1 //===-- PPC32ISelDAGToDAG.cpp - PPC32 pattern matching inst selector ------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Chris Lattner 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,
11 // converting from a legalized dag to a PPC dag.
13 //===----------------------------------------------------------------------===//
16 #include "PPC32TargetMachine.h"
17 #include "PPC32ISelLowering.h"
18 #include "llvm/CodeGen/MachineInstrBuilder.h"
19 #include "llvm/CodeGen/MachineFunction.h"
20 #include "llvm/CodeGen/SSARegMap.h"
21 #include "llvm/CodeGen/SelectionDAG.h"
22 #include "llvm/CodeGen/SelectionDAGISel.h"
23 #include "llvm/Target/TargetOptions.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/Constants.h"
26 #include "llvm/GlobalValue.h"
27 #include "llvm/Support/Debug.h"
28 #include "llvm/Support/MathExtras.h"
32 Statistic<> FusedFP ("ppc-codegen", "Number of fused fp operations");
33 Statistic<> FrameOff("ppc-codegen", "Number of frame idx offsets collapsed");
35 //===--------------------------------------------------------------------===//
36 /// PPC32DAGToDAGISel - PPC32 specific code to select PPC32 machine
37 /// instructions for SelectionDAG operations.
39 class PPC32DAGToDAGISel : public SelectionDAGISel {
40 PPC32TargetLowering PPC32Lowering;
41 unsigned GlobalBaseReg;
43 PPC32DAGToDAGISel(TargetMachine &TM)
44 : SelectionDAGISel(PPC32Lowering), PPC32Lowering(TM) {}
46 virtual bool runOnFunction(Function &Fn) {
47 // Make sure we re-emit a set of the global base reg if necessary
49 return SelectionDAGISel::runOnFunction(Fn);
52 /// getI32Imm - Return a target constant with the specified value, of type
54 inline SDOperand getI32Imm(unsigned Imm) {
55 return CurDAG->getTargetConstant(Imm, MVT::i32);
58 /// getGlobalBaseReg - insert code into the entry mbb to materialize the PIC
59 /// base register. Return the virtual register that holds this value.
60 SDOperand getGlobalBaseReg();
62 // Select - Convert the specified operand from a target-independent to a
63 // target-specific node if it hasn't already been changed.
64 SDOperand Select(SDOperand Op);
66 SDNode *SelectIntImmediateExpr(SDOperand LHS, SDOperand RHS,
67 unsigned OCHi, unsigned OCLo,
68 bool IsArithmetic = false,
70 SDNode *SelectBitfieldInsert(SDNode *N);
72 /// SelectCC - Select a comparison of the specified values with the
73 /// specified condition code, returning the CR# of the expression.
74 SDOperand SelectCC(SDOperand LHS, SDOperand RHS, ISD::CondCode CC);
76 /// SelectAddr - Given the specified address, return the two operands for a
77 /// load/store instruction, and return true if it should be an indexed [r+r]
79 bool SelectAddr(SDOperand Addr, SDOperand &Op1, SDOperand &Op2);
81 SDOperand BuildSDIVSequence(SDNode *N);
82 SDOperand BuildUDIVSequence(SDNode *N);
84 /// InstructionSelectBasicBlock - This callback is invoked by
85 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
86 virtual void InstructionSelectBasicBlock(SelectionDAG &DAG);
88 virtual const char *getPassName() const {
89 return "PowerPC DAG->DAG Pattern Instruction Selection";
92 // Include the pieces autogenerated from the target description.
93 #include "PPC32GenDAGISel.inc"
96 SDOperand SelectDYNAMIC_STACKALLOC(SDOperand Op);
97 SDOperand SelectADD_PARTS(SDOperand Op);
98 SDOperand SelectSUB_PARTS(SDOperand Op);
99 SDOperand SelectSETCC(SDOperand Op);
103 /// InstructionSelectBasicBlock - This callback is invoked by
104 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
105 void PPC32DAGToDAGISel::InstructionSelectBasicBlock(SelectionDAG &DAG) {
108 // The selection process is inherently a bottom-up recursive process (users
109 // select their uses before themselves). Given infinite stack space, we
110 // could just start selecting on the root and traverse the whole graph. In
111 // practice however, this causes us to run out of stack space on large basic
112 // blocks. To avoid this problem, select the entry node, then all its uses,
113 // iteratively instead of recursively.
114 std::vector<SDOperand> Worklist;
115 Worklist.push_back(DAG.getEntryNode());
117 // Note that we can do this in the PPC target (scanning forward across token
118 // chain edges) because no nodes ever get folded across these edges. On a
119 // target like X86 which supports load/modify/store operations, this would
120 // have to be more careful.
121 while (!Worklist.empty()) {
122 SDOperand Node = Worklist.back();
125 if ((Node.Val->getOpcode() >= ISD::BUILTIN_OP_END &&
126 Node.Val->getOpcode() < PPCISD::FIRST_NUMBER) ||
127 CodeGenMap.count(Node)) continue;
129 for (SDNode::use_iterator UI = Node.Val->use_begin(),
130 E = Node.Val->use_end(); UI != E; ++UI) {
131 // Scan the values. If this use has a value that is a token chain, add it
134 for (unsigned i = 0, e = User->getNumValues(); i != e; ++i)
135 if (User->getValueType(i) == MVT::Other) {
136 Worklist.push_back(SDOperand(User, i));
141 // Finally, legalize this node.
145 // Select target instructions for the DAG.
146 DAG.setRoot(Select(DAG.getRoot()));
148 DAG.RemoveDeadNodes();
150 // Emit machine code to BB.
151 ScheduleAndEmitDAG(DAG);
154 /// getGlobalBaseReg - Output the instructions required to put the
155 /// base address to use for accessing globals into a register.
157 SDOperand PPC32DAGToDAGISel::getGlobalBaseReg() {
158 if (!GlobalBaseReg) {
159 // Insert the set of GlobalBaseReg into the first MBB of the function
160 MachineBasicBlock &FirstMBB = BB->getParent()->front();
161 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
162 SSARegMap *RegMap = BB->getParent()->getSSARegMap();
163 GlobalBaseReg = RegMap->createVirtualRegister(PPC32::GPRCRegisterClass);
164 BuildMI(FirstMBB, MBBI, PPC::MovePCtoLR, 0, PPC::LR);
165 BuildMI(FirstMBB, MBBI, PPC::MFLR, 1, GlobalBaseReg);
167 return CurDAG->getRegister(GlobalBaseReg, MVT::i32);
171 // isIntImmediate - This method tests to see if a constant operand.
172 // If so Imm will receive the 32 bit value.
173 static bool isIntImmediate(SDNode *N, unsigned& Imm) {
174 if (N->getOpcode() == ISD::Constant) {
175 Imm = cast<ConstantSDNode>(N)->getValue();
181 // isOprShiftImm - Returns true if the specified operand is a shift opcode with
182 // a immediate shift count less than 32.
183 static bool isOprShiftImm(SDNode *N, unsigned& Opc, unsigned& SH) {
184 Opc = N->getOpcode();
185 return (Opc == ISD::SHL || Opc == ISD::SRL || Opc == ISD::SRA) &&
186 isIntImmediate(N->getOperand(1).Val, SH) && SH < 32;
189 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
190 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
191 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
192 // not, since all 1s are not contiguous.
193 static bool isRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME) {
194 if (isShiftedMask_32(Val)) {
195 // look for the first non-zero bit
196 MB = CountLeadingZeros_32(Val);
197 // look for the first zero bit after the run of ones
198 ME = CountLeadingZeros_32((Val - 1) ^ Val);
201 Val = ~Val; // invert mask
202 if (isShiftedMask_32(Val)) {
203 // effectively look for the first zero bit
204 ME = CountLeadingZeros_32(Val) - 1;
205 // effectively look for the first one bit after the run of zeros
206 MB = CountLeadingZeros_32((Val - 1) ^ Val) + 1;
214 // isRotateAndMask - Returns true if Mask and Shift can be folded in to a rotate
215 // and mask opcode and mask operation.
