1 //
2 // Copyright (C) 2013 LunarG, Inc.
3 // Copyright (C) 2017 ARM Limited.
4 // Copyright (C) 2015-2018 Google, Inc.
5 //
6 // All rights reserved.
7 //
8 // Redistribution and use in source and binary forms, with or without
9 // modification, are permitted provided that the following conditions
10 // are met:
11 //
12 // Redistributions of source code must retain the above copyright
13 // notice, this list of conditions and the following disclaimer.
14 //
15 // Redistributions in binary form must reproduce the above
16 // copyright notice, this list of conditions and the following
17 // disclaimer in the documentation and/or other materials provided
18 // with the distribution.
19 //
20 // Neither the name of 3Dlabs Inc. Ltd. nor the names of its
21 // contributors may be used to endorse or promote products derived
22 // from this software without specific prior written permission.
23 //
24 // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
25 // "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
26 // LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
27 // FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
28 // COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
29 // INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
30 // BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
31 // LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
32 // CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
33 // LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
34 // ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
35 // POSSIBILITY OF SUCH DAMAGE.
36 //
37
38 //
39 // Do link-time merging and validation of intermediate representations.
40 //
41 // Basic model is that during compilation, each compilation unit (shader) is
42 // compiled into one TIntermediate instance. Then, at link time, multiple
43 // units for the same stage can be merged together, which can generate errors.
44 // Then, after all merging, a single instance of TIntermediate represents
45 // the whole stage. A final error check can be done on the resulting stage,
46 // even if no merging was done (i.e., the stage was only one compilation unit).
47 //
48
49 #include "localintermediate.h"
50 #include "../Include/InfoSink.h"
51 #include "SymbolTable.h"
52
53 namespace glslang {
54
55 //
56 // Link-time error emitter.
57 //
error(TInfoSink & infoSink,const char * message,EShLanguage unitStage)58 void TIntermediate::error(TInfoSink& infoSink, const char* message, EShLanguage unitStage)
59 {
60 #ifndef GLSLANG_WEB
61 infoSink.info.prefix(EPrefixError);
62 if (unitStage < EShLangCount)
63 infoSink.info << "Linking " << StageName(getStage()) << " and " << StageName(unitStage) << " stages: " << message << "\n";
64 else
65 infoSink.info << "Linking " << StageName(language) << " stage: " << message << "\n";
66 #endif
67
68 ++numErrors;
69 }
70
71 // Link-time warning.
warn(TInfoSink & infoSink,const char * message,EShLanguage unitStage)72 void TIntermediate::warn(TInfoSink& infoSink, const char* message, EShLanguage unitStage)
73 {
74 #ifndef GLSLANG_WEB
75 infoSink.info.prefix(EPrefixWarning);
76 if (unitStage < EShLangCount)
77 infoSink.info << "Linking " << StageName(language) << " and " << StageName(unitStage) << " stages: " << message << "\n";
78 else
79 infoSink.info << "Linking " << StageName(language) << " stage: " << message << "\n";
80 #endif
81 }
82
83 // TODO: 4.4 offset/align: "Two blocks linked together in the same program with the same block
84 // name must have the exact same set of members qualified with offset and their integral-constant
85 // expression values must be the same, or a link-time error results."
86
87 //
88 // Merge the information from 'unit' into 'this'
89 //
merge(TInfoSink & infoSink,TIntermediate & unit)90 void TIntermediate::merge(TInfoSink& infoSink, TIntermediate& unit)
91 {
92 #if !defined(GLSLANG_WEB)
93 mergeCallGraphs(infoSink, unit);
94 mergeModes(infoSink, unit);
95 mergeTrees(infoSink, unit);
96 #endif
97 }
98
99 //
100 // check that link objects between stages
101 //
mergeUniformObjects(TInfoSink & infoSink,TIntermediate & unit)102 void TIntermediate::mergeUniformObjects(TInfoSink& infoSink, TIntermediate& unit) {
103 if (unit.treeRoot == nullptr || treeRoot == nullptr)
104 return;
105
106 // Get the linker-object lists
107 TIntermSequence& linkerObjects = findLinkerObjects()->getSequence();
108 TIntermSequence unitLinkerObjects = unit.findLinkerObjects()->getSequence();
109
110 // filter unitLinkerObjects to only contain uniforms
111 auto end = std::remove_if(unitLinkerObjects.begin(), unitLinkerObjects.end(),
112 [](TIntermNode* node) {return node->getAsSymbolNode()->getQualifier().storage != EvqUniform &&
113 node->getAsSymbolNode()->getQualifier().storage != EvqBuffer; });
114 unitLinkerObjects.resize(end - unitLinkerObjects.begin());
115
116 // merge uniforms and do error checking
117 bool mergeExistingOnly = false;
118 mergeGlobalUniformBlocks(infoSink, unit, mergeExistingOnly);
119 mergeLinkerObjects(infoSink, linkerObjects, unitLinkerObjects, unit.getStage());
120 }
121
122 //
123 // do error checking on the shader boundary in / out vars
124 //
checkStageIO(TInfoSink & infoSink,TIntermediate & unit)125 void TIntermediate::checkStageIO(TInfoSink& infoSink, TIntermediate& unit) {
126 if (unit.treeRoot == nullptr || treeRoot == nullptr)
127 return;
128
129 // Get copies of the linker-object lists
130 TIntermSequence linkerObjects = findLinkerObjects()->getSequence();
131 TIntermSequence unitLinkerObjects = unit.findLinkerObjects()->getSequence();
132
133 // filter linkerObjects to only contain out variables
134 auto end = std::remove_if(linkerObjects.begin(), linkerObjects.end(),
135 [](TIntermNode* node) {return node->getAsSymbolNode()->getQualifier().storage != EvqVaryingOut; });
136 linkerObjects.resize(end - linkerObjects.begin());
137
138 // filter unitLinkerObjects to only contain in variables
139 auto unitEnd = std::remove_if(unitLinkerObjects.begin(), unitLinkerObjects.end(),
140 [](TIntermNode* node) {return node->getAsSymbolNode()->getQualifier().storage != EvqVaryingIn; });
141 unitLinkerObjects.resize(unitEnd - unitLinkerObjects.begin());
142
143 // do matching and error checking
144 mergeLinkerObjects(infoSink, linkerObjects, unitLinkerObjects, unit.getStage());
145
146 // TODO: final check; make sure that any statically used `in` have matching `out` written to
147 }
148
mergeCallGraphs(TInfoSink & infoSink,TIntermediate & unit)149 void TIntermediate::mergeCallGraphs(TInfoSink& infoSink, TIntermediate& unit)
150 {
151 if (unit.getNumEntryPoints() > 0) {
152 if (getNumEntryPoints() > 0)
153 error(infoSink, "can't handle multiple entry points per stage");
154 else {
155 entryPointName = unit.getEntryPointName();
156 entryPointMangledName = unit.getEntryPointMangledName();
157 }
158 }
159 numEntryPoints += unit.getNumEntryPoints();
160
161 callGraph.insert(callGraph.end(), unit.callGraph.begin(), unit.callGraph.end());
162 }
163
164 #if !defined(GLSLANG_WEB)
165
166 #define MERGE_MAX(member) member = std::max(member, unit.member)
167 #define MERGE_TRUE(member) if (unit.member) member = unit.member;
168
mergeModes(TInfoSink & infoSink,TIntermediate & unit)169 void TIntermediate::mergeModes(TInfoSink& infoSink, TIntermediate& unit)
170 {
171 if (language != unit.language)
172 error(infoSink, "stages must match when linking into a single stage");
173
174 if (getSource() == EShSourceNone)
175 setSource(unit.getSource());
176 if (getSource() != unit.getSource())
177 error(infoSink, "can't link compilation units from different source languages");
178
179 if (treeRoot == nullptr) {
180 profile = unit.profile;
181 version = unit.version;
182 requestedExtensions = unit.requestedExtensions;
183 } else {
184 if ((isEsProfile()) != (unit.isEsProfile()))
185 error(infoSink, "Cannot cross link ES and desktop profiles");
186 else if (unit.profile == ECompatibilityProfile)
187 profile = ECompatibilityProfile;
188 version = std::max(version, unit.version);
189 requestedExtensions.insert(unit.requestedExtensions.begin(), unit.requestedExtensions.end());
190 }
191
192 MERGE_MAX(spvVersion.spv);
193 MERGE_MAX(spvVersion.vulkanGlsl);
194 MERGE_MAX(spvVersion.vulkan);
195 MERGE_MAX(spvVersion.openGl);
196 MERGE_TRUE(spvVersion.vulkanRelaxed);
197
198 numErrors += unit.getNumErrors();
199 // Only one push_constant is allowed, mergeLinkerObjects() will ensure the push_constant
200 // is the same for all units.
