[ORC] Add std::tuple support to SimplePackedSerialization.
[llvm-project.git] / llvm / lib / Target / AMDGPU / AMDGPULibFunc.cpp
blob452142a701cd2b622967b1ea96fb0d8bdcb64615
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
11 //===----------------------------------------------------------------------===//
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/raw_ostream.h"
25 using namespace llvm;
27 static cl::opt<bool> EnableOCLManglingMismatchWA(
28 "amdgpu-enable-ocl-mangling-mismatch-workaround", cl::init(true),
29 cl::ReallyHidden,
30 cl::desc("Enable the workaround for OCL name mangling mismatch."));
32 namespace {
34 enum EManglingParam {
35 E_NONE,
36 EX_EVENT,
37 EX_FLOAT4,
38 EX_INTV4,
39 EX_RESERVEDID,
40 EX_SAMPLER,
41 EX_SIZET,
42 EX_UINT,
43 EX_UINTV4,
44 E_ANY,
45 E_CONSTPTR_ANY,
46 E_CONSTPTR_SWAPGL,
47 E_COPY,
48 E_IMAGECOORDS,
49 E_POINTEE,
50 E_SETBASE_I32,
51 E_SETBASE_U32,
52 E_MAKEBASE_UNS,
53 E_V16_OF_POINTEE,
54 E_V2_OF_POINTEE,
55 E_V3_OF_POINTEE,
56 E_V4_OF_POINTEE,
57 E_V8_OF_POINTEE,
58 E_VLTLPTR_ANY,
61 struct ManglingRule {
62 const char *Name;
63 unsigned char Lead[2];
64 unsigned char Param[5];
66 int maxLeadIndex() const { return (std::max)(Lead[0], Lead[1]); }
67 int getNumLeads() const { return (Lead[0] ? 1 : 0) + (Lead[1] ? 1 : 0); }
69 unsigned getNumArgs() const;
71 static StringMap<int> buildManglingRulesMap();
74 // Information about library functions with unmangled names.
75 class UnmangledFuncInfo {
76 const char *Name;
77 unsigned NumArgs;
79 // Table for all lib functions with unmangled names.
80 static const UnmangledFuncInfo Table[];
82 // Number of entries in Table.
83 static const unsigned TableSize;
85 static StringMap<unsigned> buildNameMap();
87 public:
88 using ID = AMDGPULibFunc::EFuncId;
89 constexpr UnmangledFuncInfo(const char *_Name, unsigned _NumArgs)
90 : Name(_Name), NumArgs(_NumArgs) {}
91 // Get index to Table by function name.
92 static bool lookup(StringRef Name, ID &Id);
93 static unsigned toIndex(ID Id) {
94 assert(static_cast<unsigned>(Id) >
95 static_cast<unsigned>(AMDGPULibFunc::EI_LAST_MANGLED) &&
96 "Invalid unmangled library function");
97 return static_cast<unsigned>(Id) - 1 -
98 static_cast<unsigned>(AMDGPULibFunc::EI_LAST_MANGLED);
100 static ID toFuncId(unsigned Index) {
101 assert(Index < TableSize &&
102 "Invalid unmangled library function");
103 return static_cast<ID>(
104 Index + 1 + static_cast<unsigned>(AMDGPULibFunc::EI_LAST_MANGLED));
106 static unsigned getNumArgs(ID Id) { return Table[toIndex(Id)].NumArgs; }
107 static StringRef getName(ID Id) { return Table[toIndex(Id)].Name; }
110 unsigned ManglingRule::getNumArgs() const {
111 unsigned I=0;
112 while (I < (sizeof Param/sizeof Param[0]) && Param[I]) ++I;
113 return I;
116 // This table describes function formal argument type rules. The order of rules
117 // corresponds to the EFuncId enum at AMDGPULibFunc.h
119 // "<func name>", { <leads> }, { <param rules> }
120 // where:
121 // <leads> - list of integers that are one-based indexes of formal argument
122 // used to mangle a function name. Other argument types are derived from types
123 // of these 'leads'. The order of integers in this list correspond to the
124 // order in which these arguments are mangled in the EDG mangling scheme. The
125 // same order should be preserved for arguments in the AMDGPULibFunc structure
126 // when it is used for mangling. For example:
127 // { "vstorea_half", {3,1}, {E_ANY,EX_SIZET,E_ANY}},
128 // will be mangled in EDG scheme as vstorea_half_<3dparam>_<1stparam>
129 // When mangling from code use:
130 // AMDGPULibFunc insc;
131 // insc.param[0] = ... // describe 3rd parameter
132 // insc.param[1] = ... // describe 1rd parameter
134 // <param rules> - list of rules used to derive all of the function formal
135 // argument types. EX_ prefixed are simple types, other derived from the
136 // latest 'lead' argument type in the order of encoding from first to last.
137 // E_ANY - use prev lead type, E_CONSTPTR_ANY - make const pointer out of
138 // prev lead type, etc. see ParamIterator::getNextParam() for details.
