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"
34 TLSModel::Model getTLSModel(const GlobalValue *GV, Reloc::Model reloc) {
35 bool isLocal = GV->hasLocalLinkage();
36 bool isDeclaration = GV->isDeclaration();
37 // FIXME: what should we do for protected and internal visibility?
38 // For variables, is internal different from hidden?
39 bool isHidden = GV->hasHiddenVisibility();
41 if (reloc == Reloc::PIC_) {
42 if (isLocal || isHidden)
43 return TLSModel::LocalDynamic;
45 return TLSModel::GeneralDynamic;
47 if (!isDeclaration || isHidden)
48 return TLSModel::LocalExec;
50 return TLSModel::InitialExec;
55 /// InitLibcallNames - Set default libcall names.
57 static void InitLibcallNames(const char **Names) {
58 Names[RTLIB::SHL_I16] = "__ashlhi3";
59 Names[RTLIB::SHL_I32] = "__ashlsi3";
60 Names[RTLIB::SHL_I64] = "__ashldi3";
61 Names[RTLIB::SHL_I128] = "__ashlti3";
62 Names[RTLIB::SRL_I16] = "__lshrhi3";
63 Names[RTLIB::SRL_I32] = "__lshrsi3";
64 Names[RTLIB::SRL_I64] = "__lshrdi3";
65 Names[RTLIB::SRL_I128] = "__lshrti3";
66 Names[RTLIB::SRA_I16] = "__ashrhi3";
67 Names[RTLIB::SRA_I32] = "__ashrsi3";
68 Names[RTLIB::SRA_I64] = "__ashrdi3";
69 Names[RTLIB::SRA_I128] = "__ashrti3";
70 Names[RTLIB::MUL_I8] = "__mulqi3";
71 Names[RTLIB::MUL_I16] = "__mulhi3";
72 Names[RTLIB::MUL_I32] = "__mulsi3";
73 Names[RTLIB::MUL_I64] = "__muldi3";
74 Names[RTLIB::MUL_I128] = "__multi3";
75 Names[RTLIB::SDIV_I8] = "__divqi3";
76 Names[RTLIB::SDIV_I16] = "__divhi3";
77 Names[RTLIB::SDIV_I32] = "__divsi3";
78 Names[RTLIB::SDIV_I64] = "__divdi3";
79 Names[RTLIB::SDIV_I128] = "__divti3";
80 Names[RTLIB::UDIV_I8] = "__udivqi3";
81 Names[RTLIB::UDIV_I16] = "__udivhi3";
82 Names[RTLIB::UDIV_I32] = "__udivsi3";
83 Names[RTLIB::UDIV_I64] = "__udivdi3";
84 Names[RTLIB::UDIV_I128] = "__udivti3";
85 Names[RTLIB::SREM_I8] = "__modqi3";
86 Names[RTLIB::SREM_I16] = "__modhi3";
87 Names[RTLIB::SREM_I32] = "__modsi3";
88 Names[RTLIB::SREM_I64] = "__moddi3";
89 Names[RTLIB::SREM_I128] = "__modti3";
90 Names[RTLIB::UREM_I8] = "__umodqi3";
91 Names[RTLIB::UREM_I16] = "__umodhi3";
92 Names[RTLIB::UREM_I32] = "__umodsi3";
93 Names[RTLIB::UREM_I64] = "__umoddi3";
94 Names[RTLIB::UREM_I128] = "__umodti3";
95 Names[RTLIB::NEG_I32] = "__negsi2";
96 Names[RTLIB::NEG_I64] = "__negdi2";
97 Names[RTLIB::ADD_F32] = "__addsf3";
98 Names[RTLIB::ADD_F64] = "__adddf3";
99 Names[RTLIB::ADD_F80] = "__addxf3";
100 Names[RTLIB::ADD_PPCF128] = "__gcc_qadd";
101 Names[RTLIB::SUB_F32] = "__subsf3";
102 Names[RTLIB::SUB_F64] = "__subdf3";
103 Names[RTLIB::SUB_F80] = "__subxf3";
104 Names[RTLIB::SUB_PPCF128] = "__gcc_qsub";
105 Names[RTLIB::MUL_F32] = "__mulsf3";
106 Names[RTLIB::MUL_F64] = "__muldf3";
107 Names[RTLIB::MUL_F80] = "__mulxf3";
108 Names[RTLIB::MUL_PPCF128] = "__gcc_qmul";
109 Names[RTLIB::DIV_F32] = "__divsf3";
110 Names[RTLIB::DIV_F64] = "__divdf3";
111 Names[RTLIB::DIV_F80] = "__divxf3";
112 Names[RTLIB::DIV_PPCF128] = "__gcc_qdiv";
113 Names[RTLIB::REM_F32] = "fmodf";
114 Names[RTLIB::REM_F64] = "fmod";
115 Names[RTLIB::REM_F80] = "fmodl";
116 Names[RTLIB::REM_PPCF128] = "fmodl";
117 Names[RTLIB::POWI_F32] = "__powisf2";
118 Names[RTLIB::POWI_F64] = "__powidf2";
119 Names[RTLIB::POWI_F80] = "__powixf2";
120 Names[RTLIB::POWI_PPCF128] = "__powitf2";
121 Names[RTLIB::SQRT_F32] = "sqrtf";
122 Names[RTLIB::SQRT_F64] = "sqrt";
123 Names[RTLIB::SQRT_F80] = "sqrtl";
124 Names[RTLIB::SQRT_PPCF128] = "sqrtl";
125 Names[RTLIB::LOG_F32] = "logf";
126 Names[RTLIB::LOG_F64] = "log";
127 Names[RTLIB::LOG_F80] = "logl";
128 Names[RTLIB::LOG_PPCF128] = "logl";
129 Names[RTLIB::LOG2_F32] = "log2f";
130 Names[RTLIB::LOG2_F64] = "log2";
131 Names[RTLIB::LOG2_F80] = "log2l";
132 Names[RTLIB::LOG2_PPCF128] = "log2l";
133 Names[RTLIB::LOG10_F32] = "log10f";
134 Names[RTLIB::LOG10_F64] = "log10";
135 Names[RTLIB::LOG10_F80] = "log10l";
136 Names[RTLIB::LOG10_PPCF128] = "log10l";
137 Names[RTLIB::EXP_F32] = "expf";
138 Names[RTLIB::EXP_F64] = "exp";
139 Names[RTLIB::EXP_F80] = "expl";
140 Names[RTLIB::EXP_PPCF128] = "expl";
141 Names[RTLIB::EXP2_F32] = "exp2f";
142 Names[RTLIB::EXP2_F64] = "exp2";
143 Names[RTLIB::EXP2_F80] = "exp2l";
144 Names[RTLIB::EXP2_PPCF128] = "exp2l";
145 Names[RTLIB::SIN_F32] = "sinf";
146 Names[RTLIB::SIN_F64] = "sin";
147 Names[RTLIB::SIN_F80] = "sinl";
148 Names[RTLIB::SIN_PPCF128] = "sinl";
149 Names[RTLIB::COS_F32] = "cosf";
150 Names[RTLIB::COS_F64] = "cos";
151 Names[RTLIB::COS_F80] = "cosl";
152 Names[RTLIB::COS_PPCF128] = "cosl";
153 Names[RTLIB::POW_F32] = "powf";
154 Names[RTLIB::POW_F64] = "pow";
155 Names[RTLIB::POW_F80] = "powl";
156 Names[RTLIB::POW_PPCF128] = "powl";
157 Names[RTLIB::CEIL_F32] = "ceilf";
158 Names[RTLIB::CEIL_F64] = "ceil";
159 Names[RTLIB::CEIL_F80] = "ceill";
160 Names[RTLIB::CEIL_PPCF128] = "ceill";
161 Names[RTLIB::TRUNC_F32] = "truncf";
162 Names[RTLIB::TRUNC_F64] = "trunc";
163 Names[RTLIB::TRUNC_F80] = "truncl";
164 Names[RTLIB::TRUNC_PPCF128] = "truncl";
165 Names[RTLIB::RINT_F32] = "rintf";
166 Names[RTLIB::RINT_F64] = "rint";
167 Names[RTLIB::RINT_F80] = "rintl";
168 Names[RTLIB::RINT_PPCF128] = "rintl";
169 Names[RTLIB::NEARBYINT_F32] = "nearbyintf";
170 Names[RTLIB::NEARBYINT_F64] = "nearbyint";
171 Names[RTLIB::NEARBYINT_F80] = "nearbyintl";
172 Names[RTLIB::NEARBYINT_PPCF128] = "nearbyintl";
173 Names[RTLIB::FLOOR_F32] = "floorf";
174 Names[RTLIB::FLOOR_F64] = "floor";
175 Names[RTLIB::FLOOR_F80] = "floorl";
176 Names[RTLIB::FLOOR_PPCF128] = "floorl";
177 Names[RTLIB::COPYSIGN_F32] = "copysignf";
178 Names[RTLIB::COPYSIGN_F64] = "copysign";
179 Names[RTLIB::COPYSIGN_F80] = "copysignl";
180 Names[RTLIB::COPYSIGN_PPCF128] = "copysignl";
181 Names[RTLIB::FPEXT_F32_F64] = "__extendsfdf2";
182 Names[RTLIB::FPEXT_F16_F32] = "__gnu_h2f_ieee";
183 Names[RTLIB::FPROUND_F32_F16] = "__gnu_f2h_ieee";
184 Names[RTLIB::FPROUND_F64_F32] = "__truncdfsf2";
185 Names[RTLIB::FPROUND_F80_F32] = "__truncxfsf2";
186 Names[RTLIB::FPROUND_PPCF128_F32] = "__trunctfsf2";
187 Names[RTLIB::FPROUND_F80_F64] = "__truncxfdf2";
188 Names[RTLIB::FPROUND_PPCF128_F64] = "__trunctfdf2";
189 Names[RTLIB::FPTOSINT_F32_I8] = "__fixsfqi";
190 Names[RTLIB::FPTOSINT_F32_I16] = "__fixsfhi";
191 Names[RTLIB::FPTOSINT_F32_I32] = "__fixsfsi";
192 Names[RTLIB::FPTOSINT_F32_I64] = "__fixsfdi";
193 Names[RTLIB::FPTOSINT_F32_I128] = "__fixsfti";
194 Names[RTLIB::FPTOSINT_F64_I8] = "__fixdfqi";
195 Names[RTLIB::FPTOSINT_F64_I16] = "__fixdfhi";
196 Names[RTLIB::FPTOSINT_F64_I32] = "__fixdfsi";
197 Names[RTLIB::FPTOSINT_F64_I64] = "__fixdfdi";
198 Names[RTLIB::FPTOSINT_F64_I128] = "__fixdfti";
199 Names[RTLIB::FPTOSINT_F80_I32] = "__fixxfsi";
200 Names[RTLIB::FPTOSINT_F80_I64] = "__fixxfdi";
201 Names[RTLIB::FPTOSINT_F80_I128] = "__fixxfti";
202 Names[RTLIB::FPTOSINT_PPCF128_I32] = "__fixtfsi";
203 Names[RTLIB::FPTOSINT_PPCF128_I64] = "__fixtfdi";
204 Names[RTLIB::FPTOSINT_PPCF128_I128] = "__fixtfti";
205 Names[RTLIB::FPTOUINT_F32_I8] = "__fixunssfqi";
206 Names[RTLIB::FPTOUINT_F32_I16] = "__fixunssfhi";
207 Names[RTLIB::FPTOUINT_F32_I32] = "__fixunssfsi";
208 Names[RTLIB::FPTOUINT_F32_I64] = "__fixunssfdi";
209 Names[RTLIB::FPTOUINT_F32_I128] = "__fixunssfti";
210 Names[RTLIB::FPTOUINT_F64_I8] = "__fixunsdfqi";
211 Names[RTLIB::FPTOUINT_F64_I16] = "__fixunsdfhi";
212 Names[RTLIB::FPTOUINT_F64_I32] = "__fixunsdfsi";
213 Names[RTLIB::FPTOUINT_F64_I64] = "__fixunsdfdi";
214 Names[RTLIB::FPTOUINT_F64_I128] = "__fixunsdfti";
215 Names[RTLIB::FPTOUINT_F80_I32] = "__fixunsxfsi";
216 Names[RTLIB::FPTOUINT_F80_I64] = "__fixunsxfdi";
217 Names[RTLIB::FPTOUINT_F80_I128] = "__fixunsxfti";
218 Names[RTLIB::FPTOUINT_PPCF128_I32] = "__fixunstfsi";
219 Names[RTLIB::FPTOUINT_PPCF128_I64] = "__fixunstfdi";
220 Names[RTLIB::FPTOUINT_PPCF128_I128] = "__fixunstfti";
221 Names[RTLIB::SINTTOFP_I32_F32] = "__floatsisf";
222 Names[RTLIB::SINTTOFP_I32_F64] = "__floatsidf";
223 Names[RTLIB::SINTTOFP_I32_F80] = "__floatsixf";
224 Names[RTLIB::SINTTOFP_I32_PPCF128] = "__floatsitf";
225 Names[RTLIB::SINTTOFP_I64_F32] = "__floatdisf";
226 Names[RTLIB::SINTTOFP_I64_F64] = "__floatdidf";
227 Names[RTLIB::SINTTOFP_I64_F80] = "__floatdixf";
228 Names[RTLIB::SINTTOFP_I64_PPCF128] = "__floatditf";
229 Names[RTLIB::SINTTOFP_I128_F32] = "__floattisf";
230 Names[RTLIB::SINTTOFP_I128_F64] = "__floattidf";
231 Names[RTLIB::SINTTOFP_I128_F80] = "__floattixf";
232 Names[RTLIB::SINTTOFP_I128_PPCF128] = "__floattitf";
233 Names[RTLIB::UINTTOFP_I32_F32] = "__floatunsisf";
234 Names[RTLIB::UINTTOFP_I32_F64] = "__floatunsidf";
235 Names[RTLIB::UINTTOFP_I32_F80] = "__floatunsixf";
236 Names[RTLIB::UINTTOFP_I32_PPCF128] = "__floatunsitf";
237 Names[RTLIB::UINTTOFP_I64_F32] = "__floatundisf";
238 Names[RTLIB::UINTTOFP_I64_F64] = "__floatundidf";
239 Names[RTLIB::UINTTOFP_I64_F80] = "__floatundixf";
240 Names[RTLIB::UINTTOFP_I64_PPCF128] = "__floatunditf";
241 Names[RTLIB::UINTTOFP_I128_F32] = "__floatuntisf";
242 Names[RTLIB::UINTTOFP_I128_F64] = "__floatuntidf";
243 Names[RTLIB::UINTTOFP_I128_F80] = "__floatuntixf";
244 Names[RTLIB::UINTTOFP_I128_PPCF128] = "__floatuntitf";
245 Names[RTLIB::OEQ_F32] = "__eqsf2";
246 Names[RTLIB::OEQ_F64] = "__eqdf2";
247 Names[RTLIB::UNE_F32] = "__nesf2";
248 Names[RTLIB::UNE_F64] = "__nedf2";
249 Names[RTLIB::OGE_F32] = "__gesf2";
250 Names[RTLIB::OGE_F64] = "__gedf2";
251 Names[RTLIB::OLT_F32] = "__ltsf2";
