1 //===-- PPC32ISelPattern.cpp - A pattern matching inst selector for PPC32 -===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Nate Begeman and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines a pattern matching instruction selector for 32 bit PowerPC.
12 //===----------------------------------------------------------------------===//
15 #include "PowerPCInstrBuilder.h"
16 #include "PowerPCInstrInfo.h"
17 #include "PPC32RegisterInfo.h"
18 #include "llvm/Constants.h" // FIXME: REMOVE
19 #include "llvm/Function.h"
20 #include "llvm/CodeGen/MachineConstantPool.h" // FIXME: REMOVE
21 #include "llvm/CodeGen/MachineFunction.h"
22 #include "llvm/CodeGen/MachineFrameInfo.h"
23 #include "llvm/CodeGen/SelectionDAG.h"
24 #include "llvm/CodeGen/SelectionDAGISel.h"
25 #include "llvm/CodeGen/SSARegMap.h"
26 #include "llvm/Target/TargetData.h"
27 #include "llvm/Target/TargetLowering.h"
28 #include "llvm/Support/Debug.h"
29 #include "llvm/Support/MathExtras.h"
30 #include "llvm/ADT/Statistic.h"
35 //===----------------------------------------------------------------------===//
36 // PPC32TargetLowering - PPC32 Implementation of the TargetLowering interface
38 class PPC32TargetLowering : public TargetLowering {
39 int VarArgsFrameIndex; // FrameIndex for start of varargs area.
40 int ReturnAddrIndex; // FrameIndex for return slot.
42 PPC32TargetLowering(TargetMachine &TM) : TargetLowering(TM) {
43 // Set up the TargetLowering object.
45 // Set up the register classes.
46 addRegisterClass(MVT::i32, PPC32::GPRCRegisterClass);
47 addRegisterClass(MVT::f32, PPC32::FPRCRegisterClass);
48 addRegisterClass(MVT::f64, PPC32::FPRCRegisterClass);
50 computeRegisterProperties();
53 /// LowerArguments - This hook must be implemented to indicate how we should
54 /// lower the arguments for the specified function, into the specified DAG.
55 virtual std::vector<SDOperand>
56 LowerArguments(Function &F, SelectionDAG &DAG);
58 /// LowerCallTo - This hook lowers an abstract call to a function into an
60 virtual std::pair<SDOperand, SDOperand>
61 LowerCallTo(SDOperand Chain, const Type *RetTy, bool isVarArg,
62 SDOperand Callee, ArgListTy &Args, SelectionDAG &DAG);
64 virtual std::pair<SDOperand, SDOperand>
65 LowerVAStart(SDOperand Chain, SelectionDAG &DAG);
67 virtual std::pair<SDOperand,SDOperand>
68 LowerVAArgNext(bool isVANext, SDOperand Chain, SDOperand VAList,
69 const Type *ArgTy, SelectionDAG &DAG);
71 virtual std::pair<SDOperand, SDOperand>
72 LowerFrameReturnAddress(bool isFrameAddr, SDOperand Chain, unsigned Depth,
78 std::vector<SDOperand>
79 PPC32TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) {
81 // add beautiful description of PPC stack frame format, or at least some docs
83 MachineFunction &MF = DAG.getMachineFunction();
84 MachineFrameInfo *MFI = MF.getFrameInfo();
85 MachineBasicBlock& BB = MF.front();
86 std::vector<SDOperand> ArgValues;
88 // Due to the rather complicated nature of the PowerPC ABI, rather than a
89 // fixed size array of physical args, for the sake of simplicity let the STL
90 // handle tracking them for us.
91 std::vector<unsigned> argVR, argPR, argOp;
92 unsigned ArgOffset = 24;
93 unsigned GPR_remaining = 8;
94 unsigned FPR_remaining = 13;
95 unsigned GPR_idx = 0, FPR_idx = 0;
96 static const unsigned GPR[] = {
97 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
98 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
100 static const unsigned FPR[] = {
101 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
102 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
105 // Add DAG nodes to load the arguments... On entry to a function on PPC,
106 // the arguments start at offset 24, although they are likely to be passed
108 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
109 SDOperand newroot, argt;
111 bool needsLoad = false;
112 MVT::ValueType ObjectVT = getValueType(I->getType());
115 default: assert(0 && "Unhandled argument type!");
121 if (GPR_remaining > 0) {
122 BuildMI(&BB, PPC::IMPLICIT_DEF, 0, GPR[GPR_idx]);
124 MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i32));
125 argt = newroot = DAG.getCopyFromReg(virtReg, MVT::i32, DAG.getRoot());
126 if (ObjectVT != MVT::i32)
127 argt = DAG.getNode(ISD::TRUNCATE, ObjectVT, newroot);
128 argVR.push_back(virtReg);
129 argPR.push_back(GPR[GPR_idx]);
130 argOp.push_back(PPC::OR);
135 case MVT::i64: ObjSize = 8;
136 // FIXME: can split 64b load between reg/mem if it is last arg in regs
137 if (GPR_remaining > 1) {
138 BuildMI(&BB, PPC::IMPLICIT_DEF, 0, GPR[GPR_idx]);
139 BuildMI(&BB, PPC::IMPLICIT_DEF, 0, GPR[GPR_idx+1]);
140 SDOperand root = DAG.getRoot();
141 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
142 root, DAG.getConstant(1, MVT::i32));
143 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
144 root, DAG.getConstant(0, MVT::i32));
146 // Create the pair of virtual registers
147 MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i32));
