1 //===-- TargetLoweringBase.cpp - Implement the TargetLoweringBase class ---===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This implements the TargetLoweringBase class.
11 //
12 //===----------------------------------------------------------------------===//
13
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/MachineInstrBuilder.h"
22 #include "llvm/CodeGen/MachineJumpTableInfo.h"
23 #include "llvm/CodeGen/StackMaps.h"
24 #include "llvm/IR/DataLayout.h"
25 #include "llvm/IR/DerivedTypes.h"
26 #include "llvm/IR/GlobalVariable.h"
27 #include "llvm/IR/Mangler.h"
28 #include "llvm/MC/MCAsmInfo.h"
29 #include "llvm/MC/MCContext.h"
30 #include "llvm/MC/MCExpr.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/MathExtras.h"
34 #include "llvm/Target/TargetLoweringObjectFile.h"
35 #include "llvm/Target/TargetMachine.h"
36 #include "llvm/Target/TargetRegisterInfo.h"
37 #include "llvm/Target/TargetSubtargetInfo.h"
38 #include <cctype>
39 using namespace llvm;
40
41 /// InitLibcallNames - Set default libcall names.
42 ///
InitLibcallNames(const char ** Names,const Triple & TT)43 static void InitLibcallNames(const char **Names, const Triple &TT) {
44 Names[RTLIB::SHL_I16] = "__ashlhi3";
45 Names[RTLIB::SHL_I32] = "__ashlsi3";
46 Names[RTLIB::SHL_I64] = "__ashldi3";
47 Names[RTLIB::SHL_I128] = "__ashlti3";
48 Names[RTLIB::SRL_I16] = "__lshrhi3";
49 Names[RTLIB::SRL_I32] = "__lshrsi3";
50 Names[RTLIB::SRL_I64] = "__lshrdi3";
51 Names[RTLIB::SRL_I128] = "__lshrti3";
52 Names[RTLIB::SRA_I16] = "__ashrhi3";
53 Names[RTLIB::SRA_I32] = "__ashrsi3";
54 Names[RTLIB::SRA_I64] = "__ashrdi3";
55 Names[RTLIB::SRA_I128] = "__ashrti3";
56 Names[RTLIB::MUL_I8] = "__mulqi3";
57 Names[RTLIB::MUL_I16] = "__mulhi3";
58 Names[RTLIB::MUL_I32] = "__mulsi3";
59 Names[RTLIB::MUL_I64] = "__muldi3";
60 Names[RTLIB::MUL_I128] = "__multi3";
61 Names[RTLIB::MULO_I32] = "__mulosi4";
62 Names[RTLIB::MULO_I64] = "__mulodi4";
63 Names[RTLIB::MULO_I128] = "__muloti4";
64 Names[RTLIB::SDIV_I8] = "__divqi3";
65 Names[RTLIB::SDIV_I16] = "__divhi3";
66 Names[RTLIB::SDIV_I32] = "__divsi3";
67 Names[RTLIB::SDIV_I64] = "__divdi3";
68 Names[RTLIB::SDIV_I128] = "__divti3";
69 Names[RTLIB::UDIV_I8] = "__udivqi3";
70 Names[RTLIB::UDIV_I16] = "__udivhi3";
71 Names[RTLIB::UDIV_I32] = "__udivsi3";
72 Names[RTLIB::UDIV_I64] = "__udivdi3";
73 Names[RTLIB::UDIV_I128] = "__udivti3";
74 Names[RTLIB::SREM_I8] = "__modqi3";
75 Names[RTLIB::SREM_I16] = "__modhi3";
76 Names[RTLIB::SREM_I32] = "__modsi3";
77 Names[RTLIB::SREM_I64] = "__moddi3";
78 Names[RTLIB::SREM_I128] = "__modti3";
79 Names[RTLIB::UREM_I8] = "__umodqi3";
80 Names[RTLIB::UREM_I16] = "__umodhi3";
81 Names[RTLIB::UREM_I32] = "__umodsi3";
82 Names[RTLIB::UREM_I64] = "__umoddi3";
83 Names[RTLIB::UREM_I128] = "__umodti3";
84
85 // These are generally not available.
86 Names[RTLIB::SDIVREM_I8] = nullptr;
87 Names[RTLIB::SDIVREM_I16] = nullptr;
88 Names[RTLIB::SDIVREM_I32] = nullptr;
89 Names[RTLIB::SDIVREM_I64] = nullptr;
90 Names[RTLIB::SDIVREM_I128] = nullptr;
91 Names[RTLIB::UDIVREM_I8] = nullptr;
92 Names[RTLIB::UDIVREM_I16] = nullptr;
93 Names[RTLIB::UDIVREM_I32] = nullptr;
94 Names[RTLIB::UDIVREM_I64] = nullptr;
95 Names[RTLIB::UDIVREM_I128] = nullptr;
96
97 Names[RTLIB::NEG_I32] = "__negsi2";
98 Names[RTLIB::NEG_I64] = "__negdi2";
99 Names[RTLIB::ADD_F32] = "__addsf3";
100 Names[RTLIB::ADD_F64] = "__adddf3";
101 Names[RTLIB::ADD_F80] = "__addxf3";
102 Names[RTLIB::ADD_F128] = "__addtf3";
103 Names[RTLIB::ADD_PPCF128] = "__gcc_qadd";
104 Names[RTLIB::SUB_F32] = "__subsf3";
105 Names[RTLIB::SUB_F64] = "__subdf3";
106 Names[RTLIB::SUB_F80] = "__subxf3";
107 Names[RTLIB::SUB_F128] = "__subtf3";
108 Names[RTLIB::SUB_PPCF128] = "__gcc_qsub";
109 Names[RTLIB::MUL_F32] = "__mulsf3";
110 Names[RTLIB::MUL_F64] = "__muldf3";
111 Names[RTLIB::MUL_F80] = "__mulxf3";
112 Names[RTLIB::MUL_F128] = "__multf3";
113 Names[RTLIB::MUL_PPCF128] = "__gcc_qmul";
114 Names[RTLIB::DIV_F32] = "__divsf3";
115 Names[RTLIB::DIV_F64] = "__divdf3";
116 Names[RTLIB::DIV_F80] = "__divxf3";
117 Names[RTLIB::DIV_F128] = "__divtf3";
118 Names[RTLIB::DIV_PPCF128] = "__gcc_qdiv";
119 Names[RTLIB::REM_F32] = "fmodf";
120 Names[RTLIB::REM_F64] = "fmod";
121 Names[RTLIB::REM_F80] = "fmodl";
122 Names[RTLIB::REM_F128] = "fmodl";
123 Names[RTLIB::REM_PPCF128] = "fmodl";
124 Names[RTLIB::FMA_F32] = "fmaf";
125 Names[RTLIB::FMA_F64] = "fma";
126 Names[RTLIB::FMA_F80] = "fmal";
127 Names[RTLIB::FMA_F128] = "fmal";
128 Names[RTLIB::FMA_PPCF128] = "fmal";
129 Names[RTLIB::POWI_F32] = "__powisf2";
130 Names[RTLIB::POWI_F64] = "__powidf2";
131 Names[RTLIB::POWI_F80] = "__powixf2";
132 Names[RTLIB::POWI_F128] = "__powitf2";
133 Names[RTLIB::POWI_PPCF128] = "__powitf2";
134 Names[RTLIB::SQRT_F32] = "sqrtf";
135 Names[RTLIB::SQRT_F64] = "sqrt";
136 Names[RTLIB::SQRT_F80] = "sqrtl";
137 Names[RTLIB::SQRT_F128] = "sqrtl";
138 Names[RTLIB::SQRT_PPCF128] = "sqrtl";
139 Names[RTLIB::LOG_F32] = "logf";
140 Names[RTLIB::LOG_F64] = "log";
141 Names[RTLIB::LOG_F80] = "logl";
142 Names[RTLIB::LOG_F128] = "logl";
143 Names[RTLIB::LOG_PPCF128] = "logl";
144 Names[RTLIB::LOG2_F32] = "log2f";
145 Names[RTLIB::LOG2_F64] = "log2";
146 Names[RTLIB::LOG2_F80] = "log2l";
147 Names[RTLIB::LOG2_F128] = "log2l";
148 Names[RTLIB::LOG2_PPCF128] = "log2l";
149 Names[RTLIB::LOG10_F32] = "log10f";
150 Names[RTLIB::LOG10_F64] = "log10";
151 Names[RTLIB::LOG10_F80] = "log10l";
152 Names[RTLIB::LOG10_F128] = "log10l";
153 Names[RTLIB::LOG10_PPCF128] = "log10l";
154 Names[RTLIB::EXP_F32] = "expf";
155 Names[RTLIB::EXP_F64] = "exp";
156 Names[RTLIB::EXP_F80] = "expl";
157 Names[RTLIB::EXP_F128] = "expl";
158 Names[RTLIB::EXP_PPCF128] = "expl";
159 Names[RTLIB::EXP2_F32] = "exp2f";
160 Names[RTLIB::EXP2_F64] = "exp2";
161 Names[RTLIB::EXP2_F80] = "exp2l";
162 Names[RTLIB::EXP2_F128] = "exp2l";
163 Names[RTLIB::EXP2_PPCF128] = "exp2l";
164 Names[RTLIB::SIN_F32] = "sinf";
165 Names[RTLIB::SIN_F64] = "sin";
166 Names[RTLIB::SIN_F80] = "sinl";
167 Names[RTLIB::SIN_F128] = "sinl";
168 Names[RTLIB::SIN_PPCF128] = "sinl";
169 Names[RTLIB::COS_F32] = "cosf";
170 Names[RTLIB::COS_F64] = "cos";
171 Names[RTLIB::COS_F80] = "cosl";
172 Names[RTLIB::COS_F128] = "cosl";
173 Names[RTLIB::COS_PPCF128] = "cosl";
174 Names[RTLIB::POW_F32] = "powf";
175 Names[RTLIB::POW_F64] = "pow";
176 Names[RTLIB::POW_F80] = "powl";
177 Names[RTLIB::POW_F128] = "powl";
178 Names[RTLIB::POW_PPCF128] = "powl";
