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/BitVector.h"
29 #include "llvm/ADT/STLExtras.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/ErrorHandling.h"
32 #include "llvm/Support/MathExtras.h"
36 /// We are in the process of implementing a new TypeLegalization action
37 /// - the promotion of vector elements. This feature is disabled by default
38 /// and only enabled using this flag.
40 AllowPromoteIntElem("promote-elements", cl::Hidden, cl::init(true),
41 cl::desc("Allow promotion of integer vector element types"));
43 /// InitLibcallNames - Set default libcall names.
45 static void InitLibcallNames(const char **Names) {
46 Names[RTLIB::SHL_I16] = "__ashlhi3";
47 Names[RTLIB::SHL_I32] = "__ashlsi3";
48 Names[RTLIB::SHL_I64] = "__ashldi3";
49 Names[RTLIB::SHL_I128] = "__ashlti3";
50 Names[RTLIB::SRL_I16] = "__lshrhi3";
51 Names[RTLIB::SRL_I32] = "__lshrsi3";
52 Names[RTLIB::SRL_I64] = "__lshrdi3";
53 Names[RTLIB::SRL_I128] = "__lshrti3";
54 Names[RTLIB::SRA_I16] = "__ashrhi3";
55 Names[RTLIB::SRA_I32] = "__ashrsi3";
56 Names[RTLIB::SRA_I64] = "__ashrdi3";
57 Names[RTLIB::SRA_I128] = "__ashrti3";
58 Names[RTLIB::MUL_I8] = "__mulqi3";
59 Names[RTLIB::MUL_I16] = "__mulhi3";
60 Names[RTLIB::MUL_I32] = "__mulsi3";
61 Names[RTLIB::MUL_I64] = "__muldi3";
62 Names[RTLIB::MUL_I128] = "__multi3";
63 Names[RTLIB::MULO_I32] = "__mulosi4";
64 Names[RTLIB::MULO_I64] = "__mulodi4";
65 Names[RTLIB::MULO_I128] = "__muloti4";
66 Names[RTLIB::SDIV_I8] = "__divqi3";
67 Names[RTLIB::SDIV_I16] = "__divhi3";
68 Names[RTLIB::SDIV_I32] = "__divsi3";
69 Names[RTLIB::SDIV_I64] = "__divdi3";
70 Names[RTLIB::SDIV_I128] = "__divti3";
71 Names[RTLIB::UDIV_I8] = "__udivqi3";
72 Names[RTLIB::UDIV_I16] = "__udivhi3";
73 Names[RTLIB::UDIV_I32] = "__udivsi3";
74 Names[RTLIB::UDIV_I64] = "__udivdi3";
75 Names[RTLIB::UDIV_I128] = "__udivti3";
76 Names[RTLIB::SREM_I8] = "__modqi3";
77 Names[RTLIB::SREM_I16] = "__modhi3";
78 Names[RTLIB::SREM_I32] = "__modsi3";
79 Names[RTLIB::SREM_I64] = "__moddi3";
80 Names[RTLIB::SREM_I128] = "__modti3";
81 Names[RTLIB::UREM_I8] = "__umodqi3";
82 Names[RTLIB::UREM_I16] = "__umodhi3";
83 Names[RTLIB::UREM_I32] = "__umodsi3";
84 Names[RTLIB::UREM_I64] = "__umoddi3";
85 Names[RTLIB::UREM_I128] = "__umodti3";
87 // These are generally not available.
88 Names[RTLIB::SDIVREM_I8] = 0;
89 Names[RTLIB::SDIVREM_I16] = 0;
90 Names[RTLIB::SDIVREM_I32] = 0;
91 Names[RTLIB::SDIVREM_I64] = 0;
92 Names[RTLIB::SDIVREM_I128] = 0;
93 Names[RTLIB::UDIVREM_I8] = 0;
94 Names[RTLIB::UDIVREM_I16] = 0;
95 Names[RTLIB::UDIVREM_I32] = 0;
96 Names[RTLIB::UDIVREM_I64] = 0;
97 Names[RTLIB::UDIVREM_I128] = 0;
99 Names[RTLIB::NEG_I32] = "__negsi2";
100 Names[RTLIB::NEG_I64] = "__negdi2";
101 Names[RTLIB::ADD_F32] = "__addsf3";
102 Names[RTLIB::ADD_F64] = "__adddf3";
103 Names[RTLIB::ADD_F80] = "__addxf3";
104 Names[RTLIB::ADD_PPCF128] = "__gcc_qadd";
105 Names[RTLIB::SUB_F32] = "__subsf3";
106 Names[RTLIB::SUB_F64] = "__subdf3";
107 Names[RTLIB::SUB_F80] = "__subxf3";
108 Names[RTLIB::SUB_PPCF128] = "__gcc_qsub";
109 Names[RTLIB::MUL_F32] = "__mulsf3";
110 Names[RTLIB::MUL_F64] = "__muldf3";
111 Names[RTLIB::MUL_F80] = "__mulxf3";
112 Names[RTLIB::MUL_PPCF128] = "__gcc_qmul";
113 Names[RTLIB::DIV_F32] = "__divsf3";
114 Names[RTLIB::DIV_F64] = "__divdf3";
115 Names[RTLIB::DIV_F80] = "__divxf3";
116 Names[RTLIB::DIV_PPCF128] = "__gcc_qdiv";
117 Names[RTLIB::REM_F32] = "fmodf";
118 Names[RTLIB::REM_F64] = "fmod";
119 Names[RTLIB::REM_F80] = "fmodl";
120 Names[RTLIB::REM_PPCF128] = "fmodl";
121 Names[RTLIB::FMA_F32] = "fmaf";
122 Names[RTLIB::FMA_F64] = "fma";
123 Names[RTLIB::FMA_F80] = "fmal";
124 Names[RTLIB::FMA_PPCF128] = "fmal";
125 Names[RTLIB::POWI_F32] = "__powisf2";
126 Names[RTLIB::POWI_F64] = "__powidf2";
127 Names[RTLIB::POWI_F80] = "__powixf2";
128 Names[RTLIB::POWI_PPCF128] = "__powitf2";
129 Names[RTLIB::SQRT_F32] = "sqrtf";
130 Names[RTLIB::SQRT_F64] = "sqrt";
131 Names[RTLIB::SQRT_F80] = "sqrtl";
132 Names[RTLIB::SQRT_PPCF128] = "sqrtl";
133 Names[RTLIB::LOG_F32] = "logf";
134 Names[RTLIB::LOG_F64] = "log";
135 Names[RTLIB::LOG_F80] = "logl";
136 Names[RTLIB::LOG_PPCF128] = "logl";
137 Names[RTLIB::LOG2_F32] = "log2f";
138 Names[RTLIB::LOG2_F64] = "log2";
139 Names[RTLIB::LOG2_F80] = "log2l";
140 Names[RTLIB::LOG2_PPCF128] = "log2l";
141 Names[RTLIB::LOG10_F32] = "log10f";
142 Names[RTLIB::LOG10_F64] = "log10";
143 Names[RTLIB::LOG10_F80] = "log10l";
144 Names[RTLIB::LOG10_PPCF128] = "log10l";
145 Names[RTLIB::EXP_F32] = "expf";
146 Names[RTLIB::EXP_F64] = "exp";
147 Names[RTLIB::EXP_F80] = "expl";
148 Names[RTLIB::EXP_PPCF128] = "expl";
149 Names[RTLIB::EXP2_F32] = "exp2f";
150 Names[RTLIB::EXP2_F64] = "exp2";
151 Names[RTLIB::EXP2_F80] = "exp2l";
152 Names[RTLIB::EXP2_PPCF128] = "exp2l";
153 Names[RTLIB::SIN_F32] = "sinf";
154 Names[RTLIB::SIN_F64] = "sin";
155 Names[RTLIB::SIN_F80] = "sinl";
156 Names[RTLIB::SIN_PPCF128] = "sinl";
157 Names[RTLIB::COS_F32] = "cosf";
158 Names[RTLIB::COS_F64] = "cos";
159 Names[RTLIB::COS_F80] = "cosl";
160 Names[RTLIB::COS_PPCF128] = "cosl";
161 Names[RTLIB::POW_F32] = "powf";
162 Names[RTLIB::POW_F64] = "pow";
163 Names[RTLIB::POW_F80] = "powl";
164 Names[RTLIB::POW_PPCF128] = "powl";
165 Names[RTLIB::CEIL_F32] = "ceilf";
166 Names[RTLIB::CEIL_F64] = "ceil";
167 Names[RTLIB::CEIL_F80] = "ceill";
168 Names[RTLIB::CEIL_PPCF128] = "ceill";
169 Names[RTLIB::TRUNC_F32] = "truncf";
170 Names[RTLIB::TRUNC_F64] = "trunc";
171 Names[RTLIB::TRUNC_F80] = "truncl";
172 Names[RTLIB::TRUNC_PPCF128] = "truncl";
173 Names[RTLIB::RINT_F32] = "rintf";
174 Names[RTLIB::RINT_F64] = "rint";
175 Names[RTLIB::RINT_F80] = "rintl";
176 Names[RTLIB::RINT_PPCF128] = "rintl";
177 Names[RTLIB::NEARBYINT_F32] = "nearbyintf";
178 Names[RTLIB::NEARBYINT_F64] = "nearbyint";
179 Names[RTLIB::NEARBYINT_F80] = "nearbyintl";
180 Names[RTLIB::NEARBYINT_PPCF128] = "nearbyintl";
181 Names[RTLIB::FLOOR_F32] = "floorf";
182 Names[RTLIB::FLOOR_F64] = "floor";
183 Names[RTLIB::FLOOR_F80] = "floorl";
184 Names[RTLIB::FLOOR_PPCF128] = "floorl";
185 Names[RTLIB::COPYSIGN_F32] = "copysignf";
186 Names[RTLIB::COPYSIGN_F64] = "copysign";
187 Names[RTLIB::COPYSIGN_F80] = "copysignl";
188 Names[RTLIB::COPYSIGN_PPCF128] = "copysignl";
189 Names[RTLIB::FPEXT_F32_F64] = "__extendsfdf2";
190 Names[RTLIB::FPEXT_F16_F32] = "__gnu_h2f_ieee";
191 Names[RTLIB::FPROUND_F32_F16] = "__gnu_f2h_ieee";
192 Names[RTLIB::FPROUND_F64_F32] = "__truncdfsf2";
193 Names[RTLIB::FPROUND_F80_F32] = "__truncxfsf2";
194 Names[RTLIB::FPROUND_PPCF128_F32] = "__trunctfsf2";
195 Names[RTLIB::FPROUND_F80_F64] = "__truncxfdf2";
196 Names[RTLIB::FPROUND_PPCF128_F64] = "__trunctfdf2";
197 Names[RTLIB::FPTOSINT_F32_I8] = "__fixsfqi";
198 Names[RTLIB::FPTOSINT_F32_I16] = "__fixsfhi";
199 Names[RTLIB::FPTOSINT_F32_I32] = "__fixsfsi";
200 Names[RTLIB::FPTOSINT_F32_I64] = "__fixsfdi";
201 Names[RTLIB::FPTOSINT_F32_I128] = "__fixsfti";
202 Names[RTLIB::FPTOSINT_F64_I8] = "__fixdfqi";
203 Names[RTLIB::FPTOSINT_F64_I16] = "__fixdfhi";
204 Names[RTLIB::FPTOSINT_F64_I32] = "__fixdfsi";
205 Names[RTLIB::FPTOSINT_F64_I64] = "__fixdfdi";
206 Names[RTLIB::FPTOSINT_F64_I128] = "__fixdfti";
207 Names[RTLIB::FPTOSINT_F80_I32] = "__fixxfsi";
208 Names[RTLIB::FPTOSINT_F80_I64] = "__fixxfdi";
209 Names[RTLIB::FPTOSINT_F80_I128] = "__fixxfti";
210 Names[RTLIB::FPTOSINT_PPCF128_I32] = "__fixtfsi";
211 Names[RTLIB::FPTOSINT_PPCF128_I64] = "__fixtfdi";
212 Names[RTLIB::FPTOSINT_PPCF128_I128] = "__fixtfti";
213 Names[RTLIB::FPTOUINT_F32_I8] = "__fixunssfqi";
214 Names[RTLIB::FPTOUINT_F32_I16] = "__fixunssfhi";
215 Names[RTLIB::FPTOUINT_F32_I32] = "__fixunssfsi";
216 Names[RTLIB::FPTOUINT_F32_I64] = "__fixunssfdi";
217 Names[RTLIB::FPTOUINT_F32_I128] = "__fixunssfti";
218 Names[RTLIB::FPTOUINT_F64_I8] = "__fixunsdfqi";
219 Names[RTLIB::FPTOUINT_F64_I16] = "__fixunsdfhi";
220 Names[RTLIB::FPTOUINT_F64_I32] = "__fixunsdfsi";
221 Names[RTLIB::FPTOUINT_F64_I64] = "__fixunsdfdi";
222 Names[RTLIB::FPTOUINT_F64_I128] = "__fixunsdfti";
223 Names[RTLIB::FPTOUINT_F80_I32] = "__fixunsxfsi";
224 Names[RTLIB::FPTOUINT_F80_I64] = "__fixunsxfdi";
225 Names[RTLIB::FPTOUINT_F80_I128] = "__fixunsxfti";
226 Names[RTLIB::FPTOUINT_PPCF128_I32] = "__fixunstfsi";
227 Names[RTLIB::FPTOUINT_PPCF128_I64] = "__fixunstfdi";
228 Names[RTLIB::FPTOUINT_PPCF128_I128] = "__fixunstfti";
229 Names[RTLIB::SINTTOFP_I32_F32] = "__floatsisf";
230 Names[RTLIB::SINTTOFP_I32_F64] = "__floatsidf";
231 Names[RTLIB::SINTTOFP_I32_F80] = "__floatsixf";
232 Names[RTLIB::SINTTOFP_I32_PPCF128] = "__floatsitf";
233 Names[RTLIB::SINTTOFP_I64_F32] = "__floatdisf";
234 Names[RTLIB::SINTTOFP_I64_F64] = "__floatdidf";
235 Names[RTLIB::SINTTOFP_I64_F80] = "__floatdixf";
236 Names[RTLIB::SINTTOFP_I64_PPCF128] = "__floatditf";
237 Names[RTLIB::SINTTOFP_I128_F32] = "__floattisf";
238 Names[RTLIB::SINTTOFP_I128_F64] = "__floattidf";
239 Names[RTLIB::SINTTOFP_I128_F80] = "__floattixf";
240 Names[RTLIB::SINTTOFP_I128_PPCF128] = "__floattitf";
241 Names[RTLIB::UINTTOFP_I32_F32] = "__floatunsisf";
242 Names[RTLIB::UINTTOFP_I32_F64] = "__floatunsidf";
243 Names[RTLIB::UINTTOFP_I32_F80] = "__floatunsixf";
244 Names[RTLIB::UINTTOFP_I32_PPCF128] = "__floatunsitf";
245 Names[RTLIB::UINTTOFP_I64_F32] = "__floatundisf";
246 Names[RTLIB::UINTTOFP_I64_F64] = "__floatundidf";
247 Names[RTLIB::UINTTOFP_I64_F80] = "__floatundixf";
248 Names[RTLIB::UINTTOFP_I64_PPCF128] = "__floatunditf";
249 Names[RTLIB::UINTTOFP_I128_F32] = "__floatuntisf";
250 Names[RTLIB::UINTTOFP_I128_F64] = "__floatuntidf";
251 Names[RTLIB::UINTTOFP_I128_F80] = "__floatuntixf";
252 Names[RTLIB::UINTTOFP_I128_PPCF128] = "__floatuntitf";
253 Names[RTLIB::OEQ_F32] = "__eqsf2";
254 Names[RTLIB::OEQ_F64] = "__eqdf2";
255 Names[RTLIB::UNE_F32] = "__nesf2";
256 Names[RTLIB::UNE_F64] = "__nedf2";
257 Names[RTLIB::OGE_F32] = "__gesf2";
258 Names[RTLIB::OGE_F64] = "__gedf2";
259 Names[RTLIB::OLT_F32] = "__ltsf2";
260 Names[RTLIB::OLT_F64] = "__ltdf2";
261 Names[RTLIB::OLE_F32] = "__lesf2";
262 Names[RTLIB::OLE_F64] = "__ledf2";
263 Names[RTLIB::OGT_F32] = "__gtsf2";
264 Names[RTLIB::OGT_F64] = "__gtdf2";
265 Names[RTLIB::UO_F32] = "__unordsf2";
266 Names[RTLIB::UO_F64] = "__unorddf2";
267 Names[RTLIB::O_F32] = "__unordsf2";
268 Names[RTLIB::O_F64] = "__unorddf2";
269 Names[RTLIB::MEMCPY] = "memcpy";
270 Names[RTLIB::MEMMOVE] = "memmove";
271 Names[RTLIB::MEMSET] = "memset";
272 Names[RTLIB::UNWIND_RESUME] = "_Unwind_Resume";
273 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_1] = "__sync_val_compare_and_swap_1";
274 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_2] = "__sync_val_compare_and_swap_2";
275 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_4] = "__sync_val_compare_and_swap_4";
276 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_8] = "__sync_val_compare_and_swap_8";
277 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_1] = "__sync_lock_test_and_set_1";
278 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_2] = "__sync_lock_test_and_set_2";
279 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_4] = "__sync_lock_test_and_set_4";
280 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_8] = "__sync_lock_test_and_set_8";
281 Names[RTLIB::SYNC_FETCH_AND_ADD_1] = "__sync_fetch_and_add_1";
282 Names[RTLIB::SYNC_FETCH_AND_ADD_2] = "__sync_fetch_and_add_2";
283 Names[RTLIB::SYNC_FETCH_AND_ADD_4] = "__sync_fetch_and_add_4";
284 Names[RTLIB::SYNC_FETCH_AND_ADD_8] = "__sync_fetch_and_add_8";
285 Names[RTLIB::SYNC_FETCH_AND_SUB_1] = "__sync_fetch_and_sub_1";
286 Names[RTLIB::SYNC_FETCH_AND_SUB_2] = "__sync_fetch_and_sub_2";
287 Names[RTLIB::SYNC_FETCH_AND_SUB_4] = "__sync_fetch_and_sub_4";
288 Names[RTLIB::SYNC_FETCH_AND_SUB_8] = "__sync_fetch_and_sub_8";
289 Names[RTLIB::SYNC_FETCH_AND_AND_1] = "__sync_fetch_and_and_1";
290 Names[RTLIB::SYNC_FETCH_AND_AND_2] = "__sync_fetch_and_and_2";
291 Names[RTLIB::SYNC_FETCH_AND_AND_4] = "__sync_fetch_and_and_4";
292 Names[RTLIB::SYNC_FETCH_AND_AND_8] = "__sync_fetch_and_and_8";
293 Names[RTLIB::SYNC_FETCH_AND_OR_1] = "__sync_fetch_and_or_1";
294 Names[RTLIB::SYNC_FETCH_AND_OR_2] = "__sync_fetch_and_or_2";
295 Names[RTLIB::SYNC_FETCH_AND_OR_4] = "__sync_fetch_and_or_4";
296 Names[RTLIB::SYNC_FETCH_AND_OR_8] = "__sync_fetch_and_or_8";
297 Names[RTLIB::SYNC_FETCH_AND_XOR_1] = "__sync_fetch_and_xor_1";
298 Names[RTLIB::SYNC_FETCH_AND_XOR_2] = "__sync_fetch_and_xor_2";
299 Names[RTLIB::SYNC_FETCH_AND_XOR_4] = "__sync_fetch_and_xor_4";
300 Names[RTLIB::SYNC_FETCH_AND_XOR_8] = "__sync_fetch_and_xor_8";
301 Names[RTLIB::SYNC_FETCH_AND_NAND_1] = "__sync_fetch_and_nand_1";
302 Names[RTLIB::SYNC_FETCH_AND_NAND_2] = "__sync_fetch_and_nand_2";
303 Names[RTLIB::SYNC_FETCH_AND_NAND_4] = "__sync_fetch_and_nand_4";
304 Names[RTLIB::SYNC_FETCH_AND_NAND_8] = "__sync_fetch_and_nand_8";
307 /// InitLibcallCallingConvs - Set default libcall CallingConvs.