140 static constexpr ManglingRule manglingRules[] = {
141 { "", {0}, {0} },
142 { "abs" , {1}, {E_ANY}},
143 { "abs_diff" , {1}, {E_ANY,E_COPY}},
144 { "acos" , {1}, {E_ANY}},
145 { "acosh" , {1}, {E_ANY}},
146 { "acospi" , {1}, {E_ANY}},
147 { "add_sat" , {1}, {E_ANY,E_COPY}},
148 { "all" , {1}, {E_ANY}},
149 { "any" , {1}, {E_ANY}},
150 { "asin" , {1}, {E_ANY}},
151 { "asinh" , {1}, {E_ANY}},
152 { "asinpi" , {1}, {E_ANY}},
153 { "async_work_group_copy" , {1}, {E_ANY,E_CONSTPTR_SWAPGL,EX_SIZET,EX_EVENT}},
154 { "async_work_group_strided_copy" , {1}, {E_ANY,E_CONSTPTR_SWAPGL,EX_SIZET,EX_SIZET,EX_EVENT}},
155 { "atan" , {1}, {E_ANY}},
156 { "atan2" , {1}, {E_ANY,E_COPY}},
157 { "atan2pi" , {1}, {E_ANY,E_COPY}},
158 { "atanh" , {1}, {E_ANY}},
159 { "atanpi" , {1}, {E_ANY}},
160 { "atomic_add" , {1}, {E_VLTLPTR_ANY,E_POINTEE}},
161 { "atomic_and" , {1}, {E_VLTLPTR_ANY,E_POINTEE}},
162 { "atomic_cmpxchg" , {1}, {E_VLTLPTR_ANY,E_POINTEE,E_POINTEE}},
163 { "atomic_dec" , {1}, {E_VLTLPTR_ANY}},
164 { "atomic_inc" , {1}, {E_VLTLPTR_ANY}},
165 { "atomic_max" , {1}, {E_VLTLPTR_ANY,E_POINTEE}},
166 { "atomic_min" , {1}, {E_VLTLPTR_ANY,E_POINTEE}},
167 { "atomic_or" , {1}, {E_VLTLPTR_ANY,E_POINTEE}},
168 { "atomic_sub" , {1}, {E_VLTLPTR_ANY,E_POINTEE}},
169 { "atomic_xchg" , {1}, {E_VLTLPTR_ANY,E_POINTEE}},
170 { "atomic_xor" , {1}, {E_VLTLPTR_ANY,E_POINTEE}},
171 { "bitselect" , {1}, {E_ANY,E_COPY,E_COPY}},
172 { "cbrt" , {1}, {E_ANY}},
173 { "ceil" , {1}, {E_ANY}},
174 { "clamp" , {1}, {E_ANY,E_COPY,E_COPY}},
175 { "clz" , {1}, {E_ANY}},
176 { "commit_read_pipe" , {1}, {E_ANY,EX_RESERVEDID}},
177 { "commit_write_pipe" , {1}, {E_ANY,EX_RESERVEDID}},
178 { "copysign" , {1}, {E_ANY,E_COPY}},
179 { "cos" , {1}, {E_ANY}},
180 { "cosh" , {1}, {E_ANY}},
181 { "cospi" , {1}, {E_ANY}},
182 { "cross" , {1}, {E_ANY,E_COPY}},
183 { "ctz" , {1}, {E_ANY}},
184 { "degrees" , {1}, {E_ANY}},
185 { "distance" , {1}, {E_ANY,E_COPY}},
186 { "divide" , {1}, {E_ANY,E_COPY}},
187 { "dot" , {1}, {E_ANY,E_COPY}},
188 { "erf" , {1}, {E_ANY}},
189 { "erfc" , {1}, {E_ANY}},
190 { "exp" , {1}, {E_ANY}},
191 { "exp10" , {1}, {E_ANY}},
192 { "exp2" , {1}, {E_ANY}},
193 { "expm1" , {1}, {E_ANY}},
194 { "fabs" , {1}, {E_ANY}},
195 { "fast_distance" , {1}, {E_ANY,E_COPY}},
196 { "fast_length" , {1}, {E_ANY}},
197 { "fast_normalize" , {1}, {E_ANY}},
198 { "fdim" , {1}, {E_ANY,E_COPY}},
199 { "floor" , {1}, {E_ANY}},
200 { "fma" , {1}, {E_ANY,E_COPY,E_COPY}},
201 { "fmax" , {1}, {E_ANY,E_COPY}},
202 { "fmin" , {1}, {E_ANY,E_COPY}},
203 { "fmod" , {1}, {E_ANY,E_COPY}},
204 { "fract" , {2}, {E_POINTEE,E_ANY}},
205 { "frexp" , {1,2}, {E_ANY,E_ANY}},
206 { "get_image_array_size" , {1}, {E_ANY}},
207 { "get_image_channel_data_type" , {1}, {E_ANY}},
208 { "get_image_channel_order" , {1}, {E_ANY}},
209 { "get_image_dim" , {1}, {E_ANY}},
210 { "get_image_height" , {1}, {E_ANY}},
211 { "get_image_width" , {1}, {E_ANY}},
212 { "get_pipe_max_packets" , {1}, {E_ANY}},
213 { "get_pipe_num_packets" , {1}, {E_ANY}},
214 { "hadd" , {1}, {E_ANY,E_COPY}},
215 { "hypot" , {1}, {E_ANY,E_COPY}},
216 { "ilogb" , {1}, {E_ANY}},
217 { "isequal" , {1}, {E_ANY,E_COPY}},
218 { "isfinite" , {1}, {E_ANY}},
219 { "isgreater" , {1}, {E_ANY,E_COPY}},
220 { "isgreaterequal" , {1}, {E_ANY,E_COPY}},
221 { "isinf" , {1}, {E_ANY}},
222 { "isless" , {1}, {E_ANY,E_COPY}},
223 { "islessequal" , {1}, {E_ANY,E_COPY}},
224 { "islessgreater" , {1}, {E_ANY,E_COPY}},
225 { "isnan" , {1}, {E_ANY}},
226 { "isnormal" , {1}, {E_ANY}},
227 { "isnotequal" , {1}, {E_ANY,E_COPY}},
228 { "isordered" , {1}, {E_ANY,E_COPY}},
229 { "isunordered" , {1}, {E_ANY,E_COPY}},
230 { "ldexp" , {1}, {E_ANY,E_SETBASE_I32}},
231 { "length" , {1}, {E_ANY}},
232 { "lgamma" , {1}, {E_ANY}},
233 { "lgamma_r" , {1,2}, {E_ANY,E_ANY}},
234 { "log" , {1}, {E_ANY}},
235 { "log10" , {1}, {E_ANY}},
236 { "log1p" , {1}, {E_ANY}},
237 { "log2" , {1}, {E_ANY}},
238 { "logb" , {1}, {E_ANY}},
239 { "mad" , {1}, {E_ANY,E_COPY,E_COPY}},
240 { "mad24" , {1}, {E_ANY,E_COPY,E_COPY}},
241 { "mad_hi" , {1}, {E_ANY,E_COPY,E_COPY}},
242 { "mad_sat" , {1}, {E_ANY,E_COPY,E_COPY}},
243 { "max" , {1}, {E_ANY,E_COPY}},
244 { "maxmag" , {1}, {E_ANY,E_COPY}},
245 { "min" , {1}, {E_ANY,E_COPY}},
