?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
zpu: wip eke out some simple instructions for load/store/add
[llvm/zpu.git] / lib / Target / SystemZ / SystemZISelLowering.cpp
blob1fc9d812b686d12b5d17bdab8c7c67cb1e899c5a
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 // Set shifts properties
63 setShiftAmountType(MVT::i64);
64
65 // Provide all sorts of operation actions
66 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
67 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
68 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote);
69
70 setLoadExtAction(ISD::SEXTLOAD, MVT::f32, Expand);
71 setLoadExtAction(ISD::ZEXTLOAD, MVT::f32, Expand);
72 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
73
74 setLoadExtAction(ISD::SEXTLOAD, MVT::f64, Expand);
75 setLoadExtAction(ISD::ZEXTLOAD, MVT::f64, Expand);
76 setLoadExtAction(ISD::EXTLOAD, MVT::f64, Expand);
77
78 setStackPointerRegisterToSaveRestore(SystemZ::R15D);
79
80 // TODO: It may be better to default to latency-oriented scheduling, however
81 // LLVM's current latency-oriented scheduler can't handle physreg definitions
82 // such as SystemZ has with PSW, so set this to the register-pressure
83 // scheduler, because it can.
84 setSchedulingPreference(Sched::RegPressure);
85
86 setBooleanContents(ZeroOrOneBooleanContent);
87
88 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
89 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
90 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
91 setOperationAction(ISD::BR_CC, MVT::i64, Custom);
92 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
93 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
94 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
95 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
96 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
97 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
98 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
99
100 setOperationAction(ISD::SDIV, MVT::i32, Expand);
101 setOperationAction(ISD::UDIV, MVT::i32, Expand);
102 setOperationAction(ISD::SDIV, MVT::i64, Expand);
103 setOperationAction(ISD::UDIV, MVT::i64, Expand);
104 setOperationAction(ISD::SREM, MVT::i32, Expand);
105 setOperationAction(ISD::UREM, MVT::i32, Expand);
106 setOperationAction(ISD::SREM, MVT::i64, Expand);
107 setOperationAction(ISD::UREM, MVT::i64, Expand);
108
109 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
110
111 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
112 setOperationAction(ISD::CTPOP, MVT::i64, Expand);
113 setOperationAction(ISD::CTTZ, MVT::i32, Expand);
114 setOperationAction(ISD::CTTZ, MVT::i64, Expand);
115 setOperationAction(ISD::CTLZ, MVT::i32, Promote);
116 setOperationAction(ISD::CTLZ, MVT::i64, Legal);
117
118 // FIXME: Can we lower these 2 efficiently?
119 setOperationAction(ISD::SETCC, MVT::i32, Expand);
120 setOperationAction(ISD::SETCC, MVT::i64, Expand);
121 setOperationAction(ISD::SETCC, MVT::f32, Expand);
122 setOperationAction(ISD::SETCC, MVT::f64, Expand);
123 setOperationAction(ISD::SELECT, MVT::i32, Expand);
124 setOperationAction(ISD::SELECT, MVT::i64, Expand);
125 setOperationAction(ISD::SELECT, MVT::f32, Expand);
126 setOperationAction(ISD::SELECT, MVT::f64, Expand);
127 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
128 setOperationAction(ISD::SELECT_CC, MVT::i64, Custom);
129 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
130 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
131
132 setOperationAction(ISD::MULHS, MVT::i64, Expand);
133 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
134
135 // FIXME: Can we support these natively?