252 Names[RTLIB::OLT_F64] = "__ltdf2";
253 Names[RTLIB::OLE_F32] = "__lesf2";
254 Names[RTLIB::OLE_F64] = "__ledf2";
255 Names[RTLIB::OGT_F32] = "__gtsf2";
256 Names[RTLIB::OGT_F64] = "__gtdf2";
257 Names[RTLIB::UO_F32] = "__unordsf2";
258 Names[RTLIB::UO_F64] = "__unorddf2";
259 Names[RTLIB::O_F32] = "__unordsf2";
260 Names[RTLIB::O_F64] = "__unorddf2";
261 Names[RTLIB::MEMCPY] = "memcpy";
262 Names[RTLIB::MEMMOVE] = "memmove";
263 Names[RTLIB::MEMSET] = "memset";
264 Names[RTLIB::UNWIND_RESUME] = "_Unwind_Resume";
265 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_1] = "__sync_val_compare_and_swap_1";
266 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_2] = "__sync_val_compare_and_swap_2";
267 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_4] = "__sync_val_compare_and_swap_4";
268 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_8] = "__sync_val_compare_and_swap_8";
269 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_1] = "__sync_lock_test_and_set_1";
270 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_2] = "__sync_lock_test_and_set_2";
271 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_4] = "__sync_lock_test_and_set_4";
272 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_8] = "__sync_lock_test_and_set_8";
273 Names[RTLIB::SYNC_FETCH_AND_ADD_1] = "__sync_fetch_and_add_1";
274 Names[RTLIB::SYNC_FETCH_AND_ADD_2] = "__sync_fetch_and_add_2";
275 Names[RTLIB::SYNC_FETCH_AND_ADD_4] = "__sync_fetch_and_add_4";
276 Names[RTLIB::SYNC_FETCH_AND_ADD_8] = "__sync_fetch_and_add_8";
277 Names[RTLIB::SYNC_FETCH_AND_SUB_1] = "__sync_fetch_and_sub_1";
278 Names[RTLIB::SYNC_FETCH_AND_SUB_2] = "__sync_fetch_and_sub_2";
279 Names[RTLIB::SYNC_FETCH_AND_SUB_4] = "__sync_fetch_and_sub_4";
280 Names[RTLIB::SYNC_FETCH_AND_SUB_8] = "__sync_fetch_and_sub_8";
281 Names[RTLIB::SYNC_FETCH_AND_AND_1] = "__sync_fetch_and_and_1";
282 Names[RTLIB::SYNC_FETCH_AND_AND_2] = "__sync_fetch_and_and_2";
283 Names[RTLIB::SYNC_FETCH_AND_AND_4] = "__sync_fetch_and_and_4";
284 Names[RTLIB::SYNC_FETCH_AND_AND_8] = "__sync_fetch_and_and_8";
285 Names[RTLIB::SYNC_FETCH_AND_OR_1] = "__sync_fetch_and_or_1";
286 Names[RTLIB::SYNC_FETCH_AND_OR_2] = "__sync_fetch_and_or_2";
287 Names[RTLIB::SYNC_FETCH_AND_OR_4] = "__sync_fetch_and_or_4";
288 Names[RTLIB::SYNC_FETCH_AND_OR_8] = "__sync_fetch_and_or_8";
289 Names[RTLIB::SYNC_FETCH_AND_XOR_1] = "__sync_fetch_and_xor_1";
290 Names[RTLIB::SYNC_FETCH_AND_XOR_2] = "__sync_fetch_and_xor_2";
291 Names[RTLIB::SYNC_FETCH_AND_XOR_4] = "__sync_fetch_and-xor_4";
292 Names[RTLIB::SYNC_FETCH_AND_XOR_8] = "__sync_fetch_and_xor_8";
293 Names[RTLIB::SYNC_FETCH_AND_NAND_1] = "__sync_fetch_and_nand_1";
294 Names[RTLIB::SYNC_FETCH_AND_NAND_2] = "__sync_fetch_and_nand_2";
295 Names[RTLIB::SYNC_FETCH_AND_NAND_4] = "__sync_fetch_and_nand_4";
296 Names[RTLIB::SYNC_FETCH_AND_NAND_8] = "__sync_fetch_and_nand_8";
299 /// InitLibcallCallingConvs - Set default libcall CallingConvs.
301 static void InitLibcallCallingConvs(CallingConv::ID *CCs) {
302 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
303 CCs[i] = CallingConv::C;
307 /// getFPEXT - Return the FPEXT_*_* value for the given types, or
308 /// UNKNOWN_LIBCALL if there is none.
309 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) {
310 if (OpVT == MVT::f32) {
311 if (RetVT == MVT::f64)
312 return FPEXT_F32_F64;
315 return UNKNOWN_LIBCALL;
318 /// getFPROUND - Return the FPROUND_*_* value for the given types, or
319 /// UNKNOWN_LIBCALL if there is none.
320 RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) {
321 if (RetVT == MVT::f32) {
322 if (OpVT == MVT::f64)
323 return FPROUND_F64_F32;
324 if (OpVT == MVT::f80)
325 return FPROUND_F80_F32;
326 if (OpVT == MVT::ppcf128)
327 return FPROUND_PPCF128_F32;
328 } else if (RetVT == MVT::f64) {
329 if (OpVT == MVT::f80)
330 return FPROUND_F80_F64;
331 if (OpVT == MVT::ppcf128)
332 return FPROUND_PPCF128_F64;
335 return UNKNOWN_LIBCALL;
338 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or
339 /// UNKNOWN_LIBCALL if there is none.
340 RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) {
341 if (OpVT == MVT::f32) {
342 if (RetVT == MVT::i8)
343 return FPTOSINT_F32_I8;
344 if (RetVT == MVT::i16)
345 return FPTOSINT_F32_I16;
346 if (RetVT == MVT::i32)
347 return FPTOSINT_F32_I32;
348 if (RetVT == MVT::i64)
349 return FPTOSINT_F32_I64;
350 if (RetVT == MVT::i128)
351 return FPTOSINT_F32_I128;
352 } else if (OpVT == MVT::f64) {
353 if (RetVT == MVT::i8)
354 return FPTOSINT_F64_I8;
355 if (RetVT == MVT::i16)
356 return FPTOSINT_F64_I16;
357 if (RetVT == MVT::i32)
358 return FPTOSINT_F64_I32;
359 if (RetVT == MVT::i64)
360 return FPTOSINT_F64_I64;
361 if (RetVT == MVT::i128)
362 return FPTOSINT_F64_I128;
363 } else if (OpVT == MVT::f80) {
364 if (RetVT == MVT::i32)
365 return FPTOSINT_F80_I32;
366 if (RetVT == MVT::i64)
367 return FPTOSINT_F80_I64;
368 if (RetVT == MVT::i128)
369 return FPTOSINT_F80_I128;
370 } else if (OpVT == MVT::ppcf128) {
371 if (RetVT == MVT::i32)
372 return FPTOSINT_PPCF128_I32;
373 if (RetVT == MVT::i64)
374 return FPTOSINT_PPCF128_I64;
375 if (RetVT == MVT::i128)
376 return FPTOSINT_PPCF128_I128;
378 return UNKNOWN_LIBCALL;
381 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or
382 /// UNKNOWN_LIBCALL if there is none.
383 RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) {
384 if (OpVT == MVT::f32) {
385 if (RetVT == MVT::i8)
386 return FPTOUINT_F32_I8;
387 if (RetVT == MVT::i16)
388 return FPTOUINT_F32_I16;
389 if (RetVT == MVT::i32)
390 return FPTOUINT_F32_I32;
391 if (RetVT == MVT::i64)
392 return FPTOUINT_F32_I64;
393 if (RetVT == MVT::i128)
394 return FPTOUINT_F32_I128;
395 } else if (OpVT == MVT::f64) {
396 if (RetVT == MVT::i8)
397 return FPTOUINT_F64_I8;
398 if (RetVT == MVT::i16)
399 return FPTOUINT_F64_I16;
400 if (RetVT == MVT::i32)
401 return FPTOUINT_F64_I32;
402 if (RetVT == MVT::i64)
403 return FPTOUINT_F64_I64;
404 if (RetVT == MVT::i128)
405 return FPTOUINT_F64_I128;
406 } else if (OpVT == MVT::f80) {
407 if (RetVT == MVT::i32)
408 return FPTOUINT_F80_I32;
409 if (RetVT == MVT::i64)
410 return FPTOUINT_F80_I64;
411 if (RetVT == MVT::i128)
412 return FPTOUINT_F80_I128;
413 } else if (OpVT == MVT::ppcf128) {
414 if (RetVT == MVT::i32)
415 return FPTOUINT_PPCF128_I32;
416 if (RetVT == MVT::i64)
417 return FPTOUINT_PPCF128_I64;
418 if (RetVT == MVT::i128)
419 return FPTOUINT_PPCF128_I128;
421 return UNKNOWN_LIBCALL;
424 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or
425 /// UNKNOWN_LIBCALL if there is none.
426 RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) {
427 if (OpVT == MVT::i32) {
428 if (RetVT == MVT::f32)
429 return SINTTOFP_I32_F32;
430 else if (RetVT == MVT::f64)
431 return SINTTOFP_I32_F64;
432 else if (RetVT == MVT::f80)
433 return SINTTOFP_I32_F80;
434 else if (RetVT == MVT::ppcf128)
435 return SINTTOFP_I32_PPCF128;
436 } else if (OpVT == MVT::i64) {
437 if (RetVT == MVT::f32)
438 return SINTTOFP_I64_F32;
439 else if (RetVT == MVT::f64)
440 return SINTTOFP_I64_F64;
441 else if (RetVT == MVT::f80)
442 return SINTTOFP_I64_F80;
443 else if (RetVT == MVT::ppcf128)
444 return SINTTOFP_I64_PPCF128;
445 } else if (OpVT == MVT::i128) {
446 if (RetVT == MVT::f32)
447 return SINTTOFP_I128_F32;
448 else if (RetVT == MVT::f64)
449 return SINTTOFP_I128_F64;
450 else if (RetVT == MVT::f80)
451 return SINTTOFP_I128_F80;
452 else if (RetVT == MVT::ppcf128)
453 return SINTTOFP_I128_PPCF128;
455 return UNKNOWN_LIBCALL;
458 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or
459 /// UNKNOWN_LIBCALL if there is none.
460 RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) {
461 if (OpVT == MVT::i32) {
462 if (RetVT == MVT::f32)
463 return UINTTOFP_I32_F32;
464 else if (RetVT == MVT::f64)
465 return UINTTOFP_I32_F64;
466 else if (RetVT == MVT::f80)
467 return UINTTOFP_I32_F80;
468 else if (RetVT == MVT::ppcf128)
469 return UINTTOFP_I32_PPCF128;
470 } else if (OpVT == MVT::i64) {
471 if (RetVT == MVT::f32)
472 return UINTTOFP_I64_F32;
473 else if (RetVT == MVT::f64)
474 return UINTTOFP_I64_F64;
475 else if (RetVT == MVT::f80)
476 return UINTTOFP_I64_F80;
477 else if (RetVT == MVT::ppcf128)
478 return UINTTOFP_I64_PPCF128;
479 } else if (OpVT == MVT::i128) {
480 if (RetVT == MVT::f32)
481 return UINTTOFP_I128_F32;
482 else if (RetVT == MVT::f64)
483 return UINTTOFP_I128_F64;
484 else if (RetVT == MVT::f80)
485 return UINTTOFP_I128_F80;
486 else if (RetVT == MVT::ppcf128)
487 return UINTTOFP_I128_PPCF128;
489 return UNKNOWN_LIBCALL;
492 /// InitCmpLibcallCCs - Set default comparison libcall CC.
494 static void InitCmpLibcallCCs(ISD::CondCode *CCs) {
495 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL);
496 CCs[RTLIB::OEQ_F32] = ISD::SETEQ;
497 CCs[RTLIB::OEQ_F64] = ISD::SETEQ;
498 CCs[RTLIB::UNE_F32] = ISD::SETNE;
499 CCs[RTLIB::UNE_F64] = ISD::SETNE;
500 CCs[RTLIB::OGE_F32] = ISD::SETGE;
501 CCs[RTLIB::OGE_F64] = ISD::SETGE;
502 CCs[RTLIB::OLT_F32] = ISD::SETLT;
503 CCs[RTLIB::OLT_F64] = ISD::SETLT;
504 CCs[RTLIB::OLE_F32] = ISD::SETLE;
505 CCs[RTLIB::OLE_F64] = ISD::SETLE;
506 CCs[RTLIB::OGT_F32] = ISD::SETGT;
507 CCs[RTLIB::OGT_F64] = ISD::SETGT;
508 CCs[RTLIB::UO_F32] = ISD::SETNE;
509 CCs[RTLIB::UO_F64] = ISD::SETNE;
510 CCs[RTLIB::O_F32] = ISD::SETEQ;
511 CCs[RTLIB::O_F64] = ISD::SETEQ;
514 /// NOTE: The constructor takes ownership of TLOF.
