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/MC/MCExpr.h"
17 #include "llvm/Target/TargetData.h"
18 #include "llvm/Target/TargetLoweringObjectFile.h"
19 #include "llvm/Target/TargetMachine.h"
20 #include "llvm/Target/TargetRegisterInfo.h"
21 #include "llvm/Target/TargetSubtarget.h"
22 #include "llvm/GlobalVariable.h"
23 #include "llvm/DerivedTypes.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineJumpTableInfo.h"
26 #include "llvm/CodeGen/MachineFunction.h"
27 #include "llvm/CodeGen/SelectionDAG.h"
28 #include "llvm/ADT/STLExtras.h"
29 #include "llvm/Support/ErrorHandling.h"
30 #include "llvm/Support/MathExtras.h"
34 TLSModel::Model getTLSModel(const GlobalValue *GV, Reloc::Model reloc) {
35 bool isLocal = GV->hasLocalLinkage();
36 bool isDeclaration = GV->isDeclaration();
37 // FIXME: what should we do for protected and internal visibility?
38 // For variables, is internal different from hidden?
39 bool isHidden = GV->hasHiddenVisibility();
41 if (reloc == Reloc::PIC_) {
42 if (isLocal || isHidden)
43 return TLSModel::LocalDynamic;
45 return TLSModel::GeneralDynamic;
47 if (!isDeclaration || isHidden)
48 return TLSModel::LocalExec;
50 return TLSModel::InitialExec;
55 /// InitLibcallNames - Set default libcall names.
57 static void InitLibcallNames(const char **Names) {
58 Names[RTLIB::SHL_I16] = "__ashlhi3";
59 Names[RTLIB::SHL_I32] = "__ashlsi3";
60 Names[RTLIB::SHL_I64] = "__ashldi3";
61 Names[RTLIB::SHL_I128] = "__ashlti3";
62 Names[RTLIB::SRL_I16] = "__lshrhi3";
63 Names[RTLIB::SRL_I32] = "__lshrsi3";
64 Names[RTLIB::SRL_I64] = "__lshrdi3";
65 Names[RTLIB::SRL_I128] = "__lshrti3";
66 Names[RTLIB::SRA_I16] = "__ashrhi3";
67 Names[RTLIB::SRA_I32] = "__ashrsi3";
68 Names[RTLIB::SRA_I64] = "__ashrdi3";
69 Names[RTLIB::SRA_I128] = "__ashrti3";
70 Names[RTLIB::MUL_I8] = "__mulqi3";
71 Names[RTLIB::MUL_I16] = "__mulhi3";
72 Names[RTLIB::MUL_I32] = "__mulsi3";
73 Names[RTLIB::MUL_I64] = "__muldi3";
74 Names[RTLIB::MUL_I128] = "__multi3";
75 Names[RTLIB::SDIV_I8] = "__divqi3";
76 Names[RTLIB::SDIV_I16] = "__divhi3";
77 Names[RTLIB::SDIV_I32] = "__divsi3";
78 Names[RTLIB::SDIV_I64] = "__divdi3";
79 Names[RTLIB::SDIV_I128] = "__divti3";
80 Names[RTLIB::UDIV_I8] = "__udivqi3";
81 Names[RTLIB::UDIV_I16] = "__udivhi3";
82 Names[RTLIB::UDIV_I32] = "__udivsi3";
83 Names[RTLIB::UDIV_I64] = "__udivdi3";
84 Names[RTLIB::UDIV_I128] = "__udivti3";
85 Names[RTLIB::SREM_I8] = "__modqi3";
86 Names[RTLIB::SREM_I16] = "__modhi3";
87 Names[RTLIB::SREM_I32] = "__modsi3";
88 Names[RTLIB::SREM_I64] = "__moddi3";
89 Names[RTLIB::SREM_I128] = "__modti3";
90 Names[RTLIB::UREM_I8] = "__umodqi3";
91 Names[RTLIB::UREM_I16] = "__umodhi3";
92 Names[RTLIB::UREM_I32] = "__umodsi3";
93 Names[RTLIB::UREM_I64] = "__umoddi3";
94 Names[RTLIB::UREM_I128] = "__umodti3";
95 Names[RTLIB::NEG_I32] = "__negsi2";
96 Names[RTLIB::NEG_I64] = "__negdi2";
97 Names[RTLIB::ADD_F32] = "__addsf3";
98 Names[RTLIB::ADD_F64] = "__adddf3";
99 Names[RTLIB::ADD_F80] = "__addxf3";
100 Names[RTLIB::ADD_PPCF128] = "__gcc_qadd";
101 Names[RTLIB::SUB_F32] = "__subsf3";
102 Names[RTLIB::SUB_F64] = "__subdf3";
103 Names[RTLIB::SUB_F80] = "__subxf3";
104 Names[RTLIB::SUB_PPCF128] = "__gcc_qsub";
105 Names[RTLIB::MUL_F32] = "__mulsf3";
106 Names[RTLIB::MUL_F64] = "__muldf3";
107 Names[RTLIB::MUL_F80] = "__mulxf3";
108 Names[RTLIB::MUL_PPCF128] = "__gcc_qmul";
109 Names[RTLIB::DIV_F32] = "__divsf3";
110 Names[RTLIB::DIV_F64] = "__divdf3";
111 Names[RTLIB::DIV_F80] = "__divxf3";
112 Names[RTLIB::DIV_PPCF128] = "__gcc_qdiv";
113 Names[RTLIB::REM_F32] = "fmodf";
114 Names[RTLIB::REM_F64] = "fmod";
115 Names[RTLIB::REM_F80] = "fmodl";
116 Names[RTLIB::REM_PPCF128] = "fmodl";
117 Names[RTLIB::POWI_F32] = "__powisf2";
118 Names[RTLIB::POWI_F64] = "__powidf2";
119 Names[RTLIB::POWI_F80] = "__powixf2";
120 Names[RTLIB::POWI_PPCF128] = "__powitf2";
121 Names[RTLIB::SQRT_F32] = "sqrtf";
122 Names[RTLIB::SQRT_F64] = "sqrt";
123 Names[RTLIB::SQRT_F80] = "sqrtl";
124 Names[RTLIB::SQRT_PPCF128] = "sqrtl";
125 Names[RTLIB::LOG_F32] = "logf";
126 Names[RTLIB::LOG_F64] = "log";
127 Names[RTLIB::LOG_F80] = "logl";
128 Names[RTLIB::LOG_PPCF128] = "logl";
129 Names[RTLIB::LOG2_F32] = "log2f";
130 Names[RTLIB::LOG2_F64] = "log2";
131 Names[RTLIB::LOG2_F80] = "log2l";
132 Names[RTLIB::LOG2_PPCF128] = "log2l";
133 Names[RTLIB::LOG10_F32] = "log10f";
134 Names[RTLIB::LOG10_F64] = "log10";
135 Names[RTLIB::LOG10_F80] = "log10l";
136 Names[RTLIB::LOG10_PPCF128] = "log10l";
137 Names[RTLIB::EXP_F32] = "expf";
138 Names[RTLIB::EXP_F64] = "exp";
139 Names[RTLIB::EXP_F80] = "expl";
140 Names[RTLIB::EXP_PPCF128] = "expl";
141 Names[RTLIB::EXP2_F32] = "exp2f";
142 Names[RTLIB::EXP2_F64] = "exp2";
143 Names[RTLIB::EXP2_F80] = "exp2l";
144 Names[RTLIB::EXP2_PPCF128] = "exp2l";
145 Names[RTLIB::SIN_F32] = "sinf";
146 Names[RTLIB::SIN_F64] = "sin";
147 Names[RTLIB::SIN_F80] = "sinl";
148 Names[RTLIB::SIN_PPCF128] = "sinl";
149 Names[RTLIB::COS_F32] = "cosf";
150 Names[RTLIB::COS_F64] = "cos";
151 Names[RTLIB::COS_F80] = "cosl";
152 Names[RTLIB::COS_PPCF128] = "cosl";
153 Names[RTLIB::POW_F32] = "powf";
154 Names[RTLIB::POW_F64] = "pow";
155 Names[RTLIB::POW_F80] = "powl";
156 Names[RTLIB::POW_PPCF128] = "powl";
157 Names[RTLIB::CEIL_F32] = "ceilf";
158 Names[RTLIB::CEIL_F64] = "ceil";
159 Names[RTLIB::CEIL_F80] = "ceill";
160 Names[RTLIB::CEIL_PPCF128] = "ceill";
161 Names[RTLIB::TRUNC_F32] = "truncf";
162 Names[RTLIB::TRUNC_F64] = "trunc";
163 Names[RTLIB::TRUNC_F80] = "truncl";
164 Names[RTLIB::TRUNC_PPCF128] = "truncl";
165 Names[RTLIB::RINT_F32] = "rintf";
166 Names[RTLIB::RINT_F64] = "rint";
167 Names[RTLIB::RINT_F80] = "rintl";
168 Names[RTLIB::RINT_PPCF128] = "rintl";
169 Names[RTLIB::NEARBYINT_F32] = "nearbyintf";
170 Names[RTLIB::NEARBYINT_F64] = "nearbyint";
171 Names[RTLIB::NEARBYINT_F80] = "nearbyintl";
172 Names[RTLIB::NEARBYINT_PPCF128] = "nearbyintl";
173 Names[RTLIB::FLOOR_F32] = "floorf";
174 Names[RTLIB::FLOOR_F64] = "floor";
175 Names[RTLIB::FLOOR_F80] = "floorl";
176 Names[RTLIB::FLOOR_PPCF128] = "floorl";
177 Names[RTLIB::FPEXT_F32_F64] = "__extendsfdf2";
178 Names[RTLIB::FPROUND_F64_F32] = "__truncdfsf2";
179 Names[RTLIB::FPROUND_F80_F32] = "__truncxfsf2";
180 Names[RTLIB::FPROUND_PPCF128_F32] = "__trunctfsf2";
181 Names[RTLIB::FPROUND_F80_F64] = "__truncxfdf2";
182 Names[RTLIB::FPROUND_PPCF128_F64] = "__trunctfdf2";
183 Names[RTLIB::FPTOSINT_F32_I8] = "__fixsfi8";
184 Names[RTLIB::FPTOSINT_F32_I16] = "__fixsfi16";
185 Names[RTLIB::FPTOSINT_F32_I32] = "__fixsfsi";
186 Names[RTLIB::FPTOSINT_F32_I64] = "__fixsfdi";
187 Names[RTLIB::FPTOSINT_F32_I128] = "__fixsfti";
188 Names[RTLIB::FPTOSINT_F64_I32] = "__fixdfsi";
189 Names[RTLIB::FPTOSINT_F64_I64] = "__fixdfdi";
190 Names[RTLIB::FPTOSINT_F64_I128] = "__fixdfti";
191 Names[RTLIB::FPTOSINT_F80_I32] = "__fixxfsi";
192 Names[RTLIB::FPTOSINT_F80_I64] = "__fixxfdi";
193 Names[RTLIB::FPTOSINT_F80_I128] = "__fixxfti";
194 Names[RTLIB::FPTOSINT_PPCF128_I32] = "__fixtfsi";
195 Names[RTLIB::FPTOSINT_PPCF128_I64] = "__fixtfdi";
196 Names[RTLIB::FPTOSINT_PPCF128_I128] = "__fixtfti";
197 Names[RTLIB::FPTOUINT_F32_I8] = "__fixunssfi8";
198 Names[RTLIB::FPTOUINT_F32_I16] = "__fixunssfi16";
199 Names[RTLIB::FPTOUINT_F32_I32] = "__fixunssfsi";
200 Names[RTLIB::FPTOUINT_F32_I64] = "__fixunssfdi";
201 Names[RTLIB::FPTOUINT_F32_I128] = "__fixunssfti";
202 Names[RTLIB::FPTOUINT_F64_I32] = "__fixunsdfsi";
203 Names[RTLIB::FPTOUINT_F64_I64] = "__fixunsdfdi";
204 Names[RTLIB::FPTOUINT_F64_I128] = "__fixunsdfti";
205 Names[RTLIB::FPTOUINT_F80_I32] = "__fixunsxfsi";
206 Names[RTLIB::FPTOUINT_F80_I64] = "__fixunsxfdi";
207 Names[RTLIB::FPTOUINT_F80_I128] = "__fixunsxfti";
208 Names[RTLIB::FPTOUINT_PPCF128_I32] = "__fixunstfsi";
209 Names[RTLIB::FPTOUINT_PPCF128_I64] = "__fixunstfdi";
210 Names[RTLIB::FPTOUINT_PPCF128_I128] = "__fixunstfti";
211 Names[RTLIB::SINTTOFP_I32_F32] = "__floatsisf";
212 Names[RTLIB::SINTTOFP_I32_F64] = "__floatsidf";
213 Names[RTLIB::SINTTOFP_I32_F80] = "__floatsixf";
214 Names[RTLIB::SINTTOFP_I32_PPCF128] = "__floatsitf";
215 Names[RTLIB::SINTTOFP_I64_F32] = "__floatdisf";
216 Names[RTLIB::SINTTOFP_I64_F64] = "__floatdidf";
217 Names[RTLIB::SINTTOFP_I64_F80] = "__floatdixf";
218 Names[RTLIB::SINTTOFP_I64_PPCF128] = "__floatditf";
219 Names[RTLIB::SINTTOFP_I128_F32] = "__floattisf";
220 Names[RTLIB::SINTTOFP_I128_F64] = "__floattidf";
221 Names[RTLIB::SINTTOFP_I128_F80] = "__floattixf";
222 Names[RTLIB::SINTTOFP_I128_PPCF128] = "__floattitf";
223 Names[RTLIB::UINTTOFP_I32_F32] = "__floatunsisf";
224 Names[RTLIB::UINTTOFP_I32_F64] = "__floatunsidf";
225 Names[RTLIB::UINTTOFP_I32_F80] = "__floatunsixf";
226 Names[RTLIB::UINTTOFP_I32_PPCF128] = "__floatunsitf";
227 Names[RTLIB::UINTTOFP_I64_F32] = "__floatundisf";
228 Names[RTLIB::UINTTOFP_I64_F64] = "__floatundidf";
229 Names[RTLIB::UINTTOFP_I64_F80] = "__floatundixf";
230 Names[RTLIB::UINTTOFP_I64_PPCF128] = "__floatunditf";
231 Names[RTLIB::UINTTOFP_I128_F32] = "__floatuntisf";
232 Names[RTLIB::UINTTOFP_I128_F64] = "__floatuntidf";
233 Names[RTLIB::UINTTOFP_I128_F80] = "__floatuntixf";
234 Names[RTLIB::UINTTOFP_I128_PPCF128] = "__floatuntitf";
235 Names[RTLIB::OEQ_F32] = "__eqsf2";
236 Names[RTLIB::OEQ_F64] = "__eqdf2";
237 Names[RTLIB::UNE_F32] = "__nesf2";
238 Names[RTLIB::UNE_F64] = "__nedf2";
239 Names[RTLIB::OGE_F32] = "__gesf2";
240 Names[RTLIB::OGE_F64] = "__gedf2";
241 Names[RTLIB::OLT_F32] = "__ltsf2";
242 Names[RTLIB::OLT_F64] = "__ltdf2";
243 Names[RTLIB::OLE_F32] = "__lesf2";
244 Names[RTLIB::OLE_F64] = "__ledf2";
245 Names[RTLIB::OGT_F32] = "__gtsf2";
246 Names[RTLIB::OGT_F64] = "__gtdf2";
247 Names[RTLIB::UO_F32] = "__unordsf2";
248 Names[RTLIB::UO_F64] = "__unorddf2";
249 Names[RTLIB::O_F32] = "__unordsf2";
250 Names[RTLIB::O_F64] = "__unorddf2";
251 Names[RTLIB::MEMCPY] = "memcpy";
252 Names[RTLIB::MEMMOVE] = "memmove";
253 Names[RTLIB::MEMSET] = "memset";
254 Names[RTLIB::UNWIND_RESUME] = "_Unwind_Resume";
257 /// InitLibcallCallingConvs - Set default libcall CallingConvs.
