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"
22 #include "llvm/Support/ErrorHandling.h"
27 //! Predicate for an unconditional branch instruction
28 inline bool isUncondBranch(const MachineInstr *I) {
29 unsigned opc = I->getOpcode();
31 return (opc == SPU::BR
36 //! Predicate for a conditional branch instruction
37 inline bool isCondBranch(const MachineInstr *I) {
38 unsigned opc = I->getOpcode();
40 return (opc == SPU::BRNZr32
41 || opc == SPU::BRNZv4i32
43 || opc == SPU::BRZv4i32
44 || opc == SPU::BRHNZr16
45 || opc == SPU::BRHNZv8i16
46 || opc == SPU::BRHZr16
47 || opc == SPU::BRHZv8i16);
51 SPUInstrInfo::SPUInstrInfo(SPUTargetMachine &tm)
52 : TargetInstrInfoImpl(SPUInsts, sizeof(SPUInsts)/sizeof(SPUInsts[0])),
54 RI(*TM.getSubtargetImpl(), *this)
58 SPUInstrInfo::isMoveInstr(const MachineInstr& MI,
61 unsigned& SrcSR, unsigned& DstSR) const {
62 SrcSR = DstSR = 0; // No sub-registers.
64 switch (MI.getOpcode()) {
79 assert(MI.getNumOperands() == 3 &&
80 MI.getOperand(0).isReg() &&
81 MI.getOperand(1).isReg() &&
82 MI.getOperand(2).isImm() &&
83 "invalid SPU ORI/ORHI/ORBI/AHI/AI/SFI/SFHI instruction!");
84 if (MI.getOperand(2).getImm() == 0) {
85 sourceReg = MI.getOperand(1).getReg();
86 destReg = MI.getOperand(0).getReg();
91 assert(MI.getNumOperands() == 3 &&
92 "wrong number of operands to AIr32");
93 if (MI.getOperand(0).isReg() &&
94 MI.getOperand(1).isReg() &&
95 (MI.getOperand(2).isImm() &&
96 MI.getOperand(2).getImm() == 0)) {
97 sourceReg = MI.getOperand(1).getReg();
98 destReg = MI.getOperand(0).getReg();
115 case SPU::ORv16i8_i8:
116 case SPU::ORv8i16_i16:
117 case SPU::ORv4i32_i32:
118 case SPU::ORv2i64_i64:
119 case SPU::ORv4f32_f32:
120 case SPU::ORv2f64_f64:
121 case SPU::ORi8_v16i8:
122 case SPU::ORi16_v8i16:
123 case SPU::ORi32_v4i32:
124 case SPU::ORi64_v2i64:
125 case SPU::ORf32_v4f32:
126 case SPU::ORf64_v2f64:
128 case SPU::ORi128_r64:
129 case SPU::ORi128_f64:
130 case SPU::ORi128_r32:
131 case SPU::ORi128_f32:
132 case SPU::ORi128_r16:
135 case SPU::ORi128_vec:
137 case SPU::ORr64_i128:
138 case SPU::ORf64_i128:
139 case SPU::ORr32_i128:
140 case SPU::ORf32_i128:
141 case SPU::ORr16_i128:
144 case SPU::ORvec_i128:
162 case SPU::ORr64_f64: {
163 assert(MI.getNumOperands() == 2 &&
164 MI.getOperand(0).isReg() &&
165 MI.getOperand(1).isReg() &&
166 "invalid SPU OR<type>_<vec> or LR instruction!");
167 if (MI.getOperand(0).getReg() == MI.getOperand(1).getReg()) {
168 sourceReg = MI.getOperand(1).getReg();
169 destReg = MI.getOperand(0).getReg();
185 assert(MI.getNumOperands() == 3 &&
186 MI.getOperand(0).isReg() &&
187 MI.getOperand(1).isReg() &&
188 MI.getOperand(2).isReg() &&
189 "invalid SPU OR(vec|r32|r64|gprc) instruction!");
190 if (MI.getOperand(1).getReg() == MI.getOperand(2).getReg()) {
191 sourceReg = MI.getOperand(1).getReg();
192 destReg = MI.getOperand(0).getReg();
202 SPUInstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
203 int &FrameIndex) const {
204 switch (MI->getOpcode()) {
215 const MachineOperand MOp1 = MI->getOperand(1);
216 const MachineOperand MOp2 = MI->getOperand(2);
217 if (MOp1.isImm() && MOp2.isFI()) {
218 FrameIndex = MOp2.getIndex();
219 return MI->getOperand(0).getReg();
228 SPUInstrInfo::isStoreToStackSlot(const MachineInstr *MI,
229 int &FrameIndex) const {
230 switch (MI->getOpcode()) {
242 const MachineOperand MOp1 = MI->getOperand(1);
243 const MachineOperand MOp2 = MI->getOperand(2);
244 if (MOp1.isImm() && MOp2.isFI()) {
245 FrameIndex = MOp2.getIndex();
246 return MI->getOperand(0).getReg();
254 bool SPUInstrInfo::copyRegToReg(MachineBasicBlock &MBB,
255 MachineBasicBlock::iterator MI,
256 unsigned DestReg, unsigned SrcReg,
257 const TargetRegisterClass *DestRC,
258 const TargetRegisterClass *SrcRC) const
260 // We support cross register class moves for our aliases, such as R3 in any
261 // reg class to any other reg class containing R3. This is required because
262 // we instruction select bitconvert i64 -> f64 as a noop for example, so our
263 // types have no specific meaning.
