ARMISelLowering.cpp revision e0703c84ddeb1a1276de4e38210c1127ef5df130
1//===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file was developed by Evan Cheng and is distributed under
6// the University of Illinois Open Source 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/CodeGen/MachineBasicBlock.h"
27#include "llvm/CodeGen/MachineFrameInfo.h"
28#include "llvm/CodeGen/MachineFunction.h"
29#include "llvm/CodeGen/MachineInstrBuilder.h"
30#include "llvm/CodeGen/SelectionDAG.h"
31#include "llvm/CodeGen/SSARegMap.h"
32#include "llvm/Target/TargetOptions.h"
33#include "llvm/ADT/VectorExtras.h"
34#include "llvm/Support/MathExtras.h"
35using namespace llvm;
36
37ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
38    : TargetLowering(TM), ARMPCLabelIndex(0) {
39  Subtarget = &TM.getSubtarget<ARMSubtarget>();
40
41  if (Subtarget->isTargetDarwin()) {
42    // Don't have these.
43    setLibcallName(RTLIB::UINTTOFP_I64_F32, NULL);
44    setLibcallName(RTLIB::UINTTOFP_I64_F64, NULL);
45
46    // Uses VFP for Thumb libfuncs if available.
47    if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
48      // Single-precision floating-point arithmetic.
49      setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
50      setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
51      setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
52      setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
53
54      // Double-precision floating-point arithmetic.
55      setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
56      setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
57      setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
58      setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
59
60      // Single-precision comparisons.
61      setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
62      setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
63      setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
64      setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
65      setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
66      setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
67      setLibcallName(RTLIB::UO_F32,  "__unordsf2vfp");
68      setLibcallName(RTLIB::O_F32,   "__unordsf2vfp");
69
70      setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
71      setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
72      setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
73      setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
74      setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
75      setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
76      setCmpLibcallCC(RTLIB::UO_F32,  ISD::SETNE);
77      setCmpLibcallCC(RTLIB::O_F32,   ISD::SETEQ);
78
79      // Double-precision comparisons.
80      setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
81      setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
82      setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
83      setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
84      setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
85      setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
86      setLibcallName(RTLIB::UO_F64,  "__unorddf2vfp");
87      setLibcallName(RTLIB::O_F64,   "__unorddf2vfp");
88
89      setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
90      setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
91      setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
92      setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
93      setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
94      setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
95      setCmpLibcallCC(RTLIB::UO_F64,  ISD::SETNE);
96      setCmpLibcallCC(RTLIB::O_F64,   ISD::SETEQ);
97
98      // Floating-point to integer conversions.
99      // i64 conversions are done via library routines even when generating VFP
100      // instructions, so use the same ones.
101      setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
102      setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
103      setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
104      setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
105
106      // Conversions between floating types.
107      setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
108      setLibcallName(RTLIB::FPEXT_F32_F64,   "__extendsfdf2vfp");
109
110      // Integer to floating-point conversions.
111      // i64 conversions are done via library routines even when generating VFP
112      // instructions, so use the same ones.
113      // FIXME: There appears to be some naming inconsistency in ARM libgcc: e.g.
114      // __floatunsidf vs. __floatunssidfvfp.
115      setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
116      setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
117      setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
118      setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
119    }
120  }
121
122  addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
123  if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb()) {
124    addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
125    addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
126  }
127  computeRegisterProperties();
128
129  // ARM does not have f32 extending load.
130  setLoadXAction(ISD::EXTLOAD, MVT::f32, Expand);
131
132  // ARM supports all 4 flavors of integer indexed load / store.
133  for (unsigned im = (unsigned)ISD::PRE_INC;
134       im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
135    setIndexedLoadAction(im,  MVT::i1,  Legal);
136    setIndexedLoadAction(im,  MVT::i8,  Legal);
137    setIndexedLoadAction(im,  MVT::i16, Legal);
138    setIndexedLoadAction(im,  MVT::i32, Legal);
139    setIndexedStoreAction(im, MVT::i1,  Legal);
140    setIndexedStoreAction(im, MVT::i8,  Legal);
141    setIndexedStoreAction(im, MVT::i16, Legal);
142    setIndexedStoreAction(im, MVT::i32, Legal);
143  }
144
145  // i64 operation support.
146  if (Subtarget->isThumb()) {
147    setOperationAction(ISD::MUL,     MVT::i64, Expand);
148    setOperationAction(ISD::MULHU,   MVT::i32, Expand);
149    setOperationAction(ISD::MULHS,   MVT::i32, Expand);
150    setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
151    setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
152  } else {
153    setOperationAction(ISD::MUL,     MVT::i64, Expand);
154    setOperationAction(ISD::MULHU,   MVT::i32, Expand);
155    if (!Subtarget->hasV6Ops())
156      setOperationAction(ISD::MULHS, MVT::i32, Expand);
157  }
158  setOperationAction(ISD::SHL_PARTS, MVT::i32, Expand);
159  setOperationAction(ISD::SRA_PARTS, MVT::i32, Expand);
160  setOperationAction(ISD::SRL_PARTS, MVT::i32, Expand);
161  setOperationAction(ISD::SRL,       MVT::i64, Custom);
162  setOperationAction(ISD::SRA,       MVT::i64, Custom);
163
164  // ARM does not have ROTL.
165  setOperationAction(ISD::ROTL,  MVT::i32, Expand);
166  setOperationAction(ISD::CTTZ , MVT::i32, Expand);
167  setOperationAction(ISD::CTPOP, MVT::i32, Expand);
168  if (!Subtarget->hasV5TOps() || Subtarget->isThumb())
169    setOperationAction(ISD::CTLZ, MVT::i32, Expand);
170
171  // Only ARMv6 has BSWAP.
172  if (!Subtarget->hasV6Ops())
173    setOperationAction(ISD::BSWAP, MVT::i32, Expand);
174
175  // These are expanded into libcalls.
176  setOperationAction(ISD::SDIV,  MVT::i32, Expand);
177  setOperationAction(ISD::UDIV,  MVT::i32, Expand);
178  setOperationAction(ISD::SREM,  MVT::i32, Expand);
179  setOperationAction(ISD::UREM,  MVT::i32, Expand);
180  setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
181  setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
182
183  // Support label based line numbers.
184  setOperationAction(ISD::LOCATION, MVT::Other, Expand);
185  setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
186
187  setOperationAction(ISD::RET,           MVT::Other, Custom);
188  setOperationAction(ISD::GlobalAddress, MVT::i32,   Custom);
189  setOperationAction(ISD::ConstantPool,  MVT::i32,   Custom);
190  setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
191  setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
192
193  // Expand mem operations genericly.
194  setOperationAction(ISD::MEMSET          , MVT::Other, Expand);
195  setOperationAction(ISD::MEMCPY          , MVT::Other, Custom);
196  setOperationAction(ISD::MEMMOVE         , MVT::Other, Expand);
197
198  // Use the default implementation.
199  setOperationAction(ISD::VASTART           , MVT::Other, Expand);
200  setOperationAction(ISD::VAARG             , MVT::Other, Expand);
201  setOperationAction(ISD::VACOPY            , MVT::Other, Expand);
202  setOperationAction(ISD::VAEND             , MVT::Other, Expand);
203  setOperationAction(ISD::STACKSAVE,          MVT::Other, Expand);
204  setOperationAction(ISD::STACKRESTORE,       MVT::Other, Expand);
205  setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32  , Expand);
206
207  if (!Subtarget->hasV6Ops()) {
208    setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
209    setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8,  Expand);
210  }
211  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
212
213  if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb())
214    // Turn f64->i64 into FMRRD iff target supports vfp2.
215    setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
216
217  setOperationAction(ISD::SETCC    , MVT::i32, Expand);
218  setOperationAction(ISD::SETCC    , MVT::f32, Expand);
219  setOperationAction(ISD::SETCC    , MVT::f64, Expand);
220  setOperationAction(ISD::SELECT   , MVT::i32, Expand);
221  setOperationAction(ISD::SELECT   , MVT::f32, Expand);
222  setOperationAction(ISD::SELECT   , MVT::f64, Expand);
223  setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
224  setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
225  setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
226
227  setOperationAction(ISD::BRCOND   , MVT::Other, Expand);
228  setOperationAction(ISD::BR_CC    , MVT::i32,   Custom);
229  setOperationAction(ISD::BR_CC    , MVT::f32,   Custom);
230  setOperationAction(ISD::BR_CC    , MVT::f64,   Custom);
231  setOperationAction(ISD::BR_JT    , MVT::Other, Custom);
232
233  setOperationAction(ISD::VASTART,       MVT::Other, Custom);
234  setOperationAction(ISD::VACOPY,        MVT::Other, Expand);
235  setOperationAction(ISD::VAEND,         MVT::Other, Expand);
236  setOperationAction(ISD::STACKSAVE,     MVT::Other, Expand);
237  setOperationAction(ISD::STACKRESTORE,  MVT::Other, Expand);
238
239  // FP Constants can't be immediates.
240  setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
241  setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
242
243  // We don't support sin/cos/fmod/copysign/pow
244  setOperationAction(ISD::FSIN     , MVT::f64, Expand);
245  setOperationAction(ISD::FSIN     , MVT::f32, Expand);
246  setOperationAction(ISD::FCOS     , MVT::f32, Expand);
247  setOperationAction(ISD::FCOS     , MVT::f64, Expand);
248  setOperationAction(ISD::FREM     , MVT::f64, Expand);
249  setOperationAction(ISD::FREM     , MVT::f32, Expand);
250  setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
251  setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
252  setOperationAction(ISD::FPOW     , MVT::f64, Expand);
253  setOperationAction(ISD::FPOW     , MVT::f32, Expand);
254
255  // int <-> fp are custom expanded into bit_convert + ARMISD ops.
