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/Support/MathExtras.h"
24 /// InitLibcallNames - Set default libcall names.
26 static void InitLibcallNames(const char **Names) {
27 Names[RTLIB::SHL_I32] = "__ashlsi3";
28 Names[RTLIB::SHL_I64] = "__ashldi3";
29 Names[RTLIB::SRL_I32] = "__lshrsi3";
30 Names[RTLIB::SRL_I64] = "__lshrdi3";
31 Names[RTLIB::SRA_I32] = "__ashrsi3";
32 Names[RTLIB::SRA_I64] = "__ashrdi3";
33 Names[RTLIB::MUL_I32] = "__mulsi3";
34 Names[RTLIB::MUL_I64] = "__muldi3";
35 Names[RTLIB::SDIV_I32] = "__divsi3";
36 Names[RTLIB::SDIV_I64] = "__divdi3";
37 Names[RTLIB::UDIV_I32] = "__udivsi3";
38 Names[RTLIB::UDIV_I64] = "__udivdi3";
39 Names[RTLIB::SREM_I32] = "__modsi3";
40 Names[RTLIB::SREM_I64] = "__moddi3";
41 Names[RTLIB::UREM_I32] = "__umodsi3";
42 Names[RTLIB::UREM_I64] = "__umoddi3";
43 Names[RTLIB::NEG_I32] = "__negsi2";
44 Names[RTLIB::NEG_I64] = "__negdi2";
45 Names[RTLIB::ADD_F32] = "__addsf3";
46 Names[RTLIB::ADD_F64] = "__adddf3";
47 Names[RTLIB::SUB_F32] = "__subsf3";
48 Names[RTLIB::SUB_F64] = "__subdf3";
49 Names[RTLIB::MUL_F32] = "__mulsf3";
50 Names[RTLIB::MUL_F64] = "__muldf3";
51 Names[RTLIB::DIV_F32] = "__divsf3";
52 Names[RTLIB::DIV_F64] = "__divdf3";
53 Names[RTLIB::REM_F32] = "fmodf";
54 Names[RTLIB::REM_F64] = "fmod";
55 Names[RTLIB::NEG_F32] = "__negsf2";
56 Names[RTLIB::NEG_F64] = "__negdf2";
57 Names[RTLIB::POWI_F32] = "__powisf2";
58 Names[RTLIB::POWI_F64] = "__powidf2";
59 Names[RTLIB::SQRT_F32] = "sqrtf";
60 Names[RTLIB::SQRT_F64] = "sqrt";
61 Names[RTLIB::SIN_F32] = "sinf";
62 Names[RTLIB::SIN_F64] = "sin";
63 Names[RTLIB::COS_F32] = "cosf";
64 Names[RTLIB::COS_F64] = "cos";
65 Names[RTLIB::FPEXT_F32_F64] = "__extendsfdf2";
66 Names[RTLIB::FPROUND_F64_F32] = "__truncdfsf2";
67 Names[RTLIB::FPTOSINT_F32_I32] = "__fixsfsi";
68 Names[RTLIB::FPTOSINT_F32_I64] = "__fixsfdi";
69 Names[RTLIB::FPTOSINT_F64_I32] = "__fixdfsi";
70 Names[RTLIB::FPTOSINT_F64_I64] = "__fixdfdi";
71 Names[RTLIB::FPTOUINT_F32_I32] = "__fixunssfsi";
72 Names[RTLIB::FPTOUINT_F32_I64] = "__fixunssfdi";
73 Names[RTLIB::FPTOUINT_F64_I32] = "__fixunsdfsi";
74 Names[RTLIB::FPTOUINT_F64_I64] = "__fixunsdfdi";
75 Names[RTLIB::SINTTOFP_I32_F32] = "__floatsisf";
76 Names[RTLIB::SINTTOFP_I32_F64] = "__floatsidf";
77 Names[RTLIB::SINTTOFP_I64_F32] = "__floatdisf";
78 Names[RTLIB::SINTTOFP_I64_F64] = "__floatdidf";
79 Names[RTLIB::UINTTOFP_I32_F32] = "__floatunsisf";
80 Names[RTLIB::UINTTOFP_I32_F64] = "__floatunsidf";
81 Names[RTLIB::UINTTOFP_I64_F32] = "__floatundisf";
82 Names[RTLIB::UINTTOFP_I64_F64] = "__floatundidf";
83 Names[RTLIB::OEQ_F32] = "__eqsf2";
84 Names[RTLIB::OEQ_F64] = "__eqdf2";
85 Names[RTLIB::UNE_F32] = "__nesf2";
86 Names[RTLIB::UNE_F64] = "__nedf2";
87 Names[RTLIB::OGE_F32] = "__gesf2";
88 Names[RTLIB::OGE_F64] = "__gedf2";
89 Names[RTLIB::OLT_F32] = "__ltsf2";
90 Names[RTLIB::OLT_F64] = "__ltdf2";
91 Names[RTLIB::OLE_F32] = "__lesf2";
92 Names[RTLIB::OLE_F64] = "__ledf2";
93 Names[RTLIB::OGT_F32] = "__gtsf2";
94 Names[RTLIB::OGT_F64] = "__gtdf2";
95 Names[RTLIB::UO_F32] = "__unordsf2";
96 Names[RTLIB::UO_F64] = "__unorddf2";
97 Names[RTLIB::O_F32] = "__unordsf2";
98 Names[RTLIB::O_F64] = "__unorddf2";
101 /// InitCmpLibcallCCs - Set default comparison libcall CC.
103 static void InitCmpLibcallCCs(ISD::CondCode *CCs) {
104 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL);
105 CCs[RTLIB::OEQ_F32] = ISD::SETEQ;
106 CCs[RTLIB::OEQ_F64] = ISD::SETEQ;
107 CCs[RTLIB::UNE_F32] = ISD::SETNE;
108 CCs[RTLIB::UNE_F64] = ISD::SETNE;
109 CCs[RTLIB::OGE_F32] = ISD::SETGE;
110 CCs[RTLIB::OGE_F64] = ISD::SETGE;
111 CCs[RTLIB::OLT_F32] = ISD::SETLT;
112 CCs[RTLIB::OLT_F64] = ISD::SETLT;
113 CCs[RTLIB::OLE_F32] = ISD::SETLE;
114 CCs[RTLIB::OLE_F64] = ISD::SETLE;
115 CCs[RTLIB::OGT_F32] = ISD::SETGT;
116 CCs[RTLIB::OGT_F64] = ISD::SETGT;
117 CCs[RTLIB::UO_F32] = ISD::SETNE;
118 CCs[RTLIB::UO_F64] = ISD::SETNE;
119 CCs[RTLIB::O_F32] = ISD::SETEQ;
120 CCs[RTLIB::O_F64] = ISD::SETEQ;
123 TargetLowering::TargetLowering(TargetMachine &tm)
124 : TM(tm), TD(TM.getTargetData()) {
125 assert(ISD::BUILTIN_OP_END <= 156 &&
126 "Fixed size array in TargetLowering is not large enough!");
127 // All operations default to being supported.
128 memset(OpActions, 0, sizeof(OpActions));
129 memset(LoadXActions, 0, sizeof(LoadXActions));
130 memset(&StoreXActions, 0, sizeof(StoreXActions));
131 // Initialize all indexed load / store to expand.
132 for (unsigned VT = 0; VT != (unsigned)MVT::LAST_VALUETYPE; ++VT) {
133 for (unsigned IM = (unsigned)ISD::PRE_INC;
134 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
135 setIndexedLoadAction(IM, (MVT::ValueType)VT, Expand);
136 setIndexedStoreAction(IM, (MVT::ValueType)VT, Expand);
140 IsLittleEndian = TD->isLittleEndian();
141 UsesGlobalOffsetTable = false;
142 ShiftAmountTy = SetCCResultTy = PointerTy = getValueType(TD->getIntPtrType());
143 ShiftAmtHandling = Undefined;
144 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE*sizeof(TargetRegisterClass*));
145 memset(TargetDAGCombineArray, 0,
146 sizeof(TargetDAGCombineArray)/sizeof(TargetDAGCombineArray[0]));
147 maxStoresPerMemset = maxStoresPerMemcpy = maxStoresPerMemmove = 8;
148 allowUnalignedMemoryAccesses = false;
149 UseUnderscoreSetJmp = false;
150 UseUnderscoreLongJmp = false;
151 SelectIsExpensive = false;
152 IntDivIsCheap = false;
153 Pow2DivIsCheap = false;
154 StackPointerRegisterToSaveRestore = 0;
155 ExceptionPointerRegister = 0;
156 ExceptionSelectorRegister = 0;
157 SchedPreferenceInfo = SchedulingForLatency;
159 JumpBufAlignment = 0;
160 IfCvtBlockSizeLimit = 2;
162 InitLibcallNames(LibcallRoutineNames);
163 InitCmpLibcallCCs(CmpLibcallCCs);
166 TargetLowering::~TargetLowering() {}
168 /// setValueTypeAction - Set the action for a particular value type. This
169 /// assumes an action has not already been set for this value type.
