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 // If we allow the promotion of vector elements using a flag,
825 // then return TypePromoteInteger on vector elements.
826 // First try to promote the elements of integer vectors. If no legal
827 // promotion was found, fallback to the widen-vector method.
828 if (mayPromoteElements)
829 for (unsigned nVT = i+1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
830 EVT SVT = (MVT::SimpleValueType)nVT;
831 // Promote vectors of integers to vectors with the same number
832 // of elements, with a wider element type.
833 if (SVT.getVectorElementType().getSizeInBits() > EltVT.getSizeInBits()
834 && SVT.getVectorNumElements() == NElts &&
835 isTypeLegal(SVT) && SVT.getScalarType().isInteger()) {
836 TransformToType[i] = SVT;
837 RegisterTypeForVT[i] = SVT;
838 NumRegistersForVT[i] = 1;
839 ValueTypeActions.setTypeAction(VT, TypePromoteInteger);
840 IsLegalWiderType = true;
845 if (IsLegalWiderType) continue;
847 // Try to widen the vector.
848 for (unsigned nVT = i+1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
849 EVT SVT = (MVT::SimpleValueType)nVT;
850 if (SVT.getVectorElementType() == EltVT &&
851 SVT.getVectorNumElements() > NElts &&
853 TransformToType[i] = SVT;
854 RegisterTypeForVT[i] = SVT;
855 NumRegistersForVT[i] = 1;
856 ValueTypeActions.setTypeAction(VT, TypeWidenVector);
857 IsLegalWiderType = true;
861 if (IsLegalWiderType) continue;
866 unsigned NumIntermediates;
867 NumRegistersForVT[i] =
868 getVectorTypeBreakdownMVT(VT, IntermediateVT, NumIntermediates,
870 RegisterTypeForVT[i] = RegisterVT;
872 EVT NVT = VT.getPow2VectorType();
874 // Type is already a power of 2. The default action is to split.
875 TransformToType[i] = MVT::Other;
876 unsigned NumElts = VT.getVectorNumElements();
877 ValueTypeActions.setTypeAction(VT,
878 NumElts > 1 ? TypeSplitVector : TypeScalarizeVector);
880 TransformToType[i] = NVT;
881 ValueTypeActions.setTypeAction(VT, TypeWidenVector);
885 // Determine the 'representative' register class for each value type.
886 // An representative register class is the largest (meaning one which is
887 // not a sub-register class / subreg register class) legal register class for
888 // a group of value types. For example, on i386, i8, i16, and i32
889 // representative would be GR32; while on x86_64 it's GR64.
890 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
891 const TargetRegisterClass* RRC;
893 tie(RRC, Cost) = findRepresentativeClass((MVT::SimpleValueType)i);
894 RepRegClassForVT[i] = RRC;
895 RepRegClassCostForVT[i] = Cost;
899 const char *TargetLowering::getTargetNodeName(unsigned Opcode) const {
904 MVT::SimpleValueType TargetLowering::getSetCCResultType(EVT VT) const {
905 return PointerTy.SimpleTy;
908 MVT::SimpleValueType TargetLowering::getCmpLibcallReturnType() const {
909 return MVT::i32; // return the default value
912 /// getVectorTypeBreakdown - Vector types are broken down into some number of
913 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
914 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
915 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
917 /// This method returns the number of registers needed, and the VT for each
918 /// register. It also returns the VT and quantity of the intermediate values
919 /// before they are promoted/expanded.
921 unsigned TargetLowering::getVectorTypeBreakdown(LLVMContext &Context, EVT VT,
923 unsigned &NumIntermediates,
924 EVT &RegisterVT) const {
925 unsigned NumElts = VT.getVectorNumElements();
927 // If there is a wider vector type with the same element type as this one,
928 // we should widen to that legal vector type. This handles things like
929 // <2 x float> -> <4 x float>.
930 if (NumElts != 1 && getTypeAction(Context, VT) == TypeWidenVector) {
931 RegisterVT = getTypeToTransformTo(Context, VT);
932 if (isTypeLegal(RegisterVT)) {
933 IntermediateVT = RegisterVT;
934 NumIntermediates = 1;
939 // Figure out the right, legal destination reg to copy into.
940 EVT EltTy = VT.getVectorElementType();
942 unsigned NumVectorRegs = 1;
944 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
945 // could break down into LHS/RHS like LegalizeDAG does.
946 if (!isPowerOf2_32(NumElts)) {
947 NumVectorRegs = NumElts;
951 // Divide the input until we get to a supported size. This will always
952 // end with a scalar if the target doesn't support vectors.
953 while (NumElts > 1 && !isTypeLegal(
954 EVT::getVectorVT(Context, EltTy, NumElts))) {
959 NumIntermediates = NumVectorRegs;
961 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts);
962 if (!isTypeLegal(NewVT))
964 IntermediateVT = NewVT;
966 EVT DestVT = getRegisterType(Context, NewVT);
968 if (DestVT.bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16.
969 return NumVectorRegs*(NewVT.getSizeInBits()/DestVT.getSizeInBits());
971 // Otherwise, promotion or legal types use the same number of registers as
972 // the vector decimated to the appropriate level.
973 return NumVectorRegs;
976 /// Get the EVTs and ArgFlags collections that represent the legalized return
977 /// type of the given function. This does not require a DAG or a return value,
978 /// and is suitable for use before any DAGs for the function are constructed.
979 /// TODO: Move this out of TargetLowering.cpp.
980 void llvm::GetReturnInfo(const Type* ReturnType, Attributes attr,
981 SmallVectorImpl<ISD::OutputArg> &Outs,
982 const TargetLowering &TLI,
983 SmallVectorImpl<uint64_t> *Offsets) {
984 SmallVector<EVT, 4> ValueVTs;
985 ComputeValueVTs(TLI, ReturnType, ValueVTs);
986 unsigned NumValues = ValueVTs.size();
987 if (NumValues == 0) return;
990 for (unsigned j = 0, f = NumValues; j != f; ++j) {
991 EVT VT = ValueVTs[j];
992 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
994 if (attr & Attribute::SExt)
995 ExtendKind = ISD::SIGN_EXTEND;
996 else if (attr & Attribute::ZExt)
997 ExtendKind = ISD::ZERO_EXTEND;
999 // FIXME: C calling convention requires the return type to be promoted to
1000 // at least 32-bit. But this is not necessary for non-C calling
1001 // conventions. The frontend should mark functions whose return values
1002 // require promoting with signext or zeroext attributes.
1003 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) {
1004 EVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32);
1005 if (VT.bitsLT(MinVT))
1009 unsigned NumParts = TLI.getNumRegisters(ReturnType->getContext(), VT);
1010 EVT PartVT = TLI.getRegisterType(ReturnType->getContext(), VT);
1011 unsigned PartSize = TLI.getTargetData()->getTypeAllocSize(
1012 PartVT.getTypeForEVT(ReturnType->getContext()));
1014 // 'inreg' on function refers to return value
1015 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1016 if (attr & Attribute::InReg)
1019 // Propagate extension type if any
1020 if (attr & Attribute::SExt)
1022 else if (attr & Attribute::ZExt)
1025 for (unsigned i = 0; i < NumParts; ++i) {
1026 Outs.push_back(ISD::OutputArg(Flags, PartVT, /*isFixed=*/true));
1028 Offsets->push_back(Offset);
1035 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1036 /// function arguments in the caller parameter area. This is the actual
1037 /// alignment, not its logarithm.
1038 unsigned TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1039 return TD->getCallFrameTypeAlignment(Ty);
1042 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1043 /// current function. The returned value is a member of the
1044 /// MachineJumpTableInfo::JTEntryKind enum.
1045 unsigned TargetLowering::getJumpTableEncoding() const {
1046 // In non-pic modes, just use the address of a block.
1047 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
1048 return MachineJumpTableInfo::EK_BlockAddress;
1050 // In PIC mode, if the target supports a GPRel32 directive, use it.
1051 if (getTargetMachine().getMCAsmInfo()->getGPRel32Directive() != 0)
1052 return MachineJumpTableInfo::EK_GPRel32BlockAddress;
1054 // Otherwise, use a label difference.
1055 return MachineJumpTableInfo::EK_LabelDifference32;
1058 SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1059 SelectionDAG &DAG) const {
1060 // If our PIC model is GP relative, use the global offset table as the base.
1061 if (getJumpTableEncoding() == MachineJumpTableInfo::EK_GPRel32BlockAddress)
1062 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy());
1066 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1067 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1070 TargetLowering::getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
1071 unsigned JTI,MCContext &Ctx) const{
1072 // The normal PIC reloc base is the label at the start of the jump table.
1073 return MCSymbolRefExpr::Create(MF->getJTISymbol(JTI, Ctx), Ctx);
1077 TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
1078 // Assume that everything is safe in static mode.
1079 if (getTargetMachine().getRelocationModel() == Reloc::Static)
1082 // In dynamic-no-pic mode, assume that known defined values are safe.
1083 if (getTargetMachine().getRelocationModel() == Reloc::DynamicNoPIC &&
1085 !GA->getGlobal()->isDeclaration() &&
1086 !GA->getGlobal()->isWeakForLinker())
1089 // Otherwise assume nothing is safe.
1093 //===----------------------------------------------------------------------===//
1094 // Optimization Methods
1095 //===----------------------------------------------------------------------===//
1097 /// ShrinkDemandedConstant - Check to see if the specified operand of the
1098 /// specified instruction is a constant integer. If so, check to see if there
1099 /// are any bits set in the constant that are not demanded. If so, shrink the
1100 /// constant and return true.
1101 bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDValue Op,
1102 const APInt &Demanded) {
1103 DebugLoc dl = Op.getDebugLoc();
1105 // FIXME: ISD::SELECT, ISD::SELECT_CC
1106 switch (Op.getOpcode()) {
1111 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
1112 if (!C) return false;
1114 if (Op.getOpcode() == ISD::XOR &&
1115 (C->getAPIntValue() | (~Demanded)).isAllOnesValue())
1118 // if we can expand it to have all bits set, do it
1119 if (C->getAPIntValue().intersects(~Demanded)) {
1120 EVT VT = Op.getValueType();
1121 SDValue New = DAG.getNode(Op.getOpcode(), dl, VT, Op.getOperand(0),
1122 DAG.getConstant(Demanded &
1125 return CombineTo(Op, New);
1135 /// ShrinkDemandedOp - Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the
1136 /// casts are free. This uses isZExtFree and ZERO_EXTEND for the widening
1137 /// cast, but it could be generalized for targets with other types of
1138 /// implicit widening casts.
1140 TargetLowering::TargetLoweringOpt::ShrinkDemandedOp(SDValue Op,
1142 const APInt &Demanded,
1144 assert(Op.getNumOperands() == 2 &&
1145 "ShrinkDemandedOp only supports binary operators!");
1146 assert(Op.getNode()->getNumValues() == 1 &&
1147 "ShrinkDemandedOp only supports nodes with one result!");
1149 // Don't do this if the node has another user, which may require the
1151 if (!Op.getNode()->hasOneUse())
1154 // Search for the smallest integer type with free casts to and from
1155 // Op's type. For expedience, just check power-of-2 integer types.
