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 = MVT::getIntegerVT(8*TD->getPointerSize());
500 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE*sizeof(TargetRegisterClass*));
501 memset(TargetDAGCombineArray, 0, array_lengthof(TargetDAGCombineArray));
502 maxStoresPerMemset = maxStoresPerMemcpy = maxStoresPerMemmove = 8;
503 allowUnalignedMemoryAccesses = false;
504 benefitFromCodePlacementOpt = false;
505 UseUnderscoreSetJmp = false;
506 UseUnderscoreLongJmp = false;
507 SelectIsExpensive = false;
508 IntDivIsCheap = false;
509 Pow2DivIsCheap = false;
510 StackPointerRegisterToSaveRestore = 0;
511 ExceptionPointerRegister = 0;
512 ExceptionSelectorRegister = 0;
513 BooleanContents = UndefinedBooleanContent;
514 SchedPreferenceInfo = SchedulingForLatency;
516 JumpBufAlignment = 0;
517 IfCvtBlockSizeLimit = 2;
518 IfCvtDupBlockSizeLimit = 0;
519 PrefLoopAlignment = 0;
521 InitLibcallNames(LibcallRoutineNames);
522 InitCmpLibcallCCs(CmpLibcallCCs);
524 // Tell Legalize whether the assembler supports DEBUG_LOC.
525 const TargetAsmInfo *TASM = TM.getTargetAsmInfo();
526 if (!TASM || !TASM->hasDotLocAndDotFile())
527 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
530 TargetLowering::~TargetLowering() {
534 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT,
535 unsigned &NumIntermediates,
537 TargetLowering* TLI) {
538 // Figure out the right, legal destination reg to copy into.
539 unsigned NumElts = VT.getVectorNumElements();
540 MVT EltTy = VT.getVectorElementType();
542 unsigned NumVectorRegs = 1;
544 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
545 // could break down into LHS/RHS like LegalizeDAG does.
546 if (!isPowerOf2_32(NumElts)) {
547 NumVectorRegs = NumElts;
551 // Divide the input until we get to a supported size. This will always
552 // end with a scalar if the target doesn't support vectors.
553 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) {
558 NumIntermediates = NumVectorRegs;
560 MVT NewVT = MVT::getVectorVT(EltTy, NumElts);
561 if (!TLI->isTypeLegal(NewVT))
563 IntermediateVT = NewVT;
565 EVT DestVT = TLI->getRegisterType(NewVT);
567 if (EVT(DestVT).bitsLT(NewVT)) {
568 // Value is expanded, e.g. i64 -> i16.
569 return NumVectorRegs*(NewVT.getSizeInBits()/DestVT.getSizeInBits());
571 // Otherwise, promotion or legal types use the same number of registers as
572 // the vector decimated to the appropriate level.
573 return NumVectorRegs;
579 /// computeRegisterProperties - Once all of the register classes are added,
580 /// this allows us to compute derived properties we expose.
581 void TargetLowering::computeRegisterProperties() {
582 assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE &&
583 "Too many value types for ValueTypeActions to hold!");
585 // Everything defaults to needing one register.
586 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
587 NumRegistersForVT[i] = 1;
588 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i;
590 // ...except isVoid, which doesn't need any registers.
591 NumRegistersForVT[MVT::isVoid] = 0;
593 // Find the largest integer register class.
594 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE;
595 for (; RegClassForVT[LargestIntReg] == 0; --LargestIntReg)
596 assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
598 // Every integer value type larger than this largest register takes twice as
599 // many registers to represent as the previous ValueType.
600 for (unsigned ExpandedReg = LargestIntReg + 1; ; ++ExpandedReg) {
601 EVT EVT = (MVT::SimpleValueType)ExpandedReg;
602 if (!EVT.isInteger())
604 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
605 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg;
606 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1);
607 ValueTypeActions.setTypeAction(EVT, Expand);
610 // Inspect all of the ValueType's smaller than the largest integer
611 // register to see which ones need promotion.
612 unsigned LegalIntReg = LargestIntReg;
613 for (unsigned IntReg = LargestIntReg - 1;
614 IntReg >= (unsigned)MVT::i1; --IntReg) {
615 EVT IVT = (MVT::SimpleValueType)IntReg;
616 if (isTypeLegal(IVT)) {
617 LegalIntReg = IntReg;
619 RegisterTypeForVT[IntReg] = TransformToType[IntReg] =
620 (MVT::SimpleValueType)LegalIntReg;
621 ValueTypeActions.setTypeAction(IVT, Promote);
625 // ppcf128 type is really two f64's.
626 if (!isTypeLegal(MVT::ppcf128)) {
627 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64];
628 RegisterTypeForVT[MVT::ppcf128] = MVT::f64;
629 TransformToType[MVT::ppcf128] = MVT::f64;
630 ValueTypeActions.setTypeAction(MVT::ppcf128, Expand);
633 // Decide how to handle f64. If the target does not have native f64 support,
634 // expand it to i64 and we will be generating soft float library calls.
635 if (!isTypeLegal(MVT::f64)) {
636 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
637 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
638 TransformToType[MVT::f64] = MVT::i64;
639 ValueTypeActions.setTypeAction(MVT::f64, Expand);
642 // Decide how to handle f32. If the target does not have native support for
643 // f32, promote it to f64 if it is legal. Otherwise, expand it to i32.
644 if (!isTypeLegal(MVT::f32)) {
645 if (isTypeLegal(MVT::f64)) {
646 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::f64];
647 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::f64];
648 TransformToType[MVT::f32] = MVT::f64;
649 ValueTypeActions.setTypeAction(MVT::f32, Promote);
651 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
652 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
653 TransformToType[MVT::f32] = MVT::i32;
654 ValueTypeActions.setTypeAction(MVT::f32, Expand);
658 // Loop over all of the vector value types to see which need transformations.
659 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
660 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
661 MVT VT = (MVT::SimpleValueType)i;
662 if (!isTypeLegal(VT)) {
665 unsigned NumIntermediates;
666 NumRegistersForVT[i] =
667 getVectorTypeBreakdownMVT(VT, IntermediateVT, NumIntermediates,
669 RegisterTypeForVT[i] = RegisterVT;
671 // Determine if there is a legal wider type.
672 bool IsLegalWiderType = false;
673 EVT EltVT = VT.getVectorElementType();
674 unsigned NElts = VT.getVectorNumElements();
675 for (unsigned nVT = i+1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
676 EVT SVT = (MVT::SimpleValueType)nVT;
677 if (isTypeLegal(SVT) && SVT.getVectorElementType() == EltVT &&
678 SVT.getVectorNumElements() > NElts) {
679 TransformToType[i] = SVT;
680 ValueTypeActions.setTypeAction(VT, Promote);
681 IsLegalWiderType = true;
685 if (!IsLegalWiderType) {
686 EVT NVT = VT.getPow2VectorType();
688 // Type is already a power of 2. The default action is to split.
689 TransformToType[i] = MVT::Other;
690 ValueTypeActions.setTypeAction(VT, Expand);
692 TransformToType[i] = NVT;
693 ValueTypeActions.setTypeAction(VT, Promote);
700 const char *TargetLowering::getTargetNodeName(unsigned Opcode) const {
705 MVT::SimpleValueType TargetLowering::getSetCCResultType(EVT VT) const {
706 return PointerTy.SimpleTy;
709 /// getVectorTypeBreakdown - Vector types are broken down into some number of
710 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
711 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
712 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
714 /// This method returns the number of registers needed, and the VT for each
715 /// register. It also returns the VT and quantity of the intermediate values
716 /// before they are promoted/expanded.
718 unsigned TargetLowering::getVectorTypeBreakdown(LLVMContext &Context, EVT VT,
720 unsigned &NumIntermediates,
721 EVT &RegisterVT) const {
722 // Figure out the right, legal destination reg to copy into.
723 unsigned NumElts = VT.getVectorNumElements();
724 EVT EltTy = VT.getVectorElementType();
726 unsigned NumVectorRegs = 1;
728 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
729 // could break down into LHS/RHS like LegalizeDAG does.
730 if (!isPowerOf2_32(NumElts)) {
731 NumVectorRegs = NumElts;
735 // Divide the input until we get to a supported size. This will always
736 // end with a scalar if the target doesn't support vectors.
737 while (NumElts > 1 && !isTypeLegal(
738 EVT::getVectorVT(Context, EltTy, NumElts))) {
743 NumIntermediates = NumVectorRegs;
745 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts);
746 if (!isTypeLegal(NewVT))
748 IntermediateVT = NewVT;
750 EVT DestVT = getRegisterType(Context, NewVT);
752 if (DestVT.bitsLT(NewVT)) {
753 // Value is expanded, e.g. i64 -> i16.
754 return NumVectorRegs*(NewVT.getSizeInBits()/DestVT.getSizeInBits());
756 // Otherwise, promotion or legal types use the same number of registers as
757 // the vector decimated to the appropriate level.
758 return NumVectorRegs;
764 /// getWidenVectorType: given a vector type, returns the type to widen to
765 /// (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
766 /// If there is no vector type that we want to widen to, returns MVT::Other
767 /// When and where to widen is target dependent based on the cost of
768 /// scalarizing vs using the wider vector type.
769 EVT TargetLowering::getWidenVectorType(EVT VT) const {
770 assert(VT.isVector());
774 // Default is not to widen until moved to LegalizeTypes
778 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
779 /// function arguments in the caller parameter area. This is the actual
780 /// alignment, not its logarithm.
