ARMISelLowering.cpp revision 69de1932b350d7cdfc0ed1f4198d6f78c7822a02
1//===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This file defines the interfaces that ARM uses to lower LLVM code into a
11// selection DAG.
12//
13//===----------------------------------------------------------------------===//
14
15#include "ARM.h"
16#include "ARMAddressingModes.h"
17#include "ARMConstantPoolValue.h"
18#include "ARMISelLowering.h"
19#include "ARMMachineFunctionInfo.h"
20#include "ARMRegisterInfo.h"
21#include "ARMSubtarget.h"
22#include "ARMTargetMachine.h"
23#include "llvm/CallingConv.h"
24#include "llvm/Constants.h"
25#include "llvm/Instruction.h"
26#include "llvm/Intrinsics.h"
27#include "llvm/GlobalValue.h"
28#include "llvm/CodeGen/MachineBasicBlock.h"
29#include "llvm/CodeGen/MachineFrameInfo.h"
30#include "llvm/CodeGen/MachineFunction.h"
31#include "llvm/CodeGen/MachineInstrBuilder.h"
32#include "llvm/CodeGen/MachineRegisterInfo.h"
33#include "llvm/CodeGen/SelectionDAG.h"
34#include "llvm/Target/TargetOptions.h"
35#include "llvm/ADT/VectorExtras.h"
36#include "llvm/Support/MathExtras.h"
37using namespace llvm;
38
39ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
40    : TargetLowering(TM), ARMPCLabelIndex(0) {
41  Subtarget = &TM.getSubtarget<ARMSubtarget>();
42
43  if (Subtarget->isTargetDarwin()) {
44    // Don't have these.
45    setLibcallName(RTLIB::UINTTOFP_I64_F32, NULL);
46    setLibcallName(RTLIB::UINTTOFP_I64_F64, NULL);
47
48    // Uses VFP for Thumb libfuncs if available.
49    if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
50      // Single-precision floating-point arithmetic.
51      setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
52      setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
53      setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
54      setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
55
56      // Double-precision floating-point arithmetic.
57      setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
58      setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
59      setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
60      setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
61
62      // Single-precision comparisons.
63      setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
64      setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
65      setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
66      setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
67      setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
68      setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
69      setLibcallName(RTLIB::UO_F32,  "__unordsf2vfp");
70      setLibcallName(RTLIB::O_F32,   "__unordsf2vfp");
71
72      setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
73      setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
74      setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
75      setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
76      setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
77      setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
78      setCmpLibcallCC(RTLIB::UO_F32,  ISD::SETNE);
79      setCmpLibcallCC(RTLIB::O_F32,   ISD::SETEQ);
80
81      // Double-precision comparisons.
82      setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
83      setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
84      setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
85      setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
86      setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
87      setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
88      setLibcallName(RTLIB::UO_F64,  "__unorddf2vfp");
89      setLibcallName(RTLIB::O_F64,   "__unorddf2vfp");
90
91      setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
92      setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
93      setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
94      setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
95      setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
96      setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
97      setCmpLibcallCC(RTLIB::UO_F64,  ISD::SETNE);
98      setCmpLibcallCC(RTLIB::O_F64,   ISD::SETEQ);
99
100      // Floating-point to integer conversions.
101      // i64 conversions are done via library routines even when generating VFP
102      // instructions, so use the same ones.
103      setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
104      setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
105      setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
106      setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
107
108      // Conversions between floating types.
109      setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
110      setLibcallName(RTLIB::FPEXT_F32_F64,   "__extendsfdf2vfp");
111
112      // Integer to floating-point conversions.
113      // i64 conversions are done via library routines even when generating VFP
114      // instructions, so use the same ones.
115      // FIXME: There appears to be some naming inconsistency in ARM libgcc: e.g.
116      // __floatunsidf vs. __floatunssidfvfp.
117      setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
118      setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
119      setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
120      setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
121    }
122  }
123
124  addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
125  if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb()) {
126    addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
127    addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
128
129    setTruncStoreAction(MVT::f64, MVT::f32, Expand);
130  }
131  computeRegisterProperties();
132
133  // ARM does not have f32 extending load.
134  setLoadXAction(ISD::EXTLOAD, MVT::f32, Expand);
135
136  // ARM does not have i1 sign extending load.
137  setLoadXAction(ISD::SEXTLOAD, MVT::i1, Promote);
138
139  // ARM supports all 4 flavors of integer indexed load / store.
140  for (unsigned im = (unsigned)ISD::PRE_INC;
141       im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
142    setIndexedLoadAction(im,  MVT::i1,  Legal);
143    setIndexedLoadAction(im,  MVT::i8,  Legal);
144    setIndexedLoadAction(im,  MVT::i16, Legal);
145    setIndexedLoadAction(im,  MVT::i32, Legal);
146    setIndexedStoreAction(im, MVT::i1,  Legal);
147    setIndexedStoreAction(im, MVT::i8,  Legal);
148    setIndexedStoreAction(im, MVT::i16, Legal);
149    setIndexedStoreAction(im, MVT::i32, Legal);
150  }
151
152  // i64 operation support.
153  if (Subtarget->isThumb()) {
154    setOperationAction(ISD::MUL,     MVT::i64, Expand);
155    setOperationAction(ISD::MULHU,   MVT::i32, Expand);
156    setOperationAction(ISD::MULHS,   MVT::i32, Expand);
157    setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
158    setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
159  } else {
160    setOperationAction(ISD::MUL,     MVT::i64, Expand);
161    setOperationAction(ISD::MULHU,   MVT::i32, Expand);
162    if (!Subtarget->hasV6Ops())
163      setOperationAction(ISD::MULHS, MVT::i32, Expand);
164  }
165  setOperationAction(ISD::SHL_PARTS, MVT::i32, Expand);
166  setOperationAction(ISD::SRA_PARTS, MVT::i32, Expand);
167  setOperationAction(ISD::SRL_PARTS, MVT::i32, Expand);
168  setOperationAction(ISD::SRL,       MVT::i64, Custom);
169  setOperationAction(ISD::SRA,       MVT::i64, Custom);
170
171  // ARM does not have ROTL.
172  setOperationAction(ISD::ROTL,  MVT::i32, Expand);
173  setOperationAction(ISD::CTTZ , MVT::i32, Expand);
174  setOperationAction(ISD::CTPOP, MVT::i32, Expand);
175  if (!Subtarget->hasV5TOps() || Subtarget->isThumb())
176    setOperationAction(ISD::CTLZ, MVT::i32, Expand);
177
178  // Only ARMv6 has BSWAP.
179  if (!Subtarget->hasV6Ops())
180    setOperationAction(ISD::BSWAP, MVT::i32, Expand);
181
182  // These are expanded into libcalls.
183  setOperationAction(ISD::SDIV,  MVT::i32, Expand);
184  setOperationAction(ISD::UDIV,  MVT::i32, Expand);
185  setOperationAction(ISD::SREM,  MVT::i32, Expand);
186  setOperationAction(ISD::UREM,  MVT::i32, Expand);
187  setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
188  setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
189
190  // Support label based line numbers.
191  setOperationAction(ISD::LOCATION, MVT::Other, Expand);
192  setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
193
194  setOperationAction(ISD::RET,           MVT::Other, Custom);
195  setOperationAction(ISD::GlobalAddress, MVT::i32,   Custom);
196  setOperationAction(ISD::ConstantPool,  MVT::i32,   Custom);
197  setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
198  setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
199
200  // Expand mem operations genericly.
201  setOperationAction(ISD::MEMSET          , MVT::Other, Expand);
202  setOperationAction(ISD::MEMCPY          , MVT::Other, Custom);
203  setOperationAction(ISD::MEMMOVE         , MVT::Other, Expand);
204
205  // Use the default implementation.
206  setOperationAction(ISD::VASTART           , MVT::Other, Custom);
207  setOperationAction(ISD::VAARG             , MVT::Other, Expand);
208  setOperationAction(ISD::VACOPY            , MVT::Other, Expand);
209  setOperationAction(ISD::VAEND             , MVT::Other, Expand);
210  setOperationAction(ISD::STACKSAVE,          MVT::Other, Expand);
211  setOperationAction(ISD::STACKRESTORE,       MVT::Other, Expand);
212  setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32  , Expand);
213
214  if (!Subtarget->hasV6Ops()) {
215    setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
216    setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8,  Expand);
217  }
218  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
219
220  if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb())
221    // Turn f64->i64 into FMRRD iff target supports vfp2.
222    setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
223
224  // We want to custom lower some of our intrinsics.
225  setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
226
227  setOperationAction(ISD::SETCC    , MVT::i32, Expand);
228  setOperationAction(ISD::SETCC    , MVT::f32, Expand);
229  setOperationAction(ISD::SETCC    , MVT::f64, Expand);
230  setOperationAction(ISD::SELECT   , MVT::i32, Expand);
231  setOperationAction(ISD::SELECT   , MVT::f32, Expand);
232  setOperationAction(ISD::SELECT   , MVT::f64, Expand);
233  setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
234  setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
235  setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
236
237  setOperationAction(ISD::BRCOND   , MVT::Other, Expand);
238  setOperationAction(ISD::BR_CC    , MVT::i32,   Custom);
239  setOperationAction(ISD::BR_CC    , MVT::f32,   Custom);
240  setOperationAction(ISD::BR_CC    , MVT::f64,   Custom);
241  setOperationAction(ISD::BR_JT    , MVT::Other, Custom);
242
243  // FP Constants can't be immediates.
244  setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
245  setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
246
247  // We don't support sin/cos/fmod/copysign/pow
248  setOperationAction(ISD::FSIN     , MVT::f64, Expand);
249  setOperationAction(ISD::FSIN     , MVT::f32, Expand);
250  setOperationAction(ISD::FCOS     , MVT::f32, Expand);
251  setOperationAction(ISD::FCOS     , MVT::f64, Expand);
252  setOperationAction(ISD::FREM     , MVT::f64, Expand);
253  setOperationAction(ISD::FREM     , MVT::f32, Expand);
254  setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
255  setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
256  setOperationAction(ISD::FPOW     , MVT::f64, Expand);
257  setOperationAction(ISD::FPOW     , MVT::f32, Expand);
258
259  // int <-> fp are custom expanded into bit_convert + ARMISD ops.
