1 //===-- TargetLowering.cpp - Implement the TargetLowering class -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This implements the TargetLowering class.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Target/TargetLowering.h"
15 #include "llvm/MC/MCAsmInfo.h"
16 #include "llvm/Target/TargetData.h"
17 #include "llvm/Target/TargetLoweringObjectFile.h"
18 #include "llvm/Target/TargetMachine.h"
19 #include "llvm/Target/TargetRegisterInfo.h"
20 #include "llvm/Target/TargetSubtarget.h"
21 #include "llvm/GlobalVariable.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/CodeGen/MachineFrameInfo.h"
24 #include "llvm/CodeGen/SelectionDAG.h"
25 #include "llvm/ADT/STLExtras.h"
26 #include "llvm/Support/ErrorHandling.h"
27 #include "llvm/Support/MathExtras.h"
31 TLSModel::Model getTLSModel(const GlobalValue *GV, Reloc::Model reloc) {
32 bool isLocal = GV->hasLocalLinkage();
33 bool isDeclaration = GV->isDeclaration();
34 // FIXME: what should we do for protected and internal visibility?
35 // For variables, is internal different from hidden?
36 bool isHidden = GV->hasHiddenVisibility();
38 if (reloc == Reloc::PIC_) {
39 if (isLocal || isHidden)
40 return TLSModel::LocalDynamic;
42 return TLSModel::GeneralDynamic;
44 if (!isDeclaration || isHidden)
45 return TLSModel::LocalExec;
47 return TLSModel::InitialExec;
52 /// InitLibcallNames - Set default libcall names.
54 static void InitLibcallNames(const char **Names) {
55 Names[RTLIB::SHL_I16] = "__ashlhi3";
56 Names[RTLIB::SHL_I32] = "__ashlsi3";
57 Names[RTLIB::SHL_I64] = "__ashldi3";
58 Names[RTLIB::SHL_I128] = "__ashlti3";
59 Names[RTLIB::SRL_I16] = "__lshrhi3";
60 Names[RTLIB::SRL_I32] = "__lshrsi3";
61 Names[RTLIB::SRL_I64] = "__lshrdi3";
62 Names[RTLIB::SRL_I128] = "__lshrti3";
63 Names[RTLIB::SRA_I16] = "__ashrhi3";
64 Names[RTLIB::SRA_I32] = "__ashrsi3";
65 Names[RTLIB::SRA_I64] = "__ashrdi3";
66 Names[RTLIB::SRA_I128] = "__ashrti3";
67 Names[RTLIB::MUL_I8] = "__mulqi3";
68 Names[RTLIB::MUL_I16] = "__mulhi3";
69 Names[RTLIB::MUL_I32] = "__mulsi3";
70 Names[RTLIB::MUL_I64] = "__muldi3";
71 Names[RTLIB::MUL_I128] = "__multi3";
72 Names[RTLIB::SDIV_I8] = "__divqi3";
73 Names[RTLIB::SDIV_I16] = "__divhi3";
74 Names[RTLIB::SDIV_I32] = "__divsi3";
75 Names[RTLIB::SDIV_I64] = "__divdi3";
76 Names[RTLIB::SDIV_I128] = "__divti3";
77 Names[RTLIB::UDIV_I8] = "__udivqi3";
78 Names[RTLIB::UDIV_I16] = "__udivhi3";
79 Names[RTLIB::UDIV_I32] = "__udivsi3";
80 Names[RTLIB::UDIV_I64] = "__udivdi3";
81 Names[RTLIB::UDIV_I128] = "__udivti3";
82 Names[RTLIB::SREM_I8] = "__modqi3";
83 Names[RTLIB::SREM_I16] = "__modhi3";
84 Names[RTLIB::SREM_I32] = "__modsi3";
85 Names[RTLIB::SREM_I64] = "__moddi3";
86 Names[RTLIB::SREM_I128] = "__modti3";
87 Names[RTLIB::UREM_I8] = "__umodqi3";
88 Names[RTLIB::UREM_I16] = "__umodhi3";
89 Names[RTLIB::UREM_I32] = "__umodsi3";
90 Names[RTLIB::UREM_I64] = "__umoddi3";
91 Names[RTLIB::UREM_I128] = "__umodti3";
92 Names[RTLIB::NEG_I32] = "__negsi2";
93 Names[RTLIB::NEG_I64] = "__negdi2";
94 Names[RTLIB::ADD_F32] = "__addsf3";
95 Names[RTLIB::ADD_F64] = "__adddf3";
96 Names[RTLIB::ADD_F80] = "__addxf3";
97 Names[RTLIB::ADD_PPCF128] = "__gcc_qadd";
98 Names[RTLIB::SUB_F32] = "__subsf3";
99 Names[RTLIB::SUB_F64] = "__subdf3";
100 Names[RTLIB::SUB_F80] = "__subxf3";
101 Names[RTLIB::SUB_PPCF128] = "__gcc_qsub";
102 Names[RTLIB::MUL_F32] = "__mulsf3";
103 Names[RTLIB::MUL_F64] = "__muldf3";
104 Names[RTLIB::MUL_F80] = "__mulxf3";
105 Names[RTLIB::MUL_PPCF128] = "__gcc_qmul";
106 Names[RTLIB::DIV_F32] = "__divsf3";
107 Names[RTLIB::DIV_F64] = "__divdf3";
108 Names[RTLIB::DIV_F80] = "__divxf3";
109 Names[RTLIB::DIV_PPCF128] = "__gcc_qdiv";
110 Names[RTLIB::REM_F32] = "fmodf";
111 Names[RTLIB::REM_F64] = "fmod";
112 Names[RTLIB::REM_F80] = "fmodl";
113 Names[RTLIB::REM_PPCF128] = "fmodl";
114 Names[RTLIB::POWI_F32] = "__powisf2";
115 Names[RTLIB::POWI_F64] = "__powidf2";
116 Names[RTLIB::POWI_F80] = "__powixf2";
117 Names[RTLIB::POWI_PPCF128] = "__powitf2";
118 Names[RTLIB::SQRT_F32] = "sqrtf";
119 Names[RTLIB::SQRT_F64] = "sqrt";
120 Names[RTLIB::SQRT_F80] = "sqrtl";
121 Names[RTLIB::SQRT_PPCF128] = "sqrtl";
122 Names[RTLIB::LOG_F32] = "logf";
123 Names[RTLIB::LOG_F64] = "log";
124 Names[RTLIB::LOG_F80] = "logl";
125 Names[RTLIB::LOG_PPCF128] = "logl";
126 Names[RTLIB::LOG2_F32] = "log2f";
127 Names[RTLIB::LOG2_F64] = "log2";
128 Names[RTLIB::LOG2_F80] = "log2l";
129 Names[RTLIB::LOG2_PPCF128] = "log2l";
130 Names[RTLIB::LOG10_F32] = "log10f";
131 Names[RTLIB::LOG10_F64] = "log10";
132 Names[RTLIB::LOG10_F80] = "log10l";
133 Names[RTLIB::LOG10_PPCF128] = "log10l";
134 Names[RTLIB::EXP_F32] = "expf";
135 Names[RTLIB::EXP_F64] = "exp";
136 Names[RTLIB::EXP_F80] = "expl";
137 Names[RTLIB::EXP_PPCF128] = "expl";
138 Names[RTLIB::EXP2_F32] = "exp2f";
139 Names[RTLIB::EXP2_F64] = "exp2";
140 Names[RTLIB::EXP2_F80] = "exp2l";
141 Names[RTLIB::EXP2_PPCF128] = "exp2l";
142 Names[RTLIB::SIN_F32] = "sinf";
143 Names[RTLIB::SIN_F64] = "sin";
144 Names[RTLIB::SIN_F80] = "sinl";
145 Names[RTLIB::SIN_PPCF128] = "sinl";
146 Names[RTLIB::COS_F32] = "cosf";
147 Names[RTLIB::COS_F64] = "cos";
148 Names[RTLIB::COS_F80] = "cosl";
149 Names[RTLIB::COS_PPCF128] = "cosl";
150 Names[RTLIB::POW_F32] = "powf";
151 Names[RTLIB::POW_F64] = "pow";
152 Names[RTLIB::POW_F80] = "powl";
153 Names[RTLIB::POW_PPCF128] = "powl";
154 Names[RTLIB::CEIL_F32] = "ceilf";
155 Names[RTLIB::CEIL_F64] = "ceil";
156 Names[RTLIB::CEIL_F80] = "ceill";
157 Names[RTLIB::CEIL_PPCF128] = "ceill";
158 Names[RTLIB::TRUNC_F32] = "truncf";
159 Names[RTLIB::TRUNC_F64] = "trunc";
160 Names[RTLIB::TRUNC_F80] = "truncl";
161 Names[RTLIB::TRUNC_PPCF128] = "truncl";
162 Names[RTLIB::RINT_F32] = "rintf";
163 Names[RTLIB::RINT_F64] = "rint";
164 Names[RTLIB::RINT_F80] = "rintl";
165 Names[RTLIB::RINT_PPCF128] = "rintl";
166 Names[RTLIB::NEARBYINT_F32] = "nearbyintf";
167 Names[RTLIB::NEARBYINT_F64] = "nearbyint";
168 Names[RTLIB::NEARBYINT_F80] = "nearbyintl";
169 Names[RTLIB::NEARBYINT_PPCF128] = "nearbyintl";
170 Names[RTLIB::FLOOR_F32] = "floorf";
171 Names[RTLIB::FLOOR_F64] = "floor";
172 Names[RTLIB::FLOOR_F80] = "floorl";
173 Names[RTLIB::FLOOR_PPCF128] = "floorl";
174 Names[RTLIB::FPEXT_F32_F64] = "__extendsfdf2";
175 Names[RTLIB::FPROUND_F64_F32] = "__truncdfsf2";
176 Names[RTLIB::FPROUND_F80_F32] = "__truncxfsf2";
177 Names[RTLIB::FPROUND_PPCF128_F32] = "__trunctfsf2";
178 Names[RTLIB::FPROUND_F80_F64] = "__truncxfdf2";
179 Names[RTLIB::FPROUND_PPCF128_F64] = "__trunctfdf2";
180 Names[RTLIB::FPTOSINT_F32_I8] = "__fixsfi8";
181 Names[RTLIB::FPTOSINT_F32_I16] = "__fixsfi16";
182 Names[RTLIB::FPTOSINT_F32_I32] = "__fixsfsi";
183 Names[RTLIB::FPTOSINT_F32_I64] = "__fixsfdi";
184 Names[RTLIB::FPTOSINT_F32_I128] = "__fixsfti";
185 Names[RTLIB::FPTOSINT_F64_I32] = "__fixdfsi";
186 Names[RTLIB::FPTOSINT_F64_I64] = "__fixdfdi";
187 Names[RTLIB::FPTOSINT_F64_I128] = "__fixdfti";
188 Names[RTLIB::FPTOSINT_F80_I32] = "__fixxfsi";
189 Names[RTLIB::FPTOSINT_F80_I64] = "__fixxfdi";
190 Names[RTLIB::FPTOSINT_F80_I128] = "__fixxfti";
191 Names[RTLIB::FPTOSINT_PPCF128_I32] = "__fixtfsi";
192 Names[RTLIB::FPTOSINT_PPCF128_I64] = "__fixtfdi";
193 Names[RTLIB::FPTOSINT_PPCF128_I128] = "__fixtfti";
194 Names[RTLIB::FPTOUINT_F32_I8] = "__fixunssfi8";
195 Names[RTLIB::FPTOUINT_F32_I16] = "__fixunssfi16";
196 Names[RTLIB::FPTOUINT_F32_I32] = "__fixunssfsi";
197 Names[RTLIB::FPTOUINT_F32_I64] = "__fixunssfdi";
198 Names[RTLIB::FPTOUINT_F32_I128] = "__fixunssfti";
199 Names[RTLIB::FPTOUINT_F64_I32] = "__fixunsdfsi";
200 Names[RTLIB::FPTOUINT_F64_I64] = "__fixunsdfdi";
201 Names[RTLIB::FPTOUINT_F64_I128] = "__fixunsdfti";
202 Names[RTLIB::FPTOUINT_F80_I32] = "__fixunsxfsi";
203 Names[RTLIB::FPTOUINT_F80_I64] = "__fixunsxfdi";
204 Names[RTLIB::FPTOUINT_F80_I128] = "__fixunsxfti";
205 Names[RTLIB::FPTOUINT_PPCF128_I32] = "__fixunstfsi";
206 Names[RTLIB::FPTOUINT_PPCF128_I64] = "__fixunstfdi";
207 Names[RTLIB::FPTOUINT_PPCF128_I128] = "__fixunstfti";
208 Names[RTLIB::SINTTOFP_I32_F32] = "__floatsisf";
209 Names[RTLIB::SINTTOFP_I32_F64] = "__floatsidf";
210 Names[RTLIB::SINTTOFP_I32_F80] = "__floatsixf";
211 Names[RTLIB::SINTTOFP_I32_PPCF128] = "__floatsitf";
212 Names[RTLIB::SINTTOFP_I64_F32] = "__floatdisf";
213 Names[RTLIB::SINTTOFP_I64_F64] = "__floatdidf";
214 Names[RTLIB::SINTTOFP_I64_F80] = "__floatdixf";
215 Names[RTLIB::SINTTOFP_I64_PPCF128] = "__floatditf";
216 Names[RTLIB::SINTTOFP_I128_F32] = "__floattisf";
217 Names[RTLIB::SINTTOFP_I128_F64] = "__floattidf";
218 Names[RTLIB::SINTTOFP_I128_F80] = "__floattixf";
219 Names[RTLIB::SINTTOFP_I128_PPCF128] = "__floattitf";
220 Names[RTLIB::UINTTOFP_I32_F32] = "__floatunsisf";
221 Names[RTLIB::UINTTOFP_I32_F64] = "__floatunsidf";
222 Names[RTLIB::UINTTOFP_I32_F80] = "__floatunsixf";
223 Names[RTLIB::UINTTOFP_I32_PPCF128] = "__floatunsitf";
224 Names[RTLIB::UINTTOFP_I64_F32] = "__floatundisf";
225 Names[RTLIB::UINTTOFP_I64_F64] = "__floatundidf";
226 Names[RTLIB::UINTTOFP_I64_F80] = "__floatundixf";
227 Names[RTLIB::UINTTOFP_I64_PPCF128] = "__floatunditf";
228 Names[RTLIB::UINTTOFP_I128_F32] = "__floatuntisf";
229 Names[RTLIB::UINTTOFP_I128_F64] = "__floatuntidf";
230 Names[RTLIB::UINTTOFP_I128_F80] = "__floatuntixf";
231 Names[RTLIB::UINTTOFP_I128_PPCF128] = "__floatuntitf";
232 Names[RTLIB::OEQ_F32] = "__eqsf2";
233 Names[RTLIB::OEQ_F64] = "__eqdf2";
234 Names[RTLIB::UNE_F32] = "__nesf2";
235 Names[RTLIB::UNE_F64] = "__nedf2";
236 Names[RTLIB::OGE_F32] = "__gesf2";
237 Names[RTLIB::OGE_F64] = "__gedf2";
238 Names[RTLIB::OLT_F32] = "__ltsf2";
239 Names[RTLIB::OLT_F64] = "__ltdf2";
240 Names[RTLIB::OLE_F32] = "__lesf2";
241 Names[RTLIB::OLE_F64] = "__ledf2";
242 Names[RTLIB::OGT_F32] = "__gtsf2";
243 Names[RTLIB::OGT_F64] = "__gtdf2";
244 Names[RTLIB::UO_F32] = "__unordsf2";
245 Names[RTLIB::UO_F64] = "__unorddf2";
246 Names[RTLIB::O_F32] = "__unordsf2";
247 Names[RTLIB::O_F64] = "__unorddf2";
248 Names[RTLIB::MEMCPY] = "memcpy";
249 Names[RTLIB::MEMMOVE] = "memmove";
250 Names[RTLIB::MEMSET] = "memset";
251 Names[RTLIB::UNWIND_RESUME] = "_Unwind_Resume";
254 /// InitLibcallCallingConvs - Set default libcall CallingConvs.