170 static void SetValueTypeAction(MVT::ValueType VT,
171 TargetLowering::LegalizeAction Action,
173 MVT::ValueType *TransformToType,
174 TargetLowering::ValueTypeActionImpl &ValueTypeActions) {
175 ValueTypeActions.setTypeAction(VT, Action);
176 if (Action == TargetLowering::Promote) {
177 MVT::ValueType PromoteTo;
179 PromoteTo = MVT::f64;
181 unsigned LargerReg = VT+1;
182 while (!TLI.isTypeLegal((MVT::ValueType)LargerReg)) {
184 assert(MVT::isInteger((MVT::ValueType)LargerReg) &&
185 "Nothing to promote to??");
187 PromoteTo = (MVT::ValueType)LargerReg;
190 assert(MVT::isInteger(VT) == MVT::isInteger(PromoteTo) &&
191 MVT::isFloatingPoint(VT) == MVT::isFloatingPoint(PromoteTo) &&
192 "Can only promote from int->int or fp->fp!");
193 assert(VT < PromoteTo && "Must promote to a larger type!");
194 TransformToType[VT] = PromoteTo;
195 } else if (Action == TargetLowering::Expand) {
196 // f32 and f64 is each expanded to corresponding integer type of same size.
198 TransformToType[VT] = MVT::i32;
199 else if (VT == MVT::f64)
200 TransformToType[VT] = MVT::i64;
202 assert((MVT::isVector(VT) || MVT::isInteger(VT)) && VT > MVT::i8 &&
203 "Cannot expand this type: target must support SOME integer reg!");
204 // Expand to the next smaller integer type!
205 TransformToType[VT] = (MVT::ValueType)(VT-1);
211 /// computeRegisterProperties - Once all of the register classes are added,
212 /// this allows us to compute derived properties we expose.
213 void TargetLowering::computeRegisterProperties() {
214 assert(MVT::LAST_VALUETYPE <= 32 &&
215 "Too many value types for ValueTypeActions to hold!");
217 // Everything defaults to one.
218 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i)
219 NumRegistersForVT[i] = 1;
221 // Find the largest integer register class.
222 unsigned LargestIntReg = MVT::i128;
223 for (; RegClassForVT[LargestIntReg] == 0; --LargestIntReg)
224 assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
226 // Every integer value type larger than this largest register takes twice as
227 // many registers to represent as the previous ValueType.
228 unsigned ExpandedReg = LargestIntReg; ++LargestIntReg;
229 for (++ExpandedReg; MVT::isInteger((MVT::ValueType)ExpandedReg);++ExpandedReg)
230 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
232 // Inspect all of the ValueType's possible, deciding how to process them.
233 for (unsigned IntReg = MVT::i1; IntReg <= MVT::i128; ++IntReg)
234 // If we are expanding this type, expand it!
235 if (getNumRegisters((MVT::ValueType)IntReg) != 1)
236 SetValueTypeAction((MVT::ValueType)IntReg, Expand, *this, TransformToType,
238 else if (!isTypeLegal((MVT::ValueType)IntReg))
239 // Otherwise, if we don't have native support, we must promote to a
241 SetValueTypeAction((MVT::ValueType)IntReg, Promote, *this,
242 TransformToType, ValueTypeActions);
244 TransformToType[(MVT::ValueType)IntReg] = (MVT::ValueType)IntReg;
246 // If the target does not have native f64 support, expand it to i64. We will
247 // be generating soft float library calls. If the target does not have native
248 // support for f32, promote it to f64 if it is legal. Otherwise, expand it to
250 if (isTypeLegal(MVT::f64))
251 TransformToType[MVT::f64] = MVT::f64;
253 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
254 SetValueTypeAction(MVT::f64, Expand, *this, TransformToType,
257 if (isTypeLegal(MVT::f32))
258 TransformToType[MVT::f32] = MVT::f32;
259 else if (isTypeLegal(MVT::f64))
260 SetValueTypeAction(MVT::f32, Promote, *this, TransformToType,
263 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
264 SetValueTypeAction(MVT::f32, Expand, *this, TransformToType,
268 // Loop over all of the legal vector value types, specifying an identity type
270 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
271 i <= MVT::LAST_VECTOR_VALUETYPE; ++i) {
272 if (isTypeLegal((MVT::ValueType)i))
273 TransformToType[i] = (MVT::ValueType)i;
275 MVT::ValueType VT1, VT2;
276 NumRegistersForVT[i] = getVectorTypeBreakdown(i, VT1, VT2);
277 SetValueTypeAction(i, Expand, *this, TransformToType,
283 const char *TargetLowering::getTargetNodeName(unsigned Opcode) const {
287 /// getVectorTypeBreakdown - Vector types are broken down into some number of
288 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
289 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
290 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
292 /// This method returns the number of registers needed, and the VT for each
293 /// register. It also returns the VT of the VectorType elements before they
294 /// are promoted/expanded.
296 unsigned TargetLowering::getVectorTypeBreakdown(MVT::ValueType VT,
297 MVT::ValueType &ElementVT,
298 MVT::ValueType &LegalElementVT) const {
299 // Figure out the right, legal destination reg to copy into.
300 unsigned NumElts = MVT::getVectorNumElements(VT);
301 MVT::ValueType EltTy = MVT::getVectorElementType(VT);
303 unsigned NumVectorRegs = 1;
305 // Divide the input until we get to a supported size. This will always
306 // end with a scalar if the target doesn't support vectors.
307 while (NumElts > 1 &&
308 !isTypeLegal(MVT::getVectorType(EltTy, NumElts))) {
313 MVT::ValueType NewVT = MVT::getVectorType(EltTy, NumElts);
314 if (!isTypeLegal(NewVT))
318 MVT::ValueType DestVT = getTypeToTransformTo(NewVT);
319 LegalElementVT = DestVT;
320 if (DestVT < NewVT) {
321 // Value is expanded, e.g. i64 -> i16.
322 return NumVectorRegs*(MVT::getSizeInBits(NewVT)/MVT::getSizeInBits(DestVT));
324 // Otherwise, promotion or legal types use the same number of registers as
325 // the vector decimated to the appropriate level.
326 return NumVectorRegs;
332 //===----------------------------------------------------------------------===//
333 // Optimization Methods
334 //===----------------------------------------------------------------------===//
336 /// ShrinkDemandedConstant - Check to see if the specified operand of the
337 /// specified instruction is a constant integer. If so, check to see if there
338 /// are any bits set in the constant that are not demanded. If so, shrink the
339 /// constant and return true.
340 bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDOperand Op,
342 // FIXME: ISD::SELECT, ISD::SELECT_CC
343 switch(Op.getOpcode()) {
348 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1)))
349 if ((~Demanded & C->getValue()) != 0) {
350 MVT::ValueType VT = Op.getValueType();
351 SDOperand New = DAG.getNode(Op.getOpcode(), VT, Op.getOperand(0),
352 DAG.getConstant(Demanded & C->getValue(),
354 return CombineTo(Op, New);
361 /// SimplifyDemandedBits - Look at Op. At this point, we know that only the
362 /// DemandedMask bits of the result of Op are ever used downstream. If we can
363 /// use this information to simplify Op, create a new simplified DAG node and
364 /// return true, returning the original and new nodes in Old and New. Otherwise,
365 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
366 /// the expression (used to simplify the caller). The KnownZero/One bits may
367 /// only be accurate for those bits in the DemandedMask.
368 bool TargetLowering::SimplifyDemandedBits(SDOperand Op, uint64_t DemandedMask,
371 TargetLoweringOpt &TLO,
372 unsigned Depth) const {
373 KnownZero = KnownOne = 0; // Don't know anything.
375 // The masks are not wide enough to represent this type! Should use APInt.