260  setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
261  setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
262  setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
263  setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
264
265  // We have target-specific dag combine patterns for the following nodes:
266  // ARMISD::FMRRD  - No need to call setTargetDAGCombine
267
268  setStackPointerRegisterToSaveRestore(ARM::SP);
269  setSchedulingPreference(SchedulingForRegPressure);
270  setIfCvtBlockSizeLimit(Subtarget->isThumb() ? 0 : 10);
271  setIfCvtDupBlockSizeLimit(Subtarget->isThumb() ? 0 : 2);
272
273  maxStoresPerMemcpy = 1;   //// temporary - rewrite interface to use type
274}
275
276
277const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
278  switch (Opcode) {
279  default: return 0;
280  case ARMISD::Wrapper:       return "ARMISD::Wrapper";
281  case ARMISD::WrapperJT:     return "ARMISD::WrapperJT";
282  case ARMISD::CALL:          return "ARMISD::CALL";
283  case ARMISD::CALL_PRED:     return "ARMISD::CALL_PRED";
284  case ARMISD::CALL_NOLINK:   return "ARMISD::CALL_NOLINK";
285  case ARMISD::tCALL:         return "ARMISD::tCALL";
286  case ARMISD::BRCOND:        return "ARMISD::BRCOND";
287  case ARMISD::BR_JT:         return "ARMISD::BR_JT";
288  case ARMISD::RET_FLAG:      return "ARMISD::RET_FLAG";
289  case ARMISD::PIC_ADD:       return "ARMISD::PIC_ADD";
290  case ARMISD::CMP:           return "ARMISD::CMP";
291  case ARMISD::CMPNZ:         return "ARMISD::CMPNZ";
292  case ARMISD::CMPFP:         return "ARMISD::CMPFP";
293  case ARMISD::CMPFPw0:       return "ARMISD::CMPFPw0";
294  case ARMISD::FMSTAT:        return "ARMISD::FMSTAT";
295  case ARMISD::CMOV:          return "ARMISD::CMOV";
296  case ARMISD::CNEG:          return "ARMISD::CNEG";
297
298  case ARMISD::FTOSI:         return "ARMISD::FTOSI";
299  case ARMISD::FTOUI:         return "ARMISD::FTOUI";
300  case ARMISD::SITOF:         return "ARMISD::SITOF";
301  case ARMISD::UITOF:         return "ARMISD::UITOF";
302
303  case ARMISD::SRL_FLAG:      return "ARMISD::SRL_FLAG";
304  case ARMISD::SRA_FLAG:      return "ARMISD::SRA_FLAG";
305  case ARMISD::RRX:           return "ARMISD::RRX";
306
307  case ARMISD::FMRRD:         return "ARMISD::FMRRD";
308  case ARMISD::FMDRR:         return "ARMISD::FMDRR";
309
310  case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
311  }
312}
313
314//===----------------------------------------------------------------------===//
315// Lowering Code
316//===----------------------------------------------------------------------===//
317
318
319/// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
320static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
321  switch (CC) {
322  default: assert(0 && "Unknown condition code!");
323  case ISD::SETNE:  return ARMCC::NE;
324  case ISD::SETEQ:  return ARMCC::EQ;
325  case ISD::SETGT:  return ARMCC::GT;
326  case ISD::SETGE:  return ARMCC::GE;
327  case ISD::SETLT:  return ARMCC::LT;
328  case ISD::SETLE:  return ARMCC::LE;
329  case ISD::SETUGT: return ARMCC::HI;
330  case ISD::SETUGE: return ARMCC::HS;
331  case ISD::SETULT: return ARMCC::LO;
332  case ISD::SETULE: return ARMCC::LS;
333  }
334}
335
336/// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC. It
337/// returns true if the operands should be inverted to form the proper
338/// comparison.
339static bool FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
340                        ARMCC::CondCodes &CondCode2) {
341  bool Invert = false;
342  CondCode2 = ARMCC::AL;
343  switch (CC) {
344  default: assert(0 && "Unknown FP condition!");
345  case ISD::SETEQ:
346  case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
347  case ISD::SETGT:
348  case ISD::SETOGT: CondCode = ARMCC::GT; break;
349  case ISD::SETGE:
350  case ISD::SETOGE: CondCode = ARMCC::GE; break;
351  case ISD::SETOLT: CondCode = ARMCC::MI; break;
352  case ISD::SETOLE: CondCode = ARMCC::GT; Invert = true; break;
353  case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
354  case ISD::SETO:   CondCode = ARMCC::VC; break;
355  case ISD::SETUO:  CondCode = ARMCC::VS; break;
356  case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
357  case ISD::SETUGT: CondCode = ARMCC::HI; break;
358  case ISD::SETUGE: CondCode = ARMCC::PL; break;
359  case ISD::SETLT:
360  case ISD::SETULT: CondCode = ARMCC::LT; break;
361  case ISD::SETLE:
362  case ISD::SETULE: CondCode = ARMCC::LE; break;
363  case ISD::SETNE:
364  case ISD::SETUNE: CondCode = ARMCC::NE; break;
365  }
366  return Invert;
367}
368
369static void
370HowToPassArgument(MVT::ValueType ObjectVT, unsigned NumGPRs,
371                  unsigned StackOffset, unsigned &NeededGPRs,
372                  unsigned &NeededStackSize, unsigned &GPRPad,
373                  unsigned &StackPad, unsigned Flags) {
374  NeededStackSize = 0;
375  NeededGPRs = 0;
376  StackPad = 0;
377  GPRPad = 0;
378  unsigned align = (Flags >> ISD::ParamFlags::OrigAlignmentOffs);
379  GPRPad = NumGPRs % ((align + 3)/4);
380  StackPad = StackOffset % align;
381  unsigned firstGPR = NumGPRs + GPRPad;
382  switch (ObjectVT) {
383  default: assert(0 && "Unhandled argument type!");
384  case MVT::i32:
385  case MVT::f32:
386    if (firstGPR < 4)
387      NeededGPRs = 1;
388    else
389      NeededStackSize = 4;
390    break;
391  case MVT::i64:
392  case MVT::f64:
393    if (firstGPR < 3)
394      NeededGPRs = 2;
395    else if (firstGPR == 3) {
396      NeededGPRs = 1;
397      NeededStackSize = 4;
398    } else
399      NeededStackSize = 8;
400  }
401}
402
403/// LowerCALL - Lowering a ISD::CALL node into a callseq_start <-
404/// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
405/// nodes.
406SDOperand ARMTargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
407  MVT::ValueType RetVT= Op.Val->getValueType(0);
408  SDOperand Chain    = Op.getOperand(0);
409  unsigned CallConv  = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
410  assert((CallConv == CallingConv::C ||
411          CallConv == CallingConv::Fast) && "unknown calling convention");
412  SDOperand Callee   = Op.getOperand(4);
413  unsigned NumOps    = (Op.getNumOperands() - 5) / 2;
414  unsigned ArgOffset = 0;   // Frame mechanisms handle retaddr slot
415  unsigned NumGPRs = 0;     // GPRs used for parameter passing.
416
417  // Count how many bytes are to be pushed on the stack.
418  unsigned NumBytes = 0;
419
420  // Add up all the space actually used.
421  for (unsigned i = 0; i < NumOps; ++i) {
422    unsigned ObjSize;
423    unsigned ObjGPRs;
424    unsigned StackPad;
425    unsigned GPRPad;
426    MVT::ValueType ObjectVT = Op.getOperand(5+2*i).getValueType();
427    unsigned Flags = Op.getConstantOperandVal(5+2*i+1);
428    HowToPassArgument(ObjectVT, NumGPRs, NumBytes, ObjGPRs, ObjSize,
429                      GPRPad, StackPad, Flags);
430    NumBytes += ObjSize + StackPad;
431    NumGPRs += ObjGPRs + GPRPad;
432  }
433
434  // Adjust the stack pointer for the new arguments...