216 static bool isRotateAndMask(SDNode *N, unsigned Mask, bool IsShiftMask,
217 unsigned &SH, unsigned &MB, unsigned &ME) {
219 unsigned Indeterminant = ~0; // bit mask marking indeterminant results
220 unsigned Opcode = N->getOpcode();
221 if (N->getNumOperands() != 2 ||
222 !isIntImmediate(N->getOperand(1).Val, Shift) || (Shift > 31))
225 if (Opcode == ISD::SHL) {
226 // apply shift left to mask if it comes first
227 if (IsShiftMask) Mask = Mask << Shift;
228 // determine which bits are made indeterminant by shift
229 Indeterminant = ~(0xFFFFFFFFu << Shift);
230 } else if (Opcode == ISD::SRA || Opcode == ISD::SRL) {
231 // apply shift right to mask if it comes first
232 if (IsShiftMask) Mask = Mask >> Shift;
233 // determine which bits are made indeterminant by shift
234 Indeterminant = ~(0xFFFFFFFFu >> Shift);
235 // adjust for the left rotate
241 // if the mask doesn't intersect any Indeterminant bits
242 if (Mask && !(Mask & Indeterminant)) {
244 // make sure the mask is still a mask (wrap arounds may not be)
245 return isRunOfOnes(Mask, MB, ME);
250 // isOpcWithIntImmediate - This method tests to see if the node is a specific
251 // opcode and that it has a immediate integer right operand.
252 // If so Imm will receive the 32 bit value.
253 static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) {
254 return N->getOpcode() == Opc && isIntImmediate(N->getOperand(1).Val, Imm);
257 // isOprNot - Returns true if the specified operand is an xor with immediate -1.
258 static bool isOprNot(SDNode *N) {
260 return isOpcWithIntImmediate(N, ISD::XOR, Imm) && (signed)Imm == -1;
263 // Immediate constant composers.
264 // Lo16 - grabs the lo 16 bits from a 32 bit constant.
265 // Hi16 - grabs the hi 16 bits from a 32 bit constant.
266 // HA16 - computes the hi bits required if the lo bits are add/subtracted in
268 static unsigned Lo16(unsigned x) { return x & 0x0000FFFF; }
269 static unsigned Hi16(unsigned x) { return Lo16(x >> 16); }
270 static unsigned HA16(unsigned x) { return Hi16((signed)x - (signed short)x); }
272 // isIntImmediate - This method tests to see if a constant operand.
273 // If so Imm will receive the 32 bit value.
274 static bool isIntImmediate(SDOperand N, unsigned& Imm) {
275 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
276 Imm = (unsigned)CN->getSignExtended();
282 /// SelectBitfieldInsert - turn an or of two masked values into
283 /// the rotate left word immediate then mask insert (rlwimi) instruction.
284 /// Returns true on success, false if the caller still needs to select OR.
286 /// Patterns matched:
287 /// 1. or shl, and 5. or and, and
288 /// 2. or and, shl 6. or shl, shr
289 /// 3. or shr, and 7. or shr, shl
291 SDNode *PPC32DAGToDAGISel::SelectBitfieldInsert(SDNode *N) {
292 bool IsRotate = false;
293 unsigned TgtMask = 0xFFFFFFFF, InsMask = 0xFFFFFFFF, SH = 0;
296 SDOperand Op0 = N->getOperand(0);
297 SDOperand Op1 = N->getOperand(1);
299 unsigned Op0Opc = Op0.getOpcode();
300 unsigned Op1Opc = Op1.getOpcode();
302 // Verify that we have the correct opcodes
303 if (ISD::SHL != Op0Opc && ISD::SRL != Op0Opc && ISD::AND != Op0Opc)
305 if (ISD::SHL != Op1Opc && ISD::SRL != Op1Opc && ISD::AND != Op1Opc)
308 // Generate Mask value for Target
309 if (isIntImmediate(Op0.getOperand(1), Value)) {
311 case ISD::SHL: TgtMask <<= Value; break;
312 case ISD::SRL: TgtMask >>= Value; break;
313 case ISD::AND: TgtMask &= Value; break;
319 // Generate Mask value for Insert
320 if (!isIntImmediate(Op1.getOperand(1), Value))
327 if (Op0Opc == ISD::SRL) IsRotate = true;
333 if (Op0Opc == ISD::SHL) IsRotate = true;
340 // If both of the inputs are ANDs and one of them has a logical shift by
341 // constant as its input, make that AND the inserted value so that we can
342 // combine the shift into the rotate part of the rlwimi instruction
343 bool IsAndWithShiftOp = false;
344 if (Op0Opc == ISD::AND && Op1Opc == ISD::AND) {
345 if (Op1.getOperand(0).getOpcode() == ISD::SHL ||
346 Op1.getOperand(0).getOpcode() == ISD::SRL) {
347 if (isIntImmediate(Op1.getOperand(0).getOperand(1), Value)) {
348 SH = Op1.getOperand(0).getOpcode() == ISD::SHL ? Value : 32 - Value;
349 IsAndWithShiftOp = true;
351 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL ||
352 Op0.getOperand(0).getOpcode() == ISD::SRL) {
353 if (isIntImmediate(Op0.getOperand(0).getOperand(1), Value)) {
355 std::swap(TgtMask, InsMask);
356 SH = Op1.getOperand(0).getOpcode() == ISD::SHL ? Value : 32 - Value;
357 IsAndWithShiftOp = true;
362 // Verify that the Target mask and Insert mask together form a full word mask
363 // and that the Insert mask is a run of set bits (which implies both are runs
364 // of set bits). Given that, Select the arguments and generate the rlwimi
367 if (((TgtMask & InsMask) == 0) && isRunOfOnes(InsMask, MB, ME)) {
368 bool fullMask = (TgtMask ^ InsMask) == 0xFFFFFFFF;
369 bool Op0IsAND = Op0Opc == ISD::AND;
370 // Check for rotlwi / rotrwi here, a special case of bitfield insert
371 // where both bitfield halves are sourced from the same value.
372 if (IsRotate && fullMask &&
373 N->getOperand(0).getOperand(0) == N->getOperand(1).getOperand(0)) {
374 Op0 = CurDAG->getTargetNode(PPC::RLWINM, MVT::i32,
375 Select(N->getOperand(0).getOperand(0)),
376 getI32Imm(SH), getI32Imm(0), getI32Imm(31));
379 SDOperand Tmp1 = (Op0IsAND && fullMask) ? Select(Op0.getOperand(0))
381 SDOperand Tmp2 = IsAndWithShiftOp ? Select(Op1.getOperand(0).getOperand(0))
382 : Select(Op1.getOperand(0));
383 Op0 = CurDAG->getTargetNode(PPC::RLWIMI, MVT::i32, Tmp1, Tmp2,
384 getI32Imm(SH), getI32Imm(MB), getI32Imm(ME));
390 // SelectIntImmediateExpr - Choose code for integer operations with an immediate
392 SDNode *PPC32DAGToDAGISel::SelectIntImmediateExpr(SDOperand LHS, SDOperand RHS,
393 unsigned OCHi, unsigned OCLo,
396 // Check to make sure this is a constant.
397 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(RHS);
398 // Exit if not a constant.
400 // Extract immediate.
401 unsigned C = (unsigned)CN->getValue();
402 // Negate if required (ISD::SUB).
404 // Get the hi and lo portions of constant.
405 unsigned Hi = IsArithmetic ? HA16(C) : Hi16(C);
406 unsigned Lo = Lo16(C);
408 // If two instructions are needed and usage indicates it would be better to
409 // load immediate into a register, bail out.
410 if (Hi && Lo && CN->use_size() > 2) return false;
412 // Select the first operand.
413 SDOperand Opr0 = Select(LHS);
415 if (Lo) // Add in the lo-part.
416 Opr0 = CurDAG->getTargetNode(OCLo, MVT::i32, Opr0, getI32Imm(Lo));
417 if (Hi) // Add in the hi-part.
418 Opr0 = CurDAG->getTargetNode(OCHi, MVT::i32, Opr0, getI32Imm(Hi));
422 /// SelectAddr - Given the specified address, return the two operands for a
423 /// load/store instruction, and return true if it should be an indexed [r+r]
425 bool PPC32DAGToDAGISel::SelectAddr(SDOperand Addr, SDOperand &Op1,
428 if (Addr.getOpcode() == ISD::ADD) {
429 if (isIntImmediate(Addr.getOperand(1), imm) && isInt16(imm)) {
430 Op1 = getI32Imm(Lo16(imm));
431 if (FrameIndexSDNode *FI =
432 dyn_cast<FrameIndexSDNode>(Addr.getOperand(0))) {
434 Op2 = CurDAG->getTargetFrameIndex(FI->getIndex(), MVT::i32);
436 Op2 = Select(Addr.getOperand(0));
440 Op1 = Select(Addr.getOperand(0));
441 Op2 = Select(Addr.getOperand(1));
442 return true; // [r+r]
446 // Now check if we're dealing with a global, and whether or not we should emit
447 // an optimized load or store for statics.
