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/TargetLowering.h"
15 #include "llvm/Target/TargetAsmInfo.h"
16 #include "llvm/Target/TargetData.h"
17 #include "llvm/Target/TargetLoweringObjectFile.h"
18 #include "llvm/Target/TargetMachine.h"
19 #include "llvm/Target/TargetRegisterInfo.h"
20 #include "llvm/Target/TargetSubtarget.h"
21 #include "llvm/GlobalVariable.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/CodeGen/MachineFrameInfo.h"
24 #include "llvm/CodeGen/SelectionDAG.h"
25 #include "llvm/ADT/StringExtras.h"
26 #include "llvm/ADT/STLExtras.h"
27 #include "llvm/Support/ErrorHandling.h"
28 #include "llvm/Support/MathExtras.h"
32 TLSModel::Model getTLSModel(const GlobalValue *GV, Reloc::Model reloc) {
33 bool isLocal = GV->hasLocalLinkage();
34 bool isDeclaration = GV->isDeclaration();
35 // FIXME: what should we do for protected and internal visibility?
36 // For variables, is internal different from hidden?
37 bool isHidden = GV->hasHiddenVisibility();
39 if (reloc == Reloc::PIC_) {
40 if (isLocal || isHidden)
41 return TLSModel::LocalDynamic;
43 return TLSModel::GeneralDynamic;
45 if (!isDeclaration || isHidden)
46 return TLSModel::LocalExec;
48 return TLSModel::InitialExec;
53 /// InitLibcallNames - Set default libcall names.
55 static void InitLibcallNames(const char **Names) {
56 Names[RTLIB::SHL_I16] = "__ashlhi3";
57 Names[RTLIB::SHL_I32] = "__ashlsi3";
58 Names[RTLIB::SHL_I64] = "__ashldi3";
59 Names[RTLIB::SHL_I128] = "__ashlti3";
60 Names[RTLIB::SRL_I16] = "__lshrhi3";
61 Names[RTLIB::SRL_I32] = "__lshrsi3";
62 Names[RTLIB::SRL_I64] = "__lshrdi3";
63 Names[RTLIB::SRL_I128] = "__lshrti3";
64 Names[RTLIB::SRA_I16] = "__ashrhi3";
65 Names[RTLIB::SRA_I32] = "__ashrsi3";
66 Names[RTLIB::SRA_I64] = "__ashrdi3";
67 Names[RTLIB::SRA_I128] = "__ashrti3";
68 Names[RTLIB::MUL_I16] = "__mulhi3";
69 Names[RTLIB::MUL_I32] = "__mulsi3";
70 Names[RTLIB::MUL_I64] = "__muldi3";
71 Names[RTLIB::MUL_I128] = "__multi3";
72 Names[RTLIB::SDIV_I16] = "__divhi3";
73 Names[RTLIB::SDIV_I32] = "__divsi3";
74 Names[RTLIB::SDIV_I64] = "__divdi3";
75 Names[RTLIB::SDIV_I128] = "__divti3";
76 Names[RTLIB::UDIV_I16] = "__udivhi3";
77 Names[RTLIB::UDIV_I32] = "__udivsi3";
78 Names[RTLIB::UDIV_I64] = "__udivdi3";
79 Names[RTLIB::UDIV_I128] = "__udivti3";
80 Names[RTLIB::SREM_I16] = "__modhi3";
81 Names[RTLIB::SREM_I32] = "__modsi3";
82 Names[RTLIB::SREM_I64] = "__moddi3";
83 Names[RTLIB::SREM_I128] = "__modti3";
84 Names[RTLIB::UREM_I16] = "__umodhi3";
85 Names[RTLIB::UREM_I32] = "__umodsi3";
86 Names[RTLIB::UREM_I64] = "__umoddi3";
87 Names[RTLIB::UREM_I128] = "__umodti3";
88 Names[RTLIB::NEG_I32] = "__negsi2";
89 Names[RTLIB::NEG_I64] = "__negdi2";
90 Names[RTLIB::ADD_F32] = "__addsf3";
91 Names[RTLIB::ADD_F64] = "__adddf3";
92 Names[RTLIB::ADD_F80] = "__addxf3";
93 Names[RTLIB::ADD_PPCF128] = "__gcc_qadd";
94 Names[RTLIB::SUB_F32] = "__subsf3";
95 Names[RTLIB::SUB_F64] = "__subdf3";
96 Names[RTLIB::SUB_F80] = "__subxf3";
97 Names[RTLIB::SUB_PPCF128] = "__gcc_qsub";
98 Names[RTLIB::MUL_F32] = "__mulsf3";
99 Names[RTLIB::MUL_F64] = "__muldf3";
100 Names[RTLIB::MUL_F80] = "__mulxf3";
101 Names[RTLIB::MUL_PPCF128] = "__gcc_qmul";
102 Names[RTLIB::DIV_F32] = "__divsf3";
103 Names[RTLIB::DIV_F64] = "__divdf3";
104 Names[RTLIB::DIV_F80] = "__divxf3";
105 Names[RTLIB::DIV_PPCF128] = "__gcc_qdiv";
106 Names[RTLIB::REM_F32] = "fmodf";
107 Names[RTLIB::REM_F64] = "fmod";
108 Names[RTLIB::REM_F80] = "fmodl";
109 Names[RTLIB::REM_PPCF128] = "fmodl";
110 Names[RTLIB::POWI_F32] = "__powisf2";
111 Names[RTLIB::POWI_F64] = "__powidf2";
112 Names[RTLIB::POWI_F80] = "__powixf2";
113 Names[RTLIB::POWI_PPCF128] = "__powitf2";
114 Names[RTLIB::SQRT_F32] = "sqrtf";
115 Names[RTLIB::SQRT_F64] = "sqrt";
116 Names[RTLIB::SQRT_F80] = "sqrtl";
117 Names[RTLIB::SQRT_PPCF128] = "sqrtl";
118 Names[RTLIB::LOG_F32] = "logf";
119 Names[RTLIB::LOG_F64] = "log";
120 Names[RTLIB::LOG_F80] = "logl";
121 Names[RTLIB::LOG_PPCF128] = "logl";
122 Names[RTLIB::LOG2_F32] = "log2f";
123 Names[RTLIB::LOG2_F64] = "log2";
124 Names[RTLIB::LOG2_F80] = "log2l";
125 Names[RTLIB::LOG2_PPCF128] = "log2l";
126 Names[RTLIB::LOG10_F32] = "log10f";
127 Names[RTLIB::LOG10_F64] = "log10";
128 Names[RTLIB::LOG10_F80] = "log10l";
129 Names[RTLIB::LOG10_PPCF128] = "log10l";
130 Names[RTLIB::EXP_F32] = "expf";
131 Names[RTLIB::EXP_F64] = "exp";
132 Names[RTLIB::EXP_F80] = "expl";
133 Names[RTLIB::EXP_PPCF128] = "expl";
134 Names[RTLIB::EXP2_F32] = "exp2f";
135 Names[RTLIB::EXP2_F64] = "exp2";
136 Names[RTLIB::EXP2_F80] = "exp2l";
137 Names[RTLIB::EXP2_PPCF128] = "exp2l";
138 Names[RTLIB::SIN_F32] = "sinf";
139 Names[RTLIB::SIN_F64] = "sin";
140 Names[RTLIB::SIN_F80] = "sinl";
141 Names[RTLIB::SIN_PPCF128] = "sinl";
142 Names[RTLIB::COS_F32] = "cosf";
143 Names[RTLIB::COS_F64] = "cos";
144 Names[RTLIB::COS_F80] = "cosl";
145 Names[RTLIB::COS_PPCF128] = "cosl";
146 Names[RTLIB::POW_F32] = "powf";
147 Names[RTLIB::POW_F64] = "pow";
148 Names[RTLIB::POW_F80] = "powl";
149 Names[RTLIB::POW_PPCF128] = "powl";
150 Names[RTLIB::CEIL_F32] = "ceilf";
151 Names[RTLIB::CEIL_F64] = "ceil";
152 Names[RTLIB::CEIL_F80] = "ceill";
153 Names[RTLIB::CEIL_PPCF128] = "ceill";
154 Names[RTLIB::TRUNC_F32] = "truncf";
155 Names[RTLIB::TRUNC_F64] = "trunc";
156 Names[RTLIB::TRUNC_F80] = "truncl";
157 Names[RTLIB::TRUNC_PPCF128] = "truncl";
158 Names[RTLIB::RINT_F32] = "rintf";
159 Names[RTLIB::RINT_F64] = "rint";
160 Names[RTLIB::RINT_F80] = "rintl";
161 Names[RTLIB::RINT_PPCF128] = "rintl";
162 Names[RTLIB::NEARBYINT_F32] = "nearbyintf";
163 Names[RTLIB::NEARBYINT_F64] = "nearbyint";
164 Names[RTLIB::NEARBYINT_F80] = "nearbyintl";
165 Names[RTLIB::NEARBYINT_PPCF128] = "nearbyintl";
166 Names[RTLIB::FLOOR_F32] = "floorf";
167 Names[RTLIB::FLOOR_F64] = "floor";
168 Names[RTLIB::FLOOR_F80] = "floorl";
169 Names[RTLIB::FLOOR_PPCF128] = "floorl";
170 Names[RTLIB::FPEXT_F32_F64] = "__extendsfdf2";
171 Names[RTLIB::FPROUND_F64_F32] = "__truncdfsf2";
172 Names[RTLIB::FPROUND_F80_F32] = "__truncxfsf2";
173 Names[RTLIB::FPROUND_PPCF128_F32] = "__trunctfsf2";
174 Names[RTLIB::FPROUND_F80_F64] = "__truncxfdf2";
175 Names[RTLIB::FPROUND_PPCF128_F64] = "__trunctfdf2";
176 Names[RTLIB::FPTOSINT_F32_I8] = "__fixsfi8";
177 Names[RTLIB::FPTOSINT_F32_I16] = "__fixsfi16";
178 Names[RTLIB::FPTOSINT_F32_I32] = "__fixsfsi";
179 Names[RTLIB::FPTOSINT_F32_I64] = "__fixsfdi";
180 Names[RTLIB::FPTOSINT_F32_I128] = "__fixsfti";
181 Names[RTLIB::FPTOSINT_F64_I32] = "__fixdfsi";
182 Names[RTLIB::FPTOSINT_F64_I64] = "__fixdfdi";
183 Names[RTLIB::FPTOSINT_F64_I128] = "__fixdfti";
184 Names[RTLIB::FPTOSINT_F80_I32] = "__fixxfsi";
185 Names[RTLIB::FPTOSINT_F80_I64] = "__fixxfdi";
186 Names[RTLIB::FPTOSINT_F80_I128] = "__fixxfti";
187 Names[RTLIB::FPTOSINT_PPCF128_I32] = "__fixtfsi";
188 Names[RTLIB::FPTOSINT_PPCF128_I64] = "__fixtfdi";
189 Names[RTLIB::FPTOSINT_PPCF128_I128] = "__fixtfti";
190 Names[RTLIB::FPTOUINT_F32_I8] = "__fixunssfi8";
191 Names[RTLIB::FPTOUINT_F32_I16] = "__fixunssfi16";
192 Names[RTLIB::FPTOUINT_F32_I32] = "__fixunssfsi";
193 Names[RTLIB::FPTOUINT_F32_I64] = "__fixunssfdi";
194 Names[RTLIB::FPTOUINT_F32_I128] = "__fixunssfti";
195 Names[RTLIB::FPTOUINT_F64_I32] = "__fixunsdfsi";
196 Names[RTLIB::FPTOUINT_F64_I64] = "__fixunsdfdi";
197 Names[RTLIB::FPTOUINT_F64_I128] = "__fixunsdfti";
198 Names[RTLIB::FPTOUINT_F80_I32] = "__fixunsxfsi";
199 Names[RTLIB::FPTOUINT_F80_I64] = "__fixunsxfdi";
200 Names[RTLIB::FPTOUINT_F80_I128] = "__fixunsxfti";
201 Names[RTLIB::FPTOUINT_PPCF128_I32] = "__fixunstfsi";
202 Names[RTLIB::FPTOUINT_PPCF128_I64] = "__fixunstfdi";
203 Names[RTLIB::FPTOUINT_PPCF128_I128] = "__fixunstfti";
204 Names[RTLIB::SINTTOFP_I32_F32] = "__floatsisf";
205 Names[RTLIB::SINTTOFP_I32_F64] = "__floatsidf";
206 Names[RTLIB::SINTTOFP_I32_F80] = "__floatsixf";
207 Names[RTLIB::SINTTOFP_I32_PPCF128] = "__floatsitf";
208 Names[RTLIB::SINTTOFP_I64_F32] = "__floatdisf";
209 Names[RTLIB::SINTTOFP_I64_F64] = "__floatdidf";
210 Names[RTLIB::SINTTOFP_I64_F80] = "__floatdixf";
211 Names[RTLIB::SINTTOFP_I64_PPCF128] = "__floatditf";
212 Names[RTLIB::SINTTOFP_I128_F32] = "__floattisf";
213 Names[RTLIB::SINTTOFP_I128_F64] = "__floattidf";
214 Names[RTLIB::SINTTOFP_I128_F80] = "__floattixf";
215 Names[RTLIB::SINTTOFP_I128_PPCF128] = "__floattitf";
216 Names[RTLIB::UINTTOFP_I32_F32] = "__floatunsisf";
217 Names[RTLIB::UINTTOFP_I32_F64] = "__floatunsidf";
218 Names[RTLIB::UINTTOFP_I32_F80] = "__floatunsixf";
219 Names[RTLIB::UINTTOFP_I32_PPCF128] = "__floatunsitf";
220 Names[RTLIB::UINTTOFP_I64_F32] = "__floatundisf";
221 Names[RTLIB::UINTTOFP_I64_F64] = "__floatundidf";
222 Names[RTLIB::UINTTOFP_I64_F80] = "__floatundixf";
223 Names[RTLIB::UINTTOFP_I64_PPCF128] = "__floatunditf";
224 Names[RTLIB::UINTTOFP_I128_F32] = "__floatuntisf";
225 Names[RTLIB::UINTTOFP_I128_F64] = "__floatuntidf";
226 Names[RTLIB::UINTTOFP_I128_F80] = "__floatuntixf";
227 Names[RTLIB::UINTTOFP_I128_PPCF128] = "__floatuntitf";
228 Names[RTLIB::OEQ_F32] = "__eqsf2";
229 Names[RTLIB::OEQ_F64] = "__eqdf2";
230 Names[RTLIB::UNE_F32] = "__nesf2";
231 Names[RTLIB::UNE_F64] = "__nedf2";
232 Names[RTLIB::OGE_F32] = "__gesf2";
233 Names[RTLIB::OGE_F64] = "__gedf2";
234 Names[RTLIB::OLT_F32] = "__ltsf2";
235 Names[RTLIB::OLT_F64] = "__ltdf2";
236 Names[RTLIB::OLE_F32] = "__lesf2";
237 Names[RTLIB::OLE_F64] = "__ledf2";
238 Names[RTLIB::OGT_F32] = "__gtsf2";
239 Names[RTLIB::OGT_F64] = "__gtdf2";
240 Names[RTLIB::UO_F32] = "__unordsf2";
241 Names[RTLIB::UO_F64] = "__unorddf2";
242 Names[RTLIB::O_F32] = "__unordsf2";
243 Names[RTLIB::O_F64] = "__unorddf2";
244 Names[RTLIB::MEMCPY] = "memcpy";
245 Names[RTLIB::MEMMOVE] = "memmove";
246 Names[RTLIB::MEMSET] = "memset";
247 Names[RTLIB::UNWIND_RESUME] = "_Unwind_Resume";
250 /// getFPEXT - Return the FPEXT_*_* value for the given types, or
251 /// UNKNOWN_LIBCALL if there is none.