1156 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1157 unsigned SmallVTBits = BitWidth - Demanded.countLeadingZeros();
1158 if (!isPowerOf2_32(SmallVTBits))
1159 SmallVTBits = NextPowerOf2(SmallVTBits);
1160 for (; SmallVTBits < BitWidth; SmallVTBits = NextPowerOf2(SmallVTBits)) {
1161 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), SmallVTBits);
1162 if (TLI.isTruncateFree(Op.getValueType(), SmallVT) &&
1163 TLI.isZExtFree(SmallVT, Op.getValueType())) {
1164 // We found a type with free casts.
1165 SDValue X = DAG.getNode(Op.getOpcode(), dl, SmallVT,
1166 DAG.getNode(ISD::TRUNCATE, dl, SmallVT,
1167 Op.getNode()->getOperand(0)),
1168 DAG.getNode(ISD::TRUNCATE, dl, SmallVT,
1169 Op.getNode()->getOperand(1)));
1170 SDValue Z = DAG.getNode(ISD::ZERO_EXTEND, dl, Op.getValueType(), X);
1171 return CombineTo(Op, Z);
1177 /// SimplifyDemandedBits - Look at Op. At this point, we know that only the
1178 /// DemandedMask bits of the result of Op are ever used downstream. If we can
1179 /// use this information to simplify Op, create a new simplified DAG node and
1180 /// return true, returning the original and new nodes in Old and New. Otherwise,
1181 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
1182 /// the expression (used to simplify the caller). The KnownZero/One bits may
1183 /// only be accurate for those bits in the DemandedMask.
1184 bool TargetLowering::SimplifyDemandedBits(SDValue Op,
1185 const APInt &DemandedMask,
1188 TargetLoweringOpt &TLO,
1189 unsigned Depth) const {
1190 unsigned BitWidth = DemandedMask.getBitWidth();
1191 assert(Op.getValueType().getScalarType().getSizeInBits() == BitWidth &&
1192 "Mask size mismatches value type size!");
1193 APInt NewMask = DemandedMask;
1194 DebugLoc dl = Op.getDebugLoc();
1196 // Don't know anything.
1197 KnownZero = KnownOne = APInt(BitWidth, 0);
1199 // Other users may use these bits.
1200 if (!Op.getNode()->hasOneUse()) {
1202 // If not at the root, Just compute the KnownZero/KnownOne bits to
1203 // simplify things downstream.
1204 TLO.DAG.ComputeMaskedBits(Op, DemandedMask, KnownZero, KnownOne, Depth);
1207 // If this is the root being simplified, allow it to have multiple uses,
1208 // just set the NewMask to all bits.
1209 NewMask = APInt::getAllOnesValue(BitWidth);
1210 } else if (DemandedMask == 0) {
1211 // Not demanding any bits from Op.
1212 if (Op.getOpcode() != ISD::UNDEF)
1213 return TLO.CombineTo(Op, TLO.DAG.getUNDEF(Op.getValueType()));
1215 } else if (Depth == 6) { // Limit search depth.
1219 APInt KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut;
1220 switch (Op.getOpcode()) {
1222 // We know all of the bits for a constant!
1223 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue() & NewMask;
1224 KnownZero = ~KnownOne & NewMask;
1225 return false; // Don't fall through, will infinitely loop.
1227 // If the RHS is a constant, check to see if the LHS would be zero without
1228 // using the bits from the RHS. Below, we use knowledge about the RHS to
1229 // simplify the LHS, here we're using information from the LHS to simplify
1231 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1232 APInt LHSZero, LHSOne;
1233 // Do not increment Depth here; that can cause an infinite loop.
1234 TLO.DAG.ComputeMaskedBits(Op.getOperand(0), NewMask,
1235 LHSZero, LHSOne, Depth);
1236 // If the LHS already has zeros where RHSC does, this and is dead.
1237 if ((LHSZero & NewMask) == (~RHSC->getAPIntValue() & NewMask))
1238 return TLO.CombineTo(Op, Op.getOperand(0));
1239 // If any of the set bits in the RHS are known zero on the LHS, shrink
1241 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & NewMask))
1245 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
1246 KnownOne, TLO, Depth+1))
1248 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1249 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownZero & NewMask,
1250 KnownZero2, KnownOne2, TLO, Depth+1))
1252 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1254 // If all of the demanded bits are known one on one side, return the other.
1255 // These bits cannot contribute to the result of the 'and'.
1256 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
1257 return TLO.CombineTo(Op, Op.getOperand(0));
1258 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
1259 return TLO.CombineTo(Op, Op.getOperand(1));
1260 // If all of the demanded bits in the inputs are known zeros, return zero.
1261 if ((NewMask & (KnownZero|KnownZero2)) == NewMask)
1262 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, Op.getValueType()));
1263 // If the RHS is a constant, see if we can simplify it.
1264 if (TLO.ShrinkDemandedConstant(Op, ~KnownZero2 & NewMask))
1266 // If the operation can be done in a smaller type, do so.
1267 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1270 // Output known-1 bits are only known if set in both the LHS & RHS.
1271 KnownOne &= KnownOne2;
1272 // Output known-0 are known to be clear if zero in either the LHS | RHS.
1273 KnownZero |= KnownZero2;
1276 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
1277 KnownOne, TLO, Depth+1))
1279 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1280 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownOne & NewMask,
1281 KnownZero2, KnownOne2, TLO, Depth+1))
1283 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1285 // If all of the demanded bits are known zero on one side, return the other.
1286 // These bits cannot contribute to the result of the 'or'.
1287 if ((NewMask & ~KnownOne2 & KnownZero) == (~KnownOne2 & NewMask))
1288 return TLO.CombineTo(Op, Op.getOperand(0));
1289 if ((NewMask & ~KnownOne & KnownZero2) == (~KnownOne & NewMask))
1290 return TLO.CombineTo(Op, Op.getOperand(1));
1291 // If all of the potentially set bits on one side are known to be set on
1292 // the other side, just use the 'other' side.
1293 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
1294 return TLO.CombineTo(Op, Op.getOperand(0));
1295 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
1296 return TLO.CombineTo(Op, Op.getOperand(1));
1297 // If the RHS is a constant, see if we can simplify it.
1298 if (TLO.ShrinkDemandedConstant(Op, NewMask))
1300 // If the operation can be done in a smaller type, do so.
1301 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1304 // Output known-0 bits are only known if clear in both the LHS & RHS.
1305 KnownZero &= KnownZero2;
1306 // Output known-1 are known to be set if set in either the LHS | RHS.
1307 KnownOne |= KnownOne2;
1310 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
1311 KnownOne, TLO, Depth+1))
1313 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1314 if (SimplifyDemandedBits(Op.getOperand(0), NewMask, KnownZero2,
1315 KnownOne2, TLO, Depth+1))
1317 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1319 // If all of the demanded bits are known zero on one side, return the other.
1320 // These bits cannot contribute to the result of the 'xor'.
1321 if ((KnownZero & NewMask) == NewMask)
1322 return TLO.CombineTo(Op, Op.getOperand(0));
1323 if ((KnownZero2 & NewMask) == NewMask)
1324 return TLO.CombineTo(Op, Op.getOperand(1));
1325 // If the operation can be done in a smaller type, do so.
1326 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1329 // If all of the unknown bits are known to be zero on one side or the other
1330 // (but not both) turn this into an *inclusive* or.
1331 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1332 if ((NewMask & ~KnownZero & ~KnownZero2) == 0)
1333 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, dl, Op.getValueType(),
1337 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1338 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
1339 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1340 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
1342 // If all of the demanded bits on one side are known, and all of the set
1343 // bits on that side are also known to be set on the other side, turn this
1344 // into an AND, as we know the bits will be cleared.
1345 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1346 if ((NewMask & (KnownZero|KnownOne)) == NewMask) { // all known
1347 if ((KnownOne & KnownOne2) == KnownOne) {
1348 EVT VT = Op.getValueType();
1349 SDValue ANDC = TLO.DAG.getConstant(~KnownOne & NewMask, VT);
1350 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, dl, VT,
1351 Op.getOperand(0), ANDC));
1355 // If the RHS is a constant, see if we can simplify it.
1356 // for XOR, we prefer to force bits to 1 if they will make a -1.
1357 // if we can't force bits, try to shrink constant
1358 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1359 APInt Expanded = C->getAPIntValue() | (~NewMask);
1360 // if we can expand it to have all bits set, do it
1361 if (Expanded.isAllOnesValue()) {
1362 if (Expanded != C->getAPIntValue()) {
1363 EVT VT = Op.getValueType();
1364 SDValue New = TLO.DAG.getNode(Op.getOpcode(), dl,VT, Op.getOperand(0),
1365 TLO.DAG.getConstant(Expanded, VT));
1366 return TLO.CombineTo(Op, New);
1368 // if it already has all the bits set, nothing to change
1369 // but don't shrink either!
1370 } else if (TLO.ShrinkDemandedConstant(Op, NewMask)) {
1375 KnownZero = KnownZeroOut;
1376 KnownOne = KnownOneOut;
1379 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero,
1380 KnownOne, TLO, Depth+1))
1382 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero2,
1383 KnownOne2, TLO, Depth+1))
1385 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1386 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1388 // If the operands are constants, see if we can simplify them.
1389 if (TLO.ShrinkDemandedConstant(Op, NewMask))
1392 // Only known if known in both the LHS and RHS.
1393 KnownOne &= KnownOne2;
1394 KnownZero &= KnownZero2;
1396 case ISD::SELECT_CC:
1397 if (SimplifyDemandedBits(Op.getOperand(3), NewMask, KnownZero,
1398 KnownOne, TLO, Depth+1))
1400 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero2,
1401 KnownOne2, TLO, Depth+1))
1403 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1404 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1406 // If the operands are constants, see if we can simplify them.
1407 if (TLO.ShrinkDemandedConstant(Op, NewMask))
1410 // Only known if known in both the LHS and RHS.
1411 KnownOne &= KnownOne2;
1412 KnownZero &= KnownZero2;
1415 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1416 unsigned ShAmt = SA->getZExtValue();
1417 SDValue InOp = Op.getOperand(0);
1419 // If the shift count is an invalid immediate, don't do anything.
1420 if (ShAmt >= BitWidth)
1423 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a
1424 // single shift. We can do this if the bottom bits (which are shifted
1425 // out) are never demanded.
1426 if (InOp.getOpcode() == ISD::SRL &&
1427 isa<ConstantSDNode>(InOp.getOperand(1))) {
1428 if (ShAmt && (NewMask & APInt::getLowBitsSet(BitWidth, ShAmt)) == 0) {
1429 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
1430 unsigned Opc = ISD::SHL;
1431 int Diff = ShAmt-C1;
1438 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
1439 EVT VT = Op.getValueType();
1440 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT,
1441 InOp.getOperand(0), NewSA));
1445 if (SimplifyDemandedBits(InOp, NewMask.lshr(ShAmt),
1446 KnownZero, KnownOne, TLO, Depth+1))
1449 // Convert (shl (anyext x, c)) to (anyext (shl x, c)) if the high bits
1450 // are not demanded. This will likely allow the anyext to be folded away.