256 static void InitLibcallCallingConvs(CallingConv::ID *CCs) {
257 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
258 CCs[i] = CallingConv::C;
262 /// getFPEXT - Return the FPEXT_*_* value for the given types, or
263 /// UNKNOWN_LIBCALL if there is none.
264 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) {
265 if (OpVT == MVT::f32) {
266 if (RetVT == MVT::f64)
267 return FPEXT_F32_F64;
269 return UNKNOWN_LIBCALL;
272 /// getFPROUND - Return the FPROUND_*_* value for the given types, or
273 /// UNKNOWN_LIBCALL if there is none.
274 RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) {
275 if (RetVT == MVT::f32) {
276 if (OpVT == MVT::f64)
277 return FPROUND_F64_F32;
278 if (OpVT == MVT::f80)
279 return FPROUND_F80_F32;
280 if (OpVT == MVT::ppcf128)
281 return FPROUND_PPCF128_F32;
282 } else if (RetVT == MVT::f64) {
283 if (OpVT == MVT::f80)
284 return FPROUND_F80_F64;
285 if (OpVT == MVT::ppcf128)
286 return FPROUND_PPCF128_F64;
288 return UNKNOWN_LIBCALL;
291 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or
292 /// UNKNOWN_LIBCALL if there is none.
293 RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) {
294 if (OpVT == MVT::f32) {
295 if (RetVT == MVT::i8)
296 return FPTOSINT_F32_I8;
297 if (RetVT == MVT::i16)
298 return FPTOSINT_F32_I16;
299 if (RetVT == MVT::i32)
300 return FPTOSINT_F32_I32;
301 if (RetVT == MVT::i64)
302 return FPTOSINT_F32_I64;
303 if (RetVT == MVT::i128)
304 return FPTOSINT_F32_I128;
305 } else if (OpVT == MVT::f64) {
306 if (RetVT == MVT::i32)
307 return FPTOSINT_F64_I32;
308 if (RetVT == MVT::i64)
309 return FPTOSINT_F64_I64;
310 if (RetVT == MVT::i128)
311 return FPTOSINT_F64_I128;
312 } else if (OpVT == MVT::f80) {
313 if (RetVT == MVT::i32)
314 return FPTOSINT_F80_I32;
315 if (RetVT == MVT::i64)
316 return FPTOSINT_F80_I64;
317 if (RetVT == MVT::i128)
318 return FPTOSINT_F80_I128;
319 } else if (OpVT == MVT::ppcf128) {
320 if (RetVT == MVT::i32)
321 return FPTOSINT_PPCF128_I32;
322 if (RetVT == MVT::i64)
323 return FPTOSINT_PPCF128_I64;
324 if (RetVT == MVT::i128)
325 return FPTOSINT_PPCF128_I128;
327 return UNKNOWN_LIBCALL;
330 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or
331 /// UNKNOWN_LIBCALL if there is none.
332 RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) {
333 if (OpVT == MVT::f32) {
334 if (RetVT == MVT::i8)
335 return FPTOUINT_F32_I8;
336 if (RetVT == MVT::i16)
337 return FPTOUINT_F32_I16;
338 if (RetVT == MVT::i32)
339 return FPTOUINT_F32_I32;
340 if (RetVT == MVT::i64)
341 return FPTOUINT_F32_I64;
342 if (RetVT == MVT::i128)
343 return FPTOUINT_F32_I128;
344 } else if (OpVT == MVT::f64) {
345 if (RetVT == MVT::i32)
346 return FPTOUINT_F64_I32;
347 if (RetVT == MVT::i64)
348 return FPTOUINT_F64_I64;
349 if (RetVT == MVT::i128)
350 return FPTOUINT_F64_I128;
351 } else if (OpVT == MVT::f80) {
352 if (RetVT == MVT::i32)
353 return FPTOUINT_F80_I32;
354 if (RetVT == MVT::i64)
355 return FPTOUINT_F80_I64;
356 if (RetVT == MVT::i128)
357 return FPTOUINT_F80_I128;
358 } else if (OpVT == MVT::ppcf128) {
359 if (RetVT == MVT::i32)
360 return FPTOUINT_PPCF128_I32;
361 if (RetVT == MVT::i64)
362 return FPTOUINT_PPCF128_I64;
363 if (RetVT == MVT::i128)
364 return FPTOUINT_PPCF128_I128;
366 return UNKNOWN_LIBCALL;
369 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or
370 /// UNKNOWN_LIBCALL if there is none.
371 RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) {
372 if (OpVT == MVT::i32) {
373 if (RetVT == MVT::f32)
374 return SINTTOFP_I32_F32;
375 else if (RetVT == MVT::f64)
376 return SINTTOFP_I32_F64;
377 else if (RetVT == MVT::f80)
378 return SINTTOFP_I32_F80;
379 else if (RetVT == MVT::ppcf128)
380 return SINTTOFP_I32_PPCF128;
381 } else if (OpVT == MVT::i64) {
382 if (RetVT == MVT::f32)
383 return SINTTOFP_I64_F32;
384 else if (RetVT == MVT::f64)
385 return SINTTOFP_I64_F64;
386 else if (RetVT == MVT::f80)
387 return SINTTOFP_I64_F80;
388 else if (RetVT == MVT::ppcf128)
389 return SINTTOFP_I64_PPCF128;
390 } else if (OpVT == MVT::i128) {
391 if (RetVT == MVT::f32)
392 return SINTTOFP_I128_F32;
393 else if (RetVT == MVT::f64)
394 return SINTTOFP_I128_F64;
395 else if (RetVT == MVT::f80)
396 return SINTTOFP_I128_F80;
397 else if (RetVT == MVT::ppcf128)
398 return SINTTOFP_I128_PPCF128;
400 return UNKNOWN_LIBCALL;
403 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or
404 /// UNKNOWN_LIBCALL if there is none.
405 RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) {
406 if (OpVT == MVT::i32) {
407 if (RetVT == MVT::f32)
408 return UINTTOFP_I32_F32;
409 else if (RetVT == MVT::f64)
410 return UINTTOFP_I32_F64;
411 else if (RetVT == MVT::f80)
412 return UINTTOFP_I32_F80;
413 else if (RetVT == MVT::ppcf128)
414 return UINTTOFP_I32_PPCF128;
415 } else if (OpVT == MVT::i64) {
416 if (RetVT == MVT::f32)
417 return UINTTOFP_I64_F32;
418 else if (RetVT == MVT::f64)
419 return UINTTOFP_I64_F64;
420 else if (RetVT == MVT::f80)
421 return UINTTOFP_I64_F80;
422 else if (RetVT == MVT::ppcf128)
423 return UINTTOFP_I64_PPCF128;
424 } else if (OpVT == MVT::i128) {
425 if (RetVT == MVT::f32)
426 return UINTTOFP_I128_F32;
427 else if (RetVT == MVT::f64)
428 return UINTTOFP_I128_F64;
429 else if (RetVT == MVT::f80)
430 return UINTTOFP_I128_F80;
431 else if (RetVT == MVT::ppcf128)
432 return UINTTOFP_I128_PPCF128;
434 return UNKNOWN_LIBCALL;
437 /// InitCmpLibcallCCs - Set default comparison libcall CC.
439 static void InitCmpLibcallCCs(ISD::CondCode *CCs) {
440 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL);
441 CCs[RTLIB::OEQ_F32] = ISD::SETEQ;
442 CCs[RTLIB::OEQ_F64] = ISD::SETEQ;
443 CCs[RTLIB::UNE_F32] = ISD::SETNE;
444 CCs[RTLIB::UNE_F64] = ISD::SETNE;
445 CCs[RTLIB::OGE_F32] = ISD::SETGE;
446 CCs[RTLIB::OGE_F64] = ISD::SETGE;
447 CCs[RTLIB::OLT_F32] = ISD::SETLT;
448 CCs[RTLIB::OLT_F64] = ISD::SETLT;
449 CCs[RTLIB::OLE_F32] = ISD::SETLE;
450 CCs[RTLIB::OLE_F64] = ISD::SETLE;
451 CCs[RTLIB::OGT_F32] = ISD::SETGT;
452 CCs[RTLIB::OGT_F64] = ISD::SETGT;
453 CCs[RTLIB::UO_F32] = ISD::SETNE;
454 CCs[RTLIB::UO_F64] = ISD::SETNE;
455 CCs[RTLIB::O_F32] = ISD::SETEQ;
456 CCs[RTLIB::O_F64] = ISD::SETEQ;
459 /// NOTE: The constructor takes ownership of TLOF.
460 TargetLowering::TargetLowering(TargetMachine &tm,TargetLoweringObjectFile *tlof)
461 : TM(tm), TD(TM.getTargetData()), TLOF(*tlof) {
462 // All operations default to being supported.
463 memset(OpActions, 0, sizeof(OpActions));
464 memset(LoadExtActions, 0, sizeof(LoadExtActions));
465 memset(TruncStoreActions, 0, sizeof(TruncStoreActions));
466 memset(IndexedModeActions, 0, sizeof(IndexedModeActions));
467 memset(ConvertActions, 0, sizeof(ConvertActions));
468 memset(CondCodeActions, 0, sizeof(CondCodeActions));
470 // Set default actions for various operations.
