1 //===-- TargetLowering.cpp - Implement the TargetLowering class -----------===//
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 implements the TargetLowering class.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Target/TargetAsmInfo.h"
15 #include "llvm/Target/TargetLowering.h"
16 #include "llvm/Target/TargetSubtarget.h"
17 #include "llvm/Target/TargetData.h"
18 #include "llvm/Target/TargetMachine.h"
19 #include "llvm/Target/TargetRegisterInfo.h"
20 #include "llvm/CallingConv.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/CodeGen/SelectionDAG.h"
23 #include "llvm/ADT/StringExtras.h"
24 #include "llvm/ADT/STLExtras.h"
25 #include "llvm/Support/MathExtras.h"
28 /// InitLibcallNames - Set default libcall names.
30 static void InitLibcallNames(const char **Names) {
31 Names[RTLIB::SHL_I32] = "__ashlsi3";
32 Names[RTLIB::SHL_I64] = "__ashldi3";
33 Names[RTLIB::SRL_I32] = "__lshrsi3";
34 Names[RTLIB::SRL_I64] = "__lshrdi3";
35 Names[RTLIB::SRA_I32] = "__ashrsi3";
36 Names[RTLIB::SRA_I64] = "__ashrdi3";
37 Names[RTLIB::MUL_I32] = "__mulsi3";
38 Names[RTLIB::MUL_I64] = "__muldi3";
39 Names[RTLIB::SDIV_I32] = "__divsi3";
40 Names[RTLIB::SDIV_I64] = "__divdi3";
41 Names[RTLIB::UDIV_I32] = "__udivsi3";
42 Names[RTLIB::UDIV_I64] = "__udivdi3";
43 Names[RTLIB::SREM_I32] = "__modsi3";
44 Names[RTLIB::SREM_I64] = "__moddi3";
45 Names[RTLIB::UREM_I32] = "__umodsi3";
46 Names[RTLIB::UREM_I64] = "__umoddi3";
47 Names[RTLIB::NEG_I32] = "__negsi2";
48 Names[RTLIB::NEG_I64] = "__negdi2";
49 Names[RTLIB::ADD_F32] = "__addsf3";
50 Names[RTLIB::ADD_F64] = "__adddf3";
51 Names[RTLIB::ADD_F80] = "__addxf3";
52 Names[RTLIB::ADD_PPCF128] = "__gcc_qadd";
53 Names[RTLIB::SUB_F32] = "__subsf3";
54 Names[RTLIB::SUB_F64] = "__subdf3";
55 Names[RTLIB::SUB_F80] = "__subxf3";
56 Names[RTLIB::SUB_PPCF128] = "__gcc_qsub";
57 Names[RTLIB::MUL_F32] = "__mulsf3";
58 Names[RTLIB::MUL_F64] = "__muldf3";
59 Names[RTLIB::MUL_F80] = "__mulxf3";
60 Names[RTLIB::MUL_PPCF128] = "__gcc_qmul";
61 Names[RTLIB::DIV_F32] = "__divsf3";
62 Names[RTLIB::DIV_F64] = "__divdf3";
63 Names[RTLIB::DIV_F80] = "__divxf3";
64 Names[RTLIB::DIV_PPCF128] = "__gcc_qdiv";
65 Names[RTLIB::REM_F32] = "fmodf";
66 Names[RTLIB::REM_F64] = "fmod";
67 Names[RTLIB::REM_F80] = "fmodl";
68 Names[RTLIB::REM_PPCF128] = "fmodl";
69 Names[RTLIB::POWI_F32] = "__powisf2";
70 Names[RTLIB::POWI_F64] = "__powidf2";
71 Names[RTLIB::POWI_F80] = "__powixf2";
72 Names[RTLIB::POWI_PPCF128] = "__powitf2";
73 Names[RTLIB::SQRT_F32] = "sqrtf";
74 Names[RTLIB::SQRT_F64] = "sqrt";
75 Names[RTLIB::SQRT_F80] = "sqrtl";
76 Names[RTLIB::SQRT_PPCF128] = "sqrtl";
77 Names[RTLIB::SIN_F32] = "sinf";
78 Names[RTLIB::SIN_F64] = "sin";
79 Names[RTLIB::SIN_F80] = "sinl";
80 Names[RTLIB::SIN_PPCF128] = "sinl";
81 Names[RTLIB::COS_F32] = "cosf";
82 Names[RTLIB::COS_F64] = "cos";
83 Names[RTLIB::COS_F80] = "cosl";
84 Names[RTLIB::COS_PPCF128] = "cosl";
85 Names[RTLIB::POW_F32] = "powf";
86 Names[RTLIB::POW_F64] = "pow";
87 Names[RTLIB::POW_F80] = "powl";
88 Names[RTLIB::POW_PPCF128] = "powl";
89 Names[RTLIB::FPEXT_F32_F64] = "__extendsfdf2";
90 Names[RTLIB::FPROUND_F64_F32] = "__truncdfsf2";
91 Names[RTLIB::FPTOSINT_F32_I32] = "__fixsfsi";
92 Names[RTLIB::FPTOSINT_F32_I64] = "__fixsfdi";
93 Names[RTLIB::FPTOSINT_F64_I32] = "__fixdfsi";
94 Names[RTLIB::FPTOSINT_F64_I64] = "__fixdfdi";
95 Names[RTLIB::FPTOSINT_F80_I64] = "__fixxfdi";
96 Names[RTLIB::FPTOSINT_PPCF128_I64] = "__fixtfdi";
97 Names[RTLIB::FPTOUINT_F32_I32] = "__fixunssfsi";
98 Names[RTLIB::FPTOUINT_F32_I64] = "__fixunssfdi";
99 Names[RTLIB::FPTOUINT_F64_I32] = "__fixunsdfsi";
100 Names[RTLIB::FPTOUINT_F64_I64] = "__fixunsdfdi";
101 Names[RTLIB::FPTOUINT_F80_I32] = "__fixunsxfsi";
102 Names[RTLIB::FPTOUINT_F80_I64] = "__fixunsxfdi";
103 Names[RTLIB::FPTOUINT_PPCF128_I64] = "__fixunstfdi";
104 Names[RTLIB::SINTTOFP_I32_F32] = "__floatsisf";
105 Names[RTLIB::SINTTOFP_I32_F64] = "__floatsidf";
106 Names[RTLIB::SINTTOFP_I64_F32] = "__floatdisf";
107 Names[RTLIB::SINTTOFP_I64_F64] = "__floatdidf";
108 Names[RTLIB::SINTTOFP_I64_F80] = "__floatdixf";
109 Names[RTLIB::SINTTOFP_I64_PPCF128] = "__floatditf";
110 Names[RTLIB::UINTTOFP_I32_F32] = "__floatunsisf";
111 Names[RTLIB::UINTTOFP_I32_F64] = "__floatunsidf";
112 Names[RTLIB::UINTTOFP_I64_F32] = "__floatundisf";
113 Names[RTLIB::UINTTOFP_I64_F64] = "__floatundidf";
114 Names[RTLIB::OEQ_F32] = "__eqsf2";
115 Names[RTLIB::OEQ_F64] = "__eqdf2";
116 Names[RTLIB::UNE_F32] = "__nesf2";
117 Names[RTLIB::UNE_F64] = "__nedf2";
118 Names[RTLIB::OGE_F32] = "__gesf2";
119 Names[RTLIB::OGE_F64] = "__gedf2";
120 Names[RTLIB::OLT_F32] = "__ltsf2";
121 Names[RTLIB::OLT_F64] = "__ltdf2";
122 Names[RTLIB::OLE_F32] = "__lesf2";
123 Names[RTLIB::OLE_F64] = "__ledf2";
124 Names[RTLIB::OGT_F32] = "__gtsf2";
125 Names[RTLIB::OGT_F64] = "__gtdf2";
126 Names[RTLIB::UO_F32] = "__unordsf2";
127 Names[RTLIB::UO_F64] = "__unorddf2";
128 Names[RTLIB::O_F32] = "__unordsf2";
129 Names[RTLIB::O_F64] = "__unorddf2";
132 /// InitCmpLibcallCCs - Set default comparison libcall CC.
134 static void InitCmpLibcallCCs(ISD::CondCode *CCs) {
135 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL);
136 CCs[RTLIB::OEQ_F32] = ISD::SETEQ;
137 CCs[RTLIB::OEQ_F64] = ISD::SETEQ;
138 CCs[RTLIB::UNE_F32] = ISD::SETNE;
139 CCs[RTLIB::UNE_F64] = ISD::SETNE;
140 CCs[RTLIB::OGE_F32] = ISD::SETGE;
141 CCs[RTLIB::OGE_F64] = ISD::SETGE;
142 CCs[RTLIB::OLT_F32] = ISD::SETLT;
143 CCs[RTLIB::OLT_F64] = ISD::SETLT;
144 CCs[RTLIB::OLE_F32] = ISD::SETLE;
145 CCs[RTLIB::OLE_F64] = ISD::SETLE;
146 CCs[RTLIB::OGT_F32] = ISD::SETGT;
147 CCs[RTLIB::OGT_F64] = ISD::SETGT;
148 CCs[RTLIB::UO_F32] = ISD::SETNE;
149 CCs[RTLIB::UO_F64] = ISD::SETNE;
150 CCs[RTLIB::O_F32] = ISD::SETEQ;
151 CCs[RTLIB::O_F64] = ISD::SETEQ;
154 TargetLowering::TargetLowering(TargetMachine &tm)
155 : TM(tm), TD(TM.getTargetData()) {
156 assert(ISD::BUILTIN_OP_END <= 156 &&
157 "Fixed size array in TargetLowering is not large enough!");
158 // All operations default to being supported.
