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 // All operations default to being supported.
402 memset(OpActions, 0, sizeof(OpActions));
403 memset(LoadExtActions, 0, sizeof(LoadExtActions));
404 memset(TruncStoreActions, 0, sizeof(TruncStoreActions));
405 memset(IndexedModeActions, 0, sizeof(IndexedModeActions));
406 memset(ConvertActions, 0, sizeof(ConvertActions));
407 memset(CondCodeActions, 0, sizeof(CondCodeActions));
409 // Set default actions for various operations.
410 for (unsigned VT = 0; VT != (unsigned)MVT::LAST_VALUETYPE; ++VT) {
411 // Default all indexed load / store to expand.
412 for (unsigned IM = (unsigned)ISD::PRE_INC;
413 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
414 setIndexedLoadAction(IM, (MVT::SimpleValueType)VT, Expand);
415 setIndexedStoreAction(IM, (MVT::SimpleValueType)VT, Expand);
418 // These operations default to expand.
419 setOperationAction(ISD::FGETSIGN, (MVT::SimpleValueType)VT, Expand);
422 // Most targets ignore the @llvm.prefetch intrinsic.
423 setOperationAction(ISD::PREFETCH, MVT::Other, Expand);
425 // ConstantFP nodes default to expand. Targets can either change this to
426 // Legal, in which case all fp constants are legal, or use addLegalFPImmediate
427 // to optimize expansions for certain constants.
428 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
429 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
430 setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
432 // These library functions default to expand.
433 setOperationAction(ISD::FLOG , MVT::f64, Expand);
434 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
435 setOperationAction(ISD::FLOG10,MVT::f64, Expand);
436 setOperationAction(ISD::FEXP , MVT::f64, Expand);
437 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
438 setOperationAction(ISD::FLOG , MVT::f32, Expand);
439 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
440 setOperationAction(ISD::FLOG10,MVT::f32, Expand);
441 setOperationAction(ISD::FEXP , MVT::f32, Expand);
442 setOperationAction(ISD::FEXP2, MVT::f32, Expand);
444 // Default ISD::TRAP to expand (which turns it into abort).
445 setOperationAction(ISD::TRAP, MVT::Other, Expand);
447 IsLittleEndian = TD->isLittleEndian();
448 UsesGlobalOffsetTable = false;
449 ShiftAmountTy = PointerTy = getValueType(TD->getIntPtrType());
450 ShiftAmtHandling = Undefined;
451 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE*sizeof(TargetRegisterClass*));
452 memset(TargetDAGCombineArray, 0, array_lengthof(TargetDAGCombineArray));
453 maxStoresPerMemset = maxStoresPerMemcpy = maxStoresPerMemmove = 8;
454 allowUnalignedMemoryAccesses = false;
455 UseUnderscoreSetJmp = false;
456 UseUnderscoreLongJmp = false;
457 SelectIsExpensive = false;
458 IntDivIsCheap = false;
459 Pow2DivIsCheap = false;
460 StackPointerRegisterToSaveRestore = 0;
461 ExceptionPointerRegister = 0;
462 ExceptionSelectorRegister = 0;
463 BooleanContents = UndefinedBooleanContent;
464 SchedPreferenceInfo = SchedulingForLatency;
466 JumpBufAlignment = 0;
467 IfCvtBlockSizeLimit = 2;
468 IfCvtDupBlockSizeLimit = 0;
469 PrefLoopAlignment = 0;
471 InitLibcallNames(LibcallRoutineNames);
472 InitCmpLibcallCCs(CmpLibcallCCs);
474 // Tell Legalize whether the assembler supports DEBUG_LOC.
475 const TargetAsmInfo *TASM = TM.getTargetAsmInfo();
476 if (!TASM || !TASM->hasDotLocAndDotFile())
477 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
480 TargetLowering::~TargetLowering() {}
482 /// computeRegisterProperties - Once all of the register classes are added,
483 /// this allows us to compute derived properties we expose.
484 void TargetLowering::computeRegisterProperties() {
485 assert(MVT::LAST_VALUETYPE <= 32 &&
486 "Too many value types for ValueTypeActions to hold!");
488 // Everything defaults to needing one register.
489 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
490 NumRegistersForVT[i] = 1;
491 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i;
493 // ...except isVoid, which doesn't need any registers.
494 NumRegistersForVT[MVT::isVoid] = 0;
496 // Find the largest integer register class.
497 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE;
498 for (; RegClassForVT[LargestIntReg] == 0; --LargestIntReg)
499 assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
501 // Every integer value type larger than this largest register takes twice as
502 // many registers to represent as the previous ValueType.
503 for (unsigned ExpandedReg = LargestIntReg + 1; ; ++ExpandedReg) {
504 MVT EVT = (MVT::SimpleValueType)ExpandedReg;
505 if (!EVT.isInteger())
507 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
508 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg;
509 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1);
510 ValueTypeActions.setTypeAction(EVT, Expand);
513 // Inspect all of the ValueType's smaller than the largest integer
514 // register to see which ones need promotion.
515 unsigned LegalIntReg = LargestIntReg;
516 for (unsigned IntReg = LargestIntReg - 1;
517 IntReg >= (unsigned)MVT::i1; --IntReg) {
518 MVT IVT = (MVT::SimpleValueType)IntReg;
519 if (isTypeLegal(IVT)) {
520 LegalIntReg = IntReg;
522 RegisterTypeForVT[IntReg] = TransformToType[IntReg] =
523 (MVT::SimpleValueType)LegalIntReg;
524 ValueTypeActions.setTypeAction(IVT, Promote);
528 // ppcf128 type is really two f64's.
529 if (!isTypeLegal(MVT::ppcf128)) {
530 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64];
531 RegisterTypeForVT[MVT::ppcf128] = MVT::f64;
532 TransformToType[MVT::ppcf128] = MVT::f64;
533 ValueTypeActions.setTypeAction(MVT::ppcf128, Expand);
536 // Decide how to handle f64. If the target does not have native f64 support,
537 // expand it to i64 and we will be generating soft float library calls.
538 if (!isTypeLegal(MVT::f64)) {
539 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
540 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
541 TransformToType[MVT::f64] = MVT::i64;
542 ValueTypeActions.setTypeAction(MVT::f64, Expand);
545 // Decide how to handle f32. If the target does not have native support for
546 // f32, promote it to f64 if it is legal. Otherwise, expand it to i32.
547 if (!isTypeLegal(MVT::f32)) {
548 if (isTypeLegal(MVT::f64)) {
549 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::f64];
550 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::f64];
551 TransformToType[MVT::f32] = MVT::f64;
552 ValueTypeActions.setTypeAction(MVT::f32, Promote);
554 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
555 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
556 TransformToType[MVT::f32] = MVT::i32;
557 ValueTypeActions.setTypeAction(MVT::f32, Expand);
561 // Loop over all of the vector value types to see which need transformations.
562 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
563 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
564 MVT VT = (MVT::SimpleValueType)i;
565 if (!isTypeLegal(VT)) {
566 MVT IntermediateVT, RegisterVT;
567 unsigned NumIntermediates;
568 NumRegistersForVT[i] =
569 getVectorTypeBreakdown(VT,
570 IntermediateVT, NumIntermediates,
572 RegisterTypeForVT[i] = RegisterVT;
574 // Determine if there is a legal wider type.
575 bool IsLegalWiderType = false;
576 MVT EltVT = VT.getVectorElementType();
577 unsigned NElts = VT.getVectorNumElements();
578 for (unsigned nVT = i+1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
579 MVT SVT = (MVT::SimpleValueType)nVT;
580 if (isTypeLegal(SVT) && SVT.getVectorElementType() == EltVT &&
581 SVT.getVectorNumElements() > NElts) {
582 TransformToType[i] = SVT;
583 ValueTypeActions.setTypeAction(VT, Promote);
584 IsLegalWiderType = true;
588 if (!IsLegalWiderType) {
589 MVT NVT = VT.getPow2VectorType();
591 // Type is already a power of 2. The default action is to split.
592 TransformToType[i] = MVT::Other;
593 ValueTypeActions.setTypeAction(VT, Expand);
595 TransformToType[i] = NVT;
596 ValueTypeActions.setTypeAction(VT, Promote);
603 const char *TargetLowering::getTargetNodeName(unsigned Opcode) const {
608 MVT TargetLowering::getSetCCResultType(MVT VT) const {
609 return getValueType(TD->getIntPtrType());
613 /// getVectorTypeBreakdown - Vector types are broken down into some number of
614 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
615 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
616 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
618 /// This method returns the number of registers needed, and the VT for each
619 /// register. It also returns the VT and quantity of the intermediate values
620 /// before they are promoted/expanded.
622 unsigned TargetLowering::getVectorTypeBreakdown(MVT VT,
624 unsigned &NumIntermediates,
625 MVT &RegisterVT) const {
626 // Figure out the right, legal destination reg to copy into.
627 unsigned NumElts = VT.getVectorNumElements();
628 MVT EltTy = VT.getVectorElementType();
630 unsigned NumVectorRegs = 1;
632 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
633 // could break down into LHS/RHS like LegalizeDAG does.