1451 if (InOp.getNode()->getOpcode() == ISD::ANY_EXTEND) {
1452 SDValue InnerOp = InOp.getNode()->getOperand(0);
1453 EVT InnerVT = InnerOp.getValueType();
1454 if ((APInt::getHighBitsSet(BitWidth,
1455 BitWidth - InnerVT.getSizeInBits()) &
1456 DemandedMask) == 0 &&
1457 isTypeDesirableForOp(ISD::SHL, InnerVT)) {
1458 EVT ShTy = getShiftAmountTy(InnerVT);
1459 if (!APInt(BitWidth, ShAmt).isIntN(ShTy.getSizeInBits()))
1462 TLO.DAG.getNode(ISD::SHL, dl, InnerVT, InnerOp,
1463 TLO.DAG.getConstant(ShAmt, ShTy));
1466 TLO.DAG.getNode(ISD::ANY_EXTEND, dl, Op.getValueType(),
1471 KnownZero <<= SA->getZExtValue();
1472 KnownOne <<= SA->getZExtValue();
1473 // low bits known zero.
1474 KnownZero |= APInt::getLowBitsSet(BitWidth, SA->getZExtValue());
1478 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1479 EVT VT = Op.getValueType();
1480 unsigned ShAmt = SA->getZExtValue();
1481 unsigned VTSize = VT.getSizeInBits();
1482 SDValue InOp = Op.getOperand(0);
1484 // If the shift count is an invalid immediate, don't do anything.
1485 if (ShAmt >= BitWidth)
1488 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a
1489 // single shift. We can do this if the top bits (which are shifted out)
1490 // are never demanded.
1491 if (InOp.getOpcode() == ISD::SHL &&
1492 isa<ConstantSDNode>(InOp.getOperand(1))) {
1493 if (ShAmt && (NewMask & APInt::getHighBitsSet(VTSize, ShAmt)) == 0) {
1494 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
1495 unsigned Opc = ISD::SRL;
1496 int Diff = ShAmt-C1;
1503 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
1504 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT,
1505 InOp.getOperand(0), NewSA));
1509 // Compute the new bits that are at the top now.
1510 if (SimplifyDemandedBits(InOp, (NewMask << ShAmt),
1511 KnownZero, KnownOne, TLO, Depth+1))
1513 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1514 KnownZero = KnownZero.lshr(ShAmt);
1515 KnownOne = KnownOne.lshr(ShAmt);
1517 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
1518 KnownZero |= HighBits; // High bits known zero.
1522 // If this is an arithmetic shift right and only the low-bit is set, we can
1523 // always convert this into a logical shr, even if the shift amount is
1524 // variable. The low bit of the shift cannot be an input sign bit unless
1525 // the shift amount is >= the size of the datatype, which is undefined.
1526 if (DemandedMask == 1)
1527 return TLO.CombineTo(Op,
1528 TLO.DAG.getNode(ISD::SRL, dl, Op.getValueType(),
1529 Op.getOperand(0), Op.getOperand(1)));
1531 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1532 EVT VT = Op.getValueType();
1533 unsigned ShAmt = SA->getZExtValue();
1535 // If the shift count is an invalid immediate, don't do anything.
1536 if (ShAmt >= BitWidth)
1539 APInt InDemandedMask = (NewMask << ShAmt);
1541 // If any of the demanded bits are produced by the sign extension, we also
1542 // demand the input sign bit.
1543 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
1544 if (HighBits.intersects(NewMask))
1545 InDemandedMask |= APInt::getSignBit(VT.getScalarType().getSizeInBits());
1547 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask,
1548 KnownZero, KnownOne, TLO, Depth+1))
1550 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1551 KnownZero = KnownZero.lshr(ShAmt);
1552 KnownOne = KnownOne.lshr(ShAmt);
1554 // Handle the sign bit, adjusted to where it is now in the mask.
1555 APInt SignBit = APInt::getSignBit(BitWidth).lshr(ShAmt);
1557 // If the input sign bit is known to be zero, or if none of the top bits
1558 // are demanded, turn this into an unsigned shift right.
1559 if (KnownZero.intersects(SignBit) || (HighBits & ~NewMask) == HighBits) {
1560 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, VT,
1563 } else if (KnownOne.intersects(SignBit)) { // New bits are known one.
1564 KnownOne |= HighBits;
1568 case ISD::SIGN_EXTEND_INREG: {
1569 EVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1571 // Sign extension. Compute the demanded bits in the result that are not
1572 // present in the input.
1574 APInt::getHighBitsSet(BitWidth,
1575 BitWidth - EVT.getScalarType().getSizeInBits());
1577 // If none of the extended bits are demanded, eliminate the sextinreg.
1578 if ((NewBits & NewMask) == 0)
1579 return TLO.CombineTo(Op, Op.getOperand(0));
1582 APInt::getSignBit(EVT.getScalarType().getSizeInBits()).zext(BitWidth);
1583 APInt InputDemandedBits =
1584 APInt::getLowBitsSet(BitWidth,
1585 EVT.getScalarType().getSizeInBits()) &
1588 // Since the sign extended bits are demanded, we know that the sign
1590 InputDemandedBits |= InSignBit;
1592 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits,
1593 KnownZero, KnownOne, TLO, Depth+1))
1595 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1597 // If the sign bit of the input is known set or clear, then we know the
1598 // top bits of the result.
1600 // If the input sign bit is known zero, convert this into a zero extension.
1601 if (KnownZero.intersects(InSignBit))
1602 return TLO.CombineTo(Op,
1603 TLO.DAG.getZeroExtendInReg(Op.getOperand(0),dl,EVT));
1605 if (KnownOne.intersects(InSignBit)) { // Input sign bit known set
1606 KnownOne |= NewBits;
1607 KnownZero &= ~NewBits;
1608 } else { // Input sign bit unknown
1609 KnownZero &= ~NewBits;
1610 KnownOne &= ~NewBits;
1614 case ISD::ZERO_EXTEND: {
1615 unsigned OperandBitWidth =
1616 Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
1617 APInt InMask = NewMask.trunc(OperandBitWidth);
1619 // If none of the top bits are demanded, convert this into an any_extend.
1621 APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask;
1622 if (!NewBits.intersects(NewMask))
1623 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND, dl,
1627 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1628 KnownZero, KnownOne, TLO, Depth+1))
1630 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1631 KnownZero = KnownZero.zext(BitWidth);
1632 KnownOne = KnownOne.zext(BitWidth);
1633 KnownZero |= NewBits;
1636 case ISD::SIGN_EXTEND: {
1637 EVT InVT = Op.getOperand(0).getValueType();
1638 unsigned InBits = InVT.getScalarType().getSizeInBits();
1639 APInt InMask = APInt::getLowBitsSet(BitWidth, InBits);
1640 APInt InSignBit = APInt::getBitsSet(BitWidth, InBits - 1, InBits);
1641 APInt NewBits = ~InMask & NewMask;
1643 // If none of the top bits are demanded, convert this into an any_extend.
1645 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND, dl,
1649 // Since some of the sign extended bits are demanded, we know that the sign
1651 APInt InDemandedBits = InMask & NewMask;
1652 InDemandedBits |= InSignBit;
1653 InDemandedBits = InDemandedBits.trunc(InBits);
1655 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero,
1656 KnownOne, TLO, Depth+1))
1658 KnownZero = KnownZero.zext(BitWidth);
1659 KnownOne = KnownOne.zext(BitWidth);
1661 // If the sign bit is known zero, convert this to a zero extend.
1662 if (KnownZero.intersects(InSignBit))
1663 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND, dl,
1667 // If the sign bit is known one, the top bits match.
1668 if (KnownOne.intersects(InSignBit)) {
1669 KnownOne |= NewBits;
1670 KnownZero &= ~NewBits;
1671 } else { // Otherwise, top bits aren't known.
1672 KnownOne &= ~NewBits;
1673 KnownZero &= ~NewBits;
1677 case ISD::ANY_EXTEND: {
1678 unsigned OperandBitWidth =
1679 Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
1680 APInt InMask = NewMask.trunc(OperandBitWidth);
1681 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1682 KnownZero, KnownOne, TLO, Depth+1))
1684 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1685 KnownZero = KnownZero.zext(BitWidth);
1686 KnownOne = KnownOne.zext(BitWidth);
1689 case ISD::TRUNCATE: {
1690 // Simplify the input, using demanded bit information, and compute the known
1691 // zero/one bits live out.
1692 unsigned OperandBitWidth =
1693 Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
1694 APInt TruncMask = NewMask.zext(OperandBitWidth);
1695 if (SimplifyDemandedBits(Op.getOperand(0), TruncMask,
1696 KnownZero, KnownOne, TLO, Depth+1))
1698 KnownZero = KnownZero.trunc(BitWidth);
1699 KnownOne = KnownOne.trunc(BitWidth);
1701 // If the input is only used by this truncate, see if we can shrink it based
1702 // on the known demanded bits.
1703 if (Op.getOperand(0).getNode()->hasOneUse()) {
1704 SDValue In = Op.getOperand(0);
1705 switch (In.getOpcode()) {
1708 // Shrink SRL by a constant if none of the high bits shifted in are
1710 if (TLO.LegalTypes() &&
1711 !isTypeDesirableForOp(ISD::SRL, Op.getValueType()))
1712 // Do not turn (vt1 truncate (vt2 srl)) into (vt1 srl) if vt1 is
1715 ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1));
1718 SDValue Shift = In.getOperand(1);
1719 if (TLO.LegalTypes()) {
1720 uint64_t ShVal = ShAmt->getZExtValue();
1722 TLO.DAG.getConstant(ShVal, getShiftAmountTy(Op.getValueType()));
1725 APInt HighBits = APInt::getHighBitsSet(OperandBitWidth,
1726 OperandBitWidth - BitWidth);
1727 HighBits = HighBits.lshr(ShAmt->getZExtValue()).trunc(BitWidth);
1729 if (ShAmt->getZExtValue() < BitWidth && !(HighBits & NewMask)) {
1730 // None of the shifted in bits are needed. Add a truncate of the
1731 // shift input, then shift it.
1732 SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE, dl,
1735 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl,
1744 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1747 case ISD::AssertZext: {
1748 // Demand all the bits of the input that are demanded in the output.
1749 // The low bits are obvious; the high bits are demanded because we're
1750 // asserting that they're zero here.
1751 if (SimplifyDemandedBits(Op.getOperand(0), NewMask,
1752 KnownZero, KnownOne, TLO, Depth+1))
1754 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1756 EVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1757 APInt InMask = APInt::getLowBitsSet(BitWidth,
1758 VT.getSizeInBits());
1759 KnownZero |= ~InMask & NewMask;
1763 // If this is an FP->Int bitcast and if the sign bit is the only thing that
1764 // is demanded, turn this into a FGETSIGN.
1765 if (NewMask == APInt::getSignBit(Op.getValueType().getSizeInBits()) &&
1766 Op.getOperand(0).getValueType().isFloatingPoint() &&
1767 !Op.getOperand(0).getValueType().isVector()) {
1768 if (isOperationLegalOrCustom(ISD::FGETSIGN, MVT::i32)) {
1769 EVT Ty = (isOperationLegalOrCustom(ISD::FGETSIGN, Op.getValueType())) ?
1770 Op.getValueType() : MVT::i32;
1771 // Make a FGETSIGN + SHL to move the sign bit into the appropriate
1772 // place. We expect the SHL to be eliminated by other optimizations.
1773 SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, dl, Ty, Op.getOperand(0));
1774 if (Ty != Op.getValueType())
1775 Sign = TLO.DAG.getNode(ISD::ZERO_EXTEND, dl, Op.getValueType(), Sign);
1776 unsigned ShVal = Op.getValueType().getSizeInBits()-1;
1777 SDValue ShAmt = TLO.DAG.getConstant(ShVal, Op.getValueType());
1778 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl,
1787 // Add, Sub, and Mul don't demand any bits in positions beyond that
1788 // of the highest bit demanded of them.