148 unsigned virtReg = MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i32))-1;
150 // Copy the extracted halves into the virtual registers
151 SDOperand argHi = DAG.getCopyFromReg(virtReg, MVT::i32, Hi);
152 SDOperand argLo = DAG.getCopyFromReg(virtReg+1, MVT::i32, Lo);
154 // Build the outgoing arg thingy
155 argt = newroot = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, argLo, argHi);
156 argVR.push_back(virtReg); argVR.push_back(virtReg+1);
157 argPR.push_back(GPR[GPR_idx]); argPR.push_back(GPR[GPR_idx+1]);
158 argOp.push_back(PPC::OR); argOp.push_back(PPC::OR);
163 case MVT::f32: ObjSize = 4;
164 case MVT::f64: ObjSize = 8;
165 if (FPR_remaining > 0) {
166 BuildMI(&BB, PPC::IMPLICIT_DEF, 0, FPR[FPR_idx]);
168 MF.getSSARegMap()->createVirtualRegister(getRegClassFor(ObjectVT));
169 argt = newroot = DAG.getCopyFromReg(virtReg, ObjectVT, DAG.getRoot());
170 argVR.push_back(virtReg);
171 argPR.push_back(FPR[FPR_idx]);
172 argOp.push_back(PPC::FMR);
181 // We need to load the argument to a virtual register if we determined above
182 // that we ran out of physical registers of the appropriate type
184 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
185 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
186 argt = newroot = DAG.getLoad(ObjectVT, DAG.getEntryNode(), FIN);
189 // Every 4 bytes of argument space consumes one of the GPRs available for
191 if (GPR_remaining > 0) {
192 unsigned delta = (GPR_remaining > 1 && ObjSize == 8) ? 2 : 1;
193 GPR_remaining -= delta;
196 ArgOffset += ObjSize;
198 DAG.setRoot(newroot.getValue(1));
199 ArgValues.push_back(argt);
202 for (int i = 0, count = argVR.size(); i < count; ++i) {
203 if (argOp[i] == PPC::FMR)
204 BuildMI(&BB, argOp[i], 1, argVR[i]).addReg(argPR[i]);
206 BuildMI(&BB, argOp[i], 2, argVR[i]).addReg(argPR[i]).addReg(argPR[i]);
209 // If the function takes variable number of arguments, make a frame index for
210 // the start of the first vararg value... for expansion of llvm.va_start.
212 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset);
217 std::pair<SDOperand, SDOperand>
218 PPC32TargetLowering::LowerCallTo(SDOperand Chain,
219 const Type *RetTy, bool isVarArg,
220 SDOperand Callee, ArgListTy &Args, SelectionDAG &DAG) {
221 // args_to_use will accumulate outgoing args for the ISD::CALL case in
222 // SelectExpr to use to put the arguments in the appropriate registers.
223 std::vector<SDOperand> args_to_use;
225 // Count how many bytes are to be pushed on the stack, including the linkage
226 // area, and parameter passing area.
227 unsigned NumBytes = 24;
230 NumBytes = 0; // Save zero bytes.
232 for (unsigned i = 0, e = Args.size(); i != e; ++i)
233 switch (getValueType(Args[i].second)) {
234 default: assert(0 && "Unknown value type!");
248 // Just to be safe, we'll always reserve the full 24 bytes of linkage area
249 // plus 32 bytes of argument space in case any called code gets funky on us.
250 if (NumBytes < 56) NumBytes = 56;
252 // Adjust the stack pointer for the new arguments...
253 // These operations are automatically eliminated by the prolog/epilog pass
254 Chain = DAG.getNode(ISD::ADJCALLSTACKDOWN, MVT::Other, Chain,
255 DAG.getConstant(NumBytes, getPointerTy()));
257 // Set up a copy of the stack pointer for use loading and storing any
258 // arguments that may not fit in the registers available for argument
260 SDOperand StackPtr = DAG.getCopyFromReg(PPC::R1, MVT::i32,
263 // Figure out which arguments are going to go in registers, and which in
264 // memory. Also, if this is a vararg function, floating point operations
265 // must be stored to our stack, and loaded into integer regs as well, if
266 // any integer regs are available for argument passing.
267 unsigned ArgOffset = 24;
268 unsigned GPR_remaining = 8;
269 unsigned FPR_remaining = 13;
270 std::vector<SDOperand> Stores;
271 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
272 // PtrOff will be used to store the current argument to the stack if a
273 // register cannot be found for it.
274 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
275 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
276 MVT::ValueType ArgVT = getValueType(Args[i].second);
279 default: assert(0 && "Unexpected ValueType for argument!");
283 // Promote the integer to 32 bits. If the input type is signed use a
284 // sign extend, otherwise use a zero extend.
285 if (Args[i].second->isSigned())
286 Args[i].first =DAG.getNode(ISD::SIGN_EXTEND, MVT::i32, Args[i].first);
288 Args[i].first =DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Args[i].first);
291 if (GPR_remaining > 0) {
292 args_to_use.push_back(Args[i].first);
295 Stores.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
296 Args[i].first, PtrOff));
301 // If we have one free GPR left, we can place the upper half of the i64
302 // in it, and store the other half to the stack. If we have two or more
303 // free GPRs, then we can pass both halves of the i64 in registers.
