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/TargetSubtarget.h"
16 #include "llvm/Target/TargetData.h"
17 #include "llvm/Target/TargetMachine.h"
18 #include "llvm/Target/MRegisterInfo.h"
19 #include "llvm/DerivedTypes.h"
20 #include "llvm/CodeGen/SelectionDAG.h"
21 #include "llvm/ADT/StringExtras.h"
22 #include "llvm/ADT/STLExtras.h"
23 #include "llvm/Support/MathExtras.h"
24 #include "llvm/Target/TargetAsmInfo.h"
25 #include "llvm/CallingConv.h"
28 /// InitLibcallNames - Set default libcall names.
30 static void InitLibcallNames(const char **Names) {
31 Names[RTLIB::SHL_I32] = "__ashlsi3";
32 Names[RTLIB::SHL_I64] = "__ashldi3";
33 Names[RTLIB::SRL_I32] = "__lshrsi3";
34 Names[RTLIB::SRL_I64] = "__lshrdi3";
35 Names[RTLIB::SRA_I32] = "__ashrsi3";
36 Names[RTLIB::SRA_I64] = "__ashrdi3";
37 Names[RTLIB::MUL_I32] = "__mulsi3";
38 Names[RTLIB::MUL_I64] = "__muldi3";
39 Names[RTLIB::SDIV_I32] = "__divsi3";
40 Names[RTLIB::SDIV_I64] = "__divdi3";
41 Names[RTLIB::UDIV_I32] = "__udivsi3";
42 Names[RTLIB::UDIV_I64] = "__udivdi3";
43 Names[RTLIB::SREM_I32] = "__modsi3";
44 Names[RTLIB::SREM_I64] = "__moddi3";
45 Names[RTLIB::UREM_I32] = "__umodsi3";
46 Names[RTLIB::UREM_I64] = "__umoddi3";
47 Names[RTLIB::NEG_I32] = "__negsi2";
48 Names[RTLIB::NEG_I64] = "__negdi2";
49 Names[RTLIB::ADD_F32] = "__addsf3";
50 Names[RTLIB::ADD_F64] = "__adddf3";
51 Names[RTLIB::ADD_PPCF128] = "__gcc_qadd";
52 Names[RTLIB::SUB_F32] = "__subsf3";
53 Names[RTLIB::SUB_F64] = "__subdf3";
54 Names[RTLIB::SUB_PPCF128] = "__gcc_qsub";
55 Names[RTLIB::MUL_F32] = "__mulsf3";
56 Names[RTLIB::MUL_F64] = "__muldf3";
57 Names[RTLIB::MUL_PPCF128] = "__gcc_qmul";
58 Names[RTLIB::DIV_F32] = "__divsf3";
59 Names[RTLIB::DIV_F64] = "__divdf3";
60 Names[RTLIB::DIV_PPCF128] = "__gcc_qdiv";
61 Names[RTLIB::REM_F32] = "fmodf";
62 Names[RTLIB::REM_F64] = "fmod";
63 Names[RTLIB::REM_PPCF128] = "fmodl";
64 Names[RTLIB::NEG_F32] = "__negsf2";
65 Names[RTLIB::NEG_F64] = "__negdf2";
66 Names[RTLIB::POWI_F32] = "__powisf2";
67 Names[RTLIB::POWI_F64] = "__powidf2";
68 Names[RTLIB::POWI_F80] = "__powixf2";
69 Names[RTLIB::POWI_PPCF128] = "__powitf2";
70 Names[RTLIB::SQRT_F32] = "sqrtf";
71 Names[RTLIB::SQRT_F64] = "sqrt";
72 Names[RTLIB::SQRT_F80] = "sqrtl";
73 Names[RTLIB::SQRT_PPCF128] = "sqrtl";
74 Names[RTLIB::SIN_F32] = "sinf";
75 Names[RTLIB::SIN_F64] = "sin";
76 Names[RTLIB::COS_F32] = "cosf";
77 Names[RTLIB::COS_F64] = "cos";
78 Names[RTLIB::POW_F32] = "powf";
79 Names[RTLIB::POW_F64] = "pow";
80 Names[RTLIB::POW_F80] = "powl";
81 Names[RTLIB::POW_PPCF128] = "powl";
82 Names[RTLIB::FPEXT_F32_F64] = "__extendsfdf2";
83 Names[RTLIB::FPROUND_F64_F32] = "__truncdfsf2";
84 Names[RTLIB::FPTOSINT_F32_I32] = "__fixsfsi";
85 Names[RTLIB::FPTOSINT_F32_I64] = "__fixsfdi";
86 Names[RTLIB::FPTOSINT_F64_I32] = "__fixdfsi";
87 Names[RTLIB::FPTOSINT_F64_I64] = "__fixdfdi";
88 Names[RTLIB::FPTOSINT_F80_I64] = "__fixxfdi";
89 Names[RTLIB::FPTOSINT_PPCF128_I64] = "__fixtfdi";
90 Names[RTLIB::FPTOUINT_F32_I32] = "__fixunssfsi";
91 Names[RTLIB::FPTOUINT_F32_I64] = "__fixunssfdi";
92 Names[RTLIB::FPTOUINT_F64_I32] = "__fixunsdfsi";
93 Names[RTLIB::FPTOUINT_F64_I64] = "__fixunsdfdi";
94 Names[RTLIB::FPTOUINT_F80_I32] = "__fixunsxfsi";
95 Names[RTLIB::FPTOUINT_F80_I64] = "__fixunsxfdi";
96 Names[RTLIB::FPTOUINT_PPCF128_I64] = "__fixunstfdi";
97 Names[RTLIB::SINTTOFP_I32_F32] = "__floatsisf";
98 Names[RTLIB::SINTTOFP_I32_F64] = "__floatsidf";
99 Names[RTLIB::SINTTOFP_I64_F32] = "__floatdisf";
100 Names[RTLIB::SINTTOFP_I64_F64] = "__floatdidf";
101 Names[RTLIB::SINTTOFP_I64_F80] = "__floatdixf";
102 Names[RTLIB::SINTTOFP_I64_PPCF128] = "__floatditf";
103 Names[RTLIB::UINTTOFP_I32_F32] = "__floatunsisf";
104 Names[RTLIB::UINTTOFP_I32_F64] = "__floatunsidf";
105 Names[RTLIB::UINTTOFP_I64_F32] = "__floatundisf";
106 Names[RTLIB::UINTTOFP_I64_F64] = "__floatundidf";
107 Names[RTLIB::OEQ_F32] = "__eqsf2";
108 Names[RTLIB::OEQ_F64] = "__eqdf2";
109 Names[RTLIB::UNE_F32] = "__nesf2";
110 Names[RTLIB::UNE_F64] = "__nedf2";
111 Names[RTLIB::OGE_F32] = "__gesf2";
112 Names[RTLIB::OGE_F64] = "__gedf2";
113 Names[RTLIB::OLT_F32] = "__ltsf2";
114 Names[RTLIB::OLT_F64] = "__ltdf2";
115 Names[RTLIB::OLE_F32] = "__lesf2";
116 Names[RTLIB::OLE_F64] = "__ledf2";
117 Names[RTLIB::OGT_F32] = "__gtsf2";
118 Names[RTLIB::OGT_F64] = "__gtdf2";
119 Names[RTLIB::UO_F32] = "__unordsf2";
120 Names[RTLIB::UO_F64] = "__unorddf2";
121 Names[RTLIB::O_F32] = "__unordsf2";
122 Names[RTLIB::O_F64] = "__unorddf2";
125 /// InitCmpLibcallCCs - Set default comparison libcall CC.
127 static void InitCmpLibcallCCs(ISD::CondCode *CCs) {
128 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL);
129 CCs[RTLIB::OEQ_F32] = ISD::SETEQ;
130 CCs[RTLIB::OEQ_F64] = ISD::SETEQ;
131 CCs[RTLIB::UNE_F32] = ISD::SETNE;
132 CCs[RTLIB::UNE_F64] = ISD::SETNE;
133 CCs[RTLIB::OGE_F32] = ISD::SETGE;
134 CCs[RTLIB::OGE_F64] = ISD::SETGE;
135 CCs[RTLIB::OLT_F32] = ISD::SETLT;
136 CCs[RTLIB::OLT_F64] = ISD::SETLT;
137 CCs[RTLIB::OLE_F32] = ISD::SETLE;
138 CCs[RTLIB::OLE_F64] = ISD::SETLE;
139 CCs[RTLIB::OGT_F32] = ISD::SETGT;
140 CCs[RTLIB::OGT_F64] = ISD::SETGT;
141 CCs[RTLIB::UO_F32] = ISD::SETNE;
142 CCs[RTLIB::UO_F64] = ISD::SETNE;
143 CCs[RTLIB::O_F32] = ISD::SETEQ;
144 CCs[RTLIB::O_F64] = ISD::SETEQ;
147 TargetLowering::TargetLowering(TargetMachine &tm)
148 : TM(tm), TD(TM.getTargetData()) {
149 assert(ISD::BUILTIN_OP_END <= 156 &&
150 "Fixed size array in TargetLowering is not large enough!");
151 // All operations default to being supported.
