1 //===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Evan Cheng and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that ARM uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
16 #include "ARMAddressingModes.h"
17 #include "ARMConstantPoolValue.h"
18 #include "ARMISelLowering.h"
19 #include "ARMMachineFunctionInfo.h"
20 #include "ARMRegisterInfo.h"
21 #include "ARMSubtarget.h"
22 #include "ARMTargetMachine.h"
23 #include "llvm/CallingConv.h"
24 #include "llvm/Constants.h"
25 #include "llvm/Instruction.h"
26 #include "llvm/CodeGen/MachineBasicBlock.h"
27 #include "llvm/CodeGen/MachineFrameInfo.h"
28 #include "llvm/CodeGen/MachineFunction.h"
29 #include "llvm/CodeGen/MachineInstrBuilder.h"
30 #include "llvm/CodeGen/SelectionDAG.h"
31 #include "llvm/CodeGen/SSARegMap.h"
32 #include "llvm/Target/TargetOptions.h"
33 #include "llvm/ADT/VectorExtras.h"
34 #include "llvm/Support/MathExtras.h"
37 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
38 : TargetLowering(TM), ARMPCLabelIndex(0) {
39 Subtarget = &TM.getSubtarget<ARMSubtarget>();
41 // Uses VFP for Thumb libfuncs if available.
42 if (!UseSoftFloat && Subtarget->isThumb() && Subtarget->hasVFP2()) {
43 // Single-precision floating-point arithmetic.
44 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
45 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
46 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
47 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
49 // Double-precision floating-point arithmetic.
50 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
51 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
52 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
53 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
55 // Single-precision comparisons.
56 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
57 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
58 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
59 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
60 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
61 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
62 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
63 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
65 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
66 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
67 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
68 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
69 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
70 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
71 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
72 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
74 // Double-precision comparisons.
75 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
76 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
77 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
78 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
79 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
80 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
81 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
82 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
84 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
85 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
86 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
87 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
88 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
89 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
90 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
91 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
93 // Floating-point to integer conversions.
94 // i64 conversions are done via library routines even when generating VFP
95 // instructions, so use the same ones.
96 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
97 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
98 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
99 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
101 // Conversions between floating types.
102 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
103 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
105 // Integer to floating-point conversions.
106 // i64 conversions are done via library routines even when generating VFP
107 // instructions, so use the same ones.
108 // FIXME: There appears to be some naming inconsistency in ARM libgcc: e.g.
109 // __floatunsidf vs. __floatunssidfvfp.
110 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
111 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
112 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
113 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
116 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
117 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb()) {
118 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
119 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
122 // ARM does not have f32 extending load.
123 setLoadXAction(ISD::EXTLOAD, MVT::f32, Expand);
125 // ARM supports all 4 flavors of integer indexed load / store.
126 for (unsigned im = (unsigned)ISD::PRE_INC;
127 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
128 setIndexedLoadAction(im, MVT::i1, Legal);
129 setIndexedLoadAction(im, MVT::i8, Legal);
130 setIndexedLoadAction(im, MVT::i16, Legal);
131 setIndexedLoadAction(im, MVT::i32, Legal);
132 setIndexedStoreAction(im, MVT::i1, Legal);
133 setIndexedStoreAction(im, MVT::i8, Legal);
134 setIndexedStoreAction(im, MVT::i16, Legal);
135 setIndexedStoreAction(im, MVT::i32, Legal);
138 // i64 operation support.
139 if (Subtarget->isThumb()) {
140 setOperationAction(ISD::MUL, MVT::i64, Expand);
141 setOperationAction(ISD::MULHU, MVT::i32, Expand);
142 setOperationAction(ISD::MULHS, MVT::i32, Expand);
144 setOperationAction(ISD::MUL, MVT::i64, Custom);
145 setOperationAction(ISD::MULHU, MVT::i32, Custom);
146 if (!Subtarget->hasV6Ops())
147 setOperationAction(ISD::MULHS, MVT::i32, Custom);
149 setOperationAction(ISD::SHL_PARTS, MVT::i32, Expand);
150 setOperationAction(ISD::SRA_PARTS, MVT::i32, Expand);
151 setOperationAction(ISD::SRL_PARTS, MVT::i32, Expand);
152 setOperationAction(ISD::SRL, MVT::i64, Custom);
153 setOperationAction(ISD::SRA, MVT::i64, Custom);
155 // ARM does not have ROTL.
156 setOperationAction(ISD::ROTL, MVT::i32, Expand);
157 setOperationAction(ISD::CTTZ , MVT::i32, Expand);
158 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
159 if (!Subtarget->hasV5TOps() || Subtarget->isThumb())
160 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
162 // These are expanded into libcalls.
163 setOperationAction(ISD::SDIV, MVT::i32, Expand);
164 setOperationAction(ISD::UDIV, MVT::i32, Expand);
165 setOperationAction(ISD::SREM, MVT::i32, Expand);
166 setOperationAction(ISD::UREM, MVT::i32, Expand);
168 // Support label based line numbers.
169 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
170 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
171 // FIXME - use subtarget debug flags
172 if (!Subtarget->isTargetDarwin())
173 setOperationAction(ISD::LABEL, MVT::Other, Expand);
175 setOperationAction(ISD::RET, MVT::Other, Custom);
176 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
177 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
179 // Expand mem operations genericly.
180 setOperationAction(ISD::MEMSET , MVT::Other, Expand);
181 setOperationAction(ISD::MEMCPY , MVT::Other, Expand);
182 setOperationAction(ISD::MEMMOVE , MVT::Other, Expand);
184 // Use the default implementation.
185 setOperationAction(ISD::VASTART , MVT::Other, Expand);
186 setOperationAction(ISD::VAARG , MVT::Other, Expand);
187 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
188 setOperationAction(ISD::VAEND , MVT::Other, Expand);
189 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
190 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
191 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Expand);
193 if (!Subtarget->hasV6Ops()) {
194 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
195 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
197 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
199 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb())
200 // Turn f64->i64 into FMRRD iff target supports vfp2.
201 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
203 setOperationAction(ISD::SETCC , MVT::i32, Expand);
204 setOperationAction(ISD::SETCC , MVT::f32, Expand);
205 setOperationAction(ISD::SETCC , MVT::f64, Expand);
206 setOperationAction(ISD::SELECT , MVT::i32, Expand);
207 setOperationAction(ISD::SELECT , MVT::f32, Expand);
208 setOperationAction(ISD::SELECT , MVT::f64, Expand);
209 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
210 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
211 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
213 setOperationAction(ISD::BRCOND , MVT::Other, Expand);
214 setOperationAction(ISD::BR_CC , MVT::i32, Custom);
215 setOperationAction(ISD::BR_CC , MVT::f32, Custom);
216 setOperationAction(ISD::BR_CC , MVT::f64, Custom);
217 setOperationAction(ISD::BR_JT , MVT::Other, Custom);
219 setOperationAction(ISD::VASTART, MVT::Other, Custom);
220 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
221 setOperationAction(ISD::VAEND, MVT::Other, Expand);
222 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
223 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
225 // FP Constants can't be immediates.