201 if (numPushConstants > 1 || unit.numPushConstants > 1)
202 error(infoSink, "Only one push_constant block is allowed per stage");
203 numPushConstants = std::min(numPushConstants + unit.numPushConstants, 1);
204
205 if (unit.invocations != TQualifier::layoutNotSet) {
206 if (invocations == TQualifier::layoutNotSet)
207 invocations = unit.invocations;
208 else if (invocations != unit.invocations)
209 error(infoSink, "number of invocations must match between compilation units");
210 }
211
212 if (vertices == TQualifier::layoutNotSet)
213 vertices = unit.vertices;
214 else if (unit.vertices != TQualifier::layoutNotSet && vertices != unit.vertices) {
215 if (language == EShLangGeometry || language == EShLangMesh)
216 error(infoSink, "Contradictory layout max_vertices values");
217 else if (language == EShLangTessControl)
218 error(infoSink, "Contradictory layout vertices values");
219 else
220 assert(0);
221 }
222 if (primitives == TQualifier::layoutNotSet)
223 primitives = unit.primitives;
224 else if (primitives != unit.primitives) {
225 if (language == EShLangMesh)
226 error(infoSink, "Contradictory layout max_primitives values");
227 else
228 assert(0);
229 }
230
231 if (inputPrimitive == ElgNone)
232 inputPrimitive = unit.inputPrimitive;
233 else if (unit.inputPrimitive != ElgNone && inputPrimitive != unit.inputPrimitive)
234 error(infoSink, "Contradictory input layout primitives");
235
236 if (outputPrimitive == ElgNone)
237 outputPrimitive = unit.outputPrimitive;
238 else if (unit.outputPrimitive != ElgNone && outputPrimitive != unit.outputPrimitive)
239 error(infoSink, "Contradictory output layout primitives");
240
241 if (originUpperLeft != unit.originUpperLeft || pixelCenterInteger != unit.pixelCenterInteger)
242 error(infoSink, "gl_FragCoord redeclarations must match across shaders");
243
244 if (vertexSpacing == EvsNone)
245 vertexSpacing = unit.vertexSpacing;
246 else if (vertexSpacing != unit.vertexSpacing)
247 error(infoSink, "Contradictory input vertex spacing");
248
249 if (vertexOrder == EvoNone)
250 vertexOrder = unit.vertexOrder;
251 else if (vertexOrder != unit.vertexOrder)
252 error(infoSink, "Contradictory triangle ordering");
253
254 MERGE_TRUE(pointMode);
255
256 for (int i = 0; i < 3; ++i) {
257 if (unit.localSizeNotDefault[i]) {
258 if (!localSizeNotDefault[i]) {
259 localSize[i] = unit.localSize[i];
260 localSizeNotDefault[i] = true;
261 }
262 else if (localSize[i] != unit.localSize[i])
263 error(infoSink, "Contradictory local size");
264 }
265
266 if (localSizeSpecId[i] == TQualifier::layoutNotSet)
267 localSizeSpecId[i] = unit.localSizeSpecId[i];
268 else if (localSizeSpecId[i] != unit.localSizeSpecId[i])
269 error(infoSink, "Contradictory local size specialization ids");
270 }
271
272 MERGE_TRUE(earlyFragmentTests);
273 MERGE_TRUE(postDepthCoverage);
274
275 if (depthLayout == EldNone)
276 depthLayout = unit.depthLayout;
277 else if (depthLayout != unit.depthLayout)
278 error(infoSink, "Contradictory depth layouts");
279
280 MERGE_TRUE(depthReplacing);
281 MERGE_TRUE(hlslFunctionality1);
282
283 blendEquations |= unit.blendEquations;
284
285 MERGE_TRUE(xfbMode);
286
287 for (size_t b = 0; b < xfbBuffers.size(); ++b) {
288 if (xfbBuffers[b].stride == TQualifier::layoutXfbStrideEnd)
289 xfbBuffers[b].stride = unit.xfbBuffers[b].stride;
290 else if (xfbBuffers[b].stride != unit.xfbBuffers[b].stride)
291 error(infoSink, "Contradictory xfb_stride");
292 xfbBuffers[b].implicitStride = std::max(xfbBuffers[b].implicitStride, unit.xfbBuffers[b].implicitStride);
293 if (unit.xfbBuffers[b].contains64BitType)
294 xfbBuffers[b].contains64BitType = true;
295 if (unit.xfbBuffers[b].contains32BitType)
296 xfbBuffers[b].contains32BitType = true;
297 if (unit.xfbBuffers[b].contains16BitType)
298 xfbBuffers[b].contains16BitType = true;
299 // TODO: 4.4 link: enhanced layouts: compare ranges
300 }
301
302 MERGE_TRUE(multiStream);
303 MERGE_TRUE(layoutOverrideCoverage);
304 MERGE_TRUE(geoPassthroughEXT);
305
306 for (unsigned int i = 0; i < unit.shiftBinding.size(); ++i) {
307 if (unit.shiftBinding[i] > 0)
308 setShiftBinding((TResourceType)i, unit.shiftBinding[i]);
309 }
310
311 for (unsigned int i = 0; i < unit.shiftBindingForSet.size(); ++i) {
312 for (auto it = unit.shiftBindingForSet[i].begin(); it != unit.shiftBindingForSet[i].end(); ++it)
313 setShiftBindingForSet((TResourceType)i, it->second, it->first);
314 }
315
316 resourceSetBinding.insert(resourceSetBinding.end(), unit.resourceSetBinding.begin(), unit.resourceSetBinding.end());
317
318 MERGE_TRUE(autoMapBindings);
319 MERGE_TRUE(autoMapLocations);
320 MERGE_TRUE(invertY);
321 MERGE_TRUE(dxPositionW);
322 MERGE_TRUE(debugInfo);
323 MERGE_TRUE(flattenUniformArrays);
324 MERGE_TRUE(useUnknownFormat);
325 MERGE_TRUE(hlslOffsets);
326 MERGE_TRUE(useStorageBuffer);
327 MERGE_TRUE(invariantAll);
328 MERGE_TRUE(hlslIoMapping);
329
330 // TODO: sourceFile
331 // TODO: sourceText
332 // TODO: processes
333
334 MERGE_TRUE(needToLegalize);
335 MERGE_TRUE(binaryDoubleOutput);
336 MERGE_TRUE(usePhysicalStorageBuffer);
337 }
338
339 //
340 // Merge the 'unit' AST into 'this' AST.
341 // That includes rationalizing the unique IDs, which were set up independently,
342 // and might have overlaps that are not the same symbol, or might have different
343 // IDs for what should be the same shared symbol.
344 //
mergeTrees(TInfoSink & infoSink,TIntermediate & unit)345 void TIntermediate::mergeTrees(TInfoSink& infoSink, TIntermediate& unit)
346 {
347 if (unit.treeRoot == nullptr)
348 return;
349
350 if (treeRoot == nullptr) {
351 treeRoot = unit.treeRoot;
352 return;
353 }
354
355 // Getting this far means we have two existing trees to merge...
356 numShaderRecordBlocks += unit.numShaderRecordBlocks;
357 numTaskNVBlocks += unit.numTaskNVBlocks;
358
359 // Get the top-level globals of each unit
360 TIntermSequence& globals = treeRoot->getAsAggregate()->getSequence();
361 TIntermSequence& unitGlobals = unit.treeRoot->getAsAggregate()->getSequence();
362
363 // Get the linker-object lists
364 TIntermSequence& linkerObjects = findLinkerObjects()->getSequence();
365 const TIntermSequence& unitLinkerObjects = unit.findLinkerObjects()->getSequence();
366
367 // Map by global name to unique ID to rationalize the same object having
368 // differing IDs in different trees.
369 TIdMaps idMaps;
370 long long idShift;
371 seedIdMap(idMaps, idShift);
372 remapIds(idMaps, idShift + 1, unit);
373
374 mergeBodies(infoSink, globals, unitGlobals);
375 bool mergeExistingOnly = false;
376 mergeGlobalUniformBlocks(infoSink, unit, mergeExistingOnly);
377 mergeLinkerObjects(infoSink, linkerObjects, unitLinkerObjects, unit.getStage());
378 ioAccessed.insert(unit.ioAccessed.begin(), unit.ioAccessed.end());
379 }
380
381 #endif
382
getNameForIdMap(TIntermSymbol * symbol)383 static const TString& getNameForIdMap(TIntermSymbol* symbol)
384 {
385 TShaderInterface si = symbol->getType().getShaderInterface();
386 if (si == EsiNone)
387 return symbol->getName();
388 else
389 return symbol->getType().getTypeName();
390 }
391
392
393
394 // Traverser that seeds an ID map with all built-ins, and tracks the
395 // maximum ID used, currently using (maximum ID + 1) as new symbol id shift seed.
396 // Level id will keep same after shifting.
397 // (It would be nice to put this in a function, but that causes warnings
398 // on having no bodies for the copy-constructor/operator=.)
399 class TBuiltInIdTraverser : public TIntermTraverser {
400 public:
TBuiltInIdTraverser(TIdMaps & idMaps)401 TBuiltInIdTraverser(TIdMaps& idMaps) : idMaps(idMaps), idShift(0) { }
402 // If it's a built in, add it to the map.
visitSymbol(TIntermSymbol * symbol)403 virtual void visitSymbol(TIntermSymbol* symbol)
404 {
405 const TQualifier& qualifier = symbol->getType().getQualifier();
406 if (qualifier.builtIn != EbvNone) {
407 TShaderInterface si = symbol->getType().getShaderInterface();
408 idMaps[si][getNameForIdMap(symbol)] = symbol->getId();
409 }
410 idShift = (symbol->getId() & ~TSymbolTable::uniqueIdMask) |
411 std::max(idShift & TSymbolTable::uniqueIdMask,
412 symbol->getId() & TSymbolTable::uniqueIdMask);
413 }
getIdShift() const414 long long getIdShift() const { return idShift; }
415 protected:
416 TBuiltInIdTraverser(TBuiltInIdTraverser&);
417 TBuiltInIdTraverser& operator=(TBuiltInIdTraverser&);
418 TIdMaps& idMaps;
419 long long idShift;
420 };
421
422 // Traverser that seeds an ID map with non-builtins.
