1 //===---- IA64ISelDAGToDAG.cpp - IA64 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 IA64,
11 // converting a legalized dag to an IA64 dag.
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "ia64-codegen"
17 #include "IA64TargetMachine.h"
18 #include "IA64ISelLowering.h"
19 #include "llvm/CodeGen/MachineInstrBuilder.h"
20 #include "llvm/CodeGen/MachineFunction.h"
21 #include "llvm/CodeGen/SelectionDAG.h"
22 #include "llvm/CodeGen/SelectionDAGISel.h"
23 #include "llvm/Target/TargetOptions.h"
24 #include "llvm/Constants.h"
25 #include "llvm/GlobalValue.h"
26 #include "llvm/Intrinsics.h"
27 #include "llvm/Support/Compiler.h"
28 #include "llvm/Support/Debug.h"
29 #include "llvm/Support/MathExtras.h"
33 //===--------------------------------------------------------------------===//
34 /// IA64DAGToDAGISel - IA64 specific code to select IA64 machine
35 /// instructions for SelectionDAG operations.
37 class IA64DAGToDAGISel : public SelectionDAGISel {
38 unsigned GlobalBaseReg;
40 explicit IA64DAGToDAGISel(IA64TargetMachine &TM)
41 : SelectionDAGISel(*TM.getTargetLowering()) {}
43 virtual bool runOnFunction(Function &Fn) {
44 // Make sure we re-emit a set of the global base reg if necessary
46 return SelectionDAGISel::runOnFunction(Fn);
49 /// getI64Imm - Return a target constant with the specified value, of type
51 inline SDValue getI64Imm(uint64_t Imm) {
52 return CurDAG->getTargetConstant(Imm, MVT::i64);
55 /// getGlobalBaseReg - insert code into the entry mbb to materialize the PIC
56 /// base register. Return the virtual register that holds this value.
57 // SDValue getGlobalBaseReg(); TODO: hmm
59 // Select - Convert the specified operand from a target-independent to a
60 // target-specific node if it hasn't already been changed.
61 SDNode *Select(SDValue N);
63 SDNode *SelectIntImmediateExpr(SDValue LHS, SDValue RHS,
64 unsigned OCHi, unsigned OCLo,
65 bool IsArithmetic = false,
67 SDNode *SelectBitfieldInsert(SDNode *N);
69 /// SelectCC - Select a comparison of the specified values with the
70 /// specified condition code, returning the CR# of the expression.
71 SDValue SelectCC(SDValue LHS, SDValue RHS, ISD::CondCode CC);
73 /// SelectAddr - Given the specified address, return the two operands for a
74 /// load/store instruction, and return true if it should be an indexed [r+r]
76 bool SelectAddr(SDValue Addr, SDValue &Op1, SDValue &Op2);
78 /// InstructionSelect - This callback is invoked by
79 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
80 virtual void InstructionSelect();
82 virtual const char *getPassName() const {
83 return "IA64 (Itanium) DAG->DAG Instruction Selector";
86 // Include the pieces autogenerated from the target description.
87 #include "IA64GenDAGISel.inc"
90 SDNode *SelectDIV(SDValue Op);
94 /// InstructionSelect - This callback is invoked by
95 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
96 void IA64DAGToDAGISel::InstructionSelect() {
99 // Select target instructions for the DAG.
101 CurDAG->RemoveDeadNodes();
104 SDNode *IA64DAGToDAGISel::SelectDIV(SDValue Op) {
105 SDNode *N = Op.getNode();
106 SDValue Chain = N->getOperand(0);
107 SDValue Tmp1 = N->getOperand(0);
108 SDValue Tmp2 = N->getOperand(1);
109 AddToISelQueue(Chain);
111 AddToISelQueue(Tmp1);
112 AddToISelQueue(Tmp2);
116 if(Tmp1.getValueType().isFloatingPoint())
119 bool isModulus=false; // is it a division or a modulus?