252 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) {
253 if (OpVT == MVT::f32) {
254 if (RetVT == MVT::f64)
255 return FPEXT_F32_F64;
257 return UNKNOWN_LIBCALL;
260 /// getFPROUND - Return the FPROUND_*_* value for the given types, or
261 /// UNKNOWN_LIBCALL if there is none.
262 RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) {
263 if (RetVT == MVT::f32) {
264 if (OpVT == MVT::f64)
265 return FPROUND_F64_F32;
266 if (OpVT == MVT::f80)
267 return FPROUND_F80_F32;
268 if (OpVT == MVT::ppcf128)
269 return FPROUND_PPCF128_F32;
270 } else if (RetVT == MVT::f64) {
271 if (OpVT == MVT::f80)
272 return FPROUND_F80_F64;
273 if (OpVT == MVT::ppcf128)
274 return FPROUND_PPCF128_F64;
276 return UNKNOWN_LIBCALL;
279 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or
280 /// UNKNOWN_LIBCALL if there is none.
281 RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) {
282 if (OpVT == MVT::f32) {
283 if (RetVT == MVT::i8)
284 return FPTOSINT_F32_I8;
285 if (RetVT == MVT::i16)
286 return FPTOSINT_F32_I16;
287 if (RetVT == MVT::i32)
288 return FPTOSINT_F32_I32;
289 if (RetVT == MVT::i64)
290 return FPTOSINT_F32_I64;
291 if (RetVT == MVT::i128)
292 return FPTOSINT_F32_I128;
293 } else if (OpVT == MVT::f64) {
294 if (RetVT == MVT::i32)
295 return FPTOSINT_F64_I32;
296 if (RetVT == MVT::i64)
297 return FPTOSINT_F64_I64;
298 if (RetVT == MVT::i128)
299 return FPTOSINT_F64_I128;
300 } else if (OpVT == MVT::f80) {
301 if (RetVT == MVT::i32)
302 return FPTOSINT_F80_I32;
303 if (RetVT == MVT::i64)
304 return FPTOSINT_F80_I64;
305 if (RetVT == MVT::i128)
306 return FPTOSINT_F80_I128;
307 } else if (OpVT == MVT::ppcf128) {
308 if (RetVT == MVT::i32)
309 return FPTOSINT_PPCF128_I32;
310 if (RetVT == MVT::i64)
311 return FPTOSINT_PPCF128_I64;
312 if (RetVT == MVT::i128)
313 return FPTOSINT_PPCF128_I128;
315 return UNKNOWN_LIBCALL;
318 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or
319 /// UNKNOWN_LIBCALL if there is none.
320 RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) {
321 if (OpVT == MVT::f32) {
322 if (RetVT == MVT::i8)
323 return FPTOUINT_F32_I8;
324 if (RetVT == MVT::i16)
325 return FPTOUINT_F32_I16;
326 if (RetVT == MVT::i32)
327 return FPTOUINT_F32_I32;
328 if (RetVT == MVT::i64)
329 return FPTOUINT_F32_I64;
330 if (RetVT == MVT::i128)
331 return FPTOUINT_F32_I128;
332 } else if (OpVT == MVT::f64) {
333 if (RetVT == MVT::i32)
334 return FPTOUINT_F64_I32;
335 if (RetVT == MVT::i64)
336 return FPTOUINT_F64_I64;
337 if (RetVT == MVT::i128)
338 return FPTOUINT_F64_I128;
339 } else if (OpVT == MVT::f80) {
340 if (RetVT == MVT::i32)
341 return FPTOUINT_F80_I32;
342 if (RetVT == MVT::i64)
343 return FPTOUINT_F80_I64;
344 if (RetVT == MVT::i128)
345 return FPTOUINT_F80_I128;
346 } else if (OpVT == MVT::ppcf128) {
347 if (RetVT == MVT::i32)
348 return FPTOUINT_PPCF128_I32;
349 if (RetVT == MVT::i64)
350 return FPTOUINT_PPCF128_I64;
351 if (RetVT == MVT::i128)
352 return FPTOUINT_PPCF128_I128;
354 return UNKNOWN_LIBCALL;
357 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or
358 /// UNKNOWN_LIBCALL if there is none.
359 RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) {
360 if (OpVT == MVT::i32) {
361 if (RetVT == MVT::f32)
362 return SINTTOFP_I32_F32;
363 else if (RetVT == MVT::f64)
364 return SINTTOFP_I32_F64;
365 else if (RetVT == MVT::f80)
366 return SINTTOFP_I32_F80;
367 else if (RetVT == MVT::ppcf128)
368 return SINTTOFP_I32_PPCF128;
369 } else if (OpVT == MVT::i64) {
370 if (RetVT == MVT::f32)
371 return SINTTOFP_I64_F32;
372 else if (RetVT == MVT::f64)
373 return SINTTOFP_I64_F64;
374 else if (RetVT == MVT::f80)
375 return SINTTOFP_I64_F80;
376 else if (RetVT == MVT::ppcf128)
377 return SINTTOFP_I64_PPCF128;
378 } else if (OpVT == MVT::i128) {
379 if (RetVT == MVT::f32)
380 return SINTTOFP_I128_F32;
381 else if (RetVT == MVT::f64)
382 return SINTTOFP_I128_F64;
383 else if (RetVT == MVT::f80)
384 return SINTTOFP_I128_F80;
385 else if (RetVT == MVT::ppcf128)
386 return SINTTOFP_I128_PPCF128;
388 return UNKNOWN_LIBCALL;
391 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or
392 /// UNKNOWN_LIBCALL if there is none.
393 RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) {
394 if (OpVT == MVT::i32) {
395 if (RetVT == MVT::f32)
396 return UINTTOFP_I32_F32;
397 else if (RetVT == MVT::f64)
398 return UINTTOFP_I32_F64;
399 else if (RetVT == MVT::f80)
400 return UINTTOFP_I32_F80;
401 else if (RetVT == MVT::ppcf128)
402 return UINTTOFP_I32_PPCF128;
403 } else if (OpVT == MVT::i64) {
404 if (RetVT == MVT::f32)
405 return UINTTOFP_I64_F32;
406 else if (RetVT == MVT::f64)
407 return UINTTOFP_I64_F64;
408 else if (RetVT == MVT::f80)
409 return UINTTOFP_I64_F80;
410 else if (RetVT == MVT::ppcf128)
411 return UINTTOFP_I64_PPCF128;
412 } else if (OpVT == MVT::i128) {
413 if (RetVT == MVT::f32)
414 return UINTTOFP_I128_F32;
415 else if (RetVT == MVT::f64)
416 return UINTTOFP_I128_F64;
417 else if (RetVT == MVT::f80)
418 return UINTTOFP_I128_F80;
419 else if (RetVT == MVT::ppcf128)
420 return UINTTOFP_I128_PPCF128;
422 return UNKNOWN_LIBCALL;
425 /// InitCmpLibcallCCs - Set default comparison libcall CC.
427 static void InitCmpLibcallCCs(ISD::CondCode *CCs) {
428 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL);
429 CCs[RTLIB::OEQ_F32] = ISD::SETEQ;
430 CCs[RTLIB::OEQ_F64] = ISD::SETEQ;
431 CCs[RTLIB::UNE_F32] = ISD::SETNE;
432 CCs[RTLIB::UNE_F64] = ISD::SETNE;
433 CCs[RTLIB::OGE_F32] = ISD::SETGE;
434 CCs[RTLIB::OGE_F64] = ISD::SETGE;
435 CCs[RTLIB::OLT_F32] = ISD::SETLT;
436 CCs[RTLIB::OLT_F64] = ISD::SETLT;
437 CCs[RTLIB::OLE_F32] = ISD::SETLE;
438 CCs[RTLIB::OLE_F64] = ISD::SETLE;
439 CCs[RTLIB::OGT_F32] = ISD::SETGT;
440 CCs[RTLIB::OGT_F64] = ISD::SETGT;
441 CCs[RTLIB::UO_F32] = ISD::SETNE;
442 CCs[RTLIB::UO_F64] = ISD::SETNE;
443 CCs[RTLIB::O_F32] = ISD::SETEQ;
444 CCs[RTLIB::O_F64] = ISD::SETEQ;
447 /// NOTE: The constructor takes ownership of TLOF.
448 TargetLowering::TargetLowering(TargetMachine &tm,TargetLoweringObjectFile *tlof)
449 : TM(tm), TD(TM.getTargetData()), TLOF(*tlof) {
450 // All operations default to being supported.
451 memset(OpActions, 0, sizeof(OpActions));
452 memset(LoadExtActions, 0, sizeof(LoadExtActions));
453 memset(TruncStoreActions, 0, sizeof(TruncStoreActions));
454 memset(IndexedModeActions, 0, sizeof(IndexedModeActions));
455 memset(ConvertActions, 0, sizeof(ConvertActions));
456 memset(CondCodeActions, 0, sizeof(CondCodeActions));
458 // Set default actions for various operations.
459 for (unsigned VT = 0; VT != (unsigned)MVT::LAST_VALUETYPE; ++VT) {
460 // Default all indexed load / store to expand.
461 for (unsigned IM = (unsigned)ISD::PRE_INC;
462 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
463 setIndexedLoadAction(IM, (MVT::SimpleValueType)VT, Expand);
464 setIndexedStoreAction(IM, (MVT::SimpleValueType)VT, Expand);
467 // These operations default to expand.
