1 //===-- TargetLowering.cpp - Implement the TargetLowering class -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This implements the TargetLowering class.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Target/TargetAsmInfo.h"
15 #include "llvm/Target/TargetLowering.h"
16 #include "llvm/Target/TargetSubtarget.h"
17 #include "llvm/Target/TargetData.h"
18 #include "llvm/Target/TargetMachine.h"
19 #include "llvm/Target/TargetRegisterInfo.h"
20 #include "llvm/GlobalVariable.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/CodeGen/MachineFrameInfo.h"
23 #include "llvm/CodeGen/SelectionDAG.h"
24 #include "llvm/ADT/StringExtras.h"
25 #include "llvm/ADT/STLExtras.h"
26 #include "llvm/Support/MathExtras.h"
29 /// InitLibcallNames - Set default libcall names.
31 static void InitLibcallNames(const char **Names) {
32 Names[RTLIB::SHL_I32] = "__ashlsi3";
33 Names[RTLIB::SHL_I64] = "__ashldi3";
34 Names[RTLIB::SHL_I128] = "__ashlti3";
35 Names[RTLIB::SRL_I32] = "__lshrsi3";
36 Names[RTLIB::SRL_I64] = "__lshrdi3";
37 Names[RTLIB::SRL_I128] = "__lshrti3";
38 Names[RTLIB::SRA_I32] = "__ashrsi3";
39 Names[RTLIB::SRA_I64] = "__ashrdi3";
40 Names[RTLIB::SRA_I128] = "__ashrti3";
41 Names[RTLIB::MUL_I32] = "__mulsi3";
42 Names[RTLIB::MUL_I64] = "__muldi3";
43 Names[RTLIB::MUL_I128] = "__multi3";
44 Names[RTLIB::SDIV_I32] = "__divsi3";
45 Names[RTLIB::SDIV_I64] = "__divdi3";
46 Names[RTLIB::SDIV_I128] = "__divti3";
47 Names[RTLIB::UDIV_I32] = "__udivsi3";
48 Names[RTLIB::UDIV_I64] = "__udivdi3";
49 Names[RTLIB::UDIV_I128] = "__udivti3";
50 Names[RTLIB::SREM_I32] = "__modsi3";
51 Names[RTLIB::SREM_I64] = "__moddi3";
52 Names[RTLIB::SREM_I128] = "__modti3";
53 Names[RTLIB::UREM_I32] = "__umodsi3";
54 Names[RTLIB::UREM_I64] = "__umoddi3";
55 Names[RTLIB::UREM_I128] = "__umodti3";
56 Names[RTLIB::NEG_I32] = "__negsi2";
57 Names[RTLIB::NEG_I64] = "__negdi2";
58 Names[RTLIB::ADD_F32] = "__addsf3";
59 Names[RTLIB::ADD_F64] = "__adddf3";
60 Names[RTLIB::ADD_F80] = "__addxf3";
61 Names[RTLIB::ADD_PPCF128] = "__gcc_qadd";
62 Names[RTLIB::SUB_F32] = "__subsf3";
63 Names[RTLIB::SUB_F64] = "__subdf3";
64 Names[RTLIB::SUB_F80] = "__subxf3";
65 Names[RTLIB::SUB_PPCF128] = "__gcc_qsub";
66 Names[RTLIB::MUL_F32] = "__mulsf3";
67 Names[RTLIB::MUL_F64] = "__muldf3";
68 Names[RTLIB::MUL_F80] = "__mulxf3";
69 Names[RTLIB::MUL_PPCF128] = "__gcc_qmul";
70 Names[RTLIB::DIV_F32] = "__divsf3";
71 Names[RTLIB::DIV_F64] = "__divdf3";
72 Names[RTLIB::DIV_F80] = "__divxf3";
73 Names[RTLIB::DIV_PPCF128] = "__gcc_qdiv";
74 Names[RTLIB::REM_F32] = "fmodf";
75 Names[RTLIB::REM_F64] = "fmod";
76 Names[RTLIB::REM_F80] = "fmodl";
77 Names[RTLIB::REM_PPCF128] = "fmodl";
78 Names[RTLIB::POWI_F32] = "__powisf2";
79 Names[RTLIB::POWI_F64] = "__powidf2";
80 Names[RTLIB::POWI_F80] = "__powixf2";
81 Names[RTLIB::POWI_PPCF128] = "__powitf2";
82 Names[RTLIB::SQRT_F32] = "sqrtf";
83 Names[RTLIB::SQRT_F64] = "sqrt";
84 Names[RTLIB::SQRT_F80] = "sqrtl";
85 Names[RTLIB::SQRT_PPCF128] = "sqrtl";
86 Names[RTLIB::LOG_F32] = "logf";
87 Names[RTLIB::LOG_F64] = "log";
88 Names[RTLIB::LOG_F80] = "logl";
89 Names[RTLIB::LOG_PPCF128] = "logl";
90 Names[RTLIB::LOG2_F32] = "log2f";
91 Names[RTLIB::LOG2_F64] = "log2";
92 Names[RTLIB::LOG2_F80] = "log2l";
93 Names[RTLIB::LOG2_PPCF128] = "log2l";
94 Names[RTLIB::LOG10_F32] = "log10f";
95 Names[RTLIB::LOG10_F64] = "log10";
96 Names[RTLIB::LOG10_F80] = "log10l";
97 Names[RTLIB::LOG10_PPCF128] = "log10l";
98 Names[RTLIB::EXP_F32] = "expf";
99 Names[RTLIB::EXP_F64] = "exp";
100 Names[RTLIB::EXP_F80] = "expl";
101 Names[RTLIB::EXP_PPCF128] = "expl";
102 Names[RTLIB::EXP2_F32] = "exp2f";
103 Names[RTLIB::EXP2_F64] = "exp2";
104 Names[RTLIB::EXP2_F80] = "exp2l";
105 Names[RTLIB::EXP2_PPCF128] = "exp2l";
106 Names[RTLIB::SIN_F32] = "sinf";
107 Names[RTLIB::SIN_F64] = "sin";
108 Names[RTLIB::SIN_F80] = "sinl";
109 Names[RTLIB::SIN_PPCF128] = "sinl";
110 Names[RTLIB::COS_F32] = "cosf";
111 Names[RTLIB::COS_F64] = "cos";
112 Names[RTLIB::COS_F80] = "cosl";
113 Names[RTLIB::COS_PPCF128] = "cosl";
114 Names[RTLIB::POW_F32] = "powf";
115 Names[RTLIB::POW_F64] = "pow";
116 Names[RTLIB::POW_F80] = "powl";
117 Names[RTLIB::POW_PPCF128] = "powl";
118 Names[RTLIB::CEIL_F32] = "ceilf";
119 Names[RTLIB::CEIL_F64] = "ceil";
120 Names[RTLIB::CEIL_F80] = "ceill";
121 Names[RTLIB::CEIL_PPCF128] = "ceill";
122 Names[RTLIB::TRUNC_F32] = "truncf";
123 Names[RTLIB::TRUNC_F64] = "trunc";
124 Names[RTLIB::TRUNC_F80] = "truncl";
125 Names[RTLIB::TRUNC_PPCF128] = "truncl";
126 Names[RTLIB::RINT_F32] = "rintf";
127 Names[RTLIB::RINT_F64] = "rint";
128 Names[RTLIB::RINT_F80] = "rintl";
129 Names[RTLIB::RINT_PPCF128] = "rintl";
130 Names[RTLIB::NEARBYINT_F32] = "nearbyintf";
131 Names[RTLIB::NEARBYINT_F64] = "nearbyint";
132 Names[RTLIB::NEARBYINT_F80] = "nearbyintl";
133 Names[RTLIB::NEARBYINT_PPCF128] = "nearbyintl";
134 Names[RTLIB::FLOOR_F32] = "floorf";
135 Names[RTLIB::FLOOR_F64] = "floor";
136 Names[RTLIB::FLOOR_F80] = "floorl";
137 Names[RTLIB::FLOOR_PPCF128] = "floorl";
138 Names[RTLIB::FPEXT_F32_F64] = "__extendsfdf2";
139 Names[RTLIB::FPROUND_F64_F32] = "__truncdfsf2";
140 Names[RTLIB::FPROUND_F80_F32] = "__truncxfsf2";
141 Names[RTLIB::FPROUND_PPCF128_F32] = "__trunctfsf2";
142 Names[RTLIB::FPROUND_F80_F64] = "__truncxfdf2";
143 Names[RTLIB::FPROUND_PPCF128_F64] = "__trunctfdf2";
144 Names[RTLIB::FPTOSINT_F32_I32] = "__fixsfsi";
145 Names[RTLIB::FPTOSINT_F32_I64] = "__fixsfdi";
146 Names[RTLIB::FPTOSINT_F32_I128] = "__fixsfti";
147 Names[RTLIB::FPTOSINT_F64_I32] = "__fixdfsi";
148 Names[RTLIB::FPTOSINT_F64_I64] = "__fixdfdi";
149 Names[RTLIB::FPTOSINT_F64_I128] = "__fixdfti";
150 Names[RTLIB::FPTOSINT_F80_I32] = "__fixxfsi";
151 Names[RTLIB::FPTOSINT_F80_I64] = "__fixxfdi";
152 Names[RTLIB::FPTOSINT_F80_I128] = "__fixxfti";
153 Names[RTLIB::FPTOSINT_PPCF128_I32] = "__fixtfsi";
154 Names[RTLIB::FPTOSINT_PPCF128_I64] = "__fixtfdi";
155 Names[RTLIB::FPTOSINT_PPCF128_I128] = "__fixtfti";
156 Names[RTLIB::FPTOUINT_F32_I32] = "__fixunssfsi";
157 Names[RTLIB::FPTOUINT_F32_I64] = "__fixunssfdi";
158 Names[RTLIB::FPTOUINT_F32_I128] = "__fixunssfti";
159 Names[RTLIB::FPTOUINT_F64_I32] = "__fixunsdfsi";
160 Names[RTLIB::FPTOUINT_F64_I64] = "__fixunsdfdi";
161 Names[RTLIB::FPTOUINT_F64_I128] = "__fixunsdfti";
162 Names[RTLIB::FPTOUINT_F80_I32] = "__fixunsxfsi";
163 Names[RTLIB::FPTOUINT_F80_I64] = "__fixunsxfdi";
164 Names[RTLIB::FPTOUINT_F80_I128] = "__fixunsxfti";
165 Names[RTLIB::FPTOUINT_PPCF128_I32] = "__fixunstfsi";
166 Names[RTLIB::FPTOUINT_PPCF128_I64] = "__fixunstfdi";
167 Names[RTLIB::FPTOUINT_PPCF128_I128] = "__fixunstfti";
168 Names[RTLIB::SINTTOFP_I32_F32] = "__floatsisf";
169 Names[RTLIB::SINTTOFP_I32_F64] = "__floatsidf";
170 Names[RTLIB::SINTTOFP_I32_F80] = "__floatsixf";
171 Names[RTLIB::SINTTOFP_I32_PPCF128] = "__floatsitf";
172 Names[RTLIB::SINTTOFP_I64_F32] = "__floatdisf";
173 Names[RTLIB::SINTTOFP_I64_F64] = "__floatdidf";
174 Names[RTLIB::SINTTOFP_I64_F80] = "__floatdixf";
175 Names[RTLIB::SINTTOFP_I64_PPCF128] = "__floatditf";
176 Names[RTLIB::SINTTOFP_I128_F32] = "__floattisf";
177 Names[RTLIB::SINTTOFP_I128_F64] = "__floattidf";
178 Names[RTLIB::SINTTOFP_I128_F80] = "__floattixf";
179 Names[RTLIB::SINTTOFP_I128_PPCF128] = "__floattitf";
180 Names[RTLIB::UINTTOFP_I32_F32] = "__floatunsisf";
181 Names[RTLIB::UINTTOFP_I32_F64] = "__floatunsidf";
182 Names[RTLIB::UINTTOFP_I32_F80] = "__floatunsixf";
183 Names[RTLIB::UINTTOFP_I32_PPCF128] = "__floatunsitf";
184 Names[RTLIB::UINTTOFP_I64_F32] = "__floatundisf";
185 Names[RTLIB::UINTTOFP_I64_F64] = "__floatundidf";
186 Names[RTLIB::UINTTOFP_I64_F80] = "__floatundixf";
187 Names[RTLIB::UINTTOFP_I64_PPCF128] = "__floatunditf";
188 Names[RTLIB::UINTTOFP_I128_F32] = "__floatuntisf";
189 Names[RTLIB::UINTTOFP_I128_F64] = "__floatuntidf";
190 Names[RTLIB::UINTTOFP_I128_F80] = "__floatuntixf";
191 Names[RTLIB::UINTTOFP_I128_PPCF128] = "__floatuntitf";
192 Names[RTLIB::OEQ_F32] = "__eqsf2";
193 Names[RTLIB::OEQ_F64] = "__eqdf2";
194 Names[RTLIB::UNE_F32] = "__nesf2";
195 Names[RTLIB::UNE_F64] = "__nedf2";
196 Names[RTLIB::OGE_F32] = "__gesf2";
197 Names[RTLIB::OGE_F64] = "__gedf2";
198 Names[RTLIB::OLT_F32] = "__ltsf2";
199 Names[RTLIB::OLT_F64] = "__ltdf2";
200 Names[RTLIB::OLE_F32] = "__lesf2";
201 Names[RTLIB::OLE_F64] = "__ledf2";
202 Names[RTLIB::OGT_F32] = "__gtsf2";
203 Names[RTLIB::OGT_F64] = "__gtdf2";
204 Names[RTLIB::UO_F32] = "__unordsf2";
205 Names[RTLIB::UO_F64] = "__unorddf2";
206 Names[RTLIB::O_F32] = "__unordsf2";
207 Names[RTLIB::O_F64] = "__unorddf2";
210 /// getFPEXT - Return the FPEXT_*_* value for the given types, or
211 /// UNKNOWN_LIBCALL if there is none.