226 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
227 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
229 // We don't support sin/cos/fmod/copysign
230 setOperationAction(ISD::FSIN , MVT::f64, Expand);
231 setOperationAction(ISD::FSIN , MVT::f32, Expand);
232 setOperationAction(ISD::FCOS , MVT::f32, Expand);
233 setOperationAction(ISD::FCOS , MVT::f64, Expand);
234 setOperationAction(ISD::FREM , MVT::f64, Expand);
235 setOperationAction(ISD::FREM , MVT::f32, Expand);
236 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
237 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
239 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
240 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
241 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
242 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
243 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
245 setStackPointerRegisterToSaveRestore(ARM::SP);
247 setSchedulingPreference(SchedulingForRegPressure);
248 computeRegisterProperties();
252 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
255 case ARMISD::Wrapper: return "ARMISD::Wrapper";
256 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
257 case ARMISD::CALL: return "ARMISD::CALL";
258 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
259 case ARMISD::tCALL: return "ARMISD::tCALL";
260 case ARMISD::BRCOND: return "ARMISD::BRCOND";
261 case ARMISD::BR_JT: return "ARMISD::BR_JT";
262 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
263 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
264 case ARMISD::CMP: return "ARMISD::CMP";
265 case ARMISD::CMPFP: return "ARMISD::CMPFP";
266 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
267 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
268 case ARMISD::CMOV: return "ARMISD::CMOV";
269 case ARMISD::CNEG: return "ARMISD::CNEG";
271 case ARMISD::FTOSI: return "ARMISD::FTOSI";
272 case ARMISD::FTOUI: return "ARMISD::FTOUI";
273 case ARMISD::SITOF: return "ARMISD::SITOF";
274 case ARMISD::UITOF: return "ARMISD::UITOF";
275 case ARMISD::MULHILOU: return "ARMISD::MULHILOU";
276 case ARMISD::MULHILOS: return "ARMISD::MULHILOS";
278 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
279 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
280 case ARMISD::RRX: return "ARMISD::RRX";
282 case ARMISD::FMRRD: return "ARMISD::FMRRD";
283 case ARMISD::FMDRR: return "ARMISD::FMDRR";
287 //===----------------------------------------------------------------------===//
289 //===----------------------------------------------------------------------===//
292 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
293 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
295 default: assert(0 && "Unknown condition code!");
296 case ISD::SETNE: return ARMCC::NE;
297 case ISD::SETEQ: return ARMCC::EQ;
298 case ISD::SETGT: return ARMCC::GT;
299 case ISD::SETGE: return ARMCC::GE;
300 case ISD::SETLT: return ARMCC::LT;
301 case ISD::SETLE: return ARMCC::LE;
302 case ISD::SETUGT: return ARMCC::HI;
303 case ISD::SETUGE: return ARMCC::HS;
304 case ISD::SETULT: return ARMCC::LO;
305 case ISD::SETULE: return ARMCC::LS;
309 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC. It
310 /// returns true if the operands should be inverted to form the proper
312 static bool FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
313 ARMCC::CondCodes &CondCode2) {
315 CondCode2 = ARMCC::AL;
317 default: assert(0 && "Unknown FP condition!");
319 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
321 case ISD::SETOGT: CondCode = ARMCC::GT; break;
323 case ISD::SETOGE: CondCode = ARMCC::GE; break;
324 case ISD::SETOLT: CondCode = ARMCC::MI; break;
325 case ISD::SETOLE: CondCode = ARMCC::GT; Invert = true; break;
326 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
327 case ISD::SETO: CondCode = ARMCC::VC; break;
328 case ISD::SETUO: CondCode = ARMCC::VS; break;
329 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
330 case ISD::SETUGT: CondCode = ARMCC::HI; break;
331 case ISD::SETUGE: CondCode = ARMCC::PL; break;
333 case ISD::SETULT: CondCode = ARMCC::LT; break;
335 case ISD::SETULE: CondCode = ARMCC::LE; break;
337 case ISD::SETUNE: CondCode = ARMCC::NE; break;
343 HowToPassArgument(MVT::ValueType ObjectVT, unsigned NumGPRs,
344 unsigned StackOffset, unsigned &NeededGPRs,
345 unsigned &NeededStackSize, unsigned &GPRPad,
346 unsigned &StackPad, unsigned Flags) {
351 unsigned align = (Flags >> ISD::ParamFlags::OrigAlignmentOffs);
352 GPRPad = NumGPRs % ((align + 3)/4);
353 StackPad = StackOffset % align;
354 unsigned firstGPR = NumGPRs + GPRPad;
356 default: assert(0 && "Unhandled argument type!");
368 else if (firstGPR == 3) {
376 /// LowerCALL - Lowering a ISD::CALL node into a callseq_start <-
377 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
379 SDOperand ARMTargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
380 MVT::ValueType RetVT= Op.Val->getValueType(0);
381 SDOperand Chain = Op.getOperand(0);
382 unsigned CallConv = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
383 assert((CallConv == CallingConv::C ||
384 CallConv == CallingConv::Fast) && "unknown calling convention");
385 SDOperand Callee = Op.getOperand(4);
386 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
387 unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
388 unsigned NumGPRs = 0; // GPRs used for parameter passing.
390 // Count how many bytes are to be pushed on the stack.
391 unsigned NumBytes = 0;
393 // Add up all the space actually used.
394 for (unsigned i = 0; i < NumOps; ++i) {
399 MVT::ValueType ObjectVT = Op.getOperand(5+2*i).getValueType();
400 unsigned Flags = Op.getConstantOperandVal(5+2*i+1);
401 HowToPassArgument(ObjectVT, NumGPRs, NumBytes, ObjGPRs, ObjSize,
402 GPRPad, StackPad, Flags);
403 NumBytes += ObjSize + StackPad;
404 NumGPRs += ObjGPRs + GPRPad;
407 // Adjust the stack pointer for the new arguments...
