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 // FIXME: Calls can't fold loads through segment registers yet.
407 if (LD->getPointerInfo().getAddrSpace() > 255)
410 // Now let's find the callseq_start.
411 while (HasCallSeq && Chain.getOpcode() != ISD::CALLSEQ_START) {
412 if (!Chain.hasOneUse())
414 Chain = Chain.getOperand(0);
417 if (!Chain.getNumOperands())
419 if (Chain.getOperand(0).getNode() == Callee.getNode())
421 if (Chain.getOperand(0).getOpcode() == ISD::TokenFactor &&
422 Callee.getValue(1).isOperandOf(Chain.getOperand(0).getNode()) &&
423 Callee.getValue(1).hasOneUse())
428 void X86DAGToDAGISel::PreprocessISelDAG() {
429 // OptForSize is used in pattern predicates that isel is matching.
430 OptForSize = MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize);
432 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
433 E = CurDAG->allnodes_end(); I != E; ) {
434 SDNode *N = I++; // Preincrement iterator to avoid invalidation issues.
436 if (OptLevel != CodeGenOpt::None &&
437 (N->getOpcode() == X86ISD::CALL ||
438 N->getOpcode() == X86ISD::TC_RETURN)) {
439 /// Also try moving call address load from outside callseq_start to just
440 /// before the call to allow it to be folded.
458 bool HasCallSeq = N->getOpcode() == X86ISD::CALL;
459 SDValue Chain = N->getOperand(0);
460 SDValue Load = N->getOperand(1);
461 if (!isCalleeLoad(Load, Chain, HasCallSeq))
463 MoveBelowOrigChain(CurDAG, Load, SDValue(N, 0), Chain);
468 // Lower fpround and fpextend nodes that target the FP stack to be store and
469 // load to the stack. This is a gross hack. We would like to simply mark
470 // these as being illegal, but when we do that, legalize produces these when
471 // it expands calls, then expands these in the same legalize pass. We would
472 // like dag combine to be able to hack on these between the call expansion
473 // and the node legalization. As such this pass basically does "really
474 // late" legalization of these inline with the X86 isel pass.
475 // FIXME: This should only happen when not compiled with -O0.
476 if (N->getOpcode() != ISD::FP_ROUND && N->getOpcode() != ISD::FP_EXTEND)
479 // If the source and destination are SSE registers, then this is a legal
480 // conversion that should not be lowered.
481 EVT SrcVT = N->getOperand(0).getValueType();
482 EVT DstVT = N->getValueType(0);
483 bool SrcIsSSE = X86Lowering.isScalarFPTypeInSSEReg(SrcVT);
484 bool DstIsSSE = X86Lowering.isScalarFPTypeInSSEReg(DstVT);
485 if (SrcIsSSE && DstIsSSE)
488 if (!SrcIsSSE && !DstIsSSE) {
489 // If this is an FPStack extension, it is a noop.
490 if (N->getOpcode() == ISD::FP_EXTEND)
492 // If this is a value-preserving FPStack truncation, it is a noop.
493 if (N->getConstantOperandVal(1))
497 // Here we could have an FP stack truncation or an FPStack <-> SSE convert.
498 // FPStack has extload and truncstore. SSE can fold direct loads into other
499 // operations. Based on this, decide what we want to do.
501 if (N->getOpcode() == ISD::FP_ROUND)
502 MemVT = DstVT; // FP_ROUND must use DstVT, we can't do a 'trunc load'.
504 MemVT = SrcIsSSE ? SrcVT : DstVT;
506 SDValue MemTmp = CurDAG->CreateStackTemporary(MemVT);
507 DebugLoc dl = N->getDebugLoc();
509 // FIXME: optimize the case where the src/dest is a load or store?
510 SDValue Store = CurDAG->getTruncStore(CurDAG->getEntryNode(), dl,
512 MemTmp, MachinePointerInfo(), MemVT,
514 SDValue Result = CurDAG->getExtLoad(ISD::EXTLOAD, DstVT, dl, Store, MemTmp,
515 MachinePointerInfo(),
516 MemVT, false, false, 0);
518 // We're about to replace all uses of the FP_ROUND/FP_EXTEND with the
519 // extload we created. This will cause general havok on the dag because
520 // anything below the conversion could be folded into other existing nodes.
521 // To avoid invalidating 'I', back it up to the convert node.
523 CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
525 // Now that we did that, the node is dead. Increment the iterator to the
526 // next node to process, then delete N.
528 CurDAG->DeleteNode(N);
533 /// EmitSpecialCodeForMain - Emit any code that needs to be executed only in
534 /// the main function.
535 void X86DAGToDAGISel::EmitSpecialCodeForMain(MachineBasicBlock *BB,
536 MachineFrameInfo *MFI) {
537 const TargetInstrInfo *TII = TM.getInstrInfo();
538 if (Subtarget->isTargetCygMing())
539 BuildMI(BB, DebugLoc(),
540 TII->get(X86::CALLpcrel32)).addExternalSymbol("__main");
543 void X86DAGToDAGISel::EmitFunctionEntryCode() {
544 // If this is main, emit special code for main.
545 if (const Function *Fn = MF->getFunction())
546 if (Fn->hasExternalLinkage() && Fn->getName() == "main")
547 EmitSpecialCodeForMain(MF->begin(), MF->getFrameInfo());
551 bool X86DAGToDAGISel::MatchSegmentBaseAddress(SDValue N,
552 X86ISelAddressMode &AM) {
553 assert(N.getOpcode() == X86ISD::SegmentBaseAddress);
554 SDValue Segment = N.getOperand(0);
556 if (AM.Segment.getNode() == 0) {
557 AM.Segment = Segment;
564 bool X86DAGToDAGISel::MatchLoad(SDValue N, X86ISelAddressMode &AM) {
565 // This optimization is valid because the GNU TLS model defines that
566 // gs:0 (or fs:0 on X86-64) contains its own address.
567 // For more information see http://people.redhat.com/drepper/tls.pdf
569 SDValue Address = N.getOperand(1);
570 if (Address.getOpcode() == X86ISD::SegmentBaseAddress &&
571 !MatchSegmentBaseAddress(Address, AM))
577 /// MatchWrapper - Try to match X86ISD::Wrapper and X86ISD::WrapperRIP nodes
578 /// into an addressing mode. These wrap things that will resolve down into a
579 /// symbol reference. If no match is possible, this returns true, otherwise it
581 bool X86DAGToDAGISel::MatchWrapper(SDValue N, X86ISelAddressMode &AM) {
582 // If the addressing mode already has a symbol as the displacement, we can
583 // never match another symbol.
584 if (AM.hasSymbolicDisplacement())
587 SDValue N0 = N.getOperand(0);
588 CodeModel::Model M = TM.getCodeModel();
590 // Handle X86-64 rip-relative addresses. We check this before checking direct
591 // folding because RIP is preferable to non-RIP accesses.
592 if (Subtarget->is64Bit() &&
593 // Under X86-64 non-small code model, GV (and friends) are 64-bits, so
594 // they cannot be folded into immediate fields.
595 // FIXME: This can be improved for kernel and other models?
596 (M == CodeModel::Small || M == CodeModel::Kernel) &&
597 // Base and index reg must be 0 in order to use %rip as base and lowering
599 !AM.hasBaseOrIndexReg() && N.getOpcode() == X86ISD::WrapperRIP) {
600 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(N0)) {
601 int64_t Offset = AM.Disp + G->getOffset();
602 if (!X86::isOffsetSuitableForCodeModel(Offset, M)) return true;
603 AM.GV = G->getGlobal();
605 AM.SymbolFlags = G->getTargetFlags();
606 } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N0)) {
607 int64_t Offset = AM.Disp + CP->getOffset();
608 if (!X86::isOffsetSuitableForCodeModel(Offset, M)) return true;
609 AM.CP = CP->getConstVal();
610 AM.Align = CP->getAlignment();
612 AM.SymbolFlags = CP->getTargetFlags();
613 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(N0)) {
614 AM.ES = S->getSymbol();
615 AM.SymbolFlags = S->getTargetFlags();
616 } else if (JumpTableSDNode *J = dyn_cast<JumpTableSDNode>(N0)) {
617 AM.JT = J->getIndex();
618 AM.SymbolFlags = J->getTargetFlags();
620 AM.BlockAddr = cast<BlockAddressSDNode>(N0)->getBlockAddress();
621 AM.SymbolFlags = cast<BlockAddressSDNode>(N0)->getTargetFlags();
624 if (N.getOpcode() == X86ISD::WrapperRIP)
625 AM.setBaseReg(CurDAG->getRegister(X86::RIP, MVT::i64));
629 // Handle the case when globals fit in our immediate field: This is true for
630 // X86-32 always and X86-64 when in -static -mcmodel=small mode. In 64-bit
631 // mode, this results in a non-RIP-relative computation.