212 RTLIB::Libcall RTLIB::getFPEXT(MVT OpVT, MVT RetVT) {
213 if (OpVT == MVT::f32) {
214 if (RetVT == MVT::f64)
215 return FPEXT_F32_F64;
217 return UNKNOWN_LIBCALL;
220 /// getFPROUND - Return the FPROUND_*_* value for the given types, or
221 /// UNKNOWN_LIBCALL if there is none.
222 RTLIB::Libcall RTLIB::getFPROUND(MVT OpVT, MVT RetVT) {
223 if (RetVT == MVT::f32) {
224 if (OpVT == MVT::f64)
225 return FPROUND_F64_F32;
226 if (OpVT == MVT::f80)
227 return FPROUND_F80_F32;
228 if (OpVT == MVT::ppcf128)
229 return FPROUND_PPCF128_F32;
230 } else if (RetVT == MVT::f64) {
231 if (OpVT == MVT::f80)
232 return FPROUND_F80_F64;
233 if (OpVT == MVT::ppcf128)
234 return FPROUND_PPCF128_F64;
236 return UNKNOWN_LIBCALL;
239 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or
240 /// UNKNOWN_LIBCALL if there is none.
241 RTLIB::Libcall RTLIB::getFPTOSINT(MVT OpVT, MVT RetVT) {
242 if (OpVT == MVT::f32) {
243 if (RetVT == MVT::i32)
244 return FPTOSINT_F32_I32;
245 if (RetVT == MVT::i64)
246 return FPTOSINT_F32_I64;
247 if (RetVT == MVT::i128)
248 return FPTOSINT_F32_I128;
249 } else if (OpVT == MVT::f64) {
250 if (RetVT == MVT::i32)
251 return FPTOSINT_F64_I32;
252 if (RetVT == MVT::i64)
253 return FPTOSINT_F64_I64;
254 if (RetVT == MVT::i128)
255 return FPTOSINT_F64_I128;
256 } else if (OpVT == MVT::f80) {
257 if (RetVT == MVT::i32)
258 return FPTOSINT_F80_I32;
259 if (RetVT == MVT::i64)
260 return FPTOSINT_F80_I64;
261 if (RetVT == MVT::i128)
262 return FPTOSINT_F80_I128;
263 } else if (OpVT == MVT::ppcf128) {
264 if (RetVT == MVT::i32)
265 return FPTOSINT_PPCF128_I32;
266 if (RetVT == MVT::i64)
267 return FPTOSINT_PPCF128_I64;
268 if (RetVT == MVT::i128)
269 return FPTOSINT_PPCF128_I128;
271 return UNKNOWN_LIBCALL;
274 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or
275 /// UNKNOWN_LIBCALL if there is none.
276 RTLIB::Libcall RTLIB::getFPTOUINT(MVT OpVT, MVT RetVT) {
277 if (OpVT == MVT::f32) {
278 if (RetVT == MVT::i32)
279 return FPTOUINT_F32_I32;
280 if (RetVT == MVT::i64)
281 return FPTOUINT_F32_I64;
282 if (RetVT == MVT::i128)
283 return FPTOUINT_F32_I128;
284 } else if (OpVT == MVT::f64) {
285 if (RetVT == MVT::i32)
286 return FPTOUINT_F64_I32;
287 if (RetVT == MVT::i64)
288 return FPTOUINT_F64_I64;
289 if (RetVT == MVT::i128)
290 return FPTOUINT_F64_I128;
291 } else if (OpVT == MVT::f80) {
292 if (RetVT == MVT::i32)
293 return FPTOUINT_F80_I32;
294 if (RetVT == MVT::i64)
295 return FPTOUINT_F80_I64;
296 if (RetVT == MVT::i128)
297 return FPTOUINT_F80_I128;
298 } else if (OpVT == MVT::ppcf128) {
299 if (RetVT == MVT::i32)
300 return FPTOUINT_PPCF128_I32;
301 if (RetVT == MVT::i64)
302 return FPTOUINT_PPCF128_I64;
303 if (RetVT == MVT::i128)
304 return FPTOUINT_PPCF128_I128;
306 return UNKNOWN_LIBCALL;
309 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or
310 /// UNKNOWN_LIBCALL if there is none.
311 RTLIB::Libcall RTLIB::getSINTTOFP(MVT OpVT, MVT RetVT) {
312 if (OpVT == MVT::i32) {
313 if (RetVT == MVT::f32)
314 return SINTTOFP_I32_F32;
315 else if (RetVT == MVT::f64)
316 return SINTTOFP_I32_F64;
317 else if (RetVT == MVT::f80)
318 return SINTTOFP_I32_F80;
319 else if (RetVT == MVT::ppcf128)
320 return SINTTOFP_I32_PPCF128;
321 } else if (OpVT == MVT::i64) {
322 if (RetVT == MVT::f32)
323 return SINTTOFP_I64_F32;
324 else if (RetVT == MVT::f64)
325 return SINTTOFP_I64_F64;
326 else if (RetVT == MVT::f80)
327 return SINTTOFP_I64_F80;
328 else if (RetVT == MVT::ppcf128)
329 return SINTTOFP_I64_PPCF128;
330 } else if (OpVT == MVT::i128) {
331 if (RetVT == MVT::f32)
332 return SINTTOFP_I128_F32;
333 else if (RetVT == MVT::f64)
334 return SINTTOFP_I128_F64;
335 else if (RetVT == MVT::f80)
336 return SINTTOFP_I128_F80;
337 else if (RetVT == MVT::ppcf128)
338 return SINTTOFP_I128_PPCF128;
340 return UNKNOWN_LIBCALL;
343 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or
344 /// UNKNOWN_LIBCALL if there is none.
345 RTLIB::Libcall RTLIB::getUINTTOFP(MVT OpVT, MVT RetVT) {
346 if (OpVT == MVT::i32) {
347 if (RetVT == MVT::f32)
348 return UINTTOFP_I32_F32;
349 else if (RetVT == MVT::f64)
350 return UINTTOFP_I32_F64;
351 else if (RetVT == MVT::f80)
352 return UINTTOFP_I32_F80;
353 else if (RetVT == MVT::ppcf128)
354 return UINTTOFP_I32_PPCF128;
355 } else if (OpVT == MVT::i64) {
356 if (RetVT == MVT::f32)
357 return UINTTOFP_I64_F32;
358 else if (RetVT == MVT::f64)
359 return UINTTOFP_I64_F64;
360 else if (RetVT == MVT::f80)
361 return UINTTOFP_I64_F80;
362 else if (RetVT == MVT::ppcf128)
363 return UINTTOFP_I64_PPCF128;
364 } else if (OpVT == MVT::i128) {
365 if (RetVT == MVT::f32)
366 return UINTTOFP_I128_F32;
367 else if (RetVT == MVT::f64)
368 return UINTTOFP_I128_F64;
369 else if (RetVT == MVT::f80)
370 return UINTTOFP_I128_F80;
371 else if (RetVT == MVT::ppcf128)
372 return UINTTOFP_I128_PPCF128;
374 return UNKNOWN_LIBCALL;
377 /// InitCmpLibcallCCs - Set default comparison libcall CC.
379 static void InitCmpLibcallCCs(ISD::CondCode *CCs) {
380 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL);
381 CCs[RTLIB::OEQ_F32] = ISD::SETEQ;
382 CCs[RTLIB::OEQ_F64] = ISD::SETEQ;
383 CCs[RTLIB::UNE_F32] = ISD::SETNE;
384 CCs[RTLIB::UNE_F64] = ISD::SETNE;
385 CCs[RTLIB::OGE_F32] = ISD::SETGE;
386 CCs[RTLIB::OGE_F64] = ISD::SETGE;
387 CCs[RTLIB::OLT_F32] = ISD::SETLT;
388 CCs[RTLIB::OLT_F64] = ISD::SETLT;
389 CCs[RTLIB::OLE_F32] = ISD::SETLE;
390 CCs[RTLIB::OLE_F64] = ISD::SETLE;
391 CCs[RTLIB::OGT_F32] = ISD::SETGT;
392 CCs[RTLIB::OGT_F64] = ISD::SETGT;
393 CCs[RTLIB::UO_F32] = ISD::SETNE;
394 CCs[RTLIB::UO_F64] = ISD::SETNE;
395 CCs[RTLIB::O_F32] = ISD::SETEQ;
396 CCs[RTLIB::O_F64] = ISD::SETEQ;
399 TargetLowering::TargetLowering(TargetMachine &tm)
400 : TM(tm), TD(TM.getTargetData()) {
401 assert(ISD::BUILTIN_OP_END <= OpActionsCapacity &&
402 "Fixed size array in TargetLowering is not large enough!");
403 // All operations default to being supported.
404 memset(OpActions, 0, sizeof(OpActions));
405 memset(LoadExtActions, 0, sizeof(LoadExtActions));
406 memset(TruncStoreActions, 0, sizeof(TruncStoreActions));
407 memset(IndexedModeActions, 0, sizeof(IndexedModeActions));
408 memset(ConvertActions, 0, sizeof(ConvertActions));
409 memset(CondCodeActions, 0, sizeof(CondCodeActions));
411 // Set default actions for various operations.
412 for (unsigned VT = 0; VT != (unsigned)MVT::LAST_VALUETYPE; ++VT) {
413 // Default all indexed load / store to expand.
414 for (unsigned IM = (unsigned)ISD::PRE_INC;
415 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
416 setIndexedLoadAction(IM, (MVT::SimpleValueType)VT, Expand);
417 setIndexedStoreAction(IM, (MVT::SimpleValueType)VT, Expand);
420 // These operations default to expand.
421 setOperationAction(ISD::FGETSIGN, (MVT::SimpleValueType)VT, Expand);
424 // Most targets ignore the @llvm.prefetch intrinsic.
