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());
660 TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
661 // Assume that everything is safe in static mode.
662 if (getTargetMachine().getRelocationModel() == Reloc::Static)
665 // In dynamic-no-pic mode, assume that known defined values are safe.
666 if (getTargetMachine().getRelocationModel() == Reloc::DynamicNoPIC &&
668 !GA->getGlobal()->isDeclaration() &&
669 !GA->getGlobal()->mayBeOverridden())
672 // Otherwise assume nothing is safe.
676 //===----------------------------------------------------------------------===//
677 // Optimization Methods
678 //===----------------------------------------------------------------------===//
680 /// ShrinkDemandedConstant - Check to see if the specified operand of the
681 /// specified instruction is a constant integer. If so, check to see if there
682 /// are any bits set in the constant that are not demanded. If so, shrink the
683 /// constant and return true.
684 bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDValue Op,
685 const APInt &Demanded) {
686 // FIXME: ISD::SELECT, ISD::SELECT_CC
687 switch(Op.getOpcode()) {
692 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1)))
693 if (C->getAPIntValue().intersects(~Demanded)) {
694 MVT VT = Op.getValueType();
695 SDValue New = DAG.getNode(Op.getOpcode(), VT, Op.getOperand(0),
696 DAG.getConstant(Demanded &
699 return CombineTo(Op, New);
706 /// SimplifyDemandedBits - Look at Op. At this point, we know that only the
707 /// DemandedMask bits of the result of Op are ever used downstream. If we can
708 /// use this information to simplify Op, create a new simplified DAG node and
709 /// return true, returning the original and new nodes in Old and New. Otherwise,
710 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
711 /// the expression (used to simplify the caller). The KnownZero/One bits may
712 /// only be accurate for those bits in the DemandedMask.
713 bool TargetLowering::SimplifyDemandedBits(SDValue Op,
714 const APInt &DemandedMask,
717 TargetLoweringOpt &TLO,
718 unsigned Depth) const {
719 unsigned BitWidth = DemandedMask.getBitWidth();
720 assert(Op.getValueSizeInBits() == BitWidth &&
721 "Mask size mismatches value type size!");
722 APInt NewMask = DemandedMask;
724 // Don't know anything.
725 KnownZero = KnownOne = APInt(BitWidth, 0);
727 // Other users may use these bits.
728 if (!Op.getNode()->hasOneUse()) {
730 // If not at the root, Just compute the KnownZero/KnownOne bits to
731 // simplify things downstream.
732 TLO.DAG.ComputeMaskedBits(Op, DemandedMask, KnownZero, KnownOne, Depth);
735 // If this is the root being simplified, allow it to have multiple uses,
736 // just set the NewMask to all bits.
737 NewMask = APInt::getAllOnesValue(BitWidth);
738 } else if (DemandedMask == 0) {
739 // Not demanding any bits from Op.
740 if (Op.getOpcode() != ISD::UNDEF)
741 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::UNDEF, Op.getValueType()));
743 } else if (Depth == 6) { // Limit search depth.
747 APInt KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut;
748 switch (Op.getOpcode()) {
750 // We know all of the bits for a constant!
751 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue() & NewMask;
752 KnownZero = ~KnownOne & NewMask;
753 return false; // Don't fall through, will infinitely loop.
755 // If the RHS is a constant, check to see if the LHS would be zero without
756 // using the bits from the RHS. Below, we use knowledge about the RHS to
757 // simplify the LHS, here we're using information from the LHS to simplify
759 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
760 APInt LHSZero, LHSOne;
761 TLO.DAG.ComputeMaskedBits(Op.getOperand(0), NewMask,
762 LHSZero, LHSOne, Depth+1);
763 // If the LHS already has zeros where RHSC does, this and is dead.
764 if ((LHSZero & NewMask) == (~RHSC->getAPIntValue() & NewMask))
765 return TLO.CombineTo(Op, Op.getOperand(0));
766 // If any of the set bits in the RHS are known zero on the LHS, shrink
768 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & NewMask))
772 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
773 KnownOne, TLO, Depth+1))
775 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
776 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownZero & NewMask,
777 KnownZero2, KnownOne2, TLO, Depth+1))
779 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
781 // If all of the demanded bits are known one on one side, return the other.
782 // These bits cannot contribute to the result of the 'and'.
783 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
784 return TLO.CombineTo(Op, Op.getOperand(0));
785 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
786 return TLO.CombineTo(Op, Op.getOperand(1));
787 // If all of the demanded bits in the inputs are known zeros, return zero.
788 if ((NewMask & (KnownZero|KnownZero2)) == NewMask)
789 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, Op.getValueType()));
790 // If the RHS is a constant, see if we can simplify it.
791 if (TLO.ShrinkDemandedConstant(Op, ~KnownZero2 & NewMask))
794 // Output known-1 bits are only known if set in both the LHS & RHS.
795 KnownOne &= KnownOne2;
796 // Output known-0 are known to be clear if zero in either the LHS | RHS.
797 KnownZero |= KnownZero2;
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), ~KnownOne & NewMask,
805 KnownZero2, 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 'or'.
811 if ((NewMask & ~KnownOne2 & KnownZero) == (~KnownOne2 & NewMask))
812 return TLO.CombineTo(Op, Op.getOperand(0));
813 if ((NewMask & ~KnownOne & KnownZero2) == (~KnownOne & NewMask))
814 return TLO.CombineTo(Op, Op.getOperand(1));
815 // If all of the potentially set bits on one side are known to be set on
816 // the other side, just use the 'other' side.
817 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
818 return TLO.CombineTo(Op, Op.getOperand(0));
819 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
820 return TLO.CombineTo(Op, Op.getOperand(1));
821 // If the RHS is a constant, see if we can simplify it.
822 if (TLO.ShrinkDemandedConstant(Op, NewMask))
825 // Output known-0 bits are only known if clear in both the LHS & RHS.
826 KnownZero &= KnownZero2;
827 // Output known-1 are known to be set if set in either the LHS | RHS.
828 KnownOne |= KnownOne2;
831 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
832 KnownOne, TLO, Depth+1))
834 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
835 if (SimplifyDemandedBits(Op.getOperand(0), NewMask, KnownZero2,
836 KnownOne2, TLO, Depth+1))
838 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
840 // If all of the demanded bits are known zero on one side, return the other.
841 // These bits cannot contribute to the result of the 'xor'.
842 if ((KnownZero & NewMask) == NewMask)
843 return TLO.CombineTo(Op, Op.getOperand(0));
844 if ((KnownZero2 & NewMask) == NewMask)
845 return TLO.CombineTo(Op, Op.getOperand(1));
847 // If all of the unknown bits are known to be zero on one side or the other
848 // (but not both) turn this into an *inclusive* or.
849 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
850 if ((NewMask & ~KnownZero & ~KnownZero2) == 0)
851 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, Op.getValueType(),
855 // Output known-0 bits are known if clear or set in both the LHS & RHS.
856 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
857 // Output known-1 are known to be set if set in only one of the LHS, RHS.
858 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
860 // If all of the demanded bits on one side are known, and all of the set
861 // bits on that side are also known to be set on the other side, turn this
862 // into an AND, as we know the bits will be cleared.
863 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
864 if ((NewMask & (KnownZero|KnownOne)) == NewMask) { // all known
865 if ((KnownOne & KnownOne2) == KnownOne) {
866 MVT VT = Op.getValueType();
867 SDValue ANDC = TLO.DAG.getConstant(~KnownOne & NewMask, VT);
868 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, VT, Op.getOperand(0),
873 // If the RHS is a constant, see if we can simplify it.