448 if (GlobalAddressSDNode *GN = dyn_cast<GlobalAddressSDNode>(Addr)) {
449 GlobalValue *GV = GN->getGlobal();
450 if (!GV->hasWeakLinkage() && !GV->isExternal()) {
451 Op1 = CurDAG->getTargetGlobalAddress(GV, MVT::i32);
453 Op2 = CurDAG->getTargetNode(PPC::ADDIS, MVT::i32, getGlobalBaseReg(),
456 Op2 = CurDAG->getTargetNode(PPC::LIS, MVT::i32, Op1);
459 } else if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Addr)) {
461 Op2 = CurDAG->getTargetFrameIndex(FI->getIndex(), MVT::i32);
463 } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Addr)) {
466 Op2 = CurDAG->getTargetNode(PPC::ADDIS, MVT::i32, getGlobalBaseReg(),Op1);
468 Op2 = CurDAG->getTargetNode(PPC::LIS, MVT::i32, Op1);
476 /// SelectCC - Select a comparison of the specified values with the specified
477 /// condition code, returning the CR# of the expression.
478 SDOperand PPC32DAGToDAGISel::SelectCC(SDOperand LHS, SDOperand RHS,
480 // Always select the LHS.
483 // Use U to determine whether the SETCC immediate range is signed or not.
484 if (MVT::isInteger(LHS.getValueType())) {
485 bool U = ISD::isUnsignedIntSetCC(CC);
487 if (isIntImmediate(RHS, Imm) &&
488 ((U && isUInt16(Imm)) || (!U && isInt16(Imm))))
489 return CurDAG->getTargetNode(U ? PPC::CMPLWI : PPC::CMPWI, MVT::i32,
490 LHS, getI32Imm(Lo16(Imm)));
491 return CurDAG->getTargetNode(U ? PPC::CMPLW : PPC::CMPW, MVT::i32,
493 } else if (LHS.getValueType() == MVT::f32) {
494 return CurDAG->getTargetNode(PPC::FCMPUS, MVT::i32, LHS, Select(RHS));
496 return CurDAG->getTargetNode(PPC::FCMPUD, MVT::i32, LHS, Select(RHS));
500 /// getBCCForSetCC - Returns the PowerPC condition branch mnemonic corresponding
502 static unsigned getBCCForSetCC(ISD::CondCode CC) {
504 default: assert(0 && "Unknown condition!"); abort();
505 case ISD::SETEQ: return PPC::BEQ;
506 case ISD::SETNE: return PPC::BNE;
508 case ISD::SETLT: return PPC::BLT;
510 case ISD::SETLE: return PPC::BLE;
512 case ISD::SETGT: return PPC::BGT;
514 case ISD::SETGE: return PPC::BGE;
519 /// getCRIdxForSetCC - Return the index of the condition register field
520 /// associated with the SetCC condition, and whether or not the field is
521 /// treated as inverted. That is, lt = 0; ge = 0 inverted.
522 static unsigned getCRIdxForSetCC(ISD::CondCode CC, bool& Inv) {
524 default: assert(0 && "Unknown condition!"); abort();
526 case ISD::SETLT: Inv = false; return 0;
528 case ISD::SETGE: Inv = true; return 0;
530 case ISD::SETGT: Inv = false; return 1;
532 case ISD::SETLE: Inv = true; return 1;
533 case ISD::SETEQ: Inv = false; return 2;
534 case ISD::SETNE: Inv = true; return 2;
539 // Structure used to return the necessary information to codegen an SDIV as
542 int m; // magic number
543 int s; // shift amount
547 unsigned int m; // magic number
548 int a; // add indicator
549 int s; // shift amount
552 /// magic - calculate the magic numbers required to codegen an integer sdiv as
553 /// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
555 static struct ms magic(int d) {
557 unsigned int ad, anc, delta, q1, r1, q2, r2, t;
558 const unsigned int two31 = 0x80000000U;
562 t = two31 + ((unsigned int)d >> 31);
563 anc = t - 1 - t%ad; // absolute value of nc
564 p = 31; // initialize p
565 q1 = two31/anc; // initialize q1 = 2p/abs(nc)
566 r1 = two31 - q1*anc; // initialize r1 = rem(2p,abs(nc))
567 q2 = two31/ad; // initialize q2 = 2p/abs(d)
568 r2 = two31 - q2*ad; // initialize r2 = rem(2p,abs(d))
571 q1 = 2*q1; // update q1 = 2p/abs(nc)
572 r1 = 2*r1; // update r1 = rem(2p/abs(nc))
573 if (r1 >= anc) { // must be unsigned comparison
577 q2 = 2*q2; // update q2 = 2p/abs(d)
578 r2 = 2*r2; // update r2 = rem(2p/abs(d))
579 if (r2 >= ad) { // must be unsigned comparison
584 } while (q1 < delta || (q1 == delta && r1 == 0));
587 if (d < 0) mag.m = -mag.m; // resulting magic number
588 mag.s = p - 32; // resulting shift
592 /// magicu - calculate the magic numbers required to codegen an integer udiv as
593 /// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
594 static struct mu magicu(unsigned d)
597 unsigned int nc, delta, q1, r1, q2, r2;
599 magu.a = 0; // initialize "add" indicator
601 p = 31; // initialize p
602 q1 = 0x80000000/nc; // initialize q1 = 2p/nc
603 r1 = 0x80000000 - q1*nc; // initialize r1 = rem(2p,nc)
604 q2 = 0x7FFFFFFF/d; // initialize q2 = (2p-1)/d
605 r2 = 0x7FFFFFFF - q2*d; // initialize r2 = rem((2p-1),d)
608 if (r1 >= nc - r1 ) {
609 q1 = 2*q1 + 1; // update q1
610 r1 = 2*r1 - nc; // update r1
613 q1 = 2*q1; // update q1
614 r1 = 2*r1; // update r1
616 if (r2 + 1 >= d - r2) {
617 if (q2 >= 0x7FFFFFFF) magu.a = 1;
618 q2 = 2*q2 + 1; // update q2
619 r2 = 2*r2 + 1 - d; // update r2
622 if (q2 >= 0x80000000) magu.a = 1;
623 q2 = 2*q2; // update q2
624 r2 = 2*r2 + 1; // update r2
627 } while (p < 64 && (q1 < delta || (q1 == delta && r1 == 0)));
628 magu.m = q2 + 1; // resulting magic number
629 magu.s = p - 32; // resulting shift
633 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
634 /// return a DAG expression to select that will generate the same value by
635 /// multiplying by a magic number. See:
636 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
637 SDOperand PPC32DAGToDAGISel::BuildSDIVSequence(SDNode *N) {
638 int d = (int)cast<ConstantSDNode>(N->getOperand(1))->getValue();
639 ms magics = magic(d);
640 // Multiply the numerator (operand 0) by the magic value
641 SDOperand Q = CurDAG->getNode(ISD::MULHS, MVT::i32, N->getOperand(0),
642 CurDAG->getConstant(magics.m, MVT::i32));
643 // If d > 0 and m < 0, add the numerator
644 if (d > 0 && magics.m < 0)
645 Q = CurDAG->getNode(ISD::ADD, MVT::i32, Q, N->getOperand(0));
646 // If d < 0 and m > 0, subtract the numerator.
