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_I32_F80] = "__floatunsixf";
139 Names[RTLIB::UINTTOFP_I32_PPCF128] = "__floatunsitf";
140 Names[RTLIB::UINTTOFP_I64_F32] = "__floatundisf";
141 Names[RTLIB::UINTTOFP_I64_F64] = "__floatundidf";
142 Names[RTLIB::UINTTOFP_I64_F80] = "__floatundixf";
143 Names[RTLIB::UINTTOFP_I64_PPCF128] = "__floatunditf";
144 Names[RTLIB::UINTTOFP_I128_F32] = "__floatuntisf";
145 Names[RTLIB::UINTTOFP_I128_F64] = "__floatuntidf";
146 Names[RTLIB::UINTTOFP_I128_F80] = "__floatuntixf";
147 Names[RTLIB::UINTTOFP_I128_PPCF128] = "__floatuntitf";
148 Names[RTLIB::OEQ_F32] = "__eqsf2";
149 Names[RTLIB::OEQ_F64] = "__eqdf2";
150 Names[RTLIB::UNE_F32] = "__nesf2";
151 Names[RTLIB::UNE_F64] = "__nedf2";
152 Names[RTLIB::OGE_F32] = "__gesf2";
153 Names[RTLIB::OGE_F64] = "__gedf2";
154 Names[RTLIB::OLT_F32] = "__ltsf2";
155 Names[RTLIB::OLT_F64] = "__ltdf2";
156 Names[RTLIB::OLE_F32] = "__lesf2";
157 Names[RTLIB::OLE_F64] = "__ledf2";
158 Names[RTLIB::OGT_F32] = "__gtsf2";
159 Names[RTLIB::OGT_F64] = "__gtdf2";
160 Names[RTLIB::UO_F32] = "__unordsf2";
161 Names[RTLIB::UO_F64] = "__unorddf2";
162 Names[RTLIB::O_F32] = "__unordsf2";
163 Names[RTLIB::O_F64] = "__unorddf2";
166 /// getFPEXT - Return the FPEXT_*_* value for the given types, or
167 /// UNKNOWN_LIBCALL if there is none.
168 RTLIB::Libcall RTLIB::getFPEXT(MVT OpVT, MVT RetVT) {
169 if (OpVT == MVT::f32) {
170 if (RetVT == MVT::f64)
171 return FPEXT_F32_F64;
173 return UNKNOWN_LIBCALL;
176 /// getFPROUND - Return the FPROUND_*_* value for the given types, or
177 /// UNKNOWN_LIBCALL if there is none.
178 RTLIB::Libcall RTLIB::getFPROUND(MVT OpVT, MVT RetVT) {
179 if (OpVT == MVT::f64) {
180 if (RetVT == MVT::f32)
181 return FPROUND_F64_F32;
183 return UNKNOWN_LIBCALL;
186 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or
187 /// UNKNOWN_LIBCALL if there is none.
188 RTLIB::Libcall RTLIB::getFPTOSINT(MVT OpVT, MVT RetVT) {
189 if (OpVT == MVT::f32) {
190 if (RetVT == MVT::i32)
191 return FPTOSINT_F32_I32;
192 if (RetVT == MVT::i64)
193 return FPTOSINT_F32_I64;
194 if (RetVT == MVT::i128)
195 return FPTOSINT_F32_I128;
196 } else if (OpVT == MVT::f64) {
197 if (RetVT == MVT::i32)
198 return FPTOSINT_F64_I32;
199 if (RetVT == MVT::i64)
200 return FPTOSINT_F64_I64;
201 if (RetVT == MVT::i128)
202 return FPTOSINT_F64_I128;
203 } else if (OpVT == MVT::f80) {
204 if (RetVT == MVT::i32)
205 return FPTOSINT_F80_I32;
206 if (RetVT == MVT::i64)
207 return FPTOSINT_F80_I64;
208 if (RetVT == MVT::i128)
209 return FPTOSINT_F80_I128;
210 } else if (OpVT == MVT::ppcf128) {
211 if (RetVT == MVT::i32)
212 return FPTOSINT_PPCF128_I32;
213 if (RetVT == MVT::i64)
214 return FPTOSINT_PPCF128_I64;
215 if (RetVT == MVT::i128)
216 return FPTOSINT_PPCF128_I128;
218 return UNKNOWN_LIBCALL;
221 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or
222 /// UNKNOWN_LIBCALL if there is none.
223 RTLIB::Libcall RTLIB::getFPTOUINT(MVT OpVT, MVT RetVT) {
224 if (OpVT == MVT::f32) {
225 if (RetVT == MVT::i32)
226 return FPTOUINT_F32_I32;
227 if (RetVT == MVT::i64)
228 return FPTOUINT_F32_I64;
229 if (RetVT == MVT::i128)
230 return FPTOUINT_F32_I128;
231 } else if (OpVT == MVT::f64) {
232 if (RetVT == MVT::i32)
233 return FPTOUINT_F64_I32;
234 if (RetVT == MVT::i64)
235 return FPTOUINT_F64_I64;
236 if (RetVT == MVT::i128)
237 return FPTOUINT_F64_I128;
238 } else if (OpVT == MVT::f80) {
239 if (RetVT == MVT::i32)
240 return FPTOUINT_F80_I32;
241 if (RetVT == MVT::i64)
242 return FPTOUINT_F80_I64;
243 if (RetVT == MVT::i128)
244 return FPTOUINT_F80_I128;
245 } else if (OpVT == MVT::ppcf128) {
246 if (RetVT == MVT::i32)
247 return FPTOUINT_PPCF128_I32;
248 if (RetVT == MVT::i64)
249 return FPTOUINT_PPCF128_I64;
250 if (RetVT == MVT::i128)
251 return FPTOUINT_PPCF128_I128;
253 return UNKNOWN_LIBCALL;
256 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or
257 /// UNKNOWN_LIBCALL if there is none.
258 RTLIB::Libcall RTLIB::getSINTTOFP(MVT OpVT, MVT RetVT) {
259 if (OpVT == MVT::i32) {
260 if (RetVT == MVT::f32)
261 return SINTTOFP_I32_F32;
262 else if (RetVT == MVT::f64)
263 return SINTTOFP_I32_F64;
264 else if (RetVT == MVT::f80)
265 return SINTTOFP_I32_F80;
266 else if (RetVT == MVT::ppcf128)
267 return SINTTOFP_I32_PPCF128;
268 } else if (OpVT == MVT::i64) {
269 if (RetVT == MVT::f32)
270 return SINTTOFP_I64_F32;
271 else if (RetVT == MVT::f64)
272 return SINTTOFP_I64_F64;
273 else if (RetVT == MVT::f80)
274 return SINTTOFP_I64_F80;
275 else if (RetVT == MVT::ppcf128)
276 return SINTTOFP_I64_PPCF128;
277 } else if (OpVT == MVT::i128) {
278 if (RetVT == MVT::f32)
279 return SINTTOFP_I128_F32;
280 else if (RetVT == MVT::f64)
281 return SINTTOFP_I128_F64;
282 else if (RetVT == MVT::f80)
283 return SINTTOFP_I128_F80;
284 else if (RetVT == MVT::ppcf128)
285 return SINTTOFP_I128_PPCF128;
287 return UNKNOWN_LIBCALL;
290 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or
291 /// UNKNOWN_LIBCALL if there is none.
292 RTLIB::Libcall RTLIB::getUINTTOFP(MVT OpVT, MVT RetVT) {
293 if (OpVT == MVT::i32) {
294 if (RetVT == MVT::f32)
295 return UINTTOFP_I32_F32;
296 else if (RetVT == MVT::f64)
297 return UINTTOFP_I32_F64;
298 else if (RetVT == MVT::f80)
299 return UINTTOFP_I32_F80;
300 else if (RetVT == MVT::ppcf128)
301 return UINTTOFP_I32_PPCF128;
302 } else if (OpVT == MVT::i64) {
303 if (RetVT == MVT::f32)
304 return UINTTOFP_I64_F32;
305 else if (RetVT == MVT::f64)
306 return UINTTOFP_I64_F64;
307 else if (RetVT == MVT::f80)
308 return UINTTOFP_I64_F80;
309 else if (RetVT == MVT::ppcf128)
310 return UINTTOFP_I64_PPCF128;
311 } else if (OpVT == MVT::i128) {
312 if (RetVT == MVT::f32)
313 return UINTTOFP_I128_F32;
314 else if (RetVT == MVT::f64)
315 return UINTTOFP_I128_F64;
316 else if (RetVT == MVT::f80)
317 return UINTTOFP_I128_F80;
318 else if (RetVT == MVT::ppcf128)
319 return UINTTOFP_I128_PPCF128;
321 return UNKNOWN_LIBCALL;
324 /// InitCmpLibcallCCs - Set default comparison libcall CC.
326 static void InitCmpLibcallCCs(ISD::CondCode *CCs) {
327 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL);
328 CCs[RTLIB::OEQ_F32] = ISD::SETEQ;
329 CCs[RTLIB::OEQ_F64] = ISD::SETEQ;
330 CCs[RTLIB::UNE_F32] = ISD::SETNE;
331 CCs[RTLIB::UNE_F64] = ISD::SETNE;
332 CCs[RTLIB::OGE_F32] = ISD::SETGE;
333 CCs[RTLIB::OGE_F64] = ISD::SETGE;
334 CCs[RTLIB::OLT_F32] = ISD::SETLT;
335 CCs[RTLIB::OLT_F64] = ISD::SETLT;
336 CCs[RTLIB::OLE_F32] = ISD::SETLE;
337 CCs[RTLIB::OLE_F64] = ISD::SETLE;
338 CCs[RTLIB::OGT_F32] = ISD::SETGT;
339 CCs[RTLIB::OGT_F64] = ISD::SETGT;
340 CCs[RTLIB::UO_F32] = ISD::SETNE;
341 CCs[RTLIB::UO_F64] = ISD::SETNE;
342 CCs[RTLIB::O_F32] = ISD::SETEQ;
343 CCs[RTLIB::O_F64] = ISD::SETEQ;
346 TargetLowering::TargetLowering(TargetMachine &tm)
347 : TM(tm), TD(TM.getTargetData()) {
348 assert(ISD::BUILTIN_OP_END <= OpActionsCapacity &&
349 "Fixed size array in TargetLowering is not large enough!");
350 // All operations default to being supported.
