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/TargetLowering.h"
15 #include "llvm/MC/MCAsmInfo.h"
16 #include "llvm/MC/MCExpr.h"
17 #include "llvm/Target/TargetData.h"
18 #include "llvm/Target/TargetLoweringObjectFile.h"
19 #include "llvm/Target/TargetMachine.h"
20 #include "llvm/Target/TargetRegisterInfo.h"
21 #include "llvm/GlobalVariable.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/CodeGen/Analysis.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineJumpTableInfo.h"
26 #include "llvm/CodeGen/MachineFunction.h"
27 #include "llvm/CodeGen/SelectionDAG.h"
28 #include "llvm/ADT/STLExtras.h"
29 #include "llvm/Support/ErrorHandling.h"
30 #include "llvm/Support/MathExtras.h"
35 TLSModel::Model getTLSModel(const GlobalValue *GV, Reloc::Model reloc) {
36 bool isLocal = GV->hasLocalLinkage();
37 bool isDeclaration = GV->isDeclaration();
38 // FIXME: what should we do for protected and internal visibility?
39 // For variables, is internal different from hidden?
40 bool isHidden = GV->hasHiddenVisibility();
42 if (reloc == Reloc::PIC_) {
43 if (isLocal || isHidden)
44 return TLSModel::LocalDynamic;
46 return TLSModel::GeneralDynamic;
48 if (!isDeclaration || isHidden)
49 return TLSModel::LocalExec;
51 return TLSModel::InitialExec;
56 /// InitLibcallNames - Set default libcall names.
58 static void InitLibcallNames(const char **Names) {
59 Names[RTLIB::SHL_I16] = "__ashlhi3";
60 Names[RTLIB::SHL_I32] = "__ashlsi3";
61 Names[RTLIB::SHL_I64] = "__ashldi3";
62 Names[RTLIB::SHL_I128] = "__ashlti3";
63 Names[RTLIB::SRL_I16] = "__lshrhi3";
64 Names[RTLIB::SRL_I32] = "__lshrsi3";
65 Names[RTLIB::SRL_I64] = "__lshrdi3";
66 Names[RTLIB::SRL_I128] = "__lshrti3";
67 Names[RTLIB::SRA_I16] = "__ashrhi3";
68 Names[RTLIB::SRA_I32] = "__ashrsi3";
69 Names[RTLIB::SRA_I64] = "__ashrdi3";
70 Names[RTLIB::SRA_I128] = "__ashrti3";
71 Names[RTLIB::MUL_I8] = "__mulqi3";
72 Names[RTLIB::MUL_I16] = "__mulhi3";
73 Names[RTLIB::MUL_I32] = "__mulsi3";
74 Names[RTLIB::MUL_I64] = "__muldi3";
75 Names[RTLIB::MUL_I128] = "__multi3";
76 Names[RTLIB::SDIV_I8] = "__divqi3";
77 Names[RTLIB::SDIV_I16] = "__divhi3";
78 Names[RTLIB::SDIV_I32] = "__divsi3";
79 Names[RTLIB::SDIV_I64] = "__divdi3";
80 Names[RTLIB::SDIV_I128] = "__divti3";
81 Names[RTLIB::UDIV_I8] = "__udivqi3";
82 Names[RTLIB::UDIV_I16] = "__udivhi3";
83 Names[RTLIB::UDIV_I32] = "__udivsi3";
84 Names[RTLIB::UDIV_I64] = "__udivdi3";
85 Names[RTLIB::UDIV_I128] = "__udivti3";
86 Names[RTLIB::SREM_I8] = "__modqi3";
87 Names[RTLIB::SREM_I16] = "__modhi3";
88 Names[RTLIB::SREM_I32] = "__modsi3";
89 Names[RTLIB::SREM_I64] = "__moddi3";
90 Names[RTLIB::SREM_I128] = "__modti3";
91 Names[RTLIB::UREM_I8] = "__umodqi3";
92 Names[RTLIB::UREM_I16] = "__umodhi3";
93 Names[RTLIB::UREM_I32] = "__umodsi3";
94 Names[RTLIB::UREM_I64] = "__umoddi3";
95 Names[RTLIB::UREM_I128] = "__umodti3";
96 Names[RTLIB::NEG_I32] = "__negsi2";
97 Names[RTLIB::NEG_I64] = "__negdi2";
98 Names[RTLIB::ADD_F32] = "__addsf3";
99 Names[RTLIB::ADD_F64] = "__adddf3";
100 Names[RTLIB::ADD_F80] = "__addxf3";
101 Names[RTLIB::ADD_PPCF128] = "__gcc_qadd";
102 Names[RTLIB::SUB_F32] = "__subsf3";
103 Names[RTLIB::SUB_F64] = "__subdf3";
104 Names[RTLIB::SUB_F80] = "__subxf3";
105 Names[RTLIB::SUB_PPCF128] = "__gcc_qsub";
106 Names[RTLIB::MUL_F32] = "__mulsf3";
107 Names[RTLIB::MUL_F64] = "__muldf3";
108 Names[RTLIB::MUL_F80] = "__mulxf3";
109 Names[RTLIB::MUL_PPCF128] = "__gcc_qmul";
110 Names[RTLIB::DIV_F32] = "__divsf3";
111 Names[RTLIB::DIV_F64] = "__divdf3";
112 Names[RTLIB::DIV_F80] = "__divxf3";
113 Names[RTLIB::DIV_PPCF128] = "__gcc_qdiv";
114 Names[RTLIB::REM_F32] = "fmodf";
115 Names[RTLIB::REM_F64] = "fmod";
116 Names[RTLIB::REM_F80] = "fmodl";
117 Names[RTLIB::REM_PPCF128] = "fmodl";
118 Names[RTLIB::POWI_F32] = "__powisf2";
119 Names[RTLIB::POWI_F64] = "__powidf2";
120 Names[RTLIB::POWI_F80] = "__powixf2";
121 Names[RTLIB::POWI_PPCF128] = "__powitf2";
122 Names[RTLIB::SQRT_F32] = "sqrtf";
123 Names[RTLIB::SQRT_F64] = "sqrt";
124 Names[RTLIB::SQRT_F80] = "sqrtl";
125 Names[RTLIB::SQRT_PPCF128] = "sqrtl";
126 Names[RTLIB::LOG_F32] = "logf";
127 Names[RTLIB::LOG_F64] = "log";
128 Names[RTLIB::LOG_F80] = "logl";
129 Names[RTLIB::LOG_PPCF128] = "logl";
130 Names[RTLIB::LOG2_F32] = "log2f";
131 Names[RTLIB::LOG2_F64] = "log2";
132 Names[RTLIB::LOG2_F80] = "log2l";
133 Names[RTLIB::LOG2_PPCF128] = "log2l";
134 Names[RTLIB::LOG10_F32] = "log10f";
135 Names[RTLIB::LOG10_F64] = "log10";
136 Names[RTLIB::LOG10_F80] = "log10l";
137 Names[RTLIB::LOG10_PPCF128] = "log10l";
138 Names[RTLIB::EXP_F32] = "expf";
139 Names[RTLIB::EXP_F64] = "exp";
140 Names[RTLIB::EXP_F80] = "expl";
141 Names[RTLIB::EXP_PPCF128] = "expl";
142 Names[RTLIB::EXP2_F32] = "exp2f";
143 Names[RTLIB::EXP2_F64] = "exp2";
144 Names[RTLIB::EXP2_F80] = "exp2l";
145 Names[RTLIB::EXP2_PPCF128] = "exp2l";
146 Names[RTLIB::SIN_F32] = "sinf";
147 Names[RTLIB::SIN_F64] = "sin";
148 Names[RTLIB::SIN_F80] = "sinl";
149 Names[RTLIB::SIN_PPCF128] = "sinl";
150 Names[RTLIB::COS_F32] = "cosf";
151 Names[RTLIB::COS_F64] = "cos";
152 Names[RTLIB::COS_F80] = "cosl";
153 Names[RTLIB::COS_PPCF128] = "cosl";
154 Names[RTLIB::POW_F32] = "powf";
155 Names[RTLIB::POW_F64] = "pow";
156 Names[RTLIB::POW_F80] = "powl";
157 Names[RTLIB::POW_PPCF128] = "powl";
158 Names[RTLIB::CEIL_F32] = "ceilf";
159 Names[RTLIB::CEIL_F64] = "ceil";
160 Names[RTLIB::CEIL_F80] = "ceill";
161 Names[RTLIB::CEIL_PPCF128] = "ceill";
162 Names[RTLIB::TRUNC_F32] = "truncf";
163 Names[RTLIB::TRUNC_F64] = "trunc";
164 Names[RTLIB::TRUNC_F80] = "truncl";
165 Names[RTLIB::TRUNC_PPCF128] = "truncl";
166 Names[RTLIB::RINT_F32] = "rintf";
167 Names[RTLIB::RINT_F64] = "rint";
168 Names[RTLIB::RINT_F80] = "rintl";
169 Names[RTLIB::RINT_PPCF128] = "rintl";
170 Names[RTLIB::NEARBYINT_F32] = "nearbyintf";
171 Names[RTLIB::NEARBYINT_F64] = "nearbyint";
172 Names[RTLIB::NEARBYINT_F80] = "nearbyintl";
173 Names[RTLIB::NEARBYINT_PPCF128] = "nearbyintl";
174 Names[RTLIB::FLOOR_F32] = "floorf";
175 Names[RTLIB::FLOOR_F64] = "floor";
176 Names[RTLIB::FLOOR_F80] = "floorl";
177 Names[RTLIB::FLOOR_PPCF128] = "floorl";
178 Names[RTLIB::COPYSIGN_F32] = "copysignf";
179 Names[RTLIB::COPYSIGN_F64] = "copysign";
180 Names[RTLIB::COPYSIGN_F80] = "copysignl";
181 Names[RTLIB::COPYSIGN_PPCF128] = "copysignl";
182 Names[RTLIB::FPEXT_F32_F64] = "__extendsfdf2";
183 Names[RTLIB::FPEXT_F16_F32] = "__gnu_h2f_ieee";
184 Names[RTLIB::FPROUND_F32_F16] = "__gnu_f2h_ieee";
185 Names[RTLIB::FPROUND_F64_F32] = "__truncdfsf2";
186 Names[RTLIB::FPROUND_F80_F32] = "__truncxfsf2";
187 Names[RTLIB::FPROUND_PPCF128_F32] = "__trunctfsf2";
188 Names[RTLIB::FPROUND_F80_F64] = "__truncxfdf2";
189 Names[RTLIB::FPROUND_PPCF128_F64] = "__trunctfdf2";
190 Names[RTLIB::FPTOSINT_F32_I8] = "__fixsfqi";
191 Names[RTLIB::FPTOSINT_F32_I16] = "__fixsfhi";
192 Names[RTLIB::FPTOSINT_F32_I32] = "__fixsfsi";
193 Names[RTLIB::FPTOSINT_F32_I64] = "__fixsfdi";
194 Names[RTLIB::FPTOSINT_F32_I128] = "__fixsfti";
195 Names[RTLIB::FPTOSINT_F64_I8] = "__fixdfqi";
196 Names[RTLIB::FPTOSINT_F64_I16] = "__fixdfhi";
197 Names[RTLIB::FPTOSINT_F64_I32] = "__fixdfsi";
198 Names[RTLIB::FPTOSINT_F64_I64] = "__fixdfdi";
199 Names[RTLIB::FPTOSINT_F64_I128] = "__fixdfti";
200 Names[RTLIB::FPTOSINT_F80_I32] = "__fixxfsi";
201 Names[RTLIB::FPTOSINT_F80_I64] = "__fixxfdi";
202 Names[RTLIB::FPTOSINT_F80_I128] = "__fixxfti";
203 Names[RTLIB::FPTOSINT_PPCF128_I32] = "__fixtfsi";
204 Names[RTLIB::FPTOSINT_PPCF128_I64] = "__fixtfdi";
205 Names[RTLIB::FPTOSINT_PPCF128_I128] = "__fixtfti";
206 Names[RTLIB::FPTOUINT_F32_I8] = "__fixunssfqi";
207 Names[RTLIB::FPTOUINT_F32_I16] = "__fixunssfhi";
208 Names[RTLIB::FPTOUINT_F32_I32] = "__fixunssfsi";
209 Names[RTLIB::FPTOUINT_F32_I64] = "__fixunssfdi";
210 Names[RTLIB::FPTOUINT_F32_I128] = "__fixunssfti";
211 Names[RTLIB::FPTOUINT_F64_I8] = "__fixunsdfqi";
212 Names[RTLIB::FPTOUINT_F64_I16] = "__fixunsdfhi";
213 Names[RTLIB::FPTOUINT_F64_I32] = "__fixunsdfsi";
214 Names[RTLIB::FPTOUINT_F64_I64] = "__fixunsdfdi";
215 Names[RTLIB::FPTOUINT_F64_I128] = "__fixunsdfti";
216 Names[RTLIB::FPTOUINT_F80_I32] = "__fixunsxfsi";
217 Names[RTLIB::FPTOUINT_F80_I64] = "__fixunsxfdi";
218 Names[RTLIB::FPTOUINT_F80_I128] = "__fixunsxfti";
219 Names[RTLIB::FPTOUINT_PPCF128_I32] = "__fixunstfsi";
220 Names[RTLIB::FPTOUINT_PPCF128_I64] = "__fixunstfdi";
221 Names[RTLIB::FPTOUINT_PPCF128_I128] = "__fixunstfti";
222 Names[RTLIB::SINTTOFP_I32_F32] = "__floatsisf";
223 Names[RTLIB::SINTTOFP_I32_F64] = "__floatsidf";
224 Names[RTLIB::SINTTOFP_I32_F80] = "__floatsixf";
225 Names[RTLIB::SINTTOFP_I32_PPCF128] = "__floatsitf";
226 Names[RTLIB::SINTTOFP_I64_F32] = "__floatdisf";
227 Names[RTLIB::SINTTOFP_I64_F64] = "__floatdidf";
228 Names[RTLIB::SINTTOFP_I64_F80] = "__floatdixf";
229 Names[RTLIB::SINTTOFP_I64_PPCF128] = "__floatditf";
230 Names[RTLIB::SINTTOFP_I128_F32] = "__floattisf";
231 Names[RTLIB::SINTTOFP_I128_F64] = "__floattidf";
232 Names[RTLIB::SINTTOFP_I128_F80] = "__floattixf";
233 Names[RTLIB::SINTTOFP_I128_PPCF128] = "__floattitf";
234 Names[RTLIB::UINTTOFP_I32_F32] = "__floatunsisf";
235 Names[RTLIB::UINTTOFP_I32_F64] = "__floatunsidf";
236 Names[RTLIB::UINTTOFP_I32_F80] = "__floatunsixf";
237 Names[RTLIB::UINTTOFP_I32_PPCF128] = "__floatunsitf";
238 Names[RTLIB::UINTTOFP_I64_F32] = "__floatundisf";
239 Names[RTLIB::UINTTOFP_I64_F64] = "__floatundidf";
240 Names[RTLIB::UINTTOFP_I64_F80] = "__floatundixf";
241 Names[RTLIB::UINTTOFP_I64_PPCF128] = "__floatunditf";
242 Names[RTLIB::UINTTOFP_I128_F32] = "__floatuntisf";
243 Names[RTLIB::UINTTOFP_I128_F64] = "__floatuntidf";
244 Names[RTLIB::UINTTOFP_I128_F80] = "__floatuntixf";
245 Names[RTLIB::UINTTOFP_I128_PPCF128] = "__floatuntitf";
246 Names[RTLIB::OEQ_F32] = "__eqsf2";
247 Names[RTLIB::OEQ_F64] = "__eqdf2";
248 Names[RTLIB::UNE_F32] = "__nesf2";
249 Names[RTLIB::UNE_F64] = "__nedf2";
250 Names[RTLIB::OGE_F32] = "__gesf2";
251 Names[RTLIB::OGE_F64] = "__gedf2";
252 Names[RTLIB::OLT_F32] = "__ltsf2";
253 Names[RTLIB::OLT_F64] = "__ltdf2";
254 Names[RTLIB::OLE_F32] = "__lesf2";
255 Names[RTLIB::OLE_F64] = "__ledf2";
256 Names[RTLIB::OGT_F32] = "__gtsf2";
257 Names[RTLIB::OGT_F64] = "__gtdf2";
258 Names[RTLIB::UO_F32] = "__unordsf2";
259 Names[RTLIB::UO_F64] = "__unorddf2";
260 Names[RTLIB::O_F32] = "__unordsf2";
261 Names[RTLIB::O_F64] = "__unorddf2";
262 Names[RTLIB::MEMCPY] = "memcpy";
263 Names[RTLIB::MEMMOVE] = "memmove";
264 Names[RTLIB::MEMSET] = "memset";
265 Names[RTLIB::UNWIND_RESUME] = "_Unwind_Resume";
266 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_1] = "__sync_val_compare_and_swap_1";
267 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_2] = "__sync_val_compare_and_swap_2";
268 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_4] = "__sync_val_compare_and_swap_4";