304 if (GPR_remaining > 0) {
305 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
306 Args[i].first, DAG.getConstant(1, MVT::i32));
307 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
308 Args[i].first, DAG.getConstant(0, MVT::i32));
309 args_to_use.push_back(Hi);
310 if (GPR_remaining > 1) {
311 args_to_use.push_back(Lo);
314 SDOperand ConstFour = DAG.getConstant(4, getPointerTy());
315 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, PtrOff, ConstFour);
316 Stores.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
321 Stores.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
322 Args[i].first, PtrOff));
328 if (FPR_remaining > 0) {
330 // FIXME: Need FunctionType information so we can conditionally
331 // store only the non-fixed arguments in a vararg function.
332 Stores.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
333 Args[i].first, PtrOff));
334 // FIXME: Need a way to communicate to the ISD::CALL select code
335 // that a particular argument is non-fixed so that we can load them
336 // into the correct GPR to shadow the FPR
338 args_to_use.push_back(Args[i].first);
340 // If we have any FPRs remaining, we may also have GPRs remaining.
341 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
343 if (GPR_remaining > 0) --GPR_remaining;
344 if (GPR_remaining > 0 && MVT::f64 == ArgVT) --GPR_remaining;
346 Stores.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
347 Args[i].first, PtrOff));
349 ArgOffset += (ArgVT == MVT::f32) ? 4 : 8;
353 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, Stores);
356 std::vector<MVT::ValueType> RetVals;
357 MVT::ValueType RetTyVT = getValueType(RetTy);
358 if (RetTyVT != MVT::isVoid)
359 RetVals.push_back(RetTyVT);
360 RetVals.push_back(MVT::Other);
362 SDOperand TheCall = SDOperand(DAG.getCall(RetVals,
363 Chain, Callee, args_to_use), 0);
364 Chain = TheCall.getValue(RetTyVT != MVT::isVoid);
365 Chain = DAG.getNode(ISD::ADJCALLSTACKUP, MVT::Other, Chain,
366 DAG.getConstant(NumBytes, getPointerTy()));
367 return std::make_pair(TheCall, Chain);
370 std::pair<SDOperand, SDOperand>
371 PPC32TargetLowering::LowerVAStart(SDOperand Chain, SelectionDAG &DAG) {
372 //vastart just returns the address of the VarArgsFrameIndex slot.
373 return std::make_pair(DAG.getFrameIndex(VarArgsFrameIndex, MVT::i32), Chain);
376 std::pair<SDOperand,SDOperand> PPC32TargetLowering::
377 LowerVAArgNext(bool isVANext, SDOperand Chain, SDOperand VAList,
378 const Type *ArgTy, SelectionDAG &DAG) {
379 MVT::ValueType ArgVT = getValueType(ArgTy);
382 Result = DAG.getLoad(ArgVT, DAG.getEntryNode(), VAList);
385 if (ArgVT == MVT::i32 || ArgVT == MVT::f32)
388 assert((ArgVT == MVT::i64 || ArgVT == MVT::f64) &&
389 "Other types should have been promoted for varargs!");
392 Result = DAG.getNode(ISD::ADD, VAList.getValueType(), VAList,
393 DAG.getConstant(Amt, VAList.getValueType()));
395 return std::make_pair(Result, Chain);
399 std::pair<SDOperand, SDOperand> PPC32TargetLowering::
400 LowerFrameReturnAddress(bool isFrameAddress, SDOperand Chain, unsigned Depth,
407 //===--------------------------------------------------------------------===//
408 /// ISel - PPC32 specific code to select PPC32 machine instructions for
409 /// SelectionDAG operations.
410 //===--------------------------------------------------------------------===//
411 class ISel : public SelectionDAGISel {
414 PPC32TargetLowering PPC32Lowering;
416 /// ExprMap - As shared expressions are codegen'd, we keep track of which
417 /// vreg the value is produced in, so we only emit one copy of each compiled
419 std::map<SDOperand, unsigned> ExprMap;
421 unsigned GlobalBaseReg;
422 bool GlobalBaseInitialized;
425 ISel(TargetMachine &TM) : SelectionDAGISel(PPC32Lowering), PPC32Lowering(TM)
428 /// runOnFunction - Override this function in order to reset our per-function
430 virtual bool runOnFunction(Function &Fn) {
431 // Make sure we re-emit a set of the global base reg if necessary
432 GlobalBaseInitialized = false;
433 return SelectionDAGISel::runOnFunction(Fn);
436 /// InstructionSelectBasicBlock - This callback is invoked by
437 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
438 virtual void InstructionSelectBasicBlock(SelectionDAG &DAG) {
440 // Codegen the basic block.
441 Select(DAG.getRoot());
443 // Clear state used for selection.
447 unsigned ISel::getGlobalBaseReg();
448 unsigned SelectExpr(SDOperand N);
449 unsigned SelectExprFP(SDOperand N, unsigned Result);
450 void Select(SDOperand N);
452 void SelectAddr(SDOperand N, unsigned& Reg, int& offset);
453 void SelectBranchCC(SDOperand N);
456 /// canUseAsImmediateForOpcode - This method returns a value indicating whether
457 /// the ConstantSDNode N can be used as an immediate to Opcode. The return
458 /// values are either 0, 1 or 2. 0 indicates that either N is not a
459 /// ConstantSDNode, or is not suitable for use by that opcode. A return value
460 /// of 1 indicates that the constant may be used in normal immediate form. A
461 /// return value of 2 indicates that the constant may be used in shifted
462 /// immediate form. If the return value is nonzero, the constant value is
465 static unsigned canUseAsImmediateForOpcode(SDOperand N, unsigned Opcode,
467 if (N.getOpcode() != ISD::Constant) return 0;
469 int v = (int)cast<ConstantSDNode>(N)->getSignExtended();
474 if (v <= 32767 && v >= -32768) { Imm = v & 0xFFFF; return 1; }
475 if ((v & 0x0000FFFF) == 0) { Imm = v >> 16; return 2; }
480 if (v >= 0 && v <= 65535) { Imm = v & 0xFFFF; return 1; }
481 if ((v & 0x0000FFFF) == 0) { Imm = v >> 16; return 2; }
484 if (v <= 32767 && v >= -32768) { Imm = v & 0xFFFF; return 1; }
491 /// getGlobalBaseReg - Output the instructions required to put the
492 /// base address to use for accessing globals into a register.