256  setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
257  setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
258  setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
259  setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
260
261  setStackPointerRegisterToSaveRestore(ARM::SP);
262  setSchedulingPreference(SchedulingForRegPressure);
263  setIfCvtBlockSizeLimit(Subtarget->isThumb() ? 0 : 10);
264  setIfCvtDupBlockSizeLimit(Subtarget->isThumb() ? 0 : 2);
265
266  maxStoresPerMemcpy = 1;   //// temporary - rewrite interface to use type
267}
268
269
270const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
271  switch (Opcode) {
272  default: return 0;
273  case ARMISD::Wrapper:       return "ARMISD::Wrapper";
274  case ARMISD::WrapperJT:     return "ARMISD::WrapperJT";
275  case ARMISD::CALL:          return "ARMISD::CALL";
276  case ARMISD::CALL_PRED:     return "ARMISD::CALL_PRED";
277  case ARMISD::CALL_NOLINK:   return "ARMISD::CALL_NOLINK";
278  case ARMISD::tCALL:         return "ARMISD::tCALL";
279  case ARMISD::BRCOND:        return "ARMISD::BRCOND";
280  case ARMISD::BR_JT:         return "ARMISD::BR_JT";
281  case ARMISD::RET_FLAG:      return "ARMISD::RET_FLAG";
282  case ARMISD::PIC_ADD:       return "ARMISD::PIC_ADD";
283  case ARMISD::CMP:           return "ARMISD::CMP";
284  case ARMISD::CMPNZ:         return "ARMISD::CMPNZ";
285  case ARMISD::CMPFP:         return "ARMISD::CMPFP";
286  case ARMISD::CMPFPw0:       return "ARMISD::CMPFPw0";
287  case ARMISD::FMSTAT:        return "ARMISD::FMSTAT";
288  case ARMISD::CMOV:          return "ARMISD::CMOV";
289  case ARMISD::CNEG:          return "ARMISD::CNEG";
290
291  case ARMISD::FTOSI:         return "ARMISD::FTOSI";
292  case ARMISD::FTOUI:         return "ARMISD::FTOUI";
293  case ARMISD::SITOF:         return "ARMISD::SITOF";
294  case ARMISD::UITOF:         return "ARMISD::UITOF";
295
296  case ARMISD::SRL_FLAG:      return "ARMISD::SRL_FLAG";
297  case ARMISD::SRA_FLAG:      return "ARMISD::SRA_FLAG";
298  case ARMISD::RRX:           return "ARMISD::RRX";
299
300  case ARMISD::FMRRD:         return "ARMISD::FMRRD";
301  case ARMISD::FMDRR:         return "ARMISD::FMDRR";
302
303  case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
304  }
305}
306
307//===----------------------------------------------------------------------===//
308// Lowering Code
309//===----------------------------------------------------------------------===//
310
311
312/// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
313static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
314  switch (CC) {
315  default: assert(0 && "Unknown condition code!");
316  case ISD::SETNE:  return ARMCC::NE;
317  case ISD::SETEQ:  return ARMCC::EQ;
318  case ISD::SETGT:  return ARMCC::GT;
319  case ISD::SETGE:  return ARMCC::GE;
320  case ISD::SETLT:  return ARMCC::LT;
321  case ISD::SETLE:  return ARMCC::LE;
322  case ISD::SETUGT: return ARMCC::HI;
323  case ISD::SETUGE: return ARMCC::HS;
324  case ISD::SETULT: return ARMCC::LO;
325  case ISD::SETULE: return ARMCC::LS;
326  }
327}
328
329/// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC. It
330/// returns true if the operands should be inverted to form the proper
331/// comparison.
332static bool FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
333                        ARMCC::CondCodes &CondCode2) {
334  bool Invert = false;
335  CondCode2 = ARMCC::AL;
336  switch (CC) {
337  default: assert(0 && "Unknown FP condition!");
338  case ISD::SETEQ:
339  case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
340  case ISD::SETGT:
341  case ISD::SETOGT: CondCode = ARMCC::GT; break;
342  case ISD::SETGE:
343  case ISD::SETOGE: CondCode = ARMCC::GE; break;
344  case ISD::SETOLT: CondCode = ARMCC::MI; break;
345  case ISD::SETOLE: CondCode = ARMCC::GT; Invert = true; break;
346  case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
347  case ISD::SETO:   CondCode = ARMCC::VC; break;
348  case ISD::SETUO:  CondCode = ARMCC::VS; break;
349  case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
350  case ISD::SETUGT: CondCode = ARMCC::HI; break;
351  case ISD::SETUGE: CondCode = ARMCC::PL; break;
352  case ISD::SETLT:
353  case ISD::SETULT: CondCode = ARMCC::LT; break;
354  case ISD::SETLE:
355  case ISD::SETULE: CondCode = ARMCC::LE; break;
356  case ISD::SETNE:
357  case ISD::SETUNE: CondCode = ARMCC::NE; break;
358  }
359  return Invert;
360}
361
362static void
363HowToPassArgument(MVT::ValueType ObjectVT, unsigned NumGPRs,
364                  unsigned StackOffset, unsigned &NeededGPRs,
365                  unsigned &NeededStackSize, unsigned &GPRPad,
366                  unsigned &StackPad, unsigned Flags) {
367  NeededStackSize = 0;
368  NeededGPRs = 0;
369  StackPad = 0;
370  GPRPad = 0;
371  unsigned align = (Flags >> ISD::ParamFlags::OrigAlignmentOffs);
372  GPRPad = NumGPRs % ((align + 3)/4);
373  StackPad = StackOffset % align;
374  unsigned firstGPR = NumGPRs + GPRPad;
375  switch (ObjectVT) {
376  default: assert(0 && "Unhandled argument type!");
377  case MVT::i32:
378  case MVT::f32:
379    if (firstGPR < 4)
380      NeededGPRs = 1;
381    else
382      NeededStackSize = 4;
383    break;
384  case MVT::i64:
385  case MVT::f64:
386    if (firstGPR < 3)
387      NeededGPRs = 2;
388    else if (firstGPR == 3) {
389      NeededGPRs = 1;
390      NeededStackSize = 4;
391    } else
392      NeededStackSize = 8;
393  }
394}
395
396/// LowerCALL - Lowering a ISD::CALL node into a callseq_start <-
397/// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
398/// nodes.
399SDOperand ARMTargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
400  MVT::ValueType RetVT= Op.Val->getValueType(0);
401  SDOperand Chain    = Op.getOperand(0);
402  unsigned CallConv  = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
403  assert((CallConv == CallingConv::C ||
404          CallConv == CallingConv::Fast) && "unknown calling convention");
405  SDOperand Callee   = Op.getOperand(4);
406  unsigned NumOps    = (Op.getNumOperands() - 5) / 2;
407  unsigned ArgOffset = 0;   // Frame mechanisms handle retaddr slot
408  unsigned NumGPRs = 0;     // GPRs used for parameter passing.
409
410  // Count how many bytes are to be pushed on the stack.
411  unsigned NumBytes = 0;
412
413  // Add up all the space actually used.
414  for (unsigned i = 0; i < NumOps; ++i) {
415    unsigned ObjSize;
416    unsigned ObjGPRs;
417    unsigned StackPad;
418    unsigned GPRPad;
419    MVT::ValueType ObjectVT = Op.getOperand(5+2*i).getValueType();
420    unsigned Flags = Op.getConstantOperandVal(5+2*i+1);
421    HowToPassArgument(ObjectVT, NumGPRs, NumBytes, ObjGPRs, ObjSize,
422                      GPRPad, StackPad, Flags);
423    NumBytes += ObjSize + StackPad;
424    NumGPRs += ObjGPRs + GPRPad;
425  }
426
427  // Adjust the stack pointer for the new arguments...
428  // These operations are automatically eliminated by the prolog/epilog pass
429  Chain = DAG.getCALLSEQ_START(Chain,
430                               DAG.getConstant(NumBytes, MVT::i32));
431
432  SDOperand StackPtr = DAG.getRegister(ARM::SP, MVT::i32);
433
434  static const unsigned GPRArgRegs[] = {
435    ARM::R0, ARM::R1, ARM::R2, ARM::R3
436  };
437
438  NumGPRs = 0;
439  std::vector<std::pair<unsigned, SDOperand> > RegsToPass;
440  std::vector<SDOperand> MemOpChains;
441  for (unsigned i = 0; i != NumOps; ++i) {
442    SDOperand Arg = Op.getOperand(5+2*i);
443    unsigned Flags = Op.getConstantOperandVal(5+2*i+1);
444    MVT::ValueType ArgVT = Arg.getValueType();
445
446    unsigned ObjSize;
447    unsigned ObjGPRs;
448    unsigned GPRPad;
449    unsigned StackPad;
450    HowToPassArgument(ArgVT, NumGPRs, ArgOffset, ObjGPRs,
451                      ObjSize, GPRPad, StackPad, Flags);
452    NumGPRs += GPRPad;
453    ArgOffset += StackPad;
454    if (ObjGPRs > 0) {
455      switch (ArgVT) {
456      default: assert(0 && "Unexpected ValueType for argument!");
457      case MVT::i32:
458        RegsToPass.push_back(std::make_pair(GPRArgRegs[NumGPRs], Arg));
459        break;
460      case MVT::f32:
461        RegsToPass.push_back(std::make_pair(GPRArgRegs[NumGPRs],
462                                 DAG.getNode(ISD::BIT_CONVERT, MVT::i32, Arg)));
463        break;
464      case MVT::i64: {
465        SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Arg,
466                                   DAG.getConstant(0, getPointerTy()));
467        SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Arg,
468                                   DAG.getConstant(1, getPointerTy()));
469        RegsToPass.push_back(std::make_pair(GPRArgRegs[NumGPRs], Lo));
470        if (ObjGPRs == 2)
471          RegsToPass.push_back(std::make_pair(GPRArgRegs[NumGPRs+1], Hi));
472        else {
473          SDOperand PtrOff= DAG.getConstant(ArgOffset, StackPtr.getValueType());
474          PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
475          MemOpChains.push_back(DAG.getStore(Chain, Hi, PtrOff, NULL, 0));
476        }
477        break;
478      }
479      case MVT::f64: {
480        SDOperand Cvt = DAG.getNode(ARMISD::FMRRD,
481                                    DAG.getVTList(MVT::i32, MVT::i32),
482                                    &Arg, 1);
483        RegsToPass.push_back(std::make_pair(GPRArgRegs[NumGPRs], Cvt));
484        if (ObjGPRs == 2)
485          RegsToPass.push_back(std::make_pair(GPRArgRegs[NumGPRs+1],
486                                              Cvt.getValue(1)));
487        else {
488          SDOperand PtrOff= DAG.getConstant(ArgOffset, StackPtr.getValueType());
489          PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
490          MemOpChains.push_back(DAG.getStore(Chain, Cvt.getValue(1), PtrOff,
491                                             NULL, 0));
492        }
493        break;
494      }
495      }
496    } else {
497      assert(ObjSize != 0);
498      SDOperand PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
499      PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
500      MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
501    }
502
503    NumGPRs += ObjGPRs;
504    ArgOffset += ObjSize;
505  }
506
507  if (!MemOpChains.empty())
508    Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
509                        &MemOpChains[0], MemOpChains.size());
510
511  // Build a sequence of copy-to-reg nodes chained together with token chain
512  // and flag operands which copy the outgoing args into the appropriate regs.