781 unsigned TargetLowering::getByValTypeAlignment(const Type *Ty) const {
782 return TD->getCallFrameTypeAlignment(Ty);
785 SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table,
786 SelectionDAG &DAG) const {
787 if (usesGlobalOffsetTable())
788 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy());
793 TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
794 // Assume that everything is safe in static mode.
795 if (getTargetMachine().getRelocationModel() == Reloc::Static)
798 // In dynamic-no-pic mode, assume that known defined values are safe.
799 if (getTargetMachine().getRelocationModel() == Reloc::DynamicNoPIC &&
801 !GA->getGlobal()->isDeclaration() &&
802 !GA->getGlobal()->isWeakForLinker())
805 // Otherwise assume nothing is safe.
809 //===----------------------------------------------------------------------===//
810 // Optimization Methods
811 //===----------------------------------------------------------------------===//
813 /// ShrinkDemandedConstant - Check to see if the specified operand of the
814 /// specified instruction is a constant integer. If so, check to see if there
815 /// are any bits set in the constant that are not demanded. If so, shrink the
816 /// constant and return true.
817 bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDValue Op,
818 const APInt &Demanded) {
819 DebugLoc dl = Op.getDebugLoc();
821 // FIXME: ISD::SELECT, ISD::SELECT_CC
822 switch (Op.getOpcode()) {
827 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
828 if (!C) return false;
830 if (Op.getOpcode() == ISD::XOR &&
831 (C->getAPIntValue() | (~Demanded)).isAllOnesValue())
834 // if we can expand it to have all bits set, do it
835 if (C->getAPIntValue().intersects(~Demanded)) {
836 EVT VT = Op.getValueType();
837 SDValue New = DAG.getNode(Op.getOpcode(), dl, VT, Op.getOperand(0),
838 DAG.getConstant(Demanded &
841 return CombineTo(Op, New);
851 /// ShrinkDemandedOp - Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the
852 /// casts are free. This uses isZExtFree and ZERO_EXTEND for the widening
853 /// cast, but it could be generalized for targets with other types of
854 /// implicit widening casts.
856 TargetLowering::TargetLoweringOpt::ShrinkDemandedOp(SDValue Op,
858 const APInt &Demanded,
860 assert(Op.getNumOperands() == 2 &&
861 "ShrinkDemandedOp only supports binary operators!");
862 assert(Op.getNode()->getNumValues() == 1 &&
863 "ShrinkDemandedOp only supports nodes with one result!");
865 // Don't do this if the node has another user, which may require the
867 if (!Op.getNode()->hasOneUse())
870 // Search for the smallest integer type with free casts to and from
871 // Op's type. For expedience, just check power-of-2 integer types.
872 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
873 unsigned SmallVTBits = BitWidth - Demanded.countLeadingZeros();
874 if (!isPowerOf2_32(SmallVTBits))
875 SmallVTBits = NextPowerOf2(SmallVTBits);
876 for (; SmallVTBits < BitWidth; SmallVTBits = NextPowerOf2(SmallVTBits)) {
877 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), SmallVTBits);
878 if (TLI.isTruncateFree(Op.getValueType(), SmallVT) &&
879 TLI.isZExtFree(SmallVT, Op.getValueType())) {
880 // We found a type with free casts.
881 SDValue X = DAG.getNode(Op.getOpcode(), dl, SmallVT,
882 DAG.getNode(ISD::TRUNCATE, dl, SmallVT,
883 Op.getNode()->getOperand(0)),
884 DAG.getNode(ISD::TRUNCATE, dl, SmallVT,
885 Op.getNode()->getOperand(1)));
886 SDValue Z = DAG.getNode(ISD::ZERO_EXTEND, dl, Op.getValueType(), X);
887 return CombineTo(Op, Z);
893 /// SimplifyDemandedBits - Look at Op. At this point, we know that only the
894 /// DemandedMask bits of the result of Op are ever used downstream. If we can
895 /// use this information to simplify Op, create a new simplified DAG node and
896 /// return true, returning the original and new nodes in Old and New. Otherwise,
897 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
898 /// the expression (used to simplify the caller). The KnownZero/One bits may
899 /// only be accurate for those bits in the DemandedMask.
900 bool TargetLowering::SimplifyDemandedBits(SDValue Op,
901 const APInt &DemandedMask,
904 TargetLoweringOpt &TLO,
905 unsigned Depth) const {
906 unsigned BitWidth = DemandedMask.getBitWidth();
907 assert(Op.getValueSizeInBits() == BitWidth &&
908 "Mask size mismatches value type size!");
909 APInt NewMask = DemandedMask;
910 DebugLoc dl = Op.getDebugLoc();
912 // Don't know anything.
913 KnownZero = KnownOne = APInt(BitWidth, 0);
915 // Other users may use these bits.
916 if (!Op.getNode()->hasOneUse()) {
918 // If not at the root, Just compute the KnownZero/KnownOne bits to
919 // simplify things downstream.
920 TLO.DAG.ComputeMaskedBits(Op, DemandedMask, KnownZero, KnownOne, Depth);
923 // If this is the root being simplified, allow it to have multiple uses,
924 // just set the NewMask to all bits.
925 NewMask = APInt::getAllOnesValue(BitWidth);
926 } else if (DemandedMask == 0) {
927 // Not demanding any bits from Op.
928 if (Op.getOpcode() != ISD::UNDEF)
929 return TLO.CombineTo(Op, TLO.DAG.getUNDEF(Op.getValueType()));
931 } else if (Depth == 6) { // Limit search depth.
935 APInt KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut;
936 switch (Op.getOpcode()) {
938 // We know all of the bits for a constant!
939 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue() & NewMask;
940 KnownZero = ~KnownOne & NewMask;
941 return false; // Don't fall through, will infinitely loop.
943 // If the RHS is a constant, check to see if the LHS would be zero without
944 // using the bits from the RHS. Below, we use knowledge about the RHS to
945 // simplify the LHS, here we're using information from the LHS to simplify
947 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
948 APInt LHSZero, LHSOne;
949 TLO.DAG.ComputeMaskedBits(Op.getOperand(0), NewMask,
950 LHSZero, LHSOne, Depth+1);
951 // If the LHS already has zeros where RHSC does, this and is dead.
952 if ((LHSZero & NewMask) == (~RHSC->getAPIntValue() & NewMask))
953 return TLO.CombineTo(Op, Op.getOperand(0));
954 // If any of the set bits in the RHS are known zero on the LHS, shrink
956 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & NewMask))
960 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
961 KnownOne, TLO, Depth+1))
963 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
964 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownZero & NewMask,
965 KnownZero2, KnownOne2, TLO, Depth+1))
967 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
969 // If all of the demanded bits are known one on one side, return the other.
970 // These bits cannot contribute to the result of the 'and'.
971 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
972 return TLO.CombineTo(Op, Op.getOperand(0));
973 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
974 return TLO.CombineTo(Op, Op.getOperand(1));
975 // If all of the demanded bits in the inputs are known zeros, return zero.
976 if ((NewMask & (KnownZero|KnownZero2)) == NewMask)
977 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, Op.getValueType()));
978 // If the RHS is a constant, see if we can simplify it.
979 if (TLO.ShrinkDemandedConstant(Op, ~KnownZero2 & NewMask))
981 // If the operation can be done in a smaller type, do so.
982 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
985 // Output known-1 bits are only known if set in both the LHS & RHS.
986 KnownOne &= KnownOne2;
987 // Output known-0 are known to be clear if zero in either the LHS | RHS.
988 KnownZero |= KnownZero2;
991 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
992 KnownOne, TLO, Depth+1))
994 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
995 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownOne & NewMask,
996 KnownZero2, KnownOne2, TLO, Depth+1))
998 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1000 // If all of the demanded bits are known zero on one side, return the other.
1001 // These bits cannot contribute to the result of the 'or'.
1002 if ((NewMask & ~KnownOne2 & KnownZero) == (~KnownOne2 & NewMask))
1003 return TLO.CombineTo(Op, Op.getOperand(0));
1004 if ((NewMask & ~KnownOne & KnownZero2) == (~KnownOne & NewMask))
1005 return TLO.CombineTo(Op, Op.getOperand(1));
1006 // If all of the potentially set bits on one side are known to be set on
1007 // the other side, just use the 'other' side.
1008 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
1009 return TLO.CombineTo(Op, Op.getOperand(0));
1010 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
1011 return TLO.CombineTo(Op, Op.getOperand(1));
1012 // If the RHS is a constant, see if we can simplify it.
1013 if (TLO.ShrinkDemandedConstant(Op, NewMask))
1015 // If the operation can be done in a smaller type, do so.
1016 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1019 // Output known-0 bits are only known if clear in both the LHS & RHS.
1020 KnownZero &= KnownZero2;
1021 // Output known-1 are known to be set if set in either the LHS | RHS.
1022 KnownOne |= KnownOne2;
1025 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
1026 KnownOne, TLO, Depth+1))
1028 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1029 if (SimplifyDemandedBits(Op.getOperand(0), NewMask, KnownZero2,
1030 KnownOne2, TLO, Depth+1))
1032 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1034 // If all of the demanded bits are known zero on one side, return the other.
1035 // These bits cannot contribute to the result of the 'xor'.
