1 //===-- TargetLowering.cpp - Implement the TargetLowering class -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This implements the TargetLowering class.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Target/TargetAsmInfo.h"
15 #include "llvm/Target/TargetLowering.h"
16 #include "llvm/Target/TargetSubtarget.h"
17 #include "llvm/Target/TargetData.h"
18 #include "llvm/Target/TargetMachine.h"
19 #include "llvm/Target/TargetRegisterInfo.h"
20 #include "llvm/GlobalVariable.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/CodeGen/MachineFrameInfo.h"
23 #include "llvm/CodeGen/SelectionDAG.h"
24 #include "llvm/ADT/StringExtras.h"
25 #include "llvm/ADT/STLExtras.h"
26 #include "llvm/Support/MathExtras.h"
29 /// InitLibcallNames - Set default libcall names.
31 static void InitLibcallNames(const char **Names) {
32 Names[RTLIB::SHL_I32] = "__ashlsi3";
33 Names[RTLIB::SHL_I64] = "__ashldi3";
34 Names[RTLIB::SHL_I128] = "__ashlti3";
35 Names[RTLIB::SRL_I32] = "__lshrsi3";
36 Names[RTLIB::SRL_I64] = "__lshrdi3";
37 Names[RTLIB::SRL_I128] = "__lshrti3";
38 Names[RTLIB::SRA_I32] = "__ashrsi3";
39 Names[RTLIB::SRA_I64] = "__ashrdi3";
40 Names[RTLIB::SRA_I128] = "__ashrti3";
41 Names[RTLIB::MUL_I32] = "__mulsi3";
42 Names[RTLIB::MUL_I64] = "__muldi3";
43 Names[RTLIB::MUL_I128] = "__multi3";
44 Names[RTLIB::SDIV_I32] = "__divsi3";
45 Names[RTLIB::SDIV_I64] = "__divdi3";
46 Names[RTLIB::SDIV_I128] = "__divti3";
47 Names[RTLIB::UDIV_I32] = "__udivsi3";
48 Names[RTLIB::UDIV_I64] = "__udivdi3";
49 Names[RTLIB::UDIV_I128] = "__udivti3";
50 Names[RTLIB::SREM_I32] = "__modsi3";
51 Names[RTLIB::SREM_I64] = "__moddi3";
52 Names[RTLIB::SREM_I128] = "__modti3";
53 Names[RTLIB::UREM_I32] = "__umodsi3";
54 Names[RTLIB::UREM_I64] = "__umoddi3";
55 Names[RTLIB::UREM_I128] = "__umodti3";
56 Names[RTLIB::NEG_I32] = "__negsi2";
57 Names[RTLIB::NEG_I64] = "__negdi2";
58 Names[RTLIB::ADD_F32] = "__addsf3";
59 Names[RTLIB::ADD_F64] = "__adddf3";
60 Names[RTLIB::ADD_F80] = "__addxf3";
61 Names[RTLIB::ADD_PPCF128] = "__gcc_qadd";
62 Names[RTLIB::SUB_F32] = "__subsf3";
63 Names[RTLIB::SUB_F64] = "__subdf3";
64 Names[RTLIB::SUB_F80] = "__subxf3";
65 Names[RTLIB::SUB_PPCF128] = "__gcc_qsub";
66 Names[RTLIB::MUL_F32] = "__mulsf3";
67 Names[RTLIB::MUL_F64] = "__muldf3";
68 Names[RTLIB::MUL_F80] = "__mulxf3";
69 Names[RTLIB::MUL_PPCF128] = "__gcc_qmul";
70 Names[RTLIB::DIV_F32] = "__divsf3";
71 Names[RTLIB::DIV_F64] = "__divdf3";
72 Names[RTLIB::DIV_F80] = "__divxf3";
73 Names[RTLIB::DIV_PPCF128] = "__gcc_qdiv";
74 Names[RTLIB::REM_F32] = "fmodf";
75 Names[RTLIB::REM_F64] = "fmod";
76 Names[RTLIB::REM_F80] = "fmodl";
77 Names[RTLIB::REM_PPCF128] = "fmodl";
78 Names[RTLIB::POWI_F32] = "__powisf2";
79 Names[RTLIB::POWI_F64] = "__powidf2";
80 Names[RTLIB::POWI_F80] = "__powixf2";
81 Names[RTLIB::POWI_PPCF128] = "__powitf2";
82 Names[RTLIB::SQRT_F32] = "sqrtf";
83 Names[RTLIB::SQRT_F64] = "sqrt";
84 Names[RTLIB::SQRT_F80] = "sqrtl";
85 Names[RTLIB::SQRT_PPCF128] = "sqrtl";
86 Names[RTLIB::SIN_F32] = "sinf";
87 Names[RTLIB::SIN_F64] = "sin";
88 Names[RTLIB::SIN_F80] = "sinl";
89 Names[RTLIB::SIN_PPCF128] = "sinl";
90 Names[RTLIB::COS_F32] = "cosf";
91 Names[RTLIB::COS_F64] = "cos";
92 Names[RTLIB::COS_F80] = "cosl";
93 Names[RTLIB::COS_PPCF128] = "cosl";
94 Names[RTLIB::POW_F32] = "powf";
95 Names[RTLIB::POW_F64] = "pow";
96 Names[RTLIB::POW_F80] = "powl";
97 Names[RTLIB::POW_PPCF128] = "powl";
98 Names[RTLIB::FPEXT_F32_F64] = "__extendsfdf2";
99 Names[RTLIB::FPROUND_F64_F32] = "__truncdfsf2";
100 Names[RTLIB::FPTOSINT_F32_I32] = "__fixsfsi";
101 Names[RTLIB::FPTOSINT_F32_I64] = "__fixsfdi";
102 Names[RTLIB::FPTOSINT_F32_I128] = "__fixsfti";
103 Names[RTLIB::FPTOSINT_F64_I32] = "__fixdfsi";
104 Names[RTLIB::FPTOSINT_F64_I64] = "__fixdfdi";
105 Names[RTLIB::FPTOSINT_F64_I128] = "__fixdfti";
106 Names[RTLIB::FPTOSINT_F80_I32] = "__fixxfsi";
107 Names[RTLIB::FPTOSINT_F80_I64] = "__fixxfdi";
108 Names[RTLIB::FPTOSINT_F80_I128] = "__fixxfti";
109 Names[RTLIB::FPTOSINT_PPCF128_I32] = "__fixtfsi";
110 Names[RTLIB::FPTOSINT_PPCF128_I64] = "__fixtfdi";
111 Names[RTLIB::FPTOSINT_PPCF128_I128] = "__fixtfti";
112 Names[RTLIB::FPTOUINT_F32_I32] = "__fixunssfsi";
113 Names[RTLIB::FPTOUINT_F32_I64] = "__fixunssfdi";
114 Names[RTLIB::FPTOUINT_F32_I128] = "__fixunssfti";
115 Names[RTLIB::FPTOUINT_F64_I32] = "__fixunsdfsi";
116 Names[RTLIB::FPTOUINT_F64_I64] = "__fixunsdfdi";
117 Names[RTLIB::FPTOUINT_F64_I128] = "__fixunsdfti";
118 Names[RTLIB::FPTOUINT_F80_I32] = "__fixunsxfsi";
119 Names[RTLIB::FPTOUINT_F80_I64] = "__fixunsxfdi";
120 Names[RTLIB::FPTOUINT_F80_I128] = "__fixunsxfti";
121 Names[RTLIB::FPTOUINT_PPCF128_I32] = "__fixunstfsi";
122 Names[RTLIB::FPTOUINT_PPCF128_I64] = "__fixunstfdi";
123 Names[RTLIB::FPTOUINT_PPCF128_I128] = "__fixunstfti";
124 Names[RTLIB::SINTTOFP_I32_F32] = "__floatsisf";
125 Names[RTLIB::SINTTOFP_I32_F64] = "__floatsidf";
126 Names[RTLIB::SINTTOFP_I32_F80] = "__floatsixf";
127 Names[RTLIB::SINTTOFP_I32_PPCF128] = "__floatsitf";
128 Names[RTLIB::SINTTOFP_I64_F32] = "__floatdisf";
129 Names[RTLIB::SINTTOFP_I64_F64] = "__floatdidf";
130 Names[RTLIB::SINTTOFP_I64_F80] = "__floatdixf";
131 Names[RTLIB::SINTTOFP_I64_PPCF128] = "__floatditf";
132 Names[RTLIB::SINTTOFP_I128_F32] = "__floattisf";
133 Names[RTLIB::SINTTOFP_I128_F64] = "__floattidf";
134 Names[RTLIB::SINTTOFP_I128_F80] = "__floattixf";
135 Names[RTLIB::SINTTOFP_I128_PPCF128] = "__floattitf";
136 Names[RTLIB::UINTTOFP_I32_F32] = "__floatunsisf";
137 Names[RTLIB::UINTTOFP_I32_F64] = "__floatunsidf";
138 Names[RTLIB::UINTTOFP_I64_F32] = "__floatundisf";
139 Names[RTLIB::UINTTOFP_I64_F64] = "__floatundidf";
140 Names[RTLIB::OEQ_F32] = "__eqsf2";
141 Names[RTLIB::OEQ_F64] = "__eqdf2";
142 Names[RTLIB::UNE_F32] = "__nesf2";
143 Names[RTLIB::UNE_F64] = "__nedf2";
144 Names[RTLIB::OGE_F32] = "__gesf2";
145 Names[RTLIB::OGE_F64] = "__gedf2";
146 Names[RTLIB::OLT_F32] = "__ltsf2";
147 Names[RTLIB::OLT_F64] = "__ltdf2";
148 Names[RTLIB::OLE_F32] = "__lesf2";
149 Names[RTLIB::OLE_F64] = "__ledf2";
150 Names[RTLIB::OGT_F32] = "__gtsf2";
151 Names[RTLIB::OGT_F64] = "__gtdf2";
152 Names[RTLIB::UO_F32] = "__unordsf2";
153 Names[RTLIB::UO_F64] = "__unorddf2";
154 Names[RTLIB::O_F32] = "__unordsf2";
155 Names[RTLIB::O_F64] = "__unorddf2";
158 /// InitCmpLibcallCCs - Set default comparison libcall CC.
