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 SetCCResultContents = UndefinedSetCCResult;
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;
575 TransformToType[i] = MVT::Other; // this isn't actually used
576 ValueTypeActions.setTypeAction(VT, Expand);
581 const char *TargetLowering::getTargetNodeName(unsigned Opcode) const {
586 MVT TargetLowering::getSetCCResultType(const SDValue &) const {
587 return getValueType(TD->getIntPtrType());
591 /// getVectorTypeBreakdown - Vector types are broken down into some number of
592 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
593 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
594 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
596 /// This method returns the number of registers needed, and the VT for each
597 /// register. It also returns the VT and quantity of the intermediate values
598 /// before they are promoted/expanded.
600 unsigned TargetLowering::getVectorTypeBreakdown(MVT VT,
602 unsigned &NumIntermediates,
603 MVT &RegisterVT) const {
604 // Figure out the right, legal destination reg to copy into.
605 unsigned NumElts = VT.getVectorNumElements();
606 MVT EltTy = VT.getVectorElementType();
608 unsigned NumVectorRegs = 1;
610 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
611 // could break down into LHS/RHS like LegalizeDAG does.
612 if (!isPowerOf2_32(NumElts)) {
613 NumVectorRegs = NumElts;
617 // Divide the input until we get to a supported size. This will always
618 // end with a scalar if the target doesn't support vectors.
619 while (NumElts > 1 && !isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) {
624 NumIntermediates = NumVectorRegs;
626 MVT NewVT = MVT::getVectorVT(EltTy, NumElts);
627 if (!isTypeLegal(NewVT))
629 IntermediateVT = NewVT;
631 MVT DestVT = getTypeToTransformTo(NewVT);
633 if (DestVT.bitsLT(NewVT)) {
634 // Value is expanded, e.g. i64 -> i16.
635 return NumVectorRegs*(NewVT.getSizeInBits()/DestVT.getSizeInBits());
637 // Otherwise, promotion or legal types use the same number of registers as
638 // the vector decimated to the appropriate level.
639 return NumVectorRegs;
645 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
646 /// function arguments in the caller parameter area. This is the actual
647 /// alignment, not its logarithm.
648 unsigned TargetLowering::getByValTypeAlignment(const Type *Ty) const {
649 return TD->getCallFrameTypeAlignment(Ty);
652 SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table,
653 SelectionDAG &DAG) const {
654 if (usesGlobalOffsetTable())
655 return DAG.getNode(ISD::GLOBAL_OFFSET_TABLE, getPointerTy());
659 //===----------------------------------------------------------------------===//
660 // Optimization Methods
661 //===----------------------------------------------------------------------===//
663 /// ShrinkDemandedConstant - Check to see if the specified operand of the
664 /// specified instruction is a constant integer. If so, check to see if there
665 /// are any bits set in the constant that are not demanded. If so, shrink the
666 /// constant and return true.
667 bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDValue Op,
668 const APInt &Demanded) {
669 // FIXME: ISD::SELECT, ISD::SELECT_CC
670 switch(Op.getOpcode()) {
675 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1)))
676 if (C->getAPIntValue().intersects(~Demanded)) {
677 MVT VT = Op.getValueType();
678 SDValue New = DAG.getNode(Op.getOpcode(), VT, Op.getOperand(0),
679 DAG.getConstant(Demanded &
682 return CombineTo(Op, New);
689 /// SimplifyDemandedBits - Look at Op. At this point, we know that only the
690 /// DemandedMask bits of the result of Op are ever used downstream. If we can
691 /// use this information to simplify Op, create a new simplified DAG node and
692 /// return true, returning the original and new nodes in Old and New. Otherwise,
693 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
694 /// the expression (used to simplify the caller). The KnownZero/One bits may
695 /// only be accurate for those bits in the DemandedMask.
696 bool TargetLowering::SimplifyDemandedBits(SDValue Op,
697 const APInt &DemandedMask,
700 TargetLoweringOpt &TLO,
701 unsigned Depth) const {
702 unsigned BitWidth = DemandedMask.getBitWidth();
703 assert(Op.getValueSizeInBits() == BitWidth &&
704 "Mask size mismatches value type size!");
705 APInt NewMask = DemandedMask;
707 // Don't know anything.
708 KnownZero = KnownOne = APInt(BitWidth, 0);
710 // Other users may use these bits.
711 if (!Op.getNode()->hasOneUse()) {
713 // If not at the root, Just compute the KnownZero/KnownOne bits to
714 // simplify things downstream.
715 TLO.DAG.ComputeMaskedBits(Op, DemandedMask, KnownZero, KnownOne, Depth);
718 // If this is the root being simplified, allow it to have multiple uses,
719 // just set the NewMask to all bits.
720 NewMask = APInt::getAllOnesValue(BitWidth);
721 } else if (DemandedMask == 0) {
722 // Not demanding any bits from Op.
723 if (Op.getOpcode() != ISD::UNDEF)
724 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::UNDEF, Op.getValueType()));
726 } else if (Depth == 6) { // Limit search depth.
730 APInt KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut;
731 switch (Op.getOpcode()) {
733 // We know all of the bits for a constant!
734 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue() & NewMask;
735 KnownZero = ~KnownOne & NewMask;
736 return false; // Don't fall through, will infinitely loop.
738 // If the RHS is a constant, check to see if the LHS would be zero without
739 // using the bits from the RHS. Below, we use knowledge about the RHS to
740 // simplify the LHS, here we're using information from the LHS to simplify
742 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
743 APInt LHSZero, LHSOne;
744 TLO.DAG.ComputeMaskedBits(Op.getOperand(0), NewMask,
745 LHSZero, LHSOne, Depth+1);
746 // If the LHS already has zeros where RHSC does, this and is dead.
747 if ((LHSZero & NewMask) == (~RHSC->getAPIntValue() & NewMask))
748 return TLO.CombineTo(Op, Op.getOperand(0));
749 // If any of the set bits in the RHS are known zero on the LHS, shrink
751 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & NewMask))
755 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
756 KnownOne, TLO, Depth+1))
758 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
759 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownZero & NewMask,
760 KnownZero2, KnownOne2, TLO, Depth+1))
762 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
764 // If all of the demanded bits are known one on one side, return the other.
765 // These bits cannot contribute to the result of the 'and'.
766 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
767 return TLO.CombineTo(Op, Op.getOperand(0));
768 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
769 return TLO.CombineTo(Op, Op.getOperand(1));
770 // If all of the demanded bits in the inputs are known zeros, return zero.
771 if ((NewMask & (KnownZero|KnownZero2)) == NewMask)
772 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, Op.getValueType()));
773 // If the RHS is a constant, see if we can simplify it.
774 if (TLO.ShrinkDemandedConstant(Op, ~KnownZero2 & NewMask))
777 // Output known-1 bits are only known if set in both the LHS & RHS.
778 KnownOne &= KnownOne2;
779 // Output known-0 are known to be clear if zero in either the LHS | RHS.
780 KnownZero |= KnownZero2;
783 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
784 KnownOne, TLO, Depth+1))
786 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
787 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownOne & NewMask,
788 KnownZero2, KnownOne2, TLO, Depth+1))
790 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
792 // If all of the demanded bits are known zero on one side, return the other.
793 // These bits cannot contribute to the result of the 'or'.
794 if ((NewMask & ~KnownOne2 & KnownZero) == (~KnownOne2 & NewMask))
795 return TLO.CombineTo(Op, Op.getOperand(0));
796 if ((NewMask & ~KnownOne & KnownZero2) == (~KnownOne & NewMask))
797 return TLO.CombineTo(Op, Op.getOperand(1));
798 // If all of the potentially set bits on one side are known to be set on
799 // the other side, just use the 'other' side.
800 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
801 return TLO.CombineTo(Op, Op.getOperand(0));
802 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
803 return TLO.CombineTo(Op, Op.getOperand(1));
804 // If the RHS is a constant, see if we can simplify it.