632 if (!Subtarget->is64Bit() ||
633 ((M == CodeModel::Small || M == CodeModel::Kernel) &&
634 TM.getRelocationModel() == Reloc::Static)) {
635 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(N0)) {
636 AM.GV = G->getGlobal();
637 AM.Disp += G->getOffset();
638 AM.SymbolFlags = G->getTargetFlags();
639 } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N0)) {
640 AM.CP = CP->getConstVal();
641 AM.Align = CP->getAlignment();
642 AM.Disp += CP->getOffset();
643 AM.SymbolFlags = CP->getTargetFlags();
644 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(N0)) {
645 AM.ES = S->getSymbol();
646 AM.SymbolFlags = S->getTargetFlags();
647 } else if (JumpTableSDNode *J = dyn_cast<JumpTableSDNode>(N0)) {
648 AM.JT = J->getIndex();
649 AM.SymbolFlags = J->getTargetFlags();
651 AM.BlockAddr = cast<BlockAddressSDNode>(N0)->getBlockAddress();
652 AM.SymbolFlags = cast<BlockAddressSDNode>(N0)->getTargetFlags();
660 /// MatchAddress - Add the specified node to the specified addressing mode,
661 /// returning true if it cannot be done. This just pattern matches for the
663 bool X86DAGToDAGISel::MatchAddress(SDValue N, X86ISelAddressMode &AM) {
664 if (MatchAddressRecursively(N, AM, 0))
667 // Post-processing: Convert lea(,%reg,2) to lea(%reg,%reg), which has
668 // a smaller encoding and avoids a scaled-index.
670 AM.BaseType == X86ISelAddressMode::RegBase &&
671 AM.Base_Reg.getNode() == 0) {
672 AM.Base_Reg = AM.IndexReg;
676 // Post-processing: Convert foo to foo(%rip), even in non-PIC mode,
677 // because it has a smaller encoding.
678 // TODO: Which other code models can use this?
679 if (TM.getCodeModel() == CodeModel::Small &&
680 Subtarget->is64Bit() &&
682 AM.BaseType == X86ISelAddressMode::RegBase &&
683 AM.Base_Reg.getNode() == 0 &&
684 AM.IndexReg.getNode() == 0 &&
685 AM.SymbolFlags == X86II::MO_NO_FLAG &&
686 AM.hasSymbolicDisplacement())
687 AM.Base_Reg = CurDAG->getRegister(X86::RIP, MVT::i64);
692 /// isLogicallyAddWithConstant - Return true if this node is semantically an
693 /// add of a value with a constantint.
694 static bool isLogicallyAddWithConstant(SDValue V, SelectionDAG *CurDAG) {
695 // Check for (add x, Cst)
696 if (V->getOpcode() == ISD::ADD)
697 return isa<ConstantSDNode>(V->getOperand(1));
699 // Check for (or x, Cst), where Cst & x == 0.
700 if (V->getOpcode() != ISD::OR ||
701 !isa<ConstantSDNode>(V->getOperand(1)))
704 // Handle "X | C" as "X + C" iff X is known to have C bits clear.
705 ConstantSDNode *CN = cast<ConstantSDNode>(V->getOperand(1));
707 // Check to see if the LHS & C is zero.
708 return CurDAG->MaskedValueIsZero(V->getOperand(0), CN->getAPIntValue());
711 bool X86DAGToDAGISel::MatchAddressRecursively(SDValue N, X86ISelAddressMode &AM,
713 bool is64Bit = Subtarget->is64Bit();
714 DebugLoc dl = N.getDebugLoc();
716 dbgs() << "MatchAddress: ";
721 return MatchAddressBase(N, AM);
723 CodeModel::Model M = TM.getCodeModel();
725 // If this is already a %rip relative address, we can only merge immediates
726 // into it. Instead of handling this in every case, we handle it here.
727 // RIP relative addressing: %rip + 32-bit displacement!
728 if (AM.isRIPRelative()) {
729 // FIXME: JumpTable and ExternalSymbol address currently don't like
730 // displacements. It isn't very important, but this should be fixed for
732 if (!AM.ES && AM.JT != -1) return true;
734 if (ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(N)) {
735 int64_t Val = AM.Disp + Cst->getSExtValue();
736 if (X86::isOffsetSuitableForCodeModel(Val, M,
737 AM.hasSymbolicDisplacement())) {
745 switch (N.getOpcode()) {
747 case ISD::Constant: {
748 uint64_t Val = cast<ConstantSDNode>(N)->getSExtValue();
750 X86::isOffsetSuitableForCodeModel(AM.Disp + Val, M,
751 AM.hasSymbolicDisplacement())) {
758 case X86ISD::SegmentBaseAddress:
759 if (!MatchSegmentBaseAddress(N, AM))
763 case X86ISD::Wrapper:
764 case X86ISD::WrapperRIP:
765 if (!MatchWrapper(N, AM))
770 if (!MatchLoad(N, AM))
774 case ISD::FrameIndex:
775 if (AM.BaseType == X86ISelAddressMode::RegBase
776 && AM.Base_Reg.getNode() == 0) {
777 AM.BaseType = X86ISelAddressMode::FrameIndexBase;
778 AM.Base_FrameIndex = cast<FrameIndexSDNode>(N)->getIndex();
784 if (AM.IndexReg.getNode() != 0 || AM.Scale != 1)
788 *CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1))) {
789 unsigned Val = CN->getZExtValue();
790 // Note that we handle x<<1 as (,x,2) rather than (x,x) here so
791 // that the base operand remains free for further matching. If
792 // the base doesn't end up getting used, a post-processing step
793 // in MatchAddress turns (,x,2) into (x,x), which is cheaper.
794 if (Val == 1 || Val == 2 || Val == 3) {
796 SDValue ShVal = N.getNode()->getOperand(0);
798 // Okay, we know that we have a scale by now. However, if the scaled
799 // value is an add of something and a constant, we can fold the
800 // constant into the disp field here.
801 if (isLogicallyAddWithConstant(ShVal, CurDAG)) {
802 AM.IndexReg = ShVal.getNode()->getOperand(0);
803 ConstantSDNode *AddVal =
804 cast<ConstantSDNode>(ShVal.getNode()->getOperand(1));
805 uint64_t Disp = AM.Disp + (AddVal->getSExtValue() << Val);
807 X86::isOffsetSuitableForCodeModel(Disp, M,
808 AM.hasSymbolicDisplacement()))
822 // A mul_lohi where we need the low part can be folded as a plain multiply.
823 if (N.getResNo() != 0) break;
826 case X86ISD::MUL_IMM:
827 // X*[3,5,9] -> X+X*[2,4,8]
828 if (AM.BaseType == X86ISelAddressMode::RegBase &&
829 AM.Base_Reg.getNode() == 0 &&
830 AM.IndexReg.getNode() == 0) {
832 *CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1)))
833 if (CN->getZExtValue() == 3 || CN->getZExtValue() == 5 ||
834 CN->getZExtValue() == 9) {
835 AM.Scale = unsigned(CN->getZExtValue())-1;
837 SDValue MulVal = N.getNode()->getOperand(0);
840 // Okay, we know that we have a scale by now. However, if the scaled
841 // value is an add of something and a constant, we can fold the
842 // constant into the disp field here.
