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/CommandLine.h"
30 #include "llvm/Support/ErrorHandling.h"
31 #include "llvm/Support/MathExtras.h"
35 /// We are in the process of implementing a new TypeLegalization action
36 /// - the promotion of vector elements. This feature is disabled by default
37 /// and only enabled using this flag.
39 AllowPromoteIntElem("promote-elements", cl::Hidden,
40 cl::desc("Allow promotion of integer vector element types"));
43 TLSModel::Model getTLSModel(const GlobalValue *GV, Reloc::Model reloc) {
44 bool isLocal = GV->hasLocalLinkage();
45 bool isDeclaration = GV->isDeclaration();
46 // FIXME: what should we do for protected and internal visibility?
47 // For variables, is internal different from hidden?
48 bool isHidden = GV->hasHiddenVisibility();
50 if (reloc == Reloc::PIC_) {
51 if (isLocal || isHidden)
52 return TLSModel::LocalDynamic;
54 return TLSModel::GeneralDynamic;
56 if (!isDeclaration || isHidden)
57 return TLSModel::LocalExec;
59 return TLSModel::InitialExec;
64 /// InitLibcallNames - Set default libcall names.
66 static void InitLibcallNames(const char **Names) {
67 Names[RTLIB::SHL_I16] = "__ashlhi3";
68 Names[RTLIB::SHL_I32] = "__ashlsi3";
69 Names[RTLIB::SHL_I64] = "__ashldi3";
70 Names[RTLIB::SHL_I128] = "__ashlti3";
71 Names[RTLIB::SRL_I16] = "__lshrhi3";
72 Names[RTLIB::SRL_I32] = "__lshrsi3";
73 Names[RTLIB::SRL_I64] = "__lshrdi3";
74 Names[RTLIB::SRL_I128] = "__lshrti3";
75 Names[RTLIB::SRA_I16] = "__ashrhi3";
76 Names[RTLIB::SRA_I32] = "__ashrsi3";
77 Names[RTLIB::SRA_I64] = "__ashrdi3";
78 Names[RTLIB::SRA_I128] = "__ashrti3";
79 Names[RTLIB::MUL_I8] = "__mulqi3";
80 Names[RTLIB::MUL_I16] = "__mulhi3";
81 Names[RTLIB::MUL_I32] = "__mulsi3";
82 Names[RTLIB::MUL_I64] = "__muldi3";
83 Names[RTLIB::MUL_I128] = "__multi3";
84 Names[RTLIB::SDIV_I8] = "__divqi3";
85 Names[RTLIB::SDIV_I16] = "__divhi3";
86 Names[RTLIB::SDIV_I32] = "__divsi3";
87 Names[RTLIB::SDIV_I64] = "__divdi3";
88 Names[RTLIB::SDIV_I128] = "__divti3";
89 Names[RTLIB::UDIV_I8] = "__udivqi3";
90 Names[RTLIB::UDIV_I16] = "__udivhi3";
91 Names[RTLIB::UDIV_I32] = "__udivsi3";
92 Names[RTLIB::UDIV_I64] = "__udivdi3";
93 Names[RTLIB::UDIV_I128] = "__udivti3";
94 Names[RTLIB::SREM_I8] = "__modqi3";
95 Names[RTLIB::SREM_I16] = "__modhi3";
96 Names[RTLIB::SREM_I32] = "__modsi3";
97 Names[RTLIB::SREM_I64] = "__moddi3";
98 Names[RTLIB::SREM_I128] = "__modti3";
99 Names[RTLIB::UREM_I8] = "__umodqi3";
100 Names[RTLIB::UREM_I16] = "__umodhi3";
101 Names[RTLIB::UREM_I32] = "__umodsi3";
102 Names[RTLIB::UREM_I64] = "__umoddi3";
103 Names[RTLIB::UREM_I128] = "__umodti3";
105 // These are generally not available.
106 Names[RTLIB::SDIVREM_I8] = 0;
107 Names[RTLIB::SDIVREM_I16] = 0;
108 Names[RTLIB::SDIVREM_I32] = 0;
109 Names[RTLIB::SDIVREM_I64] = 0;
110 Names[RTLIB::SDIVREM_I128] = 0;
111 Names[RTLIB::UDIVREM_I8] = 0;
112 Names[RTLIB::UDIVREM_I16] = 0;
113 Names[RTLIB::UDIVREM_I32] = 0;
114 Names[RTLIB::UDIVREM_I64] = 0;
115 Names[RTLIB::UDIVREM_I128] = 0;
117 Names[RTLIB::NEG_I32] = "__negsi2";
118 Names[RTLIB::NEG_I64] = "__negdi2";
119 Names[RTLIB::ADD_F32] = "__addsf3";
120 Names[RTLIB::ADD_F64] = "__adddf3";
121 Names[RTLIB::ADD_F80] = "__addxf3";
122 Names[RTLIB::ADD_PPCF128] = "__gcc_qadd";
123 Names[RTLIB::SUB_F32] = "__subsf3";
124 Names[RTLIB::SUB_F64] = "__subdf3";
125 Names[RTLIB::SUB_F80] = "__subxf3";
126 Names[RTLIB::SUB_PPCF128] = "__gcc_qsub";
127 Names[RTLIB::MUL_F32] = "__mulsf3";
128 Names[RTLIB::MUL_F64] = "__muldf3";
129 Names[RTLIB::MUL_F80] = "__mulxf3";
130 Names[RTLIB::MUL_PPCF128] = "__gcc_qmul";
131 Names[RTLIB::DIV_F32] = "__divsf3";
132 Names[RTLIB::DIV_F64] = "__divdf3";
133 Names[RTLIB::DIV_F80] = "__divxf3";
134 Names[RTLIB::DIV_PPCF128] = "__gcc_qdiv";
135 Names[RTLIB::REM_F32] = "fmodf";
136 Names[RTLIB::REM_F64] = "fmod";
137 Names[RTLIB::REM_F80] = "fmodl";
138 Names[RTLIB::REM_PPCF128] = "fmodl";
139 Names[RTLIB::POWI_F32] = "__powisf2";
140 Names[RTLIB::POWI_F64] = "__powidf2";
141 Names[RTLIB::POWI_F80] = "__powixf2";
142 Names[RTLIB::POWI_PPCF128] = "__powitf2";
143 Names[RTLIB::SQRT_F32] = "sqrtf";
144 Names[RTLIB::SQRT_F64] = "sqrt";
145 Names[RTLIB::SQRT_F80] = "sqrtl";
146 Names[RTLIB::SQRT_PPCF128] = "sqrtl";
147 Names[RTLIB::LOG_F32] = "logf";
148 Names[RTLIB::LOG_F64] = "log";
149 Names[RTLIB::LOG_F80] = "logl";
150 Names[RTLIB::LOG_PPCF128] = "logl";
151 Names[RTLIB::LOG2_F32] = "log2f";
152 Names[RTLIB::LOG2_F64] = "log2";
153 Names[RTLIB::LOG2_F80] = "log2l";
154 Names[RTLIB::LOG2_PPCF128] = "log2l";
155 Names[RTLIB::LOG10_F32] = "log10f";
156 Names[RTLIB::LOG10_F64] = "log10";
157 Names[RTLIB::LOG10_F80] = "log10l";
158 Names[RTLIB::LOG10_PPCF128] = "log10l";
159 Names[RTLIB::EXP_F32] = "expf";
160 Names[RTLIB::EXP_F64] = "exp";
161 Names[RTLIB::EXP_F80] = "expl";
162 Names[RTLIB::EXP_PPCF128] = "expl";
163 Names[RTLIB::EXP2_F32] = "exp2f";
164 Names[RTLIB::EXP2_F64] = "exp2";
165 Names[RTLIB::EXP2_F80] = "exp2l";
166 Names[RTLIB::EXP2_PPCF128] = "exp2l";
167 Names[RTLIB::SIN_F32] = "sinf";
168 Names[RTLIB::SIN_F64] = "sin";
169 Names[RTLIB::SIN_F80] = "sinl";
170 Names[RTLIB::SIN_PPCF128] = "sinl";
171 Names[RTLIB::COS_F32] = "cosf";
172 Names[RTLIB::COS_F64] = "cos";
173 Names[RTLIB::COS_F80] = "cosl";
174 Names[RTLIB::COS_PPCF128] = "cosl";
175 Names[RTLIB::POW_F32] = "powf";
176 Names[RTLIB::POW_F64] = "pow";
177 Names[RTLIB::POW_F80] = "powl";
178 Names[RTLIB::POW_PPCF128] = "powl";
179 Names[RTLIB::CEIL_F32] = "ceilf";
180 Names[RTLIB::CEIL_F64] = "ceil";
181 Names[RTLIB::CEIL_F80] = "ceill";
182 Names[RTLIB::CEIL_PPCF128] = "ceill";
183 Names[RTLIB::TRUNC_F32] = "truncf";
184 Names[RTLIB::TRUNC_F64] = "trunc";
185 Names[RTLIB::TRUNC_F80] = "truncl";
186 Names[RTLIB::TRUNC_PPCF128] = "truncl";
187 Names[RTLIB::RINT_F32] = "rintf";
188 Names[RTLIB::RINT_F64] = "rint";
189 Names[RTLIB::RINT_F80] = "rintl";
190 Names[RTLIB::RINT_PPCF128] = "rintl";
191 Names[RTLIB::NEARBYINT_F32] = "nearbyintf";
192 Names[RTLIB::NEARBYINT_F64] = "nearbyint";
193 Names[RTLIB::NEARBYINT_F80] = "nearbyintl";
194 Names[RTLIB::NEARBYINT_PPCF128] = "nearbyintl";
195 Names[RTLIB::FLOOR_F32] = "floorf";
196 Names[RTLIB::FLOOR_F64] = "floor";
197 Names[RTLIB::FLOOR_F80] = "floorl";
198 Names[RTLIB::FLOOR_PPCF128] = "floorl";
199 Names[RTLIB::COPYSIGN_F32] = "copysignf";
200 Names[RTLIB::COPYSIGN_F64] = "copysign";
201 Names[RTLIB::COPYSIGN_F80] = "copysignl";
202 Names[RTLIB::COPYSIGN_PPCF128] = "copysignl";
203 Names[RTLIB::FPEXT_F32_F64] = "__extendsfdf2";
204 Names[RTLIB::FPEXT_F16_F32] = "__gnu_h2f_ieee";
205 Names[RTLIB::FPROUND_F32_F16] = "__gnu_f2h_ieee";
206 Names[RTLIB::FPROUND_F64_F32] = "__truncdfsf2";
207 Names[RTLIB::FPROUND_F80_F32] = "__truncxfsf2";
208 Names[RTLIB::FPROUND_PPCF128_F32] = "__trunctfsf2";
209 Names[RTLIB::FPROUND_F80_F64] = "__truncxfdf2";
210 Names[RTLIB::FPROUND_PPCF128_F64] = "__trunctfdf2";
211 Names[RTLIB::FPTOSINT_F32_I8] = "__fixsfqi";
212 Names[RTLIB::FPTOSINT_F32_I16] = "__fixsfhi";
213 Names[RTLIB::FPTOSINT_F32_I32] = "__fixsfsi";
214 Names[RTLIB::FPTOSINT_F32_I64] = "__fixsfdi";
215 Names[RTLIB::FPTOSINT_F32_I128] = "__fixsfti";
216 Names[RTLIB::FPTOSINT_F64_I8] = "__fixdfqi";
217 Names[RTLIB::FPTOSINT_F64_I16] = "__fixdfhi";
218 Names[RTLIB::FPTOSINT_F64_I32] = "__fixdfsi";
219 Names[RTLIB::FPTOSINT_F64_I64] = "__fixdfdi";
220 Names[RTLIB::FPTOSINT_F64_I128] = "__fixdfti";
221 Names[RTLIB::FPTOSINT_F80_I32] = "__fixxfsi";
222 Names[RTLIB::FPTOSINT_F80_I64] = "__fixxfdi";
223 Names[RTLIB::FPTOSINT_F80_I128] = "__fixxfti";
224 Names[RTLIB::FPTOSINT_PPCF128_I32] = "__fixtfsi";
225 Names[RTLIB::FPTOSINT_PPCF128_I64] = "__fixtfdi";
226 Names[RTLIB::FPTOSINT_PPCF128_I128] = "__fixtfti";
227 Names[RTLIB::FPTOUINT_F32_I8] = "__fixunssfqi";
228 Names[RTLIB::FPTOUINT_F32_I16] = "__fixunssfhi";
229 Names[RTLIB::FPTOUINT_F32_I32] = "__fixunssfsi";
230 Names[RTLIB::FPTOUINT_F32_I64] = "__fixunssfdi";
231 Names[RTLIB::FPTOUINT_F32_I128] = "__fixunssfti";
232 Names[RTLIB::FPTOUINT_F64_I8] = "__fixunsdfqi";
233 Names[RTLIB::FPTOUINT_F64_I16] = "__fixunsdfhi";
234 Names[RTLIB::FPTOUINT_F64_I32] = "__fixunsdfsi";
235 Names[RTLIB::FPTOUINT_F64_I64] = "__fixunsdfdi";
236 Names[RTLIB::FPTOUINT_F64_I128] = "__fixunsdfti";
237 Names[RTLIB::FPTOUINT_F80_I32] = "__fixunsxfsi";
238 Names[RTLIB::FPTOUINT_F80_I64] = "__fixunsxfdi";
239 Names[RTLIB::FPTOUINT_F80_I128] = "__fixunsxfti";
240 Names[RTLIB::FPTOUINT_PPCF128_I32] = "__fixunstfsi";
241 Names[RTLIB::FPTOUINT_PPCF128_I64] = "__fixunstfdi";
242 Names[RTLIB::FPTOUINT_PPCF128_I128] = "__fixunstfti";
243 Names[RTLIB::SINTTOFP_I32_F32] = "__floatsisf";
244 Names[RTLIB::SINTTOFP_I32_F64] = "__floatsidf";
245 Names[RTLIB::SINTTOFP_I32_F80] = "__floatsixf";
246 Names[RTLIB::SINTTOFP_I32_PPCF128] = "__floatsitf";
247 Names[RTLIB::SINTTOFP_I64_F32] = "__floatdisf";
248 Names[RTLIB::SINTTOFP_I64_F64] = "__floatdidf";
249 Names[RTLIB::SINTTOFP_I64_F80] = "__floatdixf";
250 Names[RTLIB::SINTTOFP_I64_PPCF128] = "__floatditf";
251 Names[RTLIB::SINTTOFP_I128_F32] = "__floattisf";
252 Names[RTLIB::SINTTOFP_I128_F64] = "__floattidf";
253 Names[RTLIB::SINTTOFP_I128_F80] = "__floattixf";
254 Names[RTLIB::SINTTOFP_I128_PPCF128] = "__floattitf";
255 Names[RTLIB::UINTTOFP_I32_F32] = "__floatunsisf";
256 Names[RTLIB::UINTTOFP_I32_F64] = "__floatunsidf";
257 Names[RTLIB::UINTTOFP_I32_F80] = "__floatunsixf";
258 Names[RTLIB::UINTTOFP_I32_PPCF128] = "__floatunsitf";
259 Names[RTLIB::UINTTOFP_I64_F32] = "__floatundisf";
260 Names[RTLIB::UINTTOFP_I64_F64] = "__floatundidf";
261 Names[RTLIB::UINTTOFP_I64_F80] = "__floatundixf";
262 Names[RTLIB::UINTTOFP_I64_PPCF128] = "__floatunditf";
263 Names[RTLIB::UINTTOFP_I128_F32] = "__floatuntisf";
264 Names[RTLIB::UINTTOFP_I128_F64] = "__floatuntidf";
265 Names[RTLIB::UINTTOFP_I128_F80] = "__floatuntixf";
266 Names[RTLIB::UINTTOFP_I128_PPCF128] = "__floatuntitf";
267 Names[RTLIB::OEQ_F32] = "__eqsf2";
268 Names[RTLIB::OEQ_F64] = "__eqdf2";
269 Names[RTLIB::UNE_F32] = "__nesf2";
270 Names[RTLIB::UNE_F64] = "__nedf2";
271 Names[RTLIB::OGE_F32] = "__gesf2";
272 Names[RTLIB::OGE_F64] = "__gedf2";
273 Names[RTLIB::OLT_F32] = "__ltsf2";
274 Names[RTLIB::OLT_F64] = "__ltdf2";
275 Names[RTLIB::OLE_F32] = "__lesf2";
276 Names[RTLIB::OLE_F64] = "__ledf2";
277 Names[RTLIB::OGT_F32] = "__gtsf2";
278 Names[RTLIB::OGT_F64] = "__gtdf2";
279 Names[RTLIB::UO_F32] = "__unordsf2";
280 Names[RTLIB::UO_F64] = "__unorddf2";
281 Names[RTLIB::O_F32] = "__unordsf2";
282 Names[RTLIB::O_F64] = "__unorddf2";
283 Names[RTLIB::MEMCPY] = "memcpy";
284 Names[RTLIB::MEMMOVE] = "memmove";
285 Names[RTLIB::MEMSET] = "memset";
286 Names[RTLIB::UNWIND_RESUME] = "_Unwind_Resume";
287 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_1] = "__sync_val_compare_and_swap_1";
288 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_2] = "__sync_val_compare_and_swap_2";
289 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_4] = "__sync_val_compare_and_swap_4";
290 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_8] = "__sync_val_compare_and_swap_8";
291 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_1] = "__sync_lock_test_and_set_1";
292 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_2] = "__sync_lock_test_and_set_2";
293 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_4] = "__sync_lock_test_and_set_4";
294 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_8] = "__sync_lock_test_and_set_8";
295 Names[RTLIB::SYNC_FETCH_AND_ADD_1] = "__sync_fetch_and_add_1";
296 Names[RTLIB::SYNC_FETCH_AND_ADD_2] = "__sync_fetch_and_add_2";
297 Names[RTLIB::SYNC_FETCH_AND_ADD_4] = "__sync_fetch_and_add_4";
298 Names[RTLIB::SYNC_FETCH_AND_ADD_8] = "__sync_fetch_and_add_8";
299 Names[RTLIB::SYNC_FETCH_AND_SUB_1] = "__sync_fetch_and_sub_1";
300 Names[RTLIB::SYNC_FETCH_AND_SUB_2] = "__sync_fetch_and_sub_2";
301 Names[RTLIB::SYNC_FETCH_AND_SUB_4] = "__sync_fetch_and_sub_4";
302 Names[RTLIB::SYNC_FETCH_AND_SUB_8] = "__sync_fetch_and_sub_8";
303 Names[RTLIB::SYNC_FETCH_AND_AND_1] = "__sync_fetch_and_and_1";
304 Names[RTLIB::SYNC_FETCH_AND_AND_2] = "__sync_fetch_and_and_2";
305 Names[RTLIB::SYNC_FETCH_AND_AND_4] = "__sync_fetch_and_and_4";
306 Names[RTLIB::SYNC_FETCH_AND_AND_8] = "__sync_fetch_and_and_8";
307 Names[RTLIB::SYNC_FETCH_AND_OR_1] = "__sync_fetch_and_or_1";
308 Names[RTLIB::SYNC_FETCH_AND_OR_2] = "__sync_fetch_and_or_2";
309 Names[RTLIB::SYNC_FETCH_AND_OR_4] = "__sync_fetch_and_or_4";
310 Names[RTLIB::SYNC_FETCH_AND_OR_8] = "__sync_fetch_and_or_8";
311 Names[RTLIB::SYNC_FETCH_AND_XOR_1] = "__sync_fetch_and_xor_1";
312 Names[RTLIB::SYNC_FETCH_AND_XOR_2] = "__sync_fetch_and_xor_2";
313 Names[RTLIB::SYNC_FETCH_AND_XOR_4] = "__sync_fetch_and-xor_4";
314 Names[RTLIB::SYNC_FETCH_AND_XOR_8] = "__sync_fetch_and_xor_8";
315 Names[RTLIB::SYNC_FETCH_AND_NAND_1] = "__sync_fetch_and_nand_1";
316 Names[RTLIB::SYNC_FETCH_AND_NAND_2] = "__sync_fetch_and_nand_2";
317 Names[RTLIB::SYNC_FETCH_AND_NAND_4] = "__sync_fetch_and_nand_4";
318 Names[RTLIB::SYNC_FETCH_AND_NAND_8] = "__sync_fetch_and_nand_8";
321 /// InitLibcallCallingConvs - Set default libcall CallingConvs.
