1 //===-- TargetLowering.cpp - Implement the TargetLowering class -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This implements the TargetLowering class.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Target/TargetLowering.h"
15 #include "llvm/Target/TargetData.h"
16 #include "llvm/Target/TargetMachine.h"
17 #include "llvm/Target/MRegisterInfo.h"
18 #include "llvm/DerivedTypes.h"
19 #include "llvm/CodeGen/SelectionDAG.h"
20 #include "llvm/ADT/StringExtras.h"
21 #include "llvm/ADT/STLExtras.h"
22 #include "llvm/Support/MathExtras.h"
23 #include "llvm/Target/TargetAsmInfo.h"
26 /// InitLibcallNames - Set default libcall names.
28 static void InitLibcallNames(const char **Names) {
29 Names[RTLIB::SHL_I32] = "__ashlsi3";
30 Names[RTLIB::SHL_I64] = "__ashldi3";
31 Names[RTLIB::SRL_I32] = "__lshrsi3";
32 Names[RTLIB::SRL_I64] = "__lshrdi3";
33 Names[RTLIB::SRA_I32] = "__ashrsi3";
34 Names[RTLIB::SRA_I64] = "__ashrdi3";
35 Names[RTLIB::MUL_I32] = "__mulsi3";
36 Names[RTLIB::MUL_I64] = "__muldi3";
37 Names[RTLIB::SDIV_I32] = "__divsi3";
38 Names[RTLIB::SDIV_I64] = "__divdi3";
39 Names[RTLIB::UDIV_I32] = "__udivsi3";
40 Names[RTLIB::UDIV_I64] = "__udivdi3";
41 Names[RTLIB::SREM_I32] = "__modsi3";
42 Names[RTLIB::SREM_I64] = "__moddi3";
43 Names[RTLIB::UREM_I32] = "__umodsi3";
44 Names[RTLIB::UREM_I64] = "__umoddi3";
45 Names[RTLIB::NEG_I32] = "__negsi2";
46 Names[RTLIB::NEG_I64] = "__negdi2";
47 Names[RTLIB::ADD_F32] = "__addsf3";
48 Names[RTLIB::ADD_F64] = "__adddf3";
49 Names[RTLIB::SUB_F32] = "__subsf3";
50 Names[RTLIB::SUB_F64] = "__subdf3";
51 Names[RTLIB::MUL_F32] = "__mulsf3";
52 Names[RTLIB::MUL_F64] = "__muldf3";
53 Names[RTLIB::DIV_F32] = "__divsf3";
54 Names[RTLIB::DIV_F64] = "__divdf3";
55 Names[RTLIB::REM_F32] = "fmodf";
56 Names[RTLIB::REM_F64] = "fmod";
57 Names[RTLIB::NEG_F32] = "__negsf2";
58 Names[RTLIB::NEG_F64] = "__negdf2";
59 Names[RTLIB::POWI_F32] = "__powisf2";
60 Names[RTLIB::POWI_F64] = "__powidf2";
61 Names[RTLIB::SQRT_F32] = "sqrtf";
62 Names[RTLIB::SQRT_F64] = "sqrt";
63 Names[RTLIB::SIN_F32] = "sinf";
64 Names[RTLIB::SIN_F64] = "sin";
65 Names[RTLIB::COS_F32] = "cosf";
66 Names[RTLIB::COS_F64] = "cos";
67 Names[RTLIB::FPEXT_F32_F64] = "__extendsfdf2";
68 Names[RTLIB::FPROUND_F64_F32] = "__truncdfsf2";
69 Names[RTLIB::FPTOSINT_F32_I32] = "__fixsfsi";
70 Names[RTLIB::FPTOSINT_F32_I64] = "__fixsfdi";
71 Names[RTLIB::FPTOSINT_F64_I32] = "__fixdfsi";
72 Names[RTLIB::FPTOSINT_F64_I64] = "__fixdfdi";
73 Names[RTLIB::FPTOSINT_LD_I64] = "__fixxfdi";
74 Names[RTLIB::FPTOUINT_F32_I32] = "__fixunssfsi";
75 Names[RTLIB::FPTOUINT_F32_I64] = "__fixunssfdi";
76 Names[RTLIB::FPTOUINT_F64_I32] = "__fixunsdfsi";
77 Names[RTLIB::FPTOUINT_F64_I64] = "__fixunsdfdi";
78 Names[RTLIB::FPTOUINT_LD_I32] = "__fixunsxfsi";
79 Names[RTLIB::FPTOUINT_LD_I64] = "__fixunsxfdi";
80 Names[RTLIB::SINTTOFP_I32_F32] = "__floatsisf";
81 Names[RTLIB::SINTTOFP_I32_F64] = "__floatsidf";
82 Names[RTLIB::SINTTOFP_I64_F32] = "__floatdisf";
83 Names[RTLIB::SINTTOFP_I64_F64] = "__floatdidf";
84 Names[RTLIB::SINTTOFP_I64_LD] = "__floatdixf";
85 Names[RTLIB::UINTTOFP_I32_F32] = "__floatunsisf";
86 Names[RTLIB::UINTTOFP_I32_F64] = "__floatunsidf";
87 Names[RTLIB::UINTTOFP_I64_F32] = "__floatundisf";
88 Names[RTLIB::UINTTOFP_I64_F64] = "__floatundidf";
89 Names[RTLIB::OEQ_F32] = "__eqsf2";
90 Names[RTLIB::OEQ_F64] = "__eqdf2";
91 Names[RTLIB::UNE_F32] = "__nesf2";
92 Names[RTLIB::UNE_F64] = "__nedf2";
93 Names[RTLIB::OGE_F32] = "__gesf2";
94 Names[RTLIB::OGE_F64] = "__gedf2";
95 Names[RTLIB::OLT_F32] = "__ltsf2";
96 Names[RTLIB::OLT_F64] = "__ltdf2";
97 Names[RTLIB::OLE_F32] = "__lesf2";
98 Names[RTLIB::OLE_F64] = "__ledf2";
99 Names[RTLIB::OGT_F32] = "__gtsf2";
100 Names[RTLIB::OGT_F64] = "__gtdf2";
101 Names[RTLIB::UO_F32] = "__unordsf2";
102 Names[RTLIB::UO_F64] = "__unorddf2";
103 Names[RTLIB::O_F32] = "__unordsf2";
104 Names[RTLIB::O_F64] = "__unorddf2";
107 /// InitCmpLibcallCCs - Set default comparison libcall CC.
109 static void InitCmpLibcallCCs(ISD::CondCode *CCs) {
110 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL);
111 CCs[RTLIB::OEQ_F32] = ISD::SETEQ;
112 CCs[RTLIB::OEQ_F64] = ISD::SETEQ;
113 CCs[RTLIB::UNE_F32] = ISD::SETNE;
114 CCs[RTLIB::UNE_F64] = ISD::SETNE;
115 CCs[RTLIB::OGE_F32] = ISD::SETGE;
116 CCs[RTLIB::OGE_F64] = ISD::SETGE;
117 CCs[RTLIB::OLT_F32] = ISD::SETLT;
118 CCs[RTLIB::OLT_F64] = ISD::SETLT;
119 CCs[RTLIB::OLE_F32] = ISD::SETLE;
120 CCs[RTLIB::OLE_F64] = ISD::SETLE;
121 CCs[RTLIB::OGT_F32] = ISD::SETGT;
122 CCs[RTLIB::OGT_F64] = ISD::SETGT;
123 CCs[RTLIB::UO_F32] = ISD::SETNE;
124 CCs[RTLIB::UO_F64] = ISD::SETNE;
125 CCs[RTLIB::O_F32] = ISD::SETEQ;
126 CCs[RTLIB::O_F64] = ISD::SETEQ;
129 TargetLowering::TargetLowering(TargetMachine &tm)
130 : TM(tm), TD(TM.getTargetData()) {
131 assert(ISD::BUILTIN_OP_END <= 156 &&
132 "Fixed size array in TargetLowering is not large enough!");
133 // All operations default to being supported.
134 memset(OpActions, 0, sizeof(OpActions));
135 memset(LoadXActions, 0, sizeof(LoadXActions));
136 memset(&StoreXActions, 0, sizeof(StoreXActions));
137 memset(&IndexedModeActions, 0, sizeof(IndexedModeActions));
138 memset(&ConvertActions, 0, sizeof(ConvertActions));
140 // Set all indexed load / store to expand.
