelAddressMode AM;
1154 if (MatchAddress(N, AM))
1157 EVT VT = N.getValueType();
1158 if (AM.BaseType == X86ISelAddressMode::RegBase) {
1159 if (!AM.Base_Reg.getNode())
1160 AM.Base_Reg = CurDAG->getRegister(0, VT);
1163 if (!AM.IndexReg.getNode())
1164 AM.IndexReg = CurDAG->getRegister(0, VT);
1167 // This list of opcodes are all the nodes that have an "addr:$ptr" operand
1168 // that are not a MemSDNode, and thus don't have proper addrspace info.
1169 Parent->getOpcode() != ISD::PREFETCH &&
1170 Parent->getOpcode() != ISD::INTRINSIC_W_CHAIN && // unaligned loads, fixme
1171 Parent->getOpcode() != ISD::INTRINSIC_VOID && // nontemporal stores.
1172 Parent->getOpcode() != X86ISD::VZEXT_LOAD &&
1173 Parent->getOpcode() != X86ISD::FLD &&
1174 Parent->getOpcode() != X86ISD::FILD &&
1175 Parent->getOpcode() != X86ISD::FILD_FLAG &&
1176 Parent->getOpcode() != X86ISD::FP_TO_INT16_IN_MEM &&
1177 Parent->getOpcode() != X86ISD::FP_TO_INT32_IN_MEM &&
1178 Parent->getOpcode() != X86ISD::FP_TO_INT64_IN_MEM &&
1179 Parent->getOpcode() != X86ISD::FST) {
1180 unsigned AddrSpace =
1181 cast<MemSDNode>(Parent)->getPointerInfo().getAddrSpace();
1182 // AddrSpace 256 -> GS, 257 -> FS.
1183 if (AddrSpace == 256)
1184 AM.Segment = CurDAG->getRegister(X86::GS, VT);
1185 if (AddrSpace == 257)
1186 AM.Segment = CurDAG->getRegister(X86::FS, VT);
1190 getAddressOperands(AM, Base, Scale, Index, Disp, Segment);
1194 /// SelectScalarSSELoad - Match a scalar SSE load. In particular, we want to
1195 /// match a load whose top elements are either undef or zeros. The load flavor
1196 /// is derived from the type of N, which is either v4f32 or v2f64.
1199 /// PatternChainNode: this is the matched node that has a chain input and
1201 bool X86DAGToDAGISel::SelectScalarSSELoad(SDNode *Root,
1202 SDValue N, SDValue &Base,
1203 SDValue &Scale, SDValue &Index,
1204 SDValue &Disp, SDValue &Segment,
1205 SDValue &PatternNodeWithChain) {
1206 if (N.getOpcode() == ISD::SCALAR_TO_VECTOR) {
1207 PatternNodeWithChain = N.getOperand(0);
1208 if (ISD::isNON_EXTLoad(PatternNodeWithChain.getNode()) &&
1209 PatternNodeWithChain.hasOneUse() &&
1210 IsProfitableToFold(N.getOperand(0), N.getNode(), Root) &&
1211 IsLegalToFold(N.getOperand(0), N.getNode(), Root, OptLevel)) {
1212 LoadSDNode *LD = cast<LoadSDNode>(PatternNodeWithChain);
1213 if (!SelectAddr(LD, LD->getBasePtr(), Base, Scale, Index, Disp, Segment))
1219 // Also handle the case where we explicitly require zeros in the top
1220 // elements. This is a vector shuffle from the zero vector.
1221 if (N.getOpcode() == X86ISD::VZEXT_MOVL && N.getNode()->hasOneUse() &&
1222 // Check to see if the top elements are all zeros (or bitcast of zeros).
1223 N.getOperand(0).getOpcode() == ISD::SCALAR_TO_VECTOR &&
1224 N.getOperand(0).getNode()->hasOneUse() &&
1225 ISD::isNON_EXTLoad(N.getOperand(0).getOperand(0).getNode()) &&
1226 N.getOperand(0).getOperand(0).hasOneUse() &&
1227 IsProfitableToFold(N.getOperand(0), N.getNode(), Root) &&
1228 IsLegalToFold(N.getOperand(0), N.getNode(), Root, OptLevel)) {
1229 // Okay, this is a zero extending load. Fold it.
