1 //===-- TargetLoweringBase.cpp - Implement the TargetLoweringBase 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 TargetLoweringBase class.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Target/TargetLowering.h"
15 #include "llvm/ADT/BitVector.h"
16 #include "llvm/ADT/STLExtras.h"
17 #include "llvm/ADT/Triple.h"
18 #include "llvm/CodeGen/Analysis.h"
19 #include "llvm/CodeGen/MachineFrameInfo.h"
20 #include "llvm/CodeGen/MachineFunction.h"
21 #include "llvm/CodeGen/MachineJumpTableInfo.h"
22 #include "llvm/IR/DataLayout.h"
23 #include "llvm/IR/DerivedTypes.h"
24 #include "llvm/IR/GlobalVariable.h"
25 #include "llvm/MC/MCAsmInfo.h"
26 #include "llvm/MC/MCExpr.h"
27 #include "llvm/Support/CommandLine.h"
28 #include "llvm/Support/ErrorHandling.h"
29 #include "llvm/Support/MathExtras.h"
30 #include "llvm/Target/TargetLoweringObjectFile.h"
31 #include "llvm/Target/TargetMachine.h"
32 #include "llvm/Target/TargetRegisterInfo.h"
36 /// InitLibcallNames - Set default libcall names.
38 static void InitLibcallNames(const char **Names, const TargetMachine &TM) {
39 Names[RTLIB::SHL_I16] = "__ashlhi3";
40 Names[RTLIB::SHL_I32] = "__ashlsi3";
41 Names[RTLIB::SHL_I64] = "__ashldi3";
42 Names[RTLIB::SHL_I128] = "__ashlti3";
43 Names[RTLIB::SRL_I16] = "__lshrhi3";
44 Names[RTLIB::SRL_I32] = "__lshrsi3";
45 Names[RTLIB::SRL_I64] = "__lshrdi3";
46 Names[RTLIB::SRL_I128] = "__lshrti3";
47 Names[RTLIB::SRA_I16] = "__ashrhi3";
48 Names[RTLIB::SRA_I32] = "__ashrsi3";
49 Names[RTLIB::SRA_I64] = "__ashrdi3";
50 Names[RTLIB::SRA_I128] = "__ashrti3";
51 Names[RTLIB::MUL_I8] = "__mulqi3";
52 Names[RTLIB::MUL_I16] = "__mulhi3";
53 Names[RTLIB::MUL_I32] = "__mulsi3";
54 Names[RTLIB::MUL_I64] = "__muldi3";
55 Names[RTLIB::MUL_I128] = "__multi3";
56 Names[RTLIB::MULO_I32] = "__mulosi4";
57 Names[RTLIB::MULO_I64] = "__mulodi4";
58 Names[RTLIB::MULO_I128] = "__muloti4";
59 Names[RTLIB::SDIV_I8] = "__divqi3";
60 Names[RTLIB::SDIV_I16] = "__divhi3";
61 Names[RTLIB::SDIV_I32] = "__divsi3";
62 Names[RTLIB::SDIV_I64] = "__divdi3";
63 Names[RTLIB::SDIV_I128] = "__divti3";
64 Names[RTLIB::UDIV_I8] = "__udivqi3";
65 Names[RTLIB::UDIV_I16] = "__udivhi3";
66 Names[RTLIB::UDIV_I32] = "__udivsi3";
67 Names[RTLIB::UDIV_I64] = "__udivdi3";
68 Names[RTLIB::UDIV_I128] = "__udivti3";
69 Names[RTLIB::SREM_I8] = "__modqi3";
70 Names[RTLIB::SREM_I16] = "__modhi3";
71 Names[RTLIB::SREM_I32] = "__modsi3";
72 Names[RTLIB::SREM_I64] = "__moddi3";
73 Names[RTLIB::SREM_I128] = "__modti3";
74 Names[RTLIB::UREM_I8] = "__umodqi3";
75 Names[RTLIB::UREM_I16] = "__umodhi3";
76 Names[RTLIB::UREM_I32] = "__umodsi3";
77 Names[RTLIB::UREM_I64] = "__umoddi3";
78 Names[RTLIB::UREM_I128] = "__umodti3";
80 // These are generally not available.
81 Names[RTLIB::SDIVREM_I8] = 0;
82 Names[RTLIB::SDIVREM_I16] = 0;
83 Names[RTLIB::SDIVREM_I32] = 0;
84 Names[RTLIB::SDIVREM_I64] = 0;
85 Names[RTLIB::SDIVREM_I128] = 0;
86 Names[RTLIB::UDIVREM_I8] = 0;
87 Names[RTLIB::UDIVREM_I16] = 0;
88 Names[RTLIB::UDIVREM_I32] = 0;
89 Names[RTLIB::UDIVREM_I64] = 0;
90 Names[RTLIB::UDIVREM_I128] = 0;
92 Names[RTLIB::NEG_I32] = "__negsi2";
93 Names[RTLIB::NEG_I64] = "__negdi2";
94 Names[RTLIB::ADD_F32] = "__addsf3";
95 Names[RTLIB::ADD_F64] = "__adddf3";
96 Names[RTLIB::ADD_F80] = "__addxf3";
97 Names[RTLIB::ADD_F128] = "__addtf3";
98 Names[RTLIB::ADD_PPCF128] = "__gcc_qadd";
99 Names[RTLIB::SUB_F32] = "__subsf3";
100 Names[RTLIB::SUB_F64] = "__subdf3";
101 Names[RTLIB::SUB_F80] = "__subxf3";
102 Names[RTLIB::SUB_F128] = "__subtf3";
103 Names[RTLIB::SUB_PPCF128] = "__gcc_qsub";
104 Names[RTLIB::MUL_F32] = "__mulsf3";
105 Names[RTLIB::MUL_F64] = "__muldf3";
106 Names[RTLIB::MUL_F80] = "__mulxf3";
107 Names[RTLIB::MUL_F128] = "__multf3";
108 Names[RTLIB::MUL_PPCF128] = "__gcc_qmul";
109 Names[RTLIB::DIV_F32] = "__divsf3";
110 Names[RTLIB::DIV_F64] = "__divdf3";
111 Names[RTLIB::DIV_F80] = "__divxf3";
112 Names[RTLIB::DIV_F128] = "__divtf3";
113 Names[RTLIB::DIV_PPCF128] = "__gcc_qdiv";
114 Names[RTLIB::REM_F32] = "fmodf";
115 Names[RTLIB::REM_F64] = "fmod";
116 Names[RTLIB::REM_F80] = "fmodl";
117 Names[RTLIB::REM_F128] = "fmodl";
118 Names[RTLIB::REM_PPCF128] = "fmodl";
119 Names[RTLIB::FMA_F32] = "fmaf";
120 Names[RTLIB::FMA_F64] = "fma";
121 Names[RTLIB::FMA_F80] = "fmal";
122 Names[RTLIB::FMA_F128] = "fmal";
123 Names[RTLIB::FMA_PPCF128] = "fmal";
124 Names[RTLIB::POWI_F32] = "__powisf2";
125 Names[RTLIB::POWI_F64] = "__powidf2";
126 Names[RTLIB::POWI_F80] = "__powixf2";
127 Names[RTLIB::POWI_F128] = "__powitf2";
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_F128] = "sqrtl";
133 Names[RTLIB::SQRT_PPCF128] = "sqrtl";
134 Names[RTLIB::LOG_F32] = "logf";
135 Names[RTLIB::LOG_F64] = "log";
136 Names[RTLIB::LOG_F80] = "logl";
137 Names[RTLIB::LOG_F128] = "logl";
138 Names[RTLIB::LOG_PPCF128] = "logl";
139 Names[RTLIB::LOG2_F32] = "log2f";
140 Names[RTLIB::LOG2_F64] = "log2";
141 Names[RTLIB::LOG2_F80] = "log2l";
142 Names[RTLIB::LOG2_F128] = "log2l";
143 Names[RTLIB::LOG2_PPCF128] = "log2l";
144 Names[RTLIB::LOG10_F32] = "log10f";
145 Names[RTLIB::LOG10_F64] = "log10";
146 Names[RTLIB::LOG10_F80] = "log10l";
147 Names[RTLIB::LOG10_F128] = "log10l";
148 Names[RTLIB::LOG10_PPCF128] = "log10l";
149 Names[RTLIB::EXP_F32] = "expf";
150 Names[RTLIB::EXP_F64] = "exp";
151 Names[RTLIB::EXP_F80] = "expl";
152 Names[RTLIB::EXP_F128] = "expl";
153 Names[RTLIB::EXP_PPCF128] = "expl";
154 Names[RTLIB::EXP2_F32] = "exp2f";
