1 //===---- IA64ISelDAGToDAG.cpp - IA64 pattern matching inst selector ------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Duraid Madina and is distributed under
6 // the University of Illinois Open Source 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 //===----------------------------------------------------------------------===//
16 #include "IA64TargetMachine.h"
17 #include "IA64ISelLowering.h"
18 #include "llvm/CodeGen/MachineInstrBuilder.h"
19 #include "llvm/CodeGen/MachineFunction.h"
20 #include "llvm/CodeGen/SSARegMap.h"
21 #include "llvm/CodeGen/SelectionDAG.h"
22 #include "llvm/CodeGen/SelectionDAGISel.h"
23 #include "llvm/Target/TargetOptions.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/Constants.h"
26 #include "llvm/GlobalValue.h"
27 #include "llvm/Intrinsics.h"
28 #include "llvm/Support/Debug.h"
29 #include "llvm/Support/MathExtras.h"
35 Statistic FusedFP ("ia64-codegen", "Number of fused fp operations");
36 Statistic FrameOff("ia64-codegen", "Number of frame idx offsets collapsed");
38 //===--------------------------------------------------------------------===//
39 /// IA64DAGToDAGISel - IA64 specific code to select IA64 machine
40 /// instructions for SelectionDAG operations.
42 class IA64DAGToDAGISel : public SelectionDAGISel {
43 IA64TargetLowering IA64Lowering;
44 unsigned GlobalBaseReg;
46 IA64DAGToDAGISel(IA64TargetMachine &TM)
47 : SelectionDAGISel(IA64Lowering), IA64Lowering(*TM.getTargetLowering()) {}
49 virtual bool runOnFunction(Function &Fn) {
50 // Make sure we re-emit a set of the global base reg if necessary
52 return SelectionDAGISel::runOnFunction(Fn);
55 /// getI64Imm - Return a target constant with the specified value, of type
57 inline SDOperand getI64Imm(uint64_t Imm) {
58 return CurDAG->getTargetConstant(Imm, MVT::i64);
61 /// getGlobalBaseReg - insert code into the entry mbb to materialize the PIC
62 /// base register. Return the virtual register that holds this value.
63 // SDOperand getGlobalBaseReg(); TODO: hmm
65 // Select - Convert the specified operand from a target-independent to a
66 // target-specific node if it hasn't already been changed.
67 SDNode *Select(SDOperand N);
69 SDNode *SelectIntImmediateExpr(SDOperand LHS, SDOperand RHS,
70 unsigned OCHi, unsigned OCLo,
71 bool IsArithmetic = false,
73 SDNode *SelectBitfieldInsert(SDNode *N);
75 /// SelectCC - Select a comparison of the specified values with the
76 /// specified condition code, returning the CR# of the expression.
77 SDOperand SelectCC(SDOperand LHS, SDOperand RHS, ISD::CondCode CC);
79 /// SelectAddr - Given the specified address, return the two operands for a
80 /// load/store instruction, and return true if it should be an indexed [r+r]
82 bool SelectAddr(SDOperand Addr, SDOperand &Op1, SDOperand &Op2);
84 /// InstructionSelectBasicBlock - This callback is invoked by
85 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
86 virtual void InstructionSelectBasicBlock(SelectionDAG &DAG);
88 virtual const char *getPassName() const {
89 return "IA64 (Itanium) DAG->DAG Instruction Selector";
92 // Include the pieces autogenerated from the target description.
93 #include "IA64GenDAGISel.inc"
96 SDNode *SelectDIV(SDOperand Op);
100 /// InstructionSelectBasicBlock - This callback is invoked by
101 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
102 void IA64DAGToDAGISel::InstructionSelectBasicBlock(SelectionDAG &DAG) {
105 // Select target instructions for the DAG.
106 DAG.setRoot(SelectRoot(DAG.getRoot()));
107 DAG.RemoveDeadNodes();
109 // Emit machine code to BB.