1789 APInt LoMask = APInt::getLowBitsSet(BitWidth,
1790 BitWidth - NewMask.countLeadingZeros());
1791 if (SimplifyDemandedBits(Op.getOperand(0), LoMask, KnownZero2,
1792 KnownOne2, TLO, Depth+1))
1794 if (SimplifyDemandedBits(Op.getOperand(1), LoMask, KnownZero2,
1795 KnownOne2, TLO, Depth+1))
1797 // See if the operation should be performed at a smaller bit width.
1798 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1803 // Just use ComputeMaskedBits to compute output bits.
1804 TLO.DAG.ComputeMaskedBits(Op, NewMask, KnownZero, KnownOne, Depth);
1808 // If we know the value of all of the demanded bits, return this as a
1810 if ((NewMask & (KnownZero|KnownOne)) == NewMask)
1811 return TLO.CombineTo(Op, TLO.DAG.getConstant(KnownOne, Op.getValueType()));
1816 /// computeMaskedBitsForTargetNode - Determine which of the bits specified
1817 /// in Mask are known to be either zero or one and return them in the
1818 /// KnownZero/KnownOne bitsets.
1819 void TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
1823 const SelectionDAG &DAG,
1824 unsigned Depth) const {
1825 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1826 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1827 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1828 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1829 "Should use MaskedValueIsZero if you don't know whether Op"
1830 " is a target node!");
1831 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
1834 /// ComputeNumSignBitsForTargetNode - This method can be implemented by
1835 /// targets that want to expose additional information about sign bits to the
1837 unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
1838 unsigned Depth) const {
1839 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1840 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1841 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1842 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1843 "Should use ComputeNumSignBits if you don't know whether Op"
1844 " is a target node!");
1848 /// ValueHasExactlyOneBitSet - Test if the given value is known to have exactly
1849 /// one bit set. This differs from ComputeMaskedBits in that it doesn't need to
1850 /// determine which bit is set.
1852 static bool ValueHasExactlyOneBitSet(SDValue Val, const SelectionDAG &DAG) {
1853 // A left-shift of a constant one will have exactly one bit set, because
1854 // shifting the bit off the end is undefined.
1855 if (Val.getOpcode() == ISD::SHL)
1856 if (ConstantSDNode *C =
1857 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0)))
1858 if (C->getAPIntValue() == 1)
1861 // Similarly, a right-shift of a constant sign-bit will have exactly
1863 if (Val.getOpcode() == ISD::SRL)
1864 if (ConstantSDNode *C =
1865 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0)))
1866 if (C->getAPIntValue().isSignBit())
1869 // More could be done here, though the above checks are enough
1870 // to handle some common cases.
1872 // Fall back to ComputeMaskedBits to catch other known cases.
1873 EVT OpVT = Val.getValueType();
1874 unsigned BitWidth = OpVT.getScalarType().getSizeInBits();
1875 APInt Mask = APInt::getAllOnesValue(BitWidth);
1876 APInt KnownZero, KnownOne;
1877 DAG.ComputeMaskedBits(Val, Mask, KnownZero, KnownOne);
1878 return (KnownZero.countPopulation() == BitWidth - 1) &&
1879 (KnownOne.countPopulation() == 1);
1882 /// SimplifySetCC - Try to simplify a setcc built with the specified operands
1883 /// and cc. If it is unable to simplify it, return a null SDValue.
1885 TargetLowering::SimplifySetCC(EVT VT, SDValue N0, SDValue N1,
1886 ISD::CondCode Cond, bool foldBooleans,
1887 DAGCombinerInfo &DCI, DebugLoc dl) const {
1888 SelectionDAG &DAG = DCI.DAG;
1890 // These setcc operations always fold.
1894 case ISD::SETFALSE2: return DAG.getConstant(0, VT);
1896 case ISD::SETTRUE2: return DAG.getConstant(1, VT);
1899 // Ensure that the constant occurs on the RHS, and fold constant
1901 if (isa<ConstantSDNode>(N0.getNode()))
1902 return DAG.getSetCC(dl, VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
1904 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) {
1905 const APInt &C1 = N1C->getAPIntValue();
1907 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an
1908 // equality comparison, then we're just comparing whether X itself is
1910 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) &&
1911 N0.getOperand(0).getOpcode() == ISD::CTLZ &&
1912 N0.getOperand(1).getOpcode() == ISD::Constant) {
1914 = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
1915 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1916 ShAmt == Log2_32(N0.getValueType().getSizeInBits())) {
1917 if ((C1 == 0) == (Cond == ISD::SETEQ)) {
1918 // (srl (ctlz x), 5) == 0 -> X != 0
1919 // (srl (ctlz x), 5) != 1 -> X != 0
1922 // (srl (ctlz x), 5) != 0 -> X == 0
1923 // (srl (ctlz x), 5) == 1 -> X == 0
1926 SDValue Zero = DAG.getConstant(0, N0.getValueType());
1927 return DAG.getSetCC(dl, VT, N0.getOperand(0).getOperand(0),
1933 // Look through truncs that don't change the value of a ctpop.
1934 if (N0.hasOneUse() && N0.getOpcode() == ISD::TRUNCATE)
1935 CTPOP = N0.getOperand(0);
1937 if (CTPOP.hasOneUse() && CTPOP.getOpcode() == ISD::CTPOP &&
1938 (N0 == CTPOP || N0.getValueType().getSizeInBits() >
1939 Log2_32_Ceil(CTPOP.getValueType().getSizeInBits()))) {
1940 EVT CTVT = CTPOP.getValueType();
1941 SDValue CTOp = CTPOP.getOperand(0);
1943 // (ctpop x) u< 2 -> (x & x-1) == 0
1944 // (ctpop x) u> 1 -> (x & x-1) != 0
1945 if ((Cond == ISD::SETULT && C1 == 2) || (Cond == ISD::SETUGT && C1 == 1)){
1946 SDValue Sub = DAG.getNode(ISD::SUB, dl, CTVT, CTOp,
1947 DAG.getConstant(1, CTVT));
1948 SDValue And = DAG.getNode(ISD::AND, dl, CTVT, CTOp, Sub);
1949 ISD::CondCode CC = Cond == ISD::SETULT ? ISD::SETEQ : ISD::SETNE;
1950 return DAG.getSetCC(dl, VT, And, DAG.getConstant(0, CTVT), CC);
1953 // TODO: (ctpop x) == 1 -> x && (x & x-1) == 0 iff ctpop is illegal.
1956 // (zext x) == C --> x == (trunc C)
1957 if (DCI.isBeforeLegalize() && N0->hasOneUse() &&
1958 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1959 unsigned MinBits = N0.getValueSizeInBits();
1961 if (N0->getOpcode() == ISD::ZERO_EXTEND) {
1963 MinBits = N0->getOperand(0).getValueSizeInBits();
1964 PreZExt = N0->getOperand(0);
1965 } else if (N0->getOpcode() == ISD::AND) {
1966 // DAGCombine turns costly ZExts into ANDs
1967 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0->getOperand(1)))
1968 if ((C->getAPIntValue()+1).isPowerOf2()) {
1969 MinBits = C->getAPIntValue().countTrailingOnes();
1970 PreZExt = N0->getOperand(0);
1972 } else if (LoadSDNode *LN0 = dyn_cast<LoadSDNode>(N0)) {
1974 if (LN0->getExtensionType() == ISD::ZEXTLOAD) {
1975 MinBits = LN0->getMemoryVT().getSizeInBits();
1980 // Make sure we're not loosing bits from the constant.
1981 if (MinBits < C1.getBitWidth() && MinBits > C1.getActiveBits()) {
1982 EVT MinVT = EVT::getIntegerVT(*DAG.getContext(), MinBits);
1983 if (isTypeDesirableForOp(ISD::SETCC, MinVT)) {
1984 // Will get folded away.
1985 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, MinVT, PreZExt);
1986 SDValue C = DAG.getConstant(C1.trunc(MinBits), MinVT);
1987 return DAG.getSetCC(dl, VT, Trunc, C, Cond);
1992 // If the LHS is '(and load, const)', the RHS is 0,
1993 // the test is for equality or unsigned, and all 1 bits of the const are
1994 // in the same partial word, see if we can shorten the load.
1995 if (DCI.isBeforeLegalize() &&
1996 N0.getOpcode() == ISD::AND && C1 == 0 &&
1997 N0.getNode()->hasOneUse() &&
1998 isa<LoadSDNode>(N0.getOperand(0)) &&
1999 N0.getOperand(0).getNode()->hasOneUse() &&
2000 isa<ConstantSDNode>(N0.getOperand(1))) {
2001 LoadSDNode *Lod = cast<LoadSDNode>(N0.getOperand(0));
2003 unsigned bestWidth = 0, bestOffset = 0;
2004 if (!Lod->isVolatile() && Lod->isUnindexed()) {
2005 unsigned origWidth = N0.getValueType().getSizeInBits();
2006 unsigned maskWidth = origWidth;
2007 // We can narrow (e.g.) 16-bit extending loads on 32-bit target to
2008 // 8 bits, but have to be careful...
2009 if (Lod->getExtensionType() != ISD::NON_EXTLOAD)
2010 origWidth = Lod->getMemoryVT().getSizeInBits();
2012 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
2013 for (unsigned width = origWidth / 2; width>=8; width /= 2) {
2014 APInt newMask = APInt::getLowBitsSet(maskWidth, width);
2015 for (unsigned offset=0; offset<origWidth/width; offset++) {
2016 if ((newMask & Mask) == Mask) {
2017 if (!TD->isLittleEndian())
2018 bestOffset = (origWidth/width - offset - 1) * (width/8);
2020 bestOffset = (uint64_t)offset * (width/8);
2021 bestMask = Mask.lshr(offset * (width/8) * 8);
2025 newMask = newMask << width;
2030 EVT newVT = EVT::getIntegerVT(*DAG.getContext(), bestWidth);
2031 if (newVT.isRound()) {
2032 EVT PtrType = Lod->getOperand(1).getValueType();
2033 SDValue Ptr = Lod->getBasePtr();
2034 if (bestOffset != 0)
2035 Ptr = DAG.getNode(ISD::ADD, dl, PtrType, Lod->getBasePtr(),
2036 DAG.getConstant(bestOffset, PtrType));
2037 unsigned NewAlign = MinAlign(Lod->getAlignment(), bestOffset);
2038 SDValue NewLoad = DAG.getLoad(newVT, dl, Lod->getChain(), Ptr,
2039 Lod->getPointerInfo().getWithOffset(bestOffset),
2040 false, false, NewAlign);
2041 return DAG.getSetCC(dl, VT,
2042 DAG.getNode(ISD::AND, dl, newVT, NewLoad,
2043 DAG.getConstant(bestMask.trunc(bestWidth),
2045 DAG.getConstant(0LL, newVT), Cond);
2050 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
2051 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
2052 unsigned InSize = N0.getOperand(0).getValueType().getSizeInBits();
2054 // If the comparison constant has bits in the upper part, the
2055 // zero-extended value could never match.
2056 if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(),
2057 C1.getBitWidth() - InSize))) {
2061 case ISD::SETEQ: return DAG.getConstant(0, VT);
2064 case ISD::SETNE: return DAG.getConstant(1, VT);
2067 // True if the sign bit of C1 is set.
