1 //===-- SPUISelDAGToDAG.cpp - CellSPU pattern matching inst selector ------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines a pattern matching instruction selector for the Cell SPU,
11 // converting from a legalized dag to a SPU-target dag.
13 //===----------------------------------------------------------------------===//
16 #include "SPUTargetMachine.h"
17 #include "SPUHazardRecognizers.h"
18 #include "SPUFrameInfo.h"
19 #include "SPURegisterNames.h"
20 #include "SPUTargetMachine.h"
21 #include "llvm/CodeGen/MachineConstantPool.h"
22 #include "llvm/CodeGen/MachineInstrBuilder.h"
23 #include "llvm/CodeGen/MachineFunction.h"
24 #include "llvm/CodeGen/SelectionDAG.h"
25 #include "llvm/CodeGen/SelectionDAGISel.h"
26 #include "llvm/CodeGen/PseudoSourceValue.h"
27 #include "llvm/Target/TargetOptions.h"
28 #include "llvm/ADT/Statistic.h"
29 #include "llvm/Constants.h"
30 #include "llvm/GlobalValue.h"
31 #include "llvm/Intrinsics.h"
32 #include "llvm/LLVMContext.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/ErrorHandling.h"
35 #include "llvm/Support/MathExtras.h"
36 #include "llvm/Support/Compiler.h"
37 #include "llvm/Support/raw_ostream.h"
42 //! ConstantSDNode predicate for i32 sign-extended, 10-bit immediates
44 isI32IntS10Immediate(ConstantSDNode *CN)
46 return isInt<10>(CN->getSExtValue());
49 //! ConstantSDNode predicate for i32 unsigned 10-bit immediate values
51 isI32IntU10Immediate(ConstantSDNode *CN)
53 return isUInt<10>(CN->getSExtValue());
56 //! ConstantSDNode predicate for i16 sign-extended, 10-bit immediate values
58 isI16IntS10Immediate(ConstantSDNode *CN)
60 return isInt<10>(CN->getSExtValue());
63 //! ConstantSDNode predicate for i16 unsigned 10-bit immediate values
65 isI16IntU10Immediate(ConstantSDNode *CN)
67 return isUInt<10>((short) CN->getZExtValue());
70 //! ConstantSDNode predicate for signed 16-bit values
72 \arg CN The constant SelectionDAG node holding the value
73 \arg Imm The returned 16-bit value, if returning true
75 This predicate tests the value in \a CN to see whether it can be
76 represented as a 16-bit, sign-extended quantity. Returns true if
80 isIntS16Immediate(ConstantSDNode *CN, short &Imm)
82 EVT vt = CN->getValueType(0);
83 Imm = (short) CN->getZExtValue();
84 if (vt.getSimpleVT() >= MVT::i1 && vt.getSimpleVT() <= MVT::i16) {
86 } else if (vt == MVT::i32) {
87 int32_t i_val = (int32_t) CN->getZExtValue();
88 short s_val = (short) i_val;
89 return i_val == s_val;
91 int64_t i_val = (int64_t) CN->getZExtValue();
92 short s_val = (short) i_val;
93 return i_val == s_val;
99 //! ConstantFPSDNode predicate for representing floats as 16-bit sign ext.
101 isFPS16Immediate(ConstantFPSDNode *FPN, short &Imm)
103 EVT vt = FPN->getValueType(0);
104 if (vt == MVT::f32) {
105 int val = FloatToBits(FPN->getValueAPF().convertToFloat());
106 int sval = (int) ((val << 16) >> 16);
114 //===------------------------------------------------------------------===//
115 //! EVT to "useful stuff" mapping structure:
117 struct valtype_map_s {
119 unsigned ldresult_ins; /// LDRESULT instruction (0 = undefined)
120 bool ldresult_imm; /// LDRESULT instruction requires immediate?
121 unsigned lrinst; /// LR instruction
124 const valtype_map_s valtype_map[] = {
125 { MVT::i8, SPU::ORBIr8, true, SPU::LRr8 },
126 { MVT::i16, SPU::ORHIr16, true, SPU::LRr16 },
127 { MVT::i32, SPU::ORIr32, true, SPU::LRr32 },
128 { MVT::i64, SPU::ORr64, false, SPU::LRr64 },
129 { MVT::f32, SPU::ORf32, false, SPU::LRf32 },
130 { MVT::f64, SPU::ORf64, false, SPU::LRf64 },
131 // vector types... (sigh!)
132 { MVT::v16i8, 0, false, SPU::LRv16i8 },
133 { MVT::v8i16, 0, false, SPU::LRv8i16 },
134 { MVT::v4i32, 0, false, SPU::LRv4i32 },
135 { MVT::v2i64, 0, false, SPU::LRv2i64 },
136 { MVT::v4f32, 0, false, SPU::LRv4f32 },
137 { MVT::v2f64, 0, false, SPU::LRv2f64 }
140 const size_t n_valtype_map = sizeof(valtype_map) / sizeof(valtype_map[0]);
142 const valtype_map_s *getValueTypeMapEntry(EVT VT)
144 const valtype_map_s *retval = 0;
145 for (size_t i = 0; i < n_valtype_map; ++i) {
146 if (valtype_map[i].VT == VT) {
147 retval = valtype_map + i;
155 report_fatal_error("SPUISelDAGToDAG.cpp: getValueTypeMapEntry returns"
156 "NULL for " + Twine(VT.getEVTString()));
163 //! Generate the carry-generate shuffle mask.
164 SDValue getCarryGenerateShufMask(SelectionDAG &DAG, DebugLoc dl) {
165 SmallVector<SDValue, 16 > ShufBytes;
167 // Create the shuffle mask for "rotating" the borrow up one register slot
168 // once the borrow is generated.
169 ShufBytes.push_back(DAG.getConstant(0x04050607, MVT::i32));
170 ShufBytes.push_back(DAG.getConstant(0x80808080, MVT::i32));
171 ShufBytes.push_back(DAG.getConstant(0x0c0d0e0f, MVT::i32));
172 ShufBytes.push_back(DAG.getConstant(0x80808080, MVT::i32));
174 return DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
175 &ShufBytes[0], ShufBytes.size());
178 //! Generate the borrow-generate shuffle mask
179 SDValue getBorrowGenerateShufMask(SelectionDAG &DAG, DebugLoc dl) {
180 SmallVector<SDValue, 16 > ShufBytes;
182 // Create the shuffle mask for "rotating" the borrow up one register slot
183 // once the borrow is generated.
