1 //===-- SparcISelLowering.cpp - Sparc DAG Lowering Implementation ---------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the interfaces that Sparc uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #include "SparcISelLowering.h"
16 #include "SparcMachineFunctionInfo.h"
17 #include "SparcTargetMachine.h"
18 #include "llvm/CodeGen/CallingConvLower.h"
19 #include "llvm/CodeGen/MachineFrameInfo.h"
20 #include "llvm/CodeGen/MachineFunction.h"
21 #include "llvm/CodeGen/MachineInstrBuilder.h"
22 #include "llvm/CodeGen/MachineRegisterInfo.h"
23 #include "llvm/CodeGen/SelectionDAG.h"
24 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
25 #include "llvm/IR/DerivedTypes.h"
26 #include "llvm/IR/Function.h"
27 #include "llvm/IR/Module.h"
28 #include "llvm/Support/ErrorHandling.h"
32 //===----------------------------------------------------------------------===//
33 // Calling Convention Implementation
34 //===----------------------------------------------------------------------===//
36 static bool CC_Sparc_Assign_SRet(unsigned &ValNo, MVT &ValVT,
37 MVT &LocVT, CCValAssign::LocInfo &LocInfo,
38 ISD::ArgFlagsTy &ArgFlags, CCState &State)
40 assert (ArgFlags.isSRet());
42 //Assign SRet argument
43 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
49 static bool CC_Sparc_Assign_f64(unsigned &ValNo, MVT &ValVT,
50 MVT &LocVT, CCValAssign::LocInfo &LocInfo,
51 ISD::ArgFlagsTy &ArgFlags, CCState &State)
53 static const uint16_t RegList[] = {
54 SP::I0, SP::I1, SP::I2, SP::I3, SP::I4, SP::I5
56 //Try to get first reg
57 if (unsigned Reg = State.AllocateReg(RegList, 6)) {
58 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
60 //Assign whole thing in stack
61 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
62 State.AllocateStack(8,4),
67 //Try to get second reg
68 if (unsigned Reg = State.AllocateReg(RegList, 6))
69 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
71 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
72 State.AllocateStack(4,4),
77 // Allocate a full-sized argument for the 64-bit ABI.
78 static bool CC_Sparc64_Full(unsigned &ValNo, MVT &ValVT,
79 MVT &LocVT, CCValAssign::LocInfo &LocInfo,
80 ISD::ArgFlagsTy &ArgFlags, CCState &State) {
81 assert((LocVT == MVT::f32 || LocVT.getSizeInBits() == 64) &&
82 "Can't handle non-64 bits locations");
84 // Stack space is allocated for all arguments starting from [%fp+BIAS+128].
85 unsigned Offset = State.AllocateStack(8, 8);
88 if (LocVT == MVT::i64 && Offset < 6*8)
89 // Promote integers to %i0-%i5.
90 Reg = SP::I0 + Offset/8;
91 else if (LocVT == MVT::f64 && Offset < 16*8)
92 // Promote doubles to %d0-%d30. (Which LLVM calls D0-D15).
93 Reg = SP::D0 + Offset/8;
94 else if (LocVT == MVT::f32 && Offset < 16*8)
95 // Promote floats to %f1, %f3, ...
96 Reg = SP::F1 + Offset/4;
98 // Promote to register when possible, otherwise use the stack slot.
100 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
104 // This argument goes on the stack in an 8-byte slot.
105 // When passing floats, LocVT is smaller than 8 bytes. Adjust the offset to
106 // the right-aligned float. The first 4 bytes of the stack slot are undefined.
107 if (LocVT == MVT::f32)
110 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
114 // Allocate a half-sized argument for the 64-bit ABI.
116 // This is used when passing { float, int } structs by value in registers.
117 static bool CC_Sparc64_Half(unsigned &ValNo, MVT &ValVT,
118 MVT &LocVT, CCValAssign::LocInfo &LocInfo,
119 ISD::ArgFlagsTy &ArgFlags, CCState &State) {
120 assert(LocVT.getSizeInBits() == 32 && "Can't handle non-32 bits locations");
121 unsigned Offset = State.AllocateStack(4, 4);
123 if (LocVT == MVT::f32 && Offset < 16*8) {
124 // Promote floats to %f0-%f31.
125 State.addLoc(CCValAssign::getReg(ValNo, ValVT, SP::F0 + Offset/4,
130 if (LocVT == MVT::i32 && Offset < 6*8) {
131 // Promote integers to %i0-%i5, using half the register.
132 unsigned Reg = SP::I0 + Offset/8;
134 LocInfo = CCValAssign::AExt;
136 // Set the Custom bit if this i32 goes in the high bits of a register.
138 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg,
141 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
145 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
149 #include "SparcGenCallingConv.inc"
152 SparcTargetLowering::LowerReturn(SDValue Chain,
153 CallingConv::ID CallConv, bool IsVarArg,
154 const SmallVectorImpl<ISD::OutputArg> &Outs,
155 const SmallVectorImpl<SDValue> &OutVals,
156 DebugLoc DL, SelectionDAG &DAG) const {
157 if (Subtarget->is64Bit())
158 return LowerReturn_64(Chain, CallConv, IsVarArg, Outs, OutVals, DL, DAG);
159 return LowerReturn_32(Chain, CallConv, IsVarArg, Outs, OutVals, DL, DAG);
163 SparcTargetLowering::LowerReturn_32(SDValue Chain,
164 CallingConv::ID CallConv, bool IsVarArg,
165 const SmallVectorImpl<ISD::OutputArg> &Outs,
166 const SmallVectorImpl<SDValue> &OutVals,
167 DebugLoc DL, SelectionDAG &DAG) const {
168 MachineFunction &MF = DAG.getMachineFunction();
170 // CCValAssign - represent the assignment of the return value to locations.
171 SmallVector<CCValAssign, 16> RVLocs;
173 // CCState - Info about the registers and stack slot.
174 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
175 DAG.getTarget(), RVLocs, *DAG.getContext());
177 // Analyze return values.
178 CCInfo.AnalyzeReturn(Outs, RetCC_Sparc32);
181 SmallVector<SDValue, 4> RetOps(1, Chain);
182 // Make room for the return address offset.
183 RetOps.push_back(SDValue());
185 // Copy the result values into the output registers.
186 for (unsigned i = 0; i != RVLocs.size(); ++i) {
187 CCValAssign &VA = RVLocs[i];
188 assert(VA.isRegLoc() && "Can only return in registers!");
190 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(),
193 // Guarantee that all emitted copies are stuck together with flags.
194 Flag = Chain.getValue(1);
195 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
198 unsigned RetAddrOffset = 8; //Call Inst + Delay Slot
199 // If the function returns a struct, copy the SRetReturnReg to I0
200 if (MF.getFunction()->hasStructRetAttr()) {
201 SparcMachineFunctionInfo *SFI = MF.getInfo<SparcMachineFunctionInfo>();
202 unsigned Reg = SFI->getSRetReturnReg();
204 llvm_unreachable("sret virtual register not created in the entry block");
205 SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
206 Chain = DAG.getCopyToReg(Chain, DL, SP::I0, Val, Flag);
207 Flag = Chain.getValue(1);
208 RetOps.push_back(DAG.getRegister(SP::I0, getPointerTy()));
209 RetAddrOffset = 12; // CallInst + Delay Slot + Unimp
212 RetOps[0] = Chain; // Update chain.
213 RetOps[1] = DAG.getConstant(RetAddrOffset, MVT::i32);
215 // Add the flag if we have it.