805 if (TLO.ShrinkDemandedConstant(Op, NewMask))
808 // Output known-0 bits are only known if clear in both the LHS & RHS.
809 KnownZero &= KnownZero2;
810 // Output known-1 are known to be set if set in either the LHS | RHS.
811 KnownOne |= KnownOne2;
814 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
815 KnownOne, TLO, Depth+1))
817 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
818 if (SimplifyDemandedBits(Op.getOperand(0), NewMask, KnownZero2,
819 KnownOne2, TLO, Depth+1))
821 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
823 // If all of the demanded bits are known zero on one side, return the other.
824 // These bits cannot contribute to the result of the 'xor'.
825 if ((KnownZero & NewMask) == NewMask)
826 return TLO.CombineTo(Op, Op.getOperand(0));
827 if ((KnownZero2 & NewMask) == NewMask)
828 return TLO.CombineTo(Op, Op.getOperand(1));
830 // If all of the unknown bits are known to be zero on one side or the other
831 // (but not both) turn this into an *inclusive* or.
832 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
833 if ((NewMask & ~KnownZero & ~KnownZero2) == 0)
834 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, Op.getValueType(),
838 // Output known-0 bits are known if clear or set in both the LHS & RHS.
839 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
840 // Output known-1 are known to be set if set in only one of the LHS, RHS.
841 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
843 // If all of the demanded bits on one side are known, and all of the set
844 // bits on that side are also known to be set on the other side, turn this
845 // into an AND, as we know the bits will be cleared.
846 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
847 if ((NewMask & (KnownZero|KnownOne)) == NewMask) { // all known
848 if ((KnownOne & KnownOne2) == KnownOne) {
849 MVT VT = Op.getValueType();
850 SDValue ANDC = TLO.DAG.getConstant(~KnownOne & NewMask, VT);
851 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, VT, Op.getOperand(0),
856 // If the RHS is a constant, see if we can simplify it.
857 // for XOR, we prefer to force bits to 1 if they will make a -1.
858 // if we can't force bits, try to shrink constant
859 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
860 APInt Expanded = C->getAPIntValue() | (~NewMask);
861 // if we can expand it to have all bits set, do it
862 if (Expanded.isAllOnesValue()) {
863 if (Expanded != C->getAPIntValue()) {
864 MVT VT = Op.getValueType();
865 SDValue New = TLO.DAG.getNode(Op.getOpcode(), VT, Op.getOperand(0),
866 TLO.DAG.getConstant(Expanded, VT));
867 return TLO.CombineTo(Op, New);
869 // if it already has all the bits set, nothing to change
870 // but don't shrink either!
871 } else if (TLO.ShrinkDemandedConstant(Op, NewMask)) {
876 KnownZero = KnownZeroOut;
877 KnownOne = KnownOneOut;
880 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero,
881 KnownOne, TLO, Depth+1))
883 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero2,
884 KnownOne2, TLO, Depth+1))
886 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
887 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
889 // If the operands are constants, see if we can simplify them.
890 if (TLO.ShrinkDemandedConstant(Op, NewMask))
893 // Only known if known in both the LHS and RHS.
894 KnownOne &= KnownOne2;
895 KnownZero &= KnownZero2;
898 if (SimplifyDemandedBits(Op.getOperand(3), NewMask, KnownZero,
899 KnownOne, TLO, Depth+1))
901 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero2,
902 KnownOne2, TLO, Depth+1))
904 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
905 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
907 // If the operands are constants, see if we can simplify them.
908 if (TLO.ShrinkDemandedConstant(Op, NewMask))
911 // Only known if known in both the LHS and RHS.
912 KnownOne &= KnownOne2;
913 KnownZero &= KnownZero2;
916 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
917 unsigned ShAmt = SA->getZExtValue();
918 SDValue InOp = Op.getOperand(0);
920 // If the shift count is an invalid immediate, don't do anything.
921 if (ShAmt >= BitWidth)
924 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a
925 // single shift. We can do this if the bottom bits (which are shifted
926 // out) are never demanded.
927 if (InOp.getOpcode() == ISD::SRL &&
928 isa<ConstantSDNode>(InOp.getOperand(1))) {
929 if (ShAmt && (NewMask & APInt::getLowBitsSet(BitWidth, ShAmt)) == 0) {
930 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
931 unsigned Opc = ISD::SHL;
939 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
940 MVT VT = Op.getValueType();
941 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, VT,
942 InOp.getOperand(0), NewSA));
946 if (SimplifyDemandedBits(Op.getOperand(0), NewMask.lshr(ShAmt),
947 KnownZero, KnownOne, TLO, Depth+1))
949 KnownZero <<= SA->getZExtValue();
950 KnownOne <<= SA->getZExtValue();
951 // low bits known zero.
952 KnownZero |= APInt::getLowBitsSet(BitWidth, SA->getZExtValue());
956 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
957 MVT VT = Op.getValueType();
958 unsigned ShAmt = SA->getZExtValue();
959 unsigned VTSize = VT.getSizeInBits();
960 SDValue InOp = Op.getOperand(0);
962 // If the shift count is an invalid immediate, don't do anything.
963 if (ShAmt >= BitWidth)
966 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a
967 // single shift. We can do this if the top bits (which are shifted out)
968 // are never demanded.
969 if (InOp.getOpcode() == ISD::SHL &&
970 isa<ConstantSDNode>(InOp.getOperand(1))) {
971 if (ShAmt && (NewMask & APInt::getHighBitsSet(VTSize, ShAmt)) == 0) {
972 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
973 unsigned Opc = ISD::SRL;
981 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
982 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, VT,
983 InOp.getOperand(0), NewSA));
987 // Compute the new bits that are at the top now.
988 if (SimplifyDemandedBits(InOp, (NewMask << ShAmt),
989 KnownZero, KnownOne, TLO, Depth+1))
991 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
992 KnownZero = KnownZero.lshr(ShAmt);
993 KnownOne = KnownOne.lshr(ShAmt);
995 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
996 KnownZero |= HighBits; // High bits known zero.
1000 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1001 MVT VT = Op.getValueType();
1002 unsigned ShAmt = SA->getZExtValue();
1004 // If the shift count is an invalid immediate, don't do anything.
1005 if (ShAmt >= BitWidth)
1008 APInt InDemandedMask = (NewMask << ShAmt);
1010 // If any of the demanded bits are produced by the sign extension, we also
1011 // demand the input sign bit.
1012 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
1013 if (HighBits.intersects(NewMask))
1014 InDemandedMask |= APInt::getSignBit(VT.getSizeInBits());
1016 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask,
1017 KnownZero, KnownOne, TLO, Depth+1))
1019 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1020 KnownZero = KnownZero.lshr(ShAmt);
1021 KnownOne = KnownOne.lshr(ShAmt);
1023 // Handle the sign bit, adjusted to where it is now in the mask.
1024 APInt SignBit = APInt::getSignBit(BitWidth).lshr(ShAmt);
1026 // If the input sign bit is known to be zero, or if none of the top bits
1027 // are demanded, turn this into an unsigned shift right.
1028 if (KnownZero.intersects(SignBit) || (HighBits & ~NewMask) == HighBits) {
1029 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, VT, Op.getOperand(0),
1031 } else if (KnownOne.intersects(SignBit)) { // New bits are known one.
1032 KnownOne |= HighBits;
1036 case ISD::SIGN_EXTEND_INREG: {
1037 MVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1039 // Sign extension. Compute the demanded bits in the result that are not
1040 // present in the input.
1041 APInt NewBits = APInt::getHighBitsSet(BitWidth,
1042 BitWidth - EVT.getSizeInBits()) &
1045 // If none of the extended bits are demanded, eliminate the sextinreg.