843 if (MulVal.getNode()->getOpcode() == ISD::ADD && MulVal.hasOneUse() &&
844 isa<ConstantSDNode>(MulVal.getNode()->getOperand(1))) {
845 Reg = MulVal.getNode()->getOperand(0);
846 ConstantSDNode *AddVal =
847 cast<ConstantSDNode>(MulVal.getNode()->getOperand(1));
848 uint64_t Disp = AM.Disp + AddVal->getSExtValue() *
851 X86::isOffsetSuitableForCodeModel(Disp, M,
852 AM.hasSymbolicDisplacement()))
855 Reg = N.getNode()->getOperand(0);
857 Reg = N.getNode()->getOperand(0);
860 AM.IndexReg = AM.Base_Reg = Reg;
867 // Given A-B, if A can be completely folded into the address and
868 // the index field with the index field unused, use -B as the index.
869 // This is a win if a has multiple parts that can be folded into
870 // the address. Also, this saves a mov if the base register has
871 // other uses, since it avoids a two-address sub instruction, however
872 // it costs an additional mov if the index register has other uses.
874 // Add an artificial use to this node so that we can keep track of
875 // it if it gets CSE'd with a different node.
876 HandleSDNode Handle(N);
878 // Test if the LHS of the sub can be folded.
879 X86ISelAddressMode Backup = AM;
880 if (MatchAddressRecursively(N.getNode()->getOperand(0), AM, Depth+1)) {
884 // Test if the index field is free for use.
885 if (AM.IndexReg.getNode() || AM.isRIPRelative()) {
891 SDValue RHS = Handle.getValue().getNode()->getOperand(1);
892 // If the RHS involves a register with multiple uses, this
893 // transformation incurs an extra mov, due to the neg instruction
894 // clobbering its operand.
895 if (!RHS.getNode()->hasOneUse() ||
896 RHS.getNode()->getOpcode() == ISD::CopyFromReg ||
897 RHS.getNode()->getOpcode() == ISD::TRUNCATE ||
898 RHS.getNode()->getOpcode() == ISD::ANY_EXTEND ||
899 (RHS.getNode()->getOpcode() == ISD::ZERO_EXTEND &&
900 RHS.getNode()->getOperand(0).getValueType() == MVT::i32))
902 // If the base is a register with multiple uses, this
903 // transformation may save a mov.
904 if ((AM.BaseType == X86ISelAddressMode::RegBase &&
905 AM.Base_Reg.getNode() &&
906 !AM.Base_Reg.getNode()->hasOneUse()) ||
907 AM.BaseType == X86ISelAddressMode::FrameIndexBase)
909 // If the folded LHS was interesting, this transformation saves
910 // address arithmetic.
911 if ((AM.hasSymbolicDisplacement() && !Backup.hasSymbolicDisplacement()) +
912 ((AM.Disp != 0) && (Backup.Disp == 0)) +
913 (AM.Segment.getNode() && !Backup.Segment.getNode()) >= 2)
915 // If it doesn't look like it may be an overall win, don't do it.
921 // Ok, the transformation is legal and appears profitable. Go for it.
922 SDValue Zero = CurDAG->getConstant(0, N.getValueType());
923 SDValue Neg = CurDAG->getNode(ISD::SUB, dl, N.getValueType(), Zero, RHS);
927 // Insert the new nodes into the topological ordering.
928 if (Zero.getNode()->getNodeId() == -1 ||
929 Zero.getNode()->getNodeId() > N.getNode()->getNodeId()) {
930 CurDAG->RepositionNode(N.getNode(), Zero.getNode());
931 Zero.getNode()->setNodeId(N.getNode()->getNodeId());
933 if (Neg.getNode()->getNodeId() == -1 ||
934 Neg.getNode()->getNodeId() > N.getNode()->getNodeId()) {
935 CurDAG->RepositionNode(N.getNode(), Neg.getNode());
936 Neg.getNode()->setNodeId(N.getNode()->getNodeId());
942 // Add an artificial use to this node so that we can keep track of
943 // it if it gets CSE'd with a different node.
944 HandleSDNode Handle(N);
945 SDValue LHS = Handle.getValue().getNode()->getOperand(0);
946 SDValue RHS = Handle.getValue().getNode()->getOperand(1);
948 X86ISelAddressMode Backup = AM;
949 if (!MatchAddressRecursively(LHS, AM, Depth+1) &&
950 !MatchAddressRecursively(RHS, AM, Depth+1))
953 LHS = Handle.getValue().getNode()->getOperand(0);
954 RHS = Handle.getValue().getNode()->getOperand(1);
956 // Try again after commuting the operands.
957 if (!MatchAddressRecursively(RHS, AM, Depth+1) &&
958 !MatchAddressRecursively(LHS, AM, Depth+1))
961 LHS = Handle.getValue().getNode()->getOperand(0);
962 RHS = Handle.getValue().getNode()->getOperand(1);
964 // If we couldn't fold both operands into the address at the same time,
965 // see if we can just put each operand into a register and fold at least
967 if (AM.BaseType == X86ISelAddressMode::RegBase &&
968 !AM.Base_Reg.getNode() &&
969 !AM.IndexReg.getNode()) {
979 // Handle "X | C" as "X + C" iff X is known to have C bits clear.
980 if (isLogicallyAddWithConstant(N, CurDAG)) {
981 X86ISelAddressMode Backup = AM;
982 ConstantSDNode *CN = cast<ConstantSDNode>(N.getOperand(1));
983 uint64_t Offset = CN->getSExtValue();
985 // Start with the LHS as an addr mode.
986 if (!MatchAddressRecursively(N.getOperand(0), AM, Depth+1) &&
987 // Address could not have picked a GV address for the displacement.
989 // On x86-64, the resultant disp must fit in 32-bits.
991 X86::isOffsetSuitableForCodeModel(AM.Disp + Offset, M,
992 AM.hasSymbolicDisplacement()))) {
1001 // Perform some heroic transforms on an and of a constant-count shift
1002 // with a constant to enable use of the scaled offset field.
1004 SDValue Shift = N.getOperand(0);
1005 if (Shift.getNumOperands() != 2) break;
1007 // Scale must not be used already.
1008 if (AM.IndexReg.getNode() != 0 || AM.Scale != 1) break;
1010 SDValue X = Shift.getOperand(0);
1011 ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N.getOperand(1));
1012 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(Shift.getOperand(1));
1013 if (!C1 || !C2) break;
1015 // Handle "(X >> (8-C1)) & C2" as "(X >> 8) & 0xff)" if safe. This
1016 // allows us to convert the shift and and into an h-register extract and
1018 if (Shift.getOpcode() == ISD::SRL && Shift.hasOneUse()) {
1019 unsigned ScaleLog = 8 - C1->getZExtValue();
1020 if (ScaleLog > 0 && ScaleLog < 4 &&
1021 C2->getZExtValue() == (UINT64_C(0xff) << ScaleLog)) {
1022 SDValue Eight = CurDAG->getConstant(8, MVT::i8);
1023 SDValue Mask = CurDAG->getConstant(0xff, N.getValueType());
1024 SDValue Srl = CurDAG->getNode(ISD::SRL, dl, N.getValueType(),
1026 SDValue And = CurDAG->getNode(ISD::AND, dl, N.getValueType(),
1028 SDValue ShlCount = CurDAG->getConstant(ScaleLog, MVT::i8);
1029 SDValue Shl = CurDAG->getNode(ISD::SHL, dl, N.getValueType(),
1032 // Insert the new nodes into the topological ordering.