217 RetOps.push_back(Flag);
219 return DAG.getNode(SPISD::RET_FLAG, DL, MVT::Other,
220 &RetOps[0], RetOps.size());
223 // Lower return values for the 64-bit ABI.
224 // Return values are passed the exactly the same way as function arguments.
226 SparcTargetLowering::LowerReturn_64(SDValue Chain,
227 CallingConv::ID CallConv, bool IsVarArg,
228 const SmallVectorImpl<ISD::OutputArg> &Outs,
229 const SmallVectorImpl<SDValue> &OutVals,
230 DebugLoc DL, SelectionDAG &DAG) const {
231 // CCValAssign - represent the assignment of the return value to locations.
232 SmallVector<CCValAssign, 16> RVLocs;
234 // CCState - Info about the registers and stack slot.
235 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
236 DAG.getTarget(), RVLocs, *DAG.getContext());
238 // Analyze return values.
239 CCInfo.AnalyzeReturn(Outs, CC_Sparc64);
242 SmallVector<SDValue, 4> RetOps(1, Chain);
244 // The second operand on the return instruction is the return address offset.
245 // The return address is always %i7+8 with the 64-bit ABI.
246 RetOps.push_back(DAG.getConstant(8, MVT::i32));
248 // Copy the result values into the output registers.
249 for (unsigned i = 0; i != RVLocs.size(); ++i) {
250 CCValAssign &VA = RVLocs[i];
251 assert(VA.isRegLoc() && "Can only return in registers!");
252 SDValue OutVal = OutVals[i];
254 // Integer return values must be sign or zero extended by the callee.
255 switch (VA.getLocInfo()) {
256 case CCValAssign::SExt:
257 OutVal = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), OutVal);
259 case CCValAssign::ZExt:
260 OutVal = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), OutVal);
262 case CCValAssign::AExt:
263 OutVal = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), OutVal);
268 // The custom bit on an i32 return value indicates that it should be passed
269 // in the high bits of the register.
270 if (VA.getValVT() == MVT::i32 && VA.needsCustom()) {
271 OutVal = DAG.getNode(ISD::SHL, DL, MVT::i64, OutVal,
272 DAG.getConstant(32, MVT::i32));
274 // The next value may go in the low bits of the same register.
275 // Handle both at once.
276 if (i+1 < RVLocs.size() && RVLocs[i+1].getLocReg() == VA.getLocReg()) {
277 SDValue NV = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, OutVals[i+1]);
278 OutVal = DAG.getNode(ISD::OR, DL, MVT::i64, OutVal, NV);
279 // Skip the next value, it's already done.
284 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), OutVal, Flag);
286 // Guarantee that all emitted copies are stuck together with flags.
287 Flag = Chain.getValue(1);
288 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
291 RetOps[0] = Chain; // Update chain.
293 // Add the flag if we have it.
295 RetOps.push_back(Flag);
297 return DAG.getNode(SPISD::RET_FLAG, DL, MVT::Other,
298 &RetOps[0], RetOps.size());
301 SDValue SparcTargetLowering::
302 LowerFormalArguments(SDValue Chain,
303 CallingConv::ID CallConv,
305 const SmallVectorImpl<ISD::InputArg> &Ins,
308 SmallVectorImpl<SDValue> &InVals) const {
309 if (Subtarget->is64Bit())
310 return LowerFormalArguments_64(Chain, CallConv, IsVarArg, Ins,
312 return LowerFormalArguments_32(Chain, CallConv, IsVarArg, Ins,
316 /// LowerFormalArguments32 - V8 uses a very simple ABI, where all values are
317 /// passed in either one or two GPRs, including FP values. TODO: we should
318 /// pass FP values in FP registers for fastcc functions.
319 SDValue SparcTargetLowering::
320 LowerFormalArguments_32(SDValue Chain,
321 CallingConv::ID CallConv,
323 const SmallVectorImpl<ISD::InputArg> &Ins,
326 SmallVectorImpl<SDValue> &InVals) const {
327 MachineFunction &MF = DAG.getMachineFunction();
328 MachineRegisterInfo &RegInfo = MF.getRegInfo();
329 SparcMachineFunctionInfo *FuncInfo = MF.getInfo<SparcMachineFunctionInfo>();
331 // Assign locations to all of the incoming arguments.
332 SmallVector<CCValAssign, 16> ArgLocs;
333 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
334 getTargetMachine(), ArgLocs, *DAG.getContext());
335 CCInfo.AnalyzeFormalArguments(Ins, CC_Sparc32);
337 const unsigned StackOffset = 92;
339 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
340 CCValAssign &VA = ArgLocs[i];
342 if (i == 0 && Ins[i].Flags.isSRet()) {
343 //Get SRet from [%fp+64]
344 int FrameIdx = MF.getFrameInfo()->CreateFixedObject(4, 64, true);
345 SDValue FIPtr = DAG.getFrameIndex(FrameIdx, MVT::i32);
346 SDValue Arg = DAG.getLoad(MVT::i32, dl, Chain, FIPtr,
347 MachinePointerInfo(),
348 false, false, false, 0);
349 InVals.push_back(Arg);
354 if (VA.needsCustom()) {
355 assert(VA.getLocVT() == MVT::f64);
356 unsigned VRegHi = RegInfo.createVirtualRegister(&SP::IntRegsRegClass);
357 MF.getRegInfo().addLiveIn(VA.getLocReg(), VRegHi);
358 SDValue HiVal = DAG.getCopyFromReg(Chain, dl, VRegHi, MVT::i32);
361 CCValAssign &NextVA = ArgLocs[++i];
364 if (NextVA.isMemLoc()) {
365 int FrameIdx = MF.getFrameInfo()->
366 CreateFixedObject(4, StackOffset+NextVA.getLocMemOffset(),true);
367 SDValue FIPtr = DAG.getFrameIndex(FrameIdx, MVT::i32);
368 LoVal = DAG.getLoad(MVT::i32, dl, Chain, FIPtr,
369 MachinePointerInfo(),
370 false, false, false, 0);
372 unsigned loReg = MF.addLiveIn(NextVA.getLocReg(),
373 &SP::IntRegsRegClass);
374 LoVal = DAG.getCopyFromReg(Chain, dl, loReg, MVT::i32);
377 DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, LoVal, HiVal);
378 WholeValue = DAG.getNode(ISD::BITCAST, dl, MVT::f64, WholeValue);
379 InVals.push_back(WholeValue);
382 unsigned VReg = RegInfo.createVirtualRegister(&SP::IntRegsRegClass);
383 MF.getRegInfo().addLiveIn(VA.getLocReg(), VReg);
384 SDValue Arg = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
385 if (VA.getLocVT() == MVT::f32)
386 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Arg);
387 else if (VA.getLocVT() != MVT::i32) {
388 Arg = DAG.getNode(ISD::AssertSext, dl, MVT::i32, Arg,
389 DAG.getValueType(VA.getLocVT()));
390 Arg = DAG.getNode(ISD::TRUNCATE, dl, VA.getLocVT(), Arg);
392 InVals.push_back(Arg);
396 assert(VA.isMemLoc());
398 unsigned Offset = VA.getLocMemOffset()+StackOffset;
400 if (VA.needsCustom()) {
401 assert(VA.getValVT() == MVT::f64);
402 //If it is double-word aligned, just load.