1047 return TLO.CombineTo(Op, Op.getOperand(0));
1049 APInt InSignBit = APInt::getSignBit(EVT.getSizeInBits());
1050 InSignBit.zext(BitWidth);
1051 APInt InputDemandedBits = APInt::getLowBitsSet(BitWidth,
1052 EVT.getSizeInBits()) &
1055 // Since the sign extended bits are demanded, we know that the sign
1057 InputDemandedBits |= InSignBit;
1059 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits,
1060 KnownZero, KnownOne, TLO, Depth+1))
1062 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1064 // If the sign bit of the input is known set or clear, then we know the
1065 // top bits of the result.
1067 // If the input sign bit is known zero, convert this into a zero extension.
1068 if (KnownZero.intersects(InSignBit))
1069 return TLO.CombineTo(Op,
1070 TLO.DAG.getZeroExtendInReg(Op.getOperand(0), EVT));
1072 if (KnownOne.intersects(InSignBit)) { // Input sign bit known set
1073 KnownOne |= NewBits;
1074 KnownZero &= ~NewBits;
1075 } else { // Input sign bit unknown
1076 KnownZero &= ~NewBits;
1077 KnownOne &= ~NewBits;
1081 case ISD::ZERO_EXTEND: {
1082 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits();
1083 APInt InMask = NewMask;
1084 InMask.trunc(OperandBitWidth);
1086 // If none of the top bits are demanded, convert this into an any_extend.
1088 APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask;
1089 if (!NewBits.intersects(NewMask))
1090 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND,
1094 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1095 KnownZero, KnownOne, TLO, Depth+1))
1097 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1098 KnownZero.zext(BitWidth);
1099 KnownOne.zext(BitWidth);
1100 KnownZero |= NewBits;
1103 case ISD::SIGN_EXTEND: {
1104 MVT InVT = Op.getOperand(0).getValueType();
1105 unsigned InBits = InVT.getSizeInBits();
1106 APInt InMask = APInt::getLowBitsSet(BitWidth, InBits);
1107 APInt InSignBit = APInt::getBitsSet(BitWidth, InBits - 1, InBits);
1108 APInt NewBits = ~InMask & NewMask;
1110 // If none of the top bits are demanded, convert this into an any_extend.
1112 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND,Op.getValueType(),
1115 // Since some of the sign extended bits are demanded, we know that the sign
1117 APInt InDemandedBits = InMask & NewMask;
1118 InDemandedBits |= InSignBit;
1119 InDemandedBits.trunc(InBits);
1121 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero,
1122 KnownOne, TLO, Depth+1))
1124 KnownZero.zext(BitWidth);
1125 KnownOne.zext(BitWidth);
1127 // If the sign bit is known zero, convert this to a zero extend.
1128 if (KnownZero.intersects(InSignBit))
1129 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND,
1133 // If the sign bit is known one, the top bits match.
1134 if (KnownOne.intersects(InSignBit)) {
1135 KnownOne |= NewBits;
1136 KnownZero &= ~NewBits;
1137 } else { // Otherwise, top bits aren't known.
1138 KnownOne &= ~NewBits;
1139 KnownZero &= ~NewBits;
1143 case ISD::ANY_EXTEND: {
1144 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits();
1145 APInt InMask = NewMask;
1146 InMask.trunc(OperandBitWidth);
1147 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1148 KnownZero, KnownOne, TLO, Depth+1))
1150 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1151 KnownZero.zext(BitWidth);
1152 KnownOne.zext(BitWidth);
1155 case ISD::TRUNCATE: {
1156 // Simplify the input, using demanded bit information, and compute the known
1157 // zero/one bits live out.
1158 APInt TruncMask = NewMask;
1159 TruncMask.zext(Op.getOperand(0).getValueSizeInBits());
1160 if (SimplifyDemandedBits(Op.getOperand(0), TruncMask,
1161 KnownZero, KnownOne, TLO, Depth+1))
1163 KnownZero.trunc(BitWidth);
1164 KnownOne.trunc(BitWidth);
1166 // If the input is only used by this truncate, see if we can shrink it based
1167 // on the known demanded bits.
1168 if (Op.getOperand(0).getNode()->hasOneUse()) {
1169 SDValue In = Op.getOperand(0);
1170 unsigned InBitWidth = In.getValueSizeInBits();
1171 switch (In.getOpcode()) {
1174 // Shrink SRL by a constant if none of the high bits shifted in are
1176 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1))){
1177 APInt HighBits = APInt::getHighBitsSet(InBitWidth,
1178 InBitWidth - BitWidth);
1179 HighBits = HighBits.lshr(ShAmt->getZExtValue());
1180 HighBits.trunc(BitWidth);
1182 if (ShAmt->getZExtValue() < BitWidth && !(HighBits & NewMask)) {
1183 // None of the shifted in bits are needed. Add a truncate of the
1184 // shift input, then shift it.
1185 SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE,
1188 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL,Op.getValueType(),
1189 NewTrunc, In.getOperand(1)));
1196 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1199 case ISD::AssertZext: {
1200 MVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1201 APInt InMask = APInt::getLowBitsSet(BitWidth,
1202 VT.getSizeInBits());
1203 if (SimplifyDemandedBits(Op.getOperand(0), InMask & NewMask,
1204 KnownZero, KnownOne, TLO, Depth+1))
1206 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1207 KnownZero |= ~InMask & NewMask;
1210 case ISD::BIT_CONVERT:
1212 // If this is an FP->Int bitcast and if the sign bit is the only thing that
1213 // is demanded, turn this into a FGETSIGN.
1214 if (NewMask == MVT::getIntegerVTSignBit(Op.getValueType()) &&
1215 MVT::isFloatingPoint(Op.getOperand(0).getValueType()) &&
1216 !MVT::isVector(Op.getOperand(0).getValueType())) {
1217 // Only do this xform if FGETSIGN is valid or if before legalize.
1218 if (!TLO.AfterLegalize ||
1219 isOperationLegal(ISD::FGETSIGN, Op.getValueType())) {
1220 // Make a FGETSIGN + SHL to move the sign bit into the appropriate
1221 // place. We expect the SHL to be eliminated by other optimizations.
1222 SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, Op.getValueType(),
1224 unsigned ShVal = Op.getValueType().getSizeInBits()-1;
1225 SDValue ShAmt = TLO.DAG.getConstant(ShVal, getShiftAmountTy());
1226 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, Op.getValueType(),
1233 // Just use ComputeMaskedBits to compute output bits.
1234 TLO.DAG.ComputeMaskedBits(Op, NewMask, KnownZero, KnownOne, Depth);
1238 // If we know the value of all of the demanded bits, return this as a
1240 if ((NewMask & (KnownZero|KnownOne)) == NewMask)
1241 return TLO.CombineTo(Op, TLO.DAG.getConstant(KnownOne, Op.getValueType()));
1246 /// computeMaskedBitsForTargetNode - Determine which of the bits specified
1247 /// in Mask are known to be either zero or one and return them in the
1248 /// KnownZero/KnownOne bitsets.
1249 void TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
1253 const SelectionDAG &DAG,
1254 unsigned Depth) const {
1255 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1256 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1257 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1258 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1259 "Should use MaskedValueIsZero if you don't know whether Op"
1260 " is a target node!");
1261 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
1264 /// ComputeNumSignBitsForTargetNode - This method can be implemented by
1265 /// targets that want to expose additional information about sign bits to the
1267 unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
1268 unsigned Depth) const {
1269 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1270 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1271 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1272 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1273 "Should use ComputeNumSignBits if you don't know whether Op"
1274 " is a target node!");
1279 /// SimplifySetCC - Try to simplify a setcc built with the specified operands
1280 /// and cc. If it is unable to simplify it, return a null SDValue.
1282 TargetLowering::SimplifySetCC(MVT VT, SDValue N0, SDValue N1,
1283 ISD::CondCode Cond, bool foldBooleans,
1284 DAGCombinerInfo &DCI) const {
1285 SelectionDAG &DAG = DCI.DAG;
1287 // These setcc operations always fold.
