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 *Op, SDValue N, SDValue &Base,
201 SDValue &Scale, SDValue &Index, SDValue &Disp,
203 bool SelectLEAAddr(SDNode *Op, SDValue N, SDValue &Base,
204 SDValue &Scale, SDValue &Index, SDValue &Disp,
206 bool SelectTLSADDRAddr(SDNode *Op, 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 (const Value *Src = LD->getSrcValue())
408 if (const PointerType *PT = dyn_cast<PointerType>(Src->getType()))
409 if (PT->getAddressSpace() >= 256)
412 // Now let's find the callseq_start.
413 while (HasCallSeq && Chain.getOpcode() != ISD::CALLSEQ_START) {
414 if (!Chain.hasOneUse())
416 Chain = Chain.getOperand(0);
419 if (!Chain.getNumOperands())
421 if (Chain.getOperand(0).getNode() == Callee.getNode())
423 if (Chain.getOperand(0).getOpcode() == ISD::TokenFactor &&
424 Callee.getValue(1).isOperandOf(Chain.getOperand(0).getNode()) &&
425 Callee.getValue(1).hasOneUse())
430 void X86DAGToDAGISel::PreprocessISelDAG() {
431 // OptForSize is used in pattern predicates that isel is matching.
432 OptForSize = MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize);
434 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
435 E = CurDAG->allnodes_end(); I != E; ) {
436 SDNode *N = I++; // Preincrement iterator to avoid invalidation issues.
438 if (OptLevel != CodeGenOpt::None &&
439 (N->getOpcode() == X86ISD::CALL ||
440 N->getOpcode() == X86ISD::TC_RETURN)) {
441 /// Also try moving call address load from outside callseq_start to just
442 /// before the call to allow it to be folded.
460 bool HasCallSeq = N->getOpcode() == X86ISD::CALL;
461 SDValue Chain = N->getOperand(0);
462 SDValue Load = N->getOperand(1);
463 if (!isCalleeLoad(Load, Chain, HasCallSeq))
465 MoveBelowOrigChain(CurDAG, Load, SDValue(N, 0), Chain);
470 // Lower fpround and fpextend nodes that target the FP stack to be store and
471 // load to the stack. This is a gross hack. We would like to simply mark
472 // these as being illegal, but when we do that, legalize produces these when
473 // it expands calls, then expands these in the same legalize pass. We would
474 // like dag combine to be able to hack on these between the call expansion
475 // and the node legalization. As such this pass basically does "really
476 // late" legalization of these inline with the X86 isel pass.
477 // FIXME: This should only happen when not compiled with -O0.
478 if (N->getOpcode() != ISD::FP_ROUND && N->getOpcode() != ISD::FP_EXTEND)
481 // If the source and destination are SSE registers, then this is a legal
482 // conversion that should not be lowered.
483 EVT SrcVT = N->getOperand(0).getValueType();
484 EVT DstVT = N->getValueType(0);
485 bool SrcIsSSE = X86Lowering.isScalarFPTypeInSSEReg(SrcVT);
486 bool DstIsSSE = X86Lowering.isScalarFPTypeInSSEReg(DstVT);
487 if (SrcIsSSE && DstIsSSE)
490 if (!SrcIsSSE && !DstIsSSE) {
491 // If this is an FPStack extension, it is a noop.
492 if (N->getOpcode() == ISD::FP_EXTEND)
494 // If this is a value-preserving FPStack truncation, it is a noop.
495 if (N->getConstantOperandVal(1))
499 // Here we could have an FP stack truncation or an FPStack <-> SSE convert.
500 // FPStack has extload and truncstore. SSE can fold direct loads into other
501 // operations. Based on this, decide what we want to do.
503 if (N->getOpcode() == ISD::FP_ROUND)
504 MemVT = DstVT; // FP_ROUND must use DstVT, we can't do a 'trunc load'.
506 MemVT = SrcIsSSE ? SrcVT : DstVT;
508 SDValue MemTmp = CurDAG->CreateStackTemporary(MemVT);
509 DebugLoc dl = N->getDebugLoc();
511 // FIXME: optimize the case where the src/dest is a load or store?
512 SDValue Store = CurDAG->getTruncStore(CurDAG->getEntryNode(), dl,
514 MemTmp, NULL, 0, MemVT,
516 SDValue Result = CurDAG->getExtLoad(ISD::EXTLOAD, DstVT, dl, Store, MemTmp,
517 NULL, 0, MemVT, false, false, 0);
519 // We're about to replace all uses of the FP_ROUND/FP_EXTEND with the
520 // extload we created. This will cause general havok on the dag because
521 // anything below the conversion could be folded into other existing nodes.
522 // To avoid invalidating 'I', back it up to the convert node.
524 CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
526 // Now that we did that, the node is dead. Increment the iterator to the
527 // next node to process, then delete N.
529 CurDAG->DeleteNode(N);
534 /// EmitSpecialCodeForMain - Emit any code that needs to be executed only in
535 /// the main function.
536 void X86DAGToDAGISel::EmitSpecialCodeForMain(MachineBasicBlock *BB,
537 MachineFrameInfo *MFI) {
538 const TargetInstrInfo *TII = TM.getInstrInfo();
539 if (Subtarget->isTargetCygMing())
540 BuildMI(BB, DebugLoc(),
541 TII->get(X86::CALLpcrel32)).addExternalSymbol("__main");
544 void X86DAGToDAGISel::EmitFunctionEntryCode() {
545 // If this is main, emit special code for main.
546 if (const Function *Fn = MF->getFunction())
547 if (Fn->hasExternalLinkage() && Fn->getName() == "main")
548 EmitSpecialCodeForMain(MF->begin(), MF->getFrameInfo());
552 bool X86DAGToDAGISel::MatchSegmentBaseAddress(SDValue N,
553 X86ISelAddressMode &AM) {
554 assert(N.getOpcode() == X86ISD::SegmentBaseAddress);
555 SDValue Segment = N.getOperand(0);
557 if (AM.Segment.getNode() == 0) {
558 AM.Segment = Segment;
565 bool X86DAGToDAGISel::MatchLoad(SDValue N, X86ISelAddressMode &AM) {
566 // This optimization is valid because the GNU TLS model defines that
567 // gs:0 (or fs:0 on X86-64) contains its own address.
568 // For more information see http://people.redhat.com/drepper/tls.pdf
570 SDValue Address = N.getOperand(1);
571 if (Address.getOpcode() == X86ISD::SegmentBaseAddress &&
572 !MatchSegmentBaseAddress(Address, AM))
578 /// MatchWrapper - Try to match X86ISD::Wrapper and X86ISD::WrapperRIP nodes
579 /// into an addressing mode. These wrap things that will resolve down into a
580 /// symbol reference. If no match is possible, this returns true, otherwise it
582 bool X86DAGToDAGISel::MatchWrapper(SDValue N, X86ISelAddressMode &AM) {
583 // If the addressing mode already has a symbol as the displacement, we can
584 // never match another symbol.
585 if (AM.hasSymbolicDisplacement())
588 SDValue N0 = N.getOperand(0);
589 CodeModel::Model M = TM.getCodeModel();
591 // Handle X86-64 rip-relative addresses. We check this before checking direct
592 // folding because RIP is preferable to non-RIP accesses.
593 if (Subtarget->is64Bit() &&
594 // Under X86-64 non-small code model, GV (and friends) are 64-bits, so
595 // they cannot be folded into immediate fields.
596 // FIXME: This can be improved for kernel and other models?