269 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_8] = "__sync_val_compare_and_swap_8";
270 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_1] = "__sync_lock_test_and_set_1";
271 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_2] = "__sync_lock_test_and_set_2";
272 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_4] = "__sync_lock_test_and_set_4";
273 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_8] = "__sync_lock_test_and_set_8";
274 Names[RTLIB::SYNC_FETCH_AND_ADD_1] = "__sync_fetch_and_add_1";
275 Names[RTLIB::SYNC_FETCH_AND_ADD_2] = "__sync_fetch_and_add_2";
276 Names[RTLIB::SYNC_FETCH_AND_ADD_4] = "__sync_fetch_and_add_4";
277 Names[RTLIB::SYNC_FETCH_AND_ADD_8] = "__sync_fetch_and_add_8";
278 Names[RTLIB::SYNC_FETCH_AND_SUB_1] = "__sync_fetch_and_sub_1";
279 Names[RTLIB::SYNC_FETCH_AND_SUB_2] = "__sync_fetch_and_sub_2";
280 Names[RTLIB::SYNC_FETCH_AND_SUB_4] = "__sync_fetch_and_sub_4";
281 Names[RTLIB::SYNC_FETCH_AND_SUB_8] = "__sync_fetch_and_sub_8";
282 Names[RTLIB::SYNC_FETCH_AND_AND_1] = "__sync_fetch_and_and_1";
283 Names[RTLIB::SYNC_FETCH_AND_AND_2] = "__sync_fetch_and_and_2";
284 Names[RTLIB::SYNC_FETCH_AND_AND_4] = "__sync_fetch_and_and_4";
285 Names[RTLIB::SYNC_FETCH_AND_AND_8] = "__sync_fetch_and_and_8";
286 Names[RTLIB::SYNC_FETCH_AND_OR_1] = "__sync_fetch_and_or_1";
287 Names[RTLIB::SYNC_FETCH_AND_OR_2] = "__sync_fetch_and_or_2";
288 Names[RTLIB::SYNC_FETCH_AND_OR_4] = "__sync_fetch_and_or_4";
289 Names[RTLIB::SYNC_FETCH_AND_OR_8] = "__sync_fetch_and_or_8";
290 Names[RTLIB::SYNC_FETCH_AND_XOR_1] = "__sync_fetch_and_xor_1";
291 Names[RTLIB::SYNC_FETCH_AND_XOR_2] = "__sync_fetch_and_xor_2";
292 Names[RTLIB::SYNC_FETCH_AND_XOR_4] = "__sync_fetch_and-xor_4";
293 Names[RTLIB::SYNC_FETCH_AND_XOR_8] = "__sync_fetch_and_xor_8";
294 Names[RTLIB::SYNC_FETCH_AND_NAND_1] = "__sync_fetch_and_nand_1";
295 Names[RTLIB::SYNC_FETCH_AND_NAND_2] = "__sync_fetch_and_nand_2";
296 Names[RTLIB::SYNC_FETCH_AND_NAND_4] = "__sync_fetch_and_nand_4";
297 Names[RTLIB::SYNC_FETCH_AND_NAND_8] = "__sync_fetch_and_nand_8";
300 /// InitLibcallCallingConvs - Set default libcall CallingConvs.
302 static void InitLibcallCallingConvs(CallingConv::ID *CCs) {
303 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
304 CCs[i] = CallingConv::C;
308 /// getFPEXT - Return the FPEXT_*_* value for the given types, or
309 /// UNKNOWN_LIBCALL if there is none.
310 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) {
311 if (OpVT == MVT::f32) {
312 if (RetVT == MVT::f64)
313 return FPEXT_F32_F64;
316 return UNKNOWN_LIBCALL;
319 /// getFPROUND - Return the FPROUND_*_* value for the given types, or
320 /// UNKNOWN_LIBCALL if there is none.
321 RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) {
322 if (RetVT == MVT::f32) {
323 if (OpVT == MVT::f64)
324 return FPROUND_F64_F32;
325 if (OpVT == MVT::f80)
326 return FPROUND_F80_F32;
327 if (OpVT == MVT::ppcf128)
328 return FPROUND_PPCF128_F32;
329 } else if (RetVT == MVT::f64) {
330 if (OpVT == MVT::f80)
331 return FPROUND_F80_F64;
332 if (OpVT == MVT::ppcf128)
333 return FPROUND_PPCF128_F64;
336 return UNKNOWN_LIBCALL;
339 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or
340 /// UNKNOWN_LIBCALL if there is none.
341 RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) {
342 if (OpVT == MVT::f32) {
343 if (RetVT == MVT::i8)
344 return FPTOSINT_F32_I8;
345 if (RetVT == MVT::i16)
346 return FPTOSINT_F32_I16;
347 if (RetVT == MVT::i32)
348 return FPTOSINT_F32_I32;
349 if (RetVT == MVT::i64)
350 return FPTOSINT_F32_I64;
351 if (RetVT == MVT::i128)
352 return FPTOSINT_F32_I128;
353 } else if (OpVT == MVT::f64) {
354 if (RetVT == MVT::i8)
355 return FPTOSINT_F64_I8;
356 if (RetVT == MVT::i16)
357 return FPTOSINT_F64_I16;
358 if (RetVT == MVT::i32)
359 return FPTOSINT_F64_I32;
360 if (RetVT == MVT::i64)
361 return FPTOSINT_F64_I64;
362 if (RetVT == MVT::i128)
363 return FPTOSINT_F64_I128;
364 } else if (OpVT == MVT::f80) {
365 if (RetVT == MVT::i32)
366 return FPTOSINT_F80_I32;
367 if (RetVT == MVT::i64)
368 return FPTOSINT_F80_I64;
369 if (RetVT == MVT::i128)
370 return FPTOSINT_F80_I128;
371 } else if (OpVT == MVT::ppcf128) {
372 if (RetVT == MVT::i32)
373 return FPTOSINT_PPCF128_I32;
374 if (RetVT == MVT::i64)
375 return FPTOSINT_PPCF128_I64;
376 if (RetVT == MVT::i128)
377 return FPTOSINT_PPCF128_I128;
379 return UNKNOWN_LIBCALL;
382 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or
383 /// UNKNOWN_LIBCALL if there is none.
384 RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) {
385 if (OpVT == MVT::f32) {
386 if (RetVT == MVT::i8)
387 return FPTOUINT_F32_I8;
388 if (RetVT == MVT::i16)
389 return FPTOUINT_F32_I16;
390 if (RetVT == MVT::i32)
391 return FPTOUINT_F32_I32;
392 if (RetVT == MVT::i64)
393 return FPTOUINT_F32_I64;
394 if (RetVT == MVT::i128)
395 return FPTOUINT_F32_I128;
396 } else if (OpVT == MVT::f64) {
397 if (RetVT == MVT::i8)
398 return FPTOUINT_F64_I8;
399 if (RetVT == MVT::i16)
400 return FPTOUINT_F64_I16;
401 if (RetVT == MVT::i32)
402 return FPTOUINT_F64_I32;
403 if (RetVT == MVT::i64)
404 return FPTOUINT_F64_I64;
405 if (RetVT == MVT::i128)
406 return FPTOUINT_F64_I128;
407 } else if (OpVT == MVT::f80) {
408 if (RetVT == MVT::i32)
409 return FPTOUINT_F80_I32;
410 if (RetVT == MVT::i64)
411 return FPTOUINT_F80_I64;
412 if (RetVT == MVT::i128)
413 return FPTOUINT_F80_I128;
414 } else if (OpVT == MVT::ppcf128) {
415 if (RetVT == MVT::i32)
416 return FPTOUINT_PPCF128_I32;
417 if (RetVT == MVT::i64)
418 return FPTOUINT_PPCF128_I64;
419 if (RetVT == MVT::i128)
420 return FPTOUINT_PPCF128_I128;
422 return UNKNOWN_LIBCALL;
425 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or
426 /// UNKNOWN_LIBCALL if there is none.
427 RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) {
428 if (OpVT == MVT::i32) {
429 if (RetVT == MVT::f32)
430 return SINTTOFP_I32_F32;
431 else if (RetVT == MVT::f64)
432 return SINTTOFP_I32_F64;
433 else if (RetVT == MVT::f80)
434 return SINTTOFP_I32_F80;
435 else if (RetVT == MVT::ppcf128)
436 return SINTTOFP_I32_PPCF128;
437 } else if (OpVT == MVT::i64) {
438 if (RetVT == MVT::f32)
439 return SINTTOFP_I64_F32;
440 else if (RetVT == MVT::f64)
441 return SINTTOFP_I64_F64;
442 else if (RetVT == MVT::f80)
443 return SINTTOFP_I64_F80;
444 else if (RetVT == MVT::ppcf128)
445 return SINTTOFP_I64_PPCF128;
446 } else if (OpVT == MVT::i128) {
447 if (RetVT == MVT::f32)
448 return SINTTOFP_I128_F32;
449 else if (RetVT == MVT::f64)
450 return SINTTOFP_I128_F64;
451 else if (RetVT == MVT::f80)
452 return SINTTOFP_I128_F80;
453 else if (RetVT == MVT::ppcf128)
454 return SINTTOFP_I128_PPCF128;
456 return UNKNOWN_LIBCALL;
459 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or
460 /// UNKNOWN_LIBCALL if there is none.
461 RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) {
462 if (OpVT == MVT::i32) {
463 if (RetVT == MVT::f32)
464 return UINTTOFP_I32_F32;
465 else if (RetVT == MVT::f64)
466 return UINTTOFP_I32_F64;
467 else if (RetVT == MVT::f80)
468 return UINTTOFP_I32_F80;
469 else if (RetVT == MVT::ppcf128)
470 return UINTTOFP_I32_PPCF128;
471 } else if (OpVT == MVT::i64) {
472 if (RetVT == MVT::f32)
473 return UINTTOFP_I64_F32;
474 else if (RetVT == MVT::f64)
475 return UINTTOFP_I64_F64;
476 else if (RetVT == MVT::f80)
477 return UINTTOFP_I64_F80;
478 else if (RetVT == MVT::ppcf128)
479 return UINTTOFP_I64_PPCF128;
480 } else if (OpVT == MVT::i128) {
481 if (RetVT == MVT::f32)
482 return UINTTOFP_I128_F32;
483 else if (RetVT == MVT::f64)
484 return UINTTOFP_I128_F64;
485 else if (RetVT == MVT::f80)
486 return UINTTOFP_I128_F80;
487 else if (RetVT == MVT::ppcf128)
488 return UINTTOFP_I128_PPCF128;
490 return UNKNOWN_LIBCALL;
493 /// InitCmpLibcallCCs - Set default comparison libcall CC.
495 static void InitCmpLibcallCCs(ISD::CondCode *CCs) {
496 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL);
497 CCs[RTLIB::OEQ_F32] = ISD::SETEQ;
498 CCs[RTLIB::OEQ_F64] = ISD::SETEQ;
499 CCs[RTLIB::UNE_F32] = ISD::SETNE;
500 CCs[RTLIB::UNE_F64] = ISD::SETNE;
501 CCs[RTLIB::OGE_F32] = ISD::SETGE;
502 CCs[RTLIB::OGE_F64] = ISD::SETGE;
503 CCs[RTLIB::OLT_F32] = ISD::SETLT;
504 CCs[RTLIB::OLT_F64] = ISD::SETLT;
505 CCs[RTLIB::OLE_F32] = ISD::SETLE;
506 CCs[RTLIB::OLE_F64] = ISD::SETLE;
507 CCs[RTLIB::OGT_F32] = ISD::SETGT;
508 CCs[RTLIB::OGT_F64] = ISD::SETGT;
509 CCs[RTLIB::UO_F32] = ISD::SETNE;
510 CCs[RTLIB::UO_F64] = ISD::SETNE;
511 CCs[RTLIB::O_F32] = ISD::SETEQ;
512 CCs[RTLIB::O_F64] = ISD::SETEQ;
515 /// NOTE: The constructor takes ownership of TLOF.
516 TargetLowering::TargetLowering(const TargetMachine &tm,
517 const TargetLoweringObjectFile *tlof)
518 : TM(tm), TD(TM.getTargetData()), TLOF(*tlof) {
519 // All operations default to being supported.
520 memset(OpActions, 0, sizeof(OpActions));
521 memset(LoadExtActions, 0, sizeof(LoadExtActions));
522 memset(TruncStoreActions, 0, sizeof(TruncStoreActions));
523 memset(IndexedModeActions, 0, sizeof(IndexedModeActions));
524 memset(CondCodeActions, 0, sizeof(CondCodeActions));
526 // Set default actions for various operations.
527 for (unsigned VT = 0; VT != (unsigned)MVT::LAST_VALUETYPE; ++VT) {
528 // Default all indexed load / store to expand.
529 for (unsigned IM = (unsigned)ISD::PRE_INC;
530 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
531 setIndexedLoadAction(IM, (MVT::SimpleValueType)VT, Expand);
532 setIndexedStoreAction(IM, (MVT::SimpleValueType)VT, Expand);
535 // These operations default to expand.
536 setOperationAction(ISD::FGETSIGN, (MVT::SimpleValueType)VT, Expand);
537 setOperationAction(ISD::CONCAT_VECTORS, (MVT::SimpleValueType)VT, Expand);
540 // Most targets ignore the @llvm.prefetch intrinsic.
541 setOperationAction(ISD::PREFETCH, MVT::Other, Expand);
543 // ConstantFP nodes default to expand. Targets can either change this to
544 // Legal, in which case all fp constants are legal, or use isFPImmLegal()
545 // to optimize expansions for certain constants.
546 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
547 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
548 setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
550 // These library functions default to expand.
551 setOperationAction(ISD::FLOG , MVT::f64, Expand);
552 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
553 setOperationAction(ISD::FLOG10,MVT::f64, Expand);
554 setOperationAction(ISD::FEXP , MVT::f64, Expand);
555 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
556 setOperationAction(ISD::FLOG , MVT::f32, Expand);
557 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
558 setOperationAction(ISD::FLOG10,MVT::f32, Expand);
559 setOperationAction(ISD::FEXP , MVT::f32, Expand);
560 setOperationAction(ISD::FEXP2, MVT::f32, Expand);
562 // Default ISD::TRAP to expand (which turns it into abort).
