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::MULO_I32] = "__mulosi4";
85 Names[RTLIB::MULO_I64] = "__mulodi4";
86 Names[RTLIB::MULO_I128] = "__muloti4";
87 Names[RTLIB::SDIV_I8] = "__divqi3";
88 Names[RTLIB::SDIV_I16] = "__divhi3";
89 Names[RTLIB::SDIV_I32] = "__divsi3";
90 Names[RTLIB::SDIV_I64] = "__divdi3";
91 Names[RTLIB::SDIV_I128] = "__divti3";
92 Names[RTLIB::UDIV_I8] = "__udivqi3";
93 Names[RTLIB::UDIV_I16] = "__udivhi3";
94 Names[RTLIB::UDIV_I32] = "__udivsi3";
95 Names[RTLIB::UDIV_I64] = "__udivdi3";
96 Names[RTLIB::UDIV_I128] = "__udivti3";
97 Names[RTLIB::SREM_I8] = "__modqi3";
98 Names[RTLIB::SREM_I16] = "__modhi3";
99 Names[RTLIB::SREM_I32] = "__modsi3";
100 Names[RTLIB::SREM_I64] = "__moddi3";
101 Names[RTLIB::SREM_I128] = "__modti3";
102 Names[RTLIB::UREM_I8] = "__umodqi3";
103 Names[RTLIB::UREM_I16] = "__umodhi3";
104 Names[RTLIB::UREM_I32] = "__umodsi3";
105 Names[RTLIB::UREM_I64] = "__umoddi3";
106 Names[RTLIB::UREM_I128] = "__umodti3";
108 // These are generally not available.
109 Names[RTLIB::SDIVREM_I8] = 0;
110 Names[RTLIB::SDIVREM_I16] = 0;
111 Names[RTLIB::SDIVREM_I32] = 0;
112 Names[RTLIB::SDIVREM_I64] = 0;
113 Names[RTLIB::SDIVREM_I128] = 0;
114 Names[RTLIB::UDIVREM_I8] = 0;
115 Names[RTLIB::UDIVREM_I16] = 0;
116 Names[RTLIB::UDIVREM_I32] = 0;
117 Names[RTLIB::UDIVREM_I64] = 0;
118 Names[RTLIB::UDIVREM_I128] = 0;
120 Names[RTLIB::NEG_I32] = "__negsi2";
121 Names[RTLIB::NEG_I64] = "__negdi2";
122 Names[RTLIB::ADD_F32] = "__addsf3";
123 Names[RTLIB::ADD_F64] = "__adddf3";
124 Names[RTLIB::ADD_F80] = "__addxf3";
125 Names[RTLIB::ADD_PPCF128] = "__gcc_qadd";
126 Names[RTLIB::SUB_F32] = "__subsf3";
127 Names[RTLIB::SUB_F64] = "__subdf3";
128 Names[RTLIB::SUB_F80] = "__subxf3";
129 Names[RTLIB::SUB_PPCF128] = "__gcc_qsub";
130 Names[RTLIB::MUL_F32] = "__mulsf3";
131 Names[RTLIB::MUL_F64] = "__muldf3";
132 Names[RTLIB::MUL_F80] = "__mulxf3";
133 Names[RTLIB::MUL_PPCF128] = "__gcc_qmul";
134 Names[RTLIB::DIV_F32] = "__divsf3";
135 Names[RTLIB::DIV_F64] = "__divdf3";
136 Names[RTLIB::DIV_F80] = "__divxf3";
137 Names[RTLIB::DIV_PPCF128] = "__gcc_qdiv";
138 Names[RTLIB::REM_F32] = "fmodf";
139 Names[RTLIB::REM_F64] = "fmod";
140 Names[RTLIB::REM_F80] = "fmodl";
141 Names[RTLIB::REM_PPCF128] = "fmodl";
142 Names[RTLIB::POWI_F32] = "__powisf2";
143 Names[RTLIB::POWI_F64] = "__powidf2";
144 Names[RTLIB::POWI_F80] = "__powixf2";
145 Names[RTLIB::POWI_PPCF128] = "__powitf2";
146 Names[RTLIB::SQRT_F32] = "sqrtf";
147 Names[RTLIB::SQRT_F64] = "sqrt";
148 Names[RTLIB::SQRT_F80] = "sqrtl";
149 Names[RTLIB::SQRT_PPCF128] = "sqrtl";
150 Names[RTLIB::LOG_F32] = "logf";
151 Names[RTLIB::LOG_F64] = "log";
152 Names[RTLIB::LOG_F80] = "logl";
153 Names[RTLIB::LOG_PPCF128] = "logl";
154 Names[RTLIB::LOG2_F32] = "log2f";
155 Names[RTLIB::LOG2_F64] = "log2";
156 Names[RTLIB::LOG2_F80] = "log2l";
157 Names[RTLIB::LOG2_PPCF128] = "log2l";
158 Names[RTLIB::LOG10_F32] = "log10f";
159 Names[RTLIB::LOG10_F64] = "log10";
160 Names[RTLIB::LOG10_F80] = "log10l";
161 Names[RTLIB::LOG10_PPCF128] = "log10l";
162 Names[RTLIB::EXP_F32] = "expf";
163 Names[RTLIB::EXP_F64] = "exp";
164 Names[RTLIB::EXP_F80] = "expl";
165 Names[RTLIB::EXP_PPCF128] = "expl";
166 Names[RTLIB::EXP2_F32] = "exp2f";
167 Names[RTLIB::EXP2_F64] = "exp2";
168 Names[RTLIB::EXP2_F80] = "exp2l";
169 Names[RTLIB::EXP2_PPCF128] = "exp2l";
170 Names[RTLIB::SIN_F32] = "sinf";
171 Names[RTLIB::SIN_F64] = "sin";
172 Names[RTLIB::SIN_F80] = "sinl";
173 Names[RTLIB::SIN_PPCF128] = "sinl";
174 Names[RTLIB::COS_F32] = "cosf";
175 Names[RTLIB::COS_F64] = "cos";
176 Names[RTLIB::COS_F80] = "cosl";
177 Names[RTLIB::COS_PPCF128] = "cosl";
178 Names[RTLIB::POW_F32] = "powf";
179 Names[RTLIB::POW_F64] = "pow";
180 Names[RTLIB::POW_F80] = "powl";
181 Names[RTLIB::POW_PPCF128] = "powl";
182 Names[RTLIB::CEIL_F32] = "ceilf";
183 Names[RTLIB::CEIL_F64] = "ceil";
184 Names[RTLIB::CEIL_F80] = "ceill";
185 Names[RTLIB::CEIL_PPCF128] = "ceill";
186 Names[RTLIB::TRUNC_F32] = "truncf";
187 Names[RTLIB::TRUNC_F64] = "trunc";
188 Names[RTLIB::TRUNC_F80] = "truncl";
189 Names[RTLIB::TRUNC_PPCF128] = "truncl";
190 Names[RTLIB::RINT_F32] = "rintf";
191 Names[RTLIB::RINT_F64] = "rint";
192 Names[RTLIB::RINT_F80] = "rintl";
193 Names[RTLIB::RINT_PPCF128] = "rintl";
194 Names[RTLIB::NEARBYINT_F32] = "nearbyintf";
195 Names[RTLIB::NEARBYINT_F64] = "nearbyint";
196 Names[RTLIB::NEARBYINT_F80] = "nearbyintl";
197 Names[RTLIB::NEARBYINT_PPCF128] = "nearbyintl";
198 Names[RTLIB::FLOOR_F32] = "floorf";
199 Names[RTLIB::FLOOR_F64] = "floor";
200 Names[RTLIB::FLOOR_F80] = "floorl";
201 Names[RTLIB::FLOOR_PPCF128] = "floorl";
202 Names[RTLIB::COPYSIGN_F32] = "copysignf";
203 Names[RTLIB::COPYSIGN_F64] = "copysign";
204 Names[RTLIB::COPYSIGN_F80] = "copysignl";
205 Names[RTLIB::COPYSIGN_PPCF128] = "copysignl";
206 Names[RTLIB::FPEXT_F32_F64] = "__extendsfdf2";
207 Names[RTLIB::FPEXT_F16_F32] = "__gnu_h2f_ieee";
208 Names[RTLIB::FPROUND_F32_F16] = "__gnu_f2h_ieee";
209 Names[RTLIB::FPROUND_F64_F32] = "__truncdfsf2";
210 Names[RTLIB::FPROUND_F80_F32] = "__truncxfsf2";
211 Names[RTLIB::FPROUND_PPCF128_F32] = "__trunctfsf2";
212 Names[RTLIB::FPROUND_F80_F64] = "__truncxfdf2";
213 Names[RTLIB::FPROUND_PPCF128_F64] = "__trunctfdf2";
214 Names[RTLIB::FPTOSINT_F32_I8] = "__fixsfqi";
215 Names[RTLIB::FPTOSINT_F32_I16] = "__fixsfhi";
216 Names[RTLIB::FPTOSINT_F32_I32] = "__fixsfsi";
217 Names[RTLIB::FPTOSINT_F32_I64] = "__fixsfdi";
218 Names[RTLIB::FPTOSINT_F32_I128] = "__fixsfti";
219 Names[RTLIB::FPTOSINT_F64_I8] = "__fixdfqi";
220 Names[RTLIB::FPTOSINT_F64_I16] = "__fixdfhi";
221 Names[RTLIB::FPTOSINT_F64_I32] = "__fixdfsi";
222 Names[RTLIB::FPTOSINT_F64_I64] = "__fixdfdi";
223 Names[RTLIB::FPTOSINT_F64_I128] = "__fixdfti";
224 Names[RTLIB::FPTOSINT_F80_I32] = "__fixxfsi";
225 Names[RTLIB::FPTOSINT_F80_I64] = "__fixxfdi";
226 Names[RTLIB::FPTOSINT_F80_I128] = "__fixxfti";
227 Names[RTLIB::FPTOSINT_PPCF128_I32] = "__fixtfsi";
228 Names[RTLIB::FPTOSINT_PPCF128_I64] = "__fixtfdi";
229 Names[RTLIB::FPTOSINT_PPCF128_I128] = "__fixtfti";
230 Names[RTLIB::FPTOUINT_F32_I8] = "__fixunssfqi";
231 Names[RTLIB::FPTOUINT_F32_I16] = "__fixunssfhi";
232 Names[RTLIB::FPTOUINT_F32_I32] = "__fixunssfsi";
233 Names[RTLIB::FPTOUINT_F32_I64] = "__fixunssfdi";
234 Names[RTLIB::FPTOUINT_F32_I128] = "__fixunssfti";
235 Names[RTLIB::FPTOUINT_F64_I8] = "__fixunsdfqi";
236 Names[RTLIB::FPTOUINT_F64_I16] = "__fixunsdfhi";
237 Names[RTLIB::FPTOUINT_F64_I32] = "__fixunsdfsi";
238 Names[RTLIB::FPTOUINT_F64_I64] = "__fixunsdfdi";
239 Names[RTLIB::FPTOUINT_F64_I128] = "__fixunsdfti";
240 Names[RTLIB::FPTOUINT_F80_I32] = "__fixunsxfsi";
241 Names[RTLIB::FPTOUINT_F80_I64] = "__fixunsxfdi";
242 Names[RTLIB::FPTOUINT_F80_I128] = "__fixunsxfti";
243 Names[RTLIB::FPTOUINT_PPCF128_I32] = "__fixunstfsi";
244 Names[RTLIB::FPTOUINT_PPCF128_I64] = "__fixunstfdi";
245 Names[RTLIB::FPTOUINT_PPCF128_I128] = "__fixunstfti";
246 Names[RTLIB::SINTTOFP_I32_F32] = "__floatsisf";
247 Names[RTLIB::SINTTOFP_I32_F64] = "__floatsidf";
248 Names[RTLIB::SINTTOFP_I32_F80] = "__floatsixf";
249 Names[RTLIB::SINTTOFP_I32_PPCF128] = "__floatsitf";
250 Names[RTLIB::SINTTOFP_I64_F32] = "__floatdisf";
251 Names[RTLIB::SINTTOFP_I64_F64] = "__floatdidf";
252 Names[RTLIB::SINTTOFP_I64_F80] = "__floatdixf";
253 Names[RTLIB::SINTTOFP_I64_PPCF128] = "__floatditf";
254 Names[RTLIB::SINTTOFP_I128_F32] = "__floattisf";
255 Names[RTLIB::SINTTOFP_I128_F64] = "__floattidf";
256 Names[RTLIB::SINTTOFP_I128_F80] = "__floattixf";
257 Names[RTLIB::SINTTOFP_I128_PPCF128] = "__floattitf";
258 Names[RTLIB::UINTTOFP_I32_F32] = "__floatunsisf";
259 Names[RTLIB::UINTTOFP_I32_F64] = "__floatunsidf";
260 Names[RTLIB::UINTTOFP_I32_F80] = "__floatunsixf";
261 Names[RTLIB::UINTTOFP_I32_PPCF128] = "__floatunsitf";
262 Names[RTLIB::UINTTOFP_I64_F32] = "__floatundisf";
263 Names[RTLIB::UINTTOFP_I64_F64] = "__floatundidf";
264 Names[RTLIB::UINTTOFP_I64_F80] = "__floatundixf";
265 Names[RTLIB::UINTTOFP_I64_PPCF128] = "__floatunditf";
266 Names[RTLIB::UINTTOFP_I128_F32] = "__floatuntisf";
267 Names[RTLIB::UINTTOFP_I128_F64] = "__floatuntidf";
268 Names[RTLIB::UINTTOFP_I128_F80] = "__floatuntixf";
269 Names[RTLIB::UINTTOFP_I128_PPCF128] = "__floatuntitf";
270 Names[RTLIB::OEQ_F32] = "__eqsf2";
271 Names[RTLIB::OEQ_F64] = "__eqdf2";
272 Names[RTLIB::UNE_F32] = "__nesf2";
273 Names[RTLIB::UNE_F64] = "__nedf2";
274 Names[RTLIB::OGE_F32] = "__gesf2";
275 Names[RTLIB::OGE_F64] = "__gedf2";
276 Names[RTLIB::OLT_F32] = "__ltsf2";
277 Names[RTLIB::OLT_F64] = "__ltdf2";
278 Names[RTLIB::OLE_F32] = "__lesf2";
279 Names[RTLIB::OLE_F64] = "__ledf2";
280 Names[RTLIB::OGT_F32] = "__gtsf2";
281 Names[RTLIB::OGT_F64] = "__gtdf2";
282 Names[RTLIB::UO_F32] = "__unordsf2";
283 Names[RTLIB::UO_F64] = "__unorddf2";
284 Names[RTLIB::O_F32] = "__unordsf2";
285 Names[RTLIB::O_F64] = "__unorddf2";
286 Names[RTLIB::MEMCPY] = "memcpy";
287 Names[RTLIB::MEMMOVE] = "memmove";
288 Names[RTLIB::MEMSET] = "memset";
289 Names[RTLIB::UNWIND_RESUME] = "_Unwind_Resume";
290 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_1] = "__sync_val_compare_and_swap_1";
291 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_2] = "__sync_val_compare_and_swap_2";
292 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_4] = "__sync_val_compare_and_swap_4";
293 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_8] = "__sync_val_compare_and_swap_8";
294 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_1] = "__sync_lock_test_and_set_1";
295 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_2] = "__sync_lock_test_and_set_2";
296 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_4] = "__sync_lock_test_and_set_4";
297 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_8] = "__sync_lock_test_and_set_8";
298 Names[RTLIB::SYNC_FETCH_AND_ADD_1] = "__sync_fetch_and_add_1";
299 Names[RTLIB::SYNC_FETCH_AND_ADD_2] = "__sync_fetch_and_add_2";
300 Names[RTLIB::SYNC_FETCH_AND_ADD_4] = "__sync_fetch_and_add_4";
301 Names[RTLIB::SYNC_FETCH_AND_ADD_8] = "__sync_fetch_and_add_8";
302 Names[RTLIB::SYNC_FETCH_AND_SUB_1] = "__sync_fetch_and_sub_1";
303 Names[RTLIB::SYNC_FETCH_AND_SUB_2] = "__sync_fetch_and_sub_2";
304 Names[RTLIB::SYNC_FETCH_AND_SUB_4] = "__sync_fetch_and_sub_4";
305 Names[RTLIB::SYNC_FETCH_AND_SUB_8] = "__sync_fetch_and_sub_8";
306 Names[RTLIB::SYNC_FETCH_AND_AND_1] = "__sync_fetch_and_and_1";
307 Names[RTLIB::SYNC_FETCH_AND_AND_2] = "__sync_fetch_and_and_2";
308 Names[RTLIB::SYNC_FETCH_AND_AND_4] = "__sync_fetch_and_and_4";
309 Names[RTLIB::SYNC_FETCH_AND_AND_8] = "__sync_fetch_and_and_8";
310 Names[RTLIB::SYNC_FETCH_AND_OR_1] = "__sync_fetch_and_or_1";
311 Names[RTLIB::SYNC_FETCH_AND_OR_2] = "__sync_fetch_and_or_2";
312 Names[RTLIB::SYNC_FETCH_AND_OR_4] = "__sync_fetch_and_or_4";
313 Names[RTLIB::SYNC_FETCH_AND_OR_8] = "__sync_fetch_and_or_8";
314 Names[RTLIB::SYNC_FETCH_AND_XOR_1] = "__sync_fetch_and_xor_1";
315 Names[RTLIB::SYNC_FETCH_AND_XOR_2] = "__sync_fetch_and_xor_2";
316 Names[RTLIB::SYNC_FETCH_AND_XOR_4] = "__sync_fetch_and-xor_4";
317 Names[RTLIB::SYNC_FETCH_AND_XOR_8] = "__sync_fetch_and_xor_8";
318 Names[RTLIB::SYNC_FETCH_AND_NAND_1] = "__sync_fetch_and_nand_1";
319 Names[RTLIB::SYNC_FETCH_AND_NAND_2] = "__sync_fetch_and_nand_2";
320 Names[RTLIB::SYNC_FETCH_AND_NAND_4] = "__sync_fetch_and_nand_4";
321 Names[RTLIB::SYNC_FETCH_AND_NAND_8] = "__sync_fetch_and_nand_8";
324 /// InitLibcallCallingConvs - Set default libcall CallingConvs.
