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::CEIL_F32] = "ceilf";
99 Names[RTLIB::CEIL_F64] = "ceil";
100 Names[RTLIB::CEIL_F80] = "ceill";
101 Names[RTLIB::CEIL_PPCF128] = "ceill";
102 Names[RTLIB::TRUNC_F32] = "truncf";
103 Names[RTLIB::TRUNC_F64] = "trunc";
104 Names[RTLIB::TRUNC_F80] = "truncl";
105 Names[RTLIB::TRUNC_PPCF128] = "truncl";
106 Names[RTLIB::RINT_F32] = "rintf";
107 Names[RTLIB::RINT_F64] = "rint";
108 Names[RTLIB::RINT_F80] = "rintl";
109 Names[RTLIB::RINT_PPCF128] = "rintl";
110 Names[RTLIB::NEARBYINT_F32] = "nearbyintf";
111 Names[RTLIB::NEARBYINT_F64] = "nearbyint";
112 Names[RTLIB::NEARBYINT_F80] = "nearbyintl";
113 Names[RTLIB::NEARBYINT_PPCF128] = "nearbyintl";
114 Names[RTLIB::FLOOR_F32] = "floorf";
115 Names[RTLIB::FLOOR_F64] = "floor";
116 Names[RTLIB::FLOOR_F80] = "floorl";
117 Names[RTLIB::FLOOR_PPCF128] = "floorl";
118 Names[RTLIB::FPEXT_F32_F64] = "__extendsfdf2";
119 Names[RTLIB::FPROUND_F64_F32] = "__truncdfsf2";
120 Names[RTLIB::FPROUND_F80_F32] = "__truncxfsf2";
121 Names[RTLIB::FPROUND_PPCF128_F32] = "__trunctfsf2";
122 Names[RTLIB::FPROUND_F80_F64] = "__truncxfdf2";
123 Names[RTLIB::FPROUND_PPCF128_F64] = "__trunctfdf2";
124 Names[RTLIB::FPTOSINT_F32_I32] = "__fixsfsi";
125 Names[RTLIB::FPTOSINT_F32_I64] = "__fixsfdi";
126 Names[RTLIB::FPTOSINT_F32_I128] = "__fixsfti";
127 Names[RTLIB::FPTOSINT_F64_I32] = "__fixdfsi";
128 Names[RTLIB::FPTOSINT_F64_I64] = "__fixdfdi";
129 Names[RTLIB::FPTOSINT_F64_I128] = "__fixdfti";
130 Names[RTLIB::FPTOSINT_F80_I32] = "__fixxfsi";
131 Names[RTLIB::FPTOSINT_F80_I64] = "__fixxfdi";
132 Names[RTLIB::FPTOSINT_F80_I128] = "__fixxfti";
133 Names[RTLIB::FPTOSINT_PPCF128_I32] = "__fixtfsi";
134 Names[RTLIB::FPTOSINT_PPCF128_I64] = "__fixtfdi";
135 Names[RTLIB::FPTOSINT_PPCF128_I128] = "__fixtfti";
136 Names[RTLIB::FPTOUINT_F32_I32] = "__fixunssfsi";
137 Names[RTLIB::FPTOUINT_F32_I64] = "__fixunssfdi";
138 Names[RTLIB::FPTOUINT_F32_I128] = "__fixunssfti";
139 Names[RTLIB::FPTOUINT_F64_I32] = "__fixunsdfsi";
140 Names[RTLIB::FPTOUINT_F64_I64] = "__fixunsdfdi";
141 Names[RTLIB::FPTOUINT_F64_I128] = "__fixunsdfti";
142 Names[RTLIB::FPTOUINT_F80_I32] = "__fixunsxfsi";
143 Names[RTLIB::FPTOUINT_F80_I64] = "__fixunsxfdi";
144 Names[RTLIB::FPTOUINT_F80_I128] = "__fixunsxfti";
145 Names[RTLIB::FPTOUINT_PPCF128_I32] = "__fixunstfsi";
146 Names[RTLIB::FPTOUINT_PPCF128_I64] = "__fixunstfdi";
147 Names[RTLIB::FPTOUINT_PPCF128_I128] = "__fixunstfti";
148 Names[RTLIB::SINTTOFP_I32_F32] = "__floatsisf";
149 Names[RTLIB::SINTTOFP_I32_F64] = "__floatsidf";
150 Names[RTLIB::SINTTOFP_I32_F80] = "__floatsixf";
151 Names[RTLIB::SINTTOFP_I32_PPCF128] = "__floatsitf";
152 Names[RTLIB::SINTTOFP_I64_F32] = "__floatdisf";
153 Names[RTLIB::SINTTOFP_I64_F64] = "__floatdidf";
154 Names[RTLIB::SINTTOFP_I64_F80] = "__floatdixf";
155 Names[RTLIB::SINTTOFP_I64_PPCF128] = "__floatditf";
156 Names[RTLIB::SINTTOFP_I128_F32] = "__floattisf";
157 Names[RTLIB::SINTTOFP_I128_F64] = "__floattidf";
158 Names[RTLIB::SINTTOFP_I128_F80] = "__floattixf";
159 Names[RTLIB::SINTTOFP_I128_PPCF128] = "__floattitf";
160 Names[RTLIB::UINTTOFP_I32_F32] = "__floatunsisf";
161 Names[RTLIB::UINTTOFP_I32_F64] = "__floatunsidf";
162 Names[RTLIB::UINTTOFP_I32_F80] = "__floatunsixf";
163 Names[RTLIB::UINTTOFP_I32_PPCF128] = "__floatunsitf";
164 Names[RTLIB::UINTTOFP_I64_F32] = "__floatundisf";
165 Names[RTLIB::UINTTOFP_I64_F64] = "__floatundidf";
166 Names[RTLIB::UINTTOFP_I64_F80] = "__floatundixf";
167 Names[RTLIB::UINTTOFP_I64_PPCF128] = "__floatunditf";
168 Names[RTLIB::UINTTOFP_I128_F32] = "__floatuntisf";
169 Names[RTLIB::UINTTOFP_I128_F64] = "__floatuntidf";
170 Names[RTLIB::UINTTOFP_I128_F80] = "__floatuntixf";
171 Names[RTLIB::UINTTOFP_I128_PPCF128] = "__floatuntitf";
172 Names[RTLIB::OEQ_F32] = "__eqsf2";
173 Names[RTLIB::OEQ_F64] = "__eqdf2";
174 Names[RTLIB::UNE_F32] = "__nesf2";
175 Names[RTLIB::UNE_F64] = "__nedf2";
176 Names[RTLIB::OGE_F32] = "__gesf2";
177 Names[RTLIB::OGE_F64] = "__gedf2";
178 Names[RTLIB::OLT_F32] = "__ltsf2";
179 Names[RTLIB::OLT_F64] = "__ltdf2";
180 Names[RTLIB::OLE_F32] = "__lesf2";
181 Names[RTLIB::OLE_F64] = "__ledf2";
182 Names[RTLIB::OGT_F32] = "__gtsf2";
183 Names[RTLIB::OGT_F64] = "__gtdf2";
184 Names[RTLIB::UO_F32] = "__unordsf2";
185 Names[RTLIB::UO_F64] = "__unorddf2";
186 Names[RTLIB::O_F32] = "__unordsf2";
187 Names[RTLIB::O_F64] = "__unorddf2";
190 /// getFPEXT - Return the FPEXT_*_* value for the given types, or
191 /// UNKNOWN_LIBCALL if there is none.
192 RTLIB::Libcall RTLIB::getFPEXT(MVT OpVT, MVT RetVT) {
193 if (OpVT == MVT::f32) {
194 if (RetVT == MVT::f64)
195 return FPEXT_F32_F64;
197 return UNKNOWN_LIBCALL;
200 /// getFPROUND - Return the FPROUND_*_* value for the given types, or
201 /// UNKNOWN_LIBCALL if there is none.
202 RTLIB::Libcall RTLIB::getFPROUND(MVT OpVT, MVT RetVT) {
203 if (RetVT == MVT::f32) {
204 if (OpVT == MVT::f64)
205 return FPROUND_F64_F32;
206 if (OpVT == MVT::f80)
207 return FPROUND_F80_F32;
208 if (OpVT == MVT::ppcf128)
209 return FPROUND_PPCF128_F32;
210 } else if (RetVT == MVT::f64) {
211 if (OpVT == MVT::f80)
212 return FPROUND_F80_F64;
213 if (OpVT == MVT::ppcf128)
214 return FPROUND_PPCF128_F64;
216 return UNKNOWN_LIBCALL;
219 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or
220 /// UNKNOWN_LIBCALL if there is none.
221 RTLIB::Libcall RTLIB::getFPTOSINT(MVT OpVT, MVT RetVT) {
222 if (OpVT == MVT::f32) {
223 if (RetVT == MVT::i32)
224 return FPTOSINT_F32_I32;
225 if (RetVT == MVT::i64)
226 return FPTOSINT_F32_I64;
227 if (RetVT == MVT::i128)
228 return FPTOSINT_F32_I128;
229 } else if (OpVT == MVT::f64) {
230 if (RetVT == MVT::i32)
231 return FPTOSINT_F64_I32;
232 if (RetVT == MVT::i64)
233 return FPTOSINT_F64_I64;
234 if (RetVT == MVT::i128)
235 return FPTOSINT_F64_I128;
236 } else if (OpVT == MVT::f80) {
237 if (RetVT == MVT::i32)
238 return FPTOSINT_F80_I32;
239 if (RetVT == MVT::i64)
240 return FPTOSINT_F80_I64;
241 if (RetVT == MVT::i128)
242 return FPTOSINT_F80_I128;
243 } else if (OpVT == MVT::ppcf128) {
244 if (RetVT == MVT::i32)
245 return FPTOSINT_PPCF128_I32;
246 if (RetVT == MVT::i64)
247 return FPTOSINT_PPCF128_I64;
248 if (RetVT == MVT::i128)
249 return FPTOSINT_PPCF128_I128;
251 return UNKNOWN_LIBCALL;
254 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or
255 /// UNKNOWN_LIBCALL if there is none.