471 for (unsigned VT = 0; VT != (unsigned)MVT::LAST_VALUETYPE; ++VT) {
472 // Default all indexed load / store to expand.
473 for (unsigned IM = (unsigned)ISD::PRE_INC;
474 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
475 setIndexedLoadAction(IM, (MVT::SimpleValueType)VT, Expand);
476 setIndexedStoreAction(IM, (MVT::SimpleValueType)VT, Expand);
479 // These operations default to expand.
480 setOperationAction(ISD::FGETSIGN, (MVT::SimpleValueType)VT, Expand);
481 setOperationAction(ISD::CONCAT_VECTORS, (MVT::SimpleValueType)VT, Expand);
484 // Most targets ignore the @llvm.prefetch intrinsic.
485 setOperationAction(ISD::PREFETCH, MVT::Other, Expand);
487 // ConstantFP nodes default to expand. Targets can either change this to
488 // Legal, in which case all fp constants are legal, or use isFPImmLegal()
489 // to optimize expansions for certain constants.
490 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
491 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
492 setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
494 // These library functions default to expand.
495 setOperationAction(ISD::FLOG , MVT::f64, Expand);
496 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
497 setOperationAction(ISD::FLOG10,MVT::f64, Expand);
498 setOperationAction(ISD::FEXP , MVT::f64, Expand);
499 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
500 setOperationAction(ISD::FLOG , MVT::f32, Expand);
501 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
502 setOperationAction(ISD::FLOG10,MVT::f32, Expand);
503 setOperationAction(ISD::FEXP , MVT::f32, Expand);
504 setOperationAction(ISD::FEXP2, MVT::f32, Expand);
506 // Default ISD::TRAP to expand (which turns it into abort).
507 setOperationAction(ISD::TRAP, MVT::Other, Expand);
509 IsLittleEndian = TD->isLittleEndian();
510 UsesGlobalOffsetTable = false;
511 ShiftAmountTy = PointerTy = MVT::getIntegerVT(8*TD->getPointerSize());
512 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE*sizeof(TargetRegisterClass*));
513 memset(TargetDAGCombineArray, 0, array_lengthof(TargetDAGCombineArray));
514 maxStoresPerMemset = maxStoresPerMemcpy = maxStoresPerMemmove = 8;
515 benefitFromCodePlacementOpt = false;
516 UseUnderscoreSetJmp = false;
517 UseUnderscoreLongJmp = false;
518 SelectIsExpensive = false;
519 IntDivIsCheap = false;
520 Pow2DivIsCheap = false;
521 StackPointerRegisterToSaveRestore = 0;
522 ExceptionPointerRegister = 0;
523 ExceptionSelectorRegister = 0;
524 BooleanContents = UndefinedBooleanContent;
525 SchedPreferenceInfo = SchedulingForLatency;
527 JumpBufAlignment = 0;
528 IfCvtBlockSizeLimit = 2;
529 IfCvtDupBlockSizeLimit = 0;
530 PrefLoopAlignment = 0;
532 InitLibcallNames(LibcallRoutineNames);
533 InitCmpLibcallCCs(CmpLibcallCCs);
534 InitLibcallCallingConvs(LibcallCallingConvs);
537 TargetLowering::~TargetLowering() {
541 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT,
542 unsigned &NumIntermediates,
544 TargetLowering* TLI) {
545 // Figure out the right, legal destination reg to copy into.
546 unsigned NumElts = VT.getVectorNumElements();
547 MVT EltTy = VT.getVectorElementType();
549 unsigned NumVectorRegs = 1;
551 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
552 // could break down into LHS/RHS like LegalizeDAG does.
553 if (!isPowerOf2_32(NumElts)) {
554 NumVectorRegs = NumElts;
558 // Divide the input until we get to a supported size. This will always
559 // end with a scalar if the target doesn't support vectors.
560 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) {
565 NumIntermediates = NumVectorRegs;
567 MVT NewVT = MVT::getVectorVT(EltTy, NumElts);
568 if (!TLI->isTypeLegal(NewVT))
570 IntermediateVT = NewVT;
572 EVT DestVT = TLI->getRegisterType(NewVT);
574 if (EVT(DestVT).bitsLT(NewVT)) {
575 // Value is expanded, e.g. i64 -> i16.
576 return NumVectorRegs*(NewVT.getSizeInBits()/DestVT.getSizeInBits());
578 // Otherwise, promotion or legal types use the same number of registers as
579 // the vector decimated to the appropriate level.
580 return NumVectorRegs;
586 /// computeRegisterProperties - Once all of the register classes are added,
587 /// this allows us to compute derived properties we expose.
588 void TargetLowering::computeRegisterProperties() {
589 assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE &&
590 "Too many value types for ValueTypeActions to hold!");
592 // Everything defaults to needing one register.
593 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
594 NumRegistersForVT[i] = 1;
595 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i;
597 // ...except isVoid, which doesn't need any registers.
598 NumRegistersForVT[MVT::isVoid] = 0;
600 // Find the largest integer register class.
601 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE;
602 for (; RegClassForVT[LargestIntReg] == 0; --LargestIntReg)
603 assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
605 // Every integer value type larger than this largest register takes twice as
606 // many registers to represent as the previous ValueType.
607 for (unsigned ExpandedReg = LargestIntReg + 1; ; ++ExpandedReg) {
608 EVT ExpandedVT = (MVT::SimpleValueType)ExpandedReg;
609 if (!ExpandedVT.isInteger())
611 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
612 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg;
613 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1);
614 ValueTypeActions.setTypeAction(ExpandedVT, Expand);
617 // Inspect all of the ValueType's smaller than the largest integer
618 // register to see which ones need promotion.
619 unsigned LegalIntReg = LargestIntReg;
620 for (unsigned IntReg = LargestIntReg - 1;
621 IntReg >= (unsigned)MVT::i1; --IntReg) {
622 EVT IVT = (MVT::SimpleValueType)IntReg;
623 if (isTypeLegal(IVT)) {
624 LegalIntReg = IntReg;
626 RegisterTypeForVT[IntReg] = TransformToType[IntReg] =
627 (MVT::SimpleValueType)LegalIntReg;
628 ValueTypeActions.setTypeAction(IVT, Promote);
632 // ppcf128 type is really two f64's.
633 if (!isTypeLegal(MVT::ppcf128)) {
634 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64];
635 RegisterTypeForVT[MVT::ppcf128] = MVT::f64;
636 TransformToType[MVT::ppcf128] = MVT::f64;
637 ValueTypeActions.setTypeAction(MVT::ppcf128, Expand);
640 // Decide how to handle f64. If the target does not have native f64 support,
641 // expand it to i64 and we will be generating soft float library calls.
642 if (!isTypeLegal(MVT::f64)) {
643 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
644 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
645 TransformToType[MVT::f64] = MVT::i64;
646 ValueTypeActions.setTypeAction(MVT::f64, Expand);
649 // Decide how to handle f32. If the target does not have native support for
650 // f32, promote it to f64 if it is legal. Otherwise, expand it to i32.
651 if (!isTypeLegal(MVT::f32)) {
652 if (isTypeLegal(MVT::f64)) {
653 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::f64];
654 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::f64];
655 TransformToType[MVT::f32] = MVT::f64;
656 ValueTypeActions.setTypeAction(MVT::f32, Promote);
658 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
659 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
660 TransformToType[MVT::f32] = MVT::i32;
661 ValueTypeActions.setTypeAction(MVT::f32, Expand);
665 // Loop over all of the vector value types to see which need transformations.
666 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
667 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
668 MVT VT = (MVT::SimpleValueType)i;
669 if (!isTypeLegal(VT)) {
672 unsigned NumIntermediates;
673 NumRegistersForVT[i] =
674 getVectorTypeBreakdownMVT(VT, IntermediateVT, NumIntermediates,
676 RegisterTypeForVT[i] = RegisterVT;
678 // Determine if there is a legal wider type.
679 bool IsLegalWiderType = false;
680 EVT EltVT = VT.getVectorElementType();
681 unsigned NElts = VT.getVectorNumElements();
682 for (unsigned nVT = i+1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
683 EVT SVT = (MVT::SimpleValueType)nVT;
684 if (isTypeLegal(SVT) && SVT.getVectorElementType() == EltVT &&
685 SVT.getVectorNumElements() > NElts) {
686 TransformToType[i] = SVT;
687 ValueTypeActions.setTypeAction(VT, Promote);
688 IsLegalWiderType = true;
692 if (!IsLegalWiderType) {
693 EVT NVT = VT.getPow2VectorType();
695 // Type is already a power of 2. The default action is to split.
696 TransformToType[i] = MVT::Other;
697 ValueTypeActions.setTypeAction(VT, Expand);
699 TransformToType[i] = NVT;
700 ValueTypeActions.setTypeAction(VT, Promote);
707 const char *TargetLowering::getTargetNodeName(unsigned Opcode) const {
712 MVT::SimpleValueType TargetLowering::getSetCCResultType(EVT VT) const {
713 return PointerTy.SimpleTy;
716 /// getVectorTypeBreakdown - Vector types are broken down into some number of
717 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
718 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
719 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
721 /// This method returns the number of registers needed, and the VT for each
722 /// register. It also returns the VT and quantity of the intermediate values
723 /// before they are promoted/expanded.
725 unsigned TargetLowering::getVectorTypeBreakdown(LLVMContext &Context, EVT VT,
727 unsigned &NumIntermediates,
728 EVT &RegisterVT) const {
729 // Figure out the right, legal destination reg to copy into.
730 unsigned NumElts = VT.getVectorNumElements();
731 EVT EltTy = VT.getVectorElementType();
733 unsigned NumVectorRegs = 1;
735 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
736 // could break down into LHS/RHS like LegalizeDAG does.
737 if (!isPowerOf2_32(NumElts)) {
738 NumVectorRegs = NumElts;
742 // Divide the input until we get to a supported size. This will always
743 // end with a scalar if the target doesn't support vectors.
744 while (NumElts > 1 && !isTypeLegal(
745 EVT::getVectorVT(Context, EltTy, NumElts))) {
750 NumIntermediates = NumVectorRegs;
752 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts);
753 if (!isTypeLegal(NewVT))
755 IntermediateVT = NewVT;
757 EVT DestVT = getRegisterType(Context, NewVT);
759 if (DestVT.bitsLT(NewVT)) {
760 // Value is expanded, e.g. i64 -> i16.
761 return NumVectorRegs*(NewVT.getSizeInBits()/DestVT.getSizeInBits());
763 // Otherwise, promotion or legal types use the same number of registers as
764 // the vector decimated to the appropriate level.
765 return NumVectorRegs;
771 /// getWidenVectorType: given a vector type, returns the type to widen to
772 /// (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
773 /// If there is no vector type that we want to widen to, returns MVT::Other
774 /// When and where to widen is target dependent based on the cost of
775 /// scalarizing vs using the wider vector type.
776 EVT TargetLowering::getWidenVectorType(EVT VT) const {
777 assert(VT.isVector());
781 // Default is not to widen until moved to LegalizeTypes
785 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
786 /// function arguments in the caller parameter area. This is the actual
787 /// alignment, not its logarithm.
788 unsigned TargetLowering::getByValTypeAlignment(const Type *Ty) const {
789 return TD->getCallFrameTypeAlignment(Ty);
792 SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table,
793 SelectionDAG &DAG) const {
794 if (usesGlobalOffsetTable())
795 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy());
800 TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
801 // Assume that everything is safe in static mode.
802 if (getTargetMachine().getRelocationModel() == Reloc::Static)
805 // In dynamic-no-pic mode, assume that known defined values are safe.
806 if (getTargetMachine().getRelocationModel() == Reloc::DynamicNoPIC &&
808 !GA->getGlobal()->isDeclaration() &&
809 !GA->getGlobal()->isWeakForLinker())
812 // Otherwise assume nothing is safe.
816 //===----------------------------------------------------------------------===//
817 // Optimization Methods
818 //===----------------------------------------------------------------------===//
820 /// ShrinkDemandedConstant - Check to see if the specified operand of the
821 /// specified instruction is a constant integer. If so, check to see if there
822 /// are any bits set in the constant that are not demanded. If so, shrink the
823 /// constant and return true.