326 static void InitLibcallCallingConvs(CallingConv::ID *CCs) {
327 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
328 CCs[i] = CallingConv::C;
332 /// getFPEXT - Return the FPEXT_*_* value for the given types, or
333 /// UNKNOWN_LIBCALL if there is none.
334 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) {
335 if (OpVT == MVT::f32) {
336 if (RetVT == MVT::f64)
337 return FPEXT_F32_F64;
340 return UNKNOWN_LIBCALL;
343 /// getFPROUND - Return the FPROUND_*_* value for the given types, or
344 /// UNKNOWN_LIBCALL if there is none.
345 RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) {
346 if (RetVT == MVT::f32) {
347 if (OpVT == MVT::f64)
348 return FPROUND_F64_F32;
349 if (OpVT == MVT::f80)
350 return FPROUND_F80_F32;
351 if (OpVT == MVT::ppcf128)
352 return FPROUND_PPCF128_F32;
353 } else if (RetVT == MVT::f64) {
354 if (OpVT == MVT::f80)
355 return FPROUND_F80_F64;
356 if (OpVT == MVT::ppcf128)
357 return FPROUND_PPCF128_F64;
360 return UNKNOWN_LIBCALL;
363 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or
364 /// UNKNOWN_LIBCALL if there is none.
365 RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) {
366 if (OpVT == MVT::f32) {
367 if (RetVT == MVT::i8)
368 return FPTOSINT_F32_I8;
369 if (RetVT == MVT::i16)
370 return FPTOSINT_F32_I16;
371 if (RetVT == MVT::i32)
372 return FPTOSINT_F32_I32;
373 if (RetVT == MVT::i64)
374 return FPTOSINT_F32_I64;
375 if (RetVT == MVT::i128)
376 return FPTOSINT_F32_I128;
377 } else if (OpVT == MVT::f64) {
378 if (RetVT == MVT::i8)
379 return FPTOSINT_F64_I8;
380 if (RetVT == MVT::i16)
381 return FPTOSINT_F64_I16;
382 if (RetVT == MVT::i32)
383 return FPTOSINT_F64_I32;
384 if (RetVT == MVT::i64)
385 return FPTOSINT_F64_I64;
386 if (RetVT == MVT::i128)
387 return FPTOSINT_F64_I128;
388 } else if (OpVT == MVT::f80) {
389 if (RetVT == MVT::i32)
390 return FPTOSINT_F80_I32;
391 if (RetVT == MVT::i64)
392 return FPTOSINT_F80_I64;
393 if (RetVT == MVT::i128)
394 return FPTOSINT_F80_I128;
395 } else if (OpVT == MVT::ppcf128) {
396 if (RetVT == MVT::i32)
397 return FPTOSINT_PPCF128_I32;
398 if (RetVT == MVT::i64)
399 return FPTOSINT_PPCF128_I64;
400 if (RetVT == MVT::i128)
401 return FPTOSINT_PPCF128_I128;
403 return UNKNOWN_LIBCALL;
406 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or
407 /// UNKNOWN_LIBCALL if there is none.
408 RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) {
409 if (OpVT == MVT::f32) {
410 if (RetVT == MVT::i8)
411 return FPTOUINT_F32_I8;
412 if (RetVT == MVT::i16)
413 return FPTOUINT_F32_I16;
414 if (RetVT == MVT::i32)
415 return FPTOUINT_F32_I32;
416 if (RetVT == MVT::i64)
417 return FPTOUINT_F32_I64;
418 if (RetVT == MVT::i128)
419 return FPTOUINT_F32_I128;
420 } else if (OpVT == MVT::f64) {
421 if (RetVT == MVT::i8)
422 return FPTOUINT_F64_I8;
423 if (RetVT == MVT::i16)
424 return FPTOUINT_F64_I16;
425 if (RetVT == MVT::i32)
426 return FPTOUINT_F64_I32;
427 if (RetVT == MVT::i64)
428 return FPTOUINT_F64_I64;
429 if (RetVT == MVT::i128)
430 return FPTOUINT_F64_I128;
431 } else if (OpVT == MVT::f80) {
432 if (RetVT == MVT::i32)
433 return FPTOUINT_F80_I32;
434 if (RetVT == MVT::i64)
435 return FPTOUINT_F80_I64;
436 if (RetVT == MVT::i128)
437 return FPTOUINT_F80_I128;
438 } else if (OpVT == MVT::ppcf128) {
439 if (RetVT == MVT::i32)
440 return FPTOUINT_PPCF128_I32;
441 if (RetVT == MVT::i64)
442 return FPTOUINT_PPCF128_I64;
443 if (RetVT == MVT::i128)
444 return FPTOUINT_PPCF128_I128;
446 return UNKNOWN_LIBCALL;
449 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or
450 /// UNKNOWN_LIBCALL if there is none.
451 RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) {
452 if (OpVT == MVT::i32) {
453 if (RetVT == MVT::f32)
454 return SINTTOFP_I32_F32;
455 else if (RetVT == MVT::f64)
456 return SINTTOFP_I32_F64;
457 else if (RetVT == MVT::f80)
458 return SINTTOFP_I32_F80;
459 else if (RetVT == MVT::ppcf128)
460 return SINTTOFP_I32_PPCF128;
461 } else if (OpVT == MVT::i64) {
462 if (RetVT == MVT::f32)
463 return SINTTOFP_I64_F32;
464 else if (RetVT == MVT::f64)
465 return SINTTOFP_I64_F64;
466 else if (RetVT == MVT::f80)
467 return SINTTOFP_I64_F80;
468 else if (RetVT == MVT::ppcf128)
469 return SINTTOFP_I64_PPCF128;
470 } else if (OpVT == MVT::i128) {
471 if (RetVT == MVT::f32)
472 return SINTTOFP_I128_F32;
473 else if (RetVT == MVT::f64)
474 return SINTTOFP_I128_F64;
475 else if (RetVT == MVT::f80)
476 return SINTTOFP_I128_F80;
477 else if (RetVT == MVT::ppcf128)
478 return SINTTOFP_I128_PPCF128;
480 return UNKNOWN_LIBCALL;
483 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or
484 /// UNKNOWN_LIBCALL if there is none.
485 RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) {
486 if (OpVT == MVT::i32) {
487 if (RetVT == MVT::f32)
488 return UINTTOFP_I32_F32;
489 else if (RetVT == MVT::f64)
490 return UINTTOFP_I32_F64;
491 else if (RetVT == MVT::f80)
492 return UINTTOFP_I32_F80;
493 else if (RetVT == MVT::ppcf128)
494 return UINTTOFP_I32_PPCF128;
495 } else if (OpVT == MVT::i64) {
496 if (RetVT == MVT::f32)
497 return UINTTOFP_I64_F32;
498 else if (RetVT == MVT::f64)
499 return UINTTOFP_I64_F64;
500 else if (RetVT == MVT::f80)
501 return UINTTOFP_I64_F80;
502 else if (RetVT == MVT::ppcf128)
503 return UINTTOFP_I64_PPCF128;
504 } else if (OpVT == MVT::i128) {
505 if (RetVT == MVT::f32)
506 return UINTTOFP_I128_F32;
507 else if (RetVT == MVT::f64)
508 return UINTTOFP_I128_F64;
509 else if (RetVT == MVT::f80)
510 return UINTTOFP_I128_F80;
511 else if (RetVT == MVT::ppcf128)
512 return UINTTOFP_I128_PPCF128;
514 return UNKNOWN_LIBCALL;
517 /// InitCmpLibcallCCs - Set default comparison libcall CC.
519 static void InitCmpLibcallCCs(ISD::CondCode *CCs) {
520 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL);
521 CCs[RTLIB::OEQ_F32] = ISD::SETEQ;
522 CCs[RTLIB::OEQ_F64] = ISD::SETEQ;
523 CCs[RTLIB::UNE_F32] = ISD::SETNE;
524 CCs[RTLIB::UNE_F64] = ISD::SETNE;
525 CCs[RTLIB::OGE_F32] = ISD::SETGE;
526 CCs[RTLIB::OGE_F64] = ISD::SETGE;
527 CCs[RTLIB::OLT_F32] = ISD::SETLT;
528 CCs[RTLIB::OLT_F64] = ISD::SETLT;
529 CCs[RTLIB::OLE_F32] = ISD::SETLE;
530 CCs[RTLIB::OLE_F64] = ISD::SETLE;
531 CCs[RTLIB::OGT_F32] = ISD::SETGT;
532 CCs[RTLIB::OGT_F64] = ISD::SETGT;
533 CCs[RTLIB::UO_F32] = ISD::SETNE;
534 CCs[RTLIB::UO_F64] = ISD::SETNE;
535 CCs[RTLIB::O_F32] = ISD::SETEQ;
536 CCs[RTLIB::O_F64] = ISD::SETEQ;
539 /// NOTE: The constructor takes ownership of TLOF.
540 TargetLowering::TargetLowering(const TargetMachine &tm,
541 const TargetLoweringObjectFile *tlof)
542 : TM(tm), TD(TM.getTargetData()), TLOF(*tlof),
543 mayPromoteElements(AllowPromoteIntElem) {
544 // All operations default to being supported.
545 memset(OpActions, 0, sizeof(OpActions));
546 memset(LoadExtActions, 0, sizeof(LoadExtActions));
547 memset(TruncStoreActions, 0, sizeof(TruncStoreActions));
548 memset(IndexedModeActions, 0, sizeof(IndexedModeActions));
549 memset(CondCodeActions, 0, sizeof(CondCodeActions));
551 // Set default actions for various operations.
552 for (unsigned VT = 0; VT != (unsigned)MVT::LAST_VALUETYPE; ++VT) {
553 // Default all indexed load / store to expand.
554 for (unsigned IM = (unsigned)ISD::PRE_INC;
555 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
556 setIndexedLoadAction(IM, (MVT::SimpleValueType)VT, Expand);
557 setIndexedStoreAction(IM, (MVT::SimpleValueType)VT, Expand);
560 // These operations default to expand.
561 setOperationAction(ISD::FGETSIGN, (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::CONCAT_VECTORS, (MVT::SimpleValueType)VT, Expand);
565 // Most targets ignore the @llvm.prefetch intrinsic.
566 setOperationAction(ISD::PREFETCH, MVT::Other, Expand);
568 // ConstantFP nodes default to expand. Targets can either change this to
569 // Legal, in which case all fp constants are legal, or use isFPImmLegal()
570 // to optimize expansions for certain constants.
571 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
572 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
573 setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
575 // These library functions default to expand.
576 setOperationAction(ISD::FLOG , MVT::f64, Expand);
577 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
578 setOperationAction(ISD::FLOG10,MVT::f64, Expand);
579 setOperationAction(ISD::FEXP , MVT::f64, Expand);
580 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
581 setOperationAction(ISD::FLOG , MVT::f32, Expand);
582 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
583 setOperationAction(ISD::FLOG10,MVT::f32, Expand);
584 setOperationAction(ISD::FEXP , MVT::f32, Expand);
585 setOperationAction(ISD::FEXP2, MVT::f32, Expand);
587 // Default ISD::TRAP to expand (which turns it into abort).
588 setOperationAction(ISD::TRAP, MVT::Other, Expand);
590 IsLittleEndian = TD->isLittleEndian();
591 PointerTy = MVT::getIntegerVT(8*TD->getPointerSize());
592 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE*sizeof(TargetRegisterClass*));
593 memset(TargetDAGCombineArray, 0, array_lengthof(TargetDAGCombineArray));
594 maxStoresPerMemset = maxStoresPerMemcpy = maxStoresPerMemmove = 8;
595 maxStoresPerMemsetOptSize = maxStoresPerMemcpyOptSize
596 = maxStoresPerMemmoveOptSize = 4;
597 benefitFromCodePlacementOpt = false;
598 UseUnderscoreSetJmp = false;
599 UseUnderscoreLongJmp = false;
600 SelectIsExpensive = false;
601 IntDivIsCheap = false;
602 Pow2DivIsCheap = false;
603 JumpIsExpensive = false;
604 StackPointerRegisterToSaveRestore = 0;
605 ExceptionPointerRegister = 0;
606 ExceptionSelectorRegister = 0;
607 BooleanContents = UndefinedBooleanContent;
608 SchedPreferenceInfo = Sched::Latency;
610 JumpBufAlignment = 0;
611 MinFunctionAlignment = 0;
612 PrefFunctionAlignment = 0;
613 PrefLoopAlignment = 0;
614 MinStackArgumentAlignment = 1;
615 ShouldFoldAtomicFences = false;
617 InitLibcallNames(LibcallRoutineNames);
618 InitCmpLibcallCCs(CmpLibcallCCs);
619 InitLibcallCallingConvs(LibcallCallingConvs);
622 TargetLowering::~TargetLowering() {
626 MVT TargetLowering::getShiftAmountTy(EVT LHSTy) const {
627 return MVT::getIntegerVT(8*TD->getPointerSize());
630 /// canOpTrap - Returns true if the operation can trap for the value type.
631 /// VT must be a legal type.
632 bool TargetLowering::canOpTrap(unsigned Op, EVT VT) const {
633 assert(isTypeLegal(VT));
648 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT,
649 unsigned &NumIntermediates,
651 TargetLowering *TLI) {
652 // Figure out the right, legal destination reg to copy into.
653 unsigned NumElts = VT.getVectorNumElements();
654 MVT EltTy = VT.getVectorElementType();
656 unsigned NumVectorRegs = 1;
658 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
659 // could break down into LHS/RHS like LegalizeDAG does.
660 if (!isPowerOf2_32(NumElts)) {
661 NumVectorRegs = NumElts;
665 // Divide the input until we get to a supported size. This will always
666 // end with a scalar if the target doesn't support vectors.