136 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
137 setOperationAction(ISD::SRL_PARTS, MVT::i64, Expand);
138 setOperationAction(ISD::SHL_PARTS, MVT::i64, Expand);
139 setOperationAction(ISD::SRA_PARTS, MVT::i64, Expand);
140
141 // Lower some FP stuff
142 setOperationAction(ISD::FSIN, MVT::f32, Expand);
143 setOperationAction(ISD::FSIN, MVT::f64, Expand);
144 setOperationAction(ISD::FCOS, MVT::f32, Expand);
145 setOperationAction(ISD::FCOS, MVT::f64, Expand);
146 setOperationAction(ISD::FREM, MVT::f32, Expand);
147 setOperationAction(ISD::FREM, MVT::f64, Expand);
148
149 // We have only 64-bit bitconverts
150 setOperationAction(ISD::BIT_CONVERT, MVT::f32, Expand);
151 setOperationAction(ISD::BIT_CONVERT, MVT::i32, Expand);
152
153 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
154 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
155 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
156 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
157
158 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
159 }
160
161 SDValue SystemZTargetLowering::LowerOperation(SDValue Op,
162 SelectionDAG &DAG) const {
163 switch (Op.getOpcode()) {
164 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
165 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
166 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
167 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
168 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
169 default:
170 llvm_unreachable("Should not custom lower this!");
171 return SDValue();
172 }
173 }
174
175 bool SystemZTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
176 if (UseSoftFloat || (VT != MVT::f32 && VT != MVT::f64))
177 return false;
178
179 // +0.0 lzer
180 // +0.0f lzdr
181 // -0.0 lzer + lner
182 // -0.0f lzdr + lndr
183 return Imm.isZero() || Imm.isNegZero();
184 }
185
186 //===----------------------------------------------------------------------===//
187 // SystemZ Inline Assembly Support
188 //===----------------------------------------------------------------------===//
189
190 /// getConstraintType - Given a constraint letter, return the type of
191 /// constraint it is for this target.
192 TargetLowering::ConstraintType
193 SystemZTargetLowering::getConstraintType(const std::string &Constraint) const {
194 if (Constraint.size() == 1) {
195 switch (Constraint[0]) {
196 case 'r':
197 return C_RegisterClass;
198 default:
199 break;
200 }
201 }
202 return TargetLowering::getConstraintType(Constraint);
203 }
204
205 std::pair<unsigned, const TargetRegisterClass*>
206 SystemZTargetLowering::
207 getRegForInlineAsmConstraint(const std::string &Constraint,
208 EVT VT) const {
209 if (Constraint.size() == 1) {
210 // GCC Constraint Letters
211 switch (Constraint[0]) {
212 default: break;
213 case 'r': // GENERAL_REGS
214 if (VT == MVT::i32)
215 return std::make_pair(0U, SystemZ::GR32RegisterClass);
216 else if (VT == MVT::i128)
217 return std::make_pair(0U, SystemZ::GR128RegisterClass);
218
219 return std::make_pair(0U, SystemZ::GR64RegisterClass);
220 }
221 }
222
223 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
224 }
225
226 //===----------------------------------------------------------------------===//
227 // Calling Convention Implementation
228 //===----------------------------------------------------------------------===//
229
230 #include "SystemZGenCallingConv.inc"
231
232 SDValue
233 SystemZTargetLowering::LowerFormalArguments(SDValue Chain,
234 CallingConv::ID CallConv,
235 bool isVarArg,
236 const SmallVectorImpl<ISD::InputArg>
237 &Ins,
238 DebugLoc dl,
239 SelectionDAG &DAG,
240 SmallVectorImpl<SDValue> &InVals)
241 const {
242
243 switch (CallConv) {
244 default:
245 llvm_unreachable("Unsupported calling convention");
246 case CallingConv::C:
247 case CallingConv::Fast:
248 return LowerCCCArguments(Chain, CallConv, isVarArg, Ins, dl, DAG, InVals);
249 }
250 }
251
252 SDValue
253 SystemZTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
254 CallingConv::ID CallConv, bool isVarArg,
255 bool &isTailCall,
256 const SmallVectorImpl<ISD::OutputArg> &Outs,
257 const SmallVectorImpl<SDValue> &OutVals,
258 const SmallVectorImpl<ISD::InputArg> &Ins,
259 DebugLoc dl, SelectionDAG &DAG,
260 SmallVectorImpl<SDValue> &InVals) const {
261 // SystemZ target does not yet support tail call optimization.