824 bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDValue Op,
825 const APInt &Demanded) {
826 DebugLoc dl = Op.getDebugLoc();
828 // FIXME: ISD::SELECT, ISD::SELECT_CC
829 switch (Op.getOpcode()) {
834 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
835 if (!C) return false;
837 if (Op.getOpcode() == ISD::XOR &&
838 (C->getAPIntValue() | (~Demanded)).isAllOnesValue())
841 // if we can expand it to have all bits set, do it
842 if (C->getAPIntValue().intersects(~Demanded)) {
843 EVT VT = Op.getValueType();
844 SDValue New = DAG.getNode(Op.getOpcode(), dl, VT, Op.getOperand(0),
845 DAG.getConstant(Demanded &
848 return CombineTo(Op, New);
858 /// ShrinkDemandedOp - Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the
859 /// casts are free. This uses isZExtFree and ZERO_EXTEND for the widening
860 /// cast, but it could be generalized for targets with other types of
861 /// implicit widening casts.
863 TargetLowering::TargetLoweringOpt::ShrinkDemandedOp(SDValue Op,
865 const APInt &Demanded,
867 assert(Op.getNumOperands() == 2 &&
868 "ShrinkDemandedOp only supports binary operators!");
869 assert(Op.getNode()->getNumValues() == 1 &&
870 "ShrinkDemandedOp only supports nodes with one result!");
872 // Don't do this if the node has another user, which may require the
874 if (!Op.getNode()->hasOneUse())
877 // Search for the smallest integer type with free casts to and from
878 // Op's type. For expedience, just check power-of-2 integer types.
879 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
880 unsigned SmallVTBits = BitWidth - Demanded.countLeadingZeros();
881 if (!isPowerOf2_32(SmallVTBits))
882 SmallVTBits = NextPowerOf2(SmallVTBits);
883 for (; SmallVTBits < BitWidth; SmallVTBits = NextPowerOf2(SmallVTBits)) {
884 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), SmallVTBits);
885 if (TLI.isTruncateFree(Op.getValueType(), SmallVT) &&
886 TLI.isZExtFree(SmallVT, Op.getValueType())) {
887 // We found a type with free casts.
888 SDValue X = DAG.getNode(Op.getOpcode(), dl, SmallVT,
889 DAG.getNode(ISD::TRUNCATE, dl, SmallVT,
890 Op.getNode()->getOperand(0)),
891 DAG.getNode(ISD::TRUNCATE, dl, SmallVT,
892 Op.getNode()->getOperand(1)));
893 SDValue Z = DAG.getNode(ISD::ZERO_EXTEND, dl, Op.getValueType(), X);
894 return CombineTo(Op, Z);
900 /// SimplifyDemandedBits - Look at Op. At this point, we know that only the
901 /// DemandedMask bits of the result of Op are ever used downstream. If we can
902 /// use this information to simplify Op, create a new simplified DAG node and
903 /// return true, returning the original and new nodes in Old and New. Otherwise,
904 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
905 /// the expression (used to simplify the caller). The KnownZero/One bits may
906 /// only be accurate for those bits in the DemandedMask.
907 bool TargetLowering::SimplifyDemandedBits(SDValue Op,
908 const APInt &DemandedMask,
911 TargetLoweringOpt &TLO,
912 unsigned Depth) const {
913 unsigned BitWidth = DemandedMask.getBitWidth();
914 assert(Op.getValueType().getScalarType().getSizeInBits() == BitWidth &&
915 "Mask size mismatches value type size!");
916 APInt NewMask = DemandedMask;
917 DebugLoc dl = Op.getDebugLoc();
919 // Don't know anything.
920 KnownZero = KnownOne = APInt(BitWidth, 0);
922 // Other users may use these bits.
923 if (!Op.getNode()->hasOneUse()) {
925 // If not at the root, Just compute the KnownZero/KnownOne bits to
926 // simplify things downstream.
927 TLO.DAG.ComputeMaskedBits(Op, DemandedMask, KnownZero, KnownOne, Depth);
930 // If this is the root being simplified, allow it to have multiple uses,
931 // just set the NewMask to all bits.
932 NewMask = APInt::getAllOnesValue(BitWidth);
933 } else if (DemandedMask == 0) {
934 // Not demanding any bits from Op.
935 if (Op.getOpcode() != ISD::UNDEF)
936 return TLO.CombineTo(Op, TLO.DAG.getUNDEF(Op.getValueType()));
938 } else if (Depth == 6) { // Limit search depth.
942 APInt KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut;
943 switch (Op.getOpcode()) {
945 // We know all of the bits for a constant!
946 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue() & NewMask;
947 KnownZero = ~KnownOne & NewMask;
948 return false; // Don't fall through, will infinitely loop.
950 // If the RHS is a constant, check to see if the LHS would be zero without
951 // using the bits from the RHS. Below, we use knowledge about the RHS to
952 // simplify the LHS, here we're using information from the LHS to simplify
954 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
955 APInt LHSZero, LHSOne;
956 TLO.DAG.ComputeMaskedBits(Op.getOperand(0), NewMask,
957 LHSZero, LHSOne, Depth+1);
958 // If the LHS already has zeros where RHSC does, this and is dead.
959 if ((LHSZero & NewMask) == (~RHSC->getAPIntValue() & NewMask))
960 return TLO.CombineTo(Op, Op.getOperand(0));
961 // If any of the set bits in the RHS are known zero on the LHS, shrink
963 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & NewMask))
967 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
968 KnownOne, TLO, Depth+1))
970 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
971 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownZero & NewMask,
972 KnownZero2, KnownOne2, TLO, Depth+1))
974 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
976 // If all of the demanded bits are known one on one side, return the other.
977 // These bits cannot contribute to the result of the 'and'.
978 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
979 return TLO.CombineTo(Op, Op.getOperand(0));
980 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
981 return TLO.CombineTo(Op, Op.getOperand(1));
982 // If all of the demanded bits in the inputs are known zeros, return zero.
983 if ((NewMask & (KnownZero|KnownZero2)) == NewMask)
984 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, Op.getValueType()));
985 // If the RHS is a constant, see if we can simplify it.
986 if (TLO.ShrinkDemandedConstant(Op, ~KnownZero2 & NewMask))
988 // If the operation can be done in a smaller type, do so.
989 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
992 // Output known-1 bits are only known if set in both the LHS & RHS.
993 KnownOne &= KnownOne2;
994 // Output known-0 are known to be clear if zero in either the LHS | RHS.
995 KnownZero |= KnownZero2;
998 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
999 KnownOne, TLO, Depth+1))
1001 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1002 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownOne & NewMask,
1003 KnownZero2, KnownOne2, TLO, Depth+1))
1005 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1007 // If all of the demanded bits are known zero on one side, return the other.
1008 // These bits cannot contribute to the result of the 'or'.
1009 if ((NewMask & ~KnownOne2 & KnownZero) == (~KnownOne2 & NewMask))
1010 return TLO.CombineTo(Op, Op.getOperand(0));
1011 if ((NewMask & ~KnownOne & KnownZero2) == (~KnownOne & NewMask))
1012 return TLO.CombineTo(Op, Op.getOperand(1));
1013 // If all of the potentially set bits on one side are known to be set on
1014 // the other side, just use the 'other' side.
1015 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
1016 return TLO.CombineTo(Op, Op.getOperand(0));
1017 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
1018 return TLO.CombineTo(Op, Op.getOperand(1));
1019 // If the RHS is a constant, see if we can simplify it.
1020 if (TLO.ShrinkDemandedConstant(Op, NewMask))
1022 // If the operation can be done in a smaller type, do so.
1023 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1026 // Output known-0 bits are only known if clear in both the LHS & RHS.
1027 KnownZero &= KnownZero2;
1028 // Output known-1 are known to be set if set in either the LHS | RHS.
1029 KnownOne |= KnownOne2;
1032 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
1033 KnownOne, TLO, Depth+1))
1035 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1036 if (SimplifyDemandedBits(Op.getOperand(0), NewMask, KnownZero2,
1037 KnownOne2, TLO, Depth+1))
1039 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1041 // If all of the demanded bits are known zero on one side, return the other.
1042 // These bits cannot contribute to the result of the 'xor'.
1043 if ((KnownZero & NewMask) == NewMask)
1044 return TLO.CombineTo(Op, Op.getOperand(0));
1045 if ((KnownZero2 & NewMask) == NewMask)
1046 return TLO.CombineTo(Op, Op.getOperand(1));
1047 // If the operation can be done in a smaller type, do so.
1048 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1051 // If all of the unknown bits are known to be zero on one side or the other
1052 // (but not both) turn this into an *inclusive* or.
1053 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1054 if ((NewMask & ~KnownZero & ~KnownZero2) == 0)
1055 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, dl, Op.getValueType(),
1059 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1060 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
1061 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1062 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
1064 // If all of the demanded bits on one side are known, and all of the set
1065 // bits on that side are also known to be set on the other side, turn this
1066 // into an AND, as we know the bits will be cleared.
1067 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1068 if ((NewMask & (KnownZero|KnownOne)) == NewMask) { // all known
1069 if ((KnownOne & KnownOne2) == KnownOne) {
1070 EVT VT = Op.getValueType();
1071 SDValue ANDC = TLO.DAG.getConstant(~KnownOne & NewMask, VT);
1072 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, dl, VT,
1073 Op.getOperand(0), ANDC));
1077 // If the RHS is a constant, see if we can simplify it.
1078 // for XOR, we prefer to force bits to 1 if they will make a -1.
1079 // if we can't force bits, try to shrink constant
1080 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1081 APInt Expanded = C->getAPIntValue() | (~NewMask);
1082 // if we can expand it to have all bits set, do it
1083 if (Expanded.isAllOnesValue()) {
1084 if (Expanded != C->getAPIntValue()) {
1085 EVT VT = Op.getValueType();
1086 SDValue New = TLO.DAG.getNode(Op.getOpcode(), dl,VT, Op.getOperand(0),
1087 TLO.DAG.getConstant(Expanded, VT));
1088 return TLO.CombineTo(Op, New);
1090 // if it already has all the bits set, nothing to change
1091 // but don't shrink either!
1092 } else if (TLO.ShrinkDemandedConstant(Op, NewMask)) {
1097 KnownZero = KnownZeroOut;
1098 KnownOne = KnownOneOut;
1101 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero,
1102 KnownOne, TLO, Depth+1))
1104 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero2,
1105 KnownOne2, TLO, Depth+1))
1107 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1108 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1110 // If the operands are constants, see if we can simplify them.
1111 if (TLO.ShrinkDemandedConstant(Op, NewMask))
1114 // Only known if known in both the LHS and RHS.
1115 KnownOne &= KnownOne2;
1116 KnownZero &= KnownZero2;
1118 case ISD::SELECT_CC:
1119 if (SimplifyDemandedBits(Op.getOperand(3), NewMask, KnownZero,
1120 KnownOne, TLO, Depth+1))
1122 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero2,
1123 KnownOne2, TLO, Depth+1))
1125 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1126 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1128 // If the operands are constants, see if we can simplify them.
1129 if (TLO.ShrinkDemandedConstant(Op, NewMask))
1132 // Only known if known in both the LHS and RHS.
1133 KnownOne &= KnownOne2;
1134 KnownZero &= KnownZero2;
1137 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1138 unsigned ShAmt = SA->getZExtValue();
1139 SDValue InOp = Op.getOperand(0);
1141 // If the shift count is an invalid immediate, don't do anything.
1142 if (ShAmt >= BitWidth)
1145 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a
1146 // single shift. We can do this if the bottom bits (which are shifted
1147 // out) are never demanded.
1148 if (InOp.getOpcode() == ISD::SRL &&
1149 isa<ConstantSDNode>(InOp.getOperand(1))) {
1150 if (ShAmt && (NewMask & APInt::getLowBitsSet(BitWidth, ShAmt)) == 0) {
1151 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
1152 unsigned Opc = ISD::SHL;
1153 int Diff = ShAmt-C1;
1160 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
1161 EVT VT = Op.getValueType();
1162 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT,
1163 InOp.getOperand(0), NewSA));
1167 if (SimplifyDemandedBits(Op.getOperand(0), NewMask.lshr(ShAmt),
1168 KnownZero, KnownOne, TLO, Depth+1))
1170 KnownZero <<= SA->getZExtValue();
1171 KnownOne <<= SA->getZExtValue();
1172 // low bits known zero.
1173 KnownZero |= APInt::getLowBitsSet(BitWidth, SA->getZExtValue());
1177 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1178 EVT VT = Op.getValueType();
1179 unsigned ShAmt = SA->getZExtValue();
1180 unsigned VTSize = VT.getSizeInBits();
1181 SDValue InOp = Op.getOperand(0);
1183 // If the shift count is an invalid immediate, don't do anything.
1184 if (ShAmt >= BitWidth)
1187 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a
1188 // single shift. We can do this if the top bits (which are shifted out)
1189 // are never demanded.