309 static void InitLibcallCallingConvs(CallingConv::ID *CCs) {
310 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
311 CCs[i] = CallingConv::C;
315 /// getFPEXT - Return the FPEXT_*_* value for the given types, or
316 /// UNKNOWN_LIBCALL if there is none.
317 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) {
318 if (OpVT == MVT::f32) {
319 if (RetVT == MVT::f64)
320 return FPEXT_F32_F64;
323 return UNKNOWN_LIBCALL;
326 /// getFPROUND - Return the FPROUND_*_* value for the given types, or
327 /// UNKNOWN_LIBCALL if there is none.
328 RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) {
329 if (RetVT == MVT::f32) {
330 if (OpVT == MVT::f64)
331 return FPROUND_F64_F32;
332 if (OpVT == MVT::f80)
333 return FPROUND_F80_F32;
334 if (OpVT == MVT::ppcf128)
335 return FPROUND_PPCF128_F32;
336 } else if (RetVT == MVT::f64) {
337 if (OpVT == MVT::f80)
338 return FPROUND_F80_F64;
339 if (OpVT == MVT::ppcf128)
340 return FPROUND_PPCF128_F64;
343 return UNKNOWN_LIBCALL;
346 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or
347 /// UNKNOWN_LIBCALL if there is none.
348 RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) {
349 if (OpVT == MVT::f32) {
350 if (RetVT == MVT::i8)
351 return FPTOSINT_F32_I8;
352 if (RetVT == MVT::i16)
353 return FPTOSINT_F32_I16;
354 if (RetVT == MVT::i32)
355 return FPTOSINT_F32_I32;
356 if (RetVT == MVT::i64)
357 return FPTOSINT_F32_I64;
358 if (RetVT == MVT::i128)
359 return FPTOSINT_F32_I128;
360 } else if (OpVT == MVT::f64) {
361 if (RetVT == MVT::i8)
362 return FPTOSINT_F64_I8;
363 if (RetVT == MVT::i16)
364 return FPTOSINT_F64_I16;
365 if (RetVT == MVT::i32)
366 return FPTOSINT_F64_I32;
367 if (RetVT == MVT::i64)
368 return FPTOSINT_F64_I64;
369 if (RetVT == MVT::i128)
370 return FPTOSINT_F64_I128;
371 } else if (OpVT == MVT::f80) {
372 if (RetVT == MVT::i32)
373 return FPTOSINT_F80_I32;
374 if (RetVT == MVT::i64)
375 return FPTOSINT_F80_I64;
376 if (RetVT == MVT::i128)
377 return FPTOSINT_F80_I128;
378 } else if (OpVT == MVT::ppcf128) {
379 if (RetVT == MVT::i32)
380 return FPTOSINT_PPCF128_I32;
381 if (RetVT == MVT::i64)
382 return FPTOSINT_PPCF128_I64;
383 if (RetVT == MVT::i128)
384 return FPTOSINT_PPCF128_I128;
386 return UNKNOWN_LIBCALL;
389 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or
390 /// UNKNOWN_LIBCALL if there is none.
391 RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) {
392 if (OpVT == MVT::f32) {
393 if (RetVT == MVT::i8)
394 return FPTOUINT_F32_I8;
395 if (RetVT == MVT::i16)
396 return FPTOUINT_F32_I16;
397 if (RetVT == MVT::i32)
398 return FPTOUINT_F32_I32;
399 if (RetVT == MVT::i64)
400 return FPTOUINT_F32_I64;
401 if (RetVT == MVT::i128)
402 return FPTOUINT_F32_I128;
403 } else if (OpVT == MVT::f64) {
404 if (RetVT == MVT::i8)
405 return FPTOUINT_F64_I8;
406 if (RetVT == MVT::i16)
407 return FPTOUINT_F64_I16;
408 if (RetVT == MVT::i32)
409 return FPTOUINT_F64_I32;
410 if (RetVT == MVT::i64)
411 return FPTOUINT_F64_I64;
412 if (RetVT == MVT::i128)
413 return FPTOUINT_F64_I128;
414 } else if (OpVT == MVT::f80) {
415 if (RetVT == MVT::i32)
416 return FPTOUINT_F80_I32;
417 if (RetVT == MVT::i64)
418 return FPTOUINT_F80_I64;
419 if (RetVT == MVT::i128)
420 return FPTOUINT_F80_I128;
421 } else if (OpVT == MVT::ppcf128) {
422 if (RetVT == MVT::i32)
423 return FPTOUINT_PPCF128_I32;
424 if (RetVT == MVT::i64)
425 return FPTOUINT_PPCF128_I64;
426 if (RetVT == MVT::i128)
427 return FPTOUINT_PPCF128_I128;
429 return UNKNOWN_LIBCALL;
432 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or
433 /// UNKNOWN_LIBCALL if there is none.
434 RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) {
435 if (OpVT == MVT::i32) {
436 if (RetVT == MVT::f32)
437 return SINTTOFP_I32_F32;
438 else if (RetVT == MVT::f64)
439 return SINTTOFP_I32_F64;
440 else if (RetVT == MVT::f80)
441 return SINTTOFP_I32_F80;
442 else if (RetVT == MVT::ppcf128)
443 return SINTTOFP_I32_PPCF128;
444 } else if (OpVT == MVT::i64) {
445 if (RetVT == MVT::f32)
446 return SINTTOFP_I64_F32;
447 else if (RetVT == MVT::f64)
448 return SINTTOFP_I64_F64;
449 else if (RetVT == MVT::f80)
450 return SINTTOFP_I64_F80;
451 else if (RetVT == MVT::ppcf128)
452 return SINTTOFP_I64_PPCF128;
453 } else if (OpVT == MVT::i128) {
454 if (RetVT == MVT::f32)
455 return SINTTOFP_I128_F32;
456 else if (RetVT == MVT::f64)
457 return SINTTOFP_I128_F64;
458 else if (RetVT == MVT::f80)
459 return SINTTOFP_I128_F80;
460 else if (RetVT == MVT::ppcf128)
461 return SINTTOFP_I128_PPCF128;
463 return UNKNOWN_LIBCALL;
466 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or
467 /// UNKNOWN_LIBCALL if there is none.
468 RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) {
469 if (OpVT == MVT::i32) {
470 if (RetVT == MVT::f32)
471 return UINTTOFP_I32_F32;
472 else if (RetVT == MVT::f64)
473 return UINTTOFP_I32_F64;
474 else if (RetVT == MVT::f80)
475 return UINTTOFP_I32_F80;
476 else if (RetVT == MVT::ppcf128)
477 return UINTTOFP_I32_PPCF128;
478 } else if (OpVT == MVT::i64) {
479 if (RetVT == MVT::f32)
480 return UINTTOFP_I64_F32;
481 else if (RetVT == MVT::f64)
482 return UINTTOFP_I64_F64;
483 else if (RetVT == MVT::f80)
484 return UINTTOFP_I64_F80;
485 else if (RetVT == MVT::ppcf128)
486 return UINTTOFP_I64_PPCF128;
487 } else if (OpVT == MVT::i128) {
488 if (RetVT == MVT::f32)
489 return UINTTOFP_I128_F32;
490 else if (RetVT == MVT::f64)
491 return UINTTOFP_I128_F64;
492 else if (RetVT == MVT::f80)
493 return UINTTOFP_I128_F80;
494 else if (RetVT == MVT::ppcf128)
495 return UINTTOFP_I128_PPCF128;
497 return UNKNOWN_LIBCALL;
500 /// InitCmpLibcallCCs - Set default comparison libcall CC.
502 static void InitCmpLibcallCCs(ISD::CondCode *CCs) {
503 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL);
504 CCs[RTLIB::OEQ_F32] = ISD::SETEQ;
505 CCs[RTLIB::OEQ_F64] = ISD::SETEQ;
506 CCs[RTLIB::UNE_F32] = ISD::SETNE;
507 CCs[RTLIB::UNE_F64] = ISD::SETNE;
508 CCs[RTLIB::OGE_F32] = ISD::SETGE;
509 CCs[RTLIB::OGE_F64] = ISD::SETGE;
510 CCs[RTLIB::OLT_F32] = ISD::SETLT;
511 CCs[RTLIB::OLT_F64] = ISD::SETLT;
512 CCs[RTLIB::OLE_F32] = ISD::SETLE;
513 CCs[RTLIB::OLE_F64] = ISD::SETLE;
514 CCs[RTLIB::OGT_F32] = ISD::SETGT;
515 CCs[RTLIB::OGT_F64] = ISD::SETGT;
516 CCs[RTLIB::UO_F32] = ISD::SETNE;
517 CCs[RTLIB::UO_F64] = ISD::SETNE;
518 CCs[RTLIB::O_F32] = ISD::SETEQ;
519 CCs[RTLIB::O_F64] = ISD::SETEQ;
522 /// NOTE: The constructor takes ownership of TLOF.
523 TargetLowering::TargetLowering(const TargetMachine &tm,
524 const TargetLoweringObjectFile *tlof)
525 : TM(tm), TD(TM.getTargetData()), TLOF(*tlof),
526 mayPromoteElements(AllowPromoteIntElem) {
527 // All operations default to being supported.
528 memset(OpActions, 0, sizeof(OpActions));
529 memset(LoadExtActions, 0, sizeof(LoadExtActions));
530 memset(TruncStoreActions, 0, sizeof(TruncStoreActions));
531 memset(IndexedModeActions, 0, sizeof(IndexedModeActions));
532 memset(CondCodeActions, 0, sizeof(CondCodeActions));
534 // Set default actions for various operations.
535 for (unsigned VT = 0; VT != (unsigned)MVT::LAST_VALUETYPE; ++VT) {
536 // Default all indexed load / store to expand.
537 for (unsigned IM = (unsigned)ISD::PRE_INC;
538 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
539 setIndexedLoadAction(IM, (MVT::SimpleValueType)VT, Expand);
540 setIndexedStoreAction(IM, (MVT::SimpleValueType)VT, Expand);
543 // These operations default to expand.
544 setOperationAction(ISD::FGETSIGN, (MVT::SimpleValueType)VT, Expand);
545 setOperationAction(ISD::CONCAT_VECTORS, (MVT::SimpleValueType)VT, Expand);
548 // Most targets ignore the @llvm.prefetch intrinsic.
549 setOperationAction(ISD::PREFETCH, MVT::Other, Expand);
551 // ConstantFP nodes default to expand. Targets can either change this to
552 // Legal, in which case all fp constants are legal, or use isFPImmLegal()
553 // to optimize expansions for certain constants.
554 setOperationAction(ISD::ConstantFP, MVT::f16, Expand);
555 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
556 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
557 setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
559 // These library functions default to expand.
560 setOperationAction(ISD::FLOG , MVT::f16, Expand);
561 setOperationAction(ISD::FLOG2, MVT::f16, Expand);
562 setOperationAction(ISD::FLOG10, MVT::f16, Expand);
563 setOperationAction(ISD::FEXP , MVT::f16, Expand);
564 setOperationAction(ISD::FEXP2, MVT::f16, Expand);
565 setOperationAction(ISD::FFLOOR, MVT::f16, Expand);
566 setOperationAction(ISD::FNEARBYINT, MVT::f16, Expand);
567 setOperationAction(ISD::FCEIL, MVT::f16, Expand);
568 setOperationAction(ISD::FRINT, MVT::f16, Expand);
569 setOperationAction(ISD::FTRUNC, MVT::f16, Expand);
570 setOperationAction(ISD::FLOG , MVT::f32, Expand);
571 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
572 setOperationAction(ISD::FLOG10, MVT::f32, Expand);
573 setOperationAction(ISD::FEXP , MVT::f32, Expand);
574 setOperationAction(ISD::FEXP2, MVT::f32, Expand);
575 setOperationAction(ISD::FFLOOR, MVT::f32, Expand);
576 setOperationAction(ISD::FNEARBYINT, MVT::f32, Expand);
577 setOperationAction(ISD::FCEIL, MVT::f32, Expand);
578 setOperationAction(ISD::FRINT, MVT::f32, Expand);
579 setOperationAction(ISD::FTRUNC, MVT::f32, Expand);
580 setOperationAction(ISD::FLOG , MVT::f64, Expand);
581 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
582 setOperationAction(ISD::FLOG10, MVT::f64, Expand);
583 setOperationAction(ISD::FEXP , MVT::f64, Expand);
584 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
585 setOperationAction(ISD::FFLOOR, MVT::f64, Expand);
586 setOperationAction(ISD::FNEARBYINT, MVT::f64, Expand);
587 setOperationAction(ISD::FCEIL, MVT::f64, Expand);
588 setOperationAction(ISD::FRINT, MVT::f64, Expand);
589 setOperationAction(ISD::FTRUNC, MVT::f64, Expand);
591 // Default ISD::TRAP to expand (which turns it into abort).