647 if (d < 0 && magics.m > 0)
648 Q = CurDAG->getNode(ISD::SUB, MVT::i32, Q, N->getOperand(0));
649 // Shift right algebraic if shift value is nonzero
651 Q = CurDAG->getNode(ISD::SRA, MVT::i32, Q,
652 CurDAG->getConstant(magics.s, MVT::i32));
653 // Extract the sign bit and add it to the quotient
655 CurDAG->getNode(ISD::SRL, MVT::i32, Q, CurDAG->getConstant(31, MVT::i32));
656 return CurDAG->getNode(ISD::ADD, MVT::i32, Q, T);
659 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
660 /// return a DAG expression to select that will generate the same value by
661 /// multiplying by a magic number. See:
662 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
663 SDOperand PPC32DAGToDAGISel::BuildUDIVSequence(SDNode *N) {
664 unsigned d = (unsigned)cast<ConstantSDNode>(N->getOperand(1))->getValue();
665 mu magics = magicu(d);
666 // Multiply the numerator (operand 0) by the magic value
667 SDOperand Q = CurDAG->getNode(ISD::MULHU, MVT::i32, N->getOperand(0),
668 CurDAG->getConstant(magics.m, MVT::i32));
670 return CurDAG->getNode(ISD::SRL, MVT::i32, Q,
671 CurDAG->getConstant(magics.s, MVT::i32));
673 SDOperand NPQ = CurDAG->getNode(ISD::SUB, MVT::i32, N->getOperand(0), Q);
674 NPQ = CurDAG->getNode(ISD::SRL, MVT::i32, NPQ,
675 CurDAG->getConstant(1, MVT::i32));
676 NPQ = CurDAG->getNode(ISD::ADD, MVT::i32, NPQ, Q);
677 return CurDAG->getNode(ISD::SRL, MVT::i32, NPQ,
678 CurDAG->getConstant(magics.s-1, MVT::i32));
682 SDOperand PPC32DAGToDAGISel::SelectDYNAMIC_STACKALLOC(SDOperand Op) {
685 // FIXME: We are currently ignoring the requested alignment for handling
686 // greater than the stack alignment. This will need to be revisited at some
687 // point. Align = N.getOperand(2);
688 if (!isa<ConstantSDNode>(N->getOperand(2)) ||
689 cast<ConstantSDNode>(N->getOperand(2))->getValue() != 0) {
690 std::cerr << "Cannot allocate stack object with greater alignment than"
691 << " the stack alignment yet!";
694 SDOperand Chain = Select(N->getOperand(0));
695 SDOperand Amt = Select(N->getOperand(1));
697 SDOperand R1Reg = CurDAG->getRegister(PPC::R1, MVT::i32);
699 SDOperand R1Val = CurDAG->getCopyFromReg(Chain, PPC::R1, MVT::i32);
700 Chain = R1Val.getValue(1);
702 // Subtract the amount (guaranteed to be a multiple of the stack alignment)
703 // from the stack pointer, giving us the result pointer.
704 SDOperand Result = CurDAG->getTargetNode(PPC::SUBF, MVT::i32, Amt, R1Val);
706 // Copy this result back into R1.
707 Chain = CurDAG->getNode(ISD::CopyToReg, MVT::Other, Chain, R1Reg, Result);
709 // Copy this result back out of R1 to make sure we're not using the stack
710 // space without decrementing the stack pointer.
711 Result = CurDAG->getCopyFromReg(Chain, PPC::R1, MVT::i32);
713 // Finally, replace the DYNAMIC_STACKALLOC with the copyfromreg.
714 CodeGenMap[Op.getValue(0)] = Result;
715 CodeGenMap[Op.getValue(1)] = Result.getValue(1);
716 return SDOperand(Result.Val, Op.ResNo);
719 SDOperand PPC32DAGToDAGISel::SelectADD_PARTS(SDOperand Op) {
721 SDOperand LHSL = Select(N->getOperand(0));
722 SDOperand LHSH = Select(N->getOperand(1));
725 bool ME = false, ZE = false;
726 if (isIntImmediate(N->getOperand(3), Imm)) {
727 ME = (signed)Imm == -1;
731 std::vector<SDOperand> Result;
732 SDOperand CarryFromLo;
733 if (isIntImmediate(N->getOperand(2), Imm) &&
734 ((signed)Imm >= -32768 || (signed)Imm < 32768)) {
735 // Codegen the low 32 bits of the add. Interestingly, there is no
736 // shifted form of add immediate carrying.
737 CarryFromLo = CurDAG->getTargetNode(PPC::ADDIC, MVT::i32, MVT::Flag,
738 LHSL, getI32Imm(Imm));
740 CarryFromLo = CurDAG->getTargetNode(PPC::ADDC, MVT::i32, MVT::Flag,
741 LHSL, Select(N->getOperand(2)));
743 CarryFromLo = CarryFromLo.getValue(1);
745 // Codegen the high 32 bits, adding zero, minus one, or the full value
746 // along with the carry flag produced by addc/addic.
749 ResultHi = CurDAG->getTargetNode(PPC::ADDZE, MVT::i32, LHSH, CarryFromLo);
751 ResultHi = CurDAG->getTargetNode(PPC::ADDME, MVT::i32, LHSH, CarryFromLo);
753 ResultHi = CurDAG->getTargetNode(PPC::ADDE, MVT::i32, LHSH,
754 Select(N->getOperand(3)), CarryFromLo);
755 Result.push_back(CarryFromLo.getValue(0));
756 Result.push_back(ResultHi);
758 CodeGenMap[Op.getValue(0)] = Result[0];
759 CodeGenMap[Op.getValue(1)] = Result[1];
760 return Result[Op.ResNo];
762 SDOperand PPC32DAGToDAGISel::SelectSUB_PARTS(SDOperand Op) {
764 SDOperand LHSL = Select(N->getOperand(0));
765 SDOperand LHSH = Select(N->getOperand(1));
766 SDOperand RHSL = Select(N->getOperand(2));
767 SDOperand RHSH = Select(N->getOperand(3));
769 std::vector<SDOperand> Result;
770 Result.push_back(CurDAG->getTargetNode(PPC::SUBFC, MVT::i32, MVT::Flag,
772 Result.push_back(CurDAG->getTargetNode(PPC::SUBFE, MVT::i32, RHSH, LHSH,
773 Result[0].getValue(1)));
774 CodeGenMap[Op.getValue(0)] = Result[0];
775 CodeGenMap[Op.getValue(1)] = Result[1];
776 return Result[Op.ResNo];
779 SDOperand PPC32DAGToDAGISel::SelectSETCC(SDOperand Op) {
782 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
783 if (isIntImmediate(N->getOperand(1), Imm)) {
784 // We can codegen setcc op, imm very efficiently compared to a brcond.
785 // Check for those cases here.
788 SDOperand Op = Select(N->getOperand(0));
790 default: assert(0 && "Unhandled SetCC condition"); abort();
792 Op = CurDAG->getTargetNode(PPC::CNTLZW, MVT::i32, Op);
793 CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Op, getI32Imm(27),
794 getI32Imm(5), getI32Imm(31));
797 SDOperand AD = CurDAG->getTargetNode(PPC::ADDIC, MVT::i32, MVT::Flag,
799 CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, AD, Op, AD.getValue(1));
803 CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Op, getI32Imm(1),
804 getI32Imm(31), getI32Imm(31));
807 SDOperand T = CurDAG->getTargetNode(PPC::NEG, MVT::i32, Op);
808 T = CurDAG->getTargetNode(PPC::ANDC, MVT::i32, T, Op);;
809 CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, T, getI32Imm(1),
810 getI32Imm(31), getI32Imm(31));
814 return SDOperand(N, 0);
815 } else if (Imm == ~0U) { // setcc op, -1
816 SDOperand Op = Select(N->getOperand(0));
818 default: assert(0 && "Unhandled SetCC condition"); abort();
820 Op = CurDAG->getTargetNode(PPC::ADDIC, MVT::i32, MVT::Flag,
822 CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32,
823 CurDAG->getTargetNode(PPC::LI, MVT::i32,
828 Op = CurDAG->getTargetNode(PPC::NOR, MVT::i32, Op, Op);
829 SDOperand AD = CurDAG->getTargetNode(PPC::ADDIC, MVT::i32, MVT::Flag,
831 CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, AD, Op, AD.getValue(1));
835 SDOperand AD = CurDAG->getTargetNode(PPC::ADDI, MVT::i32, Op,
837 SDOperand AN = CurDAG->getTargetNode(PPC::AND, MVT::i32, AD, Op);
838 CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, AN, getI32Imm(1),
839 getI32Imm(31), getI32Imm(31));
843 Op = CurDAG->getTargetNode(PPC::RLWINM, MVT::i32, Op, getI32Imm(1),
844 getI32Imm(31), getI32Imm(31));
845 CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Op, getI32Imm(1));
848 return SDOperand(N, 0);
853 unsigned Idx = getCRIdxForSetCC(CC, Inv);
854 SDOperand CCReg = SelectCC(N->getOperand(0), N->getOperand(1), CC);
857 // Force the ccreg into CR7.
