1 //===-- TargetLowering.cpp - Implement the TargetLowering class -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This implements the TargetLowering class.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Target/TargetAsmInfo.h"
15 #include "llvm/Target/TargetLowering.h"
16 #include "llvm/Target/TargetSubtarget.h"
17 #include "llvm/Target/TargetData.h"
18 #include "llvm/Target/TargetMachine.h"
19 #include "llvm/Target/TargetRegisterInfo.h"
20 #include "llvm/GlobalVariable.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/CodeGen/MachineFrameInfo.h"
23 #include "llvm/CodeGen/SelectionDAG.h"
24 #include "llvm/ADT/StringExtras.h"
25 #include "llvm/ADT/STLExtras.h"
26 #include "llvm/Support/MathExtras.h"
30 TLSModel::Model getTLSModel(const GlobalValue *GV, Reloc::Model reloc) {
31 bool isLocal = GV->hasLocalLinkage();
32 bool isDeclaration = GV->isDeclaration();
33 // FIXME: what should we do for protected and internal visibility?
34 // For variables, is internal different from hidden?
35 bool isHidden = GV->hasHiddenVisibility();
37 if (reloc == Reloc::PIC_) {
38 if (isLocal || isHidden)
39 return TLSModel::LocalDynamic;
41 return TLSModel::GeneralDynamic;
43 if (!isDeclaration || isHidden)
44 return TLSModel::LocalExec;
46 return TLSModel::InitialExec;
51 /// InitLibcallNames - Set default libcall names.
53 static void InitLibcallNames(const char **Names) {
54 Names[RTLIB::SHL_I16] = "__ashlhi3";
55 Names[RTLIB::SHL_I32] = "__ashlsi3";
56 Names[RTLIB::SHL_I64] = "__ashldi3";
57 Names[RTLIB::SHL_I128] = "__ashlti3";
58 Names[RTLIB::SRL_I16] = "__lshrhi3";
59 Names[RTLIB::SRL_I32] = "__lshrsi3";
60 Names[RTLIB::SRL_I64] = "__lshrdi3";
61 Names[RTLIB::SRL_I128] = "__lshrti3";
62 Names[RTLIB::SRA_I16] = "__ashrhi3";
63 Names[RTLIB::SRA_I32] = "__ashrsi3";
64 Names[RTLIB::SRA_I64] = "__ashrdi3";
65 Names[RTLIB::SRA_I128] = "__ashrti3";
66 Names[RTLIB::MUL_I16] = "__mulhi3";
67 Names[RTLIB::MUL_I32] = "__mulsi3";
68 Names[RTLIB::MUL_I64] = "__muldi3";
69 Names[RTLIB::MUL_I128] = "__multi3";
70 Names[RTLIB::SDIV_I16] = "__divhi3";
71 Names[RTLIB::SDIV_I32] = "__divsi3";
72 Names[RTLIB::SDIV_I64] = "__divdi3";
73 Names[RTLIB::SDIV_I128] = "__divti3";
74 Names[RTLIB::UDIV_I16] = "__udivhi3";
75 Names[RTLIB::UDIV_I32] = "__udivsi3";
76 Names[RTLIB::UDIV_I64] = "__udivdi3";
77 Names[RTLIB::UDIV_I128] = "__udivti3";
78 Names[RTLIB::SREM_I16] = "__modhi3";
79 Names[RTLIB::SREM_I32] = "__modsi3";
80 Names[RTLIB::SREM_I64] = "__moddi3";
81 Names[RTLIB::SREM_I128] = "__modti3";
82 Names[RTLIB::UREM_I16] = "__umodhi3";
83 Names[RTLIB::UREM_I32] = "__umodsi3";
84 Names[RTLIB::UREM_I64] = "__umoddi3";
85 Names[RTLIB::UREM_I128] = "__umodti3";
86 Names[RTLIB::NEG_I32] = "__negsi2";
87 Names[RTLIB::NEG_I64] = "__negdi2";
88 Names[RTLIB::ADD_F32] = "__addsf3";
89 Names[RTLIB::ADD_F64] = "__adddf3";
90 Names[RTLIB::ADD_F80] = "__addxf3";
91 Names[RTLIB::ADD_PPCF128] = "__gcc_qadd";
92 Names[RTLIB::SUB_F32] = "__subsf3";
93 Names[RTLIB::SUB_F64] = "__subdf3";
94 Names[RTLIB::SUB_F80] = "__subxf3";
95 Names[RTLIB::SUB_PPCF128] = "__gcc_qsub";
96 Names[RTLIB::MUL_F32] = "__mulsf3";
97 Names[RTLIB::MUL_F64] = "__muldf3";
98 Names[RTLIB::MUL_F80] = "__mulxf3";
99 Names[RTLIB::MUL_PPCF128] = "__gcc_qmul";
100 Names[RTLIB::DIV_F32] = "__divsf3";
101 Names[RTLIB::DIV_F64] = "__divdf3";
102 Names[RTLIB::DIV_F80] = "__divxf3";
103 Names[RTLIB::DIV_PPCF128] = "__gcc_qdiv";
104 Names[RTLIB::REM_F32] = "fmodf";
105 Names[RTLIB::REM_F64] = "fmod";
106 Names[RTLIB::REM_F80] = "fmodl";
107 Names[RTLIB::REM_PPCF128] = "fmodl";
108 Names[RTLIB::POWI_F32] = "__powisf2";
109 Names[RTLIB::POWI_F64] = "__powidf2";
110 Names[RTLIB::POWI_F80] = "__powixf2";
111 Names[RTLIB::POWI_PPCF128] = "__powitf2";
112 Names[RTLIB::SQRT_F32] = "sqrtf";
113 Names[RTLIB::SQRT_F64] = "sqrt";
114 Names[RTLIB::SQRT_F80] = "sqrtl";
115 Names[RTLIB::SQRT_PPCF128] = "sqrtl";
116 Names[RTLIB::LOG_F32] = "logf";
117 Names[RTLIB::LOG_F64] = "log";
118 Names[RTLIB::LOG_F80] = "logl";
119 Names[RTLIB::LOG_PPCF128] = "logl";
120 Names[RTLIB::LOG2_F32] = "log2f";
121 Names[RTLIB::LOG2_F64] = "log2";
122 Names[RTLIB::LOG2_F80] = "log2l";
123 Names[RTLIB::LOG2_PPCF128] = "log2l";
124 Names[RTLIB::LOG10_F32] = "log10f";
125 Names[RTLIB::LOG10_F64] = "log10";
126 Names[RTLIB::LOG10_F80] = "log10l";
127 Names[RTLIB::LOG10_PPCF128] = "log10l";
128 Names[RTLIB::EXP_F32] = "expf";
129 Names[RTLIB::EXP_F64] = "exp";
130 Names[RTLIB::EXP_F80] = "expl";
131 Names[RTLIB::EXP_PPCF128] = "expl";
132 Names[RTLIB::EXP2_F32] = "exp2f";
133 Names[RTLIB::EXP2_F64] = "exp2";
134 Names[RTLIB::EXP2_F80] = "exp2l";
135 Names[RTLIB::EXP2_PPCF128] = "exp2l";
136 Names[RTLIB::SIN_F32] = "sinf";
137 Names[RTLIB::SIN_F64] = "sin";
138 Names[RTLIB::SIN_F80] = "sinl";
139 Names[RTLIB::SIN_PPCF128] = "sinl";
140 Names[RTLIB::COS_F32] = "cosf";
141 Names[RTLIB::COS_F64] = "cos";
142 Names[RTLIB::COS_F80] = "cosl";
143 Names[RTLIB::COS_PPCF128] = "cosl";
144 Names[RTLIB::POW_F32] = "powf";
145 Names[RTLIB::POW_F64] = "pow";
146 Names[RTLIB::POW_F80] = "powl";
147 Names[RTLIB::POW_PPCF128] = "powl";
148 Names[RTLIB::CEIL_F32] = "ceilf";
149 Names[RTLIB::CEIL_F64] = "ceil";
150 Names[RTLIB::CEIL_F80] = "ceill";
151 Names[RTLIB::CEIL_PPCF128] = "ceill";
152 Names[RTLIB::TRUNC_F32] = "truncf";
153 Names[RTLIB::TRUNC_F64] = "trunc";
154 Names[RTLIB::TRUNC_F80] = "truncl";
155 Names[RTLIB::TRUNC_PPCF128] = "truncl";
156 Names[RTLIB::RINT_F32] = "rintf";
157 Names[RTLIB::RINT_F64] = "rint";
158 Names[RTLIB::RINT_F80] = "rintl";
159 Names[RTLIB::RINT_PPCF128] = "rintl";
160 Names[RTLIB::NEARBYINT_F32] = "nearbyintf";
161 Names[RTLIB::NEARBYINT_F64] = "nearbyint";
162 Names[RTLIB::NEARBYINT_F80] = "nearbyintl";
163 Names[RTLIB::NEARBYINT_PPCF128] = "nearbyintl";
164 Names[RTLIB::FLOOR_F32] = "floorf";
165 Names[RTLIB::FLOOR_F64] = "floor";
166 Names[RTLIB::FLOOR_F80] = "floorl";
167 Names[RTLIB::FLOOR_PPCF128] = "floorl";
168 Names[RTLIB::FPEXT_F32_F64] = "__extendsfdf2";
169 Names[RTLIB::FPROUND_F64_F32] = "__truncdfsf2";
170 Names[RTLIB::FPROUND_F80_F32] = "__truncxfsf2";
171 Names[RTLIB::FPROUND_PPCF128_F32] = "__trunctfsf2";
172 Names[RTLIB::FPROUND_F80_F64] = "__truncxfdf2";
173 Names[RTLIB::FPROUND_PPCF128_F64] = "__trunctfdf2";
174 Names[RTLIB::FPTOSINT_F32_I8] = "__fixsfi8";
175 Names[RTLIB::FPTOSINT_F32_I16] = "__fixsfi16";
176 Names[RTLIB::FPTOSINT_F32_I32] = "__fixsfsi";
177 Names[RTLIB::FPTOSINT_F32_I64] = "__fixsfdi";
178 Names[RTLIB::FPTOSINT_F32_I128] = "__fixsfti";
179 Names[RTLIB::FPTOSINT_F64_I32] = "__fixdfsi";
180 Names[RTLIB::FPTOSINT_F64_I64] = "__fixdfdi";
181 Names[RTLIB::FPTOSINT_F64_I128] = "__fixdfti";
182 Names[RTLIB::FPTOSINT_F80_I32] = "__fixxfsi";
183 Names[RTLIB::FPTOSINT_F80_I64] = "__fixxfdi";
184 Names[RTLIB::FPTOSINT_F80_I128] = "__fixxfti";
185 Names[RTLIB::FPTOSINT_PPCF128_I32] = "__fixtfsi";
186 Names[RTLIB::FPTOSINT_PPCF128_I64] = "__fixtfdi";
187 Names[RTLIB::FPTOSINT_PPCF128_I128] = "__fixtfti";
188 Names[RTLIB::FPTOUINT_F32_I8] = "__fixunssfi8";
189 Names[RTLIB::FPTOUINT_F32_I16] = "__fixunssfi16";
190 Names[RTLIB::FPTOUINT_F32_I32] = "__fixunssfsi";
191 Names[RTLIB::FPTOUINT_F32_I64] = "__fixunssfdi";
192 Names[RTLIB::FPTOUINT_F32_I128] = "__fixunssfti";
193 Names[RTLIB::FPTOUINT_F64_I32] = "__fixunsdfsi";
194 Names[RTLIB::FPTOUINT_F64_I64] = "__fixunsdfdi";
195 Names[RTLIB::FPTOUINT_F64_I128] = "__fixunsdfti";
196 Names[RTLIB::FPTOUINT_F80_I32] = "__fixunsxfsi";
197 Names[RTLIB::FPTOUINT_F80_I64] = "__fixunsxfdi";
198 Names[RTLIB::FPTOUINT_F80_I128] = "__fixunsxfti";
199 Names[RTLIB::FPTOUINT_PPCF128_I32] = "__fixunstfsi";
200 Names[RTLIB::FPTOUINT_PPCF128_I64] = "__fixunstfdi";
201 Names[RTLIB::FPTOUINT_PPCF128_I128] = "__fixunstfti";
202 Names[RTLIB::SINTTOFP_I32_F32] = "__floatsisf";
203 Names[RTLIB::SINTTOFP_I32_F64] = "__floatsidf";
204 Names[RTLIB::SINTTOFP_I32_F80] = "__floatsixf";
205 Names[RTLIB::SINTTOFP_I32_PPCF128] = "__floatsitf";
206 Names[RTLIB::SINTTOFP_I64_F32] = "__floatdisf";
207 Names[RTLIB::SINTTOFP_I64_F64] = "__floatdidf";
208 Names[RTLIB::SINTTOFP_I64_F80] = "__floatdixf";
209 Names[RTLIB::SINTTOFP_I64_PPCF128] = "__floatditf";
210 Names[RTLIB::SINTTOFP_I128_F32] = "__floattisf";
211 Names[RTLIB::SINTTOFP_I128_F64] = "__floattidf";
212 Names[RTLIB::SINTTOFP_I128_F80] = "__floattixf";
213 Names[RTLIB::SINTTOFP_I128_PPCF128] = "__floattitf";
214 Names[RTLIB::UINTTOFP_I32_F32] = "__floatunsisf";
215 Names[RTLIB::UINTTOFP_I32_F64] = "__floatunsidf";
216 Names[RTLIB::UINTTOFP_I32_F80] = "__floatunsixf";
217 Names[RTLIB::UINTTOFP_I32_PPCF128] = "__floatunsitf";
218 Names[RTLIB::UINTTOFP_I64_F32] = "__floatundisf";
219 Names[RTLIB::UINTTOFP_I64_F64] = "__floatundidf";
220 Names[RTLIB::UINTTOFP_I64_F80] = "__floatundixf";
221 Names[RTLIB::UINTTOFP_I64_PPCF128] = "__floatunditf";
222 Names[RTLIB::UINTTOFP_I128_F32] = "__floatuntisf";
223 Names[RTLIB::UINTTOFP_I128_F64] = "__floatuntidf";
224 Names[RTLIB::UINTTOFP_I128_F80] = "__floatuntixf";
225 Names[RTLIB::UINTTOFP_I128_PPCF128] = "__floatuntitf";
226 Names[RTLIB::OEQ_F32] = "__eqsf2";
227 Names[RTLIB::OEQ_F64] = "__eqdf2";
228 Names[RTLIB::UNE_F32] = "__nesf2";
229 Names[RTLIB::UNE_F64] = "__nedf2";
230 Names[RTLIB::OGE_F32] = "__gesf2";
231 Names[RTLIB::OGE_F64] = "__gedf2";
232 Names[RTLIB::OLT_F32] = "__ltsf2";
233 Names[RTLIB::OLT_F64] = "__ltdf2";
234 Names[RTLIB::OLE_F32] = "__lesf2";
235 Names[RTLIB::OLE_F64] = "__ledf2";
236 Names[RTLIB::OGT_F32] = "__gtsf2";
237 Names[RTLIB::OGT_F64] = "__gtdf2";
238 Names[RTLIB::UO_F32] = "__unordsf2";
239 Names[RTLIB::UO_F64] = "__unorddf2";
240 Names[RTLIB::O_F32] = "__unordsf2";
241 Names[RTLIB::O_F64] = "__unorddf2";
242 Names[RTLIB::UNWIND_RESUME] = "_Unwind_Resume";
245 /// getFPEXT - Return the FPEXT_*_* value for the given types, or
246 /// UNKNOWN_LIBCALL if there is none.
