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.
11 // Magic number generation for integer divide from the PowerPC Compiler Writer's
12 // Guide, section 3.2.3.5
14 //===----------------------------------------------------------------------===//
17 #include "PowerPCInstrBuilder.h"
18 #include "PowerPCInstrInfo.h"
19 #include "PPC32TargetMachine.h"
20 #include "PPC32ISelLowering.h"
21 #include "llvm/Constants.h"
22 #include "llvm/Function.h"
23 #include "llvm/CodeGen/MachineConstantPool.h"
24 #include "llvm/CodeGen/MachineFunction.h"
25 #include "llvm/CodeGen/MachineFrameInfo.h"
26 #include "llvm/CodeGen/SelectionDAG.h"
27 #include "llvm/CodeGen/SelectionDAGISel.h"
28 #include "llvm/CodeGen/SSARegMap.h"
29 #include "llvm/Target/TargetData.h"
30 #include "llvm/Target/TargetOptions.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/MathExtras.h"
33 #include "llvm/ADT/Statistic.h"
39 Statistic<> Recorded("ppc-codegen", "Number of recording ops emitted");
40 Statistic<> FusedFP ("ppc-codegen", "Number of fused fp operations");
41 Statistic<> FrameOff("ppc-codegen", "Number of frame idx offsets collapsed");
43 //===--------------------------------------------------------------------===//
44 // ISel - PPC32 specific code to select PPC32 machine instructions for
45 // SelectionDAG operations.
46 //===--------------------------------------------------------------------===//
48 class ISel : public SelectionDAGISel {
49 PPC32TargetLowering PPC32Lowering;
50 SelectionDAG *ISelDAG; // Hack to support us having a dag->dag transform
51 // for sdiv and udiv until it is put into the future
54 /// ExprMap - As shared expressions are codegen'd, we keep track of which
55 /// vreg the value is produced in, so we only emit one copy of each compiled
57 std::map<SDOperand, unsigned> ExprMap;
59 unsigned GlobalBaseReg;
60 bool GlobalBaseInitialized;
63 ISel(TargetMachine &TM) : SelectionDAGISel(PPC32Lowering), PPC32Lowering(TM),
66 /// runOnFunction - Override this function in order to reset our per-function
68 virtual bool runOnFunction(Function &Fn) {
69 // Make sure we re-emit a set of the global base reg if necessary
70 GlobalBaseInitialized = false;
71 return SelectionDAGISel::runOnFunction(Fn);
74 /// InstructionSelectBasicBlock - This callback is invoked by
75 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
76 virtual void InstructionSelectBasicBlock(SelectionDAG &DAG) {
78 // Codegen the basic block.
80 Select(DAG.getRoot());
82 // Clear state used for selection.
87 // convenience functions for virtual register creation
88 inline unsigned MakeIntReg() {
89 return RegMap->createVirtualRegister(PPC32::GPRCRegisterClass);
91 inline unsigned MakeFPReg() {
92 return RegMap->createVirtualRegister(PPC32::FPRCRegisterClass);
95 // dag -> dag expanders for integer divide by constant
96 SDOperand BuildSDIVSequence(SDOperand N);
97 SDOperand BuildUDIVSequence(SDOperand N);
99 unsigned getGlobalBaseReg();
100 unsigned getConstDouble(double floatVal, unsigned Result);
101 void MoveCRtoGPR(unsigned CCReg, ISD::CondCode CC, unsigned Result);
102 bool SelectBitfieldInsert(SDOperand OR, unsigned Result);
103 unsigned FoldIfWideZeroExtend(SDOperand N);
104 unsigned SelectCC(SDOperand LHS, SDOperand RHS, ISD::CondCode CC);
105 bool SelectIntImmediateExpr(SDOperand N, unsigned Result,
106 unsigned OCHi, unsigned OCLo,
107 bool IsArithmetic = false, bool Negate = false);
108 unsigned SelectExpr(SDOperand N, bool Recording=false);
109 void Select(SDOperand N);
111 unsigned SelectAddr(SDOperand N, unsigned& Reg, int& offset);
112 void SelectBranchCC(SDOperand N);
114 virtual const char *getPassName() const {
115 return "PowerPC Pattern Instruction Selection";
119 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
120 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
121 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
122 // not, since all 1s are not contiguous.
123 static bool isRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME) {
124 if (isShiftedMask_32(Val)) {
125 // look for the first non-zero bit
126 MB = CountLeadingZeros_32(Val);
127 // look for the first zero bit after the run of ones
128 ME = CountLeadingZeros_32((Val - 1) ^ Val);
130 } else if (isShiftedMask_32(Val = ~Val)) { // invert mask
131 // effectively look for the first zero bit
132 ME = CountLeadingZeros_32(Val) - 1;
133 // effectively look for the first one bit after the run of zeros
134 MB = CountLeadingZeros_32((Val - 1) ^ Val) + 1;
141 // isRotateAndMask - Returns true if Mask and Shift can be folded in to a rotate
142 // and mask opcode and mask operation.
143 static bool isRotateAndMask(unsigned Opcode, unsigned Shift, unsigned Mask,
145 unsigned &SH, unsigned &MB, unsigned &ME) {
146 if (Shift > 31) return false;
147 unsigned Indeterminant = ~0; // bit mask marking indeterminant results
149 if (Opcode == ISD::SHL) { // shift left
150 // apply shift to mask if it comes first
151 if (IsShiftMask) Mask = Mask << Shift;
152 // determine which bits are made indeterminant by shift
153 Indeterminant = ~(0xFFFFFFFFu << Shift);
154 } else if (Opcode == ISD::SRA || Opcode == ISD::SRL) { // shift rights
155 // apply shift to mask if it comes first
156 if (IsShiftMask) Mask = Mask >> Shift;
157 // determine which bits are made indeterminant by shift
158 Indeterminant = ~(0xFFFFFFFFu >> Shift);
159 // adjust for the left rotate
163 // if the mask doesn't intersect any Indeterminant bits
164 if (Mask && !(Mask & Indeterminant)) {
166 // make sure the mask is still a mask (wrap arounds may not be)
167 return isRunOfOnes(Mask, MB, ME);
174 // isIntImmediate - This method tests to see if a constant operand.
175 // If so Imm will receive the 32 bit value.
176 static bool isIntImmediate(SDOperand N, unsigned& Imm) {
178 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
180 Imm = (unsigned)CN->getValue();
188 // isOpcWithIntImmediate - This method tests to see if the node is a specific
189 // opcode and that it has a immediate integer right operand.
190 // If so Imm will receive the 32 bit value.
191 static bool isOpcWithIntImmediate(SDOperand N, unsigned Opc, unsigned& Imm) {
192 return N.getOpcode() == Opc && isIntImmediate(N.getOperand(1), Imm);
195 // isOprShiftImm - Returns true if the specified operand is a shift opcode with
196 // a immediate shift count less than 32.
197 static bool isOprShiftImm(SDOperand N, unsigned& Opc, unsigned& SH) {
199 return (Opc == ISD::SHL || Opc == ISD::SRL || Opc == ISD::SRA) &&
200 isIntImmediate(N.getOperand(1), SH) && SH < 32;
203 // isOprNot - Returns true if the specified operand is an xor with immediate -1.
204 static bool isOprNot(SDOperand N) {
206 return isOpcWithIntImmediate(N, ISD::XOR, Imm) && (signed)Imm == -1;
209 // Immediate constant composers.
210 // Lo16 - grabs the lo 16 bits from a 32 bit constant.
211 // Hi16 - grabs the hi 16 bits from a 32 bit constant.
212 // HA16 - computes the hi bits required if the lo bits are add/subtracted in
214 static unsigned Lo16(unsigned x) { return x & 0x0000FFFF; }
215 static unsigned Hi16(unsigned x) { return Lo16(x >> 16); }
216 static unsigned HA16(unsigned x) { return Hi16((signed)x - (signed short)x); }
218 /// NodeHasRecordingVariant - If SelectExpr can always produce code for
219 /// NodeOpcode that also sets CR0 as a side effect, return true. Otherwise,
221 static bool NodeHasRecordingVariant(unsigned NodeOpcode) {
223 default: return false;
230 /// getBCCForSetCC - Returns the PowerPC condition branch mnemonic corresponding
232 static unsigned getBCCForSetCC(ISD::CondCode CC) {
234 default: assert(0 && "Unknown condition!"); abort();
235 case ISD::SETEQ: return PPC::BEQ;
236 case ISD::SETNE: return PPC::BNE;
238 case ISD::SETLT: return PPC::BLT;
240 case ISD::SETLE: return PPC::BLE;
242 case ISD::SETGT: return PPC::BGT;
244 case ISD::SETGE: return PPC::BGE;
249 /// getCRIdxForSetCC - Return the index of the condition register field
250 /// associated with the SetCC condition, and whether or not the field is
251 /// treated as inverted. That is, lt = 0; ge = 0 inverted.
252 static unsigned getCRIdxForSetCC(ISD::CondCode CC, bool& Inv) {
254 default: assert(0 && "Unknown condition!"); abort();
256 case ISD::SETLT: Inv = false; return 0;
258 case ISD::SETGE: Inv = true; return 0;
260 case ISD::SETGT: Inv = false; return 1;
262 case ISD::SETLE: Inv = true; return 1;
263 case ISD::SETEQ: Inv = false; return 2;
264 case ISD::SETNE: Inv = true; return 2;
269 /// IndexedOpForOp - Return the indexed variant for each of the PowerPC load
270 /// and store immediate instructions.