597 (M == CodeModel::Small || M == CodeModel::Kernel) &&
598 // Base and index reg must be 0 in order to use %rip as base and lowering
600 !AM.hasBaseOrIndexReg() && N.getOpcode() == X86ISD::WrapperRIP) {
601 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(N0)) {
602 int64_t Offset = AM.Disp + G->getOffset();
603 if (!X86::isOffsetSuitableForCodeModel(Offset, M)) return true;
604 AM.GV = G->getGlobal();
606 AM.SymbolFlags = G->getTargetFlags();
607 } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N0)) {
608 int64_t Offset = AM.Disp + CP->getOffset();
609 if (!X86::isOffsetSuitableForCodeModel(Offset, M)) return true;
610 AM.CP = CP->getConstVal();
611 AM.Align = CP->getAlignment();
613 AM.SymbolFlags = CP->getTargetFlags();
614 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(N0)) {
615 AM.ES = S->getSymbol();
616 AM.SymbolFlags = S->getTargetFlags();
617 } else if (JumpTableSDNode *J = dyn_cast<JumpTableSDNode>(N0)) {
618 AM.JT = J->getIndex();
619 AM.SymbolFlags = J->getTargetFlags();
621 AM.BlockAddr = cast<BlockAddressSDNode>(N0)->getBlockAddress();
622 AM.SymbolFlags = cast<BlockAddressSDNode>(N0)->getTargetFlags();
625 if (N.getOpcode() == X86ISD::WrapperRIP)
626 AM.setBaseReg(CurDAG->getRegister(X86::RIP, MVT::i64));
630 // Handle the case when globals fit in our immediate field: This is true for
631 // X86-32 always and X86-64 when in -static -mcmodel=small mode. In 64-bit
632 // mode, this results in a non-RIP-relative computation.
633 if (!Subtarget->is64Bit() ||
634 ((M == CodeModel::Small || M == CodeModel::Kernel) &&
635 TM.getRelocationModel() == Reloc::Static)) {
636 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(N0)) {
637 AM.GV = G->getGlobal();
638 AM.Disp += G->getOffset();
639 AM.SymbolFlags = G->getTargetFlags();
640 } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N0)) {
641 AM.CP = CP->getConstVal();
642 AM.Align = CP->getAlignment();
643 AM.Disp += CP->getOffset();
644 AM.SymbolFlags = CP->getTargetFlags();
645 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(N0)) {
646 AM.ES = S->getSymbol();
647 AM.SymbolFlags = S->getTargetFlags();
648 } else if (JumpTableSDNode *J = dyn_cast<JumpTableSDNode>(N0)) {
649 AM.JT = J->getIndex();
650 AM.SymbolFlags = J->getTargetFlags();
652 AM.BlockAddr = cast<BlockAddressSDNode>(N0)->getBlockAddress();
653 AM.SymbolFlags = cast<BlockAddressSDNode>(N0)->getTargetFlags();
661 /// MatchAddress - Add the specified node to the specified addressing mode,
662 /// returning true if it cannot be done. This just pattern matches for the
664 bool X86DAGToDAGISel::MatchAddress(SDValue N, X86ISelAddressMode &AM) {
665 if (MatchAddressRecursively(N, AM, 0))
668 // Post-processing: Convert lea(,%reg,2) to lea(%reg,%reg), which has
669 // a smaller encoding and avoids a scaled-index.
671 AM.BaseType == X86ISelAddressMode::RegBase &&
672 AM.Base_Reg.getNode() == 0) {
673 AM.Base_Reg = AM.IndexReg;
677 // Post-processing: Convert foo to foo(%rip), even in non-PIC mode,
678 // because it has a smaller encoding.
679 // TODO: Which other code models can use this?
680 if (TM.getCodeModel() == CodeModel::Small &&
681 Subtarget->is64Bit() &&
683 AM.BaseType == X86ISelAddressMode::RegBase &&
684 AM.Base_Reg.getNode() == 0 &&
685 AM.IndexReg.getNode() == 0 &&
686 AM.SymbolFlags == X86II::MO_NO_FLAG &&
687 AM.hasSymbolicDisplacement())
688 AM.Base_Reg = CurDAG->getRegister(X86::RIP, MVT::i64);
693 /// isLogicallyAddWithConstant - Return true if this node is semantically an
694 /// add of a value with a constantint.
695 static bool isLogicallyAddWithConstant(SDValue V, SelectionDAG *CurDAG) {
696 // Check for (add x, Cst)
697 if (V->getOpcode() == ISD::ADD)
698 return isa<ConstantSDNode>(V->getOperand(1));
700 // Check for (or x, Cst), where Cst & x == 0.
701 if (V->getOpcode() != ISD::OR ||
702 !isa<ConstantSDNode>(V->getOperand(1)))
705 // Handle "X | C" as "X + C" iff X is known to have C bits clear.
706 ConstantSDNode *CN = cast<ConstantSDNode>(V->getOperand(1));
708 // Check to see if the LHS & C is zero.
709 return CurDAG->MaskedValueIsZero(V->getOperand(0), CN->getAPIntValue());
712 bool X86DAGToDAGISel::MatchAddressRecursively(SDValue N, X86ISelAddressMode &AM,
714 bool is64Bit = Subtarget->is64Bit();
715 DebugLoc dl = N.getDebugLoc();
717 dbgs() << "MatchAddress: ";
722 return MatchAddressBase(N, AM);
724 CodeModel::Model M = TM.getCodeModel();
726 // If this is already a %rip relative address, we can only merge immediates
727 // into it. Instead of handling this in every case, we handle it here.
728 // RIP relative addressing: %rip + 32-bit displacement!
729 if (AM.isRIPRelative()) {
730 // FIXME: JumpTable and ExternalSymbol address currently don't like
731 // displacements. It isn't very important, but this should be fixed for
733 if (!AM.ES && AM.JT != -1) return true;
735 if (ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(N)) {
736 int64_t Val = AM.Disp + Cst->getSExtValue();
737 if (X86::isOffsetSuitableForCodeModel(Val, M,
738 AM.hasSymbolicDisplacement())) {
746 switch (N.getOpcode()) {
748 case ISD::Constant: {
749 uint64_t Val = cast<ConstantSDNode>(N)->getSExtValue();
751 X86::isOffsetSuitableForCodeModel(AM.Disp + Val, M,
752 AM.hasSymbolicDisplacement())) {
759 case X86ISD::SegmentBaseAddress:
760 if (!MatchSegmentBaseAddress(N, AM))
764 case X86ISD::Wrapper:
765 case X86ISD::WrapperRIP:
766 if (!MatchWrapper(N, AM))
771 if (!MatchLoad(N, AM))
775 case ISD::FrameIndex:
776 if (AM.BaseType == X86ISelAddressMode::RegBase
777 && AM.Base_Reg.getNode() == 0) {
778 AM.BaseType = X86ISelAddressMode::FrameIndexBase;
779 AM.Base_FrameIndex = cast<FrameIndexSDNode>(N)->getIndex();
785 if (AM.IndexReg.getNode() != 0 || AM.Scale != 1)
789 *CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1))) {
790 unsigned Val = CN->getZExtValue();
791 // Note that we handle x<<1 as (,x,2) rather than (x,x) here so
792 // that the base operand remains free for further matching. If
793 // the base doesn't end up getting used, a post-processing step
794 // in MatchAddress turns (,x,2) into (x,x), which is cheaper.
795 if (Val == 1 || Val == 2 || Val == 3) {
797 SDValue ShVal = N.getNode()->getOperand(0);
799 // Okay, we know that we have a scale by now. However, if the scaled
800 // value is an add of something and a constant, we can fold the
801 // constant into the disp field here.
802 if (isLogicallyAddWithConstant(ShVal, CurDAG)) {
803 AM.IndexReg = ShVal.getNode()->getOperand(0);
804 ConstantSDNode *AddVal =
805 cast<ConstantSDNode>(ShVal.getNode()->getOperand(1));
806 uint64_t Disp = AM.Disp + (AddVal->getSExtValue() << Val);
808 X86::isOffsetSuitableForCodeModel(Disp, M,
809 AM.hasSymbolicDisplacement()))
823 // A mul_lohi where we need the low part can be folded as a plain multiply.