874 // for XOR, we prefer to force bits to 1 if they will make a -1.
875 // if we can't force bits, try to shrink constant
876 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
877 APInt Expanded = C->getAPIntValue() | (~NewMask);
878 // if we can expand it to have all bits set, do it
879 if (Expanded.isAllOnesValue()) {
880 if (Expanded != C->getAPIntValue()) {
881 MVT VT = Op.getValueType();
882 SDValue New = TLO.DAG.getNode(Op.getOpcode(), VT, Op.getOperand(0),
883 TLO.DAG.getConstant(Expanded, VT));
884 return TLO.CombineTo(Op, New);
886 // if it already has all the bits set, nothing to change
887 // but don't shrink either!
888 } else if (TLO.ShrinkDemandedConstant(Op, NewMask)) {
893 KnownZero = KnownZeroOut;
894 KnownOne = KnownOneOut;
897 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero,
898 KnownOne, TLO, Depth+1))
900 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero2,
901 KnownOne2, TLO, Depth+1))
903 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
904 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
906 // If the operands are constants, see if we can simplify them.
907 if (TLO.ShrinkDemandedConstant(Op, NewMask))
910 // Only known if known in both the LHS and RHS.
911 KnownOne &= KnownOne2;
912 KnownZero &= KnownZero2;
915 if (SimplifyDemandedBits(Op.getOperand(3), NewMask, KnownZero,
916 KnownOne, TLO, Depth+1))
918 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero2,
919 KnownOne2, TLO, Depth+1))
921 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
922 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
924 // If the operands are constants, see if we can simplify them.
925 if (TLO.ShrinkDemandedConstant(Op, NewMask))
928 // Only known if known in both the LHS and RHS.
929 KnownOne &= KnownOne2;
930 KnownZero &= KnownZero2;
933 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
934 unsigned ShAmt = SA->getZExtValue();
935 SDValue InOp = Op.getOperand(0);
937 // If the shift count is an invalid immediate, don't do anything.
938 if (ShAmt >= BitWidth)
941 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a
942 // single shift. We can do this if the bottom bits (which are shifted
943 // out) are never demanded.
944 if (InOp.getOpcode() == ISD::SRL &&
945 isa<ConstantSDNode>(InOp.getOperand(1))) {
946 if (ShAmt && (NewMask & APInt::getLowBitsSet(BitWidth, ShAmt)) == 0) {
947 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
948 unsigned Opc = ISD::SHL;
956 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
957 MVT VT = Op.getValueType();
958 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, VT,
959 InOp.getOperand(0), NewSA));
963 if (SimplifyDemandedBits(Op.getOperand(0), NewMask.lshr(ShAmt),
964 KnownZero, KnownOne, TLO, Depth+1))
966 KnownZero <<= SA->getZExtValue();
967 KnownOne <<= SA->getZExtValue();
968 // low bits known zero.
969 KnownZero |= APInt::getLowBitsSet(BitWidth, SA->getZExtValue());
973 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
974 MVT VT = Op.getValueType();
975 unsigned ShAmt = SA->getZExtValue();
976 unsigned VTSize = VT.getSizeInBits();
977 SDValue InOp = Op.getOperand(0);
979 // If the shift count is an invalid immediate, don't do anything.
980 if (ShAmt >= BitWidth)
983 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a
984 // single shift. We can do this if the top bits (which are shifted out)
985 // are never demanded.
986 if (InOp.getOpcode() == ISD::SHL &&
987 isa<ConstantSDNode>(InOp.getOperand(1))) {
988 if (ShAmt && (NewMask & APInt::getHighBitsSet(VTSize, ShAmt)) == 0) {
989 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
990 unsigned Opc = ISD::SRL;
998 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
999 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, VT,
1000 InOp.getOperand(0), NewSA));
1004 // Compute the new bits that are at the top now.
1005 if (SimplifyDemandedBits(InOp, (NewMask << ShAmt),
1006 KnownZero, KnownOne, TLO, Depth+1))
1008 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1009 KnownZero = KnownZero.lshr(ShAmt);
1010 KnownOne = KnownOne.lshr(ShAmt);
1012 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
1013 KnownZero |= HighBits; // High bits known zero.
1017 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1018 MVT VT = Op.getValueType();
1019 unsigned ShAmt = SA->getZExtValue();
1021 // If the shift count is an invalid immediate, don't do anything.
1022 if (ShAmt >= BitWidth)
1025 APInt InDemandedMask = (NewMask << ShAmt);
1027 // If any of the demanded bits are produced by the sign extension, we also
1028 // demand the input sign bit.
1029 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
1030 if (HighBits.intersects(NewMask))
1031 InDemandedMask |= APInt::getSignBit(VT.getSizeInBits());
1033 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask,
1034 KnownZero, KnownOne, TLO, Depth+1))
1036 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1037 KnownZero = KnownZero.lshr(ShAmt);
1038 KnownOne = KnownOne.lshr(ShAmt);
1040 // Handle the sign bit, adjusted to where it is now in the mask.
1041 APInt SignBit = APInt::getSignBit(BitWidth).lshr(ShAmt);
1043 // If the input sign bit is known to be zero, or if none of the top bits
1044 // are demanded, turn this into an unsigned shift right.
1045 if (KnownZero.intersects(SignBit) || (HighBits & ~NewMask) == HighBits) {
1046 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, VT, Op.getOperand(0),
1048 } else if (KnownOne.intersects(SignBit)) { // New bits are known one.
1049 KnownOne |= HighBits;
1053 case ISD::SIGN_EXTEND_INREG: {
1054 MVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1056 // Sign extension. Compute the demanded bits in the result that are not
1057 // present in the input.
1058 APInt NewBits = APInt::getHighBitsSet(BitWidth,
1059 BitWidth - EVT.getSizeInBits()) &
1062 // If none of the extended bits are demanded, eliminate the sextinreg.
1064 return TLO.CombineTo(Op, Op.getOperand(0));
1066 APInt InSignBit = APInt::getSignBit(EVT.getSizeInBits());
1067 InSignBit.zext(BitWidth);
1068 APInt InputDemandedBits = APInt::getLowBitsSet(BitWidth,
1069 EVT.getSizeInBits()) &
1072 // Since the sign extended bits are demanded, we know that the sign
1074 InputDemandedBits |= InSignBit;
1076 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits,
1077 KnownZero, KnownOne, TLO, Depth+1))
1079 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1081 // If the sign bit of the input is known set or clear, then we know the
1082 // top bits of the result.
1084 // If the input sign bit is known zero, convert this into a zero extension.
1085 if (KnownZero.intersects(InSignBit))
1086 return TLO.CombineTo(Op,
1087 TLO.DAG.getZeroExtendInReg(Op.getOperand(0), EVT));
1089 if (KnownOne.intersects(InSignBit)) { // Input sign bit known set
1090 KnownOne |= NewBits;
1091 KnownZero &= ~NewBits;
1092 } else { // Input sign bit unknown
1093 KnownZero &= ~NewBits;
1094 KnownOne &= ~NewBits;
1098 case ISD::ZERO_EXTEND: {
1099 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits();
1100 APInt InMask = NewMask;
1101 InMask.trunc(OperandBitWidth);
1103 // If none of the top bits are demanded, convert this into an any_extend.