265 DebugLoc DL = DebugLoc::getUnknownLoc();
266 if (MI != MBB.end()) DL = MI->getDebugLoc();
268 if (DestRC == SPU::R8CRegisterClass) {
269 BuildMI(MBB, MI, DL, get(SPU::LRr8), DestReg).addReg(SrcReg);
270 } else if (DestRC == SPU::R16CRegisterClass) {
271 BuildMI(MBB, MI, DL, get(SPU::LRr16), DestReg).addReg(SrcReg);
272 } else if (DestRC == SPU::R32CRegisterClass) {
273 BuildMI(MBB, MI, DL, get(SPU::LRr32), DestReg).addReg(SrcReg);
274 } else if (DestRC == SPU::R32FPRegisterClass) {
275 BuildMI(MBB, MI, DL, get(SPU::LRf32), DestReg).addReg(SrcReg);
276 } else if (DestRC == SPU::R64CRegisterClass) {
277 BuildMI(MBB, MI, DL, get(SPU::LRr64), DestReg).addReg(SrcReg);
278 } else if (DestRC == SPU::R64FPRegisterClass) {
279 BuildMI(MBB, MI, DL, get(SPU::LRf64), DestReg).addReg(SrcReg);
280 } else if (DestRC == SPU::GPRCRegisterClass) {
281 BuildMI(MBB, MI, DL, get(SPU::LRr128), DestReg).addReg(SrcReg);
282 } else if (DestRC == SPU::VECREGRegisterClass) {
283 BuildMI(MBB, MI, DL, get(SPU::LRv16i8), DestReg).addReg(SrcReg);
285 // Attempt to copy unknown/unsupported register class!
293 SPUInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
294 MachineBasicBlock::iterator MI,
295 unsigned SrcReg, bool isKill, int FrameIdx,
296 const TargetRegisterClass *RC) const
299 bool isValidFrameIdx = (FrameIdx < SPUFrameInfo::maxFrameOffset());
300 if (RC == SPU::GPRCRegisterClass) {
301 opc = (isValidFrameIdx ? SPU::STQDr128 : SPU::STQXr128);
302 } else if (RC == SPU::R64CRegisterClass) {
303 opc = (isValidFrameIdx ? SPU::STQDr64 : SPU::STQXr64);
304 } else if (RC == SPU::R64FPRegisterClass) {
305 opc = (isValidFrameIdx ? SPU::STQDr64 : SPU::STQXr64);
306 } else if (RC == SPU::R32CRegisterClass) {
307 opc = (isValidFrameIdx ? SPU::STQDr32 : SPU::STQXr32);
308 } else if (RC == SPU::R32FPRegisterClass) {
309 opc = (isValidFrameIdx ? SPU::STQDr32 : SPU::STQXr32);
310 } else if (RC == SPU::R16CRegisterClass) {
311 opc = (isValidFrameIdx ? SPU::STQDr16 : SPU::STQXr16);
312 } else if (RC == SPU::R8CRegisterClass) {
313 opc = (isValidFrameIdx ? SPU::STQDr8 : SPU::STQXr8);
314 } else if (RC == SPU::VECREGRegisterClass) {
315 opc = (isValidFrameIdx) ? SPU::STQDv16i8 : SPU::STQXv16i8;
317 llvm_unreachable("Unknown regclass!");
320 DebugLoc DL = DebugLoc::getUnknownLoc();
321 if (MI != MBB.end()) DL = MI->getDebugLoc();
322 addFrameReference(BuildMI(MBB, MI, DL, get(opc))
323 .addReg(SrcReg, getKillRegState(isKill)), FrameIdx);