1036 if ((KnownZero & NewMask) == NewMask)
1037 return TLO.CombineTo(Op, Op.getOperand(0));
1038 if ((KnownZero2 & NewMask) == NewMask)
1039 return TLO.CombineTo(Op, Op.getOperand(1));
1040 // If the operation can be done in a smaller type, do so.
1041 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1044 // If all of the unknown bits are known to be zero on one side or the other
1045 // (but not both) turn this into an *inclusive* or.
1046 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1047 if ((NewMask & ~KnownZero & ~KnownZero2) == 0)
1048 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, dl, Op.getValueType(),
1052 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1053 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
1054 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1055 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
1057 // If all of the demanded bits on one side are known, and all of the set
1058 // bits on that side are also known to be set on the other side, turn this
1059 // into an AND, as we know the bits will be cleared.
1060 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1061 if ((NewMask & (KnownZero|KnownOne)) == NewMask) { // all known
1062 if ((KnownOne & KnownOne2) == KnownOne) {
1063 EVT VT = Op.getValueType();
1064 SDValue ANDC = TLO.DAG.getConstant(~KnownOne & NewMask, VT);
1065 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, dl, VT,
1066 Op.getOperand(0), ANDC));
1070 // If the RHS is a constant, see if we can simplify it.
1071 // for XOR, we prefer to force bits to 1 if they will make a -1.
1072 // if we can't force bits, try to shrink constant
1073 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1074 APInt Expanded = C->getAPIntValue() | (~NewMask);
1075 // if we can expand it to have all bits set, do it
1076 if (Expanded.isAllOnesValue()) {
1077 if (Expanded != C->getAPIntValue()) {
1078 EVT VT = Op.getValueType();
1079 SDValue New = TLO.DAG.getNode(Op.getOpcode(), dl,VT, Op.getOperand(0),
1080 TLO.DAG.getConstant(Expanded, VT));
1081 return TLO.CombineTo(Op, New);
1083 // if it already has all the bits set, nothing to change
1084 // but don't shrink either!
1085 } else if (TLO.ShrinkDemandedConstant(Op, NewMask)) {
1090 KnownZero = KnownZeroOut;
1091 KnownOne = KnownOneOut;
1094 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero,
1095 KnownOne, TLO, Depth+1))
1097 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero2,
1098 KnownOne2, TLO, Depth+1))
1100 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1101 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1103 // If the operands are constants, see if we can simplify them.
1104 if (TLO.ShrinkDemandedConstant(Op, NewMask))
1107 // Only known if known in both the LHS and RHS.
1108 KnownOne &= KnownOne2;
1109 KnownZero &= KnownZero2;
1111 case ISD::SELECT_CC:
1112 if (SimplifyDemandedBits(Op.getOperand(3), NewMask, KnownZero,
1113 KnownOne, TLO, Depth+1))
1115 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero2,
1116 KnownOne2, TLO, Depth+1))
1118 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1119 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1121 // If the operands are constants, see if we can simplify them.
1122 if (TLO.ShrinkDemandedConstant(Op, NewMask))
1125 // Only known if known in both the LHS and RHS.
1126 KnownOne &= KnownOne2;
1127 KnownZero &= KnownZero2;
1130 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1131 unsigned ShAmt = SA->getZExtValue();
1132 SDValue InOp = Op.getOperand(0);
1134 // If the shift count is an invalid immediate, don't do anything.
1135 if (ShAmt >= BitWidth)
1138 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a
1139 // single shift. We can do this if the bottom bits (which are shifted
1140 // out) are never demanded.
1141 if (InOp.getOpcode() == ISD::SRL &&
1142 isa<ConstantSDNode>(InOp.getOperand(1))) {
1143 if (ShAmt && (NewMask & APInt::getLowBitsSet(BitWidth, ShAmt)) == 0) {
1144 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
1145 unsigned Opc = ISD::SHL;
1146 int Diff = ShAmt-C1;
1153 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
1154 EVT VT = Op.getValueType();
1155 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT,
1156 InOp.getOperand(0), NewSA));
1160 if (SimplifyDemandedBits(Op.getOperand(0), NewMask.lshr(ShAmt),
1161 KnownZero, KnownOne, TLO, Depth+1))
1163 KnownZero <<= SA->getZExtValue();
1164 KnownOne <<= SA->getZExtValue();
1165 // low bits known zero.
1166 KnownZero |= APInt::getLowBitsSet(BitWidth, SA->getZExtValue());
1170 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1171 EVT VT = Op.getValueType();
1172 unsigned ShAmt = SA->getZExtValue();
1173 unsigned VTSize = VT.getSizeInBits();
1174 SDValue InOp = Op.getOperand(0);
1176 // If the shift count is an invalid immediate, don't do anything.
1177 if (ShAmt >= BitWidth)
1180 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a
1181 // single shift. We can do this if the top bits (which are shifted out)
1182 // are never demanded.
1183 if (InOp.getOpcode() == ISD::SHL &&
1184 isa<ConstantSDNode>(InOp.getOperand(1))) {
1185 if (ShAmt && (NewMask & APInt::getHighBitsSet(VTSize, ShAmt)) == 0) {
1186 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
1187 unsigned Opc = ISD::SRL;
1188 int Diff = ShAmt-C1;
1195 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
1196 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT,
1197 InOp.getOperand(0), NewSA));
1201 // Compute the new bits that are at the top now.
1202 if (SimplifyDemandedBits(InOp, (NewMask << ShAmt),
1203 KnownZero, KnownOne, TLO, Depth+1))
1205 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1206 KnownZero = KnownZero.lshr(ShAmt);
1207 KnownOne = KnownOne.lshr(ShAmt);
1209 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
1210 KnownZero |= HighBits; // High bits known zero.
1214 // If this is an arithmetic shift right and only the low-bit is set, we can
1215 // always convert this into a logical shr, even if the shift amount is
1216 // variable. The low bit of the shift cannot be an input sign bit unless
1217 // the shift amount is >= the size of the datatype, which is undefined.
1218 if (DemandedMask == 1)
1219 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, Op.getValueType(),
1220 Op.getOperand(0), Op.getOperand(1)));
1222 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1223 EVT VT = Op.getValueType();
1224 unsigned ShAmt = SA->getZExtValue();
1226 // If the shift count is an invalid immediate, don't do anything.
1227 if (ShAmt >= BitWidth)
1230 APInt InDemandedMask = (NewMask << ShAmt);
1232 // If any of the demanded bits are produced by the sign extension, we also
1233 // demand the input sign bit.
1234 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
1235 if (HighBits.intersects(NewMask))
1236 InDemandedMask |= APInt::getSignBit(VT.getSizeInBits());
1238 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask,
1239 KnownZero, KnownOne, TLO, Depth+1))
1241 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1242 KnownZero = KnownZero.lshr(ShAmt);
1243 KnownOne = KnownOne.lshr(ShAmt);
1245 // Handle the sign bit, adjusted to where it is now in the mask.
1246 APInt SignBit = APInt::getSignBit(BitWidth).lshr(ShAmt);
1248 // If the input sign bit is known to be zero, or if none of the top bits
1249 // are demanded, turn this into an unsigned shift right.
1250 if (KnownZero.intersects(SignBit) || (HighBits & ~NewMask) == HighBits) {
1251 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, VT,
1254 } else if (KnownOne.intersects(SignBit)) { // New bits are known one.
1255 KnownOne |= HighBits;
1259 case ISD::SIGN_EXTEND_INREG: {
1260 EVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1262 // Sign extension. Compute the demanded bits in the result that are not
1263 // present in the input.
1264 APInt NewBits = APInt::getHighBitsSet(BitWidth,
1265 BitWidth - EVT.getSizeInBits()) &
1268 // If none of the extended bits are demanded, eliminate the sextinreg.
1270 return TLO.CombineTo(Op, Op.getOperand(0));
1272 APInt InSignBit = APInt::getSignBit(EVT.getSizeInBits());
1273 InSignBit.zext(BitWidth);
1274 APInt InputDemandedBits = APInt::getLowBitsSet(BitWidth,
1275 EVT.getSizeInBits()) &
1278 // Since the sign extended bits are demanded, we know that the sign
1280 InputDemandedBits |= InSignBit;
1282 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits,
1283 KnownZero, KnownOne, TLO, Depth+1))
1285 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1287 // If the sign bit of the input is known set or clear, then we know the
1288 // top bits of the result.
1290 // If the input sign bit is known zero, convert this into a zero extension.
1291 if (KnownZero.intersects(InSignBit))
1292 return TLO.CombineTo(Op,
1293 TLO.DAG.getZeroExtendInReg(Op.getOperand(0),dl,EVT));
1295 if (KnownOne.intersects(InSignBit)) { // Input sign bit known set
1296 KnownOne |= NewBits;
1297 KnownZero &= ~NewBits;
1298 } else { // Input sign bit unknown
1299 KnownZero &= ~NewBits;
1300 KnownOne &= ~NewBits;
1304 case ISD::ZERO_EXTEND: {
1305 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits();
1306 APInt InMask = NewMask;
1307 InMask.trunc(OperandBitWidth);
1309 // If none of the top bits are demanded, convert this into an any_extend.