408 // These operations are automatically eliminated by the prolog/epilog pass
409 Chain = DAG.getCALLSEQ_START(Chain,
410 DAG.getConstant(NumBytes, MVT::i32));
412 SDOperand StackPtr = DAG.getRegister(ARM::SP, MVT::i32);
414 static const unsigned GPRArgRegs[] = {
415 ARM::R0, ARM::R1, ARM::R2, ARM::R3
419 std::vector<std::pair<unsigned, SDOperand> > RegsToPass;
420 std::vector<SDOperand> MemOpChains;
421 for (unsigned i = 0; i != NumOps; ++i) {
422 SDOperand Arg = Op.getOperand(5+2*i);
423 unsigned Flags = Op.getConstantOperandVal(5+2*i+1);
424 MVT::ValueType ArgVT = Arg.getValueType();
430 HowToPassArgument(ArgVT, NumGPRs, ArgOffset, ObjGPRs,
431 ObjSize, GPRPad, StackPad, Flags);
433 ArgOffset += StackPad;
436 default: assert(0 && "Unexpected ValueType for argument!");
438 RegsToPass.push_back(std::make_pair(GPRArgRegs[NumGPRs], Arg));
441 RegsToPass.push_back(std::make_pair(GPRArgRegs[NumGPRs],
442 DAG.getNode(ISD::BIT_CONVERT, MVT::i32, Arg)));
445 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Arg,
446 DAG.getConstant(0, getPointerTy()));
447 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Arg,
448 DAG.getConstant(1, getPointerTy()));
449 RegsToPass.push_back(std::make_pair(GPRArgRegs[NumGPRs], Lo));
451 RegsToPass.push_back(std::make_pair(GPRArgRegs[NumGPRs+1], Hi));
453 SDOperand PtrOff= DAG.getConstant(ArgOffset, StackPtr.getValueType());
454 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
455 MemOpChains.push_back(DAG.getStore(Chain, Hi, PtrOff, NULL, 0));
460 SDOperand Cvt = DAG.getNode(ARMISD::FMRRD,
461 DAG.getVTList(MVT::i32, MVT::i32),
463 RegsToPass.push_back(std::make_pair(GPRArgRegs[NumGPRs], Cvt));
465 RegsToPass.push_back(std::make_pair(GPRArgRegs[NumGPRs+1],
468 SDOperand PtrOff= DAG.getConstant(ArgOffset, StackPtr.getValueType());
469 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
470 MemOpChains.push_back(DAG.getStore(Chain, Cvt.getValue(1), PtrOff,
477 assert(ObjSize != 0);
478 SDOperand PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
479 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
480 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
484 ArgOffset += ObjSize;
487 if (!MemOpChains.empty())
488 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
489 &MemOpChains[0], MemOpChains.size());
491 // Build a sequence of copy-to-reg nodes chained together with token chain
492 // and flag operands which copy the outgoing args into the appropriate regs.
494 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
495 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
497 InFlag = Chain.getValue(1);
500 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
501 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
502 // node so that legalize doesn't hack it.
503 bool isDirect = false;
504 bool isARMFunc = false;
505 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
506 GlobalValue *GV = G->getGlobal();
508 bool isExt = (GV->isDeclaration() || GV->hasWeakLinkage() ||
509 GV->hasLinkOnceLinkage());
510 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
511 getTargetMachine().getRelocationModel() != Reloc::Static;
512 isARMFunc = !Subtarget->isThumb() || isStub;
513 // tBX takes a register source operand.
514 if (isARMFunc && Subtarget->isThumb() && !Subtarget->hasV5TOps()) {
515 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMPCLabelIndex,
517 SDOperand CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 2);
518 CPAddr = DAG.getNode(ARMISD::Wrapper, MVT::i32, CPAddr);
519 Callee = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), CPAddr, NULL, 0);
520 SDOperand PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
521 Callee = DAG.getNode(ARMISD::PIC_ADD, getPointerTy(), Callee, PICLabel);
523 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
524 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
526 bool isStub = Subtarget->isTargetDarwin() &&
527 getTargetMachine().getRelocationModel() != Reloc::Static;
528 isARMFunc = !Subtarget->isThumb() || isStub;
529 // tBX takes a register source operand.
530 const char *Sym = S->getSymbol();
531 if (isARMFunc && Subtarget->isThumb() && !Subtarget->hasV5TOps()) {
532 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(Sym, ARMPCLabelIndex,
534 SDOperand CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 2);
535 CPAddr = DAG.getNode(ARMISD::Wrapper, MVT::i32, CPAddr);
536 Callee = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), CPAddr, NULL, 0);
537 SDOperand PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
538 Callee = DAG.getNode(ARMISD::PIC_ADD, getPointerTy(), Callee, PICLabel);
540 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
543 std::vector<MVT::ValueType> NodeTys;
544 NodeTys.push_back(MVT::Other); // Returns a chain
545 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
547 std::vector<SDOperand> Ops;
548 Ops.push_back(Chain);
549 Ops.push_back(Callee);
551 // Add argument registers to the end of the list so that they are known live
553 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
554 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
555 RegsToPass[i].second.getValueType()));
557 // FIXME: handle tail calls differently.
559 if (Subtarget->isThumb()) {
560 if (!Subtarget->hasV5TOps() && (!isDirect || isARMFunc))
561 CallOpc = ARMISD::CALL_NOLINK;
563 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
565 CallOpc = (isDirect || Subtarget->hasV5TOps())
566 ? ARMISD::CALL : ARMISD::CALL_NOLINK;
569 Ops.push_back(InFlag);
570 Chain = DAG.getNode(CallOpc, NodeTys, &Ops[0], Ops.size());
571 InFlag = Chain.getValue(1);
573 SDOperand CSOps[] = { Chain, DAG.getConstant(NumBytes, MVT::i32), InFlag };
574 Chain = DAG.getNode(ISD::CALLSEQ_END,
575 DAG.getNodeValueTypes(MVT::Other, MVT::Flag),
576 ((RetVT != MVT::Other) ? 2 : 1), CSOps, 3);
577 if (RetVT != MVT::Other)
578 InFlag = Chain.getValue(1);
580 std::vector<SDOperand> ResultVals;
583 // If the call has results, copy the values out of the ret val registers.
585 default: assert(0 && "Unexpected ret value!");
589 Chain = DAG.getCopyFromReg(Chain, ARM::R0, MVT::i32, InFlag).getValue(1);
590 ResultVals.push_back(Chain.getValue(0));
591 if (Op.Val->getValueType(1) == MVT::i32) {
592 // Returns a i64 value.