513  SDOperand InFlag;
514  for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
515    Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
516                             InFlag);
517    InFlag = Chain.getValue(1);
518  }
519
520  // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
521  // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
522  // node so that legalize doesn't hack it.
523  bool isDirect = false;
524  bool isARMFunc = false;
525  bool isLocalARMFunc = false;
526  if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
527    GlobalValue *GV = G->getGlobal();
528    isDirect = true;
529    bool isExt = (GV->isDeclaration() || GV->hasWeakLinkage() ||
530                  GV->hasLinkOnceLinkage());
531    bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
532                   getTargetMachine().getRelocationModel() != Reloc::Static;
533    isARMFunc = !Subtarget->isThumb() || isStub;
534    // ARM call to a local ARM function is predicable.
535    isLocalARMFunc = !Subtarget->isThumb() && !isExt;
536    // tBX takes a register source operand.
537    if (isARMFunc && Subtarget->isThumb() && !Subtarget->hasV5TOps()) {
538      ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMPCLabelIndex,
539                                                           ARMCP::CPStub, 4);
540      SDOperand CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 2);
541      CPAddr = DAG.getNode(ARMISD::Wrapper, MVT::i32, CPAddr);
542      Callee = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), CPAddr, NULL, 0);
543      SDOperand PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
544      Callee = DAG.getNode(ARMISD::PIC_ADD, getPointerTy(), Callee, PICLabel);
545   } else
546      Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
547  } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
548    isDirect = true;
549    bool isStub = Subtarget->isTargetDarwin() &&
550                  getTargetMachine().getRelocationModel() != Reloc::Static;
551    isARMFunc = !Subtarget->isThumb() || isStub;
552    // tBX takes a register source operand.
553    const char *Sym = S->getSymbol();
554    if (isARMFunc && Subtarget->isThumb() && !Subtarget->hasV5TOps()) {
555      ARMConstantPoolValue *CPV = new ARMConstantPoolValue(Sym, ARMPCLabelIndex,
556                                                           ARMCP::CPStub, 4);
557      SDOperand CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 2);
558      CPAddr = DAG.getNode(ARMISD::Wrapper, MVT::i32, CPAddr);
559      Callee = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), CPAddr, NULL, 0);
560      SDOperand PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
561      Callee = DAG.getNode(ARMISD::PIC_ADD, getPointerTy(), Callee, PICLabel);
562    } else
563      Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
564  }
565
566  // FIXME: handle tail calls differently.
567  unsigned CallOpc;
568  if (Subtarget->isThumb()) {
569    if (!Subtarget->hasV5TOps() && (!isDirect || isARMFunc))
570      CallOpc = ARMISD::CALL_NOLINK;
571    else
572      CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
573  } else {
574    CallOpc = (isDirect || Subtarget->hasV5TOps())
575      ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
576      : ARMISD::CALL_NOLINK;
577  }
578  if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb()) {
579    // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
580    Chain = DAG.getCopyToReg(Chain, ARM::LR,
581                             DAG.getNode(ISD::UNDEF, MVT::i32), InFlag);
582    InFlag = Chain.getValue(1);
583  }
584
585  std::vector<MVT::ValueType> NodeTys;
586  NodeTys.push_back(MVT::Other);   // Returns a chain
587  NodeTys.push_back(MVT::Flag);    // Returns a flag for retval copy to use.
588
589  std::vector<SDOperand> Ops;
590  Ops.push_back(Chain);
591  Ops.push_back(Callee);
592
593  // Add argument registers to the end of the list so that they are known live
594  // into the call.
595  for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
596    Ops.push_back(DAG.getRegister(RegsToPass[i].first,
597                                  RegsToPass[i].second.getValueType()));
598
599  if (InFlag.Val)
600    Ops.push_back(InFlag);
601  Chain = DAG.getNode(CallOpc, NodeTys, &Ops[0], Ops.size());
602  InFlag = Chain.getValue(1);
603
604  SDOperand CSOps[] = { Chain, DAG.getConstant(NumBytes, MVT::i32), InFlag };
605  Chain = DAG.getNode(ISD::CALLSEQ_END,
606                      DAG.getNodeValueTypes(MVT::Other, MVT::Flag),
607                      ((RetVT != MVT::Other) ? 2 : 1), CSOps, 3);
608  if (RetVT != MVT::Other)
609    InFlag = Chain.getValue(1);
610
611  std::vector<SDOperand> ResultVals;
612  NodeTys.clear();
613
614  // If the call has results, copy the values out of the ret val registers.
615  switch (RetVT) {
616  default: assert(0 && "Unexpected ret value!");
617  case MVT::Other:
618    break;
619  case MVT::i32:
620    Chain = DAG.getCopyFromReg(Chain, ARM::R0, MVT::i32, InFlag).getValue(1);
621    ResultVals.push_back(Chain.getValue(0));
622    if (Op.Val->getValueType(1) == MVT::i32) {
623      // Returns a i64 value.
624      Chain = DAG.getCopyFromReg(Chain, ARM::R1, MVT::i32,
625                                 Chain.getValue(2)).getValue(1);
626      ResultVals.push_back(Chain.getValue(0));
627      NodeTys.push_back(MVT::i32);
628    }
629    NodeTys.push_back(MVT::i32);
630    break;
631  case MVT::f32:
632    Chain = DAG.getCopyFromReg(Chain, ARM::R0, MVT::i32, InFlag).getValue(1);
633    ResultVals.push_back(DAG.getNode(ISD::BIT_CONVERT, MVT::f32,
634                                     Chain.getValue(0)));
635    NodeTys.push_back(MVT::f32);
636    break;
637  case MVT::f64: {
638    SDOperand Lo = DAG.getCopyFromReg(Chain, ARM::R0, MVT::i32, InFlag);
639    SDOperand Hi = DAG.getCopyFromReg(Lo, ARM::R1, MVT::i32, Lo.getValue(2));
640    ResultVals.push_back(DAG.getNode(ARMISD::FMDRR, MVT::f64, Lo, Hi));
641    NodeTys.push_back(MVT::f64);
642    break;
643  }
644  }
645
646  NodeTys.push_back(MVT::Other);
647
648  if (ResultVals.empty())
649    return Chain;
650
651  ResultVals.push_back(Chain);
652  SDOperand Res = DAG.getNode(ISD::MERGE_VALUES, NodeTys, &ResultVals[0],
653                              ResultVals.size());
654  return Res.getValue(Op.ResNo);
655}
656
657static SDOperand LowerRET(SDOperand Op, SelectionDAG &DAG) {
658  SDOperand Copy;
659  SDOperand Chain = Op.getOperand(0);
660  switch(Op.getNumOperands()) {
661  default:
662    assert(0 && "Do not know how to return this many arguments!");
663    abort();
664  case 1: {
665    SDOperand LR = DAG.getRegister(ARM::LR, MVT::i32);
666    return DAG.getNode(ARMISD::RET_FLAG, MVT::Other, Chain);
667  }
668  case 3:
669    Op = Op.getOperand(1);
670    if (Op.getValueType() == MVT::f32) {
671      Op = DAG.getNode(ISD::BIT_CONVERT, MVT::i32, Op);
672    } else if (Op.getValueType() == MVT::f64) {
673      // Legalize ret f64 -> ret 2 x i32.  We always have fmrrd if f64 is
674      // available.
675      Op = DAG.getNode(ARMISD::FMRRD, DAG.getVTList(MVT::i32, MVT::i32), &Op,1);
676      SDOperand Sign = DAG.getConstant(0, MVT::i32);
677      return DAG.getNode(ISD::RET, MVT::Other, Chain, Op, Sign,
678                         Op.getValue(1), Sign);
679    }
680    Copy = DAG.getCopyToReg(Chain, ARM::R0, Op, SDOperand());
681    if (DAG.getMachineFunction().liveout_empty())
682      DAG.getMachineFunction().addLiveOut(ARM::R0);
683    break;
684  case 5:
685    Copy = DAG.getCopyToReg(Chain, ARM::R1, Op.getOperand(3), SDOperand());
686    Copy = DAG.getCopyToReg(Copy, ARM::R0, Op.getOperand(1), Copy.getValue(1));
687    // If we haven't noted the R0+R1 are live out, do so now.
688    if (DAG.getMachineFunction().liveout_empty()) {
689      DAG.getMachineFunction().addLiveOut(ARM::R0);
690      DAG.getMachineFunction().addLiveOut(ARM::R1);
691    }
692    break;
693  }
694
695  //We must use RET_FLAG instead of BRIND because BRIND doesn't have a flag
696  return DAG.getNode(ARMISD::RET_FLAG, MVT::Other, Copy, Copy.getValue(1));
697}
698
699// ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
700// their target countpart wrapped in the ARMISD::Wrapper node. Suppose N is
701// one of the above mentioned nodes. It has to be wrapped because otherwise
702// Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
703// be used to form addressing mode. These wrapped nodes will be selected
704// into MOVi.