1311 APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask;
1312 if (!NewBits.intersects(NewMask))
1313 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND, dl,
1317 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1318 KnownZero, KnownOne, TLO, Depth+1))
1320 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1321 KnownZero.zext(BitWidth);
1322 KnownOne.zext(BitWidth);
1323 KnownZero |= NewBits;
1326 case ISD::SIGN_EXTEND: {
1327 EVT InVT = Op.getOperand(0).getValueType();
1328 unsigned InBits = InVT.getSizeInBits();
1329 APInt InMask = APInt::getLowBitsSet(BitWidth, InBits);
1330 APInt InSignBit = APInt::getBitsSet(BitWidth, InBits - 1, InBits);
1331 APInt NewBits = ~InMask & NewMask;
1333 // If none of the top bits are demanded, convert this into an any_extend.
1335 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND, dl,
1339 // Since some of the sign extended bits are demanded, we know that the sign
1341 APInt InDemandedBits = InMask & NewMask;
1342 InDemandedBits |= InSignBit;
1343 InDemandedBits.trunc(InBits);
1345 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero,
1346 KnownOne, TLO, Depth+1))
1348 KnownZero.zext(BitWidth);
1349 KnownOne.zext(BitWidth);
1351 // If the sign bit is known zero, convert this to a zero extend.
1352 if (KnownZero.intersects(InSignBit))
1353 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND, dl,
1357 // If the sign bit is known one, the top bits match.
1358 if (KnownOne.intersects(InSignBit)) {
1359 KnownOne |= NewBits;
1360 KnownZero &= ~NewBits;
1361 } else { // Otherwise, top bits aren't known.
1362 KnownOne &= ~NewBits;
1363 KnownZero &= ~NewBits;
1367 case ISD::ANY_EXTEND: {
1368 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits();
1369 APInt InMask = NewMask;
1370 InMask.trunc(OperandBitWidth);
1371 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1372 KnownZero, KnownOne, TLO, Depth+1))
1374 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1375 KnownZero.zext(BitWidth);
1376 KnownOne.zext(BitWidth);
1379 case ISD::TRUNCATE: {
1380 // Simplify the input, using demanded bit information, and compute the known
1381 // zero/one bits live out.
1382 APInt TruncMask = NewMask;
1383 TruncMask.zext(Op.getOperand(0).getValueSizeInBits());
1384 if (SimplifyDemandedBits(Op.getOperand(0), TruncMask,
1385 KnownZero, KnownOne, TLO, Depth+1))
1387 KnownZero.trunc(BitWidth);
1388 KnownOne.trunc(BitWidth);
1390 // If the input is only used by this truncate, see if we can shrink it based
1391 // on the known demanded bits.
1392 if (Op.getOperand(0).getNode()->hasOneUse()) {
1393 SDValue In = Op.getOperand(0);
1394 unsigned InBitWidth = In.getValueSizeInBits();
1395 switch (In.getOpcode()) {
1398 // Shrink SRL by a constant if none of the high bits shifted in are
1400 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1))){
1401 APInt HighBits = APInt::getHighBitsSet(InBitWidth,
1402 InBitWidth - BitWidth);
1403 HighBits = HighBits.lshr(ShAmt->getZExtValue());
1404 HighBits.trunc(BitWidth);
1406 if (ShAmt->getZExtValue() < BitWidth && !(HighBits & NewMask)) {
1407 // None of the shifted in bits are needed. Add a truncate of the
1408 // shift input, then shift it.
1409 SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE, dl,
1412 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl,
1422 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1425 case ISD::AssertZext: {
1426 EVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1427 APInt InMask = APInt::getLowBitsSet(BitWidth,
1428 VT.getSizeInBits());
1429 if (SimplifyDemandedBits(Op.getOperand(0), InMask & NewMask,
1430 KnownZero, KnownOne, TLO, Depth+1))
1432 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1433 KnownZero |= ~InMask & NewMask;
1436 case ISD::BIT_CONVERT:
1438 // If this is an FP->Int bitcast and if the sign bit is the only thing that
1439 // is demanded, turn this into a FGETSIGN.
1440 if (NewMask == EVT::getIntegerVTSignBit(Op.getValueType()) &&
1441 MVT::isFloatingPoint(Op.getOperand(0).getValueType()) &&
1442 !MVT::isVector(Op.getOperand(0).getValueType())) {
1443 // Only do this xform if FGETSIGN is valid or if before legalize.
1444 if (!TLO.AfterLegalize ||
1445 isOperationLegal(ISD::FGETSIGN, Op.getValueType())) {
1446 // Make a FGETSIGN + SHL to move the sign bit into the appropriate
1447 // place. We expect the SHL to be eliminated by other optimizations.
1448 SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, Op.getValueType(),
1450 unsigned ShVal = Op.getValueType().getSizeInBits()-1;
1451 SDValue ShAmt = TLO.DAG.getConstant(ShVal, getShiftAmountTy());
1452 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, Op.getValueType(),
1461 // Add, Sub, and Mul don't demand any bits in positions beyond that
1462 // of the highest bit demanded of them.
1463 APInt LoMask = APInt::getLowBitsSet(BitWidth,
1464 BitWidth - NewMask.countLeadingZeros());
1465 if (SimplifyDemandedBits(Op.getOperand(0), LoMask, KnownZero2,
1466 KnownOne2, TLO, Depth+1))
1468 if (SimplifyDemandedBits(Op.getOperand(1), LoMask, KnownZero2,
1469 KnownOne2, TLO, Depth+1))
1471 // See if the operation should be performed at a smaller bit width.
1472 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1477 // Just use ComputeMaskedBits to compute output bits.
1478 TLO.DAG.ComputeMaskedBits(Op, NewMask, KnownZero, KnownOne, Depth);
1482 // If we know the value of all of the demanded bits, return this as a
1484 if ((NewMask & (KnownZero|KnownOne)) == NewMask)
1485 return TLO.CombineTo(Op, TLO.DAG.getConstant(KnownOne, Op.getValueType()));
1490 /// computeMaskedBitsForTargetNode - Determine which of the bits specified
1491 /// in Mask are known to be either zero or one and return them in the
1492 /// KnownZero/KnownOne bitsets.
1493 void TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
1497 const SelectionDAG &DAG,
1498 unsigned Depth) const {
1499 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1500 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1501 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1502 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1503 "Should use MaskedValueIsZero if you don't know whether Op"
1504 " is a target node!");
1505 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
1508 /// ComputeNumSignBitsForTargetNode - This method can be implemented by
1509 /// targets that want to expose additional information about sign bits to the
1511 unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
1512 unsigned Depth) const {
1513 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1514 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1515 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1516 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1517 "Should use ComputeNumSignBits if you don't know whether Op"
1518 " is a target node!");
1522 /// ValueHasExactlyOneBitSet - Test if the given value is known to have exactly
1523 /// one bit set. This differs from ComputeMaskedBits in that it doesn't need to
1524 /// determine which bit is set.
1526 static bool ValueHasExactlyOneBitSet(SDValue Val, const SelectionDAG &DAG) {
1527 // A left-shift of a constant one will have exactly one bit set, because
1528 // shifting the bit off the end is undefined.
1529 if (Val.getOpcode() == ISD::SHL)
1530 if (ConstantSDNode *C =
1531 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0)))
1532 if (C->getAPIntValue() == 1)
1535 // Similarly, a right-shift of a constant sign-bit will have exactly
1537 if (Val.getOpcode() == ISD::SRL)
1538 if (ConstantSDNode *C =
1539 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0)))
1540 if (C->getAPIntValue().isSignBit())
1543 // More could be done here, though the above checks are enough
1544 // to handle some common cases.
1546 // Fall back to ComputeMaskedBits to catch other known cases.
1547 EVT OpVT = Val.getValueType();
1548 unsigned BitWidth = OpVT.getSizeInBits();
1549 APInt Mask = APInt::getAllOnesValue(BitWidth);
1550 APInt KnownZero, KnownOne;
1551 DAG.ComputeMaskedBits(Val, Mask, KnownZero, KnownOne);
1552 return (KnownZero.countPopulation() == BitWidth - 1) &&
1553 (KnownOne.countPopulation() == 1);
1556 /// SimplifySetCC - Try to simplify a setcc built with the specified operands
1557 /// and cc. If it is unable to simplify it, return a null SDValue.
1559 TargetLowering::SimplifySetCC(EVT VT, SDValue N0, SDValue N1,
1560 ISD::CondCode Cond, bool foldBooleans,
1561 DAGCombinerInfo &DCI, DebugLoc dl) const {
1562 SelectionDAG &DAG = DCI.DAG;
1563 LLVMContext &Context = *DAG.getContext();
1565 // These setcc operations always fold.
1569 case ISD::SETFALSE2: return DAG.getConstant(0, VT);
1571 case ISD::SETTRUE2: return DAG.getConstant(1, VT);
1574 if (isa<ConstantSDNode>(N0.getNode())) {
1575 // Ensure that the constant occurs on the RHS, and fold constant
1577 return DAG.getSetCC(dl, VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
1580 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) {
1581 const APInt &C1 = N1C->getAPIntValue();
1583 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an
1584 // equality comparison, then we're just comparing whether X itself is
1586 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) &&
1587 N0.getOperand(0).getOpcode() == ISD::CTLZ &&
1588 N0.getOperand(1).getOpcode() == ISD::Constant) {
1589 unsigned ShAmt = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
1590 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1591 ShAmt == Log2_32(N0.getValueType().getSizeInBits())) {
1592 if ((C1 == 0) == (Cond == ISD::SETEQ)) {
1593 // (srl (ctlz x), 5) == 0 -> X != 0
1594 // (srl (ctlz x), 5) != 1 -> X != 0
1597 // (srl (ctlz x), 5) != 0 -> X == 0
1598 // (srl (ctlz x), 5) == 1 -> X == 0
1601 SDValue Zero = DAG.getConstant(0, N0.getValueType());
1602 return DAG.getSetCC(dl, VT, N0.getOperand(0).getOperand(0),
1607 // If the LHS is '(and load, const)', the RHS is 0,
1608 // the test is for equality or unsigned, and all 1 bits of the const are
1609 // in the same partial word, see if we can shorten the load.