515 TargetLowering::TargetLowering(const TargetMachine &tm,
516 const TargetLoweringObjectFile *tlof)
517 : TM(tm), TD(TM.getTargetData()), TLOF(*tlof) {
518 // All operations default to being supported.
519 memset(OpActions, 0, sizeof(OpActions));
520 memset(LoadExtActions, 0, sizeof(LoadExtActions));
521 memset(TruncStoreActions, 0, sizeof(TruncStoreActions));
522 memset(IndexedModeActions, 0, sizeof(IndexedModeActions));
523 memset(CondCodeActions, 0, sizeof(CondCodeActions));
525 // Set default actions for various operations.
526 for (unsigned VT = 0; VT != (unsigned)MVT::LAST_VALUETYPE; ++VT) {
527 // Default all indexed load / store to expand.
528 for (unsigned IM = (unsigned)ISD::PRE_INC;
529 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
530 setIndexedLoadAction(IM, (MVT::SimpleValueType)VT, Expand);
531 setIndexedStoreAction(IM, (MVT::SimpleValueType)VT, Expand);
534 // These operations default to expand.
535 setOperationAction(ISD::FGETSIGN, (MVT::SimpleValueType)VT, Expand);
536 setOperationAction(ISD::CONCAT_VECTORS, (MVT::SimpleValueType)VT, Expand);
539 // Most targets ignore the @llvm.prefetch intrinsic.
540 setOperationAction(ISD::PREFETCH, MVT::Other, Expand);
542 // ConstantFP nodes default to expand. Targets can either change this to
543 // Legal, in which case all fp constants are legal, or use isFPImmLegal()
544 // to optimize expansions for certain constants.
545 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
546 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
547 setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
549 // These library functions default to expand.
550 setOperationAction(ISD::FLOG , MVT::f64, Expand);
551 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
552 setOperationAction(ISD::FLOG10,MVT::f64, Expand);
553 setOperationAction(ISD::FEXP , MVT::f64, Expand);
554 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
555 setOperationAction(ISD::FLOG , MVT::f32, Expand);
556 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
557 setOperationAction(ISD::FLOG10,MVT::f32, Expand);
558 setOperationAction(ISD::FEXP , MVT::f32, Expand);
559 setOperationAction(ISD::FEXP2, MVT::f32, Expand);
561 // Default ISD::TRAP to expand (which turns it into abort).
562 setOperationAction(ISD::TRAP, MVT::Other, Expand);
564 IsLittleEndian = TD->isLittleEndian();
565 ShiftAmountTy = PointerTy = MVT::getIntegerVT(8*TD->getPointerSize());
566 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE*sizeof(TargetRegisterClass*));
567 memset(TargetDAGCombineArray, 0, array_lengthof(TargetDAGCombineArray));
568 maxStoresPerMemset = maxStoresPerMemcpy = maxStoresPerMemmove = 8;
569 benefitFromCodePlacementOpt = false;
570 UseUnderscoreSetJmp = false;
571 UseUnderscoreLongJmp = false;
572 SelectIsExpensive = false;
573 IntDivIsCheap = false;
574 Pow2DivIsCheap = false;
575 StackPointerRegisterToSaveRestore = 0;
576 ExceptionPointerRegister = 0;
577 ExceptionSelectorRegister = 0;
578 BooleanContents = UndefinedBooleanContent;
579 SchedPreferenceInfo = Sched::Latency;
581 JumpBufAlignment = 0;
582 PrefLoopAlignment = 0;
583 MinStackArgumentAlignment = 1;
584 ShouldFoldAtomicFences = false;
586 InitLibcallNames(LibcallRoutineNames);
587 InitCmpLibcallCCs(CmpLibcallCCs);
588 InitLibcallCallingConvs(LibcallCallingConvs);
591 TargetLowering::~TargetLowering() {
595 /// canOpTrap - Returns true if the operation can trap for the value type.
596 /// VT must be a legal type.
597 bool TargetLowering::canOpTrap(unsigned Op, EVT VT) const {
598 assert(isTypeLegal(VT));
613 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT,
614 unsigned &NumIntermediates,
616 TargetLowering *TLI) {
617 // Figure out the right, legal destination reg to copy into.
618 unsigned NumElts = VT.getVectorNumElements();
619 MVT EltTy = VT.getVectorElementType();
621 unsigned NumVectorRegs = 1;
623 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
624 // could break down into LHS/RHS like LegalizeDAG does.
625 if (!isPowerOf2_32(NumElts)) {
626 NumVectorRegs = NumElts;
630 // Divide the input until we get to a supported size. This will always
631 // end with a scalar if the target doesn't support vectors.
632 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) {
637 NumIntermediates = NumVectorRegs;
639 MVT NewVT = MVT::getVectorVT(EltTy, NumElts);
640 if (!TLI->isTypeLegal(NewVT))
642 IntermediateVT = NewVT;
644 EVT DestVT = TLI->getRegisterType(NewVT);
646 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16.
647 return NumVectorRegs*(NewVT.getSizeInBits()/DestVT.getSizeInBits());
649 // Otherwise, promotion or legal types use the same number of registers as
650 // the vector decimated to the appropriate level.
651 return NumVectorRegs;
654 /// isLegalRC - Return true if the value types that can be represented by the
655 /// specified register class are all legal.
656 bool TargetLowering::isLegalRC(const TargetRegisterClass *RC) const {
657 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
665 /// hasLegalSuperRegRegClasses - Return true if the specified register class
666 /// has one or more super-reg register classes that are legal.
667 bool TargetLowering::hasLegalSuperRegRegClasses(const TargetRegisterClass *RC) {
668 if (*RC->superregclasses_begin() == 0)
670 for (TargetRegisterInfo::regclass_iterator I = RC->superregclasses_begin(),
671 E = RC->superregclasses_end(); I != E; ++I) {
672 const TargetRegisterClass *RRC = *I;
679 /// findRepresentativeClass - Return the largest legal super-reg register class
680 /// of the specified register class.
681 const TargetRegisterClass *
682 TargetLowering::findRepresentativeClass(const TargetRegisterClass *RC) {
685 const TargetRegisterClass *BestRC = RC;
686 for (TargetRegisterInfo::regclass_iterator I = RC->superregclasses_begin(),
687 E = RC->superregclasses_end(); I != E; ++I) {
688 const TargetRegisterClass *RRC = *I;
689 if (RRC->isASubClass() || !isLegalRC(RRC))
691 if (!hasLegalSuperRegRegClasses(RRC))
698 /// computeRegisterProperties - Once all of the register classes are added,
699 /// this allows us to compute derived properties we expose.
700 void TargetLowering::computeRegisterProperties() {
701 assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE &&
702 "Too many value types for ValueTypeActions to hold!");
704 // Everything defaults to needing one register.
705 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
706 NumRegistersForVT[i] = 1;
707 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i;
709 // ...except isVoid, which doesn't need any registers.
710 NumRegistersForVT[MVT::isVoid] = 0;
712 // Find the largest integer register class.
713 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE;
714 for (; RegClassForVT[LargestIntReg] == 0; --LargestIntReg)
715 assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
717 // Every integer value type larger than this largest register takes twice as
718 // many registers to represent as the previous ValueType.
719 for (unsigned ExpandedReg = LargestIntReg + 1; ; ++ExpandedReg) {
720 EVT ExpandedVT = (MVT::SimpleValueType)ExpandedReg;
721 if (!ExpandedVT.isInteger())
723 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
724 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg;
725 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1);
726 ValueTypeActions.setTypeAction(ExpandedVT, Expand);
729 // Inspect all of the ValueType's smaller than the largest integer
730 // register to see which ones need promotion.
731 unsigned LegalIntReg = LargestIntReg;
732 for (unsigned IntReg = LargestIntReg - 1;
733 IntReg >= (unsigned)MVT::i1; --IntReg) {
734 EVT IVT = (MVT::SimpleValueType)IntReg;
735 if (isTypeLegal(IVT)) {
736 LegalIntReg = IntReg;
738 RegisterTypeForVT[IntReg] = TransformToType[IntReg] =
739 (MVT::SimpleValueType)LegalIntReg;
740 ValueTypeActions.setTypeAction(IVT, Promote);
744 // ppcf128 type is really two f64's.
745 if (!isTypeLegal(MVT::ppcf128)) {
746 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64];
747 RegisterTypeForVT[MVT::ppcf128] = MVT::f64;
748 TransformToType[MVT::ppcf128] = MVT::f64;
749 ValueTypeActions.setTypeAction(MVT::ppcf128, Expand);
752 // Decide how to handle f64. If the target does not have native f64 support,
753 // expand it to i64 and we will be generating soft float library calls.
754 if (!isTypeLegal(MVT::f64)) {
755 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
756 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
757 TransformToType[MVT::f64] = MVT::i64;
758 ValueTypeActions.setTypeAction(MVT::f64, Expand);
761 // Decide how to handle f32. If the target does not have native support for
762 // f32, promote it to f64 if it is legal. Otherwise, expand it to i32.
763 if (!isTypeLegal(MVT::f32)) {
764 if (isTypeLegal(MVT::f64)) {
765 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::f64];
766 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::f64];
767 TransformToType[MVT::f32] = MVT::f64;
768 ValueTypeActions.setTypeAction(MVT::f32, Promote);
770 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
771 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
772 TransformToType[MVT::f32] = MVT::i32;
773 ValueTypeActions.setTypeAction(MVT::f32, Expand);
777 // Loop over all of the vector value types to see which need transformations.
778 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
779 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
780 MVT VT = (MVT::SimpleValueType)i;
781 if (isTypeLegal(VT)) continue;
785 unsigned NumIntermediates;
786 NumRegistersForVT[i] =
787 getVectorTypeBreakdownMVT(VT, IntermediateVT, NumIntermediates,
789 RegisterTypeForVT[i] = RegisterVT;
791 // Determine if there is a legal wider type.
792 bool IsLegalWiderType = false;
793 EVT EltVT = VT.getVectorElementType();
794 unsigned NElts = VT.getVectorNumElements();
795 for (unsigned nVT = i+1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
796 EVT SVT = (MVT::SimpleValueType)nVT;
797 if (isTypeSynthesizable(SVT) && SVT.getVectorElementType() == EltVT &&
798 SVT.getVectorNumElements() > NElts && NElts != 1) {
799 TransformToType[i] = SVT;
800 ValueTypeActions.setTypeAction(VT, Promote);
801 IsLegalWiderType = true;
805 if (!IsLegalWiderType) {
806 EVT NVT = VT.getPow2VectorType();
808 // Type is already a power of 2. The default action is to split.
809 TransformToType[i] = MVT::Other;
810 ValueTypeActions.setTypeAction(VT, Expand);
812 TransformToType[i] = NVT;
813 ValueTypeActions.setTypeAction(VT, Promote);
818 // Determine the 'representative' register class for each value type.
819 // An representative register class is the largest (meaning one which is
820 // not a sub-register class / subreg register class) legal register class for
821 // a group of value types. For example, on i386, i8, i16, and i32
822 // representative would be GR32; while on x86_64 it's GR64.
823 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i)
824 RepRegClassForVT[i] = findRepresentativeClass(RegClassForVT[i]);
827 const char *TargetLowering::getTargetNodeName(unsigned Opcode) const {
832 MVT::SimpleValueType TargetLowering::getSetCCResultType(EVT VT) const {
833 return PointerTy.SimpleTy;
836 MVT::SimpleValueType TargetLowering::getCmpLibcallReturnType() const {
837 return MVT::i32; // return the default value
840 /// getVectorTypeBreakdown - Vector types are broken down into some number of
841 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
842 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
843 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
845 /// This method returns the number of registers needed, and the VT for each
846 /// register. It also returns the VT and quantity of the intermediate values
847 /// before they are promoted/expanded.
849 unsigned TargetLowering::getVectorTypeBreakdown(LLVMContext &Context, EVT VT,
851 unsigned &NumIntermediates,
852 EVT &RegisterVT) const {
853 // Figure out the right, legal destination reg to copy into.
854 unsigned NumElts = VT.getVectorNumElements();
855 EVT EltTy = VT.getVectorElementType();
857 unsigned NumVectorRegs = 1;
859 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
860 // could break down into LHS/RHS like LegalizeDAG does.
861 if (!isPowerOf2_32(NumElts)) {
862 NumVectorRegs = NumElts;
866 // Divide the input until we get to a supported size. This will always
867 // end with a scalar if the target doesn't support vectors.
868 while (NumElts > 1 && !isTypeLegal(
869 EVT::getVectorVT(Context, EltTy, NumElts))) {
874 NumIntermediates = NumVectorRegs;
876 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts);
877 if (!isTypeLegal(NewVT))
879 IntermediateVT = NewVT;
881 EVT DestVT = getRegisterType(Context, NewVT);
883 if (DestVT.bitsLT(NewVT)) {
884 // Value is expanded, e.g. i64 -> i16.
885 return NumVectorRegs*(NewVT.getSizeInBits()/DestVT.getSizeInBits());
887 // Otherwise, promotion or legal types use the same number of registers as
888 // the vector decimated to the appropriate level.
889 return NumVectorRegs;
895 /// Get the EVTs and ArgFlags collections that represent the legalized return
896 /// type of the given function. This does not require a DAG or a return value,
897 /// and is suitable for use before any DAGs for the function are constructed.
898 /// TODO: Move this out of TargetLowering.cpp.
899 void llvm::GetReturnInfo(const Type* ReturnType, Attributes attr,
900 SmallVectorImpl<ISD::OutputArg> &Outs,
901 const TargetLowering &TLI,
902 SmallVectorImpl<uint64_t> *Offsets) {
903 SmallVector<EVT, 4> ValueVTs;
904 ComputeValueVTs(TLI, ReturnType, ValueVTs);
905 unsigned NumValues = ValueVTs.size();
906 if (NumValues == 0) return;
909 for (unsigned j = 0, f = NumValues; j != f; ++j) {
910 EVT VT = ValueVTs[j];
911 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
913 if (attr & Attribute::SExt)
914 ExtendKind = ISD::SIGN_EXTEND;
915 else if (attr & Attribute::ZExt)
916 ExtendKind = ISD::ZERO_EXTEND;
918 // FIXME: C calling convention requires the return type to be promoted to
919 // at least 32-bit. But this is not necessary for non-C calling
920 // conventions. The frontend should mark functions whose return values
921 // require promoting with signext or zeroext attributes.
