1 //===-- AMDGPULibFunc.cpp -------------------------------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file contains utility functions to work with Itanium mangled names
10 //
11 //===----------------------------------------------------------------------===//
12
13 #include "AMDGPULibFunc.h"
14 #include "AMDGPU.h"
15 #include "llvm/ADT/StringExtras.h"
16 #include "llvm/ADT/StringMap.h"
17 #include "llvm/ADT/StringSwitch.h"
18 #include "llvm/IR/DerivedTypes.h"
19 #include "llvm/IR/Function.h"
20 #include "llvm/IR/Module.h"
21 #include "llvm/IR/ValueSymbolTable.h"
22 #include "llvm/Support/CommandLine.h"
23 #include "llvm/Support/ModRef.h"
24 #include "llvm/Support/raw_ostream.h"
25
26 using namespace llvm;
27
28 static cl::opt<bool> EnableOCLManglingMismatchWA(
29 "amdgpu-enable-ocl-mangling-mismatch-workaround", cl::init(true),
30 cl::ReallyHidden,
31 cl::desc("Enable the workaround for OCL name mangling mismatch."));
32
33 namespace {
34
35 enum EManglingParam {
36 E_NONE,
37 EX_EVENT,
38 EX_FLOAT4,
39 EX_INTV4,
40 EX_RESERVEDID,
41 EX_SAMPLER,
42 EX_SIZET,
43 EX_UINT,
44 EX_UINTV4,
45 E_ANY,
46 E_CONSTPTR_ANY,
47 E_CONSTPTR_SWAPGL,
48 E_COPY,
49 E_IMAGECOORDS,
50 E_POINTEE,
51 E_SETBASE_I32,
52 E_SETBASE_U32,
53 E_MAKEBASE_UNS,
54 E_V16_OF_POINTEE,
55 E_V2_OF_POINTEE,
56 E_V3_OF_POINTEE,
57 E_V4_OF_POINTEE,
58 E_V8_OF_POINTEE,
59 E_VLTLPTR_ANY,
60 };
61
62 struct ManglingRule {
63 const char *Name;
64 unsigned char Lead[2];
65 unsigned char Param[5];
66
maxLeadIndex__anonbcd55a340111::ManglingRule67 int maxLeadIndex() const { return (std::max)(Lead[0], Lead[1]); }
getNumLeads__anonbcd55a340111::ManglingRule68 int getNumLeads() const { return (Lead[0] ? 1 : 0) + (Lead[1] ? 1 : 0); }
69
70 unsigned getNumArgs() const;
71
72 static StringMap<int> buildManglingRulesMap();
73 };
74
75 // Information about library functions with unmangled names.
76 class UnmangledFuncInfo {
77 const char *Name;
78 unsigned NumArgs;
79
80 // Table for all lib functions with unmangled names.
81 static const UnmangledFuncInfo Table[];
82
83 // Number of entries in Table.
84 static const unsigned TableSize;
85
86 static StringMap<unsigned> buildNameMap();
87
88 public:
89 using ID = AMDGPULibFunc::EFuncId;
UnmangledFuncInfo(const char * _Name,unsigned _NumArgs)90 constexpr UnmangledFuncInfo(const char *_Name, unsigned _NumArgs)
91 : Name(_Name), NumArgs(_NumArgs) {}
92 // Get index to Table by function name.
93 static bool lookup(StringRef Name, ID &Id);
toIndex(ID Id)94 static unsigned toIndex(ID Id) {
95 assert(static_cast<unsigned>(Id) >
96 static_cast<unsigned>(AMDGPULibFunc::EI_LAST_MANGLED) &&
97 "Invalid unmangled library function");
98 return static_cast<unsigned>(Id) - 1 -
99 static_cast<unsigned>(AMDGPULibFunc::EI_LAST_MANGLED);
100 }
toFuncId(unsigned Index)101 static ID toFuncId(unsigned Index) {
102 assert(Index < TableSize &&
103 "Invalid unmangled library function");
104 return static_cast<ID>(
105 Index + 1 + static_cast<unsigned>(AMDGPULibFunc::EI_LAST_MANGLED));
106 }
getNumArgs(ID Id)107 static unsigned getNumArgs(ID Id) { return Table[toIndex(Id)].NumArgs; }
getName(ID Id)108 static StringRef getName(ID Id) { return Table[toIndex(Id)].Name; }
109 };
110
getNumArgs() const111 unsigned ManglingRule::getNumArgs() const {
112 unsigned I=0;
113 while (I < (sizeof Param/sizeof Param[0]) && Param[I]) ++I;
114 return I;
115 }
116
117 // This table describes function formal argument type rules. The order of rules
118 // corresponds to the EFuncId enum at AMDGPULibFunc.h
119 //
120 // "<func name>", { <leads> }, { <param rules> }
121 // where:
122 // <leads> - list of integers that are one-based indexes of formal argument
123 // used to mangle a function name. Other argument types are derived from types
124 // of these 'leads'. The order of integers in this list correspond to the
125 // order in which these arguments are mangled in the EDG mangling scheme. The
126 // same order should be preserved for arguments in the AMDGPULibFunc structure
127 // when it is used for mangling. For example:
128 // { "vstorea_half", {3,1}, {E_ANY,EX_SIZET,E_ANY}},
129 // will be mangled in EDG scheme as vstorea_half_<3dparam>_<1stparam>
130 // When mangling from code use:
131 // AMDGPULibFunc insc;
132 // insc.param[0] = ... // describe 3rd parameter
133 // insc.param[1] = ... // describe 1rd parameter
134 //
135 // <param rules> - list of rules used to derive all of the function formal
136 // argument types. EX_ prefixed are simple types, other derived from the
137 // latest 'lead' argument type in the order of encoding from first to last.
138 // E_ANY - use prev lead type, E_CONSTPTR_ANY - make const pointer out of
139 // prev lead type, etc. see ParamIterator::getNextParam() for details.