435  // These operations are automatically eliminated by the prolog/epilog pass
436  Chain = DAG.getCALLSEQ_START(Chain,
437                               DAG.getConstant(NumBytes, MVT::i32));
438
439  SDOperand StackPtr = DAG.getRegister(ARM::SP, MVT::i32);
440
441  static const unsigned GPRArgRegs[] = {
442    ARM::R0, ARM::R1, ARM::R2, ARM::R3
443  };
444
445  NumGPRs = 0;
446  std::vector<std::pair<unsigned, SDOperand> > RegsToPass;
447  std::vector<SDOperand> MemOpChains;
448  for (unsigned i = 0; i != NumOps; ++i) {
449    SDOperand Arg = Op.getOperand(5+2*i);
450    unsigned Flags = Op.getConstantOperandVal(5+2*i+1);
451    MVT::ValueType ArgVT = Arg.getValueType();
452
453    unsigned ObjSize;
454    unsigned ObjGPRs;
455    unsigned GPRPad;
456    unsigned StackPad;
457    HowToPassArgument(ArgVT, NumGPRs, ArgOffset, ObjGPRs,
458                      ObjSize, GPRPad, StackPad, Flags);
459    NumGPRs += GPRPad;
460    ArgOffset += StackPad;
461    if (ObjGPRs > 0) {
462      switch (ArgVT) {
463      default: assert(0 && "Unexpected ValueType for argument!");
464      case MVT::i32:
465        RegsToPass.push_back(std::make_pair(GPRArgRegs[NumGPRs], Arg));
466        break;
467      case MVT::f32:
468        RegsToPass.push_back(std::make_pair(GPRArgRegs[NumGPRs],
469                                 DAG.getNode(ISD::BIT_CONVERT, MVT::i32, Arg)));
470        break;
471      case MVT::i64: {
472        SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Arg,
473                                   DAG.getConstant(0, getPointerTy()));
474        SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Arg,
475                                   DAG.getConstant(1, getPointerTy()));
476        RegsToPass.push_back(std::make_pair(GPRArgRegs[NumGPRs], Lo));
477        if (ObjGPRs == 2)
478          RegsToPass.push_back(std::make_pair(GPRArgRegs[NumGPRs+1], Hi));
479        else {
480          SDOperand PtrOff= DAG.getConstant(ArgOffset, StackPtr.getValueType());
481          PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
482          MemOpChains.push_back(DAG.getStore(Chain, Hi, PtrOff, NULL, 0));
483        }
484        break;
485      }
486      case MVT::f64: {
487        SDOperand Cvt = DAG.getNode(ARMISD::FMRRD,
488                                    DAG.getVTList(MVT::i32, MVT::i32),
489                                    &Arg, 1);
490        RegsToPass.push_back(std::make_pair(GPRArgRegs[NumGPRs], Cvt));
491        if (ObjGPRs == 2)
492          RegsToPass.push_back(std::make_pair(GPRArgRegs[NumGPRs+1],
493                                              Cvt.getValue(1)));
494        else {
495          SDOperand PtrOff= DAG.getConstant(ArgOffset, StackPtr.getValueType());
496          PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
497          MemOpChains.push_back(DAG.getStore(Chain, Cvt.getValue(1), PtrOff,
498                                             NULL, 0));
499        }
500        break;
501      }
502      }
503    } else {
504      assert(ObjSize != 0);
505      SDOperand PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
506      PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
507      MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
508    }
509
510    NumGPRs += ObjGPRs;
511    ArgOffset += ObjSize;
512  }
513
514  if (!MemOpChains.empty())
515    Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
516                        &MemOpChains[0], MemOpChains.size());
517
518  // Build a sequence of copy-to-reg nodes chained together with token chain
519  // and flag operands which copy the outgoing args into the appropriate regs.
520  SDOperand InFlag;
521  for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
522    Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
523                             InFlag);
524    InFlag = Chain.getValue(1);
525  }
526
527  // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
528  // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
529  // node so that legalize doesn't hack it.
530  bool isDirect = false;
531  bool isARMFunc = false;
532  bool isLocalARMFunc = false;
533  if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
534    GlobalValue *GV = G->getGlobal();
535    isDirect = true;
536    bool isExt = (GV->isDeclaration() || GV->hasWeakLinkage() ||
537                  GV->hasLinkOnceLinkage());
538    bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
539                   getTargetMachine().getRelocationModel() != Reloc::Static;
540    isARMFunc = !Subtarget->isThumb() || isStub;
541    // ARM call to a local ARM function is predicable.
542    isLocalARMFunc = !Subtarget->isThumb() && !isExt;
543    // tBX takes a register source operand.
544    if (isARMFunc && Subtarget->isThumb() && !Subtarget->hasV5TOps()) {
545      ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMPCLabelIndex,
546                                                           ARMCP::CPStub, 4);
547      SDOperand CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 2);
548      CPAddr = DAG.getNode(ARMISD::Wrapper, MVT::i32, CPAddr);
549      Callee = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), CPAddr, NULL, 0);
550      SDOperand PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
551      Callee = DAG.getNode(ARMISD::PIC_ADD, getPointerTy(), Callee, PICLabel);
552   } else
553      Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
554  } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
555    isDirect = true;
556    bool isStub = Subtarget->isTargetDarwin() &&
557                  getTargetMachine().getRelocationModel() != Reloc::Static;
558    isARMFunc = !Subtarget->isThumb() || isStub;
559    // tBX takes a register source operand.
560    const char *Sym = S->getSymbol();
561    if (isARMFunc && Subtarget->isThumb() && !Subtarget->hasV5TOps()) {
562      ARMConstantPoolValue *CPV = new ARMConstantPoolValue(Sym, ARMPCLabelIndex,
563                                                           ARMCP::CPStub, 4);
564      SDOperand CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 2);
565      CPAddr = DAG.getNode(ARMISD::Wrapper, MVT::i32, CPAddr);
566      Callee = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), CPAddr, NULL, 0);
567      SDOperand PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
568      Callee = DAG.getNode(ARMISD::PIC_ADD, getPointerTy(), Callee, PICLabel);
569    } else
570      Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
571  }
572
573  // FIXME: handle tail calls differently.
574  unsigned CallOpc;
575  if (Subtarget->isThumb()) {
576    if (!Subtarget->hasV5TOps() && (!isDirect || isARMFunc))
577      CallOpc = ARMISD::CALL_NOLINK;
578    else
579      CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
580  } else {
581    CallOpc = (isDirect || Subtarget->hasV5TOps())
582      ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
583      : ARMISD::CALL_NOLINK;
584  }
585  if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb()) {
586    // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
587    Chain = DAG.getCopyToReg(Chain, ARM::LR,
588                             DAG.getNode(ISD::UNDEF, MVT::i32), InFlag);
589    InFlag = Chain.getValue(1);
590  }
591
592  std::vector<MVT::ValueType> NodeTys;
593  NodeTys.push_back(MVT::Other);   // Returns a chain
594  NodeTys.push_back(MVT::Flag);    // Returns a flag for retval copy to use.
595
596  std::vector<SDOperand> Ops;
597  Ops.push_back(Chain);
598  Ops.push_back(Callee);
599
600  // Add argument registers to the end of the list so that they are known live
601  // into the call.
602  for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
603    Ops.push_back(DAG.getRegister(RegsToPass[i].first,
604                                  RegsToPass[i].second.getValueType()));
605
606  if (InFlag.Val)
607    Ops.push_back(InFlag);
608  Chain = DAG.getNode(CallOpc, NodeTys, &Ops[0], Ops.size());
609  InFlag = Chain.getValue(1);
610
611  Chain = DAG.getCALLSEQ_END(Chain,
612                             DAG.getConstant(NumBytes, MVT::i32),
613                             DAG.getConstant(0, MVT::i32),
614                             InFlag);
615  if (RetVT != MVT::Other)
616    InFlag = Chain.getValue(1);
617
618  std::vector<SDOperand> ResultVals;
619  NodeTys.clear();
620
621  // If the call has results, copy the values out of the ret val registers.
622  switch (RetVT) {
623  default: assert(0 && "Unexpected ret value!");
624  case MVT::Other:
625    break;
626  case MVT::i32:
627    Chain = DAG.getCopyFromReg(Chain, ARM::R0, MVT::i32, InFlag).getValue(1);
628    ResultVals.push_back(Chain.getValue(0));
629    if (Op.Val->getValueType(1) == MVT::i32) {
630      // Returns a i64 value.
631      Chain = DAG.getCopyFromReg(Chain, ARM::R1, MVT::i32,
632                                 Chain.getValue(2)).getValue(1);
633      ResultVals.push_back(Chain.getValue(0));
634      NodeTys.push_back(MVT::i32);
635    }
636    NodeTys.push_back(MVT::i32);
637    break;
638  case MVT::f32:
639    Chain = DAG.getCopyFromReg(Chain, ARM::R0, MVT::i32, InFlag).getValue(1);
640    ResultVals.push_back(DAG.getNode(ISD::BIT_CONVERT, MVT::f32,
641                                     Chain.getValue(0)));
642    NodeTys.push_back(MVT::f32);
643    break;
644  case MVT::f64: {
645    SDOperand Lo = DAG.getCopyFromReg(Chain, ARM::R0, MVT::i32, InFlag);
646    SDOperand Hi = DAG.getCopyFromReg(Lo, ARM::R1, MVT::i32, Lo.getValue(2));
647    ResultVals.push_back(DAG.getNode(ARMISD::FMDRR, MVT::f64, Lo, Hi));
648    NodeTys.push_back(MVT::f64);
649    break;
650  }
651  }
652
653  NodeTys.push_back(MVT::Other);
654
655  if (ResultVals.empty())
656    return Chain;
657
658  ResultVals.push_back(Chain);
659  SDOperand Res = DAG.getNode(ISD::MERGE_VALUES, NodeTys, &ResultVals[0],
660                              ResultVals.size());
661  return Res.getValue(Op.ResNo);
662}
663
664static SDOperand LowerRET(SDOperand Op, SelectionDAG &DAG) {
665  SDOperand Copy;
666  SDOperand Chain = Op.getOperand(0);
667  switch(Op.getNumOperands()) {
668  default:
669    assert(0 && "Do not know how to return this many arguments!");
670    abort();
671  case 1: {
672    SDOperand LR = DAG.getRegister(ARM::LR, MVT::i32);
673    return DAG.getNode(ARMISD::RET_FLAG, MVT::Other, Chain);
674  }
675  case 3:
676    Op = Op.getOperand(1);
677    if (Op.getValueType() == MVT::f32) {
678      Op = DAG.getNode(ISD::BIT_CONVERT, MVT::i32, Op);
679    } else if (Op.getValueType() == MVT::f64) {
680      // Legalize ret f64 -> ret 2 x i32.  We always have fmrrd if f64 is
681      // available.
682      Op = DAG.getNode(ARMISD::FMRRD, DAG.getVTList(MVT::i32, MVT::i32), &Op,1);
683      SDOperand Sign = DAG.getConstant(0, MVT::i32);
684      return DAG.getNode(ISD::RET, MVT::Other, Chain, Op, Sign,
685                         Op.getValue(1), Sign);
686    }
687    Copy = DAG.getCopyToReg(Chain, ARM::R0, Op, SDOperand());
688    if (DAG.getMachineFunction().getRegInfo().liveout_empty())
689      DAG.getMachineFunction().getRegInfo().addLiveOut(ARM::R0);
690    break;
691  case 5:
692    Copy = DAG.getCopyToReg(Chain, ARM::R1, Op.getOperand(3), SDOperand());
693    Copy = DAG.getCopyToReg(Copy, ARM::R0, Op.getOperand(1), Copy.getValue(1));
694    // If we haven't noted the R0+R1 are live out, do so now.
