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/CodeGen/Analysis.h"
18 #include "llvm/CodeGen/MachineFrameInfo.h"
19 #include "llvm/CodeGen/MachineFunction.h"
20 #include "llvm/CodeGen/MachineJumpTableInfo.h"
21 #include "llvm/IR/DataLayout.h"
22 #include "llvm/IR/DerivedTypes.h"
23 #include "llvm/IR/GlobalVariable.h"
24 #include "llvm/MC/MCAsmInfo.h"
25 #include "llvm/MC/MCExpr.h"
26 #include "llvm/Support/CommandLine.h"
27 #include "llvm/Support/ErrorHandling.h"
28 #include "llvm/Support/MathExtras.h"
29 #include "llvm/Target/TargetLoweringObjectFile.h"
30 #include "llvm/Target/TargetMachine.h"
31 #include "llvm/Target/TargetRegisterInfo.h"
35 /// InitLibcallNames - Set default libcall names.
37 static void InitLibcallNames(const char **Names) {
38 Names[RTLIB::SHL_I16] = "__ashlhi3";
39 Names[RTLIB::SHL_I32] = "__ashlsi3";
40 Names[RTLIB::SHL_I64] = "__ashldi3";
41 Names[RTLIB::SHL_I128] = "__ashlti3";
42 Names[RTLIB::SRL_I16] = "__lshrhi3";
43 Names[RTLIB::SRL_I32] = "__lshrsi3";
44 Names[RTLIB::SRL_I64] = "__lshrdi3";
45 Names[RTLIB::SRL_I128] = "__lshrti3";
46 Names[RTLIB::SRA_I16] = "__ashrhi3";
47 Names[RTLIB::SRA_I32] = "__ashrsi3";
48 Names[RTLIB::SRA_I64] = "__ashrdi3";
49 Names[RTLIB::SRA_I128] = "__ashrti3";
50 Names[RTLIB::MUL_I8] = "__mulqi3";
51 Names[RTLIB::MUL_I16] = "__mulhi3";
52 Names[RTLIB::MUL_I32] = "__mulsi3";
53 Names[RTLIB::MUL_I64] = "__muldi3";
54 Names[RTLIB::MUL_I128] = "__multi3";
55 Names[RTLIB::MULO_I32] = "__mulosi4";
56 Names[RTLIB::MULO_I64] = "__mulodi4";
57 Names[RTLIB::MULO_I128] = "__muloti4";
58 Names[RTLIB::SDIV_I8] = "__divqi3";
59 Names[RTLIB::SDIV_I16] = "__divhi3";
60 Names[RTLIB::SDIV_I32] = "__divsi3";
61 Names[RTLIB::SDIV_I64] = "__divdi3";
62 Names[RTLIB::SDIV_I128] = "__divti3";
63 Names[RTLIB::UDIV_I8] = "__udivqi3";
64 Names[RTLIB::UDIV_I16] = "__udivhi3";
65 Names[RTLIB::UDIV_I32] = "__udivsi3";
66 Names[RTLIB::UDIV_I64] = "__udivdi3";
67 Names[RTLIB::UDIV_I128] = "__udivti3";
68 Names[RTLIB::SREM_I8] = "__modqi3";
69 Names[RTLIB::SREM_I16] = "__modhi3";
70 Names[RTLIB::SREM_I32] = "__modsi3";
71 Names[RTLIB::SREM_I64] = "__moddi3";
72 Names[RTLIB::SREM_I128] = "__modti3";
73 Names[RTLIB::UREM_I8] = "__umodqi3";
74 Names[RTLIB::UREM_I16] = "__umodhi3";
75 Names[RTLIB::UREM_I32] = "__umodsi3";
76 Names[RTLIB::UREM_I64] = "__umoddi3";
77 Names[RTLIB::UREM_I128] = "__umodti3";
79 // These are generally not available.
80 Names[RTLIB::SDIVREM_I8] = 0;
81 Names[RTLIB::SDIVREM_I16] = 0;
82 Names[RTLIB::SDIVREM_I32] = 0;
83 Names[RTLIB::SDIVREM_I64] = 0;
84 Names[RTLIB::SDIVREM_I128] = 0;
85 Names[RTLIB::UDIVREM_I8] = 0;
86 Names[RTLIB::UDIVREM_I16] = 0;
87 Names[RTLIB::UDIVREM_I32] = 0;
88 Names[RTLIB::UDIVREM_I64] = 0;
89 Names[RTLIB::UDIVREM_I128] = 0;
91 Names[RTLIB::NEG_I32] = "__negsi2";
92 Names[RTLIB::NEG_I64] = "__negdi2";
93 Names[RTLIB::ADD_F32] = "__addsf3";
94 Names[RTLIB::ADD_F64] = "__adddf3";
95 Names[RTLIB::ADD_F80] = "__addxf3";
96 Names[RTLIB::ADD_F128] = "__addtf3";
97 Names[RTLIB::ADD_PPCF128] = "__gcc_qadd";
98 Names[RTLIB::SUB_F32] = "__subsf3";
99 Names[RTLIB::SUB_F64] = "__subdf3";
100 Names[RTLIB::SUB_F80] = "__subxf3";
101 Names[RTLIB::SUB_F128] = "__subtf3";
102 Names[RTLIB::SUB_PPCF128] = "__gcc_qsub";
103 Names[RTLIB::MUL_F32] = "__mulsf3";
104 Names[RTLIB::MUL_F64] = "__muldf3";
105 Names[RTLIB::MUL_F80] = "__mulxf3";
106 Names[RTLIB::MUL_F128] = "__multf3";
107 Names[RTLIB::MUL_PPCF128] = "__gcc_qmul";
108 Names[RTLIB::DIV_F32] = "__divsf3";
109 Names[RTLIB::DIV_F64] = "__divdf3";
110 Names[RTLIB::DIV_F80] = "__divxf3";
111 Names[RTLIB::DIV_F128] = "__divtf3";
112 Names[RTLIB::DIV_PPCF128] = "__gcc_qdiv";
113 Names[RTLIB::REM_F32] = "fmodf";
114 Names[RTLIB::REM_F64] = "fmod";
115 Names[RTLIB::REM_F80] = "fmodl";
116 Names[RTLIB::REM_F128] = "fmodl";
117 Names[RTLIB::REM_PPCF128] = "fmodl";
118 Names[RTLIB::FMA_F32] = "fmaf";
119 Names[RTLIB::FMA_F64] = "fma";
120 Names[RTLIB::FMA_F80] = "fmal";
121 Names[RTLIB::FMA_F128] = "fmal";
122 Names[RTLIB::FMA_PPCF128] = "fmal";
123 Names[RTLIB::POWI_F32] = "__powisf2";
124 Names[RTLIB::POWI_F64] = "__powidf2";
125 Names[RTLIB::POWI_F80] = "__powixf2";
126 Names[RTLIB::POWI_F128] = "__powitf2";
127 Names[RTLIB::POWI_PPCF128] = "__powitf2";
128 Names[RTLIB::SQRT_F32] = "sqrtf";
129 Names[RTLIB::SQRT_F64] = "sqrt";
130 Names[RTLIB::SQRT_F80] = "sqrtl";
131 Names[RTLIB::SQRT_F128] = "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_F128] = "logl";
137 Names[RTLIB::LOG_PPCF128] = "logl";
138 Names[RTLIB::LOG2_F32] = "log2f";
139 Names[RTLIB::LOG2_F64] = "log2";
140 Names[RTLIB::LOG2_F80] = "log2l";
141 Names[RTLIB::LOG2_F128] = "log2l";
142 Names[RTLIB::LOG2_PPCF128] = "log2l";
143 Names[RTLIB::LOG10_F32] = "log10f";
144 Names[RTLIB::LOG10_F64] = "log10";
145 Names[RTLIB::LOG10_F80] = "log10l";
146 Names[RTLIB::LOG10_F128] = "log10l";
147 Names[RTLIB::LOG10_PPCF128] = "log10l";
148 Names[RTLIB::EXP_F32] = "expf";
149 Names[RTLIB::EXP_F64] = "exp";