634 if (!isPowerOf2_32(NumElts)) {
635 NumVectorRegs = NumElts;
639 // Divide the input until we get to a supported size. This will always
640 // end with a scalar if the target doesn't support vectors.
641 while (NumElts > 1 && !isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) {
646 NumIntermediates = NumVectorRegs;
648 MVT NewVT = MVT::getVectorVT(EltTy, NumElts);
649 if (!isTypeLegal(NewVT))
651 IntermediateVT = NewVT;
653 MVT DestVT = getTypeToTransformTo(NewVT);
655 if (DestVT.bitsLT(NewVT)) {
656 // Value is expanded, e.g. i64 -> i16.
657 return NumVectorRegs*(NewVT.getSizeInBits()/DestVT.getSizeInBits());
659 // Otherwise, promotion or legal types use the same number of registers as
660 // the vector decimated to the appropriate level.
661 return NumVectorRegs;
667 /// getWidenVectorType: given a vector type, returns the type to widen to
668 /// (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
669 /// If there is no vector type that we want to widen to, returns MVT::Other
670 /// When and where to widen is target dependent based on the cost of
671 /// scalarizing vs using the wider vector type.
672 MVT TargetLowering::getWidenVectorType(MVT VT) {
673 assert(VT.isVector());
677 // Default is not to widen until moved to LegalizeTypes
681 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
682 /// function arguments in the caller parameter area. This is the actual
683 /// alignment, not its logarithm.
684 unsigned TargetLowering::getByValTypeAlignment(const Type *Ty) const {
685 return TD->getCallFrameTypeAlignment(Ty);
688 SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table,
689 SelectionDAG &DAG) const {
690 if (usesGlobalOffsetTable())
691 return DAG.getNode(ISD::GLOBAL_OFFSET_TABLE, getPointerTy());
696 TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
697 // Assume that everything is safe in static mode.
698 if (getTargetMachine().getRelocationModel() == Reloc::Static)
701 // In dynamic-no-pic mode, assume that known defined values are safe.
702 if (getTargetMachine().getRelocationModel() == Reloc::DynamicNoPIC &&
704 !GA->getGlobal()->isDeclaration() &&
705 !GA->getGlobal()->mayBeOverridden())
708 // Otherwise assume nothing is safe.
712 //===----------------------------------------------------------------------===//
713 // Optimization Methods
714 //===----------------------------------------------------------------------===//
716 /// ShrinkDemandedConstant - Check to see if the specified operand of the
717 /// specified instruction is a constant integer. If so, check to see if there
718 /// are any bits set in the constant that are not demanded. If so, shrink the
719 /// constant and return true.
720 bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDValue Op,
721 const APInt &Demanded) {
722 // FIXME: ISD::SELECT, ISD::SELECT_CC
723 switch(Op.getOpcode()) {
728 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1)))
729 if (C->getAPIntValue().intersects(~Demanded)) {
730 MVT VT = Op.getValueType();
731 SDValue New = DAG.getNode(Op.getOpcode(), VT, Op.getOperand(0),
732 DAG.getConstant(Demanded &
735 return CombineTo(Op, New);
742 /// SimplifyDemandedBits - Look at Op. At this point, we know that only the
743 /// DemandedMask bits of the result of Op are ever used downstream. If we can
744 /// use this information to simplify Op, create a new simplified DAG node and
745 /// return true, returning the original and new nodes in Old and New. Otherwise,
746 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
747 /// the expression (used to simplify the caller). The KnownZero/One bits may
748 /// only be accurate for those bits in the DemandedMask.
749 bool TargetLowering::SimplifyDemandedBits(SDValue Op,
750 const APInt &DemandedMask,
753 TargetLoweringOpt &TLO,
754 unsigned Depth) const {
755 unsigned BitWidth = DemandedMask.getBitWidth();
756 assert(Op.getValueSizeInBits() == BitWidth &&
757 "Mask size mismatches value type size!");
758 APInt NewMask = DemandedMask;
760 // Don't know anything.
761 KnownZero = KnownOne = APInt(BitWidth, 0);
763 // Other users may use these bits.
764 if (!Op.getNode()->hasOneUse()) {
766 // If not at the root, Just compute the KnownZero/KnownOne bits to
767 // simplify things downstream.
768 TLO.DAG.ComputeMaskedBits(Op, DemandedMask, KnownZero, KnownOne, Depth);
771 // If this is the root being simplified, allow it to have multiple uses,
772 // just set the NewMask to all bits.
773 NewMask = APInt::getAllOnesValue(BitWidth);
774 } else if (DemandedMask == 0) {
775 // Not demanding any bits from Op.
776 if (Op.getOpcode() != ISD::UNDEF)
777 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::UNDEF, Op.getValueType()));
779 } else if (Depth == 6) { // Limit search depth.
783 APInt KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut;
784 switch (Op.getOpcode()) {
786 // We know all of the bits for a constant!
787 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue() & NewMask;
788 KnownZero = ~KnownOne & NewMask;
789 return false; // Don't fall through, will infinitely loop.
791 // If the RHS is a constant, check to see if the LHS would be zero without
792 // using the bits from the RHS. Below, we use knowledge about the RHS to
793 // simplify the LHS, here we're using information from the LHS to simplify
795 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
796 APInt LHSZero, LHSOne;
797 TLO.DAG.ComputeMaskedBits(Op.getOperand(0), NewMask,
798 LHSZero, LHSOne, Depth+1);
799 // If the LHS already has zeros where RHSC does, this and is dead.
800 if ((LHSZero & NewMask) == (~RHSC->getAPIntValue() & NewMask))
801 return TLO.CombineTo(Op, Op.getOperand(0));
802 // If any of the set bits in the RHS are known zero on the LHS, shrink
804 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & NewMask))
808 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
809 KnownOne, TLO, Depth+1))
811 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
812 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownZero & NewMask,
813 KnownZero2, KnownOne2, TLO, Depth+1))
815 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
817 // If all of the demanded bits are known one on one side, return the other.
818 // These bits cannot contribute to the result of the 'and'.
819 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
820 return TLO.CombineTo(Op, Op.getOperand(0));
821 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
822 return TLO.CombineTo(Op, Op.getOperand(1));
823 // If all of the demanded bits in the inputs are known zeros, return zero.
824 if ((NewMask & (KnownZero|KnownZero2)) == NewMask)
825 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, Op.getValueType()));
826 // If the RHS is a constant, see if we can simplify it.
827 if (TLO.ShrinkDemandedConstant(Op, ~KnownZero2 & NewMask))
830 // Output known-1 bits are only known if set in both the LHS & RHS.
831 KnownOne &= KnownOne2;
832 // Output known-0 are known to be clear if zero in either the LHS | RHS.
833 KnownZero |= KnownZero2;
836 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
837 KnownOne, TLO, Depth+1))
839 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
840 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownOne & NewMask,
841 KnownZero2, KnownOne2, TLO, Depth+1))
843 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
845 // If all of the demanded bits are known zero on one side, return the other.
846 // These bits cannot contribute to the result of the 'or'.
847 if ((NewMask & ~KnownOne2 & KnownZero) == (~KnownOne2 & NewMask))
848 return TLO.CombineTo(Op, Op.getOperand(0));
849 if ((NewMask & ~KnownOne & KnownZero2) == (~KnownOne & NewMask))
850 return TLO.CombineTo(Op, Op.getOperand(1));
851 // If all of the potentially set bits on one side are known to be set on
852 // the other side, just use the 'other' side.
853 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
854 return TLO.CombineTo(Op, Op.getOperand(0));
855 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
856 return TLO.CombineTo(Op, Op.getOperand(1));
857 // If the RHS is a constant, see if we can simplify it.
858 if (TLO.ShrinkDemandedConstant(Op, NewMask))
861 // Output known-0 bits are only known if clear in both the LHS & RHS.
862 KnownZero &= KnownZero2;
863 // Output known-1 are known to be set if set in either the LHS | RHS.
864 KnownOne |= KnownOne2;
867 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
868 KnownOne, TLO, Depth+1))
870 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
871 if (SimplifyDemandedBits(Op.getOperand(0), NewMask, KnownZero2,
872 KnownOne2, TLO, Depth+1))
874 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
876 // If all of the demanded bits are known zero on one side, return the other.
877 // These bits cannot contribute to the result of the 'xor'.
878 if ((KnownZero & NewMask) == NewMask)
879 return TLO.CombineTo(Op, Op.getOperand(0));
880 if ((KnownZero2 & NewMask) == NewMask)
881 return TLO.CombineTo(Op, Op.getOperand(1));
883 // If all of the unknown bits are known to be zero on one side or the other
884 // (but not both) turn this into an *inclusive* or.
885 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
886 if ((NewMask & ~KnownZero & ~KnownZero2) == 0)
887 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, Op.getValueType(),
891 // Output known-0 bits are known if clear or set in both the LHS & RHS.
892 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
893 // Output known-1 are known to be set if set in only one of the LHS, RHS.
894 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
896 // If all of the demanded bits on one side are known, and all of the set
897 // bits on that side are also known to be set on the other side, turn this
898 // into an AND, as we know the bits will be cleared.