403 if (Offset % 8 == 0) {
404 int FI = MF.getFrameInfo()->CreateFixedObject(8,
407 SDValue FIPtr = DAG.getFrameIndex(FI, getPointerTy());
408 SDValue Load = DAG.getLoad(VA.getValVT(), dl, Chain, FIPtr,
409 MachinePointerInfo(),
410 false,false, false, 0);
411 InVals.push_back(Load);
415 int FI = MF.getFrameInfo()->CreateFixedObject(4,
418 SDValue FIPtr = DAG.getFrameIndex(FI, getPointerTy());
419 SDValue HiVal = DAG.getLoad(MVT::i32, dl, Chain, FIPtr,
420 MachinePointerInfo(),
421 false, false, false, 0);
422 int FI2 = MF.getFrameInfo()->CreateFixedObject(4,
425 SDValue FIPtr2 = DAG.getFrameIndex(FI2, getPointerTy());
427 SDValue LoVal = DAG.getLoad(MVT::i32, dl, Chain, FIPtr2,
428 MachinePointerInfo(),
429 false, false, false, 0);
432 DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, LoVal, HiVal);
433 WholeValue = DAG.getNode(ISD::BITCAST, dl, MVT::f64, WholeValue);
434 InVals.push_back(WholeValue);
438 int FI = MF.getFrameInfo()->CreateFixedObject(4,
441 SDValue FIPtr = DAG.getFrameIndex(FI, getPointerTy());
443 if (VA.getValVT() == MVT::i32 || VA.getValVT() == MVT::f32) {
444 Load = DAG.getLoad(VA.getValVT(), dl, Chain, FIPtr,
445 MachinePointerInfo(),
446 false, false, false, 0);
448 ISD::LoadExtType LoadOp = ISD::SEXTLOAD;
449 // Sparc is big endian, so add an offset based on the ObjectVT.
450 unsigned Offset = 4-std::max(1U, VA.getValVT().getSizeInBits()/8);
451 FIPtr = DAG.getNode(ISD::ADD, dl, MVT::i32, FIPtr,
452 DAG.getConstant(Offset, MVT::i32));
453 Load = DAG.getExtLoad(LoadOp, dl, MVT::i32, Chain, FIPtr,
454 MachinePointerInfo(),
455 VA.getValVT(), false, false,0);
456 Load = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Load);
458 InVals.push_back(Load);
461 if (MF.getFunction()->hasStructRetAttr()) {
462 //Copy the SRet Argument to SRetReturnReg
463 SparcMachineFunctionInfo *SFI = MF.getInfo<SparcMachineFunctionInfo>();
464 unsigned Reg = SFI->getSRetReturnReg();
466 Reg = MF.getRegInfo().createVirtualRegister(&SP::IntRegsRegClass);
467 SFI->setSRetReturnReg(Reg);
469 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
470 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
473 // Store remaining ArgRegs to the stack if this is a varargs function.
475 static const uint16_t ArgRegs[] = {
476 SP::I0, SP::I1, SP::I2, SP::I3, SP::I4, SP::I5
478 unsigned NumAllocated = CCInfo.getFirstUnallocated(ArgRegs, 6);
479 const uint16_t *CurArgReg = ArgRegs+NumAllocated, *ArgRegEnd = ArgRegs+6;
480 unsigned ArgOffset = CCInfo.getNextStackOffset();
481 if (NumAllocated == 6)
482 ArgOffset += StackOffset;
485 ArgOffset = 68+4*NumAllocated;
488 // Remember the vararg offset for the va_start implementation.
489 FuncInfo->setVarArgsFrameOffset(ArgOffset);
491 std::vector<SDValue> OutChains;
493 for (; CurArgReg != ArgRegEnd; ++CurArgReg) {
494 unsigned VReg = RegInfo.createVirtualRegister(&SP::IntRegsRegClass);
495 MF.getRegInfo().addLiveIn(*CurArgReg, VReg);
496 SDValue Arg = DAG.getCopyFromReg(DAG.getRoot(), dl, VReg, MVT::i32);
498 int FrameIdx = MF.getFrameInfo()->CreateFixedObject(4, ArgOffset,
500 SDValue FIPtr = DAG.getFrameIndex(FrameIdx, MVT::i32);
502 OutChains.push_back(DAG.getStore(DAG.getRoot(), dl, Arg, FIPtr,
503 MachinePointerInfo(),
508 if (!OutChains.empty()) {
509 OutChains.push_back(Chain);
510 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
511 &OutChains[0], OutChains.size());
518 // Lower formal arguments for the 64 bit ABI.
519 SDValue SparcTargetLowering::
520 LowerFormalArguments_64(SDValue Chain,
521 CallingConv::ID CallConv,
523 const SmallVectorImpl<ISD::InputArg> &Ins,
526 SmallVectorImpl<SDValue> &InVals) const {
527 MachineFunction &MF = DAG.getMachineFunction();
529 // Analyze arguments according to CC_Sparc64.
530 SmallVector<CCValAssign, 16> ArgLocs;
531 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
532 getTargetMachine(), ArgLocs, *DAG.getContext());
533 CCInfo.AnalyzeFormalArguments(Ins, CC_Sparc64);
535 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
536 CCValAssign &VA = ArgLocs[i];
538 // This argument is passed in a register.
539 // All integer register arguments are promoted by the caller to i64.
541 // Create a virtual register for the promoted live-in value.
542 unsigned VReg = MF.addLiveIn(VA.getLocReg(),
543 getRegClassFor(VA.getLocVT()));
544 SDValue Arg = DAG.getCopyFromReg(Chain, DL, VReg, VA.getLocVT());
546 // Get the high bits for i32 struct elements.
547 if (VA.getValVT() == MVT::i32 && VA.needsCustom())
548 Arg = DAG.getNode(ISD::SRL, DL, VA.getLocVT(), Arg,
549 DAG.getConstant(32, MVT::i32));
551 // The caller promoted the argument, so insert an Assert?ext SDNode so we
552 // won't promote the value again in this function.
553 switch (VA.getLocInfo()) {
554 case CCValAssign::SExt:
555 Arg = DAG.getNode(ISD::AssertSext, DL, VA.getLocVT(), Arg,
556 DAG.getValueType(VA.getValVT()));
558 case CCValAssign::ZExt:
559 Arg = DAG.getNode(ISD::AssertZext, DL, VA.getLocVT(), Arg,
560 DAG.getValueType(VA.getValVT()));
566 // Truncate the register down to the argument type.
568 Arg = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Arg);
570 InVals.push_back(Arg);
574 // The registers are exhausted. This argument was passed on the stack.
575 assert(VA.isMemLoc());
576 // The CC_Sparc64_Full/Half functions compute stack offsets relative to the
577 // beginning of the arguments area at %fp+BIAS+128.
578 unsigned Offset = VA.getLocMemOffset() + 128;
579 unsigned ValSize = VA.getValVT().getSizeInBits() / 8;
580 // Adjust offset for extended arguments, SPARC is big-endian.
581 // The caller will have written the full slot with extended bytes, but we
582 // prefer our own extending loads.
584 Offset += 8 - ValSize;
585 int FI = MF.getFrameInfo()->CreateFixedObject(ValSize, Offset, true);
586 InVals.push_back(DAG.getLoad(VA.getValVT(), DL, Chain,
587 DAG.getFrameIndex(FI, getPointerTy()),
588 MachinePointerInfo::getFixedStack(FI),
589 false, false, false, 0));
595 SparcTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
596 SmallVectorImpl<SDValue> &InVals) const {
597 SelectionDAG &DAG = CLI.DAG;
598 DebugLoc &dl = CLI.DL;
599 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
600 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
601 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
602 SDValue Chain = CLI.Chain;
603 SDValue Callee = CLI.Callee;
604 bool &isTailCall = CLI.IsTailCall;
605 CallingConv::ID CallConv = CLI.CallConv;
606 bool isVarArg = CLI.IsVarArg;
608 // Sparc target does not yet support tail call optimization.