1230 LoadSDNode *LD = cast<LoadSDNode>(N.getOperand(0).getOperand(0));
1231 if (!SelectAddr(LD, LD->getBasePtr(), Base, Scale, Index, Disp, Segment))
1233 PatternNodeWithChain = SDValue(LD, 0);
1240 /// SelectLEAAddr - it calls SelectAddr and determines if the maximal addressing
1241 /// mode it matches can be cost effectively emitted as an LEA instruction.
1242 bool X86DAGToDAGISel::SelectLEAAddr(SDValue N,
1243 SDValue &Base, SDValue &Scale,
1244 SDValue &Index, SDValue &Disp,
1246 X86ISelAddressMode AM;
1248 // Set AM.Segment to prevent MatchAddress from using one. LEA doesn't support
1250 SDValue Copy = AM.Segment;
1251 SDValue T = CurDAG->getRegister(0, MVT::i32);
1253 if (MatchAddress(N, AM))
1255 assert (T == AM.Segment);
1258 EVT VT = N.getValueType();
1259 unsigned Complexity = 0;
1260 if (AM.BaseType == X86ISelAddressMode::RegBase)
1261 if (AM.Base_Reg.getNode())
1264 AM.Base_Reg = CurDAG->getRegister(0, VT);
1265 else if (AM.BaseType == X86ISelAddressMode::FrameIndexBase)
1268 if (AM.IndexReg.getNode())
1271 AM.IndexReg = CurDAG->getRegister(0, VT);
1273 // Don't match just leal(,%reg,2). It's cheaper to do addl %reg, %reg, or with
1278 // FIXME: We are artificially lowering the criteria to turn ADD %reg, $GA
1279 // to a LEA. This is determined with some expermentation but is by no means
1280 // optimal (especially for code size consideration). LEA is nice because of
1281 // its three-address nature. Tweak the cost function again when we can run
1282 // convertToThreeAddress() at register allocation time.
1283 if (AM.hasSymbolicDisplacement()) {
1284 // For X86-64, we should always use lea to materialize RIP relative
1286 if (Subtarget->is64Bit())
1292 if (AM.Disp && (AM.Base_Reg.getNode() || AM.IndexReg.getNode()))
1295 // If it isn't worth using an LEA, reject it.
1296 if (Complexity <= 2)
1299 getAddressOperands(AM, Base, Scale, Index, Disp, Segment);
1303 /// SelectTLSADDRAddr - This is only run on TargetGlobalTLSAddress nodes.
1304 bool X86DAGToDAGISel::SelectTLSADDRAddr(SDValue N, SDValue &Base,
1305 SDValue &Scale, SDValue &Index,
1306 SDValue &Disp, SDValue &Segment) {
1307 assert(N.getOpcode() == ISD::TargetGlobalTLSAddress);
1308 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(N);
1310 X86ISelAddressMode AM;
1311 AM.GV = GA->getGlobal();
1312 AM.Disp += GA->getOffset();
1313 AM.Base_Reg = CurDAG->getRegister(0, N.getValueType());
1314 AM.SymbolFlags = GA->getTargetFlags();
1316 if (N.getValueType() == MVT::i32) {
1318 AM.IndexReg = CurDAG->getRegister(X86::EBX, MVT::i32);
1320 AM.IndexReg = CurDAG->getRegister(0, MVT::i64);
1323 getAddressOperands(AM, Base, Scale, Index, Disp, Segment);
1328 bool X86DAGToDAGISel::TryFoldLoad(SDNode *P, SDValue N,
1329 SDValue &Base, SDValue &Scale,
1330 SDValue &Index, SDValue &Disp,
1332 if (!ISD::isNON_EXTLoad(N.getNode()) ||
1333 !IsProfitableToFold(N, P, P) ||
1334 !IsLegalToFold(N, P, P, OptLevel))
1337 return SelectAddr(N.getNode(),
1338 N.getOperand(1), Base, Scale, Index, Disp, Segment);
1341 /// getGlobalBaseReg - Return an SDNode that returns the value of
1342 /// the global base register. Output instructions required to
1343 /// initialize the global base register, if necessary.