705static SDOperand LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
706  MVT::ValueType PtrVT = Op.getValueType();
707  ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
708  SDOperand Res;
709  if (CP->isMachineConstantPoolEntry())
710    Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
711                                    CP->getAlignment());
712  else
713    Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
714                                    CP->getAlignment());
715  return DAG.getNode(ARMISD::Wrapper, MVT::i32, Res);
716}
717
718// Lower ISD::GlobalTLSAddress using the "general dynamic" model
719SDOperand
720ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
721                                                 SelectionDAG &DAG) {
722  MVT::ValueType PtrVT = getPointerTy();
723  unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
724  ARMConstantPoolValue *CPV =
725    new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex, ARMCP::CPValue,
726                             PCAdj, "tlsgd", true);
727  SDOperand Argument = DAG.getTargetConstantPool(CPV, PtrVT, 2);
728  Argument = DAG.getNode(ARMISD::Wrapper, MVT::i32, Argument);
729  Argument = DAG.getLoad(PtrVT, DAG.getEntryNode(), Argument, NULL, 0);
730  SDOperand Chain = Argument.getValue(1);
731
732  SDOperand PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
733  Argument = DAG.getNode(ARMISD::PIC_ADD, PtrVT, Argument, PICLabel);
734
735  // call __tls_get_addr.
736  ArgListTy Args;
737  ArgListEntry Entry;
738  Entry.Node = Argument;
739  Entry.Ty = (const Type *) Type::Int32Ty;
740  Args.push_back(Entry);
741  std::pair<SDOperand, SDOperand> CallResult =
742    LowerCallTo(Chain, (const Type *) Type::Int32Ty, false, false,
743                CallingConv::C, false,
744                DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG);
745  return CallResult.first;
746}
747
748// Lower ISD::GlobalTLSAddress using the "initial exec" or
749// "local exec" model.
750SDOperand
751ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
752                                            SelectionDAG &DAG) {
753  GlobalValue *GV = GA->getGlobal();
754  SDOperand Offset;
755  SDOperand Chain = DAG.getEntryNode();
756  MVT::ValueType PtrVT = getPointerTy();
757  // Get the Thread Pointer
758  SDOperand ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, PtrVT);
759
760  if (GV->isDeclaration()){
761    // initial exec model
762    unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
763    ARMConstantPoolValue *CPV =
764      new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex, ARMCP::CPValue,
765                               PCAdj, "gottpoff", true);
766    Offset = DAG.getTargetConstantPool(CPV, PtrVT, 2);
767    Offset = DAG.getNode(ARMISD::Wrapper, MVT::i32, Offset);
768    Offset = DAG.getLoad(PtrVT, Chain, Offset, NULL, 0);
769    Chain = Offset.getValue(1);
770
771    SDOperand PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
772    Offset = DAG.getNode(ARMISD::PIC_ADD, PtrVT, Offset, PICLabel);
773
774    Offset = DAG.getLoad(PtrVT, Chain, Offset, NULL, 0);
775  } else {
776    // local exec model
777    ARMConstantPoolValue *CPV =
778      new ARMConstantPoolValue(GV, ARMCP::CPValue, "tpoff");
779    Offset = DAG.getTargetConstantPool(CPV, PtrVT, 2);
780    Offset = DAG.getNode(ARMISD::Wrapper, MVT::i32, Offset);
781    Offset = DAG.getLoad(PtrVT, Chain, Offset, NULL, 0);
782  }
783
784  // The address of the thread local variable is the add of the thread
785  // pointer with the offset of the variable.
786  return DAG.getNode(ISD::ADD, PtrVT, ThreadPointer, Offset);
787}
788
789SDOperand
790ARMTargetLowering::LowerGlobalTLSAddress(SDOperand Op, SelectionDAG &DAG) {
791  // TODO: implement the "local dynamic" model
792  assert(Subtarget->isTargetELF() &&
793         "TLS not implemented for non-ELF targets");
794  GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
795  // If the relocation model is PIC, use the "General Dynamic" TLS Model,
796  // otherwise use the "Local Exec" TLS Model
797  if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
798    return LowerToTLSGeneralDynamicModel(GA, DAG);
799  else
800    return LowerToTLSExecModels(GA, DAG);
801}
802
803SDOperand ARMTargetLowering::LowerGlobalAddressELF(SDOperand Op,
804                                                   SelectionDAG &DAG) {
805  MVT::ValueType PtrVT = getPointerTy();
806  GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
807  Reloc::Model RelocM = getTargetMachine().getRelocationModel();
808  if (RelocM == Reloc::PIC_) {
809    bool UseGOTOFF = GV->hasInternalLinkage() || GV->hasHiddenVisibility();
810    ARMConstantPoolValue *CPV =
811      new ARMConstantPoolValue(GV, ARMCP::CPValue, UseGOTOFF ? "GOTOFF":"GOT");
812    SDOperand CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 2);
813    CPAddr = DAG.getNode(ARMISD::Wrapper, MVT::i32, CPAddr);
814    SDOperand Result = DAG.getLoad(PtrVT, DAG.getEntryNode(), CPAddr, NULL, 0);
815    SDOperand Chain = Result.getValue(1);
816    SDOperand GOT = DAG.getNode(ISD::GLOBAL_OFFSET_TABLE, PtrVT);
817    Result = DAG.getNode(ISD::ADD, PtrVT, Result, GOT);
818    if (!UseGOTOFF)
819      Result = DAG.getLoad(PtrVT, Chain, Result, NULL, 0);
820    return Result;
821  } else {
822    SDOperand CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 2);
823    CPAddr = DAG.getNode(ARMISD::Wrapper, MVT::i32, CPAddr);
824    return DAG.getLoad(PtrVT, DAG.getEntryNode(), CPAddr, NULL, 0);
825  }
826}
827
828/// GVIsIndirectSymbol - true if the GV will be accessed via an indirect symbol
829/// even in non-static mode.
830static bool GVIsIndirectSymbol(GlobalValue *GV, Reloc::Model RelocM) {
831  return RelocM != Reloc::Static &&
832    (GV->hasWeakLinkage() || GV->hasLinkOnceLinkage() ||
833     (GV->isDeclaration() && !GV->hasNotBeenReadFromBitcode()));
834}
835
836SDOperand ARMTargetLowering::LowerGlobalAddressDarwin(SDOperand Op,
837                                                      SelectionDAG &DAG) {
838  MVT::ValueType PtrVT = getPointerTy();
839  GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
840  Reloc::Model RelocM = getTargetMachine().getRelocationModel();
841  bool IsIndirect = GVIsIndirectSymbol(GV, RelocM);
842  SDOperand CPAddr;
843  if (RelocM == Reloc::Static)
844    CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 2);
845  else {
846    unsigned PCAdj = (RelocM != Reloc::PIC_)
847      ? 0 : (Subtarget->isThumb() ? 4 : 8);
848    ARMCP::ARMCPKind Kind = IsIndirect ? ARMCP::CPNonLazyPtr
849      : ARMCP::CPValue;
850    ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMPCLabelIndex,
851                                                         Kind, PCAdj);
852    CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 2);
853  }
854  CPAddr = DAG.getNode(ARMISD::Wrapper, MVT::i32, CPAddr);
855
856  SDOperand Result = DAG.getLoad(PtrVT, DAG.getEntryNode(), CPAddr, NULL, 0);
857  SDOperand Chain = Result.getValue(1);
858
859  if (RelocM == Reloc::PIC_) {
860    SDOperand PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
861    Result = DAG.getNode(ARMISD::PIC_ADD, PtrVT, Result, PICLabel);
862  }
863  if (IsIndirect)
864    Result = DAG.getLoad(PtrVT, Chain, Result, NULL, 0);
865
866  return Result;
867}
868
869SDOperand ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDOperand Op,
870                                                      SelectionDAG &DAG){
871  assert(Subtarget->isTargetELF() &&
872         "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
873  MVT::ValueType PtrVT = getPointerTy();
874  unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
875  ARMConstantPoolValue *CPV = new ARMConstantPoolValue("_GLOBAL_OFFSET_TABLE_",
876                                                       ARMPCLabelIndex,
877                                                       ARMCP::CPValue, PCAdj);
878  SDOperand CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 2);
879  CPAddr = DAG.getNode(ARMISD::Wrapper, MVT::i32, CPAddr);
880  SDOperand Result = DAG.getLoad(PtrVT, DAG.getEntryNode(), CPAddr, NULL, 0);
881  SDOperand PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
882  return DAG.getNode(ARMISD::PIC_ADD, PtrVT, Result, PICLabel);
883}
884
885static SDOperand LowerVASTART(SDOperand Op, SelectionDAG &DAG,
886                              unsigned VarArgsFrameIndex) {
887  // vastart just stores the address of the VarArgsFrameIndex slot into the
888  // memory location argument.