323 static void InitLibcallCallingConvs(CallingConv::ID *CCs) {
324 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
325 CCs[i] = CallingConv::C;
329 /// getFPEXT - Return the FPEXT_*_* value for the given types, or
330 /// UNKNOWN_LIBCALL if there is none.
331 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) {
332 if (OpVT == MVT::f32) {
333 if (RetVT == MVT::f64)
334 return FPEXT_F32_F64;
337 return UNKNOWN_LIBCALL;
340 /// getFPROUND - Return the FPROUND_*_* value for the given types, or
341 /// UNKNOWN_LIBCALL if there is none.
342 RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) {
343 if (RetVT == MVT::f32) {
344 if (OpVT == MVT::f64)
345 return FPROUND_F64_F32;
346 if (OpVT == MVT::f80)
347 return FPROUND_F80_F32;
348 if (OpVT == MVT::ppcf128)
349 return FPROUND_PPCF128_F32;
350 } else if (RetVT == MVT::f64) {
351 if (OpVT == MVT::f80)
352 return FPROUND_F80_F64;
353 if (OpVT == MVT::ppcf128)
354 return FPROUND_PPCF128_F64;
357 return UNKNOWN_LIBCALL;
360 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or
361 /// UNKNOWN_LIBCALL if there is none.
362 RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) {
363 if (OpVT == MVT::f32) {
364 if (RetVT == MVT::i8)
365 return FPTOSINT_F32_I8;
366 if (RetVT == MVT::i16)
367 return FPTOSINT_F32_I16;
368 if (RetVT == MVT::i32)
369 return FPTOSINT_F32_I32;
370 if (RetVT == MVT::i64)
371 return FPTOSINT_F32_I64;
372 if (RetVT == MVT::i128)
373 return FPTOSINT_F32_I128;
374 } else if (OpVT == MVT::f64) {
375 if (RetVT == MVT::i8)
376 return FPTOSINT_F64_I8;
377 if (RetVT == MVT::i16)
378 return FPTOSINT_F64_I16;
379 if (RetVT == MVT::i32)
380 return FPTOSINT_F64_I32;
381 if (RetVT == MVT::i64)
382 return FPTOSINT_F64_I64;
383 if (RetVT == MVT::i128)
384 return FPTOSINT_F64_I128;
385 } else if (OpVT == MVT::f80) {
386 if (RetVT == MVT::i32)
387 return FPTOSINT_F80_I32;
388 if (RetVT == MVT::i64)
389 return FPTOSINT_F80_I64;
390 if (RetVT == MVT::i128)
391 return FPTOSINT_F80_I128;
392 } else if (OpVT == MVT::ppcf128) {
393 if (RetVT == MVT::i32)
394 return FPTOSINT_PPCF128_I32;
395 if (RetVT == MVT::i64)
396 return FPTOSINT_PPCF128_I64;
397 if (RetVT == MVT::i128)
398 return FPTOSINT_PPCF128_I128;
400 return UNKNOWN_LIBCALL;
403 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or
404 /// UNKNOWN_LIBCALL if there is none.
405 RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) {
406 if (OpVT == MVT::f32) {
407 if (RetVT == MVT::i8)
408 return FPTOUINT_F32_I8;
409 if (RetVT == MVT::i16)
410 return FPTOUINT_F32_I16;
411 if (RetVT == MVT::i32)
412 return FPTOUINT_F32_I32;
413 if (RetVT == MVT::i64)
414 return FPTOUINT_F32_I64;
415 if (RetVT == MVT::i128)
416 return FPTOUINT_F32_I128;
417 } else if (OpVT == MVT::f64) {
418 if (RetVT == MVT::i8)
419 return FPTOUINT_F64_I8;
420 if (RetVT == MVT::i16)
421 return FPTOUINT_F64_I16;
422 if (RetVT == MVT::i32)
423 return FPTOUINT_F64_I32;
424 if (RetVT == MVT::i64)
425 return FPTOUINT_F64_I64;
426 if (RetVT == MVT::i128)
427 return FPTOUINT_F64_I128;
428 } else if (OpVT == MVT::f80) {
429 if (RetVT == MVT::i32)
430 return FPTOUINT_F80_I32;
431 if (RetVT == MVT::i64)
432 return FPTOUINT_F80_I64;
433 if (RetVT == MVT::i128)
434 return FPTOUINT_F80_I128;
435 } else if (OpVT == MVT::ppcf128) {
436 if (RetVT == MVT::i32)
437 return FPTOUINT_PPCF128_I32;
438 if (RetVT == MVT::i64)
439 return FPTOUINT_PPCF128_I64;
440 if (RetVT == MVT::i128)
441 return FPTOUINT_PPCF128_I128;
443 return UNKNOWN_LIBCALL;
446 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or
447 /// UNKNOWN_LIBCALL if there is none.
448 RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) {
449 if (OpVT == MVT::i32) {
450 if (RetVT == MVT::f32)
451 return SINTTOFP_I32_F32;
452 else if (RetVT == MVT::f64)
453 return SINTTOFP_I32_F64;
454 else if (RetVT == MVT::f80)
455 return SINTTOFP_I32_F80;
456 else if (RetVT == MVT::ppcf128)
457 return SINTTOFP_I32_PPCF128;
458 } else if (OpVT == MVT::i64) {
459 if (RetVT == MVT::f32)
460 return SINTTOFP_I64_F32;
461 else if (RetVT == MVT::f64)
462 return SINTTOFP_I64_F64;
463 else if (RetVT == MVT::f80)
464 return SINTTOFP_I64_F80;
465 else if (RetVT == MVT::ppcf128)
466 return SINTTOFP_I64_PPCF128;
467 } else if (OpVT == MVT::i128) {
468 if (RetVT == MVT::f32)
469 return SINTTOFP_I128_F32;
470 else if (RetVT == MVT::f64)
471 return SINTTOFP_I128_F64;
472 else if (RetVT == MVT::f80)
473 return SINTTOFP_I128_F80;
474 else if (RetVT == MVT::ppcf128)
475 return SINTTOFP_I128_PPCF128;
477 return UNKNOWN_LIBCALL;
480 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or
481 /// UNKNOWN_LIBCALL if there is none.
482 RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) {
483 if (OpVT == MVT::i32) {
484 if (RetVT == MVT::f32)
485 return UINTTOFP_I32_F32;
486 else if (RetVT == MVT::f64)
487 return UINTTOFP_I32_F64;
488 else if (RetVT == MVT::f80)
489 return UINTTOFP_I32_F80;
490 else if (RetVT == MVT::ppcf128)
491 return UINTTOFP_I32_PPCF128;
492 } else if (OpVT == MVT::i64) {
493 if (RetVT == MVT::f32)
494 return UINTTOFP_I64_F32;
495 else if (RetVT == MVT::f64)
496 return UINTTOFP_I64_F64;
497 else if (RetVT == MVT::f80)
498 return UINTTOFP_I64_F80;
499 else if (RetVT == MVT::ppcf128)
500 return UINTTOFP_I64_PPCF128;
501 } else if (OpVT == MVT::i128) {
502 if (RetVT == MVT::f32)
503 return UINTTOFP_I128_F32;
504 else if (RetVT == MVT::f64)
505 return UINTTOFP_I128_F64;
506 else if (RetVT == MVT::f80)
507 return UINTTOFP_I128_F80;
508 else if (RetVT == MVT::ppcf128)
509 return UINTTOFP_I128_PPCF128;
511 return UNKNOWN_LIBCALL;
514 /// InitCmpLibcallCCs - Set default comparison libcall CC.
516 static void InitCmpLibcallCCs(ISD::CondCode *CCs) {
517 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL);
518 CCs[RTLIB::OEQ_F32] = ISD::SETEQ;
519 CCs[RTLIB::OEQ_F64] = ISD::SETEQ;
520 CCs[RTLIB::UNE_F32] = ISD::SETNE;
521 CCs[RTLIB::UNE_F64] = ISD::SETNE;
522 CCs[RTLIB::OGE_F32] = ISD::SETGE;
523 CCs[RTLIB::OGE_F64] = ISD::SETGE;
524 CCs[RTLIB::OLT_F32] = ISD::SETLT;
525 CCs[RTLIB::OLT_F64] = ISD::SETLT;
526 CCs[RTLIB::OLE_F32] = ISD::SETLE;
527 CCs[RTLIB::OLE_F64] = ISD::SETLE;
528 CCs[RTLIB::OGT_F32] = ISD::SETGT;
529 CCs[RTLIB::OGT_F64] = ISD::SETGT;
530 CCs[RTLIB::UO_F32] = ISD::SETNE;
531 CCs[RTLIB::UO_F64] = ISD::SETNE;
532 CCs[RTLIB::O_F32] = ISD::SETEQ;
533 CCs[RTLIB::O_F64] = ISD::SETEQ;
536 /// NOTE: The constructor takes ownership of TLOF.
537 TargetLowering::TargetLowering(const TargetMachine &tm,
538 const TargetLoweringObjectFile *tlof)
539 : TM(tm), TD(TM.getTargetData()), TLOF(*tlof),
540 mayPromoteElements(AllowPromoteIntElem) {
541 // All operations default to being supported.
542 memset(OpActions, 0, sizeof(OpActions));
543 memset(LoadExtActions, 0, sizeof(LoadExtActions));
544 memset(TruncStoreActions, 0, sizeof(TruncStoreActions));
545 memset(IndexedModeActions, 0, sizeof(IndexedModeActions));
546 memset(CondCodeActions, 0, sizeof(CondCodeActions));
548 // Set default actions for various operations.
549 for (unsigned VT = 0; VT != (unsigned)MVT::LAST_VALUETYPE; ++VT) {
550 // Default all indexed load / store to expand.
551 for (unsigned IM = (unsigned)ISD::PRE_INC;
552 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
553 setIndexedLoadAction(IM, (MVT::SimpleValueType)VT, Expand);
554 setIndexedStoreAction(IM, (MVT::SimpleValueType)VT, Expand);
557 // These operations default to expand.
558 setOperationAction(ISD::FGETSIGN, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::CONCAT_VECTORS, (MVT::SimpleValueType)VT, Expand);
562 // Most targets ignore the @llvm.prefetch intrinsic.
563 setOperationAction(ISD::PREFETCH, MVT::Other, Expand);
565 // ConstantFP nodes default to expand. Targets can either change this to
566 // Legal, in which case all fp constants are legal, or use isFPImmLegal()
567 // to optimize expansions for certain constants.
568 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
569 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
570 setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
572 // These library functions default to expand.
573 setOperationAction(ISD::FLOG , MVT::f64, Expand);
574 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
575 setOperationAction(ISD::FLOG10,MVT::f64, Expand);
576 setOperationAction(ISD::FEXP , MVT::f64, Expand);
577 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
578 setOperationAction(ISD::FLOG , MVT::f32, Expand);
579 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
580 setOperationAction(ISD::FLOG10,MVT::f32, Expand);
581 setOperationAction(ISD::FEXP , MVT::f32, Expand);
582 setOperationAction(ISD::FEXP2, MVT::f32, Expand);
584 // Default ISD::TRAP to expand (which turns it into abort).