423 // (It would be nice to put this in a function, but that causes warnings
424 // on having no bodies for the copy-constructor/operator=.)
425 class TUserIdTraverser : public TIntermTraverser {
426 public:
TUserIdTraverser(TIdMaps & idMaps)427 TUserIdTraverser(TIdMaps& idMaps) : idMaps(idMaps) { }
428 // If its a non-built-in global, add it to the map.
visitSymbol(TIntermSymbol * symbol)429 virtual void visitSymbol(TIntermSymbol* symbol)
430 {
431 const TQualifier& qualifier = symbol->getType().getQualifier();
432 if (qualifier.builtIn == EbvNone) {
433 TShaderInterface si = symbol->getType().getShaderInterface();
434 idMaps[si][getNameForIdMap(symbol)] = symbol->getId();
435 }
436 }
437
438 protected:
439 TUserIdTraverser(TUserIdTraverser&);
440 TUserIdTraverser& operator=(TUserIdTraverser&);
441 TIdMaps& idMaps; // over biggest id
442 };
443
444 // Initialize the the ID map with what we know of 'this' AST.
seedIdMap(TIdMaps & idMaps,long long & idShift)445 void TIntermediate::seedIdMap(TIdMaps& idMaps, long long& idShift)
446 {
447 // all built-ins everywhere need to align on IDs and contribute to the max ID
448 TBuiltInIdTraverser builtInIdTraverser(idMaps);
449 treeRoot->traverse(&builtInIdTraverser);
450 idShift = builtInIdTraverser.getIdShift() & TSymbolTable::uniqueIdMask;
451
452 // user variables in the linker object list need to align on ids
453 TUserIdTraverser userIdTraverser(idMaps);
454 findLinkerObjects()->traverse(&userIdTraverser);
455 }
456
457 // Traverser to map an AST ID to what was known from the seeding AST.
458 // (It would be nice to put this in a function, but that causes warnings
459 // on having no bodies for the copy-constructor/operator=.)
460 class TRemapIdTraverser : public TIntermTraverser {
461 public:
TRemapIdTraverser(const TIdMaps & idMaps,long long idShift)462 TRemapIdTraverser(const TIdMaps& idMaps, long long idShift) : idMaps(idMaps), idShift(idShift) { }
463 // Do the mapping:
464 // - if the same symbol, adopt the 'this' ID
465 // - otherwise, ensure a unique ID by shifting to a new space
visitSymbol(TIntermSymbol * symbol)466 virtual void visitSymbol(TIntermSymbol* symbol)
467 {
468 const TQualifier& qualifier = symbol->getType().getQualifier();
469 bool remapped = false;
470 if (qualifier.isLinkable() || qualifier.builtIn != EbvNone) {
471 TShaderInterface si = symbol->getType().getShaderInterface();
472 auto it = idMaps[si].find(getNameForIdMap(symbol));
473 if (it != idMaps[si].end()) {
474 uint64_t id = (symbol->getId() & ~TSymbolTable::uniqueIdMask) |
475 (it->second & TSymbolTable::uniqueIdMask);
476 symbol->changeId(id);
477 remapped = true;
478 }
479 }
480 if (!remapped)
481 symbol->changeId(symbol->getId() + idShift);
482 }
483 protected:
484 TRemapIdTraverser(TRemapIdTraverser&);
485 TRemapIdTraverser& operator=(TRemapIdTraverser&);
486 const TIdMaps& idMaps;
487 long long idShift;
488 };
489
remapIds(const TIdMaps & idMaps,long long idShift,TIntermediate & unit)490 void TIntermediate::remapIds(const TIdMaps& idMaps, long long idShift, TIntermediate& unit)
491 {
492 // Remap all IDs to either share or be unique, as dictated by the idMap and idShift.
493 TRemapIdTraverser idTraverser(idMaps, idShift);
494 unit.getTreeRoot()->traverse(&idTraverser);
495 }
496
497 //
498 // Merge the function bodies and global-level initializers from unitGlobals into globals.
499 // Will error check duplication of function bodies for the same signature.
500 //
mergeBodies(TInfoSink & infoSink,TIntermSequence & globals,const TIntermSequence & unitGlobals)501 void TIntermediate::mergeBodies(TInfoSink& infoSink, TIntermSequence& globals, const TIntermSequence& unitGlobals)
502 {
503 // TODO: link-time performance: Processing in alphabetical order will be faster
504
505 // Error check the global objects, not including the linker objects
506 for (unsigned int child = 0; child < globals.size() - 1; ++child) {
507 for (unsigned int unitChild = 0; unitChild < unitGlobals.size() - 1; ++unitChild) {
508 TIntermAggregate* body = globals[child]->getAsAggregate();
509 TIntermAggregate* unitBody = unitGlobals[unitChild]->getAsAggregate();
510 if (body && unitBody && body->getOp() == EOpFunction && unitBody->getOp() == EOpFunction && body->getName() == unitBody->getName()) {
511 error(infoSink, "Multiple function bodies in multiple compilation units for the same signature in the same stage:");
512 infoSink.info << " " << globals[child]->getAsAggregate()->getName() << "\n";
513 }
514 }
515 }
516
517 // Merge the global objects, just in front of the linker objects
518 globals.insert(globals.end() - 1, unitGlobals.begin(), unitGlobals.end() - 1);
519 }
520
isSameInterface(TIntermSymbol * symbol,EShLanguage stage,TIntermSymbol * unitSymbol,EShLanguage unitStage)521 static inline bool isSameInterface(TIntermSymbol* symbol, EShLanguage stage, TIntermSymbol* unitSymbol, EShLanguage unitStage) {
522 return // 1) same stage and same shader interface
523 (stage == unitStage && symbol->getType().getShaderInterface() == unitSymbol->getType().getShaderInterface()) ||
524 // 2) accross stages and both are uniform or buffer
525 (symbol->getQualifier().storage == EvqUniform && unitSymbol->getQualifier().storage == EvqUniform) ||
526 (symbol->getQualifier().storage == EvqBuffer && unitSymbol->getQualifier().storage == EvqBuffer) ||
527 // 3) in/out matched across stage boundary
528 (stage < unitStage && symbol->getQualifier().storage == EvqVaryingOut && unitSymbol->getQualifier().storage == EvqVaryingIn) ||
529 (unitStage < stage && symbol->getQualifier().storage == EvqVaryingIn && unitSymbol->getQualifier().storage == EvqVaryingOut);
530 }
531
532 //
533 // Global Unfiform block stores any default uniforms (i.e. uniforms without a block)
534 // If two linked stages declare the same member, they are meant to be the same uniform
535 // and need to be in the same block
536 // merge the members of different stages to allow them to be linked properly
537 // as a single block
538 //
mergeGlobalUniformBlocks(TInfoSink & infoSink,TIntermediate & unit,bool mergeExistingOnly)539 void TIntermediate::mergeGlobalUniformBlocks(TInfoSink& infoSink, TIntermediate& unit, bool mergeExistingOnly)
540 {
541 TIntermSequence& linkerObjects = findLinkerObjects()->getSequence();
542 TIntermSequence& unitLinkerObjects = unit.findLinkerObjects()->getSequence();
543
544 // build lists of default blocks from the intermediates
545 TIntermSequence defaultBlocks;
546 TIntermSequence unitDefaultBlocks;
547
548 auto filter = [](TIntermSequence& list, TIntermNode* node) {
549 if (node->getAsSymbolNode()->getQualifier().defaultBlock) {
550 list.push_back(node);
551 }
552 };
553
554 std::for_each(linkerObjects.begin(), linkerObjects.end(),
555 [&defaultBlocks, &filter](TIntermNode* node) {
556 filter(defaultBlocks, node);
557 });
558 std::for_each(unitLinkerObjects.begin(), unitLinkerObjects.end(),
559 [&unitDefaultBlocks, &filter](TIntermNode* node) {
560 filter(unitDefaultBlocks, node);
561 });
562
563 auto itUnitBlock = unitDefaultBlocks.begin();
564 for (; itUnitBlock != unitDefaultBlocks.end(); itUnitBlock++) {
565
566 bool add = !mergeExistingOnly;
567 auto itBlock = defaultBlocks.begin();
568
569 for (; itBlock != defaultBlocks.end(); itBlock++) {
570 TIntermSymbol* block = (*itBlock)->getAsSymbolNode();
571 TIntermSymbol* unitBlock = (*itUnitBlock)->getAsSymbolNode();
572
573 assert(block && unitBlock);
574
575 // if the two default blocks match, then merge their definitions
576 if (block->getType().getTypeName() == unitBlock->getType().getTypeName() &&
577 block->getQualifier().storage == unitBlock->getQualifier().storage) {
578 add = false;
579 mergeBlockDefinitions(infoSink, block, unitBlock, &unit);
580 }
581 }
582 if (add) {