2068 return DAG.getConstant(C1.isNegative(), VT);
2071 // True if the sign bit of C1 isn't set.
2072 return DAG.getConstant(C1.isNonNegative(), VT);
2078 // Otherwise, we can perform the comparison with the low bits.
2086 EVT newVT = N0.getOperand(0).getValueType();
2087 if (DCI.isBeforeLegalizeOps() ||
2088 (isOperationLegal(ISD::SETCC, newVT) &&
2089 getCondCodeAction(Cond, newVT)==Legal))
2090 return DAG.getSetCC(dl, VT, N0.getOperand(0),
2091 DAG.getConstant(C1.trunc(InSize), newVT),
2096 break; // todo, be more careful with signed comparisons
2098 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
2099 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
2100 EVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
2101 unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits();
2102 EVT ExtDstTy = N0.getValueType();
2103 unsigned ExtDstTyBits = ExtDstTy.getSizeInBits();
2105 // If the constant doesn't fit into the number of bits for the source of
2106 // the sign extension, it is impossible for both sides to be equal.
2107 if (C1.getMinSignedBits() > ExtSrcTyBits)
2108 return DAG.getConstant(Cond == ISD::SETNE, VT);
2111 EVT Op0Ty = N0.getOperand(0).getValueType();
2112 if (Op0Ty == ExtSrcTy) {
2113 ZextOp = N0.getOperand(0);
2115 APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits);
2116 ZextOp = DAG.getNode(ISD::AND, dl, Op0Ty, N0.getOperand(0),
2117 DAG.getConstant(Imm, Op0Ty));
2119 if (!DCI.isCalledByLegalizer())
2120 DCI.AddToWorklist(ZextOp.getNode());
2121 // Otherwise, make this a use of a zext.
2122 return DAG.getSetCC(dl, VT, ZextOp,
2123 DAG.getConstant(C1 & APInt::getLowBitsSet(
2128 } else if ((N1C->isNullValue() || N1C->getAPIntValue() == 1) &&
2129 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
2130 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC
2131 if (N0.getOpcode() == ISD::SETCC &&
2132 isTypeLegal(VT) && VT.bitsLE(N0.getValueType())) {
2133 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getAPIntValue() != 1);
2135 return DAG.getNode(ISD::TRUNCATE, dl, VT, N0);
2136 // Invert the condition.
2137 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
2138 CC = ISD::getSetCCInverse(CC,
2139 N0.getOperand(0).getValueType().isInteger());
2140 return DAG.getSetCC(dl, VT, N0.getOperand(0), N0.getOperand(1), CC);
2143 if ((N0.getOpcode() == ISD::XOR ||
2144 (N0.getOpcode() == ISD::AND &&
2145 N0.getOperand(0).getOpcode() == ISD::XOR &&
2146 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
2147 isa<ConstantSDNode>(N0.getOperand(1)) &&
2148 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue() == 1) {
2149 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We
2150 // can only do this if the top bits are known zero.
2151 unsigned BitWidth = N0.getValueSizeInBits();
2152 if (DAG.MaskedValueIsZero(N0,
2153 APInt::getHighBitsSet(BitWidth,
2155 // Okay, get the un-inverted input value.
2157 if (N0.getOpcode() == ISD::XOR)
2158 Val = N0.getOperand(0);
2160 assert(N0.getOpcode() == ISD::AND &&
2161 N0.getOperand(0).getOpcode() == ISD::XOR);
2162 // ((X^1)&1)^1 -> X & 1
2163 Val = DAG.getNode(ISD::AND, dl, N0.getValueType(),
2164 N0.getOperand(0).getOperand(0),
2168 return DAG.getSetCC(dl, VT, Val, N1,
2169 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
2171 } else if (N1C->getAPIntValue() == 1 &&
2173 getBooleanContents() == ZeroOrOneBooleanContent)) {
2175 if (Op0.getOpcode() == ISD::TRUNCATE)
2176 Op0 = Op0.getOperand(0);
2178 if ((Op0.getOpcode() == ISD::XOR) &&
2179 Op0.getOperand(0).getOpcode() == ISD::SETCC &&
2180 Op0.getOperand(1).getOpcode() == ISD::SETCC) {
2181 // (xor (setcc), (setcc)) == / != 1 -> (setcc) != / == (setcc)
2182 Cond = (Cond == ISD::SETEQ) ? ISD::SETNE : ISD::SETEQ;
2183 return DAG.getSetCC(dl, VT, Op0.getOperand(0), Op0.getOperand(1),
2185 } else if (Op0.getOpcode() == ISD::AND &&
2186 isa<ConstantSDNode>(Op0.getOperand(1)) &&
2187 cast<ConstantSDNode>(Op0.getOperand(1))->getAPIntValue() == 1) {
2188 // If this is (X&1) == / != 1, normalize it to (X&1) != / == 0.
2189 if (Op0.getValueType().bitsGT(VT))
2190 Op0 = DAG.getNode(ISD::AND, dl, VT,
2191 DAG.getNode(ISD::TRUNCATE, dl, VT, Op0.getOperand(0)),
2192 DAG.getConstant(1, VT));
2193 else if (Op0.getValueType().bitsLT(VT))
2194 Op0 = DAG.getNode(ISD::AND, dl, VT,
2195 DAG.getNode(ISD::ANY_EXTEND, dl, VT, Op0.getOperand(0)),
2196 DAG.getConstant(1, VT));
2198 return DAG.getSetCC(dl, VT, Op0,
2199 DAG.getConstant(0, Op0.getValueType()),
2200 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
2205 APInt MinVal, MaxVal;
2206 unsigned OperandBitSize = N1C->getValueType(0).getSizeInBits();
2207 if (ISD::isSignedIntSetCC(Cond)) {
2208 MinVal = APInt::getSignedMinValue(OperandBitSize);
2209 MaxVal = APInt::getSignedMaxValue(OperandBitSize);
2211 MinVal = APInt::getMinValue(OperandBitSize);
2212 MaxVal = APInt::getMaxValue(OperandBitSize);
2215 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
2216 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
2217 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true
2218 // X >= C0 --> X > (C0-1)
2219 return DAG.getSetCC(dl, VT, N0,
2220 DAG.getConstant(C1-1, N1.getValueType()),
2221 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
2224 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
2225 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true
2226 // X <= C0 --> X < (C0+1)
2227 return DAG.getSetCC(dl, VT, N0,
2228 DAG.getConstant(C1+1, N1.getValueType()),
2229 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
2232 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
2233 return DAG.getConstant(0, VT); // X < MIN --> false
2234 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal)
2235 return DAG.getConstant(1, VT); // X >= MIN --> true
2236 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal)
2237 return DAG.getConstant(0, VT); // X > MAX --> false
2238 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal)
2239 return DAG.getConstant(1, VT); // X <= MAX --> true
2241 // Canonicalize setgt X, Min --> setne X, Min
2242 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
2243 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
2244 // Canonicalize setlt X, Max --> setne X, Max
2245 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
2246 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
2248 // If we have setult X, 1, turn it into seteq X, 0
2249 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
2250 return DAG.getSetCC(dl, VT, N0,
2251 DAG.getConstant(MinVal, N0.getValueType()),
2253 // If we have setugt X, Max-1, turn it into seteq X, Max
2254 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
2255 return DAG.getSetCC(dl, VT, N0,
2256 DAG.getConstant(MaxVal, N0.getValueType()),
2259 // If we have "setcc X, C0", check to see if we can shrink the immediate
2262 // SETUGT X, SINTMAX -> SETLT X, 0
2263 if (Cond == ISD::SETUGT &&
2264 C1 == APInt::getSignedMaxValue(OperandBitSize))
2265 return DAG.getSetCC(dl, VT, N0,
2266 DAG.getConstant(0, N1.getValueType()),
2269 // SETULT X, SINTMIN -> SETGT X, -1
2270 if (Cond == ISD::SETULT &&
2271 C1 == APInt::getSignedMinValue(OperandBitSize)) {
2272 SDValue ConstMinusOne =
2273 DAG.getConstant(APInt::getAllOnesValue(OperandBitSize),
2275 return DAG.getSetCC(dl, VT, N0, ConstMinusOne, ISD::SETGT);
2278 // Fold bit comparisons when we can.
2279 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
2280 (VT == N0.getValueType() ||
2281 (isTypeLegal(VT) && VT.bitsLE(N0.getValueType()))) &&
2282 N0.getOpcode() == ISD::AND)
2283 if (ConstantSDNode *AndRHS =
2284 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2285 EVT ShiftTy = DCI.isBeforeLegalize() ?
2286 getPointerTy() : getShiftAmountTy(N0.getValueType());
2287 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
2288 // Perform the xform if the AND RHS is a single bit.
2289 if (AndRHS->getAPIntValue().isPowerOf2()) {
2290 return DAG.getNode(ISD::TRUNCATE, dl, VT,
2291 DAG.getNode(ISD::SRL, dl, N0.getValueType(), N0,
2292 DAG.getConstant(AndRHS->getAPIntValue().logBase2(), ShiftTy)));
2294 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getAPIntValue()) {
2295 // (X & 8) == 8 --> (X & 8) >> 3
2296 // Perform the xform if C1 is a single bit.
2297 if (C1.isPowerOf2()) {
2298 return DAG.getNode(ISD::TRUNCATE, dl, VT,
2299 DAG.getNode(ISD::SRL, dl, N0.getValueType(), N0,
2300 DAG.getConstant(C1.logBase2(), ShiftTy)));
2306 if (isa<ConstantFPSDNode>(N0.getNode())) {
2307 // Constant fold or commute setcc.
2308 SDValue O = DAG.FoldSetCC(VT, N0, N1, Cond, dl);
2309 if (O.getNode()) return O;
2310 } else if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1.getNode())) {
2311 // If the RHS of an FP comparison is a constant, simplify it away in
2313 if (CFP->getValueAPF().isNaN()) {
2314 // If an operand is known to be a nan, we can fold it.
2315 switch (ISD::getUnorderedFlavor(Cond)) {
2316 default: llvm_unreachable("Unknown flavor!");
2317 case 0: // Known false.
2318 return DAG.getConstant(0, VT);
2319 case 1: // Known true.
2320 return DAG.getConstant(1, VT);
2321 case 2: // Undefined.
2322 return DAG.getUNDEF(VT);
2326 // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the
2327 // constant if knowing that the operand is non-nan is enough. We prefer to
2328 // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to
2330 if (Cond == ISD::SETO || Cond == ISD::SETUO)
2331 return DAG.getSetCC(dl, VT, N0, N0, Cond);
2333 // If the condition is not legal, see if we can find an equivalent one
2335 if (!isCondCodeLegal(Cond, N0.getValueType())) {
2336 // If the comparison was an awkward floating-point == or != and one of
2337 // the comparison operands is infinity or negative infinity, convert the
2338 // condition to a less-awkward <= or >=.
