1 //===- SPUInstrInfo.cpp - Cell SPU Instruction Information ----------------===//
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 contains the Cell SPU implementation of the TargetInstrInfo class.
12 //===----------------------------------------------------------------------===//
14 #include "SPURegisterNames.h"
15 #include "SPUInstrInfo.h"
16 #include "SPUInstrBuilder.h"
17 #include "SPUTargetMachine.h"
18 #include "SPUGenInstrInfo.inc"
19 #include "llvm/CodeGen/MachineInstrBuilder.h"
20 #include "llvm/Support/Streams.h"
21 #include "llvm/Support/Debug.h"
26 //! Predicate for an unconditional branch instruction
27 inline bool isUncondBranch(const MachineInstr *I) {
28 unsigned opc = I->getOpcode();
30 return (opc == SPU::BR
35 //! Predicate for a conditional branch instruction
36 inline bool isCondBranch(const MachineInstr *I) {
37 unsigned opc = I->getOpcode();
39 return (opc == SPU::BRNZr32
40 || opc == SPU::BRNZv4i32
42 || opc == SPU::BRZv4i32
43 || opc == SPU::BRHNZr16
44 || opc == SPU::BRHNZv8i16
45 || opc == SPU::BRHZr16
46 || opc == SPU::BRHZv8i16);
50 SPUInstrInfo::SPUInstrInfo(SPUTargetMachine &tm)
51 : TargetInstrInfoImpl(SPUInsts, sizeof(SPUInsts)/sizeof(SPUInsts[0])),
53 RI(*TM.getSubtargetImpl(), *this)
56 /// getPointerRegClass - Return the register class to use to hold pointers.
57 /// This is used for addressing modes.
58 const TargetRegisterClass *
59 SPUInstrInfo::getPointerRegClass() const
61 return &SPU::R32CRegClass;
65 SPUInstrInfo::isMoveInstr(const MachineInstr& MI,
68 unsigned& SrcSR, unsigned& DstSR) const {
69 SrcSR = DstSR = 0; // No sub-registers.
71 // Primarily, ORI and OR are generated by copyRegToReg. But, there are other
72 // cases where we can safely say that what's being done is really a move
73 // (see how PowerPC does this -- it's the model for this code too.)
74 switch (MI.getOpcode()) {
89 assert(MI.getNumOperands() == 3 &&
90 MI.getOperand(0).isReg() &&
91 MI.getOperand(1).isReg() &&
92 MI.getOperand(2).isImm() &&
93 "invalid SPU ORI/ORHI/ORBI/AHI/AI/SFI/SFHI instruction!");
94 if (MI.getOperand(2).getImm() == 0) {
95 sourceReg = MI.getOperand(1).getReg();
96 destReg = MI.getOperand(0).getReg();
101 assert(MI.getNumOperands() == 3 &&
102 "wrong number of operands to AIr32");
103 if (MI.getOperand(0).isReg() &&
104 MI.getOperand(1).isReg() &&
105 (MI.getOperand(2).isImm() &&
106 MI.getOperand(2).getImm() == 0)) {
107 sourceReg = MI.getOperand(1).getReg();
108 destReg = MI.getOperand(0).getReg();
125 case SPU::ORv16i8_i8:
126 case SPU::ORv8i16_i16:
127 case SPU::ORv4i32_i32:
128 case SPU::ORv2i64_i64:
129 case SPU::ORv4f32_f32:
130 case SPU::ORv2f64_f64:
131 case SPU::ORi8_v16i8:
132 case SPU::ORi16_v8i16:
133 case SPU::ORi32_v4i32:
134 case SPU::ORi64_v2i64:
135 case SPU::ORf32_v4f32:
136 case SPU::ORf64_v2f64:
138 case SPU::ORi128_r64:
139 case SPU::ORi128_f64:
140 case SPU::ORi128_r32:
141 case SPU::ORi128_f32:
142 case SPU::ORi128_r16:
144 case SPU::ORi128_vec:
145 case SPU::ORr64_i128:
146 case SPU::ORf64_i128:
147 case SPU::ORr32_i128:
148 case SPU::ORf32_i128:
149 case SPU::ORr16_i128:
151 case SPU::ORvec_i128:
168 case SPU::ORr64_f64: {
169 assert(MI.getNumOperands() == 2 &&
170 MI.getOperand(0).isReg() &&
171 MI.getOperand(1).isReg() &&
172 "invalid SPU OR<type>_<vec> or LR instruction!");
173 if (MI.getOperand(0).getReg() == MI.getOperand(1).getReg()) {
174 sourceReg = MI.getOperand(0).getReg();
175 destReg = MI.getOperand(0).getReg();
190 assert(MI.getNumOperands() == 3 &&
191 MI.getOperand(0).isReg() &&
192 MI.getOperand(1).isReg() &&
193 MI.getOperand(2).isReg() &&
194 "invalid SPU OR(vec|r32|r64|gprc) instruction!");
195 if (MI.getOperand(1).getReg() == MI.getOperand(2).getReg()) {
196 sourceReg = MI.getOperand(1).getReg();
197 destReg = MI.getOperand(0).getReg();
207 SPUInstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
208 int &FrameIndex) const {
209 switch (MI->getOpcode()) {
220 const MachineOperand MOp1 = MI->getOperand(1);
221 const MachineOperand MOp2 = MI->getOperand(2);
222 if (MOp1.isImm() && MOp2.isFI()) {
223 FrameIndex = MOp2.getIndex();
224 return MI->getOperand(0).getReg();
233 SPUInstrInfo::isStoreToStackSlot(const MachineInstr *MI,
234 int &FrameIndex) const {
235 switch (MI->getOpcode()) {
247 const MachineOperand MOp1 = MI->getOperand(1);
248 const MachineOperand MOp2 = MI->getOperand(2);
249 if (MOp1.isImm() && MOp2.isFI()) {
250 FrameIndex = MOp2.getIndex();
251 return MI->getOperand(0).getReg();
259 bool SPUInstrInfo::copyRegToReg(MachineBasicBlock &MBB,
260 MachineBasicBlock::iterator MI,
261 unsigned DestReg, unsigned SrcReg,
262 const TargetRegisterClass *DestRC,
263 const TargetRegisterClass *SrcRC) const
265 // We support cross register class moves for our aliases, such as R3 in any
266 // reg class to any other reg class containing R3. This is required because
267 // we instruction select bitconvert i64 -> f64 as a noop for example, so our
268 // types have no specific meaning.
270 if (DestRC == SPU::R8CRegisterClass) {
271 BuildMI(MBB, MI, get(SPU::LRr8), DestReg).addReg(SrcReg);
272 } else if (DestRC == SPU::R16CRegisterClass) {
273 BuildMI(MBB, MI, get(SPU::LRr16), DestReg).addReg(SrcReg);
274 } else if (DestRC == SPU::R32CRegisterClass) {
275 BuildMI(MBB, MI, get(SPU::LRr32), DestReg).addReg(SrcReg);
276 } else if (DestRC == SPU::R32FPRegisterClass) {
277 BuildMI(MBB, MI, get(SPU::LRf32), DestReg).addReg(SrcReg);
278 } else if (DestRC == SPU::R64CRegisterClass) {
279 BuildMI(MBB, MI, get(SPU::LRr64), DestReg).addReg(SrcReg);
280 } else if (DestRC == SPU::R64FPRegisterClass) {
281 BuildMI(MBB, MI, get(SPU::LRf64), DestReg).addReg(SrcReg);
282 } else if (DestRC == SPU::GPRCRegisterClass) {
283 BuildMI(MBB, MI, get(SPU::LRr128), DestReg).addReg(SrcReg);
284 } else if (DestRC == SPU::VECREGRegisterClass) {
285 BuildMI(MBB, MI, get(SPU::LRv16i8), DestReg).addReg(SrcReg);
287 // Attempt to copy unknown/unsupported register class!