889  MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
890  SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
891  SrcValueSDNode *SV = cast<SrcValueSDNode>(Op.getOperand(2));
892  return DAG.getStore(Op.getOperand(0), FR, Op.getOperand(1), SV->getValue(),
893                      SV->getOffset());
894}
895
896static SDOperand LowerFORMAL_ARGUMENT(SDOperand Op, SelectionDAG &DAG,
897                                      unsigned *vRegs, unsigned ArgNo,
898                                      unsigned &NumGPRs, unsigned &ArgOffset) {
899  MachineFunction &MF = DAG.getMachineFunction();
900  MVT::ValueType ObjectVT = Op.getValue(ArgNo).getValueType();
901  SDOperand Root = Op.getOperand(0);
902  std::vector<SDOperand> ArgValues;
903  SSARegMap *RegMap = MF.getSSARegMap();
904
905  static const unsigned GPRArgRegs[] = {
906    ARM::R0, ARM::R1, ARM::R2, ARM::R3
907  };
908
909  unsigned ObjSize;
910  unsigned ObjGPRs;
911  unsigned GPRPad;
912  unsigned StackPad;
913  unsigned Flags = Op.getConstantOperandVal(ArgNo + 3);
914  HowToPassArgument(ObjectVT, NumGPRs, ArgOffset, ObjGPRs,
915                    ObjSize, GPRPad, StackPad, Flags);
916  NumGPRs += GPRPad;
917  ArgOffset += StackPad;
918
919  SDOperand ArgValue;
920  if (ObjGPRs == 1) {
921    unsigned VReg = RegMap->createVirtualRegister(&ARM::GPRRegClass);
922    MF.addLiveIn(GPRArgRegs[NumGPRs], VReg);
923    vRegs[NumGPRs] = VReg;
924    ArgValue = DAG.getCopyFromReg(Root, VReg, MVT::i32);
925    if (ObjectVT == MVT::f32)
926      ArgValue = DAG.getNode(ISD::BIT_CONVERT, MVT::f32, ArgValue);
927  } else if (ObjGPRs == 2) {
928    unsigned VReg = RegMap->createVirtualRegister(&ARM::GPRRegClass);
929    MF.addLiveIn(GPRArgRegs[NumGPRs], VReg);
930    vRegs[NumGPRs] = VReg;
931    ArgValue = DAG.getCopyFromReg(Root, VReg, MVT::i32);
932
933    VReg = RegMap->createVirtualRegister(&ARM::GPRRegClass);
934    MF.addLiveIn(GPRArgRegs[NumGPRs+1], VReg);
935    vRegs[NumGPRs+1] = VReg;
936    SDOperand ArgValue2 = DAG.getCopyFromReg(Root, VReg, MVT::i32);
937
938    if (ObjectVT == MVT::i64)
939      ArgValue = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, ArgValue, ArgValue2);
940    else
941      ArgValue = DAG.getNode(ARMISD::FMDRR, MVT::f64, ArgValue, ArgValue2);
942  }
943  NumGPRs += ObjGPRs;
944
945  if (ObjSize) {
946    // If the argument is actually used, emit a load from the right stack
947    // slot.
948    if (!Op.Val->hasNUsesOfValue(0, ArgNo)) {
949      MachineFrameInfo *MFI = MF.getFrameInfo();
950      int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
951      SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
952      if (ObjGPRs == 0)
953        ArgValue = DAG.getLoad(ObjectVT, Root, FIN, NULL, 0);
954      else {
955        SDOperand ArgValue2 =
956          DAG.getLoad(MVT::i32, Root, FIN, NULL, 0);
957        if (ObjectVT == MVT::i64)
958          ArgValue= DAG.getNode(ISD::BUILD_PAIR, MVT::i64, ArgValue, ArgValue2);
959        else
960          ArgValue= DAG.getNode(ARMISD::FMDRR, MVT::f64, ArgValue, ArgValue2);
961      }
962    } else {
963      // Don't emit a dead load.
964      ArgValue = DAG.getNode(ISD::UNDEF, ObjectVT);
965    }
966
967    ArgOffset += ObjSize;   // Move on to the next argument.
968  }
969
970  return ArgValue;
971}
972
973SDOperand
974ARMTargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) {
975  std::vector<SDOperand> ArgValues;
976  SDOperand Root = Op.getOperand(0);
977  unsigned ArgOffset = 0;   // Frame mechanisms handle retaddr slot
978  unsigned NumGPRs = 0;     // GPRs used for parameter passing.
979  unsigned VRegs[4];
980
981  unsigned NumArgs = Op.Val->getNumValues()-1;
982  for (unsigned ArgNo = 0; ArgNo < NumArgs; ++ArgNo)
983    ArgValues.push_back(LowerFORMAL_ARGUMENT(Op, DAG, VRegs, ArgNo,
984                                             NumGPRs, ArgOffset));
985
986  bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
987  if (isVarArg) {
988    static const unsigned GPRArgRegs[] = {
989      ARM::R0, ARM::R1, ARM::R2, ARM::R3
990    };
991
992    MachineFunction &MF = DAG.getMachineFunction();
993    SSARegMap *RegMap = MF.getSSARegMap();
994    MachineFrameInfo *MFI = MF.getFrameInfo();
995    ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
996    unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
997    unsigned VARegSize = (4 - NumGPRs) * 4;
998    unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
999    if (VARegSaveSize) {
1000      // If this function is vararg, store any remaining integer argument regs
1001      // to their spots on the stack so that they may be loaded by deferencing
1002      // the result of va_next.
1003      AFI->setVarArgsRegSaveSize(VARegSaveSize);
1004      VarArgsFrameIndex = MFI->CreateFixedObject(VARegSaveSize, ArgOffset +
1005                                                 VARegSaveSize - VARegSize);
1006      SDOperand FIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
1007
1008      SmallVector<SDOperand, 4> MemOps;
1009      for (; NumGPRs < 4; ++NumGPRs) {
1010        unsigned VReg = RegMap->createVirtualRegister(&ARM::GPRRegClass);
1011        MF.addLiveIn(GPRArgRegs[NumGPRs], VReg);
1012        SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::i32);
1013        SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1014        MemOps.push_back(Store);
1015        FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1016                          DAG.getConstant(4, getPointerTy()));
1017      }
1018      if (!MemOps.empty())
1019        Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1020                           &MemOps[0], MemOps.size());
1021    } else
1022      // This will point to the next argument passed via stack.
1023      VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset);
1024  }
1025
1026  ArgValues.push_back(Root);
1027
1028  // Return the new list of results.
1029  std::vector<MVT::ValueType> RetVT(Op.Val->value_begin(),
1030                                    Op.Val->value_end());
1031  return DAG.getNode(ISD::MERGE_VALUES, RetVT, &ArgValues[0], ArgValues.size());
1032}
1033
1034/// isFloatingPointZero - Return true if this is +0.0.
1035static bool isFloatingPointZero(SDOperand Op) {
1036  if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
1037    return CFP->getValueAPF().isPosZero();
1038  else if (ISD::isEXTLoad(Op.Val) || ISD::isNON_EXTLoad(Op.Val)) {
1039    // Maybe this has already been legalized into the constant pool?
1040    if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
1041      SDOperand WrapperOp = Op.getOperand(1).getOperand(0);
1042      if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
1043        if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
1044          return CFP->getValueAPF().isPosZero();
1045    }
1046  }
1047  return false;
1048}
1049
1050static bool isLegalCmpImmediate(unsigned C, bool isThumb) {
1051  return ( isThumb && (C & ~255U) == 0) ||
1052         (!isThumb && ARM_AM::getSOImmVal(C) != -1);
1053}
1054
1055/// Returns appropriate ARM CMP (cmp) and corresponding condition code for
1056/// the given operands.
1057static SDOperand getARMCmp(SDOperand LHS, SDOperand RHS, ISD::CondCode CC,
1058                           SDOperand &ARMCC, SelectionDAG &DAG, bool isThumb) {
1059  if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.Val)) {
1060    unsigned C = RHSC->getValue();
1061    if (!isLegalCmpImmediate(C, isThumb)) {
1062      // Constant does not fit, try adjusting it by one?
1063      switch (CC) {
1064      default: break;
1065      case ISD::SETLT:
1066      case ISD::SETGE:
1067        if (isLegalCmpImmediate(C-1, isThumb)) {
1068          CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1069          RHS = DAG.getConstant(C-1, MVT::i32);
1070        }
1071        break;
1072      case ISD::SETULT:
1073      case ISD::SETUGE:
1074        if (C > 0 && isLegalCmpImmediate(C-1, isThumb)) {
1075          CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1076          RHS = DAG.getConstant(C-1, MVT::i32);
1077        }
1078        break;
1079      case ISD::SETLE:
1080      case ISD::SETGT:
1081        if (isLegalCmpImmediate(C+1, isThumb)) {
1082          CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1083          RHS = DAG.getConstant(C+1, MVT::i32);
1084        }
1085        break;
1086      case ISD::SETULE:
1087      case ISD::SETUGT:
1088        if (C < 0xffffffff && isLegalCmpImmediate(C+1, isThumb)) {
1089          CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1090          RHS = DAG.getConstant(C+1, MVT::i32);
1091        }
1092        break;
1093      }
1094    }
1095  }
1096
1097  ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
1098  ARMISD::NodeType CompareType;
1099  switch (CondCode) {
1100  default:
1101    CompareType = ARMISD::CMP;
1102    break;
1103  case ARMCC::EQ:
1104  case ARMCC::NE:
1105  case ARMCC::MI:
1106  case ARMCC::PL:
1107    // Uses only N and Z Flags
1108    CompareType = ARMISD::CMPNZ;
1109    break;
1110  }
1111  ARMCC = DAG.getConstant(CondCode, MVT::i32);
1112  return DAG.getNode(CompareType, MVT::Flag, LHS, RHS);
1113}
1114
1115/// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
1116static SDOperand getVFPCmp(SDOperand LHS, SDOperand RHS, SelectionDAG &DAG) {
1117  SDOperand Cmp;
1118  if (!isFloatingPointZero(RHS))
1119    Cmp = DAG.getNode(ARMISD::CMPFP, MVT::Flag, LHS, RHS);
1120  else
1121    Cmp = DAG.getNode(ARMISD::CMPFPw0, MVT::Flag, LHS);
1122  return DAG.getNode(ARMISD::FMSTAT, MVT::Flag, Cmp);
1123}
1124
1125static SDOperand LowerSELECT_CC(SDOperand Op, SelectionDAG &DAG,
1126                                const ARMSubtarget *ST) {
1127  MVT::ValueType VT = Op.getValueType();
1128  SDOperand LHS = Op.getOperand(0);
1129  SDOperand RHS = Op.getOperand(1);
1130  ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1131  SDOperand TrueVal = Op.getOperand(2);
1132  SDOperand FalseVal = Op.getOperand(3);
1133
1134  if (LHS.getValueType() == MVT::i32) {
1135    SDOperand ARMCC;
1136    SDOperand CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1137    SDOperand Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb());
1138    return DAG.getNode(ARMISD::CMOV, VT, FalseVal, TrueVal, ARMCC, CCR, Cmp);
1139  }
1140
1141  ARMCC::CondCodes CondCode, CondCode2;
1142  if (FPCCToARMCC(CC, CondCode, CondCode2))
1143    std::swap(TrueVal, FalseVal);
1144
1145  SDOperand ARMCC = DAG.getConstant(CondCode, MVT::i32);
1146  SDOperand CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1147  SDOperand Cmp = getVFPCmp(LHS, RHS, DAG);
1148  SDOperand Result = DAG.getNode(ARMISD::CMOV, VT, FalseVal, TrueVal,
1149                                 ARMCC, CCR, Cmp);
1150  if (CondCode2 != ARMCC::AL) {
1151    SDOperand ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
1152    // FIXME: Needs another CMP because flag can have but one use.