1345 SDNode *X86DAGToDAGISel::getGlobalBaseReg() {
1346 unsigned GlobalBaseReg = getInstrInfo()->getGlobalBaseReg(MF);
1347 return CurDAG->getRegister(GlobalBaseReg, TLI.getPointerTy()).getNode();
1350 SDNode *X86DAGToDAGISel::SelectAtomic64(SDNode *Node, unsigned Opc) {
1351 SDValue Chain = Node->getOperand(0);
1352 SDValue In1 = Node->getOperand(1);
1353 SDValue In2L = Node->getOperand(2);
1354 SDValue In2H = Node->getOperand(3);
1355 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1356 if (!SelectAddr(Node, In1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4))
1358 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
1359 MemOp[0] = cast<MemSDNode>(Node)->getMemOperand();
1360 const SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, In2L, In2H, Chain};
1361 SDNode *ResNode = CurDAG->getMachineNode(Opc, Node->getDebugLoc(),
1362 MVT::i32, MVT::i32, MVT::Other, Ops,
1363 array_lengthof(Ops));
1364 cast<MachineSDNode>(ResNode)->setMemRefs(MemOp, MemOp + 1);
1368 SDNode *X86DAGToDAGISel::SelectAtomicLoadAdd(SDNode *Node, EVT NVT) {
1369 if (Node->hasAnyUseOfValue(0))
1372 // Optimize common patterns for __sync_add_and_fetch and
1373 // __sync_sub_and_fetch where the result is not used. This allows us
1374 // to use "lock" version of add, sub, inc, dec instructions.
1375 // FIXME: Do not use special instructions but instead add the "lock"
1376 // prefix to the target node somehow. The extra information will then be
1377 // transferred to machine instruction and it denotes the prefix.
1378 SDValue Chain = Node->getOperand(0);
1379 SDValue Ptr = Node->getOperand(1);
1380 SDValue Val = Node->getOperand(2);
1381 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1382 if (!SelectAddr(Node, Ptr, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4))
1385 bool isInc = false, isDec = false, isSub = false, isCN = false;
1386 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Val);
1389 int64_t CNVal = CN->getSExtValue();
1392 else if (CNVal == -1)
1394 else if (CNVal >= 0)
1395 Val = CurDAG->getTargetConstant(CNVal, NVT);
1398 Val = CurDAG->getTargetConstant(-CNVal, NVT);
1400 } else if (Val.hasOneUse() &&
1401 Val.getOpcode() == ISD::SUB &&
1402 X86::isZeroNode(Val.getOperand(0))) {
1404 Val = Val.getOperand(1);
1408 switch (NVT.getSimpleVT().SimpleTy) {
1412 Opc = X86::LOCK_INC8m;
1414 Opc = X86::LOCK_DEC8m;
1417 Opc = X86::LOCK_SUB8mi;
1419 Opc = X86::LOCK_SUB8mr;
1422 Opc = X86::LOCK_ADD8mi;
1424 Opc = X86::LOCK_ADD8mr;
1429 Opc = X86::LOCK_INC16m;
1431 Opc = X86::LOCK_DEC16m;
1434 if (immSext8(Val.getNode()))
1435 Opc = X86::LOCK_SUB16mi8;
1437 Opc = X86::LOCK_SUB16mi;
1439 Opc = X86::LOCK_SUB16mr;
1442 if (immSext8(Val.getNode()))
1443 Opc = X86::LOCK_ADD16mi8;
1445 Opc = X86::LOCK_ADD16mi;
1447 Opc = X86::LOCK_ADD16mr;
1452 Opc = X86::LOCK_INC32m;
1454 Opc = X86::LOCK_DEC32m;
1457 if (immSext8(Val.getNode()))
1458 Opc = X86::LOCK_SUB32mi8;
1460 Opc = X86::LOCK_SUB32mi;
1462 Opc = X86::LOCK_SUB32mr;
1465 if (immSext8(Val.getNode()))
1466 Opc = X86::LOCK_ADD32mi8;
1468 Opc = X86::LOCK_ADD32mi;
1470 Opc = X86::LOCK_ADD32mr;
1475 Opc = X86::LOCK_INC64m;
1477 Opc = X86::LOCK_DEC64m;
1479 Opc = X86::LOCK_SUB64mr;
1481 if (immSext8(Val.getNode()))
1482 Opc = X86::LOCK_SUB64mi8;
1483 else if (i64immSExt32(Val.getNode()))
1484 Opc = X86::LOCK_SUB64mi32;
1487 Opc = X86::LOCK_ADD64mr;
1489 if (immSext8(Val.getNode()))
1490 Opc = X86::LOCK_ADD64mi8;
1491 else if (i64immSExt32(Val.getNode()))
1492 Opc = X86::LOCK_ADD64mi32;
1498 DebugLoc dl = Node->getDebugLoc();
1499 SDValue Undef = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1501 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
1502 MemOp[0] = cast<MemSDNode>(Node)->getMemOperand();
1503 if (isInc || isDec) {
1504 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Chain };
1505 SDValue Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 6), 0);
1506 cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
1507 SDValue RetVals[] = { Undef, Ret };
1508 return CurDAG->getMergeValues(RetVals, 2, dl).getNode();
1510 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Val, Chain };
1511 SDValue Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 7), 0);
1512 cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
1513 SDValue RetVals[] = { Undef, Ret };
1514 return CurDAG->getMergeValues(RetVals, 2, dl).getNode();
1518 /// HasNoSignedComparisonUses - Test whether the given X86ISD::CMP node has
1519 /// any uses which require the SF or OF bits to be accurate.
