?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Revert r354244 "[DAGCombiner] Eliminate dead stores to stack."
[llvm-complete.git] / lib / Target / AMDGPU / AMDGPULibFunc.cpp
blob09322594dc8e0a896adcdc0e910a0e10c734ced2
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 "AMDGPU.h"
14 #include "AMDGPULibFunc.h"
15 #include <llvm/ADT/SmallString.h>
16 #include <llvm/ADT/SmallVector.h>
17 #include <llvm/ADT/StringSwitch.h>
18 #include "llvm/IR/Attributes.h"
19 #include "llvm/IR/DerivedTypes.h"
20 #include "llvm/IR/Function.h"
21 #include "llvm/IR/Module.h"
22 #include "llvm/IR/ValueSymbolTable.h"
23 #include <llvm/Support/raw_ostream.h>
24 #include <string>
25
26 using namespace llvm;
27
28 namespace {
29
30 enum EManglingParam {
31 E_NONE,
32 EX_EVENT,
33 EX_FLOAT4,
34 EX_INTV4,
35 EX_RESERVEDID,
36 EX_SAMPLER,
37 EX_SIZET,
38 EX_UINT,
39 EX_UINTV4,
40 E_ANY,
41 E_CONSTPTR_ANY,
42 E_CONSTPTR_SWAPGL,
43 E_COPY,
44 E_IMAGECOORDS,
45 E_POINTEE,
46 E_SETBASE_I32,
47 E_SETBASE_U32,
48 E_MAKEBASE_UNS,
49 E_V16_OF_POINTEE,
50 E_V2_OF_POINTEE,
51 E_V3_OF_POINTEE,
52 E_V4_OF_POINTEE,
53 E_V8_OF_POINTEE,
54 E_VLTLPTR_ANY,
55 };
56
57 struct ManglingRule {
58 StringRef const Name;
59 unsigned char Lead[2];
60 unsigned char Param[5];
61
62 int maxLeadIndex() const { return (std::max)(Lead[0], Lead[1]); }
63 int getNumLeads() const { return (Lead[0] ? 1 : 0) + (Lead[1] ? 1 : 0); }
64
65 unsigned getNumArgs() const;
66 };
67
68 // Information about library functions with unmangled names.
69 class UnmangledFuncInfo {
70 StringRef const Name;
71 unsigned NumArgs;
72
73 // Table for all lib functions with unmangled names.
74 static const UnmangledFuncInfo Table[];
75
76 // Number of entries in Table.
77 static const unsigned TableSize;
78
79 // Map function name to index.
80 class NameMap : public StringMap<unsigned> {
81 public:
82 NameMap() {
83 for (unsigned I = 0; I != TableSize; ++I)
84 (*this)[Table[I].Name] = I;
85 }
86 };
87 friend class NameMap;
88 static NameMap Map;
89
90 public:
91 using ID = AMDGPULibFunc::EFuncId;
92 UnmangledFuncInfo(StringRef _Name, unsigned _NumArgs)
93 : Name(_Name), NumArgs(_NumArgs) {}
94 // Get index to Table by function name.
95 static bool lookup(StringRef Name, ID &Id);
96 static unsigned toIndex(ID Id) {
97 assert(static_cast<unsigned>(Id) >
98 static_cast<unsigned>(AMDGPULibFunc::EI_LAST_MANGLED) &&
99 "Invalid unmangled library function");
100 return static_cast<unsigned>(Id) - 1 -
101 static_cast<unsigned>(AMDGPULibFunc::EI_LAST_MANGLED);
102 }
103 static ID toFuncId(unsigned Index) {
104 assert(Index < TableSize && "Invalid unmangled library function");
105 return static_cast<ID>(
106 Index + 1 + static_cast<unsigned>(AMDGPULibFunc::EI_LAST_MANGLED));
107 }
108 static unsigned getNumArgs(ID Id) { return Table[toIndex(Id)].NumArgs; }
109 static StringRef getName(ID Id) { return Table[toIndex(Id)].Name; }
110 };
111
112 unsigned ManglingRule::getNumArgs() const {
113 unsigned I=0;
114 while (I < (sizeof Param/sizeof Param[0]) && Param[I]) ++I;
115 return I;
116 }
117
118 // This table describes function formal argument type rules. The order of rules
119 // corresponds to the EFuncId enum at AMDGPULibFunc.h
120 //
121 // "<func name>", { <leads> }, { <param rules> }
122 // where:
123 // <leads> - list of integers that are one-based indexes of formal argument
124 // used to mangle a function name. Other argument types are derived from types
125 // of these 'leads'. The order of integers in this list correspond to the
126 // order in which these arguments are mangled in the EDG mangling scheme. The
127 // same order should be preserved for arguments in the AMDGPULibFunc structure
128 // when it is used for mangling. For example:
129 // { "vstorea_half", {3,1}, {E_ANY,EX_SIZET,E_ANY}},
130 // will be mangled in EDG scheme as vstorea_half_<3dparam>_<1stparam>
131 // When mangling from code use:
132 // AMDGPULibFunc insc;
133 // insc.param[0] = ... // describe 3rd parameter
134 // insc.param[1] = ... // describe 1rd parameter
135 //
136 // <param rules> - list of rules used to derive all of the function formal
137 // argument types. EX_ prefixed are simple types, other derived from the
138 // latest 'lead' argument type in the order of encoding from first to last.
139 // E_ANY - use prev lead type, E_CONSTPTR_ANY - make const pointer out of
140 // prev lead type, etc. see ParamIterator::getNextParam() for details.