667 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) {
672 NumIntermediates = NumVectorRegs;
674 MVT NewVT = MVT::getVectorVT(EltTy, NumElts);
675 if (!TLI->isTypeLegal(NewVT))
677 IntermediateVT = NewVT;
679 unsigned NewVTSize = NewVT.getSizeInBits();
681 // Convert sizes such as i33 to i64.
682 if (!isPowerOf2_32(NewVTSize))
683 NewVTSize = NextPowerOf2(NewVTSize);
685 EVT DestVT = TLI->getRegisterType(NewVT);
687 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16.
688 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits());
690 // Otherwise, promotion or legal types use the same number of registers as
691 // the vector decimated to the appropriate level.
692 return NumVectorRegs;
695 /// isLegalRC - Return true if the value types that can be represented by the
696 /// specified register class are all legal.
697 bool TargetLowering::isLegalRC(const TargetRegisterClass *RC) const {
698 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
706 /// hasLegalSuperRegRegClasses - Return true if the specified register class
707 /// has one or more super-reg register classes that are legal.
709 TargetLowering::hasLegalSuperRegRegClasses(const TargetRegisterClass *RC) const{
710 if (*RC->superregclasses_begin() == 0)
712 for (TargetRegisterInfo::regclass_iterator I = RC->superregclasses_begin(),
713 E = RC->superregclasses_end(); I != E; ++I) {
714 const TargetRegisterClass *RRC = *I;
721 /// findRepresentativeClass - Return the largest legal super-reg register class
722 /// of the register class for the specified type and its associated "cost".
723 std::pair<const TargetRegisterClass*, uint8_t>
724 TargetLowering::findRepresentativeClass(EVT VT) const {
725 const TargetRegisterClass *RC = RegClassForVT[VT.getSimpleVT().SimpleTy];
727 return std::make_pair(RC, 0);
728 const TargetRegisterClass *BestRC = RC;
729 for (TargetRegisterInfo::regclass_iterator I = RC->superregclasses_begin(),
730 E = RC->superregclasses_end(); I != E; ++I) {
731 const TargetRegisterClass *RRC = *I;
732 if (RRC->isASubClass() || !isLegalRC(RRC))
734 if (!hasLegalSuperRegRegClasses(RRC))
735 return std::make_pair(RRC, 1);
738 return std::make_pair(BestRC, 1);
742 /// computeRegisterProperties - Once all of the register classes are added,
743 /// this allows us to compute derived properties we expose.
744 void TargetLowering::computeRegisterProperties() {
745 assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE &&
746 "Too many value types for ValueTypeActions to hold!");
748 // Everything defaults to needing one register.
749 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
750 NumRegistersForVT[i] = 1;
751 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i;
753 // ...except isVoid, which doesn't need any registers.
754 NumRegistersForVT[MVT::isVoid] = 0;
756 // Find the largest integer register class.
757 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE;
758 for (; RegClassForVT[LargestIntReg] == 0; --LargestIntReg)
759 assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
761 // Every integer value type larger than this largest register takes twice as
762 // many registers to represent as the previous ValueType.
763 for (unsigned ExpandedReg = LargestIntReg + 1; ; ++ExpandedReg) {
764 EVT ExpandedVT = (MVT::SimpleValueType)ExpandedReg;
765 if (!ExpandedVT.isInteger())
767 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
768 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg;
769 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1);
770 ValueTypeActions.setTypeAction(ExpandedVT, TypeExpandInteger);
773 // Inspect all of the ValueType's smaller than the largest integer
774 // register to see which ones need promotion.
775 unsigned LegalIntReg = LargestIntReg;
776 for (unsigned IntReg = LargestIntReg - 1;
777 IntReg >= (unsigned)MVT::i1; --IntReg) {
778 EVT IVT = (MVT::SimpleValueType)IntReg;
779 if (isTypeLegal(IVT)) {
780 LegalIntReg = IntReg;
782 RegisterTypeForVT[IntReg] = TransformToType[IntReg] =
783 (MVT::SimpleValueType)LegalIntReg;
784 ValueTypeActions.setTypeAction(IVT, TypePromoteInteger);
788 // ppcf128 type is really two f64's.
789 if (!isTypeLegal(MVT::ppcf128)) {
790 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64];
791 RegisterTypeForVT[MVT::ppcf128] = MVT::f64;
792 TransformToType[MVT::ppcf128] = MVT::f64;
793 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat);
796 // Decide how to handle f64. If the target does not have native f64 support,
797 // expand it to i64 and we will be generating soft float library calls.
798 if (!isTypeLegal(MVT::f64)) {
799 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
800 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
801 TransformToType[MVT::f64] = MVT::i64;
802 ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat);
805 // Decide how to handle f32. If the target does not have native support for
806 // f32, promote it to f64 if it is legal. Otherwise, expand it to i32.
807 if (!isTypeLegal(MVT::f32)) {
808 if (isTypeLegal(MVT::f64)) {
809 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::f64];
810 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::f64];
811 TransformToType[MVT::f32] = MVT::f64;
812 ValueTypeActions.setTypeAction(MVT::f32, TypePromoteInteger);
814 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
815 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
816 TransformToType[MVT::f32] = MVT::i32;
817 ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat);
821 // Loop over all of the vector value types to see which need transformations.
822 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
823 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
824 MVT VT = (MVT::SimpleValueType)i;
825 if (isTypeLegal(VT)) continue;
827 // Determine if there is a legal wider type. If so, we should promote to
828 // that wider vector type.
829 EVT EltVT = VT.getVectorElementType();
830 unsigned NElts = VT.getVectorNumElements();
832 bool IsLegalWiderType = false;
833 // If we allow the promotion of vector elements using a flag,
834 // then return TypePromoteInteger on vector elements.
835 // First try to promote the elements of integer vectors. If no legal
836 // promotion was found, fallback to the widen-vector method.
837 if (mayPromoteElements)
838 for (unsigned nVT = i+1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
839 EVT SVT = (MVT::SimpleValueType)nVT;
840 // Promote vectors of integers to vectors with the same number
841 // of elements, with a wider element type.
842 if (SVT.getVectorElementType().getSizeInBits() > EltVT.getSizeInBits()
843 && SVT.getVectorNumElements() == NElts &&
844 isTypeLegal(SVT) && SVT.getScalarType().isInteger()) {
845 TransformToType[i] = SVT;
846 RegisterTypeForVT[i] = SVT;
847 NumRegistersForVT[i] = 1;
848 ValueTypeActions.setTypeAction(VT, TypePromoteInteger);
849 IsLegalWiderType = true;
854 if (IsLegalWiderType) continue;
856 // Try to widen the vector.
857 for (unsigned nVT = i+1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
858 EVT SVT = (MVT::SimpleValueType)nVT;
859 if (SVT.getVectorElementType() == EltVT &&
860 SVT.getVectorNumElements() > NElts &&
862 TransformToType[i] = SVT;
863 RegisterTypeForVT[i] = SVT;
864 NumRegistersForVT[i] = 1;
865 ValueTypeActions.setTypeAction(VT, TypeWidenVector);
866 IsLegalWiderType = true;
870 if (IsLegalWiderType) continue;
875 unsigned NumIntermediates;
876 NumRegistersForVT[i] =
877 getVectorTypeBreakdownMVT(VT, IntermediateVT, NumIntermediates,
879 RegisterTypeForVT[i] = RegisterVT;
881 EVT NVT = VT.getPow2VectorType();
883 // Type is already a power of 2. The default action is to split.
884 TransformToType[i] = MVT::Other;
885 unsigned NumElts = VT.getVectorNumElements();
886 ValueTypeActions.setTypeAction(VT,
887 NumElts > 1 ? TypeSplitVector : TypeScalarizeVector);
889 TransformToType[i] = NVT;
890 ValueTypeActions.setTypeAction(VT, TypeWidenVector);
894 // Determine the 'representative' register class for each value type.
895 // An representative register class is the largest (meaning one which is
896 // not a sub-register class / subreg register class) legal register class for
897 // a group of value types. For example, on i386, i8, i16, and i32
898 // representative would be GR32; while on x86_64 it's GR64.
899 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
900 const TargetRegisterClass* RRC;
902 tie(RRC, Cost) = findRepresentativeClass((MVT::SimpleValueType)i);
903 RepRegClassForVT[i] = RRC;
904 RepRegClassCostForVT[i] = Cost;
908 const char *TargetLowering::getTargetNodeName(unsigned Opcode) const {
913 MVT::SimpleValueType TargetLowering::getSetCCResultType(EVT VT) const {
914 return PointerTy.SimpleTy;
917 MVT::SimpleValueType TargetLowering::getCmpLibcallReturnType() const {
918 return MVT::i32; // return the default value
921 /// getVectorTypeBreakdown - Vector types are broken down into some number of
922 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
923 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
924 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
926 /// This method returns the number of registers needed, and the VT for each
927 /// register. It also returns the VT and quantity of the intermediate values
928 /// before they are promoted/expanded.
930 unsigned TargetLowering::getVectorTypeBreakdown(LLVMContext &Context, EVT VT,
932 unsigned &NumIntermediates,
933 EVT &RegisterVT) const {
934 unsigned NumElts = VT.getVectorNumElements();
936 // If there is a wider vector type with the same element type as this one,
937 // we should widen to that legal vector type. This handles things like
938 // <2 x float> -> <4 x float>.
939 if (NumElts != 1 && getTypeAction(Context, VT) == TypeWidenVector) {
940 RegisterVT = getTypeToTransformTo(Context, VT);
941 if (isTypeLegal(RegisterVT)) {
942 IntermediateVT = RegisterVT;
943 NumIntermediates = 1;
948 // Figure out the right, legal destination reg to copy into.
949 EVT EltTy = VT.getVectorElementType();
951 unsigned NumVectorRegs = 1;
953 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
954 // could break down into LHS/RHS like LegalizeDAG does.
955 if (!isPowerOf2_32(NumElts)) {
956 NumVectorRegs = NumElts;
960 // Divide the input until we get to a supported size. This will always
961 // end with a scalar if the target doesn't support vectors.
962 while (NumElts > 1 && !isTypeLegal(
963 EVT::getVectorVT(Context, EltTy, NumElts))) {
968 NumIntermediates = NumVectorRegs;
970 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts);
971 if (!isTypeLegal(NewVT))
973 IntermediateVT = NewVT;
975 EVT DestVT = getRegisterType(Context, NewVT);
977 unsigned NewVTSize = NewVT.getSizeInBits();
979 // Convert sizes such as i33 to i64.
980 if (!isPowerOf2_32(NewVTSize))
981 NewVTSize = NextPowerOf2(NewVTSize);
983 if (DestVT.bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16.
984 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits());
986 // Otherwise, promotion or legal types use the same number of registers as
987 // the vector decimated to the appropriate level.
988 return NumVectorRegs;
991 /// Get the EVTs and ArgFlags collections that represent the legalized return
992 /// type of the given function. This does not require a DAG or a return value,
993 /// and is suitable for use before any DAGs for the function are constructed.
994 /// TODO: Move this out of TargetLowering.cpp.
995 void llvm::GetReturnInfo(const Type* ReturnType, Attributes attr,
996 SmallVectorImpl<ISD::OutputArg> &Outs,
997 const TargetLowering &TLI,
998 SmallVectorImpl<uint64_t> *Offsets) {
999 SmallVector<EVT, 4> ValueVTs;
1000 ComputeValueVTs(TLI, ReturnType, ValueVTs);
1001 unsigned NumValues = ValueVTs.size();
1002 if (NumValues == 0) return;
1003 unsigned Offset = 0;
1005 for (unsigned j = 0, f = NumValues; j != f; ++j) {
1006 EVT VT = ValueVTs[j];
1007 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
1009 if (attr & Attribute::SExt)
1010 ExtendKind = ISD::SIGN_EXTEND;
1011 else if (attr & Attribute::ZExt)
1012 ExtendKind = ISD::ZERO_EXTEND;
1014 // FIXME: C calling convention requires the return type to be promoted to
1015 // at least 32-bit. But this is not necessary for non-C calling
1016 // conventions. The frontend should mark functions whose return values
1017 // require promoting with signext or zeroext attributes.
1018 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) {
1019 EVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32);
1020 if (VT.bitsLT(MinVT))
1024 unsigned NumParts = TLI.getNumRegisters(ReturnType->getContext(), VT);
1025 EVT PartVT = TLI.getRegisterType(ReturnType->getContext(), VT);
1026 unsigned PartSize = TLI.getTargetData()->getTypeAllocSize(
1027 PartVT.getTypeForEVT(ReturnType->getContext()));
1029 // 'inreg' on function refers to return value
1030 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1031 if (attr & Attribute::InReg)
1034 // Propagate extension type if any
1035 if (attr & Attribute::SExt)
1037 else if (attr & Attribute::ZExt)
1040 for (unsigned i = 0; i < NumParts; ++i) {
1041 Outs.push_back(ISD::OutputArg(Flags, PartVT, /*isFixed=*/true));
1043 Offsets->push_back(Offset);
1050 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1051 /// function arguments in the caller parameter area. This is the actual
1052 /// alignment, not its logarithm.
1053 unsigned TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1054 return TD->getCallFrameTypeAlignment(Ty);
1057 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1058 /// current function. The returned value is a member of the
1059 /// MachineJumpTableInfo::JTEntryKind enum.
1060 unsigned TargetLowering::getJumpTableEncoding() const {
1061 // In non-pic modes, just use the address of a block.
1062 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
1063 return MachineJumpTableInfo::EK_BlockAddress;
1065 // In PIC mode, if the target supports a GPRel32 directive, use it.
1066 if (getTargetMachine().getMCAsmInfo()->getGPRel32Directive() != 0)
1067 return MachineJumpTableInfo::EK_GPRel32BlockAddress;
1069 // Otherwise, use a label difference.
1070 return MachineJumpTableInfo::EK_LabelDifference32;
1073 SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1074 SelectionDAG &DAG) const {
1075 // If our PIC model is GP relative, use the global offset table as the base.
1076 if (getJumpTableEncoding() == MachineJumpTableInfo::EK_GPRel32BlockAddress)
1077 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy());
1081 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1082 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1085 TargetLowering::getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
1086 unsigned JTI,MCContext &Ctx) const{
1087 // The normal PIC reloc base is the label at the start of the jump table.
1088 return MCSymbolRefExpr::Create(MF->getJTISymbol(JTI, Ctx), Ctx);
1092 TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
1093 // Assume that everything is safe in static mode.
1094 if (getTargetMachine().getRelocationModel() == Reloc::Static)
1097 // In dynamic-no-pic mode, assume that known defined values are safe.
1098 if (getTargetMachine().getRelocationModel() == Reloc::DynamicNoPIC &&
1100 !GA->getGlobal()->isDeclaration() &&
1101 !GA->getGlobal()->isWeakForLinker())
1104 // Otherwise assume nothing is safe.
1108 //===----------------------------------------------------------------------===//
1109 // Optimization Methods
1110 //===----------------------------------------------------------------------===//
1112 /// ShrinkDemandedConstant - Check to see if the specified operand of the
1113 /// specified instruction is a constant integer. If so, check to see if there
1114 /// are any bits set in the constant that are not demanded. If so, shrink the
1115 /// constant and return true.
1116 bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDValue Op,
1117 const APInt &Demanded) {
1118 DebugLoc dl = Op.getDebugLoc();
1120 // FIXME: ISD::SELECT, ISD::SELECT_CC
1121 switch (Op.getOpcode()) {
1126 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
1127 if (!C) return false;
1129 if (Op.getOpcode() == ISD::XOR &&
1130 (C->getAPIntValue() | (~Demanded)).isAllOnesValue())
1133 // if we can expand it to have all bits set, do it
1134 if (C->getAPIntValue().intersects(~Demanded)) {
1135 EVT VT = Op.getValueType();
1136 SDValue New = DAG.getNode(Op.getOpcode(), dl, VT, Op.getOperand(0),
1137 DAG.getConstant(Demanded &
1140 return CombineTo(Op, New);
1150 /// ShrinkDemandedOp - Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the
1151 /// casts are free. This uses isZExtFree and ZERO_EXTEND for the widening
1152 /// cast, but it could be generalized for targets with other types of
1153 /// implicit widening casts.