376 if (Op.getValueType() == MVT::i128)
379 // Other users may use these bits.
380 if (!Op.Val->hasOneUse()) {
382 // If not at the root, Just compute the KnownZero/KnownOne bits to
383 // simplify things downstream.
384 TLO.DAG.ComputeMaskedBits(Op, DemandedMask, KnownZero, KnownOne, Depth);
387 // If this is the root being simplified, allow it to have multiple uses,
388 // just set the DemandedMask to all bits.
389 DemandedMask = MVT::getIntVTBitMask(Op.getValueType());
390 } else if (DemandedMask == 0) {
391 // Not demanding any bits from Op.
392 if (Op.getOpcode() != ISD::UNDEF)
393 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::UNDEF, Op.getValueType()));
395 } else if (Depth == 6) { // Limit search depth.
399 uint64_t KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut;
400 switch (Op.getOpcode()) {
402 // We know all of the bits for a constant!
403 KnownOne = cast<ConstantSDNode>(Op)->getValue() & DemandedMask;
404 KnownZero = ~KnownOne & DemandedMask;
405 return false; // Don't fall through, will infinitely loop.
407 // If the RHS is a constant, check to see if the LHS would be zero without
408 // using the bits from the RHS. Below, we use knowledge about the RHS to
409 // simplify the LHS, here we're using information from the LHS to simplify
411 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
412 uint64_t LHSZero, LHSOne;
413 TLO.DAG.ComputeMaskedBits(Op.getOperand(0), DemandedMask,
414 LHSZero, LHSOne, Depth+1);
415 // If the LHS already has zeros where RHSC does, this and is dead.
416 if ((LHSZero & DemandedMask) == (~RHSC->getValue() & DemandedMask))
417 return TLO.CombineTo(Op, Op.getOperand(0));
418 // If any of the set bits in the RHS are known zero on the LHS, shrink
420 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & DemandedMask))
424 if (SimplifyDemandedBits(Op.getOperand(1), DemandedMask, KnownZero,
425 KnownOne, TLO, Depth+1))
427 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
428 if (SimplifyDemandedBits(Op.getOperand(0), DemandedMask & ~KnownZero,
429 KnownZero2, KnownOne2, TLO, Depth+1))
431 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
433 // If all of the demanded bits are known one on one side, return the other.
434 // These bits cannot contribute to the result of the 'and'.
435 if ((DemandedMask & ~KnownZero2 & KnownOne)==(DemandedMask & ~KnownZero2))
436 return TLO.CombineTo(Op, Op.getOperand(0));
437 if ((DemandedMask & ~KnownZero & KnownOne2)==(DemandedMask & ~KnownZero))
438 return TLO.CombineTo(Op, Op.getOperand(1));
439 // If all of the demanded bits in the inputs are known zeros, return zero.
440 if ((DemandedMask & (KnownZero|KnownZero2)) == DemandedMask)
441 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, Op.getValueType()));
442 // If the RHS is a constant, see if we can simplify it.
443 if (TLO.ShrinkDemandedConstant(Op, DemandedMask & ~KnownZero2))
446 // Output known-1 bits are only known if set in both the LHS & RHS.
447 KnownOne &= KnownOne2;
448 // Output known-0 are known to be clear if zero in either the LHS | RHS.
449 KnownZero |= KnownZero2;
452 if (SimplifyDemandedBits(Op.getOperand(1), DemandedMask, KnownZero,
453 KnownOne, TLO, Depth+1))
455 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
456 if (SimplifyDemandedBits(Op.getOperand(0), DemandedMask & ~KnownOne,
457 KnownZero2, KnownOne2, TLO, Depth+1))
459 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
461 // If all of the demanded bits are known zero on one side, return the other.
462 // These bits cannot contribute to the result of the 'or'.
463 if ((DemandedMask & ~KnownOne2 & KnownZero) == (DemandedMask & ~KnownOne2))
464 return TLO.CombineTo(Op, Op.getOperand(0));
465 if ((DemandedMask & ~KnownOne & KnownZero2) == (DemandedMask & ~KnownOne))
466 return TLO.CombineTo(Op, Op.getOperand(1));
467 // If all of the potentially set bits on one side are known to be set on
468 // the other side, just use the 'other' side.
469 if ((DemandedMask & (~KnownZero) & KnownOne2) ==
470 (DemandedMask & (~KnownZero)))
471 return TLO.CombineTo(Op, Op.getOperand(0));
472 if ((DemandedMask & (~KnownZero2) & KnownOne) ==
473 (DemandedMask & (~KnownZero2)))
474 return TLO.CombineTo(Op, Op.getOperand(1));
475 // If the RHS is a constant, see if we can simplify it.
476 if (TLO.ShrinkDemandedConstant(Op, DemandedMask))
479 // Output known-0 bits are only known if clear in both the LHS & RHS.
480 KnownZero &= KnownZero2;
481 // Output known-1 are known to be set if set in either the LHS | RHS.
482 KnownOne |= KnownOne2;
485 if (SimplifyDemandedBits(Op.getOperand(1), DemandedMask, KnownZero,
486 KnownOne, TLO, Depth+1))
488 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
489 if (SimplifyDemandedBits(Op.getOperand(0), DemandedMask, KnownZero2,
490 KnownOne2, TLO, Depth+1))
492 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
494 // If all of the demanded bits are known zero on one side, return the other.
495 // These bits cannot contribute to the result of the 'xor'.
496 if ((DemandedMask & KnownZero) == DemandedMask)
497 return TLO.CombineTo(Op, Op.getOperand(0));
498 if ((DemandedMask & KnownZero2) == DemandedMask)
499 return TLO.CombineTo(Op, Op.getOperand(1));
501 // If all of the unknown bits are known to be zero on one side or the other
502 // (but not both) turn this into an *inclusive* or.
503 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
504 if ((DemandedMask & ~KnownZero & ~KnownZero2) == 0)
505 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, Op.getValueType(),
509 // Output known-0 bits are known if clear or set in both the LHS & RHS.
510 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
511 // Output known-1 are known to be set if set in only one of the LHS, RHS.
512 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
514 // If all of the demanded bits on one side are known, and all of the set
515 // bits on that side are also known to be set on the other side, turn this
516 // into an AND, as we know the bits will be cleared.
517 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
518 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask) { // all known
519 if ((KnownOne & KnownOne2) == KnownOne) {
520 MVT::ValueType VT = Op.getValueType();
521 SDOperand ANDC = TLO.DAG.getConstant(~KnownOne & DemandedMask, VT);
522 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, VT, Op.getOperand(0),
527 // If the RHS is a constant, see if we can simplify it.
528 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
529 if (TLO.ShrinkDemandedConstant(Op, DemandedMask))
532 KnownZero = KnownZeroOut;
533 KnownOne = KnownOneOut;
536 // If we know the result of a setcc has the top bits zero, use this info.
537 if (getSetCCResultContents() == TargetLowering::ZeroOrOneSetCCResult)
538 KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL);
541 if (SimplifyDemandedBits(Op.getOperand(2), DemandedMask, KnownZero,
542 KnownOne, TLO, Depth+1))
544 if (SimplifyDemandedBits(Op.getOperand(1), DemandedMask, KnownZero2,
545 KnownOne2, TLO, Depth+1))
547 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
548 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
550 // If the operands are constants, see if we can simplify them.
551 if (TLO.ShrinkDemandedConstant(Op, DemandedMask))
554 // Only known if known in both the LHS and RHS.
555 KnownOne &= KnownOne2;
556 KnownZero &= KnownZero2;
559 if (SimplifyDemandedBits(Op.getOperand(3), DemandedMask, KnownZero,
560 KnownOne, TLO, Depth+1))
562 if (SimplifyDemandedBits(Op.getOperand(2), DemandedMask, KnownZero2,
563 KnownOne2, TLO, Depth+1))
565 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
566 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
568 // If the operands are constants, see if we can simplify them.
569 if (TLO.ShrinkDemandedConstant(Op, DemandedMask))
572 // Only known if known in both the LHS and RHS.