140
141 static constexpr ManglingRule manglingRules[] = {
142 { "", {0}, {0} },
143 { "abs" , {1}, {E_ANY}},
144 { "abs_diff" , {1}, {E_ANY,E_COPY}},
145 { "acos" , {1}, {E_ANY}},
146 { "acosh" , {1}, {E_ANY}},
147 { "acospi" , {1}, {E_ANY}},
148 { "add_sat" , {1}, {E_ANY,E_COPY}},
149 { "all" , {1}, {E_ANY}},
150 { "any" , {1}, {E_ANY}},
151 { "asin" , {1}, {E_ANY}},
152 { "asinh" , {1}, {E_ANY}},
153 { "asinpi" , {1}, {E_ANY}},
154 { "async_work_group_copy" , {1}, {E_ANY,E_CONSTPTR_SWAPGL,EX_SIZET,EX_EVENT}},
155 { "async_work_group_strided_copy" , {1}, {E_ANY,E_CONSTPTR_SWAPGL,EX_SIZET,EX_SIZET,EX_EVENT}},
156 { "atan" , {1}, {E_ANY}},
157 { "atan2" , {1}, {E_ANY,E_COPY}},
158 { "atan2pi" , {1}, {E_ANY,E_COPY}},
159 { "atanh" , {1}, {E_ANY}},
160 { "atanpi" , {1}, {E_ANY}},
161 { "atomic_add" , {1}, {E_VLTLPTR_ANY,E_POINTEE}},
162 { "atomic_and" , {1}, {E_VLTLPTR_ANY,E_POINTEE}},
163 { "atomic_cmpxchg" , {1}, {E_VLTLPTR_ANY,E_POINTEE,E_POINTEE}},
164 { "atomic_dec" , {1}, {E_VLTLPTR_ANY}},
165 { "atomic_inc" , {1}, {E_VLTLPTR_ANY}},
166 { "atomic_max" , {1}, {E_VLTLPTR_ANY,E_POINTEE}},
167 { "atomic_min" , {1}, {E_VLTLPTR_ANY,E_POINTEE}},
168 { "atomic_or" , {1}, {E_VLTLPTR_ANY,E_POINTEE}},
169 { "atomic_sub" , {1}, {E_VLTLPTR_ANY,E_POINTEE}},
170 { "atomic_xchg" , {1}, {E_VLTLPTR_ANY,E_POINTEE}},
171 { "atomic_xor" , {1}, {E_VLTLPTR_ANY,E_POINTEE}},
172 { "bitselect" , {1}, {E_ANY,E_COPY,E_COPY}},
173 { "cbrt" , {1}, {E_ANY}},
174 { "ceil" , {1}, {E_ANY}},
175 { "clamp" , {1}, {E_ANY,E_COPY,E_COPY}},
176 { "clz" , {1}, {E_ANY}},
177 { "commit_read_pipe" , {1}, {E_ANY,EX_RESERVEDID}},
178 { "commit_write_pipe" , {1}, {E_ANY,EX_RESERVEDID}},
179 { "copysign" , {1}, {E_ANY,E_COPY}},
180 { "cos" , {1}, {E_ANY}},
181 { "cosh" , {1}, {E_ANY}},
182 { "cospi" , {1}, {E_ANY}},
183 { "cross" , {1}, {E_ANY,E_COPY}},
184 { "ctz" , {1}, {E_ANY}},
185 { "degrees" , {1}, {E_ANY}},
186 { "distance" , {1}, {E_ANY,E_COPY}},
187 { "divide" , {1}, {E_ANY,E_COPY}},
188 { "dot" , {1}, {E_ANY,E_COPY}},
189 { "erf" , {1}, {E_ANY}},
190 { "erfc" , {1}, {E_ANY}},
191 { "exp" , {1}, {E_ANY}},
192 { "exp10" , {1}, {E_ANY}},
193 { "exp2" , {1}, {E_ANY}},
194 { "expm1" , {1}, {E_ANY}},
195 { "fabs" , {1}, {E_ANY}},
196 { "fast_distance" , {1}, {E_ANY,E_COPY}},
197 { "fast_length" , {1}, {E_ANY}},
198 { "fast_normalize" , {1}, {E_ANY}},
199 { "fdim" , {1}, {E_ANY,E_COPY}},
200 { "floor" , {1}, {E_ANY}},
201 { "fma" , {1}, {E_ANY,E_COPY,E_COPY}},
202 { "fmax" , {1}, {E_ANY,E_COPY}},
203 { "fmin" , {1}, {E_ANY,E_COPY}},
204 { "fmod" , {1}, {E_ANY,E_COPY}},
205 { "fract" , {2}, {E_POINTEE,E_ANY}},
206 { "frexp" , {1,2}, {E_ANY,E_ANY}},
207 { "get_image_array_size" , {1}, {E_ANY}},
208 { "get_image_channel_data_type" , {1}, {E_ANY}},
209 { "get_image_channel_order" , {1}, {E_ANY}},
210 { "get_image_dim" , {1}, {E_ANY}},
211 { "get_image_height" , {1}, {E_ANY}},
212 { "get_image_width" , {1}, {E_ANY}},
213 { "get_pipe_max_packets" , {1}, {E_ANY}},
214 { "get_pipe_num_packets" , {1}, {E_ANY}},
215 { "hadd" , {1}, {E_ANY,E_COPY}},
216 { "hypot" , {1}, {E_ANY,E_COPY}},
217 { "ilogb" , {1}, {E_ANY}},
218 { "isequal" , {1}, {E_ANY,E_COPY}},
219 { "isfinite" , {1}, {E_ANY}},
220 { "isgreater" , {1}, {E_ANY,E_COPY}},
221 { "isgreaterequal" , {1}, {E_ANY,E_COPY}},
222 { "isinf" , {1}, {E_ANY}},
223 { "isless" , {1}, {E_ANY,E_COPY}},
224 { "islessequal" , {1}, {E_ANY,E_COPY}},
225 { "islessgreater" , {1}, {E_ANY,E_COPY}},
226 { "isnan" , {1}, {E_ANY}},
227 { "isnormal" , {1}, {E_ANY}},
228 { "isnotequal" , {1}, {E_ANY,E_COPY}},
229 { "isordered" , {1}, {E_ANY,E_COPY}},
230 { "isunordered" , {1}, {E_ANY,E_COPY}},
231 { "ldexp" , {1}, {E_ANY,E_SETBASE_I32}},
232 { "length" , {1}, {E_ANY}},
233 { "lgamma" , {1}, {E_ANY}},
234 { "lgamma_r" , {1,2}, {E_ANY,E_ANY}},
235 { "log" , {1}, {E_ANY}},
236 { "log10" , {1}, {E_ANY}},
237 { "log1p" , {1}, {E_ANY}},
238 { "log2" , {1}, {E_ANY}},
239 { "logb" , {1}, {E_ANY}},
240 { "mad" , {1}, {E_ANY,E_COPY,E_COPY}},
241 { "mad24" , {1}, {E_ANY,E_COPY,E_COPY}},
242 { "mad_hi" , {1}, {E_ANY,E_COPY,E_COPY}},
243 { "mad_sat" , {1}, {E_ANY,E_COPY,E_COPY}},
244 { "max" , {1}, {E_ANY,E_COPY}},
245 { "maxmag" , {1}, {E_ANY,E_COPY}},
246 { "min" , {1}, {E_ANY,E_COPY}},