159 memset(OpActions, 0, sizeof(OpActions));
160 memset(LoadXActions, 0, sizeof(LoadXActions));
161 memset(TruncStoreActions, 0, sizeof(TruncStoreActions));
162 memset(IndexedModeActions, 0, sizeof(IndexedModeActions));
163 memset(ConvertActions, 0, sizeof(ConvertActions));
165 // Set default actions for various operations.
166 for (unsigned VT = 0; VT != (unsigned)MVT::LAST_VALUETYPE; ++VT) {
167 // Default all indexed load / store to expand.
168 for (unsigned IM = (unsigned)ISD::PRE_INC;
169 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
170 setIndexedLoadAction(IM, (MVT::ValueType)VT, Expand);
171 setIndexedStoreAction(IM, (MVT::ValueType)VT, Expand);
174 // These operations default to expand.
175 setOperationAction(ISD::FGETSIGN, (MVT::ValueType)VT, Expand);
178 // ConstantFP nodes default to expand. Targets can either change this to
179 // Legal, in which case all fp constants are legal, or use addLegalFPImmediate
180 // to optimize expansions for certain constants.
181 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
182 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
183 setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
185 // Default ISD::TRAP to expand (which turns it into abort).
186 setOperationAction(ISD::TRAP, MVT::Other, Expand);
188 IsLittleEndian = TD->isLittleEndian();
189 UsesGlobalOffsetTable = false;
190 ShiftAmountTy = SetCCResultTy = PointerTy = getValueType(TD->getIntPtrType());
191 ShiftAmtHandling = Undefined;
192 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE*sizeof(TargetRegisterClass*));
193 memset(TargetDAGCombineArray, 0, array_lengthof(TargetDAGCombineArray));
194 maxStoresPerMemset = maxStoresPerMemcpy = maxStoresPerMemmove = 8;
195 allowUnalignedMemoryAccesses = false;
196 UseUnderscoreSetJmp = false;
197 UseUnderscoreLongJmp = false;
198 SelectIsExpensive = false;
199 IntDivIsCheap = false;
200 Pow2DivIsCheap = false;
201 StackPointerRegisterToSaveRestore = 0;
202 ExceptionPointerRegister = 0;
203 ExceptionSelectorRegister = 0;
204 SetCCResultContents = UndefinedSetCCResult;
205 SchedPreferenceInfo = SchedulingForLatency;
207 JumpBufAlignment = 0;
208 IfCvtBlockSizeLimit = 2;
210 InitLibcallNames(LibcallRoutineNames);
211 InitCmpLibcallCCs(CmpLibcallCCs);
213 // Tell Legalize whether the assembler supports DEBUG_LOC.
214 if (!TM.getTargetAsmInfo()->hasDotLocAndDotFile())
215 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
218 TargetLowering::~TargetLowering() {}
221 SDOperand TargetLowering::LowerMEMCPY(SDOperand Op, SelectionDAG &DAG) {
222 assert(getSubtarget() && "Subtarget not defined");
223 SDOperand ChainOp = Op.getOperand(0);
224 SDOperand DestOp = Op.getOperand(1);
225 SDOperand SourceOp = Op.getOperand(2);
226 SDOperand CountOp = Op.getOperand(3);
227 SDOperand AlignOp = Op.getOperand(4);
228 SDOperand AlwaysInlineOp = Op.getOperand(5);
230 bool AlwaysInline = (bool)cast<ConstantSDNode>(AlwaysInlineOp)->getValue();
231 unsigned Align = (unsigned)cast<ConstantSDNode>(AlignOp)->getValue();
232 if (Align == 0) Align = 1;
234 // If size is unknown, call memcpy.
235 ConstantSDNode *I = dyn_cast<ConstantSDNode>(CountOp);
237 assert(!AlwaysInline && "Cannot inline copy of unknown size");
238 return LowerMEMCPYCall(ChainOp, DestOp, SourceOp, CountOp, DAG);
241 // If not DWORD aligned or if size is more than threshold, then call memcpy.
242 // The libc version is likely to be faster for the following cases. It can
243 // use the address value and run time information about the CPU.
244 // With glibc 2.6.1 on a core 2, coping an array of 100M longs was 30% faster
245 unsigned Size = I->getValue();
247 (Size <= getSubtarget()->getMaxInlineSizeThreshold() &&
249 return LowerMEMCPYInline(ChainOp, DestOp, SourceOp, Size, Align, DAG);
250 return LowerMEMCPYCall(ChainOp, DestOp, SourceOp, CountOp, DAG);
254 SDOperand TargetLowering::LowerMEMCPYCall(SDOperand Chain,
259 MVT::ValueType IntPtr = getPointerTy();
260 TargetLowering::ArgListTy Args;
261 TargetLowering::ArgListEntry Entry;
262 Entry.Ty = getTargetData()->getIntPtrType();
263 Entry.Node = Dest; Args.push_back(Entry);
264 Entry.Node = Source; Args.push_back(Entry);
265 Entry.Node = Count; Args.push_back(Entry);
266 std::pair<SDOperand,SDOperand> CallResult =
267 LowerCallTo(Chain, Type::VoidTy, false, false, false, CallingConv::C,
268 false, DAG.getExternalSymbol("memcpy", IntPtr), Args, DAG);
269 return CallResult.second;
273 /// computeRegisterProperties - Once all of the register classes are added,
274 /// this allows us to compute derived properties we expose.
275 void TargetLowering::computeRegisterProperties() {
276 assert(MVT::LAST_VALUETYPE <= 32 &&
277 "Too many value types for ValueTypeActions to hold!");
279 // Everything defaults to needing one register.
280 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
281 NumRegistersForVT[i] = 1;
282 RegisterTypeForVT[i] = TransformToType[i] = i;
284 // ...except isVoid, which doesn't need any registers.
285 NumRegistersForVT[MVT::isVoid] = 0;
287 // Find the largest integer register class.
288 unsigned LargestIntReg = MVT::i128;
289 for (; RegClassForVT[LargestIntReg] == 0; --LargestIntReg)
290 assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
292 // Every integer value type larger than this largest register takes twice as
293 // many registers to represent as the previous ValueType.
294 for (MVT::ValueType ExpandedReg = LargestIntReg + 1;
295 MVT::isInteger(ExpandedReg); ++ExpandedReg) {
296 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
297 RegisterTypeForVT[ExpandedReg] = LargestIntReg;
298 TransformToType[ExpandedReg] = ExpandedReg - 1;
299 ValueTypeActions.setTypeAction(ExpandedReg, Expand);
302 // Inspect all of the ValueType's smaller than the largest integer
303 // register to see which ones need promotion.
304 MVT::ValueType LegalIntReg = LargestIntReg;
305 for (MVT::ValueType IntReg = LargestIntReg - 1;
306 IntReg >= MVT::i1; --IntReg) {
307 if (isTypeLegal(IntReg)) {
308 LegalIntReg = IntReg;
310 RegisterTypeForVT[IntReg] = TransformToType[IntReg] = LegalIntReg;
311 ValueTypeActions.setTypeAction(IntReg, Promote);
315 // ppcf128 type is really two f64's.
316 if (!isTypeLegal(MVT::ppcf128)) {
317 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64];
318 RegisterTypeForVT[MVT::ppcf128] = MVT::f64;
319 TransformToType[MVT::ppcf128] = MVT::f64;
320 ValueTypeActions.setTypeAction(MVT::ppcf128, Expand);
323 // Decide how to handle f64. If the target does not have native f64 support,
324 // expand it to i64 and we will be generating soft float library calls.
325 if (!isTypeLegal(MVT::f64)) {
326 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
327 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
328 TransformToType[MVT::f64] = MVT::i64;
329 ValueTypeActions.setTypeAction(MVT::f64, Expand);
332 // Decide how to handle f32. If the target does not have native support for
333 // f32, promote it to f64 if it is legal. Otherwise, expand it to i32.
334 if (!isTypeLegal(MVT::f32)) {
335 if (isTypeLegal(MVT::f64)) {
336 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::f64];
337 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::f64];
338 TransformToType[MVT::f32] = MVT::f64;
339 ValueTypeActions.setTypeAction(MVT::f32, Promote);
341 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
342 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
343 TransformToType[MVT::f32] = MVT::i32;
344 ValueTypeActions.setTypeAction(MVT::f32, Expand);
348 // Loop over all of the vector value types to see which need transformations.
349 for (MVT::ValueType i = MVT::FIRST_VECTOR_VALUETYPE;
350 i <= MVT::LAST_VECTOR_VALUETYPE; ++i) {
351 if (!isTypeLegal(i)) {
352 MVT::ValueType IntermediateVT, RegisterVT;
353 unsigned NumIntermediates;
354 NumRegistersForVT[i] =
355 getVectorTypeBreakdown(i,
356 IntermediateVT, NumIntermediates,
358 RegisterTypeForVT[i] = RegisterVT;
359 TransformToType[i] = MVT::Other; // this isn't actually used
360 ValueTypeActions.setTypeAction(i, Expand);
365 const char *TargetLowering::getTargetNodeName(unsigned Opcode) const {
369 /// getVectorTypeBreakdown - Vector types are broken down into some number of
370 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
371 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
372 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
374 /// This method returns the number of registers needed, and the VT for each
375 /// register. It also returns the VT and quantity of the intermediate values
376 /// before they are promoted/expanded.