1153    SDOperand Cmp2 = getVFPCmp(LHS, RHS, DAG);
1154    Result = DAG.getNode(ARMISD::CMOV, VT, Result, TrueVal, ARMCC2, CCR, Cmp2);
1155  }
1156  return Result;
1157}
1158
1159static SDOperand LowerBR_CC(SDOperand Op, SelectionDAG &DAG,
1160                            const ARMSubtarget *ST) {
1161  SDOperand  Chain = Op.getOperand(0);
1162  ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
1163  SDOperand    LHS = Op.getOperand(2);
1164  SDOperand    RHS = Op.getOperand(3);
1165  SDOperand   Dest = Op.getOperand(4);
1166
1167  if (LHS.getValueType() == MVT::i32) {
1168    SDOperand ARMCC;
1169    SDOperand CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1170    SDOperand Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb());
1171    return DAG.getNode(ARMISD::BRCOND, MVT::Other, Chain, Dest, ARMCC, CCR,Cmp);
1172  }
1173
1174  assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
1175  ARMCC::CondCodes CondCode, CondCode2;
1176  if (FPCCToARMCC(CC, CondCode, CondCode2))
1177    // Swap the LHS/RHS of the comparison if needed.
1178    std::swap(LHS, RHS);
1179
1180  SDOperand Cmp = getVFPCmp(LHS, RHS, DAG);
1181  SDOperand ARMCC = DAG.getConstant(CondCode, MVT::i32);
1182  SDOperand CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1183  SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
1184  SDOperand Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
1185  SDOperand Res = DAG.getNode(ARMISD::BRCOND, VTList, Ops, 5);
1186  if (CondCode2 != ARMCC::AL) {
1187    ARMCC = DAG.getConstant(CondCode2, MVT::i32);
1188    SDOperand Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
1189    Res = DAG.getNode(ARMISD::BRCOND, VTList, Ops, 5);
1190  }
1191  return Res;
1192}
1193
1194SDOperand ARMTargetLowering::LowerBR_JT(SDOperand Op, SelectionDAG &DAG) {
1195  SDOperand Chain = Op.getOperand(0);
1196  SDOperand Table = Op.getOperand(1);
1197  SDOperand Index = Op.getOperand(2);
1198
1199  MVT::ValueType PTy = getPointerTy();
1200  JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
1201  ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
1202  SDOperand UId =  DAG.getConstant(AFI->createJumpTableUId(), PTy);
1203  SDOperand JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
1204  Table = DAG.getNode(ARMISD::WrapperJT, MVT::i32, JTI, UId);
1205  Index = DAG.getNode(ISD::MUL, PTy, Index, DAG.getConstant(4, PTy));
1206  SDOperand Addr = DAG.getNode(ISD::ADD, PTy, Index, Table);
1207  bool isPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_;
1208  Addr = DAG.getLoad(isPIC ? (MVT::ValueType)MVT::i32 : PTy,
1209                     Chain, Addr, NULL, 0);
1210  Chain = Addr.getValue(1);
1211  if (isPIC)
1212    Addr = DAG.getNode(ISD::ADD, PTy, Addr, Table);
1213  return DAG.getNode(ARMISD::BR_JT, MVT::Other, Chain, Addr, JTI, UId);
1214}
1215
1216static SDOperand LowerFP_TO_INT(SDOperand Op, SelectionDAG &DAG) {
1217  unsigned Opc =
1218    Op.getOpcode() == ISD::FP_TO_SINT ? ARMISD::FTOSI : ARMISD::FTOUI;
1219  Op = DAG.getNode(Opc, MVT::f32, Op.getOperand(0));
1220  return DAG.getNode(ISD::BIT_CONVERT, MVT::i32, Op);
1221}
1222
1223static SDOperand LowerINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
1224  MVT::ValueType VT = Op.getValueType();
1225  unsigned Opc =
1226    Op.getOpcode() == ISD::SINT_TO_FP ? ARMISD::SITOF : ARMISD::UITOF;
1227
1228  Op = DAG.getNode(ISD::BIT_CONVERT, MVT::f32, Op.getOperand(0));
1229  return DAG.getNode(Opc, VT, Op);
1230}
1231
1232static SDOperand LowerFCOPYSIGN(SDOperand Op, SelectionDAG &DAG) {
1233  // Implement fcopysign with a fabs and a conditional fneg.
1234  SDOperand Tmp0 = Op.getOperand(0);
1235  SDOperand Tmp1 = Op.getOperand(1);
1236  MVT::ValueType VT = Op.getValueType();
1237  MVT::ValueType SrcVT = Tmp1.getValueType();
1238  SDOperand AbsVal = DAG.getNode(ISD::FABS, VT, Tmp0);
1239  SDOperand Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG);
1240  SDOperand ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
1241  SDOperand CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1242  return DAG.getNode(ARMISD::CNEG, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
1243}
1244
1245static SDOperand LowerBIT_CONVERT(SDOperand Op, SelectionDAG &DAG) {
1246  // Turn f64->i64 into FMRRD.
1247  assert(Op.getValueType() == MVT::i64 &&
1248         Op.getOperand(0).getValueType() == MVT::f64);
1249
1250  Op = Op.getOperand(0);
1251  SDOperand Cvt = DAG.getNode(ARMISD::FMRRD, DAG.getVTList(MVT::i32, MVT::i32),
1252                              &Op, 1);
1253
1254  // Merge the pieces into a single i64 value.
1255  return DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Cvt, Cvt.getValue(1));
1256}
1257
1258static SDOperand LowerSRx(SDOperand Op, SelectionDAG &DAG,
1259                          const ARMSubtarget *ST) {
1260  assert(Op.getValueType() == MVT::i64 &&
1261         (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SRA) &&
1262         "Unknown shift to lower!");
1263
1264  // We only lower SRA, SRL of 1 here, all others use generic lowering.
1265  if (!isa<ConstantSDNode>(Op.getOperand(1)) ||
1266      cast<ConstantSDNode>(Op.getOperand(1))->getValue() != 1)
1267    return SDOperand();
1268
1269  // If we are in thumb mode, we don't have RRX.
1270  if (ST->isThumb()) return SDOperand();
1271
1272  // Okay, we have a 64-bit SRA or SRL of 1.  Lower this to an RRX expr.
1273  SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
1274                             DAG.getConstant(0, MVT::i32));
1275  SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
1276                             DAG.getConstant(1, MVT::i32));
1277
1278  // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
1279  // captures the result into a carry flag.
1280  unsigned Opc = Op.getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
1281  Hi = DAG.getNode(Opc, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
1282
1283  // The low part is an ARMISD::RRX operand, which shifts the carry in.
1284  Lo = DAG.getNode(ARMISD::RRX, MVT::i32, Lo, Hi.getValue(1));
1285
1286  // Merge the pieces into a single i64 value.
1287  return DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Lo, Hi);
1288}
1289
1290SDOperand ARMTargetLowering::LowerMEMCPY(SDOperand Op, SelectionDAG &DAG) {
1291  SDOperand ChainOp = Op.getOperand(0);
1292  SDOperand DestOp = Op.getOperand(1);
1293  SDOperand SourceOp = Op.getOperand(2);
1294  SDOperand CountOp = Op.getOperand(3);
1295  SDOperand AlignOp = Op.getOperand(4);
1296  SDOperand AlwaysInlineOp = Op.getOperand(5);
1297
1298  bool AlwaysInline = (bool)cast<ConstantSDNode>(AlwaysInlineOp)->getValue();
1299  unsigned Align = (unsigned)cast<ConstantSDNode>(AlignOp)->getValue();
1300  if (Align == 0) Align = 1;
1301
1302  // If size is unknown, call memcpy.
1303  ConstantSDNode *I = dyn_cast<ConstantSDNode>(CountOp);
1304  if (!I) {
1305    assert(!AlwaysInline && "Cannot inline copy of unknown size");
1306    return LowerMEMCPYCall(ChainOp, DestOp, SourceOp, CountOp, DAG);
1307  }
1308
1309  // If not DWORD aligned or if size is more than threshold, then call memcpy.
1310  // The libc version is likely to be faster for the these cases. It can
1311  // use the address value and run time information about the CPU.
1312  // With glibc 2.6.1 on a core 2, coping an array of 100M longs was 30% faster
1313  unsigned Size = I->getValue();
1314  if (AlwaysInline ||
1315      (Size <= Subtarget->getMaxInlineSizeThreshold() &&
1316       (Align & 3) == 0))
1317    return LowerMEMCPYInline(ChainOp, DestOp, SourceOp, Size, Align, DAG);
1318  return LowerMEMCPYCall(ChainOp, DestOp, SourceOp, CountOp, DAG);
1319}
1320
1321SDOperand ARMTargetLowering::LowerMEMCPYCall(SDOperand Chain,
1322                                             SDOperand Dest,
1323                                             SDOperand Source,
1324                                             SDOperand Count,
1325                                             SelectionDAG &DAG) {
1326  MVT::ValueType IntPtr = getPointerTy();
1327  TargetLowering::ArgListTy Args;
1328  TargetLowering::ArgListEntry Entry;
1329  Entry.Ty = getTargetData()->getIntPtrType();
1330  Entry.Node = Dest; Args.push_back(Entry);
1331  Entry.Node = Source; Args.push_back(Entry);
1332  Entry.Node = Count; Args.push_back(Entry);
1333  std::pair<SDOperand,SDOperand> CallResult =
1334      LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
1335                  DAG.getExternalSymbol("memcpy", IntPtr), Args, DAG);
1336  return CallResult.second;
1337}
1338
1339SDOperand ARMTargetLowering::LowerMEMCPYInline(SDOperand Chain,
1340                                               SDOperand Dest,
1341                                               SDOperand Source,
1342                                               unsigned Size,
1343                                               unsigned Align,
1344                                               SelectionDAG &DAG) {
1345  // Do repeated 4-byte loads and stores. To be improved.