122 switch(N->getOpcode()) {
124 case ISD::SDIV: isModulus=false; isSigned=true; break;
125 case ISD::UDIV: isModulus=false; isSigned=false; break;
127 case ISD::SREM: isModulus=true; isSigned=true; break;
128 case ISD::UREM: isModulus=true; isSigned=false; break;
131 // TODO: check for integer divides by powers of 2 (or other simple patterns?)
133 SDValue TmpPR, TmpPR2;
134 SDValue TmpF1, TmpF2, TmpF3, TmpF4, TmpF5, TmpF6, TmpF7, TmpF8;
135 SDValue TmpF9, TmpF10,TmpF11,TmpF12,TmpF13,TmpF14,TmpF15;
138 // we'll need copies of F0 and F1
139 SDValue F0 = CurDAG->getRegister(IA64::F0, MVT::f64);
140 SDValue F1 = CurDAG->getRegister(IA64::F1, MVT::f64);
142 // OK, emit some code:
145 // first, load the inputs into FP regs.
147 SDValue(CurDAG->getTargetNode(IA64::SETFSIG, MVT::f64, Tmp1), 0);
148 Chain = TmpF1.getValue(1);
150 SDValue(CurDAG->getTargetNode(IA64::SETFSIG, MVT::f64, Tmp2), 0);
151 Chain = TmpF2.getValue(1);
153 // next, convert the inputs to FP
156 SDValue(CurDAG->getTargetNode(IA64::FCVTXF, MVT::f64, TmpF1), 0);
157 Chain = TmpF3.getValue(1);
159 SDValue(CurDAG->getTargetNode(IA64::FCVTXF, MVT::f64, TmpF2), 0);
160 Chain = TmpF4.getValue(1);
161 } else { // is unsigned
163 SDValue(CurDAG->getTargetNode(IA64::FCVTXUFS1, MVT::f64, TmpF1), 0);
164 Chain = TmpF3.getValue(1);
166 SDValue(CurDAG->getTargetNode(IA64::FCVTXUFS1, MVT::f64, TmpF2), 0);
167 Chain = TmpF4.getValue(1);
170 } else { // this is an FP divide/remainder, so we 'leak' some temp
171 // regs and assign TmpF3=Tmp1, TmpF4=Tmp2
176 // we start by computing an approximate reciprocal (good to 9 bits?)
177 // note, this instruction writes _both_ TmpF5 (answer) and TmpPR (predicate)
179 TmpF5 = SDValue(CurDAG->getTargetNode(IA64::FRCPAS0, MVT::f64, MVT::i1,
182 TmpF5 = SDValue(CurDAG->getTargetNode(IA64::FRCPAS1, MVT::f64, MVT::i1,
185 TmpPR = TmpF5.getValue(1);
186 Chain = TmpF5.getValue(2);
189 if(isModulus) { // for remainders, it'll be handy to have
190 // copies of -input_b
191 minusB = SDValue(CurDAG->getTargetNode(IA64::SUB, MVT::i64,
192 CurDAG->getRegister(IA64::r0, MVT::i64), Tmp2), 0);
193 Chain = minusB.getValue(1);
196 SDValue TmpE0, TmpY1, TmpE1, TmpY2;
198 SDValue OpsE0[] = { TmpF4, TmpF5, F1, TmpPR };
199 TmpE0 = SDValue(CurDAG->getTargetNode(IA64::CFNMAS1, MVT::f64,
201 Chain = TmpE0.getValue(1);
202 SDValue OpsY1[] = { TmpF5, TmpE0, TmpF5, TmpPR };
203 TmpY1 = SDValue(CurDAG->getTargetNode(IA64::CFMAS1, MVT::f64,
205 Chain = TmpY1.getValue(1);
206 SDValue OpsE1[] = { TmpE0, TmpE0, F0, TmpPR };
207 TmpE1 = SDValue(CurDAG->getTargetNode(IA64::CFMAS1, MVT::f64,
209 Chain = TmpE1.getValue(1);
210 SDValue OpsY2[] = { TmpY1, TmpE1, TmpY1, TmpPR };
211 TmpY2 = SDValue(CurDAG->getTargetNode(IA64::CFMAS1, MVT::f64,
213 Chain = TmpY2.getValue(1);
215 if(isFP) { // if this is an FP divide, we finish up here and exit early
217 assert(0 && "Sorry, try another FORTRAN compiler.");