155 Names[RTLIB::EXP2_F64] = "exp2";
156 Names[RTLIB::EXP2_F80] = "exp2l";
157 Names[RTLIB::EXP2_F128] = "exp2l";
158 Names[RTLIB::EXP2_PPCF128] = "exp2l";
159 Names[RTLIB::SIN_F32] = "sinf";
160 Names[RTLIB::SIN_F64] = "sin";
161 Names[RTLIB::SIN_F80] = "sinl";
162 Names[RTLIB::SIN_F128] = "sinl";
163 Names[RTLIB::SIN_PPCF128] = "sinl";
164 Names[RTLIB::COS_F32] = "cosf";
165 Names[RTLIB::COS_F64] = "cos";
166 Names[RTLIB::COS_F80] = "cosl";
167 Names[RTLIB::COS_F128] = "cosl";
168 Names[RTLIB::COS_PPCF128] = "cosl";
169 Names[RTLIB::POW_F32] = "powf";
170 Names[RTLIB::POW_F64] = "pow";
171 Names[RTLIB::POW_F80] = "powl";
172 Names[RTLIB::POW_F128] = "powl";
173 Names[RTLIB::POW_PPCF128] = "powl";
174 Names[RTLIB::CEIL_F32] = "ceilf";
175 Names[RTLIB::CEIL_F64] = "ceil";
176 Names[RTLIB::CEIL_F80] = "ceill";
177 Names[RTLIB::CEIL_F128] = "ceill";
178 Names[RTLIB::CEIL_PPCF128] = "ceill";
179 Names[RTLIB::TRUNC_F32] = "truncf";
180 Names[RTLIB::TRUNC_F64] = "trunc";
181 Names[RTLIB::TRUNC_F80] = "truncl";
182 Names[RTLIB::TRUNC_F128] = "truncl";
183 Names[RTLIB::TRUNC_PPCF128] = "truncl";
184 Names[RTLIB::RINT_F32] = "rintf";
185 Names[RTLIB::RINT_F64] = "rint";
186 Names[RTLIB::RINT_F80] = "rintl";
187 Names[RTLIB::RINT_F128] = "rintl";
188 Names[RTLIB::RINT_PPCF128] = "rintl";
189 Names[RTLIB::NEARBYINT_F32] = "nearbyintf";
190 Names[RTLIB::NEARBYINT_F64] = "nearbyint";
191 Names[RTLIB::NEARBYINT_F80] = "nearbyintl";
192 Names[RTLIB::NEARBYINT_F128] = "nearbyintl";
193 Names[RTLIB::NEARBYINT_PPCF128] = "nearbyintl";
194 Names[RTLIB::ROUND_F32] = "roundf";
195 Names[RTLIB::ROUND_F64] = "round";
196 Names[RTLIB::ROUND_F80] = "roundl";
197 Names[RTLIB::ROUND_F128] = "roundl";
198 Names[RTLIB::ROUND_PPCF128] = "roundl";
199 Names[RTLIB::FLOOR_F32] = "floorf";
200 Names[RTLIB::FLOOR_F64] = "floor";
201 Names[RTLIB::FLOOR_F80] = "floorl";
202 Names[RTLIB::FLOOR_F128] = "floorl";
203 Names[RTLIB::FLOOR_PPCF128] = "floorl";
204 Names[RTLIB::COPYSIGN_F32] = "copysignf";
205 Names[RTLIB::COPYSIGN_F64] = "copysign";
206 Names[RTLIB::COPYSIGN_F80] = "copysignl";
207 Names[RTLIB::COPYSIGN_F128] = "copysignl";
208 Names[RTLIB::COPYSIGN_PPCF128] = "copysignl";
209 Names[RTLIB::FPEXT_F64_F128] = "__extenddftf2";
210 Names[RTLIB::FPEXT_F32_F128] = "__extendsftf2";
211 Names[RTLIB::FPEXT_F32_F64] = "__extendsfdf2";
212 Names[RTLIB::FPEXT_F16_F32] = "__gnu_h2f_ieee";
213 Names[RTLIB::FPROUND_F32_F16] = "__gnu_f2h_ieee";
214 Names[RTLIB::FPROUND_F64_F32] = "__truncdfsf2";
215 Names[RTLIB::FPROUND_F80_F32] = "__truncxfsf2";
216 Names[RTLIB::FPROUND_F128_F32] = "__trunctfsf2";
217 Names[RTLIB::FPROUND_PPCF128_F32] = "__trunctfsf2";
218 Names[RTLIB::FPROUND_F80_F64] = "__truncxfdf2";
219 Names[RTLIB::FPROUND_F128_F64] = "__trunctfdf2";
220 Names[RTLIB::FPROUND_PPCF128_F64] = "__trunctfdf2";
221 Names[RTLIB::FPTOSINT_F32_I8] = "__fixsfqi";
222 Names[RTLIB::FPTOSINT_F32_I16] = "__fixsfhi";
223 Names[RTLIB::FPTOSINT_F32_I32] = "__fixsfsi";
224 Names[RTLIB::FPTOSINT_F32_I64] = "__fixsfdi";
225 Names[RTLIB::FPTOSINT_F32_I128] = "__fixsfti";
226 Names[RTLIB::FPTOSINT_F64_I8] = "__fixdfqi";
227 Names[RTLIB::FPTOSINT_F64_I16] = "__fixdfhi";
228 Names[RTLIB::FPTOSINT_F64_I32] = "__fixdfsi";
229 Names[RTLIB::FPTOSINT_F64_I64] = "__fixdfdi";
230 Names[RTLIB::FPTOSINT_F64_I128] = "__fixdfti";
231 Names[RTLIB::FPTOSINT_F80_I32] = "__fixxfsi";
232 Names[RTLIB::FPTOSINT_F80_I64] = "__fixxfdi";
233 Names[RTLIB::FPTOSINT_F80_I128] = "__fixxfti";
234 Names[RTLIB::FPTOSINT_F128_I32] = "__fixtfsi";
235 Names[RTLIB::FPTOSINT_F128_I64] = "__fixtfdi";
236 Names[RTLIB::FPTOSINT_F128_I128] = "__fixtfti";
237 Names[RTLIB::FPTOSINT_PPCF128_I32] = "__fixtfsi";
238 Names[RTLIB::FPTOSINT_PPCF128_I64] = "__fixtfdi";
239 Names[RTLIB::FPTOSINT_PPCF128_I128] = "__fixtfti";
240 Names[RTLIB::FPTOUINT_F32_I8] = "__fixunssfqi";
241 Names[RTLIB::FPTOUINT_F32_I16] = "__fixunssfhi";
242 Names[RTLIB::FPTOUINT_F32_I32] = "__fixunssfsi";
243 Names[RTLIB::FPTOUINT_F32_I64] = "__fixunssfdi";
244 Names[RTLIB::FPTOUINT_F32_I128] = "__fixunssfti";
245 Names[RTLIB::FPTOUINT_F64_I8] = "__fixunsdfqi";
246 Names[RTLIB::FPTOUINT_F64_I16] = "__fixunsdfhi";
247 Names[RTLIB::FPTOUINT_F64_I32] = "__fixunsdfsi";
248 Names[RTLIB::FPTOUINT_F64_I64] = "__fixunsdfdi";
249 Names[RTLIB::FPTOUINT_F64_I128] = "__fixunsdfti";
250 Names[RTLIB::FPTOUINT_F80_I32] = "__fixunsxfsi";
251 Names[RTLIB::FPTOUINT_F80_I64] = "__fixunsxfdi";
252 Names[RTLIB::FPTOUINT_F80_I128] = "__fixunsxfti";
253 Names[RTLIB::FPTOUINT_F128_I32] = "__fixunstfsi";
254 Names[RTLIB::FPTOUINT_F128_I64] = "__fixunstfdi";
255 Names[RTLIB::FPTOUINT_F128_I128] = "__fixunstfti";
256 Names[RTLIB::FPTOUINT_PPCF128_I32] = "__fixunstfsi";
257 Names[RTLIB::FPTOUINT_PPCF128_I64] = "__fixunstfdi";
258 Names[RTLIB::FPTOUINT_PPCF128_I128] = "__fixunstfti";
259 Names[RTLIB::SINTTOFP_I32_F32] = "__floatsisf";
260 Names[RTLIB::SINTTOFP_I32_F64] = "__floatsidf";
261 Names[RTLIB::SINTTOFP_I32_F80] = "__floatsixf";
262 Names[RTLIB::SINTTOFP_I32_F128] = "__floatsitf";
263 Names[RTLIB::SINTTOFP_I32_PPCF128] = "__floatsitf";
264 Names[RTLIB::SINTTOFP_I64_F32] = "__floatdisf";
265 Names[RTLIB::SINTTOFP_I64_F64] = "__floatdidf";
266 Names[RTLIB::SINTTOFP_I64_F80] = "__floatdixf";
267 Names[RTLIB::SINTTOFP_I64_F128] = "__floatditf";
268 Names[RTLIB::SINTTOFP_I64_PPCF128] = "__floatditf";
269 Names[RTLIB::SINTTOFP_I128_F32] = "__floattisf";
270 Names[RTLIB::SINTTOFP_I128_F64] = "__floattidf";