262 isTailCall = false;
263
264 switch (CallConv) {
265 default:
266 llvm_unreachable("Unsupported calling convention");
267 case CallingConv::Fast:
268 case CallingConv::C:
269 return LowerCCCCallTo(Chain, Callee, CallConv, isVarArg, isTailCall,
270 Outs, OutVals, Ins, dl, DAG, InVals);
271 }
272 }
273
274 /// LowerCCCArguments - transform physical registers into virtual registers and
275 /// generate load operations for arguments places on the stack.
276 // FIXME: struct return stuff
277 // FIXME: varargs
278 SDValue
279 SystemZTargetLowering::LowerCCCArguments(SDValue Chain,
280 CallingConv::ID CallConv,
281 bool isVarArg,
282 const SmallVectorImpl<ISD::InputArg>
283 &Ins,
284 DebugLoc dl,
285 SelectionDAG &DAG,
286 SmallVectorImpl<SDValue> &InVals)
287 const {
288
289 MachineFunction &MF = DAG.getMachineFunction();
290 MachineFrameInfo *MFI = MF.getFrameInfo();
291 MachineRegisterInfo &RegInfo = MF.getRegInfo();
292
293 // Assign locations to all of the incoming arguments.
294 SmallVector<CCValAssign, 16> ArgLocs;
295 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
296 ArgLocs, *DAG.getContext());
297 CCInfo.AnalyzeFormalArguments(Ins, CC_SystemZ);
298
299 if (isVarArg)
300 report_fatal_error("Varargs not supported yet");
301
302 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
303 SDValue ArgValue;
304 CCValAssign &VA = ArgLocs[i];
305 EVT LocVT = VA.getLocVT();
306 if (VA.isRegLoc()) {
307 // Arguments passed in registers
308 TargetRegisterClass *RC;
309 switch (LocVT.getSimpleVT().SimpleTy) {
310 default:
311 #ifndef NDEBUG
312 errs() << "LowerFormalArguments Unhandled argument type: "
313 << LocVT.getSimpleVT().SimpleTy
314 << "\n";
315 #endif
316 llvm_unreachable(0);
317 case MVT::i64:
318 RC = SystemZ::GR64RegisterClass;
319 break;
320 case MVT::f32:
321 RC = SystemZ::FP32RegisterClass;
322 break;
323 case MVT::f64:
324 RC = SystemZ::FP64RegisterClass;
325 break;
326 }
327
328 unsigned VReg = RegInfo.createVirtualRegister(RC);
329 RegInfo.addLiveIn(VA.getLocReg(), VReg);
330 ArgValue = DAG.getCopyFromReg(Chain, dl, VReg, LocVT);
331 } else {
332 // Sanity check
333 assert(VA.isMemLoc());
334
335 // Create the nodes corresponding to a load from this parameter slot.
336 // Create the frame index object for this incoming parameter...
337 int FI = MFI->CreateFixedObject(LocVT.getSizeInBits()/8,
338 VA.getLocMemOffset(), true);
339
340 // Create the SelectionDAG nodes corresponding to a load
341 // from this parameter
342 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
343 ArgValue = DAG.getLoad(LocVT, dl, Chain, FIN,
344 MachinePointerInfo::getFixedStack(FI),
345 false, false, 0);
346 }
347
348 // If this is an 8/16/32-bit value, it is really passed promoted to 64
349 // bits. Insert an assert[sz]ext to capture this, then truncate to the
350 // right size.