351 memset(OpActions, 0, sizeof(OpActions));
352 memset(LoadXActions, 0, sizeof(LoadXActions));
353 memset(TruncStoreActions, 0, sizeof(TruncStoreActions));
354 memset(IndexedModeActions, 0, sizeof(IndexedModeActions));
355 memset(ConvertActions, 0, sizeof(ConvertActions));
357 // Set default actions for various operations.
358 for (unsigned VT = 0; VT != (unsigned)MVT::LAST_VALUETYPE; ++VT) {
359 // Default all indexed load / store to expand.
360 for (unsigned IM = (unsigned)ISD::PRE_INC;
361 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
362 setIndexedLoadAction(IM, (MVT::SimpleValueType)VT, Expand);
363 setIndexedStoreAction(IM, (MVT::SimpleValueType)VT, Expand);
366 // These operations default to expand.
367 setOperationAction(ISD::FGETSIGN, (MVT::SimpleValueType)VT, Expand);
370 // Most targets ignore the @llvm.prefetch intrinsic.
371 setOperationAction(ISD::PREFETCH, MVT::Other, Expand);
373 // ConstantFP nodes default to expand. Targets can either change this to
374 // Legal, in which case all fp constants are legal, or use addLegalFPImmediate
375 // to optimize expansions for certain constants.
376 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
377 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
378 setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
380 // Default ISD::TRAP to expand (which turns it into abort).
381 setOperationAction(ISD::TRAP, MVT::Other, Expand);
383 IsLittleEndian = TD->isLittleEndian();
384 UsesGlobalOffsetTable = false;
385 ShiftAmountTy = PointerTy = getValueType(TD->getIntPtrType());
386 ShiftAmtHandling = Undefined;
387 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE*sizeof(TargetRegisterClass*));
388 memset(TargetDAGCombineArray, 0, array_lengthof(TargetDAGCombineArray));
389 maxStoresPerMemset = maxStoresPerMemcpy = maxStoresPerMemmove = 8;
390 allowUnalignedMemoryAccesses = false;
391 UseUnderscoreSetJmp = false;
392 UseUnderscoreLongJmp = false;
393 SelectIsExpensive = false;
394 IntDivIsCheap = false;
395 Pow2DivIsCheap = false;
396 StackPointerRegisterToSaveRestore = 0;
397 ExceptionPointerRegister = 0;
398 ExceptionSelectorRegister = 0;
399 SetCCResultContents = UndefinedSetCCResult;
400 SchedPreferenceInfo = SchedulingForLatency;
402 JumpBufAlignment = 0;
403 IfCvtBlockSizeLimit = 2;
404 IfCvtDupBlockSizeLimit = 0;
405 PrefLoopAlignment = 0;
407 InitLibcallNames(LibcallRoutineNames);
408 InitCmpLibcallCCs(CmpLibcallCCs);
410 // Tell Legalize whether the assembler supports DEBUG_LOC.
411 if (!TM.getTargetAsmInfo()->hasDotLocAndDotFile())
412 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
415 TargetLowering::~TargetLowering() {}
417 /// computeRegisterProperties - Once all of the register classes are added,
418 /// this allows us to compute derived properties we expose.
419 void TargetLowering::computeRegisterProperties() {
420 assert(MVT::LAST_VALUETYPE <= 32 &&
421 "Too many value types for ValueTypeActions to hold!");
423 // Everything defaults to needing one register.
424 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
425 NumRegistersForVT[i] = 1;
426 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i;
428 // ...except isVoid, which doesn't need any registers.
429 NumRegistersForVT[MVT::isVoid] = 0;
431 // Find the largest integer register class.
432 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE;
433 for (; RegClassForVT[LargestIntReg] == 0; --LargestIntReg)
434 assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
436 // Every integer value type larger than this largest register takes twice as
437 // many registers to represent as the previous ValueType.
438 for (unsigned ExpandedReg = LargestIntReg + 1; ; ++ExpandedReg) {
439 MVT EVT = (MVT::SimpleValueType)ExpandedReg;
440 if (!EVT.isInteger())
442 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
443 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg;
444 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1);
445 ValueTypeActions.setTypeAction(EVT, Expand);
448 // Inspect all of the ValueType's smaller than the largest integer
449 // register to see which ones need promotion.
450 unsigned LegalIntReg = LargestIntReg;
451 for (unsigned IntReg = LargestIntReg - 1;
452 IntReg >= (unsigned)MVT::i1; --IntReg) {
453 MVT IVT = (MVT::SimpleValueType)IntReg;
454 if (isTypeLegal(IVT)) {
455 LegalIntReg = IntReg;
457 RegisterTypeForVT[IntReg] = TransformToType[IntReg] =
458 (MVT::SimpleValueType)LegalIntReg;
459 ValueTypeActions.setTypeAction(IVT, Promote);
463 // ppcf128 type is really two f64's.
464 if (!isTypeLegal(MVT::ppcf128)) {
465 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64];
466 RegisterTypeForVT[MVT::ppcf128] = MVT::f64;
467 TransformToType[MVT::ppcf128] = MVT::f64;
468 ValueTypeActions.setTypeAction(MVT::ppcf128, Expand);
471 // Decide how to handle f64. If the target does not have native f64 support,
472 // expand it to i64 and we will be generating soft float library calls.
473 if (!isTypeLegal(MVT::f64)) {
474 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
475 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
476 TransformToType[MVT::f64] = MVT::i64;
477 ValueTypeActions.setTypeAction(MVT::f64, Expand);
480 // Decide how to handle f32. If the target does not have native support for
481 // f32, promote it to f64 if it is legal. Otherwise, expand it to i32.
482 if (!isTypeLegal(MVT::f32)) {
483 if (isTypeLegal(MVT::f64)) {
484 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::f64];
485 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::f64];
486 TransformToType[MVT::f32] = MVT::f64;
487 ValueTypeActions.setTypeAction(MVT::f32, Promote);
489 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
490 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
491 TransformToType[MVT::f32] = MVT::i32;
492 ValueTypeActions.setTypeAction(MVT::f32, Expand);
496 // Loop over all of the vector value types to see which need transformations.
497 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
498 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
499 MVT VT = (MVT::SimpleValueType)i;
500 if (!isTypeLegal(VT)) {
501 MVT IntermediateVT, RegisterVT;
502 unsigned NumIntermediates;
503 NumRegistersForVT[i] =
504 getVectorTypeBreakdown(VT,
505 IntermediateVT, NumIntermediates,
507 RegisterTypeForVT[i] = RegisterVT;
508 TransformToType[i] = MVT::Other; // this isn't actually used
509 ValueTypeActions.setTypeAction(VT, Expand);
514 const char *TargetLowering::getTargetNodeName(unsigned Opcode) const {
519 MVT TargetLowering::getSetCCResultType(const SDValue &) const {
520 return getValueType(TD->getIntPtrType());
524 /// getVectorTypeBreakdown - Vector types are broken down into some number of
525 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
526 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
527 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
529 /// This method returns the number of registers needed, and the VT for each
530 /// register. It also returns the VT and quantity of the intermediate values
531 /// before they are promoted/expanded.
533 unsigned TargetLowering::getVectorTypeBreakdown(MVT VT,
535 unsigned &NumIntermediates,
536 MVT &RegisterVT) const {
537 // Figure out the right, legal destination reg to copy into.
538 unsigned NumElts = VT.getVectorNumElements();
539 MVT EltTy = VT.getVectorElementType();
541 unsigned NumVectorRegs = 1;
543 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
544 // could break down into LHS/RHS like LegalizeDAG does.
545 if (!isPowerOf2_32(NumElts)) {
546 NumVectorRegs = NumElts;
550 // Divide the input until we get to a supported size. This will always
551 // end with a scalar if the target doesn't support vectors.
552 while (NumElts > 1 && !isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) {
557 NumIntermediates = NumVectorRegs;
559 MVT NewVT = MVT::getVectorVT(EltTy, NumElts);
560 if (!isTypeLegal(NewVT))
562 IntermediateVT = NewVT;
564 MVT DestVT = getTypeToTransformTo(NewVT);
566 if (DestVT.bitsLT(NewVT)) {
567 // Value is expanded, e.g. i64 -> i16.
568 return NumVectorRegs*(NewVT.getSizeInBits()/DestVT.getSizeInBits());
570 // Otherwise, promotion or legal types use the same number of registers as
571 // the vector decimated to the appropriate level.
572 return NumVectorRegs;
578 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
579 /// function arguments in the caller parameter area. This is the actual
580 /// alignment, not its logarithm.
581 unsigned TargetLowering::getByValTypeAlignment(const Type *Ty) const {
582 return TD->getCallFrameTypeAlignment(Ty);
585 SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table,
586 SelectionDAG &DAG) const {
587 if (usesGlobalOffsetTable())
588 return DAG.getNode(ISD::GLOBAL_OFFSET_TABLE, getPointerTy());
592 //===----------------------------------------------------------------------===//
593 // Optimization Methods
594 //===----------------------------------------------------------------------===//
596 /// ShrinkDemandedConstant - Check to see if the specified operand of the
597 /// specified instruction is a constant integer. If so, check to see if there
598 /// are any bits set in the constant that are not demanded. If so, shrink the
599 /// constant and return true.