563 setOperationAction(ISD::TRAP, MVT::Other, Expand);
565 IsLittleEndian = TD->isLittleEndian();
566 ShiftAmountTy = PointerTy = MVT::getIntegerVT(8*TD->getPointerSize());
567 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE*sizeof(TargetRegisterClass*));
568 memset(TargetDAGCombineArray, 0, array_lengthof(TargetDAGCombineArray));
569 maxStoresPerMemset = maxStoresPerMemcpy = maxStoresPerMemmove = 8;
570 maxStoresPerMemsetOptSize = maxStoresPerMemcpyOptSize
571 = maxStoresPerMemmoveOptSize = 4;
572 benefitFromCodePlacementOpt = false;
573 UseUnderscoreSetJmp = false;
574 UseUnderscoreLongJmp = false;
575 SelectIsExpensive = false;
576 IntDivIsCheap = false;
577 Pow2DivIsCheap = false;
578 JumpIsExpensive = false;
579 StackPointerRegisterToSaveRestore = 0;
580 ExceptionPointerRegister = 0;
581 ExceptionSelectorRegister = 0;
582 BooleanContents = UndefinedBooleanContent;
583 SchedPreferenceInfo = Sched::Latency;
585 JumpBufAlignment = 0;
586 PrefLoopAlignment = 0;
587 MinStackArgumentAlignment = 1;
588 ShouldFoldAtomicFences = false;
590 InitLibcallNames(LibcallRoutineNames);
591 InitCmpLibcallCCs(CmpLibcallCCs);
592 InitLibcallCallingConvs(LibcallCallingConvs);
595 TargetLowering::~TargetLowering() {
599 /// canOpTrap - Returns true if the operation can trap for the value type.
600 /// VT must be a legal type.
601 bool TargetLowering::canOpTrap(unsigned Op, EVT VT) const {
602 assert(isTypeLegal(VT));
617 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT,
618 unsigned &NumIntermediates,
620 TargetLowering *TLI) {
621 // Figure out the right, legal destination reg to copy into.
622 unsigned NumElts = VT.getVectorNumElements();
623 MVT EltTy = VT.getVectorElementType();
625 unsigned NumVectorRegs = 1;
627 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
628 // could break down into LHS/RHS like LegalizeDAG does.
629 if (!isPowerOf2_32(NumElts)) {
630 NumVectorRegs = NumElts;
634 // Divide the input until we get to a supported size. This will always
635 // end with a scalar if the target doesn't support vectors.
636 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) {
641 NumIntermediates = NumVectorRegs;
643 MVT NewVT = MVT::getVectorVT(EltTy, NumElts);
644 if (!TLI->isTypeLegal(NewVT))
646 IntermediateVT = NewVT;
648 EVT DestVT = TLI->getRegisterType(NewVT);
650 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16.
651 return NumVectorRegs*(NewVT.getSizeInBits()/DestVT.getSizeInBits());
653 // Otherwise, promotion or legal types use the same number of registers as
654 // the vector decimated to the appropriate level.
655 return NumVectorRegs;
658 /// isLegalRC - Return true if the value types that can be represented by the
659 /// specified register class are all legal.
660 bool TargetLowering::isLegalRC(const TargetRegisterClass *RC) const {
661 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
669 /// hasLegalSuperRegRegClasses - Return true if the specified register class
670 /// has one or more super-reg register classes that are legal.
672 TargetLowering::hasLegalSuperRegRegClasses(const TargetRegisterClass *RC) const{
673 if (*RC->superregclasses_begin() == 0)
675 for (TargetRegisterInfo::regclass_iterator I = RC->superregclasses_begin(),
676 E = RC->superregclasses_end(); I != E; ++I) {
677 const TargetRegisterClass *RRC = *I;
684 /// findRepresentativeClass - Return the largest legal super-reg register class
685 /// of the register class for the specified type and its associated "cost".
686 std::pair<const TargetRegisterClass*, uint8_t>
687 TargetLowering::findRepresentativeClass(EVT VT) const {
688 const TargetRegisterClass *RC = RegClassForVT[VT.getSimpleVT().SimpleTy];
690 return std::make_pair(RC, 0);
691 const TargetRegisterClass *BestRC = RC;
692 for (TargetRegisterInfo::regclass_iterator I = RC->superregclasses_begin(),
693 E = RC->superregclasses_end(); I != E; ++I) {
694 const TargetRegisterClass *RRC = *I;
695 if (RRC->isASubClass() || !isLegalRC(RRC))
697 if (!hasLegalSuperRegRegClasses(RRC))
698 return std::make_pair(RRC, 1);
701 return std::make_pair(BestRC, 1);
705 /// computeRegisterProperties - Once all of the register classes are added,
706 /// this allows us to compute derived properties we expose.
707 void TargetLowering::computeRegisterProperties() {
708 assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE &&
709 "Too many value types for ValueTypeActions to hold!");
711 // Everything defaults to needing one register.
712 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
713 NumRegistersForVT[i] = 1;
714 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i;
716 // ...except isVoid, which doesn't need any registers.
717 NumRegistersForVT[MVT::isVoid] = 0;
719 // Find the largest integer register class.
720 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE;
721 for (; RegClassForVT[LargestIntReg] == 0; --LargestIntReg)
722 assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
724 // Every integer value type larger than this largest register takes twice as
725 // many registers to represent as the previous ValueType.
726 for (unsigned ExpandedReg = LargestIntReg + 1; ; ++ExpandedReg) {
727 EVT ExpandedVT = (MVT::SimpleValueType)ExpandedReg;
728 if (!ExpandedVT.isInteger())
730 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
731 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg;
732 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1);
733 ValueTypeActions.setTypeAction(ExpandedVT, Expand);
736 // Inspect all of the ValueType's smaller than the largest integer
737 // register to see which ones need promotion.
738 unsigned LegalIntReg = LargestIntReg;
739 for (unsigned IntReg = LargestIntReg - 1;
740 IntReg >= (unsigned)MVT::i1; --IntReg) {
741 EVT IVT = (MVT::SimpleValueType)IntReg;
742 if (isTypeLegal(IVT)) {
743 LegalIntReg = IntReg;
745 RegisterTypeForVT[IntReg] = TransformToType[IntReg] =
746 (MVT::SimpleValueType)LegalIntReg;
747 ValueTypeActions.setTypeAction(IVT, Promote);
751 // ppcf128 type is really two f64's.
752 if (!isTypeLegal(MVT::ppcf128)) {
753 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64];
754 RegisterTypeForVT[MVT::ppcf128] = MVT::f64;
755 TransformToType[MVT::ppcf128] = MVT::f64;
756 ValueTypeActions.setTypeAction(MVT::ppcf128, Expand);
759 // Decide how to handle f64. If the target does not have native f64 support,
760 // expand it to i64 and we will be generating soft float library calls.
761 if (!isTypeLegal(MVT::f64)) {
762 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
763 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
764 TransformToType[MVT::f64] = MVT::i64;
765 ValueTypeActions.setTypeAction(MVT::f64, Expand);
768 // Decide how to handle f32. If the target does not have native support for
769 // f32, promote it to f64 if it is legal. Otherwise, expand it to i32.
770 if (!isTypeLegal(MVT::f32)) {
771 if (isTypeLegal(MVT::f64)) {
772 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::f64];
773 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::f64];
774 TransformToType[MVT::f32] = MVT::f64;
775 ValueTypeActions.setTypeAction(MVT::f32, Promote);
777 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
778 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
779 TransformToType[MVT::f32] = MVT::i32;
780 ValueTypeActions.setTypeAction(MVT::f32, Expand);
784 // Loop over all of the vector value types to see which need transformations.
785 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
786 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
787 MVT VT = (MVT::SimpleValueType)i;
788 if (isTypeLegal(VT)) continue;
790 // Determine if there is a legal wider type. If so, we should promote to
791 // that wider vector type.
792 EVT EltVT = VT.getVectorElementType();
793 unsigned NElts = VT.getVectorNumElements();
795 bool IsLegalWiderType = false;
796 for (unsigned nVT = i+1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
797 EVT SVT = (MVT::SimpleValueType)nVT;
798 if (SVT.getVectorElementType() == EltVT &&
799 SVT.getVectorNumElements() > NElts &&
801 TransformToType[i] = SVT;
802 RegisterTypeForVT[i] = SVT;
803 NumRegistersForVT[i] = 1;
804 ValueTypeActions.setTypeAction(VT, Promote);
805 IsLegalWiderType = true;
809 if (IsLegalWiderType) continue;
814 unsigned NumIntermediates;
815 NumRegistersForVT[i] =
816 getVectorTypeBreakdownMVT(VT, IntermediateVT, NumIntermediates,
818 RegisterTypeForVT[i] = RegisterVT;
820 EVT NVT = VT.getPow2VectorType();
822 // Type is already a power of 2. The default action is to split.
823 TransformToType[i] = MVT::Other;
824 ValueTypeActions.setTypeAction(VT, Expand);
826 TransformToType[i] = NVT;
827 ValueTypeActions.setTypeAction(VT, Promote);
831 // Determine the 'representative' register class for each value type.
832 // An representative register class is the largest (meaning one which is
833 // not a sub-register class / subreg register class) legal register class for
834 // a group of value types. For example, on i386, i8, i16, and i32
835 // representative would be GR32; while on x86_64 it's GR64.
836 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
837 const TargetRegisterClass* RRC;
839 tie(RRC, Cost) = findRepresentativeClass((MVT::SimpleValueType)i);
840 RepRegClassForVT[i] = RRC;
841 RepRegClassCostForVT[i] = Cost;
845 const char *TargetLowering::getTargetNodeName(unsigned Opcode) const {
850 MVT::SimpleValueType TargetLowering::getSetCCResultType(EVT VT) const {
851 return PointerTy.SimpleTy;
854 MVT::SimpleValueType TargetLowering::getCmpLibcallReturnType() const {
855 return MVT::i32; // return the default value
858 /// getVectorTypeBreakdown - Vector types are broken down into some number of
859 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
860 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
861 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
863 /// This method returns the number of registers needed, and the VT for each
864 /// register. It also returns the VT and quantity of the intermediate values
865 /// before they are promoted/expanded.
867 unsigned TargetLowering::getVectorTypeBreakdown(LLVMContext &Context, EVT VT,
869 unsigned &NumIntermediates,
870 EVT &RegisterVT) const {
871 unsigned NumElts = VT.getVectorNumElements();
873 // If there is a wider vector type with the same element type as this one,
874 // we should widen to that legal vector type. This handles things like
875 // <2 x float> -> <4 x float>.
876 if (NumElts != 1 && getTypeAction(VT) == Promote) {
877 RegisterVT = getTypeToTransformTo(Context, VT);
878 if (isTypeLegal(RegisterVT)) {
879 IntermediateVT = RegisterVT;
880 NumIntermediates = 1;
885 // Figure out the right, legal destination reg to copy into.
886 EVT EltTy = VT.getVectorElementType();
888 unsigned NumVectorRegs = 1;
890 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
891 // could break down into LHS/RHS like LegalizeDAG does.
892 if (!isPowerOf2_32(NumElts)) {
893 NumVectorRegs = NumElts;
897 // Divide the input until we get to a supported size. This will always
898 // end with a scalar if the target doesn't support vectors.
899 while (NumElts > 1 && !isTypeLegal(
900 EVT::getVectorVT(Context, EltTy, NumElts))) {
905 NumIntermediates = NumVectorRegs;
907 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts);
908 if (!isTypeLegal(NewVT))
910 IntermediateVT = NewVT;
912 EVT DestVT = getRegisterType(Context, NewVT);
914 if (DestVT.bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16.
915 return NumVectorRegs*(NewVT.getSizeInBits()/DestVT.getSizeInBits());
917 // Otherwise, promotion or legal types use the same number of registers as
918 // the vector decimated to the appropriate level.
919 return NumVectorRegs;
922 /// Get the EVTs and ArgFlags collections that represent the legalized return
923 /// type of the given function. This does not require a DAG or a return value,
924 /// and is suitable for use before any DAGs for the function are constructed.
925 /// TODO: Move this out of TargetLowering.cpp.
926 void llvm::GetReturnInfo(const Type* ReturnType, Attributes attr,
927 SmallVectorImpl<ISD::OutputArg> &Outs,
928 const TargetLowering &TLI,
929 SmallVectorImpl<uint64_t> *Offsets) {
930 SmallVector<EVT, 4> ValueVTs;
931 ComputeValueVTs(TLI, ReturnType, ValueVTs);
932 unsigned NumValues = ValueVTs.size();
933 if (NumValues == 0) return;
936 for (unsigned j = 0, f = NumValues; j != f; ++j) {
937 EVT VT = ValueVTs[j];
938 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
940 if (attr & Attribute::SExt)
941 ExtendKind = ISD::SIGN_EXTEND;
942 else if (attr & Attribute::ZExt)
943 ExtendKind = ISD::ZERO_EXTEND;
945 // FIXME: C calling convention requires the return type to be promoted to
946 // at least 32-bit. But this is not necessary for non-C calling
947 // conventions. The frontend should mark functions whose return values
948 // require promoting with signext or zeroext attributes.
949 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) {
950 EVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32);
951 if (VT.bitsLT(MinVT))
955 unsigned NumParts = TLI.getNumRegisters(ReturnType->getContext(), VT);
956 EVT PartVT = TLI.getRegisterType(ReturnType->getContext(), VT);
957 unsigned PartSize = TLI.getTargetData()->getTypeAllocSize(
958 PartVT.getTypeForEVT(ReturnType->getContext()));
960 // 'inreg' on function refers to return value
961 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
962 if (attr & Attribute::InReg)
965 // Propagate extension type if any
966 if (attr & Attribute::SExt)
968 else if (attr & Attribute::ZExt)
971 for (unsigned i = 0; i < NumParts; ++i) {
972 Outs.push_back(ISD::OutputArg(Flags, PartVT, /*isFixed=*/true));
974 Offsets->push_back(Offset);
981 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
982 /// function arguments in the caller parameter area. This is the actual
983 /// alignment, not its logarithm.
984 unsigned TargetLowering::getByValTypeAlignment(const Type *Ty) const {
985 return TD->getCallFrameTypeAlignment(Ty);
988 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
989 /// current function. The returned value is a member of the
990 /// MachineJumpTableInfo::JTEntryKind enum.
991 unsigned TargetLowering::getJumpTableEncoding() const {
992 // In non-pic modes, just use the address of a block.
993 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
994 return MachineJumpTableInfo::EK_BlockAddress;
996 // In PIC mode, if the target supports a GPRel32 directive, use it.
997 if (getTargetMachine().getMCAsmInfo()->getGPRel32Directive() != 0)
998 return MachineJumpTableInfo::EK_GPRel32BlockAddress;
1000 // Otherwise, use a label difference.
1001 return MachineJumpTableInfo::EK_LabelDifference32;
1004 SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1005 SelectionDAG &DAG) const {
1006 // If our PIC model is GP relative, use the global offset table as the base.
1007 if (getJumpTableEncoding() == MachineJumpTableInfo::EK_GPRel32BlockAddress)
1008 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy());
1012 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1013 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1016 TargetLowering::getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
1017 unsigned JTI,MCContext &Ctx) const{
1018 // The normal PIC reloc base is the label at the start of the jump table.