592 setOperationAction(ISD::TRAP, MVT::Other, Expand);
594 IsLittleEndian = TD->isLittleEndian();
595 PointerTy = MVT::getIntegerVT(8*TD->getPointerSize());
596 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE*sizeof(TargetRegisterClass*));
597 memset(TargetDAGCombineArray, 0, array_lengthof(TargetDAGCombineArray));
598 maxStoresPerMemset = maxStoresPerMemcpy = maxStoresPerMemmove = 8;
599 maxStoresPerMemsetOptSize = maxStoresPerMemcpyOptSize
600 = maxStoresPerMemmoveOptSize = 4;
601 benefitFromCodePlacementOpt = false;
602 UseUnderscoreSetJmp = false;
603 UseUnderscoreLongJmp = false;
604 SelectIsExpensive = false;
605 IntDivIsCheap = false;
606 Pow2DivIsCheap = false;
607 JumpIsExpensive = false;
608 predictableSelectIsExpensive = false;
609 StackPointerRegisterToSaveRestore = 0;
610 ExceptionPointerRegister = 0;
611 ExceptionSelectorRegister = 0;
612 BooleanContents = UndefinedBooleanContent;
613 BooleanVectorContents = UndefinedBooleanContent;
614 SchedPreferenceInfo = Sched::ILP;
616 JumpBufAlignment = 0;
617 MinFunctionAlignment = 0;
618 PrefFunctionAlignment = 0;
619 PrefLoopAlignment = 0;
620 MinStackArgumentAlignment = 1;
621 ShouldFoldAtomicFences = false;
622 InsertFencesForAtomic = false;
624 InitLibcallNames(LibcallRoutineNames);
625 InitCmpLibcallCCs(CmpLibcallCCs);
626 InitLibcallCallingConvs(LibcallCallingConvs);
629 TargetLowering::~TargetLowering() {
633 MVT TargetLowering::getShiftAmountTy(EVT LHSTy) const {
634 return MVT::getIntegerVT(8*TD->getPointerSize());
637 /// canOpTrap - Returns true if the operation can trap for the value type.
638 /// VT must be a legal type.
639 bool TargetLowering::canOpTrap(unsigned Op, EVT VT) const {
640 assert(isTypeLegal(VT));
655 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT,
656 unsigned &NumIntermediates,
658 TargetLowering *TLI) {
659 // Figure out the right, legal destination reg to copy into.
660 unsigned NumElts = VT.getVectorNumElements();
661 MVT EltTy = VT.getVectorElementType();
663 unsigned NumVectorRegs = 1;
665 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
666 // could break down into LHS/RHS like LegalizeDAG does.
667 if (!isPowerOf2_32(NumElts)) {
668 NumVectorRegs = NumElts;
672 // Divide the input until we get to a supported size. This will always
673 // end with a scalar if the target doesn't support vectors.
674 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) {
679 NumIntermediates = NumVectorRegs;
681 MVT NewVT = MVT::getVectorVT(EltTy, NumElts);
682 if (!TLI->isTypeLegal(NewVT))
684 IntermediateVT = NewVT;
686 unsigned NewVTSize = NewVT.getSizeInBits();
688 // Convert sizes such as i33 to i64.
689 if (!isPowerOf2_32(NewVTSize))
690 NewVTSize = NextPowerOf2(NewVTSize);
692 EVT DestVT = TLI->getRegisterType(NewVT);
694 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16.
695 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits());
697 // Otherwise, promotion or legal types use the same number of registers as
698 // the vector decimated to the appropriate level.
699 return NumVectorRegs;
702 /// isLegalRC - Return true if the value types that can be represented by the
703 /// specified register class are all legal.
704 bool TargetLowering::isLegalRC(const TargetRegisterClass *RC) const {
705 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
713 /// findRepresentativeClass - Return the largest legal super-reg register class
714 /// of the register class for the specified type and its associated "cost".
715 std::pair<const TargetRegisterClass*, uint8_t>
716 TargetLowering::findRepresentativeClass(EVT VT) const {
717 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
718 const TargetRegisterClass *RC = RegClassForVT[VT.getSimpleVT().SimpleTy];
720 return std::make_pair(RC, 0);
722 // Compute the set of all super-register classes.
723 BitVector SuperRegRC(TRI->getNumRegClasses());
724 for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI)
725 SuperRegRC.setBitsInMask(RCI.getMask());
727 // Find the first legal register class with the largest spill size.
728 const TargetRegisterClass *BestRC = RC;
729 for (int i = SuperRegRC.find_first(); i >= 0; i = SuperRegRC.find_next(i)) {
730 const TargetRegisterClass *SuperRC = TRI->getRegClass(i);
731 // We want the largest possible spill size.
732 if (SuperRC->getSize() <= BestRC->getSize())
734 if (!isLegalRC(SuperRC))
738 return std::make_pair(BestRC, 1);
741 /// computeRegisterProperties - Once all of the register classes are added,
742 /// this allows us to compute derived properties we expose.
743 void TargetLowering::computeRegisterProperties() {
744 assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE &&
745 "Too many value types for ValueTypeActions to hold!");
747 // Everything defaults to needing one register.
748 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
749 NumRegistersForVT[i] = 1;
750 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i;
752 // ...except isVoid, which doesn't need any registers.
753 NumRegistersForVT[MVT::isVoid] = 0;
755 // Find the largest integer register class.
756 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE;
757 for (; RegClassForVT[LargestIntReg] == 0; --LargestIntReg)
758 assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
760 // Every integer value type larger than this largest register takes twice as
761 // many registers to represent as the previous ValueType.
762 for (unsigned ExpandedReg = LargestIntReg + 1; ; ++ExpandedReg) {
763 EVT ExpandedVT = (MVT::SimpleValueType)ExpandedReg;
764 if (!ExpandedVT.isInteger())
766 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
767 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg;
768 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1);
769 ValueTypeActions.setTypeAction(ExpandedVT, TypeExpandInteger);
772 // Inspect all of the ValueType's smaller than the largest integer
773 // register to see which ones need promotion.
774 unsigned LegalIntReg = LargestIntReg;
775 for (unsigned IntReg = LargestIntReg - 1;
776 IntReg >= (unsigned)MVT::i1; --IntReg) {
777 EVT IVT = (MVT::SimpleValueType)IntReg;
778 if (isTypeLegal(IVT)) {
779 LegalIntReg = IntReg;
781 RegisterTypeForVT[IntReg] = TransformToType[IntReg] =
782 (MVT::SimpleValueType)LegalIntReg;
783 ValueTypeActions.setTypeAction(IVT, TypePromoteInteger);
787 // ppcf128 type is really two f64's.
788 if (!isTypeLegal(MVT::ppcf128)) {
789 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64];
790 RegisterTypeForVT[MVT::ppcf128] = MVT::f64;
791 TransformToType[MVT::ppcf128] = MVT::f64;
792 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat);
795 // Decide how to handle f64. If the target does not have native f64 support,
796 // expand it to i64 and we will be generating soft float library calls.
797 if (!isTypeLegal(MVT::f64)) {
798 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
799 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
800 TransformToType[MVT::f64] = MVT::i64;
801 ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat);
804 // Decide how to handle f32. If the target does not have native support for
805 // f32, promote it to f64 if it is legal. Otherwise, expand it to i32.
806 if (!isTypeLegal(MVT::f32)) {
807 if (isTypeLegal(MVT::f64)) {
808 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::f64];
809 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::f64];
810 TransformToType[MVT::f32] = MVT::f64;
811 ValueTypeActions.setTypeAction(MVT::f32, TypePromoteInteger);
813 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
814 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
815 TransformToType[MVT::f32] = MVT::i32;
816 ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat);
820 // Loop over all of the vector value types to see which need transformations.
821 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
822 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
823 MVT VT = (MVT::SimpleValueType)i;
824 if (isTypeLegal(VT)) continue;
826 // Determine if there is a legal wider type. If so, we should promote to
827 // that wider vector type.
828 EVT EltVT = VT.getVectorElementType();
829 unsigned NElts = VT.getVectorNumElements();
831 bool IsLegalWiderType = false;
832 // If we allow the promotion of vector elements using a flag,
833 // then return TypePromoteInteger on vector elements.
834 // First try to promote the elements of integer vectors. If no legal
835 // promotion was found, fallback to the widen-vector method.
836 if (mayPromoteElements)
837 for (unsigned nVT = i+1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
838 EVT SVT = (MVT::SimpleValueType)nVT;
839 // Promote vectors of integers to vectors with the same number
840 // of elements, with a wider element type.
841 if (SVT.getVectorElementType().getSizeInBits() > EltVT.getSizeInBits()
842 && SVT.getVectorNumElements() == NElts &&
843 isTypeLegal(SVT) && SVT.getScalarType().isInteger()) {
844 TransformToType[i] = SVT;
845 RegisterTypeForVT[i] = SVT;
846 NumRegistersForVT[i] = 1;
847 ValueTypeActions.setTypeAction(VT, TypePromoteInteger);
848 IsLegalWiderType = true;
853 if (IsLegalWiderType) continue;
855 // Try to widen the vector.
856 for (unsigned nVT = i+1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
857 EVT SVT = (MVT::SimpleValueType)nVT;
858 if (SVT.getVectorElementType() == EltVT &&
859 SVT.getVectorNumElements() > NElts &&
861 TransformToType[i] = SVT;
862 RegisterTypeForVT[i] = SVT;
863 NumRegistersForVT[i] = 1;
864 ValueTypeActions.setTypeAction(VT, TypeWidenVector);
865 IsLegalWiderType = true;
869 if (IsLegalWiderType) continue;
874 unsigned NumIntermediates;
875 NumRegistersForVT[i] =
876 getVectorTypeBreakdownMVT(VT, IntermediateVT, NumIntermediates,
878 RegisterTypeForVT[i] = RegisterVT;
880 EVT NVT = VT.getPow2VectorType();
882 // Type is already a power of 2. The default action is to split.
883 TransformToType[i] = MVT::Other;
884 unsigned NumElts = VT.getVectorNumElements();
885 ValueTypeActions.setTypeAction(VT,
886 NumElts > 1 ? TypeSplitVector : TypeScalarizeVector);
888 TransformToType[i] = NVT;
889 ValueTypeActions.setTypeAction(VT, TypeWidenVector);
893 // Determine the 'representative' register class for each value type.
894 // An representative register class is the largest (meaning one which is
895 // not a sub-register class / subreg register class) legal register class for
896 // a group of value types. For example, on i386, i8, i16, and i32
897 // representative would be GR32; while on x86_64 it's GR64.
898 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
899 const TargetRegisterClass* RRC;
901 tie(RRC, Cost) = findRepresentativeClass((MVT::SimpleValueType)i);
902 RepRegClassForVT[i] = RRC;
903 RepRegClassCostForVT[i] = Cost;
907 const char *TargetLowering::getTargetNodeName(unsigned Opcode) const {
912 EVT TargetLowering::getSetCCResultType(EVT VT) const {
913 assert(!VT.isVector() && "No default SetCC type for vectors!");
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 // or a promoted vector type that has the same number of elements which
938 // are wider, then we should convert to that legal vector type.
939 // This handles things like <2 x float> -> <4 x float> and
940 // <4 x i1> -> <4 x i32>.
941 LegalizeTypeAction TA = getTypeAction(Context, VT);
942 if (NumElts != 1 && (TA == TypeWidenVector || TA == TypePromoteInteger)) {
943 RegisterVT = getTypeToTransformTo(Context, VT);
944 if (isTypeLegal(RegisterVT)) {
945 IntermediateVT = RegisterVT;
946 NumIntermediates = 1;
951 // Figure out the right, legal destination reg to copy into.
952 EVT EltTy = VT.getVectorElementType();
954 unsigned NumVectorRegs = 1;
956 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
957 // could break down into LHS/RHS like LegalizeDAG does.
958 if (!isPowerOf2_32(NumElts)) {
959 NumVectorRegs = NumElts;
963 // Divide the input until we get to a supported size. This will always
964 // end with a scalar if the target doesn't support vectors.
965 while (NumElts > 1 && !isTypeLegal(
966 EVT::getVectorVT(Context, EltTy, NumElts))) {
971 NumIntermediates = NumVectorRegs;
973 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts);
974 if (!isTypeLegal(NewVT))
976 IntermediateVT = NewVT;
978 EVT DestVT = getRegisterType(Context, NewVT);
980 unsigned NewVTSize = NewVT.getSizeInBits();
982 // Convert sizes such as i33 to i64.
983 if (!isPowerOf2_32(NewVTSize))
984 NewVTSize = NextPowerOf2(NewVTSize);
986 if (DestVT.bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16.
987 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits());
989 // Otherwise, promotion or legal types use the same number of registers as
990 // the vector decimated to the appropriate level.
991 return NumVectorRegs;
994 /// Get the EVTs and ArgFlags collections that represent the legalized return
995 /// type of the given function. This does not require a DAG or a return value,
996 /// and is suitable for use before any DAGs for the function are constructed.
997 /// TODO: Move this out of TargetLowering.cpp.
998 void llvm::GetReturnInfo(Type* ReturnType, Attributes attr,
999 SmallVectorImpl<ISD::OutputArg> &Outs,
1000 const TargetLowering &TLI,
1001 SmallVectorImpl<uint64_t> *Offsets) {
1002 SmallVector<EVT, 4> ValueVTs;
1003 ComputeValueVTs(TLI, ReturnType, ValueVTs);
1004 unsigned NumValues = ValueVTs.size();
1005 if (NumValues == 0) return;
1006 unsigned Offset = 0;
1008 for (unsigned j = 0, f = NumValues; j != f; ++j) {
1009 EVT VT = ValueVTs[j];
1010 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
1012 if (attr & Attribute::SExt)
1013 ExtendKind = ISD::SIGN_EXTEND;
1014 else if (attr & Attribute::ZExt)
1015 ExtendKind = ISD::ZERO_EXTEND;
1017 // FIXME: C calling convention requires the return type to be promoted to
1018 // at least 32-bit. But this is not necessary for non-C calling
1019 // conventions. The frontend should mark functions whose return values
1020 // require promoting with signext or zeroext attributes.
1021 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) {
1022 EVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32);
1023 if (VT.bitsLT(MinVT))
1027 unsigned NumParts = TLI.getNumRegisters(ReturnType->getContext(), VT);
1028 EVT PartVT = TLI.getRegisterType(ReturnType->getContext(), VT);
1029 unsigned PartSize = TLI.getTargetData()->getTypeAllocSize(
1030 PartVT.getTypeForEVT(ReturnType->getContext()));
1032 // 'inreg' on function refers to return value
1033 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1034 if (attr & Attribute::InReg)
1037 // Propagate extension type if any
1038 if (attr & Attribute::SExt)
1040 else if (attr & Attribute::ZExt)
1043 for (unsigned i = 0; i < NumParts; ++i) {
1044 Outs.push_back(ISD::OutputArg(Flags, PartVT, /*isFixed=*/true));
1046 Offsets->push_back(Offset);
1053 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1054 /// function arguments in the caller parameter area. This is the actual
1055 /// alignment, not its logarithm.
1056 unsigned TargetLowering::getByValTypeAlignment(Type *Ty) const {
1057 return TD->getCallFrameTypeAlignment(Ty);
1060 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1061 /// current function. The returned value is a member of the
1062 /// MachineJumpTableInfo::JTEntryKind enum.
1063 unsigned TargetLowering::getJumpTableEncoding() const {
1064 // In non-pic modes, just use the address of a block.
1065 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
1066 return MachineJumpTableInfo::EK_BlockAddress;
1068 // In PIC mode, if the target supports a GPRel32 directive, use it.
1069 if (getTargetMachine().getMCAsmInfo()->getGPRel32Directive() != 0)
1070 return MachineJumpTableInfo::EK_GPRel32BlockAddress;
1072 // Otherwise, use a label difference.
1073 return MachineJumpTableInfo::EK_LabelDifference32;
1076 SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1077 SelectionDAG &DAG) const {
1078 // If our PIC model is GP relative, use the global offset table as the base.
1079 unsigned JTEncoding = getJumpTableEncoding();
1081 if ((JTEncoding == MachineJumpTableInfo::EK_GPRel64BlockAddress) ||
1082 (JTEncoding == MachineJumpTableInfo::EK_GPRel32BlockAddress))
1083 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy());
1088 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1089 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1092 TargetLowering::getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
1093 unsigned JTI,MCContext &Ctx) const{
1094 // The normal PIC reloc base is the label at the start of the jump table.
1095 return MCSymbolRefExpr::Create(MF->getJTISymbol(JTI, Ctx), Ctx);
1099 TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
1100 // Assume that everything is safe in static mode.
1101 if (getTargetMachine().getRelocationModel() == Reloc::Static)
1104 // In dynamic-no-pic mode, assume that known defined values are safe.
1105 if (getTargetMachine().getRelocationModel() == Reloc::DynamicNoPIC &&
1107 !GA->getGlobal()->isDeclaration() &&
1108 !GA->getGlobal()->isWeakForLinker())
1111 // Otherwise assume nothing is safe.
1115 //===----------------------------------------------------------------------===//
1116 // Optimization Methods
1117 //===----------------------------------------------------------------------===//
1119 /// ShrinkDemandedConstant - Check to see if the specified operand of the
1120 /// specified instruction is a constant integer. If so, check to see if there
1121 /// are any bits set in the constant that are not demanded. If so, shrink the
1122 /// constant and return true.