695    if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
696      DAG.getMachineFunction().getRegInfo().addLiveOut(ARM::R0);
697      DAG.getMachineFunction().getRegInfo().addLiveOut(ARM::R1);
698    }
699    break;
700  }
701
702  //We must use RET_FLAG instead of BRIND because BRIND doesn't have a flag
703  return DAG.getNode(ARMISD::RET_FLAG, MVT::Other, Copy, Copy.getValue(1));
704}
705
706// ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
707// their target countpart wrapped in the ARMISD::Wrapper node. Suppose N is
708// one of the above mentioned nodes. It has to be wrapped because otherwise
709// Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
710// be used to form addressing mode. These wrapped nodes will be selected
711// into MOVi.
712static SDOperand LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
713  MVT::ValueType PtrVT = Op.getValueType();
714  ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
715  SDOperand Res;
716  if (CP->isMachineConstantPoolEntry())
717    Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
718                                    CP->getAlignment());
719  else
720    Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
721                                    CP->getAlignment());
722  return DAG.getNode(ARMISD::Wrapper, MVT::i32, Res);
723}
724
725// Lower ISD::GlobalTLSAddress using the "general dynamic" model
726SDOperand
727ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
728                                                 SelectionDAG &DAG) {
729  MVT::ValueType PtrVT = getPointerTy();
730  unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
731  ARMConstantPoolValue *CPV =
732    new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex, ARMCP::CPValue,
733                             PCAdj, "tlsgd", true);
734  SDOperand Argument = DAG.getTargetConstantPool(CPV, PtrVT, 2);
735  Argument = DAG.getNode(ARMISD::Wrapper, MVT::i32, Argument);
736  Argument = DAG.getLoad(PtrVT, DAG.getEntryNode(), Argument, NULL, 0);
737  SDOperand Chain = Argument.getValue(1);
738
739  SDOperand PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
740  Argument = DAG.getNode(ARMISD::PIC_ADD, PtrVT, Argument, PICLabel);
741
742  // call __tls_get_addr.
743  ArgListTy Args;
744  ArgListEntry Entry;
745  Entry.Node = Argument;
746  Entry.Ty = (const Type *) Type::Int32Ty;
747  Args.push_back(Entry);
748  std::pair<SDOperand, SDOperand> CallResult =
749    LowerCallTo(Chain, (const Type *) Type::Int32Ty, false, false,
750                CallingConv::C, false,
751                DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG);
752  return CallResult.first;
753}
754
755// Lower ISD::GlobalTLSAddress using the "initial exec" or
756// "local exec" model.
757SDOperand
758ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
759                                            SelectionDAG &DAG) {
760  GlobalValue *GV = GA->getGlobal();
761  SDOperand Offset;
762  SDOperand Chain = DAG.getEntryNode();
763  MVT::ValueType PtrVT = getPointerTy();
764  // Get the Thread Pointer
765  SDOperand ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, PtrVT);
766
767  if (GV->isDeclaration()){
768    // initial exec model
769    unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
770    ARMConstantPoolValue *CPV =
771      new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex, ARMCP::CPValue,
772                               PCAdj, "gottpoff", true);
773    Offset = DAG.getTargetConstantPool(CPV, PtrVT, 2);
774    Offset = DAG.getNode(ARMISD::Wrapper, MVT::i32, Offset);
775    Offset = DAG.getLoad(PtrVT, Chain, Offset, NULL, 0);
776    Chain = Offset.getValue(1);
777
778    SDOperand PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
779    Offset = DAG.getNode(ARMISD::PIC_ADD, PtrVT, Offset, PICLabel);
780
781    Offset = DAG.getLoad(PtrVT, Chain, Offset, NULL, 0);
782  } else {
783    // local exec model
784    ARMConstantPoolValue *CPV =
785      new ARMConstantPoolValue(GV, ARMCP::CPValue, "tpoff");
786    Offset = DAG.getTargetConstantPool(CPV, PtrVT, 2);
787    Offset = DAG.getNode(ARMISD::Wrapper, MVT::i32, Offset);
788    Offset = DAG.getLoad(PtrVT, Chain, Offset, NULL, 0);
789  }
790
791  // The address of the thread local variable is the add of the thread
792  // pointer with the offset of the variable.
793  return DAG.getNode(ISD::ADD, PtrVT, ThreadPointer, Offset);
794}
795
796SDOperand
797ARMTargetLowering::LowerGlobalTLSAddress(SDOperand Op, SelectionDAG &DAG) {
798  // TODO: implement the "local dynamic" model
799  assert(Subtarget->isTargetELF() &&
800         "TLS not implemented for non-ELF targets");
801  GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
802  // If the relocation model is PIC, use the "General Dynamic" TLS Model,
803  // otherwise use the "Local Exec" TLS Model
804  if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
805    return LowerToTLSGeneralDynamicModel(GA, DAG);
806  else
807    return LowerToTLSExecModels(GA, DAG);
808}
809
810SDOperand ARMTargetLowering::LowerGlobalAddressELF(SDOperand Op,
811                                                   SelectionDAG &DAG) {
812  MVT::ValueType PtrVT = getPointerTy();
813  GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
814  Reloc::Model RelocM = getTargetMachine().getRelocationModel();
815  if (RelocM == Reloc::PIC_) {
816    bool UseGOTOFF = GV->hasInternalLinkage() || GV->hasHiddenVisibility();
817    ARMConstantPoolValue *CPV =
818      new ARMConstantPoolValue(GV, ARMCP::CPValue, UseGOTOFF ? "GOTOFF":"GOT");
819    SDOperand CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 2);
820    CPAddr = DAG.getNode(ARMISD::Wrapper, MVT::i32, CPAddr);
821    SDOperand Result = DAG.getLoad(PtrVT, DAG.getEntryNode(), CPAddr, NULL, 0);
822    SDOperand Chain = Result.getValue(1);
823    SDOperand GOT = DAG.getNode(ISD::GLOBAL_OFFSET_TABLE, PtrVT);
824    Result = DAG.getNode(ISD::ADD, PtrVT, Result, GOT);
825    if (!UseGOTOFF)
826      Result = DAG.getLoad(PtrVT, Chain, Result, NULL, 0);
827    return Result;
828  } else {
829    SDOperand CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 2);
830    CPAddr = DAG.getNode(ARMISD::Wrapper, MVT::i32, CPAddr);
831    return DAG.getLoad(PtrVT, DAG.getEntryNode(), CPAddr, NULL, 0);
832  }
833}
834
835/// GVIsIndirectSymbol - true if the GV will be accessed via an indirect symbol
836/// even in non-static mode.
837static bool GVIsIndirectSymbol(GlobalValue *GV, Reloc::Model RelocM) {
838  return RelocM != Reloc::Static &&
839    (GV->hasWeakLinkage() || GV->hasLinkOnceLinkage() ||
840     (GV->isDeclaration() && !GV->hasNotBeenReadFromBitcode()));
841}
842
843SDOperand ARMTargetLowering::LowerGlobalAddressDarwin(SDOperand Op,
844                                                      SelectionDAG &DAG) {
845  MVT::ValueType PtrVT = getPointerTy();
846  GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
847  Reloc::Model RelocM = getTargetMachine().getRelocationModel();
848  bool IsIndirect = GVIsIndirectSymbol(GV, RelocM);
849  SDOperand CPAddr;
850  if (RelocM == Reloc::Static)
851    CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 2);
852  else {
853    unsigned PCAdj = (RelocM != Reloc::PIC_)
854      ? 0 : (Subtarget->isThumb() ? 4 : 8);
855    ARMCP::ARMCPKind Kind = IsIndirect ? ARMCP::CPNonLazyPtr
856      : ARMCP::CPValue;
857    ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMPCLabelIndex,
858                                                         Kind, PCAdj);
859    CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 2);
860  }
861  CPAddr = DAG.getNode(ARMISD::Wrapper, MVT::i32, CPAddr);
862
863  SDOperand Result = DAG.getLoad(PtrVT, DAG.getEntryNode(), CPAddr, NULL, 0);
864  SDOperand Chain = Result.getValue(1);
865
866  if (RelocM == Reloc::PIC_) {
867    SDOperand PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
868    Result = DAG.getNode(ARMISD::PIC_ADD, PtrVT, Result, PICLabel);
869  }
870  if (IsIndirect)
871    Result = DAG.getLoad(PtrVT, Chain, Result, NULL, 0);
872
873  return Result;
874}
875
876SDOperand ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDOperand Op,
877                                                      SelectionDAG &DAG){
878  assert(Subtarget->isTargetELF() &&
879         "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
880  MVT::ValueType PtrVT = getPointerTy();
881  unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
882  ARMConstantPoolValue *CPV = new ARMConstantPoolValue("_GLOBAL_OFFSET_TABLE_",
883                                                       ARMPCLabelIndex,
884                                                       ARMCP::CPValue, PCAdj);
885  SDOperand CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 2);
886  CPAddr = DAG.getNode(ARMISD::Wrapper, MVT::i32, CPAddr);
887  SDOperand Result = DAG.getLoad(PtrVT, DAG.getEntryNode(), CPAddr, NULL, 0);
888  SDOperand PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
889  return DAG.getNode(ARMISD::PIC_ADD, PtrVT, Result, PICLabel);
890}
891
892static SDOperand LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
893  MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
894  unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
895  switch (IntNo) {
896  default: return SDOperand();    // Don't custom lower most intrinsics.
897  case Intrinsic::arm_thread_pointer:
898      return DAG.getNode(ARMISD::THREAD_POINTER, PtrVT);
899  }
900}
901
902static SDOperand LowerVASTART(SDOperand Op, SelectionDAG &DAG,
903                              unsigned VarArgsFrameIndex) {
904  // vastart just stores the address of the VarArgsFrameIndex slot into the
905  // memory location argument.