583 // push back on original list; won't change the size of the list we're iterating over
584 linkerObjects.push_back(*itUnitBlock);
585 }
586 }
587 }
588
mergeBlockDefinitions(TInfoSink & infoSink,TIntermSymbol * block,TIntermSymbol * unitBlock,TIntermediate * unit)589 void TIntermediate::mergeBlockDefinitions(TInfoSink& infoSink, TIntermSymbol* block, TIntermSymbol* unitBlock, TIntermediate* unit) {
590
591 if (block->getType().getTypeName() != unitBlock->getType().getTypeName() ||
592 block->getType().getBasicType() != unitBlock->getType().getBasicType() ||
593 block->getQualifier().storage != unitBlock->getQualifier().storage ||
594 block->getQualifier().layoutSet != unitBlock->getQualifier().layoutSet) {
595 // different block names likely means different blocks
596 return;
597 }
598
599 // merge the struct
600 // order of declarations doesn't matter and they matched based on member name
601 TTypeList* memberList = block->getType().getWritableStruct();
602 TTypeList* unitMemberList = unitBlock->getType().getWritableStruct();
603
604 // keep track of which members have changed position
605 // so we don't have to search the array again
606 std::map<unsigned int, unsigned int> memberIndexUpdates;
607
608 size_t memberListStartSize = memberList->size();
609 for (unsigned int i = 0; i < unitMemberList->size(); ++i) {
610 bool merge = true;
611 for (unsigned int j = 0; j < memberListStartSize; ++j) {
612 if ((*memberList)[j].type->getFieldName() == (*unitMemberList)[i].type->getFieldName()) {
613 merge = false;
614 const TType* memberType = (*memberList)[j].type;
615 const TType* unitMemberType = (*unitMemberList)[i].type;
616
617 // compare types
618 // don't need as many checks as when merging symbols, since
619 // initializers and most qualifiers are stripped when the member is moved into the block
620 if ((*memberType) != (*unitMemberType)) {
621 error(infoSink, "Types must match:");
622 infoSink.info << " " << memberType->getFieldName() << ": ";
623 infoSink.info << "\"" << memberType->getCompleteString() << "\" versus ";
624 infoSink.info << "\"" << unitMemberType->getCompleteString() << "\"\n";
625 }
626
627 memberIndexUpdates[i] = j;
628 }
629 }
630 if (merge) {
631 memberList->push_back((*unitMemberList)[i]);
632 memberIndexUpdates[i] = (unsigned int)memberList->size() - 1;
633 }
634 }
635
636 // update symbol node in unit tree,
637 // and other nodes that may reference it
638 class TMergeBlockTraverser : public TIntermTraverser {
639 public:
640 TMergeBlockTraverser(const TIntermSymbol* newSym)
641 : newSymbol(newSym), newType(nullptr), unit(nullptr), memberIndexUpdates(nullptr)
642 {
643 }
644 TMergeBlockTraverser(const TIntermSymbol* newSym, const glslang::TType* unitType, glslang::TIntermediate* unit,
645 const std::map<unsigned int, unsigned int>* memberIdxUpdates)
646 : TIntermTraverser(false, true), newSymbol(newSym), newType(unitType), unit(unit), memberIndexUpdates(memberIdxUpdates)
647 {
648 }
649 virtual ~TMergeBlockTraverser() {}
650
651 const TIntermSymbol* newSymbol;
652 const glslang::TType* newType; // shallow copy of the new type
653 glslang::TIntermediate* unit; // intermediate that is being updated
654 const std::map<unsigned int, unsigned int>* memberIndexUpdates;
655
656 virtual void visitSymbol(TIntermSymbol* symbol)
657 {
658 if (newSymbol->getAccessName() == symbol->getAccessName() &&
659 newSymbol->getQualifier().getBlockStorage() == symbol->getQualifier().getBlockStorage()) {
660 // Each symbol node may have a local copy of the block structure.
661 // Update those structures to match the new one post-merge
662 *(symbol->getWritableType().getWritableStruct()) = *(newSymbol->getType().getStruct());
663 }
664 }
665
666 virtual bool visitBinary(TVisit, glslang::TIntermBinary* node)
667 {
668 if (!unit || !newType || !memberIndexUpdates || memberIndexUpdates->empty())
669 return true;
670
671 if (node->getOp() == EOpIndexDirectStruct && node->getLeft()->getType() == *newType) {
672 // this is a dereference to a member of the block since the
673 // member list changed, need to update this to point to the
674 // right index
675 assert(node->getRight()->getAsConstantUnion());
676
677 glslang::TIntermConstantUnion* constNode = node->getRight()->getAsConstantUnion();
678 unsigned int memberIdx = constNode->getConstArray()[0].getUConst();
679 unsigned int newIdx = memberIndexUpdates->at(memberIdx);
680 TIntermTyped* newConstNode = unit->addConstantUnion(newIdx, node->getRight()->getLoc());
681
682 node->setRight(newConstNode);
683 delete constNode;
684
685 return true;
686 }
687 return true;
688 }
689 };
690
691 // 'this' may have symbols that are using the old block structure, so traverse the tree to update those
692 // in 'visitSymbol'
693 TMergeBlockTraverser finalLinkTraverser(block);
694 getTreeRoot()->traverse(&finalLinkTraverser);
695
696 // The 'unit' intermediate needs the block structures update, but also structure entry indices
697 // may have changed from the old block to the new one that it was merged into, so update those
698 // in 'visitBinary'
699 TType newType;
700 newType.shallowCopy(block->getType());
701 TMergeBlockTraverser unitFinalLinkTraverser(block, &newType, unit, &memberIndexUpdates);
702 unit->getTreeRoot()->traverse(&unitFinalLinkTraverser);
703
704 // update the member list
705 (*unitMemberList) = (*memberList);
706 }
707
708 //
709 // Merge the linker objects from unitLinkerObjects into linkerObjects.
710 // Duplication is expected and filtered out, but contradictions are an error.
711 //
mergeLinkerObjects(TInfoSink & infoSink,TIntermSequence & linkerObjects,const TIntermSequence & unitLinkerObjects,EShLanguage unitStage)712 void TIntermediate::mergeLinkerObjects(TInfoSink& infoSink, TIntermSequence& linkerObjects, const TIntermSequence& unitLinkerObjects, EShLanguage unitStage)
713 {
714 // Error check and merge the linker objects (duplicates should not be created)
715 std::size_t initialNumLinkerObjects = linkerObjects.size();
716 for (unsigned int unitLinkObj = 0; unitLinkObj < unitLinkerObjects.size(); ++unitLinkObj) {
717 bool merge = true;
718 for (std::size_t linkObj = 0; linkObj < initialNumLinkerObjects; ++linkObj) {
719 TIntermSymbol* symbol = linkerObjects[linkObj]->getAsSymbolNode();
720 TIntermSymbol* unitSymbol = unitLinkerObjects[unitLinkObj]->getAsSymbolNode();
721 assert(symbol && unitSymbol);
722
723 bool isSameSymbol = false;
724 // If they are both blocks in the same shader interface,
725 // match by the block-name, not the identifier name.
726 if (symbol->getType().getBasicType() == EbtBlock && unitSymbol->getType().getBasicType() == EbtBlock) {
727 if (isSameInterface(symbol, getStage(), unitSymbol, unitStage)) {
728 isSameSymbol = symbol->getType().getTypeName() == unitSymbol->getType().getTypeName();
729 }
730 }
731 else if (symbol->getName() == unitSymbol->getName())
732 isSameSymbol = true;
733
734 if (isSameSymbol) {
735 // filter out copy
736 merge = false;
737
738 // but if one has an initializer and the other does not, update
739 // the initializer
740 if (symbol->getConstArray().empty() && ! unitSymbol->getConstArray().empty())
741 symbol->setConstArray(unitSymbol->getConstArray());
742
743 // Similarly for binding
744 if (! symbol->getQualifier().hasBinding() && unitSymbol->getQualifier().hasBinding())
745 symbol->getQualifier().layoutBinding = unitSymbol->getQualifier().layoutBinding;
746
747 // Similarly for location
748 if (!symbol->getQualifier().hasLocation() && unitSymbol->getQualifier().hasLocation()) {
749 symbol->getQualifier().layoutLocation = unitSymbol->getQualifier().layoutLocation;
750 }
751
752 // Update implicit array sizes
753 mergeImplicitArraySizes(symbol->getWritableType(), unitSymbol->getType());
754
755 // Check for consistent types/qualification/initializers etc.