219 SDValue TmpE2, TmpY3, TmpQ0, TmpR0;
221 SDValue OpsE2[] = { TmpE1, TmpE1, F0, TmpPR };
222 TmpE2 = SDValue(CurDAG->getTargetNode(IA64::CFMAS1, MVT::f64,
224 Chain = TmpE2.getValue(1);
225 SDValue OpsY3[] = { TmpY2, TmpE2, TmpY2, TmpPR };
226 TmpY3 = SDValue(CurDAG->getTargetNode(IA64::CFMAS1, MVT::f64,
228 Chain = TmpY3.getValue(1);
229 SDValue OpsQ0[] = { Tmp1, TmpY3, F0, TmpPR };
231 SDValue(CurDAG->getTargetNode(IA64::CFMADS1, MVT::f64, // double prec!
233 Chain = TmpQ0.getValue(1);
234 SDValue OpsR0[] = { Tmp2, TmpQ0, Tmp1, TmpPR };
236 SDValue(CurDAG->getTargetNode(IA64::CFNMADS1, MVT::f64, // double prec!
238 Chain = TmpR0.getValue(1);
240 // we want Result to have the same target register as the frcpa, so
241 // we two-address hack it. See the comment "for this to work..." on
242 // page 48 of Intel application note #245415
243 SDValue Ops[] = { TmpF5, TmpY3, TmpR0, TmpQ0, TmpPR };
244 Result = CurDAG->getTargetNode(IA64::TCFMADS0, MVT::f64, // d.p. s0 rndg!
246 Chain = SDValue(Result, 1);
247 return Result; // XXX: early exit!
248 } else { // this is *not* an FP divide, so there's a bit left to do:
250 SDValue TmpQ2, TmpR2, TmpQ3, TmpQ;
252 SDValue OpsQ2[] = { TmpF3, TmpY2, F0, TmpPR };
253 TmpQ2 = SDValue(CurDAG->getTargetNode(IA64::CFMAS1, MVT::f64,
255 Chain = TmpQ2.getValue(1);
256 SDValue OpsR2[] = { TmpF4, TmpQ2, TmpF3, TmpPR };
257 TmpR2 = SDValue(CurDAG->getTargetNode(IA64::CFNMAS1, MVT::f64,
259 Chain = TmpR2.getValue(1);
261 // we want TmpQ3 to have the same target register as the frcpa? maybe we
262 // should two-address hack it. See the comment "for this to work..." on page
263 // 48 of Intel application note #245415
264 SDValue OpsQ3[] = { TmpF5, TmpR2, TmpY2, TmpQ2, TmpPR };
265 TmpQ3 = SDValue(CurDAG->getTargetNode(IA64::TCFMAS1, MVT::f64,
267 Chain = TmpQ3.getValue(1);
269 // STORY: without these two-address instructions (TCFMAS1 and TCFMADS0)
270 // the FPSWA won't be able to help out in the case of large/tiny
271 // arguments. Other fun bugs may also appear, e.g. 0/x = x, not 0.
274 TmpQ = SDValue(CurDAG->getTargetNode(IA64::FCVTFXTRUNCS1,
275 MVT::f64, TmpQ3), 0);
277 TmpQ = SDValue(CurDAG->getTargetNode(IA64::FCVTFXUTRUNCS1,
278 MVT::f64, TmpQ3), 0);
280 Chain = TmpQ.getValue(1);
284 SDValue(CurDAG->getTargetNode(IA64::SETFSIG, MVT::f64, minusB), 0);
285 Chain = FPminusB.getValue(1);
287 SDValue(CurDAG->getTargetNode(IA64::XMAL, MVT::f64,
288 TmpQ, FPminusB, TmpF1), 0);
289 Chain = Remainder.getValue(1);
290 Result = CurDAG->getTargetNode(IA64::GETFSIG, MVT::i64, Remainder);
291 Chain = SDValue(Result, 1);
292 } else { // just an integer divide
293 Result = CurDAG->getTargetNode(IA64::GETFSIG, MVT::i64, TmpQ);
294 Chain = SDValue(Result, 1);
298 } // wasn't an FP divide
301 // Select - Convert the specified operand from a target-independent to a
302 // target-specific node if it hasn't already been changed.