585 setOperationAction(ISD::TRAP, MVT::Other, Expand);
587 IsLittleEndian = TD->isLittleEndian();
588 PointerTy = MVT::getIntegerVT(8*TD->getPointerSize());
589 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE*sizeof(TargetRegisterClass*));
590 memset(TargetDAGCombineArray, 0, array_lengthof(TargetDAGCombineArray));
591 maxStoresPerMemset = maxStoresPerMemcpy = maxStoresPerMemmove = 8;
592 maxStoresPerMemsetOptSize = maxStoresPerMemcpyOptSize
593 = maxStoresPerMemmoveOptSize = 4;
594 benefitFromCodePlacementOpt = false;
595 UseUnderscoreSetJmp = false;
596 UseUnderscoreLongJmp = false;
597 SelectIsExpensive = false;
598 IntDivIsCheap = false;
599 Pow2DivIsCheap = false;
600 JumpIsExpensive = false;
601 StackPointerRegisterToSaveRestore = 0;
602 ExceptionPointerRegister = 0;
603 ExceptionSelectorRegister = 0;
604 BooleanContents = UndefinedBooleanContent;
605 SchedPreferenceInfo = Sched::Latency;
607 JumpBufAlignment = 0;
608 MinFunctionAlignment = 0;
609 PrefFunctionAlignment = 0;
610 PrefLoopAlignment = 0;
611 MinStackArgumentAlignment = 1;
612 ShouldFoldAtomicFences = false;
614 InitLibcallNames(LibcallRoutineNames);
615 InitCmpLibcallCCs(CmpLibcallCCs);
616 InitLibcallCallingConvs(LibcallCallingConvs);
619 TargetLowering::~TargetLowering() {
623 MVT TargetLowering::getShiftAmountTy(EVT LHSTy) const {
624 return MVT::getIntegerVT(8*TD->getPointerSize());
627 /// canOpTrap - Returns true if the operation can trap for the value type.
628 /// VT must be a legal type.
629 bool TargetLowering::canOpTrap(unsigned Op, EVT VT) const {
630 assert(isTypeLegal(VT));
645 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT,
646 unsigned &NumIntermediates,
648 TargetLowering *TLI) {
649 // Figure out the right, legal destination reg to copy into.
650 unsigned NumElts = VT.getVectorNumElements();
651 MVT EltTy = VT.getVectorElementType();
653 unsigned NumVectorRegs = 1;
655 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
656 // could break down into LHS/RHS like LegalizeDAG does.
657 if (!isPowerOf2_32(NumElts)) {
658 NumVectorRegs = NumElts;
662 // Divide the input until we get to a supported size. This will always
663 // end with a scalar if the target doesn't support vectors.
664 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) {
669 NumIntermediates = NumVectorRegs;
671 MVT NewVT = MVT::getVectorVT(EltTy, NumElts);
672 if (!TLI->isTypeLegal(NewVT))
674 IntermediateVT = NewVT;
676 EVT DestVT = TLI->getRegisterType(NewVT);
678 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16.
679 return NumVectorRegs*(NewVT.getSizeInBits()/DestVT.getSizeInBits());
681 // Otherwise, promotion or legal types use the same number of registers as
682 // the vector decimated to the appropriate level.
683 return NumVectorRegs;
686 /// isLegalRC - Return true if the value types that can be represented by the
687 /// specified register class are all legal.
688 bool TargetLowering::isLegalRC(const TargetRegisterClass *RC) const {
689 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
697 /// hasLegalSuperRegRegClasses - Return true if the specified register class
698 /// has one or more super-reg register classes that are legal.
700 TargetLowering::hasLegalSuperRegRegClasses(const TargetRegisterClass *RC) const{
701 if (*RC->superregclasses_begin() == 0)
703 for (TargetRegisterInfo::regclass_iterator I = RC->superregclasses_begin(),
704 E = RC->superregclasses_end(); I != E; ++I) {
705 const TargetRegisterClass *RRC = *I;
712 /// findRepresentativeClass - Return the largest legal super-reg register class
713 /// of the register class for the specified type and its associated "cost".
714 std::pair<const TargetRegisterClass*, uint8_t>
715 TargetLowering::findRepresentativeClass(EVT VT) const {
716 const TargetRegisterClass *RC = RegClassForVT[VT.getSimpleVT().SimpleTy];
718 return std::make_pair(RC, 0);
719 const TargetRegisterClass *BestRC = RC;
720 for (TargetRegisterInfo::regclass_iterator I = RC->superregclasses_begin(),
721 E = RC->superregclasses_end(); I != E; ++I) {
722 const TargetRegisterClass *RRC = *I;
723 if (RRC->isASubClass() || !isLegalRC(RRC))
725 if (!hasLegalSuperRegRegClasses(RRC))
726 return std::make_pair(RRC, 1);
729 return std::make_pair(BestRC, 1);
733 /// computeRegisterProperties - Once all of the register classes are added,
734 /// this allows us to compute derived properties we expose.
735 void TargetLowering::computeRegisterProperties() {
736 assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE &&
737 "Too many value types for ValueTypeActions to hold!");
739 // Everything defaults to needing one register.
740 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
741 NumRegistersForVT[i] = 1;
742 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i;
744 // ...except isVoid, which doesn't need any registers.
745 NumRegistersForVT[MVT::isVoid] = 0;
747 // Find the largest integer register class.
748 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE;
749 for (; RegClassForVT[LargestIntReg] == 0; --LargestIntReg)
750 assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
752 // Every integer value type larger than this largest register takes twice as
753 // many registers to represent as the previous ValueType.
754 for (unsigned ExpandedReg = LargestIntReg + 1; ; ++ExpandedReg) {
755 EVT ExpandedVT = (MVT::SimpleValueType)ExpandedReg;
756 if (!ExpandedVT.isInteger())
758 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
759 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg;
760 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1);
761 ValueTypeActions.setTypeAction(ExpandedVT, TypeExpandInteger);
764 // Inspect all of the ValueType's smaller than the largest integer
765 // register to see which ones need promotion.
766 unsigned LegalIntReg = LargestIntReg;
767 for (unsigned IntReg = LargestIntReg - 1;
768 IntReg >= (unsigned)MVT::i1; --IntReg) {
769 EVT IVT = (MVT::SimpleValueType)IntReg;
770 if (isTypeLegal(IVT)) {
771 LegalIntReg = IntReg;
773 RegisterTypeForVT[IntReg] = TransformToType[IntReg] =
774 (MVT::SimpleValueType)LegalIntReg;
775 ValueTypeActions.setTypeAction(IVT, TypePromoteInteger);
779 // ppcf128 type is really two f64's.
780 if (!isTypeLegal(MVT::ppcf128)) {
781 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64];
782 RegisterTypeForVT[MVT::ppcf128] = MVT::f64;
783 TransformToType[MVT::ppcf128] = MVT::f64;
784 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat);
787 // Decide how to handle f64. If the target does not have native f64 support,
788 // expand it to i64 and we will be generating soft float library calls.
789 if (!isTypeLegal(MVT::f64)) {
790 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
791 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
792 TransformToType[MVT::f64] = MVT::i64;
793 ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat);
796 // Decide how to handle f32. If the target does not have native support for
797 // f32, promote it to f64 if it is legal. Otherwise, expand it to i32.
798 if (!isTypeLegal(MVT::f32)) {
799 if (isTypeLegal(MVT::f64)) {
800 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::f64];
801 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::f64];
802 TransformToType[MVT::f32] = MVT::f64;
803 ValueTypeActions.setTypeAction(MVT::f32, TypePromoteInteger);
805 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
806 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
807 TransformToType[MVT::f32] = MVT::i32;
808 ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat);
812 // Loop over all of the vector value types to see which need transformations.
813 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
814 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
815 MVT VT = (MVT::SimpleValueType)i;
816 if (isTypeLegal(VT)) continue;
818 // Determine if there is a legal wider type. If so, we should promote to
819 // that wider vector type.
820 EVT EltVT = VT.getVectorElementType();
821 unsigned NElts = VT.getVectorNumElements();
823 bool IsLegalWiderType = false;
824 for (unsigned nVT = i+1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
825 EVT SVT = (MVT::SimpleValueType)nVT;
827 // If we allow the promotion of vector elements using a flag,
828 // then return TypePromoteInteger on vector elements.
829 if (mayPromoteElements) {
830 // Promote vectors of integers to vectors with the same number
831 // of elements, with a wider element type.
832 if (SVT.getVectorElementType().getSizeInBits() > EltVT.getSizeInBits()
833 && SVT.getVectorNumElements() == NElts &&
834 isTypeLegal(SVT) && SVT.getScalarType().isInteger()) {
835 TransformToType[i] = SVT;
836 RegisterTypeForVT[i] = SVT;
837 NumRegistersForVT[i] = 1;
838 ValueTypeActions.setTypeAction(VT, TypePromoteInteger);
839 IsLegalWiderType = true;
844 if (SVT.getVectorElementType() == EltVT &&
845 SVT.getVectorNumElements() > NElts &&
847 TransformToType[i] = SVT;
848 RegisterTypeForVT[i] = SVT;
849 NumRegistersForVT[i] = 1;
850 ValueTypeActions.setTypeAction(VT, TypeWidenVector);
851 IsLegalWiderType = true;
855 if (IsLegalWiderType) continue;
860 unsigned NumIntermediates;
861 NumRegistersForVT[i] =
862 getVectorTypeBreakdownMVT(VT, IntermediateVT, NumIntermediates,
864 RegisterTypeForVT[i] = RegisterVT;
866 EVT NVT = VT.getPow2VectorType();
868 // Type is already a power of 2. The default action is to split.
869 TransformToType[i] = MVT::Other;
870 unsigned NumElts = VT.getVectorNumElements();
871 ValueTypeActions.setTypeAction(VT,
872 NumElts > 1 ? TypeSplitVector : TypeScalarizeVector);
874 TransformToType[i] = NVT;
875 ValueTypeActions.setTypeAction(VT, TypeWidenVector);
879 // Determine the 'representative' register class for each value type.
880 // An representative register class is the largest (meaning one which is
881 // not a sub-register class / subreg register class) legal register class for
882 // a group of value types. For example, on i386, i8, i16, and i32
883 // representative would be GR32; while on x86_64 it's GR64.
884 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
885 const TargetRegisterClass* RRC;
887 tie(RRC, Cost) = findRepresentativeClass((MVT::SimpleValueType)i);
888 RepRegClassForVT[i] = RRC;
889 RepRegClassCostForVT[i] = Cost;
893 const char *TargetLowering::getTargetNodeName(unsigned Opcode) const {
898 MVT::SimpleValueType TargetLowering::getSetCCResultType(EVT VT) const {
899 return PointerTy.SimpleTy;
902 MVT::SimpleValueType TargetLowering::getCmpLibcallReturnType() const {
903 return MVT::i32; // return the default value
906 /// getVectorTypeBreakdown - Vector types are broken down into some number of
907 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
908 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
909 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
911 /// This method returns the number of registers needed, and the VT for each
912 /// register. It also returns the VT and quantity of the intermediate values
913 /// before they are promoted/expanded.
915 unsigned TargetLowering::getVectorTypeBreakdown(LLVMContext &Context, EVT VT,
917 unsigned &NumIntermediates,
918 EVT &RegisterVT) const {
919 unsigned NumElts = VT.getVectorNumElements();
921 // If there is a wider vector type with the same element type as this one,
922 // we should widen to that legal vector type. This handles things like
923 // <2 x float> -> <4 x float>.
924 if (NumElts != 1 && getTypeAction(Context, VT) == TypeWidenVector) {
925 RegisterVT = getTypeToTransformTo(Context, VT);
926 if (isTypeLegal(RegisterVT)) {
927 IntermediateVT = RegisterVT;
928 NumIntermediates = 1;
933 // Figure out the right, legal destination reg to copy into.
934 EVT EltTy = VT.getVectorElementType();
936 unsigned NumVectorRegs = 1;
938 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
939 // could break down into LHS/RHS like LegalizeDAG does.
940 if (!isPowerOf2_32(NumElts)) {
941 NumVectorRegs = NumElts;
945 // Divide the input until we get to a supported size. This will always
946 // end with a scalar if the target doesn't support vectors.
947 while (NumElts > 1 && !isTypeLegal(
948 EVT::getVectorVT(Context, EltTy, NumElts))) {
953 NumIntermediates = NumVectorRegs;
955 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts);
956 if (!isTypeLegal(NewVT))
958 IntermediateVT = NewVT;
960 EVT DestVT = getRegisterType(Context, NewVT);
962 if (DestVT.bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16.
963 return NumVectorRegs*(NewVT.getSizeInBits()/DestVT.getSizeInBits());
965 // Otherwise, promotion or legal types use the same number of registers as
966 // the vector decimated to the appropriate level.
967 return NumVectorRegs;
970 /// Get the EVTs and ArgFlags collections that represent the legalized return
971 /// type of the given function. This does not require a DAG or a return value,
972 /// and is suitable for use before any DAGs for the function are constructed.
973 /// TODO: Move this out of TargetLowering.cpp.
974 void llvm::GetReturnInfo(const Type* ReturnType, Attributes attr,
975 SmallVectorImpl<ISD::OutputArg> &Outs,
976 const TargetLowering &TLI,
977 SmallVectorImpl<uint64_t> *Offsets) {
978 SmallVector<EVT, 4> ValueVTs;
979 ComputeValueVTs(TLI, ReturnType, ValueVTs);
980 unsigned NumValues = ValueVTs.size();
981 if (NumValues == 0) return;
984 for (unsigned j = 0, f = NumValues; j != f; ++j) {
985 EVT VT = ValueVTs[j];
986 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
988 if (attr & Attribute::SExt)
989 ExtendKind = ISD::SIGN_EXTEND;
990 else if (attr & Attribute::ZExt)
991 ExtendKind = ISD::ZERO_EXTEND;
993 // FIXME: C calling convention requires the return type to be promoted to
994 // at least 32-bit. But this is not necessary for non-C calling
995 // conventions. The frontend should mark functions whose return values
996 // require promoting with signext or zeroext attributes.
997 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) {
998 EVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32);
999 if (VT.bitsLT(MinVT))
1003 unsigned NumParts = TLI.getNumRegisters(ReturnType->getContext(), VT);
1004 EVT PartVT = TLI.getRegisterType(ReturnType->getContext(), VT);
1005 unsigned PartSize = TLI.getTargetData()->getTypeAllocSize(
1006 PartVT.getTypeForEVT(ReturnType->getContext()));
1008 // 'inreg' on function refers to return value
1009 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1010 if (attr & Attribute::InReg)
1013 // Propagate extension type if any
1014 if (attr & Attribute::SExt)
1016 else if (attr & Attribute::ZExt)
1019 for (unsigned i = 0; i < NumParts; ++i) {
1020 Outs.push_back(ISD::OutputArg(Flags, PartVT, /*isFixed=*/true));
1022 Offsets->push_back(Offset);
1029 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1030 /// function arguments in the caller parameter area. This is the actual
1031 /// alignment, not its logarithm.
1032 unsigned TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1033 return TD->getCallFrameTypeAlignment(Ty);
1036 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1037 /// current function. The returned value is a member of the
1038 /// MachineJumpTableInfo::JTEntryKind enum.
1039 unsigned TargetLowering::getJumpTableEncoding() const {
1040 // In non-pic modes, just use the address of a block.
1041 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
1042 return MachineJumpTableInfo::EK_BlockAddress;
1044 // In PIC mode, if the target supports a GPRel32 directive, use it.
1045 if (getTargetMachine().getMCAsmInfo()->getGPRel32Directive() != 0)
1046 return MachineJumpTableInfo::EK_GPRel32BlockAddress;
1048 // Otherwise, use a label difference.
1049 return MachineJumpTableInfo::EK_LabelDifference32;
1052 SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1053 SelectionDAG &DAG) const {
1054 // If our PIC model is GP relative, use the global offset table as the base.
1055 if (getJumpTableEncoding() == MachineJumpTableInfo::EK_GPRel32BlockAddress)
1056 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy());
1060 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1061 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1064 TargetLowering::getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
1065 unsigned JTI,MCContext &Ctx) const{
1066 // The normal PIC reloc base is the label at the start of the jump table.
1067 return MCSymbolRefExpr::Create(MF->getJTISymbol(JTI, Ctx), Ctx);
1071 TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
1072 // Assume that everything is safe in static mode.
1073 if (getTargetMachine().getRelocationModel() == Reloc::Static)
1076 // In dynamic-no-pic mode, assume that known defined values are safe.
1077 if (getTargetMachine().getRelocationModel() == Reloc::DynamicNoPIC &&
1079 !GA->getGlobal()->isDeclaration() &&
1080 !GA->getGlobal()->isWeakForLinker())
1083 // Otherwise assume nothing is safe.
1087 //===----------------------------------------------------------------------===//
1088 // Optimization Methods
1089 //===----------------------------------------------------------------------===//
1091 /// ShrinkDemandedConstant - Check to see if the specified operand of the
1092 /// specified instruction is a constant integer. If so, check to see if there
1093 /// are any bits set in the constant that are not demanded. If so, shrink the
1094 /// constant and return true.