593 Chain = DAG.getCopyFromReg(Chain, ARM::R1, MVT::i32,
594 Chain.getValue(2)).getValue(1);
595 ResultVals.push_back(Chain.getValue(0));
596 NodeTys.push_back(MVT::i32);
598 NodeTys.push_back(MVT::i32);
601 Chain = DAG.getCopyFromReg(Chain, ARM::R0, MVT::i32, InFlag).getValue(1);
602 ResultVals.push_back(DAG.getNode(ISD::BIT_CONVERT, MVT::f32,
604 NodeTys.push_back(MVT::f32);
607 SDOperand Lo = DAG.getCopyFromReg(Chain, ARM::R0, MVT::i32, InFlag);
608 SDOperand Hi = DAG.getCopyFromReg(Lo, ARM::R1, MVT::i32, Lo.getValue(2));
609 ResultVals.push_back(DAG.getNode(ARMISD::FMDRR, MVT::f64, Lo, Hi));
610 NodeTys.push_back(MVT::f64);
615 NodeTys.push_back(MVT::Other);
617 if (ResultVals.empty())
620 ResultVals.push_back(Chain);
621 SDOperand Res = DAG.getNode(ISD::MERGE_VALUES, NodeTys, &ResultVals[0],
623 return Res.getValue(Op.ResNo);
626 static SDOperand LowerRET(SDOperand Op, SelectionDAG &DAG) {
628 SDOperand Chain = Op.getOperand(0);
629 switch(Op.getNumOperands()) {
631 assert(0 && "Do not know how to return this many arguments!");
634 SDOperand LR = DAG.getRegister(ARM::LR, MVT::i32);
635 return DAG.getNode(ARMISD::RET_FLAG, MVT::Other, Chain);
638 Op = Op.getOperand(1);
639 if (Op.getValueType() == MVT::f32) {
640 Op = DAG.getNode(ISD::BIT_CONVERT, MVT::i32, Op);
641 } else if (Op.getValueType() == MVT::f64) {
642 // Recursively legalize f64 -> i64.
643 Op = DAG.getNode(ISD::BIT_CONVERT, MVT::i64, Op);
644 return DAG.getNode(ISD::RET, MVT::Other, Chain, Op,
645 DAG.getConstant(0, MVT::i32));
647 Copy = DAG.getCopyToReg(Chain, ARM::R0, Op, SDOperand());
648 if (DAG.getMachineFunction().liveout_empty())
649 DAG.getMachineFunction().addLiveOut(ARM::R0);
652 Copy = DAG.getCopyToReg(Chain, ARM::R1, Op.getOperand(3), SDOperand());
653 Copy = DAG.getCopyToReg(Copy, ARM::R0, Op.getOperand(1), Copy.getValue(1));
654 // If we haven't noted the R0+R1 are live out, do so now.
655 if (DAG.getMachineFunction().liveout_empty()) {
656 DAG.getMachineFunction().addLiveOut(ARM::R0);
657 DAG.getMachineFunction().addLiveOut(ARM::R1);
662 //We must use RET_FLAG instead of BRIND because BRIND doesn't have a flag
663 return DAG.getNode(ARMISD::RET_FLAG, MVT::Other, Copy, Copy.getValue(1));
666 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
667 // their target countpart wrapped in the ARMISD::Wrapper node. Suppose N is
668 // one of the above mentioned nodes. It has to be wrapped because otherwise
669 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
670 // be used to form addressing mode. These wrapped nodes will be selected
672 static SDOperand LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
673 MVT::ValueType PtrVT = Op.getValueType();
674 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
676 if (CP->isMachineConstantPoolEntry())
677 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
680 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
682 return DAG.getNode(ARMISD::Wrapper, MVT::i32, Res);
685 /// GVIsIndirectSymbol - true if the GV will be accessed via an indirect symbol
686 /// even in dynamic-no-pic mode.
687 static bool GVIsIndirectSymbol(GlobalValue *GV) {
688 return (GV->hasWeakLinkage() || GV->hasLinkOnceLinkage() ||
689 (GV->isDeclaration() && !GV->hasNotBeenReadFromBytecode()));
692 SDOperand ARMTargetLowering::LowerGlobalAddress(SDOperand Op,
694 MVT::ValueType PtrVT = getPointerTy();
695 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
696 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
697 bool IsIndirect = Subtarget->isTargetDarwin() && GVIsIndirectSymbol(GV);
699 if (RelocM == Reloc::Static)
700 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 2);
702 unsigned PCAdj = (RelocM != Reloc::PIC_)
703 ? 0 : (Subtarget->isThumb() ? 4 : 8);
704 ARMCP::ARMCPKind Kind = IsIndirect ? ARMCP::CPNonLazyPtr
706 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMPCLabelIndex,
708 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 2);
710 CPAddr = DAG.getNode(ARMISD::Wrapper, MVT::i32, CPAddr);
712 SDOperand Result = DAG.getLoad(PtrVT, DAG.getEntryNode(), CPAddr, NULL, 0);
713 SDOperand Chain = Result.getValue(1);
715 if (RelocM == Reloc::PIC_) {
716 SDOperand PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
717 Result = DAG.getNode(ARMISD::PIC_ADD, PtrVT, Result, PICLabel);
720 Result = DAG.getLoad(PtrVT, Chain, Result, NULL, 0);
725 static SDOperand LowerVASTART(SDOperand Op, SelectionDAG &DAG,
726 unsigned VarArgsFrameIndex) {
727 // vastart just stores the address of the VarArgsFrameIndex slot into the
728 // memory location argument.
729 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
730 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
731 SrcValueSDNode *SV = cast<SrcValueSDNode>(Op.getOperand(2));
732 return DAG.getStore(Op.getOperand(0), FR, Op.getOperand(1), SV->getValue(),
736 static SDOperand LowerFORMAL_ARGUMENT(SDOperand Op, SelectionDAG &DAG,
737 unsigned *vRegs, unsigned ArgNo,
738 unsigned &NumGPRs, unsigned &ArgOffset) {
739 MachineFunction &MF = DAG.getMachineFunction();
740 MVT::ValueType ObjectVT = Op.getValue(ArgNo).getValueType();
741 SDOperand Root = Op.getOperand(0);
742 std::vector<SDOperand> ArgValues;
743 SSARegMap *RegMap = MF.getSSARegMap();
745 static const unsigned GPRArgRegs[] = {
746 ARM::R0, ARM::R1, ARM::R2, ARM::R3
753 unsigned Flags = Op.getConstantOperandVal(ArgNo + 3);
754 HowToPassArgument(ObjectVT, NumGPRs, ArgOffset, ObjGPRs,
755 ObjSize, GPRPad, StackPad, Flags);
757 ArgOffset += StackPad;
761 unsigned VReg = RegMap->createVirtualRegister(&ARM::GPRRegClass);
762 MF.addLiveIn(GPRArgRegs[NumGPRs], VReg);
763 vRegs[NumGPRs] = VReg;
764 ArgValue = DAG.getCopyFromReg(Root, VReg, MVT::i32);
765 if (ObjectVT == MVT::f32)
766 ArgValue = DAG.getNode(ISD::BIT_CONVERT, MVT::f32, ArgValue);
767 } else if (ObjGPRs == 2) {
768 unsigned VReg = RegMap->createVirtualRegister(&ARM::GPRRegClass);
769 MF.addLiveIn(GPRArgRegs[NumGPRs], VReg);
770 vRegs[NumGPRs] = VReg;
771 ArgValue = DAG.getCopyFromReg(Root, VReg, MVT::i32);
773 VReg = RegMap->createVirtualRegister(&ARM::GPRRegClass);
774 MF.addLiveIn(GPRArgRegs[NumGPRs+1], VReg);
775 vRegs[NumGPRs+1] = VReg;
776 SDOperand ArgValue2 = DAG.getCopyFromReg(Root, VReg, MVT::i32);
778 if (ObjectVT == MVT::i64)
779 ArgValue = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, ArgValue, ArgValue2);
781 ArgValue = DAG.getNode(ARMISD::FMDRR, MVT::f64, ArgValue, ArgValue2);
786 // If the argument is actually used, emit a load from the right stack
788 if (!Op.Val->hasNUsesOfValue(0, ArgNo)) {
789 MachineFrameInfo *MFI = MF.getFrameInfo();
790 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
791 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
793 ArgValue = DAG.getLoad(ObjectVT, Root, FIN, NULL, 0);
795 SDOperand ArgValue2 =
796 DAG.getLoad(MVT::i32, Root, FIN, NULL, 0);
797 if (ObjectVT == MVT::i64)
798 ArgValue= DAG.getNode(ISD::BUILD_PAIR, MVT::i64, ArgValue, ArgValue2);
800 ArgValue= DAG.getNode(ARMISD::FMDRR, MVT::f64, ArgValue, ArgValue2);
803 // Don't emit a dead load.