824 if (N.getResNo() != 0) break;
827 case X86ISD::MUL_IMM:
828 // X*[3,5,9] -> X+X*[2,4,8]
829 if (AM.BaseType == X86ISelAddressMode::RegBase &&
830 AM.Base_Reg.getNode() == 0 &&
831 AM.IndexReg.getNode() == 0) {
833 *CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1)))
834 if (CN->getZExtValue() == 3 || CN->getZExtValue() == 5 ||
835 CN->getZExtValue() == 9) {
836 AM.Scale = unsigned(CN->getZExtValue())-1;
838 SDValue MulVal = N.getNode()->getOperand(0);
841 // Okay, we know that we have a scale by now. However, if the scaled
842 // value is an add of something and a constant, we can fold the
843 // constant into the disp field here.
844 if (MulVal.getNode()->getOpcode() == ISD::ADD && MulVal.hasOneUse() &&
845 isa<ConstantSDNode>(MulVal.getNode()->getOperand(1))) {
846 Reg = MulVal.getNode()->getOperand(0);
847 ConstantSDNode *AddVal =
848 cast<ConstantSDNode>(MulVal.getNode()->getOperand(1));
849 uint64_t Disp = AM.Disp + AddVal->getSExtValue() *
852 X86::isOffsetSuitableForCodeModel(Disp, M,
853 AM.hasSymbolicDisplacement()))
856 Reg = N.getNode()->getOperand(0);
858 Reg = N.getNode()->getOperand(0);
861 AM.IndexReg = AM.Base_Reg = Reg;
868 // Given A-B, if A can be completely folded into the address and
869 // the index field with the index field unused, use -B as the index.
870 // This is a win if a has multiple parts that can be folded into
871 // the address. Also, this saves a mov if the base register has
872 // other uses, since it avoids a two-address sub instruction, however
873 // it costs an additional mov if the index register has other uses.
875 // Add an artificial use to this node so that we can keep track of
876 // it if it gets CSE'd with a different node.
877 HandleSDNode Handle(N);
879 // Test if the LHS of the sub can be folded.
880 X86ISelAddressMode Backup = AM;
881 if (MatchAddressRecursively(N.getNode()->getOperand(0), AM, Depth+1)) {
885 // Test if the index field is free for use.
886 if (AM.IndexReg.getNode() || AM.isRIPRelative()) {
892 SDValue RHS = Handle.getValue().getNode()->getOperand(1);
893 // If the RHS involves a register with multiple uses, this
894 // transformation incurs an extra mov, due to the neg instruction
895 // clobbering its operand.
896 if (!RHS.getNode()->hasOneUse() ||
897 RHS.getNode()->getOpcode() == ISD::CopyFromReg ||
898 RHS.getNode()->getOpcode() == ISD::TRUNCATE ||
899 RHS.getNode()->getOpcode() == ISD::ANY_EXTEND ||
900 (RHS.getNode()->getOpcode() == ISD::ZERO_EXTEND &&
901 RHS.getNode()->getOperand(0).getValueType() == MVT::i32))
903 // If the base is a register with multiple uses, this
904 // transformation may save a mov.
905 if ((AM.BaseType == X86ISelAddressMode::RegBase &&
906 AM.Base_Reg.getNode() &&
907 !AM.Base_Reg.getNode()->hasOneUse()) ||
908 AM.BaseType == X86ISelAddressMode::FrameIndexBase)
910 // If the folded LHS was interesting, this transformation saves
911 // address arithmetic.
912 if ((AM.hasSymbolicDisplacement() && !Backup.hasSymbolicDisplacement()) +
913 ((AM.Disp != 0) && (Backup.Disp == 0)) +
914 (AM.Segment.getNode() && !Backup.Segment.getNode()) >= 2)
916 // If it doesn't look like it may be an overall win, don't do it.
922 // Ok, the transformation is legal and appears profitable. Go for it.
923 SDValue Zero = CurDAG->getConstant(0, N.getValueType());
924 SDValue Neg = CurDAG->getNode(ISD::SUB, dl, N.getValueType(), Zero, RHS);
928 // Insert the new nodes into the topological ordering.
929 if (Zero.getNode()->getNodeId() == -1 ||
930 Zero.getNode()->getNodeId() > N.getNode()->getNodeId()) {
931 CurDAG->RepositionNode(N.getNode(), Zero.getNode());
932 Zero.getNode()->setNodeId(N.getNode()->getNodeId());
934 if (Neg.getNode()->getNodeId() == -1 ||
935 Neg.getNode()->getNodeId() > N.getNode()->getNodeId()) {
936 CurDAG->RepositionNode(N.getNode(), Neg.getNode());
937 Neg.getNode()->setNodeId(N.getNode()->getNodeId());
943 // Add an artificial use to this node so that we can keep track of
944 // it if it gets CSE'd with a different node.
945 HandleSDNode Handle(N);
946 SDValue LHS = Handle.getValue().getNode()->getOperand(0);
947 SDValue RHS = Handle.getValue().getNode()->getOperand(1);
949 X86ISelAddressMode Backup = AM;
950 if (!MatchAddressRecursively(LHS, AM, Depth+1) &&
951 !MatchAddressRecursively(RHS, AM, Depth+1))
954 LHS = Handle.getValue().getNode()->getOperand(0);
955 RHS = Handle.getValue().getNode()->getOperand(1);
957 // Try again after commuting the operands.
958 if (!MatchAddressRecursively(RHS, AM, Depth+1) &&
959 !MatchAddressRecursively(LHS, AM, Depth+1))
962 LHS = Handle.getValue().getNode()->getOperand(0);
963 RHS = Handle.getValue().getNode()->getOperand(1);
965 // If we couldn't fold both operands into the address at the same time,
966 // see if we can just put each operand into a register and fold at least
968 if (AM.BaseType == X86ISelAddressMode::RegBase &&
969 !AM.Base_Reg.getNode() &&
970 !AM.IndexReg.getNode()) {
980 // Handle "X | C" as "X + C" iff X is known to have C bits clear.
981 if (isLogicallyAddWithConstant(N, CurDAG)) {
982 X86ISelAddressMode Backup = AM;
983 ConstantSDNode *CN = cast<ConstantSDNode>(N.getOperand(1));
984 uint64_t Offset = CN->getSExtValue();
986 // Start with the LHS as an addr mode.
987 if (!MatchAddressRecursively(N.getOperand(0), AM, Depth+1) &&
988 // Address could not have picked a GV address for the displacement.
990 // On x86-64, the resultant disp must fit in 32-bits.
992 X86::isOffsetSuitableForCodeModel(AM.Disp + Offset, M,
993 AM.hasSymbolicDisplacement()))) {
1002 // Perform some heroic transforms on an and of a constant-count shift
1003 // with a constant to enable use of the scaled offset field.
1005 SDValue Shift = N.getOperand(0);
1006 if (Shift.getNumOperands() != 2) break;
1008 // Scale must not be used already.
1009 if (AM.IndexReg.getNode() != 0 || AM.Scale != 1) break;
1011 SDValue X = Shift.getOperand(0);
1012 ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N.getOperand(1));
1013 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(Shift.getOperand(1));
1014 if (!C1 || !C2) break;
1016 // Handle "(X >> (8-C1)) & C2" as "(X >> 8) & 0xff)" if safe. This
1017 // allows us to convert the shift and and into an h-register extract and
1019 if (Shift.getOpcode() == ISD::SRL && Shift.hasOneUse()) {
1020 unsigned ScaleLog = 8 - C1->getZExtValue();
1021 if (ScaleLog > 0 && ScaleLog < 4 &&
1022 C2->getZExtValue() == (UINT64_C(0xff) << ScaleLog)) {
1023 SDValue Eight = CurDAG->getConstant(8, MVT::i8);
1024 SDValue Mask = CurDAG->getConstant(0xff, N.getValueType());
1025 SDValue Srl = CurDAG->getNode(ISD::SRL, dl, N.getValueType(),
1027 SDValue And = CurDAG->getNode(ISD::AND, dl, N.getValueType(),
1029 SDValue ShlCount = CurDAG->getConstant(ScaleLog, MVT::i8);
1030 SDValue Shl = CurDAG->getNode(ISD::SHL, dl, N.getValueType(),
1033 // Insert the new nodes into the topological ordering.