259 static void InitLibcallCallingConvs(CallingConv::ID *CCs) {
260 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
261 CCs[i] = CallingConv::C;
265 /// getFPEXT - Return the FPEXT_*_* value for the given types, or
266 /// UNKNOWN_LIBCALL if there is none.
267 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) {
268 if (OpVT == MVT::f32) {
269 if (RetVT == MVT::f64)
270 return FPEXT_F32_F64;
272 return UNKNOWN_LIBCALL;
275 /// getFPROUND - Return the FPROUND_*_* value for the given types, or
276 /// UNKNOWN_LIBCALL if there is none.
277 RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) {
278 if (RetVT == MVT::f32) {
279 if (OpVT == MVT::f64)
280 return FPROUND_F64_F32;
281 if (OpVT == MVT::f80)
282 return FPROUND_F80_F32;
283 if (OpVT == MVT::ppcf128)
284 return FPROUND_PPCF128_F32;
285 } else if (RetVT == MVT::f64) {
286 if (OpVT == MVT::f80)
287 return FPROUND_F80_F64;
288 if (OpVT == MVT::ppcf128)
289 return FPROUND_PPCF128_F64;
291 return UNKNOWN_LIBCALL;
294 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or
295 /// UNKNOWN_LIBCALL if there is none.
296 RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) {
297 if (OpVT == MVT::f32) {
298 if (RetVT == MVT::i8)
299 return FPTOSINT_F32_I8;
300 if (RetVT == MVT::i16)
301 return FPTOSINT_F32_I16;
302 if (RetVT == MVT::i32)
303 return FPTOSINT_F32_I32;
304 if (RetVT == MVT::i64)
305 return FPTOSINT_F32_I64;
306 if (RetVT == MVT::i128)
307 return FPTOSINT_F32_I128;
308 } else if (OpVT == MVT::f64) {
309 if (RetVT == MVT::i32)
310 return FPTOSINT_F64_I32;
311 if (RetVT == MVT::i64)
312 return FPTOSINT_F64_I64;
313 if (RetVT == MVT::i128)
314 return FPTOSINT_F64_I128;
315 } else if (OpVT == MVT::f80) {
316 if (RetVT == MVT::i32)
317 return FPTOSINT_F80_I32;
318 if (RetVT == MVT::i64)
319 return FPTOSINT_F80_I64;
320 if (RetVT == MVT::i128)
321 return FPTOSINT_F80_I128;
322 } else if (OpVT == MVT::ppcf128) {
323 if (RetVT == MVT::i32)
324 return FPTOSINT_PPCF128_I32;
325 if (RetVT == MVT::i64)
326 return FPTOSINT_PPCF128_I64;
327 if (RetVT == MVT::i128)
328 return FPTOSINT_PPCF128_I128;
330 return UNKNOWN_LIBCALL;
333 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or
334 /// UNKNOWN_LIBCALL if there is none.
335 RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) {
336 if (OpVT == MVT::f32) {
337 if (RetVT == MVT::i8)
338 return FPTOUINT_F32_I8;
339 if (RetVT == MVT::i16)
340 return FPTOUINT_F32_I16;
341 if (RetVT == MVT::i32)
342 return FPTOUINT_F32_I32;
343 if (RetVT == MVT::i64)
344 return FPTOUINT_F32_I64;
345 if (RetVT == MVT::i128)
346 return FPTOUINT_F32_I128;
347 } else if (OpVT == MVT::f64) {
348 if (RetVT == MVT::i32)
349 return FPTOUINT_F64_I32;
350 if (RetVT == MVT::i64)
351 return FPTOUINT_F64_I64;
352 if (RetVT == MVT::i128)
353 return FPTOUINT_F64_I128;
354 } else if (OpVT == MVT::f80) {
355 if (RetVT == MVT::i32)
356 return FPTOUINT_F80_I32;
357 if (RetVT == MVT::i64)
358 return FPTOUINT_F80_I64;
359 if (RetVT == MVT::i128)
360 return FPTOUINT_F80_I128;
361 } else if (OpVT == MVT::ppcf128) {
362 if (RetVT == MVT::i32)
363 return FPTOUINT_PPCF128_I32;
364 if (RetVT == MVT::i64)
365 return FPTOUINT_PPCF128_I64;
366 if (RetVT == MVT::i128)
367 return FPTOUINT_PPCF128_I128;
369 return UNKNOWN_LIBCALL;
372 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or
373 /// UNKNOWN_LIBCALL if there is none.
374 RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) {
375 if (OpVT == MVT::i32) {
376 if (RetVT == MVT::f32)
377 return SINTTOFP_I32_F32;
378 else if (RetVT == MVT::f64)
379 return SINTTOFP_I32_F64;
380 else if (RetVT == MVT::f80)
381 return SINTTOFP_I32_F80;
382 else if (RetVT == MVT::ppcf128)
383 return SINTTOFP_I32_PPCF128;
384 } else if (OpVT == MVT::i64) {
385 if (RetVT == MVT::f32)
386 return SINTTOFP_I64_F32;
387 else if (RetVT == MVT::f64)
388 return SINTTOFP_I64_F64;
389 else if (RetVT == MVT::f80)
390 return SINTTOFP_I64_F80;
391 else if (RetVT == MVT::ppcf128)
392 return SINTTOFP_I64_PPCF128;
393 } else if (OpVT == MVT::i128) {
394 if (RetVT == MVT::f32)
395 return SINTTOFP_I128_F32;
396 else if (RetVT == MVT::f64)
397 return SINTTOFP_I128_F64;
398 else if (RetVT == MVT::f80)
399 return SINTTOFP_I128_F80;
400 else if (RetVT == MVT::ppcf128)
401 return SINTTOFP_I128_PPCF128;
403 return UNKNOWN_LIBCALL;
406 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or
407 /// UNKNOWN_LIBCALL if there is none.
408 RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) {
409 if (OpVT == MVT::i32) {
410 if (RetVT == MVT::f32)
411 return UINTTOFP_I32_F32;
412 else if (RetVT == MVT::f64)
413 return UINTTOFP_I32_F64;
414 else if (RetVT == MVT::f80)
415 return UINTTOFP_I32_F80;
416 else if (RetVT == MVT::ppcf128)
417 return UINTTOFP_I32_PPCF128;
418 } else if (OpVT == MVT::i64) {
419 if (RetVT == MVT::f32)
420 return UINTTOFP_I64_F32;
421 else if (RetVT == MVT::f64)
422 return UINTTOFP_I64_F64;
423 else if (RetVT == MVT::f80)
424 return UINTTOFP_I64_F80;
425 else if (RetVT == MVT::ppcf128)
426 return UINTTOFP_I64_PPCF128;
427 } else if (OpVT == MVT::i128) {
428 if (RetVT == MVT::f32)
429 return UINTTOFP_I128_F32;
430 else if (RetVT == MVT::f64)
431 return UINTTOFP_I128_F64;
432 else if (RetVT == MVT::f80)
433 return UINTTOFP_I128_F80;
434 else if (RetVT == MVT::ppcf128)
435 return UINTTOFP_I128_PPCF128;
437 return UNKNOWN_LIBCALL;
440 /// InitCmpLibcallCCs - Set default comparison libcall CC.
442 static void InitCmpLibcallCCs(ISD::CondCode *CCs) {
443 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL);
444 CCs[RTLIB::OEQ_F32] = ISD::SETEQ;
445 CCs[RTLIB::OEQ_F64] = ISD::SETEQ;
446 CCs[RTLIB::UNE_F32] = ISD::SETNE;
447 CCs[RTLIB::UNE_F64] = ISD::SETNE;
448 CCs[RTLIB::OGE_F32] = ISD::SETGE;
449 CCs[RTLIB::OGE_F64] = ISD::SETGE;
450 CCs[RTLIB::OLT_F32] = ISD::SETLT;
451 CCs[RTLIB::OLT_F64] = ISD::SETLT;
452 CCs[RTLIB::OLE_F32] = ISD::SETLE;
453 CCs[RTLIB::OLE_F64] = ISD::SETLE;
454 CCs[RTLIB::OGT_F32] = ISD::SETGT;
455 CCs[RTLIB::OGT_F64] = ISD::SETGT;
456 CCs[RTLIB::UO_F32] = ISD::SETNE;
457 CCs[RTLIB::UO_F64] = ISD::SETNE;
458 CCs[RTLIB::O_F32] = ISD::SETEQ;
459 CCs[RTLIB::O_F64] = ISD::SETEQ;
462 /// NOTE: The constructor takes ownership of TLOF.
463 TargetLowering::TargetLowering(TargetMachine &tm,TargetLoweringObjectFile *tlof)
464 : TM(tm), TD(TM.getTargetData()), TLOF(*tlof) {
465 // All operations default to being supported.
466 memset(OpActions, 0, sizeof(OpActions));
467 memset(LoadExtActions, 0, sizeof(LoadExtActions));
468 memset(TruncStoreActions, 0, sizeof(TruncStoreActions));
469 memset(IndexedModeActions, 0, sizeof(IndexedModeActions));
470 memset(ConvertActions, 0, sizeof(ConvertActions));
471 memset(CondCodeActions, 0, sizeof(CondCodeActions));
473 // Set default actions for various operations.
474 for (unsigned VT = 0; VT != (unsigned)MVT::LAST_VALUETYPE; ++VT) {
475 // Default all indexed load / store to expand.
476 for (unsigned IM = (unsigned)ISD::PRE_INC;
477 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
478 setIndexedLoadAction(IM, (MVT::SimpleValueType)VT, Expand);
479 setIndexedStoreAction(IM, (MVT::SimpleValueType)VT, Expand);
482 // These operations default to expand.
483 setOperationAction(ISD::FGETSIGN, (MVT::SimpleValueType)VT, Expand);
484 setOperationAction(ISD::CONCAT_VECTORS, (MVT::SimpleValueType)VT, Expand);
487 // Most targets ignore the @llvm.prefetch intrinsic.