160 static void InitCmpLibcallCCs(ISD::CondCode *CCs) {
161 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL);
162 CCs[RTLIB::OEQ_F32] = ISD::SETEQ;
163 CCs[RTLIB::OEQ_F64] = ISD::SETEQ;
164 CCs[RTLIB::UNE_F32] = ISD::SETNE;
165 CCs[RTLIB::UNE_F64] = ISD::SETNE;
166 CCs[RTLIB::OGE_F32] = ISD::SETGE;
167 CCs[RTLIB::OGE_F64] = ISD::SETGE;
168 CCs[RTLIB::OLT_F32] = ISD::SETLT;
169 CCs[RTLIB::OLT_F64] = ISD::SETLT;
170 CCs[RTLIB::OLE_F32] = ISD::SETLE;
171 CCs[RTLIB::OLE_F64] = ISD::SETLE;
172 CCs[RTLIB::OGT_F32] = ISD::SETGT;
173 CCs[RTLIB::OGT_F64] = ISD::SETGT;
174 CCs[RTLIB::UO_F32] = ISD::SETNE;
175 CCs[RTLIB::UO_F64] = ISD::SETNE;
176 CCs[RTLIB::O_F32] = ISD::SETEQ;
177 CCs[RTLIB::O_F64] = ISD::SETEQ;
180 TargetLowering::TargetLowering(TargetMachine &tm)
181 : TM(tm), TD(TM.getTargetData()) {
182 assert(ISD::BUILTIN_OP_END <= OpActionsCapacity &&
183 "Fixed size array in TargetLowering is not large enough!");
184 // All operations default to being supported.
185 memset(OpActions, 0, sizeof(OpActions));
186 memset(LoadXActions, 0, sizeof(LoadXActions));
187 memset(TruncStoreActions, 0, sizeof(TruncStoreActions));
188 memset(IndexedModeActions, 0, sizeof(IndexedModeActions));
189 memset(ConvertActions, 0, sizeof(ConvertActions));
191 // Set default actions for various operations.
192 for (unsigned VT = 0; VT != (unsigned)MVT::LAST_VALUETYPE; ++VT) {
193 // Default all indexed load / store to expand.
194 for (unsigned IM = (unsigned)ISD::PRE_INC;
195 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
196 setIndexedLoadAction(IM, (MVT::SimpleValueType)VT, Expand);
197 setIndexedStoreAction(IM, (MVT::SimpleValueType)VT, Expand);
200 // These operations default to expand.
201 setOperationAction(ISD::FGETSIGN, (MVT::SimpleValueType)VT, Expand);
204 // Most targets ignore the @llvm.prefetch intrinsic.
205 setOperationAction(ISD::PREFETCH, MVT::Other, Expand);
207 // ConstantFP nodes default to expand. Targets can either change this to
208 // Legal, in which case all fp constants are legal, or use addLegalFPImmediate
209 // to optimize expansions for certain constants.
210 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
211 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
212 setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
214 // Default ISD::TRAP to expand (which turns it into abort).
215 setOperationAction(ISD::TRAP, MVT::Other, Expand);
217 IsLittleEndian = TD->isLittleEndian();
218 UsesGlobalOffsetTable = false;
219 ShiftAmountTy = PointerTy = getValueType(TD->getIntPtrType());
220 ShiftAmtHandling = Undefined;
221 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE*sizeof(TargetRegisterClass*));
222 memset(TargetDAGCombineArray, 0, array_lengthof(TargetDAGCombineArray));
223 maxStoresPerMemset = maxStoresPerMemcpy = maxStoresPerMemmove = 8;
224 allowUnalignedMemoryAccesses = false;
225 UseUnderscoreSetJmp = false;
226 UseUnderscoreLongJmp = false;
227 SelectIsExpensive = false;
228 IntDivIsCheap = false;
229 Pow2DivIsCheap = false;
230 StackPointerRegisterToSaveRestore = 0;
231 ExceptionPointerRegister = 0;
232 ExceptionSelectorRegister = 0;
233 SetCCResultContents = UndefinedSetCCResult;
234 SchedPreferenceInfo = SchedulingForLatency;
236 JumpBufAlignment = 0;
237 IfCvtBlockSizeLimit = 2;
238 IfCvtDupBlockSizeLimit = 0;
239 PrefLoopAlignment = 0;
241 InitLibcallNames(LibcallRoutineNames);
242 InitCmpLibcallCCs(CmpLibcallCCs);
244 // Tell Legalize whether the assembler supports DEBUG_LOC.
245 if (!TM.getTargetAsmInfo()->hasDotLocAndDotFile())
246 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
249 TargetLowering::~TargetLowering() {}
251 /// computeRegisterProperties - Once all of the register classes are added,
252 /// this allows us to compute derived properties we expose.
253 void TargetLowering::computeRegisterProperties() {
254 assert(MVT::LAST_VALUETYPE <= 32 &&
255 "Too many value types for ValueTypeActions to hold!");
257 // Everything defaults to needing one register.
258 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
259 NumRegistersForVT[i] = 1;
260 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i;
262 // ...except isVoid, which doesn't need any registers.
263 NumRegistersForVT[MVT::isVoid] = 0;
265 // Find the largest integer register class.
266 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE;
267 for (; RegClassForVT[LargestIntReg] == 0; --LargestIntReg)
268 assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
270 // Every integer value type larger than this largest register takes twice as
271 // many registers to represent as the previous ValueType.
272 for (unsigned ExpandedReg = LargestIntReg + 1; ; ++ExpandedReg) {
273 MVT EVT = (MVT::SimpleValueType)ExpandedReg;
274 if (!EVT.isInteger())
276 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
277 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg;
278 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1);
279 ValueTypeActions.setTypeAction(EVT, Expand);
282 // Inspect all of the ValueType's smaller than the largest integer
283 // register to see which ones need promotion.
284 unsigned LegalIntReg = LargestIntReg;
285 for (unsigned IntReg = LargestIntReg - 1;
286 IntReg >= (unsigned)MVT::i1; --IntReg) {
287 MVT IVT = (MVT::SimpleValueType)IntReg;
288 if (isTypeLegal(IVT)) {
289 LegalIntReg = IntReg;
291 RegisterTypeForVT[IntReg] = TransformToType[IntReg] =
292 (MVT::SimpleValueType)LegalIntReg;
293 ValueTypeActions.setTypeAction(IVT, Promote);
297 // ppcf128 type is really two f64's.
298 if (!isTypeLegal(MVT::ppcf128)) {
299 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64];
300 RegisterTypeForVT[MVT::ppcf128] = MVT::f64;
301 TransformToType[MVT::ppcf128] = MVT::f64;
302 ValueTypeActions.setTypeAction(MVT::ppcf128, Expand);
305 // Decide how to handle f64. If the target does not have native f64 support,
306 // expand it to i64 and we will be generating soft float library calls.
307 if (!isTypeLegal(MVT::f64)) {
308 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
309 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
310 TransformToType[MVT::f64] = MVT::i64;
311 ValueTypeActions.setTypeAction(MVT::f64, Expand);
314 // Decide how to handle f32. If the target does not have native support for
315 // f32, promote it to f64 if it is legal. Otherwise, expand it to i32.
316 if (!isTypeLegal(MVT::f32)) {
317 if (isTypeLegal(MVT::f64)) {
318 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::f64];
319 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::f64];
320 TransformToType[MVT::f32] = MVT::f64;
321 ValueTypeActions.setTypeAction(MVT::f32, Promote);
323 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
324 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
325 TransformToType[MVT::f32] = MVT::i32;
326 ValueTypeActions.setTypeAction(MVT::f32, Expand);
330 // Loop over all of the vector value types to see which need transformations.
331 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
332 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
333 MVT VT = (MVT::SimpleValueType)i;
334 if (!isTypeLegal(VT)) {
335 MVT IntermediateVT, RegisterVT;
336 unsigned NumIntermediates;
337 NumRegistersForVT[i] =
338 getVectorTypeBreakdown(VT,
339 IntermediateVT, NumIntermediates,
341 RegisterTypeForVT[i] = RegisterVT;
342 TransformToType[i] = MVT::Other; // this isn't actually used
343 ValueTypeActions.setTypeAction(VT, Expand);
348 const char *TargetLowering::getTargetNodeName(unsigned Opcode) const {
353 MVT TargetLowering::getSetCCResultType(const SDOperand &) const {
354 return getValueType(TD->getIntPtrType());
358 /// getVectorTypeBreakdown - Vector types are broken down into some number of
359 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
360 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
361 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
363 /// This method returns the number of registers needed, and the VT for each
364 /// register. It also returns the VT and quantity of the intermediate values
365 /// before they are promoted/expanded.
367 unsigned TargetLowering::getVectorTypeBreakdown(MVT VT,
369 unsigned &NumIntermediates,
370 MVT &RegisterVT) const {
371 // Figure out the right, legal destination reg to copy into.
372 unsigned NumElts = VT.getVectorNumElements();
373 MVT EltTy = VT.getVectorElementType();
375 unsigned NumVectorRegs = 1;
377 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
378 // could break down into LHS/RHS like LegalizeDAG does.
379 if (!isPowerOf2_32(NumElts)) {
380 NumVectorRegs = NumElts;
384 // Divide the input until we get to a supported size. This will always
385 // end with a scalar if the target doesn't support vectors.
386 while (NumElts > 1 && !isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) {
391 NumIntermediates = NumVectorRegs;
393 MVT NewVT = MVT::getVectorVT(EltTy, NumElts);
394 if (!isTypeLegal(NewVT))
396 IntermediateVT = NewVT;
398 MVT DestVT = getTypeToTransformTo(NewVT);
400 if (DestVT.bitsLT(NewVT)) {
401 // Value is expanded, e.g. i64 -> i16.
