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 "SPUISelLowering.h"
18 #include "SPUHazardRecognizers.h"
19 #include "SPUFrameInfo.h"
20 #include "SPURegisterNames.h"
21 #include "SPUTargetMachine.h"
22 #include "llvm/CodeGen/MachineConstantPool.h"
23 #include "llvm/CodeGen/MachineInstrBuilder.h"
24 #include "llvm/CodeGen/MachineFunction.h"
25 #include "llvm/CodeGen/SelectionDAG.h"
26 #include "llvm/CodeGen/SelectionDAGISel.h"
27 #include "llvm/CodeGen/PseudoSourceValue.h"
28 #include "llvm/Target/TargetOptions.h"
29 #include "llvm/ADT/Statistic.h"
30 #include "llvm/Constants.h"
31 #include "llvm/GlobalValue.h"
32 #include "llvm/Intrinsics.h"
33 #include "llvm/LLVMContext.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/ErrorHandling.h"
36 #include "llvm/Support/MathExtras.h"
37 #include "llvm/Support/Compiler.h"
38 #include "llvm/Support/raw_ostream.h"
43 //! ConstantSDNode predicate for i32 sign-extended, 10-bit immediates
45 isI64IntS10Immediate(ConstantSDNode *CN)
47 return isS10Constant(CN->getSExtValue());
50 //! ConstantSDNode predicate for i32 sign-extended, 10-bit immediates
52 isI32IntS10Immediate(ConstantSDNode *CN)
54 return isS10Constant(CN->getSExtValue());
57 //! ConstantSDNode predicate for i32 unsigned 10-bit immediate values
59 isI32IntU10Immediate(ConstantSDNode *CN)
61 return isU10Constant(CN->getSExtValue());
64 //! ConstantSDNode predicate for i16 sign-extended, 10-bit immediate values
66 isI16IntS10Immediate(ConstantSDNode *CN)
68 return isS10Constant(CN->getSExtValue());
71 //! SDNode predicate for i16 sign-extended, 10-bit immediate values
73 isI16IntS10Immediate(SDNode *N)
75 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N);
76 return (CN != 0 && isI16IntS10Immediate(CN));
79 //! ConstantSDNode predicate for i16 unsigned 10-bit immediate values
81 isI16IntU10Immediate(ConstantSDNode *CN)
83 return isU10Constant((short) CN->getZExtValue());
86 //! SDNode predicate for i16 sign-extended, 10-bit immediate values
88 isI16IntU10Immediate(SDNode *N)
90 return (N->getOpcode() == ISD::Constant
91 && isI16IntU10Immediate(cast<ConstantSDNode>(N)));
94 //! ConstantSDNode predicate for signed 16-bit values
96 \arg CN The constant SelectionDAG node holding the value
97 \arg Imm The returned 16-bit value, if returning true
99 This predicate tests the value in \a CN to see whether it can be
100 represented as a 16-bit, sign-extended quantity. Returns true if
104 isIntS16Immediate(ConstantSDNode *CN, short &Imm)
106 EVT vt = CN->getValueType(0);
107 Imm = (short) CN->getZExtValue();
108 if (vt.getSimpleVT() >= MVT::i1 && vt.getSimpleVT() <= MVT::i16) {
110 } else if (vt == MVT::i32) {
111 int32_t i_val = (int32_t) CN->getZExtValue();
112 short s_val = (short) i_val;
113 return i_val == s_val;
115 int64_t i_val = (int64_t) CN->getZExtValue();
116 short s_val = (short) i_val;
117 return i_val == s_val;
123 //! SDNode predicate for signed 16-bit values.
125 isIntS16Immediate(SDNode *N, short &Imm)
127 return (N->getOpcode() == ISD::Constant
128 && isIntS16Immediate(cast<ConstantSDNode>(N), Imm));
131 //! ConstantFPSDNode predicate for representing floats as 16-bit sign ext.
133 isFPS16Immediate(ConstantFPSDNode *FPN, short &Imm)
135 EVT vt = FPN->getValueType(0);
136 if (vt == MVT::f32) {
137 int val = FloatToBits(FPN->getValueAPF().convertToFloat());
138 int sval = (int) ((val << 16) >> 16);
147 isHighLow(const SDValue &Op)
149 return (Op.getOpcode() == SPUISD::IndirectAddr
150 && ((Op.getOperand(0).getOpcode() == SPUISD::Hi
151 && Op.getOperand(1).getOpcode() == SPUISD::Lo)
152 || (Op.getOperand(0).getOpcode() == SPUISD::Lo
153 && Op.getOperand(1).getOpcode() == SPUISD::Hi)));
156 //===------------------------------------------------------------------===//
157 //! EVT to "useful stuff" mapping structure:
159 struct valtype_map_s {
161 unsigned ldresult_ins; /// LDRESULT instruction (0 = undefined)
162 bool ldresult_imm; /// LDRESULT instruction requires immediate?
163 unsigned lrinst; /// LR instruction
166 const valtype_map_s valtype_map[] = {
167 { MVT::i8, SPU::ORBIr8, true, SPU::LRr8 },
168 { MVT::i16, SPU::ORHIr16, true, SPU::LRr16 },
169 { MVT::i32, SPU::ORIr32, true, SPU::LRr32 },
170 { MVT::i64, SPU::ORr64, false, SPU::LRr64 },
171 { MVT::f32, SPU::ORf32, false, SPU::LRf32 },
172 { MVT::f64, SPU::ORf64, false, SPU::LRf64 },
173 // vector types... (sigh!)
174 { MVT::v16i8, 0, false, SPU::LRv16i8 },
175 { MVT::v8i16, 0, false, SPU::LRv8i16 },
176 { MVT::v4i32, 0, false, SPU::LRv4i32 },
177 { MVT::v2i64, 0, false, SPU::LRv2i64 },
178 { MVT::v4f32, 0, false, SPU::LRv4f32 },
179 { MVT::v2f64, 0, false, SPU::LRv2f64 }
182 const size_t n_valtype_map = sizeof(valtype_map) / sizeof(valtype_map[0]);
184 const valtype_map_s *getValueTypeMapEntry(EVT VT)
186 const valtype_map_s *retval = 0;
187 for (size_t i = 0; i < n_valtype_map; ++i) {
188 if (valtype_map[i].VT == VT) {
189 retval = valtype_map + i;
198 raw_string_ostream Msg(msg);
199 Msg << "SPUISelDAGToDAG.cpp: getValueTypeMapEntry returns NULL for "
200 << VT.getEVTString();
201 llvm_report_error(Msg.str());
208 //! Generate the carry-generate shuffle mask.