611 // Analyze operands of the call, assigning locations to each operand.
612 SmallVector<CCValAssign, 16> ArgLocs;
613 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
614 DAG.getTarget(), ArgLocs, *DAG.getContext());
615 CCInfo.AnalyzeCallOperands(Outs, CC_Sparc32);
617 // Get the size of the outgoing arguments stack space requirement.
618 unsigned ArgsSize = CCInfo.getNextStackOffset();
620 // Keep stack frames 8-byte aligned.
621 ArgsSize = (ArgsSize+7) & ~7;
623 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
625 //Create local copies for byval args.
626 SmallVector<SDValue, 8> ByValArgs;
627 for (unsigned i = 0, e = Outs.size(); i != e; ++i) {
628 ISD::ArgFlagsTy Flags = Outs[i].Flags;
629 if (!Flags.isByVal())
632 SDValue Arg = OutVals[i];
633 unsigned Size = Flags.getByValSize();
634 unsigned Align = Flags.getByValAlign();
636 int FI = MFI->CreateStackObject(Size, Align, false);
637 SDValue FIPtr = DAG.getFrameIndex(FI, getPointerTy());
638 SDValue SizeNode = DAG.getConstant(Size, MVT::i32);
640 Chain = DAG.getMemcpy(Chain, dl, FIPtr, Arg, SizeNode, Align,
642 (Size <= 32), //AlwaysInline if size <= 32
643 MachinePointerInfo(), MachinePointerInfo());
644 ByValArgs.push_back(FIPtr);
647 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(ArgsSize, true));
649 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
650 SmallVector<SDValue, 8> MemOpChains;
652 const unsigned StackOffset = 92;
653 bool hasStructRetAttr = false;
654 // Walk the register/memloc assignments, inserting copies/loads.
655 for (unsigned i = 0, realArgIdx = 0, byvalArgIdx = 0, e = ArgLocs.size();
658 CCValAssign &VA = ArgLocs[i];
659 SDValue Arg = OutVals[realArgIdx];
661 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
663 //Use local copy if it is a byval arg.
665 Arg = ByValArgs[byvalArgIdx++];
667 // Promote the value if needed.
668 switch (VA.getLocInfo()) {
669 default: llvm_unreachable("Unknown loc info!");
670 case CCValAssign::Full: break;
671 case CCValAssign::SExt:
672 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
674 case CCValAssign::ZExt:
675 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
677 case CCValAssign::AExt:
678 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
680 case CCValAssign::BCvt:
681 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
685 if (Flags.isSRet()) {
686 assert(VA.needsCustom());
687 // store SRet argument in %sp+64
688 SDValue StackPtr = DAG.getRegister(SP::O6, MVT::i32);
689 SDValue PtrOff = DAG.getIntPtrConstant(64);
690 PtrOff = DAG.getNode(ISD::ADD, dl, MVT::i32, StackPtr, PtrOff);
691 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
692 MachinePointerInfo(),
694 hasStructRetAttr = true;
698 if (VA.needsCustom()) {
699 assert(VA.getLocVT() == MVT::f64);
702 unsigned Offset = VA.getLocMemOffset() + StackOffset;
703 //if it is double-word aligned, just store.
704 if (Offset % 8 == 0) {
705 SDValue StackPtr = DAG.getRegister(SP::O6, MVT::i32);
706 SDValue PtrOff = DAG.getIntPtrConstant(Offset);
707 PtrOff = DAG.getNode(ISD::ADD, dl, MVT::i32, StackPtr, PtrOff);
708 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
709 MachinePointerInfo(),
715 SDValue StackPtr = DAG.CreateStackTemporary(MVT::f64, MVT::i32);
716 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl,
717 Arg, StackPtr, MachinePointerInfo(),
719 // Sparc is big-endian, so the high part comes first.
720 SDValue Hi = DAG.getLoad(MVT::i32, dl, Store, StackPtr,
721 MachinePointerInfo(), false, false, false, 0);
722 // Increment the pointer to the other half.
723 StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
724 DAG.getIntPtrConstant(4));
725 // Load the low part.
726 SDValue Lo = DAG.getLoad(MVT::i32, dl, Store, StackPtr,
727 MachinePointerInfo(), false, false, false, 0);
730 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Hi));
732 CCValAssign &NextVA = ArgLocs[++i];
733 if (NextVA.isRegLoc()) {
734 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), Lo));
736 //Store the low part in stack.
737 unsigned Offset = NextVA.getLocMemOffset() + StackOffset;
738 SDValue StackPtr = DAG.getRegister(SP::O6, MVT::i32);
739 SDValue PtrOff = DAG.getIntPtrConstant(Offset);
740 PtrOff = DAG.getNode(ISD::ADD, dl, MVT::i32, StackPtr, PtrOff);
741 MemOpChains.push_back(DAG.getStore(Chain, dl, Lo, PtrOff,
742 MachinePointerInfo(),
746 unsigned Offset = VA.getLocMemOffset() + StackOffset;
747 // Store the high part.
748 SDValue StackPtr = DAG.getRegister(SP::O6, MVT::i32);
749 SDValue PtrOff = DAG.getIntPtrConstant(Offset);
750 PtrOff = DAG.getNode(ISD::ADD, dl, MVT::i32, StackPtr, PtrOff);
751 MemOpChains.push_back(DAG.getStore(Chain, dl, Hi, PtrOff,
752 MachinePointerInfo(),
754 // Store the low part.
755 PtrOff = DAG.getIntPtrConstant(Offset+4);
756 PtrOff = DAG.getNode(ISD::ADD, dl, MVT::i32, StackPtr, PtrOff);
757 MemOpChains.push_back(DAG.getStore(Chain, dl, Lo, PtrOff,
758 MachinePointerInfo(),
764 // Arguments that can be passed on register must be kept at
767 if (VA.getLocVT() != MVT::f32) {
768 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
771 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Arg);
772 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
776 assert(VA.isMemLoc());
778 // Create a store off the stack pointer for this argument.
779 SDValue StackPtr = DAG.getRegister(SP::O6, MVT::i32);
780 SDValue PtrOff = DAG.getIntPtrConstant(VA.getLocMemOffset()+StackOffset);
781 PtrOff = DAG.getNode(ISD::ADD, dl, MVT::i32, StackPtr, PtrOff);
782 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
783 MachinePointerInfo(),
788 // Emit all stores, make sure the occur before any copies into physregs.
789 if (!MemOpChains.empty())
790 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
791 &MemOpChains[0], MemOpChains.size());
793 // Build a sequence of copy-to-reg nodes chained together with token
794 // chain and flag operands which copy the outgoing args into registers.
795 // The InFlag in necessary since all emitted instructions must be
798 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
799 unsigned Reg = RegsToPass[i].first;
800 // Remap I0->I7 -> O0->O7.
801 if (Reg >= SP::I0 && Reg <= SP::I7)
802 Reg = Reg-SP::I0+SP::O0;
804 Chain = DAG.getCopyToReg(Chain, dl, Reg, RegsToPass[i].second, InFlag);
805 InFlag = Chain.getValue(1);
808 unsigned SRetArgSize = (hasStructRetAttr)? getSRetArgSize(DAG, Callee):0;
810 // If the callee is a GlobalAddress node (quite common, every direct call is)
811 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
812 // Likewise ExternalSymbol -> TargetExternalSymbol.