152 memset(OpActions, 0, sizeof(OpActions));
153 memset(LoadXActions, 0, sizeof(LoadXActions));
154 memset(&StoreXActions, 0, sizeof(StoreXActions));
155 memset(&IndexedModeActions, 0, sizeof(IndexedModeActions));
156 memset(&ConvertActions, 0, sizeof(ConvertActions));
158 // Set all indexed load / store to expand.
159 for (unsigned VT = 0; VT != (unsigned)MVT::LAST_VALUETYPE; ++VT) {
160 for (unsigned IM = (unsigned)ISD::PRE_INC;
161 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
162 setIndexedLoadAction(IM, (MVT::ValueType)VT, Expand);
163 setIndexedStoreAction(IM, (MVT::ValueType)VT, Expand);
167 IsLittleEndian = TD->isLittleEndian();
168 UsesGlobalOffsetTable = false;
169 ShiftAmountTy = SetCCResultTy = PointerTy = getValueType(TD->getIntPtrType());
170 ShiftAmtHandling = Undefined;
171 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE*sizeof(TargetRegisterClass*));
172 memset(TargetDAGCombineArray, 0, array_lengthof(TargetDAGCombineArray));
173 maxStoresPerMemset = maxStoresPerMemcpy = maxStoresPerMemmove = 8;
174 allowUnalignedMemoryAccesses = false;
175 UseUnderscoreSetJmp = false;
176 UseUnderscoreLongJmp = false;
177 SelectIsExpensive = false;
178 IntDivIsCheap = false;
179 Pow2DivIsCheap = false;
180 StackPointerRegisterToSaveRestore = 0;
181 ExceptionPointerRegister = 0;
182 ExceptionSelectorRegister = 0;
183 SetCCResultContents = UndefinedSetCCResult;
184 SchedPreferenceInfo = SchedulingForLatency;
186 JumpBufAlignment = 0;
187 IfCvtBlockSizeLimit = 2;
189 InitLibcallNames(LibcallRoutineNames);
190 InitCmpLibcallCCs(CmpLibcallCCs);
192 // Tell Legalize whether the assembler supports DEBUG_LOC.
193 if (!TM.getTargetAsmInfo()->hasDotLocAndDotFile())
194 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
197 TargetLowering::~TargetLowering() {}
200 SDOperand TargetLowering::LowerMEMCPY(SDOperand Op, SelectionDAG &DAG) {
201 assert(getSubtarget() && "Subtarget not defined");
202 SDOperand ChainOp = Op.getOperand(0);
203 SDOperand DestOp = Op.getOperand(1);
204 SDOperand SourceOp = Op.getOperand(2);
205 SDOperand CountOp = Op.getOperand(3);
206 SDOperand AlignOp = Op.getOperand(4);
207 SDOperand AlwaysInlineOp = Op.getOperand(5);
209 bool AlwaysInline = (bool)cast<ConstantSDNode>(AlwaysInlineOp)->getValue();
210 unsigned Align = (unsigned)cast<ConstantSDNode>(AlignOp)->getValue();
211 if (Align == 0) Align = 1;
213 // If size is unknown, call memcpy.
214 ConstantSDNode *I = dyn_cast<ConstantSDNode>(CountOp);
216 assert(!AlwaysInline && "Cannot inline copy of unknown size");
217 return LowerMEMCPYCall(ChainOp, DestOp, SourceOp, CountOp, DAG);
220 // If not DWORD aligned or if size is more than threshold, then call memcpy.
221 // The libc version is likely to be faster for the following cases. It can
222 // use the address value and run time information about the CPU.
223 // With glibc 2.6.1 on a core 2, coping an array of 100M longs was 30% faster
224 unsigned Size = I->getValue();
226 (Size <= getSubtarget()->getMaxInlineSizeThreshold() &&
228 return LowerMEMCPYInline(ChainOp, DestOp, SourceOp, Size, Align, DAG);
229 return LowerMEMCPYCall(ChainOp, DestOp, SourceOp, CountOp, DAG);
233 SDOperand TargetLowering::LowerMEMCPYCall(SDOperand Chain,
238 MVT::ValueType IntPtr = getPointerTy();
239 TargetLowering::ArgListTy Args;
240 TargetLowering::ArgListEntry Entry;
241 Entry.Ty = getTargetData()->getIntPtrType();
242 Entry.Node = Dest; Args.push_back(Entry);
243 Entry.Node = Source; Args.push_back(Entry);
244 Entry.Node = Count; Args.push_back(Entry);
245 std::pair<SDOperand,SDOperand> CallResult =
246 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
247 DAG.getExternalSymbol("memcpy", IntPtr), Args, DAG);
248 return CallResult.second;
252 /// computeRegisterProperties - Once all of the register classes are added,
253 /// this allows us to compute derived properties we expose.
254 void TargetLowering::computeRegisterProperties() {
255 assert(MVT::LAST_VALUETYPE <= 32 &&
256 "Too many value types for ValueTypeActions to hold!");
258 // Everything defaults to needing one register.
259 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
260 NumRegistersForVT[i] = 1;
261 RegisterTypeForVT[i] = TransformToType[i] = i;
263 // ...except isVoid, which doesn't need any registers.
264 NumRegistersForVT[MVT::isVoid] = 0;
266 // Find the largest integer register class.
267 unsigned LargestIntReg = MVT::i128;
268 for (; RegClassForVT[LargestIntReg] == 0; --LargestIntReg)
269 assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
271 // Every integer value type larger than this largest register takes twice as
272 // many registers to represent as the previous ValueType.
273 for (MVT::ValueType ExpandedReg = LargestIntReg + 1;
274 MVT::isInteger(ExpandedReg); ++ExpandedReg) {
275 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
276 RegisterTypeForVT[ExpandedReg] = LargestIntReg;
277 TransformToType[ExpandedReg] = ExpandedReg - 1;
278 ValueTypeActions.setTypeAction(ExpandedReg, Expand);
281 // Inspect all of the ValueType's smaller than the largest integer
282 // register to see which ones need promotion.
283 MVT::ValueType LegalIntReg = LargestIntReg;
284 for (MVT::ValueType IntReg = LargestIntReg - 1;
285 IntReg >= MVT::i1; --IntReg) {
286 if (isTypeLegal(IntReg)) {
287 LegalIntReg = IntReg;
289 RegisterTypeForVT[IntReg] = TransformToType[IntReg] = LegalIntReg;
290 ValueTypeActions.setTypeAction(IntReg, Promote);
294 // ppcf128 type is really two f64's.
295 if (!isTypeLegal(MVT::ppcf128)) {
296 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64];
297 RegisterTypeForVT[MVT::ppcf128] = MVT::f64;
298 TransformToType[MVT::ppcf128] = MVT::f64;
299 ValueTypeActions.setTypeAction(MVT::ppcf128, Expand);
302 // Decide how to handle f64. If the target does not have native f64 support,
303 // expand it to i64 and we will be generating soft float library calls.
304 if (!isTypeLegal(MVT::f64)) {
305 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
306 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
307 TransformToType[MVT::f64] = MVT::i64;
308 ValueTypeActions.setTypeAction(MVT::f64, Expand);
311 // Decide how to handle f32. If the target does not have native support for
312 // f32, promote it to f64 if it is legal. Otherwise, expand it to i32.
313 if (!isTypeLegal(MVT::f32)) {
314 if (isTypeLegal(MVT::f64)) {
315 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::f64];
316 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::f64];
317 TransformToType[MVT::f32] = MVT::f64;
318 ValueTypeActions.setTypeAction(MVT::f32, Promote);
320 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
321 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
322 TransformToType[MVT::f32] = MVT::i32;
323 ValueTypeActions.setTypeAction(MVT::f32, Expand);
327 // Loop over all of the vector value types to see which need transformations.
328 for (MVT::ValueType i = MVT::FIRST_VECTOR_VALUETYPE;
329 i <= MVT::LAST_VECTOR_VALUETYPE; ++i) {
330 if (!isTypeLegal(i)) {
331 MVT::ValueType IntermediateVT, RegisterVT;
332 unsigned NumIntermediates;
333 NumRegistersForVT[i] =
334 getVectorTypeBreakdown(i,
335 IntermediateVT, NumIntermediates,
337 RegisterTypeForVT[i] = RegisterVT;
338 TransformToType[i] = MVT::Other; // this isn't actually used
339 ValueTypeActions.setTypeAction(i, Expand);
344 const char *TargetLowering::getTargetNodeName(unsigned Opcode) const {
348 /// getVectorTypeBreakdown - Vector types are broken down into some number of
349 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
350 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
351 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
353 /// This method returns the number of registers needed, and the VT for each
354 /// register. It also returns the VT and quantity of the intermediate values
355 /// before they are promoted/expanded.
