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(LoadXActions, 0, sizeof(LoadXActions));
406 memset(TruncStoreActions, 0, sizeof(TruncStoreActions));
407 memset(IndexedModeActions, 0, sizeof(IndexedModeActions));
408 memset(ConvertActions, 0, sizeof(ConvertActions));
410 // Set default actions for various operations.
411 for (unsigned VT = 0; VT != (unsigned)MVT::LAST_VALUETYPE; ++VT) {
412 // Default all indexed load / store to expand.
413 for (unsigned IM = (unsigned)ISD::PRE_INC;
414 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
415 setIndexedLoadAction(IM, (MVT::SimpleValueType)VT, Expand);
416 setIndexedStoreAction(IM, (MVT::SimpleValueType)VT, Expand);
419 // These operations default to expand.
420 setOperationAction(ISD::FGETSIGN, (MVT::SimpleValueType)VT, Expand);
423 // Most targets ignore the @llvm.prefetch intrinsic.
424 setOperationAction(ISD::PREFETCH, MVT::Other, Expand);
426 // ConstantFP nodes default to expand. Targets can either change this to
427 // Legal, in which case all fp constants are legal, or use addLegalFPImmediate
428 // to optimize expansions for certain constants.
429 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
430 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
431 setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
433 // Default ISD::TRAP to expand (which turns it into abort).
434 setOperationAction(ISD::TRAP, MVT::Other, Expand);
436 IsLittleEndian = TD->isLittleEndian();
437 UsesGlobalOffsetTable = false;
438 ShiftAmountTy = PointerTy = getValueType(TD->getIntPtrType());
439 ShiftAmtHandling = Undefined;
440 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE*sizeof(TargetRegisterClass*));
441 memset(TargetDAGCombineArray, 0, array_lengthof(TargetDAGCombineArray));
442 maxStoresPerMemset = maxStoresPerMemcpy = maxStoresPerMemmove = 8;
443 allowUnalignedMemoryAccesses = false;
444 UseUnderscoreSetJmp = false;
445 UseUnderscoreLongJmp = false;
446 SelectIsExpensive = false;
447 IntDivIsCheap = false;
448 Pow2DivIsCheap = false;
449 StackPointerRegisterToSaveRestore = 0;
450 ExceptionPointerRegister = 0;
451 ExceptionSelectorRegister = 0;
452 SetCCResultContents = UndefinedSetCCResult;
453 SchedPreferenceInfo = SchedulingForLatency;
455 JumpBufAlignment = 0;
456 IfCvtBlockSizeLimit = 2;
457 IfCvtDupBlockSizeLimit = 0;
458 PrefLoopAlignment = 0;
460 InitLibcallNames(LibcallRoutineNames);
461 InitCmpLibcallCCs(CmpLibcallCCs);
463 // Tell Legalize whether the assembler supports DEBUG_LOC.
464 if (!TM.getTargetAsmInfo()->hasDotLocAndDotFile())
465 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
468 TargetLowering::~TargetLowering() {}
470 /// computeRegisterProperties - Once all of the register classes are added,
471 /// this allows us to compute derived properties we expose.
472 void TargetLowering::computeRegisterProperties() {
473 assert(MVT::LAST_VALUETYPE <= 32 &&
474 "Too many value types for ValueTypeActions to hold!");
476 // Everything defaults to needing one register.
477 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
478 NumRegistersForVT[i] = 1;
479 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i;
481 // ...except isVoid, which doesn't need any registers.
482 NumRegistersForVT[MVT::isVoid] = 0;
484 // Find the largest integer register class.
485 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE;
486 for (; RegClassForVT[LargestIntReg] == 0; --LargestIntReg)
487 assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
489 // Every integer value type larger than this largest register takes twice as
490 // many registers to represent as the previous ValueType.
491 for (unsigned ExpandedReg = LargestIntReg + 1; ; ++ExpandedReg) {
492 MVT EVT = (MVT::SimpleValueType)ExpandedReg;
493 if (!EVT.isInteger())
495 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
496 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg;
497 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1);
498 ValueTypeActions.setTypeAction(EVT, Expand);
501 // Inspect all of the ValueType's smaller than the largest integer
502 // register to see which ones need promotion.
503 unsigned LegalIntReg = LargestIntReg;
504 for (unsigned IntReg = LargestIntReg - 1;
505 IntReg >= (unsigned)MVT::i1; --IntReg) {
506 MVT IVT = (MVT::SimpleValueType)IntReg;
507 if (isTypeLegal(IVT)) {
508 LegalIntReg = IntReg;
510 RegisterTypeForVT[IntReg] = TransformToType[IntReg] =
511 (MVT::SimpleValueType)LegalIntReg;
512 ValueTypeActions.setTypeAction(IVT, Promote);
516 // ppcf128 type is really two f64's.
517 if (!isTypeLegal(MVT::ppcf128)) {
518 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64];
519 RegisterTypeForVT[MVT::ppcf128] = MVT::f64;
520 TransformToType[MVT::ppcf128] = MVT::f64;
521 ValueTypeActions.setTypeAction(MVT::ppcf128, Expand);
524 // Decide how to handle f64. If the target does not have native f64 support,
525 // expand it to i64 and we will be generating soft float library calls.
526 if (!isTypeLegal(MVT::f64)) {
527 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
528 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
529 TransformToType[MVT::f64] = MVT::i64;
530 ValueTypeActions.setTypeAction(MVT::f64, Expand);
533 // Decide how to handle f32. If the target does not have native support for
534 // f32, promote it to f64 if it is legal. Otherwise, expand it to i32.
535 if (!isTypeLegal(MVT::f32)) {
536 if (isTypeLegal(MVT::f64)) {
537 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::f64];
538 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::f64];
539 TransformToType[MVT::f32] = MVT::f64;
540 ValueTypeActions.setTypeAction(MVT::f32, Promote);
542 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
543 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
544 TransformToType[MVT::f32] = MVT::i32;
545 ValueTypeActions.setTypeAction(MVT::f32, Expand);
549 // Loop over all of the vector value types to see which need transformations.
550 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
551 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
552 MVT VT = (MVT::SimpleValueType)i;
553 if (!isTypeLegal(VT)) {
554 MVT IntermediateVT, RegisterVT;
555 unsigned NumIntermediates;
556 NumRegistersForVT[i] =
557 getVectorTypeBreakdown(VT,
558 IntermediateVT, NumIntermediates,
560 RegisterTypeForVT[i] = RegisterVT;
561 TransformToType[i] = MVT::Other; // this isn't actually used
562 ValueTypeActions.setTypeAction(VT, Expand);
567 const char *TargetLowering::getTargetNodeName(unsigned Opcode) const {
572 MVT TargetLowering::getSetCCResultType(const SDValue &) const {
573 return getValueType(TD->getIntPtrType());
577 /// getVectorTypeBreakdown - Vector types are broken down into some number of
578 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
579 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
580 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
582 /// This method returns the number of registers needed, and the VT for each
583 /// register. It also returns the VT and quantity of the intermediate values
584 /// before they are promoted/expanded.
586 unsigned TargetLowering::getVectorTypeBreakdown(MVT VT,
588 unsigned &NumIntermediates,
589 MVT &RegisterVT) const {
590 // Figure out the right, legal destination reg to copy into.
591 unsigned NumElts = VT.getVectorNumElements();
592 MVT EltTy = VT.getVectorElementType();
594 unsigned NumVectorRegs = 1;
596 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
597 // could break down into LHS/RHS like LegalizeDAG does.
598 if (!isPowerOf2_32(NumElts)) {
599 NumVectorRegs = NumElts;
603 // Divide the input until we get to a supported size. This will always
604 // end with a scalar if the target doesn't support vectors.
605 while (NumElts > 1 && !isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) {
610 NumIntermediates = NumVectorRegs;
612 MVT NewVT = MVT::getVectorVT(EltTy, NumElts);
613 if (!isTypeLegal(NewVT))
615 IntermediateVT = NewVT;
617 MVT DestVT = getTypeToTransformTo(NewVT);
619 if (DestVT.bitsLT(NewVT)) {
620 // Value is expanded, e.g. i64 -> i16.
621 return NumVectorRegs*(NewVT.getSizeInBits()/DestVT.getSizeInBits());
623 // Otherwise, promotion or legal types use the same number of registers as
624 // the vector decimated to the appropriate level.
625 return NumVectorRegs;
631 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
632 /// function arguments in the caller parameter area. This is the actual
633 /// alignment, not its logarithm.