813 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
814 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl, MVT::i32);
815 else if (ExternalSymbolSDNode *E = dyn_cast<ExternalSymbolSDNode>(Callee))
816 Callee = DAG.getTargetExternalSymbol(E->getSymbol(), MVT::i32);
818 // Returns a chain & a flag for retval copy to use
819 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
820 SmallVector<SDValue, 8> Ops;
821 Ops.push_back(Chain);
822 Ops.push_back(Callee);
823 if (hasStructRetAttr)
824 Ops.push_back(DAG.getTargetConstant(SRetArgSize, MVT::i32));
825 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
826 unsigned Reg = RegsToPass[i].first;
827 if (Reg >= SP::I0 && Reg <= SP::I7)
828 Reg = Reg-SP::I0+SP::O0;
830 Ops.push_back(DAG.getRegister(Reg, RegsToPass[i].second.getValueType()));
832 if (InFlag.getNode())
833 Ops.push_back(InFlag);
835 Chain = DAG.getNode(SPISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
836 InFlag = Chain.getValue(1);
838 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(ArgsSize, true),
839 DAG.getIntPtrConstant(0, true), InFlag);
840 InFlag = Chain.getValue(1);
842 // Assign locations to each value returned by this call.
843 SmallVector<CCValAssign, 16> RVLocs;
844 CCState RVInfo(CallConv, isVarArg, DAG.getMachineFunction(),
845 DAG.getTarget(), RVLocs, *DAG.getContext());
847 RVInfo.AnalyzeCallResult(Ins, RetCC_Sparc32);
849 // Copy all of the result registers out of their specified physreg.
850 for (unsigned i = 0; i != RVLocs.size(); ++i) {
851 unsigned Reg = RVLocs[i].getLocReg();
853 // Remap I0->I7 -> O0->O7.
854 if (Reg >= SP::I0 && Reg <= SP::I7)
855 Reg = Reg-SP::I0+SP::O0;
857 Chain = DAG.getCopyFromReg(Chain, dl, Reg,
858 RVLocs[i].getValVT(), InFlag).getValue(1);
859 InFlag = Chain.getValue(2);
860 InVals.push_back(Chain.getValue(0));
867 SparcTargetLowering::getSRetArgSize(SelectionDAG &DAG, SDValue Callee) const
869 const Function *CalleeFn = 0;
870 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
871 CalleeFn = dyn_cast<Function>(G->getGlobal());
872 } else if (ExternalSymbolSDNode *E =
873 dyn_cast<ExternalSymbolSDNode>(Callee)) {
874 const Function *Fn = DAG.getMachineFunction().getFunction();
875 const Module *M = Fn->getParent();
876 CalleeFn = M->getFunction(E->getSymbol());
882 assert(CalleeFn->hasStructRetAttr() &&
883 "Callee does not have the StructRet attribute.");
885 PointerType *Ty = cast<PointerType>(CalleeFn->arg_begin()->getType());
886 Type *ElementTy = Ty->getElementType();
887 return getDataLayout()->getTypeAllocSize(ElementTy);
890 //===----------------------------------------------------------------------===//
891 // TargetLowering Implementation
892 //===----------------------------------------------------------------------===//
894 /// IntCondCCodeToICC - Convert a DAG integer condition code to a SPARC ICC
896 static SPCC::CondCodes IntCondCCodeToICC(ISD::CondCode CC) {
898 default: llvm_unreachable("Unknown integer condition code!");
899 case ISD::SETEQ: return SPCC::ICC_E;
900 case ISD::SETNE: return SPCC::ICC_NE;
901 case ISD::SETLT: return SPCC::ICC_L;
902 case ISD::SETGT: return SPCC::ICC_G;
903 case ISD::SETLE: return SPCC::ICC_LE;
904 case ISD::SETGE: return SPCC::ICC_GE;
905 case ISD::SETULT: return SPCC::ICC_CS;
906 case ISD::SETULE: return SPCC::ICC_LEU;
907 case ISD::SETUGT: return SPCC::ICC_GU;
908 case ISD::SETUGE: return SPCC::ICC_CC;
912 /// FPCondCCodeToFCC - Convert a DAG floatingp oint condition code to a SPARC
914 static SPCC::CondCodes FPCondCCodeToFCC(ISD::CondCode CC) {
916 default: llvm_unreachable("Unknown fp condition code!");
918 case ISD::SETOEQ: return SPCC::FCC_E;
920 case ISD::SETUNE: return SPCC::FCC_NE;
922 case ISD::SETOLT: return SPCC::FCC_L;
924 case ISD::SETOGT: return SPCC::FCC_G;
926 case ISD::SETOLE: return SPCC::FCC_LE;
928 case ISD::SETOGE: return SPCC::FCC_GE;
929 case ISD::SETULT: return SPCC::FCC_UL;
930 case ISD::SETULE: return SPCC::FCC_ULE;
931 case ISD::SETUGT: return SPCC::FCC_UG;
932 case ISD::SETUGE: return SPCC::FCC_UGE;
933 case ISD::SETUO: return SPCC::FCC_U;
934 case ISD::SETO: return SPCC::FCC_O;
935 case ISD::SETONE: return SPCC::FCC_LG;
936 case ISD::SETUEQ: return SPCC::FCC_UE;
940 SparcTargetLowering::SparcTargetLowering(TargetMachine &TM)
941 : TargetLowering(TM, new TargetLoweringObjectFileELF()) {
942 Subtarget = &TM.getSubtarget<SparcSubtarget>();
944 // Set up the register classes.
945 addRegisterClass(MVT::i32, &SP::IntRegsRegClass);
946 addRegisterClass(MVT::f32, &SP::FPRegsRegClass);
947 addRegisterClass(MVT::f64, &SP::DFPRegsRegClass);
948 if (Subtarget->is64Bit())
949 addRegisterClass(MVT::i64, &SP::I64RegsRegClass);
951 // Turn FP extload into load/fextend
952 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
953 // Sparc doesn't have i1 sign extending load
954 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
955 // Turn FP truncstore into trunc + store.
956 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
958 // Custom legalize GlobalAddress nodes into LO/HI parts.
959 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
960 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
961 setOperationAction(ISD::ConstantPool , MVT::i32, Custom);
963 // Sparc doesn't have sext_inreg, replace them with shl/sra
964 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
965 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Expand);
966 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
968 // Sparc has no REM or DIVREM operations.
969 setOperationAction(ISD::UREM, MVT::i32, Expand);
970 setOperationAction(ISD::SREM, MVT::i32, Expand);
971 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
972 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
974 // Custom expand fp<->sint
975 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
976 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
979 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
980 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
982 setOperationAction(ISD::BITCAST, MVT::f32, Expand);
983 setOperationAction(ISD::BITCAST, MVT::i32, Expand);
985 // Sparc has no select or setcc: expand to SELECT_CC.
986 setOperationAction(ISD::SELECT, MVT::i32, Expand);
987 setOperationAction(ISD::SELECT, MVT::f32, Expand);
988 setOperationAction(ISD::SELECT, MVT::f64, Expand);
989 setOperationAction(ISD::SETCC, MVT::i32, Expand);
990 setOperationAction(ISD::SETCC, MVT::f32, Expand);
991 setOperationAction(ISD::SETCC, MVT::f64, Expand);
993 // Sparc doesn't have BRCOND either, it has BR_CC.