922 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) {
923 EVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32);
924 if (VT.bitsLT(MinVT))
928 unsigned NumParts = TLI.getNumRegisters(ReturnType->getContext(), VT);
929 EVT PartVT = TLI.getRegisterType(ReturnType->getContext(), VT);
930 unsigned PartSize = TLI.getTargetData()->getTypeAllocSize(
931 PartVT.getTypeForEVT(ReturnType->getContext()));
933 // 'inreg' on function refers to return value
934 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
935 if (attr & Attribute::InReg)
938 // Propagate extension type if any
939 if (attr & Attribute::SExt)
941 else if (attr & Attribute::ZExt)
944 for (unsigned i = 0; i < NumParts; ++i) {
945 Outs.push_back(ISD::OutputArg(Flags, PartVT, /*isFixed=*/true));
947 Offsets->push_back(Offset);
954 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
955 /// function arguments in the caller parameter area. This is the actual
956 /// alignment, not its logarithm.
957 unsigned TargetLowering::getByValTypeAlignment(const Type *Ty) const {
958 return TD->getCallFrameTypeAlignment(Ty);
961 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
962 /// current function. The returned value is a member of the
963 /// MachineJumpTableInfo::JTEntryKind enum.
964 unsigned TargetLowering::getJumpTableEncoding() const {
965 // In non-pic modes, just use the address of a block.
966 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
967 return MachineJumpTableInfo::EK_BlockAddress;
969 // In PIC mode, if the target supports a GPRel32 directive, use it.
970 if (getTargetMachine().getMCAsmInfo()->getGPRel32Directive() != 0)
971 return MachineJumpTableInfo::EK_GPRel32BlockAddress;
973 // Otherwise, use a label difference.
974 return MachineJumpTableInfo::EK_LabelDifference32;
977 SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table,
978 SelectionDAG &DAG) const {
979 // If our PIC model is GP relative, use the global offset table as the base.
980 if (getJumpTableEncoding() == MachineJumpTableInfo::EK_GPRel32BlockAddress)
981 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy());
985 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
986 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
989 TargetLowering::getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
990 unsigned JTI,MCContext &Ctx) const{
991 // The normal PIC reloc base is the label at the start of the jump table.
992 return MCSymbolRefExpr::Create(MF->getJTISymbol(JTI, Ctx), Ctx);
996 TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
997 // Assume that everything is safe in static mode.
998 if (getTargetMachine().getRelocationModel() == Reloc::Static)
1001 // In dynamic-no-pic mode, assume that known defined values are safe.
1002 if (getTargetMachine().getRelocationModel() == Reloc::DynamicNoPIC &&
1004 !GA->getGlobal()->isDeclaration() &&
1005 !GA->getGlobal()->isWeakForLinker())
1008 // Otherwise assume nothing is safe.
1012 //===----------------------------------------------------------------------===//
1013 // Optimization Methods
1014 //===----------------------------------------------------------------------===//
1016 /// ShrinkDemandedConstant - Check to see if the specified operand of the
1017 /// specified instruction is a constant integer. If so, check to see if there
1018 /// are any bits set in the constant that are not demanded. If so, shrink the
1019 /// constant and return true.
1020 bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDValue Op,
1021 const APInt &Demanded) {
1022 DebugLoc dl = Op.getDebugLoc();
1024 // FIXME: ISD::SELECT, ISD::SELECT_CC
1025 switch (Op.getOpcode()) {
1030 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
1031 if (!C) return false;
1033 if (Op.getOpcode() == ISD::XOR &&
1034 (C->getAPIntValue() | (~Demanded)).isAllOnesValue())
1037 // if we can expand it to have all bits set, do it
1038 if (C->getAPIntValue().intersects(~Demanded)) {
1039 EVT VT = Op.getValueType();
1040 SDValue New = DAG.getNode(Op.getOpcode(), dl, VT, Op.getOperand(0),
1041 DAG.getConstant(Demanded &
1044 return CombineTo(Op, New);
1054 /// ShrinkDemandedOp - Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the
1055 /// casts are free. This uses isZExtFree and ZERO_EXTEND for the widening
1056 /// cast, but it could be generalized for targets with other types of
1057 /// implicit widening casts.
1059 TargetLowering::TargetLoweringOpt::ShrinkDemandedOp(SDValue Op,
1061 const APInt &Demanded,
1063 assert(Op.getNumOperands() == 2 &&
1064 "ShrinkDemandedOp only supports binary operators!");
1065 assert(Op.getNode()->getNumValues() == 1 &&
1066 "ShrinkDemandedOp only supports nodes with one result!");
1068 // Don't do this if the node has another user, which may require the
1070 if (!Op.getNode()->hasOneUse())
1073 // Search for the smallest integer type with free casts to and from
1074 // Op's type. For expedience, just check power-of-2 integer types.
1075 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1076 unsigned SmallVTBits = BitWidth - Demanded.countLeadingZeros();
1077 if (!isPowerOf2_32(SmallVTBits))
1078 SmallVTBits = NextPowerOf2(SmallVTBits);
1079 for (; SmallVTBits < BitWidth; SmallVTBits = NextPowerOf2(SmallVTBits)) {
1080 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), SmallVTBits);
1081 if (TLI.isTruncateFree(Op.getValueType(), SmallVT) &&
1082 TLI.isZExtFree(SmallVT, Op.getValueType())) {
1083 // We found a type with free casts.
1084 SDValue X = DAG.getNode(Op.getOpcode(), dl, SmallVT,
1085 DAG.getNode(ISD::TRUNCATE, dl, SmallVT,
1086 Op.getNode()->getOperand(0)),
1087 DAG.getNode(ISD::TRUNCATE, dl, SmallVT,
1088 Op.getNode()->getOperand(1)));
1089 SDValue Z = DAG.getNode(ISD::ZERO_EXTEND, dl, Op.getValueType(), X);
1090 return CombineTo(Op, Z);
1096 /// SimplifyDemandedBits - Look at Op. At this point, we know that only the
1097 /// DemandedMask bits of the result of Op are ever used downstream. If we can
1098 /// use this information to simplify Op, create a new simplified DAG node and
1099 /// return true, returning the original and new nodes in Old and New. Otherwise,
1100 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
1101 /// the expression (used to simplify the caller). The KnownZero/One bits may
1102 /// only be accurate for those bits in the DemandedMask.
1103 bool TargetLowering::SimplifyDemandedBits(SDValue Op,
1104 const APInt &DemandedMask,
1107 TargetLoweringOpt &TLO,
1108 unsigned Depth) const {
1109 unsigned BitWidth = DemandedMask.getBitWidth();
1110 assert(Op.getValueType().getScalarType().getSizeInBits() == BitWidth &&
1111 "Mask size mismatches value type size!");
1112 APInt NewMask = DemandedMask;
1113 DebugLoc dl = Op.getDebugLoc();
1115 // Don't know anything.
1116 KnownZero = KnownOne = APInt(BitWidth, 0);
1118 // Other users may use these bits.
1119 if (!Op.getNode()->hasOneUse()) {
1121 // If not at the root, Just compute the KnownZero/KnownOne bits to
1122 // simplify things downstream.
1123 TLO.DAG.ComputeMaskedBits(Op, DemandedMask, KnownZero, KnownOne, Depth);
1126 // If this is the root being simplified, allow it to have multiple uses,
1127 // just set the NewMask to all bits.
1128 NewMask = APInt::getAllOnesValue(BitWidth);
1129 } else if (DemandedMask == 0) {
1130 // Not demanding any bits from Op.
1131 if (Op.getOpcode() != ISD::UNDEF)
1132 return TLO.CombineTo(Op, TLO.DAG.getUNDEF(Op.getValueType()));
1134 } else if (Depth == 6) { // Limit search depth.
1138 APInt KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut;
1139 switch (Op.getOpcode()) {
1141 // We know all of the bits for a constant!
1142 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue() & NewMask;
1143 KnownZero = ~KnownOne & NewMask;
1144 return false; // Don't fall through, will infinitely loop.
1146 // If the RHS is a constant, check to see if the LHS would be zero without
1147 // using the bits from the RHS. Below, we use knowledge about the RHS to
1148 // simplify the LHS, here we're using information from the LHS to simplify
1150 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1151 APInt LHSZero, LHSOne;
1152 TLO.DAG.ComputeMaskedBits(Op.getOperand(0), NewMask,
1153 LHSZero, LHSOne, Depth+1);
1154 // If the LHS already has zeros where RHSC does, this and is dead.
1155 if ((LHSZero & NewMask) == (~RHSC->getAPIntValue() & NewMask))
1156 return TLO.CombineTo(Op, Op.getOperand(0));
1157 // If any of the set bits in the RHS are known zero on the LHS, shrink
1159 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & NewMask))
1163 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
1164 KnownOne, TLO, Depth+1))
1166 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1167 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownZero & NewMask,
1168 KnownZero2, KnownOne2, TLO, Depth+1))
1170 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1172 // If all of the demanded bits are known one on one side, return the other.
1173 // These bits cannot contribute to the result of the 'and'.
1174 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
1175 return TLO.CombineTo(Op, Op.getOperand(0));
1176 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
1177 return TLO.CombineTo(Op, Op.getOperand(1));
1178 // If all of the demanded bits in the inputs are known zeros, return zero.
1179 if ((NewMask & (KnownZero|KnownZero2)) == NewMask)
1180 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, Op.getValueType()));
1181 // If the RHS is a constant, see if we can simplify it.
1182 if (TLO.ShrinkDemandedConstant(Op, ~KnownZero2 & NewMask))
1184 // If the operation can be done in a smaller type, do so.
1185 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1188 // Output known-1 bits are only known if set in both the LHS & RHS.
1189 KnownOne &= KnownOne2;
1190 // Output known-0 are known to be clear if zero in either the LHS | RHS.
1191 KnownZero |= KnownZero2;
1194 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
1195 KnownOne, TLO, Depth+1))
1197 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1198 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownOne & NewMask,
1199 KnownZero2, KnownOne2, TLO, Depth+1))
1201 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1203 // If all of the demanded bits are known zero on one side, return the other.
1204 // These bits cannot contribute to the result of the 'or'.
1205 if ((NewMask & ~KnownOne2 & KnownZero) == (~KnownOne2 & NewMask))
1206 return TLO.CombineTo(Op, Op.getOperand(0));
1207 if ((NewMask & ~KnownOne & KnownZero2) == (~KnownOne & NewMask))
1208 return TLO.CombineTo(Op, Op.getOperand(1));
1209 // If all of the potentially set bits on one side are known to be set on
1210 // the other side, just use the 'other' side.
1211 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
1212 return TLO.CombineTo(Op, Op.getOperand(0));
1213 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
1214 return TLO.CombineTo(Op, Op.getOperand(1));
1215 // If the RHS is a constant, see if we can simplify it.
1216 if (TLO.ShrinkDemandedConstant(Op, NewMask))
1218 // If the operation can be done in a smaller type, do so.
1219 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1222 // Output known-0 bits are only known if clear in both the LHS & RHS.
1223 KnownZero &= KnownZero2;
1224 // Output known-1 are known to be set if set in either the LHS | RHS.
1225 KnownOne |= KnownOne2;
1228 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
1229 KnownOne, TLO, Depth+1))
1231 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1232 if (SimplifyDemandedBits(Op.getOperand(0), NewMask, KnownZero2,
1233 KnownOne2, TLO, Depth+1))
1235 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1237 // If all of the demanded bits are known zero on one side, return the other.
1238 // These bits cannot contribute to the result of the 'xor'.
1239 if ((KnownZero & NewMask) == NewMask)
1240 return TLO.CombineTo(Op, Op.getOperand(0));
1241 if ((KnownZero2 & NewMask) == NewMask)
1242 return TLO.CombineTo(Op, Op.getOperand(1));
1243 // If the operation can be done in a smaller type, do so.
1244 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1247 // If all of the unknown bits are known to be zero on one side or the other
1248 // (but not both) turn this into an *inclusive* or.
1249 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1250 if ((NewMask & ~KnownZero & ~KnownZero2) == 0)
1251 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, dl, Op.getValueType(),
1255 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1256 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
1257 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1258 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
1260 // If all of the demanded bits on one side are known, and all of the set
1261 // bits on that side are also known to be set on the other side, turn this
1262 // into an AND, as we know the bits will be cleared.
1263 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1264 if ((NewMask & (KnownZero|KnownOne)) == NewMask) { // all known
1265 if ((KnownOne & KnownOne2) == KnownOne) {
1266 EVT VT = Op.getValueType();
1267 SDValue ANDC = TLO.DAG.getConstant(~KnownOne & NewMask, VT);
1268 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, dl, VT,
1269 Op.getOperand(0), ANDC));
1273 // If the RHS is a constant, see if we can simplify it.
1274 // for XOR, we prefer to force bits to 1 if they will make a -1.
1275 // if we can't force bits, try to shrink constant
1276 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1277 APInt Expanded = C->getAPIntValue() | (~NewMask);
1278 // if we can expand it to have all bits set, do it
1279 if (Expanded.isAllOnesValue()) {
1280 if (Expanded != C->getAPIntValue()) {
1281 EVT VT = Op.getValueType();
1282 SDValue New = TLO.DAG.getNode(Op.getOpcode(), dl,VT, Op.getOperand(0),
1283 TLO.DAG.getConstant(Expanded, VT));
1284 return TLO.CombineTo(Op, New);
1286 // if it already has all the bits set, nothing to change
1287 // but don't shrink either!
1288 } else if (TLO.ShrinkDemandedConstant(Op, NewMask)) {
1293 KnownZero = KnownZeroOut;
1294 KnownOne = KnownOneOut;
1297 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero,
1298 KnownOne, TLO, Depth+1))
1300 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero2,
1301 KnownOne2, TLO, Depth+1))
1303 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1304 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1306 // If the operands are constants, see if we can simplify them.