179 Names[RTLIB::CEIL_F32] = "ceilf";
180 Names[RTLIB::CEIL_F64] = "ceil";
181 Names[RTLIB::CEIL_F80] = "ceill";
182 Names[RTLIB::CEIL_F128] = "ceill";
183 Names[RTLIB::CEIL_PPCF128] = "ceill";
184 Names[RTLIB::TRUNC_F32] = "truncf";
185 Names[RTLIB::TRUNC_F64] = "trunc";
186 Names[RTLIB::TRUNC_F80] = "truncl";
187 Names[RTLIB::TRUNC_F128] = "truncl";
188 Names[RTLIB::TRUNC_PPCF128] = "truncl";
189 Names[RTLIB::RINT_F32] = "rintf";
190 Names[RTLIB::RINT_F64] = "rint";
191 Names[RTLIB::RINT_F80] = "rintl";
192 Names[RTLIB::RINT_F128] = "rintl";
193 Names[RTLIB::RINT_PPCF128] = "rintl";
194 Names[RTLIB::NEARBYINT_F32] = "nearbyintf";
195 Names[RTLIB::NEARBYINT_F64] = "nearbyint";
196 Names[RTLIB::NEARBYINT_F80] = "nearbyintl";
197 Names[RTLIB::NEARBYINT_F128] = "nearbyintl";
198 Names[RTLIB::NEARBYINT_PPCF128] = "nearbyintl";
199 Names[RTLIB::ROUND_F32] = "roundf";
200 Names[RTLIB::ROUND_F64] = "round";
201 Names[RTLIB::ROUND_F80] = "roundl";
202 Names[RTLIB::ROUND_F128] = "roundl";
203 Names[RTLIB::ROUND_PPCF128] = "roundl";
204 Names[RTLIB::FLOOR_F32] = "floorf";
205 Names[RTLIB::FLOOR_F64] = "floor";
206 Names[RTLIB::FLOOR_F80] = "floorl";
207 Names[RTLIB::FLOOR_F128] = "floorl";
208 Names[RTLIB::FLOOR_PPCF128] = "floorl";
209 Names[RTLIB::FMIN_F32] = "fminf";
210 Names[RTLIB::FMIN_F64] = "fmin";
211 Names[RTLIB::FMIN_F80] = "fminl";
212 Names[RTLIB::FMIN_F128] = "fminl";
213 Names[RTLIB::FMIN_PPCF128] = "fminl";
214 Names[RTLIB::FMAX_F32] = "fmaxf";
215 Names[RTLIB::FMAX_F64] = "fmax";
216 Names[RTLIB::FMAX_F80] = "fmaxl";
217 Names[RTLIB::FMAX_F128] = "fmaxl";
218 Names[RTLIB::FMAX_PPCF128] = "fmaxl";
219 Names[RTLIB::ROUND_F32] = "roundf";
220 Names[RTLIB::ROUND_F64] = "round";
221 Names[RTLIB::ROUND_F80] = "roundl";
222 Names[RTLIB::ROUND_F128] = "roundl";
223 Names[RTLIB::ROUND_PPCF128] = "roundl";
224 Names[RTLIB::COPYSIGN_F32] = "copysignf";
225 Names[RTLIB::COPYSIGN_F64] = "copysign";
226 Names[RTLIB::COPYSIGN_F80] = "copysignl";
227 Names[RTLIB::COPYSIGN_F128] = "copysignl";
228 Names[RTLIB::COPYSIGN_PPCF128] = "copysignl";
229 Names[RTLIB::FPEXT_F64_F128] = "__extenddftf2";
230 Names[RTLIB::FPEXT_F32_F128] = "__extendsftf2";
231 Names[RTLIB::FPEXT_F32_F64] = "__extendsfdf2";
232 Names[RTLIB::FPEXT_F16_F32] = "__gnu_h2f_ieee";
233 Names[RTLIB::FPROUND_F32_F16] = "__gnu_f2h_ieee";
234 Names[RTLIB::FPROUND_F64_F16] = "__truncdfhf2";
235 Names[RTLIB::FPROUND_F80_F16] = "__truncxfhf2";
236 Names[RTLIB::FPROUND_F128_F16] = "__trunctfhf2";
237 Names[RTLIB::FPROUND_PPCF128_F16] = "__trunctfhf2";
238 Names[RTLIB::FPROUND_F64_F32] = "__truncdfsf2";
239 Names[RTLIB::FPROUND_F80_F32] = "__truncxfsf2";
240 Names[RTLIB::FPROUND_F128_F32] = "__trunctfsf2";
241 Names[RTLIB::FPROUND_PPCF128_F32] = "__trunctfsf2";
242 Names[RTLIB::FPROUND_F80_F64] = "__truncxfdf2";
243 Names[RTLIB::FPROUND_F128_F64] = "__trunctfdf2";
244 Names[RTLIB::FPROUND_PPCF128_F64] = "__trunctfdf2";
245 Names[RTLIB::FPTOSINT_F32_I8] = "__fixsfqi";
246 Names[RTLIB::FPTOSINT_F32_I16] = "__fixsfhi";
247 Names[RTLIB::FPTOSINT_F32_I32] = "__fixsfsi";
248 Names[RTLIB::FPTOSINT_F32_I64] = "__fixsfdi";
249 Names[RTLIB::FPTOSINT_F32_I128] = "__fixsfti";
250 Names[RTLIB::FPTOSINT_F64_I8] = "__fixdfqi";
251 Names[RTLIB::FPTOSINT_F64_I16] = "__fixdfhi";
252 Names[RTLIB::FPTOSINT_F64_I32] = "__fixdfsi";
253 Names[RTLIB::FPTOSINT_F64_I64] = "__fixdfdi";
254 Names[RTLIB::FPTOSINT_F64_I128] = "__fixdfti";
255 Names[RTLIB::FPTOSINT_F80_I32] = "__fixxfsi";
256 Names[RTLIB::FPTOSINT_F80_I64] = "__fixxfdi";
257 Names[RTLIB::FPTOSINT_F80_I128] = "__fixxfti";
258 Names[RTLIB::FPTOSINT_F128_I32] = "__fixtfsi";
259 Names[RTLIB::FPTOSINT_F128_I64] = "__fixtfdi";
260 Names[RTLIB::FPTOSINT_F128_I128] = "__fixtfti";
261 Names[RTLIB::FPTOSINT_PPCF128_I32] = "__fixtfsi";
262 Names[RTLIB::FPTOSINT_PPCF128_I64] = "__fixtfdi";
263 Names[RTLIB::FPTOSINT_PPCF128_I128] = "__fixtfti";
264 Names[RTLIB::FPTOUINT_F32_I8] = "__fixunssfqi";
265 Names[RTLIB::FPTOUINT_F32_I16] = "__fixunssfhi";
266 Names[RTLIB::FPTOUINT_F32_I32] = "__fixunssfsi";
267 Names[RTLIB::FPTOUINT_F32_I64] = "__fixunssfdi";
268 Names[RTLIB::FPTOUINT_F32_I128] = "__fixunssfti";
269 Names[RTLIB::FPTOUINT_F64_I8] = "__fixunsdfqi";
270 Names[RTLIB::FPTOUINT_F64_I16] = "__fixunsdfhi";
271 Names[RTLIB::FPTOUINT_F64_I32] = "__fixunsdfsi";
272 Names[RTLIB::FPTOUINT_F64_I64] = "__fixunsdfdi";
273 Names[RTLIB::FPTOUINT_F64_I128] = "__fixunsdfti";
274 Names[RTLIB::FPTOUINT_F80_I32] = "__fixunsxfsi";
275 Names[RTLIB::FPTOUINT_F80_I64] = "__fixunsxfdi";
276 Names[RTLIB::FPTOUINT_F80_I128] = "__fixunsxfti";
277 Names[RTLIB::FPTOUINT_F128_I32] = "__fixunstfsi";
278 Names[RTLIB::FPTOUINT_F128_I64] = "__fixunstfdi";
279 Names[RTLIB::FPTOUINT_F128_I128] = "__fixunstfti";
280 Names[RTLIB::FPTOUINT_PPCF128_I32] = "__fixunstfsi";
281 Names[RTLIB::FPTOUINT_PPCF128_I64] = "__fixunstfdi";
282 Names[RTLIB::FPTOUINT_PPCF128_I128] = "__fixunstfti";
283 Names[RTLIB::SINTTOFP_I32_F32] = "__floatsisf";
284 Names[RTLIB::SINTTOFP_I32_F64] = "__floatsidf";
285 Names[RTLIB::SINTTOFP_I32_F80] = "__floatsixf";
286 Names[RTLIB::SINTTOFP_I32_F128] = "__floatsitf";
287 Names[RTLIB::SINTTOFP_I32_PPCF128] = "__floatsitf";
288 Names[RTLIB::SINTTOFP_I64_F32] = "__floatdisf";
289 Names[RTLIB::SINTTOFP_I64_F64] = "__floatdidf";
290 Names[RTLIB::SINTTOFP_I64_F80] = "__floatdixf";
291 Names[RTLIB::SINTTOFP_I64_F128] = "__floatditf";
292 Names[RTLIB::SINTTOFP_I64_PPCF128] = "__floatditf";
293 Names[RTLIB::SINTTOFP_I128_F32] = "__floattisf";
294 Names[RTLIB::SINTTOFP_I128_F64] = "__floattidf";
295 Names[RTLIB::SINTTOFP_I128_F80] = "__floattixf";
296 Names[RTLIB::SINTTOFP_I128_F128] = "__floattitf";
297 Names[RTLIB::SINTTOFP_I128_PPCF128] = "__floattitf";
298 Names[RTLIB::UINTTOFP_I32_F32] = "__floatunsisf";
299 Names[RTLIB::UINTTOFP_I32_F64] = "__floatunsidf";
300 Names[RTLIB::UINTTOFP_I32_F80] = "__floatunsixf";
301 Names[RTLIB::UINTTOFP_I32_F128] = "__floatunsitf";
302 Names[RTLIB::UINTTOFP_I32_PPCF128] = "__floatunsitf";
303 Names[RTLIB::UINTTOFP_I64_F32] = "__floatundisf";
304 Names[RTLIB::UINTTOFP_I64_F64] = "__floatundidf";
305 Names[RTLIB::UINTTOFP_I64_F80] = "__floatundixf";
306 Names[RTLIB::UINTTOFP_I64_F128] = "__floatunditf";
307 Names[RTLIB::UINTTOFP_I64_PPCF128] = "__floatunditf";
308 Names[RTLIB::UINTTOFP_I128_F32] = "__floatuntisf";
309 Names[RTLIB::UINTTOFP_I128_F64] = "__floatuntidf";
310 Names[RTLIB::UINTTOFP_I128_F80] = "__floatuntixf";