994 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
995 setOperationAction(ISD::BRIND, MVT::Other, Expand);
996 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
997 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
998 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
999 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
1001 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
1002 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
1003 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
1005 if (Subtarget->is64Bit()) {
1006 setOperationAction(ISD::BR_CC, MVT::i64, Custom);
1007 setOperationAction(ISD::SELECT_CC, MVT::i64, Custom);
1010 // FIXME: There are instructions available for ATOMIC_FENCE
1011 // on SparcV8 and later.
1012 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
1013 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
1015 setOperationAction(ISD::FSIN , MVT::f64, Expand);
1016 setOperationAction(ISD::FCOS , MVT::f64, Expand);
1017 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
1018 setOperationAction(ISD::FREM , MVT::f64, Expand);
1019 setOperationAction(ISD::FMA , MVT::f64, Expand);
1020 setOperationAction(ISD::FSIN , MVT::f32, Expand);
1021 setOperationAction(ISD::FCOS , MVT::f32, Expand);
1022 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
1023 setOperationAction(ISD::FREM , MVT::f32, Expand);
1024 setOperationAction(ISD::FMA , MVT::f32, Expand);
1025 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
1026 setOperationAction(ISD::CTTZ , MVT::i32, Expand);
1027 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
1028 setOperationAction(ISD::CTLZ , MVT::i32, Expand);
1029 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
1030 setOperationAction(ISD::ROTL , MVT::i32, Expand);
1031 setOperationAction(ISD::ROTR , MVT::i32, Expand);
1032 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
1033 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
1034 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
1035 setOperationAction(ISD::FPOW , MVT::f64, Expand);
1036 setOperationAction(ISD::FPOW , MVT::f32, Expand);
1038 setOperationAction(ISD::SHL_PARTS, MVT::i32, Expand);
1039 setOperationAction(ISD::SRA_PARTS, MVT::i32, Expand);
1040 setOperationAction(ISD::SRL_PARTS, MVT::i32, Expand);
1042 // FIXME: Sparc provides these multiplies, but we don't have them yet.
1043 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
1044 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
1046 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
1048 // VASTART needs to be custom lowered to use the VarArgsFrameIndex.
1049 setOperationAction(ISD::VASTART , MVT::Other, Custom);
1050 // VAARG needs to be lowered to not do unaligned accesses for doubles.
1051 setOperationAction(ISD::VAARG , MVT::Other, Custom);
1053 // Use the default implementation.
1054 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
1055 setOperationAction(ISD::VAEND , MVT::Other, Expand);
1056 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
1057 setOperationAction(ISD::STACKRESTORE , MVT::Other, Expand);
1058 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Custom);
1060 // No debug info support yet.
1061 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
1063 setStackPointerRegisterToSaveRestore(SP::O6);
1065 if (TM.getSubtarget<SparcSubtarget>().isV9())
1066 setOperationAction(ISD::CTPOP, MVT::i32, Legal);
1068 setMinFunctionAlignment(2);
1070 computeRegisterProperties();
1073 const char *SparcTargetLowering::getTargetNodeName(unsigned Opcode) const {
1076 case SPISD::CMPICC: return "SPISD::CMPICC";
1077 case SPISD::CMPFCC: return "SPISD::CMPFCC";
1078 case SPISD::BRICC: return "SPISD::BRICC";
1079 case SPISD::BRXCC: return "SPISD::BRXCC";
1080 case SPISD::BRFCC: return "SPISD::BRFCC";
1081 case SPISD::SELECT_ICC: return "SPISD::SELECT_ICC";
1082 case SPISD::SELECT_XCC: return "SPISD::SELECT_XCC";
1083 case SPISD::SELECT_FCC: return "SPISD::SELECT_FCC";
1084 case SPISD::Hi: return "SPISD::Hi";
1085 case SPISD::Lo: return "SPISD::Lo";
1086 case SPISD::FTOI: return "SPISD::FTOI";
1087 case SPISD::ITOF: return "SPISD::ITOF";
1088 case SPISD::CALL: return "SPISD::CALL";
1089 case SPISD::RET_FLAG: return "SPISD::RET_FLAG";
1090 case SPISD::GLOBAL_BASE_REG: return "SPISD::GLOBAL_BASE_REG";
1091 case SPISD::FLUSHW: return "SPISD::FLUSHW";
1095 /// isMaskedValueZeroForTargetNode - Return true if 'Op & Mask' is known to
1096 /// be zero. Op is expected to be a target specific node. Used by DAG
1098 void SparcTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
1101 const SelectionDAG &DAG,
1102 unsigned Depth) const {
1103 APInt KnownZero2, KnownOne2;
1104 KnownZero = KnownOne = APInt(KnownZero.getBitWidth(), 0);
1106 switch (Op.getOpcode()) {
1108 case SPISD::SELECT_ICC:
1109 case SPISD::SELECT_XCC:
1110 case SPISD::SELECT_FCC:
1111 DAG.ComputeMaskedBits(Op.getOperand(1), KnownZero, KnownOne, Depth+1);
1112 DAG.ComputeMaskedBits(Op.getOperand(0), KnownZero2, KnownOne2, Depth+1);
1113 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1114 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1116 // Only known if known in both the LHS and RHS.
1117 KnownOne &= KnownOne2;
1118 KnownZero &= KnownZero2;
1123 // Look at LHS/RHS/CC and see if they are a lowered setcc instruction. If so
1124 // set LHS/RHS and SPCC to the LHS/RHS of the setcc and SPCC to the condition.
1125 static void LookThroughSetCC(SDValue &LHS, SDValue &RHS,
1126 ISD::CondCode CC, unsigned &SPCC) {
1127 if (isa<ConstantSDNode>(RHS) &&
1128 cast<ConstantSDNode>(RHS)->isNullValue() &&
1130 (((LHS.getOpcode() == SPISD::SELECT_ICC ||
1131 LHS.getOpcode() == SPISD::SELECT_XCC) &&
1132 LHS.getOperand(3).getOpcode() == SPISD::CMPICC) ||
1133 (LHS.getOpcode() == SPISD::SELECT_FCC &&
1134 LHS.getOperand(3).getOpcode() == SPISD::CMPFCC)) &&
1135 isa<ConstantSDNode>(LHS.getOperand(0)) &&
1136 isa<ConstantSDNode>(LHS.getOperand(1)) &&
1137 cast<ConstantSDNode>(LHS.getOperand(0))->isOne() &&
1138 cast<ConstantSDNode>(LHS.getOperand(1))->isNullValue()) {
1139 SDValue CMPCC = LHS.getOperand(3);
1140 SPCC = cast<ConstantSDNode>(LHS.getOperand(2))->getZExtValue();
1141 LHS = CMPCC.getOperand(0);
1142 RHS = CMPCC.getOperand(1);
1146 SDValue SparcTargetLowering::LowerGlobalAddress(SDValue Op,
1147 SelectionDAG &DAG) const {
1148 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1149 // FIXME there isn't really any debug info here
1150 DebugLoc dl = Op.getDebugLoc();
1151 SDValue GA = DAG.getTargetGlobalAddress(GV, dl, MVT::i32);
1152 SDValue Hi = DAG.getNode(SPISD::Hi, dl, MVT::i32, GA);
1153 SDValue Lo = DAG.getNode(SPISD::Lo, dl, MVT::i32, GA);
1155 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
1156 return DAG.getNode(ISD::ADD, dl, MVT::i32, Lo, Hi);
1158 SDValue GlobalBase = DAG.getNode(SPISD::GLOBAL_BASE_REG, dl,
1160 SDValue RelAddr = DAG.getNode(ISD::ADD, dl, MVT::i32, Lo, Hi);
1161 SDValue AbsAddr = DAG.getNode(ISD::ADD, dl, MVT::i32,
1162 GlobalBase, RelAddr);
1163 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
1164 AbsAddr, MachinePointerInfo(), false, false, false, 0);
1167 SDValue SparcTargetLowering::LowerConstantPool(SDValue Op,
1168 SelectionDAG &DAG) const {
1169 ConstantPoolSDNode *N = cast<ConstantPoolSDNode>(Op);
1170 // FIXME there isn't really any debug info here
1171 DebugLoc dl = Op.getDebugLoc();
1172 const Constant *C = N->getConstVal();
1173 SDValue CP = DAG.getTargetConstantPool(C, MVT::i32, N->getAlignment());
1174 SDValue Hi = DAG.getNode(SPISD::Hi, dl, MVT::i32, CP);
1175 SDValue Lo = DAG.getNode(SPISD::Lo, dl, MVT::i32, CP);
1176 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
1177 return DAG.getNode(ISD::ADD, dl, MVT::i32, Lo, Hi);
1179 SDValue GlobalBase = DAG.getNode(SPISD::GLOBAL_BASE_REG, dl,
1181 SDValue RelAddr = DAG.getNode(ISD::ADD, dl, MVT::i32, Lo, Hi);
1182 SDValue AbsAddr = DAG.getNode(ISD::ADD, dl, MVT::i32,
1183 GlobalBase, RelAddr);
1184 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
1185 AbsAddr, MachinePointerInfo(), false, false, false, 0);
1188 static SDValue LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
1189 DebugLoc dl = Op.getDebugLoc();
1190 // Convert the fp value to integer in an FP register.