378 unsigned TargetLowering::getVectorTypeBreakdown(MVT::ValueType VT,
379 MVT::ValueType &IntermediateVT,
380 unsigned &NumIntermediates,
381 MVT::ValueType &RegisterVT) const {
382 // Figure out the right, legal destination reg to copy into.
383 unsigned NumElts = MVT::getVectorNumElements(VT);
384 MVT::ValueType EltTy = MVT::getVectorElementType(VT);
386 unsigned NumVectorRegs = 1;
388 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
389 // could break down into LHS/RHS like LegalizeDAG does.
390 if (!isPowerOf2_32(NumElts)) {
391 NumVectorRegs = NumElts;
395 // Divide the input until we get to a supported size. This will always
396 // end with a scalar if the target doesn't support vectors.
397 while (NumElts > 1 &&
398 !isTypeLegal(MVT::getVectorType(EltTy, NumElts))) {
403 NumIntermediates = NumVectorRegs;
405 MVT::ValueType NewVT = MVT::getVectorType(EltTy, NumElts);
406 if (!isTypeLegal(NewVT))
408 IntermediateVT = NewVT;
410 MVT::ValueType DestVT = getTypeToTransformTo(NewVT);
412 if (DestVT < NewVT) {
413 // Value is expanded, e.g. i64 -> i16.
414 return NumVectorRegs*(MVT::getSizeInBits(NewVT)/MVT::getSizeInBits(DestVT));
416 // Otherwise, promotion or legal types use the same number of registers as
417 // the vector decimated to the appropriate level.
418 return NumVectorRegs;
424 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
425 /// function arguments in the caller parameter area.
426 unsigned TargetLowering::getByValTypeAlignment(const Type *Ty) const {
427 return Log2_32(TD->getCallFrameTypeAlignment(Ty));
430 SDOperand TargetLowering::getPICJumpTableRelocBase(SDOperand Table,
431 SelectionDAG &DAG) const {
432 if (usesGlobalOffsetTable())
433 return DAG.getNode(ISD::GLOBAL_OFFSET_TABLE, getPointerTy());
437 //===----------------------------------------------------------------------===//
438 // Optimization Methods
439 //===----------------------------------------------------------------------===//
441 /// ShrinkDemandedConstant - Check to see if the specified operand of the
442 /// specified instruction is a constant integer. If so, check to see if there
443 /// are any bits set in the constant that are not demanded. If so, shrink the
444 /// constant and return true.
445 bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDOperand Op,
446 const APInt &Demanded) {
447 // FIXME: ISD::SELECT, ISD::SELECT_CC
448 switch(Op.getOpcode()) {
453 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1)))
454 if (C->getAPIntValue().intersects(~Demanded)) {
455 MVT::ValueType VT = Op.getValueType();
456 SDOperand New = DAG.getNode(Op.getOpcode(), VT, Op.getOperand(0),
457 DAG.getConstant(Demanded &
460 return CombineTo(Op, New);
467 /// SimplifyDemandedBits - Look at Op. At this point, we know that only the
468 /// DemandedMask bits of the result of Op are ever used downstream. If we can
469 /// use this information to simplify Op, create a new simplified DAG node and
470 /// return true, returning the original and new nodes in Old and New. Otherwise,
471 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
472 /// the expression (used to simplify the caller). The KnownZero/One bits may
473 /// only be accurate for those bits in the DemandedMask.
474 bool TargetLowering::SimplifyDemandedBits(SDOperand Op,
475 const APInt &DemandedMask,
478 TargetLoweringOpt &TLO,
479 unsigned Depth) const {
480 unsigned BitWidth = DemandedMask.getBitWidth();
481 assert(Op.getValueSizeInBits() == BitWidth &&
482 "Mask size mismatches value type size!");
483 APInt NewMask = DemandedMask;
485 // Don't know anything.
486 KnownZero = KnownOne = APInt(BitWidth, 0);
488 // Other users may use these bits.
489 if (!Op.Val->hasOneUse()) {
491 // If not at the root, Just compute the KnownZero/KnownOne bits to
492 // simplify things downstream.
493 TLO.DAG.ComputeMaskedBits(Op, DemandedMask, KnownZero, KnownOne, Depth);
496 // If this is the root being simplified, allow it to have multiple uses,
497 // just set the NewMask to all bits.
498 NewMask = APInt::getAllOnesValue(BitWidth);
499 } else if (DemandedMask == 0) {
500 // Not demanding any bits from Op.
501 if (Op.getOpcode() != ISD::UNDEF)
502 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::UNDEF, Op.getValueType()));
504 } else if (Depth == 6) { // Limit search depth.
508 APInt KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut;
509 switch (Op.getOpcode()) {
511 // We know all of the bits for a constant!
512 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue() & NewMask;
513 KnownZero = ~KnownOne & NewMask;
514 return false; // Don't fall through, will infinitely loop.
516 // If the RHS is a constant, check to see if the LHS would be zero without
517 // using the bits from the RHS. Below, we use knowledge about the RHS to
518 // simplify the LHS, here we're using information from the LHS to simplify
520 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
521 APInt LHSZero, LHSOne;
522 TLO.DAG.ComputeMaskedBits(Op.getOperand(0), NewMask,
523 LHSZero, LHSOne, Depth+1);
524 // If the LHS already has zeros where RHSC does, this and is dead.
525 if ((LHSZero & NewMask) == (~RHSC->getAPIntValue() & NewMask))
526 return TLO.CombineTo(Op, Op.getOperand(0));
527 // If any of the set bits in the RHS are known zero on the LHS, shrink
529 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & NewMask))
533 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
534 KnownOne, TLO, Depth+1))
536 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
537 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownZero & NewMask,
538 KnownZero2, KnownOne2, TLO, Depth+1))
540 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
542 // If all of the demanded bits are known one on one side, return the other.
543 // These bits cannot contribute to the result of the 'and'.
544 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
545 return TLO.CombineTo(Op, Op.getOperand(0));
546 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
547 return TLO.CombineTo(Op, Op.getOperand(1));
548 // If all of the demanded bits in the inputs are known zeros, return zero.
549 if ((NewMask & (KnownZero|KnownZero2)) == NewMask)
550 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, Op.getValueType()));
551 // If the RHS is a constant, see if we can simplify it.
552 if (TLO.ShrinkDemandedConstant(Op, ~KnownZero2 & NewMask))
555 // Output known-1 bits are only known if set in both the LHS & RHS.
556 KnownOne &= KnownOne2;
557 // Output known-0 are known to be clear if zero in either the LHS | RHS.
558 KnownZero |= KnownZero2;
561 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
562 KnownOne, TLO, Depth+1))
564 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
565 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownOne & NewMask,
566 KnownZero2, KnownOne2, TLO, Depth+1))
568 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
570 // If all of the demanded bits are known zero on one side, return the other.
571 // These bits cannot contribute to the result of the 'or'.
572 if ((NewMask & ~KnownOne2 & KnownZero) == (~KnownOne2 & NewMask))
573 return TLO.CombineTo(Op, Op.getOperand(0));
574 if ((NewMask & ~KnownOne & KnownZero2) == (~KnownOne & NewMask))
575 return TLO.CombineTo(Op, Op.getOperand(1));
576 // If all of the potentially set bits on one side are known to be set on
577 // the other side, just use the 'other' side.
578 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
579 return TLO.CombineTo(Op, Op.getOperand(0));
580 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
581 return TLO.CombineTo(Op, Op.getOperand(1));
582 // If the RHS is a constant, see if we can simplify it.
583 if (TLO.ShrinkDemandedConstant(Op, NewMask))
586 // Output known-0 bits are only known if clear in both the LHS & RHS.
587 KnownZero &= KnownZero2;
588 // Output known-1 are known to be set if set in either the LHS | RHS.
589 KnownOne |= KnownOne2;
592 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
593 KnownOne, TLO, Depth+1))
595 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
596 if (SimplifyDemandedBits(Op.getOperand(0), NewMask, KnownZero2,
597 KnownOne2, TLO, Depth+1))
599 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
601 // If all of the demanded bits are known zero on one side, return the other.
602 // These bits cannot contribute to the result of the 'xor'.
603 if ((KnownZero & NewMask) == NewMask)
604 return TLO.CombineTo(Op, Op.getOperand(0));
605 if ((KnownZero2 & NewMask) == NewMask)
606 return TLO.CombineTo(Op, Op.getOperand(1));
608 // If all of the unknown bits are known to be zero on one side or the other
609 // (but not both) turn this into an *inclusive* or.
610 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
611 if ((NewMask & ~KnownZero & ~KnownZero2) == 0)
612 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, Op.getValueType(),
616 // Output known-0 bits are known if clear or set in both the LHS & RHS.
617 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
618 // Output known-1 are known to be set if set in only one of the LHS, RHS.