425 setOperationAction(ISD::PREFETCH, MVT::Other, Expand);
427 // ConstantFP nodes default to expand. Targets can either change this to
428 // Legal, in which case all fp constants are legal, or use addLegalFPImmediate
429 // to optimize expansions for certain constants.
430 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
431 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
432 setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
434 // These library functions default to expand.
435 setOperationAction(ISD::FLOG , MVT::f64, Expand);
436 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
437 setOperationAction(ISD::FLOG10,MVT::f64, Expand);
438 setOperationAction(ISD::FEXP , MVT::f64, Expand);
439 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
440 setOperationAction(ISD::FLOG , MVT::f32, Expand);
441 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
442 setOperationAction(ISD::FLOG10,MVT::f32, Expand);
443 setOperationAction(ISD::FEXP , MVT::f32, Expand);
444 setOperationAction(ISD::FEXP2, MVT::f32, Expand);
446 // Default ISD::TRAP to expand (which turns it into abort).
447 setOperationAction(ISD::TRAP, MVT::Other, Expand);
449 IsLittleEndian = TD->isLittleEndian();
450 UsesGlobalOffsetTable = false;
451 ShiftAmountTy = PointerTy = getValueType(TD->getIntPtrType());
452 ShiftAmtHandling = Undefined;
453 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE*sizeof(TargetRegisterClass*));
454 memset(TargetDAGCombineArray, 0, array_lengthof(TargetDAGCombineArray));
455 maxStoresPerMemset = maxStoresPerMemcpy = maxStoresPerMemmove = 8;
456 allowUnalignedMemoryAccesses = false;
457 UseUnderscoreSetJmp = false;
458 UseUnderscoreLongJmp = false;
459 SelectIsExpensive = false;
460 IntDivIsCheap = false;
461 Pow2DivIsCheap = false;
462 StackPointerRegisterToSaveRestore = 0;
463 ExceptionPointerRegister = 0;
464 ExceptionSelectorRegister = 0;
465 BooleanContents = UndefinedBooleanContent;
466 SchedPreferenceInfo = SchedulingForLatency;
468 JumpBufAlignment = 0;
469 IfCvtBlockSizeLimit = 2;
470 IfCvtDupBlockSizeLimit = 0;
471 PrefLoopAlignment = 0;
473 InitLibcallNames(LibcallRoutineNames);
474 InitCmpLibcallCCs(CmpLibcallCCs);
476 // Tell Legalize whether the assembler supports DEBUG_LOC.
477 const TargetAsmInfo *TASM = TM.getTargetAsmInfo();
478 if (!TASM || !TASM->hasDotLocAndDotFile())
479 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
482 TargetLowering::~TargetLowering() {}
484 /// computeRegisterProperties - Once all of the register classes are added,
485 /// this allows us to compute derived properties we expose.
486 void TargetLowering::computeRegisterProperties() {
487 assert(MVT::LAST_VALUETYPE <= 32 &&
488 "Too many value types for ValueTypeActions to hold!");
490 // Everything defaults to needing one register.
491 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
492 NumRegistersForVT[i] = 1;
493 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i;
495 // ...except isVoid, which doesn't need any registers.
496 NumRegistersForVT[MVT::isVoid] = 0;
498 // Find the largest integer register class.
499 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE;
500 for (; RegClassForVT[LargestIntReg] == 0; --LargestIntReg)
501 assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
503 // Every integer value type larger than this largest register takes twice as
504 // many registers to represent as the previous ValueType.
505 for (unsigned ExpandedReg = LargestIntReg + 1; ; ++ExpandedReg) {
506 MVT EVT = (MVT::SimpleValueType)ExpandedReg;
507 if (!EVT.isInteger())
509 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
510 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg;
511 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1);
512 ValueTypeActions.setTypeAction(EVT, Expand);
515 // Inspect all of the ValueType's smaller than the largest integer
516 // register to see which ones need promotion.
517 unsigned LegalIntReg = LargestIntReg;
518 for (unsigned IntReg = LargestIntReg - 1;
519 IntReg >= (unsigned)MVT::i1; --IntReg) {
520 MVT IVT = (MVT::SimpleValueType)IntReg;
521 if (isTypeLegal(IVT)) {
522 LegalIntReg = IntReg;
524 RegisterTypeForVT[IntReg] = TransformToType[IntReg] =
525 (MVT::SimpleValueType)LegalIntReg;
526 ValueTypeActions.setTypeAction(IVT, Promote);
530 // ppcf128 type is really two f64's.
531 if (!isTypeLegal(MVT::ppcf128)) {
532 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64];
533 RegisterTypeForVT[MVT::ppcf128] = MVT::f64;
534 TransformToType[MVT::ppcf128] = MVT::f64;
535 ValueTypeActions.setTypeAction(MVT::ppcf128, Expand);
538 // Decide how to handle f64. If the target does not have native f64 support,
539 // expand it to i64 and we will be generating soft float library calls.
540 if (!isTypeLegal(MVT::f64)) {
541 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
542 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
543 TransformToType[MVT::f64] = MVT::i64;
544 ValueTypeActions.setTypeAction(MVT::f64, Expand);
547 // Decide how to handle f32. If the target does not have native support for
548 // f32, promote it to f64 if it is legal. Otherwise, expand it to i32.
549 if (!isTypeLegal(MVT::f32)) {
550 if (isTypeLegal(MVT::f64)) {
551 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::f64];
552 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::f64];
553 TransformToType[MVT::f32] = MVT::f64;
554 ValueTypeActions.setTypeAction(MVT::f32, Promote);
556 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
557 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
558 TransformToType[MVT::f32] = MVT::i32;
559 ValueTypeActions.setTypeAction(MVT::f32, Expand);
563 // Loop over all of the vector value types to see which need transformations.
564 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
565 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
566 MVT VT = (MVT::SimpleValueType)i;
567 if (!isTypeLegal(VT)) {
568 MVT IntermediateVT, RegisterVT;
569 unsigned NumIntermediates;
570 NumRegistersForVT[i] =
571 getVectorTypeBreakdown(VT,
572 IntermediateVT, NumIntermediates,
574 RegisterTypeForVT[i] = RegisterVT;
576 // Determine if there is a legal wider type.
577 bool IsLegalWiderType = false;
578 MVT EltVT = VT.getVectorElementType();
579 unsigned NElts = VT.getVectorNumElements();
580 for (unsigned nVT = i+1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
581 MVT SVT = (MVT::SimpleValueType)nVT;
582 if (isTypeLegal(SVT) && SVT.getVectorElementType() == EltVT &&
583 SVT.getVectorNumElements() > NElts) {
584 TransformToType[i] = SVT;
585 ValueTypeActions.setTypeAction(VT, Promote);
586 IsLegalWiderType = true;
590 if (!IsLegalWiderType) {
591 MVT NVT = VT.getPow2VectorType();
593 // Type is already a power of 2. The default action is to split.
594 TransformToType[i] = MVT::Other;
595 ValueTypeActions.setTypeAction(VT, Expand);
597 TransformToType[i] = NVT;
598 ValueTypeActions.setTypeAction(VT, Promote);
605 const char *TargetLowering::getTargetNodeName(unsigned Opcode) const {
610 MVT TargetLowering::getSetCCResultType(const SDValue &) const {
611 return getValueType(TD->getIntPtrType());
615 /// getVectorTypeBreakdown - Vector types are broken down into some number of
616 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
617 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
618 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
620 /// This method returns the number of registers needed, and the VT for each
621 /// register. It also returns the VT and quantity of the intermediate values
622 /// before they are promoted/expanded.
624 unsigned TargetLowering::getVectorTypeBreakdown(MVT VT,
626 unsigned &NumIntermediates,
627 MVT &RegisterVT) const {
628 // Figure out the right, legal destination reg to copy into.
629 unsigned NumElts = VT.getVectorNumElements();
630 MVT EltTy = VT.getVectorElementType();
632 unsigned NumVectorRegs = 1;
634 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
635 // could break down into LHS/RHS like LegalizeDAG does.
636 if (!isPowerOf2_32(NumElts)) {
637 NumVectorRegs = NumElts;
641 // Divide the input until we get to a supported size. This will always
642 // end with a scalar if the target doesn't support vectors.
643 while (NumElts > 1 && !isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) {
648 NumIntermediates = NumVectorRegs;
650 MVT NewVT = MVT::getVectorVT(EltTy, NumElts);
651 if (!isTypeLegal(NewVT))
653 IntermediateVT = NewVT;
655 MVT DestVT = getTypeToTransformTo(NewVT);
657 if (DestVT.bitsLT(NewVT)) {
658 // Value is expanded, e.g. i64 -> i16.
659 return NumVectorRegs*(NewVT.getSizeInBits()/DestVT.getSizeInBits());
661 // Otherwise, promotion or legal types use the same number of registers as
662 // the vector decimated to the appropriate level.
663 return NumVectorRegs;
669 /// getWidenVectorType: given a vector type, returns the type to widen to
670 /// (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
671 /// If there is no vector type that we want to widen to, returns MVT::Other
672 /// When and where to widen is target dependent based on the cost of
673 /// scalarizing vs using the wider vector type.
674 MVT TargetLowering::getWidenVectorType(MVT VT) {
675 assert(VT.isVector());
679 // Default is not to widen until moved to LegalizeTypes
683 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
684 /// function arguments in the caller parameter area. This is the actual
685 /// alignment, not its logarithm.
686 unsigned TargetLowering::getByValTypeAlignment(const Type *Ty) const {
687 return TD->getCallFrameTypeAlignment(Ty);
690 SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table,
691 SelectionDAG &DAG) const {
692 if (usesGlobalOffsetTable())
693 return DAG.getNode(ISD::GLOBAL_OFFSET_TABLE, getPointerTy());
698 TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
699 // Assume that everything is safe in static mode.
700 if (getTargetMachine().getRelocationModel() == Reloc::Static)
703 // In dynamic-no-pic mode, assume that known defined values are safe.
704 if (getTargetMachine().getRelocationModel() == Reloc::DynamicNoPIC &&
706 !GA->getGlobal()->isDeclaration() &&
707 !GA->getGlobal()->mayBeOverridden())
710 // Otherwise assume nothing is safe.
714 //===----------------------------------------------------------------------===//
715 // Optimization Methods
716 //===----------------------------------------------------------------------===//
718 /// ShrinkDemandedConstant - Check to see if the specified operand of the
719 /// specified instruction is a constant integer. If so, check to see if there
720 /// are any bits set in the constant that are not demanded. If so, shrink the
721 /// constant and return true.
722 bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDValue Op,
723 const APInt &Demanded) {
724 // FIXME: ISD::SELECT, ISD::SELECT_CC
725 switch(Op.getOpcode()) {
730 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1)))
731 if (C->getAPIntValue().intersects(~Demanded)) {
732 MVT VT = Op.getValueType();
733 SDValue New = DAG.getNode(Op.getOpcode(), VT, Op.getOperand(0),
734 DAG.getConstant(Demanded &
737 return CombineTo(Op, New);
744 /// SimplifyDemandedBits - Look at Op. At this point, we know that only the
745 /// DemandedMask bits of the result of Op are ever used downstream. If we can
746 /// use this information to simplify Op, create a new simplified DAG node and
747 /// return true, returning the original and new nodes in Old and New. Otherwise,
748 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
749 /// the expression (used to simplify the caller). The KnownZero/One bits may
750 /// only be accurate for those bits in the DemandedMask.
751 bool TargetLowering::SimplifyDemandedBits(SDValue Op,
752 const APInt &DemandedMask,
755 TargetLoweringOpt &TLO,
756 unsigned Depth) const {
757 unsigned BitWidth = DemandedMask.getBitWidth();
758 assert(Op.getValueSizeInBits() == BitWidth &&
759 "Mask size mismatches value type size!");
760 APInt NewMask = DemandedMask;
762 // Don't know anything.