141 for (unsigned VT = 0; VT != (unsigned)MVT::LAST_VALUETYPE; ++VT) {
142 for (unsigned IM = (unsigned)ISD::PRE_INC;
143 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
144 setIndexedLoadAction(IM, (MVT::ValueType)VT, Expand);
145 setIndexedStoreAction(IM, (MVT::ValueType)VT, Expand);
149 IsLittleEndian = TD->isLittleEndian();
150 UsesGlobalOffsetTable = false;
151 ShiftAmountTy = SetCCResultTy = PointerTy = getValueType(TD->getIntPtrType());
152 ShiftAmtHandling = Undefined;
153 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE*sizeof(TargetRegisterClass*));
154 memset(TargetDAGCombineArray, 0, array_lengthof(TargetDAGCombineArray));
155 maxStoresPerMemset = maxStoresPerMemcpy = maxStoresPerMemmove = 8;
156 allowUnalignedMemoryAccesses = false;
157 UseUnderscoreSetJmp = false;
158 UseUnderscoreLongJmp = false;
159 SelectIsExpensive = false;
160 IntDivIsCheap = false;
161 Pow2DivIsCheap = false;
162 StackPointerRegisterToSaveRestore = 0;
163 ExceptionPointerRegister = 0;
164 ExceptionSelectorRegister = 0;
165 SetCCResultContents = UndefinedSetCCResult;
166 SchedPreferenceInfo = SchedulingForLatency;
168 JumpBufAlignment = 0;
169 IfCvtBlockSizeLimit = 2;
171 InitLibcallNames(LibcallRoutineNames);
172 InitCmpLibcallCCs(CmpLibcallCCs);
174 // Tell Legalize whether the assembler supports DEBUG_LOC.
175 if (!TM.getTargetAsmInfo()->hasDotLocAndDotFile())
176 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
179 TargetLowering::~TargetLowering() {}
181 /// computeRegisterProperties - Once all of the register classes are added,
182 /// this allows us to compute derived properties we expose.
183 void TargetLowering::computeRegisterProperties() {
184 assert(MVT::LAST_VALUETYPE <= 32 &&
185 "Too many value types for ValueTypeActions to hold!");
187 // Everything defaults to needing one register.
188 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
189 NumRegistersForVT[i] = 1;
190 RegisterTypeForVT[i] = TransformToType[i] = i;
192 // ...except isVoid, which doesn't need any registers.
193 NumRegistersForVT[MVT::isVoid] = 0;
195 // Find the largest integer register class.
196 unsigned LargestIntReg = MVT::i128;
197 for (; RegClassForVT[LargestIntReg] == 0; --LargestIntReg)
198 assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
200 // Every integer value type larger than this largest register takes twice as
201 // many registers to represent as the previous ValueType.
202 for (MVT::ValueType ExpandedReg = LargestIntReg + 1;
203 MVT::isInteger(ExpandedReg); ++ExpandedReg) {
204 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
205 RegisterTypeForVT[ExpandedReg] = LargestIntReg;
206 TransformToType[ExpandedReg] = ExpandedReg - 1;
207 ValueTypeActions.setTypeAction(ExpandedReg, Expand);
210 // Inspect all of the ValueType's smaller than the largest integer
211 // register to see which ones need promotion.
212 MVT::ValueType LegalIntReg = LargestIntReg;
213 for (MVT::ValueType IntReg = LargestIntReg - 1;
214 IntReg >= MVT::i1; --IntReg) {
215 if (isTypeLegal(IntReg)) {
216 LegalIntReg = IntReg;
218 RegisterTypeForVT[IntReg] = TransformToType[IntReg] = LegalIntReg;
219 ValueTypeActions.setTypeAction(IntReg, Promote);
223 // Decide how to handle f64. If the target does not have native f64 support,
224 // expand it to i64 and we will be generating soft float library calls.
225 if (!isTypeLegal(MVT::f64)) {
226 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
227 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
228 TransformToType[MVT::f64] = MVT::i64;
229 ValueTypeActions.setTypeAction(MVT::f64, Expand);
232 // Decide how to handle f32. If the target does not have native support for
233 // f32, promote it to f64 if it is legal. Otherwise, expand it to i32.
234 if (!isTypeLegal(MVT::f32)) {
235 if (isTypeLegal(MVT::f64)) {
236 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::f64];
237 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::f64];
238 TransformToType[MVT::f32] = MVT::f64;
239 ValueTypeActions.setTypeAction(MVT::f32, Promote);
241 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
242 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
243 TransformToType[MVT::f32] = MVT::i32;
244 ValueTypeActions.setTypeAction(MVT::f32, Expand);
248 // Loop over all of the vector value types to see which need transformations.
249 for (MVT::ValueType i = MVT::FIRST_VECTOR_VALUETYPE;
250 i <= MVT::LAST_VECTOR_VALUETYPE; ++i) {
251 if (!isTypeLegal(i)) {
252 MVT::ValueType IntermediateVT, RegisterVT;
253 unsigned NumIntermediates;
254 NumRegistersForVT[i] =
255 getVectorTypeBreakdown(i,
256 IntermediateVT, NumIntermediates,
258 RegisterTypeForVT[i] = RegisterVT;
259 TransformToType[i] = MVT::Other; // this isn't actually used
260 ValueTypeActions.setTypeAction(i, Expand);
265 const char *TargetLowering::getTargetNodeName(unsigned Opcode) const {
269 /// getVectorTypeBreakdown - Vector types are broken down into some number of
270 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
271 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
272 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
274 /// This method returns the number of registers needed, and the VT for each
275 /// register. It also returns the VT and quantity of the intermediate values
276 /// before they are promoted/expanded.
278 unsigned TargetLowering::getVectorTypeBreakdown(MVT::ValueType VT,
279 MVT::ValueType &IntermediateVT,
280 unsigned &NumIntermediates,
281 MVT::ValueType &RegisterVT) const {
282 // Figure out the right, legal destination reg to copy into.
283 unsigned NumElts = MVT::getVectorNumElements(VT);
284 MVT::ValueType EltTy = MVT::getVectorElementType(VT);
286 unsigned NumVectorRegs = 1;
288 // Divide the input until we get to a supported size. This will always
289 // end with a scalar if the target doesn't support vectors.
290 while (NumElts > 1 &&
291 !isTypeLegal(MVT::getVectorType(EltTy, NumElts))) {
296 NumIntermediates = NumVectorRegs;
298 MVT::ValueType NewVT = MVT::getVectorType(EltTy, NumElts);
299 if (!isTypeLegal(NewVT))
301 IntermediateVT = NewVT;
303 MVT::ValueType DestVT = getTypeToTransformTo(NewVT);
305 if (DestVT < NewVT) {
306 // Value is expanded, e.g. i64 -> i16.
307 return NumVectorRegs*(MVT::getSizeInBits(NewVT)/MVT::getSizeInBits(DestVT));
309 // Otherwise, promotion or legal types use the same number of registers as
310 // the vector decimated to the appropriate level.
311 return NumVectorRegs;
317 //===----------------------------------------------------------------------===//
318 // Optimization Methods
319 //===----------------------------------------------------------------------===//
321 /// ShrinkDemandedConstant - Check to see if the specified operand of the
322 /// specified instruction is a constant integer. If so, check to see if there
323 /// are any bits set in the constant that are not demanded. If so, shrink the
324 /// constant and return true.
325 bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDOperand Op,
327 // FIXME: ISD::SELECT, ISD::SELECT_CC
328 switch(Op.getOpcode()) {
333 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1)))
334 if ((~Demanded & C->getValue()) != 0) {
335 MVT::ValueType VT = Op.getValueType();
336 SDOperand New = DAG.getNode(Op.getOpcode(), VT, Op.getOperand(0),
337 DAG.getConstant(Demanded & C->getValue(),
339 return CombineTo(Op, New);
346 /// SimplifyDemandedBits - Look at Op. At this point, we know that only the
347 /// DemandedMask bits of the result of Op are ever used downstream. If we can
348 /// use this information to simplify Op, create a new simplified DAG node and
349 /// return true, returning the original and new nodes in Old and New. Otherwise,
350 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
351 /// the expression (used to simplify the caller). The KnownZero/One bits may
352 /// only be accurate for those bits in the DemandedMask.