1105 APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask;
1106 if (!NewBits.intersects(NewMask))
1107 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND,
1111 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1112 KnownZero, KnownOne, TLO, Depth+1))
1114 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1115 KnownZero.zext(BitWidth);
1116 KnownOne.zext(BitWidth);
1117 KnownZero |= NewBits;
1120 case ISD::SIGN_EXTEND: {
1121 MVT InVT = Op.getOperand(0).getValueType();
1122 unsigned InBits = InVT.getSizeInBits();
1123 APInt InMask = APInt::getLowBitsSet(BitWidth, InBits);
1124 APInt InSignBit = APInt::getBitsSet(BitWidth, InBits - 1, InBits);
1125 APInt NewBits = ~InMask & NewMask;
1127 // If none of the top bits are demanded, convert this into an any_extend.
1129 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND,Op.getValueType(),
1132 // Since some of the sign extended bits are demanded, we know that the sign
1134 APInt InDemandedBits = InMask & NewMask;
1135 InDemandedBits |= InSignBit;
1136 InDemandedBits.trunc(InBits);
1138 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero,
1139 KnownOne, TLO, Depth+1))
1141 KnownZero.zext(BitWidth);
1142 KnownOne.zext(BitWidth);
1144 // If the sign bit is known zero, convert this to a zero extend.
1145 if (KnownZero.intersects(InSignBit))
1146 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND,
1150 // If the sign bit is known one, the top bits match.
1151 if (KnownOne.intersects(InSignBit)) {
1152 KnownOne |= NewBits;
1153 KnownZero &= ~NewBits;
1154 } else { // Otherwise, top bits aren't known.
1155 KnownOne &= ~NewBits;
1156 KnownZero &= ~NewBits;
1160 case ISD::ANY_EXTEND: {
1161 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits();
1162 APInt InMask = NewMask;
1163 InMask.trunc(OperandBitWidth);
1164 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1165 KnownZero, KnownOne, TLO, Depth+1))
1167 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1168 KnownZero.zext(BitWidth);
1169 KnownOne.zext(BitWidth);
1172 case ISD::TRUNCATE: {
1173 // Simplify the input, using demanded bit information, and compute the known
1174 // zero/one bits live out.
1175 APInt TruncMask = NewMask;
1176 TruncMask.zext(Op.getOperand(0).getValueSizeInBits());
1177 if (SimplifyDemandedBits(Op.getOperand(0), TruncMask,
1178 KnownZero, KnownOne, TLO, Depth+1))
1180 KnownZero.trunc(BitWidth);
1181 KnownOne.trunc(BitWidth);
1183 // If the input is only used by this truncate, see if we can shrink it based
1184 // on the known demanded bits.
1185 if (Op.getOperand(0).getNode()->hasOneUse()) {
1186 SDValue In = Op.getOperand(0);
1187 unsigned InBitWidth = In.getValueSizeInBits();
1188 switch (In.getOpcode()) {
1191 // Shrink SRL by a constant if none of the high bits shifted in are
1193 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1))){
1194 APInt HighBits = APInt::getHighBitsSet(InBitWidth,
1195 InBitWidth - BitWidth);
1196 HighBits = HighBits.lshr(ShAmt->getZExtValue());
1197 HighBits.trunc(BitWidth);
1199 if (ShAmt->getZExtValue() < BitWidth && !(HighBits & NewMask)) {
1200 // None of the shifted in bits are needed. Add a truncate of the
1201 // shift input, then shift it.
1202 SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE,
1205 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL,Op.getValueType(),
1206 NewTrunc, In.getOperand(1)));
1213 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1216 case ISD::AssertZext: {
1217 MVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1218 APInt InMask = APInt::getLowBitsSet(BitWidth,
1219 VT.getSizeInBits());
1220 if (SimplifyDemandedBits(Op.getOperand(0), InMask & NewMask,
1221 KnownZero, KnownOne, TLO, Depth+1))
1223 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1224 KnownZero |= ~InMask & NewMask;
1227 case ISD::BIT_CONVERT:
1229 // If this is an FP->Int bitcast and if the sign bit is the only thing that
1230 // is demanded, turn this into a FGETSIGN.
1231 if (NewMask == MVT::getIntegerVTSignBit(Op.getValueType()) &&
1232 MVT::isFloatingPoint(Op.getOperand(0).getValueType()) &&
1233 !MVT::isVector(Op.getOperand(0).getValueType())) {
1234 // Only do this xform if FGETSIGN is valid or if before legalize.
1235 if (!TLO.AfterLegalize ||
1236 isOperationLegal(ISD::FGETSIGN, Op.getValueType())) {
1237 // Make a FGETSIGN + SHL to move the sign bit into the appropriate
1238 // place. We expect the SHL to be eliminated by other optimizations.
1239 SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, Op.getValueType(),
1241 unsigned ShVal = Op.getValueType().getSizeInBits()-1;
1242 SDValue ShAmt = TLO.DAG.getConstant(ShVal, getShiftAmountTy());
1243 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, Op.getValueType(),
1250 // Just use ComputeMaskedBits to compute output bits.
1251 TLO.DAG.ComputeMaskedBits(Op, NewMask, KnownZero, KnownOne, Depth);
1255 // If we know the value of all of the demanded bits, return this as a
1257 if ((NewMask & (KnownZero|KnownOne)) == NewMask)
1258 return TLO.CombineTo(Op, TLO.DAG.getConstant(KnownOne, Op.getValueType()));
1263 /// computeMaskedBitsForTargetNode - Determine which of the bits specified
1264 /// in Mask are known to be either zero or one and return them in the
1265 /// KnownZero/KnownOne bitsets.
1266 void TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
1270 const SelectionDAG &DAG,
1271 unsigned Depth) const {
1272 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1273 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1274 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1275 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1276 "Should use MaskedValueIsZero if you don't know whether Op"
1277 " is a target node!");
1278 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
1281 /// ComputeNumSignBitsForTargetNode - This method can be implemented by
1282 /// targets that want to expose additional information about sign bits to the
1284 unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
1285 unsigned Depth) const {
1286 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1287 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1288 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1289 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1290 "Should use ComputeNumSignBits if you don't know whether Op"
1291 " is a target node!");
1296 /// SimplifySetCC - Try to simplify a setcc built with the specified operands
1297 /// and cc. If it is unable to simplify it, return a null SDValue.
1299 TargetLowering::SimplifySetCC(MVT VT, SDValue N0, SDValue N1,
1300 ISD::CondCode Cond, bool foldBooleans,
1301 DAGCombinerInfo &DCI) const {
1302 SelectionDAG &DAG = DCI.DAG;
1304 // These setcc operations always fold.
1308 case ISD::SETFALSE2: return DAG.getConstant(0, VT);
1310 case ISD::SETTRUE2: return DAG.getConstant(1, VT);
1313 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) {
1314 const APInt &C1 = N1C->getAPIntValue();
1315 if (isa<ConstantSDNode>(N0.getNode())) {
1316 return DAG.FoldSetCC(VT, N0, N1, Cond);
1318 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an
1319 // equality comparison, then we're just comparing whether X itself is
1321 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) &&
1322 N0.getOperand(0).getOpcode() == ISD::CTLZ &&
1323 N0.getOperand(1).getOpcode() == ISD::Constant) {
1324 unsigned ShAmt = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
1325 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1326 ShAmt == Log2_32(N0.getValueType().getSizeInBits())) {
1327 if ((C1 == 0) == (Cond == ISD::SETEQ)) {
1328 // (srl (ctlz x), 5) == 0 -> X != 0
1329 // (srl (ctlz x), 5) != 1 -> X != 0
1332 // (srl (ctlz x), 5) != 0 -> X == 0
1333 // (srl (ctlz x), 5) == 1 -> X == 0
1336 SDValue Zero = DAG.getConstant(0, N0.getValueType());
1337 return DAG.getSetCC(VT, N0.getOperand(0).getOperand(0),
1342 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
1343 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
1344 unsigned InSize = N0.getOperand(0).getValueType().getSizeInBits();
1346 // If the comparison constant has bits in the upper part, the
1347 // zero-extended value could never match.