634 unsigned TargetLowering::getByValTypeAlignment(const Type *Ty) const {
635 return TD->getCallFrameTypeAlignment(Ty);
638 SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table,
639 SelectionDAG &DAG) const {
640 if (usesGlobalOffsetTable())
641 return DAG.getNode(ISD::GLOBAL_OFFSET_TABLE, getPointerTy());
645 //===----------------------------------------------------------------------===//
646 // Optimization Methods
647 //===----------------------------------------------------------------------===//
649 /// ShrinkDemandedConstant - Check to see if the specified operand of the
650 /// specified instruction is a constant integer. If so, check to see if there
651 /// are any bits set in the constant that are not demanded. If so, shrink the
652 /// constant and return true.
653 bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDValue Op,
654 const APInt &Demanded) {
655 // FIXME: ISD::SELECT, ISD::SELECT_CC
656 switch(Op.getOpcode()) {
661 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1)))
662 if (C->getAPIntValue().intersects(~Demanded)) {
663 MVT VT = Op.getValueType();
664 SDValue New = DAG.getNode(Op.getOpcode(), VT, Op.getOperand(0),
665 DAG.getConstant(Demanded &
668 return CombineTo(Op, New);
675 /// SimplifyDemandedBits - Look at Op. At this point, we know that only the
676 /// DemandedMask bits of the result of Op are ever used downstream. If we can
677 /// use this information to simplify Op, create a new simplified DAG node and
678 /// return true, returning the original and new nodes in Old and New. Otherwise,
679 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
680 /// the expression (used to simplify the caller). The KnownZero/One bits may
681 /// only be accurate for those bits in the DemandedMask.
682 bool TargetLowering::SimplifyDemandedBits(SDValue Op,
683 const APInt &DemandedMask,
686 TargetLoweringOpt &TLO,
687 unsigned Depth) const {
688 unsigned BitWidth = DemandedMask.getBitWidth();
689 assert(Op.getValueSizeInBits() == BitWidth &&
690 "Mask size mismatches value type size!");
691 APInt NewMask = DemandedMask;
693 // Don't know anything.
694 KnownZero = KnownOne = APInt(BitWidth, 0);
696 // Other users may use these bits.
697 if (!Op.getNode()->hasOneUse()) {
699 // If not at the root, Just compute the KnownZero/KnownOne bits to
700 // simplify things downstream.
701 TLO.DAG.ComputeMaskedBits(Op, DemandedMask, KnownZero, KnownOne, Depth);
704 // If this is the root being simplified, allow it to have multiple uses,
705 // just set the NewMask to all bits.
706 NewMask = APInt::getAllOnesValue(BitWidth);
707 } else if (DemandedMask == 0) {
708 // Not demanding any bits from Op.
709 if (Op.getOpcode() != ISD::UNDEF)
710 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::UNDEF, Op.getValueType()));
712 } else if (Depth == 6) { // Limit search depth.
716 APInt KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut;
717 switch (Op.getOpcode()) {
719 // We know all of the bits for a constant!
720 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue() & NewMask;
721 KnownZero = ~KnownOne & NewMask;
722 return false; // Don't fall through, will infinitely loop.
724 // If the RHS is a constant, check to see if the LHS would be zero without
725 // using the bits from the RHS. Below, we use knowledge about the RHS to
726 // simplify the LHS, here we're using information from the LHS to simplify
728 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
729 APInt LHSZero, LHSOne;
730 TLO.DAG.ComputeMaskedBits(Op.getOperand(0), NewMask,
731 LHSZero, LHSOne, Depth+1);
732 // If the LHS already has zeros where RHSC does, this and is dead.
733 if ((LHSZero & NewMask) == (~RHSC->getAPIntValue() & NewMask))
734 return TLO.CombineTo(Op, Op.getOperand(0));
735 // If any of the set bits in the RHS are known zero on the LHS, shrink
737 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & NewMask))
741 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
742 KnownOne, TLO, Depth+1))
744 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
745 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownZero & NewMask,
746 KnownZero2, KnownOne2, TLO, Depth+1))
748 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
750 // If all of the demanded bits are known one on one side, return the other.
751 // These bits cannot contribute to the result of the 'and'.
752 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
753 return TLO.CombineTo(Op, Op.getOperand(0));
754 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
755 return TLO.CombineTo(Op, Op.getOperand(1));
756 // If all of the demanded bits in the inputs are known zeros, return zero.
757 if ((NewMask & (KnownZero|KnownZero2)) == NewMask)
758 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, Op.getValueType()));
759 // If the RHS is a constant, see if we can simplify it.
760 if (TLO.ShrinkDemandedConstant(Op, ~KnownZero2 & NewMask))
763 // Output known-1 bits are only known if set in both the LHS & RHS.
764 KnownOne &= KnownOne2;
765 // Output known-0 are known to be clear if zero in either the LHS | RHS.
766 KnownZero |= KnownZero2;
769 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
770 KnownOne, TLO, Depth+1))
772 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
773 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownOne & NewMask,
774 KnownZero2, KnownOne2, TLO, Depth+1))
776 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
778 // If all of the demanded bits are known zero on one side, return the other.
779 // These bits cannot contribute to the result of the 'or'.
780 if ((NewMask & ~KnownOne2 & KnownZero) == (~KnownOne2 & NewMask))
781 return TLO.CombineTo(Op, Op.getOperand(0));
782 if ((NewMask & ~KnownOne & KnownZero2) == (~KnownOne & NewMask))
783 return TLO.CombineTo(Op, Op.getOperand(1));
784 // If all of the potentially set bits on one side are known to be set on
785 // the other side, just use the 'other' side.
786 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
787 return TLO.CombineTo(Op, Op.getOperand(0));
788 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
789 return TLO.CombineTo(Op, Op.getOperand(1));
790 // If the RHS is a constant, see if we can simplify it.
791 if (TLO.ShrinkDemandedConstant(Op, NewMask))
794 // Output known-0 bits are only known if clear in both the LHS & RHS.
795 KnownZero &= KnownZero2;
796 // Output known-1 are known to be set if set in either the LHS | RHS.
797 KnownOne |= KnownOne2;
800 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
801 KnownOne, TLO, Depth+1))
803 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
804 if (SimplifyDemandedBits(Op.getOperand(0), NewMask, KnownZero2,
805 KnownOne2, TLO, Depth+1))
807 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
809 // If all of the demanded bits are known zero on one side, return the other.
810 // These bits cannot contribute to the result of the 'xor'.
811 if ((KnownZero & NewMask) == NewMask)
812 return TLO.CombineTo(Op, Op.getOperand(0));
813 if ((KnownZero2 & NewMask) == NewMask)
814 return TLO.CombineTo(Op, Op.getOperand(1));
816 // If all of the unknown bits are known to be zero on one side or the other
817 // (but not both) turn this into an *inclusive* or.
818 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
819 if ((NewMask & ~KnownZero & ~KnownZero2) == 0)
820 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, Op.getValueType(),
824 // Output known-0 bits are known if clear or set in both the LHS & RHS.
825 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
826 // Output known-1 are known to be set if set in only one of the LHS, RHS.
827 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
829 // If all of the demanded bits on one side are known, and all of the set
830 // bits on that side are also known to be set on the other side, turn this
831 // into an AND, as we know the bits will be cleared.
832 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
833 if ((NewMask & (KnownZero|KnownOne)) == NewMask) { // all known
834 if ((KnownOne & KnownOne2) == KnownOne) {
835 MVT VT = Op.getValueType();
836 SDValue ANDC = TLO.DAG.getConstant(~KnownOne & NewMask, VT);
837 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, VT, Op.getOperand(0),
842 // If the RHS is a constant, see if we can simplify it.
843 // for XOR, we prefer to force bits to 1 if they will make a -1.
844 // if we can't force bits, try to shrink constant
845 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
846 APInt Expanded = C->getAPIntValue() | (~NewMask);
847 // if we can expand it to have all bits set, do it
848 if (Expanded.isAllOnesValue()) {
849 if (Expanded != C->getAPIntValue()) {
850 MVT VT = Op.getValueType();
851 SDValue New = TLO.DAG.getNode(Op.getOpcode(), VT, Op.getOperand(0),
852 TLO.DAG.getConstant(Expanded, VT));
853 return TLO.CombineTo(Op, New);
855 // if it already has all the bits set, nothing to change
856 // but don't shrink either!
857 } else if (TLO.ShrinkDemandedConstant(Op, NewMask)) {
862 KnownZero = KnownZeroOut;
863 KnownOne = KnownOneOut;
866 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero,
867 KnownOne, TLO, Depth+1))
869 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero2,
870 KnownOne2, TLO, Depth+1))
872 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
873 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
875 // If the operands are constants, see if we can simplify them.