247 { "minmag" , {1}, {E_ANY,E_COPY}},
248 { "mix" , {1}, {E_ANY,E_COPY,E_COPY}},
249 { "modf" , {2}, {E_POINTEE,E_ANY}},
250 { "mul24" , {1}, {E_ANY,E_COPY}},
251 { "mul_hi" , {1}, {E_ANY,E_COPY}},
252 { "nan" , {1}, {E_ANY}},
253 { "nextafter" , {1}, {E_ANY,E_COPY}},
254 { "normalize" , {1}, {E_ANY}},
255 { "popcount" , {1}, {E_ANY}},
256 { "pow" , {1}, {E_ANY,E_COPY}},
257 { "pown" , {1}, {E_ANY,E_SETBASE_I32}},
258 { "powr" , {1}, {E_ANY,E_COPY}},
259 { "prefetch" , {1}, {E_CONSTPTR_ANY,EX_SIZET}},
260 { "radians" , {1}, {E_ANY}},
261 { "recip" , {1}, {E_ANY}},
262 { "remainder" , {1}, {E_ANY,E_COPY}},
263 { "remquo" , {1,3}, {E_ANY,E_COPY,E_ANY}},
264 { "reserve_read_pipe" , {1}, {E_ANY,EX_UINT}},
265 { "reserve_write_pipe" , {1}, {E_ANY,EX_UINT}},
266 { "rhadd" , {1}, {E_ANY,E_COPY}},
267 { "rint" , {1}, {E_ANY}},
268 { "rootn" , {1}, {E_ANY,E_SETBASE_I32}},
269 { "rotate" , {1}, {E_ANY,E_COPY}},
270 { "round" , {1}, {E_ANY}},
271 { "rsqrt" , {1}, {E_ANY}},
272 { "select" , {1,3}, {E_ANY,E_COPY,E_ANY}},
273 { "shuffle" , {1,2}, {E_ANY,E_ANY}},
274 { "shuffle2" , {1,3}, {E_ANY,E_COPY,E_ANY}},
275 { "sign" , {1}, {E_ANY}},
276 { "signbit" , {1}, {E_ANY}},
277 { "sin" , {1}, {E_ANY}},
278 { "sincos" , {2}, {E_POINTEE,E_ANY}},
279 { "sinh" , {1}, {E_ANY}},
280 { "sinpi" , {1}, {E_ANY}},
281 { "smoothstep" , {1}, {E_ANY,E_COPY,E_COPY}},
282 { "sqrt" , {1}, {E_ANY}},
283 { "step" , {1}, {E_ANY,E_COPY}},
284 { "sub_group_broadcast" , {1}, {E_ANY,EX_UINT}},
285 { "sub_group_commit_read_pipe" , {1}, {E_ANY,EX_RESERVEDID}},
286 { "sub_group_commit_write_pipe" , {1}, {E_ANY,EX_RESERVEDID}},
287 { "sub_group_reduce_add" , {1}, {E_ANY}},
288 { "sub_group_reduce_max" , {1}, {E_ANY}},
289 { "sub_group_reduce_min" , {1}, {E_ANY}},
290 { "sub_group_reserve_read_pipe" , {1}, {E_ANY,EX_UINT}},
291 { "sub_group_reserve_write_pipe" , {1}, {E_ANY,EX_UINT}},
292 { "sub_group_scan_exclusive_add" , {1}, {E_ANY}},
293 { "sub_group_scan_exclusive_max" , {1}, {E_ANY}},
294 { "sub_group_scan_exclusive_min" , {1}, {E_ANY}},
295 { "sub_group_scan_inclusive_add" , {1}, {E_ANY}},
296 { "sub_group_scan_inclusive_max" , {1}, {E_ANY}},
297 { "sub_group_scan_inclusive_min" , {1}, {E_ANY}},
298 { "sub_sat" , {1}, {E_ANY,E_COPY}},
299 { "tan" , {1}, {E_ANY}},
300 { "tanh" , {1}, {E_ANY}},
301 { "tanpi" , {1}, {E_ANY}},
302 { "tgamma" , {1}, {E_ANY}},
303 { "trunc" , {1}, {E_ANY}},
304 { "upsample" , {1}, {E_ANY,E_MAKEBASE_UNS}},
305 { "vec_step" , {1}, {E_ANY}},
306 { "vstore" , {3}, {E_POINTEE,EX_SIZET,E_ANY}},
307 { "vstore16" , {3}, {E_V16_OF_POINTEE,EX_SIZET,E_ANY}},
308 { "vstore2" , {3}, {E_V2_OF_POINTEE,EX_SIZET,E_ANY}},
309 { "vstore3" , {3}, {E_V3_OF_POINTEE,EX_SIZET,E_ANY}},
310 { "vstore4" , {3}, {E_V4_OF_POINTEE,EX_SIZET,E_ANY}},
311 { "vstore8" , {3}, {E_V8_OF_POINTEE,EX_SIZET,E_ANY}},
312 { "work_group_commit_read_pipe" , {1}, {E_ANY,EX_RESERVEDID}},
313 { "work_group_commit_write_pipe" , {1}, {E_ANY,EX_RESERVEDID}},
314 { "work_group_reduce_add" , {1}, {E_ANY}},
315 { "work_group_reduce_max" , {1}, {E_ANY}},
316 { "work_group_reduce_min" , {1}, {E_ANY}},
317 { "work_group_reserve_read_pipe" , {1}, {E_ANY,EX_UINT}},
318 { "work_group_reserve_write_pipe" , {1}, {E_ANY,EX_UINT}},
319 { "work_group_scan_exclusive_add" , {1}, {E_ANY}},
320 { "work_group_scan_exclusive_max" , {1}, {E_ANY}},
321 { "work_group_scan_exclusive_min" , {1}, {E_ANY}},
322 { "work_group_scan_inclusive_add" , {1}, {E_ANY}},
323 { "work_group_scan_inclusive_max" , {1}, {E_ANY}},
324 { "work_group_scan_inclusive_min" , {1}, {E_ANY}},
325 { "write_imagef" , {1}, {E_ANY,E_IMAGECOORDS,EX_FLOAT4}},
326 { "write_imagei" , {1}, {E_ANY,E_IMAGECOORDS,EX_INTV4}},
327 { "write_imageui" , {1}, {E_ANY,E_IMAGECOORDS,EX_UINTV4}},
328 { "ncos" , {1}, {E_ANY} },
329 { "nexp2" , {1}, {E_ANY} },
330 { "nfma" , {1}, {E_ANY, E_COPY, E_COPY} },
331 { "nlog2" , {1}, {E_ANY} },
332 { "nrcp" , {1}, {E_ANY} },
333 { "nrsqrt" , {1}, {E_ANY} },
334 { "nsin" , {1}, {E_ANY} },
335 { "nsqrt" , {1}, {E_ANY} },
336 { "ftz" , {1}, {E_ANY} },
337 { "fldexp" , {1}, {E_ANY, EX_UINT} },
338 { "class" , {1}, {E_ANY, EX_UINT} },
339 { "rcbrt" , {1}, {E_ANY} },
340 };
341
342 // Library functions with unmangled name.