271 Names[RTLIB::SINTTOFP_I128_F80] = "__floattixf";
272 Names[RTLIB::SINTTOFP_I128_F128] = "__floattitf";
273 Names[RTLIB::SINTTOFP_I128_PPCF128] = "__floattitf";
274 Names[RTLIB::UINTTOFP_I32_F32] = "__floatunsisf";
275 Names[RTLIB::UINTTOFP_I32_F64] = "__floatunsidf";
276 Names[RTLIB::UINTTOFP_I32_F80] = "__floatunsixf";
277 Names[RTLIB::UINTTOFP_I32_F128] = "__floatunsitf";
278 Names[RTLIB::UINTTOFP_I32_PPCF128] = "__floatunsitf";
279 Names[RTLIB::UINTTOFP_I64_F32] = "__floatundisf";
280 Names[RTLIB::UINTTOFP_I64_F64] = "__floatundidf";
281 Names[RTLIB::UINTTOFP_I64_F80] = "__floatundixf";
282 Names[RTLIB::UINTTOFP_I64_F128] = "__floatunditf";
283 Names[RTLIB::UINTTOFP_I64_PPCF128] = "__floatunditf";
284 Names[RTLIB::UINTTOFP_I128_F32] = "__floatuntisf";
285 Names[RTLIB::UINTTOFP_I128_F64] = "__floatuntidf";
286 Names[RTLIB::UINTTOFP_I128_F80] = "__floatuntixf";
287 Names[RTLIB::UINTTOFP_I128_F128] = "__floatuntitf";
288 Names[RTLIB::UINTTOFP_I128_PPCF128] = "__floatuntitf";
289 Names[RTLIB::OEQ_F32] = "__eqsf2";
290 Names[RTLIB::OEQ_F64] = "__eqdf2";
291 Names[RTLIB::OEQ_F128] = "__eqtf2";
292 Names[RTLIB::UNE_F32] = "__nesf2";
293 Names[RTLIB::UNE_F64] = "__nedf2";
294 Names[RTLIB::UNE_F128] = "__netf2";
295 Names[RTLIB::OGE_F32] = "__gesf2";
296 Names[RTLIB::OGE_F64] = "__gedf2";
297 Names[RTLIB::OGE_F128] = "__getf2";
298 Names[RTLIB::OLT_F32] = "__ltsf2";
299 Names[RTLIB::OLT_F64] = "__ltdf2";
300 Names[RTLIB::OLT_F128] = "__lttf2";
301 Names[RTLIB::OLE_F32] = "__lesf2";
302 Names[RTLIB::OLE_F64] = "__ledf2";
303 Names[RTLIB::OLE_F128] = "__letf2";
304 Names[RTLIB::OGT_F32] = "__gtsf2";
305 Names[RTLIB::OGT_F64] = "__gtdf2";
306 Names[RTLIB::OGT_F128] = "__gttf2";
307 Names[RTLIB::UO_F32] = "__unordsf2";
308 Names[RTLIB::UO_F64] = "__unorddf2";
309 Names[RTLIB::UO_F128] = "__unordtf2";
310 Names[RTLIB::O_F32] = "__unordsf2";
311 Names[RTLIB::O_F64] = "__unorddf2";
312 Names[RTLIB::O_F128] = "__unordtf2";
313 Names[RTLIB::MEMCPY] = "memcpy";
314 Names[RTLIB::MEMMOVE] = "memmove";
315 Names[RTLIB::MEMSET] = "memset";
316 Names[RTLIB::UNWIND_RESUME] = "_Unwind_Resume";
317 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_1] = "__sync_val_compare_and_swap_1";
318 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_2] = "__sync_val_compare_and_swap_2";
319 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_4] = "__sync_val_compare_and_swap_4";
320 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_8] = "__sync_val_compare_and_swap_8";
321 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_16] = "__sync_val_compare_and_swap_16";
322 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_1] = "__sync_lock_test_and_set_1";
323 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_2] = "__sync_lock_test_and_set_2";
324 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_4] = "__sync_lock_test_and_set_4";
325 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_8] = "__sync_lock_test_and_set_8";
326 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_16] = "__sync_lock_test_and_set_16";
327 Names[RTLIB::SYNC_FETCH_AND_ADD_1] = "__sync_fetch_and_add_1";
328 Names[RTLIB::SYNC_FETCH_AND_ADD_2] = "__sync_fetch_and_add_2";
329 Names[RTLIB::SYNC_FETCH_AND_ADD_4] = "__sync_fetch_and_add_4";
330 Names[RTLIB::SYNC_FETCH_AND_ADD_8] = "__sync_fetch_and_add_8";
331 Names[RTLIB::SYNC_FETCH_AND_ADD_16] = "__sync_fetch_and_add_16";
332 Names[RTLIB::SYNC_FETCH_AND_SUB_1] = "__sync_fetch_and_sub_1";
333 Names[RTLIB::SYNC_FETCH_AND_SUB_2] = "__sync_fetch_and_sub_2";
334 Names[RTLIB::SYNC_FETCH_AND_SUB_4] = "__sync_fetch_and_sub_4";
335 Names[RTLIB::SYNC_FETCH_AND_SUB_8] = "__sync_fetch_and_sub_8";
336 Names[RTLIB::SYNC_FETCH_AND_SUB_16] = "__sync_fetch_and_sub_16";
337 Names[RTLIB::SYNC_FETCH_AND_AND_1] = "__sync_fetch_and_and_1";
338 Names[RTLIB::SYNC_FETCH_AND_AND_2] = "__sync_fetch_and_and_2";
339 Names[RTLIB::SYNC_FETCH_AND_AND_4] = "__sync_fetch_and_and_4";
340 Names[RTLIB::SYNC_FETCH_AND_AND_8] = "__sync_fetch_and_and_8";
341 Names[RTLIB::SYNC_FETCH_AND_AND_16] = "__sync_fetch_and_and_16";
342 Names[RTLIB::SYNC_FETCH_AND_OR_1] = "__sync_fetch_and_or_1";
343 Names[RTLIB::SYNC_FETCH_AND_OR_2] = "__sync_fetch_and_or_2";
344 Names[RTLIB::SYNC_FETCH_AND_OR_4] = "__sync_fetch_and_or_4";
345 Names[RTLIB::SYNC_FETCH_AND_OR_8] = "__sync_fetch_and_or_8";
346 Names[RTLIB::SYNC_FETCH_AND_OR_16] = "__sync_fetch_and_or_16";
347 Names[RTLIB::SYNC_FETCH_AND_XOR_1] = "__sync_fetch_and_xor_1";
348 Names[RTLIB::SYNC_FETCH_AND_XOR_2] = "__sync_fetch_and_xor_2";
349 Names[RTLIB::SYNC_FETCH_AND_XOR_4] = "__sync_fetch_and_xor_4";
350 Names[RTLIB::SYNC_FETCH_AND_XOR_8] = "__sync_fetch_and_xor_8";
351 Names[RTLIB::SYNC_FETCH_AND_XOR_16] = "__sync_fetch_and_xor_16";
352 Names[RTLIB::SYNC_FETCH_AND_NAND_1] = "__sync_fetch_and_nand_1";
353 Names[RTLIB::SYNC_FETCH_AND_NAND_2] = "__sync_fetch_and_nand_2";
354 Names[RTLIB::SYNC_FETCH_AND_NAND_4] = "__sync_fetch_and_nand_4";
355 Names[RTLIB::SYNC_FETCH_AND_NAND_8] = "__sync_fetch_and_nand_8";
356 Names[RTLIB::SYNC_FETCH_AND_NAND_16] = "__sync_fetch_and_nand_16";
358 if (Triple(TM.getTargetTriple()).getEnvironment() == Triple::GNU) {
359 Names[RTLIB::SINCOS_F32] = "sincosf";
360 Names[RTLIB::SINCOS_F64] = "sincos";
361 Names[RTLIB::SINCOS_F80] = "sincosl";
362 Names[RTLIB::SINCOS_F128] = "sincosl";
363 Names[RTLIB::SINCOS_PPCF128] = "sincosl";
365 // These are generally not available.