141
142 static const ManglingRule manglingRules[] = {
143 { StringRef(), {0}, {0} },
144 { "abs" , {1}, {E_ANY}},
145 { "abs_diff" , {1}, {E_ANY,E_COPY}},
146 { "acos" , {1}, {E_ANY}},
147 { "acosh" , {1}, {E_ANY}},
148 { "acospi" , {1}, {E_ANY}},
149 { "add_sat" , {1}, {E_ANY,E_COPY}},
150 { "all" , {1}, {E_ANY}},
151 { "any" , {1}, {E_ANY}},
152 { "asin" , {1}, {E_ANY}},
153 { "asinh" , {1}, {E_ANY}},
154 { "asinpi" , {1}, {E_ANY}},
155 { "async_work_group_copy" , {1}, {E_ANY,E_CONSTPTR_SWAPGL,EX_SIZET,EX_EVENT}},
156 { "async_work_group_strided_copy" , {1}, {E_ANY,E_CONSTPTR_SWAPGL,EX_SIZET,EX_SIZET,EX_EVENT}},
157 { "atan" , {1}, {E_ANY}},
158 { "atan2" , {1}, {E_ANY,E_COPY}},
159 { "atan2pi" , {1}, {E_ANY,E_COPY}},
160 { "atanh" , {1}, {E_ANY}},
161 { "atanpi" , {1}, {E_ANY}},
162 { "atomic_add" , {1}, {E_VLTLPTR_ANY,E_POINTEE}},
163 { "atomic_and" , {1}, {E_VLTLPTR_ANY,E_POINTEE}},
164 { "atomic_cmpxchg" , {1}, {E_VLTLPTR_ANY,E_POINTEE,E_POINTEE}},
165 { "atomic_dec" , {1}, {E_VLTLPTR_ANY}},
166 { "atomic_inc" , {1}, {E_VLTLPTR_ANY}},
167 { "atomic_max" , {1}, {E_VLTLPTR_ANY,E_POINTEE}},
168 { "atomic_min" , {1}, {E_VLTLPTR_ANY,E_POINTEE}},
169 { "atomic_or" , {1}, {E_VLTLPTR_ANY,E_POINTEE}},
170 { "atomic_sub" , {1}, {E_VLTLPTR_ANY,E_POINTEE}},
171 { "atomic_xchg" , {1}, {E_VLTLPTR_ANY,E_POINTEE}},
172 { "atomic_xor" , {1}, {E_VLTLPTR_ANY,E_POINTEE}},
173 { "bitselect" , {1}, {E_ANY,E_COPY,E_COPY}},
174 { "cbrt" , {1}, {E_ANY}},
175 { "ceil" , {1}, {E_ANY}},
176 { "clamp" , {1}, {E_ANY,E_COPY,E_COPY}},
177 { "clz" , {1}, {E_ANY}},
178 { "commit_read_pipe" , {1}, {E_ANY,EX_RESERVEDID}},
179 { "commit_write_pipe" , {1}, {E_ANY,EX_RESERVEDID}},
180 { "copysign" , {1}, {E_ANY,E_COPY}},
181 { "cos" , {1}, {E_ANY}},
182 { "cosh" , {1}, {E_ANY}},
183 { "cospi" , {1}, {E_ANY}},
184 { "cross" , {1}, {E_ANY,E_COPY}},
185 { "ctz" , {1}, {E_ANY}},
186 { "degrees" , {1}, {E_ANY}},
187 { "distance" , {1}, {E_ANY,E_COPY}},
188 { "divide" , {1}, {E_ANY,E_COPY}},
189 { "dot" , {1}, {E_ANY,E_COPY}},
190 { "erf" , {1}, {E_ANY}},
191 { "erfc" , {1}, {E_ANY}},
192 { "exp" , {1}, {E_ANY}},
193 { "exp10" , {1}, {E_ANY}},
194 { "exp2" , {1}, {E_ANY}},
195 { "expm1" , {1}, {E_ANY}},
196 { "fabs" , {1}, {E_ANY}},
197 { "fast_distance" , {1}, {E_ANY,E_COPY}},
198 { "fast_length" , {1}, {E_ANY}},
199 { "fast_normalize" , {1}, {E_ANY}},
200 { "fdim" , {1}, {E_ANY,E_COPY}},
201 { "floor" , {1}, {E_ANY}},
202 { "fma" , {1}, {E_ANY,E_COPY,E_COPY}},
203 { "fmax" , {1}, {E_ANY,E_COPY}},
204 { "fmin" , {1}, {E_ANY,E_COPY}},
205 { "fmod" , {1}, {E_ANY,E_COPY}},
206 { "fract" , {2}, {E_POINTEE,E_ANY}},
207 { "frexp" , {1,2}, {E_ANY,E_ANY}},
208 { "get_image_array_size" , {1}, {E_ANY}},
209 { "get_image_channel_data_type" , {1}, {E_ANY}},
210 { "get_image_channel_order" , {1}, {E_ANY}},
211 { "get_image_dim" , {1}, {E_ANY}},
212 { "get_image_height" , {1}, {E_ANY}},
213 { "get_image_width" , {1}, {E_ANY}},
214 { "get_pipe_max_packets" , {1}, {E_ANY}},
215 { "get_pipe_num_packets" , {1}, {E_ANY}},
216 { "hadd" , {1}, {E_ANY,E_COPY}},
217 { "hypot" , {1}, {E_ANY,E_COPY}},
218 { "ilogb" , {1}, {E_ANY}},
219 { "isequal" , {1}, {E_ANY,E_COPY}},
220 { "isfinite" , {1}, {E_ANY}},
221 { "isgreater" , {1}, {E_ANY,E_COPY}},
222 { "isgreaterequal" , {1}, {E_ANY,E_COPY}},
223 { "isinf" , {1}, {E_ANY}},
224 { "isless" , {1}, {E_ANY,E_COPY}},
225 { "islessequal" , {1}, {E_ANY,E_COPY}},
226 { "islessgreater" , {1}, {E_ANY,E_COPY}},
227 { "isnan" , {1}, {E_ANY}},
228 { "isnormal" , {1}, {E_ANY}},
229 { "isnotequal" , {1}, {E_ANY,E_COPY}},
230 { "isordered" , {1}, {E_ANY,E_COPY}},
231 { "isunordered" , {1}, {E_ANY,E_COPY}},
232 { "ldexp" , {1}, {E_ANY,E_SETBASE_I32}},
233 { "length" , {1}, {E_ANY}},
234 { "lgamma" , {1}, {E_ANY}},
235 { "lgamma_r" , {1,2}, {E_ANY,E_ANY}},
236 { "log" , {1}, {E_ANY}},
237 { "log10" , {1}, {E_ANY}},
238 { "log1p" , {1}, {E_ANY}},
239 { "log2" , {1}, {E_ANY}},
240 { "logb" , {1}, {E_ANY}},
241 { "mad" , {1}, {E_ANY,E_COPY,E_COPY}},
242 { "mad24" , {1}, {E_ANY,E_COPY,E_COPY}},
243 { "mad_hi" , {1}, {E_ANY,E_COPY,E_COPY}},
244 { "mad_sat" , {1}, {E_ANY,E_COPY,E_COPY}},
245 { "max" , {1}, {E_ANY,E_COPY}},
246 { "maxmag" , {1}, {E_ANY,E_COPY}},
247 { "min" , {1}, {E_ANY,E_COPY}},