343 const UnmangledFuncInfo UnmangledFuncInfo::Table[] = {
344 {"__read_pipe_2", 4},
345 {"__read_pipe_4", 6},
346 {"__write_pipe_2", 4},
347 {"__write_pipe_4", 6},
348 };
349
350 const unsigned UnmangledFuncInfo::TableSize =
351 std::size(UnmangledFuncInfo::Table);
352
getRetType(AMDGPULibFunc::EFuncId id,const AMDGPULibFunc::Param (& Leads)[2])353 static AMDGPULibFunc::Param getRetType(AMDGPULibFunc::EFuncId id,
354 const AMDGPULibFunc::Param (&Leads)[2]) {
355 AMDGPULibFunc::Param Res = Leads[0];
356 // TBD - This switch may require to be extended for other intrinsics
357 switch (id) {
358 case AMDGPULibFunc::EI_SINCOS:
359 Res.PtrKind = AMDGPULibFunc::BYVALUE;
360 break;
361 default:
362 break;
363 }
364 return Res;
365 }
366
367 class ParamIterator {
368 const AMDGPULibFunc::Param (&Leads)[2];
369 const ManglingRule& Rule;
370 int Index = 0;
371 public:
ParamIterator(const AMDGPULibFunc::Param (& leads)[2],const ManglingRule & rule)372 ParamIterator(const AMDGPULibFunc::Param (&leads)[2],
373 const ManglingRule& rule)
374 : Leads(leads), Rule(rule) {}
375
376 AMDGPULibFunc::Param getNextParam();
377 };
378
getNextParam()379 AMDGPULibFunc::Param ParamIterator::getNextParam() {
380 AMDGPULibFunc::Param P;
381 if (Index >= int(sizeof Rule.Param/sizeof Rule.Param[0])) return P;
382
383 const char R = Rule.Param[Index];
384 switch (R) {
385 case E_NONE: break;
386 case EX_UINT:
387 P.ArgType = AMDGPULibFunc::U32; break;
388 case EX_INTV4:
389 P.ArgType = AMDGPULibFunc::I32; P.VectorSize = 4; break;
390 case EX_UINTV4:
391 P.ArgType = AMDGPULibFunc::U32; P.VectorSize = 4; break;
392 case EX_FLOAT4:
393 P.ArgType = AMDGPULibFunc::F32; P.VectorSize = 4; break;
394 case EX_SIZET:
395 P.ArgType = AMDGPULibFunc::U64; break;
396 case EX_EVENT:
397 P.ArgType = AMDGPULibFunc::EVENT; break;
398 case EX_SAMPLER:
399 P.ArgType = AMDGPULibFunc::SAMPLER; break;
400 case EX_RESERVEDID: break; // TBD
401 default:
402 if (Index == (Rule.Lead[1] - 1)) P = Leads[1];
403 else P = Leads[0];
404
405 switch (R) {
406 case E_ANY:
407 case E_COPY: break;
408
409 case E_POINTEE:
410 P.PtrKind = AMDGPULibFunc::BYVALUE; break;
411 case E_V2_OF_POINTEE:
412 P.VectorSize = 2; P.PtrKind = AMDGPULibFunc::BYVALUE; break;
413 case E_V3_OF_POINTEE:
414 P.VectorSize = 3; P.PtrKind = AMDGPULibFunc::BYVALUE; break;
415 case E_V4_OF_POINTEE:
416 P.VectorSize = 4; P.PtrKind = AMDGPULibFunc::BYVALUE; break;
417 case E_V8_OF_POINTEE:
418 P.VectorSize = 8; P.PtrKind = AMDGPULibFunc::BYVALUE; break;
419 case E_V16_OF_POINTEE:
420 P.VectorSize = 16; P.PtrKind = AMDGPULibFunc::BYVALUE; break;
421 case E_CONSTPTR_ANY:
422 P.PtrKind |= AMDGPULibFunc::CONST; break;
423 case E_VLTLPTR_ANY:
424 P.PtrKind |= AMDGPULibFunc::VOLATILE; break;
425 case E_SETBASE_I32:
426 P.ArgType = AMDGPULibFunc::I32; break;
427 case E_SETBASE_U32:
428 P.ArgType = AMDGPULibFunc::U32; break;
429
430 case E_MAKEBASE_UNS:
431 P.ArgType &= ~AMDGPULibFunc::BASE_TYPE_MASK;
432 P.ArgType |= AMDGPULibFunc::UINT;
433 break;
434
435 case E_IMAGECOORDS:
436 switch (P.ArgType) {
437 case AMDGPULibFunc::IMG1DA: P.VectorSize = 2; break;
438 case AMDGPULibFunc::IMG1DB: P.VectorSize = 1; break;
439 case AMDGPULibFunc::IMG2DA: P.VectorSize = 4; break;
440 case AMDGPULibFunc::IMG1D: P.VectorSize = 1; break;
441 case AMDGPULibFunc::IMG2D: P.VectorSize = 2; break;
442 case AMDGPULibFunc::IMG3D: P.VectorSize = 4; break;
443 }
444 P.PtrKind = AMDGPULibFunc::BYVALUE;
445 P.ArgType = AMDGPULibFunc::I32;
446 break;
447
448 case E_CONSTPTR_SWAPGL: {
449 unsigned AS = AMDGPULibFunc::getAddrSpaceFromEPtrKind(P.PtrKind);
450 switch (AS) {
451 case AMDGPUAS::GLOBAL_ADDRESS: AS = AMDGPUAS::LOCAL_ADDRESS; break;
452 case AMDGPUAS::LOCAL_ADDRESS: AS = AMDGPUAS::GLOBAL_ADDRESS; break;
453 }
454 P.PtrKind = AMDGPULibFunc::getEPtrKindFromAddrSpace(AS);
455 P.PtrKind |= AMDGPULibFunc::CONST;
456 break;
457 }
458
459 default:
460 llvm_unreachable("Unhandled param rule");
461 }
462 }
463 ++Index;
464 return P;
465 }
466
drop_front(StringRef & str,size_t n=1)467 inline static void drop_front(StringRef& str, size_t n = 1) {
468 str = str.drop_front(n);
469 }
470
eatTerm(StringRef & mangledName,const char c)471 static bool eatTerm(StringRef& mangledName, const char c) {
472 if (mangledName.front() == c) {
473 drop_front(mangledName);
474 return true;
475 }
476 return false;
477 }
478
479 template <size_t N>
eatTerm(StringRef & mangledName,const char (& str)[N])480 static bool eatTerm(StringRef& mangledName, const char (&str)[N]) {
481 if (mangledName.starts_with(StringRef(str, N - 1))) {
482 drop_front(mangledName, N-1);
483 return true;
484 }
485 return false;
486 }
487
eatNumber(StringRef & s)488 static int eatNumber(StringRef& s) {
489 size_t const savedSize = s.size();
490 int n = 0;
491 while (!s.empty() && isDigit(s.front())) {
492 n = n*10 + s.front() - '0';
493 drop_front(s);
494 }
495 return s.size() < savedSize ? n : -1;
496 }
497
eatLengthPrefixedName(StringRef & mangledName)498 static StringRef eatLengthPrefixedName(StringRef& mangledName) {
499 int const Len = eatNumber(mangledName);
500 if (Len <= 0 || static_cast<size_t>(Len) > mangledName.size())
501 return StringRef();
502 StringRef Res = mangledName.substr(0, Len);
503 drop_front(mangledName, Len);
504 return Res;
505 }
506
507 } // end anonymous namespace
508
AMDGPUMangledLibFunc()509 AMDGPUMangledLibFunc::AMDGPUMangledLibFunc() {
510 FuncId = EI_NONE;
511 FKind = NOPFX;
512 Leads[0].reset();
513 Leads[1].reset();
514 Name.clear();
515 }
516
AMDGPUUnmangledLibFunc()517 AMDGPUUnmangledLibFunc::AMDGPUUnmangledLibFunc() {
518 FuncId = EI_NONE;
519 FuncTy = nullptr;
520 }
521