366 Names[RTLIB::SINCOS_F32] = 0;
367 Names[RTLIB::SINCOS_F64] = 0;
368 Names[RTLIB::SINCOS_F80] = 0;
369 Names[RTLIB::SINCOS_F128] = 0;
370 Names[RTLIB::SINCOS_PPCF128] = 0;
373 if (Triple(TM.getTargetTriple()).getOS() != Triple::OpenBSD) {
374 Names[RTLIB::STACKPROTECTOR_CHECK_FAIL] = "__stack_chk_fail";
376 // These are generally not available.
377 Names[RTLIB::STACKPROTECTOR_CHECK_FAIL] = 0;
381 /// InitLibcallCallingConvs - Set default libcall CallingConvs.
383 static void InitLibcallCallingConvs(CallingConv::ID *CCs) {
384 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
385 CCs[i] = CallingConv::C;
389 /// getFPEXT - Return the FPEXT_*_* value for the given types, or
390 /// UNKNOWN_LIBCALL if there is none.
391 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) {
392 if (OpVT == MVT::f32) {
393 if (RetVT == MVT::f64)
394 return FPEXT_F32_F64;
395 if (RetVT == MVT::f128)
396 return FPEXT_F32_F128;
397 } else if (OpVT == MVT::f64) {
398 if (RetVT == MVT::f128)
399 return FPEXT_F64_F128;
402 return UNKNOWN_LIBCALL;
405 /// getFPROUND - Return the FPROUND_*_* value for the given types, or
406 /// UNKNOWN_LIBCALL if there is none.
407 RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) {
408 if (RetVT == MVT::f32) {
409 if (OpVT == MVT::f64)
410 return FPROUND_F64_F32;
411 if (OpVT == MVT::f80)
412 return FPROUND_F80_F32;
413 if (OpVT == MVT::f128)
414 return FPROUND_F128_F32;
415 if (OpVT == MVT::ppcf128)
416 return FPROUND_PPCF128_F32;
417 } else if (RetVT == MVT::f64) {
418 if (OpVT == MVT::f80)
419 return FPROUND_F80_F64;
420 if (OpVT == MVT::f128)
421 return FPROUND_F128_F64;
422 if (OpVT == MVT::ppcf128)
423 return FPROUND_PPCF128_F64;
426 return UNKNOWN_LIBCALL;
429 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or
430 /// UNKNOWN_LIBCALL if there is none.
431 RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) {
432 if (OpVT == MVT::f32) {
433 if (RetVT == MVT::i8)
434 return FPTOSINT_F32_I8;
435 if (RetVT == MVT::i16)
436 return FPTOSINT_F32_I16;
437 if (RetVT == MVT::i32)
438 return FPTOSINT_F32_I32;
439 if (RetVT == MVT::i64)
440 return FPTOSINT_F32_I64;
441 if (RetVT == MVT::i128)
442 return FPTOSINT_F32_I128;
443 } else if (OpVT == MVT::f64) {
444 if (RetVT == MVT::i8)
445 return FPTOSINT_F64_I8;
446 if (RetVT == MVT::i16)
447 return FPTOSINT_F64_I16;
448 if (RetVT == MVT::i32)
449 return FPTOSINT_F64_I32;
450 if (RetVT == MVT::i64)
451 return FPTOSINT_F64_I64;
452 if (RetVT == MVT::i128)
453 return FPTOSINT_F64_I128;
454 } else if (OpVT == MVT::f80) {
455 if (RetVT == MVT::i32)
456 return FPTOSINT_F80_I32;
457 if (RetVT == MVT::i64)
458 return FPTOSINT_F80_I64;
459 if (RetVT == MVT::i128)
460 return FPTOSINT_F80_I128;
461 } else if (OpVT == MVT::f128) {
462 if (RetVT == MVT::i32)
463 return FPTOSINT_F128_I32;
464 if (RetVT == MVT::i64)
465 return FPTOSINT_F128_I64;
466 if (RetVT == MVT::i128)
467 return FPTOSINT_F128_I128;
468 } else if (OpVT == MVT::ppcf128) {
469 if (RetVT == MVT::i32)
470 return FPTOSINT_PPCF128_I32;
471 if (RetVT == MVT::i64)
472 return FPTOSINT_PPCF128_I64;
473 if (RetVT == MVT::i128)
474 return FPTOSINT_PPCF128_I128;
476 return UNKNOWN_LIBCALL;
479 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or
480 /// UNKNOWN_LIBCALL if there is none.
481 RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) {
482 if (OpVT == MVT::f32) {
483 if (RetVT == MVT::i8)
484 return FPTOUINT_F32_I8;
485 if (RetVT == MVT::i16)
486 return FPTOUINT_F32_I16;
487 if (RetVT == MVT::i32)
488 return FPTOUINT_F32_I32;
489 if (RetVT == MVT::i64)
490 return FPTOUINT_F32_I64;
491 if (RetVT == MVT::i128)
492 return FPTOUINT_F32_I128;
493 } else if (OpVT == MVT::f64) {
494 if (RetVT == MVT::i8)
495 return FPTOUINT_F64_I8;
496 if (RetVT == MVT::i16)
497 return FPTOUINT_F64_I16;
498 if (RetVT == MVT::i32)
499 return FPTOUINT_F64_I32;
500 if (RetVT == MVT::i64)
501 return FPTOUINT_F64_I64;
502 if (RetVT == MVT::i128)
503 return FPTOUINT_F64_I128;
504 } else if (OpVT == MVT::f80) {
505 if (RetVT == MVT::i32)
506 return FPTOUINT_F80_I32;
507 if (RetVT == MVT::i64)
508 return FPTOUINT_F80_I64;
509 if (RetVT == MVT::i128)
510 return FPTOUINT_F80_I128;
511 } else if (OpVT == MVT::f128) {
512 if (RetVT == MVT::i32)
513 return FPTOUINT_F128_I32;
514 if (RetVT == MVT::i64)
515 return FPTOUINT_F128_I64;
516 if (RetVT == MVT::i128)
517 return FPTOUINT_F128_I128;
518 } else if (OpVT == MVT::ppcf128) {
519 if (RetVT == MVT::i32)
520 return FPTOUINT_PPCF128_I32;
521 if (RetVT == MVT::i64)
522 return FPTOUINT_PPCF128_I64;
523 if (RetVT == MVT::i128)
524 return FPTOUINT_PPCF128_I128;
526 return UNKNOWN_LIBCALL;
529 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or
530 /// UNKNOWN_LIBCALL if there is none.