353 bool TargetLowering::SimplifyDemandedBits(SDOperand Op, uint64_t DemandedMask,
356 TargetLoweringOpt &TLO,
357 unsigned Depth) const {
358 KnownZero = KnownOne = 0; // Don't know anything.
360 // The masks are not wide enough to represent this type! Should use APInt.
361 if (Op.getValueType() == MVT::i128)
364 // Other users may use these bits.
365 if (!Op.Val->hasOneUse()) {
367 // If not at the root, Just compute the KnownZero/KnownOne bits to
368 // simplify things downstream.
369 TLO.DAG.ComputeMaskedBits(Op, DemandedMask, KnownZero, KnownOne, Depth);
372 // If this is the root being simplified, allow it to have multiple uses,
373 // just set the DemandedMask to all bits.
374 DemandedMask = MVT::getIntVTBitMask(Op.getValueType());
375 } else if (DemandedMask == 0) {
376 // Not demanding any bits from Op.
377 if (Op.getOpcode() != ISD::UNDEF)
378 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::UNDEF, Op.getValueType()));
380 } else if (Depth == 6) { // Limit search depth.
384 uint64_t KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut;
385 switch (Op.getOpcode()) {
387 // We know all of the bits for a constant!
388 KnownOne = cast<ConstantSDNode>(Op)->getValue() & DemandedMask;
389 KnownZero = ~KnownOne & DemandedMask;
390 return false; // Don't fall through, will infinitely loop.
392 // If the RHS is a constant, check to see if the LHS would be zero without
393 // using the bits from the RHS. Below, we use knowledge about the RHS to
394 // simplify the LHS, here we're using information from the LHS to simplify
396 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
397 uint64_t LHSZero, LHSOne;
398 TLO.DAG.ComputeMaskedBits(Op.getOperand(0), DemandedMask,
399 LHSZero, LHSOne, Depth+1);
400 // If the LHS already has zeros where RHSC does, this and is dead.
401 if ((LHSZero & DemandedMask) == (~RHSC->getValue() & DemandedMask))
402 return TLO.CombineTo(Op, Op.getOperand(0));
403 // If any of the set bits in the RHS are known zero on the LHS, shrink
405 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & DemandedMask))
409 if (SimplifyDemandedBits(Op.getOperand(1), DemandedMask, KnownZero,
410 KnownOne, TLO, Depth+1))
412 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
413 if (SimplifyDemandedBits(Op.getOperand(0), DemandedMask & ~KnownZero,
414 KnownZero2, KnownOne2, TLO, Depth+1))
416 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
418 // If all of the demanded bits are known one on one side, return the other.
419 // These bits cannot contribute to the result of the 'and'.
420 if ((DemandedMask & ~KnownZero2 & KnownOne)==(DemandedMask & ~KnownZero2))
421 return TLO.CombineTo(Op, Op.getOperand(0));
422 if ((DemandedMask & ~KnownZero & KnownOne2)==(DemandedMask & ~KnownZero))
423 return TLO.CombineTo(Op, Op.getOperand(1));
424 // If all of the demanded bits in the inputs are known zeros, return zero.
425 if ((DemandedMask & (KnownZero|KnownZero2)) == DemandedMask)
426 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, Op.getValueType()));
427 // If the RHS is a constant, see if we can simplify it.
428 if (TLO.ShrinkDemandedConstant(Op, DemandedMask & ~KnownZero2))
431 // Output known-1 bits are only known if set in both the LHS & RHS.
432 KnownOne &= KnownOne2;
433 // Output known-0 are known to be clear if zero in either the LHS | RHS.
434 KnownZero |= KnownZero2;
437 if (SimplifyDemandedBits(Op.getOperand(1), DemandedMask, KnownZero,
438 KnownOne, TLO, Depth+1))
440 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
441 if (SimplifyDemandedBits(Op.getOperand(0), DemandedMask & ~KnownOne,
442 KnownZero2, KnownOne2, TLO, Depth+1))
444 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
446 // If all of the demanded bits are known zero on one side, return the other.
447 // These bits cannot contribute to the result of the 'or'.
448 if ((DemandedMask & ~KnownOne2 & KnownZero) == (DemandedMask & ~KnownOne2))
449 return TLO.CombineTo(Op, Op.getOperand(0));
450 if ((DemandedMask & ~KnownOne & KnownZero2) == (DemandedMask & ~KnownOne))
451 return TLO.CombineTo(Op, Op.getOperand(1));
452 // If all of the potentially set bits on one side are known to be set on
453 // the other side, just use the 'other' side.
454 if ((DemandedMask & (~KnownZero) & KnownOne2) ==
455 (DemandedMask & (~KnownZero)))
456 return TLO.CombineTo(Op, Op.getOperand(0));
457 if ((DemandedMask & (~KnownZero2) & KnownOne) ==
458 (DemandedMask & (~KnownZero2)))
459 return TLO.CombineTo(Op, Op.getOperand(1));
460 // If the RHS is a constant, see if we can simplify it.
461 if (TLO.ShrinkDemandedConstant(Op, DemandedMask))
464 // Output known-0 bits are only known if clear in both the LHS & RHS.
465 KnownZero &= KnownZero2;
466 // Output known-1 are known to be set if set in either the LHS | RHS.
467 KnownOne |= KnownOne2;
470 if (SimplifyDemandedBits(Op.getOperand(1), DemandedMask, KnownZero,
471 KnownOne, TLO, Depth+1))
473 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
474 if (SimplifyDemandedBits(Op.getOperand(0), DemandedMask, KnownZero2,
475 KnownOne2, TLO, Depth+1))
477 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
479 // If all of the demanded bits are known zero on one side, return the other.
480 // These bits cannot contribute to the result of the 'xor'.
481 if ((DemandedMask & KnownZero) == DemandedMask)
482 return TLO.CombineTo(Op, Op.getOperand(0));
483 if ((DemandedMask & KnownZero2) == DemandedMask)
484 return TLO.CombineTo(Op, Op.getOperand(1));
486 // If all of the unknown bits are known to be zero on one side or the other
487 // (but not both) turn this into an *inclusive* or.
488 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
489 if ((DemandedMask & ~KnownZero & ~KnownZero2) == 0)
490 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, Op.getValueType(),
494 // Output known-0 bits are known if clear or set in both the LHS & RHS.
495 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
496 // Output known-1 are known to be set if set in only one of the LHS, RHS.
497 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
499 // If all of the demanded bits on one side are known, and all of the set
500 // bits on that side are also known to be set on the other side, turn this
501 // into an AND, as we know the bits will be cleared.
502 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
503 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask) { // all known
504 if ((KnownOne & KnownOne2) == KnownOne) {
505 MVT::ValueType VT = Op.getValueType();
506 SDOperand ANDC = TLO.DAG.getConstant(~KnownOne & DemandedMask, VT);
507 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, VT, Op.getOperand(0),
512 // If the RHS is a constant, see if we can simplify it.
513 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
514 if (TLO.ShrinkDemandedConstant(Op, DemandedMask))
517 KnownZero = KnownZeroOut;
518 KnownOne = KnownOneOut;
521 // If we know the result of a setcc has the top bits zero, use this info.
522 if (getSetCCResultContents() == TargetLowering::ZeroOrOneSetCCResult)
523 KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL);
526 if (SimplifyDemandedBits(Op.getOperand(2), DemandedMask, KnownZero,
527 KnownOne, TLO, Depth+1))
529 if (SimplifyDemandedBits(Op.getOperand(1), DemandedMask, KnownZero2,
530 KnownOne2, TLO, Depth+1))
532 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
533 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
535 // If the operands are constants, see if we can simplify them.
536 if (TLO.ShrinkDemandedConstant(Op, DemandedMask))
539 // Only known if known in both the LHS and RHS.
540 KnownOne &= KnownOne2;
541 KnownZero &= KnownZero2;
544 if (SimplifyDemandedBits(Op.getOperand(3), DemandedMask, KnownZero,
545 KnownOne, TLO, Depth+1))
547 if (SimplifyDemandedBits(Op.getOperand(2), DemandedMask, KnownZero2,
548 KnownOne2, TLO, Depth+1))
550 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
551 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
553 // If the operands are constants, see if we can simplify them.
554 if (TLO.ShrinkDemandedConstant(Op, DemandedMask))
557 // Only known if known in both the LHS and RHS.