600 bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDValue Op,
601 const APInt &Demanded) {
602 // FIXME: ISD::SELECT, ISD::SELECT_CC
603 switch(Op.getOpcode()) {
608 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1)))
609 if (C->getAPIntValue().intersects(~Demanded)) {
610 MVT VT = Op.getValueType();
611 SDValue New = DAG.getNode(Op.getOpcode(), VT, Op.getOperand(0),
612 DAG.getConstant(Demanded &
615 return CombineTo(Op, New);
622 /// SimplifyDemandedBits - Look at Op. At this point, we know that only the
623 /// DemandedMask bits of the result of Op are ever used downstream. If we can
624 /// use this information to simplify Op, create a new simplified DAG node and
625 /// return true, returning the original and new nodes in Old and New. Otherwise,
626 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
627 /// the expression (used to simplify the caller). The KnownZero/One bits may
628 /// only be accurate for those bits in the DemandedMask.
629 bool TargetLowering::SimplifyDemandedBits(SDValue Op,
630 const APInt &DemandedMask,
633 TargetLoweringOpt &TLO,
634 unsigned Depth) const {
635 unsigned BitWidth = DemandedMask.getBitWidth();
636 assert(Op.getValueSizeInBits() == BitWidth &&
637 "Mask size mismatches value type size!");
638 APInt NewMask = DemandedMask;
640 // Don't know anything.
641 KnownZero = KnownOne = APInt(BitWidth, 0);
643 // Other users may use these bits.
644 if (!Op.Val->hasOneUse()) {
646 // If not at the root, Just compute the KnownZero/KnownOne bits to
647 // simplify things downstream.
648 TLO.DAG.ComputeMaskedBits(Op, DemandedMask, KnownZero, KnownOne, Depth);
651 // If this is the root being simplified, allow it to have multiple uses,
652 // just set the NewMask to all bits.
653 NewMask = APInt::getAllOnesValue(BitWidth);
654 } else if (DemandedMask == 0) {
655 // Not demanding any bits from Op.
656 if (Op.getOpcode() != ISD::UNDEF)
657 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::UNDEF, Op.getValueType()));
659 } else if (Depth == 6) { // Limit search depth.
663 APInt KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut;
664 switch (Op.getOpcode()) {
666 // We know all of the bits for a constant!
667 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue() & NewMask;
668 KnownZero = ~KnownOne & NewMask;
669 return false; // Don't fall through, will infinitely loop.
671 // If the RHS is a constant, check to see if the LHS would be zero without
672 // using the bits from the RHS. Below, we use knowledge about the RHS to
673 // simplify the LHS, here we're using information from the LHS to simplify
675 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
676 APInt LHSZero, LHSOne;
677 TLO.DAG.ComputeMaskedBits(Op.getOperand(0), NewMask,
678 LHSZero, LHSOne, Depth+1);
679 // If the LHS already has zeros where RHSC does, this and is dead.
680 if ((LHSZero & NewMask) == (~RHSC->getAPIntValue() & NewMask))
681 return TLO.CombineTo(Op, Op.getOperand(0));
682 // If any of the set bits in the RHS are known zero on the LHS, shrink
684 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & NewMask))
688 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
689 KnownOne, TLO, Depth+1))
691 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
692 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownZero & NewMask,
693 KnownZero2, KnownOne2, TLO, Depth+1))
695 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
697 // If all of the demanded bits are known one on one side, return the other.
698 // These bits cannot contribute to the result of the 'and'.
699 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
700 return TLO.CombineTo(Op, Op.getOperand(0));
701 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
702 return TLO.CombineTo(Op, Op.getOperand(1));
703 // If all of the demanded bits in the inputs are known zeros, return zero.
704 if ((NewMask & (KnownZero|KnownZero2)) == NewMask)
705 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, Op.getValueType()));
706 // If the RHS is a constant, see if we can simplify it.
707 if (TLO.ShrinkDemandedConstant(Op, ~KnownZero2 & NewMask))
710 // Output known-1 bits are only known if set in both the LHS & RHS.
711 KnownOne &= KnownOne2;
712 // Output known-0 are known to be clear if zero in either the LHS | RHS.
713 KnownZero |= KnownZero2;
716 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
717 KnownOne, TLO, Depth+1))
719 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
720 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownOne & NewMask,
721 KnownZero2, KnownOne2, TLO, Depth+1))
723 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
725 // If all of the demanded bits are known zero on one side, return the other.
726 // These bits cannot contribute to the result of the 'or'.
727 if ((NewMask & ~KnownOne2 & KnownZero) == (~KnownOne2 & NewMask))
728 return TLO.CombineTo(Op, Op.getOperand(0));
729 if ((NewMask & ~KnownOne & KnownZero2) == (~KnownOne & NewMask))
730 return TLO.CombineTo(Op, Op.getOperand(1));
731 // If all of the potentially set bits on one side are known to be set on
732 // the other side, just use the 'other' side.
733 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
734 return TLO.CombineTo(Op, Op.getOperand(0));
735 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
736 return TLO.CombineTo(Op, Op.getOperand(1));
737 // If the RHS is a constant, see if we can simplify it.
738 if (TLO.ShrinkDemandedConstant(Op, NewMask))
741 // Output known-0 bits are only known if clear in both the LHS & RHS.
742 KnownZero &= KnownZero2;
743 // Output known-1 are known to be set if set in either the LHS | RHS.
744 KnownOne |= KnownOne2;
747 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
748 KnownOne, TLO, Depth+1))
750 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
751 if (SimplifyDemandedBits(Op.getOperand(0), NewMask, KnownZero2,
752 KnownOne2, TLO, Depth+1))
754 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
756 // If all of the demanded bits are known zero on one side, return the other.
757 // These bits cannot contribute to the result of the 'xor'.
758 if ((KnownZero & NewMask) == NewMask)
759 return TLO.CombineTo(Op, Op.getOperand(0));
760 if ((KnownZero2 & NewMask) == NewMask)
761 return TLO.CombineTo(Op, Op.getOperand(1));
763 // If all of the unknown bits are known to be zero on one side or the other
764 // (but not both) turn this into an *inclusive* or.
765 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
766 if ((NewMask & ~KnownZero & ~KnownZero2) == 0)
767 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, Op.getValueType(),
771 // Output known-0 bits are known if clear or set in both the LHS & RHS.
772 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
773 // Output known-1 are known to be set if set in only one of the LHS, RHS.
774 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
776 // If all of the demanded bits on one side are known, and all of the set
777 // bits on that side are also known to be set on the other side, turn this
778 // into an AND, as we know the bits will be cleared.
779 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
780 if ((NewMask & (KnownZero|KnownOne)) == NewMask) { // all known
781 if ((KnownOne & KnownOne2) == KnownOne) {
782 MVT VT = Op.getValueType();
783 SDValue ANDC = TLO.DAG.getConstant(~KnownOne & NewMask, VT);
784 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, VT, Op.getOperand(0),
789 // If the RHS is a constant, see if we can simplify it.
790 // for XOR, we prefer to force bits to 1 if they will make a -1.
791 // if we can't force bits, try to shrink constant
792 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
793 APInt Expanded = C->getAPIntValue() | (~NewMask);
794 // if we can expand it to have all bits set, do it
795 if (Expanded.isAllOnesValue()) {
796 if (Expanded != C->getAPIntValue()) {
797 MVT VT = Op.getValueType();
798 SDValue New = TLO.DAG.getNode(Op.getOpcode(), VT, Op.getOperand(0),
799 TLO.DAG.getConstant(Expanded, VT));
800 return TLO.CombineTo(Op, New);
802 // if it already has all the bits set, nothing to change
803 // but don't shrink either!
804 } else if (TLO.ShrinkDemandedConstant(Op, NewMask)) {
809 KnownZero = KnownZeroOut;
810 KnownOne = KnownOneOut;
813 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero,
814 KnownOne, TLO, Depth+1))
816 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero2,
817 KnownOne2, TLO, Depth+1))
819 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
820 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
822 // If the operands are constants, see if we can simplify them.
823 if (TLO.ShrinkDemandedConstant(Op, NewMask))
826 // Only known if known in both the LHS and RHS.
827 KnownOne &= KnownOne2;
828 KnownZero &= KnownZero2;
831 if (SimplifyDemandedBits(Op.getOperand(3), NewMask, KnownZero,
832 KnownOne, TLO, Depth+1))
834 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero2,
835 KnownOne2, TLO, Depth+1))
837 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
838 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
840 // If the operands are constants, see if we can simplify them.
841 if (TLO.ShrinkDemandedConstant(Op, NewMask))
844 // Only known if known in both the LHS and RHS.
845 KnownOne &= KnownOne2;
846 KnownZero &= KnownZero2;
849 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
850 unsigned ShAmt = SA->getValue();
851 SDValue InOp = Op.getOperand(0);
853 // If the shift count is an invalid immediate, don't do anything.
854 if (ShAmt >= BitWidth)
857 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a
858 // single shift. We can do this if the bottom bits (which are shifted
859 // out) are never demanded.
860 if (InOp.getOpcode() == ISD::SRL &&
861 isa<ConstantSDNode>(InOp.getOperand(1))) {
862 if (ShAmt && (NewMask & APInt::getLowBitsSet(BitWidth, ShAmt)) == 0) {
863 unsigned C1 = cast<ConstantSDNode>(InOp.getOperand(1))->getValue();
864 unsigned Opc = ISD::SHL;
872 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
873 MVT VT = Op.getValueType();
874 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, VT,
875 InOp.getOperand(0), NewSA));
879 if (SimplifyDemandedBits(Op.getOperand(0), NewMask.lshr(ShAmt),
880 KnownZero, KnownOne, TLO, Depth+1))
882 KnownZero <<= SA->getValue();
883 KnownOne <<= SA->getValue();
884 // low bits known zero.