357 unsigned TargetLowering::getVectorTypeBreakdown(MVT::ValueType VT,
358 MVT::ValueType &IntermediateVT,
359 unsigned &NumIntermediates,
360 MVT::ValueType &RegisterVT) const {
361 // Figure out the right, legal destination reg to copy into.
362 unsigned NumElts = MVT::getVectorNumElements(VT);
363 MVT::ValueType EltTy = MVT::getVectorElementType(VT);
365 unsigned NumVectorRegs = 1;
367 // Divide the input until we get to a supported size. This will always
368 // end with a scalar if the target doesn't support vectors.
369 while (NumElts > 1 &&
370 !isTypeLegal(MVT::getVectorType(EltTy, NumElts))) {
375 NumIntermediates = NumVectorRegs;
377 MVT::ValueType NewVT = MVT::getVectorType(EltTy, NumElts);
378 if (!isTypeLegal(NewVT))
380 IntermediateVT = NewVT;
382 MVT::ValueType DestVT = getTypeToTransformTo(NewVT);
384 if (DestVT < NewVT) {
385 // Value is expanded, e.g. i64 -> i16.
386 return NumVectorRegs*(MVT::getSizeInBits(NewVT)/MVT::getSizeInBits(DestVT));
388 // Otherwise, promotion or legal types use the same number of registers as
389 // the vector decimated to the appropriate level.
390 return NumVectorRegs;
396 //===----------------------------------------------------------------------===//
397 // Optimization Methods
398 //===----------------------------------------------------------------------===//
400 /// ShrinkDemandedConstant - Check to see if the specified operand of the
401 /// specified instruction is a constant integer. If so, check to see if there
402 /// are any bits set in the constant that are not demanded. If so, shrink the
403 /// constant and return true.
404 bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDOperand Op,
406 // FIXME: ISD::SELECT, ISD::SELECT_CC
407 switch(Op.getOpcode()) {
412 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1)))
413 if ((~Demanded & C->getValue()) != 0) {
414 MVT::ValueType VT = Op.getValueType();
415 SDOperand New = DAG.getNode(Op.getOpcode(), VT, Op.getOperand(0),
416 DAG.getConstant(Demanded & C->getValue(),
418 return CombineTo(Op, New);
425 /// SimplifyDemandedBits - Look at Op. At this point, we know that only the
426 /// DemandedMask bits of the result of Op are ever used downstream. If we can
427 /// use this information to simplify Op, create a new simplified DAG node and
428 /// return true, returning the original and new nodes in Old and New. Otherwise,
429 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
430 /// the expression (used to simplify the caller). The KnownZero/One bits may
431 /// only be accurate for those bits in the DemandedMask.
432 bool TargetLowering::SimplifyDemandedBits(SDOperand Op, uint64_t DemandedMask,
435 TargetLoweringOpt &TLO,
436 unsigned Depth) const {
437 KnownZero = KnownOne = 0; // Don't know anything.
439 // The masks are not wide enough to represent this type! Should use APInt.
440 if (Op.getValueType() == MVT::i128)
443 // Other users may use these bits.
444 if (!Op.Val->hasOneUse()) {
446 // If not at the root, Just compute the KnownZero/KnownOne bits to
447 // simplify things downstream.
448 TLO.DAG.ComputeMaskedBits(Op, DemandedMask, KnownZero, KnownOne, Depth);
451 // If this is the root being simplified, allow it to have multiple uses,
452 // just set the DemandedMask to all bits.
453 DemandedMask = MVT::getIntVTBitMask(Op.getValueType());
454 } else if (DemandedMask == 0) {
455 // Not demanding any bits from Op.
456 if (Op.getOpcode() != ISD::UNDEF)
457 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::UNDEF, Op.getValueType()));
459 } else if (Depth == 6) { // Limit search depth.
463 uint64_t KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut;
464 switch (Op.getOpcode()) {
466 // We know all of the bits for a constant!
467 KnownOne = cast<ConstantSDNode>(Op)->getValue() & DemandedMask;
468 KnownZero = ~KnownOne & DemandedMask;
469 return false; // Don't fall through, will infinitely loop.
471 // If the RHS is a constant, check to see if the LHS would be zero without
472 // using the bits from the RHS. Below, we use knowledge about the RHS to
473 // simplify the LHS, here we're using information from the LHS to simplify
475 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
476 uint64_t LHSZero, LHSOne;
477 TLO.DAG.ComputeMaskedBits(Op.getOperand(0), DemandedMask,
478 LHSZero, LHSOne, Depth+1);
479 // If the LHS already has zeros where RHSC does, this and is dead.
480 if ((LHSZero & DemandedMask) == (~RHSC->getValue() & DemandedMask))
481 return TLO.CombineTo(Op, Op.getOperand(0));
482 // If any of the set bits in the RHS are known zero on the LHS, shrink
484 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & DemandedMask))
488 if (SimplifyDemandedBits(Op.getOperand(1), DemandedMask, KnownZero,
489 KnownOne, TLO, Depth+1))
491 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
492 if (SimplifyDemandedBits(Op.getOperand(0), DemandedMask & ~KnownZero,
493 KnownZero2, KnownOne2, TLO, Depth+1))
495 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
497 // If all of the demanded bits are known one on one side, return the other.
498 // These bits cannot contribute to the result of the 'and'.
499 if ((DemandedMask & ~KnownZero2 & KnownOne)==(DemandedMask & ~KnownZero2))
500 return TLO.CombineTo(Op, Op.getOperand(0));
501 if ((DemandedMask & ~KnownZero & KnownOne2)==(DemandedMask & ~KnownZero))
502 return TLO.CombineTo(Op, Op.getOperand(1));
503 // If all of the demanded bits in the inputs are known zeros, return zero.
504 if ((DemandedMask & (KnownZero|KnownZero2)) == DemandedMask)
505 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, Op.getValueType()));
506 // If the RHS is a constant, see if we can simplify it.
507 if (TLO.ShrinkDemandedConstant(Op, DemandedMask & ~KnownZero2))
510 // Output known-1 bits are only known if set in both the LHS & RHS.
511 KnownOne &= KnownOne2;
512 // Output known-0 are known to be clear if zero in either the LHS | RHS.
513 KnownZero |= KnownZero2;
516 if (SimplifyDemandedBits(Op.getOperand(1), DemandedMask, KnownZero,
517 KnownOne, TLO, Depth+1))
519 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
520 if (SimplifyDemandedBits(Op.getOperand(0), DemandedMask & ~KnownOne,
521 KnownZero2, KnownOne2, TLO, Depth+1))
523 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
525 // If all of the demanded bits are known zero on one side, return the other.
526 // These bits cannot contribute to the result of the 'or'.
527 if ((DemandedMask & ~KnownOne2 & KnownZero) == (DemandedMask & ~KnownOne2))
528 return TLO.CombineTo(Op, Op.getOperand(0));
529 if ((DemandedMask & ~KnownOne & KnownZero2) == (DemandedMask & ~KnownOne))
530 return TLO.CombineTo(Op, Op.getOperand(1));
531 // If all of the potentially set bits on one side are known to be set on
532 // the other side, just use the 'other' side.
533 if ((DemandedMask & (~KnownZero) & KnownOne2) ==
534 (DemandedMask & (~KnownZero)))
535 return TLO.CombineTo(Op, Op.getOperand(0));
536 if ((DemandedMask & (~KnownZero2) & KnownOne) ==
537 (DemandedMask & (~KnownZero2)))
538 return TLO.CombineTo(Op, Op.getOperand(1));
539 // If the RHS is a constant, see if we can simplify it.
540 if (TLO.ShrinkDemandedConstant(Op, DemandedMask))
543 // Output known-0 bits are only known if clear in both the LHS & RHS.
544 KnownZero &= KnownZero2;
545 // Output known-1 are known to be set if set in either the LHS | RHS.
546 KnownOne |= KnownOne2;
549 if (SimplifyDemandedBits(Op.getOperand(1), DemandedMask, KnownZero,
550 KnownOne, TLO, Depth+1))
552 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
553 if (SimplifyDemandedBits(Op.getOperand(0), DemandedMask, KnownZero2,
554 KnownOne2, TLO, Depth+1))
556 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
558 // If all of the demanded bits are known zero on one side, return the other.
559 // These bits cannot contribute to the result of the 'xor'.
560 if ((DemandedMask & KnownZero) == DemandedMask)
561 return TLO.CombineTo(Op, Op.getOperand(0));
562 if ((DemandedMask & KnownZero2) == DemandedMask)
563 return TLO.CombineTo(Op, Op.getOperand(1));
565 // If all of the unknown bits are known to be zero on one side or the other
566 // (but not both) turn this into an *inclusive* or.
567 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
568 if ((DemandedMask & ~KnownZero & ~KnownZero2) == 0)
569 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, Op.getValueType(),
573 // Output known-0 bits are known if clear or set in both the LHS & RHS.