110 ScheduleAndEmitDAG(DAG);
113 SDNode *IA64DAGToDAGISel::SelectDIV(SDOperand Op) {
115 SDOperand Chain = N->getOperand(0);
116 SDOperand Tmp1 = N->getOperand(0);
117 SDOperand Tmp2 = N->getOperand(1);
118 AddToISelQueue(Chain);
120 AddToISelQueue(Tmp1);
121 AddToISelQueue(Tmp2);
125 if(MVT::isFloatingPoint(Tmp1.getValueType()))
128 bool isModulus=false; // is it a division or a modulus?
131 switch(N->getOpcode()) {
133 case ISD::SDIV: isModulus=false; isSigned=true; break;
134 case ISD::UDIV: isModulus=false; isSigned=false; break;
136 case ISD::SREM: isModulus=true; isSigned=true; break;
137 case ISD::UREM: isModulus=true; isSigned=false; break;
140 // TODO: check for integer divides by powers of 2 (or other simple patterns?)
142 SDOperand TmpPR, TmpPR2;
143 SDOperand TmpF1, TmpF2, TmpF3, TmpF4, TmpF5, TmpF6, TmpF7, TmpF8;
144 SDOperand TmpF9, TmpF10,TmpF11,TmpF12,TmpF13,TmpF14,TmpF15;
147 // we'll need copies of F0 and F1
148 SDOperand F0 = CurDAG->getRegister(IA64::F0, MVT::f64);
149 SDOperand F1 = CurDAG->getRegister(IA64::F1, MVT::f64);
151 // OK, emit some code:
154 // first, load the inputs into FP regs.
156 SDOperand(CurDAG->getTargetNode(IA64::SETFSIG, MVT::f64, Tmp1), 0);
157 Chain = TmpF1.getValue(1);
159 SDOperand(CurDAG->getTargetNode(IA64::SETFSIG, MVT::f64, Tmp2), 0);
160 Chain = TmpF2.getValue(1);
162 // next, convert the inputs to FP
165 SDOperand(CurDAG->getTargetNode(IA64::FCVTXF, MVT::f64, TmpF1), 0);
166 Chain = TmpF3.getValue(1);
168 SDOperand(CurDAG->getTargetNode(IA64::FCVTXF, MVT::f64, TmpF2), 0);
169 Chain = TmpF4.getValue(1);
170 } else { // is unsigned
172 SDOperand(CurDAG->getTargetNode(IA64::FCVTXUFS1, MVT::f64, TmpF1), 0);
173 Chain = TmpF3.getValue(1);
175 SDOperand(CurDAG->getTargetNode(IA64::FCVTXUFS1, MVT::f64, TmpF2), 0);
176 Chain = TmpF4.getValue(1);
179 } else { // this is an FP divide/remainder, so we 'leak' some temp
180 // regs and assign TmpF3=Tmp1, TmpF4=Tmp2
185 // we start by computing an approximate reciprocal (good to 9 bits?)
186 // note, this instruction writes _both_ TmpF5 (answer) and TmpPR (predicate)
188 TmpF5 = SDOperand(CurDAG->getTargetNode(IA64::FRCPAS0, MVT::f64, MVT::i1,
191 TmpF5 = SDOperand(CurDAG->getTargetNode(IA64::FRCPAS1, MVT::f64, MVT::i1,
194 TmpPR = TmpF5.getValue(1);
195 Chain = TmpF5.getValue(2);
198 if(isModulus) { // for remainders, it'll be handy to have
199 // copies of -input_b
200 minusB = SDOperand(CurDAG->getTargetNode(IA64::SUB, MVT::i64,
201 CurDAG->getRegister(IA64::r0, MVT::i64), Tmp2), 0);
202 Chain = minusB.getValue(1);
205 SDOperand TmpE0, TmpY1, TmpE1, TmpY2;
207 SDOperand OpsE0[] = { TmpF4, TmpF5, F1, TmpPR };
208 TmpE0 = SDOperand(CurDAG->getTargetNode(IA64::CFNMAS1, MVT::f64,