1346  assert((Align & 3) == 0 && "Expected 4-byte aligned addresses!");
1347  unsigned BytesLeft = Size & 3;
1348  unsigned NumMemOps = Size >> 2;
1349  unsigned EmittedNumMemOps = 0;
1350  unsigned SrcOff = 0, DstOff = 0;
1351  MVT::ValueType VT = MVT::i32;
1352  unsigned VTSize = 4;
1353  unsigned i = 0;
1354  const unsigned MAX_LOADS_IN_LDM = 6;
1355  SDOperand TFOps[MAX_LOADS_IN_LDM];
1356  SDOperand Loads[MAX_LOADS_IN_LDM];
1357
1358  // Emit up to MAX_LOADS_IN_LDM loads, then a TokenFactor barrier, then the
1359  // same number of stores.  The loads and stores will get combined into
1360  // ldm/stm later on.
1361  while (EmittedNumMemOps < NumMemOps) {
1362    for (i = 0;
1363         i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
1364      Loads[i] = DAG.getLoad(VT, Chain,
1365                             DAG.getNode(ISD::ADD, MVT::i32, Source,
1366                                         DAG.getConstant(SrcOff, MVT::i32)),
1367                             NULL, 0);
1368      TFOps[i] = Loads[i].getValue(1);
1369      SrcOff += VTSize;
1370    }
1371    Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, &TFOps[0], i);
1372
1373    for (i = 0;
1374         i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
1375      TFOps[i] = DAG.getStore(Chain, Loads[i],
1376                           DAG.getNode(ISD::ADD, MVT::i32, Dest,
1377                                       DAG.getConstant(DstOff, MVT::i32)),
1378                           NULL, 0);
1379      DstOff += VTSize;
1380    }
1381    Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, &TFOps[0], i);
1382
1383    EmittedNumMemOps += i;
1384  }
1385
1386  if (BytesLeft == 0)
1387    return Chain;
1388
1389  // Issue loads / stores for the trailing (1 - 3) bytes.
1390  unsigned BytesLeftSave = BytesLeft;
1391  i = 0;
1392  while (BytesLeft) {
1393    if (BytesLeft >= 2) {
1394      VT = MVT::i16;
1395      VTSize = 2;
1396    } else {
1397      VT = MVT::i8;
1398      VTSize = 1;
1399    }
1400
1401    Loads[i] = DAG.getLoad(VT, Chain,
1402                           DAG.getNode(ISD::ADD, MVT::i32, Source,
1403                                       DAG.getConstant(SrcOff, MVT::i32)),
1404                           NULL, 0);
1405    TFOps[i] = Loads[i].getValue(1);
1406    ++i;
1407    SrcOff += VTSize;
1408    BytesLeft -= VTSize;
1409  }
1410  Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, &TFOps[0], i);
1411
1412  i = 0;
1413  BytesLeft = BytesLeftSave;
1414  while (BytesLeft) {
1415    if (BytesLeft >= 2) {
1416      VT = MVT::i16;
1417      VTSize = 2;
1418    } else {
1419      VT = MVT::i8;
1420      VTSize = 1;
1421    }
1422
1423    TFOps[i] = DAG.getStore(Chain, Loads[i],
1424                            DAG.getNode(ISD::ADD, MVT::i32, Dest,
1425                                        DAG.getConstant(DstOff, MVT::i32)),
1426                            NULL, 0);
1427    ++i;
1428    DstOff += VTSize;
1429    BytesLeft -= VTSize;
1430  }
1431  return DAG.getNode(ISD::TokenFactor, MVT::Other, &TFOps[0], i);
1432}
1433
1434SDOperand ARMTargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
1435  switch (Op.getOpcode()) {
1436  default: assert(0 && "Don't know how to custom lower this!"); abort();
1437  case ISD::ConstantPool:  return LowerConstantPool(Op, DAG);
1438  case ISD::GlobalAddress:
1439    return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
1440      LowerGlobalAddressELF(Op, DAG);
1441  case ISD::GlobalTLSAddress:   return LowerGlobalTLSAddress(Op, DAG);
1442  case ISD::CALL:          return LowerCALL(Op, DAG);
1443  case ISD::RET:           return LowerRET(Op, DAG);
1444  case ISD::SELECT_CC:     return LowerSELECT_CC(Op, DAG, Subtarget);
1445  case ISD::BR_CC:         return LowerBR_CC(Op, DAG, Subtarget);
1446  case ISD::BR_JT:         return LowerBR_JT(Op, DAG);
1447  case ISD::VASTART:       return LowerVASTART(Op, DAG, VarArgsFrameIndex);
1448  case ISD::SINT_TO_FP:
1449  case ISD::UINT_TO_FP:    return LowerINT_TO_FP(Op, DAG);
1450  case ISD::FP_TO_SINT:
1451  case ISD::FP_TO_UINT:    return LowerFP_TO_INT(Op, DAG);
1452  case ISD::FCOPYSIGN:     return LowerFCOPYSIGN(Op, DAG);
1453  case ISD::BIT_CONVERT:   return LowerBIT_CONVERT(Op, DAG);
1454  case ISD::SRL:
1455  case ISD::SRA:           return LowerSRx(Op, DAG, Subtarget);
1456  case ISD::FORMAL_ARGUMENTS:
1457    return LowerFORMAL_ARGUMENTS(Op, DAG);
1458  case ISD::RETURNADDR:    break;
1459  case ISD::FRAMEADDR:     break;
1460  case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
1461  case ISD::MEMCPY:        return LowerMEMCPY(Op, DAG);
1462  }
1463  return SDOperand();
1464}
1465
1466//===----------------------------------------------------------------------===//
1467//                           ARM Scheduler Hooks
1468//===----------------------------------------------------------------------===//
1469
1470MachineBasicBlock *
1471ARMTargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
1472                                           MachineBasicBlock *BB) {
1473  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1474  switch (MI->getOpcode()) {
1475  default: assert(false && "Unexpected instr type to insert");
1476  case ARM::tMOVCCr: {
1477    // To "insert" a SELECT_CC instruction, we actually have to insert the
1478    // diamond control-flow pattern.  The incoming instruction knows the
1479    // destination vreg to set, the condition code register to branch on, the
1480    // true/false values to select between, and a branch opcode to use.
1481    const BasicBlock *LLVM_BB = BB->getBasicBlock();
1482    ilist<MachineBasicBlock>::iterator It = BB;
1483    ++It;
1484
1485    //  thisMBB:
1486    //  ...
1487    //   TrueVal = ...
1488    //   cmpTY ccX, r1, r2
1489    //   bCC copy1MBB
1490    //   fallthrough --> copy0MBB
1491    MachineBasicBlock *thisMBB  = BB;
1492    MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
1493    MachineBasicBlock *sinkMBB  = new MachineBasicBlock(LLVM_BB);
1494    BuildMI(BB, TII->get(ARM::tBcc)).addMBB(sinkMBB)
1495      .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
1496    MachineFunction *F = BB->getParent();
1497    F->getBasicBlockList().insert(It, copy0MBB);
1498    F->getBasicBlockList().insert(It, sinkMBB);
1499    // Update machine-CFG edges by first adding all successors of the current
1500    // block to the new block which will contain the Phi node for the select.
1501    for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
1502        e = BB->succ_end(); i != e; ++i)
1503      sinkMBB->addSuccessor(*i);
1504    // Next, remove all successors of the current block, and add the true
1505    // and fallthrough blocks as its successors.
1506    while(!BB->succ_empty())
1507      BB->removeSuccessor(BB->succ_begin());
1508    BB->addSuccessor(copy0MBB);
1509    BB->addSuccessor(sinkMBB);
1510
1511    //  copy0MBB:
1512    //   %FalseValue = ...
1513    //   # fallthrough to sinkMBB
1514    BB = copy0MBB;
1515
1516    // Update machine-CFG edges
1517    BB->addSuccessor(sinkMBB);
1518
1519    //  sinkMBB:
1520    //   %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
1521    //  ...
1522    BB = sinkMBB;
1523    BuildMI(BB, TII->get(ARM::PHI), MI->getOperand(0).getReg())
1524      .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
1525      .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
1526
1527    delete MI;   // The pseudo instruction is gone now.
1528    return BB;
1529  }
1530  }
1531}
1532
1533//===----------------------------------------------------------------------===//
1534//                           ARM Optimization Hooks
1535//===----------------------------------------------------------------------===//
1536
1537/// isLegalAddressImmediate - Return true if the integer value can be used
1538/// as the offset of the target addressing mode for load / store of the
1539/// given type.
1540static bool isLegalAddressImmediate(int64_t V, MVT::ValueType VT,
1541                                    const ARMSubtarget *Subtarget) {
1542  if (V == 0)
1543    return true;
1544
1545  if (Subtarget->isThumb()) {
1546    if (V < 0)
1547      return false;
1548
1549    unsigned Scale = 1;
1550    switch (VT) {
1551    default: return false;
1552    case MVT::i1:
1553    case MVT::i8:
1554      // Scale == 1;
1555      break;
1556    case MVT::i16:
1557      // Scale == 2;
1558      Scale = 2;
1559      break;
1560    case MVT::i32:
1561      // Scale == 4;
1562      Scale = 4;
1563      break;
1564    }
1565
1566    if ((V & (Scale - 1)) != 0)
1567      return false;
1568    V /= Scale;
1569    return V == V & ((1LL << 5) - 1);
1570  }
1571
1572  if (V < 0)
1573    V = - V;
1574  switch (VT) {
1575  default: return false;
1576  case MVT::i1:
1577  case MVT::i8:
1578  case MVT::i32:
1579    // +- imm12
1580    return V == V & ((1LL << 12) - 1);
1581  case MVT::i16:
1582    // +- imm8
1583    return V == V & ((1LL << 8) - 1);
1584  case MVT::f32:
1585  case MVT::f64:
1586    if (!Subtarget->hasVFP2())
1587      return false;
1588    if ((V & 3) != 0)
1589      return false;
1590    V >>= 2;
1591    return V == V & ((1LL << 8) - 1);
1592  }
1593}
1594
1595/// isLegalAddressingMode - Return true if the addressing mode represented
1596/// by AM is legal for this target, for a load/store of the specified type.