247 RTLIB::Libcall RTLIB::getFPEXT(MVT OpVT, MVT RetVT) {
248 if (OpVT == MVT::f32) {
249 if (RetVT == MVT::f64)
250 return FPEXT_F32_F64;
252 return UNKNOWN_LIBCALL;
255 /// getFPROUND - Return the FPROUND_*_* value for the given types, or
256 /// UNKNOWN_LIBCALL if there is none.
257 RTLIB::Libcall RTLIB::getFPROUND(MVT OpVT, MVT RetVT) {
258 if (RetVT == MVT::f32) {
259 if (OpVT == MVT::f64)
260 return FPROUND_F64_F32;
261 if (OpVT == MVT::f80)
262 return FPROUND_F80_F32;
263 if (OpVT == MVT::ppcf128)
264 return FPROUND_PPCF128_F32;
265 } else if (RetVT == MVT::f64) {
266 if (OpVT == MVT::f80)
267 return FPROUND_F80_F64;
268 if (OpVT == MVT::ppcf128)
269 return FPROUND_PPCF128_F64;
271 return UNKNOWN_LIBCALL;
274 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or
275 /// UNKNOWN_LIBCALL if there is none.
276 RTLIB::Libcall RTLIB::getFPTOSINT(MVT OpVT, MVT RetVT) {
277 if (OpVT == MVT::f32) {
278 if (RetVT == MVT::i8)
279 return FPTOSINT_F32_I8;
280 if (RetVT == MVT::i16)
281 return FPTOSINT_F32_I16;
282 if (RetVT == MVT::i32)
283 return FPTOSINT_F32_I32;
284 if (RetVT == MVT::i64)
285 return FPTOSINT_F32_I64;
286 if (RetVT == MVT::i128)
287 return FPTOSINT_F32_I128;
288 } else if (OpVT == MVT::f64) {
289 if (RetVT == MVT::i32)
290 return FPTOSINT_F64_I32;
291 if (RetVT == MVT::i64)
292 return FPTOSINT_F64_I64;
293 if (RetVT == MVT::i128)
294 return FPTOSINT_F64_I128;
295 } else if (OpVT == MVT::f80) {
296 if (RetVT == MVT::i32)
297 return FPTOSINT_F80_I32;
298 if (RetVT == MVT::i64)
299 return FPTOSINT_F80_I64;
300 if (RetVT == MVT::i128)
301 return FPTOSINT_F80_I128;
302 } else if (OpVT == MVT::ppcf128) {
303 if (RetVT == MVT::i32)
304 return FPTOSINT_PPCF128_I32;
305 if (RetVT == MVT::i64)
306 return FPTOSINT_PPCF128_I64;
307 if (RetVT == MVT::i128)
308 return FPTOSINT_PPCF128_I128;
310 return UNKNOWN_LIBCALL;
313 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or
314 /// UNKNOWN_LIBCALL if there is none.
315 RTLIB::Libcall RTLIB::getFPTOUINT(MVT OpVT, MVT RetVT) {
316 if (OpVT == MVT::f32) {
317 if (RetVT == MVT::i8)
318 return FPTOUINT_F32_I8;
319 if (RetVT == MVT::i16)
320 return FPTOUINT_F32_I16;
321 if (RetVT == MVT::i32)
322 return FPTOUINT_F32_I32;
323 if (RetVT == MVT::i64)
324 return FPTOUINT_F32_I64;
325 if (RetVT == MVT::i128)
326 return FPTOUINT_F32_I128;
327 } else if (OpVT == MVT::f64) {
328 if (RetVT == MVT::i32)
329 return FPTOUINT_F64_I32;
330 if (RetVT == MVT::i64)
331 return FPTOUINT_F64_I64;
332 if (RetVT == MVT::i128)
333 return FPTOUINT_F64_I128;
334 } else if (OpVT == MVT::f80) {
335 if (RetVT == MVT::i32)
336 return FPTOUINT_F80_I32;
337 if (RetVT == MVT::i64)
338 return FPTOUINT_F80_I64;
339 if (RetVT == MVT::i128)
340 return FPTOUINT_F80_I128;
341 } else if (OpVT == MVT::ppcf128) {
342 if (RetVT == MVT::i32)
343 return FPTOUINT_PPCF128_I32;
344 if (RetVT == MVT::i64)
345 return FPTOUINT_PPCF128_I64;
346 if (RetVT == MVT::i128)
347 return FPTOUINT_PPCF128_I128;
349 return UNKNOWN_LIBCALL;
352 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or
353 /// UNKNOWN_LIBCALL if there is none.
354 RTLIB::Libcall RTLIB::getSINTTOFP(MVT OpVT, MVT RetVT) {
355 if (OpVT == MVT::i32) {
356 if (RetVT == MVT::f32)
357 return SINTTOFP_I32_F32;
358 else if (RetVT == MVT::f64)
359 return SINTTOFP_I32_F64;
360 else if (RetVT == MVT::f80)
361 return SINTTOFP_I32_F80;
362 else if (RetVT == MVT::ppcf128)
363 return SINTTOFP_I32_PPCF128;
364 } else if (OpVT == MVT::i64) {
365 if (RetVT == MVT::f32)
366 return SINTTOFP_I64_F32;
367 else if (RetVT == MVT::f64)
368 return SINTTOFP_I64_F64;
369 else if (RetVT == MVT::f80)
370 return SINTTOFP_I64_F80;
371 else if (RetVT == MVT::ppcf128)
372 return SINTTOFP_I64_PPCF128;
373 } else if (OpVT == MVT::i128) {
374 if (RetVT == MVT::f32)
375 return SINTTOFP_I128_F32;
376 else if (RetVT == MVT::f64)
377 return SINTTOFP_I128_F64;
378 else if (RetVT == MVT::f80)
379 return SINTTOFP_I128_F80;
380 else if (RetVT == MVT::ppcf128)
381 return SINTTOFP_I128_PPCF128;
383 return UNKNOWN_LIBCALL;
386 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or
387 /// UNKNOWN_LIBCALL if there is none.
388 RTLIB::Libcall RTLIB::getUINTTOFP(MVT OpVT, MVT RetVT) {
389 if (OpVT == MVT::i32) {
390 if (RetVT == MVT::f32)
391 return UINTTOFP_I32_F32;
392 else if (RetVT == MVT::f64)
393 return UINTTOFP_I32_F64;
394 else if (RetVT == MVT::f80)
395 return UINTTOFP_I32_F80;
396 else if (RetVT == MVT::ppcf128)
397 return UINTTOFP_I32_PPCF128;
398 } else if (OpVT == MVT::i64) {
399 if (RetVT == MVT::f32)
400 return UINTTOFP_I64_F32;
401 else if (RetVT == MVT::f64)
402 return UINTTOFP_I64_F64;
403 else if (RetVT == MVT::f80)
404 return UINTTOFP_I64_F80;
405 else if (RetVT == MVT::ppcf128)
406 return UINTTOFP_I64_PPCF128;
407 } else if (OpVT == MVT::i128) {
408 if (RetVT == MVT::f32)
409 return UINTTOFP_I128_F32;
410 else if (RetVT == MVT::f64)
411 return UINTTOFP_I128_F64;
412 else if (RetVT == MVT::f80)
413 return UINTTOFP_I128_F80;
414 else if (RetVT == MVT::ppcf128)
415 return UINTTOFP_I128_PPCF128;
417 return UNKNOWN_LIBCALL;
420 /// InitCmpLibcallCCs - Set default comparison libcall CC.
422 static void InitCmpLibcallCCs(ISD::CondCode *CCs) {
423 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL);
424 CCs[RTLIB::OEQ_F32] = ISD::SETEQ;
425 CCs[RTLIB::OEQ_F64] = ISD::SETEQ;
426 CCs[RTLIB::UNE_F32] = ISD::SETNE;
427 CCs[RTLIB::UNE_F64] = ISD::SETNE;
428 CCs[RTLIB::OGE_F32] = ISD::SETGE;
429 CCs[RTLIB::OGE_F64] = ISD::SETGE;
430 CCs[RTLIB::OLT_F32] = ISD::SETLT;
431 CCs[RTLIB::OLT_F64] = ISD::SETLT;
432 CCs[RTLIB::OLE_F32] = ISD::SETLE;
433 CCs[RTLIB::OLE_F64] = ISD::SETLE;
434 CCs[RTLIB::OGT_F32] = ISD::SETGT;
435 CCs[RTLIB::OGT_F64] = ISD::SETGT;
436 CCs[RTLIB::UO_F32] = ISD::SETNE;
437 CCs[RTLIB::UO_F64] = ISD::SETNE;
438 CCs[RTLIB::O_F32] = ISD::SETEQ;
439 CCs[RTLIB::O_F64] = ISD::SETEQ;
442 TargetLowering::TargetLowering(TargetMachine &tm)
443 : TM(tm), TD(TM.getTargetData()) {
444 // All operations default to being supported.
445 memset(OpActions, 0, sizeof(OpActions));
446 memset(LoadExtActions, 0, sizeof(LoadExtActions));
447 memset(TruncStoreActions, 0, sizeof(TruncStoreActions));
448 memset(IndexedModeActions, 0, sizeof(IndexedModeActions));
449 memset(ConvertActions, 0, sizeof(ConvertActions));
450 memset(CondCodeActions, 0, sizeof(CondCodeActions));
452 // Set default actions for various operations.
453 for (unsigned VT = 0; VT != (unsigned)MVT::LAST_VALUETYPE; ++VT) {
454 // Default all indexed load / store to expand.
455 for (unsigned IM = (unsigned)ISD::PRE_INC;
456 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
457 setIndexedLoadAction(IM, (MVT::SimpleValueType)VT, Expand);
458 setIndexedStoreAction(IM, (MVT::SimpleValueType)VT, Expand);
461 // These operations default to expand.