494 unsigned ISel::getGlobalBaseReg() {
495 if (!GlobalBaseInitialized) {
496 // Insert the set of GlobalBaseReg into the first MBB of the function
497 MachineBasicBlock &FirstMBB = BB->getParent()->front();
498 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
499 GlobalBaseReg = MakeReg(MVT::i32);
500 BuildMI(FirstMBB, MBBI, PPC::MovePCtoLR, 0, PPC::LR);
501 BuildMI(FirstMBB, MBBI, PPC::MFLR, 1, GlobalBaseReg).addReg(PPC::LR);
502 GlobalBaseInitialized = true;
504 return GlobalBaseReg;
507 //Check to see if the load is a constant offset from a base register
508 void ISel::SelectAddr(SDOperand N, unsigned& Reg, int& offset)
515 void ISel::SelectBranchCC(SDOperand N)
517 assert(N.getOpcode() == ISD::BRCOND && "Not a BranchCC???");
518 MachineBasicBlock *Dest =
519 cast<BasicBlockSDNode>(N.getOperand(2))->getBasicBlock();
522 Select(N.getOperand(0)); //chain
523 SDOperand CC = N.getOperand(1);
525 //Giveup and do the stupid thing
526 unsigned Tmp1 = SelectExpr(CC);
527 BuildMI(BB, PPC::BNE, 2).addReg(Tmp1).addMBB(Dest);
531 unsigned ISel::SelectExprFP(SDOperand N, unsigned Result)
533 unsigned Tmp1, Tmp2, Tmp3;
535 SDNode *Node = N.Val;
536 MVT::ValueType DestType = N.getValueType();
537 unsigned opcode = N.getOpcode();
542 assert(0 && "Node not handled!\n");
548 assert (DestType == MVT::f32 &&
549 N.getOperand(0).getValueType() == MVT::f64 &&
550 "only f64 to f32 conversion supported here");
551 Tmp1 = SelectExpr(N.getOperand(0));
552 BuildMI(BB, PPC::FRSP, 1, Result).addReg(Tmp1);
556 assert (DestType == MVT::f64 &&
557 N.getOperand(0).getValueType() == MVT::f32 &&
558 "only f32 to f64 conversion supported here");
559 Tmp1 = SelectExpr(N.getOperand(0));
560 BuildMI(BB, PPC::FMR, 1, Result).addReg(Tmp1);
563 case ISD::CopyFromReg:
565 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
566 Tmp1 = dyn_cast<RegSDNode>(Node)->getReg();
567 BuildMI(BB, PPC::FMR, 1, Result).addReg(Tmp1);
572 MVT::ValueType TypeBeingLoaded = (ISD::LOAD == opcode) ?
573 Node->getValueType(0) : cast<MVTSDNode>(Node)->getExtraValueType();
575 // Make sure we generate both values.
577 ExprMap[N.getValue(1)] = 1; // Generate the token
579 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
581 SDOperand Chain = N.getOperand(0);
582 SDOperand Address = N.getOperand(1);
585 switch (TypeBeingLoaded) {
586 default: assert(0 && "Cannot fp load this type!");
587 case MVT::f32: Opc = PPC::LFS; break;
588 case MVT::f64: Opc = PPC::LFD; break;
591 if(Address.getOpcode() == ISD::FrameIndex) {
592 BuildMI(BB, Opc, 2, Result)
593 .addFrameIndex(cast<FrameIndexSDNode>(Address)->getIndex())
597 SelectAddr(Address, Tmp1, offset);
598 BuildMI(BB, Opc, 2, Result).addSImm(offset).addReg(Tmp1);
603 case ISD::ConstantFP:
604 assert(0 && "ISD::ConstantFP Unimplemented");
612 case ISD::MUL: Opc = DestType == MVT::f64 ? PPC::FMUL : PPC::FMULS; break;
613 case ISD::ADD: Opc = DestType == MVT::f64 ? PPC::FADD : PPC::FADDS; break;
614 case ISD::SUB: Opc = DestType == MVT::f64 ? PPC::FSUB : PPC::FSUBS; break;
615 case ISD::SDIV: Opc = DestType == MVT::f64 ? PPC::FDIV : PPC::FDIVS; break;
618 Tmp1 = SelectExpr(N.getOperand(0));
619 Tmp2 = SelectExpr(N.getOperand(1));
620 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
623 case ISD::UINT_TO_FP:
624 case ISD::SINT_TO_FP:
627 assert(0 && "should not get here");
631 unsigned ISel::SelectExpr(SDOperand N) {
633 unsigned Tmp1, Tmp2, Tmp3;
635 unsigned opcode = N.getOpcode();
637 SDNode *Node = N.Val;
638 MVT::ValueType DestType = N.getValueType();
640 unsigned &Reg = ExprMap[N];
643 if (N.getOpcode() != ISD::CALL && N.getOpcode() != ISD::ADD_PARTS &&
644 N.getOpcode() != ISD::SUB_PARTS)
645 Reg = Result = (N.getValueType() != MVT::Other) ?
646 MakeReg(N.getValueType()) : 1;
648 // If this is a call instruction, make sure to prepare ALL of the result
649 // values as well as the chain.