1597bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
1598                                              const Type *Ty) const {
1599  if (!isLegalAddressImmediate(AM.BaseOffs, getValueType(Ty), Subtarget))
1600    return false;
1601
1602  // Can never fold addr of global into load/store.
1603  if (AM.BaseGV)
1604    return false;
1605
1606  switch (AM.Scale) {
1607  case 0:  // no scale reg, must be "r+i" or "r", or "i".
1608    break;
1609  case 1:
1610    if (Subtarget->isThumb())
1611      return false;
1612    // FALL THROUGH.
1613  default:
1614    // ARM doesn't support any R+R*scale+imm addr modes.
1615    if (AM.BaseOffs)
1616      return false;
1617
1618    int Scale = AM.Scale;
1619    switch (getValueType(Ty)) {
1620    default: return false;
1621    case MVT::i1:
1622    case MVT::i8:
1623    case MVT::i32:
1624    case MVT::i64:
1625      // This assumes i64 is legalized to a pair of i32. If not (i.e.
1626      // ldrd / strd are used, then its address mode is same as i16.
1627      // r + r
1628      if (Scale < 0) Scale = -Scale;
1629      if (Scale == 1)
1630        return true;
1631      // r + r << imm
1632      return isPowerOf2_32(Scale & ~1);
1633    case MVT::i16:
1634      // r + r
1635      if (((unsigned)AM.HasBaseReg + Scale) <= 2)
1636        return true;
1637      return false;
1638
1639    case MVT::isVoid:
1640      // Note, we allow "void" uses (basically, uses that aren't loads or
1641      // stores), because arm allows folding a scale into many arithmetic
1642      // operations.  This should be made more precise and revisited later.
1643
1644      // Allow r << imm, but the imm has to be a multiple of two.
1645      if (AM.Scale & 1) return false;
1646      return isPowerOf2_32(AM.Scale);
1647    }
1648    break;
1649  }
1650  return true;
1651}
1652
1653
1654static bool getIndexedAddressParts(SDNode *Ptr, MVT::ValueType VT,
1655                                   bool isSEXTLoad, SDOperand &Base,
1656                                   SDOperand &Offset, bool &isInc,
1657                                   SelectionDAG &DAG) {
1658  if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
1659    return false;
1660
1661  if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
1662    // AddressingMode 3
1663    Base = Ptr->getOperand(0);
1664    if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
1665      int RHSC = (int)RHS->getValue();
1666      if (RHSC < 0 && RHSC > -256) {
1667        isInc = false;
1668        Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
1669        return true;
1670      }
1671    }
1672    isInc = (Ptr->getOpcode() == ISD::ADD);
1673    Offset = Ptr->getOperand(1);
1674    return true;
1675  } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
1676    // AddressingMode 2
1677    if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
1678      int RHSC = (int)RHS->getValue();
1679      if (RHSC < 0 && RHSC > -0x1000) {
1680        isInc = false;
1681        Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
1682        Base = Ptr->getOperand(0);
1683        return true;
1684      }
1685    }
1686
1687    if (Ptr->getOpcode() == ISD::ADD) {
1688      isInc = true;
1689      ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
1690      if (ShOpcVal != ARM_AM::no_shift) {
1691        Base = Ptr->getOperand(1);
1692        Offset = Ptr->getOperand(0);
1693      } else {
1694        Base = Ptr->getOperand(0);
1695        Offset = Ptr->getOperand(1);
1696      }
1697      return true;
1698    }
1699
1700    isInc = (Ptr->getOpcode() == ISD::ADD);
1701    Base = Ptr->getOperand(0);
1702    Offset = Ptr->getOperand(1);
1703    return true;
1704  }
1705
1706  // FIXME: Use FLDM / FSTM to emulate indexed FP load / store.
1707  return false;
1708}
1709
1710/// getPreIndexedAddressParts - returns true by value, base pointer and
1711/// offset pointer and addressing mode by reference if the node's address
1712/// can be legally represented as pre-indexed load / store address.
1713bool
1714ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDOperand &Base,
1715                                             SDOperand &Offset,
1716                                             ISD::MemIndexedMode &AM,
1717                                             SelectionDAG &DAG) {
1718  if (Subtarget->isThumb())
1719    return false;
1720
1721  MVT::ValueType VT;
1722  SDOperand Ptr;
1723  bool isSEXTLoad = false;
1724  if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1725    Ptr = LD->getBasePtr();
1726    VT  = LD->getLoadedVT();
1727    isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
1728  } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1729    Ptr = ST->getBasePtr();
1730    VT  = ST->getStoredVT();
1731  } else
1732    return false;
1733
1734  bool isInc;
1735  bool isLegal = getIndexedAddressParts(Ptr.Val, VT, isSEXTLoad, Base, Offset,
1736                                        isInc, DAG);
1737  if (isLegal) {
1738    AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
1739    return true;
1740  }
1741  return false;
1742}
1743
1744/// getPostIndexedAddressParts - returns true by value, base pointer and
1745/// offset pointer and addressing mode by reference if this node can be
1746/// combined with a load / store to form a post-indexed load / store.
1747bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
1748                                                   SDOperand &Base,
1749                                                   SDOperand &Offset,
1750                                                   ISD::MemIndexedMode &AM,
1751                                                   SelectionDAG &DAG) {
1752  if (Subtarget->isThumb())
1753    return false;
1754
1755  MVT::ValueType VT;
1756  SDOperand Ptr;
1757  bool isSEXTLoad = false;
1758  if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1759    VT  = LD->getLoadedVT();
1760    isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
1761  } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1762    VT  = ST->getStoredVT();
1763  } else
1764    return false;
1765
1766  bool isInc;
1767  bool isLegal = getIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
1768                                        isInc, DAG);
1769  if (isLegal) {
1770    AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
1771    return true;
1772  }
1773  return false;
1774}
1775
1776void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
1777                                                       uint64_t Mask,
1778                                                       uint64_t &KnownZero,
1779                                                       uint64_t &KnownOne,
1780                                                       const SelectionDAG &DAG,
1781                                                       unsigned Depth) const {
1782  KnownZero = 0;
1783  KnownOne = 0;
1784  switch (Op.getOpcode()) {
1785  default: break;
1786  case ARMISD::CMOV: {
1787    // Bits are known zero/one if known on the LHS and RHS.
1788    DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
1789    if (KnownZero == 0 && KnownOne == 0) return;
1790
1791    uint64_t KnownZeroRHS, KnownOneRHS;
1792    DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
1793                          KnownZeroRHS, KnownOneRHS, Depth+1);
1794    KnownZero &= KnownZeroRHS;
1795    KnownOne  &= KnownOneRHS;
1796    return;
1797  }
1798  }
1799}
1800
1801//===----------------------------------------------------------------------===//
1802//                           ARM Inline Assembly Support
1803//===----------------------------------------------------------------------===//
1804
1805/// getConstraintType - Given a constraint letter, return the type of
1806/// constraint it is for this target.
1807ARMTargetLowering::ConstraintType
1808ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
1809  if (Constraint.size() == 1) {
1810    switch (Constraint[0]) {
1811    default:  break;
1812    case 'l': return C_RegisterClass;
1813    case 'w': return C_RegisterClass;
1814    }
1815  }
1816  return TargetLowering::getConstraintType(Constraint);
1817}
1818
1819std::pair<unsigned, const TargetRegisterClass*>
1820ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
1821                                                MVT::ValueType VT) const {
1822  if (Constraint.size() == 1) {
1823    // GCC RS6000 Constraint Letters
1824    switch (Constraint[0]) {
1825    case 'l':
1826    // FIXME: in thumb mode, 'l' is only low-regs.
1827    // FALL THROUGH.
1828    case 'r':
1829      return std::make_pair(0U, ARM::GPRRegisterClass);
1830    case 'w':
1831      if (VT == MVT::f32)
1832        return std::make_pair(0U, ARM::SPRRegisterClass);
1833      if (VT == MVT::f64)
1834        return std::make_pair(0U, ARM::DPRRegisterClass);
1835      break;
1836    }
1837  }
1838  return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
1839}
1840
1841std::vector<unsigned> ARMTargetLowering::
1842getRegClassForInlineAsmConstraint(const std::string &Constraint,
1843                                  MVT::ValueType VT) const {
1844  if (Constraint.size() != 1)
1845    return std::vector<unsigned>();
1846
1847  switch (Constraint[0]) {      // GCC ARM Constraint Letters
1848  default: break;
1849  case 'l':
1850  case 'r':
1851    return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
1852                                 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
1853                                 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
1854                                 ARM::R12, ARM::LR, 0);
1855  case 'w':
1856    if (VT == MVT::f32)
1857      return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
1858                                   ARM::S4, ARM::S5, ARM::S6, ARM::S7,
1859                                   ARM::S8, ARM::S9, ARM::S10, ARM::S11,
1860                                   ARM::S12,ARM::S13,ARM::S14,ARM::S15,
1861                                   ARM::S16,ARM::S17,ARM::S18,ARM::S19,
1862                                   ARM::S20,ARM::S21,ARM::S22,ARM::S23,
1863                                   ARM::S24,ARM::S25,ARM::S26,ARM::S27,
1864                                   ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
1865    if (VT == MVT::f64)
1866      return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
1867                                   ARM::D4, ARM::D5, ARM::D6, ARM::D7,
1868                                   ARM::D8, ARM::D9, ARM::D10,ARM::D11,
1869                                   ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
1870      break;
1871  }
1872
1873  return std::vector<unsigned>();
1874}
1875