256 RTLIB::Libcall RTLIB::getFPTOUINT(MVT OpVT, MVT RetVT) {
257 if (OpVT == MVT::f32) {
258 if (RetVT == MVT::i32)
259 return FPTOUINT_F32_I32;
260 if (RetVT == MVT::i64)
261 return FPTOUINT_F32_I64;
262 if (RetVT == MVT::i128)
263 return FPTOUINT_F32_I128;
264 } else if (OpVT == MVT::f64) {
265 if (RetVT == MVT::i32)
266 return FPTOUINT_F64_I32;
267 if (RetVT == MVT::i64)
268 return FPTOUINT_F64_I64;
269 if (RetVT == MVT::i128)
270 return FPTOUINT_F64_I128;
271 } else if (OpVT == MVT::f80) {
272 if (RetVT == MVT::i32)
273 return FPTOUINT_F80_I32;
274 if (RetVT == MVT::i64)
275 return FPTOUINT_F80_I64;
276 if (RetVT == MVT::i128)
277 return FPTOUINT_F80_I128;
278 } else if (OpVT == MVT::ppcf128) {
279 if (RetVT == MVT::i32)
280 return FPTOUINT_PPCF128_I32;
281 if (RetVT == MVT::i64)
282 return FPTOUINT_PPCF128_I64;
283 if (RetVT == MVT::i128)
284 return FPTOUINT_PPCF128_I128;
286 return UNKNOWN_LIBCALL;
289 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or
290 /// UNKNOWN_LIBCALL if there is none.
291 RTLIB::Libcall RTLIB::getSINTTOFP(MVT OpVT, MVT RetVT) {
292 if (OpVT == MVT::i32) {
293 if (RetVT == MVT::f32)
294 return SINTTOFP_I32_F32;
295 else if (RetVT == MVT::f64)
296 return SINTTOFP_I32_F64;
297 else if (RetVT == MVT::f80)
298 return SINTTOFP_I32_F80;
299 else if (RetVT == MVT::ppcf128)
300 return SINTTOFP_I32_PPCF128;
301 } else if (OpVT == MVT::i64) {
302 if (RetVT == MVT::f32)
303 return SINTTOFP_I64_F32;
304 else if (RetVT == MVT::f64)
305 return SINTTOFP_I64_F64;
306 else if (RetVT == MVT::f80)
307 return SINTTOFP_I64_F80;
308 else if (RetVT == MVT::ppcf128)
309 return SINTTOFP_I64_PPCF128;
310 } else if (OpVT == MVT::i128) {
311 if (RetVT == MVT::f32)
312 return SINTTOFP_I128_F32;
313 else if (RetVT == MVT::f64)
314 return SINTTOFP_I128_F64;
315 else if (RetVT == MVT::f80)
316 return SINTTOFP_I128_F80;
317 else if (RetVT == MVT::ppcf128)
318 return SINTTOFP_I128_PPCF128;
320 return UNKNOWN_LIBCALL;
323 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or
324 /// UNKNOWN_LIBCALL if there is none.
325 RTLIB::Libcall RTLIB::getUINTTOFP(MVT OpVT, MVT RetVT) {
326 if (OpVT == MVT::i32) {
327 if (RetVT == MVT::f32)
328 return UINTTOFP_I32_F32;
329 else if (RetVT == MVT::f64)
330 return UINTTOFP_I32_F64;
331 else if (RetVT == MVT::f80)
332 return UINTTOFP_I32_F80;
333 else if (RetVT == MVT::ppcf128)
334 return UINTTOFP_I32_PPCF128;
335 } else if (OpVT == MVT::i64) {
336 if (RetVT == MVT::f32)
337 return UINTTOFP_I64_F32;
338 else if (RetVT == MVT::f64)
339 return UINTTOFP_I64_F64;
340 else if (RetVT == MVT::f80)
341 return UINTTOFP_I64_F80;
342 else if (RetVT == MVT::ppcf128)
343 return UINTTOFP_I64_PPCF128;
344 } else if (OpVT == MVT::i128) {
345 if (RetVT == MVT::f32)
346 return UINTTOFP_I128_F32;
347 else if (RetVT == MVT::f64)
348 return UINTTOFP_I128_F64;
349 else if (RetVT == MVT::f80)
350 return UINTTOFP_I128_F80;
351 else if (RetVT == MVT::ppcf128)
352 return UINTTOFP_I128_PPCF128;
354 return UNKNOWN_LIBCALL;
357 /// InitCmpLibcallCCs - Set default comparison libcall CC.
359 static void InitCmpLibcallCCs(ISD::CondCode *CCs) {
360 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL);
361 CCs[RTLIB::OEQ_F32] = ISD::SETEQ;
362 CCs[RTLIB::OEQ_F64] = ISD::SETEQ;
363 CCs[RTLIB::UNE_F32] = ISD::SETNE;
364 CCs[RTLIB::UNE_F64] = ISD::SETNE;
365 CCs[RTLIB::OGE_F32] = ISD::SETGE;
366 CCs[RTLIB::OGE_F64] = ISD::SETGE;
367 CCs[RTLIB::OLT_F32] = ISD::SETLT;
368 CCs[RTLIB::OLT_F64] = ISD::SETLT;
369 CCs[RTLIB::OLE_F32] = ISD::SETLE;
370 CCs[RTLIB::OLE_F64] = ISD::SETLE;
371 CCs[RTLIB::OGT_F32] = ISD::SETGT;
372 CCs[RTLIB::OGT_F64] = ISD::SETGT;
373 CCs[RTLIB::UO_F32] = ISD::SETNE;
374 CCs[RTLIB::UO_F64] = ISD::SETNE;
375 CCs[RTLIB::O_F32] = ISD::SETEQ;
376 CCs[RTLIB::O_F64] = ISD::SETEQ;
379 TargetLowering::TargetLowering(TargetMachine &tm)
380 : TM(tm), TD(TM.getTargetData()) {
381 assert(ISD::BUILTIN_OP_END <= OpActionsCapacity &&
382 "Fixed size array in TargetLowering is not large enough!");
383 // All operations default to being supported.
384 memset(OpActions, 0, sizeof(OpActions));
385 memset(LoadXActions, 0, sizeof(LoadXActions));
386 memset(TruncStoreActions, 0, sizeof(TruncStoreActions));
387 memset(IndexedModeActions, 0, sizeof(IndexedModeActions));
388 memset(ConvertActions, 0, sizeof(ConvertActions));
390 // Set default actions for various operations.
391 for (unsigned VT = 0; VT != (unsigned)MVT::LAST_VALUETYPE; ++VT) {
392 // Default all indexed load / store to expand.
393 for (unsigned IM = (unsigned)ISD::PRE_INC;
394 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
395 setIndexedLoadAction(IM, (MVT::SimpleValueType)VT, Expand);
396 setIndexedStoreAction(IM, (MVT::SimpleValueType)VT, Expand);
399 // These operations default to expand.
400 setOperationAction(ISD::FGETSIGN, (MVT::SimpleValueType)VT, Expand);
403 // Most targets ignore the @llvm.prefetch intrinsic.
404 setOperationAction(ISD::PREFETCH, MVT::Other, Expand);
406 // ConstantFP nodes default to expand. Targets can either change this to
407 // Legal, in which case all fp constants are legal, or use addLegalFPImmediate
408 // to optimize expansions for certain constants.
409 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
410 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
411 setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
413 // Default ISD::TRAP to expand (which turns it into abort).
414 setOperationAction(ISD::TRAP, MVT::Other, Expand);
416 IsLittleEndian = TD->isLittleEndian();
417 UsesGlobalOffsetTable = false;
418 ShiftAmountTy = PointerTy = getValueType(TD->getIntPtrType());
419 ShiftAmtHandling = Undefined;
420 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE*sizeof(TargetRegisterClass*));
421 memset(TargetDAGCombineArray, 0, array_lengthof(TargetDAGCombineArray));
422 maxStoresPerMemset = maxStoresPerMemcpy = maxStoresPerMemmove = 8;
423 allowUnalignedMemoryAccesses = false;
424 UseUnderscoreSetJmp = false;
425 UseUnderscoreLongJmp = false;
426 SelectIsExpensive = false;
427 IntDivIsCheap = false;
428 Pow2DivIsCheap = false;
429 StackPointerRegisterToSaveRestore = 0;
430 ExceptionPointerRegister = 0;
431 ExceptionSelectorRegister = 0;
432 SetCCResultContents = UndefinedSetCCResult;
433 SchedPreferenceInfo = SchedulingForLatency;
435 JumpBufAlignment = 0;
436 IfCvtBlockSizeLimit = 2;
437 IfCvtDupBlockSizeLimit = 0;
438 PrefLoopAlignment = 0;
440 InitLibcallNames(LibcallRoutineNames);
441 InitCmpLibcallCCs(CmpLibcallCCs);
443 // Tell Legalize whether the assembler supports DEBUG_LOC.
444 if (!TM.getTargetAsmInfo()->hasDotLocAndDotFile())
445 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
448 TargetLowering::~TargetLowering() {}
450 /// computeRegisterProperties - Once all of the register classes are added,
451 /// this allows us to compute derived properties we expose.
452 void TargetLowering::computeRegisterProperties() {
453 assert(MVT::LAST_VALUETYPE <= 32 &&
454 "Too many value types for ValueTypeActions to hold!");
456 // Everything defaults to needing one register.
457 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
458 NumRegistersForVT[i] = 1;
459 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i;
461 // ...except isVoid, which doesn't need any registers.
462 NumRegistersForVT[MVT::isVoid] = 0;
464 // Find the largest integer register class.
465 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE;
466 for (; RegClassForVT[LargestIntReg] == 0; --LargestIntReg)
467 assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
469 // Every integer value type larger than this largest register takes twice as
470 // many registers to represent as the previous ValueType.
471 for (unsigned ExpandedReg = LargestIntReg + 1; ; ++ExpandedReg) {
472 MVT EVT = (MVT::SimpleValueType)ExpandedReg;
473 if (!EVT.isInteger())
475 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
476 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg;
477 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1);
478 ValueTypeActions.setTypeAction(EVT, Expand);
481 // Inspect all of the ValueType's smaller than the largest integer
482 // register to see which ones need promotion.
483 unsigned LegalIntReg = LargestIntReg;
484 for (unsigned IntReg = LargestIntReg - 1;
485 IntReg >= (unsigned)MVT::i1; --IntReg) {
486 MVT IVT = (MVT::SimpleValueType)IntReg;
487 if (isTypeLegal(IVT)) {
488 LegalIntReg = IntReg;
490 RegisterTypeForVT[IntReg] = TransformToType[IntReg] =
491 (MVT::SimpleValueType)LegalIntReg;
492 ValueTypeActions.setTypeAction(IVT, Promote);
496 // ppcf128 type is really two f64's.
497 if (!isTypeLegal(MVT::ppcf128)) {
498 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64];
499 RegisterTypeForVT[MVT::ppcf128] = MVT::f64;
500 TransformToType[MVT::ppcf128] = MVT::f64;
501 ValueTypeActions.setTypeAction(MVT::ppcf128, Expand);
504 // Decide how to handle f64. If the target does not have native f64 support,
505 // expand it to i64 and we will be generating soft float library calls.