488 setOperationAction(ISD::PREFETCH, MVT::Other, Expand);
490 // ConstantFP nodes default to expand. Targets can either change this to
491 // Legal, in which case all fp constants are legal, or use isFPImmLegal()
492 // to optimize expansions for certain constants.
493 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
494 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
495 setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
497 // These library functions default to expand.
498 setOperationAction(ISD::FLOG , MVT::f64, Expand);
499 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
500 setOperationAction(ISD::FLOG10,MVT::f64, Expand);
501 setOperationAction(ISD::FEXP , MVT::f64, Expand);
502 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
503 setOperationAction(ISD::FLOG , MVT::f32, Expand);
504 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
505 setOperationAction(ISD::FLOG10,MVT::f32, Expand);
506 setOperationAction(ISD::FEXP , MVT::f32, Expand);
507 setOperationAction(ISD::FEXP2, MVT::f32, Expand);
509 // Default ISD::TRAP to expand (which turns it into abort).
510 setOperationAction(ISD::TRAP, MVT::Other, Expand);
512 IsLittleEndian = TD->isLittleEndian();
513 ShiftAmountTy = PointerTy = MVT::getIntegerVT(8*TD->getPointerSize());
514 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE*sizeof(TargetRegisterClass*));
515 memset(TargetDAGCombineArray, 0, array_lengthof(TargetDAGCombineArray));
516 maxStoresPerMemset = maxStoresPerMemcpy = maxStoresPerMemmove = 8;
517 benefitFromCodePlacementOpt = false;
518 UseUnderscoreSetJmp = false;
519 UseUnderscoreLongJmp = false;
520 SelectIsExpensive = false;
521 IntDivIsCheap = false;
522 Pow2DivIsCheap = false;
523 StackPointerRegisterToSaveRestore = 0;
524 ExceptionPointerRegister = 0;
525 ExceptionSelectorRegister = 0;
526 BooleanContents = UndefinedBooleanContent;
527 SchedPreferenceInfo = SchedulingForLatency;
529 JumpBufAlignment = 0;
530 IfCvtBlockSizeLimit = 2;
531 IfCvtDupBlockSizeLimit = 0;
532 PrefLoopAlignment = 0;
534 InitLibcallNames(LibcallRoutineNames);
535 InitCmpLibcallCCs(CmpLibcallCCs);
536 InitLibcallCallingConvs(LibcallCallingConvs);
539 TargetLowering::~TargetLowering() {
543 /// canOpTrap - Returns true if the operation can trap for the value type.
544 /// VT must be a legal type.
545 bool TargetLowering::canOpTrap(unsigned Op, EVT VT) const {
546 assert(isTypeLegal(VT));
561 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT,
562 unsigned &NumIntermediates,
564 TargetLowering* TLI) {
565 // Figure out the right, legal destination reg to copy into.
566 unsigned NumElts = VT.getVectorNumElements();
567 MVT EltTy = VT.getVectorElementType();
569 unsigned NumVectorRegs = 1;
571 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
572 // could break down into LHS/RHS like LegalizeDAG does.
573 if (!isPowerOf2_32(NumElts)) {
574 NumVectorRegs = NumElts;
578 // Divide the input until we get to a supported size. This will always
579 // end with a scalar if the target doesn't support vectors.
580 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) {
585 NumIntermediates = NumVectorRegs;
587 MVT NewVT = MVT::getVectorVT(EltTy, NumElts);
588 if (!TLI->isTypeLegal(NewVT))
590 IntermediateVT = NewVT;
592 EVT DestVT = TLI->getRegisterType(NewVT);
594 if (EVT(DestVT).bitsLT(NewVT)) {
595 // Value is expanded, e.g. i64 -> i16.
596 return NumVectorRegs*(NewVT.getSizeInBits()/DestVT.getSizeInBits());
598 // Otherwise, promotion or legal types use the same number of registers as
599 // the vector decimated to the appropriate level.
600 return NumVectorRegs;
606 /// computeRegisterProperties - Once all of the register classes are added,
607 /// this allows us to compute derived properties we expose.
608 void TargetLowering::computeRegisterProperties() {
609 assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE &&
610 "Too many value types for ValueTypeActions to hold!");
612 // Everything defaults to needing one register.
613 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
614 NumRegistersForVT[i] = 1;
615 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i;
617 // ...except isVoid, which doesn't need any registers.
618 NumRegistersForVT[MVT::isVoid] = 0;
620 // Find the largest integer register class.
621 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE;
622 for (; RegClassForVT[LargestIntReg] == 0; --LargestIntReg)
623 assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
625 // Every integer value type larger than this largest register takes twice as
626 // many registers to represent as the previous ValueType.
627 for (unsigned ExpandedReg = LargestIntReg + 1; ; ++ExpandedReg) {
628 EVT ExpandedVT = (MVT::SimpleValueType)ExpandedReg;
629 if (!ExpandedVT.isInteger())
631 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
632 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg;
633 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1);
634 ValueTypeActions.setTypeAction(ExpandedVT, Expand);
637 // Inspect all of the ValueType's smaller than the largest integer
638 // register to see which ones need promotion.
639 unsigned LegalIntReg = LargestIntReg;
640 for (unsigned IntReg = LargestIntReg - 1;
641 IntReg >= (unsigned)MVT::i1; --IntReg) {
642 EVT IVT = (MVT::SimpleValueType)IntReg;
643 if (isTypeLegal(IVT)) {
644 LegalIntReg = IntReg;
646 RegisterTypeForVT[IntReg] = TransformToType[IntReg] =
647 (MVT::SimpleValueType)LegalIntReg;
648 ValueTypeActions.setTypeAction(IVT, Promote);
652 // ppcf128 type is really two f64's.
653 if (!isTypeLegal(MVT::ppcf128)) {
654 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64];
655 RegisterTypeForVT[MVT::ppcf128] = MVT::f64;
656 TransformToType[MVT::ppcf128] = MVT::f64;
657 ValueTypeActions.setTypeAction(MVT::ppcf128, Expand);
660 // Decide how to handle f64. If the target does not have native f64 support,
661 // expand it to i64 and we will be generating soft float library calls.
662 if (!isTypeLegal(MVT::f64)) {
663 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
664 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
665 TransformToType[MVT::f64] = MVT::i64;
666 ValueTypeActions.setTypeAction(MVT::f64, Expand);
669 // Decide how to handle f32. If the target does not have native support for
670 // f32, promote it to f64 if it is legal. Otherwise, expand it to i32.
671 if (!isTypeLegal(MVT::f32)) {
672 if (isTypeLegal(MVT::f64)) {
673 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::f64];
674 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::f64];
675 TransformToType[MVT::f32] = MVT::f64;
676 ValueTypeActions.setTypeAction(MVT::f32, Promote);
678 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
679 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
680 TransformToType[MVT::f32] = MVT::i32;
681 ValueTypeActions.setTypeAction(MVT::f32, Expand);
685 // Loop over all of the vector value types to see which need transformations.
686 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
687 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
688 MVT VT = (MVT::SimpleValueType)i;
689 if (!isTypeLegal(VT)) {
692 unsigned NumIntermediates;
693 NumRegistersForVT[i] =
694 getVectorTypeBreakdownMVT(VT, IntermediateVT, NumIntermediates,
696 RegisterTypeForVT[i] = RegisterVT;
698 // Determine if there is a legal wider type.
699 bool IsLegalWiderType = false;
700 EVT EltVT = VT.getVectorElementType();
701 unsigned NElts = VT.getVectorNumElements();
702 for (unsigned nVT = i+1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
703 EVT SVT = (MVT::SimpleValueType)nVT;
704 if (isTypeLegal(SVT) && SVT.getVectorElementType() == EltVT &&
705 SVT.getVectorNumElements() > NElts && NElts != 1) {
706 TransformToType[i] = SVT;
707 ValueTypeActions.setTypeAction(VT, Promote);
708 IsLegalWiderType = true;
712 if (!IsLegalWiderType) {
713 EVT NVT = VT.getPow2VectorType();
715 // Type is already a power of 2. The default action is to split.
716 TransformToType[i] = MVT::Other;
717 ValueTypeActions.setTypeAction(VT, Expand);
719 TransformToType[i] = NVT;
720 ValueTypeActions.setTypeAction(VT, Promote);
727 const char *TargetLowering::getTargetNodeName(unsigned Opcode) const {
732 MVT::SimpleValueType TargetLowering::getSetCCResultType(EVT VT) const {
733 return PointerTy.SimpleTy;
736 MVT::SimpleValueType TargetLowering::getCmpLibcallReturnType() const {
737 return MVT::i32; // return the default value
740 /// getVectorTypeBreakdown - Vector types are broken down into some number of
741 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
742 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
743 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
745 /// This method returns the number of registers needed, and the VT for each
746 /// register. It also returns the VT and quantity of the intermediate values
747 /// before they are promoted/expanded.
749 unsigned TargetLowering::getVectorTypeBreakdown(LLVMContext &Context, EVT VT,
751 unsigned &NumIntermediates,
752 EVT &RegisterVT) const {
753 // Figure out the right, legal destination reg to copy into.
754 unsigned NumElts = VT.getVectorNumElements();
755 EVT EltTy = VT.getVectorElementType();
757 unsigned NumVectorRegs = 1;
759 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
760 // could break down into LHS/RHS like LegalizeDAG does.
761 if (!isPowerOf2_32(NumElts)) {
762 NumVectorRegs = NumElts;
766 // Divide the input until we get to a supported size. This will always
767 // end with a scalar if the target doesn't support vectors.
768 while (NumElts > 1 && !isTypeLegal(
769 EVT::getVectorVT(Context, EltTy, NumElts))) {
774 NumIntermediates = NumVectorRegs;
776 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts);
777 if (!isTypeLegal(NewVT))
779 IntermediateVT = NewVT;
781 EVT DestVT = getRegisterType(Context, NewVT);
783 if (DestVT.bitsLT(NewVT)) {
784 // Value is expanded, e.g. i64 -> i16.
785 return NumVectorRegs*(NewVT.getSizeInBits()/DestVT.getSizeInBits());
787 // Otherwise, promotion or legal types use the same number of registers as
788 // the vector decimated to the appropriate level.
789 return NumVectorRegs;
795 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
796 /// function arguments in the caller parameter area. This is the actual
797 /// alignment, not its logarithm.
798 unsigned TargetLowering::getByValTypeAlignment(const Type *Ty) const {
799 return TD->getCallFrameTypeAlignment(Ty);
802 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
803 /// current function. The returned value is a member of the
804 /// MachineJumpTableInfo::JTEntryKind enum.
805 unsigned TargetLowering::getJumpTableEncoding() const {
806 // In non-pic modes, just use the address of a block.
807 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
808 return MachineJumpTableInfo::EK_BlockAddress;
810 // In PIC mode, if the target supports a GPRel32 directive, use it.
811 if (getTargetMachine().getMCAsmInfo()->getGPRel32Directive() != 0)
812 return MachineJumpTableInfo::EK_GPRel32BlockAddress;
814 // Otherwise, use a label difference.
815 return MachineJumpTableInfo::EK_LabelDifference32;
818 SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table,
819 SelectionDAG &DAG) const {
820 // If our PIC model is GP relative, use the global offset table as the base.
821 if (getJumpTableEncoding() == MachineJumpTableInfo::EK_GPRel32BlockAddress)
822 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy());
826 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
827 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
830 TargetLowering::getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
831 unsigned JTI,MCContext &Ctx) const{
832 // The normal PIC reloc base is the label at the start of the jump table.
833 return MCSymbolRefExpr::Create(MF->getJTISymbol(JTI, Ctx), Ctx);
837 TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
838 // Assume that everything is safe in static mode.
839 if (getTargetMachine().getRelocationModel() == Reloc::Static)
842 // In dynamic-no-pic mode, assume that known defined values are safe.
843 if (getTargetMachine().getRelocationModel() == Reloc::DynamicNoPIC &&
845 !GA->getGlobal()->isDeclaration() &&
846 !GA->getGlobal()->isWeakForLinker())
849 // Otherwise assume nothing is safe.
853 //===----------------------------------------------------------------------===//
854 // Optimization Methods
855 //===----------------------------------------------------------------------===//
857 /// ShrinkDemandedConstant - Check to see if the specified operand of the
858 /// specified instruction is a constant integer. If so, check to see if there
859 /// are any bits set in the constant that are not demanded. If so, shrink the
860 /// constant and return true.
861 bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDValue Op,
862 const APInt &Demanded) {
863 DebugLoc dl = Op.getDebugLoc();
865 // FIXME: ISD::SELECT, ISD::SELECT_CC
866 switch (Op.getOpcode()) {
871 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
872 if (!C) return false;
874 if (Op.getOpcode() == ISD::XOR &&
875 (C->getAPIntValue() | (~Demanded)).isAllOnesValue())
878 // if we can expand it to have all bits set, do it
879 if (C->getAPIntValue().intersects(~Demanded)) {
880 EVT VT = Op.getValueType();
881 SDValue New = DAG.getNode(Op.getOpcode(), dl, VT, Op.getOperand(0),
882 DAG.getConstant(Demanded &
885 return CombineTo(Op, New);
895 /// ShrinkDemandedOp - Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the
896 /// casts are free. This uses isZExtFree and ZERO_EXTEND for the widening
897 /// cast, but it could be generalized for targets with other types of
898 /// implicit widening casts.
