1 //===-- SystemZISelLowering.cpp - SystemZ 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 implements the SystemZTargetLowering class.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #define DEBUG_TYPE "systemz-lower"
15 
16 #include "SystemZISelLowering.h"
17 #include "SystemZ.h"
18 #include "SystemZTargetMachine.h"
19 #include "SystemZSubtarget.h"
20 #include "llvm/DerivedTypes.h"
21 #include "llvm/Function.h"
22 #include "llvm/Intrinsics.h"
23 #include "llvm/CallingConv.h"
24 #include "llvm/GlobalVariable.h"
25 #include "llvm/GlobalAlias.h"
26 #include "llvm/CodeGen/CallingConvLower.h"
27 #include "llvm/CodeGen/MachineFrameInfo.h"
28 #include "llvm/CodeGen/MachineFunction.h"
29 #include "llvm/CodeGen/MachineInstrBuilder.h"
30 #include "llvm/CodeGen/MachineRegisterInfo.h"
31 #include "llvm/CodeGen/PseudoSourceValue.h"
32 #include "llvm/CodeGen/SelectionDAGISel.h"
33 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
34 #include "llvm/CodeGen/ValueTypes.h"
35 #include "llvm/Target/TargetOptions.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Support/ErrorHandling.h"
38 #include "llvm/Support/raw_ostream.h"
39 #include "llvm/ADT/VectorExtras.h"
40 using namespace llvm;
41 
42 SystemZTargetLowering::SystemZTargetLowering(SystemZTargetMachine &tm) :
43   TargetLowering(tm, new TargetLoweringObjectFileELF()),
44   Subtarget(*tm.getSubtargetImpl()), TM(tm) {
45 
46   RegInfo = TM.getRegisterInfo();
47 
48   // Set up the register classes.
49   addRegisterClass(MVT::i32,  SystemZ::GR32RegisterClass);
50   addRegisterClass(MVT::i64,  SystemZ::GR64RegisterClass);
51   addRegisterClass(MVT::v2i32,SystemZ::GR64PRegisterClass);
52   addRegisterClass(MVT::v2i64,SystemZ::GR128RegisterClass);
53 
54   if (!UseSoftFloat) {
55     addRegisterClass(MVT::f32, SystemZ::FP32RegisterClass);
56     addRegisterClass(MVT::f64, SystemZ::FP64RegisterClass);
57   }
58 
59   // Compute derived properties from the register classes
60   computeRegisterProperties();
61 
62   // Provide all sorts of operation actions
63   setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
64   setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
65   setLoadExtAction(ISD::EXTLOAD,  MVT::i1, Promote);
66 
67   setLoadExtAction(ISD::SEXTLOAD, MVT::f32, Expand);
68   setLoadExtAction(ISD::ZEXTLOAD, MVT::f32, Expand);
69   setLoadExtAction(ISD::EXTLOAD,  MVT::f32, Expand);
70 
71   setLoadExtAction(ISD::SEXTLOAD, MVT::f64, Expand);
72   setLoadExtAction(ISD::ZEXTLOAD, MVT::f64, Expand);
73   setLoadExtAction(ISD::EXTLOAD,  MVT::f64, Expand);
74 
75   setStackPointerRegisterToSaveRestore(SystemZ::R15D);
76 
77   // TODO: It may be better to default to latency-oriented scheduling, however
78   // LLVM's current latency-oriented scheduler can't handle physreg definitions
79   // such as SystemZ has with PSW, so set this to the register-pressure
80   // scheduler, because it can.