506 if (!isTypeLegal(MVT::f64)) {
507 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
508 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
509 TransformToType[MVT::f64] = MVT::i64;
510 ValueTypeActions.setTypeAction(MVT::f64, Expand);
513 // Decide how to handle f32. If the target does not have native support for
514 // f32, promote it to f64 if it is legal. Otherwise, expand it to i32.
515 if (!isTypeLegal(MVT::f32)) {
516 if (isTypeLegal(MVT::f64)) {
517 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::f64];
518 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::f64];
519 TransformToType[MVT::f32] = MVT::f64;
520 ValueTypeActions.setTypeAction(MVT::f32, Promote);
522 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
523 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
524 TransformToType[MVT::f32] = MVT::i32;
525 ValueTypeActions.setTypeAction(MVT::f32, Expand);
529 // Loop over all of the vector value types to see which need transformations.
530 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
531 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
532 MVT VT = (MVT::SimpleValueType)i;
533 if (!isTypeLegal(VT)) {
534 MVT IntermediateVT, RegisterVT;
535 unsigned NumIntermediates;
536 NumRegistersForVT[i] =
537 getVectorTypeBreakdown(VT,
538 IntermediateVT, NumIntermediates,
540 RegisterTypeForVT[i] = RegisterVT;
541 TransformToType[i] = MVT::Other; // this isn't actually used
542 ValueTypeActions.setTypeAction(VT, Expand);
547 const char *TargetLowering::getTargetNodeName(unsigned Opcode) const {
552 MVT TargetLowering::getSetCCResultType(const SDValue &) const {
553 return getValueType(TD->getIntPtrType());
557 /// getVectorTypeBreakdown - Vector types are broken down into some number of
558 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
559 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
560 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
562 /// This method returns the number of registers needed, and the VT for each
563 /// register. It also returns the VT and quantity of the intermediate values
564 /// before they are promoted/expanded.
566 unsigned TargetLowering::getVectorTypeBreakdown(MVT VT,
568 unsigned &NumIntermediates,
569 MVT &RegisterVT) const {
570 // Figure out the right, legal destination reg to copy into.
571 unsigned NumElts = VT.getVectorNumElements();
572 MVT EltTy = VT.getVectorElementType();
574 unsigned NumVectorRegs = 1;
576 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
577 // could break down into LHS/RHS like LegalizeDAG does.
578 if (!isPowerOf2_32(NumElts)) {
579 NumVectorRegs = NumElts;
583 // Divide the input until we get to a supported size. This will always
584 // end with a scalar if the target doesn't support vectors.
585 while (NumElts > 1 && !isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) {
590 NumIntermediates = NumVectorRegs;
592 MVT NewVT = MVT::getVectorVT(EltTy, NumElts);
593 if (!isTypeLegal(NewVT))
595 IntermediateVT = NewVT;
597 MVT DestVT = getTypeToTransformTo(NewVT);
599 if (DestVT.bitsLT(NewVT)) {
600 // Value is expanded, e.g. i64 -> i16.
601 return NumVectorRegs*(NewVT.getSizeInBits()/DestVT.getSizeInBits());
603 // Otherwise, promotion or legal types use the same number of registers as
604 // the vector decimated to the appropriate level.
605 return NumVectorRegs;
611 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
612 /// function arguments in the caller parameter area. This is the actual
613 /// alignment, not its logarithm.
614 unsigned TargetLowering::getByValTypeAlignment(const Type *Ty) const {
615 return TD->getCallFrameTypeAlignment(Ty);
618 SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table,
619 SelectionDAG &DAG) const {
620 if (usesGlobalOffsetTable())
621 return DAG.getNode(ISD::GLOBAL_OFFSET_TABLE, getPointerTy());
625 //===----------------------------------------------------------------------===//
626 // Optimization Methods
627 //===----------------------------------------------------------------------===//
629 /// ShrinkDemandedConstant - Check to see if the specified operand of the
630 /// specified instruction is a constant integer. If so, check to see if there
631 /// are any bits set in the constant that are not demanded. If so, shrink the
632 /// constant and return true.
633 bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDValue Op,
634 const APInt &Demanded) {
635 // FIXME: ISD::SELECT, ISD::SELECT_CC
636 switch(Op.getOpcode()) {
641 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1)))
642 if (C->getAPIntValue().intersects(~Demanded)) {
643 MVT VT = Op.getValueType();
644 SDValue New = DAG.getNode(Op.getOpcode(), VT, Op.getOperand(0),
645 DAG.getConstant(Demanded &
648 return CombineTo(Op, New);
655 /// SimplifyDemandedBits - Look at Op. At this point, we know that only the
656 /// DemandedMask bits of the result of Op are ever used downstream. If we can
657 /// use this information to simplify Op, create a new simplified DAG node and
658 /// return true, returning the original and new nodes in Old and New. Otherwise,
659 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
660 /// the expression (used to simplify the caller). The KnownZero/One bits may
661 /// only be accurate for those bits in the DemandedMask.
662 bool TargetLowering::SimplifyDemandedBits(SDValue Op,
663 const APInt &DemandedMask,
666 TargetLoweringOpt &TLO,
667 unsigned Depth) const {
668 unsigned BitWidth = DemandedMask.getBitWidth();
669 assert(Op.getValueSizeInBits() == BitWidth &&
670 "Mask size mismatches value type size!");
671 APInt NewMask = DemandedMask;
673 // Don't know anything.
674 KnownZero = KnownOne = APInt(BitWidth, 0);
676 // Other users may use these bits.
677 if (!Op.getNode()->hasOneUse()) {
679 // If not at the root, Just compute the KnownZero/KnownOne bits to
680 // simplify things downstream.
681 TLO.DAG.ComputeMaskedBits(Op, DemandedMask, KnownZero, KnownOne, Depth);
684 // If this is the root being simplified, allow it to have multiple uses,
685 // just set the NewMask to all bits.
686 NewMask = APInt::getAllOnesValue(BitWidth);
687 } else if (DemandedMask == 0) {
688 // Not demanding any bits from Op.
689 if (Op.getOpcode() != ISD::UNDEF)
690 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::UNDEF, Op.getValueType()));
692 } else if (Depth == 6) { // Limit search depth.
696 APInt KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut;
697 switch (Op.getOpcode()) {
699 // We know all of the bits for a constant!
700 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue() & NewMask;
701 KnownZero = ~KnownOne & NewMask;
702 return false; // Don't fall through, will infinitely loop.
704 // If the RHS is a constant, check to see if the LHS would be zero without
705 // using the bits from the RHS. Below, we use knowledge about the RHS to
706 // simplify the LHS, here we're using information from the LHS to simplify
708 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
709 APInt LHSZero, LHSOne;
710 TLO.DAG.ComputeMaskedBits(Op.getOperand(0), NewMask,
711 LHSZero, LHSOne, Depth+1);
712 // If the LHS already has zeros where RHSC does, this and is dead.
713 if ((LHSZero & NewMask) == (~RHSC->getAPIntValue() & NewMask))
714 return TLO.CombineTo(Op, Op.getOperand(0));
715 // If any of the set bits in the RHS are known zero on the LHS, shrink
717 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & NewMask))
721 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
722 KnownOne, TLO, Depth+1))
724 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
725 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownZero & NewMask,
726 KnownZero2, KnownOne2, TLO, Depth+1))
728 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
730 // If all of the demanded bits are known one on one side, return the other.
731 // These bits cannot contribute to the result of the 'and'.
732 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
733 return TLO.CombineTo(Op, Op.getOperand(0));
734 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
735 return TLO.CombineTo(Op, Op.getOperand(1));
736 // If all of the demanded bits in the inputs are known zeros, return zero.
737 if ((NewMask & (KnownZero|KnownZero2)) == NewMask)
738 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, Op.getValueType()));
739 // If the RHS is a constant, see if we can simplify it.
740 if (TLO.ShrinkDemandedConstant(Op, ~KnownZero2 & NewMask))
743 // Output known-1 bits are only known if set in both the LHS & RHS.
744 KnownOne &= KnownOne2;
745 // Output known-0 are known to be clear if zero in either the LHS | RHS.
746 KnownZero |= KnownZero2;
749 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
750 KnownOne, TLO, Depth+1))
752 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
753 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownOne & NewMask,
754 KnownZero2, KnownOne2, TLO, Depth+1))
756 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
758 // If all of the demanded bits are known zero on one side, return the other.
759 // These bits cannot contribute to the result of the 'or'.
760 if ((NewMask & ~KnownOne2 & KnownZero) == (~KnownOne2 & NewMask))
761 return TLO.CombineTo(Op, Op.getOperand(0));
762 if ((NewMask & ~KnownOne & KnownZero2) == (~KnownOne & NewMask))
763 return TLO.CombineTo(Op, Op.getOperand(1));
764 // If all of the potentially set bits on one side are known to be set on
765 // the other side, just use the 'other' side.
766 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
767 return TLO.CombineTo(Op, Op.getOperand(0));
768 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
769 return TLO.CombineTo(Op, Op.getOperand(1));
770 // If the RHS is a constant, see if we can simplify it.
771 if (TLO.ShrinkDemandedConstant(Op, NewMask))
774 // Output known-0 bits are only known if clear in both the LHS & RHS.
775 KnownZero &= KnownZero2;
776 // Output known-1 are known to be set if set in either the LHS | RHS.
777 KnownOne |= KnownOne2;
780 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
781 KnownOne, TLO, Depth+1))
783 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
784 if (SimplifyDemandedBits(Op.getOperand(0), NewMask, KnownZero2,
785 KnownOne2, TLO, Depth+1))
787 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
789 // If all of the demanded bits are known zero on one side, return the other.
790 // These bits cannot contribute to the result of the 'xor'.
791 if ((KnownZero & NewMask) == NewMask)
792 return TLO.CombineTo(Op, Op.getOperand(0));
793 if ((KnownZero2 & NewMask) == NewMask)
794 return TLO.CombineTo(Op, Op.getOperand(1));
796 // If all of the unknown bits are known to be zero on one side or the other
797 // (but not both) turn this into an *inclusive* or.
798 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
799 if ((NewMask & ~KnownZero & ~KnownZero2) == 0)
800 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, Op.getValueType(),
804 // Output known-0 bits are known if clear or set in both the LHS & RHS.
805 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
806 // Output known-1 are known to be set if set in only one of the LHS, RHS.