1520 static bool HasNoSignedComparisonUses(SDNode *N) {
1521 // Examine each user of the node.
1522 for (SDNode::use_iterator UI = N->use_begin(),
1523 UE = N->use_end(); UI != UE; ++UI) {
1524 // Only examine CopyToReg uses.
1525 if (UI->getOpcode() != ISD::CopyToReg)
1527 // Only examine CopyToReg uses that copy to EFLAGS.
1528 if (cast<RegisterSDNode>(UI->getOperand(1))->getReg() !=
1531 // Examine each user of the CopyToReg use.
1532 for (SDNode::use_iterator FlagUI = UI->use_begin(),
1533 FlagUE = UI->use_end(); FlagUI != FlagUE; ++FlagUI) {
1534 // Only examine the Flag result.
1535 if (FlagUI.getUse().getResNo() != 1) continue;
1536 // Anything unusual: assume conservatively.
1537 if (!FlagUI->isMachineOpcode()) return false;
1538 // Examine the opcode of the user.
1539 switch (FlagUI->getMachineOpcode()) {
1540 // These comparisons don't treat the most significant bit specially.
1541 case X86::SETAr: case X86::SETAEr: case X86::SETBr: case X86::SETBEr:
1542 case X86::SETEr: case X86::SETNEr: case X86::SETPr: case X86::SETNPr:
1543 case X86::SETAm: case X86::SETAEm: case X86::SETBm: case X86::SETBEm:
1544 case X86::SETEm: case X86::SETNEm: case X86::SETPm: case X86::SETNPm:
1545 case X86::JA_4: case X86::JAE_4: case X86::JB_4: case X86::JBE_4:
1546 case X86::JE_4: case X86::JNE_4: case X86::JP_4: case X86::JNP_4:
1547 case X86::CMOVA16rr: case X86::CMOVA16rm:
1548 case X86::CMOVA32rr: case X86::CMOVA32rm:
1549 case X86::CMOVA64rr: case X86::CMOVA64rm:
1550 case X86::CMOVAE16rr: case X86::CMOVAE16rm:
1551 case X86::CMOVAE32rr: case X86::CMOVAE32rm:
1552 case X86::CMOVAE64rr: case X86::CMOVAE64rm:
1553 case X86::CMOVB16rr: case X86::CMOVB16rm:
1554 case X86::CMOVB32rr: case X86::CMOVB32rm:
1555 case X86::CMOVB64rr: case X86::CMOVB64rm:
1556 case X86::CMOVBE16rr: case X86::CMOVBE16rm:
1557 case X86::CMOVBE32rr: case X86::CMOVBE32rm:
1558 case X86::CMOVBE64rr: case X86::CMOVBE64rm:
1559 case X86::CMOVE16rr: case X86::CMOVE16rm:
1560 case X86::CMOVE32rr: case X86::CMOVE32rm:
1561 case X86::CMOVE64rr: case X86::CMOVE64rm:
1562 case X86::CMOVNE16rr: case X86::CMOVNE16rm:
1563 case X86::CMOVNE32rr: case X86::CMOVNE32rm:
1564 case X86::CMOVNE64rr: case X86::CMOVNE64rm:
1565 case X86::CMOVNP16rr: case X86::CMOVNP16rm:
1566 case X86::CMOVNP32rr: case X86::CMOVNP32rm:
1567 case X86::CMOVNP64rr: case X86::CMOVNP64rm:
1568 case X86::CMOVP16rr: case X86::CMOVP16rm:
1569 case X86::CMOVP32rr: case X86::CMOVP32rm:
1570 case X86::CMOVP64rr: case X86::CMOVP64rm:
1572 // Anything else: assume conservatively.