1033 if (Eight.getNode()->getNodeId() == -1 ||
1034 Eight.getNode()->getNodeId() > X.getNode()->getNodeId()) {
1035 CurDAG->RepositionNode(X.getNode(), Eight.getNode());
1036 Eight.getNode()->setNodeId(X.getNode()->getNodeId());
1038 if (Mask.getNode()->getNodeId() == -1 ||
1039 Mask.getNode()->getNodeId() > X.getNode()->getNodeId()) {
1040 CurDAG->RepositionNode(X.getNode(), Mask.getNode());
1041 Mask.getNode()->setNodeId(X.getNode()->getNodeId());
1043 if (Srl.getNode()->getNodeId() == -1 ||
1044 Srl.getNode()->getNodeId() > Shift.getNode()->getNodeId()) {
1045 CurDAG->RepositionNode(Shift.getNode(), Srl.getNode());
1046 Srl.getNode()->setNodeId(Shift.getNode()->getNodeId());
1048 if (And.getNode()->getNodeId() == -1 ||
1049 And.getNode()->getNodeId() > N.getNode()->getNodeId()) {
1050 CurDAG->RepositionNode(N.getNode(), And.getNode());
1051 And.getNode()->setNodeId(N.getNode()->getNodeId());
1053 if (ShlCount.getNode()->getNodeId() == -1 ||
1054 ShlCount.getNode()->getNodeId() > X.getNode()->getNodeId()) {
1055 CurDAG->RepositionNode(X.getNode(), ShlCount.getNode());
1056 ShlCount.getNode()->setNodeId(N.getNode()->getNodeId());
1058 if (Shl.getNode()->getNodeId() == -1 ||
1059 Shl.getNode()->getNodeId() > N.getNode()->getNodeId()) {
1060 CurDAG->RepositionNode(N.getNode(), Shl.getNode());
1061 Shl.getNode()->setNodeId(N.getNode()->getNodeId());
1063 CurDAG->ReplaceAllUsesWith(N, Shl);
1065 AM.Scale = (1 << ScaleLog);
1070 // Handle "(X << C1) & C2" as "(X & (C2>>C1)) << C1" if safe and if this
1071 // allows us to fold the shift into this addressing mode.
1072 if (Shift.getOpcode() != ISD::SHL) break;
1074 // Not likely to be profitable if either the AND or SHIFT node has more
1075 // than one use (unless all uses are for address computation). Besides,
1076 // isel mechanism requires their node ids to be reused.
1077 if (!N.hasOneUse() || !Shift.hasOneUse())
1080 // Verify that the shift amount is something we can fold.
1081 unsigned ShiftCst = C1->getZExtValue();
1082 if (ShiftCst != 1 && ShiftCst != 2 && ShiftCst != 3)
1085 // Get the new AND mask, this folds to a constant.
1086 SDValue NewANDMask = CurDAG->getNode(ISD::SRL, dl, N.getValueType(),
1087 SDValue(C2, 0), SDValue(C1, 0));
1088 SDValue NewAND = CurDAG->getNode(ISD::AND, dl, N.getValueType(), X,
1090 SDValue NewSHIFT = CurDAG->getNode(ISD::SHL, dl, N.getValueType(),
1091 NewAND, SDValue(C1, 0));
1093 // Insert the new nodes into the topological ordering.
1094 if (C1->getNodeId() > X.getNode()->getNodeId()) {
1095 CurDAG->RepositionNode(X.getNode(), C1);
1096 C1->setNodeId(X.getNode()->getNodeId());
1098 if (NewANDMask.getNode()->getNodeId() == -1 ||
1099 NewANDMask.getNode()->getNodeId() > X.getNode()->getNodeId()) {
1100 CurDAG->RepositionNode(X.getNode(), NewANDMask.getNode());
1101 NewANDMask.getNode()->setNodeId(X.getNode()->getNodeId());
1103 if (NewAND.getNode()->getNodeId() == -1 ||
1104 NewAND.getNode()->getNodeId() > Shift.getNode()->getNodeId()) {
1105 CurDAG->RepositionNode(Shift.getNode(), NewAND.getNode());
1106 NewAND.getNode()->setNodeId(Shift.getNode()->getNodeId());
1108 if (NewSHIFT.getNode()->getNodeId() == -1 ||
1109 NewSHIFT.getNode()->getNodeId() > N.getNode()->getNodeId()) {
1110 CurDAG->RepositionNode(N.getNode(), NewSHIFT.getNode());
1111 NewSHIFT.getNode()->setNodeId(N.getNode()->getNodeId());
1114 CurDAG->ReplaceAllUsesWith(N, NewSHIFT);
1116 AM.Scale = 1 << ShiftCst;
1117 AM.IndexReg = NewAND;
1122 return MatchAddressBase(N, AM);
1125 /// MatchAddressBase - Helper for MatchAddress. Add the specified node to the
1126 /// specified addressing mode without any further recursion.
1127 bool X86DAGToDAGISel::MatchAddressBase(SDValue N, X86ISelAddressMode &AM) {
1128 // Is the base register already occupied?
1129 if (AM.BaseType != X86ISelAddressMode::RegBase || AM.Base_Reg.getNode()) {
1130 // If so, check to see if the scale index register is set.
1131 if (AM.IndexReg.getNode() == 0) {
1137 // Otherwise, we cannot select it.
1141 // Default, generate it as a register.
1142 AM.BaseType = X86ISelAddressMode::RegBase;
1147 /// SelectAddr - returns true if it is able pattern match an addressing mode.
1148 /// It returns the operands which make up the maximal addressing mode it can
1149 /// match by reference.
1151 /// Parent is the parent node of the addr operand that is being matched. It
1152 /// is always a load, store, atomic node, or null. It is only null when
1153 /// checking memory operands for inline asm nodes.
1154 bool X86DAGToDAGISel::SelectAddr(SDNode *Parent, SDValue N, SDValue &Base,
1155 SDValue &Scale, SDValue &Index,
1156 SDValue &Disp, SDValue &Segment) {
1157 X86ISelAddressMode AM;
1158 if (MatchAddress(N, AM))
1161 EVT VT = N.getValueType();
1162 if (AM.BaseType == X86ISelAddressMode::RegBase) {
1163 if (!AM.Base_Reg.getNode())
1164 AM.Base_Reg = CurDAG->getRegister(0, VT);
1167 if (!AM.IndexReg.getNode())
1168 AM.IndexReg = CurDAG->getRegister(0, VT);
1171 // This list of opcodes are all the nodes that have an "addr:$ptr" operand
1172 // that are not a MemSDNode, and thus don't have proper addrspace info.
1173 Parent->getOpcode() != ISD::PREFETCH &&
1174 Parent->getOpcode() != ISD::INTRINSIC_W_CHAIN && // unaligned loads, fixme
1175 Parent->getOpcode() != ISD::INTRINSIC_VOID && // nontemporal stores.
1176 Parent->getOpcode() != X86ISD::VZEXT_LOAD &&
1177 Parent->getOpcode() != X86ISD::FLD &&
1178 Parent->getOpcode() != X86ISD::FILD &&
1179 Parent->getOpcode() != X86ISD::FILD_FLAG &&
1180 Parent->getOpcode() != X86ISD::FP_TO_INT16_IN_MEM &&
1181 Parent->getOpcode() != X86ISD::FP_TO_INT32_IN_MEM &&
1182 Parent->getOpcode() != X86ISD::FP_TO_INT64_IN_MEM &&
1183 Parent->getOpcode() != X86ISD::FST) {
1184 unsigned AddrSpace =
1185 cast<MemSDNode>(Parent)->getPointerInfo().getAddrSpace();
1186 // AddrSpace 256 -> GS, 257 -> FS.
1187 if (AddrSpace == 256)
1188 AM.Segment = CurDAG->getRegister(X86::GS, VT);
1189 if (AddrSpace == 257)
1190 AM.Segment = CurDAG->getRegister(X86::FS, VT);
1194 getAddressOperands(AM, Base, Scale, Index, Disp, Segment);
1198 /// SelectScalarSSELoad - Match a scalar SSE load. In particular, we want to
1199 /// match a load whose top elements are either undef or zeros. The load flavor
1200 /// is derived from the type of N, which is either v4f32 or v2f64.