885 KnownZero |= APInt::getLowBitsSet(BitWidth, SA->getValue());
889 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
890 MVT VT = Op.getValueType();
891 unsigned ShAmt = SA->getValue();
892 unsigned VTSize = VT.getSizeInBits();
893 SDValue InOp = Op.getOperand(0);
895 // If the shift count is an invalid immediate, don't do anything.
896 if (ShAmt >= BitWidth)
899 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a
900 // single shift. We can do this if the top bits (which are shifted out)
901 // are never demanded.
902 if (InOp.getOpcode() == ISD::SHL &&
903 isa<ConstantSDNode>(InOp.getOperand(1))) {
904 if (ShAmt && (NewMask & APInt::getHighBitsSet(VTSize, ShAmt)) == 0) {
905 unsigned C1 = cast<ConstantSDNode>(InOp.getOperand(1))->getValue();
906 unsigned Opc = ISD::SRL;
914 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
915 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, VT,
916 InOp.getOperand(0), NewSA));
920 // Compute the new bits that are at the top now.
921 if (SimplifyDemandedBits(InOp, (NewMask << ShAmt),
922 KnownZero, KnownOne, TLO, Depth+1))
924 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
925 KnownZero = KnownZero.lshr(ShAmt);
926 KnownOne = KnownOne.lshr(ShAmt);
928 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
929 KnownZero |= HighBits; // High bits known zero.
933 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
934 MVT VT = Op.getValueType();
935 unsigned ShAmt = SA->getValue();
937 // If the shift count is an invalid immediate, don't do anything.
938 if (ShAmt >= BitWidth)
941 APInt InDemandedMask = (NewMask << ShAmt);
943 // If any of the demanded bits are produced by the sign extension, we also
944 // demand the input sign bit.
945 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
946 if (HighBits.intersects(NewMask))
947 InDemandedMask |= APInt::getSignBit(VT.getSizeInBits());
949 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask,
950 KnownZero, KnownOne, TLO, Depth+1))
952 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
953 KnownZero = KnownZero.lshr(ShAmt);
954 KnownOne = KnownOne.lshr(ShAmt);
956 // Handle the sign bit, adjusted to where it is now in the mask.
957 APInt SignBit = APInt::getSignBit(BitWidth).lshr(ShAmt);
959 // If the input sign bit is known to be zero, or if none of the top bits
960 // are demanded, turn this into an unsigned shift right.
961 if (KnownZero.intersects(SignBit) || (HighBits & ~NewMask) == HighBits) {
962 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, VT, Op.getOperand(0),
964 } else if (KnownOne.intersects(SignBit)) { // New bits are known one.
965 KnownOne |= HighBits;
969 case ISD::SIGN_EXTEND_INREG: {
970 MVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
972 // Sign extension. Compute the demanded bits in the result that are not
973 // present in the input.
974 APInt NewBits = APInt::getHighBitsSet(BitWidth,
975 BitWidth - EVT.getSizeInBits()) &
978 // If none of the extended bits are demanded, eliminate the sextinreg.
980 return TLO.CombineTo(Op, Op.getOperand(0));
982 APInt InSignBit = APInt::getSignBit(EVT.getSizeInBits());
983 InSignBit.zext(BitWidth);
984 APInt InputDemandedBits = APInt::getLowBitsSet(BitWidth,
985 EVT.getSizeInBits()) &
988 // Since the sign extended bits are demanded, we know that the sign
990 InputDemandedBits |= InSignBit;
992 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits,
993 KnownZero, KnownOne, TLO, Depth+1))
995 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
997 // If the sign bit of the input is known set or clear, then we know the
998 // top bits of the result.
1000 // If the input sign bit is known zero, convert this into a zero extension.
1001 if (KnownZero.intersects(InSignBit))
1002 return TLO.CombineTo(Op,
1003 TLO.DAG.getZeroExtendInReg(Op.getOperand(0), EVT));
1005 if (KnownOne.intersects(InSignBit)) { // Input sign bit known set
1006 KnownOne |= NewBits;
1007 KnownZero &= ~NewBits;
1008 } else { // Input sign bit unknown
1009 KnownZero &= ~NewBits;
1010 KnownOne &= ~NewBits;
1014 case ISD::ZERO_EXTEND: {
1015 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits();
1016 APInt InMask = NewMask;
1017 InMask.trunc(OperandBitWidth);
1019 // If none of the top bits are demanded, convert this into an any_extend.
1021 APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask;
1022 if (!NewBits.intersects(NewMask))
1023 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND,
1027 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1028 KnownZero, KnownOne, TLO, Depth+1))
1030 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1031 KnownZero.zext(BitWidth);
1032 KnownOne.zext(BitWidth);
1033 KnownZero |= NewBits;
1036 case ISD::SIGN_EXTEND: {
1037 MVT InVT = Op.getOperand(0).getValueType();
1038 unsigned InBits = InVT.getSizeInBits();
1039 APInt InMask = APInt::getLowBitsSet(BitWidth, InBits);
1040 APInt InSignBit = APInt::getBitsSet(BitWidth, InBits - 1, InBits);
1041 APInt NewBits = ~InMask & NewMask;
1043 // If none of the top bits are demanded, convert this into an any_extend.
1045 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND,Op.getValueType(),
1048 // Since some of the sign extended bits are demanded, we know that the sign
1050 APInt InDemandedBits = InMask & NewMask;
1051 InDemandedBits |= InSignBit;
1052 InDemandedBits.trunc(InBits);
1054 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero,
1055 KnownOne, TLO, Depth+1))
1057 KnownZero.zext(BitWidth);
1058 KnownOne.zext(BitWidth);
1060 // If the sign bit is known zero, convert this to a zero extend.
1061 if (KnownZero.intersects(InSignBit))
1062 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND,
1066 // If the sign bit is known one, the top bits match.
1067 if (KnownOne.intersects(InSignBit)) {
1068 KnownOne |= NewBits;
1069 KnownZero &= ~NewBits;
1070 } else { // Otherwise, top bits aren't known.
1071 KnownOne &= ~NewBits;
1072 KnownZero &= ~NewBits;
1076 case ISD::ANY_EXTEND: {
1077 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits();
1078 APInt InMask = NewMask;
1079 InMask.trunc(OperandBitWidth);
1080 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1081 KnownZero, KnownOne, TLO, Depth+1))
1083 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1084 KnownZero.zext(BitWidth);
1085 KnownOne.zext(BitWidth);
1088 case ISD::TRUNCATE: {
1089 // Simplify the input, using demanded bit information, and compute the known
1090 // zero/one bits live out.
1091 APInt TruncMask = NewMask;
1092 TruncMask.zext(Op.getOperand(0).getValueSizeInBits());
1093 if (SimplifyDemandedBits(Op.getOperand(0), TruncMask,
1094 KnownZero, KnownOne, TLO, Depth+1))
1096 KnownZero.trunc(BitWidth);
1097 KnownOne.trunc(BitWidth);
1099 // If the input is only used by this truncate, see if we can shrink it based
1100 // on the known demanded bits.
1101 if (Op.getOperand(0).Val->hasOneUse()) {
1102 SDValue In = Op.getOperand(0);
1103 unsigned InBitWidth = In.getValueSizeInBits();
1104 switch (In.getOpcode()) {
1107 // Shrink SRL by a constant if none of the high bits shifted in are
1109 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1))){
1110 APInt HighBits = APInt::getHighBitsSet(InBitWidth,
1111 InBitWidth - BitWidth);
1112 HighBits = HighBits.lshr(ShAmt->getValue());
1113 HighBits.trunc(BitWidth);
1115 if (ShAmt->getValue() < BitWidth && !(HighBits & NewMask)) {
1116 // None of the shifted in bits are needed. Add a truncate of the
1117 // shift input, then shift it.
1118 SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE,
1121 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL,Op.getValueType(),
1122 NewTrunc, In.getOperand(1)));
1129 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1132 case ISD::AssertZext: {
1133 MVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1134 APInt InMask = APInt::getLowBitsSet(BitWidth,
1135 VT.getSizeInBits());
1136 if (SimplifyDemandedBits(Op.getOperand(0), InMask & NewMask,
1137 KnownZero, KnownOne, TLO, Depth+1))
1139 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1140 KnownZero |= ~InMask & NewMask;
1143 case ISD::BIT_CONVERT:
1145 // If this is an FP->Int bitcast and if the sign bit is the only thing that
1146 // is demanded, turn this into a FGETSIGN.
1147 if (NewMask == MVT::getIntegerVTSignBit(Op.getValueType()) &&
1148 MVT::isFloatingPoint(Op.getOperand(0).getValueType()) &&
1149 !MVT::isVector(Op.getOperand(0).getValueType())) {
1150 // Only do this xform if FGETSIGN is valid or if before legalize.
1151 if (!TLO.AfterLegalize ||
1152 isOperationLegal(ISD::FGETSIGN, Op.getValueType())) {
1153 // Make a FGETSIGN + SHL to move the sign bit into the appropriate
1154 // place. We expect the SHL to be eliminated by other optimizations.
1155 SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, Op.getValueType(),
1157 unsigned ShVal = Op.getValueType().getSizeInBits()-1;
1158 SDValue ShAmt = TLO.DAG.getConstant(ShVal, getShiftAmountTy());
1159 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, Op.getValueType(),
1166 // Just use ComputeMaskedBits to compute output bits.
1167 TLO.DAG.ComputeMaskedBits(Op, NewMask, KnownZero, KnownOne, Depth);
1171 // If we know the value of all of the demanded bits, return this as a
1173 if ((NewMask & (KnownZero|KnownOne)) == NewMask)
1174 return TLO.CombineTo(Op, TLO.DAG.getConstant(KnownOne, Op.getValueType()));
1179 /// computeMaskedBitsForTargetNode - Determine which of the bits specified
1180 /// in Mask are known to be either zero or one and return them in the
1181 /// KnownZero/KnownOne bitsets.