574 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
575 // Output known-1 are known to be set if set in only one of the LHS, RHS.
576 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
578 // If all of the demanded bits on one side are known, and all of the set
579 // bits on that side are also known to be set on the other side, turn this
580 // into an AND, as we know the bits will be cleared.
581 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
582 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask) { // all known
583 if ((KnownOne & KnownOne2) == KnownOne) {
584 MVT::ValueType VT = Op.getValueType();
585 SDOperand ANDC = TLO.DAG.getConstant(~KnownOne & DemandedMask, VT);
586 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, VT, Op.getOperand(0),
591 // If the RHS is a constant, see if we can simplify it.
592 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
593 if (TLO.ShrinkDemandedConstant(Op, DemandedMask))
596 KnownZero = KnownZeroOut;
597 KnownOne = KnownOneOut;
600 // If we know the result of a setcc has the top bits zero, use this info.
601 if (getSetCCResultContents() == TargetLowering::ZeroOrOneSetCCResult)
602 KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL);
605 if (SimplifyDemandedBits(Op.getOperand(2), DemandedMask, KnownZero,
606 KnownOne, TLO, Depth+1))
608 if (SimplifyDemandedBits(Op.getOperand(1), DemandedMask, KnownZero2,
609 KnownOne2, TLO, Depth+1))
611 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
612 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
614 // If the operands are constants, see if we can simplify them.
615 if (TLO.ShrinkDemandedConstant(Op, DemandedMask))
618 // Only known if known in both the LHS and RHS.
619 KnownOne &= KnownOne2;
620 KnownZero &= KnownZero2;
623 if (SimplifyDemandedBits(Op.getOperand(3), DemandedMask, KnownZero,
624 KnownOne, TLO, Depth+1))
626 if (SimplifyDemandedBits(Op.getOperand(2), DemandedMask, KnownZero2,
627 KnownOne2, TLO, Depth+1))
629 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
630 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
632 // If the operands are constants, see if we can simplify them.
633 if (TLO.ShrinkDemandedConstant(Op, DemandedMask))
636 // Only known if known in both the LHS and RHS.
637 KnownOne &= KnownOne2;
638 KnownZero &= KnownZero2;
641 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
642 unsigned ShAmt = SA->getValue();
643 SDOperand InOp = Op.getOperand(0);
645 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a
646 // single shift. We can do this if the bottom bits (which are shifted
647 // out) are never demanded.
648 if (InOp.getOpcode() == ISD::SRL &&
649 isa<ConstantSDNode>(InOp.getOperand(1))) {
650 if (ShAmt && (DemandedMask & ((1ULL << ShAmt)-1)) == 0) {
651 unsigned C1 = cast<ConstantSDNode>(InOp.getOperand(1))->getValue();
652 unsigned Opc = ISD::SHL;
660 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
661 MVT::ValueType VT = Op.getValueType();
662 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, VT,
663 InOp.getOperand(0), NewSA));
667 if (SimplifyDemandedBits(Op.getOperand(0), DemandedMask >> ShAmt,
668 KnownZero, KnownOne, TLO, Depth+1))
670 KnownZero <<= SA->getValue();
671 KnownOne <<= SA->getValue();
672 KnownZero |= (1ULL << SA->getValue())-1; // low bits known zero.
676 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
677 MVT::ValueType VT = Op.getValueType();
678 unsigned ShAmt = SA->getValue();
679 uint64_t TypeMask = MVT::getIntVTBitMask(VT);
680 unsigned VTSize = MVT::getSizeInBits(VT);
681 SDOperand InOp = Op.getOperand(0);
683 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a
684 // single shift. We can do this if the top bits (which are shifted out)
685 // are never demanded.
686 if (InOp.getOpcode() == ISD::SHL &&
687 isa<ConstantSDNode>(InOp.getOperand(1))) {
688 if (ShAmt && (DemandedMask & (~0ULL << (VTSize-ShAmt))) == 0) {
689 unsigned C1 = cast<ConstantSDNode>(InOp.getOperand(1))->getValue();
690 unsigned Opc = ISD::SRL;
698 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
699 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, VT,
700 InOp.getOperand(0), NewSA));
704 // Compute the new bits that are at the top now.
705 if (SimplifyDemandedBits(InOp, (DemandedMask << ShAmt) & TypeMask,
706 KnownZero, KnownOne, TLO, Depth+1))
708 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
709 KnownZero &= TypeMask;
710 KnownOne &= TypeMask;
714 uint64_t HighBits = (1ULL << ShAmt)-1;
715 HighBits <<= VTSize - ShAmt;
716 KnownZero |= HighBits; // High bits known zero.
720 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
721 MVT::ValueType VT = Op.getValueType();
722 unsigned ShAmt = SA->getValue();
724 // Compute the new bits that are at the top now.
725 uint64_t TypeMask = MVT::getIntVTBitMask(VT);
727 uint64_t InDemandedMask = (DemandedMask << ShAmt) & TypeMask;
729 // If any of the demanded bits are produced by the sign extension, we also
730 // demand the input sign bit.
731 uint64_t HighBits = (1ULL << ShAmt)-1;
732 HighBits <<= MVT::getSizeInBits(VT) - ShAmt;
733 if (HighBits & DemandedMask)
734 InDemandedMask |= MVT::getIntVTSignBit(VT);
736 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask,
737 KnownZero, KnownOne, TLO, Depth+1))
739 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
740 KnownZero &= TypeMask;
741 KnownOne &= TypeMask;
745 // Handle the sign bits.
746 uint64_t SignBit = MVT::getIntVTSignBit(VT);
747 SignBit >>= ShAmt; // Adjust to where it is now in the mask.
749 // If the input sign bit is known to be zero, or if none of the top bits
750 // are demanded, turn this into an unsigned shift right.
751 if ((KnownZero & SignBit) || (HighBits & ~DemandedMask) == HighBits) {
752 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, VT, Op.getOperand(0),
754 } else if (KnownOne & SignBit) { // New bits are known one.
755 KnownOne |= HighBits;
759 case ISD::SIGN_EXTEND_INREG: {
760 MVT::ValueType EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
762 // Sign extension. Compute the demanded bits in the result that are not
763 // present in the input.
764 uint64_t NewBits = ~MVT::getIntVTBitMask(EVT) & DemandedMask;
766 // If none of the extended bits are demanded, eliminate the sextinreg.
768 return TLO.CombineTo(Op, Op.getOperand(0));
770 uint64_t InSignBit = MVT::getIntVTSignBit(EVT);
771 int64_t InputDemandedBits = DemandedMask & MVT::getIntVTBitMask(EVT);
773 // Since the sign extended bits are demanded, we know that the sign
775 InputDemandedBits |= InSignBit;
777 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits,
778 KnownZero, KnownOne, TLO, Depth+1))
780 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
782 // If the sign bit of the input is known set or clear, then we know the
783 // top bits of the result.
785 // If the input sign bit is known zero, convert this into a zero extension.
786 if (KnownZero & InSignBit)
787 return TLO.CombineTo(Op,
788 TLO.DAG.getZeroExtendInReg(Op.getOperand(0), EVT));
790 if (KnownOne & InSignBit) { // Input sign bit known set
792 KnownZero &= ~NewBits;
793 } else { // Input sign bit unknown
794 KnownZero &= ~NewBits;
795 KnownOne &= ~NewBits;
802 MVT::ValueType VT = Op.getValueType();
803 unsigned LowBits = Log2_32(MVT::getSizeInBits(VT))+1;
804 KnownZero = ~((1ULL << LowBits)-1) & MVT::getIntVTBitMask(VT);
809 if (ISD::isZEXTLoad(Op.Val)) {
810 LoadSDNode *LD = cast<LoadSDNode>(Op);
811 MVT::ValueType VT = LD->getLoadedVT();
812 KnownZero |= ~MVT::getIntVTBitMask(VT) & DemandedMask;
816 case ISD::ZERO_EXTEND: {
817 uint64_t InMask = MVT::getIntVTBitMask(Op.getOperand(0).getValueType());
819 // If none of the top bits are demanded, convert this into an any_extend.
820 uint64_t NewBits = (~InMask) & DemandedMask;
822 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND,
826 if (SimplifyDemandedBits(Op.getOperand(0), DemandedMask & InMask,
827 KnownZero, KnownOne, TLO, Depth+1))
829 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
830 KnownZero |= NewBits;
833 case ISD::SIGN_EXTEND: {
834 MVT::ValueType InVT = Op.getOperand(0).getValueType();
835 uint64_t InMask = MVT::getIntVTBitMask(InVT);
836 uint64_t InSignBit = MVT::getIntVTSignBit(InVT);
837 uint64_t NewBits = (~InMask) & DemandedMask;
839 // If none of the top bits are demanded, convert this into an any_extend.