1191 assert(Op.getValueType() == MVT::i32);
1192 Op = DAG.getNode(SPISD::FTOI, dl, MVT::f32, Op.getOperand(0));
1193 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
1196 static SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1197 DebugLoc dl = Op.getDebugLoc();
1198 assert(Op.getOperand(0).getValueType() == MVT::i32);
1199 SDValue Tmp = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
1200 // Convert the int value to FP in an FP register.
1201 return DAG.getNode(SPISD::ITOF, dl, Op.getValueType(), Tmp);
1204 static SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG) {
1205 SDValue Chain = Op.getOperand(0);
1206 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
1207 SDValue LHS = Op.getOperand(2);
1208 SDValue RHS = Op.getOperand(3);
1209 SDValue Dest = Op.getOperand(4);
1210 DebugLoc dl = Op.getDebugLoc();
1211 unsigned Opc, SPCC = ~0U;
1213 // If this is a br_cc of a "setcc", and if the setcc got lowered into
1214 // an CMP[IF]CC/SELECT_[IF]CC pair, find the original compared values.
1215 LookThroughSetCC(LHS, RHS, CC, SPCC);
1217 // Get the condition flag.
1218 SDValue CompareFlag;
1219 if (LHS.getValueType().isInteger()) {
1220 EVT VTs[] = { LHS.getValueType(), MVT::Glue };
1221 SDValue Ops[2] = { LHS, RHS };
1222 CompareFlag = DAG.getNode(SPISD::CMPICC, dl, VTs, Ops, 2).getValue(1);
1223 if (SPCC == ~0U) SPCC = IntCondCCodeToICC(CC);
1224 // 32-bit compares use the icc flags, 64-bit uses the xcc flags.
1225 Opc = LHS.getValueType() == MVT::i32 ? SPISD::BRICC : SPISD::BRXCC;
1227 CompareFlag = DAG.getNode(SPISD::CMPFCC, dl, MVT::Glue, LHS, RHS);
1228 if (SPCC == ~0U) SPCC = FPCondCCodeToFCC(CC);
1231 return DAG.getNode(Opc, dl, MVT::Other, Chain, Dest,
1232 DAG.getConstant(SPCC, MVT::i32), CompareFlag);
1235 static SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) {
1236 SDValue LHS = Op.getOperand(0);
1237 SDValue RHS = Op.getOperand(1);
1238 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1239 SDValue TrueVal = Op.getOperand(2);
1240 SDValue FalseVal = Op.getOperand(3);
1241 DebugLoc dl = Op.getDebugLoc();
1242 unsigned Opc, SPCC = ~0U;
1244 // If this is a select_cc of a "setcc", and if the setcc got lowered into
1245 // an CMP[IF]CC/SELECT_[IF]CC pair, find the original compared values.
1246 LookThroughSetCC(LHS, RHS, CC, SPCC);
1248 SDValue CompareFlag;
1249 if (LHS.getValueType().isInteger()) {
1250 // subcc returns a value
1251 EVT VTs[] = { LHS.getValueType(), MVT::Glue };
1252 SDValue Ops[2] = { LHS, RHS };
1253 CompareFlag = DAG.getNode(SPISD::CMPICC, dl, VTs, Ops, 2).getValue(1);
1254 Opc = LHS.getValueType() == MVT::i32 ?
1255 SPISD::SELECT_ICC : SPISD::SELECT_XCC;
1256 if (SPCC == ~0U) SPCC = IntCondCCodeToICC(CC);
1258 CompareFlag = DAG.getNode(SPISD::CMPFCC, dl, MVT::Glue, LHS, RHS);
1259 Opc = SPISD::SELECT_FCC;
1260 if (SPCC == ~0U) SPCC = FPCondCCodeToFCC(CC);
1262 return DAG.getNode(Opc, dl, TrueVal.getValueType(), TrueVal, FalseVal,
1263 DAG.getConstant(SPCC, MVT::i32), CompareFlag);
1266 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG,
1267 const SparcTargetLowering &TLI) {
1268 MachineFunction &MF = DAG.getMachineFunction();
1269 SparcMachineFunctionInfo *FuncInfo = MF.getInfo<SparcMachineFunctionInfo>();
1271 // vastart just stores the address of the VarArgsFrameIndex slot into the
1272 // memory location argument.
1273 DebugLoc dl = Op.getDebugLoc();
1275 DAG.getNode(ISD::ADD, dl, MVT::i32,
1276 DAG.getRegister(SP::I6, MVT::i32),
1277 DAG.getConstant(FuncInfo->getVarArgsFrameOffset(),
1279 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1280 return DAG.getStore(Op.getOperand(0), dl, Offset, Op.getOperand(1),
1281 MachinePointerInfo(SV), false, false, 0);
1284 static SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) {
1285 SDNode *Node = Op.getNode();
1286 EVT VT = Node->getValueType(0);
1287 SDValue InChain = Node->getOperand(0);
1288 SDValue VAListPtr = Node->getOperand(1);
1289 const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
1290 DebugLoc dl = Node->getDebugLoc();
1291 SDValue VAList = DAG.getLoad(MVT::i32, dl, InChain, VAListPtr,
1292 MachinePointerInfo(SV), false, false, false, 0);
1293 // Increment the pointer, VAList, to the next vaarg
1294 SDValue NextPtr = DAG.getNode(ISD::ADD, dl, MVT::i32, VAList,
1295 DAG.getConstant(VT.getSizeInBits()/8,
1297 // Store the incremented VAList to the legalized pointer
1298 InChain = DAG.getStore(VAList.getValue(1), dl, NextPtr,
1299 VAListPtr, MachinePointerInfo(SV), false, false, 0);
1300 // Load the actual argument out of the pointer VAList, unless this is an
1303 return DAG.getLoad(VT, dl, InChain, VAList, MachinePointerInfo(),
1304 false, false, false, 0);
1306 // Otherwise, load it as i64, then do a bitconvert.