619 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
621 // If all of the demanded bits on one side are known, and all of the set
622 // bits on that side are also known to be set on the other side, turn this
623 // into an AND, as we know the bits will be cleared.
624 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
625 if ((NewMask & (KnownZero|KnownOne)) == NewMask) { // all known
626 if ((KnownOne & KnownOne2) == KnownOne) {
627 MVT::ValueType VT = Op.getValueType();
628 SDOperand ANDC = TLO.DAG.getConstant(~KnownOne & NewMask, VT);
629 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, VT, Op.getOperand(0),
634 // If the RHS is a constant, see if we can simplify it.
635 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
636 if (TLO.ShrinkDemandedConstant(Op, NewMask))
639 KnownZero = KnownZeroOut;
640 KnownOne = KnownOneOut;
643 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero,
644 KnownOne, TLO, Depth+1))
646 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero2,
647 KnownOne2, TLO, Depth+1))
649 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
650 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
652 // If the operands are constants, see if we can simplify them.
653 if (TLO.ShrinkDemandedConstant(Op, NewMask))
656 // Only known if known in both the LHS and RHS.
657 KnownOne &= KnownOne2;
658 KnownZero &= KnownZero2;
661 if (SimplifyDemandedBits(Op.getOperand(3), NewMask, KnownZero,
662 KnownOne, TLO, Depth+1))
664 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero2,
665 KnownOne2, TLO, Depth+1))
667 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
668 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
670 // If the operands are constants, see if we can simplify them.
671 if (TLO.ShrinkDemandedConstant(Op, NewMask))
674 // Only known if known in both the LHS and RHS.
675 KnownOne &= KnownOne2;
676 KnownZero &= KnownZero2;
679 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
680 unsigned ShAmt = SA->getValue();
681 SDOperand InOp = Op.getOperand(0);
683 // If the shift count is an invalid immediate, don't do anything.
684 if (ShAmt >= BitWidth)
687 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a
688 // single shift. We can do this if the bottom bits (which are shifted
689 // out) are never demanded.
690 if (InOp.getOpcode() == ISD::SRL &&
691 isa<ConstantSDNode>(InOp.getOperand(1))) {
692 if (ShAmt && (NewMask & APInt::getLowBitsSet(BitWidth, ShAmt)) == 0) {
693 unsigned C1 = cast<ConstantSDNode>(InOp.getOperand(1))->getValue();
694 unsigned Opc = ISD::SHL;
702 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
703 MVT::ValueType VT = Op.getValueType();
704 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, VT,
705 InOp.getOperand(0), NewSA));
709 if (SimplifyDemandedBits(Op.getOperand(0), NewMask.lshr(ShAmt),
710 KnownZero, KnownOne, TLO, Depth+1))
712 KnownZero <<= SA->getValue();
713 KnownOne <<= SA->getValue();
714 // low bits known zero.
715 KnownZero |= APInt::getLowBitsSet(BitWidth, SA->getValue());
719 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
720 MVT::ValueType VT = Op.getValueType();
721 unsigned ShAmt = SA->getValue();
722 unsigned VTSize = MVT::getSizeInBits(VT);
723 SDOperand InOp = Op.getOperand(0);
725 // If the shift count is an invalid immediate, don't do anything.
726 if (ShAmt >= BitWidth)
729 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a
730 // single shift. We can do this if the top bits (which are shifted out)
731 // are never demanded.
732 if (InOp.getOpcode() == ISD::SHL &&
733 isa<ConstantSDNode>(InOp.getOperand(1))) {
734 if (ShAmt && (NewMask & APInt::getHighBitsSet(VTSize, ShAmt)) == 0) {
735 unsigned C1 = cast<ConstantSDNode>(InOp.getOperand(1))->getValue();
736 unsigned Opc = ISD::SRL;
744 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
745 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, VT,
746 InOp.getOperand(0), NewSA));
750 // Compute the new bits that are at the top now.
751 if (SimplifyDemandedBits(InOp, (NewMask << ShAmt),
752 KnownZero, KnownOne, TLO, Depth+1))
754 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
755 KnownZero = KnownZero.lshr(ShAmt);
756 KnownOne = KnownOne.lshr(ShAmt);
758 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
759 KnownZero |= HighBits; // High bits known zero.
763 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
764 MVT::ValueType VT = Op.getValueType();
765 unsigned ShAmt = SA->getValue();
767 // If the shift count is an invalid immediate, don't do anything.
768 if (ShAmt >= BitWidth)
771 APInt InDemandedMask = (NewMask << ShAmt);
773 // If any of the demanded bits are produced by the sign extension, we also
774 // demand the input sign bit.
775 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
776 if (HighBits.intersects(NewMask))
777 InDemandedMask |= APInt::getSignBit(MVT::getSizeInBits(VT));
779 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask,
780 KnownZero, KnownOne, TLO, Depth+1))
782 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
783 KnownZero = KnownZero.lshr(ShAmt);
784 KnownOne = KnownOne.lshr(ShAmt);
786 // Handle the sign bit, adjusted to where it is now in the mask.
787 APInt SignBit = APInt::getSignBit(BitWidth).lshr(ShAmt);
789 // If the input sign bit is known to be zero, or if none of the top bits
790 // are demanded, turn this into an unsigned shift right.
791 if (KnownZero.intersects(SignBit) || (HighBits & ~NewMask) == HighBits) {
792 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, VT, Op.getOperand(0),
794 } else if (KnownOne.intersects(SignBit)) { // New bits are known one.
795 KnownOne |= HighBits;
799 case ISD::SIGN_EXTEND_INREG: {
800 MVT::ValueType EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
802 // Sign extension. Compute the demanded bits in the result that are not
803 // present in the input.
804 APInt NewBits = APInt::getHighBitsSet(BitWidth,
805 BitWidth - MVT::getSizeInBits(EVT)) &
808 // If none of the extended bits are demanded, eliminate the sextinreg.
810 return TLO.CombineTo(Op, Op.getOperand(0));
812 APInt InSignBit = APInt::getSignBit(MVT::getSizeInBits(EVT));
813 InSignBit.zext(BitWidth);
814 APInt InputDemandedBits = APInt::getLowBitsSet(BitWidth,
815 MVT::getSizeInBits(EVT)) &
818 // Since the sign extended bits are demanded, we know that the sign
820 InputDemandedBits |= InSignBit;
822 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits,
823 KnownZero, KnownOne, TLO, Depth+1))
825 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
827 // If the sign bit of the input is known set or clear, then we know the
828 // top bits of the result.
830 // If the input sign bit is known zero, convert this into a zero extension.
831 if (KnownZero.intersects(InSignBit))
832 return TLO.CombineTo(Op,
833 TLO.DAG.getZeroExtendInReg(Op.getOperand(0), EVT));
835 if (KnownOne.intersects(InSignBit)) { // Input sign bit known set
837 KnownZero &= ~NewBits;
838 } else { // Input sign bit unknown
839 KnownZero &= ~NewBits;
840 KnownOne &= ~NewBits;
844 case ISD::ZERO_EXTEND: {
845 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits();
846 APInt InMask = NewMask;
847 InMask.trunc(OperandBitWidth);
849 // If none of the top bits are demanded, convert this into an any_extend.
851 APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask;
852 if (!NewBits.intersects(NewMask))
853 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND,
857 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
858 KnownZero, KnownOne, TLO, Depth+1))
860 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
861 KnownZero.zext(BitWidth);
862 KnownOne.zext(BitWidth);
863 KnownZero |= NewBits;
866 case ISD::SIGN_EXTEND: {
867 MVT::ValueType InVT = Op.getOperand(0).getValueType();
868 unsigned InBits = MVT::getSizeInBits(InVT);
869 APInt InMask = APInt::getLowBitsSet(BitWidth, InBits);
870 APInt InSignBit = APInt::getLowBitsSet(BitWidth, InBits);
871 APInt NewBits = ~InMask & NewMask;
873 // If none of the top bits are demanded, convert this into an any_extend.
875 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND,Op.getValueType(),
878 // Since some of the sign extended bits are demanded, we know that the sign
880 APInt InDemandedBits = InMask & NewMask;
881 InDemandedBits |= InSignBit;
882 InDemandedBits.trunc(InBits);
884 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero,
885 KnownOne, TLO, Depth+1))
887 KnownZero.zext(BitWidth);
888 KnownOne.zext(BitWidth);
890 // If the sign bit is known zero, convert this to a zero extend.
891 if (KnownZero.intersects(InSignBit))
892 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND,
896 // If the sign bit is known one, the top bits match.
897 if (KnownOne.intersects(InSignBit)) {
899 KnownZero &= ~NewBits;
900 } else { // Otherwise, top bits aren't known.
901 KnownOne &= ~NewBits;
902 KnownZero &= ~NewBits;
906 case ISD::ANY_EXTEND: {
907 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits();
908 APInt InMask = NewMask;
909 InMask.trunc(OperandBitWidth);
910 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
911 KnownZero, KnownOne, TLO, Depth+1))
913 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
914 KnownZero.zext(BitWidth);
915 KnownOne.zext(BitWidth);
918 case ISD::TRUNCATE: {
919 // Simplify the input, using demanded bit information, and compute the known
920 // zero/one bits live out.