1034 if (Eight.getNode()->getNodeId() == -1 ||
1035 Eight.getNode()->getNodeId() > X.getNode()->getNodeId()) {
1036 CurDAG->RepositionNode(X.getNode(), Eight.getNode());
1037 Eight.getNode()->setNodeId(X.getNode()->getNodeId());
1039 if (Mask.getNode()->getNodeId() == -1 ||
1040 Mask.getNode()->getNodeId() > X.getNode()->getNodeId()) {
1041 CurDAG->RepositionNode(X.getNode(), Mask.getNode());
1042 Mask.getNode()->setNodeId(X.getNode()->getNodeId());
1044 if (Srl.getNode()->getNodeId() == -1 ||
1045 Srl.getNode()->getNodeId() > Shift.getNode()->getNodeId()) {
1046 CurDAG->RepositionNode(Shift.getNode(), Srl.getNode());
1047 Srl.getNode()->setNodeId(Shift.getNode()->getNodeId());
1049 if (And.getNode()->getNodeId() == -1 ||
1050 And.getNode()->getNodeId() > N.getNode()->getNodeId()) {
1051 CurDAG->RepositionNode(N.getNode(), And.getNode());
1052 And.getNode()->setNodeId(N.getNode()->getNodeId());
1054 if (ShlCount.getNode()->getNodeId() == -1 ||
1055 ShlCount.getNode()->getNodeId() > X.getNode()->getNodeId()) {
1056 CurDAG->RepositionNode(X.getNode(), ShlCount.getNode());
1057 ShlCount.getNode()->setNodeId(N.getNode()->getNodeId());
1059 if (Shl.getNode()->getNodeId() == -1 ||
1060 Shl.getNode()->getNodeId() > N.getNode()->getNodeId()) {
1061 CurDAG->RepositionNode(N.getNode(), Shl.getNode());
1062 Shl.getNode()->setNodeId(N.getNode()->getNodeId());
1064 CurDAG->ReplaceAllUsesWith(N, Shl);
1066 AM.Scale = (1 << ScaleLog);
1071 // Handle "(X << C1) & C2" as "(X & (C2>>C1)) << C1" if safe and if this
1072 // allows us to fold the shift into this addressing mode.
1073 if (Shift.getOpcode() != ISD::SHL) break;
1075 // Not likely to be profitable if either the AND or SHIFT node has more
1076 // than one use (unless all uses are for address computation). Besides,
1077 // isel mechanism requires their node ids to be reused.
1078 if (!N.hasOneUse() || !Shift.hasOneUse())
1081 // Verify that the shift amount is something we can fold.
1082 unsigned ShiftCst = C1->getZExtValue();
1083 if (ShiftCst != 1 && ShiftCst != 2 && ShiftCst != 3)
1086 // Get the new AND mask, this folds to a constant.
1087 SDValue NewANDMask = CurDAG->getNode(ISD::SRL, dl, N.getValueType(),
1088 SDValue(C2, 0), SDValue(C1, 0));
1089 SDValue NewAND = CurDAG->getNode(ISD::AND, dl, N.getValueType(), X,
1091 SDValue NewSHIFT = CurDAG->getNode(ISD::SHL, dl, N.getValueType(),
1092 NewAND, SDValue(C1, 0));
1094 // Insert the new nodes into the topological ordering.
1095 if (C1->getNodeId() > X.getNode()->getNodeId()) {
1096 CurDAG->RepositionNode(X.getNode(), C1);
1097 C1->setNodeId(X.getNode()->getNodeId());
1099 if (NewANDMask.getNode()->getNodeId() == -1 ||
1100 NewANDMask.getNode()->getNodeId() > X.getNode()->getNodeId()) {
1101 CurDAG->RepositionNode(X.getNode(), NewANDMask.getNode());
1102 NewANDMask.getNode()->setNodeId(X.getNode()->getNodeId());
1104 if (NewAND.getNode()->getNodeId() == -1 ||
1105 NewAND.getNode()->getNodeId() > Shift.getNode()->getNodeId()) {
1106 CurDAG->RepositionNode(Shift.getNode(), NewAND.getNode());
1107 NewAND.getNode()->setNodeId(Shift.getNode()->getNodeId());
1109 if (NewSHIFT.getNode()->getNodeId() == -1 ||
1110 NewSHIFT.getNode()->getNodeId() > N.getNode()->getNodeId()) {
1111 CurDAG->RepositionNode(N.getNode(), NewSHIFT.getNode());
1112 NewSHIFT.getNode()->setNodeId(N.getNode()->getNodeId());
1115 CurDAG->ReplaceAllUsesWith(N, NewSHIFT);
1117 AM.Scale = 1 << ShiftCst;
1118 AM.IndexReg = NewAND;
1123 return MatchAddressBase(N, AM);
1126 /// MatchAddressBase - Helper for MatchAddress. Add the specified node to the
1127 /// specified addressing mode without any further recursion.
1128 bool X86DAGToDAGISel::MatchAddressBase(SDValue N, X86ISelAddressMode &AM) {
1129 // Is the base register already occupied?
1130 if (AM.BaseType != X86ISelAddressMode::RegBase || AM.Base_Reg.getNode()) {
1131 // If so, check to see if the scale index register is set.
1132 if (AM.IndexReg.getNode() == 0) {
1138 // Otherwise, we cannot select it.
1142 // Default, generate it as a register.
1143 AM.BaseType = X86ISelAddressMode::RegBase;
1148 /// SelectAddr - returns true if it is able pattern match an addressing mode.
1149 /// It returns the operands which make up the maximal addressing mode it can
1150 /// match by reference.
1151 bool X86DAGToDAGISel::SelectAddr(SDNode *Op, SDValue N, SDValue &Base,
1152 SDValue &Scale, SDValue &Index,
1153 SDValue &Disp, SDValue &Segment) {
1154 X86ISelAddressMode AM;
1155 if (MatchAddress(N, AM))
1158 EVT VT = N.getValueType();
1159 if (AM.BaseType == X86ISelAddressMode::RegBase) {
1160 if (!AM.Base_Reg.getNode())
1161 AM.Base_Reg = CurDAG->getRegister(0, VT);
1164 if (!AM.IndexReg.getNode())
1165 AM.IndexReg = CurDAG->getRegister(0, VT);
1167 getAddressOperands(AM, Base, Scale, Index, Disp, Segment);
1171 /// SelectScalarSSELoad - Match a scalar SSE load. In particular, we want to
1172 /// match a load whose top elements are either undef or zeros. The load flavor
1173 /// is derived from the type of N, which is either v4f32 or v2f64.
1176 /// PatternChainNode: this is the matched node that has a chain input and
1178 bool X86DAGToDAGISel::SelectScalarSSELoad(SDNode *Root,
1179 SDValue N, SDValue &Base,
1180 SDValue &Scale, SDValue &Index,
1181 SDValue &Disp, SDValue &Segment,
1182 SDValue &PatternNodeWithChain) {
1183 if (N.getOpcode() == ISD::SCALAR_TO_VECTOR) {
1184 PatternNodeWithChain = N.getOperand(0);
1185 if (ISD::isNON_EXTLoad(PatternNodeWithChain.getNode()) &&
1186 PatternNodeWithChain.hasOneUse() &&
1187 IsProfitableToFold(N.getOperand(0), N.getNode(), Root) &&
1188 IsLegalToFold(N.getOperand(0), N.getNode(), Root, OptLevel)) {
1189 LoadSDNode *LD = cast<LoadSDNode>(PatternNodeWithChain);
1190 if (!SelectAddr(Root, LD->getBasePtr(), Base, Scale, Index, Disp,Segment))
1196 // Also handle the case where we explicitly require zeros in the top
1197 // elements. This is a vector shuffle from the zero vector.