1348 if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(),
1349 C1.getBitWidth() - InSize))) {
1353 case ISD::SETEQ: return DAG.getConstant(0, VT);
1356 case ISD::SETNE: return DAG.getConstant(1, VT);
1359 // True if the sign bit of C1 is set.
1360 return DAG.getConstant(C1.isNegative(), VT);
1363 // True if the sign bit of C1 isn't set.
1364 return DAG.getConstant(C1.isNonNegative(), VT);
1370 // Otherwise, we can perform the comparison with the low bits.
1378 return DAG.getSetCC(VT, N0.getOperand(0),
1379 DAG.getConstant(APInt(C1).trunc(InSize),
1380 N0.getOperand(0).getValueType()),
1383 break; // todo, be more careful with signed comparisons
1385 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
1386 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1387 MVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
1388 unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits();
1389 MVT ExtDstTy = N0.getValueType();
1390 unsigned ExtDstTyBits = ExtDstTy.getSizeInBits();
1392 // If the extended part has any inconsistent bits, it cannot ever
1393 // compare equal. In other words, they have to be all ones or all
1396 APInt::getHighBitsSet(ExtDstTyBits, ExtDstTyBits - ExtSrcTyBits);
1397 if ((C1 & ExtBits) != 0 && (C1 & ExtBits) != ExtBits)
1398 return DAG.getConstant(Cond == ISD::SETNE, VT);
1401 MVT Op0Ty = N0.getOperand(0).getValueType();
1402 if (Op0Ty == ExtSrcTy) {
1403 ZextOp = N0.getOperand(0);
1405 APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits);
1406 ZextOp = DAG.getNode(ISD::AND, Op0Ty, N0.getOperand(0),
1407 DAG.getConstant(Imm, Op0Ty));
1409 if (!DCI.isCalledByLegalizer())
1410 DCI.AddToWorklist(ZextOp.getNode());
1411 // Otherwise, make this a use of a zext.
1412 return DAG.getSetCC(VT, ZextOp,
1413 DAG.getConstant(C1 & APInt::getLowBitsSet(
1418 } else if ((N1C->isNullValue() || N1C->getAPIntValue() == 1) &&
1419 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1421 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC
1422 if (N0.getOpcode() == ISD::SETCC) {
1423 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getZExtValue() != 1);
1427 // Invert the condition.
1428 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
1429 CC = ISD::getSetCCInverse(CC,
1430 N0.getOperand(0).getValueType().isInteger());
1431 return DAG.getSetCC(VT, N0.getOperand(0), N0.getOperand(1), CC);
1434 if ((N0.getOpcode() == ISD::XOR ||
1435 (N0.getOpcode() == ISD::AND &&
1436 N0.getOperand(0).getOpcode() == ISD::XOR &&
1437 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
1438 isa<ConstantSDNode>(N0.getOperand(1)) &&
1439 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue() == 1) {
1440 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We
1441 // can only do this if the top bits are known zero.
1442 unsigned BitWidth = N0.getValueSizeInBits();
1443 if (DAG.MaskedValueIsZero(N0,
1444 APInt::getHighBitsSet(BitWidth,
1446 // Okay, get the un-inverted input value.
1448 if (N0.getOpcode() == ISD::XOR)
1449 Val = N0.getOperand(0);
1451 assert(N0.getOpcode() == ISD::AND &&
1452 N0.getOperand(0).getOpcode() == ISD::XOR);
1453 // ((X^1)&1)^1 -> X & 1
1454 Val = DAG.getNode(ISD::AND, N0.getValueType(),
1455 N0.getOperand(0).getOperand(0),
1458 return DAG.getSetCC(VT, Val, N1,
1459 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
1464 APInt MinVal, MaxVal;
1465 unsigned OperandBitSize = N1C->getValueType(0).getSizeInBits();
1466 if (ISD::isSignedIntSetCC(Cond)) {
1467 MinVal = APInt::getSignedMinValue(OperandBitSize);
1468 MaxVal = APInt::getSignedMaxValue(OperandBitSize);
1470 MinVal = APInt::getMinValue(OperandBitSize);
1471 MaxVal = APInt::getMaxValue(OperandBitSize);
1474 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
1475 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
1476 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true
1477 // X >= C0 --> X > (C0-1)
1478 return DAG.getSetCC(VT, N0, DAG.getConstant(C1-1, N1.getValueType()),
1479 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
1482 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
1483 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true
1484 // X <= C0 --> X < (C0+1)
1485 return DAG.getSetCC(VT, N0, DAG.getConstant(C1+1, N1.getValueType()),
1486 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
1489 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
1490 return DAG.getConstant(0, VT); // X < MIN --> false
1491 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal)
1492 return DAG.getConstant(1, VT); // X >= MIN --> true
1493 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal)
1494 return DAG.getConstant(0, VT); // X > MAX --> false
1495 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal)
1496 return DAG.getConstant(1, VT); // X <= MAX --> true
1498 // Canonicalize setgt X, Min --> setne X, Min
1499 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
1500 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
1501 // Canonicalize setlt X, Max --> setne X, Max
1502 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
1503 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
1505 // If we have setult X, 1, turn it into seteq X, 0
1506 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
1507 return DAG.getSetCC(VT, N0, DAG.getConstant(MinVal, N0.getValueType()),
1509 // If we have setugt X, Max-1, turn it into seteq X, Max
1510 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
1511 return DAG.getSetCC(VT, N0, DAG.getConstant(MaxVal, N0.getValueType()),
1514 // If we have "setcc X, C0", check to see if we can shrink the immediate
1517 // SETUGT X, SINTMAX -> SETLT X, 0
1518 if (Cond == ISD::SETUGT && OperandBitSize != 1 &&
1519 C1 == (~0ULL >> (65-OperandBitSize)))
1520 return DAG.getSetCC(VT, N0, DAG.getConstant(0, N1.getValueType()),
1523 // FIXME: Implement the rest of these.
1525 // Fold bit comparisons when we can.
1526 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1527 VT == N0.getValueType() && N0.getOpcode() == ISD::AND)
1528 if (ConstantSDNode *AndRHS =
1529 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1530 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
1531 // Perform the xform if the AND RHS is a single bit.
1532 if (isPowerOf2_64(AndRHS->getZExtValue())) {
1533 return DAG.getNode(ISD::SRL, VT, N0,
1534 DAG.getConstant(Log2_64(AndRHS->getZExtValue()),
1535 getShiftAmountTy()));
1537 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getZExtValue()) {
1538 // (X & 8) == 8 --> (X & 8) >> 3
1539 // Perform the xform if C1 is a single bit.
1540 if (C1.isPowerOf2()) {
1541 return DAG.getNode(ISD::SRL, VT, N0,
1542 DAG.getConstant(C1.logBase2(), getShiftAmountTy()));
1547 } else if (isa<ConstantSDNode>(N0.getNode())) {
1548 // Ensure that the constant occurs on the RHS.
1549 return DAG.getSetCC(VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
1552 if (isa<ConstantFPSDNode>(N0.getNode())) {
1553 // Constant fold or commute setcc.
1554 SDValue O = DAG.FoldSetCC(VT, N0, N1, Cond);
1555 if (O.getNode()) return O;
1556 } else if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1.getNode())) {
1557 // If the RHS of an FP comparison is a constant, simplify it away in
1559 if (CFP->getValueAPF().isNaN()) {
1560 // If an operand is known to be a nan, we can fold it.