899 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
900 if ((NewMask & (KnownZero|KnownOne)) == NewMask) { // all known
901 if ((KnownOne & KnownOne2) == KnownOne) {
902 MVT VT = Op.getValueType();
903 SDValue ANDC = TLO.DAG.getConstant(~KnownOne & NewMask, VT);
904 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, VT, Op.getOperand(0),
909 // If the RHS is a constant, see if we can simplify it.
910 // for XOR, we prefer to force bits to 1 if they will make a -1.
911 // if we can't force bits, try to shrink constant
912 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
913 APInt Expanded = C->getAPIntValue() | (~NewMask);
914 // if we can expand it to have all bits set, do it
915 if (Expanded.isAllOnesValue()) {
916 if (Expanded != C->getAPIntValue()) {
917 MVT VT = Op.getValueType();
918 SDValue New = TLO.DAG.getNode(Op.getOpcode(), VT, Op.getOperand(0),
919 TLO.DAG.getConstant(Expanded, VT));
920 return TLO.CombineTo(Op, New);
922 // if it already has all the bits set, nothing to change
923 // but don't shrink either!
924 } else if (TLO.ShrinkDemandedConstant(Op, NewMask)) {
929 KnownZero = KnownZeroOut;
930 KnownOne = KnownOneOut;
933 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero,
934 KnownOne, TLO, Depth+1))
936 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero2,
937 KnownOne2, TLO, Depth+1))
939 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
940 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
942 // If the operands are constants, see if we can simplify them.
943 if (TLO.ShrinkDemandedConstant(Op, NewMask))
946 // Only known if known in both the LHS and RHS.
947 KnownOne &= KnownOne2;
948 KnownZero &= KnownZero2;
951 if (SimplifyDemandedBits(Op.getOperand(3), NewMask, KnownZero,
952 KnownOne, TLO, Depth+1))
954 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero2,
955 KnownOne2, TLO, Depth+1))
957 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
958 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
960 // If the operands are constants, see if we can simplify them.
961 if (TLO.ShrinkDemandedConstant(Op, NewMask))
964 // Only known if known in both the LHS and RHS.
965 KnownOne &= KnownOne2;
966 KnownZero &= KnownZero2;
969 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
970 unsigned ShAmt = SA->getZExtValue();
971 SDValue InOp = Op.getOperand(0);
973 // If the shift count is an invalid immediate, don't do anything.
974 if (ShAmt >= BitWidth)
977 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a
978 // single shift. We can do this if the bottom bits (which are shifted
979 // out) are never demanded.
980 if (InOp.getOpcode() == ISD::SRL &&
981 isa<ConstantSDNode>(InOp.getOperand(1))) {
982 if (ShAmt && (NewMask & APInt::getLowBitsSet(BitWidth, ShAmt)) == 0) {
983 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
984 unsigned Opc = ISD::SHL;
992 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
993 MVT VT = Op.getValueType();
994 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, VT,
995 InOp.getOperand(0), NewSA));
999 if (SimplifyDemandedBits(Op.getOperand(0), NewMask.lshr(ShAmt),
1000 KnownZero, KnownOne, TLO, Depth+1))
1002 KnownZero <<= SA->getZExtValue();
1003 KnownOne <<= SA->getZExtValue();
1004 // low bits known zero.
1005 KnownZero |= APInt::getLowBitsSet(BitWidth, SA->getZExtValue());
1009 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1010 MVT VT = Op.getValueType();
1011 unsigned ShAmt = SA->getZExtValue();
1012 unsigned VTSize = VT.getSizeInBits();
1013 SDValue InOp = Op.getOperand(0);
1015 // If the shift count is an invalid immediate, don't do anything.
1016 if (ShAmt >= BitWidth)
1019 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a
1020 // single shift. We can do this if the top bits (which are shifted out)
1021 // are never demanded.
1022 if (InOp.getOpcode() == ISD::SHL &&
1023 isa<ConstantSDNode>(InOp.getOperand(1))) {
1024 if (ShAmt && (NewMask & APInt::getHighBitsSet(VTSize, ShAmt)) == 0) {
1025 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
1026 unsigned Opc = ISD::SRL;
1027 int Diff = ShAmt-C1;
1034 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
1035 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, VT,
1036 InOp.getOperand(0), NewSA));
1040 // Compute the new bits that are at the top now.
1041 if (SimplifyDemandedBits(InOp, (NewMask << ShAmt),
1042 KnownZero, KnownOne, TLO, Depth+1))
1044 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1045 KnownZero = KnownZero.lshr(ShAmt);
1046 KnownOne = KnownOne.lshr(ShAmt);
1048 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
1049 KnownZero |= HighBits; // High bits known zero.
1053 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1054 MVT VT = Op.getValueType();
1055 unsigned ShAmt = SA->getZExtValue();
1057 // If the shift count is an invalid immediate, don't do anything.
1058 if (ShAmt >= BitWidth)
1061 APInt InDemandedMask = (NewMask << ShAmt);
1063 // If any of the demanded bits are produced by the sign extension, we also
1064 // demand the input sign bit.
1065 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
1066 if (HighBits.intersects(NewMask))
1067 InDemandedMask |= APInt::getSignBit(VT.getSizeInBits());
1069 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask,
1070 KnownZero, KnownOne, TLO, Depth+1))
1072 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1073 KnownZero = KnownZero.lshr(ShAmt);
1074 KnownOne = KnownOne.lshr(ShAmt);
1076 // Handle the sign bit, adjusted to where it is now in the mask.
1077 APInt SignBit = APInt::getSignBit(BitWidth).lshr(ShAmt);
1079 // If the input sign bit is known to be zero, or if none of the top bits
1080 // are demanded, turn this into an unsigned shift right.
1081 if (KnownZero.intersects(SignBit) || (HighBits & ~NewMask) == HighBits) {
1082 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, VT, Op.getOperand(0),
1084 } else if (KnownOne.intersects(SignBit)) { // New bits are known one.
1085 KnownOne |= HighBits;
1089 case ISD::SIGN_EXTEND_INREG: {
1090 MVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1092 // Sign extension. Compute the demanded bits in the result that are not
1093 // present in the input.
1094 APInt NewBits = APInt::getHighBitsSet(BitWidth,
1095 BitWidth - EVT.getSizeInBits()) &
1098 // If none of the extended bits are demanded, eliminate the sextinreg.
1100 return TLO.CombineTo(Op, Op.getOperand(0));
1102 APInt InSignBit = APInt::getSignBit(EVT.getSizeInBits());
1103 InSignBit.zext(BitWidth);
1104 APInt InputDemandedBits = APInt::getLowBitsSet(BitWidth,
1105 EVT.getSizeInBits()) &
1108 // Since the sign extended bits are demanded, we know that the sign
1110 InputDemandedBits |= InSignBit;
1112 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits,
1113 KnownZero, KnownOne, TLO, Depth+1))
1115 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1117 // If the sign bit of the input is known set or clear, then we know the
1118 // top bits of the result.
1120 // If the input sign bit is known zero, convert this into a zero extension.
1121 if (KnownZero.intersects(InSignBit))
1122 return TLO.CombineTo(Op,
1123 TLO.DAG.getZeroExtendInReg(Op.getOperand(0), EVT));
1125 if (KnownOne.intersects(InSignBit)) { // Input sign bit known set
1126 KnownOne |= NewBits;
1127 KnownZero &= ~NewBits;
1128 } else { // Input sign bit unknown
1129 KnownZero &= ~NewBits;
1130 KnownOne &= ~NewBits;
1134 case ISD::ZERO_EXTEND: {
1135 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits();
1136 APInt InMask = NewMask;
1137 InMask.trunc(OperandBitWidth);
1139 // If none of the top bits are demanded, convert this into an any_extend.
1141 APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask;
1142 if (!NewBits.intersects(NewMask))
1143 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND,
1147 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1148 KnownZero, KnownOne, TLO, Depth+1))
1150 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1151 KnownZero.zext(BitWidth);
1152 KnownOne.zext(BitWidth);
1153 KnownZero |= NewBits;
1156 case ISD::SIGN_EXTEND: {
1157 MVT InVT = Op.getOperand(0).getValueType();
1158 unsigned InBits = InVT.getSizeInBits();
1159 APInt InMask = APInt::getLowBitsSet(BitWidth, InBits);
1160 APInt InSignBit = APInt::getBitsSet(BitWidth, InBits - 1, InBits);
1161 APInt NewBits = ~InMask & NewMask;
1163 // If none of the top bits are demanded, convert this into an any_extend.
1165 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND,Op.getValueType(),
1168 // Since some of the sign extended bits are demanded, we know that the sign
1170 APInt InDemandedBits = InMask & NewMask;
1171 InDemandedBits |= InSignBit;
1172 InDemandedBits.trunc(InBits);
1174 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero,
1175 KnownOne, TLO, Depth+1))
1177 KnownZero.zext(BitWidth);
1178 KnownOne.zext(BitWidth);
1180 // If the sign bit is known zero, convert this to a zero extend.