210 Chain = TmpE0.getValue(1);
211 SDOperand OpsY1[] = { TmpF5, TmpE0, TmpF5, TmpPR };
212 TmpY1 = SDOperand(CurDAG->getTargetNode(IA64::CFMAS1, MVT::f64,
214 Chain = TmpY1.getValue(1);
215 SDOperand OpsE1[] = { TmpE0, TmpE0, F0, TmpPR };
216 TmpE1 = SDOperand(CurDAG->getTargetNode(IA64::CFMAS1, MVT::f64,
218 Chain = TmpE1.getValue(1);
219 SDOperand OpsY2[] = { TmpY1, TmpE1, TmpY1, TmpPR };
220 TmpY2 = SDOperand(CurDAG->getTargetNode(IA64::CFMAS1, MVT::f64,
222 Chain = TmpY2.getValue(1);
224 if(isFP) { // if this is an FP divide, we finish up here and exit early
226 assert(0 && "Sorry, try another FORTRAN compiler.");
228 SDOperand TmpE2, TmpY3, TmpQ0, TmpR0;
230 SDOperand OpsE2[] = { TmpE1, TmpE1, F0, TmpPR };
231 TmpE2 = SDOperand(CurDAG->getTargetNode(IA64::CFMAS1, MVT::f64,
233 Chain = TmpE2.getValue(1);
234 SDOperand OpsY3[] = { TmpY2, TmpE2, TmpY2, TmpPR };
235 TmpY3 = SDOperand(CurDAG->getTargetNode(IA64::CFMAS1, MVT::f64,
237 Chain = TmpY3.getValue(1);
238 SDOperand OpsQ0[] = { Tmp1, TmpY3, F0, TmpPR };
240 SDOperand(CurDAG->getTargetNode(IA64::CFMADS1, MVT::f64, // double prec!
242 Chain = TmpQ0.getValue(1);
243 SDOperand OpsR0[] = { Tmp2, TmpQ0, Tmp1, TmpPR };
245 SDOperand(CurDAG->getTargetNode(IA64::CFNMADS1, MVT::f64, // double prec!
247 Chain = TmpR0.getValue(1);
249 // we want Result to have the same target register as the frcpa, so
250 // we two-address hack it. See the comment "for this to work..." on
251 // page 48 of Intel application note #245415
252 SDOperand Ops[] = { TmpF5, TmpY3, TmpR0, TmpQ0, TmpPR };
253 Result = CurDAG->getTargetNode(IA64::TCFMADS0, MVT::f64, // d.p. s0 rndg!
255 Chain = SDOperand(Result, 1);
256 return Result; // XXX: early exit!
257 } else { // this is *not* an FP divide, so there's a bit left to do:
259 SDOperand TmpQ2, TmpR2, TmpQ3, TmpQ;
261 SDOperand OpsQ2[] = { TmpF3, TmpY2, F0, TmpPR };
262 TmpQ2 = SDOperand(CurDAG->getTargetNode(IA64::CFMAS1, MVT::f64,
264 Chain = TmpQ2.getValue(1);
265 SDOperand OpsR2[] = { TmpF4, TmpQ2, TmpF3, TmpPR };
266 TmpR2 = SDOperand(CurDAG->getTargetNode(IA64::CFNMAS1, MVT::f64,
268 Chain = TmpR2.getValue(1);
270 // we want TmpQ3 to have the same target register as the frcpa? maybe we
271 // should two-address hack it. See the comment "for this to work..." on page
272 // 48 of Intel application note #245415
273 SDOperand OpsQ3[] = { TmpF5, TmpR2, TmpY2, TmpQ2, TmpPR };
274 TmpQ3 = SDOperand(CurDAG->getTargetNode(IA64::TCFMAS1, MVT::f64,
276 Chain = TmpQ3.getValue(1);
278 // STORY: without these two-address instructions (TCFMAS1 and TCFMADS0)
279 // the FPSWA won't be able to help out in the case of large/tiny
280 // arguments. Other fun bugs may also appear, e.g. 0/x = x, not 0.