1573 default: return false;
1580 SDNode *X86DAGToDAGISel::Select(SDNode *Node) {
1581 EVT NVT = Node->getValueType(0);
1583 unsigned Opcode = Node->getOpcode();
1584 DebugLoc dl = Node->getDebugLoc();
1586 DEBUG(dbgs() << "Selecting: "; Node->dump(CurDAG); dbgs() << '\n');
1588 if (Node->isMachineOpcode()) {
1589 DEBUG(dbgs() << "== "; Node->dump(CurDAG); dbgs() << '\n');
1590 return NULL; // Already selected.
1595 case X86ISD::GlobalBaseReg:
1596 return getGlobalBaseReg();
1598 case X86ISD::ATOMOR64_DAG:
1599 return SelectAtomic64(Node, X86::ATOMOR6432);
1600 case X86ISD::ATOMXOR64_DAG:
1601 return SelectAtomic64(Node, X86::ATOMXOR6432);
1602 case X86ISD::ATOMADD64_DAG:
1603 return SelectAtomic64(Node, X86::ATOMADD6432);
1604 case X86ISD::ATOMSUB64_DAG:
1605 return SelectAtomic64(Node, X86::ATOMSUB6432);
1606 case X86ISD::ATOMNAND64_DAG:
1607 return SelectAtomic64(Node, X86::ATOMNAND6432);
1608 case X86ISD::ATOMAND64_DAG:
1609 return SelectAtomic64(Node, X86::ATOMAND6432);
1610 case X86ISD::ATOMSWAP64_DAG:
1611 return SelectAtomic64(Node, X86::ATOMSWAP6432);
1613 case ISD::ATOMIC_LOAD_ADD: {
1614 SDNode *RetVal = SelectAtomicLoadAdd(Node, NVT);
1620 case ISD::SMUL_LOHI:
1621 case ISD::UMUL_LOHI: {
1622 SDValue N0 = Node->getOperand(0);
1623 SDValue N1 = Node->getOperand(1);
1625 bool isSigned = Opcode == ISD::SMUL_LOHI;
1627 switch (NVT.getSimpleVT().SimpleTy) {
1628 default: llvm_unreachable("Unsupported VT!");
1629 case MVT::i8: Opc = X86::MUL8r; MOpc = X86::MUL8m; break;
1630 case MVT::i16: Opc = X86::MUL16r; MOpc = X86::MUL16m; break;
1631 case MVT::i32: Opc = X86::MUL32r; MOpc = X86::MUL32m; break;
1632 case MVT::i64: Opc = X86::MUL64r; MOpc = X86::MUL64m; break;
1635 switch (NVT.getSimpleVT().SimpleTy) {
1636 default: llvm_unreachable("Unsupported VT!");
1637 case MVT::i8: Opc = X86::IMUL8r; MOpc = X86::IMUL8m; break;
1638 case MVT::i16: Opc = X86::IMUL16r; MOpc = X86::IMUL16m; break;
1639 case MVT::i32: Opc = X86::IMUL32r; MOpc = X86::IMUL32m; break;
1640 case MVT::i64: Opc = X86::IMUL64r; MOpc = X86::IMUL64m; break;
1644 unsigned LoReg, HiReg;
1645 switch (NVT.getSimpleVT().SimpleTy) {
1646 default: llvm_unreachable("Unsupported VT!");
1647 case MVT::i8: LoReg = X86::AL; HiReg = X86::AH; break;
1648 case MVT::i16: LoReg = X86::AX; HiReg = X86::DX; break;
1649 case MVT::i32: LoReg = X86::EAX; HiReg = X86::EDX; break;
1650 case MVT::i64: LoReg = X86::RAX; HiReg = X86::RDX; break;
1653 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1654 bool foldedLoad = TryFoldLoad(Node, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
1655 // Multiply is commmutative.
1657 foldedLoad = TryFoldLoad(Node, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
1662 SDValue InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, LoReg,
1663 N0, SDValue()).getValue(1);
1666 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
1669 CurDAG->getMachineNode(MOpc, dl, MVT::Other, MVT::Flag, Ops,
1670 array_lengthof(Ops));
1671 InFlag = SDValue(CNode, 1);
1672 // Update the chain.
1673 ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
1676 SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Flag, N1, InFlag), 0);
1679 // Prevent use of AH in a REX instruction by referencing AX instead.