921 APInt TruncMask = NewMask;
922 TruncMask.zext(Op.getOperand(0).getValueSizeInBits());
923 if (SimplifyDemandedBits(Op.getOperand(0), TruncMask,
924 KnownZero, KnownOne, TLO, Depth+1))
926 KnownZero.trunc(BitWidth);
927 KnownOne.trunc(BitWidth);
929 // If the input is only used by this truncate, see if we can shrink it based
930 // on the known demanded bits.
931 if (Op.getOperand(0).Val->hasOneUse()) {
932 SDOperand In = Op.getOperand(0);
933 unsigned InBitWidth = In.getValueSizeInBits();
934 switch (In.getOpcode()) {
937 // Shrink SRL by a constant if none of the high bits shifted in are
939 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1))){
940 APInt HighBits = APInt::getHighBitsSet(InBitWidth,
941 InBitWidth - BitWidth);
942 HighBits = HighBits.lshr(ShAmt->getValue());
943 HighBits.trunc(BitWidth);
945 if (ShAmt->getValue() < BitWidth && !(HighBits & NewMask)) {
946 // None of the shifted in bits are needed. Add a truncate of the
947 // shift input, then shift it.
948 SDOperand NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE,
951 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL,Op.getValueType(),
952 NewTrunc, In.getOperand(1)));
959 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
962 case ISD::AssertZext: {
963 MVT::ValueType VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
964 APInt InMask = APInt::getLowBitsSet(BitWidth,
965 MVT::getSizeInBits(VT));
966 if (SimplifyDemandedBits(Op.getOperand(0), InMask & NewMask,
967 KnownZero, KnownOne, TLO, Depth+1))
969 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
970 KnownZero |= ~InMask & NewMask;
973 case ISD::BIT_CONVERT:
975 // If this is an FP->Int bitcast and if the sign bit is the only thing that
976 // is demanded, turn this into a FGETSIGN.
977 if (NewMask == MVT::getIntVTSignBit(Op.getValueType()) &&
978 MVT::isFloatingPoint(Op.getOperand(0).getValueType()) &&
979 !MVT::isVector(Op.getOperand(0).getValueType())) {
980 // Only do this xform if FGETSIGN is valid or if before legalize.
981 if (!TLO.AfterLegalize ||
982 isOperationLegal(ISD::FGETSIGN, Op.getValueType())) {
983 // Make a FGETSIGN + SHL to move the sign bit into the appropriate
984 // place. We expect the SHL to be eliminated by other optimizations.
985 SDOperand Sign = TLO.DAG.getNode(ISD::FGETSIGN, Op.getValueType(),
987 unsigned ShVal = MVT::getSizeInBits(Op.getValueType())-1;
988 SDOperand ShAmt = TLO.DAG.getConstant(ShVal, getShiftAmountTy());
989 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, Op.getValueType(),
997 case ISD::INTRINSIC_WO_CHAIN:
998 case ISD::INTRINSIC_W_CHAIN:
999 case ISD::INTRINSIC_VOID:
1006 // Just use ComputeMaskedBits to compute output bits.
1007 TLO.DAG.ComputeMaskedBits(Op, NewMask, KnownZero, KnownOne, Depth);
1011 // If we know the value of all of the demanded bits, return this as a
1013 if ((NewMask & (KnownZero|KnownOne)) == NewMask)
1014 return TLO.CombineTo(Op, TLO.DAG.getConstant(KnownOne, Op.getValueType()));
1019 /// computeMaskedBitsForTargetNode - Determine which of the bits specified
1020 /// in Mask are known to be either zero or one and return them in the
1021 /// KnownZero/KnownOne bitsets.
1022 void TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
1026 const SelectionDAG &DAG,
1027 unsigned Depth) const {
1028 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1029 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1030 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1031 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1032 "Should use MaskedValueIsZero if you don't know whether Op"
1033 " is a target node!");
1034 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
1037 /// ComputeNumSignBitsForTargetNode - This method can be implemented by
1038 /// targets that want to expose additional information about sign bits to the
1040 unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDOperand Op,
1041 unsigned Depth) const {
1042 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1043 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1044 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1045 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1046 "Should use ComputeNumSignBits if you don't know whether Op"
1047 " is a target node!");
1052 /// SimplifySetCC - Try to simplify a setcc built with the specified operands
1053 /// and cc. If it is unable to simplify it, return a null SDOperand.
1055 TargetLowering::SimplifySetCC(MVT::ValueType VT, SDOperand N0, SDOperand N1,
1056 ISD::CondCode Cond, bool foldBooleans,
1057 DAGCombinerInfo &DCI) const {
1058 SelectionDAG &DAG = DCI.DAG;
1060 // These setcc operations always fold.
1064 case ISD::SETFALSE2: return DAG.getConstant(0, VT);
1066 case ISD::SETTRUE2: return DAG.getConstant(1, VT);
1069 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val)) {
1070 uint64_t C1 = N1C->getValue();
1071 if (isa<ConstantSDNode>(N0.Val)) {
1072 return DAG.FoldSetCC(VT, N0, N1, Cond);
1074 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an
1075 // equality comparison, then we're just comparing whether X itself is
1077 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) &&
1078 N0.getOperand(0).getOpcode() == ISD::CTLZ &&
1079 N0.getOperand(1).getOpcode() == ISD::Constant) {
1080 unsigned ShAmt = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1081 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1082 ShAmt == Log2_32(MVT::getSizeInBits(N0.getValueType()))) {
1083 if ((C1 == 0) == (Cond == ISD::SETEQ)) {
1084 // (srl (ctlz x), 5) == 0 -> X != 0
1085 // (srl (ctlz x), 5) != 1 -> X != 0
1088 // (srl (ctlz x), 5) != 0 -> X == 0
1089 // (srl (ctlz x), 5) == 1 -> X == 0
1092 SDOperand Zero = DAG.getConstant(0, N0.getValueType());
1093 return DAG.getSetCC(VT, N0.getOperand(0).getOperand(0),
1098 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
1099 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
1100 unsigned InSize = MVT::getSizeInBits(N0.getOperand(0).getValueType());
1102 // If the comparison constant has bits in the upper part, the
1103 // zero-extended value could never match.
1104 if (C1 & (~0ULL << InSize)) {
1105 unsigned VSize = MVT::getSizeInBits(N0.getValueType());
1109 case ISD::SETEQ: return DAG.getConstant(0, VT);
1112 case ISD::SETNE: return DAG.getConstant(1, VT);
1115 // True if the sign bit of C1 is set.
1116 return DAG.getConstant((C1 & (1ULL << (VSize-1))) != 0, VT);
1119 // True if the sign bit of C1 isn't set.
1120 return DAG.getConstant((C1 & (1ULL << (VSize-1))) == 0, VT);
1126 // Otherwise, we can perform the comparison with the low bits.
1134 return DAG.getSetCC(VT, N0.getOperand(0),
1135 DAG.getConstant(C1, N0.getOperand(0).getValueType()),
1138 break; // todo, be more careful with signed comparisons
1140 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
1141 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1142 MVT::ValueType ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
1143 unsigned ExtSrcTyBits = MVT::getSizeInBits(ExtSrcTy);
1144 MVT::ValueType ExtDstTy = N0.getValueType();
1145 unsigned ExtDstTyBits = MVT::getSizeInBits(ExtDstTy);
1147 // If the extended part has any inconsistent bits, it cannot ever
1148 // compare equal. In other words, they have to be all ones or all
1151 (~0ULL >> (64-ExtSrcTyBits)) & (~0ULL << (ExtDstTyBits-1));
1152 if ((C1 & ExtBits) != 0 && (C1 & ExtBits) != ExtBits)
1153 return DAG.getConstant(Cond == ISD::SETNE, VT);
1156 MVT::ValueType Op0Ty = N0.getOperand(0).getValueType();
1157 if (Op0Ty == ExtSrcTy) {
1158 ZextOp = N0.getOperand(0);
1160 int64_t Imm = ~0ULL >> (64-ExtSrcTyBits);
1161 ZextOp = DAG.getNode(ISD::AND, Op0Ty, N0.getOperand(0),
1162 DAG.getConstant(Imm, Op0Ty));
1164 if (!DCI.isCalledByLegalizer())
1165 DCI.AddToWorklist(ZextOp.Val);
1166 // Otherwise, make this a use of a zext.
1167 return DAG.getSetCC(VT, ZextOp,
1168 DAG.getConstant(C1 & (~0ULL>>(64-ExtSrcTyBits)),
1171 } else if ((N1C->getValue() == 0 || N1C->getValue() == 1) &&
1172 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1174 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC
1175 if (N0.getOpcode() == ISD::SETCC) {
1176 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getValue() != 1);
1180 // Invert the condition.
1181 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
1182 CC = ISD::getSetCCInverse(CC,
1183 MVT::isInteger(N0.getOperand(0).getValueType()));
1184 return DAG.getSetCC(VT, N0.getOperand(0), N0.getOperand(1), CC);
1187 if ((N0.getOpcode() == ISD::XOR ||
1188 (N0.getOpcode() == ISD::AND &&
1189 N0.getOperand(0).getOpcode() == ISD::XOR &&
1190 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
1191 isa<ConstantSDNode>(N0.getOperand(1)) &&
1192 cast<ConstantSDNode>(N0.getOperand(1))->getValue() == 1) {
1193 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We
1194 // can only do this if the top bits are known zero.