271 static unsigned IndexedOpForOp(unsigned Opcode) {
273 default: assert(0 && "Unknown opcode!"); abort();
274 case PPC::LBZ: return PPC::LBZX; case PPC::STB: return PPC::STBX;
275 case PPC::LHZ: return PPC::LHZX; case PPC::STH: return PPC::STHX;
276 case PPC::LHA: return PPC::LHAX; case PPC::STW: return PPC::STWX;
277 case PPC::LWZ: return PPC::LWZX; case PPC::STFS: return PPC::STFSX;
278 case PPC::LFS: return PPC::LFSX; case PPC::STFD: return PPC::STFDX;
279 case PPC::LFD: return PPC::LFDX;
284 // Structure used to return the necessary information to codegen an SDIV as
287 int m; // magic number
288 int s; // shift amount
292 unsigned int m; // magic number
293 int a; // add indicator
294 int s; // shift amount
297 /// magic - calculate the magic numbers required to codegen an integer sdiv as
298 /// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
300 static struct ms magic(int d) {
302 unsigned int ad, anc, delta, q1, r1, q2, r2, t;
303 const unsigned int two31 = 0x80000000U;
307 t = two31 + ((unsigned int)d >> 31);
308 anc = t - 1 - t%ad; // absolute value of nc
309 p = 31; // initialize p
310 q1 = two31/anc; // initialize q1 = 2p/abs(nc)
311 r1 = two31 - q1*anc; // initialize r1 = rem(2p,abs(nc))
312 q2 = two31/ad; // initialize q2 = 2p/abs(d)
313 r2 = two31 - q2*ad; // initialize r2 = rem(2p,abs(d))
316 q1 = 2*q1; // update q1 = 2p/abs(nc)
317 r1 = 2*r1; // update r1 = rem(2p/abs(nc))
318 if (r1 >= anc) { // must be unsigned comparison
322 q2 = 2*q2; // update q2 = 2p/abs(d)
323 r2 = 2*r2; // update r2 = rem(2p/abs(d))
324 if (r2 >= ad) { // must be unsigned comparison
329 } while (q1 < delta || (q1 == delta && r1 == 0));
332 if (d < 0) mag.m = -mag.m; // resulting magic number
333 mag.s = p - 32; // resulting shift
337 /// magicu - calculate the magic numbers required to codegen an integer udiv as
338 /// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
339 static struct mu magicu(unsigned d)
342 unsigned int nc, delta, q1, r1, q2, r2;
344 magu.a = 0; // initialize "add" indicator
346 p = 31; // initialize p
347 q1 = 0x80000000/nc; // initialize q1 = 2p/nc
348 r1 = 0x80000000 - q1*nc; // initialize r1 = rem(2p,nc)
349 q2 = 0x7FFFFFFF/d; // initialize q2 = (2p-1)/d
350 r2 = 0x7FFFFFFF - q2*d; // initialize r2 = rem((2p-1),d)
353 if (r1 >= nc - r1 ) {
354 q1 = 2*q1 + 1; // update q1
355 r1 = 2*r1 - nc; // update r1
358 q1 = 2*q1; // update q1
359 r1 = 2*r1; // update r1
361 if (r2 + 1 >= d - r2) {
362 if (q2 >= 0x7FFFFFFF) magu.a = 1;
363 q2 = 2*q2 + 1; // update q2
364 r2 = 2*r2 + 1 - d; // update r2
367 if (q2 >= 0x80000000) magu.a = 1;
368 q2 = 2*q2; // update q2
369 r2 = 2*r2 + 1; // update r2
372 } while (p < 64 && (q1 < delta || (q1 == delta && r1 == 0)));
373 magu.m = q2 + 1; // resulting magic number
374 magu.s = p - 32; // resulting shift
379 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
380 /// return a DAG expression to select that will generate the same value by
381 /// multiplying by a magic number. See:
382 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
383 SDOperand ISel::BuildSDIVSequence(SDOperand N) {
384 int d = (int)cast<ConstantSDNode>(N.getOperand(1))->getSignExtended();
385 ms magics = magic(d);
386 // Multiply the numerator (operand 0) by the magic value
387 SDOperand Q = ISelDAG->getNode(ISD::MULHS, MVT::i32, N.getOperand(0),
388 ISelDAG->getConstant(magics.m, MVT::i32));
389 // If d > 0 and m < 0, add the numerator
390 if (d > 0 && magics.m < 0)
391 Q = ISelDAG->getNode(ISD::ADD, MVT::i32, Q, N.getOperand(0));
392 // If d < 0 and m > 0, subtract the numerator.
393 if (d < 0 && magics.m > 0)
394 Q = ISelDAG->getNode(ISD::SUB, MVT::i32, Q, N.getOperand(0));
395 // Shift right algebraic if shift value is nonzero
397 Q = ISelDAG->getNode(ISD::SRA, MVT::i32, Q,
398 ISelDAG->getConstant(magics.s, MVT::i32));
399 // Extract the sign bit and add it to the quotient
401 ISelDAG->getNode(ISD::SRL, MVT::i32, Q, ISelDAG->getConstant(31, MVT::i32));
402 return ISelDAG->getNode(ISD::ADD, MVT::i32, Q, T);
405 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
406 /// return a DAG expression to select that will generate the same value by
407 /// multiplying by a magic number. See:
408 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
409 SDOperand ISel::BuildUDIVSequence(SDOperand N) {
411 (unsigned)cast<ConstantSDNode>(N.getOperand(1))->getSignExtended();
412 mu magics = magicu(d);
413 // Multiply the numerator (operand 0) by the magic value
414 SDOperand Q = ISelDAG->getNode(ISD::MULHU, MVT::i32, N.getOperand(0),
415 ISelDAG->getConstant(magics.m, MVT::i32));
417 Q = ISelDAG->getNode(ISD::SRL, MVT::i32, Q,
418 ISelDAG->getConstant(magics.s, MVT::i32));
420 SDOperand NPQ = ISelDAG->getNode(ISD::SUB, MVT::i32, N.getOperand(0), Q);
421 NPQ = ISelDAG->getNode(ISD::SRL, MVT::i32, NPQ,
422 ISelDAG->getConstant(1, MVT::i32));
423 NPQ = ISelDAG->getNode(ISD::ADD, MVT::i32, NPQ, Q);
424 Q = ISelDAG->getNode(ISD::SRL, MVT::i32, NPQ,
425 ISelDAG->getConstant(magics.s-1, MVT::i32));
430 /// getGlobalBaseReg - Output the instructions required to put the
431 /// base address to use for accessing globals into a register.
433 unsigned ISel::getGlobalBaseReg() {
434 if (!GlobalBaseInitialized) {
435 // Insert the set of GlobalBaseReg into the first MBB of the function
436 MachineBasicBlock &FirstMBB = BB->getParent()->front();
437 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
438 GlobalBaseReg = MakeIntReg();
439 BuildMI(FirstMBB, MBBI, PPC::MovePCtoLR, 0, PPC::LR);
440 BuildMI(FirstMBB, MBBI, PPC::MFLR, 1, GlobalBaseReg).addReg(PPC::LR);
441 GlobalBaseInitialized = true;
443 return GlobalBaseReg;
446 /// getConstDouble - Loads a floating point value into a register, via the
447 /// Constant Pool. Optionally takes a register in which to load the value.
448 unsigned ISel::getConstDouble(double doubleVal, unsigned Result=0) {
449 unsigned Tmp1 = MakeIntReg();
450 if (0 == Result) Result = MakeFPReg();
451 MachineConstantPool *CP = BB->getParent()->getConstantPool();
452 ConstantFP *CFP = ConstantFP::get(Type::DoubleTy, doubleVal);
453 unsigned CPI = CP->getConstantPoolIndex(CFP);
455 BuildMI(BB, PPC::ADDIS, 2, Tmp1).addReg(getGlobalBaseReg())
456 .addConstantPoolIndex(CPI);
458 BuildMI(BB, PPC::LIS, 1, Tmp1).addConstantPoolIndex(CPI);
459 BuildMI(BB, PPC::LFD, 2, Result).addConstantPoolIndex(CPI).addReg(Tmp1);
463 /// MoveCRtoGPR - Move CCReg[Idx] to the least significant bit of Result. If
464 /// Inv is true, then invert the result.
465 void ISel::MoveCRtoGPR(unsigned CCReg, ISD::CondCode CC, unsigned Result){
467 unsigned IntCR = MakeIntReg();
468 unsigned Idx = getCRIdxForSetCC(CC, Inv);
469 BuildMI(BB, PPC::MCRF, 1, PPC::CR7).addReg(CCReg);
471 TLI.getTargetMachine().getSubtarget<PPCSubtarget>().isGigaProcessor();
472 BuildMI(BB, GPOpt ? PPC::MFOCRF : PPC::MFCR, 1, IntCR).addReg(PPC::CR7);
474 unsigned Tmp1 = MakeIntReg();
475 BuildMI(BB, PPC::RLWINM, 4, Tmp1).addReg(IntCR).addImm(32-(3-Idx))
476 .addImm(31).addImm(31);
477 BuildMI(BB, PPC::XORI, 2, Result).addReg(Tmp1).addImm(1);
479 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(IntCR).addImm(32-(3-Idx))
480 .addImm(31).addImm(31);
484 /// SelectBitfieldInsert - turn an or of two masked values into
485 /// the rotate left word immediate then mask insert (rlwimi) instruction.
486 /// Returns true on success, false if the caller still needs to select OR.