558 KnownOne &= KnownOne2;
559 KnownZero &= KnownZero2;
562 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
563 unsigned ShAmt = SA->getValue();
564 SDOperand InOp = Op.getOperand(0);
566 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a
567 // single shift. We can do this if the bottom bits (which are shifted
568 // out) are never demanded.
569 if (InOp.getOpcode() == ISD::SRL &&
570 isa<ConstantSDNode>(InOp.getOperand(1))) {
571 if (ShAmt && (DemandedMask & ((1ULL << ShAmt)-1)) == 0) {
572 unsigned C1 = cast<ConstantSDNode>(InOp.getOperand(1))->getValue();
573 unsigned Opc = ISD::SHL;
581 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
582 MVT::ValueType VT = Op.getValueType();
583 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, VT,
584 InOp.getOperand(0), NewSA));
588 if (SimplifyDemandedBits(Op.getOperand(0), DemandedMask >> ShAmt,
589 KnownZero, KnownOne, TLO, Depth+1))
591 KnownZero <<= SA->getValue();
592 KnownOne <<= SA->getValue();
593 KnownZero |= (1ULL << SA->getValue())-1; // low bits known zero.
597 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
598 MVT::ValueType VT = Op.getValueType();
599 unsigned ShAmt = SA->getValue();
600 uint64_t TypeMask = MVT::getIntVTBitMask(VT);
601 unsigned VTSize = MVT::getSizeInBits(VT);
602 SDOperand InOp = Op.getOperand(0);
604 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a
605 // single shift. We can do this if the top bits (which are shifted out)
606 // are never demanded.
607 if (InOp.getOpcode() == ISD::SHL &&
608 isa<ConstantSDNode>(InOp.getOperand(1))) {
609 if (ShAmt && (DemandedMask & (~0ULL << (VTSize-ShAmt))) == 0) {
610 unsigned C1 = cast<ConstantSDNode>(InOp.getOperand(1))->getValue();
611 unsigned Opc = ISD::SRL;
619 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
620 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, VT,
621 InOp.getOperand(0), NewSA));
625 // Compute the new bits that are at the top now.
626 if (SimplifyDemandedBits(InOp, (DemandedMask << ShAmt) & TypeMask,
627 KnownZero, KnownOne, TLO, Depth+1))
629 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
630 KnownZero &= TypeMask;
631 KnownOne &= TypeMask;
635 uint64_t HighBits = (1ULL << ShAmt)-1;
636 HighBits <<= VTSize - ShAmt;
637 KnownZero |= HighBits; // High bits known zero.
641 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
642 MVT::ValueType VT = Op.getValueType();
643 unsigned ShAmt = SA->getValue();
645 // Compute the new bits that are at the top now.
646 uint64_t TypeMask = MVT::getIntVTBitMask(VT);
648 uint64_t InDemandedMask = (DemandedMask << ShAmt) & TypeMask;
650 // If any of the demanded bits are produced by the sign extension, we also
651 // demand the input sign bit.
652 uint64_t HighBits = (1ULL << ShAmt)-1;
653 HighBits <<= MVT::getSizeInBits(VT) - ShAmt;
654 if (HighBits & DemandedMask)
655 InDemandedMask |= MVT::getIntVTSignBit(VT);
657 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask,
658 KnownZero, KnownOne, TLO, Depth+1))
660 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
661 KnownZero &= TypeMask;
662 KnownOne &= TypeMask;
666 // Handle the sign bits.
667 uint64_t SignBit = MVT::getIntVTSignBit(VT);
668 SignBit >>= ShAmt; // Adjust to where it is now in the mask.
670 // If the input sign bit is known to be zero, or if none of the top bits
671 // are demanded, turn this into an unsigned shift right.
672 if ((KnownZero & SignBit) || (HighBits & ~DemandedMask) == HighBits) {
673 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, VT, Op.getOperand(0),
675 } else if (KnownOne & SignBit) { // New bits are known one.
676 KnownOne |= HighBits;
680 case ISD::SIGN_EXTEND_INREG: {
681 MVT::ValueType EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
683 // Sign extension. Compute the demanded bits in the result that are not
684 // present in the input.
685 uint64_t NewBits = ~MVT::getIntVTBitMask(EVT) & DemandedMask;
687 // If none of the extended bits are demanded, eliminate the sextinreg.
689 return TLO.CombineTo(Op, Op.getOperand(0));
691 uint64_t InSignBit = MVT::getIntVTSignBit(EVT);
692 int64_t InputDemandedBits = DemandedMask & MVT::getIntVTBitMask(EVT);
694 // Since the sign extended bits are demanded, we know that the sign
696 InputDemandedBits |= InSignBit;
698 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits,
699 KnownZero, KnownOne, TLO, Depth+1))
701 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
703 // If the sign bit of the input is known set or clear, then we know the
704 // top bits of the result.
706 // If the input sign bit is known zero, convert this into a zero extension.
707 if (KnownZero & InSignBit)
708 return TLO.CombineTo(Op,
709 TLO.DAG.getZeroExtendInReg(Op.getOperand(0), EVT));
711 if (KnownOne & InSignBit) { // Input sign bit known set
713 KnownZero &= ~NewBits;
714 } else { // Input sign bit unknown
715 KnownZero &= ~NewBits;
716 KnownOne &= ~NewBits;
723 MVT::ValueType VT = Op.getValueType();
724 unsigned LowBits = Log2_32(MVT::getSizeInBits(VT))+1;
725 KnownZero = ~((1ULL << LowBits)-1) & MVT::getIntVTBitMask(VT);
730 if (ISD::isZEXTLoad(Op.Val)) {
731 LoadSDNode *LD = cast<LoadSDNode>(Op);
732 MVT::ValueType VT = LD->getLoadedVT();
733 KnownZero |= ~MVT::getIntVTBitMask(VT) & DemandedMask;
737 case ISD::ZERO_EXTEND: {
738 uint64_t InMask = MVT::getIntVTBitMask(Op.getOperand(0).getValueType());
740 // If none of the top bits are demanded, convert this into an any_extend.
741 uint64_t NewBits = (~InMask) & DemandedMask;
743 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND,
747 if (SimplifyDemandedBits(Op.getOperand(0), DemandedMask & InMask,
748 KnownZero, KnownOne, TLO, Depth+1))
750 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
751 KnownZero |= NewBits;
754 case ISD::SIGN_EXTEND: {
755 MVT::ValueType InVT = Op.getOperand(0).getValueType();
756 uint64_t InMask = MVT::getIntVTBitMask(InVT);
757 uint64_t InSignBit = MVT::getIntVTSignBit(InVT);
758 uint64_t NewBits = (~InMask) & DemandedMask;
760 // If none of the top bits are demanded, convert this into an any_extend.
762 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND,Op.getValueType(),
765 // Since some of the sign extended bits are demanded, we know that the sign
767 uint64_t InDemandedBits = DemandedMask & InMask;
768 InDemandedBits |= InSignBit;
770 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero,
771 KnownOne, TLO, Depth+1))
774 // If the sign bit is known zero, convert this to a zero extend.
775 if (KnownZero & InSignBit)
776 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND,
780 // If the sign bit is known one, the top bits match.
781 if (KnownOne & InSignBit) {
783 KnownZero &= ~NewBits;
784 } else { // Otherwise, top bits aren't known.
785 KnownOne &= ~NewBits;
786 KnownZero &= ~NewBits;
790 case ISD::ANY_EXTEND: {
791 uint64_t InMask = MVT::getIntVTBitMask(Op.getOperand(0).getValueType());
792 if (SimplifyDemandedBits(Op.getOperand(0), DemandedMask & InMask,
793 KnownZero, KnownOne, TLO, Depth+1))
795 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
798 case ISD::TRUNCATE: {
799 // Simplify the input, using demanded bit information, and compute the known
800 // zero/one bits live out.
801 if (SimplifyDemandedBits(Op.getOperand(0), DemandedMask,
802 KnownZero, KnownOne, TLO, Depth+1))
805 // If the input is only used by this truncate, see if we can shrink it based
806 // on the known demanded bits.