248 { "minmag" , {1}, {E_ANY,E_COPY}},
249 { "mix" , {1}, {E_ANY,E_COPY,E_COPY}},
250 { "modf" , {2}, {E_POINTEE,E_ANY}},
251 { "mul24" , {1}, {E_ANY,E_COPY}},
252 { "mul_hi" , {1}, {E_ANY,E_COPY}},
253 { "nan" , {1}, {E_ANY}},
254 { "nextafter" , {1}, {E_ANY,E_COPY}},
255 { "normalize" , {1}, {E_ANY}},
256 { "popcount" , {1}, {E_ANY}},
257 { "pow" , {1}, {E_ANY,E_COPY}},
258 { "pown" , {1}, {E_ANY,E_SETBASE_I32}},
259 { "powr" , {1}, {E_ANY,E_COPY}},
260 { "prefetch" , {1}, {E_CONSTPTR_ANY,EX_SIZET}},
261 { "radians" , {1}, {E_ANY}},
262 { "recip" , {1}, {E_ANY}},
263 { "remainder" , {1}, {E_ANY,E_COPY}},
264 { "remquo" , {1,3}, {E_ANY,E_COPY,E_ANY}},
265 { "reserve_read_pipe" , {1}, {E_ANY,EX_UINT}},
266 { "reserve_write_pipe" , {1}, {E_ANY,EX_UINT}},
267 { "rhadd" , {1}, {E_ANY,E_COPY}},
268 { "rint" , {1}, {E_ANY}},
269 { "rootn" , {1}, {E_ANY,E_SETBASE_I32}},
270 { "rotate" , {1}, {E_ANY,E_COPY}},
271 { "round" , {1}, {E_ANY}},
272 { "rsqrt" , {1}, {E_ANY}},
273 { "select" , {1,3}, {E_ANY,E_COPY,E_ANY}},
274 { "shuffle" , {1,2}, {E_ANY,E_ANY}},
275 { "shuffle2" , {1,3}, {E_ANY,E_COPY,E_ANY}},
276 { "sign" , {1}, {E_ANY}},
277 { "signbit" , {1}, {E_ANY}},
278 { "sin" , {1}, {E_ANY}},
279 { "sincos" , {2}, {E_POINTEE,E_ANY}},
280 { "sinh" , {1}, {E_ANY}},
281 { "sinpi" , {1}, {E_ANY}},
282 { "smoothstep" , {1}, {E_ANY,E_COPY,E_COPY}},
283 { "sqrt" , {1}, {E_ANY}},
284 { "step" , {1}, {E_ANY,E_COPY}},
285 { "sub_group_broadcast" , {1}, {E_ANY,EX_UINT}},
286 { "sub_group_commit_read_pipe" , {1}, {E_ANY,EX_RESERVEDID}},
287 { "sub_group_commit_write_pipe" , {1}, {E_ANY,EX_RESERVEDID}},
288 { "sub_group_reduce_add" , {1}, {E_ANY}},
289 { "sub_group_reduce_max" , {1}, {E_ANY}},
290 { "sub_group_reduce_min" , {1}, {E_ANY}},
291 { "sub_group_reserve_read_pipe" , {1}, {E_ANY,EX_UINT}},
292 { "sub_group_reserve_write_pipe" , {1}, {E_ANY,EX_UINT}},
293 { "sub_group_scan_exclusive_add" , {1}, {E_ANY}},
294 { "sub_group_scan_exclusive_max" , {1}, {E_ANY}},
295 { "sub_group_scan_exclusive_min" , {1}, {E_ANY}},
296 { "sub_group_scan_inclusive_add" , {1}, {E_ANY}},
297 { "sub_group_scan_inclusive_max" , {1}, {E_ANY}},
298 { "sub_group_scan_inclusive_min" , {1}, {E_ANY}},
299 { "sub_sat" , {1}, {E_ANY,E_COPY}},
300 { "tan" , {1}, {E_ANY}},
301 { "tanh" , {1}, {E_ANY}},
302 { "tanpi" , {1}, {E_ANY}},
303 { "tgamma" , {1}, {E_ANY}},
304 { "trunc" , {1}, {E_ANY}},
305 { "upsample" , {1}, {E_ANY,E_MAKEBASE_UNS}},
306 { "vec_step" , {1}, {E_ANY}},
307 { "vstore" , {3}, {E_POINTEE,EX_SIZET,E_ANY}},
308 { "vstore16" , {3}, {E_V16_OF_POINTEE,EX_SIZET,E_ANY}},
309 { "vstore2" , {3}, {E_V2_OF_POINTEE,EX_SIZET,E_ANY}},
310 { "vstore3" , {3}, {E_V3_OF_POINTEE,EX_SIZET,E_ANY}},
311 { "vstore4" , {3}, {E_V4_OF_POINTEE,EX_SIZET,E_ANY}},
312 { "vstore8" , {3}, {E_V8_OF_POINTEE,EX_SIZET,E_ANY}},
313 { "work_group_commit_read_pipe" , {1}, {E_ANY,EX_RESERVEDID}},
314 { "work_group_commit_write_pipe" , {1}, {E_ANY,EX_RESERVEDID}},
315 { "work_group_reduce_add" , {1}, {E_ANY}},
316 { "work_group_reduce_max" , {1}, {E_ANY}},
317 { "work_group_reduce_min" , {1}, {E_ANY}},
318 { "work_group_reserve_read_pipe" , {1}, {E_ANY,EX_UINT}},
319 { "work_group_reserve_write_pipe" , {1}, {E_ANY,EX_UINT}},
320 { "work_group_scan_exclusive_add" , {1}, {E_ANY}},
321 { "work_group_scan_exclusive_max" , {1}, {E_ANY}},
322 { "work_group_scan_exclusive_min" , {1}, {E_ANY}},
323 { "work_group_scan_inclusive_add" , {1}, {E_ANY}},
324 { "work_group_scan_inclusive_max" , {1}, {E_ANY}},
325 { "work_group_scan_inclusive_min" , {1}, {E_ANY}},
326 { "write_imagef" , {1}, {E_ANY,E_IMAGECOORDS,EX_FLOAT4}},
327 { "write_imagei" , {1}, {E_ANY,E_IMAGECOORDS,EX_INTV4}},
328 { "write_imageui" , {1}, {E_ANY,E_IMAGECOORDS,EX_UINTV4}},
329 { "ncos" , {1}, {E_ANY} },
330 { "nexp2" , {1}, {E_ANY} },
331 { "nfma" , {1}, {E_ANY, E_COPY, E_COPY} },
332 { "nlog2" , {1}, {E_ANY} },
333 { "nrcp" , {1}, {E_ANY} },
334 { "nrsqrt" , {1}, {E_ANY} },
335 { "nsin" , {1}, {E_ANY} },
336 { "nsqrt" , {1}, {E_ANY} },
337 { "ftz" , {1}, {E_ANY} },
338 { "fldexp" , {1}, {E_ANY, EX_UINT} },
339 { "class" , {1}, {E_ANY, EX_UINT} },
340 { "rcbrt" , {1}, {E_ANY} },
341 };
342
343 // Library functions with unmangled name.