246 { "minmag" , {1}, {E_ANY,E_COPY}},
247 { "mix" , {1}, {E_ANY,E_COPY,E_COPY}},
248 { "modf" , {2}, {E_POINTEE,E_ANY}},
249 { "mul24" , {1}, {E_ANY,E_COPY}},
250 { "mul_hi" , {1}, {E_ANY,E_COPY}},
251 { "nan" , {1}, {E_ANY}},
252 { "nextafter" , {1}, {E_ANY,E_COPY}},
253 { "normalize" , {1}, {E_ANY}},
254 { "popcount" , {1}, {E_ANY}},
255 { "pow" , {1}, {E_ANY,E_COPY}},
256 { "pown" , {1}, {E_ANY,E_SETBASE_I32}},
257 { "powr" , {1}, {E_ANY,E_COPY}},
258 { "prefetch" , {1}, {E_CONSTPTR_ANY,EX_SIZET}},
259 { "radians" , {1}, {E_ANY}},
260 { "recip" , {1}, {E_ANY}},
261 { "remainder" , {1}, {E_ANY,E_COPY}},
262 { "remquo" , {1,3}, {E_ANY,E_COPY,E_ANY}},
263 { "reserve_read_pipe" , {1}, {E_ANY,EX_UINT}},
264 { "reserve_write_pipe" , {1}, {E_ANY,EX_UINT}},
265 { "rhadd" , {1}, {E_ANY,E_COPY}},
266 { "rint" , {1}, {E_ANY}},
267 { "rootn" , {1}, {E_ANY,E_SETBASE_I32}},
268 { "rotate" , {1}, {E_ANY,E_COPY}},
269 { "round" , {1}, {E_ANY}},
270 { "rsqrt" , {1}, {E_ANY}},
271 { "select" , {1,3}, {E_ANY,E_COPY,E_ANY}},
272 { "shuffle" , {1,2}, {E_ANY,E_ANY}},
273 { "shuffle2" , {1,3}, {E_ANY,E_COPY,E_ANY}},
274 { "sign" , {1}, {E_ANY}},
275 { "signbit" , {1}, {E_ANY}},
276 { "sin" , {1}, {E_ANY}},
277 { "sincos" , {2}, {E_POINTEE,E_ANY}},
278 { "sinh" , {1}, {E_ANY}},
279 { "sinpi" , {1}, {E_ANY}},
280 { "smoothstep" , {1}, {E_ANY,E_COPY,E_COPY}},
281 { "sqrt" , {1}, {E_ANY}},
282 { "step" , {1}, {E_ANY,E_COPY}},
283 { "sub_group_broadcast" , {1}, {E_ANY,EX_UINT}},
284 { "sub_group_commit_read_pipe" , {1}, {E_ANY,EX_RESERVEDID}},
285 { "sub_group_commit_write_pipe" , {1}, {E_ANY,EX_RESERVEDID}},
286 { "sub_group_reduce_add" , {1}, {E_ANY}},
287 { "sub_group_reduce_max" , {1}, {E_ANY}},
288 { "sub_group_reduce_min" , {1}, {E_ANY}},
289 { "sub_group_reserve_read_pipe" , {1}, {E_ANY,EX_UINT}},
290 { "sub_group_reserve_write_pipe" , {1}, {E_ANY,EX_UINT}},
291 { "sub_group_scan_exclusive_add" , {1}, {E_ANY}},
292 { "sub_group_scan_exclusive_max" , {1}, {E_ANY}},
293 { "sub_group_scan_exclusive_min" , {1}, {E_ANY}},
294 { "sub_group_scan_inclusive_add" , {1}, {E_ANY}},
295 { "sub_group_scan_inclusive_max" , {1}, {E_ANY}},
296 { "sub_group_scan_inclusive_min" , {1}, {E_ANY}},
297 { "sub_sat" , {1}, {E_ANY,E_COPY}},
298 { "tan" , {1}, {E_ANY}},
299 { "tanh" , {1}, {E_ANY}},
300 { "tanpi" , {1}, {E_ANY}},
301 { "tgamma" , {1}, {E_ANY}},
302 { "trunc" , {1}, {E_ANY}},
303 { "upsample" , {1}, {E_ANY,E_MAKEBASE_UNS}},
304 { "vec_step" , {1}, {E_ANY}},
305 { "vstore" , {3}, {E_POINTEE,EX_SIZET,E_ANY}},
306 { "vstore16" , {3}, {E_V16_OF_POINTEE,EX_SIZET,E_ANY}},
307 { "vstore2" , {3}, {E_V2_OF_POINTEE,EX_SIZET,E_ANY}},
308 { "vstore3" , {3}, {E_V3_OF_POINTEE,EX_SIZET,E_ANY}},
309 { "vstore4" , {3}, {E_V4_OF_POINTEE,EX_SIZET,E_ANY}},
310 { "vstore8" , {3}, {E_V8_OF_POINTEE,EX_SIZET,E_ANY}},
311 { "work_group_commit_read_pipe" , {1}, {E_ANY,EX_RESERVEDID}},
312 { "work_group_commit_write_pipe" , {1}, {E_ANY,EX_RESERVEDID}},
313 { "work_group_reduce_add" , {1}, {E_ANY}},
314 { "work_group_reduce_max" , {1}, {E_ANY}},
315 { "work_group_reduce_min" , {1}, {E_ANY}},
316 { "work_group_reserve_read_pipe" , {1}, {E_ANY,EX_UINT}},
317 { "work_group_reserve_write_pipe" , {1}, {E_ANY,EX_UINT}},
318 { "work_group_scan_exclusive_add" , {1}, {E_ANY}},
319 { "work_group_scan_exclusive_max" , {1}, {E_ANY}},
320 { "work_group_scan_exclusive_min" , {1}, {E_ANY}},
321 { "work_group_scan_inclusive_add" , {1}, {E_ANY}},
322 { "work_group_scan_inclusive_max" , {1}, {E_ANY}},
323 { "work_group_scan_inclusive_min" , {1}, {E_ANY}},
324 { "write_imagef" , {1}, {E_ANY,E_IMAGECOORDS,EX_FLOAT4}},
325 { "write_imagei" , {1}, {E_ANY,E_IMAGECOORDS,EX_INTV4}},
326 { "write_imageui" , {1}, {E_ANY,E_IMAGECOORDS,EX_UINTV4}},
327 { "ncos" , {1}, {E_ANY} },
328 { "nexp2" , {1}, {E_ANY} },
329 { "nfma" , {1}, {E_ANY, E_COPY, E_COPY} },
330 { "nlog2" , {1}, {E_ANY} },
331 { "nrcp" , {1}, {E_ANY} },
332 { "nrsqrt" , {1}, {E_ANY} },
333 { "nsin" , {1}, {E_ANY} },
334 { "nsqrt" , {1}, {E_ANY} },
335 { "ftz" , {1}, {E_ANY} },
336 { "fldexp" , {1}, {E_ANY, EX_UINT} },
337 { "class" , {1}, {E_ANY, EX_UINT} },
338 { "rcbrt" , {1}, {E_ANY} },
341 // Library functions with unmangled name.