756 mergeErrorCheck(infoSink, *symbol, *unitSymbol, unitStage);
757 }
758 // If different symbols, verify they arn't push_constant since there can only be one per stage
759 else if (symbol->getQualifier().isPushConstant() && unitSymbol->getQualifier().isPushConstant() && getStage() == unitStage)
760 error(infoSink, "Only one push_constant block is allowed per stage");
761 }
762 if (merge) {
763 linkerObjects.push_back(unitLinkerObjects[unitLinkObj]);
764
765 // for anonymous blocks, check that their members don't conflict with other names
766 if (unitLinkerObjects[unitLinkObj]->getAsSymbolNode()->getBasicType() == EbtBlock &&
767 IsAnonymous(unitLinkerObjects[unitLinkObj]->getAsSymbolNode()->getName())) {
768 for (std::size_t linkObj = 0; linkObj < initialNumLinkerObjects; ++linkObj) {
769 TIntermSymbol* symbol = linkerObjects[linkObj]->getAsSymbolNode();
770 TIntermSymbol* unitSymbol = unitLinkerObjects[unitLinkObj]->getAsSymbolNode();
771 assert(symbol && unitSymbol);
772
773 auto checkName = [this, unitSymbol, &infoSink](const TString& name) {
774 for (unsigned int i = 0; i < unitSymbol->getType().getStruct()->size(); ++i) {
775 if (name == (*unitSymbol->getType().getStruct())[i].type->getFieldName()
776 && !((*unitSymbol->getType().getStruct())[i].type->getQualifier().hasLocation()
777 || unitSymbol->getType().getQualifier().hasLocation())
778 ) {
779 error(infoSink, "Anonymous member name used for global variable or other anonymous member: ");
780 infoSink.info << (*unitSymbol->getType().getStruct())[i].type->getCompleteString() << "\n";
781 }
782 }
783 };
784
785 if (isSameInterface(symbol, getStage(), unitSymbol, unitStage)) {
786 checkName(symbol->getName());
787
788 // check members of other anonymous blocks
789 if (symbol->getBasicType() == EbtBlock && IsAnonymous(symbol->getName())) {
790 for (unsigned int i = 0; i < symbol->getType().getStruct()->size(); ++i) {
791 checkName((*symbol->getType().getStruct())[i].type->getFieldName());
792 }
793 }
794 }
795 }
796 }
797 }
798 }
799 }
800
801 // TODO 4.5 link functionality: cull distance array size checking
802
803 // Recursively merge the implicit array sizes through the objects' respective type trees.
mergeImplicitArraySizes(TType & type,const TType & unitType)804 void TIntermediate::mergeImplicitArraySizes(TType& type, const TType& unitType)
805 {
806 if (type.isUnsizedArray()) {
807 if (unitType.isUnsizedArray()) {
808 type.updateImplicitArraySize(unitType.getImplicitArraySize());
809 if (unitType.isArrayVariablyIndexed())
810 type.setArrayVariablyIndexed();
811 } else if (unitType.isSizedArray())
812 type.changeOuterArraySize(unitType.getOuterArraySize());
813 }
814
815 // Type mismatches are caught and reported after this, just be careful for now.
816 if (! type.isStruct() || ! unitType.isStruct() || type.getStruct()->size() != unitType.getStruct()->size())
817 return;
818
819 for (int i = 0; i < (int)type.getStruct()->size(); ++i)
820 mergeImplicitArraySizes(*(*type.getStruct())[i].type, *(*unitType.getStruct())[i].type);
821 }
822
823 //
824 // Compare two global objects from two compilation units and see if they match
825 // well enough. Rules can be different for intra- vs. cross-stage matching.
826 //
827 // This function only does one of intra- or cross-stage matching per call.
828 //
mergeErrorCheck(TInfoSink & infoSink,const TIntermSymbol & symbol,const TIntermSymbol & unitSymbol,EShLanguage unitStage)829 void TIntermediate::mergeErrorCheck(TInfoSink& infoSink, const TIntermSymbol& symbol, const TIntermSymbol& unitSymbol, EShLanguage unitStage)
830 {
831 #if !defined(GLSLANG_WEB)
832 bool crossStage = getStage() != unitStage;
833 bool writeTypeComparison = false;
834 bool errorReported = false;
835 bool printQualifiers = false;
836 bool printPrecision = false;
837 bool printType = false;
838
839 // Types have to match
840 {
841 // but, we make an exception if one is an implicit array and the other is sized
842 // or if the array sizes differ because of the extra array dimension on some in/out boundaries
843 bool arraysMatch = false;
844 if (isIoResizeArray(symbol.getType(), getStage()) || isIoResizeArray(unitSymbol.getType(), unitStage)) {
845 // if the arrays have an extra dimension because of the stage.
846 // compare dimensions while ignoring the outer dimension
847 unsigned int firstDim = isIoResizeArray(symbol.getType(), getStage()) ? 1 : 0;
848 unsigned int numDim = symbol.getArraySizes()
849 ? symbol.getArraySizes()->getNumDims() : 0;
850 unsigned int unitFirstDim = isIoResizeArray(unitSymbol.getType(), unitStage) ? 1 : 0;
851 unsigned int unitNumDim = unitSymbol.getArraySizes()
852 ? unitSymbol.getArraySizes()->getNumDims() : 0;
853 arraysMatch = (numDim - firstDim) == (unitNumDim - unitFirstDim);
854 // check that array sizes match as well
855 for (unsigned int i = 0; i < (numDim - firstDim) && arraysMatch; i++) {
856 if (symbol.getArraySizes()->getDimSize(firstDim + i) !=
857 unitSymbol.getArraySizes()->getDimSize(unitFirstDim + i)) {
858 arraysMatch = false;
859 break;
860 }
861 }
862 }
863 else {
864 arraysMatch = symbol.getType().sameArrayness(unitSymbol.getType()) ||
865 (symbol.getType().isArray() && unitSymbol.getType().isArray() &&
866 (symbol.getType().isUnsizedArray() || unitSymbol.getType().isUnsizedArray()));
867 }
868
869 int lpidx = -1;
870 int rpidx = -1;
871 if (!symbol.getType().sameElementType(unitSymbol.getType(), &lpidx, &rpidx)) {
872 if (lpidx >= 0 && rpidx >= 0) {
873 error(infoSink, "Member names and types must match:", unitStage);
874 infoSink.info << " Block: " << symbol.getType().getTypeName() << "\n";
875 infoSink.info << " " << StageName(getStage()) << " stage: \""
876 << (*symbol.getType().getStruct())[lpidx].type->getCompleteString(true, false, false, true,
877 (*symbol.getType().getStruct())[lpidx].type->getFieldName()) << "\"\n";
878 infoSink.info << " " << StageName(unitStage) << " stage: \""
879 << (*unitSymbol.getType().getStruct())[rpidx].type->getCompleteString(true, false, false, true,
880 (*unitSymbol.getType().getStruct())[rpidx].type->getFieldName()) << "\"\n";
881 errorReported = true;
882 } else if (lpidx >= 0 && rpidx == -1) {
883 TString errmsg = StageName(getStage());
884 errmsg.append(" block member has no corresponding member in ").append(StageName(unitStage)).append(" block:");
885 error(infoSink, errmsg.c_str(), unitStage);
886 infoSink.info << " " << StageName(getStage()) << " stage: Block: " << symbol.getType().getTypeName() << ", Member: "
887 << (*symbol.getType().getStruct())[lpidx].type->getFieldName() << "\n";
888 infoSink.info << " " << StageName(unitStage) << " stage: Block: " << unitSymbol.getType().getTypeName() << ", Member: n/a \n";
889 errorReported = true;
890 } else if (lpidx == -1 && rpidx >= 0) {
891 TString errmsg = StageName(unitStage);
892 errmsg.append(" block member has no corresponding member in ").append(StageName(getStage())).append(" block:");
893 error(infoSink, errmsg.c_str(), unitStage);
894 infoSink.info << " " << StageName(unitStage) << " stage: Block: " << unitSymbol.getType().getTypeName() << ", Member: "
895 << (*unitSymbol.getType().getStruct())[rpidx].type->getFieldName() << "\n";
896 infoSink.info << " " << StageName(getStage()) << " stage: Block: " << symbol.getType().getTypeName() << ", Member: n/a \n";
897 errorReported = true;
898 } else {
899 error(infoSink, "Types must match:", unitStage);
900 writeTypeComparison = true;
901 printType = true;
902 }
903 } else if (!arraysMatch) {
904 error(infoSink, "Array sizes must be compatible:", unitStage);
905 writeTypeComparison = true;
906 printType = true;
907 } else if (!symbol.getType().sameTypeParameters(unitSymbol.getType())) {
908 error(infoSink, "Type parameters must match:", unitStage);
909 writeTypeComparison = true;
910 printType = true;
911 }
912 }
913
914 // Interface block member-wise layout qualifiers have to match
915 if (symbol.getType().getBasicType() == EbtBlock && unitSymbol.getType().getBasicType() == EbtBlock &&
916 symbol.getType().getStruct() && unitSymbol.getType().getStruct() &&
917 symbol.getType().sameStructType(unitSymbol.getType())) {
918 unsigned int li = 0;
919 unsigned int ri = 0;
920 while (li < symbol.getType().getStruct()->size() && ri < unitSymbol.getType().getStruct()->size()) {
921 if ((*symbol.getType().getStruct())[li].type->hiddenMember()) {
922 ++li;
923 continue;
924 }
925 if ((*unitSymbol.getType().getStruct())[ri].type->hiddenMember()) {
926 ++ri;
927 continue;
928 }
929 const TQualifier& qualifier = (*symbol.getType().getStruct())[li].type->getQualifier();
930 const TQualifier & unitQualifier = (*unitSymbol.getType().getStruct())[ri].type->getQualifier();
931 bool layoutQualifierError = false;
932 if (qualifier.layoutMatrix != unitQualifier.layoutMatrix) {
933 error(infoSink, "Interface block member layout matrix qualifier must match:", unitStage);
934 layoutQualifierError = true;
935 }
936 if (qualifier.layoutOffset != unitQualifier.layoutOffset) {
937 error(infoSink, "Interface block member layout offset qualifier must match:", unitStage);
938 layoutQualifierError = true;
939 }
940 if (qualifier.layoutAlign != unitQualifier.layoutAlign) {
941 error(infoSink, "Interface block member layout align qualifier must match:", unitStage);
942 layoutQualifierError = true;
943 }
944 if (qualifier.layoutLocation != unitQualifier.layoutLocation) {
945 error(infoSink, "Interface block member layout location qualifier must match:", unitStage);
946 layoutQualifierError = true;
947 }
948 if (qualifier.layoutComponent != unitQualifier.layoutComponent) {
949 error(infoSink, "Interface block member layout component qualifier must match:", unitStage);
950 layoutQualifierError = true;
951 }
952 if (layoutQualifierError) {
953 infoSink.info << " " << StageName(getStage()) << " stage: Block: " << symbol.getType().getTypeName() << ", Member: "
954 << (*symbol.getType().getStruct())[li].type->getFieldName() << " \""
955 << (*symbol.getType().getStruct())[li].type->getCompleteString(true, true, false, false) << "\"\n";
956 infoSink.info << " " << StageName(unitStage) << " stage: Block: " << unitSymbol.getType().getTypeName() << ", Member: "
957 << (*unitSymbol.getType().getStruct())[ri].type->getFieldName() << " \""
958 << (*unitSymbol.getType().getStruct())[ri].type->getCompleteString(true, true, false, false) << "\"\n";
959 errorReported = true;
960 }
961 ++li;
962 ++ri;
963 }
964 }
965
966 bool isInOut = crossStage &&
967 ((symbol.getQualifier().storage == EvqVaryingIn && unitSymbol.getQualifier().storage == EvqVaryingOut) ||
968 (symbol.getQualifier().storage == EvqVaryingOut && unitSymbol.getQualifier().storage == EvqVaryingIn));
969
970 // Qualifiers have to (almost) match
971 // Storage...