468 setOperationAction(ISD::FGETSIGN, (MVT::SimpleValueType)VT, Expand);
469 setOperationAction(ISD::CONCAT_VECTORS, (MVT::SimpleValueType)VT, Expand);
472 // Most targets ignore the @llvm.prefetch intrinsic.
473 setOperationAction(ISD::PREFETCH, MVT::Other, Expand);
475 // ConstantFP nodes default to expand. Targets can either change this to
476 // Legal, in which case all fp constants are legal, or use addLegalFPImmediate
477 // to optimize expansions for certain constants.
478 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
479 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
480 setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
482 // These library functions default to expand.
483 setOperationAction(ISD::FLOG , MVT::f64, Expand);
484 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
485 setOperationAction(ISD::FLOG10,MVT::f64, Expand);
486 setOperationAction(ISD::FEXP , MVT::f64, Expand);
487 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
488 setOperationAction(ISD::FLOG , MVT::f32, Expand);
489 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
490 setOperationAction(ISD::FLOG10,MVT::f32, Expand);
491 setOperationAction(ISD::FEXP , MVT::f32, Expand);
492 setOperationAction(ISD::FEXP2, MVT::f32, Expand);
494 // Default ISD::TRAP to expand (which turns it into abort).
495 setOperationAction(ISD::TRAP, MVT::Other, Expand);
497 IsLittleEndian = TD->isLittleEndian();
498 UsesGlobalOffsetTable = false;
499 ShiftAmountTy = PointerTy =
500 getValueType(TD->getIntPtrType()).getSimpleVT().SimpleTy;
501 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE*sizeof(TargetRegisterClass*));
502 memset(TargetDAGCombineArray, 0, array_lengthof(TargetDAGCombineArray));
503 maxStoresPerMemset = maxStoresPerMemcpy = maxStoresPerMemmove = 8;
504 allowUnalignedMemoryAccesses = false;
505 benefitFromCodePlacementOpt = false;
506 UseUnderscoreSetJmp = false;
507 UseUnderscoreLongJmp = false;
508 SelectIsExpensive = false;
509 IntDivIsCheap = false;
510 Pow2DivIsCheap = false;
511 StackPointerRegisterToSaveRestore = 0;
512 ExceptionPointerRegister = 0;
513 ExceptionSelectorRegister = 0;
514 BooleanContents = UndefinedBooleanContent;
515 SchedPreferenceInfo = SchedulingForLatency;
517 JumpBufAlignment = 0;
518 IfCvtBlockSizeLimit = 2;
519 IfCvtDupBlockSizeLimit = 0;
520 PrefLoopAlignment = 0;
522 InitLibcallNames(LibcallRoutineNames);
523 InitCmpLibcallCCs(CmpLibcallCCs);
525 // Tell Legalize whether the assembler supports DEBUG_LOC.
526 const TargetAsmInfo *TASM = TM.getTargetAsmInfo();
527 if (!TASM || !TASM->hasDotLocAndDotFile())
528 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
531 TargetLowering::~TargetLowering() {
535 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT,
536 unsigned &NumIntermediates,
538 TargetLowering* TLI) {
539 // Figure out the right, legal destination reg to copy into.
540 unsigned NumElts = VT.getVectorNumElements();
541 MVT EltTy = VT.getVectorElementType();
543 unsigned NumVectorRegs = 1;
545 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
546 // could break down into LHS/RHS like LegalizeDAG does.
547 if (!isPowerOf2_32(NumElts)) {
548 NumVectorRegs = NumElts;
552 // Divide the input until we get to a supported size. This will always
553 // end with a scalar if the target doesn't support vectors.
554 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) {
559 NumIntermediates = NumVectorRegs;
561 MVT NewVT = MVT::getVectorVT(EltTy, NumElts);
562 if (!TLI->isTypeLegal(NewVT))
564 IntermediateVT = NewVT;
566 EVT DestVT = TLI->getRegisterType(NewVT);
568 if (EVT(DestVT).bitsLT(NewVT)) {
569 // Value is expanded, e.g. i64 -> i16.
570 return NumVectorRegs*(NewVT.getSizeInBits()/DestVT.getSizeInBits());
572 // Otherwise, promotion or legal types use the same number of registers as
573 // the vector decimated to the appropriate level.
574 return NumVectorRegs;
580 /// computeRegisterProperties - Once all of the register classes are added,
581 /// this allows us to compute derived properties we expose.
582 void TargetLowering::computeRegisterProperties() {
583 assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE &&
584 "Too many value types for ValueTypeActions to hold!");
586 // Everything defaults to needing one register.
587 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
588 NumRegistersForVT[i] = 1;
589 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i;
591 // ...except isVoid, which doesn't need any registers.
592 NumRegistersForVT[MVT::isVoid] = 0;
594 // Find the largest integer register class.
595 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE;
596 for (; RegClassForVT[LargestIntReg] == 0; --LargestIntReg)
597 assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
599 // Every integer value type larger than this largest register takes twice as
600 // many registers to represent as the previous ValueType.
601 for (unsigned ExpandedReg = LargestIntReg + 1; ; ++ExpandedReg) {
602 EVT EVT = (MVT::SimpleValueType)ExpandedReg;
603 if (!EVT.isInteger())
605 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
606 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg;
607 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1);
608 ValueTypeActions.setTypeAction(EVT, Expand);
611 // Inspect all of the ValueType's smaller than the largest integer
612 // register to see which ones need promotion.
613 unsigned LegalIntReg = LargestIntReg;
614 for (unsigned IntReg = LargestIntReg - 1;
615 IntReg >= (unsigned)MVT::i1; --IntReg) {
616 EVT IVT = (MVT::SimpleValueType)IntReg;
617 if (isTypeLegal(IVT)) {
618 LegalIntReg = IntReg;
620 RegisterTypeForVT[IntReg] = TransformToType[IntReg] =
621 (MVT::SimpleValueType)LegalIntReg;
622 ValueTypeActions.setTypeAction(IVT, Promote);
626 // ppcf128 type is really two f64's.
627 if (!isTypeLegal(MVT::ppcf128)) {
628 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64];
629 RegisterTypeForVT[MVT::ppcf128] = MVT::f64;
630 TransformToType[MVT::ppcf128] = MVT::f64;
631 ValueTypeActions.setTypeAction(MVT::ppcf128, Expand);
634 // Decide how to handle f64. If the target does not have native f64 support,
635 // expand it to i64 and we will be generating soft float library calls.
636 if (!isTypeLegal(MVT::f64)) {
637 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
638 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
639 TransformToType[MVT::f64] = MVT::i64;
640 ValueTypeActions.setTypeAction(MVT::f64, Expand);
643 // Decide how to handle f32. If the target does not have native support for
644 // f32, promote it to f64 if it is legal. Otherwise, expand it to i32.
645 if (!isTypeLegal(MVT::f32)) {
646 if (isTypeLegal(MVT::f64)) {
647 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::f64];
648 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::f64];
649 TransformToType[MVT::f32] = MVT::f64;
650 ValueTypeActions.setTypeAction(MVT::f32, Promote);
652 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
653 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
654 TransformToType[MVT::f32] = MVT::i32;
655 ValueTypeActions.setTypeAction(MVT::f32, Expand);
659 // Loop over all of the vector value types to see which need transformations.
660 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
661 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
662 MVT VT = (MVT::SimpleValueType)i;
663 if (!isTypeLegal(VT)) {
666 unsigned NumIntermediates;
667 NumRegistersForVT[i] =
668 getVectorTypeBreakdownMVT(VT, IntermediateVT, NumIntermediates,
670 RegisterTypeForVT[i] = RegisterVT;
672 // Determine if there is a legal wider type.
673 bool IsLegalWiderType = false;
674 EVT EltVT = VT.getVectorElementType();
675 unsigned NElts = VT.getVectorNumElements();
676 for (unsigned nVT = i+1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
677 EVT SVT = (MVT::SimpleValueType)nVT;
678 if (isTypeLegal(SVT) && SVT.getVectorElementType() == EltVT &&
679 SVT.getVectorNumElements() > NElts) {
680 TransformToType[i] = SVT;
681 ValueTypeActions.setTypeAction(VT, Promote);
682 IsLegalWiderType = true;
686 if (!IsLegalWiderType) {
687 EVT NVT = VT.getPow2VectorType();
689 // Type is already a power of 2. The default action is to split.
690 TransformToType[i] = MVT::Other;
691 ValueTypeActions.setTypeAction(VT, Expand);
693 TransformToType[i] = NVT;
694 ValueTypeActions.setTypeAction(VT, Promote);
701 const char *TargetLowering::getTargetNodeName(unsigned Opcode) const {
706 MVT::SimpleValueType TargetLowering::getSetCCResultType(EVT VT) const {
707 return getValueType(TD->getIntPtrType()).getSimpleVT().SimpleTy;
710 /// getVectorTypeBreakdown - Vector types are broken down into some number of
711 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
712 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
713 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
715 /// This method returns the number of registers needed, and the VT for each
716 /// register. It also returns the VT and quantity of the intermediate values
717 /// before they are promoted/expanded.
719 unsigned TargetLowering::getVectorTypeBreakdown(LLVMContext &Context, EVT VT,
721 unsigned &NumIntermediates,
722 EVT &RegisterVT) const {
723 // Figure out the right, legal destination reg to copy into.
724 unsigned NumElts = VT.getVectorNumElements();
725 EVT EltTy = VT.getVectorElementType();
727 unsigned NumVectorRegs = 1;
729 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
730 // could break down into LHS/RHS like LegalizeDAG does.
731 if (!isPowerOf2_32(NumElts)) {
732 NumVectorRegs = NumElts;
736 // Divide the input until we get to a supported size. This will always
737 // end with a scalar if the target doesn't support vectors.
738 while (NumElts > 1 && !isTypeLegal(
739 EVT::getVectorVT(Context, EltTy, NumElts))) {
744 NumIntermediates = NumVectorRegs;
746 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts);
747 if (!isTypeLegal(NewVT))
749 IntermediateVT = NewVT;
751 EVT DestVT = getRegisterType(Context, NewVT);
753 if (DestVT.bitsLT(NewVT)) {
754 // Value is expanded, e.g. i64 -> i16.
755 return NumVectorRegs*(NewVT.getSizeInBits()/DestVT.getSizeInBits());
757 // Otherwise, promotion or legal types use the same number of registers as
758 // the vector decimated to the appropriate level.
759 return NumVectorRegs;
765 /// getWidenVectorType: given a vector type, returns the type to widen to
766 /// (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
767 /// If there is no vector type that we want to widen to, returns MVT::Other
768 /// When and where to widen is target dependent based on the cost of
769 /// scalarizing vs using the wider vector type.
770 EVT TargetLowering::getWidenVectorType(EVT VT) const {
771 assert(VT.isVector());
775 // Default is not to widen until moved to LegalizeTypes
779 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
780 /// function arguments in the caller parameter area. This is the actual
781 /// alignment, not its logarithm.
782 unsigned TargetLowering::getByValTypeAlignment(const Type *Ty) const {
783 return TD->getCallFrameTypeAlignment(Ty);
786 SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table,
787 SelectionDAG &DAG) const {
788 if (usesGlobalOffsetTable())
789 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy());
794 TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
795 // Assume that everything is safe in static mode.
796 if (getTargetMachine().getRelocationModel() == Reloc::Static)
799 // In dynamic-no-pic mode, assume that known defined values are safe.
800 if (getTargetMachine().getRelocationModel() == Reloc::DynamicNoPIC &&
802 !GA->getGlobal()->isDeclaration() &&
803 !GA->getGlobal()->isWeakForLinker())
806 // Otherwise assume nothing is safe.
810 //===----------------------------------------------------------------------===//
811 // Optimization Methods
812 //===----------------------------------------------------------------------===//
814 /// ShrinkDemandedConstant - Check to see if the specified operand of the
815 /// specified instruction is a constant integer. If so, check to see if there
816 /// are any bits set in the constant that are not demanded. If so, shrink the
817 /// constant and return true.
818 bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDValue Op,
819 const APInt &Demanded) {
820 DebugLoc dl = Op.getDebugLoc();
822 // FIXME: ISD::SELECT, ISD::SELECT_CC
823 switch (Op.getOpcode()) {
828 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
829 if (!C) return false;
831 if (Op.getOpcode() == ISD::XOR &&
832 (C->getAPIntValue() | (~Demanded)).isAllOnesValue())
835 // if we can expand it to have all bits set, do it
836 if (C->getAPIntValue().intersects(~Demanded)) {
837 EVT VT = Op.getValueType();
838 SDValue New = DAG.getNode(Op.getOpcode(), dl, VT, Op.getOperand(0),
839 DAG.getConstant(Demanded &
842 return CombineTo(Op, New);
852 /// ShrinkDemandedOp - Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the
853 /// casts are free. This uses isZExtFree and ZERO_EXTEND for the widening
854 /// cast, but it could be generalized for targets with other types of
855 /// implicit widening casts.