344 const UnmangledFuncInfo UnmangledFuncInfo::Table[] = {
345 {"__read_pipe_2", 4},
346 {"__read_pipe_4", 6},
347 {"__write_pipe_2", 4},
348 {"__write_pipe_4", 6},
349 };
350
351 const unsigned UnmangledFuncInfo::TableSize =
352 sizeof(UnmangledFuncInfo::Table) / sizeof(UnmangledFuncInfo::Table[0]);
353
354 UnmangledFuncInfo::NameMap UnmangledFuncInfo::Map;
355
356 static const struct ManglingRulesMap : public StringMap<int> {
357 ManglingRulesMap()
358 : StringMap<int>(sizeof(manglingRules)/sizeof(manglingRules[0])) {
359 int Id = 0;
360 for (auto Rule : manglingRules)
361 insert({ Rule.Name, Id++ });
362 }
363 } manglingRulesMap;
364
365 static AMDGPULibFunc::Param getRetType(AMDGPULibFunc::EFuncId id,
366 const AMDGPULibFunc::Param (&Leads)[2]) {
367 AMDGPULibFunc::Param Res = Leads[0];
368 // TBD - This switch may require to be extended for other intriniscs
369 switch (id) {
370 case AMDGPULibFunc::EI_SINCOS:
371 Res.PtrKind = AMDGPULibFunc::BYVALUE;
372 break;
373 default:
374 break;
375 }
376 return Res;
377 }
378
379 class ParamIterator {
380 const AMDGPULibFunc::Param (&Leads)[2];
381 const ManglingRule& Rule;
382 int Index;
383 public:
384 ParamIterator(const AMDGPULibFunc::Param (&leads)[2],
385 const ManglingRule& rule)
386 : Leads(leads), Rule(rule), Index(0) {}
387
388 AMDGPULibFunc::Param getNextParam();
389 };
390
391 AMDGPULibFunc::Param ParamIterator::getNextParam() {
392 AMDGPULibFunc::Param P;
393 if (Index >= int(sizeof Rule.Param/sizeof Rule.Param[0])) return P;
394
395 const char R = Rule.Param[Index];
396 switch (R) {
397 case E_NONE: break;
398 case EX_UINT:
399 P.ArgType = AMDGPULibFunc::U32; break;
400 case EX_INTV4:
401 P.ArgType = AMDGPULibFunc::I32; P.VectorSize = 4; break;
402 case EX_UINTV4:
403 P.ArgType = AMDGPULibFunc::U32; P.VectorSize = 4; break;
404 case EX_FLOAT4:
405 P.ArgType = AMDGPULibFunc::F32; P.VectorSize = 4; break;
406 case EX_SIZET:
407 P.ArgType = AMDGPULibFunc::U64; break;
408 case EX_EVENT:
409 P.ArgType = AMDGPULibFunc::EVENT; break;
410 case EX_SAMPLER:
411 P.ArgType = AMDGPULibFunc::SAMPLER; break;
412 case EX_RESERVEDID: break; // TBD
413 default:
414 if (Index == (Rule.Lead[1] - 1)) P = Leads[1];
415 else P = Leads[0];
416
417 switch (R) {
418 case E_ANY:
419 case E_COPY: break;
420
421 case E_POINTEE:
422 P.PtrKind = AMDGPULibFunc::BYVALUE; break;
423 case E_V2_OF_POINTEE:
424 P.VectorSize = 2; P.PtrKind = AMDGPULibFunc::BYVALUE; break;
425 case E_V3_OF_POINTEE:
426 P.VectorSize = 3; P.PtrKind = AMDGPULibFunc::BYVALUE; break;
427 case E_V4_OF_POINTEE:
428 P.VectorSize = 4; P.PtrKind = AMDGPULibFunc::BYVALUE; break;
429 case E_V8_OF_POINTEE:
430 P.VectorSize = 8; P.PtrKind = AMDGPULibFunc::BYVALUE; break;
431 case E_V16_OF_POINTEE:
432 P.VectorSize = 16; P.PtrKind = AMDGPULibFunc::BYVALUE; break;
433 case E_CONSTPTR_ANY:
434 P.PtrKind |= AMDGPULibFunc::CONST; break;
435 case E_VLTLPTR_ANY:
436 P.PtrKind |= AMDGPULibFunc::VOLATILE; break;
437 case E_SETBASE_I32:
438 P.ArgType = AMDGPULibFunc::I32; break;
439 case E_SETBASE_U32:
440 P.ArgType = AMDGPULibFunc::U32; break;
441
442 case E_MAKEBASE_UNS:
443 P.ArgType &= ~AMDGPULibFunc::BASE_TYPE_MASK;
444 P.ArgType |= AMDGPULibFunc::UINT;
445 break;
446
447 case E_IMAGECOORDS:
448 switch (P.ArgType) {
449 case AMDGPULibFunc::IMG1DA: P.VectorSize = 2; break;
450 case AMDGPULibFunc::IMG1DB: P.VectorSize = 1; break;
451 case AMDGPULibFunc::IMG2DA: P.VectorSize = 4; break;
452 case AMDGPULibFunc::IMG1D: P.VectorSize = 1; break;
453 case AMDGPULibFunc::IMG2D: P.VectorSize = 2; break;
454 case AMDGPULibFunc::IMG3D: P.VectorSize = 4; break;
455 }
456 P.PtrKind = AMDGPULibFunc::BYVALUE;
457 P.ArgType = AMDGPULibFunc::I32;
458 break;
459
460 case E_CONSTPTR_SWAPGL: {
461 unsigned AS = AMDGPULibFunc::getAddrSpaceFromEPtrKind(P.PtrKind);
462 switch (AS) {
463 case AMDGPUAS::GLOBAL_ADDRESS: AS = AMDGPUAS::LOCAL_ADDRESS; break;
464 case AMDGPUAS::LOCAL_ADDRESS: AS = AMDGPUAS::GLOBAL_ADDRESS; break;
465 }
466 P.PtrKind = AMDGPULibFunc::getEPtrKindFromAddrSpace(AS);
467 P.PtrKind |= AMDGPULibFunc::CONST;
468 break;
469 }
470
471 default: llvm_unreachable("Unhandeled param rule");
472 }
473 }
474 ++Index;
475 return P;
476 }
477
478 inline static void drop_front(StringRef& str, size_t n = 1) {
479 str = str.drop_front(n);
480 }
481
482 static bool eatTerm(StringRef& mangledName, const char c) {
483 if (mangledName.front() == c) {
484 drop_front(mangledName);
485 return true;