351 if (VA.getLocInfo() == CCValAssign::SExt)
352 ArgValue = DAG.getNode(ISD::AssertSext, dl, LocVT, ArgValue,
353 DAG.getValueType(VA.getValVT()));
354 else if (VA.getLocInfo() == CCValAssign::ZExt)
355 ArgValue = DAG.getNode(ISD::AssertZext, dl, LocVT, ArgValue,
356 DAG.getValueType(VA.getValVT()));
357
358 if (VA.getLocInfo() != CCValAssign::Full)
359 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
360
361 InVals.push_back(ArgValue);
362 }
363
364 return Chain;
365 }
366
367 /// LowerCCCCallTo - functions arguments are copied from virtual regs to
368 /// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted.
369 /// TODO: sret.
370 SDValue
371 SystemZTargetLowering::LowerCCCCallTo(SDValue Chain, SDValue Callee,
372 CallingConv::ID CallConv, bool isVarArg,
373 bool isTailCall,
374 const SmallVectorImpl<ISD::OutputArg>
375 &Outs,
376 const SmallVectorImpl<SDValue> &OutVals,
377 const SmallVectorImpl<ISD::InputArg> &Ins,
378 DebugLoc dl, SelectionDAG &DAG,
379 SmallVectorImpl<SDValue> &InVals) const {
380
381 MachineFunction &MF = DAG.getMachineFunction();
382
383 // Offset to first argument stack slot.
384 const unsigned FirstArgOffset = 160;
385
386 // Analyze operands of the call, assigning locations to each operand.
387 SmallVector<CCValAssign, 16> ArgLocs;
388 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
389 ArgLocs, *DAG.getContext());
390
391 CCInfo.AnalyzeCallOperands(Outs, CC_SystemZ);
392
393 // Get a count of how many bytes are to be pushed on the stack.
394 unsigned NumBytes = CCInfo.getNextStackOffset();
395
396 Chain = DAG.getCALLSEQ_START(Chain ,DAG.getConstant(NumBytes,
397 getPointerTy(), true));
398
399 SmallVector<std::pair<unsigned, SDValue>, 4> RegsToPass;
400 SmallVector<SDValue, 12> MemOpChains;
401 SDValue StackPtr;
402
403 // Walk the register/memloc assignments, inserting copies/loads.
404 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
405 CCValAssign &VA = ArgLocs[i];
406
407 SDValue Arg = OutVals[i];
408
409 // Promote the value if needed.
410 switch (VA.getLocInfo()) {
411 default: assert(0 && "Unknown loc info!");
412 case CCValAssign::Full: break;
413 case CCValAssign::SExt:
414 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
415 break;
416 case CCValAssign::ZExt:
417 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
418 break;
419 case CCValAssign::AExt:
420 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
421 break;
422 }
423
424 // Arguments that can be passed on register must be kept at RegsToPass
425 // vector
426 if (VA.isRegLoc()) {
427 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
428 } else {
429 assert(VA.isMemLoc());
430
431 if (StackPtr.getNode() == 0)
432 StackPtr =
433 DAG.getCopyFromReg(Chain, dl,
434 (RegInfo->hasFP(MF) ?
435 SystemZ::R11D : SystemZ::R15D),
436 getPointerTy());
437
438 unsigned Offset = FirstArgOffset + VA.getLocMemOffset();
439 SDValue PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(),
440 StackPtr,
441 DAG.getIntPtrConstant(Offset));
442
443 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
444 MachinePointerInfo(),
445 false, false, 0));
446 }
447 }
448
449 // Transform all store nodes into one single node because all store nodes are
450 // independent of each other.
451 if (!MemOpChains.empty())
452 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
453 &MemOpChains[0], MemOpChains.size());
454
455 // Build a sequence of copy-to-reg nodes chained together with token chain and
456 // flag operands which copy the outgoing args into registers. The InFlag in
457 // necessary since all emited instructions must be stuck together.
458 SDValue InFlag;
459 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
460 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
461 RegsToPass[i].second, InFlag);
462 InFlag = Chain.getValue(1);
463 }
464
465 // If the callee is a GlobalAddress node (quite common, every direct call is)
466 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
467 // Likewise ExternalSymbol -> TargetExternalSymbol.