1182 void TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
1186 const SelectionDAG &DAG,
1187 unsigned Depth) const {
1188 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1189 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1190 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1191 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1192 "Should use MaskedValueIsZero if you don't know whether Op"
1193 " is a target node!");
1194 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
1197 /// ComputeNumSignBitsForTargetNode - This method can be implemented by
1198 /// targets that want to expose additional information about sign bits to the
1200 unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
1201 unsigned Depth) const {
1202 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1203 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1204 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1205 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1206 "Should use ComputeNumSignBits if you don't know whether Op"
1207 " is a target node!");
1212 /// SimplifySetCC - Try to simplify a setcc built with the specified operands
1213 /// and cc. If it is unable to simplify it, return a null SDValue.
1215 TargetLowering::SimplifySetCC(MVT VT, SDValue N0, SDValue N1,
1216 ISD::CondCode Cond, bool foldBooleans,
1217 DAGCombinerInfo &DCI) const {
1218 SelectionDAG &DAG = DCI.DAG;
1220 // These setcc operations always fold.
1224 case ISD::SETFALSE2: return DAG.getConstant(0, VT);
1226 case ISD::SETTRUE2: return DAG.getConstant(1, VT);
1229 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val)) {
1230 const APInt &C1 = N1C->getAPIntValue();
1231 if (isa<ConstantSDNode>(N0.Val)) {
1232 return DAG.FoldSetCC(VT, N0, N1, Cond);
1234 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an
1235 // equality comparison, then we're just comparing whether X itself is
1237 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) &&
1238 N0.getOperand(0).getOpcode() == ISD::CTLZ &&
1239 N0.getOperand(1).getOpcode() == ISD::Constant) {
1240 unsigned ShAmt = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1241 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1242 ShAmt == Log2_32(N0.getValueType().getSizeInBits())) {
1243 if ((C1 == 0) == (Cond == ISD::SETEQ)) {
1244 // (srl (ctlz x), 5) == 0 -> X != 0
1245 // (srl (ctlz x), 5) != 1 -> X != 0
1248 // (srl (ctlz x), 5) != 0 -> X == 0
1249 // (srl (ctlz x), 5) == 1 -> X == 0
1252 SDValue Zero = DAG.getConstant(0, N0.getValueType());
1253 return DAG.getSetCC(VT, N0.getOperand(0).getOperand(0),
1258 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
1259 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
1260 unsigned InSize = N0.getOperand(0).getValueType().getSizeInBits();
1262 // If the comparison constant has bits in the upper part, the
1263 // zero-extended value could never match.
1264 if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(),
1265 C1.getBitWidth() - InSize))) {
1269 case ISD::SETEQ: return DAG.getConstant(0, VT);
1272 case ISD::SETNE: return DAG.getConstant(1, VT);
1275 // True if the sign bit of C1 is set.
1276 return DAG.getConstant(C1.isNegative(), VT);
1279 // True if the sign bit of C1 isn't set.
1280 return DAG.getConstant(C1.isNonNegative(), VT);
1286 // Otherwise, we can perform the comparison with the low bits.
1294 return DAG.getSetCC(VT, N0.getOperand(0),
1295 DAG.getConstant(APInt(C1).trunc(InSize),
1296 N0.getOperand(0).getValueType()),
1299 break; // todo, be more careful with signed comparisons
1301 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
1302 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1303 MVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
1304 unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits();
1305 MVT ExtDstTy = N0.getValueType();
1306 unsigned ExtDstTyBits = ExtDstTy.getSizeInBits();
1308 // If the extended part has any inconsistent bits, it cannot ever
1309 // compare equal. In other words, they have to be all ones or all
1312 APInt::getHighBitsSet(ExtDstTyBits, ExtDstTyBits - ExtSrcTyBits);
1313 if ((C1 & ExtBits) != 0 && (C1 & ExtBits) != ExtBits)
1314 return DAG.getConstant(Cond == ISD::SETNE, VT);
1317 MVT Op0Ty = N0.getOperand(0).getValueType();
1318 if (Op0Ty == ExtSrcTy) {
1319 ZextOp = N0.getOperand(0);
1321 APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits);
1322 ZextOp = DAG.getNode(ISD::AND, Op0Ty, N0.getOperand(0),
1323 DAG.getConstant(Imm, Op0Ty));
1325 if (!DCI.isCalledByLegalizer())
1326 DCI.AddToWorklist(ZextOp.Val);
1327 // Otherwise, make this a use of a zext.
1328 return DAG.getSetCC(VT, ZextOp,
1329 DAG.getConstant(C1 & APInt::getLowBitsSet(
1334 } else if ((N1C->isNullValue() || N1C->getAPIntValue() == 1) &&
1335 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1337 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC
1338 if (N0.getOpcode() == ISD::SETCC) {
1339 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getValue() != 1);
1343 // Invert the condition.
1344 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
1345 CC = ISD::getSetCCInverse(CC,
1346 N0.getOperand(0).getValueType().isInteger());
1347 return DAG.getSetCC(VT, N0.getOperand(0), N0.getOperand(1), CC);
1350 if ((N0.getOpcode() == ISD::XOR ||
1351 (N0.getOpcode() == ISD::AND &&
1352 N0.getOperand(0).getOpcode() == ISD::XOR &&
1353 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
1354 isa<ConstantSDNode>(N0.getOperand(1)) &&
1355 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue() == 1) {
1356 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We
1357 // can only do this if the top bits are known zero.
1358 unsigned BitWidth = N0.getValueSizeInBits();
1359 if (DAG.MaskedValueIsZero(N0,
1360 APInt::getHighBitsSet(BitWidth,
1362 // Okay, get the un-inverted input value.
1364 if (N0.getOpcode() == ISD::XOR)
1365 Val = N0.getOperand(0);
1367 assert(N0.getOpcode() == ISD::AND &&
1368 N0.getOperand(0).getOpcode() == ISD::XOR);
1369 // ((X^1)&1)^1 -> X & 1
1370 Val = DAG.getNode(ISD::AND, N0.getValueType(),
1371 N0.getOperand(0).getOperand(0),
1374 return DAG.getSetCC(VT, Val, N1,
1375 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
1380 APInt MinVal, MaxVal;
1381 unsigned OperandBitSize = N1C->getValueType(0).getSizeInBits();
1382 if (ISD::isSignedIntSetCC(Cond)) {
1383 MinVal = APInt::getSignedMinValue(OperandBitSize);
1384 MaxVal = APInt::getSignedMaxValue(OperandBitSize);
1386 MinVal = APInt::getMinValue(OperandBitSize);
1387 MaxVal = APInt::getMaxValue(OperandBitSize);
1390 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
1391 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
1392 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true
1393 // X >= C0 --> X > (C0-1)
1394 return DAG.getSetCC(VT, N0, DAG.getConstant(C1-1, N1.getValueType()),
1395 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
1398 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
1399 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true
1400 // X <= C0 --> X < (C0+1)
1401 return DAG.getSetCC(VT, N0, DAG.getConstant(C1+1, N1.getValueType()),
1402 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
1405 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
1406 return DAG.getConstant(0, VT); // X < MIN --> false
1407 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal)
1408 return DAG.getConstant(1, VT); // X >= MIN --> true
1409 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal)
1410 return DAG.getConstant(0, VT); // X > MAX --> false
1411 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal)
1412 return DAG.getConstant(1, VT); // X <= MAX --> true
1414 // Canonicalize setgt X, Min --> setne X, Min
1415 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
1416 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
1417 // Canonicalize setlt X, Max --> setne X, Max
1418 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
1419 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
1421 // If we have setult X, 1, turn it into seteq X, 0
1422 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
1423 return DAG.getSetCC(VT, N0, DAG.getConstant(MinVal, N0.getValueType()),
1425 // If we have setugt X, Max-1, turn it into seteq X, Max
1426 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
1427 return DAG.getSetCC(VT, N0, DAG.getConstant(MaxVal, N0.getValueType()),
1430 // If we have "setcc X, C0", check to see if we can shrink the immediate
1433 // SETUGT X, SINTMAX -> SETLT X, 0
1434 if (Cond == ISD::SETUGT && OperandBitSize != 1 &&
1435 C1 == (~0ULL >> (65-OperandBitSize)))
1436 return DAG.getSetCC(VT, N0, DAG.getConstant(0, N1.getValueType()),
1439 // FIXME: Implement the rest of these.
1441 // Fold bit comparisons when we can.
1442 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1443 VT == N0.getValueType() && N0.getOpcode() == ISD::AND)
1444 if (ConstantSDNode *AndRHS =
1445 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1446 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
1447 // Perform the xform if the AND RHS is a single bit.
1448 if (isPowerOf2_64(AndRHS->getValue())) {
1449 return DAG.getNode(ISD::SRL, VT, N0,
1450 DAG.getConstant(Log2_64(AndRHS->getValue()),
1451 getShiftAmountTy()));
1453 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getValue()) {
1454 // (X & 8) == 8 --> (X & 8) >> 3
1455 // Perform the xform if C1 is a single bit.
1456 if (C1.isPowerOf2()) {
1457 return DAG.getNode(ISD::SRL, VT, N0,
1458 DAG.getConstant(C1.logBase2(), getShiftAmountTy()));
1463 } else if (isa<ConstantSDNode>(N0.Val)) {
1464 // Ensure that the constant occurs on the RHS.
1465 return DAG.getSetCC(VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
1468 if (isa<ConstantFPSDNode>(N0.Val)) {
1469 // Constant fold or commute setcc.
1470 SDValue O = DAG.FoldSetCC(VT, N0, N1, Cond);
1471 if (O.Val) return O;
1472 } else if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1.Val)) {
1473 // If the RHS of an FP comparison is a constant, simplify it away in
1475 if (CFP->getValueAPF().isNaN()) {
1476 // If an operand is known to be a nan, we can fold it.
1477 switch (ISD::getUnorderedFlavor(Cond)) {
1478 default: assert(0 && "Unknown flavor!");
1479 case 0: // Known false.