1198 if (N.getOpcode() == X86ISD::VZEXT_MOVL && N.getNode()->hasOneUse() &&
1199 // Check to see if the top elements are all zeros (or bitcast of zeros).
1200 N.getOperand(0).getOpcode() == ISD::SCALAR_TO_VECTOR &&
1201 N.getOperand(0).getNode()->hasOneUse() &&
1202 ISD::isNON_EXTLoad(N.getOperand(0).getOperand(0).getNode()) &&
1203 N.getOperand(0).getOperand(0).hasOneUse() &&
1204 IsProfitableToFold(N.getOperand(0), N.getNode(), Root) &&
1205 IsLegalToFold(N.getOperand(0), N.getNode(), Root, OptLevel)) {
1206 // Okay, this is a zero extending load. Fold it.
1207 LoadSDNode *LD = cast<LoadSDNode>(N.getOperand(0).getOperand(0));
1208 if (!SelectAddr(Root, LD->getBasePtr(), Base, Scale, Index, Disp, Segment))
1210 PatternNodeWithChain = SDValue(LD, 0);
1217 /// SelectLEAAddr - it calls SelectAddr and determines if the maximal addressing
1218 /// mode it matches can be cost effectively emitted as an LEA instruction.
1219 bool X86DAGToDAGISel::SelectLEAAddr(SDNode *Op, SDValue N,
1220 SDValue &Base, SDValue &Scale,
1221 SDValue &Index, SDValue &Disp,
1223 X86ISelAddressMode AM;
1225 // Set AM.Segment to prevent MatchAddress from using one. LEA doesn't support
1227 SDValue Copy = AM.Segment;
1228 SDValue T = CurDAG->getRegister(0, MVT::i32);
1230 if (MatchAddress(N, AM))
1232 assert (T == AM.Segment);
1235 EVT VT = N.getValueType();
1236 unsigned Complexity = 0;
1237 if (AM.BaseType == X86ISelAddressMode::RegBase)
1238 if (AM.Base_Reg.getNode())
1241 AM.Base_Reg = CurDAG->getRegister(0, VT);
1242 else if (AM.BaseType == X86ISelAddressMode::FrameIndexBase)
1245 if (AM.IndexReg.getNode())
1248 AM.IndexReg = CurDAG->getRegister(0, VT);
1250 // Don't match just leal(,%reg,2). It's cheaper to do addl %reg, %reg, or with
1255 // FIXME: We are artificially lowering the criteria to turn ADD %reg, $GA
1256 // to a LEA. This is determined with some expermentation but is by no means
1257 // optimal (especially for code size consideration). LEA is nice because of
1258 // its three-address nature. Tweak the cost function again when we can run
1259 // convertToThreeAddress() at register allocation time.
1260 if (AM.hasSymbolicDisplacement()) {
1261 // For X86-64, we should always use lea to materialize RIP relative
1263 if (Subtarget->is64Bit())
1269 if (AM.Disp && (AM.Base_Reg.getNode() || AM.IndexReg.getNode()))
1272 // If it isn't worth using an LEA, reject it.
1273 if (Complexity <= 2)
1276 getAddressOperands(AM, Base, Scale, Index, Disp, Segment);
1280 /// SelectTLSADDRAddr - This is only run on TargetGlobalTLSAddress nodes.
1281 bool X86DAGToDAGISel::SelectTLSADDRAddr(SDNode *Op, SDValue N, SDValue &Base,
1282 SDValue &Scale, SDValue &Index,
1283 SDValue &Disp, SDValue &Segment) {
1284 assert(N.getOpcode() == ISD::TargetGlobalTLSAddress);
1285 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(N);
1287 X86ISelAddressMode AM;
1288 AM.GV = GA->getGlobal();
1289 AM.Disp += GA->getOffset();
1290 AM.Base_Reg = CurDAG->getRegister(0, N.getValueType());
1291 AM.SymbolFlags = GA->getTargetFlags();
1293 if (N.getValueType() == MVT::i32) {
1295 AM.IndexReg = CurDAG->getRegister(X86::EBX, MVT::i32);
1297 AM.IndexReg = CurDAG->getRegister(0, MVT::i64);
1300 getAddressOperands(AM, Base, Scale, Index, Disp, Segment);
1305 bool X86DAGToDAGISel::TryFoldLoad(SDNode *P, SDValue N,
1306 SDValue &Base, SDValue &Scale,
1307 SDValue &Index, SDValue &Disp,
1309 if (!ISD::isNON_EXTLoad(N.getNode()) ||
1310 !IsProfitableToFold(N, P, P) ||
1311 !IsLegalToFold(N, P, P, OptLevel))
1314 return SelectAddr(P, N.getOperand(1), Base, Scale, Index, Disp, Segment);
1317 /// getGlobalBaseReg - Return an SDNode that returns the value of
1318 /// the global base register. Output instructions required to
1319 /// initialize the global base register, if necessary.
1321 SDNode *X86DAGToDAGISel::getGlobalBaseReg() {
1322 unsigned GlobalBaseReg = getInstrInfo()->getGlobalBaseReg(MF);
1323 return CurDAG->getRegister(GlobalBaseReg, TLI.getPointerTy()).getNode();
1326 SDNode *X86DAGToDAGISel::SelectAtomic64(SDNode *Node, unsigned Opc) {
1327 SDValue Chain = Node->getOperand(0);
1328 SDValue In1 = Node->getOperand(1);
1329 SDValue In2L = Node->getOperand(2);
1330 SDValue In2H = Node->getOperand(3);
1331 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1332 if (!SelectAddr(In1.getNode(), In1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4))
1334 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
1335 MemOp[0] = cast<MemSDNode>(Node)->getMemOperand();
1336 const SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, In2L, In2H, Chain};
1337 SDNode *ResNode = CurDAG->getMachineNode(Opc, Node->getDebugLoc(),
1338 MVT::i32, MVT::i32, MVT::Other, Ops,
1339 array_lengthof(Ops));
1340 cast<MachineSDNode>(ResNode)->setMemRefs(MemOp, MemOp + 1);
1344 SDNode *X86DAGToDAGISel::SelectAtomicLoadAdd(SDNode *Node, EVT NVT) {
1345 if (Node->hasAnyUseOfValue(0))
1348 // Optimize common patterns for __sync_add_and_fetch and
1349 // __sync_sub_and_fetch where the result is not used. This allows us
1350 // to use "lock" version of add, sub, inc, dec instructions.
1351 // FIXME: Do not use special instructions but instead add the "lock"
1352 // prefix to the target node somehow. The extra information will then be
1353 // transferred to machine instruction and it denotes the prefix.