804 ArgValue = DAG.getNode(ISD::UNDEF, ObjectVT);
807 ArgOffset += ObjSize; // Move on to the next argument.
814 ARMTargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) {
815 std::vector<SDOperand> ArgValues;
816 SDOperand Root = Op.getOperand(0);
817 unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
818 unsigned NumGPRs = 0; // GPRs used for parameter passing.
821 unsigned NumArgs = Op.Val->getNumValues()-1;
822 for (unsigned ArgNo = 0; ArgNo < NumArgs; ++ArgNo)
823 ArgValues.push_back(LowerFORMAL_ARGUMENT(Op, DAG, VRegs, ArgNo,
824 NumGPRs, ArgOffset));
826 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
828 static const unsigned GPRArgRegs[] = {
829 ARM::R0, ARM::R1, ARM::R2, ARM::R3
832 MachineFunction &MF = DAG.getMachineFunction();
833 SSARegMap *RegMap = MF.getSSARegMap();
834 MachineFrameInfo *MFI = MF.getFrameInfo();
835 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
836 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
837 unsigned VARegSize = (4 - NumGPRs) * 4;
838 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
840 // If this function is vararg, store any remaining integer argument regs
841 // to their spots on the stack so that they may be loaded by deferencing
842 // the result of va_next.
843 AFI->setVarArgsRegSaveSize(VARegSaveSize);
844 VarArgsFrameIndex = MFI->CreateFixedObject(VARegSaveSize, ArgOffset +
845 VARegSaveSize - VARegSize);
846 SDOperand FIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
848 SmallVector<SDOperand, 4> MemOps;
849 for (; NumGPRs < 4; ++NumGPRs) {
850 unsigned VReg = RegMap->createVirtualRegister(&ARM::GPRRegClass);
851 MF.addLiveIn(GPRArgRegs[NumGPRs], VReg);
852 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::i32);
853 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
854 MemOps.push_back(Store);
855 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
856 DAG.getConstant(4, getPointerTy()));
859 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
860 &MemOps[0], MemOps.size());
862 // This will point to the next argument passed via stack.
863 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset);
866 ArgValues.push_back(Root);
868 // Return the new list of results.
869 std::vector<MVT::ValueType> RetVT(Op.Val->value_begin(),
870 Op.Val->value_end());
871 return DAG.getNode(ISD::MERGE_VALUES, RetVT, &ArgValues[0], ArgValues.size());
874 /// isFloatingPointZero - Return true if this is +0.0.
875 static bool isFloatingPointZero(SDOperand Op) {
876 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
877 return CFP->isExactlyValue(0.0);
878 else if (ISD::isEXTLoad(Op.Val) || ISD::isNON_EXTLoad(Op.Val)) {
879 // Maybe this has already been legalized into the constant pool?
880 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
881 SDOperand WrapperOp = Op.getOperand(1).getOperand(0);
882 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
883 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
884 return CFP->isExactlyValue(0.0);
890 static bool isLegalCmpImmediate(unsigned C, bool isThumb) {
891 return ( isThumb && (C & ~255U) == 0) ||
892 (!isThumb && ARM_AM::getSOImmVal(C) != -1);
895 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
896 /// the given operands.
897 static SDOperand getARMCmp(SDOperand LHS, SDOperand RHS, ISD::CondCode CC,
898 SDOperand &ARMCC, SelectionDAG &DAG, bool isThumb) {
899 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.Val)) {
900 unsigned C = RHSC->getValue();
901 if (!isLegalCmpImmediate(C, isThumb)) {
902 // Constant does not fit, try adjusting it by one?
907 if (isLegalCmpImmediate(C-1, isThumb)) {
908 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
909 RHS = DAG.getConstant(C-1, MVT::i32);
914 if (C > 0 && isLegalCmpImmediate(C-1, isThumb)) {
915 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
916 RHS = DAG.getConstant(C-1, MVT::i32);
921 if (isLegalCmpImmediate(C+1, isThumb)) {
922 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
923 RHS = DAG.getConstant(C+1, MVT::i32);
928 if (C < 0xffffffff && isLegalCmpImmediate(C+1, isThumb)) {
929 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
930 RHS = DAG.getConstant(C+1, MVT::i32);
937 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
938 ARMCC = DAG.getConstant(CondCode, MVT::i32);
939 return DAG.getNode(ARMISD::CMP, MVT::Flag, LHS, RHS);
942 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
943 static SDOperand getVFPCmp(SDOperand LHS, SDOperand RHS, SelectionDAG &DAG) {
945 if (!isFloatingPointZero(RHS))
946 Cmp = DAG.getNode(ARMISD::CMPFP, MVT::Flag, LHS, RHS);
948 Cmp = DAG.getNode(ARMISD::CMPFPw0, MVT::Flag, LHS);
949 return DAG.getNode(ARMISD::FMSTAT, MVT::Flag, Cmp);
952 static SDOperand LowerSELECT_CC(SDOperand Op, SelectionDAG &DAG,
953 const ARMSubtarget *ST) {
954 MVT::ValueType VT = Op.getValueType();
955 SDOperand LHS = Op.getOperand(0);
956 SDOperand RHS = Op.getOperand(1);
957 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
958 SDOperand TrueVal = Op.getOperand(2);
959 SDOperand FalseVal = Op.getOperand(3);
961 if (LHS.getValueType() == MVT::i32) {
963 SDOperand Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb());
964 return DAG.getNode(ARMISD::CMOV, VT, FalseVal, TrueVal, ARMCC, Cmp);
967 ARMCC::CondCodes CondCode, CondCode2;
968 if (FPCCToARMCC(CC, CondCode, CondCode2))
969 std::swap(TrueVal, FalseVal);
971 SDOperand ARMCC = DAG.getConstant(CondCode, MVT::i32);
972 SDOperand Cmp = getVFPCmp(LHS, RHS, DAG);
973 SDOperand Result = DAG.getNode(ARMISD::CMOV, VT, FalseVal, TrueVal,
975 if (CondCode2 != ARMCC::AL) {
976 SDOperand ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
977 // FIXME: Needs another CMP because flag can have but one use.