311 Names[RTLIB::UINTTOFP_I128_F128] = "__floatuntitf";
312 Names[RTLIB::UINTTOFP_I128_PPCF128] = "__floatuntitf";
313 Names[RTLIB::OEQ_F32] = "__eqsf2";
314 Names[RTLIB::OEQ_F64] = "__eqdf2";
315 Names[RTLIB::OEQ_F128] = "__eqtf2";
316 Names[RTLIB::UNE_F32] = "__nesf2";
317 Names[RTLIB::UNE_F64] = "__nedf2";
318 Names[RTLIB::UNE_F128] = "__netf2";
319 Names[RTLIB::OGE_F32] = "__gesf2";
320 Names[RTLIB::OGE_F64] = "__gedf2";
321 Names[RTLIB::OGE_F128] = "__getf2";
322 Names[RTLIB::OLT_F32] = "__ltsf2";
323 Names[RTLIB::OLT_F64] = "__ltdf2";
324 Names[RTLIB::OLT_F128] = "__lttf2";
325 Names[RTLIB::OLE_F32] = "__lesf2";
326 Names[RTLIB::OLE_F64] = "__ledf2";
327 Names[RTLIB::OLE_F128] = "__letf2";
328 Names[RTLIB::OGT_F32] = "__gtsf2";
329 Names[RTLIB::OGT_F64] = "__gtdf2";
330 Names[RTLIB::OGT_F128] = "__gttf2";
331 Names[RTLIB::UO_F32] = "__unordsf2";
332 Names[RTLIB::UO_F64] = "__unorddf2";
333 Names[RTLIB::UO_F128] = "__unordtf2";
334 Names[RTLIB::O_F32] = "__unordsf2";
335 Names[RTLIB::O_F64] = "__unorddf2";
336 Names[RTLIB::O_F128] = "__unordtf2";
337 Names[RTLIB::MEMCPY] = "memcpy";
338 Names[RTLIB::MEMMOVE] = "memmove";
339 Names[RTLIB::MEMSET] = "memset";
340 Names[RTLIB::UNWIND_RESUME] = "_Unwind_Resume";
341 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_1] = "__sync_val_compare_and_swap_1";
342 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_2] = "__sync_val_compare_and_swap_2";
343 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_4] = "__sync_val_compare_and_swap_4";
344 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_8] = "__sync_val_compare_and_swap_8";
345 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_16] = "__sync_val_compare_and_swap_16";
346 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_1] = "__sync_lock_test_and_set_1";
347 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_2] = "__sync_lock_test_and_set_2";
348 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_4] = "__sync_lock_test_and_set_4";
349 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_8] = "__sync_lock_test_and_set_8";
350 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_16] = "__sync_lock_test_and_set_16";
351 Names[RTLIB::SYNC_FETCH_AND_ADD_1] = "__sync_fetch_and_add_1";
352 Names[RTLIB::SYNC_FETCH_AND_ADD_2] = "__sync_fetch_and_add_2";
353 Names[RTLIB::SYNC_FETCH_AND_ADD_4] = "__sync_fetch_and_add_4";
354 Names[RTLIB::SYNC_FETCH_AND_ADD_8] = "__sync_fetch_and_add_8";
355 Names[RTLIB::SYNC_FETCH_AND_ADD_16] = "__sync_fetch_and_add_16";
356 Names[RTLIB::SYNC_FETCH_AND_SUB_1] = "__sync_fetch_and_sub_1";
357 Names[RTLIB::SYNC_FETCH_AND_SUB_2] = "__sync_fetch_and_sub_2";
358 Names[RTLIB::SYNC_FETCH_AND_SUB_4] = "__sync_fetch_and_sub_4";
359 Names[RTLIB::SYNC_FETCH_AND_SUB_8] = "__sync_fetch_and_sub_8";
360 Names[RTLIB::SYNC_FETCH_AND_SUB_16] = "__sync_fetch_and_sub_16";
361 Names[RTLIB::SYNC_FETCH_AND_AND_1] = "__sync_fetch_and_and_1";
362 Names[RTLIB::SYNC_FETCH_AND_AND_2] = "__sync_fetch_and_and_2";
363 Names[RTLIB::SYNC_FETCH_AND_AND_4] = "__sync_fetch_and_and_4";
364 Names[RTLIB::SYNC_FETCH_AND_AND_8] = "__sync_fetch_and_and_8";
365 Names[RTLIB::SYNC_FETCH_AND_AND_16] = "__sync_fetch_and_and_16";
366 Names[RTLIB::SYNC_FETCH_AND_OR_1] = "__sync_fetch_and_or_1";
367 Names[RTLIB::SYNC_FETCH_AND_OR_2] = "__sync_fetch_and_or_2";
368 Names[RTLIB::SYNC_FETCH_AND_OR_4] = "__sync_fetch_and_or_4";
369 Names[RTLIB::SYNC_FETCH_AND_OR_8] = "__sync_fetch_and_or_8";
370 Names[RTLIB::SYNC_FETCH_AND_OR_16] = "__sync_fetch_and_or_16";
371 Names[RTLIB::SYNC_FETCH_AND_XOR_1] = "__sync_fetch_and_xor_1";
372 Names[RTLIB::SYNC_FETCH_AND_XOR_2] = "__sync_fetch_and_xor_2";
373 Names[RTLIB::SYNC_FETCH_AND_XOR_4] = "__sync_fetch_and_xor_4";
374 Names[RTLIB::SYNC_FETCH_AND_XOR_8] = "__sync_fetch_and_xor_8";
375 Names[RTLIB::SYNC_FETCH_AND_XOR_16] = "__sync_fetch_and_xor_16";
376 Names[RTLIB::SYNC_FETCH_AND_NAND_1] = "__sync_fetch_and_nand_1";
377 Names[RTLIB::SYNC_FETCH_AND_NAND_2] = "__sync_fetch_and_nand_2";
378 Names[RTLIB::SYNC_FETCH_AND_NAND_4] = "__sync_fetch_and_nand_4";
379 Names[RTLIB::SYNC_FETCH_AND_NAND_8] = "__sync_fetch_and_nand_8";
380 Names[RTLIB::SYNC_FETCH_AND_NAND_16] = "__sync_fetch_and_nand_16";
381 Names[RTLIB::SYNC_FETCH_AND_MAX_1] = "__sync_fetch_and_max_1";
382 Names[RTLIB::SYNC_FETCH_AND_MAX_2] = "__sync_fetch_and_max_2";
383 Names[RTLIB::SYNC_FETCH_AND_MAX_4] = "__sync_fetch_and_max_4";
384 Names[RTLIB::SYNC_FETCH_AND_MAX_8] = "__sync_fetch_and_max_8";
385 Names[RTLIB::SYNC_FETCH_AND_MAX_16] = "__sync_fetch_and_max_16";
386 Names[RTLIB::SYNC_FETCH_AND_UMAX_1] = "__sync_fetch_and_umax_1";
387 Names[RTLIB::SYNC_FETCH_AND_UMAX_2] = "__sync_fetch_and_umax_2";
388 Names[RTLIB::SYNC_FETCH_AND_UMAX_4] = "__sync_fetch_and_umax_4";
389 Names[RTLIB::SYNC_FETCH_AND_UMAX_8] = "__sync_fetch_and_umax_8";
390 Names[RTLIB::SYNC_FETCH_AND_UMAX_16] = "__sync_fetch_and_umax_16";
391 Names[RTLIB::SYNC_FETCH_AND_MIN_1] = "__sync_fetch_and_min_1";
392 Names[RTLIB::SYNC_FETCH_AND_MIN_2] = "__sync_fetch_and_min_2";
393 Names[RTLIB::SYNC_FETCH_AND_MIN_4] = "__sync_fetch_and_min_4";
394 Names[RTLIB::SYNC_FETCH_AND_MIN_8] = "__sync_fetch_and_min_8";
395 Names[RTLIB::SYNC_FETCH_AND_MIN_16] = "__sync_fetch_and_min_16";
396 Names[RTLIB::SYNC_FETCH_AND_UMIN_1] = "__sync_fetch_and_umin_1";
397 Names[RTLIB::SYNC_FETCH_AND_UMIN_2] = "__sync_fetch_and_umin_2";
398 Names[RTLIB::SYNC_FETCH_AND_UMIN_4] = "__sync_fetch_and_umin_4";
399 Names[RTLIB::SYNC_FETCH_AND_UMIN_8] = "__sync_fetch_and_umin_8";
400 Names[RTLIB::SYNC_FETCH_AND_UMIN_16] = "__sync_fetch_and_umin_16";
401
402 if (TT.getEnvironment() == Triple::GNU) {
403 Names[RTLIB::SINCOS_F32] = "sincosf";
404 Names[RTLIB::SINCOS_F64] = "sincos";
405 Names[RTLIB::SINCOS_F80] = "sincosl";
406 Names[RTLIB::SINCOS_F128] = "sincosl";
407 Names[RTLIB::SINCOS_PPCF128] = "sincosl";
408 } else {
409 // These are generally not available.
410 Names[RTLIB::SINCOS_F32] = nullptr;
411 Names[RTLIB::SINCOS_F64] = nullptr;
412 Names[RTLIB::SINCOS_F80] = nullptr;
413 Names[RTLIB::SINCOS_F128] = nullptr;
414 Names[RTLIB::SINCOS_PPCF128] = nullptr;
415 }
416
417 if (!TT.isOSOpenBSD()) {
418 Names[RTLIB::STACKPROTECTOR_CHECK_FAIL] = "__stack_chk_fail";
419 } else {
420 // These are generally not available.
421 Names[RTLIB::STACKPROTECTOR_CHECK_FAIL] = nullptr;
422 }
423 }
424
425 /// InitLibcallCallingConvs - Set default libcall CallingConvs.