81   setSchedulingPreference(Sched::RegPressure);
82 
83   setBooleanContents(ZeroOrOneBooleanContent);
84 
85   setOperationAction(ISD::BR_JT,            MVT::Other, Expand);
86   setOperationAction(ISD::BRCOND,           MVT::Other, Expand);
87   setOperationAction(ISD::BR_CC,            MVT::i32, Custom);
88   setOperationAction(ISD::BR_CC,            MVT::i64, Custom);
89   setOperationAction(ISD::BR_CC,            MVT::f32, Custom);
90   setOperationAction(ISD::BR_CC,            MVT::f64, Custom);
91   setOperationAction(ISD::ConstantPool,     MVT::i32, Custom);
92   setOperationAction(ISD::ConstantPool,     MVT::i64, Custom);
93   setOperationAction(ISD::GlobalAddress,    MVT::i64, Custom);
94   setOperationAction(ISD::JumpTable,        MVT::i64, Custom);
95   setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
96 
97   setOperationAction(ISD::SDIV,             MVT::i32, Expand);
98   setOperationAction(ISD::UDIV,             MVT::i32, Expand);
99   setOperationAction(ISD::SDIV,             MVT::i64, Expand);
100   setOperationAction(ISD::UDIV,             MVT::i64, Expand);
101   setOperationAction(ISD::SREM,             MVT::i32, Expand);
102   setOperationAction(ISD::UREM,             MVT::i32, Expand);
103   setOperationAction(ISD::SREM,             MVT::i64, Expand);
104   setOperationAction(ISD::UREM,             MVT::i64, Expand);
105 
106   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
107 
108   setOperationAction(ISD::CTPOP,            MVT::i32, Expand);
109   setOperationAction(ISD::CTPOP,            MVT::i64, Expand);
110   setOperationAction(ISD::CTTZ,             MVT::i32, Expand);
111   setOperationAction(ISD::CTTZ,             MVT::i64, Expand);
112   setOperationAction(ISD::CTLZ,             MVT::i32, Promote);
113   setOperationAction(ISD::CTLZ,             MVT::i64, Legal);
114 
115   // FIXME: Can we lower these 2 efficiently?
116   setOperationAction(ISD::SETCC,            MVT::i32, Expand);
117   setOperationAction(ISD::SETCC,            MVT::i64, Expand);
118   setOperationAction(ISD::SETCC,            MVT::f32, Expand);
119   setOperationAction(ISD::SETCC,            MVT::f64, Expand);
120   setOperationAction(ISD::SELECT,           MVT::i32, Expand);
121   setOperationAction(ISD::SELECT,           MVT::i64, Expand);
122   setOperationAction(ISD::SELECT,           MVT::f32, Expand);
123   setOperationAction(ISD::SELECT,           MVT::f64, Expand);
124   setOperationAction(ISD::SELECT_CC,        MVT::i32, Custom);
125   setOperationAction(ISD::SELECT_CC,        MVT::i64, Custom);
126   setOperationAction(ISD::SELECT_CC,        MVT::f32, Custom);
127   setOperationAction(ISD::SELECT_CC,        MVT::f64, Custom);
128 
129   setOperationAction(ISD::MULHS,            MVT::i64, Expand);
130   setOperationAction(ISD::SMUL_LOHI,        MVT::i64, Expand);
131 
132   // FIXME: Can we support these natively?
133   setOperationAction(ISD::UMUL_LOHI,        MVT::i64, Expand);
134   setOperationAction(ISD::SRL_PARTS,        MVT::i64, Expand);
135   setOperationAction(ISD::SHL_PARTS,        MVT::i64, Expand);
136   setOperationAction(ISD::SRA_PARTS,        MVT::i64, Expand);
137 
138   // Lower some FP stuff
139   setOperationAction(ISD::FSIN,             MVT::f32, Expand);
140   setOperationAction(ISD::FSIN,             MVT::f64, Expand);
141   setOperationAction(ISD::FCOS,             MVT::f32, Expand);
142   setOperationAction(ISD::FCOS,             MVT::f64, Expand);
143   setOperationAction(ISD::FREM,             MVT::f32, Expand);
144   setOperationAction(ISD::FREM,             MVT::f64, Expand);
145 
146   // We have only 64-bit bitconverts
147   setOperationAction(ISD::BITCAST,          MVT::f32, Expand);
148   setOperationAction(ISD::BITCAST,          MVT::i32, Expand);
149 
150   setOperationAction(ISD::UINT_TO_FP,       MVT::i32, Expand);
151   setOperationAction(ISD::UINT_TO_FP,       MVT::i64, Expand);
152   setOperationAction(ISD::FP_TO_UINT,       MVT::i32, Expand);
153   setOperationAction(ISD::FP_TO_UINT,       MVT::i64, Expand);
154 
155   setTruncStoreAction(MVT::f64, MVT::f32, Expand);
156 
157   setMinFunctionAlignment(1);
158 }
159 
160 SDValue SystemZTargetLowering::LowerOperation(SDValue Op,
161                                               SelectionDAG &DAG) const {
162   switch (Op.getOpcode()) {
163   case ISD::BR_CC:            return LowerBR_CC(Op, DAG);
164   case ISD::SELECT_CC:        return LowerSELECT_CC(Op, DAG);
165   case ISD::GlobalAddress:    return LowerGlobalAddress(Op, DAG);
166   case ISD::JumpTable:        return LowerJumpTable(Op, DAG);
167   case ISD::ConstantPool:     return LowerConstantPool(Op, DAG);
168   default:
169     llvm_unreachable("Should not custom lower this!");
170     return SDValue();
171   }
172 }
173 
174 bool SystemZTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
175   if (UseSoftFloat || (VT != MVT::f32 && VT != MVT::f64))
176     return false;
177 
178   // +0.0  lzer
179   // +0.0f lzdr
180   // -0.0  lzer + lner
181   // -0.0f lzdr + lndr
182   return Imm.isZero() || Imm.isNegZero();
183 }
184 
185 //===----------------------------------------------------------------------===//
186 //                       SystemZ Inline Assembly Support
187 //===----------------------------------------------------------------------===//
188 
189 /// getConstraintType - Given a constraint letter, return the type of
190 /// constraint it is for this target.