AMDGPUMangledLibFunc(EFuncId id,const AMDGPUMangledLibFunc & copyFrom)522 AMDGPUMangledLibFunc::AMDGPUMangledLibFunc(
523 EFuncId id, const AMDGPUMangledLibFunc ©From) {
524 FuncId = id;
525 FKind = copyFrom.FKind;
526 Leads[0] = copyFrom.Leads[0];
527 Leads[1] = copyFrom.Leads[1];
528 }
529
AMDGPUMangledLibFunc(EFuncId id,FunctionType * FT,bool SignedInts)530 AMDGPUMangledLibFunc::AMDGPUMangledLibFunc(EFuncId id, FunctionType *FT,
531 bool SignedInts) {
532 FuncId = id;
533 unsigned NumArgs = FT->getNumParams();
534 if (NumArgs >= 1)
535 Leads[0] = Param::getFromTy(FT->getParamType(0), SignedInts);
536 if (NumArgs >= 2)
537 Leads[1] = Param::getFromTy(FT->getParamType(1), SignedInts);
538 }
539
540 ///////////////////////////////////////////////////////////////////////////////
541 // Demangling
542
parseVecSize(StringRef & mangledName)543 static int parseVecSize(StringRef& mangledName) {
544 size_t const Len = eatNumber(mangledName);
545 switch (Len) {
546 case 2: case 3: case 4: case 8: case 16:
547 return Len;
548 default:
549 break;
550 }
551 return 1;
552 }
553
parseNamePrefix(StringRef & mangledName)554 static AMDGPULibFunc::ENamePrefix parseNamePrefix(StringRef& mangledName) {
555 std::pair<StringRef, StringRef> const P = mangledName.split('_');
556 AMDGPULibFunc::ENamePrefix Pfx =
557 StringSwitch<AMDGPULibFunc::ENamePrefix>(P.first)
558 .Case("native", AMDGPULibFunc::NATIVE)
559 .Case("half" , AMDGPULibFunc::HALF)
560 .Default(AMDGPULibFunc::NOPFX);
561
562 if (Pfx != AMDGPULibFunc::NOPFX)
563 mangledName = P.second;
564
565 return Pfx;
566 }
567
buildManglingRulesMap()568 StringMap<int> ManglingRule::buildManglingRulesMap() {
569 StringMap<int> Map(std::size(manglingRules));
570 int Id = 0;
571 for (auto Rule : manglingRules)
572 Map.insert({Rule.Name, Id++});
573 return Map;
574 }
575
parseUnmangledName(StringRef FullName)576 bool AMDGPUMangledLibFunc::parseUnmangledName(StringRef FullName) {
577 static const StringMap<int> manglingRulesMap =
578 ManglingRule::buildManglingRulesMap();
579 FuncId = static_cast<EFuncId>(manglingRulesMap.lookup(FullName));
580 return FuncId != EI_NONE;
581 }
582
583 ///////////////////////////////////////////////////////////////////////////////
584 // Itanium Demangling
585
586 namespace {
587 struct ItaniumParamParser {
588 AMDGPULibFunc::Param Prev;
589 bool parseItaniumParam(StringRef& param, AMDGPULibFunc::Param &res);
590 };
591 } // namespace
592
parseItaniumParam(StringRef & param,AMDGPULibFunc::Param & res)593 bool ItaniumParamParser::parseItaniumParam(StringRef& param,
594 AMDGPULibFunc::Param &res) {
595 res.reset();
596 if (param.empty()) return false;
597
598 // parse pointer prefix
599 if (eatTerm(param, 'P')) {
600 if (eatTerm(param, 'K')) res.PtrKind |= AMDGPULibFunc::CONST;
601 if (eatTerm(param, 'V')) res.PtrKind |= AMDGPULibFunc::VOLATILE;
602 unsigned AS;
603 if (!eatTerm(param, "U3AS")) {
604 AS = 0;
605 } else {
606 AS = param.front() - '0';
607 drop_front(param, 1);
608 }
609 res.PtrKind |= AMDGPULibFuncBase::getEPtrKindFromAddrSpace(AS);
610 } else {
611 res.PtrKind = AMDGPULibFunc::BYVALUE;
612 }
613
614 // parse vector size
615 if (eatTerm(param,"Dv")) {
616 res.VectorSize = parseVecSize(param);
617 if (res.VectorSize==1 || !eatTerm(param, '_')) return false;
618 }
619
620 // parse type
621 char const TC = param.front();
622 if (isDigit(TC)) {
623 res.ArgType = StringSwitch<AMDGPULibFunc::EType>
624 (eatLengthPrefixedName(param))
625 .Case("ocl_image1darray" , AMDGPULibFunc::IMG1DA)
626 .Case("ocl_image1dbuffer", AMDGPULibFunc::IMG1DB)
627 .Case("ocl_image2darray" , AMDGPULibFunc::IMG2DA)
628 .Case("ocl_image1d" , AMDGPULibFunc::IMG1D)
629 .Case("ocl_image2d" , AMDGPULibFunc::IMG2D)
630 .Case("ocl_image3d" , AMDGPULibFunc::IMG3D)
631 .Case("ocl_event" , AMDGPULibFunc::DUMMY)
632 .Case("ocl_sampler" , AMDGPULibFunc::DUMMY)
633 .Default(AMDGPULibFunc::DUMMY);
634 } else {
635 drop_front(param);
636 switch (TC) {
637 case 'h': res.ArgType = AMDGPULibFunc::U8; break;
638 case 't': res.ArgType = AMDGPULibFunc::U16; break;
639 case 'j': res.ArgType = AMDGPULibFunc::U32; break;
640 case 'm': res.ArgType = AMDGPULibFunc::U64; break;
641 case 'c': res.ArgType = AMDGPULibFunc::I8; break;
642 case 's': res.ArgType = AMDGPULibFunc::I16; break;
643 case 'i': res.ArgType = AMDGPULibFunc::I32; break;
644 case 'l': res.ArgType = AMDGPULibFunc::I64; break;
645 case 'f': res.ArgType = AMDGPULibFunc::F32; break;
646 case 'd': res.ArgType = AMDGPULibFunc::F64; break;
647 case 'D': if (!eatTerm(param, 'h')) return false;
648 res.ArgType = AMDGPULibFunc::F16; break;
649 case 'S':
650 if (!eatTerm(param, '_')) {
651 eatNumber(param);
652 if (!eatTerm(param, '_')) return false;
653 }
654 res.VectorSize = Prev.VectorSize;
655 res.ArgType = Prev.ArgType;
656 break;
657 default:;
658 }
659 }
660 if (res.ArgType == 0) return false;
661 Prev.VectorSize = res.VectorSize;
662 Prev.ArgType = res.ArgType;
663 return true;
664 }
665
parseFuncName(StringRef & mangledName)666 bool AMDGPUMangledLibFunc::parseFuncName(StringRef &mangledName) {
667 StringRef Name = eatLengthPrefixedName(mangledName);
668 FKind = parseNamePrefix(Name);
669 if (!parseUnmangledName(Name))
670 return false;
671
672 const ManglingRule& Rule = manglingRules[FuncId];
673 ItaniumParamParser Parser;
674 for (int I=0; I < Rule.maxLeadIndex(); ++I) {
675 Param P;
676 if (!Parser.parseItaniumParam(mangledName, P))
677 return false;
678
679 if ((I + 1) == Rule.Lead[0]) Leads[0] = P;
680 if ((I + 1) == Rule.Lead[1]) Leads[1] = P;
681 }
682 return true;
683 }
684
parseFuncName(StringRef & Name)685 bool AMDGPUUnmangledLibFunc::parseFuncName(StringRef &Name) {
686 if (!UnmangledFuncInfo::lookup(Name, FuncId))
687 return false;