1155 TargetLowering::TargetLoweringOpt::ShrinkDemandedOp(SDValue Op,
1157 const APInt &Demanded,
1159 assert(Op.getNumOperands() == 2 &&
1160 "ShrinkDemandedOp only supports binary operators!");
1161 assert(Op.getNode()->getNumValues() == 1 &&
1162 "ShrinkDemandedOp only supports nodes with one result!");
1164 // Don't do this if the node has another user, which may require the
1166 if (!Op.getNode()->hasOneUse())
1169 // Search for the smallest integer type with free casts to and from
1170 // Op's type. For expedience, just check power-of-2 integer types.
1171 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1172 unsigned SmallVTBits = BitWidth - Demanded.countLeadingZeros();
1173 if (!isPowerOf2_32(SmallVTBits))
1174 SmallVTBits = NextPowerOf2(SmallVTBits);
1175 for (; SmallVTBits < BitWidth; SmallVTBits = NextPowerOf2(SmallVTBits)) {
1176 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), SmallVTBits);
1177 if (TLI.isTruncateFree(Op.getValueType(), SmallVT) &&
1178 TLI.isZExtFree(SmallVT, Op.getValueType())) {
1179 // We found a type with free casts.
1180 SDValue X = DAG.getNode(Op.getOpcode(), dl, SmallVT,
1181 DAG.getNode(ISD::TRUNCATE, dl, SmallVT,
1182 Op.getNode()->getOperand(0)),
1183 DAG.getNode(ISD::TRUNCATE, dl, SmallVT,
1184 Op.getNode()->getOperand(1)));
1185 SDValue Z = DAG.getNode(ISD::ZERO_EXTEND, dl, Op.getValueType(), X);
1186 return CombineTo(Op, Z);
1192 /// SimplifyDemandedBits - Look at Op. At this point, we know that only the
1193 /// DemandedMask bits of the result of Op are ever used downstream. If we can
1194 /// use this information to simplify Op, create a new simplified DAG node and
1195 /// return true, returning the original and new nodes in Old and New. Otherwise,
1196 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
1197 /// the expression (used to simplify the caller). The KnownZero/One bits may
1198 /// only be accurate for those bits in the DemandedMask.
1199 bool TargetLowering::SimplifyDemandedBits(SDValue Op,
1200 const APInt &DemandedMask,
1203 TargetLoweringOpt &TLO,
1204 unsigned Depth) const {
1205 unsigned BitWidth = DemandedMask.getBitWidth();
1206 assert(Op.getValueType().getScalarType().getSizeInBits() == BitWidth &&
1207 "Mask size mismatches value type size!");
1208 APInt NewMask = DemandedMask;
1209 DebugLoc dl = Op.getDebugLoc();
1211 // Don't know anything.
1212 KnownZero = KnownOne = APInt(BitWidth, 0);
1214 // Other users may use these bits.
1215 if (!Op.getNode()->hasOneUse()) {
1217 // If not at the root, Just compute the KnownZero/KnownOne bits to
1218 // simplify things downstream.
1219 TLO.DAG.ComputeMaskedBits(Op, DemandedMask, KnownZero, KnownOne, Depth);
1222 // If this is the root being simplified, allow it to have multiple uses,
1223 // just set the NewMask to all bits.
1224 NewMask = APInt::getAllOnesValue(BitWidth);
1225 } else if (DemandedMask == 0) {
1226 // Not demanding any bits from Op.
1227 if (Op.getOpcode() != ISD::UNDEF)
1228 return TLO.CombineTo(Op, TLO.DAG.getUNDEF(Op.getValueType()));
1230 } else if (Depth == 6) { // Limit search depth.
1234 APInt KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut;
1235 switch (Op.getOpcode()) {
1237 // We know all of the bits for a constant!
1238 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue() & NewMask;
1239 KnownZero = ~KnownOne & NewMask;
1240 return false; // Don't fall through, will infinitely loop.
1242 // If the RHS is a constant, check to see if the LHS would be zero without
1243 // using the bits from the RHS. Below, we use knowledge about the RHS to
1244 // simplify the LHS, here we're using information from the LHS to simplify
1246 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1247 APInt LHSZero, LHSOne;
1248 // Do not increment Depth here; that can cause an infinite loop.
1249 TLO.DAG.ComputeMaskedBits(Op.getOperand(0), NewMask,
1250 LHSZero, LHSOne, Depth);
1251 // If the LHS already has zeros where RHSC does, this and is dead.
1252 if ((LHSZero & NewMask) == (~RHSC->getAPIntValue() & NewMask))
1253 return TLO.CombineTo(Op, Op.getOperand(0));
1254 // If any of the set bits in the RHS are known zero on the LHS, shrink
1256 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & NewMask))
1260 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
1261 KnownOne, TLO, Depth+1))
1263 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1264 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownZero & NewMask,
1265 KnownZero2, KnownOne2, TLO, Depth+1))
1267 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1269 // If all of the demanded bits are known one on one side, return the other.
1270 // These bits cannot contribute to the result of the 'and'.
1271 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
1272 return TLO.CombineTo(Op, Op.getOperand(0));
1273 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
1274 return TLO.CombineTo(Op, Op.getOperand(1));
1275 // If all of the demanded bits in the inputs are known zeros, return zero.
1276 if ((NewMask & (KnownZero|KnownZero2)) == NewMask)
1277 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, Op.getValueType()));
1278 // If the RHS is a constant, see if we can simplify it.
1279 if (TLO.ShrinkDemandedConstant(Op, ~KnownZero2 & NewMask))
1281 // If the operation can be done in a smaller type, do so.
1282 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1285 // Output known-1 bits are only known if set in both the LHS & RHS.
1286 KnownOne &= KnownOne2;
1287 // Output known-0 are known to be clear if zero in either the LHS | RHS.
1288 KnownZero |= KnownZero2;
1291 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
1292 KnownOne, TLO, Depth+1))
1294 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1295 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownOne & NewMask,
1296 KnownZero2, KnownOne2, TLO, Depth+1))
1298 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1300 // If all of the demanded bits are known zero on one side, return the other.
1301 // These bits cannot contribute to the result of the 'or'.
1302 if ((NewMask & ~KnownOne2 & KnownZero) == (~KnownOne2 & NewMask))
1303 return TLO.CombineTo(Op, Op.getOperand(0));
1304 if ((NewMask & ~KnownOne & KnownZero2) == (~KnownOne & NewMask))
1305 return TLO.CombineTo(Op, Op.getOperand(1));
1306 // If all of the potentially set bits on one side are known to be set on
1307 // the other side, just use the 'other' side.
1308 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
1309 return TLO.CombineTo(Op, Op.getOperand(0));
1310 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
1311 return TLO.CombineTo(Op, Op.getOperand(1));
1312 // If the RHS is a constant, see if we can simplify it.
1313 if (TLO.ShrinkDemandedConstant(Op, NewMask))
1315 // If the operation can be done in a smaller type, do so.
1316 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1319 // Output known-0 bits are only known if clear in both the LHS & RHS.
1320 KnownZero &= KnownZero2;
1321 // Output known-1 are known to be set if set in either the LHS | RHS.
1322 KnownOne |= KnownOne2;
1325 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
1326 KnownOne, TLO, Depth+1))
1328 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1329 if (SimplifyDemandedBits(Op.getOperand(0), NewMask, KnownZero2,
1330 KnownOne2, TLO, Depth+1))
1332 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1334 // If all of the demanded bits are known zero on one side, return the other.
1335 // These bits cannot contribute to the result of the 'xor'.
1336 if ((KnownZero & NewMask) == NewMask)
1337 return TLO.CombineTo(Op, Op.getOperand(0));
1338 if ((KnownZero2 & NewMask) == NewMask)
1339 return TLO.CombineTo(Op, Op.getOperand(1));
1340 // If the operation can be done in a smaller type, do so.
1341 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1344 // If all of the unknown bits are known to be zero on one side or the other
1345 // (but not both) turn this into an *inclusive* or.
1346 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1347 if ((NewMask & ~KnownZero & ~KnownZero2) == 0)
1348 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, dl, Op.getValueType(),
1352 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1353 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
1354 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1355 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
1357 // If all of the demanded bits on one side are known, and all of the set
1358 // bits on that side are also known to be set on the other side, turn this
1359 // into an AND, as we know the bits will be cleared.
1360 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1361 if ((NewMask & (KnownZero|KnownOne)) == NewMask) { // all known
1362 if ((KnownOne & KnownOne2) == KnownOne) {
1363 EVT VT = Op.getValueType();
1364 SDValue ANDC = TLO.DAG.getConstant(~KnownOne & NewMask, VT);
1365 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, dl, VT,
1366 Op.getOperand(0), ANDC));
1370 // If the RHS is a constant, see if we can simplify it.
1371 // for XOR, we prefer to force bits to 1 if they will make a -1.
1372 // if we can't force bits, try to shrink constant
1373 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1374 APInt Expanded = C->getAPIntValue() | (~NewMask);
1375 // if we can expand it to have all bits set, do it
1376 if (Expanded.isAllOnesValue()) {
1377 if (Expanded != C->getAPIntValue()) {
1378 EVT VT = Op.getValueType();
1379 SDValue New = TLO.DAG.getNode(Op.getOpcode(), dl,VT, Op.getOperand(0),
1380 TLO.DAG.getConstant(Expanded, VT));
1381 return TLO.CombineTo(Op, New);
1383 // if it already has all the bits set, nothing to change
1384 // but don't shrink either!
1385 } else if (TLO.ShrinkDemandedConstant(Op, NewMask)) {
1390 KnownZero = KnownZeroOut;
1391 KnownOne = KnownOneOut;
1394 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero,
1395 KnownOne, TLO, Depth+1))
1397 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero2,
1398 KnownOne2, TLO, Depth+1))
1400 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1401 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1403 // If the operands are constants, see if we can simplify them.
1404 if (TLO.ShrinkDemandedConstant(Op, NewMask))
1407 // Only known if known in both the LHS and RHS.
1408 KnownOne &= KnownOne2;
1409 KnownZero &= KnownZero2;
1411 case ISD::SELECT_CC:
1412 if (SimplifyDemandedBits(Op.getOperand(3), NewMask, KnownZero,
1413 KnownOne, TLO, Depth+1))
1415 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero2,
1416 KnownOne2, TLO, Depth+1))
1418 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1419 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1421 // If the operands are constants, see if we can simplify them.
1422 if (TLO.ShrinkDemandedConstant(Op, NewMask))
1425 // Only known if known in both the LHS and RHS.
1426 KnownOne &= KnownOne2;
1427 KnownZero &= KnownZero2;
1430 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1431 unsigned ShAmt = SA->getZExtValue();
1432 SDValue InOp = Op.getOperand(0);
1434 // If the shift count is an invalid immediate, don't do anything.
1435 if (ShAmt >= BitWidth)
1438 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a
1439 // single shift. We can do this if the bottom bits (which are shifted
1440 // out) are never demanded.
1441 if (InOp.getOpcode() == ISD::SRL &&
1442 isa<ConstantSDNode>(InOp.getOperand(1))) {
1443 if (ShAmt && (NewMask & APInt::getLowBitsSet(BitWidth, ShAmt)) == 0) {
1444 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
1445 unsigned Opc = ISD::SHL;
1446 int Diff = ShAmt-C1;
1453 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
1454 EVT VT = Op.getValueType();
1455 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT,
1456 InOp.getOperand(0), NewSA));
1460 if (SimplifyDemandedBits(InOp, NewMask.lshr(ShAmt),
1461 KnownZero, KnownOne, TLO, Depth+1))
1464 // Convert (shl (anyext x, c)) to (anyext (shl x, c)) if the high bits
1465 // are not demanded. This will likely allow the anyext to be folded away.
1466 if (InOp.getNode()->getOpcode() == ISD::ANY_EXTEND) {
1467 SDValue InnerOp = InOp.getNode()->getOperand(0);
1468 EVT InnerVT = InnerOp.getValueType();
1469 if ((APInt::getHighBitsSet(BitWidth,
1470 BitWidth - InnerVT.getSizeInBits()) &
1471 DemandedMask) == 0 &&
1472 isTypeDesirableForOp(ISD::SHL, InnerVT)) {
1473 EVT ShTy = getShiftAmountTy(InnerVT);
1474 if (!APInt(BitWidth, ShAmt).isIntN(ShTy.getSizeInBits()))
1477 TLO.DAG.getNode(ISD::SHL, dl, InnerVT, InnerOp,
1478 TLO.DAG.getConstant(ShAmt, ShTy));
1481 TLO.DAG.getNode(ISD::ANY_EXTEND, dl, Op.getValueType(),
1486 KnownZero <<= SA->getZExtValue();
1487 KnownOne <<= SA->getZExtValue();
1488 // low bits known zero.
1489 KnownZero |= APInt::getLowBitsSet(BitWidth, SA->getZExtValue());
1493 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1494 EVT VT = Op.getValueType();
1495 unsigned ShAmt = SA->getZExtValue();
1496 unsigned VTSize = VT.getSizeInBits();
1497 SDValue InOp = Op.getOperand(0);
1499 // If the shift count is an invalid immediate, don't do anything.
1500 if (ShAmt >= BitWidth)
1503 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a
1504 // single shift. We can do this if the top bits (which are shifted out)
1505 // are never demanded.
1506 if (InOp.getOpcode() == ISD::SHL &&
1507 isa<ConstantSDNode>(InOp.getOperand(1))) {
1508 if (ShAmt && (NewMask & APInt::getHighBitsSet(VTSize, ShAmt)) == 0) {
1509 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
1510 unsigned Opc = ISD::SRL;
1511 int Diff = ShAmt-C1;
1518 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
1519 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT,
1520 InOp.getOperand(0), NewSA));
1524 // Compute the new bits that are at the top now.
1525 if (SimplifyDemandedBits(InOp, (NewMask << ShAmt),
1526 KnownZero, KnownOne, TLO, Depth+1))
1528 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1529 KnownZero = KnownZero.lshr(ShAmt);
1530 KnownOne = KnownOne.lshr(ShAmt);
1532 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
1533 KnownZero |= HighBits; // High bits known zero.
1537 // If this is an arithmetic shift right and only the low-bit is set, we can
1538 // always convert this into a logical shr, even if the shift amount is
1539 // variable. The low bit of the shift cannot be an input sign bit unless
1540 // the shift amount is >= the size of the datatype, which is undefined.
1541 if (DemandedMask == 1)
1542 return TLO.CombineTo(Op,
1543 TLO.DAG.getNode(ISD::SRL, dl, Op.getValueType(),
1544 Op.getOperand(0), Op.getOperand(1)));
1546 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1547 EVT VT = Op.getValueType();
1548 unsigned ShAmt = SA->getZExtValue();
1550 // If the shift count is an invalid immediate, don't do anything.
1551 if (ShAmt >= BitWidth)
1554 APInt InDemandedMask = (NewMask << ShAmt);
1556 // If any of the demanded bits are produced by the sign extension, we also
1557 // demand the input sign bit.
1558 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
1559 if (HighBits.intersects(NewMask))
1560 InDemandedMask |= APInt::getSignBit(VT.getScalarType().getSizeInBits());
1562 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask,
1563 KnownZero, KnownOne, TLO, Depth+1))
1565 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1566 KnownZero = KnownZero.lshr(ShAmt);
1567 KnownOne = KnownOne.lshr(ShAmt);
1569 // Handle the sign bit, adjusted to where it is now in the mask.
1570 APInt SignBit = APInt::getSignBit(BitWidth).lshr(ShAmt);
1572 // If the input sign bit is known to be zero, or if none of the top bits
1573 // are demanded, turn this into an unsigned shift right.
1574 if (KnownZero.intersects(SignBit) || (HighBits & ~NewMask) == HighBits) {
1575 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, VT,
1578 } else if (KnownOne.intersects(SignBit)) { // New bits are known one.
1579 KnownOne |= HighBits;
1583 case ISD::SIGN_EXTEND_INREG: {
1584 EVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1586 // Sign extension. Compute the demanded bits in the result that are not
1587 // present in the input.