1610 if (DCI.isBeforeLegalize() &&
1611 N0.getOpcode() == ISD::AND && C1 == 0 &&
1612 N0.getNode()->hasOneUse() &&
1613 isa<LoadSDNode>(N0.getOperand(0)) &&
1614 N0.getOperand(0).getNode()->hasOneUse() &&
1615 isa<ConstantSDNode>(N0.getOperand(1))) {
1616 LoadSDNode *Lod = cast<LoadSDNode>(N0.getOperand(0));
1617 uint64_t bestMask = 0;
1618 unsigned bestWidth = 0, bestOffset = 0;
1619 if (!Lod->isVolatile() && Lod->isUnindexed() &&
1620 // FIXME: This uses getZExtValue() below so it only works on i64 and
1622 N0.getValueType().getSizeInBits() <= 64) {
1623 unsigned origWidth = N0.getValueType().getSizeInBits();
1624 // We can narrow (e.g.) 16-bit extending loads on 32-bit target to
1625 // 8 bits, but have to be careful...
1626 if (Lod->getExtensionType() != ISD::NON_EXTLOAD)
1627 origWidth = Lod->getMemoryVT().getSizeInBits();
1628 uint64_t Mask =cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
1629 for (unsigned width = origWidth / 2; width>=8; width /= 2) {
1630 uint64_t newMask = (1ULL << width) - 1;
1631 for (unsigned offset=0; offset<origWidth/width; offset++) {
1632 if ((newMask & Mask) == Mask) {
1633 if (!TD->isLittleEndian())
1634 bestOffset = (origWidth/width - offset - 1) * (width/8);
1636 bestOffset = (uint64_t)offset * (width/8);
1637 bestMask = Mask >> (offset * (width/8) * 8);
1641 newMask = newMask << width;
1646 EVT newVT = EVT::getIntegerVT(Context, bestWidth);
1647 if (newVT.isRound()) {
1648 EVT PtrType = Lod->getOperand(1).getValueType();
1649 SDValue Ptr = Lod->getBasePtr();
1650 if (bestOffset != 0)
1651 Ptr = DAG.getNode(ISD::ADD, dl, PtrType, Lod->getBasePtr(),
1652 DAG.getConstant(bestOffset, PtrType));
1653 unsigned NewAlign = MinAlign(Lod->getAlignment(), bestOffset);
1654 SDValue NewLoad = DAG.getLoad(newVT, dl, Lod->getChain(), Ptr,
1656 Lod->getSrcValueOffset() + bestOffset,
1658 return DAG.getSetCC(dl, VT,
1659 DAG.getNode(ISD::AND, dl, newVT, NewLoad,
1660 DAG.getConstant(bestMask, newVT)),
1661 DAG.getConstant(0LL, newVT), Cond);
1666 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
1667 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
1668 unsigned InSize = N0.getOperand(0).getValueType().getSizeInBits();
1670 // If the comparison constant has bits in the upper part, the
1671 // zero-extended value could never match.
1672 if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(),
1673 C1.getBitWidth() - InSize))) {
1677 case ISD::SETEQ: return DAG.getConstant(0, VT);
1680 case ISD::SETNE: return DAG.getConstant(1, VT);
1683 // True if the sign bit of C1 is set.
1684 return DAG.getConstant(C1.isNegative(), VT);
1687 // True if the sign bit of C1 isn't set.
1688 return DAG.getConstant(C1.isNonNegative(), VT);
1694 // Otherwise, we can perform the comparison with the low bits.
1702 EVT newVT = N0.getOperand(0).getValueType();
1703 if (DCI.isBeforeLegalizeOps() ||
1704 (isOperationLegal(ISD::SETCC, newVT) &&
1705 getCondCodeAction(Cond, newVT)==Legal))
1706 return DAG.getSetCC(dl, VT, N0.getOperand(0),
1707 DAG.getConstant(APInt(C1).trunc(InSize), newVT),
1712 break; // todo, be more careful with signed comparisons
1714 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
1715 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1716 EVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
1717 unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits();
1718 EVT ExtDstTy = N0.getValueType();
1719 unsigned ExtDstTyBits = ExtDstTy.getSizeInBits();
1721 // If the extended part has any inconsistent bits, it cannot ever
1722 // compare equal. In other words, they have to be all ones or all
1725 APInt::getHighBitsSet(ExtDstTyBits, ExtDstTyBits - ExtSrcTyBits);
1726 if ((C1 & ExtBits) != 0 && (C1 & ExtBits) != ExtBits)
1727 return DAG.getConstant(Cond == ISD::SETNE, VT);
1730 EVT Op0Ty = N0.getOperand(0).getValueType();
1731 if (Op0Ty == ExtSrcTy) {
1732 ZextOp = N0.getOperand(0);
1734 APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits);
1735 ZextOp = DAG.getNode(ISD::AND, dl, Op0Ty, N0.getOperand(0),
1736 DAG.getConstant(Imm, Op0Ty));
1738 if (!DCI.isCalledByLegalizer())
1739 DCI.AddToWorklist(ZextOp.getNode());
1740 // Otherwise, make this a use of a zext.
1741 return DAG.getSetCC(dl, VT, ZextOp,
1742 DAG.getConstant(C1 & APInt::getLowBitsSet(
1747 } else if ((N1C->isNullValue() || N1C->getAPIntValue() == 1) &&
1748 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1750 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC
1751 if (N0.getOpcode() == ISD::SETCC) {
1752 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getZExtValue() != 1);
1756 // Invert the condition.
1757 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
1758 CC = ISD::getSetCCInverse(CC,
1759 N0.getOperand(0).getValueType().isInteger());
1760 return DAG.getSetCC(dl, VT, N0.getOperand(0), N0.getOperand(1), CC);
1763 if ((N0.getOpcode() == ISD::XOR ||
1764 (N0.getOpcode() == ISD::AND &&
1765 N0.getOperand(0).getOpcode() == ISD::XOR &&
1766 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
1767 isa<ConstantSDNode>(N0.getOperand(1)) &&
1768 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue() == 1) {
1769 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We
1770 // can only do this if the top bits are known zero.
1771 unsigned BitWidth = N0.getValueSizeInBits();
1772 if (DAG.MaskedValueIsZero(N0,
1773 APInt::getHighBitsSet(BitWidth,
1775 // Okay, get the un-inverted input value.
1777 if (N0.getOpcode() == ISD::XOR)
1778 Val = N0.getOperand(0);
1780 assert(N0.getOpcode() == ISD::AND &&
1781 N0.getOperand(0).getOpcode() == ISD::XOR);
1782 // ((X^1)&1)^1 -> X & 1
1783 Val = DAG.getNode(ISD::AND, dl, N0.getValueType(),
1784 N0.getOperand(0).getOperand(0),
1787 return DAG.getSetCC(dl, VT, Val, N1,
1788 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
1793 APInt MinVal, MaxVal;
1794 unsigned OperandBitSize = N1C->getValueType(0).getSizeInBits();
1795 if (ISD::isSignedIntSetCC(Cond)) {
1796 MinVal = APInt::getSignedMinValue(OperandBitSize);
1797 MaxVal = APInt::getSignedMaxValue(OperandBitSize);
1799 MinVal = APInt::getMinValue(OperandBitSize);
1800 MaxVal = APInt::getMaxValue(OperandBitSize);
1803 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
1804 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
1805 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true
1806 // X >= C0 --> X > (C0-1)
1807 return DAG.getSetCC(dl, VT, N0,
1808 DAG.getConstant(C1-1, N1.getValueType()),
1809 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
1812 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
1813 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true
1814 // X <= C0 --> X < (C0+1)
1815 return DAG.getSetCC(dl, VT, N0,
1816 DAG.getConstant(C1+1, N1.getValueType()),
1817 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
1820 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
1821 return DAG.getConstant(0, VT); // X < MIN --> false
1822 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal)
1823 return DAG.getConstant(1, VT); // X >= MIN --> true
1824 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal)
1825 return DAG.getConstant(0, VT); // X > MAX --> false
1826 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal)
1827 return DAG.getConstant(1, VT); // X <= MAX --> true
1829 // Canonicalize setgt X, Min --> setne X, Min
1830 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
1831 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
1832 // Canonicalize setlt X, Max --> setne X, Max
1833 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
1834 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
1836 // If we have setult X, 1, turn it into seteq X, 0
1837 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
1838 return DAG.getSetCC(dl, VT, N0,
1839 DAG.getConstant(MinVal, N0.getValueType()),
1841 // If we have setugt X, Max-1, turn it into seteq X, Max
1842 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
1843 return DAG.getSetCC(dl, VT, N0,
1844 DAG.getConstant(MaxVal, N0.getValueType()),
1847 // If we have "setcc X, C0", check to see if we can shrink the immediate
1850 // SETUGT X, SINTMAX -> SETLT X, 0
1851 if (Cond == ISD::SETUGT &&
1852 C1 == APInt::getSignedMaxValue(OperandBitSize))
1853 return DAG.getSetCC(dl, VT, N0,
1854 DAG.getConstant(0, N1.getValueType()),
1857 // SETULT X, SINTMIN -> SETGT X, -1
1858 if (Cond == ISD::SETULT &&
1859 C1 == APInt::getSignedMinValue(OperandBitSize)) {
1860 SDValue ConstMinusOne =
1861 DAG.getConstant(APInt::getAllOnesValue(OperandBitSize),
1863 return DAG.getSetCC(dl, VT, N0, ConstMinusOne, ISD::SETGT);
1866 // Fold bit comparisons when we can.