468 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
469 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl, getPointerTy());
470 else if (ExternalSymbolSDNode *E = dyn_cast<ExternalSymbolSDNode>(Callee))
471 Callee = DAG.getTargetExternalSymbol(E->getSymbol(), getPointerTy());
472
473 // Returns a chain & a flag for retval copy to use.
474 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
475 SmallVector<SDValue, 8> Ops;
476 Ops.push_back(Chain);
477 Ops.push_back(Callee);
478
479 // Add argument registers to the end of the list so that they are
480 // known live into the call.
481 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
482 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
483 RegsToPass[i].second.getValueType()));
484
485 if (InFlag.getNode())
486 Ops.push_back(InFlag);
487
488 Chain = DAG.getNode(SystemZISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
489 InFlag = Chain.getValue(1);
490
491 // Create the CALLSEQ_END node.
492 Chain = DAG.getCALLSEQ_END(Chain,
493 DAG.getConstant(NumBytes, getPointerTy(), true),
494 DAG.getConstant(0, getPointerTy(), true),
495 InFlag);
496 InFlag = Chain.getValue(1);
497
498 // Handle result values, copying them out of physregs into vregs that we
499 // return.
500 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl,
501 DAG, InVals);
502 }
503
504 /// LowerCallResult - Lower the result values of a call into the
505 /// appropriate copies out of appropriate physical registers.
506 ///
507 SDValue
508 SystemZTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
509 CallingConv::ID CallConv, bool isVarArg,
510 const SmallVectorImpl<ISD::InputArg>
511 &Ins,
512 DebugLoc dl, SelectionDAG &DAG,
513 SmallVectorImpl<SDValue> &InVals) const {
514
515 // Assign locations to each value returned by this call.
516 SmallVector<CCValAssign, 16> RVLocs;
517 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
518 *DAG.getContext());
519
520 CCInfo.AnalyzeCallResult(Ins, RetCC_SystemZ);
521
522 // Copy all of the result registers out of their specified physreg.
523 for (unsigned i = 0; i != RVLocs.size(); ++i) {
524 CCValAssign &VA = RVLocs[i];
525
526 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
527 VA.getLocVT(), InFlag).getValue(1);
528 SDValue RetValue = Chain.getValue(0);
529 InFlag = Chain.getValue(2);
530
531 // If this is an 8/16/32-bit value, it is really passed promoted to 64
532 // bits. Insert an assert[sz]ext to capture this, then truncate to the
533 // right size.
534 if (VA.getLocInfo() == CCValAssign::SExt)
535 RetValue = DAG.getNode(ISD::AssertSext, dl, VA.getLocVT(), RetValue,
536 DAG.getValueType(VA.getValVT()));
537 else if (VA.getLocInfo() == CCValAssign::ZExt)
538 RetValue = DAG.getNode(ISD::AssertZext, dl, VA.getLocVT(), RetValue,
539 DAG.getValueType(VA.getValVT()));
540
541 if (VA.getLocInfo() != CCValAssign::Full)
542 RetValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), RetValue);
543
544 InVals.push_back(RetValue);
545 }
546
547 return Chain;
548 }
549
550
551 SDValue
552 SystemZTargetLowering::LowerReturn(SDValue Chain,
553 CallingConv::ID CallConv, bool isVarArg,
554 const SmallVectorImpl<ISD::OutputArg> &Outs,
555 const SmallVectorImpl<SDValue> &OutVals,
556 DebugLoc dl, SelectionDAG &DAG) const {
557
558 // CCValAssign - represent the assignment of the return value to a location
559 SmallVector<CCValAssign, 16> RVLocs;
560
561 // CCState - Info about the registers and stack slot.
562 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
563 RVLocs, *DAG.getContext());
564
565 // Analize return values.