807 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
809 // If all of the demanded bits on one side are known, and all of the set
810 // bits on that side are also known to be set on the other side, turn this
811 // into an AND, as we know the bits will be cleared.
812 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
813 if ((NewMask & (KnownZero|KnownOne)) == NewMask) { // all known
814 if ((KnownOne & KnownOne2) == KnownOne) {
815 MVT VT = Op.getValueType();
816 SDValue ANDC = TLO.DAG.getConstant(~KnownOne & NewMask, VT);
817 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, VT, Op.getOperand(0),
822 // If the RHS is a constant, see if we can simplify it.
823 // for XOR, we prefer to force bits to 1 if they will make a -1.
824 // if we can't force bits, try to shrink constant
825 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
826 APInt Expanded = C->getAPIntValue() | (~NewMask);
827 // if we can expand it to have all bits set, do it
828 if (Expanded.isAllOnesValue()) {
829 if (Expanded != C->getAPIntValue()) {
830 MVT VT = Op.getValueType();
831 SDValue New = TLO.DAG.getNode(Op.getOpcode(), VT, Op.getOperand(0),
832 TLO.DAG.getConstant(Expanded, VT));
833 return TLO.CombineTo(Op, New);
835 // if it already has all the bits set, nothing to change
836 // but don't shrink either!
837 } else if (TLO.ShrinkDemandedConstant(Op, NewMask)) {
842 KnownZero = KnownZeroOut;
843 KnownOne = KnownOneOut;
846 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero,
847 KnownOne, TLO, Depth+1))
849 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero2,
850 KnownOne2, TLO, Depth+1))
852 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
853 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
855 // If the operands are constants, see if we can simplify them.
856 if (TLO.ShrinkDemandedConstant(Op, NewMask))
859 // Only known if known in both the LHS and RHS.
860 KnownOne &= KnownOne2;
861 KnownZero &= KnownZero2;
864 if (SimplifyDemandedBits(Op.getOperand(3), NewMask, KnownZero,
865 KnownOne, TLO, Depth+1))
867 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero2,
868 KnownOne2, TLO, Depth+1))
870 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
871 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
873 // If the operands are constants, see if we can simplify them.
874 if (TLO.ShrinkDemandedConstant(Op, NewMask))
877 // Only known if known in both the LHS and RHS.
878 KnownOne &= KnownOne2;
879 KnownZero &= KnownZero2;
882 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
883 unsigned ShAmt = SA->getValue();
884 SDValue InOp = Op.getOperand(0);
886 // If the shift count is an invalid immediate, don't do anything.
887 if (ShAmt >= BitWidth)
890 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a
891 // single shift. We can do this if the bottom bits (which are shifted
892 // out) are never demanded.
893 if (InOp.getOpcode() == ISD::SRL &&
894 isa<ConstantSDNode>(InOp.getOperand(1))) {
895 if (ShAmt && (NewMask & APInt::getLowBitsSet(BitWidth, ShAmt)) == 0) {
896 unsigned C1 = cast<ConstantSDNode>(InOp.getOperand(1))->getValue();
897 unsigned Opc = ISD::SHL;
905 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
906 MVT VT = Op.getValueType();
907 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, VT,
908 InOp.getOperand(0), NewSA));
912 if (SimplifyDemandedBits(Op.getOperand(0), NewMask.lshr(ShAmt),
913 KnownZero, KnownOne, TLO, Depth+1))
915 KnownZero <<= SA->getValue();
916 KnownOne <<= SA->getValue();
917 // low bits known zero.
918 KnownZero |= APInt::getLowBitsSet(BitWidth, SA->getValue());
922 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
923 MVT VT = Op.getValueType();
924 unsigned ShAmt = SA->getValue();
925 unsigned VTSize = VT.getSizeInBits();
926 SDValue InOp = Op.getOperand(0);
928 // If the shift count is an invalid immediate, don't do anything.
929 if (ShAmt >= BitWidth)
932 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a
933 // single shift. We can do this if the top bits (which are shifted out)
934 // are never demanded.
935 if (InOp.getOpcode() == ISD::SHL &&
936 isa<ConstantSDNode>(InOp.getOperand(1))) {
937 if (ShAmt && (NewMask & APInt::getHighBitsSet(VTSize, ShAmt)) == 0) {
938 unsigned C1 = cast<ConstantSDNode>(InOp.getOperand(1))->getValue();
939 unsigned Opc = ISD::SRL;
947 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
948 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, VT,
949 InOp.getOperand(0), NewSA));
953 // Compute the new bits that are at the top now.
954 if (SimplifyDemandedBits(InOp, (NewMask << ShAmt),
955 KnownZero, KnownOne, TLO, Depth+1))
957 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
958 KnownZero = KnownZero.lshr(ShAmt);
959 KnownOne = KnownOne.lshr(ShAmt);
961 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
962 KnownZero |= HighBits; // High bits known zero.
966 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
967 MVT VT = Op.getValueType();
968 unsigned ShAmt = SA->getValue();
970 // If the shift count is an invalid immediate, don't do anything.
971 if (ShAmt >= BitWidth)
974 APInt InDemandedMask = (NewMask << ShAmt);
976 // If any of the demanded bits are produced by the sign extension, we also
977 // demand the input sign bit.
978 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
979 if (HighBits.intersects(NewMask))
980 InDemandedMask |= APInt::getSignBit(VT.getSizeInBits());
982 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask,
983 KnownZero, KnownOne, TLO, Depth+1))
985 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
986 KnownZero = KnownZero.lshr(ShAmt);
987 KnownOne = KnownOne.lshr(ShAmt);
989 // Handle the sign bit, adjusted to where it is now in the mask.
990 APInt SignBit = APInt::getSignBit(BitWidth).lshr(ShAmt);
992 // If the input sign bit is known to be zero, or if none of the top bits
993 // are demanded, turn this into an unsigned shift right.
994 if (KnownZero.intersects(SignBit) || (HighBits & ~NewMask) == HighBits) {
995 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, VT, Op.getOperand(0),
997 } else if (KnownOne.intersects(SignBit)) { // New bits are known one.
998 KnownOne |= HighBits;
1002 case ISD::SIGN_EXTEND_INREG: {
1003 MVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1005 // Sign extension. Compute the demanded bits in the result that are not
1006 // present in the input.
1007 APInt NewBits = APInt::getHighBitsSet(BitWidth,
1008 BitWidth - EVT.getSizeInBits()) &
1011 // If none of the extended bits are demanded, eliminate the sextinreg.
1013 return TLO.CombineTo(Op, Op.getOperand(0));
1015 APInt InSignBit = APInt::getSignBit(EVT.getSizeInBits());
1016 InSignBit.zext(BitWidth);
1017 APInt InputDemandedBits = APInt::getLowBitsSet(BitWidth,
1018 EVT.getSizeInBits()) &
1021 // Since the sign extended bits are demanded, we know that the sign
1023 InputDemandedBits |= InSignBit;
1025 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits,
1026 KnownZero, KnownOne, TLO, Depth+1))
1028 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1030 // If the sign bit of the input is known set or clear, then we know the
1031 // top bits of the result.
1033 // If the input sign bit is known zero, convert this into a zero extension.
1034 if (KnownZero.intersects(InSignBit))
1035 return TLO.CombineTo(Op,
1036 TLO.DAG.getZeroExtendInReg(Op.getOperand(0), EVT));
1038 if (KnownOne.intersects(InSignBit)) { // Input sign bit known set
1039 KnownOne |= NewBits;
1040 KnownZero &= ~NewBits;
1041 } else { // Input sign bit unknown
1042 KnownZero &= ~NewBits;
1043 KnownOne &= ~NewBits;
1047 case ISD::ZERO_EXTEND: {
1048 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits();
1049 APInt InMask = NewMask;
1050 InMask.trunc(OperandBitWidth);
1052 // If none of the top bits are demanded, convert this into an any_extend.
1054 APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask;
1055 if (!NewBits.intersects(NewMask))
1056 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND,
1060 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1061 KnownZero, KnownOne, TLO, Depth+1))
1063 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1064 KnownZero.zext(BitWidth);
1065 KnownOne.zext(BitWidth);
1066 KnownZero |= NewBits;
1069 case ISD::SIGN_EXTEND: {
1070 MVT InVT = Op.getOperand(0).getValueType();
1071 unsigned InBits = InVT.getSizeInBits();
1072 APInt InMask = APInt::getLowBitsSet(BitWidth, InBits);
1073 APInt InSignBit = APInt::getBitsSet(BitWidth, InBits - 1, InBits);
1074 APInt NewBits = ~InMask & NewMask;
1076 // If none of the top bits are demanded, convert this into an any_extend.
1078 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND,Op.getValueType(),
1081 // Since some of the sign extended bits are demanded, we know that the sign
1083 APInt InDemandedBits = InMask & NewMask;
1084 InDemandedBits |= InSignBit;
1085 InDemandedBits.trunc(InBits);
1087 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero,
1088 KnownOne, TLO, Depth+1))
1090 KnownZero.zext(BitWidth);
1091 KnownOne.zext(BitWidth);
1093 // If the sign bit is known zero, convert this to a zero extend.
1094 if (KnownZero.intersects(InSignBit))
1095 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND,
1099 // If the sign bit is known one, the top bits match.
1100 if (KnownOne.intersects(InSignBit)) {
1101 KnownOne |= NewBits;
1102 KnownZero &= ~NewBits;
1103 } else { // Otherwise, top bits aren't known.
1104 KnownOne &= ~NewBits;
1105 KnownZero &= ~NewBits;
1109 case ISD::ANY_EXTEND: {
1110 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits();
1111 APInt InMask = NewMask;
1112 InMask.trunc(OperandBitWidth);
1113 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1114 KnownZero, KnownOne, TLO, Depth+1))
1116 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1117 KnownZero.zext(BitWidth);
1118 KnownOne.zext(BitWidth);
1121 case ISD::TRUNCATE: {
1122 // Simplify the input, using demanded bit information, and compute the known
1123 // zero/one bits live out.
1124 APInt TruncMask = NewMask;
1125 TruncMask.zext(Op.getOperand(0).getValueSizeInBits());
1126 if (SimplifyDemandedBits(Op.getOperand(0), TruncMask,
1127 KnownZero, KnownOne, TLO, Depth+1))
1129 KnownZero.trunc(BitWidth);
1130 KnownOne.trunc(BitWidth);
1132 // If the input is only used by this truncate, see if we can shrink it based
1133 // on the known demanded bits.