2339 if (CFP->getValueAPF().isInfinity()) {
2340 if (CFP->getValueAPF().isNegative()) {
2341 if (Cond == ISD::SETOEQ &&
2342 isCondCodeLegal(ISD::SETOLE, N0.getValueType()))
2343 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLE);
2344 if (Cond == ISD::SETUEQ &&
2345 isCondCodeLegal(ISD::SETOLE, N0.getValueType()))
2346 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULE);
2347 if (Cond == ISD::SETUNE &&
2348 isCondCodeLegal(ISD::SETUGT, N0.getValueType()))
2349 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGT);
2350 if (Cond == ISD::SETONE &&
2351 isCondCodeLegal(ISD::SETUGT, N0.getValueType()))
2352 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGT);
2354 if (Cond == ISD::SETOEQ &&
2355 isCondCodeLegal(ISD::SETOGE, N0.getValueType()))
2356 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGE);
2357 if (Cond == ISD::SETUEQ &&
2358 isCondCodeLegal(ISD::SETOGE, N0.getValueType()))
2359 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGE);
2360 if (Cond == ISD::SETUNE &&
2361 isCondCodeLegal(ISD::SETULT, N0.getValueType()))
2362 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULT);
2363 if (Cond == ISD::SETONE &&
2364 isCondCodeLegal(ISD::SETULT, N0.getValueType()))
2365 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLT);
2372 // We can always fold X == X for integer setcc's.
2373 if (N0.getValueType().isInteger())
2374 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
2375 unsigned UOF = ISD::getUnorderedFlavor(Cond);
2376 if (UOF == 2) // FP operators that are undefined on NaNs.
2377 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
2378 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
2379 return DAG.getConstant(UOF, VT);
2380 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
2381 // if it is not already.
2382 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
2383 if (NewCond != Cond)
2384 return DAG.getSetCC(dl, VT, N0, N1, NewCond);
2387 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
2388 N0.getValueType().isInteger()) {
2389 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
2390 N0.getOpcode() == ISD::XOR) {
2391 // Simplify (X+Y) == (X+Z) --> Y == Z
2392 if (N0.getOpcode() == N1.getOpcode()) {
2393 if (N0.getOperand(0) == N1.getOperand(0))
2394 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(1), Cond);
2395 if (N0.getOperand(1) == N1.getOperand(1))
2396 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(0), Cond);
2397 if (DAG.isCommutativeBinOp(N0.getOpcode())) {
2398 // If X op Y == Y op X, try other combinations.
2399 if (N0.getOperand(0) == N1.getOperand(1))
2400 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(0),
2402 if (N0.getOperand(1) == N1.getOperand(0))
2403 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(1),
2408 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) {
2409 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2410 // Turn (X+C1) == C2 --> X == C2-C1
2411 if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse()) {
2412 return DAG.getSetCC(dl, VT, N0.getOperand(0),
2413 DAG.getConstant(RHSC->getAPIntValue()-
2414 LHSR->getAPIntValue(),
2415 N0.getValueType()), Cond);
2418 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
2419 if (N0.getOpcode() == ISD::XOR)
2420 // If we know that all of the inverted bits are zero, don't bother
2421 // performing the inversion.
2422 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue()))
2424 DAG.getSetCC(dl, VT, N0.getOperand(0),
2425 DAG.getConstant(LHSR->getAPIntValue() ^
2426 RHSC->getAPIntValue(),
2431 // Turn (C1-X) == C2 --> X == C1-C2
2432 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
2433 if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse()) {
2435 DAG.getSetCC(dl, VT, N0.getOperand(1),
2436 DAG.getConstant(SUBC->getAPIntValue() -
2437 RHSC->getAPIntValue(),
2444 // Simplify (X+Z) == X --> Z == 0
2445 if (N0.getOperand(0) == N1)
2446 return DAG.getSetCC(dl, VT, N0.getOperand(1),
2447 DAG.getConstant(0, N0.getValueType()), Cond);
2448 if (N0.getOperand(1) == N1) {
2449 if (DAG.isCommutativeBinOp(N0.getOpcode()))
2450 return DAG.getSetCC(dl, VT, N0.getOperand(0),
2451 DAG.getConstant(0, N0.getValueType()), Cond);
2452 else if (N0.getNode()->hasOneUse()) {
2453 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
2454 // (Z-X) == X --> Z == X<<1
2455 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(),
2457 DAG.getConstant(1, getShiftAmountTy(N1.getValueType())));
2458 if (!DCI.isCalledByLegalizer())
2459 DCI.AddToWorklist(SH.getNode());
2460 return DAG.getSetCC(dl, VT, N0.getOperand(0), SH, Cond);
2465 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
2466 N1.getOpcode() == ISD::XOR) {
2467 // Simplify X == (X+Z) --> Z == 0
2468 if (N1.getOperand(0) == N0) {
2469 return DAG.getSetCC(dl, VT, N1.getOperand(1),
2470 DAG.getConstant(0, N1.getValueType()), Cond);
2471 } else if (N1.getOperand(1) == N0) {
2472 if (DAG.isCommutativeBinOp(N1.getOpcode())) {
2473 return DAG.getSetCC(dl, VT, N1.getOperand(0),
2474 DAG.getConstant(0, N1.getValueType()), Cond);
2475 } else if (N1.getNode()->hasOneUse()) {
2476 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
2477 // X == (Z-X) --> X<<1 == Z
2478 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(), N0,
2479 DAG.getConstant(1, getShiftAmountTy(N0.getValueType())));
2480 if (!DCI.isCalledByLegalizer())
2481 DCI.AddToWorklist(SH.getNode());
2482 return DAG.getSetCC(dl, VT, SH, N1.getOperand(0), Cond);
2487 // Simplify x&y == y to x&y != 0 if y has exactly one bit set.
2488 // Note that where y is variable and is known to have at most
2489 // one bit set (for example, if it is z&1) we cannot do this;
2490 // the expressions are not equivalent when y==0.
2491 if (N0.getOpcode() == ISD::AND)
2492 if (N0.getOperand(0) == N1 || N0.getOperand(1) == N1) {
2493 if (ValueHasExactlyOneBitSet(N1, DAG)) {
2494 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true);
2495 SDValue Zero = DAG.getConstant(0, N1.getValueType());
2496 return DAG.getSetCC(dl, VT, N0, Zero, Cond);
2499 if (N1.getOpcode() == ISD::AND)
2500 if (N1.getOperand(0) == N0 || N1.getOperand(1) == N0) {
2501 if (ValueHasExactlyOneBitSet(N0, DAG)) {
2502 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true);
2503 SDValue Zero = DAG.getConstant(0, N0.getValueType());
2504 return DAG.getSetCC(dl, VT, N1, Zero, Cond);
2509 // Fold away ALL boolean setcc's.
2511 if (N0.getValueType() == MVT::i1 && foldBooleans) {
2513 default: llvm_unreachable("Unknown integer setcc!");
2514 case ISD::SETEQ: // X == Y -> ~(X^Y)
2515 Temp = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1);
2516 N0 = DAG.getNOT(dl, Temp, MVT::i1);
2517 if (!DCI.isCalledByLegalizer())
2518 DCI.AddToWorklist(Temp.getNode());
2520 case ISD::SETNE: // X != Y --> (X^Y)
2521 N0 = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1);
2523 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> ~X & Y
2524 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> ~X & Y
2525 Temp = DAG.getNOT(dl, N0, MVT::i1);
2526 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N1, Temp);
2527 if (!DCI.isCalledByLegalizer())
2528 DCI.AddToWorklist(Temp.getNode());
2530 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> ~Y & X
2531 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> ~Y & X
2532 Temp = DAG.getNOT(dl, N1, MVT::i1);
2533 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N0, Temp);
2534 if (!DCI.isCalledByLegalizer())
2535 DCI.AddToWorklist(Temp.getNode());
2537 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> ~X | Y
2538 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> ~X | Y
2539 Temp = DAG.getNOT(dl, N0, MVT::i1);
2540 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N1, Temp);
2541 if (!DCI.isCalledByLegalizer())
2542 DCI.AddToWorklist(Temp.getNode());
2544 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> ~Y | X
2545 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> ~Y | X
2546 Temp = DAG.getNOT(dl, N1, MVT::i1);
2547 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N0, Temp);
2550 if (VT != MVT::i1) {
2551 if (!DCI.isCalledByLegalizer())
2552 DCI.AddToWorklist(N0.getNode());
2553 // FIXME: If running after legalize, we probably can't do this.
2554 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, N0);
2559 // Could not fold it.
2563 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
2564 /// node is a GlobalAddress + offset.
2565 bool TargetLowering::isGAPlusOffset(SDNode *N, const GlobalValue *&GA,
2566 int64_t &Offset) const {
2567 if (isa<GlobalAddressSDNode>(N)) {
2568 GlobalAddressSDNode *GASD = cast<GlobalAddressSDNode>(N);
2569 GA = GASD->getGlobal();
2570 Offset += GASD->getOffset();
2574 if (N->getOpcode() == ISD::ADD) {
2575 SDValue N1 = N->getOperand(0);
2576 SDValue N2 = N->getOperand(1);
2577 if (isGAPlusOffset(N1.getNode(), GA, Offset)) {
2578 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
2580 Offset += V->getSExtValue();
2583 } else if (isGAPlusOffset(N2.getNode(), GA, Offset)) {
2584 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
2586 Offset += V->getSExtValue();
2596 SDValue TargetLowering::
2597 PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
2598 // Default implementation: no optimization.
2602 //===----------------------------------------------------------------------===//
2603 // Inline Assembler Implementation Methods
2604 //===----------------------------------------------------------------------===//
2607 TargetLowering::ConstraintType
2608 TargetLowering::getConstraintType(const std::string &Constraint) const {
2609 // FIXME: lots more standard ones to handle.
2610 if (Constraint.size() == 1) {
2611 switch (Constraint[0]) {
2613 case 'r': return C_RegisterClass;
2615 case 'o': // offsetable
2616 case 'V': // not offsetable
2618 case 'i': // Simple Integer or Relocatable Constant
2619 case 'n': // Simple Integer
2620 case 'E': // Floating Point Constant
2621 case 'F': // Floating Point Constant
2622 case 's': // Relocatable Constant
2623 case 'p': // Address.
2624 case 'X': // Allow ANY value.
2625 case 'I': // Target registers.
2639 if (Constraint.size() > 1 && Constraint[0] == '{' &&
2640 Constraint[Constraint.size()-1] == '}')
2645 /// LowerXConstraint - try to replace an X constraint, which matches anything,
2646 /// with another that has more specific requirements based on the type of the
2647 /// corresponding operand.
2648 const char *TargetLowering::LowerXConstraint(EVT ConstraintVT) const{
2649 if (ConstraintVT.isInteger())
2651 if (ConstraintVT.isFloatingPoint())
2652 return "f"; // works for many targets
2656 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
2657 /// vector. If it is invalid, don't add anything to Ops.
2658 void TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
2659 std::string &Constraint,
2660 std::vector<SDValue> &Ops,
2661 SelectionDAG &DAG) const {
2663 if (Constraint.length() > 1) return;
2665 char ConstraintLetter = Constraint[0];
2666 switch (ConstraintLetter) {
2668 case 'X': // Allows any operand; labels (basic block) use this.
2669 if (Op.getOpcode() == ISD::BasicBlock) {
2674 case 'i': // Simple Integer or Relocatable Constant
2675 case 'n': // Simple Integer
2676 case 's': { // Relocatable Constant
2677 // These operands are interested in values of the form (GV+C), where C may
2678 // be folded in as an offset of GV, or it may be explicitly added. Also, it
2679 // is possible and fine if either GV or C are missing.
2680 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2681 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
2683 // If we have "(add GV, C)", pull out GV/C
2684 if (Op.getOpcode() == ISD::ADD) {
2685 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
2686 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
2687 if (C == 0 || GA == 0) {
2688 C = dyn_cast<ConstantSDNode>(Op.getOperand(0));
2689 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1));
2691 if (C == 0 || GA == 0)
2695 // If we find a valid operand, map to the TargetXXX version so that the
2696 // value itself doesn't get selected.