1561 switch (ISD::getUnorderedFlavor(Cond)) {
1562 default: assert(0 && "Unknown flavor!");
1563 case 0: // Known false.
1564 return DAG.getConstant(0, VT);
1565 case 1: // Known true.
1566 return DAG.getConstant(1, VT);
1567 case 2: // Undefined.
1568 return DAG.getNode(ISD::UNDEF, VT);
1572 // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the
1573 // constant if knowing that the operand is non-nan is enough. We prefer to
1574 // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to
1576 if (Cond == ISD::SETO || Cond == ISD::SETUO)
1577 return DAG.getSetCC(VT, N0, N0, Cond);
1581 // We can always fold X == X for integer setcc's.
1582 if (N0.getValueType().isInteger())
1583 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
1584 unsigned UOF = ISD::getUnorderedFlavor(Cond);
1585 if (UOF == 2) // FP operators that are undefined on NaNs.
1586 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
1587 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
1588 return DAG.getConstant(UOF, VT);
1589 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
1590 // if it is not already.
1591 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
1592 if (NewCond != Cond)
1593 return DAG.getSetCC(VT, N0, N1, NewCond);
1596 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1597 N0.getValueType().isInteger()) {
1598 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
1599 N0.getOpcode() == ISD::XOR) {
1600 // Simplify (X+Y) == (X+Z) --> Y == Z
1601 if (N0.getOpcode() == N1.getOpcode()) {
1602 if (N0.getOperand(0) == N1.getOperand(0))
1603 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(1), Cond);
1604 if (N0.getOperand(1) == N1.getOperand(1))
1605 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(0), Cond);
1606 if (DAG.isCommutativeBinOp(N0.getOpcode())) {
1607 // If X op Y == Y op X, try other combinations.
1608 if (N0.getOperand(0) == N1.getOperand(1))
1609 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(0), Cond);
1610 if (N0.getOperand(1) == N1.getOperand(0))
1611 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(1), Cond);
1615 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) {
1616 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1617 // Turn (X+C1) == C2 --> X == C2-C1
1618 if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse()) {
1619 return DAG.getSetCC(VT, N0.getOperand(0),
1620 DAG.getConstant(RHSC->getAPIntValue()-
1621 LHSR->getAPIntValue(),
1622 N0.getValueType()), Cond);
1625 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
1626 if (N0.getOpcode() == ISD::XOR)
1627 // If we know that all of the inverted bits are zero, don't bother
1628 // performing the inversion.
1629 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue()))
1631 DAG.getSetCC(VT, N0.getOperand(0),
1632 DAG.getConstant(LHSR->getAPIntValue() ^
1633 RHSC->getAPIntValue(),
1638 // Turn (C1-X) == C2 --> X == C1-C2
1639 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
1640 if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse()) {
1642 DAG.getSetCC(VT, N0.getOperand(1),
1643 DAG.getConstant(SUBC->getAPIntValue() -
1644 RHSC->getAPIntValue(),
1651 // Simplify (X+Z) == X --> Z == 0
1652 if (N0.getOperand(0) == N1)
1653 return DAG.getSetCC(VT, N0.getOperand(1),
1654 DAG.getConstant(0, N0.getValueType()), Cond);
1655 if (N0.getOperand(1) == N1) {
1656 if (DAG.isCommutativeBinOp(N0.getOpcode()))
1657 return DAG.getSetCC(VT, N0.getOperand(0),
1658 DAG.getConstant(0, N0.getValueType()), Cond);
1659 else if (N0.getNode()->hasOneUse()) {
1660 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
1661 // (Z-X) == X --> Z == X<<1
1662 SDValue SH = DAG.getNode(ISD::SHL, N1.getValueType(),
1664 DAG.getConstant(1, getShiftAmountTy()));
1665 if (!DCI.isCalledByLegalizer())
1666 DCI.AddToWorklist(SH.getNode());
1667 return DAG.getSetCC(VT, N0.getOperand(0), SH, Cond);
1672 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
1673 N1.getOpcode() == ISD::XOR) {
1674 // Simplify X == (X+Z) --> Z == 0
1675 if (N1.getOperand(0) == N0) {
1676 return DAG.getSetCC(VT, N1.getOperand(1),
1677 DAG.getConstant(0, N1.getValueType()), Cond);
1678 } else if (N1.getOperand(1) == N0) {
1679 if (DAG.isCommutativeBinOp(N1.getOpcode())) {
1680 return DAG.getSetCC(VT, N1.getOperand(0),
1681 DAG.getConstant(0, N1.getValueType()), Cond);
1682 } else if (N1.getNode()->hasOneUse()) {
1683 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
1684 // X == (Z-X) --> X<<1 == Z
1685 SDValue SH = DAG.getNode(ISD::SHL, N1.getValueType(), N0,
1686 DAG.getConstant(1, getShiftAmountTy()));
1687 if (!DCI.isCalledByLegalizer())
1688 DCI.AddToWorklist(SH.getNode());
1689 return DAG.getSetCC(VT, SH, N1.getOperand(0), Cond);
1695 // Fold away ALL boolean setcc's.
1697 if (N0.getValueType() == MVT::i1 && foldBooleans) {
1699 default: assert(0 && "Unknown integer setcc!");
1700 case ISD::SETEQ: // X == Y -> (X^Y)^1
1701 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
1702 N0 = DAG.getNode(ISD::XOR, MVT::i1, Temp, DAG.getConstant(1, MVT::i1));
1703 if (!DCI.isCalledByLegalizer())
1704 DCI.AddToWorklist(Temp.getNode());
1706 case ISD::SETNE: // X != Y --> (X^Y)
1707 N0 = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
1709 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> X^1 & Y
1710 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> X^1 & Y
1711 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
1712 N0 = DAG.getNode(ISD::AND, MVT::i1, N1, Temp);
1713 if (!DCI.isCalledByLegalizer())
1714 DCI.AddToWorklist(Temp.getNode());
1716 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> Y^1 & X
1717 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> Y^1 & X
1718 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
1719 N0 = DAG.getNode(ISD::AND, MVT::i1, N0, Temp);
1720 if (!DCI.isCalledByLegalizer())
1721 DCI.AddToWorklist(Temp.getNode());
1723 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> X^1 | Y
1724 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> X^1 | Y
1725 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
1726 N0 = DAG.getNode(ISD::OR, MVT::i1, N1, Temp);
1727 if (!DCI.isCalledByLegalizer())
1728 DCI.AddToWorklist(Temp.getNode());
1730 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> Y^1 | X
1731 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> Y^1 | X
1732 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
1733 N0 = DAG.getNode(ISD::OR, MVT::i1, N0, Temp);
1736 if (VT != MVT::i1) {
1737 if (!DCI.isCalledByLegalizer())
1738 DCI.AddToWorklist(N0.getNode());
1739 // FIXME: If running after legalize, we probably can't do this.
1740 N0 = DAG.getNode(ISD::ZERO_EXTEND, VT, N0);
1745 // Could not fold it.
1749 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
1750 /// node is a GlobalAddress + offset.
1751 bool TargetLowering::isGAPlusOffset(SDNode *N, GlobalValue* &GA,
1752 int64_t &Offset) const {
1753 if (isa<GlobalAddressSDNode>(N)) {
1754 GlobalAddressSDNode *GASD = cast<GlobalAddressSDNode>(N);
1755 GA = GASD->getGlobal();
1756 Offset += GASD->getOffset();
1760 if (N->getOpcode() == ISD::ADD) {
1761 SDValue N1 = N->getOperand(0);
1762 SDValue N2 = N->getOperand(1);
1763 if (isGAPlusOffset(N1.getNode(), GA, Offset)) {
1764 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
1766 Offset += V->getSExtValue();
1769 } else if (isGAPlusOffset(N2.getNode(), GA, Offset)) {
1770 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
1772 Offset += V->getSExtValue();
1781 /// isConsecutiveLoad - Return true if LD (which must be a LoadSDNode) is
1782 /// loading 'Bytes' bytes from a location that is 'Dist' units away from the
1783 /// location that the 'Base' load is loading from.