763 KnownZero = KnownOne = APInt(BitWidth, 0);
765 // Other users may use these bits.
766 if (!Op.getNode()->hasOneUse()) {
768 // If not at the root, Just compute the KnownZero/KnownOne bits to
769 // simplify things downstream.
770 TLO.DAG.ComputeMaskedBits(Op, DemandedMask, KnownZero, KnownOne, Depth);
773 // If this is the root being simplified, allow it to have multiple uses,
774 // just set the NewMask to all bits.
775 NewMask = APInt::getAllOnesValue(BitWidth);
776 } else if (DemandedMask == 0) {
777 // Not demanding any bits from Op.
778 if (Op.getOpcode() != ISD::UNDEF)
779 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::UNDEF, Op.getValueType()));
781 } else if (Depth == 6) { // Limit search depth.
785 APInt KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut;
786 switch (Op.getOpcode()) {
788 // We know all of the bits for a constant!
789 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue() & NewMask;
790 KnownZero = ~KnownOne & NewMask;
791 return false; // Don't fall through, will infinitely loop.
793 // If the RHS is a constant, check to see if the LHS would be zero without
794 // using the bits from the RHS. Below, we use knowledge about the RHS to
795 // simplify the LHS, here we're using information from the LHS to simplify
797 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
798 APInt LHSZero, LHSOne;
799 TLO.DAG.ComputeMaskedBits(Op.getOperand(0), NewMask,
800 LHSZero, LHSOne, Depth+1);
801 // If the LHS already has zeros where RHSC does, this and is dead.
802 if ((LHSZero & NewMask) == (~RHSC->getAPIntValue() & NewMask))
803 return TLO.CombineTo(Op, Op.getOperand(0));
804 // If any of the set bits in the RHS are known zero on the LHS, shrink
806 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & NewMask))
810 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
811 KnownOne, TLO, Depth+1))
813 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
814 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownZero & NewMask,
815 KnownZero2, KnownOne2, TLO, Depth+1))
817 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
819 // If all of the demanded bits are known one on one side, return the other.
820 // These bits cannot contribute to the result of the 'and'.
821 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
822 return TLO.CombineTo(Op, Op.getOperand(0));
823 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
824 return TLO.CombineTo(Op, Op.getOperand(1));
825 // If all of the demanded bits in the inputs are known zeros, return zero.
826 if ((NewMask & (KnownZero|KnownZero2)) == NewMask)
827 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, Op.getValueType()));
828 // If the RHS is a constant, see if we can simplify it.
829 if (TLO.ShrinkDemandedConstant(Op, ~KnownZero2 & NewMask))
832 // Output known-1 bits are only known if set in both the LHS & RHS.
833 KnownOne &= KnownOne2;
834 // Output known-0 are known to be clear if zero in either the LHS | RHS.
835 KnownZero |= KnownZero2;
838 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
839 KnownOne, TLO, Depth+1))
841 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
842 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownOne & NewMask,
843 KnownZero2, KnownOne2, TLO, Depth+1))
845 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
847 // If all of the demanded bits are known zero on one side, return the other.
848 // These bits cannot contribute to the result of the 'or'.
849 if ((NewMask & ~KnownOne2 & KnownZero) == (~KnownOne2 & NewMask))
850 return TLO.CombineTo(Op, Op.getOperand(0));
851 if ((NewMask & ~KnownOne & KnownZero2) == (~KnownOne & NewMask))
852 return TLO.CombineTo(Op, Op.getOperand(1));
853 // If all of the potentially set bits on one side are known to be set on
854 // the other side, just use the 'other' side.
855 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
856 return TLO.CombineTo(Op, Op.getOperand(0));
857 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
858 return TLO.CombineTo(Op, Op.getOperand(1));
859 // If the RHS is a constant, see if we can simplify it.
860 if (TLO.ShrinkDemandedConstant(Op, NewMask))
863 // Output known-0 bits are only known if clear in both the LHS & RHS.
864 KnownZero &= KnownZero2;
865 // Output known-1 are known to be set if set in either the LHS | RHS.
866 KnownOne |= KnownOne2;
869 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
870 KnownOne, TLO, Depth+1))
872 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
873 if (SimplifyDemandedBits(Op.getOperand(0), NewMask, KnownZero2,
874 KnownOne2, TLO, Depth+1))
876 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
878 // If all of the demanded bits are known zero on one side, return the other.
879 // These bits cannot contribute to the result of the 'xor'.
880 if ((KnownZero & NewMask) == NewMask)
881 return TLO.CombineTo(Op, Op.getOperand(0));
882 if ((KnownZero2 & NewMask) == NewMask)
883 return TLO.CombineTo(Op, Op.getOperand(1));
885 // If all of the unknown bits are known to be zero on one side or the other
886 // (but not both) turn this into an *inclusive* or.
887 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
888 if ((NewMask & ~KnownZero & ~KnownZero2) == 0)
889 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, Op.getValueType(),
893 // Output known-0 bits are known if clear or set in both the LHS & RHS.
894 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
895 // Output known-1 are known to be set if set in only one of the LHS, RHS.
896 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
898 // If all of the demanded bits on one side are known, and all of the set
899 // bits on that side are also known to be set on the other side, turn this
900 // into an AND, as we know the bits will be cleared.
901 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
902 if ((NewMask & (KnownZero|KnownOne)) == NewMask) { // all known
903 if ((KnownOne & KnownOne2) == KnownOne) {
904 MVT VT = Op.getValueType();
905 SDValue ANDC = TLO.DAG.getConstant(~KnownOne & NewMask, VT);
906 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, VT, Op.getOperand(0),
911 // If the RHS is a constant, see if we can simplify it.
912 // for XOR, we prefer to force bits to 1 if they will make a -1.
913 // if we can't force bits, try to shrink constant
914 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
915 APInt Expanded = C->getAPIntValue() | (~NewMask);
916 // if we can expand it to have all bits set, do it
917 if (Expanded.isAllOnesValue()) {
918 if (Expanded != C->getAPIntValue()) {
919 MVT VT = Op.getValueType();
920 SDValue New = TLO.DAG.getNode(Op.getOpcode(), VT, Op.getOperand(0),
921 TLO.DAG.getConstant(Expanded, VT));
922 return TLO.CombineTo(Op, New);
924 // if it already has all the bits set, nothing to change
925 // but don't shrink either!
926 } else if (TLO.ShrinkDemandedConstant(Op, NewMask)) {
931 KnownZero = KnownZeroOut;
932 KnownOne = KnownOneOut;
935 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero,
936 KnownOne, TLO, Depth+1))
938 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero2,
939 KnownOne2, TLO, Depth+1))
941 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
942 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
944 // If the operands are constants, see if we can simplify them.
945 if (TLO.ShrinkDemandedConstant(Op, NewMask))
948 // Only known if known in both the LHS and RHS.
949 KnownOne &= KnownOne2;
950 KnownZero &= KnownZero2;
953 if (SimplifyDemandedBits(Op.getOperand(3), NewMask, KnownZero,
954 KnownOne, TLO, Depth+1))
956 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero2,
957 KnownOne2, TLO, Depth+1))
959 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
960 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
962 // If the operands are constants, see if we can simplify them.
963 if (TLO.ShrinkDemandedConstant(Op, NewMask))
966 // Only known if known in both the LHS and RHS.
967 KnownOne &= KnownOne2;
968 KnownZero &= KnownZero2;
971 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
972 unsigned ShAmt = SA->getZExtValue();
973 SDValue InOp = Op.getOperand(0);
975 // If the shift count is an invalid immediate, don't do anything.
976 if (ShAmt >= BitWidth)
979 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a
980 // single shift. We can do this if the bottom bits (which are shifted
981 // out) are never demanded.
982 if (InOp.getOpcode() == ISD::SRL &&
983 isa<ConstantSDNode>(InOp.getOperand(1))) {
984 if (ShAmt && (NewMask & APInt::getLowBitsSet(BitWidth, ShAmt)) == 0) {
985 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
986 unsigned Opc = ISD::SHL;
994 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
995 MVT VT = Op.getValueType();
996 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, VT,
997 InOp.getOperand(0), NewSA));
1001 if (SimplifyDemandedBits(Op.getOperand(0), NewMask.lshr(ShAmt),
1002 KnownZero, KnownOne, TLO, Depth+1))
1004 KnownZero <<= SA->getZExtValue();
1005 KnownOne <<= SA->getZExtValue();
1006 // low bits known zero.
1007 KnownZero |= APInt::getLowBitsSet(BitWidth, SA->getZExtValue());
1011 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1012 MVT VT = Op.getValueType();
1013 unsigned ShAmt = SA->getZExtValue();
1014 unsigned VTSize = VT.getSizeInBits();
1015 SDValue InOp = Op.getOperand(0);
1017 // If the shift count is an invalid immediate, don't do anything.
1018 if (ShAmt >= BitWidth)
1021 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a
1022 // single shift. We can do this if the top bits (which are shifted out)
1023 // are never demanded.
1024 if (InOp.getOpcode() == ISD::SHL &&
1025 isa<ConstantSDNode>(InOp.getOperand(1))) {
1026 if (ShAmt && (NewMask & APInt::getHighBitsSet(VTSize, ShAmt)) == 0) {
1027 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
1028 unsigned Opc = ISD::SRL;
1029 int Diff = ShAmt-C1;
1036 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
1037 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, VT,
1038 InOp.getOperand(0), NewSA));
1042 // Compute the new bits that are at the top now.
1043 if (SimplifyDemandedBits(InOp, (NewMask << ShAmt),
1044 KnownZero, KnownOne, TLO, Depth+1))
1046 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1047 KnownZero = KnownZero.lshr(ShAmt);
1048 KnownOne = KnownOne.lshr(ShAmt);
1050 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
1051 KnownZero |= HighBits; // High bits known zero.
1055 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1056 MVT VT = Op.getValueType();
1057 unsigned ShAmt = SA->getZExtValue();
1059 // If the shift count is an invalid immediate, don't do anything.
1060 if (ShAmt >= BitWidth)
1063 APInt InDemandedMask = (NewMask << ShAmt);
1065 // If any of the demanded bits are produced by the sign extension, we also
1066 // demand the input sign bit.
1067 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
1068 if (HighBits.intersects(NewMask))
1069 InDemandedMask |= APInt::getSignBit(VT.getSizeInBits());
1071 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask,
1072 KnownZero, KnownOne, TLO, Depth+1))
1074 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1075 KnownZero = KnownZero.lshr(ShAmt);
1076 KnownOne = KnownOne.lshr(ShAmt);
1078 // Handle the sign bit, adjusted to where it is now in the mask.
1079 APInt SignBit = APInt::getSignBit(BitWidth).lshr(ShAmt);
1081 // If the input sign bit is known to be zero, or if none of the top bits
1082 // are demanded, turn this into an unsigned shift right.
1083 if (KnownZero.intersects(SignBit) || (HighBits & ~NewMask) == HighBits) {
1084 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, VT, Op.getOperand(0),
1086 } else if (KnownOne.intersects(SignBit)) { // New bits are known one.
1087 KnownOne |= HighBits;
1091 case ISD::SIGN_EXTEND_INREG: {
1092 MVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1094 // Sign extension. Compute the demanded bits in the result that are not
1095 // present in the input.
1096 APInt NewBits = APInt::getHighBitsSet(BitWidth,
1097 BitWidth - EVT.getSizeInBits()) &
1100 // If none of the extended bits are demanded, eliminate the sextinreg.
1102 return TLO.CombineTo(Op, Op.getOperand(0));
1104 APInt InSignBit = APInt::getSignBit(EVT.getSizeInBits());
1105 InSignBit.zext(BitWidth);
1106 APInt InputDemandedBits = APInt::getLowBitsSet(BitWidth,
1107 EVT.getSizeInBits()) &
1110 // Since the sign extended bits are demanded, we know that the sign
1112 InputDemandedBits |= InSignBit;
1114 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits,
1115 KnownZero, KnownOne, TLO, Depth+1))
1117 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1119 // If the sign bit of the input is known set or clear, then we know the
1120 // top bits of the result.