1190 if (InOp.getOpcode() == ISD::SHL &&
1191 isa<ConstantSDNode>(InOp.getOperand(1))) {
1192 if (ShAmt && (NewMask & APInt::getHighBitsSet(VTSize, ShAmt)) == 0) {
1193 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
1194 unsigned Opc = ISD::SRL;
1195 int Diff = ShAmt-C1;
1202 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
1203 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT,
1204 InOp.getOperand(0), NewSA));
1208 // Compute the new bits that are at the top now.
1209 if (SimplifyDemandedBits(InOp, (NewMask << ShAmt),
1210 KnownZero, KnownOne, TLO, Depth+1))
1212 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1213 KnownZero = KnownZero.lshr(ShAmt);
1214 KnownOne = KnownOne.lshr(ShAmt);
1216 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
1217 KnownZero |= HighBits; // High bits known zero.
1221 // If this is an arithmetic shift right and only the low-bit is set, we can
1222 // always convert this into a logical shr, even if the shift amount is
1223 // variable. The low bit of the shift cannot be an input sign bit unless
1224 // the shift amount is >= the size of the datatype, which is undefined.
1225 if (DemandedMask == 1)
1226 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, Op.getValueType(),
1227 Op.getOperand(0), Op.getOperand(1)));
1229 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1230 EVT VT = Op.getValueType();
1231 unsigned ShAmt = SA->getZExtValue();
1233 // If the shift count is an invalid immediate, don't do anything.
1234 if (ShAmt >= BitWidth)
1237 APInt InDemandedMask = (NewMask << ShAmt);
1239 // If any of the demanded bits are produced by the sign extension, we also
1240 // demand the input sign bit.
1241 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
1242 if (HighBits.intersects(NewMask))
1243 InDemandedMask |= APInt::getSignBit(VT.getScalarType().getSizeInBits());
1245 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask,
1246 KnownZero, KnownOne, TLO, Depth+1))
1248 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1249 KnownZero = KnownZero.lshr(ShAmt);
1250 KnownOne = KnownOne.lshr(ShAmt);
1252 // Handle the sign bit, adjusted to where it is now in the mask.
1253 APInt SignBit = APInt::getSignBit(BitWidth).lshr(ShAmt);
1255 // If the input sign bit is known to be zero, or if none of the top bits
1256 // are demanded, turn this into an unsigned shift right.
1257 if (KnownZero.intersects(SignBit) || (HighBits & ~NewMask) == HighBits) {
1258 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, VT,
1261 } else if (KnownOne.intersects(SignBit)) { // New bits are known one.
1262 KnownOne |= HighBits;
1266 case ISD::SIGN_EXTEND_INREG: {
1267 EVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1269 // Sign extension. Compute the demanded bits in the result that are not
1270 // present in the input.
1271 APInt NewBits = APInt::getHighBitsSet(BitWidth,
1272 BitWidth - EVT.getSizeInBits()) &
1275 // If none of the extended bits are demanded, eliminate the sextinreg.
1277 return TLO.CombineTo(Op, Op.getOperand(0));
1279 APInt InSignBit = APInt::getSignBit(EVT.getSizeInBits());
1280 InSignBit.zext(BitWidth);
1281 APInt InputDemandedBits = APInt::getLowBitsSet(BitWidth,
1282 EVT.getSizeInBits()) &
1285 // Since the sign extended bits are demanded, we know that the sign
1287 InputDemandedBits |= InSignBit;
1289 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits,
1290 KnownZero, KnownOne, TLO, Depth+1))
1292 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1294 // If the sign bit of the input is known set or clear, then we know the
1295 // top bits of the result.
1297 // If the input sign bit is known zero, convert this into a zero extension.
1298 if (KnownZero.intersects(InSignBit))
1299 return TLO.CombineTo(Op,
1300 TLO.DAG.getZeroExtendInReg(Op.getOperand(0),dl,EVT));
1302 if (KnownOne.intersects(InSignBit)) { // Input sign bit known set
1303 KnownOne |= NewBits;
1304 KnownZero &= ~NewBits;
1305 } else { // Input sign bit unknown
1306 KnownZero &= ~NewBits;
1307 KnownOne &= ~NewBits;
1311 case ISD::ZERO_EXTEND: {
1312 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits();
1313 APInt InMask = NewMask;
1314 InMask.trunc(OperandBitWidth);
1316 // If none of the top bits are demanded, convert this into an any_extend.
1318 APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask;
1319 if (!NewBits.intersects(NewMask))
1320 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND, dl,
1324 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1325 KnownZero, KnownOne, TLO, Depth+1))
1327 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1328 KnownZero.zext(BitWidth);
1329 KnownOne.zext(BitWidth);
1330 KnownZero |= NewBits;
1333 case ISD::SIGN_EXTEND: {
1334 EVT InVT = Op.getOperand(0).getValueType();
1335 unsigned InBits = InVT.getSizeInBits();
1336 APInt InMask = APInt::getLowBitsSet(BitWidth, InBits);
1337 APInt InSignBit = APInt::getBitsSet(BitWidth, InBits - 1, InBits);
1338 APInt NewBits = ~InMask & NewMask;
1340 // If none of the top bits are demanded, convert this into an any_extend.
1342 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND, dl,
1346 // Since some of the sign extended bits are demanded, we know that the sign
1348 APInt InDemandedBits = InMask & NewMask;
1349 InDemandedBits |= InSignBit;
1350 InDemandedBits.trunc(InBits);
1352 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero,
1353 KnownOne, TLO, Depth+1))
1355 KnownZero.zext(BitWidth);
1356 KnownOne.zext(BitWidth);
1358 // If the sign bit is known zero, convert this to a zero extend.
1359 if (KnownZero.intersects(InSignBit))
1360 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND, dl,
1364 // If the sign bit is known one, the top bits match.
1365 if (KnownOne.intersects(InSignBit)) {
1366 KnownOne |= NewBits;
1367 KnownZero &= ~NewBits;
1368 } else { // Otherwise, top bits aren't known.
1369 KnownOne &= ~NewBits;
1370 KnownZero &= ~NewBits;
1374 case ISD::ANY_EXTEND: {
1375 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits();
1376 APInt InMask = NewMask;
1377 InMask.trunc(OperandBitWidth);
1378 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1379 KnownZero, KnownOne, TLO, Depth+1))
1381 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1382 KnownZero.zext(BitWidth);
1383 KnownOne.zext(BitWidth);
1386 case ISD::TRUNCATE: {
1387 // Simplify the input, using demanded bit information, and compute the known
1388 // zero/one bits live out.
1389 APInt TruncMask = NewMask;
1390 TruncMask.zext(Op.getOperand(0).getValueSizeInBits());
1391 if (SimplifyDemandedBits(Op.getOperand(0), TruncMask,
1392 KnownZero, KnownOne, TLO, Depth+1))
1394 KnownZero.trunc(BitWidth);
1395 KnownOne.trunc(BitWidth);
1397 // If the input is only used by this truncate, see if we can shrink it based
1398 // on the known demanded bits.
1399 if (Op.getOperand(0).getNode()->hasOneUse()) {
1400 SDValue In = Op.getOperand(0);
1401 unsigned InBitWidth = In.getValueSizeInBits();
1402 switch (In.getOpcode()) {
1405 // Shrink SRL by a constant if none of the high bits shifted in are
1407 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1))){
1408 APInt HighBits = APInt::getHighBitsSet(InBitWidth,
1409 InBitWidth - BitWidth);
1410 HighBits = HighBits.lshr(ShAmt->getZExtValue());
1411 HighBits.trunc(BitWidth);
1413 if (ShAmt->getZExtValue() < BitWidth && !(HighBits & NewMask)) {
1414 // None of the shifted in bits are needed. Add a truncate of the
1415 // shift input, then shift it.
1416 SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE, dl,
1419 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl,
1429 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1432 case ISD::AssertZext: {
1433 EVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1434 APInt InMask = APInt::getLowBitsSet(BitWidth,
1435 VT.getSizeInBits());
1436 if (SimplifyDemandedBits(Op.getOperand(0), InMask & NewMask,
1437 KnownZero, KnownOne, TLO, Depth+1))
1439 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1440 KnownZero |= ~InMask & NewMask;
1443 case ISD::BIT_CONVERT:
1445 // If this is an FP->Int bitcast and if the sign bit is the only thing that
1446 // is demanded, turn this into a FGETSIGN.
1447 if (NewMask == EVT::getIntegerVTSignBit(Op.getValueType()) &&
1448 MVT::isFloatingPoint(Op.getOperand(0).getValueType()) &&
1449 !MVT::isVector(Op.getOperand(0).getValueType())) {
1450 // Only do this xform if FGETSIGN is valid or if before legalize.
1451 if (!TLO.AfterLegalize ||
1452 isOperationLegal(ISD::FGETSIGN, Op.getValueType())) {
1453 // Make a FGETSIGN + SHL to move the sign bit into the appropriate
1454 // place. We expect the SHL to be eliminated by other optimizations.
1455 SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, Op.getValueType(),
1457 unsigned ShVal = Op.getValueType().getSizeInBits()-1;
1458 SDValue ShAmt = TLO.DAG.getConstant(ShVal, getShiftAmountTy());
1459 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, Op.getValueType(),
1468 // Add, Sub, and Mul don't demand any bits in positions beyond that
1469 // of the highest bit demanded of them.
1470 APInt LoMask = APInt::getLowBitsSet(BitWidth,
1471 BitWidth - NewMask.countLeadingZeros());
1472 if (SimplifyDemandedBits(Op.getOperand(0), LoMask, KnownZero2,
1473 KnownOne2, TLO, Depth+1))
1475 if (SimplifyDemandedBits(Op.getOperand(1), LoMask, KnownZero2,
1476 KnownOne2, TLO, Depth+1))
1478 // See if the operation should be performed at a smaller bit width.
1479 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1484 // Just use ComputeMaskedBits to compute output bits.
1485 TLO.DAG.ComputeMaskedBits(Op, NewMask, KnownZero, KnownOne, Depth);
1489 // If we know the value of all of the demanded bits, return this as a
1491 if ((NewMask & (KnownZero|KnownOne)) == NewMask)
1492 return TLO.CombineTo(Op, TLO.DAG.getConstant(KnownOne, Op.getValueType()));
1497 /// computeMaskedBitsForTargetNode - Determine which of the bits specified
1498 /// in Mask are known to be either zero or one and return them in the
1499 /// KnownZero/KnownOne bitsets.
1500 void TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
1504 const SelectionDAG &DAG,
1505 unsigned Depth) const {
1506 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1507 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1508 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1509 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1510 "Should use MaskedValueIsZero if you don't know whether Op"
1511 " is a target node!");
1512 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
1515 /// ComputeNumSignBitsForTargetNode - This method can be implemented by
1516 /// targets that want to expose additional information about sign bits to the
1518 unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
1519 unsigned Depth) const {
1520 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1521 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1522 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1523 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1524 "Should use ComputeNumSignBits if you don't know whether Op"
1525 " is a target node!");
1529 /// ValueHasExactlyOneBitSet - Test if the given value is known to have exactly
1530 /// one bit set. This differs from ComputeMaskedBits in that it doesn't need to
1531 /// determine which bit is set.
1533 static bool ValueHasExactlyOneBitSet(SDValue Val, const SelectionDAG &DAG) {
1534 // A left-shift of a constant one will have exactly one bit set, because
1535 // shifting the bit off the end is undefined.
1536 if (Val.getOpcode() == ISD::SHL)
1537 if (ConstantSDNode *C =
1538 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0)))
1539 if (C->getAPIntValue() == 1)
1542 // Similarly, a right-shift of a constant sign-bit will have exactly
1544 if (Val.getOpcode() == ISD::SRL)
1545 if (ConstantSDNode *C =
1546 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0)))
1547 if (C->getAPIntValue().isSignBit())
1550 // More could be done here, though the above checks are enough
1551 // to handle some common cases.
1553 // Fall back to ComputeMaskedBits to catch other known cases.
1554 EVT OpVT = Val.getValueType();
1555 unsigned BitWidth = OpVT.getSizeInBits();
1556 APInt Mask = APInt::getAllOnesValue(BitWidth);
1557 APInt KnownZero, KnownOne;
1558 DAG.ComputeMaskedBits(Val, Mask, KnownZero, KnownOne);
1559 return (KnownZero.countPopulation() == BitWidth - 1) &&
1560 (KnownOne.countPopulation() == 1);
1563 /// SimplifySetCC - Try to simplify a setcc built with the specified operands
1564 /// and cc. If it is unable to simplify it, return a null SDValue.
1566 TargetLowering::SimplifySetCC(EVT VT, SDValue N0, SDValue N1,
1567 ISD::CondCode Cond, bool foldBooleans,
1568 DAGCombinerInfo &DCI, DebugLoc dl) const {
1569 SelectionDAG &DAG = DCI.DAG;
1570 LLVMContext &Context = *DAG.getContext();
1572 // These setcc operations always fold.