326 void SPUInstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
328 SmallVectorImpl<MachineOperand> &Addr,
329 const TargetRegisterClass *RC,
330 SmallVectorImpl<MachineInstr*> &NewMIs) const {
331 llvm_report_error("storeRegToAddr() invoked!");
333 if (Addr[0].isFI()) {
334 /* do what storeRegToStackSlot does here */
337 if (RC == SPU::GPRCRegisterClass) {
338 /* Opc = PPC::STW; */
339 } else if (RC == SPU::R16CRegisterClass) {
340 /* Opc = PPC::STD; */
341 } else if (RC == SPU::R32CRegisterClass) {
342 /* Opc = PPC::STFD; */
343 } else if (RC == SPU::R32FPRegisterClass) {
344 /* Opc = PPC::STFD; */
345 } else if (RC == SPU::R64FPRegisterClass) {
346 /* Opc = PPC::STFS; */
347 } else if (RC == SPU::VECREGRegisterClass) {
348 /* Opc = PPC::STVX; */
350 llvm_unreachable("Unknown regclass!");
352 DebugLoc DL = DebugLoc::getUnknownLoc();
353 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc))
354 .addReg(SrcReg, getKillRegState(isKill));
355 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
356 MIB.addOperand(Addr[i]);
357 NewMIs.push_back(MIB);
362 SPUInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
363 MachineBasicBlock::iterator MI,
364 unsigned DestReg, int FrameIdx,
365 const TargetRegisterClass *RC) const
368 bool isValidFrameIdx = (FrameIdx < SPUFrameInfo::maxFrameOffset());
369 if (RC == SPU::GPRCRegisterClass) {
370 opc = (isValidFrameIdx ? SPU::LQDr128 : SPU::LQXr128);
371 } else if (RC == SPU::R64CRegisterClass) {
372 opc = (isValidFrameIdx ? SPU::LQDr64 : SPU::LQXr64);
373 } else if (RC == SPU::R64FPRegisterClass) {
374 opc = (isValidFrameIdx ? SPU::LQDr64 : SPU::LQXr64);
375 } else if (RC == SPU::R32CRegisterClass) {
376 opc = (isValidFrameIdx ? SPU::LQDr32 : SPU::LQXr32);
377 } else if (RC == SPU::R32FPRegisterClass) {
378 opc = (isValidFrameIdx ? SPU::LQDr32 : SPU::LQXr32);
379 } else if (RC == SPU::R16CRegisterClass) {
380 opc = (isValidFrameIdx ? SPU::LQDr16 : SPU::LQXr16);
381 } else if (RC == SPU::R8CRegisterClass) {
382 opc = (isValidFrameIdx ? SPU::LQDr8 : SPU::LQXr8);
383 } else if (RC == SPU::VECREGRegisterClass) {
384 opc = (isValidFrameIdx) ? SPU::LQDv16i8 : SPU::LQXv16i8;
386 llvm_unreachable("Unknown regclass in loadRegFromStackSlot!");
389 DebugLoc DL = DebugLoc::getUnknownLoc();
390 if (MI != MBB.end()) DL = MI->getDebugLoc();
391 addFrameReference(BuildMI(MBB, MI, DL, get(opc), DestReg), FrameIdx);
395 \note We are really pessimistic here about what kind of a load we're doing.