1291 case ISD::SETFALSE2: return DAG.getConstant(0, VT);
1293 case ISD::SETTRUE2: return DAG.getConstant(1, VT);
1296 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) {
1297 const APInt &C1 = N1C->getAPIntValue();
1298 if (isa<ConstantSDNode>(N0.getNode())) {
1299 return DAG.FoldSetCC(VT, N0, N1, Cond);
1301 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an
1302 // equality comparison, then we're just comparing whether X itself is
1304 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) &&
1305 N0.getOperand(0).getOpcode() == ISD::CTLZ &&
1306 N0.getOperand(1).getOpcode() == ISD::Constant) {
1307 unsigned ShAmt = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
1308 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1309 ShAmt == Log2_32(N0.getValueType().getSizeInBits())) {
1310 if ((C1 == 0) == (Cond == ISD::SETEQ)) {
1311 // (srl (ctlz x), 5) == 0 -> X != 0
1312 // (srl (ctlz x), 5) != 1 -> X != 0
1315 // (srl (ctlz x), 5) != 0 -> X == 0
1316 // (srl (ctlz x), 5) == 1 -> X == 0
1319 SDValue Zero = DAG.getConstant(0, N0.getValueType());
1320 return DAG.getSetCC(VT, N0.getOperand(0).getOperand(0),
1325 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
1326 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
1327 unsigned InSize = N0.getOperand(0).getValueType().getSizeInBits();
1329 // If the comparison constant has bits in the upper part, the
1330 // zero-extended value could never match.
1331 if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(),
1332 C1.getBitWidth() - InSize))) {
1336 case ISD::SETEQ: return DAG.getConstant(0, VT);
1339 case ISD::SETNE: return DAG.getConstant(1, VT);
1342 // True if the sign bit of C1 is set.
1343 return DAG.getConstant(C1.isNegative(), VT);
1346 // True if the sign bit of C1 isn't set.
1347 return DAG.getConstant(C1.isNonNegative(), VT);
1353 // Otherwise, we can perform the comparison with the low bits.
1361 return DAG.getSetCC(VT, N0.getOperand(0),
1362 DAG.getConstant(APInt(C1).trunc(InSize),
1363 N0.getOperand(0).getValueType()),
1366 break; // todo, be more careful with signed comparisons
1368 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
1369 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1370 MVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
1371 unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits();
1372 MVT ExtDstTy = N0.getValueType();
1373 unsigned ExtDstTyBits = ExtDstTy.getSizeInBits();
1375 // If the extended part has any inconsistent bits, it cannot ever
1376 // compare equal. In other words, they have to be all ones or all
1379 APInt::getHighBitsSet(ExtDstTyBits, ExtDstTyBits - ExtSrcTyBits);
1380 if ((C1 & ExtBits) != 0 && (C1 & ExtBits) != ExtBits)
1381 return DAG.getConstant(Cond == ISD::SETNE, VT);
1384 MVT Op0Ty = N0.getOperand(0).getValueType();
1385 if (Op0Ty == ExtSrcTy) {
1386 ZextOp = N0.getOperand(0);
1388 APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits);
1389 ZextOp = DAG.getNode(ISD::AND, Op0Ty, N0.getOperand(0),
1390 DAG.getConstant(Imm, Op0Ty));
1392 if (!DCI.isCalledByLegalizer())
1393 DCI.AddToWorklist(ZextOp.getNode());
1394 // Otherwise, make this a use of a zext.
1395 return DAG.getSetCC(VT, ZextOp,
1396 DAG.getConstant(C1 & APInt::getLowBitsSet(
1401 } else if ((N1C->isNullValue() || N1C->getAPIntValue() == 1) &&
1402 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1404 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC
1405 if (N0.getOpcode() == ISD::SETCC) {
1406 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getZExtValue() != 1);
1410 // Invert the condition.
1411 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
1412 CC = ISD::getSetCCInverse(CC,
1413 N0.getOperand(0).getValueType().isInteger());
1414 return DAG.getSetCC(VT, N0.getOperand(0), N0.getOperand(1), CC);
1417 if ((N0.getOpcode() == ISD::XOR ||
1418 (N0.getOpcode() == ISD::AND &&
1419 N0.getOperand(0).getOpcode() == ISD::XOR &&
1420 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
1421 isa<ConstantSDNode>(N0.getOperand(1)) &&
1422 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue() == 1) {
1423 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We
1424 // can only do this if the top bits are known zero.
1425 unsigned BitWidth = N0.getValueSizeInBits();
1426 if (DAG.MaskedValueIsZero(N0,
1427 APInt::getHighBitsSet(BitWidth,
1429 // Okay, get the un-inverted input value.
1431 if (N0.getOpcode() == ISD::XOR)
1432 Val = N0.getOperand(0);
1434 assert(N0.getOpcode() == ISD::AND &&
1435 N0.getOperand(0).getOpcode() == ISD::XOR);
1436 // ((X^1)&1)^1 -> X & 1
1437 Val = DAG.getNode(ISD::AND, N0.getValueType(),
1438 N0.getOperand(0).getOperand(0),
1441 return DAG.getSetCC(VT, Val, N1,
1442 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
1447 APInt MinVal, MaxVal;
1448 unsigned OperandBitSize = N1C->getValueType(0).getSizeInBits();
1449 if (ISD::isSignedIntSetCC(Cond)) {
1450 MinVal = APInt::getSignedMinValue(OperandBitSize);
1451 MaxVal = APInt::getSignedMaxValue(OperandBitSize);
1453 MinVal = APInt::getMinValue(OperandBitSize);
1454 MaxVal = APInt::getMaxValue(OperandBitSize);
1457 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
1458 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
1459 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true
1460 // X >= C0 --> X > (C0-1)
1461 return DAG.getSetCC(VT, N0, DAG.getConstant(C1-1, N1.getValueType()),
1462 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
1465 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
1466 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true
1467 // X <= C0 --> X < (C0+1)
1468 return DAG.getSetCC(VT, N0, DAG.getConstant(C1+1, N1.getValueType()),
1469 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
1472 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
1473 return DAG.getConstant(0, VT); // X < MIN --> false
1474 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal)
1475 return DAG.getConstant(1, VT); // X >= MIN --> true
1476 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal)
1477 return DAG.getConstant(0, VT); // X > MAX --> false
1478 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal)
1479 return DAG.getConstant(1, VT); // X <= MAX --> true
1481 // Canonicalize setgt X, Min --> setne X, Min
1482 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
1483 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
1484 // Canonicalize setlt X, Max --> setne X, Max
1485 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
1486 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
1488 // If we have setult X, 1, turn it into seteq X, 0
1489 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
1490 return DAG.getSetCC(VT, N0, DAG.getConstant(MinVal, N0.getValueType()),
1492 // If we have setugt X, Max-1, turn it into seteq X, Max
1493 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
1494 return DAG.getSetCC(VT, N0, DAG.getConstant(MaxVal, N0.getValueType()),
1497 // If we have "setcc X, C0", check to see if we can shrink the immediate
1500 // SETUGT X, SINTMAX -> SETLT X, 0
1501 if (Cond == ISD::SETUGT && OperandBitSize != 1 &&
1502 C1 == (~0ULL >> (65-OperandBitSize)))
1503 return DAG.getSetCC(VT, N0, DAG.getConstant(0, N1.getValueType()),
1506 // FIXME: Implement the rest of these.
1508 // Fold bit comparisons when we can.
1509 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1510 VT == N0.getValueType() && N0.getOpcode() == ISD::AND)
1511 if (ConstantSDNode *AndRHS =
1512 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1513 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
1514 // Perform the xform if the AND RHS is a single bit.