858 SDOperand CR7Reg = CurDAG->getRegister(PPC::CR7, MVT::i32);
860 std::vector<MVT::ValueType> VTs;
861 VTs.push_back(MVT::Other);
862 VTs.push_back(MVT::Flag); // NONSTANDARD CopyToReg node: defines a flag
863 std::vector<SDOperand> Ops;
864 Ops.push_back(CurDAG->getEntryNode());
865 Ops.push_back(CR7Reg);
866 Ops.push_back(CCReg);
867 CCReg = CurDAG->getNode(ISD::CopyToReg, VTs, Ops).getValue(1);
869 if (TLI.getTargetMachine().getSubtarget<PPCSubtarget>().isGigaProcessor())
870 IntCR = CurDAG->getTargetNode(PPC::MFOCRF, MVT::i32, CR7Reg, CCReg);
872 IntCR = CurDAG->getTargetNode(PPC::MFCR, MVT::i32, CCReg);
875 CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, IntCR,
876 getI32Imm(32-(3-Idx)), getI32Imm(31), getI32Imm(31));
879 CurDAG->getTargetNode(PPC::RLWINM, MVT::i32, IntCR,
880 getI32Imm(32-(3-Idx)), getI32Imm(31),getI32Imm(31));
881 CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Tmp, getI32Imm(1));
884 return SDOperand(N, 0);
887 // Select - Convert the specified operand from a target-independent to a
888 // target-specific node if it hasn't already been changed.
889 SDOperand PPC32DAGToDAGISel::Select(SDOperand Op) {
891 if (N->getOpcode() >= ISD::BUILTIN_OP_END &&
892 N->getOpcode() < PPCISD::FIRST_NUMBER)
893 return Op; // Already selected.
895 // If this has already been converted, use it.
896 std::map<SDOperand, SDOperand>::iterator CGMI = CodeGenMap.find(Op);
897 if (CGMI != CodeGenMap.end()) return CGMI->second;
899 switch (N->getOpcode()) {
901 case ISD::DYNAMIC_STACKALLOC: return SelectDYNAMIC_STACKALLOC(Op);
902 case ISD::ADD_PARTS: return SelectADD_PARTS(Op);
903 case ISD::SUB_PARTS: return SelectSUB_PARTS(Op);
904 case ISD::SETCC: return SelectSETCC(Op);
905 case ISD::TokenFactor: {
907 if (N->getNumOperands() == 2) {
908 SDOperand Op0 = Select(N->getOperand(0));
909 SDOperand Op1 = Select(N->getOperand(1));
910 New = CurDAG->getNode(ISD::TokenFactor, MVT::Other, Op0, Op1);
912 std::vector<SDOperand> Ops;
913 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
914 Ops.push_back(Select(N->getOperand(i)));
915 New = CurDAG->getNode(ISD::TokenFactor, MVT::Other, Ops);
918 if (!N->hasOneUse()) CodeGenMap[Op] = New;
921 case ISD::CopyFromReg: {
922 SDOperand Chain = Select(N->getOperand(0));
923 if (Chain == N->getOperand(0)) return Op; // No change
924 SDOperand New = CurDAG->getCopyFromReg(Chain,
925 cast<RegisterSDNode>(N->getOperand(1))->getReg(), N->getValueType(0));
926 return New.getValue(Op.ResNo);
928 case ISD::CopyToReg: {
929 SDOperand Chain = Select(N->getOperand(0));
930 SDOperand Reg = N->getOperand(1);
931 SDOperand Val = Select(N->getOperand(2));
932 SDOperand New = CurDAG->getNode(ISD::CopyToReg, MVT::Other,
934 if (!N->hasOneUse()) CodeGenMap[Op] = New;
938 if (N->getValueType(0) == MVT::i32)
939 CurDAG->SelectNodeTo(N, PPC::IMPLICIT_DEF_GPR, MVT::i32);
940 else if (N->getValueType(0) == MVT::f32)
941 CurDAG->SelectNodeTo(N, PPC::IMPLICIT_DEF_F4, MVT::f32);
943 CurDAG->SelectNodeTo(N, PPC::IMPLICIT_DEF_F8, MVT::f64);
944 return SDOperand(N, 0);
945 case ISD::FrameIndex: {
946 int FI = cast<FrameIndexSDNode>(N)->getIndex();
947 CurDAG->SelectNodeTo(N, PPC::ADDI, MVT::i32,
948 CurDAG->getTargetFrameIndex(FI, MVT::i32),
950 return SDOperand(N, 0);
952 case ISD::ConstantPool: {
953 Constant *C = cast<ConstantPoolSDNode>(N)->get();
954 SDOperand Tmp, CPI = CurDAG->getTargetConstantPool(C, MVT::i32);
956 Tmp = CurDAG->getTargetNode(PPC::ADDIS, MVT::i32, getGlobalBaseReg(),CPI);
958 Tmp = CurDAG->getTargetNode(PPC::LIS, MVT::i32, CPI);
959 CurDAG->SelectNodeTo(N, PPC::LA, MVT::i32, Tmp, CPI);
960 return SDOperand(N, 0);
962 case ISD::GlobalAddress: {
963 GlobalValue *GV = cast<GlobalAddressSDNode>(N)->getGlobal();
965 SDOperand GA = CurDAG->getTargetGlobalAddress(GV, MVT::i32);
967 Tmp = CurDAG->getTargetNode(PPC::ADDIS, MVT::i32, getGlobalBaseReg(), GA);
969 Tmp = CurDAG->getTargetNode(PPC::LIS, MVT::i32, GA);
971 if (GV->hasWeakLinkage() || GV->isExternal())
972 CurDAG->SelectNodeTo(N, PPC::LWZ, MVT::i32, GA, Tmp);
974 CurDAG->SelectNodeTo(N, PPC::LA, MVT::i32, Tmp, GA);
975 return SDOperand(N, 0);
979 SDOperand Comparison = Select(N->getOperand(0));
980 // Extend the comparison to 64-bits.