566 CCInfo.AnalyzeReturn(Outs, RetCC_SystemZ);
567
568 // If this is the first return lowered for this function, add the regs to the
569 // liveout set for the function.
570 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
571 for (unsigned i = 0; i != RVLocs.size(); ++i)
572 if (RVLocs[i].isRegLoc())
573 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
574 }
575
576 SDValue Flag;
577
578 // Copy the result values into the output registers.
579 for (unsigned i = 0; i != RVLocs.size(); ++i) {
580 CCValAssign &VA = RVLocs[i];
581 SDValue ResValue = OutVals[i];
582 assert(VA.isRegLoc() && "Can only return in registers!");
583
584 // If this is an 8/16/32-bit value, it is really should be passed promoted
585 // to 64 bits.
586 if (VA.getLocInfo() == CCValAssign::SExt)
587 ResValue = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ResValue);
588 else if (VA.getLocInfo() == CCValAssign::ZExt)
589 ResValue = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ResValue);
590 else if (VA.getLocInfo() == CCValAssign::AExt)
591 ResValue = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ResValue);
592
593 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ResValue, Flag);
594
595 // Guarantee that all emitted copies are stuck together,
596 // avoiding something bad.
597 Flag = Chain.getValue(1);
598 }
599
600 if (Flag.getNode())
601 return DAG.getNode(SystemZISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
602
603 // Return Void
604 return DAG.getNode(SystemZISD::RET_FLAG, dl, MVT::Other, Chain);
605 }
606
607 SDValue SystemZTargetLowering::EmitCmp(SDValue LHS, SDValue RHS,
608 ISD::CondCode CC, SDValue &SystemZCC,
609 SelectionDAG &DAG) const {
610 // FIXME: Emit a test if RHS is zero
611
612 bool isUnsigned = false;
613 SystemZCC::CondCodes TCC;
614 switch (CC) {
615 default:
616 llvm_unreachable("Invalid integer condition!");
617 case ISD::SETEQ:
618 case ISD::SETOEQ:
619 TCC = SystemZCC::E;
620 break;
621 case ISD::SETUEQ:
622 TCC = SystemZCC::NLH;
623 break;
624 case ISD::SETNE:
625 case ISD::SETONE:
626 TCC = SystemZCC::NE;
627 break;
628 case ISD::SETUNE:
629 TCC = SystemZCC::LH;
630 break;
631 case ISD::SETO:
632 TCC = SystemZCC::O;
633 break;
634 case ISD::SETUO:
635 TCC = SystemZCC::NO;
636 break;
637 case ISD::SETULE:
638 if (LHS.getValueType().isFloatingPoint()) {
639 TCC = SystemZCC::NH;
640 break;
641 }
642 isUnsigned = true; // FALLTHROUGH
643 case ISD::SETLE:
644 case ISD::SETOLE:
645 TCC = SystemZCC::LE;
646 break;
647 case ISD::SETUGE:
648 if (LHS.getValueType().isFloatingPoint()) {
649 TCC = SystemZCC::NL;
650 break;
651 }
652 isUnsigned = true; // FALLTHROUGH
653 case ISD::SETGE:
654 case ISD::SETOGE:
655 TCC = SystemZCC::HE;
656 break;
657 case ISD::SETUGT:
658 if (LHS.getValueType().isFloatingPoint()) {
659 TCC = SystemZCC::NLE;
660 break;
661 }