906  MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
907  SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
908  const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
909  return DAG.getStore(Op.getOperand(0), FR, Op.getOperand(1), SV, 0);
910}
911
912static SDOperand LowerFORMAL_ARGUMENT(SDOperand Op, SelectionDAG &DAG,
913                                      unsigned *vRegs, unsigned ArgNo,
914                                      unsigned &NumGPRs, unsigned &ArgOffset) {
915  MachineFunction &MF = DAG.getMachineFunction();
916  MVT::ValueType ObjectVT = Op.getValue(ArgNo).getValueType();
917  SDOperand Root = Op.getOperand(0);
918  std::vector<SDOperand> ArgValues;
919  MachineRegisterInfo &RegInfo = MF.getRegInfo();
920
921  static const unsigned GPRArgRegs[] = {
922    ARM::R0, ARM::R1, ARM::R2, ARM::R3
923  };
924
925  unsigned ObjSize;
926  unsigned ObjGPRs;
927  unsigned GPRPad;
928  unsigned StackPad;
929  unsigned Flags = Op.getConstantOperandVal(ArgNo + 3);
930  HowToPassArgument(ObjectVT, NumGPRs, ArgOffset, ObjGPRs,
931                    ObjSize, GPRPad, StackPad, Flags);
932  NumGPRs += GPRPad;
933  ArgOffset += StackPad;
934
935  SDOperand ArgValue;
936  if (ObjGPRs == 1) {
937    unsigned VReg = RegInfo.createVirtualRegister(&ARM::GPRRegClass);
938    RegInfo.addLiveIn(GPRArgRegs[NumGPRs], VReg);
939    vRegs[NumGPRs] = VReg;
940    ArgValue = DAG.getCopyFromReg(Root, VReg, MVT::i32);
941    if (ObjectVT == MVT::f32)
942      ArgValue = DAG.getNode(ISD::BIT_CONVERT, MVT::f32, ArgValue);
943  } else if (ObjGPRs == 2) {
944    unsigned VReg = RegInfo.createVirtualRegister(&ARM::GPRRegClass);
945    RegInfo.addLiveIn(GPRArgRegs[NumGPRs], VReg);
946    vRegs[NumGPRs] = VReg;
947    ArgValue = DAG.getCopyFromReg(Root, VReg, MVT::i32);
948
949    VReg = RegInfo.createVirtualRegister(&ARM::GPRRegClass);
950    RegInfo.addLiveIn(GPRArgRegs[NumGPRs+1], VReg);
951    vRegs[NumGPRs+1] = VReg;
952    SDOperand ArgValue2 = DAG.getCopyFromReg(Root, VReg, MVT::i32);
953
954    assert(ObjectVT != MVT::i64 && "i64 should already be lowered");
955    ArgValue = DAG.getNode(ARMISD::FMDRR, MVT::f64, ArgValue, ArgValue2);
956  }
957  NumGPRs += ObjGPRs;
958
959  if (ObjSize) {
960    // If the argument is actually used, emit a load from the right stack
961    // slot.
962    if (!Op.Val->hasNUsesOfValue(0, ArgNo)) {
963      MachineFrameInfo *MFI = MF.getFrameInfo();
964      int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
965      SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
966      if (ObjGPRs == 0)
967        ArgValue = DAG.getLoad(ObjectVT, Root, FIN, NULL, 0);
968      else {
969        SDOperand ArgValue2 = DAG.getLoad(MVT::i32, Root, FIN, NULL, 0);
970        assert(ObjectVT != MVT::i64 && "i64 should already be lowered");
971        ArgValue = DAG.getNode(ARMISD::FMDRR, MVT::f64, ArgValue, ArgValue2);
972      }
973    } else {
974      // Don't emit a dead load.
975      ArgValue = DAG.getNode(ISD::UNDEF, ObjectVT);
976    }
977
978    ArgOffset += ObjSize;   // Move on to the next argument.
979  }
980
981  return ArgValue;
982}
983
984SDOperand
985ARMTargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) {
986  std::vector<SDOperand> ArgValues;
987  SDOperand Root = Op.getOperand(0);
988  unsigned ArgOffset = 0;   // Frame mechanisms handle retaddr slot
989  unsigned NumGPRs = 0;     // GPRs used for parameter passing.
990  unsigned VRegs[4];
991
992  unsigned NumArgs = Op.Val->getNumValues()-1;
993  for (unsigned ArgNo = 0; ArgNo < NumArgs; ++ArgNo)
994    ArgValues.push_back(LowerFORMAL_ARGUMENT(Op, DAG, VRegs, ArgNo,
995                                             NumGPRs, ArgOffset));
996
997  bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
998  if (isVarArg) {
999    static const unsigned GPRArgRegs[] = {
1000      ARM::R0, ARM::R1, ARM::R2, ARM::R3
1001    };
1002
1003    MachineFunction &MF = DAG.getMachineFunction();
1004    MachineRegisterInfo &RegInfo = MF.getRegInfo();
1005    MachineFrameInfo *MFI = MF.getFrameInfo();
1006    ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1007    unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
1008    unsigned VARegSize = (4 - NumGPRs) * 4;
1009    unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
1010    if (VARegSaveSize) {
1011      // If this function is vararg, store any remaining integer argument regs
1012      // to their spots on the stack so that they may be loaded by deferencing
1013      // the result of va_next.
1014      AFI->setVarArgsRegSaveSize(VARegSaveSize);
1015      VarArgsFrameIndex = MFI->CreateFixedObject(VARegSaveSize, ArgOffset +
1016                                                 VARegSaveSize - VARegSize);
1017      SDOperand FIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
1018
1019      SmallVector<SDOperand, 4> MemOps;
1020      for (; NumGPRs < 4; ++NumGPRs) {
1021        unsigned VReg = RegInfo.createVirtualRegister(&ARM::GPRRegClass);
1022        RegInfo.addLiveIn(GPRArgRegs[NumGPRs], VReg);
1023        SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::i32);
1024        SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1025        MemOps.push_back(Store);
1026        FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1027                          DAG.getConstant(4, getPointerTy()));
1028      }
1029      if (!MemOps.empty())
1030        Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1031                           &MemOps[0], MemOps.size());
1032    } else
1033      // This will point to the next argument passed via stack.
1034      VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset);
1035  }
1036
1037  ArgValues.push_back(Root);
1038
1039  // Return the new list of results.
1040  std::vector<MVT::ValueType> RetVT(Op.Val->value_begin(),
1041                                    Op.Val->value_end());
1042  return DAG.getNode(ISD::MERGE_VALUES, RetVT, &ArgValues[0], ArgValues.size());
1043}
1044
1045/// isFloatingPointZero - Return true if this is +0.0.
1046static bool isFloatingPointZero(SDOperand Op) {
1047  if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
1048    return CFP->getValueAPF().isPosZero();
1049  else if (ISD::isEXTLoad(Op.Val) || ISD::isNON_EXTLoad(Op.Val)) {
1050    // Maybe this has already been legalized into the constant pool?
1051    if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
1052      SDOperand WrapperOp = Op.getOperand(1).getOperand(0);
1053      if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
1054        if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
1055          return CFP->getValueAPF().isPosZero();
1056    }
1057  }
1058  return false;
1059}
1060
1061static bool isLegalCmpImmediate(unsigned C, bool isThumb) {
1062  return ( isThumb && (C & ~255U) == 0) ||
1063         (!isThumb && ARM_AM::getSOImmVal(C) != -1);
1064}
1065
1066/// Returns appropriate ARM CMP (cmp) and corresponding condition code for
1067/// the given operands.
1068static SDOperand getARMCmp(SDOperand LHS, SDOperand RHS, ISD::CondCode CC,
1069                           SDOperand &ARMCC, SelectionDAG &DAG, bool isThumb) {
1070  if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.Val)) {
1071    unsigned C = RHSC->getValue();
1072    if (!isLegalCmpImmediate(C, isThumb)) {
1073      // Constant does not fit, try adjusting it by one?
1074      switch (CC) {
1075      default: break;
1076      case ISD::SETLT:
1077      case ISD::SETGE:
1078        if (isLegalCmpImmediate(C-1, isThumb)) {
1079          CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1080          RHS = DAG.getConstant(C-1, MVT::i32);
1081        }
1082        break;
1083      case ISD::SETULT:
1084      case ISD::SETUGE:
1085        if (C > 0 && isLegalCmpImmediate(C-1, isThumb)) {
1086          CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1087          RHS = DAG.getConstant(C-1, MVT::i32);
1088        }
1089        break;
1090      case ISD::SETLE:
1091      case ISD::SETGT:
1092        if (isLegalCmpImmediate(C+1, isThumb)) {
1093          CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1094          RHS = DAG.getConstant(C+1, MVT::i32);
1095        }
1096        break;
1097      case ISD::SETULE:
1098      case ISD::SETUGT:
1099        if (C < 0xffffffff && isLegalCmpImmediate(C+1, isThumb)) {
1100          CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1101          RHS = DAG.getConstant(C+1, MVT::i32);
1102        }
1103        break;
1104      }
1105    }
1106  }
1107
1108  ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
1109  ARMISD::NodeType CompareType;
1110  switch (CondCode) {
1111  default:
1112    CompareType = ARMISD::CMP;
1113    break;
1114  case ARMCC::EQ:
1115  case ARMCC::NE:
1116  case ARMCC::MI:
1117  case ARMCC::PL:
1118    // Uses only N and Z Flags
1119    CompareType = ARMISD::CMPNZ;
1120    break;
1121  }
1122  ARMCC = DAG.getConstant(CondCode, MVT::i32);
1123  return DAG.getNode(CompareType, MVT::Flag, LHS, RHS);
1124}
1125
1126/// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
1127static SDOperand getVFPCmp(SDOperand LHS, SDOperand RHS, SelectionDAG &DAG) {
1128  SDOperand Cmp;
1129  if (!isFloatingPointZero(RHS))
1130    Cmp = DAG.getNode(ARMISD::CMPFP, MVT::Flag, LHS, RHS);
1131  else
1132    Cmp = DAG.getNode(ARMISD::CMPFPw0, MVT::Flag, LHS);
1133  return DAG.getNode(ARMISD::FMSTAT, MVT::Flag, Cmp);
1134}
1135
1136static SDOperand LowerSELECT_CC(SDOperand Op, SelectionDAG &DAG,
1137                                const ARMSubtarget *ST) {
1138  MVT::ValueType VT = Op.getValueType();
1139  SDOperand LHS = Op.getOperand(0);
1140  SDOperand RHS = Op.getOperand(1);
1141  ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1142  SDOperand TrueVal = Op.getOperand(2);
1143  SDOperand FalseVal = Op.getOperand(3);
1144
1145  if (LHS.getValueType() == MVT::i32) {
1146    SDOperand ARMCC;
1147    SDOperand CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1148    SDOperand Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb());
1149    return DAG.getNode(ARMISD::CMOV, VT, FalseVal, TrueVal, ARMCC, CCR, Cmp);
1150  }
1151
1152  ARMCC::CondCodes CondCode, CondCode2;
1153  if (FPCCToARMCC(CC, CondCode, CondCode2))
1154    std::swap(TrueVal, FalseVal);
1155
1156  SDOperand ARMCC = DAG.getConstant(CondCode, MVT::i32);
1157  SDOperand CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1158  SDOperand Cmp = getVFPCmp(LHS, RHS, DAG);
1159  SDOperand Result = DAG.getNode(ARMISD::CMOV, VT, FalseVal, TrueVal,
1160                                 ARMCC, CCR, Cmp);
1161  if (CondCode2 != ARMCC::AL) {
1162    SDOperand ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
1163    // FIXME: Needs another CMP because flag can have but one use.