688 setName(Name);
689 return true;
690 }
691
parse(StringRef FuncName,AMDGPULibFunc & F)692 bool AMDGPULibFunc::parse(StringRef FuncName, AMDGPULibFunc &F) {
693 if (FuncName.empty()) {
694 F.Impl = std::unique_ptr<AMDGPULibFuncImpl>();
695 return false;
696 }
697
698 if (eatTerm(FuncName, "_Z"))
699 F.Impl = std::make_unique<AMDGPUMangledLibFunc>();
700 else
701 F.Impl = std::make_unique<AMDGPUUnmangledLibFunc>();
702 if (F.Impl->parseFuncName(FuncName))
703 return true;
704
705 F.Impl = std::unique_ptr<AMDGPULibFuncImpl>();
706 return false;
707 }
708
getUnmangledName(StringRef mangledName)709 StringRef AMDGPUMangledLibFunc::getUnmangledName(StringRef mangledName) {
710 StringRef S = mangledName;
711 if (eatTerm(S, "_Z"))
712 return eatLengthPrefixedName(S);
713 return StringRef();
714 }
715
716 ///////////////////////////////////////////////////////////////////////////////
717 // Mangling
718
719 template <typename Stream>
writeName(Stream & OS) const720 void AMDGPUMangledLibFunc::writeName(Stream &OS) const {
721 const char *Pfx = "";
722 switch (FKind) {
723 case NATIVE: Pfx = "native_"; break;
724 case HALF: Pfx = "half_"; break;
725 default: break;
726 }
727 if (!Name.empty()) {
728 OS << Pfx << Name;
729 } else if (FuncId != EI_NONE) {
730 OS << Pfx;
731 const StringRef& S = manglingRules[FuncId].Name;
732 OS.write(S.data(), S.size());
733 }
734 }
735
mangle() const736 std::string AMDGPUMangledLibFunc::mangle() const { return mangleNameItanium(); }
737
738 ///////////////////////////////////////////////////////////////////////////////
739 // Itanium Mangling
740
getItaniumTypeName(AMDGPULibFunc::EType T)741 static const char *getItaniumTypeName(AMDGPULibFunc::EType T) {
742 switch (T) {
743 case AMDGPULibFunc::U8: return "h";
744 case AMDGPULibFunc::U16: return "t";
745 case AMDGPULibFunc::U32: return "j";
746 case AMDGPULibFunc::U64: return "m";
747 case AMDGPULibFunc::I8: return "c";
748 case AMDGPULibFunc::I16: return "s";
749 case AMDGPULibFunc::I32: return "i";
750 case AMDGPULibFunc::I64: return "l";
751 case AMDGPULibFunc::F16: return "Dh";
752 case AMDGPULibFunc::F32: return "f";
753 case AMDGPULibFunc::F64: return "d";
754 case AMDGPULibFunc::IMG1DA: return "16ocl_image1darray";
755 case AMDGPULibFunc::IMG1DB: return "17ocl_image1dbuffer";
756 case AMDGPULibFunc::IMG2DA: return "16ocl_image2darray";
757 case AMDGPULibFunc::IMG1D: return "11ocl_image1d";
758 case AMDGPULibFunc::IMG2D: return "11ocl_image2d";
759 case AMDGPULibFunc::IMG3D: return "11ocl_image3d";
760 case AMDGPULibFunc::SAMPLER: return "11ocl_sampler";
761 case AMDGPULibFunc::EVENT: return "9ocl_event";
762 default:
763 llvm_unreachable("Unhandled param type");
764 }
765 return nullptr;
766 }
767
768 namespace {
769 // Itanium mangling ABI says:
770 // "5.1.8. Compression
771 // ... Each non-terminal in the grammar for which <substitution> appears on the
772 // right-hand side is both a source of future substitutions and a candidate
773 // for being substituted. There are two exceptions that appear to be
774 // substitution candidates from the grammar, but are explicitly excluded:
775 // 1. <builtin-type> other than vendor extended types ..."
776
777 // For the purpose of functions the following productions make sense for the
778 // substitution:
779 // <type> ::= <builtin-type>
780 // ::= <class-enum-type>
781 // ::= <array-type>
782 // ::=<CV-qualifiers> <type>
783 // ::= P <type> # pointer-to
784 // ::= <substitution>
785 //
786 // Note that while types like images, samplers and events are by the ABI encoded
787 // using <class-enum-type> production rule they're not used for substitution
788 // because clang consider them as builtin types.
789 //
790 // DvNN_ type is GCC extension for vectors and is a subject for the
791 // substitution.
792
793 class ItaniumMangler {
794 SmallVector<AMDGPULibFunc::Param, 10> Str; // list of accumulated substitutions
795 bool UseAddrSpace;
796
findSubst(const AMDGPULibFunc::Param & P) const797 int findSubst(const AMDGPULibFunc::Param& P) const {
798 for(unsigned I = 0; I < Str.size(); ++I) {
799 const AMDGPULibFunc::Param& T = Str[I];
800 if (P.PtrKind == T.PtrKind &&
801 P.VectorSize == T.VectorSize &&
802 P.ArgType == T.ArgType) {
803 return I;
804 }
805 }
806 return -1;
807 }
808
809 template <typename Stream>
trySubst(Stream & os,const AMDGPULibFunc::Param & p)810 bool trySubst(Stream& os, const AMDGPULibFunc::Param& p) {
811 int const subst = findSubst(p);
812 if (subst < 0) return false;
813 // Substitutions are mangled as S(XX)?_ where XX is a hexadecimal number
814 // 0 1 2
815 // S_ S0_ S1_
816 if (subst == 0) os << "S_";
817 else os << 'S' << (subst-1) << '_';
818 return true;
819 }
820
821 public:
ItaniumMangler(bool useAddrSpace)822 ItaniumMangler(bool useAddrSpace)
823 : UseAddrSpace(useAddrSpace) {}
824
825 template <typename Stream>
operator ()(Stream & os,AMDGPULibFunc::Param p)826 void operator()(Stream& os, AMDGPULibFunc::Param p) {
827
828 // Itanium mangling ABI 5.1.8. Compression:
829 // Logically, the substitutable components of a mangled name are considered
830 // left-to-right, components before the composite structure of which they
831 // are a part. If a component has been encountered before, it is substituted
832 // as described below. This decision is independent of whether its components
833 // have been substituted, so an implementation may optimize by considering
834 // large structures for substitution before their components. If a component
835 // has not been encountered before, its mangling is identified, and it is
836 // added to a dictionary of substitution candidates. No entity is added to
837 // the dictionary twice.