841 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND,Op.getValueType(),
844 // Since some of the sign extended bits are demanded, we know that the sign
846 uint64_t InDemandedBits = DemandedMask & InMask;
847 InDemandedBits |= InSignBit;
849 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero,
850 KnownOne, TLO, Depth+1))
853 // If the sign bit is known zero, convert this to a zero extend.
854 if (KnownZero & InSignBit)
855 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND,
859 // If the sign bit is known one, the top bits match.
860 if (KnownOne & InSignBit) {
862 KnownZero &= ~NewBits;
863 } else { // Otherwise, top bits aren't known.
864 KnownOne &= ~NewBits;
865 KnownZero &= ~NewBits;
869 case ISD::ANY_EXTEND: {
870 uint64_t InMask = MVT::getIntVTBitMask(Op.getOperand(0).getValueType());
871 if (SimplifyDemandedBits(Op.getOperand(0), DemandedMask & InMask,
872 KnownZero, KnownOne, TLO, Depth+1))
874 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
877 case ISD::TRUNCATE: {
878 // Simplify the input, using demanded bit information, and compute the known
879 // zero/one bits live out.
880 if (SimplifyDemandedBits(Op.getOperand(0), DemandedMask,
881 KnownZero, KnownOne, TLO, Depth+1))
884 // If the input is only used by this truncate, see if we can shrink it based
885 // on the known demanded bits.
886 if (Op.getOperand(0).Val->hasOneUse()) {
887 SDOperand In = Op.getOperand(0);
888 switch (In.getOpcode()) {
891 // Shrink SRL by a constant if none of the high bits shifted in are
893 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1))){
894 uint64_t HighBits = MVT::getIntVTBitMask(In.getValueType());
895 HighBits &= ~MVT::getIntVTBitMask(Op.getValueType());
896 HighBits >>= ShAmt->getValue();
898 if (ShAmt->getValue() < MVT::getSizeInBits(Op.getValueType()) &&
899 (DemandedMask & HighBits) == 0) {
900 // None of the shifted in bits are needed. Add a truncate of the
901 // shift input, then shift it.
902 SDOperand NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE,
905 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL,Op.getValueType(),
906 NewTrunc, In.getOperand(1)));
913 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
914 uint64_t OutMask = MVT::getIntVTBitMask(Op.getValueType());
915 KnownZero &= OutMask;
919 case ISD::AssertZext: {
920 MVT::ValueType VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
921 uint64_t InMask = MVT::getIntVTBitMask(VT);
922 if (SimplifyDemandedBits(Op.getOperand(0), DemandedMask & InMask,
923 KnownZero, KnownOne, TLO, Depth+1))
925 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
926 KnownZero |= ~InMask & DemandedMask;
931 case ISD::INTRINSIC_WO_CHAIN:
932 case ISD::INTRINSIC_W_CHAIN:
933 case ISD::INTRINSIC_VOID:
934 // Just use ComputeMaskedBits to compute output bits.
935 TLO.DAG.ComputeMaskedBits(Op, DemandedMask, KnownZero, KnownOne, Depth);
939 // If we know the value of all of the demanded bits, return this as a
941 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask)
942 return TLO.CombineTo(Op, TLO.DAG.getConstant(KnownOne, Op.getValueType()));
947 /// computeMaskedBitsForTargetNode - Determine which of the bits specified
948 /// in Mask are known to be either zero or one and return them in the
949 /// KnownZero/KnownOne bitsets.
950 void TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
954 const SelectionDAG &DAG,
955 unsigned Depth) const {
956 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
957 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
958 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
959 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
960 "Should use MaskedValueIsZero if you don't know whether Op"
961 " is a target node!");
966 /// ComputeNumSignBitsForTargetNode - This method can be implemented by
967 /// targets that want to expose additional information about sign bits to the
969 unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDOperand Op,
970 unsigned Depth) const {
971 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
972 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
973 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
974 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
975 "Should use ComputeNumSignBits if you don't know whether Op"
976 " is a target node!");
981 /// SimplifySetCC - Try to simplify a setcc built with the specified operands
982 /// and cc. If it is unable to simplify it, return a null SDOperand.
984 TargetLowering::SimplifySetCC(MVT::ValueType VT, SDOperand N0, SDOperand N1,
985 ISD::CondCode Cond, bool foldBooleans,
986 DAGCombinerInfo &DCI) const {
987 SelectionDAG &DAG = DCI.DAG;
989 // These setcc operations always fold.
993 case ISD::SETFALSE2: return DAG.getConstant(0, VT);
995 case ISD::SETTRUE2: return DAG.getConstant(1, VT);
998 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val)) {
999 uint64_t C1 = N1C->getValue();
1000 if (isa<ConstantSDNode>(N0.Val)) {
1001 return DAG.FoldSetCC(VT, N0, N1, Cond);
1003 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an
1004 // equality comparison, then we're just comparing whether X itself is
1006 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) &&
1007 N0.getOperand(0).getOpcode() == ISD::CTLZ &&
1008 N0.getOperand(1).getOpcode() == ISD::Constant) {
1009 unsigned ShAmt = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1010 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1011 ShAmt == Log2_32(MVT::getSizeInBits(N0.getValueType()))) {
1012 if ((C1 == 0) == (Cond == ISD::SETEQ)) {
1013 // (srl (ctlz x), 5) == 0 -> X != 0
1014 // (srl (ctlz x), 5) != 1 -> X != 0
1017 // (srl (ctlz x), 5) != 0 -> X == 0
1018 // (srl (ctlz x), 5) == 1 -> X == 0
1021 SDOperand Zero = DAG.getConstant(0, N0.getValueType());
1022 return DAG.getSetCC(VT, N0.getOperand(0).getOperand(0),
1027 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
1028 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
1029 unsigned InSize = MVT::getSizeInBits(N0.getOperand(0).getValueType());
1031 // If the comparison constant has bits in the upper part, the
1032 // zero-extended value could never match.
1033 if (C1 & (~0ULL << InSize)) {
1034 unsigned VSize = MVT::getSizeInBits(N0.getValueType());
1038 case ISD::SETEQ: return DAG.getConstant(0, VT);
1041 case ISD::SETNE: return DAG.getConstant(1, VT);
1044 // True if the sign bit of C1 is set.
1045 return DAG.getConstant((C1 & (1ULL << (VSize-1))) != 0, VT);
1048 // True if the sign bit of C1 isn't set.
1049 return DAG.getConstant((C1 & (1ULL << (VSize-1))) == 0, VT);
1055 // Otherwise, we can perform the comparison with the low bits.
1063 return DAG.getSetCC(VT, N0.getOperand(0),
1064 DAG.getConstant(C1, N0.getOperand(0).getValueType()),
1067 break; // todo, be more careful with signed comparisons
1069 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
1070 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1071 MVT::ValueType ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
1072 unsigned ExtSrcTyBits = MVT::getSizeInBits(ExtSrcTy);
1073 MVT::ValueType ExtDstTy = N0.getValueType();
1074 unsigned ExtDstTyBits = MVT::getSizeInBits(ExtDstTy);
1076 // If the extended part has any inconsistent bits, it cannot ever
1077 // compare equal. In other words, they have to be all ones or all
1080 (~0ULL >> (64-ExtSrcTyBits)) & (~0ULL << (ExtDstTyBits-1));
1081 if ((C1 & ExtBits) != 0 && (C1 & ExtBits) != ExtBits)
1082 return DAG.getConstant(Cond == ISD::SETNE, VT);
1085 MVT::ValueType Op0Ty = N0.getOperand(0).getValueType();
1086 if (Op0Ty == ExtSrcTy) {
1087 ZextOp = N0.getOperand(0);
1089 int64_t Imm = ~0ULL >> (64-ExtSrcTyBits);
1090 ZextOp = DAG.getNode(ISD::AND, Op0Ty, N0.getOperand(0),
1091 DAG.getConstant(Imm, Op0Ty));
1093 if (!DCI.isCalledByLegalizer())
1094 DCI.AddToWorklist(ZextOp.Val);
1095 // Otherwise, make this a use of a zext.
1096 return DAG.getSetCC(VT, ZextOp,
1097 DAG.getConstant(C1 & (~0ULL>>(64-ExtSrcTyBits)),
1100 } else if ((N1C->getValue() == 0 || N1C->getValue() == 1) &&
1101 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1103 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC
1104 if (N0.getOpcode() == ISD::SETCC) {
1105 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getValue() != 1);
1109 // Invert the condition.
1110 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
1111 CC = ISD::getSetCCInverse(CC,
1112 MVT::isInteger(N0.getOperand(0).getValueType()));
1113 return DAG.getSetCC(VT, N0.getOperand(0), N0.getOperand(1), CC);
1116 if ((N0.getOpcode() == ISD::XOR ||
1117 (N0.getOpcode() == ISD::AND &&
1118 N0.getOperand(0).getOpcode() == ISD::XOR &&
1119 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
1120 isa<ConstantSDNode>(N0.getOperand(1)) &&
1121 cast<ConstantSDNode>(N0.getOperand(1))->getValue() == 1) {
1122 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We
1123 // can only do this if the top bits are known zero.