1095 bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDValue Op,
1096 const APInt &Demanded) {
1097 DebugLoc dl = Op.getDebugLoc();
1099 // FIXME: ISD::SELECT, ISD::SELECT_CC
1100 switch (Op.getOpcode()) {
1105 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
1106 if (!C) return false;
1108 if (Op.getOpcode() == ISD::XOR &&
1109 (C->getAPIntValue() | (~Demanded)).isAllOnesValue())
1112 // if we can expand it to have all bits set, do it
1113 if (C->getAPIntValue().intersects(~Demanded)) {
1114 EVT VT = Op.getValueType();
1115 SDValue New = DAG.getNode(Op.getOpcode(), dl, VT, Op.getOperand(0),
1116 DAG.getConstant(Demanded &
1119 return CombineTo(Op, New);
1129 /// ShrinkDemandedOp - Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the
1130 /// casts are free. This uses isZExtFree and ZERO_EXTEND for the widening
1131 /// cast, but it could be generalized for targets with other types of
1132 /// implicit widening casts.
1134 TargetLowering::TargetLoweringOpt::ShrinkDemandedOp(SDValue Op,
1136 const APInt &Demanded,
1138 assert(Op.getNumOperands() == 2 &&
1139 "ShrinkDemandedOp only supports binary operators!");
1140 assert(Op.getNode()->getNumValues() == 1 &&
1141 "ShrinkDemandedOp only supports nodes with one result!");
1143 // Don't do this if the node has another user, which may require the
1145 if (!Op.getNode()->hasOneUse())
1148 // Search for the smallest integer type with free casts to and from
1149 // Op's type. For expedience, just check power-of-2 integer types.
1150 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1151 unsigned SmallVTBits = BitWidth - Demanded.countLeadingZeros();
1152 if (!isPowerOf2_32(SmallVTBits))
1153 SmallVTBits = NextPowerOf2(SmallVTBits);
1154 for (; SmallVTBits < BitWidth; SmallVTBits = NextPowerOf2(SmallVTBits)) {
1155 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), SmallVTBits);
1156 if (TLI.isTruncateFree(Op.getValueType(), SmallVT) &&
1157 TLI.isZExtFree(SmallVT, Op.getValueType())) {
1158 // We found a type with free casts.
1159 SDValue X = DAG.getNode(Op.getOpcode(), dl, SmallVT,
1160 DAG.getNode(ISD::TRUNCATE, dl, SmallVT,
1161 Op.getNode()->getOperand(0)),
1162 DAG.getNode(ISD::TRUNCATE, dl, SmallVT,
1163 Op.getNode()->getOperand(1)));
1164 SDValue Z = DAG.getNode(ISD::ZERO_EXTEND, dl, Op.getValueType(), X);
1165 return CombineTo(Op, Z);
1171 /// SimplifyDemandedBits - Look at Op. At this point, we know that only the
1172 /// DemandedMask bits of the result of Op are ever used downstream. If we can
1173 /// use this information to simplify Op, create a new simplified DAG node and
1174 /// return true, returning the original and new nodes in Old and New. Otherwise,
1175 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
1176 /// the expression (used to simplify the caller). The KnownZero/One bits may
1177 /// only be accurate for those bits in the DemandedMask.
1178 bool TargetLowering::SimplifyDemandedBits(SDValue Op,
1179 const APInt &DemandedMask,
1182 TargetLoweringOpt &TLO,
1183 unsigned Depth) const {
1184 unsigned BitWidth = DemandedMask.getBitWidth();
1185 assert(Op.getValueType().getScalarType().getSizeInBits() == BitWidth &&
1186 "Mask size mismatches value type size!");
1187 APInt NewMask = DemandedMask;
1188 DebugLoc dl = Op.getDebugLoc();
1190 // Don't know anything.
1191 KnownZero = KnownOne = APInt(BitWidth, 0);
1193 // Other users may use these bits.
1194 if (!Op.getNode()->hasOneUse()) {
1196 // If not at the root, Just compute the KnownZero/KnownOne bits to
1197 // simplify things downstream.
1198 TLO.DAG.ComputeMaskedBits(Op, DemandedMask, KnownZero, KnownOne, Depth);
1201 // If this is the root being simplified, allow it to have multiple uses,
1202 // just set the NewMask to all bits.
1203 NewMask = APInt::getAllOnesValue(BitWidth);
1204 } else if (DemandedMask == 0) {
1205 // Not demanding any bits from Op.
1206 if (Op.getOpcode() != ISD::UNDEF)
1207 return TLO.CombineTo(Op, TLO.DAG.getUNDEF(Op.getValueType()));
1209 } else if (Depth == 6) { // Limit search depth.
1213 APInt KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut;
1214 switch (Op.getOpcode()) {
1216 // We know all of the bits for a constant!
1217 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue() & NewMask;
1218 KnownZero = ~KnownOne & NewMask;
1219 return false; // Don't fall through, will infinitely loop.
1221 // If the RHS is a constant, check to see if the LHS would be zero without
1222 // using the bits from the RHS. Below, we use knowledge about the RHS to
1223 // simplify the LHS, here we're using information from the LHS to simplify
1225 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1226 APInt LHSZero, LHSOne;
1227 // Do not increment Depth here; that can cause an infinite loop.
1228 TLO.DAG.ComputeMaskedBits(Op.getOperand(0), NewMask,
1229 LHSZero, LHSOne, Depth);
1230 // If the LHS already has zeros where RHSC does, this and is dead.
1231 if ((LHSZero & NewMask) == (~RHSC->getAPIntValue() & NewMask))
1232 return TLO.CombineTo(Op, Op.getOperand(0));
1233 // If any of the set bits in the RHS are known zero on the LHS, shrink
1235 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & NewMask))
1239 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
1240 KnownOne, TLO, Depth+1))
1242 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1243 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownZero & NewMask,
1244 KnownZero2, KnownOne2, TLO, Depth+1))
1246 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1248 // If all of the demanded bits are known one on one side, return the other.
1249 // These bits cannot contribute to the result of the 'and'.
1250 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
1251 return TLO.CombineTo(Op, Op.getOperand(0));
1252 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
1253 return TLO.CombineTo(Op, Op.getOperand(1));
1254 // If all of the demanded bits in the inputs are known zeros, return zero.
1255 if ((NewMask & (KnownZero|KnownZero2)) == NewMask)
1256 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, Op.getValueType()));
1257 // If the RHS is a constant, see if we can simplify it.
1258 if (TLO.ShrinkDemandedConstant(Op, ~KnownZero2 & NewMask))
1260 // If the operation can be done in a smaller type, do so.
1261 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1264 // Output known-1 bits are only known if set in both the LHS & RHS.
1265 KnownOne &= KnownOne2;
1266 // Output known-0 are known to be clear if zero in either the LHS | RHS.
1267 KnownZero |= KnownZero2;
1270 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
1271 KnownOne, TLO, Depth+1))
1273 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1274 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownOne & NewMask,
1275 KnownZero2, KnownOne2, TLO, Depth+1))
1277 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1279 // If all of the demanded bits are known zero on one side, return the other.
1280 // These bits cannot contribute to the result of the 'or'.
1281 if ((NewMask & ~KnownOne2 & KnownZero) == (~KnownOne2 & NewMask))
1282 return TLO.CombineTo(Op, Op.getOperand(0));
1283 if ((NewMask & ~KnownOne & KnownZero2) == (~KnownOne & NewMask))
1284 return TLO.CombineTo(Op, Op.getOperand(1));
1285 // If all of the potentially set bits on one side are known to be set on
1286 // the other side, just use the 'other' side.
1287 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
1288 return TLO.CombineTo(Op, Op.getOperand(0));
1289 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
1290 return TLO.CombineTo(Op, Op.getOperand(1));
1291 // If the RHS is a constant, see if we can simplify it.
1292 if (TLO.ShrinkDemandedConstant(Op, NewMask))
1294 // If the operation can be done in a smaller type, do so.
1295 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1298 // Output known-0 bits are only known if clear in both the LHS & RHS.
1299 KnownZero &= KnownZero2;
1300 // Output known-1 are known to be set if set in either the LHS | RHS.
1301 KnownOne |= KnownOne2;
1304 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
1305 KnownOne, TLO, Depth+1))
1307 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1308 if (SimplifyDemandedBits(Op.getOperand(0), NewMask, KnownZero2,
1309 KnownOne2, TLO, Depth+1))
1311 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1313 // If all of the demanded bits are known zero on one side, return the other.
1314 // These bits cannot contribute to the result of the 'xor'.
1315 if ((KnownZero & NewMask) == NewMask)
1316 return TLO.CombineTo(Op, Op.getOperand(0));
1317 if ((KnownZero2 & NewMask) == NewMask)
1318 return TLO.CombineTo(Op, Op.getOperand(1));
1319 // If the operation can be done in a smaller type, do so.
1320 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1323 // If all of the unknown bits are known to be zero on one side or the other
1324 // (but not both) turn this into an *inclusive* or.
1325 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1326 if ((NewMask & ~KnownZero & ~KnownZero2) == 0)
1327 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, dl, Op.getValueType(),
1331 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1332 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
1333 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1334 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
1336 // If all of the demanded bits on one side are known, and all of the set
1337 // bits on that side are also known to be set on the other side, turn this
1338 // into an AND, as we know the bits will be cleared.
1339 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1340 if ((NewMask & (KnownZero|KnownOne)) == NewMask) { // all known
1341 if ((KnownOne & KnownOne2) == KnownOne) {
1342 EVT VT = Op.getValueType();
1343 SDValue ANDC = TLO.DAG.getConstant(~KnownOne & NewMask, VT);
1344 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, dl, VT,
1345 Op.getOperand(0), ANDC));
1349 // If the RHS is a constant, see if we can simplify it.
1350 // for XOR, we prefer to force bits to 1 if they will make a -1.
1351 // if we can't force bits, try to shrink constant
1352 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1353 APInt Expanded = C->getAPIntValue() | (~NewMask);
1354 // if we can expand it to have all bits set, do it
1355 if (Expanded.isAllOnesValue()) {
1356 if (Expanded != C->getAPIntValue()) {
1357 EVT VT = Op.getValueType();
1358 SDValue New = TLO.DAG.getNode(Op.getOpcode(), dl,VT, Op.getOperand(0),
1359 TLO.DAG.getConstant(Expanded, VT));
1360 return TLO.CombineTo(Op, New);
1362 // if it already has all the bits set, nothing to change
1363 // but don't shrink either!
1364 } else if (TLO.ShrinkDemandedConstant(Op, NewMask)) {
1369 KnownZero = KnownZeroOut;
1370 KnownOne = KnownOneOut;
1373 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero,
1374 KnownOne, TLO, Depth+1))
1376 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero2,
1377 KnownOne2, TLO, Depth+1))
1379 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1380 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1382 // If the operands are constants, see if we can simplify them.
1383 if (TLO.ShrinkDemandedConstant(Op, NewMask))
1386 // Only known if known in both the LHS and RHS.
1387 KnownOne &= KnownOne2;
1388 KnownZero &= KnownZero2;
1390 case ISD::SELECT_CC:
1391 if (SimplifyDemandedBits(Op.getOperand(3), NewMask, KnownZero,
1392 KnownOne, TLO, Depth+1))
1394 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero2,
1395 KnownOne2, TLO, Depth+1))
1397 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1398 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1400 // If the operands are constants, see if we can simplify them.
1401 if (TLO.ShrinkDemandedConstant(Op, NewMask))
1404 // Only known if known in both the LHS and RHS.
1405 KnownOne &= KnownOne2;
1406 KnownZero &= KnownZero2;
1409 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1410 unsigned ShAmt = SA->getZExtValue();
1411 SDValue InOp = Op.getOperand(0);
1413 // If the shift count is an invalid immediate, don't do anything.
1414 if (ShAmt >= BitWidth)
1417 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a
1418 // single shift. We can do this if the bottom bits (which are shifted
1419 // out) are never demanded.
1420 if (InOp.getOpcode() == ISD::SRL &&
1421 isa<ConstantSDNode>(InOp.getOperand(1))) {
1422 if (ShAmt && (NewMask & APInt::getLowBitsSet(BitWidth, ShAmt)) == 0) {
1423 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
1424 unsigned Opc = ISD::SHL;
1425 int Diff = ShAmt-C1;
1432 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
1433 EVT VT = Op.getValueType();
1434 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT,
1435 InOp.getOperand(0), NewSA));
1439 if (SimplifyDemandedBits(InOp, NewMask.lshr(ShAmt),
1440 KnownZero, KnownOne, TLO, Depth+1))
1443 // Convert (shl (anyext x, c)) to (anyext (shl x, c)) if the high bits
1444 // are not demanded. This will likely allow the anyext to be folded away.
1445 if (InOp.getNode()->getOpcode() == ISD::ANY_EXTEND) {
1446 SDValue InnerOp = InOp.getNode()->getOperand(0);
1447 EVT InnerVT = InnerOp.getValueType();
1448 if ((APInt::getHighBitsSet(BitWidth,
1449 BitWidth - InnerVT.getSizeInBits()) &
1450 DemandedMask) == 0 &&
1451 isTypeDesirableForOp(ISD::SHL, InnerVT)) {
1452 EVT ShTy = getShiftAmountTy(InnerVT);
1453 if (!APInt(BitWidth, ShAmt).isIntN(ShTy.getSizeInBits()))
1456 TLO.DAG.getNode(ISD::SHL, dl, InnerVT, InnerOp,
1457 TLO.DAG.getConstant(ShAmt, ShTy));
1460 TLO.DAG.getNode(ISD::ANY_EXTEND, dl, Op.getValueType(),
1465 KnownZero <<= SA->getZExtValue();
1466 KnownOne <<= SA->getZExtValue();
1467 // low bits known zero.
1468 KnownZero |= APInt::getLowBitsSet(BitWidth, SA->getZExtValue());
1472 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1473 EVT VT = Op.getValueType();
1474 unsigned ShAmt = SA->getZExtValue();
1475 unsigned VTSize = VT.getSizeInBits();
1476 SDValue InOp = Op.getOperand(0);
1478 // If the shift count is an invalid immediate, don't do anything.
1479 if (ShAmt >= BitWidth)
1482 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a
1483 // single shift. We can do this if the top bits (which are shifted out)
1484 // are never demanded.
1485 if (InOp.getOpcode() == ISD::SHL &&
1486 isa<ConstantSDNode>(InOp.getOperand(1))) {
1487 if (ShAmt && (NewMask & APInt::getHighBitsSet(VTSize, ShAmt)) == 0) {
1488 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
1489 unsigned Opc = ISD::SRL;
1490 int Diff = ShAmt-C1;
1497 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
1498 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT,
1499 InOp.getOperand(0), NewSA));
1503 // Compute the new bits that are at the top now.
1504 if (SimplifyDemandedBits(InOp, (NewMask << ShAmt),
1505 KnownZero, KnownOne, TLO, Depth+1))
1507 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1508 KnownZero = KnownZero.lshr(ShAmt);
1509 KnownOne = KnownOne.lshr(ShAmt);
1511 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
1512 KnownZero |= HighBits; // High bits known zero.
1516 // If this is an arithmetic shift right and only the low-bit is set, we can
1517 // always convert this into a logical shr, even if the shift amount is
1518 // variable. The low bit of the shift cannot be an input sign bit unless
1519 // the shift amount is >= the size of the datatype, which is undefined.
1520 if (DemandedMask == 1)
1521 return TLO.CombineTo(Op,
1522 TLO.DAG.getNode(ISD::SRL, dl, Op.getValueType(),
1523 Op.getOperand(0), Op.getOperand(1)));
1525 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1526 EVT VT = Op.getValueType();
1527 unsigned ShAmt = SA->getZExtValue();
1529 // If the shift count is an invalid immediate, don't do anything.
1530 if (ShAmt >= BitWidth)
1533 APInt InDemandedMask = (NewMask << ShAmt);
1535 // If any of the demanded bits are produced by the sign extension, we also
1536 // demand the input sign bit.