857 TargetLowering::TargetLoweringOpt::ShrinkDemandedOp(SDValue Op,
859 const APInt &Demanded,
861 assert(Op.getNumOperands() == 2 &&
862 "ShrinkDemandedOp only supports binary operators!");
863 assert(Op.getNode()->getNumValues() == 1 &&
864 "ShrinkDemandedOp only supports nodes with one result!");
866 // Don't do this if the node has another user, which may require the
868 if (!Op.getNode()->hasOneUse())
871 // Search for the smallest integer type with free casts to and from
872 // Op's type. For expedience, just check power-of-2 integer types.
873 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
874 unsigned SmallVTBits = BitWidth - Demanded.countLeadingZeros();
875 if (!isPowerOf2_32(SmallVTBits))
876 SmallVTBits = NextPowerOf2(SmallVTBits);
877 for (; SmallVTBits < BitWidth; SmallVTBits = NextPowerOf2(SmallVTBits)) {
878 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), SmallVTBits);
879 if (TLI.isTruncateFree(Op.getValueType(), SmallVT) &&
880 TLI.isZExtFree(SmallVT, Op.getValueType())) {
881 // We found a type with free casts.
882 SDValue X = DAG.getNode(Op.getOpcode(), dl, SmallVT,
883 DAG.getNode(ISD::TRUNCATE, dl, SmallVT,
884 Op.getNode()->getOperand(0)),
885 DAG.getNode(ISD::TRUNCATE, dl, SmallVT,
886 Op.getNode()->getOperand(1)));
887 SDValue Z = DAG.getNode(ISD::ZERO_EXTEND, dl, Op.getValueType(), X);
888 return CombineTo(Op, Z);
894 /// SimplifyDemandedBits - Look at Op. At this point, we know that only the
895 /// DemandedMask bits of the result of Op are ever used downstream. If we can
896 /// use this information to simplify Op, create a new simplified DAG node and
897 /// return true, returning the original and new nodes in Old and New. Otherwise,
898 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
899 /// the expression (used to simplify the caller). The KnownZero/One bits may
900 /// only be accurate for those bits in the DemandedMask.
901 bool TargetLowering::SimplifyDemandedBits(SDValue Op,
902 const APInt &DemandedMask,
905 TargetLoweringOpt &TLO,
906 unsigned Depth) const {
907 unsigned BitWidth = DemandedMask.getBitWidth();
908 assert(Op.getValueSizeInBits() == BitWidth &&
909 "Mask size mismatches value type size!");
910 APInt NewMask = DemandedMask;
911 DebugLoc dl = Op.getDebugLoc();
913 // Don't know anything.
914 KnownZero = KnownOne = APInt(BitWidth, 0);
916 // Other users may use these bits.
917 if (!Op.getNode()->hasOneUse()) {
919 // If not at the root, Just compute the KnownZero/KnownOne bits to
920 // simplify things downstream.
921 TLO.DAG.ComputeMaskedBits(Op, DemandedMask, KnownZero, KnownOne, Depth);
924 // If this is the root being simplified, allow it to have multiple uses,
925 // just set the NewMask to all bits.
926 NewMask = APInt::getAllOnesValue(BitWidth);
927 } else if (DemandedMask == 0) {
928 // Not demanding any bits from Op.
929 if (Op.getOpcode() != ISD::UNDEF)
930 return TLO.CombineTo(Op, TLO.DAG.getUNDEF(Op.getValueType()));
932 } else if (Depth == 6) { // Limit search depth.
936 APInt KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut;
937 switch (Op.getOpcode()) {
939 // We know all of the bits for a constant!
940 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue() & NewMask;
941 KnownZero = ~KnownOne & NewMask;
942 return false; // Don't fall through, will infinitely loop.
944 // If the RHS is a constant, check to see if the LHS would be zero without
945 // using the bits from the RHS. Below, we use knowledge about the RHS to
946 // simplify the LHS, here we're using information from the LHS to simplify
948 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
949 APInt LHSZero, LHSOne;
950 TLO.DAG.ComputeMaskedBits(Op.getOperand(0), NewMask,
951 LHSZero, LHSOne, Depth+1);
952 // If the LHS already has zeros where RHSC does, this and is dead.
953 if ((LHSZero & NewMask) == (~RHSC->getAPIntValue() & NewMask))
954 return TLO.CombineTo(Op, Op.getOperand(0));
955 // If any of the set bits in the RHS are known zero on the LHS, shrink
957 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & NewMask))
961 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
962 KnownOne, TLO, Depth+1))
964 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
965 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownZero & NewMask,
966 KnownZero2, KnownOne2, TLO, Depth+1))
968 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
970 // If all of the demanded bits are known one on one side, return the other.
971 // These bits cannot contribute to the result of the 'and'.
972 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
973 return TLO.CombineTo(Op, Op.getOperand(0));
974 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
975 return TLO.CombineTo(Op, Op.getOperand(1));
976 // If all of the demanded bits in the inputs are known zeros, return zero.
977 if ((NewMask & (KnownZero|KnownZero2)) == NewMask)
978 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, Op.getValueType()));
979 // If the RHS is a constant, see if we can simplify it.
980 if (TLO.ShrinkDemandedConstant(Op, ~KnownZero2 & NewMask))
982 // If the operation can be done in a smaller type, do so.
983 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
986 // Output known-1 bits are only known if set in both the LHS & RHS.
987 KnownOne &= KnownOne2;
988 // Output known-0 are known to be clear if zero in either the LHS | RHS.
989 KnownZero |= KnownZero2;
992 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
993 KnownOne, TLO, Depth+1))
995 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
996 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownOne & NewMask,
997 KnownZero2, KnownOne2, TLO, Depth+1))
999 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1001 // If all of the demanded bits are known zero on one side, return the other.
1002 // These bits cannot contribute to the result of the 'or'.
1003 if ((NewMask & ~KnownOne2 & KnownZero) == (~KnownOne2 & NewMask))
1004 return TLO.CombineTo(Op, Op.getOperand(0));
1005 if ((NewMask & ~KnownOne & KnownZero2) == (~KnownOne & NewMask))
1006 return TLO.CombineTo(Op, Op.getOperand(1));
1007 // If all of the potentially set bits on one side are known to be set on
1008 // the other side, just use the 'other' side.
1009 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
1010 return TLO.CombineTo(Op, Op.getOperand(0));
1011 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
1012 return TLO.CombineTo(Op, Op.getOperand(1));
1013 // If the RHS is a constant, see if we can simplify it.
1014 if (TLO.ShrinkDemandedConstant(Op, NewMask))
1016 // If the operation can be done in a smaller type, do so.
1017 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1020 // Output known-0 bits are only known if clear in both the LHS & RHS.
1021 KnownZero &= KnownZero2;
1022 // Output known-1 are known to be set if set in either the LHS | RHS.
1023 KnownOne |= KnownOne2;
1026 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
1027 KnownOne, TLO, Depth+1))
1029 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1030 if (SimplifyDemandedBits(Op.getOperand(0), NewMask, KnownZero2,
1031 KnownOne2, TLO, Depth+1))
1033 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1035 // If all of the demanded bits are known zero on one side, return the other.
1036 // These bits cannot contribute to the result of the 'xor'.
1037 if ((KnownZero & NewMask) == NewMask)
1038 return TLO.CombineTo(Op, Op.getOperand(0));
1039 if ((KnownZero2 & NewMask) == NewMask)
1040 return TLO.CombineTo(Op, Op.getOperand(1));
1041 // If the operation can be done in a smaller type, do so.
1042 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1045 // If all of the unknown bits are known to be zero on one side or the other
1046 // (but not both) turn this into an *inclusive* or.
1047 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1048 if ((NewMask & ~KnownZero & ~KnownZero2) == 0)
1049 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, dl, Op.getValueType(),
1053 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1054 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
1055 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1056 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
1058 // If all of the demanded bits on one side are known, and all of the set
1059 // bits on that side are also known to be set on the other side, turn this
1060 // into an AND, as we know the bits will be cleared.
1061 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1062 if ((NewMask & (KnownZero|KnownOne)) == NewMask) { // all known
1063 if ((KnownOne & KnownOne2) == KnownOne) {
1064 EVT VT = Op.getValueType();
1065 SDValue ANDC = TLO.DAG.getConstant(~KnownOne & NewMask, VT);
1066 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, dl, VT,
1067 Op.getOperand(0), ANDC));
1071 // If the RHS is a constant, see if we can simplify it.
1072 // for XOR, we prefer to force bits to 1 if they will make a -1.
1073 // if we can't force bits, try to shrink constant
1074 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1075 APInt Expanded = C->getAPIntValue() | (~NewMask);
1076 // if we can expand it to have all bits set, do it
1077 if (Expanded.isAllOnesValue()) {
1078 if (Expanded != C->getAPIntValue()) {
1079 EVT VT = Op.getValueType();
1080 SDValue New = TLO.DAG.getNode(Op.getOpcode(), dl,VT, Op.getOperand(0),
1081 TLO.DAG.getConstant(Expanded, VT));
1082 return TLO.CombineTo(Op, New);
1084 // if it already has all the bits set, nothing to change
1085 // but don't shrink either!
1086 } else if (TLO.ShrinkDemandedConstant(Op, NewMask)) {
1091 KnownZero = KnownZeroOut;
1092 KnownOne = KnownOneOut;
1095 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero,
1096 KnownOne, TLO, Depth+1))
1098 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero2,
1099 KnownOne2, TLO, Depth+1))
1101 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1102 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1104 // If the operands are constants, see if we can simplify them.
1105 if (TLO.ShrinkDemandedConstant(Op, NewMask))
1108 // Only known if known in both the LHS and RHS.
1109 KnownOne &= KnownOne2;
1110 KnownZero &= KnownZero2;
1112 case ISD::SELECT_CC:
1113 if (SimplifyDemandedBits(Op.getOperand(3), NewMask, KnownZero,
1114 KnownOne, TLO, Depth+1))
1116 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero2,
1117 KnownOne2, TLO, Depth+1))
1119 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1120 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1122 // If the operands are constants, see if we can simplify them.
1123 if (TLO.ShrinkDemandedConstant(Op, NewMask))
1126 // Only known if known in both the LHS and RHS.
1127 KnownOne &= KnownOne2;
1128 KnownZero &= KnownZero2;
1131 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1132 unsigned ShAmt = SA->getZExtValue();
1133 SDValue InOp = Op.getOperand(0);
1135 // If the shift count is an invalid immediate, don't do anything.
1136 if (ShAmt >= BitWidth)
1139 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a
1140 // single shift. We can do this if the bottom bits (which are shifted
1141 // out) are never demanded.
1142 if (InOp.getOpcode() == ISD::SRL &&
1143 isa<ConstantSDNode>(InOp.getOperand(1))) {
1144 if (ShAmt && (NewMask & APInt::getLowBitsSet(BitWidth, ShAmt)) == 0) {
1145 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
1146 unsigned Opc = ISD::SHL;
1147 int Diff = ShAmt-C1;
1154 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
1155 EVT VT = Op.getValueType();
1156 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT,
1157 InOp.getOperand(0), NewSA));
1161 if (SimplifyDemandedBits(Op.getOperand(0), NewMask.lshr(ShAmt),
1162 KnownZero, KnownOne, TLO, Depth+1))
1164 KnownZero <<= SA->getZExtValue();
1165 KnownOne <<= SA->getZExtValue();
1166 // low bits known zero.
1167 KnownZero |= APInt::getLowBitsSet(BitWidth, SA->getZExtValue());
1171 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1172 EVT VT = Op.getValueType();
1173 unsigned ShAmt = SA->getZExtValue();
1174 unsigned VTSize = VT.getSizeInBits();
1175 SDValue InOp = Op.getOperand(0);
1177 // If the shift count is an invalid immediate, don't do anything.
1178 if (ShAmt >= BitWidth)
1181 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a
1182 // single shift. We can do this if the top bits (which are shifted out)
1183 // are never demanded.
1184 if (InOp.getOpcode() == ISD::SHL &&
1185 isa<ConstantSDNode>(InOp.getOperand(1))) {
1186 if (ShAmt && (NewMask & APInt::getHighBitsSet(VTSize, ShAmt)) == 0) {
1187 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
1188 unsigned Opc = ISD::SRL;
1189 int Diff = ShAmt-C1;
1196 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
1197 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT,
1198 InOp.getOperand(0), NewSA));
1202 // Compute the new bits that are at the top now.