150 Names[RTLIB::EXP_F80] = "expl";
151 Names[RTLIB::EXP_F128] = "expl";
152 Names[RTLIB::EXP_PPCF128] = "expl";
153 Names[RTLIB::EXP2_F32] = "exp2f";
154 Names[RTLIB::EXP2_F64] = "exp2";
155 Names[RTLIB::EXP2_F80] = "exp2l";
156 Names[RTLIB::EXP2_F128] = "exp2l";
157 Names[RTLIB::EXP2_PPCF128] = "exp2l";
158 Names[RTLIB::SIN_F32] = "sinf";
159 Names[RTLIB::SIN_F64] = "sin";
160 Names[RTLIB::SIN_F80] = "sinl";
161 Names[RTLIB::SIN_F128] = "sinl";
162 Names[RTLIB::SIN_PPCF128] = "sinl";
163 Names[RTLIB::COS_F32] = "cosf";
164 Names[RTLIB::COS_F64] = "cos";
165 Names[RTLIB::COS_F80] = "cosl";
166 Names[RTLIB::COS_F128] = "cosl";
167 Names[RTLIB::COS_PPCF128] = "cosl";
168 Names[RTLIB::POW_F32] = "powf";
169 Names[RTLIB::POW_F64] = "pow";
170 Names[RTLIB::POW_F80] = "powl";
171 Names[RTLIB::POW_F128] = "powl";
172 Names[RTLIB::POW_PPCF128] = "powl";
173 Names[RTLIB::CEIL_F32] = "ceilf";
174 Names[RTLIB::CEIL_F64] = "ceil";
175 Names[RTLIB::CEIL_F80] = "ceill";
176 Names[RTLIB::CEIL_F128] = "ceill";
177 Names[RTLIB::CEIL_PPCF128] = "ceill";
178 Names[RTLIB::TRUNC_F32] = "truncf";
179 Names[RTLIB::TRUNC_F64] = "trunc";
180 Names[RTLIB::TRUNC_F80] = "truncl";
181 Names[RTLIB::TRUNC_F128] = "truncl";
182 Names[RTLIB::TRUNC_PPCF128] = "truncl";
183 Names[RTLIB::RINT_F32] = "rintf";
184 Names[RTLIB::RINT_F64] = "rint";
185 Names[RTLIB::RINT_F80] = "rintl";
186 Names[RTLIB::RINT_F128] = "rintl";
187 Names[RTLIB::RINT_PPCF128] = "rintl";
188 Names[RTLIB::NEARBYINT_F32] = "nearbyintf";
189 Names[RTLIB::NEARBYINT_F64] = "nearbyint";
190 Names[RTLIB::NEARBYINT_F80] = "nearbyintl";
191 Names[RTLIB::NEARBYINT_F128] = "nearbyintl";
192 Names[RTLIB::NEARBYINT_PPCF128] = "nearbyintl";
193 Names[RTLIB::FLOOR_F32] = "floorf";
194 Names[RTLIB::FLOOR_F64] = "floor";
195 Names[RTLIB::FLOOR_F80] = "floorl";
196 Names[RTLIB::FLOOR_F128] = "floorl";
197 Names[RTLIB::FLOOR_PPCF128] = "floorl";
198 Names[RTLIB::COPYSIGN_F32] = "copysignf";
199 Names[RTLIB::COPYSIGN_F64] = "copysign";
200 Names[RTLIB::COPYSIGN_F80] = "copysignl";
201 Names[RTLIB::COPYSIGN_F128] = "copysignl";
202 Names[RTLIB::COPYSIGN_PPCF128] = "copysignl";
203 Names[RTLIB::FPEXT_F64_F128] = "__extenddftf2";
204 Names[RTLIB::FPEXT_F32_F128] = "__extendsftf2";
205 Names[RTLIB::FPEXT_F32_F64] = "__extendsfdf2";
206 Names[RTLIB::FPEXT_F16_F32] = "__gnu_h2f_ieee";
207 Names[RTLIB::FPROUND_F32_F16] = "__gnu_f2h_ieee";
208 Names[RTLIB::FPROUND_F64_F32] = "__truncdfsf2";
209 Names[RTLIB::FPROUND_F80_F32] = "__truncxfsf2";
210 Names[RTLIB::FPROUND_F128_F32] = "__trunctfsf2";
211 Names[RTLIB::FPROUND_PPCF128_F32] = "__trunctfsf2";
212 Names[RTLIB::FPROUND_F80_F64] = "__truncxfdf2";
213 Names[RTLIB::FPROUND_F128_F64] = "__trunctfdf2";
214 Names[RTLIB::FPROUND_PPCF128_F64] = "__trunctfdf2";
215 Names[RTLIB::FPTOSINT_F32_I8] = "__fixsfqi";
216 Names[RTLIB::FPTOSINT_F32_I16] = "__fixsfhi";
217 Names[RTLIB::FPTOSINT_F32_I32] = "__fixsfsi";
218 Names[RTLIB::FPTOSINT_F32_I64] = "__fixsfdi";
219 Names[RTLIB::FPTOSINT_F32_I128] = "__fixsfti";
220 Names[RTLIB::FPTOSINT_F64_I8] = "__fixdfqi";
221 Names[RTLIB::FPTOSINT_F64_I16] = "__fixdfhi";
222 Names[RTLIB::FPTOSINT_F64_I32] = "__fixdfsi";
223 Names[RTLIB::FPTOSINT_F64_I64] = "__fixdfdi";
224 Names[RTLIB::FPTOSINT_F64_I128] = "__fixdfti";
225 Names[RTLIB::FPTOSINT_F80_I32] = "__fixxfsi";
226 Names[RTLIB::FPTOSINT_F80_I64] = "__fixxfdi";
227 Names[RTLIB::FPTOSINT_F80_I128] = "__fixxfti";
228 Names[RTLIB::FPTOSINT_F128_I32] = "__fixtfsi";
229 Names[RTLIB::FPTOSINT_F128_I64] = "__fixtfdi";
230 Names[RTLIB::FPTOSINT_F128_I128] = "__fixtfti";
231 Names[RTLIB::FPTOSINT_PPCF128_I32] = "__fixtfsi";
232 Names[RTLIB::FPTOSINT_PPCF128_I64] = "__fixtfdi";
233 Names[RTLIB::FPTOSINT_PPCF128_I128] = "__fixtfti";
234 Names[RTLIB::FPTOUINT_F32_I8] = "__fixunssfqi";
235 Names[RTLIB::FPTOUINT_F32_I16] = "__fixunssfhi";
236 Names[RTLIB::FPTOUINT_F32_I32] = "__fixunssfsi";
237 Names[RTLIB::FPTOUINT_F32_I64] = "__fixunssfdi";
238 Names[RTLIB::FPTOUINT_F32_I128] = "__fixunssfti";
239 Names[RTLIB::FPTOUINT_F64_I8] = "__fixunsdfqi";
240 Names[RTLIB::FPTOUINT_F64_I16] = "__fixunsdfhi";
241 Names[RTLIB::FPTOUINT_F64_I32] = "__fixunsdfsi";
242 Names[RTLIB::FPTOUINT_F64_I64] = "__fixunsdfdi";
243 Names[RTLIB::FPTOUINT_F64_I128] = "__fixunsdfti";
244 Names[RTLIB::FPTOUINT_F80_I32] = "__fixunsxfsi";
245 Names[RTLIB::FPTOUINT_F80_I64] = "__fixunsxfdi";
246 Names[RTLIB::FPTOUINT_F80_I128] = "__fixunsxfti";
247 Names[RTLIB::FPTOUINT_F128_I32] = "__fixunstfsi";
248 Names[RTLIB::FPTOUINT_F128_I64] = "__fixunstfdi";
249 Names[RTLIB::FPTOUINT_F128_I128] = "__fixunstfti";
250 Names[RTLIB::FPTOUINT_PPCF128_I32] = "__fixunstfsi";
251 Names[RTLIB::FPTOUINT_PPCF128_I64] = "__fixunstfdi";
252 Names[RTLIB::FPTOUINT_PPCF128_I128] = "__fixunstfti";
253 Names[RTLIB::SINTTOFP_I32_F32] = "__floatsisf";
254 Names[RTLIB::SINTTOFP_I32_F64] = "__floatsidf";
255 Names[RTLIB::SINTTOFP_I32_F80] = "__floatsixf";
256 Names[RTLIB::SINTTOFP_I32_F128] = "__floatsitf";