426 ///
InitLibcallCallingConvs(CallingConv::ID * CCs)427 static void InitLibcallCallingConvs(CallingConv::ID *CCs) {
428 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
429 CCs[i] = CallingConv::C;
430 }
431 }
432
433 /// getFPEXT - Return the FPEXT_*_* value for the given types, or
434 /// UNKNOWN_LIBCALL if there is none.
getFPEXT(EVT OpVT,EVT RetVT)435 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) {
436 if (OpVT == MVT::f16) {
437 if (RetVT == MVT::f32)
438 return FPEXT_F16_F32;
439 } else if (OpVT == MVT::f32) {
440 if (RetVT == MVT::f64)
441 return FPEXT_F32_F64;
442 if (RetVT == MVT::f128)
443 return FPEXT_F32_F128;
444 } else if (OpVT == MVT::f64) {
445 if (RetVT == MVT::f128)
446 return FPEXT_F64_F128;
447 }
448
449 return UNKNOWN_LIBCALL;
450 }
451
452 /// getFPROUND - Return the FPROUND_*_* value for the given types, or
453 /// UNKNOWN_LIBCALL if there is none.
getFPROUND(EVT OpVT,EVT RetVT)454 RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) {
455 if (RetVT == MVT::f16) {
456 if (OpVT == MVT::f32)
457 return FPROUND_F32_F16;
458 if (OpVT == MVT::f64)
459 return FPROUND_F64_F16;
460 if (OpVT == MVT::f80)
461 return FPROUND_F80_F16;
462 if (OpVT == MVT::f128)
463 return FPROUND_F128_F16;
464 if (OpVT == MVT::ppcf128)
465 return FPROUND_PPCF128_F16;
466 } else if (RetVT == MVT::f32) {
467 if (OpVT == MVT::f64)
468 return FPROUND_F64_F32;
469 if (OpVT == MVT::f80)
470 return FPROUND_F80_F32;
471 if (OpVT == MVT::f128)
472 return FPROUND_F128_F32;
473 if (OpVT == MVT::ppcf128)
474 return FPROUND_PPCF128_F32;
475 } else if (RetVT == MVT::f64) {
476 if (OpVT == MVT::f80)
477 return FPROUND_F80_F64;
478 if (OpVT == MVT::f128)
479 return FPROUND_F128_F64;
480 if (OpVT == MVT::ppcf128)
481 return FPROUND_PPCF128_F64;
482 }
483
484 return UNKNOWN_LIBCALL;
485 }
486
487 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or
488 /// UNKNOWN_LIBCALL if there is none.
getFPTOSINT(EVT OpVT,EVT RetVT)489 RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) {
490 if (OpVT == MVT::f32) {
491 if (RetVT == MVT::i8)
492 return FPTOSINT_F32_I8;
493 if (RetVT == MVT::i16)
494 return FPTOSINT_F32_I16;
495 if (RetVT == MVT::i32)
496 return FPTOSINT_F32_I32;
497 if (RetVT == MVT::i64)
498 return FPTOSINT_F32_I64;
499 if (RetVT == MVT::i128)
500 return FPTOSINT_F32_I128;
501 } else if (OpVT == MVT::f64) {
502 if (RetVT == MVT::i8)
503 return FPTOSINT_F64_I8;
504 if (RetVT == MVT::i16)
505 return FPTOSINT_F64_I16;
506 if (RetVT == MVT::i32)
507 return FPTOSINT_F64_I32;
508 if (RetVT == MVT::i64)
509 return FPTOSINT_F64_I64;
510 if (RetVT == MVT::i128)
511 return FPTOSINT_F64_I128;
512 } else if (OpVT == MVT::f80) {
513 if (RetVT == MVT::i32)
514 return FPTOSINT_F80_I32;
515 if (RetVT == MVT::i64)
516 return FPTOSINT_F80_I64;
517 if (RetVT == MVT::i128)
518 return FPTOSINT_F80_I128;
519 } else if (OpVT == MVT::f128) {
520 if (RetVT == MVT::i32)
521 return FPTOSINT_F128_I32;
522 if (RetVT == MVT::i64)
523 return FPTOSINT_F128_I64;
524 if (RetVT == MVT::i128)
525 return FPTOSINT_F128_I128;
526 } else if (OpVT == MVT::ppcf128) {
527 if (RetVT == MVT::i32)
528 return FPTOSINT_PPCF128_I32;
529 if (RetVT == MVT::i64)
530 return FPTOSINT_PPCF128_I64;
531 if (RetVT == MVT::i128)
532 return FPTOSINT_PPCF128_I128;
533 }
534 return UNKNOWN_LIBCALL;
535 }
536
537 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or
538 /// UNKNOWN_LIBCALL if there is none.
getFPTOUINT(EVT OpVT,EVT RetVT)539 RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) {
540 if (OpVT == MVT::f32) {
541 if (RetVT == MVT::i8)
542 return FPTOUINT_F32_I8;
543 if (RetVT == MVT::i16)
544 return FPTOUINT_F32_I16;
545 if (RetVT == MVT::i32)
546 return FPTOUINT_F32_I32;
547 if (RetVT == MVT::i64)
548 return FPTOUINT_F32_I64;
549 if (RetVT == MVT::i128)
550 return FPTOUINT_F32_I128;
551 } else if (OpVT == MVT::f64) {
552 if (RetVT == MVT::i8)
553 return FPTOUINT_F64_I8;
554 if (RetVT == MVT::i16)
555 return FPTOUINT_F64_I16;
556 if (RetVT == MVT::i32)
557 return FPTOUINT_F64_I32;
558 if (RetVT == MVT::i64)
559 return FPTOUINT_F64_I64;
560 if (RetVT == MVT::i128)
561 return FPTOUINT_F64_I128;
562 } else if (OpVT == MVT::f80) {
563 if (RetVT == MVT::i32)
564 return FPTOUINT_F80_I32;
565 if (RetVT == MVT::i64)
566 return FPTOUINT_F80_I64;
567 if (RetVT == MVT::i128)
568 return FPTOUINT_F80_I128;
569 } else if (OpVT == MVT::f128) {
570 if (RetVT == MVT::i32)
571 return FPTOUINT_F128_I32;
572 if (RetVT == MVT::i64)
573 return FPTOUINT_F128_I64;
574 if (RetVT == MVT::i128)
575 return FPTOUINT_F128_I128;
576 } else if (OpVT == MVT::ppcf128) {
577 if (RetVT == MVT::i32)
578 return FPTOUINT_PPCF128_I32;
579 if (RetVT == MVT::i64)
580 return FPTOUINT_PPCF128_I64;
581 if (RetVT == MVT::i128)
582 return FPTOUINT_PPCF128_I128;
583 }
584 return UNKNOWN_LIBCALL;
585 }
586
587 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or
588 /// UNKNOWN_LIBCALL if there is none.
getSINTTOFP(EVT OpVT,EVT RetVT)589 RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) {
590 if (OpVT == MVT::i32) {
591 if (RetVT == MVT::f32)
592 return SINTTOFP_I32_F32;
593 if (RetVT == MVT::f64)
594 return SINTTOFP_I32_F64;
595 if (RetVT == MVT::f80)
596 return SINTTOFP_I32_F80;
597 if (RetVT == MVT::f128)
598 return SINTTOFP_I32_F128;
599 if (RetVT == MVT::ppcf128)
600 return SINTTOFP_I32_PPCF128;
601 } else if (OpVT == MVT::i64) {
602 if (RetVT == MVT::f32)
603 return SINTTOFP_I64_F32;
604 if (RetVT == MVT::f64)
605 return SINTTOFP_I64_F64;
606 if (RetVT == MVT::f80)
607 return SINTTOFP_I64_F80;
608 if (RetVT == MVT::f128)
609 return SINTTOFP_I64_F128;
610 if (RetVT == MVT::ppcf128)
611 return SINTTOFP_I64_PPCF128;
612 } else if (OpVT == MVT::i128) {
613 if (RetVT == MVT::f32)
614 return SINTTOFP_I128_F32;
615 if (RetVT == MVT::f64)
616 return SINTTOFP_I128_F64;
617 if (RetVT == MVT::f80)
618 return SINTTOFP_I128_F80;
619 if (RetVT == MVT::f128)
620 return SINTTOFP_I128_F128;
621 if (RetVT == MVT::ppcf128)
622 return SINTTOFP_I128_PPCF128;
623 }
624 return UNKNOWN_LIBCALL;
625 }
626
627 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or
628 /// UNKNOWN_LIBCALL if there is none.
getUINTTOFP(EVT OpVT,EVT RetVT)629 RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) {
630 if (OpVT == MVT::i32) {
631 if (RetVT == MVT::f32)
632 return UINTTOFP_I32_F32;
633 if (RetVT == MVT::f64)
634 return UINTTOFP_I32_F64;
635 if (RetVT == MVT::f80)
636 return UINTTOFP_I32_F80;
637 if (RetVT == MVT::f128)
638 return UINTTOFP_I32_F128;
639 if (RetVT == MVT::ppcf128)
640 return UINTTOFP_I32_PPCF128;
641 } else if (OpVT == MVT::i64) {
642 if (RetVT == MVT::f32)
643 return UINTTOFP_I64_F32;
644 if (RetVT == MVT::f64)
645 return UINTTOFP_I64_F64;
646 if (RetVT == MVT::f80)
647 return UINTTOFP_I64_F80;
648 if (RetVT == MVT::f128)
649 return UINTTOFP_I64_F128;
650 if (RetVT == MVT::ppcf128)
651 return UINTTOFP_I64_PPCF128;
652 } else if (OpVT == MVT::i128) {
653 if (RetVT == MVT::f32)
654 return UINTTOFP_I128_F32;
655 if (RetVT == MVT::f64)
656 return UINTTOFP_I128_F64;
657 if (RetVT == MVT::f80)
658 return UINTTOFP_I128_F80;
659 if (RetVT == MVT::f128)
660 return UINTTOFP_I128_F128;
661 if (RetVT == MVT::ppcf128)
662 return UINTTOFP_I128_PPCF128;
663 }
664 return UNKNOWN_LIBCALL;
665 }
666
getATOMIC(unsigned Opc,MVT VT)667 RTLIB::Libcall RTLIB::getATOMIC(unsigned Opc, MVT VT) {
668 #define OP_TO_LIBCALL(Name, Enum) \
669 case Name: \
670 switch (VT.SimpleTy) { \
671 default: \
672 return UNKNOWN_LIBCALL; \
673 case MVT::i8: \
674 return Enum##_1; \
675 case MVT::i16: \
676 return Enum##_2; \
677 case MVT::i32: \
678 return Enum##_4; \
679 case MVT::i64: \
680 return Enum##_8; \
681 case MVT::i128: \
682 return Enum##_16; \
683 }
684
685 switch (Opc) {
686 OP_TO_LIBCALL(ISD::ATOMIC_SWAP, SYNC_LOCK_TEST_AND_SET)
687 OP_TO_LIBCALL(ISD::ATOMIC_CMP_SWAP, SYNC_VAL_COMPARE_AND_SWAP)
688 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_ADD, SYNC_FETCH_AND_ADD)
689 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_SUB, SYNC_FETCH_AND_SUB)
690 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_AND, SYNC_FETCH_AND_AND)
691 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_OR, SYNC_FETCH_AND_OR)
692 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_XOR, SYNC_FETCH_AND_XOR)
693 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_NAND, SYNC_FETCH_AND_NAND)
694 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MAX, SYNC_FETCH_AND_MAX)
695 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMAX, SYNC_FETCH_AND_UMAX)
696 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MIN, SYNC_FETCH_AND_MIN)
697 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMIN, SYNC_FETCH_AND_UMIN)
698 }
699
700 #undef OP_TO_LIBCALL
701
702 return UNKNOWN_LIBCALL;
703 }
704
705 /// InitCmpLibcallCCs - Set default comparison libcall CC.