283 TmpQ = SDOperand(CurDAG->getTargetNode(IA64::FCVTFXTRUNCS1,
284 MVT::f64, TmpQ3), 0);
286 TmpQ = SDOperand(CurDAG->getTargetNode(IA64::FCVTFXUTRUNCS1,
287 MVT::f64, TmpQ3), 0);
289 Chain = TmpQ.getValue(1);
293 SDOperand(CurDAG->getTargetNode(IA64::SETFSIG, MVT::f64, minusB), 0);
294 Chain = FPminusB.getValue(1);
295 SDOperand Remainder =
296 SDOperand(CurDAG->getTargetNode(IA64::XMAL, MVT::f64,
297 TmpQ, FPminusB, TmpF1), 0);
298 Chain = Remainder.getValue(1);
299 Result = CurDAG->getTargetNode(IA64::GETFSIG, MVT::i64, Remainder);
300 Chain = SDOperand(Result, 1);
301 } else { // just an integer divide
302 Result = CurDAG->getTargetNode(IA64::GETFSIG, MVT::i64, TmpQ);
303 Chain = SDOperand(Result, 1);
307 } // wasn't an FP divide
310 // Select - Convert the specified operand from a target-independent to a
311 // target-specific node if it hasn't already been changed.
312 SDNode *IA64DAGToDAGISel::Select(SDOperand Op) {
314 if (N->getOpcode() >= ISD::BUILTIN_OP_END &&
315 N->getOpcode() < IA64ISD::FIRST_NUMBER)
316 return NULL; // Already selected.
318 switch (N->getOpcode()) {
321 case IA64ISD::BRCALL: { // XXX: this is also a hack!