1181 if (KnownZero.intersects(InSignBit))
1182 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND,
1186 // If the sign bit is known one, the top bits match.
1187 if (KnownOne.intersects(InSignBit)) {
1188 KnownOne |= NewBits;
1189 KnownZero &= ~NewBits;
1190 } else { // Otherwise, top bits aren't known.
1191 KnownOne &= ~NewBits;
1192 KnownZero &= ~NewBits;
1196 case ISD::ANY_EXTEND: {
1197 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits();
1198 APInt InMask = NewMask;
1199 InMask.trunc(OperandBitWidth);
1200 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1201 KnownZero, KnownOne, TLO, Depth+1))
1203 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1204 KnownZero.zext(BitWidth);
1205 KnownOne.zext(BitWidth);
1208 case ISD::TRUNCATE: {
1209 // Simplify the input, using demanded bit information, and compute the known
1210 // zero/one bits live out.
1211 APInt TruncMask = NewMask;
1212 TruncMask.zext(Op.getOperand(0).getValueSizeInBits());
1213 if (SimplifyDemandedBits(Op.getOperand(0), TruncMask,
1214 KnownZero, KnownOne, TLO, Depth+1))
1216 KnownZero.trunc(BitWidth);
1217 KnownOne.trunc(BitWidth);
1219 // If the input is only used by this truncate, see if we can shrink it based
1220 // on the known demanded bits.
1221 if (Op.getOperand(0).getNode()->hasOneUse()) {
1222 SDValue In = Op.getOperand(0);
1223 unsigned InBitWidth = In.getValueSizeInBits();
1224 switch (In.getOpcode()) {
1227 // Shrink SRL by a constant if none of the high bits shifted in are
1229 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1))){
1230 APInt HighBits = APInt::getHighBitsSet(InBitWidth,
1231 InBitWidth - BitWidth);
1232 HighBits = HighBits.lshr(ShAmt->getZExtValue());
1233 HighBits.trunc(BitWidth);
1235 if (ShAmt->getZExtValue() < BitWidth && !(HighBits & NewMask)) {
1236 // None of the shifted in bits are needed. Add a truncate of the
1237 // shift input, then shift it.
1238 SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE,
1241 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL,Op.getValueType(),
1242 NewTrunc, In.getOperand(1)));
1249 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1252 case ISD::AssertZext: {
1253 MVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1254 APInt InMask = APInt::getLowBitsSet(BitWidth,
1255 VT.getSizeInBits());
1256 if (SimplifyDemandedBits(Op.getOperand(0), InMask & NewMask,
1257 KnownZero, KnownOne, TLO, Depth+1))
1259 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1260 KnownZero |= ~InMask & NewMask;
1263 case ISD::BIT_CONVERT:
1265 // If this is an FP->Int bitcast and if the sign bit is the only thing that
1266 // is demanded, turn this into a FGETSIGN.
1267 if (NewMask == MVT::getIntegerVTSignBit(Op.getValueType()) &&
1268 MVT::isFloatingPoint(Op.getOperand(0).getValueType()) &&
1269 !MVT::isVector(Op.getOperand(0).getValueType())) {
1270 // Only do this xform if FGETSIGN is valid or if before legalize.
1271 if (!TLO.AfterLegalize ||
1272 isOperationLegal(ISD::FGETSIGN, Op.getValueType())) {
1273 // Make a FGETSIGN + SHL to move the sign bit into the appropriate
1274 // place. We expect the SHL to be eliminated by other optimizations.
1275 SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, Op.getValueType(),
1277 unsigned ShVal = Op.getValueType().getSizeInBits()-1;
1278 SDValue ShAmt = TLO.DAG.getConstant(ShVal, getShiftAmountTy());
1279 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, Op.getValueType(),
1286 // Just use ComputeMaskedBits to compute output bits.
1287 TLO.DAG.ComputeMaskedBits(Op, NewMask, KnownZero, KnownOne, Depth);
1291 // If we know the value of all of the demanded bits, return this as a
1293 if ((NewMask & (KnownZero|KnownOne)) == NewMask)
1294 return TLO.CombineTo(Op, TLO.DAG.getConstant(KnownOne, Op.getValueType()));
1299 /// computeMaskedBitsForTargetNode - Determine which of the bits specified
1300 /// in Mask are known to be either zero or one and return them in the
1301 /// KnownZero/KnownOne bitsets.
1302 void TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
1306 const SelectionDAG &DAG,
1307 unsigned Depth) const {
1308 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1309 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1310 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1311 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1312 "Should use MaskedValueIsZero if you don't know whether Op"
1313 " is a target node!");
1314 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
1317 /// ComputeNumSignBitsForTargetNode - This method can be implemented by
1318 /// targets that want to expose additional information about sign bits to the
1320 unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
1321 unsigned Depth) const {
1322 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1323 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1324 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1325 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1326 "Should use ComputeNumSignBits if you don't know whether Op"
1327 " is a target node!");
1332 /// SimplifySetCC - Try to simplify a setcc built with the specified operands
1333 /// and cc. If it is unable to simplify it, return a null SDValue.
1335 TargetLowering::SimplifySetCC(MVT VT, SDValue N0, SDValue N1,
1336 ISD::CondCode Cond, bool foldBooleans,
1337 DAGCombinerInfo &DCI) const {
1338 SelectionDAG &DAG = DCI.DAG;
1340 // These setcc operations always fold.
1344 case ISD::SETFALSE2: return DAG.getConstant(0, VT);
1346 case ISD::SETTRUE2: return DAG.getConstant(1, VT);
1349 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) {
1350 const APInt &C1 = N1C->getAPIntValue();
1351 if (isa<ConstantSDNode>(N0.getNode())) {
1352 return DAG.FoldSetCC(VT, N0, N1, Cond);
1354 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an
1355 // equality comparison, then we're just comparing whether X itself is
1357 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) &&
1358 N0.getOperand(0).getOpcode() == ISD::CTLZ &&
1359 N0.getOperand(1).getOpcode() == ISD::Constant) {
1360 unsigned ShAmt = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
1361 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1362 ShAmt == Log2_32(N0.getValueType().getSizeInBits())) {
1363 if ((C1 == 0) == (Cond == ISD::SETEQ)) {
1364 // (srl (ctlz x), 5) == 0 -> X != 0
1365 // (srl (ctlz x), 5) != 1 -> X != 0
1368 // (srl (ctlz x), 5) != 0 -> X == 0
1369 // (srl (ctlz x), 5) == 1 -> X == 0
1372 SDValue Zero = DAG.getConstant(0, N0.getValueType());
1373 return DAG.getSetCC(VT, N0.getOperand(0).getOperand(0),
1378 // If the LHS is '(and load, const)', the RHS is 0,
1379 // the test is for equality or unsigned, and all 1 bits of the const are
1380 // in the same partial word, see if we can shorten the load.
1381 if (DCI.isBeforeLegalize() &&
1382 N0.getOpcode() == ISD::AND && C1 == 0 &&
1383 isa<LoadSDNode>(N0.getOperand(0)) &&
1384 N0.getOperand(0).getNode()->hasOneUse() &&
1385 isa<ConstantSDNode>(N0.getOperand(1))) {
1386 LoadSDNode *Lod = cast<LoadSDNode>(N0.getOperand(0));
1387 uint64_t Mask = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
1388 uint64_t bestMask = 0;
1389 unsigned bestWidth = 0, bestOffset = 0;
1390 if (!Lod->isVolatile() && Lod->isUnindexed()) {
1391 unsigned origWidth = N0.getValueType().getSizeInBits();
1392 // We can narrow (e.g.) 16-bit extending loads on 32-bit target to
1393 // 8 bits, but have to be careful...
1394 if (Lod->getExtensionType() != ISD::NON_EXTLOAD)
1395 origWidth = Lod->getMemoryVT().getSizeInBits();
1396 for (unsigned width = origWidth / 2; width>=8; width /= 2) {
1397 uint64_t newMask = (1ULL << width) - 1;
1398 for (unsigned offset=0; offset<origWidth/width; offset++) {
1399 if ((newMask & Mask)==Mask) {
1400 if (!TD->isLittleEndian())
1401 bestOffset = (origWidth/width - offset - 1) * (width/8);
1403 bestOffset = (uint64_t)offset * (width/8);
1404 bestMask = Mask >> (offset * (width/8) * 8);
1408 newMask = newMask << width;
1413 MVT newVT = MVT::getIntegerVT(bestWidth);
1414 if (newVT.isRound()) {
1415 MVT PtrType = Lod->getOperand(1).getValueType();
1416 SDValue Ptr = Lod->getBasePtr();
1417 if (bestOffset != 0)
1418 Ptr = DAG.getNode(ISD::ADD, PtrType, Lod->getBasePtr(),
1419 DAG.getConstant(bestOffset, PtrType));
1420 unsigned NewAlign = MinAlign(Lod->getAlignment(), bestOffset);
1421 SDValue NewLoad = DAG.getLoad(newVT, Lod->getChain(), Ptr,
1423 Lod->getSrcValueOffset() + bestOffset,
1425 return DAG.getSetCC(VT, DAG.getNode(ISD::AND, newVT, NewLoad,
1426 DAG.getConstant(bestMask, newVT)),
1427 DAG.getConstant(0LL, newVT), Cond);
1432 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
1433 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
1434 unsigned InSize = N0.getOperand(0).getValueType().getSizeInBits();
1436 // If the comparison constant has bits in the upper part, the
1437 // zero-extended value could never match.