573 KnownOne &= KnownOne2;
574 KnownZero &= KnownZero2;
577 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
578 unsigned ShAmt = SA->getValue();
579 SDOperand InOp = Op.getOperand(0);
581 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a
582 // single shift. We can do this if the bottom bits (which are shifted
583 // out) are never demanded.
584 if (InOp.getOpcode() == ISD::SRL &&
585 isa<ConstantSDNode>(InOp.getOperand(1))) {
586 if (ShAmt && (DemandedMask & ((1ULL << ShAmt)-1)) == 0) {
587 unsigned C1 = cast<ConstantSDNode>(InOp.getOperand(1))->getValue();
588 unsigned Opc = ISD::SHL;
596 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
597 MVT::ValueType VT = Op.getValueType();
598 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, VT,
599 InOp.getOperand(0), NewSA));
603 if (SimplifyDemandedBits(Op.getOperand(0), DemandedMask >> ShAmt,
604 KnownZero, KnownOne, TLO, Depth+1))
606 KnownZero <<= SA->getValue();
607 KnownOne <<= SA->getValue();
608 KnownZero |= (1ULL << SA->getValue())-1; // low bits known zero.
612 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
613 MVT::ValueType VT = Op.getValueType();
614 unsigned ShAmt = SA->getValue();
615 uint64_t TypeMask = MVT::getIntVTBitMask(VT);
616 unsigned VTSize = MVT::getSizeInBits(VT);
617 SDOperand InOp = Op.getOperand(0);
619 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a
620 // single shift. We can do this if the top bits (which are shifted out)
621 // are never demanded.
622 if (InOp.getOpcode() == ISD::SHL &&
623 isa<ConstantSDNode>(InOp.getOperand(1))) {
624 if (ShAmt && (DemandedMask & (~0ULL << (VTSize-ShAmt))) == 0) {
625 unsigned C1 = cast<ConstantSDNode>(InOp.getOperand(1))->getValue();
626 unsigned Opc = ISD::SRL;
634 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
635 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, VT,
636 InOp.getOperand(0), NewSA));
640 // Compute the new bits that are at the top now.
641 if (SimplifyDemandedBits(InOp, (DemandedMask << ShAmt) & TypeMask,
642 KnownZero, KnownOne, TLO, Depth+1))
644 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
645 KnownZero &= TypeMask;
646 KnownOne &= TypeMask;
650 uint64_t HighBits = (1ULL << ShAmt)-1;
651 HighBits <<= VTSize - ShAmt;
652 KnownZero |= HighBits; // High bits known zero.
656 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
657 MVT::ValueType VT = Op.getValueType();
658 unsigned ShAmt = SA->getValue();
660 // Compute the new bits that are at the top now.
661 uint64_t TypeMask = MVT::getIntVTBitMask(VT);
663 uint64_t InDemandedMask = (DemandedMask << ShAmt) & TypeMask;
665 // If any of the demanded bits are produced by the sign extension, we also
666 // demand the input sign bit.
667 uint64_t HighBits = (1ULL << ShAmt)-1;
668 HighBits <<= MVT::getSizeInBits(VT) - ShAmt;
669 if (HighBits & DemandedMask)
670 InDemandedMask |= MVT::getIntVTSignBit(VT);
672 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask,
673 KnownZero, KnownOne, TLO, Depth+1))
675 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
676 KnownZero &= TypeMask;
677 KnownOne &= TypeMask;
681 // Handle the sign bits.
682 uint64_t SignBit = MVT::getIntVTSignBit(VT);
683 SignBit >>= ShAmt; // Adjust to where it is now in the mask.
685 // If the input sign bit is known to be zero, or if none of the top bits
686 // are demanded, turn this into an unsigned shift right.
687 if ((KnownZero & SignBit) || (HighBits & ~DemandedMask) == HighBits) {
688 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, VT, Op.getOperand(0),
690 } else if (KnownOne & SignBit) { // New bits are known one.
691 KnownOne |= HighBits;
695 case ISD::SIGN_EXTEND_INREG: {
696 MVT::ValueType EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
698 // Sign extension. Compute the demanded bits in the result that are not
699 // present in the input.
700 uint64_t NewBits = ~MVT::getIntVTBitMask(EVT) & DemandedMask;
702 // If none of the extended bits are demanded, eliminate the sextinreg.
704 return TLO.CombineTo(Op, Op.getOperand(0));
706 uint64_t InSignBit = MVT::getIntVTSignBit(EVT);
707 int64_t InputDemandedBits = DemandedMask & MVT::getIntVTBitMask(EVT);
709 // Since the sign extended bits are demanded, we know that the sign
711 InputDemandedBits |= InSignBit;
713 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits,
714 KnownZero, KnownOne, TLO, Depth+1))
716 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
718 // If the sign bit of the input is known set or clear, then we know the
719 // top bits of the result.
721 // If the input sign bit is known zero, convert this into a zero extension.
722 if (KnownZero & InSignBit)
723 return TLO.CombineTo(Op,
724 TLO.DAG.getZeroExtendInReg(Op.getOperand(0), EVT));
726 if (KnownOne & InSignBit) { // Input sign bit known set
728 KnownZero &= ~NewBits;
729 } else { // Input sign bit unknown
730 KnownZero &= ~NewBits;
731 KnownOne &= ~NewBits;
738 MVT::ValueType VT = Op.getValueType();
739 unsigned LowBits = Log2_32(MVT::getSizeInBits(VT))+1;
740 KnownZero = ~((1ULL << LowBits)-1) & MVT::getIntVTBitMask(VT);
745 if (ISD::isZEXTLoad(Op.Val)) {
746 LoadSDNode *LD = cast<LoadSDNode>(Op);
747 MVT::ValueType VT = LD->getLoadedVT();
748 KnownZero |= ~MVT::getIntVTBitMask(VT) & DemandedMask;
752 case ISD::ZERO_EXTEND: {
753 uint64_t InMask = MVT::getIntVTBitMask(Op.getOperand(0).getValueType());
755 // If none of the top bits are demanded, convert this into an any_extend.
756 uint64_t NewBits = (~InMask) & DemandedMask;
758 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND,
762 if (SimplifyDemandedBits(Op.getOperand(0), DemandedMask & InMask,
763 KnownZero, KnownOne, TLO, Depth+1))
765 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
766 KnownZero |= NewBits;
769 case ISD::SIGN_EXTEND: {
770 MVT::ValueType InVT = Op.getOperand(0).getValueType();
771 uint64_t InMask = MVT::getIntVTBitMask(InVT);
772 uint64_t InSignBit = MVT::getIntVTSignBit(InVT);
773 uint64_t NewBits = (~InMask) & DemandedMask;
775 // If none of the top bits are demanded, convert this into an any_extend.
777 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND,Op.getValueType(),
780 // Since some of the sign extended bits are demanded, we know that the sign
782 uint64_t InDemandedBits = DemandedMask & InMask;
783 InDemandedBits |= InSignBit;
785 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero,
786 KnownOne, TLO, Depth+1))
789 // If the sign bit is known zero, convert this to a zero extend.
790 if (KnownZero & InSignBit)
791 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND,
795 // If the sign bit is known one, the top bits match.
796 if (KnownOne & InSignBit) {
798 KnownZero &= ~NewBits;
799 } else { // Otherwise, top bits aren't known.
800 KnownOne &= ~NewBits;
801 KnownZero &= ~NewBits;
805 case ISD::ANY_EXTEND: {
806 uint64_t InMask = MVT::getIntVTBitMask(Op.getOperand(0).getValueType());
807 if (SimplifyDemandedBits(Op.getOperand(0), DemandedMask & InMask,
808 KnownZero, KnownOne, TLO, Depth+1))
810 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
813 case ISD::TRUNCATE: {
814 // Simplify the input, using demanded bit information, and compute the known
815 // zero/one bits live out.
816 if (SimplifyDemandedBits(Op.getOperand(0), DemandedMask,
817 KnownZero, KnownOne, TLO, Depth+1))
820 // If the input is only used by this truncate, see if we can shrink it based
821 // on the known demanded bits.
822 if (Op.getOperand(0).Val->hasOneUse()) {
823 SDOperand In = Op.getOperand(0);
824 switch (In.getOpcode()) {
827 // Shrink SRL by a constant if none of the high bits shifted in are
829 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1))){
830 uint64_t HighBits = MVT::getIntVTBitMask(In.getValueType());
831 HighBits &= ~MVT::getIntVTBitMask(Op.getValueType());
832 HighBits >>= ShAmt->getValue();
834 if (ShAmt->getValue() < MVT::getSizeInBits(Op.getValueType()) &&
835 (DemandedMask & HighBits) == 0) {
836 // None of the shifted in bits are needed. Add a truncate of the
837 // shift input, then shift it.