209 SDValue getCarryGenerateShufMask(SelectionDAG &DAG, DebugLoc dl) {
210 SmallVector<SDValue, 16 > ShufBytes;
212 // Create the shuffle mask for "rotating" the borrow up one register slot
213 // once the borrow is generated.
214 ShufBytes.push_back(DAG.getConstant(0x04050607, MVT::i32));
215 ShufBytes.push_back(DAG.getConstant(0x80808080, MVT::i32));
216 ShufBytes.push_back(DAG.getConstant(0x0c0d0e0f, MVT::i32));
217 ShufBytes.push_back(DAG.getConstant(0x80808080, MVT::i32));
219 return DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
220 &ShufBytes[0], ShufBytes.size());
223 //! Generate the borrow-generate shuffle mask
224 SDValue getBorrowGenerateShufMask(SelectionDAG &DAG, DebugLoc dl) {
225 SmallVector<SDValue, 16 > ShufBytes;
227 // Create the shuffle mask for "rotating" the borrow up one register slot
228 // once the borrow is generated.
229 ShufBytes.push_back(DAG.getConstant(0x04050607, MVT::i32));
230 ShufBytes.push_back(DAG.getConstant(0xc0c0c0c0, MVT::i32));
231 ShufBytes.push_back(DAG.getConstant(0x0c0d0e0f, MVT::i32));
232 ShufBytes.push_back(DAG.getConstant(0xc0c0c0c0, MVT::i32));
234 return DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
235 &ShufBytes[0], ShufBytes.size());
238 //===------------------------------------------------------------------===//
239 /// SPUDAGToDAGISel - Cell SPU-specific code to select SPU machine
240 /// instructions for SelectionDAG operations.
242 class SPUDAGToDAGISel :
243 public SelectionDAGISel
245 SPUTargetMachine &TM;
246 SPUTargetLowering &SPUtli;
247 unsigned GlobalBaseReg;
250 explicit SPUDAGToDAGISel(SPUTargetMachine &tm) :
251 SelectionDAGISel(tm),
253 SPUtli(*tm.getTargetLowering())
256 virtual bool runOnMachineFunction(MachineFunction &MF) {
257 // Make sure we re-emit a set of the global base reg if necessary
259 SelectionDAGISel::runOnMachineFunction(MF);
263 /// getI32Imm - Return a target constant with the specified value, of type
265 inline SDValue getI32Imm(uint32_t Imm) {
266 return CurDAG->getTargetConstant(Imm, MVT::i32);
269 /// getI64Imm - Return a target constant with the specified value, of type
271 inline SDValue getI64Imm(uint64_t Imm) {
272 return CurDAG->getTargetConstant(Imm, MVT::i64);
275 /// getSmallIPtrImm - Return a target constant of pointer type.
276 inline SDValue getSmallIPtrImm(unsigned Imm) {
277 return CurDAG->getTargetConstant(Imm, SPUtli.getPointerTy());
280 SDNode *emitBuildVector(SDValue build_vec) {
281 EVT vecVT = build_vec.getValueType();
282 EVT eltVT = vecVT.getVectorElementType();
283 SDNode *bvNode = build_vec.getNode();
284 DebugLoc dl = bvNode->getDebugLoc();
286 // Check to see if this vector can be represented as a CellSPU immediate
287 // constant by invoking all of the instruction selection predicates:
288 if (((vecVT == MVT::v8i16) &&
289 (SPU::get_vec_i16imm(bvNode, *CurDAG, MVT::i16).getNode() != 0)) ||
290 ((vecVT == MVT::v4i32) &&
291 ((SPU::get_vec_i16imm(bvNode, *CurDAG, MVT::i32).getNode() != 0) ||
292 (SPU::get_ILHUvec_imm(bvNode, *CurDAG, MVT::i32).getNode() != 0) ||
293 (SPU::get_vec_u18imm(bvNode, *CurDAG, MVT::i32).getNode() != 0) ||
294 (SPU::get_v4i32_imm(bvNode, *CurDAG).getNode() != 0))) ||
295 ((vecVT == MVT::v2i64) &&
296 ((SPU::get_vec_i16imm(bvNode, *CurDAG, MVT::i64).getNode() != 0) ||
297 (SPU::get_ILHUvec_imm(bvNode, *CurDAG, MVT::i64).getNode() != 0) ||
298 (SPU::get_vec_u18imm(bvNode, *CurDAG, MVT::i64).getNode() != 0))))
299 return Select(build_vec);
301 // No, need to emit a constant pool spill:
302 std::vector<Constant*> CV;
304 for (size_t i = 0; i < build_vec.getNumOperands(); ++i) {
305 ConstantSDNode *V = dyn_cast<ConstantSDNode > (build_vec.getOperand(i));
306 CV.push_back(const_cast<ConstantInt *> (V->getConstantIntValue()));
309 Constant *CP = ConstantVector::get(CV);
310 SDValue CPIdx = CurDAG->getConstantPool(CP, SPUtli.getPointerTy());
311 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
312 SDValue CGPoolOffset =
313 SPU::LowerConstantPool(CPIdx, *CurDAG,
314 SPUtli.getSPUTargetMachine());
315 return SelectCode(CurDAG->getLoad(build_vec.getValueType(), dl,
316 CurDAG->getEntryNode(), CGPoolOffset,
317 PseudoSourceValue::getConstantPool(), 0,
321 /// Select - Convert the specified operand from a target-independent to a
322 /// target-specific node if it hasn't already been changed.