191 TargetLowering::ConstraintType
192 SystemZTargetLowering::getConstraintType(const std::string &Constraint) const {
193   if (Constraint.size() == 1) {
194     switch (Constraint[0]) {
195     case 'r':
196       return C_RegisterClass;
197     default:
198       break;
199     }
200   }
201   return TargetLowering::getConstraintType(Constraint);
202 }
203 
204 std::pair<unsigned, const TargetRegisterClass*>
205 SystemZTargetLowering::
206 getRegForInlineAsmConstraint(const std::string &Constraint,
207                              EVT VT) const {
208   if (Constraint.size() == 1) {
209     // GCC Constraint Letters
210     switch (Constraint[0]) {
211     default: break;
212     case 'r':   // GENERAL_REGS
213       if (VT == MVT::i32)
214         return std::make_pair(0U, SystemZ::GR32RegisterClass);
215       else if (VT == MVT::i128)
216         return std::make_pair(0U, SystemZ::GR128RegisterClass);
217 
218       return std::make_pair(0U, SystemZ::GR64RegisterClass);
219     }
220   }
221 
222   return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
223 }
224 
225 //===----------------------------------------------------------------------===//
226 //                      Calling Convention Implementation
227 //===----------------------------------------------------------------------===//
228 
229 #include "SystemZGenCallingConv.inc"
230 
231 SDValue
232 SystemZTargetLowering::LowerFormalArguments(SDValue Chain,
233                                             CallingConv::ID CallConv,
234                                             bool isVarArg,
235                                             const SmallVectorImpl<ISD::InputArg>
236                                               &Ins,
237                                             DebugLoc dl,
238                                             SelectionDAG &DAG,
239                                             SmallVectorImpl<SDValue> &InVals)
240                                               const {
241 
242   switch (CallConv) {
243   default:
244     llvm_unreachable("Unsupported calling convention");
245   case CallingConv::C:
246   case CallingConv::Fast:
247     return LowerCCCArguments(Chain, CallConv, isVarArg, Ins, dl, DAG, InVals);
248   }
249 }
250 
251 SDValue
252 SystemZTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
253                                  CallingConv::ID CallConv, bool isVarArg,
254                                  bool &isTailCall,
255                                  const SmallVectorImpl<ISD::OutputArg> &Outs,
256                                  const SmallVectorImpl<SDValue> &OutVals,
257                                  const SmallVectorImpl<ISD::InputArg> &Ins,
258                                  DebugLoc dl, SelectionDAG &DAG,
259                                  SmallVectorImpl<SDValue> &InVals) const {
260   // SystemZ target does not yet support tail call optimization.
261   isTailCall = false;
262 
263   switch (CallConv) {
264   default:
265     llvm_unreachable("Unsupported calling convention");
266   case CallingConv::Fast:
267   case CallingConv::C:
268     return LowerCCCCallTo(Chain, Callee, CallConv, isVarArg, isTailCall,
269                           Outs, OutVals, Ins, dl, DAG, InVals);
270   }
271 }
272 
273 /// LowerCCCArguments - transform physical registers into virtual registers and
274 /// generate load operations for arguments places on the stack.
275 // FIXME: struct return stuff
276 // FIXME: varargs
277 SDValue
278 SystemZTargetLowering::LowerCCCArguments(SDValue Chain,
279                                          CallingConv::ID CallConv,
280                                          bool isVarArg,
281                                          const SmallVectorImpl<ISD::InputArg>
282                                            &Ins,
283                                          DebugLoc dl,
284                                          SelectionDAG &DAG,
285                                          SmallVectorImpl<SDValue> &InVals)
286                                            const {
287 
288   MachineFunction &MF = DAG.getMachineFunction();
289   MachineFrameInfo *MFI = MF.getFrameInfo();
290   MachineRegisterInfo &RegInfo = MF.getRegInfo();
291 
292   // Assign locations to all of the incoming arguments.