1195 unsigned BitWidth = N0.getValueSizeInBits();
1196 if (DAG.MaskedValueIsZero(N0,
1197 APInt::getHighBitsSet(BitWidth,
1199 // Okay, get the un-inverted input value.
1201 if (N0.getOpcode() == ISD::XOR)
1202 Val = N0.getOperand(0);
1204 assert(N0.getOpcode() == ISD::AND &&
1205 N0.getOperand(0).getOpcode() == ISD::XOR);
1206 // ((X^1)&1)^1 -> X & 1
1207 Val = DAG.getNode(ISD::AND, N0.getValueType(),
1208 N0.getOperand(0).getOperand(0),
1211 return DAG.getSetCC(VT, Val, N1,
1212 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
1217 uint64_t MinVal, MaxVal;
1218 unsigned OperandBitSize = MVT::getSizeInBits(N1C->getValueType(0));
1219 if (ISD::isSignedIntSetCC(Cond)) {
1220 MinVal = 1ULL << (OperandBitSize-1);
1221 if (OperandBitSize != 1) // Avoid X >> 64, which is undefined.
1222 MaxVal = ~0ULL >> (65-OperandBitSize);
1227 MaxVal = ~0ULL >> (64-OperandBitSize);
1230 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
1231 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
1232 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true
1233 --C1; // X >= C0 --> X > (C0-1)
1234 return DAG.getSetCC(VT, N0, DAG.getConstant(C1, N1.getValueType()),
1235 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
1238 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
1239 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true
1240 ++C1; // X <= C0 --> X < (C0+1)
1241 return DAG.getSetCC(VT, N0, DAG.getConstant(C1, N1.getValueType()),
1242 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
1245 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
1246 return DAG.getConstant(0, VT); // X < MIN --> false
1247 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal)
1248 return DAG.getConstant(1, VT); // X >= MIN --> true
1249 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal)
1250 return DAG.getConstant(0, VT); // X > MAX --> false
1251 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal)
1252 return DAG.getConstant(1, VT); // X <= MAX --> true
1254 // Canonicalize setgt X, Min --> setne X, Min
1255 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
1256 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
1257 // Canonicalize setlt X, Max --> setne X, Max
1258 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
1259 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
1261 // If we have setult X, 1, turn it into seteq X, 0
1262 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
1263 return DAG.getSetCC(VT, N0, DAG.getConstant(MinVal, N0.getValueType()),
1265 // If we have setugt X, Max-1, turn it into seteq X, Max
1266 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
1267 return DAG.getSetCC(VT, N0, DAG.getConstant(MaxVal, N0.getValueType()),
1270 // If we have "setcc X, C0", check to see if we can shrink the immediate
1273 // SETUGT X, SINTMAX -> SETLT X, 0
1274 if (Cond == ISD::SETUGT && OperandBitSize != 1 &&
1275 C1 == (~0ULL >> (65-OperandBitSize)))
1276 return DAG.getSetCC(VT, N0, DAG.getConstant(0, N1.getValueType()),
1279 // FIXME: Implement the rest of these.
1281 // Fold bit comparisons when we can.
1282 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1283 VT == N0.getValueType() && N0.getOpcode() == ISD::AND)
1284 if (ConstantSDNode *AndRHS =
1285 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1286 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
1287 // Perform the xform if the AND RHS is a single bit.
1288 if (isPowerOf2_64(AndRHS->getValue())) {
1289 return DAG.getNode(ISD::SRL, VT, N0,
1290 DAG.getConstant(Log2_64(AndRHS->getValue()),
1291 getShiftAmountTy()));
1293 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getValue()) {
1294 // (X & 8) == 8 --> (X & 8) >> 3
1295 // Perform the xform if C1 is a single bit.
1296 if (isPowerOf2_64(C1)) {
1297 return DAG.getNode(ISD::SRL, VT, N0,
1298 DAG.getConstant(Log2_64(C1), getShiftAmountTy()));
1303 } else if (isa<ConstantSDNode>(N0.Val)) {
1304 // Ensure that the constant occurs on the RHS.
1305 return DAG.getSetCC(VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
1308 if (isa<ConstantFPSDNode>(N0.Val)) {
1309 // Constant fold or commute setcc.
1310 SDOperand O = DAG.FoldSetCC(VT, N0, N1, Cond);
1311 if (O.Val) return O;
1312 } else if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1.Val)) {
1313 // If the RHS of an FP comparison is a constant, simplify it away in
1315 if (CFP->getValueAPF().isNaN()) {
1316 // If an operand is known to be a nan, we can fold it.
1317 switch (ISD::getUnorderedFlavor(Cond)) {
1318 default: assert(0 && "Unknown flavor!");
1319 case 0: // Known false.
1320 return DAG.getConstant(0, VT);
1321 case 1: // Known true.
1322 return DAG.getConstant(1, VT);
1323 case 2: // Undefined.
1324 return DAG.getNode(ISD::UNDEF, VT);
1328 // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the
1329 // constant if knowing that the operand is non-nan is enough. We prefer to
1330 // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to
1332 if (Cond == ISD::SETO || Cond == ISD::SETUO)
1333 return DAG.getSetCC(VT, N0, N0, Cond);
1337 // We can always fold X == X for integer setcc's.
1338 if (MVT::isInteger(N0.getValueType()))
1339 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
1340 unsigned UOF = ISD::getUnorderedFlavor(Cond);
1341 if (UOF == 2) // FP operators that are undefined on NaNs.
1342 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
1343 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
1344 return DAG.getConstant(UOF, VT);
1345 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
1346 // if it is not already.
1347 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
1348 if (NewCond != Cond)
1349 return DAG.getSetCC(VT, N0, N1, NewCond);
1352 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1353 MVT::isInteger(N0.getValueType())) {
1354 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
1355 N0.getOpcode() == ISD::XOR) {
1356 // Simplify (X+Y) == (X+Z) --> Y == Z
1357 if (N0.getOpcode() == N1.getOpcode()) {
1358 if (N0.getOperand(0) == N1.getOperand(0))
1359 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(1), Cond);
1360 if (N0.getOperand(1) == N1.getOperand(1))
1361 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(0), Cond);
1362 if (DAG.isCommutativeBinOp(N0.getOpcode())) {
1363 // If X op Y == Y op X, try other combinations.
1364 if (N0.getOperand(0) == N1.getOperand(1))
1365 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(0), Cond);
1366 if (N0.getOperand(1) == N1.getOperand(0))
1367 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(1), Cond);
1371 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) {
1372 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1373 // Turn (X+C1) == C2 --> X == C2-C1
1374 if (N0.getOpcode() == ISD::ADD && N0.Val->hasOneUse()) {
1375 return DAG.getSetCC(VT, N0.getOperand(0),
1376 DAG.getConstant(RHSC->getValue()-LHSR->getValue(),
1377 N0.getValueType()), Cond);
1380 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
1381 if (N0.getOpcode() == ISD::XOR)
1382 // If we know that all of the inverted bits are zero, don't bother
1383 // performing the inversion.
1384 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue()))
1386 DAG.getSetCC(VT, N0.getOperand(0),
1387 DAG.getConstant(LHSR->getAPIntValue() ^
1388 RHSC->getAPIntValue(),
1393 // Turn (C1-X) == C2 --> X == C1-C2
1394 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
1395 if (N0.getOpcode() == ISD::SUB && N0.Val->hasOneUse()) {
1397 DAG.getSetCC(VT, N0.getOperand(1),
1398 DAG.getConstant(SUBC->getAPIntValue() -
1399 RHSC->getAPIntValue(),
1406 // Simplify (X+Z) == X --> Z == 0
1407 if (N0.getOperand(0) == N1)
1408 return DAG.getSetCC(VT, N0.getOperand(1),
1409 DAG.getConstant(0, N0.getValueType()), Cond);
1410 if (N0.getOperand(1) == N1) {
1411 if (DAG.isCommutativeBinOp(N0.getOpcode()))
1412 return DAG.getSetCC(VT, N0.getOperand(0),
1413 DAG.getConstant(0, N0.getValueType()), Cond);
1414 else if (N0.Val->hasOneUse()) {
1415 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
1416 // (Z-X) == X --> Z == X<<1
1417 SDOperand SH = DAG.getNode(ISD::SHL, N1.getValueType(),
1419 DAG.getConstant(1, getShiftAmountTy()));
1420 if (!DCI.isCalledByLegalizer())
1421 DCI.AddToWorklist(SH.Val);
1422 return DAG.getSetCC(VT, N0.getOperand(0), SH, Cond);
1427 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
1428 N1.getOpcode() == ISD::XOR) {
1429 // Simplify X == (X+Z) --> Z == 0
1430 if (N1.getOperand(0) == N0) {
1431 return DAG.getSetCC(VT, N1.getOperand(1),
1432 DAG.getConstant(0, N1.getValueType()), Cond);
1433 } else if (N1.getOperand(1) == N0) {
1434 if (DAG.isCommutativeBinOp(N1.getOpcode())) {
1435 return DAG.getSetCC(VT, N1.getOperand(0),
1436 DAG.getConstant(0, N1.getValueType()), Cond);
1437 } else if (N1.Val->hasOneUse()) {
1438 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
1439 // X == (Z-X) --> X<<1 == Z
1440 SDOperand SH = DAG.getNode(ISD::SHL, N1.getValueType(), N0,
1441 DAG.getConstant(1, getShiftAmountTy()));
1442 if (!DCI.isCalledByLegalizer())
1443 DCI.AddToWorklist(SH.Val);
1444 return DAG.getSetCC(VT, SH, N1.getOperand(0), Cond);
1450 // Fold away ALL boolean setcc's.