1537 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
1538 if (HighBits.intersects(NewMask))
1539 InDemandedMask |= APInt::getSignBit(VT.getScalarType().getSizeInBits());
1541 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask,
1542 KnownZero, KnownOne, TLO, Depth+1))
1544 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1545 KnownZero = KnownZero.lshr(ShAmt);
1546 KnownOne = KnownOne.lshr(ShAmt);
1548 // Handle the sign bit, adjusted to where it is now in the mask.
1549 APInt SignBit = APInt::getSignBit(BitWidth).lshr(ShAmt);
1551 // If the input sign bit is known to be zero, or if none of the top bits
1552 // are demanded, turn this into an unsigned shift right.
1553 if (KnownZero.intersects(SignBit) || (HighBits & ~NewMask) == HighBits) {
1554 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, VT,
1557 } else if (KnownOne.intersects(SignBit)) { // New bits are known one.
1558 KnownOne |= HighBits;
1562 case ISD::SIGN_EXTEND_INREG: {
1563 EVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1565 // Sign extension. Compute the demanded bits in the result that are not
1566 // present in the input.
1568 APInt::getHighBitsSet(BitWidth,
1569 BitWidth - EVT.getScalarType().getSizeInBits());
1571 // If none of the extended bits are demanded, eliminate the sextinreg.
1572 if ((NewBits & NewMask) == 0)
1573 return TLO.CombineTo(Op, Op.getOperand(0));
1576 APInt::getSignBit(EVT.getScalarType().getSizeInBits()).zext(BitWidth);
1577 APInt InputDemandedBits =
1578 APInt::getLowBitsSet(BitWidth,
1579 EVT.getScalarType().getSizeInBits()) &
1582 // Since the sign extended bits are demanded, we know that the sign
1584 InputDemandedBits |= InSignBit;
1586 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits,
1587 KnownZero, KnownOne, TLO, Depth+1))
1589 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1591 // If the sign bit of the input is known set or clear, then we know the
1592 // top bits of the result.
1594 // If the input sign bit is known zero, convert this into a zero extension.
1595 if (KnownZero.intersects(InSignBit))
1596 return TLO.CombineTo(Op,
1597 TLO.DAG.getZeroExtendInReg(Op.getOperand(0),dl,EVT));
1599 if (KnownOne.intersects(InSignBit)) { // Input sign bit known set
1600 KnownOne |= NewBits;
1601 KnownZero &= ~NewBits;
1602 } else { // Input sign bit unknown
1603 KnownZero &= ~NewBits;
1604 KnownOne &= ~NewBits;
1608 case ISD::ZERO_EXTEND: {
1609 unsigned OperandBitWidth =
1610 Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
1611 APInt InMask = NewMask.trunc(OperandBitWidth);
1613 // If none of the top bits are demanded, convert this into an any_extend.
1615 APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask;
1616 if (!NewBits.intersects(NewMask))
1617 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND, dl,
1621 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1622 KnownZero, KnownOne, TLO, Depth+1))
1624 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1625 KnownZero = KnownZero.zext(BitWidth);
1626 KnownOne = KnownOne.zext(BitWidth);
1627 KnownZero |= NewBits;
1630 case ISD::SIGN_EXTEND: {
1631 EVT InVT = Op.getOperand(0).getValueType();
1632 unsigned InBits = InVT.getScalarType().getSizeInBits();
1633 APInt InMask = APInt::getLowBitsSet(BitWidth, InBits);
1634 APInt InSignBit = APInt::getBitsSet(BitWidth, InBits - 1, InBits);
1635 APInt NewBits = ~InMask & NewMask;
1637 // If none of the top bits are demanded, convert this into an any_extend.
1639 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND, dl,
1643 // Since some of the sign extended bits are demanded, we know that the sign
1645 APInt InDemandedBits = InMask & NewMask;
1646 InDemandedBits |= InSignBit;
1647 InDemandedBits = InDemandedBits.trunc(InBits);
1649 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero,
1650 KnownOne, TLO, Depth+1))
1652 KnownZero = KnownZero.zext(BitWidth);
1653 KnownOne = KnownOne.zext(BitWidth);
1655 // If the sign bit is known zero, convert this to a zero extend.
1656 if (KnownZero.intersects(InSignBit))
1657 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND, dl,
1661 // If the sign bit is known one, the top bits match.
1662 if (KnownOne.intersects(InSignBit)) {
1663 KnownOne |= NewBits;
1664 KnownZero &= ~NewBits;
1665 } else { // Otherwise, top bits aren't known.
1666 KnownOne &= ~NewBits;
1667 KnownZero &= ~NewBits;
1671 case ISD::ANY_EXTEND: {
1672 unsigned OperandBitWidth =
1673 Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
1674 APInt InMask = NewMask.trunc(OperandBitWidth);
1675 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1676 KnownZero, KnownOne, TLO, Depth+1))
1678 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1679 KnownZero = KnownZero.zext(BitWidth);
1680 KnownOne = KnownOne.zext(BitWidth);
1683 case ISD::TRUNCATE: {
1684 // Simplify the input, using demanded bit information, and compute the known
1685 // zero/one bits live out.
1686 unsigned OperandBitWidth =
1687 Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
1688 APInt TruncMask = NewMask.zext(OperandBitWidth);
1689 if (SimplifyDemandedBits(Op.getOperand(0), TruncMask,
1690 KnownZero, KnownOne, TLO, Depth+1))
1692 KnownZero = KnownZero.trunc(BitWidth);
1693 KnownOne = KnownOne.trunc(BitWidth);
1695 // If the input is only used by this truncate, see if we can shrink it based
1696 // on the known demanded bits.
1697 if (Op.getOperand(0).getNode()->hasOneUse()) {
1698 SDValue In = Op.getOperand(0);
1699 switch (In.getOpcode()) {
1702 // Shrink SRL by a constant if none of the high bits shifted in are
1704 if (TLO.LegalTypes() &&
1705 !isTypeDesirableForOp(ISD::SRL, Op.getValueType()))
1706 // Do not turn (vt1 truncate (vt2 srl)) into (vt1 srl) if vt1 is
1709 ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1));
1712 SDValue Shift = In.getOperand(1);
1713 if (TLO.LegalTypes()) {
1714 uint64_t ShVal = ShAmt->getZExtValue();
1716 TLO.DAG.getConstant(ShVal, getShiftAmountTy(Op.getValueType()));
1719 APInt HighBits = APInt::getHighBitsSet(OperandBitWidth,
1720 OperandBitWidth - BitWidth);
1721 HighBits = HighBits.lshr(ShAmt->getZExtValue()).trunc(BitWidth);
1723 if (ShAmt->getZExtValue() < BitWidth && !(HighBits & NewMask)) {
1724 // None of the shifted in bits are needed. Add a truncate of the
1725 // shift input, then shift it.
1726 SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE, dl,
1729 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl,
1738 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1741 case ISD::AssertZext: {
1742 // Demand all the bits of the input that are demanded in the output.
1743 // The low bits are obvious; the high bits are demanded because we're
1744 // asserting that they're zero here.
1745 if (SimplifyDemandedBits(Op.getOperand(0), NewMask,
1746 KnownZero, KnownOne, TLO, Depth+1))
1748 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1750 EVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1751 APInt InMask = APInt::getLowBitsSet(BitWidth,
1752 VT.getSizeInBits());
1753 KnownZero |= ~InMask & NewMask;
1757 // If this is an FP->Int bitcast and if the sign bit is the only thing that
1758 // is demanded, turn this into a FGETSIGN.
1759 if (NewMask == APInt::getSignBit(Op.getValueType().getSizeInBits()) &&
1760 Op.getOperand(0).getValueType().isFloatingPoint() &&
1761 !Op.getOperand(0).getValueType().isVector()) {
1762 bool OpVTLegal = isOperationLegalOrCustom(ISD::FGETSIGN, Op.getValueType());
1763 bool i32Legal = isOperationLegalOrCustom(ISD::FGETSIGN, MVT::i32);
1764 if (OpVTLegal || i32Legal) {
1765 EVT Ty = OpVTLegal ? Op.getValueType() : MVT::i32;
1766 // Make a FGETSIGN + SHL to move the sign bit into the appropriate
1767 // place. We expect the SHL to be eliminated by other optimizations.
1768 SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, dl, Ty, Op.getOperand(0));
1770 Sign = TLO.DAG.getNode(ISD::ZERO_EXTEND, dl, Op.getValueType(), Sign);
1771 unsigned ShVal = Op.getValueType().getSizeInBits()-1;
1772 SDValue ShAmt = TLO.DAG.getConstant(ShVal, Op.getValueType());
1773 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl,
1782 // Add, Sub, and Mul don't demand any bits in positions beyond that
1783 // of the highest bit demanded of them.
1784 APInt LoMask = APInt::getLowBitsSet(BitWidth,
1785 BitWidth - NewMask.countLeadingZeros());
1786 if (SimplifyDemandedBits(Op.getOperand(0), LoMask, KnownZero2,
1787 KnownOne2, TLO, Depth+1))
1789 if (SimplifyDemandedBits(Op.getOperand(1), LoMask, KnownZero2,
1790 KnownOne2, TLO, Depth+1))
1792 // See if the operation should be performed at a smaller bit width.
1793 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1798 // Just use ComputeMaskedBits to compute output bits.
1799 TLO.DAG.ComputeMaskedBits(Op, NewMask, KnownZero, KnownOne, Depth);
1803 // If we know the value of all of the demanded bits, return this as a
1805 if ((NewMask & (KnownZero|KnownOne)) == NewMask)
1806 return TLO.CombineTo(Op, TLO.DAG.getConstant(KnownOne, Op.getValueType()));
1811 /// computeMaskedBitsForTargetNode - Determine which of the bits specified
1812 /// in Mask are known to be either zero or one and return them in the
1813 /// KnownZero/KnownOne bitsets.
1814 void TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
1818 const SelectionDAG &DAG,
1819 unsigned Depth) const {
1820 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1821 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1822 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1823 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1824 "Should use MaskedValueIsZero if you don't know whether Op"
1825 " is a target node!");
1826 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
1829 /// ComputeNumSignBitsForTargetNode - This method can be implemented by
1830 /// targets that want to expose additional information about sign bits to the
1832 unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
1833 unsigned Depth) const {
1834 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1835 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1836 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1837 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1838 "Should use ComputeNumSignBits if you don't know whether Op"
1839 " is a target node!");
1843 /// ValueHasExactlyOneBitSet - Test if the given value is known to have exactly
1844 /// one bit set. This differs from ComputeMaskedBits in that it doesn't need to
1845 /// determine which bit is set.
1847 static bool ValueHasExactlyOneBitSet(SDValue Val, const SelectionDAG &DAG) {
1848 // A left-shift of a constant one will have exactly one bit set, because
1849 // shifting the bit off the end is undefined.
1850 if (Val.getOpcode() == ISD::SHL)
1851 if (ConstantSDNode *C =
1852 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0)))
1853 if (C->getAPIntValue() == 1)
1856 // Similarly, a right-shift of a constant sign-bit will have exactly
1858 if (Val.getOpcode() == ISD::SRL)
1859 if (ConstantSDNode *C =
1860 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0)))
1861 if (C->getAPIntValue().isSignBit())
1864 // More could be done here, though the above checks are enough
1865 // to handle some common cases.
1867 // Fall back to ComputeMaskedBits to catch other known cases.
1868 EVT OpVT = Val.getValueType();
1869 unsigned BitWidth = OpVT.getScalarType().getSizeInBits();
1870 APInt Mask = APInt::getAllOnesValue(BitWidth);
1871 APInt KnownZero, KnownOne;
1872 DAG.ComputeMaskedBits(Val, Mask, KnownZero, KnownOne);
1873 return (KnownZero.countPopulation() == BitWidth - 1) &&
1874 (KnownOne.countPopulation() == 1);
1877 /// SimplifySetCC - Try to simplify a setcc built with the specified operands
1878 /// and cc. If it is unable to simplify it, return a null SDValue.
1880 TargetLowering::SimplifySetCC(EVT VT, SDValue N0, SDValue N1,
1881 ISD::CondCode Cond, bool foldBooleans,
1882 DAGCombinerInfo &DCI, DebugLoc dl) const {
1883 SelectionDAG &DAG = DCI.DAG;
1885 // These setcc operations always fold.
1889 case ISD::SETFALSE2: return DAG.getConstant(0, VT);
1891 case ISD::SETTRUE2: return DAG.getConstant(1, VT);
1894 // Ensure that the constant occurs on the RHS, and fold constant
1896 if (isa<ConstantSDNode>(N0.getNode()))
1897 return DAG.getSetCC(dl, VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
1899 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) {
1900 const APInt &C1 = N1C->getAPIntValue();
1902 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an
1903 // equality comparison, then we're just comparing whether X itself is
1905 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) &&
1906 N0.getOperand(0).getOpcode() == ISD::CTLZ &&
1907 N0.getOperand(1).getOpcode() == ISD::Constant) {
1909 = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
1910 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1911 ShAmt == Log2_32(N0.getValueType().getSizeInBits())) {
1912 if ((C1 == 0) == (Cond == ISD::SETEQ)) {
1913 // (srl (ctlz x), 5) == 0 -> X != 0
1914 // (srl (ctlz x), 5) != 1 -> X != 0
1917 // (srl (ctlz x), 5) != 0 -> X == 0
1918 // (srl (ctlz x), 5) == 1 -> X == 0
1921 SDValue Zero = DAG.getConstant(0, N0.getValueType());
1922 return DAG.getSetCC(dl, VT, N0.getOperand(0).getOperand(0),
1928 // Look through truncs that don't change the value of a ctpop.
1929 if (N0.hasOneUse() && N0.getOpcode() == ISD::TRUNCATE)
1930 CTPOP = N0.getOperand(0);
1932 if (CTPOP.hasOneUse() && CTPOP.getOpcode() == ISD::CTPOP &&
1933 (N0 == CTPOP || N0.getValueType().getSizeInBits() >
1934 Log2_32_Ceil(CTPOP.getValueType().getSizeInBits()))) {
1935 EVT CTVT = CTPOP.getValueType();
1936 SDValue CTOp = CTPOP.getOperand(0);
1938 // (ctpop x) u< 2 -> (x & x-1) == 0
1939 // (ctpop x) u> 1 -> (x & x-1) != 0
1940 if ((Cond == ISD::SETULT && C1 == 2) || (Cond == ISD::SETUGT && C1 == 1)){
1941 SDValue Sub = DAG.getNode(ISD::SUB, dl, CTVT, CTOp,
1942 DAG.getConstant(1, CTVT));
1943 SDValue And = DAG.getNode(ISD::AND, dl, CTVT, CTOp, Sub);
1944 ISD::CondCode CC = Cond == ISD::SETULT ? ISD::SETEQ : ISD::SETNE;
1945 return DAG.getSetCC(dl, VT, And, DAG.getConstant(0, CTVT), CC);
1948 // TODO: (ctpop x) == 1 -> x && (x & x-1) == 0 iff ctpop is illegal.
1951 // (zext x) == C --> x == (trunc C)
1952 if (DCI.isBeforeLegalize() && N0->hasOneUse() &&
1953 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1954 unsigned MinBits = N0.getValueSizeInBits();
1956 if (N0->getOpcode() == ISD::ZERO_EXTEND) {
1958 MinBits = N0->getOperand(0).getValueSizeInBits();
1959 PreZExt = N0->getOperand(0);
1960 } else if (N0->getOpcode() == ISD::AND) {
1961 // DAGCombine turns costly ZExts into ANDs
1962 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0->getOperand(1)))
1963 if ((C->getAPIntValue()+1).isPowerOf2()) {
1964 MinBits = C->getAPIntValue().countTrailingOnes();
1965 PreZExt = N0->getOperand(0);
1967 } else if (LoadSDNode *LN0 = dyn_cast<LoadSDNode>(N0)) {
1969 if (LN0->getExtensionType() == ISD::ZEXTLOAD) {
1970 MinBits = LN0->getMemoryVT().getSizeInBits();
1975 // Make sure we're not loosing bits from the constant.
1976 if (MinBits < C1.getBitWidth() && MinBits > C1.getActiveBits()) {
1977 EVT MinVT = EVT::getIntegerVT(*DAG.getContext(), MinBits);
1978 if (isTypeDesirableForOp(ISD::SETCC, MinVT)) {
1979 // Will get folded away.