1134 if (Op.getOperand(0).getNode()->hasOneUse()) {
1135 SDValue In = Op.getOperand(0);
1136 unsigned InBitWidth = In.getValueSizeInBits();
1137 switch (In.getOpcode()) {
1140 // Shrink SRL by a constant if none of the high bits shifted in are
1142 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1))){
1143 APInt HighBits = APInt::getHighBitsSet(InBitWidth,
1144 InBitWidth - BitWidth);
1145 HighBits = HighBits.lshr(ShAmt->getValue());
1146 HighBits.trunc(BitWidth);
1148 if (ShAmt->getValue() < BitWidth && !(HighBits & NewMask)) {
1149 // None of the shifted in bits are needed. Add a truncate of the
1150 // shift input, then shift it.
1151 SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE,
1154 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL,Op.getValueType(),
1155 NewTrunc, In.getOperand(1)));
1162 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1165 case ISD::AssertZext: {
1166 MVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1167 APInt InMask = APInt::getLowBitsSet(BitWidth,
1168 VT.getSizeInBits());
1169 if (SimplifyDemandedBits(Op.getOperand(0), InMask & NewMask,
1170 KnownZero, KnownOne, TLO, Depth+1))
1172 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1173 KnownZero |= ~InMask & NewMask;
1176 case ISD::BIT_CONVERT:
1178 // If this is an FP->Int bitcast and if the sign bit is the only thing that
1179 // is demanded, turn this into a FGETSIGN.
1180 if (NewMask == MVT::getIntegerVTSignBit(Op.getValueType()) &&
1181 MVT::isFloatingPoint(Op.getOperand(0).getValueType()) &&
1182 !MVT::isVector(Op.getOperand(0).getValueType())) {
1183 // Only do this xform if FGETSIGN is valid or if before legalize.
1184 if (!TLO.AfterLegalize ||
1185 isOperationLegal(ISD::FGETSIGN, Op.getValueType())) {
1186 // Make a FGETSIGN + SHL to move the sign bit into the appropriate
1187 // place. We expect the SHL to be eliminated by other optimizations.
1188 SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, Op.getValueType(),
1190 unsigned ShVal = Op.getValueType().getSizeInBits()-1;
1191 SDValue ShAmt = TLO.DAG.getConstant(ShVal, getShiftAmountTy());
1192 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, Op.getValueType(),
1199 // Just use ComputeMaskedBits to compute output bits.
1200 TLO.DAG.ComputeMaskedBits(Op, NewMask, KnownZero, KnownOne, Depth);
1204 // If we know the value of all of the demanded bits, return this as a
1206 if ((NewMask & (KnownZero|KnownOne)) == NewMask)
1207 return TLO.CombineTo(Op, TLO.DAG.getConstant(KnownOne, Op.getValueType()));
1212 /// computeMaskedBitsForTargetNode - Determine which of the bits specified
1213 /// in Mask are known to be either zero or one and return them in the
1214 /// KnownZero/KnownOne bitsets.
1215 void TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
1219 const SelectionDAG &DAG,
1220 unsigned Depth) const {
1221 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1222 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1223 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1224 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1225 "Should use MaskedValueIsZero if you don't know whether Op"
1226 " is a target node!");
1227 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
1230 /// ComputeNumSignBitsForTargetNode - This method can be implemented by
1231 /// targets that want to expose additional information about sign bits to the
1233 unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
1234 unsigned Depth) const {
1235 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1236 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1237 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1238 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1239 "Should use ComputeNumSignBits if you don't know whether Op"
1240 " is a target node!");
1245 /// SimplifySetCC - Try to simplify a setcc built with the specified operands
1246 /// and cc. If it is unable to simplify it, return a null SDValue.
1248 TargetLowering::SimplifySetCC(MVT VT, SDValue N0, SDValue N1,
1249 ISD::CondCode Cond, bool foldBooleans,
1250 DAGCombinerInfo &DCI) const {
1251 SelectionDAG &DAG = DCI.DAG;
1253 // These setcc operations always fold.
1257 case ISD::SETFALSE2: return DAG.getConstant(0, VT);
1259 case ISD::SETTRUE2: return DAG.getConstant(1, VT);
1262 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) {
1263 const APInt &C1 = N1C->getAPIntValue();
1264 if (isa<ConstantSDNode>(N0.getNode())) {
1265 return DAG.FoldSetCC(VT, N0, N1, Cond);
1267 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an
1268 // equality comparison, then we're just comparing whether X itself is
1270 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) &&
1271 N0.getOperand(0).getOpcode() == ISD::CTLZ &&
1272 N0.getOperand(1).getOpcode() == ISD::Constant) {
1273 unsigned ShAmt = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1274 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1275 ShAmt == Log2_32(N0.getValueType().getSizeInBits())) {
1276 if ((C1 == 0) == (Cond == ISD::SETEQ)) {
1277 // (srl (ctlz x), 5) == 0 -> X != 0
1278 // (srl (ctlz x), 5) != 1 -> X != 0
1281 // (srl (ctlz x), 5) != 0 -> X == 0
1282 // (srl (ctlz x), 5) == 1 -> X == 0
1285 SDValue Zero = DAG.getConstant(0, N0.getValueType());
1286 return DAG.getSetCC(VT, N0.getOperand(0).getOperand(0),
1291 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
1292 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
1293 unsigned InSize = N0.getOperand(0).getValueType().getSizeInBits();
1295 // If the comparison constant has bits in the upper part, the
1296 // zero-extended value could never match.
1297 if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(),
1298 C1.getBitWidth() - InSize))) {
1302 case ISD::SETEQ: return DAG.getConstant(0, VT);
1305 case ISD::SETNE: return DAG.getConstant(1, VT);
1308 // True if the sign bit of C1 is set.
1309 return DAG.getConstant(C1.isNegative(), VT);
1312 // True if the sign bit of C1 isn't set.
1313 return DAG.getConstant(C1.isNonNegative(), VT);
1319 // Otherwise, we can perform the comparison with the low bits.
1327 return DAG.getSetCC(VT, N0.getOperand(0),
1328 DAG.getConstant(APInt(C1).trunc(InSize),
1329 N0.getOperand(0).getValueType()),
1332 break; // todo, be more careful with signed comparisons
1334 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
1335 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1336 MVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
1337 unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits();
1338 MVT ExtDstTy = N0.getValueType();
1339 unsigned ExtDstTyBits = ExtDstTy.getSizeInBits();
1341 // If the extended part has any inconsistent bits, it cannot ever
1342 // compare equal. In other words, they have to be all ones or all
1345 APInt::getHighBitsSet(ExtDstTyBits, ExtDstTyBits - ExtSrcTyBits);
1346 if ((C1 & ExtBits) != 0 && (C1 & ExtBits) != ExtBits)
1347 return DAG.getConstant(Cond == ISD::SETNE, VT);
1350 MVT Op0Ty = N0.getOperand(0).getValueType();
1351 if (Op0Ty == ExtSrcTy) {
1352 ZextOp = N0.getOperand(0);
1354 APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits);
1355 ZextOp = DAG.getNode(ISD::AND, Op0Ty, N0.getOperand(0),
1356 DAG.getConstant(Imm, Op0Ty));
1358 if (!DCI.isCalledByLegalizer())
1359 DCI.AddToWorklist(ZextOp.getNode());
1360 // Otherwise, make this a use of a zext.
1361 return DAG.getSetCC(VT, ZextOp,
1362 DAG.getConstant(C1 & APInt::getLowBitsSet(
1367 } else if ((N1C->isNullValue() || N1C->getAPIntValue() == 1) &&
1368 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1370 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC
1371 if (N0.getOpcode() == ISD::SETCC) {
1372 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getValue() != 1);
1376 // Invert the condition.
1377 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
1378 CC = ISD::getSetCCInverse(CC,
1379 N0.getOperand(0).getValueType().isInteger());
1380 return DAG.getSetCC(VT, N0.getOperand(0), N0.getOperand(1), CC);
1383 if ((N0.getOpcode() == ISD::XOR ||
1384 (N0.getOpcode() == ISD::AND &&
1385 N0.getOperand(0).getOpcode() == ISD::XOR &&
1386 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
1387 isa<ConstantSDNode>(N0.getOperand(1)) &&
1388 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue() == 1) {
1389 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We
1390 // can only do this if the top bits are known zero.
1391 unsigned BitWidth = N0.getValueSizeInBits();
1392 if (DAG.MaskedValueIsZero(N0,
1393 APInt::getHighBitsSet(BitWidth,
1395 // Okay, get the un-inverted input value.
1397 if (N0.getOpcode() == ISD::XOR)
1398 Val = N0.getOperand(0);
1400 assert(N0.getOpcode() == ISD::AND &&
1401 N0.getOperand(0).getOpcode() == ISD::XOR);
1402 // ((X^1)&1)^1 -> X & 1
1403 Val = DAG.getNode(ISD::AND, N0.getValueType(),
1404 N0.getOperand(0).getOperand(0),
1407 return DAG.getSetCC(VT, Val, N1,
1408 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
1413 APInt MinVal, MaxVal;
1414 unsigned OperandBitSize = N1C->getValueType(0).getSizeInBits();
1415 if (ISD::isSignedIntSetCC(Cond)) {
1416 MinVal = APInt::getSignedMinValue(OperandBitSize);
1417 MaxVal = APInt::getSignedMaxValue(OperandBitSize);
1419 MinVal = APInt::getMinValue(OperandBitSize);
1420 MaxVal = APInt::getMaxValue(OperandBitSize);
1423 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
1424 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
1425 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true
1426 // X >= C0 --> X > (C0-1)
1427 return DAG.getSetCC(VT, N0, DAG.getConstant(C1-1, N1.getValueType()),
1428 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
1431 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
1432 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true
1433 // X <= C0 --> X < (C0+1)
1434 return DAG.getSetCC(VT, N0, DAG.getConstant(C1+1, N1.getValueType()),
1435 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
1438 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
1439 return DAG.getConstant(0, VT); // X < MIN --> false
1440 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal)
1441 return DAG.getConstant(1, VT); // X >= MIN --> true
1442 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal)
1443 return DAG.getConstant(0, VT); // X > MAX --> false
1444 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal)
1445 return DAG.getConstant(1, VT); // X <= MAX --> true
1447 // Canonicalize setgt X, Min --> setne X, Min
1448 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
1449 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
1450 // Canonicalize setlt X, Max --> setne X, Max
1451 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
1452 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
1454 // If we have setult X, 1, turn it into seteq X, 0
1455 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
1456 return DAG.getSetCC(VT, N0, DAG.getConstant(MinVal, N0.getValueType()),
1458 // If we have setugt X, Max-1, turn it into seteq X, Max
1459 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
1460 return DAG.getSetCC(VT, N0, DAG.getConstant(MaxVal, N0.getValueType()),
1463 // If we have "setcc X, C0", check to see if we can shrink the immediate
1466 // SETUGT X, SINTMAX -> SETLT X, 0
1467 if (Cond == ISD::SETUGT && OperandBitSize != 1 &&
1468 C1 == (~0ULL >> (65-OperandBitSize)))
1469 return DAG.getSetCC(VT, N0, DAG.getConstant(0, N1.getValueType()),
1472 // FIXME: Implement the rest of these.