402 return NumVectorRegs*(NewVT.getSizeInBits()/DestVT.getSizeInBits());
404 // Otherwise, promotion or legal types use the same number of registers as
405 // the vector decimated to the appropriate level.
406 return NumVectorRegs;
412 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
413 /// function arguments in the caller parameter area. This is the actual
414 /// alignment, not its logarithm.
415 unsigned TargetLowering::getByValTypeAlignment(const Type *Ty) const {
416 return TD->getCallFrameTypeAlignment(Ty);
419 SDOperand TargetLowering::getPICJumpTableRelocBase(SDOperand Table,
420 SelectionDAG &DAG) const {
421 if (usesGlobalOffsetTable())
422 return DAG.getNode(ISD::GLOBAL_OFFSET_TABLE, getPointerTy());
426 //===----------------------------------------------------------------------===//
427 // Optimization Methods
428 //===----------------------------------------------------------------------===//
430 /// ShrinkDemandedConstant - Check to see if the specified operand of the
431 /// specified instruction is a constant integer. If so, check to see if there
432 /// are any bits set in the constant that are not demanded. If so, shrink the
433 /// constant and return true.
434 bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDOperand Op,
435 const APInt &Demanded) {
436 // FIXME: ISD::SELECT, ISD::SELECT_CC
437 switch(Op.getOpcode()) {
442 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1)))
443 if (C->getAPIntValue().intersects(~Demanded)) {
444 MVT VT = Op.getValueType();
445 SDOperand New = DAG.getNode(Op.getOpcode(), VT, Op.getOperand(0),
446 DAG.getConstant(Demanded &
449 return CombineTo(Op, New);
456 /// SimplifyDemandedBits - Look at Op. At this point, we know that only the
457 /// DemandedMask bits of the result of Op are ever used downstream. If we can
458 /// use this information to simplify Op, create a new simplified DAG node and
459 /// return true, returning the original and new nodes in Old and New. Otherwise,
460 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
461 /// the expression (used to simplify the caller). The KnownZero/One bits may
462 /// only be accurate for those bits in the DemandedMask.
463 bool TargetLowering::SimplifyDemandedBits(SDOperand Op,
464 const APInt &DemandedMask,
467 TargetLoweringOpt &TLO,
468 unsigned Depth) const {
469 unsigned BitWidth = DemandedMask.getBitWidth();
470 assert(Op.getValueSizeInBits() == BitWidth &&
471 "Mask size mismatches value type size!");
472 APInt NewMask = DemandedMask;
474 // Don't know anything.
475 KnownZero = KnownOne = APInt(BitWidth, 0);
477 // Other users may use these bits.
478 if (!Op.Val->hasOneUse()) {
480 // If not at the root, Just compute the KnownZero/KnownOne bits to
481 // simplify things downstream.
482 TLO.DAG.ComputeMaskedBits(Op, DemandedMask, KnownZero, KnownOne, Depth);
485 // If this is the root being simplified, allow it to have multiple uses,
486 // just set the NewMask to all bits.
487 NewMask = APInt::getAllOnesValue(BitWidth);
488 } else if (DemandedMask == 0) {
489 // Not demanding any bits from Op.
490 if (Op.getOpcode() != ISD::UNDEF)
491 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::UNDEF, Op.getValueType()));
493 } else if (Depth == 6) { // Limit search depth.
497 APInt KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut;
498 switch (Op.getOpcode()) {
500 // We know all of the bits for a constant!
501 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue() & NewMask;
502 KnownZero = ~KnownOne & NewMask;
503 return false; // Don't fall through, will infinitely loop.
505 // If the RHS is a constant, check to see if the LHS would be zero without
506 // using the bits from the RHS. Below, we use knowledge about the RHS to
507 // simplify the LHS, here we're using information from the LHS to simplify
509 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
510 APInt LHSZero, LHSOne;
511 TLO.DAG.ComputeMaskedBits(Op.getOperand(0), NewMask,
512 LHSZero, LHSOne, Depth+1);
513 // If the LHS already has zeros where RHSC does, this and is dead.
514 if ((LHSZero & NewMask) == (~RHSC->getAPIntValue() & NewMask))
515 return TLO.CombineTo(Op, Op.getOperand(0));
516 // If any of the set bits in the RHS are known zero on the LHS, shrink
518 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & NewMask))
522 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
523 KnownOne, TLO, Depth+1))
525 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
526 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownZero & NewMask,
527 KnownZero2, KnownOne2, TLO, Depth+1))
529 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
531 // If all of the demanded bits are known one on one side, return the other.
532 // These bits cannot contribute to the result of the 'and'.
533 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
534 return TLO.CombineTo(Op, Op.getOperand(0));
535 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
536 return TLO.CombineTo(Op, Op.getOperand(1));
537 // If all of the demanded bits in the inputs are known zeros, return zero.
538 if ((NewMask & (KnownZero|KnownZero2)) == NewMask)
539 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, Op.getValueType()));
540 // If the RHS is a constant, see if we can simplify it.
541 if (TLO.ShrinkDemandedConstant(Op, ~KnownZero2 & NewMask))
544 // Output known-1 bits are only known if set in both the LHS & RHS.
545 KnownOne &= KnownOne2;
546 // Output known-0 are known to be clear if zero in either the LHS | RHS.
547 KnownZero |= KnownZero2;
550 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
551 KnownOne, TLO, Depth+1))
553 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
554 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownOne & NewMask,
555 KnownZero2, KnownOne2, TLO, Depth+1))
557 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
559 // If all of the demanded bits are known zero on one side, return the other.
560 // These bits cannot contribute to the result of the 'or'.
561 if ((NewMask & ~KnownOne2 & KnownZero) == (~KnownOne2 & NewMask))
562 return TLO.CombineTo(Op, Op.getOperand(0));
563 if ((NewMask & ~KnownOne & KnownZero2) == (~KnownOne & NewMask))
564 return TLO.CombineTo(Op, Op.getOperand(1));
565 // If all of the potentially set bits on one side are known to be set on
566 // the other side, just use the 'other' side.
567 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
568 return TLO.CombineTo(Op, Op.getOperand(0));
569 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
570 return TLO.CombineTo(Op, Op.getOperand(1));
571 // If the RHS is a constant, see if we can simplify it.
572 if (TLO.ShrinkDemandedConstant(Op, NewMask))
575 // Output known-0 bits are only known if clear in both the LHS & RHS.
576 KnownZero &= KnownZero2;
577 // Output known-1 are known to be set if set in either the LHS | RHS.
578 KnownOne |= KnownOne2;
581 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
582 KnownOne, TLO, Depth+1))
584 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
585 if (SimplifyDemandedBits(Op.getOperand(0), NewMask, KnownZero2,
586 KnownOne2, TLO, Depth+1))
588 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
590 // If all of the demanded bits are known zero on one side, return the other.
591 // These bits cannot contribute to the result of the 'xor'.
592 if ((KnownZero & NewMask) == NewMask)
593 return TLO.CombineTo(Op, Op.getOperand(0));
594 if ((KnownZero2 & NewMask) == NewMask)
595 return TLO.CombineTo(Op, Op.getOperand(1));
597 // If all of the unknown bits are known to be zero on one side or the other
598 // (but not both) turn this into an *inclusive* or.
599 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
600 if ((NewMask & ~KnownZero & ~KnownZero2) == 0)
601 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, Op.getValueType(),
605 // Output known-0 bits are known if clear or set in both the LHS & RHS.
606 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
607 // Output known-1 are known to be set if set in only one of the LHS, RHS.
608 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
610 // If all of the demanded bits on one side are known, and all of the set
611 // bits on that side are also known to be set on the other side, turn this
612 // into an AND, as we know the bits will be cleared.
613 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
614 if ((NewMask & (KnownZero|KnownOne)) == NewMask) { // all known
615 if ((KnownOne & KnownOne2) == KnownOne) {
616 MVT VT = Op.getValueType();
617 SDOperand ANDC = TLO.DAG.getConstant(~KnownOne & NewMask, VT);
618 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, VT, Op.getOperand(0),
623 // If the RHS is a constant, see if we can simplify it.
624 // for XOR, we prefer to force bits to 1 if they will make a -1.
625 // if we can't force bits, try to shrink constant
626 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
627 APInt Expanded = C->getAPIntValue() | (~NewMask);
628 // if we can expand it to have all bits set, do it
629 if (Expanded.isAllOnesValue()) {
630 if (Expanded != C->getAPIntValue()) {
631 MVT VT = Op.getValueType();
632 SDOperand New = TLO.DAG.getNode(Op.getOpcode(), VT, Op.getOperand(0),
633 TLO.DAG.getConstant(Expanded, VT));
634 return TLO.CombineTo(Op, New);
636 // if it already has all the bits set, nothing to change
637 // but don't shrink either!
638 } else if (TLO.ShrinkDemandedConstant(Op, NewMask)) {
643 KnownZero = KnownZeroOut;
644 KnownOne = KnownOneOut;
647 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero,
648 KnownOne, TLO, Depth+1))
650 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero2,
651 KnownOne2, TLO, Depth+1))
653 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
654 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
656 // If the operands are constants, see if we can simplify them.
657 if (TLO.ShrinkDemandedConstant(Op, NewMask))
660 // Only known if known in both the LHS and RHS.
661 KnownOne &= KnownOne2;
662 KnownZero &= KnownZero2;
665 if (SimplifyDemandedBits(Op.getOperand(3), NewMask, KnownZero,
666 KnownOne, TLO, Depth+1))
668 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero2,
669 KnownOne2, TLO, Depth+1))
671 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
672 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
674 // If the operands are constants, see if we can simplify them.
675 if (TLO.ShrinkDemandedConstant(Op, NewMask))
678 // Only known if known in both the LHS and RHS.
679 KnownOne &= KnownOne2;
680 KnownZero &= KnownZero2;
683 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
684 unsigned ShAmt = SA->getValue();
685 SDOperand InOp = Op.getOperand(0);
687 // If the shift count is an invalid immediate, don't do anything.
688 if (ShAmt >= BitWidth)
691 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a
692 // single shift. We can do this if the bottom bits (which are shifted
693 // out) are never demanded.