978 SDOperand Cmp2 = getVFPCmp(LHS, RHS, DAG);
979 Result = DAG.getNode(ARMISD::CMOV, VT, Result, TrueVal, ARMCC2, Cmp2);
984 static SDOperand LowerBR_CC(SDOperand Op, SelectionDAG &DAG,
985 const ARMSubtarget *ST) {
986 SDOperand Chain = Op.getOperand(0);
987 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
988 SDOperand LHS = Op.getOperand(2);
989 SDOperand RHS = Op.getOperand(3);
990 SDOperand Dest = Op.getOperand(4);
992 if (LHS.getValueType() == MVT::i32) {
994 SDOperand Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb());
995 return DAG.getNode(ARMISD::BRCOND, MVT::Other, Chain, Dest, ARMCC, Cmp);
998 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
999 ARMCC::CondCodes CondCode, CondCode2;
1000 if (FPCCToARMCC(CC, CondCode, CondCode2))
1001 // Swap the LHS/RHS of the comparison if needed.
1002 std::swap(LHS, RHS);
1004 SDOperand Cmp = getVFPCmp(LHS, RHS, DAG);
1005 SDOperand ARMCC = DAG.getConstant(CondCode, MVT::i32);
1006 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
1007 SDOperand Ops[] = { Chain, Dest, ARMCC, Cmp };
1008 SDOperand Res = DAG.getNode(ARMISD::BRCOND, VTList, Ops, 4);
1009 if (CondCode2 != ARMCC::AL) {
1010 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
1011 SDOperand Ops[] = { Res, Dest, ARMCC, Res.getValue(1) };
1012 Res = DAG.getNode(ARMISD::BRCOND, VTList, Ops, 4);
1017 SDOperand ARMTargetLowering::LowerBR_JT(SDOperand Op, SelectionDAG &DAG) {
1018 SDOperand Chain = Op.getOperand(0);
1019 SDOperand Table = Op.getOperand(1);
1020 SDOperand Index = Op.getOperand(2);
1022 MVT::ValueType PTy = getPointerTy();
1023 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
1024 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
1025 SDOperand UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
1026 SDOperand JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
1027 Table = DAG.getNode(ARMISD::WrapperJT, MVT::i32, JTI, UId);
1028 Index = DAG.getNode(ISD::MUL, PTy, Index, DAG.getConstant(4, PTy));
1029 SDOperand Addr = DAG.getNode(ISD::ADD, PTy, Index, Table);
1030 bool isPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_;
1031 Addr = DAG.getLoad(isPIC ? MVT::i32 : PTy, Chain, Addr, NULL, 0);
1032 Chain = Addr.getValue(1);
1034 Addr = DAG.getNode(ISD::ADD, PTy, Addr, Table);
1035 return DAG.getNode(ARMISD::BR_JT, MVT::Other, Chain, Addr, JTI, UId);
1038 static SDOperand LowerFP_TO_INT(SDOperand Op, SelectionDAG &DAG) {
1040 Op.getOpcode() == ISD::FP_TO_SINT ? ARMISD::FTOSI : ARMISD::FTOUI;
1041 Op = DAG.getNode(Opc, MVT::f32, Op.getOperand(0));
1042 return DAG.getNode(ISD::BIT_CONVERT, MVT::i32, Op);
1045 static SDOperand LowerINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
1046 MVT::ValueType VT = Op.getValueType();
1048 Op.getOpcode() == ISD::SINT_TO_FP ? ARMISD::SITOF : ARMISD::UITOF;
1050 Op = DAG.getNode(ISD::BIT_CONVERT, MVT::f32, Op.getOperand(0));
1051 return DAG.getNode(Opc, VT, Op);
1054 static SDOperand LowerFCOPYSIGN(SDOperand Op, SelectionDAG &DAG) {
1055 // Implement fcopysign with a fabs and a conditional fneg.
1056 SDOperand Tmp0 = Op.getOperand(0);
1057 SDOperand Tmp1 = Op.getOperand(1);
1058 MVT::ValueType VT = Op.getValueType();
1059 MVT::ValueType SrcVT = Tmp1.getValueType();
1060 SDOperand AbsVal = DAG.getNode(ISD::FABS, VT, Tmp0);
1061 SDOperand Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG);
1062 SDOperand ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
1063 return DAG.getNode(ARMISD::CNEG, VT, AbsVal, AbsVal, ARMCC, Cmp);
1066 static SDOperand LowerBIT_CONVERT(SDOperand Op, SelectionDAG &DAG) {
1067 // Turn f64->i64 into FMRRD.
1068 assert(Op.getValueType() == MVT::i64 &&
1069 Op.getOperand(0).getValueType() == MVT::f64);
1071 Op = Op.getOperand(0);
1072 SDOperand Cvt = DAG.getNode(ARMISD::FMRRD, DAG.getVTList(MVT::i32, MVT::i32),
1075 // Merge the pieces into a single i64 value.
1076 return DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Cvt, Cvt.getValue(1));
1079 static SDOperand LowerMUL(SDOperand Op, SelectionDAG &DAG) {
1080 // FIXME: All this code is target-independent. Create a new target-indep
1081 // MULHILO node and move this code to the legalizer.
1083 assert(Op.getValueType() == MVT::i64 && "Only handles i64 expand right now!");
1085 SDOperand LL = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
1086 DAG.getConstant(0, MVT::i32));
1087 SDOperand RL = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(1),
1088 DAG.getConstant(0, MVT::i32));
1090 const TargetLowering &TL = DAG.getTargetLoweringInfo();
1091 unsigned LHSSB = TL.ComputeNumSignBits(Op.getOperand(0));
1092 unsigned RHSSB = TL.ComputeNumSignBits(Op.getOperand(1));
1095 // Figure out how to lower this multiply.
1096 if (LHSSB >= 33 && RHSSB >= 33) {
1097 // If the input values are both sign extended, we can emit a mulhs+mul.