1203 /// PatternChainNode: this is the matched node that has a chain input and
1205 bool X86DAGToDAGISel::SelectScalarSSELoad(SDNode *Root,
1206 SDValue N, SDValue &Base,
1207 SDValue &Scale, SDValue &Index,
1208 SDValue &Disp, SDValue &Segment,
1209 SDValue &PatternNodeWithChain) {
1210 if (N.getOpcode() == ISD::SCALAR_TO_VECTOR) {
1211 PatternNodeWithChain = N.getOperand(0);
1212 if (ISD::isNON_EXTLoad(PatternNodeWithChain.getNode()) &&
1213 PatternNodeWithChain.hasOneUse() &&
1214 IsProfitableToFold(N.getOperand(0), N.getNode(), Root) &&
1215 IsLegalToFold(N.getOperand(0), N.getNode(), Root, OptLevel)) {
1216 LoadSDNode *LD = cast<LoadSDNode>(PatternNodeWithChain);
1217 if (!SelectAddr(LD, LD->getBasePtr(), Base, Scale, Index, Disp, Segment))
1223 // Also handle the case where we explicitly require zeros in the top
1224 // elements. This is a vector shuffle from the zero vector.
1225 if (N.getOpcode() == X86ISD::VZEXT_MOVL && N.getNode()->hasOneUse() &&
1226 // Check to see if the top elements are all zeros (or bitcast of zeros).
1227 N.getOperand(0).getOpcode() == ISD::SCALAR_TO_VECTOR &&
1228 N.getOperand(0).getNode()->hasOneUse() &&
1229 ISD::isNON_EXTLoad(N.getOperand(0).getOperand(0).getNode()) &&
1230 N.getOperand(0).getOperand(0).hasOneUse() &&
1231 IsProfitableToFold(N.getOperand(0), N.getNode(), Root) &&
1232 IsLegalToFold(N.getOperand(0), N.getNode(), Root, OptLevel)) {
1233 // Okay, this is a zero extending load. Fold it.
1234 LoadSDNode *LD = cast<LoadSDNode>(N.getOperand(0).getOperand(0));
1235 if (!SelectAddr(LD, LD->getBasePtr(), Base, Scale, Index, Disp, Segment))
1237 PatternNodeWithChain = SDValue(LD, 0);
1244 /// SelectLEAAddr - it calls SelectAddr and determines if the maximal addressing
1245 /// mode it matches can be cost effectively emitted as an LEA instruction.
1246 bool X86DAGToDAGISel::SelectLEAAddr(SDValue N,
1247 SDValue &Base, SDValue &Scale,
1248 SDValue &Index, SDValue &Disp,
1250 X86ISelAddressMode AM;
1252 // Set AM.Segment to prevent MatchAddress from using one. LEA doesn't support
1254 SDValue Copy = AM.Segment;
1255 SDValue T = CurDAG->getRegister(0, MVT::i32);
1257 if (MatchAddress(N, AM))
1259 assert (T == AM.Segment);
1262 EVT VT = N.getValueType();
1263 unsigned Complexity = 0;
1264 if (AM.BaseType == X86ISelAddressMode::RegBase)
1265 if (AM.Base_Reg.getNode())
1268 AM.Base_Reg = CurDAG->getRegister(0, VT);
1269 else if (AM.BaseType == X86ISelAddressMode::FrameIndexBase)
1272 if (AM.IndexReg.getNode())
1275 AM.IndexReg = CurDAG->getRegister(0, VT);
1277 // Don't match just leal(,%reg,2). It's cheaper to do addl %reg, %reg, or with
1282 // FIXME: We are artificially lowering the criteria to turn ADD %reg, $GA
1283 // to a LEA. This is determined with some expermentation but is by no means
1284 // optimal (especially for code size consideration). LEA is nice because of
1285 // its three-address nature. Tweak the cost function again when we can run
1286 // convertToThreeAddress() at register allocation time.
1287 if (AM.hasSymbolicDisplacement()) {
1288 // For X86-64, we should always use lea to materialize RIP relative
1290 if (Subtarget->is64Bit())
1296 if (AM.Disp && (AM.Base_Reg.getNode() || AM.IndexReg.getNode()))
1299 // If it isn't worth using an LEA, reject it.
1300 if (Complexity <= 2)
1303 getAddressOperands(AM, Base, Scale, Index, Disp, Segment);
1307 /// SelectTLSADDRAddr - This is only run on TargetGlobalTLSAddress nodes.
1308 bool X86DAGToDAGISel::SelectTLSADDRAddr(SDValue N, SDValue &Base,
1309 SDValue &Scale, SDValue &Index,
1310 SDValue &Disp, SDValue &Segment) {
1311 assert(N.getOpcode() == ISD::TargetGlobalTLSAddress);
1312 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(N);
1314 X86ISelAddressMode AM;
1315 AM.GV = GA->getGlobal();
1316 AM.Disp += GA->getOffset();
1317 AM.Base_Reg = CurDAG->getRegister(0, N.getValueType());
1318 AM.SymbolFlags = GA->getTargetFlags();
1320 if (N.getValueType() == MVT::i32) {
1322 AM.IndexReg = CurDAG->getRegister(X86::EBX, MVT::i32);
1324 AM.IndexReg = CurDAG->getRegister(0, MVT::i64);
1327 getAddressOperands(AM, Base, Scale, Index, Disp, Segment);
1332 bool X86DAGToDAGISel::TryFoldLoad(SDNode *P, SDValue N,
1333 SDValue &Base, SDValue &Scale,
1334 SDValue &Index, SDValue &Disp,
1336 if (!ISD::isNON_EXTLoad(N.getNode()) ||
1337 !IsProfitableToFold(N, P, P) ||
1338 !IsLegalToFold(N, P, P, OptLevel))
1341 return SelectAddr(N.getNode(),
1342 N.getOperand(1), Base, Scale, Index, Disp, Segment);
1345 /// getGlobalBaseReg - Return an SDNode that returns the value of
1346 /// the global base register. Output instructions required to
1347 /// initialize the global base register, if necessary.
1349 SDNode *X86DAGToDAGISel::getGlobalBaseReg() {
1350 unsigned GlobalBaseReg = getInstrInfo()->getGlobalBaseReg(MF);
1351 return CurDAG->getRegister(GlobalBaseReg, TLI.getPointerTy()).getNode();
1354 SDNode *X86DAGToDAGISel::SelectAtomic64(SDNode *Node, unsigned Opc) {
1355 SDValue Chain = Node->getOperand(0);
1356 SDValue In1 = Node->getOperand(1);
1357 SDValue In2L = Node->getOperand(2);
1358 SDValue In2H = Node->getOperand(3);
1359 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1360 if (!SelectAddr(Node, In1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4))
1362 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
1363 MemOp[0] = cast<MemSDNode>(Node)->getMemOperand();
1364 const SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, In2L, In2H, Chain};
1365 SDNode *ResNode = CurDAG->getMachineNode(Opc, Node->getDebugLoc(),
1366 MVT::i32, MVT::i32, MVT::Other, Ops,
1367 array_lengthof(Ops));
1368 cast<MachineSDNode>(ResNode)->setMemRefs(MemOp, MemOp + 1);
1372 SDNode *X86DAGToDAGISel::SelectAtomicLoadAdd(SDNode *Node, EVT NVT) {
1373 if (Node->hasAnyUseOfValue(0))
1376 // Optimize common patterns for __sync_add_and_fetch and
1377 // __sync_sub_and_fetch where the result is not used. This allows us
1378 // to use "lock" version of add, sub, inc, dec instructions.
1379 // FIXME: Do not use special instructions but instead add the "lock"
1380 // prefix to the target node somehow. The extra information will then be
1381 // transferred to machine instruction and it denotes the prefix.