1019 return MCSymbolRefExpr::Create(MF->getJTISymbol(JTI, Ctx), Ctx);
1023 TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
1024 // Assume that everything is safe in static mode.
1025 if (getTargetMachine().getRelocationModel() == Reloc::Static)
1028 // In dynamic-no-pic mode, assume that known defined values are safe.
1029 if (getTargetMachine().getRelocationModel() == Reloc::DynamicNoPIC &&
1031 !GA->getGlobal()->isDeclaration() &&
1032 !GA->getGlobal()->isWeakForLinker())
1035 // Otherwise assume nothing is safe.
1039 //===----------------------------------------------------------------------===//
1040 // Optimization Methods
1041 //===----------------------------------------------------------------------===//
1043 /// ShrinkDemandedConstant - Check to see if the specified operand of the
1044 /// specified instruction is a constant integer. If so, check to see if there
1045 /// are any bits set in the constant that are not demanded. If so, shrink the
1046 /// constant and return true.
1047 bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDValue Op,
1048 const APInt &Demanded) {
1049 DebugLoc dl = Op.getDebugLoc();
1051 // FIXME: ISD::SELECT, ISD::SELECT_CC
1052 switch (Op.getOpcode()) {
1057 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
1058 if (!C) return false;
1060 if (Op.getOpcode() == ISD::XOR &&
1061 (C->getAPIntValue() | (~Demanded)).isAllOnesValue())
1064 // if we can expand it to have all bits set, do it
1065 if (C->getAPIntValue().intersects(~Demanded)) {
1066 EVT VT = Op.getValueType();
1067 SDValue New = DAG.getNode(Op.getOpcode(), dl, VT, Op.getOperand(0),
1068 DAG.getConstant(Demanded &
1071 return CombineTo(Op, New);
1081 /// ShrinkDemandedOp - Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the
1082 /// casts are free. This uses isZExtFree and ZERO_EXTEND for the widening
1083 /// cast, but it could be generalized for targets with other types of
1084 /// implicit widening casts.
1086 TargetLowering::TargetLoweringOpt::ShrinkDemandedOp(SDValue Op,
1088 const APInt &Demanded,
1090 assert(Op.getNumOperands() == 2 &&
1091 "ShrinkDemandedOp only supports binary operators!");
1092 assert(Op.getNode()->getNumValues() == 1 &&
1093 "ShrinkDemandedOp only supports nodes with one result!");
1095 // Don't do this if the node has another user, which may require the
1097 if (!Op.getNode()->hasOneUse())
1100 // Search for the smallest integer type with free casts to and from
1101 // Op's type. For expedience, just check power-of-2 integer types.
1102 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1103 unsigned SmallVTBits = BitWidth - Demanded.countLeadingZeros();
1104 if (!isPowerOf2_32(SmallVTBits))
1105 SmallVTBits = NextPowerOf2(SmallVTBits);
1106 for (; SmallVTBits < BitWidth; SmallVTBits = NextPowerOf2(SmallVTBits)) {
1107 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), SmallVTBits);
1108 if (TLI.isTruncateFree(Op.getValueType(), SmallVT) &&
1109 TLI.isZExtFree(SmallVT, Op.getValueType())) {
1110 // We found a type with free casts.
1111 SDValue X = DAG.getNode(Op.getOpcode(), dl, SmallVT,
1112 DAG.getNode(ISD::TRUNCATE, dl, SmallVT,
1113 Op.getNode()->getOperand(0)),
1114 DAG.getNode(ISD::TRUNCATE, dl, SmallVT,
1115 Op.getNode()->getOperand(1)));
1116 SDValue Z = DAG.getNode(ISD::ZERO_EXTEND, dl, Op.getValueType(), X);
1117 return CombineTo(Op, Z);
1123 /// SimplifyDemandedBits - Look at Op. At this point, we know that only the
1124 /// DemandedMask bits of the result of Op are ever used downstream. If we can
1125 /// use this information to simplify Op, create a new simplified DAG node and
1126 /// return true, returning the original and new nodes in Old and New. Otherwise,
1127 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
1128 /// the expression (used to simplify the caller). The KnownZero/One bits may
1129 /// only be accurate for those bits in the DemandedMask.
1130 bool TargetLowering::SimplifyDemandedBits(SDValue Op,
1131 const APInt &DemandedMask,
1134 TargetLoweringOpt &TLO,
1135 unsigned Depth) const {
1136 unsigned BitWidth = DemandedMask.getBitWidth();
1137 assert(Op.getValueType().getScalarType().getSizeInBits() == BitWidth &&
1138 "Mask size mismatches value type size!");
1139 APInt NewMask = DemandedMask;
1140 DebugLoc dl = Op.getDebugLoc();
1142 // Don't know anything.
1143 KnownZero = KnownOne = APInt(BitWidth, 0);
1145 // Other users may use these bits.
1146 if (!Op.getNode()->hasOneUse()) {
1148 // If not at the root, Just compute the KnownZero/KnownOne bits to
1149 // simplify things downstream.
1150 TLO.DAG.ComputeMaskedBits(Op, DemandedMask, KnownZero, KnownOne, Depth);
1153 // If this is the root being simplified, allow it to have multiple uses,
1154 // just set the NewMask to all bits.
1155 NewMask = APInt::getAllOnesValue(BitWidth);
1156 } else if (DemandedMask == 0) {
1157 // Not demanding any bits from Op.
1158 if (Op.getOpcode() != ISD::UNDEF)
1159 return TLO.CombineTo(Op, TLO.DAG.getUNDEF(Op.getValueType()));
1161 } else if (Depth == 6) { // Limit search depth.
1165 APInt KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut;
1166 switch (Op.getOpcode()) {
1168 // We know all of the bits for a constant!
1169 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue() & NewMask;
1170 KnownZero = ~KnownOne & NewMask;
1171 return false; // Don't fall through, will infinitely loop.
1173 // If the RHS is a constant, check to see if the LHS would be zero without
1174 // using the bits from the RHS. Below, we use knowledge about the RHS to
1175 // simplify the LHS, here we're using information from the LHS to simplify
1177 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1178 APInt LHSZero, LHSOne;
1179 // Do not increment Depth here; that can cause an infinite loop.
1180 TLO.DAG.ComputeMaskedBits(Op.getOperand(0), NewMask,
1181 LHSZero, LHSOne, Depth);
1182 // If the LHS already has zeros where RHSC does, this and is dead.
1183 if ((LHSZero & NewMask) == (~RHSC->getAPIntValue() & NewMask))
1184 return TLO.CombineTo(Op, Op.getOperand(0));
1185 // If any of the set bits in the RHS are known zero on the LHS, shrink
1187 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & NewMask))
1191 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
1192 KnownOne, TLO, Depth+1))
1194 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1195 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownZero & NewMask,
1196 KnownZero2, KnownOne2, TLO, Depth+1))
1198 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1200 // If all of the demanded bits are known one on one side, return the other.
1201 // These bits cannot contribute to the result of the 'and'.
1202 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
1203 return TLO.CombineTo(Op, Op.getOperand(0));
1204 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
1205 return TLO.CombineTo(Op, Op.getOperand(1));
1206 // If all of the demanded bits in the inputs are known zeros, return zero.
1207 if ((NewMask & (KnownZero|KnownZero2)) == NewMask)
1208 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, Op.getValueType()));
1209 // If the RHS is a constant, see if we can simplify it.
1210 if (TLO.ShrinkDemandedConstant(Op, ~KnownZero2 & NewMask))
1212 // If the operation can be done in a smaller type, do so.
1213 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1216 // Output known-1 bits are only known if set in both the LHS & RHS.
1217 KnownOne &= KnownOne2;
1218 // Output known-0 are known to be clear if zero in either the LHS | RHS.
1219 KnownZero |= KnownZero2;
1222 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
1223 KnownOne, TLO, Depth+1))
1225 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1226 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownOne & NewMask,
1227 KnownZero2, KnownOne2, TLO, Depth+1))
1229 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1231 // If all of the demanded bits are known zero on one side, return the other.
1232 // These bits cannot contribute to the result of the 'or'.
1233 if ((NewMask & ~KnownOne2 & KnownZero) == (~KnownOne2 & NewMask))
1234 return TLO.CombineTo(Op, Op.getOperand(0));
1235 if ((NewMask & ~KnownOne & KnownZero2) == (~KnownOne & NewMask))
1236 return TLO.CombineTo(Op, Op.getOperand(1));
1237 // If all of the potentially set bits on one side are known to be set on
1238 // the other side, just use the 'other' side.
1239 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
1240 return TLO.CombineTo(Op, Op.getOperand(0));
1241 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
1242 return TLO.CombineTo(Op, Op.getOperand(1));
1243 // If the RHS is a constant, see if we can simplify it.
1244 if (TLO.ShrinkDemandedConstant(Op, NewMask))
1246 // If the operation can be done in a smaller type, do so.
1247 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1250 // Output known-0 bits are only known if clear in both the LHS & RHS.
1251 KnownZero &= KnownZero2;
1252 // Output known-1 are known to be set if set in either the LHS | RHS.
1253 KnownOne |= KnownOne2;
1256 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
1257 KnownOne, TLO, Depth+1))
1259 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1260 if (SimplifyDemandedBits(Op.getOperand(0), NewMask, KnownZero2,
1261 KnownOne2, TLO, Depth+1))
1263 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1265 // If all of the demanded bits are known zero on one side, return the other.
1266 // These bits cannot contribute to the result of the 'xor'.
1267 if ((KnownZero & NewMask) == NewMask)
1268 return TLO.CombineTo(Op, Op.getOperand(0));
1269 if ((KnownZero2 & NewMask) == NewMask)
1270 return TLO.CombineTo(Op, Op.getOperand(1));
1271 // If the operation can be done in a smaller type, do so.
1272 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1275 // If all of the unknown bits are known to be zero on one side or the other
1276 // (but not both) turn this into an *inclusive* or.
1277 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1278 if ((NewMask & ~KnownZero & ~KnownZero2) == 0)
1279 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, dl, Op.getValueType(),
1283 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1284 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
1285 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1286 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
1288 // If all of the demanded bits on one side are known, and all of the set
1289 // bits on that side are also known to be set on the other side, turn this
1290 // into an AND, as we know the bits will be cleared.
1291 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1292 if ((NewMask & (KnownZero|KnownOne)) == NewMask) { // all known
1293 if ((KnownOne & KnownOne2) == KnownOne) {
1294 EVT VT = Op.getValueType();
1295 SDValue ANDC = TLO.DAG.getConstant(~KnownOne & NewMask, VT);
1296 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, dl, VT,
1297 Op.getOperand(0), ANDC));
1301 // If the RHS is a constant, see if we can simplify it.
1302 // for XOR, we prefer to force bits to 1 if they will make a -1.
1303 // if we can't force bits, try to shrink constant
1304 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1305 APInt Expanded = C->getAPIntValue() | (~NewMask);
1306 // if we can expand it to have all bits set, do it
1307 if (Expanded.isAllOnesValue()) {
1308 if (Expanded != C->getAPIntValue()) {
1309 EVT VT = Op.getValueType();
1310 SDValue New = TLO.DAG.getNode(Op.getOpcode(), dl,VT, Op.getOperand(0),
1311 TLO.DAG.getConstant(Expanded, VT));
1312 return TLO.CombineTo(Op, New);
1314 // if it already has all the bits set, nothing to change
1315 // but don't shrink either!
1316 } else if (TLO.ShrinkDemandedConstant(Op, NewMask)) {
1321 KnownZero = KnownZeroOut;
1322 KnownOne = KnownOneOut;
1325 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero,
1326 KnownOne, TLO, Depth+1))
1328 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero2,
1329 KnownOne2, TLO, Depth+1))
1331 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1332 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1334 // If the operands are constants, see if we can simplify them.
1335 if (TLO.ShrinkDemandedConstant(Op, NewMask))
1338 // Only known if known in both the LHS and RHS.
1339 KnownOne &= KnownOne2;
1340 KnownZero &= KnownZero2;
1342 case ISD::SELECT_CC:
1343 if (SimplifyDemandedBits(Op.getOperand(3), NewMask, KnownZero,
1344 KnownOne, TLO, Depth+1))
1346 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero2,
1347 KnownOne2, TLO, Depth+1))
1349 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1350 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1352 // If the operands are constants, see if we can simplify them.
1353 if (TLO.ShrinkDemandedConstant(Op, NewMask))
1356 // Only known if known in both the LHS and RHS.
1357 KnownOne &= KnownOne2;
1358 KnownZero &= KnownZero2;
1361 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1362 unsigned ShAmt = SA->getZExtValue();
1363 SDValue InOp = Op.getOperand(0);
1365 // If the shift count is an invalid immediate, don't do anything.
1366 if (ShAmt >= BitWidth)
1369 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a
1370 // single shift. We can do this if the bottom bits (which are shifted
1371 // out) are never demanded.
1372 if (InOp.getOpcode() == ISD::SRL &&
1373 isa<ConstantSDNode>(InOp.getOperand(1))) {
1374 if (ShAmt && (NewMask & APInt::getLowBitsSet(BitWidth, ShAmt)) == 0) {
1375 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
1376 unsigned Opc = ISD::SHL;
1377 int Diff = ShAmt-C1;
1384 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
1385 EVT VT = Op.getValueType();
1386 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT,
1387 InOp.getOperand(0), NewSA));
1391 if (SimplifyDemandedBits(InOp, NewMask.lshr(ShAmt),
1392 KnownZero, KnownOne, TLO, Depth+1))
1395 // Convert (shl (anyext x, c)) to (anyext (shl x, c)) if the high bits
1396 // are not demanded. This will likely allow the anyext to be folded away.
1397 if (InOp.getNode()->getOpcode() == ISD::ANY_EXTEND) {
1398 SDValue InnerOp = InOp.getNode()->getOperand(0);
1399 EVT InnerVT = InnerOp.getValueType();
1400 if ((APInt::getHighBitsSet(BitWidth,
1401 BitWidth - InnerVT.getSizeInBits()) &
1402 DemandedMask) == 0 &&
1403 isTypeDesirableForOp(ISD::SHL, InnerVT)) {
1404 EVT ShTy = getShiftAmountTy();
1405 if (!APInt(BitWidth, ShAmt).isIntN(ShTy.getSizeInBits()))
1408 TLO.DAG.getNode(ISD::SHL, dl, InnerVT, InnerOp,
1409 TLO.DAG.getConstant(ShAmt, ShTy));
1412 TLO.DAG.getNode(ISD::ANY_EXTEND, dl, Op.getValueType(),
1417 KnownZero <<= SA->getZExtValue();
1418 KnownOne <<= SA->getZExtValue();
1419 // low bits known zero.
1420 KnownZero |= APInt::getLowBitsSet(BitWidth, SA->getZExtValue());
1424 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1425 EVT VT = Op.getValueType();
1426 unsigned ShAmt = SA->getZExtValue();
1427 unsigned VTSize = VT.getSizeInBits();
1428 SDValue InOp = Op.getOperand(0);
1430 // If the shift count is an invalid immediate, don't do anything.
1431 if (ShAmt >= BitWidth)
1434 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a
1435 // single shift. We can do this if the top bits (which are shifted out)
1436 // are never demanded.
1437 if (InOp.getOpcode() == ISD::SHL &&
1438 isa<ConstantSDNode>(InOp.getOperand(1))) {
1439 if (ShAmt && (NewMask & APInt::getHighBitsSet(VTSize, ShAmt)) == 0) {
1440 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
1441 unsigned Opc = ISD::SRL;
1442 int Diff = ShAmt-C1;
1449 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
1450 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT,
1451 InOp.getOperand(0), NewSA));
1455 // Compute the new bits that are at the top now.