257 Names[RTLIB::SINTTOFP_I32_PPCF128] = "__floatsitf";
258 Names[RTLIB::SINTTOFP_I64_F32] = "__floatdisf";
259 Names[RTLIB::SINTTOFP_I64_F64] = "__floatdidf";
260 Names[RTLIB::SINTTOFP_I64_F80] = "__floatdixf";
261 Names[RTLIB::SINTTOFP_I64_F128] = "__floatditf";
262 Names[RTLIB::SINTTOFP_I64_PPCF128] = "__floatditf";
263 Names[RTLIB::SINTTOFP_I128_F32] = "__floattisf";
264 Names[RTLIB::SINTTOFP_I128_F64] = "__floattidf";
265 Names[RTLIB::SINTTOFP_I128_F80] = "__floattixf";
266 Names[RTLIB::SINTTOFP_I128_F128] = "__floattitf";
267 Names[RTLIB::SINTTOFP_I128_PPCF128] = "__floattitf";
268 Names[RTLIB::UINTTOFP_I32_F32] = "__floatunsisf";
269 Names[RTLIB::UINTTOFP_I32_F64] = "__floatunsidf";
270 Names[RTLIB::UINTTOFP_I32_F80] = "__floatunsixf";
271 Names[RTLIB::UINTTOFP_I32_F128] = "__floatunsitf";
272 Names[RTLIB::UINTTOFP_I32_PPCF128] = "__floatunsitf";
273 Names[RTLIB::UINTTOFP_I64_F32] = "__floatundisf";
274 Names[RTLIB::UINTTOFP_I64_F64] = "__floatundidf";
275 Names[RTLIB::UINTTOFP_I64_F80] = "__floatundixf";
276 Names[RTLIB::UINTTOFP_I64_F128] = "__floatunditf";
277 Names[RTLIB::UINTTOFP_I64_PPCF128] = "__floatunditf";
278 Names[RTLIB::UINTTOFP_I128_F32] = "__floatuntisf";
279 Names[RTLIB::UINTTOFP_I128_F64] = "__floatuntidf";
280 Names[RTLIB::UINTTOFP_I128_F80] = "__floatuntixf";
281 Names[RTLIB::UINTTOFP_I128_F128] = "__floatuntitf";
282 Names[RTLIB::UINTTOFP_I128_PPCF128] = "__floatuntitf";
283 Names[RTLIB::OEQ_F32] = "__eqsf2";
284 Names[RTLIB::OEQ_F64] = "__eqdf2";
285 Names[RTLIB::OEQ_F128] = "__eqtf2";
286 Names[RTLIB::UNE_F32] = "__nesf2";
287 Names[RTLIB::UNE_F64] = "__nedf2";
288 Names[RTLIB::UNE_F128] = "__netf2";
289 Names[RTLIB::OGE_F32] = "__gesf2";
290 Names[RTLIB::OGE_F64] = "__gedf2";
291 Names[RTLIB::OGE_F128] = "__getf2";
292 Names[RTLIB::OLT_F32] = "__ltsf2";
293 Names[RTLIB::OLT_F64] = "__ltdf2";
294 Names[RTLIB::OLT_F128] = "__lttf2";
295 Names[RTLIB::OLE_F32] = "__lesf2";
296 Names[RTLIB::OLE_F64] = "__ledf2";
297 Names[RTLIB::OLE_F128] = "__letf2";
298 Names[RTLIB::OGT_F32] = "__gtsf2";
299 Names[RTLIB::OGT_F64] = "__gtdf2";
300 Names[RTLIB::OGT_F128] = "__gttf2";
301 Names[RTLIB::UO_F32] = "__unordsf2";
302 Names[RTLIB::UO_F64] = "__unorddf2";
303 Names[RTLIB::UO_F128] = "__unordtf2";
304 Names[RTLIB::O_F32] = "__unordsf2";
305 Names[RTLIB::O_F64] = "__unorddf2";
306 Names[RTLIB::O_F128] = "__unordtf2";
307 Names[RTLIB::MEMCPY] = "memcpy";
308 Names[RTLIB::MEMMOVE] = "memmove";
309 Names[RTLIB::MEMSET] = "memset";
310 Names[RTLIB::UNWIND_RESUME] = "_Unwind_Resume";
311 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_1] = "__sync_val_compare_and_swap_1";
312 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_2] = "__sync_val_compare_and_swap_2";
313 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_4] = "__sync_val_compare_and_swap_4";
314 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_8] = "__sync_val_compare_and_swap_8";
315 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_1] = "__sync_lock_test_and_set_1";
316 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_2] = "__sync_lock_test_and_set_2";
317 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_4] = "__sync_lock_test_and_set_4";
318 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_8] = "__sync_lock_test_and_set_8";
319 Names[RTLIB::SYNC_FETCH_AND_ADD_1] = "__sync_fetch_and_add_1";
320 Names[RTLIB::SYNC_FETCH_AND_ADD_2] = "__sync_fetch_and_add_2";
321 Names[RTLIB::SYNC_FETCH_AND_ADD_4] = "__sync_fetch_and_add_4";
322 Names[RTLIB::SYNC_FETCH_AND_ADD_8] = "__sync_fetch_and_add_8";
323 Names[RTLIB::SYNC_FETCH_AND_SUB_1] = "__sync_fetch_and_sub_1";
324 Names[RTLIB::SYNC_FETCH_AND_SUB_2] = "__sync_fetch_and_sub_2";
325 Names[RTLIB::SYNC_FETCH_AND_SUB_4] = "__sync_fetch_and_sub_4";
326 Names[RTLIB::SYNC_FETCH_AND_SUB_8] = "__sync_fetch_and_sub_8";
327 Names[RTLIB::SYNC_FETCH_AND_AND_1] = "__sync_fetch_and_and_1";
328 Names[RTLIB::SYNC_FETCH_AND_AND_2] = "__sync_fetch_and_and_2";
329 Names[RTLIB::SYNC_FETCH_AND_AND_4] = "__sync_fetch_and_and_4";
330 Names[RTLIB::SYNC_FETCH_AND_AND_8] = "__sync_fetch_and_and_8";
331 Names[RTLIB::SYNC_FETCH_AND_OR_1] = "__sync_fetch_and_or_1";
332 Names[RTLIB::SYNC_FETCH_AND_OR_2] = "__sync_fetch_and_or_2";
333 Names[RTLIB::SYNC_FETCH_AND_OR_4] = "__sync_fetch_and_or_4";
334 Names[RTLIB::SYNC_FETCH_AND_OR_8] = "__sync_fetch_and_or_8";
335 Names[RTLIB::SYNC_FETCH_AND_XOR_1] = "__sync_fetch_and_xor_1";
336 Names[RTLIB::SYNC_FETCH_AND_XOR_2] = "__sync_fetch_and_xor_2";
337 Names[RTLIB::SYNC_FETCH_AND_XOR_4] = "__sync_fetch_and_xor_4";
338 Names[RTLIB::SYNC_FETCH_AND_XOR_8] = "__sync_fetch_and_xor_8";
339 Names[RTLIB::SYNC_FETCH_AND_NAND_1] = "__sync_fetch_and_nand_1";
340 Names[RTLIB::SYNC_FETCH_AND_NAND_2] = "__sync_fetch_and_nand_2";
341 Names[RTLIB::SYNC_FETCH_AND_NAND_4] = "__sync_fetch_and_nand_4";
342 Names[RTLIB::SYNC_FETCH_AND_NAND_8] = "__sync_fetch_and_nand_8";
344 // These are generally not available.