694 if (InOp.getOpcode() == ISD::SRL &&
695 isa<ConstantSDNode>(InOp.getOperand(1))) {
696 if (ShAmt && (NewMask & APInt::getLowBitsSet(BitWidth, ShAmt)) == 0) {
697 unsigned C1 = cast<ConstantSDNode>(InOp.getOperand(1))->getValue();
698 unsigned Opc = ISD::SHL;
706 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
707 MVT VT = Op.getValueType();
708 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, VT,
709 InOp.getOperand(0), NewSA));
713 if (SimplifyDemandedBits(Op.getOperand(0), NewMask.lshr(ShAmt),
714 KnownZero, KnownOne, TLO, Depth+1))
716 KnownZero <<= SA->getValue();
717 KnownOne <<= SA->getValue();
718 // low bits known zero.
719 KnownZero |= APInt::getLowBitsSet(BitWidth, SA->getValue());
723 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
724 MVT VT = Op.getValueType();
725 unsigned ShAmt = SA->getValue();
726 unsigned VTSize = VT.getSizeInBits();
727 SDOperand InOp = Op.getOperand(0);
729 // If the shift count is an invalid immediate, don't do anything.
730 if (ShAmt >= BitWidth)
733 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a
734 // single shift. We can do this if the top bits (which are shifted out)
735 // are never demanded.
736 if (InOp.getOpcode() == ISD::SHL &&
737 isa<ConstantSDNode>(InOp.getOperand(1))) {
738 if (ShAmt && (NewMask & APInt::getHighBitsSet(VTSize, ShAmt)) == 0) {
739 unsigned C1 = cast<ConstantSDNode>(InOp.getOperand(1))->getValue();
740 unsigned Opc = ISD::SRL;
748 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
749 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, VT,
750 InOp.getOperand(0), NewSA));
754 // Compute the new bits that are at the top now.
755 if (SimplifyDemandedBits(InOp, (NewMask << ShAmt),
756 KnownZero, KnownOne, TLO, Depth+1))
758 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
759 KnownZero = KnownZero.lshr(ShAmt);
760 KnownOne = KnownOne.lshr(ShAmt);
762 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
763 KnownZero |= HighBits; // High bits known zero.
767 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
768 MVT VT = Op.getValueType();
769 unsigned ShAmt = SA->getValue();
771 // If the shift count is an invalid immediate, don't do anything.
772 if (ShAmt >= BitWidth)
775 APInt InDemandedMask = (NewMask << ShAmt);
777 // If any of the demanded bits are produced by the sign extension, we also
778 // demand the input sign bit.
779 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
780 if (HighBits.intersects(NewMask))
781 InDemandedMask |= APInt::getSignBit(VT.getSizeInBits());
783 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask,
784 KnownZero, KnownOne, TLO, Depth+1))
786 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
787 KnownZero = KnownZero.lshr(ShAmt);
788 KnownOne = KnownOne.lshr(ShAmt);
790 // Handle the sign bit, adjusted to where it is now in the mask.
791 APInt SignBit = APInt::getSignBit(BitWidth).lshr(ShAmt);
793 // If the input sign bit is known to be zero, or if none of the top bits
794 // are demanded, turn this into an unsigned shift right.
795 if (KnownZero.intersects(SignBit) || (HighBits & ~NewMask) == HighBits) {
796 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, VT, Op.getOperand(0),
798 } else if (KnownOne.intersects(SignBit)) { // New bits are known one.
799 KnownOne |= HighBits;
803 case ISD::SIGN_EXTEND_INREG: {
804 MVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
806 // Sign extension. Compute the demanded bits in the result that are not
807 // present in the input.
808 APInt NewBits = APInt::getHighBitsSet(BitWidth,
809 BitWidth - EVT.getSizeInBits()) &
812 // If none of the extended bits are demanded, eliminate the sextinreg.
814 return TLO.CombineTo(Op, Op.getOperand(0));
816 APInt InSignBit = APInt::getSignBit(EVT.getSizeInBits());
817 InSignBit.zext(BitWidth);
818 APInt InputDemandedBits = APInt::getLowBitsSet(BitWidth,
819 EVT.getSizeInBits()) &
822 // Since the sign extended bits are demanded, we know that the sign
824 InputDemandedBits |= InSignBit;
826 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits,
827 KnownZero, KnownOne, TLO, Depth+1))
829 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
831 // If the sign bit of the input is known set or clear, then we know the
832 // top bits of the result.
834 // If the input sign bit is known zero, convert this into a zero extension.
835 if (KnownZero.intersects(InSignBit))
836 return TLO.CombineTo(Op,
837 TLO.DAG.getZeroExtendInReg(Op.getOperand(0), EVT));
839 if (KnownOne.intersects(InSignBit)) { // Input sign bit known set
841 KnownZero &= ~NewBits;
842 } else { // Input sign bit unknown
843 KnownZero &= ~NewBits;
844 KnownOne &= ~NewBits;
848 case ISD::ZERO_EXTEND: {
849 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits();
850 APInt InMask = NewMask;
851 InMask.trunc(OperandBitWidth);
853 // If none of the top bits are demanded, convert this into an any_extend.
855 APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask;
856 if (!NewBits.intersects(NewMask))
857 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND,
861 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
862 KnownZero, KnownOne, TLO, Depth+1))
864 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
865 KnownZero.zext(BitWidth);
866 KnownOne.zext(BitWidth);
867 KnownZero |= NewBits;
870 case ISD::SIGN_EXTEND: {
871 MVT InVT = Op.getOperand(0).getValueType();
872 unsigned InBits = InVT.getSizeInBits();
873 APInt InMask = APInt::getLowBitsSet(BitWidth, InBits);
874 APInt InSignBit = APInt::getBitsSet(BitWidth, InBits - 1, InBits);
875 APInt NewBits = ~InMask & NewMask;
877 // If none of the top bits are demanded, convert this into an any_extend.
879 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND,Op.getValueType(),
882 // Since some of the sign extended bits are demanded, we know that the sign
884 APInt InDemandedBits = InMask & NewMask;
885 InDemandedBits |= InSignBit;
886 InDemandedBits.trunc(InBits);
888 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero,
889 KnownOne, TLO, Depth+1))
891 KnownZero.zext(BitWidth);
892 KnownOne.zext(BitWidth);
894 // If the sign bit is known zero, convert this to a zero extend.
895 if (KnownZero.intersects(InSignBit))
896 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND,
900 // If the sign bit is known one, the top bits match.
901 if (KnownOne.intersects(InSignBit)) {
903 KnownZero &= ~NewBits;
904 } else { // Otherwise, top bits aren't known.
905 KnownOne &= ~NewBits;
906 KnownZero &= ~NewBits;
910 case ISD::ANY_EXTEND: {
911 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits();
912 APInt InMask = NewMask;
913 InMask.trunc(OperandBitWidth);
914 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
915 KnownZero, KnownOne, TLO, Depth+1))
917 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
918 KnownZero.zext(BitWidth);
919 KnownOne.zext(BitWidth);
922 case ISD::TRUNCATE: {
923 // Simplify the input, using demanded bit information, and compute the known
924 // zero/one bits live out.
925 APInt TruncMask = NewMask;
926 TruncMask.zext(Op.getOperand(0).getValueSizeInBits());
927 if (SimplifyDemandedBits(Op.getOperand(0), TruncMask,
928 KnownZero, KnownOne, TLO, Depth+1))
930 KnownZero.trunc(BitWidth);
931 KnownOne.trunc(BitWidth);
933 // If the input is only used by this truncate, see if we can shrink it based
934 // on the known demanded bits.
935 if (Op.getOperand(0).Val->hasOneUse()) {
936 SDOperand In = Op.getOperand(0);
937 unsigned InBitWidth = In.getValueSizeInBits();
938 switch (In.getOpcode()) {
941 // Shrink SRL by a constant if none of the high bits shifted in are
943 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1))){
944 APInt HighBits = APInt::getHighBitsSet(InBitWidth,
945 InBitWidth - BitWidth);
946 HighBits = HighBits.lshr(ShAmt->getValue());
947 HighBits.trunc(BitWidth);
949 if (ShAmt->getValue() < BitWidth && !(HighBits & NewMask)) {
950 // None of the shifted in bits are needed. Add a truncate of the
951 // shift input, then shift it.
952 SDOperand NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE,
955 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL,Op.getValueType(),
956 NewTrunc, In.getOperand(1)));
963 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
966 case ISD::AssertZext: {
967 MVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
968 APInt InMask = APInt::getLowBitsSet(BitWidth,
970 if (SimplifyDemandedBits(Op.getOperand(0), InMask & NewMask,
971 KnownZero, KnownOne, TLO, Depth+1))
973 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
974 KnownZero |= ~InMask & NewMask;
977 case ISD::BIT_CONVERT:
979 // If this is an FP->Int bitcast and if the sign bit is the only thing that
980 // is demanded, turn this into a FGETSIGN.
981 if (NewMask == MVT::getIntegerVTSignBit(Op.getValueType()) &&
982 MVT::isFloatingPoint(Op.getOperand(0).getValueType()) &&
983 !MVT::isVector(Op.getOperand(0).getValueType())) {
984 // Only do this xform if FGETSIGN is valid or if before legalize.
985 if (!TLO.AfterLegalize ||
986 isOperationLegal(ISD::FGETSIGN, Op.getValueType())) {
987 // Make a FGETSIGN + SHL to move the sign bit into the appropriate
988 // place. We expect the SHL to be eliminated by other optimizations.
989 SDOperand Sign = TLO.DAG.getNode(ISD::FGETSIGN, Op.getValueType(),
991 unsigned ShVal = Op.getValueType().getSizeInBits()-1;
992 SDOperand ShAmt = TLO.DAG.getConstant(ShVal, getShiftAmountTy());
993 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, Op.getValueType(),
1000 // Just use ComputeMaskedBits to compute output bits.