486 }
487 return false;
488 }
489
490 template <size_t N>
491 static bool eatTerm(StringRef& mangledName, const char (&str)[N]) {
492 if (mangledName.startswith(StringRef(str, N-1))) {
493 drop_front(mangledName, N-1);
494 return true;
495 }
496 return false;
497 }
498
499 static inline bool isDigit(char c) { return c >= '0' && c <= '9'; }
500
501 static int eatNumber(StringRef& s) {
502 size_t const savedSize = s.size();
503 int n = 0;
504 while (!s.empty() && isDigit(s.front())) {
505 n = n*10 + s.front() - '0';
506 drop_front(s);
507 }
508 return s.size() < savedSize ? n : -1;
509 }
510
511 static StringRef eatLengthPrefixedName(StringRef& mangledName) {
512 int const Len = eatNumber(mangledName);
513 if (Len <= 0 || static_cast<size_t>(Len) > mangledName.size())
514 return StringRef();
515 StringRef Res = mangledName.substr(0, Len);
516 drop_front(mangledName, Len);
517 return Res;
518 }
519
520 } // end anonymous namespace
521
522 AMDGPUMangledLibFunc::AMDGPUMangledLibFunc() {
523 FuncId = EI_NONE;
524 FKind = NOPFX;
525 Leads[0].reset();
526 Leads[1].reset();
527 Name.clear();
528 }
529
530 AMDGPUUnmangledLibFunc::AMDGPUUnmangledLibFunc() {
531 FuncId = EI_NONE;
532 FuncTy = nullptr;
533 }
534
535 AMDGPUMangledLibFunc::AMDGPUMangledLibFunc(
536 EFuncId id, const AMDGPUMangledLibFunc &copyFrom) {
537 FuncId = id;
538 FKind = copyFrom.FKind;
539 Leads[0] = copyFrom.Leads[0];
540 Leads[1] = copyFrom.Leads[1];
541 }
542
543 ///////////////////////////////////////////////////////////////////////////////
544 // Demangling
545
546 static int parseVecSize(StringRef& mangledName) {
547 size_t const Len = eatNumber(mangledName);
548 switch (Len) {
549 case 2: case 3: case 4: case 8: case 16:
550 return Len;
551 default:
552 break;
553 }
554 return 1;
555 }
556
557 static AMDGPULibFunc::ENamePrefix parseNamePrefix(StringRef& mangledName) {
558 std::pair<StringRef, StringRef> const P = mangledName.split('_');
559 AMDGPULibFunc::ENamePrefix Pfx =
560 StringSwitch<AMDGPULibFunc::ENamePrefix>(P.first)
561 .Case("native", AMDGPULibFunc::NATIVE)
562 .Case("half" , AMDGPULibFunc::HALF)
563 .Default(AMDGPULibFunc::NOPFX);
564
565 if (Pfx != AMDGPULibFunc::NOPFX)
566 mangledName = P.second;
567
568 return Pfx;
569 }
570
571 bool AMDGPUMangledLibFunc::parseUnmangledName(StringRef FullName) {
572 FuncId = static_cast<EFuncId>(manglingRulesMap.lookup(FullName));
573 return FuncId != EI_NONE;
574 }
575
576 ///////////////////////////////////////////////////////////////////////////////
577 // Itanium Demangling
578
579 namespace {
580 struct ItaniumParamParser {
581 AMDGPULibFunc::Param Prev;
582 bool parseItaniumParam(StringRef& param, AMDGPULibFunc::Param &res);
583 };
584 } // namespace
585
586 bool ItaniumParamParser::parseItaniumParam(StringRef& param,
587 AMDGPULibFunc::Param &res) {
588 res.reset();
589 if (param.empty()) return false;
590
591 // parse pointer prefix
592 if (eatTerm(param, 'P')) {
593 if (eatTerm(param, 'K')) res.PtrKind |= AMDGPULibFunc::CONST;
594 if (eatTerm(param, 'V')) res.PtrKind |= AMDGPULibFunc::VOLATILE;
595 unsigned AS;
596 if (!eatTerm(param, "U3AS")) {
597 AS = 0;
598 } else {
599 AS = param.front() - '0';
600 drop_front(param, 1);
601 }
602 res.PtrKind |= AMDGPULibFuncBase::getEPtrKindFromAddrSpace(AS);
603 } else {
604 res.PtrKind = AMDGPULibFunc::BYVALUE;
605 }
606
607 // parse vector size
608 if (eatTerm(param,"Dv")) {
609 res.VectorSize = parseVecSize(param);
610 if (res.VectorSize==1 || !eatTerm(param, '_')) return false;
611 }
612
613 // parse type
614 char const TC = param.front();
615 if (::isDigit(TC)) {
616 res.ArgType = StringSwitch<AMDGPULibFunc::EType>
617 (eatLengthPrefixedName(param))
618 .Case("ocl_image1darray" , AMDGPULibFunc::IMG1DA)
619 .Case("ocl_image1dbuffer", AMDGPULibFunc::IMG1DB)
620 .Case("ocl_image2darray" , AMDGPULibFunc::IMG2DA)
621 .Case("ocl_image1d" , AMDGPULibFunc::IMG1D)
622 .Case("ocl_image2d" , AMDGPULibFunc::IMG2D)
623 .Case("ocl_image3d" , AMDGPULibFunc::IMG3D)
624 .Case("ocl_event" , AMDGPULibFunc::DUMMY)
625 .Case("ocl_sampler" , AMDGPULibFunc::DUMMY)
626 .Default(AMDGPULibFunc::DUMMY);
627 } else {
628 drop_front(param);
629 switch (TC) {
630 case 'h': res.ArgType = AMDGPULibFunc::U8; break;
631 case 't': res.ArgType = AMDGPULibFunc::U16; break;
632 case 'j': res.ArgType = AMDGPULibFunc::U32; break;
633 case 'm': res.ArgType = AMDGPULibFunc::U64; break;
634 case 'c': res.ArgType = AMDGPULibFunc::I8; break;