184 ShufBytes.push_back(DAG.getConstant(0x04050607, MVT::i32));
185 ShufBytes.push_back(DAG.getConstant(0xc0c0c0c0, MVT::i32));
186 ShufBytes.push_back(DAG.getConstant(0x0c0d0e0f, MVT::i32));
187 ShufBytes.push_back(DAG.getConstant(0xc0c0c0c0, MVT::i32));
189 return DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
190 &ShufBytes[0], ShufBytes.size());
193 //===------------------------------------------------------------------===//
194 /// SPUDAGToDAGISel - Cell SPU-specific code to select SPU machine
195 /// instructions for SelectionDAG operations.
197 class SPUDAGToDAGISel :
198 public SelectionDAGISel
200 const SPUTargetMachine &TM;
201 const SPUTargetLowering &SPUtli;
202 unsigned GlobalBaseReg;
205 explicit SPUDAGToDAGISel(SPUTargetMachine &tm) :
206 SelectionDAGISel(tm),
208 SPUtli(*tm.getTargetLowering())
211 virtual bool runOnMachineFunction(MachineFunction &MF) {
212 // Make sure we re-emit a set of the global base reg if necessary
214 SelectionDAGISel::runOnMachineFunction(MF);
218 /// getI32Imm - Return a target constant with the specified value, of type
220 inline SDValue getI32Imm(uint32_t Imm) {
221 return CurDAG->getTargetConstant(Imm, MVT::i32);
224 /// getSmallIPtrImm - Return a target constant of pointer type.
225 inline SDValue getSmallIPtrImm(unsigned Imm) {
226 return CurDAG->getTargetConstant(Imm, SPUtli.getPointerTy());
229 SDNode *emitBuildVector(SDNode *bvNode) {
230 EVT vecVT = bvNode->getValueType(0);
231 DebugLoc dl = bvNode->getDebugLoc();
233 // Check to see if this vector can be represented as a CellSPU immediate
234 // constant by invoking all of the instruction selection predicates:
235 if (((vecVT == MVT::v8i16) &&
236 (SPU::get_vec_i16imm(bvNode, *CurDAG, MVT::i16).getNode() != 0)) ||
237 ((vecVT == MVT::v4i32) &&
238 ((SPU::get_vec_i16imm(bvNode, *CurDAG, MVT::i32).getNode() != 0) ||
239 (SPU::get_ILHUvec_imm(bvNode, *CurDAG, MVT::i32).getNode() != 0) ||
240 (SPU::get_vec_u18imm(bvNode, *CurDAG, MVT::i32).getNode() != 0) ||
241 (SPU::get_v4i32_imm(bvNode, *CurDAG).getNode() != 0))) ||
242 ((vecVT == MVT::v2i64) &&
243 ((SPU::get_vec_i16imm(bvNode, *CurDAG, MVT::i64).getNode() != 0) ||
244 (SPU::get_ILHUvec_imm(bvNode, *CurDAG, MVT::i64).getNode() != 0) ||
245 (SPU::get_vec_u18imm(bvNode, *CurDAG, MVT::i64).getNode() != 0)))) {
246 HandleSDNode Dummy(SDValue(bvNode, 0));
247 if (SDNode *N = Select(bvNode))
249 return Dummy.getValue().getNode();
252 // No, need to emit a constant pool spill:
253 std::vector<Constant*> CV;
255 for (size_t i = 0; i < bvNode->getNumOperands(); ++i) {
256 ConstantSDNode *V = cast<ConstantSDNode > (bvNode->getOperand(i));
257 CV.push_back(const_cast<ConstantInt *>(V->getConstantIntValue()));
260 const Constant *CP = ConstantVector::get(CV);
261 SDValue CPIdx = CurDAG->getConstantPool(CP, SPUtli.getPointerTy());
262 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
263 SDValue CGPoolOffset =
264 SPU::LowerConstantPool(CPIdx, *CurDAG, TM);
266 HandleSDNode Dummy(CurDAG->getLoad(vecVT, dl,
267 CurDAG->getEntryNode(), CGPoolOffset,
268 PseudoSourceValue::getConstantPool(),0,
269 false, false, Alignment));
270 CurDAG->ReplaceAllUsesWith(SDValue(bvNode, 0), Dummy.getValue());
271 if (SDNode *N = SelectCode(Dummy.getValue().getNode()))
273 return Dummy.getValue().getNode();
276 /// Select - Convert the specified operand from a target-independent to a
277 /// target-specific node if it hasn't already been changed.
278 SDNode *Select(SDNode *N);
280 //! Emit the instruction sequence for i64 shl
281 SDNode *SelectSHLi64(SDNode *N, EVT OpVT);
283 //! Emit the instruction sequence for i64 srl
284 SDNode *SelectSRLi64(SDNode *N, EVT OpVT);
286 //! Emit the instruction sequence for i64 sra
287 SDNode *SelectSRAi64(SDNode *N, EVT OpVT);
289 //! Emit the necessary sequence for loading i64 constants:
290 SDNode *SelectI64Constant(SDNode *N, EVT OpVT, DebugLoc dl);
292 //! Alternate instruction emit sequence for loading i64 constants
293 SDNode *SelectI64Constant(uint64_t i64const, EVT OpVT, DebugLoc dl);
295 //! Returns true if the address N is an A-form (local store) address
296 bool SelectAFormAddr(SDNode *Op, SDValue N, SDValue &Base,
299 //! D-form address predicate
300 bool SelectDFormAddr(SDNode *Op, SDValue N, SDValue &Base,
303 /// Alternate D-form address using i7 offset predicate
304 bool SelectDForm2Addr(SDNode *Op, SDValue N, SDValue &Disp,
307 /// D-form address selection workhorse
308 bool DFormAddressPredicate(SDNode *Op, SDValue N, SDValue &Disp,
309 SDValue &Base, int minOffset, int maxOffset);
311 //! Address predicate if N can be expressed as an indexed [r+r] operation.