323 SDNode *Select(SDValue Op);
325 //! Emit the instruction sequence for i64 shl
326 SDNode *SelectSHLi64(SDValue &Op, EVT OpVT);
328 //! Emit the instruction sequence for i64 srl
329 SDNode *SelectSRLi64(SDValue &Op, EVT OpVT);
331 //! Emit the instruction sequence for i64 sra
332 SDNode *SelectSRAi64(SDValue &Op, EVT OpVT);
334 //! Emit the necessary sequence for loading i64 constants:
335 SDNode *SelectI64Constant(SDValue &Op, EVT OpVT, DebugLoc dl);
337 //! Alternate instruction emit sequence for loading i64 constants
338 SDNode *SelectI64Constant(uint64_t i64const, EVT OpVT, DebugLoc dl);
340 //! Returns true if the address N is an A-form (local store) address
341 bool SelectAFormAddr(SDValue Op, SDValue N, SDValue &Base,
344 //! D-form address predicate
345 bool SelectDFormAddr(SDValue Op, SDValue N, SDValue &Base,
348 /// Alternate D-form address using i7 offset predicate
349 bool SelectDForm2Addr(SDValue Op, SDValue N, SDValue &Disp,
352 /// D-form address selection workhorse
353 bool DFormAddressPredicate(SDValue Op, SDValue N, SDValue &Disp,
354 SDValue &Base, int minOffset, int maxOffset);
356 //! Address predicate if N can be expressed as an indexed [r+r] operation.
357 bool SelectXFormAddr(SDValue Op, SDValue N, SDValue &Base,
360 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
361 /// inline asm expressions.
362 virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
364 std::vector<SDValue> &OutOps) {
366 switch (ConstraintCode) {
367 default: return true;
369 if (!SelectDFormAddr(Op, Op, Op0, Op1)
370 && !SelectAFormAddr(Op, Op, Op0, Op1))
371 SelectXFormAddr(Op, Op, Op0, Op1);
373 case 'o': // offsetable
374 if (!SelectDFormAddr(Op, Op, Op0, Op1)
375 && !SelectAFormAddr(Op, Op, Op0, Op1)) {
377 Op1 = getSmallIPtrImm(0);
380 case 'v': // not offsetable
382 llvm_unreachable("InlineAsmMemoryOperand 'v' constraint not handled.");
384 SelectAddrIdxOnly(Op, Op, Op0, Op1);
389 OutOps.push_back(Op0);
390 OutOps.push_back(Op1);
394 /// InstructionSelect - This callback is invoked by
395 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
396 virtual void InstructionSelect();
398 virtual const char *getPassName() const {
399 return "Cell SPU DAG->DAG Pattern Instruction Selection";
402 /// CreateTargetHazardRecognizer - Return the hazard recognizer to use for
403 /// this target when scheduling the DAG.
404 virtual ScheduleHazardRecognizer *CreateTargetHazardRecognizer() {
405 const TargetInstrInfo *II = TM.getInstrInfo();
406 assert(II && "No InstrInfo?");
407 return new SPUHazardRecognizer(*II);
410 // Include the pieces autogenerated from the target description.
411 #include "SPUGenDAGISel.inc"
415 /// InstructionSelect - This callback is invoked by
416 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
418 SPUDAGToDAGISel::InstructionSelect()
422 // Select target instructions for the DAG.
424 CurDAG->RemoveDeadNodes();
428 \arg Op The ISD instruction operand
429 \arg N The address to be tested
430 \arg Base The base address
431 \arg Index The base address index
434 SPUDAGToDAGISel::SelectAFormAddr(SDValue Op, SDValue N, SDValue &Base,
436 // These match the addr256k operand type:
437 EVT OffsVT = MVT::i16;
438 SDValue Zero = CurDAG->getTargetConstant(0, OffsVT);
440 switch (N.getOpcode()) {
442 case ISD::ConstantPool:
443 case ISD::GlobalAddress:
444 llvm_report_error("SPU SelectAFormAddr: Constant/Pool/Global not lowered.");
447 case ISD::TargetConstant:
448 case ISD::TargetGlobalAddress:
449 case ISD::TargetJumpTable:
450 llvm_report_error("SPUSelectAFormAddr: Target Constant/Pool/Global "
451 "not wrapped as A-form address.");
454 case SPUISD::AFormAddr:
455 // Just load from memory if there's only a single use of the location,
456 // otherwise, this will get handled below with D-form offset addresses
458 SDValue Op0 = N.getOperand(0);
459 switch (Op0.getOpcode()) {
460 case ISD::TargetConstantPool:
461 case ISD::TargetJumpTable:
466 case ISD::TargetGlobalAddress: {
467 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op0);
468 GlobalValue *GV = GSDN->getGlobal();
469 if (GV->getAlignment() == 16) {
484 SPUDAGToDAGISel::SelectDForm2Addr(SDValue Op, SDValue N, SDValue &Disp,
486 const int minDForm2Offset = -(1 << 7);
487 const int maxDForm2Offset = (1 << 7) - 1;
488 return DFormAddressPredicate(Op, N, Disp, Base, minDForm2Offset,
493 \arg Op The ISD instruction (ignored)
494 \arg N The address to be tested
495 \arg Base Base address register/pointer
496 \arg Index Base address index
498 Examine the input address by a base register plus a signed 10-bit
499 displacement, [r+I10] (D-form address).
501 \return true if \a N is a D-form address with \a Base and \a Index set
502 to non-empty SDValue instances.