1589 APInt::getHighBitsSet(BitWidth,
1590 BitWidth - EVT.getScalarType().getSizeInBits());
1592 // If none of the extended bits are demanded, eliminate the sextinreg.
1593 if ((NewBits & NewMask) == 0)
1594 return TLO.CombineTo(Op, Op.getOperand(0));
1597 APInt::getSignBit(EVT.getScalarType().getSizeInBits()).zext(BitWidth);
1598 APInt InputDemandedBits =
1599 APInt::getLowBitsSet(BitWidth,
1600 EVT.getScalarType().getSizeInBits()) &
1603 // Since the sign extended bits are demanded, we know that the sign
1605 InputDemandedBits |= InSignBit;
1607 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits,
1608 KnownZero, KnownOne, TLO, Depth+1))
1610 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1612 // If the sign bit of the input is known set or clear, then we know the
1613 // top bits of the result.
1615 // If the input sign bit is known zero, convert this into a zero extension.
1616 if (KnownZero.intersects(InSignBit))
1617 return TLO.CombineTo(Op,
1618 TLO.DAG.getZeroExtendInReg(Op.getOperand(0),dl,EVT));
1620 if (KnownOne.intersects(InSignBit)) { // Input sign bit known set
1621 KnownOne |= NewBits;
1622 KnownZero &= ~NewBits;
1623 } else { // Input sign bit unknown
1624 KnownZero &= ~NewBits;
1625 KnownOne &= ~NewBits;
1629 case ISD::ZERO_EXTEND: {
1630 unsigned OperandBitWidth =
1631 Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
1632 APInt InMask = NewMask.trunc(OperandBitWidth);
1634 // If none of the top bits are demanded, convert this into an any_extend.
1636 APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask;
1637 if (!NewBits.intersects(NewMask))
1638 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND, dl,
1642 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1643 KnownZero, KnownOne, TLO, Depth+1))
1645 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1646 KnownZero = KnownZero.zext(BitWidth);
1647 KnownOne = KnownOne.zext(BitWidth);
1648 KnownZero |= NewBits;
1651 case ISD::SIGN_EXTEND: {
1652 EVT InVT = Op.getOperand(0).getValueType();
1653 unsigned InBits = InVT.getScalarType().getSizeInBits();
1654 APInt InMask = APInt::getLowBitsSet(BitWidth, InBits);
1655 APInt InSignBit = APInt::getBitsSet(BitWidth, InBits - 1, InBits);
1656 APInt NewBits = ~InMask & NewMask;
1658 // If none of the top bits are demanded, convert this into an any_extend.
1660 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND, dl,
1664 // Since some of the sign extended bits are demanded, we know that the sign
1666 APInt InDemandedBits = InMask & NewMask;
1667 InDemandedBits |= InSignBit;
1668 InDemandedBits = InDemandedBits.trunc(InBits);
1670 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero,
1671 KnownOne, TLO, Depth+1))
1673 KnownZero = KnownZero.zext(BitWidth);
1674 KnownOne = KnownOne.zext(BitWidth);
1676 // If the sign bit is known zero, convert this to a zero extend.
1677 if (KnownZero.intersects(InSignBit))
1678 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND, dl,
1682 // If the sign bit is known one, the top bits match.
1683 if (KnownOne.intersects(InSignBit)) {
1684 KnownOne |= NewBits;
1685 KnownZero &= ~NewBits;
1686 } else { // Otherwise, top bits aren't known.
1687 KnownOne &= ~NewBits;
1688 KnownZero &= ~NewBits;
1692 case ISD::ANY_EXTEND: {
1693 unsigned OperandBitWidth =
1694 Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
1695 APInt InMask = NewMask.trunc(OperandBitWidth);
1696 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1697 KnownZero, KnownOne, TLO, Depth+1))
1699 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1700 KnownZero = KnownZero.zext(BitWidth);
1701 KnownOne = KnownOne.zext(BitWidth);
1704 case ISD::TRUNCATE: {
1705 // Simplify the input, using demanded bit information, and compute the known
1706 // zero/one bits live out.
1707 unsigned OperandBitWidth =
1708 Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
1709 APInt TruncMask = NewMask.zext(OperandBitWidth);
1710 if (SimplifyDemandedBits(Op.getOperand(0), TruncMask,
1711 KnownZero, KnownOne, TLO, Depth+1))
1713 KnownZero = KnownZero.trunc(BitWidth);
1714 KnownOne = KnownOne.trunc(BitWidth);
1716 // If the input is only used by this truncate, see if we can shrink it based
1717 // on the known demanded bits.
1718 if (Op.getOperand(0).getNode()->hasOneUse()) {
1719 SDValue In = Op.getOperand(0);
1720 switch (In.getOpcode()) {
1723 // Shrink SRL by a constant if none of the high bits shifted in are
1725 if (TLO.LegalTypes() &&
1726 !isTypeDesirableForOp(ISD::SRL, Op.getValueType()))
1727 // Do not turn (vt1 truncate (vt2 srl)) into (vt1 srl) if vt1 is
1730 ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1));
1733 SDValue Shift = In.getOperand(1);
1734 if (TLO.LegalTypes()) {
1735 uint64_t ShVal = ShAmt->getZExtValue();
1737 TLO.DAG.getConstant(ShVal, getShiftAmountTy(Op.getValueType()));
1740 APInt HighBits = APInt::getHighBitsSet(OperandBitWidth,
1741 OperandBitWidth - BitWidth);
1742 HighBits = HighBits.lshr(ShAmt->getZExtValue()).trunc(BitWidth);
1744 if (ShAmt->getZExtValue() < BitWidth && !(HighBits & NewMask)) {
1745 // None of the shifted in bits are needed. Add a truncate of the
1746 // shift input, then shift it.
1747 SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE, dl,
1750 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl,
1759 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1762 case ISD::AssertZext: {
1763 // Demand all the bits of the input that are demanded in the output.
1764 // The low bits are obvious; the high bits are demanded because we're
1765 // asserting that they're zero here.
1766 if (SimplifyDemandedBits(Op.getOperand(0), NewMask,
1767 KnownZero, KnownOne, TLO, Depth+1))
1769 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1771 EVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1772 APInt InMask = APInt::getLowBitsSet(BitWidth,
1773 VT.getSizeInBits());
1774 KnownZero |= ~InMask & NewMask;
1778 // If this is an FP->Int bitcast and if the sign bit is the only
1779 // thing demanded, turn this into a FGETSIGN.
1780 if (!Op.getOperand(0).getValueType().isVector() &&
1781 NewMask == APInt::getSignBit(Op.getValueType().getSizeInBits()) &&
1782 Op.getOperand(0).getValueType().isFloatingPoint()) {
1783 bool OpVTLegal = isOperationLegalOrCustom(ISD::FGETSIGN, Op.getValueType());
1784 bool i32Legal = isOperationLegalOrCustom(ISD::FGETSIGN, MVT::i32);
1785 if ((OpVTLegal || i32Legal) && Op.getValueType().isSimple()) {
1786 EVT Ty = OpVTLegal ? Op.getValueType() : MVT::i32;
1787 // Make a FGETSIGN + SHL to move the sign bit into the appropriate
1788 // place. We expect the SHL to be eliminated by other optimizations.
1789 SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, dl, Ty, Op.getOperand(0));
1790 unsigned OpVTSizeInBits = Op.getValueType().getSizeInBits();
1791 if (!OpVTLegal && OpVTSizeInBits > 32)
1792 Sign = TLO.DAG.getNode(ISD::ZERO_EXTEND, dl, Op.getValueType(), Sign);
1793 unsigned ShVal = Op.getValueType().getSizeInBits()-1;
1794 SDValue ShAmt = TLO.DAG.getConstant(ShVal, Op.getValueType());
1795 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl,
1804 // Add, Sub, and Mul don't demand any bits in positions beyond that
1805 // of the highest bit demanded of them.
1806 APInt LoMask = APInt::getLowBitsSet(BitWidth,
1807 BitWidth - NewMask.countLeadingZeros());
1808 if (SimplifyDemandedBits(Op.getOperand(0), LoMask, KnownZero2,
1809 KnownOne2, TLO, Depth+1))
1811 if (SimplifyDemandedBits(Op.getOperand(1), LoMask, KnownZero2,
1812 KnownOne2, TLO, Depth+1))
1814 // See if the operation should be performed at a smaller bit width.
1815 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1820 // Just use ComputeMaskedBits to compute output bits.
1821 TLO.DAG.ComputeMaskedBits(Op, NewMask, KnownZero, KnownOne, Depth);
1825 // If we know the value of all of the demanded bits, return this as a
1827 if ((NewMask & (KnownZero|KnownOne)) == NewMask)
1828 return TLO.CombineTo(Op, TLO.DAG.getConstant(KnownOne, Op.getValueType()));
1833 /// computeMaskedBitsForTargetNode - Determine which of the bits specified
1834 /// in Mask are known to be either zero or one and return them in the
1835 /// KnownZero/KnownOne bitsets.
1836 void TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
1840 const SelectionDAG &DAG,
1841 unsigned Depth) const {
1842 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1843 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1844 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1845 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1846 "Should use MaskedValueIsZero if you don't know whether Op"
1847 " is a target node!");
1848 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
1851 /// ComputeNumSignBitsForTargetNode - This method can be implemented by
1852 /// targets that want to expose additional information about sign bits to the
1854 unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
1855 unsigned Depth) const {
1856 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1857 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1858 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1859 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1860 "Should use ComputeNumSignBits if you don't know whether Op"
1861 " is a target node!");
1865 /// ValueHasExactlyOneBitSet - Test if the given value is known to have exactly
1866 /// one bit set. This differs from ComputeMaskedBits in that it doesn't need to
1867 /// determine which bit is set.
1869 static bool ValueHasExactlyOneBitSet(SDValue Val, const SelectionDAG &DAG) {
1870 // A left-shift of a constant one will have exactly one bit set, because
1871 // shifting the bit off the end is undefined.
1872 if (Val.getOpcode() == ISD::SHL)
1873 if (ConstantSDNode *C =
1874 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0)))
1875 if (C->getAPIntValue() == 1)
1878 // Similarly, a right-shift of a constant sign-bit will have exactly
1880 if (Val.getOpcode() == ISD::SRL)
1881 if (ConstantSDNode *C =
1882 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0)))
1883 if (C->getAPIntValue().isSignBit())
1886 // More could be done here, though the above checks are enough
1887 // to handle some common cases.
1889 // Fall back to ComputeMaskedBits to catch other known cases.
1890 EVT OpVT = Val.getValueType();
1891 unsigned BitWidth = OpVT.getScalarType().getSizeInBits();
1892 APInt Mask = APInt::getAllOnesValue(BitWidth);
1893 APInt KnownZero, KnownOne;
1894 DAG.ComputeMaskedBits(Val, Mask, KnownZero, KnownOne);
1895 return (KnownZero.countPopulation() == BitWidth - 1) &&
1896 (KnownOne.countPopulation() == 1);
1899 /// SimplifySetCC - Try to simplify a setcc built with the specified operands
1900 /// and cc. If it is unable to simplify it, return a null SDValue.
1902 TargetLowering::SimplifySetCC(EVT VT, SDValue N0, SDValue N1,
1903 ISD::CondCode Cond, bool foldBooleans,
1904 DAGCombinerInfo &DCI, DebugLoc dl) const {
1905 SelectionDAG &DAG = DCI.DAG;
1907 // These setcc operations always fold.
1911 case ISD::SETFALSE2: return DAG.getConstant(0, VT);
1913 case ISD::SETTRUE2: return DAG.getConstant(1, VT);
1916 // Ensure that the constant occurs on the RHS, and fold constant
1918 if (isa<ConstantSDNode>(N0.getNode()))
1919 return DAG.getSetCC(dl, VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
1921 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) {
1922 const APInt &C1 = N1C->getAPIntValue();
1924 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an
1925 // equality comparison, then we're just comparing whether X itself is
1927 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) &&
1928 N0.getOperand(0).getOpcode() == ISD::CTLZ &&
1929 N0.getOperand(1).getOpcode() == ISD::Constant) {
1931 = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
1932 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1933 ShAmt == Log2_32(N0.getValueType().getSizeInBits())) {
1934 if ((C1 == 0) == (Cond == ISD::SETEQ)) {
1935 // (srl (ctlz x), 5) == 0 -> X != 0
1936 // (srl (ctlz x), 5) != 1 -> X != 0
1939 // (srl (ctlz x), 5) != 0 -> X == 0
1940 // (srl (ctlz x), 5) == 1 -> X == 0
1943 SDValue Zero = DAG.getConstant(0, N0.getValueType());
1944 return DAG.getSetCC(dl, VT, N0.getOperand(0).getOperand(0),
1950 // Look through truncs that don't change the value of a ctpop.
1951 if (N0.hasOneUse() && N0.getOpcode() == ISD::TRUNCATE)
1952 CTPOP = N0.getOperand(0);
1954 if (CTPOP.hasOneUse() && CTPOP.getOpcode() == ISD::CTPOP &&
1955 (N0 == CTPOP || N0.getValueType().getSizeInBits() >
1956 Log2_32_Ceil(CTPOP.getValueType().getSizeInBits()))) {
1957 EVT CTVT = CTPOP.getValueType();
1958 SDValue CTOp = CTPOP.getOperand(0);
1960 // (ctpop x) u< 2 -> (x & x-1) == 0
1961 // (ctpop x) u> 1 -> (x & x-1) != 0
1962 if ((Cond == ISD::SETULT && C1 == 2) || (Cond == ISD::SETUGT && C1 == 1)){
1963 SDValue Sub = DAG.getNode(ISD::SUB, dl, CTVT, CTOp,
1964 DAG.getConstant(1, CTVT));
1965 SDValue And = DAG.getNode(ISD::AND, dl, CTVT, CTOp, Sub);
1966 ISD::CondCode CC = Cond == ISD::SETULT ? ISD::SETEQ : ISD::SETNE;
1967 return DAG.getSetCC(dl, VT, And, DAG.getConstant(0, CTVT), CC);
1970 // TODO: (ctpop x) == 1 -> x && (x & x-1) == 0 iff ctpop is illegal.
1973 // (zext x) == C --> x == (trunc C)
1974 if (DCI.isBeforeLegalize() && N0->hasOneUse() &&
1975 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1976 unsigned MinBits = N0.getValueSizeInBits();
1978 if (N0->getOpcode() == ISD::ZERO_EXTEND) {
1980 MinBits = N0->getOperand(0).getValueSizeInBits();
1981 PreZExt = N0->getOperand(0);
1982 } else if (N0->getOpcode() == ISD::AND) {
1983 // DAGCombine turns costly ZExts into ANDs
1984 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0->getOperand(1)))
1985 if ((C->getAPIntValue()+1).isPowerOf2()) {
1986 MinBits = C->getAPIntValue().countTrailingOnes();
1987 PreZExt = N0->getOperand(0);
1989 } else if (LoadSDNode *LN0 = dyn_cast<LoadSDNode>(N0)) {
1991 if (LN0->getExtensionType() == ISD::ZEXTLOAD) {
1992 MinBits = LN0->getMemoryVT().getSizeInBits();
1997 // Make sure we're not loosing bits from the constant.
1998 if (MinBits < C1.getBitWidth() && MinBits > C1.getActiveBits()) {
1999 EVT MinVT = EVT::getIntegerVT(*DAG.getContext(), MinBits);
2000 if (isTypeDesirableForOp(ISD::SETCC, MinVT)) {
2001 // Will get folded away.
2002 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, MinVT, PreZExt);
2003 SDValue C = DAG.getConstant(C1.trunc(MinBits), MinVT);
2004 return DAG.getSetCC(dl, VT, Trunc, C, Cond);
2009 // If the LHS is '(and load, const)', the RHS is 0,
2010 // the test is for equality or unsigned, and all 1 bits of the const are
2011 // in the same partial word, see if we can shorten the load.