312 bool SelectXFormAddr(SDNode *Op, SDValue N, SDValue &Base,
315 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
316 /// inline asm expressions.
317 virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
319 std::vector<SDValue> &OutOps) {
321 switch (ConstraintCode) {
322 default: return true;
324 if (!SelectDFormAddr(Op.getNode(), Op, Op0, Op1)
325 && !SelectAFormAddr(Op.getNode(), Op, Op0, Op1))
326 SelectXFormAddr(Op.getNode(), Op, Op0, Op1);
328 case 'o': // offsetable
329 if (!SelectDFormAddr(Op.getNode(), Op, Op0, Op1)
330 && !SelectAFormAddr(Op.getNode(), Op, Op0, Op1)) {
332 Op1 = getSmallIPtrImm(0);
335 case 'v': // not offsetable
337 llvm_unreachable("InlineAsmMemoryOperand 'v' constraint not handled.");
339 SelectAddrIdxOnly(Op, Op, Op0, Op1);
344 OutOps.push_back(Op0);
345 OutOps.push_back(Op1);
349 virtual const char *getPassName() const {
350 return "Cell SPU DAG->DAG Pattern Instruction Selection";
353 /// CreateTargetHazardRecognizer - Return the hazard recognizer to use for
354 /// this target when scheduling the DAG.
355 virtual ScheduleHazardRecognizer *CreateTargetHazardRecognizer() {
356 const TargetInstrInfo *II = TM.getInstrInfo();
357 assert(II && "No InstrInfo?");
358 return new SPUHazardRecognizer(*II);
362 SDValue getRC( MVT );
364 // Include the pieces autogenerated from the target description.
365 #include "SPUGenDAGISel.inc"
370 \arg Op The ISD instruction operand
371 \arg N The address to be tested
372 \arg Base The base address
373 \arg Index The base address index
376 SPUDAGToDAGISel::SelectAFormAddr(SDNode *Op, SDValue N, SDValue &Base,
378 // These match the addr256k operand type:
379 EVT OffsVT = MVT::i16;
380 SDValue Zero = CurDAG->getTargetConstant(0, OffsVT);
382 switch (N.getOpcode()) {
384 case ISD::ConstantPool:
385 case ISD::GlobalAddress:
386 report_fatal_error("SPU SelectAFormAddr: Constant/Pool/Global not lowered.");
389 case ISD::TargetConstant:
390 case ISD::TargetGlobalAddress:
391 case ISD::TargetJumpTable:
392 report_fatal_error("SPUSelectAFormAddr: Target Constant/Pool/Global "
393 "not wrapped as A-form address.");
396 case SPUISD::AFormAddr:
397 // Just load from memory if there's only a single use of the location,
398 // otherwise, this will get handled below with D-form offset addresses
400 SDValue Op0 = N.getOperand(0);
401 switch (Op0.getOpcode()) {
402 case ISD::TargetConstantPool:
403 case ISD::TargetJumpTable:
408 case ISD::TargetGlobalAddress: {
409 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op0);
410 const GlobalValue *GV = GSDN->getGlobal();
411 if (GV->getAlignment() == 16) {
426 SPUDAGToDAGISel::SelectDForm2Addr(SDNode *Op, SDValue N, SDValue &Disp,
428 const int minDForm2Offset = -(1 << 7);
429 const int maxDForm2Offset = (1 << 7) - 1;
430 return DFormAddressPredicate(Op, N, Disp, Base, minDForm2Offset,
435 \arg Op The ISD instruction (ignored)
436 \arg N The address to be tested
437 \arg Base Base address register/pointer
438 \arg Index Base address index
440 Examine the input address by a base register plus a signed 10-bit
441 displacement, [r+I10] (D-form address).
443 \return true if \a N is a D-form address with \a Base and \a Index set
444 to non-empty SDValue instances.
447 SPUDAGToDAGISel::SelectDFormAddr(SDNode *Op, SDValue N, SDValue &Base,
449 return DFormAddressPredicate(Op, N, Base, Index,
450 SPUFrameInfo::minFrameOffset(),
451 SPUFrameInfo::maxFrameOffset());
455 SPUDAGToDAGISel::DFormAddressPredicate(SDNode *Op, SDValue N, SDValue &Base,
456 SDValue &Index, int minOffset,
458 unsigned Opc = N.getOpcode();
459 EVT PtrTy = SPUtli.getPointerTy();
461 if (Opc == ISD::FrameIndex) {
462 // Stack frame index must be less than 512 (divided by 16):
463 FrameIndexSDNode *FIN = cast<FrameIndexSDNode>(N);
464 int FI = int(FIN->getIndex());
465 DEBUG(errs() << "SelectDFormAddr: ISD::FrameIndex = "
467 if (SPUFrameInfo::FItoStackOffset(FI) < maxOffset) {
468 Base = CurDAG->getTargetConstant(0, PtrTy);
469 Index = CurDAG->getTargetFrameIndex(FI, PtrTy);
472 } else if (Opc == ISD::ADD) {
473 // Generated by getelementptr
474 const SDValue Op0 = N.getOperand(0);
475 const SDValue Op1 = N.getOperand(1);
477 if ((Op0.getOpcode() == SPUISD::Hi && Op1.getOpcode() == SPUISD::Lo)
478 || (Op1.getOpcode() == SPUISD::Hi && Op0.getOpcode() == SPUISD::Lo)) {
479 Base = CurDAG->getTargetConstant(0, PtrTy);
482 } else if (Op1.getOpcode() == ISD::Constant
483 || Op1.getOpcode() == ISD::TargetConstant) {
484 ConstantSDNode *CN = cast<ConstantSDNode>(Op1);
485 int32_t offset = int32_t(CN->getSExtValue());
487 if (Op0.getOpcode() == ISD::FrameIndex) {
488 FrameIndexSDNode *FIN = cast<FrameIndexSDNode>(Op0);
489 int FI = int(FIN->getIndex());
490 DEBUG(errs() << "SelectDFormAddr: ISD::ADD offset = " << offset
491 << " frame index = " << FI << "\n");
493 if (SPUFrameInfo::FItoStackOffset(FI) < maxOffset) {