295 SPUInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
296 MachineBasicBlock::iterator MI,
297 unsigned SrcReg, bool isKill, int FrameIdx,
298 const TargetRegisterClass *RC) const
301 bool isValidFrameIdx = (FrameIdx < SPUFrameInfo::maxFrameOffset());
302 if (RC == SPU::GPRCRegisterClass) {
303 opc = (isValidFrameIdx ? SPU::STQDr128 : SPU::STQXr128);
304 } else if (RC == SPU::R64CRegisterClass) {
305 opc = (isValidFrameIdx ? SPU::STQDr64 : SPU::STQXr64);
306 } else if (RC == SPU::R64FPRegisterClass) {
307 opc = (isValidFrameIdx ? SPU::STQDr64 : SPU::STQXr64);
308 } else if (RC == SPU::R32CRegisterClass) {
309 opc = (isValidFrameIdx ? SPU::STQDr32 : SPU::STQXr32);
310 } else if (RC == SPU::R32FPRegisterClass) {
311 opc = (isValidFrameIdx ? SPU::STQDr32 : SPU::STQXr32);
312 } else if (RC == SPU::R16CRegisterClass) {
313 opc = (isValidFrameIdx ? SPU::STQDr16 : SPU::STQXr16);
314 } else if (RC == SPU::R8CRegisterClass) {
315 opc = (isValidFrameIdx ? SPU::STQDr8 : SPU::STQXr8);
316 } else if (RC == SPU::VECREGRegisterClass) {
317 opc = (isValidFrameIdx) ? SPU::STQDv16i8 : SPU::STQXv16i8;
319 assert(0 && "Unknown regclass!");
323 addFrameReference(BuildMI(MBB, MI, get(opc))
324 .addReg(SrcReg, false, false, isKill), FrameIdx);
327 void SPUInstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
329 SmallVectorImpl<MachineOperand> &Addr,
330 const TargetRegisterClass *RC,
331 SmallVectorImpl<MachineInstr*> &NewMIs) const {
332 cerr << "storeRegToAddr() invoked!\n";
335 if (Addr[0].isFI()) {
336 /* do what storeRegToStackSlot does here */
339 if (RC == SPU::GPRCRegisterClass) {
340 /* Opc = PPC::STW; */
341 } else if (RC == SPU::R16CRegisterClass) {
342 /* Opc = PPC::STD; */
343 } else if (RC == SPU::R32CRegisterClass) {
344 /* Opc = PPC::STFD; */
345 } else if (RC == SPU::R32FPRegisterClass) {
346 /* Opc = PPC::STFD; */
347 } else if (RC == SPU::R64FPRegisterClass) {
348 /* Opc = PPC::STFS; */
349 } else if (RC == SPU::VECREGRegisterClass) {
350 /* Opc = PPC::STVX; */
352 assert(0 && "Unknown regclass!");
355 MachineInstrBuilder MIB = BuildMI(MF, get(Opc))
356 .addReg(SrcReg, false, false, isKill);
357 for (unsigned i = 0, e = Addr.size(); i != e; ++i) {
358 MachineOperand &MO = Addr[i];
360 MIB.addReg(MO.getReg());
362 MIB.addImm(MO.getImm());
364 MIB.addFrameIndex(MO.getIndex());
366 NewMIs.push_back(MIB);
371 SPUInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
372 MachineBasicBlock::iterator MI,
373 unsigned DestReg, int FrameIdx,
374 const TargetRegisterClass *RC) const
377 bool isValidFrameIdx = (FrameIdx < SPUFrameInfo::maxFrameOffset());
378 if (RC == SPU::GPRCRegisterClass) {
379 opc = (isValidFrameIdx ? SPU::LQDr128 : SPU::LQXr128);
380 } else if (RC == SPU::R64CRegisterClass) {
381 opc = (isValidFrameIdx ? SPU::LQDr64 : SPU::LQXr64);
382 } else if (RC == SPU::R64FPRegisterClass) {
383 opc = (isValidFrameIdx ? SPU::LQDr64 : SPU::LQXr64);
384 } else if (RC == SPU::R32CRegisterClass) {
385 opc = (isValidFrameIdx ? SPU::LQDr32 : SPU::LQXr32);
386 } else if (RC == SPU::R32FPRegisterClass) {
387 opc = (isValidFrameIdx ? SPU::LQDr32 : SPU::LQXr32);
388 } else if (RC == SPU::R16CRegisterClass) {
389 opc = (isValidFrameIdx ? SPU::LQDr16 : SPU::LQXr16);
390 } else if (RC == SPU::R8CRegisterClass) {
391 opc = (isValidFrameIdx ? SPU::LQDr8 : SPU::LQXr8);
392 } else if (RC == SPU::VECREGRegisterClass) {
393 opc = (isValidFrameIdx) ? SPU::LQDv16i8 : SPU::LQXv16i8;
395 assert(0 && "Unknown regclass in loadRegFromStackSlot!");
399 addFrameReference(BuildMI(MBB, MI, get(opc)).addReg(DestReg), FrameIdx);
403 \note We are really pessimistic here about what kind of a load we're doing.