1307 if (TLO.ShrinkDemandedConstant(Op, NewMask))
1310 // Only known if known in both the LHS and RHS.
1311 KnownOne &= KnownOne2;
1312 KnownZero &= KnownZero2;
1314 case ISD::SELECT_CC:
1315 if (SimplifyDemandedBits(Op.getOperand(3), NewMask, KnownZero,
1316 KnownOne, TLO, Depth+1))
1318 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero2,
1319 KnownOne2, TLO, Depth+1))
1321 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1322 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1324 // If the operands are constants, see if we can simplify them.
1325 if (TLO.ShrinkDemandedConstant(Op, NewMask))
1328 // Only known if known in both the LHS and RHS.
1329 KnownOne &= KnownOne2;
1330 KnownZero &= KnownZero2;
1333 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1334 unsigned ShAmt = SA->getZExtValue();
1335 SDValue InOp = Op.getOperand(0);
1337 // If the shift count is an invalid immediate, don't do anything.
1338 if (ShAmt >= BitWidth)
1341 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a
1342 // single shift. We can do this if the bottom bits (which are shifted
1343 // out) are never demanded.
1344 if (InOp.getOpcode() == ISD::SRL &&
1345 isa<ConstantSDNode>(InOp.getOperand(1))) {
1346 if (ShAmt && (NewMask & APInt::getLowBitsSet(BitWidth, ShAmt)) == 0) {
1347 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
1348 unsigned Opc = ISD::SHL;
1349 int Diff = ShAmt-C1;
1356 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
1357 EVT VT = Op.getValueType();
1358 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT,
1359 InOp.getOperand(0), NewSA));
1363 if (SimplifyDemandedBits(Op.getOperand(0), NewMask.lshr(ShAmt),
1364 KnownZero, KnownOne, TLO, Depth+1))
1366 KnownZero <<= SA->getZExtValue();
1367 KnownOne <<= SA->getZExtValue();
1368 // low bits known zero.
1369 KnownZero |= APInt::getLowBitsSet(BitWidth, SA->getZExtValue());
1373 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1374 EVT VT = Op.getValueType();
1375 unsigned ShAmt = SA->getZExtValue();
1376 unsigned VTSize = VT.getSizeInBits();
1377 SDValue InOp = Op.getOperand(0);
1379 // If the shift count is an invalid immediate, don't do anything.
1380 if (ShAmt >= BitWidth)
1383 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a
1384 // single shift. We can do this if the top bits (which are shifted out)
1385 // are never demanded.
1386 if (InOp.getOpcode() == ISD::SHL &&
1387 isa<ConstantSDNode>(InOp.getOperand(1))) {
1388 if (ShAmt && (NewMask & APInt::getHighBitsSet(VTSize, ShAmt)) == 0) {
1389 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
1390 unsigned Opc = ISD::SRL;
1391 int Diff = ShAmt-C1;
1398 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
1399 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT,
1400 InOp.getOperand(0), NewSA));
1404 // Compute the new bits that are at the top now.
1405 if (SimplifyDemandedBits(InOp, (NewMask << ShAmt),
1406 KnownZero, KnownOne, TLO, Depth+1))
1408 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1409 KnownZero = KnownZero.lshr(ShAmt);
1410 KnownOne = KnownOne.lshr(ShAmt);
1412 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
1413 KnownZero |= HighBits; // High bits known zero.
1417 // If this is an arithmetic shift right and only the low-bit is set, we can
1418 // always convert this into a logical shr, even if the shift amount is
1419 // variable. The low bit of the shift cannot be an input sign bit unless
1420 // the shift amount is >= the size of the datatype, which is undefined.
1421 if (DemandedMask == 1)
1422 return TLO.CombineTo(Op,
1423 TLO.DAG.getNode(ISD::SRL, dl, Op.getValueType(),
1424 Op.getOperand(0), Op.getOperand(1)));
1426 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1427 EVT VT = Op.getValueType();
1428 unsigned ShAmt = SA->getZExtValue();
1430 // If the shift count is an invalid immediate, don't do anything.
1431 if (ShAmt >= BitWidth)
1434 APInt InDemandedMask = (NewMask << ShAmt);
1436 // If any of the demanded bits are produced by the sign extension, we also
1437 // demand the input sign bit.
1438 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
1439 if (HighBits.intersects(NewMask))
1440 InDemandedMask |= APInt::getSignBit(VT.getScalarType().getSizeInBits());
1442 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask,
1443 KnownZero, KnownOne, TLO, Depth+1))
1445 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1446 KnownZero = KnownZero.lshr(ShAmt);
1447 KnownOne = KnownOne.lshr(ShAmt);
1449 // Handle the sign bit, adjusted to where it is now in the mask.
1450 APInt SignBit = APInt::getSignBit(BitWidth).lshr(ShAmt);
1452 // If the input sign bit is known to be zero, or if none of the top bits
1453 // are demanded, turn this into an unsigned shift right.
1454 if (KnownZero.intersects(SignBit) || (HighBits & ~NewMask) == HighBits) {
1455 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, VT,
1458 } else if (KnownOne.intersects(SignBit)) { // New bits are known one.
1459 KnownOne |= HighBits;
1463 case ISD::SIGN_EXTEND_INREG: {
1464 EVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1466 // Sign extension. Compute the demanded bits in the result that are not
1467 // present in the input.
1469 APInt::getHighBitsSet(BitWidth,
1470 BitWidth - EVT.getScalarType().getSizeInBits()) &
1473 // If none of the extended bits are demanded, eliminate the sextinreg.
1475 return TLO.CombineTo(Op, Op.getOperand(0));
1477 APInt InSignBit = APInt::getSignBit(EVT.getScalarType().getSizeInBits());
1478 InSignBit.zext(BitWidth);
1479 APInt InputDemandedBits =
1480 APInt::getLowBitsSet(BitWidth,
1481 EVT.getScalarType().getSizeInBits()) &
1484 // Since the sign extended bits are demanded, we know that the sign
1486 InputDemandedBits |= InSignBit;
1488 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits,
1489 KnownZero, KnownOne, TLO, Depth+1))
1491 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1493 // If the sign bit of the input is known set or clear, then we know the
1494 // top bits of the result.
1496 // If the input sign bit is known zero, convert this into a zero extension.
1497 if (KnownZero.intersects(InSignBit))
1498 return TLO.CombineTo(Op,
1499 TLO.DAG.getZeroExtendInReg(Op.getOperand(0),dl,EVT));
1501 if (KnownOne.intersects(InSignBit)) { // Input sign bit known set
1502 KnownOne |= NewBits;
1503 KnownZero &= ~NewBits;
1504 } else { // Input sign bit unknown
1505 KnownZero &= ~NewBits;
1506 KnownOne &= ~NewBits;
1510 case ISD::ZERO_EXTEND: {
1511 unsigned OperandBitWidth =
1512 Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
1513 APInt InMask = NewMask;
1514 InMask.trunc(OperandBitWidth);
1516 // If none of the top bits are demanded, convert this into an any_extend.
1518 APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask;
1519 if (!NewBits.intersects(NewMask))
1520 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND, dl,
1524 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1525 KnownZero, KnownOne, TLO, Depth+1))
1527 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1528 KnownZero.zext(BitWidth);
1529 KnownOne.zext(BitWidth);
1530 KnownZero |= NewBits;
1533 case ISD::SIGN_EXTEND: {
1534 EVT InVT = Op.getOperand(0).getValueType();
1535 unsigned InBits = InVT.getScalarType().getSizeInBits();
1536 APInt InMask = APInt::getLowBitsSet(BitWidth, InBits);
1537 APInt InSignBit = APInt::getBitsSet(BitWidth, InBits - 1, InBits);
1538 APInt NewBits = ~InMask & NewMask;
1540 // If none of the top bits are demanded, convert this into an any_extend.
1542 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND, dl,
1546 // Since some of the sign extended bits are demanded, we know that the sign
1548 APInt InDemandedBits = InMask & NewMask;
1549 InDemandedBits |= InSignBit;
1550 InDemandedBits.trunc(InBits);
1552 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero,
1553 KnownOne, TLO, Depth+1))
1555 KnownZero.zext(BitWidth);
1556 KnownOne.zext(BitWidth);
1558 // If the sign bit is known zero, convert this to a zero extend.
1559 if (KnownZero.intersects(InSignBit))
1560 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND, dl,
1564 // If the sign bit is known one, the top bits match.
1565 if (KnownOne.intersects(InSignBit)) {
1566 KnownOne |= NewBits;
1567 KnownZero &= ~NewBits;
1568 } else { // Otherwise, top bits aren't known.
1569 KnownOne &= ~NewBits;
1570 KnownZero &= ~NewBits;
1574 case ISD::ANY_EXTEND: {
1575 unsigned OperandBitWidth =
1576 Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
1577 APInt InMask = NewMask;
1578 InMask.trunc(OperandBitWidth);
1579 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1580 KnownZero, KnownOne, TLO, Depth+1))
1582 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1583 KnownZero.zext(BitWidth);
1584 KnownOne.zext(BitWidth);
1587 case ISD::TRUNCATE: {
1588 // Simplify the input, using demanded bit information, and compute the known
1589 // zero/one bits live out.
1590 unsigned OperandBitWidth =
1591 Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
1592 APInt TruncMask = NewMask;
1593 TruncMask.zext(OperandBitWidth);
1594 if (SimplifyDemandedBits(Op.getOperand(0), TruncMask,
1595 KnownZero, KnownOne, TLO, Depth+1))
1597 KnownZero.trunc(BitWidth);
1598 KnownOne.trunc(BitWidth);
1600 // If the input is only used by this truncate, see if we can shrink it based
1601 // on the known demanded bits.
1602 if (Op.getOperand(0).getNode()->hasOneUse()) {
1603 SDValue In = Op.getOperand(0);
1604 switch (In.getOpcode()) {
1607 // Shrink SRL by a constant if none of the high bits shifted in are
1609 if (TLO.LegalTypes() &&
1610 !isTypeDesirableForOp(ISD::SRL, Op.getValueType()))
1611 // Do not turn (vt1 truncate (vt2 srl)) into (vt1 srl) if vt1 is
1614 ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1));
1617 APInt HighBits = APInt::getHighBitsSet(OperandBitWidth,
1618 OperandBitWidth - BitWidth);
1619 HighBits = HighBits.lshr(ShAmt->getZExtValue());
1620 HighBits.trunc(BitWidth);
1622 if (ShAmt->getZExtValue() < BitWidth && !(HighBits & NewMask)) {
1623 // None of the shifted in bits are needed. Add a truncate of the
1624 // shift input, then shift it.
1625 SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE, dl,
1628 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl,
1637 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1640 case ISD::AssertZext: {
1641 // Demand all the bits of the input that are demanded in the output.
1642 // The low bits are obvious; the high bits are demanded because we're
1643 // asserting that they're zero here.
1644 if (SimplifyDemandedBits(Op.getOperand(0), NewMask,
1645 KnownZero, KnownOne, TLO, Depth+1))
1647 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1649 EVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1650 APInt InMask = APInt::getLowBitsSet(BitWidth,
1651 VT.getSizeInBits());
1652 KnownZero |= ~InMask & NewMask;
1655 case ISD::BIT_CONVERT:
1657 // If this is an FP->Int bitcast and if the sign bit is the only thing that
1658 // is demanded, turn this into a FGETSIGN.
1659 if (NewMask == EVT::getIntegerVTSignBit(Op.getValueType()) &&
1660 MVT::isFloatingPoint(Op.getOperand(0).getValueType()) &&
1661 !MVT::isVector(Op.getOperand(0).getValueType())) {
1662 // Only do this xform if FGETSIGN is valid or if before legalize.
1663 if (!TLO.AfterLegalize ||
1664 isOperationLegal(ISD::FGETSIGN, Op.getValueType())) {
1665 // Make a FGETSIGN + SHL to move the sign bit into the appropriate
1666 // place. We expect the SHL to be eliminated by other optimizations.
1667 SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, Op.getValueType(),
1669 unsigned ShVal = Op.getValueType().getSizeInBits()-1;
1670 SDValue ShAmt = TLO.DAG.getConstant(ShVal, getShiftAmountTy());
1671 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, Op.getValueType(),
1680 // Add, Sub, and Mul don't demand any bits in positions beyond that
1681 // of the highest bit demanded of them.
1682 APInt LoMask = APInt::getLowBitsSet(BitWidth,
1683 BitWidth - NewMask.countLeadingZeros());
1684 if (SimplifyDemandedBits(Op.getOperand(0), LoMask, KnownZero2,
1685 KnownOne2, TLO, Depth+1))
1687 if (SimplifyDemandedBits(Op.getOperand(1), LoMask, KnownZero2,
1688 KnownOne2, TLO, Depth+1))
1690 // See if the operation should be performed at a smaller bit width.
1691 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1696 // Just use ComputeMaskedBits to compute output bits.
1697 TLO.DAG.ComputeMaskedBits(Op, NewMask, KnownZero, KnownOne, Depth);
1701 // If we know the value of all of the demanded bits, return this as a
1703 if ((NewMask & (KnownZero|KnownOne)) == NewMask)
1704 return TLO.CombineTo(Op, TLO.DAG.getConstant(KnownOne, Op.getValueType()));
1709 /// computeMaskedBitsForTargetNode - Determine which of the bits specified
1710 /// in Mask are known to be either zero or one and return them in the
1711 /// KnownZero/KnownOne bitsets.
1712 void TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
1716 const SelectionDAG &DAG,
1717 unsigned Depth) const {
1718 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1719 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1720 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1721 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1722 "Should use MaskedValueIsZero if you don't know whether Op"
1723 " is a target node!");
1724 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
1727 /// ComputeNumSignBitsForTargetNode - This method can be implemented by
1728 /// targets that want to expose additional information about sign bits to the
1730 unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
1731 unsigned Depth) const {
1732 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1733 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1734 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1735 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1736 "Should use ComputeNumSignBits if you don't know whether Op"
1737 " is a target node!");
1741 /// ValueHasExactlyOneBitSet - Test if the given value is known to have exactly
1742 /// one bit set. This differs from ComputeMaskedBits in that it doesn't need to
1743 /// determine which bit is set.