293   SmallVector<CCValAssign, 16> ArgLocs;
294   CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
295                  ArgLocs, *DAG.getContext());
296   CCInfo.AnalyzeFormalArguments(Ins, CC_SystemZ);
297 
298   if (isVarArg)
299     report_fatal_error("Varargs not supported yet");
300 
301   for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
302     SDValue ArgValue;
303     CCValAssign &VA = ArgLocs[i];
304     EVT LocVT = VA.getLocVT();
305     if (VA.isRegLoc()) {
306       // Arguments passed in registers
307       TargetRegisterClass *RC;
308       switch (LocVT.getSimpleVT().SimpleTy) {
309       default:
310 #ifndef NDEBUG
311         errs() << "LowerFormalArguments Unhandled argument type: "
312              << LocVT.getSimpleVT().SimpleTy
313              << "\n";
314 #endif
315         llvm_unreachable(0);
316       case MVT::i64:
317         RC = SystemZ::GR64RegisterClass;
318         break;
319       case MVT::f32:
320         RC = SystemZ::FP32RegisterClass;
321         break;
322       case MVT::f64:
323         RC = SystemZ::FP64RegisterClass;
324         break;
325       }
326 
327       unsigned VReg = RegInfo.createVirtualRegister(RC);
328       RegInfo.addLiveIn(VA.getLocReg(), VReg);
329       ArgValue = DAG.getCopyFromReg(Chain, dl, VReg, LocVT);
330     } else {
331       // Sanity check
332       assert(VA.isMemLoc());
333 
334       // Create the nodes corresponding to a load from this parameter slot.
335       // Create the frame index object for this incoming parameter...
336       int FI = MFI->CreateFixedObject(LocVT.getSizeInBits()/8,
337                                       VA.getLocMemOffset(), true);
338 
339       // Create the SelectionDAG nodes corresponding to a load
340       // from this parameter
341       SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
342       ArgValue = DAG.getLoad(LocVT, dl, Chain, FIN,
343                              MachinePointerInfo::getFixedStack(FI),
344                              false, false, 0);
345     }
346 
347     // If this is an 8/16/32-bit value, it is really passed promoted to 64
348     // bits. Insert an assert[sz]ext to capture this, then truncate to the
349     // right size.
350     if (VA.getLocInfo() == CCValAssign::SExt)
351       ArgValue = DAG.getNode(ISD::AssertSext, dl, LocVT, ArgValue,
352                              DAG.getValueType(VA.getValVT()));
353     else if (VA.getLocInfo() == CCValAssign::ZExt)
354       ArgValue = DAG.getNode(ISD::AssertZext, dl, LocVT, ArgValue,
355                              DAG.getValueType(VA.getValVT()));
356 
357     if (VA.getLocInfo() != CCValAssign::Full)
358       ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
359 
360     InVals.push_back(ArgValue);
361   }
362 
363   return Chain;
364 }
365 
366 /// LowerCCCCallTo - functions arguments are copied from virtual regs to
367 /// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted.
368 /// TODO: sret.
369 SDValue
370 SystemZTargetLowering::LowerCCCCallTo(SDValue Chain, SDValue Callee,
371                                       CallingConv::ID CallConv, bool isVarArg,
372                                       bool isTailCall,
373                                       const SmallVectorImpl<ISD::OutputArg>
374                                         &Outs,
375                                       const SmallVectorImpl<SDValue> &OutVals,
376                                       const SmallVectorImpl<ISD::InputArg> &Ins,
377                                       DebugLoc dl, SelectionDAG &DAG,
378                                       SmallVectorImpl<SDValue> &InVals) const {
379   MachineFunction &MF = DAG.getMachineFunction();
380   const TargetFrameLowering *TFI = TM.getFrameLowering();
381 
382   // Offset to first argument stack slot.
383   const unsigned FirstArgOffset = 160;
384 
385   // Analyze operands of the call, assigning locations to each operand.
386   SmallVector<CCValAssign, 16> ArgLocs;
387   CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
388                  ArgLocs, *DAG.getContext());
389 
390   CCInfo.AnalyzeCallOperands(Outs, CC_SystemZ);
391 
392   // Get a count of how many bytes are to be pushed on the stack.
393   unsigned NumBytes = CCInfo.getNextStackOffset();
394 
395   Chain = DAG.getCALLSEQ_START(Chain ,DAG.getConstant(NumBytes,
396                                                       getPointerTy(), true));
397 
398   SmallVector<std::pair<unsigned, SDValue>, 4> RegsToPass;
399   SmallVector<SDValue, 12> MemOpChains;
400   SDValue StackPtr;
401 
402   // Walk the register/memloc assignments, inserting copies/loads.