1307 SDValue V = DAG.getLoad(MVT::i64, dl, InChain, VAList, MachinePointerInfo(),
1308 false, false, false, 0);
1310 // Bit-Convert the value to f64.
1312 DAG.getNode(ISD::BITCAST, dl, MVT::f64, V),
1315 return DAG.getMergeValues(Ops, 2, dl);
1318 static SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) {
1319 SDValue Chain = Op.getOperand(0); // Legalize the chain.
1320 SDValue Size = Op.getOperand(1); // Legalize the size.
1321 DebugLoc dl = Op.getDebugLoc();
1323 unsigned SPReg = SP::O6;
1324 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, MVT::i32);
1325 SDValue NewSP = DAG.getNode(ISD::SUB, dl, MVT::i32, SP, Size); // Value
1326 Chain = DAG.getCopyToReg(SP.getValue(1), dl, SPReg, NewSP); // Output chain
1328 // The resultant pointer is actually 16 words from the bottom of the stack,
1329 // to provide a register spill area.
1330 SDValue NewVal = DAG.getNode(ISD::ADD, dl, MVT::i32, NewSP,
1331 DAG.getConstant(96, MVT::i32));
1332 SDValue Ops[2] = { NewVal, Chain };
1333 return DAG.getMergeValues(Ops, 2, dl);
1337 static SDValue getFLUSHW(SDValue Op, SelectionDAG &DAG) {
1338 DebugLoc dl = Op.getDebugLoc();
1339 SDValue Chain = DAG.getNode(SPISD::FLUSHW,
1340 dl, MVT::Other, DAG.getEntryNode());
1344 static SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
1345 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
1346 MFI->setFrameAddressIsTaken(true);
1348 EVT VT = Op.getValueType();
1349 DebugLoc dl = Op.getDebugLoc();
1350 unsigned FrameReg = SP::I6;
1352 uint64_t depth = Op.getConstantOperandVal(0);
1356 FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
1358 // flush first to make sure the windowed registers' values are in stack
1359 SDValue Chain = getFLUSHW(Op, DAG);
1360 FrameAddr = DAG.getCopyFromReg(Chain, dl, FrameReg, VT);
1362 for (uint64_t i = 0; i != depth; ++i) {
1363 SDValue Ptr = DAG.getNode(ISD::ADD,
1365 FrameAddr, DAG.getIntPtrConstant(56));
1366 FrameAddr = DAG.getLoad(MVT::i32, dl,
1369 MachinePointerInfo(), false, false, false, 0);
1375 static SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
1376 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
1377 MFI->setReturnAddressIsTaken(true);
1379 EVT VT = Op.getValueType();
1380 DebugLoc dl = Op.getDebugLoc();
1381 unsigned RetReg = SP::I7;
1383 uint64_t depth = Op.getConstantOperandVal(0);
1387 RetAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, RetReg, VT);
1389 // flush first to make sure the windowed registers' values are in stack
1390 SDValue Chain = getFLUSHW(Op, DAG);
1391 RetAddr = DAG.getCopyFromReg(Chain, dl, SP::I6, VT);
1393 for (uint64_t i = 0; i != depth; ++i) {
1394 SDValue Ptr = DAG.getNode(ISD::ADD,
1397 DAG.getIntPtrConstant((i == depth-1)?60:56));
1398 RetAddr = DAG.getLoad(MVT::i32, dl,
1401 MachinePointerInfo(), false, false, false, 0);
1407 SDValue SparcTargetLowering::
1408 LowerOperation(SDValue Op, SelectionDAG &DAG) const {
1409 switch (Op.getOpcode()) {
1410 default: llvm_unreachable("Should not custom lower this!");
1411 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
1412 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
1413 case ISD::GlobalTLSAddress:
1414 llvm_unreachable("TLS not implemented for Sparc.");
1415 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
1416 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
1417 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
1418 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
1419 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
1420 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
1421 case ISD::VASTART: return LowerVASTART(Op, DAG, *this);
1422 case ISD::VAARG: return LowerVAARG(Op, DAG);
1423 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
1428 SparcTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
1429 MachineBasicBlock *BB) const {
1430 const TargetInstrInfo &TII = *getTargetMachine().getInstrInfo();
1433 DebugLoc dl = MI->getDebugLoc();
1434 // Figure out the conditional branch opcode to use for this select_cc.
1435 switch (MI->getOpcode()) {
1436 default: llvm_unreachable("Unknown SELECT_CC!");
1437 case SP::SELECT_CC_Int_ICC:
1438 case SP::SELECT_CC_FP_ICC:
1439 case SP::SELECT_CC_DFP_ICC:
1440 BROpcode = SP::BCOND;
1442 case SP::SELECT_CC_Int_FCC:
1443 case SP::SELECT_CC_FP_FCC:
1444 case SP::SELECT_CC_DFP_FCC:
1445 BROpcode = SP::FBCOND;
1449 CC = (SPCC::CondCodes)MI->getOperand(3).getImm();
1451 // To "insert" a SELECT_CC instruction, we actually have to insert the diamond
1452 // control-flow pattern. The incoming instruction knows the destination vreg
1453 // to set, the condition code register to branch on, the true/false values to
1454 // select between, and a branch opcode to use.
1455 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1456 MachineFunction::iterator It = BB;
1463 // fallthrough --> copy0MBB
1464 MachineBasicBlock *thisMBB = BB;
1465 MachineFunction *F = BB->getParent();
1466 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
1467 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
1468 F->insert(It, copy0MBB);
1469 F->insert(It, sinkMBB);
1471 // Transfer the remainder of BB and its successor edges to sinkMBB.
1472 sinkMBB->splice(sinkMBB->begin(), BB,
1473 llvm::next(MachineBasicBlock::iterator(MI)),
1475 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
1477 // Add the true and fallthrough blocks as its successors.
1478 BB->addSuccessor(copy0MBB);
1479 BB->addSuccessor(sinkMBB);
1481 BuildMI(BB, dl, TII.get(BROpcode)).addMBB(sinkMBB).addImm(CC);
1484 // %FalseValue = ...
1485 // # fallthrough to sinkMBB
1488 // Update machine-CFG edges
1489 BB->addSuccessor(sinkMBB);
1492 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
1495 BuildMI(*BB, BB->begin(), dl, TII.get(SP::PHI), MI->getOperand(0).getReg())
1496 .addReg(MI->getOperand(2).getReg()).addMBB(copy0MBB)
1497 .addReg(MI->getOperand(1).getReg()).addMBB(thisMBB);
1499 MI->eraseFromParent(); // The pseudo instruction is gone now.
1503 //===----------------------------------------------------------------------===//
1504 // Sparc Inline Assembly Support
1505 //===----------------------------------------------------------------------===//
1507 /// getConstraintType - Given a constraint letter, return the type of
1508 /// constraint it is for this target.
1509 SparcTargetLowering::ConstraintType
1510 SparcTargetLowering::getConstraintType(const std::string &Constraint) const {
1511 if (Constraint.size() == 1) {
1512 switch (Constraint[0]) {
1514 case 'r': return C_RegisterClass;
1518 return TargetLowering::getConstraintType(Constraint);
1521 std::pair<unsigned, const TargetRegisterClass*>
1522 SparcTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
1524 if (Constraint.size() == 1) {
1525 switch (Constraint[0]) {
1527 return std::make_pair(0U, &SP::IntRegsRegClass);
1531 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
1535 SparcTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
1536 // The Sparc target isn't yet aware of offsets.