1480 return DAG.getConstant(0, VT);
1481 case 1: // Known true.
1482 return DAG.getConstant(1, VT);
1483 case 2: // Undefined.
1484 return DAG.getNode(ISD::UNDEF, VT);
1488 // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the
1489 // constant if knowing that the operand is non-nan is enough. We prefer to
1490 // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to
1492 if (Cond == ISD::SETO || Cond == ISD::SETUO)
1493 return DAG.getSetCC(VT, N0, N0, Cond);
1497 // We can always fold X == X for integer setcc's.
1498 if (N0.getValueType().isInteger())
1499 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
1500 unsigned UOF = ISD::getUnorderedFlavor(Cond);
1501 if (UOF == 2) // FP operators that are undefined on NaNs.
1502 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
1503 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
1504 return DAG.getConstant(UOF, VT);
1505 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
1506 // if it is not already.
1507 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
1508 if (NewCond != Cond)
1509 return DAG.getSetCC(VT, N0, N1, NewCond);
1512 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1513 N0.getValueType().isInteger()) {
1514 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
1515 N0.getOpcode() == ISD::XOR) {
1516 // Simplify (X+Y) == (X+Z) --> Y == Z
1517 if (N0.getOpcode() == N1.getOpcode()) {
1518 if (N0.getOperand(0) == N1.getOperand(0))
1519 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(1), Cond);
1520 if (N0.getOperand(1) == N1.getOperand(1))
1521 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(0), Cond);
1522 if (DAG.isCommutativeBinOp(N0.getOpcode())) {
1523 // If X op Y == Y op X, try other combinations.
1524 if (N0.getOperand(0) == N1.getOperand(1))
1525 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(0), Cond);
1526 if (N0.getOperand(1) == N1.getOperand(0))
1527 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(1), Cond);
1531 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) {
1532 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1533 // Turn (X+C1) == C2 --> X == C2-C1
1534 if (N0.getOpcode() == ISD::ADD && N0.Val->hasOneUse()) {
1535 return DAG.getSetCC(VT, N0.getOperand(0),
1536 DAG.getConstant(RHSC->getValue()-LHSR->getValue(),
1537 N0.getValueType()), Cond);
1540 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
1541 if (N0.getOpcode() == ISD::XOR)
1542 // If we know that all of the inverted bits are zero, don't bother
1543 // performing the inversion.
1544 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue()))
1546 DAG.getSetCC(VT, N0.getOperand(0),
1547 DAG.getConstant(LHSR->getAPIntValue() ^
1548 RHSC->getAPIntValue(),
1553 // Turn (C1-X) == C2 --> X == C1-C2
1554 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
1555 if (N0.getOpcode() == ISD::SUB && N0.Val->hasOneUse()) {
1557 DAG.getSetCC(VT, N0.getOperand(1),
1558 DAG.getConstant(SUBC->getAPIntValue() -
1559 RHSC->getAPIntValue(),
1566 // Simplify (X+Z) == X --> Z == 0
1567 if (N0.getOperand(0) == N1)
1568 return DAG.getSetCC(VT, N0.getOperand(1),
1569 DAG.getConstant(0, N0.getValueType()), Cond);
1570 if (N0.getOperand(1) == N1) {
1571 if (DAG.isCommutativeBinOp(N0.getOpcode()))
1572 return DAG.getSetCC(VT, N0.getOperand(0),
1573 DAG.getConstant(0, N0.getValueType()), Cond);
1574 else if (N0.Val->hasOneUse()) {
1575 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
1576 // (Z-X) == X --> Z == X<<1
1577 SDValue SH = DAG.getNode(ISD::SHL, N1.getValueType(),
1579 DAG.getConstant(1, getShiftAmountTy()));
1580 if (!DCI.isCalledByLegalizer())
1581 DCI.AddToWorklist(SH.Val);
1582 return DAG.getSetCC(VT, N0.getOperand(0), SH, Cond);
1587 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
1588 N1.getOpcode() == ISD::XOR) {
1589 // Simplify X == (X+Z) --> Z == 0
1590 if (N1.getOperand(0) == N0) {
1591 return DAG.getSetCC(VT, N1.getOperand(1),
1592 DAG.getConstant(0, N1.getValueType()), Cond);
1593 } else if (N1.getOperand(1) == N0) {
1594 if (DAG.isCommutativeBinOp(N1.getOpcode())) {
1595 return DAG.getSetCC(VT, N1.getOperand(0),
1596 DAG.getConstant(0, N1.getValueType()), Cond);
1597 } else if (N1.Val->hasOneUse()) {
1598 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
1599 // X == (Z-X) --> X<<1 == Z
1600 SDValue SH = DAG.getNode(ISD::SHL, N1.getValueType(), N0,
1601 DAG.getConstant(1, getShiftAmountTy()));
1602 if (!DCI.isCalledByLegalizer())
1603 DCI.AddToWorklist(SH.Val);
1604 return DAG.getSetCC(VT, SH, N1.getOperand(0), Cond);
1610 // Fold away ALL boolean setcc's.
1612 if (N0.getValueType() == MVT::i1 && foldBooleans) {
1614 default: assert(0 && "Unknown integer setcc!");
1615 case ISD::SETEQ: // X == Y -> (X^Y)^1
1616 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
1617 N0 = DAG.getNode(ISD::XOR, MVT::i1, Temp, DAG.getConstant(1, MVT::i1));
1618 if (!DCI.isCalledByLegalizer())
1619 DCI.AddToWorklist(Temp.Val);
1621 case ISD::SETNE: // X != Y --> (X^Y)
1622 N0 = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
1624 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> X^1 & Y
1625 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> X^1 & Y
1626 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
1627 N0 = DAG.getNode(ISD::AND, MVT::i1, N1, Temp);
1628 if (!DCI.isCalledByLegalizer())
1629 DCI.AddToWorklist(Temp.Val);
1631 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> Y^1 & X
1632 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> Y^1 & X
1633 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
1634 N0 = DAG.getNode(ISD::AND, MVT::i1, N0, Temp);
1635 if (!DCI.isCalledByLegalizer())
1636 DCI.AddToWorklist(Temp.Val);
1638 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> X^1 | Y
1639 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> X^1 | Y
1640 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
1641 N0 = DAG.getNode(ISD::OR, MVT::i1, N1, Temp);
1642 if (!DCI.isCalledByLegalizer())
1643 DCI.AddToWorklist(Temp.Val);
1645 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> Y^1 | X
1646 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> Y^1 | X
1647 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
1648 N0 = DAG.getNode(ISD::OR, MVT::i1, N0, Temp);
1651 if (VT != MVT::i1) {
1652 if (!DCI.isCalledByLegalizer())
1653 DCI.AddToWorklist(N0.Val);
1654 // FIXME: If running after legalize, we probably can't do this.
1655 N0 = DAG.getNode(ISD::ZERO_EXTEND, VT, N0);
1660 // Could not fold it.
1664 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
1665 /// node is a GlobalAddress + offset.
1666 bool TargetLowering::isGAPlusOffset(SDNode *N, GlobalValue* &GA,
1667 int64_t &Offset) const {
1668 if (isa<GlobalAddressSDNode>(N)) {
1669 GlobalAddressSDNode *GASD = cast<GlobalAddressSDNode>(N);
1670 GA = GASD->getGlobal();
1671 Offset += GASD->getOffset();
1675 if (N->getOpcode() == ISD::ADD) {
1676 SDValue N1 = N->getOperand(0);
1677 SDValue N2 = N->getOperand(1);
1678 if (isGAPlusOffset(N1.Val, GA, Offset)) {
1679 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
1681 Offset += V->getSignExtended();
1684 } else if (isGAPlusOffset(N2.Val, GA, Offset)) {
1685 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
1687 Offset += V->getSignExtended();
1696 /// isConsecutiveLoad - Return true if LD (which must be a LoadSDNode) is
1697 /// loading 'Bytes' bytes from a location that is 'Dist' units away from the
1698 /// location that the 'Base' load is loading from.
1699 bool TargetLowering::isConsecutiveLoad(SDNode *LD, SDNode *Base,
1700 unsigned Bytes, int Dist,
1701 const MachineFrameInfo *MFI) const {
1702 if (LD->getOperand(0).Val != Base->getOperand(0).Val)
1704 MVT VT = LD->getValueType(0);
1705 if (VT.getSizeInBits() / 8 != Bytes)
1708 SDValue Loc = LD->getOperand(1);
1709 SDValue BaseLoc = Base->getOperand(1);
1710 if (Loc.getOpcode() == ISD::FrameIndex) {
1711 if (BaseLoc.getOpcode() != ISD::FrameIndex)
1713 int FI = cast<FrameIndexSDNode>(Loc)->getIndex();
1714 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
1715 int FS = MFI->getObjectSize(FI);
1716 int BFS = MFI->getObjectSize(BFI);
1717 if (FS != BFS || FS != (int)Bytes) return false;
1718 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Bytes);
1721 GlobalValue *GV1 = NULL;
1722 GlobalValue *GV2 = NULL;
1723 int64_t Offset1 = 0;
1724 int64_t Offset2 = 0;
1725 bool isGA1 = isGAPlusOffset(Loc.Val, GV1, Offset1);
1726 bool isGA2 = isGAPlusOffset(BaseLoc.Val, GV2, Offset2);
1727 if (isGA1 && isGA2 && GV1 == GV2)
1728 return Offset1 == (Offset2 + Dist*Bytes);
1733 SDValue TargetLowering::
1734 PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
1735 // Default implementation: no optimization.
1739 //===----------------------------------------------------------------------===//
1740 // Inline Assembler Implementation Methods
1741 //===----------------------------------------------------------------------===//
1744 TargetLowering::ConstraintType
1745 TargetLowering::getConstraintType(const std::string &Constraint) const {
1746 // FIXME: lots more standard ones to handle.