345 Names[RTLIB::SINCOS_F32] = 0;
346 Names[RTLIB::SINCOS_F64] = 0;
347 Names[RTLIB::SINCOS_F80] = 0;
348 Names[RTLIB::SINCOS_F128] = 0;
349 Names[RTLIB::SINCOS_PPCF128] = 0;
352 /// InitLibcallCallingConvs - Set default libcall CallingConvs.
354 static void InitLibcallCallingConvs(CallingConv::ID *CCs) {
355 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
356 CCs[i] = CallingConv::C;
360 /// getFPEXT - Return the FPEXT_*_* value for the given types, or
361 /// UNKNOWN_LIBCALL if there is none.
362 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) {
363 if (OpVT == MVT::f32) {
364 if (RetVT == MVT::f64)
365 return FPEXT_F32_F64;
366 if (RetVT == MVT::f128)
367 return FPEXT_F32_F128;
368 } else if (OpVT == MVT::f64) {
369 if (RetVT == MVT::f128)
370 return FPEXT_F64_F128;
373 return UNKNOWN_LIBCALL;
376 /// getFPROUND - Return the FPROUND_*_* value for the given types, or
377 /// UNKNOWN_LIBCALL if there is none.
378 RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) {
379 if (RetVT == MVT::f32) {
380 if (OpVT == MVT::f64)
381 return FPROUND_F64_F32;
382 if (OpVT == MVT::f80)
383 return FPROUND_F80_F32;
384 if (OpVT == MVT::f128)
385 return FPROUND_F128_F32;
386 if (OpVT == MVT::ppcf128)
387 return FPROUND_PPCF128_F32;
388 } else if (RetVT == MVT::f64) {
389 if (OpVT == MVT::f80)
390 return FPROUND_F80_F64;
391 if (OpVT == MVT::f128)
392 return FPROUND_F128_F64;
393 if (OpVT == MVT::ppcf128)
394 return FPROUND_PPCF128_F64;
397 return UNKNOWN_LIBCALL;
400 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or
401 /// UNKNOWN_LIBCALL if there is none.
402 RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) {
403 if (OpVT == MVT::f32) {
404 if (RetVT == MVT::i8)
405 return FPTOSINT_F32_I8;
406 if (RetVT == MVT::i16)
407 return FPTOSINT_F32_I16;
408 if (RetVT == MVT::i32)
409 return FPTOSINT_F32_I32;
410 if (RetVT == MVT::i64)
411 return FPTOSINT_F32_I64;
412 if (RetVT == MVT::i128)
413 return FPTOSINT_F32_I128;
414 } else if (OpVT == MVT::f64) {
415 if (RetVT == MVT::i8)
416 return FPTOSINT_F64_I8;
417 if (RetVT == MVT::i16)
418 return FPTOSINT_F64_I16;
419 if (RetVT == MVT::i32)
420 return FPTOSINT_F64_I32;
421 if (RetVT == MVT::i64)
422 return FPTOSINT_F64_I64;
423 if (RetVT == MVT::i128)
424 return FPTOSINT_F64_I128;
425 } else if (OpVT == MVT::f80) {
426 if (RetVT == MVT::i32)
427 return FPTOSINT_F80_I32;
428 if (RetVT == MVT::i64)
429 return FPTOSINT_F80_I64;
430 if (RetVT == MVT::i128)
431 return FPTOSINT_F80_I128;
432 } else if (OpVT == MVT::f128) {
433 if (RetVT == MVT::i32)
434 return FPTOSINT_F128_I32;
435 if (RetVT == MVT::i64)
436 return FPTOSINT_F128_I64;
437 if (RetVT == MVT::i128)
438 return FPTOSINT_F128_I128;
439 } else if (OpVT == MVT::ppcf128) {
440 if (RetVT == MVT::i32)
441 return FPTOSINT_PPCF128_I32;
442 if (RetVT == MVT::i64)
443 return FPTOSINT_PPCF128_I64;
444 if (RetVT == MVT::i128)
445 return FPTOSINT_PPCF128_I128;
447 return UNKNOWN_LIBCALL;
450 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or
451 /// UNKNOWN_LIBCALL if there is none.
452 RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) {
453 if (OpVT == MVT::f32) {
454 if (RetVT == MVT::i8)
455 return FPTOUINT_F32_I8;
456 if (RetVT == MVT::i16)
457 return FPTOUINT_F32_I16;
458 if (RetVT == MVT::i32)
459 return FPTOUINT_F32_I32;
460 if (RetVT == MVT::i64)
461 return FPTOUINT_F32_I64;
462 if (RetVT == MVT::i128)
463 return FPTOUINT_F32_I128;
464 } else if (OpVT == MVT::f64) {
465 if (RetVT == MVT::i8)
466 return FPTOUINT_F64_I8;
467 if (RetVT == MVT::i16)
468 return FPTOUINT_F64_I16;
469 if (RetVT == MVT::i32)
470 return FPTOUINT_F64_I32;
471 if (RetVT == MVT::i64)
472 return FPTOUINT_F64_I64;
473 if (RetVT == MVT::i128)
474 return FPTOUINT_F64_I128;
475 } else if (OpVT == MVT::f80) {
476 if (RetVT == MVT::i32)
477 return FPTOUINT_F80_I32;
478 if (RetVT == MVT::i64)
479 return FPTOUINT_F80_I64;
480 if (RetVT == MVT::i128)
481 return FPTOUINT_F80_I128;
482 } else if (OpVT == MVT::f128) {
483 if (RetVT == MVT::i32)
484 return FPTOUINT_F128_I32;
485 if (RetVT == MVT::i64)
486 return FPTOUINT_F128_I64;
487 if (RetVT == MVT::i128)
488 return FPTOUINT_F128_I128;
489 } else if (OpVT == MVT::ppcf128) {
490 if (RetVT == MVT::i32)
491 return FPTOUINT_PPCF128_I32;
492 if (RetVT == MVT::i64)
493 return FPTOUINT_PPCF128_I64;
494 if (RetVT == MVT::i128)
495 return FPTOUINT_PPCF128_I128;
497 return UNKNOWN_LIBCALL;
500 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or
501 /// UNKNOWN_LIBCALL if there is none.