635 case 's': res.ArgType = AMDGPULibFunc::I16; break;
636 case 'i': res.ArgType = AMDGPULibFunc::I32; break;
637 case 'l': res.ArgType = AMDGPULibFunc::I64; break;
638 case 'f': res.ArgType = AMDGPULibFunc::F32; break;
639 case 'd': res.ArgType = AMDGPULibFunc::F64; break;
640 case 'D': if (!eatTerm(param, 'h')) return false;
641 res.ArgType = AMDGPULibFunc::F16; break;
642 case 'S':
643 if (!eatTerm(param, '_')) {
644 eatNumber(param);
645 if (!eatTerm(param, '_')) return false;
646 }
647 res.VectorSize = Prev.VectorSize;
648 res.ArgType = Prev.ArgType;
649 break;
650 default:;
651 }
652 }
653 if (res.ArgType == 0) return false;
654 Prev.VectorSize = res.VectorSize;
655 Prev.ArgType = res.ArgType;
656 return true;
657 }
658
659 bool AMDGPUMangledLibFunc::parseFuncName(StringRef &mangledName) {
660 StringRef Name = eatLengthPrefixedName(mangledName);
661 FKind = parseNamePrefix(Name);
662 if (!parseUnmangledName(Name))
663 return false;
664
665 const ManglingRule& Rule = manglingRules[FuncId];
666 ItaniumParamParser Parser;
667 for (int I=0; I < Rule.maxLeadIndex(); ++I) {
668 Param P;
669 if (!Parser.parseItaniumParam(mangledName, P))
670 return false;
671
672 if ((I + 1) == Rule.Lead[0]) Leads[0] = P;
673 if ((I + 1) == Rule.Lead[1]) Leads[1] = P;
674 }
675 return true;
676 }
677
678 bool AMDGPUUnmangledLibFunc::parseFuncName(StringRef &Name) {
679 if (!UnmangledFuncInfo::lookup(Name, FuncId))
680 return false;
681 setName(Name);
682 return true;
683 }
684
685 bool AMDGPULibFunc::parse(StringRef FuncName, AMDGPULibFunc &F) {
686 if (FuncName.empty()) {
687 F.Impl = std::unique_ptr<AMDGPULibFuncImpl>();
688 return false;
689 }
690
691 if (eatTerm(FuncName, "_Z"))
692 F.Impl = make_unique<AMDGPUMangledLibFunc>();
693 else
694 F.Impl = make_unique<AMDGPUUnmangledLibFunc>();
695 if (F.Impl->parseFuncName(FuncName))
696 return true;
697
698 F.Impl = std::unique_ptr<AMDGPULibFuncImpl>();
699 return false;
700 }
701
702 StringRef AMDGPUMangledLibFunc::getUnmangledName(StringRef mangledName) {
703 StringRef S = mangledName;
704 if (eatTerm(S, "_Z"))
705 return eatLengthPrefixedName(S);
706 return StringRef();
707 }
708
709 ///////////////////////////////////////////////////////////////////////////////
710 // Mangling
711
712 template <typename Stream>
713 void AMDGPUMangledLibFunc::writeName(Stream &OS) const {
714 const char *Pfx = "";
715 switch (FKind) {
716 case NATIVE: Pfx = "native_"; break;
717 case HALF: Pfx = "half_"; break;
718 default: break;
719 }
720 if (!Name.empty()) {
721 OS << Pfx << Name;
722 } else if (FuncId != EI_NONE) {
723 OS << Pfx;
724 const StringRef& S = manglingRules[FuncId].Name;
725 OS.write(S.data(), S.size());
726 }
727 }
728
729 std::string AMDGPUMangledLibFunc::mangle() const { return mangleNameItanium(); }
730
731 ///////////////////////////////////////////////////////////////////////////////
732 // Itanium Mangling
733
734 static const char *getItaniumTypeName(AMDGPULibFunc::EType T) {
735 switch (T) {
736 case AMDGPULibFunc::U8: return "h";
737 case AMDGPULibFunc::U16: return "t";
738 case AMDGPULibFunc::U32: return "j";
739 case AMDGPULibFunc::U64: return "m";
740 case AMDGPULibFunc::I8: return "c";
741 case AMDGPULibFunc::I16: return "s";
742 case AMDGPULibFunc::I32: return "i";
743 case AMDGPULibFunc::I64: return "l";
744 case AMDGPULibFunc::F16: return "Dh";
745 case AMDGPULibFunc::F32: return "f";
746 case AMDGPULibFunc::F64: return "d";
747 case AMDGPULibFunc::IMG1DA: return "16ocl_image1darray";
748 case AMDGPULibFunc::IMG1DB: return "17ocl_image1dbuffer";
749 case AMDGPULibFunc::IMG2DA: return "16ocl_image2darray";
750 case AMDGPULibFunc::IMG1D: return "11ocl_image1d";
751 case AMDGPULibFunc::IMG2D: return "11ocl_image2d";
752 case AMDGPULibFunc::IMG3D: return "11ocl_image3d";
753 case AMDGPULibFunc::SAMPLER: return "11ocl_sampler";
754 case AMDGPULibFunc::EVENT: return "9ocl_event";
755 default: llvm_unreachable("Unhandeled param type");
756 }
757 return nullptr;
758 }
759
760 namespace {
761 // Itanium mangling ABI says:
762 // "5.1.8. Compression
763 // ... Each non-terminal in the grammar for which <substitution> appears on the
764 // right-hand side is both a source of future substitutions and a candidate
765 // for being substituted. There are two exceptions that appear to be
766 // substitution candidates from the grammar, but are explicitly excluded:
767 // 1. <builtin-type> other than vendor extended types ..."