1680 if (HiReg == X86::AH && Subtarget->is64Bit() &&
1681 !SDValue(Node, 1).use_empty()) {
1682 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1683 X86::AX, MVT::i16, InFlag);
1684 InFlag = Result.getValue(2);
1685 // Get the low part if needed. Don't use getCopyFromReg for aliasing
1687 if (!SDValue(Node, 0).use_empty())
1688 ReplaceUses(SDValue(Node, 1),
1689 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
1691 // Shift AX down 8 bits.
1692 Result = SDValue(CurDAG->getMachineNode(X86::SHR16ri, dl, MVT::i16,
1694 CurDAG->getTargetConstant(8, MVT::i8)), 0);
1695 // Then truncate it down to i8.
1696 ReplaceUses(SDValue(Node, 1),
1697 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
1699 // Copy the low half of the result, if it is needed.
1700 if (!SDValue(Node, 0).use_empty()) {
1701 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1702 LoReg, NVT, InFlag);
1703 InFlag = Result.getValue(2);
1704 ReplaceUses(SDValue(Node, 0), Result);
1705 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
1707 // Copy the high half of the result, if it is needed.
1708 if (!SDValue(Node, 1).use_empty()) {
1709 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1710 HiReg, NVT, InFlag);
1711 InFlag = Result.getValue(2);
1712 ReplaceUses(SDValue(Node, 1), Result);
1713 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
1720 case ISD::UDIVREM: {
1721 SDValue N0 = Node->getOperand(0);
1722 SDValue N1 = Node->getOperand(1);
1724 bool isSigned = Opcode == ISD::SDIVREM;
1726 switch (NVT.getSimpleVT().SimpleTy) {
1727 default: llvm_unreachable("Unsupported VT!");
1728 case MVT::i8: Opc = X86::DIV8r; MOpc = X86::DIV8m; break;
1729 case MVT::i16: Opc = X86::DIV16r; MOpc = X86::DIV16m; break;
1730 case MVT::i32: Opc = X86::DIV32r; MOpc = X86::DIV32m; break;
1731 case MVT::i64: Opc = X86::DIV64r; MOpc = X86::DIV64m; break;
1734 switch (NVT.getSimpleVT().SimpleTy) {
1735 default: llvm_unreachable("Unsupported VT!");
1736 case MVT::i8: Opc = X86::IDIV8r; MOpc = X86::IDIV8m; break;
1737 case MVT::i16: Opc = X86::IDIV16r; MOpc = X86::IDIV16m; break;
1738 case MVT::i32: Opc = X86::IDIV32r; MOpc = X86::IDIV32m; break;
1739 case MVT::i64: Opc = X86::IDIV64r; MOpc = X86::IDIV64m; break;
1743 unsigned LoReg, HiReg, ClrReg;
1744 unsigned ClrOpcode, SExtOpcode;
1745 switch (NVT.getSimpleVT().SimpleTy) {
1746 default: llvm_unreachable("Unsupported VT!");
1748 LoReg = X86::AL; ClrReg = HiReg = X86::AH;
1750 SExtOpcode = X86::CBW;
1753 LoReg = X86::AX; HiReg = X86::DX;
1754 ClrOpcode = X86::MOV16r0; ClrReg = X86::DX;
1755 SExtOpcode = X86::CWD;
1758 LoReg = X86::EAX; ClrReg = HiReg = X86::EDX;
1759 ClrOpcode = X86::MOV32r0;
1760 SExtOpcode = X86::CDQ;
1763 LoReg = X86::RAX; ClrReg = HiReg = X86::RDX;
1764 ClrOpcode = X86::MOV64r0;
1765 SExtOpcode = X86::CQO;
1769 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1770 bool foldedLoad = TryFoldLoad(Node, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
1771 bool signBitIsZero = CurDAG->SignBitIsZero(N0);
1774 if (NVT == MVT::i8 && (!isSigned || signBitIsZero)) {
1775 // Special case for div8, just use a move with zero extension to AX to
1776 // clear the upper 8 bits (AH).
1777 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Move, Chain;
1778 if (TryFoldLoad(Node, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4)) {
1779 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N0.getOperand(0) };
1781 SDValue(CurDAG->getMachineNode(X86::MOVZX16rm8, dl, MVT::i16,
1783 array_lengthof(Ops)), 0);
1784 Chain = Move.getValue(1);
1785 ReplaceUses(N0.getValue(1), Chain);
1788 SDValue(CurDAG->getMachineNode(X86::MOVZX16rr8, dl, MVT::i16, N0),0);
1789 Chain = CurDAG->getEntryNode();
1791 Chain = CurDAG->getCopyToReg(Chain, dl, X86::AX, Move, SDValue());
1792 InFlag = Chain.getValue(1);
1795 CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl,
1796 LoReg, N0, SDValue()).getValue(1);
1797 if (isSigned && !signBitIsZero) {
1798 // Sign extend the low part into the high part.