403   for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
404     CCValAssign &VA = ArgLocs[i];
405 
406     SDValue Arg = OutVals[i];
407 
408     // Promote the value if needed.
409     switch (VA.getLocInfo()) {
410       default: assert(0 && "Unknown loc info!");
411       case CCValAssign::Full: break;
412       case CCValAssign::SExt:
413         Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
414         break;
415       case CCValAssign::ZExt:
416         Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
417         break;
418       case CCValAssign::AExt:
419         Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
420         break;
421     }
422 
423     // Arguments that can be passed on register must be kept at RegsToPass
424     // vector
425     if (VA.isRegLoc()) {
426       RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
427     } else {
428       assert(VA.isMemLoc());
429 
430       if (StackPtr.getNode() == 0)
431         StackPtr =
432           DAG.getCopyFromReg(Chain, dl,
433                              (TFI->hasFP(MF) ?
434                               SystemZ::R11D : SystemZ::R15D),
435                              getPointerTy());
436 
437       unsigned Offset = FirstArgOffset + VA.getLocMemOffset();
438       SDValue PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(),
439                                    StackPtr,
440                                    DAG.getIntPtrConstant(Offset));
441 
442       MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
443                                          MachinePointerInfo(),
444                                          false, false, 0));
445     }
446   }
447 
448   // Transform all store nodes into one single node because all store nodes are
449   // independent of each other.
450   if (!MemOpChains.empty())
451     Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
452                         &MemOpChains[0], MemOpChains.size());
453 
454   // Build a sequence of copy-to-reg nodes chained together with token chain and
455   // flag operands which copy the outgoing args into registers.  The InFlag in
456   // necessary since all emitted instructions must be stuck together.
457   SDValue InFlag;
458   for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
459     Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
460                              RegsToPass[i].second, InFlag);
461     InFlag = Chain.getValue(1);
462   }
463 
464   // If the callee is a GlobalAddress node (quite common, every direct call is)
465   // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
466   // Likewise ExternalSymbol -> TargetExternalSymbol.
467   if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
468     Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl, getPointerTy());
469   else if (ExternalSymbolSDNode *E = dyn_cast<ExternalSymbolSDNode>(Callee))
470     Callee = DAG.getTargetExternalSymbol(E->getSymbol(), getPointerTy());
471 
472   // Returns a chain & a flag for retval copy to use.
473   SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
474   SmallVector<SDValue, 8> Ops;
475   Ops.push_back(Chain);
476   Ops.push_back(Callee);
477 
478   // Add argument registers to the end of the list so that they are
479   // known live into the call.
480   for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
481     Ops.push_back(DAG.getRegister(RegsToPass[i].first,
482                                   RegsToPass[i].second.getValueType()));
483 
484   if (InFlag.getNode())
485     Ops.push_back(InFlag);
486 
487   Chain = DAG.getNode(SystemZISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
488   InFlag = Chain.getValue(1);
489 
490   // Create the CALLSEQ_END node.
491   Chain = DAG.getCALLSEQ_END(Chain,
492                              DAG.getConstant(NumBytes, getPointerTy(), true),
493                              DAG.getConstant(0, getPointerTy(), true),
494                              InFlag);
495   InFlag = Chain.getValue(1);
496 
497   // Handle result values, copying them out of physregs into vregs that we
498   // return.
499   return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl,
500                          DAG, InVals);
501 }
502 
503 /// LowerCallResult - Lower the result values of a call into the
504 /// appropriate copies out of appropriate physical registers.
505 ///
506 SDValue
507 SystemZTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
508                                        CallingConv::ID CallConv, bool isVarArg,
509                                        const SmallVectorImpl<ISD::InputArg>
510                                          &Ins,
511                                        DebugLoc dl, SelectionDAG &DAG,
512                                        SmallVectorImpl<SDValue> &InVals) const {
513 
514   // Assign locations to each value returned by this call.
515   SmallVector<CCValAssign, 16> RVLocs;
516   CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
517                  *DAG.getContext());
518 
519   CCInfo.AnalyzeCallResult(Ins, RetCC_SystemZ);
520 
521   // Copy all of the result registers out of their specified physreg.