706 ///
InitCmpLibcallCCs(ISD::CondCode * CCs)707 static void InitCmpLibcallCCs(ISD::CondCode *CCs) {
708 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL);
709 CCs[RTLIB::OEQ_F32] = ISD::SETEQ;
710 CCs[RTLIB::OEQ_F64] = ISD::SETEQ;
711 CCs[RTLIB::OEQ_F128] = ISD::SETEQ;
712 CCs[RTLIB::UNE_F32] = ISD::SETNE;
713 CCs[RTLIB::UNE_F64] = ISD::SETNE;
714 CCs[RTLIB::UNE_F128] = ISD::SETNE;
715 CCs[RTLIB::OGE_F32] = ISD::SETGE;
716 CCs[RTLIB::OGE_F64] = ISD::SETGE;
717 CCs[RTLIB::OGE_F128] = ISD::SETGE;
718 CCs[RTLIB::OLT_F32] = ISD::SETLT;
719 CCs[RTLIB::OLT_F64] = ISD::SETLT;
720 CCs[RTLIB::OLT_F128] = ISD::SETLT;
721 CCs[RTLIB::OLE_F32] = ISD::SETLE;
722 CCs[RTLIB::OLE_F64] = ISD::SETLE;
723 CCs[RTLIB::OLE_F128] = ISD::SETLE;
724 CCs[RTLIB::OGT_F32] = ISD::SETGT;
725 CCs[RTLIB::OGT_F64] = ISD::SETGT;
726 CCs[RTLIB::OGT_F128] = ISD::SETGT;
727 CCs[RTLIB::UO_F32] = ISD::SETNE;
728 CCs[RTLIB::UO_F64] = ISD::SETNE;
729 CCs[RTLIB::UO_F128] = ISD::SETNE;
730 CCs[RTLIB::O_F32] = ISD::SETEQ;
731 CCs[RTLIB::O_F64] = ISD::SETEQ;
732 CCs[RTLIB::O_F128] = ISD::SETEQ;
733 }
734
735 /// NOTE: The TargetMachine owns TLOF.
TargetLoweringBase(const TargetMachine & tm)736 TargetLoweringBase::TargetLoweringBase(const TargetMachine &tm) : TM(tm) {
737 initActions();
738
739 // Perform these initializations only once.
740 IsLittleEndian = getDataLayout()->isLittleEndian();
741 MaxStoresPerMemset = MaxStoresPerMemcpy = MaxStoresPerMemmove = 8;
742 MaxStoresPerMemsetOptSize = MaxStoresPerMemcpyOptSize
743 = MaxStoresPerMemmoveOptSize = 4;
744 UseUnderscoreSetJmp = false;
745 UseUnderscoreLongJmp = false;
746 SelectIsExpensive = false;
747 HasMultipleConditionRegisters = false;
748 HasExtractBitsInsn = false;
749 IntDivIsCheap = false;
750 FsqrtIsCheap = false;
751 Pow2SDivIsCheap = false;
752 JumpIsExpensive = false;
753 PredictableSelectIsExpensive = false;
754 MaskAndBranchFoldingIsLegal = false;
755 EnableExtLdPromotion = false;
756 HasFloatingPointExceptions = true;
757 StackPointerRegisterToSaveRestore = 0;
758 ExceptionPointerRegister = 0;
759 ExceptionSelectorRegister = 0;
760 BooleanContents = UndefinedBooleanContent;
761 BooleanFloatContents = UndefinedBooleanContent;
762 BooleanVectorContents = UndefinedBooleanContent;
763 SchedPreferenceInfo = Sched::ILP;
764 JumpBufSize = 0;
765 JumpBufAlignment = 0;
766 MinFunctionAlignment = 0;
767 PrefFunctionAlignment = 0;
768 PrefLoopAlignment = 0;
769 MinStackArgumentAlignment = 1;
770 InsertFencesForAtomic = false;
771 MinimumJumpTableEntries = 4;
772
773 InitLibcallNames(LibcallRoutineNames, Triple(TM.getTargetTriple()));
774 InitCmpLibcallCCs(CmpLibcallCCs);
775 InitLibcallCallingConvs(LibcallCallingConvs);
776 }
777
initActions()778 void TargetLoweringBase::initActions() {
779 // All operations default to being supported.
780 memset(OpActions, 0, sizeof(OpActions));
781 memset(LoadExtActions, 0, sizeof(LoadExtActions));
782 memset(TruncStoreActions, 0, sizeof(TruncStoreActions));
783 memset(IndexedModeActions, 0, sizeof(IndexedModeActions));
784 memset(CondCodeActions, 0, sizeof(CondCodeActions));
785 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE*sizeof(TargetRegisterClass*));
786 memset(TargetDAGCombineArray, 0, array_lengthof(TargetDAGCombineArray));
787
788 // Set default actions for various operations.
789 for (MVT VT : MVT::all_valuetypes()) {
790 // Default all indexed load / store to expand.
791 for (unsigned IM = (unsigned)ISD::PRE_INC;
792 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) {
793 setIndexedLoadAction(IM, VT, Expand);
794 setIndexedStoreAction(IM, VT, Expand);
795 }
796
797 // Most backends expect to see the node which just returns the value loaded.
798 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Expand);
799
800 // These operations default to expand.
801 setOperationAction(ISD::FGETSIGN, VT, Expand);
802 setOperationAction(ISD::CONCAT_VECTORS, VT, Expand);
803 setOperationAction(ISD::FMINNUM, VT, Expand);
804 setOperationAction(ISD::FMAXNUM, VT, Expand);
805 setOperationAction(ISD::FMAD, VT, Expand);
806
807 // These library functions default to expand.
808 setOperationAction(ISD::FROUND, VT, Expand);
809
810 // These operations default to expand for vector types.
811 if (VT.isVector()) {
812 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
813 setOperationAction(ISD::ANY_EXTEND_VECTOR_INREG, VT, Expand);
814 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Expand);
815 setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Expand);
816 }
817 }
818
819 // Most targets ignore the @llvm.prefetch intrinsic.
820 setOperationAction(ISD::PREFETCH, MVT::Other, Expand);
821
822 // ConstantFP nodes default to expand. Targets can either change this to
823 // Legal, in which case all fp constants are legal, or use isFPImmLegal()
824 // to optimize expansions for certain constants.
825 setOperationAction(ISD::ConstantFP, MVT::f16, Expand);
826 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
827 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
828 setOperationAction(ISD::ConstantFP, MVT::f80, Expand);
829 setOperationAction(ISD::ConstantFP, MVT::f128, Expand);
830
831 // These library functions default to expand.
832 for (MVT VT : {MVT::f32, MVT::f64, MVT::f128}) {
833 setOperationAction(ISD::FLOG , VT, Expand);
834 setOperationAction(ISD::FLOG2, VT, Expand);
835 setOperationAction(ISD::FLOG10, VT, Expand);
836 setOperationAction(ISD::FEXP , VT, Expand);
837 setOperationAction(ISD::FEXP2, VT, Expand);
838 setOperationAction(ISD::FFLOOR, VT, Expand);
839 setOperationAction(ISD::FMINNUM, VT, Expand);
840 setOperationAction(ISD::FMAXNUM, VT, Expand);
841 setOperationAction(ISD::FNEARBYINT, VT, Expand);
842 setOperationAction(ISD::FCEIL, VT, Expand);
843 setOperationAction(ISD::FRINT, VT, Expand);
844 setOperationAction(ISD::FTRUNC, VT, Expand);
845 setOperationAction(ISD::FROUND, VT, Expand);
846 }
847
848 // Default ISD::TRAP to expand (which turns it into abort).
849 setOperationAction(ISD::TRAP, MVT::Other, Expand);
850
851 // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand"
852 // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP.
853 //
854 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Expand);
855 }
856
getPointerTy(uint32_t AS) const857 MVT TargetLoweringBase::getPointerTy(uint32_t AS) const {
858 return MVT::getIntegerVT(getPointerSizeInBits(AS));
859 }
860
getPointerSizeInBits(uint32_t AS) const861 unsigned TargetLoweringBase::getPointerSizeInBits(uint32_t AS) const {
862 return getDataLayout()->getPointerSizeInBits(AS);
863 }
864
getPointerTypeSizeInBits(Type * Ty) const865 unsigned TargetLoweringBase::getPointerTypeSizeInBits(Type *Ty) const {
866 assert(Ty->isPointerTy());
867 return getPointerSizeInBits(Ty->getPointerAddressSpace());
868 }
869
getScalarShiftAmountTy(EVT LHSTy) const870 MVT TargetLoweringBase::getScalarShiftAmountTy(EVT LHSTy) const {
871 return MVT::getIntegerVT(8 * getDataLayout()->getPointerSize(0));
872 }
873
getShiftAmountTy(EVT LHSTy) const874 EVT TargetLoweringBase::getShiftAmountTy(EVT LHSTy) const {
875 assert(LHSTy.isInteger() && "Shift amount is not an integer type!");
876 if (LHSTy.isVector())
877 return LHSTy;
878 return getScalarShiftAmountTy(LHSTy);
879 }
880
881 /// canOpTrap - Returns true if the operation can trap for the value type.
882 /// VT must be a legal type.
canOpTrap(unsigned Op,EVT VT) const883 bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const {
884 assert(isTypeLegal(VT));
885 switch (Op) {
886 default:
887 return false;
888 case ISD::FDIV:
889 case ISD::FREM:
890 case ISD::SDIV:
891 case ISD::UDIV:
892 case ISD::SREM:
893 case ISD::UREM:
894 return true;
895 }
896 }
897
898 TargetLoweringBase::LegalizeKind
getTypeConversion(LLVMContext & Context,EVT VT) const899 TargetLoweringBase::getTypeConversion(LLVMContext &Context, EVT VT) const {
900 // If this is a simple type, use the ComputeRegisterProp mechanism.
901 if (VT.isSimple()) {
902 MVT SVT = VT.getSimpleVT();
903 assert((unsigned)SVT.SimpleTy < array_lengthof(TransformToType));
904 MVT NVT = TransformToType[SVT.SimpleTy];
905 LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT);
906
907 assert((LA == TypeLegal || LA == TypeSoftenFloat ||
908 ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger) &&
909 "Promote may not follow Expand or Promote");
910
911 if (LA == TypeSplitVector)
912 return LegalizeKind(LA,
913 EVT::getVectorVT(Context, SVT.getVectorElementType(),
914 SVT.getVectorNumElements() / 2));
915 if (LA == TypeScalarizeVector)
916 return LegalizeKind(LA, SVT.getVectorElementType());
917 return LegalizeKind(LA, NVT);
918 }
919
920 // Handle Extended Scalar Types.