1515 if (isPowerOf2_64(AndRHS->getZExtValue())) {
1516 return DAG.getNode(ISD::SRL, VT, N0,
1517 DAG.getConstant(Log2_64(AndRHS->getZExtValue()),
1518 getShiftAmountTy()));
1520 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getZExtValue()) {
1521 // (X & 8) == 8 --> (X & 8) >> 3
1522 // Perform the xform if C1 is a single bit.
1523 if (C1.isPowerOf2()) {
1524 return DAG.getNode(ISD::SRL, VT, N0,
1525 DAG.getConstant(C1.logBase2(), getShiftAmountTy()));
1530 } else if (isa<ConstantSDNode>(N0.getNode())) {
1531 // Ensure that the constant occurs on the RHS.
1532 return DAG.getSetCC(VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
1535 if (isa<ConstantFPSDNode>(N0.getNode())) {
1536 // Constant fold or commute setcc.
1537 SDValue O = DAG.FoldSetCC(VT, N0, N1, Cond);
1538 if (O.getNode()) return O;
1539 } else if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1.getNode())) {
1540 // If the RHS of an FP comparison is a constant, simplify it away in
1542 if (CFP->getValueAPF().isNaN()) {
1543 // If an operand is known to be a nan, we can fold it.
1544 switch (ISD::getUnorderedFlavor(Cond)) {
1545 default: assert(0 && "Unknown flavor!");
1546 case 0: // Known false.
1547 return DAG.getConstant(0, VT);
1548 case 1: // Known true.
1549 return DAG.getConstant(1, VT);
1550 case 2: // Undefined.
1551 return DAG.getNode(ISD::UNDEF, VT);
1555 // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the
1556 // constant if knowing that the operand is non-nan is enough. We prefer to
1557 // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to
1559 if (Cond == ISD::SETO || Cond == ISD::SETUO)
1560 return DAG.getSetCC(VT, N0, N0, Cond);
1564 // We can always fold X == X for integer setcc's.
1565 if (N0.getValueType().isInteger())
1566 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
1567 unsigned UOF = ISD::getUnorderedFlavor(Cond);
1568 if (UOF == 2) // FP operators that are undefined on NaNs.
1569 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
1570 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
1571 return DAG.getConstant(UOF, VT);
1572 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
1573 // if it is not already.
1574 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
1575 if (NewCond != Cond)
1576 return DAG.getSetCC(VT, N0, N1, NewCond);
1579 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1580 N0.getValueType().isInteger()) {
1581 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
1582 N0.getOpcode() == ISD::XOR) {
1583 // Simplify (X+Y) == (X+Z) --> Y == Z
1584 if (N0.getOpcode() == N1.getOpcode()) {
1585 if (N0.getOperand(0) == N1.getOperand(0))
1586 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(1), Cond);
1587 if (N0.getOperand(1) == N1.getOperand(1))
1588 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(0), Cond);
1589 if (DAG.isCommutativeBinOp(N0.getOpcode())) {
1590 // If X op Y == Y op X, try other combinations.
1591 if (N0.getOperand(0) == N1.getOperand(1))
1592 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(0), Cond);
1593 if (N0.getOperand(1) == N1.getOperand(0))
1594 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(1), Cond);
1598 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) {
1599 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1600 // Turn (X+C1) == C2 --> X == C2-C1
1601 if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse()) {
1602 return DAG.getSetCC(VT, N0.getOperand(0),
1603 DAG.getConstant(RHSC->getAPIntValue()-
1604 LHSR->getAPIntValue(),
1605 N0.getValueType()), Cond);
1608 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
1609 if (N0.getOpcode() == ISD::XOR)
1610 // If we know that all of the inverted bits are zero, don't bother
1611 // performing the inversion.
1612 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue()))
1614 DAG.getSetCC(VT, N0.getOperand(0),
1615 DAG.getConstant(LHSR->getAPIntValue() ^
1616 RHSC->getAPIntValue(),
1621 // Turn (C1-X) == C2 --> X == C1-C2
1622 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
1623 if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse()) {
1625 DAG.getSetCC(VT, N0.getOperand(1),
1626 DAG.getConstant(SUBC->getAPIntValue() -
1627 RHSC->getAPIntValue(),
1634 // Simplify (X+Z) == X --> Z == 0
1635 if (N0.getOperand(0) == N1)
1636 return DAG.getSetCC(VT, N0.getOperand(1),
1637 DAG.getConstant(0, N0.getValueType()), Cond);
1638 if (N0.getOperand(1) == N1) {
1639 if (DAG.isCommutativeBinOp(N0.getOpcode()))
1640 return DAG.getSetCC(VT, N0.getOperand(0),
1641 DAG.getConstant(0, N0.getValueType()), Cond);
1642 else if (N0.getNode()->hasOneUse()) {
1643 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
1644 // (Z-X) == X --> Z == X<<1
1645 SDValue SH = DAG.getNode(ISD::SHL, N1.getValueType(),
1647 DAG.getConstant(1, getShiftAmountTy()));
1648 if (!DCI.isCalledByLegalizer())
1649 DCI.AddToWorklist(SH.getNode());
1650 return DAG.getSetCC(VT, N0.getOperand(0), SH, Cond);
1655 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
1656 N1.getOpcode() == ISD::XOR) {
1657 // Simplify X == (X+Z) --> Z == 0
1658 if (N1.getOperand(0) == N0) {
1659 return DAG.getSetCC(VT, N1.getOperand(1),
1660 DAG.getConstant(0, N1.getValueType()), Cond);
1661 } else if (N1.getOperand(1) == N0) {
1662 if (DAG.isCommutativeBinOp(N1.getOpcode())) {
1663 return DAG.getSetCC(VT, N1.getOperand(0),
1664 DAG.getConstant(0, N1.getValueType()), Cond);
1665 } else if (N1.getNode()->hasOneUse()) {
1666 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
1667 // X == (Z-X) --> X<<1 == Z
1668 SDValue SH = DAG.getNode(ISD::SHL, N1.getValueType(), N0,
1669 DAG.getConstant(1, getShiftAmountTy()));
1670 if (!DCI.isCalledByLegalizer())
1671 DCI.AddToWorklist(SH.getNode());
1672 return DAG.getSetCC(VT, SH, N1.getOperand(0), Cond);
1678 // Fold away ALL boolean setcc's.
1680 if (N0.getValueType() == MVT::i1 && foldBooleans) {
1682 default: assert(0 && "Unknown integer setcc!");
1683 case ISD::SETEQ: // X == Y -> (X^Y)^1
1684 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
1685 N0 = DAG.getNode(ISD::XOR, MVT::i1, Temp, DAG.getConstant(1, MVT::i1));
1686 if (!DCI.isCalledByLegalizer())
1687 DCI.AddToWorklist(Temp.getNode());
1689 case ISD::SETNE: // X != Y --> (X^Y)
1690 N0 = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
1692 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> X^1 & Y
1693 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> X^1 & Y
1694 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
1695 N0 = DAG.getNode(ISD::AND, MVT::i1, N1, Temp);
1696 if (!DCI.isCalledByLegalizer())
1697 DCI.AddToWorklist(Temp.getNode());
1699 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> Y^1 & X
1700 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> Y^1 & X
1701 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
1702 N0 = DAG.getNode(ISD::AND, MVT::i1, N0, Temp);
1703 if (!DCI.isCalledByLegalizer())
1704 DCI.AddToWorklist(Temp.getNode());
1706 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> X^1 | Y
1707 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> X^1 | Y
1708 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
1709 N0 = DAG.getNode(ISD::OR, MVT::i1, N1, Temp);
1710 if (!DCI.isCalledByLegalizer())
1711 DCI.AddToWorklist(Temp.getNode());
1713 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> Y^1 | X
1714 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> Y^1 | X
1715 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
1716 N0 = DAG.getNode(ISD::OR, MVT::i1, N0, Temp);
1719 if (VT != MVT::i1) {
1720 if (!DCI.isCalledByLegalizer())
1721 DCI.AddToWorklist(N0.getNode());
1722 // FIXME: If running after legalize, we probably can't do this.