981 if (Comparison.getValueType() == MVT::f32)
982 Comparison = CurDAG->getTargetNode(PPC::FMRSD, MVT::f64, Comparison);
984 unsigned Opc = N->getValueType(0) == MVT::f32 ? PPC::FSELS : PPC::FSELD;
985 CurDAG->SelectNodeTo(N, Opc, N->getValueType(0), Comparison,
986 Select(N->getOperand(1)), Select(N->getOperand(2)));
987 return SDOperand(N, 0);
990 CurDAG->SelectNodeTo(N, PPC::FCFID, N->getValueType(0),
991 Select(N->getOperand(0)));
992 return SDOperand(N, 0);
994 CurDAG->SelectNodeTo(N, PPC::FCTIDZ, N->getValueType(0),
995 Select(N->getOperand(0)));
996 return SDOperand(N, 0);
998 CurDAG->SelectNodeTo(N, PPC::FCTIWZ, N->getValueType(0),
999 Select(N->getOperand(0)));
1000 return SDOperand(N, 0);
1002 MVT::ValueType Ty = N->getValueType(0);
1003 if (!NoExcessFPPrecision) { // Match FMA ops
1004 if (N->getOperand(0).getOpcode() == ISD::FMUL &&
1005 N->getOperand(0).Val->hasOneUse()) {
1006 ++FusedFP; // Statistic
1007 CurDAG->SelectNodeTo(N, Ty == MVT::f64 ? PPC::FMADD : PPC::FMADDS, Ty,
1008 Select(N->getOperand(0).getOperand(0)),
1009 Select(N->getOperand(0).getOperand(1)),
1010 Select(N->getOperand(1)));
1011 return SDOperand(N, 0);
1012 } else if (N->getOperand(1).getOpcode() == ISD::FMUL &&
1013 N->getOperand(1).hasOneUse()) {
1014 ++FusedFP; // Statistic
1015 CurDAG->SelectNodeTo(N, Ty == MVT::f64 ? PPC::FMADD : PPC::FMADDS, Ty,
1016 Select(N->getOperand(1).getOperand(0)),
1017 Select(N->getOperand(1).getOperand(1)),
1018 Select(N->getOperand(0)));
1019 return SDOperand(N, 0);
1023 CurDAG->SelectNodeTo(N, Ty == MVT::f64 ? PPC::FADD : PPC::FADDS, Ty,
1024 Select(N->getOperand(0)), Select(N->getOperand(1)));
1025 return SDOperand(N, 0);
1028 MVT::ValueType Ty = N->getValueType(0);
1030 if (!NoExcessFPPrecision) { // Match FMA ops
1031 if (N->getOperand(0).getOpcode() == ISD::FMUL &&
1032 N->getOperand(0).Val->hasOneUse()) {
1033 ++FusedFP; // Statistic
1034 CurDAG->SelectNodeTo(N, Ty == MVT::f64 ? PPC::FMSUB : PPC::FMSUBS, Ty,
1035 Select(N->getOperand(0).getOperand(0)),
1036 Select(N->getOperand(0).getOperand(1)),
1037 Select(N->getOperand(1)));
1038 return SDOperand(N, 0);
1039 } else if (N->getOperand(1).getOpcode() == ISD::FMUL &&
1040 N->getOperand(1).Val->hasOneUse()) {
1041 ++FusedFP; // Statistic
1042 CurDAG->SelectNodeTo(N, Ty == MVT::f64 ? PPC::FNMSUB : PPC::FNMSUBS, Ty,
1043 Select(N->getOperand(1).getOperand(0)),
1044 Select(N->getOperand(1).getOperand(1)),
1045 Select(N->getOperand(0)));
1046 return SDOperand(N, 0);
1049 CurDAG->SelectNodeTo(N, Ty == MVT::f64 ? PPC::FSUB : PPC::FSUBS, Ty,
1050 Select(N->getOperand(0)),
1051 Select(N->getOperand(1)));
1052 return SDOperand(N, 0);
1056 if (isIntImmediate(N->getOperand(1), Imm)) {
1057 if ((signed)Imm > 0 && isPowerOf2_32(Imm)) {
1059 CurDAG->getTargetNode(PPC::SRAWI, MVT::i32, MVT::Flag,
1060 Select(N->getOperand(0)),
1061 getI32Imm(Log2_32(Imm)));
1062 CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32,
1063 Op.getValue(0), Op.getValue(1));
1064 return SDOperand(N, 0);
1065 } else if ((signed)Imm < 0 && isPowerOf2_32(-Imm)) {
1067 CurDAG->getTargetNode(PPC::SRAWI, MVT::i32, MVT::Flag,
1068 Select(N->getOperand(0)),
1069 getI32Imm(Log2_32(-Imm)));
1071 CurDAG->getTargetNode(PPC::ADDZE, MVT::i32, Op.getValue(0),
1073 CurDAG->SelectNodeTo(N, PPC::NEG, MVT::i32, PT);
1074 return SDOperand(N, 0);
1076 SDOperand Result = Select(BuildSDIVSequence(N));
1077 CodeGenMap[Op] = Result;
1082 // Other cases are autogenerated.
1086 // If this is a divide by constant, we can emit code using some magic
1087 // constants to implement it as a multiply instead.
1089 if (isIntImmediate(N->getOperand(1), Imm) && Imm) {
1090 SDOperand Result = Select(BuildUDIVSequence(N));
1091 CodeGenMap[Op] = Result;
1095 // Other cases are autogenerated.
1100 // If this is an and of a value rotated between 0 and 31 bits and then and'd
1101 // with a mask, emit rlwinm
1102 if (isIntImmediate(N->getOperand(1), Imm) && (isShiftedMask_32(Imm) ||
1103 isShiftedMask_32(~Imm))) {
1105 unsigned SH, MB, ME;
1106 if (isRotateAndMask(N->getOperand(0).Val, Imm, false, SH, MB, ME)) {
1107 Val = Select(N->getOperand(0).getOperand(0));
1109 Val = Select(N->getOperand(0));
1110 isRunOfOnes(Imm, MB, ME);
1113 CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Val, getI32Imm(SH),
1114 getI32Imm(MB), getI32Imm(ME));
1115 return SDOperand(N, 0);
1118 // Other cases are autogenerated.
1122 if (SDNode *I = SelectBitfieldInsert(N))
1123 return CodeGenMap[Op] = SDOperand(I, 0);
1125 if (SDNode *I = SelectIntImmediateExpr(N->getOperand(0),
1127 PPC::ORIS, PPC::ORI))
1128 return CodeGenMap[Op] = SDOperand(I, 0);
1130 // Other cases are autogenerated.
1133 unsigned Imm, SH, MB, ME;
1134 if (isOpcWithIntImmediate(N->getOperand(0).Val, ISD::AND, Imm) &&
1135 isRotateAndMask(N, Imm, true, SH, MB, ME))
1136 CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32,
1137 Select(N->getOperand(0).getOperand(0)),
1138 getI32Imm(SH), getI32Imm(MB), getI32Imm(ME));
1139 else if (isIntImmediate(N->getOperand(1), Imm))
1140 CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Select(N->getOperand(0)),
1141 getI32Imm(Imm), getI32Imm(0), getI32Imm(31-Imm));
1143 CurDAG->SelectNodeTo(N, PPC::SLW, MVT::i32, Select(N->getOperand(0)),
1144 Select(N->getOperand(1)));
1145 return SDOperand(N, 0);
1148 unsigned Imm, SH, MB, ME;
1149 if (isOpcWithIntImmediate(N->getOperand(0).Val, ISD::AND, Imm) &&
1150 isRotateAndMask(N, Imm, true, SH, MB, ME))
1151 CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32,
1152 Select(N->getOperand(0).getOperand(0)),
1153 getI32Imm(SH & 0x1F), getI32Imm(MB), getI32Imm(ME));
1154 else if (isIntImmediate(N->getOperand(1), Imm))
1155 CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Select(N->getOperand(0)),
1156 getI32Imm((32-Imm) & 0x1F), getI32Imm(Imm),
1159 CurDAG->SelectNodeTo(N, PPC::SRW, MVT::i32, Select(N->getOperand(0)),
1160 Select(N->getOperand(1)));
1161 return SDOperand(N, 0);
1164 unsigned Imm, SH, MB, ME;
1165 if (isOpcWithIntImmediate(N->getOperand(0).Val, ISD::AND, Imm) &&
1166 isRotateAndMask(N, Imm, true, SH, MB, ME))
1167 CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32,
1168 Select(N->getOperand(0).getOperand(0)),
1169 getI32Imm(SH), getI32Imm(MB), getI32Imm(ME));
1170 else if (isIntImmediate(N->getOperand(1), Imm))
1171 CurDAG->SelectNodeTo(N, PPC::SRAWI, MVT::i32, Select(N->getOperand(0)),
1174 CurDAG->SelectNodeTo(N, PPC::SRAW, MVT::i32, Select(N->getOperand(0)),
1175 Select(N->getOperand(1)));
1176 return SDOperand(N, 0);
1179 unsigned Opc = N->getValueType(0) == MVT::f32 ? PPC::FMULS : PPC::FMUL;
1180 CurDAG->SelectNodeTo(N, Opc, N->getValueType(0), Select(N->getOperand(0)),
1181 Select(N->getOperand(1)));
1182 return SDOperand(N, 0);
1185 unsigned Opc = N->getValueType(0) == MVT::f32 ? PPC::FDIVS : PPC::FDIV;
1186 CurDAG->SelectNodeTo(N, Opc, N->getValueType(0), Select(N->getOperand(0)),
1187 Select(N->getOperand(1)));
1188 return SDOperand(N, 0);
1191 if (N->getValueType(0) == MVT::f32)
1192 CurDAG->SelectNodeTo(N, PPC::FABSS, MVT::f32, Select(N->getOperand(0)));
1194 CurDAG->SelectNodeTo(N, PPC::FABSD, MVT::f64, Select(N->getOperand(0)));
1195 return SDOperand(N, 0);
1196 case ISD::FP_EXTEND:
1197 assert(MVT::f64 == N->getValueType(0) &&
1198 MVT::f32 == N->getOperand(0).getValueType() && "Illegal FP_EXTEND");
1199 // We need to emit an FMR to make sure that the result has the right value
1201 CurDAG->SelectNodeTo(N, PPC::FMRSD, MVT::f64, Select(N->getOperand(0)));
1202 return SDOperand(N, 0);
1204 assert(MVT::f32 == N->getValueType(0) &&
1205 MVT::f64 == N->getOperand(0).getValueType() && "Illegal FP_ROUND");
1206 CurDAG->SelectNodeTo(N, PPC::FRSP, MVT::f32, Select(N->getOperand(0)));
1207 return SDOperand(N, 0);
1209 SDOperand Val = Select(N->getOperand(0));
1210 MVT::ValueType Ty = N->getValueType(0);
1211 if (Val.Val->hasOneUse()) {
1213 switch (Val.isTargetOpcode() ? Val.getTargetOpcode() : 0) {
1214 default: Opc = 0; break;
1215 case PPC::FABSS: Opc = PPC::FNABSS; break;
1216 case PPC::FABSD: Opc = PPC::FNABSD; break;
1217 case PPC::FMADD: Opc = PPC::FNMADD; break;
1218 case PPC::FMADDS: Opc = PPC::FNMADDS; break;
1219 case PPC::FMSUB: Opc = PPC::FNMSUB; break;
1220 case PPC::FMSUBS: Opc = PPC::FNMSUBS; break;
1222 // If we inverted the opcode, then emit the new instruction with the
1223 // inverted opcode and the original instruction's operands. Otherwise,
1224 // fall through and generate a fneg instruction.