488 /// Patterns matched:
489 /// 1. or shl, and 5. or and, and
490 /// 2. or and, shl 6. or shl, shr
491 /// 3. or shr, and 7. or shr, shl
493 bool ISel::SelectBitfieldInsert(SDOperand OR, unsigned Result) {
494 bool IsRotate = false;
495 unsigned TgtMask = 0xFFFFFFFF, InsMask = 0xFFFFFFFF, Amount = 0;
498 SDOperand Op0 = OR.getOperand(0);
499 SDOperand Op1 = OR.getOperand(1);
501 unsigned Op0Opc = Op0.getOpcode();
502 unsigned Op1Opc = Op1.getOpcode();
504 // Verify that we have the correct opcodes
505 if (ISD::SHL != Op0Opc && ISD::SRL != Op0Opc && ISD::AND != Op0Opc)
507 if (ISD::SHL != Op1Opc && ISD::SRL != Op1Opc && ISD::AND != Op1Opc)
510 // Generate Mask value for Target
511 if (isIntImmediate(Op0.getOperand(1), Value)) {
513 case ISD::SHL: TgtMask <<= Value; break;
514 case ISD::SRL: TgtMask >>= Value; break;
515 case ISD::AND: TgtMask &= Value; break;
521 // Generate Mask value for Insert
522 if (isIntImmediate(Op1.getOperand(1), Value)) {
527 if (Op0Opc == ISD::SRL) IsRotate = true;
533 if (Op0Opc == ISD::SHL) IsRotate = true;
545 // If both of the inputs are ANDs and one of them has a logical shift by
546 // constant as its input, make that the inserted value so that we can combine
547 // the shift into the rotate part of the rlwimi instruction
548 if (Op0Opc == ISD::AND && Op1Opc == ISD::AND) {
549 if (Op1.getOperand(0).getOpcode() == ISD::SHL ||
550 Op1.getOperand(0).getOpcode() == ISD::SRL) {
551 if (isIntImmediate(Op1.getOperand(0).getOperand(1), Value)) {
552 Amount = Op1.getOperand(0).getOpcode() == ISD::SHL ?
554 Tmp3 = SelectExpr(Op1.getOperand(0).getOperand(0));
556 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL ||
557 Op0.getOperand(0).getOpcode() == ISD::SRL) {
558 if (isIntImmediate(Op0.getOperand(0).getOperand(1), Value)) {
560 std::swap(TgtMask, InsMask);
561 Amount = Op1.getOperand(0).getOpcode() == ISD::SHL ?
563 Tmp3 = SelectExpr(Op1.getOperand(0).getOperand(0));
568 // Verify that the Target mask and Insert mask together form a full word mask
569 // and that the Insert mask is a run of set bits (which implies both are runs
570 // of set bits). Given that, Select the arguments and generate the rlwimi
573 if (((TgtMask & InsMask) == 0) && isRunOfOnes(InsMask, MB, ME)) {
575 bool fullMask = (TgtMask ^ InsMask) == 0xFFFFFFFF;
576 // Check for rotlwi / rotrwi here, a special case of bitfield insert
577 // where both bitfield halves are sourced from the same value.
578 if (IsRotate && fullMask &&
579 OR.getOperand(0).getOperand(0) == OR.getOperand(1).getOperand(0)) {
580 Tmp1 = SelectExpr(OR.getOperand(0).getOperand(0));
581 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp1).addImm(Amount)
582 .addImm(0).addImm(31);
585 if (Op0Opc == ISD::AND && fullMask)
586 Tmp1 = SelectExpr(Op0.getOperand(0));
588 Tmp1 = SelectExpr(Op0);
589 Tmp2 = Tmp3 ? Tmp3 : SelectExpr(Op1.getOperand(0));
590 BuildMI(BB, PPC::RLWIMI, 5, Result).addReg(Tmp1).addReg(Tmp2)
591 .addImm(Amount).addImm(MB).addImm(ME);
597 /// FoldIfWideZeroExtend - 32 bit PowerPC implicit masks shift amounts to the
598 /// low six bits. If the shift amount is an ISD::AND node with a mask that is
599 /// wider than the implicit mask, then we can get rid of the AND and let the
600 /// shift do the mask.
601 unsigned ISel::FoldIfWideZeroExtend(SDOperand N) {
603 if (isOpcWithIntImmediate(N, ISD::AND, C) && isMask_32(C) && C > 63)
604 return SelectExpr(N.getOperand(0));
606 return SelectExpr(N);
609 unsigned ISel::SelectCC(SDOperand LHS, SDOperand RHS, ISD::CondCode CC) {
610 unsigned Result, Tmp1, Tmp2;
611 bool AlreadySelected = false;
612 static const unsigned CompareOpcodes[] =
613 { PPC::FCMPU, PPC::FCMPU, PPC::CMPW, PPC::CMPLW };
615 // Allocate a condition register for this expression
616 Result = RegMap->createVirtualRegister(PPC32::CRRCRegisterClass);
618 // Use U to determine whether the SETCC immediate range is signed or not.
619 bool U = ISD::isUnsignedIntSetCC(CC);
620 if (isIntImmediate(RHS, Tmp2) &&
621 ((U && isUInt16(Tmp2)) || (!U && isInt16(Tmp2)))) {
623 // For comparisons against zero, we can implicity set CR0 if a recording
624 // variant (e.g. 'or.' instead of 'or') of the instruction that defines
625 // operand zero of the SetCC node is available.
627 NodeHasRecordingVariant(LHS.getOpcode()) && LHS.Val->hasOneUse()) {
628 RecordSuccess = false;
629 Tmp1 = SelectExpr(LHS, true);
632 BuildMI(BB, PPC::MCRF, 1, Result).addReg(PPC::CR0);
635 AlreadySelected = true;
637 // If we could not implicitly set CR0, then emit a compare immediate
639 if (!AlreadySelected) Tmp1 = SelectExpr(LHS);
641 BuildMI(BB, PPC::CMPLWI, 2, Result).addReg(Tmp1).addImm(Tmp2);
643 BuildMI(BB, PPC::CMPWI, 2, Result).addReg(Tmp1).addSImm(Tmp2);
645 bool IsInteger = MVT::isInteger(LHS.getValueType());
646 unsigned CompareOpc = CompareOpcodes[2 * IsInteger + U];
647 Tmp1 = SelectExpr(LHS);
648 Tmp2 = SelectExpr(RHS);
649 BuildMI(BB, CompareOpc, 2, Result).addReg(Tmp1).addReg(Tmp2);
654 /// Check to see if the load is a constant offset from a base register.
655 unsigned ISel::SelectAddr(SDOperand N, unsigned& Reg, int& offset)
657 unsigned imm = 0, opcode = N.getOpcode();
658 if (N.getOpcode() == ISD::ADD) {
659 bool isFrame = N.getOperand(0).getOpcode() == ISD::FrameIndex;
660 if (isIntImmediate(N.getOperand(1), imm) && isInt16(imm)) {
664 Reg = cast<FrameIndexSDNode>(N.getOperand(0))->getIndex();
667 Reg = SelectExpr(N.getOperand(0));
671 Reg = SelectExpr(N.getOperand(0));
672 offset = SelectExpr(N.getOperand(1));
676 // Now check if we're dealing with a global, and whether or not we should emit
677 // an optimized load or store for statics.
678 if(GlobalAddressSDNode *GN = dyn_cast<GlobalAddressSDNode>(N)) {
679 GlobalValue *GV = GN->getGlobal();
680 if (!GV->hasWeakLinkage() && !GV->isExternal()) {
681 unsigned GlobalHi = MakeIntReg();
683 BuildMI(BB, PPC::ADDIS, 2, GlobalHi).addReg(getGlobalBaseReg())
684 .addGlobalAddress(GV);
686 BuildMI(BB, PPC::LIS, 1, GlobalHi).addGlobalAddress(GV);
697 void ISel::SelectBranchCC(SDOperand N)
699 MachineBasicBlock *Dest =
700 cast<BasicBlockSDNode>(N.getOperand(4))->getBasicBlock();
702 Select(N.getOperand(0)); //chain
703 ISD::CondCode CC = cast<CondCodeSDNode>(N.getOperand(1))->get();
704 unsigned CCReg = SelectCC(N.getOperand(2), N.getOperand(3), CC);
705 unsigned Opc = getBCCForSetCC(CC);
707 // Iterate to the next basic block
708 ilist<MachineBasicBlock>::iterator It = BB;
711 // If this is a two way branch, then grab the fallthrough basic block argument
712 // and build a PowerPC branch pseudo-op, suitable for long branch conversion
713 // if necessary by the branch selection pass. Otherwise, emit a standard
714 // conditional branch.
715 if (N.getOpcode() == ISD::BRTWOWAY_CC) {
716 MachineBasicBlock *Fallthrough =
717 cast<BasicBlockSDNode>(N.getOperand(5))->getBasicBlock();
719 BuildMI(BB, PPC::COND_BRANCH, 4).addReg(CCReg).addImm(Opc)
720 .addMBB(Dest).addMBB(Fallthrough);
721 if (Fallthrough != It)
722 BuildMI(BB, PPC::B, 1).addMBB(Fallthrough);
724 if (Fallthrough != It) {
725 Opc = PPC32InstrInfo::invertPPCBranchOpcode(Opc);
726 BuildMI(BB, PPC::COND_BRANCH, 4).addReg(CCReg).addImm(Opc)
727 .addMBB(Fallthrough).addMBB(Dest);
731 // If the fallthrough path is off the end of the function, which would be
732 // undefined behavior, set it to be the same as the current block because
733 // we have nothing better to set it to, and leaving it alone will cause the
734 // PowerPC Branch Selection pass to crash.
735 if (It == BB->getParent()->end()) It = Dest;
736 BuildMI(BB, PPC::COND_BRANCH, 4).addReg(CCReg).addImm(Opc)
737 .addMBB(Dest).addMBB(It);
742 // SelectIntImmediateExpr - Choose code for opcodes with immediate value.
743 bool ISel::SelectIntImmediateExpr(SDOperand N, unsigned Result,
744 unsigned OCHi, unsigned OCLo,
745 bool IsArithmetic, bool Negate) {
747 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1));
748 // exit if not a constant
749 if (!CN) return false;
751 unsigned C = (unsigned)CN->getValue();
752 // negate if required (ISD::SUB)
754 // get the hi and lo portions of constant
755 unsigned Hi = IsArithmetic ? HA16(C) : Hi16(C);
756 unsigned Lo = Lo16(C);
757 // assume no intermediate result from lo instruction (same as final result)
758 unsigned Tmp = Result;
759 // check if two instructions are needed
761 // exit if usage indicates it would be better to load immediate into a
763 if (CN->use_size() > 2) return false;
764 // need intermediate result for two instructions
768 unsigned Opr0 = SelectExpr(N.getOperand(0));
769 // is a lo instruction needed
771 // generate instruction for lo portion
772 BuildMI(BB, OCLo, 2, Tmp).addReg(Opr0).addImm(Lo);
773 // need to switch out first operand for hi instruction
776 // is a hi instruction needed
778 // generate instruction for hi portion
779 BuildMI(BB, OCHi, 2, Result).addReg(Opr0).addImm(Hi);
784 unsigned ISel::SelectExpr(SDOperand N, bool Recording) {
786 unsigned Tmp1, Tmp2, Tmp3;
788 unsigned opcode = N.getOpcode();
790 SDNode *Node = N.Val;
791 MVT::ValueType DestType = N.getValueType();
793 if (Node->getOpcode() == ISD::CopyFromReg) {
794 unsigned Reg = cast<RegisterSDNode>(Node->getOperand(1))->getReg();
795 // Just use the specified register as our input.