505 SPUDAGToDAGISel::SelectDFormAddr(SDValue Op, SDValue N, SDValue &Base,
507 return DFormAddressPredicate(Op, N, Base, Index,
508 SPUFrameInfo::minFrameOffset(),
509 SPUFrameInfo::maxFrameOffset());
513 SPUDAGToDAGISel::DFormAddressPredicate(SDValue Op, SDValue N, SDValue &Base,
514 SDValue &Index, int minOffset,
516 unsigned Opc = N.getOpcode();
517 EVT PtrTy = SPUtli.getPointerTy();
519 if (Opc == ISD::FrameIndex) {
520 // Stack frame index must be less than 512 (divided by 16):
521 FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(N);
522 int FI = int(FIN->getIndex());
523 DEBUG(errs() << "SelectDFormAddr: ISD::FrameIndex = "
525 if (SPUFrameInfo::FItoStackOffset(FI) < maxOffset) {
526 Base = CurDAG->getTargetConstant(0, PtrTy);
527 Index = CurDAG->getTargetFrameIndex(FI, PtrTy);
530 } else if (Opc == ISD::ADD) {
531 // Generated by getelementptr
532 const SDValue Op0 = N.getOperand(0);
533 const SDValue Op1 = N.getOperand(1);
535 if ((Op0.getOpcode() == SPUISD::Hi && Op1.getOpcode() == SPUISD::Lo)
536 || (Op1.getOpcode() == SPUISD::Hi && Op0.getOpcode() == SPUISD::Lo)) {
537 Base = CurDAG->getTargetConstant(0, PtrTy);
540 } else if (Op1.getOpcode() == ISD::Constant
541 || Op1.getOpcode() == ISD::TargetConstant) {
542 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Op1);
543 int32_t offset = int32_t(CN->getSExtValue());
545 if (Op0.getOpcode() == ISD::FrameIndex) {
546 FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Op0);
547 int FI = int(FIN->getIndex());
548 DEBUG(errs() << "SelectDFormAddr: ISD::ADD offset = " << offset
549 << " frame index = " << FI << "\n");
551 if (SPUFrameInfo::FItoStackOffset(FI) < maxOffset) {
552 Base = CurDAG->getTargetConstant(offset, PtrTy);
553 Index = CurDAG->getTargetFrameIndex(FI, PtrTy);
556 } else if (offset > minOffset && offset < maxOffset) {
557 Base = CurDAG->getTargetConstant(offset, PtrTy);
561 } else if (Op0.getOpcode() == ISD::Constant
562 || Op0.getOpcode() == ISD::TargetConstant) {
563 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Op0);
564 int32_t offset = int32_t(CN->getSExtValue());
566 if (Op1.getOpcode() == ISD::FrameIndex) {
567 FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Op1);
568 int FI = int(FIN->getIndex());
569 DEBUG(errs() << "SelectDFormAddr: ISD::ADD offset = " << offset
570 << " frame index = " << FI << "\n");
572 if (SPUFrameInfo::FItoStackOffset(FI) < maxOffset) {
573 Base = CurDAG->getTargetConstant(offset, PtrTy);
574 Index = CurDAG->getTargetFrameIndex(FI, PtrTy);
577 } else if (offset > minOffset && offset < maxOffset) {
578 Base = CurDAG->getTargetConstant(offset, PtrTy);
583 } else if (Opc == SPUISD::IndirectAddr) {
584 // Indirect with constant offset -> D-Form address
585 const SDValue Op0 = N.getOperand(0);
586 const SDValue Op1 = N.getOperand(1);
588 if (Op0.getOpcode() == SPUISD::Hi
589 && Op1.getOpcode() == SPUISD::Lo) {
590 // (SPUindirect (SPUhi <arg>, 0), (SPUlo <arg>, 0))
591 Base = CurDAG->getTargetConstant(0, PtrTy);
594 } else if (isa<ConstantSDNode>(Op0) || isa<ConstantSDNode>(Op1)) {
598 if (isa<ConstantSDNode>(Op1)) {
599 ConstantSDNode *CN = cast<ConstantSDNode>(Op1);
600 offset = int32_t(CN->getSExtValue());
602 } else if (isa<ConstantSDNode>(Op0)) {
603 ConstantSDNode *CN = cast<ConstantSDNode>(Op0);
604 offset = int32_t(CN->getSExtValue());
608 if (offset >= minOffset && offset <= maxOffset) {
609 Base = CurDAG->getTargetConstant(offset, PtrTy);
614 } else if (Opc == SPUISD::AFormAddr) {
615 Base = CurDAG->getTargetConstant(0, N.getValueType());
618 } else if (Opc == SPUISD::LDRESULT) {
619 Base = CurDAG->getTargetConstant(0, N.getValueType());
622 } else if (Opc == ISD::Register || Opc == ISD::CopyFromReg) {
623 unsigned OpOpc = Op.getOpcode();
625 if (OpOpc == ISD::STORE || OpOpc == ISD::LOAD) {
626 // Direct load/store without getelementptr
629 // Get the register from CopyFromReg
630 if (Opc == ISD::CopyFromReg)
631 Addr = N.getOperand(1);
633 Addr = N; // Register
635 Offs = ((OpOpc == ISD::STORE) ? Op.getOperand(3) : Op.getOperand(2));
637 if (Offs.getOpcode() == ISD::Constant || Offs.getOpcode() == ISD::UNDEF) {
638 if (Offs.getOpcode() == ISD::UNDEF)
639 Offs = CurDAG->getTargetConstant(0, Offs.getValueType());
646 /* If otherwise unadorned, default to D-form address with 0 offset: */
647 if (Opc == ISD::CopyFromReg) {
648 Index = N.getOperand(1);
653 Base = CurDAG->getTargetConstant(0, Index.getValueType());
662 \arg Op The ISD instruction operand
663 \arg N The address operand
664 \arg Base The base pointer operand
665 \arg Index The offset/index operand
667 If the address \a N can be expressed as an A-form or D-form address, returns
668 false. Otherwise, creates two operands, Base and Index that will become the
669 (r)(r) X-form address.
672 SPUDAGToDAGISel::SelectXFormAddr(SDValue Op, SDValue N, SDValue &Base,
674 if (!SelectAFormAddr(Op, N, Base, Index)
675 && !SelectDFormAddr(Op, N, Base, Index)) {
676 // If the address is neither A-form or D-form, punt and use an X-form
678 Base = N.getOperand(1);
679 Index = N.getOperand(0);
686 //! Convert the operand from a target-independent to a target-specific node
690 SPUDAGToDAGISel::Select(SDValue Op) {
691 SDNode *N = Op.getNode();
692 unsigned Opc = N->getOpcode();
695 EVT OpVT = Op.getValueType();
697 DebugLoc dl = N->getDebugLoc();
699 if (N->isMachineOpcode()) {
700 return NULL; // Already selected.