405 void SPUInstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
406 SmallVectorImpl<MachineOperand> &Addr,
407 const TargetRegisterClass *RC,
408 SmallVectorImpl<MachineInstr*> &NewMIs)
410 cerr << "loadRegToAddr() invoked!\n";
413 if (Addr[0].isFI()) {
414 /* do what loadRegFromStackSlot does here... */
417 if (RC == SPU::R8CRegisterClass) {
418 /* do brilliance here */
419 } else if (RC == SPU::R16CRegisterClass) {
420 /* Opc = PPC::LWZ; */
421 } else if (RC == SPU::R32CRegisterClass) {
423 } else if (RC == SPU::R32FPRegisterClass) {
424 /* Opc = PPC::LFD; */
425 } else if (RC == SPU::R64FPRegisterClass) {
426 /* Opc = PPC::LFS; */
427 } else if (RC == SPU::VECREGRegisterClass) {
428 /* Opc = PPC::LVX; */
429 } else if (RC == SPU::GPRCRegisterClass) {
430 /* Opc = something else! */
432 assert(0 && "Unknown regclass!");
435 MachineInstrBuilder MIB = BuildMI(MF, get(Opc), DestReg);
436 for (unsigned i = 0, e = Addr.size(); i != e; ++i) {
437 MachineOperand &MO = Addr[i];
439 MIB.addReg(MO.getReg());
441 MIB.addImm(MO.getImm());
443 MIB.addFrameIndex(MO.getIndex());
445 NewMIs.push_back(MIB);
449 //! Return true if the specified load or store can be folded
451 SPUInstrInfo::canFoldMemoryOperand(const MachineInstr *MI,
452 const SmallVectorImpl<unsigned> &Ops) const {
453 if (Ops.size() != 1) return false;
455 // Make sure this is a reg-reg copy.
456 unsigned Opc = MI->getOpcode();
469 if (MI->getOperand(1).getReg() == MI->getOperand(2).getReg())
477 /// foldMemoryOperand - SPU, like PPC, can only fold spills into
478 /// copy instructions, turning them into load/store instructions.
480 SPUInstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
482 const SmallVectorImpl<unsigned> &Ops,
483 int FrameIndex) const
485 if (Ops.size() != 1) return 0;
487 unsigned OpNum = Ops[0];
488 unsigned Opc = MI->getOpcode();
489 MachineInstr *NewMI = 0;
502 if (OpNum == 0) { // move -> store
503 unsigned InReg = MI->getOperand(1).getReg();
504 bool isKill = MI->getOperand(1).isKill();
505 if (FrameIndex < SPUFrameInfo::maxFrameOffset()) {
506 MachineInstrBuilder MIB = BuildMI(MF, get(SPU::STQDr32));
508 MIB.addReg(InReg, false, false, isKill);
509 NewMI = addFrameReference(MIB, FrameIndex);
511 } else { // move -> load
512 unsigned OutReg = MI->getOperand(0).getReg();
513 bool isDead = MI->getOperand(0).isDead();
514 MachineInstrBuilder MIB = BuildMI(MF, get(Opc));
516 MIB.addReg(OutReg, true, false, false, isDead);
517 Opc = (FrameIndex < SPUFrameInfo::maxFrameOffset())
518 ? SPU::STQDr32 : SPU::STQXr32;
519 NewMI = addFrameReference(MIB, FrameIndex);
529 \note This code was kiped from PPC. There may be more branch analysis for
530 CellSPU than what's currently done here.
533 SPUInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB,
534 MachineBasicBlock *&FBB,
535 SmallVectorImpl<MachineOperand> &Cond) const {
536 // If the block has no terminators, it just falls into the block after it.
537 MachineBasicBlock::iterator I = MBB.end();
538 if (I == MBB.begin() || !isUnpredicatedTerminator(--I))
541 // Get the last instruction in the block.
542 MachineInstr *LastInst = I;
544 // If there is only one terminator instruction, process it.
545 if (I == MBB.begin() || !isUnpredicatedTerminator(--I)) {
546 if (isUncondBranch(LastInst)) {
547 TBB = LastInst->getOperand(0).getMBB();
549 } else if (isCondBranch(LastInst)) {
550 // Block ends with fall-through condbranch.
551 TBB = LastInst->getOperand(1).getMBB();
552 DEBUG(cerr << "Pushing LastInst: ");
553 DEBUG(LastInst->dump());
554 Cond.push_back(MachineOperand::CreateImm(LastInst->getOpcode()));
555 Cond.push_back(LastInst->getOperand(0));
558 // Otherwise, don't know what this is.
562 // Get the instruction before it if it's a terminator.