1745 static bool ValueHasExactlyOneBitSet(SDValue Val, const SelectionDAG &DAG) {
1746 // A left-shift of a constant one will have exactly one bit set, because
1747 // shifting the bit off the end is undefined.
1748 if (Val.getOpcode() == ISD::SHL)
1749 if (ConstantSDNode *C =
1750 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0)))
1751 if (C->getAPIntValue() == 1)
1754 // Similarly, a right-shift of a constant sign-bit will have exactly
1756 if (Val.getOpcode() == ISD::SRL)
1757 if (ConstantSDNode *C =
1758 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0)))
1759 if (C->getAPIntValue().isSignBit())
1762 // More could be done here, though the above checks are enough
1763 // to handle some common cases.
1765 // Fall back to ComputeMaskedBits to catch other known cases.
1766 EVT OpVT = Val.getValueType();
1767 unsigned BitWidth = OpVT.getScalarType().getSizeInBits();
1768 APInt Mask = APInt::getAllOnesValue(BitWidth);
1769 APInt KnownZero, KnownOne;
1770 DAG.ComputeMaskedBits(Val, Mask, KnownZero, KnownOne);
1771 return (KnownZero.countPopulation() == BitWidth - 1) &&
1772 (KnownOne.countPopulation() == 1);
1775 /// SimplifySetCC - Try to simplify a setcc built with the specified operands
1776 /// and cc. If it is unable to simplify it, return a null SDValue.
1778 TargetLowering::SimplifySetCC(EVT VT, SDValue N0, SDValue N1,
1779 ISD::CondCode Cond, bool foldBooleans,
1780 DAGCombinerInfo &DCI, DebugLoc dl) const {
1781 SelectionDAG &DAG = DCI.DAG;
1782 LLVMContext &Context = *DAG.getContext();
1784 // These setcc operations always fold.
1788 case ISD::SETFALSE2: return DAG.getConstant(0, VT);
1790 case ISD::SETTRUE2: return DAG.getConstant(1, VT);
1793 if (isa<ConstantSDNode>(N0.getNode())) {
1794 // Ensure that the constant occurs on the RHS, and fold constant
1796 return DAG.getSetCC(dl, VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
1799 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) {
1800 const APInt &C1 = N1C->getAPIntValue();
1802 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an
1803 // equality comparison, then we're just comparing whether X itself is
1805 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) &&
1806 N0.getOperand(0).getOpcode() == ISD::CTLZ &&
1807 N0.getOperand(1).getOpcode() == ISD::Constant) {
1809 = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
1810 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1811 ShAmt == Log2_32(N0.getValueType().getSizeInBits())) {
1812 if ((C1 == 0) == (Cond == ISD::SETEQ)) {
1813 // (srl (ctlz x), 5) == 0 -> X != 0
1814 // (srl (ctlz x), 5) != 1 -> X != 0
1817 // (srl (ctlz x), 5) != 0 -> X == 0
1818 // (srl (ctlz x), 5) == 1 -> X == 0
1821 SDValue Zero = DAG.getConstant(0, N0.getValueType());
1822 return DAG.getSetCC(dl, VT, N0.getOperand(0).getOperand(0),
1827 // If the LHS is '(and load, const)', the RHS is 0,
1828 // the test is for equality or unsigned, and all 1 bits of the const are
1829 // in the same partial word, see if we can shorten the load.
1830 if (DCI.isBeforeLegalize() &&
1831 N0.getOpcode() == ISD::AND && C1 == 0 &&
1832 N0.getNode()->hasOneUse() &&
1833 isa<LoadSDNode>(N0.getOperand(0)) &&
1834 N0.getOperand(0).getNode()->hasOneUse() &&
1835 isa<ConstantSDNode>(N0.getOperand(1))) {
1836 LoadSDNode *Lod = cast<LoadSDNode>(N0.getOperand(0));
1838 unsigned bestWidth = 0, bestOffset = 0;
1839 if (!Lod->isVolatile() && Lod->isUnindexed()) {
1840 unsigned origWidth = N0.getValueType().getSizeInBits();
1841 unsigned maskWidth = origWidth;
1842 // We can narrow (e.g.) 16-bit extending loads on 32-bit target to
1843 // 8 bits, but have to be careful...
1844 if (Lod->getExtensionType() != ISD::NON_EXTLOAD)
1845 origWidth = Lod->getMemoryVT().getSizeInBits();
1847 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
1848 for (unsigned width = origWidth / 2; width>=8; width /= 2) {
1849 APInt newMask = APInt::getLowBitsSet(maskWidth, width);
1850 for (unsigned offset=0; offset<origWidth/width; offset++) {
1851 if ((newMask & Mask) == Mask) {
1852 if (!TD->isLittleEndian())
1853 bestOffset = (origWidth/width - offset - 1) * (width/8);
1855 bestOffset = (uint64_t)offset * (width/8);
1856 bestMask = Mask.lshr(offset * (width/8) * 8);
1860 newMask = newMask << width;
1865 EVT newVT = EVT::getIntegerVT(Context, bestWidth);
1866 if (newVT.isRound()) {
1867 EVT PtrType = Lod->getOperand(1).getValueType();
1868 SDValue Ptr = Lod->getBasePtr();
1869 if (bestOffset != 0)
1870 Ptr = DAG.getNode(ISD::ADD, dl, PtrType, Lod->getBasePtr(),
1871 DAG.getConstant(bestOffset, PtrType));
1872 unsigned NewAlign = MinAlign(Lod->getAlignment(), bestOffset);
1873 SDValue NewLoad = DAG.getLoad(newVT, dl, Lod->getChain(), Ptr,
1875 Lod->getSrcValueOffset() + bestOffset,
1876 false, false, NewAlign);
1877 return DAG.getSetCC(dl, VT,
1878 DAG.getNode(ISD::AND, dl, newVT, NewLoad,
1879 DAG.getConstant(bestMask.trunc(bestWidth),
1881 DAG.getConstant(0LL, newVT), Cond);
1886 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
1887 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
1888 unsigned InSize = N0.getOperand(0).getValueType().getSizeInBits();
1890 // If the comparison constant has bits in the upper part, the
1891 // zero-extended value could never match.
1892 if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(),
1893 C1.getBitWidth() - InSize))) {
1897 case ISD::SETEQ: return DAG.getConstant(0, VT);
1900 case ISD::SETNE: return DAG.getConstant(1, VT);
1903 // True if the sign bit of C1 is set.
1904 return DAG.getConstant(C1.isNegative(), VT);
1907 // True if the sign bit of C1 isn't set.
1908 return DAG.getConstant(C1.isNonNegative(), VT);
1914 // Otherwise, we can perform the comparison with the low bits.
1922 EVT newVT = N0.getOperand(0).getValueType();
1923 if (DCI.isBeforeLegalizeOps() ||
1924 (isOperationLegal(ISD::SETCC, newVT) &&
1925 getCondCodeAction(Cond, newVT)==Legal))
1926 return DAG.getSetCC(dl, VT, N0.getOperand(0),
1927 DAG.getConstant(APInt(C1).trunc(InSize), newVT),
1932 break; // todo, be more careful with signed comparisons
1934 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
1935 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1936 EVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
1937 unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits();
1938 EVT ExtDstTy = N0.getValueType();
1939 unsigned ExtDstTyBits = ExtDstTy.getSizeInBits();
1941 // If the extended part has any inconsistent bits, it cannot ever
1942 // compare equal. In other words, they have to be all ones or all
1945 APInt::getHighBitsSet(ExtDstTyBits, ExtDstTyBits - ExtSrcTyBits);
1946 if ((C1 & ExtBits) != 0 && (C1 & ExtBits) != ExtBits)
1947 return DAG.getConstant(Cond == ISD::SETNE, VT);
1950 EVT Op0Ty = N0.getOperand(0).getValueType();
1951 if (Op0Ty == ExtSrcTy) {
1952 ZextOp = N0.getOperand(0);
1954 APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits);
1955 ZextOp = DAG.getNode(ISD::AND, dl, Op0Ty, N0.getOperand(0),
1956 DAG.getConstant(Imm, Op0Ty));
1958 if (!DCI.isCalledByLegalizer())
1959 DCI.AddToWorklist(ZextOp.getNode());
1960 // Otherwise, make this a use of a zext.
1961 return DAG.getSetCC(dl, VT, ZextOp,
1962 DAG.getConstant(C1 & APInt::getLowBitsSet(
1967 } else if ((N1C->isNullValue() || N1C->getAPIntValue() == 1) &&
1968 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1969 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC
1970 if (N0.getOpcode() == ISD::SETCC &&
1971 isTypeLegal(VT) && VT.bitsLE(N0.getValueType())) {
1972 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getAPIntValue() != 1);
1974 return DAG.getNode(ISD::TRUNCATE, dl, VT, N0);
1975 // Invert the condition.
1976 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
1977 CC = ISD::getSetCCInverse(CC,
1978 N0.getOperand(0).getValueType().isInteger());
1979 return DAG.getSetCC(dl, VT, N0.getOperand(0), N0.getOperand(1), CC);
1982 if ((N0.getOpcode() == ISD::XOR ||
1983 (N0.getOpcode() == ISD::AND &&
1984 N0.getOperand(0).getOpcode() == ISD::XOR &&
1985 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
1986 isa<ConstantSDNode>(N0.getOperand(1)) &&
1987 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue() == 1) {
1988 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We
1989 // can only do this if the top bits are known zero.
1990 unsigned BitWidth = N0.getValueSizeInBits();
1991 if (DAG.MaskedValueIsZero(N0,
1992 APInt::getHighBitsSet(BitWidth,
1994 // Okay, get the un-inverted input value.
1996 if (N0.getOpcode() == ISD::XOR)
1997 Val = N0.getOperand(0);
1999 assert(N0.getOpcode() == ISD::AND &&
2000 N0.getOperand(0).getOpcode() == ISD::XOR);
2001 // ((X^1)&1)^1 -> X & 1
2002 Val = DAG.getNode(ISD::AND, dl, N0.getValueType(),
2003 N0.getOperand(0).getOperand(0),
2007 return DAG.getSetCC(dl, VT, Val, N1,
2008 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
2010 } else if (N1C->getAPIntValue() == 1 &&
2012 getBooleanContents() == ZeroOrOneBooleanContent)) {
2014 if (Op0.getOpcode() == ISD::TRUNCATE)
2015 Op0 = Op0.getOperand(0);
2017 if ((Op0.getOpcode() == ISD::XOR) &&
2018 Op0.getOperand(0).getOpcode() == ISD::SETCC &&
2019 Op0.getOperand(1).getOpcode() == ISD::SETCC) {
2020 // (xor (setcc), (setcc)) == / != 1 -> (setcc) != / == (setcc)
2021 Cond = (Cond == ISD::SETEQ) ? ISD::SETNE : ISD::SETEQ;
2022 return DAG.getSetCC(dl, VT, Op0.getOperand(0), Op0.getOperand(1),
2024 } else if (Op0.getOpcode() == ISD::AND &&
2025 isa<ConstantSDNode>(Op0.getOperand(1)) &&
2026 cast<ConstantSDNode>(Op0.getOperand(1))->getAPIntValue() == 1) {
2027 // If this is (X&1) == / != 1, normalize it to (X&1) != / == 0.
2028 if (Op0.getValueType().bitsGT(VT))
2029 Op0 = DAG.getNode(ISD::AND, dl, VT,
2030 DAG.getNode(ISD::TRUNCATE, dl, VT, Op0.getOperand(0)),
2031 DAG.getConstant(1, VT));
2032 else if (Op0.getValueType().bitsLT(VT))
2033 Op0 = DAG.getNode(ISD::AND, dl, VT,
2034 DAG.getNode(ISD::ANY_EXTEND, dl, VT, Op0.getOperand(0)),
2035 DAG.getConstant(1, VT));
2037 return DAG.getSetCC(dl, VT, Op0,
2038 DAG.getConstant(0, Op0.getValueType()),
2039 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
2044 APInt MinVal, MaxVal;
2045 unsigned OperandBitSize = N1C->getValueType(0).getSizeInBits();
2046 if (ISD::isSignedIntSetCC(Cond)) {
2047 MinVal = APInt::getSignedMinValue(OperandBitSize);
2048 MaxVal = APInt::getSignedMaxValue(OperandBitSize);
2050 MinVal = APInt::getMinValue(OperandBitSize);
2051 MaxVal = APInt::getMaxValue(OperandBitSize);
2054 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
2055 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
2056 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true
2057 // X >= C0 --> X > (C0-1)
2058 return DAG.getSetCC(dl, VT, N0,
2059 DAG.getConstant(C1-1, N1.getValueType()),
2060 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
2063 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
2064 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true
2065 // X <= C0 --> X < (C0+1)
2066 return DAG.getSetCC(dl, VT, N0,
2067 DAG.getConstant(C1+1, N1.getValueType()),
2068 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
2071 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
2072 return DAG.getConstant(0, VT); // X < MIN --> false
2073 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal)
2074 return DAG.getConstant(1, VT); // X >= MIN --> true
2075 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal)
2076 return DAG.getConstant(0, VT); // X > MAX --> false
2077 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal)
2078 return DAG.getConstant(1, VT); // X <= MAX --> true
2080 // Canonicalize setgt X, Min --> setne X, Min
2081 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
2082 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
2083 // Canonicalize setlt X, Max --> setne X, Max
2084 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
2085 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
2087 // If we have setult X, 1, turn it into seteq X, 0
2088 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
2089 return DAG.getSetCC(dl, VT, N0,
2090 DAG.getConstant(MinVal, N0.getValueType()),
2092 // If we have setugt X, Max-1, turn it into seteq X, Max
2093 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
2094 return DAG.getSetCC(dl, VT, N0,
2095 DAG.getConstant(MaxVal, N0.getValueType()),
2098 // If we have "setcc X, C0", check to see if we can shrink the immediate
2101 // SETUGT X, SINTMAX -> SETLT X, 0
2102 if (Cond == ISD::SETUGT &&
2103 C1 == APInt::getSignedMaxValue(OperandBitSize))
2104 return DAG.getSetCC(dl, VT, N0,
2105 DAG.getConstant(0, N1.getValueType()),
2108 // SETULT X, SINTMIN -> SETGT X, -1
2109 if (Cond == ISD::SETULT &&
2110 C1 == APInt::getSignedMinValue(OperandBitSize)) {
2111 SDValue ConstMinusOne =
2112 DAG.getConstant(APInt::getAllOnesValue(OperandBitSize),
2114 return DAG.getSetCC(dl, VT, N0, ConstMinusOne, ISD::SETGT);
2117 // Fold bit comparisons when we can.