796 if (MRegisterInfo::isVirtualRegister(Reg) || Reg == PPC::R1)
800 unsigned &Reg = ExprMap[N];
803 switch (N.getOpcode()) {
805 Reg = Result = (N.getValueType() != MVT::Other) ?
806 MakeReg(N.getValueType()) : 1;
810 // If this is a call instruction, make sure to prepare ALL of the result
811 // values as well as the chain.
812 if (Node->getNumValues() == 1)
813 Reg = Result = 1; // Void call, just a chain.
815 Result = MakeReg(Node->getValueType(0));
816 ExprMap[N.getValue(0)] = Result;
817 for (unsigned i = 1, e = N.Val->getNumValues()-1; i != e; ++i)
818 ExprMap[N.getValue(i)] = MakeReg(Node->getValueType(i));
819 ExprMap[SDOperand(Node, Node->getNumValues()-1)] = 1;
827 Result = MakeReg(Node->getValueType(0));
828 ExprMap[N.getValue(0)] = Result;
829 for (unsigned i = 1, e = N.Val->getNumValues(); i != e; ++i)
830 ExprMap[N.getValue(i)] = MakeReg(Node->getValueType(i));
836 Node->dump(); std::cerr << '\n';
837 assert(0 && "Node not handled!\n");
839 BuildMI(BB, PPC::IMPLICIT_DEF, 0, Result);
841 case ISD::DYNAMIC_STACKALLOC:
842 // Generate both result values. FIXME: Need a better commment here?
844 ExprMap[N.getValue(1)] = 1;
846 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
848 // FIXME: We are currently ignoring the requested alignment for handling
849 // greater than the stack alignment. This will need to be revisited at some
850 // point. Align = N.getOperand(2);
851 if (!isa<ConstantSDNode>(N.getOperand(2)) ||
852 cast<ConstantSDNode>(N.getOperand(2))->getValue() != 0) {
853 std::cerr << "Cannot allocate stack object with greater alignment than"
854 << " the stack alignment yet!";
857 Select(N.getOperand(0));
858 Tmp1 = SelectExpr(N.getOperand(1));
859 // Subtract size from stack pointer, thereby allocating some space.
860 BuildMI(BB, PPC::SUBF, 2, PPC::R1).addReg(Tmp1).addReg(PPC::R1);
861 // Put a pointer to the space into the result register by copying the SP
862 BuildMI(BB, PPC::OR, 2, Result).addReg(PPC::R1).addReg(PPC::R1);
865 case ISD::ConstantPool:
866 Tmp1 = cast<ConstantPoolSDNode>(N)->getIndex();
869 BuildMI(BB, PPC::ADDIS, 2, Tmp2).addReg(getGlobalBaseReg())
870 .addConstantPoolIndex(Tmp1);
872 BuildMI(BB, PPC::LIS, 1, Tmp2).addConstantPoolIndex(Tmp1);
873 BuildMI(BB, PPC::LA, 2, Result).addReg(Tmp2).addConstantPoolIndex(Tmp1);
876 case ISD::FrameIndex:
877 Tmp1 = cast<FrameIndexSDNode>(N)->getIndex();
878 addFrameReference(BuildMI(BB, PPC::ADDI, 2, Result), (int)Tmp1, 0, false);
881 case ISD::GlobalAddress: {
882 GlobalValue *GV = cast<GlobalAddressSDNode>(N)->getGlobal();
885 BuildMI(BB, PPC::ADDIS, 2, Tmp1).addReg(getGlobalBaseReg())
886 .addGlobalAddress(GV);
888 BuildMI(BB, PPC::LIS, 1, Tmp1).addGlobalAddress(GV);
889 if (GV->hasWeakLinkage() || GV->isExternal()) {
890 BuildMI(BB, PPC::LWZ, 2, Result).addGlobalAddress(GV).addReg(Tmp1);
892 BuildMI(BB, PPC::LA, 2, Result).addReg(Tmp1).addGlobalAddress(GV);
900 case ISD::SEXTLOAD: {
901 MVT::ValueType TypeBeingLoaded = (ISD::LOAD == opcode) ?
902 Node->getValueType(0) : cast<VTSDNode>(Node->getOperand(3))->getVT();
903 bool sext = (ISD::SEXTLOAD == opcode);
905 // Make sure we generate both values.
907 ExprMap[N.getValue(1)] = 1; // Generate the token
909 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
911 SDOperand Chain = N.getOperand(0);
912 SDOperand Address = N.getOperand(1);
915 switch (TypeBeingLoaded) {
916 default: Node->dump(); assert(0 && "Cannot load this type!");
917 case MVT::i1: Opc = PPC::LBZ; break;
918 case MVT::i8: Opc = PPC::LBZ; break;
919 case MVT::i16: Opc = sext ? PPC::LHA : PPC::LHZ; break;
920 case MVT::i32: Opc = PPC::LWZ; break;
921 case MVT::f32: Opc = PPC::LFS; break;
922 case MVT::f64: Opc = PPC::LFD; break;
925 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Address)) {
927 int CPI = CP->getIndex();
929 BuildMI(BB, PPC::ADDIS, 2, Tmp1).addReg(getGlobalBaseReg())
930 .addConstantPoolIndex(CPI);
932 BuildMI(BB, PPC::LIS, 1, Tmp1).addConstantPoolIndex(CPI);
933 BuildMI(BB, Opc, 2, Result).addConstantPoolIndex(CPI).addReg(Tmp1);
934 } else if (Address.getOpcode() == ISD::FrameIndex) {
935 Tmp1 = cast<FrameIndexSDNode>(Address)->getIndex();
936 addFrameReference(BuildMI(BB, Opc, 2, Result), (int)Tmp1);
939 switch(SelectAddr(Address, Tmp1, offset)) {
940 default: assert(0 && "Unhandled return value from SelectAddr");
941 case 0: // imm offset, no frame, no index
942 BuildMI(BB, Opc, 2, Result).addSImm(offset).addReg(Tmp1);
944 case 1: // imm offset + frame index
945 addFrameReference(BuildMI(BB, Opc, 2, Result), (int)Tmp1, offset);
947 case 2: // base+index addressing
948 Opc = IndexedOpForOp(Opc);
949 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(offset);
952 GlobalAddressSDNode *GN = cast<GlobalAddressSDNode>(Address);
953 GlobalValue *GV = GN->getGlobal();
954 BuildMI(BB, Opc, 2, Result).addGlobalAddress(GV).addReg(Tmp1);
963 unsigned GPR_idx = 0, FPR_idx = 0;
964 static const unsigned GPR[] = {
965 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
966 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
968 static const unsigned FPR[] = {
969 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
970 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
973 // Lower the chain for this call.
974 Select(N.getOperand(0));
975 ExprMap[N.getValue(Node->getNumValues()-1)] = 1;
977 MachineInstr *CallMI;
978 // Emit the correct call instruction based on the type of symbol called.