2697 if (GA) { // Either &GV or &GV+C
2698 if (ConstraintLetter != 'n') {
2699 int64_t Offs = GA->getOffset();
2700 if (C) Offs += C->getZExtValue();
2701 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(),
2702 C ? C->getDebugLoc() : DebugLoc(),
2703 Op.getValueType(), Offs));
2707 if (C) { // just C, no GV.
2708 // Simple constants are not allowed for 's'.
2709 if (ConstraintLetter != 's') {
2710 // gcc prints these as sign extended. Sign extend value to 64 bits
2711 // now; without this it would get ZExt'd later in
2712 // ScheduleDAGSDNodes::EmitNode, which is very generic.
2713 Ops.push_back(DAG.getTargetConstant(C->getAPIntValue().getSExtValue(),
2723 std::vector<unsigned> TargetLowering::
2724 getRegClassForInlineAsmConstraint(const std::string &Constraint,
2726 return std::vector<unsigned>();
2730 std::pair<unsigned, const TargetRegisterClass*> TargetLowering::
2731 getRegForInlineAsmConstraint(const std::string &Constraint,
2733 if (Constraint[0] != '{')
2734 return std::make_pair(0u, static_cast<TargetRegisterClass*>(0));
2735 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?");
2737 // Remove the braces from around the name.
2738 StringRef RegName(Constraint.data()+1, Constraint.size()-2);
2740 // Figure out which register class contains this reg.
2741 const TargetRegisterInfo *RI = TM.getRegisterInfo();
2742 for (TargetRegisterInfo::regclass_iterator RCI = RI->regclass_begin(),
2743 E = RI->regclass_end(); RCI != E; ++RCI) {
2744 const TargetRegisterClass *RC = *RCI;
2746 // If none of the value types for this register class are valid, we
2747 // can't use it. For example, 64-bit reg classes on 32-bit targets.
2748 bool isLegal = false;
2749 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
2751 if (isTypeLegal(*I)) {
2757 if (!isLegal) continue;
2759 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end();
2761 if (RegName.equals_lower(RI->getName(*I)))
2762 return std::make_pair(*I, RC);
2766 return std::make_pair(0u, static_cast<const TargetRegisterClass*>(0));
2769 //===----------------------------------------------------------------------===//
2770 // Constraint Selection.
2772 /// isMatchingInputConstraint - Return true of this is an input operand that is
2773 /// a matching constraint like "4".
2774 bool TargetLowering::AsmOperandInfo::isMatchingInputConstraint() const {
2775 assert(!ConstraintCode.empty() && "No known constraint!");
2776 return isdigit(ConstraintCode[0]);
2779 /// getMatchedOperand - If this is an input matching constraint, this method
2780 /// returns the output operand it matches.
2781 unsigned TargetLowering::AsmOperandInfo::getMatchedOperand() const {
2782 assert(!ConstraintCode.empty() && "No known constraint!");
2783 return atoi(ConstraintCode.c_str());
2787 /// ParseConstraints - Split up the constraint string from the inline
2788 /// assembly value into the specific constraints and their prefixes,
2789 /// and also tie in the associated operand values.
2790 /// If this returns an empty vector, and if the constraint string itself
2791 /// isn't empty, there was an error parsing.
2792 TargetLowering::AsmOperandInfoVector TargetLowering::ParseConstraints(
2793 ImmutableCallSite CS) const {
2794 /// ConstraintOperands - Information about all of the constraints.
2795 AsmOperandInfoVector ConstraintOperands;
2796 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
2797 unsigned maCount = 0; // Largest number of multiple alternative constraints.
2799 // Do a prepass over the constraints, canonicalizing them, and building up the
2800 // ConstraintOperands list.
2801 InlineAsm::ConstraintInfoVector
2802 ConstraintInfos = IA->ParseConstraints();
2804 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
2805 unsigned ResNo = 0; // ResNo - The result number of the next output.
2807 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
2808 ConstraintOperands.push_back(AsmOperandInfo(ConstraintInfos[i]));
2809 AsmOperandInfo &OpInfo = ConstraintOperands.back();
2811 // Update multiple alternative constraint count.
2812 if (OpInfo.multipleAlternatives.size() > maCount)
2813 maCount = OpInfo.multipleAlternatives.size();
2815 OpInfo.ConstraintVT = MVT::Other;
2817 // Compute the value type for each operand.
2818 switch (OpInfo.Type) {
2819 case InlineAsm::isOutput:
2820 // Indirect outputs just consume an argument.
2821 if (OpInfo.isIndirect) {
2822 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
2826 // The return value of the call is this value. As such, there is no
2827 // corresponding argument.
2828 assert(!CS.getType()->isVoidTy() &&
2830 if (const StructType *STy = dyn_cast<StructType>(CS.getType())) {
2831 OpInfo.ConstraintVT = getValueType(STy->getElementType(ResNo));
2833 assert(ResNo == 0 && "Asm only has one result!");
2834 OpInfo.ConstraintVT = getValueType(CS.getType());
2838 case InlineAsm::isInput:
2839 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
2841 case InlineAsm::isClobber:
2846 if (OpInfo.CallOperandVal) {
2847 const llvm::Type *OpTy = OpInfo.CallOperandVal->getType();
2848 if (OpInfo.isIndirect) {
2849 const llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy);
2851 report_fatal_error("Indirect operand for inline asm not a pointer!");
2852 OpTy = PtrTy->getElementType();
2855 // Look for vector wrapped in a struct. e.g. { <16 x i8> }.
2856 if (const StructType *STy = dyn_cast<StructType>(OpTy))
2857 if (STy->getNumElements() == 1)
2858 OpTy = STy->getElementType(0);
2860 // If OpTy is not a single value, it may be a struct/union that we
2861 // can tile with integers.
2862 if (!OpTy->isSingleValueType() && OpTy->isSized()) {
2863 unsigned BitSize = TD->getTypeSizeInBits(OpTy);
2872 OpInfo.ConstraintVT =
2873 EVT::getEVT(IntegerType::get(OpTy->getContext(), BitSize), true);
2876 } else if (dyn_cast<PointerType>(OpTy)) {
2877 OpInfo.ConstraintVT = MVT::getIntegerVT(8*TD->getPointerSize());
2879 OpInfo.ConstraintVT = EVT::getEVT(OpTy, true);
2884 // If we have multiple alternative constraints, select the best alternative.
2885 if (ConstraintInfos.size()) {
2887 unsigned bestMAIndex = 0;
2888 int bestWeight = -1;
2889 // weight: -1 = invalid match, and 0 = so-so match to 5 = good match.
2892 // Compute the sums of the weights for each alternative, keeping track
2893 // of the best (highest weight) one so far.
2894 for (maIndex = 0; maIndex < maCount; ++maIndex) {
2896 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size();
2897 cIndex != eIndex; ++cIndex) {
2898 AsmOperandInfo& OpInfo = ConstraintOperands[cIndex];
2899 if (OpInfo.Type == InlineAsm::isClobber)
2902 // If this is an output operand with a matching input operand,
2903 // look up the matching input. If their types mismatch, e.g. one
2904 // is an integer, the other is floating point, or their sizes are
2905 // different, flag it as an maCantMatch.
2906 if (OpInfo.hasMatchingInput()) {
2907 AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
2908 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
2909 if ((OpInfo.ConstraintVT.isInteger() !=
2910 Input.ConstraintVT.isInteger()) ||
2911 (OpInfo.ConstraintVT.getSizeInBits() !=
2912 Input.ConstraintVT.getSizeInBits())) {
2913 weightSum = -1; // Can't match.
2918 weight = getMultipleConstraintMatchWeight(OpInfo, maIndex);
2923 weightSum += weight;
2926 if (weightSum > bestWeight) {
2927 bestWeight = weightSum;
2928 bestMAIndex = maIndex;
2932 // Now select chosen alternative in each constraint.
2933 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size();
2934 cIndex != eIndex; ++cIndex) {
2935 AsmOperandInfo& cInfo = ConstraintOperands[cIndex];
2936 if (cInfo.Type == InlineAsm::isClobber)
2938 cInfo.selectAlternative(bestMAIndex);
2943 // Check and hook up tied operands, choose constraint code to use.
2944 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size();
2945 cIndex != eIndex; ++cIndex) {
2946 AsmOperandInfo& OpInfo = ConstraintOperands[cIndex];
2948 // If this is an output operand with a matching input operand, look up the
2949 // matching input. If their types mismatch, e.g. one is an integer, the
2950 // other is floating point, or their sizes are different, flag it as an
2952 if (OpInfo.hasMatchingInput()) {
2953 AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
2955 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
2956 if ((OpInfo.ConstraintVT.isInteger() !=
2957 Input.ConstraintVT.isInteger()) ||
2958 (OpInfo.ConstraintVT.getSizeInBits() !=
2959 Input.ConstraintVT.getSizeInBits())) {
2960 report_fatal_error("Unsupported asm: input constraint"
2961 " with a matching output constraint of"
2962 " incompatible type!");
2969 return ConstraintOperands;
2973 /// getConstraintGenerality - Return an integer indicating how general CT
2975 static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) {
2977 default: llvm_unreachable("Unknown constraint type!");
2978 case TargetLowering::C_Other:
2979 case TargetLowering::C_Unknown:
2981 case TargetLowering::C_Register:
2983 case TargetLowering::C_RegisterClass:
2985 case TargetLowering::C_Memory:
2990 /// Examine constraint type and operand type and determine a weight value.
2991 /// This object must already have been set up with the operand type
2992 /// and the current alternative constraint selected.
2993 TargetLowering::ConstraintWeight
2994 TargetLowering::getMultipleConstraintMatchWeight(
2995 AsmOperandInfo &info, int maIndex) const {
2996 InlineAsm::ConstraintCodeVector *rCodes;
2997 if (maIndex >= (int)info.multipleAlternatives.size())
2998 rCodes = &info.Codes;
3000 rCodes = &info.multipleAlternatives[maIndex].Codes;
3001 ConstraintWeight BestWeight = CW_Invalid;
3003 // Loop over the options, keeping track of the most general one.
3004 for (unsigned i = 0, e = rCodes->size(); i != e; ++i) {
3005 ConstraintWeight weight =
3006 getSingleConstraintMatchWeight(info, (*rCodes)[i].c_str());
3007 if (weight > BestWeight)
3008 BestWeight = weight;
3014 /// Examine constraint type and operand type and determine a weight value.
3015 /// This object must already have been set up with the operand type
3016 /// and the current alternative constraint selected.
3017 TargetLowering::ConstraintWeight
3018 TargetLowering::getSingleConstraintMatchWeight(
3019 AsmOperandInfo &info, const char *constraint) const {
3020 ConstraintWeight weight = CW_Invalid;
3021 Value *CallOperandVal = info.CallOperandVal;
3022 // If we don't have a value, we can't do a match,
3023 // but allow it at the lowest weight.
3024 if (CallOperandVal == NULL)
3026 // Look at the constraint type.
3027 switch (*constraint) {
3028 case 'i': // immediate integer.
3029 case 'n': // immediate integer with a known value.
3030 if (isa<ConstantInt>(CallOperandVal))
3031 weight = CW_Constant;
3033 case 's': // non-explicit intregal immediate.