494 Base = CurDAG->getTargetConstant(offset, PtrTy);
495 Index = CurDAG->getTargetFrameIndex(FI, PtrTy);
498 } else if (offset > minOffset && offset < maxOffset) {
499 Base = CurDAG->getTargetConstant(offset, PtrTy);
503 } else if (Op0.getOpcode() == ISD::Constant
504 || Op0.getOpcode() == ISD::TargetConstant) {
505 ConstantSDNode *CN = cast<ConstantSDNode>(Op0);
506 int32_t offset = int32_t(CN->getSExtValue());
508 if (Op1.getOpcode() == ISD::FrameIndex) {
509 FrameIndexSDNode *FIN = cast<FrameIndexSDNode>(Op1);
510 int FI = int(FIN->getIndex());
511 DEBUG(errs() << "SelectDFormAddr: ISD::ADD offset = " << offset
512 << " frame index = " << FI << "\n");
514 if (SPUFrameInfo::FItoStackOffset(FI) < maxOffset) {
515 Base = CurDAG->getTargetConstant(offset, PtrTy);
516 Index = CurDAG->getTargetFrameIndex(FI, PtrTy);
519 } else if (offset > minOffset && offset < maxOffset) {
520 Base = CurDAG->getTargetConstant(offset, PtrTy);
525 } else if (Opc == SPUISD::IndirectAddr) {
526 // Indirect with constant offset -> D-Form address
527 const SDValue Op0 = N.getOperand(0);
528 const SDValue Op1 = N.getOperand(1);
530 if (Op0.getOpcode() == SPUISD::Hi
531 && Op1.getOpcode() == SPUISD::Lo) {
532 // (SPUindirect (SPUhi <arg>, 0), (SPUlo <arg>, 0))
533 Base = CurDAG->getTargetConstant(0, PtrTy);
536 } else if (isa<ConstantSDNode>(Op0) || isa<ConstantSDNode>(Op1)) {
540 if (isa<ConstantSDNode>(Op1)) {
541 ConstantSDNode *CN = cast<ConstantSDNode>(Op1);
542 offset = int32_t(CN->getSExtValue());
544 } else if (isa<ConstantSDNode>(Op0)) {
545 ConstantSDNode *CN = cast<ConstantSDNode>(Op0);
546 offset = int32_t(CN->getSExtValue());
550 if (offset >= minOffset && offset <= maxOffset) {
551 Base = CurDAG->getTargetConstant(offset, PtrTy);
556 } else if (Opc == SPUISD::AFormAddr) {
557 Base = CurDAG->getTargetConstant(0, N.getValueType());
560 } else if (Opc == SPUISD::LDRESULT) {
561 Base = CurDAG->getTargetConstant(0, N.getValueType());
564 } else if (Opc == ISD::Register
565 ||Opc == ISD::CopyFromReg
567 ||Opc == ISD::Constant) {
568 unsigned OpOpc = Op->getOpcode();
570 if (OpOpc == ISD::STORE || OpOpc == ISD::LOAD) {
571 // Direct load/store without getelementptr
574 Offs = ((OpOpc == ISD::STORE) ? Op->getOperand(3) : Op->getOperand(2));
576 if (Offs.getOpcode() == ISD::Constant || Offs.getOpcode() == ISD::UNDEF) {
577 if (Offs.getOpcode() == ISD::UNDEF)
578 Offs = CurDAG->getTargetConstant(0, Offs.getValueType());
585 /* If otherwise unadorned, default to D-form address with 0 offset: */
586 if (Opc == ISD::CopyFromReg) {
587 Index = N.getOperand(1);
592 Base = CurDAG->getTargetConstant(0, Index.getValueType());
601 \arg Op The ISD instruction operand
602 \arg N The address operand
603 \arg Base The base pointer operand
604 \arg Index The offset/index operand
606 If the address \a N can be expressed as an A-form or D-form address, returns
607 false. Otherwise, creates two operands, Base and Index that will become the
608 (r)(r) X-form address.
611 SPUDAGToDAGISel::SelectXFormAddr(SDNode *Op, SDValue N, SDValue &Base,
613 if (!SelectAFormAddr(Op, N, Base, Index)
614 && !SelectDFormAddr(Op, N, Base, Index)) {
615 // If the address is neither A-form or D-form, punt and use an X-form
617 Base = N.getOperand(1);
618 Index = N.getOperand(0);
626 Utility function to use with COPY_TO_REGCLASS instructions. Returns a SDValue
627 to be used as the last parameter of a
628 CurDAG->getMachineNode(COPY_TO_REGCLASS,..., ) function call
629 \arg VT the value type for which we want a register class
631 SDValue SPUDAGToDAGISel::getRC( MVT VT ) {
632 switch( VT.SimpleTy ) {
634 return CurDAG->getTargetConstant(SPU::R32CRegClass.getID(), MVT::i32);
637 return CurDAG->getTargetConstant(SPU::R64CRegClass.getID(), MVT::i32);
640 return CurDAG->getTargetConstant(SPU::VECREGRegClass.getID(), MVT::i32);
643 assert( false && "add a new case here" );
648 //! Convert the operand from a target-independent to a target-specific node
652 SPUDAGToDAGISel::Select(SDNode *N) {
653 unsigned Opc = N->getOpcode();
656 EVT OpVT = N->getValueType(0);
658 DebugLoc dl = N->getDebugLoc();
660 if (N->isMachineOpcode())
661 return NULL; // Already selected.
663 if (Opc == ISD::FrameIndex) {
664 int FI = cast<FrameIndexSDNode>(N)->getIndex();
665 SDValue TFI = CurDAG->getTargetFrameIndex(FI, N->getValueType(0));
666 SDValue Imm0 = CurDAG->getTargetConstant(0, N->getValueType(0));
675 Ops[0] = CurDAG->getRegister(SPU::R1, N->getValueType(0));
676 Ops[1] = SDValue(CurDAG->getMachineNode(SPU::ILAr32, dl,
677 N->getValueType(0), TFI, Imm0),
681 } else if (Opc == ISD::Constant && OpVT == MVT::i64) {
682 // Catch the i64 constants that end up here. Note: The backend doesn't
683 // attempt to legalize the constant (it's useless because DAGCombiner
684 // will insert 64-bit constants and we can't stop it).