563 MachineInstr *SecondLastInst = I;
565 // If there are three terminators, we don't know what sort of block this is.
566 if (SecondLastInst && I != MBB.begin() &&
567 isUnpredicatedTerminator(--I))
570 // If the block ends with a conditional and unconditional branch, handle it.
571 if (isCondBranch(SecondLastInst) && isUncondBranch(LastInst)) {
572 TBB = SecondLastInst->getOperand(1).getMBB();
573 DEBUG(cerr << "Pushing SecondLastInst: ");
574 DEBUG(SecondLastInst->dump());
575 Cond.push_back(MachineOperand::CreateImm(SecondLastInst->getOpcode()));
576 Cond.push_back(SecondLastInst->getOperand(0));
577 FBB = LastInst->getOperand(0).getMBB();
581 // If the block ends with two unconditional branches, handle it. The second
582 // one is not executed, so remove it.
583 if (isUncondBranch(SecondLastInst) && isUncondBranch(LastInst)) {
584 TBB = SecondLastInst->getOperand(0).getMBB();
586 I->eraseFromParent();
590 // Otherwise, can't handle this.
595 SPUInstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
596 MachineBasicBlock::iterator I = MBB.end();
597 if (I == MBB.begin())
600 if (!isCondBranch(I) && !isUncondBranch(I))
603 // Remove the first branch.
604 DEBUG(cerr << "Removing branch: ");
606 I->eraseFromParent();
608 if (I == MBB.begin())
612 if (!(isCondBranch(I) || isUncondBranch(I)))
615 // Remove the second branch.
616 DEBUG(cerr << "Removing second branch: ");
618 I->eraseFromParent();
623 SPUInstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
624 MachineBasicBlock *FBB,
625 const SmallVectorImpl<MachineOperand> &Cond) const {
626 // Shouldn't be a fall through.
627 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
628 assert((Cond.size() == 2 || Cond.size() == 0) &&
629 "SPU branch conditions have two components!");
634 // Unconditional branch
635 MachineInstrBuilder MIB = BuildMI(&MBB, get(SPU::BR));
638 DEBUG(cerr << "Inserted one-way uncond branch: ");
639 DEBUG((*MIB).dump());
641 // Conditional branch
642 MachineInstrBuilder MIB = BuildMI(&MBB, get(Cond[0].getImm()));
643 MIB.addReg(Cond[1].getReg()).addMBB(TBB);
645 DEBUG(cerr << "Inserted one-way cond branch: ");
646 DEBUG((*MIB).dump());
650 MachineInstrBuilder MIB = BuildMI(&MBB, get(Cond[0].getImm()));
651 MachineInstrBuilder MIB2 = BuildMI(&MBB, get(SPU::BR));
653 // Two-way Conditional Branch.
654 MIB.addReg(Cond[1].getReg()).addMBB(TBB);
657 DEBUG(cerr << "Inserted conditional branch: ");
658 DEBUG((*MIB).dump());
659 DEBUG(cerr << "part 2: ");
660 DEBUG((*MIB2).dump());
666 SPUInstrInfo::BlockHasNoFallThrough(const MachineBasicBlock &MBB) const {
667 return (!MBB.empty() && isUncondBranch(&MBB.back()));
669 //! Reverses a branch's condition, returning false on success.
671 SPUInstrInfo::ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond)
673 // Pretty brainless way of inverting the condition, but it works, considering
674 // there are only two conditions...
676 unsigned Opc; //! The incoming opcode
677 unsigned RevCondOpc; //! The reversed condition opcode
679 { SPU::BRNZr32, SPU::BRZr32 },
680 { SPU::BRNZv4i32, SPU::BRZv4i32 },
681 { SPU::BRZr32, SPU::BRNZr32 },
682 { SPU::BRZv4i32, SPU::BRNZv4i32 },
683 { SPU::BRHNZr16, SPU::BRHZr16 },
684 { SPU::BRHNZv8i16, SPU::BRHZv8i16 },
685 { SPU::BRHZr16, SPU::BRHNZr16 },
686 { SPU::BRHZv8i16, SPU::BRHNZv8i16 }
689 unsigned Opc = unsigned(Cond[0].getImm());
690 // Pretty dull mapping between the two conditions that SPU can generate:
691 for (int i = sizeof(revconds)/sizeof(revconds[0]) - 1; i >= 0; --i) {
692 if (revconds[i].Opc == Opc) {
693 Cond[0].setImm(revconds[i].RevCondOpc);