1001 TLO.DAG.ComputeMaskedBits(Op, NewMask, KnownZero, KnownOne, Depth);
1005 // If we know the value of all of the demanded bits, return this as a
1007 if ((NewMask & (KnownZero|KnownOne)) == NewMask)
1008 return TLO.CombineTo(Op, TLO.DAG.getConstant(KnownOne, Op.getValueType()));
1013 /// computeMaskedBitsForTargetNode - Determine which of the bits specified
1014 /// in Mask are known to be either zero or one and return them in the
1015 /// KnownZero/KnownOne bitsets.
1016 void TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
1020 const SelectionDAG &DAG,
1021 unsigned Depth) const {
1022 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1023 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1024 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1025 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1026 "Should use MaskedValueIsZero if you don't know whether Op"
1027 " is a target node!");
1028 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
1031 /// ComputeNumSignBitsForTargetNode - This method can be implemented by
1032 /// targets that want to expose additional information about sign bits to the
1034 unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDOperand Op,
1035 unsigned Depth) const {
1036 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1037 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1038 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1039 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1040 "Should use ComputeNumSignBits if you don't know whether Op"
1041 " is a target node!");
1046 /// SimplifySetCC - Try to simplify a setcc built with the specified operands
1047 /// and cc. If it is unable to simplify it, return a null SDOperand.
1049 TargetLowering::SimplifySetCC(MVT VT, SDOperand N0, SDOperand N1,
1050 ISD::CondCode Cond, bool foldBooleans,
1051 DAGCombinerInfo &DCI) const {
1052 SelectionDAG &DAG = DCI.DAG;
1054 // These setcc operations always fold.
1058 case ISD::SETFALSE2: return DAG.getConstant(0, VT);
1060 case ISD::SETTRUE2: return DAG.getConstant(1, VT);
1063 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val)) {
1064 const APInt &C1 = N1C->getAPIntValue();
1065 if (isa<ConstantSDNode>(N0.Val)) {
1066 return DAG.FoldSetCC(VT, N0, N1, Cond);
1068 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an
1069 // equality comparison, then we're just comparing whether X itself is
1071 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) &&
1072 N0.getOperand(0).getOpcode() == ISD::CTLZ &&
1073 N0.getOperand(1).getOpcode() == ISD::Constant) {
1074 unsigned ShAmt = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1075 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1076 ShAmt == Log2_32(N0.getValueType().getSizeInBits())) {
1077 if ((C1 == 0) == (Cond == ISD::SETEQ)) {
1078 // (srl (ctlz x), 5) == 0 -> X != 0
1079 // (srl (ctlz x), 5) != 1 -> X != 0
1082 // (srl (ctlz x), 5) != 0 -> X == 0
1083 // (srl (ctlz x), 5) == 1 -> X == 0
1086 SDOperand Zero = DAG.getConstant(0, N0.getValueType());
1087 return DAG.getSetCC(VT, N0.getOperand(0).getOperand(0),
1092 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
1093 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
1094 unsigned InSize = N0.getOperand(0).getValueType().getSizeInBits();
1096 // If the comparison constant has bits in the upper part, the
1097 // zero-extended value could never match.
1098 if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(),
1099 C1.getBitWidth() - InSize))) {
1103 case ISD::SETEQ: return DAG.getConstant(0, VT);
1106 case ISD::SETNE: return DAG.getConstant(1, VT);
1109 // True if the sign bit of C1 is set.
1110 return DAG.getConstant(C1.isNegative(), VT);
1113 // True if the sign bit of C1 isn't set.
1114 return DAG.getConstant(C1.isNonNegative(), VT);
1120 // Otherwise, we can perform the comparison with the low bits.
1128 return DAG.getSetCC(VT, N0.getOperand(0),
1129 DAG.getConstant(APInt(C1).trunc(InSize),
1130 N0.getOperand(0).getValueType()),
1133 break; // todo, be more careful with signed comparisons
1135 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
1136 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1137 MVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
1138 unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits();
1139 MVT ExtDstTy = N0.getValueType();
1140 unsigned ExtDstTyBits = ExtDstTy.getSizeInBits();
1142 // If the extended part has any inconsistent bits, it cannot ever
1143 // compare equal. In other words, they have to be all ones or all
1146 APInt::getHighBitsSet(ExtDstTyBits, ExtDstTyBits - ExtSrcTyBits);
1147 if ((C1 & ExtBits) != 0 && (C1 & ExtBits) != ExtBits)
1148 return DAG.getConstant(Cond == ISD::SETNE, VT);
1151 MVT Op0Ty = N0.getOperand(0).getValueType();
1152 if (Op0Ty == ExtSrcTy) {
1153 ZextOp = N0.getOperand(0);
1155 APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits);
1156 ZextOp = DAG.getNode(ISD::AND, Op0Ty, N0.getOperand(0),
1157 DAG.getConstant(Imm, Op0Ty));
1159 if (!DCI.isCalledByLegalizer())
1160 DCI.AddToWorklist(ZextOp.Val);
1161 // Otherwise, make this a use of a zext.
1162 return DAG.getSetCC(VT, ZextOp,
1163 DAG.getConstant(C1 & APInt::getLowBitsSet(
1168 } else if ((N1C->isNullValue() || N1C->getAPIntValue() == 1) &&
1169 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1171 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC
1172 if (N0.getOpcode() == ISD::SETCC) {
1173 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getValue() != 1);
1177 // Invert the condition.
1178 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
1179 CC = ISD::getSetCCInverse(CC,
1180 N0.getOperand(0).getValueType().isInteger());
1181 return DAG.getSetCC(VT, N0.getOperand(0), N0.getOperand(1), CC);
1184 if ((N0.getOpcode() == ISD::XOR ||
1185 (N0.getOpcode() == ISD::AND &&
1186 N0.getOperand(0).getOpcode() == ISD::XOR &&
1187 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
1188 isa<ConstantSDNode>(N0.getOperand(1)) &&
1189 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue() == 1) {
1190 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We
1191 // can only do this if the top bits are known zero.
1192 unsigned BitWidth = N0.getValueSizeInBits();
1193 if (DAG.MaskedValueIsZero(N0,
1194 APInt::getHighBitsSet(BitWidth,
1196 // Okay, get the un-inverted input value.
1198 if (N0.getOpcode() == ISD::XOR)
1199 Val = N0.getOperand(0);
1201 assert(N0.getOpcode() == ISD::AND &&
1202 N0.getOperand(0).getOpcode() == ISD::XOR);
1203 // ((X^1)&1)^1 -> X & 1
1204 Val = DAG.getNode(ISD::AND, N0.getValueType(),
1205 N0.getOperand(0).getOperand(0),
1208 return DAG.getSetCC(VT, Val, N1,
1209 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
1214 APInt MinVal, MaxVal;
1215 unsigned OperandBitSize = N1C->getValueType(0).getSizeInBits();
1216 if (ISD::isSignedIntSetCC(Cond)) {
1217 MinVal = APInt::getSignedMinValue(OperandBitSize);
1218 MaxVal = APInt::getSignedMaxValue(OperandBitSize);
1220 MinVal = APInt::getMinValue(OperandBitSize);
1221 MaxVal = APInt::getMaxValue(OperandBitSize);
1224 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
1225 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
1226 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true
1227 // X >= C0 --> X > (C0-1)
1228 return DAG.getSetCC(VT, N0, DAG.getConstant(C1-1, N1.getValueType()),
1229 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
1232 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
1233 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true
1234 // X <= C0 --> X < (C0+1)
1235 return DAG.getSetCC(VT, N0, DAG.getConstant(C1+1, N1.getValueType()),
1236 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
1239 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
1240 return DAG.getConstant(0, VT); // X < MIN --> false
1241 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal)
1242 return DAG.getConstant(1, VT); // X >= MIN --> true
1243 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal)
1244 return DAG.getConstant(0, VT); // X > MAX --> false
1245 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal)
1246 return DAG.getConstant(1, VT); // X <= MAX --> true
1248 // Canonicalize setgt X, Min --> setne X, Min
1249 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
1250 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
1251 // Canonicalize setlt X, Max --> setne X, Max
1252 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
1253 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
1255 // If we have setult X, 1, turn it into seteq X, 0
1256 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
1257 return DAG.getSetCC(VT, N0, DAG.getConstant(MinVal, N0.getValueType()),
1259 // If we have setugt X, Max-1, turn it into seteq X, Max
1260 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
1261 return DAG.getSetCC(VT, N0, DAG.getConstant(MaxVal, N0.getValueType()),
1264 // If we have "setcc X, C0", check to see if we can shrink the immediate
1267 // SETUGT X, SINTMAX -> SETLT X, 0
1268 if (Cond == ISD::SETUGT && OperandBitSize != 1 &&
1269 C1 == (~0ULL >> (65-OperandBitSize)))
1270 return DAG.getSetCC(VT, N0, DAG.getConstant(0, N1.getValueType()),
1273 // FIXME: Implement the rest of these.
1275 // Fold bit comparisons when we can.
1276 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1277 VT == N0.getValueType() && N0.getOpcode() == ISD::AND)
1278 if (ConstantSDNode *AndRHS =
1279 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1280 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
1281 // Perform the xform if the AND RHS is a single bit.
1282 if (isPowerOf2_64(AndRHS->getValue())) {
1283 return DAG.getNode(ISD::SRL, VT, N0,
1284 DAG.getConstant(Log2_64(AndRHS->getValue()),
1285 getShiftAmountTy()));
1287 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getValue()) {
1288 // (X & 8) == 8 --> (X & 8) >> 3
1289 // Perform the xform if C1 is a single bit.
1290 if (C1.isPowerOf2()) {
1291 return DAG.getNode(ISD::SRL, VT, N0,
1292 DAG.getConstant(C1.logBase2(), getShiftAmountTy()));
1297 } else if (isa<ConstantSDNode>(N0.Val)) {
1298 // Ensure that the constant occurs on the RHS.
1299 return DAG.getSetCC(VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
1302 if (isa<ConstantFPSDNode>(N0.Val)) {
1303 // Constant fold or commute setcc.