807 if (Op.getOperand(0).Val->hasOneUse()) {
808 SDOperand In = Op.getOperand(0);
809 switch (In.getOpcode()) {
812 // Shrink SRL by a constant if none of the high bits shifted in are
814 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1))){
815 uint64_t HighBits = MVT::getIntVTBitMask(In.getValueType());
816 HighBits &= ~MVT::getIntVTBitMask(Op.getValueType());
817 HighBits >>= ShAmt->getValue();
819 if (ShAmt->getValue() < MVT::getSizeInBits(Op.getValueType()) &&
820 (DemandedMask & HighBits) == 0) {
821 // None of the shifted in bits are needed. Add a truncate of the
822 // shift input, then shift it.
823 SDOperand NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE,
826 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL,Op.getValueType(),
827 NewTrunc, In.getOperand(1)));
834 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
835 uint64_t OutMask = MVT::getIntVTBitMask(Op.getValueType());
836 KnownZero &= OutMask;
840 case ISD::AssertZext: {
841 MVT::ValueType VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
842 uint64_t InMask = MVT::getIntVTBitMask(VT);
843 if (SimplifyDemandedBits(Op.getOperand(0), DemandedMask & InMask,
844 KnownZero, KnownOne, TLO, Depth+1))
846 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
847 KnownZero |= ~InMask & DemandedMask;
852 case ISD::INTRINSIC_WO_CHAIN:
853 case ISD::INTRINSIC_W_CHAIN:
854 case ISD::INTRINSIC_VOID:
855 // Just use ComputeMaskedBits to compute output bits.
856 TLO.DAG.ComputeMaskedBits(Op, DemandedMask, KnownZero, KnownOne, Depth);
860 // If we know the value of all of the demanded bits, return this as a
862 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask)
863 return TLO.CombineTo(Op, TLO.DAG.getConstant(KnownOne, Op.getValueType()));
868 /// computeMaskedBitsForTargetNode - Determine which of the bits specified
869 /// in Mask are known to be either zero or one and return them in the
870 /// KnownZero/KnownOne bitsets.
871 void TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
875 const SelectionDAG &DAG,
876 unsigned Depth) const {
877 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
878 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
879 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
880 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
881 "Should use MaskedValueIsZero if you don't know whether Op"
882 " is a target node!");
887 /// ComputeNumSignBitsForTargetNode - This method can be implemented by
888 /// targets that want to expose additional information about sign bits to the
890 unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDOperand Op,
891 unsigned Depth) const {
892 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
893 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
894 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
895 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
896 "Should use ComputeNumSignBits if you don't know whether Op"
897 " is a target node!");
902 /// SimplifySetCC - Try to simplify a setcc built with the specified operands
903 /// and cc. If it is unable to simplify it, return a null SDOperand.
905 TargetLowering::SimplifySetCC(MVT::ValueType VT, SDOperand N0, SDOperand N1,
906 ISD::CondCode Cond, bool foldBooleans,
907 DAGCombinerInfo &DCI) const {
908 SelectionDAG &DAG = DCI.DAG;
910 // These setcc operations always fold.
914 case ISD::SETFALSE2: return DAG.getConstant(0, VT);
916 case ISD::SETTRUE2: return DAG.getConstant(1, VT);
919 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val)) {
920 uint64_t C1 = N1C->getValue();
921 if (isa<ConstantSDNode>(N0.Val)) {
922 return DAG.FoldSetCC(VT, N0, N1, Cond);
924 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an
925 // equality comparison, then we're just comparing whether X itself is
927 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) &&
928 N0.getOperand(0).getOpcode() == ISD::CTLZ &&
929 N0.getOperand(1).getOpcode() == ISD::Constant) {
930 unsigned ShAmt = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
931 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
932 ShAmt == Log2_32(MVT::getSizeInBits(N0.getValueType()))) {
933 if ((C1 == 0) == (Cond == ISD::SETEQ)) {
934 // (srl (ctlz x), 5) == 0 -> X != 0
935 // (srl (ctlz x), 5) != 1 -> X != 0
938 // (srl (ctlz x), 5) != 0 -> X == 0
939 // (srl (ctlz x), 5) == 1 -> X == 0
942 SDOperand Zero = DAG.getConstant(0, N0.getValueType());
943 return DAG.getSetCC(VT, N0.getOperand(0).getOperand(0),
948 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
949 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
950 unsigned InSize = MVT::getSizeInBits(N0.getOperand(0).getValueType());
952 // If the comparison constant has bits in the upper part, the
953 // zero-extended value could never match.
954 if (C1 & (~0ULL << InSize)) {
955 unsigned VSize = MVT::getSizeInBits(N0.getValueType());
959 case ISD::SETEQ: return DAG.getConstant(0, VT);
962 case ISD::SETNE: return DAG.getConstant(1, VT);
965 // True if the sign bit of C1 is set.
966 return DAG.getConstant((C1 & (1ULL << (VSize-1))) != 0, VT);
969 // True if the sign bit of C1 isn't set.
970 return DAG.getConstant((C1 & (1ULL << (VSize-1))) == 0, VT);
976 // Otherwise, we can perform the comparison with the low bits.
984 return DAG.getSetCC(VT, N0.getOperand(0),
985 DAG.getConstant(C1, N0.getOperand(0).getValueType()),
988 break; // todo, be more careful with signed comparisons
990 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
991 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
992 MVT::ValueType ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
993 unsigned ExtSrcTyBits = MVT::getSizeInBits(ExtSrcTy);
994 MVT::ValueType ExtDstTy = N0.getValueType();
995 unsigned ExtDstTyBits = MVT::getSizeInBits(ExtDstTy);
997 // If the extended part has any inconsistent bits, it cannot ever
998 // compare equal. In other words, they have to be all ones or all
1001 (~0ULL >> (64-ExtSrcTyBits)) & (~0ULL << (ExtDstTyBits-1));
1002 if ((C1 & ExtBits) != 0 && (C1 & ExtBits) != ExtBits)
1003 return DAG.getConstant(Cond == ISD::SETNE, VT);
1006 MVT::ValueType Op0Ty = N0.getOperand(0).getValueType();
1007 if (Op0Ty == ExtSrcTy) {
1008 ZextOp = N0.getOperand(0);
1010 int64_t Imm = ~0ULL >> (64-ExtSrcTyBits);
1011 ZextOp = DAG.getNode(ISD::AND, Op0Ty, N0.getOperand(0),
1012 DAG.getConstant(Imm, Op0Ty));
1014 if (!DCI.isCalledByLegalizer())
1015 DCI.AddToWorklist(ZextOp.Val);
1016 // Otherwise, make this a use of a zext.
1017 return DAG.getSetCC(VT, ZextOp,
1018 DAG.getConstant(C1 & (~0ULL>>(64-ExtSrcTyBits)),
1021 } else if ((N1C->getValue() == 0 || N1C->getValue() == 1) &&
1022 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1024 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC
1025 if (N0.getOpcode() == ISD::SETCC) {
1026 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getValue() != 1);
1030 // Invert the condition.
1031 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
1032 CC = ISD::getSetCCInverse(CC,
1033 MVT::isInteger(N0.getOperand(0).getValueType()));
1034 return DAG.getSetCC(VT, N0.getOperand(0), N0.getOperand(1), CC);
1037 if ((N0.getOpcode() == ISD::XOR ||
1038 (N0.getOpcode() == ISD::AND &&
1039 N0.getOperand(0).getOpcode() == ISD::XOR &&
1040 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
1041 isa<ConstantSDNode>(N0.getOperand(1)) &&
1042 cast<ConstantSDNode>(N0.getOperand(1))->getValue() == 1) {
1043 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We
1044 // can only do this if the top bits are known zero.
1045 if (DAG.MaskedValueIsZero(N0,
1046 MVT::getIntVTBitMask(N0.getValueType())-1)){
1047 // Okay, get the un-inverted input value.
1049 if (N0.getOpcode() == ISD::XOR)
1050 Val = N0.getOperand(0);
1052 assert(N0.getOpcode() == ISD::AND &&
1053 N0.getOperand(0).getOpcode() == ISD::XOR);
1054 // ((X^1)&1)^1 -> X & 1
1055 Val = DAG.getNode(ISD::AND, N0.getValueType(),
1056 N0.getOperand(0).getOperand(0),
1059 return DAG.getSetCC(VT, Val, N1,
1060 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
1065 uint64_t MinVal, MaxVal;
1066 unsigned OperandBitSize = MVT::getSizeInBits(N1C->getValueType(0));
1067 if (ISD::isSignedIntSetCC(Cond)) {
1068 MinVal = 1ULL << (OperandBitSize-1);
1069 if (OperandBitSize != 1) // Avoid X >> 64, which is undefined.