876 if (TLO.ShrinkDemandedConstant(Op, NewMask))
879 // Only known if known in both the LHS and RHS.
880 KnownOne &= KnownOne2;
881 KnownZero &= KnownZero2;
884 if (SimplifyDemandedBits(Op.getOperand(3), NewMask, KnownZero,
885 KnownOne, TLO, Depth+1))
887 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero2,
888 KnownOne2, TLO, Depth+1))
890 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
891 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
893 // If the operands are constants, see if we can simplify them.
894 if (TLO.ShrinkDemandedConstant(Op, NewMask))
897 // Only known if known in both the LHS and RHS.
898 KnownOne &= KnownOne2;
899 KnownZero &= KnownZero2;
902 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
903 unsigned ShAmt = SA->getZExtValue();
904 SDValue InOp = Op.getOperand(0);
906 // If the shift count is an invalid immediate, don't do anything.
907 if (ShAmt >= BitWidth)
910 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a
911 // single shift. We can do this if the bottom bits (which are shifted
912 // out) are never demanded.
913 if (InOp.getOpcode() == ISD::SRL &&
914 isa<ConstantSDNode>(InOp.getOperand(1))) {
915 if (ShAmt && (NewMask & APInt::getLowBitsSet(BitWidth, ShAmt)) == 0) {
916 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
917 unsigned Opc = ISD::SHL;
925 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
926 MVT VT = Op.getValueType();
927 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, VT,
928 InOp.getOperand(0), NewSA));
932 if (SimplifyDemandedBits(Op.getOperand(0), NewMask.lshr(ShAmt),
933 KnownZero, KnownOne, TLO, Depth+1))
935 KnownZero <<= SA->getZExtValue();
936 KnownOne <<= SA->getZExtValue();
937 // low bits known zero.
938 KnownZero |= APInt::getLowBitsSet(BitWidth, SA->getZExtValue());
942 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
943 MVT VT = Op.getValueType();
944 unsigned ShAmt = SA->getZExtValue();
945 unsigned VTSize = VT.getSizeInBits();
946 SDValue InOp = Op.getOperand(0);
948 // If the shift count is an invalid immediate, don't do anything.
949 if (ShAmt >= BitWidth)
952 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a
953 // single shift. We can do this if the top bits (which are shifted out)
954 // are never demanded.
955 if (InOp.getOpcode() == ISD::SHL &&
956 isa<ConstantSDNode>(InOp.getOperand(1))) {
957 if (ShAmt && (NewMask & APInt::getHighBitsSet(VTSize, ShAmt)) == 0) {
958 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
959 unsigned Opc = ISD::SRL;
967 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
968 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, VT,
969 InOp.getOperand(0), NewSA));
973 // Compute the new bits that are at the top now.
974 if (SimplifyDemandedBits(InOp, (NewMask << ShAmt),
975 KnownZero, KnownOne, TLO, Depth+1))
977 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
978 KnownZero = KnownZero.lshr(ShAmt);
979 KnownOne = KnownOne.lshr(ShAmt);
981 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
982 KnownZero |= HighBits; // High bits known zero.
986 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
987 MVT VT = Op.getValueType();
988 unsigned ShAmt = SA->getZExtValue();
990 // If the shift count is an invalid immediate, don't do anything.
991 if (ShAmt >= BitWidth)
994 APInt InDemandedMask = (NewMask << ShAmt);
996 // If any of the demanded bits are produced by the sign extension, we also
997 // demand the input sign bit.
998 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
999 if (HighBits.intersects(NewMask))
1000 InDemandedMask |= APInt::getSignBit(VT.getSizeInBits());
1002 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask,
1003 KnownZero, KnownOne, TLO, Depth+1))
1005 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1006 KnownZero = KnownZero.lshr(ShAmt);
1007 KnownOne = KnownOne.lshr(ShAmt);
1009 // Handle the sign bit, adjusted to where it is now in the mask.
1010 APInt SignBit = APInt::getSignBit(BitWidth).lshr(ShAmt);
1012 // If the input sign bit is known to be zero, or if none of the top bits
1013 // are demanded, turn this into an unsigned shift right.
1014 if (KnownZero.intersects(SignBit) || (HighBits & ~NewMask) == HighBits) {
1015 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, VT, Op.getOperand(0),
1017 } else if (KnownOne.intersects(SignBit)) { // New bits are known one.
1018 KnownOne |= HighBits;
1022 case ISD::SIGN_EXTEND_INREG: {
1023 MVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1025 // Sign extension. Compute the demanded bits in the result that are not
1026 // present in the input.
1027 APInt NewBits = APInt::getHighBitsSet(BitWidth,
1028 BitWidth - EVT.getSizeInBits()) &
1031 // If none of the extended bits are demanded, eliminate the sextinreg.
1033 return TLO.CombineTo(Op, Op.getOperand(0));
1035 APInt InSignBit = APInt::getSignBit(EVT.getSizeInBits());
1036 InSignBit.zext(BitWidth);
1037 APInt InputDemandedBits = APInt::getLowBitsSet(BitWidth,
1038 EVT.getSizeInBits()) &
1041 // Since the sign extended bits are demanded, we know that the sign
1043 InputDemandedBits |= InSignBit;
1045 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits,
1046 KnownZero, KnownOne, TLO, Depth+1))
1048 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1050 // If the sign bit of the input is known set or clear, then we know the
1051 // top bits of the result.
1053 // If the input sign bit is known zero, convert this into a zero extension.
1054 if (KnownZero.intersects(InSignBit))
1055 return TLO.CombineTo(Op,
1056 TLO.DAG.getZeroExtendInReg(Op.getOperand(0), EVT));
1058 if (KnownOne.intersects(InSignBit)) { // Input sign bit known set
1059 KnownOne |= NewBits;
1060 KnownZero &= ~NewBits;
1061 } else { // Input sign bit unknown
1062 KnownZero &= ~NewBits;
1063 KnownOne &= ~NewBits;
1067 case ISD::ZERO_EXTEND: {
1068 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits();
1069 APInt InMask = NewMask;
1070 InMask.trunc(OperandBitWidth);
1072 // If none of the top bits are demanded, convert this into an any_extend.
1074 APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask;
1075 if (!NewBits.intersects(NewMask))
1076 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND,
1080 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1081 KnownZero, KnownOne, TLO, Depth+1))
1083 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1084 KnownZero.zext(BitWidth);
1085 KnownOne.zext(BitWidth);
1086 KnownZero |= NewBits;
1089 case ISD::SIGN_EXTEND: {
1090 MVT InVT = Op.getOperand(0).getValueType();
1091 unsigned InBits = InVT.getSizeInBits();
1092 APInt InMask = APInt::getLowBitsSet(BitWidth, InBits);
1093 APInt InSignBit = APInt::getBitsSet(BitWidth, InBits - 1, InBits);
1094 APInt NewBits = ~InMask & NewMask;
1096 // If none of the top bits are demanded, convert this into an any_extend.
1098 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND,Op.getValueType(),
1101 // Since some of the sign extended bits are demanded, we know that the sign
1103 APInt InDemandedBits = InMask & NewMask;
1104 InDemandedBits |= InSignBit;
1105 InDemandedBits.trunc(InBits);
1107 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero,
1108 KnownOne, TLO, Depth+1))
1110 KnownZero.zext(BitWidth);
1111 KnownOne.zext(BitWidth);
1113 // If the sign bit is known zero, convert this to a zero extend.
1114 if (KnownZero.intersects(InSignBit))
1115 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND,
1119 // If the sign bit is known one, the top bits match.
1120 if (KnownOne.intersects(InSignBit)) {
1121 KnownOne |= NewBits;
1122 KnownZero &= ~NewBits;
1123 } else { // Otherwise, top bits aren't known.
1124 KnownOne &= ~NewBits;
1125 KnownZero &= ~NewBits;
1129 case ISD::ANY_EXTEND: {
1130 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits();
1131 APInt InMask = NewMask;
1132 InMask.trunc(OperandBitWidth);
1133 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1134 KnownZero, KnownOne, TLO, Depth+1))
1136 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1137 KnownZero.zext(BitWidth);
1138 KnownOne.zext(BitWidth);
1141 case ISD::TRUNCATE: {
1142 // Simplify the input, using demanded bit information, and compute the known
1143 // zero/one bits live out.