1438 if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(),
1439 C1.getBitWidth() - InSize))) {
1443 case ISD::SETEQ: return DAG.getConstant(0, VT);
1446 case ISD::SETNE: return DAG.getConstant(1, VT);
1449 // True if the sign bit of C1 is set.
1450 return DAG.getConstant(C1.isNegative(), VT);
1453 // True if the sign bit of C1 isn't set.
1454 return DAG.getConstant(C1.isNonNegative(), VT);
1460 // Otherwise, we can perform the comparison with the low bits.
1468 return DAG.getSetCC(VT, N0.getOperand(0),
1469 DAG.getConstant(APInt(C1).trunc(InSize),
1470 N0.getOperand(0).getValueType()),
1473 break; // todo, be more careful with signed comparisons
1475 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
1476 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1477 MVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
1478 unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits();
1479 MVT ExtDstTy = N0.getValueType();
1480 unsigned ExtDstTyBits = ExtDstTy.getSizeInBits();
1482 // If the extended part has any inconsistent bits, it cannot ever
1483 // compare equal. In other words, they have to be all ones or all
1486 APInt::getHighBitsSet(ExtDstTyBits, ExtDstTyBits - ExtSrcTyBits);
1487 if ((C1 & ExtBits) != 0 && (C1 & ExtBits) != ExtBits)
1488 return DAG.getConstant(Cond == ISD::SETNE, VT);
1491 MVT Op0Ty = N0.getOperand(0).getValueType();
1492 if (Op0Ty == ExtSrcTy) {
1493 ZextOp = N0.getOperand(0);
1495 APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits);
1496 ZextOp = DAG.getNode(ISD::AND, Op0Ty, N0.getOperand(0),
1497 DAG.getConstant(Imm, Op0Ty));
1499 if (!DCI.isCalledByLegalizer())
1500 DCI.AddToWorklist(ZextOp.getNode());
1501 // Otherwise, make this a use of a zext.
1502 return DAG.getSetCC(VT, ZextOp,
1503 DAG.getConstant(C1 & APInt::getLowBitsSet(
1508 } else if ((N1C->isNullValue() || N1C->getAPIntValue() == 1) &&
1509 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1511 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC
1512 if (N0.getOpcode() == ISD::SETCC) {
1513 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getZExtValue() != 1);
1517 // Invert the condition.
1518 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
1519 CC = ISD::getSetCCInverse(CC,
1520 N0.getOperand(0).getValueType().isInteger());
1521 return DAG.getSetCC(VT, N0.getOperand(0), N0.getOperand(1), CC);
1524 if ((N0.getOpcode() == ISD::XOR ||
1525 (N0.getOpcode() == ISD::AND &&
1526 N0.getOperand(0).getOpcode() == ISD::XOR &&
1527 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
1528 isa<ConstantSDNode>(N0.getOperand(1)) &&
1529 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue() == 1) {
1530 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We
1531 // can only do this if the top bits are known zero.
1532 unsigned BitWidth = N0.getValueSizeInBits();
1533 if (DAG.MaskedValueIsZero(N0,
1534 APInt::getHighBitsSet(BitWidth,
1536 // Okay, get the un-inverted input value.
1538 if (N0.getOpcode() == ISD::XOR)
1539 Val = N0.getOperand(0);
1541 assert(N0.getOpcode() == ISD::AND &&
1542 N0.getOperand(0).getOpcode() == ISD::XOR);
1543 // ((X^1)&1)^1 -> X & 1
1544 Val = DAG.getNode(ISD::AND, N0.getValueType(),
1545 N0.getOperand(0).getOperand(0),
1548 return DAG.getSetCC(VT, Val, N1,
1549 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
1554 APInt MinVal, MaxVal;
1555 unsigned OperandBitSize = N1C->getValueType(0).getSizeInBits();
1556 if (ISD::isSignedIntSetCC(Cond)) {
1557 MinVal = APInt::getSignedMinValue(OperandBitSize);
1558 MaxVal = APInt::getSignedMaxValue(OperandBitSize);
1560 MinVal = APInt::getMinValue(OperandBitSize);
1561 MaxVal = APInt::getMaxValue(OperandBitSize);
1564 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
1565 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
1566 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true
1567 // X >= C0 --> X > (C0-1)
1568 return DAG.getSetCC(VT, N0, DAG.getConstant(C1-1, N1.getValueType()),
1569 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
1572 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
1573 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true
1574 // X <= C0 --> X < (C0+1)
1575 return DAG.getSetCC(VT, N0, DAG.getConstant(C1+1, N1.getValueType()),
1576 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
1579 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
1580 return DAG.getConstant(0, VT); // X < MIN --> false
1581 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal)
1582 return DAG.getConstant(1, VT); // X >= MIN --> true
1583 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal)
1584 return DAG.getConstant(0, VT); // X > MAX --> false
1585 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal)
1586 return DAG.getConstant(1, VT); // X <= MAX --> true
1588 // Canonicalize setgt X, Min --> setne X, Min
1589 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
1590 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
1591 // Canonicalize setlt X, Max --> setne X, Max
1592 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
1593 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
1595 // If we have setult X, 1, turn it into seteq X, 0
1596 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
1597 return DAG.getSetCC(VT, N0, DAG.getConstant(MinVal, N0.getValueType()),
1599 // If we have setugt X, Max-1, turn it into seteq X, Max
1600 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
1601 return DAG.getSetCC(VT, N0, DAG.getConstant(MaxVal, N0.getValueType()),
1604 // If we have "setcc X, C0", check to see if we can shrink the immediate
1607 // SETUGT X, SINTMAX -> SETLT X, 0
1608 if (Cond == ISD::SETUGT &&
1609 C1 == APInt::getSignedMaxValue(OperandBitSize))
1610 return DAG.getSetCC(VT, N0, DAG.getConstant(0, N1.getValueType()),
1613 // SETULT X, SINTMIN -> SETGT X, -1
1614 if (Cond == ISD::SETULT &&
1615 C1 == APInt::getSignedMinValue(OperandBitSize)) {
1616 SDValue ConstMinusOne =
1617 DAG.getConstant(APInt::getAllOnesValue(OperandBitSize),
1619 return DAG.getSetCC(VT, N0, ConstMinusOne, ISD::SETGT);
1622 // Fold bit comparisons when we can.
1623 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1624 VT == N0.getValueType() && N0.getOpcode() == ISD::AND)
1625 if (ConstantSDNode *AndRHS =
1626 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1627 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
1628 // Perform the xform if the AND RHS is a single bit.
1629 if (isPowerOf2_64(AndRHS->getZExtValue())) {
1630 return DAG.getNode(ISD::SRL, VT, N0,
1631 DAG.getConstant(Log2_64(AndRHS->getZExtValue()),
1632 getShiftAmountTy()));
1634 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getZExtValue()) {
1635 // (X & 8) == 8 --> (X & 8) >> 3
1636 // Perform the xform if C1 is a single bit.
1637 if (C1.isPowerOf2()) {
1638 return DAG.getNode(ISD::SRL, VT, N0,
1639 DAG.getConstant(C1.logBase2(), getShiftAmountTy()));
1644 } else if (isa<ConstantSDNode>(N0.getNode())) {
1645 // Ensure that the constant occurs on the RHS.
1646 return DAG.getSetCC(VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
1649 if (isa<ConstantFPSDNode>(N0.getNode())) {
1650 // Constant fold or commute setcc.
1651 SDValue O = DAG.FoldSetCC(VT, N0, N1, Cond);
1652 if (O.getNode()) return O;
1653 } else if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1.getNode())) {
1654 // If the RHS of an FP comparison is a constant, simplify it away in
1656 if (CFP->getValueAPF().isNaN()) {
1657 // If an operand is known to be a nan, we can fold it.
1658 switch (ISD::getUnorderedFlavor(Cond)) {
1659 default: assert(0 && "Unknown flavor!");
1660 case 0: // Known false.
1661 return DAG.getConstant(0, VT);
1662 case 1: // Known true.
1663 return DAG.getConstant(1, VT);
1664 case 2: // Undefined.
1665 return DAG.getNode(ISD::UNDEF, VT);
1669 // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the
1670 // constant if knowing that the operand is non-nan is enough. We prefer to
1671 // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to
1673 if (Cond == ISD::SETO || Cond == ISD::SETUO)
1674 return DAG.getSetCC(VT, N0, N0, Cond);
1678 // We can always fold X == X for integer setcc's.
1679 if (N0.getValueType().isInteger())
1680 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
1681 unsigned UOF = ISD::getUnorderedFlavor(Cond);
1682 if (UOF == 2) // FP operators that are undefined on NaNs.