1203 if (SimplifyDemandedBits(InOp, (NewMask << ShAmt),
1204 KnownZero, KnownOne, TLO, Depth+1))
1206 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1207 KnownZero = KnownZero.lshr(ShAmt);
1208 KnownOne = KnownOne.lshr(ShAmt);
1210 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
1211 KnownZero |= HighBits; // High bits known zero.
1215 // If this is an arithmetic shift right and only the low-bit is set, we can
1216 // always convert this into a logical shr, even if the shift amount is
1217 // variable. The low bit of the shift cannot be an input sign bit unless
1218 // the shift amount is >= the size of the datatype, which is undefined.
1219 if (DemandedMask == 1)
1220 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, Op.getValueType(),
1221 Op.getOperand(0), Op.getOperand(1)));
1223 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1224 EVT VT = Op.getValueType();
1225 unsigned ShAmt = SA->getZExtValue();
1227 // If the shift count is an invalid immediate, don't do anything.
1228 if (ShAmt >= BitWidth)
1231 APInt InDemandedMask = (NewMask << ShAmt);
1233 // If any of the demanded bits are produced by the sign extension, we also
1234 // demand the input sign bit.
1235 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
1236 if (HighBits.intersects(NewMask))
1237 InDemandedMask |= APInt::getSignBit(VT.getSizeInBits());
1239 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask,
1240 KnownZero, KnownOne, TLO, Depth+1))
1242 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1243 KnownZero = KnownZero.lshr(ShAmt);
1244 KnownOne = KnownOne.lshr(ShAmt);
1246 // Handle the sign bit, adjusted to where it is now in the mask.
1247 APInt SignBit = APInt::getSignBit(BitWidth).lshr(ShAmt);
1249 // If the input sign bit is known to be zero, or if none of the top bits
1250 // are demanded, turn this into an unsigned shift right.
1251 if (KnownZero.intersects(SignBit) || (HighBits & ~NewMask) == HighBits) {
1252 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, VT,
1255 } else if (KnownOne.intersects(SignBit)) { // New bits are known one.
1256 KnownOne |= HighBits;
1260 case ISD::SIGN_EXTEND_INREG: {
1261 EVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1263 // Sign extension. Compute the demanded bits in the result that are not
1264 // present in the input.
1265 APInt NewBits = APInt::getHighBitsSet(BitWidth,
1266 BitWidth - EVT.getSizeInBits()) &
1269 // If none of the extended bits are demanded, eliminate the sextinreg.
1271 return TLO.CombineTo(Op, Op.getOperand(0));
1273 APInt InSignBit = APInt::getSignBit(EVT.getSizeInBits());
1274 InSignBit.zext(BitWidth);
1275 APInt InputDemandedBits = APInt::getLowBitsSet(BitWidth,
1276 EVT.getSizeInBits()) &
1279 // Since the sign extended bits are demanded, we know that the sign
1281 InputDemandedBits |= InSignBit;
1283 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits,
1284 KnownZero, KnownOne, TLO, Depth+1))
1286 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1288 // If the sign bit of the input is known set or clear, then we know the
1289 // top bits of the result.
1291 // If the input sign bit is known zero, convert this into a zero extension.
1292 if (KnownZero.intersects(InSignBit))
1293 return TLO.CombineTo(Op,
1294 TLO.DAG.getZeroExtendInReg(Op.getOperand(0),dl,EVT));
1296 if (KnownOne.intersects(InSignBit)) { // Input sign bit known set
1297 KnownOne |= NewBits;
1298 KnownZero &= ~NewBits;
1299 } else { // Input sign bit unknown
1300 KnownZero &= ~NewBits;
1301 KnownOne &= ~NewBits;
1305 case ISD::ZERO_EXTEND: {
1306 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits();
1307 APInt InMask = NewMask;
1308 InMask.trunc(OperandBitWidth);
1310 // If none of the top bits are demanded, convert this into an any_extend.
1312 APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask;
1313 if (!NewBits.intersects(NewMask))
1314 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND, dl,
1318 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1319 KnownZero, KnownOne, TLO, Depth+1))
1321 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1322 KnownZero.zext(BitWidth);
1323 KnownOne.zext(BitWidth);
1324 KnownZero |= NewBits;
1327 case ISD::SIGN_EXTEND: {
1328 EVT InVT = Op.getOperand(0).getValueType();
1329 unsigned InBits = InVT.getSizeInBits();
1330 APInt InMask = APInt::getLowBitsSet(BitWidth, InBits);
1331 APInt InSignBit = APInt::getBitsSet(BitWidth, InBits - 1, InBits);
1332 APInt NewBits = ~InMask & NewMask;
1334 // If none of the top bits are demanded, convert this into an any_extend.
1336 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND, dl,
1340 // Since some of the sign extended bits are demanded, we know that the sign
1342 APInt InDemandedBits = InMask & NewMask;
1343 InDemandedBits |= InSignBit;
1344 InDemandedBits.trunc(InBits);
1346 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero,
1347 KnownOne, TLO, Depth+1))
1349 KnownZero.zext(BitWidth);
1350 KnownOne.zext(BitWidth);
1352 // If the sign bit is known zero, convert this to a zero extend.
1353 if (KnownZero.intersects(InSignBit))
1354 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND, dl,
1358 // If the sign bit is known one, the top bits match.
1359 if (KnownOne.intersects(InSignBit)) {
1360 KnownOne |= NewBits;
1361 KnownZero &= ~NewBits;
1362 } else { // Otherwise, top bits aren't known.
1363 KnownOne &= ~NewBits;
1364 KnownZero &= ~NewBits;
1368 case ISD::ANY_EXTEND: {
1369 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits();
1370 APInt InMask = NewMask;
1371 InMask.trunc(OperandBitWidth);
1372 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1373 KnownZero, KnownOne, TLO, Depth+1))
1375 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1376 KnownZero.zext(BitWidth);
1377 KnownOne.zext(BitWidth);
1380 case ISD::TRUNCATE: {
1381 // Simplify the input, using demanded bit information, and compute the known
1382 // zero/one bits live out.
1383 APInt TruncMask = NewMask;
1384 TruncMask.zext(Op.getOperand(0).getValueSizeInBits());
1385 if (SimplifyDemandedBits(Op.getOperand(0), TruncMask,
1386 KnownZero, KnownOne, TLO, Depth+1))
1388 KnownZero.trunc(BitWidth);
1389 KnownOne.trunc(BitWidth);
1391 // If the input is only used by this truncate, see if we can shrink it based
1392 // on the known demanded bits.
1393 if (Op.getOperand(0).getNode()->hasOneUse()) {
1394 SDValue In = Op.getOperand(0);
1395 unsigned InBitWidth = In.getValueSizeInBits();
1396 switch (In.getOpcode()) {
1399 // Shrink SRL by a constant if none of the high bits shifted in are
1401 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1))){
1402 APInt HighBits = APInt::getHighBitsSet(InBitWidth,
1403 InBitWidth - BitWidth);
1404 HighBits = HighBits.lshr(ShAmt->getZExtValue());
1405 HighBits.trunc(BitWidth);
1407 if (ShAmt->getZExtValue() < BitWidth && !(HighBits & NewMask)) {
1408 // None of the shifted in bits are needed. Add a truncate of the
1409 // shift input, then shift it.
1410 SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE, dl,
1413 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl,
1423 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1426 case ISD::AssertZext: {
1427 EVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1428 APInt InMask = APInt::getLowBitsSet(BitWidth,
1429 VT.getSizeInBits());
1430 if (SimplifyDemandedBits(Op.getOperand(0), InMask & NewMask,
1431 KnownZero, KnownOne, TLO, Depth+1))
1433 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1434 KnownZero |= ~InMask & NewMask;
1437 case ISD::BIT_CONVERT:
1439 // If this is an FP->Int bitcast and if the sign bit is the only thing that
1440 // is demanded, turn this into a FGETSIGN.
1441 if (NewMask == EVT::getIntegerVTSignBit(Op.getValueType()) &&
1442 MVT::isFloatingPoint(Op.getOperand(0).getValueType()) &&
1443 !MVT::isVector(Op.getOperand(0).getValueType())) {
1444 // Only do this xform if FGETSIGN is valid or if before legalize.
1445 if (!TLO.AfterLegalize ||
1446 isOperationLegal(ISD::FGETSIGN, Op.getValueType())) {
1447 // Make a FGETSIGN + SHL to move the sign bit into the appropriate
1448 // place. We expect the SHL to be eliminated by other optimizations.
1449 SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, Op.getValueType(),
1451 unsigned ShVal = Op.getValueType().getSizeInBits()-1;
1452 SDValue ShAmt = TLO.DAG.getConstant(ShVal, getShiftAmountTy());
1453 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, Op.getValueType(),
1462 // Add, Sub, and Mul don't demand any bits in positions beyond that
1463 // of the highest bit demanded of them.
1464 APInt LoMask = APInt::getLowBitsSet(BitWidth,
1465 BitWidth - NewMask.countLeadingZeros());
1466 if (SimplifyDemandedBits(Op.getOperand(0), LoMask, KnownZero2,
1467 KnownOne2, TLO, Depth+1))
1469 if (SimplifyDemandedBits(Op.getOperand(1), LoMask, KnownZero2,
1470 KnownOne2, TLO, Depth+1))
1472 // See if the operation should be performed at a smaller bit width.
1473 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1478 // Just use ComputeMaskedBits to compute output bits.
1479 TLO.DAG.ComputeMaskedBits(Op, NewMask, KnownZero, KnownOne, Depth);
1483 // If we know the value of all of the demanded bits, return this as a
1485 if ((NewMask & (KnownZero|KnownOne)) == NewMask)
1486 return TLO.CombineTo(Op, TLO.DAG.getConstant(KnownOne, Op.getValueType()));
1491 /// computeMaskedBitsForTargetNode - Determine which of the bits specified
1492 /// in Mask are known to be either zero or one and return them in the
1493 /// KnownZero/KnownOne bitsets.
1494 void TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
1498 const SelectionDAG &DAG,
1499 unsigned Depth) const {
1500 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1501 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1502 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1503 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1504 "Should use MaskedValueIsZero if you don't know whether Op"
1505 " is a target node!");
1506 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
1509 /// ComputeNumSignBitsForTargetNode - This method can be implemented by
1510 /// targets that want to expose additional information about sign bits to the
1512 unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
1513 unsigned Depth) const {
1514 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1515 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1516 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1517 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1518 "Should use ComputeNumSignBits if you don't know whether Op"
1519 " is a target node!");
1523 /// ValueHasExactlyOneBitSet - Test if the given value is known to have exactly
1524 /// one bit set. This differs from ComputeMaskedBits in that it doesn't need to
1525 /// determine which bit is set.
1527 static bool ValueHasExactlyOneBitSet(SDValue Val, const SelectionDAG &DAG) {
1528 // A left-shift of a constant one will have exactly one bit set, because
1529 // shifting the bit off the end is undefined.
1530 if (Val.getOpcode() == ISD::SHL)
1531 if (ConstantSDNode *C =
1532 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0)))
1533 if (C->getAPIntValue() == 1)
1536 // Similarly, a right-shift of a constant sign-bit will have exactly
1538 if (Val.getOpcode() == ISD::SRL)
1539 if (ConstantSDNode *C =
1540 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0)))
1541 if (C->getAPIntValue().isSignBit())
1544 // More could be done here, though the above checks are enough
1545 // to handle some common cases.
1547 // Fall back to ComputeMaskedBits to catch other known cases.
1548 EVT OpVT = Val.getValueType();
1549 unsigned BitWidth = OpVT.getSizeInBits();
1550 APInt Mask = APInt::getAllOnesValue(BitWidth);
1551 APInt KnownZero, KnownOne;
1552 DAG.ComputeMaskedBits(Val, Mask, KnownZero, KnownOne);
1553 return (KnownZero.countPopulation() == BitWidth - 1) &&
1554 (KnownOne.countPopulation() == 1);
1557 /// SimplifySetCC - Try to simplify a setcc built with the specified operands
1558 /// and cc. If it is unable to simplify it, return a null SDValue.
1560 TargetLowering::SimplifySetCC(EVT VT, SDValue N0, SDValue N1,
1561 ISD::CondCode Cond, bool foldBooleans,
1562 DAGCombinerInfo &DCI, DebugLoc dl) const {
1563 SelectionDAG &DAG = DCI.DAG;
1564 LLVMContext &Context = *DAG.getContext();
1566 // These setcc operations always fold.