685 return SelectI64Constant(N, OpVT, N->getDebugLoc());
686 } else if ((Opc == ISD::ZERO_EXTEND || Opc == ISD::ANY_EXTEND)
687 && OpVT == MVT::i64) {
688 SDValue Op0 = N->getOperand(0);
689 EVT Op0VT = Op0.getValueType();
690 EVT Op0VecVT = EVT::getVectorVT(*CurDAG->getContext(),
691 Op0VT, (128 / Op0VT.getSizeInBits()));
692 EVT OpVecVT = EVT::getVectorVT(*CurDAG->getContext(),
693 OpVT, (128 / OpVT.getSizeInBits()));
696 switch (Op0VT.getSimpleVT().SimpleTy) {
698 report_fatal_error("CellSPU Select: Unhandled zero/any extend EVT");
701 shufMask = CurDAG->getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
702 CurDAG->getConstant(0x80808080, MVT::i32),
703 CurDAG->getConstant(0x00010203, MVT::i32),
704 CurDAG->getConstant(0x80808080, MVT::i32),
705 CurDAG->getConstant(0x08090a0b, MVT::i32));
709 shufMask = CurDAG->getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
710 CurDAG->getConstant(0x80808080, MVT::i32),
711 CurDAG->getConstant(0x80800203, MVT::i32),
712 CurDAG->getConstant(0x80808080, MVT::i32),
713 CurDAG->getConstant(0x80800a0b, MVT::i32));
717 shufMask = CurDAG->getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
718 CurDAG->getConstant(0x80808080, MVT::i32),
719 CurDAG->getConstant(0x80808003, MVT::i32),
720 CurDAG->getConstant(0x80808080, MVT::i32),
721 CurDAG->getConstant(0x8080800b, MVT::i32));
725 SDNode *shufMaskLoad = emitBuildVector(shufMask.getNode());
727 HandleSDNode PromoteScalar(CurDAG->getNode(SPUISD::PREFSLOT2VEC, dl,
731 if (SDNode *N = SelectCode(PromoteScalar.getValue().getNode()))
732 PromScalar = SDValue(N, 0);
734 PromScalar = PromoteScalar.getValue();
736 SDValue zextShuffle =
737 CurDAG->getNode(SPUISD::SHUFB, dl, OpVecVT,
738 PromScalar, PromScalar,
739 SDValue(shufMaskLoad, 0));
741 HandleSDNode Dummy2(zextShuffle);
742 if (SDNode *N = SelectCode(Dummy2.getValue().getNode()))
743 zextShuffle = SDValue(N, 0);
745 zextShuffle = Dummy2.getValue();
746 HandleSDNode Dummy(CurDAG->getNode(SPUISD::VEC2PREFSLOT, dl, OpVT,
749 CurDAG->ReplaceAllUsesWith(N, Dummy.getValue().getNode());
750 SelectCode(Dummy.getValue().getNode());
751 return Dummy.getValue().getNode();
752 } else if (Opc == ISD::ADD && (OpVT == MVT::i64 || OpVT == MVT::v2i64)) {
754 emitBuildVector(getCarryGenerateShufMask(*CurDAG, dl).getNode());
756 HandleSDNode Dummy(CurDAG->getNode(SPUISD::ADD64_MARKER, dl, OpVT,
757 N->getOperand(0), N->getOperand(1),
758 SDValue(CGLoad, 0)));
760 CurDAG->ReplaceAllUsesWith(N, Dummy.getValue().getNode());
761 if (SDNode *N = SelectCode(Dummy.getValue().getNode()))
763 return Dummy.getValue().getNode();
764 } else if (Opc == ISD::SUB && (OpVT == MVT::i64 || OpVT == MVT::v2i64)) {
766 emitBuildVector(getBorrowGenerateShufMask(*CurDAG, dl).getNode());
768 HandleSDNode Dummy(CurDAG->getNode(SPUISD::SUB64_MARKER, dl, OpVT,
769 N->getOperand(0), N->getOperand(1),
770 SDValue(CGLoad, 0)));
772 CurDAG->ReplaceAllUsesWith(N, Dummy.getValue().getNode());
773 if (SDNode *N = SelectCode(Dummy.getValue().getNode()))
775 return Dummy.getValue().getNode();
776 } else if (Opc == ISD::MUL && (OpVT == MVT::i64 || OpVT == MVT::v2i64)) {
778 emitBuildVector(getCarryGenerateShufMask(*CurDAG, dl).getNode());
780 HandleSDNode Dummy(CurDAG->getNode(SPUISD::MUL64_MARKER, dl, OpVT,
781 N->getOperand(0), N->getOperand(1),
782 SDValue(CGLoad, 0)));
783 CurDAG->ReplaceAllUsesWith(N, Dummy.getValue().getNode());
784 if (SDNode *N = SelectCode(Dummy.getValue().getNode()))
786 return Dummy.getValue().getNode();
787 } else if (Opc == ISD::TRUNCATE) {
788 SDValue Op0 = N->getOperand(0);
789 if ((Op0.getOpcode() == ISD::SRA || Op0.getOpcode() == ISD::SRL)
791 && Op0.getValueType() == MVT::i64) {
792 // Catch (truncate:i32 ([sra|srl]:i64 arg, c), where c >= 32
794 // Take advantage of the fact that the upper 32 bits are in the
795 // i32 preferred slot and avoid shuffle gymnastics:
796 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Op0.getOperand(1));
798 unsigned shift_amt = unsigned(CN->getZExtValue());
800 if (shift_amt >= 32) {
802 CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS, dl, OpVT,
803 Op0.getOperand(0), getRC(MVT::i32));
807 // Take care of the additional shift, if present:
808 SDValue shift = CurDAG->getTargetConstant(shift_amt, MVT::i32);
809 unsigned Opc = SPU::ROTMAIr32_i32;
811 if (Op0.getOpcode() == ISD::SRL)
814 hi32 = CurDAG->getMachineNode(Opc, dl, OpVT, SDValue(hi32, 0),
822 } else if (Opc == ISD::SHL) {
823 if (OpVT == MVT::i64)
824 return SelectSHLi64(N, OpVT);
825 } else if (Opc == ISD::SRL) {
826 if (OpVT == MVT::i64)
827 return SelectSRLi64(N, OpVT);
828 } else if (Opc == ISD::SRA) {
829 if (OpVT == MVT::i64)
830 return SelectSRAi64(N, OpVT);
831 } else if (Opc == ISD::FNEG
832 && (OpVT == MVT::f64 || OpVT == MVT::v2f64)) {
833 DebugLoc dl = N->getDebugLoc();
834 // Check if the pattern is a special form of DFNMS:
835 // (fneg (fsub (fmul R64FP:$rA, R64FP:$rB), R64FP:$rC))
836 SDValue Op0 = N->getOperand(0);
837 if (Op0.getOpcode() == ISD::FSUB) {
838 SDValue Op00 = Op0.getOperand(0);
839 if (Op00.getOpcode() == ISD::FMUL) {
840 unsigned Opc = SPU::DFNMSf64;
841 if (OpVT == MVT::v2f64)
842 Opc = SPU::DFNMSv2f64;
844 return CurDAG->getMachineNode(Opc, dl, OpVT,
851 SDValue negConst = CurDAG->getConstant(0x8000000000000000ULL, MVT::i64);
852 SDNode *signMask = 0;
853 unsigned Opc = SPU::XORfneg64;
855 if (OpVT == MVT::f64) {
856 signMask = SelectI64Constant(negConst.getNode(), MVT::i64, dl);
857 } else if (OpVT == MVT::v2f64) {
858 Opc = SPU::XORfnegvec;
859 signMask = emitBuildVector(CurDAG->getNode(ISD::BUILD_VECTOR, dl,
861 negConst, negConst).getNode());
864 return CurDAG->getMachineNode(Opc, dl, OpVT,
865 N->getOperand(0), SDValue(signMask, 0));
866 } else if (Opc == ISD::FABS) {
867 if (OpVT == MVT::f64) {
868 SDNode *signMask = SelectI64Constant(0x7fffffffffffffffULL, MVT::i64, dl);
869 return CurDAG->getMachineNode(SPU::ANDfabs64, dl, OpVT,
870 N->getOperand(0), SDValue(signMask, 0));
871 } else if (OpVT == MVT::v2f64) {
872 SDValue absConst = CurDAG->getConstant(0x7fffffffffffffffULL, MVT::i64);
873 SDValue absVec = CurDAG->getNode(ISD::BUILD_VECTOR, dl, MVT::v2i64,
875 SDNode *signMask = emitBuildVector(absVec.getNode());
876 return CurDAG->getMachineNode(SPU::ANDfabsvec, dl, OpVT,
877 N->getOperand(0), SDValue(signMask, 0));
879 } else if (Opc == SPUISD::LDRESULT) {
880 // Custom select instructions for LDRESULT
881 EVT VT = N->getValueType(0);
882 SDValue Arg = N->getOperand(0);
883 SDValue Chain = N->getOperand(1);
885 const valtype_map_s *vtm = getValueTypeMapEntry(VT);
887 if (vtm->ldresult_ins == 0) {
888 report_fatal_error("LDRESULT for unsupported type: " +
889 Twine(VT.getEVTString()));
892 Opc = vtm->ldresult_ins;
893 if (vtm->ldresult_imm) {
894 SDValue Zero = CurDAG->getTargetConstant(0, VT);
896 Result = CurDAG->getMachineNode(Opc, dl, VT, MVT::Other, Arg, Zero, Chain);
898 Result = CurDAG->getMachineNode(Opc, dl, VT, MVT::Other, Arg, Arg, Chain);
902 } else if (Opc == SPUISD::IndirectAddr) {
903 // Look at the operands: SelectCode() will catch the cases that aren't
904 // specifically handled here.
906 // SPUInstrInfo catches the following patterns:
907 // (SPUindirect (SPUhi ...), (SPUlo ...))
908 // (SPUindirect $sp, imm)
909 EVT VT = N->getValueType(0);
910 SDValue Op0 = N->getOperand(0);
911 SDValue Op1 = N->getOperand(1);
914 if ((Op0.getOpcode() != SPUISD::Hi && Op1.getOpcode() != SPUISD::Lo)
915 || (Op0.getOpcode() == ISD::Register
916 && ((RN = dyn_cast<RegisterSDNode>(Op0.getNode())) != 0
917 && RN->getReg() != SPU::R1))) {
920 if (Op1.getOpcode() == ISD::Constant) {
921 ConstantSDNode *CN = cast<ConstantSDNode>(Op1);
922 Op1 = CurDAG->getTargetConstant(CN->getSExtValue(), VT);
923 if (isInt<10>(CN->getSExtValue())) {
927 Ops[1] = SDValue(CurDAG->getMachineNode(SPU::ILr32, dl,
940 return CurDAG->SelectNodeTo(N, NewOpc, OpVT, Ops, n_ops);
942 return CurDAG->getMachineNode(NewOpc, dl, OpVT, Ops, n_ops);
944 return SelectCode(N);
948 * Emit the instruction sequence for i64 left shifts. The basic algorithm
949 * is to fill the bottom two word slots with zeros so that zeros are shifted
950 * in as the entire quadword is shifted left.
952 * \note This code could also be used to implement v2i64 shl.
954 * @param Op The shl operand
955 * @param OpVT Op's machine value value type (doesn't need to be passed, but
956 * makes life easier.)