462 setOperationAction(ISD::FGETSIGN, (MVT::SimpleValueType)VT, Expand);
463 setOperationAction(ISD::CONCAT_VECTORS, (MVT::SimpleValueType)VT, Expand);
466 // Most targets ignore the @llvm.prefetch intrinsic.
467 setOperationAction(ISD::PREFETCH, MVT::Other, Expand);
469 // ConstantFP nodes default to expand. Targets can either change this to
470 // Legal, in which case all fp constants are legal, or use addLegalFPImmediate
471 // to optimize expansions for certain constants.
472 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
473 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
474 setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
476 // These library functions default to expand.
477 setOperationAction(ISD::FLOG , MVT::f64, Expand);
478 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
479 setOperationAction(ISD::FLOG10,MVT::f64, Expand);
480 setOperationAction(ISD::FEXP , MVT::f64, Expand);
481 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
482 setOperationAction(ISD::FLOG , MVT::f32, Expand);
483 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
484 setOperationAction(ISD::FLOG10,MVT::f32, Expand);
485 setOperationAction(ISD::FEXP , MVT::f32, Expand);
486 setOperationAction(ISD::FEXP2, MVT::f32, Expand);
488 // Default ISD::TRAP to expand (which turns it into abort).
489 setOperationAction(ISD::TRAP, MVT::Other, Expand);
491 IsLittleEndian = TD->isLittleEndian();
492 UsesGlobalOffsetTable = false;
493 ShiftAmountTy = PointerTy = getValueType(TD->getIntPtrType());
494 ShiftAmtHandling = Undefined;
495 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE*sizeof(TargetRegisterClass*));
496 memset(TargetDAGCombineArray, 0, array_lengthof(TargetDAGCombineArray));
497 maxStoresPerMemset = maxStoresPerMemcpy = maxStoresPerMemmove = 8;
498 allowUnalignedMemoryAccesses = false;
499 benefitFromCodePlacementOpt = false;
500 UseUnderscoreSetJmp = false;
501 UseUnderscoreLongJmp = false;
502 SelectIsExpensive = false;
503 IntDivIsCheap = false;
504 Pow2DivIsCheap = false;
505 StackPointerRegisterToSaveRestore = 0;
506 ExceptionPointerRegister = 0;
507 ExceptionSelectorRegister = 0;
508 BooleanContents = UndefinedBooleanContent;
509 SchedPreferenceInfo = SchedulingForLatency;
511 JumpBufAlignment = 0;
512 IfCvtBlockSizeLimit = 2;
513 IfCvtDupBlockSizeLimit = 0;
514 PrefLoopAlignment = 0;
516 InitLibcallNames(LibcallRoutineNames);
517 InitCmpLibcallCCs(CmpLibcallCCs);
519 // Tell Legalize whether the assembler supports DEBUG_LOC.
520 const TargetAsmInfo *TASM = TM.getTargetAsmInfo();
521 if (!TASM || !TASM->hasDotLocAndDotFile())
522 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
525 TargetLowering::~TargetLowering() {}
527 /// computeRegisterProperties - Once all of the register classes are added,
528 /// this allows us to compute derived properties we expose.
529 void TargetLowering::computeRegisterProperties() {
530 assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE &&
531 "Too many value types for ValueTypeActions to hold!");
533 // Everything defaults to needing one register.
534 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
535 NumRegistersForVT[i] = 1;
536 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i;
538 // ...except isVoid, which doesn't need any registers.
539 NumRegistersForVT[MVT::isVoid] = 0;
541 // Find the largest integer register class.
542 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE;
543 for (; RegClassForVT[LargestIntReg] == 0; --LargestIntReg)
544 assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
546 // Every integer value type larger than this largest register takes twice as
547 // many registers to represent as the previous ValueType.
548 for (unsigned ExpandedReg = LargestIntReg + 1; ; ++ExpandedReg) {
549 MVT EVT = (MVT::SimpleValueType)ExpandedReg;
550 if (!EVT.isInteger())
552 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
553 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg;
554 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1);
555 ValueTypeActions.setTypeAction(EVT, Expand);
558 // Inspect all of the ValueType's smaller than the largest integer
559 // register to see which ones need promotion.
560 unsigned LegalIntReg = LargestIntReg;
561 for (unsigned IntReg = LargestIntReg - 1;
562 IntReg >= (unsigned)MVT::i1; --IntReg) {
563 MVT IVT = (MVT::SimpleValueType)IntReg;
564 if (isTypeLegal(IVT)) {
565 LegalIntReg = IntReg;
567 RegisterTypeForVT[IntReg] = TransformToType[IntReg] =
568 (MVT::SimpleValueType)LegalIntReg;
569 ValueTypeActions.setTypeAction(IVT, Promote);
573 // ppcf128 type is really two f64's.
574 if (!isTypeLegal(MVT::ppcf128)) {
575 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64];
576 RegisterTypeForVT[MVT::ppcf128] = MVT::f64;
577 TransformToType[MVT::ppcf128] = MVT::f64;
578 ValueTypeActions.setTypeAction(MVT::ppcf128, Expand);
581 // Decide how to handle f64. If the target does not have native f64 support,
582 // expand it to i64 and we will be generating soft float library calls.
583 if (!isTypeLegal(MVT::f64)) {
584 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
585 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
586 TransformToType[MVT::f64] = MVT::i64;
587 ValueTypeActions.setTypeAction(MVT::f64, Expand);
590 // Decide how to handle f32. If the target does not have native support for
591 // f32, promote it to f64 if it is legal. Otherwise, expand it to i32.
592 if (!isTypeLegal(MVT::f32)) {
593 if (isTypeLegal(MVT::f64)) {
594 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::f64];
595 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::f64];
596 TransformToType[MVT::f32] = MVT::f64;
597 ValueTypeActions.setTypeAction(MVT::f32, Promote);
599 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
600 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
601 TransformToType[MVT::f32] = MVT::i32;
602 ValueTypeActions.setTypeAction(MVT::f32, Expand);
606 // Loop over all of the vector value types to see which need transformations.
607 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
608 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
609 MVT VT = (MVT::SimpleValueType)i;
610 if (!isTypeLegal(VT)) {
611 MVT IntermediateVT, RegisterVT;
612 unsigned NumIntermediates;
613 NumRegistersForVT[i] =
614 getVectorTypeBreakdown(VT,
615 IntermediateVT, NumIntermediates,
617 RegisterTypeForVT[i] = RegisterVT;
619 // Determine if there is a legal wider type.
620 bool IsLegalWiderType = false;
621 MVT EltVT = VT.getVectorElementType();
622 unsigned NElts = VT.getVectorNumElements();
623 for (unsigned nVT = i+1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
624 MVT SVT = (MVT::SimpleValueType)nVT;
625 if (isTypeLegal(SVT) && SVT.getVectorElementType() == EltVT &&
626 SVT.getVectorNumElements() > NElts) {
627 TransformToType[i] = SVT;
628 ValueTypeActions.setTypeAction(VT, Promote);
629 IsLegalWiderType = true;
633 if (!IsLegalWiderType) {
634 MVT NVT = VT.getPow2VectorType();
636 // Type is already a power of 2. The default action is to split.
637 TransformToType[i] = MVT::Other;
638 ValueTypeActions.setTypeAction(VT, Expand);
640 TransformToType[i] = NVT;
641 ValueTypeActions.setTypeAction(VT, Promote);
648 const char *TargetLowering::getTargetNodeName(unsigned Opcode) const {
653 MVT TargetLowering::getSetCCResultType(MVT VT) const {
654 return getValueType(TD->getIntPtrType());
658 /// getVectorTypeBreakdown - Vector types are broken down into some number of
659 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
660 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
661 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
663 /// This method returns the number of registers needed, and the VT for each
664 /// register. It also returns the VT and quantity of the intermediate values
665 /// before they are promoted/expanded.
667 unsigned TargetLowering::getVectorTypeBreakdown(MVT VT,
669 unsigned &NumIntermediates,
670 MVT &RegisterVT) const {
671 // Figure out the right, legal destination reg to copy into.
672 unsigned NumElts = VT.getVectorNumElements();
673 MVT EltTy = VT.getVectorElementType();
675 unsigned NumVectorRegs = 1;
677 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
678 // could break down into LHS/RHS like LegalizeDAG does.
679 if (!isPowerOf2_32(NumElts)) {
680 NumVectorRegs = NumElts;
684 // Divide the input until we get to a supported size. This will always
685 // end with a scalar if the target doesn't support vectors.
686 while (NumElts > 1 && !isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) {
691 NumIntermediates = NumVectorRegs;
693 MVT NewVT = MVT::getVectorVT(EltTy, NumElts);
694 if (!isTypeLegal(NewVT))
696 IntermediateVT = NewVT;
698 MVT DestVT = getRegisterType(NewVT);
700 if (DestVT.bitsLT(NewVT)) {
701 // Value is expanded, e.g. i64 -> i16.
702 return NumVectorRegs*(NewVT.getSizeInBits()/DestVT.getSizeInBits());
704 // Otherwise, promotion or legal types use the same number of registers as
705 // the vector decimated to the appropriate level.
706 return NumVectorRegs;
712 /// getWidenVectorType: given a vector type, returns the type to widen to
713 /// (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
714 /// If there is no vector type that we want to widen to, returns MVT::Other
715 /// When and where to widen is target dependent based on the cost of
716 /// scalarizing vs using the wider vector type.
717 MVT TargetLowering::getWidenVectorType(MVT VT) const {
718 assert(VT.isVector());
722 // Default is not to widen until moved to LegalizeTypes
726 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
727 /// function arguments in the caller parameter area. This is the actual
728 /// alignment, not its logarithm.
729 unsigned TargetLowering::getByValTypeAlignment(const Type *Ty) const {
730 return TD->getCallFrameTypeAlignment(Ty);
733 SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table,
734 SelectionDAG &DAG) const {
735 if (usesGlobalOffsetTable())
736 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy());
741 TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
742 // Assume that everything is safe in static mode.
743 if (getTargetMachine().getRelocationModel() == Reloc::Static)
746 // In dynamic-no-pic mode, assume that known defined values are safe.
747 if (getTargetMachine().getRelocationModel() == Reloc::DynamicNoPIC &&
749 !GA->getGlobal()->isDeclaration() &&
750 !GA->getGlobal()->isWeakForLinker())
753 // Otherwise assume nothing is safe.
757 //===----------------------------------------------------------------------===//
758 // Optimization Methods
759 //===----------------------------------------------------------------------===//
761 /// ShrinkDemandedConstant - Check to see if the specified operand of the
762 /// specified instruction is a constant integer. If so, check to see if there
763 /// are any bits set in the constant that are not demanded. If so, shrink the
764 /// constant and return true.
765 bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDValue Op,
766 const APInt &Demanded) {
767 DebugLoc dl = Op.getDebugLoc();
769 // FIXME: ISD::SELECT, ISD::SELECT_CC
770 switch (Op.getOpcode()) {
775 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
776 if (!C) return false;
778 if (Op.getOpcode() == ISD::XOR &&
779 (C->getAPIntValue() | (~Demanded)).isAllOnesValue())
782 // if we can expand it to have all bits set, do it
783 if (C->getAPIntValue().intersects(~Demanded)) {
784 MVT VT = Op.getValueType();
785 SDValue New = DAG.getNode(Op.getOpcode(), dl, VT, Op.getOperand(0),
786 DAG.getConstant(Demanded &
789 return CombineTo(Op, New);
799 /// ShrinkDemandedOp - Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the
800 /// casts are free. This uses isZExtFree and ZERO_EXTEND for the widening
801 /// cast, but it could be generalized for targets with other types of
802 /// implicit widening casts.
804 TargetLowering::TargetLoweringOpt::ShrinkDemandedOp(SDValue Op,
806 const APInt &Demanded,
808 assert(Op.getNumOperands() == 2 &&
809 "ShrinkDemandedOp only supports binary operators!");
810 assert(Op.getNode()->getNumValues() == 1 &&
811 "ShrinkDemandedOp only supports nodes with one result!");
813 // Don't do this if the node has another user, which may require the
815 if (!Op.getNode()->hasOneUse())
818 // Search for the smallest integer type with free casts to and from
819 // Op's type. For expedience, just check power-of-2 integer types.
820 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
821 unsigned SmallVTBits = BitWidth - Demanded.countLeadingZeros();
822 if (!isPowerOf2_32(SmallVTBits))
823 SmallVTBits = NextPowerOf2(SmallVTBits);
824 for (; SmallVTBits < BitWidth; SmallVTBits = NextPowerOf2(SmallVTBits)) {
825 MVT SmallVT = MVT::getIntegerVT(SmallVTBits);
826 if (TLI.isTruncateFree(Op.getValueType(), SmallVT) &&
827 TLI.isZExtFree(SmallVT, Op.getValueType())) {
828 // We found a type with free casts.