1576 case ISD::SETFALSE2: return DAG.getConstant(0, VT);
1578 case ISD::SETTRUE2: return DAG.getConstant(1, VT);
1581 if (isa<ConstantSDNode>(N0.getNode())) {
1582 // Ensure that the constant occurs on the RHS, and fold constant
1584 return DAG.getSetCC(dl, VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
1587 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) {
1588 const APInt &C1 = N1C->getAPIntValue();
1590 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an
1591 // equality comparison, then we're just comparing whether X itself is
1593 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) &&
1594 N0.getOperand(0).getOpcode() == ISD::CTLZ &&
1595 N0.getOperand(1).getOpcode() == ISD::Constant) {
1596 unsigned ShAmt = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
1597 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1598 ShAmt == Log2_32(N0.getValueType().getSizeInBits())) {
1599 if ((C1 == 0) == (Cond == ISD::SETEQ)) {
1600 // (srl (ctlz x), 5) == 0 -> X != 0
1601 // (srl (ctlz x), 5) != 1 -> X != 0
1604 // (srl (ctlz x), 5) != 0 -> X == 0
1605 // (srl (ctlz x), 5) == 1 -> X == 0
1608 SDValue Zero = DAG.getConstant(0, N0.getValueType());
1609 return DAG.getSetCC(dl, VT, N0.getOperand(0).getOperand(0),
1614 // If the LHS is '(and load, const)', the RHS is 0,
1615 // the test is for equality or unsigned, and all 1 bits of the const are
1616 // in the same partial word, see if we can shorten the load.
1617 if (DCI.isBeforeLegalize() &&
1618 N0.getOpcode() == ISD::AND && C1 == 0 &&
1619 N0.getNode()->hasOneUse() &&
1620 isa<LoadSDNode>(N0.getOperand(0)) &&
1621 N0.getOperand(0).getNode()->hasOneUse() &&
1622 isa<ConstantSDNode>(N0.getOperand(1))) {
1623 LoadSDNode *Lod = cast<LoadSDNode>(N0.getOperand(0));
1624 uint64_t bestMask = 0;
1625 unsigned bestWidth = 0, bestOffset = 0;
1626 if (!Lod->isVolatile() && Lod->isUnindexed() &&
1627 // FIXME: This uses getZExtValue() below so it only works on i64 and
1629 N0.getValueType().getSizeInBits() <= 64) {
1630 unsigned origWidth = N0.getValueType().getSizeInBits();
1631 // We can narrow (e.g.) 16-bit extending loads on 32-bit target to
1632 // 8 bits, but have to be careful...
1633 if (Lod->getExtensionType() != ISD::NON_EXTLOAD)
1634 origWidth = Lod->getMemoryVT().getSizeInBits();
1635 uint64_t Mask =cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
1636 for (unsigned width = origWidth / 2; width>=8; width /= 2) {
1637 uint64_t newMask = (1ULL << width) - 1;
1638 for (unsigned offset=0; offset<origWidth/width; offset++) {
1639 if ((newMask & Mask) == Mask) {
1640 if (!TD->isLittleEndian())
1641 bestOffset = (origWidth/width - offset - 1) * (width/8);
1643 bestOffset = (uint64_t)offset * (width/8);
1644 bestMask = Mask >> (offset * (width/8) * 8);
1648 newMask = newMask << width;
1653 EVT newVT = EVT::getIntegerVT(Context, bestWidth);
1654 if (newVT.isRound()) {
1655 EVT PtrType = Lod->getOperand(1).getValueType();
1656 SDValue Ptr = Lod->getBasePtr();
1657 if (bestOffset != 0)
1658 Ptr = DAG.getNode(ISD::ADD, dl, PtrType, Lod->getBasePtr(),
1659 DAG.getConstant(bestOffset, PtrType));
1660 unsigned NewAlign = MinAlign(Lod->getAlignment(), bestOffset);
1661 SDValue NewLoad = DAG.getLoad(newVT, dl, Lod->getChain(), Ptr,
1663 Lod->getSrcValueOffset() + bestOffset,
1665 return DAG.getSetCC(dl, VT,
1666 DAG.getNode(ISD::AND, dl, newVT, NewLoad,
1667 DAG.getConstant(bestMask, newVT)),
1668 DAG.getConstant(0LL, newVT), Cond);
1673 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
1674 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
1675 unsigned InSize = N0.getOperand(0).getValueType().getSizeInBits();
1677 // If the comparison constant has bits in the upper part, the
1678 // zero-extended value could never match.
1679 if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(),
1680 C1.getBitWidth() - InSize))) {
1684 case ISD::SETEQ: return DAG.getConstant(0, VT);
1687 case ISD::SETNE: return DAG.getConstant(1, VT);
1690 // True if the sign bit of C1 is set.
1691 return DAG.getConstant(C1.isNegative(), VT);
1694 // True if the sign bit of C1 isn't set.
1695 return DAG.getConstant(C1.isNonNegative(), VT);
1701 // Otherwise, we can perform the comparison with the low bits.
1709 EVT newVT = N0.getOperand(0).getValueType();
1710 if (DCI.isBeforeLegalizeOps() ||
1711 (isOperationLegal(ISD::SETCC, newVT) &&
1712 getCondCodeAction(Cond, newVT)==Legal))
1713 return DAG.getSetCC(dl, VT, N0.getOperand(0),
1714 DAG.getConstant(APInt(C1).trunc(InSize), newVT),
1719 break; // todo, be more careful with signed comparisons
1721 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
1722 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1723 EVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
1724 unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits();
1725 EVT ExtDstTy = N0.getValueType();
1726 unsigned ExtDstTyBits = ExtDstTy.getSizeInBits();
1728 // If the extended part has any inconsistent bits, it cannot ever
1729 // compare equal. In other words, they have to be all ones or all
1732 APInt::getHighBitsSet(ExtDstTyBits, ExtDstTyBits - ExtSrcTyBits);
1733 if ((C1 & ExtBits) != 0 && (C1 & ExtBits) != ExtBits)
1734 return DAG.getConstant(Cond == ISD::SETNE, VT);
1737 EVT Op0Ty = N0.getOperand(0).getValueType();
1738 if (Op0Ty == ExtSrcTy) {
1739 ZextOp = N0.getOperand(0);
1741 APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits);
1742 ZextOp = DAG.getNode(ISD::AND, dl, Op0Ty, N0.getOperand(0),
1743 DAG.getConstant(Imm, Op0Ty));
1745 if (!DCI.isCalledByLegalizer())
1746 DCI.AddToWorklist(ZextOp.getNode());
1747 // Otherwise, make this a use of a zext.
1748 return DAG.getSetCC(dl, VT, ZextOp,
1749 DAG.getConstant(C1 & APInt::getLowBitsSet(
1754 } else if ((N1C->isNullValue() || N1C->getAPIntValue() == 1) &&
1755 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1757 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC
1758 if (N0.getOpcode() == ISD::SETCC) {
1759 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getZExtValue() != 1);
1763 // Invert the condition.
1764 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
1765 CC = ISD::getSetCCInverse(CC,
1766 N0.getOperand(0).getValueType().isInteger());
1767 return DAG.getSetCC(dl, VT, N0.getOperand(0), N0.getOperand(1), CC);
1770 if ((N0.getOpcode() == ISD::XOR ||
1771 (N0.getOpcode() == ISD::AND &&
1772 N0.getOperand(0).getOpcode() == ISD::XOR &&
1773 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
1774 isa<ConstantSDNode>(N0.getOperand(1)) &&
1775 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue() == 1) {
1776 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We
1777 // can only do this if the top bits are known zero.
1778 unsigned BitWidth = N0.getValueSizeInBits();
1779 if (DAG.MaskedValueIsZero(N0,
1780 APInt::getHighBitsSet(BitWidth,
1782 // Okay, get the un-inverted input value.
1784 if (N0.getOpcode() == ISD::XOR)
1785 Val = N0.getOperand(0);
1787 assert(N0.getOpcode() == ISD::AND &&
1788 N0.getOperand(0).getOpcode() == ISD::XOR);
1789 // ((X^1)&1)^1 -> X & 1
1790 Val = DAG.getNode(ISD::AND, dl, N0.getValueType(),
1791 N0.getOperand(0).getOperand(0),
1794 return DAG.getSetCC(dl, VT, Val, N1,
1795 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
1800 APInt MinVal, MaxVal;
1801 unsigned OperandBitSize = N1C->getValueType(0).getSizeInBits();
1802 if (ISD::isSignedIntSetCC(Cond)) {
1803 MinVal = APInt::getSignedMinValue(OperandBitSize);
1804 MaxVal = APInt::getSignedMaxValue(OperandBitSize);
1806 MinVal = APInt::getMinValue(OperandBitSize);
1807 MaxVal = APInt::getMaxValue(OperandBitSize);
1810 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
1811 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
1812 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true
1813 // X >= C0 --> X > (C0-1)
1814 return DAG.getSetCC(dl, VT, N0,
1815 DAG.getConstant(C1-1, N1.getValueType()),
1816 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
1819 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
1820 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true
1821 // X <= C0 --> X < (C0+1)
1822 return DAG.getSetCC(dl, VT, N0,
1823 DAG.getConstant(C1+1, N1.getValueType()),
1824 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
1827 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
1828 return DAG.getConstant(0, VT); // X < MIN --> false
1829 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal)
1830 return DAG.getConstant(1, VT); // X >= MIN --> true
1831 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal)
1832 return DAG.getConstant(0, VT); // X > MAX --> false
1833 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal)
1834 return DAG.getConstant(1, VT); // X <= MAX --> true
1836 // Canonicalize setgt X, Min --> setne X, Min
1837 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
1838 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
1839 // Canonicalize setlt X, Max --> setne X, Max
1840 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
1841 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
1843 // If we have setult X, 1, turn it into seteq X, 0
1844 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
1845 return DAG.getSetCC(dl, VT, N0,
1846 DAG.getConstant(MinVal, N0.getValueType()),
1848 // If we have setugt X, Max-1, turn it into seteq X, Max
1849 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
1850 return DAG.getSetCC(dl, VT, N0,
1851 DAG.getConstant(MaxVal, N0.getValueType()),
1854 // If we have "setcc X, C0", check to see if we can shrink the immediate
1857 // SETUGT X, SINTMAX -> SETLT X, 0
1858 if (Cond == ISD::SETUGT &&
1859 C1 == APInt::getSignedMaxValue(OperandBitSize))
1860 return DAG.getSetCC(dl, VT, N0,
1861 DAG.getConstant(0, N1.getValueType()),
1864 // SETULT X, SINTMIN -> SETGT X, -1
1865 if (Cond == ISD::SETULT &&
1866 C1 == APInt::getSignedMinValue(OperandBitSize)) {
1867 SDValue ConstMinusOne =
1868 DAG.getConstant(APInt::getAllOnesValue(OperandBitSize),
1870 return DAG.getSetCC(dl, VT, N0, ConstMinusOne, ISD::SETGT);
1873 // Fold bit comparisons when we can.
1874 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1875 VT == N0.getValueType() && N0.getOpcode() == ISD::AND)
1876 if (ConstantSDNode *AndRHS =
1877 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1878 EVT ShiftTy = DCI.isBeforeLegalize() ?
1879 getPointerTy() : getShiftAmountTy();
1880 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
1881 // Perform the xform if the AND RHS is a single bit.
1882 if (isPowerOf2_64(AndRHS->getZExtValue())) {
1883 return DAG.getNode(ISD::SRL, dl, VT, N0,
1884 DAG.getConstant(Log2_64(AndRHS->getZExtValue()),
1887 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getZExtValue()) {
1888 // (X & 8) == 8 --> (X & 8) >> 3
1889 // Perform the xform if C1 is a single bit.
1890 if (C1.isPowerOf2()) {
1891 return DAG.getNode(ISD::SRL, dl, VT, N0,
1892 DAG.getConstant(C1.logBase2(), ShiftTy));
1898 if (isa<ConstantFPSDNode>(N0.getNode())) {
1899 // Constant fold or commute setcc.
1900 SDValue O = DAG.FoldSetCC(VT, N0, N1, Cond, dl);
1901 if (O.getNode()) return O;
1902 } else if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1.getNode())) {
1903 // If the RHS of an FP comparison is a constant, simplify it away in
1905 if (CFP->getValueAPF().isNaN()) {
1906 // If an operand is known to be a nan, we can fold it.
1907 switch (ISD::getUnorderedFlavor(Cond)) {
1908 default: llvm_unreachable("Unknown flavor!");
1909 case 0: // Known false.