531 RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) {
532 if (OpVT == MVT::i32) {
533 if (RetVT == MVT::f32)
534 return SINTTOFP_I32_F32;
535 if (RetVT == MVT::f64)
536 return SINTTOFP_I32_F64;
537 if (RetVT == MVT::f80)
538 return SINTTOFP_I32_F80;
539 if (RetVT == MVT::f128)
540 return SINTTOFP_I32_F128;
541 if (RetVT == MVT::ppcf128)
542 return SINTTOFP_I32_PPCF128;
543 } else if (OpVT == MVT::i64) {
544 if (RetVT == MVT::f32)
545 return SINTTOFP_I64_F32;
546 if (RetVT == MVT::f64)
547 return SINTTOFP_I64_F64;
548 if (RetVT == MVT::f80)
549 return SINTTOFP_I64_F80;
550 if (RetVT == MVT::f128)
551 return SINTTOFP_I64_F128;
552 if (RetVT == MVT::ppcf128)
553 return SINTTOFP_I64_PPCF128;
554 } else if (OpVT == MVT::i128) {
555 if (RetVT == MVT::f32)
556 return SINTTOFP_I128_F32;
557 if (RetVT == MVT::f64)
558 return SINTTOFP_I128_F64;
559 if (RetVT == MVT::f80)
560 return SINTTOFP_I128_F80;
561 if (RetVT == MVT::f128)
562 return SINTTOFP_I128_F128;
563 if (RetVT == MVT::ppcf128)
564 return SINTTOFP_I128_PPCF128;
566 return UNKNOWN_LIBCALL;
569 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or
570 /// UNKNOWN_LIBCALL if there is none.
571 RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) {
572 if (OpVT == MVT::i32) {
573 if (RetVT == MVT::f32)
574 return UINTTOFP_I32_F32;
575 if (RetVT == MVT::f64)
576 return UINTTOFP_I32_F64;
577 if (RetVT == MVT::f80)
578 return UINTTOFP_I32_F80;
579 if (RetVT == MVT::f128)
580 return UINTTOFP_I32_F128;
581 if (RetVT == MVT::ppcf128)
582 return UINTTOFP_I32_PPCF128;
583 } else if (OpVT == MVT::i64) {
584 if (RetVT == MVT::f32)
585 return UINTTOFP_I64_F32;
586 if (RetVT == MVT::f64)
587 return UINTTOFP_I64_F64;
588 if (RetVT == MVT::f80)
589 return UINTTOFP_I64_F80;
590 if (RetVT == MVT::f128)
591 return UINTTOFP_I64_F128;
592 if (RetVT == MVT::ppcf128)
593 return UINTTOFP_I64_PPCF128;
594 } else if (OpVT == MVT::i128) {
595 if (RetVT == MVT::f32)
596 return UINTTOFP_I128_F32;
597 if (RetVT == MVT::f64)
598 return UINTTOFP_I128_F64;
599 if (RetVT == MVT::f80)
600 return UINTTOFP_I128_F80;
601 if (RetVT == MVT::f128)
602 return UINTTOFP_I128_F128;
603 if (RetVT == MVT::ppcf128)
604 return UINTTOFP_I128_PPCF128;
606 return UNKNOWN_LIBCALL;
609 /// InitCmpLibcallCCs - Set default comparison libcall CC.
611 static void InitCmpLibcallCCs(ISD::CondCode *CCs) {
612 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL);
613 CCs[RTLIB::OEQ_F32] = ISD::SETEQ;
614 CCs[RTLIB::OEQ_F64] = ISD::SETEQ;
615 CCs[RTLIB::OEQ_F128] = ISD::SETEQ;
616 CCs[RTLIB::UNE_F32] = ISD::SETNE;
617 CCs[RTLIB::UNE_F64] = ISD::SETNE;
618 CCs[RTLIB::UNE_F128] = ISD::SETNE;
619 CCs[RTLIB::OGE_F32] = ISD::SETGE;
620 CCs[RTLIB::OGE_F64] = ISD::SETGE;
621 CCs[RTLIB::OGE_F128] = ISD::SETGE;
622 CCs[RTLIB::OLT_F32] = ISD::SETLT;
623 CCs[RTLIB::OLT_F64] = ISD::SETLT;
624 CCs[RTLIB::OLT_F128] = ISD::SETLT;
625 CCs[RTLIB::OLE_F32] = ISD::SETLE;
626 CCs[RTLIB::OLE_F64] = ISD::SETLE;
627 CCs[RTLIB::OLE_F128] = ISD::SETLE;
628 CCs[RTLIB::OGT_F32] = ISD::SETGT;
629 CCs[RTLIB::OGT_F64] = ISD::SETGT;
630 CCs[RTLIB::OGT_F128] = ISD::SETGT;
631 CCs[RTLIB::UO_F32] = ISD::SETNE;
632 CCs[RTLIB::UO_F64] = ISD::SETNE;
633 CCs[RTLIB::UO_F128] = ISD::SETNE;
634 CCs[RTLIB::O_F32] = ISD::SETEQ;
635 CCs[RTLIB::O_F64] = ISD::SETEQ;
636 CCs[RTLIB::O_F128] = ISD::SETEQ;
639 /// NOTE: The constructor takes ownership of TLOF.
640 TargetLoweringBase::TargetLoweringBase(const TargetMachine &tm,
641 const TargetLoweringObjectFile *tlof)
642 : TM(tm), TD(TM.getDataLayout()), TLOF(*tlof) {
645 // Perform these initializations only once.
646 IsLittleEndian = TD->isLittleEndian();
647 PointerTy = MVT::getIntegerVT(8*TD->getPointerSize(0));
648 MaxStoresPerMemset = MaxStoresPerMemcpy = MaxStoresPerMemmove = 8;
649 MaxStoresPerMemsetOptSize = MaxStoresPerMemcpyOptSize
650 = MaxStoresPerMemmoveOptSize = 4;
651 UseUnderscoreSetJmp = false;
652 UseUnderscoreLongJmp = false;
653 SelectIsExpensive = false;
654 IntDivIsCheap = false;
655 Pow2DivIsCheap = false;
656 JumpIsExpensive = false;
657 PredictableSelectIsExpensive = false;
658 StackPointerRegisterToSaveRestore = 0;
659 ExceptionPointerRegister = 0;
660 ExceptionSelectorRegister = 0;
661 BooleanContents = UndefinedBooleanContent;
662 BooleanVectorContents = UndefinedBooleanContent;
663 SchedPreferenceInfo = Sched::ILP;
665 JumpBufAlignment = 0;
666 MinFunctionAlignment = 0;
667 PrefFunctionAlignment = 0;
668 PrefLoopAlignment = 0;
669 MinStackArgumentAlignment = 1;
670 InsertFencesForAtomic = false;
671 SupportJumpTables = true;
672 MinimumJumpTableEntries = 4;
674 InitLibcallNames(LibcallRoutineNames, TM);
675 InitCmpLibcallCCs(CmpLibcallCCs);
676 InitLibcallCallingConvs(LibcallCallingConvs);
679 TargetLoweringBase::~TargetLoweringBase() {
683 void TargetLoweringBase::initActions() {
684 // All operations default to being supported.
685 memset(OpActions, 0, sizeof(OpActions));
686 memset(LoadExtActions, 0, sizeof(LoadExtActions));
687 memset(TruncStoreActions, 0, sizeof(TruncStoreActions));
688 memset(IndexedModeActions, 0, sizeof(IndexedModeActions));
689 memset(CondCodeActions, 0, sizeof(CondCodeActions));
690 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE*sizeof(TargetRegisterClass*));
691 memset(TargetDAGCombineArray, 0, array_lengthof(TargetDAGCombineArray));
693 // Set default actions for various operations.
694 for (unsigned VT = 0; VT != (unsigned)MVT::LAST_VALUETYPE; ++VT) {
695 // Default all indexed load / store to expand.