1570 case ISD::SETFALSE2: return DAG.getConstant(0, VT);
1572 case ISD::SETTRUE2: return DAG.getConstant(1, VT);
1575 if (isa<ConstantSDNode>(N0.getNode())) {
1576 // Ensure that the constant occurs on the RHS, and fold constant
1578 return DAG.getSetCC(dl, VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
1581 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) {
1582 const APInt &C1 = N1C->getAPIntValue();
1584 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an
1585 // equality comparison, then we're just comparing whether X itself is
1587 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) &&
1588 N0.getOperand(0).getOpcode() == ISD::CTLZ &&
1589 N0.getOperand(1).getOpcode() == ISD::Constant) {
1590 unsigned ShAmt = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
1591 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1592 ShAmt == Log2_32(N0.getValueType().getSizeInBits())) {
1593 if ((C1 == 0) == (Cond == ISD::SETEQ)) {
1594 // (srl (ctlz x), 5) == 0 -> X != 0
1595 // (srl (ctlz x), 5) != 1 -> X != 0
1598 // (srl (ctlz x), 5) != 0 -> X == 0
1599 // (srl (ctlz x), 5) == 1 -> X == 0
1602 SDValue Zero = DAG.getConstant(0, N0.getValueType());
1603 return DAG.getSetCC(dl, VT, N0.getOperand(0).getOperand(0),
1608 // If the LHS is '(and load, const)', the RHS is 0,
1609 // the test is for equality or unsigned, and all 1 bits of the const are
1610 // in the same partial word, see if we can shorten the load.
1611 if (DCI.isBeforeLegalize() &&
1612 N0.getOpcode() == ISD::AND && C1 == 0 &&
1613 N0.getNode()->hasOneUse() &&
1614 isa<LoadSDNode>(N0.getOperand(0)) &&
1615 N0.getOperand(0).getNode()->hasOneUse() &&
1616 isa<ConstantSDNode>(N0.getOperand(1))) {
1617 LoadSDNode *Lod = cast<LoadSDNode>(N0.getOperand(0));
1618 uint64_t bestMask = 0;
1619 unsigned bestWidth = 0, bestOffset = 0;
1620 if (!Lod->isVolatile() && Lod->isUnindexed() &&
1621 // FIXME: This uses getZExtValue() below so it only works on i64 and
1623 N0.getValueType().getSizeInBits() <= 64) {
1624 unsigned origWidth = N0.getValueType().getSizeInBits();
1625 // We can narrow (e.g.) 16-bit extending loads on 32-bit target to
1626 // 8 bits, but have to be careful...
1627 if (Lod->getExtensionType() != ISD::NON_EXTLOAD)
1628 origWidth = Lod->getMemoryVT().getSizeInBits();
1629 uint64_t Mask =cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
1630 for (unsigned width = origWidth / 2; width>=8; width /= 2) {
1631 uint64_t newMask = (1ULL << width) - 1;
1632 for (unsigned offset=0; offset<origWidth/width; offset++) {
1633 if ((newMask & Mask) == Mask) {
1634 if (!TD->isLittleEndian())
1635 bestOffset = (origWidth/width - offset - 1) * (width/8);
1637 bestOffset = (uint64_t)offset * (width/8);
1638 bestMask = Mask >> (offset * (width/8) * 8);
1642 newMask = newMask << width;
1647 EVT newVT = EVT::getIntegerVT(Context, bestWidth);
1648 if (newVT.isRound()) {
1649 EVT PtrType = Lod->getOperand(1).getValueType();
1650 SDValue Ptr = Lod->getBasePtr();
1651 if (bestOffset != 0)
1652 Ptr = DAG.getNode(ISD::ADD, dl, PtrType, Lod->getBasePtr(),
1653 DAG.getConstant(bestOffset, PtrType));
1654 unsigned NewAlign = MinAlign(Lod->getAlignment(), bestOffset);
1655 SDValue NewLoad = DAG.getLoad(newVT, dl, Lod->getChain(), Ptr,
1657 Lod->getSrcValueOffset() + bestOffset,
1659 return DAG.getSetCC(dl, VT,
1660 DAG.getNode(ISD::AND, dl, newVT, NewLoad,
1661 DAG.getConstant(bestMask, newVT)),
1662 DAG.getConstant(0LL, newVT), Cond);
1667 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
1668 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
1669 unsigned InSize = N0.getOperand(0).getValueType().getSizeInBits();
1671 // If the comparison constant has bits in the upper part, the
1672 // zero-extended value could never match.
1673 if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(),
1674 C1.getBitWidth() - InSize))) {
1678 case ISD::SETEQ: return DAG.getConstant(0, VT);
1681 case ISD::SETNE: return DAG.getConstant(1, VT);
1684 // True if the sign bit of C1 is set.
1685 return DAG.getConstant(C1.isNegative(), VT);
1688 // True if the sign bit of C1 isn't set.
1689 return DAG.getConstant(C1.isNonNegative(), VT);
1695 // Otherwise, we can perform the comparison with the low bits.
1703 EVT newVT = N0.getOperand(0).getValueType();
1704 if (DCI.isBeforeLegalizeOps() ||
1705 (isOperationLegal(ISD::SETCC, newVT) &&
1706 getCondCodeAction(Cond, newVT)==Legal))
1707 return DAG.getSetCC(dl, VT, N0.getOperand(0),
1708 DAG.getConstant(APInt(C1).trunc(InSize), newVT),
1713 break; // todo, be more careful with signed comparisons
1715 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
1716 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1717 EVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
1718 unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits();
1719 EVT ExtDstTy = N0.getValueType();
1720 unsigned ExtDstTyBits = ExtDstTy.getSizeInBits();
1722 // If the extended part has any inconsistent bits, it cannot ever
1723 // compare equal. In other words, they have to be all ones or all
1726 APInt::getHighBitsSet(ExtDstTyBits, ExtDstTyBits - ExtSrcTyBits);
1727 if ((C1 & ExtBits) != 0 && (C1 & ExtBits) != ExtBits)
1728 return DAG.getConstant(Cond == ISD::SETNE, VT);
1731 EVT Op0Ty = N0.getOperand(0).getValueType();
1732 if (Op0Ty == ExtSrcTy) {
1733 ZextOp = N0.getOperand(0);
1735 APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits);
1736 ZextOp = DAG.getNode(ISD::AND, dl, Op0Ty, N0.getOperand(0),
1737 DAG.getConstant(Imm, Op0Ty));
1739 if (!DCI.isCalledByLegalizer())
1740 DCI.AddToWorklist(ZextOp.getNode());
1741 // Otherwise, make this a use of a zext.
1742 return DAG.getSetCC(dl, VT, ZextOp,
1743 DAG.getConstant(C1 & APInt::getLowBitsSet(
1748 } else if ((N1C->isNullValue() || N1C->getAPIntValue() == 1) &&
1749 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1751 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC
1752 if (N0.getOpcode() == ISD::SETCC) {
1753 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getZExtValue() != 1);
1757 // Invert the condition.
1758 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
1759 CC = ISD::getSetCCInverse(CC,
1760 N0.getOperand(0).getValueType().isInteger());
1761 return DAG.getSetCC(dl, VT, N0.getOperand(0), N0.getOperand(1), CC);
1764 if ((N0.getOpcode() == ISD::XOR ||
1765 (N0.getOpcode() == ISD::AND &&
1766 N0.getOperand(0).getOpcode() == ISD::XOR &&
1767 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
1768 isa<ConstantSDNode>(N0.getOperand(1)) &&
1769 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue() == 1) {
1770 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We
1771 // can only do this if the top bits are known zero.
1772 unsigned BitWidth = N0.getValueSizeInBits();
1773 if (DAG.MaskedValueIsZero(N0,
1774 APInt::getHighBitsSet(BitWidth,
1776 // Okay, get the un-inverted input value.
1778 if (N0.getOpcode() == ISD::XOR)
1779 Val = N0.getOperand(0);
1781 assert(N0.getOpcode() == ISD::AND &&
1782 N0.getOperand(0).getOpcode() == ISD::XOR);
1783 // ((X^1)&1)^1 -> X & 1
1784 Val = DAG.getNode(ISD::AND, dl, N0.getValueType(),
1785 N0.getOperand(0).getOperand(0),
1788 return DAG.getSetCC(dl, VT, Val, N1,
1789 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
1794 APInt MinVal, MaxVal;
1795 unsigned OperandBitSize = N1C->getValueType(0).getSizeInBits();
1796 if (ISD::isSignedIntSetCC(Cond)) {
1797 MinVal = APInt::getSignedMinValue(OperandBitSize);
1798 MaxVal = APInt::getSignedMaxValue(OperandBitSize);
1800 MinVal = APInt::getMinValue(OperandBitSize);
1801 MaxVal = APInt::getMaxValue(OperandBitSize);
1804 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
1805 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
1806 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true
1807 // X >= C0 --> X > (C0-1)
1808 return DAG.getSetCC(dl, VT, N0,
1809 DAG.getConstant(C1-1, N1.getValueType()),
1810 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
1813 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
1814 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true
1815 // X <= C0 --> X < (C0+1)
1816 return DAG.getSetCC(dl, VT, N0,
1817 DAG.getConstant(C1+1, N1.getValueType()),
1818 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
1821 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
1822 return DAG.getConstant(0, VT); // X < MIN --> false
1823 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal)
1824 return DAG.getConstant(1, VT); // X >= MIN --> true
1825 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal)
1826 return DAG.getConstant(0, VT); // X > MAX --> false
1827 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal)
1828 return DAG.getConstant(1, VT); // X <= MAX --> true
1830 // Canonicalize setgt X, Min --> setne X, Min
1831 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
1832 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
1833 // Canonicalize setlt X, Max --> setne X, Max
1834 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
1835 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
1837 // If we have setult X, 1, turn it into seteq X, 0
1838 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
1839 return DAG.getSetCC(dl, VT, N0,
1840 DAG.getConstant(MinVal, N0.getValueType()),
1842 // If we have setugt X, Max-1, turn it into seteq X, Max
1843 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
1844 return DAG.getSetCC(dl, VT, N0,
1845 DAG.getConstant(MaxVal, N0.getValueType()),
1848 // If we have "setcc X, C0", check to see if we can shrink the immediate
1851 // SETUGT X, SINTMAX -> SETLT X, 0
1852 if (Cond == ISD::SETUGT &&
1853 C1 == APInt::getSignedMaxValue(OperandBitSize))
1854 return DAG.getSetCC(dl, VT, N0,
1855 DAG.getConstant(0, N1.getValueType()),
1858 // SETULT X, SINTMIN -> SETGT X, -1
1859 if (Cond == ISD::SETULT &&
1860 C1 == APInt::getSignedMinValue(OperandBitSize)) {
1861 SDValue ConstMinusOne =
1862 DAG.getConstant(APInt::getAllOnesValue(OperandBitSize),
1864 return DAG.getSetCC(dl, VT, N0, ConstMinusOne, ISD::SETGT);
1867 // Fold bit comparisons when we can.
1868 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1869 VT == N0.getValueType() && N0.getOpcode() == ISD::AND)
1870 if (ConstantSDNode *AndRHS =
1871 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1872 EVT ShiftTy = DCI.isBeforeLegalize() ?
1873 getPointerTy() : getShiftAmountTy();
1874 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
1875 // Perform the xform if the AND RHS is a single bit.
1876 if (isPowerOf2_64(AndRHS->getZExtValue())) {
1877 return DAG.getNode(ISD::SRL, dl, VT, N0,
1878 DAG.getConstant(Log2_64(AndRHS->getZExtValue()),
1881 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getZExtValue()) {
1882 // (X & 8) == 8 --> (X & 8) >> 3
1883 // Perform the xform if C1 is a single bit.
1884 if (C1.isPowerOf2()) {
1885 return DAG.getNode(ISD::SRL, dl, VT, N0,
1886 DAG.getConstant(C1.logBase2(), ShiftTy));
1892 if (isa<ConstantFPSDNode>(N0.getNode())) {
1893 // Constant fold or commute setcc.
1894 SDValue O = DAG.FoldSetCC(VT, N0, N1, Cond, dl);
1895 if (O.getNode()) return O;
1896 } else if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1.getNode())) {
1897 // If the RHS of an FP comparison is a constant, simplify it away in
1899 if (CFP->getValueAPF().isNaN()) {
1900 // If an operand is known to be a nan, we can fold it.
1901 switch (ISD::getUnorderedFlavor(Cond)) {
1902 default: llvm_unreachable("Unknown flavor!");
1903 case 0: // Known false.
1904 return DAG.getConstant(0, VT);
1905 case 1: // Known true.
1906 return DAG.getConstant(1, VT);
1907 case 2: // Undefined.