1123 bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDValue Op,
1124 const APInt &Demanded) {
1125 DebugLoc dl = Op.getDebugLoc();
1127 // FIXME: ISD::SELECT, ISD::SELECT_CC
1128 switch (Op.getOpcode()) {
1133 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
1134 if (!C) return false;
1136 if (Op.getOpcode() == ISD::XOR &&
1137 (C->getAPIntValue() | (~Demanded)).isAllOnesValue())
1140 // if we can expand it to have all bits set, do it
1141 if (C->getAPIntValue().intersects(~Demanded)) {
1142 EVT VT = Op.getValueType();
1143 SDValue New = DAG.getNode(Op.getOpcode(), dl, VT, Op.getOperand(0),
1144 DAG.getConstant(Demanded &
1147 return CombineTo(Op, New);
1157 /// ShrinkDemandedOp - Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the
1158 /// casts are free. This uses isZExtFree and ZERO_EXTEND for the widening
1159 /// cast, but it could be generalized for targets with other types of
1160 /// implicit widening casts.
1162 TargetLowering::TargetLoweringOpt::ShrinkDemandedOp(SDValue Op,
1164 const APInt &Demanded,
1166 assert(Op.getNumOperands() == 2 &&
1167 "ShrinkDemandedOp only supports binary operators!");
1168 assert(Op.getNode()->getNumValues() == 1 &&
1169 "ShrinkDemandedOp only supports nodes with one result!");
1171 // Don't do this if the node has another user, which may require the
1173 if (!Op.getNode()->hasOneUse())
1176 // Search for the smallest integer type with free casts to and from
1177 // Op's type. For expedience, just check power-of-2 integer types.
1178 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1179 unsigned SmallVTBits = BitWidth - Demanded.countLeadingZeros();
1180 if (!isPowerOf2_32(SmallVTBits))
1181 SmallVTBits = NextPowerOf2(SmallVTBits);
1182 for (; SmallVTBits < BitWidth; SmallVTBits = NextPowerOf2(SmallVTBits)) {
1183 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), SmallVTBits);
1184 if (TLI.isTruncateFree(Op.getValueType(), SmallVT) &&
1185 TLI.isZExtFree(SmallVT, Op.getValueType())) {
1186 // We found a type with free casts.
1187 SDValue X = DAG.getNode(Op.getOpcode(), dl, SmallVT,
1188 DAG.getNode(ISD::TRUNCATE, dl, SmallVT,
1189 Op.getNode()->getOperand(0)),
1190 DAG.getNode(ISD::TRUNCATE, dl, SmallVT,
1191 Op.getNode()->getOperand(1)));
1192 SDValue Z = DAG.getNode(ISD::ZERO_EXTEND, dl, Op.getValueType(), X);
1193 return CombineTo(Op, Z);
1199 /// SimplifyDemandedBits - Look at Op. At this point, we know that only the
1200 /// DemandedMask bits of the result of Op are ever used downstream. If we can
1201 /// use this information to simplify Op, create a new simplified DAG node and
1202 /// return true, returning the original and new nodes in Old and New. Otherwise,
1203 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
1204 /// the expression (used to simplify the caller). The KnownZero/One bits may
1205 /// only be accurate for those bits in the DemandedMask.
1206 bool TargetLowering::SimplifyDemandedBits(SDValue Op,
1207 const APInt &DemandedMask,
1210 TargetLoweringOpt &TLO,
1211 unsigned Depth) const {
1212 unsigned BitWidth = DemandedMask.getBitWidth();
1213 assert(Op.getValueType().getScalarType().getSizeInBits() == BitWidth &&
1214 "Mask size mismatches value type size!");
1215 APInt NewMask = DemandedMask;
1216 DebugLoc dl = Op.getDebugLoc();
1218 // Don't know anything.
1219 KnownZero = KnownOne = APInt(BitWidth, 0);
1221 // Other users may use these bits.
1222 if (!Op.getNode()->hasOneUse()) {
1224 // If not at the root, Just compute the KnownZero/KnownOne bits to
1225 // simplify things downstream.
1226 TLO.DAG.ComputeMaskedBits(Op, KnownZero, KnownOne, Depth);
1229 // If this is the root being simplified, allow it to have multiple uses,
1230 // just set the NewMask to all bits.
1231 NewMask = APInt::getAllOnesValue(BitWidth);
1232 } else if (DemandedMask == 0) {
1233 // Not demanding any bits from Op.
1234 if (Op.getOpcode() != ISD::UNDEF)
1235 return TLO.CombineTo(Op, TLO.DAG.getUNDEF(Op.getValueType()));
1237 } else if (Depth == 6) { // Limit search depth.
1241 APInt KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut;
1242 switch (Op.getOpcode()) {
1244 // We know all of the bits for a constant!
1245 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue();
1246 KnownZero = ~KnownOne;
1247 return false; // Don't fall through, will infinitely loop.
1249 // If the RHS is a constant, check to see if the LHS would be zero without
1250 // using the bits from the RHS. Below, we use knowledge about the RHS to
1251 // simplify the LHS, here we're using information from the LHS to simplify
1253 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1254 APInt LHSZero, LHSOne;
1255 // Do not increment Depth here; that can cause an infinite loop.
1256 TLO.DAG.ComputeMaskedBits(Op.getOperand(0), LHSZero, LHSOne, Depth);
1257 // If the LHS already has zeros where RHSC does, this and is dead.
1258 if ((LHSZero & NewMask) == (~RHSC->getAPIntValue() & NewMask))
1259 return TLO.CombineTo(Op, Op.getOperand(0));
1260 // If any of the set bits in the RHS are known zero on the LHS, shrink
1262 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & NewMask))
1266 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
1267 KnownOne, TLO, Depth+1))
1269 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1270 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownZero & NewMask,
1271 KnownZero2, KnownOne2, TLO, Depth+1))
1273 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1275 // If all of the demanded bits are known one on one side, return the other.
1276 // These bits cannot contribute to the result of the 'and'.
1277 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
1278 return TLO.CombineTo(Op, Op.getOperand(0));
1279 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
1280 return TLO.CombineTo(Op, Op.getOperand(1));
1281 // If all of the demanded bits in the inputs are known zeros, return zero.
1282 if ((NewMask & (KnownZero|KnownZero2)) == NewMask)
1283 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, Op.getValueType()));
1284 // If the RHS is a constant, see if we can simplify it.
1285 if (TLO.ShrinkDemandedConstant(Op, ~KnownZero2 & NewMask))
1287 // If the operation can be done in a smaller type, do so.
1288 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1291 // Output known-1 bits are only known if set in both the LHS & RHS.
1292 KnownOne &= KnownOne2;
1293 // Output known-0 are known to be clear if zero in either the LHS | RHS.
1294 KnownZero |= KnownZero2;
1297 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
1298 KnownOne, TLO, Depth+1))
1300 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1301 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownOne & NewMask,
1302 KnownZero2, KnownOne2, TLO, Depth+1))
1304 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1306 // If all of the demanded bits are known zero on one side, return the other.
1307 // These bits cannot contribute to the result of the 'or'.
1308 if ((NewMask & ~KnownOne2 & KnownZero) == (~KnownOne2 & NewMask))
1309 return TLO.CombineTo(Op, Op.getOperand(0));
1310 if ((NewMask & ~KnownOne & KnownZero2) == (~KnownOne & NewMask))
1311 return TLO.CombineTo(Op, Op.getOperand(1));
1312 // If all of the potentially set bits on one side are known to be set on
1313 // the other side, just use the 'other' side.
1314 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
1315 return TLO.CombineTo(Op, Op.getOperand(0));
1316 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
1317 return TLO.CombineTo(Op, Op.getOperand(1));
1318 // If the RHS is a constant, see if we can simplify it.
1319 if (TLO.ShrinkDemandedConstant(Op, NewMask))
1321 // If the operation can be done in a smaller type, do so.
1322 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1325 // Output known-0 bits are only known if clear in both the LHS & RHS.
1326 KnownZero &= KnownZero2;
1327 // Output known-1 are known to be set if set in either the LHS | RHS.
1328 KnownOne |= KnownOne2;
1331 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
1332 KnownOne, TLO, Depth+1))
1334 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1335 if (SimplifyDemandedBits(Op.getOperand(0), NewMask, KnownZero2,
1336 KnownOne2, TLO, Depth+1))
1338 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1340 // If all of the demanded bits are known zero on one side, return the other.
1341 // These bits cannot contribute to the result of the 'xor'.
1342 if ((KnownZero & NewMask) == NewMask)
1343 return TLO.CombineTo(Op, Op.getOperand(0));
1344 if ((KnownZero2 & NewMask) == NewMask)
1345 return TLO.CombineTo(Op, Op.getOperand(1));
1346 // If the operation can be done in a smaller type, do so.
1347 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1350 // If all of the unknown bits are known to be zero on one side or the other
1351 // (but not both) turn this into an *inclusive* or.
1352 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1353 if ((NewMask & ~KnownZero & ~KnownZero2) == 0)
1354 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, dl, Op.getValueType(),
1358 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1359 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
1360 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1361 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
1363 // If all of the demanded bits on one side are known, and all of the set
1364 // bits on that side are also known to be set on the other side, turn this
1365 // into an AND, as we know the bits will be cleared.
1366 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1367 // NB: it is okay if more bits are known than are requested
1368 if ((NewMask & (KnownZero|KnownOne)) == NewMask) { // all known on one side
1369 if (KnownOne == KnownOne2) { // set bits are the same on both sides
1370 EVT VT = Op.getValueType();
1371 SDValue ANDC = TLO.DAG.getConstant(~KnownOne & NewMask, VT);
1372 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, dl, VT,
1373 Op.getOperand(0), ANDC));
1377 // If the RHS is a constant, see if we can simplify it.
1378 // for XOR, we prefer to force bits to 1 if they will make a -1.
1379 // if we can't force bits, try to shrink constant
1380 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1381 APInt Expanded = C->getAPIntValue() | (~NewMask);
1382 // if we can expand it to have all bits set, do it
1383 if (Expanded.isAllOnesValue()) {
1384 if (Expanded != C->getAPIntValue()) {
1385 EVT VT = Op.getValueType();
1386 SDValue New = TLO.DAG.getNode(Op.getOpcode(), dl,VT, Op.getOperand(0),
1387 TLO.DAG.getConstant(Expanded, VT));
1388 return TLO.CombineTo(Op, New);
1390 // if it already has all the bits set, nothing to change
1391 // but don't shrink either!
1392 } else if (TLO.ShrinkDemandedConstant(Op, NewMask)) {
1397 KnownZero = KnownZeroOut;
1398 KnownOne = KnownOneOut;
1401 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero,
1402 KnownOne, TLO, Depth+1))
1404 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero2,
1405 KnownOne2, TLO, Depth+1))
1407 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1408 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1410 // If the operands are constants, see if we can simplify them.
1411 if (TLO.ShrinkDemandedConstant(Op, NewMask))
1414 // Only known if known in both the LHS and RHS.
1415 KnownOne &= KnownOne2;
1416 KnownZero &= KnownZero2;
1418 case ISD::SELECT_CC:
1419 if (SimplifyDemandedBits(Op.getOperand(3), NewMask, KnownZero,
1420 KnownOne, TLO, Depth+1))
1422 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero2,
1423 KnownOne2, TLO, Depth+1))
1425 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1426 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1428 // If the operands are constants, see if we can simplify them.
1429 if (TLO.ShrinkDemandedConstant(Op, NewMask))
1432 // Only known if known in both the LHS and RHS.
1433 KnownOne &= KnownOne2;
1434 KnownZero &= KnownZero2;
1437 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1438 unsigned ShAmt = SA->getZExtValue();
1439 SDValue InOp = Op.getOperand(0);
1441 // If the shift count is an invalid immediate, don't do anything.
1442 if (ShAmt >= BitWidth)
1445 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a
1446 // single shift. We can do this if the bottom bits (which are shifted
1447 // out) are never demanded.
1448 if (InOp.getOpcode() == ISD::SRL &&
1449 isa<ConstantSDNode>(InOp.getOperand(1))) {
1450 if (ShAmt && (NewMask & APInt::getLowBitsSet(BitWidth, ShAmt)) == 0) {
1451 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
1452 unsigned Opc = ISD::SHL;
1453 int Diff = ShAmt-C1;
1460 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
1461 EVT VT = Op.getValueType();
1462 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT,
1463 InOp.getOperand(0), NewSA));
1467 if (SimplifyDemandedBits(InOp, NewMask.lshr(ShAmt),
1468 KnownZero, KnownOne, TLO, Depth+1))
1471 // Convert (shl (anyext x, c)) to (anyext (shl x, c)) if the high bits
1472 // are not demanded. This will likely allow the anyext to be folded away.
1473 if (InOp.getNode()->getOpcode() == ISD::ANY_EXTEND) {
1474 SDValue InnerOp = InOp.getNode()->getOperand(0);
1475 EVT InnerVT = InnerOp.getValueType();
1476 unsigned InnerBits = InnerVT.getSizeInBits();
1477 if (ShAmt < InnerBits && NewMask.lshr(InnerBits) == 0 &&
1478 isTypeDesirableForOp(ISD::SHL, InnerVT)) {
1479 EVT ShTy = getShiftAmountTy(InnerVT);
1480 if (!APInt(BitWidth, ShAmt).isIntN(ShTy.getSizeInBits()))
1483 TLO.DAG.getNode(ISD::SHL, dl, InnerVT, InnerOp,
1484 TLO.DAG.getConstant(ShAmt, ShTy));
1487 TLO.DAG.getNode(ISD::ANY_EXTEND, dl, Op.getValueType(),
1492 KnownZero <<= SA->getZExtValue();
1493 KnownOne <<= SA->getZExtValue();
1494 // low bits known zero.
1495 KnownZero |= APInt::getLowBitsSet(BitWidth, SA->getZExtValue());
1499 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1500 EVT VT = Op.getValueType();
1501 unsigned ShAmt = SA->getZExtValue();
1502 unsigned VTSize = VT.getSizeInBits();
1503 SDValue InOp = Op.getOperand(0);
1505 // If the shift count is an invalid immediate, don't do anything.
1506 if (ShAmt >= BitWidth)
1509 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a
1510 // single shift. We can do this if the top bits (which are shifted out)
1511 // are never demanded.
1512 if (InOp.getOpcode() == ISD::SHL &&
1513 isa<ConstantSDNode>(InOp.getOperand(1))) {
1514 if (ShAmt && (NewMask & APInt::getHighBitsSet(VTSize, ShAmt)) == 0) {
1515 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
1516 unsigned Opc = ISD::SRL;
1517 int Diff = ShAmt-C1;
1524 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
1525 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT,
1526 InOp.getOperand(0), NewSA));
1530 // Compute the new bits that are at the top now.
1531 if (SimplifyDemandedBits(InOp, (NewMask << ShAmt),
1532 KnownZero, KnownOne, TLO, Depth+1))
1534 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1535 KnownZero = KnownZero.lshr(ShAmt);
1536 KnownOne = KnownOne.lshr(ShAmt);
1538 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
1539 KnownZero |= HighBits; // High bits known zero.
1543 // If this is an arithmetic shift right and only the low-bit is set, we can
1544 // always convert this into a logical shr, even if the shift amount is
1545 // variable. The low bit of the shift cannot be an input sign bit unless
1546 // the shift amount is >= the size of the datatype, which is undefined.
1548 return TLO.CombineTo(Op,
1549 TLO.DAG.getNode(ISD::SRL, dl, Op.getValueType(),
1550 Op.getOperand(0), Op.getOperand(1)));
1552 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1553 EVT VT = Op.getValueType();
1554 unsigned ShAmt = SA->getZExtValue();
1556 // If the shift count is an invalid immediate, don't do anything.
1557 if (ShAmt >= BitWidth)
1560 APInt InDemandedMask = (NewMask << ShAmt);
1562 // If any of the demanded bits are produced by the sign extension, we also
1563 // demand the input sign bit.
1564 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
1565 if (HighBits.intersects(NewMask))
1566 InDemandedMask |= APInt::getSignBit(VT.getScalarType().getSizeInBits());
1568 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask,
1569 KnownZero, KnownOne, TLO, Depth+1))
1571 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1572 KnownZero = KnownZero.lshr(ShAmt);
1573 KnownOne = KnownOne.lshr(ShAmt);
1575 // Handle the sign bit, adjusted to where it is now in the mask.
1576 APInt SignBit = APInt::getSignBit(BitWidth).lshr(ShAmt);
1578 // If the input sign bit is known to be zero, or if none of the top bits
1579 // are demanded, turn this into an unsigned shift right.
1580 if (KnownZero.intersects(SignBit) || (HighBits & ~NewMask) == HighBits) {
1581 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, VT,
1584 } else if (KnownOne.intersects(SignBit)) { // New bits are known one.
1585 KnownOne |= HighBits;
1589 case ISD::SIGN_EXTEND_INREG: {
1590 EVT ExVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1592 APInt MsbMask = APInt::getHighBitsSet(BitWidth, 1);
1593 // If we only care about the highest bit, don't bother shifting right.