1784 bool TargetLowering::isConsecutiveLoad(SDNode *LD, SDNode *Base,
1785 unsigned Bytes, int Dist,
1786 const MachineFrameInfo *MFI) const {
1787 if (LD->getOperand(0).getNode() != Base->getOperand(0).getNode())
1789 MVT VT = LD->getValueType(0);
1790 if (VT.getSizeInBits() / 8 != Bytes)
1793 SDValue Loc = LD->getOperand(1);
1794 SDValue BaseLoc = Base->getOperand(1);
1795 if (Loc.getOpcode() == ISD::FrameIndex) {
1796 if (BaseLoc.getOpcode() != ISD::FrameIndex)
1798 int FI = cast<FrameIndexSDNode>(Loc)->getIndex();
1799 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
1800 int FS = MFI->getObjectSize(FI);
1801 int BFS = MFI->getObjectSize(BFI);
1802 if (FS != BFS || FS != (int)Bytes) return false;
1803 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Bytes);
1806 GlobalValue *GV1 = NULL;
1807 GlobalValue *GV2 = NULL;
1808 int64_t Offset1 = 0;
1809 int64_t Offset2 = 0;
1810 bool isGA1 = isGAPlusOffset(Loc.getNode(), GV1, Offset1);
1811 bool isGA2 = isGAPlusOffset(BaseLoc.getNode(), GV2, Offset2);
1812 if (isGA1 && isGA2 && GV1 == GV2)
1813 return Offset1 == (Offset2 + Dist*Bytes);
1818 SDValue TargetLowering::
1819 PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
1820 // Default implementation: no optimization.
1824 //===----------------------------------------------------------------------===//
1825 // Inline Assembler Implementation Methods
1826 //===----------------------------------------------------------------------===//
1829 TargetLowering::ConstraintType
1830 TargetLowering::getConstraintType(const std::string &Constraint) const {
1831 // FIXME: lots more standard ones to handle.
1832 if (Constraint.size() == 1) {
1833 switch (Constraint[0]) {
1835 case 'r': return C_RegisterClass;
1837 case 'o': // offsetable
1838 case 'V': // not offsetable
1840 case 'i': // Simple Integer or Relocatable Constant
1841 case 'n': // Simple Integer
1842 case 's': // Relocatable Constant
1843 case 'X': // Allow ANY value.
1844 case 'I': // Target registers.
1856 if (Constraint.size() > 1 && Constraint[0] == '{' &&
1857 Constraint[Constraint.size()-1] == '}')
1862 /// LowerXConstraint - try to replace an X constraint, which matches anything,
1863 /// with another that has more specific requirements based on the type of the
1864 /// corresponding operand.
1865 const char *TargetLowering::LowerXConstraint(MVT ConstraintVT) const{
1866 if (ConstraintVT.isInteger())
1868 if (ConstraintVT.isFloatingPoint())
1869 return "f"; // works for many targets
1873 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
1874 /// vector. If it is invalid, don't add anything to Ops.
1875 void TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
1876 char ConstraintLetter,
1878 std::vector<SDValue> &Ops,
1879 SelectionDAG &DAG) const {
1880 switch (ConstraintLetter) {
1882 case 'X': // Allows any operand; labels (basic block) use this.
1883 if (Op.getOpcode() == ISD::BasicBlock) {
1888 case 'i': // Simple Integer or Relocatable Constant
1889 case 'n': // Simple Integer
1890 case 's': { // Relocatable Constant
1891 // These operands are interested in values of the form (GV+C), where C may
1892 // be folded in as an offset of GV, or it may be explicitly added. Also, it
1893 // is possible and fine if either GV or C are missing.
1894 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
1895 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
1897 // If we have "(add GV, C)", pull out GV/C
1898 if (Op.getOpcode() == ISD::ADD) {
1899 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
1900 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
1901 if (C == 0 || GA == 0) {
1902 C = dyn_cast<ConstantSDNode>(Op.getOperand(0));
1903 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1));
1905 if (C == 0 || GA == 0)
1909 // If we find a valid operand, map to the TargetXXX version so that the
1910 // value itself doesn't get selected.
1911 if (GA) { // Either &GV or &GV+C
1912 if (ConstraintLetter != 'n') {
1913 int64_t Offs = GA->getOffset();
1914 if (C) Offs += C->getZExtValue();
1915 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(),
1916 Op.getValueType(), Offs));
1920 if (C) { // just C, no GV.
1921 // Simple constants are not allowed for 's'.
1922 if (ConstraintLetter != 's') {
1923 Ops.push_back(DAG.getTargetConstant(C->getAPIntValue(),
1924 Op.getValueType()));
1933 std::vector<unsigned> TargetLowering::
1934 getRegClassForInlineAsmConstraint(const std::string &Constraint,
1936 return std::vector<unsigned>();
1940 std::pair<unsigned, const TargetRegisterClass*> TargetLowering::
1941 getRegForInlineAsmConstraint(const std::string &Constraint,
1943 if (Constraint[0] != '{')
1944 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
1945 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?");
1947 // Remove the braces from around the name.
1948 std::string RegName(Constraint.begin()+1, Constraint.end()-1);
1950 // Figure out which register class contains this reg.
1951 const TargetRegisterInfo *RI = TM.getRegisterInfo();
1952 for (TargetRegisterInfo::regclass_iterator RCI = RI->regclass_begin(),
1953 E = RI->regclass_end(); RCI != E; ++RCI) {
1954 const TargetRegisterClass *RC = *RCI;
1956 // If none of the the value types for this register class are valid, we
1957 // can't use it. For example, 64-bit reg classes on 32-bit targets.
1958 bool isLegal = false;
1959 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
1961 if (isTypeLegal(*I)) {
1967 if (!isLegal) continue;
1969 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end();
1971 if (StringsEqualNoCase(RegName, RI->get(*I).AsmName))
1972 return std::make_pair(*I, RC);
1976 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
1979 //===----------------------------------------------------------------------===//
1980 // Constraint Selection.
1982 /// isMatchingInputConstraint - Return true of this is an input operand that is
1983 /// a matching constraint like "4".
1984 bool TargetLowering::AsmOperandInfo::isMatchingInputConstraint() const {
1985 assert(!ConstraintCode.empty() && "No known constraint!");
1986 return isdigit(ConstraintCode[0]);
1989 /// getMatchedOperand - If this is an input matching constraint, this method
1990 /// returns the output operand it matches.
1991 unsigned TargetLowering::AsmOperandInfo::getMatchedOperand() const {
1992 assert(!ConstraintCode.empty() && "No known constraint!");
1993 return atoi(ConstraintCode.c_str());
1997 /// getConstraintGenerality - Return an integer indicating how general CT
1999 static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) {
2001 default: assert(0 && "Unknown constraint type!");
2002 case TargetLowering::C_Other:
2003 case TargetLowering::C_Unknown:
2005 case TargetLowering::C_Register:
2007 case TargetLowering::C_RegisterClass:
2009 case TargetLowering::C_Memory:
2014 /// ChooseConstraint - If there are multiple different constraints that we
2015 /// could pick for this operand (e.g. "imr") try to pick the 'best' one.