900 TargetLowering::TargetLoweringOpt::ShrinkDemandedOp(SDValue Op,
902 const APInt &Demanded,
904 assert(Op.getNumOperands() == 2 &&
905 "ShrinkDemandedOp only supports binary operators!");
906 assert(Op.getNode()->getNumValues() == 1 &&
907 "ShrinkDemandedOp only supports nodes with one result!");
909 // Don't do this if the node has another user, which may require the
911 if (!Op.getNode()->hasOneUse())
914 // Search for the smallest integer type with free casts to and from
915 // Op's type. For expedience, just check power-of-2 integer types.
916 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
917 unsigned SmallVTBits = BitWidth - Demanded.countLeadingZeros();
918 if (!isPowerOf2_32(SmallVTBits))
919 SmallVTBits = NextPowerOf2(SmallVTBits);
920 for (; SmallVTBits < BitWidth; SmallVTBits = NextPowerOf2(SmallVTBits)) {
921 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), SmallVTBits);
922 if (TLI.isTruncateFree(Op.getValueType(), SmallVT) &&
923 TLI.isZExtFree(SmallVT, Op.getValueType())) {
924 // We found a type with free casts.
925 SDValue X = DAG.getNode(Op.getOpcode(), dl, SmallVT,
926 DAG.getNode(ISD::TRUNCATE, dl, SmallVT,
927 Op.getNode()->getOperand(0)),
928 DAG.getNode(ISD::TRUNCATE, dl, SmallVT,
929 Op.getNode()->getOperand(1)));
930 SDValue Z = DAG.getNode(ISD::ZERO_EXTEND, dl, Op.getValueType(), X);
931 return CombineTo(Op, Z);
937 /// SimplifyDemandedBits - Look at Op. At this point, we know that only the
938 /// DemandedMask bits of the result of Op are ever used downstream. If we can
939 /// use this information to simplify Op, create a new simplified DAG node and
940 /// return true, returning the original and new nodes in Old and New. Otherwise,
941 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
942 /// the expression (used to simplify the caller). The KnownZero/One bits may
943 /// only be accurate for those bits in the DemandedMask.
944 bool TargetLowering::SimplifyDemandedBits(SDValue Op,
945 const APInt &DemandedMask,
948 TargetLoweringOpt &TLO,
949 unsigned Depth) const {
950 unsigned BitWidth = DemandedMask.getBitWidth();
951 assert(Op.getValueType().getScalarType().getSizeInBits() == BitWidth &&
952 "Mask size mismatches value type size!");
953 APInt NewMask = DemandedMask;
954 DebugLoc dl = Op.getDebugLoc();
956 // Don't know anything.
957 KnownZero = KnownOne = APInt(BitWidth, 0);
959 // Other users may use these bits.
960 if (!Op.getNode()->hasOneUse()) {
962 // If not at the root, Just compute the KnownZero/KnownOne bits to
963 // simplify things downstream.
964 TLO.DAG.ComputeMaskedBits(Op, DemandedMask, KnownZero, KnownOne, Depth);
967 // If this is the root being simplified, allow it to have multiple uses,
968 // just set the NewMask to all bits.
969 NewMask = APInt::getAllOnesValue(BitWidth);
970 } else if (DemandedMask == 0) {
971 // Not demanding any bits from Op.
972 if (Op.getOpcode() != ISD::UNDEF)
973 return TLO.CombineTo(Op, TLO.DAG.getUNDEF(Op.getValueType()));
975 } else if (Depth == 6) { // Limit search depth.
979 APInt KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut;
980 switch (Op.getOpcode()) {
982 // We know all of the bits for a constant!
983 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue() & NewMask;
984 KnownZero = ~KnownOne & NewMask;
985 return false; // Don't fall through, will infinitely loop.
987 // If the RHS is a constant, check to see if the LHS would be zero without
988 // using the bits from the RHS. Below, we use knowledge about the RHS to
989 // simplify the LHS, here we're using information from the LHS to simplify
991 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
992 APInt LHSZero, LHSOne;
993 TLO.DAG.ComputeMaskedBits(Op.getOperand(0), NewMask,
994 LHSZero, LHSOne, Depth+1);
995 // If the LHS already has zeros where RHSC does, this and is dead.
996 if ((LHSZero & NewMask) == (~RHSC->getAPIntValue() & NewMask))
997 return TLO.CombineTo(Op, Op.getOperand(0));
998 // If any of the set bits in the RHS are known zero on the LHS, shrink
1000 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & NewMask))
1004 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
1005 KnownOne, TLO, Depth+1))
1007 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1008 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownZero & NewMask,
1009 KnownZero2, KnownOne2, TLO, Depth+1))
1011 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1013 // If all of the demanded bits are known one on one side, return the other.
1014 // These bits cannot contribute to the result of the 'and'.
1015 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
1016 return TLO.CombineTo(Op, Op.getOperand(0));
1017 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
1018 return TLO.CombineTo(Op, Op.getOperand(1));
1019 // If all of the demanded bits in the inputs are known zeros, return zero.
1020 if ((NewMask & (KnownZero|KnownZero2)) == NewMask)
1021 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, Op.getValueType()));
1022 // If the RHS is a constant, see if we can simplify it.
1023 if (TLO.ShrinkDemandedConstant(Op, ~KnownZero2 & NewMask))
1025 // If the operation can be done in a smaller type, do so.
1026 if (TLO.ShrinkOps && TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1029 // Output known-1 bits are only known if set in both the LHS & RHS.
1030 KnownOne &= KnownOne2;
1031 // Output known-0 are known to be clear if zero in either the LHS | RHS.
1032 KnownZero |= KnownZero2;
1035 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
1036 KnownOne, TLO, Depth+1))
1038 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1039 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownOne & NewMask,
1040 KnownZero2, KnownOne2, TLO, Depth+1))
1042 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1044 // If all of the demanded bits are known zero on one side, return the other.
1045 // These bits cannot contribute to the result of the 'or'.
1046 if ((NewMask & ~KnownOne2 & KnownZero) == (~KnownOne2 & NewMask))
1047 return TLO.CombineTo(Op, Op.getOperand(0));
1048 if ((NewMask & ~KnownOne & KnownZero2) == (~KnownOne & NewMask))
1049 return TLO.CombineTo(Op, Op.getOperand(1));
1050 // If all of the potentially set bits on one side are known to be set on
1051 // the other side, just use the 'other' side.
1052 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
1053 return TLO.CombineTo(Op, Op.getOperand(0));
1054 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
1055 return TLO.CombineTo(Op, Op.getOperand(1));
1056 // If the RHS is a constant, see if we can simplify it.
1057 if (TLO.ShrinkDemandedConstant(Op, NewMask))
1059 // If the operation can be done in a smaller type, do so.
1060 if (TLO.ShrinkOps && TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1063 // Output known-0 bits are only known if clear in both the LHS & RHS.
1064 KnownZero &= KnownZero2;
1065 // Output known-1 are known to be set if set in either the LHS | RHS.
1066 KnownOne |= KnownOne2;
1069 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
1070 KnownOne, TLO, Depth+1))
1072 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1073 if (SimplifyDemandedBits(Op.getOperand(0), NewMask, KnownZero2,
1074 KnownOne2, TLO, Depth+1))
1076 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1078 // If all of the demanded bits are known zero on one side, return the other.
1079 // These bits cannot contribute to the result of the 'xor'.
1080 if ((KnownZero & NewMask) == NewMask)
1081 return TLO.CombineTo(Op, Op.getOperand(0));
1082 if ((KnownZero2 & NewMask) == NewMask)
1083 return TLO.CombineTo(Op, Op.getOperand(1));
1084 // If the operation can be done in a smaller type, do so.
1085 if (TLO.ShrinkOps && TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1088 // If all of the unknown bits are known to be zero on one side or the other
1089 // (but not both) turn this into an *inclusive* or.
1090 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1091 if ((NewMask & ~KnownZero & ~KnownZero2) == 0)
1092 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, dl, Op.getValueType(),
1096 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1097 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
1098 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1099 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
1101 // If all of the demanded bits on one side are known, and all of the set
1102 // bits on that side are also known to be set on the other side, turn this
1103 // into an AND, as we know the bits will be cleared.
1104 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1105 if ((NewMask & (KnownZero|KnownOne)) == NewMask) { // all known
1106 if ((KnownOne & KnownOne2) == KnownOne) {
1107 EVT VT = Op.getValueType();
1108 SDValue ANDC = TLO.DAG.getConstant(~KnownOne & NewMask, VT);
1109 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, dl, VT,
1110 Op.getOperand(0), ANDC));
1114 // If the RHS is a constant, see if we can simplify it.
1115 // for XOR, we prefer to force bits to 1 if they will make a -1.
1116 // if we can't force bits, try to shrink constant
1117 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1118 APInt Expanded = C->getAPIntValue() | (~NewMask);
1119 // if we can expand it to have all bits set, do it
1120 if (Expanded.isAllOnesValue()) {
1121 if (Expanded != C->getAPIntValue()) {
1122 EVT VT = Op.getValueType();
1123 SDValue New = TLO.DAG.getNode(Op.getOpcode(), dl,VT, Op.getOperand(0),
1124 TLO.DAG.getConstant(Expanded, VT));
1125 return TLO.CombineTo(Op, New);
1127 // if it already has all the bits set, nothing to change
1128 // but don't shrink either!
1129 } else if (TLO.ShrinkDemandedConstant(Op, NewMask)) {
1134 KnownZero = KnownZeroOut;
1135 KnownOne = KnownOneOut;
1138 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero,
1139 KnownOne, TLO, Depth+1))
1141 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero2,
1142 KnownOne2, TLO, Depth+1))
1144 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1145 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1147 // If the operands are constants, see if we can simplify them.
1148 if (TLO.ShrinkDemandedConstant(Op, NewMask))
1151 // Only known if known in both the LHS and RHS.
1152 KnownOne &= KnownOne2;
1153 KnownZero &= KnownZero2;
1155 case ISD::SELECT_CC:
1156 if (SimplifyDemandedBits(Op.getOperand(3), NewMask, KnownZero,
1157 KnownOne, TLO, Depth+1))
1159 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero2,
1160 KnownOne2, TLO, Depth+1))
1162 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1163 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1165 // If the operands are constants, see if we can simplify them.
1166 if (TLO.ShrinkDemandedConstant(Op, NewMask))
1169 // Only known if known in both the LHS and RHS.
1170 KnownOne &= KnownOne2;
1171 KnownZero &= KnownZero2;
1174 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1175 unsigned ShAmt = SA->getZExtValue();
1176 SDValue InOp = Op.getOperand(0);
1178 // If the shift count is an invalid immediate, don't do anything.
1179 if (ShAmt >= BitWidth)
1182 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a
1183 // single shift. We can do this if the bottom bits (which are shifted
1184 // out) are never demanded.
1185 if (InOp.getOpcode() == ISD::SRL &&
1186 isa<ConstantSDNode>(InOp.getOperand(1))) {
1187 if (ShAmt && (NewMask & APInt::getLowBitsSet(BitWidth, ShAmt)) == 0) {
1188 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
1189 unsigned Opc = ISD::SHL;
1190 int Diff = ShAmt-C1;
1197 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
1198 EVT VT = Op.getValueType();
1199 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT,
1200 InOp.getOperand(0), NewSA));
1204 if (SimplifyDemandedBits(Op.getOperand(0), NewMask.lshr(ShAmt),
1205 KnownZero, KnownOne, TLO, Depth+1))
1207 KnownZero <<= SA->getZExtValue();
1208 KnownOne <<= SA->getZExtValue();
1209 // low bits known zero.
1210 KnownZero |= APInt::getLowBitsSet(BitWidth, SA->getZExtValue());
1214 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1215 EVT VT = Op.getValueType();
1216 unsigned ShAmt = SA->getZExtValue();
1217 unsigned VTSize = VT.getSizeInBits();
1218 SDValue InOp = Op.getOperand(0);
1220 // If the shift count is an invalid immediate, don't do anything.
1221 if (ShAmt >= BitWidth)
1224 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a
1225 // single shift. We can do this if the top bits (which are shifted out)
1226 // are never demanded.
1227 if (InOp.getOpcode() == ISD::SHL &&
1228 isa<ConstantSDNode>(InOp.getOperand(1))) {
1229 if (ShAmt && (NewMask & APInt::getHighBitsSet(VTSize, ShAmt)) == 0) {
1230 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
1231 unsigned Opc = ISD::SRL;
1232 int Diff = ShAmt-C1;
1239 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
1240 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT,
1241 InOp.getOperand(0), NewSA));
1245 // Compute the new bits that are at the top now.