1144 APInt TruncMask = NewMask;
1145 TruncMask.zext(Op.getOperand(0).getValueSizeInBits());
1146 if (SimplifyDemandedBits(Op.getOperand(0), TruncMask,
1147 KnownZero, KnownOne, TLO, Depth+1))
1149 KnownZero.trunc(BitWidth);
1150 KnownOne.trunc(BitWidth);
1152 // If the input is only used by this truncate, see if we can shrink it based
1153 // on the known demanded bits.
1154 if (Op.getOperand(0).getNode()->hasOneUse()) {
1155 SDValue In = Op.getOperand(0);
1156 unsigned InBitWidth = In.getValueSizeInBits();
1157 switch (In.getOpcode()) {
1160 // Shrink SRL by a constant if none of the high bits shifted in are
1162 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1))){
1163 APInt HighBits = APInt::getHighBitsSet(InBitWidth,
1164 InBitWidth - BitWidth);
1165 HighBits = HighBits.lshr(ShAmt->getZExtValue());
1166 HighBits.trunc(BitWidth);
1168 if (ShAmt->getZExtValue() < BitWidth && !(HighBits & NewMask)) {
1169 // None of the shifted in bits are needed. Add a truncate of the
1170 // shift input, then shift it.
1171 SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE,
1174 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL,Op.getValueType(),
1175 NewTrunc, In.getOperand(1)));
1182 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1185 case ISD::AssertZext: {
1186 MVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1187 APInt InMask = APInt::getLowBitsSet(BitWidth,
1188 VT.getSizeInBits());
1189 if (SimplifyDemandedBits(Op.getOperand(0), InMask & NewMask,
1190 KnownZero, KnownOne, TLO, Depth+1))
1192 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1193 KnownZero |= ~InMask & NewMask;
1196 case ISD::BIT_CONVERT:
1198 // If this is an FP->Int bitcast and if the sign bit is the only thing that
1199 // is demanded, turn this into a FGETSIGN.
1200 if (NewMask == MVT::getIntegerVTSignBit(Op.getValueType()) &&
1201 MVT::isFloatingPoint(Op.getOperand(0).getValueType()) &&
1202 !MVT::isVector(Op.getOperand(0).getValueType())) {
1203 // Only do this xform if FGETSIGN is valid or if before legalize.
1204 if (!TLO.AfterLegalize ||
1205 isOperationLegal(ISD::FGETSIGN, Op.getValueType())) {
1206 // Make a FGETSIGN + SHL to move the sign bit into the appropriate
1207 // place. We expect the SHL to be eliminated by other optimizations.
1208 SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, Op.getValueType(),
1210 unsigned ShVal = Op.getValueType().getSizeInBits()-1;
1211 SDValue ShAmt = TLO.DAG.getConstant(ShVal, getShiftAmountTy());
1212 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, Op.getValueType(),
1219 // Just use ComputeMaskedBits to compute output bits.
1220 TLO.DAG.ComputeMaskedBits(Op, NewMask, KnownZero, KnownOne, Depth);
1224 // If we know the value of all of the demanded bits, return this as a
1226 if ((NewMask & (KnownZero|KnownOne)) == NewMask)
1227 return TLO.CombineTo(Op, TLO.DAG.getConstant(KnownOne, Op.getValueType()));
1232 /// computeMaskedBitsForTargetNode - Determine which of the bits specified
1233 /// in Mask are known to be either zero or one and return them in the
1234 /// KnownZero/KnownOne bitsets.
1235 void TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
1239 const SelectionDAG &DAG,
1240 unsigned Depth) const {
1241 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1242 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1243 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1244 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1245 "Should use MaskedValueIsZero if you don't know whether Op"
1246 " is a target node!");
1247 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
1250 /// ComputeNumSignBitsForTargetNode - This method can be implemented by
1251 /// targets that want to expose additional information about sign bits to the
1253 unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
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 ComputeNumSignBits if you don't know whether Op"
1260 " is a target node!");
1265 /// SimplifySetCC - Try to simplify a setcc built with the specified operands
1266 /// and cc. If it is unable to simplify it, return a null SDValue.
1268 TargetLowering::SimplifySetCC(MVT VT, SDValue N0, SDValue N1,
1269 ISD::CondCode Cond, bool foldBooleans,
1270 DAGCombinerInfo &DCI) const {
1271 SelectionDAG &DAG = DCI.DAG;
1273 // These setcc operations always fold.
1277 case ISD::SETFALSE2: return DAG.getConstant(0, VT);
1279 case ISD::SETTRUE2: return DAG.getConstant(1, VT);
1282 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) {
1283 const APInt &C1 = N1C->getAPIntValue();
1284 if (isa<ConstantSDNode>(N0.getNode())) {
1285 return DAG.FoldSetCC(VT, N0, N1, Cond);
1287 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an
1288 // equality comparison, then we're just comparing whether X itself is
1290 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) &&
1291 N0.getOperand(0).getOpcode() == ISD::CTLZ &&
1292 N0.getOperand(1).getOpcode() == ISD::Constant) {
1293 unsigned ShAmt = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
1294 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1295 ShAmt == Log2_32(N0.getValueType().getSizeInBits())) {
1296 if ((C1 == 0) == (Cond == ISD::SETEQ)) {
1297 // (srl (ctlz x), 5) == 0 -> X != 0
1298 // (srl (ctlz x), 5) != 1 -> X != 0
1301 // (srl (ctlz x), 5) != 0 -> X == 0
1302 // (srl (ctlz x), 5) == 1 -> X == 0
1305 SDValue Zero = DAG.getConstant(0, N0.getValueType());
1306 return DAG.getSetCC(VT, N0.getOperand(0).getOperand(0),
1311 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
1312 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
1313 unsigned InSize = N0.getOperand(0).getValueType().getSizeInBits();
1315 // If the comparison constant has bits in the upper part, the
1316 // zero-extended value could never match.
1317 if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(),
1318 C1.getBitWidth() - InSize))) {
1322 case ISD::SETEQ: return DAG.getConstant(0, VT);
1325 case ISD::SETNE: return DAG.getConstant(1, VT);
1328 // True if the sign bit of C1 is set.
1329 return DAG.getConstant(C1.isNegative(), VT);
1332 // True if the sign bit of C1 isn't set.
1333 return DAG.getConstant(C1.isNonNegative(), VT);
1339 // Otherwise, we can perform the comparison with the low bits.
1347 return DAG.getSetCC(VT, N0.getOperand(0),
1348 DAG.getConstant(APInt(C1).trunc(InSize),
1349 N0.getOperand(0).getValueType()),
1352 break; // todo, be more careful with signed comparisons
1354 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
1355 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1356 MVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
1357 unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits();
1358 MVT ExtDstTy = N0.getValueType();
1359 unsigned ExtDstTyBits = ExtDstTy.getSizeInBits();
1361 // If the extended part has any inconsistent bits, it cannot ever
1362 // compare equal. In other words, they have to be all ones or all
1365 APInt::getHighBitsSet(ExtDstTyBits, ExtDstTyBits - ExtSrcTyBits);
1366 if ((C1 & ExtBits) != 0 && (C1 & ExtBits) != ExtBits)
1367 return DAG.getConstant(Cond == ISD::SETNE, VT);
1370 MVT Op0Ty = N0.getOperand(0).getValueType();
1371 if (Op0Ty == ExtSrcTy) {
1372 ZextOp = N0.getOperand(0);
1374 APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits);
1375 ZextOp = DAG.getNode(ISD::AND, Op0Ty, N0.getOperand(0),
1376 DAG.getConstant(Imm, Op0Ty));
1378 if (!DCI.isCalledByLegalizer())
1379 DCI.AddToWorklist(ZextOp.getNode());
1380 // Otherwise, make this a use of a zext.
1381 return DAG.getSetCC(VT, ZextOp,
1382 DAG.getConstant(C1 & APInt::getLowBitsSet(
1387 } else if ((N1C->isNullValue() || N1C->getAPIntValue() == 1) &&
1388 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1390 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC
1391 if (N0.getOpcode() == ISD::SETCC) {
1392 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getZExtValue() != 1);
1396 // Invert the condition.
1397 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
1398 CC = ISD::getSetCCInverse(CC,
1399 N0.getOperand(0).getValueType().isInteger());
1400 return DAG.getSetCC(VT, N0.getOperand(0), N0.getOperand(1), CC);
1403 if ((N0.getOpcode() == ISD::XOR ||
1404 (N0.getOpcode() == ISD::AND &&
1405 N0.getOperand(0).getOpcode() == ISD::XOR &&
1406 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
1407 isa<ConstantSDNode>(N0.getOperand(1)) &&
1408 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue() == 1) {
1409 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We
1410 // can only do this if the top bits are known zero.