703 if (Opc == ISD::FrameIndex) {
704 int FI = cast<FrameIndexSDNode>(N)->getIndex();
705 SDValue TFI = CurDAG->getTargetFrameIndex(FI, Op.getValueType());
706 SDValue Imm0 = CurDAG->getTargetConstant(0, Op.getValueType());
715 Ops[0] = CurDAG->getRegister(SPU::R1, Op.getValueType());
716 Ops[1] = SDValue(CurDAG->getMachineNode(SPU::ILAr32, dl,
717 Op.getValueType(), TFI, Imm0),
721 } else if (Opc == ISD::Constant && OpVT == MVT::i64) {
722 // Catch the i64 constants that end up here. Note: The backend doesn't
723 // attempt to legalize the constant (it's useless because DAGCombiner
724 // will insert 64-bit constants and we can't stop it).
725 return SelectI64Constant(Op, OpVT, Op.getDebugLoc());
726 } else if ((Opc == ISD::ZERO_EXTEND || Opc == ISD::ANY_EXTEND)
727 && OpVT == MVT::i64) {
728 SDValue Op0 = Op.getOperand(0);
729 EVT Op0VT = Op0.getValueType();
730 EVT Op0VecVT = EVT::getVectorVT(*CurDAG->getContext(),
731 Op0VT, (128 / Op0VT.getSizeInBits()));
732 EVT OpVecVT = EVT::getVectorVT(*CurDAG->getContext(),
733 OpVT, (128 / OpVT.getSizeInBits()));
736 switch (Op0VT.getSimpleVT().SimpleTy) {
738 llvm_report_error("CellSPU Select: Unhandled zero/any extend EVT");
741 shufMask = CurDAG->getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
742 CurDAG->getConstant(0x80808080, MVT::i32),
743 CurDAG->getConstant(0x00010203, MVT::i32),
744 CurDAG->getConstant(0x80808080, MVT::i32),
745 CurDAG->getConstant(0x08090a0b, MVT::i32));
749 shufMask = CurDAG->getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
750 CurDAG->getConstant(0x80808080, MVT::i32),
751 CurDAG->getConstant(0x80800203, MVT::i32),
752 CurDAG->getConstant(0x80808080, MVT::i32),
753 CurDAG->getConstant(0x80800a0b, MVT::i32));
757 shufMask = CurDAG->getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
758 CurDAG->getConstant(0x80808080, MVT::i32),
759 CurDAG->getConstant(0x80808003, MVT::i32),
760 CurDAG->getConstant(0x80808080, MVT::i32),
761 CurDAG->getConstant(0x8080800b, MVT::i32));
765 SDNode *shufMaskLoad = emitBuildVector(shufMask);
766 SDNode *PromoteScalar =
767 SelectCode(CurDAG->getNode(SPUISD::PREFSLOT2VEC, dl, Op0VecVT, Op0));
769 SDValue zextShuffle =
770 CurDAG->getNode(SPUISD::SHUFB, dl, OpVecVT,
771 SDValue(PromoteScalar, 0),
772 SDValue(PromoteScalar, 0),
773 SDValue(shufMaskLoad, 0));
775 // N.B.: BIT_CONVERT replaces and updates the zextShuffle node, so we
776 // re-use it in the VEC2PREFSLOT selection without needing to explicitly
777 // call SelectCode (it's already done for us.)
778 SelectCode(CurDAG->getNode(ISD::BIT_CONVERT, dl, OpVecVT, zextShuffle));
779 return SelectCode(CurDAG->getNode(SPUISD::VEC2PREFSLOT, dl, OpVT,
781 } else if (Opc == ISD::ADD && (OpVT == MVT::i64 || OpVT == MVT::v2i64)) {
783 emitBuildVector(getCarryGenerateShufMask(*CurDAG, dl));
785 return SelectCode(CurDAG->getNode(SPUISD::ADD64_MARKER, dl, OpVT,
786 Op.getOperand(0), Op.getOperand(1),
787 SDValue(CGLoad, 0)));
788 } else if (Opc == ISD::SUB && (OpVT == MVT::i64 || OpVT == MVT::v2i64)) {
790 emitBuildVector(getBorrowGenerateShufMask(*CurDAG, dl));
792 return SelectCode(CurDAG->getNode(SPUISD::SUB64_MARKER, dl, OpVT,
793 Op.getOperand(0), Op.getOperand(1),
794 SDValue(CGLoad, 0)));
795 } else if (Opc == ISD::MUL && (OpVT == MVT::i64 || OpVT == MVT::v2i64)) {
797 emitBuildVector(getCarryGenerateShufMask(*CurDAG, dl));
799 return SelectCode(CurDAG->getNode(SPUISD::MUL64_MARKER, dl, OpVT,
800 Op.getOperand(0), Op.getOperand(1),
801 SDValue(CGLoad, 0)));
802 } else if (Opc == ISD::TRUNCATE) {
803 SDValue Op0 = Op.getOperand(0);
804 if ((Op0.getOpcode() == ISD::SRA || Op0.getOpcode() == ISD::SRL)
806 && Op0.getValueType() == MVT::i64) {
807 // Catch (truncate:i32 ([sra|srl]:i64 arg, c), where c >= 32
809 // Take advantage of the fact that the upper 32 bits are in the
810 // i32 preferred slot and avoid shuffle gymnastics:
811 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Op0.getOperand(1));
813 unsigned shift_amt = unsigned(CN->getZExtValue());
815 if (shift_amt >= 32) {
817 CurDAG->getMachineNode(SPU::ORr32_r64, dl, OpVT,
822 // Take care of the additional shift, if present:
823 SDValue shift = CurDAG->getTargetConstant(shift_amt, MVT::i32);
824 unsigned Opc = SPU::ROTMAIr32_i32;
826 if (Op0.getOpcode() == ISD::SRL)
829 hi32 = CurDAG->getMachineNode(Opc, dl, OpVT, SDValue(hi32, 0),
837 } else if (Opc == ISD::SHL) {
838 if (OpVT == MVT::i64) {
839 return SelectSHLi64(Op, OpVT);
841 } else if (Opc == ISD::SRL) {
842 if (OpVT == MVT::i64) {
843 return SelectSRLi64(Op, OpVT);
845 } else if (Opc == ISD::SRA) {
846 if (OpVT == MVT::i64) {
847 return SelectSRAi64(Op, OpVT);
849 } else if (Opc == ISD::FNEG
850 && (OpVT == MVT::f64 || OpVT == MVT::v2f64)) {
851 DebugLoc dl = Op.getDebugLoc();
852 // Check if the pattern is a special form of DFNMS:
853 // (fneg (fsub (fmul R64FP:$rA, R64FP:$rB), R64FP:$rC))
854 SDValue Op0 = Op.getOperand(0);
855 if (Op0.getOpcode() == ISD::FSUB) {
856 SDValue Op00 = Op0.getOperand(0);
857 if (Op00.getOpcode() == ISD::FMUL) {
858 unsigned Opc = SPU::DFNMSf64;
859 if (OpVT == MVT::v2f64)
860 Opc = SPU::DFNMSv2f64;
862 return CurDAG->getMachineNode(Opc, dl, OpVT,
869 SDValue negConst = CurDAG->getConstant(0x8000000000000000ULL, MVT::i64);
870 SDNode *signMask = 0;
871 unsigned Opc = SPU::XORfneg64;
873 if (OpVT == MVT::f64) {
874 signMask = SelectI64Constant(negConst, MVT::i64, dl);
875 } else if (OpVT == MVT::v2f64) {
876 Opc = SPU::XORfnegvec;
877 signMask = emitBuildVector(CurDAG->getNode(ISD::BUILD_VECTOR, dl,
879 negConst, negConst));
882 return CurDAG->getMachineNode(Opc, dl, OpVT,
883 Op.getOperand(0), SDValue(signMask, 0));
884 } else if (Opc == ISD::FABS) {
885 if (OpVT == MVT::f64) {
886 SDNode *signMask = SelectI64Constant(0x7fffffffffffffffULL, MVT::i64, dl);
887 return CurDAG->getMachineNode(SPU::ANDfabs64, dl, OpVT,
888 Op.getOperand(0), SDValue(signMask, 0));
889 } else if (OpVT == MVT::v2f64) {
890 SDValue absConst = CurDAG->getConstant(0x7fffffffffffffffULL, MVT::i64);
891 SDValue absVec = CurDAG->getNode(ISD::BUILD_VECTOR, dl, MVT::v2i64,
893 SDNode *signMask = emitBuildVector(absVec);
894 return CurDAG->getMachineNode(SPU::ANDfabsvec, dl, OpVT,
895 Op.getOperand(0), SDValue(signMask, 0));
897 } else if (Opc == SPUISD::LDRESULT) {
898 // Custom select instructions for LDRESULT
899 EVT VT = N->getValueType(0);
900 SDValue Arg = N->getOperand(0);
901 SDValue Chain = N->getOperand(1);
903 const valtype_map_s *vtm = getValueTypeMapEntry(VT);
905 if (vtm->ldresult_ins == 0) {
907 raw_string_ostream Msg(msg);
908 Msg << "LDRESULT for unsupported type: "
909 << VT.getEVTString();
910 llvm_report_error(Msg.str());
913 Opc = vtm->ldresult_ins;
914 if (vtm->ldresult_imm) {
915 SDValue Zero = CurDAG->getTargetConstant(0, VT);
917 Result = CurDAG->getMachineNode(Opc, dl, VT, MVT::Other, Arg, Zero, Chain);
919 Result = CurDAG->getMachineNode(Opc, dl, VT, MVT::Other, Arg, Arg, Chain);
923 } else if (Opc == SPUISD::IndirectAddr) {
924 // Look at the operands: SelectCode() will catch the cases that aren't
925 // specifically handled here.
927 // SPUInstrInfo catches the following patterns:
928 // (SPUindirect (SPUhi ...), (SPUlo ...))
929 // (SPUindirect $sp, imm)
930 EVT VT = Op.getValueType();
931 SDValue Op0 = N->getOperand(0);
932 SDValue Op1 = N->getOperand(1);
935 if ((Op0.getOpcode() != SPUISD::Hi && Op1.getOpcode() != SPUISD::Lo)
936 || (Op0.getOpcode() == ISD::Register
937 && ((RN = dyn_cast<RegisterSDNode>(Op0.getNode())) != 0
938 && RN->getReg() != SPU::R1))) {
940 if (Op1.getOpcode() == ISD::Constant) {
941 ConstantSDNode *CN = cast<ConstantSDNode>(Op1);
942 Op1 = CurDAG->getTargetConstant(CN->getSExtValue(), VT);
943 NewOpc = (isI32IntS10Immediate(CN) ? SPU::AIr32 : SPU::Ar32);
953 return CurDAG->SelectNodeTo(N, NewOpc, OpVT, Ops, n_ops);
955 return CurDAG->getMachineNode(NewOpc, dl, OpVT, Ops, n_ops);
957 return SelectCode(Op);
961 * Emit the instruction sequence for i64 left shifts. The basic algorithm
962 * is to fill the bottom two word slots with zeros so that zeros are shifted
963 * in as the entire quadword is shifted left.
965 * \note This code could also be used to implement v2i64 shl.
967 * @param Op The shl operand
968 * @param OpVT Op's machine value value type (doesn't need to be passed, but
969 * makes life easier.)