303 SDNode *IA64DAGToDAGISel::Select(SDValue Op) {
304 SDNode *N = Op.getNode();
305 if (N->isMachineOpcode())
306 return NULL; // Already selected.
308 switch (N->getOpcode()) {
311 case IA64ISD::BRCALL: { // XXX: this is also a hack!
312 SDValue Chain = N->getOperand(0);
313 SDValue InFlag; // Null incoming flag value.
315 AddToISelQueue(Chain);
316 if(N->getNumOperands()==3) { // we have an incoming chain, callee and flag
317 InFlag = N->getOperand(2);
318 AddToISelQueue(InFlag);
324 // if we can call directly, do so
325 if (GlobalAddressSDNode *GASD =
326 dyn_cast<GlobalAddressSDNode>(N->getOperand(1))) {
327 CallOpcode = IA64::BRCALL_IPREL_GA;
328 CallOperand = CurDAG->getTargetGlobalAddress(GASD->getGlobal(), MVT::i64);
329 } else if (isa<ExternalSymbolSDNode>(N->getOperand(1))) {
330 // FIXME: we currently NEED this case for correctness, to avoid
331 // "non-pic code with imm reloc.n against dynamic symbol" errors
332 CallOpcode = IA64::BRCALL_IPREL_ES;
333 CallOperand = N->getOperand(1);
335 // otherwise we need to load the function descriptor,
336 // load the branch target (function)'s entry point and GP,
337 // branch (call) then restore the GP
338 SDValue FnDescriptor = N->getOperand(1);
339 AddToISelQueue(FnDescriptor);
341 // load the branch target's entry point [mem] and
343 SDValue targetEntryPoint=
344 SDValue(CurDAG->getTargetNode(IA64::LD8, MVT::i64, MVT::Other,
345 FnDescriptor, CurDAG->getEntryNode()), 0);
346 Chain = targetEntryPoint.getValue(1);
347 SDValue targetGPAddr=
348 SDValue(CurDAG->getTargetNode(IA64::ADDS, MVT::i64,
350 CurDAG->getConstant(8, MVT::i64)), 0);
351 Chain = targetGPAddr.getValue(1);
353 SDValue(CurDAG->getTargetNode(IA64::LD8, MVT::i64,MVT::Other,
354 targetGPAddr, CurDAG->getEntryNode()), 0);
355 Chain = targetGP.getValue(1);
357 Chain = CurDAG->getCopyToReg(Chain, IA64::r1, targetGP, InFlag);
358 InFlag = Chain.getValue(1);
359 Chain = CurDAG->getCopyToReg(Chain, IA64::B6,
360 targetEntryPoint, InFlag); // FLAG these?
361 InFlag = Chain.getValue(1);
363 CallOperand = CurDAG->getRegister(IA64::B6, MVT::i64);
364 CallOpcode = IA64::BRCALL_INDIRECT;
367 // Finally, once everything is setup, emit the call itself
368 if (InFlag.getNode())
369 Chain = SDValue(CurDAG->getTargetNode(CallOpcode, MVT::Other, MVT::Flag,
370 CallOperand, InFlag), 0);
371 else // there might be no arguments
372 Chain = SDValue(CurDAG->getTargetNode(CallOpcode, MVT::Other, MVT::Flag,
373 CallOperand, Chain), 0);
374 InFlag = Chain.getValue(1);
376 std::vector<SDValue> CallResults;
378 CallResults.push_back(Chain);
379 CallResults.push_back(InFlag);
381 for (unsigned i = 0, e = CallResults.size(); i != e; ++i)
382 ReplaceUses(Op.getValue(i), CallResults[i]);
386 case IA64ISD::GETFD: {
387 SDValue Input = N->getOperand(0);
388 AddToISelQueue(Input);
389 return CurDAG->getTargetNode(IA64::GETFD, MVT::i64, Input);
397 return SelectDIV(Op);