1474 // Fold bit comparisons when we can.
1475 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1476 VT == N0.getValueType() && N0.getOpcode() == ISD::AND)
1477 if (ConstantSDNode *AndRHS =
1478 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1479 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
1480 // Perform the xform if the AND RHS is a single bit.
1481 if (isPowerOf2_64(AndRHS->getValue())) {
1482 return DAG.getNode(ISD::SRL, VT, N0,
1483 DAG.getConstant(Log2_64(AndRHS->getValue()),
1484 getShiftAmountTy()));
1486 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getValue()) {
1487 // (X & 8) == 8 --> (X & 8) >> 3
1488 // Perform the xform if C1 is a single bit.
1489 if (C1.isPowerOf2()) {
1490 return DAG.getNode(ISD::SRL, VT, N0,
1491 DAG.getConstant(C1.logBase2(), getShiftAmountTy()));
1496 } else if (isa<ConstantSDNode>(N0.getNode())) {
1497 // Ensure that the constant occurs on the RHS.
1498 return DAG.getSetCC(VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
1501 if (isa<ConstantFPSDNode>(N0.getNode())) {
1502 // Constant fold or commute setcc.
1503 SDValue O = DAG.FoldSetCC(VT, N0, N1, Cond);
1504 if (O.getNode()) return O;
1505 } else if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1.getNode())) {
1506 // If the RHS of an FP comparison is a constant, simplify it away in
1508 if (CFP->getValueAPF().isNaN()) {
1509 // If an operand is known to be a nan, we can fold it.
1510 switch (ISD::getUnorderedFlavor(Cond)) {
1511 default: assert(0 && "Unknown flavor!");
1512 case 0: // Known false.
1513 return DAG.getConstant(0, VT);
1514 case 1: // Known true.
1515 return DAG.getConstant(1, VT);
1516 case 2: // Undefined.
1517 return DAG.getNode(ISD::UNDEF, VT);
1521 // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the
1522 // constant if knowing that the operand is non-nan is enough. We prefer to
1523 // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to
1525 if (Cond == ISD::SETO || Cond == ISD::SETUO)
1526 return DAG.getSetCC(VT, N0, N0, Cond);
1530 // We can always fold X == X for integer setcc's.
1531 if (N0.getValueType().isInteger())
1532 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
1533 unsigned UOF = ISD::getUnorderedFlavor(Cond);
1534 if (UOF == 2) // FP operators that are undefined on NaNs.
1535 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
1536 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
1537 return DAG.getConstant(UOF, VT);
1538 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
1539 // if it is not already.
1540 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
1541 if (NewCond != Cond)
1542 return DAG.getSetCC(VT, N0, N1, NewCond);
1545 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1546 N0.getValueType().isInteger()) {
1547 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
1548 N0.getOpcode() == ISD::XOR) {
1549 // Simplify (X+Y) == (X+Z) --> Y == Z
1550 if (N0.getOpcode() == N1.getOpcode()) {
1551 if (N0.getOperand(0) == N1.getOperand(0))
1552 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(1), Cond);
1553 if (N0.getOperand(1) == N1.getOperand(1))
1554 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(0), Cond);
1555 if (DAG.isCommutativeBinOp(N0.getOpcode())) {
1556 // If X op Y == Y op X, try other combinations.
1557 if (N0.getOperand(0) == N1.getOperand(1))
1558 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(0), Cond);
1559 if (N0.getOperand(1) == N1.getOperand(0))
1560 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(1), Cond);
1564 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) {
1565 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1566 // Turn (X+C1) == C2 --> X == C2-C1
1567 if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse()) {
1568 return DAG.getSetCC(VT, N0.getOperand(0),
1569 DAG.getConstant(RHSC->getValue()-LHSR->getValue(),
1570 N0.getValueType()), Cond);
1573 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
1574 if (N0.getOpcode() == ISD::XOR)
1575 // If we know that all of the inverted bits are zero, don't bother
1576 // performing the inversion.
1577 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue()))
1579 DAG.getSetCC(VT, N0.getOperand(0),
1580 DAG.getConstant(LHSR->getAPIntValue() ^
1581 RHSC->getAPIntValue(),
1586 // Turn (C1-X) == C2 --> X == C1-C2
1587 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
1588 if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse()) {
1590 DAG.getSetCC(VT, N0.getOperand(1),
1591 DAG.getConstant(SUBC->getAPIntValue() -
1592 RHSC->getAPIntValue(),
1599 // Simplify (X+Z) == X --> Z == 0
1600 if (N0.getOperand(0) == N1)
1601 return DAG.getSetCC(VT, N0.getOperand(1),
1602 DAG.getConstant(0, N0.getValueType()), Cond);
1603 if (N0.getOperand(1) == N1) {
1604 if (DAG.isCommutativeBinOp(N0.getOpcode()))
1605 return DAG.getSetCC(VT, N0.getOperand(0),
1606 DAG.getConstant(0, N0.getValueType()), Cond);
1607 else if (N0.getNode()->hasOneUse()) {
1608 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
1609 // (Z-X) == X --> Z == X<<1
1610 SDValue SH = DAG.getNode(ISD::SHL, N1.getValueType(),
1612 DAG.getConstant(1, getShiftAmountTy()));
1613 if (!DCI.isCalledByLegalizer())
1614 DCI.AddToWorklist(SH.getNode());
1615 return DAG.getSetCC(VT, N0.getOperand(0), SH, Cond);
1620 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
1621 N1.getOpcode() == ISD::XOR) {
1622 // Simplify X == (X+Z) --> Z == 0
1623 if (N1.getOperand(0) == N0) {
1624 return DAG.getSetCC(VT, N1.getOperand(1),
1625 DAG.getConstant(0, N1.getValueType()), Cond);
1626 } else if (N1.getOperand(1) == N0) {
1627 if (DAG.isCommutativeBinOp(N1.getOpcode())) {
1628 return DAG.getSetCC(VT, N1.getOperand(0),
1629 DAG.getConstant(0, N1.getValueType()), Cond);
1630 } else if (N1.getNode()->hasOneUse()) {
1631 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
1632 // X == (Z-X) --> X<<1 == Z
1633 SDValue SH = DAG.getNode(ISD::SHL, N1.getValueType(), N0,
1634 DAG.getConstant(1, getShiftAmountTy()));
1635 if (!DCI.isCalledByLegalizer())
1636 DCI.AddToWorklist(SH.getNode());
1637 return DAG.getSetCC(VT, SH, N1.getOperand(0), Cond);
1643 // Fold away ALL boolean setcc's.
1645 if (N0.getValueType() == MVT::i1 && foldBooleans) {
1647 default: assert(0 && "Unknown integer setcc!");
1648 case ISD::SETEQ: // X == Y -> (X^Y)^1
1649 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
1650 N0 = DAG.getNode(ISD::XOR, MVT::i1, Temp, DAG.getConstant(1, MVT::i1));
1651 if (!DCI.isCalledByLegalizer())
1652 DCI.AddToWorklist(Temp.getNode());
1654 case ISD::SETNE: // X != Y --> (X^Y)
1655 N0 = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
1657 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> X^1 & Y
1658 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> X^1 & Y
1659 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
1660 N0 = DAG.getNode(ISD::AND, MVT::i1, N1, Temp);
1661 if (!DCI.isCalledByLegalizer())
1662 DCI.AddToWorklist(Temp.getNode());
1664 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> Y^1 & X
1665 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> Y^1 & X
1666 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
1667 N0 = DAG.getNode(ISD::AND, MVT::i1, N0, Temp);
1668 if (!DCI.isCalledByLegalizer())
1669 DCI.AddToWorklist(Temp.getNode());
1671 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> X^1 | Y
1672 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> X^1 | Y
1673 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
1674 N0 = DAG.getNode(ISD::OR, MVT::i1, N1, Temp);
1675 if (!DCI.isCalledByLegalizer())
1676 DCI.AddToWorklist(Temp.getNode());
1678 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> Y^1 | X
1679 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> Y^1 | X
1680 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
1681 N0 = DAG.getNode(ISD::OR, MVT::i1, N0, Temp);
1684 if (VT != MVT::i1) {
1685 if (!DCI.isCalledByLegalizer())
1686 DCI.AddToWorklist(N0.getNode());
1687 // FIXME: If running after legalize, we probably can't do this.
1688 N0 = DAG.getNode(ISD::ZERO_EXTEND, VT, N0);
1693 // Could not fold it.
1697 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
1698 /// node is a GlobalAddress + offset.
1699 bool TargetLowering::isGAPlusOffset(SDNode *N, GlobalValue* &GA,
1700 int64_t &Offset) const {
1701 if (isa<GlobalAddressSDNode>(N)) {
1702 GlobalAddressSDNode *GASD = cast<GlobalAddressSDNode>(N);
1703 GA = GASD->getGlobal();
1704 Offset += GASD->getOffset();
1708 if (N->getOpcode() == ISD::ADD) {
1709 SDValue N1 = N->getOperand(0);
1710 SDValue N2 = N->getOperand(1);
1711 if (isGAPlusOffset(N1.getNode(), GA, Offset)) {
1712 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
1714 Offset += V->getSignExtended();
1717 } else if (isGAPlusOffset(N2.getNode(), GA, Offset)) {
1718 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
1720 Offset += V->getSignExtended();
1729 /// isConsecutiveLoad - Return true if LD (which must be a LoadSDNode) is
1730 /// loading 'Bytes' bytes from a location that is 'Dist' units away from the
1731 /// location that the 'Base' load is loading from.