2012 if (DCI.isBeforeLegalize() &&
2013 N0.getOpcode() == ISD::AND && C1 == 0 &&
2014 N0.getNode()->hasOneUse() &&
2015 isa<LoadSDNode>(N0.getOperand(0)) &&
2016 N0.getOperand(0).getNode()->hasOneUse() &&
2017 isa<ConstantSDNode>(N0.getOperand(1))) {
2018 LoadSDNode *Lod = cast<LoadSDNode>(N0.getOperand(0));
2020 unsigned bestWidth = 0, bestOffset = 0;
2021 if (!Lod->isVolatile() && Lod->isUnindexed()) {
2022 unsigned origWidth = N0.getValueType().getSizeInBits();
2023 unsigned maskWidth = origWidth;
2024 // We can narrow (e.g.) 16-bit extending loads on 32-bit target to
2025 // 8 bits, but have to be careful...
2026 if (Lod->getExtensionType() != ISD::NON_EXTLOAD)
2027 origWidth = Lod->getMemoryVT().getSizeInBits();
2029 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
2030 for (unsigned width = origWidth / 2; width>=8; width /= 2) {
2031 APInt newMask = APInt::getLowBitsSet(maskWidth, width);
2032 for (unsigned offset=0; offset<origWidth/width; offset++) {
2033 if ((newMask & Mask) == Mask) {
2034 if (!TD->isLittleEndian())
2035 bestOffset = (origWidth/width - offset - 1) * (width/8);
2037 bestOffset = (uint64_t)offset * (width/8);
2038 bestMask = Mask.lshr(offset * (width/8) * 8);
2042 newMask = newMask << width;
2047 EVT newVT = EVT::getIntegerVT(*DAG.getContext(), bestWidth);
2048 if (newVT.isRound()) {
2049 EVT PtrType = Lod->getOperand(1).getValueType();
2050 SDValue Ptr = Lod->getBasePtr();
2051 if (bestOffset != 0)
2052 Ptr = DAG.getNode(ISD::ADD, dl, PtrType, Lod->getBasePtr(),
2053 DAG.getConstant(bestOffset, PtrType));
2054 unsigned NewAlign = MinAlign(Lod->getAlignment(), bestOffset);
2055 SDValue NewLoad = DAG.getLoad(newVT, dl, Lod->getChain(), Ptr,
2056 Lod->getPointerInfo().getWithOffset(bestOffset),
2057 false, false, NewAlign);
2058 return DAG.getSetCC(dl, VT,
2059 DAG.getNode(ISD::AND, dl, newVT, NewLoad,
2060 DAG.getConstant(bestMask.trunc(bestWidth),
2062 DAG.getConstant(0LL, newVT), Cond);
2067 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
2068 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
2069 unsigned InSize = N0.getOperand(0).getValueType().getSizeInBits();
2071 // If the comparison constant has bits in the upper part, the
2072 // zero-extended value could never match.
2073 if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(),
2074 C1.getBitWidth() - InSize))) {
2078 case ISD::SETEQ: return DAG.getConstant(0, VT);
2081 case ISD::SETNE: return DAG.getConstant(1, VT);
2084 // True if the sign bit of C1 is set.
2085 return DAG.getConstant(C1.isNegative(), VT);
2088 // True if the sign bit of C1 isn't set.
2089 return DAG.getConstant(C1.isNonNegative(), VT);
2095 // Otherwise, we can perform the comparison with the low bits.
2103 EVT newVT = N0.getOperand(0).getValueType();
2104 if (DCI.isBeforeLegalizeOps() ||
2105 (isOperationLegal(ISD::SETCC, newVT) &&
2106 getCondCodeAction(Cond, newVT)==Legal))
2107 return DAG.getSetCC(dl, VT, N0.getOperand(0),
2108 DAG.getConstant(C1.trunc(InSize), newVT),
2113 break; // todo, be more careful with signed comparisons
2115 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
2116 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
2117 EVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
2118 unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits();
2119 EVT ExtDstTy = N0.getValueType();
2120 unsigned ExtDstTyBits = ExtDstTy.getSizeInBits();
2122 // If the constant doesn't fit into the number of bits for the source of
2123 // the sign extension, it is impossible for both sides to be equal.
2124 if (C1.getMinSignedBits() > ExtSrcTyBits)
2125 return DAG.getConstant(Cond == ISD::SETNE, VT);
2128 EVT Op0Ty = N0.getOperand(0).getValueType();
2129 if (Op0Ty == ExtSrcTy) {
2130 ZextOp = N0.getOperand(0);
2132 APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits);
2133 ZextOp = DAG.getNode(ISD::AND, dl, Op0Ty, N0.getOperand(0),
2134 DAG.getConstant(Imm, Op0Ty));
2136 if (!DCI.isCalledByLegalizer())
2137 DCI.AddToWorklist(ZextOp.getNode());
2138 // Otherwise, make this a use of a zext.
2139 return DAG.getSetCC(dl, VT, ZextOp,
2140 DAG.getConstant(C1 & APInt::getLowBitsSet(
2145 } else if ((N1C->isNullValue() || N1C->getAPIntValue() == 1) &&
2146 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
2147 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC
2148 if (N0.getOpcode() == ISD::SETCC &&
2149 isTypeLegal(VT) && VT.bitsLE(N0.getValueType())) {
2150 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getAPIntValue() != 1);
2152 return DAG.getNode(ISD::TRUNCATE, dl, VT, N0);
2153 // Invert the condition.
2154 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
2155 CC = ISD::getSetCCInverse(CC,
2156 N0.getOperand(0).getValueType().isInteger());
2157 return DAG.getSetCC(dl, VT, N0.getOperand(0), N0.getOperand(1), CC);
2160 if ((N0.getOpcode() == ISD::XOR ||
2161 (N0.getOpcode() == ISD::AND &&
2162 N0.getOperand(0).getOpcode() == ISD::XOR &&
2163 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
2164 isa<ConstantSDNode>(N0.getOperand(1)) &&
2165 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue() == 1) {
2166 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We
2167 // can only do this if the top bits are known zero.
2168 unsigned BitWidth = N0.getValueSizeInBits();
2169 if (DAG.MaskedValueIsZero(N0,
2170 APInt::getHighBitsSet(BitWidth,
2172 // Okay, get the un-inverted input value.
2174 if (N0.getOpcode() == ISD::XOR)
2175 Val = N0.getOperand(0);
2177 assert(N0.getOpcode() == ISD::AND &&
2178 N0.getOperand(0).getOpcode() == ISD::XOR);
2179 // ((X^1)&1)^1 -> X & 1
2180 Val = DAG.getNode(ISD::AND, dl, N0.getValueType(),
2181 N0.getOperand(0).getOperand(0),
2185 return DAG.getSetCC(dl, VT, Val, N1,
2186 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
2188 } else if (N1C->getAPIntValue() == 1 &&
2190 getBooleanContents() == ZeroOrOneBooleanContent)) {
2192 if (Op0.getOpcode() == ISD::TRUNCATE)
2193 Op0 = Op0.getOperand(0);
2195 if ((Op0.getOpcode() == ISD::XOR) &&
2196 Op0.getOperand(0).getOpcode() == ISD::SETCC &&
2197 Op0.getOperand(1).getOpcode() == ISD::SETCC) {
2198 // (xor (setcc), (setcc)) == / != 1 -> (setcc) != / == (setcc)
2199 Cond = (Cond == ISD::SETEQ) ? ISD::SETNE : ISD::SETEQ;
2200 return DAG.getSetCC(dl, VT, Op0.getOperand(0), Op0.getOperand(1),
2202 } else if (Op0.getOpcode() == ISD::AND &&
2203 isa<ConstantSDNode>(Op0.getOperand(1)) &&
2204 cast<ConstantSDNode>(Op0.getOperand(1))->getAPIntValue() == 1) {
2205 // If this is (X&1) == / != 1, normalize it to (X&1) != / == 0.
2206 if (Op0.getValueType().bitsGT(VT))
2207 Op0 = DAG.getNode(ISD::AND, dl, VT,
2208 DAG.getNode(ISD::TRUNCATE, dl, VT, Op0.getOperand(0)),
2209 DAG.getConstant(1, VT));
2210 else if (Op0.getValueType().bitsLT(VT))
2211 Op0 = DAG.getNode(ISD::AND, dl, VT,
2212 DAG.getNode(ISD::ANY_EXTEND, dl, VT, Op0.getOperand(0)),
2213 DAG.getConstant(1, VT));
2215 return DAG.getSetCC(dl, VT, Op0,
2216 DAG.getConstant(0, Op0.getValueType()),
2217 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
2222 APInt MinVal, MaxVal;
2223 unsigned OperandBitSize = N1C->getValueType(0).getSizeInBits();
2224 if (ISD::isSignedIntSetCC(Cond)) {
2225 MinVal = APInt::getSignedMinValue(OperandBitSize);
2226 MaxVal = APInt::getSignedMaxValue(OperandBitSize);
2228 MinVal = APInt::getMinValue(OperandBitSize);
2229 MaxVal = APInt::getMaxValue(OperandBitSize);
2232 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
2233 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
2234 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true
2235 // X >= C0 --> X > (C0-1)
2236 return DAG.getSetCC(dl, VT, N0,
2237 DAG.getConstant(C1-1, N1.getValueType()),
2238 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
2241 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
2242 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true
2243 // X <= C0 --> X < (C0+1)
2244 return DAG.getSetCC(dl, VT, N0,
2245 DAG.getConstant(C1+1, N1.getValueType()),
2246 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
2249 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
2250 return DAG.getConstant(0, VT); // X < MIN --> false
2251 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal)
2252 return DAG.getConstant(1, VT); // X >= MIN --> true
2253 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal)
2254 return DAG.getConstant(0, VT); // X > MAX --> false
2255 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal)
2256 return DAG.getConstant(1, VT); // X <= MAX --> true
2258 // Canonicalize setgt X, Min --> setne X, Min
2259 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
2260 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
2261 // Canonicalize setlt X, Max --> setne X, Max
2262 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
2263 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
2265 // If we have setult X, 1, turn it into seteq X, 0
2266 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
2267 return DAG.getSetCC(dl, VT, N0,
2268 DAG.getConstant(MinVal, N0.getValueType()),
2270 // If we have setugt X, Max-1, turn it into seteq X, Max
2271 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
2272 return DAG.getSetCC(dl, VT, N0,
2273 DAG.getConstant(MaxVal, N0.getValueType()),
2276 // If we have "setcc X, C0", check to see if we can shrink the immediate
2279 // SETUGT X, SINTMAX -> SETLT X, 0
2280 if (Cond == ISD::SETUGT &&
2281 C1 == APInt::getSignedMaxValue(OperandBitSize))
2282 return DAG.getSetCC(dl, VT, N0,
2283 DAG.getConstant(0, N1.getValueType()),
2286 // SETULT X, SINTMIN -> SETGT X, -1
2287 if (Cond == ISD::SETULT &&
2288 C1 == APInt::getSignedMinValue(OperandBitSize)) {
2289 SDValue ConstMinusOne =
2290 DAG.getConstant(APInt::getAllOnesValue(OperandBitSize),
2292 return DAG.getSetCC(dl, VT, N0, ConstMinusOne, ISD::SETGT);
2295 // Fold bit comparisons when we can.
2296 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
2297 (VT == N0.getValueType() ||
2298 (isTypeLegal(VT) && VT.bitsLE(N0.getValueType()))) &&
2299 N0.getOpcode() == ISD::AND)
2300 if (ConstantSDNode *AndRHS =
2301 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2302 EVT ShiftTy = DCI.isBeforeLegalize() ?
2303 getPointerTy() : getShiftAmountTy(N0.getValueType());
2304 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
2305 // Perform the xform if the AND RHS is a single bit.
2306 if (AndRHS->getAPIntValue().isPowerOf2()) {
2307 return DAG.getNode(ISD::TRUNCATE, dl, VT,
2308 DAG.getNode(ISD::SRL, dl, N0.getValueType(), N0,
2309 DAG.getConstant(AndRHS->getAPIntValue().logBase2(), ShiftTy)));
2311 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getAPIntValue()) {
2312 // (X & 8) == 8 --> (X & 8) >> 3
2313 // Perform the xform if C1 is a single bit.
2314 if (C1.isPowerOf2()) {
2315 return DAG.getNode(ISD::TRUNCATE, dl, VT,
2316 DAG.getNode(ISD::SRL, dl, N0.getValueType(), N0,
2317 DAG.getConstant(C1.logBase2(), ShiftTy)));
2323 if (isa<ConstantFPSDNode>(N0.getNode())) {
2324 // Constant fold or commute setcc.
2325 SDValue O = DAG.FoldSetCC(VT, N0, N1, Cond, dl);
2326 if (O.getNode()) return O;
2327 } else if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1.getNode())) {
2328 // If the RHS of an FP comparison is a constant, simplify it away in
2330 if (CFP->getValueAPF().isNaN()) {
2331 // If an operand is known to be a nan, we can fold it.
2332 switch (ISD::getUnorderedFlavor(Cond)) {
2333 default: llvm_unreachable("Unknown flavor!");
2334 case 0: // Known false.
2335 return DAG.getConstant(0, VT);
2336 case 1: // Known true.
2337 return DAG.getConstant(1, VT);
2338 case 2: // Undefined.
2339 return DAG.getUNDEF(VT);
2343 // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the
2344 // constant if knowing that the operand is non-nan is enough. We prefer to
2345 // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to
2347 if (Cond == ISD::SETO || Cond == ISD::SETUO)
2348 return DAG.getSetCC(dl, VT, N0, N0, Cond);
2350 // If the condition is not legal, see if we can find an equivalent one
2352 if (!isCondCodeLegal(Cond, N0.getValueType())) {
2353 // If the comparison was an awkward floating-point == or != and one of
2354 // the comparison operands is infinity or negative infinity, convert the
2355 // condition to a less-awkward <= or >=.
2356 if (CFP->getValueAPF().isInfinity()) {
2357 if (CFP->getValueAPF().isNegative()) {
2358 if (Cond == ISD::SETOEQ &&
2359 isCondCodeLegal(ISD::SETOLE, N0.getValueType()))
2360 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLE);
2361 if (Cond == ISD::SETUEQ &&
2362 isCondCodeLegal(ISD::SETOLE, N0.getValueType()))
2363 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULE);
2364 if (Cond == ISD::SETUNE &&
2365 isCondCodeLegal(ISD::SETUGT, N0.getValueType()))
2366 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGT);
2367 if (Cond == ISD::SETONE &&
2368 isCondCodeLegal(ISD::SETUGT, N0.getValueType()))
2369 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGT);
2371 if (Cond == ISD::SETOEQ &&
2372 isCondCodeLegal(ISD::SETOGE, N0.getValueType()))
2373 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGE);
2374 if (Cond == ISD::SETUEQ &&
2375 isCondCodeLegal(ISD::SETOGE, N0.getValueType()))
2376 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGE);
2377 if (Cond == ISD::SETUNE &&
2378 isCondCodeLegal(ISD::SETULT, N0.getValueType()))
2379 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULT);
2380 if (Cond == ISD::SETONE &&
2381 isCondCodeLegal(ISD::SETULT, N0.getValueType()))
2382 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLT);
2389 // We can always fold X == X for integer setcc's.
2390 if (N0.getValueType().isInteger())
2391 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
2392 unsigned UOF = ISD::getUnorderedFlavor(Cond);
2393 if (UOF == 2) // FP operators that are undefined on NaNs.
2394 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
2395 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
2396 return DAG.getConstant(UOF, VT);
2397 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
2398 // if it is not already.
2399 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
2400 if (NewCond != Cond)
2401 return DAG.getSetCC(dl, VT, N0, N1, NewCond);
2404 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
2405 N0.getValueType().isInteger()) {
2406 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
2407 N0.getOpcode() == ISD::XOR) {
2408 // Simplify (X+Y) == (X+Z) --> Y == Z
2409 if (N0.getOpcode() == N1.getOpcode()) {
2410 if (N0.getOperand(0) == N1.getOperand(0))
2411 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(1), Cond);
2412 if (N0.getOperand(1) == N1.getOperand(1))
2413 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(0), Cond);
2414 if (DAG.isCommutativeBinOp(N0.getOpcode())) {
2415 // If X op Y == Y op X, try other combinations.