1456 if (SimplifyDemandedBits(InOp, (NewMask << ShAmt),
1457 KnownZero, KnownOne, TLO, Depth+1))
1459 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1460 KnownZero = KnownZero.lshr(ShAmt);
1461 KnownOne = KnownOne.lshr(ShAmt);
1463 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
1464 KnownZero |= HighBits; // High bits known zero.
1468 // If this is an arithmetic shift right and only the low-bit is set, we can
1469 // always convert this into a logical shr, even if the shift amount is
1470 // variable. The low bit of the shift cannot be an input sign bit unless
1471 // the shift amount is >= the size of the datatype, which is undefined.
1472 if (DemandedMask == 1)
1473 return TLO.CombineTo(Op,
1474 TLO.DAG.getNode(ISD::SRL, dl, Op.getValueType(),
1475 Op.getOperand(0), Op.getOperand(1)));
1477 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1478 EVT VT = Op.getValueType();
1479 unsigned ShAmt = SA->getZExtValue();
1481 // If the shift count is an invalid immediate, don't do anything.
1482 if (ShAmt >= BitWidth)
1485 APInt InDemandedMask = (NewMask << ShAmt);
1487 // If any of the demanded bits are produced by the sign extension, we also
1488 // demand the input sign bit.
1489 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
1490 if (HighBits.intersects(NewMask))
1491 InDemandedMask |= APInt::getSignBit(VT.getScalarType().getSizeInBits());
1493 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask,
1494 KnownZero, KnownOne, TLO, Depth+1))
1496 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1497 KnownZero = KnownZero.lshr(ShAmt);
1498 KnownOne = KnownOne.lshr(ShAmt);
1500 // Handle the sign bit, adjusted to where it is now in the mask.
1501 APInt SignBit = APInt::getSignBit(BitWidth).lshr(ShAmt);
1503 // If the input sign bit is known to be zero, or if none of the top bits
1504 // are demanded, turn this into an unsigned shift right.
1505 if (KnownZero.intersects(SignBit) || (HighBits & ~NewMask) == HighBits) {
1506 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, VT,
1509 } else if (KnownOne.intersects(SignBit)) { // New bits are known one.
1510 KnownOne |= HighBits;
1514 case ISD::SIGN_EXTEND_INREG: {
1515 EVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1517 // Sign extension. Compute the demanded bits in the result that are not
1518 // present in the input.
1520 APInt::getHighBitsSet(BitWidth,
1521 BitWidth - EVT.getScalarType().getSizeInBits());
1523 // If none of the extended bits are demanded, eliminate the sextinreg.
1524 if ((NewBits & NewMask) == 0)
1525 return TLO.CombineTo(Op, Op.getOperand(0));
1528 APInt::getSignBit(EVT.getScalarType().getSizeInBits()).zext(BitWidth);
1529 APInt InputDemandedBits =
1530 APInt::getLowBitsSet(BitWidth,
1531 EVT.getScalarType().getSizeInBits()) &
1534 // Since the sign extended bits are demanded, we know that the sign
1536 InputDemandedBits |= InSignBit;
1538 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits,
1539 KnownZero, KnownOne, TLO, Depth+1))
1541 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1543 // If the sign bit of the input is known set or clear, then we know the
1544 // top bits of the result.
1546 // If the input sign bit is known zero, convert this into a zero extension.
1547 if (KnownZero.intersects(InSignBit))
1548 return TLO.CombineTo(Op,
1549 TLO.DAG.getZeroExtendInReg(Op.getOperand(0),dl,EVT));
1551 if (KnownOne.intersects(InSignBit)) { // Input sign bit known set
1552 KnownOne |= NewBits;
1553 KnownZero &= ~NewBits;
1554 } else { // Input sign bit unknown
1555 KnownZero &= ~NewBits;
1556 KnownOne &= ~NewBits;
1560 case ISD::ZERO_EXTEND: {
1561 unsigned OperandBitWidth =
1562 Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
1563 APInt InMask = NewMask.trunc(OperandBitWidth);
1565 // If none of the top bits are demanded, convert this into an any_extend.
1567 APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask;
1568 if (!NewBits.intersects(NewMask))
1569 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND, dl,
1573 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1574 KnownZero, KnownOne, TLO, Depth+1))
1576 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1577 KnownZero = KnownZero.zext(BitWidth);
1578 KnownOne = KnownOne.zext(BitWidth);
1579 KnownZero |= NewBits;
1582 case ISD::SIGN_EXTEND: {
1583 EVT InVT = Op.getOperand(0).getValueType();
1584 unsigned InBits = InVT.getScalarType().getSizeInBits();
1585 APInt InMask = APInt::getLowBitsSet(BitWidth, InBits);
1586 APInt InSignBit = APInt::getBitsSet(BitWidth, InBits - 1, InBits);
1587 APInt NewBits = ~InMask & NewMask;
1589 // If none of the top bits are demanded, convert this into an any_extend.
1591 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND, dl,
1595 // Since some of the sign extended bits are demanded, we know that the sign
1597 APInt InDemandedBits = InMask & NewMask;
1598 InDemandedBits |= InSignBit;
1599 InDemandedBits = InDemandedBits.trunc(InBits);
1601 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero,
1602 KnownOne, TLO, Depth+1))
1604 KnownZero = KnownZero.zext(BitWidth);
1605 KnownOne = KnownOne.zext(BitWidth);
1607 // If the sign bit is known zero, convert this to a zero extend.
1608 if (KnownZero.intersects(InSignBit))
1609 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND, dl,
1613 // If the sign bit is known one, the top bits match.
1614 if (KnownOne.intersects(InSignBit)) {
1615 KnownOne |= NewBits;
1616 KnownZero &= ~NewBits;
1617 } else { // Otherwise, top bits aren't known.
1618 KnownOne &= ~NewBits;
1619 KnownZero &= ~NewBits;
1623 case ISD::ANY_EXTEND: {
1624 unsigned OperandBitWidth =
1625 Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
1626 APInt InMask = NewMask.trunc(OperandBitWidth);
1627 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1628 KnownZero, KnownOne, TLO, Depth+1))
1630 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1631 KnownZero = KnownZero.zext(BitWidth);
1632 KnownOne = KnownOne.zext(BitWidth);
1635 case ISD::TRUNCATE: {
1636 // Simplify the input, using demanded bit information, and compute the known
1637 // zero/one bits live out.
1638 unsigned OperandBitWidth =
1639 Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
1640 APInt TruncMask = NewMask.zext(OperandBitWidth);
1641 if (SimplifyDemandedBits(Op.getOperand(0), TruncMask,
1642 KnownZero, KnownOne, TLO, Depth+1))
1644 KnownZero = KnownZero.trunc(BitWidth);
1645 KnownOne = KnownOne.trunc(BitWidth);
1647 // If the input is only used by this truncate, see if we can shrink it based
1648 // on the known demanded bits.
1649 if (Op.getOperand(0).getNode()->hasOneUse()) {
1650 SDValue In = Op.getOperand(0);
1651 switch (In.getOpcode()) {
1654 // Shrink SRL by a constant if none of the high bits shifted in are
1656 if (TLO.LegalTypes() &&
1657 !isTypeDesirableForOp(ISD::SRL, Op.getValueType()))
1658 // Do not turn (vt1 truncate (vt2 srl)) into (vt1 srl) if vt1 is
1661 ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1));
1664 APInt HighBits = APInt::getHighBitsSet(OperandBitWidth,
1665 OperandBitWidth - BitWidth);
1666 HighBits = HighBits.lshr(ShAmt->getZExtValue()).trunc(BitWidth);
1668 if (ShAmt->getZExtValue() < BitWidth && !(HighBits & NewMask)) {
1669 // None of the shifted in bits are needed. Add a truncate of the
1670 // shift input, then shift it.
1671 SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE, dl,
1674 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl,
1683 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1686 case ISD::AssertZext: {
1687 // Demand all the bits of the input that are demanded in the output.
1688 // The low bits are obvious; the high bits are demanded because we're
1689 // asserting that they're zero here.
1690 if (SimplifyDemandedBits(Op.getOperand(0), NewMask,
1691 KnownZero, KnownOne, TLO, Depth+1))
1693 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1695 EVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1696 APInt InMask = APInt::getLowBitsSet(BitWidth,
1697 VT.getSizeInBits());
1698 KnownZero |= ~InMask & NewMask;
1703 // If this is an FP->Int bitcast and if the sign bit is the only thing that
1704 // is demanded, turn this into a FGETSIGN.
1705 if (NewMask == EVT::getIntegerVTSignBit(Op.getValueType()) &&
1706 MVT::isFloatingPoint(Op.getOperand(0).getValueType()) &&
1707 !MVT::isVector(Op.getOperand(0).getValueType())) {
1708 // Only do this xform if FGETSIGN is valid or if before legalize.
1709 if (!TLO.AfterLegalize ||
1710 isOperationLegal(ISD::FGETSIGN, Op.getValueType())) {
1711 // Make a FGETSIGN + SHL to move the sign bit into the appropriate
1712 // place. We expect the SHL to be eliminated by other optimizations.
1713 SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, Op.getValueType(),
1715 unsigned ShVal = Op.getValueType().getSizeInBits()-1;
1716 SDValue ShAmt = TLO.DAG.getConstant(ShVal, getShiftAmountTy());
1717 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, Op.getValueType(),
1726 // Add, Sub, and Mul don't demand any bits in positions beyond that
1727 // of the highest bit demanded of them.
1728 APInt LoMask = APInt::getLowBitsSet(BitWidth,
1729 BitWidth - NewMask.countLeadingZeros());
1730 if (SimplifyDemandedBits(Op.getOperand(0), LoMask, KnownZero2,
1731 KnownOne2, TLO, Depth+1))
1733 if (SimplifyDemandedBits(Op.getOperand(1), LoMask, KnownZero2,
1734 KnownOne2, TLO, Depth+1))
1736 // See if the operation should be performed at a smaller bit width.
1737 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1742 // Just use ComputeMaskedBits to compute output bits.
1743 TLO.DAG.ComputeMaskedBits(Op, NewMask, KnownZero, KnownOne, Depth);
1747 // If we know the value of all of the demanded bits, return this as a
1749 if ((NewMask & (KnownZero|KnownOne)) == NewMask)
1750 return TLO.CombineTo(Op, TLO.DAG.getConstant(KnownOne, Op.getValueType()));
1755 /// computeMaskedBitsForTargetNode - Determine which of the bits specified
1756 /// in Mask are known to be either zero or one and return them in the
1757 /// KnownZero/KnownOne bitsets.
1758 void TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
1762 const SelectionDAG &DAG,
1763 unsigned Depth) const {
1764 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1765 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1766 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1767 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1768 "Should use MaskedValueIsZero if you don't know whether Op"
1769 " is a target node!");
1770 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
1773 /// ComputeNumSignBitsForTargetNode - This method can be implemented by
1774 /// targets that want to expose additional information about sign bits to the
1776 unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
1777 unsigned Depth) const {
1778 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1779 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1780 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1781 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1782 "Should use ComputeNumSignBits if you don't know whether Op"
1783 " is a target node!");
1787 /// ValueHasExactlyOneBitSet - Test if the given value is known to have exactly
1788 /// one bit set. This differs from ComputeMaskedBits in that it doesn't need to
1789 /// determine which bit is set.
1791 static bool ValueHasExactlyOneBitSet(SDValue Val, const SelectionDAG &DAG) {
1792 // A left-shift of a constant one will have exactly one bit set, because
1793 // shifting the bit off the end is undefined.
1794 if (Val.getOpcode() == ISD::SHL)
1795 if (ConstantSDNode *C =
1796 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0)))
1797 if (C->getAPIntValue() == 1)
1800 // Similarly, a right-shift of a constant sign-bit will have exactly
1802 if (Val.getOpcode() == ISD::SRL)
1803 if (ConstantSDNode *C =
1804 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0)))
1805 if (C->getAPIntValue().isSignBit())
1808 // More could be done here, though the above checks are enough
1809 // to handle some common cases.
1811 // Fall back to ComputeMaskedBits to catch other known cases.
1812 EVT OpVT = Val.getValueType();
1813 unsigned BitWidth = OpVT.getScalarType().getSizeInBits();
1814 APInt Mask = APInt::getAllOnesValue(BitWidth);
1815 APInt KnownZero, KnownOne;
1816 DAG.ComputeMaskedBits(Val, Mask, KnownZero, KnownOne);
1817 return (KnownZero.countPopulation() == BitWidth - 1) &&
1818 (KnownOne.countPopulation() == 1);
1821 /// SimplifySetCC - Try to simplify a setcc built with the specified operands
1822 /// and cc. If it is unable to simplify it, return a null SDValue.
1824 TargetLowering::SimplifySetCC(EVT VT, SDValue N0, SDValue N1,
1825 ISD::CondCode Cond, bool foldBooleans,
1826 DAGCombinerInfo &DCI, DebugLoc dl) const {
1827 SelectionDAG &DAG = DCI.DAG;
1828 LLVMContext &Context = *DAG.getContext();
1830 // These setcc operations always fold.
1834 case ISD::SETFALSE2: return DAG.getConstant(0, VT);
1836 case ISD::SETTRUE2: return DAG.getConstant(1, VT);
1839 if (isa<ConstantSDNode>(N0.getNode())) {
1840 // Ensure that the constant occurs on the RHS, and fold constant
1842 return DAG.getSetCC(dl, VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
1845 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) {
1846 const APInt &C1 = N1C->getAPIntValue();
1848 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an
1849 // equality comparison, then we're just comparing whether X itself is
1851 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) &&
1852 N0.getOperand(0).getOpcode() == ISD::CTLZ &&
1853 N0.getOperand(1).getOpcode() == ISD::Constant) {
1855 = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
1856 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1857 ShAmt == Log2_32(N0.getValueType().getSizeInBits())) {
1858 if ((C1 == 0) == (Cond == ISD::SETEQ)) {
1859 // (srl (ctlz x), 5) == 0 -> X != 0
1860 // (srl (ctlz x), 5) != 1 -> X != 0
1863 // (srl (ctlz x), 5) != 0 -> X == 0
1864 // (srl (ctlz x), 5) == 1 -> X == 0
1867 SDValue Zero = DAG.getConstant(0, N0.getValueType());
1868 return DAG.getSetCC(dl, VT, N0.getOperand(0).getOperand(0),
1873 // If the LHS is '(and load, const)', the RHS is 0,
1874 // the test is for equality or unsigned, and all 1 bits of the const are
1875 // in the same partial word, see if we can shorten the load.
1876 if (DCI.isBeforeLegalize() &&
1877 N0.getOpcode() == ISD::AND && C1 == 0 &&
1878 N0.getNode()->hasOneUse() &&
1879 isa<LoadSDNode>(N0.getOperand(0)) &&
1880 N0.getOperand(0).getNode()->hasOneUse() &&
1881 isa<ConstantSDNode>(N0.getOperand(1))) {
1882 LoadSDNode *Lod = cast<LoadSDNode>(N0.getOperand(0));
1884 unsigned bestWidth = 0, bestOffset = 0;
1885 if (!Lod->isVolatile() && Lod->isUnindexed()) {
1886 unsigned origWidth = N0.getValueType().getSizeInBits();
1887 unsigned maskWidth = origWidth;
1888 // We can narrow (e.g.) 16-bit extending loads on 32-bit target to
1889 // 8 bits, but have to be careful...