838 AMDGPULibFunc::Param Ptr;
839
840 if (p.PtrKind) {
841 if (trySubst(os, p)) return;
842 os << 'P';
843 if (p.PtrKind & AMDGPULibFunc::CONST) os << 'K';
844 if (p.PtrKind & AMDGPULibFunc::VOLATILE) os << 'V';
845 unsigned AS = UseAddrSpace
846 ? AMDGPULibFuncBase::getAddrSpaceFromEPtrKind(p.PtrKind)
847 : 0;
848 if (EnableOCLManglingMismatchWA || AS != 0)
849 os << "U3AS" << AS;
850 Ptr = p;
851 p.PtrKind = 0;
852 }
853
854 if (p.VectorSize > 1) {
855 if (trySubst(os, p)) goto exit;
856 Str.push_back(p);
857 os << "Dv" << static_cast<unsigned>(p.VectorSize) << '_';
858 }
859
860 os << getItaniumTypeName((AMDGPULibFunc::EType)p.ArgType);
861
862 exit:
863 if (Ptr.ArgType) Str.push_back(Ptr);
864 }
865 };
866 } // namespace
867
mangleNameItanium() const868 std::string AMDGPUMangledLibFunc::mangleNameItanium() const {
869 SmallString<128> Buf;
870 raw_svector_ostream S(Buf);
871 SmallString<128> NameBuf;
872 raw_svector_ostream Name(NameBuf);
873 writeName(Name);
874 const StringRef& NameStr = Name.str();
875 S << "_Z" << static_cast<int>(NameStr.size()) << NameStr;
876
877 ItaniumMangler Mangler(true);
878 ParamIterator I(Leads, manglingRules[FuncId]);
879 Param P;
880 while ((P = I.getNextParam()).ArgType != 0)
881 Mangler(S, P);
882 return std::string(S.str());
883 }
884
885 ///////////////////////////////////////////////////////////////////////////////
886 // Misc
887
getFromTy(Type * Ty,bool Signed)888 AMDGPULibFuncBase::Param AMDGPULibFuncBase::Param::getFromTy(Type *Ty,
889 bool Signed) {
890 Param P;
891 if (FixedVectorType *VT = dyn_cast<FixedVectorType>(Ty)) {
892 P.VectorSize = VT->getNumElements();
893 Ty = VT->getElementType();
894 }
895
896 switch (Ty->getTypeID()) {
897 case Type::FloatTyID:
898 P.ArgType = AMDGPULibFunc::F32;
899 break;
900 case Type::DoubleTyID:
901 P.ArgType = AMDGPULibFunc::F64;
902 break;
903 case Type::HalfTyID:
904 P.ArgType = AMDGPULibFunc::F16;
905 break;
906 case Type::IntegerTyID:
907 switch (cast<IntegerType>(Ty)->getBitWidth()) {
908 case 8:
909 P.ArgType = Signed ? AMDGPULibFunc::I8 : AMDGPULibFunc::U8;
910 break;
911 case 16:
912 P.ArgType = Signed ? AMDGPULibFunc::I16 : AMDGPULibFunc::U16;
913 break;
914 case 32:
915 P.ArgType = Signed ? AMDGPULibFunc::I32 : AMDGPULibFunc::U32;
916 break;
917 case 64:
918 P.ArgType = Signed ? AMDGPULibFunc::I64 : AMDGPULibFunc::U64;
919 break;
920 default:
921 llvm_unreachable("unhandled libcall argument type");
922 }
923
924 break;
925 default:
926 llvm_unreachable("unhandled libcall argument type");
927 }
928
929 return P;
930 }
931
getIntrinsicParamType(LLVMContext & C,const AMDGPULibFunc::Param & P,bool useAddrSpace)932 static Type* getIntrinsicParamType(
933 LLVMContext& C,
934 const AMDGPULibFunc::Param& P,
935 bool useAddrSpace) {
936 Type* T = nullptr;
937 switch (P.ArgType) {
938 case AMDGPULibFunc::U8:
939 case AMDGPULibFunc::I8: T = Type::getInt8Ty(C); break;
940 case AMDGPULibFunc::U16:
941 case AMDGPULibFunc::I16: T = Type::getInt16Ty(C); break;
942 case AMDGPULibFunc::U32:
943 case AMDGPULibFunc::I32: T = Type::getInt32Ty(C); break;
944 case AMDGPULibFunc::U64:
945 case AMDGPULibFunc::I64: T = Type::getInt64Ty(C); break;
946 case AMDGPULibFunc::F16: T = Type::getHalfTy(C); break;
947 case AMDGPULibFunc::F32: T = Type::getFloatTy(C); break;
948 case AMDGPULibFunc::F64: T = Type::getDoubleTy(C); break;
949
950 case AMDGPULibFunc::IMG1DA:
951 case AMDGPULibFunc::IMG1DB:
952 case AMDGPULibFunc::IMG2DA:
953 case AMDGPULibFunc::IMG1D:
954 case AMDGPULibFunc::IMG2D:
955 case AMDGPULibFunc::IMG3D:
956 T = StructType::create(C,"ocl_image")->getPointerTo(); break;
957 case AMDGPULibFunc::SAMPLER:
958 T = StructType::create(C,"ocl_sampler")->getPointerTo(); break;
959 case AMDGPULibFunc::EVENT:
960 T = StructType::create(C,"ocl_event")->getPointerTo(); break;
961 default:
962 llvm_unreachable("Unhandled param type");
963 return nullptr;
964 }
965 if (P.VectorSize > 1)
966 T = FixedVectorType::get(T, P.VectorSize);
967 if (P.PtrKind != AMDGPULibFunc::BYVALUE)
968 T = useAddrSpace ? T->getPointerTo((P.PtrKind & AMDGPULibFunc::ADDR_SPACE)
969 - 1)
970 : T->getPointerTo();
971 return T;
972 }
973
getFunctionType(Module & M) const974 FunctionType *AMDGPUMangledLibFunc::getFunctionType(Module &M) const {
975 LLVMContext& C = M.getContext();
976 std::vector<Type*> Args;
977 ParamIterator I(Leads, manglingRules[FuncId]);
978 Param P;
979 while ((P=I.getNextParam()).ArgType != 0)
980 Args.push_back(getIntrinsicParamType(C, P, true));
981
982 return FunctionType::get(
983 getIntrinsicParamType(C, getRetType(FuncId, Leads), true),
984 Args, false);
985 }
986
getNumArgs() const987 unsigned AMDGPUMangledLibFunc::getNumArgs() const {
988 return manglingRules[FuncId].getNumArgs();
989 }
990
getNumArgs() const991 unsigned AMDGPUUnmangledLibFunc::getNumArgs() const {