1452 if (N0.getValueType() == MVT::i1 && foldBooleans) {
1454 default: assert(0 && "Unknown integer setcc!");
1455 case ISD::SETEQ: // X == Y -> (X^Y)^1
1456 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
1457 N0 = DAG.getNode(ISD::XOR, MVT::i1, Temp, DAG.getConstant(1, MVT::i1));
1458 if (!DCI.isCalledByLegalizer())
1459 DCI.AddToWorklist(Temp.Val);
1461 case ISD::SETNE: // X != Y --> (X^Y)
1462 N0 = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
1464 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> X^1 & Y
1465 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> X^1 & Y
1466 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
1467 N0 = DAG.getNode(ISD::AND, MVT::i1, N1, Temp);
1468 if (!DCI.isCalledByLegalizer())
1469 DCI.AddToWorklist(Temp.Val);
1471 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> Y^1 & X
1472 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> Y^1 & X
1473 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
1474 N0 = DAG.getNode(ISD::AND, MVT::i1, N0, Temp);
1475 if (!DCI.isCalledByLegalizer())
1476 DCI.AddToWorklist(Temp.Val);
1478 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> X^1 | Y
1479 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> X^1 | Y
1480 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
1481 N0 = DAG.getNode(ISD::OR, MVT::i1, N1, Temp);
1482 if (!DCI.isCalledByLegalizer())
1483 DCI.AddToWorklist(Temp.Val);
1485 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> Y^1 | X
1486 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> Y^1 | X
1487 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
1488 N0 = DAG.getNode(ISD::OR, MVT::i1, N0, Temp);
1491 if (VT != MVT::i1) {
1492 if (!DCI.isCalledByLegalizer())
1493 DCI.AddToWorklist(N0.Val);
1494 // FIXME: If running after legalize, we probably can't do this.
1495 N0 = DAG.getNode(ISD::ZERO_EXTEND, VT, N0);
1500 // Could not fold it.
1504 SDOperand TargetLowering::
1505 PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
1506 // Default implementation: no optimization.
1510 //===----------------------------------------------------------------------===//
1511 // Inline Assembler Implementation Methods
1512 //===----------------------------------------------------------------------===//
1514 TargetLowering::ConstraintType
1515 TargetLowering::getConstraintType(const std::string &Constraint) const {
1516 // FIXME: lots more standard ones to handle.
1517 if (Constraint.size() == 1) {
1518 switch (Constraint[0]) {
1520 case 'r': return C_RegisterClass;
1522 case 'o': // offsetable
1523 case 'V': // not offsetable
1525 case 'i': // Simple Integer or Relocatable Constant
1526 case 'n': // Simple Integer
1527 case 's': // Relocatable Constant
1528 case 'X': // Allow ANY value.
1529 case 'I': // Target registers.
1541 if (Constraint.size() > 1 && Constraint[0] == '{' &&
1542 Constraint[Constraint.size()-1] == '}')
1547 /// LowerXConstraint - try to replace an X constraint, which matches anything,
1548 /// with another that has more specific requirements based on the type of the
1549 /// corresponding operand.
1550 void TargetLowering::lowerXConstraint(MVT::ValueType ConstraintVT,
1551 std::string& s) const {
1552 if (MVT::isInteger(ConstraintVT))
1554 else if (MVT::isFloatingPoint(ConstraintVT))
1555 s = "f"; // works for many targets
1560 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
1561 /// vector. If it is invalid, don't add anything to Ops.
1562 void TargetLowering::LowerAsmOperandForConstraint(SDOperand Op,
1563 char ConstraintLetter,
1564 std::vector<SDOperand> &Ops,
1565 SelectionDAG &DAG) {
1566 switch (ConstraintLetter) {
1568 case 'X': // Allows any operand; labels (basic block) use this.
1569 if (Op.getOpcode() == ISD::BasicBlock) {
1574 case 'i': // Simple Integer or Relocatable Constant
1575 case 'n': // Simple Integer
1576 case 's': { // Relocatable Constant
1577 // These operands are interested in values of the form (GV+C), where C may
1578 // be folded in as an offset of GV, or it may be explicitly added. Also, it
1579 // is possible and fine if either GV or C are missing.
1580 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
1581 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
1583 // If we have "(add GV, C)", pull out GV/C
1584 if (Op.getOpcode() == ISD::ADD) {
1585 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
1586 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
1587 if (C == 0 || GA == 0) {
1588 C = dyn_cast<ConstantSDNode>(Op.getOperand(0));
1589 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1));
1591 if (C == 0 || GA == 0)
1595 // If we find a valid operand, map to the TargetXXX version so that the
1596 // value itself doesn't get selected.
1597 if (GA) { // Either &GV or &GV+C
1598 if (ConstraintLetter != 'n') {
1599 int64_t Offs = GA->getOffset();
1600 if (C) Offs += C->getValue();
1601 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(),
1602 Op.getValueType(), Offs));
1606 if (C) { // just C, no GV.
1607 // Simple constants are not allowed for 's'.
1608 if (ConstraintLetter != 's') {
1609 Ops.push_back(DAG.getTargetConstant(C->getValue(), Op.getValueType()));
1618 std::vector<unsigned> TargetLowering::
1619 getRegClassForInlineAsmConstraint(const std::string &Constraint,
1620 MVT::ValueType VT) const {
1621 return std::vector<unsigned>();
1625 std::pair<unsigned, const TargetRegisterClass*> TargetLowering::
1626 getRegForInlineAsmConstraint(const std::string &Constraint,
1627 MVT::ValueType VT) const {
1628 if (Constraint[0] != '{')
1629 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
1630 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?");
1632 // Remove the braces from around the name.
1633 std::string RegName(Constraint.begin()+1, Constraint.end()-1);
1635 // Figure out which register class contains this reg.
1636 const TargetRegisterInfo *RI = TM.getRegisterInfo();
1637 for (TargetRegisterInfo::regclass_iterator RCI = RI->regclass_begin(),
1638 E = RI->regclass_end(); RCI != E; ++RCI) {
1639 const TargetRegisterClass *RC = *RCI;
1641 // If none of the the value types for this register class are valid, we
1642 // can't use it. For example, 64-bit reg classes on 32-bit targets.
1643 bool isLegal = false;
1644 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
1646 if (isTypeLegal(*I)) {
1652 if (!isLegal) continue;
1654 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end();
1656 if (StringsEqualNoCase(RegName, RI->get(*I).AsmName))
1657 return std::make_pair(*I, RC);
1661 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
1664 //===----------------------------------------------------------------------===//
1665 // Loop Strength Reduction hooks
1666 //===----------------------------------------------------------------------===//
1668 /// isLegalAddressingMode - Return true if the addressing mode represented
1669 /// by AM is legal for this target, for a load/store of the specified type.
1670 bool TargetLowering::isLegalAddressingMode(const AddrMode &AM,
1671 const Type *Ty) const {
1672 // The default implementation of this implements a conservative RISCy, r+r and
1675 // Allows a sign-extended 16-bit immediate field.
1676 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
1679 // No global is ever allowed as a base.
1683 // Only support r+r,
1685 case 0: // "r+i" or just "i", depending on HasBaseReg.
1688 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
1690 // Otherwise we have r+r or r+i.
1693 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
1695 // Allow 2*r as r+r.