1800 SDValue(CurDAG->getMachineNode(SExtOpcode, dl, MVT::Flag, InFlag),0);
1802 // Zero out the high part, effectively zero extending the input.
1804 SDValue(CurDAG->getMachineNode(ClrOpcode, dl, NVT), 0);
1805 InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, ClrReg,
1806 ClrNode, InFlag).getValue(1);
1811 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
1814 CurDAG->getMachineNode(MOpc, dl, MVT::Other, MVT::Flag, Ops,
1815 array_lengthof(Ops));
1816 InFlag = SDValue(CNode, 1);
1817 // Update the chain.
1818 ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
1821 SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Flag, N1, InFlag), 0);
1824 // Prevent use of AH in a REX instruction by referencing AX instead.
1825 // Shift it down 8 bits.
1826 if (HiReg == X86::AH && Subtarget->is64Bit() &&
1827 !SDValue(Node, 1).use_empty()) {
1828 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1829 X86::AX, MVT::i16, InFlag);
1830 InFlag = Result.getValue(2);
1832 // If we also need AL (the quotient), get it by extracting a subreg from
1833 // Result. The fast register allocator does not like multiple CopyFromReg
1834 // nodes using aliasing registers.
1835 if (!SDValue(Node, 0).use_empty())
1836 ReplaceUses(SDValue(Node, 0),
1837 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
1839 // Shift AX right by 8 bits instead of using AH.
1840 Result = SDValue(CurDAG->getMachineNode(X86::SHR16ri, dl, MVT::i16,
1842 CurDAG->getTargetConstant(8, MVT::i8)),
1844 ReplaceUses(SDValue(Node, 1),
1845 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
1847 // Copy the division (low) result, if it is needed.
1848 if (!SDValue(Node, 0).use_empty()) {
1849 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1850 LoReg, NVT, InFlag);
1851 InFlag = Result.getValue(2);
1852 ReplaceUses(SDValue(Node, 0), Result);
1853 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
1855 // Copy the remainder (high) result, if it is needed.
1856 if (!SDValue(Node, 1).use_empty()) {
1857 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1858 HiReg, NVT, InFlag);
1859 InFlag = Result.getValue(2);
1860 ReplaceUses(SDValue(Node, 1), Result);
1861 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
1867 SDValue N0 = Node->getOperand(0);
1868 SDValue N1 = Node->getOperand(1);
1870 // Look for (X86cmp (and $op, $imm), 0) and see if we can convert it to
1871 // use a smaller encoding.
1872 if (N0.getOpcode() == ISD::TRUNCATE && N0.hasOneUse() &&
1873 HasNoSignedComparisonUses(Node))
1874 // Look past the truncate if CMP is the only use of it.
1875 N0 = N0.getOperand(0);
1876 if (N0.getNode()->getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
1877 N0.getValueType() != MVT::i8 &&
1878 X86::isZeroNode(N1)) {
1879 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getNode()->getOperand(1));
1882 // For example, convert "testl %eax, $8" to "testb %al, $8"
1883 if ((C->getZExtValue() & ~UINT64_C(0xff)) == 0 &&
1884 (!(C->getZExtValue() & 0x80) ||
1885 HasNoSignedComparisonUses(Node))) {
1886 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), MVT::i8);
1887 SDValue Reg = N0.getNode()->getOperand(0);
1889 // On x86-32, only the ABCD registers have 8-bit subregisters.
1890 if (!Subtarget->is64Bit()) {
1891 TargetRegisterClass *TRC = 0;
1892 switch (N0.getValueType().getSimpleVT().SimpleTy) {
1893 case MVT::i32: TRC = &X86::GR32_ABCDRegClass; break;
1894 case MVT::i16: TRC = &X86::GR16_ABCDRegClass; break;
1895 default: llvm_unreachable("Unsupported TEST operand type!");
1897 SDValue RC = CurDAG->getTargetConstant(TRC->getID(), MVT::i32);
1898 Reg = SDValue(CurDAG->getMachineNode(X86::COPY_TO_REGCLASS, dl,
1899 Reg.getValueType(), Reg, RC), 0);
1902 // Extract the l-register.