322 SDOperand Chain = N->getOperand(0);
323 SDOperand InFlag; // Null incoming flag value.
325 AddToISelQueue(Chain);
326 if(N->getNumOperands()==3) { // we have an incoming chain, callee and flag
327 InFlag = N->getOperand(2);
328 AddToISelQueue(InFlag);
332 SDOperand CallOperand;
334 // if we can call directly, do so
335 if (GlobalAddressSDNode *GASD =
336 dyn_cast<GlobalAddressSDNode>(N->getOperand(1))) {
337 CallOpcode = IA64::BRCALL_IPREL_GA;
338 CallOperand = CurDAG->getTargetGlobalAddress(GASD->getGlobal(), MVT::i64);
339 } else if (isa<ExternalSymbolSDNode>(N->getOperand(1))) {
340 // FIXME: we currently NEED this case for correctness, to avoid
341 // "non-pic code with imm reloc.n against dynamic symbol" errors
342 CallOpcode = IA64::BRCALL_IPREL_ES;
343 CallOperand = N->getOperand(1);
345 // otherwise we need to load the function descriptor,
346 // load the branch target (function)'s entry point and GP,
347 // branch (call) then restore the GP
348 SDOperand FnDescriptor = N->getOperand(1);
349 AddToISelQueue(FnDescriptor);
351 // load the branch target's entry point [mem] and
353 SDOperand targetEntryPoint=
354 SDOperand(CurDAG->getTargetNode(IA64::LD8, MVT::i64, FnDescriptor), 0);
355 Chain = targetEntryPoint.getValue(1);
356 SDOperand targetGPAddr=
357 SDOperand(CurDAG->getTargetNode(IA64::ADDS, MVT::i64,
358 FnDescriptor, CurDAG->getConstant(8, MVT::i64)), 0);
359 Chain = targetGPAddr.getValue(1);
361 SDOperand(CurDAG->getTargetNode(IA64::LD8, MVT::i64, targetGPAddr), 0);
362 Chain = targetGP.getValue(1);
364 Chain = CurDAG->getCopyToReg(Chain, IA64::r1, targetGP, InFlag);
365 InFlag = Chain.getValue(1);
366 Chain = CurDAG->getCopyToReg(Chain, IA64::B6, targetEntryPoint, InFlag); // FLAG these?
367 InFlag = Chain.getValue(1);
369 CallOperand = CurDAG->getRegister(IA64::B6, MVT::i64);
370 CallOpcode = IA64::BRCALL_INDIRECT;
373 // Finally, once everything is setup, emit the call itself
375 Chain = SDOperand(CurDAG->getTargetNode(CallOpcode, MVT::Other, MVT::Flag,
376 CallOperand, InFlag), 0);
377 else // there might be no arguments
378 Chain = SDOperand(CurDAG->getTargetNode(CallOpcode, MVT::Other, MVT::Flag,
379 CallOperand, Chain), 0);
380 InFlag = Chain.getValue(1);
382 std::vector<SDOperand> CallResults;
384 CallResults.push_back(Chain);
385 CallResults.push_back(InFlag);
387 for (unsigned i = 0, e = CallResults.size(); i != e; ++i)
388 ReplaceUses(Op.getValue(i), CallResults[i]);
392 case IA64ISD::GETFD: {
393 SDOperand Input = N->getOperand(0);
394 AddToISelQueue(Input);
395 return CurDAG->getTargetNode(IA64::GETFD, MVT::i64, Input);
403 return SelectDIV(Op);
405 case ISD::TargetConstantFP: {
406 SDOperand Chain = CurDAG->getEntryNode(); // this is a constant, so..
409 if (cast<ConstantFPSDNode>(N)->isExactlyValue(+0.0)) {
410 V = CurDAG->getCopyFromReg(Chain, IA64::F0, MVT::f64);
411 } else if (cast<ConstantFPSDNode>(N)->isExactlyValue(+1.0)) {
412 V = CurDAG->getCopyFromReg(Chain, IA64::F1, MVT::f64);
414 assert(0 && "Unexpected FP constant!");
416 ReplaceUses(SDOperand(N, 0), V);
420 case ISD::FrameIndex: { // TODO: reduce creepyness
421 int FI = cast<FrameIndexSDNode>(N)->getIndex();
423 return CurDAG->SelectNodeTo(N, IA64::MOV, MVT::i64,
424 CurDAG->getTargetFrameIndex(FI, MVT::i64));
426 return CurDAG->getTargetNode(IA64::MOV, MVT::i64,
427 CurDAG->getTargetFrameIndex(FI, MVT::i64));
430 case ISD::ConstantPool: { // TODO: nuke the constant pool
431 // (ia64 doesn't need one)
432 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(N);
433 Constant *C = CP->getConstVal();
434 SDOperand CPI = CurDAG->getTargetConstantPool(C, MVT::i64,
436 return CurDAG->getTargetNode(IA64::ADDL_GA, MVT::i64, // ?
437 CurDAG->getRegister(IA64::r1, MVT::i64), CPI);
440 case ISD::GlobalAddress: {
441 GlobalValue *GV = cast<GlobalAddressSDNode>(N)->getGlobal();
442 SDOperand GA = CurDAG->getTargetGlobalAddress(GV, MVT::i64);
443 SDOperand Tmp = SDOperand(CurDAG->getTargetNode(IA64::ADDL_GA, MVT::i64,
444 CurDAG->getRegister(IA64::r1, MVT::i64), GA), 0);
445 return CurDAG->getTargetNode(IA64::LD8, MVT::i64, Tmp);
448 /* XXX case ISD::ExternalSymbol: {
449 SDOperand EA = CurDAG->getTargetExternalSymbol(cast<ExternalSymbolSDNode>(N)->getSymbol(),
451 SDOperand Tmp = CurDAG->getTargetNode(IA64::ADDL_EA, MVT::i64,
452 CurDAG->getRegister(IA64::r1, MVT::i64), EA);
453 return CurDAG->getTargetNode(IA64::LD8, MVT::i64, Tmp);
457 case ISD::LOAD: { // FIXME: load -1, not 1, for bools?
458 LoadSDNode *LD = cast<LoadSDNode>(N);
459 SDOperand Chain = LD->getChain();
460 SDOperand Address = LD->getBasePtr();
461 AddToISelQueue(Chain);
462 AddToISelQueue(Address);
464 MVT::ValueType TypeBeingLoaded = LD->getLoadedVT();
466 switch (TypeBeingLoaded) {
471 assert(0 && "Cannot load this type!");
472 case MVT::i1: { // this is a bool
473 Opc = IA64::LD1; // first we load a byte, then compare for != 0
474 if(N->getValueType(0) == MVT::i1) { // XXX: early exit!