992 return UnmangledFuncInfo::getNumArgs(FuncId);
993 }
994
getName() const995 std::string AMDGPUMangledLibFunc::getName() const {
996 SmallString<128> Buf;
997 raw_svector_ostream OS(Buf);
998 writeName(OS);
999 return std::string(OS.str());
1000 }
1001
isCompatibleSignature(const FunctionType * FuncTy) const1002 bool AMDGPULibFunc::isCompatibleSignature(const FunctionType *FuncTy) const {
1003 // TODO: Validate types make sense
1004 return !FuncTy->isVarArg() && FuncTy->getNumParams() == getNumArgs();
1005 }
1006
getFunction(Module * M,const AMDGPULibFunc & fInfo)1007 Function *AMDGPULibFunc::getFunction(Module *M, const AMDGPULibFunc &fInfo) {
1008 std::string FuncName = fInfo.mangle();
1009 Function *F = dyn_cast_or_null<Function>(
1010 M->getValueSymbolTable().lookup(FuncName));
1011 if (!F || F->isDeclaration())
1012 return nullptr;
1013
1014 if (F->hasFnAttribute(Attribute::NoBuiltin))
1015 return nullptr;
1016
1017 if (!fInfo.isCompatibleSignature(F->getFunctionType()))
1018 return nullptr;
1019
1020 return F;
1021 }
1022
getOrInsertFunction(Module * M,const AMDGPULibFunc & fInfo)1023 FunctionCallee AMDGPULibFunc::getOrInsertFunction(Module *M,
1024 const AMDGPULibFunc &fInfo) {
1025 std::string const FuncName = fInfo.mangle();
1026 Function *F = dyn_cast_or_null<Function>(
1027 M->getValueSymbolTable().lookup(FuncName));
1028
1029 if (F) {
1030 if (F->hasFnAttribute(Attribute::NoBuiltin))
1031 return nullptr;
1032 if (!F->isDeclaration() &&
1033 fInfo.isCompatibleSignature(F->getFunctionType()))
1034 return F;
1035 }
1036
1037 FunctionType *FuncTy = fInfo.getFunctionType(*M);
1038
1039 bool hasPtr = false;
1040 for (FunctionType::param_iterator
1041 PI = FuncTy->param_begin(),
1042 PE = FuncTy->param_end();
1043 PI != PE; ++PI) {
1044 const Type* argTy = static_cast<const Type*>(*PI);
1045 if (argTy->isPointerTy()) {
1046 hasPtr = true;
1047 break;
1048 }
1049 }
1050
1051 FunctionCallee C;
1052 if (hasPtr) {
1053 // Do not set extra attributes for functions with pointer arguments.
1054 C = M->getOrInsertFunction(FuncName, FuncTy);
1055 } else {
1056 AttributeList Attr;
1057 LLVMContext &Ctx = M->getContext();
1058 Attr = Attr.addFnAttribute(
1059 Ctx, Attribute::getWithMemoryEffects(Ctx, MemoryEffects::readOnly()));
1060 Attr = Attr.addFnAttribute(Ctx, Attribute::NoUnwind);
1061 C = M->getOrInsertFunction(FuncName, FuncTy, Attr);
1062 }
1063
1064 return C;
1065 }
1066
buildNameMap()1067 StringMap<unsigned> UnmangledFuncInfo::buildNameMap() {
1068 StringMap<unsigned> Map;
1069 for (unsigned I = 0; I != TableSize; ++I)
1070 Map[Table[I].Name] = I;
1071 return Map;
1072 }
1073
lookup(StringRef Name,ID & Id)1074 bool UnmangledFuncInfo::lookup(StringRef Name, ID &Id) {
1075 static const StringMap<unsigned> Map = buildNameMap();
1076 auto Loc = Map.find(Name);
1077 if (Loc != Map.end()) {
1078 Id = toFuncId(Loc->second);
1079 return true;
1080 }
1081 Id = AMDGPULibFunc::EI_NONE;
1082 return false;
1083 }
1084
AMDGPULibFunc(const AMDGPULibFunc & F)1085 AMDGPULibFunc::AMDGPULibFunc(const AMDGPULibFunc &F) {
1086 if (auto *MF = dyn_cast<AMDGPUMangledLibFunc>(F.Impl.get()))
1087 Impl = std::make_unique<AMDGPUMangledLibFunc>(*MF);
1088 else if (auto *UMF = dyn_cast<AMDGPUUnmangledLibFunc>(F.Impl.get()))
1089 Impl = std::make_unique<AMDGPUUnmangledLibFunc>(*UMF);
1090 else
1091 Impl = std::unique_ptr<AMDGPULibFuncImpl>();
1092 }
1093
operator =(const AMDGPULibFunc & F)1094 AMDGPULibFunc &AMDGPULibFunc::operator=(const AMDGPULibFunc &F) {
1095 if (this == &F)
1096 return *this;
1097 new (this) AMDGPULibFunc(F);
1098 return *this;
1099 }
1100
AMDGPULibFunc(EFuncId Id,const AMDGPULibFunc & CopyFrom)1101 AMDGPULibFunc::AMDGPULibFunc(EFuncId Id, const AMDGPULibFunc &CopyFrom) {
1102 assert(AMDGPULibFuncBase::isMangled(Id) && CopyFrom.isMangled() &&
1103 "not supported");
1104 Impl = std::make_unique<AMDGPUMangledLibFunc>(
1105 Id, *cast<AMDGPUMangledLibFunc>(CopyFrom.Impl.get()));
1106 }
1107
AMDGPULibFunc(EFuncId Id,FunctionType * FT,bool SignedInts)1108 AMDGPULibFunc::AMDGPULibFunc(EFuncId Id, FunctionType *FT, bool SignedInts) {
1109 Impl = std::make_unique<AMDGPUMangledLibFunc>(Id, FT, SignedInts);
1110 }
1111
AMDGPULibFunc(StringRef Name,FunctionType * FT)1112 AMDGPULibFunc::AMDGPULibFunc(StringRef Name, FunctionType *FT) {
1113 Impl = std::make_unique<AMDGPUUnmangledLibFunc>(Name, FT);
1114 }
1115
initMangled()1116 void AMDGPULibFunc::initMangled() {
1117 Impl = std::make_unique<AMDGPUMangledLibFunc>();
1118 }
1119
getLeads()1120 AMDGPULibFunc::Param *AMDGPULibFunc::getLeads() {
1121 if (!Impl)
1122 initMangled();
1123 return cast<AMDGPUMangledLibFunc>(Impl.get())->Leads;
1124 }
1125
getLeads() const1126 const AMDGPULibFunc::Param *AMDGPULibFunc::getLeads() const {
1127 return cast<const AMDGPUMangledLibFunc>(Impl.get())->Leads;
1128 }
1129