957 * @return The SDNode with the entire instruction sequence
960 SPUDAGToDAGISel::SelectSHLi64(SDNode *N, EVT OpVT) {
961 SDValue Op0 = N->getOperand(0);
962 EVT VecVT = EVT::getVectorVT(*CurDAG->getContext(),
963 OpVT, (128 / OpVT.getSizeInBits()));
964 SDValue ShiftAmt = N->getOperand(1);
965 EVT ShiftAmtVT = ShiftAmt.getValueType();
966 SDNode *VecOp0, *SelMask, *ZeroFill, *Shift = 0;
968 DebugLoc dl = N->getDebugLoc();
970 VecOp0 = CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS, dl, VecVT,
971 Op0, getRC(MVT::v2i64) );
972 SelMaskVal = CurDAG->getTargetConstant(0xff00ULL, MVT::i16);
973 SelMask = CurDAG->getMachineNode(SPU::FSMBIv2i64, dl, VecVT, SelMaskVal);
974 ZeroFill = CurDAG->getMachineNode(SPU::ILv2i64, dl, VecVT,
975 CurDAG->getTargetConstant(0, OpVT));
976 VecOp0 = CurDAG->getMachineNode(SPU::SELBv2i64, dl, VecVT,
977 SDValue(ZeroFill, 0),
979 SDValue(SelMask, 0));
981 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(ShiftAmt)) {
982 unsigned bytes = unsigned(CN->getZExtValue()) >> 3;
983 unsigned bits = unsigned(CN->getZExtValue()) & 7;
987 CurDAG->getMachineNode(SPU::SHLQBYIv2i64, dl, VecVT,
989 CurDAG->getTargetConstant(bytes, ShiftAmtVT));
994 CurDAG->getMachineNode(SPU::SHLQBIIv2i64, dl, VecVT,
995 SDValue((Shift != 0 ? Shift : VecOp0), 0),
996 CurDAG->getTargetConstant(bits, ShiftAmtVT));
1000 CurDAG->getMachineNode(SPU::ROTMIr32, dl, ShiftAmtVT,
1002 CurDAG->getTargetConstant(3, ShiftAmtVT));
1004 CurDAG->getMachineNode(SPU::ANDIr32, dl, ShiftAmtVT,
1006 CurDAG->getTargetConstant(7, ShiftAmtVT));
1008 CurDAG->getMachineNode(SPU::SHLQBYv2i64, dl, VecVT,
1009 SDValue(VecOp0, 0), SDValue(Bytes, 0));
1011 CurDAG->getMachineNode(SPU::SHLQBIv2i64, dl, VecVT,
1012 SDValue(Shift, 0), SDValue(Bits, 0));
1015 return CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS, dl,
1016 OpVT, SDValue(Shift, 0), getRC(MVT::i64));
1020 * Emit the instruction sequence for i64 logical right shifts.
1022 * @param Op The shl operand
1023 * @param OpVT Op's machine value value type (doesn't need to be passed, but
1024 * makes life easier.)
1025 * @return The SDNode with the entire instruction sequence
1028 SPUDAGToDAGISel::SelectSRLi64(SDNode *N, EVT OpVT) {
1029 SDValue Op0 = N->getOperand(0);
1030 EVT VecVT = EVT::getVectorVT(*CurDAG->getContext(),
1031 OpVT, (128 / OpVT.getSizeInBits()));
1032 SDValue ShiftAmt = N->getOperand(1);
1033 EVT ShiftAmtVT = ShiftAmt.getValueType();
1034 SDNode *VecOp0, *Shift = 0;
1035 DebugLoc dl = N->getDebugLoc();
1037 VecOp0 = CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS, dl, VecVT,
1038 Op0, getRC(MVT::v2i64) );
1040 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(ShiftAmt)) {
1041 unsigned bytes = unsigned(CN->getZExtValue()) >> 3;
1042 unsigned bits = unsigned(CN->getZExtValue()) & 7;
1046 CurDAG->getMachineNode(SPU::ROTQMBYIv2i64, dl, VecVT,
1048 CurDAG->getTargetConstant(bytes, ShiftAmtVT));
1053 CurDAG->getMachineNode(SPU::ROTQMBIIv2i64, dl, VecVT,
1054 SDValue((Shift != 0 ? Shift : VecOp0), 0),
1055 CurDAG->getTargetConstant(bits, ShiftAmtVT));
1059 CurDAG->getMachineNode(SPU::ROTMIr32, dl, ShiftAmtVT,
1061 CurDAG->getTargetConstant(3, ShiftAmtVT));
1063 CurDAG->getMachineNode(SPU::ANDIr32, dl, ShiftAmtVT,
1065 CurDAG->getTargetConstant(7, ShiftAmtVT));
1067 // Ensure that the shift amounts are negated!
1068 Bytes = CurDAG->getMachineNode(SPU::SFIr32, dl, ShiftAmtVT,
1070 CurDAG->getTargetConstant(0, ShiftAmtVT));
1072 Bits = CurDAG->getMachineNode(SPU::SFIr32, dl, ShiftAmtVT,
1074 CurDAG->getTargetConstant(0, ShiftAmtVT));
1077 CurDAG->getMachineNode(SPU::ROTQMBYv2i64, dl, VecVT,
1078 SDValue(VecOp0, 0), SDValue(Bytes, 0));
1080 CurDAG->getMachineNode(SPU::ROTQMBIv2i64, dl, VecVT,
1081 SDValue(Shift, 0), SDValue(Bits, 0));
1084 return CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS, dl,
1085 OpVT, SDValue(Shift, 0), getRC(MVT::i64));
1089 * Emit the instruction sequence for i64 arithmetic right shifts.
1091 * @param Op The shl operand
1092 * @param OpVT Op's machine value value type (doesn't need to be passed, but
1093 * makes life easier.)