650 if (N.getOpcode() == ISD::CALL) {
651 if (Node->getNumValues() == 1)
652 Reg = Result = 1; // Void call, just a chain.
654 Result = MakeReg(Node->getValueType(0));
655 ExprMap[N.getValue(0)] = Result;
656 for (unsigned i = 1, e = N.Val->getNumValues()-1; i != e; ++i)
657 ExprMap[N.getValue(i)] = MakeReg(Node->getValueType(i));
658 ExprMap[SDOperand(Node, Node->getNumValues()-1)] = 1;
661 Result = MakeReg(Node->getValueType(0));
662 ExprMap[N.getValue(0)] = Result;
663 for (unsigned i = 1, e = N.Val->getNumValues(); i != e; ++i)
664 ExprMap[N.getValue(i)] = MakeReg(Node->getValueType(i));
668 if (DestType == MVT::f64 || DestType == MVT::f32)
669 return SelectExprFP(N, Result);
674 assert(0 && "Node not handled!\n");
676 case ISD::DYNAMIC_STACKALLOC:
677 // Generate both result values. FIXME: Need a better commment here?
679 ExprMap[N.getValue(1)] = 1;
681 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
683 // FIXME: We are currently ignoring the requested alignment for handling
684 // greater than the stack alignment. This will need to be revisited at some
685 // point. Align = N.getOperand(2);
686 if (!isa<ConstantSDNode>(N.getOperand(2)) ||
687 cast<ConstantSDNode>(N.getOperand(2))->getValue() != 0) {
688 std::cerr << "Cannot allocate stack object with greater alignment than"
689 << " the stack alignment yet!";
692 Select(N.getOperand(0));
693 Tmp1 = SelectExpr(N.getOperand(1));
694 // Subtract size from stack pointer, thereby allocating some space.
695 BuildMI(BB, PPC::SUBF, 2, PPC::R1).addReg(Tmp1).addReg(PPC::R1);
696 // Put a pointer to the space into the result register by copying the SP
697 BuildMI(BB, PPC::OR, 2, Result).addReg(PPC::R1).addReg(PPC::R1);
700 case ISD::ConstantPool:
701 Tmp1 = cast<ConstantPoolSDNode>(N)->getIndex();
702 Tmp2 = MakeReg(MVT::i32);
703 BuildMI(BB, PPC::LOADHiAddr, 2, Tmp2).addReg(getGlobalBaseReg())
704 .addConstantPoolIndex(Tmp1);
705 BuildMI(BB, PPC::LA, 2, Result).addReg(Tmp2).addConstantPoolIndex(Tmp1);
708 case ISD::FrameIndex:
709 assert(0 && "ISD::FrameIndex Unimplemented");
712 case ISD::GlobalAddress: {
713 GlobalValue *GV = cast<GlobalAddressSDNode>(N)->getGlobal();
714 Tmp1 = MakeReg(MVT::i32);
715 BuildMI(BB, PPC::LOADHiAddr, 2, Tmp1).addReg(getGlobalBaseReg())
716 .addGlobalAddress(GV);
717 if (GV->hasWeakLinkage() || GV->isExternal()) {
718 BuildMI(BB, PPC::LWZ, 2, Result).addGlobalAddress(GV).addReg(Tmp1);
720 BuildMI(BB, PPC::LA, 2, Result).addReg(Tmp1).addGlobalAddress(GV);
728 case ISD::SEXTLOAD: {
729 bool sext = (ISD::SEXTLOAD == opcode);
730 bool byte = (MVT::i8 == Node->getValueType(0));
731 MVT::ValueType TypeBeingLoaded = (ISD::LOAD == opcode) ?
732 Node->getValueType(0) : cast<MVTSDNode>(Node)->getExtraValueType();
734 // Make sure we generate both values.
736 ExprMap[N.getValue(1)] = 1; // Generate the token
738 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
740 SDOperand Chain = N.getOperand(0);
741 SDOperand Address = N.getOperand(1);
744 switch (TypeBeingLoaded) {
745 default: assert(0 && "Cannot load this type!");
746 case MVT::i1: Opc = PPC::LBZ; break;
747 case MVT::i8: Opc = PPC::LBZ; break;
748 case MVT::i16: Opc = sext ? PPC::LHA : PPC::LHZ; break;
749 case MVT::i32: Opc = PPC::LWZ; break;
752 // Since there's no load byte & sign extend instruction we have to split
753 // byte SEXTLOADs into lbz + extsb. This requires we make a temp register.
756 Result = MakeReg(MVT::i32);
758 if(Address.getOpcode() == ISD::FrameIndex) {
759 BuildMI(BB, Opc, 2, Result)
760 .addFrameIndex(cast<FrameIndexSDNode>(Address)->getIndex())
764 SelectAddr(Address, Tmp1, offset);
765 BuildMI(BB, Opc, 2, Result).addSImm(offset).addReg(Tmp1);
768 BuildMI(BB, PPC::EXTSB, 1, Tmp3).addReg(Result);
775 // Lower the chain for this call.