1908 return DAG.getUNDEF(VT);
1912 // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the
1913 // constant if knowing that the operand is non-nan is enough. We prefer to
1914 // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to
1916 if (Cond == ISD::SETO || Cond == ISD::SETUO)
1917 return DAG.getSetCC(dl, VT, N0, N0, Cond);
1921 // We can always fold X == X for integer setcc's.
1922 if (N0.getValueType().isInteger())
1923 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
1924 unsigned UOF = ISD::getUnorderedFlavor(Cond);
1925 if (UOF == 2) // FP operators that are undefined on NaNs.
1926 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
1927 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
1928 return DAG.getConstant(UOF, VT);
1929 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
1930 // if it is not already.
1931 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
1932 if (NewCond != Cond)
1933 return DAG.getSetCC(dl, VT, N0, N1, NewCond);
1936 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1937 N0.getValueType().isInteger()) {
1938 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
1939 N0.getOpcode() == ISD::XOR) {
1940 // Simplify (X+Y) == (X+Z) --> Y == Z
1941 if (N0.getOpcode() == N1.getOpcode()) {
1942 if (N0.getOperand(0) == N1.getOperand(0))
1943 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(1), Cond);
1944 if (N0.getOperand(1) == N1.getOperand(1))
1945 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(0), Cond);
1946 if (DAG.isCommutativeBinOp(N0.getOpcode())) {
1947 // If X op Y == Y op X, try other combinations.
1948 if (N0.getOperand(0) == N1.getOperand(1))
1949 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(0),
1951 if (N0.getOperand(1) == N1.getOperand(0))
1952 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(1),
1957 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) {
1958 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1959 // Turn (X+C1) == C2 --> X == C2-C1
1960 if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse()) {
1961 return DAG.getSetCC(dl, VT, N0.getOperand(0),
1962 DAG.getConstant(RHSC->getAPIntValue()-
1963 LHSR->getAPIntValue(),
1964 N0.getValueType()), Cond);
1967 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
1968 if (N0.getOpcode() == ISD::XOR)
1969 // If we know that all of the inverted bits are zero, don't bother
1970 // performing the inversion.
1971 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue()))
1973 DAG.getSetCC(dl, VT, N0.getOperand(0),
1974 DAG.getConstant(LHSR->getAPIntValue() ^
1975 RHSC->getAPIntValue(),
1980 // Turn (C1-X) == C2 --> X == C1-C2
1981 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
1982 if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse()) {
1984 DAG.getSetCC(dl, VT, N0.getOperand(1),
1985 DAG.getConstant(SUBC->getAPIntValue() -
1986 RHSC->getAPIntValue(),
1993 // Simplify (X+Z) == X --> Z == 0
1994 if (N0.getOperand(0) == N1)
1995 return DAG.getSetCC(dl, VT, N0.getOperand(1),
1996 DAG.getConstant(0, N0.getValueType()), Cond);
1997 if (N0.getOperand(1) == N1) {
1998 if (DAG.isCommutativeBinOp(N0.getOpcode()))
1999 return DAG.getSetCC(dl, VT, N0.getOperand(0),
2000 DAG.getConstant(0, N0.getValueType()), Cond);
2001 else if (N0.getNode()->hasOneUse()) {
2002 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
2003 // (Z-X) == X --> Z == X<<1
2004 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(),
2006 DAG.getConstant(1, getShiftAmountTy()));
2007 if (!DCI.isCalledByLegalizer())
2008 DCI.AddToWorklist(SH.getNode());
2009 return DAG.getSetCC(dl, VT, N0.getOperand(0), SH, Cond);
2014 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
2015 N1.getOpcode() == ISD::XOR) {
2016 // Simplify X == (X+Z) --> Z == 0
2017 if (N1.getOperand(0) == N0) {
2018 return DAG.getSetCC(dl, VT, N1.getOperand(1),
2019 DAG.getConstant(0, N1.getValueType()), Cond);
2020 } else if (N1.getOperand(1) == N0) {
2021 if (DAG.isCommutativeBinOp(N1.getOpcode())) {
2022 return DAG.getSetCC(dl, VT, N1.getOperand(0),
2023 DAG.getConstant(0, N1.getValueType()), Cond);
2024 } else if (N1.getNode()->hasOneUse()) {
2025 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
2026 // X == (Z-X) --> X<<1 == Z
2027 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(), N0,
2028 DAG.getConstant(1, getShiftAmountTy()));
2029 if (!DCI.isCalledByLegalizer())
2030 DCI.AddToWorklist(SH.getNode());
2031 return DAG.getSetCC(dl, VT, SH, N1.getOperand(0), Cond);
2036 // Simplify x&y == y to x&y != 0 if y has exactly one bit set.
2037 // Note that where y is variable and is known to have at most
2038 // one bit set (for example, if it is z&1) we cannot do this;
2039 // the expressions are not equivalent when y==0.
2040 if (N0.getOpcode() == ISD::AND)
2041 if (N0.getOperand(0) == N1 || N0.getOperand(1) == N1) {
2042 if (ValueHasExactlyOneBitSet(N1, DAG)) {
2043 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true);
2044 SDValue Zero = DAG.getConstant(0, N1.getValueType());
2045 return DAG.getSetCC(dl, VT, N0, Zero, Cond);
2048 if (N1.getOpcode() == ISD::AND)
2049 if (N1.getOperand(0) == N0 || N1.getOperand(1) == N0) {
2050 if (ValueHasExactlyOneBitSet(N0, DAG)) {
2051 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true);
2052 SDValue Zero = DAG.getConstant(0, N0.getValueType());
2053 return DAG.getSetCC(dl, VT, N1, Zero, Cond);
2058 // Fold away ALL boolean setcc's.
2060 if (N0.getValueType() == MVT::i1 && foldBooleans) {
2062 default: llvm_unreachable("Unknown integer setcc!");
2063 case ISD::SETEQ: // X == Y -> ~(X^Y)
2064 Temp = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1);
2065 N0 = DAG.getNOT(dl, Temp, MVT::i1);
2066 if (!DCI.isCalledByLegalizer())
2067 DCI.AddToWorklist(Temp.getNode());
2069 case ISD::SETNE: // X != Y --> (X^Y)
2070 N0 = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1);
2072 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> ~X & Y
2073 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> ~X & Y
2074 Temp = DAG.getNOT(dl, N0, MVT::i1);
2075 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N1, Temp);
2076 if (!DCI.isCalledByLegalizer())
2077 DCI.AddToWorklist(Temp.getNode());
2079 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> ~Y & X
2080 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> ~Y & X
2081 Temp = DAG.getNOT(dl, N1, MVT::i1);
2082 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N0, Temp);
2083 if (!DCI.isCalledByLegalizer())
2084 DCI.AddToWorklist(Temp.getNode());
2086 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> ~X | Y
2087 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> ~X | Y
2088 Temp = DAG.getNOT(dl, N0, MVT::i1);
2089 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N1, Temp);
2090 if (!DCI.isCalledByLegalizer())
2091 DCI.AddToWorklist(Temp.getNode());
2093 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> ~Y | X
2094 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> ~Y | X
2095 Temp = DAG.getNOT(dl, N1, MVT::i1);
2096 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N0, Temp);
2099 if (VT != MVT::i1) {
2100 if (!DCI.isCalledByLegalizer())
2101 DCI.AddToWorklist(N0.getNode());
2102 // FIXME: If running after legalize, we probably can't do this.
2103 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, N0);
2108 // Could not fold it.
2112 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
2113 /// node is a GlobalAddress + offset.
2114 bool TargetLowering::isGAPlusOffset(SDNode *N, GlobalValue* &GA,
2115 int64_t &Offset) const {
2116 if (isa<GlobalAddressSDNode>(N)) {
2117 GlobalAddressSDNode *GASD = cast<GlobalAddressSDNode>(N);
2118 GA = GASD->getGlobal();
2119 Offset += GASD->getOffset();
2123 if (N->getOpcode() == ISD::ADD) {
2124 SDValue N1 = N->getOperand(0);
2125 SDValue N2 = N->getOperand(1);
2126 if (isGAPlusOffset(N1.getNode(), GA, Offset)) {
2127 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
2129 Offset += V->getSExtValue();
2132 } else if (isGAPlusOffset(N2.getNode(), GA, Offset)) {
2133 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
2135 Offset += V->getSExtValue();
2144 /// isConsecutiveLoad - Return true if LD is loading 'Bytes' bytes from a
2145 /// location that is 'Dist' units away from the location that the 'Base' load
2146 /// is loading from.
2147 bool TargetLowering::isConsecutiveLoad(LoadSDNode *LD, LoadSDNode *Base,
2148 unsigned Bytes, int Dist,
2149 const MachineFrameInfo *MFI) const {
2150 if (LD->getChain() != Base->getChain())
2152 EVT VT = LD->getValueType(0);
2153 if (VT.getSizeInBits() / 8 != Bytes)
2156 SDValue Loc = LD->getOperand(1);
2157 SDValue BaseLoc = Base->getOperand(1);
2158 if (Loc.getOpcode() == ISD::FrameIndex) {
2159 if (BaseLoc.getOpcode() != ISD::FrameIndex)
2161 int FI = cast<FrameIndexSDNode>(Loc)->getIndex();
2162 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
2163 int FS = MFI->getObjectSize(FI);
2164 int BFS = MFI->getObjectSize(BFI);
2165 if (FS != BFS || FS != (int)Bytes) return false;
2166 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Bytes);
2168 if (Loc.getOpcode() == ISD::ADD && Loc.getOperand(0) == BaseLoc) {
2169 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Loc.getOperand(1));
2170 if (V && (V->getSExtValue() == Dist*Bytes))
2174 GlobalValue *GV1 = NULL;
2175 GlobalValue *GV2 = NULL;
2176 int64_t Offset1 = 0;
2177 int64_t Offset2 = 0;
2178 bool isGA1 = isGAPlusOffset(Loc.getNode(), GV1, Offset1);
2179 bool isGA2 = isGAPlusOffset(BaseLoc.getNode(), GV2, Offset2);
2180 if (isGA1 && isGA2 && GV1 == GV2)
2181 return Offset1 == (Offset2 + Dist*Bytes);
2186 SDValue TargetLowering::
2187 PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
2188 // Default implementation: no optimization.
2192 //===----------------------------------------------------------------------===//
2193 // Inline Assembler Implementation Methods
2194 //===----------------------------------------------------------------------===//
2197 TargetLowering::ConstraintType
2198 TargetLowering::getConstraintType(const std::string &Constraint) const {
2199 // FIXME: lots more standard ones to handle.
2200 if (Constraint.size() == 1) {
2201 switch (Constraint[0]) {
2203 case 'r': return C_RegisterClass;
2205 case 'o': // offsetable
2206 case 'V': // not offsetable
2208 case 'i': // Simple Integer or Relocatable Constant
2209 case 'n': // Simple Integer
2210 case 's': // Relocatable Constant
2211 case 'X': // Allow ANY value.
2212 case 'I': // Target registers.
2224 if (Constraint.size() > 1 && Constraint[0] == '{' &&
2225 Constraint[Constraint.size()-1] == '}')
2230 /// LowerXConstraint - try to replace an X constraint, which matches anything,
2231 /// with another that has more specific requirements based on the type of the
2232 /// corresponding operand.
2233 const char *TargetLowering::LowerXConstraint(EVT ConstraintVT) const{
2234 if (ConstraintVT.isInteger())
2236 if (ConstraintVT.isFloatingPoint())
2237 return "f"; // works for many targets
2241 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
2242 /// vector. If it is invalid, don't add anything to Ops.
2243 void TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
2244 char ConstraintLetter,
2246 std::vector<SDValue> &Ops,
2247 SelectionDAG &DAG) const {
2248 switch (ConstraintLetter) {
2250 case 'X': // Allows any operand; labels (basic block) use this.
2251 if (Op.getOpcode() == ISD::BasicBlock) {
2256 case 'i': // Simple Integer or Relocatable Constant
2257 case 'n': // Simple Integer
2258 case 's': { // Relocatable Constant
2259 // These operands are interested in values of the form (GV+C), where C may
2260 // be folded in as an offset of GV, or it may be explicitly added. Also, it
2261 // is possible and fine if either GV or C are missing.
2262 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2263 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
2265 // If we have "(add GV, C)", pull out GV/C
2266 if (Op.getOpcode() == ISD::ADD) {
2267 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
2268 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
2269 if (C == 0 || GA == 0) {
2270 C = dyn_cast<ConstantSDNode>(Op.getOperand(0));
2271 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1));
2273 if (C == 0 || GA == 0)
2277 // If we find a valid operand, map to the TargetXXX version so that the
2278 // value itself doesn't get selected.