397 void SPUInstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
398 SmallVectorImpl<MachineOperand> &Addr,
399 const TargetRegisterClass *RC,
400 SmallVectorImpl<MachineInstr*> &NewMIs)
402 llvm_report_error("loadRegToAddr() invoked!");
404 if (Addr[0].isFI()) {
405 /* do what loadRegFromStackSlot does here... */
408 if (RC == SPU::R8CRegisterClass) {
409 /* do brilliance here */
410 } else if (RC == SPU::R16CRegisterClass) {
411 /* Opc = PPC::LWZ; */
412 } else if (RC == SPU::R32CRegisterClass) {
414 } else if (RC == SPU::R32FPRegisterClass) {
415 /* Opc = PPC::LFD; */
416 } else if (RC == SPU::R64FPRegisterClass) {
417 /* Opc = PPC::LFS; */
418 } else if (RC == SPU::VECREGRegisterClass) {
419 /* Opc = PPC::LVX; */
420 } else if (RC == SPU::GPRCRegisterClass) {
421 /* Opc = something else! */
423 llvm_unreachable("Unknown regclass!");
425 DebugLoc DL = DebugLoc::getUnknownLoc();
426 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), DestReg);
427 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
428 MIB.addOperand(Addr[i]);
429 NewMIs.push_back(MIB);
433 //! Return true if the specified load or store can be folded
435 SPUInstrInfo::canFoldMemoryOperand(const MachineInstr *MI,
436 const SmallVectorImpl<unsigned> &Ops) const {
437 if (Ops.size() != 1) return false;
439 // Make sure this is a reg-reg copy.
440 unsigned Opc = MI->getOpcode();
453 if (MI->getOperand(1).getReg() == MI->getOperand(2).getReg())
461 /// foldMemoryOperand - SPU, like PPC, can only fold spills into
462 /// copy instructions, turning them into load/store instructions.
464 SPUInstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
466 const SmallVectorImpl<unsigned> &Ops,
467 int FrameIndex) const
469 if (Ops.size() != 1) return 0;
471 unsigned OpNum = Ops[0];
472 unsigned Opc = MI->getOpcode();
473 MachineInstr *NewMI = 0;
486 if (OpNum == 0) { // move -> store
487 unsigned InReg = MI->getOperand(1).getReg();
488 bool isKill = MI->getOperand(1).isKill();
489 bool isUndef = MI->getOperand(1).isUndef();
490 if (FrameIndex < SPUFrameInfo::maxFrameOffset()) {
491 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(),
494 MIB.addReg(InReg, getKillRegState(isKill) | getUndefRegState(isUndef));
495 NewMI = addFrameReference(MIB, FrameIndex);
497 } else { // move -> load
498 unsigned OutReg = MI->getOperand(0).getReg();
499 bool isDead = MI->getOperand(0).isDead();
500 bool isUndef = MI->getOperand(0).isUndef();
501 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc));
503 MIB.addReg(OutReg, RegState::Define | getDeadRegState(isDead) |
504 getUndefRegState(isUndef));
505 Opc = (FrameIndex < SPUFrameInfo::maxFrameOffset())
506 ? SPU::STQDr32 : SPU::STQXr32;
507 NewMI = addFrameReference(MIB, FrameIndex);
517 \note This code was kiped from PPC. There may be more branch analysis for
518 CellSPU than what's currently done here.
521 SPUInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB,
522 MachineBasicBlock *&FBB,
523 SmallVectorImpl<MachineOperand> &Cond,
524 bool AllowModify) const {
525 // If the block has no terminators, it just falls into the block after it.
526 MachineBasicBlock::iterator I = MBB.end();
527 if (I == MBB.begin() || !isUnpredicatedTerminator(--I))
530 // Get the last instruction in the block.
531 MachineInstr *LastInst = I;
533 // If there is only one terminator instruction, process it.
534 if (I == MBB.begin() || !isUnpredicatedTerminator(--I)) {
535 if (isUncondBranch(LastInst)) {
536 TBB = LastInst->getOperand(0).getMBB();
538 } else if (isCondBranch(LastInst)) {
539 // Block ends with fall-through condbranch.
540 TBB = LastInst->getOperand(1).getMBB();
541 DEBUG(cerr << "Pushing LastInst: ");
542 DEBUG(LastInst->dump());
543 Cond.push_back(MachineOperand::CreateImm(LastInst->getOpcode()));
544 Cond.push_back(LastInst->getOperand(0));
547 // Otherwise, don't know what this is.
551 // Get the instruction before it if it's a terminator.
552 MachineInstr *SecondLastInst = I;
554 // If there are three terminators, we don't know what sort of block this is.
555 if (SecondLastInst && I != MBB.begin() &&
556 isUnpredicatedTerminator(--I))
559 // If the block ends with a conditional and unconditional branch, handle it.