829 SDValue X = DAG.getNode(Op.getOpcode(), dl, SmallVT,
830 DAG.getNode(ISD::TRUNCATE, dl, SmallVT,
831 Op.getNode()->getOperand(0)),
832 DAG.getNode(ISD::TRUNCATE, dl, SmallVT,
833 Op.getNode()->getOperand(1)));
834 SDValue Z = DAG.getNode(ISD::ZERO_EXTEND, dl, Op.getValueType(), X);
835 return CombineTo(Op, Z);
841 /// SimplifyDemandedBits - Look at Op. At this point, we know that only the
842 /// DemandedMask bits of the result of Op are ever used downstream. If we can
843 /// use this information to simplify Op, create a new simplified DAG node and
844 /// return true, returning the original and new nodes in Old and New. Otherwise,
845 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
846 /// the expression (used to simplify the caller). The KnownZero/One bits may
847 /// only be accurate for those bits in the DemandedMask.
848 bool TargetLowering::SimplifyDemandedBits(SDValue Op,
849 const APInt &DemandedMask,
852 TargetLoweringOpt &TLO,
853 unsigned Depth) const {
854 unsigned BitWidth = DemandedMask.getBitWidth();
855 assert(Op.getValueSizeInBits() == BitWidth &&
856 "Mask size mismatches value type size!");
857 APInt NewMask = DemandedMask;
858 DebugLoc dl = Op.getDebugLoc();
860 // Don't know anything.
861 KnownZero = KnownOne = APInt(BitWidth, 0);
863 // Other users may use these bits.
864 if (!Op.getNode()->hasOneUse()) {
866 // If not at the root, Just compute the KnownZero/KnownOne bits to
867 // simplify things downstream.
868 TLO.DAG.ComputeMaskedBits(Op, DemandedMask, KnownZero, KnownOne, Depth);
871 // If this is the root being simplified, allow it to have multiple uses,
872 // just set the NewMask to all bits.
873 NewMask = APInt::getAllOnesValue(BitWidth);
874 } else if (DemandedMask == 0) {
875 // Not demanding any bits from Op.
876 if (Op.getOpcode() != ISD::UNDEF)
877 return TLO.CombineTo(Op, TLO.DAG.getUNDEF(Op.getValueType()));
879 } else if (Depth == 6) { // Limit search depth.
883 APInt KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut;
884 switch (Op.getOpcode()) {
886 // We know all of the bits for a constant!
887 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue() & NewMask;
888 KnownZero = ~KnownOne & NewMask;
889 return false; // Don't fall through, will infinitely loop.
891 // If the RHS is a constant, check to see if the LHS would be zero without
892 // using the bits from the RHS. Below, we use knowledge about the RHS to
893 // simplify the LHS, here we're using information from the LHS to simplify
895 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
896 APInt LHSZero, LHSOne;
897 TLO.DAG.ComputeMaskedBits(Op.getOperand(0), NewMask,
898 LHSZero, LHSOne, Depth+1);
899 // If the LHS already has zeros where RHSC does, this and is dead.
900 if ((LHSZero & NewMask) == (~RHSC->getAPIntValue() & NewMask))
901 return TLO.CombineTo(Op, Op.getOperand(0));
902 // If any of the set bits in the RHS are known zero on the LHS, shrink
904 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & NewMask))
908 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
909 KnownOne, TLO, Depth+1))
911 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
912 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownZero & NewMask,
913 KnownZero2, KnownOne2, TLO, Depth+1))
915 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
917 // If all of the demanded bits are known one on one side, return the other.
918 // These bits cannot contribute to the result of the 'and'.
919 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
920 return TLO.CombineTo(Op, Op.getOperand(0));
921 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
922 return TLO.CombineTo(Op, Op.getOperand(1));
923 // If all of the demanded bits in the inputs are known zeros, return zero.
924 if ((NewMask & (KnownZero|KnownZero2)) == NewMask)
925 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, Op.getValueType()));
926 // If the RHS is a constant, see if we can simplify it.
927 if (TLO.ShrinkDemandedConstant(Op, ~KnownZero2 & NewMask))
929 // If the operation can be done in a smaller type, do so.
930 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
933 // Output known-1 bits are only known if set in both the LHS & RHS.
934 KnownOne &= KnownOne2;
935 // Output known-0 are known to be clear if zero in either the LHS | RHS.
936 KnownZero |= KnownZero2;
939 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
940 KnownOne, TLO, Depth+1))
942 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
943 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownOne & NewMask,
944 KnownZero2, KnownOne2, TLO, Depth+1))
946 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
948 // If all of the demanded bits are known zero on one side, return the other.
949 // These bits cannot contribute to the result of the 'or'.
950 if ((NewMask & ~KnownOne2 & KnownZero) == (~KnownOne2 & NewMask))
951 return TLO.CombineTo(Op, Op.getOperand(0));
952 if ((NewMask & ~KnownOne & KnownZero2) == (~KnownOne & NewMask))
953 return TLO.CombineTo(Op, Op.getOperand(1));
954 // If all of the potentially set bits on one side are known to be set on
955 // the other side, just use the 'other' side.
956 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
957 return TLO.CombineTo(Op, Op.getOperand(0));
958 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
959 return TLO.CombineTo(Op, Op.getOperand(1));
960 // If the RHS is a constant, see if we can simplify it.
961 if (TLO.ShrinkDemandedConstant(Op, NewMask))
963 // If the operation can be done in a smaller type, do so.
964 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
967 // Output known-0 bits are only known if clear in both the LHS & RHS.
968 KnownZero &= KnownZero2;
969 // Output known-1 are known to be set if set in either the LHS | RHS.
970 KnownOne |= KnownOne2;
973 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
974 KnownOne, TLO, Depth+1))
976 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
977 if (SimplifyDemandedBits(Op.getOperand(0), NewMask, KnownZero2,
978 KnownOne2, TLO, Depth+1))
980 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
982 // If all of the demanded bits are known zero on one side, return the other.
983 // These bits cannot contribute to the result of the 'xor'.
984 if ((KnownZero & NewMask) == NewMask)
985 return TLO.CombineTo(Op, Op.getOperand(0));
986 if ((KnownZero2 & NewMask) == NewMask)
987 return TLO.CombineTo(Op, Op.getOperand(1));
988 // If the operation can be done in a smaller type, do so.
989 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
992 // If all of the unknown bits are known to be zero on one side or the other
993 // (but not both) turn this into an *inclusive* or.
994 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
995 if ((NewMask & ~KnownZero & ~KnownZero2) == 0)
996 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, dl, Op.getValueType(),
1000 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1001 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
1002 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1003 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
1005 // If all of the demanded bits on one side are known, and all of the set
1006 // bits on that side are also known to be set on the other side, turn this
1007 // into an AND, as we know the bits will be cleared.
1008 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1009 if ((NewMask & (KnownZero|KnownOne)) == NewMask) { // all known
1010 if ((KnownOne & KnownOne2) == KnownOne) {
1011 MVT VT = Op.getValueType();
1012 SDValue ANDC = TLO.DAG.getConstant(~KnownOne & NewMask, VT);
1013 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, dl, VT,
1014 Op.getOperand(0), ANDC));
1018 // If the RHS is a constant, see if we can simplify it.
1019 // for XOR, we prefer to force bits to 1 if they will make a -1.
1020 // if we can't force bits, try to shrink constant
1021 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1022 APInt Expanded = C->getAPIntValue() | (~NewMask);
1023 // if we can expand it to have all bits set, do it
1024 if (Expanded.isAllOnesValue()) {
1025 if (Expanded != C->getAPIntValue()) {
1026 MVT VT = Op.getValueType();
1027 SDValue New = TLO.DAG.getNode(Op.getOpcode(), dl,VT, Op.getOperand(0),
1028 TLO.DAG.getConstant(Expanded, VT));
1029 return TLO.CombineTo(Op, New);
1031 // if it already has all the bits set, nothing to change
1032 // but don't shrink either!
1033 } else if (TLO.ShrinkDemandedConstant(Op, NewMask)) {
1038 KnownZero = KnownZeroOut;
1039 KnownOne = KnownOneOut;
1042 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero,
1043 KnownOne, TLO, Depth+1))
1045 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero2,
1046 KnownOne2, TLO, Depth+1))
1048 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1049 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1051 // If the operands are constants, see if we can simplify them.
1052 if (TLO.ShrinkDemandedConstant(Op, NewMask))
1055 // Only known if known in both the LHS and RHS.
1056 KnownOne &= KnownOne2;
1057 KnownZero &= KnownZero2;
1059 case ISD::SELECT_CC:
1060 if (SimplifyDemandedBits(Op.getOperand(3), NewMask, KnownZero,
1061 KnownOne, TLO, Depth+1))
1063 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero2,
1064 KnownOne2, TLO, Depth+1))
1066 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1067 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1069 // If the operands are constants, see if we can simplify them.
1070 if (TLO.ShrinkDemandedConstant(Op, NewMask))
1073 // Only known if known in both the LHS and RHS.
1074 KnownOne &= KnownOne2;
1075 KnownZero &= KnownZero2;
1078 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1079 unsigned ShAmt = SA->getZExtValue();
1080 SDValue InOp = Op.getOperand(0);
1082 // If the shift count is an invalid immediate, don't do anything.
1083 if (ShAmt >= BitWidth)
1086 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a
1087 // single shift. We can do this if the bottom bits (which are shifted
1088 // out) are never demanded.
1089 if (InOp.getOpcode() == ISD::SRL &&
1090 isa<ConstantSDNode>(InOp.getOperand(1))) {
1091 if (ShAmt && (NewMask & APInt::getLowBitsSet(BitWidth, ShAmt)) == 0) {
1092 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
1093 unsigned Opc = ISD::SHL;
1094 int Diff = ShAmt-C1;
1101 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
1102 MVT VT = Op.getValueType();
1103 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT,
1104 InOp.getOperand(0), NewSA));
1108 if (SimplifyDemandedBits(Op.getOperand(0), NewMask.lshr(ShAmt),
1109 KnownZero, KnownOne, TLO, Depth+1))
1111 KnownZero <<= SA->getZExtValue();
1112 KnownOne <<= SA->getZExtValue();
1113 // low bits known zero.
1114 KnownZero |= APInt::getLowBitsSet(BitWidth, SA->getZExtValue());
1118 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1119 MVT VT = Op.getValueType();
1120 unsigned ShAmt = SA->getZExtValue();
1121 unsigned VTSize = VT.getSizeInBits();
1122 SDValue InOp = Op.getOperand(0);
1124 // If the shift count is an invalid immediate, don't do anything.
1125 if (ShAmt >= BitWidth)
1128 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a
1129 // single shift. We can do this if the top bits (which are shifted out)
1130 // are never demanded.
1131 if (InOp.getOpcode() == ISD::SHL &&
1132 isa<ConstantSDNode>(InOp.getOperand(1))) {
1133 if (ShAmt && (NewMask & APInt::getHighBitsSet(VTSize, ShAmt)) == 0) {
1134 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
1135 unsigned Opc = ISD::SRL;
1136 int Diff = ShAmt-C1;
1143 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
1144 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT,
1145 InOp.getOperand(0), NewSA));
1149 // Compute the new bits that are at the top now.
1150 if (SimplifyDemandedBits(InOp, (NewMask << ShAmt),
1151 KnownZero, KnownOne, TLO, Depth+1))
1153 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1154 KnownZero = KnownZero.lshr(ShAmt);
1155 KnownOne = KnownOne.lshr(ShAmt);
1157 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
1158 KnownZero |= HighBits; // High bits known zero.
1162 // If this is an arithmetic shift right and only the low-bit is set, we can
1163 // always convert this into a logical shr, even if the shift amount is
1164 // variable. The low bit of the shift cannot be an input sign bit unless
1165 // the shift amount is >= the size of the datatype, which is undefined.
1166 if (DemandedMask == 1)
1167 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, Op.getValueType(),
1168 Op.getOperand(0), Op.getOperand(1)));
1170 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1171 MVT VT = Op.getValueType();
1172 unsigned ShAmt = SA->getZExtValue();
1174 // If the shift count is an invalid immediate, don't do anything.
1175 if (ShAmt >= BitWidth)
1178 APInt InDemandedMask = (NewMask << ShAmt);
1180 // If any of the demanded bits are produced by the sign extension, we also
1181 // demand the input sign bit.