1246 if (SimplifyDemandedBits(InOp, (NewMask << ShAmt),
1247 KnownZero, KnownOne, TLO, Depth+1))
1249 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1250 KnownZero = KnownZero.lshr(ShAmt);
1251 KnownOne = KnownOne.lshr(ShAmt);
1253 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
1254 KnownZero |= HighBits; // High bits known zero.
1258 // If this is an arithmetic shift right and only the low-bit is set, we can
1259 // always convert this into a logical shr, even if the shift amount is
1260 // variable. The low bit of the shift cannot be an input sign bit unless
1261 // the shift amount is >= the size of the datatype, which is undefined.
1262 if (DemandedMask == 1)
1263 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, Op.getValueType(),
1264 Op.getOperand(0), Op.getOperand(1)));
1266 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1267 EVT VT = Op.getValueType();
1268 unsigned ShAmt = SA->getZExtValue();
1270 // If the shift count is an invalid immediate, don't do anything.
1271 if (ShAmt >= BitWidth)
1274 APInt InDemandedMask = (NewMask << ShAmt);
1276 // If any of the demanded bits are produced by the sign extension, we also
1277 // demand the input sign bit.
1278 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
1279 if (HighBits.intersects(NewMask))
1280 InDemandedMask |= APInt::getSignBit(VT.getScalarType().getSizeInBits());
1282 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask,
1283 KnownZero, KnownOne, TLO, Depth+1))
1285 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1286 KnownZero = KnownZero.lshr(ShAmt);
1287 KnownOne = KnownOne.lshr(ShAmt);
1289 // Handle the sign bit, adjusted to where it is now in the mask.
1290 APInt SignBit = APInt::getSignBit(BitWidth).lshr(ShAmt);
1292 // If the input sign bit is known to be zero, or if none of the top bits
1293 // are demanded, turn this into an unsigned shift right.
1294 if (KnownZero.intersects(SignBit) || (HighBits & ~NewMask) == HighBits) {
1295 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, VT,
1298 } else if (KnownOne.intersects(SignBit)) { // New bits are known one.
1299 KnownOne |= HighBits;
1303 case ISD::SIGN_EXTEND_INREG: {
1304 EVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1306 // Sign extension. Compute the demanded bits in the result that are not
1307 // present in the input.
1309 APInt::getHighBitsSet(BitWidth,
1310 BitWidth - EVT.getScalarType().getSizeInBits()) &
1313 // If none of the extended bits are demanded, eliminate the sextinreg.
1315 return TLO.CombineTo(Op, Op.getOperand(0));
1317 APInt InSignBit = APInt::getSignBit(EVT.getScalarType().getSizeInBits());
1318 InSignBit.zext(BitWidth);
1319 APInt InputDemandedBits =
1320 APInt::getLowBitsSet(BitWidth,
1321 EVT.getScalarType().getSizeInBits()) &
1324 // Since the sign extended bits are demanded, we know that the sign
1326 InputDemandedBits |= InSignBit;
1328 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits,
1329 KnownZero, KnownOne, TLO, Depth+1))
1331 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1333 // If the sign bit of the input is known set or clear, then we know the
1334 // top bits of the result.
1336 // If the input sign bit is known zero, convert this into a zero extension.
1337 if (KnownZero.intersects(InSignBit))
1338 return TLO.CombineTo(Op,
1339 TLO.DAG.getZeroExtendInReg(Op.getOperand(0),dl,EVT));
1341 if (KnownOne.intersects(InSignBit)) { // Input sign bit known set
1342 KnownOne |= NewBits;
1343 KnownZero &= ~NewBits;
1344 } else { // Input sign bit unknown
1345 KnownZero &= ~NewBits;
1346 KnownOne &= ~NewBits;
1350 case ISD::ZERO_EXTEND: {
1351 unsigned OperandBitWidth =
1352 Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
1353 APInt InMask = NewMask;
1354 InMask.trunc(OperandBitWidth);
1356 // If none of the top bits are demanded, convert this into an any_extend.
1358 APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask;
1359 if (!NewBits.intersects(NewMask))
1360 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND, dl,
1364 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1365 KnownZero, KnownOne, TLO, Depth+1))
1367 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1368 KnownZero.zext(BitWidth);
1369 KnownOne.zext(BitWidth);
1370 KnownZero |= NewBits;
1373 case ISD::SIGN_EXTEND: {
1374 EVT InVT = Op.getOperand(0).getValueType();
1375 unsigned InBits = InVT.getScalarType().getSizeInBits();
1376 APInt InMask = APInt::getLowBitsSet(BitWidth, InBits);
1377 APInt InSignBit = APInt::getBitsSet(BitWidth, InBits - 1, InBits);
1378 APInt NewBits = ~InMask & NewMask;
1380 // If none of the top bits are demanded, convert this into an any_extend.
1382 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND, dl,
1386 // Since some of the sign extended bits are demanded, we know that the sign
1388 APInt InDemandedBits = InMask & NewMask;
1389 InDemandedBits |= InSignBit;
1390 InDemandedBits.trunc(InBits);
1392 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero,
1393 KnownOne, TLO, Depth+1))
1395 KnownZero.zext(BitWidth);
1396 KnownOne.zext(BitWidth);
1398 // If the sign bit is known zero, convert this to a zero extend.
1399 if (KnownZero.intersects(InSignBit))
1400 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND, dl,
1404 // If the sign bit is known one, the top bits match.
1405 if (KnownOne.intersects(InSignBit)) {
1406 KnownOne |= NewBits;
1407 KnownZero &= ~NewBits;
1408 } else { // Otherwise, top bits aren't known.
1409 KnownOne &= ~NewBits;
1410 KnownZero &= ~NewBits;
1414 case ISD::ANY_EXTEND: {
1415 unsigned OperandBitWidth =
1416 Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
1417 APInt InMask = NewMask;
1418 InMask.trunc(OperandBitWidth);
1419 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1420 KnownZero, KnownOne, TLO, Depth+1))
1422 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1423 KnownZero.zext(BitWidth);
1424 KnownOne.zext(BitWidth);
1427 case ISD::TRUNCATE: {
1428 // Simplify the input, using demanded bit information, and compute the known
1429 // zero/one bits live out.
1430 unsigned OperandBitWidth =
1431 Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
1432 APInt TruncMask = NewMask;
1433 TruncMask.zext(OperandBitWidth);
1434 if (SimplifyDemandedBits(Op.getOperand(0), TruncMask,
1435 KnownZero, KnownOne, TLO, Depth+1))
1437 KnownZero.trunc(BitWidth);
1438 KnownOne.trunc(BitWidth);
1440 // If the input is only used by this truncate, see if we can shrink it based
1441 // on the known demanded bits.
1442 if (Op.getOperand(0).getNode()->hasOneUse()) {
1443 SDValue In = Op.getOperand(0);
1444 switch (In.getOpcode()) {
1447 // Shrink SRL by a constant if none of the high bits shifted in are
1449 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1))){
1450 APInt HighBits = APInt::getHighBitsSet(OperandBitWidth,
1451 OperandBitWidth - BitWidth);
1452 HighBits = HighBits.lshr(ShAmt->getZExtValue());
1453 HighBits.trunc(BitWidth);
1455 if (ShAmt->getZExtValue() < BitWidth && !(HighBits & NewMask)) {
1456 // None of the shifted in bits are needed. Add a truncate of the
1457 // shift input, then shift it.
1458 SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE, dl,
1461 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl,
1471 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1474 case ISD::AssertZext: {
1475 EVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1476 APInt InMask = APInt::getLowBitsSet(BitWidth,
1477 VT.getSizeInBits());
1478 if (SimplifyDemandedBits(Op.getOperand(0), InMask & NewMask,
1479 KnownZero, KnownOne, TLO, Depth+1))
1481 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1482 KnownZero |= ~InMask & NewMask;
1485 case ISD::BIT_CONVERT:
1487 // If this is an FP->Int bitcast and if the sign bit is the only thing that
1488 // is demanded, turn this into a FGETSIGN.
1489 if (NewMask == EVT::getIntegerVTSignBit(Op.getValueType()) &&
1490 MVT::isFloatingPoint(Op.getOperand(0).getValueType()) &&
1491 !MVT::isVector(Op.getOperand(0).getValueType())) {
1492 // Only do this xform if FGETSIGN is valid or if before legalize.
1493 if (!TLO.AfterLegalize ||
1494 isOperationLegal(ISD::FGETSIGN, Op.getValueType())) {
1495 // Make a FGETSIGN + SHL to move the sign bit into the appropriate
1496 // place. We expect the SHL to be eliminated by other optimizations.
1497 SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, Op.getValueType(),
1499 unsigned ShVal = Op.getValueType().getSizeInBits()-1;
1500 SDValue ShAmt = TLO.DAG.getConstant(ShVal, getShiftAmountTy());
1501 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, Op.getValueType(),
1510 // Add, Sub, and Mul don't demand any bits in positions beyond that
1511 // of the highest bit demanded of them.
1512 APInt LoMask = APInt::getLowBitsSet(BitWidth,
1513 BitWidth - NewMask.countLeadingZeros());
1514 if (SimplifyDemandedBits(Op.getOperand(0), LoMask, KnownZero2,
1515 KnownOne2, TLO, Depth+1))
1517 if (SimplifyDemandedBits(Op.getOperand(1), LoMask, KnownZero2,
1518 KnownOne2, TLO, Depth+1))
1520 // See if the operation should be performed at a smaller bit width.
1521 if (TLO.ShrinkOps && TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1526 // Just use ComputeMaskedBits to compute output bits.
1527 TLO.DAG.ComputeMaskedBits(Op, NewMask, KnownZero, KnownOne, Depth);
1531 // If we know the value of all of the demanded bits, return this as a
1533 if ((NewMask & (KnownZero|KnownOne)) == NewMask)
1534 return TLO.CombineTo(Op, TLO.DAG.getConstant(KnownOne, Op.getValueType()));
1539 /// computeMaskedBitsForTargetNode - Determine which of the bits specified
1540 /// in Mask are known to be either zero or one and return them in the
1541 /// KnownZero/KnownOne bitsets.
1542 void TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
1546 const SelectionDAG &DAG,
1547 unsigned Depth) const {
1548 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1549 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1550 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1551 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1552 "Should use MaskedValueIsZero if you don't know whether Op"
1553 " is a target node!");
1554 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
1557 /// ComputeNumSignBitsForTargetNode - This method can be implemented by
1558 /// targets that want to expose additional information about sign bits to the
1560 unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
1561 unsigned Depth) const {
1562 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1563 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1564 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1565 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1566 "Should use ComputeNumSignBits if you don't know whether Op"
1567 " is a target node!");
1571 /// ValueHasExactlyOneBitSet - Test if the given value is known to have exactly
1572 /// one bit set. This differs from ComputeMaskedBits in that it doesn't need to
1573 /// determine which bit is set.
1575 static bool ValueHasExactlyOneBitSet(SDValue Val, const SelectionDAG &DAG) {
1576 // A left-shift of a constant one will have exactly one bit set, because
1577 // shifting the bit off the end is undefined.
1578 if (Val.getOpcode() == ISD::SHL)
1579 if (ConstantSDNode *C =
1580 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0)))
1581 if (C->getAPIntValue() == 1)
1584 // Similarly, a right-shift of a constant sign-bit will have exactly
1586 if (Val.getOpcode() == ISD::SRL)
1587 if (ConstantSDNode *C =
1588 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0)))
1589 if (C->getAPIntValue().isSignBit())
1592 // More could be done here, though the above checks are enough
1593 // to handle some common cases.
1595 // Fall back to ComputeMaskedBits to catch other known cases.
1596 EVT OpVT = Val.getValueType();
1597 unsigned BitWidth = OpVT.getScalarType().getSizeInBits();
1598 APInt Mask = APInt::getAllOnesValue(BitWidth);
1599 APInt KnownZero, KnownOne;
1600 DAG.ComputeMaskedBits(Val, Mask, KnownZero, KnownOne);
1601 return (KnownZero.countPopulation() == BitWidth - 1) &&
1602 (KnownOne.countPopulation() == 1);
1605 /// SimplifySetCC - Try to simplify a setcc built with the specified operands
1606 /// and cc. If it is unable to simplify it, return a null SDValue.
1608 TargetLowering::SimplifySetCC(EVT VT, SDValue N0, SDValue N1,
1609 ISD::CondCode Cond, bool foldBooleans,
1610 DAGCombinerInfo &DCI, DebugLoc dl) const {
1611 SelectionDAG &DAG = DCI.DAG;
1612 LLVMContext &Context = *DAG.getContext();
1614 // These setcc operations always fold.