1980 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, MinVT, PreZExt);
1981 SDValue C = DAG.getConstant(C1.trunc(MinBits), MinVT);
1982 return DAG.getSetCC(dl, VT, Trunc, C, Cond);
1987 // If the LHS is '(and load, const)', the RHS is 0,
1988 // the test is for equality or unsigned, and all 1 bits of the const are
1989 // in the same partial word, see if we can shorten the load.
1990 if (DCI.isBeforeLegalize() &&
1991 N0.getOpcode() == ISD::AND && C1 == 0 &&
1992 N0.getNode()->hasOneUse() &&
1993 isa<LoadSDNode>(N0.getOperand(0)) &&
1994 N0.getOperand(0).getNode()->hasOneUse() &&
1995 isa<ConstantSDNode>(N0.getOperand(1))) {
1996 LoadSDNode *Lod = cast<LoadSDNode>(N0.getOperand(0));
1998 unsigned bestWidth = 0, bestOffset = 0;
1999 if (!Lod->isVolatile() && Lod->isUnindexed()) {
2000 unsigned origWidth = N0.getValueType().getSizeInBits();
2001 unsigned maskWidth = origWidth;
2002 // We can narrow (e.g.) 16-bit extending loads on 32-bit target to
2003 // 8 bits, but have to be careful...
2004 if (Lod->getExtensionType() != ISD::NON_EXTLOAD)
2005 origWidth = Lod->getMemoryVT().getSizeInBits();
2007 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
2008 for (unsigned width = origWidth / 2; width>=8; width /= 2) {
2009 APInt newMask = APInt::getLowBitsSet(maskWidth, width);
2010 for (unsigned offset=0; offset<origWidth/width; offset++) {
2011 if ((newMask & Mask) == Mask) {
2012 if (!TD->isLittleEndian())
2013 bestOffset = (origWidth/width - offset - 1) * (width/8);
2015 bestOffset = (uint64_t)offset * (width/8);
2016 bestMask = Mask.lshr(offset * (width/8) * 8);
2020 newMask = newMask << width;
2025 EVT newVT = EVT::getIntegerVT(*DAG.getContext(), bestWidth);
2026 if (newVT.isRound()) {
2027 EVT PtrType = Lod->getOperand(1).getValueType();
2028 SDValue Ptr = Lod->getBasePtr();
2029 if (bestOffset != 0)
2030 Ptr = DAG.getNode(ISD::ADD, dl, PtrType, Lod->getBasePtr(),
2031 DAG.getConstant(bestOffset, PtrType));
2032 unsigned NewAlign = MinAlign(Lod->getAlignment(), bestOffset);
2033 SDValue NewLoad = DAG.getLoad(newVT, dl, Lod->getChain(), Ptr,
2034 Lod->getPointerInfo().getWithOffset(bestOffset),
2035 false, false, NewAlign);
2036 return DAG.getSetCC(dl, VT,
2037 DAG.getNode(ISD::AND, dl, newVT, NewLoad,
2038 DAG.getConstant(bestMask.trunc(bestWidth),
2040 DAG.getConstant(0LL, newVT), Cond);
2045 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
2046 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
2047 unsigned InSize = N0.getOperand(0).getValueType().getSizeInBits();
2049 // If the comparison constant has bits in the upper part, the
2050 // zero-extended value could never match.
2051 if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(),
2052 C1.getBitWidth() - InSize))) {
2056 case ISD::SETEQ: return DAG.getConstant(0, VT);
2059 case ISD::SETNE: return DAG.getConstant(1, VT);
2062 // True if the sign bit of C1 is set.
2063 return DAG.getConstant(C1.isNegative(), VT);
2066 // True if the sign bit of C1 isn't set.
2067 return DAG.getConstant(C1.isNonNegative(), VT);
2073 // Otherwise, we can perform the comparison with the low bits.
2081 EVT newVT = N0.getOperand(0).getValueType();
2082 if (DCI.isBeforeLegalizeOps() ||
2083 (isOperationLegal(ISD::SETCC, newVT) &&
2084 getCondCodeAction(Cond, newVT)==Legal))
2085 return DAG.getSetCC(dl, VT, N0.getOperand(0),
2086 DAG.getConstant(C1.trunc(InSize), newVT),
2091 break; // todo, be more careful with signed comparisons
2093 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
2094 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
2095 EVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
2096 unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits();
2097 EVT ExtDstTy = N0.getValueType();
2098 unsigned ExtDstTyBits = ExtDstTy.getSizeInBits();
2100 // If the constant doesn't fit into the number of bits for the source of
2101 // the sign extension, it is impossible for both sides to be equal.
2102 if (C1.getMinSignedBits() > ExtSrcTyBits)
2103 return DAG.getConstant(Cond == ISD::SETNE, VT);
2106 EVT Op0Ty = N0.getOperand(0).getValueType();
2107 if (Op0Ty == ExtSrcTy) {
2108 ZextOp = N0.getOperand(0);
2110 APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits);
2111 ZextOp = DAG.getNode(ISD::AND, dl, Op0Ty, N0.getOperand(0),
2112 DAG.getConstant(Imm, Op0Ty));
2114 if (!DCI.isCalledByLegalizer())
2115 DCI.AddToWorklist(ZextOp.getNode());
2116 // Otherwise, make this a use of a zext.
2117 return DAG.getSetCC(dl, VT, ZextOp,
2118 DAG.getConstant(C1 & APInt::getLowBitsSet(
2123 } else if ((N1C->isNullValue() || N1C->getAPIntValue() == 1) &&
2124 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
2125 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC
2126 if (N0.getOpcode() == ISD::SETCC &&
2127 isTypeLegal(VT) && VT.bitsLE(N0.getValueType())) {
2128 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getAPIntValue() != 1);
2130 return DAG.getNode(ISD::TRUNCATE, dl, VT, N0);
2131 // Invert the condition.
2132 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
2133 CC = ISD::getSetCCInverse(CC,
2134 N0.getOperand(0).getValueType().isInteger());
2135 return DAG.getSetCC(dl, VT, N0.getOperand(0), N0.getOperand(1), CC);
2138 if ((N0.getOpcode() == ISD::XOR ||
2139 (N0.getOpcode() == ISD::AND &&
2140 N0.getOperand(0).getOpcode() == ISD::XOR &&
2141 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
2142 isa<ConstantSDNode>(N0.getOperand(1)) &&
2143 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue() == 1) {
2144 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We
2145 // can only do this if the top bits are known zero.
2146 unsigned BitWidth = N0.getValueSizeInBits();
2147 if (DAG.MaskedValueIsZero(N0,
2148 APInt::getHighBitsSet(BitWidth,
2150 // Okay, get the un-inverted input value.
2152 if (N0.getOpcode() == ISD::XOR)
2153 Val = N0.getOperand(0);
2155 assert(N0.getOpcode() == ISD::AND &&
2156 N0.getOperand(0).getOpcode() == ISD::XOR);
2157 // ((X^1)&1)^1 -> X & 1
2158 Val = DAG.getNode(ISD::AND, dl, N0.getValueType(),
2159 N0.getOperand(0).getOperand(0),
2163 return DAG.getSetCC(dl, VT, Val, N1,
2164 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
2166 } else if (N1C->getAPIntValue() == 1 &&
2168 getBooleanContents() == ZeroOrOneBooleanContent)) {
2170 if (Op0.getOpcode() == ISD::TRUNCATE)
2171 Op0 = Op0.getOperand(0);
2173 if ((Op0.getOpcode() == ISD::XOR) &&
2174 Op0.getOperand(0).getOpcode() == ISD::SETCC &&
2175 Op0.getOperand(1).getOpcode() == ISD::SETCC) {
2176 // (xor (setcc), (setcc)) == / != 1 -> (setcc) != / == (setcc)
2177 Cond = (Cond == ISD::SETEQ) ? ISD::SETNE : ISD::SETEQ;
2178 return DAG.getSetCC(dl, VT, Op0.getOperand(0), Op0.getOperand(1),
2180 } else if (Op0.getOpcode() == ISD::AND &&
2181 isa<ConstantSDNode>(Op0.getOperand(1)) &&
2182 cast<ConstantSDNode>(Op0.getOperand(1))->getAPIntValue() == 1) {
2183 // If this is (X&1) == / != 1, normalize it to (X&1) != / == 0.
2184 if (Op0.getValueType().bitsGT(VT))
2185 Op0 = DAG.getNode(ISD::AND, dl, VT,
2186 DAG.getNode(ISD::TRUNCATE, dl, VT, Op0.getOperand(0)),
2187 DAG.getConstant(1, VT));
2188 else if (Op0.getValueType().bitsLT(VT))
2189 Op0 = DAG.getNode(ISD::AND, dl, VT,
2190 DAG.getNode(ISD::ANY_EXTEND, dl, VT, Op0.getOperand(0)),
2191 DAG.getConstant(1, VT));
2193 return DAG.getSetCC(dl, VT, Op0,
2194 DAG.getConstant(0, Op0.getValueType()),
2195 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
2200 APInt MinVal, MaxVal;
2201 unsigned OperandBitSize = N1C->getValueType(0).getSizeInBits();
2202 if (ISD::isSignedIntSetCC(Cond)) {
2203 MinVal = APInt::getSignedMinValue(OperandBitSize);
2204 MaxVal = APInt::getSignedMaxValue(OperandBitSize);
2206 MinVal = APInt::getMinValue(OperandBitSize);
2207 MaxVal = APInt::getMaxValue(OperandBitSize);
2210 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
2211 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
2212 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true
2213 // X >= C0 --> X > (C0-1)
2214 return DAG.getSetCC(dl, VT, N0,
2215 DAG.getConstant(C1-1, N1.getValueType()),
2216 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
2219 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
2220 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true
2221 // X <= C0 --> X < (C0+1)
2222 return DAG.getSetCC(dl, VT, N0,
2223 DAG.getConstant(C1+1, N1.getValueType()),
2224 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
2227 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
2228 return DAG.getConstant(0, VT); // X < MIN --> false
2229 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal)
2230 return DAG.getConstant(1, VT); // X >= MIN --> true
2231 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal)
2232 return DAG.getConstant(0, VT); // X > MAX --> false
2233 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal)
2234 return DAG.getConstant(1, VT); // X <= MAX --> true
2236 // Canonicalize setgt X, Min --> setne X, Min
2237 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
2238 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
2239 // Canonicalize setlt X, Max --> setne X, Max
2240 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
2241 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
2243 // If we have setult X, 1, turn it into seteq X, 0
2244 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
2245 return DAG.getSetCC(dl, VT, N0,
2246 DAG.getConstant(MinVal, N0.getValueType()),
2248 // If we have setugt X, Max-1, turn it into seteq X, Max
2249 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
2250 return DAG.getSetCC(dl, VT, N0,
2251 DAG.getConstant(MaxVal, N0.getValueType()),
2254 // If we have "setcc X, C0", check to see if we can shrink the immediate
2257 // SETUGT X, SINTMAX -> SETLT X, 0
2258 if (Cond == ISD::SETUGT &&
2259 C1 == APInt::getSignedMaxValue(OperandBitSize))
2260 return DAG.getSetCC(dl, VT, N0,
2261 DAG.getConstant(0, N1.getValueType()),
2264 // SETULT X, SINTMIN -> SETGT X, -1
2265 if (Cond == ISD::SETULT &&
2266 C1 == APInt::getSignedMinValue(OperandBitSize)) {
2267 SDValue ConstMinusOne =
2268 DAG.getConstant(APInt::getAllOnesValue(OperandBitSize),
2270 return DAG.getSetCC(dl, VT, N0, ConstMinusOne, ISD::SETGT);
2273 // Fold bit comparisons when we can.
2274 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
2275 (VT == N0.getValueType() ||
2276 (isTypeLegal(VT) && VT.bitsLE(N0.getValueType()))) &&
2277 N0.getOpcode() == ISD::AND)
2278 if (ConstantSDNode *AndRHS =
2279 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2280 EVT ShiftTy = DCI.isBeforeLegalize() ?
2281 getPointerTy() : getShiftAmountTy(N0.getValueType());
2282 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
2283 // Perform the xform if the AND RHS is a single bit.
2284 if (AndRHS->getAPIntValue().isPowerOf2()) {
2285 return DAG.getNode(ISD::TRUNCATE, dl, VT,
2286 DAG.getNode(ISD::SRL, dl, N0.getValueType(), N0,
2287 DAG.getConstant(AndRHS->getAPIntValue().logBase2(), ShiftTy)));
2289 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getAPIntValue()) {
2290 // (X & 8) == 8 --> (X & 8) >> 3
2291 // Perform the xform if C1 is a single bit.
2292 if (C1.isPowerOf2()) {
2293 return DAG.getNode(ISD::TRUNCATE, dl, VT,
2294 DAG.getNode(ISD::SRL, dl, N0.getValueType(), N0,
2295 DAG.getConstant(C1.logBase2(), ShiftTy)));
2301 if (isa<ConstantFPSDNode>(N0.getNode())) {
2302 // Constant fold or commute setcc.
2303 SDValue O = DAG.FoldSetCC(VT, N0, N1, Cond, dl);
2304 if (O.getNode()) return O;
2305 } else if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1.getNode())) {
2306 // If the RHS of an FP comparison is a constant, simplify it away in
2308 if (CFP->getValueAPF().isNaN()) {
2309 // If an operand is known to be a nan, we can fold it.
2310 switch (ISD::getUnorderedFlavor(Cond)) {
2311 default: llvm_unreachable("Unknown flavor!");
2312 case 0: // Known false.
2313 return DAG.getConstant(0, VT);
2314 case 1: // Known true.
2315 return DAG.getConstant(1, VT);
2316 case 2: // Undefined.
2317 return DAG.getUNDEF(VT);
2321 // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the
2322 // constant if knowing that the operand is non-nan is enough. We prefer to
2323 // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to
2325 if (Cond == ISD::SETO || Cond == ISD::SETUO)
2326 return DAG.getSetCC(dl, VT, N0, N0, Cond);
2328 // If the condition is not legal, see if we can find an equivalent one
2330 if (!isCondCodeLegal(Cond, N0.getValueType())) {
2331 // If the comparison was an awkward floating-point == or != and one of
2332 // the comparison operands is infinity or negative infinity, convert the
2333 // condition to a less-awkward <= or >=.
2334 if (CFP->getValueAPF().isInfinity()) {
2335 if (CFP->getValueAPF().isNegative()) {
2336 if (Cond == ISD::SETOEQ &&
2337 isCondCodeLegal(ISD::SETOLE, N0.getValueType()))
2338 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLE);
2339 if (Cond == ISD::SETUEQ &&
2340 isCondCodeLegal(ISD::SETOLE, N0.getValueType()))
2341 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULE);
2342 if (Cond == ISD::SETUNE &&
2343 isCondCodeLegal(ISD::SETUGT, N0.getValueType()))
2344 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGT);
2345 if (Cond == ISD::SETONE &&
2346 isCondCodeLegal(ISD::SETUGT, N0.getValueType()))
2347 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGT);
2349 if (Cond == ISD::SETOEQ &&
2350 isCondCodeLegal(ISD::SETOGE, N0.getValueType()))
2351 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGE);
2352 if (Cond == ISD::SETUEQ &&
2353 isCondCodeLegal(ISD::SETOGE, N0.getValueType()))
2354 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGE);
2355 if (Cond == ISD::SETUNE &&
2356 isCondCodeLegal(ISD::SETULT, N0.getValueType()))
2357 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULT);
2358 if (Cond == ISD::SETONE &&
2359 isCondCodeLegal(ISD::SETULT, N0.getValueType()))
2360 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLT);
2367 // We can always fold X == X for integer setcc's.
2368 if (N0.getValueType().isInteger())
2369 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
2370 unsigned UOF = ISD::getUnorderedFlavor(Cond);
2371 if (UOF == 2) // FP operators that are undefined on NaNs.
2372 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
2373 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
2374 return DAG.getConstant(UOF, VT);
2375 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
2376 // if it is not already.