1182 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
1183 if (HighBits.intersects(NewMask))
1184 InDemandedMask |= APInt::getSignBit(VT.getSizeInBits());
1186 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask,
1187 KnownZero, KnownOne, TLO, Depth+1))
1189 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1190 KnownZero = KnownZero.lshr(ShAmt);
1191 KnownOne = KnownOne.lshr(ShAmt);
1193 // Handle the sign bit, adjusted to where it is now in the mask.
1194 APInt SignBit = APInt::getSignBit(BitWidth).lshr(ShAmt);
1196 // If the input sign bit is known to be zero, or if none of the top bits
1197 // are demanded, turn this into an unsigned shift right.
1198 if (KnownZero.intersects(SignBit) || (HighBits & ~NewMask) == HighBits) {
1199 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, VT,
1202 } else if (KnownOne.intersects(SignBit)) { // New bits are known one.
1203 KnownOne |= HighBits;
1207 case ISD::SIGN_EXTEND_INREG: {
1208 MVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1210 // Sign extension. Compute the demanded bits in the result that are not
1211 // present in the input.
1212 APInt NewBits = APInt::getHighBitsSet(BitWidth,
1213 BitWidth - EVT.getSizeInBits()) &
1216 // If none of the extended bits are demanded, eliminate the sextinreg.
1218 return TLO.CombineTo(Op, Op.getOperand(0));
1220 APInt InSignBit = APInt::getSignBit(EVT.getSizeInBits());
1221 InSignBit.zext(BitWidth);
1222 APInt InputDemandedBits = APInt::getLowBitsSet(BitWidth,
1223 EVT.getSizeInBits()) &
1226 // Since the sign extended bits are demanded, we know that the sign
1228 InputDemandedBits |= InSignBit;
1230 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits,
1231 KnownZero, KnownOne, TLO, Depth+1))
1233 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1235 // If the sign bit of the input is known set or clear, then we know the
1236 // top bits of the result.
1238 // If the input sign bit is known zero, convert this into a zero extension.
1239 if (KnownZero.intersects(InSignBit))
1240 return TLO.CombineTo(Op,
1241 TLO.DAG.getZeroExtendInReg(Op.getOperand(0),dl,EVT));
1243 if (KnownOne.intersects(InSignBit)) { // Input sign bit known set
1244 KnownOne |= NewBits;
1245 KnownZero &= ~NewBits;
1246 } else { // Input sign bit unknown
1247 KnownZero &= ~NewBits;
1248 KnownOne &= ~NewBits;
1252 case ISD::ZERO_EXTEND: {
1253 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits();
1254 APInt InMask = NewMask;
1255 InMask.trunc(OperandBitWidth);
1257 // If none of the top bits are demanded, convert this into an any_extend.
1259 APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask;
1260 if (!NewBits.intersects(NewMask))
1261 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND, dl,
1265 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1266 KnownZero, KnownOne, TLO, Depth+1))
1268 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1269 KnownZero.zext(BitWidth);
1270 KnownOne.zext(BitWidth);
1271 KnownZero |= NewBits;
1274 case ISD::SIGN_EXTEND: {
1275 MVT InVT = Op.getOperand(0).getValueType();
1276 unsigned InBits = InVT.getSizeInBits();
1277 APInt InMask = APInt::getLowBitsSet(BitWidth, InBits);
1278 APInt InSignBit = APInt::getBitsSet(BitWidth, InBits - 1, InBits);
1279 APInt NewBits = ~InMask & NewMask;
1281 // If none of the top bits are demanded, convert this into an any_extend.
1283 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND, dl,
1287 // Since some of the sign extended bits are demanded, we know that the sign
1289 APInt InDemandedBits = InMask & NewMask;
1290 InDemandedBits |= InSignBit;
1291 InDemandedBits.trunc(InBits);
1293 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero,
1294 KnownOne, TLO, Depth+1))
1296 KnownZero.zext(BitWidth);
1297 KnownOne.zext(BitWidth);
1299 // If the sign bit is known zero, convert this to a zero extend.
1300 if (KnownZero.intersects(InSignBit))
1301 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND, dl,
1305 // If the sign bit is known one, the top bits match.
1306 if (KnownOne.intersects(InSignBit)) {
1307 KnownOne |= NewBits;
1308 KnownZero &= ~NewBits;
1309 } else { // Otherwise, top bits aren't known.
1310 KnownOne &= ~NewBits;
1311 KnownZero &= ~NewBits;
1315 case ISD::ANY_EXTEND: {
1316 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits();
1317 APInt InMask = NewMask;
1318 InMask.trunc(OperandBitWidth);
1319 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1320 KnownZero, KnownOne, TLO, Depth+1))
1322 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1323 KnownZero.zext(BitWidth);
1324 KnownOne.zext(BitWidth);
1327 case ISD::TRUNCATE: {
1328 // Simplify the input, using demanded bit information, and compute the known
1329 // zero/one bits live out.
1330 APInt TruncMask = NewMask;
1331 TruncMask.zext(Op.getOperand(0).getValueSizeInBits());
1332 if (SimplifyDemandedBits(Op.getOperand(0), TruncMask,
1333 KnownZero, KnownOne, TLO, Depth+1))
1335 KnownZero.trunc(BitWidth);
1336 KnownOne.trunc(BitWidth);
1338 // If the input is only used by this truncate, see if we can shrink it based
1339 // on the known demanded bits.
1340 if (Op.getOperand(0).getNode()->hasOneUse()) {
1341 SDValue In = Op.getOperand(0);
1342 unsigned InBitWidth = In.getValueSizeInBits();
1343 switch (In.getOpcode()) {
1346 // Shrink SRL by a constant if none of the high bits shifted in are
1348 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1))){
1349 APInt HighBits = APInt::getHighBitsSet(InBitWidth,
1350 InBitWidth - BitWidth);
1351 HighBits = HighBits.lshr(ShAmt->getZExtValue());
1352 HighBits.trunc(BitWidth);
1354 if (ShAmt->getZExtValue() < BitWidth && !(HighBits & NewMask)) {
1355 // None of the shifted in bits are needed. Add a truncate of the
1356 // shift input, then shift it.
1357 SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE, dl,
1360 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl,
1370 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1373 case ISD::AssertZext: {
1374 MVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1375 APInt InMask = APInt::getLowBitsSet(BitWidth,
1376 VT.getSizeInBits());
1377 if (SimplifyDemandedBits(Op.getOperand(0), InMask & NewMask,
1378 KnownZero, KnownOne, TLO, Depth+1))
1380 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1381 KnownZero |= ~InMask & NewMask;
1384 case ISD::BIT_CONVERT:
1386 // If this is an FP->Int bitcast and if the sign bit is the only thing that
1387 // is demanded, turn this into a FGETSIGN.
1388 if (NewMask == MVT::getIntegerVTSignBit(Op.getValueType()) &&
1389 MVT::isFloatingPoint(Op.getOperand(0).getValueType()) &&
1390 !MVT::isVector(Op.getOperand(0).getValueType())) {
1391 // Only do this xform if FGETSIGN is valid or if before legalize.
1392 if (!TLO.AfterLegalize ||
1393 isOperationLegal(ISD::FGETSIGN, Op.getValueType())) {
1394 // Make a FGETSIGN + SHL to move the sign bit into the appropriate
1395 // place. We expect the SHL to be eliminated by other optimizations.
1396 SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, Op.getValueType(),
1398 unsigned ShVal = Op.getValueType().getSizeInBits()-1;
1399 SDValue ShAmt = TLO.DAG.getConstant(ShVal, getShiftAmountTy());
1400 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, Op.getValueType(),
1409 // Add, Sub, and Mul don't demand any bits in positions beyond that
1410 // of the highest bit demanded of them.
1411 APInt LoMask = APInt::getLowBitsSet(BitWidth,
1412 BitWidth - NewMask.countLeadingZeros());
1413 if (SimplifyDemandedBits(Op.getOperand(0), LoMask, KnownZero2,
1414 KnownOne2, TLO, Depth+1))
1416 if (SimplifyDemandedBits(Op.getOperand(1), LoMask, KnownZero2,
1417 KnownOne2, TLO, Depth+1))
1419 // See if the operation should be performed at a smaller bit width.
1420 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1425 // Just use ComputeMaskedBits to compute output bits.
1426 TLO.DAG.ComputeMaskedBits(Op, NewMask, KnownZero, KnownOne, Depth);
1430 // If we know the value of all of the demanded bits, return this as a
1432 if ((NewMask & (KnownZero|KnownOne)) == NewMask)
1433 return TLO.CombineTo(Op, TLO.DAG.getConstant(KnownOne, Op.getValueType()));
1438 /// computeMaskedBitsForTargetNode - Determine which of the bits specified
1439 /// in Mask are known to be either zero or one and return them in the
1440 /// KnownZero/KnownOne bitsets.
1441 void TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
1445 const SelectionDAG &DAG,
1446 unsigned Depth) const {
1447 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1448 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1449 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1450 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1451 "Should use MaskedValueIsZero if you don't know whether Op"
1452 " is a target node!");
1453 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
1456 /// ComputeNumSignBitsForTargetNode - This method can be implemented by
1457 /// targets that want to expose additional information about sign bits to the
1459 unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
1460 unsigned Depth) const {
1461 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1462 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1463 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1464 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1465 "Should use ComputeNumSignBits if you don't know whether Op"
1466 " is a target node!");
1470 /// ValueHasExactlyOneBitSet - Test if the given value is known to have exactly
1471 /// one bit set. This differs from ComputeMaskedBits in that it doesn't need to
1472 /// determine which bit is set.
1474 static bool ValueHasExactlyOneBitSet(SDValue Val, const SelectionDAG &DAG) {
1475 // A left-shift of a constant one will have exactly one bit set, because
1476 // shifting the bit off the end is undefined.
1477 if (Val.getOpcode() == ISD::SHL)
1478 if (ConstantSDNode *C =
1479 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0)))
1480 if (C->getAPIntValue() == 1)
1483 // Similarly, a right-shift of a constant sign-bit will have exactly
1485 if (Val.getOpcode() == ISD::SRL)
1486 if (ConstantSDNode *C =
1487 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0)))
1488 if (C->getAPIntValue().isSignBit())
1491 // More could be done here, though the above checks are enough
1492 // to handle some common cases.
1494 // Fall back to ComputeMaskedBits to catch other known cases.
1495 MVT OpVT = Val.getValueType();
1496 unsigned BitWidth = OpVT.getSizeInBits();
1497 APInt Mask = APInt::getAllOnesValue(BitWidth);
1498 APInt KnownZero, KnownOne;
1499 DAG.ComputeMaskedBits(Val, Mask, KnownZero, KnownOne);
1500 return (KnownZero.countPopulation() == BitWidth - 1) &&
1501 (KnownOne.countPopulation() == 1);
1504 /// SimplifySetCC - Try to simplify a setcc built with the specified operands
1505 /// and cc. If it is unable to simplify it, return a null SDValue.
1507 TargetLowering::SimplifySetCC(MVT VT, SDValue N0, SDValue N1,
1508 ISD::CondCode Cond, bool foldBooleans,
1509 DAGCombinerInfo &DCI, DebugLoc dl) const {
1510 SelectionDAG &DAG = DCI.DAG;
1512 // These setcc operations always fold.
1516 case ISD::SETFALSE2: return DAG.getConstant(0, VT);
1518 case ISD::SETTRUE2: return DAG.getConstant(1, VT);
1521 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) {
1522 const APInt &C1 = N1C->getAPIntValue();
1523 if (isa<ConstantSDNode>(N0.getNode())) {
1524 return DAG.FoldSetCC(VT, N0, N1, Cond, dl);
1526 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an
1527 // equality comparison, then we're just comparing whether X itself is
1529 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) &&
1530 N0.getOperand(0).getOpcode() == ISD::CTLZ &&
1531 N0.getOperand(1).getOpcode() == ISD::Constant) {
1532 unsigned ShAmt = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
1533 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1534 ShAmt == Log2_32(N0.getValueType().getSizeInBits())) {
1535 if ((C1 == 0) == (Cond == ISD::SETEQ)) {
1536 // (srl (ctlz x), 5) == 0 -> X != 0
1537 // (srl (ctlz x), 5) != 1 -> X != 0
1540 // (srl (ctlz x), 5) != 0 -> X == 0
1541 // (srl (ctlz x), 5) == 1 -> X == 0
1544 SDValue Zero = DAG.getConstant(0, N0.getValueType());
1545 return DAG.getSetCC(dl, VT, N0.getOperand(0).getOperand(0),
1550 // If the LHS is '(and load, const)', the RHS is 0,
1551 // the test is for equality or unsigned, and all 1 bits of the const are
1552 // in the same partial word, see if we can shorten the load.