1594 if (MsbMask == DemandedMask) {
1595 unsigned ShAmt = ExVT.getScalarType().getSizeInBits();
1596 SDValue InOp = Op.getOperand(0);
1598 // Compute the correct shift amount type, which must be getShiftAmountTy
1599 // for scalar types after legalization.
1600 EVT ShiftAmtTy = Op.getValueType();
1601 if (TLO.LegalTypes() && !ShiftAmtTy.isVector())
1602 ShiftAmtTy = getShiftAmountTy(ShiftAmtTy);
1604 SDValue ShiftAmt = TLO.DAG.getConstant(BitWidth - ShAmt, ShiftAmtTy);
1605 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl,
1606 Op.getValueType(), InOp, ShiftAmt));
1609 // Sign extension. Compute the demanded bits in the result that are not
1610 // present in the input.
1612 APInt::getHighBitsSet(BitWidth,
1613 BitWidth - ExVT.getScalarType().getSizeInBits());
1615 // If none of the extended bits are demanded, eliminate the sextinreg.
1616 if ((NewBits & NewMask) == 0)
1617 return TLO.CombineTo(Op, Op.getOperand(0));
1620 APInt::getSignBit(ExVT.getScalarType().getSizeInBits()).zext(BitWidth);
1621 APInt InputDemandedBits =
1622 APInt::getLowBitsSet(BitWidth,
1623 ExVT.getScalarType().getSizeInBits()) &
1626 // Since the sign extended bits are demanded, we know that the sign
1628 InputDemandedBits |= InSignBit;
1630 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits,
1631 KnownZero, KnownOne, TLO, Depth+1))
1633 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1635 // If the sign bit of the input is known set or clear, then we know the
1636 // top bits of the result.
1638 // If the input sign bit is known zero, convert this into a zero extension.
1639 if (KnownZero.intersects(InSignBit))
1640 return TLO.CombineTo(Op,
1641 TLO.DAG.getZeroExtendInReg(Op.getOperand(0),dl,ExVT));
1643 if (KnownOne.intersects(InSignBit)) { // Input sign bit known set
1644 KnownOne |= NewBits;
1645 KnownZero &= ~NewBits;
1646 } else { // Input sign bit unknown
1647 KnownZero &= ~NewBits;
1648 KnownOne &= ~NewBits;
1652 case ISD::ZERO_EXTEND: {
1653 unsigned OperandBitWidth =
1654 Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
1655 APInt InMask = NewMask.trunc(OperandBitWidth);
1657 // If none of the top bits are demanded, convert this into an any_extend.
1659 APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask;
1660 if (!NewBits.intersects(NewMask))
1661 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND, dl,
1665 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1666 KnownZero, KnownOne, TLO, Depth+1))
1668 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1669 KnownZero = KnownZero.zext(BitWidth);
1670 KnownOne = KnownOne.zext(BitWidth);
1671 KnownZero |= NewBits;
1674 case ISD::SIGN_EXTEND: {
1675 EVT InVT = Op.getOperand(0).getValueType();
1676 unsigned InBits = InVT.getScalarType().getSizeInBits();
1677 APInt InMask = APInt::getLowBitsSet(BitWidth, InBits);
1678 APInt InSignBit = APInt::getBitsSet(BitWidth, InBits - 1, InBits);
1679 APInt NewBits = ~InMask & NewMask;
1681 // If none of the top bits are demanded, convert this into an any_extend.
1683 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND, dl,
1687 // Since some of the sign extended bits are demanded, we know that the sign
1689 APInt InDemandedBits = InMask & NewMask;
1690 InDemandedBits |= InSignBit;
1691 InDemandedBits = InDemandedBits.trunc(InBits);
1693 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero,
1694 KnownOne, TLO, Depth+1))
1696 KnownZero = KnownZero.zext(BitWidth);
1697 KnownOne = KnownOne.zext(BitWidth);
1699 // If the sign bit is known zero, convert this to a zero extend.
1700 if (KnownZero.intersects(InSignBit))
1701 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND, dl,
1705 // If the sign bit is known one, the top bits match.
1706 if (KnownOne.intersects(InSignBit)) {
1707 KnownOne |= NewBits;
1708 assert((KnownZero & NewBits) == 0);
1709 } else { // Otherwise, top bits aren't known.
1710 assert((KnownOne & NewBits) == 0);
1711 assert((KnownZero & NewBits) == 0);
1715 case ISD::ANY_EXTEND: {
1716 unsigned OperandBitWidth =
1717 Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
1718 APInt InMask = NewMask.trunc(OperandBitWidth);
1719 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
1720 KnownZero, KnownOne, TLO, Depth+1))
1722 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1723 KnownZero = KnownZero.zext(BitWidth);
1724 KnownOne = KnownOne.zext(BitWidth);
1727 case ISD::TRUNCATE: {
1728 // Simplify the input, using demanded bit information, and compute the known
1729 // zero/one bits live out.
1730 unsigned OperandBitWidth =
1731 Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
1732 APInt TruncMask = NewMask.zext(OperandBitWidth);
1733 if (SimplifyDemandedBits(Op.getOperand(0), TruncMask,
1734 KnownZero, KnownOne, TLO, Depth+1))
1736 KnownZero = KnownZero.trunc(BitWidth);
1737 KnownOne = KnownOne.trunc(BitWidth);
1739 // If the input is only used by this truncate, see if we can shrink it based
1740 // on the known demanded bits.
1741 if (Op.getOperand(0).getNode()->hasOneUse()) {
1742 SDValue In = Op.getOperand(0);
1743 switch (In.getOpcode()) {
1746 // Shrink SRL by a constant if none of the high bits shifted in are
1748 if (TLO.LegalTypes() &&
1749 !isTypeDesirableForOp(ISD::SRL, Op.getValueType()))
1750 // Do not turn (vt1 truncate (vt2 srl)) into (vt1 srl) if vt1 is
1753 ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1));
1756 SDValue Shift = In.getOperand(1);
1757 if (TLO.LegalTypes()) {
1758 uint64_t ShVal = ShAmt->getZExtValue();
1760 TLO.DAG.getConstant(ShVal, getShiftAmountTy(Op.getValueType()));
1763 APInt HighBits = APInt::getHighBitsSet(OperandBitWidth,
1764 OperandBitWidth - BitWidth);
1765 HighBits = HighBits.lshr(ShAmt->getZExtValue()).trunc(BitWidth);
1767 if (ShAmt->getZExtValue() < BitWidth && !(HighBits & NewMask)) {
1768 // None of the shifted in bits are needed. Add a truncate of the
1769 // shift input, then shift it.
1770 SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE, dl,
1773 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl,
1782 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1785 case ISD::AssertZext: {
1786 // AssertZext demands all of the high bits, plus any of the low bits
1787 // demanded by its users.
1788 EVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1789 APInt InMask = APInt::getLowBitsSet(BitWidth,
1790 VT.getSizeInBits());
1791 if (SimplifyDemandedBits(Op.getOperand(0), ~InMask | NewMask,
1792 KnownZero, KnownOne, TLO, Depth+1))
1794 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1796 KnownZero |= ~InMask & NewMask;
1800 // If this is an FP->Int bitcast and if the sign bit is the only
1801 // thing demanded, turn this into a FGETSIGN.
1802 if (!TLO.LegalOperations() &&
1803 !Op.getValueType().isVector() &&
1804 !Op.getOperand(0).getValueType().isVector() &&
1805 NewMask == APInt::getSignBit(Op.getValueType().getSizeInBits()) &&
1806 Op.getOperand(0).getValueType().isFloatingPoint()) {
1807 bool OpVTLegal = isOperationLegalOrCustom(ISD::FGETSIGN, Op.getValueType());
1808 bool i32Legal = isOperationLegalOrCustom(ISD::FGETSIGN, MVT::i32);
1809 if ((OpVTLegal || i32Legal) && Op.getValueType().isSimple()) {
1810 EVT Ty = OpVTLegal ? Op.getValueType() : MVT::i32;
1811 // Make a FGETSIGN + SHL to move the sign bit into the appropriate
1812 // place. We expect the SHL to be eliminated by other optimizations.
1813 SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, dl, Ty, Op.getOperand(0));
1814 unsigned OpVTSizeInBits = Op.getValueType().getSizeInBits();
1815 if (!OpVTLegal && OpVTSizeInBits > 32)
1816 Sign = TLO.DAG.getNode(ISD::ZERO_EXTEND, dl, Op.getValueType(), Sign);
1817 unsigned ShVal = Op.getValueType().getSizeInBits()-1;
1818 SDValue ShAmt = TLO.DAG.getConstant(ShVal, Op.getValueType());
1819 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl,
1828 // Add, Sub, and Mul don't demand any bits in positions beyond that
1829 // of the highest bit demanded of them.
1830 APInt LoMask = APInt::getLowBitsSet(BitWidth,
1831 BitWidth - NewMask.countLeadingZeros());
1832 if (SimplifyDemandedBits(Op.getOperand(0), LoMask, KnownZero2,
1833 KnownOne2, TLO, Depth+1))
1835 if (SimplifyDemandedBits(Op.getOperand(1), LoMask, KnownZero2,
1836 KnownOne2, TLO, Depth+1))
1838 // See if the operation should be performed at a smaller bit width.
1839 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1844 // Just use ComputeMaskedBits to compute output bits.
1845 TLO.DAG.ComputeMaskedBits(Op, KnownZero, KnownOne, Depth);
1849 // If we know the value of all of the demanded bits, return this as a
1851 if ((NewMask & (KnownZero|KnownOne)) == NewMask)
1852 return TLO.CombineTo(Op, TLO.DAG.getConstant(KnownOne, Op.getValueType()));
1857 /// computeMaskedBitsForTargetNode - Determine which of the bits specified
1858 /// in Mask are known to be either zero or one and return them in the
1859 /// KnownZero/KnownOne bitsets.
1860 void TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
1863 const SelectionDAG &DAG,
1864 unsigned Depth) const {
1865 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1866 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1867 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1868 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1869 "Should use MaskedValueIsZero if you don't know whether Op"
1870 " is a target node!");
1871 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0);
1874 /// ComputeNumSignBitsForTargetNode - This method can be implemented by
1875 /// targets that want to expose additional information about sign bits to the
1877 unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
1878 unsigned Depth) const {
1879 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1880 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1881 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1882 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1883 "Should use ComputeNumSignBits if you don't know whether Op"
1884 " is a target node!");
1888 /// ValueHasExactlyOneBitSet - Test if the given value is known to have exactly
1889 /// one bit set. This differs from ComputeMaskedBits in that it doesn't need to
1890 /// determine which bit is set.
1892 static bool ValueHasExactlyOneBitSet(SDValue Val, const SelectionDAG &DAG) {
1893 // A left-shift of a constant one will have exactly one bit set, because
1894 // shifting the bit off the end is undefined.
1895 if (Val.getOpcode() == ISD::SHL)
1896 if (ConstantSDNode *C =
1897 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0)))
1898 if (C->getAPIntValue() == 1)
1901 // Similarly, a right-shift of a constant sign-bit will have exactly
1903 if (Val.getOpcode() == ISD::SRL)
1904 if (ConstantSDNode *C =
1905 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0)))
1906 if (C->getAPIntValue().isSignBit())
1909 // More could be done here, though the above checks are enough
1910 // to handle some common cases.
1912 // Fall back to ComputeMaskedBits to catch other known cases.
1913 EVT OpVT = Val.getValueType();
1914 unsigned BitWidth = OpVT.getScalarType().getSizeInBits();
1915 APInt KnownZero, KnownOne;
1916 DAG.ComputeMaskedBits(Val, KnownZero, KnownOne);
1917 return (KnownZero.countPopulation() == BitWidth - 1) &&
1918 (KnownOne.countPopulation() == 1);
1921 /// SimplifySetCC - Try to simplify a setcc built with the specified operands
1922 /// and cc. If it is unable to simplify it, return a null SDValue.
1924 TargetLowering::SimplifySetCC(EVT VT, SDValue N0, SDValue N1,
1925 ISD::CondCode Cond, bool foldBooleans,
1926 DAGCombinerInfo &DCI, DebugLoc dl) const {
1927 SelectionDAG &DAG = DCI.DAG;
1929 // These setcc operations always fold.
1933 case ISD::SETFALSE2: return DAG.getConstant(0, VT);
1935 case ISD::SETTRUE2: return DAG.getConstant(1, VT);
1938 // Ensure that the constant occurs on the RHS, and fold constant
1940 if (isa<ConstantSDNode>(N0.getNode()))
1941 return DAG.getSetCC(dl, VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
1943 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) {
1944 const APInt &C1 = N1C->getAPIntValue();
1946 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an
1947 // equality comparison, then we're just comparing whether X itself is
1949 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) &&
1950 N0.getOperand(0).getOpcode() == ISD::CTLZ &&
1951 N0.getOperand(1).getOpcode() == ISD::Constant) {
1953 = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
1954 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1955 ShAmt == Log2_32(N0.getValueType().getSizeInBits())) {
1956 if ((C1 == 0) == (Cond == ISD::SETEQ)) {
1957 // (srl (ctlz x), 5) == 0 -> X != 0
1958 // (srl (ctlz x), 5) != 1 -> X != 0
1961 // (srl (ctlz x), 5) != 0 -> X == 0
1962 // (srl (ctlz x), 5) == 1 -> X == 0
1965 SDValue Zero = DAG.getConstant(0, N0.getValueType());
1966 return DAG.getSetCC(dl, VT, N0.getOperand(0).getOperand(0),
1972 // Look through truncs that don't change the value of a ctpop.
1973 if (N0.hasOneUse() && N0.getOpcode() == ISD::TRUNCATE)
1974 CTPOP = N0.getOperand(0);
1976 if (CTPOP.hasOneUse() && CTPOP.getOpcode() == ISD::CTPOP &&
1977 (N0 == CTPOP || N0.getValueType().getSizeInBits() >
1978 Log2_32_Ceil(CTPOP.getValueType().getSizeInBits()))) {
1979 EVT CTVT = CTPOP.getValueType();
1980 SDValue CTOp = CTPOP.getOperand(0);
1982 // (ctpop x) u< 2 -> (x & x-1) == 0
1983 // (ctpop x) u> 1 -> (x & x-1) != 0
1984 if ((Cond == ISD::SETULT && C1 == 2) || (Cond == ISD::SETUGT && C1 == 1)){
1985 SDValue Sub = DAG.getNode(ISD::SUB, dl, CTVT, CTOp,
1986 DAG.getConstant(1, CTVT));
1987 SDValue And = DAG.getNode(ISD::AND, dl, CTVT, CTOp, Sub);
1988 ISD::CondCode CC = Cond == ISD::SETULT ? ISD::SETEQ : ISD::SETNE;
1989 return DAG.getSetCC(dl, VT, And, DAG.getConstant(0, CTVT), CC);
1992 // TODO: (ctpop x) == 1 -> x && (x & x-1) == 0 iff ctpop is illegal.
1995 // (zext x) == C --> x == (trunc C)
1996 if (DCI.isBeforeLegalize() && N0->hasOneUse() &&
1997 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1998 unsigned MinBits = N0.getValueSizeInBits();
2000 if (N0->getOpcode() == ISD::ZERO_EXTEND) {
2002 MinBits = N0->getOperand(0).getValueSizeInBits();
2003 PreZExt = N0->getOperand(0);
2004 } else if (N0->getOpcode() == ISD::AND) {
2005 // DAGCombine turns costly ZExts into ANDs
2006 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0->getOperand(1)))
2007 if ((C->getAPIntValue()+1).isPowerOf2()) {
2008 MinBits = C->getAPIntValue().countTrailingOnes();
2009 PreZExt = N0->getOperand(0);
2011 } else if (LoadSDNode *LN0 = dyn_cast<LoadSDNode>(N0)) {
2013 if (LN0->getExtensionType() == ISD::ZEXTLOAD) {
2014 MinBits = LN0->getMemoryVT().getSizeInBits();
2019 // Make sure we're not loosing bits from the constant.
2020 if (MinBits < C1.getBitWidth() && MinBits > C1.getActiveBits()) {
2021 EVT MinVT = EVT::getIntegerVT(*DAG.getContext(), MinBits);
2022 if (isTypeDesirableForOp(ISD::SETCC, MinVT)) {
2023 // Will get folded away.
2024 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, MinVT, PreZExt);
2025 SDValue C = DAG.getConstant(C1.trunc(MinBits), MinVT);
2026 return DAG.getSetCC(dl, VT, Trunc, C, Cond);
2031 // If the LHS is '(and load, const)', the RHS is 0,
2032 // the test is for equality or unsigned, and all 1 bits of the const are
2033 // in the same partial word, see if we can shorten the load.