1124 if (DAG.MaskedValueIsZero(N0,
1125 MVT::getIntVTBitMask(N0.getValueType())-1)){
1126 // Okay, get the un-inverted input value.
1128 if (N0.getOpcode() == ISD::XOR)
1129 Val = N0.getOperand(0);
1131 assert(N0.getOpcode() == ISD::AND &&
1132 N0.getOperand(0).getOpcode() == ISD::XOR);
1133 // ((X^1)&1)^1 -> X & 1
1134 Val = DAG.getNode(ISD::AND, N0.getValueType(),
1135 N0.getOperand(0).getOperand(0),
1138 return DAG.getSetCC(VT, Val, N1,
1139 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
1144 uint64_t MinVal, MaxVal;
1145 unsigned OperandBitSize = MVT::getSizeInBits(N1C->getValueType(0));
1146 if (ISD::isSignedIntSetCC(Cond)) {
1147 MinVal = 1ULL << (OperandBitSize-1);
1148 if (OperandBitSize != 1) // Avoid X >> 64, which is undefined.
1149 MaxVal = ~0ULL >> (65-OperandBitSize);
1154 MaxVal = ~0ULL >> (64-OperandBitSize);
1157 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
1158 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
1159 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true
1160 --C1; // X >= C0 --> X > (C0-1)
1161 return DAG.getSetCC(VT, N0, DAG.getConstant(C1, N1.getValueType()),
1162 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
1165 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
1166 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true
1167 ++C1; // X <= C0 --> X < (C0+1)
1168 return DAG.getSetCC(VT, N0, DAG.getConstant(C1, N1.getValueType()),
1169 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
1172 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
1173 return DAG.getConstant(0, VT); // X < MIN --> false
1174 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal)
1175 return DAG.getConstant(1, VT); // X >= MIN --> true
1176 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal)
1177 return DAG.getConstant(0, VT); // X > MAX --> false
1178 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal)
1179 return DAG.getConstant(1, VT); // X <= MAX --> true
1181 // Canonicalize setgt X, Min --> setne X, Min
1182 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
1183 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
1184 // Canonicalize setlt X, Max --> setne X, Max
1185 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
1186 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
1188 // If we have setult X, 1, turn it into seteq X, 0
1189 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
1190 return DAG.getSetCC(VT, N0, DAG.getConstant(MinVal, N0.getValueType()),
1192 // If we have setugt X, Max-1, turn it into seteq X, Max
1193 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
1194 return DAG.getSetCC(VT, N0, DAG.getConstant(MaxVal, N0.getValueType()),
1197 // If we have "setcc X, C0", check to see if we can shrink the immediate
1200 // SETUGT X, SINTMAX -> SETLT X, 0
1201 if (Cond == ISD::SETUGT && OperandBitSize != 1 &&
1202 C1 == (~0ULL >> (65-OperandBitSize)))
1203 return DAG.getSetCC(VT, N0, DAG.getConstant(0, N1.getValueType()),
1206 // FIXME: Implement the rest of these.
1208 // Fold bit comparisons when we can.
1209 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1210 VT == N0.getValueType() && N0.getOpcode() == ISD::AND)
1211 if (ConstantSDNode *AndRHS =
1212 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1213 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
1214 // Perform the xform if the AND RHS is a single bit.
1215 if (isPowerOf2_64(AndRHS->getValue())) {
1216 return DAG.getNode(ISD::SRL, VT, N0,
1217 DAG.getConstant(Log2_64(AndRHS->getValue()),
1218 getShiftAmountTy()));
1220 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getValue()) {
1221 // (X & 8) == 8 --> (X & 8) >> 3
1222 // Perform the xform if C1 is a single bit.
1223 if (isPowerOf2_64(C1)) {
1224 return DAG.getNode(ISD::SRL, VT, N0,
1225 DAG.getConstant(Log2_64(C1), getShiftAmountTy()));
1230 } else if (isa<ConstantSDNode>(N0.Val)) {
1231 // Ensure that the constant occurs on the RHS.
1232 return DAG.getSetCC(VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
1235 if (isa<ConstantFPSDNode>(N0.Val)) {
1236 // Constant fold or commute setcc.
1237 SDOperand O = DAG.FoldSetCC(VT, N0, N1, Cond);
1238 if (O.Val) return O;
1242 // We can always fold X == X for integer setcc's.
1243 if (MVT::isInteger(N0.getValueType()))
1244 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
1245 unsigned UOF = ISD::getUnorderedFlavor(Cond);
1246 if (UOF == 2) // FP operators that are undefined on NaNs.
1247 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
1248 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
1249 return DAG.getConstant(UOF, VT);
1250 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
1251 // if it is not already.
1252 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
1253 if (NewCond != Cond)
1254 return DAG.getSetCC(VT, N0, N1, NewCond);
1257 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1258 MVT::isInteger(N0.getValueType())) {
1259 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
1260 N0.getOpcode() == ISD::XOR) {
1261 // Simplify (X+Y) == (X+Z) --> Y == Z
1262 if (N0.getOpcode() == N1.getOpcode()) {
1263 if (N0.getOperand(0) == N1.getOperand(0))
1264 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(1), Cond);
1265 if (N0.getOperand(1) == N1.getOperand(1))
1266 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(0), Cond);
1267 if (DAG.isCommutativeBinOp(N0.getOpcode())) {
1268 // If X op Y == Y op X, try other combinations.
1269 if (N0.getOperand(0) == N1.getOperand(1))
1270 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(0), Cond);
1271 if (N0.getOperand(1) == N1.getOperand(0))
1272 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(1), Cond);
1276 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) {
1277 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1278 // Turn (X+C1) == C2 --> X == C2-C1
1279 if (N0.getOpcode() == ISD::ADD && N0.Val->hasOneUse()) {
1280 return DAG.getSetCC(VT, N0.getOperand(0),
1281 DAG.getConstant(RHSC->getValue()-LHSR->getValue(),
1282 N0.getValueType()), Cond);
1285 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
1286 if (N0.getOpcode() == ISD::XOR)
1287 // If we know that all of the inverted bits are zero, don't bother
1288 // performing the inversion.
1289 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getValue()))
1290 return DAG.getSetCC(VT, N0.getOperand(0),
1291 DAG.getConstant(LHSR->getValue()^RHSC->getValue(),
1292 N0.getValueType()), Cond);
1295 // Turn (C1-X) == C2 --> X == C1-C2
1296 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
1297 if (N0.getOpcode() == ISD::SUB && N0.Val->hasOneUse()) {
1298 return DAG.getSetCC(VT, N0.getOperand(1),
1299 DAG.getConstant(SUBC->getValue()-RHSC->getValue(),
1300 N0.getValueType()), Cond);
1305 // Simplify (X+Z) == X --> Z == 0
1306 if (N0.getOperand(0) == N1)
1307 return DAG.getSetCC(VT, N0.getOperand(1),
1308 DAG.getConstant(0, N0.getValueType()), Cond);
1309 if (N0.getOperand(1) == N1) {
1310 if (DAG.isCommutativeBinOp(N0.getOpcode()))
1311 return DAG.getSetCC(VT, N0.getOperand(0),
1312 DAG.getConstant(0, N0.getValueType()), Cond);
1313 else if (N0.Val->hasOneUse()) {
1314 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
1315 // (Z-X) == X --> Z == X<<1
1316 SDOperand SH = DAG.getNode(ISD::SHL, N1.getValueType(),
1318 DAG.getConstant(1, getShiftAmountTy()));
1319 if (!DCI.isCalledByLegalizer())
1320 DCI.AddToWorklist(SH.Val);
1321 return DAG.getSetCC(VT, N0.getOperand(0), SH, Cond);
1326 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
1327 N1.getOpcode() == ISD::XOR) {
1328 // Simplify X == (X+Z) --> Z == 0
1329 if (N1.getOperand(0) == N0) {
1330 return DAG.getSetCC(VT, N1.getOperand(1),
1331 DAG.getConstant(0, N1.getValueType()), Cond);
1332 } else if (N1.getOperand(1) == N0) {
1333 if (DAG.isCommutativeBinOp(N1.getOpcode())) {
1334 return DAG.getSetCC(VT, N1.getOperand(0),
1335 DAG.getConstant(0, N1.getValueType()), Cond);
1336 } else if (N1.Val->hasOneUse()) {
1337 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
1338 // X == (Z-X) --> X<<1 == Z
1339 SDOperand SH = DAG.getNode(ISD::SHL, N1.getValueType(), N0,
1340 DAG.getConstant(1, getShiftAmountTy()));
1341 if (!DCI.isCalledByLegalizer())
1342 DCI.AddToWorklist(SH.Val);
1343 return DAG.getSetCC(VT, SH, N1.getOperand(0), Cond);
1349 // Fold away ALL boolean setcc's.