2416 if (N0.getOperand(0) == N1.getOperand(1))
2417 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(0),
2419 if (N0.getOperand(1) == N1.getOperand(0))
2420 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(1),
2425 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) {
2426 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2427 // Turn (X+C1) == C2 --> X == C2-C1
2428 if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse()) {
2429 return DAG.getSetCC(dl, VT, N0.getOperand(0),
2430 DAG.getConstant(RHSC->getAPIntValue()-
2431 LHSR->getAPIntValue(),
2432 N0.getValueType()), Cond);
2435 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
2436 if (N0.getOpcode() == ISD::XOR)
2437 // If we know that all of the inverted bits are zero, don't bother
2438 // performing the inversion.
2439 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue()))
2441 DAG.getSetCC(dl, VT, N0.getOperand(0),
2442 DAG.getConstant(LHSR->getAPIntValue() ^
2443 RHSC->getAPIntValue(),
2448 // Turn (C1-X) == C2 --> X == C1-C2
2449 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
2450 if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse()) {
2452 DAG.getSetCC(dl, VT, N0.getOperand(1),
2453 DAG.getConstant(SUBC->getAPIntValue() -
2454 RHSC->getAPIntValue(),
2461 // Simplify (X+Z) == X --> Z == 0
2462 if (N0.getOperand(0) == N1)
2463 return DAG.getSetCC(dl, VT, N0.getOperand(1),
2464 DAG.getConstant(0, N0.getValueType()), Cond);
2465 if (N0.getOperand(1) == N1) {
2466 if (DAG.isCommutativeBinOp(N0.getOpcode()))
2467 return DAG.getSetCC(dl, VT, N0.getOperand(0),
2468 DAG.getConstant(0, N0.getValueType()), Cond);
2469 else if (N0.getNode()->hasOneUse()) {
2470 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
2471 // (Z-X) == X --> Z == X<<1
2472 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(),
2474 DAG.getConstant(1, getShiftAmountTy(N1.getValueType())));
2475 if (!DCI.isCalledByLegalizer())
2476 DCI.AddToWorklist(SH.getNode());
2477 return DAG.getSetCC(dl, VT, N0.getOperand(0), SH, Cond);
2482 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
2483 N1.getOpcode() == ISD::XOR) {
2484 // Simplify X == (X+Z) --> Z == 0
2485 if (N1.getOperand(0) == N0) {
2486 return DAG.getSetCC(dl, VT, N1.getOperand(1),
2487 DAG.getConstant(0, N1.getValueType()), Cond);
2488 } else if (N1.getOperand(1) == N0) {
2489 if (DAG.isCommutativeBinOp(N1.getOpcode())) {
2490 return DAG.getSetCC(dl, VT, N1.getOperand(0),
2491 DAG.getConstant(0, N1.getValueType()), Cond);
2492 } else if (N1.getNode()->hasOneUse()) {
2493 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
2494 // X == (Z-X) --> X<<1 == Z
2495 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(), N0,
2496 DAG.getConstant(1, getShiftAmountTy(N0.getValueType())));
2497 if (!DCI.isCalledByLegalizer())
2498 DCI.AddToWorklist(SH.getNode());
2499 return DAG.getSetCC(dl, VT, SH, N1.getOperand(0), Cond);
2504 // Simplify x&y == y to x&y != 0 if y has exactly one bit set.
2505 // Note that where y is variable and is known to have at most
2506 // one bit set (for example, if it is z&1) we cannot do this;
2507 // the expressions are not equivalent when y==0.
2508 if (N0.getOpcode() == ISD::AND)
2509 if (N0.getOperand(0) == N1 || N0.getOperand(1) == N1) {
2510 if (ValueHasExactlyOneBitSet(N1, DAG)) {
2511 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true);
2512 SDValue Zero = DAG.getConstant(0, N1.getValueType());
2513 return DAG.getSetCC(dl, VT, N0, Zero, Cond);
2516 if (N1.getOpcode() == ISD::AND)
2517 if (N1.getOperand(0) == N0 || N1.getOperand(1) == N0) {
2518 if (ValueHasExactlyOneBitSet(N0, DAG)) {
2519 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true);
2520 SDValue Zero = DAG.getConstant(0, N0.getValueType());
2521 return DAG.getSetCC(dl, VT, N1, Zero, Cond);
2526 // Fold away ALL boolean setcc's.
2528 if (N0.getValueType() == MVT::i1 && foldBooleans) {
2530 default: llvm_unreachable("Unknown integer setcc!");
2531 case ISD::SETEQ: // X == Y -> ~(X^Y)
2532 Temp = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1);
2533 N0 = DAG.getNOT(dl, Temp, MVT::i1);
2534 if (!DCI.isCalledByLegalizer())
2535 DCI.AddToWorklist(Temp.getNode());
2537 case ISD::SETNE: // X != Y --> (X^Y)
2538 N0 = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1);
2540 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> ~X & Y
2541 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> ~X & Y
2542 Temp = DAG.getNOT(dl, N0, MVT::i1);
2543 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N1, Temp);
2544 if (!DCI.isCalledByLegalizer())
2545 DCI.AddToWorklist(Temp.getNode());
2547 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> ~Y & X
2548 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> ~Y & X
2549 Temp = DAG.getNOT(dl, N1, MVT::i1);
2550 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N0, Temp);
2551 if (!DCI.isCalledByLegalizer())
2552 DCI.AddToWorklist(Temp.getNode());
2554 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> ~X | Y
2555 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> ~X | Y
2556 Temp = DAG.getNOT(dl, N0, MVT::i1);
2557 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N1, Temp);
2558 if (!DCI.isCalledByLegalizer())
2559 DCI.AddToWorklist(Temp.getNode());
2561 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> ~Y | X
2562 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> ~Y | X
2563 Temp = DAG.getNOT(dl, N1, MVT::i1);
2564 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N0, Temp);
2567 if (VT != MVT::i1) {
2568 if (!DCI.isCalledByLegalizer())
2569 DCI.AddToWorklist(N0.getNode());
2570 // FIXME: If running after legalize, we probably can't do this.
2571 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, N0);
2576 // Could not fold it.
2580 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
2581 /// node is a GlobalAddress + offset.
2582 bool TargetLowering::isGAPlusOffset(SDNode *N, const GlobalValue *&GA,
2583 int64_t &Offset) const {
2584 if (isa<GlobalAddressSDNode>(N)) {
2585 GlobalAddressSDNode *GASD = cast<GlobalAddressSDNode>(N);
2586 GA = GASD->getGlobal();
2587 Offset += GASD->getOffset();
2591 if (N->getOpcode() == ISD::ADD) {
2592 SDValue N1 = N->getOperand(0);
2593 SDValue N2 = N->getOperand(1);
2594 if (isGAPlusOffset(N1.getNode(), GA, Offset)) {
2595 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
2597 Offset += V->getSExtValue();
2600 } else if (isGAPlusOffset(N2.getNode(), GA, Offset)) {
2601 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
2603 Offset += V->getSExtValue();
2613 SDValue TargetLowering::
2614 PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
2615 // Default implementation: no optimization.
2619 //===----------------------------------------------------------------------===//
2620 // Inline Assembler Implementation Methods
2621 //===----------------------------------------------------------------------===//
2624 TargetLowering::ConstraintType
2625 TargetLowering::getConstraintType(const std::string &Constraint) const {
2626 // FIXME: lots more standard ones to handle.
2627 if (Constraint.size() == 1) {
2628 switch (Constraint[0]) {
2630 case 'r': return C_RegisterClass;
2632 case 'o': // offsetable
2633 case 'V': // not offsetable
2635 case 'i': // Simple Integer or Relocatable Constant
2636 case 'n': // Simple Integer
2637 case 'E': // Floating Point Constant
2638 case 'F': // Floating Point Constant
2639 case 's': // Relocatable Constant
2640 case 'p': // Address.
2641 case 'X': // Allow ANY value.
2642 case 'I': // Target registers.
2656 if (Constraint.size() > 1 && Constraint[0] == '{' &&
2657 Constraint[Constraint.size()-1] == '}')
2662 /// LowerXConstraint - try to replace an X constraint, which matches anything,
2663 /// with another that has more specific requirements based on the type of the
2664 /// corresponding operand.
2665 const char *TargetLowering::LowerXConstraint(EVT ConstraintVT) const{
2666 if (ConstraintVT.isInteger())
2668 if (ConstraintVT.isFloatingPoint())
2669 return "f"; // works for many targets
2673 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
2674 /// vector. If it is invalid, don't add anything to Ops.
2675 void TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
2676 std::string &Constraint,
2677 std::vector<SDValue> &Ops,
2678 SelectionDAG &DAG) const {
2680 if (Constraint.length() > 1) return;
2682 char ConstraintLetter = Constraint[0];
2683 switch (ConstraintLetter) {
2685 case 'X': // Allows any operand; labels (basic block) use this.
2686 if (Op.getOpcode() == ISD::BasicBlock) {
2691 case 'i': // Simple Integer or Relocatable Constant
2692 case 'n': // Simple Integer
2693 case 's': { // Relocatable Constant
2694 // These operands are interested in values of the form (GV+C), where C may
2695 // be folded in as an offset of GV, or it may be explicitly added. Also, it
2696 // is possible and fine if either GV or C are missing.
2697 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2698 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
2700 // If we have "(add GV, C)", pull out GV/C
2701 if (Op.getOpcode() == ISD::ADD) {
2702 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
2703 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
2704 if (C == 0 || GA == 0) {
2705 C = dyn_cast<ConstantSDNode>(Op.getOperand(0));
2706 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1));
2708 if (C == 0 || GA == 0)
2712 // If we find a valid operand, map to the TargetXXX version so that the
2713 // value itself doesn't get selected.
2714 if (GA) { // Either &GV or &GV+C
2715 if (ConstraintLetter != 'n') {
2716 int64_t Offs = GA->getOffset();
2717 if (C) Offs += C->getZExtValue();
2718 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(),
2719 C ? C->getDebugLoc() : DebugLoc(),
2720 Op.getValueType(), Offs));
2724 if (C) { // just C, no GV.
2725 // Simple constants are not allowed for 's'.
2726 if (ConstraintLetter != 's') {
2727 // gcc prints these as sign extended. Sign extend value to 64 bits
2728 // now; without this it would get ZExt'd later in
2729 // ScheduleDAGSDNodes::EmitNode, which is very generic.
2730 Ops.push_back(DAG.getTargetConstant(C->getAPIntValue().getSExtValue(),
2740 std::vector<unsigned> TargetLowering::
2741 getRegClassForInlineAsmConstraint(const std::string &Constraint,
2743 return std::vector<unsigned>();
2747 std::pair<unsigned, const TargetRegisterClass*> TargetLowering::
2748 getRegForInlineAsmConstraint(const std::string &Constraint,
2750 if (Constraint[0] != '{')
2751 return std::make_pair(0u, static_cast<TargetRegisterClass*>(0));
2752 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?");
2754 // Remove the braces from around the name.
2755 StringRef RegName(Constraint.data()+1, Constraint.size()-2);
2757 // Figure out which register class contains this reg.
2758 const TargetRegisterInfo *RI = TM.getRegisterInfo();
2759 for (TargetRegisterInfo::regclass_iterator RCI = RI->regclass_begin(),
2760 E = RI->regclass_end(); RCI != E; ++RCI) {
2761 const TargetRegisterClass *RC = *RCI;
2763 // If none of the value types for this register class are valid, we
2764 // can't use it. For example, 64-bit reg classes on 32-bit targets.
2765 bool isLegal = false;
2766 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
2768 if (isTypeLegal(*I)) {
2774 if (!isLegal) continue;
2776 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end();
2778 if (RegName.equals_lower(RI->getName(*I)))
2779 return std::make_pair(*I, RC);
2783 return std::make_pair(0u, static_cast<const TargetRegisterClass*>(0));
2786 //===----------------------------------------------------------------------===//
2787 // Constraint Selection.
2789 /// isMatchingInputConstraint - Return true of this is an input operand that is
2790 /// a matching constraint like "4".
2791 bool TargetLowering::AsmOperandInfo::isMatchingInputConstraint() const {
2792 assert(!ConstraintCode.empty() && "No known constraint!");
2793 return isdigit(ConstraintCode[0]);
2796 /// getMatchedOperand - If this is an input matching constraint, this method
2797 /// returns the output operand it matches.
2798 unsigned TargetLowering::AsmOperandInfo::getMatchedOperand() const {
2799 assert(!ConstraintCode.empty() && "No known constraint!");
2800 return atoi(ConstraintCode.c_str());
2804 /// ParseConstraints - Split up the constraint string from the inline
2805 /// assembly value into the specific constraints and their prefixes,
2806 /// and also tie in the associated operand values.
2807 /// If this returns an empty vector, and if the constraint string itself
2808 /// isn't empty, there was an error parsing.
2809 TargetLowering::AsmOperandInfoVector TargetLowering::ParseConstraints(
2810 ImmutableCallSite CS) const {
2811 /// ConstraintOperands - Information about all of the constraints.
2812 AsmOperandInfoVector ConstraintOperands;
2813 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
2814 unsigned maCount = 0; // Largest number of multiple alternative constraints.
2816 // Do a prepass over the constraints, canonicalizing them, and building up the
2817 // ConstraintOperands list.
2818 InlineAsm::ConstraintInfoVector
2819 ConstraintInfos = IA->ParseConstraints();
2821 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
2822 unsigned ResNo = 0; // ResNo - The result number of the next output.
2824 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
2825 ConstraintOperands.push_back(AsmOperandInfo(ConstraintInfos[i]));
2826 AsmOperandInfo &OpInfo = ConstraintOperands.back();
2828 // Update multiple alternative constraint count.
2829 if (OpInfo.multipleAlternatives.size() > maCount)
2830 maCount = OpInfo.multipleAlternatives.size();
2832 OpInfo.ConstraintVT = MVT::Other;
2834 // Compute the value type for each operand.
2835 switch (OpInfo.Type) {
2836 case InlineAsm::isOutput:
2837 // Indirect outputs just consume an argument.
2838 if (OpInfo.isIndirect) {
2839 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
2843 // The return value of the call is this value. As such, there is no
2844 // corresponding argument.
2845 assert(!CS.getType()->isVoidTy() &&
2847 if (const StructType *STy = dyn_cast<StructType>(CS.getType())) {
2848 OpInfo.ConstraintVT = getValueType(STy->getElementType(ResNo));
2850 assert(ResNo == 0 && "Asm only has one result!");
2851 OpInfo.ConstraintVT = getValueType(CS.getType());
2855 case InlineAsm::isInput:
2856 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
2858 case InlineAsm::isClobber:
2863 if (OpInfo.CallOperandVal) {
2864 const llvm::Type *OpTy = OpInfo.CallOperandVal->getType();
2865 if (OpInfo.isIndirect) {
2866 const llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy);
2868 report_fatal_error("Indirect operand for inline asm not a pointer!");
2869 OpTy = PtrTy->getElementType();
2872 // Look for vector wrapped in a struct. e.g. { <16 x i8> }.
2873 if (const StructType *STy = dyn_cast<StructType>(OpTy))
2874 if (STy->getNumElements() == 1)
2875 OpTy = STy->getElementType(0);
2877 // If OpTy is not a single value, it may be a struct/union that we
2878 // can tile with integers.
2879 if (!OpTy->isSingleValueType() && OpTy->isSized()) {
2880 unsigned BitSize = TD->getTypeSizeInBits(OpTy);
2889 OpInfo.ConstraintVT =
2890 EVT::getEVT(IntegerType::get(OpTy->getContext(), BitSize), true);
2893 } else if (dyn_cast<PointerType>(OpTy)) {
2894 OpInfo.ConstraintVT = MVT::getIntegerVT(8*TD->getPointerSize());
2896 OpInfo.ConstraintVT = EVT::getEVT(OpTy, true);
2901 // If we have multiple alternative constraints, select the best alternative.
2902 if (ConstraintInfos.size()) {
2904 unsigned bestMAIndex = 0;
2905 int bestWeight = -1;
2906 // weight: -1 = invalid match, and 0 = so-so match to 5 = good match.