972 if (!isInOut && symbol.getQualifier().storage != unitSymbol.getQualifier().storage) {
973 error(infoSink, "Storage qualifiers must match:", unitStage);
974 writeTypeComparison = true;
975 printQualifiers = true;
976 }
977
978 // Uniform and buffer blocks must either both have an instance name, or
979 // must both be anonymous. The names don't need to match though.
980 if (symbol.getQualifier().isUniformOrBuffer() &&
981 (IsAnonymous(symbol.getName()) != IsAnonymous(unitSymbol.getName()))) {
982 error(infoSink, "Matched Uniform or Storage blocks must all be anonymous,"
983 " or all be named:", unitStage);
984 writeTypeComparison = true;
985 }
986
987 if (symbol.getQualifier().storage == unitSymbol.getQualifier().storage &&
988 (IsAnonymous(symbol.getName()) != IsAnonymous(unitSymbol.getName()) ||
989 (!IsAnonymous(symbol.getName()) && symbol.getName() != unitSymbol.getName()))) {
990 warn(infoSink, "Matched shader interfaces are using different instance names.", unitStage);
991 writeTypeComparison = true;
992 }
993
994 // Precision...
995 if (!isInOut && symbol.getQualifier().precision != unitSymbol.getQualifier().precision) {
996 error(infoSink, "Precision qualifiers must match:", unitStage);
997 writeTypeComparison = true;
998 printPrecision = true;
999 }
1000
1001 // Invariance...
1002 if (! crossStage && symbol.getQualifier().invariant != unitSymbol.getQualifier().invariant) {
1003 error(infoSink, "Presence of invariant qualifier must match:", unitStage);
1004 writeTypeComparison = true;
1005 printQualifiers = true;
1006 }
1007
1008 // Precise...
1009 if (! crossStage && symbol.getQualifier().isNoContraction() != unitSymbol.getQualifier().isNoContraction()) {
1010 error(infoSink, "Presence of precise qualifier must match:", unitStage);
1011 writeTypeComparison = true;
1012 printPrecision = true;
1013 }
1014
1015 // Auxiliary and interpolation...
1016 // "interpolation qualification (e.g., flat) and auxiliary qualification (e.g. centroid) may differ.
1017 // These mismatches are allowed between any pair of stages ...
1018 // those provided in the fragment shader supersede those provided in previous stages."
1019 if (!crossStage &&
1020 (symbol.getQualifier().centroid != unitSymbol.getQualifier().centroid ||
1021 symbol.getQualifier().smooth != unitSymbol.getQualifier().smooth ||
1022 symbol.getQualifier().flat != unitSymbol.getQualifier().flat ||
1023 symbol.getQualifier().isSample()!= unitSymbol.getQualifier().isSample() ||
1024 symbol.getQualifier().isPatch() != unitSymbol.getQualifier().isPatch() ||
1025 symbol.getQualifier().isNonPerspective() != unitSymbol.getQualifier().isNonPerspective())) {
1026 error(infoSink, "Interpolation and auxiliary storage qualifiers must match:", unitStage);
1027 writeTypeComparison = true;
1028 printQualifiers = true;
1029 }
1030
1031 // Memory...
1032 bool memoryQualifierError = false;
1033 if (symbol.getQualifier().coherent != unitSymbol.getQualifier().coherent) {
1034 error(infoSink, "Memory coherent qualifier must match:", unitStage);
1035 memoryQualifierError = true;
1036 }
1037 if (symbol.getQualifier().devicecoherent != unitSymbol.getQualifier().devicecoherent) {
1038 error(infoSink, "Memory devicecoherent qualifier must match:", unitStage);
1039 memoryQualifierError = true;
1040 }
1041 if (symbol.getQualifier().queuefamilycoherent != unitSymbol.getQualifier().queuefamilycoherent) {
1042 error(infoSink, "Memory queuefamilycoherent qualifier must match:", unitStage);
1043 memoryQualifierError = true;
1044 }
1045 if (symbol.getQualifier().workgroupcoherent != unitSymbol.getQualifier().workgroupcoherent) {
1046 error(infoSink, "Memory workgroupcoherent qualifier must match:", unitStage);
1047 memoryQualifierError = true;
1048 }
1049 if (symbol.getQualifier().subgroupcoherent != unitSymbol.getQualifier().subgroupcoherent) {
1050 error(infoSink, "Memory subgroupcoherent qualifier must match:", unitStage);
1051 memoryQualifierError = true;
1052 }
1053 if (symbol.getQualifier().shadercallcoherent != unitSymbol.getQualifier().shadercallcoherent) {
1054 error(infoSink, "Memory shadercallcoherent qualifier must match:", unitStage);
1055 memoryQualifierError = true;
1056 }
1057 if (symbol.getQualifier().nonprivate != unitSymbol.getQualifier().nonprivate) {
1058 error(infoSink, "Memory nonprivate qualifier must match:", unitStage);
1059 memoryQualifierError = true;
1060 }
1061 if (symbol.getQualifier().volatil != unitSymbol.getQualifier().volatil) {
1062 error(infoSink, "Memory volatil qualifier must match:", unitStage);
1063 memoryQualifierError = true;
1064 }
1065 if (symbol.getQualifier().restrict != unitSymbol.getQualifier().restrict) {
1066 error(infoSink, "Memory restrict qualifier must match:", unitStage);
1067 memoryQualifierError = true;
1068 }
1069 if (symbol.getQualifier().readonly != unitSymbol.getQualifier().readonly) {
1070 error(infoSink, "Memory readonly qualifier must match:", unitStage);
1071 memoryQualifierError = true;
1072 }
1073 if (symbol.getQualifier().writeonly != unitSymbol.getQualifier().writeonly) {
1074 error(infoSink, "Memory writeonly qualifier must match:", unitStage);
1075 memoryQualifierError = true;
1076 }
1077 if (memoryQualifierError) {
1078 writeTypeComparison = true;
1079 printQualifiers = true;
1080 }
1081
1082 // Layouts...
1083 // TODO: 4.4 enhanced layouts: Generalize to include offset/align: current spec
1084 // requires separate user-supplied offset from actual computed offset, but
1085 // current implementation only has one offset.