1304 SDOperand O = DAG.FoldSetCC(VT, N0, N1, Cond);
1305 if (O.Val) return O;
1306 } else if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1.Val)) {
1307 // If the RHS of an FP comparison is a constant, simplify it away in
1309 if (CFP->getValueAPF().isNaN()) {
1310 // If an operand is known to be a nan, we can fold it.
1311 switch (ISD::getUnorderedFlavor(Cond)) {
1312 default: assert(0 && "Unknown flavor!");
1313 case 0: // Known false.
1314 return DAG.getConstant(0, VT);
1315 case 1: // Known true.
1316 return DAG.getConstant(1, VT);
1317 case 2: // Undefined.
1318 return DAG.getNode(ISD::UNDEF, VT);
1322 // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the
1323 // constant if knowing that the operand is non-nan is enough. We prefer to
1324 // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to
1326 if (Cond == ISD::SETO || Cond == ISD::SETUO)
1327 return DAG.getSetCC(VT, N0, N0, Cond);
1331 // We can always fold X == X for integer setcc's.
1332 if (N0.getValueType().isInteger())
1333 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
1334 unsigned UOF = ISD::getUnorderedFlavor(Cond);
1335 if (UOF == 2) // FP operators that are undefined on NaNs.
1336 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
1337 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
1338 return DAG.getConstant(UOF, VT);
1339 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
1340 // if it is not already.
1341 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
1342 if (NewCond != Cond)
1343 return DAG.getSetCC(VT, N0, N1, NewCond);
1346 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1347 N0.getValueType().isInteger()) {
1348 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
1349 N0.getOpcode() == ISD::XOR) {
1350 // Simplify (X+Y) == (X+Z) --> Y == Z
1351 if (N0.getOpcode() == N1.getOpcode()) {
1352 if (N0.getOperand(0) == N1.getOperand(0))
1353 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(1), Cond);
1354 if (N0.getOperand(1) == N1.getOperand(1))
1355 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(0), Cond);
1356 if (DAG.isCommutativeBinOp(N0.getOpcode())) {
1357 // If X op Y == Y op X, try other combinations.
1358 if (N0.getOperand(0) == N1.getOperand(1))
1359 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(0), Cond);
1360 if (N0.getOperand(1) == N1.getOperand(0))
1361 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(1), Cond);
1365 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) {
1366 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1367 // Turn (X+C1) == C2 --> X == C2-C1
1368 if (N0.getOpcode() == ISD::ADD && N0.Val->hasOneUse()) {
1369 return DAG.getSetCC(VT, N0.getOperand(0),
1370 DAG.getConstant(RHSC->getValue()-LHSR->getValue(),
1371 N0.getValueType()), Cond);
1374 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
1375 if (N0.getOpcode() == ISD::XOR)
1376 // If we know that all of the inverted bits are zero, don't bother
1377 // performing the inversion.
1378 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue()))
1380 DAG.getSetCC(VT, N0.getOperand(0),
1381 DAG.getConstant(LHSR->getAPIntValue() ^
1382 RHSC->getAPIntValue(),
1387 // Turn (C1-X) == C2 --> X == C1-C2
1388 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
1389 if (N0.getOpcode() == ISD::SUB && N0.Val->hasOneUse()) {
1391 DAG.getSetCC(VT, N0.getOperand(1),
1392 DAG.getConstant(SUBC->getAPIntValue() -
1393 RHSC->getAPIntValue(),
1400 // Simplify (X+Z) == X --> Z == 0
1401 if (N0.getOperand(0) == N1)
1402 return DAG.getSetCC(VT, N0.getOperand(1),
1403 DAG.getConstant(0, N0.getValueType()), Cond);
1404 if (N0.getOperand(1) == N1) {
1405 if (DAG.isCommutativeBinOp(N0.getOpcode()))
1406 return DAG.getSetCC(VT, N0.getOperand(0),
1407 DAG.getConstant(0, N0.getValueType()), Cond);
1408 else if (N0.Val->hasOneUse()) {
1409 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
1410 // (Z-X) == X --> Z == X<<1
1411 SDOperand SH = DAG.getNode(ISD::SHL, N1.getValueType(),
1413 DAG.getConstant(1, getShiftAmountTy()));
1414 if (!DCI.isCalledByLegalizer())
1415 DCI.AddToWorklist(SH.Val);
1416 return DAG.getSetCC(VT, N0.getOperand(0), SH, Cond);
1421 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
1422 N1.getOpcode() == ISD::XOR) {
1423 // Simplify X == (X+Z) --> Z == 0
1424 if (N1.getOperand(0) == N0) {
1425 return DAG.getSetCC(VT, N1.getOperand(1),
1426 DAG.getConstant(0, N1.getValueType()), Cond);
1427 } else if (N1.getOperand(1) == N0) {
1428 if (DAG.isCommutativeBinOp(N1.getOpcode())) {
1429 return DAG.getSetCC(VT, N1.getOperand(0),
1430 DAG.getConstant(0, N1.getValueType()), Cond);
1431 } else if (N1.Val->hasOneUse()) {
1432 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
1433 // X == (Z-X) --> X<<1 == Z
1434 SDOperand SH = DAG.getNode(ISD::SHL, N1.getValueType(), N0,
1435 DAG.getConstant(1, getShiftAmountTy()));
1436 if (!DCI.isCalledByLegalizer())
1437 DCI.AddToWorklist(SH.Val);
1438 return DAG.getSetCC(VT, SH, N1.getOperand(0), Cond);
1444 // Fold away ALL boolean setcc's.
1446 if (N0.getValueType() == MVT::i1 && foldBooleans) {
1448 default: assert(0 && "Unknown integer setcc!");
1449 case ISD::SETEQ: // X == Y -> (X^Y)^1
1450 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
1451 N0 = DAG.getNode(ISD::XOR, MVT::i1, Temp, DAG.getConstant(1, MVT::i1));
1452 if (!DCI.isCalledByLegalizer())
1453 DCI.AddToWorklist(Temp.Val);
1455 case ISD::SETNE: // X != Y --> (X^Y)
1456 N0 = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
1458 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> X^1 & Y
1459 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> X^1 & Y
1460 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
1461 N0 = DAG.getNode(ISD::AND, MVT::i1, N1, Temp);
1462 if (!DCI.isCalledByLegalizer())
1463 DCI.AddToWorklist(Temp.Val);
1465 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> Y^1 & X
1466 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> Y^1 & X
1467 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
1468 N0 = DAG.getNode(ISD::AND, MVT::i1, N0, Temp);
1469 if (!DCI.isCalledByLegalizer())
1470 DCI.AddToWorklist(Temp.Val);
1472 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> X^1 | Y
1473 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> X^1 | Y
1474 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
1475 N0 = DAG.getNode(ISD::OR, MVT::i1, N1, Temp);
1476 if (!DCI.isCalledByLegalizer())
1477 DCI.AddToWorklist(Temp.Val);
1479 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> Y^1 | X
1480 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> Y^1 | X
1481 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
1482 N0 = DAG.getNode(ISD::OR, MVT::i1, N0, Temp);
1485 if (VT != MVT::i1) {
1486 if (!DCI.isCalledByLegalizer())
1487 DCI.AddToWorklist(N0.Val);
1488 // FIXME: If running after legalize, we probably can't do this.
1489 N0 = DAG.getNode(ISD::ZERO_EXTEND, VT, N0);
1494 // Could not fold it.
1498 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
1499 /// node is a GlobalAddress + offset.
1500 bool TargetLowering::isGAPlusOffset(SDNode *N, GlobalValue* &GA,
1501 int64_t &Offset) const {
1502 if (isa<GlobalAddressSDNode>(N)) {
1503 GlobalAddressSDNode *GASD = cast<GlobalAddressSDNode>(N);
1504 GA = GASD->getGlobal();
1505 Offset += GASD->getOffset();
1509 if (N->getOpcode() == ISD::ADD) {
1510 SDOperand N1 = N->getOperand(0);
1511 SDOperand N2 = N->getOperand(1);
1512 if (isGAPlusOffset(N1.Val, GA, Offset)) {
1513 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
1515 Offset += V->getSignExtended();
1518 } else if (isGAPlusOffset(N2.Val, GA, Offset)) {
1519 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
1521 Offset += V->getSignExtended();
1530 /// isConsecutiveLoad - Return true if LD (which must be a LoadSDNode) is
1531 /// loading 'Bytes' bytes from a location that is 'Dist' units away from the
1532 /// location that the 'Base' load is loading from.
1533 bool TargetLowering::isConsecutiveLoad(SDNode *LD, SDNode *Base,
1534 unsigned Bytes, int Dist,
1535 const MachineFrameInfo *MFI) const {
1536 if (LD->getOperand(0).Val != Base->getOperand(0).Val)
1538 MVT VT = LD->getValueType(0);
1539 if (VT.getSizeInBits() / 8 != Bytes)
1542 SDOperand Loc = LD->getOperand(1);
1543 SDOperand BaseLoc = Base->getOperand(1);
1544 if (Loc.getOpcode() == ISD::FrameIndex) {
1545 if (BaseLoc.getOpcode() != ISD::FrameIndex)
1547 int FI = cast<FrameIndexSDNode>(Loc)->getIndex();
1548 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
1549 int FS = MFI->getObjectSize(FI);
1550 int BFS = MFI->getObjectSize(BFI);
1551 if (FS != BFS || FS != (int)Bytes) return false;
1552 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Bytes);
1555 GlobalValue *GV1 = NULL;
1556 GlobalValue *GV2 = NULL;
1557 int64_t Offset1 = 0;
1558 int64_t Offset2 = 0;
1559 bool isGA1 = isGAPlusOffset(Loc.Val, GV1, Offset1);
1560 bool isGA2 = isGAPlusOffset(BaseLoc.Val, GV2, Offset2);
1561 if (isGA1 && isGA2 && GV1 == GV2)
1562 return Offset1 == (Offset2 + Dist*Bytes);
1567 SDOperand TargetLowering::
1568 PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
1569 // Default implementation: no optimization.