2377 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
2378 if (NewCond != Cond)
2379 return DAG.getSetCC(dl, VT, N0, N1, NewCond);
2382 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
2383 N0.getValueType().isInteger()) {
2384 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
2385 N0.getOpcode() == ISD::XOR) {
2386 // Simplify (X+Y) == (X+Z) --> Y == Z
2387 if (N0.getOpcode() == N1.getOpcode()) {
2388 if (N0.getOperand(0) == N1.getOperand(0))
2389 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(1), Cond);
2390 if (N0.getOperand(1) == N1.getOperand(1))
2391 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(0), Cond);
2392 if (DAG.isCommutativeBinOp(N0.getOpcode())) {
2393 // If X op Y == Y op X, try other combinations.
2394 if (N0.getOperand(0) == N1.getOperand(1))
2395 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(0),
2397 if (N0.getOperand(1) == N1.getOperand(0))
2398 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(1),
2403 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) {
2404 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2405 // Turn (X+C1) == C2 --> X == C2-C1
2406 if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse()) {
2407 return DAG.getSetCC(dl, VT, N0.getOperand(0),
2408 DAG.getConstant(RHSC->getAPIntValue()-
2409 LHSR->getAPIntValue(),
2410 N0.getValueType()), Cond);
2413 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
2414 if (N0.getOpcode() == ISD::XOR)
2415 // If we know that all of the inverted bits are zero, don't bother
2416 // performing the inversion.
2417 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue()))
2419 DAG.getSetCC(dl, VT, N0.getOperand(0),
2420 DAG.getConstant(LHSR->getAPIntValue() ^
2421 RHSC->getAPIntValue(),
2426 // Turn (C1-X) == C2 --> X == C1-C2
2427 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
2428 if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse()) {
2430 DAG.getSetCC(dl, VT, N0.getOperand(1),
2431 DAG.getConstant(SUBC->getAPIntValue() -
2432 RHSC->getAPIntValue(),
2439 // Simplify (X+Z) == X --> Z == 0
2440 if (N0.getOperand(0) == N1)
2441 return DAG.getSetCC(dl, VT, N0.getOperand(1),
2442 DAG.getConstant(0, N0.getValueType()), Cond);
2443 if (N0.getOperand(1) == N1) {
2444 if (DAG.isCommutativeBinOp(N0.getOpcode()))
2445 return DAG.getSetCC(dl, VT, N0.getOperand(0),
2446 DAG.getConstant(0, N0.getValueType()), Cond);
2447 else if (N0.getNode()->hasOneUse()) {
2448 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
2449 // (Z-X) == X --> Z == X<<1
2450 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(),
2452 DAG.getConstant(1, getShiftAmountTy(N1.getValueType())));
2453 if (!DCI.isCalledByLegalizer())
2454 DCI.AddToWorklist(SH.getNode());
2455 return DAG.getSetCC(dl, VT, N0.getOperand(0), SH, Cond);
2460 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
2461 N1.getOpcode() == ISD::XOR) {
2462 // Simplify X == (X+Z) --> Z == 0
2463 if (N1.getOperand(0) == N0) {
2464 return DAG.getSetCC(dl, VT, N1.getOperand(1),
2465 DAG.getConstant(0, N1.getValueType()), Cond);
2466 } else if (N1.getOperand(1) == N0) {
2467 if (DAG.isCommutativeBinOp(N1.getOpcode())) {
2468 return DAG.getSetCC(dl, VT, N1.getOperand(0),
2469 DAG.getConstant(0, N1.getValueType()), Cond);
2470 } else if (N1.getNode()->hasOneUse()) {
2471 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
2472 // X == (Z-X) --> X<<1 == Z
2473 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(), N0,
2474 DAG.getConstant(1, getShiftAmountTy(N0.getValueType())));
2475 if (!DCI.isCalledByLegalizer())
2476 DCI.AddToWorklist(SH.getNode());
2477 return DAG.getSetCC(dl, VT, SH, N1.getOperand(0), Cond);
2482 // Simplify x&y == y to x&y != 0 if y has exactly one bit set.
2483 // Note that where y is variable and is known to have at most
2484 // one bit set (for example, if it is z&1) we cannot do this;
2485 // the expressions are not equivalent when y==0.
2486 if (N0.getOpcode() == ISD::AND)
2487 if (N0.getOperand(0) == N1 || N0.getOperand(1) == N1) {
2488 if (ValueHasExactlyOneBitSet(N1, DAG)) {
2489 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true);
2490 SDValue Zero = DAG.getConstant(0, N1.getValueType());
2491 return DAG.getSetCC(dl, VT, N0, Zero, Cond);
2494 if (N1.getOpcode() == ISD::AND)
2495 if (N1.getOperand(0) == N0 || N1.getOperand(1) == N0) {
2496 if (ValueHasExactlyOneBitSet(N0, DAG)) {
2497 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true);
2498 SDValue Zero = DAG.getConstant(0, N0.getValueType());
2499 return DAG.getSetCC(dl, VT, N1, Zero, Cond);
2504 // Fold away ALL boolean setcc's.
2506 if (N0.getValueType() == MVT::i1 && foldBooleans) {
2508 default: llvm_unreachable("Unknown integer setcc!");
2509 case ISD::SETEQ: // X == Y -> ~(X^Y)
2510 Temp = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1);
2511 N0 = DAG.getNOT(dl, Temp, MVT::i1);
2512 if (!DCI.isCalledByLegalizer())
2513 DCI.AddToWorklist(Temp.getNode());
2515 case ISD::SETNE: // X != Y --> (X^Y)
2516 N0 = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1);
2518 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> ~X & Y
2519 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> ~X & Y
2520 Temp = DAG.getNOT(dl, N0, MVT::i1);
2521 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N1, Temp);
2522 if (!DCI.isCalledByLegalizer())
2523 DCI.AddToWorklist(Temp.getNode());
2525 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> ~Y & X
2526 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> ~Y & X
2527 Temp = DAG.getNOT(dl, N1, MVT::i1);
2528 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N0, Temp);
2529 if (!DCI.isCalledByLegalizer())
2530 DCI.AddToWorklist(Temp.getNode());
2532 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> ~X | Y
2533 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> ~X | Y
2534 Temp = DAG.getNOT(dl, N0, MVT::i1);
2535 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N1, Temp);
2536 if (!DCI.isCalledByLegalizer())
2537 DCI.AddToWorklist(Temp.getNode());
2539 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> ~Y | X
2540 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> ~Y | X
2541 Temp = DAG.getNOT(dl, N1, MVT::i1);
2542 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N0, Temp);
2545 if (VT != MVT::i1) {
2546 if (!DCI.isCalledByLegalizer())
2547 DCI.AddToWorklist(N0.getNode());
2548 // FIXME: If running after legalize, we probably can't do this.
2549 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, N0);
2554 // Could not fold it.
2558 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
2559 /// node is a GlobalAddress + offset.
2560 bool TargetLowering::isGAPlusOffset(SDNode *N, const GlobalValue *&GA,
2561 int64_t &Offset) const {
2562 if (isa<GlobalAddressSDNode>(N)) {
2563 GlobalAddressSDNode *GASD = cast<GlobalAddressSDNode>(N);
2564 GA = GASD->getGlobal();
2565 Offset += GASD->getOffset();
2569 if (N->getOpcode() == ISD::ADD) {
2570 SDValue N1 = N->getOperand(0);
2571 SDValue N2 = N->getOperand(1);
2572 if (isGAPlusOffset(N1.getNode(), GA, Offset)) {
2573 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
2575 Offset += V->getSExtValue();
2578 } else if (isGAPlusOffset(N2.getNode(), GA, Offset)) {
2579 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
2581 Offset += V->getSExtValue();
2591 SDValue TargetLowering::
2592 PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
2593 // Default implementation: no optimization.
2597 //===----------------------------------------------------------------------===//
2598 // Inline Assembler Implementation Methods
2599 //===----------------------------------------------------------------------===//
2602 TargetLowering::ConstraintType
2603 TargetLowering::getConstraintType(const std::string &Constraint) const {
2604 // FIXME: lots more standard ones to handle.
2605 if (Constraint.size() == 1) {
2606 switch (Constraint[0]) {
2608 case 'r': return C_RegisterClass;
2610 case 'o': // offsetable
2611 case 'V': // not offsetable
2613 case 'i': // Simple Integer or Relocatable Constant
2614 case 'n': // Simple Integer
2615 case 'E': // Floating Point Constant
2616 case 'F': // Floating Point Constant
2617 case 's': // Relocatable Constant
2618 case 'p': // Address.
2619 case 'X': // Allow ANY value.
2620 case 'I': // Target registers.
2634 if (Constraint.size() > 1 && Constraint[0] == '{' &&
2635 Constraint[Constraint.size()-1] == '}')
2640 /// LowerXConstraint - try to replace an X constraint, which matches anything,
2641 /// with another that has more specific requirements based on the type of the
2642 /// corresponding operand.
2643 const char *TargetLowering::LowerXConstraint(EVT ConstraintVT) const{
2644 if (ConstraintVT.isInteger())
2646 if (ConstraintVT.isFloatingPoint())
2647 return "f"; // works for many targets
2651 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
2652 /// vector. If it is invalid, don't add anything to Ops.
2653 void TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
2654 char ConstraintLetter,
2655 std::vector<SDValue> &Ops,
2656 SelectionDAG &DAG) const {
2657 switch (ConstraintLetter) {
2659 case 'X': // Allows any operand; labels (basic block) use this.
2660 if (Op.getOpcode() == ISD::BasicBlock) {
2665 case 'i': // Simple Integer or Relocatable Constant
2666 case 'n': // Simple Integer
2667 case 's': { // Relocatable Constant
2668 // These operands are interested in values of the form (GV+C), where C may
2669 // be folded in as an offset of GV, or it may be explicitly added. Also, it
2670 // is possible and fine if either GV or C are missing.
2671 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2672 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
2674 // If we have "(add GV, C)", pull out GV/C
2675 if (Op.getOpcode() == ISD::ADD) {
2676 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
2677 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
2678 if (C == 0 || GA == 0) {
2679 C = dyn_cast<ConstantSDNode>(Op.getOperand(0));
2680 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1));
2682 if (C == 0 || GA == 0)
2686 // If we find a valid operand, map to the TargetXXX version so that the
2687 // value itself doesn't get selected.
2688 if (GA) { // Either &GV or &GV+C
2689 if (ConstraintLetter != 'n') {
2690 int64_t Offs = GA->getOffset();
2691 if (C) Offs += C->getZExtValue();
2692 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(),
2693 C ? C->getDebugLoc() : DebugLoc(),
2694 Op.getValueType(), Offs));
2698 if (C) { // just C, no GV.
2699 // Simple constants are not allowed for 's'.
2700 if (ConstraintLetter != 's') {
2701 // gcc prints these as sign extended. Sign extend value to 64 bits
2702 // now; without this it would get ZExt'd later in
2703 // ScheduleDAGSDNodes::EmitNode, which is very generic.
2704 Ops.push_back(DAG.getTargetConstant(C->getAPIntValue().getSExtValue(),
2714 std::vector<unsigned> TargetLowering::
2715 getRegClassForInlineAsmConstraint(const std::string &Constraint,
2717 return std::vector<unsigned>();
2721 std::pair<unsigned, const TargetRegisterClass*> TargetLowering::
2722 getRegForInlineAsmConstraint(const std::string &Constraint,
2724 if (Constraint[0] != '{')
2725 return std::make_pair(0u, static_cast<TargetRegisterClass*>(0));
2726 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?");
2728 // Remove the braces from around the name.
2729 StringRef RegName(Constraint.data()+1, Constraint.size()-2);
2731 // Figure out which register class contains this reg.
2732 const TargetRegisterInfo *RI = TM.getRegisterInfo();
2733 for (TargetRegisterInfo::regclass_iterator RCI = RI->regclass_begin(),
2734 E = RI->regclass_end(); RCI != E; ++RCI) {
2735 const TargetRegisterClass *RC = *RCI;
2737 // If none of the value types for this register class are valid, we
2738 // can't use it. For example, 64-bit reg classes on 32-bit targets.
2739 bool isLegal = false;
2740 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
2742 if (isTypeLegal(*I)) {
2748 if (!isLegal) continue;
2750 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end();
2752 if (RegName.equals_lower(RI->getName(*I)))
2753 return std::make_pair(*I, RC);
2757 return std::make_pair(0u, static_cast<const TargetRegisterClass*>(0));
2760 //===----------------------------------------------------------------------===//
2761 // Constraint Selection.
2763 /// isMatchingInputConstraint - Return true of this is an input operand that is
2764 /// a matching constraint like "4".
2765 bool TargetLowering::AsmOperandInfo::isMatchingInputConstraint() const {
2766 assert(!ConstraintCode.empty() && "No known constraint!");
2767 return isdigit(ConstraintCode[0]);
2770 /// getMatchedOperand - If this is an input matching constraint, this method
2771 /// returns the output operand it matches.
2772 unsigned TargetLowering::AsmOperandInfo::getMatchedOperand() const {
2773 assert(!ConstraintCode.empty() && "No known constraint!");
2774 return atoi(ConstraintCode.c_str());
2778 /// ParseConstraints - Split up the constraint string from the inline
2779 /// assembly value into the specific constraints and their prefixes,
2780 /// and also tie in the associated operand values.
2781 /// If this returns an empty vector, and if the constraint string itself
2782 /// isn't empty, there was an error parsing.
2783 TargetLowering::AsmOperandInfoVector TargetLowering::ParseConstraints(
2784 ImmutableCallSite CS) const {
2785 /// ConstraintOperands - Information about all of the constraints.
2786 AsmOperandInfoVector ConstraintOperands;
2787 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
2788 unsigned maCount = 0; // Largest number of multiple alternative constraints.
2790 // Do a prepass over the constraints, canonicalizing them, and building up the
2791 // ConstraintOperands list.
2792 InlineAsm::ConstraintInfoVector
2793 ConstraintInfos = IA->ParseConstraints();
2795 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
2796 unsigned ResNo = 0; // ResNo - The result number of the next output.
2798 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
2799 ConstraintOperands.push_back(AsmOperandInfo(ConstraintInfos[i]));
2800 AsmOperandInfo &OpInfo = ConstraintOperands.back();
2802 // Update multiple alternative constraint count.
2803 if (OpInfo.multipleAlternatives.size() > maCount)
2804 maCount = OpInfo.multipleAlternatives.size();
2806 OpInfo.ConstraintVT = MVT::Other;
2808 // Compute the value type for each operand.
2809 switch (OpInfo.Type) {
2810 case InlineAsm::isOutput:
2811 // Indirect outputs just consume an argument.
2812 if (OpInfo.isIndirect) {
2813 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
2817 // The return value of the call is this value. As such, there is no
2818 // corresponding argument.
2819 assert(!CS.getType()->isVoidTy() &&
2821 if (const StructType *STy = dyn_cast<StructType>(CS.getType())) {
2822 OpInfo.ConstraintVT = getValueType(STy->getElementType(ResNo));
2824 assert(ResNo == 0 && "Asm only has one result!");
2825 OpInfo.ConstraintVT = getValueType(CS.getType());
2829 case InlineAsm::isInput:
2830 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
2832 case InlineAsm::isClobber:
2837 if (OpInfo.CallOperandVal) {
2838 const llvm::Type *OpTy = OpInfo.CallOperandVal->getType();
2839 if (OpInfo.isIndirect) {
2840 const llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy);
2842 report_fatal_error("Indirect operand for inline asm not a pointer!");
2843 OpTy = PtrTy->getElementType();
2846 // Look for vector wrapped in a struct. e.g. { <16 x i8> }.
2847 if (const StructType *STy = dyn_cast<StructType>(OpTy))
2848 if (STy->getNumElements() == 1)
2849 OpTy = STy->getElementType(0);
2851 // If OpTy is not a single value, it may be a struct/union that we
2852 // can tile with integers.
2853 if (!OpTy->isSingleValueType() && OpTy->isSized()) {
2854 unsigned BitSize = TD->getTypeSizeInBits(OpTy);
2863 OpInfo.ConstraintVT =
2864 EVT::getEVT(IntegerType::get(OpTy->getContext(), BitSize), true);
2867 } else if (dyn_cast<PointerType>(OpTy)) {
2868 OpInfo.ConstraintVT = MVT::getIntegerVT(8*TD->getPointerSize());
2870 OpInfo.ConstraintVT = EVT::getEVT(OpTy, true);
2875 // If we have multiple alternative constraints, select the best alternative.