696 for (unsigned IM = (unsigned)ISD::PRE_INC;
697 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
698 setIndexedLoadAction(IM, (MVT::SimpleValueType)VT, Expand);
699 setIndexedStoreAction(IM, (MVT::SimpleValueType)VT, Expand);
702 // These operations default to expand.
703 setOperationAction(ISD::FGETSIGN, (MVT::SimpleValueType)VT, Expand);
704 setOperationAction(ISD::CONCAT_VECTORS, (MVT::SimpleValueType)VT, Expand);
706 // These library functions default to expand.
707 setOperationAction(ISD::FROUND, (MVT::SimpleValueType)VT, Expand);
709 // These operations default to expand for vector types.
710 if (VT >= MVT::FIRST_VECTOR_VALUETYPE &&
711 VT <= MVT::LAST_VECTOR_VALUETYPE)
712 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
715 // Most targets ignore the @llvm.prefetch intrinsic.
716 setOperationAction(ISD::PREFETCH, MVT::Other, Expand);
718 // ConstantFP nodes default to expand. Targets can either change this to
719 // Legal, in which case all fp constants are legal, or use isFPImmLegal()
720 // to optimize expansions for certain constants.
721 setOperationAction(ISD::ConstantFP, MVT::f16, Expand);
722 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
723 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
724 setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
725 setOperationAction(ISD::ConstantFP, MVT::f128, Expand);
727 // These library functions default to expand.
728 setOperationAction(ISD::FLOG , MVT::f16, Expand);
729 setOperationAction(ISD::FLOG2, MVT::f16, Expand);
730 setOperationAction(ISD::FLOG10, MVT::f16, Expand);
731 setOperationAction(ISD::FEXP , MVT::f16, Expand);
732 setOperationAction(ISD::FEXP2, MVT::f16, Expand);
733 setOperationAction(ISD::FFLOOR, MVT::f16, Expand);
734 setOperationAction(ISD::FNEARBYINT, MVT::f16, Expand);
735 setOperationAction(ISD::FCEIL, MVT::f16, Expand);
736 setOperationAction(ISD::FRINT, MVT::f16, Expand);
737 setOperationAction(ISD::FTRUNC, MVT::f16, Expand);
738 setOperationAction(ISD::FLOG , MVT::f32, Expand);
739 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
740 setOperationAction(ISD::FLOG10, MVT::f32, Expand);
741 setOperationAction(ISD::FEXP , MVT::f32, Expand);
742 setOperationAction(ISD::FEXP2, MVT::f32, Expand);
743 setOperationAction(ISD::FFLOOR, MVT::f32, Expand);
744 setOperationAction(ISD::FNEARBYINT, MVT::f32, Expand);
745 setOperationAction(ISD::FCEIL, MVT::f32, Expand);
746 setOperationAction(ISD::FRINT, MVT::f32, Expand);
747 setOperationAction(ISD::FTRUNC, MVT::f32, Expand);
748 setOperationAction(ISD::FLOG , MVT::f64, Expand);
749 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
750 setOperationAction(ISD::FLOG10, MVT::f64, Expand);
751 setOperationAction(ISD::FEXP , MVT::f64, Expand);
752 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
753 setOperationAction(ISD::FFLOOR, MVT::f64, Expand);
754 setOperationAction(ISD::FNEARBYINT, MVT::f64, Expand);
755 setOperationAction(ISD::FCEIL, MVT::f64, Expand);
756 setOperationAction(ISD::FRINT, MVT::f64, Expand);
757 setOperationAction(ISD::FTRUNC, MVT::f64, Expand);
758 setOperationAction(ISD::FLOG , MVT::f128, Expand);
759 setOperationAction(ISD::FLOG2, MVT::f128, Expand);
760 setOperationAction(ISD::FLOG10, MVT::f128, Expand);
761 setOperationAction(ISD::FEXP , MVT::f128, Expand);
762 setOperationAction(ISD::FEXP2, MVT::f128, Expand);
763 setOperationAction(ISD::FFLOOR, MVT::f128, Expand);
764 setOperationAction(ISD::FNEARBYINT, MVT::f128, Expand);
765 setOperationAction(ISD::FCEIL, MVT::f128, Expand);
766 setOperationAction(ISD::FRINT, MVT::f128, Expand);
767 setOperationAction(ISD::FTRUNC, MVT::f128, Expand);
769 // Default ISD::TRAP to expand (which turns it into abort).
770 setOperationAction(ISD::TRAP, MVT::Other, Expand);
772 // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand"
773 // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP.
775 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Expand);
778 MVT TargetLoweringBase::getPointerTy(uint32_t AS) const {
779 return MVT::getIntegerVT(getPointerSizeInBits(AS));
782 unsigned TargetLoweringBase::getPointerSizeInBits(uint32_t AS) const {
783 return TD->getPointerSizeInBits(AS);
786 unsigned TargetLoweringBase::getPointerTypeSizeInBits(Type *Ty) const {
787 assert(Ty->isPointerTy());
788 return getPointerSizeInBits(Ty->getPointerAddressSpace());
791 MVT TargetLoweringBase::getScalarShiftAmountTy(EVT LHSTy) const {
792 return MVT::getIntegerVT(8*TD->getPointerSize(0));
795 EVT TargetLoweringBase::getShiftAmountTy(EVT LHSTy) const {
796 assert(LHSTy.isInteger() && "Shift amount is not an integer type!");
797 if (LHSTy.isVector())
799 return getScalarShiftAmountTy(LHSTy);
802 /// canOpTrap - Returns true if the operation can trap for the value type.
803 /// VT must be a legal type.
804 bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const {
805 assert(isTypeLegal(VT));
820 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT,
821 unsigned &NumIntermediates,
823 TargetLoweringBase *TLI) {
824 // Figure out the right, legal destination reg to copy into.
825 unsigned NumElts = VT.getVectorNumElements();
826 MVT EltTy = VT.getVectorElementType();
828 unsigned NumVectorRegs = 1;
830 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
831 // could break down into LHS/RHS like LegalizeDAG does.
832 if (!isPowerOf2_32(NumElts)) {
833 NumVectorRegs = NumElts;
837 // Divide the input until we get to a supported size. This will always
838 // end with a scalar if the target doesn't support vectors.
839 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) {
844 NumIntermediates = NumVectorRegs;
846 MVT NewVT = MVT::getVectorVT(EltTy, NumElts);
847 if (!TLI->isTypeLegal(NewVT))
849 IntermediateVT = NewVT;
851 unsigned NewVTSize = NewVT.getSizeInBits();
853 // Convert sizes such as i33 to i64.
854 if (!isPowerOf2_32(NewVTSize))
855 NewVTSize = NextPowerOf2(NewVTSize);
857 MVT DestVT = TLI->getRegisterType(NewVT);
859 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16.
860 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits());
862 // Otherwise, promotion or legal types use the same number of registers as
863 // the vector decimated to the appropriate level.
864 return NumVectorRegs;
867 /// isLegalRC - Return true if the value types that can be represented by the
868 /// specified register class are all legal.
869 bool TargetLoweringBase::isLegalRC(const TargetRegisterClass *RC) const {
870 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
878 /// findRepresentativeClass - Return the largest legal super-reg register class
879 /// of the register class for the specified type and its associated "cost".
880 std::pair<const TargetRegisterClass*, uint8_t>
881 TargetLoweringBase::findRepresentativeClass(MVT VT) const {
882 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
883 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy];
885 return std::make_pair(RC, 0);
887 // Compute the set of all super-register classes.
888 BitVector SuperRegRC(TRI->getNumRegClasses());
889 for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI)
890 SuperRegRC.setBitsInMask(RCI.getMask());
892 // Find the first legal register class with the largest spill size.
893 const TargetRegisterClass *BestRC = RC;
894 for (int i = SuperRegRC.find_first(); i >= 0; i = SuperRegRC.find_next(i)) {
895 const TargetRegisterClass *SuperRC = TRI->getRegClass(i);
896 // We want the largest possible spill size.
897 if (SuperRC->getSize() <= BestRC->getSize())
899 if (!isLegalRC(SuperRC))
903 return std::make_pair(BestRC, 1);
906 /// computeRegisterProperties - Once all of the register classes are added,
907 /// this allows us to compute derived properties we expose.