1618 case ISD::SETFALSE2: return DAG.getConstant(0, VT);
1620 case ISD::SETTRUE2: return DAG.getConstant(1, VT);
1623 if (isa<ConstantSDNode>(N0.getNode())) {
1624 // Ensure that the constant occurs on the RHS, and fold constant
1626 return DAG.getSetCC(dl, VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
1629 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) {
1630 const APInt &C1 = N1C->getAPIntValue();
1632 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an
1633 // equality comparison, then we're just comparing whether X itself is
1635 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) &&
1636 N0.getOperand(0).getOpcode() == ISD::CTLZ &&
1637 N0.getOperand(1).getOpcode() == ISD::Constant) {
1639 = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
1640 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1641 ShAmt == Log2_32(N0.getValueType().getSizeInBits())) {
1642 if ((C1 == 0) == (Cond == ISD::SETEQ)) {
1643 // (srl (ctlz x), 5) == 0 -> X != 0
1644 // (srl (ctlz x), 5) != 1 -> X != 0
1647 // (srl (ctlz x), 5) != 0 -> X == 0
1648 // (srl (ctlz x), 5) == 1 -> X == 0
1651 SDValue Zero = DAG.getConstant(0, N0.getValueType());
1652 return DAG.getSetCC(dl, VT, N0.getOperand(0).getOperand(0),
1657 // If the LHS is '(and load, const)', the RHS is 0,
1658 // the test is for equality or unsigned, and all 1 bits of the const are
1659 // in the same partial word, see if we can shorten the load.
1660 if (DCI.isBeforeLegalize() &&
1661 N0.getOpcode() == ISD::AND && C1 == 0 &&
1662 N0.getNode()->hasOneUse() &&
1663 isa<LoadSDNode>(N0.getOperand(0)) &&
1664 N0.getOperand(0).getNode()->hasOneUse() &&
1665 isa<ConstantSDNode>(N0.getOperand(1))) {
1666 LoadSDNode *Lod = cast<LoadSDNode>(N0.getOperand(0));
1668 unsigned bestWidth = 0, bestOffset = 0;
1669 if (!Lod->isVolatile() && Lod->isUnindexed()) {
1670 unsigned origWidth = N0.getValueType().getSizeInBits();
1671 unsigned maskWidth = origWidth;
1672 // We can narrow (e.g.) 16-bit extending loads on 32-bit target to
1673 // 8 bits, but have to be careful...
1674 if (Lod->getExtensionType() != ISD::NON_EXTLOAD)
1675 origWidth = Lod->getMemoryVT().getSizeInBits();
1677 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
1678 for (unsigned width = origWidth / 2; width>=8; width /= 2) {
1679 APInt newMask = APInt::getLowBitsSet(maskWidth, width);
1680 for (unsigned offset=0; offset<origWidth/width; offset++) {
1681 if ((newMask & Mask) == Mask) {
1682 if (!TD->isLittleEndian())
1683 bestOffset = (origWidth/width - offset - 1) * (width/8);
1685 bestOffset = (uint64_t)offset * (width/8);
1686 bestMask = Mask.lshr(offset * (width/8) * 8);
1690 newMask = newMask << width;
1695 EVT newVT = EVT::getIntegerVT(Context, bestWidth);
1696 if (newVT.isRound()) {
1697 EVT PtrType = Lod->getOperand(1).getValueType();
1698 SDValue Ptr = Lod->getBasePtr();
1699 if (bestOffset != 0)
1700 Ptr = DAG.getNode(ISD::ADD, dl, PtrType, Lod->getBasePtr(),
1701 DAG.getConstant(bestOffset, PtrType));
1702 unsigned NewAlign = MinAlign(Lod->getAlignment(), bestOffset);
1703 SDValue NewLoad = DAG.getLoad(newVT, dl, Lod->getChain(), Ptr,
1705 Lod->getSrcValueOffset() + bestOffset,
1706 false, false, NewAlign);
1707 return DAG.getSetCC(dl, VT,
1708 DAG.getNode(ISD::AND, dl, newVT, NewLoad,
1709 DAG.getConstant(bestMask.trunc(bestWidth),
1711 DAG.getConstant(0LL, newVT), Cond);
1716 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
1717 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
1718 unsigned InSize = N0.getOperand(0).getValueType().getSizeInBits();
1720 // If the comparison constant has bits in the upper part, the
1721 // zero-extended value could never match.
1722 if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(),
1723 C1.getBitWidth() - InSize))) {
1727 case ISD::SETEQ: return DAG.getConstant(0, VT);
1730 case ISD::SETNE: return DAG.getConstant(1, VT);
1733 // True if the sign bit of C1 is set.
1734 return DAG.getConstant(C1.isNegative(), VT);
1737 // True if the sign bit of C1 isn't set.
1738 return DAG.getConstant(C1.isNonNegative(), VT);
1744 // Otherwise, we can perform the comparison with the low bits.
1752 EVT newVT = N0.getOperand(0).getValueType();
1753 if (DCI.isBeforeLegalizeOps() ||
1754 (isOperationLegal(ISD::SETCC, newVT) &&
1755 getCondCodeAction(Cond, newVT)==Legal))
1756 return DAG.getSetCC(dl, VT, N0.getOperand(0),
1757 DAG.getConstant(APInt(C1).trunc(InSize), newVT),
1762 break; // todo, be more careful with signed comparisons
1764 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
1765 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1766 EVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
1767 unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits();
1768 EVT ExtDstTy = N0.getValueType();
1769 unsigned ExtDstTyBits = ExtDstTy.getSizeInBits();
1771 // If the extended part has any inconsistent bits, it cannot ever
1772 // compare equal. In other words, they have to be all ones or all
1775 APInt::getHighBitsSet(ExtDstTyBits, ExtDstTyBits - ExtSrcTyBits);
1776 if ((C1 & ExtBits) != 0 && (C1 & ExtBits) != ExtBits)
1777 return DAG.getConstant(Cond == ISD::SETNE, VT);
1780 EVT Op0Ty = N0.getOperand(0).getValueType();
1781 if (Op0Ty == ExtSrcTy) {
1782 ZextOp = N0.getOperand(0);
1784 APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits);
1785 ZextOp = DAG.getNode(ISD::AND, dl, Op0Ty, N0.getOperand(0),
1786 DAG.getConstant(Imm, Op0Ty));
1788 if (!DCI.isCalledByLegalizer())
1789 DCI.AddToWorklist(ZextOp.getNode());
1790 // Otherwise, make this a use of a zext.
1791 return DAG.getSetCC(dl, VT, ZextOp,
1792 DAG.getConstant(C1 & APInt::getLowBitsSet(
1797 } else if ((N1C->isNullValue() || N1C->getAPIntValue() == 1) &&
1798 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1799 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC
1800 if (N0.getOpcode() == ISD::SETCC &&
1801 isTypeLegal(VT) && VT.bitsLE(N0.getValueType())) {
1802 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getAPIntValue() != 1);
1804 return DAG.getNode(ISD::TRUNCATE, dl, VT, N0);
1805 // Invert the condition.
1806 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
1807 CC = ISD::getSetCCInverse(CC,
1808 N0.getOperand(0).getValueType().isInteger());
1809 return DAG.getSetCC(dl, VT, N0.getOperand(0), N0.getOperand(1), CC);
1812 if ((N0.getOpcode() == ISD::XOR ||
1813 (N0.getOpcode() == ISD::AND &&
1814 N0.getOperand(0).getOpcode() == ISD::XOR &&
1815 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
1816 isa<ConstantSDNode>(N0.getOperand(1)) &&
1817 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue() == 1) {
1818 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We
1819 // can only do this if the top bits are known zero.
1820 unsigned BitWidth = N0.getValueSizeInBits();
1821 if (DAG.MaskedValueIsZero(N0,
1822 APInt::getHighBitsSet(BitWidth,
1824 // Okay, get the un-inverted input value.
1826 if (N0.getOpcode() == ISD::XOR)
1827 Val = N0.getOperand(0);
1829 assert(N0.getOpcode() == ISD::AND &&
1830 N0.getOperand(0).getOpcode() == ISD::XOR);
1831 // ((X^1)&1)^1 -> X & 1
1832 Val = DAG.getNode(ISD::AND, dl, N0.getValueType(),
1833 N0.getOperand(0).getOperand(0),
1837 return DAG.getSetCC(dl, VT, Val, N1,
1838 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
1840 } else if (N1C->getAPIntValue() == 1 &&
1842 getBooleanContents() == ZeroOrOneBooleanContent)) {
1844 if (Op0.getOpcode() == ISD::TRUNCATE)
1845 Op0 = Op0.getOperand(0);
1847 if ((Op0.getOpcode() == ISD::XOR) &&
1848 Op0.getOperand(0).getOpcode() == ISD::SETCC &&
1849 Op0.getOperand(1).getOpcode() == ISD::SETCC) {
1850 // (xor (setcc), (setcc)) == / != 1 -> (setcc) != / == (setcc)
1851 Cond = (Cond == ISD::SETEQ) ? ISD::SETNE : ISD::SETEQ;
1852 return DAG.getSetCC(dl, VT, Op0.getOperand(0), Op0.getOperand(1),
1854 } else if (Op0.getOpcode() == ISD::AND &&
1855 isa<ConstantSDNode>(Op0.getOperand(1)) &&
1856 cast<ConstantSDNode>(Op0.getOperand(1))->getAPIntValue() == 1) {
1857 // If this is (X&1) == / != 1, normalize it to (X&1) != / == 0.
1858 if (Op0.getValueType() != VT)
1859 Op0 = DAG.getNode(ISD::AND, dl, VT,
1860 DAG.getNode(ISD::TRUNCATE, dl, VT, Op0.getOperand(0)),
1861 DAG.getConstant(1, VT));
1862 return DAG.getSetCC(dl, VT, Op0,
1863 DAG.getConstant(0, Op0.getValueType()),
1864 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
1869 APInt MinVal, MaxVal;
1870 unsigned OperandBitSize = N1C->getValueType(0).getSizeInBits();
1871 if (ISD::isSignedIntSetCC(Cond)) {
1872 MinVal = APInt::getSignedMinValue(OperandBitSize);
1873 MaxVal = APInt::getSignedMaxValue(OperandBitSize);
1875 MinVal = APInt::getMinValue(OperandBitSize);
1876 MaxVal = APInt::getMaxValue(OperandBitSize);
1879 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
1880 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
1881 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true
1882 // X >= C0 --> X > (C0-1)
1883 return DAG.getSetCC(dl, VT, N0,
1884 DAG.getConstant(C1-1, N1.getValueType()),
1885 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
1888 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
1889 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true
1890 // X <= C0 --> X < (C0+1)
1891 return DAG.getSetCC(dl, VT, N0,
1892 DAG.getConstant(C1+1, N1.getValueType()),
1893 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
1896 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
1897 return DAG.getConstant(0, VT); // X < MIN --> false
1898 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal)
1899 return DAG.getConstant(1, VT); // X >= MIN --> true
1900 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal)
1901 return DAG.getConstant(0, VT); // X > MAX --> false
1902 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal)
1903 return DAG.getConstant(1, VT); // X <= MAX --> true
1905 // Canonicalize setgt X, Min --> setne X, Min
1906 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
1907 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
1908 // Canonicalize setlt X, Max --> setne X, Max
1909 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
1910 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
1912 // If we have setult X, 1, turn it into seteq X, 0
1913 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
1914 return DAG.getSetCC(dl, VT, N0,
1915 DAG.getConstant(MinVal, N0.getValueType()),
1917 // If we have setugt X, Max-1, turn it into seteq X, Max
1918 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
1919 return DAG.getSetCC(dl, VT, N0,
1920 DAG.getConstant(MaxVal, N0.getValueType()),
1923 // If we have "setcc X, C0", check to see if we can shrink the immediate
1926 // SETUGT X, SINTMAX -> SETLT X, 0
1927 if (Cond == ISD::SETUGT &&
1928 C1 == APInt::getSignedMaxValue(OperandBitSize))
1929 return DAG.getSetCC(dl, VT, N0,
1930 DAG.getConstant(0, N1.getValueType()),
1933 // SETULT X, SINTMIN -> SETGT X, -1
1934 if (Cond == ISD::SETULT &&
1935 C1 == APInt::getSignedMinValue(OperandBitSize)) {
1936 SDValue ConstMinusOne =
1937 DAG.getConstant(APInt::getAllOnesValue(OperandBitSize),
1939 return DAG.getSetCC(dl, VT, N0, ConstMinusOne, ISD::SETGT);
1942 // Fold bit comparisons when we can.
1943 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1944 (VT == N0.getValueType() ||
1945 (isTypeLegal(VT) && VT.bitsLE(N0.getValueType()))) &&
1946 N0.getOpcode() == ISD::AND)
1947 if (ConstantSDNode *AndRHS =
1948 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1949 EVT ShiftTy = DCI.isBeforeLegalize() ?
1950 getPointerTy() : getShiftAmountTy();
1951 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
1952 // Perform the xform if the AND RHS is a single bit.
1953 if (AndRHS->getAPIntValue().isPowerOf2()) {
1954 return DAG.getNode(ISD::TRUNCATE, dl, VT,
1955 DAG.getNode(ISD::SRL, dl, N0.getValueType(), N0,
1956 DAG.getConstant(AndRHS->getAPIntValue().logBase2(), ShiftTy)));
1958 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getAPIntValue()) {
1959 // (X & 8) == 8 --> (X & 8) >> 3
1960 // Perform the xform if C1 is a single bit.