1122 // If the input sign bit is known zero, convert this into a zero extension.
1123 if (KnownZero.intersects(InSignBit))
1124 return TLO.CombineTo(Op,
1125 TLO.DAG.getZeroExtendInReg(Op.getOperand(0), EVT));
1127 if (KnownOne.intersects(InSignBit)) { // Input sign bit known set
1128 KnownOne |= NewBits;
1129 KnownZero &= ~NewBits;
1130 } else { // Input sign bit unknown
1131 KnownZero &= ~NewBits;
1132 KnownOne &= ~NewBits;
1136 case ISD::ZERO_EXTEND: {
1137 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits();
1138 APInt InMask = NewMask;
1139 InMask.trunc(OperandBitWidth);
1141 // If none of the top bits are demanded, convert this into an any_extend.
1143 APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask;
1144 if (!NewBits.intersects(NewMask))
1145 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND,
1149 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1150 KnownZero, KnownOne, TLO, Depth+1))
1152 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1153 KnownZero.zext(BitWidth);
1154 KnownOne.zext(BitWidth);
1155 KnownZero |= NewBits;
1158 case ISD::SIGN_EXTEND: {
1159 MVT InVT = Op.getOperand(0).getValueType();
1160 unsigned InBits = InVT.getSizeInBits();
1161 APInt InMask = APInt::getLowBitsSet(BitWidth, InBits);
1162 APInt InSignBit = APInt::getBitsSet(BitWidth, InBits - 1, InBits);
1163 APInt NewBits = ~InMask & NewMask;
1165 // If none of the top bits are demanded, convert this into an any_extend.
1167 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND,Op.getValueType(),
1170 // Since some of the sign extended bits are demanded, we know that the sign
1172 APInt InDemandedBits = InMask & NewMask;
1173 InDemandedBits |= InSignBit;
1174 InDemandedBits.trunc(InBits);
1176 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero,
1177 KnownOne, TLO, Depth+1))
1179 KnownZero.zext(BitWidth);
1180 KnownOne.zext(BitWidth);
1182 // If the sign bit is known zero, convert this to a zero extend.
1183 if (KnownZero.intersects(InSignBit))
1184 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND,
1188 // If the sign bit is known one, the top bits match.
1189 if (KnownOne.intersects(InSignBit)) {
1190 KnownOne |= NewBits;
1191 KnownZero &= ~NewBits;
1192 } else { // Otherwise, top bits aren't known.
1193 KnownOne &= ~NewBits;
1194 KnownZero &= ~NewBits;
1198 case ISD::ANY_EXTEND: {
1199 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits();
1200 APInt InMask = NewMask;
1201 InMask.trunc(OperandBitWidth);
1202 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1203 KnownZero, KnownOne, TLO, Depth+1))
1205 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1206 KnownZero.zext(BitWidth);
1207 KnownOne.zext(BitWidth);
1210 case ISD::TRUNCATE: {
1211 // Simplify the input, using demanded bit information, and compute the known
1212 // zero/one bits live out.
1213 APInt TruncMask = NewMask;
1214 TruncMask.zext(Op.getOperand(0).getValueSizeInBits());
1215 if (SimplifyDemandedBits(Op.getOperand(0), TruncMask,
1216 KnownZero, KnownOne, TLO, Depth+1))
1218 KnownZero.trunc(BitWidth);
1219 KnownOne.trunc(BitWidth);
1221 // If the input is only used by this truncate, see if we can shrink it based
1222 // on the known demanded bits.
1223 if (Op.getOperand(0).getNode()->hasOneUse()) {
1224 SDValue In = Op.getOperand(0);
1225 unsigned InBitWidth = In.getValueSizeInBits();
1226 switch (In.getOpcode()) {
1229 // Shrink SRL by a constant if none of the high bits shifted in are
1231 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1))){
1232 APInt HighBits = APInt::getHighBitsSet(InBitWidth,
1233 InBitWidth - BitWidth);
1234 HighBits = HighBits.lshr(ShAmt->getZExtValue());
1235 HighBits.trunc(BitWidth);
1237 if (ShAmt->getZExtValue() < BitWidth && !(HighBits & NewMask)) {
1238 // None of the shifted in bits are needed. Add a truncate of the
1239 // shift input, then shift it.
1240 SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE,
1243 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL,Op.getValueType(),
1244 NewTrunc, In.getOperand(1)));
1251 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1254 case ISD::AssertZext: {
1255 MVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1256 APInt InMask = APInt::getLowBitsSet(BitWidth,
1257 VT.getSizeInBits());
1258 if (SimplifyDemandedBits(Op.getOperand(0), InMask & NewMask,
1259 KnownZero, KnownOne, TLO, Depth+1))
1261 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1262 KnownZero |= ~InMask & NewMask;
1265 case ISD::BIT_CONVERT:
1267 // If this is an FP->Int bitcast and if the sign bit is the only thing that
1268 // is demanded, turn this into a FGETSIGN.
1269 if (NewMask == MVT::getIntegerVTSignBit(Op.getValueType()) &&
1270 MVT::isFloatingPoint(Op.getOperand(0).getValueType()) &&
1271 !MVT::isVector(Op.getOperand(0).getValueType())) {
1272 // Only do this xform if FGETSIGN is valid or if before legalize.
1273 if (!TLO.AfterLegalize ||
1274 isOperationLegal(ISD::FGETSIGN, Op.getValueType())) {
1275 // Make a FGETSIGN + SHL to move the sign bit into the appropriate
1276 // place. We expect the SHL to be eliminated by other optimizations.
1277 SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, Op.getValueType(),
1279 unsigned ShVal = Op.getValueType().getSizeInBits()-1;
1280 SDValue ShAmt = TLO.DAG.getConstant(ShVal, getShiftAmountTy());
1281 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, Op.getValueType(),
1288 // Just use ComputeMaskedBits to compute output bits.
1289 TLO.DAG.ComputeMaskedBits(Op, NewMask, KnownZero, KnownOne, Depth);
1293 // If we know the value of all of the demanded bits, return this as a
1295 if ((NewMask & (KnownZero|KnownOne)) == NewMask)
1296 return TLO.CombineTo(Op, TLO.DAG.getConstant(KnownOne, Op.getValueType()));
1301 /// computeMaskedBitsForTargetNode - Determine which of the bits specified
1302 /// in Mask are known to be either zero or one and return them in the
1303 /// KnownZero/KnownOne bitsets.
1304 void TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
1308 const SelectionDAG &DAG,
1309 unsigned Depth) const {
1310 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1311 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1312 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1313 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1314 "Should use MaskedValueIsZero if you don't know whether Op"
1315 " is a target node!");
1316 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
1319 /// ComputeNumSignBitsForTargetNode - This method can be implemented by
1320 /// targets that want to expose additional information about sign bits to the
1322 unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
1323 unsigned Depth) const {
1324 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1325 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1326 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1327 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1328 "Should use ComputeNumSignBits if you don't know whether Op"
1329 " is a target node!");
1334 /// SimplifySetCC - Try to simplify a setcc built with the specified operands
1335 /// and cc. If it is unable to simplify it, return a null SDValue.
1337 TargetLowering::SimplifySetCC(MVT VT, SDValue N0, SDValue N1,
1338 ISD::CondCode Cond, bool foldBooleans,
1339 DAGCombinerInfo &DCI) const {
1340 SelectionDAG &DAG = DCI.DAG;
1342 // These setcc operations always fold.
1346 case ISD::SETFALSE2: return DAG.getConstant(0, VT);
1348 case ISD::SETTRUE2: return DAG.getConstant(1, VT);
1351 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) {
1352 const APInt &C1 = N1C->getAPIntValue();
1353 if (isa<ConstantSDNode>(N0.getNode())) {
1354 return DAG.FoldSetCC(VT, N0, N1, Cond);
1356 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an
1357 // equality comparison, then we're just comparing whether X itself is
1359 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) &&
1360 N0.getOperand(0).getOpcode() == ISD::CTLZ &&
1361 N0.getOperand(1).getOpcode() == ISD::Constant) {
1362 unsigned ShAmt = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
1363 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1364 ShAmt == Log2_32(N0.getValueType().getSizeInBits())) {
1365 if ((C1 == 0) == (Cond == ISD::SETEQ)) {
1366 // (srl (ctlz x), 5) == 0 -> X != 0
1367 // (srl (ctlz x), 5) != 1 -> X != 0
1370 // (srl (ctlz x), 5) != 0 -> X == 0
1371 // (srl (ctlz x), 5) == 1 -> X == 0
1374 SDValue Zero = DAG.getConstant(0, N0.getValueType());
1375 return DAG.getSetCC(VT, N0.getOperand(0).getOperand(0),
1380 // If the LHS is '(and load, const)', the RHS is 0,
1381 // the test is for equality or unsigned, and all 1 bits of the const are
1382 // in the same partial word, see if we can shorten the load.
1383 if (DCI.isBeforeLegalize() &&
1384 N0.getOpcode() == ISD::AND && C1 == 0 &&
1385 isa<LoadSDNode>(N0.getOperand(0)) &&
1386 N0.getOperand(0).getNode()->hasOneUse() &&
1387 isa<ConstantSDNode>(N0.getOperand(1))) {
1388 LoadSDNode *Lod = cast<LoadSDNode>(N0.getOperand(0));
1389 uint64_t Mask = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
1390 uint64_t bestMask = 0;
1391 unsigned bestWidth = 0, bestOffset = 0;
1392 if (!Lod->isVolatile() && Lod->isUnindexed()) {
1393 unsigned origWidth = N0.getValueType().getSizeInBits();
1394 // We can narrow (e.g.) 16-bit extending loads on 32-bit target to
1395 // 8 bits, but have to be careful...
1396 if (Lod->getExtensionType() != ISD::NON_EXTLOAD)
1397 origWidth = Lod->getMemoryVT().getSizeInBits();
1398 for (unsigned width = origWidth / 2; width>=8; width /= 2) {
1399 uint64_t newMask = (1ULL << width) - 1;
1400 for (unsigned offset=0; offset<origWidth/width; offset++) {
1401 if ((newMask & Mask)==Mask) {
1402 if (!TD->isLittleEndian())
1403 bestOffset = (origWidth/width - offset - 1) * (width/8);
1405 bestOffset = (uint64_t)offset * (width/8);
1406 bestMask = Mask >> (offset * (width/8) * 8);
1410 newMask = newMask << width;
1415 MVT newVT = MVT::getIntegerVT(bestWidth);
1416 if (newVT.isRound()) {
1417 MVT PtrType = Lod->getOperand(1).getValueType();
1418 SDValue Ptr = Lod->getBasePtr();
1419 if (bestOffset != 0)
1420 Ptr = DAG.getNode(ISD::ADD, PtrType, Lod->getBasePtr(),
1421 DAG.getConstant(bestOffset, PtrType));
1422 unsigned NewAlign = MinAlign(Lod->getAlignment(), bestOffset);
1423 SDValue NewLoad = DAG.getLoad(newVT, Lod->getChain(), Ptr,
1425 Lod->getSrcValueOffset() + bestOffset,
1427 return DAG.getSetCC(VT, DAG.getNode(ISD::AND, newVT, NewLoad,
1428 DAG.getConstant(bestMask, newVT)),
1429 DAG.getConstant(0LL, newVT), Cond);
1434 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
1435 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
1436 unsigned InSize = N0.getOperand(0).getValueType().getSizeInBits();
1438 // If the comparison constant has bits in the upper part, the
1439 // zero-extended value could never match.