2034 if (DCI.isBeforeLegalize() &&
2035 N0.getOpcode() == ISD::AND && C1 == 0 &&
2036 N0.getNode()->hasOneUse() &&
2037 isa<LoadSDNode>(N0.getOperand(0)) &&
2038 N0.getOperand(0).getNode()->hasOneUse() &&
2039 isa<ConstantSDNode>(N0.getOperand(1))) {
2040 LoadSDNode *Lod = cast<LoadSDNode>(N0.getOperand(0));
2042 unsigned bestWidth = 0, bestOffset = 0;
2043 if (!Lod->isVolatile() && Lod->isUnindexed()) {
2044 unsigned origWidth = N0.getValueType().getSizeInBits();
2045 unsigned maskWidth = origWidth;
2046 // We can narrow (e.g.) 16-bit extending loads on 32-bit target to
2047 // 8 bits, but have to be careful...
2048 if (Lod->getExtensionType() != ISD::NON_EXTLOAD)
2049 origWidth = Lod->getMemoryVT().getSizeInBits();
2051 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
2052 for (unsigned width = origWidth / 2; width>=8; width /= 2) {
2053 APInt newMask = APInt::getLowBitsSet(maskWidth, width);
2054 for (unsigned offset=0; offset<origWidth/width; offset++) {
2055 if ((newMask & Mask) == Mask) {
2056 if (!TD->isLittleEndian())
2057 bestOffset = (origWidth/width - offset - 1) * (width/8);
2059 bestOffset = (uint64_t)offset * (width/8);
2060 bestMask = Mask.lshr(offset * (width/8) * 8);
2064 newMask = newMask << width;
2069 EVT newVT = EVT::getIntegerVT(*DAG.getContext(), bestWidth);
2070 if (newVT.isRound()) {
2071 EVT PtrType = Lod->getOperand(1).getValueType();
2072 SDValue Ptr = Lod->getBasePtr();
2073 if (bestOffset != 0)
2074 Ptr = DAG.getNode(ISD::ADD, dl, PtrType, Lod->getBasePtr(),
2075 DAG.getConstant(bestOffset, PtrType));
2076 unsigned NewAlign = MinAlign(Lod->getAlignment(), bestOffset);
2077 SDValue NewLoad = DAG.getLoad(newVT, dl, Lod->getChain(), Ptr,
2078 Lod->getPointerInfo().getWithOffset(bestOffset),
2079 false, false, false, NewAlign);
2080 return DAG.getSetCC(dl, VT,
2081 DAG.getNode(ISD::AND, dl, newVT, NewLoad,
2082 DAG.getConstant(bestMask.trunc(bestWidth),
2084 DAG.getConstant(0LL, newVT), Cond);
2089 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
2090 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
2091 unsigned InSize = N0.getOperand(0).getValueType().getSizeInBits();
2093 // If the comparison constant has bits in the upper part, the
2094 // zero-extended value could never match.
2095 if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(),
2096 C1.getBitWidth() - InSize))) {
2100 case ISD::SETEQ: return DAG.getConstant(0, VT);
2103 case ISD::SETNE: return DAG.getConstant(1, VT);
2106 // True if the sign bit of C1 is set.
2107 return DAG.getConstant(C1.isNegative(), VT);
2110 // True if the sign bit of C1 isn't set.
2111 return DAG.getConstant(C1.isNonNegative(), VT);
2117 // Otherwise, we can perform the comparison with the low bits.
2125 EVT newVT = N0.getOperand(0).getValueType();
2126 if (DCI.isBeforeLegalizeOps() ||
2127 (isOperationLegal(ISD::SETCC, newVT) &&
2128 getCondCodeAction(Cond, newVT)==Legal))
2129 return DAG.getSetCC(dl, VT, N0.getOperand(0),
2130 DAG.getConstant(C1.trunc(InSize), newVT),
2135 break; // todo, be more careful with signed comparisons
2137 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
2138 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
2139 EVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
2140 unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits();
2141 EVT ExtDstTy = N0.getValueType();
2142 unsigned ExtDstTyBits = ExtDstTy.getSizeInBits();
2144 // If the constant doesn't fit into the number of bits for the source of
2145 // the sign extension, it is impossible for both sides to be equal.
2146 if (C1.getMinSignedBits() > ExtSrcTyBits)
2147 return DAG.getConstant(Cond == ISD::SETNE, VT);
2150 EVT Op0Ty = N0.getOperand(0).getValueType();
2151 if (Op0Ty == ExtSrcTy) {
2152 ZextOp = N0.getOperand(0);
2154 APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits);
2155 ZextOp = DAG.getNode(ISD::AND, dl, Op0Ty, N0.getOperand(0),
2156 DAG.getConstant(Imm, Op0Ty));
2158 if (!DCI.isCalledByLegalizer())
2159 DCI.AddToWorklist(ZextOp.getNode());
2160 // Otherwise, make this a use of a zext.
2161 return DAG.getSetCC(dl, VT, ZextOp,
2162 DAG.getConstant(C1 & APInt::getLowBitsSet(
2167 } else if ((N1C->isNullValue() || N1C->getAPIntValue() == 1) &&
2168 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
2169 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC
2170 if (N0.getOpcode() == ISD::SETCC &&
2171 isTypeLegal(VT) && VT.bitsLE(N0.getValueType())) {
2172 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getAPIntValue() != 1);
2174 return DAG.getNode(ISD::TRUNCATE, dl, VT, N0);
2175 // Invert the condition.
2176 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
2177 CC = ISD::getSetCCInverse(CC,
2178 N0.getOperand(0).getValueType().isInteger());
2179 return DAG.getSetCC(dl, VT, N0.getOperand(0), N0.getOperand(1), CC);
2182 if ((N0.getOpcode() == ISD::XOR ||
2183 (N0.getOpcode() == ISD::AND &&
2184 N0.getOperand(0).getOpcode() == ISD::XOR &&
2185 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
2186 isa<ConstantSDNode>(N0.getOperand(1)) &&
2187 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue() == 1) {
2188 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We
2189 // can only do this if the top bits are known zero.
2190 unsigned BitWidth = N0.getValueSizeInBits();
2191 if (DAG.MaskedValueIsZero(N0,
2192 APInt::getHighBitsSet(BitWidth,
2194 // Okay, get the un-inverted input value.
2196 if (N0.getOpcode() == ISD::XOR)
2197 Val = N0.getOperand(0);
2199 assert(N0.getOpcode() == ISD::AND &&
2200 N0.getOperand(0).getOpcode() == ISD::XOR);
2201 // ((X^1)&1)^1 -> X & 1
2202 Val = DAG.getNode(ISD::AND, dl, N0.getValueType(),
2203 N0.getOperand(0).getOperand(0),
2207 return DAG.getSetCC(dl, VT, Val, N1,
2208 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
2210 } else if (N1C->getAPIntValue() == 1 &&
2212 getBooleanContents(false) == ZeroOrOneBooleanContent)) {
2214 if (Op0.getOpcode() == ISD::TRUNCATE)
2215 Op0 = Op0.getOperand(0);
2217 if ((Op0.getOpcode() == ISD::XOR) &&
2218 Op0.getOperand(0).getOpcode() == ISD::SETCC &&
2219 Op0.getOperand(1).getOpcode() == ISD::SETCC) {
2220 // (xor (setcc), (setcc)) == / != 1 -> (setcc) != / == (setcc)
2221 Cond = (Cond == ISD::SETEQ) ? ISD::SETNE : ISD::SETEQ;
2222 return DAG.getSetCC(dl, VT, Op0.getOperand(0), Op0.getOperand(1),
2224 } else if (Op0.getOpcode() == ISD::AND &&
2225 isa<ConstantSDNode>(Op0.getOperand(1)) &&
2226 cast<ConstantSDNode>(Op0.getOperand(1))->getAPIntValue() == 1) {
2227 // If this is (X&1) == / != 1, normalize it to (X&1) != / == 0.
2228 if (Op0.getValueType().bitsGT(VT))
2229 Op0 = DAG.getNode(ISD::AND, dl, VT,
2230 DAG.getNode(ISD::TRUNCATE, dl, VT, Op0.getOperand(0)),
2231 DAG.getConstant(1, VT));
2232 else if (Op0.getValueType().bitsLT(VT))
2233 Op0 = DAG.getNode(ISD::AND, dl, VT,
2234 DAG.getNode(ISD::ANY_EXTEND, dl, VT, Op0.getOperand(0)),
2235 DAG.getConstant(1, VT));
2237 return DAG.getSetCC(dl, VT, Op0,
2238 DAG.getConstant(0, Op0.getValueType()),
2239 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
2244 APInt MinVal, MaxVal;
2245 unsigned OperandBitSize = N1C->getValueType(0).getSizeInBits();
2246 if (ISD::isSignedIntSetCC(Cond)) {
2247 MinVal = APInt::getSignedMinValue(OperandBitSize);
2248 MaxVal = APInt::getSignedMaxValue(OperandBitSize);
2250 MinVal = APInt::getMinValue(OperandBitSize);
2251 MaxVal = APInt::getMaxValue(OperandBitSize);
2254 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
2255 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
2256 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true
2257 // X >= C0 --> X > (C0-1)
2258 return DAG.getSetCC(dl, VT, N0,
2259 DAG.getConstant(C1-1, N1.getValueType()),
2260 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
2263 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
2264 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true
2265 // X <= C0 --> X < (C0+1)
2266 return DAG.getSetCC(dl, VT, N0,
2267 DAG.getConstant(C1+1, N1.getValueType()),
2268 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
2271 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
2272 return DAG.getConstant(0, VT); // X < MIN --> false
2273 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal)
2274 return DAG.getConstant(1, VT); // X >= MIN --> true
2275 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal)
2276 return DAG.getConstant(0, VT); // X > MAX --> false
2277 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal)
2278 return DAG.getConstant(1, VT); // X <= MAX --> true
2280 // Canonicalize setgt X, Min --> setne X, Min
2281 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
2282 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
2283 // Canonicalize setlt X, Max --> setne X, Max
2284 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
2285 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
2287 // If we have setult X, 1, turn it into seteq X, 0
2288 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
2289 return DAG.getSetCC(dl, VT, N0,
2290 DAG.getConstant(MinVal, N0.getValueType()),
2292 // If we have setugt X, Max-1, turn it into seteq X, Max
2293 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
2294 return DAG.getSetCC(dl, VT, N0,
2295 DAG.getConstant(MaxVal, N0.getValueType()),
2298 // If we have "setcc X, C0", check to see if we can shrink the immediate
2301 // SETUGT X, SINTMAX -> SETLT X, 0
2302 if (Cond == ISD::SETUGT &&
2303 C1 == APInt::getSignedMaxValue(OperandBitSize))
2304 return DAG.getSetCC(dl, VT, N0,
2305 DAG.getConstant(0, N1.getValueType()),
2308 // SETULT X, SINTMIN -> SETGT X, -1
2309 if (Cond == ISD::SETULT &&
2310 C1 == APInt::getSignedMinValue(OperandBitSize)) {
2311 SDValue ConstMinusOne =
2312 DAG.getConstant(APInt::getAllOnesValue(OperandBitSize),
2314 return DAG.getSetCC(dl, VT, N0, ConstMinusOne, ISD::SETGT);
2317 // Fold bit comparisons when we can.
2318 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
2319 (VT == N0.getValueType() ||
2320 (isTypeLegal(VT) && VT.bitsLE(N0.getValueType()))) &&
2321 N0.getOpcode() == ISD::AND)
2322 if (ConstantSDNode *AndRHS =
2323 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2324 EVT ShiftTy = DCI.isBeforeLegalize() ?
2325 getPointerTy() : getShiftAmountTy(N0.getValueType());
2326 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
2327 // Perform the xform if the AND RHS is a single bit.
2328 if (AndRHS->getAPIntValue().isPowerOf2()) {
2329 return DAG.getNode(ISD::TRUNCATE, dl, VT,
2330 DAG.getNode(ISD::SRL, dl, N0.getValueType(), N0,
2331 DAG.getConstant(AndRHS->getAPIntValue().logBase2(), ShiftTy)));
2333 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getAPIntValue()) {
2334 // (X & 8) == 8 --> (X & 8) >> 3
2335 // Perform the xform if C1 is a single bit.
2336 if (C1.isPowerOf2()) {
2337 return DAG.getNode(ISD::TRUNCATE, dl, VT,
2338 DAG.getNode(ISD::SRL, dl, N0.getValueType(), N0,
2339 DAG.getConstant(C1.logBase2(), ShiftTy)));
2345 if (isa<ConstantFPSDNode>(N0.getNode())) {
2346 // Constant fold or commute setcc.
2347 SDValue O = DAG.FoldSetCC(VT, N0, N1, Cond, dl);
2348 if (O.getNode()) return O;
2349 } else if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1.getNode())) {
2350 // If the RHS of an FP comparison is a constant, simplify it away in
2352 if (CFP->getValueAPF().isNaN()) {
2353 // If an operand is known to be a nan, we can fold it.
2354 switch (ISD::getUnorderedFlavor(Cond)) {
2355 default: llvm_unreachable("Unknown flavor!");
2356 case 0: // Known false.
2357 return DAG.getConstant(0, VT);
2358 case 1: // Known true.
2359 return DAG.getConstant(1, VT);
2360 case 2: // Undefined.
2361 return DAG.getUNDEF(VT);
2365 // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the
2366 // constant if knowing that the operand is non-nan is enough. We prefer to
2367 // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to
2369 if (Cond == ISD::SETO || Cond == ISD::SETUO)
2370 return DAG.getSetCC(dl, VT, N0, N0, Cond);
2372 // If the condition is not legal, see if we can find an equivalent one
2374 if (!isCondCodeLegal(Cond, N0.getValueType())) {
2375 // If the comparison was an awkward floating-point == or != and one of
2376 // the comparison operands is infinity or negative infinity, convert the
2377 // condition to a less-awkward <= or >=.
2378 if (CFP->getValueAPF().isInfinity()) {
2379 if (CFP->getValueAPF().isNegative()) {
2380 if (Cond == ISD::SETOEQ &&
2381 isCondCodeLegal(ISD::SETOLE, N0.getValueType()))
2382 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLE);
2383 if (Cond == ISD::SETUEQ &&
2384 isCondCodeLegal(ISD::SETOLE, N0.getValueType()))
2385 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULE);
2386 if (Cond == ISD::SETUNE &&
2387 isCondCodeLegal(ISD::SETUGT, N0.getValueType()))
2388 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGT);
2389 if (Cond == ISD::SETONE &&
2390 isCondCodeLegal(ISD::SETUGT, N0.getValueType()))
2391 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGT);
2393 if (Cond == ISD::SETOEQ &&
2394 isCondCodeLegal(ISD::SETOGE, N0.getValueType()))
2395 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGE);
2396 if (Cond == ISD::SETUEQ &&
2397 isCondCodeLegal(ISD::SETOGE, N0.getValueType()))
2398 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGE);
2399 if (Cond == ISD::SETUNE &&
2400 isCondCodeLegal(ISD::SETULT, N0.getValueType()))
2401 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULT);
2402 if (Cond == ISD::SETONE &&
2403 isCondCodeLegal(ISD::SETULT, N0.getValueType()))
2404 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLT);
2411 // We can always fold X == X for integer setcc's.
2412 if (N0.getValueType().isInteger()) {
2413 switch (getBooleanContents(N0.getValueType().isVector())) {
2414 case UndefinedBooleanContent:
2415 case ZeroOrOneBooleanContent:
2416 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
2417 case ZeroOrNegativeOneBooleanContent:
2418 return DAG.getConstant(ISD::isTrueWhenEqual(Cond) ? -1 : 0, VT);
2421 unsigned UOF = ISD::getUnorderedFlavor(Cond);
2422 if (UOF == 2) // FP operators that are undefined on NaNs.
2423 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
2424 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
2425 return DAG.getConstant(UOF, VT);
2426 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
2427 // if it is not already.
2428 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
2429 if (NewCond != Cond)
2430 return DAG.getSetCC(dl, VT, N0, N1, NewCond);
2433 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
2434 N0.getValueType().isInteger()) {
2435 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
2436 N0.getOpcode() == ISD::XOR) {
2437 // Simplify (X+Y) == (X+Z) --> Y == Z
2438 if (N0.getOpcode() == N1.getOpcode()) {
2439 if (N0.getOperand(0) == N1.getOperand(0))
2440 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(1), Cond);
2441 if (N0.getOperand(1) == N1.getOperand(1))
2442 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(0), Cond);
2443 if (DAG.isCommutativeBinOp(N0.getOpcode())) {
2444 // If X op Y == Y op X, try other combinations.
2445 if (N0.getOperand(0) == N1.getOperand(1))
2446 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(0),
2448 if (N0.getOperand(1) == N1.getOperand(0))
2449 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(1),
2454 // If RHS is a legal immediate value for a compare instruction, we need
2455 // to be careful about increasing register pressure needlessly.