1683 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
1684 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
1685 return DAG.getConstant(UOF, VT);
1686 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
1687 // if it is not already.
1688 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
1689 if (NewCond != Cond)
1690 return DAG.getSetCC(VT, N0, N1, NewCond);
1693 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1694 N0.getValueType().isInteger()) {
1695 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
1696 N0.getOpcode() == ISD::XOR) {
1697 // Simplify (X+Y) == (X+Z) --> Y == Z
1698 if (N0.getOpcode() == N1.getOpcode()) {
1699 if (N0.getOperand(0) == N1.getOperand(0))
1700 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(1), Cond);
1701 if (N0.getOperand(1) == N1.getOperand(1))
1702 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(0), Cond);
1703 if (DAG.isCommutativeBinOp(N0.getOpcode())) {
1704 // If X op Y == Y op X, try other combinations.
1705 if (N0.getOperand(0) == N1.getOperand(1))
1706 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(0), Cond);
1707 if (N0.getOperand(1) == N1.getOperand(0))
1708 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(1), Cond);
1712 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) {
1713 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1714 // Turn (X+C1) == C2 --> X == C2-C1
1715 if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse()) {
1716 return DAG.getSetCC(VT, N0.getOperand(0),
1717 DAG.getConstant(RHSC->getAPIntValue()-
1718 LHSR->getAPIntValue(),
1719 N0.getValueType()), Cond);
1722 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
1723 if (N0.getOpcode() == ISD::XOR)
1724 // If we know that all of the inverted bits are zero, don't bother
1725 // performing the inversion.
1726 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue()))
1728 DAG.getSetCC(VT, N0.getOperand(0),
1729 DAG.getConstant(LHSR->getAPIntValue() ^
1730 RHSC->getAPIntValue(),
1735 // Turn (C1-X) == C2 --> X == C1-C2
1736 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
1737 if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse()) {
1739 DAG.getSetCC(VT, N0.getOperand(1),
1740 DAG.getConstant(SUBC->getAPIntValue() -
1741 RHSC->getAPIntValue(),
1748 // Simplify (X+Z) == X --> Z == 0
1749 if (N0.getOperand(0) == N1)
1750 return DAG.getSetCC(VT, N0.getOperand(1),
1751 DAG.getConstant(0, N0.getValueType()), Cond);
1752 if (N0.getOperand(1) == N1) {
1753 if (DAG.isCommutativeBinOp(N0.getOpcode()))
1754 return DAG.getSetCC(VT, N0.getOperand(0),
1755 DAG.getConstant(0, N0.getValueType()), Cond);
1756 else if (N0.getNode()->hasOneUse()) {
1757 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
1758 // (Z-X) == X --> Z == X<<1
1759 SDValue SH = DAG.getNode(ISD::SHL, N1.getValueType(),
1761 DAG.getConstant(1, getShiftAmountTy()));
1762 if (!DCI.isCalledByLegalizer())
1763 DCI.AddToWorklist(SH.getNode());
1764 return DAG.getSetCC(VT, N0.getOperand(0), SH, Cond);
1769 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
1770 N1.getOpcode() == ISD::XOR) {
1771 // Simplify X == (X+Z) --> Z == 0
1772 if (N1.getOperand(0) == N0) {
1773 return DAG.getSetCC(VT, N1.getOperand(1),
1774 DAG.getConstant(0, N1.getValueType()), Cond);
1775 } else if (N1.getOperand(1) == N0) {
1776 if (DAG.isCommutativeBinOp(N1.getOpcode())) {
1777 return DAG.getSetCC(VT, N1.getOperand(0),
1778 DAG.getConstant(0, N1.getValueType()), Cond);
1779 } else if (N1.getNode()->hasOneUse()) {
1780 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
1781 // X == (Z-X) --> X<<1 == Z
1782 SDValue SH = DAG.getNode(ISD::SHL, N1.getValueType(), N0,
1783 DAG.getConstant(1, getShiftAmountTy()));
1784 if (!DCI.isCalledByLegalizer())
1785 DCI.AddToWorklist(SH.getNode());
1786 return DAG.getSetCC(VT, SH, N1.getOperand(0), Cond);
1792 // Fold away ALL boolean setcc's.
1794 if (N0.getValueType() == MVT::i1 && foldBooleans) {
1796 default: assert(0 && "Unknown integer setcc!");
1797 case ISD::SETEQ: // X == Y -> (X^Y)^1
1798 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
1799 N0 = DAG.getNode(ISD::XOR, MVT::i1, Temp, DAG.getConstant(1, MVT::i1));
1800 if (!DCI.isCalledByLegalizer())
1801 DCI.AddToWorklist(Temp.getNode());
1803 case ISD::SETNE: // X != Y --> (X^Y)
1804 N0 = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
1806 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> X^1 & Y
1807 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> X^1 & Y
1808 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
1809 N0 = DAG.getNode(ISD::AND, MVT::i1, N1, Temp);
1810 if (!DCI.isCalledByLegalizer())
1811 DCI.AddToWorklist(Temp.getNode());
1813 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> Y^1 & X
1814 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> Y^1 & X
1815 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
1816 N0 = DAG.getNode(ISD::AND, MVT::i1, N0, Temp);
1817 if (!DCI.isCalledByLegalizer())
1818 DCI.AddToWorklist(Temp.getNode());
1820 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> X^1 | Y
1821 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> X^1 | Y
1822 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
1823 N0 = DAG.getNode(ISD::OR, MVT::i1, N1, Temp);
1824 if (!DCI.isCalledByLegalizer())
1825 DCI.AddToWorklist(Temp.getNode());
1827 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> Y^1 | X
1828 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> Y^1 | X
1829 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
1830 N0 = DAG.getNode(ISD::OR, MVT::i1, N0, Temp);
1833 if (VT != MVT::i1) {
1834 if (!DCI.isCalledByLegalizer())
1835 DCI.AddToWorklist(N0.getNode());
1836 // FIXME: If running after legalize, we probably can't do this.
1837 N0 = DAG.getNode(ISD::ZERO_EXTEND, VT, N0);
1842 // Could not fold it.
1846 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
1847 /// node is a GlobalAddress + offset.
1848 bool TargetLowering::isGAPlusOffset(SDNode *N, GlobalValue* &GA,
1849 int64_t &Offset) const {
1850 if (isa<GlobalAddressSDNode>(N)) {
1851 GlobalAddressSDNode *GASD = cast<GlobalAddressSDNode>(N);
1852 GA = GASD->getGlobal();
1853 Offset += GASD->getOffset();
1857 if (N->getOpcode() == ISD::ADD) {
1858 SDValue N1 = N->getOperand(0);
1859 SDValue N2 = N->getOperand(1);
1860 if (isGAPlusOffset(N1.getNode(), GA, Offset)) {
1861 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
1863 Offset += V->getSExtValue();
1866 } else if (isGAPlusOffset(N2.getNode(), GA, Offset)) {
1867 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
1869 Offset += V->getSExtValue();
1878 /// isConsecutiveLoad - Return true if LD (which must be a LoadSDNode) is
1879 /// loading 'Bytes' bytes from a location that is 'Dist' units away from the
1880 /// location that the 'Base' load is loading from.
1881 bool TargetLowering::isConsecutiveLoad(SDNode *LD, SDNode *Base,
1882 unsigned Bytes, int Dist,
1883 const MachineFrameInfo *MFI) const {
1884 if (LD->getOperand(0).getNode() != Base->getOperand(0).getNode())
1886 MVT VT = LD->getValueType(0);
1887 if (VT.getSizeInBits() / 8 != Bytes)
1890 SDValue Loc = LD->getOperand(1);
1891 SDValue BaseLoc = Base->getOperand(1);
1892 if (Loc.getOpcode() == ISD::FrameIndex) {
1893 if (BaseLoc.getOpcode() != ISD::FrameIndex)
1895 int FI = cast<FrameIndexSDNode>(Loc)->getIndex();
1896 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
1897 int FS = MFI->getObjectSize(FI);
1898 int BFS = MFI->getObjectSize(BFI);
1899 if (FS != BFS || FS != (int)Bytes) return false;
1900 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Bytes);
1903 GlobalValue *GV1 = NULL;
1904 GlobalValue *GV2 = NULL;
1905 int64_t Offset1 = 0;
1906 int64_t Offset2 = 0;
1907 bool isGA1 = isGAPlusOffset(Loc.getNode(), GV1, Offset1);
1908 bool isGA2 = isGAPlusOffset(BaseLoc.getNode(), GV2, Offset2);
1909 if (isGA1 && isGA2 && GV1 == GV2)
1910 return Offset1 == (Offset2 + Dist*Bytes);
1915 SDValue TargetLowering::
1916 PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
1917 // Default implementation: no optimization.
1921 //===----------------------------------------------------------------------===//
1922 // Inline Assembler Implementation Methods
1923 //===----------------------------------------------------------------------===//
1926 TargetLowering::ConstraintType
1927 TargetLowering::getConstraintType(const std::string &Constraint) const {
1928 // FIXME: lots more standard ones to handle.