1747 if (Constraint.size() == 1) {
1748 switch (Constraint[0]) {
1750 case 'r': return C_RegisterClass;
1752 case 'o': // offsetable
1753 case 'V': // not offsetable
1755 case 'i': // Simple Integer or Relocatable Constant
1756 case 'n': // Simple Integer
1757 case 's': // Relocatable Constant
1758 case 'X': // Allow ANY value.
1759 case 'I': // Target registers.
1771 if (Constraint.size() > 1 && Constraint[0] == '{' &&
1772 Constraint[Constraint.size()-1] == '}')
1777 /// LowerXConstraint - try to replace an X constraint, which matches anything,
1778 /// with another that has more specific requirements based on the type of the
1779 /// corresponding operand.
1780 const char *TargetLowering::LowerXConstraint(MVT ConstraintVT) const{
1781 if (ConstraintVT.isInteger())
1783 if (ConstraintVT.isFloatingPoint())
1784 return "f"; // works for many targets
1788 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
1789 /// vector. If it is invalid, don't add anything to Ops.
1790 void TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
1791 char ConstraintLetter,
1792 std::vector<SDValue> &Ops,
1793 SelectionDAG &DAG) const {
1794 switch (ConstraintLetter) {
1796 case 'X': // Allows any operand; labels (basic block) use this.
1797 if (Op.getOpcode() == ISD::BasicBlock) {
1802 case 'i': // Simple Integer or Relocatable Constant
1803 case 'n': // Simple Integer
1804 case 's': { // Relocatable Constant
1805 // These operands are interested in values of the form (GV+C), where C may
1806 // be folded in as an offset of GV, or it may be explicitly added. Also, it
1807 // is possible and fine if either GV or C are missing.
1808 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
1809 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
1811 // If we have "(add GV, C)", pull out GV/C
1812 if (Op.getOpcode() == ISD::ADD) {
1813 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
1814 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
1815 if (C == 0 || GA == 0) {
1816 C = dyn_cast<ConstantSDNode>(Op.getOperand(0));
1817 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1));
1819 if (C == 0 || GA == 0)
1823 // If we find a valid operand, map to the TargetXXX version so that the
1824 // value itself doesn't get selected.
1825 if (GA) { // Either &GV or &GV+C
1826 if (ConstraintLetter != 'n') {
1827 int64_t Offs = GA->getOffset();
1828 if (C) Offs += C->getValue();
1829 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(),
1830 Op.getValueType(), Offs));
1834 if (C) { // just C, no GV.
1835 // Simple constants are not allowed for 's'.
1836 if (ConstraintLetter != 's') {
1837 Ops.push_back(DAG.getTargetConstant(C->getValue(), Op.getValueType()));
1846 std::vector<unsigned> TargetLowering::
1847 getRegClassForInlineAsmConstraint(const std::string &Constraint,
1849 return std::vector<unsigned>();
1853 std::pair<unsigned, const TargetRegisterClass*> TargetLowering::
1854 getRegForInlineAsmConstraint(const std::string &Constraint,
1856 if (Constraint[0] != '{')
1857 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
1858 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?");
1860 // Remove the braces from around the name.
1861 std::string RegName(Constraint.begin()+1, Constraint.end()-1);
1863 // Figure out which register class contains this reg.
1864 const TargetRegisterInfo *RI = TM.getRegisterInfo();
1865 for (TargetRegisterInfo::regclass_iterator RCI = RI->regclass_begin(),
1866 E = RI->regclass_end(); RCI != E; ++RCI) {
1867 const TargetRegisterClass *RC = *RCI;
1869 // If none of the the value types for this register class are valid, we
1870 // can't use it. For example, 64-bit reg classes on 32-bit targets.
1871 bool isLegal = false;
1872 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
1874 if (isTypeLegal(*I)) {
1880 if (!isLegal) continue;
1882 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end();
1884 if (StringsEqualNoCase(RegName, RI->get(*I).AsmName))
1885 return std::make_pair(*I, RC);
1889 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
1892 //===----------------------------------------------------------------------===//
1893 // Constraint Selection.
1895 /// getConstraintGenerality - Return an integer indicating how general CT
1897 static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) {
1899 default: assert(0 && "Unknown constraint type!");
1900 case TargetLowering::C_Other:
1901 case TargetLowering::C_Unknown:
1903 case TargetLowering::C_Register:
1905 case TargetLowering::C_RegisterClass:
1907 case TargetLowering::C_Memory:
1912 /// ChooseConstraint - If there are multiple different constraints that we
1913 /// could pick for this operand (e.g. "imr") try to pick the 'best' one.
1914 /// This is somewhat tricky: constraints fall into four classes:
1915 /// Other -> immediates and magic values
1916 /// Register -> one specific register
1917 /// RegisterClass -> a group of regs
1918 /// Memory -> memory
1919 /// Ideally, we would pick the most specific constraint possible: if we have
1920 /// something that fits into a register, we would pick it. The problem here
1921 /// is that if we have something that could either be in a register or in
1922 /// memory that use of the register could cause selection of *other*
1923 /// operands to fail: they might only succeed if we pick memory. Because of
1924 /// this the heuristic we use is:
1926 /// 1) If there is an 'other' constraint, and if the operand is valid for
1927 /// that constraint, use it. This makes us take advantage of 'i'
1928 /// constraints when available.
1929 /// 2) Otherwise, pick the most general constraint present. This prefers
1930 /// 'm' over 'r', for example.
1932 static void ChooseConstraint(TargetLowering::AsmOperandInfo &OpInfo,
1933 const TargetLowering &TLI,
1934 SDValue Op, SelectionDAG *DAG) {
1935 assert(OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options");
1936 unsigned BestIdx = 0;
1937 TargetLowering::ConstraintType BestType = TargetLowering::C_Unknown;
1938 int BestGenerality = -1;
1940 // Loop over the options, keeping track of the most general one.
1941 for (unsigned i = 0, e = OpInfo.Codes.size(); i != e; ++i) {
1942 TargetLowering::ConstraintType CType =
1943 TLI.getConstraintType(OpInfo.Codes[i]);
1945 // If this is an 'other' constraint, see if the operand is valid for it.
1946 // For example, on X86 we might have an 'rI' constraint. If the operand
1947 // is an integer in the range [0..31] we want to use I (saving a load
1948 // of a register), otherwise we must use 'r'.
1949 if (CType == TargetLowering::C_Other && Op.Val) {
1950 assert(OpInfo.Codes[i].size() == 1 &&
1951 "Unhandled multi-letter 'other' constraint");
1952 std::vector<SDValue> ResultOps;
1953 TLI.LowerAsmOperandForConstraint(Op, OpInfo.Codes[i][0],
1955 if (!ResultOps.empty()) {
1962 // This constraint letter is more general than the previous one, use it.
1963 int Generality = getConstraintGenerality(CType);
1964 if (Generality > BestGenerality) {
1967 BestGenerality = Generality;
1971 OpInfo.ConstraintCode = OpInfo.Codes[BestIdx];
1972 OpInfo.ConstraintType = BestType;
1975 /// ComputeConstraintToUse - Determines the constraint code and constraint
1976 /// type to use for the specific AsmOperandInfo, setting
1977 /// OpInfo.ConstraintCode and OpInfo.ConstraintType.
1978 void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo,
1980 SelectionDAG *DAG) const {
1981 assert(!OpInfo.Codes.empty() && "Must have at least one constraint");
1983 // Single-letter constraints ('r') are very common.
1984 if (OpInfo.Codes.size() == 1) {
1985 OpInfo.ConstraintCode = OpInfo.Codes[0];
1986 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
1988 ChooseConstraint(OpInfo, *this, Op, DAG);
1991 // 'X' matches anything.
1992 if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) {
1993 // Labels and constants are handled elsewhere ('X' is the only thing
1994 // that matches labels).
1995 if (isa<BasicBlock>(OpInfo.CallOperandVal) ||
1996 isa<ConstantInt>(OpInfo.CallOperandVal))
1999 // Otherwise, try to resolve it to something we know about by looking at
2000 // the actual operand type.
2001 if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) {
2002 OpInfo.ConstraintCode = Repl;
2003 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
2008 //===----------------------------------------------------------------------===//
2009 // Loop Strength Reduction hooks
2010 //===----------------------------------------------------------------------===//
2012 /// isLegalAddressingMode - Return true if the addressing mode represented
2013 /// by AM is legal for this target, for a load/store of the specified type.
2014 bool TargetLowering::isLegalAddressingMode(const AddrMode &AM,
2015 const Type *Ty) const {
2016 // The default implementation of this implements a conservative RISCy, r+r and
2019 // Allows a sign-extended 16-bit immediate field.
2020 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
2023 // No global is ever allowed as a base.
2027 // Only support r+r,
2029 case 0: // "r+i" or just "i", depending on HasBaseReg.
2032 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
2034 // Otherwise we have r+r or r+i.
2037 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
2039 // Allow 2*r as r+r.