475 return CurDAG->SelectNodeTo(N, IA64::CMPNE, MVT::i1, MVT::Other,
476 SDOperand(CurDAG->getTargetNode(Opc, MVT::i64, Address), 0),
477 CurDAG->getRegister(IA64::r0, MVT::i64),
480 /* otherwise, we want to load a bool into something bigger: LD1
481 will do that for us, so we just fall through */
483 case MVT::i8: Opc = IA64::LD1; break;
484 case MVT::i16: Opc = IA64::LD2; break;
485 case MVT::i32: Opc = IA64::LD4; break;
486 case MVT::i64: Opc = IA64::LD8; break;
488 case MVT::f32: Opc = IA64::LDF4; break;
489 case MVT::f64: Opc = IA64::LDF8; break;
492 // TODO: comment this
493 return CurDAG->SelectNodeTo(N, Opc, N->getValueType(0), MVT::Other,
498 StoreSDNode *ST = cast<StoreSDNode>(N);
499 SDOperand Address = ST->getBasePtr();
500 SDOperand Chain = ST->getChain();
501 AddToISelQueue(Address);
502 AddToISelQueue(Chain);
505 if (ISD::isNON_TRUNCStore(N)) {
506 switch (N->getOperand(1).getValueType()) {
507 default: assert(0 && "unknown type in store");
508 case MVT::i1: { // this is a bool
509 Opc = IA64::ST1; // we store either 0 or 1 as a byte
511 SDOperand Initial = CurDAG->getCopyFromReg(Chain, IA64::r0, MVT::i64);
512 Chain = Initial.getValue(1);
513 // then load 1 into the same reg iff the predicate to store is 1
514 SDOperand Tmp = ST->getValue();
516 Tmp = SDOperand(CurDAG->getTargetNode(IA64::TPCADDS, MVT::i64, Initial,
517 CurDAG->getTargetConstant(1, MVT::i64),
519 return CurDAG->SelectNodeTo(N, Opc, MVT::Other, Address, Tmp, Chain);
521 case MVT::i64: Opc = IA64::ST8; break;
522 case MVT::f64: Opc = IA64::STF8; break;
524 } else { // Truncating store
525 switch(ST->getStoredVT()) {
526 default: assert(0 && "unknown type in truncstore");
527 case MVT::i8: Opc = IA64::ST1; break;
528 case MVT::i16: Opc = IA64::ST2; break;
529 case MVT::i32: Opc = IA64::ST4; break;
530 case MVT::f32: Opc = IA64::STF4; break;
534 SDOperand N1 = N->getOperand(1);
535 SDOperand N2 = N->getOperand(2);
538 return CurDAG->SelectNodeTo(N, Opc, MVT::Other, N2, N1, Chain);
542 SDOperand Chain = N->getOperand(0);
543 SDOperand CC = N->getOperand(1);
544 AddToISelQueue(Chain);
546 MachineBasicBlock *Dest =
547 cast<BasicBlockSDNode>(N->getOperand(2))->getBasicBlock();
548 //FIXME - we do NOT need long branches all the time
549 return CurDAG->SelectNodeTo(N, IA64::BRLCOND_NOTCALL, MVT::Other, CC,
550 CurDAG->getBasicBlock(Dest), Chain);
553 case ISD::CALLSEQ_START:
554 case ISD::CALLSEQ_END: {
555 int64_t Amt = cast<ConstantSDNode>(N->getOperand(1))->getValue();
556 unsigned Opc = N->getOpcode() == ISD::CALLSEQ_START ?
557 IA64::ADJUSTCALLSTACKDOWN : IA64::ADJUSTCALLSTACKUP;
558 SDOperand N0 = N->getOperand(0);
560 return CurDAG->SelectNodeTo(N, Opc, MVT::Other, getI64Imm(Amt), N0);
564 // FIXME: we don't need long branches all the time!
565 SDOperand N0 = N->getOperand(0);
567 return CurDAG->SelectNodeTo(N, IA64::BRL_NOTCALL, MVT::Other,
568 N->getOperand(1), N0);
571 return SelectCode(Op);
575 /// createIA64DAGToDAGInstructionSelector - This pass converts a legalized DAG
576 /// into an IA64-specific DAG, ready for instruction scheduling.
579 *llvm::createIA64DAGToDAGInstructionSelector(IA64TargetMachine &TM) {
580 return new IA64DAGToDAGISel(TM);