1070 MaxVal = ~0ULL >> (65-OperandBitSize);
1075 MaxVal = ~0ULL >> (64-OperandBitSize);
1078 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
1079 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
1080 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true
1081 --C1; // X >= C0 --> X > (C0-1)
1082 return DAG.getSetCC(VT, N0, DAG.getConstant(C1, N1.getValueType()),
1083 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
1086 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
1087 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true
1088 ++C1; // X <= C0 --> X < (C0+1)
1089 return DAG.getSetCC(VT, N0, DAG.getConstant(C1, N1.getValueType()),
1090 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
1093 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
1094 return DAG.getConstant(0, VT); // X < MIN --> false
1095 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal)
1096 return DAG.getConstant(1, VT); // X >= MIN --> true
1097 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal)
1098 return DAG.getConstant(0, VT); // X > MAX --> false
1099 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal)
1100 return DAG.getConstant(1, VT); // X <= MAX --> true
1102 // Canonicalize setgt X, Min --> setne X, Min
1103 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
1104 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
1105 // Canonicalize setlt X, Max --> setne X, Max
1106 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
1107 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
1109 // If we have setult X, 1, turn it into seteq X, 0
1110 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
1111 return DAG.getSetCC(VT, N0, DAG.getConstant(MinVal, N0.getValueType()),
1113 // If we have setugt X, Max-1, turn it into seteq X, Max
1114 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
1115 return DAG.getSetCC(VT, N0, DAG.getConstant(MaxVal, N0.getValueType()),
1118 // If we have "setcc X, C0", check to see if we can shrink the immediate
1121 // SETUGT X, SINTMAX -> SETLT X, 0
1122 if (Cond == ISD::SETUGT && OperandBitSize != 1 &&
1123 C1 == (~0ULL >> (65-OperandBitSize)))
1124 return DAG.getSetCC(VT, N0, DAG.getConstant(0, N1.getValueType()),
1127 // FIXME: Implement the rest of these.
1129 // Fold bit comparisons when we can.
1130 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1131 VT == N0.getValueType() && N0.getOpcode() == ISD::AND)
1132 if (ConstantSDNode *AndRHS =
1133 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1134 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
1135 // Perform the xform if the AND RHS is a single bit.
1136 if (isPowerOf2_64(AndRHS->getValue())) {
1137 return DAG.getNode(ISD::SRL, VT, N0,
1138 DAG.getConstant(Log2_64(AndRHS->getValue()),
1139 getShiftAmountTy()));
1141 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getValue()) {
1142 // (X & 8) == 8 --> (X & 8) >> 3
1143 // Perform the xform if C1 is a single bit.
1144 if (isPowerOf2_64(C1)) {
1145 return DAG.getNode(ISD::SRL, VT, N0,
1146 DAG.getConstant(Log2_64(C1), getShiftAmountTy()));
1151 } else if (isa<ConstantSDNode>(N0.Val)) {
1152 // Ensure that the constant occurs on the RHS.
1153 return DAG.getSetCC(VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
1156 if (isa<ConstantFPSDNode>(N0.Val)) {
1157 // Constant fold or commute setcc.
1158 SDOperand O = DAG.FoldSetCC(VT, N0, N1, Cond);
1159 if (O.Val) return O;
1163 // We can always fold X == X for integer setcc's.
1164 if (MVT::isInteger(N0.getValueType()))
1165 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
1166 unsigned UOF = ISD::getUnorderedFlavor(Cond);
1167 if (UOF == 2) // FP operators that are undefined on NaNs.
1168 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
1169 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
1170 return DAG.getConstant(UOF, VT);
1171 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
1172 // if it is not already.
1173 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
1174 if (NewCond != Cond)
1175 return DAG.getSetCC(VT, N0, N1, NewCond);
1178 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1179 MVT::isInteger(N0.getValueType())) {
1180 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
1181 N0.getOpcode() == ISD::XOR) {
1182 // Simplify (X+Y) == (X+Z) --> Y == Z
1183 if (N0.getOpcode() == N1.getOpcode()) {
1184 if (N0.getOperand(0) == N1.getOperand(0))
1185 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(1), Cond);
1186 if (N0.getOperand(1) == N1.getOperand(1))
1187 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(0), Cond);
1188 if (DAG.isCommutativeBinOp(N0.getOpcode())) {
1189 // If X op Y == Y op X, try other combinations.
1190 if (N0.getOperand(0) == N1.getOperand(1))
1191 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(0), Cond);
1192 if (N0.getOperand(1) == N1.getOperand(0))
1193 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(1), Cond);
1197 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) {
1198 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1199 // Turn (X+C1) == C2 --> X == C2-C1
1200 if (N0.getOpcode() == ISD::ADD && N0.Val->hasOneUse()) {
1201 return DAG.getSetCC(VT, N0.getOperand(0),
1202 DAG.getConstant(RHSC->getValue()-LHSR->getValue(),
1203 N0.getValueType()), Cond);
1206 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
1207 if (N0.getOpcode() == ISD::XOR)
1208 // If we know that all of the inverted bits are zero, don't bother
1209 // performing the inversion.
1210 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getValue()))
1211 return DAG.getSetCC(VT, N0.getOperand(0),
1212 DAG.getConstant(LHSR->getValue()^RHSC->getValue(),
1213 N0.getValueType()), Cond);
1216 // Turn (C1-X) == C2 --> X == C1-C2
1217 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
1218 if (N0.getOpcode() == ISD::SUB && N0.Val->hasOneUse()) {
1219 return DAG.getSetCC(VT, N0.getOperand(1),
1220 DAG.getConstant(SUBC->getValue()-RHSC->getValue(),
1221 N0.getValueType()), Cond);
1226 // Simplify (X+Z) == X --> Z == 0
1227 if (N0.getOperand(0) == N1)
1228 return DAG.getSetCC(VT, N0.getOperand(1),
1229 DAG.getConstant(0, N0.getValueType()), Cond);
1230 if (N0.getOperand(1) == N1) {
1231 if (DAG.isCommutativeBinOp(N0.getOpcode()))
1232 return DAG.getSetCC(VT, N0.getOperand(0),
1233 DAG.getConstant(0, N0.getValueType()), Cond);
1234 else if (N0.Val->hasOneUse()) {
1235 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
1236 // (Z-X) == X --> Z == X<<1
1237 SDOperand SH = DAG.getNode(ISD::SHL, N1.getValueType(),
1239 DAG.getConstant(1, getShiftAmountTy()));
1240 if (!DCI.isCalledByLegalizer())
1241 DCI.AddToWorklist(SH.Val);
1242 return DAG.getSetCC(VT, N0.getOperand(0), SH, Cond);
1247 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
1248 N1.getOpcode() == ISD::XOR) {
1249 // Simplify X == (X+Z) --> Z == 0
1250 if (N1.getOperand(0) == N0) {
1251 return DAG.getSetCC(VT, N1.getOperand(1),
1252 DAG.getConstant(0, N1.getValueType()), Cond);
1253 } else if (N1.getOperand(1) == N0) {
1254 if (DAG.isCommutativeBinOp(N1.getOpcode())) {
1255 return DAG.getSetCC(VT, N1.getOperand(0),
1256 DAG.getConstant(0, N1.getValueType()), Cond);
1257 } else if (N1.Val->hasOneUse()) {
1258 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
1259 // X == (Z-X) --> X<<1 == Z
1260 SDOperand SH = DAG.getNode(ISD::SHL, N1.getValueType(), N0,
1261 DAG.getConstant(1, getShiftAmountTy()));
1262 if (!DCI.isCalledByLegalizer())
1263 DCI.AddToWorklist(SH.Val);
1264 return DAG.getSetCC(VT, SH, N1.getOperand(0), Cond);
1270 // Fold away ALL boolean setcc's.