838 SDOperand NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE,
841 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL,Op.getValueType(),
842 NewTrunc, In.getOperand(1)));
849 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
850 uint64_t OutMask = MVT::getIntVTBitMask(Op.getValueType());
851 KnownZero &= OutMask;
855 case ISD::AssertZext: {
856 MVT::ValueType VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
857 uint64_t InMask = MVT::getIntVTBitMask(VT);
858 if (SimplifyDemandedBits(Op.getOperand(0), DemandedMask & InMask,
859 KnownZero, KnownOne, TLO, Depth+1))
861 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
862 KnownZero |= ~InMask & DemandedMask;
867 case ISD::INTRINSIC_WO_CHAIN:
868 case ISD::INTRINSIC_W_CHAIN:
869 case ISD::INTRINSIC_VOID:
870 // Just use ComputeMaskedBits to compute output bits.
871 TLO.DAG.ComputeMaskedBits(Op, DemandedMask, KnownZero, KnownOne, Depth);
875 // If we know the value of all of the demanded bits, return this as a
877 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask)
878 return TLO.CombineTo(Op, TLO.DAG.getConstant(KnownOne, Op.getValueType()));
883 /// computeMaskedBitsForTargetNode - Determine which of the bits specified
884 /// in Mask are known to be either zero or one and return them in the
885 /// KnownZero/KnownOne bitsets.
886 void TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
890 const SelectionDAG &DAG,
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 MaskedValueIsZero if you don't know whether Op"
897 " is a target node!");
902 /// ComputeNumSignBitsForTargetNode - This method can be implemented by
903 /// targets that want to expose additional information about sign bits to the
905 unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDOperand Op,
906 unsigned Depth) const {
907 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
908 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
909 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
910 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
911 "Should use ComputeNumSignBits if you don't know whether Op"
912 " is a target node!");
917 /// SimplifySetCC - Try to simplify a setcc built with the specified operands
918 /// and cc. If it is unable to simplify it, return a null SDOperand.
920 TargetLowering::SimplifySetCC(MVT::ValueType VT, SDOperand N0, SDOperand N1,
921 ISD::CondCode Cond, bool foldBooleans,
922 DAGCombinerInfo &DCI) const {
923 SelectionDAG &DAG = DCI.DAG;
925 // These setcc operations always fold.
929 case ISD::SETFALSE2: return DAG.getConstant(0, VT);
931 case ISD::SETTRUE2: return DAG.getConstant(1, VT);
934 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val)) {
935 uint64_t C1 = N1C->getValue();
936 if (isa<ConstantSDNode>(N0.Val)) {
937 return DAG.FoldSetCC(VT, N0, N1, Cond);
939 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an
940 // equality comparison, then we're just comparing whether X itself is
942 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) &&
943 N0.getOperand(0).getOpcode() == ISD::CTLZ &&
944 N0.getOperand(1).getOpcode() == ISD::Constant) {
945 unsigned ShAmt = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
946 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
947 ShAmt == Log2_32(MVT::getSizeInBits(N0.getValueType()))) {
948 if ((C1 == 0) == (Cond == ISD::SETEQ)) {
949 // (srl (ctlz x), 5) == 0 -> X != 0
950 // (srl (ctlz x), 5) != 1 -> X != 0
953 // (srl (ctlz x), 5) != 0 -> X == 0
954 // (srl (ctlz x), 5) == 1 -> X == 0
957 SDOperand Zero = DAG.getConstant(0, N0.getValueType());
958 return DAG.getSetCC(VT, N0.getOperand(0).getOperand(0),
963 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
964 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
965 unsigned InSize = MVT::getSizeInBits(N0.getOperand(0).getValueType());
967 // If the comparison constant has bits in the upper part, the
968 // zero-extended value could never match.
969 if (C1 & (~0ULL << InSize)) {
970 unsigned VSize = MVT::getSizeInBits(N0.getValueType());
974 case ISD::SETEQ: return DAG.getConstant(0, VT);
977 case ISD::SETNE: return DAG.getConstant(1, VT);
980 // True if the sign bit of C1 is set.
981 return DAG.getConstant((C1 & (1ULL << (VSize-1))) != 0, VT);
984 // True if the sign bit of C1 isn't set.
985 return DAG.getConstant((C1 & (1ULL << (VSize-1))) == 0, VT);
991 // Otherwise, we can perform the comparison with the low bits.
999 return DAG.getSetCC(VT, N0.getOperand(0),
1000 DAG.getConstant(C1, N0.getOperand(0).getValueType()),
1003 break; // todo, be more careful with signed comparisons
1005 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
1006 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1007 MVT::ValueType ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
1008 unsigned ExtSrcTyBits = MVT::getSizeInBits(ExtSrcTy);
1009 MVT::ValueType ExtDstTy = N0.getValueType();
1010 unsigned ExtDstTyBits = MVT::getSizeInBits(ExtDstTy);
1012 // If the extended part has any inconsistent bits, it cannot ever
1013 // compare equal. In other words, they have to be all ones or all
1016 (~0ULL >> (64-ExtSrcTyBits)) & (~0ULL << (ExtDstTyBits-1));
1017 if ((C1 & ExtBits) != 0 && (C1 & ExtBits) != ExtBits)
1018 return DAG.getConstant(Cond == ISD::SETNE, VT);
1021 MVT::ValueType Op0Ty = N0.getOperand(0).getValueType();
1022 if (Op0Ty == ExtSrcTy) {
1023 ZextOp = N0.getOperand(0);
1025 int64_t Imm = ~0ULL >> (64-ExtSrcTyBits);
1026 ZextOp = DAG.getNode(ISD::AND, Op0Ty, N0.getOperand(0),
1027 DAG.getConstant(Imm, Op0Ty));
1029 if (!DCI.isCalledByLegalizer())
1030 DCI.AddToWorklist(ZextOp.Val);
1031 // Otherwise, make this a use of a zext.
1032 return DAG.getSetCC(VT, ZextOp,
1033 DAG.getConstant(C1 & (~0ULL>>(64-ExtSrcTyBits)),
1036 } else if ((N1C->getValue() == 0 || N1C->getValue() == 1) &&
1037 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1039 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC
1040 if (N0.getOpcode() == ISD::SETCC) {
1041 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getValue() != 1);
1045 // Invert the condition.
1046 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
1047 CC = ISD::getSetCCInverse(CC,
1048 MVT::isInteger(N0.getOperand(0).getValueType()));
1049 return DAG.getSetCC(VT, N0.getOperand(0), N0.getOperand(1), CC);
1052 if ((N0.getOpcode() == ISD::XOR ||
1053 (N0.getOpcode() == ISD::AND &&
1054 N0.getOperand(0).getOpcode() == ISD::XOR &&
1055 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
1056 isa<ConstantSDNode>(N0.getOperand(1)) &&
1057 cast<ConstantSDNode>(N0.getOperand(1))->getValue() == 1) {
1058 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We
1059 // can only do this if the top bits are known zero.
1060 if (DAG.MaskedValueIsZero(N0,
1061 MVT::getIntVTBitMask(N0.getValueType())-1)){
1062 // Okay, get the un-inverted input value.
1064 if (N0.getOpcode() == ISD::XOR)
1065 Val = N0.getOperand(0);
1067 assert(N0.getOpcode() == ISD::AND &&
1068 N0.getOperand(0).getOpcode() == ISD::XOR);
1069 // ((X^1)&1)^1 -> X & 1
1070 Val = DAG.getNode(ISD::AND, N0.getValueType(),
1071 N0.getOperand(0).getOperand(0),
1074 return DAG.getSetCC(VT, Val, N1,
1075 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
1080 uint64_t MinVal, MaxVal;
1081 unsigned OperandBitSize = MVT::getSizeInBits(N1C->getValueType(0));
1082 if (ISD::isSignedIntSetCC(Cond)) {
1083 MinVal = 1ULL << (OperandBitSize-1);
1084 if (OperandBitSize != 1) // Avoid X >> 64, which is undefined.