1553 if (DCI.isBeforeLegalize() &&
1554 N0.getOpcode() == ISD::AND && C1 == 0 &&
1555 N0.getNode()->hasOneUse() &&
1556 isa<LoadSDNode>(N0.getOperand(0)) &&
1557 N0.getOperand(0).getNode()->hasOneUse() &&
1558 isa<ConstantSDNode>(N0.getOperand(1))) {
1559 LoadSDNode *Lod = cast<LoadSDNode>(N0.getOperand(0));
1560 uint64_t bestMask = 0;
1561 unsigned bestWidth = 0, bestOffset = 0;
1562 if (!Lod->isVolatile() && Lod->isUnindexed() &&
1563 // FIXME: This uses getZExtValue() below so it only works on i64 and
1565 N0.getValueType().getSizeInBits() <= 64) {
1566 unsigned origWidth = N0.getValueType().getSizeInBits();
1567 // We can narrow (e.g.) 16-bit extending loads on 32-bit target to
1568 // 8 bits, but have to be careful...
1569 if (Lod->getExtensionType() != ISD::NON_EXTLOAD)
1570 origWidth = Lod->getMemoryVT().getSizeInBits();
1571 uint64_t Mask =cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
1572 for (unsigned width = origWidth / 2; width>=8; width /= 2) {
1573 uint64_t newMask = (1ULL << width) - 1;
1574 for (unsigned offset=0; offset<origWidth/width; offset++) {
1575 if ((newMask & Mask) == Mask) {
1576 if (!TD->isLittleEndian())
1577 bestOffset = (origWidth/width - offset - 1) * (width/8);
1579 bestOffset = (uint64_t)offset * (width/8);
1580 bestMask = Mask >> (offset * (width/8) * 8);
1584 newMask = newMask << width;
1589 MVT newVT = MVT::getIntegerVT(bestWidth);
1590 if (newVT.isRound()) {
1591 MVT PtrType = Lod->getOperand(1).getValueType();
1592 SDValue Ptr = Lod->getBasePtr();
1593 if (bestOffset != 0)
1594 Ptr = DAG.getNode(ISD::ADD, dl, PtrType, Lod->getBasePtr(),
1595 DAG.getConstant(bestOffset, PtrType));
1596 unsigned NewAlign = MinAlign(Lod->getAlignment(), bestOffset);
1597 SDValue NewLoad = DAG.getLoad(newVT, dl, Lod->getChain(), Ptr,
1599 Lod->getSrcValueOffset() + bestOffset,
1601 return DAG.getSetCC(dl, VT,
1602 DAG.getNode(ISD::AND, dl, newVT, NewLoad,
1603 DAG.getConstant(bestMask, newVT)),
1604 DAG.getConstant(0LL, newVT), Cond);
1609 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
1610 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
1611 unsigned InSize = N0.getOperand(0).getValueType().getSizeInBits();
1613 // If the comparison constant has bits in the upper part, the
1614 // zero-extended value could never match.
1615 if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(),
1616 C1.getBitWidth() - InSize))) {
1620 case ISD::SETEQ: return DAG.getConstant(0, VT);
1623 case ISD::SETNE: return DAG.getConstant(1, VT);
1626 // True if the sign bit of C1 is set.
1627 return DAG.getConstant(C1.isNegative(), VT);
1630 // True if the sign bit of C1 isn't set.
1631 return DAG.getConstant(C1.isNonNegative(), VT);
1637 // Otherwise, we can perform the comparison with the low bits.
1645 return DAG.getSetCC(dl, VT, N0.getOperand(0),
1646 DAG.getConstant(APInt(C1).trunc(InSize),
1647 N0.getOperand(0).getValueType()),
1650 break; // todo, be more careful with signed comparisons
1652 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
1653 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1654 MVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
1655 unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits();
1656 MVT ExtDstTy = N0.getValueType();
1657 unsigned ExtDstTyBits = ExtDstTy.getSizeInBits();
1659 // If the extended part has any inconsistent bits, it cannot ever
1660 // compare equal. In other words, they have to be all ones or all
1663 APInt::getHighBitsSet(ExtDstTyBits, ExtDstTyBits - ExtSrcTyBits);
1664 if ((C1 & ExtBits) != 0 && (C1 & ExtBits) != ExtBits)
1665 return DAG.getConstant(Cond == ISD::SETNE, VT);
1668 MVT Op0Ty = N0.getOperand(0).getValueType();
1669 if (Op0Ty == ExtSrcTy) {
1670 ZextOp = N0.getOperand(0);
1672 APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits);
1673 ZextOp = DAG.getNode(ISD::AND, dl, Op0Ty, N0.getOperand(0),
1674 DAG.getConstant(Imm, Op0Ty));
1676 if (!DCI.isCalledByLegalizer())
1677 DCI.AddToWorklist(ZextOp.getNode());
1678 // Otherwise, make this a use of a zext.
1679 return DAG.getSetCC(dl, VT, ZextOp,
1680 DAG.getConstant(C1 & APInt::getLowBitsSet(
1685 } else if ((N1C->isNullValue() || N1C->getAPIntValue() == 1) &&
1686 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1688 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC
1689 if (N0.getOpcode() == ISD::SETCC) {
1690 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getZExtValue() != 1);
1694 // Invert the condition.
1695 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
1696 CC = ISD::getSetCCInverse(CC,
1697 N0.getOperand(0).getValueType().isInteger());
1698 return DAG.getSetCC(dl, VT, N0.getOperand(0), N0.getOperand(1), CC);
1701 if ((N0.getOpcode() == ISD::XOR ||
1702 (N0.getOpcode() == ISD::AND &&
1703 N0.getOperand(0).getOpcode() == ISD::XOR &&
1704 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
1705 isa<ConstantSDNode>(N0.getOperand(1)) &&
1706 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue() == 1) {
1707 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We
1708 // can only do this if the top bits are known zero.
1709 unsigned BitWidth = N0.getValueSizeInBits();
1710 if (DAG.MaskedValueIsZero(N0,
1711 APInt::getHighBitsSet(BitWidth,
1713 // Okay, get the un-inverted input value.
1715 if (N0.getOpcode() == ISD::XOR)
1716 Val = N0.getOperand(0);
1718 assert(N0.getOpcode() == ISD::AND &&
1719 N0.getOperand(0).getOpcode() == ISD::XOR);
1720 // ((X^1)&1)^1 -> X & 1
1721 Val = DAG.getNode(ISD::AND, dl, N0.getValueType(),
1722 N0.getOperand(0).getOperand(0),
1725 return DAG.getSetCC(dl, VT, Val, N1,
1726 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
1731 APInt MinVal, MaxVal;
1732 unsigned OperandBitSize = N1C->getValueType(0).getSizeInBits();
1733 if (ISD::isSignedIntSetCC(Cond)) {
1734 MinVal = APInt::getSignedMinValue(OperandBitSize);
1735 MaxVal = APInt::getSignedMaxValue(OperandBitSize);
1737 MinVal = APInt::getMinValue(OperandBitSize);
1738 MaxVal = APInt::getMaxValue(OperandBitSize);
1741 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
1742 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
1743 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true
1744 // X >= C0 --> X > (C0-1)
1745 return DAG.getSetCC(dl, VT, N0,
1746 DAG.getConstant(C1-1, N1.getValueType()),
1747 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
1750 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
1751 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true
1752 // X <= C0 --> X < (C0+1)
1753 return DAG.getSetCC(dl, VT, N0,
1754 DAG.getConstant(C1+1, N1.getValueType()),
1755 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
1758 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
1759 return DAG.getConstant(0, VT); // X < MIN --> false
1760 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal)
1761 return DAG.getConstant(1, VT); // X >= MIN --> true
1762 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal)
1763 return DAG.getConstant(0, VT); // X > MAX --> false
1764 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal)
1765 return DAG.getConstant(1, VT); // X <= MAX --> true
1767 // Canonicalize setgt X, Min --> setne X, Min
1768 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
1769 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
1770 // Canonicalize setlt X, Max --> setne X, Max
1771 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
1772 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
1774 // If we have setult X, 1, turn it into seteq X, 0
1775 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
1776 return DAG.getSetCC(dl, VT, N0,
1777 DAG.getConstant(MinVal, N0.getValueType()),
1779 // If we have setugt X, Max-1, turn it into seteq X, Max
1780 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
1781 return DAG.getSetCC(dl, VT, N0,
1782 DAG.getConstant(MaxVal, N0.getValueType()),
1785 // If we have "setcc X, C0", check to see if we can shrink the immediate
1788 // SETUGT X, SINTMAX -> SETLT X, 0
1789 if (Cond == ISD::SETUGT &&
1790 C1 == APInt::getSignedMaxValue(OperandBitSize))
1791 return DAG.getSetCC(dl, VT, N0,
1792 DAG.getConstant(0, N1.getValueType()),
1795 // SETULT X, SINTMIN -> SETGT X, -1
1796 if (Cond == ISD::SETULT &&
1797 C1 == APInt::getSignedMinValue(OperandBitSize)) {
1798 SDValue ConstMinusOne =
1799 DAG.getConstant(APInt::getAllOnesValue(OperandBitSize),
1801 return DAG.getSetCC(dl, VT, N0, ConstMinusOne, ISD::SETGT);
1804 // Fold bit comparisons when we can.
1805 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1806 VT == N0.getValueType() && N0.getOpcode() == ISD::AND)
1807 if (ConstantSDNode *AndRHS =
1808 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1809 MVT ShiftTy = DCI.isBeforeLegalize() ?
1810 getPointerTy() : getShiftAmountTy();
1811 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
1812 // Perform the xform if the AND RHS is a single bit.
1813 if (isPowerOf2_64(AndRHS->getZExtValue())) {
1814 return DAG.getNode(ISD::SRL, dl, VT, N0,
1815 DAG.getConstant(Log2_64(AndRHS->getZExtValue()),
1818 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getZExtValue()) {
1819 // (X & 8) == 8 --> (X & 8) >> 3
1820 // Perform the xform if C1 is a single bit.
1821 if (C1.isPowerOf2()) {
1822 return DAG.getNode(ISD::SRL, dl, VT, N0,
1823 DAG.getConstant(C1.logBase2(), ShiftTy));
1828 } else if (isa<ConstantSDNode>(N0.getNode())) {
1829 // Ensure that the constant occurs on the RHS.
1830 return DAG.getSetCC(dl, VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
1833 if (isa<ConstantFPSDNode>(N0.getNode())) {
1834 // Constant fold or commute setcc.
1835 SDValue O = DAG.FoldSetCC(VT, N0, N1, Cond, dl);
1836 if (O.getNode()) return O;
1837 } else if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1.getNode())) {
1838 // If the RHS of an FP comparison is a constant, simplify it away in
1840 if (CFP->getValueAPF().isNaN()) {
1841 // If an operand is known to be a nan, we can fold it.
1842 switch (ISD::getUnorderedFlavor(Cond)) {
1843 default: assert(0 && "Unknown flavor!");
1844 case 0: // Known false.
1845 return DAG.getConstant(0, VT);
1846 case 1: // Known true.
1847 return DAG.getConstant(1, VT);
1848 case 2: // Undefined.
1849 return DAG.getUNDEF(VT);
1853 // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the
1854 // constant if knowing that the operand is non-nan is enough. We prefer to
1855 // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to
1857 if (Cond == ISD::SETO || Cond == ISD::SETUO)
1858 return DAG.getSetCC(dl, VT, N0, N0, Cond);
1862 // We can always fold X == X for integer setcc's.
1863 if (N0.getValueType().isInteger())
1864 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
1865 unsigned UOF = ISD::getUnorderedFlavor(Cond);
1866 if (UOF == 2) // FP operators that are undefined on NaNs.
1867 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
1868 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
1869 return DAG.getConstant(UOF, VT);
1870 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
1871 // if it is not already.
1872 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
1873 if (NewCond != Cond)
1874 return DAG.getSetCC(dl, VT, N0, N1, NewCond);
1877 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1878 N0.getValueType().isInteger()) {
1879 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
1880 N0.getOpcode() == ISD::XOR) {
1881 // Simplify (X+Y) == (X+Z) --> Y == Z
1882 if (N0.getOpcode() == N1.getOpcode()) {
1883 if (N0.getOperand(0) == N1.getOperand(0))
1884 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(1), Cond);
1885 if (N0.getOperand(1) == N1.getOperand(1))
1886 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(0), Cond);
1887 if (DAG.isCommutativeBinOp(N0.getOpcode())) {
1888 // If X op Y == Y op X, try other combinations.