970 * @return The SDNode with the entire instruction sequence
973 SPUDAGToDAGISel::SelectSHLi64(SDValue &Op, EVT OpVT) {
974 SDValue Op0 = Op.getOperand(0);
975 EVT VecVT = EVT::getVectorVT(*CurDAG->getContext(),
976 OpVT, (128 / OpVT.getSizeInBits()));
977 SDValue ShiftAmt = Op.getOperand(1);
978 EVT ShiftAmtVT = ShiftAmt.getValueType();
979 SDNode *VecOp0, *SelMask, *ZeroFill, *Shift = 0;
981 DebugLoc dl = Op.getDebugLoc();
983 VecOp0 = CurDAG->getMachineNode(SPU::ORv2i64_i64, dl, VecVT, Op0);
984 SelMaskVal = CurDAG->getTargetConstant(0xff00ULL, MVT::i16);
985 SelMask = CurDAG->getMachineNode(SPU::FSMBIv2i64, dl, VecVT, SelMaskVal);
986 ZeroFill = CurDAG->getMachineNode(SPU::ILv2i64, dl, VecVT,
987 CurDAG->getTargetConstant(0, OpVT));
988 VecOp0 = CurDAG->getMachineNode(SPU::SELBv2i64, dl, VecVT,
989 SDValue(ZeroFill, 0),
991 SDValue(SelMask, 0));
993 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(ShiftAmt)) {
994 unsigned bytes = unsigned(CN->getZExtValue()) >> 3;
995 unsigned bits = unsigned(CN->getZExtValue()) & 7;
999 CurDAG->getMachineNode(SPU::SHLQBYIv2i64, dl, VecVT,
1001 CurDAG->getTargetConstant(bytes, ShiftAmtVT));
1006 CurDAG->getMachineNode(SPU::SHLQBIIv2i64, dl, VecVT,
1007 SDValue((Shift != 0 ? Shift : VecOp0), 0),
1008 CurDAG->getTargetConstant(bits, ShiftAmtVT));
1012 CurDAG->getMachineNode(SPU::ROTMIr32, dl, ShiftAmtVT,
1014 CurDAG->getTargetConstant(3, ShiftAmtVT));
1016 CurDAG->getMachineNode(SPU::ANDIr32, dl, ShiftAmtVT,
1018 CurDAG->getTargetConstant(7, ShiftAmtVT));
1020 CurDAG->getMachineNode(SPU::SHLQBYv2i64, dl, VecVT,
1021 SDValue(VecOp0, 0), SDValue(Bytes, 0));
1023 CurDAG->getMachineNode(SPU::SHLQBIv2i64, dl, VecVT,
1024 SDValue(Shift, 0), SDValue(Bits, 0));
1027 return CurDAG->getMachineNode(SPU::ORi64_v2i64, dl, OpVT, SDValue(Shift, 0));
1031 * Emit the instruction sequence for i64 logical right shifts.
1033 * @param Op The shl operand
1034 * @param OpVT Op's machine value value type (doesn't need to be passed, but
1035 * makes life easier.)
1036 * @return The SDNode with the entire instruction sequence
1039 SPUDAGToDAGISel::SelectSRLi64(SDValue &Op, EVT OpVT) {
1040 SDValue Op0 = Op.getOperand(0);
1041 EVT VecVT = EVT::getVectorVT(*CurDAG->getContext(),
1042 OpVT, (128 / OpVT.getSizeInBits()));
1043 SDValue ShiftAmt = Op.getOperand(1);
1044 EVT ShiftAmtVT = ShiftAmt.getValueType();
1045 SDNode *VecOp0, *Shift = 0;
1046 DebugLoc dl = Op.getDebugLoc();
1048 VecOp0 = CurDAG->getMachineNode(SPU::ORv2i64_i64, dl, VecVT, Op0);
1050 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(ShiftAmt)) {
1051 unsigned bytes = unsigned(CN->getZExtValue()) >> 3;
1052 unsigned bits = unsigned(CN->getZExtValue()) & 7;
1056 CurDAG->getMachineNode(SPU::ROTQMBYIv2i64, dl, VecVT,
1058 CurDAG->getTargetConstant(bytes, ShiftAmtVT));
1063 CurDAG->getMachineNode(SPU::ROTQMBIIv2i64, dl, VecVT,
1064 SDValue((Shift != 0 ? Shift : VecOp0), 0),
1065 CurDAG->getTargetConstant(bits, ShiftAmtVT));
1069 CurDAG->getMachineNode(SPU::ROTMIr32, dl, ShiftAmtVT,
1071 CurDAG->getTargetConstant(3, ShiftAmtVT));
1073 CurDAG->getMachineNode(SPU::ANDIr32, dl, ShiftAmtVT,
1075 CurDAG->getTargetConstant(7, ShiftAmtVT));
1077 // Ensure that the shift amounts are negated!
1078 Bytes = CurDAG->getMachineNode(SPU::SFIr32, dl, ShiftAmtVT,
1080 CurDAG->getTargetConstant(0, ShiftAmtVT));
1082 Bits = CurDAG->getMachineNode(SPU::SFIr32, dl, ShiftAmtVT,
1084 CurDAG->getTargetConstant(0, ShiftAmtVT));
1087 CurDAG->getMachineNode(SPU::ROTQMBYv2i64, dl, VecVT,
1088 SDValue(VecOp0, 0), SDValue(Bytes, 0));
1090 CurDAG->getMachineNode(SPU::ROTQMBIv2i64, dl, VecVT,
1091 SDValue(Shift, 0), SDValue(Bits, 0));
1094 return CurDAG->getMachineNode(SPU::ORi64_v2i64, dl, OpVT, SDValue(Shift, 0));
1098 * Emit the instruction sequence for i64 arithmetic right shifts.
1100 * @param Op The shl operand
1101 * @param OpVT Op's machine value value type (doesn't need to be passed, but
1102 * makes life easier.)