1440 if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(),
1441 C1.getBitWidth() - InSize))) {
1445 case ISD::SETEQ: return DAG.getConstant(0, VT);
1448 case ISD::SETNE: return DAG.getConstant(1, VT);
1451 // True if the sign bit of C1 is set.
1452 return DAG.getConstant(C1.isNegative(), VT);
1455 // True if the sign bit of C1 isn't set.
1456 return DAG.getConstant(C1.isNonNegative(), VT);
1462 // Otherwise, we can perform the comparison with the low bits.
1470 return DAG.getSetCC(VT, N0.getOperand(0),
1471 DAG.getConstant(APInt(C1).trunc(InSize),
1472 N0.getOperand(0).getValueType()),
1475 break; // todo, be more careful with signed comparisons
1477 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
1478 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1479 MVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
1480 unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits();
1481 MVT ExtDstTy = N0.getValueType();
1482 unsigned ExtDstTyBits = ExtDstTy.getSizeInBits();
1484 // If the extended part has any inconsistent bits, it cannot ever
1485 // compare equal. In other words, they have to be all ones or all
1488 APInt::getHighBitsSet(ExtDstTyBits, ExtDstTyBits - ExtSrcTyBits);
1489 if ((C1 & ExtBits) != 0 && (C1 & ExtBits) != ExtBits)
1490 return DAG.getConstant(Cond == ISD::SETNE, VT);
1493 MVT Op0Ty = N0.getOperand(0).getValueType();
1494 if (Op0Ty == ExtSrcTy) {
1495 ZextOp = N0.getOperand(0);
1497 APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits);
1498 ZextOp = DAG.getNode(ISD::AND, Op0Ty, N0.getOperand(0),
1499 DAG.getConstant(Imm, Op0Ty));
1501 if (!DCI.isCalledByLegalizer())
1502 DCI.AddToWorklist(ZextOp.getNode());
1503 // Otherwise, make this a use of a zext.
1504 return DAG.getSetCC(VT, ZextOp,
1505 DAG.getConstant(C1 & APInt::getLowBitsSet(
1510 } else if ((N1C->isNullValue() || N1C->getAPIntValue() == 1) &&
1511 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1513 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC
1514 if (N0.getOpcode() == ISD::SETCC) {
1515 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getZExtValue() != 1);
1519 // Invert the condition.
1520 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
1521 CC = ISD::getSetCCInverse(CC,
1522 N0.getOperand(0).getValueType().isInteger());
1523 return DAG.getSetCC(VT, N0.getOperand(0), N0.getOperand(1), CC);
1526 if ((N0.getOpcode() == ISD::XOR ||
1527 (N0.getOpcode() == ISD::AND &&
1528 N0.getOperand(0).getOpcode() == ISD::XOR &&
1529 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
1530 isa<ConstantSDNode>(N0.getOperand(1)) &&
1531 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue() == 1) {
1532 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We
1533 // can only do this if the top bits are known zero.
1534 unsigned BitWidth = N0.getValueSizeInBits();
1535 if (DAG.MaskedValueIsZero(N0,
1536 APInt::getHighBitsSet(BitWidth,
1538 // Okay, get the un-inverted input value.
1540 if (N0.getOpcode() == ISD::XOR)
1541 Val = N0.getOperand(0);
1543 assert(N0.getOpcode() == ISD::AND &&
1544 N0.getOperand(0).getOpcode() == ISD::XOR);
1545 // ((X^1)&1)^1 -> X & 1
1546 Val = DAG.getNode(ISD::AND, N0.getValueType(),
1547 N0.getOperand(0).getOperand(0),
1550 return DAG.getSetCC(VT, Val, N1,
1551 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
1556 APInt MinVal, MaxVal;
1557 unsigned OperandBitSize = N1C->getValueType(0).getSizeInBits();
1558 if (ISD::isSignedIntSetCC(Cond)) {
1559 MinVal = APInt::getSignedMinValue(OperandBitSize);
1560 MaxVal = APInt::getSignedMaxValue(OperandBitSize);
1562 MinVal = APInt::getMinValue(OperandBitSize);
1563 MaxVal = APInt::getMaxValue(OperandBitSize);
1566 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
1567 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
1568 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true
1569 // X >= C0 --> X > (C0-1)
1570 return DAG.getSetCC(VT, N0, DAG.getConstant(C1-1, N1.getValueType()),
1571 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
1574 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
1575 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true
1576 // X <= C0 --> X < (C0+1)
1577 return DAG.getSetCC(VT, N0, DAG.getConstant(C1+1, N1.getValueType()),
1578 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
1581 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
1582 return DAG.getConstant(0, VT); // X < MIN --> false
1583 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal)
1584 return DAG.getConstant(1, VT); // X >= MIN --> true
1585 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal)
1586 return DAG.getConstant(0, VT); // X > MAX --> false
1587 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal)
1588 return DAG.getConstant(1, VT); // X <= MAX --> true
1590 // Canonicalize setgt X, Min --> setne X, Min
1591 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
1592 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
1593 // Canonicalize setlt X, Max --> setne X, Max
1594 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
1595 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
1597 // If we have setult X, 1, turn it into seteq X, 0
1598 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
1599 return DAG.getSetCC(VT, N0, DAG.getConstant(MinVal, N0.getValueType()),
1601 // If we have setugt X, Max-1, turn it into seteq X, Max
1602 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
1603 return DAG.getSetCC(VT, N0, DAG.getConstant(MaxVal, N0.getValueType()),
1606 // If we have "setcc X, C0", check to see if we can shrink the immediate
1609 // SETUGT X, SINTMAX -> SETLT X, 0
1610 if (Cond == ISD::SETUGT &&
1611 C1 == APInt::getSignedMaxValue(OperandBitSize))
1612 return DAG.getSetCC(VT, N0, DAG.getConstant(0, N1.getValueType()),
1615 // SETULT X, SINTMIN -> SETGT X, -1
1616 if (Cond == ISD::SETULT &&
1617 C1 == APInt::getSignedMinValue(OperandBitSize)) {
1618 SDValue ConstMinusOne =
1619 DAG.getConstant(APInt::getAllOnesValue(OperandBitSize),
1621 return DAG.getSetCC(VT, N0, ConstMinusOne, ISD::SETGT);
1624 // Fold bit comparisons when we can.
1625 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1626 VT == N0.getValueType() && N0.getOpcode() == ISD::AND)
1627 if (ConstantSDNode *AndRHS =
1628 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1629 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
1630 // Perform the xform if the AND RHS is a single bit.
1631 if (isPowerOf2_64(AndRHS->getZExtValue())) {
1632 return DAG.getNode(ISD::SRL, VT, N0,
1633 DAG.getConstant(Log2_64(AndRHS->getZExtValue()),
1634 getShiftAmountTy()));
1636 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getZExtValue()) {
1637 // (X & 8) == 8 --> (X & 8) >> 3
1638 // Perform the xform if C1 is a single bit.
1639 if (C1.isPowerOf2()) {
1640 return DAG.getNode(ISD::SRL, VT, N0,
1641 DAG.getConstant(C1.logBase2(), getShiftAmountTy()));
1646 } else if (isa<ConstantSDNode>(N0.getNode())) {
1647 // Ensure that the constant occurs on the RHS.
1648 return DAG.getSetCC(VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
1651 if (isa<ConstantFPSDNode>(N0.getNode())) {
1652 // Constant fold or commute setcc.
1653 SDValue O = DAG.FoldSetCC(VT, N0, N1, Cond);
1654 if (O.getNode()) return O;
1655 } else if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1.getNode())) {
1656 // If the RHS of an FP comparison is a constant, simplify it away in
1658 if (CFP->getValueAPF().isNaN()) {
1659 // If an operand is known to be a nan, we can fold it.
1660 switch (ISD::getUnorderedFlavor(Cond)) {
1661 default: assert(0 && "Unknown flavor!");
1662 case 0: // Known false.
1663 return DAG.getConstant(0, VT);
1664 case 1: // Known true.
1665 return DAG.getConstant(1, VT);
1666 case 2: // Undefined.
1667 return DAG.getNode(ISD::UNDEF, VT);
1671 // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the
1672 // constant if knowing that the operand is non-nan is enough. We prefer to
1673 // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to
1675 if (Cond == ISD::SETO || Cond == ISD::SETUO)
1676 return DAG.getSetCC(VT, N0, N0, Cond);
1680 // We can always fold X == X for integer setcc's.
1681 if (N0.getValueType().isInteger())
1682 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
1683 unsigned UOF = ISD::getUnorderedFlavor(Cond);
1684 if (UOF == 2) // FP operators that are undefined on NaNs.
1685 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
1686 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
1687 return DAG.getConstant(UOF, VT);
1688 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
1689 // if it is not already.
1690 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
1691 if (NewCond != Cond)
1692 return DAG.getSetCC(VT, N0, N1, NewCond);
1695 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1696 N0.getValueType().isInteger()) {
1697 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
1698 N0.getOpcode() == ISD::XOR) {
1699 // Simplify (X+Y) == (X+Z) --> Y == Z
1700 if (N0.getOpcode() == N1.getOpcode()) {
1701 if (N0.getOperand(0) == N1.getOperand(0))
1702 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(1), Cond);
1703 if (N0.getOperand(1) == N1.getOperand(1))
1704 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(0), Cond);
1705 if (DAG.isCommutativeBinOp(N0.getOpcode())) {
1706 // If X op Y == Y op X, try other combinations.
1707 if (N0.getOperand(0) == N1.getOperand(1))
1708 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(0), Cond);
1709 if (N0.getOperand(1) == N1.getOperand(0))
1710 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(1), Cond);
1714 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) {
1715 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1716 // Turn (X+C1) == C2 --> X == C2-C1
1717 if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse()) {
1718 return DAG.getSetCC(VT, N0.getOperand(0),
1719 DAG.getConstant(RHSC->getAPIntValue()-
1720 LHSR->getAPIntValue(),
1721 N0.getValueType()), Cond);
1724 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
1725 if (N0.getOpcode() == ISD::XOR)
1726 // If we know that all of the inverted bits are zero, don't bother
1727 // performing the inversion.
1728 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue()))
1730 DAG.getSetCC(VT, N0.getOperand(0),
1731 DAG.getConstant(LHSR->getAPIntValue() ^
1732 RHSC->getAPIntValue(),
1737 // Turn (C1-X) == C2 --> X == C1-C2
1738 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
1739 if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse()) {
1741 DAG.getSetCC(VT, N0.getOperand(1),
1742 DAG.getConstant(SUBC->getAPIntValue() -
1743 RHSC->getAPIntValue(),
1750 // Simplify (X+Z) == X --> Z == 0
1751 if (N0.getOperand(0) == N1)
1752 return DAG.getSetCC(VT, N0.getOperand(1),
1753 DAG.getConstant(0, N0.getValueType()), Cond);
1754 if (N0.getOperand(1) == N1) {
1755 if (DAG.isCommutativeBinOp(N0.getOpcode()))
1756 return DAG.getSetCC(VT, N0.getOperand(0),
1757 DAG.getConstant(0, N0.getValueType()), Cond);
1758 else if (N0.getNode()->hasOneUse()) {
1759 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
1760 // (Z-X) == X --> Z == X<<1
1761 SDValue SH = DAG.getNode(ISD::SHL, N1.getValueType(),
1763 DAG.getConstant(1, getShiftAmountTy()));
1764 if (!DCI.isCalledByLegalizer())
1765 DCI.AddToWorklist(SH.getNode());
1766 return DAG.getSetCC(VT, N0.getOperand(0), SH, Cond);
1771 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
1772 N1.getOpcode() == ISD::XOR) {
1773 // Simplify X == (X+Z) --> Z == 0
1774 if (N1.getOperand(0) == N0) {
1775 return DAG.getSetCC(VT, N1.getOperand(1),
1776 DAG.getConstant(0, N1.getValueType()), Cond);
1777 } else if (N1.getOperand(1) == N0) {
1778 if (DAG.isCommutativeBinOp(N1.getOpcode())) {
1779 return DAG.getSetCC(VT, N1.getOperand(0),
1780 DAG.getConstant(0, N1.getValueType()), Cond);
1781 } else if (N1.getNode()->hasOneUse()) {
1782 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
1783 // X == (Z-X) --> X<<1 == Z
1784 SDValue SH = DAG.getNode(ISD::SHL, N1.getValueType(), N0,
1785 DAG.getConstant(1, getShiftAmountTy()));
1786 if (!DCI.isCalledByLegalizer())
1787 DCI.AddToWorklist(SH.getNode());
1788 return DAG.getSetCC(VT, SH, N1.getOperand(0), Cond);
1794 // Fold away ALL boolean setcc's.