1094 * @return The SDNode with the entire instruction sequence
1097 SPUDAGToDAGISel::SelectSRAi64(SDNode *N, EVT OpVT) {
1098 // Promote Op0 to vector
1099 EVT VecVT = EVT::getVectorVT(*CurDAG->getContext(),
1100 OpVT, (128 / OpVT.getSizeInBits()));
1101 SDValue ShiftAmt = N->getOperand(1);
1102 EVT ShiftAmtVT = ShiftAmt.getValueType();
1103 DebugLoc dl = N->getDebugLoc();
1106 CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS, dl,
1107 VecVT, N->getOperand(0), getRC(MVT::v2i64));
1109 SDValue SignRotAmt = CurDAG->getTargetConstant(31, ShiftAmtVT);
1111 CurDAG->getMachineNode(SPU::ROTMAIv2i64_i32, dl, MVT::v2i64,
1112 SDValue(VecOp0, 0), SignRotAmt);
1113 SDNode *UpperHalfSign =
1114 CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS, dl,
1115 MVT::i32, SDValue(SignRot, 0), getRC(MVT::i32));
1117 SDNode *UpperHalfSignMask =
1118 CurDAG->getMachineNode(SPU::FSM64r32, dl, VecVT, SDValue(UpperHalfSign, 0));
1119 SDNode *UpperLowerMask =
1120 CurDAG->getMachineNode(SPU::FSMBIv2i64, dl, VecVT,
1121 CurDAG->getTargetConstant(0xff00ULL, MVT::i16));
1122 SDNode *UpperLowerSelect =
1123 CurDAG->getMachineNode(SPU::SELBv2i64, dl, VecVT,
1124 SDValue(UpperHalfSignMask, 0),
1126 SDValue(UpperLowerMask, 0));
1130 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(ShiftAmt)) {
1131 unsigned bytes = unsigned(CN->getZExtValue()) >> 3;
1132 unsigned bits = unsigned(CN->getZExtValue()) & 7;
1137 CurDAG->getMachineNode(SPU::ROTQBYIv2i64, dl, VecVT,
1138 SDValue(UpperLowerSelect, 0),
1139 CurDAG->getTargetConstant(bytes, ShiftAmtVT));
1145 CurDAG->getMachineNode(SPU::ROTQBIIv2i64, dl, VecVT,
1146 SDValue((Shift != 0 ? Shift : UpperLowerSelect), 0),
1147 CurDAG->getTargetConstant(bits, ShiftAmtVT));
1151 CurDAG->getMachineNode(SPU::SFIr32, dl, ShiftAmtVT,
1152 ShiftAmt, CurDAG->getTargetConstant(0, ShiftAmtVT));
1155 CurDAG->getMachineNode(SPU::ROTQBYBIv2i64_r32, dl, VecVT,
1156 SDValue(UpperLowerSelect, 0), SDValue(NegShift, 0));
1158 CurDAG->getMachineNode(SPU::ROTQBIv2i64, dl, VecVT,
1159 SDValue(Shift, 0), SDValue(NegShift, 0));
1162 return CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS, dl,
1163 OpVT, SDValue(Shift, 0), getRC(MVT::i64));
1167 Do the necessary magic necessary to load a i64 constant
1169 SDNode *SPUDAGToDAGISel::SelectI64Constant(SDNode *N, EVT OpVT,
1171 ConstantSDNode *CN = cast<ConstantSDNode>(N);
1172 return SelectI64Constant(CN->getZExtValue(), OpVT, dl);
1175 SDNode *SPUDAGToDAGISel::SelectI64Constant(uint64_t Value64, EVT OpVT,
1177 EVT OpVecVT = EVT::getVectorVT(*CurDAG->getContext(), OpVT, 2);
1179 SPU::LowerV2I64Splat(OpVecVT, *CurDAG, Value64, dl);
1181 // Here's where it gets interesting, because we have to parse out the
1182 // subtree handed back in i64vec:
1184 if (i64vec.getOpcode() == ISD::BIT_CONVERT) {
1185 // The degenerate case where the upper and lower bits in the splat are
1187 SDValue Op0 = i64vec.getOperand(0);
1189 ReplaceUses(i64vec, Op0);
1190 return CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS, dl, OpVT,
1191 SDValue(emitBuildVector(Op0.getNode()), 0),
1193 } else if (i64vec.getOpcode() == SPUISD::SHUFB) {
1194 SDValue lhs = i64vec.getOperand(0);
1195 SDValue rhs = i64vec.getOperand(1);
1196 SDValue shufmask = i64vec.getOperand(2);
1198 if (lhs.getOpcode() == ISD::BIT_CONVERT) {
1199 ReplaceUses(lhs, lhs.getOperand(0));
1200 lhs = lhs.getOperand(0);
1203 SDNode *lhsNode = (lhs.getNode()->isMachineOpcode()
1205 : emitBuildVector(lhs.getNode()));
1207 if (rhs.getOpcode() == ISD::BIT_CONVERT) {
1208 ReplaceUses(rhs, rhs.getOperand(0));
1209 rhs = rhs.getOperand(0);
1212 SDNode *rhsNode = (rhs.getNode()->isMachineOpcode()
1214 : emitBuildVector(rhs.getNode()));
1216 if (shufmask.getOpcode() == ISD::BIT_CONVERT) {
1217 ReplaceUses(shufmask, shufmask.getOperand(0));
1218 shufmask = shufmask.getOperand(0);
1221 SDNode *shufMaskNode = (shufmask.getNode()->isMachineOpcode()
1222 ? shufmask.getNode()
1223 : emitBuildVector(shufmask.getNode()));
1226 CurDAG->getNode(SPUISD::SHUFB, dl, OpVecVT,
1227 SDValue(lhsNode, 0), SDValue(rhsNode, 0),
1228 SDValue(shufMaskNode, 0));
1229 HandleSDNode Dummy(shufNode);
1230 SDNode *SN = SelectCode(Dummy.getValue().getNode());
1231 if (SN == 0) SN = Dummy.getValue().getNode();
1233 return CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS, dl,
1234 OpVT, SDValue(SN, 0), getRC(MVT::i64));
1235 } else if (i64vec.getOpcode() == ISD::BUILD_VECTOR) {
1236 return CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS, dl, OpVT,
1237 SDValue(emitBuildVector(i64vec.getNode()), 0),
1240 report_fatal_error("SPUDAGToDAGISel::SelectI64Constant: Unhandled i64vec"
1245 /// createSPUISelDag - This pass converts a legalized DAG into a
1246 /// SPU-specific DAG, ready for instruction scheduling.
1248 FunctionPass *llvm::createSPUISelDag(SPUTargetMachine &TM) {
1249 return new SPUDAGToDAGISel(TM);