1351 if (N0.getValueType() == MVT::i1 && foldBooleans) {
1353 default: assert(0 && "Unknown integer setcc!");
1354 case ISD::SETEQ: // X == Y -> (X^Y)^1
1355 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
1356 N0 = DAG.getNode(ISD::XOR, MVT::i1, Temp, DAG.getConstant(1, MVT::i1));
1357 if (!DCI.isCalledByLegalizer())
1358 DCI.AddToWorklist(Temp.Val);
1360 case ISD::SETNE: // X != Y --> (X^Y)
1361 N0 = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
1363 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> X^1 & Y
1364 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> X^1 & Y
1365 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
1366 N0 = DAG.getNode(ISD::AND, MVT::i1, N1, Temp);
1367 if (!DCI.isCalledByLegalizer())
1368 DCI.AddToWorklist(Temp.Val);
1370 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> Y^1 & X
1371 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> Y^1 & X
1372 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
1373 N0 = DAG.getNode(ISD::AND, MVT::i1, N0, Temp);
1374 if (!DCI.isCalledByLegalizer())
1375 DCI.AddToWorklist(Temp.Val);
1377 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> X^1 | Y
1378 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> X^1 | Y
1379 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
1380 N0 = DAG.getNode(ISD::OR, MVT::i1, N1, Temp);
1381 if (!DCI.isCalledByLegalizer())
1382 DCI.AddToWorklist(Temp.Val);
1384 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> Y^1 | X
1385 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> Y^1 | X
1386 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
1387 N0 = DAG.getNode(ISD::OR, MVT::i1, N0, Temp);
1390 if (VT != MVT::i1) {
1391 if (!DCI.isCalledByLegalizer())
1392 DCI.AddToWorklist(N0.Val);
1393 // FIXME: If running after legalize, we probably can't do this.
1394 N0 = DAG.getNode(ISD::ZERO_EXTEND, VT, N0);
1399 // Could not fold it.
1403 SDOperand TargetLowering::
1404 PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
1405 // Default implementation: no optimization.
1409 //===----------------------------------------------------------------------===//
1410 // Inline Assembler Implementation Methods
1411 //===----------------------------------------------------------------------===//
1413 TargetLowering::ConstraintType
1414 TargetLowering::getConstraintType(const std::string &Constraint) const {
1415 // FIXME: lots more standard ones to handle.
1416 if (Constraint.size() == 1) {
1417 switch (Constraint[0]) {
1419 case 'r': return C_RegisterClass;
1421 case 'o': // offsetable
1422 case 'V': // not offsetable
1424 case 'i': // Simple Integer or Relocatable Constant
1425 case 'n': // Simple Integer
1426 case 's': // Relocatable Constant
1427 case 'X': // Allow ANY value.
1428 case 'I': // Target registers.
1440 if (Constraint.size() > 1 && Constraint[0] == '{' &&
1441 Constraint[Constraint.size()-1] == '}')
1446 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
1447 /// vector. If it is invalid, don't add anything to Ops.
1448 void TargetLowering::LowerAsmOperandForConstraint(SDOperand Op,
1449 char ConstraintLetter,
1450 std::vector<SDOperand> &Ops,
1451 SelectionDAG &DAG) {
1452 switch (ConstraintLetter) {
1454 case 'X': // Allows any operand; labels (basic block) use this.
1455 if (Op.getOpcode() == ISD::BasicBlock) {
1460 case 'i': // Simple Integer or Relocatable Constant
1461 case 'n': // Simple Integer
1462 case 's': { // Relocatable Constant
1463 // These operands are interested in values of the form (GV+C), where C may
1464 // be folded in as an offset of GV, or it may be explicitly added. Also, it
1465 // is possible and fine if either GV or C are missing.
1466 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
1467 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
1469 // If we have "(add GV, C)", pull out GV/C
1470 if (Op.getOpcode() == ISD::ADD) {
1471 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
1472 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
1473 if (C == 0 || GA == 0) {
1474 C = dyn_cast<ConstantSDNode>(Op.getOperand(0));
1475 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1));
1477 if (C == 0 || GA == 0)
1481 // If we find a valid operand, map to the TargetXXX version so that the
1482 // value itself doesn't get selected.
1483 if (GA) { // Either &GV or &GV+C
1484 if (ConstraintLetter != 'n') {
1485 int64_t Offs = GA->getOffset();
1486 if (C) Offs += C->getValue();
1487 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(),
1488 Op.getValueType(), Offs));
1492 if (C) { // just C, no GV.
1493 // Simple constants are not allowed for 's'.
1494 if (ConstraintLetter != 's') {
1495 Ops.push_back(DAG.getTargetConstant(C->getValue(), Op.getValueType()));
1504 std::vector<unsigned> TargetLowering::
1505 getRegClassForInlineAsmConstraint(const std::string &Constraint,
1506 MVT::ValueType VT) const {
1507 return std::vector<unsigned>();
1511 std::pair<unsigned, const TargetRegisterClass*> TargetLowering::
1512 getRegForInlineAsmConstraint(const std::string &Constraint,
1513 MVT::ValueType VT) const {
1514 if (Constraint[0] != '{')
1515 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
1516 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?");
1518 // Remove the braces from around the name.
1519 std::string RegName(Constraint.begin()+1, Constraint.end()-1);
1521 // Figure out which register class contains this reg.
1522 const MRegisterInfo *RI = TM.getRegisterInfo();
1523 for (MRegisterInfo::regclass_iterator RCI = RI->regclass_begin(),
1524 E = RI->regclass_end(); RCI != E; ++RCI) {
1525 const TargetRegisterClass *RC = *RCI;
1527 // If none of the the value types for this register class are valid, we
1528 // can't use it. For example, 64-bit reg classes on 32-bit targets.
1529 bool isLegal = false;
1530 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
1532 if (isTypeLegal(*I)) {
1538 if (!isLegal) continue;
1540 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end();
1542 if (StringsEqualNoCase(RegName, RI->get(*I).Name))
1543 return std::make_pair(*I, RC);
1547 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
1550 //===----------------------------------------------------------------------===//
1551 // Loop Strength Reduction hooks
1552 //===----------------------------------------------------------------------===//
1554 /// isLegalAddressingMode - Return true if the addressing mode represented
1555 /// by AM is legal for this target, for a load/store of the specified type.
1556 bool TargetLowering::isLegalAddressingMode(const AddrMode &AM,
1557 const Type *Ty) const {
1558 // The default implementation of this implements a conservative RISCy, r+r and
1561 // Allows a sign-extended 16-bit immediate field.
1562 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
1565 // No global is ever allowed as a base.
1569 // Only support r+r,
1571 case 0: // "r+i" or just "i", depending on HasBaseReg.
1574 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
1576 // Otherwise we have r+r or r+i.
1579 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
1581 // Allow 2*r as r+r.