1867 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1868 VT == N0.getValueType() && N0.getOpcode() == ISD::AND)
1869 if (ConstantSDNode *AndRHS =
1870 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1871 EVT ShiftTy = DCI.isBeforeLegalize() ?
1872 getPointerTy() : getShiftAmountTy();
1873 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
1874 // Perform the xform if the AND RHS is a single bit.
1875 if (isPowerOf2_64(AndRHS->getZExtValue())) {
1876 return DAG.getNode(ISD::SRL, dl, VT, N0,
1877 DAG.getConstant(Log2_64(AndRHS->getZExtValue()),
1880 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getZExtValue()) {
1881 // (X & 8) == 8 --> (X & 8) >> 3
1882 // Perform the xform if C1 is a single bit.
1883 if (C1.isPowerOf2()) {
1884 return DAG.getNode(ISD::SRL, dl, VT, N0,
1885 DAG.getConstant(C1.logBase2(), ShiftTy));
1891 if (isa<ConstantFPSDNode>(N0.getNode())) {
1892 // Constant fold or commute setcc.
1893 SDValue O = DAG.FoldSetCC(VT, N0, N1, Cond, dl);
1894 if (O.getNode()) return O;
1895 } else if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1.getNode())) {
1896 // If the RHS of an FP comparison is a constant, simplify it away in
1898 if (CFP->getValueAPF().isNaN()) {
1899 // If an operand is known to be a nan, we can fold it.
1900 switch (ISD::getUnorderedFlavor(Cond)) {
1901 default: llvm_unreachable("Unknown flavor!");
1902 case 0: // Known false.
1903 return DAG.getConstant(0, VT);
1904 case 1: // Known true.
1905 return DAG.getConstant(1, VT);
1906 case 2: // Undefined.
1907 return DAG.getUNDEF(VT);
1911 // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the
1912 // constant if knowing that the operand is non-nan is enough. We prefer to
1913 // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to
1915 if (Cond == ISD::SETO || Cond == ISD::SETUO)
1916 return DAG.getSetCC(dl, VT, N0, N0, Cond);
1920 // We can always fold X == X for integer setcc's.
1921 if (N0.getValueType().isInteger())
1922 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
1923 unsigned UOF = ISD::getUnorderedFlavor(Cond);
1924 if (UOF == 2) // FP operators that are undefined on NaNs.
1925 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
1926 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
1927 return DAG.getConstant(UOF, VT);
1928 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
1929 // if it is not already.
1930 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
1931 if (NewCond != Cond)
1932 return DAG.getSetCC(dl, VT, N0, N1, NewCond);
1935 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1936 N0.getValueType().isInteger()) {
1937 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
1938 N0.getOpcode() == ISD::XOR) {
1939 // Simplify (X+Y) == (X+Z) --> Y == Z
1940 if (N0.getOpcode() == N1.getOpcode()) {
1941 if (N0.getOperand(0) == N1.getOperand(0))
1942 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(1), Cond);
1943 if (N0.getOperand(1) == N1.getOperand(1))
1944 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(0), Cond);
1945 if (DAG.isCommutativeBinOp(N0.getOpcode())) {
1946 // If X op Y == Y op X, try other combinations.
1947 if (N0.getOperand(0) == N1.getOperand(1))
1948 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(0),
1950 if (N0.getOperand(1) == N1.getOperand(0))
1951 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(1),
1956 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) {
1957 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1958 // Turn (X+C1) == C2 --> X == C2-C1
1959 if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse()) {
1960 return DAG.getSetCC(dl, VT, N0.getOperand(0),
1961 DAG.getConstant(RHSC->getAPIntValue()-
1962 LHSR->getAPIntValue(),
1963 N0.getValueType()), Cond);
1966 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
1967 if (N0.getOpcode() == ISD::XOR)
1968 // If we know that all of the inverted bits are zero, don't bother
1969 // performing the inversion.
1970 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue()))
1972 DAG.getSetCC(dl, VT, N0.getOperand(0),
1973 DAG.getConstant(LHSR->getAPIntValue() ^
1974 RHSC->getAPIntValue(),
1979 // Turn (C1-X) == C2 --> X == C1-C2
1980 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
1981 if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse()) {
1983 DAG.getSetCC(dl, VT, N0.getOperand(1),
1984 DAG.getConstant(SUBC->getAPIntValue() -
1985 RHSC->getAPIntValue(),
1992 // Simplify (X+Z) == X --> Z == 0
1993 if (N0.getOperand(0) == N1)
1994 return DAG.getSetCC(dl, VT, N0.getOperand(1),
1995 DAG.getConstant(0, N0.getValueType()), Cond);
1996 if (N0.getOperand(1) == N1) {
1997 if (DAG.isCommutativeBinOp(N0.getOpcode()))
1998 return DAG.getSetCC(dl, VT, N0.getOperand(0),
1999 DAG.getConstant(0, N0.getValueType()), Cond);
2000 else if (N0.getNode()->hasOneUse()) {
2001 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
2002 // (Z-X) == X --> Z == X<<1
2003 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(),
2005 DAG.getConstant(1, getShiftAmountTy()));
2006 if (!DCI.isCalledByLegalizer())
2007 DCI.AddToWorklist(SH.getNode());
2008 return DAG.getSetCC(dl, VT, N0.getOperand(0), SH, Cond);
2013 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
2014 N1.getOpcode() == ISD::XOR) {
2015 // Simplify X == (X+Z) --> Z == 0
2016 if (N1.getOperand(0) == N0) {
2017 return DAG.getSetCC(dl, VT, N1.getOperand(1),
2018 DAG.getConstant(0, N1.getValueType()), Cond);
2019 } else if (N1.getOperand(1) == N0) {
2020 if (DAG.isCommutativeBinOp(N1.getOpcode())) {
2021 return DAG.getSetCC(dl, VT, N1.getOperand(0),
2022 DAG.getConstant(0, N1.getValueType()), Cond);
2023 } else if (N1.getNode()->hasOneUse()) {
2024 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
2025 // X == (Z-X) --> X<<1 == Z
2026 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(), N0,
2027 DAG.getConstant(1, getShiftAmountTy()));
2028 if (!DCI.isCalledByLegalizer())
2029 DCI.AddToWorklist(SH.getNode());
2030 return DAG.getSetCC(dl, VT, SH, N1.getOperand(0), Cond);
2035 // Simplify x&y == y to x&y != 0 if y has exactly one bit set.
2036 // Note that where y is variable and is known to have at most
2037 // one bit set (for example, if it is z&1) we cannot do this;
2038 // the expressions are not equivalent when y==0.
2039 if (N0.getOpcode() == ISD::AND)
2040 if (N0.getOperand(0) == N1 || N0.getOperand(1) == N1) {
2041 if (ValueHasExactlyOneBitSet(N1, DAG)) {
2042 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true);
2043 SDValue Zero = DAG.getConstant(0, N1.getValueType());
2044 return DAG.getSetCC(dl, VT, N0, Zero, Cond);
2047 if (N1.getOpcode() == ISD::AND)
2048 if (N1.getOperand(0) == N0 || N1.getOperand(1) == N0) {
2049 if (ValueHasExactlyOneBitSet(N0, DAG)) {
2050 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true);
2051 SDValue Zero = DAG.getConstant(0, N0.getValueType());
2052 return DAG.getSetCC(dl, VT, N1, Zero, Cond);
2057 // Fold away ALL boolean setcc's.
2059 if (N0.getValueType() == MVT::i1 && foldBooleans) {
2061 default: llvm_unreachable("Unknown integer setcc!");
2062 case ISD::SETEQ: // X == Y -> ~(X^Y)
2063 Temp = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1);
2064 N0 = DAG.getNOT(dl, Temp, MVT::i1);
2065 if (!DCI.isCalledByLegalizer())
2066 DCI.AddToWorklist(Temp.getNode());
2068 case ISD::SETNE: // X != Y --> (X^Y)
2069 N0 = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1);
2071 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> ~X & Y
2072 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> ~X & Y
2073 Temp = DAG.getNOT(dl, N0, MVT::i1);
2074 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N1, Temp);
2075 if (!DCI.isCalledByLegalizer())
2076 DCI.AddToWorklist(Temp.getNode());
2078 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> ~Y & X
2079 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> ~Y & X
2080 Temp = DAG.getNOT(dl, N1, MVT::i1);
2081 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N0, Temp);
2082 if (!DCI.isCalledByLegalizer())
2083 DCI.AddToWorklist(Temp.getNode());
2085 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> ~X | Y
2086 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> ~X | Y
2087 Temp = DAG.getNOT(dl, N0, MVT::i1);
2088 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N1, Temp);
2089 if (!DCI.isCalledByLegalizer())
2090 DCI.AddToWorklist(Temp.getNode());
2092 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> ~Y | X
2093 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> ~Y | X
2094 Temp = DAG.getNOT(dl, N1, MVT::i1);
2095 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N0, Temp);
2098 if (VT != MVT::i1) {
2099 if (!DCI.isCalledByLegalizer())
2100 DCI.AddToWorklist(N0.getNode());
2101 // FIXME: If running after legalize, we probably can't do this.