1164    SDOperand Cmp2 = getVFPCmp(LHS, RHS, DAG);
1165    Result = DAG.getNode(ARMISD::CMOV, VT, Result, TrueVal, ARMCC2, CCR, Cmp2);
1166  }
1167  return Result;
1168}
1169
1170static SDOperand LowerBR_CC(SDOperand Op, SelectionDAG &DAG,
1171                            const ARMSubtarget *ST) {
1172  SDOperand  Chain = Op.getOperand(0);
1173  ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
1174  SDOperand    LHS = Op.getOperand(2);
1175  SDOperand    RHS = Op.getOperand(3);
1176  SDOperand   Dest = Op.getOperand(4);
1177
1178  if (LHS.getValueType() == MVT::i32) {
1179    SDOperand ARMCC;
1180    SDOperand CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1181    SDOperand Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb());
1182    return DAG.getNode(ARMISD::BRCOND, MVT::Other, Chain, Dest, ARMCC, CCR,Cmp);
1183  }
1184
1185  assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
1186  ARMCC::CondCodes CondCode, CondCode2;
1187  if (FPCCToARMCC(CC, CondCode, CondCode2))
1188    // Swap the LHS/RHS of the comparison if needed.
1189    std::swap(LHS, RHS);
1190
1191  SDOperand Cmp = getVFPCmp(LHS, RHS, DAG);
1192  SDOperand ARMCC = DAG.getConstant(CondCode, MVT::i32);
1193  SDOperand CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1194  SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
1195  SDOperand Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
1196  SDOperand Res = DAG.getNode(ARMISD::BRCOND, VTList, Ops, 5);
1197  if (CondCode2 != ARMCC::AL) {
1198    ARMCC = DAG.getConstant(CondCode2, MVT::i32);
1199    SDOperand Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
1200    Res = DAG.getNode(ARMISD::BRCOND, VTList, Ops, 5);
1201  }
1202  return Res;
1203}
1204
1205SDOperand ARMTargetLowering::LowerBR_JT(SDOperand Op, SelectionDAG &DAG) {
1206  SDOperand Chain = Op.getOperand(0);
1207  SDOperand Table = Op.getOperand(1);
1208  SDOperand Index = Op.getOperand(2);
1209
1210  MVT::ValueType PTy = getPointerTy();
1211  JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
1212  ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
1213  SDOperand UId =  DAG.getConstant(AFI->createJumpTableUId(), PTy);
1214  SDOperand JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
1215  Table = DAG.getNode(ARMISD::WrapperJT, MVT::i32, JTI, UId);
1216  Index = DAG.getNode(ISD::MUL, PTy, Index, DAG.getConstant(4, PTy));
1217  SDOperand Addr = DAG.getNode(ISD::ADD, PTy, Index, Table);
1218  bool isPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_;
1219  Addr = DAG.getLoad(isPIC ? (MVT::ValueType)MVT::i32 : PTy,
1220                     Chain, Addr, NULL, 0);
1221  Chain = Addr.getValue(1);
1222  if (isPIC)
1223    Addr = DAG.getNode(ISD::ADD, PTy, Addr, Table);
1224  return DAG.getNode(ARMISD::BR_JT, MVT::Other, Chain, Addr, JTI, UId);
1225}
1226
1227static SDOperand LowerFP_TO_INT(SDOperand Op, SelectionDAG &DAG) {
1228  unsigned Opc =
1229    Op.getOpcode() == ISD::FP_TO_SINT ? ARMISD::FTOSI : ARMISD::FTOUI;
1230  Op = DAG.getNode(Opc, MVT::f32, Op.getOperand(0));
1231  return DAG.getNode(ISD::BIT_CONVERT, MVT::i32, Op);
1232}
1233
1234static SDOperand LowerINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
1235  MVT::ValueType VT = Op.getValueType();
1236  unsigned Opc =
1237    Op.getOpcode() == ISD::SINT_TO_FP ? ARMISD::SITOF : ARMISD::UITOF;
1238
1239  Op = DAG.getNode(ISD::BIT_CONVERT, MVT::f32, Op.getOperand(0));
1240  return DAG.getNode(Opc, VT, Op);
1241}
1242
1243static SDOperand LowerFCOPYSIGN(SDOperand Op, SelectionDAG &DAG) {
1244  // Implement fcopysign with a fabs and a conditional fneg.
1245  SDOperand Tmp0 = Op.getOperand(0);
1246  SDOperand Tmp1 = Op.getOperand(1);
1247  MVT::ValueType VT = Op.getValueType();
1248  MVT::ValueType SrcVT = Tmp1.getValueType();
1249  SDOperand AbsVal = DAG.getNode(ISD::FABS, VT, Tmp0);
1250  SDOperand Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG);
1251  SDOperand ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
1252  SDOperand CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1253  return DAG.getNode(ARMISD::CNEG, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
1254}
1255
1256SDOperand ARMTargetLowering::LowerMEMCPYInline(SDOperand Chain,
1257                                               SDOperand Dest,
1258                                               SDOperand Source,
1259                                               unsigned Size,
1260                                               unsigned Align,
1261                                               SelectionDAG &DAG) {
1262  // Do repeated 4-byte loads and stores. To be improved.
1263  assert((Align & 3) == 0 && "Expected 4-byte aligned addresses!");
1264  unsigned BytesLeft = Size & 3;
1265  unsigned NumMemOps = Size >> 2;
1266  unsigned EmittedNumMemOps = 0;
1267  unsigned SrcOff = 0, DstOff = 0;
1268  MVT::ValueType VT = MVT::i32;
1269  unsigned VTSize = 4;
1270  unsigned i = 0;
1271  const unsigned MAX_LOADS_IN_LDM = 6;
1272  SDOperand TFOps[MAX_LOADS_IN_LDM];
1273  SDOperand Loads[MAX_LOADS_IN_LDM];
1274
1275  // Emit up to MAX_LOADS_IN_LDM loads, then a TokenFactor barrier, then the
1276  // same number of stores.  The loads and stores will get combined into
1277  // ldm/stm later on.
1278  while (EmittedNumMemOps < NumMemOps) {
1279    for (i = 0;
1280         i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
1281      Loads[i] = DAG.getLoad(VT, Chain,
1282                             DAG.getNode(ISD::ADD, MVT::i32, Source,
1283                                         DAG.getConstant(SrcOff, MVT::i32)),
1284                             NULL, 0);
1285      TFOps[i] = Loads[i].getValue(1);
1286      SrcOff += VTSize;
1287    }
1288    Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, &TFOps[0], i);
1289
1290    for (i = 0;
1291         i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
1292      TFOps[i] = DAG.getStore(Chain, Loads[i],
1293                           DAG.getNode(ISD::ADD, MVT::i32, Dest,
1294                                       DAG.getConstant(DstOff, MVT::i32)),
1295                           NULL, 0);
1296      DstOff += VTSize;
1297    }
1298    Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, &TFOps[0], i);
1299
1300    EmittedNumMemOps += i;
1301  }
1302
1303  if (BytesLeft == 0)
1304    return Chain;
1305
1306  // Issue loads / stores for the trailing (1 - 3) bytes.
1307  unsigned BytesLeftSave = BytesLeft;
1308  i = 0;
1309  while (BytesLeft) {
1310    if (BytesLeft >= 2) {
1311      VT = MVT::i16;
1312      VTSize = 2;
1313    } else {
1314      VT = MVT::i8;
1315      VTSize = 1;
1316    }
1317
1318    Loads[i] = DAG.getLoad(VT, Chain,
1319                           DAG.getNode(ISD::ADD, MVT::i32, Source,
1320                                       DAG.getConstant(SrcOff, MVT::i32)),
1321                           NULL, 0);
1322    TFOps[i] = Loads[i].getValue(1);
1323    ++i;
1324    SrcOff += VTSize;
1325    BytesLeft -= VTSize;
1326  }
1327  Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, &TFOps[0], i);
1328
1329  i = 0;
1330  BytesLeft = BytesLeftSave;
1331  while (BytesLeft) {
1332    if (BytesLeft >= 2) {
1333      VT = MVT::i16;
1334      VTSize = 2;
1335    } else {
1336      VT = MVT::i8;
1337      VTSize = 1;
1338    }
1339
1340    TFOps[i] = DAG.getStore(Chain, Loads[i],
1341                            DAG.getNode(ISD::ADD, MVT::i32, Dest,
1342                                        DAG.getConstant(DstOff, MVT::i32)),
1343                            NULL, 0);
1344    ++i;
1345    DstOff += VTSize;
1346    BytesLeft -= VTSize;
1347  }
1348  return DAG.getNode(ISD::TokenFactor, MVT::Other, &TFOps[0], i);
1349}
1350
1351static SDNode *ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
1352  // Turn f64->i64 into FMRRD.