1929 if (Constraint.size() == 1) {
1930 switch (Constraint[0]) {
1932 case 'r': return C_RegisterClass;
1934 case 'o': // offsetable
1935 case 'V': // not offsetable
1937 case 'i': // Simple Integer or Relocatable Constant
1938 case 'n': // Simple Integer
1939 case 's': // Relocatable Constant
1940 case 'X': // Allow ANY value.
1941 case 'I': // Target registers.
1953 if (Constraint.size() > 1 && Constraint[0] == '{' &&
1954 Constraint[Constraint.size()-1] == '}')
1959 /// LowerXConstraint - try to replace an X constraint, which matches anything,
1960 /// with another that has more specific requirements based on the type of the
1961 /// corresponding operand.
1962 const char *TargetLowering::LowerXConstraint(MVT ConstraintVT) const{
1963 if (ConstraintVT.isInteger())
1965 if (ConstraintVT.isFloatingPoint())
1966 return "f"; // works for many targets
1970 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
1971 /// vector. If it is invalid, don't add anything to Ops.
1972 void TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
1973 char ConstraintLetter,
1975 std::vector<SDValue> &Ops,
1976 SelectionDAG &DAG) const {
1977 switch (ConstraintLetter) {
1979 case 'X': // Allows any operand; labels (basic block) use this.
1980 if (Op.getOpcode() == ISD::BasicBlock) {
1985 case 'i': // Simple Integer or Relocatable Constant
1986 case 'n': // Simple Integer
1987 case 's': { // Relocatable Constant
1988 // These operands are interested in values of the form (GV+C), where C may
1989 // be folded in as an offset of GV, or it may be explicitly added. Also, it
1990 // is possible and fine if either GV or C are missing.
1991 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
1992 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
1994 // If we have "(add GV, C)", pull out GV/C
1995 if (Op.getOpcode() == ISD::ADD) {
1996 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
1997 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
1998 if (C == 0 || GA == 0) {
1999 C = dyn_cast<ConstantSDNode>(Op.getOperand(0));
2000 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1));
2002 if (C == 0 || GA == 0)
2006 // If we find a valid operand, map to the TargetXXX version so that the
2007 // value itself doesn't get selected.
2008 if (GA) { // Either &GV or &GV+C
2009 if (ConstraintLetter != 'n') {
2010 int64_t Offs = GA->getOffset();
2011 if (C) Offs += C->getZExtValue();
2012 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(),
2013 Op.getValueType(), Offs));
2017 if (C) { // just C, no GV.
2018 // Simple constants are not allowed for 's'.
2019 if (ConstraintLetter != 's') {
2020 Ops.push_back(DAG.getTargetConstant(C->getAPIntValue(),
2021 Op.getValueType()));
2030 std::vector<unsigned> TargetLowering::
2031 getRegClassForInlineAsmConstraint(const std::string &Constraint,
2033 return std::vector<unsigned>();
2037 std::pair<unsigned, const TargetRegisterClass*> TargetLowering::
2038 getRegForInlineAsmConstraint(const std::string &Constraint,
2040 if (Constraint[0] != '{')
2041 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
2042 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?");
2044 // Remove the braces from around the name.
2045 std::string RegName(Constraint.begin()+1, Constraint.end()-1);
2047 // Figure out which register class contains this reg.
2048 const TargetRegisterInfo *RI = TM.getRegisterInfo();
2049 for (TargetRegisterInfo::regclass_iterator RCI = RI->regclass_begin(),
2050 E = RI->regclass_end(); RCI != E; ++RCI) {
2051 const TargetRegisterClass *RC = *RCI;
2053 // If none of the the value types for this register class are valid, we
2054 // can't use it. For example, 64-bit reg classes on 32-bit targets.
2055 bool isLegal = false;
2056 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
2058 if (isTypeLegal(*I)) {
2064 if (!isLegal) continue;
2066 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end();
2068 if (StringsEqualNoCase(RegName, RI->get(*I).AsmName))
2069 return std::make_pair(*I, RC);
2073 return std::pair<unsigned, const TargetRegisterClass*>(0, 0);
2076 //===----------------------------------------------------------------------===//
2077 // Constraint Selection.
2079 /// isMatchingInputConstraint - Return true of this is an input operand that is
2080 /// a matching constraint like "4".
2081 bool TargetLowering::AsmOperandInfo::isMatchingInputConstraint() const {
2082 assert(!ConstraintCode.empty() && "No known constraint!");
2083 return isdigit(ConstraintCode[0]);
2086 /// getMatchedOperand - If this is an input matching constraint, this method
2087 /// returns the output operand it matches.
2088 unsigned TargetLowering::AsmOperandInfo::getMatchedOperand() const {
2089 assert(!ConstraintCode.empty() && "No known constraint!");
2090 return atoi(ConstraintCode.c_str());
2094 /// getConstraintGenerality - Return an integer indicating how general CT
2096 static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) {
2098 default: assert(0 && "Unknown constraint type!");
2099 case TargetLowering::C_Other:
2100 case TargetLowering::C_Unknown:
2102 case TargetLowering::C_Register:
2104 case TargetLowering::C_RegisterClass:
2106 case TargetLowering::C_Memory:
2111 /// ChooseConstraint - If there are multiple different constraints that we
2112 /// could pick for this operand (e.g. "imr") try to pick the 'best' one.
2113 /// This is somewhat tricky: constraints fall into four classes:
2114 /// Other -> immediates and magic values
2115 /// Register -> one specific register
2116 /// RegisterClass -> a group of regs
2117 /// Memory -> memory
2118 /// Ideally, we would pick the most specific constraint possible: if we have
2119 /// something that fits into a register, we would pick it. The problem here
2120 /// is that if we have something that could either be in a register or in
2121 /// memory that use of the register could cause selection of *other*
2122 /// operands to fail: they might only succeed if we pick memory. Because of
2123 /// this the heuristic we use is:
2125 /// 1) If there is an 'other' constraint, and if the operand is valid for
2126 /// that constraint, use it. This makes us take advantage of 'i'
2127 /// constraints when available.
2128 /// 2) Otherwise, pick the most general constraint present. This prefers
2129 /// 'm' over 'r', for example.
2131 static void ChooseConstraint(TargetLowering::AsmOperandInfo &OpInfo,
2132 bool hasMemory, const TargetLowering &TLI,
2133 SDValue Op, SelectionDAG *DAG) {
2134 assert(OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options");
2135 unsigned BestIdx = 0;
2136 TargetLowering::ConstraintType BestType = TargetLowering::C_Unknown;
2137 int BestGenerality = -1;
2139 // Loop over the options, keeping track of the most general one.
2140 for (unsigned i = 0, e = OpInfo.Codes.size(); i != e; ++i) {
2141 TargetLowering::ConstraintType CType =
2142 TLI.getConstraintType(OpInfo.Codes[i]);
2144 // If this is an 'other' constraint, see if the operand is valid for it.
2145 // For example, on X86 we might have an 'rI' constraint. If the operand
2146 // is an integer in the range [0..31] we want to use I (saving a load
2147 // of a register), otherwise we must use 'r'.
2148 if (CType == TargetLowering::C_Other && Op.getNode()) {
2149 assert(OpInfo.Codes[i].size() == 1 &&
2150 "Unhandled multi-letter 'other' constraint");
2151 std::vector<SDValue> ResultOps;
2152 TLI.LowerAsmOperandForConstraint(Op, OpInfo.Codes[i][0], hasMemory,
2154 if (!ResultOps.empty()) {
2161 // This constraint letter is more general than the previous one, use it.
2162 int Generality = getConstraintGenerality(CType);
2163 if (Generality > BestGenerality) {
2166 BestGenerality = Generality;
2170 OpInfo.ConstraintCode = OpInfo.Codes[BestIdx];
2171 OpInfo.ConstraintType = BestType;
2174 /// ComputeConstraintToUse - Determines the constraint code and constraint
2175 /// type to use for the specific AsmOperandInfo, setting
2176 /// OpInfo.ConstraintCode and OpInfo.ConstraintType.
2177 void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo,
2180 SelectionDAG *DAG) const {
2181 assert(!OpInfo.Codes.empty() && "Must have at least one constraint");
2183 // Single-letter constraints ('r') are very common.
2184 if (OpInfo.Codes.size() == 1) {
2185 OpInfo.ConstraintCode = OpInfo.Codes[0];
2186 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
2188 ChooseConstraint(OpInfo, hasMemory, *this, Op, DAG);
2191 // 'X' matches anything.
2192 if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) {
2193 // Labels and constants are handled elsewhere ('X' is the only thing
2194 // that matches labels).
2195 if (isa<BasicBlock>(OpInfo.CallOperandVal) ||
2196 isa<ConstantInt>(OpInfo.CallOperandVal))
2199 // Otherwise, try to resolve it to something we know about by looking at
2200 // the actual operand type.
2201 if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) {
2202 OpInfo.ConstraintCode = Repl;
2203 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
2208 //===----------------------------------------------------------------------===//
2209 // Loop Strength Reduction hooks
2210 //===----------------------------------------------------------------------===//
2212 /// isLegalAddressingMode - Return true if the addressing mode represented
2213 /// by AM is legal for this target, for a load/store of the specified type.