1085 MaxVal = ~0ULL >> (65-OperandBitSize);
1090 MaxVal = ~0ULL >> (64-OperandBitSize);
1093 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
1094 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
1095 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true
1096 --C1; // X >= C0 --> X > (C0-1)
1097 return DAG.getSetCC(VT, N0, DAG.getConstant(C1, N1.getValueType()),
1098 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
1101 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
1102 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true
1103 ++C1; // X <= C0 --> X < (C0+1)
1104 return DAG.getSetCC(VT, N0, DAG.getConstant(C1, N1.getValueType()),
1105 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
1108 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
1109 return DAG.getConstant(0, VT); // X < MIN --> false
1110 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal)
1111 return DAG.getConstant(1, VT); // X >= MIN --> true
1112 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal)
1113 return DAG.getConstant(0, VT); // X > MAX --> false
1114 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal)
1115 return DAG.getConstant(1, VT); // X <= MAX --> true
1117 // Canonicalize setgt X, Min --> setne X, Min
1118 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
1119 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
1120 // Canonicalize setlt X, Max --> setne X, Max
1121 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
1122 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
1124 // If we have setult X, 1, turn it into seteq X, 0
1125 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
1126 return DAG.getSetCC(VT, N0, DAG.getConstant(MinVal, N0.getValueType()),
1128 // If we have setugt X, Max-1, turn it into seteq X, Max
1129 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
1130 return DAG.getSetCC(VT, N0, DAG.getConstant(MaxVal, N0.getValueType()),
1133 // If we have "setcc X, C0", check to see if we can shrink the immediate
1136 // SETUGT X, SINTMAX -> SETLT X, 0
1137 if (Cond == ISD::SETUGT && OperandBitSize != 1 &&
1138 C1 == (~0ULL >> (65-OperandBitSize)))
1139 return DAG.getSetCC(VT, N0, DAG.getConstant(0, N1.getValueType()),
1142 // FIXME: Implement the rest of these.
1144 // Fold bit comparisons when we can.
1145 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1146 VT == N0.getValueType() && N0.getOpcode() == ISD::AND)
1147 if (ConstantSDNode *AndRHS =
1148 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1149 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
1150 // Perform the xform if the AND RHS is a single bit.
1151 if (isPowerOf2_64(AndRHS->getValue())) {
1152 return DAG.getNode(ISD::SRL, VT, N0,
1153 DAG.getConstant(Log2_64(AndRHS->getValue()),
1154 getShiftAmountTy()));
1156 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getValue()) {
1157 // (X & 8) == 8 --> (X & 8) >> 3
1158 // Perform the xform if C1 is a single bit.
1159 if (isPowerOf2_64(C1)) {
1160 return DAG.getNode(ISD::SRL, VT, N0,
1161 DAG.getConstant(Log2_64(C1), getShiftAmountTy()));
1166 } else if (isa<ConstantSDNode>(N0.Val)) {
1167 // Ensure that the constant occurs on the RHS.
1168 return DAG.getSetCC(VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
1171 if (isa<ConstantFPSDNode>(N0.Val)) {
1172 // Constant fold or commute setcc.
1173 SDOperand O = DAG.FoldSetCC(VT, N0, N1, Cond);
1174 if (O.Val) return O;
1178 // We can always fold X == X for integer setcc's.
1179 if (MVT::isInteger(N0.getValueType()))
1180 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
1181 unsigned UOF = ISD::getUnorderedFlavor(Cond);
1182 if (UOF == 2) // FP operators that are undefined on NaNs.
1183 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
1184 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
1185 return DAG.getConstant(UOF, VT);
1186 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
1187 // if it is not already.
1188 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
1189 if (NewCond != Cond)
1190 return DAG.getSetCC(VT, N0, N1, NewCond);
1193 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1194 MVT::isInteger(N0.getValueType())) {
1195 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
1196 N0.getOpcode() == ISD::XOR) {
1197 // Simplify (X+Y) == (X+Z) --> Y == Z
1198 if (N0.getOpcode() == N1.getOpcode()) {
1199 if (N0.getOperand(0) == N1.getOperand(0))
1200 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(1), Cond);
1201 if (N0.getOperand(1) == N1.getOperand(1))
1202 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(0), Cond);
1203 if (DAG.isCommutativeBinOp(N0.getOpcode())) {
1204 // If X op Y == Y op X, try other combinations.
1205 if (N0.getOperand(0) == N1.getOperand(1))
1206 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(0), Cond);
1207 if (N0.getOperand(1) == N1.getOperand(0))
1208 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(1), Cond);
1212 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) {
1213 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1214 // Turn (X+C1) == C2 --> X == C2-C1
1215 if (N0.getOpcode() == ISD::ADD && N0.Val->hasOneUse()) {
1216 return DAG.getSetCC(VT, N0.getOperand(0),
1217 DAG.getConstant(RHSC->getValue()-LHSR->getValue(),
1218 N0.getValueType()), Cond);
1221 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
1222 if (N0.getOpcode() == ISD::XOR)
1223 // If we know that all of the inverted bits are zero, don't bother
1224 // performing the inversion.
1225 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getValue()))
1226 return DAG.getSetCC(VT, N0.getOperand(0),
1227 DAG.getConstant(LHSR->getValue()^RHSC->getValue(),
1228 N0.getValueType()), Cond);
1231 // Turn (C1-X) == C2 --> X == C1-C2
1232 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
1233 if (N0.getOpcode() == ISD::SUB && N0.Val->hasOneUse()) {
1234 return DAG.getSetCC(VT, N0.getOperand(1),
1235 DAG.getConstant(SUBC->getValue()-RHSC->getValue(),
1236 N0.getValueType()), Cond);
1241 // Simplify (X+Z) == X --> Z == 0
1242 if (N0.getOperand(0) == N1)
1243 return DAG.getSetCC(VT, N0.getOperand(1),
1244 DAG.getConstant(0, N0.getValueType()), Cond);
1245 if (N0.getOperand(1) == N1) {
1246 if (DAG.isCommutativeBinOp(N0.getOpcode()))
1247 return DAG.getSetCC(VT, N0.getOperand(0),
1248 DAG.getConstant(0, N0.getValueType()), Cond);
1249 else if (N0.Val->hasOneUse()) {
1250 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
1251 // (Z-X) == X --> Z == X<<1
1252 SDOperand SH = DAG.getNode(ISD::SHL, N1.getValueType(),
1254 DAG.getConstant(1, getShiftAmountTy()));
1255 if (!DCI.isCalledByLegalizer())
1256 DCI.AddToWorklist(SH.Val);
1257 return DAG.getSetCC(VT, N0.getOperand(0), SH, Cond);
1262 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
1263 N1.getOpcode() == ISD::XOR) {
1264 // Simplify X == (X+Z) --> Z == 0
1265 if (N1.getOperand(0) == N0) {
1266 return DAG.getSetCC(VT, N1.getOperand(1),
1267 DAG.getConstant(0, N1.getValueType()), Cond);
1268 } else if (N1.getOperand(1) == N0) {
1269 if (DAG.isCommutativeBinOp(N1.getOpcode())) {
1270 return DAG.getSetCC(VT, N1.getOperand(0),
1271 DAG.getConstant(0, N1.getValueType()), Cond);
1272 } else if (N1.Val->hasOneUse()) {
1273 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
1274 // X == (Z-X) --> X<<1 == Z
1275 SDOperand SH = DAG.getNode(ISD::SHL, N1.getValueType(), N0,
1276 DAG.getConstant(1, getShiftAmountTy()));
1277 if (!DCI.isCalledByLegalizer())
1278 DCI.AddToWorklist(SH.Val);
1279 return DAG.getSetCC(VT, SH, N1.getOperand(0), Cond);
1285 // Fold away ALL boolean setcc's.