1732 bool TargetLowering::isConsecutiveLoad(SDNode *LD, SDNode *Base,
1733 unsigned Bytes, int Dist,
1734 const MachineFrameInfo *MFI) const {
1735 if (LD->getOperand(0).getNode() != Base->getOperand(0).getNode())
1737 MVT VT = LD->getValueType(0);
1738 if (VT.getSizeInBits() / 8 != Bytes)
1741 SDValue Loc = LD->getOperand(1);
1742 SDValue BaseLoc = Base->getOperand(1);
1743 if (Loc.getOpcode() == ISD::FrameIndex) {
1744 if (BaseLoc.getOpcode() != ISD::FrameIndex)
1746 int FI = cast<FrameIndexSDNode>(Loc)->getIndex();
1747 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
1748 int FS = MFI->getObjectSize(FI);
1749 int BFS = MFI->getObjectSize(BFI);
1750 if (FS != BFS || FS != (int)Bytes) return false;
1751 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Bytes);
1754 GlobalValue *GV1 = NULL;
1755 GlobalValue *GV2 = NULL;
1756 int64_t Offset1 = 0;
1757 int64_t Offset2 = 0;
1758 bool isGA1 = isGAPlusOffset(Loc.getNode(), GV1, Offset1);
1759 bool isGA2 = isGAPlusOffset(BaseLoc.getNode(), GV2, Offset2);
1760 if (isGA1 && isGA2 && GV1 == GV2)
1761 return Offset1 == (Offset2 + Dist*Bytes);
1766 SDValue TargetLowering::
1767 PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
1768 // Default implementation: no optimization.
1772 //===----------------------------------------------------------------------===//
1773 // Inline Assembler Implementation Methods
1774 //===----------------------------------------------------------------------===//
1777 TargetLowering::ConstraintType
1778 TargetLowering::getConstraintType(const std::string &Constraint) const {
1779 // FIXME: lots more standard ones to handle.
1780 if (Constraint.size() == 1) {
1781 switch (Constraint[0]) {
1783 case 'r': return C_RegisterClass;
1785 case 'o': // offsetable
1786 case 'V': // not offsetable
1788 case 'i': // Simple Integer or Relocatable Constant
1789 case 'n': // Simple Integer
1790 case 's': // Relocatable Constant
1791 case 'X': // Allow ANY value.
1792 case 'I': // Target registers.
1804 if (Constraint.size() > 1 && Constraint[0] == '{' &&
1805 Constraint[Constraint.size()-1] == '}')
1810 /// LowerXConstraint - try to replace an X constraint, which matches anything,
1811 /// with another that has more specific requirements based on the type of the
1812 /// corresponding operand.
1813 const char *TargetLowering::LowerXConstraint(MVT ConstraintVT) const{
1814 if (ConstraintVT.isInteger())
1816 if (ConstraintVT.isFloatingPoint())
1817 return "f"; // works for many targets
1821 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
1822 /// vector. If it is invalid, don't add anything to Ops.
1823 void TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
1824 char ConstraintLetter,
1825 std::vector<SDValue> &Ops,
1826 SelectionDAG &DAG) const {
1827 switch (ConstraintLetter) {
1829 case 'X': // Allows any operand; labels (basic block) use this.
1830 if (Op.getOpcode() == ISD::BasicBlock) {
1835 case 'i': // Simple Integer or Relocatable Constant
1836 case 'n': // Simple Integer
1837 case 's': { // Relocatable Constant
1838 // These operands are interested in values of the form (GV+C), where C may
1839 // be folded in as an offset of GV, or it may be explicitly added. Also, it
1840 // is possible and fine if either GV or C are missing.
1841 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
1842 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
1844 // If we have "(add GV, C)", pull out GV/C
1845 if (Op.getOpcode() == ISD::ADD) {
1846 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
1847 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
1848 if (C == 0 || GA == 0) {
1849 C = dyn_cast<ConstantSDNode>(Op.getOperand(0));
1850 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1));
1852 if (C == 0 || GA == 0)
1856 // If we find a valid operand, map to the TargetXXX version so that the
1857 // value itself doesn't get selected.
1858 if (GA) { // Either &GV or &GV+C
1859 if (ConstraintLetter != 'n') {
1860 int64_t Offs = GA->getOffset();
1861 if (C) Offs += C->getValue();
1862 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(),
1863 Op.getValueType(), Offs));
1867 if (C) { // just C, no GV.
1868 // Simple constants are not allowed for 's'.
1869 if (ConstraintLetter != 's') {
1870 Ops.push_back(DAG.getTargetConstant(C->getValue(), Op.getValueType()));
1879 std::vector<unsigned> TargetLowering::
1880 getRegClassForInlineAsmConstraint(const std::string &Constraint,
1882 return std::vector<unsigned>();
1886 std::pair<unsigned, const TargetRegisterClass*> TargetLowering::
1887 getRegForInlineAsmConstraint(const std::string &Constraint,
1889 if (Constraint[0] != '{')
1890 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
1891 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?");
1893 // Remove the braces from around the name.
1894 std::string RegName(Constraint.begin()+1, Constraint.end()-1);
1896 // Figure out which register class contains this reg.
1897 const TargetRegisterInfo *RI = TM.getRegisterInfo();
1898 for (TargetRegisterInfo::regclass_iterator RCI = RI->regclass_begin(),
1899 E = RI->regclass_end(); RCI != E; ++RCI) {
1900 const TargetRegisterClass *RC = *RCI;
1902 // If none of the the value types for this register class are valid, we
1903 // can't use it. For example, 64-bit reg classes on 32-bit targets.
1904 bool isLegal = false;
1905 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
1907 if (isTypeLegal(*I)) {
1913 if (!isLegal) continue;
1915 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end();
1917 if (StringsEqualNoCase(RegName, RI->get(*I).AsmName))
1918 return std::make_pair(*I, RC);
1922 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
1925 //===----------------------------------------------------------------------===//
1926 // Constraint Selection.
1928 /// getConstraintGenerality - Return an integer indicating how general CT
1930 static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) {
1932 default: assert(0 && "Unknown constraint type!");
1933 case TargetLowering::C_Other:
1934 case TargetLowering::C_Unknown:
1936 case TargetLowering::C_Register:
1938 case TargetLowering::C_RegisterClass:
1940 case TargetLowering::C_Memory:
1945 /// ChooseConstraint - If there are multiple different constraints that we
1946 /// could pick for this operand (e.g. "imr") try to pick the 'best' one.
1947 /// This is somewhat tricky: constraints fall into four classes:
1948 /// Other -> immediates and magic values
1949 /// Register -> one specific register
1950 /// RegisterClass -> a group of regs
1951 /// Memory -> memory
1952 /// Ideally, we would pick the most specific constraint possible: if we have
1953 /// something that fits into a register, we would pick it. The problem here
1954 /// is that if we have something that could either be in a register or in
1955 /// memory that use of the register could cause selection of *other*
1956 /// operands to fail: they might only succeed if we pick memory. Because of
1957 /// this the heuristic we use is:
1959 /// 1) If there is an 'other' constraint, and if the operand is valid for
1960 /// that constraint, use it. This makes us take advantage of 'i'
1961 /// constraints when available.
1962 /// 2) Otherwise, pick the most general constraint present. This prefers
1963 /// 'm' over 'r', for example.
1965 static void ChooseConstraint(TargetLowering::AsmOperandInfo &OpInfo,
1966 const TargetLowering &TLI,
1967 SDValue Op, SelectionDAG *DAG) {
1968 assert(OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options");
1969 unsigned BestIdx = 0;
1970 TargetLowering::ConstraintType BestType = TargetLowering::C_Unknown;
1971 int BestGenerality = -1;
1973 // Loop over the options, keeping track of the most general one.
1974 for (unsigned i = 0, e = OpInfo.Codes.size(); i != e; ++i) {
1975 TargetLowering::ConstraintType CType =
1976 TLI.getConstraintType(OpInfo.Codes[i]);
1978 // If this is an 'other' constraint, see if the operand is valid for it.
1979 // For example, on X86 we might have an 'rI' constraint. If the operand
1980 // is an integer in the range [0..31] we want to use I (saving a load
1981 // of a register), otherwise we must use 'r'.
1982 if (CType == TargetLowering::C_Other && Op.getNode()) {
1983 assert(OpInfo.Codes[i].size() == 1 &&
1984 "Unhandled multi-letter 'other' constraint");
1985 std::vector<SDValue> ResultOps;
1986 TLI.LowerAsmOperandForConstraint(Op, OpInfo.Codes[i][0],
1988 if (!ResultOps.empty()) {
1995 // This constraint letter is more general than the previous one, use it.
1996 int Generality = getConstraintGenerality(CType);
1997 if (Generality > BestGenerality) {
2000 BestGenerality = Generality;
2004 OpInfo.ConstraintCode = OpInfo.Codes[BestIdx];
2005 OpInfo.ConstraintType = BestType;
2008 /// ComputeConstraintToUse - Determines the constraint code and constraint
2009 /// type to use for the specific AsmOperandInfo, setting
2010 /// OpInfo.ConstraintCode and OpInfo.ConstraintType.
2011 void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo,
2013 SelectionDAG *DAG) const {
2014 assert(!OpInfo.Codes.empty() && "Must have at least one constraint");
2016 // Single-letter constraints ('r') are very common.
2017 if (OpInfo.Codes.size() == 1) {
2018 OpInfo.ConstraintCode = OpInfo.Codes[0];
2019 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
2021 ChooseConstraint(OpInfo, *this, Op, DAG);
2024 // 'X' matches anything.
2025 if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) {
2026 // Labels and constants are handled elsewhere ('X' is the only thing
2027 // that matches labels).
2028 if (isa<BasicBlock>(OpInfo.CallOperandVal) ||
2029 isa<ConstantInt>(OpInfo.CallOperandVal))
2032 // Otherwise, try to resolve it to something we know about by looking at
2033 // the actual operand type.
2034 if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) {
2035 OpInfo.ConstraintCode = Repl;
2036 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
2041 //===----------------------------------------------------------------------===//
2042 // Loop Strength Reduction hooks
2043 //===----------------------------------------------------------------------===//
2045 /// isLegalAddressingMode - Return true if the addressing mode represented
2046 /// by AM is legal for this target, for a load/store of the specified type.
2047 bool TargetLowering::isLegalAddressingMode(const AddrMode &AM,
2048 const Type *Ty) const {
2049 // The default implementation of this implements a conservative RISCy, r+r and
2052 // Allows a sign-extended 16-bit immediate field.
2053 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
2056 // No global is ever allowed as a base.
2060 // Only support r+r,
2062 case 0: // "r+i" or just "i", depending on HasBaseReg.
2065 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
2067 // Otherwise we have r+r or r+i.
2070 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
2072 // Allow 2*r as r+r.