399 case ISD::TargetConstantFP: {
400 SDValue Chain = CurDAG->getEntryNode(); // this is a constant, so..
403 ConstantFPSDNode* N2 = cast<ConstantFPSDNode>(N);
404 if (N2->getValueAPF().isPosZero()) {
405 V = CurDAG->getCopyFromReg(Chain, IA64::F0, MVT::f64);
406 } else if (N2->isExactlyValue(N2->getValueType(0) == MVT::f32 ?
407 APFloat(+1.0f) : APFloat(+1.0))) {
408 V = CurDAG->getCopyFromReg(Chain, IA64::F1, MVT::f64);
410 assert(0 && "Unexpected FP constant!");
412 ReplaceUses(SDValue(N, 0), V);
416 case ISD::FrameIndex: { // TODO: reduce creepyness
417 int FI = cast<FrameIndexSDNode>(N)->getIndex();
419 return CurDAG->SelectNodeTo(N, IA64::MOV, MVT::i64,
420 CurDAG->getTargetFrameIndex(FI, MVT::i64));
422 return CurDAG->getTargetNode(IA64::MOV, MVT::i64,
423 CurDAG->getTargetFrameIndex(FI, MVT::i64));
426 case ISD::ConstantPool: { // TODO: nuke the constant pool
427 // (ia64 doesn't need one)
428 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(N);
429 Constant *C = CP->getConstVal();
430 SDValue CPI = CurDAG->getTargetConstantPool(C, MVT::i64,
432 return CurDAG->getTargetNode(IA64::ADDL_GA, MVT::i64, // ?
433 CurDAG->getRegister(IA64::r1, MVT::i64), CPI);
436 case ISD::GlobalAddress: {
437 GlobalValue *GV = cast<GlobalAddressSDNode>(N)->getGlobal();
438 SDValue GA = CurDAG->getTargetGlobalAddress(GV, MVT::i64);
440 SDValue(CurDAG->getTargetNode(IA64::ADDL_GA, MVT::i64,
441 CurDAG->getRegister(IA64::r1,
443 return CurDAG->getTargetNode(IA64::LD8, MVT::i64, MVT::Other, Tmp,
444 CurDAG->getEntryNode());
448 case ISD::ExternalSymbol: {
449 SDValue EA = CurDAG->getTargetExternalSymbol(
450 cast<ExternalSymbolSDNode>(N)->getSymbol(),
452 SDValue Tmp = CurDAG->getTargetNode(IA64::ADDL_EA, MVT::i64,
453 CurDAG->getRegister(IA64::r1,
456 return CurDAG->getTargetNode(IA64::LD8, MVT::i64, Tmp);
460 case ISD::LOAD: { // FIXME: load -1, not 1, for bools?
461 LoadSDNode *LD = cast<LoadSDNode>(N);
462 SDValue Chain = LD->getChain();
463 SDValue Address = LD->getBasePtr();
464 AddToISelQueue(Chain);
465 AddToISelQueue(Address);
467 MVT TypeBeingLoaded = LD->getMemoryVT();
469 switch (TypeBeingLoaded.getSimpleVT()) {
474 assert(0 && "Cannot load this type!");
475 case MVT::i1: { // this is a bool
476 Opc = IA64::LD1; // first we load a byte, then compare for != 0
477 if(N->getValueType(0) == MVT::i1) { // XXX: early exit!