2016 /// This is somewhat tricky: constraints fall into four classes:
2017 /// Other -> immediates and magic values
2018 /// Register -> one specific register
2019 /// RegisterClass -> a group of regs
2020 /// Memory -> memory
2021 /// Ideally, we would pick the most specific constraint possible: if we have
2022 /// something that fits into a register, we would pick it. The problem here
2023 /// is that if we have something that could either be in a register or in
2024 /// memory that use of the register could cause selection of *other*
2025 /// operands to fail: they might only succeed if we pick memory. Because of
2026 /// this the heuristic we use is:
2028 /// 1) If there is an 'other' constraint, and if the operand is valid for
2029 /// that constraint, use it. This makes us take advantage of 'i'
2030 /// constraints when available.
2031 /// 2) Otherwise, pick the most general constraint present. This prefers
2032 /// 'm' over 'r', for example.
2034 static void ChooseConstraint(TargetLowering::AsmOperandInfo &OpInfo,
2035 bool hasMemory, const TargetLowering &TLI,
2036 SDValue Op, SelectionDAG *DAG) {
2037 assert(OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options");
2038 unsigned BestIdx = 0;
2039 TargetLowering::ConstraintType BestType = TargetLowering::C_Unknown;
2040 int BestGenerality = -1;
2042 // Loop over the options, keeping track of the most general one.
2043 for (unsigned i = 0, e = OpInfo.Codes.size(); i != e; ++i) {
2044 TargetLowering::ConstraintType CType =
2045 TLI.getConstraintType(OpInfo.Codes[i]);
2047 // If this is an 'other' constraint, see if the operand is valid for it.
2048 // For example, on X86 we might have an 'rI' constraint. If the operand
2049 // is an integer in the range [0..31] we want to use I (saving a load
2050 // of a register), otherwise we must use 'r'.
2051 if (CType == TargetLowering::C_Other && Op.getNode()) {
2052 assert(OpInfo.Codes[i].size() == 1 &&
2053 "Unhandled multi-letter 'other' constraint");
2054 std::vector<SDValue> ResultOps;
2055 TLI.LowerAsmOperandForConstraint(Op, OpInfo.Codes[i][0], hasMemory,
2057 if (!ResultOps.empty()) {
2064 // This constraint letter is more general than the previous one, use it.
2065 int Generality = getConstraintGenerality(CType);
2066 if (Generality > BestGenerality) {
2069 BestGenerality = Generality;
2073 OpInfo.ConstraintCode = OpInfo.Codes[BestIdx];
2074 OpInfo.ConstraintType = BestType;
2077 /// ComputeConstraintToUse - Determines the constraint code and constraint
2078 /// type to use for the specific AsmOperandInfo, setting
2079 /// OpInfo.ConstraintCode and OpInfo.ConstraintType.
2080 void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo,
2083 SelectionDAG *DAG) const {
2084 assert(!OpInfo.Codes.empty() && "Must have at least one constraint");
2086 // Single-letter constraints ('r') are very common.
2087 if (OpInfo.Codes.size() == 1) {
2088 OpInfo.ConstraintCode = OpInfo.Codes[0];
2089 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
2091 ChooseConstraint(OpInfo, hasMemory, *this, Op, DAG);
2094 // 'X' matches anything.
2095 if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) {
2096 // Labels and constants are handled elsewhere ('X' is the only thing
2097 // that matches labels).
2098 if (isa<BasicBlock>(OpInfo.CallOperandVal) ||
2099 isa<ConstantInt>(OpInfo.CallOperandVal))
2102 // Otherwise, try to resolve it to something we know about by looking at
2103 // the actual operand type.
2104 if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) {
2105 OpInfo.ConstraintCode = Repl;
2106 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
2111 //===----------------------------------------------------------------------===//
2112 // Loop Strength Reduction hooks
2113 //===----------------------------------------------------------------------===//
2115 /// isLegalAddressingMode - Return true if the addressing mode represented
2116 /// by AM is legal for this target, for a load/store of the specified type.
2117 bool TargetLowering::isLegalAddressingMode(const AddrMode &AM,
2118 const Type *Ty) const {
2119 // The default implementation of this implements a conservative RISCy, r+r and
2122 // Allows a sign-extended 16-bit immediate field.
2123 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
2126 // No global is ever allowed as a base.
2130 // Only support r+r,
2132 case 0: // "r+i" or just "i", depending on HasBaseReg.
2135 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
2137 // Otherwise we have r+r or r+i.
2140 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
2142 // Allow 2*r as r+r.
2149 // Magic for divide replacement
2152 int64_t m; // magic number
2153 int64_t s; // shift amount
2157 uint64_t m; // magic number
2158 int64_t a; // add indicator
2159 int64_t s; // shift amount
2162 /// magic - calculate the magic numbers required to codegen an integer sdiv as
2163 /// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
2165 static ms magic32(int32_t d) {
2167 uint32_t ad, anc, delta, q1, r1, q2, r2, t;
2168 const uint32_t two31 = 0x80000000U;
2172 t = two31 + ((uint32_t)d >> 31);
2173 anc = t - 1 - t%ad; // absolute value of nc
2174 p = 31; // initialize p
2175 q1 = two31/anc; // initialize q1 = 2p/abs(nc)
2176 r1 = two31 - q1*anc; // initialize r1 = rem(2p,abs(nc))
2177 q2 = two31/ad; // initialize q2 = 2p/abs(d)
2178 r2 = two31 - q2*ad; // initialize r2 = rem(2p,abs(d))
2181 q1 = 2*q1; // update q1 = 2p/abs(nc)
2182 r1 = 2*r1; // update r1 = rem(2p/abs(nc))
2183 if (r1 >= anc) { // must be unsigned comparison
2187 q2 = 2*q2; // update q2 = 2p/abs(d)
2188 r2 = 2*r2; // update r2 = rem(2p/abs(d))
2189 if (r2 >= ad) { // must be unsigned comparison
2194 } while (q1 < delta || (q1 == delta && r1 == 0));
2196 mag.m = (int32_t)(q2 + 1); // make sure to sign extend
2197 if (d < 0) mag.m = -mag.m; // resulting magic number
2198 mag.s = p - 32; // resulting shift
2202 /// magicu - calculate the magic numbers required to codegen an integer udiv as
2203 /// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
2204 static mu magicu32(uint32_t d) {
2206 uint32_t nc, delta, q1, r1, q2, r2;
2208 magu.a = 0; // initialize "add" indicator
2210 p = 31; // initialize p
2211 q1 = 0x80000000/nc; // initialize q1 = 2p/nc
2212 r1 = 0x80000000 - q1*nc; // initialize r1 = rem(2p,nc)
2213 q2 = 0x7FFFFFFF/d; // initialize q2 = (2p-1)/d