1890 if (Lod->getExtensionType() != ISD::NON_EXTLOAD)
1891 origWidth = Lod->getMemoryVT().getSizeInBits();
1893 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
1894 for (unsigned width = origWidth / 2; width>=8; width /= 2) {
1895 APInt newMask = APInt::getLowBitsSet(maskWidth, width);
1896 for (unsigned offset=0; offset<origWidth/width; offset++) {
1897 if ((newMask & Mask) == Mask) {
1898 if (!TD->isLittleEndian())
1899 bestOffset = (origWidth/width - offset - 1) * (width/8);
1901 bestOffset = (uint64_t)offset * (width/8);
1902 bestMask = Mask.lshr(offset * (width/8) * 8);
1906 newMask = newMask << width;
1911 EVT newVT = EVT::getIntegerVT(Context, bestWidth);
1912 if (newVT.isRound()) {
1913 EVT PtrType = Lod->getOperand(1).getValueType();
1914 SDValue Ptr = Lod->getBasePtr();
1915 if (bestOffset != 0)
1916 Ptr = DAG.getNode(ISD::ADD, dl, PtrType, Lod->getBasePtr(),
1917 DAG.getConstant(bestOffset, PtrType));
1918 unsigned NewAlign = MinAlign(Lod->getAlignment(), bestOffset);
1919 SDValue NewLoad = DAG.getLoad(newVT, dl, Lod->getChain(), Ptr,
1920 Lod->getPointerInfo().getWithOffset(bestOffset),
1921 false, false, NewAlign);
1922 return DAG.getSetCC(dl, VT,
1923 DAG.getNode(ISD::AND, dl, newVT, NewLoad,
1924 DAG.getConstant(bestMask.trunc(bestWidth),
1926 DAG.getConstant(0LL, newVT), Cond);
1931 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
1932 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
1933 unsigned InSize = N0.getOperand(0).getValueType().getSizeInBits();
1935 // If the comparison constant has bits in the upper part, the
1936 // zero-extended value could never match.
1937 if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(),
1938 C1.getBitWidth() - InSize))) {
1942 case ISD::SETEQ: return DAG.getConstant(0, VT);
1945 case ISD::SETNE: return DAG.getConstant(1, VT);
1948 // True if the sign bit of C1 is set.
1949 return DAG.getConstant(C1.isNegative(), VT);
1952 // True if the sign bit of C1 isn't set.
1953 return DAG.getConstant(C1.isNonNegative(), VT);
1959 // Otherwise, we can perform the comparison with the low bits.
1967 EVT newVT = N0.getOperand(0).getValueType();
1968 if (DCI.isBeforeLegalizeOps() ||
1969 (isOperationLegal(ISD::SETCC, newVT) &&
1970 getCondCodeAction(Cond, newVT)==Legal))
1971 return DAG.getSetCC(dl, VT, N0.getOperand(0),
1972 DAG.getConstant(C1.trunc(InSize), newVT),
1977 break; // todo, be more careful with signed comparisons
1979 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
1980 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1981 EVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
1982 unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits();
1983 EVT ExtDstTy = N0.getValueType();
1984 unsigned ExtDstTyBits = ExtDstTy.getSizeInBits();
1986 // If the constant doesn't fit into the number of bits for the source of
1987 // the sign extension, it is impossible for both sides to be equal.
1988 if (C1.getMinSignedBits() > ExtSrcTyBits)
1989 return DAG.getConstant(Cond == ISD::SETNE, VT);
1992 EVT Op0Ty = N0.getOperand(0).getValueType();
1993 if (Op0Ty == ExtSrcTy) {
1994 ZextOp = N0.getOperand(0);
1996 APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits);
1997 ZextOp = DAG.getNode(ISD::AND, dl, Op0Ty, N0.getOperand(0),
1998 DAG.getConstant(Imm, Op0Ty));
2000 if (!DCI.isCalledByLegalizer())
2001 DCI.AddToWorklist(ZextOp.getNode());
2002 // Otherwise, make this a use of a zext.
2003 return DAG.getSetCC(dl, VT, ZextOp,
2004 DAG.getConstant(C1 & APInt::getLowBitsSet(
2009 } else if ((N1C->isNullValue() || N1C->getAPIntValue() == 1) &&
2010 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
2011 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC
2012 if (N0.getOpcode() == ISD::SETCC &&
2013 isTypeLegal(VT) && VT.bitsLE(N0.getValueType())) {
2014 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getAPIntValue() != 1);
2016 return DAG.getNode(ISD::TRUNCATE, dl, VT, N0);
2017 // Invert the condition.
2018 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
2019 CC = ISD::getSetCCInverse(CC,
2020 N0.getOperand(0).getValueType().isInteger());
2021 return DAG.getSetCC(dl, VT, N0.getOperand(0), N0.getOperand(1), CC);
2024 if ((N0.getOpcode() == ISD::XOR ||
2025 (N0.getOpcode() == ISD::AND &&
2026 N0.getOperand(0).getOpcode() == ISD::XOR &&
2027 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
2028 isa<ConstantSDNode>(N0.getOperand(1)) &&
2029 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue() == 1) {
2030 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We
2031 // can only do this if the top bits are known zero.
2032 unsigned BitWidth = N0.getValueSizeInBits();
2033 if (DAG.MaskedValueIsZero(N0,
2034 APInt::getHighBitsSet(BitWidth,
2036 // Okay, get the un-inverted input value.
2038 if (N0.getOpcode() == ISD::XOR)
2039 Val = N0.getOperand(0);
2041 assert(N0.getOpcode() == ISD::AND &&
2042 N0.getOperand(0).getOpcode() == ISD::XOR);
2043 // ((X^1)&1)^1 -> X & 1
2044 Val = DAG.getNode(ISD::AND, dl, N0.getValueType(),
2045 N0.getOperand(0).getOperand(0),
2049 return DAG.getSetCC(dl, VT, Val, N1,
2050 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
2052 } else if (N1C->getAPIntValue() == 1 &&
2054 getBooleanContents() == ZeroOrOneBooleanContent)) {
2056 if (Op0.getOpcode() == ISD::TRUNCATE)
2057 Op0 = Op0.getOperand(0);
2059 if ((Op0.getOpcode() == ISD::XOR) &&
2060 Op0.getOperand(0).getOpcode() == ISD::SETCC &&
2061 Op0.getOperand(1).getOpcode() == ISD::SETCC) {
2062 // (xor (setcc), (setcc)) == / != 1 -> (setcc) != / == (setcc)
2063 Cond = (Cond == ISD::SETEQ) ? ISD::SETNE : ISD::SETEQ;
2064 return DAG.getSetCC(dl, VT, Op0.getOperand(0), Op0.getOperand(1),
2066 } else if (Op0.getOpcode() == ISD::AND &&
2067 isa<ConstantSDNode>(Op0.getOperand(1)) &&
2068 cast<ConstantSDNode>(Op0.getOperand(1))->getAPIntValue() == 1) {
2069 // If this is (X&1) == / != 1, normalize it to (X&1) != / == 0.
2070 if (Op0.getValueType().bitsGT(VT))
2071 Op0 = DAG.getNode(ISD::AND, dl, VT,
2072 DAG.getNode(ISD::TRUNCATE, dl, VT, Op0.getOperand(0)),
2073 DAG.getConstant(1, VT));
2074 else if (Op0.getValueType().bitsLT(VT))
2075 Op0 = DAG.getNode(ISD::AND, dl, VT,
2076 DAG.getNode(ISD::ANY_EXTEND, dl, VT, Op0.getOperand(0)),
2077 DAG.getConstant(1, VT));
2079 return DAG.getSetCC(dl, VT, Op0,
2080 DAG.getConstant(0, Op0.getValueType()),
2081 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
2086 APInt MinVal, MaxVal;
2087 unsigned OperandBitSize = N1C->getValueType(0).getSizeInBits();
2088 if (ISD::isSignedIntSetCC(Cond)) {
2089 MinVal = APInt::getSignedMinValue(OperandBitSize);
2090 MaxVal = APInt::getSignedMaxValue(OperandBitSize);
2092 MinVal = APInt::getMinValue(OperandBitSize);
2093 MaxVal = APInt::getMaxValue(OperandBitSize);
2096 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
2097 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
2098 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true
2099 // X >= C0 --> X > (C0-1)
2100 return DAG.getSetCC(dl, VT, N0,
2101 DAG.getConstant(C1-1, N1.getValueType()),
2102 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
2105 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
2106 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true
2107 // X <= C0 --> X < (C0+1)
2108 return DAG.getSetCC(dl, VT, N0,
2109 DAG.getConstant(C1+1, N1.getValueType()),
2110 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
2113 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
2114 return DAG.getConstant(0, VT); // X < MIN --> false
2115 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal)
2116 return DAG.getConstant(1, VT); // X >= MIN --> true
2117 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal)
2118 return DAG.getConstant(0, VT); // X > MAX --> false
2119 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal)
2120 return DAG.getConstant(1, VT); // X <= MAX --> true
2122 // Canonicalize setgt X, Min --> setne X, Min
2123 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
2124 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
2125 // Canonicalize setlt X, Max --> setne X, Max
2126 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
2127 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
2129 // If we have setult X, 1, turn it into seteq X, 0
2130 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
2131 return DAG.getSetCC(dl, VT, N0,
2132 DAG.getConstant(MinVal, N0.getValueType()),
2134 // If we have setugt X, Max-1, turn it into seteq X, Max
2135 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
2136 return DAG.getSetCC(dl, VT, N0,
2137 DAG.getConstant(MaxVal, N0.getValueType()),
2140 // If we have "setcc X, C0", check to see if we can shrink the immediate
2143 // SETUGT X, SINTMAX -> SETLT X, 0
2144 if (Cond == ISD::SETUGT &&
2145 C1 == APInt::getSignedMaxValue(OperandBitSize))
2146 return DAG.getSetCC(dl, VT, N0,
2147 DAG.getConstant(0, N1.getValueType()),
2150 // SETULT X, SINTMIN -> SETGT X, -1
2151 if (Cond == ISD::SETULT &&
2152 C1 == APInt::getSignedMinValue(OperandBitSize)) {
2153 SDValue ConstMinusOne =
2154 DAG.getConstant(APInt::getAllOnesValue(OperandBitSize),
2156 return DAG.getSetCC(dl, VT, N0, ConstMinusOne, ISD::SETGT);
2159 // Fold bit comparisons when we can.
2160 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
2161 (VT == N0.getValueType() ||
2162 (isTypeLegal(VT) && VT.bitsLE(N0.getValueType()))) &&
2163 N0.getOpcode() == ISD::AND)
2164 if (ConstantSDNode *AndRHS =
2165 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2166 EVT ShiftTy = DCI.isBeforeLegalize() ?
2167 getPointerTy() : getShiftAmountTy();
2168 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
2169 // Perform the xform if the AND RHS is a single bit.
2170 if (AndRHS->getAPIntValue().isPowerOf2()) {
2171 return DAG.getNode(ISD::TRUNCATE, dl, VT,
2172 DAG.getNode(ISD::SRL, dl, N0.getValueType(), N0,
2173 DAG.getConstant(AndRHS->getAPIntValue().logBase2(), ShiftTy)));
2175 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getAPIntValue()) {
2176 // (X & 8) == 8 --> (X & 8) >> 3
2177 // Perform the xform if C1 is a single bit.
2178 if (C1.isPowerOf2()) {
2179 return DAG.getNode(ISD::TRUNCATE, dl, VT,
2180 DAG.getNode(ISD::SRL, dl, N0.getValueType(), N0,
2181 DAG.getConstant(C1.logBase2(), ShiftTy)));
2187 if (isa<ConstantFPSDNode>(N0.getNode())) {
2188 // Constant fold or commute setcc.
2189 SDValue O = DAG.FoldSetCC(VT, N0, N1, Cond, dl);
2190 if (O.getNode()) return O;
2191 } else if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1.getNode())) {
2192 // If the RHS of an FP comparison is a constant, simplify it away in
2194 if (CFP->getValueAPF().isNaN()) {
2195 // If an operand is known to be a nan, we can fold it.
2196 switch (ISD::getUnorderedFlavor(Cond)) {
2197 default: llvm_unreachable("Unknown flavor!");
2198 case 0: // Known false.
2199 return DAG.getConstant(0, VT);
2200 case 1: // Known true.
2201 return DAG.getConstant(1, VT);
2202 case 2: // Undefined.
2203 return DAG.getUNDEF(VT);
2207 // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the
2208 // constant if knowing that the operand is non-nan is enough. We prefer to
2209 // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to
2211 if (Cond == ISD::SETO || Cond == ISD::SETUO)
2212 return DAG.getSetCC(dl, VT, N0, N0, Cond);
2214 // If the condition is not legal, see if we can find an equivalent one
2216 if (!isCondCodeLegal(Cond, N0.getValueType())) {
2217 // If the comparison was an awkward floating-point == or != and one of
2218 // the comparison operands is infinity or negative infinity, convert the
2219 // condition to a less-awkward <= or >=.
2220 if (CFP->getValueAPF().isInfinity()) {
2221 if (CFP->getValueAPF().isNegative()) {
2222 if (Cond == ISD::SETOEQ &&
2223 isCondCodeLegal(ISD::SETOLE, N0.getValueType()))
2224 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLE);
2225 if (Cond == ISD::SETUEQ &&
2226 isCondCodeLegal(ISD::SETOLE, N0.getValueType()))
2227 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULE);
2228 if (Cond == ISD::SETUNE &&
2229 isCondCodeLegal(ISD::SETUGT, N0.getValueType()))
2230 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGT);
2231 if (Cond == ISD::SETONE &&
2232 isCondCodeLegal(ISD::SETUGT, N0.getValueType()))
2233 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGT);
2235 if (Cond == ISD::SETOEQ &&
2236 isCondCodeLegal(ISD::SETOGE, N0.getValueType()))
2237 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGE);
2238 if (Cond == ISD::SETUEQ &&
2239 isCondCodeLegal(ISD::SETOGE, N0.getValueType()))
2240 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGE);
2241 if (Cond == ISD::SETUNE &&
2242 isCondCodeLegal(ISD::SETULT, N0.getValueType()))
2243 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULT);
2244 if (Cond == ISD::SETONE &&
2245 isCondCodeLegal(ISD::SETULT, N0.getValueType()))
2246 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLT);
2253 // We can always fold X == X for integer setcc's.
2254 if (N0.getValueType().isInteger())
2255 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
2256 unsigned UOF = ISD::getUnorderedFlavor(Cond);
2257 if (UOF == 2) // FP operators that are undefined on NaNs.
2258 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
2259 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
2260 return DAG.getConstant(UOF, VT);
2261 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
2262 // if it is not already.
2263 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
2264 if (NewCond != Cond)
2265 return DAG.getSetCC(dl, VT, N0, N1, NewCond);
2268 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
2269 N0.getValueType().isInteger()) {
2270 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
2271 N0.getOpcode() == ISD::XOR) {
2272 // Simplify (X+Y) == (X+Z) --> Y == Z
2273 if (N0.getOpcode() == N1.getOpcode()) {
2274 if (N0.getOperand(0) == N1.getOperand(0))
2275 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(1), Cond);
2276 if (N0.getOperand(1) == N1.getOperand(1))
2277 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(0), Cond);
2278 if (DAG.isCommutativeBinOp(N0.getOpcode())) {
2279 // If X op Y == Y op X, try other combinations.