1354 SDValue Chain = Node->getOperand(0);
1355 SDValue Ptr = Node->getOperand(1);
1356 SDValue Val = Node->getOperand(2);
1357 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1358 if (!SelectAddr(Ptr.getNode(), Ptr, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4))
1361 bool isInc = false, isDec = false, isSub = false, isCN = false;
1362 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Val);
1365 int64_t CNVal = CN->getSExtValue();
1368 else if (CNVal == -1)
1370 else if (CNVal >= 0)
1371 Val = CurDAG->getTargetConstant(CNVal, NVT);
1374 Val = CurDAG->getTargetConstant(-CNVal, NVT);
1376 } else if (Val.hasOneUse() &&
1377 Val.getOpcode() == ISD::SUB &&
1378 X86::isZeroNode(Val.getOperand(0))) {
1380 Val = Val.getOperand(1);
1384 switch (NVT.getSimpleVT().SimpleTy) {
1388 Opc = X86::LOCK_INC8m;
1390 Opc = X86::LOCK_DEC8m;
1393 Opc = X86::LOCK_SUB8mi;
1395 Opc = X86::LOCK_SUB8mr;
1398 Opc = X86::LOCK_ADD8mi;
1400 Opc = X86::LOCK_ADD8mr;
1405 Opc = X86::LOCK_INC16m;
1407 Opc = X86::LOCK_DEC16m;
1410 if (immSext8(Val.getNode()))
1411 Opc = X86::LOCK_SUB16mi8;
1413 Opc = X86::LOCK_SUB16mi;
1415 Opc = X86::LOCK_SUB16mr;
1418 if (immSext8(Val.getNode()))
1419 Opc = X86::LOCK_ADD16mi8;
1421 Opc = X86::LOCK_ADD16mi;
1423 Opc = X86::LOCK_ADD16mr;
1428 Opc = X86::LOCK_INC32m;
1430 Opc = X86::LOCK_DEC32m;
1433 if (immSext8(Val.getNode()))
1434 Opc = X86::LOCK_SUB32mi8;
1436 Opc = X86::LOCK_SUB32mi;
1438 Opc = X86::LOCK_SUB32mr;
1441 if (immSext8(Val.getNode()))
1442 Opc = X86::LOCK_ADD32mi8;
1444 Opc = X86::LOCK_ADD32mi;
1446 Opc = X86::LOCK_ADD32mr;
1451 Opc = X86::LOCK_INC64m;
1453 Opc = X86::LOCK_DEC64m;
1455 Opc = X86::LOCK_SUB64mr;
1457 if (immSext8(Val.getNode()))
1458 Opc = X86::LOCK_SUB64mi8;
1459 else if (i64immSExt32(Val.getNode()))
1460 Opc = X86::LOCK_SUB64mi32;
1463 Opc = X86::LOCK_ADD64mr;
1465 if (immSext8(Val.getNode()))
1466 Opc = X86::LOCK_ADD64mi8;
1467 else if (i64immSExt32(Val.getNode()))
1468 Opc = X86::LOCK_ADD64mi32;
1474 DebugLoc dl = Node->getDebugLoc();
1475 SDValue Undef = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1477 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
1478 MemOp[0] = cast<MemSDNode>(Node)->getMemOperand();
1479 if (isInc || isDec) {
1480 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Chain };
1481 SDValue Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 6), 0);
1482 cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
1483 SDValue RetVals[] = { Undef, Ret };
1484 return CurDAG->getMergeValues(RetVals, 2, dl).getNode();
1486 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Val, Chain };
1487 SDValue Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 7), 0);
1488 cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
1489 SDValue RetVals[] = { Undef, Ret };
1490 return CurDAG->getMergeValues(RetVals, 2, dl).getNode();
1494 /// HasNoSignedComparisonUses - Test whether the given X86ISD::CMP node has
1495 /// any uses which require the SF or OF bits to be accurate.
1496 static bool HasNoSignedComparisonUses(SDNode *N) {
1497 // Examine each user of the node.
1498 for (SDNode::use_iterator UI = N->use_begin(),
1499 UE = N->use_end(); UI != UE; ++UI) {
1500 // Only examine CopyToReg uses.
1501 if (UI->getOpcode() != ISD::CopyToReg)
1503 // Only examine CopyToReg uses that copy to EFLAGS.
1504 if (cast<RegisterSDNode>(UI->getOperand(1))->getReg() !=
1507 // Examine each user of the CopyToReg use.
1508 for (SDNode::use_iterator FlagUI = UI->use_begin(),
1509 FlagUE = UI->use_end(); FlagUI != FlagUE; ++FlagUI) {
1510 // Only examine the Flag result.
1511 if (FlagUI.getUse().getResNo() != 1) continue;
1512 // Anything unusual: assume conservatively.
1513 if (!FlagUI->isMachineOpcode()) return false;
1514 // Examine the opcode of the user.
1515 switch (FlagUI->getMachineOpcode()) {
1516 // These comparisons don't treat the most significant bit specially.
1517 case X86::SETAr: case X86::SETAEr: case X86::SETBr: case X86::SETBEr:
1518 case X86::SETEr: case X86::SETNEr: case X86::SETPr: case X86::SETNPr:
1519 case X86::SETAm: case X86::SETAEm: case X86::SETBm: case X86::SETBEm:
1520 case X86::SETEm: case X86::SETNEm: case X86::SETPm: case X86::SETNPm:
1521 case X86::JA_4: case X86::JAE_4: case X86::JB_4: case X86::JBE_4:
1522 case X86::JE_4: case X86::JNE_4: case X86::JP_4: case X86::JNP_4:
1523 case X86::CMOVA16rr: case X86::CMOVA16rm:
1524 case X86::CMOVA32rr: case X86::CMOVA32rm:
1525 case X86::CMOVA64rr: case X86::CMOVA64rm:
1526 case X86::CMOVAE16rr: case X86::CMOVAE16rm:
1527 case X86::CMOVAE32rr: case X86::CMOVAE32rm:
1528 case X86::CMOVAE64rr: case X86::CMOVAE64rm:
1529 case X86::CMOVB16rr: case X86::CMOVB16rm:
1530 case X86::CMOVB32rr: case X86::CMOVB32rm:
1531 case X86::CMOVB64rr: case X86::CMOVB64rm:
1532 case X86::CMOVBE16rr: case X86::CMOVBE16rm:
1533 case X86::CMOVBE32rr: case X86::CMOVBE32rm:
1534 case X86::CMOVBE64rr: case X86::CMOVBE64rm:
1535 case X86::CMOVE16rr: case X86::CMOVE16rm:
1536 case X86::CMOVE32rr: case X86::CMOVE32rm:
1537 case X86::CMOVE64rr: case X86::CMOVE64rm:
1538 case X86::CMOVNE16rr: case X86::CMOVNE16rm:
1539 case X86::CMOVNE32rr: case X86::CMOVNE32rm:
1540 case X86::CMOVNE64rr: case X86::CMOVNE64rm:
1541 case X86::CMOVNP16rr: case X86::CMOVNP16rm:
1542 case X86::CMOVNP32rr: case X86::CMOVNP32rm:
1543 case X86::CMOVNP64rr: case X86::CMOVNP64rm:
1544 case X86::CMOVP16rr: case X86::CMOVP16rm:
1545 case X86::CMOVP32rr: case X86::CMOVP32rm:
1546 case X86::CMOVP64rr: case X86::CMOVP64rm:
1548 // Anything else: assume conservatively.
1549 default: return false;
1556 SDNode *X86DAGToDAGISel::Select(SDNode *Node) {
1557 EVT NVT = Node->getValueType(0);
1559 unsigned Opcode = Node->getOpcode();
1560 DebugLoc dl = Node->getDebugLoc();
1562 DEBUG(dbgs() << "Selecting: "; Node->dump(CurDAG); dbgs() << '\n');
1564 if (Node->isMachineOpcode()) {
1565 DEBUG(dbgs() << "== "; Node->dump(CurDAG); dbgs() << '\n');
1566 return NULL; // Already selected.