776 Select(N.getOperand(0));
777 ExprMap[N.getValue(Node->getNumValues()-1)] = 1;
779 // get the virtual reg for each argument
780 std::vector<unsigned> VRegs;
781 for(int i = 2, e = Node->getNumOperands(); i < e; ++i)
782 VRegs.push_back(SelectExpr(N.getOperand(i)));
784 // The ABI specifies that the first 32 bytes of args may be passed in GPRs,
785 // and that 13 FPRs may also be used for passing any floating point args.
786 int GPR_remaining = 8, FPR_remaining = 13;
787 unsigned GPR_idx = 0, FPR_idx = 0;
788 static const unsigned GPR[] = {
789 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
790 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
792 static const unsigned FPR[] = {
793 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6,
794 PPC::F7, PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12,
798 // move the vregs into the appropriate architected register or stack slot
799 for(int i = 0, e = VRegs.size(); i < e; ++i) {
800 unsigned OperandType = N.getOperand(i+2).getValueType();
801 switch(OperandType) {
804 N.getOperand(i).Val->dump();
805 std::cerr << "Type for " << i << " is: " <<
806 N.getOperand(i+2).getValueType() << "\n";
807 assert(0 && "Unknown value type for call");
812 if (GPR_remaining > 0)
813 BuildMI(BB, PPC::OR, 2, GPR[GPR_idx]).addReg(VRegs[i])
818 if (FPR_remaining > 0) {
819 BuildMI(BB, PPC::FMR, 1, FPR[FPR_idx]).addReg(VRegs[i]);
825 // All arguments consume GPRs available for argument passing
826 if (GPR_remaining > 0) {
830 if (MVT::f64 == OperandType && GPR_remaining > 0) {
836 // Emit the correct call instruction based on the type of symbol called.
837 if (GlobalAddressSDNode *GASD =
838 dyn_cast<GlobalAddressSDNode>(N.getOperand(1))) {
839 BuildMI(BB, PPC::CALLpcrel, 1).addGlobalAddress(GASD->getGlobal(), true);
840 } else if (ExternalSymbolSDNode *ESSDN =
841 dyn_cast<ExternalSymbolSDNode>(N.getOperand(1))) {
842 BuildMI(BB, PPC::CALLpcrel, 1).addExternalSymbol(ESSDN->getSymbol(), true);
844 Tmp1 = SelectExpr(N.getOperand(1));
845 BuildMI(BB, PPC::OR, 2, PPC::R12).addReg(Tmp1).addReg(Tmp1);
846 BuildMI(BB, PPC::MTCTR, 1).addReg(PPC::R12);
847 BuildMI(BB, PPC::CALLindirect, 3).addImm(20).addImm(0).addReg(PPC::R12);
850 switch (Node->getValueType(0)) {
851 default: assert(0 && "Unknown value type for call result!");
852 case MVT::Other: return 1;
857 BuildMI(BB, PPC::OR, 2, Result).addReg(PPC::R3).addReg(PPC::R3);
858 if (Node->getValueType(1) == MVT::i32)
859 BuildMI(BB, PPC::OR, 2, Result+1).addReg(PPC::R4).addReg(PPC::R4);
863 BuildMI(BB, PPC::FMR, 1, Result).addReg(PPC::F1);
866 return Result+N.ResNo;
869 case ISD::SIGN_EXTEND:
870 case ISD::SIGN_EXTEND_INREG:
871 Tmp1 = SelectExpr(N.getOperand(0));
872 switch(cast<MVTSDNode>(Node)->getExtraValueType()) {
873 default: Node->dump(); assert(0 && "Unhandled SIGN_EXTEND type"); break;
875 BuildMI(BB, PPC::EXTSH, 1, Result).addReg(Tmp1);
878 BuildMI(BB, PPC::EXTSB, 1, Result).addReg(Tmp1);
883 case ISD::ZERO_EXTEND_INREG:
884 Tmp1 = SelectExpr(N.getOperand(0));
885 switch(cast<MVTSDNode>(Node)->getExtraValueType()) {
886 default: Node->dump(); assert(0 && "Unhandled ZERO_EXTEND type"); break;
887 case MVT::i16: Tmp2 = 16; break;
888 case MVT::i8: Tmp2 = 24; break;
889 case MVT::i1: Tmp2 = 31; break;
891 BuildMI(BB, PPC::RLWINM, 5, Result).addReg(Tmp1).addImm(0).addImm(0)
892 .addImm(Tmp2).addImm(31);
895 case ISD::CopyFromReg:
897 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
898 Tmp1 = dyn_cast<RegSDNode>(Node)->getReg();
899 BuildMI(BB, PPC::OR, 2, Result).addReg(Tmp1).addReg(Tmp1);
903 Tmp1 = SelectExpr(N.getOperand(0));
904 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
905 Tmp2 = CN->getValue() & 0x1F;
906 BuildMI(BB, PPC::RLWINM, 5, Result).addReg(Tmp1).addImm(Tmp2).addImm(0)
909 Tmp2 = SelectExpr(N.getOperand(1));
910 BuildMI(BB, PPC::SLW, 2, Result).addReg(Tmp1).addReg(Tmp2);
915 Tmp1 = SelectExpr(N.getOperand(0));
916 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
917 Tmp2 = CN->getValue() & 0x1F;