1226 if (Opc == PPC::FNABSS || Opc == PPC::FNABSD)
1227 CurDAG->SelectNodeTo(N, Opc, Ty, Val.getOperand(0));
1229 CurDAG->SelectNodeTo(N, Opc, Ty, Val.getOperand(0),
1230 Val.getOperand(1), Val.getOperand(2));
1231 return SDOperand(N, 0);
1235 CurDAG->SelectNodeTo(N, PPC::FNEGS, MVT::f32, Val);
1237 CurDAG->SelectNodeTo(N, PPC::FNEGD, MVT::f64, Val);
1238 return SDOperand(N, 0);
1241 MVT::ValueType Ty = N->getValueType(0);
1242 CurDAG->SelectNodeTo(N, Ty == MVT::f64 ? PPC::FSQRT : PPC::FSQRTS, Ty,
1243 Select(N->getOperand(0)));
1244 return SDOperand(N, 0);
1250 case ISD::SEXTLOAD: {
1252 bool isIdx = SelectAddr(N->getOperand(1), Op1, Op2);
1254 MVT::ValueType TypeBeingLoaded = (N->getOpcode() == ISD::LOAD) ?
1255 N->getValueType(0) : cast<VTSDNode>(N->getOperand(3))->getVT();
1257 switch (TypeBeingLoaded) {
1258 default: N->dump(); assert(0 && "Cannot load this type!");
1260 case MVT::i8: Opc = isIdx ? PPC::LBZX : PPC::LBZ; break;
1262 if (N->getOpcode() == ISD::SEXTLOAD) { // SEXT load?
1263 Opc = isIdx ? PPC::LHAX : PPC::LHA;
1265 Opc = isIdx ? PPC::LHZX : PPC::LHZ;
1268 case MVT::i32: Opc = isIdx ? PPC::LWZX : PPC::LWZ; break;
1269 case MVT::f32: Opc = isIdx ? PPC::LFSX : PPC::LFS; break;
1270 case MVT::f64: Opc = isIdx ? PPC::LFDX : PPC::LFD; break;
1273 // If this is an f32 -> f64 load, emit the f32 load, then use an 'extending
1275 if (TypeBeingLoaded != MVT::f32 || N->getOpcode() == ISD::LOAD) {
1276 CurDAG->SelectNodeTo(N, Opc, N->getValueType(0), MVT::Other,
1277 Op1, Op2, Select(N->getOperand(0)));
1278 return SDOperand(N, Op.ResNo);
1280 std::vector<SDOperand> Ops;
1283 Ops.push_back(Select(N->getOperand(0)));
1284 SDOperand Res = CurDAG->getTargetNode(Opc, MVT::f32, MVT::Other, Ops);
1285 SDOperand Ext = CurDAG->getTargetNode(PPC::FMRSD, MVT::f64, Res);
1286 CodeGenMap[Op.getValue(0)] = Ext;
1287 CodeGenMap[Op.getValue(1)] = Res.getValue(1);
1289 return Res.getValue(1);
1295 case ISD::TRUNCSTORE:
1297 SDOperand AddrOp1, AddrOp2;
1298 bool isIdx = SelectAddr(N->getOperand(2), AddrOp1, AddrOp2);
1301 if (N->getOpcode() == ISD::STORE) {
1302 switch (N->getOperand(1).getValueType()) {
1303 default: assert(0 && "unknown Type in store");
1304 case MVT::i32: Opc = isIdx ? PPC::STWX : PPC::STW; break;
1305 case MVT::f64: Opc = isIdx ? PPC::STFDX : PPC::STFD; break;
1306 case MVT::f32: Opc = isIdx ? PPC::STFSX : PPC::STFS; break;
1308 } else { //ISD::TRUNCSTORE
1309 switch(cast<VTSDNode>(N->getOperand(4))->getVT()) {
1310 default: assert(0 && "unknown Type in store");
1311 case MVT::i8: Opc = isIdx ? PPC::STBX : PPC::STB; break;
1312 case MVT::i16: Opc = isIdx ? PPC::STHX : PPC::STH; break;
1316 CurDAG->SelectNodeTo(N, Opc, MVT::Other, Select(N->getOperand(1)),
1317 AddrOp1, AddrOp2, Select(N->getOperand(0)));
1318 return SDOperand(N, 0);
1321 case ISD::SELECT_CC: {
1322 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
1324 // handle the setcc cases here. select_cc lhs, 0, 1, 0, cc
1325 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N->getOperand(1)))
1326 if (ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N->getOperand(2)))
1327 if (ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N->getOperand(3)))
1328 if (N1C->isNullValue() && N3C->isNullValue() &&
1329 N2C->getValue() == 1ULL && CC == ISD::SETNE) {
1330 SDOperand LHS = Select(N->getOperand(0));
1332 CurDAG->getTargetNode(PPC::ADDIC, MVT::i32, MVT::Flag,
1333 LHS, getI32Imm(~0U));
1334 CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, Tmp, LHS,
1336 return SDOperand(N, 0);
1339 SDOperand CCReg = SelectCC(N->getOperand(0), N->getOperand(1), CC);
1340 unsigned BROpc = getBCCForSetCC(CC);
1342 bool isFP = MVT::isFloatingPoint(N->getValueType(0));
1343 unsigned SelectCCOp;
1344 if (MVT::isInteger(N->getValueType(0)))
1345 SelectCCOp = PPC::SELECT_CC_Int;
1346 else if (N->getValueType(0) == MVT::f32)
1347 SelectCCOp = PPC::SELECT_CC_F4;
1349 SelectCCOp = PPC::SELECT_CC_F8;
1350 CurDAG->SelectNodeTo(N, SelectCCOp, N->getValueType(0), CCReg,
1351 Select(N->getOperand(2)), Select(N->getOperand(3)),
1353 return SDOperand(N, 0);
1356 case ISD::CALLSEQ_START:
1357 case ISD::CALLSEQ_END: {
1358 unsigned Amt = cast<ConstantSDNode>(N->getOperand(1))->getValue();
1359 unsigned Opc = N->getOpcode() == ISD::CALLSEQ_START ?
1360 PPC::ADJCALLSTACKDOWN : PPC::ADJCALLSTACKUP;
1361 CurDAG->SelectNodeTo(N, Opc, MVT::Other,
1362 getI32Imm(Amt), Select(N->getOperand(0)));
1363 return SDOperand(N, 0);
1366 case ISD::TAILCALL: {
1367 SDOperand Chain = Select(N->getOperand(0));
1369 unsigned CallOpcode;
1370 std::vector<SDOperand> CallOperands;
1372 if (GlobalAddressSDNode *GASD =
1373 dyn_cast<GlobalAddressSDNode>(N->getOperand(1))) {
1374 CallOpcode = PPC::CALLpcrel;
1375 CallOperands.push_back(CurDAG->getTargetGlobalAddress(GASD->getGlobal(),
1377 } else if (ExternalSymbolSDNode *ESSDN =
1378 dyn_cast<ExternalSymbolSDNode>(N->getOperand(1))) {
1379 CallOpcode = PPC::CALLpcrel;
1380 CallOperands.push_back(N->getOperand(1));
1382 // Copy the callee address into the CTR register.