2046 // Magic for divide replacement
2049 int64_t m; // magic number
2050 int64_t s; // shift amount
2054 uint64_t m; // magic number
2055 int64_t a; // add indicator
2056 int64_t s; // shift amount
2059 /// magic - calculate the magic numbers required to codegen an integer sdiv as
2060 /// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
2062 static ms magic32(int32_t d) {
2064 uint32_t ad, anc, delta, q1, r1, q2, r2, t;
2065 const uint32_t two31 = 0x80000000U;
2069 t = two31 + ((uint32_t)d >> 31);
2070 anc = t - 1 - t%ad; // absolute value of nc
2071 p = 31; // initialize p
2072 q1 = two31/anc; // initialize q1 = 2p/abs(nc)
2073 r1 = two31 - q1*anc; // initialize r1 = rem(2p,abs(nc))
2074 q2 = two31/ad; // initialize q2 = 2p/abs(d)
2075 r2 = two31 - q2*ad; // initialize r2 = rem(2p,abs(d))
2078 q1 = 2*q1; // update q1 = 2p/abs(nc)
2079 r1 = 2*r1; // update r1 = rem(2p/abs(nc))
2080 if (r1 >= anc) { // must be unsigned comparison
2084 q2 = 2*q2; // update q2 = 2p/abs(d)
2085 r2 = 2*r2; // update r2 = rem(2p/abs(d))
2086 if (r2 >= ad) { // must be unsigned comparison
2091 } while (q1 < delta || (q1 == delta && r1 == 0));
2093 mag.m = (int32_t)(q2 + 1); // make sure to sign extend
2094 if (d < 0) mag.m = -mag.m; // resulting magic number
2095 mag.s = p - 32; // resulting shift
2099 /// magicu - calculate the magic numbers required to codegen an integer udiv as
2100 /// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
2101 static mu magicu32(uint32_t d) {
2103 uint32_t nc, delta, q1, r1, q2, r2;
2105 magu.a = 0; // initialize "add" indicator
2107 p = 31; // initialize p
2108 q1 = 0x80000000/nc; // initialize q1 = 2p/nc
2109 r1 = 0x80000000 - q1*nc; // initialize r1 = rem(2p,nc)
2110 q2 = 0x7FFFFFFF/d; // initialize q2 = (2p-1)/d
2111 r2 = 0x7FFFFFFF - q2*d; // initialize r2 = rem((2p-1),d)
2114 if (r1 >= nc - r1 ) {
2115 q1 = 2*q1 + 1; // update q1
2116 r1 = 2*r1 - nc; // update r1
2119 q1 = 2*q1; // update q1
2120 r1 = 2*r1; // update r1
2122 if (r2 + 1 >= d - r2) {
2123 if (q2 >= 0x7FFFFFFF) magu.a = 1;
2124 q2 = 2*q2 + 1; // update q2
2125 r2 = 2*r2 + 1 - d; // update r2
2128 if (q2 >= 0x80000000) magu.a = 1;
2129 q2 = 2*q2; // update q2
2130 r2 = 2*r2 + 1; // update r2
2133 } while (p < 64 && (q1 < delta || (q1 == delta && r1 == 0)));
2134 magu.m = q2 + 1; // resulting magic number
2135 magu.s = p - 32; // resulting shift
2139 /// magic - calculate the magic numbers required to codegen an integer sdiv as
2140 /// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
2142 static ms magic64(int64_t d) {
2144 uint64_t ad, anc, delta, q1, r1, q2, r2, t;
2145 const uint64_t two63 = 9223372036854775808ULL; // 2^63
2148 ad = d >= 0 ? d : -d;
2149 t = two63 + ((uint64_t)d >> 63);
2150 anc = t - 1 - t%ad; // absolute value of nc
2151 p = 63; // initialize p
2152 q1 = two63/anc; // initialize q1 = 2p/abs(nc)
2153 r1 = two63 - q1*anc; // initialize r1 = rem(2p,abs(nc))
2154 q2 = two63/ad; // initialize q2 = 2p/abs(d)
2155 r2 = two63 - q2*ad; // initialize r2 = rem(2p,abs(d))
2158 q1 = 2*q1; // update q1 = 2p/abs(nc)
2159 r1 = 2*r1; // update r1 = rem(2p/abs(nc))
2160 if (r1 >= anc) { // must be unsigned comparison
2164 q2 = 2*q2; // update q2 = 2p/abs(d)
2165 r2 = 2*r2; // update r2 = rem(2p/abs(d))
2166 if (r2 >= ad) { // must be unsigned comparison
2171 } while (q1 < delta || (q1 == delta && r1 == 0));
2174 if (d < 0) mag.m = -mag.m; // resulting magic number
2175 mag.s = p - 64; // resulting shift
2179 /// magicu - calculate the magic numbers required to codegen an integer udiv as
2180 /// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
2181 static mu magicu64(uint64_t d)
2184 uint64_t nc, delta, q1, r1, q2, r2;
2186 magu.a = 0; // initialize "add" indicator
2188 p = 63; // initialize p
2189 q1 = 0x8000000000000000ull/nc; // initialize q1 = 2p/nc
2190 r1 = 0x8000000000000000ull - q1*nc; // initialize r1 = rem(2p,nc)
2191 q2 = 0x7FFFFFFFFFFFFFFFull/d; // initialize q2 = (2p-1)/d
2192 r2 = 0x7FFFFFFFFFFFFFFFull - q2*d; // initialize r2 = rem((2p-1),d)
2195 if (r1 >= nc - r1 ) {
2196 q1 = 2*q1 + 1; // update q1
2197 r1 = 2*r1 - nc; // update r1
2200 q1 = 2*q1; // update q1
2201 r1 = 2*r1; // update r1
2203 if (r2 + 1 >= d - r2) {
2204 if (q2 >= 0x7FFFFFFFFFFFFFFFull) magu.a = 1;
2205 q2 = 2*q2 + 1; // update q2
2206 r2 = 2*r2 + 1 - d; // update r2
2209 if (q2 >= 0x8000000000000000ull) magu.a = 1;
2210 q2 = 2*q2; // update q2
2211 r2 = 2*r2 + 1; // update r2
2214 } while (p < 128 && (q1 < delta || (q1 == delta && r1 == 0)));
2215 magu.m = q2 + 1; // resulting magic number
2216 magu.s = p - 64; // resulting shift
2220 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
2221 /// return a DAG expression to select that will generate the same value by
2222 /// multiplying by a magic number. See:
2223 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
2224 SDValue TargetLowering::BuildSDIV(SDNode *N, SelectionDAG &DAG,
2225 std::vector<SDNode*>* Created) const {
2226 MVT VT = N->getValueType(0);
2228 // Check to see if we can do this.
2229 if (!isTypeLegal(VT) || (VT != MVT::i32 && VT != MVT::i64))
2230 return SDValue(); // BuildSDIV only operates on i32 or i64
2232 int64_t d = cast<ConstantSDNode>(N->getOperand(1))->getSignExtended();
2233 ms magics = (VT == MVT::i32) ? magic32(d) : magic64(d);
2235 // Multiply the numerator (operand 0) by the magic value
2237 if (isOperationLegal(ISD::MULHS, VT))
2238 Q = DAG.getNode(ISD::MULHS, VT, N->getOperand(0),
2239 DAG.getConstant(magics.m, VT));
2240 else if (isOperationLegal(ISD::SMUL_LOHI, VT))
2241 Q = SDValue(DAG.getNode(ISD::SMUL_LOHI, DAG.getVTList(VT, VT),
2243 DAG.getConstant(magics.m, VT)).Val, 1);
2245 return SDValue(); // No mulhs or equvialent
2246 // If d > 0 and m < 0, add the numerator
2247 if (d > 0 && magics.m < 0) {
2248 Q = DAG.getNode(ISD::ADD, VT, Q, N->getOperand(0));
2250 Created->push_back(Q.Val);
2252 // If d < 0 and m > 0, subtract the numerator.
2253 if (d < 0 && magics.m > 0) {
2254 Q = DAG.getNode(ISD::SUB, VT, Q, N->getOperand(0));
2256 Created->push_back(Q.Val);
2258 // Shift right algebraic if shift value is nonzero
2260 Q = DAG.getNode(ISD::SRA, VT, Q,
2261 DAG.getConstant(magics.s, getShiftAmountTy()));
2263 Created->push_back(Q.Val);
2265 // Extract the sign bit and add it to the quotient
2267 DAG.getNode(ISD::SRL, VT, Q, DAG.getConstant(VT.getSizeInBits()-1,
2268 getShiftAmountTy()));
2270 Created->push_back(T.Val);
2271 return DAG.getNode(ISD::ADD, VT, Q, T);
2274 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
2275 /// return a DAG expression to select that will generate the same value by
2276 /// multiplying by a magic number. See:
2277 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
2278 SDValue TargetLowering::BuildUDIV(SDNode *N, SelectionDAG &DAG,
2279 std::vector<SDNode*>* Created) const {
2280 MVT VT = N->getValueType(0);
2282 // Check to see if we can do this.
2283 if (!isTypeLegal(VT) || (VT != MVT::i32 && VT != MVT::i64))
2284 return SDValue(); // BuildUDIV only operates on i32 or i64
2286 uint64_t d = cast<ConstantSDNode>(N->getOperand(1))->getValue();
2287 mu magics = (VT == MVT::i32) ? magicu32(d) : magicu64(d);
2289 // Multiply the numerator (operand 0) by the magic value
2291 if (isOperationLegal(ISD::MULHU, VT))
2292 Q = DAG.getNode(ISD::MULHU, VT, N->getOperand(0),
2293 DAG.getConstant(magics.m, VT));
2294 else if (isOperationLegal(ISD::UMUL_LOHI, VT))
2295 Q = SDValue(DAG.getNode(ISD::UMUL_LOHI, DAG.getVTList(VT, VT),
2297 DAG.getConstant(magics.m, VT)).Val, 1);
2299 return SDValue(); // No mulhu or equvialent
2301 Created->push_back(Q.Val);
2303 if (magics.a == 0) {
2304 return DAG.getNode(ISD::SRL, VT, Q,
2305 DAG.getConstant(magics.s, getShiftAmountTy()));
2307 SDValue NPQ = DAG.getNode(ISD::SUB, VT, N->getOperand(0), Q);
2309 Created->push_back(NPQ.Val);
2310 NPQ = DAG.getNode(ISD::SRL, VT, NPQ,
2311 DAG.getConstant(1, getShiftAmountTy()));
2313 Created->push_back(NPQ.Val);
2314 NPQ = DAG.getNode(ISD::ADD, VT, NPQ, Q);
2316 Created->push_back(NPQ.Val);
2317 return DAG.getNode(ISD::SRL, VT, NPQ,
2318 DAG.getConstant(magics.s-1, getShiftAmountTy()));