1411 unsigned BitWidth = N0.getValueSizeInBits();
1412 if (DAG.MaskedValueIsZero(N0,
1413 APInt::getHighBitsSet(BitWidth,
1415 // Okay, get the un-inverted input value.
1417 if (N0.getOpcode() == ISD::XOR)
1418 Val = N0.getOperand(0);
1420 assert(N0.getOpcode() == ISD::AND &&
1421 N0.getOperand(0).getOpcode() == ISD::XOR);
1422 // ((X^1)&1)^1 -> X & 1
1423 Val = DAG.getNode(ISD::AND, N0.getValueType(),
1424 N0.getOperand(0).getOperand(0),
1427 return DAG.getSetCC(VT, Val, N1,
1428 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
1433 APInt MinVal, MaxVal;
1434 unsigned OperandBitSize = N1C->getValueType(0).getSizeInBits();
1435 if (ISD::isSignedIntSetCC(Cond)) {
1436 MinVal = APInt::getSignedMinValue(OperandBitSize);
1437 MaxVal = APInt::getSignedMaxValue(OperandBitSize);
1439 MinVal = APInt::getMinValue(OperandBitSize);
1440 MaxVal = APInt::getMaxValue(OperandBitSize);
1443 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
1444 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
1445 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true
1446 // X >= C0 --> X > (C0-1)
1447 return DAG.getSetCC(VT, N0, DAG.getConstant(C1-1, N1.getValueType()),
1448 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
1451 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
1452 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true
1453 // X <= C0 --> X < (C0+1)
1454 return DAG.getSetCC(VT, N0, DAG.getConstant(C1+1, N1.getValueType()),
1455 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
1458 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
1459 return DAG.getConstant(0, VT); // X < MIN --> false
1460 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal)
1461 return DAG.getConstant(1, VT); // X >= MIN --> true
1462 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal)
1463 return DAG.getConstant(0, VT); // X > MAX --> false
1464 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal)
1465 return DAG.getConstant(1, VT); // X <= MAX --> true
1467 // Canonicalize setgt X, Min --> setne X, Min
1468 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
1469 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
1470 // Canonicalize setlt X, Max --> setne X, Max
1471 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
1472 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
1474 // If we have setult X, 1, turn it into seteq X, 0
1475 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
1476 return DAG.getSetCC(VT, N0, DAG.getConstant(MinVal, N0.getValueType()),
1478 // If we have setugt X, Max-1, turn it into seteq X, Max
1479 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
1480 return DAG.getSetCC(VT, N0, DAG.getConstant(MaxVal, N0.getValueType()),
1483 // If we have "setcc X, C0", check to see if we can shrink the immediate
1486 // SETUGT X, SINTMAX -> SETLT X, 0
1487 if (Cond == ISD::SETUGT && OperandBitSize != 1 &&
1488 C1 == (~0ULL >> (65-OperandBitSize)))
1489 return DAG.getSetCC(VT, N0, DAG.getConstant(0, N1.getValueType()),
1492 // FIXME: Implement the rest of these.
1494 // Fold bit comparisons when we can.
1495 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1496 VT == N0.getValueType() && N0.getOpcode() == ISD::AND)
1497 if (ConstantSDNode *AndRHS =
1498 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1499 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
1500 // Perform the xform if the AND RHS is a single bit.
1501 if (isPowerOf2_64(AndRHS->getZExtValue())) {
1502 return DAG.getNode(ISD::SRL, VT, N0,
1503 DAG.getConstant(Log2_64(AndRHS->getZExtValue()),
1504 getShiftAmountTy()));
1506 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getZExtValue()) {
1507 // (X & 8) == 8 --> (X & 8) >> 3
1508 // Perform the xform if C1 is a single bit.
1509 if (C1.isPowerOf2()) {
1510 return DAG.getNode(ISD::SRL, VT, N0,
1511 DAG.getConstant(C1.logBase2(), getShiftAmountTy()));
1516 } else if (isa<ConstantSDNode>(N0.getNode())) {
1517 // Ensure that the constant occurs on the RHS.
1518 return DAG.getSetCC(VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
1521 if (isa<ConstantFPSDNode>(N0.getNode())) {
1522 // Constant fold or commute setcc.
1523 SDValue O = DAG.FoldSetCC(VT, N0, N1, Cond);
1524 if (O.getNode()) return O;
1525 } else if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1.getNode())) {
1526 // If the RHS of an FP comparison is a constant, simplify it away in
1528 if (CFP->getValueAPF().isNaN()) {
1529 // If an operand is known to be a nan, we can fold it.
1530 switch (ISD::getUnorderedFlavor(Cond)) {
1531 default: assert(0 && "Unknown flavor!");
1532 case 0: // Known false.
1533 return DAG.getConstant(0, VT);
1534 case 1: // Known true.
1535 return DAG.getConstant(1, VT);
1536 case 2: // Undefined.
1537 return DAG.getNode(ISD::UNDEF, VT);
1541 // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the
1542 // constant if knowing that the operand is non-nan is enough. We prefer to
1543 // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to
1545 if (Cond == ISD::SETO || Cond == ISD::SETUO)
1546 return DAG.getSetCC(VT, N0, N0, Cond);
1550 // We can always fold X == X for integer setcc's.
1551 if (N0.getValueType().isInteger())
1552 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
1553 unsigned UOF = ISD::getUnorderedFlavor(Cond);
1554 if (UOF == 2) // FP operators that are undefined on NaNs.
1555 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
1556 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
1557 return DAG.getConstant(UOF, VT);
1558 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
1559 // if it is not already.
1560 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
1561 if (NewCond != Cond)
1562 return DAG.getSetCC(VT, N0, N1, NewCond);
1565 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1566 N0.getValueType().isInteger()) {
1567 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
1568 N0.getOpcode() == ISD::XOR) {
1569 // Simplify (X+Y) == (X+Z) --> Y == Z
1570 if (N0.getOpcode() == N1.getOpcode()) {
1571 if (N0.getOperand(0) == N1.getOperand(0))
1572 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(1), Cond);
1573 if (N0.getOperand(1) == N1.getOperand(1))
1574 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(0), Cond);
1575 if (DAG.isCommutativeBinOp(N0.getOpcode())) {
1576 // If X op Y == Y op X, try other combinations.
1577 if (N0.getOperand(0) == N1.getOperand(1))
1578 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(0), Cond);
1579 if (N0.getOperand(1) == N1.getOperand(0))
1580 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(1), Cond);
1584 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) {
1585 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1586 // Turn (X+C1) == C2 --> X == C2-C1
1587 if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse()) {
1588 return DAG.getSetCC(VT, N0.getOperand(0),
1589 DAG.getConstant(RHSC->getAPIntValue()-
1590 LHSR->getAPIntValue(),
1591 N0.getValueType()), Cond);
1594 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
1595 if (N0.getOpcode() == ISD::XOR)
1596 // If we know that all of the inverted bits are zero, don't bother
1597 // performing the inversion.
1598 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue()))
1600 DAG.getSetCC(VT, N0.getOperand(0),
1601 DAG.getConstant(LHSR->getAPIntValue() ^
1602 RHSC->getAPIntValue(),
1607 // Turn (C1-X) == C2 --> X == C1-C2
1608 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
1609 if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse()) {
1611 DAG.getSetCC(VT, N0.getOperand(1),
1612 DAG.getConstant(SUBC->getAPIntValue() -
1613 RHSC->getAPIntValue(),
1620 // Simplify (X+Z) == X --> Z == 0
1621 if (N0.getOperand(0) == N1)
1622 return DAG.getSetCC(VT, N0.getOperand(1),
1623 DAG.getConstant(0, N0.getValueType()), Cond);
1624 if (N0.getOperand(1) == N1) {
1625 if (DAG.isCommutativeBinOp(N0.getOpcode()))
1626 return DAG.getSetCC(VT, N0.getOperand(0),
1627 DAG.getConstant(0, N0.getValueType()), Cond);
1628 else if (N0.getNode()->hasOneUse()) {
1629 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
1630 // (Z-X) == X --> Z == X<<1
1631 SDValue SH = DAG.getNode(ISD::SHL, N1.getValueType(),
1633 DAG.getConstant(1, getShiftAmountTy()));
1634 if (!DCI.isCalledByLegalizer())
1635 DCI.AddToWorklist(SH.getNode());
1636 return DAG.getSetCC(VT, N0.getOperand(0), SH, Cond);
1641 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
1642 N1.getOpcode() == ISD::XOR) {
1643 // Simplify X == (X+Z) --> Z == 0
1644 if (N1.getOperand(0) == N0) {
1645 return DAG.getSetCC(VT, N1.getOperand(1),
1646 DAG.getConstant(0, N1.getValueType()), Cond);
1647 } else if (N1.getOperand(1) == N0) {
1648 if (DAG.isCommutativeBinOp(N1.getOpcode())) {
1649 return DAG.getSetCC(VT, N1.getOperand(0),
1650 DAG.getConstant(0, N1.getValueType()), Cond);
1651 } else if (N1.getNode()->hasOneUse()) {
1652 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
1653 // X == (Z-X) --> X<<1 == Z
1654 SDValue SH = DAG.getNode(ISD::SHL, N1.getValueType(), N0,
1655 DAG.getConstant(1, getShiftAmountTy()));
1656 if (!DCI.isCalledByLegalizer())
1657 DCI.AddToWorklist(SH.getNode());
1658 return DAG.getSetCC(VT, SH, N1.getOperand(0), Cond);
1664 // Fold away ALL boolean setcc's.