2214 bool TargetLowering::isLegalAddressingMode(const AddrMode &AM,
2215 const Type *Ty) const {
2216 // The default implementation of this implements a conservative RISCy, r+r and
2219 // Allows a sign-extended 16-bit immediate field.
2220 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
2223 // No global is ever allowed as a base.
2227 // Only support r+r,
2229 case 0: // "r+i" or just "i", depending on HasBaseReg.
2232 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
2234 // Otherwise we have r+r or r+i.
2237 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
2239 // Allow 2*r as r+r.
2247 APInt m; // magic number
2248 bool a; // add indicator
2249 unsigned s; // shift amount
2252 /// magicu - calculate the magic numbers required to codegen an integer udiv as
2253 /// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
2254 static mu magicu(const APInt& d) {
2256 APInt nc, delta, q1, r1, q2, r2;
2258 magu.a = 0; // initialize "add" indicator
2259 APInt allOnes = APInt::getAllOnesValue(d.getBitWidth());
2260 APInt signedMin = APInt::getSignedMinValue(d.getBitWidth());
2261 APInt signedMax = APInt::getSignedMaxValue(d.getBitWidth());
2263 nc = allOnes - (-d).urem(d);
2264 p = d.getBitWidth() - 1; // initialize p
2265 q1 = signedMin.udiv(nc); // initialize q1 = 2p/nc
2266 r1 = signedMin - q1*nc; // initialize r1 = rem(2p,nc)
2267 q2 = signedMax.udiv(d); // initialize q2 = (2p-1)/d
2268 r2 = signedMax - q2*d; // initialize r2 = rem((2p-1),d)
2271 if (r1.uge(nc - r1)) {
2272 q1 = q1 + q1 + 1; // update q1
2273 r1 = r1 + r1 - nc; // update r1
2276 q1 = q1+q1; // update q1
2277 r1 = r1+r1; // update r1
2279 if ((r2 + 1).uge(d - r2)) {
2280 if (q2.uge(signedMax)) magu.a = 1;
2281 q2 = q2+q2 + 1; // update q2
2282 r2 = r2+r2 + 1 - d; // update r2
2285 if (q2.uge(signedMin)) magu.a = 1;
2286 q2 = q2+q2; // update q2
2287 r2 = r2+r2 + 1; // update r2
2290 } while (p < d.getBitWidth()*2 &&
2291 (q1.ult(delta) || (q1 == delta && r1 == 0)));
2292 magu.m = q2 + 1; // resulting magic number
2293 magu.s = p - d.getBitWidth(); // resulting shift
2297 // Magic for divide replacement
2299 APInt m; // magic number
2300 unsigned s; // shift amount
2303 /// magic - calculate the magic numbers required to codegen an integer sdiv as
2304 /// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
2306 static ms magic(const APInt& d) {
2308 APInt ad, anc, delta, q1, r1, q2, r2, t;
2309 APInt allOnes = APInt::getAllOnesValue(d.getBitWidth());
2310 APInt signedMin = APInt::getSignedMinValue(d.getBitWidth());
2311 APInt signedMax = APInt::getSignedMaxValue(d.getBitWidth());
2315 t = signedMin + (d.lshr(d.getBitWidth() - 1));
2316 anc = t - 1 - t.urem(ad); // absolute value of nc
2317 p = d.getBitWidth() - 1; // initialize p
2318 q1 = signedMin.udiv(anc); // initialize q1 = 2p/abs(nc)
2319 r1 = signedMin - q1*anc; // initialize r1 = rem(2p,abs(nc))
2320 q2 = signedMin.udiv(ad); // initialize q2 = 2p/abs(d)
2321 r2 = signedMin - q2*ad; // initialize r2 = rem(2p,abs(d))
2324 q1 = q1<<1; // update q1 = 2p/abs(nc)
2325 r1 = r1<<1; // update r1 = rem(2p/abs(nc))
2326 if (r1.uge(anc)) { // must be unsigned comparison
2330 q2 = q2<<1; // update q2 = 2p/abs(d)
2331 r2 = r2<<1; // update r2 = rem(2p/abs(d))
2332 if (r2.uge(ad)) { // must be unsigned comparison
2337 } while (q1.ule(delta) || (q1 == delta && r1 == 0));
2340 if (d.isNegative()) mag.m = -mag.m; // resulting magic number
2341 mag.s = p - d.getBitWidth(); // resulting shift
2345 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
2346 /// return a DAG expression to select that will generate the same value by
2347 /// multiplying by a magic number. See:
2348 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
2349 SDValue TargetLowering::BuildSDIV(SDNode *N, SelectionDAG &DAG,
2350 std::vector<SDNode*>* Created) const {
2351 MVT VT = N->getValueType(0);
2353 // Check to see if we can do this.
2354 // FIXME: We should be more aggressive here.
2355 if (!isTypeLegal(VT))
2358 APInt d = cast<ConstantSDNode>(N->getOperand(1))->getAPIntValue();
2359 ms magics = magic(d);
2361 // Multiply the numerator (operand 0) by the magic value
2362 // FIXME: We should support doing a MUL in a wider type
2364 if (isOperationLegal(ISD::MULHS, VT))
2365 Q = DAG.getNode(ISD::MULHS, VT, N->getOperand(0),
2366 DAG.getConstant(magics.m, VT));
2367 else if (isOperationLegal(ISD::SMUL_LOHI, VT))
2368 Q = SDValue(DAG.getNode(ISD::SMUL_LOHI, DAG.getVTList(VT, VT),
2370 DAG.getConstant(magics.m, VT)).getNode(), 1);
2372 return SDValue(); // No mulhs or equvialent
2373 // If d > 0 and m < 0, add the numerator
2374 if (d.isStrictlyPositive() && magics.m.isNegative()) {
2375 Q = DAG.getNode(ISD::ADD, VT, Q, N->getOperand(0));
2377 Created->push_back(Q.getNode());
2379 // If d < 0 and m > 0, subtract the numerator.
2380 if (d.isNegative() && magics.m.isStrictlyPositive()) {
2381 Q = DAG.getNode(ISD::SUB, VT, Q, N->getOperand(0));
2383 Created->push_back(Q.getNode());
2385 // Shift right algebraic if shift value is nonzero
2387 Q = DAG.getNode(ISD::SRA, VT, Q,
2388 DAG.getConstant(magics.s, getShiftAmountTy()));
2390 Created->push_back(Q.getNode());
2392 // Extract the sign bit and add it to the quotient
2394 DAG.getNode(ISD::SRL, VT, Q, DAG.getConstant(VT.getSizeInBits()-1,
2395 getShiftAmountTy()));
2397 Created->push_back(T.getNode());
2398 return DAG.getNode(ISD::ADD, VT, Q, T);
2401 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
2402 /// return a DAG expression to select that will generate the same value by
2403 /// multiplying by a magic number. See:
2404 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
2405 SDValue TargetLowering::BuildUDIV(SDNode *N, SelectionDAG &DAG,
2406 std::vector<SDNode*>* Created) const {
2407 MVT VT = N->getValueType(0);
2409 // Check to see if we can do this.
2410 // FIXME: We should be more aggressive here.
2411 if (!isTypeLegal(VT))
2414 // FIXME: We should use a narrower constant when the upper
2415 // bits are known to be zero.
2416 ConstantSDNode *N1C = cast<ConstantSDNode>(N->getOperand(1));
2417 mu magics = magicu(N1C->getAPIntValue());
2419 // Multiply the numerator (operand 0) by the magic value
2420 // FIXME: We should support doing a MUL in a wider type
2422 if (isOperationLegal(ISD::MULHU, VT))
2423 Q = DAG.getNode(ISD::MULHU, VT, N->getOperand(0),
2424 DAG.getConstant(magics.m, VT));
2425 else if (isOperationLegal(ISD::UMUL_LOHI, VT))
2426 Q = SDValue(DAG.getNode(ISD::UMUL_LOHI, DAG.getVTList(VT, VT),
2428 DAG.getConstant(magics.m, VT)).getNode(), 1);
2430 return SDValue(); // No mulhu or equvialent
2432 Created->push_back(Q.getNode());
2434 if (magics.a == 0) {
2435 assert(magics.s < N1C->getAPIntValue().getBitWidth() &&
2436 "We shouldn't generate an undefined shift!");
2437 return DAG.getNode(ISD::SRL, VT, Q,
2438 DAG.getConstant(magics.s, getShiftAmountTy()));
2440 SDValue NPQ = DAG.getNode(ISD::SUB, VT, N->getOperand(0), Q);
2442 Created->push_back(NPQ.getNode());
2443 NPQ = DAG.getNode(ISD::SRL, VT, NPQ,
2444 DAG.getConstant(1, getShiftAmountTy()));
2446 Created->push_back(NPQ.getNode());
2447 NPQ = DAG.getNode(ISD::ADD, VT, NPQ, Q);
2449 Created->push_back(NPQ.getNode());
2450 return DAG.getNode(ISD::SRL, VT, NPQ,
2451 DAG.getConstant(magics.s-1, getShiftAmountTy()));