662 isUnsigned = true; // FALLTHROUGH
663 case ISD::SETGT:
664 case ISD::SETOGT:
665 TCC = SystemZCC::H;
666 break;
667 case ISD::SETULT:
668 if (LHS.getValueType().isFloatingPoint()) {
669 TCC = SystemZCC::NHE;
670 break;
671 }
672 isUnsigned = true; // FALLTHROUGH
673 case ISD::SETLT:
674 case ISD::SETOLT:
675 TCC = SystemZCC::L;
676 break;
677 }
678
679 SystemZCC = DAG.getConstant(TCC, MVT::i32);
680
681 DebugLoc dl = LHS.getDebugLoc();
682 return DAG.getNode((isUnsigned ? SystemZISD::UCMP : SystemZISD::CMP),
683 dl, MVT::i64, LHS, RHS);
684 }
685
686
687 SDValue SystemZTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
688 SDValue Chain = Op.getOperand(0);
689 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
690 SDValue LHS = Op.getOperand(2);
691 SDValue RHS = Op.getOperand(3);
692 SDValue Dest = Op.getOperand(4);
693 DebugLoc dl = Op.getDebugLoc();
694
695 SDValue SystemZCC;
696 SDValue Flag = EmitCmp(LHS, RHS, CC, SystemZCC, DAG);
697 return DAG.getNode(SystemZISD::BRCOND, dl, Op.getValueType(),
698 Chain, Dest, SystemZCC, Flag);
699 }
700
701 SDValue SystemZTargetLowering::LowerSELECT_CC(SDValue Op,
702 SelectionDAG &DAG) const {
703 SDValue LHS = Op.getOperand(0);
704 SDValue RHS = Op.getOperand(1);
705 SDValue TrueV = Op.getOperand(2);
706 SDValue FalseV = Op.getOperand(3);
707 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
708 DebugLoc dl = Op.getDebugLoc();
709
710 SDValue SystemZCC;
711 SDValue Flag = EmitCmp(LHS, RHS, CC, SystemZCC, DAG);
712
713 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
714 SmallVector<SDValue, 4> Ops;
715 Ops.push_back(TrueV);
716 Ops.push_back(FalseV);
717 Ops.push_back(SystemZCC);
718 Ops.push_back(Flag);
719
720 return DAG.getNode(SystemZISD::SELECT, dl, VTs, &Ops[0], Ops.size());
721 }
722
723 SDValue SystemZTargetLowering::LowerGlobalAddress(SDValue Op,
724 SelectionDAG &DAG) const {
725 DebugLoc dl = Op.getDebugLoc();
726 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
727 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
728
729 bool IsPic = getTargetMachine().getRelocationModel() == Reloc::PIC_;
730 bool ExtraLoadRequired =
731 Subtarget.GVRequiresExtraLoad(GV, getTargetMachine(), false);
732
733 SDValue Result;
734 if (!IsPic && !ExtraLoadRequired) {
735 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
736 Offset = 0;
737 } else {
738 unsigned char OpFlags = 0;
739 if (ExtraLoadRequired)
740 OpFlags = SystemZII::MO_GOTENT;
741
742 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
743 }
744
745 Result = DAG.getNode(SystemZISD::PCRelativeWrapper, dl,
746 getPointerTy(), Result);
747
748 if (ExtraLoadRequired)
749 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
750 MachinePointerInfo::getGOT(), false, false, 0);
751
752 // If there was a non-zero offset that we didn't fold, create an explicit
753 // addition for it.