342 const UnmangledFuncInfo UnmangledFuncInfo::Table[] = {
343 {"__read_pipe_2", 4},
344 {"__read_pipe_4", 6},
345 {"__write_pipe_2", 4},
346 {"__write_pipe_4", 6},
349 const unsigned UnmangledFuncInfo::TableSize =
350 array_lengthof(UnmangledFuncInfo::Table);
352 static AMDGPULibFunc::Param getRetType(AMDGPULibFunc::EFuncId id,
353 const AMDGPULibFunc::Param (&Leads)[2]) {
354 AMDGPULibFunc::Param Res = Leads[0];
355 // TBD - This switch may require to be extended for other intriniscs
356 switch (id) {
357 case AMDGPULibFunc::EI_SINCOS:
358 Res.PtrKind = AMDGPULibFunc::BYVALUE;
359 break;
360 default:
361 break;
363 return Res;
366 class ParamIterator {
367 const AMDGPULibFunc::Param (&Leads)[2];
368 const ManglingRule& Rule;
369 int Index;
370 public:
371 ParamIterator(const AMDGPULibFunc::Param (&leads)[2],
372 const ManglingRule& rule)
373 : Leads(leads), Rule(rule), Index(0) {}
375 AMDGPULibFunc::Param getNextParam();
378 AMDGPULibFunc::Param ParamIterator::getNextParam() {
379 AMDGPULibFunc::Param P;
380 if (Index >= int(sizeof Rule.Param/sizeof Rule.Param[0])) return P;
382 const char R = Rule.Param[Index];
383 switch (R) {
384 case E_NONE: break;
385 case EX_UINT:
386 P.ArgType = AMDGPULibFunc::U32; break;
387 case EX_INTV4:
388 P.ArgType = AMDGPULibFunc::I32; P.VectorSize = 4; break;
389 case EX_UINTV4:
390 P.ArgType = AMDGPULibFunc::U32; P.VectorSize = 4; break;
391 case EX_FLOAT4:
392 P.ArgType = AMDGPULibFunc::F32; P.VectorSize = 4; break;
393 case EX_SIZET:
394 P.ArgType = AMDGPULibFunc::U64; break;
395 case EX_EVENT:
396 P.ArgType = AMDGPULibFunc::EVENT; break;
397 case EX_SAMPLER:
398 P.ArgType = AMDGPULibFunc::SAMPLER; break;
399 case EX_RESERVEDID: break; // TBD
400 default:
401 if (Index == (Rule.Lead[1] - 1)) P = Leads[1];
402 else P = Leads[0];
404 switch (R) {
405 case E_ANY:
406 case E_COPY: break;
408 case E_POINTEE:
409 P.PtrKind = AMDGPULibFunc::BYVALUE; break;
410 case E_V2_OF_POINTEE:
411 P.VectorSize = 2; P.PtrKind = AMDGPULibFunc::BYVALUE; break;
412 case E_V3_OF_POINTEE:
413 P.VectorSize = 3; P.PtrKind = AMDGPULibFunc::BYVALUE; break;
414 case E_V4_OF_POINTEE:
415 P.VectorSize = 4; P.PtrKind = AMDGPULibFunc::BYVALUE; break;
416 case E_V8_OF_POINTEE:
417 P.VectorSize = 8; P.PtrKind = AMDGPULibFunc::BYVALUE; break;
418 case E_V16_OF_POINTEE:
419 P.VectorSize = 16; P.PtrKind = AMDGPULibFunc::BYVALUE; break;
420 case E_CONSTPTR_ANY:
421 P.PtrKind |= AMDGPULibFunc::CONST; break;
422 case E_VLTLPTR_ANY:
423 P.PtrKind |= AMDGPULibFunc::VOLATILE; break;
424 case E_SETBASE_I32:
425 P.ArgType = AMDGPULibFunc::I32; break;
426 case E_SETBASE_U32:
427 P.ArgType = AMDGPULibFunc::U32; break;
429 case E_MAKEBASE_UNS:
430 P.ArgType &= ~AMDGPULibFunc::BASE_TYPE_MASK;
431 P.ArgType |= AMDGPULibFunc::UINT;
432 break;
434 case E_IMAGECOORDS:
435 switch (P.ArgType) {
436 case AMDGPULibFunc::IMG1DA: P.VectorSize = 2; break;
437 case AMDGPULibFunc::IMG1DB: P.VectorSize = 1; break;
438 case AMDGPULibFunc::IMG2DA: P.VectorSize = 4; break;
439 case AMDGPULibFunc::IMG1D: P.VectorSize = 1; break;
440 case AMDGPULibFunc::IMG2D: P.VectorSize = 2; break;
441 case AMDGPULibFunc::IMG3D: P.VectorSize = 4; break;
443 P.PtrKind = AMDGPULibFunc::BYVALUE;
444 P.ArgType = AMDGPULibFunc::I32;
445 break;
447 case E_CONSTPTR_SWAPGL: {
448 unsigned AS = AMDGPULibFunc::getAddrSpaceFromEPtrKind(P.PtrKind);
449 switch (AS) {
450 case AMDGPUAS::GLOBAL_ADDRESS: AS = AMDGPUAS::LOCAL_ADDRESS; break;
451 case AMDGPUAS::LOCAL_ADDRESS: AS = AMDGPUAS::GLOBAL_ADDRESS; break;
453 P.PtrKind = AMDGPULibFunc::getEPtrKindFromAddrSpace(AS);
454 P.PtrKind |= AMDGPULibFunc::CONST;
455 break;
458 default: llvm_unreachable("Unhandeled param rule");
461 ++Index;
462 return P;
465 inline static void drop_front(StringRef& str, size_t n = 1) {
466 str = str.drop_front(n);
469 static bool eatTerm(StringRef& mangledName, const char c) {
470 if (mangledName.front() == c) {
471 drop_front(mangledName);
472 return true;
474 return false;
477 template <size_t N>
478 static bool eatTerm(StringRef& mangledName, const char (&str)[N]) {
479 if (mangledName.startswith(StringRef(str, N-1))) {
480 drop_front(mangledName, N-1);
481 return true;
483 return false;
486 static int eatNumber(StringRef& s) {
487 size_t const savedSize = s.size();