921 if (!VT.isVector()) {
922 assert(VT.isInteger() && "Float types must be simple");
923 unsigned BitSize = VT.getSizeInBits();
924 // First promote to a power-of-two size, then expand if necessary.
925 if (BitSize < 8 || !isPowerOf2_32(BitSize)) {
926 EVT NVT = VT.getRoundIntegerType(Context);
927 assert(NVT != VT && "Unable to round integer VT");
928 LegalizeKind NextStep = getTypeConversion(Context, NVT);
929 // Avoid multi-step promotion.
930 if (NextStep.first == TypePromoteInteger)
931 return NextStep;
932 // Return rounded integer type.
933 return LegalizeKind(TypePromoteInteger, NVT);
934 }
935
936 return LegalizeKind(TypeExpandInteger,
937 EVT::getIntegerVT(Context, VT.getSizeInBits() / 2));
938 }
939
940 // Handle vector types.
941 unsigned NumElts = VT.getVectorNumElements();
942 EVT EltVT = VT.getVectorElementType();
943
944 // Vectors with only one element are always scalarized.
945 if (NumElts == 1)
946 return LegalizeKind(TypeScalarizeVector, EltVT);
947
948 // Try to widen vector elements until the element type is a power of two and
949 // promote it to a legal type later on, for example:
950 // <3 x i8> -> <4 x i8> -> <4 x i32>
951 if (EltVT.isInteger()) {
952 // Vectors with a number of elements that is not a power of two are always
953 // widened, for example <3 x i8> -> <4 x i8>.
954 if (!VT.isPow2VectorType()) {
955 NumElts = (unsigned)NextPowerOf2(NumElts);
956 EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts);
957 return LegalizeKind(TypeWidenVector, NVT);
958 }
959
960 // Examine the element type.
961 LegalizeKind LK = getTypeConversion(Context, EltVT);
962
963 // If type is to be expanded, split the vector.
964 // <4 x i140> -> <2 x i140>
965 if (LK.first == TypeExpandInteger)
966 return LegalizeKind(TypeSplitVector,
967 EVT::getVectorVT(Context, EltVT, NumElts / 2));
968
969 // Promote the integer element types until a legal vector type is found
970 // or until the element integer type is too big. If a legal type was not
971 // found, fallback to the usual mechanism of widening/splitting the
972 // vector.
973 EVT OldEltVT = EltVT;
974 while (1) {
975 // Increase the bitwidth of the element to the next pow-of-two
976 // (which is greater than 8 bits).
977 EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits())
978 .getRoundIntegerType(Context);
979
980 // Stop trying when getting a non-simple element type.
981 // Note that vector elements may be greater than legal vector element
982 // types. Example: X86 XMM registers hold 64bit element on 32bit
983 // systems.
984 if (!EltVT.isSimple())
985 break;
986
987 // Build a new vector type and check if it is legal.
988 MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts);
989 // Found a legal promoted vector type.
990 if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal)
991 return LegalizeKind(TypePromoteInteger,
992 EVT::getVectorVT(Context, EltVT, NumElts));
993 }
994
995 // Reset the type to the unexpanded type if we did not find a legal vector
996 // type with a promoted vector element type.
997 EltVT = OldEltVT;
998 }
999
1000 // Try to widen the vector until a legal type is found.
1001 // If there is no wider legal type, split the vector.
1002 while (1) {
1003 // Round up to the next power of 2.
1004 NumElts = (unsigned)NextPowerOf2(NumElts);
1005
1006 // If there is no simple vector type with this many elements then there
1007 // cannot be a larger legal vector type. Note that this assumes that
1008 // there are no skipped intermediate vector types in the simple types.
1009 if (!EltVT.isSimple())
1010 break;
1011 MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts);
1012 if (LargerVector == MVT())
1013 break;
1014
1015 // If this type is legal then widen the vector.
1016 if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal)
1017 return LegalizeKind(TypeWidenVector, LargerVector);
1018 }
1019
1020 // Widen odd vectors to next power of two.
1021 if (!VT.isPow2VectorType()) {
1022 EVT NVT = VT.getPow2VectorType(Context);
1023 return LegalizeKind(TypeWidenVector, NVT);
1024 }
1025
1026 // Vectors with illegal element types are expanded.
1027 EVT NVT = EVT::getVectorVT(Context, EltVT, VT.getVectorNumElements() / 2);
1028 return LegalizeKind(TypeSplitVector, NVT);
1029 }
1030
getVectorTypeBreakdownMVT(MVT VT,MVT & IntermediateVT,unsigned & NumIntermediates,MVT & RegisterVT,TargetLoweringBase * TLI)1031 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT,
1032 unsigned &NumIntermediates,
1033 MVT &RegisterVT,
1034 TargetLoweringBase *TLI) {
1035 // Figure out the right, legal destination reg to copy into.
1036 unsigned NumElts = VT.getVectorNumElements();
1037 MVT EltTy = VT.getVectorElementType();
1038
1039 unsigned NumVectorRegs = 1;
1040
1041 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
1042 // could break down into LHS/RHS like LegalizeDAG does.
1043 if (!isPowerOf2_32(NumElts)) {
1044 NumVectorRegs = NumElts;
1045 NumElts = 1;
1046 }
1047
1048 // Divide the input until we get to a supported size. This will always
1049 // end with a scalar if the target doesn't support vectors.
1050 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) {
1051 NumElts >>= 1;
1052 NumVectorRegs <<= 1;
1053 }
1054
1055 NumIntermediates = NumVectorRegs;
1056
1057 MVT NewVT = MVT::getVectorVT(EltTy, NumElts);
1058 if (!TLI->isTypeLegal(NewVT))
1059 NewVT = EltTy;
1060 IntermediateVT = NewVT;
1061
1062 unsigned NewVTSize = NewVT.getSizeInBits();
1063
1064 // Convert sizes such as i33 to i64.
1065 if (!isPowerOf2_32(NewVTSize))
1066 NewVTSize = NextPowerOf2(NewVTSize);
1067
1068 MVT DestVT = TLI->getRegisterType(NewVT);
1069 RegisterVT = DestVT;
1070 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16.
1071 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits());
1072
1073 // Otherwise, promotion or legal types use the same number of registers as
1074 // the vector decimated to the appropriate level.
1075 return NumVectorRegs;
1076 }
1077
1078 /// isLegalRC - Return true if the value types that can be represented by the
1079 /// specified register class are all legal.
isLegalRC(const TargetRegisterClass * RC) const1080 bool TargetLoweringBase::isLegalRC(const TargetRegisterClass *RC) const {
1081 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
1082 I != E; ++I) {
1083 if (isTypeLegal(*I))
1084 return true;
1085 }
1086 return false;
1087 }
1088
1089 /// Replace/modify any TargetFrameIndex operands with a targte-dependent
1090 /// sequence of memory operands that is recognized by PrologEpilogInserter.
1091 MachineBasicBlock*
emitPatchPoint(MachineInstr * MI,MachineBasicBlock * MBB) const1092 TargetLoweringBase::emitPatchPoint(MachineInstr *MI,
1093 MachineBasicBlock *MBB) const {
1094 MachineFunction &MF = *MI->getParent()->getParent();
1095
1096 // MI changes inside this loop as we grow operands.
1097 for(unsigned OperIdx = 0; OperIdx != MI->getNumOperands(); ++OperIdx) {
1098 MachineOperand &MO = MI->getOperand(OperIdx);
1099 if (!MO.isFI())
1100 continue;
1101
1102 // foldMemoryOperand builds a new MI after replacing a single FI operand
1103 // with the canonical set of five x86 addressing-mode operands.
1104 int FI = MO.getIndex();
1105 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), MI->getDesc());
1106
1107 // Copy operands before the frame-index.
1108 for (unsigned i = 0; i < OperIdx; ++i)
1109 MIB.addOperand(MI->getOperand(i));
1110 // Add frame index operands: direct-mem-ref tag, #FI, offset.
1111 MIB.addImm(StackMaps::DirectMemRefOp);
1112 MIB.addOperand(MI->getOperand(OperIdx));
1113 MIB.addImm(0);
1114 // Copy the operands after the frame index.
1115 for (unsigned i = OperIdx + 1; i != MI->getNumOperands(); ++i)
1116 MIB.addOperand(MI->getOperand(i));
1117
1118 // Inherit previous memory operands.
1119 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
1120 assert(MIB->mayLoad() && "Folded a stackmap use to a non-load!");
1121
1122 // Add a new memory operand for this FI.
1123 const MachineFrameInfo &MFI = *MF.getFrameInfo();
1124 assert(MFI.getObjectOffset(FI) != -1);
1125
1126 unsigned Flags = MachineMemOperand::MOLoad;
1127 if (MI->getOpcode() == TargetOpcode::STATEPOINT) {
1128 Flags |= MachineMemOperand::MOStore;
1129 Flags |= MachineMemOperand::MOVolatile;
1130 }
1131 MachineMemOperand *MMO = MF.getMachineMemOperand(
1132 MachinePointerInfo::getFixedStack(FI), Flags,
1133 TM.getDataLayout()->getPointerSize(), MFI.getObjectAlignment(FI));
1134 MIB->addMemOperand(MF, MMO);
1135
1136 // Replace the instruction and update the operand index.
1137 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
1138 OperIdx += (MIB->getNumOperands() - MI->getNumOperands()) - 1;
1139 MI->eraseFromParent();
1140 MI = MIB;
1141 }
1142 return MBB;
1143 }
1144
1145 /// findRepresentativeClass - Return the largest legal super-reg register class
1146 /// of the register class for the specified type and its associated "cost".
1147 // This function is in TargetLowering because it uses RegClassForVT which would
1148 // need to be moved to TargetRegisterInfo and would necessitate moving
1149 // isTypeLegal over as well - a massive change that would just require
1150 // TargetLowering having a TargetRegisterInfo class member that it would use.
1151 std::pair<const TargetRegisterClass *, uint8_t>
findRepresentativeClass(const TargetRegisterInfo * TRI,MVT VT) const1152 TargetLoweringBase::findRepresentativeClass(const TargetRegisterInfo *TRI,
1153 MVT VT) const {
1154 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy];
1155 if (!RC)
1156 return std::make_pair(RC, 0);
1157
1158 // Compute the set of all super-register classes.