1796 if (N0.getValueType() == MVT::i1 && foldBooleans) {
1798 default: assert(0 && "Unknown integer setcc!");
1799 case ISD::SETEQ: // X == Y -> (X^Y)^1
1800 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
1801 N0 = DAG.getNode(ISD::XOR, MVT::i1, Temp, DAG.getConstant(1, MVT::i1));
1802 if (!DCI.isCalledByLegalizer())
1803 DCI.AddToWorklist(Temp.getNode());
1805 case ISD::SETNE: // X != Y --> (X^Y)
1806 N0 = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
1808 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> X^1 & Y
1809 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> X^1 & Y
1810 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
1811 N0 = DAG.getNode(ISD::AND, MVT::i1, N1, Temp);
1812 if (!DCI.isCalledByLegalizer())
1813 DCI.AddToWorklist(Temp.getNode());
1815 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> Y^1 & X
1816 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> Y^1 & X
1817 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
1818 N0 = DAG.getNode(ISD::AND, MVT::i1, N0, Temp);
1819 if (!DCI.isCalledByLegalizer())
1820 DCI.AddToWorklist(Temp.getNode());
1822 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> X^1 | Y
1823 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> X^1 | Y
1824 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
1825 N0 = DAG.getNode(ISD::OR, MVT::i1, N1, Temp);
1826 if (!DCI.isCalledByLegalizer())
1827 DCI.AddToWorklist(Temp.getNode());
1829 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> Y^1 | X
1830 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> Y^1 | X
1831 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
1832 N0 = DAG.getNode(ISD::OR, MVT::i1, N0, Temp);
1835 if (VT != MVT::i1) {
1836 if (!DCI.isCalledByLegalizer())
1837 DCI.AddToWorklist(N0.getNode());
1838 // FIXME: If running after legalize, we probably can't do this.
1839 N0 = DAG.getNode(ISD::ZERO_EXTEND, VT, N0);
1844 // Could not fold it.
1848 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
1849 /// node is a GlobalAddress + offset.
1850 bool TargetLowering::isGAPlusOffset(SDNode *N, GlobalValue* &GA,
1851 int64_t &Offset) const {
1852 if (isa<GlobalAddressSDNode>(N)) {
1853 GlobalAddressSDNode *GASD = cast<GlobalAddressSDNode>(N);
1854 GA = GASD->getGlobal();
1855 Offset += GASD->getOffset();
1859 if (N->getOpcode() == ISD::ADD) {
1860 SDValue N1 = N->getOperand(0);
1861 SDValue N2 = N->getOperand(1);
1862 if (isGAPlusOffset(N1.getNode(), GA, Offset)) {
1863 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
1865 Offset += V->getSExtValue();
1868 } else if (isGAPlusOffset(N2.getNode(), GA, Offset)) {
1869 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
1871 Offset += V->getSExtValue();
1880 /// isConsecutiveLoad - Return true if LD (which must be a LoadSDNode) is
1881 /// loading 'Bytes' bytes from a location that is 'Dist' units away from the
1882 /// location that the 'Base' load is loading from.
1883 bool TargetLowering::isConsecutiveLoad(SDNode *LD, SDNode *Base,
1884 unsigned Bytes, int Dist,
1885 const MachineFrameInfo *MFI) const {
1886 if (LD->getOperand(0).getNode() != Base->getOperand(0).getNode())
1888 MVT VT = LD->getValueType(0);
1889 if (VT.getSizeInBits() / 8 != Bytes)
1892 SDValue Loc = LD->getOperand(1);
1893 SDValue BaseLoc = Base->getOperand(1);
1894 if (Loc.getOpcode() == ISD::FrameIndex) {
1895 if (BaseLoc.getOpcode() != ISD::FrameIndex)
1897 int FI = cast<FrameIndexSDNode>(Loc)->getIndex();
1898 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
1899 int FS = MFI->getObjectSize(FI);
1900 int BFS = MFI->getObjectSize(BFI);
1901 if (FS != BFS || FS != (int)Bytes) return false;
1902 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Bytes);
1905 GlobalValue *GV1 = NULL;
1906 GlobalValue *GV2 = NULL;
1907 int64_t Offset1 = 0;
1908 int64_t Offset2 = 0;
1909 bool isGA1 = isGAPlusOffset(Loc.getNode(), GV1, Offset1);
1910 bool isGA2 = isGAPlusOffset(BaseLoc.getNode(), GV2, Offset2);
1911 if (isGA1 && isGA2 && GV1 == GV2)
1912 return Offset1 == (Offset2 + Dist*Bytes);
1917 SDValue TargetLowering::
1918 PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
1919 // Default implementation: no optimization.
1923 //===----------------------------------------------------------------------===//
1924 // Inline Assembler Implementation Methods
1925 //===----------------------------------------------------------------------===//
1928 TargetLowering::ConstraintType
1929 TargetLowering::getConstraintType(const std::string &Constraint) const {
1930 // FIXME: lots more standard ones to handle.
1931 if (Constraint.size() == 1) {
1932 switch (Constraint[0]) {
1934 case 'r': return C_RegisterClass;
1936 case 'o': // offsetable
1937 case 'V': // not offsetable
1939 case 'i': // Simple Integer or Relocatable Constant
1940 case 'n': // Simple Integer
1941 case 's': // Relocatable Constant
1942 case 'X': // Allow ANY value.
1943 case 'I': // Target registers.
1955 if (Constraint.size() > 1 && Constraint[0] == '{' &&
1956 Constraint[Constraint.size()-1] == '}')
1961 /// LowerXConstraint - try to replace an X constraint, which matches anything,
1962 /// with another that has more specific requirements based on the type of the
1963 /// corresponding operand.
1964 const char *TargetLowering::LowerXConstraint(MVT ConstraintVT) const{
1965 if (ConstraintVT.isInteger())
1967 if (ConstraintVT.isFloatingPoint())
1968 return "f"; // works for many targets
1972 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
1973 /// vector. If it is invalid, don't add anything to Ops.
1974 void TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
1975 char ConstraintLetter,
1977 std::vector<SDValue> &Ops,
1978 SelectionDAG &DAG) const {
1979 switch (ConstraintLetter) {
1981 case 'X': // Allows any operand; labels (basic block) use this.
1982 if (Op.getOpcode() == ISD::BasicBlock) {
1987 case 'i': // Simple Integer or Relocatable Constant
1988 case 'n': // Simple Integer
1989 case 's': { // Relocatable Constant
1990 // These operands are interested in values of the form (GV+C), where C may
1991 // be folded in as an offset of GV, or it may be explicitly added. Also, it
1992 // is possible and fine if either GV or C are missing.
1993 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
1994 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
1996 // If we have "(add GV, C)", pull out GV/C
1997 if (Op.getOpcode() == ISD::ADD) {
1998 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
1999 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
2000 if (C == 0 || GA == 0) {
2001 C = dyn_cast<ConstantSDNode>(Op.getOperand(0));
2002 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1));
2004 if (C == 0 || GA == 0)
2008 // If we find a valid operand, map to the TargetXXX version so that the
2009 // value itself doesn't get selected.
2010 if (GA) { // Either &GV or &GV+C
2011 if (ConstraintLetter != 'n') {
2012 int64_t Offs = GA->getOffset();
2013 if (C) Offs += C->getZExtValue();
2014 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(),
2015 Op.getValueType(), Offs));
2019 if (C) { // just C, no GV.
2020 // Simple constants are not allowed for 's'.
2021 if (ConstraintLetter != 's') {
2022 Ops.push_back(DAG.getTargetConstant(C->getAPIntValue(),
2023 Op.getValueType()));
2032 std::vector<unsigned> TargetLowering::
2033 getRegClassForInlineAsmConstraint(const std::string &Constraint,
2035 return std::vector<unsigned>();
2039 std::pair<unsigned, const TargetRegisterClass*> TargetLowering::
2040 getRegForInlineAsmConstraint(const std::string &Constraint,
2042 if (Constraint[0] != '{')
2043 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
2044 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?");
2046 // Remove the braces from around the name.
2047 std::string RegName(Constraint.begin()+1, Constraint.end()-1);
2049 // Figure out which register class contains this reg.
2050 const TargetRegisterInfo *RI = TM.getRegisterInfo();
2051 for (TargetRegisterInfo::regclass_iterator RCI = RI->regclass_begin(),
2052 E = RI->regclass_end(); RCI != E; ++RCI) {
2053 const TargetRegisterClass *RC = *RCI;
2055 // If none of the the value types for this register class are valid, we
2056 // can't use it. For example, 64-bit reg classes on 32-bit targets.
2057 bool isLegal = false;
2058 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
2060 if (isTypeLegal(*I)) {
2066 if (!isLegal) continue;
2068 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end();
2070 if (StringsEqualNoCase(RegName, RI->get(*I).AsmName))
2071 return std::make_pair(*I, RC);
2075 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
2078 //===----------------------------------------------------------------------===//
2079 // Constraint Selection.
2081 /// isMatchingInputConstraint - Return true of this is an input operand that is
2082 /// a matching constraint like "4".
2083 bool TargetLowering::AsmOperandInfo::isMatchingInputConstraint() const {
2084 assert(!ConstraintCode.empty() && "No known constraint!");
2085 return isdigit(ConstraintCode[0]);
2088 /// getMatchedOperand - If this is an input matching constraint, this method
2089 /// returns the output operand it matches.
2090 unsigned TargetLowering::AsmOperandInfo::getMatchedOperand() const {
2091 assert(!ConstraintCode.empty() && "No known constraint!");
2092 return atoi(ConstraintCode.c_str());
2096 /// getConstraintGenerality - Return an integer indicating how general CT
2098 static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) {
2100 default: assert(0 && "Unknown constraint type!");
2101 case TargetLowering::C_Other:
2102 case TargetLowering::C_Unknown:
2104 case TargetLowering::C_Register:
2106 case TargetLowering::C_RegisterClass:
2108 case TargetLowering::C_Memory:
2113 /// ChooseConstraint - If there are multiple different constraints that we
2114 /// could pick for this operand (e.g. "imr") try to pick the 'best' one.
2115 /// This is somewhat tricky: constraints fall into four classes:
2116 /// Other -> immediates and magic values
2117 /// Register -> one specific register
2118 /// RegisterClass -> a group of regs
2119 /// Memory -> memory
2120 /// Ideally, we would pick the most specific constraint possible: if we have
2121 /// something that fits into a register, we would pick it. The problem here
2122 /// is that if we have something that could either be in a register or in
2123 /// memory that use of the register could cause selection of *other*
2124 /// operands to fail: they might only succeed if we pick memory. Because of
2125 /// this the heuristic we use is:
2127 /// 1) If there is an 'other' constraint, and if the operand is valid for
2128 /// that constraint, use it. This makes us take advantage of 'i'
2129 /// constraints when available.
2130 /// 2) Otherwise, pick the most general constraint present. This prefers
2131 /// 'm' over 'r', for example.