2456 bool LegalRHSImm = false;
2458 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) {
2459 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2460 // Turn (X+C1) == C2 --> X == C2-C1
2461 if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse()) {
2462 return DAG.getSetCC(dl, VT, N0.getOperand(0),
2463 DAG.getConstant(RHSC->getAPIntValue()-
2464 LHSR->getAPIntValue(),
2465 N0.getValueType()), Cond);
2468 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
2469 if (N0.getOpcode() == ISD::XOR)
2470 // If we know that all of the inverted bits are zero, don't bother
2471 // performing the inversion.
2472 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue()))
2474 DAG.getSetCC(dl, VT, N0.getOperand(0),
2475 DAG.getConstant(LHSR->getAPIntValue() ^
2476 RHSC->getAPIntValue(),
2481 // Turn (C1-X) == C2 --> X == C1-C2
2482 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
2483 if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse()) {
2485 DAG.getSetCC(dl, VT, N0.getOperand(1),
2486 DAG.getConstant(SUBC->getAPIntValue() -
2487 RHSC->getAPIntValue(),
2493 // Could RHSC fold directly into a compare?
2494 if (RHSC->getValueType(0).getSizeInBits() <= 64)
2495 LegalRHSImm = isLegalICmpImmediate(RHSC->getSExtValue());
2498 // Simplify (X+Z) == X --> Z == 0
2499 // Don't do this if X is an immediate that can fold into a cmp
2500 // instruction and X+Z has other uses. It could be an induction variable
2501 // chain, and the transform would increase register pressure.
2502 if (!LegalRHSImm || N0.getNode()->hasOneUse()) {
2503 if (N0.getOperand(0) == N1)
2504 return DAG.getSetCC(dl, VT, N0.getOperand(1),
2505 DAG.getConstant(0, N0.getValueType()), Cond);
2506 if (N0.getOperand(1) == N1) {
2507 if (DAG.isCommutativeBinOp(N0.getOpcode()))
2508 return DAG.getSetCC(dl, VT, N0.getOperand(0),
2509 DAG.getConstant(0, N0.getValueType()), Cond);
2510 else if (N0.getNode()->hasOneUse()) {
2511 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
2512 // (Z-X) == X --> Z == X<<1
2513 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(), N1,
2514 DAG.getConstant(1, getShiftAmountTy(N1.getValueType())));
2515 if (!DCI.isCalledByLegalizer())
2516 DCI.AddToWorklist(SH.getNode());
2517 return DAG.getSetCC(dl, VT, N0.getOperand(0), SH, Cond);
2523 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
2524 N1.getOpcode() == ISD::XOR) {
2525 // Simplify X == (X+Z) --> Z == 0
2526 if (N1.getOperand(0) == N0) {
2527 return DAG.getSetCC(dl, VT, N1.getOperand(1),
2528 DAG.getConstant(0, N1.getValueType()), Cond);
2529 } else if (N1.getOperand(1) == N0) {
2530 if (DAG.isCommutativeBinOp(N1.getOpcode())) {
2531 return DAG.getSetCC(dl, VT, N1.getOperand(0),
2532 DAG.getConstant(0, N1.getValueType()), Cond);
2533 } else if (N1.getNode()->hasOneUse()) {
2534 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
2535 // X == (Z-X) --> X<<1 == Z
2536 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(), N0,
2537 DAG.getConstant(1, getShiftAmountTy(N0.getValueType())));
2538 if (!DCI.isCalledByLegalizer())
2539 DCI.AddToWorklist(SH.getNode());
2540 return DAG.getSetCC(dl, VT, SH, N1.getOperand(0), Cond);
2545 // Simplify x&y == y to x&y != 0 if y has exactly one bit set.
2546 // Note that where y is variable and is known to have at most
2547 // one bit set (for example, if it is z&1) we cannot do this;
2548 // the expressions are not equivalent when y==0.
2549 if (N0.getOpcode() == ISD::AND)
2550 if (N0.getOperand(0) == N1 || N0.getOperand(1) == N1) {
2551 if (ValueHasExactlyOneBitSet(N1, DAG)) {
2552 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true);
2553 SDValue Zero = DAG.getConstant(0, N1.getValueType());
2554 return DAG.getSetCC(dl, VT, N0, Zero, Cond);
2557 if (N1.getOpcode() == ISD::AND)
2558 if (N1.getOperand(0) == N0 || N1.getOperand(1) == N0) {
2559 if (ValueHasExactlyOneBitSet(N0, DAG)) {
2560 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true);
2561 SDValue Zero = DAG.getConstant(0, N0.getValueType());
2562 return DAG.getSetCC(dl, VT, N1, Zero, Cond);
2567 // Fold away ALL boolean setcc's.
2569 if (N0.getValueType() == MVT::i1 && foldBooleans) {
2571 default: llvm_unreachable("Unknown integer setcc!");
2572 case ISD::SETEQ: // X == Y -> ~(X^Y)
2573 Temp = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1);
2574 N0 = DAG.getNOT(dl, Temp, MVT::i1);
2575 if (!DCI.isCalledByLegalizer())
2576 DCI.AddToWorklist(Temp.getNode());
2578 case ISD::SETNE: // X != Y --> (X^Y)
2579 N0 = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1);
2581 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> ~X & Y
2582 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> ~X & Y
2583 Temp = DAG.getNOT(dl, N0, MVT::i1);
2584 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N1, Temp);
2585 if (!DCI.isCalledByLegalizer())
2586 DCI.AddToWorklist(Temp.getNode());
2588 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> ~Y & X
2589 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> ~Y & X
2590 Temp = DAG.getNOT(dl, N1, MVT::i1);
2591 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N0, Temp);
2592 if (!DCI.isCalledByLegalizer())
2593 DCI.AddToWorklist(Temp.getNode());
2595 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> ~X | Y
2596 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> ~X | Y
2597 Temp = DAG.getNOT(dl, N0, MVT::i1);
2598 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N1, Temp);
2599 if (!DCI.isCalledByLegalizer())
2600 DCI.AddToWorklist(Temp.getNode());
2602 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> ~Y | X
2603 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> ~Y | X
2604 Temp = DAG.getNOT(dl, N1, MVT::i1);
2605 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N0, Temp);
2608 if (VT != MVT::i1) {
2609 if (!DCI.isCalledByLegalizer())
2610 DCI.AddToWorklist(N0.getNode());
2611 // FIXME: If running after legalize, we probably can't do this.
2612 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, N0);
2617 // Could not fold it.
2621 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
2622 /// node is a GlobalAddress + offset.
2623 bool TargetLowering::isGAPlusOffset(SDNode *N, const GlobalValue *&GA,
2624 int64_t &Offset) const {
2625 if (isa<GlobalAddressSDNode>(N)) {
2626 GlobalAddressSDNode *GASD = cast<GlobalAddressSDNode>(N);
2627 GA = GASD->getGlobal();
2628 Offset += GASD->getOffset();
2632 if (N->getOpcode() == ISD::ADD) {
2633 SDValue N1 = N->getOperand(0);
2634 SDValue N2 = N->getOperand(1);
2635 if (isGAPlusOffset(N1.getNode(), GA, Offset)) {
2636 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
2638 Offset += V->getSExtValue();
2641 } else if (isGAPlusOffset(N2.getNode(), GA, Offset)) {
2642 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
2644 Offset += V->getSExtValue();
2654 SDValue TargetLowering::
2655 PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
2656 // Default implementation: no optimization.
2660 //===----------------------------------------------------------------------===//
2661 // Inline Assembler Implementation Methods
2662 //===----------------------------------------------------------------------===//
2665 TargetLowering::ConstraintType
2666 TargetLowering::getConstraintType(const std::string &Constraint) const {
2667 if (Constraint.size() == 1) {
2668 switch (Constraint[0]) {
2670 case 'r': return C_RegisterClass;
2672 case 'o': // offsetable
2673 case 'V': // not offsetable
2675 case 'i': // Simple Integer or Relocatable Constant
2676 case 'n': // Simple Integer
2677 case 'E': // Floating Point Constant
2678 case 'F': // Floating Point Constant
2679 case 's': // Relocatable Constant
2680 case 'p': // Address.
2681 case 'X': // Allow ANY value.
2682 case 'I': // Target registers.
2696 if (Constraint.size() > 1 && Constraint[0] == '{' &&
2697 Constraint[Constraint.size()-1] == '}')
2702 /// LowerXConstraint - try to replace an X constraint, which matches anything,
2703 /// with another that has more specific requirements based on the type of the
2704 /// corresponding operand.
2705 const char *TargetLowering::LowerXConstraint(EVT ConstraintVT) const{
2706 if (ConstraintVT.isInteger())
2708 if (ConstraintVT.isFloatingPoint())
2709 return "f"; // works for many targets
2713 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
2714 /// vector. If it is invalid, don't add anything to Ops.
2715 void TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
2716 std::string &Constraint,
2717 std::vector<SDValue> &Ops,
2718 SelectionDAG &DAG) const {
2720 if (Constraint.length() > 1) return;
2722 char ConstraintLetter = Constraint[0];
2723 switch (ConstraintLetter) {
2725 case 'X': // Allows any operand; labels (basic block) use this.
2726 if (Op.getOpcode() == ISD::BasicBlock) {
2731 case 'i': // Simple Integer or Relocatable Constant
2732 case 'n': // Simple Integer
2733 case 's': { // Relocatable Constant
2734 // These operands are interested in values of the form (GV+C), where C may
2735 // be folded in as an offset of GV, or it may be explicitly added. Also, it
2736 // is possible and fine if either GV or C are missing.
2737 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2738 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
2740 // If we have "(add GV, C)", pull out GV/C
2741 if (Op.getOpcode() == ISD::ADD) {
2742 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
2743 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
2744 if (C == 0 || GA == 0) {
2745 C = dyn_cast<ConstantSDNode>(Op.getOperand(0));
2746 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1));
2748 if (C == 0 || GA == 0)
2752 // If we find a valid operand, map to the TargetXXX version so that the
2753 // value itself doesn't get selected.
2754 if (GA) { // Either &GV or &GV+C
2755 if (ConstraintLetter != 'n') {
2756 int64_t Offs = GA->getOffset();
2757 if (C) Offs += C->getZExtValue();
2758 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(),
2759 C ? C->getDebugLoc() : DebugLoc(),
2760 Op.getValueType(), Offs));
2764 if (C) { // just C, no GV.
2765 // Simple constants are not allowed for 's'.
2766 if (ConstraintLetter != 's') {
2767 // gcc prints these as sign extended. Sign extend value to 64 bits
2768 // now; without this it would get ZExt'd later in
2769 // ScheduleDAGSDNodes::EmitNode, which is very generic.
2770 Ops.push_back(DAG.getTargetConstant(C->getAPIntValue().getSExtValue(),
2780 std::pair<unsigned, const TargetRegisterClass*> TargetLowering::
2781 getRegForInlineAsmConstraint(const std::string &Constraint,
2783 if (Constraint[0] != '{')
2784 return std::make_pair(0u, static_cast<TargetRegisterClass*>(0));
2785 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?");
2787 // Remove the braces from around the name.
2788 StringRef RegName(Constraint.data()+1, Constraint.size()-2);
2790 // Figure out which register class contains this reg.
2791 const TargetRegisterInfo *RI = TM.getRegisterInfo();
2792 for (TargetRegisterInfo::regclass_iterator RCI = RI->regclass_begin(),
2793 E = RI->regclass_end(); RCI != E; ++RCI) {
2794 const TargetRegisterClass *RC = *RCI;
2796 // If none of the value types for this register class are valid, we
2797 // can't use it. For example, 64-bit reg classes on 32-bit targets.
2801 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end();
2803 if (RegName.equals_lower(RI->getName(*I)))
2804 return std::make_pair(*I, RC);
2808 return std::make_pair(0u, static_cast<const TargetRegisterClass*>(0));
2811 //===----------------------------------------------------------------------===//
2812 // Constraint Selection.
2814 /// isMatchingInputConstraint - Return true of this is an input operand that is
2815 /// a matching constraint like "4".
2816 bool TargetLowering::AsmOperandInfo::isMatchingInputConstraint() const {
2817 assert(!ConstraintCode.empty() && "No known constraint!");
2818 return isdigit(ConstraintCode[0]);
2821 /// getMatchedOperand - If this is an input matching constraint, this method
2822 /// returns the output operand it matches.
2823 unsigned TargetLowering::AsmOperandInfo::getMatchedOperand() const {
2824 assert(!ConstraintCode.empty() && "No known constraint!");
2825 return atoi(ConstraintCode.c_str());
2829 /// ParseConstraints - Split up the constraint string from the inline
2830 /// assembly value into the specific constraints and their prefixes,
2831 /// and also tie in the associated operand values.
2832 /// If this returns an empty vector, and if the constraint string itself
2833 /// isn't empty, there was an error parsing.
2834 TargetLowering::AsmOperandInfoVector TargetLowering::ParseConstraints(
2835 ImmutableCallSite CS) const {
2836 /// ConstraintOperands - Information about all of the constraints.
2837 AsmOperandInfoVector ConstraintOperands;
2838 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
2839 unsigned maCount = 0; // Largest number of multiple alternative constraints.
2841 // Do a prepass over the constraints, canonicalizing them, and building up the
2842 // ConstraintOperands list.
2843 InlineAsm::ConstraintInfoVector
2844 ConstraintInfos = IA->ParseConstraints();
2846 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
2847 unsigned ResNo = 0; // ResNo - The result number of the next output.
2849 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
2850 ConstraintOperands.push_back(AsmOperandInfo(ConstraintInfos[i]));
2851 AsmOperandInfo &OpInfo = ConstraintOperands.back();
2853 // Update multiple alternative constraint count.
2854 if (OpInfo.multipleAlternatives.size() > maCount)
2855 maCount = OpInfo.multipleAlternatives.size();
2857 OpInfo.ConstraintVT = MVT::Other;
2859 // Compute the value type for each operand.
2860 switch (OpInfo.Type) {
2861 case InlineAsm::isOutput:
2862 // Indirect outputs just consume an argument.
2863 if (OpInfo.isIndirect) {
2864 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
2868 // The return value of the call is this value. As such, there is no
2869 // corresponding argument.
2870 assert(!CS.getType()->isVoidTy() &&
2872 if (StructType *STy = dyn_cast<StructType>(CS.getType())) {
2873 OpInfo.ConstraintVT = getValueType(STy->getElementType(ResNo));
2875 assert(ResNo == 0 && "Asm only has one result!");
2876 OpInfo.ConstraintVT = getValueType(CS.getType());
2880 case InlineAsm::isInput:
2881 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
2883 case InlineAsm::isClobber:
2888 if (OpInfo.CallOperandVal) {
2889 llvm::Type *OpTy = OpInfo.CallOperandVal->getType();
2890 if (OpInfo.isIndirect) {
2891 llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy);
2893 report_fatal_error("Indirect operand for inline asm not a pointer!");
2894 OpTy = PtrTy->getElementType();
2897 // Look for vector wrapped in a struct. e.g. { <16 x i8> }.
2898 if (StructType *STy = dyn_cast<StructType>(OpTy))
2899 if (STy->getNumElements() == 1)
2900 OpTy = STy->getElementType(0);
2902 // If OpTy is not a single value, it may be a struct/union that we
2903 // can tile with integers.
2904 if (!OpTy->isSingleValueType() && OpTy->isSized()) {
2905 unsigned BitSize = TD->getTypeSizeInBits(OpTy);
2914 OpInfo.ConstraintVT =
2915 EVT::getEVT(IntegerType::get(OpTy->getContext(), BitSize), true);
2918 } else if (dyn_cast<PointerType>(OpTy)) {
2919 OpInfo.ConstraintVT = MVT::getIntegerVT(8*TD->getPointerSize());
2921 OpInfo.ConstraintVT = EVT::getEVT(OpTy, true);
2926 // If we have multiple alternative constraints, select the best alternative.
2927 if (ConstraintInfos.size()) {
2929 unsigned bestMAIndex = 0;
2930 int bestWeight = -1;
2931 // weight: -1 = invalid match, and 0 = so-so match to 5 = good match.
2934 // Compute the sums of the weights for each alternative, keeping track
2935 // of the best (highest weight) one so far.
2936 for (maIndex = 0; maIndex < maCount; ++maIndex) {
2938 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size();
2939 cIndex != eIndex; ++cIndex) {
2940 AsmOperandInfo& OpInfo = ConstraintOperands[cIndex];
2941 if (OpInfo.Type == InlineAsm::isClobber)
2944 // If this is an output operand with a matching input operand,
2945 // look up the matching input. If their types mismatch, e.g. one
2946 // is an integer, the other is floating point, or their sizes are
2947 // different, flag it as an maCantMatch.
2948 if (OpInfo.hasMatchingInput()) {
2949 AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
2950 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
2951 if ((OpInfo.ConstraintVT.isInteger() !=
2952 Input.ConstraintVT.isInteger()) ||
2953 (OpInfo.ConstraintVT.getSizeInBits() !=
2954 Input.ConstraintVT.getSizeInBits())) {
2955 weightSum = -1; // Can't match.