560 if (isCondBranch(SecondLastInst) && isUncondBranch(LastInst)) {
561 TBB = SecondLastInst->getOperand(1).getMBB();
562 DEBUG(cerr << "Pushing SecondLastInst: ");
563 DEBUG(SecondLastInst->dump());
564 Cond.push_back(MachineOperand::CreateImm(SecondLastInst->getOpcode()));
565 Cond.push_back(SecondLastInst->getOperand(0));
566 FBB = LastInst->getOperand(0).getMBB();
570 // If the block ends with two unconditional branches, handle it. The second
571 // one is not executed, so remove it.
572 if (isUncondBranch(SecondLastInst) && isUncondBranch(LastInst)) {
573 TBB = SecondLastInst->getOperand(0).getMBB();
576 I->eraseFromParent();
580 // Otherwise, can't handle this.
585 SPUInstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
586 MachineBasicBlock::iterator I = MBB.end();
587 if (I == MBB.begin())
590 if (!isCondBranch(I) && !isUncondBranch(I))
593 // Remove the first branch.
594 DEBUG(cerr << "Removing branch: ");
596 I->eraseFromParent();
598 if (I == MBB.begin())
602 if (!(isCondBranch(I) || isUncondBranch(I)))
605 // Remove the second branch.
606 DEBUG(cerr << "Removing second branch: ");
608 I->eraseFromParent();
613 SPUInstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
614 MachineBasicBlock *FBB,
615 const SmallVectorImpl<MachineOperand> &Cond) const {
616 // FIXME this should probably have a DebugLoc argument
617 DebugLoc dl = DebugLoc::getUnknownLoc();
618 // Shouldn't be a fall through.
619 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
620 assert((Cond.size() == 2 || Cond.size() == 0) &&
621 "SPU branch conditions have two components!");
626 // Unconditional branch
627 MachineInstrBuilder MIB = BuildMI(&MBB, dl, get(SPU::BR));
630 DEBUG(cerr << "Inserted one-way uncond branch: ");
631 DEBUG((*MIB).dump());
633 // Conditional branch
634 MachineInstrBuilder MIB = BuildMI(&MBB, dl, get(Cond[0].getImm()));
635 MIB.addReg(Cond[1].getReg()).addMBB(TBB);
637 DEBUG(cerr << "Inserted one-way cond branch: ");
638 DEBUG((*MIB).dump());
642 MachineInstrBuilder MIB = BuildMI(&MBB, dl, get(Cond[0].getImm()));
643 MachineInstrBuilder MIB2 = BuildMI(&MBB, dl, get(SPU::BR));
645 // Two-way Conditional Branch.
646 MIB.addReg(Cond[1].getReg()).addMBB(TBB);
649 DEBUG(cerr << "Inserted conditional branch: ");
650 DEBUG((*MIB).dump());
651 DEBUG(cerr << "part 2: ");
652 DEBUG((*MIB2).dump());
658 SPUInstrInfo::BlockHasNoFallThrough(const MachineBasicBlock &MBB) const {
659 return (!MBB.empty() && isUncondBranch(&MBB.back()));
661 //! Reverses a branch's condition, returning false on success.
663 SPUInstrInfo::ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond)
665 // Pretty brainless way of inverting the condition, but it works, considering
666 // there are only two conditions...
668 unsigned Opc; //! The incoming opcode
669 unsigned RevCondOpc; //! The reversed condition opcode
671 { SPU::BRNZr32, SPU::BRZr32 },
672 { SPU::BRNZv4i32, SPU::BRZv4i32 },
673 { SPU::BRZr32, SPU::BRNZr32 },
674 { SPU::BRZv4i32, SPU::BRNZv4i32 },
675 { SPU::BRHNZr16, SPU::BRHZr16 },
676 { SPU::BRHNZv8i16, SPU::BRHZv8i16 },
677 { SPU::BRHZr16, SPU::BRHNZr16 },
678 { SPU::BRHZv8i16, SPU::BRHNZv8i16 }
681 unsigned Opc = unsigned(Cond[0].getImm());
682 // Pretty dull mapping between the two conditions that SPU can generate:
683 for (int i = sizeof(revconds)/sizeof(revconds[0]) - 1; i >= 0; --i) {
684 if (revconds[i].Opc == Opc) {
685 Cond[0].setImm(revconds[i].RevCondOpc);