1159 BitVector SuperRegRC(TRI->getNumRegClasses());
1160 for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI)
1161 SuperRegRC.setBitsInMask(RCI.getMask());
1162
1163 // Find the first legal register class with the largest spill size.
1164 const TargetRegisterClass *BestRC = RC;
1165 for (int i = SuperRegRC.find_first(); i >= 0; i = SuperRegRC.find_next(i)) {
1166 const TargetRegisterClass *SuperRC = TRI->getRegClass(i);
1167 // We want the largest possible spill size.
1168 if (SuperRC->getSize() <= BestRC->getSize())
1169 continue;
1170 if (!isLegalRC(SuperRC))
1171 continue;
1172 BestRC = SuperRC;
1173 }
1174 return std::make_pair(BestRC, 1);
1175 }
1176
1177 /// computeRegisterProperties - Once all of the register classes are added,
1178 /// this allows us to compute derived properties we expose.
computeRegisterProperties(const TargetRegisterInfo * TRI)1179 void TargetLoweringBase::computeRegisterProperties(
1180 const TargetRegisterInfo *TRI) {
1181 static_assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE,
1182 "Too many value types for ValueTypeActions to hold!");
1183
1184 // Everything defaults to needing one register.
1185 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
1186 NumRegistersForVT[i] = 1;
1187 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i;
1188 }
1189 // ...except isVoid, which doesn't need any registers.
1190 NumRegistersForVT[MVT::isVoid] = 0;
1191
1192 // Find the largest integer register class.
1193 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE;
1194 for (; RegClassForVT[LargestIntReg] == nullptr; --LargestIntReg)
1195 assert(LargestIntReg != MVT::i1 && "No integer registers defined!");
1196
1197 // Every integer value type larger than this largest register takes twice as
1198 // many registers to represent as the previous ValueType.
1199 for (unsigned ExpandedReg = LargestIntReg + 1;
1200 ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) {
1201 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1];
1202 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg;
1203 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1);
1204 ValueTypeActions.setTypeAction((MVT::SimpleValueType)ExpandedReg,
1205 TypeExpandInteger);
1206 }
1207
1208 // Inspect all of the ValueType's smaller than the largest integer
1209 // register to see which ones need promotion.
1210 unsigned LegalIntReg = LargestIntReg;
1211 for (unsigned IntReg = LargestIntReg - 1;
1212 IntReg >= (unsigned)MVT::i1; --IntReg) {
1213 MVT IVT = (MVT::SimpleValueType)IntReg;
1214 if (isTypeLegal(IVT)) {
1215 LegalIntReg = IntReg;
1216 } else {
1217 RegisterTypeForVT[IntReg] = TransformToType[IntReg] =
1218 (const MVT::SimpleValueType)LegalIntReg;
1219 ValueTypeActions.setTypeAction(IVT, TypePromoteInteger);
1220 }
1221 }
1222
1223 // ppcf128 type is really two f64's.
1224 if (!isTypeLegal(MVT::ppcf128)) {
1225 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64];
1226 RegisterTypeForVT[MVT::ppcf128] = MVT::f64;
1227 TransformToType[MVT::ppcf128] = MVT::f64;
1228 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat);
1229 }
1230
1231 // Decide how to handle f128. If the target does not have native f128 support,
1232 // expand it to i128 and we will be generating soft float library calls.
1233 if (!isTypeLegal(MVT::f128)) {
1234 NumRegistersForVT[MVT::f128] = NumRegistersForVT[MVT::i128];
1235 RegisterTypeForVT[MVT::f128] = RegisterTypeForVT[MVT::i128];
1236 TransformToType[MVT::f128] = MVT::i128;
1237 ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat);
1238 }
1239
1240 // Decide how to handle f64. If the target does not have native f64 support,
1241 // expand it to i64 and we will be generating soft float library calls.
1242 if (!isTypeLegal(MVT::f64)) {
1243 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64];
1244 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64];
1245 TransformToType[MVT::f64] = MVT::i64;
1246 ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat);
1247 }
1248
1249 // Decide how to handle f32. If the target does not have native f32 support,
1250 // expand it to i32 and we will be generating soft float library calls.
1251 if (!isTypeLegal(MVT::f32)) {
1252 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32];
1253 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32];
1254 TransformToType[MVT::f32] = MVT::i32;
1255 ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat);
1256 }
1257
1258 if (!isTypeLegal(MVT::f16)) {
1259 // If the target has native f32 support, promote f16 operations to f32. If
1260 // f32 is not supported, generate soft float library calls.
1261 if (isTypeLegal(MVT::f32)) {
1262 NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::f32];
1263 RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::f32];
1264 TransformToType[MVT::f16] = MVT::f32;
1265 ValueTypeActions.setTypeAction(MVT::f16, TypePromoteFloat);
1266 } else {
1267 NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::i16];
1268 RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::i16];
1269 TransformToType[MVT::f16] = MVT::i16;
1270 ValueTypeActions.setTypeAction(MVT::f16, TypeSoftenFloat);
1271 }
1272 }
1273
1274 // Loop over all of the vector value types to see which need transformations.
1275 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE;
1276 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
1277 MVT VT = (MVT::SimpleValueType) i;
1278 if (isTypeLegal(VT))
1279 continue;
1280
1281 MVT EltVT = VT.getVectorElementType();
1282 unsigned NElts = VT.getVectorNumElements();
1283 bool IsLegalWiderType = false;
1284 LegalizeTypeAction PreferredAction = getPreferredVectorAction(VT);
1285 switch (PreferredAction) {
1286 case TypePromoteInteger: {
1287 // Try to promote the elements of integer vectors. If no legal
1288 // promotion was found, fall through to the widen-vector method.
1289 for (unsigned nVT = i + 1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
1290 MVT SVT = (MVT::SimpleValueType) nVT;
1291 // Promote vectors of integers to vectors with the same number
1292 // of elements, with a wider element type.
1293 if (SVT.getVectorElementType().getSizeInBits() > EltVT.getSizeInBits()
1294 && SVT.getVectorNumElements() == NElts && isTypeLegal(SVT)
1295 && SVT.getScalarType().isInteger()) {
1296 TransformToType[i] = SVT;
1297 RegisterTypeForVT[i] = SVT;
1298 NumRegistersForVT[i] = 1;
1299 ValueTypeActions.setTypeAction(VT, TypePromoteInteger);
1300 IsLegalWiderType = true;
1301 break;
1302 }
1303 }
1304 if (IsLegalWiderType)
1305 break;
1306 }
1307 case TypeWidenVector: {
1308 // Try to widen the vector.
1309 for (unsigned nVT = i + 1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
1310 MVT SVT = (MVT::SimpleValueType) nVT;
1311 if (SVT.getVectorElementType() == EltVT
1312 && SVT.getVectorNumElements() > NElts && isTypeLegal(SVT)) {
1313 TransformToType[i] = SVT;
1314 RegisterTypeForVT[i] = SVT;
1315 NumRegistersForVT[i] = 1;
1316 ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1317 IsLegalWiderType = true;
1318 break;
1319 }
1320 }
1321 if (IsLegalWiderType)
1322 break;
1323 }
1324 case TypeSplitVector:
1325 case TypeScalarizeVector: {
1326 MVT IntermediateVT;
1327 MVT RegisterVT;
1328 unsigned NumIntermediates;
1329 NumRegistersForVT[i] = getVectorTypeBreakdownMVT(VT, IntermediateVT,
1330 NumIntermediates, RegisterVT, this);
1331 RegisterTypeForVT[i] = RegisterVT;
1332
1333 MVT NVT = VT.getPow2VectorType();
1334 if (NVT == VT) {
1335 // Type is already a power of 2. The default action is to split.
1336 TransformToType[i] = MVT::Other;
1337 if (PreferredAction == TypeScalarizeVector)
1338 ValueTypeActions.setTypeAction(VT, TypeScalarizeVector);
1339 else if (PreferredAction == TypeSplitVector)
1340 ValueTypeActions.setTypeAction(VT, TypeSplitVector);
1341 else
1342 // Set type action according to the number of elements.
1343 ValueTypeActions.setTypeAction(VT, NElts == 1 ? TypeScalarizeVector
1344 : TypeSplitVector);
1345 } else {
1346 TransformToType[i] = NVT;
1347 ValueTypeActions.setTypeAction(VT, TypeWidenVector);
1348 }
1349 break;
1350 }
1351 default:
1352 llvm_unreachable("Unknown vector legalization action!");
1353 }
1354 }
1355
1356 // Determine the 'representative' register class for each value type.
1357 // An representative register class is the largest (meaning one which is
1358 // not a sub-register class / subreg register class) legal register class for
1359 // a group of value types. For example, on i386, i8, i16, and i32
1360 // representative would be GR32; while on x86_64 it's GR64.
1361 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) {
1362 const TargetRegisterClass* RRC;
1363 uint8_t Cost;
1364 std::tie(RRC, Cost) = findRepresentativeClass(TRI, (MVT::SimpleValueType)i);
1365 RepRegClassForVT[i] = RRC;
1366 RepRegClassCostForVT[i] = Cost;
1367 }
1368 }
1369
getSetCCResultType(LLVMContext &,EVT VT) const1370 EVT TargetLoweringBase::getSetCCResultType(LLVMContext &, EVT VT) const {
1371 assert(!VT.isVector() && "No default SetCC type for vectors!");
1372 return getPointerTy(0).SimpleTy;
1373 }
1374
getCmpLibcallReturnType() const1375 MVT::SimpleValueType TargetLoweringBase::getCmpLibcallReturnType() const {
1376 return MVT::i32; // return the default value
1377 }
1378
1379 /// getVectorTypeBreakdown - Vector types are broken down into some number of
1380 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
1381 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
1382 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
1383 ///
1384 /// This method returns the number of registers needed, and the VT for each
1385 /// register. It also returns the VT and quantity of the intermediate values
1386 /// before they are promoted/expanded.
1387 ///
getVectorTypeBreakdown(LLVMContext & Context,EVT VT,EVT & IntermediateVT,unsigned & NumIntermediates,MVT & RegisterVT) const1388 unsigned TargetLoweringBase::getVectorTypeBreakdown(LLVMContext &Context, EVT VT,
1389 EVT &IntermediateVT,
1390 unsigned &NumIntermediates,
1391 MVT &RegisterVT) const {
1392 unsigned NumElts = VT.getVectorNumElements();
1393
1394 // If there is a wider vector type with the same element type as this one,
1395 // or a promoted vector type that has the same number of elements which
1396 // are wider, then we should convert to that legal vector type.