522   for (unsigned i = 0; i != RVLocs.size(); ++i) {
523     CCValAssign &VA = RVLocs[i];
524 
525     Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
526                                VA.getLocVT(), InFlag).getValue(1);
527     SDValue RetValue = Chain.getValue(0);
528     InFlag = Chain.getValue(2);
529 
530     // If this is an 8/16/32-bit value, it is really passed promoted to 64
531     // bits. Insert an assert[sz]ext to capture this, then truncate to the
532     // right size.
533     if (VA.getLocInfo() == CCValAssign::SExt)
534       RetValue = DAG.getNode(ISD::AssertSext, dl, VA.getLocVT(), RetValue,
535                              DAG.getValueType(VA.getValVT()));
536     else if (VA.getLocInfo() == CCValAssign::ZExt)
537       RetValue = DAG.getNode(ISD::AssertZext, dl, VA.getLocVT(), RetValue,
538                              DAG.getValueType(VA.getValVT()));
539 
540     if (VA.getLocInfo() != CCValAssign::Full)
541       RetValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), RetValue);
542 
543     InVals.push_back(RetValue);
544   }
545 
546   return Chain;
547 }
548 
549 
550 SDValue
551 SystemZTargetLowering::LowerReturn(SDValue Chain,
552                                    CallingConv::ID CallConv, bool isVarArg,
553                                    const SmallVectorImpl<ISD::OutputArg> &Outs,
554                                    const SmallVectorImpl<SDValue> &OutVals,
555                                    DebugLoc dl, SelectionDAG &DAG) const {
556 
557   // CCValAssign - represent the assignment of the return value to a location
558   SmallVector<CCValAssign, 16> RVLocs;
559 
560   // CCState - Info about the registers and stack slot.
561   CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
562                  RVLocs, *DAG.getContext());
563 
564   // Analize return values.
565   CCInfo.AnalyzeReturn(Outs, RetCC_SystemZ);
566 
567   // If this is the first return lowered for this function, add the regs to the
568   // liveout set for the function.
569   if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
570     for (unsigned i = 0; i != RVLocs.size(); ++i)
571       if (RVLocs[i].isRegLoc())
572         DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
573   }
574 
575   SDValue Flag;
576 
577   // Copy the result values into the output registers.
578   for (unsigned i = 0; i != RVLocs.size(); ++i) {
579     CCValAssign &VA = RVLocs[i];
580     SDValue ResValue = OutVals[i];
581     assert(VA.isRegLoc() && "Can only return in registers!");
582 
583     // If this is an 8/16/32-bit value, it is really should be passed promoted
584     // to 64 bits.
585     if (VA.getLocInfo() == CCValAssign::SExt)
586       ResValue = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ResValue);
587     else if (VA.getLocInfo() == CCValAssign::ZExt)
588       ResValue = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ResValue);
589     else if (VA.getLocInfo() == CCValAssign::AExt)
590       ResValue = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ResValue);
591 
592     Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ResValue, Flag);
593 
594     // Guarantee that all emitted copies are stuck together,
595     // avoiding something bad.
596     Flag = Chain.getValue(1);
597   }
598 
599   if (Flag.getNode())
600     return DAG.getNode(SystemZISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
601 
602   // Return Void
603   return DAG.getNode(SystemZISD::RET_FLAG, dl, MVT::Other, Chain);
604 }
605 
606 SDValue SystemZTargetLowering::EmitCmp(SDValue LHS, SDValue RHS,
607                                        ISD::CondCode CC, SDValue &SystemZCC,
608                                        SelectionDAG &DAG) const {
609   // FIXME: Emit a test if RHS is zero
610 
611   bool isUnsigned = false;
612   SystemZCC::CondCodes TCC;
613   switch (CC) {
614   default:
615     llvm_unreachable("Invalid integer condition!");
616   case ISD::SETEQ:
617   case ISD::SETOEQ:
618     TCC = SystemZCC::E;
619     break;
620   case ISD::SETUEQ:
621     TCC = SystemZCC::NLH;
622     break;
623   case ISD::SETNE:
624   case ISD::SETONE:
625     TCC = SystemZCC::NE;
626     break;
627   case ISD::SETUNE:
628     TCC = SystemZCC::LH;
629     break;
630   case ISD::SETO:
631     TCC = SystemZCC::O;
632     break;
633   case ISD::SETUO:
634     TCC = SystemZCC::NO;
635     break;
636   case ISD::SETULE:
637     if (LHS.getValueType().isFloatingPoint()) {
638       TCC = SystemZCC::NH;
639       break;
640     }
641     isUnsigned = true;   // FALLTHROUGH
642   case ISD::SETLE:
643   case ISD::SETOLE:
644     TCC = SystemZCC::LE;
645     break;
646   case ISD::SETUGE:
647     if (LHS.getValueType().isFloatingPoint()) {
648       TCC = SystemZCC::NL;
649       break;
650     }
651     isUnsigned = true;   // FALLTHROUGH
652   case ISD::SETGE:
653   case ISD::SETOGE:
654     TCC = SystemZCC::HE;
655     break;
656   case ISD::SETUGT:
657     if (LHS.getValueType().isFloatingPoint()) {
658       TCC = SystemZCC::NLE;
659       break;
660     }
661     isUnsigned = true;  // FALLTHROUGH
662   case ISD::SETGT:
663   case ISD::SETOGT:
664     TCC = SystemZCC::H;
665     break;
666   case ISD::SETULT:
667     if (LHS.getValueType().isFloatingPoint()) {
668       TCC = SystemZCC::NHE;
669       break;
670     }
671     isUnsigned = true;  // FALLTHROUGH
672   case ISD::SETLT:
673   case ISD::SETOLT:
674     TCC = SystemZCC::L;
675     break;
676   }
677 
678   SystemZCC = DAG.getConstant(TCC, MVT::i32);
679 
680   DebugLoc dl = LHS.getDebugLoc();
681   return DAG.getNode((isUnsigned ? SystemZISD::UCMP : SystemZISD::CMP),
682                      dl, MVT::i64, LHS, RHS);
683 }
684 
685 
686 SDValue SystemZTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
687   SDValue Chain = Op.getOperand(0);
688   ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
689   SDValue LHS   = Op.getOperand(2);
690   SDValue RHS   = Op.getOperand(3);
691   SDValue Dest  = Op.getOperand(4);
692   DebugLoc dl   = Op.getDebugLoc();
693 
694   SDValue SystemZCC;
695   SDValue Flag = EmitCmp(LHS, RHS, CC, SystemZCC, DAG);
696   return DAG.getNode(SystemZISD::BRCOND, dl, Op.getValueType(),
697                      Chain, Dest, SystemZCC, Flag);
698 }
699 
700 SDValue SystemZTargetLowering::LowerSELECT_CC(SDValue Op,
701                                               SelectionDAG &DAG) const {
702   SDValue LHS    = Op.getOperand(0);
703   SDValue RHS    = Op.getOperand(1);
704   SDValue TrueV  = Op.getOperand(2);
705   SDValue FalseV = Op.getOperand(3);
706   ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
707   DebugLoc dl   = Op.getDebugLoc();
708 
709   SDValue SystemZCC;
710   SDValue Flag = EmitCmp(LHS, RHS, CC, SystemZCC, DAG);
711 
712   SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
713   SmallVector<SDValue, 4> Ops;
714   Ops.push_back(TrueV);
715   Ops.push_back(FalseV);
716   Ops.push_back(SystemZCC);
717   Ops.push_back(Flag);
718 
719   return DAG.getNode(SystemZISD::SELECT, dl, VTs, &Ops[0], Ops.size());
720 }
721 
722 SDValue SystemZTargetLowering::LowerGlobalAddress(SDValue Op,
723                                                   SelectionDAG &DAG) const {
724   DebugLoc dl = Op.getDebugLoc();
725   const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
726   int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
727 
728   bool IsPic = getTargetMachine().getRelocationModel() == Reloc::PIC_;
729   bool ExtraLoadRequired =
730     Subtarget.GVRequiresExtraLoad(GV, getTargetMachine(), false);
731 
732   SDValue Result;
733   if (!IsPic && !ExtraLoadRequired) {
734     Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
735     Offset = 0;
736   } else {
737     unsigned char OpFlags = 0;
738     if (ExtraLoadRequired)
739       OpFlags = SystemZII::MO_GOTENT;
740 
741     Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
742   }
743 
744   Result = DAG.getNode(SystemZISD::PCRelativeWrapper, dl,
745                        getPointerTy(), Result);
746 
747   if (ExtraLoadRequired)
748     Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
749                          MachinePointerInfo::getGOT(), false, false, 0);
750 
751   // If there was a non-zero offset that we didn't fold, create an explicit
752   // addition for it.