1723 N0 = DAG.getNode(ISD::ZERO_EXTEND, VT, N0);
1728 // Could not fold it.
1732 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
1733 /// node is a GlobalAddress + offset.
1734 bool TargetLowering::isGAPlusOffset(SDNode *N, GlobalValue* &GA,
1735 int64_t &Offset) const {
1736 if (isa<GlobalAddressSDNode>(N)) {
1737 GlobalAddressSDNode *GASD = cast<GlobalAddressSDNode>(N);
1738 GA = GASD->getGlobal();
1739 Offset += GASD->getOffset();
1743 if (N->getOpcode() == ISD::ADD) {
1744 SDValue N1 = N->getOperand(0);
1745 SDValue N2 = N->getOperand(1);
1746 if (isGAPlusOffset(N1.getNode(), GA, Offset)) {
1747 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
1749 Offset += V->getSExtValue();
1752 } else if (isGAPlusOffset(N2.getNode(), GA, Offset)) {
1753 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
1755 Offset += V->getSExtValue();
1764 /// isConsecutiveLoad - Return true if LD (which must be a LoadSDNode) is
1765 /// loading 'Bytes' bytes from a location that is 'Dist' units away from the
1766 /// location that the 'Base' load is loading from.
1767 bool TargetLowering::isConsecutiveLoad(SDNode *LD, SDNode *Base,
1768 unsigned Bytes, int Dist,
1769 const MachineFrameInfo *MFI) const {
1770 if (LD->getOperand(0).getNode() != Base->getOperand(0).getNode())
1772 MVT VT = LD->getValueType(0);
1773 if (VT.getSizeInBits() / 8 != Bytes)
1776 SDValue Loc = LD->getOperand(1);
1777 SDValue BaseLoc = Base->getOperand(1);
1778 if (Loc.getOpcode() == ISD::FrameIndex) {
1779 if (BaseLoc.getOpcode() != ISD::FrameIndex)
1781 int FI = cast<FrameIndexSDNode>(Loc)->getIndex();
1782 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
1783 int FS = MFI->getObjectSize(FI);
1784 int BFS = MFI->getObjectSize(BFI);
1785 if (FS != BFS || FS != (int)Bytes) return false;
1786 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Bytes);
1789 GlobalValue *GV1 = NULL;
1790 GlobalValue *GV2 = NULL;
1791 int64_t Offset1 = 0;
1792 int64_t Offset2 = 0;
1793 bool isGA1 = isGAPlusOffset(Loc.getNode(), GV1, Offset1);
1794 bool isGA2 = isGAPlusOffset(BaseLoc.getNode(), GV2, Offset2);
1795 if (isGA1 && isGA2 && GV1 == GV2)
1796 return Offset1 == (Offset2 + Dist*Bytes);
1801 SDValue TargetLowering::
1802 PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
1803 // Default implementation: no optimization.
1807 //===----------------------------------------------------------------------===//
1808 // Inline Assembler Implementation Methods
1809 //===----------------------------------------------------------------------===//
1812 TargetLowering::ConstraintType
1813 TargetLowering::getConstraintType(const std::string &Constraint) const {
1814 // FIXME: lots more standard ones to handle.
1815 if (Constraint.size() == 1) {
1816 switch (Constraint[0]) {
1818 case 'r': return C_RegisterClass;
1820 case 'o': // offsetable
1821 case 'V': // not offsetable
1823 case 'i': // Simple Integer or Relocatable Constant
1824 case 'n': // Simple Integer
1825 case 's': // Relocatable Constant
1826 case 'X': // Allow ANY value.
1827 case 'I': // Target registers.
1839 if (Constraint.size() > 1 && Constraint[0] == '{' &&
1840 Constraint[Constraint.size()-1] == '}')
1845 /// LowerXConstraint - try to replace an X constraint, which matches anything,
1846 /// with another that has more specific requirements based on the type of the
1847 /// corresponding operand.
1848 const char *TargetLowering::LowerXConstraint(MVT ConstraintVT) const{
1849 if (ConstraintVT.isInteger())
1851 if (ConstraintVT.isFloatingPoint())
1852 return "f"; // works for many targets
1856 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
1857 /// vector. If it is invalid, don't add anything to Ops.
1858 void TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
1859 char ConstraintLetter,
1861 std::vector<SDValue> &Ops,
1862 SelectionDAG &DAG) const {
1863 switch (ConstraintLetter) {
1865 case 'X': // Allows any operand; labels (basic block) use this.
1866 if (Op.getOpcode() == ISD::BasicBlock) {
1871 case 'i': // Simple Integer or Relocatable Constant
1872 case 'n': // Simple Integer
1873 case 's': { // Relocatable Constant
1874 // These operands are interested in values of the form (GV+C), where C may
1875 // be folded in as an offset of GV, or it may be explicitly added. Also, it
1876 // is possible and fine if either GV or C are missing.
1877 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
1878 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
1880 // If we have "(add GV, C)", pull out GV/C
1881 if (Op.getOpcode() == ISD::ADD) {
1882 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
1883 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
1884 if (C == 0 || GA == 0) {
1885 C = dyn_cast<ConstantSDNode>(Op.getOperand(0));
1886 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1));
1888 if (C == 0 || GA == 0)
1892 // If we find a valid operand, map to the TargetXXX version so that the
1893 // value itself doesn't get selected.
1894 if (GA) { // Either &GV or &GV+C
1895 if (ConstraintLetter != 'n') {
1896 int64_t Offs = GA->getOffset();
1897 if (C) Offs += C->getZExtValue();
1898 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(),
1899 Op.getValueType(), Offs));
1903 if (C) { // just C, no GV.
1904 // Simple constants are not allowed for 's'.
1905 if (ConstraintLetter != 's') {
1906 Ops.push_back(DAG.getTargetConstant(C->getAPIntValue(),
1907 Op.getValueType()));
1916 std::vector<unsigned> TargetLowering::
1917 getRegClassForInlineAsmConstraint(const std::string &Constraint,
1919 return std::vector<unsigned>();
1923 std::pair<unsigned, const TargetRegisterClass*> TargetLowering::
1924 getRegForInlineAsmConstraint(const std::string &Constraint,
1926 if (Constraint[0] != '{')
1927 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
1928 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?");
1930 // Remove the braces from around the name.
1931 std::string RegName(Constraint.begin()+1, Constraint.end()-1);
1933 // Figure out which register class contains this reg.
1934 const TargetRegisterInfo *RI = TM.getRegisterInfo();
1935 for (TargetRegisterInfo::regclass_iterator RCI = RI->regclass_begin(),
1936 E = RI->regclass_end(); RCI != E; ++RCI) {
1937 const TargetRegisterClass *RC = *RCI;
1939 // If none of the the value types for this register class are valid, we
1940 // can't use it. For example, 64-bit reg classes on 32-bit targets.
1941 bool isLegal = false;
1942 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
1944 if (isTypeLegal(*I)) {
1950 if (!isLegal) continue;
1952 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end();
1954 if (StringsEqualNoCase(RegName, RI->get(*I).AsmName))
1955 return std::make_pair(*I, RC);
1959 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
1962 //===----------------------------------------------------------------------===//
1963 // Constraint Selection.
1965 /// isMatchingConstraint - Return true of this is an input operand that is a
1966 /// matching constraint like "4".
1967 bool TargetLowering::AsmOperandInfo::isMatchingConstraint() const {
1968 assert(!ConstraintCode.empty() && "No known constraint!");
1969 return isdigit(ConstraintCode[0]);
1972 /// getMatchedOperand - If this is an input matching constraint, this method
1973 /// returns the output operand it matches.