1287 if (N0.getValueType() == MVT::i1 && foldBooleans) {
1289 default: assert(0 && "Unknown integer setcc!");
1290 case ISD::SETEQ: // X == Y -> (X^Y)^1
1291 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
1292 N0 = DAG.getNode(ISD::XOR, MVT::i1, Temp, DAG.getConstant(1, MVT::i1));
1293 if (!DCI.isCalledByLegalizer())
1294 DCI.AddToWorklist(Temp.Val);
1296 case ISD::SETNE: // X != Y --> (X^Y)
1297 N0 = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
1299 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> X^1 & Y
1300 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> X^1 & Y
1301 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
1302 N0 = DAG.getNode(ISD::AND, MVT::i1, N1, Temp);
1303 if (!DCI.isCalledByLegalizer())
1304 DCI.AddToWorklist(Temp.Val);
1306 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> Y^1 & X
1307 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> Y^1 & X
1308 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
1309 N0 = DAG.getNode(ISD::AND, MVT::i1, N0, Temp);
1310 if (!DCI.isCalledByLegalizer())
1311 DCI.AddToWorklist(Temp.Val);
1313 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> X^1 | Y
1314 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> X^1 | Y
1315 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
1316 N0 = DAG.getNode(ISD::OR, MVT::i1, N1, Temp);
1317 if (!DCI.isCalledByLegalizer())
1318 DCI.AddToWorklist(Temp.Val);
1320 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> Y^1 | X
1321 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> Y^1 | X
1322 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
1323 N0 = DAG.getNode(ISD::OR, MVT::i1, N0, Temp);
1326 if (VT != MVT::i1) {
1327 if (!DCI.isCalledByLegalizer())
1328 DCI.AddToWorklist(N0.Val);
1329 // FIXME: If running after legalize, we probably can't do this.
1330 N0 = DAG.getNode(ISD::ZERO_EXTEND, VT, N0);
1335 // Could not fold it.
1339 SDOperand TargetLowering::
1340 PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
1341 // Default implementation: no optimization.
1345 //===----------------------------------------------------------------------===//
1346 // Inline Assembler Implementation Methods
1347 //===----------------------------------------------------------------------===//
1349 TargetLowering::ConstraintType
1350 TargetLowering::getConstraintType(const std::string &Constraint) const {
1351 // FIXME: lots more standard ones to handle.
1352 if (Constraint.size() == 1) {
1353 switch (Constraint[0]) {
1355 case 'r': return C_RegisterClass;
1357 case 'o': // offsetable
1358 case 'V': // not offsetable
1360 case 'i': // Simple Integer or Relocatable Constant
1361 case 'n': // Simple Integer
1362 case 's': // Relocatable Constant
1363 case 'X': // Allow ANY value.
1364 case 'I': // Target registers.
1376 if (Constraint.size() > 1 && Constraint[0] == '{' &&
1377 Constraint[Constraint.size()-1] == '}')
1382 /// isOperandValidForConstraint - Return the specified operand (possibly
1383 /// modified) if the specified SDOperand is valid for the specified target
1384 /// constraint letter, otherwise return null.
1385 SDOperand TargetLowering::isOperandValidForConstraint(SDOperand Op,
1386 char ConstraintLetter,
1387 SelectionDAG &DAG) {
1388 switch (ConstraintLetter) {
1390 case 'i': // Simple Integer or Relocatable Constant
1391 case 'n': // Simple Integer
1392 case 's': // Relocatable Constant
1393 case 'X': { // Allows any operand.
1394 // These operands are interested in values of the form (GV+C), where C may
1395 // be folded in as an offset of GV, or it may be explicitly added. Also, it
1396 // is possible and fine if either GV or C are missing.
1397 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
1398 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
1400 // If we have "(add GV, C)", pull out GV/C
1401 if (Op.getOpcode() == ISD::ADD) {
1402 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
1403 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
1404 if (C == 0 || GA == 0) {
1405 C = dyn_cast<ConstantSDNode>(Op.getOperand(0));
1406 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1));
1408 if (C == 0 || GA == 0)
1412 // If we find a valid operand, map to the TargetXXX version so that the
1413 // value itself doesn't get selected.
1414 if (GA) { // Either &GV or &GV+C
1415 if (ConstraintLetter != 'n') {
1416 int64_t Offs = GA->getOffset();
1417 if (C) Offs += C->getValue();
1418 return DAG.getTargetGlobalAddress(GA->getGlobal(), Op.getValueType(),
1422 if (C) { // just C, no GV.
1423 // Simple constants are not allowed for 's'.
1424 if (ConstraintLetter != 's')
1425 return DAG.getTargetConstant(C->getValue(), Op.getValueType());
1430 return SDOperand(0,0);
1433 std::vector<unsigned> TargetLowering::
1434 getRegClassForInlineAsmConstraint(const std::string &Constraint,
1435 MVT::ValueType VT) const {
1436 return std::vector<unsigned>();
1440 std::pair<unsigned, const TargetRegisterClass*> TargetLowering::
1441 getRegForInlineAsmConstraint(const std::string &Constraint,
1442 MVT::ValueType VT) const {
1443 if (Constraint[0] != '{')
1444 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
1445 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?");
1447 // Remove the braces from around the name.
1448 std::string RegName(Constraint.begin()+1, Constraint.end()-1);
1450 // Figure out which register class contains this reg.
1451 const MRegisterInfo *RI = TM.getRegisterInfo();
1452 for (MRegisterInfo::regclass_iterator RCI = RI->regclass_begin(),
1453 E = RI->regclass_end(); RCI != E; ++RCI) {
1454 const TargetRegisterClass *RC = *RCI;
1456 // If none of the the value types for this register class are valid, we
1457 // can't use it. For example, 64-bit reg classes on 32-bit targets.
1458 bool isLegal = false;
1459 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
1461 if (isTypeLegal(*I)) {
1467 if (!isLegal) continue;
1469 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end();
1471 if (StringsEqualNoCase(RegName, RI->get(*I).Name))
1472 return std::make_pair(*I, RC);
1476 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
1479 //===----------------------------------------------------------------------===//
1480 // Loop Strength Reduction hooks
1481 //===----------------------------------------------------------------------===//
1483 /// isLegalAddressingMode - Return true if the addressing mode represented
1484 /// by AM is legal for this target, for a load/store of the specified type.
1485 bool TargetLowering::isLegalAddressingMode(const AddrMode &AM,
1486 const Type *Ty) const {
1487 // The default implementation of this implements a conservative RISCy, r+r and
1490 // Allows a sign-extended 16-bit immediate field.
1491 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
1494 // No global is ever allowed as a base.
1498 // Only support r+r,
1500 case 0: // "r+i" or just "i", depending on HasBaseReg.
1503 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
1505 // Otherwise we have r+r or r+i.
1508 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
1510 // Allow 2*r as r+r.