918 BuildMI(BB, PPC::RLWINM, 5, Result).addReg(Tmp1).addImm(32-Tmp2)
919 .addImm(Tmp2).addImm(31);
921 Tmp2 = SelectExpr(N.getOperand(1));
922 BuildMI(BB, PPC::SRW, 2, Result).addReg(Tmp1).addReg(Tmp2);
927 Tmp1 = SelectExpr(N.getOperand(0));
928 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
929 Tmp2 = CN->getValue() & 0x1F;
930 BuildMI(BB, PPC::SRAWI, 2, Result).addReg(Tmp1).addImm(Tmp2);
932 Tmp2 = SelectExpr(N.getOperand(1));
933 BuildMI(BB, PPC::SRAW, 2, Result).addReg(Tmp1).addReg(Tmp2);
938 assert (DestType == MVT::i32 && "Only do arithmetic on i32s!");
939 Tmp1 = SelectExpr(N.getOperand(0));
940 switch(canUseAsImmediateForOpcode(N.getOperand(1), opcode, Tmp2)) {
941 default: assert(0 && "unhandled result code");
942 case 0: // No immediate
943 Tmp2 = SelectExpr(N.getOperand(1));
944 BuildMI(BB, PPC::ADD, 2, Result).addReg(Tmp1).addReg(Tmp2);
946 case 1: // Low immediate
947 BuildMI(BB, PPC::ADDI, 2, Result).addReg(Tmp1).addSImm(Tmp2);
949 case 2: // Shifted immediate
950 BuildMI(BB, PPC::ADDIS, 2, Result).addReg(Tmp1).addSImm(Tmp2);
958 assert (DestType == MVT::i32 && "Only do arithmetic on i32s!");
959 Tmp1 = SelectExpr(N.getOperand(0));
960 switch(canUseAsImmediateForOpcode(N.getOperand(1), opcode, Tmp2)) {
961 default: assert(0 && "unhandled result code");
962 case 0: // No immediate
963 Tmp2 = SelectExpr(N.getOperand(1));
965 case ISD::AND: Opc = PPC::AND; break;
966 case ISD::OR: Opc = PPC::OR; break;
967 case ISD::XOR: Opc = PPC::XOR; break;
969 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
971 case 1: // Low immediate
973 case ISD::AND: Opc = PPC::ANDIo; break;
974 case ISD::OR: Opc = PPC::ORI; break;
975 case ISD::XOR: Opc = PPC::XORI; break;
977 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addImm(Tmp2);
979 case 2: // Shifted immediate
981 case ISD::AND: Opc = PPC::ANDISo; break;
982 case ISD::OR: Opc = PPC::ORIS; break;
983 case ISD::XOR: Opc = PPC::XORIS; break;
985 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addImm(Tmp2);
991 assert (DestType == MVT::i32 && "Only do arithmetic on i32s!");
992 Tmp1 = SelectExpr(N.getOperand(0));
993 Tmp2 = SelectExpr(N.getOperand(1));
994 BuildMI(BB, PPC::SUBF, 2, Result).addReg(Tmp2).addReg(Tmp1);
998 assert (DestType == MVT::i32 && "Only do arithmetic on i32s!");
999 Tmp1 = SelectExpr(N.getOperand(0));
1000 if (1 == canUseAsImmediateForOpcode(N.getOperand(1), opcode, Tmp2))
1001 BuildMI(BB, PPC::MULLI, 2, Result).addReg(Tmp1).addSImm(Tmp2);
1003 Tmp2 = SelectExpr(N.getOperand(1));
1004 BuildMI(BB, PPC::MULLW, 2, Result).addReg(Tmp1).addReg(Tmp2);
1008 case ISD::ADD_PARTS:
1009 case ISD::SUB_PARTS: {
1010 assert(N.getNumOperands() == 4 && N.getValueType() == MVT::i32 &&
1011 "Not an i64 add/sub!");
1012 // Emit all of the operands.
1013 std::vector<unsigned> InVals;
1014 for (unsigned i = 0, e = N.getNumOperands(); i != e; ++i)
1015 InVals.push_back(SelectExpr(N.getOperand(i)));
1016 if (N.getOpcode() == ISD::ADD_PARTS) {
1017 BuildMI(BB, PPC::ADDC, 2, Result+1).addReg(InVals[1]).addReg(InVals[3]);
1018 BuildMI(BB, PPC::ADDE, 2, Result).addReg(InVals[0]).addReg(InVals[2]);
1020 BuildMI(BB, PPC::SUBFC, 2, Result+1).addReg(InVals[1]).addReg(InVals[3]);
1021 BuildMI(BB, PPC::SUBFE, 2, Result).addReg(InVals[0]).addReg(InVals[2]);
1023 return Result+N.ResNo;
1032 case ISD::FP_TO_UINT:
1033 case ISD::FP_TO_SINT:
1043 switch (N.getValueType()) {
1044 default: assert(0 && "Cannot use constants of this type!");
1046 BuildMI(BB, PPC::LI, 1, Result)
1047 .addSImm(!cast<ConstantSDNode>(N)->isNullValue());
1051 int v = (int)cast<ConstantSDNode>(N)->getSignExtended();
1052 if (v < 32768 && v >= -32768) {
1053 BuildMI(BB, PPC::LI, 1, Result).addSImm(v);
1055 Tmp1 = MakeReg(MVT::i32);
1056 BuildMI(BB, PPC::LIS, 1, Tmp1).addSImm(v >> 16);
1057 BuildMI(BB, PPC::ORI, 2, Result).addReg(Tmp1).addImm(v & 0xFFFF);
1067 void ISel::Select(SDOperand N) {
1068 unsigned Tmp1, Tmp2, Opc;
1069 unsigned opcode = N.getOpcode();
1071 if (!ExprMap.insert(std::make_pair(N, 1)).second)
1072 return; // Already selected.