1666 if (N0.getValueType() == MVT::i1 && foldBooleans) {
1668 default: assert(0 && "Unknown integer setcc!");
1669 case ISD::SETEQ: // X == Y -> (X^Y)^1
1670 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
1671 N0 = DAG.getNode(ISD::XOR, MVT::i1, Temp, DAG.getConstant(1, MVT::i1));
1672 if (!DCI.isCalledByLegalizer())
1673 DCI.AddToWorklist(Temp.getNode());
1675 case ISD::SETNE: // X != Y --> (X^Y)
1676 N0 = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
1678 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> X^1 & Y
1679 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> X^1 & Y
1680 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
1681 N0 = DAG.getNode(ISD::AND, MVT::i1, N1, Temp);
1682 if (!DCI.isCalledByLegalizer())
1683 DCI.AddToWorklist(Temp.getNode());
1685 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> Y^1 & X
1686 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> Y^1 & X
1687 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
1688 N0 = DAG.getNode(ISD::AND, MVT::i1, N0, Temp);
1689 if (!DCI.isCalledByLegalizer())
1690 DCI.AddToWorklist(Temp.getNode());
1692 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> X^1 | Y
1693 case ISD::SETGE: // X >=s 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::OR, MVT::i1, N1, Temp);
1696 if (!DCI.isCalledByLegalizer())
1697 DCI.AddToWorklist(Temp.getNode());
1699 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> Y^1 | X
1700 case ISD::SETLE: // X <=s 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::OR, MVT::i1, N0, Temp);
1705 if (VT != MVT::i1) {
1706 if (!DCI.isCalledByLegalizer())
1707 DCI.AddToWorklist(N0.getNode());
1708 // FIXME: If running after legalize, we probably can't do this.
1709 N0 = DAG.getNode(ISD::ZERO_EXTEND, VT, N0);
1714 // Could not fold it.
1718 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
1719 /// node is a GlobalAddress + offset.
1720 bool TargetLowering::isGAPlusOffset(SDNode *N, GlobalValue* &GA,
1721 int64_t &Offset) const {
1722 if (isa<GlobalAddressSDNode>(N)) {
1723 GlobalAddressSDNode *GASD = cast<GlobalAddressSDNode>(N);
1724 GA = GASD->getGlobal();
1725 Offset += GASD->getOffset();
1729 if (N->getOpcode() == ISD::ADD) {
1730 SDValue N1 = N->getOperand(0);
1731 SDValue N2 = N->getOperand(1);
1732 if (isGAPlusOffset(N1.getNode(), GA, Offset)) {
1733 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
1735 Offset += V->getSignExtended();
1738 } else if (isGAPlusOffset(N2.getNode(), GA, Offset)) {
1739 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
1741 Offset += V->getSignExtended();
1750 /// isConsecutiveLoad - Return true if LD (which must be a LoadSDNode) is
1751 /// loading 'Bytes' bytes from a location that is 'Dist' units away from the
1752 /// location that the 'Base' load is loading from.
1753 bool TargetLowering::isConsecutiveLoad(SDNode *LD, SDNode *Base,
1754 unsigned Bytes, int Dist,
1755 const MachineFrameInfo *MFI) const {
1756 if (LD->getOperand(0).getNode() != Base->getOperand(0).getNode())
1758 MVT VT = LD->getValueType(0);
1759 if (VT.getSizeInBits() / 8 != Bytes)
1762 SDValue Loc = LD->getOperand(1);
1763 SDValue BaseLoc = Base->getOperand(1);
1764 if (Loc.getOpcode() == ISD::FrameIndex) {
1765 if (BaseLoc.getOpcode() != ISD::FrameIndex)
1767 int FI = cast<FrameIndexSDNode>(Loc)->getIndex();
1768 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
1769 int FS = MFI->getObjectSize(FI);
1770 int BFS = MFI->getObjectSize(BFI);
1771 if (FS != BFS || FS != (int)Bytes) return false;
1772 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Bytes);
1775 GlobalValue *GV1 = NULL;
1776 GlobalValue *GV2 = NULL;
1777 int64_t Offset1 = 0;
1778 int64_t Offset2 = 0;
1779 bool isGA1 = isGAPlusOffset(Loc.getNode(), GV1, Offset1);
1780 bool isGA2 = isGAPlusOffset(BaseLoc.getNode(), GV2, Offset2);
1781 if (isGA1 && isGA2 && GV1 == GV2)
1782 return Offset1 == (Offset2 + Dist*Bytes);
1787 SDValue TargetLowering::
1788 PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
1789 // Default implementation: no optimization.
1793 //===----------------------------------------------------------------------===//
1794 // Inline Assembler Implementation Methods
1795 //===----------------------------------------------------------------------===//
1798 TargetLowering::ConstraintType
1799 TargetLowering::getConstraintType(const std::string &Constraint) const {
1800 // FIXME: lots more standard ones to handle.
1801 if (Constraint.size() == 1) {
1802 switch (Constraint[0]) {
1804 case 'r': return C_RegisterClass;
1806 case 'o': // offsetable
1807 case 'V': // not offsetable
1809 case 'i': // Simple Integer or Relocatable Constant
1810 case 'n': // Simple Integer
1811 case 's': // Relocatable Constant
1812 case 'X': // Allow ANY value.
1813 case 'I': // Target registers.
1825 if (Constraint.size() > 1 && Constraint[0] == '{' &&
1826 Constraint[Constraint.size()-1] == '}')
1831 /// LowerXConstraint - try to replace an X constraint, which matches anything,
1832 /// with another that has more specific requirements based on the type of the
1833 /// corresponding operand.
1834 const char *TargetLowering::LowerXConstraint(MVT ConstraintVT) const{
1835 if (ConstraintVT.isInteger())
1837 if (ConstraintVT.isFloatingPoint())
1838 return "f"; // works for many targets
1842 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
1843 /// vector. If it is invalid, don't add anything to Ops.
1844 void TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
1845 char ConstraintLetter,
1846 std::vector<SDValue> &Ops,
1847 SelectionDAG &DAG) const {
1848 switch (ConstraintLetter) {
1850 case 'X': // Allows any operand; labels (basic block) use this.
1851 if (Op.getOpcode() == ISD::BasicBlock) {
1856 case 'i': // Simple Integer or Relocatable Constant
1857 case 'n': // Simple Integer
1858 case 's': { // Relocatable Constant
1859 // These operands are interested in values of the form (GV+C), where C may
1860 // be folded in as an offset of GV, or it may be explicitly added. Also, it
1861 // is possible and fine if either GV or C are missing.
1862 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
1863 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
1865 // If we have "(add GV, C)", pull out GV/C
1866 if (Op.getOpcode() == ISD::ADD) {
1867 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
1868 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
1869 if (C == 0 || GA == 0) {
1870 C = dyn_cast<ConstantSDNode>(Op.getOperand(0));
1871 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1));
1873 if (C == 0 || GA == 0)
1877 // If we find a valid operand, map to the TargetXXX version so that the
1878 // value itself doesn't get selected.
1879 if (GA) { // Either &GV or &GV+C
1880 if (ConstraintLetter != 'n') {
1881 int64_t Offs = GA->getOffset();
1882 if (C) Offs += C->getZExtValue();
1883 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(),
1884 Op.getValueType(), Offs));
1888 if (C) { // just C, no GV.
1889 // Simple constants are not allowed for 's'.