768
769 // For the purpose of functions the following productions make sence for the
770 // substitution:
771 // <type> ::= <builtin-type>
772 // ::= <class-enum-type>
773 // ::= <array-type>
774 // ::=<CV-qualifiers> <type>
775 // ::= P <type> # pointer-to
776 // ::= <substitution>
777 //
778 // Note that while types like images, samplers and events are by the ABI encoded
779 // using <class-enum-type> production rule they're not used for substitution
780 // because clang consider them as builtin types.
781 //
782 // DvNN_ type is GCC extension for vectors and is a subject for the substitution.
783
784
785 class ItaniumMangler {
786 SmallVector<AMDGPULibFunc::Param, 10> Str; // list of accumulated substituions
787 bool UseAddrSpace;
788
789 int findSubst(const AMDGPULibFunc::Param& P) const {
790 for(unsigned I = 0; I < Str.size(); ++I) {
791 const AMDGPULibFunc::Param& T = Str[I];
792 if (P.PtrKind == T.PtrKind &&
793 P.VectorSize == T.VectorSize &&
794 P.ArgType == T.ArgType) {
795 return I;
796 }
797 }
798 return -1;
799 }
800
801 template <typename Stream>
802 bool trySubst(Stream& os, const AMDGPULibFunc::Param& p) {
803 int const subst = findSubst(p);
804 if (subst < 0) return false;
805 // Substitutions are mangled as S(XX)?_ where XX is a hexadecimal number
806 // 0 1 2
807 // S_ S0_ S1_
808 if (subst == 0) os << "S_";
809 else os << 'S' << (subst-1) << '_';
810 return true;
811 }
812
813 public:
814 ItaniumMangler(bool useAddrSpace)
815 : UseAddrSpace(useAddrSpace) {}
816
817 template <typename Stream>
818 void operator()(Stream& os, AMDGPULibFunc::Param p) {
819
820 // Itanium mangling ABI 5.1.8. Compression:
821 // Logically, the substitutable components of a mangled name are considered
822 // left-to-right, components before the composite structure of which they
823 // are a part. If a component has been encountered before, it is substituted
824 // as described below. This decision is independent of whether its components
825 // have been substituted, so an implementation may optimize by considering
826 // large structures for substitution before their components. If a component
827 // has not been encountered before, its mangling is identified, and it is
828 // added to a dictionary of substitution candidates. No entity is added to
829 // the dictionary twice.
830 AMDGPULibFunc::Param Ptr;
831
832 if (p.PtrKind) {
833 if (trySubst(os, p)) return;
834 os << 'P';
835 if (p.PtrKind & AMDGPULibFunc::CONST) os << 'K';
836 if (p.PtrKind & AMDGPULibFunc::VOLATILE) os << 'V';
837 unsigned AS = UseAddrSpace
838 ? AMDGPULibFuncBase::getAddrSpaceFromEPtrKind(p.PtrKind)
839 : 0;
840 if (AS != 0) os << "U3AS" << AS;
841 Ptr = p;
842 p.PtrKind = 0;
843 }
844
845 if (p.VectorSize > 1) {
846 if (trySubst(os, p)) goto exit;
847 Str.push_back(p);
848 os << "Dv" << static_cast<unsigned>(p.VectorSize) << '_';
849 }
850
851 os << getItaniumTypeName((AMDGPULibFunc::EType)p.ArgType);
852
853 exit:
854 if (Ptr.ArgType) Str.push_back(Ptr);
855 }
856 };
857 } // namespace
858
859 std::string AMDGPUMangledLibFunc::mangleNameItanium() const {
860 SmallString<128> Buf;
861 raw_svector_ostream S(Buf);
862 SmallString<128> NameBuf;
863 raw_svector_ostream Name(NameBuf);
864 writeName(Name);
865 const StringRef& NameStr = Name.str();
866 S << "_Z" << static_cast<int>(NameStr.size()) << NameStr;
867
868 ItaniumMangler Mangler(true);
869 ParamIterator I(Leads, manglingRules[FuncId]);
870 Param P;
871 while ((P = I.getNextParam()).ArgType != 0)
872 Mangler(S, P);
873 return S.str();
874 }
875
876 ///////////////////////////////////////////////////////////////////////////////
877 // Misc
878
879 static Type* getIntrinsicParamType(
880 LLVMContext& C,
881 const AMDGPULibFunc::Param& P,
882 bool useAddrSpace) {
883 Type* T = nullptr;
884 switch (P.ArgType) {
885 case AMDGPULibFunc::U8:
886 case AMDGPULibFunc::I8: T = Type::getInt8Ty(C); break;
887 case AMDGPULibFunc::U16:
888 case AMDGPULibFunc::I16: T = Type::getInt16Ty(C); break;
889 case AMDGPULibFunc::U32:
890 case AMDGPULibFunc::I32: T = Type::getInt32Ty(C); break;
891 case AMDGPULibFunc::U64:
892 case AMDGPULibFunc::I64: T = Type::getInt64Ty(C); break;
893 case AMDGPULibFunc::F16: T = Type::getHalfTy(C); break;
894 case AMDGPULibFunc::F32: T = Type::getFloatTy(C); break;
895 case AMDGPULibFunc::F64: T = Type::getDoubleTy(C); break;
896
897 case AMDGPULibFunc::IMG1DA:
898 case AMDGPULibFunc::IMG1DB:
899 case AMDGPULibFunc::IMG2DA:
900 case AMDGPULibFunc::IMG1D:
901 case AMDGPULibFunc::IMG2D:
902 case AMDGPULibFunc::IMG3D:
903 T = StructType::create(C,"ocl_image")->getPointerTo(); break;
904 case AMDGPULibFunc::SAMPLER:
905 T = StructType::create(C,"ocl_sampler")->getPointerTo(); break;
906 case AMDGPULibFunc::EVENT:
907 T = StructType::create(C,"ocl_event")->getPointerTo(); break;
908 default:
909 llvm_unreachable("Unhandeled param type");
910 return nullptr;
911 }
912 if (P.VectorSize > 1)
913 T = VectorType::get(T, P.VectorSize);
914 if (P.PtrKind != AMDGPULibFunc::BYVALUE)
915 T = useAddrSpace ? T->getPointerTo((P.PtrKind & AMDGPULibFunc::ADDR_SPACE)
916 - 1)
917 : T->getPointerTo();
918 return T;
919 }
920
921 FunctionType *AMDGPUMangledLibFunc::getFunctionType(Module &M) const {
922 LLVMContext& C = M.getContext();
923 std::vector<Type*> Args;
924 ParamIterator I(Leads, manglingRules[FuncId]);
925 Param P;
926 while ((P=I.getNextParam()).ArgType != 0)
927 Args.push_back(getIntrinsicParamType(C, P, true));
928
929 return FunctionType::get(
930 getIntrinsicParamType(C, getRetType(FuncId, Leads), true),
931 Args, false);
932 }
933
934 unsigned AMDGPUMangledLibFunc::getNumArgs() const {
935 return manglingRules[FuncId].getNumArgs();
936 }
937
938 unsigned AMDGPUUnmangledLibFunc::getNumArgs() const {
939 return UnmangledFuncInfo::getNumArgs(FuncId);
940 }
941
942 std::string AMDGPUMangledLibFunc::getName() const {
943 SmallString<128> Buf;
944 raw_svector_ostream OS(Buf);
945 writeName(OS);
946 return OS.str();
947 }
948
949 Function *AMDGPULibFunc::getFunction(Module *M, const AMDGPULibFunc &fInfo) {
950 std::string FuncName = fInfo.mangle();
951 Function *F = dyn_cast_or_null<Function>(
952 M->getValueSymbolTable().lookup(FuncName));
953
954 // check formal with actual types conformance
955 if (F && !F->isDeclaration()
956 && !F->isVarArg()
957 && F->arg_size() == fInfo.getNumArgs()) {
958 return F;
959 }
960 return nullptr;
961 }
962
963 FunctionCallee AMDGPULibFunc::getOrInsertFunction(Module *M,
964 const AMDGPULibFunc &fInfo) {
965 std::string const FuncName = fInfo.mangle();
966 Function *F = dyn_cast_or_null<Function>(
967 M->getValueSymbolTable().lookup(FuncName));
968
969 // check formal with actual types conformance
970 if (F && !F->isDeclaration()
971 && !F->isVarArg()
972 && F->arg_size() == fInfo.getNumArgs()) {
973 return F;
974 }
975
976 FunctionType *FuncTy = fInfo.getFunctionType(*M);
977
978 bool hasPtr = false;
979 for (FunctionType::param_iterator
980 PI = FuncTy->param_begin(),
981 PE = FuncTy->param_end();
982 PI != PE; ++PI) {
983 const Type* argTy = static_cast<const Type*>(*PI);
984 if (argTy->isPointerTy()) {
985 hasPtr = true;
986 break;
987 }
988 }
989
990 FunctionCallee C;
991 if (hasPtr) {
992 // Do not set extra attributes for functions with pointer arguments.
993 C = M->getOrInsertFunction(FuncName, FuncTy);
994 } else {
995 AttributeList Attr;
996 LLVMContext &Ctx = M->getContext();
997 Attr = Attr.addAttribute(Ctx, AttributeList::FunctionIndex,
998 Attribute::ReadOnly);
999 Attr = Attr.addAttribute(Ctx, AttributeList::FunctionIndex,
1000 Attribute::NoUnwind);
1001 C = M->getOrInsertFunction(FuncName, FuncTy, Attr);
1002 }
1003
1004 return C;
1005 }
1006
1007 bool UnmangledFuncInfo::lookup(StringRef Name, ID &Id) {
1008 auto Loc = Map.find(Name);
1009 if (Loc != Map.end()) {
1010 Id = toFuncId(Loc->second);
1011 return true;
1012 }
1013 Id = AMDGPULibFunc::EI_NONE;
1014 return false;
1015 }
1016
1017 AMDGPULibFunc::AMDGPULibFunc(const AMDGPULibFunc &F) {
1018 if (auto *MF = dyn_cast<AMDGPUMangledLibFunc>(F.Impl.get()))
1019 Impl.reset(new AMDGPUMangledLibFunc(*MF));
1020 else if (auto *UMF = dyn_cast<AMDGPUUnmangledLibFunc>(F.Impl.get()))
1021 Impl.reset(new AMDGPUUnmangledLibFunc(*UMF));
1022 else
1023 Impl = std::unique_ptr<AMDGPULibFuncImpl>();
1024 }
1025
1026 AMDGPULibFunc &AMDGPULibFunc::operator=(const AMDGPULibFunc &F) {
1027 if (this == &F)
1028 return *this;
1029 new (this) AMDGPULibFunc(F);
1030 return *this;
1031 }
1032
1033 AMDGPULibFunc::AMDGPULibFunc(EFuncId Id, const AMDGPULibFunc &CopyFrom) {
1034 assert(AMDGPULibFuncBase::isMangled(Id) && CopyFrom.isMangled() &&
1035 "not supported");
1036 Impl.reset(new AMDGPUMangledLibFunc(
1037 Id, *cast<AMDGPUMangledLibFunc>(CopyFrom.Impl.get())));
1038 }
1039
1040 AMDGPULibFunc::AMDGPULibFunc(StringRef Name, FunctionType *FT) {
1041 Impl.reset(new AMDGPUUnmangledLibFunc(Name, FT));
1042 }
1043
1044 void AMDGPULibFunc::initMangled() { Impl.reset(new AMDGPUMangledLibFunc()); }
1045
1046 AMDGPULibFunc::Param *AMDGPULibFunc::getLeads() {
1047 if (!Impl)
1048 initMangled();
1049 return cast<AMDGPUMangledLibFunc>(Impl.get())->Leads;
1050 }
1051
1052 const AMDGPULibFunc::Param *AMDGPULibFunc::getLeads() const {
1053 return cast<const AMDGPUMangledLibFunc>(Impl.get())->Leads;
1054 }
The low level virtual machine