908 void TargetLoweringBase::computeRegisterProperties() {
909 assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE &&
910 "Too many value types for ValueTypeActions to hold!");
912 // Everything defaults to needing one register.
913 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
914 NumRegistersForVT[i] = 1;
915 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i;
917 // ...except isVoid, which doesn't need any registers.
918 NumRegistersForVT[MVT::isVoid] = 0;
920 // Find the largest integer register class.
921 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE;
922 for (; RegClassForVT[LargestIntReg] == 0; --LargestIntReg)
923 assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
925 // Every integer value type larger than this largest register takes twice as
926 // many registers to represent as the previous ValueType.
927 for (unsigned ExpandedReg = LargestIntReg + 1;
928 ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) {
929 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
930 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg;
931 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1);
932 ValueTypeActions.setTypeAction((MVT::SimpleValueType)ExpandedReg,
936 // Inspect all of the ValueType's smaller than the largest integer
937 // register to see which ones need promotion.
938 unsigned LegalIntReg = LargestIntReg;
939 for (unsigned IntReg = LargestIntReg - 1;
940 IntReg >= (unsigned)MVT::i1; --IntReg) {
941 MVT IVT = (MVT::SimpleValueType)IntReg;
942 if (isTypeLegal(IVT)) {
943 LegalIntReg = IntReg;
945 RegisterTypeForVT[IntReg] = TransformToType[IntReg] =
946 (const MVT::SimpleValueType)LegalIntReg;
947 ValueTypeActions.setTypeAction(IVT, TypePromoteInteger);
951 // ppcf128 type is really two f64's.
952 if (!isTypeLegal(MVT::ppcf128)) {
953 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64];
954 RegisterTypeForVT[MVT::ppcf128] = MVT::f64;
955 TransformToType[MVT::ppcf128] = MVT::f64;
956 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat);
959 // Decide how to handle f128. If the target does not have native f128 support,
960 // expand it to i128 and we will be generating soft float library calls.
961 if (!isTypeLegal(MVT::f128)) {
962 NumRegistersForVT[MVT::f128] = NumRegistersForVT[MVT::i128];
963 RegisterTypeForVT[MVT::f128] = RegisterTypeForVT[MVT::i128];
964 TransformToType[MVT::f128] = MVT::i128;
965 ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat);
968 // Decide how to handle f64. If the target does not have native f64 support,
969 // expand it to i64 and we will be generating soft float library calls.
970 if (!isTypeLegal(MVT::f64)) {
971 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
972 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
973 TransformToType[MVT::f64] = MVT::i64;
974 ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat);
977 // Decide how to handle f32. If the target does not have native support for
978 // f32, promote it to f64 if it is legal. Otherwise, expand it to i32.
979 if (!isTypeLegal(MVT::f32)) {
980 if (isTypeLegal(MVT::f64)) {
981 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::f64];
982 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::f64];
983 TransformToType[MVT::f32] = MVT::f64;
984 ValueTypeActions.setTypeAction(MVT::f32, TypePromoteInteger);
986 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
987 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
988 TransformToType[MVT::f32] = MVT::i32;
989 ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat);
993 // Loop over all of the vector value types to see which need transformations.
994 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
995 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
996 MVT VT = (MVT::SimpleValueType)i;
997 if (isTypeLegal(VT)) continue;
999 // Determine if there is a legal wider type. If so, we should promote to
1000 // that wider vector type.
1001 MVT EltVT = VT.getVectorElementType();
1002 unsigned NElts = VT.getVectorNumElements();
1003 if (NElts != 1 && !shouldSplitVectorElementType(EltVT)) {
1004 bool IsLegalWiderType = false;
1005 // First try to promote the elements of integer vectors. If no legal
1006 // promotion was found, fallback to the widen-vector method.
1007 for (unsigned nVT = i+1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
1008 MVT SVT = (MVT::SimpleValueType)nVT;
1009 // Promote vectors of integers to vectors with the same number
1010 // of elements, with a wider element type.
1011 if (SVT.getVectorElementType().getSizeInBits() > EltVT.getSizeInBits()
1012 && SVT.getVectorNumElements() == NElts &&
1013 isTypeLegal(SVT) && SVT.getScalarType().isInteger()) {
1014 TransformToType[i] = SVT;
1015 RegisterTypeForVT[i] = SVT;
1016 NumRegistersForVT[i] = 1;
1017 ValueTypeActions.setTypeAction(VT, TypePromoteInteger);
1018 IsLegalWiderType = true;
1023 if (IsLegalWiderType) continue;
1025 // Try to widen the vector.
1026 for (unsigned nVT = i+1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
1027 MVT SVT = (MVT::SimpleValueType)nVT;
1028 if (SVT.getVectorElementType() == EltVT &&
1029 SVT.getVectorNumElements() > NElts &&
1031 TransformToType[i] = SVT;
1032 RegisterTypeForVT[i] = SVT;
1033 NumRegistersForVT[i] = 1;
1034 ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1035 IsLegalWiderType = true;
1039 if (IsLegalWiderType) continue;
1044 unsigned NumIntermediates;
1045 NumRegistersForVT[i] =
1046 getVectorTypeBreakdownMVT(VT, IntermediateVT, NumIntermediates,
1048 RegisterTypeForVT[i] = RegisterVT;
1050 MVT NVT = VT.getPow2VectorType();
1052 // Type is already a power of 2. The default action is to split.
1053 TransformToType[i] = MVT::Other;
1054 unsigned NumElts = VT.getVectorNumElements();
1055 ValueTypeActions.setTypeAction(VT,
1056 NumElts > 1 ? TypeSplitVector : TypeScalarizeVector);
1058 TransformToType[i] = NVT;
1059 ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1063 // Determine the 'representative' register class for each value type.
1064 // An representative register class is the largest (meaning one which is
1065 // not a sub-register class / subreg register class) legal register class for
1066 // a group of value types. For example, on i386, i8, i16, and i32
1067 // representative would be GR32; while on x86_64 it's GR64.
1068 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
1069 const TargetRegisterClass* RRC;
1071 tie(RRC, Cost) = findRepresentativeClass((MVT::SimpleValueType)i);
1072 RepRegClassForVT[i] = RRC;
1073 RepRegClassCostForVT[i] = Cost;
1077 EVT TargetLoweringBase::getSetCCResultType(LLVMContext &, EVT VT) const {
1078 assert(!VT.isVector() && "No default SetCC type for vectors!");
1079 return getPointerTy(0).SimpleTy;
1082 MVT::SimpleValueType TargetLoweringBase::getCmpLibcallReturnType() const {
1083 return MVT::i32; // return the default value
1086 /// getVectorTypeBreakdown - Vector types are broken down into some number of
1087 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
1088 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
1089 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
1091 /// This method returns the number of registers needed, and the VT for each
1092 /// register. It also returns the VT and quantity of the intermediate values
1093 /// before they are promoted/expanded.
1095 unsigned TargetLoweringBase::getVectorTypeBreakdown(LLVMContext &Context, EVT VT,
1096 EVT &IntermediateVT,
1097 unsigned &NumIntermediates,
1098 MVT &RegisterVT) const {
1099 unsigned NumElts = VT.getVectorNumElements();
1101 // If there is a wider vector type with the same element type as this one,
1102 // or a promoted vector type that has the same number of elements which
1103 // are wider, then we should convert to that legal vector type.
1104 // This handles things like <2 x float> -> <4 x float> and
1105 // <4 x i1> -> <4 x i32>.
1106 LegalizeTypeAction TA = getTypeAction(Context, VT);
1107 if (NumElts != 1 && (TA == TypeWidenVector || TA == TypePromoteInteger)) {
1108 EVT RegisterEVT = getTypeToTransformTo(Context, VT);
1109 if (isTypeLegal(RegisterEVT)) {
1110 IntermediateVT = RegisterEVT;
1111 RegisterVT = RegisterEVT.getSimpleVT();
1112 NumIntermediates = 1;
1117 // Figure out the right, legal destination reg to copy into.
1118 EVT EltTy = VT.getVectorElementType();
1120 unsigned NumVectorRegs = 1;
1122 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
1123 // could break down into LHS/RHS like LegalizeDAG does.