502 RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) {
503 if (OpVT == MVT::i32) {
504 if (RetVT == MVT::f32)
505 return SINTTOFP_I32_F32;
506 if (RetVT == MVT::f64)
507 return SINTTOFP_I32_F64;
508 if (RetVT == MVT::f80)
509 return SINTTOFP_I32_F80;
510 if (RetVT == MVT::f128)
511 return SINTTOFP_I32_F128;
512 if (RetVT == MVT::ppcf128)
513 return SINTTOFP_I32_PPCF128;
514 } else if (OpVT == MVT::i64) {
515 if (RetVT == MVT::f32)
516 return SINTTOFP_I64_F32;
517 if (RetVT == MVT::f64)
518 return SINTTOFP_I64_F64;
519 if (RetVT == MVT::f80)
520 return SINTTOFP_I64_F80;
521 if (RetVT == MVT::f128)
522 return SINTTOFP_I64_F128;
523 if (RetVT == MVT::ppcf128)
524 return SINTTOFP_I64_PPCF128;
525 } else if (OpVT == MVT::i128) {
526 if (RetVT == MVT::f32)
527 return SINTTOFP_I128_F32;
528 if (RetVT == MVT::f64)
529 return SINTTOFP_I128_F64;
530 if (RetVT == MVT::f80)
531 return SINTTOFP_I128_F80;
532 if (RetVT == MVT::f128)
533 return SINTTOFP_I128_F128;
534 if (RetVT == MVT::ppcf128)
535 return SINTTOFP_I128_PPCF128;
537 return UNKNOWN_LIBCALL;
540 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or
541 /// UNKNOWN_LIBCALL if there is none.
542 RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) {
543 if (OpVT == MVT::i32) {
544 if (RetVT == MVT::f32)
545 return UINTTOFP_I32_F32;
546 if (RetVT == MVT::f64)
547 return UINTTOFP_I32_F64;
548 if (RetVT == MVT::f80)
549 return UINTTOFP_I32_F80;
550 if (RetVT == MVT::f128)
551 return UINTTOFP_I32_F128;
552 if (RetVT == MVT::ppcf128)
553 return UINTTOFP_I32_PPCF128;
554 } else if (OpVT == MVT::i64) {
555 if (RetVT == MVT::f32)
556 return UINTTOFP_I64_F32;
557 if (RetVT == MVT::f64)
558 return UINTTOFP_I64_F64;
559 if (RetVT == MVT::f80)
560 return UINTTOFP_I64_F80;
561 if (RetVT == MVT::f128)
562 return UINTTOFP_I64_F128;
563 if (RetVT == MVT::ppcf128)
564 return UINTTOFP_I64_PPCF128;
565 } else if (OpVT == MVT::i128) {
566 if (RetVT == MVT::f32)
567 return UINTTOFP_I128_F32;
568 if (RetVT == MVT::f64)
569 return UINTTOFP_I128_F64;
570 if (RetVT == MVT::f80)
571 return UINTTOFP_I128_F80;
572 if (RetVT == MVT::f128)
573 return UINTTOFP_I128_F128;
574 if (RetVT == MVT::ppcf128)
575 return UINTTOFP_I128_PPCF128;
577 return UNKNOWN_LIBCALL;
580 /// InitCmpLibcallCCs - Set default comparison libcall CC.
582 static void InitCmpLibcallCCs(ISD::CondCode *CCs) {
583 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL);
584 CCs[RTLIB::OEQ_F32] = ISD::SETEQ;
585 CCs[RTLIB::OEQ_F64] = ISD::SETEQ;
586 CCs[RTLIB::OEQ_F128] = ISD::SETEQ;
587 CCs[RTLIB::UNE_F32] = ISD::SETNE;
588 CCs[RTLIB::UNE_F64] = ISD::SETNE;
589 CCs[RTLIB::UNE_F128] = ISD::SETNE;
590 CCs[RTLIB::OGE_F32] = ISD::SETGE;
591 CCs[RTLIB::OGE_F64] = ISD::SETGE;
592 CCs[RTLIB::OGE_F128] = ISD::SETGE;
593 CCs[RTLIB::OLT_F32] = ISD::SETLT;
594 CCs[RTLIB::OLT_F64] = ISD::SETLT;
595 CCs[RTLIB::OLT_F128] = ISD::SETLT;
596 CCs[RTLIB::OLE_F32] = ISD::SETLE;
597 CCs[RTLIB::OLE_F64] = ISD::SETLE;
598 CCs[RTLIB::OLE_F128] = ISD::SETLE;
599 CCs[RTLIB::OGT_F32] = ISD::SETGT;
600 CCs[RTLIB::OGT_F64] = ISD::SETGT;
601 CCs[RTLIB::OGT_F128] = ISD::SETGT;
602 CCs[RTLIB::UO_F32] = ISD::SETNE;
603 CCs[RTLIB::UO_F64] = ISD::SETNE;
604 CCs[RTLIB::UO_F128] = ISD::SETNE;
605 CCs[RTLIB::O_F32] = ISD::SETEQ;
606 CCs[RTLIB::O_F64] = ISD::SETEQ;
607 CCs[RTLIB::O_F128] = ISD::SETEQ;
610 /// NOTE: The constructor takes ownership of TLOF.
611 TargetLoweringBase::TargetLoweringBase(const TargetMachine &tm,
612 const TargetLoweringObjectFile *tlof)
613 : TM(tm), TD(TM.getDataLayout()), TLOF(*tlof) {
614 // All operations default to being supported.
615 memset(OpActions, 0, sizeof(OpActions));
616 memset(LoadExtActions, 0, sizeof(LoadExtActions));
617 memset(TruncStoreActions, 0, sizeof(TruncStoreActions));
618 memset(IndexedModeActions, 0, sizeof(IndexedModeActions));
619 memset(CondCodeActions, 0, sizeof(CondCodeActions));
621 // Set default actions for various operations.
622 for (unsigned VT = 0; VT != (unsigned)MVT::LAST_VALUETYPE; ++VT) {
623 // Default all indexed load / store to expand.
624 for (unsigned IM = (unsigned)ISD::PRE_INC;
625 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
626 setIndexedLoadAction(IM, (MVT::SimpleValueType)VT, Expand);
627 setIndexedStoreAction(IM, (MVT::SimpleValueType)VT, Expand);
630 // These operations default to expand.
631 setOperationAction(ISD::FGETSIGN, (MVT::SimpleValueType)VT, Expand);
632 setOperationAction(ISD::CONCAT_VECTORS, (MVT::SimpleValueType)VT, Expand);
635 // Most targets ignore the @llvm.prefetch intrinsic.
636 setOperationAction(ISD::PREFETCH, MVT::Other, Expand);
638 // ConstantFP nodes default to expand. Targets can either change this to
639 // Legal, in which case all fp constants are legal, or use isFPImmLegal()
640 // to optimize expansions for certain constants.
641 setOperationAction(ISD::ConstantFP, MVT::f16, Expand);
642 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
643 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
644 setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
645 setOperationAction(ISD::ConstantFP, MVT::f128, Expand);
647 // These library functions default to expand.