1889 if (N0.getOperand(0) == N1.getOperand(1))
1890 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(0),
1892 if (N0.getOperand(1) == N1.getOperand(0))
1893 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(1),
1898 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) {
1899 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1900 // Turn (X+C1) == C2 --> X == C2-C1
1901 if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse()) {
1902 return DAG.getSetCC(dl, VT, N0.getOperand(0),
1903 DAG.getConstant(RHSC->getAPIntValue()-
1904 LHSR->getAPIntValue(),
1905 N0.getValueType()), Cond);
1908 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
1909 if (N0.getOpcode() == ISD::XOR)
1910 // If we know that all of the inverted bits are zero, don't bother
1911 // performing the inversion.
1912 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue()))
1914 DAG.getSetCC(dl, VT, N0.getOperand(0),
1915 DAG.getConstant(LHSR->getAPIntValue() ^
1916 RHSC->getAPIntValue(),
1921 // Turn (C1-X) == C2 --> X == C1-C2
1922 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
1923 if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse()) {
1925 DAG.getSetCC(dl, VT, N0.getOperand(1),
1926 DAG.getConstant(SUBC->getAPIntValue() -
1927 RHSC->getAPIntValue(),
1934 // Simplify (X+Z) == X --> Z == 0
1935 if (N0.getOperand(0) == N1)
1936 return DAG.getSetCC(dl, VT, N0.getOperand(1),
1937 DAG.getConstant(0, N0.getValueType()), Cond);
1938 if (N0.getOperand(1) == N1) {
1939 if (DAG.isCommutativeBinOp(N0.getOpcode()))
1940 return DAG.getSetCC(dl, VT, N0.getOperand(0),
1941 DAG.getConstant(0, N0.getValueType()), Cond);
1942 else if (N0.getNode()->hasOneUse()) {
1943 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
1944 // (Z-X) == X --> Z == X<<1
1945 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(),
1947 DAG.getConstant(1, getShiftAmountTy()));
1948 if (!DCI.isCalledByLegalizer())
1949 DCI.AddToWorklist(SH.getNode());
1950 return DAG.getSetCC(dl, VT, N0.getOperand(0), SH, Cond);
1955 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
1956 N1.getOpcode() == ISD::XOR) {
1957 // Simplify X == (X+Z) --> Z == 0
1958 if (N1.getOperand(0) == N0) {
1959 return DAG.getSetCC(dl, VT, N1.getOperand(1),
1960 DAG.getConstant(0, N1.getValueType()), Cond);
1961 } else if (N1.getOperand(1) == N0) {
1962 if (DAG.isCommutativeBinOp(N1.getOpcode())) {
1963 return DAG.getSetCC(dl, VT, N1.getOperand(0),
1964 DAG.getConstant(0, N1.getValueType()), Cond);
1965 } else if (N1.getNode()->hasOneUse()) {
1966 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
1967 // X == (Z-X) --> X<<1 == Z
1968 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(), N0,
1969 DAG.getConstant(1, getShiftAmountTy()));
1970 if (!DCI.isCalledByLegalizer())
1971 DCI.AddToWorklist(SH.getNode());
1972 return DAG.getSetCC(dl, VT, SH, N1.getOperand(0), Cond);
1977 // Simplify x&y == y to x&y != 0 if y has exactly one bit set.
1978 // Note that where y is variable and is known to have at most
1979 // one bit set (for example, if it is z&1) we cannot do this;
1980 // the expressions are not equivalent when y==0.
1981 if (N0.getOpcode() == ISD::AND)
1982 if (N0.getOperand(0) == N1 || N0.getOperand(1) == N1) {
1983 if (ValueHasExactlyOneBitSet(N1, DAG)) {
1984 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true);
1985 SDValue Zero = DAG.getConstant(0, N1.getValueType());
1986 return DAG.getSetCC(dl, VT, N0, Zero, Cond);
1989 if (N1.getOpcode() == ISD::AND)
1990 if (N1.getOperand(0) == N0 || N1.getOperand(1) == N0) {
1991 if (ValueHasExactlyOneBitSet(N0, DAG)) {
1992 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true);
1993 SDValue Zero = DAG.getConstant(0, N0.getValueType());
1994 return DAG.getSetCC(dl, VT, N1, Zero, Cond);
1999 // Fold away ALL boolean setcc's.
2001 if (N0.getValueType() == MVT::i1 && foldBooleans) {
2003 default: assert(0 && "Unknown integer setcc!");
2004 case ISD::SETEQ: // X == Y -> ~(X^Y)
2005 Temp = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1);
2006 N0 = DAG.getNOT(dl, Temp, MVT::i1);
2007 if (!DCI.isCalledByLegalizer())
2008 DCI.AddToWorklist(Temp.getNode());
2010 case ISD::SETNE: // X != Y --> (X^Y)
2011 N0 = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1);
2013 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> ~X & Y
2014 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> ~X & Y
2015 Temp = DAG.getNOT(dl, N0, MVT::i1);
2016 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N1, Temp);
2017 if (!DCI.isCalledByLegalizer())
2018 DCI.AddToWorklist(Temp.getNode());
2020 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> ~Y & X
2021 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> ~Y & X
2022 Temp = DAG.getNOT(dl, N1, MVT::i1);
2023 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N0, Temp);
2024 if (!DCI.isCalledByLegalizer())
2025 DCI.AddToWorklist(Temp.getNode());
2027 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> ~X | Y
2028 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> ~X | Y
2029 Temp = DAG.getNOT(dl, N0, MVT::i1);
2030 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N1, Temp);
2031 if (!DCI.isCalledByLegalizer())
2032 DCI.AddToWorklist(Temp.getNode());
2034 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> ~Y | X
2035 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> ~Y | X
2036 Temp = DAG.getNOT(dl, N1, MVT::i1);
2037 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N0, Temp);
2040 if (VT != MVT::i1) {
2041 if (!DCI.isCalledByLegalizer())
2042 DCI.AddToWorklist(N0.getNode());
2043 // FIXME: If running after legalize, we probably can't do this.
2044 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, N0);
2049 // Could not fold it.
2053 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
2054 /// node is a GlobalAddress + offset.
2055 bool TargetLowering::isGAPlusOffset(SDNode *N, GlobalValue* &GA,
2056 int64_t &Offset) const {
2057 if (isa<GlobalAddressSDNode>(N)) {
2058 GlobalAddressSDNode *GASD = cast<GlobalAddressSDNode>(N);
2059 GA = GASD->getGlobal();
2060 Offset += GASD->getOffset();
2064 if (N->getOpcode() == ISD::ADD) {
2065 SDValue N1 = N->getOperand(0);
2066 SDValue N2 = N->getOperand(1);
2067 if (isGAPlusOffset(N1.getNode(), GA, Offset)) {
2068 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
2070 Offset += V->getSExtValue();
2073 } else if (isGAPlusOffset(N2.getNode(), GA, Offset)) {
2074 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
2076 Offset += V->getSExtValue();
2085 /// isConsecutiveLoad - Return true if LD is loading 'Bytes' bytes from a
2086 /// location that is 'Dist' units away from the location that the 'Base' load
2087 /// is loading from.
2088 bool TargetLowering::isConsecutiveLoad(LoadSDNode *LD, LoadSDNode *Base,
2089 unsigned Bytes, int Dist,
2090 const MachineFrameInfo *MFI) const {
2091 if (LD->getChain() != Base->getChain())
2093 MVT VT = LD->getValueType(0);
2094 if (VT.getSizeInBits() / 8 != Bytes)
2097 SDValue Loc = LD->getOperand(1);
2098 SDValue BaseLoc = Base->getOperand(1);
2099 if (Loc.getOpcode() == ISD::FrameIndex) {
2100 if (BaseLoc.getOpcode() != ISD::FrameIndex)
2102 int FI = cast<FrameIndexSDNode>(Loc)->getIndex();
2103 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
2104 int FS = MFI->getObjectSize(FI);
2105 int BFS = MFI->getObjectSize(BFI);
2106 if (FS != BFS || FS != (int)Bytes) return false;
2107 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Bytes);
2109 if (Loc.getOpcode() == ISD::ADD && Loc.getOperand(0) == BaseLoc) {
2110 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Loc.getOperand(1));
2111 if (V && (V->getSExtValue() == Dist*Bytes))
2115 GlobalValue *GV1 = NULL;
2116 GlobalValue *GV2 = NULL;
2117 int64_t Offset1 = 0;
2118 int64_t Offset2 = 0;
2119 bool isGA1 = isGAPlusOffset(Loc.getNode(), GV1, Offset1);
2120 bool isGA2 = isGAPlusOffset(BaseLoc.getNode(), GV2, Offset2);
2121 if (isGA1 && isGA2 && GV1 == GV2)
2122 return Offset1 == (Offset2 + Dist*Bytes);
2127 SDValue TargetLowering::
2128 PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
2129 // Default implementation: no optimization.
2133 //===----------------------------------------------------------------------===//
2134 // Inline Assembler Implementation Methods
2135 //===----------------------------------------------------------------------===//
2138 TargetLowering::ConstraintType
2139 TargetLowering::getConstraintType(const std::string &Constraint) const {
2140 // FIXME: lots more standard ones to handle.
2141 if (Constraint.size() == 1) {
2142 switch (Constraint[0]) {
2144 case 'r': return C_RegisterClass;
2146 case 'o': // offsetable
2147 case 'V': // not offsetable
2149 case 'i': // Simple Integer or Relocatable Constant
2150 case 'n': // Simple Integer
2151 case 's': // Relocatable Constant
2152 case 'X': // Allow ANY value.
2153 case 'I': // Target registers.
2165 if (Constraint.size() > 1 && Constraint[0] == '{' &&
2166 Constraint[Constraint.size()-1] == '}')
2171 /// LowerXConstraint - try to replace an X constraint, which matches anything,
2172 /// with another that has more specific requirements based on the type of the
2173 /// corresponding operand.
2174 const char *TargetLowering::LowerXConstraint(MVT ConstraintVT) const{
2175 if (ConstraintVT.isInteger())
2177 if (ConstraintVT.isFloatingPoint())
2178 return "f"; // works for many targets
2182 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
2183 /// vector. If it is invalid, don't add anything to Ops.
2184 void TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
2185 char ConstraintLetter,
2187 std::vector<SDValue> &Ops,
2188 SelectionDAG &DAG) const {
2189 switch (ConstraintLetter) {
2191 case 'X': // Allows any operand; labels (basic block) use this.
2192 if (Op.getOpcode() == ISD::BasicBlock) {
2197 case 'i': // Simple Integer or Relocatable Constant
2198 case 'n': // Simple Integer
2199 case 's': { // Relocatable Constant
2200 // These operands are interested in values of the form (GV+C), where C may
2201 // be folded in as an offset of GV, or it may be explicitly added. Also, it
2202 // is possible and fine if either GV or C are missing.
2203 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2204 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
2206 // If we have "(add GV, C)", pull out GV/C
2207 if (Op.getOpcode() == ISD::ADD) {
2208 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
2209 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
2210 if (C == 0 || GA == 0) {
2211 C = dyn_cast<ConstantSDNode>(Op.getOperand(0));
2212 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1));
2214 if (C == 0 || GA == 0)
2218 // If we find a valid operand, map to the TargetXXX version so that the
2219 // value itself doesn't get selected.
2220 if (GA) { // Either &GV or &GV+C
2221 if (ConstraintLetter != 'n') {
2222 int64_t Offs = GA->getOffset();
2223 if (C) Offs += C->getZExtValue();
2224 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(),
2225 Op.getValueType(), Offs));
2229 if (C) { // just C, no GV.
2230 // Simple constants are not allowed for 's'.
2231 if (ConstraintLetter != 's') {
2232 // gcc prints these as sign extended. Sign extend value to 64 bits
2233 // now; without this it would get ZExt'd later in
2234 // ScheduleDAGSDNodes::EmitNode, which is very generic.
2235 Ops.push_back(DAG.getTargetConstant(C->getAPIntValue().getSExtValue(),
2245 std::vector<unsigned> TargetLowering::
2246 getRegClassForInlineAsmConstraint(const std::string &Constraint,
2248 return std::vector<unsigned>();
2252 std::pair<unsigned, const TargetRegisterClass*> TargetLowering::
2253 getRegForInlineAsmConstraint(const std::string &Constraint,
2255 if (Constraint[0] != '{')
2256 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
2257 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?");
2259 // Remove the braces from around the name.
2260 std::string RegName(Constraint.begin()+1, Constraint.end()-1);
2262 // Figure out which register class contains this reg.
2263 const TargetRegisterInfo *RI = TM.getRegisterInfo();
2264 for (TargetRegisterInfo::regclass_iterator RCI = RI->regclass_begin(),
2265 E = RI->regclass_end(); RCI != E; ++RCI) {
2266 const TargetRegisterClass *RC = *RCI;
2268 // If none of the the value types for this register class are valid, we
2269 // can't use it. For example, 64-bit reg classes on 32-bit targets.