754 if (Offset != 0)
755 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
756 DAG.getConstant(Offset, getPointerTy()));
757
758 return Result;
759 }
760
761 // FIXME: PIC here
762 SDValue SystemZTargetLowering::LowerJumpTable(SDValue Op,
763 SelectionDAG &DAG) const {
764 DebugLoc dl = Op.getDebugLoc();
765 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
766 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
767
768 return DAG.getNode(SystemZISD::PCRelativeWrapper, dl, getPointerTy(), Result);
769 }
770
771
772 // FIXME: PIC here
773 // FIXME: This is just dirty hack. We need to lower cpool properly
774 SDValue SystemZTargetLowering::LowerConstantPool(SDValue Op,
775 SelectionDAG &DAG) const {
776 DebugLoc dl = Op.getDebugLoc();
777 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
778
779 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
780 CP->getAlignment(),
781 CP->getOffset());
782
783 return DAG.getNode(SystemZISD::PCRelativeWrapper, dl, getPointerTy(), Result);
784 }
785
786 const char *SystemZTargetLowering::getTargetNodeName(unsigned Opcode) const {
787 switch (Opcode) {
788 case SystemZISD::RET_FLAG: return "SystemZISD::RET_FLAG";
789 case SystemZISD::CALL: return "SystemZISD::CALL";
790 case SystemZISD::BRCOND: return "SystemZISD::BRCOND";
791 case SystemZISD::CMP: return "SystemZISD::CMP";
792 case SystemZISD::UCMP: return "SystemZISD::UCMP";
793 case SystemZISD::SELECT: return "SystemZISD::SELECT";
794 case SystemZISD::PCRelativeWrapper: return "SystemZISD::PCRelativeWrapper";
795 default: return NULL;
796 }
797 }
798
799 //===----------------------------------------------------------------------===//
800 // Other Lowering Code
801 //===----------------------------------------------------------------------===//
802
803 MachineBasicBlock*
804 SystemZTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
805 MachineBasicBlock *BB) const {
806 const SystemZInstrInfo &TII = *TM.getInstrInfo();
807 DebugLoc dl = MI->getDebugLoc();
808 assert((MI->getOpcode() == SystemZ::Select32 ||
809 MI->getOpcode() == SystemZ::SelectF32 ||
810 MI->getOpcode() == SystemZ::Select64 ||
811 MI->getOpcode() == SystemZ::SelectF64) &&
812 "Unexpected instr type to insert");
813
814 // To "insert" a SELECT instruction, we actually have to insert the diamond
815 // control-flow pattern. The incoming instruction knows the destination vreg
816 // to set, the condition code register to branch on, the true/false values to
817 // select between, and a branch opcode to use.
818 const BasicBlock *LLVM_BB = BB->getBasicBlock();
819 MachineFunction::iterator I = BB;
820 ++I;
821
822 // thisMBB:
823 // ...
824 // TrueVal = ...
825 // cmpTY ccX, r1, r2
826 // jCC copy1MBB
827 // fallthrough --> copy0MBB
828 MachineBasicBlock *thisMBB = BB;
829 MachineFunction *F = BB->getParent();
830 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
831 MachineBasicBlock *copy1MBB = F->CreateMachineBasicBlock(LLVM_BB);
832 SystemZCC::CondCodes CC = (SystemZCC::CondCodes)MI->getOperand(3).getImm();
833 F->insert(I, copy0MBB);
834 F->insert(I, copy1MBB);
835 // Update machine-CFG edges by transferring all successors of the current
836 // block to the new block which will contain the Phi node for the select.
837 copy1MBB->splice(copy1MBB->begin(), BB,
838 llvm::next(MachineBasicBlock::iterator(MI)),
839 BB->end());
840 copy1MBB->transferSuccessorsAndUpdatePHIs(BB);
841 // Next, add the true and fallthrough blocks as its successors.
842 BB->addSuccessor(copy0MBB);
843 BB->addSuccessor(copy1MBB);
844
845 BuildMI(BB, dl, TII.getBrCond(CC)).addMBB(copy1MBB);
846
847 // copy0MBB:
848 // %FalseValue = ...
849 // # fallthrough to copy1MBB
850 BB = copy0MBB;
851
852 // Update machine-CFG edges
853 BB->addSuccessor(copy1MBB);
854
855 // copy1MBB:
856 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
857 // ...
858 BB = copy1MBB;
859 BuildMI(*BB, BB->begin(), dl, TII.get(SystemZ::PHI),
860 MI->getOperand(0).getReg())
861 .addReg(MI->getOperand(2).getReg()).addMBB(copy0MBB)
862 .addReg(MI->getOperand(1).getReg()).addMBB(thisMBB);
863
864 MI->eraseFromParent(); // The pseudo instruction is gone now.
865 return BB;
866 }
ZPU llvm backend wip