488 int n = 0;
489 while (!s.empty() && isDigit(s.front())) {
490 n = n*10 + s.front() - '0';
491 drop_front(s);
493 return s.size() < savedSize ? n : -1;
496 static StringRef eatLengthPrefixedName(StringRef& mangledName) {
497 int const Len = eatNumber(mangledName);
498 if (Len <= 0 || static_cast<size_t>(Len) > mangledName.size())
499 return StringRef();
500 StringRef Res = mangledName.substr(0, Len);
501 drop_front(mangledName, Len);
502 return Res;
505 } // end anonymous namespace
507 AMDGPUMangledLibFunc::AMDGPUMangledLibFunc() {
508 FuncId = EI_NONE;
509 FKind = NOPFX;
510 Leads[0].reset();
511 Leads[1].reset();
512 Name.clear();
515 AMDGPUUnmangledLibFunc::AMDGPUUnmangledLibFunc() {
516 FuncId = EI_NONE;
517 FuncTy = nullptr;
520 AMDGPUMangledLibFunc::AMDGPUMangledLibFunc(
521 EFuncId id, const AMDGPUMangledLibFunc &copyFrom) {
522 FuncId = id;
523 FKind = copyFrom.FKind;
524 Leads[0] = copyFrom.Leads[0];
525 Leads[1] = copyFrom.Leads[1];
528 ///////////////////////////////////////////////////////////////////////////////
529 // Demangling
531 static int parseVecSize(StringRef& mangledName) {
532 size_t const Len = eatNumber(mangledName);
533 switch (Len) {
534 case 2: case 3: case 4: case 8: case 16:
535 return Len;
536 default:
537 break;
539 return 1;
542 static AMDGPULibFunc::ENamePrefix parseNamePrefix(StringRef& mangledName) {
543 std::pair<StringRef, StringRef> const P = mangledName.split('_');
544 AMDGPULibFunc::ENamePrefix Pfx =
545 StringSwitch<AMDGPULibFunc::ENamePrefix>(P.first)
546 .Case("native", AMDGPULibFunc::NATIVE)
547 .Case("half" , AMDGPULibFunc::HALF)
548 .Default(AMDGPULibFunc::NOPFX);
550 if (Pfx != AMDGPULibFunc::NOPFX)
551 mangledName = P.second;
553 return Pfx;
556 StringMap<int> ManglingRule::buildManglingRulesMap() {
557 StringMap<int> Map(array_lengthof(manglingRules));
558 int Id = 0;
559 for (auto Rule : manglingRules)
560 Map.insert({Rule.Name, Id++});
561 return Map;
564 bool AMDGPUMangledLibFunc::parseUnmangledName(StringRef FullName) {
565 static const StringMap<int> manglingRulesMap =
566 ManglingRule::buildManglingRulesMap();
567 FuncId = static_cast<EFuncId>(manglingRulesMap.lookup(FullName));
568 return FuncId != EI_NONE;
571 ///////////////////////////////////////////////////////////////////////////////
572 // Itanium Demangling
574 namespace {
575 struct ItaniumParamParser {
576 AMDGPULibFunc::Param Prev;
577 bool parseItaniumParam(StringRef& param, AMDGPULibFunc::Param &res);
579 } // namespace
581 bool ItaniumParamParser::parseItaniumParam(StringRef& param,
582 AMDGPULibFunc::Param &res) {
583 res.reset();
584 if (param.empty()) return false;
586 // parse pointer prefix
587 if (eatTerm(param, 'P')) {
588 if (eatTerm(param, 'K')) res.PtrKind |= AMDGPULibFunc::CONST;
589 if (eatTerm(param, 'V')) res.PtrKind |= AMDGPULibFunc::VOLATILE;
590 unsigned AS;
591 if (!eatTerm(param, "U3AS")) {
592 AS = 0;
593 } else {
594 AS = param.front() - '0';
595 drop_front(param, 1);
597 res.PtrKind |= AMDGPULibFuncBase::getEPtrKindFromAddrSpace(AS);
598 } else {
599 res.PtrKind = AMDGPULibFunc::BYVALUE;
602 // parse vector size
603 if (eatTerm(param,"Dv")) {
604 res.VectorSize = parseVecSize(param);
605 if (res.VectorSize==1 || !eatTerm(param, '_')) return false;
608 // parse type
609 char const TC = param.front();
610 if (isDigit(TC)) {
611 res.ArgType = StringSwitch<AMDGPULibFunc::EType>
612 (eatLengthPrefixedName(param))
613 .Case("ocl_image1darray" , AMDGPULibFunc::IMG1DA)
614 .Case("ocl_image1dbuffer", AMDGPULibFunc::IMG1DB)
615 .Case("ocl_image2darray" , AMDGPULibFunc::IMG2DA)
616 .Case("ocl_image1d" , AMDGPULibFunc::IMG1D)
617 .Case("ocl_image2d" , AMDGPULibFunc::IMG2D)
618 .Case("ocl_image3d" , AMDGPULibFunc::IMG3D)
619 .Case("ocl_event" , AMDGPULibFunc::DUMMY)
620 .Case("ocl_sampler" , AMDGPULibFunc::DUMMY)
621 .Default(AMDGPULibFunc::DUMMY);
622 } else {
623 drop_front(param);
624 switch (TC) {
625 case 'h': res.ArgType = AMDGPULibFunc::U8; break;
626 case 't': res.ArgType = AMDGPULibFunc::U16; break;
627 case 'j': res.ArgType = AMDGPULibFunc::U32; break;
628 case 'm': res.ArgType = AMDGPULibFunc::U64; break;
629 case 'c': res.ArgType = AMDGPULibFunc::I8; break;
630 case 's': res.ArgType = AMDGPULibFunc::I16; break;