1382 SDValue Chain = Node->getOperand(0);
1383 SDValue Ptr = Node->getOperand(1);
1384 SDValue Val = Node->getOperand(2);
1385 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1386 if (!SelectAddr(Node, Ptr, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4))
1389 bool isInc = false, isDec = false, isSub = false, isCN = false;
1390 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Val);
1393 int64_t CNVal = CN->getSExtValue();
1396 else if (CNVal == -1)
1398 else if (CNVal >= 0)
1399 Val = CurDAG->getTargetConstant(CNVal, NVT);
1402 Val = CurDAG->getTargetConstant(-CNVal, NVT);
1404 } else if (Val.hasOneUse() &&
1405 Val.getOpcode() == ISD::SUB &&
1406 X86::isZeroNode(Val.getOperand(0))) {
1408 Val = Val.getOperand(1);
1412 switch (NVT.getSimpleVT().SimpleTy) {
1416 Opc = X86::LOCK_INC8m;
1418 Opc = X86::LOCK_DEC8m;
1421 Opc = X86::LOCK_SUB8mi;
1423 Opc = X86::LOCK_SUB8mr;
1426 Opc = X86::LOCK_ADD8mi;
1428 Opc = X86::LOCK_ADD8mr;
1433 Opc = X86::LOCK_INC16m;
1435 Opc = X86::LOCK_DEC16m;
1438 if (immSext8(Val.getNode()))
1439 Opc = X86::LOCK_SUB16mi8;
1441 Opc = X86::LOCK_SUB16mi;
1443 Opc = X86::LOCK_SUB16mr;
1446 if (immSext8(Val.getNode()))
1447 Opc = X86::LOCK_ADD16mi8;
1449 Opc = X86::LOCK_ADD16mi;
1451 Opc = X86::LOCK_ADD16mr;
1456 Opc = X86::LOCK_INC32m;
1458 Opc = X86::LOCK_DEC32m;
1461 if (immSext8(Val.getNode()))
1462 Opc = X86::LOCK_SUB32mi8;
1464 Opc = X86::LOCK_SUB32mi;
1466 Opc = X86::LOCK_SUB32mr;
1469 if (immSext8(Val.getNode()))
1470 Opc = X86::LOCK_ADD32mi8;
1472 Opc = X86::LOCK_ADD32mi;
1474 Opc = X86::LOCK_ADD32mr;
1479 Opc = X86::LOCK_INC64m;
1481 Opc = X86::LOCK_DEC64m;
1483 Opc = X86::LOCK_SUB64mr;
1485 if (immSext8(Val.getNode()))
1486 Opc = X86::LOCK_SUB64mi8;
1487 else if (i64immSExt32(Val.getNode()))
1488 Opc = X86::LOCK_SUB64mi32;
1491 Opc = X86::LOCK_ADD64mr;
1493 if (immSext8(Val.getNode()))
1494 Opc = X86::LOCK_ADD64mi8;
1495 else if (i64immSExt32(Val.getNode()))
1496 Opc = X86::LOCK_ADD64mi32;
1502 DebugLoc dl = Node->getDebugLoc();
1503 SDValue Undef = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1505 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
1506 MemOp[0] = cast<MemSDNode>(Node)->getMemOperand();
1507 if (isInc || isDec) {
1508 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Chain };
1509 SDValue Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 6), 0);
1510 cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
1511 SDValue RetVals[] = { Undef, Ret };
1512 return CurDAG->getMergeValues(RetVals, 2, dl).getNode();
1514 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Val, Chain };
1515 SDValue Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 7), 0);
1516 cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
1517 SDValue RetVals[] = { Undef, Ret };
1518 return CurDAG->getMergeValues(RetVals, 2, dl).getNode();
1522 /// HasNoSignedComparisonUses - Test whether the given X86ISD::CMP node has
1523 /// any uses which require the SF or OF bits to be accurate.
1524 static bool HasNoSignedComparisonUses(SDNode *N) {
1525 // Examine each user of the node.
1526 for (SDNode::use_iterator UI = N->use_begin(),
1527 UE = N->use_end(); UI != UE; ++UI) {
1528 // Only examine CopyToReg uses.
1529 if (UI->getOpcode() != ISD::CopyToReg)
1531 // Only examine CopyToReg uses that copy to EFLAGS.
1532 if (cast<RegisterSDNode>(UI->getOperand(1))->getReg() !=
1535 // Examine each user of the CopyToReg use.
1536 for (SDNode::use_iterator FlagUI = UI->use_begin(),
1537 FlagUE = UI->use_end(); FlagUI != FlagUE; ++FlagUI) {
1538 // Only examine the Flag result.
1539 if (FlagUI.getUse().getResNo() != 1) continue;
1540 // Anything unusual: assume conservatively.
1541 if (!FlagUI->isMachineOpcode()) return false;
1542 // Examine the opcode of the user.
1543 switch (FlagUI->getMachineOpcode()) {
1544 // These comparisons don't treat the most significant bit specially.
1545 case X86::SETAr: case X86::SETAEr: case X86::SETBr: case X86::SETBEr:
1546 case X86::SETEr: case X86::SETNEr: case X86::SETPr: case X86::SETNPr:
1547 case X86::SETAm: case X86::SETAEm: case X86::SETBm: case X86::SETBEm:
1548 case X86::SETEm: case X86::SETNEm: case X86::SETPm: case X86::SETNPm:
1549 case X86::JA_4: case X86::JAE_4: case X86::JB_4: case X86::JBE_4:
1550 case X86::JE_4: case X86::JNE_4: case X86::JP_4: case X86::JNP_4:
1551 case X86::CMOVA16rr: case X86::CMOVA16rm:
1552 case X86::CMOVA32rr: case X86::CMOVA32rm:
1553 case X86::CMOVA64rr: case X86::CMOVA64rm:
1554 case X86::CMOVAE16rr: case X86::CMOVAE16rm:
1555 case X86::CMOVAE32rr: case X86::CMOVAE32rm:
1556 case X86::CMOVAE64rr: case X86::CMOVAE64rm:
1557 case X86::CMOVB16rr: case X86::CMOVB16rm:
1558 case X86::CMOVB32rr: case X86::CMOVB32rm:
1559 case X86::CMOVB64rr: case X86::CMOVB64rm:
1560 case X86::CMOVBE16rr: case X86::CMOVBE16rm:
1561 case X86::CMOVBE32rr: case X86::CMOVBE32rm:
1562 case X86::CMOVBE64rr: case X86::CMOVBE64rm:
1563 case X86::CMOVE16rr: case X86::CMOVE16rm:
1564 case X86::CMOVE32rr: case X86::CMOVE32rm:
1565 case X86::CMOVE64rr: case X86::CMOVE64rm:
1566 case X86::CMOVNE16rr: case X86::CMOVNE16rm:
1567 case X86::CMOVNE32rr: case X86::CMOVNE32rm:
1568 case X86::CMOVNE64rr: case X86::CMOVNE64rm:
1569 case X86::CMOVNP16rr: case X86::CMOVNP16rm:
1570 case X86::CMOVNP32rr: case X86::CMOVNP32rm:
1571 case X86::CMOVNP64rr: case X86::CMOVNP64rm:
1572 case X86::CMOVP16rr: case X86::CMOVP16rm:
1573 case X86::CMOVP32rr: case X86::CMOVP32rm:
1574 case X86::CMOVP64rr: case X86::CMOVP64rm:
1576 // Anything else: assume conservatively.
1577 default: return false;
1584 SDNode *X86DAGToDAGISel::Select(SDNode *Node) {
1585 EVT NVT = Node->getValueType(0);
1587 unsigned Opcode = Node->getOpcode();
1588 DebugLoc dl = Node->getDebugLoc();
1590 DEBUG(dbgs() << "Selecting: "; Node->dump(CurDAG); dbgs() << '\n');
1592 if (Node->isMachineOpcode()) {
1593 DEBUG(dbgs() << "== "; Node->dump(CurDAG); dbgs() << '\n');
1594 return NULL; // Already selected.