1397 // This handles things like <2 x float> -> <4 x float> and
1398 // <4 x i1> -> <4 x i32>.
1399 LegalizeTypeAction TA = getTypeAction(Context, VT);
1400 if (NumElts != 1 && (TA == TypeWidenVector || TA == TypePromoteInteger)) {
1401 EVT RegisterEVT = getTypeToTransformTo(Context, VT);
1402 if (isTypeLegal(RegisterEVT)) {
1403 IntermediateVT = RegisterEVT;
1404 RegisterVT = RegisterEVT.getSimpleVT();
1405 NumIntermediates = 1;
1406 return 1;
1407 }
1408 }
1409
1410 // Figure out the right, legal destination reg to copy into.
1411 EVT EltTy = VT.getVectorElementType();
1412
1413 unsigned NumVectorRegs = 1;
1414
1415 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we
1416 // could break down into LHS/RHS like LegalizeDAG does.
1417 if (!isPowerOf2_32(NumElts)) {
1418 NumVectorRegs = NumElts;
1419 NumElts = 1;
1420 }
1421
1422 // Divide the input until we get to a supported size. This will always
1423 // end with a scalar if the target doesn't support vectors.
1424 while (NumElts > 1 && !isTypeLegal(
1425 EVT::getVectorVT(Context, EltTy, NumElts))) {
1426 NumElts >>= 1;
1427 NumVectorRegs <<= 1;
1428 }
1429
1430 NumIntermediates = NumVectorRegs;
1431
1432 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts);
1433 if (!isTypeLegal(NewVT))
1434 NewVT = EltTy;
1435 IntermediateVT = NewVT;
1436
1437 MVT DestVT = getRegisterType(Context, NewVT);
1438 RegisterVT = DestVT;
1439 unsigned NewVTSize = NewVT.getSizeInBits();
1440
1441 // Convert sizes such as i33 to i64.
1442 if (!isPowerOf2_32(NewVTSize))
1443 NewVTSize = NextPowerOf2(NewVTSize);
1444
1445 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16.
1446 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits());
1447
1448 // Otherwise, promotion or legal types use the same number of registers as
1449 // the vector decimated to the appropriate level.
1450 return NumVectorRegs;
1451 }
1452
1453 /// Get the EVTs and ArgFlags collections that represent the legalized return
1454 /// type of the given function. This does not require a DAG or a return value,
1455 /// and is suitable for use before any DAGs for the function are constructed.
1456 /// TODO: Move this out of TargetLowering.cpp.
GetReturnInfo(Type * ReturnType,AttributeSet attr,SmallVectorImpl<ISD::OutputArg> & Outs,const TargetLowering & TLI)1457 void llvm::GetReturnInfo(Type* ReturnType, AttributeSet attr,
1458 SmallVectorImpl<ISD::OutputArg> &Outs,
1459 const TargetLowering &TLI) {
1460 SmallVector<EVT, 4> ValueVTs;
1461 ComputeValueVTs(TLI, ReturnType, ValueVTs);
1462 unsigned NumValues = ValueVTs.size();
1463 if (NumValues == 0) return;
1464
1465 for (unsigned j = 0, f = NumValues; j != f; ++j) {
1466 EVT VT = ValueVTs[j];
1467 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
1468
1469 if (attr.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
1470 ExtendKind = ISD::SIGN_EXTEND;
1471 else if (attr.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt))
1472 ExtendKind = ISD::ZERO_EXTEND;
1473
1474 // FIXME: C calling convention requires the return type to be promoted to
1475 // at least 32-bit. But this is not necessary for non-C calling
1476 // conventions. The frontend should mark functions whose return values
1477 // require promoting with signext or zeroext attributes.
1478 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) {
1479 MVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32);
1480 if (VT.bitsLT(MinVT))
1481 VT = MinVT;
1482 }
1483
1484 unsigned NumParts = TLI.getNumRegisters(ReturnType->getContext(), VT);
1485 MVT PartVT = TLI.getRegisterType(ReturnType->getContext(), VT);
1486
1487 // 'inreg' on function refers to return value
1488 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1489 if (attr.hasAttribute(AttributeSet::ReturnIndex, Attribute::InReg))
1490 Flags.setInReg();
1491
1492 // Propagate extension type if any
1493 if (attr.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
1494 Flags.setSExt();
1495 else if (attr.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt))
1496 Flags.setZExt();
1497
1498 for (unsigned i = 0; i < NumParts; ++i)
1499 Outs.push_back(ISD::OutputArg(Flags, PartVT, VT, /*isFixed=*/true, 0, 0));
1500 }
1501 }
1502
1503 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1504 /// function arguments in the caller parameter area. This is the actual
1505 /// alignment, not its logarithm.
getByValTypeAlignment(Type * Ty) const1506 unsigned TargetLoweringBase::getByValTypeAlignment(Type *Ty) const {
1507 return getDataLayout()->getABITypeAlignment(Ty);
1508 }
1509
1510 //===----------------------------------------------------------------------===//
1511 // TargetTransformInfo Helpers
1512 //===----------------------------------------------------------------------===//
1513
InstructionOpcodeToISD(unsigned Opcode) const1514 int TargetLoweringBase::InstructionOpcodeToISD(unsigned Opcode) const {
1515 enum InstructionOpcodes {
1516 #define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM,
1517 #define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM
1518 #include "llvm/IR/Instruction.def"
1519 };
1520 switch (static_cast<InstructionOpcodes>(Opcode)) {
1521 case Ret: return 0;
1522 case Br: return 0;
1523 case Switch: return 0;
1524 case IndirectBr: return 0;
1525 case Invoke: return 0;
1526 case Resume: return 0;
1527 case Unreachable: return 0;
1528 case Add: return ISD::ADD;
1529 case FAdd: return ISD::FADD;
1530 case Sub: return ISD::SUB;
1531 case FSub: return ISD::FSUB;
1532 case Mul: return ISD::MUL;
1533 case FMul: return ISD::FMUL;
1534 case UDiv: return ISD::UDIV;
1535 case SDiv: return ISD::SDIV;
1536 case FDiv: return ISD::FDIV;
1537 case URem: return ISD::UREM;
1538 case SRem: return ISD::SREM;
1539 case FRem: return ISD::FREM;
1540 case Shl: return ISD::SHL;
1541 case LShr: return ISD::SRL;
1542 case AShr: return ISD::SRA;
1543 case And: return ISD::AND;
1544 case Or: return ISD::OR;
1545 case Xor: return ISD::XOR;
1546 case Alloca: return 0;
1547 case Load: return ISD::LOAD;
1548 case Store: return ISD::STORE;
1549 case GetElementPtr: return 0;
1550 case Fence: return 0;
1551 case AtomicCmpXchg: return 0;
1552 case AtomicRMW: return 0;
1553 case Trunc: return ISD::TRUNCATE;
1554 case ZExt: return ISD::ZERO_EXTEND;
1555 case SExt: return ISD::SIGN_EXTEND;
1556 case FPToUI: return ISD::FP_TO_UINT;
1557 case FPToSI: return ISD::FP_TO_SINT;
1558 case UIToFP: return ISD::UINT_TO_FP;
1559 case SIToFP: return ISD::SINT_TO_FP;
1560 case FPTrunc: return ISD::FP_ROUND;
1561 case FPExt: return ISD::FP_EXTEND;
1562 case PtrToInt: return ISD::BITCAST;
1563 case IntToPtr: return ISD::BITCAST;
1564 case BitCast: return ISD::BITCAST;
1565 case AddrSpaceCast: return ISD::ADDRSPACECAST;
1566 case ICmp: return ISD::SETCC;
1567 case FCmp: return ISD::SETCC;
1568 case PHI: return 0;
1569 case Call: return 0;
1570 case Select: return ISD::SELECT;
1571 case UserOp1: return 0;
1572 case UserOp2: return 0;
1573 case VAArg: return 0;
1574 case ExtractElement: return ISD::EXTRACT_VECTOR_ELT;
1575 case InsertElement: return ISD::INSERT_VECTOR_ELT;
1576 case ShuffleVector: return ISD::VECTOR_SHUFFLE;
1577 case ExtractValue: return ISD::MERGE_VALUES;
1578 case InsertValue: return ISD::MERGE_VALUES;
1579 case LandingPad: return 0;
1580 }
1581
1582 llvm_unreachable("Unknown instruction type encountered!");
1583 }
1584
1585 std::pair<unsigned, MVT>
getTypeLegalizationCost(Type * Ty) const1586 TargetLoweringBase::getTypeLegalizationCost(Type *Ty) const {
1587 LLVMContext &C = Ty->getContext();
1588 EVT MTy = getValueType(Ty);
1589
1590 unsigned Cost = 1;
1591 // We keep legalizing the type until we find a legal kind. We assume that
1592 // the only operation that costs anything is the split. After splitting
1593 // we need to handle two types.
1594 while (true) {
1595 LegalizeKind LK = getTypeConversion(C, MTy);
1596
1597 if (LK.first == TypeLegal)
1598 return std::make_pair(Cost, MTy.getSimpleVT());
1599
1600 if (LK.first == TypeSplitVector || LK.first == TypeExpandInteger)
1601 Cost *= 2;
1602
1603 // Keep legalizing the type.
1604 MTy = LK.second;
1605 }
1606 }
1607
1608 //===----------------------------------------------------------------------===//
1609 // Loop Strength Reduction hooks
1610 //===----------------------------------------------------------------------===//
1611
1612 /// isLegalAddressingMode - Return true if the addressing mode represented
1613 /// by AM is legal for this target, for a load/store of the specified type.
isLegalAddressingMode(const AddrMode & AM,Type * Ty) const1614 bool TargetLoweringBase::isLegalAddressingMode(const AddrMode &AM,
1615 Type *Ty) const {
1616 // The default implementation of this implements a conservative RISCy, r+r and
1617 // r+i addr mode.
1618
1619 // Allows a sign-extended 16-bit immediate field.
1620 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
1621 return false;
1622
1623 // No global is ever allowed as a base.
1624 if (AM.BaseGV)
1625 return false;
1626
1627 // Only support r+r,
1628 switch (AM.Scale) {
1629 case 0: // "r+i" or just "i", depending on HasBaseReg.
1630 break;
1631 case 1:
1632 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
1633 return false;
1634 // Otherwise we have r+r or r+i.
1635 break;
1636 case 2:
1637 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
1638 return false;
1639 // Allow 2*r as r+r.
1640 break;
1641 default: // Don't allow n * r
1642 return false;
1643 }
1644
1645 return true;
1646 }
1647