631 case 'i': res.ArgType = AMDGPULibFunc::I32; break;
632 case 'l': res.ArgType = AMDGPULibFunc::I64; break;
633 case 'f': res.ArgType = AMDGPULibFunc::F32; break;
634 case 'd': res.ArgType = AMDGPULibFunc::F64; break;
635 case 'D': if (!eatTerm(param, 'h')) return false;
636 res.ArgType = AMDGPULibFunc::F16; break;
637 case 'S':
638 if (!eatTerm(param, '_')) {
639 eatNumber(param);
640 if (!eatTerm(param, '_')) return false;
642 res.VectorSize = Prev.VectorSize;
643 res.ArgType = Prev.ArgType;
644 break;
645 default:;
648 if (res.ArgType == 0) return false;
649 Prev.VectorSize = res.VectorSize;
650 Prev.ArgType = res.ArgType;
651 return true;
654 bool AMDGPUMangledLibFunc::parseFuncName(StringRef &mangledName) {
655 StringRef Name = eatLengthPrefixedName(mangledName);
656 FKind = parseNamePrefix(Name);
657 if (!parseUnmangledName(Name))
658 return false;
660 const ManglingRule& Rule = manglingRules[FuncId];
661 ItaniumParamParser Parser;
662 for (int I=0; I < Rule.maxLeadIndex(); ++I) {
663 Param P;
664 if (!Parser.parseItaniumParam(mangledName, P))
665 return false;
667 if ((I + 1) == Rule.Lead[0]) Leads[0] = P;
668 if ((I + 1) == Rule.Lead[1]) Leads[1] = P;
670 return true;
673 bool AMDGPUUnmangledLibFunc::parseFuncName(StringRef &Name) {
674 if (!UnmangledFuncInfo::lookup(Name, FuncId))
675 return false;
676 setName(Name);
677 return true;
680 bool AMDGPULibFunc::parse(StringRef FuncName, AMDGPULibFunc &F) {
681 if (FuncName.empty()) {
682 F.Impl = std::unique_ptr<AMDGPULibFuncImpl>();
683 return false;
686 if (eatTerm(FuncName, "_Z"))
687 F.Impl = std::make_unique<AMDGPUMangledLibFunc>();
688 else
689 F.Impl = std::make_unique<AMDGPUUnmangledLibFunc>();
690 if (F.Impl->parseFuncName(FuncName))
691 return true;
693 F.Impl = std::unique_ptr<AMDGPULibFuncImpl>();
694 return false;
697 StringRef AMDGPUMangledLibFunc::getUnmangledName(StringRef mangledName) {
698 StringRef S = mangledName;
699 if (eatTerm(S, "_Z"))
700 return eatLengthPrefixedName(S);
701 return StringRef();
704 ///////////////////////////////////////////////////////////////////////////////
705 // Mangling
707 template <typename Stream>
708 void AMDGPUMangledLibFunc::writeName(Stream &OS) const {
709 const char *Pfx = "";
710 switch (FKind) {
711 case NATIVE: Pfx = "native_"; break;
712 case HALF: Pfx = "half_"; break;
713 default: break;
715 if (!Name.empty()) {
716 OS << Pfx << Name;
717 } else if (FuncId != EI_NONE) {
718 OS << Pfx;
719 const StringRef& S = manglingRules[FuncId].Name;
720 OS.write(S.data(), S.size());
724 std::string AMDGPUMangledLibFunc::mangle() const { return mangleNameItanium(); }
726 ///////////////////////////////////////////////////////////////////////////////
727 // Itanium Mangling
729 static const char *getItaniumTypeName(AMDGPULibFunc::EType T) {
730 switch (T) {
731 case AMDGPULibFunc::U8: return "h";
732 case AMDGPULibFunc::U16: return "t";
733 case AMDGPULibFunc::U32: return "j";
734 case AMDGPULibFunc::U64: return "m";
735 case AMDGPULibFunc::I8: return "c";
736 case AMDGPULibFunc::I16: return "s";
737 case AMDGPULibFunc::I32: return "i";
738 case AMDGPULibFunc::I64: return "l";
739 case AMDGPULibFunc::F16: return "Dh";
740 case AMDGPULibFunc::F32: return "f";
741 case AMDGPULibFunc::F64: return "d";
742 case AMDGPULibFunc::IMG1DA: return "16ocl_image1darray";
743 case AMDGPULibFunc::IMG1DB: return "17ocl_image1dbuffer";
744 case AMDGPULibFunc::IMG2DA: return "16ocl_image2darray";
745 case AMDGPULibFunc::IMG1D: return "11ocl_image1d";
746 case AMDGPULibFunc::IMG2D: return "11ocl_image2d";
747 case AMDGPULibFunc::IMG3D: return "11ocl_image3d";
748 case AMDGPULibFunc::SAMPLER: return "11ocl_sampler";
749 case AMDGPULibFunc::EVENT: return "9ocl_event";
750 default: llvm_unreachable("Unhandeled param type");
752 return nullptr;
755 namespace {
756 // Itanium mangling ABI says:
757 // "5.1.8. Compression
758 // ... Each non-terminal in the grammar for which <substitution> appears on the
759 // right-hand side is both a source of future substitutions and a candidate
760 // for being substituted. There are two exceptions that appear to be
761 // substitution candidates from the grammar, but are explicitly excluded:
762 // 1. <builtin-type> other than vendor extended types ..."