2280 if (N0.getOperand(0) == N1.getOperand(1))
2281 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(0),
2283 if (N0.getOperand(1) == N1.getOperand(0))
2284 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(1),
2289 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) {
2290 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2291 // Turn (X+C1) == C2 --> X == C2-C1
2292 if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse()) {
2293 return DAG.getSetCC(dl, VT, N0.getOperand(0),
2294 DAG.getConstant(RHSC->getAPIntValue()-
2295 LHSR->getAPIntValue(),
2296 N0.getValueType()), Cond);
2299 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
2300 if (N0.getOpcode() == ISD::XOR)
2301 // If we know that all of the inverted bits are zero, don't bother
2302 // performing the inversion.
2303 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue()))
2305 DAG.getSetCC(dl, VT, N0.getOperand(0),
2306 DAG.getConstant(LHSR->getAPIntValue() ^
2307 RHSC->getAPIntValue(),
2312 // Turn (C1-X) == C2 --> X == C1-C2
2313 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
2314 if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse()) {
2316 DAG.getSetCC(dl, VT, N0.getOperand(1),
2317 DAG.getConstant(SUBC->getAPIntValue() -
2318 RHSC->getAPIntValue(),
2325 // Simplify (X+Z) == X --> Z == 0
2326 if (N0.getOperand(0) == N1)
2327 return DAG.getSetCC(dl, VT, N0.getOperand(1),
2328 DAG.getConstant(0, N0.getValueType()), Cond);
2329 if (N0.getOperand(1) == N1) {
2330 if (DAG.isCommutativeBinOp(N0.getOpcode()))
2331 return DAG.getSetCC(dl, VT, N0.getOperand(0),
2332 DAG.getConstant(0, N0.getValueType()), Cond);
2333 else if (N0.getNode()->hasOneUse()) {
2334 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
2335 // (Z-X) == X --> Z == X<<1
2336 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(),
2338 DAG.getConstant(1, getShiftAmountTy()));
2339 if (!DCI.isCalledByLegalizer())
2340 DCI.AddToWorklist(SH.getNode());
2341 return DAG.getSetCC(dl, VT, N0.getOperand(0), SH, Cond);
2346 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
2347 N1.getOpcode() == ISD::XOR) {
2348 // Simplify X == (X+Z) --> Z == 0
2349 if (N1.getOperand(0) == N0) {
2350 return DAG.getSetCC(dl, VT, N1.getOperand(1),
2351 DAG.getConstant(0, N1.getValueType()), Cond);
2352 } else if (N1.getOperand(1) == N0) {
2353 if (DAG.isCommutativeBinOp(N1.getOpcode())) {
2354 return DAG.getSetCC(dl, VT, N1.getOperand(0),
2355 DAG.getConstant(0, N1.getValueType()), Cond);
2356 } else if (N1.getNode()->hasOneUse()) {
2357 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
2358 // X == (Z-X) --> X<<1 == Z
2359 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(), N0,
2360 DAG.getConstant(1, getShiftAmountTy()));
2361 if (!DCI.isCalledByLegalizer())
2362 DCI.AddToWorklist(SH.getNode());
2363 return DAG.getSetCC(dl, VT, SH, N1.getOperand(0), Cond);
2368 // Simplify x&y == y to x&y != 0 if y has exactly one bit set.
2369 // Note that where y is variable and is known to have at most
2370 // one bit set (for example, if it is z&1) we cannot do this;
2371 // the expressions are not equivalent when y==0.
2372 if (N0.getOpcode() == ISD::AND)
2373 if (N0.getOperand(0) == N1 || N0.getOperand(1) == N1) {
2374 if (ValueHasExactlyOneBitSet(N1, DAG)) {
2375 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true);
2376 SDValue Zero = DAG.getConstant(0, N1.getValueType());
2377 return DAG.getSetCC(dl, VT, N0, Zero, Cond);
2380 if (N1.getOpcode() == ISD::AND)
2381 if (N1.getOperand(0) == N0 || N1.getOperand(1) == N0) {
2382 if (ValueHasExactlyOneBitSet(N0, DAG)) {
2383 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true);
2384 SDValue Zero = DAG.getConstant(0, N0.getValueType());
2385 return DAG.getSetCC(dl, VT, N1, Zero, Cond);
2390 // Fold away ALL boolean setcc's.
2392 if (N0.getValueType() == MVT::i1 && foldBooleans) {
2394 default: llvm_unreachable("Unknown integer setcc!");
2395 case ISD::SETEQ: // X == Y -> ~(X^Y)
2396 Temp = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1);
2397 N0 = DAG.getNOT(dl, Temp, MVT::i1);
2398 if (!DCI.isCalledByLegalizer())
2399 DCI.AddToWorklist(Temp.getNode());
2401 case ISD::SETNE: // X != Y --> (X^Y)
2402 N0 = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1);
2404 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> ~X & Y
2405 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> ~X & Y
2406 Temp = DAG.getNOT(dl, N0, MVT::i1);
2407 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N1, Temp);
2408 if (!DCI.isCalledByLegalizer())
2409 DCI.AddToWorklist(Temp.getNode());
2411 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> ~Y & X
2412 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> ~Y & X
2413 Temp = DAG.getNOT(dl, N1, MVT::i1);
2414 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N0, Temp);
2415 if (!DCI.isCalledByLegalizer())
2416 DCI.AddToWorklist(Temp.getNode());
2418 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> ~X | Y
2419 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> ~X | Y
2420 Temp = DAG.getNOT(dl, N0, MVT::i1);
2421 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N1, Temp);
2422 if (!DCI.isCalledByLegalizer())
2423 DCI.AddToWorklist(Temp.getNode());
2425 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> ~Y | X
2426 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> ~Y | X
2427 Temp = DAG.getNOT(dl, N1, MVT::i1);
2428 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N0, Temp);
2431 if (VT != MVT::i1) {
2432 if (!DCI.isCalledByLegalizer())
2433 DCI.AddToWorklist(N0.getNode());
2434 // FIXME: If running after legalize, we probably can't do this.
2435 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, N0);
2440 // Could not fold it.
2444 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
2445 /// node is a GlobalAddress + offset.
2446 bool TargetLowering::isGAPlusOffset(SDNode *N, const GlobalValue* &GA,
2447 int64_t &Offset) const {
2448 if (isa<GlobalAddressSDNode>(N)) {
2449 GlobalAddressSDNode *GASD = cast<GlobalAddressSDNode>(N);
2450 GA = GASD->getGlobal();
2451 Offset += GASD->getOffset();
2455 if (N->getOpcode() == ISD::ADD) {
2456 SDValue N1 = N->getOperand(0);
2457 SDValue N2 = N->getOperand(1);
2458 if (isGAPlusOffset(N1.getNode(), GA, Offset)) {
2459 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
2461 Offset += V->getSExtValue();
2464 } else if (isGAPlusOffset(N2.getNode(), GA, Offset)) {
2465 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
2467 Offset += V->getSExtValue();
2476 SDValue TargetLowering::
2477 PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
2478 // Default implementation: no optimization.
2482 //===----------------------------------------------------------------------===//
2483 // Inline Assembler Implementation Methods
2484 //===----------------------------------------------------------------------===//
2487 TargetLowering::ConstraintType
2488 TargetLowering::getConstraintType(const std::string &Constraint) const {
2489 // FIXME: lots more standard ones to handle.
2490 if (Constraint.size() == 1) {
2491 switch (Constraint[0]) {
2493 case 'r': return C_RegisterClass;
2495 case 'o': // offsetable
2496 case 'V': // not offsetable
2498 case 'i': // Simple Integer or Relocatable Constant
2499 case 'n': // Simple Integer
2500 case 'E': // Floating Point Constant
2501 case 'F': // Floating Point Constant
2502 case 's': // Relocatable Constant
2503 case 'p': // Address.
2504 case 'X': // Allow ANY value.
2505 case 'I': // Target registers.
2519 if (Constraint.size() > 1 && Constraint[0] == '{' &&
2520 Constraint[Constraint.size()-1] == '}')
2525 /// LowerXConstraint - try to replace an X constraint, which matches anything,
2526 /// with another that has more specific requirements based on the type of the
2527 /// corresponding operand.
2528 const char *TargetLowering::LowerXConstraint(EVT ConstraintVT) const{
2529 if (ConstraintVT.isInteger())
2531 if (ConstraintVT.isFloatingPoint())
2532 return "f"; // works for many targets
2536 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
2537 /// vector. If it is invalid, don't add anything to Ops.
2538 void TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
2539 char ConstraintLetter,
2540 std::vector<SDValue> &Ops,
2541 SelectionDAG &DAG) const {
2542 switch (ConstraintLetter) {
2544 case 'X': // Allows any operand; labels (basic block) use this.
2545 if (Op.getOpcode() == ISD::BasicBlock) {
2550 case 'i': // Simple Integer or Relocatable Constant
2551 case 'n': // Simple Integer
2552 case 's': { // Relocatable Constant
2553 // These operands are interested in values of the form (GV+C), where C may
2554 // be folded in as an offset of GV, or it may be explicitly added. Also, it
2555 // is possible and fine if either GV or C are missing.
2556 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2557 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
2559 // If we have "(add GV, C)", pull out GV/C
2560 if (Op.getOpcode() == ISD::ADD) {
2561 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
2562 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
2563 if (C == 0 || GA == 0) {
2564 C = dyn_cast<ConstantSDNode>(Op.getOperand(0));
2565 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1));
2567 if (C == 0 || GA == 0)
2571 // If we find a valid operand, map to the TargetXXX version so that the
2572 // value itself doesn't get selected.
2573 if (GA) { // Either &GV or &GV+C
2574 if (ConstraintLetter != 'n') {
2575 int64_t Offs = GA->getOffset();
2576 if (C) Offs += C->getZExtValue();
2577 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(),
2578 C ? C->getDebugLoc() : DebugLoc(),
2579 Op.getValueType(), Offs));
2583 if (C) { // just C, no GV.
2584 // Simple constants are not allowed for 's'.
2585 if (ConstraintLetter != 's') {
2586 // gcc prints these as sign extended. Sign extend value to 64 bits
2587 // now; without this it would get ZExt'd later in
2588 // ScheduleDAGSDNodes::EmitNode, which is very generic.
2589 Ops.push_back(DAG.getTargetConstant(C->getAPIntValue().getSExtValue(),
2599 std::vector<unsigned> TargetLowering::
2600 getRegClassForInlineAsmConstraint(const std::string &Constraint,
2602 return std::vector<unsigned>();
2606 std::pair<unsigned, const TargetRegisterClass*> TargetLowering::
2607 getRegForInlineAsmConstraint(const std::string &Constraint,
2609 if (Constraint[0] != '{')
2610 return std::make_pair(0u, static_cast<TargetRegisterClass*>(0));
2611 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?");
2613 // Remove the braces from around the name.
2614 StringRef RegName(Constraint.data()+1, Constraint.size()-2);
2616 // Figure out which register class contains this reg.
2617 const TargetRegisterInfo *RI = TM.getRegisterInfo();
2618 for (TargetRegisterInfo::regclass_iterator RCI = RI->regclass_begin(),
2619 E = RI->regclass_end(); RCI != E; ++RCI) {
2620 const TargetRegisterClass *RC = *RCI;
2622 // If none of the value types for this register class are valid, we
2623 // can't use it. For example, 64-bit reg classes on 32-bit targets.
2624 bool isLegal = false;
2625 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
2627 if (isTypeLegal(*I)) {
2633 if (!isLegal) continue;
2635 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end();
2637 if (RegName.equals_lower(RI->getName(*I)))
2638 return std::make_pair(*I, RC);
2642 return std::make_pair(0u, static_cast<const TargetRegisterClass*>(0));
2645 //===----------------------------------------------------------------------===//
2646 // Constraint Selection.
2648 /// isMatchingInputConstraint - Return true of this is an input operand that is
2649 /// a matching constraint like "4".
2650 bool TargetLowering::AsmOperandInfo::isMatchingInputConstraint() const {
2651 assert(!ConstraintCode.empty() && "No known constraint!");
2652 return isdigit(ConstraintCode[0]);
2655 /// getMatchedOperand - If this is an input matching constraint, this method
2656 /// returns the output operand it matches.
2657 unsigned TargetLowering::AsmOperandInfo::getMatchedOperand() const {
2658 assert(!ConstraintCode.empty() && "No known constraint!");
2659 return atoi(ConstraintCode.c_str());
2663 /// ParseConstraints - Split up the constraint string from the inline
2664 /// assembly value into the specific constraints and their prefixes,
2665 /// and also tie in the associated operand values.
2666 /// If this returns an empty vector, and if the constraint string itself
2667 /// isn't empty, there was an error parsing.
2668 TargetLowering::AsmOperandInfoVector TargetLowering::ParseConstraints(
2669 ImmutableCallSite CS) const {
2670 /// ConstraintOperands - Information about all of the constraints.
2671 AsmOperandInfoVector ConstraintOperands;
2672 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
2673 unsigned maCount = 0; // Largest number of multiple alternative constraints.
2675 // Do a prepass over the constraints, canonicalizing them, and building up the
2676 // ConstraintOperands list.
2677 InlineAsm::ConstraintInfoVector
2678 ConstraintInfos = IA->ParseConstraints();
2680 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
2681 unsigned ResNo = 0; // ResNo - The result number of the next output.
2683 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
2684 ConstraintOperands.push_back(AsmOperandInfo(ConstraintInfos[i]));
2685 AsmOperandInfo &OpInfo = ConstraintOperands.back();
2687 // Update multiple alternative constraint count.
2688 if (OpInfo.multipleAlternatives.size() > maCount)
2689 maCount = OpInfo.multipleAlternatives.size();
2691 OpInfo.ConstraintVT = MVT::Other;
2693 // Compute the value type for each operand.
2694 switch (OpInfo.Type) {
2695 case InlineAsm::isOutput:
2696 // Indirect outputs just consume an argument.
2697 if (OpInfo.isIndirect) {
2698 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
2702 // The return value of the call is this value. As such, there is no
2703 // corresponding argument.
2704 assert(!CS.getType()->isVoidTy() &&
2706 if (const StructType *STy = dyn_cast<StructType>(CS.getType())) {
2707 OpInfo.ConstraintVT = getValueType(STy->getElementType(ResNo));
2709 assert(ResNo == 0 && "Asm only has one result!");
2710 OpInfo.ConstraintVT = getValueType(CS.getType());
2714 case InlineAsm::isInput:
2715 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
2717 case InlineAsm::isClobber:
2722 if (OpInfo.CallOperandVal) {
2723 const llvm::Type *OpTy = OpInfo.CallOperandVal->getType();
2724 if (OpInfo.isIndirect) {
2725 const llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy);
2727 report_fatal_error("Indirect operand for inline asm not a pointer!");
2728 OpTy = PtrTy->getElementType();
2730 // If OpTy is not a single value, it may be a struct/union that we
2731 // can tile with integers.
2732 if (!OpTy->isSingleValueType() && OpTy->isSized()) {
2733 unsigned BitSize = TD->getTypeSizeInBits(OpTy);
2742 OpInfo.ConstraintVT =
2743 EVT::getEVT(IntegerType::get(OpTy->getContext(), BitSize), true);
2746 } else if (dyn_cast<PointerType>(OpTy)) {
2747 OpInfo.ConstraintVT = MVT::getIntegerVT(8*TD->getPointerSize());
2749 OpInfo.ConstraintVT = EVT::getEVT(OpTy, true);
2754 // If we have multiple alternative constraints, select the best alternative.