2270 bool isLegal = false;
2271 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
2273 if (isTypeLegal(*I)) {
2279 if (!isLegal) continue;
2281 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end();
2283 if (StringsEqualNoCase(RegName, RI->get(*I).AsmName))
2284 return std::make_pair(*I, RC);
2288 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
2291 //===----------------------------------------------------------------------===//
2292 // Constraint Selection.
2294 /// isMatchingInputConstraint - Return true of this is an input operand that is
2295 /// a matching constraint like "4".
2296 bool TargetLowering::AsmOperandInfo::isMatchingInputConstraint() const {
2297 assert(!ConstraintCode.empty() && "No known constraint!");
2298 return isdigit(ConstraintCode[0]);
2301 /// getMatchedOperand - If this is an input matching constraint, this method
2302 /// returns the output operand it matches.
2303 unsigned TargetLowering::AsmOperandInfo::getMatchedOperand() const {
2304 assert(!ConstraintCode.empty() && "No known constraint!");
2305 return atoi(ConstraintCode.c_str());
2309 /// getConstraintGenerality - Return an integer indicating how general CT
2311 static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) {
2313 default: assert(0 && "Unknown constraint type!");
2314 case TargetLowering::C_Other:
2315 case TargetLowering::C_Unknown:
2317 case TargetLowering::C_Register:
2319 case TargetLowering::C_RegisterClass:
2321 case TargetLowering::C_Memory:
2326 /// ChooseConstraint - If there are multiple different constraints that we
2327 /// could pick for this operand (e.g. "imr") try to pick the 'best' one.
2328 /// This is somewhat tricky: constraints fall into four classes:
2329 /// Other -> immediates and magic values
2330 /// Register -> one specific register
2331 /// RegisterClass -> a group of regs
2332 /// Memory -> memory
2333 /// Ideally, we would pick the most specific constraint possible: if we have
2334 /// something that fits into a register, we would pick it. The problem here
2335 /// is that if we have something that could either be in a register or in
2336 /// memory that use of the register could cause selection of *other*
2337 /// operands to fail: they might only succeed if we pick memory. Because of
2338 /// this the heuristic we use is:
2340 /// 1) If there is an 'other' constraint, and if the operand is valid for
2341 /// that constraint, use it. This makes us take advantage of 'i'
2342 /// constraints when available.
2343 /// 2) Otherwise, pick the most general constraint present. This prefers
2344 /// 'm' over 'r', for example.
2346 static void ChooseConstraint(TargetLowering::AsmOperandInfo &OpInfo,
2347 bool hasMemory, const TargetLowering &TLI,
2348 SDValue Op, SelectionDAG *DAG) {
2349 assert(OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options");
2350 unsigned BestIdx = 0;
2351 TargetLowering::ConstraintType BestType = TargetLowering::C_Unknown;
2352 int BestGenerality = -1;
2354 // Loop over the options, keeping track of the most general one.
2355 for (unsigned i = 0, e = OpInfo.Codes.size(); i != e; ++i) {
2356 TargetLowering::ConstraintType CType =
2357 TLI.getConstraintType(OpInfo.Codes[i]);
2359 // If this is an 'other' constraint, see if the operand is valid for it.
2360 // For example, on X86 we might have an 'rI' constraint. If the operand
2361 // is an integer in the range [0..31] we want to use I (saving a load
2362 // of a register), otherwise we must use 'r'.
2363 if (CType == TargetLowering::C_Other && Op.getNode()) {
2364 assert(OpInfo.Codes[i].size() == 1 &&
2365 "Unhandled multi-letter 'other' constraint");
2366 std::vector<SDValue> ResultOps;
2367 TLI.LowerAsmOperandForConstraint(Op, OpInfo.Codes[i][0], hasMemory,
2369 if (!ResultOps.empty()) {
2376 // This constraint letter is more general than the previous one, use it.
2377 int Generality = getConstraintGenerality(CType);
2378 if (Generality > BestGenerality) {
2381 BestGenerality = Generality;
2385 OpInfo.ConstraintCode = OpInfo.Codes[BestIdx];
2386 OpInfo.ConstraintType = BestType;
2389 /// ComputeConstraintToUse - Determines the constraint code and constraint
2390 /// type to use for the specific AsmOperandInfo, setting
2391 /// OpInfo.ConstraintCode and OpInfo.ConstraintType.
2392 void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo,
2395 SelectionDAG *DAG) const {
2396 assert(!OpInfo.Codes.empty() && "Must have at least one constraint");
2398 // Single-letter constraints ('r') are very common.
2399 if (OpInfo.Codes.size() == 1) {
2400 OpInfo.ConstraintCode = OpInfo.Codes[0];
2401 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
2403 ChooseConstraint(OpInfo, hasMemory, *this, Op, DAG);
2406 // 'X' matches anything.
2407 if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) {
2408 // Labels and constants are handled elsewhere ('X' is the only thing
2409 // that matches labels).
2410 if (isa<BasicBlock>(OpInfo.CallOperandVal) ||
2411 isa<ConstantInt>(OpInfo.CallOperandVal))
2414 // Otherwise, try to resolve it to something we know about by looking at
2415 // the actual operand type.
2416 if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) {
2417 OpInfo.ConstraintCode = Repl;
2418 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
2423 //===----------------------------------------------------------------------===//
2424 // Loop Strength Reduction hooks
2425 //===----------------------------------------------------------------------===//
2427 /// isLegalAddressingMode - Return true if the addressing mode represented
2428 /// by AM is legal for this target, for a load/store of the specified type.
2429 bool TargetLowering::isLegalAddressingMode(const AddrMode &AM,
2430 const Type *Ty) const {
2431 // The default implementation of this implements a conservative RISCy, r+r and
2434 // Allows a sign-extended 16-bit immediate field.
2435 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
2438 // No global is ever allowed as a base.
2442 // Only support r+r,
2444 case 0: // "r+i" or just "i", depending on HasBaseReg.
2447 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
2449 // Otherwise we have r+r or r+i.
2452 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
2454 // Allow 2*r as r+r.
2461 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
2462 /// return a DAG expression to select that will generate the same value by
2463 /// multiplying by a magic number. See:
2464 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
2465 SDValue TargetLowering::BuildSDIV(SDNode *N, SelectionDAG &DAG,
2466 std::vector<SDNode*>* Created) const {
2467 MVT VT = N->getValueType(0);
2468 DebugLoc dl= N->getDebugLoc();
2470 // Check to see if we can do this.
2471 // FIXME: We should be more aggressive here.
2472 if (!isTypeLegal(VT))
2475 APInt d = cast<ConstantSDNode>(N->getOperand(1))->getAPIntValue();
2476 APInt::ms magics = d.magic();
2478 // Multiply the numerator (operand 0) by the magic value
2479 // FIXME: We should support doing a MUL in a wider type
2481 if (isOperationLegalOrCustom(ISD::MULHS, VT))
2482 Q = DAG.getNode(ISD::MULHS, dl, VT, N->getOperand(0),
2483 DAG.getConstant(magics.m, VT));
2484 else if (isOperationLegalOrCustom(ISD::SMUL_LOHI, VT))
2485 Q = SDValue(DAG.getNode(ISD::SMUL_LOHI, dl, DAG.getVTList(VT, VT),
2487 DAG.getConstant(magics.m, VT)).getNode(), 1);
2489 return SDValue(); // No mulhs or equvialent
2490 // If d > 0 and m < 0, add the numerator
2491 if (d.isStrictlyPositive() && magics.m.isNegative()) {
2492 Q = DAG.getNode(ISD::ADD, dl, VT, Q, N->getOperand(0));
2494 Created->push_back(Q.getNode());
2496 // If d < 0 and m > 0, subtract the numerator.
2497 if (d.isNegative() && magics.m.isStrictlyPositive()) {
2498 Q = DAG.getNode(ISD::SUB, dl, VT, Q, N->getOperand(0));
2500 Created->push_back(Q.getNode());
2502 // Shift right algebraic if shift value is nonzero
2504 Q = DAG.getNode(ISD::SRA, dl, VT, Q,
2505 DAG.getConstant(magics.s, getShiftAmountTy()));
2507 Created->push_back(Q.getNode());
2509 // Extract the sign bit and add it to the quotient
2511 DAG.getNode(ISD::SRL, dl, VT, Q, DAG.getConstant(VT.getSizeInBits()-1,
2512 getShiftAmountTy()));
2514 Created->push_back(T.getNode());
2515 return DAG.getNode(ISD::ADD, dl, VT, Q, T);
2518 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
2519 /// return a DAG expression to select that will generate the same value by
2520 /// multiplying by a magic number. See:
2521 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
2522 SDValue TargetLowering::BuildUDIV(SDNode *N, SelectionDAG &DAG,
2523 std::vector<SDNode*>* Created) const {
2524 MVT VT = N->getValueType(0);
2525 DebugLoc dl = N->getDebugLoc();
2527 // Check to see if we can do this.
2528 // FIXME: We should be more aggressive here.
2529 if (!isTypeLegal(VT))
2532 // FIXME: We should use a narrower constant when the upper
2533 // bits are known to be zero.
2534 ConstantSDNode *N1C = cast<ConstantSDNode>(N->getOperand(1));
2535 APInt::mu magics = N1C->getAPIntValue().magicu();
2537 // Multiply the numerator (operand 0) by the magic value
2538 // FIXME: We should support doing a MUL in a wider type
2540 if (isOperationLegalOrCustom(ISD::MULHU, VT))
2541 Q = DAG.getNode(ISD::MULHU, dl, VT, N->getOperand(0),
2542 DAG.getConstant(magics.m, VT));
2543 else if (isOperationLegalOrCustom(ISD::UMUL_LOHI, VT))
2544 Q = SDValue(DAG.getNode(ISD::UMUL_LOHI, dl, DAG.getVTList(VT, VT),
2546 DAG.getConstant(magics.m, VT)).getNode(), 1);
2548 return SDValue(); // No mulhu or equvialent
2550 Created->push_back(Q.getNode());
2552 if (magics.a == 0) {
2553 assert(magics.s < N1C->getAPIntValue().getBitWidth() &&
2554 "We shouldn't generate an undefined shift!");
2555 return DAG.getNode(ISD::SRL, dl, VT, Q,
2556 DAG.getConstant(magics.s, getShiftAmountTy()));
2558 SDValue NPQ = DAG.getNode(ISD::SUB, dl, VT, N->getOperand(0), Q);
2560 Created->push_back(NPQ.getNode());
2561 NPQ = DAG.getNode(ISD::SRL, dl, VT, NPQ,
2562 DAG.getConstant(1, getShiftAmountTy()));
2564 Created->push_back(NPQ.getNode());
2565 NPQ = DAG.getNode(ISD::ADD, dl, VT, NPQ, Q);
2567 Created->push_back(NPQ.getNode());
2568 return DAG.getNode(ISD::SRL, dl, VT, NPQ,
2569 DAG.getConstant(magics.s-1, getShiftAmountTy()));
2573 /// IgnoreHarmlessInstructions - Ignore instructions between a CALL and RET
2574 /// node that don't prevent tail call optimization.
2575 static SDValue IgnoreHarmlessInstructions(SDValue node) {
2576 // Found call return.
2577 if (node.getOpcode() == ISD::CALL) return node;
2578 // Ignore MERGE_VALUES. Will have at least one operand.
2579 if (node.getOpcode() == ISD::MERGE_VALUES)
2580 return IgnoreHarmlessInstructions(node.getOperand(0));
2581 // Ignore ANY_EXTEND node.
2582 if (node.getOpcode() == ISD::ANY_EXTEND)
2583 return IgnoreHarmlessInstructions(node.getOperand(0));
2584 if (node.getOpcode() == ISD::TRUNCATE)
2585 return IgnoreHarmlessInstructions(node.getOperand(0));
2586 // Any other node type.
2590 bool TargetLowering::CheckTailCallReturnConstraints(CallSDNode *TheCall,
2592 unsigned NumOps = Ret.getNumOperands();
2593 // ISD::CALL results:(value0, ..., valuen, chain)
2594 // ISD::RET operands:(chain, value0, flag0, ..., valuen, flagn)
2596 // Check that operand of the RET node sources from the CALL node. The RET node
2597 // has at least two operands. Operand 0 holds the chain. Operand 1 holds the
2599 // Also we need to check that there is no code in between the call and the
2600 // return. Hence we also check that the incomming chain to the return sources
2601 // from the outgoing chain of the call.
2603 IgnoreHarmlessInstructions(Ret.getOperand(1)) == SDValue(TheCall,0) &&
2604 Ret.getOperand(0) == SDValue(TheCall, TheCall->getNumValues()-1))
2606 // void return: The RET node has the chain result value of the CALL node as
2609 Ret.getOperand(0) == SDValue(TheCall, TheCall->getNumValues()-1))