1588 // Magic for divide replacement
1591 int64_t m; // magic number
1592 int64_t s; // shift amount
1596 uint64_t m; // magic number
1597 int64_t a; // add indicator
1598 int64_t s; // shift amount
1601 /// magic - calculate the magic numbers required to codegen an integer sdiv as
1602 /// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
1604 static ms magic32(int32_t d) {
1606 uint32_t ad, anc, delta, q1, r1, q2, r2, t;
1607 const uint32_t two31 = 0x80000000U;
1611 t = two31 + ((uint32_t)d >> 31);
1612 anc = t - 1 - t%ad; // absolute value of nc
1613 p = 31; // initialize p
1614 q1 = two31/anc; // initialize q1 = 2p/abs(nc)
1615 r1 = two31 - q1*anc; // initialize r1 = rem(2p,abs(nc))
1616 q2 = two31/ad; // initialize q2 = 2p/abs(d)
1617 r2 = two31 - q2*ad; // initialize r2 = rem(2p,abs(d))
1620 q1 = 2*q1; // update q1 = 2p/abs(nc)
1621 r1 = 2*r1; // update r1 = rem(2p/abs(nc))
1622 if (r1 >= anc) { // must be unsigned comparison
1626 q2 = 2*q2; // update q2 = 2p/abs(d)
1627 r2 = 2*r2; // update r2 = rem(2p/abs(d))
1628 if (r2 >= ad) { // must be unsigned comparison
1633 } while (q1 < delta || (q1 == delta && r1 == 0));
1635 mag.m = (int32_t)(q2 + 1); // make sure to sign extend
1636 if (d < 0) mag.m = -mag.m; // resulting magic number
1637 mag.s = p - 32; // resulting shift
1641 /// magicu - calculate the magic numbers required to codegen an integer udiv as
1642 /// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
1643 static mu magicu32(uint32_t d) {
1645 uint32_t nc, delta, q1, r1, q2, r2;
1647 magu.a = 0; // initialize "add" indicator
1649 p = 31; // initialize p
1650 q1 = 0x80000000/nc; // initialize q1 = 2p/nc
1651 r1 = 0x80000000 - q1*nc; // initialize r1 = rem(2p,nc)
1652 q2 = 0x7FFFFFFF/d; // initialize q2 = (2p-1)/d
1653 r2 = 0x7FFFFFFF - q2*d; // initialize r2 = rem((2p-1),d)
1656 if (r1 >= nc - r1 ) {
1657 q1 = 2*q1 + 1; // update q1
1658 r1 = 2*r1 - nc; // update r1
1661 q1 = 2*q1; // update q1
1662 r1 = 2*r1; // update r1
1664 if (r2 + 1 >= d - r2) {
1665 if (q2 >= 0x7FFFFFFF) magu.a = 1;
1666 q2 = 2*q2 + 1; // update q2
1667 r2 = 2*r2 + 1 - d; // update r2
1670 if (q2 >= 0x80000000) magu.a = 1;
1671 q2 = 2*q2; // update q2
1672 r2 = 2*r2 + 1; // update r2
1675 } while (p < 64 && (q1 < delta || (q1 == delta && r1 == 0)));
1676 magu.m = q2 + 1; // resulting magic number
1677 magu.s = p - 32; // resulting shift
1681 /// magic - calculate the magic numbers required to codegen an integer sdiv as
1682 /// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
1684 static ms magic64(int64_t d) {
1686 uint64_t ad, anc, delta, q1, r1, q2, r2, t;
1687 const uint64_t two63 = 9223372036854775808ULL; // 2^63
1690 ad = d >= 0 ? d : -d;
1691 t = two63 + ((uint64_t)d >> 63);
1692 anc = t - 1 - t%ad; // absolute value of nc
1693 p = 63; // initialize p
1694 q1 = two63/anc; // initialize q1 = 2p/abs(nc)
1695 r1 = two63 - q1*anc; // initialize r1 = rem(2p,abs(nc))
1696 q2 = two63/ad; // initialize q2 = 2p/abs(d)
1697 r2 = two63 - q2*ad; // initialize r2 = rem(2p,abs(d))
1700 q1 = 2*q1; // update q1 = 2p/abs(nc)
1701 r1 = 2*r1; // update r1 = rem(2p/abs(nc))
1702 if (r1 >= anc) { // must be unsigned comparison
1706 q2 = 2*q2; // update q2 = 2p/abs(d)
1707 r2 = 2*r2; // update r2 = rem(2p/abs(d))
1708 if (r2 >= ad) { // must be unsigned comparison
1713 } while (q1 < delta || (q1 == delta && r1 == 0));
1716 if (d < 0) mag.m = -mag.m; // resulting magic number
1717 mag.s = p - 64; // resulting shift
1721 /// magicu - calculate the magic numbers required to codegen an integer udiv as
1722 /// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
1723 static mu magicu64(uint64_t d)
1726 uint64_t nc, delta, q1, r1, q2, r2;
1728 magu.a = 0; // initialize "add" indicator
1730 p = 63; // initialize p
1731 q1 = 0x8000000000000000ull/nc; // initialize q1 = 2p/nc
1732 r1 = 0x8000000000000000ull - q1*nc; // initialize r1 = rem(2p,nc)
1733 q2 = 0x7FFFFFFFFFFFFFFFull/d; // initialize q2 = (2p-1)/d
1734 r2 = 0x7FFFFFFFFFFFFFFFull - q2*d; // initialize r2 = rem((2p-1),d)
1737 if (r1 >= nc - r1 ) {
1738 q1 = 2*q1 + 1; // update q1
1739 r1 = 2*r1 - nc; // update r1
1742 q1 = 2*q1; // update q1
1743 r1 = 2*r1; // update r1
1745 if (r2 + 1 >= d - r2) {
1746 if (q2 >= 0x7FFFFFFFFFFFFFFFull) magu.a = 1;
1747 q2 = 2*q2 + 1; // update q2
1748 r2 = 2*r2 + 1 - d; // update r2
1751 if (q2 >= 0x8000000000000000ull) magu.a = 1;
1752 q2 = 2*q2; // update q2
1753 r2 = 2*r2 + 1; // update r2
1756 } while (p < 128 && (q1 < delta || (q1 == delta && r1 == 0)));
1757 magu.m = q2 + 1; // resulting magic number
1758 magu.s = p - 64; // resulting shift
1762 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
1763 /// return a DAG expression to select that will generate the same value by
1764 /// multiplying by a magic number. See:
1765 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
1766 SDOperand TargetLowering::BuildSDIV(SDNode *N, SelectionDAG &DAG,
1767 std::vector<SDNode*>* Created) const {
1768 MVT::ValueType VT = N->getValueType(0);
1770 // Check to see if we can do this.
1771 if (!isTypeLegal(VT) || (VT != MVT::i32 && VT != MVT::i64))
1772 return SDOperand(); // BuildSDIV only operates on i32 or i64
1774 int64_t d = cast<ConstantSDNode>(N->getOperand(1))->getSignExtended();
1775 ms magics = (VT == MVT::i32) ? magic32(d) : magic64(d);
1777 // Multiply the numerator (operand 0) by the magic value
1779 if (isOperationLegal(ISD::MULHS, VT))
1780 Q = DAG.getNode(ISD::MULHS, VT, N->getOperand(0),
1781 DAG.getConstant(magics.m, VT));
1782 else if (isOperationLegal(ISD::SMUL_LOHI, VT))
1783 Q = SDOperand(DAG.getNode(ISD::SMUL_LOHI, DAG.getVTList(VT, VT),
1785 DAG.getConstant(magics.m, VT)).Val, 1);
1787 return SDOperand(); // No mulhs or equvialent
1788 // If d > 0 and m < 0, add the numerator
1789 if (d > 0 && magics.m < 0) {
1790 Q = DAG.getNode(ISD::ADD, VT, Q, N->getOperand(0));
1792 Created->push_back(Q.Val);
1794 // If d < 0 and m > 0, subtract the numerator.
1795 if (d < 0 && magics.m > 0) {
1796 Q = DAG.getNode(ISD::SUB, VT, Q, N->getOperand(0));
1798 Created->push_back(Q.Val);
1800 // Shift right algebraic if shift value is nonzero
1802 Q = DAG.getNode(ISD::SRA, VT, Q,
1803 DAG.getConstant(magics.s, getShiftAmountTy()));
1805 Created->push_back(Q.Val);
1807 // Extract the sign bit and add it to the quotient
1809 DAG.getNode(ISD::SRL, VT, Q, DAG.getConstant(MVT::getSizeInBits(VT)-1,
1810 getShiftAmountTy()));
1812 Created->push_back(T.Val);
1813 return DAG.getNode(ISD::ADD, VT, Q, T);
1816 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
1817 /// return a DAG expression to select that will generate the same value by
1818 /// multiplying by a magic number. See:
1819 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
1820 SDOperand TargetLowering::BuildUDIV(SDNode *N, SelectionDAG &DAG,
1821 std::vector<SDNode*>* Created) const {
1822 MVT::ValueType VT = N->getValueType(0);
1824 // Check to see if we can do this.
1825 if (!isTypeLegal(VT) || (VT != MVT::i32 && VT != MVT::i64))
1826 return SDOperand(); // BuildUDIV only operates on i32 or i64
1828 uint64_t d = cast<ConstantSDNode>(N->getOperand(1))->getValue();
1829 mu magics = (VT == MVT::i32) ? magicu32(d) : magicu64(d);
1831 // Multiply the numerator (operand 0) by the magic value
1833 if (isOperationLegal(ISD::MULHU, VT))
1834 Q = DAG.getNode(ISD::MULHU, VT, N->getOperand(0),
1835 DAG.getConstant(magics.m, VT));
1836 else if (isOperationLegal(ISD::UMUL_LOHI, VT))
1837 Q = SDOperand(DAG.getNode(ISD::UMUL_LOHI, DAG.getVTList(VT, VT),
1839 DAG.getConstant(magics.m, VT)).Val, 1);
1841 return SDOperand(); // No mulhu or equvialent
1843 Created->push_back(Q.Val);
1845 if (magics.a == 0) {
1846 return DAG.getNode(ISD::SRL, VT, Q,
1847 DAG.getConstant(magics.s, getShiftAmountTy()));
1849 SDOperand NPQ = DAG.getNode(ISD::SUB, VT, N->getOperand(0), Q);
1851 Created->push_back(NPQ.Val);
1852 NPQ = DAG.getNode(ISD::SRL, VT, NPQ,
1853 DAG.getConstant(1, getShiftAmountTy()));
1855 Created->push_back(NPQ.Val);
1856 NPQ = DAG.getNode(ISD::ADD, VT, NPQ, Q);
1858 Created->push_back(NPQ.Val);
1859 return DAG.getNode(ISD::SRL, VT, NPQ,
1860 DAG.getConstant(magics.s-1, getShiftAmountTy()));