1573 //===----------------------------------------------------------------------===//
1574 // Inline Assembler Implementation Methods
1575 //===----------------------------------------------------------------------===//
1578 TargetLowering::ConstraintType
1579 TargetLowering::getConstraintType(const std::string &Constraint) const {
1580 // FIXME: lots more standard ones to handle.
1581 if (Constraint.size() == 1) {
1582 switch (Constraint[0]) {
1584 case 'r': return C_RegisterClass;
1586 case 'o': // offsetable
1587 case 'V': // not offsetable
1589 case 'i': // Simple Integer or Relocatable Constant
1590 case 'n': // Simple Integer
1591 case 's': // Relocatable Constant
1592 case 'X': // Allow ANY value.
1593 case 'I': // Target registers.
1605 if (Constraint.size() > 1 && Constraint[0] == '{' &&
1606 Constraint[Constraint.size()-1] == '}')
1611 /// LowerXConstraint - try to replace an X constraint, which matches anything,
1612 /// with another that has more specific requirements based on the type of the
1613 /// corresponding operand.
1614 const char *TargetLowering::LowerXConstraint(MVT ConstraintVT) const{
1615 if (ConstraintVT.isInteger())
1617 if (ConstraintVT.isFloatingPoint())
1618 return "f"; // works for many targets
1622 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
1623 /// vector. If it is invalid, don't add anything to Ops.
1624 void TargetLowering::LowerAsmOperandForConstraint(SDOperand Op,
1625 char ConstraintLetter,
1626 std::vector<SDOperand> &Ops,
1627 SelectionDAG &DAG) const {
1628 switch (ConstraintLetter) {
1630 case 'X': // Allows any operand; labels (basic block) use this.
1631 if (Op.getOpcode() == ISD::BasicBlock) {
1636 case 'i': // Simple Integer or Relocatable Constant
1637 case 'n': // Simple Integer
1638 case 's': { // Relocatable Constant
1639 // These operands are interested in values of the form (GV+C), where C may
1640 // be folded in as an offset of GV, or it may be explicitly added. Also, it
1641 // is possible and fine if either GV or C are missing.
1642 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
1643 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
1645 // If we have "(add GV, C)", pull out GV/C
1646 if (Op.getOpcode() == ISD::ADD) {
1647 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
1648 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
1649 if (C == 0 || GA == 0) {
1650 C = dyn_cast<ConstantSDNode>(Op.getOperand(0));
1651 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1));
1653 if (C == 0 || GA == 0)
1657 // If we find a valid operand, map to the TargetXXX version so that the
1658 // value itself doesn't get selected.
1659 if (GA) { // Either &GV or &GV+C
1660 if (ConstraintLetter != 'n') {
1661 int64_t Offs = GA->getOffset();
1662 if (C) Offs += C->getValue();
1663 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(),
1664 Op.getValueType(), Offs));
1668 if (C) { // just C, no GV.
1669 // Simple constants are not allowed for 's'.
1670 if (ConstraintLetter != 's') {
1671 Ops.push_back(DAG.getTargetConstant(C->getValue(), Op.getValueType()));
1680 std::vector<unsigned> TargetLowering::
1681 getRegClassForInlineAsmConstraint(const std::string &Constraint,
1683 return std::vector<unsigned>();
1687 std::pair<unsigned, const TargetRegisterClass*> TargetLowering::
1688 getRegForInlineAsmConstraint(const std::string &Constraint,
1690 if (Constraint[0] != '{')
1691 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
1692 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?");
1694 // Remove the braces from around the name.
1695 std::string RegName(Constraint.begin()+1, Constraint.end()-1);
1697 // Figure out which register class contains this reg.
1698 const TargetRegisterInfo *RI = TM.getRegisterInfo();
1699 for (TargetRegisterInfo::regclass_iterator RCI = RI->regclass_begin(),
1700 E = RI->regclass_end(); RCI != E; ++RCI) {
1701 const TargetRegisterClass *RC = *RCI;
1703 // If none of the the value types for this register class are valid, we
1704 // can't use it. For example, 64-bit reg classes on 32-bit targets.
1705 bool isLegal = false;
1706 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
1708 if (isTypeLegal(*I)) {
1714 if (!isLegal) continue;
1716 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end();
1718 if (StringsEqualNoCase(RegName, RI->get(*I).AsmName))
1719 return std::make_pair(*I, RC);
1723 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
1726 //===----------------------------------------------------------------------===//
1727 // Constraint Selection.
1729 /// getConstraintGenerality - Return an integer indicating how general CT
1731 static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) {
1733 default: assert(0 && "Unknown constraint type!");
1734 case TargetLowering::C_Other:
1735 case TargetLowering::C_Unknown:
1737 case TargetLowering::C_Register:
1739 case TargetLowering::C_RegisterClass:
1741 case TargetLowering::C_Memory:
1746 /// ChooseConstraint - If there are multiple different constraints that we
1747 /// could pick for this operand (e.g. "imr") try to pick the 'best' one.
1748 /// This is somewhat tricky: constraints fall into four classes:
1749 /// Other -> immediates and magic values
1750 /// Register -> one specific register
1751 /// RegisterClass -> a group of regs
1752 /// Memory -> memory
1753 /// Ideally, we would pick the most specific constraint possible: if we have
1754 /// something that fits into a register, we would pick it. The problem here
1755 /// is that if we have something that could either be in a register or in
1756 /// memory that use of the register could cause selection of *other*
1757 /// operands to fail: they might only succeed if we pick memory. Because of
1758 /// this the heuristic we use is:
1760 /// 1) If there is an 'other' constraint, and if the operand is valid for
1761 /// that constraint, use it. This makes us take advantage of 'i'
1762 /// constraints when available.
1763 /// 2) Otherwise, pick the most general constraint present. This prefers
1764 /// 'm' over 'r', for example.
1766 static void ChooseConstraint(TargetLowering::AsmOperandInfo &OpInfo,
1767 const TargetLowering &TLI,
1768 SDOperand Op, SelectionDAG *DAG) {
1769 assert(OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options");
1770 unsigned BestIdx = 0;
1771 TargetLowering::ConstraintType BestType = TargetLowering::C_Unknown;
1772 int BestGenerality = -1;
1774 // Loop over the options, keeping track of the most general one.
1775 for (unsigned i = 0, e = OpInfo.Codes.size(); i != e; ++i) {
1776 TargetLowering::ConstraintType CType =
1777 TLI.getConstraintType(OpInfo.Codes[i]);
1779 // If this is an 'other' constraint, see if the operand is valid for it.
1780 // For example, on X86 we might have an 'rI' constraint. If the operand
1781 // is an integer in the range [0..31] we want to use I (saving a load
1782 // of a register), otherwise we must use 'r'.
1783 if (CType == TargetLowering::C_Other && Op.Val) {
1784 assert(OpInfo.Codes[i].size() == 1 &&
1785 "Unhandled multi-letter 'other' constraint");
1786 std::vector<SDOperand> ResultOps;
1787 TLI.LowerAsmOperandForConstraint(Op, OpInfo.Codes[i][0],
1789 if (!ResultOps.empty()) {
1796 // This constraint letter is more general than the previous one, use it.
1797 int Generality = getConstraintGenerality(CType);
1798 if (Generality > BestGenerality) {
1801 BestGenerality = Generality;
1805 OpInfo.ConstraintCode = OpInfo.Codes[BestIdx];
1806 OpInfo.ConstraintType = BestType;
1809 /// ComputeConstraintToUse - Determines the constraint code and constraint
1810 /// type to use for the specific AsmOperandInfo, setting
1811 /// OpInfo.ConstraintCode and OpInfo.ConstraintType.
1812 void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo,
1814 SelectionDAG *DAG) const {
1815 assert(!OpInfo.Codes.empty() && "Must have at least one constraint");
1817 // Single-letter constraints ('r') are very common.
1818 if (OpInfo.Codes.size() == 1) {
1819 OpInfo.ConstraintCode = OpInfo.Codes[0];
1820 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
1822 ChooseConstraint(OpInfo, *this, Op, DAG);
1825 // 'X' matches anything.
1826 if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) {
1827 // Labels and constants are handled elsewhere ('X' is the only thing
1828 // that matches labels).
1829 if (isa<BasicBlock>(OpInfo.CallOperandVal) ||
1830 isa<ConstantInt>(OpInfo.CallOperandVal))
1833 // Otherwise, try to resolve it to something we know about by looking at
1834 // the actual operand type.
1835 if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) {
1836 OpInfo.ConstraintCode = Repl;
1837 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
1842 //===----------------------------------------------------------------------===//
1843 // Loop Strength Reduction hooks
1844 //===----------------------------------------------------------------------===//
1846 /// isLegalAddressingMode - Return true if the addressing mode represented
1847 /// by AM is legal for this target, for a load/store of the specified type.
1848 bool TargetLowering::isLegalAddressingMode(const AddrMode &AM,
1849 const Type *Ty) const {
1850 // The default implementation of this implements a conservative RISCy, r+r and
1853 // Allows a sign-extended 16-bit immediate field.
1854 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
1857 // No global is ever allowed as a base.
1861 // Only support r+r,
1863 case 0: // "r+i" or just "i", depending on HasBaseReg.
1866 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
1868 // Otherwise we have r+r or r+i.
1871 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
1873 // Allow 2*r as r+r.