1974 unsigned TargetLowering::AsmOperandInfo::getMatchedOperand() const {
1975 assert(!ConstraintCode.empty() && "No known constraint!");
1976 return atoi(ConstraintCode.c_str());
1980 /// getConstraintGenerality - Return an integer indicating how general CT
1982 static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) {
1984 default: assert(0 && "Unknown constraint type!");
1985 case TargetLowering::C_Other:
1986 case TargetLowering::C_Unknown:
1988 case TargetLowering::C_Register:
1990 case TargetLowering::C_RegisterClass:
1992 case TargetLowering::C_Memory:
1997 /// ChooseConstraint - If there are multiple different constraints that we
1998 /// could pick for this operand (e.g. "imr") try to pick the 'best' one.
1999 /// This is somewhat tricky: constraints fall into four classes:
2000 /// Other -> immediates and magic values
2001 /// Register -> one specific register
2002 /// RegisterClass -> a group of regs
2003 /// Memory -> memory
2004 /// Ideally, we would pick the most specific constraint possible: if we have
2005 /// something that fits into a register, we would pick it. The problem here
2006 /// is that if we have something that could either be in a register or in
2007 /// memory that use of the register could cause selection of *other*
2008 /// operands to fail: they might only succeed if we pick memory. Because of
2009 /// this the heuristic we use is:
2011 /// 1) If there is an 'other' constraint, and if the operand is valid for
2012 /// that constraint, use it. This makes us take advantage of 'i'
2013 /// constraints when available.
2014 /// 2) Otherwise, pick the most general constraint present. This prefers
2015 /// 'm' over 'r', for example.
2017 static void ChooseConstraint(TargetLowering::AsmOperandInfo &OpInfo,
2018 bool hasMemory, const TargetLowering &TLI,
2019 SDValue Op, SelectionDAG *DAG) {
2020 assert(OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options");
2021 unsigned BestIdx = 0;
2022 TargetLowering::ConstraintType BestType = TargetLowering::C_Unknown;
2023 int BestGenerality = -1;
2025 // Loop over the options, keeping track of the most general one.
2026 for (unsigned i = 0, e = OpInfo.Codes.size(); i != e; ++i) {
2027 TargetLowering::ConstraintType CType =
2028 TLI.getConstraintType(OpInfo.Codes[i]);
2030 // If this is an 'other' constraint, see if the operand is valid for it.
2031 // For example, on X86 we might have an 'rI' constraint. If the operand
2032 // is an integer in the range [0..31] we want to use I (saving a load
2033 // of a register), otherwise we must use 'r'.
2034 if (CType == TargetLowering::C_Other && Op.getNode()) {
2035 assert(OpInfo.Codes[i].size() == 1 &&
2036 "Unhandled multi-letter 'other' constraint");
2037 std::vector<SDValue> ResultOps;
2038 TLI.LowerAsmOperandForConstraint(Op, OpInfo.Codes[i][0], hasMemory,
2040 if (!ResultOps.empty()) {
2047 // This constraint letter is more general than the previous one, use it.
2048 int Generality = getConstraintGenerality(CType);
2049 if (Generality > BestGenerality) {
2052 BestGenerality = Generality;
2056 OpInfo.ConstraintCode = OpInfo.Codes[BestIdx];
2057 OpInfo.ConstraintType = BestType;
2060 /// ComputeConstraintToUse - Determines the constraint code and constraint
2061 /// type to use for the specific AsmOperandInfo, setting
2062 /// OpInfo.ConstraintCode and OpInfo.ConstraintType.
2063 void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo,
2066 SelectionDAG *DAG) const {
2067 assert(!OpInfo.Codes.empty() && "Must have at least one constraint");
2069 // Single-letter constraints ('r') are very common.
2070 if (OpInfo.Codes.size() == 1) {
2071 OpInfo.ConstraintCode = OpInfo.Codes[0];
2072 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
2074 ChooseConstraint(OpInfo, hasMemory, *this, Op, DAG);
2077 // 'X' matches anything.
2078 if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) {
2079 // Labels and constants are handled elsewhere ('X' is the only thing
2080 // that matches labels).
2081 if (isa<BasicBlock>(OpInfo.CallOperandVal) ||
2082 isa<ConstantInt>(OpInfo.CallOperandVal))
2085 // Otherwise, try to resolve it to something we know about by looking at
2086 // the actual operand type.
2087 if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) {
2088 OpInfo.ConstraintCode = Repl;
2089 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
2094 //===----------------------------------------------------------------------===//
2095 // Loop Strength Reduction hooks
2096 //===----------------------------------------------------------------------===//
2098 /// isLegalAddressingMode - Return true if the addressing mode represented
2099 /// by AM is legal for this target, for a load/store of the specified type.
2100 bool TargetLowering::isLegalAddressingMode(const AddrMode &AM,
2101 const Type *Ty) const {
2102 // The default implementation of this implements a conservative RISCy, r+r and
2105 // Allows a sign-extended 16-bit immediate field.
2106 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
2109 // No global is ever allowed as a base.
2113 // Only support r+r,
2115 case 0: // "r+i" or just "i", depending on HasBaseReg.
2118 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
2120 // Otherwise we have r+r or r+i.
2123 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
2125 // Allow 2*r as r+r.
2132 // Magic for divide replacement
2135 int64_t m; // magic number
2136 int64_t s; // shift amount
2140 uint64_t m; // magic number
2141 int64_t a; // add indicator
2142 int64_t s; // shift amount
2145 /// magic - calculate the magic numbers required to codegen an integer sdiv as
2146 /// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
2148 static ms magic32(int32_t d) {
2150 uint32_t ad, anc, delta, q1, r1, q2, r2, t;
2151 const uint32_t two31 = 0x80000000U;
2155 t = two31 + ((uint32_t)d >> 31);
2156 anc = t - 1 - t%ad; // absolute value of nc
2157 p = 31; // initialize p
2158 q1 = two31/anc; // initialize q1 = 2p/abs(nc)
2159 r1 = two31 - q1*anc; // initialize r1 = rem(2p,abs(nc))
2160 q2 = two31/ad; // initialize q2 = 2p/abs(d)
2161 r2 = two31 - q2*ad; // initialize r2 = rem(2p,abs(d))
2164 q1 = 2*q1; // update q1 = 2p/abs(nc)
2165 r1 = 2*r1; // update r1 = rem(2p/abs(nc))
2166 if (r1 >= anc) { // must be unsigned comparison
2170 q2 = 2*q2; // update q2 = 2p/abs(d)
2171 r2 = 2*r2; // update r2 = rem(2p/abs(d))
2172 if (r2 >= ad) { // must be unsigned comparison
2177 } while (q1 < delta || (q1 == delta && r1 == 0));
2179 mag.m = (int32_t)(q2 + 1); // make sure to sign extend
2180 if (d < 0) mag.m = -mag.m; // resulting magic number
2181 mag.s = p - 32; // resulting shift
2185 /// magicu - calculate the magic numbers required to codegen an integer udiv as
2186 /// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