1571 case X86ISD::GlobalBaseReg:
1572 return getGlobalBaseReg();
1574 case X86ISD::ATOMOR64_DAG:
1575 return SelectAtomic64(Node, X86::ATOMOR6432);
1576 case X86ISD::ATOMXOR64_DAG:
1577 return SelectAtomic64(Node, X86::ATOMXOR6432);
1578 case X86ISD::ATOMADD64_DAG:
1579 return SelectAtomic64(Node, X86::ATOMADD6432);
1580 case X86ISD::ATOMSUB64_DAG:
1581 return SelectAtomic64(Node, X86::ATOMSUB6432);
1582 case X86ISD::ATOMNAND64_DAG:
1583 return SelectAtomic64(Node, X86::ATOMNAND6432);
1584 case X86ISD::ATOMAND64_DAG:
1585 return SelectAtomic64(Node, X86::ATOMAND6432);
1586 case X86ISD::ATOMSWAP64_DAG:
1587 return SelectAtomic64(Node, X86::ATOMSWAP6432);
1589 case ISD::ATOMIC_LOAD_ADD: {
1590 SDNode *RetVal = SelectAtomicLoadAdd(Node, NVT);
1596 case ISD::SMUL_LOHI:
1597 case ISD::UMUL_LOHI: {
1598 SDValue N0 = Node->getOperand(0);
1599 SDValue N1 = Node->getOperand(1);
1601 bool isSigned = Opcode == ISD::SMUL_LOHI;
1603 switch (NVT.getSimpleVT().SimpleTy) {
1604 default: llvm_unreachable("Unsupported VT!");
1605 case MVT::i8: Opc = X86::MUL8r; MOpc = X86::MUL8m; break;
1606 case MVT::i16: Opc = X86::MUL16r; MOpc = X86::MUL16m; break;
1607 case MVT::i32: Opc = X86::MUL32r; MOpc = X86::MUL32m; break;
1608 case MVT::i64: Opc = X86::MUL64r; MOpc = X86::MUL64m; break;
1611 switch (NVT.getSimpleVT().SimpleTy) {
1612 default: llvm_unreachable("Unsupported VT!");
1613 case MVT::i8: Opc = X86::IMUL8r; MOpc = X86::IMUL8m; break;
1614 case MVT::i16: Opc = X86::IMUL16r; MOpc = X86::IMUL16m; break;
1615 case MVT::i32: Opc = X86::IMUL32r; MOpc = X86::IMUL32m; break;
1616 case MVT::i64: Opc = X86::IMUL64r; MOpc = X86::IMUL64m; break;
1620 unsigned LoReg, HiReg;
1621 switch (NVT.getSimpleVT().SimpleTy) {
1622 default: llvm_unreachable("Unsupported VT!");
1623 case MVT::i8: LoReg = X86::AL; HiReg = X86::AH; break;
1624 case MVT::i16: LoReg = X86::AX; HiReg = X86::DX; break;
1625 case MVT::i32: LoReg = X86::EAX; HiReg = X86::EDX; break;
1626 case MVT::i64: LoReg = X86::RAX; HiReg = X86::RDX; break;
1629 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1630 bool foldedLoad = TryFoldLoad(Node, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
1631 // Multiply is commmutative.
1633 foldedLoad = TryFoldLoad(Node, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
1638 SDValue InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, LoReg,
1639 N0, SDValue()).getValue(1);
1642 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
1645 CurDAG->getMachineNode(MOpc, dl, MVT::Other, MVT::Flag, Ops,
1646 array_lengthof(Ops));
1647 InFlag = SDValue(CNode, 1);
1648 // Update the chain.
1649 ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
1652 SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Flag, N1, InFlag), 0);
1655 // Prevent use of AH in a REX instruction by referencing AX instead.
1656 if (HiReg == X86::AH && Subtarget->is64Bit() &&
1657 !SDValue(Node, 1).use_empty()) {
1658 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1659 X86::AX, MVT::i16, InFlag);
1660 InFlag = Result.getValue(2);
1661 // Get the low part if needed. Don't use getCopyFromReg for aliasing
1663 if (!SDValue(Node, 0).use_empty())
1664 ReplaceUses(SDValue(Node, 1),
1665 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
1667 // Shift AX down 8 bits.
1668 Result = SDValue(CurDAG->getMachineNode(X86::SHR16ri, dl, MVT::i16,
1670 CurDAG->getTargetConstant(8, MVT::i8)), 0);
1671 // Then truncate it down to i8.
1672 ReplaceUses(SDValue(Node, 1),
1673 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
1675 // Copy the low half of the result, if it is needed.
1676 if (!SDValue(Node, 0).use_empty()) {
1677 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1678 LoReg, NVT, InFlag);
1679 InFlag = Result.getValue(2);
1680 ReplaceUses(SDValue(Node, 0), Result);
1681 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
1683 // Copy the high half of the result, if it is needed.
1684 if (!SDValue(Node, 1).use_empty()) {
1685 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1686 HiReg, NVT, InFlag);
1687 InFlag = Result.getValue(2);
1688 ReplaceUses(SDValue(Node, 1), Result);
1689 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
1696 case ISD::UDIVREM: {
1697 SDValue N0 = Node->getOperand(0);
1698 SDValue N1 = Node->getOperand(1);
1700 bool isSigned = Opcode == ISD::SDIVREM;
1702 switch (NVT.getSimpleVT().SimpleTy) {
1703 default: llvm_unreachable("Unsupported VT!");
1704 case MVT::i8: Opc = X86::DIV8r; MOpc = X86::DIV8m; break;
1705 case MVT::i16: Opc = X86::DIV16r; MOpc = X86::DIV16m; break;
1706 case MVT::i32: Opc = X86::DIV32r; MOpc = X86::DIV32m; break;
1707 case MVT::i64: Opc = X86::DIV64r; MOpc = X86::DIV64m; break;
1710 switch (NVT.getSimpleVT().SimpleTy) {
1711 default: llvm_unreachable("Unsupported VT!");
1712 case MVT::i8: Opc = X86::IDIV8r; MOpc = X86::IDIV8m; break;
1713 case MVT::i16: Opc = X86::IDIV16r; MOpc = X86::IDIV16m; break;
1714 case MVT::i32: Opc = X86::IDIV32r; MOpc = X86::IDIV32m; break;
1715 case MVT::i64: Opc = X86::IDIV64r; MOpc = X86::IDIV64m; break;
1719 unsigned LoReg, HiReg, ClrReg;
1720 unsigned ClrOpcode, SExtOpcode;
1721 switch (NVT.getSimpleVT().SimpleTy) {
1722 default: llvm_unreachable("Unsupported VT!");
1724 LoReg = X86::AL; ClrReg = HiReg = X86::AH;
1726 SExtOpcode = X86::CBW;
1729 LoReg = X86::AX; HiReg = X86::DX;
1730 ClrOpcode = X86::MOV16r0; ClrReg = X86::DX;
1731 SExtOpcode = X86::CWD;
1734 LoReg = X86::EAX; ClrReg = HiReg = X86::EDX;
1735 ClrOpcode = X86::MOV32r0;
1736 SExtOpcode = X86::CDQ;
1739 LoReg = X86::RAX; ClrReg = HiReg = X86::RDX;
1740 ClrOpcode = X86::MOV64r0;
1741 SExtOpcode = X86::CQO;
1745 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
1746 bool foldedLoad = TryFoldLoad(Node, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
1747 bool signBitIsZero = CurDAG->SignBitIsZero(N0);
1750 if (NVT == MVT::i8 && (!isSigned || signBitIsZero)) {
1751 // Special case for div8, just use a move with zero extension to AX to
1752 // clear the upper 8 bits (AH).
1753 SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Move, Chain;
1754 if (TryFoldLoad(Node, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4)) {
1755 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N0.getOperand(0) };
1757 SDValue(CurDAG->getMachineNode(X86::MOVZX16rm8, dl, MVT::i16,
1759 array_lengthof(Ops)), 0);
1760 Chain = Move.getValue(1);
1761 ReplaceUses(N0.getValue(1), Chain);
1764 SDValue(CurDAG->getMachineNode(X86::MOVZX16rr8, dl, MVT::i16, N0),0);
1765 Chain = CurDAG->getEntryNode();
1767 Chain = CurDAG->getCopyToReg(Chain, dl, X86::AX, Move, SDValue());
1768 InFlag = Chain.getValue(1);
1771 CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl,
1772 LoReg, N0, SDValue()).getValue(1);
1773 if (isSigned && !signBitIsZero) {
1774 // Sign extend the low part into the high part.
1776 SDValue(CurDAG->getMachineNode(SExtOpcode, dl, MVT::Flag, InFlag),0);
1778 // Zero out the high part, effectively zero extending the input.
1780 SDValue(CurDAG->getMachineNode(ClrOpcode, dl, NVT), 0);
1781 InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, ClrReg,
1782 ClrNode, InFlag).getValue(1);
1787 SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
1790 CurDAG->getMachineNode(MOpc, dl, MVT::Other, MVT::Flag, Ops,
1791 array_lengthof(Ops));
1792 InFlag = SDValue(CNode, 1);
1793 // Update the chain.