1272 if (N0.getValueType() == MVT::i1 && foldBooleans) {
1274 default: assert(0 && "Unknown integer setcc!");
1275 case ISD::SETEQ: // X == Y -> (X^Y)^1
1276 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
1277 N0 = DAG.getNode(ISD::XOR, MVT::i1, Temp, DAG.getConstant(1, MVT::i1));
1278 if (!DCI.isCalledByLegalizer())
1279 DCI.AddToWorklist(Temp.Val);
1281 case ISD::SETNE: // X != Y --> (X^Y)
1282 N0 = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
1284 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> X^1 & Y
1285 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> X^1 & Y
1286 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
1287 N0 = DAG.getNode(ISD::AND, MVT::i1, N1, Temp);
1288 if (!DCI.isCalledByLegalizer())
1289 DCI.AddToWorklist(Temp.Val);
1291 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> Y^1 & X
1292 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> Y^1 & X
1293 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
1294 N0 = DAG.getNode(ISD::AND, MVT::i1, N0, Temp);
1295 if (!DCI.isCalledByLegalizer())
1296 DCI.AddToWorklist(Temp.Val);
1298 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> X^1 | Y
1299 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> X^1 | Y
1300 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
1301 N0 = DAG.getNode(ISD::OR, MVT::i1, N1, Temp);
1302 if (!DCI.isCalledByLegalizer())
1303 DCI.AddToWorklist(Temp.Val);
1305 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> Y^1 | X
1306 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> Y^1 | X
1307 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
1308 N0 = DAG.getNode(ISD::OR, MVT::i1, N0, Temp);
1311 if (VT != MVT::i1) {
1312 if (!DCI.isCalledByLegalizer())
1313 DCI.AddToWorklist(N0.Val);
1314 // FIXME: If running after legalize, we probably can't do this.
1315 N0 = DAG.getNode(ISD::ZERO_EXTEND, VT, N0);
1320 // Could not fold it.
1324 SDOperand TargetLowering::
1325 PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
1326 // Default implementation: no optimization.
1330 //===----------------------------------------------------------------------===//
1331 // Inline Assembler Implementation Methods
1332 //===----------------------------------------------------------------------===//
1334 TargetLowering::ConstraintType
1335 TargetLowering::getConstraintType(const std::string &Constraint) const {
1336 // FIXME: lots more standard ones to handle.
1337 if (Constraint.size() == 1) {
1338 switch (Constraint[0]) {
1340 case 'r': return C_RegisterClass;
1342 case 'o': // offsetable
1343 case 'V': // not offsetable
1345 case 'i': // Simple Integer or Relocatable Constant
1346 case 'n': // Simple Integer
1347 case 's': // Relocatable Constant
1348 case 'X': // Allow ANY value.
1349 case 'I': // Target registers.
1361 if (Constraint.size() > 1 && Constraint[0] == '{' &&
1362 Constraint[Constraint.size()-1] == '}')
1367 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
1368 /// vector. If it is invalid, don't add anything to Ops.
1369 void TargetLowering::LowerAsmOperandForConstraint(SDOperand Op,
1370 char ConstraintLetter,
1371 std::vector<SDOperand> &Ops,
1372 SelectionDAG &DAG) {
1373 switch (ConstraintLetter) {
1375 case 'i': // Simple Integer or Relocatable Constant
1376 case 'n': // Simple Integer
1377 case 's': // Relocatable Constant
1378 case 'X': { // Allows any operand.
1379 // These operands are interested in values of the form (GV+C), where C may
1380 // be folded in as an offset of GV, or it may be explicitly added. Also, it
1381 // is possible and fine if either GV or C are missing.
1382 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
1383 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
1385 // If we have "(add GV, C)", pull out GV/C
1386 if (Op.getOpcode() == ISD::ADD) {
1387 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
1388 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
1389 if (C == 0 || GA == 0) {
1390 C = dyn_cast<ConstantSDNode>(Op.getOperand(0));
1391 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1));
1393 if (C == 0 || GA == 0)
1397 // If we find a valid operand, map to the TargetXXX version so that the
1398 // value itself doesn't get selected.
1399 if (GA) { // Either &GV or &GV+C
1400 if (ConstraintLetter != 'n') {
1401 int64_t Offs = GA->getOffset();
1402 if (C) Offs += C->getValue();
1403 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(),
1404 Op.getValueType(), Offs));
1408 if (C) { // just C, no GV.
1409 // Simple constants are not allowed for 's'.
1410 if (ConstraintLetter != 's') {
1411 Ops.push_back(DAG.getTargetConstant(C->getValue(), Op.getValueType()));
1420 std::vector<unsigned> TargetLowering::
1421 getRegClassForInlineAsmConstraint(const std::string &Constraint,
1422 MVT::ValueType VT) const {
1423 return std::vector<unsigned>();
1427 std::pair<unsigned, const TargetRegisterClass*> TargetLowering::
1428 getRegForInlineAsmConstraint(const std::string &Constraint,
1429 MVT::ValueType VT) const {
1430 if (Constraint[0] != '{')
1431 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
1432 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?");
1434 // Remove the braces from around the name.
1435 std::string RegName(Constraint.begin()+1, Constraint.end()-1);
1437 // Figure out which register class contains this reg.
1438 const MRegisterInfo *RI = TM.getRegisterInfo();
1439 for (MRegisterInfo::regclass_iterator RCI = RI->regclass_begin(),
1440 E = RI->regclass_end(); RCI != E; ++RCI) {
1441 const TargetRegisterClass *RC = *RCI;
1443 // If none of the the value types for this register class are valid, we
1444 // can't use it. For example, 64-bit reg classes on 32-bit targets.
1445 bool isLegal = false;
1446 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
1448 if (isTypeLegal(*I)) {
1454 if (!isLegal) continue;
1456 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end();
1458 if (StringsEqualNoCase(RegName, RI->get(*I).Name))
1459 return std::make_pair(*I, RC);
1463 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
1466 //===----------------------------------------------------------------------===//
1467 // Loop Strength Reduction hooks
1468 //===----------------------------------------------------------------------===//
1470 /// isLegalAddressingMode - Return true if the addressing mode represented
1471 /// by AM is legal for this target, for a load/store of the specified type.
1472 bool TargetLowering::isLegalAddressingMode(const AddrMode &AM,
1473 const Type *Ty) const {
1474 // The default implementation of this implements a conservative RISCy, r+r and
1477 // Allows a sign-extended 16-bit immediate field.
1478 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
1481 // No global is ever allowed as a base.
1485 // Only support r+r,
1487 case 0: // "r+i" or just "i", depending on HasBaseReg.
1490 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
1492 // Otherwise we have r+r or r+i.
1495 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
1497 // Allow 2*r as r+r.