2214 r2 = 0x7FFFFFFF - q2*d; // initialize r2 = rem((2p-1),d)
2217 if (r1 >= nc - r1 ) {
2218 q1 = 2*q1 + 1; // update q1
2219 r1 = 2*r1 - nc; // update r1
2222 q1 = 2*q1; // update q1
2223 r1 = 2*r1; // update r1
2225 if (r2 + 1 >= d - r2) {
2226 if (q2 >= 0x7FFFFFFF) magu.a = 1;
2227 q2 = 2*q2 + 1; // update q2
2228 r2 = 2*r2 + 1 - d; // update r2
2231 if (q2 >= 0x80000000) magu.a = 1;
2232 q2 = 2*q2; // update q2
2233 r2 = 2*r2 + 1; // update r2
2236 } while (p < 64 && (q1 < delta || (q1 == delta && r1 == 0)));
2237 magu.m = q2 + 1; // resulting magic number
2238 magu.s = p - 32; // resulting shift
2242 /// magic - calculate the magic numbers required to codegen an integer sdiv as
2243 /// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
2245 static ms magic64(int64_t d) {
2247 uint64_t ad, anc, delta, q1, r1, q2, r2, t;
2248 const uint64_t two63 = 9223372036854775808ULL; // 2^63
2251 ad = d >= 0 ? d : -d;
2252 t = two63 + ((uint64_t)d >> 63);
2253 anc = t - 1 - t%ad; // absolute value of nc
2254 p = 63; // initialize p
2255 q1 = two63/anc; // initialize q1 = 2p/abs(nc)
2256 r1 = two63 - q1*anc; // initialize r1 = rem(2p,abs(nc))
2257 q2 = two63/ad; // initialize q2 = 2p/abs(d)
2258 r2 = two63 - q2*ad; // initialize r2 = rem(2p,abs(d))
2261 q1 = 2*q1; // update q1 = 2p/abs(nc)
2262 r1 = 2*r1; // update r1 = rem(2p/abs(nc))
2263 if (r1 >= anc) { // must be unsigned comparison
2267 q2 = 2*q2; // update q2 = 2p/abs(d)
2268 r2 = 2*r2; // update r2 = rem(2p/abs(d))
2269 if (r2 >= ad) { // must be unsigned comparison
2274 } while (q1 < delta || (q1 == delta && r1 == 0));
2277 if (d < 0) mag.m = -mag.m; // resulting magic number
2278 mag.s = p - 64; // resulting shift
2282 /// magicu - calculate the magic numbers required to codegen an integer udiv as
2283 /// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
2284 static mu magicu64(uint64_t d)
2287 uint64_t nc, delta, q1, r1, q2, r2;
2289 magu.a = 0; // initialize "add" indicator
2291 p = 63; // initialize p
2292 q1 = 0x8000000000000000ull/nc; // initialize q1 = 2p/nc
2293 r1 = 0x8000000000000000ull - q1*nc; // initialize r1 = rem(2p,nc)
2294 q2 = 0x7FFFFFFFFFFFFFFFull/d; // initialize q2 = (2p-1)/d
2295 r2 = 0x7FFFFFFFFFFFFFFFull - q2*d; // initialize r2 = rem((2p-1),d)
2298 if (r1 >= nc - r1 ) {
2299 q1 = 2*q1 + 1; // update q1
2300 r1 = 2*r1 - nc; // update r1
2303 q1 = 2*q1; // update q1
2304 r1 = 2*r1; // update r1
2306 if (r2 + 1 >= d - r2) {
2307 if (q2 >= 0x7FFFFFFFFFFFFFFFull) magu.a = 1;
2308 q2 = 2*q2 + 1; // update q2
2309 r2 = 2*r2 + 1 - d; // update r2
2312 if (q2 >= 0x8000000000000000ull) magu.a = 1;
2313 q2 = 2*q2; // update q2
2314 r2 = 2*r2 + 1; // update r2
2317 } while (p < 128 && (q1 < delta || (q1 == delta && r1 == 0)));
2318 magu.m = q2 + 1; // resulting magic number
2319 magu.s = p - 64; // resulting shift
2323 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
2324 /// return a DAG expression to select that will generate the same value by
2325 /// multiplying by a magic number. See:
2326 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
2327 SDValue TargetLowering::BuildSDIV(SDNode *N, SelectionDAG &DAG,
2328 std::vector<SDNode*>* Created) const {
2329 MVT VT = N->getValueType(0);
2331 // Check to see if we can do this.
2332 if (!isTypeLegal(VT) || (VT != MVT::i32 && VT != MVT::i64))
2333 return SDValue(); // BuildSDIV only operates on i32 or i64
2335 int64_t d = cast<ConstantSDNode>(N->getOperand(1))->getSExtValue();
2336 ms magics = (VT == MVT::i32) ? magic32(d) : magic64(d);
2338 // Multiply the numerator (operand 0) by the magic value
2340 if (isOperationLegal(ISD::MULHS, VT))
2341 Q = DAG.getNode(ISD::MULHS, VT, N->getOperand(0),
2342 DAG.getConstant(magics.m, VT));
2343 else if (isOperationLegal(ISD::SMUL_LOHI, VT))
2344 Q = SDValue(DAG.getNode(ISD::SMUL_LOHI, DAG.getVTList(VT, VT),
2346 DAG.getConstant(magics.m, VT)).getNode(), 1);
2348 return SDValue(); // No mulhs or equvialent
2349 // If d > 0 and m < 0, add the numerator
2350 if (d > 0 && magics.m < 0) {
2351 Q = DAG.getNode(ISD::ADD, VT, Q, N->getOperand(0));
2353 Created->push_back(Q.getNode());
2355 // If d < 0 and m > 0, subtract the numerator.
2356 if (d < 0 && magics.m > 0) {
2357 Q = DAG.getNode(ISD::SUB, VT, Q, N->getOperand(0));
2359 Created->push_back(Q.getNode());
2361 // Shift right algebraic if shift value is nonzero
2363 Q = DAG.getNode(ISD::SRA, VT, Q,
2364 DAG.getConstant(magics.s, getShiftAmountTy()));
2366 Created->push_back(Q.getNode());
2368 // Extract the sign bit and add it to the quotient
2370 DAG.getNode(ISD::SRL, VT, Q, DAG.getConstant(VT.getSizeInBits()-1,
2371 getShiftAmountTy()));
2373 Created->push_back(T.getNode());
2374 return DAG.getNode(ISD::ADD, VT, Q, T);
2377 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
2378 /// return a DAG expression to select that will generate the same value by
2379 /// multiplying by a magic number. See:
2380 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
2381 SDValue TargetLowering::BuildUDIV(SDNode *N, SelectionDAG &DAG,
2382 std::vector<SDNode*>* Created) const {
2383 MVT VT = N->getValueType(0);
2385 // Check to see if we can do this.
2386 if (!isTypeLegal(VT) || (VT != MVT::i32 && VT != MVT::i64))
2387 return SDValue(); // BuildUDIV only operates on i32 or i64
2389 uint64_t d = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
2390 mu magics = (VT == MVT::i32) ? magicu32(d) : magicu64(d);
2392 // Multiply the numerator (operand 0) by the magic value
2394 if (isOperationLegal(ISD::MULHU, VT))
2395 Q = DAG.getNode(ISD::MULHU, VT, N->getOperand(0),
2396 DAG.getConstant(magics.m, VT));
2397 else if (isOperationLegal(ISD::UMUL_LOHI, VT))
2398 Q = SDValue(DAG.getNode(ISD::UMUL_LOHI, DAG.getVTList(VT, VT),
2400 DAG.getConstant(magics.m, VT)).getNode(), 1);
2402 return SDValue(); // No mulhu or equvialent
2404 Created->push_back(Q.getNode());
2406 if (magics.a == 0) {
2407 return DAG.getNode(ISD::SRL, VT, Q,
2408 DAG.getConstant(magics.s, getShiftAmountTy()));
2410 SDValue NPQ = DAG.getNode(ISD::SUB, VT, N->getOperand(0), Q);
2412 Created->push_back(NPQ.getNode());
2413 NPQ = DAG.getNode(ISD::SRL, VT, NPQ,
2414 DAG.getConstant(1, getShiftAmountTy()));
2416 Created->push_back(NPQ.getNode());
2417 NPQ = DAG.getNode(ISD::ADD, VT, NPQ, Q);
2419 Created->push_back(NPQ.getNode());
2420 return DAG.getNode(ISD::SRL, VT, NPQ,
2421 DAG.getConstant(magics.s-1, getShiftAmountTy()));