2102 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, N0);
2107 // Could not fold it.
2111 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
2112 /// node is a GlobalAddress + offset.
2113 bool TargetLowering::isGAPlusOffset(SDNode *N, GlobalValue* &GA,
2114 int64_t &Offset) const {
2115 if (isa<GlobalAddressSDNode>(N)) {
2116 GlobalAddressSDNode *GASD = cast<GlobalAddressSDNode>(N);
2117 GA = GASD->getGlobal();
2118 Offset += GASD->getOffset();
2122 if (N->getOpcode() == ISD::ADD) {
2123 SDValue N1 = N->getOperand(0);
2124 SDValue N2 = N->getOperand(1);
2125 if (isGAPlusOffset(N1.getNode(), GA, Offset)) {
2126 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
2128 Offset += V->getSExtValue();
2131 } else if (isGAPlusOffset(N2.getNode(), GA, Offset)) {
2132 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
2134 Offset += V->getSExtValue();
2143 /// isConsecutiveLoad - Return true if LD is loading 'Bytes' bytes from a
2144 /// location that is 'Dist' units away from the location that the 'Base' load
2145 /// is loading from.
2146 bool TargetLowering::isConsecutiveLoad(LoadSDNode *LD, LoadSDNode *Base,
2147 unsigned Bytes, int Dist,
2148 const MachineFrameInfo *MFI) const {
2149 if (LD->getChain() != Base->getChain())
2151 EVT VT = LD->getValueType(0);
2152 if (VT.getSizeInBits() / 8 != Bytes)
2155 SDValue Loc = LD->getOperand(1);
2156 SDValue BaseLoc = Base->getOperand(1);
2157 if (Loc.getOpcode() == ISD::FrameIndex) {
2158 if (BaseLoc.getOpcode() != ISD::FrameIndex)
2160 int FI = cast<FrameIndexSDNode>(Loc)->getIndex();
2161 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
2162 int FS = MFI->getObjectSize(FI);
2163 int BFS = MFI->getObjectSize(BFI);
2164 if (FS != BFS || FS != (int)Bytes) return false;
2165 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Bytes);
2167 if (Loc.getOpcode() == ISD::ADD && Loc.getOperand(0) == BaseLoc) {
2168 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Loc.getOperand(1));
2169 if (V && (V->getSExtValue() == Dist*Bytes))
2173 GlobalValue *GV1 = NULL;
2174 GlobalValue *GV2 = NULL;
2175 int64_t Offset1 = 0;
2176 int64_t Offset2 = 0;
2177 bool isGA1 = isGAPlusOffset(Loc.getNode(), GV1, Offset1);
2178 bool isGA2 = isGAPlusOffset(BaseLoc.getNode(), GV2, Offset2);
2179 if (isGA1 && isGA2 && GV1 == GV2)
2180 return Offset1 == (Offset2 + Dist*Bytes);
2185 SDValue TargetLowering::
2186 PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
2187 // Default implementation: no optimization.
2191 //===----------------------------------------------------------------------===//
2192 // Inline Assembler Implementation Methods
2193 //===----------------------------------------------------------------------===//
2196 TargetLowering::ConstraintType
2197 TargetLowering::getConstraintType(const std::string &Constraint) const {
2198 // FIXME: lots more standard ones to handle.
2199 if (Constraint.size() == 1) {
2200 switch (Constraint[0]) {
2202 case 'r': return C_RegisterClass;
2204 case 'o': // offsetable
2205 case 'V': // not offsetable
2207 case 'i': // Simple Integer or Relocatable Constant
2208 case 'n': // Simple Integer
2209 case 's': // Relocatable Constant
2210 case 'X': // Allow ANY value.
2211 case 'I': // Target registers.
2223 if (Constraint.size() > 1 && Constraint[0] == '{' &&
2224 Constraint[Constraint.size()-1] == '}')
2229 /// LowerXConstraint - try to replace an X constraint, which matches anything,
2230 /// with another that has more specific requirements based on the type of the
2231 /// corresponding operand.
2232 const char *TargetLowering::LowerXConstraint(EVT ConstraintVT) const{
2233 if (ConstraintVT.isInteger())
2235 if (ConstraintVT.isFloatingPoint())
2236 return "f"; // works for many targets
2240 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
2241 /// vector. If it is invalid, don't add anything to Ops.
2242 void TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
2243 char ConstraintLetter,
2245 std::vector<SDValue> &Ops,
2246 SelectionDAG &DAG) const {
2247 switch (ConstraintLetter) {
2249 case 'X': // Allows any operand; labels (basic block) use this.
2250 if (Op.getOpcode() == ISD::BasicBlock) {
2255 case 'i': // Simple Integer or Relocatable Constant
2256 case 'n': // Simple Integer
2257 case 's': { // Relocatable Constant
2258 // These operands are interested in values of the form (GV+C), where C may
2259 // be folded in as an offset of GV, or it may be explicitly added. Also, it
2260 // is possible and fine if either GV or C are missing.
2261 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2262 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
2264 // If we have "(add GV, C)", pull out GV/C
2265 if (Op.getOpcode() == ISD::ADD) {
2266 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
2267 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
2268 if (C == 0 || GA == 0) {
2269 C = dyn_cast<ConstantSDNode>(Op.getOperand(0));
2270 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1));
2272 if (C == 0 || GA == 0)
2276 // If we find a valid operand, map to the TargetXXX version so that the
2277 // value itself doesn't get selected.
2278 if (GA) { // Either &GV or &GV+C
2279 if (ConstraintLetter != 'n') {
2280 int64_t Offs = GA->getOffset();
2281 if (C) Offs += C->getZExtValue();
2282 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(),
2283 Op.getValueType(), Offs));
2287 if (C) { // just C, no GV.
2288 // Simple constants are not allowed for 's'.
2289 if (ConstraintLetter != 's') {
2290 // gcc prints these as sign extended. Sign extend value to 64 bits
2291 // now; without this it would get ZExt'd later in
2292 // ScheduleDAGSDNodes::EmitNode, which is very generic.
2293 Ops.push_back(DAG.getTargetConstant(C->getAPIntValue().getSExtValue(),
2303 std::vector<unsigned> TargetLowering::
2304 getRegClassForInlineAsmConstraint(const std::string &Constraint,
2306 return std::vector<unsigned>();
2310 std::pair<unsigned, const TargetRegisterClass*> TargetLowering::
2311 getRegForInlineAsmConstraint(const std::string &Constraint,
2313 if (Constraint[0] != '{')
2314 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
2315 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?");
2317 // Remove the braces from around the name.
2318 std::string RegName(Constraint.begin()+1, Constraint.end()-1);
2320 // Figure out which register class contains this reg.
2321 const TargetRegisterInfo *RI = TM.getRegisterInfo();
2322 for (TargetRegisterInfo::regclass_iterator RCI = RI->regclass_begin(),
2323 E = RI->regclass_end(); RCI != E; ++RCI) {
2324 const TargetRegisterClass *RC = *RCI;
2326 // If none of the the value types for this register class are valid, we
2327 // can't use it. For example, 64-bit reg classes on 32-bit targets.
2328 bool isLegal = false;
2329 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
2331 if (isTypeLegal(*I)) {
2337 if (!isLegal) continue;
2339 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end();
2341 if (StringsEqualNoCase(RegName, RI->get(*I).AsmName))
2342 return std::make_pair(*I, RC);
2346 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
2349 //===----------------------------------------------------------------------===//
2350 // Constraint Selection.
2352 /// isMatchingInputConstraint - Return true of this is an input operand that is
2353 /// a matching constraint like "4".
2354 bool TargetLowering::AsmOperandInfo::isMatchingInputConstraint() const {
2355 assert(!ConstraintCode.empty() && "No known constraint!");
2356 return isdigit(ConstraintCode[0]);
2359 /// getMatchedOperand - If this is an input matching constraint, this method
2360 /// returns the output operand it matches.
2361 unsigned TargetLowering::AsmOperandInfo::getMatchedOperand() const {
2362 assert(!ConstraintCode.empty() && "No known constraint!");
2363 return atoi(ConstraintCode.c_str());
2367 /// getConstraintGenerality - Return an integer indicating how general CT
2369 static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) {
2371 default: llvm_unreachable("Unknown constraint type!");
2372 case TargetLowering::C_Other:
2373 case TargetLowering::C_Unknown:
2375 case TargetLowering::C_Register:
2377 case TargetLowering::C_RegisterClass:
2379 case TargetLowering::C_Memory:
2384 /// ChooseConstraint - If there are multiple different constraints that we
2385 /// could pick for this operand (e.g. "imr") try to pick the 'best' one.