764 // For the purpose of functions the following productions make sence for the
765 // substitution:
766 // <type> ::= <builtin-type>
767 // ::= <class-enum-type>
768 // ::= <array-type>
769 // ::=<CV-qualifiers> <type>
770 // ::= P <type> # pointer-to
771 // ::= <substitution>
773 // Note that while types like images, samplers and events are by the ABI encoded
774 // using <class-enum-type> production rule they're not used for substitution
775 // because clang consider them as builtin types.
777 // DvNN_ type is GCC extension for vectors and is a subject for the substitution.
780 class ItaniumMangler {
781 SmallVector<AMDGPULibFunc::Param, 10> Str; // list of accumulated substituions
782 bool UseAddrSpace;
784 int findSubst(const AMDGPULibFunc::Param& P) const {
785 for(unsigned I = 0; I < Str.size(); ++I) {
786 const AMDGPULibFunc::Param& T = Str[I];
787 if (P.PtrKind == T.PtrKind &&
788 P.VectorSize == T.VectorSize &&
789 P.ArgType == T.ArgType) {
790 return I;
793 return -1;
796 template <typename Stream>
797 bool trySubst(Stream& os, const AMDGPULibFunc::Param& p) {
798 int const subst = findSubst(p);
799 if (subst < 0) return false;
800 // Substitutions are mangled as S(XX)?_ where XX is a hexadecimal number
801 // 0 1 2
802 // S_ S0_ S1_
803 if (subst == 0) os << "S_";
804 else os << 'S' << (subst-1) << '_';
805 return true;
808 public:
809 ItaniumMangler(bool useAddrSpace)
810 : UseAddrSpace(useAddrSpace) {}
812 template <typename Stream>
813 void operator()(Stream& os, AMDGPULibFunc::Param p) {
815 // Itanium mangling ABI 5.1.8. Compression:
816 // Logically, the substitutable components of a mangled name are considered
817 // left-to-right, components before the composite structure of which they
818 // are a part. If a component has been encountered before, it is substituted
819 // as described below. This decision is independent of whether its components
820 // have been substituted, so an implementation may optimize by considering
821 // large structures for substitution before their components. If a component
822 // has not been encountered before, its mangling is identified, and it is
823 // added to a dictionary of substitution candidates. No entity is added to
824 // the dictionary twice.
825 AMDGPULibFunc::Param Ptr;
827 if (p.PtrKind) {
828 if (trySubst(os, p)) return;
829 os << 'P';
830 if (p.PtrKind & AMDGPULibFunc::CONST) os << 'K';
831 if (p.PtrKind & AMDGPULibFunc::VOLATILE) os << 'V';
832 unsigned AS = UseAddrSpace
833 ? AMDGPULibFuncBase::getAddrSpaceFromEPtrKind(p.PtrKind)
834 : 0;
835 if (EnableOCLManglingMismatchWA || AS != 0)
836 os << "U3AS" << AS;
837 Ptr = p;
838 p.PtrKind = 0;
841 if (p.VectorSize > 1) {
842 if (trySubst(os, p)) goto exit;
843 Str.push_back(p);
844 os << "Dv" << static_cast<unsigned>(p.VectorSize) << '_';
847 os << getItaniumTypeName((AMDGPULibFunc::EType)p.ArgType);
849 exit:
850 if (Ptr.ArgType) Str.push_back(Ptr);
853 } // namespace
855 std::string AMDGPUMangledLibFunc::mangleNameItanium() const {
856 SmallString<128> Buf;
857 raw_svector_ostream S(Buf);
858 SmallString<128> NameBuf;
859 raw_svector_ostream Name(NameBuf);
860 writeName(Name);
861 const StringRef& NameStr = Name.str();
862 S << "_Z" << static_cast<int>(NameStr.size()) << NameStr;
864 ItaniumMangler Mangler(true);
865 ParamIterator I(Leads, manglingRules[FuncId]);
866 Param P;
867 while ((P = I.getNextParam()).ArgType != 0)
868 Mangler(S, P);
869 return std::string(S.str());
872 ///////////////////////////////////////////////////////////////////////////////
873 // Misc
875 static Type* getIntrinsicParamType(
876 LLVMContext& C,
877 const AMDGPULibFunc::Param& P,
878 bool useAddrSpace) {
879 Type* T = nullptr;
880 switch (P.ArgType) {
881 case AMDGPULibFunc::U8:
882 case AMDGPULibFunc::I8: T = Type::getInt8Ty(C); break;
883 case AMDGPULibFunc::U16:
884 case AMDGPULibFunc::I16: T = Type::getInt16Ty(C); break;
885 case AMDGPULibFunc::U32:
886 case AMDGPULibFunc::I32: T = Type::getInt32Ty(C); break;
887 case AMDGPULibFunc::U64:
888 case AMDGPULibFunc::I64: T = Type::getInt64Ty(C); break;
889 case AMDGPULibFunc::F16: T = Type::getHalfTy(C); break;
890 case AMDGPULibFunc::F32: T = Type::getFloatTy(C); break;
891 case AMDGPULibFunc::F64: T = Type::getDoubleTy(C); break;
893 case AMDGPULibFunc::IMG1DA:
894 case AMDGPULibFunc::IMG1DB:
895 case AMDGPULibFunc::IMG2DA:
896 case AMDGPULibFunc::IMG1D:
897 case AMDGPULibFunc::IMG2D:
898 case AMDGPULibFunc::IMG3D:
899 T = StructType::create(C,"ocl_image")->getPointerTo(); break;
900 case AMDGPULibFunc::SAMPLER:
901 T = StructType::create(C,"ocl_sampler")->getPointerTo(); break;
902 case AMDGPULibFunc::EVENT:
903 T = StructType::create(C,"ocl_event")->getPointerTo(); break;
904 default:
905 llvm_unreachable("Unhandeled param type");
906 return nullptr;
908 if (P.VectorSize > 1)
909 T = FixedVectorType::get(T, P.VectorSize);
910 if (P.PtrKind != AMDGPULibFunc::BYVALUE)
911 T = useAddrSpace ? T->getPointerTo((P.PtrKind & AMDGPULibFunc::ADDR_SPACE)
912 - 1)
913 : T->getPointerTo();
914 return T;
917 FunctionType *AMDGPUMangledLibFunc::getFunctionType(Module &M) const {
918 LLVMContext& C = M.getContext();
919 std::vector<Type*> Args;
920 ParamIterator I(Leads, manglingRules[FuncId]);
921 Param P;
922 while ((P=I.getNextParam()).ArgType != 0)
923 Args.push_back(getIntrinsicParamType(C, P, true));
925 return FunctionType::get(
926 getIntrinsicParamType(C, getRetType(FuncId, Leads), true),
927 Args, false);
930 unsigned AMDGPUMangledLibFunc::getNumArgs() const {
931 return manglingRules[FuncId].getNumArgs();
934 unsigned AMDGPUUnmangledLibFunc::getNumArgs() const {
935 return UnmangledFuncInfo::getNumArgs(FuncId);
938 std::string AMDGPUMangledLibFunc::getName() const {
939 SmallString<128> Buf;
940 raw_svector_ostream OS(Buf);
941 writeName(OS);
942 return std::string(OS.str());
945 Function *AMDGPULibFunc::getFunction(Module *M, const AMDGPULibFunc &fInfo) {
946 std::string FuncName = fInfo.mangle();
947 Function *F = dyn_cast_or_null<Function>(
948 M->getValueSymbolTable().lookup(FuncName));
950 // check formal with actual types conformance
951 if (F && !F->isDeclaration()
952 && !F->isVarArg()
953 && F->arg_size() == fInfo.getNumArgs()) {
954 return F;
956 return nullptr;
959 FunctionCallee AMDGPULibFunc::getOrInsertFunction(Module *M,
960 const AMDGPULibFunc &fInfo) {
961 std::string const FuncName = fInfo.mangle();
962 Function *F = dyn_cast_or_null<Function>(
963 M->getValueSymbolTable().lookup(FuncName));
965 // check formal with actual types conformance
966 if (F && !F->isDeclaration()
967 && !F->isVarArg()
968 && F->arg_size() == fInfo.getNumArgs()) {
969 return F;
972 FunctionType *FuncTy = fInfo.getFunctionType(*M);
974 bool hasPtr = false;
975 for (FunctionType::param_iterator
976 PI = FuncTy->param_begin(),
977 PE = FuncTy->param_end();
978 PI != PE; ++PI) {
979 const Type* argTy = static_cast<const Type*>(*PI);
980 if (argTy->isPointerTy()) {
981 hasPtr = true;
982 break;
986 FunctionCallee C;
987 if (hasPtr) {
988 // Do not set extra attributes for functions with pointer arguments.
989 C = M->getOrInsertFunction(FuncName, FuncTy);
990 } else {
991 AttributeList Attr;
992 LLVMContext &Ctx = M->getContext();
993 Attr = Attr.addFnAttribute(Ctx, Attribute::ReadOnly);
994 Attr = Attr.addFnAttribute(Ctx, Attribute::NoUnwind);
995 C = M->getOrInsertFunction(FuncName, FuncTy, Attr);
998 return C;
1001 StringMap<unsigned> UnmangledFuncInfo::buildNameMap() {
1002 StringMap<unsigned> Map;
1003 for (unsigned I = 0; I != TableSize; ++I)
1004 Map[Table[I].Name] = I;
1005 return Map;
1008 bool UnmangledFuncInfo::lookup(StringRef Name, ID &Id) {
1009 static const StringMap<unsigned> Map = buildNameMap();
1010 auto Loc = Map.find(Name);
1011 if (Loc != Map.end()) {
1012 Id = toFuncId(Loc->second);
1013 return true;
1015 Id = AMDGPULibFunc::EI_NONE;
1016 return false;
1019 AMDGPULibFunc::AMDGPULibFunc(const AMDGPULibFunc &F) {
1020 if (auto *MF = dyn_cast<AMDGPUMangledLibFunc>(F.Impl.get()))
1021 Impl.reset(new AMDGPUMangledLibFunc(*MF));
1022 else if (auto *UMF = dyn_cast<AMDGPUUnmangledLibFunc>(F.Impl.get()))
1023 Impl.reset(new AMDGPUUnmangledLibFunc(*UMF));
1024 else
1025 Impl = std::unique_ptr<AMDGPULibFuncImpl>();
1028 AMDGPULibFunc &AMDGPULibFunc::operator=(const AMDGPULibFunc &F) {
1029 if (this == &F)
1030 return *this;
1031 new (this) AMDGPULibFunc(F);
1032 return *this;
1035 AMDGPULibFunc::AMDGPULibFunc(EFuncId Id, const AMDGPULibFunc &CopyFrom) {
1036 assert(AMDGPULibFuncBase::isMangled(Id) && CopyFrom.isMangled() &&
1037 "not supported");
1038 Impl.reset(new AMDGPUMangledLibFunc(
1039 Id, *cast<AMDGPUMangledLibFunc>(CopyFrom.Impl.get())));
1042 AMDGPULibFunc::AMDGPULibFunc(StringRef Name, FunctionType *FT) {
1043 Impl.reset(new AMDGPUUnmangledLibFunc(Name, FT));
1046 void AMDGPULibFunc::initMangled() { Impl.reset(new AMDGPUMangledLibFunc()); }
1048 AMDGPULibFunc::Param *AMDGPULibFunc::getLeads() {
1049 if (!Impl)
1050 initMangled();
1051 return cast<AMDGPUMangledLibFunc>(Impl.get())->Leads;
1054 const AMDGPULibFunc::Param *AMDGPULibFunc::getLeads() const {
1055 return cast<const AMDGPUMangledLibFunc>(Impl.get())->Leads;