648 setOperationAction(ISD::FLOG , MVT::f16, Expand);
649 setOperationAction(ISD::FLOG2, MVT::f16, Expand);
650 setOperationAction(ISD::FLOG10, MVT::f16, Expand);
651 setOperationAction(ISD::FEXP , MVT::f16, Expand);
652 setOperationAction(ISD::FEXP2, MVT::f16, Expand);
653 setOperationAction(ISD::FFLOOR, MVT::f16, Expand);
654 setOperationAction(ISD::FNEARBYINT, MVT::f16, Expand);
655 setOperationAction(ISD::FCEIL, MVT::f16, Expand);
656 setOperationAction(ISD::FRINT, MVT::f16, Expand);
657 setOperationAction(ISD::FTRUNC, MVT::f16, Expand);
658 setOperationAction(ISD::FLOG , MVT::f32, Expand);
659 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
660 setOperationAction(ISD::FLOG10, MVT::f32, Expand);
661 setOperationAction(ISD::FEXP , MVT::f32, Expand);
662 setOperationAction(ISD::FEXP2, MVT::f32, Expand);
663 setOperationAction(ISD::FFLOOR, MVT::f32, Expand);
664 setOperationAction(ISD::FNEARBYINT, MVT::f32, Expand);
665 setOperationAction(ISD::FCEIL, MVT::f32, Expand);
666 setOperationAction(ISD::FRINT, MVT::f32, Expand);
667 setOperationAction(ISD::FTRUNC, MVT::f32, Expand);
668 setOperationAction(ISD::FLOG , MVT::f64, Expand);
669 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
670 setOperationAction(ISD::FLOG10, MVT::f64, Expand);
671 setOperationAction(ISD::FEXP , MVT::f64, Expand);
672 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
673 setOperationAction(ISD::FFLOOR, MVT::f64, Expand);
674 setOperationAction(ISD::FNEARBYINT, MVT::f64, Expand);
675 setOperationAction(ISD::FCEIL, MVT::f64, Expand);
676 setOperationAction(ISD::FRINT, MVT::f64, Expand);
677 setOperationAction(ISD::FTRUNC, MVT::f64, Expand);
678 setOperationAction(ISD::FLOG , MVT::f128, Expand);
679 setOperationAction(ISD::FLOG2, MVT::f128, Expand);
680 setOperationAction(ISD::FLOG10, MVT::f128, Expand);
681 setOperationAction(ISD::FEXP , MVT::f128, Expand);
682 setOperationAction(ISD::FEXP2, MVT::f128, Expand);
683 setOperationAction(ISD::FFLOOR, MVT::f128, Expand);
684 setOperationAction(ISD::FNEARBYINT, MVT::f128, Expand);
685 setOperationAction(ISD::FCEIL, MVT::f128, Expand);
686 setOperationAction(ISD::FRINT, MVT::f128, Expand);
687 setOperationAction(ISD::FTRUNC, MVT::f128, Expand);
689 // Default ISD::TRAP to expand (which turns it into abort).
690 setOperationAction(ISD::TRAP, MVT::Other, Expand);
692 // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand"
693 // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP.
695 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Expand);
697 IsLittleEndian = TD->isLittleEndian();
698 PointerTy = MVT::getIntegerVT(8*TD->getPointerSize(0));
699 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE*sizeof(TargetRegisterClass*));
700 memset(TargetDAGCombineArray, 0, array_lengthof(TargetDAGCombineArray));
701 maxStoresPerMemset = maxStoresPerMemcpy = maxStoresPerMemmove = 8;
702 maxStoresPerMemsetOptSize = maxStoresPerMemcpyOptSize
703 = maxStoresPerMemmoveOptSize = 4;
704 benefitFromCodePlacementOpt = false;
705 UseUnderscoreSetJmp = false;
706 UseUnderscoreLongJmp = false;
707 SelectIsExpensive = false;
708 IntDivIsCheap = false;
709 Pow2DivIsCheap = false;
710 JumpIsExpensive = false;
711 predictableSelectIsExpensive = false;
712 StackPointerRegisterToSaveRestore = 0;
713 ExceptionPointerRegister = 0;
714 ExceptionSelectorRegister = 0;
715 BooleanContents = UndefinedBooleanContent;
716 BooleanVectorContents = UndefinedBooleanContent;
717 SchedPreferenceInfo = Sched::ILP;
719 JumpBufAlignment = 0;
720 MinFunctionAlignment = 0;
721 PrefFunctionAlignment = 0;
722 PrefLoopAlignment = 0;
723 MinStackArgumentAlignment = 1;
724 ShouldFoldAtomicFences = false;
725 InsertFencesForAtomic = false;
726 SupportJumpTables = true;
727 MinimumJumpTableEntries = 4;
729 InitLibcallNames(LibcallRoutineNames);
730 InitCmpLibcallCCs(CmpLibcallCCs);
731 InitLibcallCallingConvs(LibcallCallingConvs);
734 TargetLoweringBase::~TargetLoweringBase() {
738 MVT TargetLoweringBase::getShiftAmountTy(EVT LHSTy) const {
739 return MVT::getIntegerVT(8*TD->getPointerSize(0));
742 /// canOpTrap - Returns true if the operation can trap for the value type.
743 /// VT must be a legal type.
744 bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const {
745 assert(isTypeLegal(VT));
760 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT,
761 unsigned &NumIntermediates,
763 TargetLoweringBase *TLI) {
764 // Figure out the right, legal destination reg to copy into.
765 unsigned NumElts = VT.getVectorNumElements();
766 MVT EltTy = VT.getVectorElementType();
768 unsigned NumVectorRegs = 1;
770 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
771 // could break down into LHS/RHS like LegalizeDAG does.
772 if (!isPowerOf2_32(NumElts)) {
773 NumVectorRegs = NumElts;
777 // Divide the input until we get to a supported size. This will always
778 // end with a scalar if the target doesn't support vectors.
779 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) {
784 NumIntermediates = NumVectorRegs;
786 MVT NewVT = MVT::getVectorVT(EltTy, NumElts);
787 if (!TLI->isTypeLegal(NewVT))
789 IntermediateVT = NewVT;
791 unsigned NewVTSize = NewVT.getSizeInBits();
793 // Convert sizes such as i33 to i64.
794 if (!isPowerOf2_32(NewVTSize))
795 NewVTSize = NextPowerOf2(NewVTSize);
797 MVT DestVT = TLI->getRegisterType(NewVT);
799 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16.
800 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits());
802 // Otherwise, promotion or legal types use the same number of registers as
803 // the vector decimated to the appropriate level.
804 return NumVectorRegs;
807 /// isLegalRC - Return true if the value types that can be represented by the
808 /// specified register class are all legal.
809 bool TargetLoweringBase::isLegalRC(const TargetRegisterClass *RC) const {
810 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
818 /// findRepresentativeClass - Return the largest legal super-reg register class
819 /// of the register class for the specified type and its associated "cost".
820 std::pair<const TargetRegisterClass*, uint8_t>
821 TargetLoweringBase::findRepresentativeClass(MVT VT) const {
822 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
823 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy];
825 return std::make_pair(RC, 0);
827 // Compute the set of all super-register classes.
828 BitVector SuperRegRC(TRI->getNumRegClasses());
829 for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI)
830 SuperRegRC.setBitsInMask(RCI.getMask());
832 // Find the first legal register class with the largest spill size.
833 const TargetRegisterClass *BestRC = RC;
834 for (int i = SuperRegRC.find_first(); i >= 0; i = SuperRegRC.find_next(i)) {
835 const TargetRegisterClass *SuperRC = TRI->getRegClass(i);
836 // We want the largest possible spill size.
837 if (SuperRC->getSize() <= BestRC->getSize())
839 if (!isLegalRC(SuperRC))
843 return std::make_pair(BestRC, 1);
846 /// computeRegisterProperties - Once all of the register classes are added,
847 /// this allows us to compute derived properties we expose.
848 void TargetLoweringBase::computeRegisterProperties() {
849 assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE &&
850 "Too many value types for ValueTypeActions to hold!");
852 // Everything defaults to needing one register.
853 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
854 NumRegistersForVT[i] = 1;
855 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i;
857 // ...except isVoid, which doesn't need any registers.
858 NumRegistersForVT[MVT::isVoid] = 0;
860 // Find the largest integer register class.