1074 SDNode *Node = N.Val;
1076 switch (Node->getOpcode()) {
1078 Node->dump(); std::cerr << "\n";
1079 assert(0 && "Node not handled yet!");
1080 case ISD::EntryToken: return; // Noop
1081 case ISD::TokenFactor:
1082 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
1083 Select(Node->getOperand(i));
1085 case ISD::ADJCALLSTACKDOWN:
1086 case ISD::ADJCALLSTACKUP:
1087 Select(N.getOperand(0));
1088 Tmp1 = cast<ConstantSDNode>(N.getOperand(1))->getValue();
1089 Opc = N.getOpcode() == ISD::ADJCALLSTACKDOWN ? PPC::ADJCALLSTACKDOWN :
1090 PPC::ADJCALLSTACKUP;
1091 BuildMI(BB, Opc, 1).addImm(Tmp1);
1094 MachineBasicBlock *Dest =
1095 cast<BasicBlockSDNode>(N.getOperand(1))->getBasicBlock();
1096 Select(N.getOperand(0));
1097 BuildMI(BB, PPC::B, 1).addMBB(Dest);
1103 case ISD::CopyToReg:
1104 Select(N.getOperand(0));
1105 Tmp1 = SelectExpr(N.getOperand(1));
1106 Tmp2 = cast<RegSDNode>(N)->getReg();
1109 if (N.getOperand(1).getValueType() == MVT::f64 ||
1110 N.getOperand(1).getValueType() == MVT::f32)
1111 BuildMI(BB, PPC::FMR, 1, Tmp2).addReg(Tmp1);
1113 BuildMI(BB, PPC::OR, 2, Tmp2).addReg(Tmp1).addReg(Tmp1);
1116 case ISD::ImplicitDef:
1117 Select(N.getOperand(0));
1118 BuildMI(BB, PPC::IMPLICIT_DEF, 0, cast<RegSDNode>(N)->getReg());
1121 switch (N.getNumOperands()) {
1123 assert(0 && "Unknown return instruction!");
1125 assert(N.getOperand(1).getValueType() == MVT::i32 &&
1126 N.getOperand(2).getValueType() == MVT::i32 &&
1127 "Unknown two-register value!");
1128 Select(N.getOperand(0));
1129 Tmp1 = SelectExpr(N.getOperand(1));
1130 Tmp2 = SelectExpr(N.getOperand(2));
1131 BuildMI(BB, PPC::OR, 2, PPC::R3).addReg(Tmp1).addReg(Tmp1);
1132 BuildMI(BB, PPC::OR, 2, PPC::R4).addReg(Tmp2).addReg(Tmp2);
1135 Select(N.getOperand(0));
1136 Tmp1 = SelectExpr(N.getOperand(1));
1137 switch (N.getOperand(1).getValueType()) {
1139 assert(0 && "Unknown return type!");
1142 BuildMI(BB, PPC::FMR, 1, PPC::F1).addReg(Tmp1);
1145 BuildMI(BB, PPC::OR, 2, PPC::R3).addReg(Tmp1).addReg(Tmp1);
1149 Select(N.getOperand(0));
1152 BuildMI(BB, PPC::BLR, 0); // Just emit a 'ret' instruction
1154 case ISD::TRUNCSTORE:
1157 SDOperand Chain = N.getOperand(0);
1158 SDOperand Value = N.getOperand(1);
1159 SDOperand Address = N.getOperand(2);
1162 Tmp1 = SelectExpr(Value); //value
1164 if (opcode == ISD::STORE) {
1165 switch(Value.getValueType()) {
1166 default: assert(0 && "unknown Type in store");
1167 case MVT::i32: Opc = PPC::STW; break;
1168 case MVT::f64: Opc = PPC::STFD; break;
1169 case MVT::f32: Opc = PPC::STFS; break;
1171 } else { //ISD::TRUNCSTORE
1172 switch(cast<MVTSDNode>(Node)->getExtraValueType()) {
1173 default: assert(0 && "unknown Type in store");
1174 case MVT::i1: //FIXME: DAG does not promote this load
1175 case MVT::i8: Opc = PPC::STB; break;
1176 case MVT::i16: Opc = PPC::STH; break;
1180 if (Address.getOpcode() == ISD::GlobalAddress)
1182 BuildMI(BB, Opc, 2).addReg(Tmp1)
1183 .addGlobalAddress(cast<GlobalAddressSDNode>(Address)->getGlobal());
1185 else if(Address.getOpcode() == ISD::FrameIndex)
1187 BuildMI(BB, Opc, 2).addReg(Tmp1)
1188 .addFrameIndex(cast<FrameIndexSDNode>(Address)->getIndex());
1193 SelectAddr(Address, Tmp2, offset);
1194 BuildMI(BB, Opc, 3).addReg(Tmp1).addImm(offset).addReg(Tmp2);
1202 case ISD::CopyFromReg:
1204 case ISD::DYNAMIC_STACKALLOC:
1209 assert(0 && "Should not be reached!");
1213 /// createPPC32PatternInstructionSelector - This pass converts an LLVM function
1214 /// into a machine code representation using pattern matching and a machine
1215 /// description file.
1217 FunctionPass *llvm::createPPC32ISelPattern(TargetMachine &TM) {
1218 return new ISel(TM);