1702 // Magic for divide replacement
1705 int64_t m; // magic number
1706 int64_t s; // shift amount
1710 uint64_t m; // magic number
1711 int64_t a; // add indicator
1712 int64_t s; // shift amount
1715 /// magic - calculate the magic numbers required to codegen an integer sdiv as
1716 /// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
1718 static ms magic32(int32_t d) {
1720 uint32_t ad, anc, delta, q1, r1, q2, r2, t;
1721 const uint32_t two31 = 0x80000000U;
1725 t = two31 + ((uint32_t)d >> 31);
1726 anc = t - 1 - t%ad; // absolute value of nc
1727 p = 31; // initialize p
1728 q1 = two31/anc; // initialize q1 = 2p/abs(nc)
1729 r1 = two31 - q1*anc; // initialize r1 = rem(2p,abs(nc))
1730 q2 = two31/ad; // initialize q2 = 2p/abs(d)
1731 r2 = two31 - q2*ad; // initialize r2 = rem(2p,abs(d))
1734 q1 = 2*q1; // update q1 = 2p/abs(nc)
1735 r1 = 2*r1; // update r1 = rem(2p/abs(nc))
1736 if (r1 >= anc) { // must be unsigned comparison
1740 q2 = 2*q2; // update q2 = 2p/abs(d)
1741 r2 = 2*r2; // update r2 = rem(2p/abs(d))
1742 if (r2 >= ad) { // must be unsigned comparison
1747 } while (q1 < delta || (q1 == delta && r1 == 0));
1749 mag.m = (int32_t)(q2 + 1); // make sure to sign extend
1750 if (d < 0) mag.m = -mag.m; // resulting magic number
1751 mag.s = p - 32; // resulting shift
1755 /// magicu - calculate the magic numbers required to codegen an integer udiv as
1756 /// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
1757 static mu magicu32(uint32_t d) {
1759 uint32_t nc, delta, q1, r1, q2, r2;
1761 magu.a = 0; // initialize "add" indicator
1763 p = 31; // initialize p
1764 q1 = 0x80000000/nc; // initialize q1 = 2p/nc
1765 r1 = 0x80000000 - q1*nc; // initialize r1 = rem(2p,nc)
1766 q2 = 0x7FFFFFFF/d; // initialize q2 = (2p-1)/d
1767 r2 = 0x7FFFFFFF - q2*d; // initialize r2 = rem((2p-1),d)
1770 if (r1 >= nc - r1 ) {
1771 q1 = 2*q1 + 1; // update q1
1772 r1 = 2*r1 - nc; // update r1
1775 q1 = 2*q1; // update q1
1776 r1 = 2*r1; // update r1
1778 if (r2 + 1 >= d - r2) {
1779 if (q2 >= 0x7FFFFFFF) magu.a = 1;
1780 q2 = 2*q2 + 1; // update q2
1781 r2 = 2*r2 + 1 - d; // update r2
1784 if (q2 >= 0x80000000) magu.a = 1;
1785 q2 = 2*q2; // update q2
1786 r2 = 2*r2 + 1; // update r2
1789 } while (p < 64 && (q1 < delta || (q1 == delta && r1 == 0)));
1790 magu.m = q2 + 1; // resulting magic number
1791 magu.s = p - 32; // resulting shift
1795 /// magic - calculate the magic numbers required to codegen an integer sdiv as
1796 /// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
1798 static ms magic64(int64_t d) {
1800 uint64_t ad, anc, delta, q1, r1, q2, r2, t;
1801 const uint64_t two63 = 9223372036854775808ULL; // 2^63
1804 ad = d >= 0 ? d : -d;
1805 t = two63 + ((uint64_t)d >> 63);
1806 anc = t - 1 - t%ad; // absolute value of nc
1807 p = 63; // initialize p
1808 q1 = two63/anc; // initialize q1 = 2p/abs(nc)
1809 r1 = two63 - q1*anc; // initialize r1 = rem(2p,abs(nc))
1810 q2 = two63/ad; // initialize q2 = 2p/abs(d)
1811 r2 = two63 - q2*ad; // initialize r2 = rem(2p,abs(d))
1814 q1 = 2*q1; // update q1 = 2p/abs(nc)
1815 r1 = 2*r1; // update r1 = rem(2p/abs(nc))
1816 if (r1 >= anc) { // must be unsigned comparison
1820 q2 = 2*q2; // update q2 = 2p/abs(d)
1821 r2 = 2*r2; // update r2 = rem(2p/abs(d))
1822 if (r2 >= ad) { // must be unsigned comparison
1827 } while (q1 < delta || (q1 == delta && r1 == 0));
1830 if (d < 0) mag.m = -mag.m; // resulting magic number
1831 mag.s = p - 64; // resulting shift
1835 /// magicu - calculate the magic numbers required to codegen an integer udiv as
1836 /// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
1837 static mu magicu64(uint64_t d)
1840 uint64_t nc, delta, q1, r1, q2, r2;
1842 magu.a = 0; // initialize "add" indicator
1844 p = 63; // initialize p
1845 q1 = 0x8000000000000000ull/nc; // initialize q1 = 2p/nc
1846 r1 = 0x8000000000000000ull - q1*nc; // initialize r1 = rem(2p,nc)
1847 q2 = 0x7FFFFFFFFFFFFFFFull/d; // initialize q2 = (2p-1)/d
1848 r2 = 0x7FFFFFFFFFFFFFFFull - q2*d; // initialize r2 = rem((2p-1),d)
1851 if (r1 >= nc - r1 ) {
1852 q1 = 2*q1 + 1; // update q1
1853 r1 = 2*r1 - nc; // update r1
1856 q1 = 2*q1; // update q1
1857 r1 = 2*r1; // update r1
1859 if (r2 + 1 >= d - r2) {
1860 if (q2 >= 0x7FFFFFFFFFFFFFFFull) magu.a = 1;
1861 q2 = 2*q2 + 1; // update q2
1862 r2 = 2*r2 + 1 - d; // update r2
1865 if (q2 >= 0x8000000000000000ull) magu.a = 1;
1866 q2 = 2*q2; // update q2
1867 r2 = 2*r2 + 1; // update r2
1870 } while (p < 128 && (q1 < delta || (q1 == delta && r1 == 0)));
1871 magu.m = q2 + 1; // resulting magic number
1872 magu.s = p - 64; // resulting shift
1876 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
1877 /// return a DAG expression to select that will generate the same value by
1878 /// multiplying by a magic number. See:
1879 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
1880 SDOperand TargetLowering::BuildSDIV(SDNode *N, SelectionDAG &DAG,
1881 std::vector<SDNode*>* Created) const {
1882 MVT::ValueType VT = N->getValueType(0);
1884 // Check to see if we can do this.
1885 if (!isTypeLegal(VT) || (VT != MVT::i32 && VT != MVT::i64))
1886 return SDOperand(); // BuildSDIV only operates on i32 or i64
1888 int64_t d = cast<ConstantSDNode>(N->getOperand(1))->getSignExtended();
1889 ms magics = (VT == MVT::i32) ? magic32(d) : magic64(d);
1891 // Multiply the numerator (operand 0) by the magic value
1893 if (isOperationLegal(ISD::MULHS, VT))
1894 Q = DAG.getNode(ISD::MULHS, VT, N->getOperand(0),
1895 DAG.getConstant(magics.m, VT));
1896 else if (isOperationLegal(ISD::SMUL_LOHI, VT))
1897 Q = SDOperand(DAG.getNode(ISD::SMUL_LOHI, DAG.getVTList(VT, VT),
1899 DAG.getConstant(magics.m, VT)).Val, 1);
1901 return SDOperand(); // No mulhs or equvialent
1902 // If d > 0 and m < 0, add the numerator
1903 if (d > 0 && magics.m < 0) {
1904 Q = DAG.getNode(ISD::ADD, VT, Q, N->getOperand(0));
1906 Created->push_back(Q.Val);
1908 // If d < 0 and m > 0, subtract the numerator.
1909 if (d < 0 && magics.m > 0) {
1910 Q = DAG.getNode(ISD::SUB, VT, Q, N->getOperand(0));
1912 Created->push_back(Q.Val);
1914 // Shift right algebraic if shift value is nonzero
1916 Q = DAG.getNode(ISD::SRA, VT, Q,
1917 DAG.getConstant(magics.s, getShiftAmountTy()));
1919 Created->push_back(Q.Val);
1921 // Extract the sign bit and add it to the quotient
1923 DAG.getNode(ISD::SRL, VT, Q, DAG.getConstant(MVT::getSizeInBits(VT)-1,
1924 getShiftAmountTy()));
1926 Created->push_back(T.Val);
1927 return DAG.getNode(ISD::ADD, VT, Q, T);
1930 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
1931 /// return a DAG expression to select that will generate the same value by
1932 /// multiplying by a magic number. See:
1933 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
1934 SDOperand TargetLowering::BuildUDIV(SDNode *N, SelectionDAG &DAG,
1935 std::vector<SDNode*>* Created) const {
1936 MVT::ValueType VT = N->getValueType(0);
1938 // Check to see if we can do this.
1939 if (!isTypeLegal(VT) || (VT != MVT::i32 && VT != MVT::i64))
1940 return SDOperand(); // BuildUDIV only operates on i32 or i64
1942 uint64_t d = cast<ConstantSDNode>(N->getOperand(1))->getValue();
1943 mu magics = (VT == MVT::i32) ? magicu32(d) : magicu64(d);
1945 // Multiply the numerator (operand 0) by the magic value
1947 if (isOperationLegal(ISD::MULHU, VT))
1948 Q = DAG.getNode(ISD::MULHU, VT, N->getOperand(0),
1949 DAG.getConstant(magics.m, VT));
1950 else if (isOperationLegal(ISD::UMUL_LOHI, VT))
1951 Q = SDOperand(DAG.getNode(ISD::UMUL_LOHI, DAG.getVTList(VT, VT),
1953 DAG.getConstant(magics.m, VT)).Val, 1);
1955 return SDOperand(); // No mulhu or equvialent
1957 Created->push_back(Q.Val);
1959 if (magics.a == 0) {
1960 return DAG.getNode(ISD::SRL, VT, Q,
1961 DAG.getConstant(magics.s, getShiftAmountTy()));
1963 SDOperand NPQ = DAG.getNode(ISD::SUB, VT, N->getOperand(0), Q);
1965 Created->push_back(NPQ.Val);
1966 NPQ = DAG.getNode(ISD::SRL, VT, NPQ,
1967 DAG.getConstant(1, getShiftAmountTy()));
1969 Created->push_back(NPQ.Val);
1970 NPQ = DAG.getNode(ISD::ADD, VT, NPQ, Q);
1972 Created->push_back(NPQ.Val);
1973 return DAG.getNode(ISD::SRL, VT, NPQ,
1974 DAG.getConstant(magics.s-1, getShiftAmountTy()));