1383 SDOperand Callee = Select(N->getOperand(1));
1384 Chain = CurDAG->getTargetNode(PPC::MTCTR, MVT::Other, Callee, Chain);
1386 // Copy the callee address into R12 on darwin.
1387 SDOperand R12 = CurDAG->getRegister(PPC::R12, MVT::i32);
1388 Chain = CurDAG->getNode(ISD::CopyToReg, MVT::Other, Chain, R12, Callee);
1390 CallOperands.push_back(getI32Imm(20)); // Information to encode indcall
1391 CallOperands.push_back(getI32Imm(0)); // Information to encode indcall
1392 CallOperands.push_back(R12);
1393 CallOpcode = PPC::CALLindirect;
1396 unsigned GPR_idx = 0, FPR_idx = 0;
1397 static const unsigned GPR[] = {
1398 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1399 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1401 static const unsigned FPR[] = {
1402 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1403 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1406 SDOperand InFlag; // Null incoming flag value.
1408 for (unsigned i = 2, e = N->getNumOperands(); i != e; ++i) {
1409 unsigned DestReg = 0;
1410 MVT::ValueType RegTy = N->getOperand(i).getValueType();
1411 if (RegTy == MVT::i32) {
1412 assert(GPR_idx < 8 && "Too many int args");
1413 DestReg = GPR[GPR_idx++];
1415 assert(MVT::isFloatingPoint(N->getOperand(i).getValueType()) &&
1416 "Unpromoted integer arg?");
1417 assert(FPR_idx < 13 && "Too many fp args");
1418 DestReg = FPR[FPR_idx++];
1421 if (N->getOperand(i).getOpcode() != ISD::UNDEF) {
1422 SDOperand Val = Select(N->getOperand(i));
1423 Chain = CurDAG->getCopyToReg(Chain, DestReg, Val, InFlag);
1424 InFlag = Chain.getValue(1);
1425 CallOperands.push_back(CurDAG->getRegister(DestReg, RegTy));
1429 // Finally, once everything is in registers to pass to the call, emit the
1432 CallOperands.push_back(InFlag); // Strong dep on register copies.
1434 CallOperands.push_back(Chain); // Weak dep on whatever occurs before
1435 Chain = CurDAG->getTargetNode(CallOpcode, MVT::Other, MVT::Flag,
1438 std::vector<SDOperand> CallResults;
1440 // If the call has results, copy the values out of the ret val registers.
1441 switch (N->getValueType(0)) {
1442 default: assert(0 && "Unexpected ret value!");
1443 case MVT::Other: break;
1445 if (N->getValueType(1) == MVT::i32) {
1446 Chain = CurDAG->getCopyFromReg(Chain, PPC::R4, MVT::i32,
1447 Chain.getValue(1)).getValue(1);
1448 CallResults.push_back(Chain.getValue(0));
1449 Chain = CurDAG->getCopyFromReg(Chain, PPC::R3, MVT::i32,
1450 Chain.getValue(2)).getValue(1);
1451 CallResults.push_back(Chain.getValue(0));
1453 Chain = CurDAG->getCopyFromReg(Chain, PPC::R3, MVT::i32,
1454 Chain.getValue(1)).getValue(1);
1455 CallResults.push_back(Chain.getValue(0));
1460 Chain = CurDAG->getCopyFromReg(Chain, PPC::F1, N->getValueType(0),
1461 Chain.getValue(1)).getValue(1);
1462 CallResults.push_back(Chain.getValue(0));
1466 CallResults.push_back(Chain);
1467 for (unsigned i = 0, e = CallResults.size(); i != e; ++i)
1468 CodeGenMap[Op.getValue(i)] = CallResults[i];
1469 return CallResults[Op.ResNo];
1472 SDOperand Chain = Select(N->getOperand(0)); // Token chain.
1474 if (N->getNumOperands() == 2) {
1475 SDOperand Val = Select(N->getOperand(1));
1476 if (N->getOperand(1).getValueType() == MVT::i32) {
1477 Chain = CurDAG->getCopyToReg(Chain, PPC::R3, Val);
1479 assert(MVT::isFloatingPoint(N->getOperand(1).getValueType()));
1480 Chain = CurDAG->getCopyToReg(Chain, PPC::F1, Val);
1482 } else if (N->getNumOperands() > 1) {
1483 assert(N->getOperand(1).getValueType() == MVT::i32 &&
1484 N->getOperand(2).getValueType() == MVT::i32 &&
1485 N->getNumOperands() == 3 && "Unknown two-register ret value!");
1486 Chain = CurDAG->getCopyToReg(Chain, PPC::R4, Select(N->getOperand(1)));
1487 Chain = CurDAG->getCopyToReg(Chain, PPC::R3, Select(N->getOperand(2)));
1490 // Finally, select this to a blr (return) instruction.
1491 CurDAG->SelectNodeTo(N, PPC::BLR, MVT::Other, Chain);
1492 return SDOperand(N, 0);
1495 CurDAG->SelectNodeTo(N, PPC::B, MVT::Other, N->getOperand(1),
1496 Select(N->getOperand(0)));
1497 return SDOperand(N, 0);
1499 case ISD::BRTWOWAY_CC: {
1500 SDOperand Chain = Select(N->getOperand(0));
1501 MachineBasicBlock *Dest =
1502 cast<BasicBlockSDNode>(N->getOperand(4))->getBasicBlock();
1503 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
1504 SDOperand CondCode = SelectCC(N->getOperand(2), N->getOperand(3), CC);
1506 // If this is a two way branch, then grab the fallthrough basic block
1507 // argument and build a PowerPC branch pseudo-op, suitable for long branch
1508 // conversion if necessary by the branch selection pass. Otherwise, emit a
1509 // standard conditional branch.
1510 if (N->getOpcode() == ISD::BRTWOWAY_CC) {
1511 SDOperand CondTrueBlock = N->getOperand(4);
1512 SDOperand CondFalseBlock = N->getOperand(5);
1514 // If the false case is the current basic block, then this is a self loop.
1515 // We do not want to emit "Loop: ... brcond Out; br Loop", as it adds an
1516 // extra dispatch group to the loop. Instead, invert the condition and
1517 // emit "Loop: ... br!cond Loop; br Out
1518 if (cast<BasicBlockSDNode>(CondFalseBlock)->getBasicBlock() == BB) {
1519 std::swap(CondTrueBlock, CondFalseBlock);
1520 CC = getSetCCInverse(CC,
1521 MVT::isInteger(N->getOperand(2).getValueType()));
1524 unsigned Opc = getBCCForSetCC(CC);
1525 SDOperand CB = CurDAG->getTargetNode(PPC::COND_BRANCH, MVT::Other,
1526 CondCode, getI32Imm(Opc),
1527 CondTrueBlock, CondFalseBlock,
1529 CurDAG->SelectNodeTo(N, PPC::B, MVT::Other, CondFalseBlock, CB);
1531 // Iterate to the next basic block
1532 ilist<MachineBasicBlock>::iterator It = BB;
1535 // If the fallthrough path is off the end of the function, which would be
1536 // undefined behavior, set it to be the same as the current block because
1537 // we have nothing better to set it to, and leaving it alone will cause
1538 // the PowerPC Branch Selection pass to crash.
1539 if (It == BB->getParent()->end()) It = Dest;
1540 CurDAG->SelectNodeTo(N, PPC::COND_BRANCH, MVT::Other, CondCode,
1541 getI32Imm(getBCCForSetCC(CC)), N->getOperand(4),
1542 CurDAG->getBasicBlock(It), Chain);
1544 return SDOperand(N, 0);
1548 return SelectCode(Op);
1552 /// createPPC32ISelDag - This pass converts a legalized DAG into a
1553 /// PowerPC-specific DAG, ready for instruction scheduling.
1555 FunctionPass *llvm::createPPC32ISelDag(TargetMachine &TM) {
1556 return new PPC32DAGToDAGISel(TM);