1517 // Magic for divide replacement
1520 int64_t m; // magic number
1521 int64_t s; // shift amount
1525 uint64_t m; // magic number
1526 int64_t a; // add indicator
1527 int64_t s; // shift amount
1530 /// magic - calculate the magic numbers required to codegen an integer sdiv as
1531 /// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
1533 static ms magic32(int32_t d) {
1535 uint32_t ad, anc, delta, q1, r1, q2, r2, t;
1536 const uint32_t two31 = 0x80000000U;
1540 t = two31 + ((uint32_t)d >> 31);
1541 anc = t - 1 - t%ad; // absolute value of nc
1542 p = 31; // initialize p
1543 q1 = two31/anc; // initialize q1 = 2p/abs(nc)
1544 r1 = two31 - q1*anc; // initialize r1 = rem(2p,abs(nc))
1545 q2 = two31/ad; // initialize q2 = 2p/abs(d)
1546 r2 = two31 - q2*ad; // initialize r2 = rem(2p,abs(d))
1549 q1 = 2*q1; // update q1 = 2p/abs(nc)
1550 r1 = 2*r1; // update r1 = rem(2p/abs(nc))
1551 if (r1 >= anc) { // must be unsigned comparison
1555 q2 = 2*q2; // update q2 = 2p/abs(d)
1556 r2 = 2*r2; // update r2 = rem(2p/abs(d))
1557 if (r2 >= ad) { // must be unsigned comparison
1562 } while (q1 < delta || (q1 == delta && r1 == 0));
1564 mag.m = (int32_t)(q2 + 1); // make sure to sign extend
1565 if (d < 0) mag.m = -mag.m; // resulting magic number
1566 mag.s = p - 32; // resulting shift
1570 /// magicu - calculate the magic numbers required to codegen an integer udiv as
1571 /// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
1572 static mu magicu32(uint32_t d) {
1574 uint32_t nc, delta, q1, r1, q2, r2;
1576 magu.a = 0; // initialize "add" indicator
1578 p = 31; // initialize p
1579 q1 = 0x80000000/nc; // initialize q1 = 2p/nc
1580 r1 = 0x80000000 - q1*nc; // initialize r1 = rem(2p,nc)
1581 q2 = 0x7FFFFFFF/d; // initialize q2 = (2p-1)/d
1582 r2 = 0x7FFFFFFF - q2*d; // initialize r2 = rem((2p-1),d)
1585 if (r1 >= nc - r1 ) {
1586 q1 = 2*q1 + 1; // update q1
1587 r1 = 2*r1 - nc; // update r1
1590 q1 = 2*q1; // update q1
1591 r1 = 2*r1; // update r1
1593 if (r2 + 1 >= d - r2) {
1594 if (q2 >= 0x7FFFFFFF) magu.a = 1;
1595 q2 = 2*q2 + 1; // update q2
1596 r2 = 2*r2 + 1 - d; // update r2
1599 if (q2 >= 0x80000000) magu.a = 1;
1600 q2 = 2*q2; // update q2
1601 r2 = 2*r2 + 1; // update r2
1604 } while (p < 64 && (q1 < delta || (q1 == delta && r1 == 0)));
1605 magu.m = q2 + 1; // resulting magic number
1606 magu.s = p - 32; // resulting shift
1610 /// magic - calculate the magic numbers required to codegen an integer sdiv as
1611 /// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
1613 static ms magic64(int64_t d) {
1615 uint64_t ad, anc, delta, q1, r1, q2, r2, t;
1616 const uint64_t two63 = 9223372036854775808ULL; // 2^63
1619 ad = d >= 0 ? d : -d;
1620 t = two63 + ((uint64_t)d >> 63);
1621 anc = t - 1 - t%ad; // absolute value of nc
1622 p = 63; // initialize p
1623 q1 = two63/anc; // initialize q1 = 2p/abs(nc)
1624 r1 = two63 - q1*anc; // initialize r1 = rem(2p,abs(nc))
1625 q2 = two63/ad; // initialize q2 = 2p/abs(d)
1626 r2 = two63 - q2*ad; // initialize r2 = rem(2p,abs(d))
1629 q1 = 2*q1; // update q1 = 2p/abs(nc)
1630 r1 = 2*r1; // update r1 = rem(2p/abs(nc))
1631 if (r1 >= anc) { // must be unsigned comparison
1635 q2 = 2*q2; // update q2 = 2p/abs(d)
1636 r2 = 2*r2; // update r2 = rem(2p/abs(d))
1637 if (r2 >= ad) { // must be unsigned comparison
1642 } while (q1 < delta || (q1 == delta && r1 == 0));
1645 if (d < 0) mag.m = -mag.m; // resulting magic number
1646 mag.s = p - 64; // resulting shift
1650 /// magicu - calculate the magic numbers required to codegen an integer udiv as
1651 /// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
1652 static mu magicu64(uint64_t d)
1655 uint64_t nc, delta, q1, r1, q2, r2;
1657 magu.a = 0; // initialize "add" indicator
1659 p = 63; // initialize p
1660 q1 = 0x8000000000000000ull/nc; // initialize q1 = 2p/nc
1661 r1 = 0x8000000000000000ull - q1*nc; // initialize r1 = rem(2p,nc)
1662 q2 = 0x7FFFFFFFFFFFFFFFull/d; // initialize q2 = (2p-1)/d
1663 r2 = 0x7FFFFFFFFFFFFFFFull - q2*d; // initialize r2 = rem((2p-1),d)
1666 if (r1 >= nc - r1 ) {
1667 q1 = 2*q1 + 1; // update q1
1668 r1 = 2*r1 - nc; // update r1
1671 q1 = 2*q1; // update q1
1672 r1 = 2*r1; // update r1
1674 if (r2 + 1 >= d - r2) {
1675 if (q2 >= 0x7FFFFFFFFFFFFFFFull) magu.a = 1;
1676 q2 = 2*q2 + 1; // update q2
1677 r2 = 2*r2 + 1 - d; // update r2
1680 if (q2 >= 0x8000000000000000ull) magu.a = 1;
1681 q2 = 2*q2; // update q2
1682 r2 = 2*r2 + 1; // update r2
1685 } while (p < 128 && (q1 < delta || (q1 == delta && r1 == 0)));
1686 magu.m = q2 + 1; // resulting magic number
1687 magu.s = p - 64; // resulting shift
1691 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
1692 /// return a DAG expression to select that will generate the same value by
1693 /// multiplying by a magic number. See:
1694 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
1695 SDOperand TargetLowering::BuildSDIV(SDNode *N, SelectionDAG &DAG,
1696 std::vector<SDNode*>* Created) const {
1697 MVT::ValueType VT = N->getValueType(0);
1699 // Check to see if we can do this.
1700 if (!isTypeLegal(VT) || (VT != MVT::i32 && VT != MVT::i64))
1701 return SDOperand(); // BuildSDIV only operates on i32 or i64
1702 if (!isOperationLegal(ISD::MULHS, VT))
1703 return SDOperand(); // Make sure the target supports MULHS.
1705 int64_t d = cast<ConstantSDNode>(N->getOperand(1))->getSignExtended();
1706 ms magics = (VT == MVT::i32) ? magic32(d) : magic64(d);
1708 // Multiply the numerator (operand 0) by the magic value
1709 SDOperand Q = DAG.getNode(ISD::MULHS, VT, N->getOperand(0),
1710 DAG.getConstant(magics.m, VT));
1711 // If d > 0 and m < 0, add the numerator
1712 if (d > 0 && magics.m < 0) {
1713 Q = DAG.getNode(ISD::ADD, VT, Q, N->getOperand(0));
1715 Created->push_back(Q.Val);
1717 // If d < 0 and m > 0, subtract the numerator.
1718 if (d < 0 && magics.m > 0) {
1719 Q = DAG.getNode(ISD::SUB, VT, Q, N->getOperand(0));
1721 Created->push_back(Q.Val);
1723 // Shift right algebraic if shift value is nonzero
1725 Q = DAG.getNode(ISD::SRA, VT, Q,
1726 DAG.getConstant(magics.s, getShiftAmountTy()));
1728 Created->push_back(Q.Val);
1730 // Extract the sign bit and add it to the quotient
1732 DAG.getNode(ISD::SRL, VT, Q, DAG.getConstant(MVT::getSizeInBits(VT)-1,
1733 getShiftAmountTy()));
1735 Created->push_back(T.Val);
1736 return DAG.getNode(ISD::ADD, VT, Q, T);
1739 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
1740 /// return a DAG expression to select that will generate the same value by
1741 /// multiplying by a magic number. See:
1742 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
1743 SDOperand TargetLowering::BuildUDIV(SDNode *N, SelectionDAG &DAG,
1744 std::vector<SDNode*>* Created) const {
1745 MVT::ValueType VT = N->getValueType(0);
1747 // Check to see if we can do this.
1748 if (!isTypeLegal(VT) || (VT != MVT::i32 && VT != MVT::i64))
1749 return SDOperand(); // BuildUDIV only operates on i32 or i64
1750 if (!isOperationLegal(ISD::MULHU, VT))
1751 return SDOperand(); // Make sure the target supports MULHU.
1753 uint64_t d = cast<ConstantSDNode>(N->getOperand(1))->getValue();
1754 mu magics = (VT == MVT::i32) ? magicu32(d) : magicu64(d);
1756 // Multiply the numerator (operand 0) by the magic value
1757 SDOperand Q = DAG.getNode(ISD::MULHU, VT, N->getOperand(0),
1758 DAG.getConstant(magics.m, VT));
1760 Created->push_back(Q.Val);
1762 if (magics.a == 0) {
1763 return DAG.getNode(ISD::SRL, VT, Q,
1764 DAG.getConstant(magics.s, getShiftAmountTy()));
1766 SDOperand NPQ = DAG.getNode(ISD::SUB, VT, N->getOperand(0), Q);
1768 Created->push_back(NPQ.Val);
1769 NPQ = DAG.getNode(ISD::SRL, VT, NPQ,
1770 DAG.getConstant(1, getShiftAmountTy()));
1772 Created->push_back(NPQ.Val);
1773 NPQ = DAG.getNode(ISD::ADD, VT, NPQ, Q);
1775 Created->push_back(NPQ.Val);
1776 return DAG.getNode(ISD::SRL, VT, NPQ,
1777 DAG.getConstant(magics.s-1, getShiftAmountTy()));