1910 return DAG.getConstant(0, VT);
1911 case 1: // Known true.
1912 return DAG.getConstant(1, VT);
1913 case 2: // Undefined.
1914 return DAG.getUNDEF(VT);
1918 // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the
1919 // constant if knowing that the operand is non-nan is enough. We prefer to
1920 // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to
1922 if (Cond == ISD::SETO || Cond == ISD::SETUO)
1923 return DAG.getSetCC(dl, VT, N0, N0, Cond);
1925 // If the condition is not legal, see if we can find an equivalent one
1927 if (!isCondCodeLegal(Cond, N0.getValueType())) {
1928 // If the comparison was an awkward floating-point == or != and one of
1929 // the comparison operands is infinity or negative infinity, convert the
1930 // condition to a less-awkward <= or >=.
1931 if (CFP->getValueAPF().isInfinity()) {
1932 if (CFP->getValueAPF().isNegative()) {
1933 if (Cond == ISD::SETOEQ &&
1934 isCondCodeLegal(ISD::SETOLE, N0.getValueType()))
1935 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLE);
1936 if (Cond == ISD::SETUEQ &&
1937 isCondCodeLegal(ISD::SETOLE, N0.getValueType()))
1938 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULE);
1939 if (Cond == ISD::SETUNE &&
1940 isCondCodeLegal(ISD::SETUGT, N0.getValueType()))
1941 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGT);
1942 if (Cond == ISD::SETONE &&
1943 isCondCodeLegal(ISD::SETUGT, N0.getValueType()))
1944 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGT);
1946 if (Cond == ISD::SETOEQ &&
1947 isCondCodeLegal(ISD::SETOGE, N0.getValueType()))
1948 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGE);
1949 if (Cond == ISD::SETUEQ &&
1950 isCondCodeLegal(ISD::SETOGE, N0.getValueType()))
1951 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGE);
1952 if (Cond == ISD::SETUNE &&
1953 isCondCodeLegal(ISD::SETULT, N0.getValueType()))
1954 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULT);
1955 if (Cond == ISD::SETONE &&
1956 isCondCodeLegal(ISD::SETULT, N0.getValueType()))
1957 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLT);
1964 // We can always fold X == X for integer setcc's.
1965 if (N0.getValueType().isInteger())
1966 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
1967 unsigned UOF = ISD::getUnorderedFlavor(Cond);
1968 if (UOF == 2) // FP operators that are undefined on NaNs.
1969 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
1970 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
1971 return DAG.getConstant(UOF, VT);
1972 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
1973 // if it is not already.
1974 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
1975 if (NewCond != Cond)
1976 return DAG.getSetCC(dl, VT, N0, N1, NewCond);
1979 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1980 N0.getValueType().isInteger()) {
1981 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
1982 N0.getOpcode() == ISD::XOR) {
1983 // Simplify (X+Y) == (X+Z) --> Y == Z
1984 if (N0.getOpcode() == N1.getOpcode()) {
1985 if (N0.getOperand(0) == N1.getOperand(0))
1986 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(1), Cond);
1987 if (N0.getOperand(1) == N1.getOperand(1))
1988 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(0), Cond);
1989 if (DAG.isCommutativeBinOp(N0.getOpcode())) {
1990 // If X op Y == Y op X, try other combinations.
1991 if (N0.getOperand(0) == N1.getOperand(1))
1992 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(0),
1994 if (N0.getOperand(1) == N1.getOperand(0))
1995 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(1),
2000 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) {
2001 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2002 // Turn (X+C1) == C2 --> X == C2-C1
2003 if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse()) {
2004 return DAG.getSetCC(dl, VT, N0.getOperand(0),
2005 DAG.getConstant(RHSC->getAPIntValue()-
2006 LHSR->getAPIntValue(),
2007 N0.getValueType()), Cond);
2010 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
2011 if (N0.getOpcode() == ISD::XOR)
2012 // If we know that all of the inverted bits are zero, don't bother
2013 // performing the inversion.
2014 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue()))
2016 DAG.getSetCC(dl, VT, N0.getOperand(0),
2017 DAG.getConstant(LHSR->getAPIntValue() ^
2018 RHSC->getAPIntValue(),
2023 // Turn (C1-X) == C2 --> X == C1-C2
2024 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
2025 if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse()) {
2027 DAG.getSetCC(dl, VT, N0.getOperand(1),
2028 DAG.getConstant(SUBC->getAPIntValue() -
2029 RHSC->getAPIntValue(),
2036 // Simplify (X+Z) == X --> Z == 0
2037 if (N0.getOperand(0) == N1)
2038 return DAG.getSetCC(dl, VT, N0.getOperand(1),
2039 DAG.getConstant(0, N0.getValueType()), Cond);
2040 if (N0.getOperand(1) == N1) {
2041 if (DAG.isCommutativeBinOp(N0.getOpcode()))
2042 return DAG.getSetCC(dl, VT, N0.getOperand(0),
2043 DAG.getConstant(0, N0.getValueType()), Cond);
2044 else if (N0.getNode()->hasOneUse()) {
2045 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
2046 // (Z-X) == X --> Z == X<<1
2047 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(),
2049 DAG.getConstant(1, getShiftAmountTy()));
2050 if (!DCI.isCalledByLegalizer())
2051 DCI.AddToWorklist(SH.getNode());
2052 return DAG.getSetCC(dl, VT, N0.getOperand(0), SH, Cond);
2057 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
2058 N1.getOpcode() == ISD::XOR) {
2059 // Simplify X == (X+Z) --> Z == 0
2060 if (N1.getOperand(0) == N0) {
2061 return DAG.getSetCC(dl, VT, N1.getOperand(1),
2062 DAG.getConstant(0, N1.getValueType()), Cond);
2063 } else if (N1.getOperand(1) == N0) {
2064 if (DAG.isCommutativeBinOp(N1.getOpcode())) {
2065 return DAG.getSetCC(dl, VT, N1.getOperand(0),
2066 DAG.getConstant(0, N1.getValueType()), Cond);
2067 } else if (N1.getNode()->hasOneUse()) {
2068 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
2069 // X == (Z-X) --> X<<1 == Z
2070 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(), N0,
2071 DAG.getConstant(1, getShiftAmountTy()));
2072 if (!DCI.isCalledByLegalizer())
2073 DCI.AddToWorklist(SH.getNode());
2074 return DAG.getSetCC(dl, VT, SH, N1.getOperand(0), Cond);
2079 // Simplify x&y == y to x&y != 0 if y has exactly one bit set.
2080 // Note that where y is variable and is known to have at most
2081 // one bit set (for example, if it is z&1) we cannot do this;
2082 // the expressions are not equivalent when y==0.
2083 if (N0.getOpcode() == ISD::AND)
2084 if (N0.getOperand(0) == N1 || N0.getOperand(1) == N1) {
2085 if (ValueHasExactlyOneBitSet(N1, DAG)) {
2086 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true);
2087 SDValue Zero = DAG.getConstant(0, N1.getValueType());
2088 return DAG.getSetCC(dl, VT, N0, Zero, Cond);
2091 if (N1.getOpcode() == ISD::AND)
2092 if (N1.getOperand(0) == N0 || N1.getOperand(1) == N0) {
2093 if (ValueHasExactlyOneBitSet(N0, DAG)) {
2094 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true);
2095 SDValue Zero = DAG.getConstant(0, N0.getValueType());
2096 return DAG.getSetCC(dl, VT, N1, Zero, Cond);
2101 // Fold away ALL boolean setcc's.
2103 if (N0.getValueType() == MVT::i1 && foldBooleans) {
2105 default: llvm_unreachable("Unknown integer setcc!");
2106 case ISD::SETEQ: // X == Y -> ~(X^Y)
2107 Temp = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1);
2108 N0 = DAG.getNOT(dl, Temp, MVT::i1);
2109 if (!DCI.isCalledByLegalizer())
2110 DCI.AddToWorklist(Temp.getNode());
2112 case ISD::SETNE: // X != Y --> (X^Y)
2113 N0 = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1);
2115 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> ~X & Y
2116 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> ~X & Y
2117 Temp = DAG.getNOT(dl, N0, MVT::i1);
2118 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N1, Temp);
2119 if (!DCI.isCalledByLegalizer())
2120 DCI.AddToWorklist(Temp.getNode());
2122 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> ~Y & X
2123 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> ~Y & X
2124 Temp = DAG.getNOT(dl, N1, MVT::i1);
2125 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N0, Temp);
2126 if (!DCI.isCalledByLegalizer())
2127 DCI.AddToWorklist(Temp.getNode());
2129 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> ~X | Y
2130 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> ~X | Y
2131 Temp = DAG.getNOT(dl, N0, MVT::i1);
2132 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N1, Temp);
2133 if (!DCI.isCalledByLegalizer())
2134 DCI.AddToWorklist(Temp.getNode());
2136 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> ~Y | X
2137 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> ~Y | X
2138 Temp = DAG.getNOT(dl, N1, MVT::i1);
2139 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N0, Temp);
2142 if (VT != MVT::i1) {
2143 if (!DCI.isCalledByLegalizer())
2144 DCI.AddToWorklist(N0.getNode());
2145 // FIXME: If running after legalize, we probably can't do this.
2146 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, N0);
2151 // Could not fold it.
2155 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
2156 /// node is a GlobalAddress + offset.
2157 bool TargetLowering::isGAPlusOffset(SDNode *N, GlobalValue* &GA,
2158 int64_t &Offset) const {
2159 if (isa<GlobalAddressSDNode>(N)) {
2160 GlobalAddressSDNode *GASD = cast<GlobalAddressSDNode>(N);
2161 GA = GASD->getGlobal();
2162 Offset += GASD->getOffset();
2166 if (N->getOpcode() == ISD::ADD) {
2167 SDValue N1 = N->getOperand(0);
2168 SDValue N2 = N->getOperand(1);
2169 if (isGAPlusOffset(N1.getNode(), GA, Offset)) {
2170 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
2172 Offset += V->getSExtValue();
2175 } else if (isGAPlusOffset(N2.getNode(), GA, Offset)) {
2176 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
2178 Offset += V->getSExtValue();
2187 SDValue TargetLowering::
2188 PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
2189 // Default implementation: no optimization.
2193 //===----------------------------------------------------------------------===//
2194 // Inline Assembler Implementation Methods
2195 //===----------------------------------------------------------------------===//
2198 TargetLowering::ConstraintType
2199 TargetLowering::getConstraintType(const std::string &Constraint) const {
2200 // FIXME: lots more standard ones to handle.
2201 if (Constraint.size() == 1) {
2202 switch (Constraint[0]) {
2204 case 'r': return C_RegisterClass;
2206 case 'o': // offsetable
2207 case 'V': // not offsetable
2209 case 'i': // Simple Integer or Relocatable Constant
2210 case 'n': // Simple Integer
2211 case 's': // Relocatable Constant
2212 case 'X': // Allow ANY value.
2213 case 'I': // Target registers.
2225 if (Constraint.size() > 1 && Constraint[0] == '{' &&
2226 Constraint[Constraint.size()-1] == '}')
2231 /// LowerXConstraint - try to replace an X constraint, which matches anything,
2232 /// with another that has more specific requirements based on the type of the
2233 /// corresponding operand.
2234 const char *TargetLowering::LowerXConstraint(EVT ConstraintVT) const{
2235 if (ConstraintVT.isInteger())
2237 if (ConstraintVT.isFloatingPoint())
2238 return "f"; // works for many targets
2242 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
2243 /// vector. If it is invalid, don't add anything to Ops.
2244 void TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
2245 char ConstraintLetter,
2247 std::vector<SDValue> &Ops,
2248 SelectionDAG &DAG) const {
2249 switch (ConstraintLetter) {
2251 case 'X': // Allows any operand; labels (basic block) use this.
2252 if (Op.getOpcode() == ISD::BasicBlock) {
2257 case 'i': // Simple Integer or Relocatable Constant
2258 case 'n': // Simple Integer
2259 case 's': { // Relocatable Constant
2260 // These operands are interested in values of the form (GV+C), where C may
2261 // be folded in as an offset of GV, or it may be explicitly added. Also, it
2262 // is possible and fine if either GV or C are missing.
2263 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2264 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
2266 // If we have "(add GV, C)", pull out GV/C
2267 if (Op.getOpcode() == ISD::ADD) {
2268 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
2269 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
2270 if (C == 0 || GA == 0) {
2271 C = dyn_cast<ConstantSDNode>(Op.getOperand(0));
2272 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1));
2274 if (C == 0 || GA == 0)
2278 // If we find a valid operand, map to the TargetXXX version so that the
2279 // value itself doesn't get selected.