1961 if (C1.isPowerOf2()) {
1962 return DAG.getNode(ISD::TRUNCATE, dl, VT,
1963 DAG.getNode(ISD::SRL, dl, N0.getValueType(), N0,
1964 DAG.getConstant(C1.logBase2(), ShiftTy)));
1970 if (isa<ConstantFPSDNode>(N0.getNode())) {
1971 // Constant fold or commute setcc.
1972 SDValue O = DAG.FoldSetCC(VT, N0, N1, Cond, dl);
1973 if (O.getNode()) return O;
1974 } else if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1.getNode())) {
1975 // If the RHS of an FP comparison is a constant, simplify it away in
1977 if (CFP->getValueAPF().isNaN()) {
1978 // If an operand is known to be a nan, we can fold it.
1979 switch (ISD::getUnorderedFlavor(Cond)) {
1980 default: llvm_unreachable("Unknown flavor!");
1981 case 0: // Known false.
1982 return DAG.getConstant(0, VT);
1983 case 1: // Known true.
1984 return DAG.getConstant(1, VT);
1985 case 2: // Undefined.
1986 return DAG.getUNDEF(VT);
1990 // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the
1991 // constant if knowing that the operand is non-nan is enough. We prefer to
1992 // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to
1994 if (Cond == ISD::SETO || Cond == ISD::SETUO)
1995 return DAG.getSetCC(dl, VT, N0, N0, Cond);
1997 // If the condition is not legal, see if we can find an equivalent one
1999 if (!isCondCodeLegal(Cond, N0.getValueType())) {
2000 // If the comparison was an awkward floating-point == or != and one of
2001 // the comparison operands is infinity or negative infinity, convert the
2002 // condition to a less-awkward <= or >=.
2003 if (CFP->getValueAPF().isInfinity()) {
2004 if (CFP->getValueAPF().isNegative()) {
2005 if (Cond == ISD::SETOEQ &&
2006 isCondCodeLegal(ISD::SETOLE, N0.getValueType()))
2007 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLE);
2008 if (Cond == ISD::SETUEQ &&
2009 isCondCodeLegal(ISD::SETOLE, N0.getValueType()))
2010 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULE);
2011 if (Cond == ISD::SETUNE &&
2012 isCondCodeLegal(ISD::SETUGT, N0.getValueType()))
2013 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGT);
2014 if (Cond == ISD::SETONE &&
2015 isCondCodeLegal(ISD::SETUGT, N0.getValueType()))
2016 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGT);
2018 if (Cond == ISD::SETOEQ &&
2019 isCondCodeLegal(ISD::SETOGE, N0.getValueType()))
2020 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGE);
2021 if (Cond == ISD::SETUEQ &&
2022 isCondCodeLegal(ISD::SETOGE, N0.getValueType()))
2023 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGE);
2024 if (Cond == ISD::SETUNE &&
2025 isCondCodeLegal(ISD::SETULT, N0.getValueType()))
2026 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULT);
2027 if (Cond == ISD::SETONE &&
2028 isCondCodeLegal(ISD::SETULT, N0.getValueType()))
2029 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLT);
2036 // We can always fold X == X for integer setcc's.
2037 if (N0.getValueType().isInteger())
2038 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
2039 unsigned UOF = ISD::getUnorderedFlavor(Cond);
2040 if (UOF == 2) // FP operators that are undefined on NaNs.
2041 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
2042 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
2043 return DAG.getConstant(UOF, VT);
2044 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
2045 // if it is not already.
2046 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
2047 if (NewCond != Cond)
2048 return DAG.getSetCC(dl, VT, N0, N1, NewCond);
2051 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
2052 N0.getValueType().isInteger()) {
2053 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
2054 N0.getOpcode() == ISD::XOR) {
2055 // Simplify (X+Y) == (X+Z) --> Y == Z
2056 if (N0.getOpcode() == N1.getOpcode()) {
2057 if (N0.getOperand(0) == N1.getOperand(0))
2058 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(1), Cond);
2059 if (N0.getOperand(1) == N1.getOperand(1))
2060 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(0), Cond);
2061 if (DAG.isCommutativeBinOp(N0.getOpcode())) {
2062 // If X op Y == Y op X, try other combinations.
2063 if (N0.getOperand(0) == N1.getOperand(1))
2064 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(0),
2066 if (N0.getOperand(1) == N1.getOperand(0))
2067 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(1),
2072 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) {
2073 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2074 // Turn (X+C1) == C2 --> X == C2-C1
2075 if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse()) {
2076 return DAG.getSetCC(dl, VT, N0.getOperand(0),
2077 DAG.getConstant(RHSC->getAPIntValue()-
2078 LHSR->getAPIntValue(),
2079 N0.getValueType()), Cond);
2082 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
2083 if (N0.getOpcode() == ISD::XOR)
2084 // If we know that all of the inverted bits are zero, don't bother
2085 // performing the inversion.
2086 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue()))
2088 DAG.getSetCC(dl, VT, N0.getOperand(0),
2089 DAG.getConstant(LHSR->getAPIntValue() ^
2090 RHSC->getAPIntValue(),
2095 // Turn (C1-X) == C2 --> X == C1-C2
2096 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
2097 if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse()) {
2099 DAG.getSetCC(dl, VT, N0.getOperand(1),
2100 DAG.getConstant(SUBC->getAPIntValue() -
2101 RHSC->getAPIntValue(),
2108 // Simplify (X+Z) == X --> Z == 0
2109 if (N0.getOperand(0) == N1)
2110 return DAG.getSetCC(dl, VT, N0.getOperand(1),
2111 DAG.getConstant(0, N0.getValueType()), Cond);
2112 if (N0.getOperand(1) == N1) {
2113 if (DAG.isCommutativeBinOp(N0.getOpcode()))
2114 return DAG.getSetCC(dl, VT, N0.getOperand(0),
2115 DAG.getConstant(0, N0.getValueType()), Cond);
2116 else if (N0.getNode()->hasOneUse()) {
2117 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
2118 // (Z-X) == X --> Z == X<<1
2119 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(),
2121 DAG.getConstant(1, getShiftAmountTy()));
2122 if (!DCI.isCalledByLegalizer())
2123 DCI.AddToWorklist(SH.getNode());
2124 return DAG.getSetCC(dl, VT, N0.getOperand(0), SH, Cond);
2129 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
2130 N1.getOpcode() == ISD::XOR) {
2131 // Simplify X == (X+Z) --> Z == 0
2132 if (N1.getOperand(0) == N0) {
2133 return DAG.getSetCC(dl, VT, N1.getOperand(1),
2134 DAG.getConstant(0, N1.getValueType()), Cond);
2135 } else if (N1.getOperand(1) == N0) {
2136 if (DAG.isCommutativeBinOp(N1.getOpcode())) {
2137 return DAG.getSetCC(dl, VT, N1.getOperand(0),
2138 DAG.getConstant(0, N1.getValueType()), Cond);
2139 } else if (N1.getNode()->hasOneUse()) {
2140 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
2141 // X == (Z-X) --> X<<1 == Z
2142 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(), N0,
2143 DAG.getConstant(1, getShiftAmountTy()));
2144 if (!DCI.isCalledByLegalizer())
2145 DCI.AddToWorklist(SH.getNode());
2146 return DAG.getSetCC(dl, VT, SH, N1.getOperand(0), Cond);
2151 // Simplify x&y == y to x&y != 0 if y has exactly one bit set.
2152 // Note that where y is variable and is known to have at most
2153 // one bit set (for example, if it is z&1) we cannot do this;
2154 // the expressions are not equivalent when y==0.
2155 if (N0.getOpcode() == ISD::AND)
2156 if (N0.getOperand(0) == N1 || N0.getOperand(1) == N1) {
2157 if (ValueHasExactlyOneBitSet(N1, DAG)) {
2158 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true);
2159 SDValue Zero = DAG.getConstant(0, N1.getValueType());
2160 return DAG.getSetCC(dl, VT, N0, Zero, Cond);
2163 if (N1.getOpcode() == ISD::AND)
2164 if (N1.getOperand(0) == N0 || N1.getOperand(1) == N0) {
2165 if (ValueHasExactlyOneBitSet(N0, DAG)) {
2166 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true);
2167 SDValue Zero = DAG.getConstant(0, N0.getValueType());
2168 return DAG.getSetCC(dl, VT, N1, Zero, Cond);
2173 // Fold away ALL boolean setcc's.
2175 if (N0.getValueType() == MVT::i1 && foldBooleans) {
2177 default: llvm_unreachable("Unknown integer setcc!");
2178 case ISD::SETEQ: // X == Y -> ~(X^Y)
2179 Temp = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1);
2180 N0 = DAG.getNOT(dl, Temp, MVT::i1);
2181 if (!DCI.isCalledByLegalizer())
2182 DCI.AddToWorklist(Temp.getNode());
2184 case ISD::SETNE: // X != Y --> (X^Y)
2185 N0 = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1);
2187 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> ~X & Y
2188 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> ~X & Y
2189 Temp = DAG.getNOT(dl, N0, MVT::i1);
2190 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N1, Temp);
2191 if (!DCI.isCalledByLegalizer())
2192 DCI.AddToWorklist(Temp.getNode());
2194 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> ~Y & X
2195 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> ~Y & X
2196 Temp = DAG.getNOT(dl, N1, MVT::i1);
2197 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N0, Temp);
2198 if (!DCI.isCalledByLegalizer())
2199 DCI.AddToWorklist(Temp.getNode());
2201 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> ~X | Y
2202 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> ~X | Y
2203 Temp = DAG.getNOT(dl, N0, MVT::i1);
2204 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N1, Temp);
2205 if (!DCI.isCalledByLegalizer())
2206 DCI.AddToWorklist(Temp.getNode());
2208 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> ~Y | X
2209 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> ~Y | X
2210 Temp = DAG.getNOT(dl, N1, MVT::i1);
2211 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N0, Temp);
2214 if (VT != MVT::i1) {
2215 if (!DCI.isCalledByLegalizer())
2216 DCI.AddToWorklist(N0.getNode());
2217 // FIXME: If running after legalize, we probably can't do this.
2218 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, N0);
2223 // Could not fold it.
2227 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
2228 /// node is a GlobalAddress + offset.
2229 bool TargetLowering::isGAPlusOffset(SDNode *N, GlobalValue* &GA,
2230 int64_t &Offset) const {
2231 if (isa<GlobalAddressSDNode>(N)) {
2232 GlobalAddressSDNode *GASD = cast<GlobalAddressSDNode>(N);
2233 GA = GASD->getGlobal();
2234 Offset += GASD->getOffset();
2238 if (N->getOpcode() == ISD::ADD) {
2239 SDValue N1 = N->getOperand(0);
2240 SDValue N2 = N->getOperand(1);
2241 if (isGAPlusOffset(N1.getNode(), GA, Offset)) {
2242 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
2244 Offset += V->getSExtValue();
2247 } else if (isGAPlusOffset(N2.getNode(), GA, Offset)) {
2248 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
2250 Offset += V->getSExtValue();
2259 SDValue TargetLowering::
2260 PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
2261 // Default implementation: no optimization.
2265 //===----------------------------------------------------------------------===//
2266 // Inline Assembler Implementation Methods
2267 //===----------------------------------------------------------------------===//
2270 TargetLowering::ConstraintType
2271 TargetLowering::getConstraintType(const std::string &Constraint) const {
2272 // FIXME: lots more standard ones to handle.
2273 if (Constraint.size() == 1) {
2274 switch (Constraint[0]) {
2276 case 'r': return C_RegisterClass;
2278 case 'o': // offsetable
2279 case 'V': // not offsetable
2281 case 'i': // Simple Integer or Relocatable Constant
2282 case 'n': // Simple Integer
2283 case 's': // Relocatable Constant
2284 case 'X': // Allow ANY value.
2285 case 'I': // Target registers.
2297 if (Constraint.size() > 1 && Constraint[0] == '{' &&
2298 Constraint[Constraint.size()-1] == '}')
2303 /// LowerXConstraint - try to replace an X constraint, which matches anything,
2304 /// with another that has more specific requirements based on the type of the
2305 /// corresponding operand.
2306 const char *TargetLowering::LowerXConstraint(EVT ConstraintVT) const{
2307 if (ConstraintVT.isInteger())
2309 if (ConstraintVT.isFloatingPoint())
2310 return "f"; // works for many targets
2314 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
2315 /// vector. If it is invalid, don't add anything to Ops.
2316 void TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
2317 char ConstraintLetter,
2319 std::vector<SDValue> &Ops,
2320 SelectionDAG &DAG) const {
2321 switch (ConstraintLetter) {
2323 case 'X': // Allows any operand; labels (basic block) use this.
2324 if (Op.getOpcode() == ISD::BasicBlock) {
2329 case 'i': // Simple Integer or Relocatable Constant
2330 case 'n': // Simple Integer
2331 case 's': { // Relocatable Constant
2332 // These operands are interested in values of the form (GV+C), where C may
2333 // be folded in as an offset of GV, or it may be explicitly added. Also, it
2334 // is possible and fine if either GV or C are missing.