2386 /// This is somewhat tricky: constraints fall into four classes:
2387 /// Other -> immediates and magic values
2388 /// Register -> one specific register
2389 /// RegisterClass -> a group of regs
2390 /// Memory -> memory
2391 /// Ideally, we would pick the most specific constraint possible: if we have
2392 /// something that fits into a register, we would pick it. The problem here
2393 /// is that if we have something that could either be in a register or in
2394 /// memory that use of the register could cause selection of *other*
2395 /// operands to fail: they might only succeed if we pick memory. Because of
2396 /// this the heuristic we use is:
2398 /// 1) If there is an 'other' constraint, and if the operand is valid for
2399 /// that constraint, use it. This makes us take advantage of 'i'
2400 /// constraints when available.
2401 /// 2) Otherwise, pick the most general constraint present. This prefers
2402 /// 'm' over 'r', for example.
2404 static void ChooseConstraint(TargetLowering::AsmOperandInfo &OpInfo,
2405 bool hasMemory, const TargetLowering &TLI,
2406 SDValue Op, SelectionDAG *DAG) {
2407 assert(OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options");
2408 unsigned BestIdx = 0;
2409 TargetLowering::ConstraintType BestType = TargetLowering::C_Unknown;
2410 int BestGenerality = -1;
2412 // Loop over the options, keeping track of the most general one.
2413 for (unsigned i = 0, e = OpInfo.Codes.size(); i != e; ++i) {
2414 TargetLowering::ConstraintType CType =
2415 TLI.getConstraintType(OpInfo.Codes[i]);
2417 // If this is an 'other' constraint, see if the operand is valid for it.
2418 // For example, on X86 we might have an 'rI' constraint. If the operand
2419 // is an integer in the range [0..31] we want to use I (saving a load
2420 // of a register), otherwise we must use 'r'.
2421 if (CType == TargetLowering::C_Other && Op.getNode()) {
2422 assert(OpInfo.Codes[i].size() == 1 &&
2423 "Unhandled multi-letter 'other' constraint");
2424 std::vector<SDValue> ResultOps;
2425 TLI.LowerAsmOperandForConstraint(Op, OpInfo.Codes[i][0], hasMemory,
2427 if (!ResultOps.empty()) {
2434 // This constraint letter is more general than the previous one, use it.
2435 int Generality = getConstraintGenerality(CType);
2436 if (Generality > BestGenerality) {
2439 BestGenerality = Generality;
2443 OpInfo.ConstraintCode = OpInfo.Codes[BestIdx];
2444 OpInfo.ConstraintType = BestType;
2447 /// ComputeConstraintToUse - Determines the constraint code and constraint
2448 /// type to use for the specific AsmOperandInfo, setting
2449 /// OpInfo.ConstraintCode and OpInfo.ConstraintType.
2450 void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo,
2453 SelectionDAG *DAG) const {
2454 assert(!OpInfo.Codes.empty() && "Must have at least one constraint");
2456 // Single-letter constraints ('r') are very common.
2457 if (OpInfo.Codes.size() == 1) {
2458 OpInfo.ConstraintCode = OpInfo.Codes[0];
2459 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
2461 ChooseConstraint(OpInfo, hasMemory, *this, Op, DAG);
2464 // 'X' matches anything.
2465 if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) {
2466 // Labels and constants are handled elsewhere ('X' is the only thing
2467 // that matches labels). For Functions, the type here is the type of
2468 // the result, which is not what we want to look at; leave them alone.
2469 Value *v = OpInfo.CallOperandVal;
2470 if (isa<BasicBlock>(v) || isa<ConstantInt>(v) || isa<Function>(v)) {
2471 OpInfo.CallOperandVal = v;
2475 // Otherwise, try to resolve it to something we know about by looking at
2476 // the actual operand type.
2477 if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) {
2478 OpInfo.ConstraintCode = Repl;
2479 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
2484 //===----------------------------------------------------------------------===//
2485 // Loop Strength Reduction hooks
2486 //===----------------------------------------------------------------------===//
2488 /// isLegalAddressingMode - Return true if the addressing mode represented
2489 /// by AM is legal for this target, for a load/store of the specified type.
2490 bool TargetLowering::isLegalAddressingMode(const AddrMode &AM,
2491 const Type *Ty) const {
2492 // The default implementation of this implements a conservative RISCy, r+r and
2495 // Allows a sign-extended 16-bit immediate field.
2496 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
2499 // No global is ever allowed as a base.
2503 // Only support r+r,
2505 case 0: // "r+i" or just "i", depending on HasBaseReg.
2508 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
2510 // Otherwise we have r+r or r+i.
2513 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
2515 // Allow 2*r as r+r.
2522 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
2523 /// return a DAG expression to select that will generate the same value by
2524 /// multiplying by a magic number. See:
2525 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
2526 SDValue TargetLowering::BuildSDIV(SDNode *N, SelectionDAG &DAG,
2527 std::vector<SDNode*>* Created) const {
2528 EVT VT = N->getValueType(0);
2529 DebugLoc dl= N->getDebugLoc();
2531 // Check to see if we can do this.
2532 // FIXME: We should be more aggressive here.
2533 if (!isTypeLegal(VT))
2536 APInt d = cast<ConstantSDNode>(N->getOperand(1))->getAPIntValue();
2537 APInt::ms magics = d.magic();
2539 // Multiply the numerator (operand 0) by the magic value
2540 // FIXME: We should support doing a MUL in a wider type
2542 if (isOperationLegalOrCustom(ISD::MULHS, VT))
2543 Q = DAG.getNode(ISD::MULHS, dl, VT, N->getOperand(0),
2544 DAG.getConstant(magics.m, VT));
2545 else if (isOperationLegalOrCustom(ISD::SMUL_LOHI, VT))
2546 Q = SDValue(DAG.getNode(ISD::SMUL_LOHI, dl, DAG.getVTList(VT, VT),
2548 DAG.getConstant(magics.m, VT)).getNode(), 1);
2550 return SDValue(); // No mulhs or equvialent
2551 // If d > 0 and m < 0, add the numerator
2552 if (d.isStrictlyPositive() && magics.m.isNegative()) {
2553 Q = DAG.getNode(ISD::ADD, dl, VT, Q, N->getOperand(0));
2555 Created->push_back(Q.getNode());
2557 // If d < 0 and m > 0, subtract the numerator.
2558 if (d.isNegative() && magics.m.isStrictlyPositive()) {
2559 Q = DAG.getNode(ISD::SUB, dl, VT, Q, N->getOperand(0));
2561 Created->push_back(Q.getNode());
2563 // Shift right algebraic if shift value is nonzero
2565 Q = DAG.getNode(ISD::SRA, dl, VT, Q,
2566 DAG.getConstant(magics.s, getShiftAmountTy()));
2568 Created->push_back(Q.getNode());
2570 // Extract the sign bit and add it to the quotient
2572 DAG.getNode(ISD::SRL, dl, VT, Q, DAG.getConstant(VT.getSizeInBits()-1,
2573 getShiftAmountTy()));
2575 Created->push_back(T.getNode());
2576 return DAG.getNode(ISD::ADD, dl, VT, Q, T);
2579 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
2580 /// return a DAG expression to select that will generate the same value by
2581 /// multiplying by a magic number. See:
2582 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
2583 SDValue TargetLowering::BuildUDIV(SDNode *N, SelectionDAG &DAG,
2584 std::vector<SDNode*>* Created) const {
2585 EVT VT = N->getValueType(0);
2586 DebugLoc dl = N->getDebugLoc();
2588 // Check to see if we can do this.
2589 // FIXME: We should be more aggressive here.
2590 if (!isTypeLegal(VT))
2593 // FIXME: We should use a narrower constant when the upper
2594 // bits are known to be zero.
2595 ConstantSDNode *N1C = cast<ConstantSDNode>(N->getOperand(1));
2596 APInt::mu magics = N1C->getAPIntValue().magicu();
2598 // Multiply the numerator (operand 0) by the magic value
2599 // FIXME: We should support doing a MUL in a wider type
2601 if (isOperationLegalOrCustom(ISD::MULHU, VT))
2602 Q = DAG.getNode(ISD::MULHU, dl, VT, N->getOperand(0),
2603 DAG.getConstant(magics.m, VT));
2604 else if (isOperationLegalOrCustom(ISD::UMUL_LOHI, VT))
2605 Q = SDValue(DAG.getNode(ISD::UMUL_LOHI, dl, DAG.getVTList(VT, VT),
2607 DAG.getConstant(magics.m, VT)).getNode(), 1);
2609 return SDValue(); // No mulhu or equvialent
2611 Created->push_back(Q.getNode());
2613 if (magics.a == 0) {
2614 assert(magics.s < N1C->getAPIntValue().getBitWidth() &&
2615 "We shouldn't generate an undefined shift!");
2616 return DAG.getNode(ISD::SRL, dl, VT, Q,
2617 DAG.getConstant(magics.s, getShiftAmountTy()));
2619 SDValue NPQ = DAG.getNode(ISD::SUB, dl, VT, N->getOperand(0), Q);
2621 Created->push_back(NPQ.getNode());
2622 NPQ = DAG.getNode(ISD::SRL, dl, VT, NPQ,
2623 DAG.getConstant(1, getShiftAmountTy()));
2625 Created->push_back(NPQ.getNode());
2626 NPQ = DAG.getNode(ISD::ADD, dl, VT, NPQ, Q);
2628 Created->push_back(NPQ.getNode());
2629 return DAG.getNode(ISD::SRL, dl, VT, NPQ,
2630 DAG.getConstant(magics.s-1, getShiftAmountTy()));