2118 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
2119 (VT == N0.getValueType() ||
2120 (isTypeLegal(VT) && VT.bitsLE(N0.getValueType()))) &&
2121 N0.getOpcode() == ISD::AND)
2122 if (ConstantSDNode *AndRHS =
2123 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2124 EVT ShiftTy = DCI.isBeforeLegalize() ?
2125 getPointerTy() : getShiftAmountTy();
2126 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
2127 // Perform the xform if the AND RHS is a single bit.
2128 if (AndRHS->getAPIntValue().isPowerOf2()) {
2129 return DAG.getNode(ISD::TRUNCATE, dl, VT,
2130 DAG.getNode(ISD::SRL, dl, N0.getValueType(), N0,
2131 DAG.getConstant(AndRHS->getAPIntValue().logBase2(), ShiftTy)));
2133 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getAPIntValue()) {
2134 // (X & 8) == 8 --> (X & 8) >> 3
2135 // Perform the xform if C1 is a single bit.
2136 if (C1.isPowerOf2()) {
2137 return DAG.getNode(ISD::TRUNCATE, dl, VT,
2138 DAG.getNode(ISD::SRL, dl, N0.getValueType(), N0,
2139 DAG.getConstant(C1.logBase2(), ShiftTy)));
2145 if (isa<ConstantFPSDNode>(N0.getNode())) {
2146 // Constant fold or commute setcc.
2147 SDValue O = DAG.FoldSetCC(VT, N0, N1, Cond, dl);
2148 if (O.getNode()) return O;
2149 } else if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1.getNode())) {
2150 // If the RHS of an FP comparison is a constant, simplify it away in
2152 if (CFP->getValueAPF().isNaN()) {
2153 // If an operand is known to be a nan, we can fold it.
2154 switch (ISD::getUnorderedFlavor(Cond)) {
2155 default: llvm_unreachable("Unknown flavor!");
2156 case 0: // Known false.
2157 return DAG.getConstant(0, VT);
2158 case 1: // Known true.
2159 return DAG.getConstant(1, VT);
2160 case 2: // Undefined.
2161 return DAG.getUNDEF(VT);
2165 // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the
2166 // constant if knowing that the operand is non-nan is enough. We prefer to
2167 // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to
2169 if (Cond == ISD::SETO || Cond == ISD::SETUO)
2170 return DAG.getSetCC(dl, VT, N0, N0, Cond);
2172 // If the condition is not legal, see if we can find an equivalent one
2174 if (!isCondCodeLegal(Cond, N0.getValueType())) {
2175 // If the comparison was an awkward floating-point == or != and one of
2176 // the comparison operands is infinity or negative infinity, convert the
2177 // condition to a less-awkward <= or >=.
2178 if (CFP->getValueAPF().isInfinity()) {
2179 if (CFP->getValueAPF().isNegative()) {
2180 if (Cond == ISD::SETOEQ &&
2181 isCondCodeLegal(ISD::SETOLE, N0.getValueType()))
2182 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLE);
2183 if (Cond == ISD::SETUEQ &&
2184 isCondCodeLegal(ISD::SETOLE, N0.getValueType()))
2185 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULE);
2186 if (Cond == ISD::SETUNE &&
2187 isCondCodeLegal(ISD::SETUGT, N0.getValueType()))
2188 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGT);
2189 if (Cond == ISD::SETONE &&
2190 isCondCodeLegal(ISD::SETUGT, N0.getValueType()))
2191 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGT);
2193 if (Cond == ISD::SETOEQ &&
2194 isCondCodeLegal(ISD::SETOGE, N0.getValueType()))
2195 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGE);
2196 if (Cond == ISD::SETUEQ &&
2197 isCondCodeLegal(ISD::SETOGE, N0.getValueType()))
2198 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGE);
2199 if (Cond == ISD::SETUNE &&
2200 isCondCodeLegal(ISD::SETULT, N0.getValueType()))
2201 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULT);
2202 if (Cond == ISD::SETONE &&
2203 isCondCodeLegal(ISD::SETULT, N0.getValueType()))
2204 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLT);
2211 // We can always fold X == X for integer setcc's.
2212 if (N0.getValueType().isInteger())
2213 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
2214 unsigned UOF = ISD::getUnorderedFlavor(Cond);
2215 if (UOF == 2) // FP operators that are undefined on NaNs.
2216 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
2217 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
2218 return DAG.getConstant(UOF, VT);
2219 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
2220 // if it is not already.
2221 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
2222 if (NewCond != Cond)
2223 return DAG.getSetCC(dl, VT, N0, N1, NewCond);
2226 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
2227 N0.getValueType().isInteger()) {
2228 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
2229 N0.getOpcode() == ISD::XOR) {
2230 // Simplify (X+Y) == (X+Z) --> Y == Z
2231 if (N0.getOpcode() == N1.getOpcode()) {
2232 if (N0.getOperand(0) == N1.getOperand(0))
2233 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(1), Cond);
2234 if (N0.getOperand(1) == N1.getOperand(1))
2235 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(0), Cond);
2236 if (DAG.isCommutativeBinOp(N0.getOpcode())) {
2237 // If X op Y == Y op X, try other combinations.
2238 if (N0.getOperand(0) == N1.getOperand(1))
2239 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(0),
2241 if (N0.getOperand(1) == N1.getOperand(0))
2242 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(1),
2247 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) {
2248 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2249 // Turn (X+C1) == C2 --> X == C2-C1
2250 if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse()) {
2251 return DAG.getSetCC(dl, VT, N0.getOperand(0),
2252 DAG.getConstant(RHSC->getAPIntValue()-
2253 LHSR->getAPIntValue(),
2254 N0.getValueType()), Cond);
2257 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
2258 if (N0.getOpcode() == ISD::XOR)
2259 // If we know that all of the inverted bits are zero, don't bother
2260 // performing the inversion.
2261 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue()))
2263 DAG.getSetCC(dl, VT, N0.getOperand(0),
2264 DAG.getConstant(LHSR->getAPIntValue() ^
2265 RHSC->getAPIntValue(),
2270 // Turn (C1-X) == C2 --> X == C1-C2
2271 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
2272 if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse()) {
2274 DAG.getSetCC(dl, VT, N0.getOperand(1),
2275 DAG.getConstant(SUBC->getAPIntValue() -
2276 RHSC->getAPIntValue(),
2283 // Simplify (X+Z) == X --> Z == 0
2284 if (N0.getOperand(0) == N1)
2285 return DAG.getSetCC(dl, VT, N0.getOperand(1),
2286 DAG.getConstant(0, N0.getValueType()), Cond);
2287 if (N0.getOperand(1) == N1) {
2288 if (DAG.isCommutativeBinOp(N0.getOpcode()))
2289 return DAG.getSetCC(dl, VT, N0.getOperand(0),
2290 DAG.getConstant(0, N0.getValueType()), Cond);
2291 else if (N0.getNode()->hasOneUse()) {
2292 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
2293 // (Z-X) == X --> Z == X<<1
2294 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(),
2296 DAG.getConstant(1, getShiftAmountTy()));
2297 if (!DCI.isCalledByLegalizer())
2298 DCI.AddToWorklist(SH.getNode());
2299 return DAG.getSetCC(dl, VT, N0.getOperand(0), SH, Cond);
2304 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
2305 N1.getOpcode() == ISD::XOR) {
2306 // Simplify X == (X+Z) --> Z == 0
2307 if (N1.getOperand(0) == N0) {
2308 return DAG.getSetCC(dl, VT, N1.getOperand(1),
2309 DAG.getConstant(0, N1.getValueType()), Cond);
2310 } else if (N1.getOperand(1) == N0) {
2311 if (DAG.isCommutativeBinOp(N1.getOpcode())) {
2312 return DAG.getSetCC(dl, VT, N1.getOperand(0),
2313 DAG.getConstant(0, N1.getValueType()), Cond);
2314 } else if (N1.getNode()->hasOneUse()) {
2315 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
2316 // X == (Z-X) --> X<<1 == Z
2317 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(), N0,
2318 DAG.getConstant(1, getShiftAmountTy()));
2319 if (!DCI.isCalledByLegalizer())
2320 DCI.AddToWorklist(SH.getNode());
2321 return DAG.getSetCC(dl, VT, SH, N1.getOperand(0), Cond);
2326 // Simplify x&y == y to x&y != 0 if y has exactly one bit set.
2327 // Note that where y is variable and is known to have at most
2328 // one bit set (for example, if it is z&1) we cannot do this;
2329 // the expressions are not equivalent when y==0.
2330 if (N0.getOpcode() == ISD::AND)
2331 if (N0.getOperand(0) == N1 || N0.getOperand(1) == N1) {
2332 if (ValueHasExactlyOneBitSet(N1, DAG)) {
2333 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true);
2334 SDValue Zero = DAG.getConstant(0, N1.getValueType());
2335 return DAG.getSetCC(dl, VT, N0, Zero, Cond);
2338 if (N1.getOpcode() == ISD::AND)
2339 if (N1.getOperand(0) == N0 || N1.getOperand(1) == N0) {
2340 if (ValueHasExactlyOneBitSet(N0, DAG)) {
2341 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true);
2342 SDValue Zero = DAG.getConstant(0, N0.getValueType());
2343 return DAG.getSetCC(dl, VT, N1, Zero, Cond);
2348 // Fold away ALL boolean setcc's.
2350 if (N0.getValueType() == MVT::i1 && foldBooleans) {
2352 default: llvm_unreachable("Unknown integer setcc!");
2353 case ISD::SETEQ: // X == Y -> ~(X^Y)
2354 Temp = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1);
2355 N0 = DAG.getNOT(dl, Temp, MVT::i1);
2356 if (!DCI.isCalledByLegalizer())
2357 DCI.AddToWorklist(Temp.getNode());
2359 case ISD::SETNE: // X != Y --> (X^Y)
2360 N0 = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1);
2362 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> ~X & Y
2363 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> ~X & Y
2364 Temp = DAG.getNOT(dl, N0, MVT::i1);
2365 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N1, Temp);
2366 if (!DCI.isCalledByLegalizer())
2367 DCI.AddToWorklist(Temp.getNode());
2369 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> ~Y & X
2370 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> ~Y & X
2371 Temp = DAG.getNOT(dl, N1, MVT::i1);
2372 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N0, Temp);
2373 if (!DCI.isCalledByLegalizer())
2374 DCI.AddToWorklist(Temp.getNode());
2376 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> ~X | Y
2377 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> ~X | Y
2378 Temp = DAG.getNOT(dl, N0, MVT::i1);
2379 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N1, Temp);
2380 if (!DCI.isCalledByLegalizer())
2381 DCI.AddToWorklist(Temp.getNode());
2383 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> ~Y | X
2384 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> ~Y | X
2385 Temp = DAG.getNOT(dl, N1, MVT::i1);
2386 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N0, Temp);
2389 if (VT != MVT::i1) {
2390 if (!DCI.isCalledByLegalizer())
2391 DCI.AddToWorklist(N0.getNode());
2392 // FIXME: If running after legalize, we probably can't do this.
2393 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, N0);
2398 // Could not fold it.
2402 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
2403 /// node is a GlobalAddress + offset.
2404 bool TargetLowering::isGAPlusOffset(SDNode *N, const GlobalValue* &GA,
2405 int64_t &Offset) const {
2406 if (isa<GlobalAddressSDNode>(N)) {
2407 GlobalAddressSDNode *GASD = cast<GlobalAddressSDNode>(N);
2408 GA = GASD->getGlobal();
2409 Offset += GASD->getOffset();
2413 if (N->getOpcode() == ISD::ADD) {
2414 SDValue N1 = N->getOperand(0);
2415 SDValue N2 = N->getOperand(1);
2416 if (isGAPlusOffset(N1.getNode(), GA, Offset)) {
2417 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
2419 Offset += V->getSExtValue();
2422 } else if (isGAPlusOffset(N2.getNode(), GA, Offset)) {
2423 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
2425 Offset += V->getSExtValue();
2434 SDValue TargetLowering::
2435 PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
2436 // Default implementation: no optimization.
2440 //===----------------------------------------------------------------------===//
2441 // Inline Assembler Implementation Methods
2442 //===----------------------------------------------------------------------===//
2445 TargetLowering::ConstraintType
2446 TargetLowering::getConstraintType(const std::string &Constraint) const {
2447 // FIXME: lots more standard ones to handle.
2448 if (Constraint.size() == 1) {
2449 switch (Constraint[0]) {
2451 case 'r': return C_RegisterClass;
2453 case 'o': // offsetable
2454 case 'V': // not offsetable
2456 case 'i': // Simple Integer or Relocatable Constant
2457 case 'n': // Simple Integer
2458 case 's': // Relocatable Constant
2459 case 'X': // Allow ANY value.
2460 case 'I': // Target registers.
2472 if (Constraint.size() > 1 && Constraint[0] == '{' &&
2473 Constraint[Constraint.size()-1] == '}')
2478 /// LowerXConstraint - try to replace an X constraint, which matches anything,
2479 /// with another that has more specific requirements based on the type of the
2480 /// corresponding operand.
2481 const char *TargetLowering::LowerXConstraint(EVT ConstraintVT) const{
2482 if (ConstraintVT.isInteger())
2484 if (ConstraintVT.isFloatingPoint())
2485 return "f"; // works for many targets
2489 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
2490 /// vector. If it is invalid, don't add anything to Ops.