979 if (GlobalAddressSDNode *GASD =
980 dyn_cast<GlobalAddressSDNode>(N.getOperand(1))) {
981 CallMI = BuildMI(PPC::CALLpcrel, 1).addGlobalAddress(GASD->getGlobal(),
983 } else if (ExternalSymbolSDNode *ESSDN =
984 dyn_cast<ExternalSymbolSDNode>(N.getOperand(1))) {
985 CallMI = BuildMI(PPC::CALLpcrel, 1).addExternalSymbol(ESSDN->getSymbol(),
988 Tmp1 = SelectExpr(N.getOperand(1));
989 BuildMI(BB, PPC::OR, 2, PPC::R12).addReg(Tmp1).addReg(Tmp1);
990 BuildMI(BB, PPC::MTCTR, 1).addReg(PPC::R12);
991 CallMI = BuildMI(PPC::CALLindirect, 3).addImm(20).addImm(0)
995 // Load the register args to virtual regs
996 std::vector<unsigned> ArgVR;
997 for(int i = 2, e = Node->getNumOperands(); i < e; ++i)
998 ArgVR.push_back(SelectExpr(N.getOperand(i)));
1000 // Copy the virtual registers into the appropriate argument register
1001 for(int i = 0, e = ArgVR.size(); i < e; ++i) {
1002 switch(N.getOperand(i+2).getValueType()) {
1003 default: Node->dump(); assert(0 && "Unknown value type for call");
1008 assert(GPR_idx < 8 && "Too many int args");
1009 if (N.getOperand(i+2).getOpcode() != ISD::UNDEF) {
1010 BuildMI(BB, PPC::OR,2,GPR[GPR_idx]).addReg(ArgVR[i]).addReg(ArgVR[i]);
1011 CallMI->addRegOperand(GPR[GPR_idx], MachineOperand::Use);
1017 assert(FPR_idx < 13 && "Too many fp args");
1018 BuildMI(BB, PPC::FMR, 1, FPR[FPR_idx]).addReg(ArgVR[i]);
1019 CallMI->addRegOperand(FPR[FPR_idx], MachineOperand::Use);
1025 // Put the call instruction in the correct place in the MachineBasicBlock
1026 BB->push_back(CallMI);
1028 switch (Node->getValueType(0)) {
1029 default: assert(0 && "Unknown value type for call result!");
1030 case MVT::Other: return 1;
1035 if (Node->getValueType(1) == MVT::i32) {
1036 BuildMI(BB, PPC::OR, 2, Result+1).addReg(PPC::R3).addReg(PPC::R3);
1037 BuildMI(BB, PPC::OR, 2, Result).addReg(PPC::R4).addReg(PPC::R4);
1039 BuildMI(BB, PPC::OR, 2, Result).addReg(PPC::R3).addReg(PPC::R3);
1044 BuildMI(BB, PPC::FMR, 1, Result).addReg(PPC::F1);
1047 return Result+N.ResNo;
1050 case ISD::SIGN_EXTEND:
1051 case ISD::SIGN_EXTEND_INREG:
1052 Tmp1 = SelectExpr(N.getOperand(0));
1053 switch(cast<VTSDNode>(Node->getOperand(1))->getVT()) {
1054 default: Node->dump(); assert(0 && "Unhandled SIGN_EXTEND type"); break;
1056 BuildMI(BB, PPC::EXTSH, 1, Result).addReg(Tmp1);
1059 BuildMI(BB, PPC::EXTSB, 1, Result).addReg(Tmp1);
1062 BuildMI(BB, PPC::SUBFIC, 2, Result).addReg(Tmp1).addSImm(0);
1067 case ISD::CopyFromReg:
1068 DestType = N.getValue(0).getValueType();
1070 Result = ExprMap[N.getValue(0)] = MakeReg(DestType);
1071 Tmp1 = dyn_cast<RegisterSDNode>(Node->getOperand(1))->getReg();
1072 if (MVT::isInteger(DestType))
1073 BuildMI(BB, PPC::OR, 2, Result).addReg(Tmp1).addReg(Tmp1);
1075 BuildMI(BB, PPC::FMR, 1, Result).addReg(Tmp1);
1079 if (isIntImmediate(N.getOperand(1), Tmp2)) {
1080 unsigned SH, MB, ME;
1081 if (isOpcWithIntImmediate(N.getOperand(0), ISD::AND, Tmp3) &&
1082 isRotateAndMask(ISD::SHL, Tmp2, Tmp3, true, SH, MB, ME)) {
1083 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1084 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp1).addImm(SH)
1085 .addImm(MB).addImm(ME);
1088 Tmp1 = SelectExpr(N.getOperand(0));
1090 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp1).addImm(Tmp2).addImm(0)
1093 Tmp1 = SelectExpr(N.getOperand(0));
1094 Tmp2 = FoldIfWideZeroExtend(N.getOperand(1));
1095 BuildMI(BB, PPC::SLW, 2, Result).addReg(Tmp1).addReg(Tmp2);
1100 if (isIntImmediate(N.getOperand(1), Tmp2)) {
1101 unsigned SH, MB, ME;
1102 if (isOpcWithIntImmediate(N.getOperand(0), ISD::AND, Tmp3) &&
1103 isRotateAndMask(ISD::SRL, Tmp2, Tmp3, true, SH, MB, ME)) {
1104 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1105 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp1).addImm(SH)
1106 .addImm(MB).addImm(ME);
1109 Tmp1 = SelectExpr(N.getOperand(0));
1111 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp1).addImm(32-Tmp2)
1112 .addImm(Tmp2).addImm(31);
1114 Tmp1 = SelectExpr(N.getOperand(0));
1115 Tmp2 = FoldIfWideZeroExtend(N.getOperand(1));
1116 BuildMI(BB, PPC::SRW, 2, Result).addReg(Tmp1).addReg(Tmp2);
1121 if (isIntImmediate(N.getOperand(1), Tmp2)) {
1122 unsigned SH, MB, ME;
1123 if (isOpcWithIntImmediate(N.getOperand(0), ISD::AND, Tmp3) &&
1124 isRotateAndMask(ISD::SRA, Tmp2, Tmp3, true, SH, MB, ME)) {
1125 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1126 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp1).addImm(SH)
1127 .addImm(MB).addImm(ME);
1130 Tmp1 = SelectExpr(N.getOperand(0));
1132 BuildMI(BB, PPC::SRAWI, 2, Result).addReg(Tmp1).addImm(Tmp2);
1134 Tmp1 = SelectExpr(N.getOperand(0));
1135 Tmp2 = FoldIfWideZeroExtend(N.getOperand(1));
1136 BuildMI(BB, PPC::SRAW, 2, Result).addReg(Tmp1).addReg(Tmp2);
1141 Tmp1 = SelectExpr(N.getOperand(0));
1142 BuildMI(BB, PPC::CNTLZW, 1, Result).addReg(Tmp1);
1146 if (!MVT::isInteger(DestType)) {
1147 if (!NoExcessFPPrecision && N.getOperand(0).getOpcode() == ISD::MUL &&
1148 N.getOperand(0).Val->hasOneUse()) {
1149 ++FusedFP; // Statistic
1150 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1151 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1152 Tmp3 = SelectExpr(N.getOperand(1));
1153 Opc = DestType == MVT::f64 ? PPC::FMADD : PPC::FMADDS;
1154 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
1157 if (!NoExcessFPPrecision && N.getOperand(1).getOpcode() == ISD::MUL &&
1158 N.getOperand(1).Val->hasOneUse()) {
1159 ++FusedFP; // Statistic
1160 Tmp1 = SelectExpr(N.getOperand(1).getOperand(0));
1161 Tmp2 = SelectExpr(N.getOperand(1).getOperand(1));
1162 Tmp3 = SelectExpr(N.getOperand(0));
1163 Opc = DestType == MVT::f64 ? PPC::FMADD : PPC::FMADDS;
1164 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
1167 Opc = DestType == MVT::f64 ? PPC::FADD : PPC::FADDS;
1168 Tmp1 = SelectExpr(N.getOperand(0));
1169 Tmp2 = SelectExpr(N.getOperand(1));
1170 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1173 if (SelectIntImmediateExpr(N, Result, PPC::ADDIS, PPC::ADDI, true))
1175 Tmp1 = SelectExpr(N.getOperand(0));
1176 Tmp2 = SelectExpr(N.getOperand(1));
1177 BuildMI(BB, PPC::ADD, 2, Result).addReg(Tmp1).addReg(Tmp2);
1181 if (isIntImmediate(N.getOperand(1), Tmp2)) {
1182 if (isShiftedMask_32(Tmp2) || isShiftedMask_32(~Tmp2)) {
1183 unsigned SH, MB, ME;
1184 Opc = Recording ? PPC::RLWINMo : PPC::RLWINM;
1186 if (isOprShiftImm(N.getOperand(0), OprOpc, Tmp3) &&
1187 isRotateAndMask(OprOpc, Tmp3, Tmp2, false, SH, MB, ME)) {
1188 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1190 Tmp1 = SelectExpr(N.getOperand(0));
1191 isRunOfOnes(Tmp2, MB, ME);
1194 BuildMI(BB, Opc, 4, Result).addReg(Tmp1).addImm(SH)
1195 .addImm(MB).addImm(ME);
1196 RecordSuccess = true;
1198 } else if (isUInt16(Tmp2)) {
1200 Tmp1 = SelectExpr(N.getOperand(0));
1201 BuildMI(BB, PPC::ANDIo, 2, Result).addReg(Tmp1).addImm(Tmp2);
1202 RecordSuccess = true;
1204 } else if (isUInt16(Tmp2)) {
1206 Tmp1 = SelectExpr(N.getOperand(0));
1207 BuildMI(BB, PPC::ANDISo, 2, Result).addReg(Tmp1).addImm(Tmp2);
1208 RecordSuccess = true;
1212 if (isOprNot(N.getOperand(1))) {
1213 Tmp1 = SelectExpr(N.getOperand(0));
1214 Tmp2 = SelectExpr(N.getOperand(1).getOperand(0));
1215 BuildMI(BB, PPC::ANDC, 2, Result).addReg(Tmp1).addReg(Tmp2);