2133 static void ChooseConstraint(TargetLowering::AsmOperandInfo &OpInfo,
2134 bool hasMemory, const TargetLowering &TLI,
2135 SDValue Op, SelectionDAG *DAG) {
2136 assert(OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options");
2137 unsigned BestIdx = 0;
2138 TargetLowering::ConstraintType BestType = TargetLowering::C_Unknown;
2139 int BestGenerality = -1;
2141 // Loop over the options, keeping track of the most general one.
2142 for (unsigned i = 0, e = OpInfo.Codes.size(); i != e; ++i) {
2143 TargetLowering::ConstraintType CType =
2144 TLI.getConstraintType(OpInfo.Codes[i]);
2146 // If this is an 'other' constraint, see if the operand is valid for it.
2147 // For example, on X86 we might have an 'rI' constraint. If the operand
2148 // is an integer in the range [0..31] we want to use I (saving a load
2149 // of a register), otherwise we must use 'r'.
2150 if (CType == TargetLowering::C_Other && Op.getNode()) {
2151 assert(OpInfo.Codes[i].size() == 1 &&
2152 "Unhandled multi-letter 'other' constraint");
2153 std::vector<SDValue> ResultOps;
2154 TLI.LowerAsmOperandForConstraint(Op, OpInfo.Codes[i][0], hasMemory,
2156 if (!ResultOps.empty()) {
2163 // This constraint letter is more general than the previous one, use it.
2164 int Generality = getConstraintGenerality(CType);
2165 if (Generality > BestGenerality) {
2168 BestGenerality = Generality;
2172 OpInfo.ConstraintCode = OpInfo.Codes[BestIdx];
2173 OpInfo.ConstraintType = BestType;
2176 /// ComputeConstraintToUse - Determines the constraint code and constraint
2177 /// type to use for the specific AsmOperandInfo, setting
2178 /// OpInfo.ConstraintCode and OpInfo.ConstraintType.
2179 void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo,
2182 SelectionDAG *DAG) const {
2183 assert(!OpInfo.Codes.empty() && "Must have at least one constraint");
2185 // Single-letter constraints ('r') are very common.
2186 if (OpInfo.Codes.size() == 1) {
2187 OpInfo.ConstraintCode = OpInfo.Codes[0];
2188 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
2190 ChooseConstraint(OpInfo, hasMemory, *this, Op, DAG);
2193 // 'X' matches anything.
2194 if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) {
2195 // Labels and constants are handled elsewhere ('X' is the only thing
2196 // that matches labels).
2197 if (isa<BasicBlock>(OpInfo.CallOperandVal) ||
2198 isa<ConstantInt>(OpInfo.CallOperandVal))
2201 // Otherwise, try to resolve it to something we know about by looking at
2202 // the actual operand type.
2203 if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) {
2204 OpInfo.ConstraintCode = Repl;
2205 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
2210 //===----------------------------------------------------------------------===//
2211 // Loop Strength Reduction hooks
2212 //===----------------------------------------------------------------------===//
2214 /// isLegalAddressingMode - Return true if the addressing mode represented
2215 /// by AM is legal for this target, for a load/store of the specified type.
2216 bool TargetLowering::isLegalAddressingMode(const AddrMode &AM,
2217 const Type *Ty) const {
2218 // The default implementation of this implements a conservative RISCy, r+r and
2221 // Allows a sign-extended 16-bit immediate field.
2222 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
2225 // No global is ever allowed as a base.
2229 // Only support r+r,
2231 case 0: // "r+i" or just "i", depending on HasBaseReg.
2234 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
2236 // Otherwise we have r+r or r+i.
2239 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
2241 // Allow 2*r as r+r.
2249 APInt m; // magic number
2250 bool a; // add indicator
2251 unsigned s; // shift amount
2254 /// magicu - calculate the magic numbers required to codegen an integer udiv as
2255 /// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
2256 static mu magicu(const APInt& d) {
2258 APInt nc, delta, q1, r1, q2, r2;
2260 magu.a = 0; // initialize "add" indicator
2261 APInt allOnes = APInt::getAllOnesValue(d.getBitWidth());
2262 APInt signedMin = APInt::getSignedMinValue(d.getBitWidth());
2263 APInt signedMax = APInt::getSignedMaxValue(d.getBitWidth());
2265 nc = allOnes - (-d).urem(d);
2266 p = d.getBitWidth() - 1; // initialize p
2267 q1 = signedMin.udiv(nc); // initialize q1 = 2p/nc
2268 r1 = signedMin - q1*nc; // initialize r1 = rem(2p,nc)
2269 q2 = signedMax.udiv(d); // initialize q2 = (2p-1)/d
2270 r2 = signedMax - q2*d; // initialize r2 = rem((2p-1),d)
2273 if (r1.uge(nc - r1)) {
2274 q1 = q1 + q1 + 1; // update q1
2275 r1 = r1 + r1 - nc; // update r1
2278 q1 = q1+q1; // update q1
2279 r1 = r1+r1; // update r1
2281 if ((r2 + 1).uge(d - r2)) {
2282 if (q2.uge(signedMax)) magu.a = 1;
2283 q2 = q2+q2 + 1; // update q2
2284 r2 = r2+r2 + 1 - d; // update r2
2287 if (q2.uge(signedMin)) magu.a = 1;
2288 q2 = q2+q2; // update q2
2289 r2 = r2+r2 + 1; // update r2
2292 } while (p < d.getBitWidth()*2 &&
2293 (q1.ult(delta) || (q1 == delta && r1 == 0)));
2294 magu.m = q2 + 1; // resulting magic number
2295 magu.s = p - d.getBitWidth(); // resulting shift
2299 // Magic for divide replacement
2301 APInt m; // magic number
2302 unsigned s; // shift amount
2305 /// magic - calculate the magic numbers required to codegen an integer sdiv as
2306 /// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
2308 static ms magic(const APInt& d) {
2310 APInt ad, anc, delta, q1, r1, q2, r2, t;
2311 APInt allOnes = APInt::getAllOnesValue(d.getBitWidth());
2312 APInt signedMin = APInt::getSignedMinValue(d.getBitWidth());
2313 APInt signedMax = APInt::getSignedMaxValue(d.getBitWidth());
2317 t = signedMin + (d.lshr(d.getBitWidth() - 1));
2318 anc = t - 1 - t.urem(ad); // absolute value of nc
2319 p = d.getBitWidth() - 1; // initialize p
2320 q1 = signedMin.udiv(anc); // initialize q1 = 2p/abs(nc)
2321 r1 = signedMin - q1*anc; // initialize r1 = rem(2p,abs(nc))
2322 q2 = signedMin.udiv(ad); // initialize q2 = 2p/abs(d)
2323 r2 = signedMin - q2*ad; // initialize r2 = rem(2p,abs(d))
2326 q1 = q1<<1; // update q1 = 2p/abs(nc)
2327 r1 = r1<<1; // update r1 = rem(2p/abs(nc))
2328 if (r1.uge(anc)) { // must be unsigned comparison
2332 q2 = q2<<1; // update q2 = 2p/abs(d)
2333 r2 = r2<<1; // update r2 = rem(2p/abs(d))
2334 if (r2.uge(ad)) { // must be unsigned comparison
2339 } while (q1.ule(delta) || (q1 == delta && r1 == 0));
2342 if (d.isNegative()) mag.m = -mag.m; // resulting magic number
2343 mag.s = p - d.getBitWidth(); // resulting shift
2347 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
2348 /// return a DAG expression to select that will generate the same value by
2349 /// multiplying by a magic number. See:
2350 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
2351 SDValue TargetLowering::BuildSDIV(SDNode *N, SelectionDAG &DAG,
2352 std::vector<SDNode*>* Created) const {
2353 MVT VT = N->getValueType(0);
2355 // Check to see if we can do this.
2356 // FIXME: We should be more aggressive here.
2357 if (!isTypeLegal(VT))
2360 APInt d = cast<ConstantSDNode>(N->getOperand(1))->getAPIntValue();
2361 ms magics = magic(d);
2363 // Multiply the numerator (operand 0) by the magic value
2364 // FIXME: We should support doing a MUL in a wider type
2366 if (isOperationLegal(ISD::MULHS, VT))
2367 Q = DAG.getNode(ISD::MULHS, VT, N->getOperand(0),
2368 DAG.getConstant(magics.m, VT));
2369 else if (isOperationLegal(ISD::SMUL_LOHI, VT))
2370 Q = SDValue(DAG.getNode(ISD::SMUL_LOHI, DAG.getVTList(VT, VT),
2372 DAG.getConstant(magics.m, VT)).getNode(), 1);
2374 return SDValue(); // No mulhs or equvialent
2375 // If d > 0 and m < 0, add the numerator
2376 if (d.isStrictlyPositive() && magics.m.isNegative()) {
2377 Q = DAG.getNode(ISD::ADD, VT, Q, N->getOperand(0));
2379 Created->push_back(Q.getNode());
2381 // If d < 0 and m > 0, subtract the numerator.
2382 if (d.isNegative() && magics.m.isStrictlyPositive()) {
2383 Q = DAG.getNode(ISD::SUB, VT, Q, N->getOperand(0));
2385 Created->push_back(Q.getNode());
2387 // Shift right algebraic if shift value is nonzero
2389 Q = DAG.getNode(ISD::SRA, VT, Q,
2390 DAG.getConstant(magics.s, getShiftAmountTy()));
2392 Created->push_back(Q.getNode());
2394 // Extract the sign bit and add it to the quotient
2396 DAG.getNode(ISD::SRL, VT, Q, DAG.getConstant(VT.getSizeInBits()-1,
2397 getShiftAmountTy()));
2399 Created->push_back(T.getNode());
2400 return DAG.getNode(ISD::ADD, VT, Q, T);
2403 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
2404 /// return a DAG expression to select that will generate the same value by
2405 /// multiplying by a magic number. See:
2406 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
2407 SDValue TargetLowering::BuildUDIV(SDNode *N, SelectionDAG &DAG,
2408 std::vector<SDNode*>* Created) const {
2409 MVT VT = N->getValueType(0);
2411 // Check to see if we can do this.
2412 // FIXME: We should be more aggressive here.
2413 if (!isTypeLegal(VT))
2416 // FIXME: We should use a narrower constant when the upper
2417 // bits are known to be zero.
2418 ConstantSDNode *N1C = cast<ConstantSDNode>(N->getOperand(1));
2419 mu magics = magicu(N1C->getAPIntValue());
2421 // Multiply the numerator (operand 0) by the magic value
2422 // FIXME: We should support doing a MUL in a wider type
2424 if (isOperationLegal(ISD::MULHU, VT))
2425 Q = DAG.getNode(ISD::MULHU, VT, N->getOperand(0),
2426 DAG.getConstant(magics.m, VT));
2427 else if (isOperationLegal(ISD::UMUL_LOHI, VT))
2428 Q = SDValue(DAG.getNode(ISD::UMUL_LOHI, DAG.getVTList(VT, VT),
2430 DAG.getConstant(magics.m, VT)).getNode(), 1);
2432 return SDValue(); // No mulhu or equvialent
2434 Created->push_back(Q.getNode());
2436 if (magics.a == 0) {
2437 assert(magics.s < N1C->getAPIntValue().getBitWidth() &&
2438 "We shouldn't generate an undefined shift!");
2439 return DAG.getNode(ISD::SRL, VT, Q,
2440 DAG.getConstant(magics.s, getShiftAmountTy()));
2442 SDValue NPQ = DAG.getNode(ISD::SUB, VT, N->getOperand(0), Q);
2444 Created->push_back(NPQ.getNode());
2445 NPQ = DAG.getNode(ISD::SRL, VT, NPQ,
2446 DAG.getConstant(1, getShiftAmountTy()));
2448 Created->push_back(NPQ.getNode());
2449 NPQ = DAG.getNode(ISD::ADD, VT, NPQ, Q);
2451 Created->push_back(NPQ.getNode());
2452 return DAG.getNode(ISD::SRL, VT, NPQ,
2453 DAG.getConstant(magics.s-1, getShiftAmountTy()));