1103 * @return The SDNode with the entire instruction sequence
1106 SPUDAGToDAGISel::SelectSRAi64(SDValue &Op, EVT OpVT) {
1107 // Promote Op0 to vector
1108 EVT VecVT = EVT::getVectorVT(*CurDAG->getContext(),
1109 OpVT, (128 / OpVT.getSizeInBits()));
1110 SDValue ShiftAmt = Op.getOperand(1);
1111 EVT ShiftAmtVT = ShiftAmt.getValueType();
1112 DebugLoc dl = Op.getDebugLoc();
1115 CurDAG->getMachineNode(SPU::ORv2i64_i64, dl, VecVT, Op.getOperand(0));
1117 SDValue SignRotAmt = CurDAG->getTargetConstant(31, ShiftAmtVT);
1119 CurDAG->getMachineNode(SPU::ROTMAIv2i64_i32, dl, MVT::v2i64,
1120 SDValue(VecOp0, 0), SignRotAmt);
1121 SDNode *UpperHalfSign =
1122 CurDAG->getMachineNode(SPU::ORi32_v4i32, dl, MVT::i32, SDValue(SignRot, 0));
1124 SDNode *UpperHalfSignMask =
1125 CurDAG->getMachineNode(SPU::FSM64r32, dl, VecVT, SDValue(UpperHalfSign, 0));
1126 SDNode *UpperLowerMask =
1127 CurDAG->getMachineNode(SPU::FSMBIv2i64, dl, VecVT,
1128 CurDAG->getTargetConstant(0xff00ULL, MVT::i16));
1129 SDNode *UpperLowerSelect =
1130 CurDAG->getMachineNode(SPU::SELBv2i64, dl, VecVT,
1131 SDValue(UpperHalfSignMask, 0),
1133 SDValue(UpperLowerMask, 0));
1137 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(ShiftAmt)) {
1138 unsigned bytes = unsigned(CN->getZExtValue()) >> 3;
1139 unsigned bits = unsigned(CN->getZExtValue()) & 7;
1144 CurDAG->getMachineNode(SPU::ROTQBYIv2i64, dl, VecVT,
1145 SDValue(UpperLowerSelect, 0),
1146 CurDAG->getTargetConstant(bytes, ShiftAmtVT));
1152 CurDAG->getMachineNode(SPU::ROTQBIIv2i64, dl, VecVT,
1153 SDValue((Shift != 0 ? Shift : UpperLowerSelect), 0),
1154 CurDAG->getTargetConstant(bits, ShiftAmtVT));
1158 CurDAG->getMachineNode(SPU::SFIr32, dl, ShiftAmtVT,
1159 ShiftAmt, CurDAG->getTargetConstant(0, ShiftAmtVT));
1162 CurDAG->getMachineNode(SPU::ROTQBYBIv2i64_r32, dl, VecVT,
1163 SDValue(UpperLowerSelect, 0), SDValue(NegShift, 0));
1165 CurDAG->getMachineNode(SPU::ROTQBIv2i64, dl, VecVT,
1166 SDValue(Shift, 0), SDValue(NegShift, 0));
1169 return CurDAG->getMachineNode(SPU::ORi64_v2i64, dl, OpVT, SDValue(Shift, 0));
1173 Do the necessary magic necessary to load a i64 constant
1175 SDNode *SPUDAGToDAGISel::SelectI64Constant(SDValue& Op, EVT OpVT,
1177 ConstantSDNode *CN = cast<ConstantSDNode>(Op.getNode());
1178 return SelectI64Constant(CN->getZExtValue(), OpVT, dl);
1181 SDNode *SPUDAGToDAGISel::SelectI64Constant(uint64_t Value64, EVT OpVT,
1183 EVT OpVecVT = EVT::getVectorVT(*CurDAG->getContext(), OpVT, 2);
1185 SPU::LowerV2I64Splat(OpVecVT, *CurDAG, Value64, dl);
1187 // Here's where it gets interesting, because we have to parse out the
1188 // subtree handed back in i64vec:
1190 if (i64vec.getOpcode() == ISD::BIT_CONVERT) {
1191 // The degenerate case where the upper and lower bits in the splat are
1193 SDValue Op0 = i64vec.getOperand(0);
1195 ReplaceUses(i64vec, Op0);
1196 return CurDAG->getMachineNode(SPU::ORi64_v2i64, dl, OpVT,
1197 SDValue(emitBuildVector(Op0), 0));
1198 } else if (i64vec.getOpcode() == SPUISD::SHUFB) {
1199 SDValue lhs = i64vec.getOperand(0);
1200 SDValue rhs = i64vec.getOperand(1);
1201 SDValue shufmask = i64vec.getOperand(2);
1203 if (lhs.getOpcode() == ISD::BIT_CONVERT) {
1204 ReplaceUses(lhs, lhs.getOperand(0));
1205 lhs = lhs.getOperand(0);
1208 SDNode *lhsNode = (lhs.getNode()->isMachineOpcode()
1210 : emitBuildVector(lhs));
1212 if (rhs.getOpcode() == ISD::BIT_CONVERT) {
1213 ReplaceUses(rhs, rhs.getOperand(0));
1214 rhs = rhs.getOperand(0);
1217 SDNode *rhsNode = (rhs.getNode()->isMachineOpcode()
1219 : emitBuildVector(rhs));
1221 if (shufmask.getOpcode() == ISD::BIT_CONVERT) {
1222 ReplaceUses(shufmask, shufmask.getOperand(0));
1223 shufmask = shufmask.getOperand(0);
1226 SDNode *shufMaskNode = (shufmask.getNode()->isMachineOpcode()
1227 ? shufmask.getNode()
1228 : emitBuildVector(shufmask));
1231 Select(CurDAG->getNode(SPUISD::SHUFB, dl, OpVecVT,
1232 SDValue(lhsNode, 0), SDValue(rhsNode, 0),
1233 SDValue(shufMaskNode, 0)));
1235 return CurDAG->getMachineNode(SPU::ORi64_v2i64, dl, OpVT,
1236 SDValue(shufNode, 0));
1237 } else if (i64vec.getOpcode() == ISD::BUILD_VECTOR) {
1238 return CurDAG->getMachineNode(SPU::ORi64_v2i64, dl, OpVT,
1239 SDValue(emitBuildVector(i64vec), 0));
1241 llvm_report_error("SPUDAGToDAGISel::SelectI64Constant: Unhandled i64vec"
1246 /// createSPUISelDag - This pass converts a legalized DAG into a
1247 /// SPU-specific DAG, ready for instruction scheduling.
1249 FunctionPass *llvm::createSPUISelDag(SPUTargetMachine &TM) {
1250 return new SPUDAGToDAGISel(TM);