1098 Lo = DAG.getNode(ISD::MUL, MVT::i32, LL, RL);
1099 Hi = DAG.getNode(ISD::MULHS, MVT::i32, LL, RL);
1100 } else if (LHSSB == 32 && RHSSB == 32 &&
1101 TL.MaskedValueIsZero(Op.getOperand(0), 0xFFFFFFFF00000000ULL) &&
1102 TL.MaskedValueIsZero(Op.getOperand(1), 0xFFFFFFFF00000000ULL)) {
1103 // If the inputs are zero extended, use mulhu.
1104 Lo = DAG.getNode(ISD::MUL, MVT::i32, LL, RL);
1105 Hi = DAG.getNode(ISD::MULHU, MVT::i32, LL, RL);
1107 SDOperand LH = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
1108 DAG.getConstant(1, MVT::i32));
1109 SDOperand RH = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(1),
1110 DAG.getConstant(1, MVT::i32));
1112 // Lo,Hi = umul LHS, RHS.
1113 SDOperand Ops[] = { LL, RL };
1114 SDOperand UMul64 = DAG.getNode(ARMISD::MULHILOU,
1115 DAG.getVTList(MVT::i32, MVT::i32), Ops, 2);
1117 Hi = UMul64.getValue(1);
1118 RH = DAG.getNode(ISD::MUL, MVT::i32, LL, RH);
1119 LH = DAG.getNode(ISD::MUL, MVT::i32, LH, RL);
1120 Hi = DAG.getNode(ISD::ADD, MVT::i32, Hi, RH);
1121 Hi = DAG.getNode(ISD::ADD, MVT::i32, Hi, LH);
1124 // Merge the pieces into a single i64 value.
1125 return DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Lo, Hi);
1128 static SDOperand LowerMULHU(SDOperand Op, SelectionDAG &DAG) {
1129 SDOperand Ops[] = { Op.getOperand(0), Op.getOperand(1) };
1130 return DAG.getNode(ARMISD::MULHILOU,
1131 DAG.getVTList(MVT::i32, MVT::i32), Ops, 2).getValue(1);
1134 static SDOperand LowerMULHS(SDOperand Op, SelectionDAG &DAG) {
1135 SDOperand Ops[] = { Op.getOperand(0), Op.getOperand(1) };
1136 return DAG.getNode(ARMISD::MULHILOS,
1137 DAG.getVTList(MVT::i32, MVT::i32), Ops, 2).getValue(1);
1140 static SDOperand LowerSRx(SDOperand Op, SelectionDAG &DAG,
1141 const ARMSubtarget *ST) {
1142 assert(Op.getValueType() == MVT::i64 &&
1143 (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SRA) &&
1144 "Unknown shift to lower!");
1146 // We only lower SRA, SRL of 1 here, all others use generic lowering.
1147 if (!isa<ConstantSDNode>(Op.getOperand(1)) ||
1148 cast<ConstantSDNode>(Op.getOperand(1))->getValue() != 1)
1151 // If we are in thumb mode, we don't have RRX.
1152 if (ST->isThumb()) return SDOperand();
1154 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
1155 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
1156 DAG.getConstant(0, MVT::i32));
1157 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
1158 DAG.getConstant(1, MVT::i32));
1160 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
1161 // captures the result into a carry flag.
1162 unsigned Opc = Op.getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
1163 Hi = DAG.getNode(Opc, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
1165 // The low part is an ARMISD::RRX operand, which shifts the carry in.
1166 Lo = DAG.getNode(ARMISD::RRX, MVT::i32, Lo, Hi.getValue(1));
1168 // Merge the pieces into a single i64 value.
1169 return DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Lo, Hi);
1172 SDOperand ARMTargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
1173 switch (Op.getOpcode()) {
1174 default: assert(0 && "Don't know how to custom lower this!"); abort();
1175 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
1176 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
1177 case ISD::CALL: return LowerCALL(Op, DAG);
1178 case ISD::RET: return LowerRET(Op, DAG);
1179 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG, Subtarget);
1180 case ISD::BR_CC: return LowerBR_CC(Op, DAG, Subtarget);
1181 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
1182 case ISD::VASTART: return LowerVASTART(Op, DAG, VarArgsFrameIndex);
1183 case ISD::SINT_TO_FP:
1184 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
1185 case ISD::FP_TO_SINT:
1186 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
1187 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
1188 case ISD::BIT_CONVERT: return LowerBIT_CONVERT(Op, DAG);
1189 case ISD::MUL: return LowerMUL(Op, DAG);
1190 case ISD::MULHU: return LowerMULHU(Op, DAG);
1191 case ISD::MULHS: return LowerMULHS(Op, DAG);
1193 case ISD::SRA: return LowerSRx(Op, DAG, Subtarget);
1194 case ISD::FORMAL_ARGUMENTS:
1195 return LowerFORMAL_ARGUMENTS(Op, DAG);
1196 case ISD::RETURNADDR: break;
1197 case ISD::FRAMEADDR: break;
1202 //===----------------------------------------------------------------------===//
1203 // ARM Scheduler Hooks
1204 //===----------------------------------------------------------------------===//
1207 ARMTargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
1208 MachineBasicBlock *BB) {
1209 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1210 switch (MI->getOpcode()) {
1211 default: assert(false && "Unexpected instr type to insert");
1212 case ARM::tMOVCCr: {
1213 // To "insert" a SELECT_CC instruction, we actually have to insert the
1214 // diamond control-flow pattern. The incoming instruction knows the
1215 // destination vreg to set, the condition code register to branch on, the
1216 // true/false values to select between, and a branch opcode to use.
1217 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1218 ilist<MachineBasicBlock>::iterator It = BB;
1224 // cmpTY ccX, r1, r2
1226 // fallthrough --> copy0MBB
1227 MachineBasicBlock *thisMBB = BB;
1228 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
1229 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
1230 BuildMI(BB, TII->get(ARM::tBcc)).addMBB(sinkMBB)
1231 .addImm(MI->getOperand(3).getImm());
1232 MachineFunction *F = BB->getParent();
1233 F->getBasicBlockList().insert(It, copy0MBB);
1234 F->getBasicBlockList().insert(It, sinkMBB);
1235 // Update machine-CFG edges by first adding all successors of the current
1236 // block to the new block which will contain the Phi node for the select.
1237 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
1238 e = BB->succ_end(); i != e; ++i)
1239 sinkMBB->addSuccessor(*i);
1240 // Next, remove all successors of the current block, and add the true
1241 // and fallthrough blocks as its successors.
1242 while(!BB->succ_empty())
1243 BB->removeSuccessor(BB->succ_begin());
1244 BB->addSuccessor(copy0MBB);
1245 BB->addSuccessor(sinkMBB);
1248 // %FalseValue = ...
1249 // # fallthrough to sinkMBB
1252 // Update machine-CFG edges
1253 BB->addSuccessor(sinkMBB);
1256 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
1259 BuildMI(BB, TII->get(ARM::PHI), MI->getOperand(0).getReg())
1260 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
1261 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
1263 delete MI; // The pseudo instruction is gone now.