1794 ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
1797 SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Flag, N1, InFlag), 0);
1800 // Prevent use of AH in a REX instruction by referencing AX instead.
1801 // Shift it down 8 bits.
1802 if (HiReg == X86::AH && Subtarget->is64Bit() &&
1803 !SDValue(Node, 1).use_empty()) {
1804 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1805 X86::AX, MVT::i16, InFlag);
1806 InFlag = Result.getValue(2);
1808 // If we also need AL (the quotient), get it by extracting a subreg from
1809 // Result. The fast register allocator does not like multiple CopyFromReg
1810 // nodes using aliasing registers.
1811 if (!SDValue(Node, 0).use_empty())
1812 ReplaceUses(SDValue(Node, 0),
1813 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
1815 // Shift AX right by 8 bits instead of using AH.
1816 Result = SDValue(CurDAG->getMachineNode(X86::SHR16ri, dl, MVT::i16,
1818 CurDAG->getTargetConstant(8, MVT::i8)),
1820 ReplaceUses(SDValue(Node, 1),
1821 CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
1823 // Copy the division (low) result, if it is needed.
1824 if (!SDValue(Node, 0).use_empty()) {
1825 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1826 LoReg, NVT, InFlag);
1827 InFlag = Result.getValue(2);
1828 ReplaceUses(SDValue(Node, 0), Result);
1829 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
1831 // Copy the remainder (high) result, if it is needed.
1832 if (!SDValue(Node, 1).use_empty()) {
1833 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
1834 HiReg, NVT, InFlag);
1835 InFlag = Result.getValue(2);
1836 ReplaceUses(SDValue(Node, 1), Result);
1837 DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
1843 SDValue N0 = Node->getOperand(0);
1844 SDValue N1 = Node->getOperand(1);
1846 // Look for (X86cmp (and $op, $imm), 0) and see if we can convert it to
1847 // use a smaller encoding.
1848 if (N0.getOpcode() == ISD::TRUNCATE && N0.hasOneUse() &&
1849 HasNoSignedComparisonUses(Node))
1850 // Look past the truncate if CMP is the only use of it.
1851 N0 = N0.getOperand(0);
1852 if (N0.getNode()->getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
1853 N0.getValueType() != MVT::i8 &&
1854 X86::isZeroNode(N1)) {
1855 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getNode()->getOperand(1));
1858 // For example, convert "testl %eax, $8" to "testb %al, $8"
1859 if ((C->getZExtValue() & ~UINT64_C(0xff)) == 0 &&
1860 (!(C->getZExtValue() & 0x80) ||
1861 HasNoSignedComparisonUses(Node))) {
1862 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), MVT::i8);
1863 SDValue Reg = N0.getNode()->getOperand(0);
1865 // On x86-32, only the ABCD registers have 8-bit subregisters.
1866 if (!Subtarget->is64Bit()) {
1867 TargetRegisterClass *TRC = 0;
1868 switch (N0.getValueType().getSimpleVT().SimpleTy) {
1869 case MVT::i32: TRC = &X86::GR32_ABCDRegClass; break;
1870 case MVT::i16: TRC = &X86::GR16_ABCDRegClass; break;
1871 default: llvm_unreachable("Unsupported TEST operand type!");
1873 SDValue RC = CurDAG->getTargetConstant(TRC->getID(), MVT::i32);
1874 Reg = SDValue(CurDAG->getMachineNode(X86::COPY_TO_REGCLASS, dl,
1875 Reg.getValueType(), Reg, RC), 0);
1878 // Extract the l-register.
1879 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl,
1883 return CurDAG->getMachineNode(X86::TEST8ri, dl, MVT::i32, Subreg, Imm);
1886 // For example, "testl %eax, $2048" to "testb %ah, $8".
1887 if ((C->getZExtValue() & ~UINT64_C(0xff00)) == 0 &&
1888 (!(C->getZExtValue() & 0x8000) ||
1889 HasNoSignedComparisonUses(Node))) {
1890 // Shift the immediate right by 8 bits.
1891 SDValue ShiftedImm = CurDAG->getTargetConstant(C->getZExtValue() >> 8,
1893 SDValue Reg = N0.getNode()->getOperand(0);
1895 // Put the value in an ABCD register.
1896 TargetRegisterClass *TRC = 0;
1897 switch (N0.getValueType().getSimpleVT().SimpleTy) {
1898 case MVT::i64: TRC = &X86::GR64_ABCDRegClass; break;
1899 case MVT::i32: TRC = &X86::GR32_ABCDRegClass; break;
1900 case MVT::i16: TRC = &X86::GR16_ABCDRegClass; break;
1901 default: llvm_unreachable("Unsupported TEST operand type!");
1903 SDValue RC = CurDAG->getTargetConstant(TRC->getID(), MVT::i32);
1904 Reg = SDValue(CurDAG->getMachineNode(X86::COPY_TO_REGCLASS, dl,
1905 Reg.getValueType(), Reg, RC), 0);
1907 // Extract the h-register.
1908 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_8bit_hi, dl,
1911 // Emit a testb. No special NOREX tricks are needed since there's
1912 // only one GPR operand!
1913 return CurDAG->getMachineNode(X86::TEST8ri, dl, MVT::i32,
1914 Subreg, ShiftedImm);
1917 // For example, "testl %eax, $32776" to "testw %ax, $32776".
1918 if ((C->getZExtValue() & ~UINT64_C(0xffff)) == 0 &&
1919 N0.getValueType() != MVT::i16 &&
1920 (!(C->getZExtValue() & 0x8000) ||
1921 HasNoSignedComparisonUses(Node))) {
1922 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), MVT::i16);
1923 SDValue Reg = N0.getNode()->getOperand(0);
1925 // Extract the 16-bit subregister.
1926 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_16bit, dl,
1930 return CurDAG->getMachineNode(X86::TEST16ri, dl, MVT::i32, Subreg, Imm);
1933 // For example, "testq %rax, $268468232" to "testl %eax, $268468232".
1934 if ((C->getZExtValue() & ~UINT64_C(0xffffffff)) == 0 &&
1935 N0.getValueType() == MVT::i64 &&
1936 (!(C->getZExtValue() & 0x80000000) ||
1937 HasNoSignedComparisonUses(Node))) {
1938 SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), MVT::i32);
1939 SDValue Reg = N0.getNode()->getOperand(0);
1941 // Extract the 32-bit subregister.
1942 SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_32bit, dl,
1946 return CurDAG->getMachineNode(X86::TEST32ri, dl, MVT::i32, Subreg, Imm);
1953 SDNode *ResNode = SelectCode(Node);
1955 DEBUG(dbgs() << "=> ";
1956 if (ResNode == NULL || ResNode == Node)
1959 ResNode->dump(CurDAG);
1965 bool X86DAGToDAGISel::
1966 SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
1967 std::vector<SDValue> &OutOps) {
1968 SDValue Op0, Op1, Op2, Op3, Op4;
1969 switch (ConstraintCode) {
1970 case 'o': // offsetable ??
1971 case 'v': // not offsetable ??
1972 default: return true;
1974 if (!SelectAddr(Op.getNode(), Op, Op0, Op1, Op2, Op3, Op4))
1979 OutOps.push_back(Op0);
1980 OutOps.push_back(Op1);
1981 OutOps.push_back(Op2);
1982 OutOps.push_back(Op3);
1983 OutOps.push_back(Op4);
1987 /// createX86ISelDag - This pass converts a legalized DAG into a
1988 /// X86-specific DAG, ready for instruction scheduling.
1990 FunctionPass *llvm::createX86ISelDag(X86TargetMachine &TM,
1991 llvm::CodeGenOpt::Level OptLevel) {
1992 return new X86DAGToDAGISel(TM, OptLevel);