2876 if (ConstraintInfos.size()) {
2878 unsigned bestMAIndex = 0;
2879 int bestWeight = -1;
2880 // weight: -1 = invalid match, and 0 = so-so match to 5 = good match.
2883 // Compute the sums of the weights for each alternative, keeping track
2884 // of the best (highest weight) one so far.
2885 for (maIndex = 0; maIndex < maCount; ++maIndex) {
2887 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size();
2888 cIndex != eIndex; ++cIndex) {
2889 AsmOperandInfo& OpInfo = ConstraintOperands[cIndex];
2890 if (OpInfo.Type == InlineAsm::isClobber)
2893 // If this is an output operand with a matching input operand,
2894 // look up the matching input. If their types mismatch, e.g. one
2895 // is an integer, the other is floating point, or their sizes are
2896 // different, flag it as an maCantMatch.
2897 if (OpInfo.hasMatchingInput()) {
2898 AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
2899 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
2900 if ((OpInfo.ConstraintVT.isInteger() !=
2901 Input.ConstraintVT.isInteger()) ||
2902 (OpInfo.ConstraintVT.getSizeInBits() !=
2903 Input.ConstraintVT.getSizeInBits())) {
2904 weightSum = -1; // Can't match.
2909 weight = getMultipleConstraintMatchWeight(OpInfo, maIndex);
2914 weightSum += weight;
2917 if (weightSum > bestWeight) {
2918 bestWeight = weightSum;
2919 bestMAIndex = maIndex;
2923 // Now select chosen alternative in each constraint.
2924 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size();
2925 cIndex != eIndex; ++cIndex) {
2926 AsmOperandInfo& cInfo = ConstraintOperands[cIndex];
2927 if (cInfo.Type == InlineAsm::isClobber)
2929 cInfo.selectAlternative(bestMAIndex);
2934 // Check and hook up tied operands, choose constraint code to use.
2935 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size();
2936 cIndex != eIndex; ++cIndex) {
2937 AsmOperandInfo& OpInfo = ConstraintOperands[cIndex];
2939 // If this is an output operand with a matching input operand, look up the
2940 // matching input. If their types mismatch, e.g. one is an integer, the
2941 // other is floating point, or their sizes are different, flag it as an
2943 if (OpInfo.hasMatchingInput()) {
2944 AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
2946 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
2947 if ((OpInfo.ConstraintVT.isInteger() !=
2948 Input.ConstraintVT.isInteger()) ||
2949 (OpInfo.ConstraintVT.getSizeInBits() !=
2950 Input.ConstraintVT.getSizeInBits())) {
2951 report_fatal_error("Unsupported asm: input constraint"
2952 " with a matching output constraint of"
2953 " incompatible type!");
2960 return ConstraintOperands;
2964 /// getConstraintGenerality - Return an integer indicating how general CT
2966 static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) {
2968 default: llvm_unreachable("Unknown constraint type!");
2969 case TargetLowering::C_Other:
2970 case TargetLowering::C_Unknown:
2972 case TargetLowering::C_Register:
2974 case TargetLowering::C_RegisterClass:
2976 case TargetLowering::C_Memory:
2981 /// Examine constraint type and operand type and determine a weight value.
2982 /// This object must already have been set up with the operand type
2983 /// and the current alternative constraint selected.
2984 TargetLowering::ConstraintWeight
2985 TargetLowering::getMultipleConstraintMatchWeight(
2986 AsmOperandInfo &info, int maIndex) const {
2987 InlineAsm::ConstraintCodeVector *rCodes;
2988 if (maIndex >= (int)info.multipleAlternatives.size())
2989 rCodes = &info.Codes;
2991 rCodes = &info.multipleAlternatives[maIndex].Codes;
2992 ConstraintWeight BestWeight = CW_Invalid;
2994 // Loop over the options, keeping track of the most general one.
2995 for (unsigned i = 0, e = rCodes->size(); i != e; ++i) {
2996 ConstraintWeight weight =
2997 getSingleConstraintMatchWeight(info, (*rCodes)[i].c_str());
2998 if (weight > BestWeight)
2999 BestWeight = weight;
3005 /// Examine constraint type and operand type and determine a weight value.
3006 /// This object must already have been set up with the operand type
3007 /// and the current alternative constraint selected.
3008 TargetLowering::ConstraintWeight
3009 TargetLowering::getSingleConstraintMatchWeight(
3010 AsmOperandInfo &info, const char *constraint) const {
3011 ConstraintWeight weight = CW_Invalid;
3012 Value *CallOperandVal = info.CallOperandVal;
3013 // If we don't have a value, we can't do a match,
3014 // but allow it at the lowest weight.
3015 if (CallOperandVal == NULL)
3017 // Look at the constraint type.
3018 switch (*constraint) {
3019 case 'i': // immediate integer.
3020 case 'n': // immediate integer with a known value.
3021 if (isa<ConstantInt>(CallOperandVal))
3022 weight = CW_Constant;
3024 case 's': // non-explicit intregal immediate.
3025 if (isa<GlobalValue>(CallOperandVal))
3026 weight = CW_Constant;
3028 case 'E': // immediate float if host format.
3029 case 'F': // immediate float.
3030 if (isa<ConstantFP>(CallOperandVal))
3031 weight = CW_Constant;
3033 case '<': // memory operand with autodecrement.
3034 case '>': // memory operand with autoincrement.
3035 case 'm': // memory operand.
3036 case 'o': // offsettable memory operand
3037 case 'V': // non-offsettable memory operand
3040 case 'r': // general register.
3041 case 'g': // general register, memory operand or immediate integer.
3042 // note: Clang converts "g" to "imr".
3043 if (CallOperandVal->getType()->isIntegerTy())
3044 weight = CW_Register;
3046 case 'X': // any operand.
3048 weight = CW_Default;
3054 /// ChooseConstraint - If there are multiple different constraints that we
3055 /// could pick for this operand (e.g. "imr") try to pick the 'best' one.
3056 /// This is somewhat tricky: constraints fall into four classes:
3057 /// Other -> immediates and magic values
3058 /// Register -> one specific register
3059 /// RegisterClass -> a group of regs
3060 /// Memory -> memory
3061 /// Ideally, we would pick the most specific constraint possible: if we have
3062 /// something that fits into a register, we would pick it. The problem here
3063 /// is that if we have something that could either be in a register or in
3064 /// memory that use of the register could cause selection of *other*
3065 /// operands to fail: they might only succeed if we pick memory. Because of
3066 /// this the heuristic we use is:
3068 /// 1) If there is an 'other' constraint, and if the operand is valid for
3069 /// that constraint, use it. This makes us take advantage of 'i'
3070 /// constraints when available.
3071 /// 2) Otherwise, pick the most general constraint present. This prefers
3072 /// 'm' over 'r', for example.
3074 static void ChooseConstraint(TargetLowering::AsmOperandInfo &OpInfo,
3075 const TargetLowering &TLI,
3076 SDValue Op, SelectionDAG *DAG) {
3077 assert(OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options");
3078 unsigned BestIdx = 0;
3079 TargetLowering::ConstraintType BestType = TargetLowering::C_Unknown;
3080 int BestGenerality = -1;
3082 // Loop over the options, keeping track of the most general one.
3083 for (unsigned i = 0, e = OpInfo.Codes.size(); i != e; ++i) {
3084 TargetLowering::ConstraintType CType =
3085 TLI.getConstraintType(OpInfo.Codes[i]);
3087 // If this is an 'other' constraint, see if the operand is valid for it.
3088 // For example, on X86 we might have an 'rI' constraint. If the operand
3089 // is an integer in the range [0..31] we want to use I (saving a load
3090 // of a register), otherwise we must use 'r'.
3091 if (CType == TargetLowering::C_Other && Op.getNode()) {
3092 assert(OpInfo.Codes[i].size() == 1 &&
3093 "Unhandled multi-letter 'other' constraint");
3094 std::vector<SDValue> ResultOps;
3095 TLI.LowerAsmOperandForConstraint(Op, OpInfo.Codes[i][0],
3097 if (!ResultOps.empty()) {
3104 // Things with matching constraints can only be registers, per gcc
3105 // documentation. This mainly affects "g" constraints.
3106 if (CType == TargetLowering::C_Memory && OpInfo.hasMatchingInput())
3109 // This constraint letter is more general than the previous one, use it.
3110 int Generality = getConstraintGenerality(CType);
3111 if (Generality > BestGenerality) {
3114 BestGenerality = Generality;
3118 OpInfo.ConstraintCode = OpInfo.Codes[BestIdx];
3119 OpInfo.ConstraintType = BestType;
3122 /// ComputeConstraintToUse - Determines the constraint code and constraint
3123 /// type to use for the specific AsmOperandInfo, setting
3124 /// OpInfo.ConstraintCode and OpInfo.ConstraintType.
3125 void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo,
3127 SelectionDAG *DAG) const {
3128 assert(!OpInfo.Codes.empty() && "Must have at least one constraint");
3130 // Single-letter constraints ('r') are very common.
3131 if (OpInfo.Codes.size() == 1) {
3132 OpInfo.ConstraintCode = OpInfo.Codes[0];
3133 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
3135 ChooseConstraint(OpInfo, *this, Op, DAG);
3138 // 'X' matches anything.
3139 if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) {
3140 // Labels and constants are handled elsewhere ('X' is the only thing
3141 // that matches labels). For Functions, the type here is the type of
3142 // the result, which is not what we want to look at; leave them alone.
3143 Value *v = OpInfo.CallOperandVal;
3144 if (isa<BasicBlock>(v) || isa<ConstantInt>(v) || isa<Function>(v)) {
3145 OpInfo.CallOperandVal = v;
3149 // Otherwise, try to resolve it to something we know about by looking at
3150 // the actual operand type.
3151 if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) {
3152 OpInfo.ConstraintCode = Repl;
3153 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
3158 //===----------------------------------------------------------------------===//
3159 // Loop Strength Reduction hooks
3160 //===----------------------------------------------------------------------===//
3162 /// isLegalAddressingMode - Return true if the addressing mode represented
3163 /// by AM is legal for this target, for a load/store of the specified type.
3164 bool TargetLowering::isLegalAddressingMode(const AddrMode &AM,
3165 const Type *Ty) const {
3166 // The default implementation of this implements a conservative RISCy, r+r and
3169 // Allows a sign-extended 16-bit immediate field.
3170 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
3173 // No global is ever allowed as a base.
3177 // Only support r+r,
3179 case 0: // "r+i" or just "i", depending on HasBaseReg.
3182 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
3184 // Otherwise we have r+r or r+i.
3187 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
3189 // Allow 2*r as r+r.
3196 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
3197 /// return a DAG expression to select that will generate the same value by
3198 /// multiplying by a magic number. See:
3199 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
3200 SDValue TargetLowering::BuildSDIV(SDNode *N, SelectionDAG &DAG,
3201 std::vector<SDNode*>* Created) const {
3202 EVT VT = N->getValueType(0);
3203 DebugLoc dl= N->getDebugLoc();
3205 // Check to see if we can do this.
3206 // FIXME: We should be more aggressive here.
3207 if (!isTypeLegal(VT))
3210 APInt d = cast<ConstantSDNode>(N->getOperand(1))->getAPIntValue();
3211 APInt::ms magics = d.magic();
3213 // Multiply the numerator (operand 0) by the magic value
3214 // FIXME: We should support doing a MUL in a wider type
3216 if (isOperationLegalOrCustom(ISD::MULHS, VT))
3217 Q = DAG.getNode(ISD::MULHS, dl, VT, N->getOperand(0),
3218 DAG.getConstant(magics.m, VT));
3219 else if (isOperationLegalOrCustom(ISD::SMUL_LOHI, VT))
3220 Q = SDValue(DAG.getNode(ISD::SMUL_LOHI, dl, DAG.getVTList(VT, VT),
3222 DAG.getConstant(magics.m, VT)).getNode(), 1);
3224 return SDValue(); // No mulhs or equvialent
3225 // If d > 0 and m < 0, add the numerator
3226 if (d.isStrictlyPositive() && magics.m.isNegative()) {
3227 Q = DAG.getNode(ISD::ADD, dl, VT, Q, N->getOperand(0));
3229 Created->push_back(Q.getNode());
3231 // If d < 0 and m > 0, subtract the numerator.
3232 if (d.isNegative() && magics.m.isStrictlyPositive()) {
3233 Q = DAG.getNode(ISD::SUB, dl, VT, Q, N->getOperand(0));
3235 Created->push_back(Q.getNode());
3237 // Shift right algebraic if shift value is nonzero
3239 Q = DAG.getNode(ISD::SRA, dl, VT, Q,
3240 DAG.getConstant(magics.s, getShiftAmountTy(Q.getValueType())));
3242 Created->push_back(Q.getNode());
3244 // Extract the sign bit and add it to the quotient
3246 DAG.getNode(ISD::SRL, dl, VT, Q, DAG.getConstant(VT.getSizeInBits()-1,
3247 getShiftAmountTy(Q.getValueType())));
3249 Created->push_back(T.getNode());
3250 return DAG.getNode(ISD::ADD, dl, VT, Q, T);
3253 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
3254 /// return a DAG expression to select that will generate the same value by
3255 /// multiplying by a magic number. See:
3256 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
3257 SDValue TargetLowering::BuildUDIV(SDNode *N, SelectionDAG &DAG,
3258 std::vector<SDNode*>* Created) const {
3259 EVT VT = N->getValueType(0);
3260 DebugLoc dl = N->getDebugLoc();
3262 // Check to see if we can do this.
3263 // FIXME: We should be more aggressive here.
3264 if (!isTypeLegal(VT))
3267 // FIXME: We should use a narrower constant when the upper
3268 // bits are known to be zero.
3269 const APInt &N1C = cast<ConstantSDNode>(N->getOperand(1))->getAPIntValue();
3270 APInt::mu magics = N1C.magicu();
3272 SDValue Q = N->getOperand(0);
3274 // If the divisor is even, we can avoid using the expensive fixup by shifting
3275 // the divided value upfront.
3276 if (magics.a != 0 && !N1C[0]) {
3277 unsigned Shift = N1C.countTrailingZeros();
3278 Q = DAG.getNode(ISD::SRL, dl, VT, Q,
3279 DAG.getConstant(Shift, getShiftAmountTy(Q.getValueType())));
3281 Created->push_back(Q.getNode());
3283 // Get magic number for the shifted divisor.
3284 magics = N1C.lshr(Shift).magicu(Shift);
3285 assert(magics.a == 0 && "Should use cheap fixup now");
3288 // Multiply the numerator (operand 0) by the magic value
3289 // FIXME: We should support doing a MUL in a wider type
3290 if (isOperationLegalOrCustom(ISD::MULHU, VT))
3291 Q = DAG.getNode(ISD::MULHU, dl, VT, Q, DAG.getConstant(magics.m, VT));
3292 else if (isOperationLegalOrCustom(ISD::UMUL_LOHI, VT))
3293 Q = SDValue(DAG.getNode(ISD::UMUL_LOHI, dl, DAG.getVTList(VT, VT), Q,
3294 DAG.getConstant(magics.m, VT)).getNode(), 1);
3296 return SDValue(); // No mulhu or equvialent
3298 Created->push_back(Q.getNode());
3300 if (magics.a == 0) {
3301 assert(magics.s < N1C.getBitWidth() &&
3302 "We shouldn't generate an undefined shift!");
3303 return DAG.getNode(ISD::SRL, dl, VT, Q,
3304 DAG.getConstant(magics.s, getShiftAmountTy(Q.getValueType())));
3306 SDValue NPQ = DAG.getNode(ISD::SUB, dl, VT, N->getOperand(0), Q);
3308 Created->push_back(NPQ.getNode());
3309 NPQ = DAG.getNode(ISD::SRL, dl, VT, NPQ,
3310 DAG.getConstant(1, getShiftAmountTy(NPQ.getValueType())));
3312 Created->push_back(NPQ.getNode());
3313 NPQ = DAG.getNode(ISD::ADD, dl, VT, NPQ, Q);
3315 Created->push_back(NPQ.getNode());
3316 return DAG.getNode(ISD::SRL, dl, VT, NPQ,
3317 DAG.getConstant(magics.s-1, getShiftAmountTy(NPQ.getValueType())));