861 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE;
862 for (; RegClassForVT[LargestIntReg] == 0; --LargestIntReg)
863 assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
865 // Every integer value type larger than this largest register takes twice as
866 // many registers to represent as the previous ValueType.
867 for (unsigned ExpandedReg = LargestIntReg + 1;
868 ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) {
869 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
870 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg;
871 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1);
872 ValueTypeActions.setTypeAction((MVT::SimpleValueType)ExpandedReg,
876 // Inspect all of the ValueType's smaller than the largest integer
877 // register to see which ones need promotion.
878 unsigned LegalIntReg = LargestIntReg;
879 for (unsigned IntReg = LargestIntReg - 1;
880 IntReg >= (unsigned)MVT::i1; --IntReg) {
881 MVT IVT = (MVT::SimpleValueType)IntReg;
882 if (isTypeLegal(IVT)) {
883 LegalIntReg = IntReg;
885 RegisterTypeForVT[IntReg] = TransformToType[IntReg] =
886 (const MVT::SimpleValueType)LegalIntReg;
887 ValueTypeActions.setTypeAction(IVT, TypePromoteInteger);
891 // ppcf128 type is really two f64's.
892 if (!isTypeLegal(MVT::ppcf128)) {
893 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64];
894 RegisterTypeForVT[MVT::ppcf128] = MVT::f64;
895 TransformToType[MVT::ppcf128] = MVT::f64;
896 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat);
899 // Decide how to handle f64. If the target does not have native f64 support,
900 // expand it to i64 and we will be generating soft float library calls.
901 if (!isTypeLegal(MVT::f64)) {
902 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
903 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
904 TransformToType[MVT::f64] = MVT::i64;
905 ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat);
908 // Decide how to handle f32. If the target does not have native support for
909 // f32, promote it to f64 if it is legal. Otherwise, expand it to i32.
910 if (!isTypeLegal(MVT::f32)) {
911 if (isTypeLegal(MVT::f64)) {
912 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::f64];
913 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::f64];
914 TransformToType[MVT::f32] = MVT::f64;
915 ValueTypeActions.setTypeAction(MVT::f32, TypePromoteInteger);
917 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
918 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
919 TransformToType[MVT::f32] = MVT::i32;
920 ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat);
924 // Loop over all of the vector value types to see which need transformations.
925 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
926 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
927 MVT VT = (MVT::SimpleValueType)i;
928 if (isTypeLegal(VT)) continue;
930 // Determine if there is a legal wider type. If so, we should promote to
931 // that wider vector type.
932 MVT EltVT = VT.getVectorElementType();
933 unsigned NElts = VT.getVectorNumElements();
934 if (NElts != 1 && !shouldSplitVectorElementType(EltVT)) {
935 bool IsLegalWiderType = false;
936 // First try to promote the elements of integer vectors. If no legal
937 // promotion was found, fallback to the widen-vector method.
938 for (unsigned nVT = i+1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
939 MVT SVT = (MVT::SimpleValueType)nVT;
940 // Promote vectors of integers to vectors with the same number
941 // of elements, with a wider element type.
942 if (SVT.getVectorElementType().getSizeInBits() > EltVT.getSizeInBits()
943 && SVT.getVectorNumElements() == NElts &&
944 isTypeLegal(SVT) && SVT.getScalarType().isInteger()) {
945 TransformToType[i] = SVT;
946 RegisterTypeForVT[i] = SVT;
947 NumRegistersForVT[i] = 1;
948 ValueTypeActions.setTypeAction(VT, TypePromoteInteger);
949 IsLegalWiderType = true;
954 if (IsLegalWiderType) continue;
956 // Try to widen the vector.
957 for (unsigned nVT = i+1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
958 MVT SVT = (MVT::SimpleValueType)nVT;
959 if (SVT.getVectorElementType() == EltVT &&
960 SVT.getVectorNumElements() > NElts &&
962 TransformToType[i] = SVT;
963 RegisterTypeForVT[i] = SVT;
964 NumRegistersForVT[i] = 1;
965 ValueTypeActions.setTypeAction(VT, TypeWidenVector);
966 IsLegalWiderType = true;
970 if (IsLegalWiderType) continue;
975 unsigned NumIntermediates;
976 NumRegistersForVT[i] =
977 getVectorTypeBreakdownMVT(VT, IntermediateVT, NumIntermediates,
979 RegisterTypeForVT[i] = RegisterVT;
981 MVT NVT = VT.getPow2VectorType();
983 // Type is already a power of 2. The default action is to split.
984 TransformToType[i] = MVT::Other;
985 unsigned NumElts = VT.getVectorNumElements();
986 ValueTypeActions.setTypeAction(VT,
987 NumElts > 1 ? TypeSplitVector : TypeScalarizeVector);
989 TransformToType[i] = NVT;
990 ValueTypeActions.setTypeAction(VT, TypeWidenVector);
994 // Determine the 'representative' register class for each value type.
995 // An representative register class is the largest (meaning one which is
996 // not a sub-register class / subreg register class) legal register class for
997 // a group of value types. For example, on i386, i8, i16, and i32
998 // representative would be GR32; while on x86_64 it's GR64.
999 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
1000 const TargetRegisterClass* RRC;
1002 tie(RRC, Cost) = findRepresentativeClass((MVT::SimpleValueType)i);
1003 RepRegClassForVT[i] = RRC;
1004 RepRegClassCostForVT[i] = Cost;
1008 EVT TargetLoweringBase::getSetCCResultType(EVT VT) const {
1009 assert(!VT.isVector() && "No default SetCC type for vectors!");
1010 return getPointerTy(0).SimpleTy;
1013 MVT::SimpleValueType TargetLoweringBase::getCmpLibcallReturnType() const {
1014 return MVT::i32; // return the default value
1017 /// getVectorTypeBreakdown - Vector types are broken down into some number of
1018 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
1019 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
1020 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
1022 /// This method returns the number of registers needed, and the VT for each
1023 /// register. It also returns the VT and quantity of the intermediate values
1024 /// before they are promoted/expanded.
1026 unsigned TargetLoweringBase::getVectorTypeBreakdown(LLVMContext &Context, EVT VT,
1027 EVT &IntermediateVT,
1028 unsigned &NumIntermediates,
1029 MVT &RegisterVT) const {
1030 unsigned NumElts = VT.getVectorNumElements();
1032 // If there is a wider vector type with the same element type as this one,
1033 // or a promoted vector type that has the same number of elements which
1034 // are wider, then we should convert to that legal vector type.
1035 // This handles things like <2 x float> -> <4 x float> and
1036 // <4 x i1> -> <4 x i32>.
1037 LegalizeTypeAction TA = getTypeAction(Context, VT);
1038 if (NumElts != 1 && (TA == TypeWidenVector || TA == TypePromoteInteger)) {
1039 EVT RegisterEVT = getTypeToTransformTo(Context, VT);
1040 if (isTypeLegal(RegisterEVT)) {
1041 IntermediateVT = RegisterEVT;
1042 RegisterVT = RegisterEVT.getSimpleVT();
1043 NumIntermediates = 1;
1048 // Figure out the right, legal destination reg to copy into.
1049 EVT EltTy = VT.getVectorElementType();
1051 unsigned NumVectorRegs = 1;
1053 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
1054 // could break down into LHS/RHS like LegalizeDAG does.
1055 if (!isPowerOf2_32(NumElts)) {
1056 NumVectorRegs = NumElts;
1060 // Divide the input until we get to a supported size. This will always
1061 // end with a scalar if the target doesn't support vectors.