1216 RecordSuccess = false;
1219 if (isOprNot(N.getOperand(0))) {
1220 Tmp1 = SelectExpr(N.getOperand(1));
1221 Tmp2 = SelectExpr(N.getOperand(0).getOperand(0));
1222 BuildMI(BB, PPC::ANDC, 2, Result).addReg(Tmp1).addReg(Tmp2);
1223 RecordSuccess = false;
1226 // emit a regular and
1227 Tmp1 = SelectExpr(N.getOperand(0));
1228 Tmp2 = SelectExpr(N.getOperand(1));
1229 Opc = Recording ? PPC::ANDo : PPC::AND;
1230 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1231 RecordSuccess = true;
1235 if (SelectBitfieldInsert(N, Result))
1237 if (SelectIntImmediateExpr(N, Result, PPC::ORIS, PPC::ORI))
1239 if (isOprNot(N.getOperand(1))) {
1240 Tmp1 = SelectExpr(N.getOperand(0));
1241 Tmp2 = SelectExpr(N.getOperand(1).getOperand(0));
1242 BuildMI(BB, PPC::ORC, 2, Result).addReg(Tmp1).addReg(Tmp2);
1243 RecordSuccess = false;
1246 if (isOprNot(N.getOperand(0))) {
1247 Tmp1 = SelectExpr(N.getOperand(1));
1248 Tmp2 = SelectExpr(N.getOperand(0).getOperand(0));
1249 BuildMI(BB, PPC::ORC, 2, Result).addReg(Tmp1).addReg(Tmp2);
1250 RecordSuccess = false;
1254 Tmp1 = SelectExpr(N.getOperand(0));
1255 Tmp2 = SelectExpr(N.getOperand(1));
1256 Opc = Recording ? PPC::ORo : PPC::OR;
1257 RecordSuccess = true;
1258 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1262 // Check for EQV: xor, (xor a, -1), b
1263 if (isOprNot(N.getOperand(0))) {
1264 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1265 Tmp2 = SelectExpr(N.getOperand(1));
1266 BuildMI(BB, PPC::EQV, 2, Result).addReg(Tmp1).addReg(Tmp2);
1269 // Check for NOT, NOR, EQV, and NAND: xor (copy, or, xor, and), -1
1271 switch(N.getOperand(0).getOpcode()) {
1273 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1274 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1275 BuildMI(BB, PPC::NOR, 2, Result).addReg(Tmp1).addReg(Tmp2);
1278 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1279 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1280 BuildMI(BB, PPC::NAND, 2, Result).addReg(Tmp1).addReg(Tmp2);
1283 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1284 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1285 BuildMI(BB, PPC::EQV, 2, Result).addReg(Tmp1).addReg(Tmp2);
1288 Tmp1 = SelectExpr(N.getOperand(0));
1289 BuildMI(BB, PPC::NOR, 2, Result).addReg(Tmp1).addReg(Tmp1);
1294 if (SelectIntImmediateExpr(N, Result, PPC::XORIS, PPC::XORI))
1297 Tmp1 = SelectExpr(N.getOperand(0));
1298 Tmp2 = SelectExpr(N.getOperand(1));
1299 BuildMI(BB, PPC::XOR, 2, Result).addReg(Tmp1).addReg(Tmp2);
1304 if (!MVT::isInteger(DestType)) {
1305 if (!NoExcessFPPrecision && N.getOperand(0).getOpcode() == ISD::MUL &&
1306 N.getOperand(0).Val->hasOneUse()) {
1307 ++FusedFP; // Statistic
1308 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1309 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1310 Tmp3 = SelectExpr(N.getOperand(1));
1311 Opc = DestType == MVT::f64 ? PPC::FMSUB : PPC::FMSUBS;
1312 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
1315 if (!NoExcessFPPrecision && N.getOperand(1).getOpcode() == ISD::MUL &&
1316 N.getOperand(1).Val->hasOneUse()) {
1317 ++FusedFP; // Statistic
1318 Tmp1 = SelectExpr(N.getOperand(1).getOperand(0));
1319 Tmp2 = SelectExpr(N.getOperand(1).getOperand(1));
1320 Tmp3 = SelectExpr(N.getOperand(0));
1321 Opc = DestType == MVT::f64 ? PPC::FNMSUB : PPC::FNMSUBS;
1322 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
1325 Opc = DestType == MVT::f64 ? PPC::FSUB : PPC::FSUBS;
1326 Tmp1 = SelectExpr(N.getOperand(0));
1327 Tmp2 = SelectExpr(N.getOperand(1));
1328 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1331 if (isIntImmediate(N.getOperand(0), Tmp1) && isInt16(Tmp1)) {
1333 Tmp2 = SelectExpr(N.getOperand(1));
1334 BuildMI(BB, PPC::SUBFIC, 2, Result).addReg(Tmp2).addSImm(Tmp1);
1337 if (SelectIntImmediateExpr(N, Result, PPC::ADDIS, PPC::ADDI, true, true))
1339 Tmp1 = SelectExpr(N.getOperand(0));
1340 Tmp2 = SelectExpr(N.getOperand(1));
1341 BuildMI(BB, PPC::SUBF, 2, Result).addReg(Tmp2).addReg(Tmp1);
1345 Tmp1 = SelectExpr(N.getOperand(0));
1346 if (isIntImmediate(N.getOperand(1), Tmp2) && isInt16(Tmp2)) {
1348 BuildMI(BB, PPC::MULLI, 2, Result).addReg(Tmp1).addSImm(Tmp2);
1350 Tmp2 = SelectExpr(N.getOperand(1));
1352 default: assert(0 && "Unknown type to ISD::MUL"); break;
1353 case MVT::i32: Opc = PPC::MULLW; break;
1354 case MVT::f32: Opc = PPC::FMULS; break;
1355 case MVT::f64: Opc = PPC::FMUL; break;
1357 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1363 Tmp1 = SelectExpr(N.getOperand(0));
1364 Tmp2 = SelectExpr(N.getOperand(1));
1365 Opc = (ISD::MULHU == opcode) ? PPC::MULHWU : PPC::MULHW;
1366 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1370 if (isIntImmediate(N.getOperand(1), Tmp3)) {
1371 if ((signed)Tmp3 > 0 && isPowerOf2_32(Tmp3)) {
1372 Tmp3 = Log2_32(Tmp3);
1373 Tmp1 = MakeIntReg();
1374 Tmp2 = SelectExpr(N.getOperand(0));
1375 BuildMI(BB, PPC::SRAWI, 2, Tmp1).addReg(Tmp2).addImm(Tmp3);
1376 BuildMI(BB, PPC::ADDZE, 1, Result).addReg(Tmp1);
1378 } else if ((signed)Tmp3 < 0 && isPowerOf2_32(-Tmp3)) {
1379 Tmp3 = Log2_32(-Tmp3);
1380 Tmp2 = SelectExpr(N.getOperand(0));
1381 Tmp1 = MakeIntReg();
1382 unsigned Tmp4 = MakeIntReg();
1383 BuildMI(BB, PPC::SRAWI, 2, Tmp1).addReg(Tmp2).addImm(Tmp3);
1384 BuildMI(BB, PPC::ADDZE, 1, Tmp4).addReg(Tmp1);
1385 BuildMI(BB, PPC::NEG, 1, Result).addReg(Tmp4);
1391 // If this is a divide by constant, we can emit code using some magic
1392 // constants to implement it as a multiply instead.
1393 if (isIntImmediate(N.getOperand(1), Tmp3)) {
1394 if (opcode == ISD::SDIV) {
1395 if ((signed)Tmp3 < -1 || (signed)Tmp3 > 1) {
1397 return SelectExpr(BuildSDIVSequence(N));
1400 if ((signed)Tmp3 > 1) {
1402 return SelectExpr(BuildUDIVSequence(N));
1406 Tmp1 = SelectExpr(N.getOperand(0));
1407 Tmp2 = SelectExpr(N.getOperand(1));
1409 default: assert(0 && "Unknown type to ISD::SDIV"); break;
1410 case MVT::i32: Opc = (ISD::UDIV == opcode) ? PPC::DIVWU : PPC::DIVW; break;
1411 case MVT::f32: Opc = PPC::FDIVS; break;
1412 case MVT::f64: Opc = PPC::FDIV; break;
1414 BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
1417 case ISD::ADD_PARTS:
1418 case ISD::SUB_PARTS: {
1419 assert(N.getNumOperands() == 4 && N.getValueType() == MVT::i32 &&
1420 "Not an i64 add/sub!");
1422 bool ME = isIntImmediate(N.getOperand(3),Tmp3) && ((signed)Tmp3 == -1);
1423 bool ZE = isIntImmediate(N.getOperand(3),Tmp3) && (Tmp3 == 0);
1424 bool IM = isIntImmediate(N.getOperand(2),Tmp3) && ((signed)Tmp3 >= -32768 ||
1425 (signed)Tmp3 < 32768);
1426 Tmp1 = SelectExpr(N.getOperand(0));
1427 Tmp2 = SelectExpr(N.getOperand(1));
1428 if (!IM || N.getOpcode() == ISD::SUB_PARTS)
1429 Tmp3 = SelectExpr(N.getOperand(2));
1430 if ((!ME && !ZE) || N.getOpcode() == ISD::SUB_PARTS)
1431 Tmp4 = SelectExpr(N.getOperand(3));
1433 if (N.getOpcode() == ISD::ADD_PARTS) {
1434 // Codegen the low 32 bits of the add. Interestingly, there is no shifted
1435 // form of add immediate carrying.
1437 BuildMI(BB, PPC::ADDIC, 2, Result).addReg(Tmp1).addSImm(Tmp3);
1439 BuildMI(BB, PPC::ADDC, 2, Result).addReg(Tmp1).addReg(Tmp3);
1440 // Codegen the high 32 bits, adding zero, minus one, or the full value
1441 // along with the carry flag produced by addc/addic to tmp2.