2960 weight = getMultipleConstraintMatchWeight(OpInfo, maIndex);
2965 weightSum += weight;
2968 if (weightSum > bestWeight) {
2969 bestWeight = weightSum;
2970 bestMAIndex = maIndex;
2974 // Now select chosen alternative in each constraint.
2975 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size();
2976 cIndex != eIndex; ++cIndex) {
2977 AsmOperandInfo& cInfo = ConstraintOperands[cIndex];
2978 if (cInfo.Type == InlineAsm::isClobber)
2980 cInfo.selectAlternative(bestMAIndex);
2985 // Check and hook up tied operands, choose constraint code to use.
2986 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size();
2987 cIndex != eIndex; ++cIndex) {
2988 AsmOperandInfo& OpInfo = ConstraintOperands[cIndex];
2990 // If this is an output operand with a matching input operand, look up the
2991 // matching input. If their types mismatch, e.g. one is an integer, the
2992 // other is floating point, or their sizes are different, flag it as an
2994 if (OpInfo.hasMatchingInput()) {
2995 AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
2997 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
2998 std::pair<unsigned, const TargetRegisterClass*> MatchRC =
2999 getRegForInlineAsmConstraint(OpInfo.ConstraintCode, OpInfo.ConstraintVT);
3000 std::pair<unsigned, const TargetRegisterClass*> InputRC =
3001 getRegForInlineAsmConstraint(Input.ConstraintCode, Input.ConstraintVT);
3002 if ((OpInfo.ConstraintVT.isInteger() !=
3003 Input.ConstraintVT.isInteger()) ||
3004 (MatchRC.second != InputRC.second)) {
3005 report_fatal_error("Unsupported asm: input constraint"
3006 " with a matching output constraint of"
3007 " incompatible type!");
3014 return ConstraintOperands;
3018 /// getConstraintGenerality - Return an integer indicating how general CT
3020 static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) {
3022 case TargetLowering::C_Other:
3023 case TargetLowering::C_Unknown:
3025 case TargetLowering::C_Register:
3027 case TargetLowering::C_RegisterClass:
3029 case TargetLowering::C_Memory:
3032 llvm_unreachable("Invalid constraint type");
3035 /// Examine constraint type and operand type and determine a weight value.
3036 /// This object must already have been set up with the operand type
3037 /// and the current alternative constraint selected.
3038 TargetLowering::ConstraintWeight
3039 TargetLowering::getMultipleConstraintMatchWeight(
3040 AsmOperandInfo &info, int maIndex) const {
3041 InlineAsm::ConstraintCodeVector *rCodes;
3042 if (maIndex >= (int)info.multipleAlternatives.size())
3043 rCodes = &info.Codes;
3045 rCodes = &info.multipleAlternatives[maIndex].Codes;
3046 ConstraintWeight BestWeight = CW_Invalid;
3048 // Loop over the options, keeping track of the most general one.
3049 for (unsigned i = 0, e = rCodes->size(); i != e; ++i) {
3050 ConstraintWeight weight =
3051 getSingleConstraintMatchWeight(info, (*rCodes)[i].c_str());
3052 if (weight > BestWeight)
3053 BestWeight = weight;
3059 /// Examine constraint type and operand type and determine a weight value.
3060 /// This object must already have been set up with the operand type
3061 /// and the current alternative constraint selected.
3062 TargetLowering::ConstraintWeight
3063 TargetLowering::getSingleConstraintMatchWeight(
3064 AsmOperandInfo &info, const char *constraint) const {
3065 ConstraintWeight weight = CW_Invalid;
3066 Value *CallOperandVal = info.CallOperandVal;
3067 // If we don't have a value, we can't do a match,
3068 // but allow it at the lowest weight.
3069 if (CallOperandVal == NULL)
3071 // Look at the constraint type.
3072 switch (*constraint) {
3073 case 'i': // immediate integer.
3074 case 'n': // immediate integer with a known value.
3075 if (isa<ConstantInt>(CallOperandVal))
3076 weight = CW_Constant;
3078 case 's': // non-explicit intregal immediate.
3079 if (isa<GlobalValue>(CallOperandVal))
3080 weight = CW_Constant;
3082 case 'E': // immediate float if host format.
3083 case 'F': // immediate float.
3084 if (isa<ConstantFP>(CallOperandVal))
3085 weight = CW_Constant;
3087 case '<': // memory operand with autodecrement.
3088 case '>': // memory operand with autoincrement.
3089 case 'm': // memory operand.
3090 case 'o': // offsettable memory operand
3091 case 'V': // non-offsettable memory operand
3094 case 'r': // general register.
3095 case 'g': // general register, memory operand or immediate integer.
3096 // note: Clang converts "g" to "imr".
3097 if (CallOperandVal->getType()->isIntegerTy())
3098 weight = CW_Register;
3100 case 'X': // any operand.
3102 weight = CW_Default;
3108 /// ChooseConstraint - If there are multiple different constraints that we
3109 /// could pick for this operand (e.g. "imr") try to pick the 'best' one.
3110 /// This is somewhat tricky: constraints fall into four classes:
3111 /// Other -> immediates and magic values
3112 /// Register -> one specific register
3113 /// RegisterClass -> a group of regs
3114 /// Memory -> memory
3115 /// Ideally, we would pick the most specific constraint possible: if we have
3116 /// something that fits into a register, we would pick it. The problem here
3117 /// is that if we have something that could either be in a register or in
3118 /// memory that use of the register could cause selection of *other*
3119 /// operands to fail: they might only succeed if we pick memory. Because of
3120 /// this the heuristic we use is:
3122 /// 1) If there is an 'other' constraint, and if the operand is valid for
3123 /// that constraint, use it. This makes us take advantage of 'i'
3124 /// constraints when available.
3125 /// 2) Otherwise, pick the most general constraint present. This prefers
3126 /// 'm' over 'r', for example.
3128 static void ChooseConstraint(TargetLowering::AsmOperandInfo &OpInfo,
3129 const TargetLowering &TLI,
3130 SDValue Op, SelectionDAG *DAG) {
3131 assert(OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options");
3132 unsigned BestIdx = 0;
3133 TargetLowering::ConstraintType BestType = TargetLowering::C_Unknown;
3134 int BestGenerality = -1;
3136 // Loop over the options, keeping track of the most general one.
3137 for (unsigned i = 0, e = OpInfo.Codes.size(); i != e; ++i) {
3138 TargetLowering::ConstraintType CType =
3139 TLI.getConstraintType(OpInfo.Codes[i]);
3141 // If this is an 'other' constraint, see if the operand is valid for it.
3142 // For example, on X86 we might have an 'rI' constraint. If the operand
3143 // is an integer in the range [0..31] we want to use I (saving a load
3144 // of a register), otherwise we must use 'r'.
3145 if (CType == TargetLowering::C_Other && Op.getNode()) {
3146 assert(OpInfo.Codes[i].size() == 1 &&
3147 "Unhandled multi-letter 'other' constraint");
3148 std::vector<SDValue> ResultOps;
3149 TLI.LowerAsmOperandForConstraint(Op, OpInfo.Codes[i],
3151 if (!ResultOps.empty()) {
3158 // Things with matching constraints can only be registers, per gcc
3159 // documentation. This mainly affects "g" constraints.
3160 if (CType == TargetLowering::C_Memory && OpInfo.hasMatchingInput())
3163 // This constraint letter is more general than the previous one, use it.
3164 int Generality = getConstraintGenerality(CType);
3165 if (Generality > BestGenerality) {
3168 BestGenerality = Generality;
3172 OpInfo.ConstraintCode = OpInfo.Codes[BestIdx];
3173 OpInfo.ConstraintType = BestType;
3176 /// ComputeConstraintToUse - Determines the constraint code and constraint
3177 /// type to use for the specific AsmOperandInfo, setting
3178 /// OpInfo.ConstraintCode and OpInfo.ConstraintType.
3179 void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo,
3181 SelectionDAG *DAG) const {
3182 assert(!OpInfo.Codes.empty() && "Must have at least one constraint");
3184 // Single-letter constraints ('r') are very common.
3185 if (OpInfo.Codes.size() == 1) {
3186 OpInfo.ConstraintCode = OpInfo.Codes[0];
3187 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
3189 ChooseConstraint(OpInfo, *this, Op, DAG);
3192 // 'X' matches anything.
3193 if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) {
3194 // Labels and constants are handled elsewhere ('X' is the only thing
3195 // that matches labels). For Functions, the type here is the type of
3196 // the result, which is not what we want to look at; leave them alone.
3197 Value *v = OpInfo.CallOperandVal;
3198 if (isa<BasicBlock>(v) || isa<ConstantInt>(v) || isa<Function>(v)) {
3199 OpInfo.CallOperandVal = v;
3203 // Otherwise, try to resolve it to something we know about by looking at
3204 // the actual operand type.
3205 if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) {
3206 OpInfo.ConstraintCode = Repl;
3207 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
3212 //===----------------------------------------------------------------------===//
3213 // Loop Strength Reduction hooks
3214 //===----------------------------------------------------------------------===//
3216 /// isLegalAddressingMode - Return true if the addressing mode represented
3217 /// by AM is legal for this target, for a load/store of the specified type.
3218 bool TargetLowering::isLegalAddressingMode(const AddrMode &AM,
3220 // The default implementation of this implements a conservative RISCy, r+r and
3223 // Allows a sign-extended 16-bit immediate field.
3224 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
3227 // No global is ever allowed as a base.
3231 // Only support r+r,
3233 case 0: // "r+i" or just "i", depending on HasBaseReg.
3236 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
3238 // Otherwise we have r+r or r+i.
3241 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
3243 // Allow 2*r as r+r.
3250 /// BuildExactDiv - Given an exact SDIV by a constant, create a multiplication
3251 /// with the multiplicative inverse of the constant.
3252 SDValue TargetLowering::BuildExactSDIV(SDValue Op1, SDValue Op2, DebugLoc dl,
3253 SelectionDAG &DAG) const {
3254 ConstantSDNode *C = cast<ConstantSDNode>(Op2);
3255 APInt d = C->getAPIntValue();
3256 assert(d != 0 && "Division by zero!");
3258 // Shift the value upfront if it is even, so the LSB is one.
3259 unsigned ShAmt = d.countTrailingZeros();
3261 // TODO: For UDIV use SRL instead of SRA.
3262 SDValue Amt = DAG.getConstant(ShAmt, getShiftAmountTy(Op1.getValueType()));
3263 Op1 = DAG.getNode(ISD::SRA, dl, Op1.getValueType(), Op1, Amt);
3267 // Calculate the multiplicative inverse, using Newton's method.
3269 while ((t = d*xn) != 1)
3270 xn *= APInt(d.getBitWidth(), 2) - t;
3272 Op2 = DAG.getConstant(xn, Op1.getValueType());
3273 return DAG.getNode(ISD::MUL, dl, Op1.getValueType(), Op1, Op2);
3276 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
3277 /// return a DAG expression to select that will generate the same value by
3278 /// multiplying by a magic number. See:
3279 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
3280 SDValue TargetLowering::
3281 BuildSDIV(SDNode *N, SelectionDAG &DAG, bool IsAfterLegalization,
3282 std::vector<SDNode*>* Created) const {
3283 EVT VT = N->getValueType(0);
3284 DebugLoc dl= N->getDebugLoc();
3286 // Check to see if we can do this.
3287 // FIXME: We should be more aggressive here.
3288 if (!isTypeLegal(VT))
3291 APInt d = cast<ConstantSDNode>(N->getOperand(1))->getAPIntValue();
3292 APInt::ms magics = d.magic();
3294 // Multiply the numerator (operand 0) by the magic value
3295 // FIXME: We should support doing a MUL in a wider type
3297 if (IsAfterLegalization ? isOperationLegal(ISD::MULHS, VT) :
3298 isOperationLegalOrCustom(ISD::MULHS, VT))
3299 Q = DAG.getNode(ISD::MULHS, dl, VT, N->getOperand(0),
3300 DAG.getConstant(magics.m, VT));
3301 else if (IsAfterLegalization ? isOperationLegal(ISD::SMUL_LOHI, VT) :
3302 isOperationLegalOrCustom(ISD::SMUL_LOHI, VT))
3303 Q = SDValue(DAG.getNode(ISD::SMUL_LOHI, dl, DAG.getVTList(VT, VT),
3305 DAG.getConstant(magics.m, VT)).getNode(), 1);
3307 return SDValue(); // No mulhs or equvialent
3308 // If d > 0 and m < 0, add the numerator
3309 if (d.isStrictlyPositive() && magics.m.isNegative()) {
3310 Q = DAG.getNode(ISD::ADD, dl, VT, Q, N->getOperand(0));
3312 Created->push_back(Q.getNode());
3314 // If d < 0 and m > 0, subtract the numerator.
3315 if (d.isNegative() && magics.m.isStrictlyPositive()) {
3316 Q = DAG.getNode(ISD::SUB, dl, VT, Q, N->getOperand(0));
3318 Created->push_back(Q.getNode());
3320 // Shift right algebraic if shift value is nonzero
3322 Q = DAG.getNode(ISD::SRA, dl, VT, Q,
3323 DAG.getConstant(magics.s, getShiftAmountTy(Q.getValueType())));
3325 Created->push_back(Q.getNode());
3327 // Extract the sign bit and add it to the quotient
3329 DAG.getNode(ISD::SRL, dl, VT, Q, DAG.getConstant(VT.getSizeInBits()-1,
3330 getShiftAmountTy(Q.getValueType())));
3332 Created->push_back(T.getNode());
3333 return DAG.getNode(ISD::ADD, dl, VT, Q, T);
3336 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
3337 /// return a DAG expression to select that will generate the same value by
3338 /// multiplying by a magic number. See:
3339 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
3340 SDValue TargetLowering::
3341 BuildUDIV(SDNode *N, SelectionDAG &DAG, bool IsAfterLegalization,
3342 std::vector<SDNode*>* Created) const {
3343 EVT VT = N->getValueType(0);
3344 DebugLoc dl = N->getDebugLoc();
3346 // Check to see if we can do this.
3347 // FIXME: We should be more aggressive here.
3348 if (!isTypeLegal(VT))
3351 // FIXME: We should use a narrower constant when the upper
3352 // bits are known to be zero.
3353 const APInt &N1C = cast<ConstantSDNode>(N->getOperand(1))->getAPIntValue();
3354 APInt::mu magics = N1C.magicu();
3356 SDValue Q = N->getOperand(0);
3358 // If the divisor is even, we can avoid using the expensive fixup by shifting
3359 // the divided value upfront.
3360 if (magics.a != 0 && !N1C[0]) {
3361 unsigned Shift = N1C.countTrailingZeros();
3362 Q = DAG.getNode(ISD::SRL, dl, VT, Q,
3363 DAG.getConstant(Shift, getShiftAmountTy(Q.getValueType())));
3365 Created->push_back(Q.getNode());
3367 // Get magic number for the shifted divisor.
3368 magics = N1C.lshr(Shift).magicu(Shift);
3369 assert(magics.a == 0 && "Should use cheap fixup now");
3372 // Multiply the numerator (operand 0) by the magic value
3373 // FIXME: We should support doing a MUL in a wider type
3374 if (IsAfterLegalization ? isOperationLegal(ISD::MULHU, VT) :
3375 isOperationLegalOrCustom(ISD::MULHU, VT))
3376 Q = DAG.getNode(ISD::MULHU, dl, VT, Q, DAG.getConstant(magics.m, VT));
3377 else if (IsAfterLegalization ? isOperationLegal(ISD::UMUL_LOHI, VT) :
3378 isOperationLegalOrCustom(ISD::UMUL_LOHI, VT))
3379 Q = SDValue(DAG.getNode(ISD::UMUL_LOHI, dl, DAG.getVTList(VT, VT), Q,
3380 DAG.getConstant(magics.m, VT)).getNode(), 1);
3382 return SDValue(); // No mulhu or equvialent
3384 Created->push_back(Q.getNode());
3386 if (magics.a == 0) {
3387 assert(magics.s < N1C.getBitWidth() &&
3388 "We shouldn't generate an undefined shift!");
3389 return DAG.getNode(ISD::SRL, dl, VT, Q,
3390 DAG.getConstant(magics.s, getShiftAmountTy(Q.getValueType())));
3392 SDValue NPQ = DAG.getNode(ISD::SUB, dl, VT, N->getOperand(0), Q);
3394 Created->push_back(NPQ.getNode());
3395 NPQ = DAG.getNode(ISD::SRL, dl, VT, NPQ,
3396 DAG.getConstant(1, getShiftAmountTy(NPQ.getValueType())));
3398 Created->push_back(NPQ.getNode());
3399 NPQ = DAG.getNode(ISD::ADD, dl, VT, NPQ, Q);
3401 Created->push_back(NPQ.getNode());
3402 return DAG.getNode(ISD::SRL, dl, VT, NPQ,
3403 DAG.getConstant(magics.s-1, getShiftAmountTy(NPQ.getValueType())));