1062 while (NumElts > 1 && !isTypeLegal(
1063 EVT::getVectorVT(Context, EltTy, NumElts))) {
1065 NumVectorRegs <<= 1;
1068 NumIntermediates = NumVectorRegs;
1070 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts);
1071 if (!isTypeLegal(NewVT))
1073 IntermediateVT = NewVT;
1075 MVT DestVT = getRegisterType(Context, NewVT);
1076 RegisterVT = DestVT;
1077 unsigned NewVTSize = NewVT.getSizeInBits();
1079 // Convert sizes such as i33 to i64.
1080 if (!isPowerOf2_32(NewVTSize))
1081 NewVTSize = NextPowerOf2(NewVTSize);
1083 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16.
1084 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits());
1086 // Otherwise, promotion or legal types use the same number of registers as
1087 // the vector decimated to the appropriate level.
1088 return NumVectorRegs;
1091 /// Get the EVTs and ArgFlags collections that represent the legalized return
1092 /// type of the given function. This does not require a DAG or a return value,
1093 /// and is suitable for use before any DAGs for the function are constructed.
1094 /// TODO: Move this out of TargetLowering.cpp.
1095 void llvm::GetReturnInfo(Type* ReturnType, AttributeSet attr,
1096 SmallVectorImpl<ISD::OutputArg> &Outs,
1097 const TargetLowering &TLI) {
1098 SmallVector<EVT, 4> ValueVTs;
1099 ComputeValueVTs(TLI, ReturnType, ValueVTs);
1100 unsigned NumValues = ValueVTs.size();
1101 if (NumValues == 0) return;
1103 for (unsigned j = 0, f = NumValues; j != f; ++j) {
1104 EVT VT = ValueVTs[j];
1105 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
1107 if (attr.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
1108 ExtendKind = ISD::SIGN_EXTEND;
1109 else if (attr.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt))
1110 ExtendKind = ISD::ZERO_EXTEND;
1112 // FIXME: C calling convention requires the return type to be promoted to
1113 // at least 32-bit. But this is not necessary for non-C calling
1114 // conventions. The frontend should mark functions whose return values
1115 // require promoting with signext or zeroext attributes.
1116 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) {
1117 MVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32);
1118 if (VT.bitsLT(MinVT))
1122 unsigned NumParts = TLI.getNumRegisters(ReturnType->getContext(), VT);
1123 MVT PartVT = TLI.getRegisterType(ReturnType->getContext(), VT);
1125 // 'inreg' on function refers to return value
1126 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1127 if (attr.hasAttribute(AttributeSet::ReturnIndex, Attribute::InReg))
1130 // Propagate extension type if any
1131 if (attr.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
1133 else if (attr.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt))
1136 for (unsigned i = 0; i < NumParts; ++i)
1137 Outs.push_back(ISD::OutputArg(Flags, PartVT, /*isFixed=*/true, 0, 0));
1141 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1142 /// function arguments in the caller parameter area. This is the actual
1143 /// alignment, not its logarithm.
1144 unsigned TargetLoweringBase::getByValTypeAlignment(Type *Ty) const {
1145 return TD->getCallFrameTypeAlignment(Ty);
1148 //===----------------------------------------------------------------------===//
1149 // TargetTransformInfo Helpers
1150 //===----------------------------------------------------------------------===//
1152 int TargetLoweringBase::InstructionOpcodeToISD(unsigned Opcode) const {
1153 enum InstructionOpcodes {
1154 #define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM,
1155 #define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM
1156 #include "llvm/IR/Instruction.def"
1158 switch (static_cast<InstructionOpcodes>(Opcode)) {
1161 case Switch: return 0;
1162 case IndirectBr: return 0;
1163 case Invoke: return 0;
1164 case Resume: return 0;
1165 case Unreachable: return 0;
1166 case Add: return ISD::ADD;
1167 case FAdd: return ISD::FADD;
1168 case Sub: return ISD::SUB;
1169 case FSub: return ISD::FSUB;
1170 case Mul: return ISD::MUL;
1171 case FMul: return ISD::FMUL;
1172 case UDiv: return ISD::UDIV;
1173 case SDiv: return ISD::UDIV;
1174 case FDiv: return ISD::FDIV;
1175 case URem: return ISD::UREM;
1176 case SRem: return ISD::SREM;
1177 case FRem: return ISD::FREM;
1178 case Shl: return ISD::SHL;
1179 case LShr: return ISD::SRL;
1180 case AShr: return ISD::SRA;
1181 case And: return ISD::AND;
1182 case Or: return ISD::OR;
1183 case Xor: return ISD::XOR;
1184 case Alloca: return 0;
1185 case Load: return ISD::LOAD;
1186 case Store: return ISD::STORE;
1187 case GetElementPtr: return 0;
1188 case Fence: return 0;
1189 case AtomicCmpXchg: return 0;
1190 case AtomicRMW: return 0;
1191 case Trunc: return ISD::TRUNCATE;
1192 case ZExt: return ISD::ZERO_EXTEND;
1193 case SExt: return ISD::SIGN_EXTEND;
1194 case FPToUI: return ISD::FP_TO_UINT;
1195 case FPToSI: return ISD::FP_TO_SINT;
1196 case UIToFP: return ISD::UINT_TO_FP;
1197 case SIToFP: return ISD::SINT_TO_FP;
1198 case FPTrunc: return ISD::FP_ROUND;
1199 case FPExt: return ISD::FP_EXTEND;
1200 case PtrToInt: return ISD::BITCAST;
1201 case IntToPtr: return ISD::BITCAST;
1202 case BitCast: return ISD::BITCAST;
1203 case ICmp: return ISD::SETCC;
1204 case FCmp: return ISD::SETCC;
1206 case Call: return 0;
1207 case Select: return ISD::SELECT;
1208 case UserOp1: return 0;
1209 case UserOp2: return 0;
1210 case VAArg: return 0;
1211 case ExtractElement: return ISD::EXTRACT_VECTOR_ELT;
1212 case InsertElement: return ISD::INSERT_VECTOR_ELT;
1213 case ShuffleVector: return ISD::VECTOR_SHUFFLE;
1214 case ExtractValue: return ISD::MERGE_VALUES;
1215 case InsertValue: return ISD::MERGE_VALUES;
1216 case LandingPad: return 0;
1219 llvm_unreachable("Unknown instruction type encountered!");
1222 std::pair<unsigned, MVT>
1223 TargetLoweringBase::getTypeLegalizationCost(Type *Ty) const {
1224 LLVMContext &C = Ty->getContext();
1225 EVT MTy = getValueType(Ty);
1228 // We keep legalizing the type until we find a legal kind. We assume that
1229 // the only operation that costs anything is the split. After splitting
1230 // we need to handle two types.
1232 LegalizeKind LK = getTypeConversion(C, MTy);
1234 if (LK.first == TypeLegal)
1235 return std::make_pair(Cost, MTy.getSimpleVT());
1237 if (LK.first == TypeSplitVector || LK.first == TypeExpandInteger)
1240 // Keep legalizing the type.
1245 //===----------------------------------------------------------------------===//
1246 // Loop Strength Reduction hooks
1247 //===----------------------------------------------------------------------===//
1249 /// isLegalAddressingMode - Return true if the addressing mode represented
1250 /// by AM is legal for this target, for a load/store of the specified type.
1251 bool TargetLoweringBase::isLegalAddressingMode(const AddrMode &AM,
1253 // The default implementation of this implements a conservative RISCy, r+r and
1256 // Allows a sign-extended 16-bit immediate field.
1257 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
1260 // No global is ever allowed as a base.
1264 // Only support r+r,
1266 case 0: // "r+i" or just "i", depending on HasBaseReg.
1269 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
1271 // Otherwise we have r+r or r+i.
1274 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
1276 // Allow 2*r as r+r.