753   if (Offset != 0)
754     Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
755                          DAG.getConstant(Offset, getPointerTy()));
756 
757   return Result;
758 }
759 
760 // FIXME: PIC here
761 SDValue SystemZTargetLowering::LowerJumpTable(SDValue Op,
762                                               SelectionDAG &DAG) const {
763   DebugLoc dl = Op.getDebugLoc();
764   JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
765   SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
766 
767   return DAG.getNode(SystemZISD::PCRelativeWrapper, dl, getPointerTy(), Result);
768 }
769 
770 
771 // FIXME: PIC here
772 // FIXME: This is just dirty hack. We need to lower cpool properly
773 SDValue SystemZTargetLowering::LowerConstantPool(SDValue Op,
774                                                  SelectionDAG &DAG) const {
775   DebugLoc dl = Op.getDebugLoc();
776   ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
777 
778   SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
779                                              CP->getAlignment(),
780                                              CP->getOffset());
781 
782   return DAG.getNode(SystemZISD::PCRelativeWrapper, dl, getPointerTy(), Result);
783 }
784 
785 const char *SystemZTargetLowering::getTargetNodeName(unsigned Opcode) const {
786   switch (Opcode) {
787   case SystemZISD::RET_FLAG:           return "SystemZISD::RET_FLAG";
788   case SystemZISD::CALL:               return "SystemZISD::CALL";
789   case SystemZISD::BRCOND:             return "SystemZISD::BRCOND";
790   case SystemZISD::CMP:                return "SystemZISD::CMP";
791   case SystemZISD::UCMP:               return "SystemZISD::UCMP";
792   case SystemZISD::SELECT:             return "SystemZISD::SELECT";
793   case SystemZISD::PCRelativeWrapper:  return "SystemZISD::PCRelativeWrapper";
794   default: return NULL;
795   }
796 }
797 
798 //===----------------------------------------------------------------------===//
799 //  Other Lowering Code
800 //===----------------------------------------------------------------------===//
801 
802 MachineBasicBlock*
803 SystemZTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
804                                                    MachineBasicBlock *BB) const {
805   const SystemZInstrInfo &TII = *TM.getInstrInfo();
806   DebugLoc dl = MI->getDebugLoc();
807   assert((MI->getOpcode() == SystemZ::Select32  ||
808           MI->getOpcode() == SystemZ::SelectF32 ||
809           MI->getOpcode() == SystemZ::Select64  ||
810           MI->getOpcode() == SystemZ::SelectF64) &&
811          "Unexpected instr type to insert");
812 
813   // To "insert" a SELECT instruction, we actually have to insert the diamond
814   // control-flow pattern.  The incoming instruction knows the destination vreg
815   // to set, the condition code register to branch on, the true/false values to
816   // select between, and a branch opcode to use.
817   const BasicBlock *LLVM_BB = BB->getBasicBlock();
818   MachineFunction::iterator I = BB;
819   ++I;
820 
821   //  thisMBB:
822   //  ...
823   //   TrueVal = ...
824   //   cmpTY ccX, r1, r2
825   //   jCC copy1MBB
826   //   fallthrough --> copy0MBB
827   MachineBasicBlock *thisMBB = BB;
828   MachineFunction *F = BB->getParent();
829   MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
830   MachineBasicBlock *copy1MBB = F->CreateMachineBasicBlock(LLVM_BB);
831   SystemZCC::CondCodes CC = (SystemZCC::CondCodes)MI->getOperand(3).getImm();
832   F->insert(I, copy0MBB);
833   F->insert(I, copy1MBB);
834   // Update machine-CFG edges by transferring all successors of the current
835   // block to the new block which will contain the Phi node for the select.
836   copy1MBB->splice(copy1MBB->begin(), BB,
837                    llvm::next(MachineBasicBlock::iterator(MI)),
838                    BB->end());
839   copy1MBB->transferSuccessorsAndUpdatePHIs(BB);
840   // Next, add the true and fallthrough blocks as its successors.
841   BB->addSuccessor(copy0MBB);
842   BB->addSuccessor(copy1MBB);
843 
844   BuildMI(BB, dl, TII.getBrCond(CC)).addMBB(copy1MBB);
845 
846   //  copy0MBB:
847   //   %FalseValue = ...
848   //   # fallthrough to copy1MBB
849   BB = copy0MBB;
850 
851   // Update machine-CFG edges
852   BB->addSuccessor(copy1MBB);
853 
854   //  copy1MBB:
855   //   %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
856   //  ...
857   BB = copy1MBB;
858   BuildMI(*BB, BB->begin(), dl, TII.get(SystemZ::PHI),
859           MI->getOperand(0).getReg())
860     .addReg(MI->getOperand(2).getReg()).addMBB(copy0MBB)
861     .addReg(MI->getOperand(1).getReg()).addMBB(thisMBB);
862 
863   MI->eraseFromParent();   // The pseudo instruction is gone now.
864   return BB;
865 }
866