2079 // Magic for divide replacement
2082 int64_t m; // magic number
2083 int64_t s; // shift amount
2087 uint64_t m; // magic number
2088 int64_t a; // add indicator
2089 int64_t s; // shift amount
2092 /// magic - calculate the magic numbers required to codegen an integer sdiv as
2093 /// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
2095 static ms magic32(int32_t d) {
2097 uint32_t ad, anc, delta, q1, r1, q2, r2, t;
2098 const uint32_t two31 = 0x80000000U;
2102 t = two31 + ((uint32_t)d >> 31);
2103 anc = t - 1 - t%ad; // absolute value of nc
2104 p = 31; // initialize p
2105 q1 = two31/anc; // initialize q1 = 2p/abs(nc)
2106 r1 = two31 - q1*anc; // initialize r1 = rem(2p,abs(nc))
2107 q2 = two31/ad; // initialize q2 = 2p/abs(d)
2108 r2 = two31 - q2*ad; // initialize r2 = rem(2p,abs(d))
2111 q1 = 2*q1; // update q1 = 2p/abs(nc)
2112 r1 = 2*r1; // update r1 = rem(2p/abs(nc))
2113 if (r1 >= anc) { // must be unsigned comparison
2117 q2 = 2*q2; // update q2 = 2p/abs(d)
2118 r2 = 2*r2; // update r2 = rem(2p/abs(d))
2119 if (r2 >= ad) { // must be unsigned comparison
2124 } while (q1 < delta || (q1 == delta && r1 == 0));
2126 mag.m = (int32_t)(q2 + 1); // make sure to sign extend
2127 if (d < 0) mag.m = -mag.m; // resulting magic number
2128 mag.s = p - 32; // resulting shift
2132 /// magicu - calculate the magic numbers required to codegen an integer udiv as
2133 /// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
2134 static mu magicu32(uint32_t d) {
2136 uint32_t nc, delta, q1, r1, q2, r2;
2138 magu.a = 0; // initialize "add" indicator
2140 p = 31; // initialize p
2141 q1 = 0x80000000/nc; // initialize q1 = 2p/nc
2142 r1 = 0x80000000 - q1*nc; // initialize r1 = rem(2p,nc)
2143 q2 = 0x7FFFFFFF/d; // initialize q2 = (2p-1)/d
2144 r2 = 0x7FFFFFFF - q2*d; // initialize r2 = rem((2p-1),d)
2147 if (r1 >= nc - r1 ) {
2148 q1 = 2*q1 + 1; // update q1
2149 r1 = 2*r1 - nc; // update r1
2152 q1 = 2*q1; // update q1
2153 r1 = 2*r1; // update r1
2155 if (r2 + 1 >= d - r2) {
2156 if (q2 >= 0x7FFFFFFF) magu.a = 1;
2157 q2 = 2*q2 + 1; // update q2
2158 r2 = 2*r2 + 1 - d; // update r2
2161 if (q2 >= 0x80000000) magu.a = 1;
2162 q2 = 2*q2; // update q2
2163 r2 = 2*r2 + 1; // update r2
2166 } while (p < 64 && (q1 < delta || (q1 == delta && r1 == 0)));
2167 magu.m = q2 + 1; // resulting magic number
2168 magu.s = p - 32; // resulting shift
2172 /// magic - calculate the magic numbers required to codegen an integer sdiv as
2173 /// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
2175 static ms magic64(int64_t d) {
2177 uint64_t ad, anc, delta, q1, r1, q2, r2, t;
2178 const uint64_t two63 = 9223372036854775808ULL; // 2^63
2181 ad = d >= 0 ? d : -d;
2182 t = two63 + ((uint64_t)d >> 63);
2183 anc = t - 1 - t%ad; // absolute value of nc
2184 p = 63; // initialize p
2185 q1 = two63/anc; // initialize q1 = 2p/abs(nc)
2186 r1 = two63 - q1*anc; // initialize r1 = rem(2p,abs(nc))
2187 q2 = two63/ad; // initialize q2 = 2p/abs(d)
2188 r2 = two63 - q2*ad; // initialize r2 = rem(2p,abs(d))
2191 q1 = 2*q1; // update q1 = 2p/abs(nc)
2192 r1 = 2*r1; // update r1 = rem(2p/abs(nc))
2193 if (r1 >= anc) { // must be unsigned comparison
2197 q2 = 2*q2; // update q2 = 2p/abs(d)
2198 r2 = 2*r2; // update r2 = rem(2p/abs(d))
2199 if (r2 >= ad) { // must be unsigned comparison
2204 } while (q1 < delta || (q1 == delta && r1 == 0));
2207 if (d < 0) mag.m = -mag.m; // resulting magic number
2208 mag.s = p - 64; // resulting shift
2212 /// magicu - calculate the magic numbers required to codegen an integer udiv as
2213 /// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
2214 static mu magicu64(uint64_t d)
2217 uint64_t nc, delta, q1, r1, q2, r2;
2219 magu.a = 0; // initialize "add" indicator
2221 p = 63; // initialize p
2222 q1 = 0x8000000000000000ull/nc; // initialize q1 = 2p/nc
2223 r1 = 0x8000000000000000ull - q1*nc; // initialize r1 = rem(2p,nc)
2224 q2 = 0x7FFFFFFFFFFFFFFFull/d; // initialize q2 = (2p-1)/d
2225 r2 = 0x7FFFFFFFFFFFFFFFull - q2*d; // initialize r2 = rem((2p-1),d)
2228 if (r1 >= nc - r1 ) {
2229 q1 = 2*q1 + 1; // update q1
2230 r1 = 2*r1 - nc; // update r1
2233 q1 = 2*q1; // update q1
2234 r1 = 2*r1; // update r1
2236 if (r2 + 1 >= d - r2) {
2237 if (q2 >= 0x7FFFFFFFFFFFFFFFull) magu.a = 1;
2238 q2 = 2*q2 + 1; // update q2
2239 r2 = 2*r2 + 1 - d; // update r2
2242 if (q2 >= 0x8000000000000000ull) magu.a = 1;
2243 q2 = 2*q2; // update q2
2244 r2 = 2*r2 + 1; // update r2
2247 } while (p < 128 && (q1 < delta || (q1 == delta && r1 == 0)));
2248 magu.m = q2 + 1; // resulting magic number
2249 magu.s = p - 64; // resulting shift
2253 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
2254 /// return a DAG expression to select that will generate the same value by
2255 /// multiplying by a magic number. See:
2256 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
2257 SDValue TargetLowering::BuildSDIV(SDNode *N, SelectionDAG &DAG,
2258 std::vector<SDNode*>* Created) const {
2259 MVT VT = N->getValueType(0);
2261 // Check to see if we can do this.
2262 if (!isTypeLegal(VT) || (VT != MVT::i32 && VT != MVT::i64))
2263 return SDValue(); // BuildSDIV only operates on i32 or i64
2265 int64_t d = cast<ConstantSDNode>(N->getOperand(1))->getSignExtended();
2266 ms magics = (VT == MVT::i32) ? magic32(d) : magic64(d);
2268 // Multiply the numerator (operand 0) by the magic value
2270 if (isOperationLegal(ISD::MULHS, VT))
2271 Q = DAG.getNode(ISD::MULHS, VT, N->getOperand(0),
2272 DAG.getConstant(magics.m, VT));
2273 else if (isOperationLegal(ISD::SMUL_LOHI, VT))
2274 Q = SDValue(DAG.getNode(ISD::SMUL_LOHI, DAG.getVTList(VT, VT),
2276 DAG.getConstant(magics.m, VT)).getNode(), 1);
2278 return SDValue(); // No mulhs or equvialent
2279 // If d > 0 and m < 0, add the numerator
2280 if (d > 0 && magics.m < 0) {
2281 Q = DAG.getNode(ISD::ADD, VT, Q, N->getOperand(0));
2283 Created->push_back(Q.getNode());
2285 // If d < 0 and m > 0, subtract the numerator.
2286 if (d < 0 && magics.m > 0) {
2287 Q = DAG.getNode(ISD::SUB, VT, Q, N->getOperand(0));
2289 Created->push_back(Q.getNode());
2291 // Shift right algebraic if shift value is nonzero
2293 Q = DAG.getNode(ISD::SRA, VT, Q,
2294 DAG.getConstant(magics.s, getShiftAmountTy()));
2296 Created->push_back(Q.getNode());
2298 // Extract the sign bit and add it to the quotient
2300 DAG.getNode(ISD::SRL, VT, Q, DAG.getConstant(VT.getSizeInBits()-1,
2301 getShiftAmountTy()));
2303 Created->push_back(T.getNode());
2304 return DAG.getNode(ISD::ADD, VT, Q, T);
2307 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
2308 /// return a DAG expression to select that will generate the same value by
2309 /// multiplying by a magic number. See:
2310 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
2311 SDValue TargetLowering::BuildUDIV(SDNode *N, SelectionDAG &DAG,
2312 std::vector<SDNode*>* Created) const {
2313 MVT VT = N->getValueType(0);
2315 // Check to see if we can do this.
2316 if (!isTypeLegal(VT) || (VT != MVT::i32 && VT != MVT::i64))
2317 return SDValue(); // BuildUDIV only operates on i32 or i64
2319 uint64_t d = cast<ConstantSDNode>(N->getOperand(1))->getValue();
2320 mu magics = (VT == MVT::i32) ? magicu32(d) : magicu64(d);
2322 // Multiply the numerator (operand 0) by the magic value
2324 if (isOperationLegal(ISD::MULHU, VT))
2325 Q = DAG.getNode(ISD::MULHU, VT, N->getOperand(0),
2326 DAG.getConstant(magics.m, VT));
2327 else if (isOperationLegal(ISD::UMUL_LOHI, VT))
2328 Q = SDValue(DAG.getNode(ISD::UMUL_LOHI, DAG.getVTList(VT, VT),
2330 DAG.getConstant(magics.m, VT)).getNode(), 1);
2332 return SDValue(); // No mulhu or equvialent
2334 Created->push_back(Q.getNode());
2336 if (magics.a == 0) {
2337 return DAG.getNode(ISD::SRL, VT, Q,
2338 DAG.getConstant(magics.s, getShiftAmountTy()));
2340 SDValue NPQ = DAG.getNode(ISD::SUB, VT, N->getOperand(0), Q);
2342 Created->push_back(NPQ.getNode());
2343 NPQ = DAG.getNode(ISD::SRL, VT, NPQ,
2344 DAG.getConstant(1, getShiftAmountTy()));
2346 Created->push_back(NPQ.getNode());
2347 NPQ = DAG.getNode(ISD::ADD, VT, NPQ, Q);
2349 Created->push_back(NPQ.getNode());
2350 return DAG.getNode(ISD::SRL, VT, NPQ,
2351 DAG.getConstant(magics.s-1, getShiftAmountTy()));