2491 void TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
2492 char ConstraintLetter,
2493 std::vector<SDValue> &Ops,
2494 SelectionDAG &DAG) const {
2495 switch (ConstraintLetter) {
2497 case 'X': // Allows any operand; labels (basic block) use this.
2498 if (Op.getOpcode() == ISD::BasicBlock) {
2503 case 'i': // Simple Integer or Relocatable Constant
2504 case 'n': // Simple Integer
2505 case 's': { // Relocatable Constant
2506 // These operands are interested in values of the form (GV+C), where C may
2507 // be folded in as an offset of GV, or it may be explicitly added. Also, it
2508 // is possible and fine if either GV or C are missing.
2509 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2510 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
2512 // If we have "(add GV, C)", pull out GV/C
2513 if (Op.getOpcode() == ISD::ADD) {
2514 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
2515 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
2516 if (C == 0 || GA == 0) {
2517 C = dyn_cast<ConstantSDNode>(Op.getOperand(0));
2518 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1));
2520 if (C == 0 || GA == 0)
2524 // If we find a valid operand, map to the TargetXXX version so that the
2525 // value itself doesn't get selected.
2526 if (GA) { // Either &GV or &GV+C
2527 if (ConstraintLetter != 'n') {
2528 int64_t Offs = GA->getOffset();
2529 if (C) Offs += C->getZExtValue();
2530 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(),
2531 C ? C->getDebugLoc() : DebugLoc(),
2532 Op.getValueType(), Offs));
2536 if (C) { // just C, no GV.
2537 // Simple constants are not allowed for 's'.
2538 if (ConstraintLetter != 's') {
2539 // gcc prints these as sign extended. Sign extend value to 64 bits
2540 // now; without this it would get ZExt'd later in
2541 // ScheduleDAGSDNodes::EmitNode, which is very generic.
2542 Ops.push_back(DAG.getTargetConstant(C->getAPIntValue().getSExtValue(),
2552 std::vector<unsigned> TargetLowering::
2553 getRegClassForInlineAsmConstraint(const std::string &Constraint,
2555 return std::vector<unsigned>();
2559 std::pair<unsigned, const TargetRegisterClass*> TargetLowering::
2560 getRegForInlineAsmConstraint(const std::string &Constraint,
2562 if (Constraint[0] != '{')
2563 return std::make_pair(0u, static_cast<TargetRegisterClass*>(0));
2564 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?");
2566 // Remove the braces from around the name.
2567 StringRef RegName(Constraint.data()+1, Constraint.size()-2);
2569 // Figure out which register class contains this reg.
2570 const TargetRegisterInfo *RI = TM.getRegisterInfo();
2571 for (TargetRegisterInfo::regclass_iterator RCI = RI->regclass_begin(),
2572 E = RI->regclass_end(); RCI != E; ++RCI) {
2573 const TargetRegisterClass *RC = *RCI;
2575 // If none of the value types for this register class are valid, we
2576 // can't use it. For example, 64-bit reg classes on 32-bit targets.
2577 bool isLegal = false;
2578 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
2580 if (isTypeLegal(*I)) {
2586 if (!isLegal) continue;
2588 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end();
2590 if (RegName.equals_lower(RI->getName(*I)))
2591 return std::make_pair(*I, RC);
2595 return std::make_pair(0u, static_cast<const TargetRegisterClass*>(0));
2598 //===----------------------------------------------------------------------===//
2599 // Constraint Selection.
2601 /// isMatchingInputConstraint - Return true of this is an input operand that is
2602 /// a matching constraint like "4".
2603 bool TargetLowering::AsmOperandInfo::isMatchingInputConstraint() const {
2604 assert(!ConstraintCode.empty() && "No known constraint!");
2605 return isdigit(ConstraintCode[0]);
2608 /// getMatchedOperand - If this is an input matching constraint, this method
2609 /// returns the output operand it matches.
2610 unsigned TargetLowering::AsmOperandInfo::getMatchedOperand() const {
2611 assert(!ConstraintCode.empty() && "No known constraint!");
2612 return atoi(ConstraintCode.c_str());
2616 /// getConstraintGenerality - Return an integer indicating how general CT
2618 static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) {
2620 default: llvm_unreachable("Unknown constraint type!");
2621 case TargetLowering::C_Other:
2622 case TargetLowering::C_Unknown:
2624 case TargetLowering::C_Register:
2626 case TargetLowering::C_RegisterClass:
2628 case TargetLowering::C_Memory:
2633 /// ChooseConstraint - If there are multiple different constraints that we
2634 /// could pick for this operand (e.g. "imr") try to pick the 'best' one.
2635 /// This is somewhat tricky: constraints fall into four classes:
2636 /// Other -> immediates and magic values
2637 /// Register -> one specific register
2638 /// RegisterClass -> a group of regs
2639 /// Memory -> memory
2640 /// Ideally, we would pick the most specific constraint possible: if we have
2641 /// something that fits into a register, we would pick it. The problem here
2642 /// is that if we have something that could either be in a register or in
2643 /// memory that use of the register could cause selection of *other*
2644 /// operands to fail: they might only succeed if we pick memory. Because of
2645 /// this the heuristic we use is:
2647 /// 1) If there is an 'other' constraint, and if the operand is valid for
2648 /// that constraint, use it. This makes us take advantage of 'i'
2649 /// constraints when available.
2650 /// 2) Otherwise, pick the most general constraint present. This prefers
2651 /// 'm' over 'r', for example.
2653 static void ChooseConstraint(TargetLowering::AsmOperandInfo &OpInfo,
2654 const TargetLowering &TLI,
2655 SDValue Op, SelectionDAG *DAG) {
2656 assert(OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options");
2657 unsigned BestIdx = 0;
2658 TargetLowering::ConstraintType BestType = TargetLowering::C_Unknown;
2659 int BestGenerality = -1;
2661 // Loop over the options, keeping track of the most general one.
2662 for (unsigned i = 0, e = OpInfo.Codes.size(); i != e; ++i) {
2663 TargetLowering::ConstraintType CType =
2664 TLI.getConstraintType(OpInfo.Codes[i]);
2666 // If this is an 'other' constraint, see if the operand is valid for it.
2667 // For example, on X86 we might have an 'rI' constraint. If the operand
2668 // is an integer in the range [0..31] we want to use I (saving a load
2669 // of a register), otherwise we must use 'r'.
2670 if (CType == TargetLowering::C_Other && Op.getNode()) {
2671 assert(OpInfo.Codes[i].size() == 1 &&
2672 "Unhandled multi-letter 'other' constraint");
2673 std::vector<SDValue> ResultOps;
2674 TLI.LowerAsmOperandForConstraint(Op, OpInfo.Codes[i][0],
2676 if (!ResultOps.empty()) {
2683 // Things with matching constraints can only be registers, per gcc
2684 // documentation. This mainly affects "g" constraints.
2685 if (CType == TargetLowering::C_Memory && OpInfo.hasMatchingInput())
2688 // This constraint letter is more general than the previous one, use it.
2689 int Generality = getConstraintGenerality(CType);
2690 if (Generality > BestGenerality) {
2693 BestGenerality = Generality;
2697 OpInfo.ConstraintCode = OpInfo.Codes[BestIdx];
2698 OpInfo.ConstraintType = BestType;
2701 /// ComputeConstraintToUse - Determines the constraint code and constraint
2702 /// type to use for the specific AsmOperandInfo, setting
2703 /// OpInfo.ConstraintCode and OpInfo.ConstraintType.
2704 void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo,
2706 SelectionDAG *DAG) const {
2707 assert(!OpInfo.Codes.empty() && "Must have at least one constraint");
2709 // Single-letter constraints ('r') are very common.
2710 if (OpInfo.Codes.size() == 1) {
2711 OpInfo.ConstraintCode = OpInfo.Codes[0];
2712 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
2714 ChooseConstraint(OpInfo, *this, Op, DAG);
2717 // 'X' matches anything.
2718 if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) {
2719 // Labels and constants are handled elsewhere ('X' is the only thing
2720 // that matches labels). For Functions, the type here is the type of
2721 // the result, which is not what we want to look at; leave them alone.
2722 Value *v = OpInfo.CallOperandVal;
2723 if (isa<BasicBlock>(v) || isa<ConstantInt>(v) || isa<Function>(v)) {
2724 OpInfo.CallOperandVal = v;
2728 // Otherwise, try to resolve it to something we know about by looking at
2729 // the actual operand type.
2730 if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) {
2731 OpInfo.ConstraintCode = Repl;
2732 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
2737 //===----------------------------------------------------------------------===//
2738 // Loop Strength Reduction hooks
2739 //===----------------------------------------------------------------------===//
2741 /// isLegalAddressingMode - Return true if the addressing mode represented
2742 /// by AM is legal for this target, for a load/store of the specified type.
2743 bool TargetLowering::isLegalAddressingMode(const AddrMode &AM,
2744 const Type *Ty) const {
2745 // The default implementation of this implements a conservative RISCy, r+r and
2748 // Allows a sign-extended 16-bit immediate field.
2749 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
2752 // No global is ever allowed as a base.
2756 // Only support r+r,
2758 case 0: // "r+i" or just "i", depending on HasBaseReg.
2761 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
2763 // Otherwise we have r+r or r+i.
2766 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
2768 // Allow 2*r as r+r.
2775 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
2776 /// return a DAG expression to select that will generate the same value by
2777 /// multiplying by a magic number. See:
2778 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
2779 SDValue TargetLowering::BuildSDIV(SDNode *N, SelectionDAG &DAG,
2780 std::vector<SDNode*>* Created) const {
2781 EVT VT = N->getValueType(0);
2782 DebugLoc dl= N->getDebugLoc();
2784 // Check to see if we can do this.
2785 // FIXME: We should be more aggressive here.
2786 if (!isTypeLegal(VT))
2789 APInt d = cast<ConstantSDNode>(N->getOperand(1))->getAPIntValue();
2790 APInt::ms magics = d.magic();
2792 // Multiply the numerator (operand 0) by the magic value
2793 // FIXME: We should support doing a MUL in a wider type
2795 if (isOperationLegalOrCustom(ISD::MULHS, VT))
2796 Q = DAG.getNode(ISD::MULHS, dl, VT, N->getOperand(0),
2797 DAG.getConstant(magics.m, VT));
2798 else if (isOperationLegalOrCustom(ISD::SMUL_LOHI, VT))
2799 Q = SDValue(DAG.getNode(ISD::SMUL_LOHI, dl, DAG.getVTList(VT, VT),
2801 DAG.getConstant(magics.m, VT)).getNode(), 1);
2803 return SDValue(); // No mulhs or equvialent
2804 // If d > 0 and m < 0, add the numerator
2805 if (d.isStrictlyPositive() && magics.m.isNegative()) {
2806 Q = DAG.getNode(ISD::ADD, dl, VT, Q, N->getOperand(0));
2808 Created->push_back(Q.getNode());
2810 // If d < 0 and m > 0, subtract the numerator.
2811 if (d.isNegative() && magics.m.isStrictlyPositive()) {
2812 Q = DAG.getNode(ISD::SUB, dl, VT, Q, N->getOperand(0));
2814 Created->push_back(Q.getNode());
2816 // Shift right algebraic if shift value is nonzero
2818 Q = DAG.getNode(ISD::SRA, dl, VT, Q,
2819 DAG.getConstant(magics.s, getShiftAmountTy()));
2821 Created->push_back(Q.getNode());
2823 // Extract the sign bit and add it to the quotient
2825 DAG.getNode(ISD::SRL, dl, VT, Q, DAG.getConstant(VT.getSizeInBits()-1,
2826 getShiftAmountTy()));
2828 Created->push_back(T.getNode());
2829 return DAG.getNode(ISD::ADD, dl, VT, Q, T);
2832 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
2833 /// return a DAG expression to select that will generate the same value by
2834 /// multiplying by a magic number. See:
2835 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
2836 SDValue TargetLowering::BuildUDIV(SDNode *N, SelectionDAG &DAG,
2837 std::vector<SDNode*>* Created) const {
2838 EVT VT = N->getValueType(0);
2839 DebugLoc dl = N->getDebugLoc();
2841 // Check to see if we can do this.
2842 // FIXME: We should be more aggressive here.
2843 if (!isTypeLegal(VT))
2846 // FIXME: We should use a narrower constant when the upper
2847 // bits are known to be zero.
2848 ConstantSDNode *N1C = cast<ConstantSDNode>(N->getOperand(1));
2849 APInt::mu magics = N1C->getAPIntValue().magicu();
2851 // Multiply the numerator (operand 0) by the magic value
2852 // FIXME: We should support doing a MUL in a wider type
2854 if (isOperationLegalOrCustom(ISD::MULHU, VT))
2855 Q = DAG.getNode(ISD::MULHU, dl, VT, N->getOperand(0),
2856 DAG.getConstant(magics.m, VT));
2857 else if (isOperationLegalOrCustom(ISD::UMUL_LOHI, VT))
2858 Q = SDValue(DAG.getNode(ISD::UMUL_LOHI, dl, DAG.getVTList(VT, VT),
2860 DAG.getConstant(magics.m, VT)).getNode(), 1);
2862 return SDValue(); // No mulhu or equvialent
2864 Created->push_back(Q.getNode());
2866 if (magics.a == 0) {
2867 assert(magics.s < N1C->getAPIntValue().getBitWidth() &&
2868 "We shouldn't generate an undefined shift!");
2869 return DAG.getNode(ISD::SRL, dl, VT, Q,
2870 DAG.getConstant(magics.s, getShiftAmountTy()));
2872 SDValue NPQ = DAG.getNode(ISD::SUB, dl, VT, N->getOperand(0), Q);
2874 Created->push_back(NPQ.getNode());
2875 NPQ = DAG.getNode(ISD::SRL, dl, VT, NPQ,
2876 DAG.getConstant(1, getShiftAmountTy()));
2878 Created->push_back(NPQ.getNode());
2879 NPQ = DAG.getNode(ISD::ADD, dl, VT, NPQ, Q);
2881 Created->push_back(NPQ.getNode());
2882 return DAG.getNode(ISD::SRL, dl, VT, NPQ,
2883 DAG.getConstant(magics.s-1, getShiftAmountTy()));