1504 // Magic for divide replacement
1507 int64_t m; // magic number
1508 int64_t s; // shift amount
1512 uint64_t m; // magic number
1513 int64_t a; // add indicator
1514 int64_t s; // shift amount
1517 /// magic - calculate the magic numbers required to codegen an integer sdiv as
1518 /// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
1520 static ms magic32(int32_t d) {
1522 uint32_t ad, anc, delta, q1, r1, q2, r2, t;
1523 const uint32_t two31 = 0x80000000U;
1527 t = two31 + ((uint32_t)d >> 31);
1528 anc = t - 1 - t%ad; // absolute value of nc
1529 p = 31; // initialize p
1530 q1 = two31/anc; // initialize q1 = 2p/abs(nc)
1531 r1 = two31 - q1*anc; // initialize r1 = rem(2p,abs(nc))
1532 q2 = two31/ad; // initialize q2 = 2p/abs(d)
1533 r2 = two31 - q2*ad; // initialize r2 = rem(2p,abs(d))
1536 q1 = 2*q1; // update q1 = 2p/abs(nc)
1537 r1 = 2*r1; // update r1 = rem(2p/abs(nc))
1538 if (r1 >= anc) { // must be unsigned comparison
1542 q2 = 2*q2; // update q2 = 2p/abs(d)
1543 r2 = 2*r2; // update r2 = rem(2p/abs(d))
1544 if (r2 >= ad) { // must be unsigned comparison
1549 } while (q1 < delta || (q1 == delta && r1 == 0));
1551 mag.m = (int32_t)(q2 + 1); // make sure to sign extend
1552 if (d < 0) mag.m = -mag.m; // resulting magic number
1553 mag.s = p - 32; // resulting shift
1557 /// magicu - calculate the magic numbers required to codegen an integer udiv as
1558 /// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
1559 static mu magicu32(uint32_t d) {
1561 uint32_t nc, delta, q1, r1, q2, r2;
1563 magu.a = 0; // initialize "add" indicator
1565 p = 31; // initialize p
1566 q1 = 0x80000000/nc; // initialize q1 = 2p/nc
1567 r1 = 0x80000000 - q1*nc; // initialize r1 = rem(2p,nc)
1568 q2 = 0x7FFFFFFF/d; // initialize q2 = (2p-1)/d
1569 r2 = 0x7FFFFFFF - q2*d; // initialize r2 = rem((2p-1),d)
1572 if (r1 >= nc - r1 ) {
1573 q1 = 2*q1 + 1; // update q1
1574 r1 = 2*r1 - nc; // update r1
1577 q1 = 2*q1; // update q1
1578 r1 = 2*r1; // update r1
1580 if (r2 + 1 >= d - r2) {
1581 if (q2 >= 0x7FFFFFFF) magu.a = 1;
1582 q2 = 2*q2 + 1; // update q2
1583 r2 = 2*r2 + 1 - d; // update r2
1586 if (q2 >= 0x80000000) magu.a = 1;
1587 q2 = 2*q2; // update q2
1588 r2 = 2*r2 + 1; // update r2
1591 } while (p < 64 && (q1 < delta || (q1 == delta && r1 == 0)));
1592 magu.m = q2 + 1; // resulting magic number
1593 magu.s = p - 32; // resulting shift
1597 /// magic - calculate the magic numbers required to codegen an integer sdiv as
1598 /// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
1600 static ms magic64(int64_t d) {
1602 uint64_t ad, anc, delta, q1, r1, q2, r2, t;
1603 const uint64_t two63 = 9223372036854775808ULL; // 2^63
1606 ad = d >= 0 ? d : -d;
1607 t = two63 + ((uint64_t)d >> 63);
1608 anc = t - 1 - t%ad; // absolute value of nc
1609 p = 63; // initialize p
1610 q1 = two63/anc; // initialize q1 = 2p/abs(nc)
1611 r1 = two63 - q1*anc; // initialize r1 = rem(2p,abs(nc))
1612 q2 = two63/ad; // initialize q2 = 2p/abs(d)
1613 r2 = two63 - q2*ad; // initialize r2 = rem(2p,abs(d))
1616 q1 = 2*q1; // update q1 = 2p/abs(nc)
1617 r1 = 2*r1; // update r1 = rem(2p/abs(nc))
1618 if (r1 >= anc) { // must be unsigned comparison
1622 q2 = 2*q2; // update q2 = 2p/abs(d)
1623 r2 = 2*r2; // update r2 = rem(2p/abs(d))
1624 if (r2 >= ad) { // must be unsigned comparison
1629 } while (q1 < delta || (q1 == delta && r1 == 0));
1632 if (d < 0) mag.m = -mag.m; // resulting magic number
1633 mag.s = p - 64; // resulting shift
1637 /// magicu - calculate the magic numbers required to codegen an integer udiv as
1638 /// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
1639 static mu magicu64(uint64_t d)
1642 uint64_t nc, delta, q1, r1, q2, r2;
1644 magu.a = 0; // initialize "add" indicator
1646 p = 63; // initialize p
1647 q1 = 0x8000000000000000ull/nc; // initialize q1 = 2p/nc
1648 r1 = 0x8000000000000000ull - q1*nc; // initialize r1 = rem(2p,nc)
1649 q2 = 0x7FFFFFFFFFFFFFFFull/d; // initialize q2 = (2p-1)/d
1650 r2 = 0x7FFFFFFFFFFFFFFFull - q2*d; // initialize r2 = rem((2p-1),d)
1653 if (r1 >= nc - r1 ) {
1654 q1 = 2*q1 + 1; // update q1
1655 r1 = 2*r1 - nc; // update r1
1658 q1 = 2*q1; // update q1
1659 r1 = 2*r1; // update r1
1661 if (r2 + 1 >= d - r2) {
1662 if (q2 >= 0x7FFFFFFFFFFFFFFFull) magu.a = 1;
1663 q2 = 2*q2 + 1; // update q2
1664 r2 = 2*r2 + 1 - d; // update r2
1667 if (q2 >= 0x8000000000000000ull) magu.a = 1;
1668 q2 = 2*q2; // update q2
1669 r2 = 2*r2 + 1; // update r2
1672 } while (p < 128 && (q1 < delta || (q1 == delta && r1 == 0)));
1673 magu.m = q2 + 1; // resulting magic number
1674 magu.s = p - 64; // resulting shift
1678 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
1679 /// return a DAG expression to select that will generate the same value by
1680 /// multiplying by a magic number. See:
1681 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
1682 SDOperand TargetLowering::BuildSDIV(SDNode *N, SelectionDAG &DAG,
1683 std::vector<SDNode*>* Created) const {
1684 MVT::ValueType VT = N->getValueType(0);
1686 // Check to see if we can do this.
1687 if (!isTypeLegal(VT) || (VT != MVT::i32 && VT != MVT::i64))
1688 return SDOperand(); // BuildSDIV only operates on i32 or i64
1689 if (!isOperationLegal(ISD::MULHS, VT))
1690 return SDOperand(); // Make sure the target supports MULHS.
1692 int64_t d = cast<ConstantSDNode>(N->getOperand(1))->getSignExtended();
1693 ms magics = (VT == MVT::i32) ? magic32(d) : magic64(d);
1695 // Multiply the numerator (operand 0) by the magic value
1696 SDOperand Q = DAG.getNode(ISD::MULHS, VT, N->getOperand(0),
1697 DAG.getConstant(magics.m, VT));
1698 // If d > 0 and m < 0, add the numerator
1699 if (d > 0 && magics.m < 0) {
1700 Q = DAG.getNode(ISD::ADD, VT, Q, N->getOperand(0));
1702 Created->push_back(Q.Val);
1704 // If d < 0 and m > 0, subtract the numerator.
1705 if (d < 0 && magics.m > 0) {
1706 Q = DAG.getNode(ISD::SUB, VT, Q, N->getOperand(0));
1708 Created->push_back(Q.Val);
1710 // Shift right algebraic if shift value is nonzero
1712 Q = DAG.getNode(ISD::SRA, VT, Q,
1713 DAG.getConstant(magics.s, getShiftAmountTy()));
1715 Created->push_back(Q.Val);
1717 // Extract the sign bit and add it to the quotient
1719 DAG.getNode(ISD::SRL, VT, Q, DAG.getConstant(MVT::getSizeInBits(VT)-1,
1720 getShiftAmountTy()));
1722 Created->push_back(T.Val);
1723 return DAG.getNode(ISD::ADD, VT, Q, T);
1726 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
1727 /// return a DAG expression to select that will generate the same value by
1728 /// multiplying by a magic number. See:
1729 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
1730 SDOperand TargetLowering::BuildUDIV(SDNode *N, SelectionDAG &DAG,
1731 std::vector<SDNode*>* Created) const {
1732 MVT::ValueType VT = N->getValueType(0);
1734 // Check to see if we can do this.
1735 if (!isTypeLegal(VT) || (VT != MVT::i32 && VT != MVT::i64))
1736 return SDOperand(); // BuildUDIV only operates on i32 or i64
1737 if (!isOperationLegal(ISD::MULHU, VT))
1738 return SDOperand(); // Make sure the target supports MULHU.
1740 uint64_t d = cast<ConstantSDNode>(N->getOperand(1))->getValue();
1741 mu magics = (VT == MVT::i32) ? magicu32(d) : magicu64(d);
1743 // Multiply the numerator (operand 0) by the magic value
1744 SDOperand Q = DAG.getNode(ISD::MULHU, VT, N->getOperand(0),
1745 DAG.getConstant(magics.m, VT));
1747 Created->push_back(Q.Val);
1749 if (magics.a == 0) {
1750 return DAG.getNode(ISD::SRL, VT, Q,
1751 DAG.getConstant(magics.s, getShiftAmountTy()));
1753 SDOperand NPQ = DAG.getNode(ISD::SUB, VT, N->getOperand(0), Q);
1755 Created->push_back(NPQ.Val);
1756 NPQ = DAG.getNode(ISD::SRL, VT, NPQ,
1757 DAG.getConstant(1, getShiftAmountTy()));
1759 Created->push_back(NPQ.Val);
1760 NPQ = DAG.getNode(ISD::ADD, VT, NPQ, Q);
1762 Created->push_back(NPQ.Val);
1763 return DAG.getNode(ISD::SRL, VT, NPQ,
1764 DAG.getConstant(magics.s-1, getShiftAmountTy()));