1599 case X86ISD::GlobalBaseReg:
1600 return getGlobalBaseReg();
1602 case X86ISD::ATOMOR64_DAG:
1603 return SelectAtomic64(Node, X86::ATOMOR6432);
1604 case X86ISD::ATOMXOR64_DAG:
1605 return SelectAtomic64(Node, X86::ATOMXOR6432);
1606 case X86ISD::ATOMADD64_DAG:
1607 return SelectAtomic64(Node, X86::ATOMADD6432);
1608 case X86ISD::ATOMSUB64_DAG:
1609 return SelectAtomic64(Node, X86::ATOMSUB6432);
1610 case X86ISD::ATOMNAND64_DAG:
1611 return SelectAtomic64(Node, X86::ATOMNAND6432);
1612 case X86ISD::ATOMAND64_DAG:
1613 return SelectAtomic64(Node, X86::ATOMAND6432);
1614 case X86ISD::ATOMSWAP64_DAG:
1615 return SelectAtomic64(Node, X86::ATOMSWAP6432);
1617 case ISD::ATOMIC_LOAD_ADD: {
1618 SDNode *RetVal = SelectAtomicLoadAdd(Node, NVT);
1624 case ISD::SMUL_LOHI:
1625 case ISD::UMUL_LOHI: {
1626 SDValue N0 = Node->getOperand(0);
1627 SDValue N1 = Node->getOperand(1);
1629 bool isSigned = Opcode == ISD::SMUL_LOHI;
1631 switch (NVT.getSimpleVT().SimpleTy) {
1632 default: llvm_unreachable("Unsupported VT!");
1633 case MVT::i8: Opc = X86::MUL8r; MOpc = X86::MUL8m; break;
1634 case MVT::i16: Opc = X86::MUL16r; MOpc = X86::MUL16m; break;
1635 case MVT::i32: Opc = X86::MUL32r; MOpc = X86::MUL32m; break;
1636 case MVT::i64: Opc = X86::MUL64r; MOpc = X86::MUL64m; break;
1639 switch (NVT.getSimpleVT().SimpleTy) {
1640 default: llvm_unreachable("Unsupported VT!");
1641 case MVT::i8: Opc = X86::IMUL8r; MOpc = X86::IMUL8m; break;
1642 case MVT::i16: Opc = X86::IMUL16r; MOpc = X86::IMUL16m; break;
1643 case MVT::i32: Opc = X86::IMUL32r; MOpc = X86::IMUL32m; break;
1644 case MVT::i64: Opc = X86::IMUL64r; MOpc = X86::IMUL64m; break;
1648 unsigned LoReg, HiReg;
1649 switch (NVT.getSimpleVT().SimpleTy) {
1650 default: llvm_unreachable("Unsupported VT!");
1651 case MVT::i8: LoReg = X86::AL; HiReg = X86::AH; break;
1652 case MVT::i16: LoReg = X86::AX; HiReg = X86::DX; break;
1653 case MVT::i32: LoReg = X86::EAX; HiReg = X86::EDX; break;
1654 case MVT::i64: LoReg = X86::RAX; HiReg = X86::RDX; break;
1657 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1658 bool foldedLoad = TryFoldLoad(Node, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
1659 // Multiply is commmutative.
1661 foldedLoad = TryFoldLoad(Node, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
1666 SDValue InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, LoReg,
1667 N0, SDValue()).getValue(1);
1670 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
1673 CurDAG->getMachineNode(MOpc, dl, MVT::Other, MVT::Flag, Ops,
1674 array_lengthof(Ops));
1675 InFlag = SDValue(CNode, 1);
1676 // Update the chain.
1677 ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
1680 SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Flag, N1, InFlag), 0);
1683 // Prevent use of AH in a REX instruction by referencing AX instead.
1684 if (HiReg == X86::AH && Subtarget->is64Bit() &&
1685 !SDValue(Node, 1).use_empty()) {
1686 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1687 X86::AX, MVT::i16, InFlag);
1688 InFlag = Result.getValue(2);
1689 // Get the low part if needed. Don't use getCopyFromReg for aliasing
1691 if (!SDValue(Node, 0).use_empty())
1692 ReplaceUses(SDValue(Node, 1),
1693 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
1695 // Shift AX down 8 bits.
1696 Result = SDValue(CurDAG->getMachineNode(X86::SHR16ri, dl, MVT::i16,
1698 CurDAG->getTargetConstant(8, MVT::i8)), 0);
1699 // Then truncate it down to i8.
1700 ReplaceUses(SDValue(Node, 1),
1701 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
1703 // Copy the low half of the result, if it is needed.
1704 if (!SDValue(Node, 0).use_empty()) {
1705 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1706 LoReg, NVT, InFlag);
1707 InFlag = Result.getValue(2);
1708 ReplaceUses(SDValue(Node, 0), Result);
1709 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
1711 // Copy the high half of the result, if it is needed.
1712 if (!SDValue(Node, 1).use_empty()) {
1713 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1714 HiReg, NVT, InFlag);
1715 InFlag = Result.getValue(2);
1716 ReplaceUses(SDValue(Node, 1), Result);
1717 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
1724 case ISD::UDIVREM: {
1725 SDValue N0 = Node->getOperand(0);
1726 SDValue N1 = Node->getOperand(1);
1728 bool isSigned = Opcode == ISD::SDIVREM;
1730 switch (NVT.getSimpleVT().SimpleTy) {
1731 default: llvm_unreachable("Unsupported VT!");
1732 case MVT::i8: Opc = X86::DIV8r; MOpc = X86::DIV8m; break;
1733 case MVT::i16: Opc = X86::DIV16r; MOpc = X86::DIV16m; break;
1734 case MVT::i32: Opc = X86::DIV32r; MOpc = X86::DIV32m; break;
1735 case MVT::i64: Opc = X86::DIV64r; MOpc = X86::DIV64m; break;
1738 switch (NVT.getSimpleVT().SimpleTy) {
1739 default: llvm_unreachable("Unsupported VT!");
1740 case MVT::i8: Opc = X86::IDIV8r; MOpc = X86::IDIV8m; break;
1741 case MVT::i16: Opc = X86::IDIV16r; MOpc = X86::IDIV16m; break;
1742 case MVT::i32: Opc = X86::IDIV32r; MOpc = X86::IDIV32m; break;
1743 case MVT::i64: Opc = X86::IDIV64r; MOpc = X86::IDIV64m; break;
1747 unsigned LoReg, HiReg, ClrReg;
1748 unsigned ClrOpcode, SExtOpcode;
1749 switch (NVT.getSimpleVT().SimpleTy) {
1750 default: llvm_unreachable("Unsupported VT!");
1752 LoReg = X86::AL; ClrReg = HiReg = X86::AH;
1754 SExtOpcode = X86::CBW;
1757 LoReg = X86::AX; HiReg = X86::DX;
1758 ClrOpcode = X86::MOV16r0; ClrReg = X86::DX;
1759 SExtOpcode = X86::CWD;
1762 LoReg = X86::EAX; ClrReg = HiReg = X86::EDX;
1763 ClrOpcode = X86::MOV32r0;
1764 SExtOpcode = X86::CDQ;
1767 LoReg = X86::RAX; ClrReg = HiReg = X86::RDX;
1768 ClrOpcode = X86::MOV64r0;
1769 SExtOpcode = X86::CQO;
1773 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1774 bool foldedLoad = TryFoldLoad(Node, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
1775 bool signBitIsZero = CurDAG->SignBitIsZero(N0);
1778 if (NVT == MVT::i8 && (!isSigned || signBitIsZero)) {
1779 // Special case for div8, just use a move with zero extension to AX to
1780 // clear the upper 8 bits (AH).
1781 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Move, Chain;
1782 if (TryFoldLoad(Node, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4)) {
1783 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N0.getOperand(0) };
1785 SDValue(CurDAG->getMachineNode(X86::MOVZX16rm8, dl, MVT::i16,
1787 array_lengthof(Ops)), 0);
1788 Chain = Move.getValue(1);
1789 ReplaceUses(N0.getValue(1), Chain);
1792 SDValue(CurDAG->getMachineNode(X86::MOVZX16rr8, dl, MVT::i16, N0),0);
1793 Chain = CurDAG->getEntryNode();
1795 Chain = CurDAG->getCopyToReg(Chain, dl, X86::AX, Move, SDValue());
1796 InFlag = Chain.getValue(1);
1799 CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl,
1800 LoReg, N0, SDValue()).getValue(1);
1801 if (isSigned && !signBitIsZero) {
1802 // Sign extend the low part into the high part.
1804 SDValue(CurDAG->getMachineNode(SExtOpcode, dl, MVT::Flag, InFlag),0);
1806 // Zero out the high part, effectively zero extending the input.