2755 if (ConstraintInfos.size()) {
2757 unsigned bestMAIndex = 0;
2758 int bestWeight = -1;
2759 // weight: -1 = invalid match, and 0 = so-so match to 5 = good match.
2762 // Compute the sums of the weights for each alternative, keeping track
2763 // of the best (highest weight) one so far.
2764 for (maIndex = 0; maIndex < maCount; ++maIndex) {
2766 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size();
2767 cIndex != eIndex; ++cIndex) {
2768 AsmOperandInfo& OpInfo = ConstraintOperands[cIndex];
2769 if (OpInfo.Type == InlineAsm::isClobber)
2772 // If this is an output operand with a matching input operand,
2773 // look up the matching input. If their types mismatch, e.g. one
2774 // is an integer, the other is floating point, or their sizes are
2775 // different, flag it as an maCantMatch.
2776 if (OpInfo.hasMatchingInput()) {
2777 AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
2778 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
2779 if ((OpInfo.ConstraintVT.isInteger() !=
2780 Input.ConstraintVT.isInteger()) ||
2781 (OpInfo.ConstraintVT.getSizeInBits() !=
2782 Input.ConstraintVT.getSizeInBits())) {
2783 weightSum = -1; // Can't match.
2788 weight = getMultipleConstraintMatchWeight(OpInfo, maIndex);
2793 weightSum += weight;
2796 if (weightSum > bestWeight) {
2797 bestWeight = weightSum;
2798 bestMAIndex = maIndex;
2802 // Now select chosen alternative in each constraint.
2803 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size();
2804 cIndex != eIndex; ++cIndex) {
2805 AsmOperandInfo& cInfo = ConstraintOperands[cIndex];
2806 if (cInfo.Type == InlineAsm::isClobber)
2808 cInfo.selectAlternative(bestMAIndex);
2813 // Check and hook up tied operands, choose constraint code to use.
2814 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size();
2815 cIndex != eIndex; ++cIndex) {
2816 AsmOperandInfo& OpInfo = ConstraintOperands[cIndex];
2818 // If this is an output operand with a matching input operand, look up the
2819 // matching input. If their types mismatch, e.g. one is an integer, the
2820 // other is floating point, or their sizes are different, flag it as an
2822 if (OpInfo.hasMatchingInput()) {
2823 AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
2825 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
2826 if ((OpInfo.ConstraintVT.isInteger() !=
2827 Input.ConstraintVT.isInteger()) ||
2828 (OpInfo.ConstraintVT.getSizeInBits() !=
2829 Input.ConstraintVT.getSizeInBits())) {
2830 report_fatal_error("Unsupported asm: input constraint"
2831 " with a matching output constraint of"
2832 " incompatible type!");
2839 return ConstraintOperands;
2843 /// getConstraintGenerality - Return an integer indicating how general CT
2845 static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) {
2847 default: llvm_unreachable("Unknown constraint type!");
2848 case TargetLowering::C_Other:
2849 case TargetLowering::C_Unknown:
2851 case TargetLowering::C_Register:
2853 case TargetLowering::C_RegisterClass:
2855 case TargetLowering::C_Memory:
2860 /// Examine constraint type and operand type and determine a weight value.
2861 /// This object must already have been set up with the operand type
2862 /// and the current alternative constraint selected.
2863 TargetLowering::ConstraintWeight
2864 TargetLowering::getMultipleConstraintMatchWeight(
2865 AsmOperandInfo &info, int maIndex) const {
2866 InlineAsm::ConstraintCodeVector *rCodes;
2867 if (maIndex >= (int)info.multipleAlternatives.size())
2868 rCodes = &info.Codes;
2870 rCodes = &info.multipleAlternatives[maIndex].Codes;
2871 ConstraintWeight BestWeight = CW_Invalid;
2873 // Loop over the options, keeping track of the most general one.
2874 for (unsigned i = 0, e = rCodes->size(); i != e; ++i) {
2875 ConstraintWeight weight =
2876 getSingleConstraintMatchWeight(info, (*rCodes)[i].c_str());
2877 if (weight > BestWeight)
2878 BestWeight = weight;
2884 /// Examine constraint type and operand type and determine a weight value.
2885 /// This object must already have been set up with the operand type
2886 /// and the current alternative constraint selected.
2887 TargetLowering::ConstraintWeight
2888 TargetLowering::getSingleConstraintMatchWeight(
2889 AsmOperandInfo &info, const char *constraint) const {
2890 ConstraintWeight weight = CW_Invalid;
2891 Value *CallOperandVal = info.CallOperandVal;
2892 // If we don't have a value, we can't do a match,
2893 // but allow it at the lowest weight.
2894 if (CallOperandVal == NULL)
2896 // Look at the constraint type.
2897 switch (*constraint) {
2898 case 'i': // immediate integer.
2899 case 'n': // immediate integer with a known value.
2900 if (isa<ConstantInt>(CallOperandVal))
2901 weight = CW_Constant;
2903 case 's': // non-explicit intregal immediate.
2904 if (isa<GlobalValue>(CallOperandVal))
2905 weight = CW_Constant;
2907 case 'E': // immediate float if host format.
2908 case 'F': // immediate float.
2909 if (isa<ConstantFP>(CallOperandVal))
2910 weight = CW_Constant;
2912 case '<': // memory operand with autodecrement.
2913 case '>': // memory operand with autoincrement.
2914 case 'm': // memory operand.
2915 case 'o': // offsettable memory operand
2916 case 'V': // non-offsettable memory operand
2919 case 'r': // general register.
2920 case 'g': // general register, memory operand or immediate integer.
2921 // note: Clang converts "g" to "imr".
2922 if (CallOperandVal->getType()->isIntegerTy())
2923 weight = CW_Register;
2925 case 'X': // any operand.
2927 weight = CW_Default;
2933 /// ChooseConstraint - If there are multiple different constraints that we
2934 /// could pick for this operand (e.g. "imr") try to pick the 'best' one.
2935 /// This is somewhat tricky: constraints fall into four classes:
2936 /// Other -> immediates and magic values
2937 /// Register -> one specific register
2938 /// RegisterClass -> a group of regs
2939 /// Memory -> memory
2940 /// Ideally, we would pick the most specific constraint possible: if we have
2941 /// something that fits into a register, we would pick it. The problem here
2942 /// is that if we have something that could either be in a register or in
2943 /// memory that use of the register could cause selection of *other*
2944 /// operands to fail: they might only succeed if we pick memory. Because of
2945 /// this the heuristic we use is:
2947 /// 1) If there is an 'other' constraint, and if the operand is valid for
2948 /// that constraint, use it. This makes us take advantage of 'i'
2949 /// constraints when available.
2950 /// 2) Otherwise, pick the most general constraint present. This prefers
2951 /// 'm' over 'r', for example.
2953 static void ChooseConstraint(TargetLowering::AsmOperandInfo &OpInfo,
2954 const TargetLowering &TLI,
2955 SDValue Op, SelectionDAG *DAG) {
2956 assert(OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options");
2957 unsigned BestIdx = 0;
2958 TargetLowering::ConstraintType BestType = TargetLowering::C_Unknown;
2959 int BestGenerality = -1;
2961 // Loop over the options, keeping track of the most general one.
2962 for (unsigned i = 0, e = OpInfo.Codes.size(); i != e; ++i) {
2963 TargetLowering::ConstraintType CType =
2964 TLI.getConstraintType(OpInfo.Codes[i]);
2966 // If this is an 'other' constraint, see if the operand is valid for it.
2967 // For example, on X86 we might have an 'rI' constraint. If the operand
2968 // is an integer in the range [0..31] we want to use I (saving a load
2969 // of a register), otherwise we must use 'r'.
2970 if (CType == TargetLowering::C_Other && Op.getNode()) {
2971 assert(OpInfo.Codes[i].size() == 1 &&
2972 "Unhandled multi-letter 'other' constraint");
2973 std::vector<SDValue> ResultOps;
2974 TLI.LowerAsmOperandForConstraint(Op, OpInfo.Codes[i][0],
2976 if (!ResultOps.empty()) {
2983 // Things with matching constraints can only be registers, per gcc
2984 // documentation. This mainly affects "g" constraints.
2985 if (CType == TargetLowering::C_Memory && OpInfo.hasMatchingInput())
2988 // This constraint letter is more general than the previous one, use it.
2989 int Generality = getConstraintGenerality(CType);
2990 if (Generality > BestGenerality) {
2993 BestGenerality = Generality;
2997 OpInfo.ConstraintCode = OpInfo.Codes[BestIdx];
2998 OpInfo.ConstraintType = BestType;
3001 /// ComputeConstraintToUse - Determines the constraint code and constraint
3002 /// type to use for the specific AsmOperandInfo, setting
3003 /// OpInfo.ConstraintCode and OpInfo.ConstraintType.
3004 void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo,
3006 SelectionDAG *DAG) const {
3007 assert(!OpInfo.Codes.empty() && "Must have at least one constraint");
3009 // Single-letter constraints ('r') are very common.
3010 if (OpInfo.Codes.size() == 1) {
3011 OpInfo.ConstraintCode = OpInfo.Codes[0];
3012 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
3014 ChooseConstraint(OpInfo, *this, Op, DAG);
3017 // 'X' matches anything.
3018 if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) {
3019 // Labels and constants are handled elsewhere ('X' is the only thing
3020 // that matches labels). For Functions, the type here is the type of
3021 // the result, which is not what we want to look at; leave them alone.
3022 Value *v = OpInfo.CallOperandVal;
3023 if (isa<BasicBlock>(v) || isa<ConstantInt>(v) || isa<Function>(v)) {
3024 OpInfo.CallOperandVal = v;
3028 // Otherwise, try to resolve it to something we know about by looking at
3029 // the actual operand type.
3030 if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) {
3031 OpInfo.ConstraintCode = Repl;
3032 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
3037 //===----------------------------------------------------------------------===//
3038 // Loop Strength Reduction hooks
3039 //===----------------------------------------------------------------------===//
3041 /// isLegalAddressingMode - Return true if the addressing mode represented
3042 /// by AM is legal for this target, for a load/store of the specified type.
3043 bool TargetLowering::isLegalAddressingMode(const AddrMode &AM,
3044 const Type *Ty) const {
3045 // The default implementation of this implements a conservative RISCy, r+r and
3048 // Allows a sign-extended 16-bit immediate field.
3049 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
3052 // No global is ever allowed as a base.
3056 // Only support r+r,
3058 case 0: // "r+i" or just "i", depending on HasBaseReg.
3061 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
3063 // Otherwise we have r+r or r+i.
3066 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
3068 // Allow 2*r as r+r.
3075 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
3076 /// return a DAG expression to select that will generate the same value by
3077 /// multiplying by a magic number. See:
3078 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
3079 SDValue TargetLowering::BuildSDIV(SDNode *N, SelectionDAG &DAG,
3080 std::vector<SDNode*>* Created) const {
3081 EVT VT = N->getValueType(0);
3082 DebugLoc dl= N->getDebugLoc();
3084 // Check to see if we can do this.
3085 // FIXME: We should be more aggressive here.
3086 if (!isTypeLegal(VT))
3089 APInt d = cast<ConstantSDNode>(N->getOperand(1))->getAPIntValue();
3090 APInt::ms magics = d.magic();
3092 // Multiply the numerator (operand 0) by the magic value
3093 // FIXME: We should support doing a MUL in a wider type
3095 if (isOperationLegalOrCustom(ISD::MULHS, VT))
3096 Q = DAG.getNode(ISD::MULHS, dl, VT, N->getOperand(0),
3097 DAG.getConstant(magics.m, VT));
3098 else if (isOperationLegalOrCustom(ISD::SMUL_LOHI, VT))
3099 Q = SDValue(DAG.getNode(ISD::SMUL_LOHI, dl, DAG.getVTList(VT, VT),
3101 DAG.getConstant(magics.m, VT)).getNode(), 1);
3103 return SDValue(); // No mulhs or equvialent
3104 // If d > 0 and m < 0, add the numerator
3105 if (d.isStrictlyPositive() && magics.m.isNegative()) {
3106 Q = DAG.getNode(ISD::ADD, dl, VT, Q, N->getOperand(0));
3108 Created->push_back(Q.getNode());
3110 // If d < 0 and m > 0, subtract the numerator.
3111 if (d.isNegative() && magics.m.isStrictlyPositive()) {
3112 Q = DAG.getNode(ISD::SUB, dl, VT, Q, N->getOperand(0));
3114 Created->push_back(Q.getNode());
3116 // Shift right algebraic if shift value is nonzero
3118 Q = DAG.getNode(ISD::SRA, dl, VT, Q,
3119 DAG.getConstant(magics.s, getShiftAmountTy()));
3121 Created->push_back(Q.getNode());
3123 // Extract the sign bit and add it to the quotient
3125 DAG.getNode(ISD::SRL, dl, VT, Q, DAG.getConstant(VT.getSizeInBits()-1,
3126 getShiftAmountTy()));
3128 Created->push_back(T.getNode());
3129 return DAG.getNode(ISD::ADD, dl, VT, Q, T);
3132 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
3133 /// return a DAG expression to select that will generate the same value by
3134 /// multiplying by a magic number. See:
3135 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
3136 SDValue TargetLowering::BuildUDIV(SDNode *N, SelectionDAG &DAG,
3137 std::vector<SDNode*>* Created) const {
3138 EVT VT = N->getValueType(0);
3139 DebugLoc dl = N->getDebugLoc();
3141 // Check to see if we can do this.
3142 // FIXME: We should be more aggressive here.
3143 if (!isTypeLegal(VT))
3146 // FIXME: We should use a narrower constant when the upper
3147 // bits are known to be zero.
3148 ConstantSDNode *N1C = cast<ConstantSDNode>(N->getOperand(1));
3149 APInt::mu magics = N1C->getAPIntValue().magicu();
3151 // Multiply the numerator (operand 0) by the magic value
3152 // FIXME: We should support doing a MUL in a wider type
3154 if (isOperationLegalOrCustom(ISD::MULHU, VT))
3155 Q = DAG.getNode(ISD::MULHU, dl, VT, N->getOperand(0),
3156 DAG.getConstant(magics.m, VT));
3157 else if (isOperationLegalOrCustom(ISD::UMUL_LOHI, VT))
3158 Q = SDValue(DAG.getNode(ISD::UMUL_LOHI, dl, DAG.getVTList(VT, VT),
3160 DAG.getConstant(magics.m, VT)).getNode(), 1);
3162 return SDValue(); // No mulhu or equvialent
3164 Created->push_back(Q.getNode());
3166 if (magics.a == 0) {
3167 assert(magics.s < N1C->getAPIntValue().getBitWidth() &&
3168 "We shouldn't generate an undefined shift!");
3169 return DAG.getNode(ISD::SRL, dl, VT, Q,
3170 DAG.getConstant(magics.s, getShiftAmountTy()));
3172 SDValue NPQ = DAG.getNode(ISD::SUB, dl, VT, N->getOperand(0), Q);
3174 Created->push_back(NPQ.getNode());
3175 NPQ = DAG.getNode(ISD::SRL, dl, VT, NPQ,
3176 DAG.getConstant(1, getShiftAmountTy()));
3178 Created->push_back(NPQ.getNode());
3179 NPQ = DAG.getNode(ISD::ADD, dl, VT, NPQ, Q);
3181 Created->push_back(NPQ.getNode());
3182 return DAG.getNode(ISD::SRL, dl, VT, NPQ,
3183 DAG.getConstant(magics.s-1, getShiftAmountTy()));