1353  assert(N->getValueType(0) == MVT::i64 &&
1354         N->getOperand(0).getValueType() == MVT::f64);
1355
1356  SDOperand Op = N->getOperand(0);
1357  SDOperand Cvt = DAG.getNode(ARMISD::FMRRD, DAG.getVTList(MVT::i32, MVT::i32),
1358                              &Op, 1);
1359
1360  // Merge the pieces into a single i64 value.
1361  return DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Cvt, Cvt.getValue(1)).Val;
1362}
1363
1364static SDNode *ExpandSRx(SDNode *N, SelectionDAG &DAG, const ARMSubtarget *ST) {
1365  assert(N->getValueType(0) == MVT::i64 &&
1366         (N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
1367         "Unknown shift to lower!");
1368
1369  // We only lower SRA, SRL of 1 here, all others use generic lowering.
1370  if (!isa<ConstantSDNode>(N->getOperand(1)) ||
1371      cast<ConstantSDNode>(N->getOperand(1))->getValue() != 1)
1372    return 0;
1373
1374  // If we are in thumb mode, we don't have RRX.
1375  if (ST->isThumb()) return 0;
1376
1377  // Okay, we have a 64-bit SRA or SRL of 1.  Lower this to an RRX expr.
1378  SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, N->getOperand(0),
1379                             DAG.getConstant(0, MVT::i32));
1380  SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, N->getOperand(0),
1381                             DAG.getConstant(1, MVT::i32));
1382
1383  // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
1384  // captures the result into a carry flag.
1385  unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
1386  Hi = DAG.getNode(Opc, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
1387
1388  // The low part is an ARMISD::RRX operand, which shifts the carry in.
1389  Lo = DAG.getNode(ARMISD::RRX, MVT::i32, Lo, Hi.getValue(1));
1390
1391  // Merge the pieces into a single i64 value.
1392 return DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Lo, Hi).Val;
1393}
1394
1395
1396SDOperand ARMTargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
1397  switch (Op.getOpcode()) {
1398  default: assert(0 && "Don't know how to custom lower this!"); abort();
1399  case ISD::ConstantPool:  return LowerConstantPool(Op, DAG);
1400  case ISD::GlobalAddress:
1401    return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
1402      LowerGlobalAddressELF(Op, DAG);
1403  case ISD::GlobalTLSAddress:   return LowerGlobalTLSAddress(Op, DAG);
1404  case ISD::CALL:          return LowerCALL(Op, DAG);
1405  case ISD::RET:           return LowerRET(Op, DAG);
1406  case ISD::SELECT_CC:     return LowerSELECT_CC(Op, DAG, Subtarget);
1407  case ISD::BR_CC:         return LowerBR_CC(Op, DAG, Subtarget);
1408  case ISD::BR_JT:         return LowerBR_JT(Op, DAG);
1409  case ISD::VASTART:       return LowerVASTART(Op, DAG, VarArgsFrameIndex);
1410  case ISD::SINT_TO_FP:
1411  case ISD::UINT_TO_FP:    return LowerINT_TO_FP(Op, DAG);
1412  case ISD::FP_TO_SINT:
1413  case ISD::FP_TO_UINT:    return LowerFP_TO_INT(Op, DAG);
1414  case ISD::FCOPYSIGN:     return LowerFCOPYSIGN(Op, DAG);
1415  case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
1416  case ISD::RETURNADDR:    break;
1417  case ISD::FRAMEADDR:     break;
1418  case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
1419  case ISD::MEMCPY:        return LowerMEMCPY(Op, DAG);
1420  case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
1421
1422
1423  // FIXME: Remove these when LegalizeDAGTypes lands.
1424  case ISD::BIT_CONVERT:   return SDOperand(ExpandBIT_CONVERT(Op.Val, DAG), 0);
1425  case ISD::SRL:
1426  case ISD::SRA:           return SDOperand(ExpandSRx(Op.Val, DAG,Subtarget),0);
1427  }
1428  return SDOperand();
1429}
1430
1431
1432/// ExpandOperationResult - Provide custom lowering hooks for expanding
1433/// operations.
1434SDNode *ARMTargetLowering::ExpandOperationResult(SDNode *N, SelectionDAG &DAG) {
1435  switch (N->getOpcode()) {
1436  default: assert(0 && "Don't know how to custom expand this!"); abort();
1437  case ISD::BIT_CONVERT:   return ExpandBIT_CONVERT(N, DAG);
1438  case ISD::SRL:
1439  case ISD::SRA:           return ExpandSRx(N, DAG, Subtarget);
1440  }
1441}
1442
1443
1444//===----------------------------------------------------------------------===//
1445//                           ARM Scheduler Hooks
1446//===----------------------------------------------------------------------===//
1447
1448MachineBasicBlock *
1449ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
1450                                           MachineBasicBlock *BB) {
1451  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1452  switch (MI->getOpcode()) {
1453  default: assert(false && "Unexpected instr type to insert");
1454  case ARM::tMOVCCr: {
1455    // To "insert" a SELECT_CC instruction, we actually have to insert the
1456    // diamond control-flow pattern.  The incoming instruction knows the
1457    // destination vreg to set, the condition code register to branch on, the
1458    // true/false values to select between, and a branch opcode to use.
1459    const BasicBlock *LLVM_BB = BB->getBasicBlock();
1460    ilist<MachineBasicBlock>::iterator It = BB;
1461    ++It;
1462
1463    //  thisMBB:
1464    //  ...
1465    //   TrueVal = ...
1466    //   cmpTY ccX, r1, r2
1467    //   bCC copy1MBB
1468    //   fallthrough --> copy0MBB
1469    MachineBasicBlock *thisMBB  = BB;
1470    MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
1471    MachineBasicBlock *sinkMBB  = new MachineBasicBlock(LLVM_BB);
1472    BuildMI(BB, TII->get(ARM::tBcc)).addMBB(sinkMBB)
1473      .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
1474    MachineFunction *F = BB->getParent();
1475    F->getBasicBlockList().insert(It, copy0MBB);
1476    F->getBasicBlockList().insert(It, sinkMBB);
1477    // Update machine-CFG edges by first adding all successors of the current
1478    // block to the new block which will contain the Phi node for the select.
1479    for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
1480        e = BB->succ_end(); i != e; ++i)
1481      sinkMBB->addSuccessor(*i);
1482    // Next, remove all successors of the current block, and add the true
1483    // and fallthrough blocks as its successors.
1484    while(!BB->succ_empty())
1485      BB->removeSuccessor(BB->succ_begin());
1486    BB->addSuccessor(copy0MBB);
1487    BB->addSuccessor(sinkMBB);
1488
1489    //  copy0MBB:
1490    //   %FalseValue = ...
1491    //   # fallthrough to sinkMBB
1492    BB = copy0MBB;
1493
1494    // Update machine-CFG edges
1495    BB->addSuccessor(sinkMBB);
1496
1497    //  sinkMBB:
1498    //   %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
1499    //  ...
1500    BB = sinkMBB;
1501    BuildMI(BB, TII->get(ARM::PHI), MI->getOperand(0).getReg())
1502      .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
1503      .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
1504
1505    delete MI;   // The pseudo instruction is gone now.
1506    return BB;
1507  }
1508  }
1509}
1510
1511//===----------------------------------------------------------------------===//
1512//                           ARM Optimization Hooks
1513//===----------------------------------------------------------------------===//
1514
1515/// PerformFMRRDCombine - Target-specific dag combine xforms for ARMISD::FMRRD.
1516static SDOperand PerformFMRRDCombine(SDNode *N,
1517                                     TargetLowering::DAGCombinerInfo &DCI) {
1518  // fmrrd(fmdrr x, y) -> x,y
1519  SDOperand InDouble = N->getOperand(0);
1520  if (InDouble.getOpcode() == ARMISD::FMDRR)
1521    return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
1522  return SDOperand();
1523}
1524
1525SDOperand ARMTargetLowering::PerformDAGCombine(SDNode *N,
1526                                               DAGCombinerInfo &DCI) const {
1527  switch (N->getOpcode()) {
1528  default: break;
1529  case ARMISD::FMRRD: return PerformFMRRDCombine(N, DCI);
1530  }
1531
1532  return SDOperand();
1533}
1534
1535
1536/// isLegalAddressImmediate - Return true if the integer value can be used
1537/// as the offset of the target addressing mode for load / store of the
1538/// given type.
1539static bool isLegalAddressImmediate(int64_t V, MVT::ValueType VT,
1540                                    const ARMSubtarget *Subtarget) {
1541  if (V == 0)
1542    return true;
1543
1544  if (Subtarget->isThumb()) {
1545    if (V < 0)
1546      return false;
1547
1548    unsigned Scale = 1;
1549    switch (VT) {
1550    default: return false;
1551    case MVT::i1:
1552    case MVT::i8:
1553      // Scale == 1;
1554      break;
1555    case MVT::i16:
1556      // Scale == 2;
1557      Scale = 2;
1558      break;
1559    case MVT::i32:
1560      // Scale == 4;
1561      Scale = 4;
1562      break;
1563    }
1564
1565    if ((V & (Scale - 1)) != 0)
1566      return false;
1567    V /= Scale;
1568    return V == V & ((1LL << 5) - 1);
1569  }
1570
1571  if (V < 0)
1572    V = - V;
1573  switch (VT) {
1574  default: return false;
1575  case MVT::i1:
1576  case MVT::i8:
1577  case MVT::i32:
1578    // +- imm12
1579    return V == V & ((1LL << 12) - 1);
1580  case MVT::i16:
1581    // +- imm8
1582    return V == V & ((1LL << 8) - 1);
1583  case MVT::f32:
1584  case MVT::f64:
1585    if (!Subtarget->hasVFP2())
1586      return false;
1587    if ((V & 3) != 0)
1588      return false;
1589    V >>= 2;
1590    return V == V & ((1LL << 8) - 1);
1591  }
1592}
1593
1594/// isLegalAddressingMode - Return true if the addressing mode represented
1595/// by AM is legal for this target, for a load/store of the specified type.