1808 SDValue(CurDAG->getMachineNode(ClrOpcode, dl, NVT), 0);
1809 InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, ClrReg,
1810 ClrNode, InFlag).getValue(1);
1815 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
1818 CurDAG->getMachineNode(MOpc, dl, MVT::Other, MVT::Flag, Ops,
1819 array_lengthof(Ops));
1820 InFlag = SDValue(CNode, 1);
1821 // Update the chain.
1822 ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
1825 SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Flag, N1, InFlag), 0);
1828 // Prevent use of AH in a REX instruction by referencing AX instead.
1829 // Shift it down 8 bits.
1830 if (HiReg == X86::AH && Subtarget->is64Bit() &&
1831 !SDValue(Node, 1).use_empty()) {
1832 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1833 X86::AX, MVT::i16, InFlag);
1834 InFlag = Result.getValue(2);
1836 // If we also need AL (the quotient), get it by extracting a subreg from
1837 // Result. The fast register allocator does not like multiple CopyFromReg
1838 // nodes using aliasing registers.
1839 if (!SDValue(Node, 0).use_empty())
1840 ReplaceUses(SDValue(Node, 0),
1841 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
1843 // Shift AX right by 8 bits instead of using AH.
1844 Result = SDValue(CurDAG->getMachineNode(X86::SHR16ri, dl, MVT::i16,
1846 CurDAG->getTargetConstant(8, MVT::i8)),
1848 ReplaceUses(SDValue(Node, 1),
1849 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
1851 // Copy the division (low) result, if it is needed.
1852 if (!SDValue(Node, 0).use_empty()) {
1853 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1854 LoReg, NVT, InFlag);
1855 InFlag = Result.getValue(2);
1856 ReplaceUses(SDValue(Node, 0), Result);
1857 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
1859 // Copy the remainder (high) result, if it is needed.
1860 if (!SDValue(Node, 1).use_empty()) {
1861 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1862 HiReg, NVT, InFlag);
1863 InFlag = Result.getValue(2);
1864 ReplaceUses(SDValue(Node, 1), Result);
1865 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
1871 SDValue N0 = Node->getOperand(0);
1872 SDValue N1 = Node->getOperand(1);
1874 // Look for (X86cmp (and $op, $imm), 0) and see if we can convert it to
1875 // use a smaller encoding.
1876 if (N0.getOpcode() == ISD::TRUNCATE && N0.hasOneUse() &&
1877 HasNoSignedComparisonUses(Node))
1878 // Look past the truncate if CMP is the only use of it.
1879 N0 = N0.getOperand(0);
1880 if (N0.getNode()->getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
1881 N0.getValueType() != MVT::i8 &&
1882 X86::isZeroNode(N1)) {
1883 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getNode()->getOperand(1));
1886 // For example, convert "testl %eax, $8" to "testb %al, $8"
1887 if ((C->getZExtValue() & ~UINT64_C(0xff)) == 0 &&
1888 (!(C->getZExtValue() & 0x80) ||
1889 HasNoSignedComparisonUses(Node))) {
1890 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), MVT::i8);
1891 SDValue Reg = N0.getNode()->getOperand(0);
1893 // On x86-32, only the ABCD registers have 8-bit subregisters.
1894 if (!Subtarget->is64Bit()) {
1895 TargetRegisterClass *TRC = 0;
1896 switch (N0.getValueType().getSimpleVT().SimpleTy) {
1897 case MVT::i32: TRC = &X86::GR32_ABCDRegClass; break;
1898 case MVT::i16: TRC = &X86::GR16_ABCDRegClass; break;
1899 default: llvm_unreachable("Unsupported TEST operand type!");
1901 SDValue RC = CurDAG->getTargetConstant(TRC->getID(), MVT::i32);
1902 Reg = SDValue(CurDAG->getMachineNode(X86::COPY_TO_REGCLASS, dl,
1903 Reg.getValueType(), Reg, RC), 0);
1906 // Extract the l-register.
1907 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl,
1911 return CurDAG->getMachineNode(X86::TEST8ri, dl, MVT::i32, Subreg, Imm);
1914 // For example, "testl %eax, $2048" to "testb %ah, $8".
1915 if ((C->getZExtValue() & ~UINT64_C(0xff00)) == 0 &&
1916 (!(C->getZExtValue() & 0x8000) ||
1917 HasNoSignedComparisonUses(Node))) {
1918 // Shift the immediate right by 8 bits.
1919 SDValue ShiftedImm = CurDAG->getTargetConstant(C->getZExtValue() >> 8,
1921 SDValue Reg = N0.getNode()->getOperand(0);
1923 // Put the value in an ABCD register.
1924 TargetRegisterClass *TRC = 0;
1925 switch (N0.getValueType().getSimpleVT().SimpleTy) {
1926 case MVT::i64: TRC = &X86::GR64_ABCDRegClass; break;
1927 case MVT::i32: TRC = &X86::GR32_ABCDRegClass; break;
1928 case MVT::i16: TRC = &X86::GR16_ABCDRegClass; break;
1929 default: llvm_unreachable("Unsupported TEST operand type!");
1931 SDValue RC = CurDAG->getTargetConstant(TRC->getID(), MVT::i32);
1932 Reg = SDValue(CurDAG->getMachineNode(X86::COPY_TO_REGCLASS, dl,
1933 Reg.getValueType(), Reg, RC), 0);
1935 // Extract the h-register.
1936 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_8bit_hi, dl,
1939 // Emit a testb. No special NOREX tricks are needed since there's
1940 // only one GPR operand!
1941 return CurDAG->getMachineNode(X86::TEST8ri, dl, MVT::i32,
1942 Subreg, ShiftedImm);
1945 // For example, "testl %eax, $32776" to "testw %ax, $32776".
1946 if ((C->getZExtValue() & ~UINT64_C(0xffff)) == 0 &&
1947 N0.getValueType() != MVT::i16 &&
1948 (!(C->getZExtValue() & 0x8000) ||
1949 HasNoSignedComparisonUses(Node))) {
1950 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), MVT::i16);
1951 SDValue Reg = N0.getNode()->getOperand(0);
1953 // Extract the 16-bit subregister.
1954 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_16bit, dl,
1958 return CurDAG->getMachineNode(X86::TEST16ri, dl, MVT::i32, Subreg, Imm);
1961 // For example, "testq %rax, $268468232" to "testl %eax, $268468232".
1962 if ((C->getZExtValue() & ~UINT64_C(0xffffffff)) == 0 &&
1963 N0.getValueType() == MVT::i64 &&
1964 (!(C->getZExtValue() & 0x80000000) ||
1965 HasNoSignedComparisonUses(Node))) {
1966 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), MVT::i32);
1967 SDValue Reg = N0.getNode()->getOperand(0);
1969 // Extract the 32-bit subregister.
1970 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_32bit, dl,
1974 return CurDAG->getMachineNode(X86::TEST32ri, dl, MVT::i32, Subreg, Imm);
1981 SDNode *ResNode = SelectCode(Node);
1983 DEBUG(dbgs() << "=> ";
1984 if (ResNode == NULL || ResNode == Node)
1987 ResNode->dump(CurDAG);
1993 bool X86DAGToDAGISel::
1994 SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
1995 std::vector<SDValue> &OutOps) {
1996 SDValue Op0, Op1, Op2, Op3, Op4;
1997 switch (ConstraintCode) {
1998 case 'o': // offsetable ??
1999 case 'v': // not offsetable ??
2000 default: return true;
2002 if (!SelectAddr(0, Op, Op0, Op1, Op2, Op3, Op4))
2007 OutOps.push_back(Op0);
2008 OutOps.push_back(Op1);
2009 OutOps.push_back(Op2);
2010 OutOps.push_back(Op3);
2011 OutOps.push_back(Op4);
2015 /// createX86ISelDag - This pass converts a legalized DAG into a
2016 /// X86-specific DAG, ready for instruction scheduling.
2018 FunctionPass *llvm::createX86ISelDag(X86TargetMachine &TM,
2019 llvm::CodeGenOpt::Level OptLevel) {
2020 return new X86DAGToDAGISel(TM, OptLevel);