1443 BuildMI(BB, PPC::ADDZE, 1, Result+1).addReg(Tmp2);
1445 BuildMI(BB, PPC::ADDME, 1, Result+1).addReg(Tmp2);
1447 BuildMI(BB, PPC::ADDE, 2, Result+1).addReg(Tmp2).addReg(Tmp4);
1449 BuildMI(BB, PPC::SUBFC, 2, Result).addReg(Tmp3).addReg(Tmp1);
1450 BuildMI(BB, PPC::SUBFE, 2, Result+1).addReg(Tmp4).addReg(Tmp2);
1452 return Result+N.ResNo;
1455 case ISD::SHL_PARTS:
1456 case ISD::SRA_PARTS:
1457 case ISD::SRL_PARTS: {
1458 assert(N.getNumOperands() == 3 && N.getValueType() == MVT::i32 &&
1459 "Not an i64 shift!");
1460 unsigned ShiftOpLo = SelectExpr(N.getOperand(0));
1461 unsigned ShiftOpHi = SelectExpr(N.getOperand(1));
1462 unsigned SHReg = FoldIfWideZeroExtend(N.getOperand(2));
1463 Tmp1 = MakeIntReg();
1464 Tmp2 = MakeIntReg();
1465 Tmp3 = MakeIntReg();
1466 unsigned Tmp4 = MakeIntReg();
1467 unsigned Tmp5 = MakeIntReg();
1468 unsigned Tmp6 = MakeIntReg();
1469 BuildMI(BB, PPC::SUBFIC, 2, Tmp1).addReg(SHReg).addSImm(32);
1470 if (ISD::SHL_PARTS == opcode) {
1471 BuildMI(BB, PPC::SLW, 2, Tmp2).addReg(ShiftOpHi).addReg(SHReg);
1472 BuildMI(BB, PPC::SRW, 2, Tmp3).addReg(ShiftOpLo).addReg(Tmp1);
1473 BuildMI(BB, PPC::OR, 2, Tmp4).addReg(Tmp2).addReg(Tmp3);
1474 BuildMI(BB, PPC::ADDI, 2, Tmp5).addReg(SHReg).addSImm(-32);
1475 BuildMI(BB, PPC::SLW, 2, Tmp6).addReg(ShiftOpLo).addReg(Tmp5);
1476 BuildMI(BB, PPC::OR, 2, Result+1).addReg(Tmp4).addReg(Tmp6);
1477 BuildMI(BB, PPC::SLW, 2, Result).addReg(ShiftOpLo).addReg(SHReg);
1478 } else if (ISD::SRL_PARTS == opcode) {
1479 BuildMI(BB, PPC::SRW, 2, Tmp2).addReg(ShiftOpLo).addReg(SHReg);
1480 BuildMI(BB, PPC::SLW, 2, Tmp3).addReg(ShiftOpHi).addReg(Tmp1);
1481 BuildMI(BB, PPC::OR, 2, Tmp4).addReg(Tmp2).addReg(Tmp3);
1482 BuildMI(BB, PPC::ADDI, 2, Tmp5).addReg(SHReg).addSImm(-32);
1483 BuildMI(BB, PPC::SRW, 2, Tmp6).addReg(ShiftOpHi).addReg(Tmp5);
1484 BuildMI(BB, PPC::OR, 2, Result).addReg(Tmp4).addReg(Tmp6);
1485 BuildMI(BB, PPC::SRW, 2, Result+1).addReg(ShiftOpHi).addReg(SHReg);
1487 MachineBasicBlock *TmpMBB = new MachineBasicBlock(BB->getBasicBlock());
1488 MachineBasicBlock *PhiMBB = new MachineBasicBlock(BB->getBasicBlock());
1489 MachineBasicBlock *OldMBB = BB;
1490 MachineFunction *F = BB->getParent();
1491 ilist<MachineBasicBlock>::iterator It = BB; ++It;
1492 F->getBasicBlockList().insert(It, TmpMBB);
1493 F->getBasicBlockList().insert(It, PhiMBB);
1494 BB->addSuccessor(TmpMBB);
1495 BB->addSuccessor(PhiMBB);
1496 BuildMI(BB, PPC::SRW, 2, Tmp2).addReg(ShiftOpLo).addReg(SHReg);
1497 BuildMI(BB, PPC::SLW, 2, Tmp3).addReg(ShiftOpHi).addReg(Tmp1);
1498 BuildMI(BB, PPC::OR, 2, Tmp4).addReg(Tmp2).addReg(Tmp3);
1499 BuildMI(BB, PPC::ADDICo, 2, Tmp5).addReg(SHReg).addSImm(-32);
1500 BuildMI(BB, PPC::SRAW, 2, Tmp6).addReg(ShiftOpHi).addReg(Tmp5);
1501 BuildMI(BB, PPC::SRAW, 2, Result+1).addReg(ShiftOpHi).addReg(SHReg);
1502 BuildMI(BB, PPC::BLE, 2).addReg(PPC::CR0).addMBB(PhiMBB);
1503 // Select correct least significant half if the shift amount > 32
1505 unsigned Tmp7 = MakeIntReg();
1506 BuildMI(BB, PPC::OR, 2, Tmp7).addReg(Tmp6).addReg(Tmp6);
1507 TmpMBB->addSuccessor(PhiMBB);
1509 BuildMI(BB, PPC::PHI, 4, Result).addReg(Tmp4).addMBB(OldMBB)
1510 .addReg(Tmp7).addMBB(TmpMBB);
1512 return Result+N.ResNo;
1515 case ISD::FP_TO_SINT: {
1516 Tmp1 = SelectExpr(N.getOperand(0));
1518 BuildMI(BB, PPC::FCTIWZ, 1, Tmp2).addReg(Tmp1);
1519 int FrameIdx = BB->getParent()->getFrameInfo()->CreateStackObject(8, 8);
1520 addFrameReference(BuildMI(BB, PPC::STFD, 3).addReg(Tmp2), FrameIdx);
1521 addFrameReference(BuildMI(BB, PPC::LWZ, 2, Result), FrameIdx, 4);
1526 ISD::CondCode CC = cast<CondCodeSDNode>(Node->getOperand(2))->get();
1527 if (isIntImmediate(Node->getOperand(1), Tmp3)) {
1528 // We can codegen setcc op, imm very efficiently compared to a brcond.
1529 // Check for those cases here.
1532 Tmp1 = SelectExpr(Node->getOperand(0));
1534 default: Node->dump(); assert(0 && "Unhandled SetCC condition"); abort();
1536 Tmp2 = MakeIntReg();
1537 BuildMI(BB, PPC::CNTLZW, 1, Tmp2).addReg(Tmp1);
1538 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp2).addImm(27)
1539 .addImm(5).addImm(31);
1542 Tmp2 = MakeIntReg();
1543 BuildMI(BB, PPC::ADDIC, 2, Tmp2).addReg(Tmp1).addSImm(-1);
1544 BuildMI(BB, PPC::SUBFE, 2, Result).addReg(Tmp2).addReg(Tmp1);
1547 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp1).addImm(1)
1548 .addImm(31).addImm(31);
1551 Tmp2 = MakeIntReg();
1552 Tmp3 = MakeIntReg();
1553 BuildMI(BB, PPC::NEG, 2, Tmp2).addReg(Tmp1);
1554 BuildMI(BB, PPC::ANDC, 2, Tmp3).addReg(Tmp2).addReg(Tmp1);
1555 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp3).addImm(1)
1556 .addImm(31).addImm(31);
1560 } else if (Tmp3 == ~0U) { // setcc op, -1
1561 Tmp1 = SelectExpr(Node->getOperand(0));
1563 default: assert(0 && "Unhandled SetCC condition"); abort();
1565 Tmp2 = MakeIntReg();
1566 Tmp3 = MakeIntReg();
1567 BuildMI(BB, PPC::ADDIC, 2, Tmp2).addReg(Tmp1).addSImm(1);
1568 BuildMI(BB, PPC::LI, 1, Tmp3).addSImm(0);
1569 BuildMI(BB, PPC::ADDZE, 1, Result).addReg(Tmp3);
1572 Tmp2 = MakeIntReg();
1573 Tmp3 = MakeIntReg();
1574 BuildMI(BB, PPC::NOR, 2, Tmp2).addReg(Tmp1).addReg(Tmp1);
1575 BuildMI(BB, PPC::ADDIC, 2, Tmp3).addReg(Tmp2).addSImm(-1);
1576 BuildMI(BB, PPC::SUBFE, 2, Result).addReg(Tmp3).addReg(Tmp2);
1579 Tmp2 = MakeIntReg();
1580 Tmp3 = MakeIntReg();
1581 BuildMI(BB, PPC::ADDI, 2, Tmp2).addReg(Tmp1).addSImm(1);
1582 BuildMI(BB, PPC::AND, 2, Tmp3).addReg(Tmp2).addReg(Tmp1);
1583 BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp3).addImm(1)
1584 .addImm(31).addImm(31);
1587 Tmp2 = MakeIntReg();
1588 BuildMI(BB, PPC::RLWINM, 4, Tmp2).addReg(Tmp1).addImm(1)
1589 .addImm(31).addImm(31);
1590 BuildMI(BB, PPC::XORI, 2, Result).addReg(Tmp2).addImm(1);
1597 unsigned CCReg = SelectCC(N.getOperand(0), N.getOperand(1), CC);
1598 MoveCRtoGPR(CCReg, CC, Result);
1602 case ISD::SELECT_CC: {
1603 ISD::CondCode CC = cast<CondCodeSDNode>(N.getOperand(4))->get();
1604 if (!MVT::isInteger(N.getOperand(0).getValueType()) &&
1605 !MVT::isInteger(N.getOperand(2).getValueType()) &&
1606 CC != ISD::SETEQ && CC != ISD::SETNE) {
1607 MVT::ValueType VT = N.getOperand(0).getValueType();
1608 unsigned TV = SelectExpr(N.getOperand(2)); // Use if TRUE
1609 unsigned FV = SelectExpr(N.getOperand(3)); // Use if FALSE
1611 ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(N.getOperand(1));
1612 if (CN && (CN->isExactlyValue(-0.0) || CN->isExactlyValue(0.0))) {
1614 default: assert(0 && "Invalid FSEL condition"); abort();
1617 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
1620 Tmp1 = SelectExpr(N.getOperand(0)); // Val to compare against
1621 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp1).addReg(TV).addReg(FV);
1625 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
1628 if (N.getOperand(0).getOpcode() == ISD::FNEG) {
1629 Tmp2 = SelectExpr(N.getOperand(0).getOperand(0));
1632 Tmp1 = SelectExpr(N.getOperand(0)); // Val to compare against
1633 BuildMI(BB, PPC::FNEG, 1, Tmp2).addReg(Tmp1);
1635 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp2).addReg(TV).addReg(FV);
1640 Opc = (MVT::f64 == VT) ? PPC::FSUB : PPC::FSUBS;
1641 Tmp1 = SelectExpr(N.getOperand(0)); // Val to compare against
1642 Tmp2 = SelectExpr(N.getOperand(1));
1645 default: assert(0 && "Invalid FSEL condition"); abort();
1648 BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp1).addReg(Tmp2);
1649 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp3).addReg(FV).addReg(TV);
1653 BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp1).addReg(Tmp2);
1654 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp3).addReg(TV).addReg(FV);
1658 BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp2).addReg(Tmp1);
1659 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp3).addReg(FV).addReg(TV);
1663 BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp2).addReg(Tmp1);
1664 BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp3).addReg(TV).addReg(FV);
1668 assert(0 && "Should never get here");
1671 // If the False value only has one use, we can generate better code by
1672 // selecting it in the fallthrough basic block rather than here, which
1673 // increases register pressure.
1674 unsigned TrueValue = SelectExpr(N.getOperand(2));
1675 unsigned FalseValue = SelectExpr(N.getOperand(3));
1676 unsigned CCReg = SelectCC(N.getOperand(0), N.getOperand(1), CC);
1677 Opc = getBCCForSetCC(CC);
1679 // Create an iterator with which to insert the MBB for copying the false
1680 // value and the MBB to hold the PHI instruction for this SetCC.
1681 MachineBasicBlock *thisMBB = BB;
1682 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1683 ilist<MachineBasicBlock>::iterator It = BB;
1689 // cmpTY ccX, r1, r2
1691 // fallthrough --> copy0MBB
1692 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
1693 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
1694 BuildMI(BB, Opc, 2).addReg(CCReg).addMBB(sinkMBB);
1695 MachineFunction *F = BB->getParent();
1696 F->getBasicBlockList().insert(It, copy0MBB);
1697 F->getBasicBlockList().insert(It, sinkMBB);
1698 // Update machine-CFG edges
1699 BB->addSuccessor(copy0MBB);
1700 BB->addSuccessor(sinkMBB);
1703 // %FalseValue = ...