2335 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2336 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
2338 // If we have "(add GV, C)", pull out GV/C
2339 if (Op.getOpcode() == ISD::ADD) {
2340 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
2341 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
2342 if (C == 0 || GA == 0) {
2343 C = dyn_cast<ConstantSDNode>(Op.getOperand(0));
2344 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1));
2346 if (C == 0 || GA == 0)
2350 // If we find a valid operand, map to the TargetXXX version so that the
2351 // value itself doesn't get selected.
2352 if (GA) { // Either &GV or &GV+C
2353 if (ConstraintLetter != 'n') {
2354 int64_t Offs = GA->getOffset();
2355 if (C) Offs += C->getZExtValue();
2356 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(),
2357 Op.getValueType(), Offs));
2361 if (C) { // just C, no GV.
2362 // Simple constants are not allowed for 's'.
2363 if (ConstraintLetter != 's') {
2364 // gcc prints these as sign extended. Sign extend value to 64 bits
2365 // now; without this it would get ZExt'd later in
2366 // ScheduleDAGSDNodes::EmitNode, which is very generic.
2367 Ops.push_back(DAG.getTargetConstant(C->getAPIntValue().getSExtValue(),
2377 std::vector<unsigned> TargetLowering::
2378 getRegClassForInlineAsmConstraint(const std::string &Constraint,
2380 return std::vector<unsigned>();
2384 std::pair<unsigned, const TargetRegisterClass*> TargetLowering::
2385 getRegForInlineAsmConstraint(const std::string &Constraint,
2387 if (Constraint[0] != '{')
2388 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
2389 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?");
2391 // Remove the braces from around the name.
2392 StringRef RegName(Constraint.data()+1, Constraint.size()-2);
2394 // Figure out which register class contains this reg.
2395 const TargetRegisterInfo *RI = TM.getRegisterInfo();
2396 for (TargetRegisterInfo::regclass_iterator RCI = RI->regclass_begin(),
2397 E = RI->regclass_end(); RCI != E; ++RCI) {
2398 const TargetRegisterClass *RC = *RCI;
2400 // If none of the value types for this register class are valid, we
2401 // can't use it. For example, 64-bit reg classes on 32-bit targets.
2402 bool isLegal = false;
2403 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
2405 if (isTypeLegal(*I)) {
2411 if (!isLegal) continue;
2413 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end();
2415 if (RegName.equals_lower(RI->getName(*I)))
2416 return std::make_pair(*I, RC);
2420 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
2423 //===----------------------------------------------------------------------===//
2424 // Constraint Selection.
2426 /// isMatchingInputConstraint - Return true of this is an input operand that is
2427 /// a matching constraint like "4".
2428 bool TargetLowering::AsmOperandInfo::isMatchingInputConstraint() const {
2429 assert(!ConstraintCode.empty() && "No known constraint!");
2430 return isdigit(ConstraintCode[0]);
2433 /// getMatchedOperand - If this is an input matching constraint, this method
2434 /// returns the output operand it matches.
2435 unsigned TargetLowering::AsmOperandInfo::getMatchedOperand() const {
2436 assert(!ConstraintCode.empty() && "No known constraint!");
2437 return atoi(ConstraintCode.c_str());
2441 /// getConstraintGenerality - Return an integer indicating how general CT
2443 static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) {
2445 default: llvm_unreachable("Unknown constraint type!");
2446 case TargetLowering::C_Other:
2447 case TargetLowering::C_Unknown:
2449 case TargetLowering::C_Register:
2451 case TargetLowering::C_RegisterClass:
2453 case TargetLowering::C_Memory:
2458 /// ChooseConstraint - If there are multiple different constraints that we
2459 /// could pick for this operand (e.g. "imr") try to pick the 'best' one.
2460 /// This is somewhat tricky: constraints fall into four classes:
2461 /// Other -> immediates and magic values
2462 /// Register -> one specific register
2463 /// RegisterClass -> a group of regs
2464 /// Memory -> memory
2465 /// Ideally, we would pick the most specific constraint possible: if we have
2466 /// something that fits into a register, we would pick it. The problem here
2467 /// is that if we have something that could either be in a register or in
2468 /// memory that use of the register could cause selection of *other*
2469 /// operands to fail: they might only succeed if we pick memory. Because of
2470 /// this the heuristic we use is:
2472 /// 1) If there is an 'other' constraint, and if the operand is valid for
2473 /// that constraint, use it. This makes us take advantage of 'i'
2474 /// constraints when available.
2475 /// 2) Otherwise, pick the most general constraint present. This prefers
2476 /// 'm' over 'r', for example.
2478 static void ChooseConstraint(TargetLowering::AsmOperandInfo &OpInfo,
2479 bool hasMemory, const TargetLowering &TLI,
2480 SDValue Op, SelectionDAG *DAG) {
2481 assert(OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options");
2482 unsigned BestIdx = 0;
2483 TargetLowering::ConstraintType BestType = TargetLowering::C_Unknown;
2484 int BestGenerality = -1;
2486 // Loop over the options, keeping track of the most general one.
2487 for (unsigned i = 0, e = OpInfo.Codes.size(); i != e; ++i) {
2488 TargetLowering::ConstraintType CType =
2489 TLI.getConstraintType(OpInfo.Codes[i]);
2491 // If this is an 'other' constraint, see if the operand is valid for it.
2492 // For example, on X86 we might have an 'rI' constraint. If the operand
2493 // is an integer in the range [0..31] we want to use I (saving a load
2494 // of a register), otherwise we must use 'r'.
2495 if (CType == TargetLowering::C_Other && Op.getNode()) {
2496 assert(OpInfo.Codes[i].size() == 1 &&
2497 "Unhandled multi-letter 'other' constraint");
2498 std::vector<SDValue> ResultOps;
2499 TLI.LowerAsmOperandForConstraint(Op, OpInfo.Codes[i][0], hasMemory,
2501 if (!ResultOps.empty()) {
2508 // This constraint letter is more general than the previous one, use it.
2509 int Generality = getConstraintGenerality(CType);
2510 if (Generality > BestGenerality) {
2513 BestGenerality = Generality;
2517 OpInfo.ConstraintCode = OpInfo.Codes[BestIdx];
2518 OpInfo.ConstraintType = BestType;
2521 /// ComputeConstraintToUse - Determines the constraint code and constraint
2522 /// type to use for the specific AsmOperandInfo, setting
2523 /// OpInfo.ConstraintCode and OpInfo.ConstraintType.
2524 void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo,
2527 SelectionDAG *DAG) const {
2528 assert(!OpInfo.Codes.empty() && "Must have at least one constraint");
2530 // Single-letter constraints ('r') are very common.
2531 if (OpInfo.Codes.size() == 1) {
2532 OpInfo.ConstraintCode = OpInfo.Codes[0];
2533 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
2535 ChooseConstraint(OpInfo, hasMemory, *this, Op, DAG);
2538 // 'X' matches anything.
2539 if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) {
2540 // Labels and constants are handled elsewhere ('X' is the only thing
2541 // that matches labels). For Functions, the type here is the type of
2542 // the result, which is not what we want to look at; leave them alone.
2543 Value *v = OpInfo.CallOperandVal;
2544 if (isa<BasicBlock>(v) || isa<ConstantInt>(v) || isa<Function>(v)) {
2545 OpInfo.CallOperandVal = v;
2549 // Otherwise, try to resolve it to something we know about by looking at
2550 // the actual operand type.
2551 if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) {
2552 OpInfo.ConstraintCode = Repl;
2553 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
2558 //===----------------------------------------------------------------------===//
2559 // Loop Strength Reduction hooks
2560 //===----------------------------------------------------------------------===//
2562 /// isLegalAddressingMode - Return true if the addressing mode represented
2563 /// by AM is legal for this target, for a load/store of the specified type.
2564 bool TargetLowering::isLegalAddressingMode(const AddrMode &AM,
2565 const Type *Ty) const {
2566 // The default implementation of this implements a conservative RISCy, r+r and
2569 // Allows a sign-extended 16-bit immediate field.
2570 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
2573 // No global is ever allowed as a base.
2577 // Only support r+r,
2579 case 0: // "r+i" or just "i", depending on HasBaseReg.
2582 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
2584 // Otherwise we have r+r or r+i.
2587 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
2589 // Allow 2*r as r+r.
2596 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
2597 /// return a DAG expression to select that will generate the same value by
2598 /// multiplying by a magic number. See:
2599 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
2600 SDValue TargetLowering::BuildSDIV(SDNode *N, SelectionDAG &DAG,
2601 std::vector<SDNode*>* Created) const {
2602 EVT VT = N->getValueType(0);
2603 DebugLoc dl= N->getDebugLoc();
2605 // Check to see if we can do this.
2606 // FIXME: We should be more aggressive here.
2607 if (!isTypeLegal(VT))
2610 APInt d = cast<ConstantSDNode>(N->getOperand(1))->getAPIntValue();
2611 APInt::ms magics = d.magic();
2613 // Multiply the numerator (operand 0) by the magic value
2614 // FIXME: We should support doing a MUL in a wider type
2616 if (isOperationLegalOrCustom(ISD::MULHS, VT))
2617 Q = DAG.getNode(ISD::MULHS, dl, VT, N->getOperand(0),
2618 DAG.getConstant(magics.m, VT));
2619 else if (isOperationLegalOrCustom(ISD::SMUL_LOHI, VT))
2620 Q = SDValue(DAG.getNode(ISD::SMUL_LOHI, dl, DAG.getVTList(VT, VT),
2622 DAG.getConstant(magics.m, VT)).getNode(), 1);
2624 return SDValue(); // No mulhs or equvialent
2625 // If d > 0 and m < 0, add the numerator
2626 if (d.isStrictlyPositive() && magics.m.isNegative()) {
2627 Q = DAG.getNode(ISD::ADD, dl, VT, Q, N->getOperand(0));
2629 Created->push_back(Q.getNode());
2631 // If d < 0 and m > 0, subtract the numerator.
2632 if (d.isNegative() && magics.m.isStrictlyPositive()) {
2633 Q = DAG.getNode(ISD::SUB, dl, VT, Q, N->getOperand(0));
2635 Created->push_back(Q.getNode());
2637 // Shift right algebraic if shift value is nonzero
2639 Q = DAG.getNode(ISD::SRA, dl, VT, Q,
2640 DAG.getConstant(magics.s, getShiftAmountTy()));
2642 Created->push_back(Q.getNode());
2644 // Extract the sign bit and add it to the quotient
2646 DAG.getNode(ISD::SRL, dl, VT, Q, DAG.getConstant(VT.getSizeInBits()-1,
2647 getShiftAmountTy()));
2649 Created->push_back(T.getNode());
2650 return DAG.getNode(ISD::ADD, dl, VT, Q, T);
2653 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
2654 /// return a DAG expression to select that will generate the same value by
2655 /// multiplying by a magic number. See:
2656 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
2657 SDValue TargetLowering::BuildUDIV(SDNode *N, SelectionDAG &DAG,
2658 std::vector<SDNode*>* Created) const {
2659 EVT VT = N->getValueType(0);
2660 DebugLoc dl = N->getDebugLoc();
2662 // Check to see if we can do this.
2663 // FIXME: We should be more aggressive here.
2664 if (!isTypeLegal(VT))
2667 // FIXME: We should use a narrower constant when the upper
2668 // bits are known to be zero.
2669 ConstantSDNode *N1C = cast<ConstantSDNode>(N->getOperand(1));
2670 APInt::mu magics = N1C->getAPIntValue().magicu();
2672 // Multiply the numerator (operand 0) by the magic value
2673 // FIXME: We should support doing a MUL in a wider type
2675 if (isOperationLegalOrCustom(ISD::MULHU, VT))
2676 Q = DAG.getNode(ISD::MULHU, dl, VT, N->getOperand(0),
2677 DAG.getConstant(magics.m, VT));
2678 else if (isOperationLegalOrCustom(ISD::UMUL_LOHI, VT))
2679 Q = SDValue(DAG.getNode(ISD::UMUL_LOHI, dl, DAG.getVTList(VT, VT),
2681 DAG.getConstant(magics.m, VT)).getNode(), 1);
2683 return SDValue(); // No mulhu or equvialent
2685 Created->push_back(Q.getNode());
2687 if (magics.a == 0) {
2688 assert(magics.s < N1C->getAPIntValue().getBitWidth() &&
2689 "We shouldn't generate an undefined shift!");
2690 return DAG.getNode(ISD::SRL, dl, VT, Q,
2691 DAG.getConstant(magics.s, getShiftAmountTy()));
2693 SDValue NPQ = DAG.getNode(ISD::SUB, dl, VT, N->getOperand(0), Q);
2695 Created->push_back(NPQ.getNode());
2696 NPQ = DAG.getNode(ISD::SRL, dl, VT, NPQ,
2697 DAG.getConstant(1, getShiftAmountTy()));
2699 Created->push_back(NPQ.getNode());
2700 NPQ = DAG.getNode(ISD::ADD, dl, VT, NPQ, Q);
2702 Created->push_back(NPQ.getNode());
2703 return DAG.getNode(ISD::SRL, dl, VT, NPQ,
2704 DAG.getConstant(magics.s-1, getShiftAmountTy()));