2279 if (GA) { // Either &GV or &GV+C
2280 if (ConstraintLetter != 'n') {
2281 int64_t Offs = GA->getOffset();
2282 if (C) Offs += C->getZExtValue();
2283 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(),
2284 Op.getValueType(), Offs));
2288 if (C) { // just C, no GV.
2289 // Simple constants are not allowed for 's'.
2290 if (ConstraintLetter != 's') {
2291 // gcc prints these as sign extended. Sign extend value to 64 bits
2292 // now; without this it would get ZExt'd later in
2293 // ScheduleDAGSDNodes::EmitNode, which is very generic.
2294 Ops.push_back(DAG.getTargetConstant(C->getAPIntValue().getSExtValue(),
2304 std::vector<unsigned> TargetLowering::
2305 getRegClassForInlineAsmConstraint(const std::string &Constraint,
2307 return std::vector<unsigned>();
2311 std::pair<unsigned, const TargetRegisterClass*> TargetLowering::
2312 getRegForInlineAsmConstraint(const std::string &Constraint,
2314 if (Constraint[0] != '{')
2315 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
2316 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?");
2318 // Remove the braces from around the name.
2319 std::string RegName(Constraint.begin()+1, Constraint.end()-1);
2321 // Figure out which register class contains this reg.
2322 const TargetRegisterInfo *RI = TM.getRegisterInfo();
2323 for (TargetRegisterInfo::regclass_iterator RCI = RI->regclass_begin(),
2324 E = RI->regclass_end(); RCI != E; ++RCI) {
2325 const TargetRegisterClass *RC = *RCI;
2327 // If none of the the value types for this register class are valid, we
2328 // can't use it. For example, 64-bit reg classes on 32-bit targets.
2329 bool isLegal = false;
2330 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
2332 if (isTypeLegal(*I)) {
2338 if (!isLegal) continue;
2340 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end();
2342 if (StringsEqualNoCase(RegName, RI->get(*I).AsmName))
2343 return std::make_pair(*I, RC);
2347 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
2350 //===----------------------------------------------------------------------===//
2351 // Constraint Selection.
2353 /// isMatchingInputConstraint - Return true of this is an input operand that is
2354 /// a matching constraint like "4".
2355 bool TargetLowering::AsmOperandInfo::isMatchingInputConstraint() const {
2356 assert(!ConstraintCode.empty() && "No known constraint!");
2357 return isdigit(ConstraintCode[0]);
2360 /// getMatchedOperand - If this is an input matching constraint, this method
2361 /// returns the output operand it matches.
2362 unsigned TargetLowering::AsmOperandInfo::getMatchedOperand() const {
2363 assert(!ConstraintCode.empty() && "No known constraint!");
2364 return atoi(ConstraintCode.c_str());
2368 /// getConstraintGenerality - Return an integer indicating how general CT
2370 static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) {
2372 default: llvm_unreachable("Unknown constraint type!");
2373 case TargetLowering::C_Other:
2374 case TargetLowering::C_Unknown:
2376 case TargetLowering::C_Register:
2378 case TargetLowering::C_RegisterClass:
2380 case TargetLowering::C_Memory:
2385 /// ChooseConstraint - If there are multiple different constraints that we
2386 /// could pick for this operand (e.g. "imr") try to pick the 'best' one.
2387 /// This is somewhat tricky: constraints fall into four classes:
2388 /// Other -> immediates and magic values
2389 /// Register -> one specific register
2390 /// RegisterClass -> a group of regs
2391 /// Memory -> memory
2392 /// Ideally, we would pick the most specific constraint possible: if we have
2393 /// something that fits into a register, we would pick it. The problem here
2394 /// is that if we have something that could either be in a register or in
2395 /// memory that use of the register could cause selection of *other*
2396 /// operands to fail: they might only succeed if we pick memory. Because of
2397 /// this the heuristic we use is:
2399 /// 1) If there is an 'other' constraint, and if the operand is valid for
2400 /// that constraint, use it. This makes us take advantage of 'i'
2401 /// constraints when available.
2402 /// 2) Otherwise, pick the most general constraint present. This prefers
2403 /// 'm' over 'r', for example.
2405 static void ChooseConstraint(TargetLowering::AsmOperandInfo &OpInfo,
2406 bool hasMemory, const TargetLowering &TLI,
2407 SDValue Op, SelectionDAG *DAG) {
2408 assert(OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options");
2409 unsigned BestIdx = 0;
2410 TargetLowering::ConstraintType BestType = TargetLowering::C_Unknown;
2411 int BestGenerality = -1;
2413 // Loop over the options, keeping track of the most general one.
2414 for (unsigned i = 0, e = OpInfo.Codes.size(); i != e; ++i) {
2415 TargetLowering::ConstraintType CType =
2416 TLI.getConstraintType(OpInfo.Codes[i]);
2418 // If this is an 'other' constraint, see if the operand is valid for it.
2419 // For example, on X86 we might have an 'rI' constraint. If the operand
2420 // is an integer in the range [0..31] we want to use I (saving a load
2421 // of a register), otherwise we must use 'r'.
2422 if (CType == TargetLowering::C_Other && Op.getNode()) {
2423 assert(OpInfo.Codes[i].size() == 1 &&
2424 "Unhandled multi-letter 'other' constraint");
2425 std::vector<SDValue> ResultOps;
2426 TLI.LowerAsmOperandForConstraint(Op, OpInfo.Codes[i][0], hasMemory,
2428 if (!ResultOps.empty()) {
2435 // This constraint letter is more general than the previous one, use it.
2436 int Generality = getConstraintGenerality(CType);
2437 if (Generality > BestGenerality) {
2440 BestGenerality = Generality;
2444 OpInfo.ConstraintCode = OpInfo.Codes[BestIdx];
2445 OpInfo.ConstraintType = BestType;
2448 /// ComputeConstraintToUse - Determines the constraint code and constraint
2449 /// type to use for the specific AsmOperandInfo, setting
2450 /// OpInfo.ConstraintCode and OpInfo.ConstraintType.
2451 void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo,
2454 SelectionDAG *DAG) const {
2455 assert(!OpInfo.Codes.empty() && "Must have at least one constraint");
2457 // Single-letter constraints ('r') are very common.
2458 if (OpInfo.Codes.size() == 1) {
2459 OpInfo.ConstraintCode = OpInfo.Codes[0];
2460 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
2462 ChooseConstraint(OpInfo, hasMemory, *this, Op, DAG);
2465 // 'X' matches anything.
2466 if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) {
2467 // Labels and constants are handled elsewhere ('X' is the only thing
2468 // that matches labels). For Functions, the type here is the type of
2469 // the result, which is not what we want to look at; leave them alone.
2470 Value *v = OpInfo.CallOperandVal;
2471 if (isa<BasicBlock>(v) || isa<ConstantInt>(v) || isa<Function>(v)) {
2472 OpInfo.CallOperandVal = v;
2476 // Otherwise, try to resolve it to something we know about by looking at
2477 // the actual operand type.
2478 if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) {
2479 OpInfo.ConstraintCode = Repl;
2480 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
2485 //===----------------------------------------------------------------------===//
2486 // Loop Strength Reduction hooks
2487 //===----------------------------------------------------------------------===//
2489 /// isLegalAddressingMode - Return true if the addressing mode represented
2490 /// by AM is legal for this target, for a load/store of the specified type.
2491 bool TargetLowering::isLegalAddressingMode(const AddrMode &AM,
2492 const Type *Ty) const {
2493 // The default implementation of this implements a conservative RISCy, r+r and
2496 // Allows a sign-extended 16-bit immediate field.
2497 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
2500 // No global is ever allowed as a base.
2504 // Only support r+r,
2506 case 0: // "r+i" or just "i", depending on HasBaseReg.
2509 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
2511 // Otherwise we have r+r or r+i.
2514 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
2516 // Allow 2*r as r+r.
2523 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
2524 /// return a DAG expression to select that will generate the same value by
2525 /// multiplying by a magic number. See:
2526 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
2527 SDValue TargetLowering::BuildSDIV(SDNode *N, SelectionDAG &DAG,
2528 std::vector<SDNode*>* Created) const {
2529 EVT VT = N->getValueType(0);
2530 DebugLoc dl= N->getDebugLoc();
2532 // Check to see if we can do this.
2533 // FIXME: We should be more aggressive here.
2534 if (!isTypeLegal(VT))
2537 APInt d = cast<ConstantSDNode>(N->getOperand(1))->getAPIntValue();
2538 APInt::ms magics = d.magic();
2540 // Multiply the numerator (operand 0) by the magic value
2541 // FIXME: We should support doing a MUL in a wider type
2543 if (isOperationLegalOrCustom(ISD::MULHS, VT))
2544 Q = DAG.getNode(ISD::MULHS, dl, VT, N->getOperand(0),
2545 DAG.getConstant(magics.m, VT));
2546 else if (isOperationLegalOrCustom(ISD::SMUL_LOHI, VT))
2547 Q = SDValue(DAG.getNode(ISD::SMUL_LOHI, dl, DAG.getVTList(VT, VT),
2549 DAG.getConstant(magics.m, VT)).getNode(), 1);
2551 return SDValue(); // No mulhs or equvialent
2552 // If d > 0 and m < 0, add the numerator
2553 if (d.isStrictlyPositive() && magics.m.isNegative()) {
2554 Q = DAG.getNode(ISD::ADD, dl, VT, Q, N->getOperand(0));
2556 Created->push_back(Q.getNode());
2558 // If d < 0 and m > 0, subtract the numerator.
2559 if (d.isNegative() && magics.m.isStrictlyPositive()) {
2560 Q = DAG.getNode(ISD::SUB, dl, VT, Q, N->getOperand(0));
2562 Created->push_back(Q.getNode());
2564 // Shift right algebraic if shift value is nonzero
2566 Q = DAG.getNode(ISD::SRA, dl, VT, Q,
2567 DAG.getConstant(magics.s, getShiftAmountTy()));
2569 Created->push_back(Q.getNode());
2571 // Extract the sign bit and add it to the quotient
2573 DAG.getNode(ISD::SRL, dl, VT, Q, DAG.getConstant(VT.getSizeInBits()-1,
2574 getShiftAmountTy()));
2576 Created->push_back(T.getNode());
2577 return DAG.getNode(ISD::ADD, dl, VT, Q, T);
2580 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
2581 /// return a DAG expression to select that will generate the same value by
2582 /// multiplying by a magic number. See:
2583 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
2584 SDValue TargetLowering::BuildUDIV(SDNode *N, SelectionDAG &DAG,
2585 std::vector<SDNode*>* Created) const {
2586 EVT VT = N->getValueType(0);
2587 DebugLoc dl = N->getDebugLoc();
2589 // Check to see if we can do this.
2590 // FIXME: We should be more aggressive here.
2591 if (!isTypeLegal(VT))
2594 // FIXME: We should use a narrower constant when the upper
2595 // bits are known to be zero.
2596 ConstantSDNode *N1C = cast<ConstantSDNode>(N->getOperand(1));
2597 APInt::mu magics = N1C->getAPIntValue().magicu();
2599 // Multiply the numerator (operand 0) by the magic value
2600 // FIXME: We should support doing a MUL in a wider type
2602 if (isOperationLegalOrCustom(ISD::MULHU, VT))
2603 Q = DAG.getNode(ISD::MULHU, dl, VT, N->getOperand(0),
2604 DAG.getConstant(magics.m, VT));
2605 else if (isOperationLegalOrCustom(ISD::UMUL_LOHI, VT))
2606 Q = SDValue(DAG.getNode(ISD::UMUL_LOHI, dl, DAG.getVTList(VT, VT),
2608 DAG.getConstant(magics.m, VT)).getNode(), 1);
2610 return SDValue(); // No mulhu or equvialent
2612 Created->push_back(Q.getNode());
2614 if (magics.a == 0) {
2615 assert(magics.s < N1C->getAPIntValue().getBitWidth() &&
2616 "We shouldn't generate an undefined shift!");
2617 return DAG.getNode(ISD::SRL, dl, VT, Q,
2618 DAG.getConstant(magics.s, getShiftAmountTy()));
2620 SDValue NPQ = DAG.getNode(ISD::SUB, dl, VT, N->getOperand(0), Q);
2622 Created->push_back(NPQ.getNode());
2623 NPQ = DAG.getNode(ISD::SRL, dl, VT, NPQ,
2624 DAG.getConstant(1, getShiftAmountTy()));
2626 Created->push_back(NPQ.getNode());
2627 NPQ = DAG.getNode(ISD::ADD, dl, VT, NPQ, Q);
2629 Created->push_back(NPQ.getNode());
2630 return DAG.getNode(ISD::SRL, dl, VT, NPQ,
2631 DAG.getConstant(magics.s-1, getShiftAmountTy()));