1903 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl,
1907 return CurDAG->getMachineNode(X86::TEST8ri, dl, MVT::i32, Subreg, Imm);
1910 // For example, "testl %eax, $2048" to "testb %ah, $8".
1911 if ((C->getZExtValue() & ~UINT64_C(0xff00)) == 0 &&
1912 (!(C->getZExtValue() & 0x8000) ||
1913 HasNoSignedComparisonUses(Node))) {
1914 // Shift the immediate right by 8 bits.
1915 SDValue ShiftedImm = CurDAG->getTargetConstant(C->getZExtValue() >> 8,
1917 SDValue Reg = N0.getNode()->getOperand(0);
1919 // Put the value in an ABCD register.
1920 TargetRegisterClass *TRC = 0;
1921 switch (N0.getValueType().getSimpleVT().SimpleTy) {
1922 case MVT::i64: TRC = &X86::GR64_ABCDRegClass; break;
1923 case MVT::i32: TRC = &X86::GR32_ABCDRegClass; break;
1924 case MVT::i16: TRC = &X86::GR16_ABCDRegClass; break;
1925 default: llvm_unreachable("Unsupported TEST operand type!");
1927 SDValue RC = CurDAG->getTargetConstant(TRC->getID(), MVT::i32);
1928 Reg = SDValue(CurDAG->getMachineNode(X86::COPY_TO_REGCLASS, dl,
1929 Reg.getValueType(), Reg, RC), 0);
1931 // Extract the h-register.
1932 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_8bit_hi, dl,
1935 // Emit a testb. No special NOREX tricks are needed since there's
1936 // only one GPR operand!
1937 return CurDAG->getMachineNode(X86::TEST8ri, dl, MVT::i32,
1938 Subreg, ShiftedImm);
1941 // For example, "testl %eax, $32776" to "testw %ax, $32776".
1942 if ((C->getZExtValue() & ~UINT64_C(0xffff)) == 0 &&
1943 N0.getValueType() != MVT::i16 &&
1944 (!(C->getZExtValue() & 0x8000) ||
1945 HasNoSignedComparisonUses(Node))) {
1946 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), MVT::i16);
1947 SDValue Reg = N0.getNode()->getOperand(0);
1949 // Extract the 16-bit subregister.
1950 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_16bit, dl,
1954 return CurDAG->getMachineNode(X86::TEST16ri, dl, MVT::i32, Subreg, Imm);
1957 // For example, "testq %rax, $268468232" to "testl %eax, $268468232".
1958 if ((C->getZExtValue() & ~UINT64_C(0xffffffff)) == 0 &&
1959 N0.getValueType() == MVT::i64 &&
1960 (!(C->getZExtValue() & 0x80000000) ||
1961 HasNoSignedComparisonUses(Node))) {
1962 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), MVT::i32);
1963 SDValue Reg = N0.getNode()->getOperand(0);
1965 // Extract the 32-bit subregister.
1966 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_32bit, dl,
1970 return CurDAG->getMachineNode(X86::TEST32ri, dl, MVT::i32, Subreg, Imm);
1977 SDNode *ResNode = SelectCode(Node);
1979 DEBUG(dbgs() << "=> ";
1980 if (ResNode == NULL || ResNode == Node)
1983 ResNode->dump(CurDAG);
1989 bool X86DAGToDAGISel::
1990 SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
1991 std::vector<SDValue> &OutOps) {
1992 SDValue Op0, Op1, Op2, Op3, Op4;
1993 switch (ConstraintCode) {
1994 case 'o': // offsetable ??
1995 case 'v': // not offsetable ??
1996 default: return true;
1998 if (!SelectAddr(0, Op, Op0, Op1, Op2, Op3, Op4))
2003 OutOps.push_back(Op0);
2004 OutOps.push_back(Op1);
2005 OutOps.push_back(Op2);
2006 OutOps.push_back(Op3);
2007 OutOps.push_back(Op4);
2011 /// createX86ISelDag - This pass converts a legalized DAG into a
2012 /// X86-specific DAG, ready for instruction scheduling.
2014 FunctionPass *llvm::createX86ISelDag(X86TargetMachine &TM,
2015 llvm::CodeGenOpt::Level OptLevel) {
2016 return new X86DAGToDAGISel(TM, OptLevel);