1086 bool layoutQualifierError = false;
1087 if (symbol.getQualifier().layoutMatrix != unitSymbol.getQualifier().layoutMatrix) {
1088 error(infoSink, "Layout matrix qualifier must match:", unitStage);
1089 layoutQualifierError = true;
1090 }
1091 if (symbol.getQualifier().layoutPacking != unitSymbol.getQualifier().layoutPacking) {
1092 error(infoSink, "Layout packing qualifier must match:", unitStage);
1093 layoutQualifierError = true;
1094 }
1095 if (symbol.getQualifier().hasLocation() && unitSymbol.getQualifier().hasLocation() && symbol.getQualifier().layoutLocation != unitSymbol.getQualifier().layoutLocation) {
1096 error(infoSink, "Layout location qualifier must match:", unitStage);
1097 layoutQualifierError = true;
1098 }
1099 if (symbol.getQualifier().layoutComponent != unitSymbol.getQualifier().layoutComponent) {
1100 error(infoSink, "Layout component qualifier must match:", unitStage);
1101 layoutQualifierError = true;
1102 }
1103 if (symbol.getQualifier().layoutIndex != unitSymbol.getQualifier().layoutIndex) {
1104 error(infoSink, "Layout index qualifier must match:", unitStage);
1105 layoutQualifierError = true;
1106 }
1107 if (symbol.getQualifier().hasBinding() && unitSymbol.getQualifier().hasBinding() && symbol.getQualifier().layoutBinding != unitSymbol.getQualifier().layoutBinding) {
1108 error(infoSink, "Layout binding qualifier must match:", unitStage);
1109 layoutQualifierError = true;
1110 }
1111 if (symbol.getQualifier().hasBinding() && (symbol.getQualifier().layoutOffset != unitSymbol.getQualifier().layoutOffset)) {
1112 error(infoSink, "Layout offset qualifier must match:", unitStage);
1113 layoutQualifierError = true;
1114 }
1115 if (layoutQualifierError) {
1116 writeTypeComparison = true;
1117 printQualifiers = true;
1118 }
1119
1120 // Initializers have to match, if both are present, and if we don't already know the types don't match
1121 if (! writeTypeComparison && ! errorReported) {
1122 if (! symbol.getConstArray().empty() && ! unitSymbol.getConstArray().empty()) {
1123 if (symbol.getConstArray() != unitSymbol.getConstArray()) {
1124 error(infoSink, "Initializers must match:", unitStage);
1125 infoSink.info << " " << symbol.getName() << "\n";
1126 }
1127 }
1128 }
1129
1130 if (writeTypeComparison) {
1131 if (symbol.getType().getBasicType() == EbtBlock && unitSymbol.getType().getBasicType() == EbtBlock &&
1132 symbol.getType().getStruct() && unitSymbol.getType().getStruct()) {
1133 if (printType) {
1134 infoSink.info << " " << StageName(getStage()) << " stage: \"" << symbol.getType().getCompleteString(true, printQualifiers, printPrecision,
1135 printType, symbol.getName(), symbol.getType().getTypeName()) << "\"\n";
1136 infoSink.info << " " << StageName(unitStage) << " stage: \"" << unitSymbol.getType().getCompleteString(true, printQualifiers, printPrecision,
1137 printType, unitSymbol.getName(), unitSymbol.getType().getTypeName()) << "\"\n";
1138 } else {
1139 infoSink.info << " " << StageName(getStage()) << " stage: Block: " << symbol.getType().getTypeName() << " Instance: " << symbol.getName()
1140 << ": \"" << symbol.getType().getCompleteString(true, printQualifiers, printPrecision, printType) << "\"\n";
1141 infoSink.info << " " << StageName(unitStage) << " stage: Block: " << unitSymbol.getType().getTypeName() << " Instance: " << unitSymbol.getName()
1142 << ": \"" << unitSymbol.getType().getCompleteString(true, printQualifiers, printPrecision, printType) << "\"\n";
1143 }
1144 } else {
1145 if (printType) {
1146 infoSink.info << " " << StageName(getStage()) << " stage: \""
1147 << symbol.getType().getCompleteString(true, printQualifiers, printPrecision, printType, symbol.getName()) << "\"\n";
1148 infoSink.info << " " << StageName(unitStage) << " stage: \""
1149 << unitSymbol.getType().getCompleteString(true, printQualifiers, printPrecision, printType, unitSymbol.getName()) << "\"\n";
1150 } else {
1151 infoSink.info << " " << StageName(getStage()) << " stage: " << symbol.getName() << " \""
1152 << symbol.getType().getCompleteString(true, printQualifiers, printPrecision, printType) << "\"\n";
1153 infoSink.info << " " << StageName(unitStage) << " stage: " << unitSymbol.getName() << " \""
1154 << unitSymbol.getType().getCompleteString(true, printQualifiers, printPrecision, printType) << "\"\n";
1155 }
1156 }
1157 }
1158 #endif
1159 }
1160
sharedBlockCheck(TInfoSink & infoSink)1161 void TIntermediate::sharedBlockCheck(TInfoSink& infoSink)
1162 {
1163 bool has_shared_block = false;
1164 bool has_shared_non_block = false;
1165 TIntermSequence& linkObjects = findLinkerObjects()->getSequence();
1166 for (size_t i = 0; i < linkObjects.size(); ++i) {
1167 const TType& type = linkObjects[i]->getAsTyped()->getType();
1168 const TQualifier& qualifier = type.getQualifier();
1169 if (qualifier.storage == glslang::EvqShared) {
1170 if (type.getBasicType() == glslang::EbtBlock)
1171 has_shared_block = true;
1172 else
1173 has_shared_non_block = true;
1174 }
1175 }
1176 if (has_shared_block && has_shared_non_block)
1177 error(infoSink, "cannot mix use of shared variables inside and outside blocks");
1178 }
1179
1180 //
1181 // Do final link-time error checking of a complete (merged) intermediate representation.
1182 // (Much error checking was done during merging).
1183 //
1184 // Also, lock in defaults of things not set, including array sizes.
1185 //
finalCheck(TInfoSink & infoSink,bool keepUncalled)1186 void TIntermediate::finalCheck(TInfoSink& infoSink, bool keepUncalled)
1187 {
1188 if (getTreeRoot() == nullptr)
1189 return;
1190
1191 if (numEntryPoints < 1) {
1192 if (getSource() == EShSourceGlsl)
1193 error(infoSink, "Missing entry point: Each stage requires one entry point");
1194 else
1195 warn(infoSink, "Entry point not found");
1196 }
1197
1198 // recursion and missing body checking
1199 checkCallGraphCycles(infoSink);
1200 checkCallGraphBodies(infoSink, keepUncalled);
1201
1202 // overlap/alias/missing I/O, etc.
1203 inOutLocationCheck(infoSink);
1204
1205 #ifndef GLSLANG_WEB
1206 if (getNumPushConstants() > 1)
1207 error(infoSink, "Only one push_constant block is allowed per stage");
1208
1209 // invocations
1210 if (invocations == TQualifier::layoutNotSet)
1211 invocations = 1;
1212
1213 if (inIoAccessed("gl_ClipDistance") && inIoAccessed("gl_ClipVertex"))
1214 error(infoSink, "Can only use one of gl_ClipDistance or gl_ClipVertex (gl_ClipDistance is preferred)");
1215 if (inIoAccessed("gl_CullDistance") && inIoAccessed("gl_ClipVertex"))
1216 error(infoSink, "Can only use one of gl_CullDistance or gl_ClipVertex (gl_ClipDistance is preferred)");
1217
1218 if (userOutputUsed() && (inIoAccessed("gl_FragColor") || inIoAccessed("gl_FragData")))
1219 error(infoSink, "Cannot use gl_FragColor or gl_FragData when using user-defined outputs");
1220 if (inIoAccessed("gl_FragColor") && inIoAccessed("gl_FragData"))
1221 error(infoSink, "Cannot use both gl_FragColor and gl_FragData");
1222
1223 for (size_t b = 0; b < xfbBuffers.size(); ++b) {
1224 if (xfbBuffers[b].contains64BitType)
1225 RoundToPow2(xfbBuffers[b].implicitStride, 8);
1226 else if (xfbBuffers[b].contains32BitType)
1227 RoundToPow2(xfbBuffers[b].implicitStride, 4);
1228 else if (xfbBuffers[b].contains16BitType)
1229 RoundToPow2(xfbBuffers[b].implicitStride, 2);
1230
1231 // "It is a compile-time or link-time error to have
1232 // any xfb_offset that overflows xfb_stride, whether stated on declarations before or after the xfb_stride, or
1233 // in different compilation units. While xfb_stride can be declared multiple times for the same buffer, it is a
1234 // compile-time or link-time error to have different values specified for the stride for the same buffer."
1235 if (xfbBuffers[b].stride != TQualifier::layoutXfbStrideEnd && xfbBuffers[b].implicitStride > xfbBuffers[b].stride) {
1236 error(infoSink, "xfb_stride is too small to hold all buffer entries:");
1237 infoSink.info.prefix(EPrefixError);
1238 infoSink.info << " xfb_buffer " << (unsigned int)b << ", xfb_stride " << xfbBuffers[b].stride << ", minimum stride needed: " << xfbBuffers[b].implicitStride << "\n";
1239 }
1240 if (xfbBuffers[b].stride == TQualifier::layoutXfbStrideEnd)
1241 xfbBuffers[b].stride = xfbBuffers[b].implicitStride;
1242
1243 // "If the buffer is capturing any
1244 // outputs with double-precision or 64-bit integer components, the stride must be a multiple of 8, otherwise it must be a
1245 // multiple of 4, or a compile-time or link-time error results."
1246 if (xfbBuffers[b].contains64BitType && ! IsMultipleOfPow2(xfbBuffers[b].stride, 8)) {
1247 error(infoSink, "xfb_stride must be multiple of 8 for buffer holding a double or 64-bit integer:");
1248 infoSink.info.prefix(EPrefixError);
1249 infoSink.info << " xfb_buffer " << (unsigned int)b << ", xfb_stride " << xfbBuffers[b].stride << "\n";
1250 } else if (xfbBuffers[b].contains32BitType && ! IsMultipleOfPow2(xfbBuffers[b].stride, 4)) {
1251 error(infoSink, "xfb_stride must be multiple of 4:");
1252 infoSink.info.prefix(EPrefixError);
1253 infoSink.info << " xfb_buffer " << (unsigned int)b << ", xfb_stride " << xfbBuffers[b].stride << "\n";
1254 }
1255 // "If the buffer is capturing any
1256 // outputs with half-precision or 16-bit integer components, the stride must be a multiple of 2"
1257 else if (xfbBuffers[b].contains16BitType && ! IsMultipleOfPow2(xfbBuffers[b].stride, 2)) {
1258 error(infoSink, "xfb_stride must be multiple of 2 for buffer holding a half float or 16-bit integer:");
1259 infoSink.info.prefix(EPrefixError);
1260 infoSink.info << " xfb_buffer " << (unsigned int)b << ", xfb_stride " << xfbBuffers[b].stride << "\n";
1261 }
1262
1263 // "The resulting stride (implicit or explicit), when divided by 4, must be less than or equal to the
1264 // implementation-dependent constant gl_MaxTransformFeedbackInterleavedComponents."