2909 // Compute the sums of the weights for each alternative, keeping track
2910 // of the best (highest weight) one so far.
2911 for (maIndex = 0; maIndex < maCount; ++maIndex) {
2913 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size();
2914 cIndex != eIndex; ++cIndex) {
2915 AsmOperandInfo& OpInfo = ConstraintOperands[cIndex];
2916 if (OpInfo.Type == InlineAsm::isClobber)
2919 // If this is an output operand with a matching input operand,
2920 // look up the matching input. If their types mismatch, e.g. one
2921 // is an integer, the other is floating point, or their sizes are
2922 // different, flag it as an maCantMatch.
2923 if (OpInfo.hasMatchingInput()) {
2924 AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
2925 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
2926 if ((OpInfo.ConstraintVT.isInteger() !=
2927 Input.ConstraintVT.isInteger()) ||
2928 (OpInfo.ConstraintVT.getSizeInBits() !=
2929 Input.ConstraintVT.getSizeInBits())) {
2930 weightSum = -1; // Can't match.
2935 weight = getMultipleConstraintMatchWeight(OpInfo, maIndex);
2940 weightSum += weight;
2943 if (weightSum > bestWeight) {
2944 bestWeight = weightSum;
2945 bestMAIndex = maIndex;
2949 // Now select chosen alternative in each constraint.
2950 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size();
2951 cIndex != eIndex; ++cIndex) {
2952 AsmOperandInfo& cInfo = ConstraintOperands[cIndex];
2953 if (cInfo.Type == InlineAsm::isClobber)
2955 cInfo.selectAlternative(bestMAIndex);
2960 // Check and hook up tied operands, choose constraint code to use.
2961 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size();
2962 cIndex != eIndex; ++cIndex) {
2963 AsmOperandInfo& OpInfo = ConstraintOperands[cIndex];
2965 // If this is an output operand with a matching input operand, look up the
2966 // matching input. If their types mismatch, e.g. one is an integer, the
2967 // other is floating point, or their sizes are different, flag it as an
2969 if (OpInfo.hasMatchingInput()) {
2970 AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
2972 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
2973 if ((OpInfo.ConstraintVT.isInteger() !=
2974 Input.ConstraintVT.isInteger()) ||
2975 (OpInfo.ConstraintVT.getSizeInBits() !=
2976 Input.ConstraintVT.getSizeInBits())) {
2977 report_fatal_error("Unsupported asm: input constraint"
2978 " with a matching output constraint of"
2979 " incompatible type!");
2986 return ConstraintOperands;
2990 /// getConstraintGenerality - Return an integer indicating how general CT
2992 static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) {
2994 default: llvm_unreachable("Unknown constraint type!");
2995 case TargetLowering::C_Other:
2996 case TargetLowering::C_Unknown:
2998 case TargetLowering::C_Register:
3000 case TargetLowering::C_RegisterClass:
3002 case TargetLowering::C_Memory:
3007 /// Examine constraint type and operand type and determine a weight value.
3008 /// This object must already have been set up with the operand type
3009 /// and the current alternative constraint selected.
3010 TargetLowering::ConstraintWeight
3011 TargetLowering::getMultipleConstraintMatchWeight(
3012 AsmOperandInfo &info, int maIndex) const {
3013 InlineAsm::ConstraintCodeVector *rCodes;
3014 if (maIndex >= (int)info.multipleAlternatives.size())
3015 rCodes = &info.Codes;
3017 rCodes = &info.multipleAlternatives[maIndex].Codes;
3018 ConstraintWeight BestWeight = CW_Invalid;
3020 // Loop over the options, keeping track of the most general one.
3021 for (unsigned i = 0, e = rCodes->size(); i != e; ++i) {
3022 ConstraintWeight weight =
3023 getSingleConstraintMatchWeight(info, (*rCodes)[i].c_str());
3024 if (weight > BestWeight)
3025 BestWeight = weight;
3031 /// Examine constraint type and operand type and determine a weight value.
3032 /// This object must already have been set up with the operand type
3033 /// and the current alternative constraint selected.
3034 TargetLowering::ConstraintWeight
3035 TargetLowering::getSingleConstraintMatchWeight(
3036 AsmOperandInfo &info, const char *constraint) const {
3037 ConstraintWeight weight = CW_Invalid;
3038 Value *CallOperandVal = info.CallOperandVal;
3039 // If we don't have a value, we can't do a match,
3040 // but allow it at the lowest weight.
3041 if (CallOperandVal == NULL)
3043 // Look at the constraint type.
3044 switch (*constraint) {
3045 case 'i': // immediate integer.
3046 case 'n': // immediate integer with a known value.
3047 if (isa<ConstantInt>(CallOperandVal))
3048 weight = CW_Constant;
3050 case 's': // non-explicit intregal immediate.
3051 if (isa<GlobalValue>(CallOperandVal))
3052 weight = CW_Constant;
3054 case 'E': // immediate float if host format.
3055 case 'F': // immediate float.
3056 if (isa<ConstantFP>(CallOperandVal))
3057 weight = CW_Constant;
3059 case '<': // memory operand with autodecrement.
3060 case '>': // memory operand with autoincrement.
3061 case 'm': // memory operand.
3062 case 'o': // offsettable memory operand
3063 case 'V': // non-offsettable memory operand
3066 case 'r': // general register.
3067 case 'g': // general register, memory operand or immediate integer.
3068 // note: Clang converts "g" to "imr".
3069 if (CallOperandVal->getType()->isIntegerTy())
3070 weight = CW_Register;
3072 case 'X': // any operand.
3074 weight = CW_Default;
3080 /// ChooseConstraint - If there are multiple different constraints that we
3081 /// could pick for this operand (e.g. "imr") try to pick the 'best' one.
3082 /// This is somewhat tricky: constraints fall into four classes:
3083 /// Other -> immediates and magic values
3084 /// Register -> one specific register
3085 /// RegisterClass -> a group of regs
3086 /// Memory -> memory
3087 /// Ideally, we would pick the most specific constraint possible: if we have
3088 /// something that fits into a register, we would pick it. The problem here
3089 /// is that if we have something that could either be in a register or in
3090 /// memory that use of the register could cause selection of *other*
3091 /// operands to fail: they might only succeed if we pick memory. Because of
3092 /// this the heuristic we use is:
3094 /// 1) If there is an 'other' constraint, and if the operand is valid for
3095 /// that constraint, use it. This makes us take advantage of 'i'
3096 /// constraints when available.
3097 /// 2) Otherwise, pick the most general constraint present. This prefers
3098 /// 'm' over 'r', for example.
3100 static void ChooseConstraint(TargetLowering::AsmOperandInfo &OpInfo,
3101 const TargetLowering &TLI,
3102 SDValue Op, SelectionDAG *DAG) {
3103 assert(OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options");
3104 unsigned BestIdx = 0;
3105 TargetLowering::ConstraintType BestType = TargetLowering::C_Unknown;
3106 int BestGenerality = -1;
3108 // Loop over the options, keeping track of the most general one.
3109 for (unsigned i = 0, e = OpInfo.Codes.size(); i != e; ++i) {
3110 TargetLowering::ConstraintType CType =
3111 TLI.getConstraintType(OpInfo.Codes[i]);
3113 // If this is an 'other' constraint, see if the operand is valid for it.
3114 // For example, on X86 we might have an 'rI' constraint. If the operand
3115 // is an integer in the range [0..31] we want to use I (saving a load
3116 // of a register), otherwise we must use 'r'.
3117 if (CType == TargetLowering::C_Other && Op.getNode()) {
3118 assert(OpInfo.Codes[i].size() == 1 &&
3119 "Unhandled multi-letter 'other' constraint");
3120 std::vector<SDValue> ResultOps;
3121 TLI.LowerAsmOperandForConstraint(Op, OpInfo.Codes[i],
3123 if (!ResultOps.empty()) {
3130 // Things with matching constraints can only be registers, per gcc
3131 // documentation. This mainly affects "g" constraints.
3132 if (CType == TargetLowering::C_Memory && OpInfo.hasMatchingInput())
3135 // This constraint letter is more general than the previous one, use it.
3136 int Generality = getConstraintGenerality(CType);
3137 if (Generality > BestGenerality) {
3140 BestGenerality = Generality;
3144 OpInfo.ConstraintCode = OpInfo.Codes[BestIdx];
3145 OpInfo.ConstraintType = BestType;
3148 /// ComputeConstraintToUse - Determines the constraint code and constraint
3149 /// type to use for the specific AsmOperandInfo, setting
3150 /// OpInfo.ConstraintCode and OpInfo.ConstraintType.
3151 void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo,
3153 SelectionDAG *DAG) const {
3154 assert(!OpInfo.Codes.empty() && "Must have at least one constraint");
3156 // Single-letter constraints ('r') are very common.
3157 if (OpInfo.Codes.size() == 1) {
3158 OpInfo.ConstraintCode = OpInfo.Codes[0];
3159 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
3161 ChooseConstraint(OpInfo, *this, Op, DAG);
3164 // 'X' matches anything.
3165 if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) {
3166 // Labels and constants are handled elsewhere ('X' is the only thing
3167 // that matches labels). For Functions, the type here is the type of
3168 // the result, which is not what we want to look at; leave them alone.
3169 Value *v = OpInfo.CallOperandVal;
3170 if (isa<BasicBlock>(v) || isa<ConstantInt>(v) || isa<Function>(v)) {
3171 OpInfo.CallOperandVal = v;
3175 // Otherwise, try to resolve it to something we know about by looking at
3176 // the actual operand type.
3177 if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) {
3178 OpInfo.ConstraintCode = Repl;
3179 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
3184 //===----------------------------------------------------------------------===//
3185 // Loop Strength Reduction hooks
3186 //===----------------------------------------------------------------------===//
3188 /// isLegalAddressingMode - Return true if the addressing mode represented
3189 /// by AM is legal for this target, for a load/store of the specified type.
3190 bool TargetLowering::isLegalAddressingMode(const AddrMode &AM,
3191 const Type *Ty) const {
3192 // The default implementation of this implements a conservative RISCy, r+r and
3195 // Allows a sign-extended 16-bit immediate field.
3196 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
3199 // No global is ever allowed as a base.
3203 // Only support r+r,
3205 case 0: // "r+i" or just "i", depending on HasBaseReg.
3208 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
3210 // Otherwise we have r+r or r+i.
3213 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
3215 // Allow 2*r as r+r.
3222 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
3223 /// return a DAG expression to select that will generate the same value by
3224 /// multiplying by a magic number. See:
3225 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
3226 SDValue TargetLowering::BuildSDIV(SDNode *N, SelectionDAG &DAG,
3227 std::vector<SDNode*>* Created) const {
3228 EVT VT = N->getValueType(0);
3229 DebugLoc dl= N->getDebugLoc();
3231 // Check to see if we can do this.
3232 // FIXME: We should be more aggressive here.
3233 if (!isTypeLegal(VT))
3236 APInt d = cast<ConstantSDNode>(N->getOperand(1))->getAPIntValue();
3237 APInt::ms magics = d.magic();
3239 // Multiply the numerator (operand 0) by the magic value
3240 // FIXME: We should support doing a MUL in a wider type
3242 if (isOperationLegalOrCustom(ISD::MULHS, VT))
3243 Q = DAG.getNode(ISD::MULHS, dl, VT, N->getOperand(0),
3244 DAG.getConstant(magics.m, VT));
3245 else if (isOperationLegalOrCustom(ISD::SMUL_LOHI, VT))
3246 Q = SDValue(DAG.getNode(ISD::SMUL_LOHI, dl, DAG.getVTList(VT, VT),
3248 DAG.getConstant(magics.m, VT)).getNode(), 1);
3250 return SDValue(); // No mulhs or equvialent
3251 // If d > 0 and m < 0, add the numerator
3252 if (d.isStrictlyPositive() && magics.m.isNegative()) {
3253 Q = DAG.getNode(ISD::ADD, dl, VT, Q, N->getOperand(0));
3255 Created->push_back(Q.getNode());
3257 // If d < 0 and m > 0, subtract the numerator.
3258 if (d.isNegative() && magics.m.isStrictlyPositive()) {
3259 Q = DAG.getNode(ISD::SUB, dl, VT, Q, N->getOperand(0));
3261 Created->push_back(Q.getNode());
3263 // Shift right algebraic if shift value is nonzero
3265 Q = DAG.getNode(ISD::SRA, dl, VT, Q,
3266 DAG.getConstant(magics.s, getShiftAmountTy(Q.getValueType())));
3268 Created->push_back(Q.getNode());
3270 // Extract the sign bit and add it to the quotient
3272 DAG.getNode(ISD::SRL, dl, VT, Q, DAG.getConstant(VT.getSizeInBits()-1,
3273 getShiftAmountTy(Q.getValueType())));
3275 Created->push_back(T.getNode());
3276 return DAG.getNode(ISD::ADD, dl, VT, Q, T);
3279 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
3280 /// return a DAG expression to select that will generate the same value by
3281 /// multiplying by a magic number. See:
3282 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
3283 SDValue TargetLowering::BuildUDIV(SDNode *N, SelectionDAG &DAG,
3284 std::vector<SDNode*>* Created) const {
3285 EVT VT = N->getValueType(0);
3286 DebugLoc dl = N->getDebugLoc();
3288 // Check to see if we can do this.
3289 // FIXME: We should be more aggressive here.
3290 if (!isTypeLegal(VT))
3293 // FIXME: We should use a narrower constant when the upper
3294 // bits are known to be zero.
3295 const APInt &N1C = cast<ConstantSDNode>(N->getOperand(1))->getAPIntValue();
3296 APInt::mu magics = N1C.magicu();
3298 SDValue Q = N->getOperand(0);
3300 // If the divisor is even, we can avoid using the expensive fixup by shifting
3301 // the divided value upfront.
3302 if (magics.a != 0 && !N1C[0]) {
3303 unsigned Shift = N1C.countTrailingZeros();
3304 Q = DAG.getNode(ISD::SRL, dl, VT, Q,
3305 DAG.getConstant(Shift, getShiftAmountTy(Q.getValueType())));
3307 Created->push_back(Q.getNode());
3309 // Get magic number for the shifted divisor.
3310 magics = N1C.lshr(Shift).magicu(Shift);
3311 assert(magics.a == 0 && "Should use cheap fixup now");
3314 // Multiply the numerator (operand 0) by the magic value
3315 // FIXME: We should support doing a MUL in a wider type
3316 if (isOperationLegalOrCustom(ISD::MULHU, VT))
3317 Q = DAG.getNode(ISD::MULHU, dl, VT, Q, DAG.getConstant(magics.m, VT));
3318 else if (isOperationLegalOrCustom(ISD::UMUL_LOHI, VT))
3319 Q = SDValue(DAG.getNode(ISD::UMUL_LOHI, dl, DAG.getVTList(VT, VT), Q,
3320 DAG.getConstant(magics.m, VT)).getNode(), 1);
3322 return SDValue(); // No mulhu or equvialent
3324 Created->push_back(Q.getNode());
3326 if (magics.a == 0) {
3327 assert(magics.s < N1C.getBitWidth() &&
3328 "We shouldn't generate an undefined shift!");
3329 return DAG.getNode(ISD::SRL, dl, VT, Q,
3330 DAG.getConstant(magics.s, getShiftAmountTy(Q.getValueType())));
3332 SDValue NPQ = DAG.getNode(ISD::SUB, dl, VT, N->getOperand(0), Q);
3334 Created->push_back(NPQ.getNode());
3335 NPQ = DAG.getNode(ISD::SRL, dl, VT, NPQ,
3336 DAG.getConstant(1, getShiftAmountTy(NPQ.getValueType())));
3338 Created->push_back(NPQ.getNode());
3339 NPQ = DAG.getNode(ISD::ADD, dl, VT, NPQ, Q);
3341 Created->push_back(NPQ.getNode());
3342 return DAG.getNode(ISD::SRL, dl, VT, NPQ,
3343 DAG.getConstant(magics.s-1, getShiftAmountTy(NPQ.getValueType())));