478 return CurDAG->SelectNodeTo(N, IA64::CMPNE, MVT::i1, MVT::Other,
479 SDValue(CurDAG->getTargetNode(Opc, MVT::i64, Address), 0),
480 CurDAG->getRegister(IA64::r0, MVT::i64),
483 /* otherwise, we want to load a bool into something bigger: LD1
484 will do that for us, so we just fall through */
486 case MVT::i8: Opc = IA64::LD1; break;
487 case MVT::i16: Opc = IA64::LD2; break;
488 case MVT::i32: Opc = IA64::LD4; break;
489 case MVT::i64: Opc = IA64::LD8; break;
491 case MVT::f32: Opc = IA64::LDF4; break;
492 case MVT::f64: Opc = IA64::LDF8; break;
495 // TODO: comment this
496 return CurDAG->SelectNodeTo(N, Opc, N->getValueType(0), MVT::Other,
501 StoreSDNode *ST = cast<StoreSDNode>(N);
502 SDValue Address = ST->getBasePtr();
503 SDValue Chain = ST->getChain();
504 AddToISelQueue(Address);
505 AddToISelQueue(Chain);
508 if (ISD::isNON_TRUNCStore(N)) {
509 switch (N->getOperand(1).getValueType().getSimpleVT()) {
510 default: assert(0 && "unknown type in store");
511 case MVT::i1: { // this is a bool
512 Opc = IA64::ST1; // we store either 0 or 1 as a byte
514 SDValue Initial = CurDAG->getCopyFromReg(Chain, IA64::r0, MVT::i64);
515 Chain = Initial.getValue(1);
516 // then load 1 into the same reg iff the predicate to store is 1
517 SDValue Tmp = ST->getValue();
520 SDValue(CurDAG->getTargetNode(IA64::TPCADDS, MVT::i64, Initial,
521 CurDAG->getTargetConstant(1,
524 return CurDAG->SelectNodeTo(N, Opc, MVT::Other, Address, Tmp, Chain);
526 case MVT::i64: Opc = IA64::ST8; break;
527 case MVT::f64: Opc = IA64::STF8; break;
529 } else { // Truncating store
530 switch(ST->getMemoryVT().getSimpleVT()) {
531 default: assert(0 && "unknown type in truncstore");
532 case MVT::i8: Opc = IA64::ST1; break;
533 case MVT::i16: Opc = IA64::ST2; break;
534 case MVT::i32: Opc = IA64::ST4; break;
535 case MVT::f32: Opc = IA64::STF4; break;
539 SDValue N1 = N->getOperand(1);
540 SDValue N2 = N->getOperand(2);
543 return CurDAG->SelectNodeTo(N, Opc, MVT::Other, N2, N1, Chain);
547 SDValue Chain = N->getOperand(0);
548 SDValue CC = N->getOperand(1);
549 AddToISelQueue(Chain);
551 MachineBasicBlock *Dest =
552 cast<BasicBlockSDNode>(N->getOperand(2))->getBasicBlock();
553 //FIXME - we do NOT need long branches all the time
554 return CurDAG->SelectNodeTo(N, IA64::BRLCOND_NOTCALL, MVT::Other, CC,
555 CurDAG->getBasicBlock(Dest), Chain);
558 case ISD::CALLSEQ_START:
559 case ISD::CALLSEQ_END: {
560 int64_t Amt = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
561 unsigned Opc = N->getOpcode() == ISD::CALLSEQ_START ?
562 IA64::ADJUSTCALLSTACKDOWN : IA64::ADJUSTCALLSTACKUP;
563 SDValue N0 = N->getOperand(0);
565 return CurDAG->SelectNodeTo(N, Opc, MVT::Other, getI64Imm(Amt), N0);
569 // FIXME: we don't need long branches all the time!
570 SDValue N0 = N->getOperand(0);
572 return CurDAG->SelectNodeTo(N, IA64::BRL_NOTCALL, MVT::Other,
573 N->getOperand(1), N0);
576 return SelectCode(Op);
580 /// createIA64DAGToDAGInstructionSelector - This pass converts a legalized DAG
581 /// into an IA64-specific DAG, ready for instruction scheduling.
584 *llvm::createIA64DAGToDAGInstructionSelector(IA64TargetMachine &TM) {
585 return new IA64DAGToDAGISel(TM);