1596bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
1597                                              const Type *Ty) const {
1598  if (!isLegalAddressImmediate(AM.BaseOffs, getValueType(Ty), Subtarget))
1599    return false;
1600
1601  // Can never fold addr of global into load/store.
1602  if (AM.BaseGV)
1603    return false;
1604
1605  switch (AM.Scale) {
1606  case 0:  // no scale reg, must be "r+i" or "r", or "i".
1607    break;
1608  case 1:
1609    if (Subtarget->isThumb())
1610      return false;
1611    // FALL THROUGH.
1612  default:
1613    // ARM doesn't support any R+R*scale+imm addr modes.
1614    if (AM.BaseOffs)
1615      return false;
1616
1617    int Scale = AM.Scale;
1618    switch (getValueType(Ty)) {
1619    default: return false;
1620    case MVT::i1:
1621    case MVT::i8:
1622    case MVT::i32:
1623    case MVT::i64:
1624      // This assumes i64 is legalized to a pair of i32. If not (i.e.
1625      // ldrd / strd are used, then its address mode is same as i16.
1626      // r + r
1627      if (Scale < 0) Scale = -Scale;
1628      if (Scale == 1)
1629        return true;
1630      // r + r << imm
1631      return isPowerOf2_32(Scale & ~1);
1632    case MVT::i16:
1633      // r + r
1634      if (((unsigned)AM.HasBaseReg + Scale) <= 2)
1635        return true;
1636      return false;
1637
1638    case MVT::isVoid:
1639      // Note, we allow "void" uses (basically, uses that aren't loads or
1640      // stores), because arm allows folding a scale into many arithmetic
1641      // operations.  This should be made more precise and revisited later.
1642
1643      // Allow r << imm, but the imm has to be a multiple of two.
1644      if (AM.Scale & 1) return false;
1645      return isPowerOf2_32(AM.Scale);
1646    }
1647    break;
1648  }
1649  return true;
1650}
1651
1652
1653static bool getIndexedAddressParts(SDNode *Ptr, MVT::ValueType VT,
1654                                   bool isSEXTLoad, SDOperand &Base,
1655                                   SDOperand &Offset, bool &isInc,
1656                                   SelectionDAG &DAG) {
1657  if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
1658    return false;
1659
1660  if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
1661    // AddressingMode 3
1662    Base = Ptr->getOperand(0);
1663    if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
1664      int RHSC = (int)RHS->getValue();
1665      if (RHSC < 0 && RHSC > -256) {
1666        isInc = false;
1667        Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
1668        return true;
1669      }
1670    }
1671    isInc = (Ptr->getOpcode() == ISD::ADD);
1672    Offset = Ptr->getOperand(1);
1673    return true;
1674  } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
1675    // AddressingMode 2
1676    if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
1677      int RHSC = (int)RHS->getValue();
1678      if (RHSC < 0 && RHSC > -0x1000) {
1679        isInc = false;
1680        Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
1681        Base = Ptr->getOperand(0);
1682        return true;
1683      }
1684    }
1685
1686    if (Ptr->getOpcode() == ISD::ADD) {
1687      isInc = true;
1688      ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
1689      if (ShOpcVal != ARM_AM::no_shift) {
1690        Base = Ptr->getOperand(1);
1691        Offset = Ptr->getOperand(0);
1692      } else {
1693        Base = Ptr->getOperand(0);
1694        Offset = Ptr->getOperand(1);
1695      }
1696      return true;
1697    }
1698
1699    isInc = (Ptr->getOpcode() == ISD::ADD);
1700    Base = Ptr->getOperand(0);
1701    Offset = Ptr->getOperand(1);
1702    return true;
1703  }
1704
1705  // FIXME: Use FLDM / FSTM to emulate indexed FP load / store.
1706  return false;
1707}
1708
1709/// getPreIndexedAddressParts - returns true by value, base pointer and
1710/// offset pointer and addressing mode by reference if the node's address
1711/// can be legally represented as pre-indexed load / store address.
1712bool
1713ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDOperand &Base,
1714                                             SDOperand &Offset,
1715                                             ISD::MemIndexedMode &AM,
1716                                             SelectionDAG &DAG) {
1717  if (Subtarget->isThumb())
1718    return false;
1719
1720  MVT::ValueType VT;
1721  SDOperand Ptr;
1722  bool isSEXTLoad = false;
1723  if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1724    Ptr = LD->getBasePtr();
1725    VT  = LD->getMemoryVT();
1726    isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
1727  } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1728    Ptr = ST->getBasePtr();
1729    VT  = ST->getMemoryVT();
1730  } else
1731    return false;
1732
1733  bool isInc;
1734  bool isLegal = getIndexedAddressParts(Ptr.Val, VT, isSEXTLoad, Base, Offset,
1735                                        isInc, DAG);
1736  if (isLegal) {
1737    AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
1738    return true;
1739  }
1740  return false;
1741}
1742
1743/// getPostIndexedAddressParts - returns true by value, base pointer and
1744/// offset pointer and addressing mode by reference if this node can be
1745/// combined with a load / store to form a post-indexed load / store.
1746bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
1747                                                   SDOperand &Base,
1748                                                   SDOperand &Offset,
1749                                                   ISD::MemIndexedMode &AM,
1750                                                   SelectionDAG &DAG) {
1751  if (Subtarget->isThumb())
1752    return false;
1753
1754  MVT::ValueType VT;
1755  SDOperand Ptr;
1756  bool isSEXTLoad = false;
1757  if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1758    VT  = LD->getMemoryVT();
1759    isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
1760  } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1761    VT  = ST->getMemoryVT();
1762  } else
1763    return false;
1764
1765  bool isInc;
1766  bool isLegal = getIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
1767                                        isInc, DAG);
1768  if (isLegal) {
1769    AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
1770    return true;
1771  }
1772  return false;
1773}
1774
1775void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
1776                                                       uint64_t Mask,
1777                                                       uint64_t &KnownZero,
1778                                                       uint64_t &KnownOne,
1779                                                       const SelectionDAG &DAG,
1780                                                       unsigned Depth) const {
1781  KnownZero = 0;
1782  KnownOne = 0;
1783  switch (Op.getOpcode()) {
1784  default: break;
1785  case ARMISD::CMOV: {
1786    // Bits are known zero/one if known on the LHS and RHS.
1787    DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
1788    if (KnownZero == 0 && KnownOne == 0) return;
1789
1790    uint64_t KnownZeroRHS, KnownOneRHS;
1791    DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
1792                          KnownZeroRHS, KnownOneRHS, Depth+1);
1793    KnownZero &= KnownZeroRHS;
1794    KnownOne  &= KnownOneRHS;
1795    return;
1796  }
1797  }
1798}
1799
1800//===----------------------------------------------------------------------===//
1801//                           ARM Inline Assembly Support
1802//===----------------------------------------------------------------------===//
1803
1804/// getConstraintType - Given a constraint letter, return the type of
1805/// constraint it is for this target.
1806ARMTargetLowering::ConstraintType
1807ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
1808  if (Constraint.size() == 1) {
1809    switch (Constraint[0]) {
1810    default:  break;
1811    case 'l': return C_RegisterClass;
1812    case 'w': return C_RegisterClass;
1813    }
1814  }
1815  return TargetLowering::getConstraintType(Constraint);
1816}
1817
1818std::pair<unsigned, const TargetRegisterClass*>
1819ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
1820                                                MVT::ValueType VT) const {
1821  if (Constraint.size() == 1) {
1822    // GCC RS6000 Constraint Letters
1823    switch (Constraint[0]) {
1824    case 'l':
1825    // FIXME: in thumb mode, 'l' is only low-regs.
1826    // FALL THROUGH.
1827    case 'r':
1828      return std::make_pair(0U, ARM::GPRRegisterClass);
1829    case 'w':
1830      if (VT == MVT::f32)
1831        return std::make_pair(0U, ARM::SPRRegisterClass);
1832      if (VT == MVT::f64)
1833        return std::make_pair(0U, ARM::DPRRegisterClass);
1834      break;
1835    }
1836  }
1837  return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
1838}
1839
1840std::vector<unsigned> ARMTargetLowering::
1841getRegClassForInlineAsmConstraint(const std::string &Constraint,
1842                                  MVT::ValueType VT) const {
1843  if (Constraint.size() != 1)
1844    return std::vector<unsigned>();
1845
1846  switch (Constraint[0]) {      // GCC ARM Constraint Letters
1847  default: break;
1848  case 'l':
1849  case 'r':
1850    return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
1851                                 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
1852                                 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
1853                                 ARM::R12, ARM::LR, 0);
1854  case 'w':
1855    if (VT == MVT::f32)
1856      return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
1857                                   ARM::S4, ARM::S5, ARM::S6, ARM::S7,
1858                                   ARM::S8, ARM::S9, ARM::S10, ARM::S11,
1859                                   ARM::S12,ARM::S13,ARM::S14,ARM::S15,
1860                                   ARM::S16,ARM::S17,ARM::S18,ARM::S19,
1861                                   ARM::S20,ARM::S21,ARM::S22,ARM::S23,
1862                                   ARM::S24,ARM::S25,ARM::S26,ARM::S27,
1863                                   ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
1864    if (VT == MVT::f64)
1865      return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
1866                                   ARM::D4, ARM::D5, ARM::D6, ARM::D7,
1867                                   ARM::D8, ARM::D9, ARM::D10,ARM::D11,
1868                                   ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
1869      break;
1870  }
1871
1872  return std::vector<unsigned>();
1873}
1874