3034 if (isa<GlobalValue>(CallOperandVal))
3035 weight = CW_Constant;
3037 case 'E': // immediate float if host format.
3038 case 'F': // immediate float.
3039 if (isa<ConstantFP>(CallOperandVal))
3040 weight = CW_Constant;
3042 case '<': // memory operand with autodecrement.
3043 case '>': // memory operand with autoincrement.
3044 case 'm': // memory operand.
3045 case 'o': // offsettable memory operand
3046 case 'V': // non-offsettable memory operand
3049 case 'r': // general register.
3050 case 'g': // general register, memory operand or immediate integer.
3051 // note: Clang converts "g" to "imr".
3052 if (CallOperandVal->getType()->isIntegerTy())
3053 weight = CW_Register;
3055 case 'X': // any operand.
3057 weight = CW_Default;
3063 /// ChooseConstraint - If there are multiple different constraints that we
3064 /// could pick for this operand (e.g. "imr") try to pick the 'best' one.
3065 /// This is somewhat tricky: constraints fall into four classes:
3066 /// Other -> immediates and magic values
3067 /// Register -> one specific register
3068 /// RegisterClass -> a group of regs
3069 /// Memory -> memory
3070 /// Ideally, we would pick the most specific constraint possible: if we have
3071 /// something that fits into a register, we would pick it. The problem here
3072 /// is that if we have something that could either be in a register or in
3073 /// memory that use of the register could cause selection of *other*
3074 /// operands to fail: they might only succeed if we pick memory. Because of
3075 /// this the heuristic we use is:
3077 /// 1) If there is an 'other' constraint, and if the operand is valid for
3078 /// that constraint, use it. This makes us take advantage of 'i'
3079 /// constraints when available.
3080 /// 2) Otherwise, pick the most general constraint present. This prefers
3081 /// 'm' over 'r', for example.
3083 static void ChooseConstraint(TargetLowering::AsmOperandInfo &OpInfo,
3084 const TargetLowering &TLI,
3085 SDValue Op, SelectionDAG *DAG) {
3086 assert(OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options");
3087 unsigned BestIdx = 0;
3088 TargetLowering::ConstraintType BestType = TargetLowering::C_Unknown;
3089 int BestGenerality = -1;
3091 // Loop over the options, keeping track of the most general one.
3092 for (unsigned i = 0, e = OpInfo.Codes.size(); i != e; ++i) {
3093 TargetLowering::ConstraintType CType =
3094 TLI.getConstraintType(OpInfo.Codes[i]);
3096 // If this is an 'other' constraint, see if the operand is valid for it.
3097 // For example, on X86 we might have an 'rI' constraint. If the operand
3098 // is an integer in the range [0..31] we want to use I (saving a load
3099 // of a register), otherwise we must use 'r'.
3100 if (CType == TargetLowering::C_Other && Op.getNode()) {
3101 assert(OpInfo.Codes[i].size() == 1 &&
3102 "Unhandled multi-letter 'other' constraint");
3103 std::vector<SDValue> ResultOps;
3104 TLI.LowerAsmOperandForConstraint(Op, OpInfo.Codes[i],
3106 if (!ResultOps.empty()) {
3113 // Things with matching constraints can only be registers, per gcc
3114 // documentation. This mainly affects "g" constraints.
3115 if (CType == TargetLowering::C_Memory && OpInfo.hasMatchingInput())
3118 // This constraint letter is more general than the previous one, use it.
3119 int Generality = getConstraintGenerality(CType);
3120 if (Generality > BestGenerality) {
3123 BestGenerality = Generality;
3127 OpInfo.ConstraintCode = OpInfo.Codes[BestIdx];
3128 OpInfo.ConstraintType = BestType;
3131 /// ComputeConstraintToUse - Determines the constraint code and constraint
3132 /// type to use for the specific AsmOperandInfo, setting
3133 /// OpInfo.ConstraintCode and OpInfo.ConstraintType.
3134 void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo,
3136 SelectionDAG *DAG) const {
3137 assert(!OpInfo.Codes.empty() && "Must have at least one constraint");
3139 // Single-letter constraints ('r') are very common.
3140 if (OpInfo.Codes.size() == 1) {
3141 OpInfo.ConstraintCode = OpInfo.Codes[0];
3142 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
3144 ChooseConstraint(OpInfo, *this, Op, DAG);
3147 // 'X' matches anything.
3148 if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) {
3149 // Labels and constants are handled elsewhere ('X' is the only thing
3150 // that matches labels). For Functions, the type here is the type of
3151 // the result, which is not what we want to look at; leave them alone.
3152 Value *v = OpInfo.CallOperandVal;
3153 if (isa<BasicBlock>(v) || isa<ConstantInt>(v) || isa<Function>(v)) {
3154 OpInfo.CallOperandVal = v;
3158 // Otherwise, try to resolve it to something we know about by looking at
3159 // the actual operand type.
3160 if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) {
3161 OpInfo.ConstraintCode = Repl;
3162 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
3167 //===----------------------------------------------------------------------===//
3168 // Loop Strength Reduction hooks
3169 //===----------------------------------------------------------------------===//
3171 /// isLegalAddressingMode - Return true if the addressing mode represented
3172 /// by AM is legal for this target, for a load/store of the specified type.
3173 bool TargetLowering::isLegalAddressingMode(const AddrMode &AM,
3174 const Type *Ty) const {
3175 // The default implementation of this implements a conservative RISCy, r+r and
3178 // Allows a sign-extended 16-bit immediate field.
3179 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
3182 // No global is ever allowed as a base.
3186 // Only support r+r,
3188 case 0: // "r+i" or just "i", depending on HasBaseReg.
3191 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
3193 // Otherwise we have r+r or r+i.
3196 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
3198 // Allow 2*r as r+r.
3205 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
3206 /// return a DAG expression to select that will generate the same value by
3207 /// multiplying by a magic number. See:
3208 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
3209 SDValue TargetLowering::BuildSDIV(SDNode *N, SelectionDAG &DAG,
3210 std::vector<SDNode*>* Created) const {
3211 EVT VT = N->getValueType(0);
3212 DebugLoc dl= N->getDebugLoc();
3214 // Check to see if we can do this.
3215 // FIXME: We should be more aggressive here.
3216 if (!isTypeLegal(VT))
3219 APInt d = cast<ConstantSDNode>(N->getOperand(1))->getAPIntValue();
3220 APInt::ms magics = d.magic();
3222 // Multiply the numerator (operand 0) by the magic value
3223 // FIXME: We should support doing a MUL in a wider type
3225 if (isOperationLegalOrCustom(ISD::MULHS, VT))
3226 Q = DAG.getNode(ISD::MULHS, dl, VT, N->getOperand(0),
3227 DAG.getConstant(magics.m, VT));
3228 else if (isOperationLegalOrCustom(ISD::SMUL_LOHI, VT))
3229 Q = SDValue(DAG.getNode(ISD::SMUL_LOHI, dl, DAG.getVTList(VT, VT),
3231 DAG.getConstant(magics.m, VT)).getNode(), 1);
3233 return SDValue(); // No mulhs or equvialent
3234 // If d > 0 and m < 0, add the numerator
3235 if (d.isStrictlyPositive() && magics.m.isNegative()) {
3236 Q = DAG.getNode(ISD::ADD, dl, VT, Q, N->getOperand(0));
3238 Created->push_back(Q.getNode());
3240 // If d < 0 and m > 0, subtract the numerator.
3241 if (d.isNegative() && magics.m.isStrictlyPositive()) {
3242 Q = DAG.getNode(ISD::SUB, dl, VT, Q, N->getOperand(0));
3244 Created->push_back(Q.getNode());
3246 // Shift right algebraic if shift value is nonzero
3248 Q = DAG.getNode(ISD::SRA, dl, VT, Q,
3249 DAG.getConstant(magics.s, getShiftAmountTy(Q.getValueType())));
3251 Created->push_back(Q.getNode());
3253 // Extract the sign bit and add it to the quotient
3255 DAG.getNode(ISD::SRL, dl, VT, Q, DAG.getConstant(VT.getSizeInBits()-1,
3256 getShiftAmountTy(Q.getValueType())));
3258 Created->push_back(T.getNode());
3259 return DAG.getNode(ISD::ADD, dl, VT, Q, T);
3262 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
3263 /// return a DAG expression to select that will generate the same value by
3264 /// multiplying by a magic number. See:
3265 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
3266 SDValue TargetLowering::BuildUDIV(SDNode *N, SelectionDAG &DAG,
3267 std::vector<SDNode*>* Created) const {
3268 EVT VT = N->getValueType(0);
3269 DebugLoc dl = N->getDebugLoc();
3271 // Check to see if we can do this.
3272 // FIXME: We should be more aggressive here.
3273 if (!isTypeLegal(VT))
3276 // FIXME: We should use a narrower constant when the upper
3277 // bits are known to be zero.
3278 const APInt &N1C = cast<ConstantSDNode>(N->getOperand(1))->getAPIntValue();
3279 APInt::mu magics = N1C.magicu();
3281 SDValue Q = N->getOperand(0);
3283 // If the divisor is even, we can avoid using the expensive fixup by shifting
3284 // the divided value upfront.
3285 if (magics.a != 0 && !N1C[0]) {
3286 unsigned Shift = N1C.countTrailingZeros();
3287 Q = DAG.getNode(ISD::SRL, dl, VT, Q,
3288 DAG.getConstant(Shift, getShiftAmountTy(Q.getValueType())));
3290 Created->push_back(Q.getNode());
3292 // Get magic number for the shifted divisor.
3293 magics = N1C.lshr(Shift).magicu(Shift);
3294 assert(magics.a == 0 && "Should use cheap fixup now");
3297 // Multiply the numerator (operand 0) by the magic value
3298 // FIXME: We should support doing a MUL in a wider type
3299 if (isOperationLegalOrCustom(ISD::MULHU, VT))
3300 Q = DAG.getNode(ISD::MULHU, dl, VT, Q, DAG.getConstant(magics.m, VT));
3301 else if (isOperationLegalOrCustom(ISD::UMUL_LOHI, VT))
3302 Q = SDValue(DAG.getNode(ISD::UMUL_LOHI, dl, DAG.getVTList(VT, VT), Q,
3303 DAG.getConstant(magics.m, VT)).getNode(), 1);
3305 return SDValue(); // No mulhu or equvialent
3307 Created->push_back(Q.getNode());
3309 if (magics.a == 0) {
3310 assert(magics.s < N1C.getBitWidth() &&
3311 "We shouldn't generate an undefined shift!");
3312 return DAG.getNode(ISD::SRL, dl, VT, Q,
3313 DAG.getConstant(magics.s, getShiftAmountTy(Q.getValueType())));
3315 SDValue NPQ = DAG.getNode(ISD::SUB, dl, VT, N->getOperand(0), Q);
3317 Created->push_back(NPQ.getNode());
3318 NPQ = DAG.getNode(ISD::SRL, dl, VT, NPQ,
3319 DAG.getConstant(1, getShiftAmountTy(NPQ.getValueType())));
3321 Created->push_back(NPQ.getNode());
3322 NPQ = DAG.getNode(ISD::ADD, dl, VT, NPQ, Q);
3324 Created->push_back(NPQ.getNode());
3325 return DAG.getNode(ISD::SRL, dl, VT, NPQ,
3326 DAG.getConstant(magics.s-1, getShiftAmountTy(NPQ.getValueType())));