2280 if (GA) { // Either &GV or &GV+C
2281 if (ConstraintLetter != 'n') {
2282 int64_t Offs = GA->getOffset();
2283 if (C) Offs += C->getZExtValue();
2284 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(),
2285 Op.getValueType(), Offs));
2289 if (C) { // just C, no GV.
2290 // Simple constants are not allowed for 's'.
2291 if (ConstraintLetter != 's') {
2292 // gcc prints these as sign extended. Sign extend value to 64 bits
2293 // now; without this it would get ZExt'd later in
2294 // ScheduleDAGSDNodes::EmitNode, which is very generic.
2295 Ops.push_back(DAG.getTargetConstant(C->getAPIntValue().getSExtValue(),
2305 std::vector<unsigned> TargetLowering::
2306 getRegClassForInlineAsmConstraint(const std::string &Constraint,
2308 return std::vector<unsigned>();
2312 std::pair<unsigned, const TargetRegisterClass*> TargetLowering::
2313 getRegForInlineAsmConstraint(const std::string &Constraint,
2315 if (Constraint[0] != '{')
2316 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
2317 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?");
2319 // Remove the braces from around the name.
2320 StringRef RegName(Constraint.data()+1, Constraint.size()-2);
2322 // Figure out which register class contains this reg.
2323 const TargetRegisterInfo *RI = TM.getRegisterInfo();
2324 for (TargetRegisterInfo::regclass_iterator RCI = RI->regclass_begin(),
2325 E = RI->regclass_end(); RCI != E; ++RCI) {
2326 const TargetRegisterClass *RC = *RCI;
2328 // If none of the the value types for this register class are valid, we
2329 // can't use it. For example, 64-bit reg classes on 32-bit targets.
2330 bool isLegal = false;
2331 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
2333 if (isTypeLegal(*I)) {
2339 if (!isLegal) continue;
2341 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end();
2343 if (RegName.equals_lower(RI->getName(*I)))
2344 return std::make_pair(*I, RC);
2348 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
2351 //===----------------------------------------------------------------------===//
2352 // Constraint Selection.
2354 /// isMatchingInputConstraint - Return true of this is an input operand that is
2355 /// a matching constraint like "4".
2356 bool TargetLowering::AsmOperandInfo::isMatchingInputConstraint() const {
2357 assert(!ConstraintCode.empty() && "No known constraint!");
2358 return isdigit(ConstraintCode[0]);
2361 /// getMatchedOperand - If this is an input matching constraint, this method
2362 /// returns the output operand it matches.
2363 unsigned TargetLowering::AsmOperandInfo::getMatchedOperand() const {
2364 assert(!ConstraintCode.empty() && "No known constraint!");
2365 return atoi(ConstraintCode.c_str());
2369 /// getConstraintGenerality - Return an integer indicating how general CT
2371 static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) {
2373 default: llvm_unreachable("Unknown constraint type!");
2374 case TargetLowering::C_Other:
2375 case TargetLowering::C_Unknown:
2377 case TargetLowering::C_Register:
2379 case TargetLowering::C_RegisterClass:
2381 case TargetLowering::C_Memory:
2386 /// ChooseConstraint - If there are multiple different constraints that we
2387 /// could pick for this operand (e.g. "imr") try to pick the 'best' one.
2388 /// This is somewhat tricky: constraints fall into four classes:
2389 /// Other -> immediates and magic values
2390 /// Register -> one specific register
2391 /// RegisterClass -> a group of regs
2392 /// Memory -> memory
2393 /// Ideally, we would pick the most specific constraint possible: if we have
2394 /// something that fits into a register, we would pick it. The problem here
2395 /// is that if we have something that could either be in a register or in
2396 /// memory that use of the register could cause selection of *other*
2397 /// operands to fail: they might only succeed if we pick memory. Because of
2398 /// this the heuristic we use is:
2400 /// 1) If there is an 'other' constraint, and if the operand is valid for
2401 /// that constraint, use it. This makes us take advantage of 'i'
2402 /// constraints when available.
2403 /// 2) Otherwise, pick the most general constraint present. This prefers
2404 /// 'm' over 'r', for example.
2406 static void ChooseConstraint(TargetLowering::AsmOperandInfo &OpInfo,
2407 bool hasMemory, const TargetLowering &TLI,
2408 SDValue Op, SelectionDAG *DAG) {
2409 assert(OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options");
2410 unsigned BestIdx = 0;
2411 TargetLowering::ConstraintType BestType = TargetLowering::C_Unknown;
2412 int BestGenerality = -1;
2414 // Loop over the options, keeping track of the most general one.
2415 for (unsigned i = 0, e = OpInfo.Codes.size(); i != e; ++i) {
2416 TargetLowering::ConstraintType CType =
2417 TLI.getConstraintType(OpInfo.Codes[i]);
2419 // If this is an 'other' constraint, see if the operand is valid for it.
2420 // For example, on X86 we might have an 'rI' constraint. If the operand
2421 // is an integer in the range [0..31] we want to use I (saving a load
2422 // of a register), otherwise we must use 'r'.
2423 if (CType == TargetLowering::C_Other && Op.getNode()) {
2424 assert(OpInfo.Codes[i].size() == 1 &&
2425 "Unhandled multi-letter 'other' constraint");
2426 std::vector<SDValue> ResultOps;
2427 TLI.LowerAsmOperandForConstraint(Op, OpInfo.Codes[i][0], hasMemory,
2429 if (!ResultOps.empty()) {
2436 // This constraint letter is more general than the previous one, use it.
2437 int Generality = getConstraintGenerality(CType);
2438 if (Generality > BestGenerality) {
2441 BestGenerality = Generality;
2445 OpInfo.ConstraintCode = OpInfo.Codes[BestIdx];
2446 OpInfo.ConstraintType = BestType;
2449 /// ComputeConstraintToUse - Determines the constraint code and constraint
2450 /// type to use for the specific AsmOperandInfo, setting
2451 /// OpInfo.ConstraintCode and OpInfo.ConstraintType.
2452 void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo,
2455 SelectionDAG *DAG) const {
2456 assert(!OpInfo.Codes.empty() && "Must have at least one constraint");
2458 // Single-letter constraints ('r') are very common.
2459 if (OpInfo.Codes.size() == 1) {
2460 OpInfo.ConstraintCode = OpInfo.Codes[0];
2461 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
2463 ChooseConstraint(OpInfo, hasMemory, *this, Op, DAG);
2466 // 'X' matches anything.
2467 if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) {
2468 // Labels and constants are handled elsewhere ('X' is the only thing
2469 // that matches labels). For Functions, the type here is the type of
2470 // the result, which is not what we want to look at; leave them alone.
2471 Value *v = OpInfo.CallOperandVal;
2472 if (isa<BasicBlock>(v) || isa<ConstantInt>(v) || isa<Function>(v)) {
2473 OpInfo.CallOperandVal = v;
2477 // Otherwise, try to resolve it to something we know about by looking at
2478 // the actual operand type.
2479 if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) {
2480 OpInfo.ConstraintCode = Repl;
2481 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
2486 //===----------------------------------------------------------------------===//
2487 // Loop Strength Reduction hooks
2488 //===----------------------------------------------------------------------===//
2490 /// isLegalAddressingMode - Return true if the addressing mode represented
2491 /// by AM is legal for this target, for a load/store of the specified type.
2492 bool TargetLowering::isLegalAddressingMode(const AddrMode &AM,
2493 const Type *Ty) const {
2494 // The default implementation of this implements a conservative RISCy, r+r and
2497 // Allows a sign-extended 16-bit immediate field.
2498 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
2501 // No global is ever allowed as a base.
2505 // Only support r+r,
2507 case 0: // "r+i" or just "i", depending on HasBaseReg.
2510 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
2512 // Otherwise we have r+r or r+i.
2515 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
2517 // Allow 2*r as r+r.
2524 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
2525 /// return a DAG expression to select that will generate the same value by
2526 /// multiplying by a magic number. See:
2527 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
2528 SDValue TargetLowering::BuildSDIV(SDNode *N, SelectionDAG &DAG,
2529 std::vector<SDNode*>* Created) const {
2530 EVT VT = N->getValueType(0);
2531 DebugLoc dl= N->getDebugLoc();
2533 // Check to see if we can do this.
2534 // FIXME: We should be more aggressive here.
2535 if (!isTypeLegal(VT))
2538 APInt d = cast<ConstantSDNode>(N->getOperand(1))->getAPIntValue();
2539 APInt::ms magics = d.magic();
2541 // Multiply the numerator (operand 0) by the magic value
2542 // FIXME: We should support doing a MUL in a wider type
2544 if (isOperationLegalOrCustom(ISD::MULHS, VT))
2545 Q = DAG.getNode(ISD::MULHS, dl, VT, N->getOperand(0),
2546 DAG.getConstant(magics.m, VT));
2547 else if (isOperationLegalOrCustom(ISD::SMUL_LOHI, VT))
2548 Q = SDValue(DAG.getNode(ISD::SMUL_LOHI, dl, DAG.getVTList(VT, VT),
2550 DAG.getConstant(magics.m, VT)).getNode(), 1);
2552 return SDValue(); // No mulhs or equvialent
2553 // If d > 0 and m < 0, add the numerator
2554 if (d.isStrictlyPositive() && magics.m.isNegative()) {
2555 Q = DAG.getNode(ISD::ADD, dl, VT, Q, N->getOperand(0));
2557 Created->push_back(Q.getNode());
2559 // If d < 0 and m > 0, subtract the numerator.
2560 if (d.isNegative() && magics.m.isStrictlyPositive()) {
2561 Q = DAG.getNode(ISD::SUB, dl, VT, Q, N->getOperand(0));
2563 Created->push_back(Q.getNode());
2565 // Shift right algebraic if shift value is nonzero
2567 Q = DAG.getNode(ISD::SRA, dl, VT, Q,
2568 DAG.getConstant(magics.s, getShiftAmountTy()));
2570 Created->push_back(Q.getNode());
2572 // Extract the sign bit and add it to the quotient
2574 DAG.getNode(ISD::SRL, dl, VT, Q, DAG.getConstant(VT.getSizeInBits()-1,
2575 getShiftAmountTy()));
2577 Created->push_back(T.getNode());
2578 return DAG.getNode(ISD::ADD, dl, VT, Q, T);
2581 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
2582 /// return a DAG expression to select that will generate the same value by
2583 /// multiplying by a magic number. See:
2584 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
2585 SDValue TargetLowering::BuildUDIV(SDNode *N, SelectionDAG &DAG,
2586 std::vector<SDNode*>* Created) const {
2587 EVT VT = N->getValueType(0);
2588 DebugLoc dl = N->getDebugLoc();
2590 // Check to see if we can do this.
2591 // FIXME: We should be more aggressive here.
2592 if (!isTypeLegal(VT))
2595 // FIXME: We should use a narrower constant when the upper
2596 // bits are known to be zero.
2597 ConstantSDNode *N1C = cast<ConstantSDNode>(N->getOperand(1));
2598 APInt::mu magics = N1C->getAPIntValue().magicu();
2600 // Multiply the numerator (operand 0) by the magic value
2601 // FIXME: We should support doing a MUL in a wider type
2603 if (isOperationLegalOrCustom(ISD::MULHU, VT))
2604 Q = DAG.getNode(ISD::MULHU, dl, VT, N->getOperand(0),
2605 DAG.getConstant(magics.m, VT));
2606 else if (isOperationLegalOrCustom(ISD::UMUL_LOHI, VT))
2607 Q = SDValue(DAG.getNode(ISD::UMUL_LOHI, dl, DAG.getVTList(VT, VT),
2609 DAG.getConstant(magics.m, VT)).getNode(), 1);
2611 return SDValue(); // No mulhu or equvialent
2613 Created->push_back(Q.getNode());
2615 if (magics.a == 0) {
2616 assert(magics.s < N1C->getAPIntValue().getBitWidth() &&
2617 "We shouldn't generate an undefined shift!");
2618 return DAG.getNode(ISD::SRL, dl, VT, Q,
2619 DAG.getConstant(magics.s, getShiftAmountTy()));
2621 SDValue NPQ = DAG.getNode(ISD::SUB, dl, VT, N->getOperand(0), Q);
2623 Created->push_back(NPQ.getNode());
2624 NPQ = DAG.getNode(ISD::SRL, dl, VT, NPQ,
2625 DAG.getConstant(1, getShiftAmountTy()));
2627 Created->push_back(NPQ.getNode());
2628 NPQ = DAG.getNode(ISD::ADD, dl, VT, NPQ, Q);
2630 Created->push_back(NPQ.getNode());
2631 return DAG.getNode(ISD::SRL, dl, VT, NPQ,
2632 DAG.getConstant(magics.s-1, getShiftAmountTy()));