1704 // # fallthrough to sinkMBB
1706 // Update machine-CFG edges
1707 BB->addSuccessor(sinkMBB);
1710 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
1713 BuildMI(BB, PPC::PHI, 4, Result).addReg(FalseValue)
1714 .addMBB(copy0MBB).addReg(TrueValue).addMBB(thisMBB);
1718 case ISD::Constant: {
1719 assert(N.getValueType() == MVT::i32 &&
1720 "Only i32 constants are legal on this target!");
1721 unsigned v = (unsigned)cast<ConstantSDNode>(N)->getValue();
1722 unsigned Hi = HA16(v);
1723 unsigned Lo = Lo16(v);
1725 Tmp1 = MakeIntReg();
1726 BuildMI(BB, PPC::LIS, 1, Tmp1).addSImm(v >> 16);
1727 BuildMI(BB, PPC::ORI, 2, Result).addReg(Tmp1).addImm(Lo);
1729 BuildMI(BB, PPC::LIS, 1, Result).addSImm(v >> 16);
1731 BuildMI(BB, PPC::LI, 1, Result).addSImm(Lo);
1736 case ISD::ConstantFP: {
1737 ConstantFPSDNode *CN = cast<ConstantFPSDNode>(N);
1738 Result = getConstDouble(CN->getValue(), Result);
1743 if (!NoExcessFPPrecision &&
1744 ISD::ADD == N.getOperand(0).getOpcode() &&
1745 N.getOperand(0).Val->hasOneUse() &&
1746 ISD::MUL == N.getOperand(0).getOperand(0).getOpcode() &&
1747 N.getOperand(0).getOperand(0).Val->hasOneUse()) {
1748 ++FusedFP; // Statistic
1749 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0).getOperand(0));
1750 Tmp2 = SelectExpr(N.getOperand(0).getOperand(0).getOperand(1));
1751 Tmp3 = SelectExpr(N.getOperand(0).getOperand(1));
1752 Opc = DestType == MVT::f64 ? PPC::FNMADD : PPC::FNMADDS;
1753 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
1754 } else if (!NoExcessFPPrecision &&
1755 ISD::ADD == N.getOperand(0).getOpcode() &&
1756 N.getOperand(0).Val->hasOneUse() &&
1757 ISD::MUL == N.getOperand(0).getOperand(1).getOpcode() &&
1758 N.getOperand(0).getOperand(1).Val->hasOneUse()) {
1759 ++FusedFP; // Statistic
1760 Tmp1 = SelectExpr(N.getOperand(0).getOperand(1).getOperand(0));
1761 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1).getOperand(1));
1762 Tmp3 = SelectExpr(N.getOperand(0).getOperand(0));
1763 Opc = DestType == MVT::f64 ? PPC::FNMADD : PPC::FNMADDS;
1764 BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
1765 } else if (ISD::FABS == N.getOperand(0).getOpcode()) {
1766 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1767 BuildMI(BB, PPC::FNABS, 1, Result).addReg(Tmp1);
1769 Tmp1 = SelectExpr(N.getOperand(0));
1770 BuildMI(BB, PPC::FNEG, 1, Result).addReg(Tmp1);
1775 Tmp1 = SelectExpr(N.getOperand(0));
1776 BuildMI(BB, PPC::FABS, 1, Result).addReg(Tmp1);
1780 Tmp1 = SelectExpr(N.getOperand(0));
1781 Opc = DestType == MVT::f64 ? PPC::FSQRT : PPC::FSQRTS;
1782 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
1786 assert (DestType == MVT::f32 &&
1787 N.getOperand(0).getValueType() == MVT::f64 &&
1788 "only f64 to f32 conversion supported here");
1789 Tmp1 = SelectExpr(N.getOperand(0));
1790 BuildMI(BB, PPC::FRSP, 1, Result).addReg(Tmp1);
1793 case ISD::FP_EXTEND:
1794 assert (DestType == MVT::f64 &&
1795 N.getOperand(0).getValueType() == MVT::f32 &&
1796 "only f32 to f64 conversion supported here");
1797 Tmp1 = SelectExpr(N.getOperand(0));
1798 BuildMI(BB, PPC::FMR, 1, Result).addReg(Tmp1);
1804 void ISel::Select(SDOperand N) {
1805 unsigned Tmp1, Tmp2, Tmp3, Opc;
1806 unsigned opcode = N.getOpcode();
1808 if (!ExprMap.insert(std::make_pair(N, 1)).second)
1809 return; // Already selected.
1811 SDNode *Node = N.Val;
1813 switch (Node->getOpcode()) {
1815 Node->dump(); std::cerr << "\n";
1816 assert(0 && "Node not handled yet!");
1817 case ISD::EntryToken: return; // Noop
1818 case ISD::TokenFactor:
1819 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
1820 Select(Node->getOperand(i));
1822 case ISD::CALLSEQ_START:
1823 case ISD::CALLSEQ_END:
1824 Select(N.getOperand(0));
1825 Tmp1 = cast<ConstantSDNode>(N.getOperand(1))->getValue();
1826 Opc = N.getOpcode() == ISD::CALLSEQ_START ? PPC::ADJCALLSTACKDOWN :
1827 PPC::ADJCALLSTACKUP;
1828 BuildMI(BB, Opc, 1).addImm(Tmp1);
1831 MachineBasicBlock *Dest =
1832 cast<BasicBlockSDNode>(N.getOperand(1))->getBasicBlock();
1833 Select(N.getOperand(0));
1834 BuildMI(BB, PPC::B, 1).addMBB(Dest);
1838 case ISD::BRTWOWAY_CC:
1841 case ISD::CopyToReg:
1842 Select(N.getOperand(0));
1843 Tmp1 = SelectExpr(N.getOperand(2));
1844 Tmp2 = cast<RegisterSDNode>(N.getOperand(1))->getReg();
1847 if (N.getOperand(2).getValueType() == MVT::f64 ||
1848 N.getOperand(2).getValueType() == MVT::f32)
1849 BuildMI(BB, PPC::FMR, 1, Tmp2).addReg(Tmp1);
1851 BuildMI(BB, PPC::OR, 2, Tmp2).addReg(Tmp1).addReg(Tmp1);
1854 case ISD::ImplicitDef:
1855 Select(N.getOperand(0));
1856 BuildMI(BB, PPC::IMPLICIT_DEF, 0,
1857 cast<RegisterSDNode>(N.getOperand(1))->getReg());
1860 switch (N.getNumOperands()) {
1862 assert(0 && "Unknown return instruction!");
1864 assert(N.getOperand(1).getValueType() == MVT::i32 &&
1865 N.getOperand(2).getValueType() == MVT::i32 &&
1866 "Unknown two-register value!");
1867 Select(N.getOperand(0));
1868 Tmp1 = SelectExpr(N.getOperand(1));
1869 Tmp2 = SelectExpr(N.getOperand(2));
1870 BuildMI(BB, PPC::OR, 2, PPC::R3).addReg(Tmp2).addReg(Tmp2);
1871 BuildMI(BB, PPC::OR, 2, PPC::R4).addReg(Tmp1).addReg(Tmp1);
1874 Select(N.getOperand(0));
1875 Tmp1 = SelectExpr(N.getOperand(1));
1876 switch (N.getOperand(1).getValueType()) {
1878 assert(0 && "Unknown return type!");
1881 BuildMI(BB, PPC::FMR, 1, PPC::F1).addReg(Tmp1);
1884 BuildMI(BB, PPC::OR, 2, PPC::R3).addReg(Tmp1).addReg(Tmp1);
1888 Select(N.getOperand(0));
1891 BuildMI(BB, PPC::BLR, 0); // Just emit a 'ret' instruction
1893 case ISD::TRUNCSTORE:
1895 SDOperand Chain = N.getOperand(0);
1896 SDOperand Value = N.getOperand(1);
1897 SDOperand Address = N.getOperand(2);
1900 Tmp1 = SelectExpr(Value); //value
1902 if (opcode == ISD::STORE) {
1903 switch(Value.getValueType()) {
1904 default: assert(0 && "unknown Type in store");
1905 case MVT::i32: Opc = PPC::STW; break;
1906 case MVT::f64: Opc = PPC::STFD; break;
1907 case MVT::f32: Opc = PPC::STFS; break;
1909 } else { //ISD::TRUNCSTORE
1910 switch(cast<VTSDNode>(Node->getOperand(4))->getVT()) {
1911 default: assert(0 && "unknown Type in store");
1913 case MVT::i8: Opc = PPC::STB; break;
1914 case MVT::i16: Opc = PPC::STH; break;
1918 if(Address.getOpcode() == ISD::FrameIndex) {
1919 Tmp2 = cast<FrameIndexSDNode>(Address)->getIndex();
1920 addFrameReference(BuildMI(BB, Opc, 3).addReg(Tmp1), (int)Tmp2);
1923 switch(SelectAddr(Address, Tmp2, offset)) {
1924 default: assert(0 && "Unhandled return value from SelectAddr");
1925 case 0: // imm offset, no frame, no index
1926 BuildMI(BB, Opc, 3).addReg(Tmp1).addSImm(offset).addReg(Tmp2);
1928 case 1: // imm offset + frame index
1929 addFrameReference(BuildMI(BB, Opc, 3).addReg(Tmp1), (int)Tmp2, offset);
1931 case 2: // base+index addressing
1932 Opc = IndexedOpForOp(Opc);
1933 BuildMI(BB, Opc, 3).addReg(Tmp1).addReg(Tmp2).addReg(offset);
1936 GlobalAddressSDNode *GN = cast<GlobalAddressSDNode>(Address);
1937 GlobalValue *GV = GN->getGlobal();
1938 BuildMI(BB, Opc, 3).addReg(Tmp1).addGlobalAddress(GV).addReg(Tmp2);
1948 case ISD::CopyFromReg:
1951 case ISD::DYNAMIC_STACKALLOC:
1956 assert(0 && "Should not be reached!");
1960 /// createPPC32PatternInstructionSelector - This pass converts an LLVM function
1961 /// into a machine code representation using pattern matching and a machine
1962 /// description file.
1964 FunctionPass *llvm::createPPC32ISelPattern(TargetMachine &TM) {
1965 return new ISel(TM);