1890 if (ConstraintLetter != 's') {
1891 Ops.push_back(DAG.getTargetConstant(C->getAPIntValue(),
1892 Op.getValueType()));
1901 std::vector<unsigned> TargetLowering::
1902 getRegClassForInlineAsmConstraint(const std::string &Constraint,
1904 return std::vector<unsigned>();
1908 std::pair<unsigned, const TargetRegisterClass*> TargetLowering::
1909 getRegForInlineAsmConstraint(const std::string &Constraint,
1911 if (Constraint[0] != '{')
1912 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
1913 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?");
1915 // Remove the braces from around the name.
1916 std::string RegName(Constraint.begin()+1, Constraint.end()-1);
1918 // Figure out which register class contains this reg.
1919 const TargetRegisterInfo *RI = TM.getRegisterInfo();
1920 for (TargetRegisterInfo::regclass_iterator RCI = RI->regclass_begin(),
1921 E = RI->regclass_end(); RCI != E; ++RCI) {
1922 const TargetRegisterClass *RC = *RCI;
1924 // If none of the the value types for this register class are valid, we
1925 // can't use it. For example, 64-bit reg classes on 32-bit targets.
1926 bool isLegal = false;
1927 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
1929 if (isTypeLegal(*I)) {
1935 if (!isLegal) continue;
1937 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end();
1939 if (StringsEqualNoCase(RegName, RI->get(*I).AsmName))
1940 return std::make_pair(*I, RC);
1944 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
1947 //===----------------------------------------------------------------------===//
1948 // Constraint Selection.
1950 /// getConstraintGenerality - Return an integer indicating how general CT
1952 static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) {
1954 default: assert(0 && "Unknown constraint type!");
1955 case TargetLowering::C_Other:
1956 case TargetLowering::C_Unknown:
1958 case TargetLowering::C_Register:
1960 case TargetLowering::C_RegisterClass:
1962 case TargetLowering::C_Memory:
1967 /// ChooseConstraint - If there are multiple different constraints that we
1968 /// could pick for this operand (e.g. "imr") try to pick the 'best' one.
1969 /// This is somewhat tricky: constraints fall into four classes:
1970 /// Other -> immediates and magic values
1971 /// Register -> one specific register
1972 /// RegisterClass -> a group of regs
1973 /// Memory -> memory
1974 /// Ideally, we would pick the most specific constraint possible: if we have
1975 /// something that fits into a register, we would pick it. The problem here
1976 /// is that if we have something that could either be in a register or in
1977 /// memory that use of the register could cause selection of *other*
1978 /// operands to fail: they might only succeed if we pick memory. Because of
1979 /// this the heuristic we use is:
1981 /// 1) If there is an 'other' constraint, and if the operand is valid for
1982 /// that constraint, use it. This makes us take advantage of 'i'
1983 /// constraints when available.
1984 /// 2) Otherwise, pick the most general constraint present. This prefers
1985 /// 'm' over 'r', for example.
1987 static void ChooseConstraint(TargetLowering::AsmOperandInfo &OpInfo,
1988 const TargetLowering &TLI,
1989 SDValue Op, SelectionDAG *DAG) {
1990 assert(OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options");
1991 unsigned BestIdx = 0;
1992 TargetLowering::ConstraintType BestType = TargetLowering::C_Unknown;
1993 int BestGenerality = -1;
1995 // Loop over the options, keeping track of the most general one.
1996 for (unsigned i = 0, e = OpInfo.Codes.size(); i != e; ++i) {
1997 TargetLowering::ConstraintType CType =
1998 TLI.getConstraintType(OpInfo.Codes[i]);
2000 // If this is an 'other' constraint, see if the operand is valid for it.
2001 // For example, on X86 we might have an 'rI' constraint. If the operand
2002 // is an integer in the range [0..31] we want to use I (saving a load
2003 // of a register), otherwise we must use 'r'.
2004 if (CType == TargetLowering::C_Other && Op.getNode()) {
2005 assert(OpInfo.Codes[i].size() == 1 &&
2006 "Unhandled multi-letter 'other' constraint");
2007 std::vector<SDValue> ResultOps;
2008 TLI.LowerAsmOperandForConstraint(Op, OpInfo.Codes[i][0],
2010 if (!ResultOps.empty()) {
2017 // This constraint letter is more general than the previous one, use it.
2018 int Generality = getConstraintGenerality(CType);
2019 if (Generality > BestGenerality) {
2022 BestGenerality = Generality;
2026 OpInfo.ConstraintCode = OpInfo.Codes[BestIdx];
2027 OpInfo.ConstraintType = BestType;
2030 /// ComputeConstraintToUse - Determines the constraint code and constraint
2031 /// type to use for the specific AsmOperandInfo, setting
2032 /// OpInfo.ConstraintCode and OpInfo.ConstraintType.
2033 void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo,
2035 SelectionDAG *DAG) const {
2036 assert(!OpInfo.Codes.empty() && "Must have at least one constraint");
2038 // Single-letter constraints ('r') are very common.
2039 if (OpInfo.Codes.size() == 1) {
2040 OpInfo.ConstraintCode = OpInfo.Codes[0];
2041 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
2043 ChooseConstraint(OpInfo, *this, Op, DAG);
2046 // 'X' matches anything.
2047 if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) {
2048 // Labels and constants are handled elsewhere ('X' is the only thing
2049 // that matches labels).
2050 if (isa<BasicBlock>(OpInfo.CallOperandVal) ||
2051 isa<ConstantInt>(OpInfo.CallOperandVal))
2054 // Otherwise, try to resolve it to something we know about by looking at
2055 // the actual operand type.
2056 if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) {
2057 OpInfo.ConstraintCode = Repl;
2058 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
2063 //===----------------------------------------------------------------------===//
2064 // Loop Strength Reduction hooks
2065 //===----------------------------------------------------------------------===//
2067 /// isLegalAddressingMode - Return true if the addressing mode represented
2068 /// by AM is legal for this target, for a load/store of the specified type.
2069 bool TargetLowering::isLegalAddressingMode(const AddrMode &AM,
2070 const Type *Ty) const {
2071 // The default implementation of this implements a conservative RISCy, r+r and
2074 // Allows a sign-extended 16-bit immediate field.
2075 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
2078 // No global is ever allowed as a base.
2082 // Only support r+r,
2084 case 0: // "r+i" or just "i", depending on HasBaseReg.
2087 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
2089 // Otherwise we have r+r or r+i.
2092 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
2094 // Allow 2*r as r+r.