1880 // Magic for divide replacement
1883 int64_t m; // magic number
1884 int64_t s; // shift amount
1888 uint64_t m; // magic number
1889 int64_t a; // add indicator
1890 int64_t s; // shift amount
1893 /// magic - calculate the magic numbers required to codegen an integer sdiv as
1894 /// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
1896 static ms magic32(int32_t d) {
1898 uint32_t ad, anc, delta, q1, r1, q2, r2, t;
1899 const uint32_t two31 = 0x80000000U;
1903 t = two31 + ((uint32_t)d >> 31);
1904 anc = t - 1 - t%ad; // absolute value of nc
1905 p = 31; // initialize p
1906 q1 = two31/anc; // initialize q1 = 2p/abs(nc)
1907 r1 = two31 - q1*anc; // initialize r1 = rem(2p,abs(nc))
1908 q2 = two31/ad; // initialize q2 = 2p/abs(d)
1909 r2 = two31 - q2*ad; // initialize r2 = rem(2p,abs(d))
1912 q1 = 2*q1; // update q1 = 2p/abs(nc)
1913 r1 = 2*r1; // update r1 = rem(2p/abs(nc))
1914 if (r1 >= anc) { // must be unsigned comparison
1918 q2 = 2*q2; // update q2 = 2p/abs(d)
1919 r2 = 2*r2; // update r2 = rem(2p/abs(d))
1920 if (r2 >= ad) { // must be unsigned comparison
1925 } while (q1 < delta || (q1 == delta && r1 == 0));
1927 mag.m = (int32_t)(q2 + 1); // make sure to sign extend
1928 if (d < 0) mag.m = -mag.m; // resulting magic number
1929 mag.s = p - 32; // resulting shift
1933 /// magicu - calculate the magic numbers required to codegen an integer udiv as
1934 /// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
1935 static mu magicu32(uint32_t d) {
1937 uint32_t nc, delta, q1, r1, q2, r2;
1939 magu.a = 0; // initialize "add" indicator
1941 p = 31; // initialize p
1942 q1 = 0x80000000/nc; // initialize q1 = 2p/nc
1943 r1 = 0x80000000 - q1*nc; // initialize r1 = rem(2p,nc)
1944 q2 = 0x7FFFFFFF/d; // initialize q2 = (2p-1)/d
1945 r2 = 0x7FFFFFFF - q2*d; // initialize r2 = rem((2p-1),d)
1948 if (r1 >= nc - r1 ) {
1949 q1 = 2*q1 + 1; // update q1
1950 r1 = 2*r1 - nc; // update r1
1953 q1 = 2*q1; // update q1
1954 r1 = 2*r1; // update r1
1956 if (r2 + 1 >= d - r2) {
1957 if (q2 >= 0x7FFFFFFF) magu.a = 1;
1958 q2 = 2*q2 + 1; // update q2
1959 r2 = 2*r2 + 1 - d; // update r2
1962 if (q2 >= 0x80000000) magu.a = 1;
1963 q2 = 2*q2; // update q2
1964 r2 = 2*r2 + 1; // update r2
1967 } while (p < 64 && (q1 < delta || (q1 == delta && r1 == 0)));
1968 magu.m = q2 + 1; // resulting magic number
1969 magu.s = p - 32; // resulting shift
1973 /// magic - calculate the magic numbers required to codegen an integer sdiv as
1974 /// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
1976 static ms magic64(int64_t d) {
1978 uint64_t ad, anc, delta, q1, r1, q2, r2, t;
1979 const uint64_t two63 = 9223372036854775808ULL; // 2^63
1982 ad = d >= 0 ? d : -d;
1983 t = two63 + ((uint64_t)d >> 63);
1984 anc = t - 1 - t%ad; // absolute value of nc
1985 p = 63; // initialize p
1986 q1 = two63/anc; // initialize q1 = 2p/abs(nc)
1987 r1 = two63 - q1*anc; // initialize r1 = rem(2p,abs(nc))
1988 q2 = two63/ad; // initialize q2 = 2p/abs(d)
1989 r2 = two63 - q2*ad; // initialize r2 = rem(2p,abs(d))
1992 q1 = 2*q1; // update q1 = 2p/abs(nc)
1993 r1 = 2*r1; // update r1 = rem(2p/abs(nc))
1994 if (r1 >= anc) { // must be unsigned comparison
1998 q2 = 2*q2; // update q2 = 2p/abs(d)
1999 r2 = 2*r2; // update r2 = rem(2p/abs(d))
2000 if (r2 >= ad) { // must be unsigned comparison
2005 } while (q1 < delta || (q1 == delta && r1 == 0));
2008 if (d < 0) mag.m = -mag.m; // resulting magic number
2009 mag.s = p - 64; // resulting shift
2013 /// magicu - calculate the magic numbers required to codegen an integer udiv as
2014 /// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
2015 static mu magicu64(uint64_t d)
2018 uint64_t nc, delta, q1, r1, q2, r2;
2020 magu.a = 0; // initialize "add" indicator
2022 p = 63; // initialize p
2023 q1 = 0x8000000000000000ull/nc; // initialize q1 = 2p/nc
2024 r1 = 0x8000000000000000ull - q1*nc; // initialize r1 = rem(2p,nc)
2025 q2 = 0x7FFFFFFFFFFFFFFFull/d; // initialize q2 = (2p-1)/d
2026 r2 = 0x7FFFFFFFFFFFFFFFull - q2*d; // initialize r2 = rem((2p-1),d)
2029 if (r1 >= nc - r1 ) {
2030 q1 = 2*q1 + 1; // update q1
2031 r1 = 2*r1 - nc; // update r1
2034 q1 = 2*q1; // update q1
2035 r1 = 2*r1; // update r1
2037 if (r2 + 1 >= d - r2) {
2038 if (q2 >= 0x7FFFFFFFFFFFFFFFull) magu.a = 1;
2039 q2 = 2*q2 + 1; // update q2
2040 r2 = 2*r2 + 1 - d; // update r2
2043 if (q2 >= 0x8000000000000000ull) magu.a = 1;
2044 q2 = 2*q2; // update q2
2045 r2 = 2*r2 + 1; // update r2
2048 } while (p < 128 && (q1 < delta || (q1 == delta && r1 == 0)));
2049 magu.m = q2 + 1; // resulting magic number
2050 magu.s = p - 64; // resulting shift
2054 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
2055 /// return a DAG expression to select that will generate the same value by
2056 /// multiplying by a magic number. See:
2057 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
2058 SDOperand TargetLowering::BuildSDIV(SDNode *N, SelectionDAG &DAG,
2059 std::vector<SDNode*>* Created) const {
2060 MVT VT = N->getValueType(0);
2062 // Check to see if we can do this.
2063 if (!isTypeLegal(VT) || (VT != MVT::i32 && VT != MVT::i64))
2064 return SDOperand(); // BuildSDIV only operates on i32 or i64
2066 int64_t d = cast<ConstantSDNode>(N->getOperand(1))->getSignExtended();
2067 ms magics = (VT == MVT::i32) ? magic32(d) : magic64(d);
2069 // Multiply the numerator (operand 0) by the magic value
2071 if (isOperationLegal(ISD::MULHS, VT))
2072 Q = DAG.getNode(ISD::MULHS, VT, N->getOperand(0),
2073 DAG.getConstant(magics.m, VT));
2074 else if (isOperationLegal(ISD::SMUL_LOHI, VT))
2075 Q = SDOperand(DAG.getNode(ISD::SMUL_LOHI, DAG.getVTList(VT, VT),
2077 DAG.getConstant(magics.m, VT)).Val, 1);
2079 return SDOperand(); // No mulhs or equvialent
2080 // If d > 0 and m < 0, add the numerator
2081 if (d > 0 && magics.m < 0) {
2082 Q = DAG.getNode(ISD::ADD, VT, Q, N->getOperand(0));
2084 Created->push_back(Q.Val);
2086 // If d < 0 and m > 0, subtract the numerator.
2087 if (d < 0 && magics.m > 0) {
2088 Q = DAG.getNode(ISD::SUB, VT, Q, N->getOperand(0));
2090 Created->push_back(Q.Val);
2092 // Shift right algebraic if shift value is nonzero
2094 Q = DAG.getNode(ISD::SRA, VT, Q,
2095 DAG.getConstant(magics.s, getShiftAmountTy()));
2097 Created->push_back(Q.Val);
2099 // Extract the sign bit and add it to the quotient
2101 DAG.getNode(ISD::SRL, VT, Q, DAG.getConstant(VT.getSizeInBits()-1,
2102 getShiftAmountTy()));
2104 Created->push_back(T.Val);
2105 return DAG.getNode(ISD::ADD, VT, Q, T);
2108 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
2109 /// return a DAG expression to select that will generate the same value by
2110 /// multiplying by a magic number. See:
2111 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
2112 SDOperand TargetLowering::BuildUDIV(SDNode *N, SelectionDAG &DAG,
2113 std::vector<SDNode*>* Created) const {
2114 MVT VT = N->getValueType(0);
2116 // Check to see if we can do this.
2117 if (!isTypeLegal(VT) || (VT != MVT::i32 && VT != MVT::i64))
2118 return SDOperand(); // BuildUDIV only operates on i32 or i64
2120 uint64_t d = cast<ConstantSDNode>(N->getOperand(1))->getValue();
2121 mu magics = (VT == MVT::i32) ? magicu32(d) : magicu64(d);
2123 // Multiply the numerator (operand 0) by the magic value
2125 if (isOperationLegal(ISD::MULHU, VT))
2126 Q = DAG.getNode(ISD::MULHU, VT, N->getOperand(0),
2127 DAG.getConstant(magics.m, VT));
2128 else if (isOperationLegal(ISD::UMUL_LOHI, VT))
2129 Q = SDOperand(DAG.getNode(ISD::UMUL_LOHI, DAG.getVTList(VT, VT),
2131 DAG.getConstant(magics.m, VT)).Val, 1);
2133 return SDOperand(); // No mulhu or equvialent
2135 Created->push_back(Q.Val);
2137 if (magics.a == 0) {
2138 return DAG.getNode(ISD::SRL, VT, Q,
2139 DAG.getConstant(magics.s, getShiftAmountTy()));
2141 SDOperand NPQ = DAG.getNode(ISD::SUB, VT, N->getOperand(0), Q);
2143 Created->push_back(NPQ.Val);
2144 NPQ = DAG.getNode(ISD::SRL, VT, NPQ,
2145 DAG.getConstant(1, getShiftAmountTy()));
2147 Created->push_back(NPQ.Val);
2148 NPQ = DAG.getNode(ISD::ADD, VT, NPQ, Q);
2150 Created->push_back(NPQ.Val);
2151 return DAG.getNode(ISD::SRL, VT, NPQ,
2152 DAG.getConstant(magics.s-1, getShiftAmountTy()));