1269 //===----------------------------------------------------------------------===//
1270 // ARM Optimization Hooks
1271 //===----------------------------------------------------------------------===//
1273 /// isLegalAddressExpression - Return true if the binary expression made up of
1274 /// specified opcode, operands, and type can be folded into target addressing
1275 /// mode for load / store of the given type.
1276 bool ARMTargetLowering::isLegalAddressExpression(unsigned Opc, Value *Op0,
1277 Value *Op1, const Type *Ty) const {
1278 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
1279 if (Opc == Instruction::Add)
1280 return isLegalAddressImmediate(Op1C->getSExtValue(), Ty);
1281 if (Opc == Instruction::Sub)
1282 return isLegalAddressImmediate(-Op1C->getSExtValue(), Ty);
1287 /// isLegalAddressImmediate - Return true if the integer value can be used
1288 /// as the offset of the target addressing mode for load / store of the
1290 bool ARMTargetLowering::isLegalAddressImmediate(int64_t V,const Type *Ty) const{
1294 MVT::ValueType VT = getValueType(Ty);
1295 if (Subtarget->isThumb()) {
1301 default: return false;
1316 if ((V & (Scale - 1)) != 0)
1319 return V == V & ((1LL << 5) - 1);
1325 default: return false;
1330 return V == V & ((1LL << 12) - 1);
1333 return V == V & ((1LL << 8) - 1);
1336 if (!Subtarget->hasVFP2())
1341 return V == V & ((1LL << 8) - 1);
1345 bool ARMTargetLowering::isLegalAddressImmediate(GlobalValue *GV) const {
1349 /// isLegalAddressScale - Return true if the integer value can be used as
1350 /// the scale of the target addressing mode for load / store of the given
1352 bool ARMTargetLowering::isLegalAddressScale(int64_t S, const Type *Ty) const {
1353 if (Subtarget->isThumb())
1356 MVT::ValueType VT = getValueType(Ty);
1358 default: return false;
1367 return isPowerOf2_32(S);
1371 static bool getIndexedAddressParts(SDNode *Ptr, MVT::ValueType VT,
1372 bool isSEXTLoad, SDOperand &Base,
1373 SDOperand &Offset, bool &isInc,
1374 SelectionDAG &DAG) {
1375 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
1378 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
1380 Base = Ptr->getOperand(0);
1381 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
1382 int RHSC = (int)RHS->getValue();
1383 if (RHSC < 0 && RHSC > -256) {
1385 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
1389 isInc = (Ptr->getOpcode() == ISD::ADD);
1390 Offset = Ptr->getOperand(1);
1392 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
1394 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
1395 int RHSC = (int)RHS->getValue();
1396 if (RHSC < 0 && RHSC > -0x1000) {
1398 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
1399 Base = Ptr->getOperand(0);
1404 if (Ptr->getOpcode() == ISD::ADD) {
1406 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
1407 if (ShOpcVal != ARM_AM::no_shift) {
1408 Base = Ptr->getOperand(1);
1409 Offset = Ptr->getOperand(0);
1411 Base = Ptr->getOperand(0);
1412 Offset = Ptr->getOperand(1);
1417 isInc = (Ptr->getOpcode() == ISD::ADD);
1418 Base = Ptr->getOperand(0);
1419 Offset = Ptr->getOperand(1);
1423 // FIXME: Use FLDM / FSTM to emulate indexed FP load / store.
1427 /// getPreIndexedAddressParts - returns true by value, base pointer and
1428 /// offset pointer and addressing mode by reference if the node's address
1429 /// can be legally represented as pre-indexed load / store address.
1431 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDOperand &Base,
1433 ISD::MemIndexedMode &AM,
1434 SelectionDAG &DAG) {
1435 if (Subtarget->isThumb())
1440 bool isSEXTLoad = false;
1441 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1442 Ptr = LD->getBasePtr();
1443 VT = LD->getLoadedVT();
1444 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
1445 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1446 Ptr = ST->getBasePtr();
1447 VT = ST->getStoredVT();
1452 bool isLegal = getIndexedAddressParts(Ptr.Val, VT, isSEXTLoad, Base, Offset,
1455 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
1461 /// getPostIndexedAddressParts - returns true by value, base pointer and
1462 /// offset pointer and addressing mode by reference if this node can be
1463 /// combined with a load / store to form a post-indexed load / store.
1464 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
1467 ISD::MemIndexedMode &AM,
1468 SelectionDAG &DAG) {
1469 if (Subtarget->isThumb())
1474 bool isSEXTLoad = false;
1475 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1476 VT = LD->getLoadedVT();
1477 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
1478 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1479 VT = ST->getStoredVT();
1484 bool isLegal = getIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
1487 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
1493 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
1495 uint64_t &KnownZero,
1497 unsigned Depth) const {
1500 switch (Op.getOpcode()) {
1502 case ARMISD::CMOV: {
1503 // Bits are known zero/one if known on the LHS and RHS.
1504 ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
1505 if (KnownZero == 0 && KnownOne == 0) return;
1507 uint64_t KnownZeroRHS, KnownOneRHS;
1508 ComputeMaskedBits(Op.getOperand(1), Mask,
1509 KnownZeroRHS, KnownOneRHS, Depth+1);
1510 KnownZero &= KnownZeroRHS;
1511 KnownOne &= KnownOneRHS;
1517 //===----------------------------------------------------------------------===//
1518 // ARM Inline Assembly Support
1519 //===----------------------------------------------------------------------===//
1521 /// getConstraintType - Given a constraint letter, return the type of
1522 /// constraint it is for this target.
1523 ARMTargetLowering::ConstraintType
1524 ARMTargetLowering::getConstraintType(char ConstraintLetter) const {
1525 switch (ConstraintLetter) {
1527 return C_RegisterClass;
1528 default: return TargetLowering::getConstraintType(ConstraintLetter);
1532 std::pair<unsigned, const TargetRegisterClass*>
1533 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
1534 MVT::ValueType VT) const {
1535 if (Constraint.size() == 1) {
1536 // GCC RS6000 Constraint Letters
1537 switch (Constraint[0]) {
1539 // FIXME: in thumb mode, 'l' is only low-regs.
1542 return std::make_pair(0U, ARM::GPRRegisterClass);
1546 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
1549 std::vector<unsigned> ARMTargetLowering::
1550 getRegClassForInlineAsmConstraint(const std::string &Constraint,
1551 MVT::ValueType VT) const {
1552 if (Constraint.size() != 1)
1553 return std::vector<unsigned>();
1555 switch (Constraint[0]) { // GCC ARM Constraint Letters
1559 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
1560 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
1561 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
1562 ARM::R12, ARM::LR, 0);
1565 return std::vector<unsigned>();