1124 if (!isPowerOf2_32(NumElts)) {
1125 NumVectorRegs = NumElts;
1129 // Divide the input until we get to a supported size. This will always
1130 // end with a scalar if the target doesn't support vectors.
1131 while (NumElts > 1 && !isTypeLegal(
1132 EVT::getVectorVT(Context, EltTy, NumElts))) {
1134 NumVectorRegs <<= 1;
1137 NumIntermediates = NumVectorRegs;
1139 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts);
1140 if (!isTypeLegal(NewVT))
1142 IntermediateVT = NewVT;
1144 MVT DestVT = getRegisterType(Context, NewVT);
1145 RegisterVT = DestVT;
1146 unsigned NewVTSize = NewVT.getSizeInBits();
1148 // Convert sizes such as i33 to i64.
1149 if (!isPowerOf2_32(NewVTSize))
1150 NewVTSize = NextPowerOf2(NewVTSize);
1152 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16.
1153 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits());
1155 // Otherwise, promotion or legal types use the same number of registers as
1156 // the vector decimated to the appropriate level.
1157 return NumVectorRegs;
1160 /// Get the EVTs and ArgFlags collections that represent the legalized return
1161 /// type of the given function. This does not require a DAG or a return value,
1162 /// and is suitable for use before any DAGs for the function are constructed.
1163 /// TODO: Move this out of TargetLowering.cpp.
1164 void llvm::GetReturnInfo(Type* ReturnType, AttributeSet attr,
1165 SmallVectorImpl<ISD::OutputArg> &Outs,
1166 const TargetLowering &TLI) {
1167 SmallVector<EVT, 4> ValueVTs;
1168 ComputeValueVTs(TLI, ReturnType, ValueVTs);
1169 unsigned NumValues = ValueVTs.size();
1170 if (NumValues == 0) return;
1172 for (unsigned j = 0, f = NumValues; j != f; ++j) {
1173 EVT VT = ValueVTs[j];
1174 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
1176 if (attr.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
1177 ExtendKind = ISD::SIGN_EXTEND;
1178 else if (attr.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt))
1179 ExtendKind = ISD::ZERO_EXTEND;
1181 // FIXME: C calling convention requires the return type to be promoted to
1182 // at least 32-bit. But this is not necessary for non-C calling
1183 // conventions. The frontend should mark functions whose return values
1184 // require promoting with signext or zeroext attributes.
1185 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) {
1186 MVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32);
1187 if (VT.bitsLT(MinVT))
1191 unsigned NumParts = TLI.getNumRegisters(ReturnType->getContext(), VT);
1192 MVT PartVT = TLI.getRegisterType(ReturnType->getContext(), VT);
1194 // 'inreg' on function refers to return value
1195 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1196 if (attr.hasAttribute(AttributeSet::ReturnIndex, Attribute::InReg))
1199 // Propagate extension type if any
1200 if (attr.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
1202 else if (attr.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt))
1205 for (unsigned i = 0; i < NumParts; ++i)
1206 Outs.push_back(ISD::OutputArg(Flags, PartVT, /*isFixed=*/true, 0, 0));
1210 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1211 /// function arguments in the caller parameter area. This is the actual
1212 /// alignment, not its logarithm.
1213 unsigned TargetLoweringBase::getByValTypeAlignment(Type *Ty) const {
1214 return TD->getCallFrameTypeAlignment(Ty);
1217 //===----------------------------------------------------------------------===//
1218 // TargetTransformInfo Helpers
1219 //===----------------------------------------------------------------------===//
1221 int TargetLoweringBase::InstructionOpcodeToISD(unsigned Opcode) const {
1222 enum InstructionOpcodes {
1223 #define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM,
1224 #define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM
1225 #include "llvm/IR/Instruction.def"
1227 switch (static_cast<InstructionOpcodes>(Opcode)) {
1230 case Switch: return 0;
1231 case IndirectBr: return 0;
1232 case Invoke: return 0;
1233 case Resume: return 0;
1234 case Unreachable: return 0;
1235 case Add: return ISD::ADD;
1236 case FAdd: return ISD::FADD;
1237 case Sub: return ISD::SUB;
1238 case FSub: return ISD::FSUB;
1239 case Mul: return ISD::MUL;
1240 case FMul: return ISD::FMUL;
1241 case UDiv: return ISD::UDIV;
1242 case SDiv: return ISD::UDIV;
1243 case FDiv: return ISD::FDIV;
1244 case URem: return ISD::UREM;
1245 case SRem: return ISD::SREM;
1246 case FRem: return ISD::FREM;
1247 case Shl: return ISD::SHL;
1248 case LShr: return ISD::SRL;
1249 case AShr: return ISD::SRA;
1250 case And: return ISD::AND;
1251 case Or: return ISD::OR;
1252 case Xor: return ISD::XOR;
1253 case Alloca: return 0;
1254 case Load: return ISD::LOAD;
1255 case Store: return ISD::STORE;
1256 case GetElementPtr: return 0;
1257 case Fence: return 0;
1258 case AtomicCmpXchg: return 0;
1259 case AtomicRMW: return 0;
1260 case Trunc: return ISD::TRUNCATE;
1261 case ZExt: return ISD::ZERO_EXTEND;
1262 case SExt: return ISD::SIGN_EXTEND;
1263 case FPToUI: return ISD::FP_TO_UINT;
1264 case FPToSI: return ISD::FP_TO_SINT;
1265 case UIToFP: return ISD::UINT_TO_FP;
1266 case SIToFP: return ISD::SINT_TO_FP;
1267 case FPTrunc: return ISD::FP_ROUND;
1268 case FPExt: return ISD::FP_EXTEND;
1269 case PtrToInt: return ISD::BITCAST;
1270 case IntToPtr: return ISD::BITCAST;
1271 case BitCast: return ISD::BITCAST;
1272 case ICmp: return ISD::SETCC;
1273 case FCmp: return ISD::SETCC;
1275 case Call: return 0;
1276 case Select: return ISD::SELECT;
1277 case UserOp1: return 0;
1278 case UserOp2: return 0;
1279 case VAArg: return 0;
1280 case ExtractElement: return ISD::EXTRACT_VECTOR_ELT;
1281 case InsertElement: return ISD::INSERT_VECTOR_ELT;
1282 case ShuffleVector: return ISD::VECTOR_SHUFFLE;
1283 case ExtractValue: return ISD::MERGE_VALUES;
1284 case InsertValue: return ISD::MERGE_VALUES;
1285 case LandingPad: return 0;
1288 llvm_unreachable("Unknown instruction type encountered!");
1291 std::pair<unsigned, MVT>
1292 TargetLoweringBase::getTypeLegalizationCost(Type *Ty) const {
1293 LLVMContext &C = Ty->getContext();
1294 EVT MTy = getValueType(Ty);
1297 // We keep legalizing the type until we find a legal kind. We assume that
1298 // the only operation that costs anything is the split. After splitting
1299 // we need to handle two types.
1301 LegalizeKind LK = getTypeConversion(C, MTy);
1303 if (LK.first == TypeLegal)
1304 return std::make_pair(Cost, MTy.getSimpleVT());
1306 if (LK.first == TypeSplitVector || LK.first == TypeExpandInteger)
1309 // Keep legalizing the type.
1314 //===----------------------------------------------------------------------===//
1315 // Loop Strength Reduction hooks
1316 //===----------------------------------------------------------------------===//
1318 /// isLegalAddressingMode - Return true if the addressing mode represented
1319 /// by AM is legal for this target, for a load/store of the specified type.
1320 bool TargetLoweringBase::isLegalAddressingMode(const AddrMode &AM,
1322 // The default implementation of this implements a conservative RISCy, r+r and
1325 // Allows a sign-extended 16-bit immediate field.
1326 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
1329 // No global is ever allowed as a base.
1333 // Only support r+r,
1335 case 0: // "r+i" or just "i", depending on HasBaseReg.
1338 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
1340 // Otherwise we have r+r or r+i.
1343 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
1345 // Allow 2*r as r+r.