2101 // Magic for divide replacement
2104 int64_t m; // magic number
2105 int64_t s; // shift amount
2109 uint64_t m; // magic number
2110 int64_t a; // add indicator
2111 int64_t s; // shift amount
2114 /// magic - calculate the magic numbers required to codegen an integer sdiv as
2115 /// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
2117 static ms magic32(int32_t d) {
2119 uint32_t ad, anc, delta, q1, r1, q2, r2, t;
2120 const uint32_t two31 = 0x80000000U;
2124 t = two31 + ((uint32_t)d >> 31);
2125 anc = t - 1 - t%ad; // absolute value of nc
2126 p = 31; // initialize p
2127 q1 = two31/anc; // initialize q1 = 2p/abs(nc)
2128 r1 = two31 - q1*anc; // initialize r1 = rem(2p,abs(nc))
2129 q2 = two31/ad; // initialize q2 = 2p/abs(d)
2130 r2 = two31 - q2*ad; // initialize r2 = rem(2p,abs(d))
2133 q1 = 2*q1; // update q1 = 2p/abs(nc)
2134 r1 = 2*r1; // update r1 = rem(2p/abs(nc))
2135 if (r1 >= anc) { // must be unsigned comparison
2139 q2 = 2*q2; // update q2 = 2p/abs(d)
2140 r2 = 2*r2; // update r2 = rem(2p/abs(d))
2141 if (r2 >= ad) { // must be unsigned comparison
2146 } while (q1 < delta || (q1 == delta && r1 == 0));
2148 mag.m = (int32_t)(q2 + 1); // make sure to sign extend
2149 if (d < 0) mag.m = -mag.m; // resulting magic number
2150 mag.s = p - 32; // resulting shift
2154 /// magicu - calculate the magic numbers required to codegen an integer udiv as
2155 /// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
2156 static mu magicu32(uint32_t d) {
2158 uint32_t nc, delta, q1, r1, q2, r2;
2160 magu.a = 0; // initialize "add" indicator
2162 p = 31; // initialize p
2163 q1 = 0x80000000/nc; // initialize q1 = 2p/nc
2164 r1 = 0x80000000 - q1*nc; // initialize r1 = rem(2p,nc)
2165 q2 = 0x7FFFFFFF/d; // initialize q2 = (2p-1)/d
2166 r2 = 0x7FFFFFFF - q2*d; // initialize r2 = rem((2p-1),d)
2169 if (r1 >= nc - r1 ) {
2170 q1 = 2*q1 + 1; // update q1
2171 r1 = 2*r1 - nc; // update r1
2174 q1 = 2*q1; // update q1
2175 r1 = 2*r1; // update r1
2177 if (r2 + 1 >= d - r2) {
2178 if (q2 >= 0x7FFFFFFF) magu.a = 1;
2179 q2 = 2*q2 + 1; // update q2
2180 r2 = 2*r2 + 1 - d; // update r2
2183 if (q2 >= 0x80000000) magu.a = 1;
2184 q2 = 2*q2; // update q2
2185 r2 = 2*r2 + 1; // update r2
2188 } while (p < 64 && (q1 < delta || (q1 == delta && r1 == 0)));
2189 magu.m = q2 + 1; // resulting magic number
2190 magu.s = p - 32; // resulting shift
2194 /// magic - calculate the magic numbers required to codegen an integer sdiv as
2195 /// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
2197 static ms magic64(int64_t d) {
2199 uint64_t ad, anc, delta, q1, r1, q2, r2, t;
2200 const uint64_t two63 = 9223372036854775808ULL; // 2^63
2203 ad = d >= 0 ? d : -d;
2204 t = two63 + ((uint64_t)d >> 63);
2205 anc = t - 1 - t%ad; // absolute value of nc
2206 p = 63; // initialize p
2207 q1 = two63/anc; // initialize q1 = 2p/abs(nc)
2208 r1 = two63 - q1*anc; // initialize r1 = rem(2p,abs(nc))
2209 q2 = two63/ad; // initialize q2 = 2p/abs(d)
2210 r2 = two63 - q2*ad; // initialize r2 = rem(2p,abs(d))
2213 q1 = 2*q1; // update q1 = 2p/abs(nc)
2214 r1 = 2*r1; // update r1 = rem(2p/abs(nc))
2215 if (r1 >= anc) { // must be unsigned comparison
2219 q2 = 2*q2; // update q2 = 2p/abs(d)
2220 r2 = 2*r2; // update r2 = rem(2p/abs(d))
2221 if (r2 >= ad) { // must be unsigned comparison
2226 } while (q1 < delta || (q1 == delta && r1 == 0));
2229 if (d < 0) mag.m = -mag.m; // resulting magic number
2230 mag.s = p - 64; // resulting shift
2234 /// magicu - calculate the magic numbers required to codegen an integer udiv as
2235 /// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
2236 static mu magicu64(uint64_t d)
2239 uint64_t nc, delta, q1, r1, q2, r2;
2241 magu.a = 0; // initialize "add" indicator
2243 p = 63; // initialize p
2244 q1 = 0x8000000000000000ull/nc; // initialize q1 = 2p/nc
2245 r1 = 0x8000000000000000ull - q1*nc; // initialize r1 = rem(2p,nc)
2246 q2 = 0x7FFFFFFFFFFFFFFFull/d; // initialize q2 = (2p-1)/d
2247 r2 = 0x7FFFFFFFFFFFFFFFull - q2*d; // initialize r2 = rem((2p-1),d)
2250 if (r1 >= nc - r1 ) {
2251 q1 = 2*q1 + 1; // update q1
2252 r1 = 2*r1 - nc; // update r1
2255 q1 = 2*q1; // update q1
2256 r1 = 2*r1; // update r1
2258 if (r2 + 1 >= d - r2) {
2259 if (q2 >= 0x7FFFFFFFFFFFFFFFull) magu.a = 1;
2260 q2 = 2*q2 + 1; // update q2
2261 r2 = 2*r2 + 1 - d; // update r2
2264 if (q2 >= 0x8000000000000000ull) magu.a = 1;
2265 q2 = 2*q2; // update q2
2266 r2 = 2*r2 + 1; // update r2
2269 } while (p < 128 && (q1 < delta || (q1 == delta && r1 == 0)));
2270 magu.m = q2 + 1; // resulting magic number
2271 magu.s = p - 64; // resulting shift
2275 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
2276 /// return a DAG expression to select that will generate the same value by
2277 /// multiplying by a magic number. See:
2278 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
2279 SDValue TargetLowering::BuildSDIV(SDNode *N, SelectionDAG &DAG,
2280 std::vector<SDNode*>* Created) const {
2281 MVT VT = N->getValueType(0);
2283 // Check to see if we can do this.
2284 if (!isTypeLegal(VT) || (VT != MVT::i32 && VT != MVT::i64))
2285 return SDValue(); // BuildSDIV only operates on i32 or i64
2287 int64_t d = cast<ConstantSDNode>(N->getOperand(1))->getSignExtended();
2288 ms magics = (VT == MVT::i32) ? magic32(d) : magic64(d);
2290 // Multiply the numerator (operand 0) by the magic value
2292 if (isOperationLegal(ISD::MULHS, VT))
2293 Q = DAG.getNode(ISD::MULHS, VT, N->getOperand(0),
2294 DAG.getConstant(magics.m, VT));
2295 else if (isOperationLegal(ISD::SMUL_LOHI, VT))
2296 Q = SDValue(DAG.getNode(ISD::SMUL_LOHI, DAG.getVTList(VT, VT),
2298 DAG.getConstant(magics.m, VT)).getNode(), 1);
2300 return SDValue(); // No mulhs or equvialent
2301 // If d > 0 and m < 0, add the numerator
2302 if (d > 0 && magics.m < 0) {
2303 Q = DAG.getNode(ISD::ADD, VT, Q, N->getOperand(0));
2305 Created->push_back(Q.getNode());
2307 // If d < 0 and m > 0, subtract the numerator.
2308 if (d < 0 && magics.m > 0) {
2309 Q = DAG.getNode(ISD::SUB, VT, Q, N->getOperand(0));
2311 Created->push_back(Q.getNode());
2313 // Shift right algebraic if shift value is nonzero
2315 Q = DAG.getNode(ISD::SRA, VT, Q,
2316 DAG.getConstant(magics.s, getShiftAmountTy()));
2318 Created->push_back(Q.getNode());
2320 // Extract the sign bit and add it to the quotient
2322 DAG.getNode(ISD::SRL, VT, Q, DAG.getConstant(VT.getSizeInBits()-1,
2323 getShiftAmountTy()));
2325 Created->push_back(T.getNode());
2326 return DAG.getNode(ISD::ADD, VT, Q, T);
2329 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
2330 /// return a DAG expression to select that will generate the same value by
2331 /// multiplying by a magic number. See:
2332 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
2333 SDValue TargetLowering::BuildUDIV(SDNode *N, SelectionDAG &DAG,
2334 std::vector<SDNode*>* Created) const {
2335 MVT VT = N->getValueType(0);
2337 // Check to see if we can do this.
2338 if (!isTypeLegal(VT) || (VT != MVT::i32 && VT != MVT::i64))
2339 return SDValue(); // BuildUDIV only operates on i32 or i64
2341 uint64_t d = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
2342 mu magics = (VT == MVT::i32) ? magicu32(d) : magicu64(d);
2344 // Multiply the numerator (operand 0) by the magic value
2346 if (isOperationLegal(ISD::MULHU, VT))
2347 Q = DAG.getNode(ISD::MULHU, VT, N->getOperand(0),
2348 DAG.getConstant(magics.m, VT));
2349 else if (isOperationLegal(ISD::UMUL_LOHI, VT))
2350 Q = SDValue(DAG.getNode(ISD::UMUL_LOHI, DAG.getVTList(VT, VT),
2352 DAG.getConstant(magics.m, VT)).getNode(), 1);
2354 return SDValue(); // No mulhu or equvialent
2356 Created->push_back(Q.getNode());
2358 if (magics.a == 0) {
2359 return DAG.getNode(ISD::SRL, VT, Q,
2360 DAG.getConstant(magics.s, getShiftAmountTy()));
2362 SDValue NPQ = DAG.getNode(ISD::SUB, VT, N->getOperand(0), Q);
2364 Created->push_back(NPQ.getNode());
2365 NPQ = DAG.getNode(ISD::SRL, VT, NPQ,
2366 DAG.getConstant(1, getShiftAmountTy()));
2368 Created->push_back(NPQ.getNode());
2369 NPQ = DAG.getNode(ISD::ADD, VT, NPQ, Q);
2371 Created->push_back(NPQ.getNode());
2372 return DAG.getNode(ISD::SRL, VT, NPQ,
2373 DAG.getConstant(magics.s-1, getShiftAmountTy()));