2187 static mu magicu32(uint32_t d) {
2189 uint32_t nc, delta, q1, r1, q2, r2;
2191 magu.a = 0; // initialize "add" indicator
2193 p = 31; // initialize p
2194 q1 = 0x80000000/nc; // initialize q1 = 2p/nc
2195 r1 = 0x80000000 - q1*nc; // initialize r1 = rem(2p,nc)
2196 q2 = 0x7FFFFFFF/d; // initialize q2 = (2p-1)/d
2197 r2 = 0x7FFFFFFF - q2*d; // initialize r2 = rem((2p-1),d)
2200 if (r1 >= nc - r1 ) {
2201 q1 = 2*q1 + 1; // update q1
2202 r1 = 2*r1 - nc; // update r1
2205 q1 = 2*q1; // update q1
2206 r1 = 2*r1; // update r1
2208 if (r2 + 1 >= d - r2) {
2209 if (q2 >= 0x7FFFFFFF) magu.a = 1;
2210 q2 = 2*q2 + 1; // update q2
2211 r2 = 2*r2 + 1 - d; // update r2
2214 if (q2 >= 0x80000000) magu.a = 1;
2215 q2 = 2*q2; // update q2
2216 r2 = 2*r2 + 1; // update r2
2219 } while (p < 64 && (q1 < delta || (q1 == delta && r1 == 0)));
2220 magu.m = q2 + 1; // resulting magic number
2221 magu.s = p - 32; // resulting shift
2225 /// magic - calculate the magic numbers required to codegen an integer sdiv as
2226 /// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
2228 static ms magic64(int64_t d) {
2230 uint64_t ad, anc, delta, q1, r1, q2, r2, t;
2231 const uint64_t two63 = 9223372036854775808ULL; // 2^63
2234 ad = d >= 0 ? d : -d;
2235 t = two63 + ((uint64_t)d >> 63);
2236 anc = t - 1 - t%ad; // absolute value of nc
2237 p = 63; // initialize p
2238 q1 = two63/anc; // initialize q1 = 2p/abs(nc)
2239 r1 = two63 - q1*anc; // initialize r1 = rem(2p,abs(nc))
2240 q2 = two63/ad; // initialize q2 = 2p/abs(d)
2241 r2 = two63 - q2*ad; // initialize r2 = rem(2p,abs(d))
2244 q1 = 2*q1; // update q1 = 2p/abs(nc)
2245 r1 = 2*r1; // update r1 = rem(2p/abs(nc))
2246 if (r1 >= anc) { // must be unsigned comparison
2250 q2 = 2*q2; // update q2 = 2p/abs(d)
2251 r2 = 2*r2; // update r2 = rem(2p/abs(d))
2252 if (r2 >= ad) { // must be unsigned comparison
2257 } while (q1 < delta || (q1 == delta && r1 == 0));
2260 if (d < 0) mag.m = -mag.m; // resulting magic number
2261 mag.s = p - 64; // resulting shift
2265 /// magicu - calculate the magic numbers required to codegen an integer udiv as
2266 /// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
2267 static mu magicu64(uint64_t d)
2270 uint64_t nc, delta, q1, r1, q2, r2;
2272 magu.a = 0; // initialize "add" indicator
2274 p = 63; // initialize p
2275 q1 = 0x8000000000000000ull/nc; // initialize q1 = 2p/nc
2276 r1 = 0x8000000000000000ull - q1*nc; // initialize r1 = rem(2p,nc)
2277 q2 = 0x7FFFFFFFFFFFFFFFull/d; // initialize q2 = (2p-1)/d
2278 r2 = 0x7FFFFFFFFFFFFFFFull - q2*d; // initialize r2 = rem((2p-1),d)
2281 if (r1 >= nc - r1 ) {
2282 q1 = 2*q1 + 1; // update q1
2283 r1 = 2*r1 - nc; // update r1
2286 q1 = 2*q1; // update q1
2287 r1 = 2*r1; // update r1
2289 if (r2 + 1 >= d - r2) {
2290 if (q2 >= 0x7FFFFFFFFFFFFFFFull) magu.a = 1;
2291 q2 = 2*q2 + 1; // update q2
2292 r2 = 2*r2 + 1 - d; // update r2
2295 if (q2 >= 0x8000000000000000ull) magu.a = 1;
2296 q2 = 2*q2; // update q2
2297 r2 = 2*r2 + 1; // update r2
2300 } while (p < 128 && (q1 < delta || (q1 == delta && r1 == 0)));
2301 magu.m = q2 + 1; // resulting magic number
2302 magu.s = p - 64; // resulting shift
2306 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
2307 /// return a DAG expression to select that will generate the same value by
2308 /// multiplying by a magic number. See:
2309 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
2310 SDValue TargetLowering::BuildSDIV(SDNode *N, SelectionDAG &DAG,
2311 std::vector<SDNode*>* Created) const {
2312 MVT VT = N->getValueType(0);
2314 // Check to see if we can do this.
2315 if (!isTypeLegal(VT) || (VT != MVT::i32 && VT != MVT::i64))
2316 return SDValue(); // BuildSDIV only operates on i32 or i64
2318 int64_t d = cast<ConstantSDNode>(N->getOperand(1))->getSExtValue();
2319 ms magics = (VT == MVT::i32) ? magic32(d) : magic64(d);
2321 // Multiply the numerator (operand 0) by the magic value
2323 if (isOperationLegal(ISD::MULHS, VT))
2324 Q = DAG.getNode(ISD::MULHS, VT, N->getOperand(0),
2325 DAG.getConstant(magics.m, VT));
2326 else if (isOperationLegal(ISD::SMUL_LOHI, VT))
2327 Q = SDValue(DAG.getNode(ISD::SMUL_LOHI, DAG.getVTList(VT, VT),
2329 DAG.getConstant(magics.m, VT)).getNode(), 1);
2331 return SDValue(); // No mulhs or equvialent
2332 // If d > 0 and m < 0, add the numerator
2333 if (d > 0 && magics.m < 0) {
2334 Q = DAG.getNode(ISD::ADD, VT, Q, N->getOperand(0));
2336 Created->push_back(Q.getNode());
2338 // If d < 0 and m > 0, subtract the numerator.
2339 if (d < 0 && magics.m > 0) {
2340 Q = DAG.getNode(ISD::SUB, VT, Q, N->getOperand(0));
2342 Created->push_back(Q.getNode());
2344 // Shift right algebraic if shift value is nonzero
2346 Q = DAG.getNode(ISD::SRA, VT, Q,
2347 DAG.getConstant(magics.s, getShiftAmountTy()));
2349 Created->push_back(Q.getNode());
2351 // Extract the sign bit and add it to the quotient
2353 DAG.getNode(ISD::SRL, VT, Q, DAG.getConstant(VT.getSizeInBits()-1,
2354 getShiftAmountTy()));
2356 Created->push_back(T.getNode());
2357 return DAG.getNode(ISD::ADD, VT, Q, T);
2360 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
2361 /// return a DAG expression to select that will generate the same value by
2362 /// multiplying by a magic number. See:
2363 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
2364 SDValue TargetLowering::BuildUDIV(SDNode *N, SelectionDAG &DAG,
2365 std::vector<SDNode*>* Created) const {
2366 MVT VT = N->getValueType(0);
2368 // Check to see if we can do this.
2369 if (!isTypeLegal(VT) || (VT != MVT::i32 && VT != MVT::i64))
2370 return SDValue(); // BuildUDIV only operates on i32 or i64
2372 uint64_t d = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
2373 mu magics = (VT == MVT::i32) ? magicu32(d) : magicu64(d);
2375 // Multiply the numerator (operand 0) by the magic value
2377 if (isOperationLegal(ISD::MULHU, VT))
2378 Q = DAG.getNode(ISD::MULHU, VT, N->getOperand(0),
2379 DAG.getConstant(magics.m, VT));
2380 else if (isOperationLegal(ISD::UMUL_LOHI, VT))
2381 Q = SDValue(DAG.getNode(ISD::UMUL_LOHI, DAG.getVTList(VT, VT),
2383 DAG.getConstant(magics.m, VT)).getNode(), 1);
2385 return SDValue(); // No mulhu or equvialent
2387 Created->push_back(Q.getNode());
2389 if (magics.a == 0) {
2390 return DAG.getNode(ISD::SRL, VT, Q,
2391 DAG.getConstant(magics.s, getShiftAmountTy()));
2393 SDValue NPQ = DAG.getNode(ISD::SUB, VT, N->getOperand(0), Q);
2395 Created->push_back(NPQ.getNode());
2396 NPQ = DAG.getNode(ISD::SRL, VT, NPQ,
2397 DAG.getConstant(1, getShiftAmountTy()));
2399 Created->push_back(NPQ.getNode());
2400 NPQ = DAG.getNode(ISD::ADD, VT, NPQ, Q);
2402 Created->push_back(NPQ.getNode());
2403 return DAG.getNode(ISD::SRL, VT, NPQ,
2404 DAG.getConstant(magics.s-1, getShiftAmountTy()));