1265 if (xfbBuffers[b].stride > (unsigned int)(4 * resources->maxTransformFeedbackInterleavedComponents)) {
1266 error(infoSink, "xfb_stride is too large:");
1267 infoSink.info.prefix(EPrefixError);
1268 infoSink.info << " xfb_buffer " << (unsigned int)b << ", components (1/4 stride) needed are " << xfbBuffers[b].stride/4 << ", gl_MaxTransformFeedbackInterleavedComponents is " << resources->maxTransformFeedbackInterleavedComponents << "\n";
1269 }
1270 }
1271
1272 switch (language) {
1273 case EShLangVertex:
1274 break;
1275 case EShLangTessControl:
1276 if (vertices == TQualifier::layoutNotSet)
1277 error(infoSink, "At least one shader must specify an output layout(vertices=...)");
1278 break;
1279 case EShLangTessEvaluation:
1280 if (getSource() == EShSourceGlsl) {
1281 if (inputPrimitive == ElgNone)
1282 error(infoSink, "At least one shader must specify an input layout primitive");
1283 if (vertexSpacing == EvsNone)
1284 vertexSpacing = EvsEqual;
1285 if (vertexOrder == EvoNone)
1286 vertexOrder = EvoCcw;
1287 }
1288 break;
1289 case EShLangGeometry:
1290 if (inputPrimitive == ElgNone)
1291 error(infoSink, "At least one shader must specify an input layout primitive");
1292 if (outputPrimitive == ElgNone)
1293 error(infoSink, "At least one shader must specify an output layout primitive");
1294 if (vertices == TQualifier::layoutNotSet)
1295 error(infoSink, "At least one shader must specify a layout(max_vertices = value)");
1296 break;
1297 case EShLangFragment:
1298 // for GL_ARB_post_depth_coverage, EarlyFragmentTest is set automatically in
1299 // ParseHelper.cpp. So if we reach here, this must be GL_EXT_post_depth_coverage
1300 // requiring explicit early_fragment_tests
1301 if (getPostDepthCoverage() && !getEarlyFragmentTests())
1302 error(infoSink, "post_depth_coverage requires early_fragment_tests");
1303 break;
1304 case EShLangCompute:
1305 sharedBlockCheck(infoSink);
1306 break;
1307 case EShLangRayGen:
1308 case EShLangIntersect:
1309 case EShLangAnyHit:
1310 case EShLangClosestHit:
1311 case EShLangMiss:
1312 case EShLangCallable:
1313 if (numShaderRecordBlocks > 1)
1314 error(infoSink, "Only one shaderRecordNV buffer block is allowed per stage");
1315 break;
1316 case EShLangMesh:
1317 // NV_mesh_shader doesn't allow use of both single-view and per-view builtins.
1318 if (inIoAccessed("gl_Position") && inIoAccessed("gl_PositionPerViewNV"))
1319 error(infoSink, "Can only use one of gl_Position or gl_PositionPerViewNV");
1320 if (inIoAccessed("gl_ClipDistance") && inIoAccessed("gl_ClipDistancePerViewNV"))
1321 error(infoSink, "Can only use one of gl_ClipDistance or gl_ClipDistancePerViewNV");
1322 if (inIoAccessed("gl_CullDistance") && inIoAccessed("gl_CullDistancePerViewNV"))
1323 error(infoSink, "Can only use one of gl_CullDistance or gl_CullDistancePerViewNV");
1324 if (inIoAccessed("gl_Layer") && inIoAccessed("gl_LayerPerViewNV"))
1325 error(infoSink, "Can only use one of gl_Layer or gl_LayerPerViewNV");
1326 if (inIoAccessed("gl_ViewportMask") && inIoAccessed("gl_ViewportMaskPerViewNV"))
1327 error(infoSink, "Can only use one of gl_ViewportMask or gl_ViewportMaskPerViewNV");
1328 if (outputPrimitive == ElgNone)
1329 error(infoSink, "At least one shader must specify an output layout primitive");
1330 if (vertices == TQualifier::layoutNotSet)
1331 error(infoSink, "At least one shader must specify a layout(max_vertices = value)");
1332 if (primitives == TQualifier::layoutNotSet)
1333 error(infoSink, "At least one shader must specify a layout(max_primitives = value)");
1334 // fall through
1335 case EShLangTask:
1336 if (numTaskNVBlocks > 1)
1337 error(infoSink, "Only one taskNV interface block is allowed per shader");
1338 if (numTaskEXTPayloads > 1)
1339 error(infoSink, "Only single variable of type taskPayloadSharedEXT is allowed per shader");
1340 sharedBlockCheck(infoSink);
1341 break;
1342 default:
1343 error(infoSink, "Unknown Stage.");
1344 break;
1345 }
1346
1347 // Process the tree for any node-specific work.
1348 class TFinalLinkTraverser : public TIntermTraverser {
1349 public:
1350 TFinalLinkTraverser() { }
1351 virtual ~TFinalLinkTraverser() { }
1352
1353 virtual void visitSymbol(TIntermSymbol* symbol)
1354 {
1355 // Implicitly size arrays.
1356 // If an unsized array is left as unsized, it effectively
1357 // becomes run-time sized.
1358 symbol->getWritableType().adoptImplicitArraySizes(false);
1359 }
1360 } finalLinkTraverser;
1361
1362 treeRoot->traverse(&finalLinkTraverser);
1363 #endif
1364 }
1365
1366 //
1367 // See if the call graph contains any static recursion, which is disallowed
1368 // by the specification.
1369 //
checkCallGraphCycles(TInfoSink & infoSink)1370 void TIntermediate::checkCallGraphCycles(TInfoSink& infoSink)
1371 {
1372 // Clear fields we'll use for this.
1373 for (TGraph::iterator call = callGraph.begin(); call != callGraph.end(); ++call) {
1374 call->visited = false;
1375 call->currentPath = false;
1376 call->errorGiven = false;
1377 }
1378
1379 //
1380 // Loop, looking for a new connected subgraph. One subgraph is handled per loop iteration.
1381 //
1382
1383 TCall* newRoot;
1384 do {
1385 // See if we have unvisited parts of the graph.
1386 newRoot = nullptr;
1387 for (TGraph::iterator call = callGraph.begin(); call != callGraph.end(); ++call) {
1388 if (! call->visited) {
1389 newRoot = &(*call);
1390 break;
1391 }
1392 }
1393
1394 // If not, we are done.
1395 if (! newRoot)
1396 break;
1397
1398 // Otherwise, we found a new subgraph, process it:
1399 // See what all can be reached by this new root, and if any of
1400 // that is recursive. This is done by depth-first traversals, seeing
1401 // if a new call is found that was already in the currentPath (a back edge),
1402 // thereby detecting recursion.
1403 std::list<TCall*> stack;
1404 newRoot->currentPath = true; // currentPath will be true iff it is on the stack
1405 stack.push_back(newRoot);
1406 while (! stack.empty()) {
1407 // get a caller
1408 TCall* call = stack.back();
1409
1410 // Add to the stack just one callee.
1411 // This algorithm always terminates, because only !visited and !currentPath causes a push
1412 // and all pushes change currentPath to true, and all pops change visited to true.
1413 TGraph::iterator child = callGraph.begin();
1414 for (; child != callGraph.end(); ++child) {
1415
1416 // If we already visited this node, its whole subgraph has already been processed, so skip it.
1417 if (child->visited)
1418 continue;
1419
1420 if (call->callee == child->caller) {
1421 if (child->currentPath) {
1422 // Then, we found a back edge
1423 if (! child->errorGiven) {
1424 error(infoSink, "Recursion detected:");
1425 infoSink.info << " " << call->callee << " calling " << child->callee << "\n";
1426 child->errorGiven = true;
1427 recursive = true;
1428 }
1429 } else {
1430 child->currentPath = true;
1431 stack.push_back(&(*child));
1432 break;
1433 }
1434 }
1435 }
1436 if (child == callGraph.end()) {
1437 // no more callees, we bottomed out, never look at this node again
1438 stack.back()->currentPath = false;
1439 stack.back()->visited = true;
1440 stack.pop_back();
1441 }
1442 } // end while, meaning nothing left to process in this subtree
1443
1444 } while (newRoot); // redundant loop check; should always exit via the 'break' above
1445 }
1446
1447 //
1448 // See which functions are reachable from the entry point and which have bodies.
1449 // Reachable ones with missing bodies are errors.
1450 // Unreachable bodies are dead code.
1451 //
checkCallGraphBodies(TInfoSink & infoSink,bool keepUncalled)1452 void TIntermediate::checkCallGraphBodies(TInfoSink& infoSink, bool keepUncalled)
1453 {
1454 // Clear fields we'll use for this.
1455 for (TGraph::iterator call = callGraph.begin(); call != callGraph.end(); ++call) {
1456 call->visited = false;
1457 call->calleeBodyPosition = -1;
1458 }
1459
1460 // The top level of the AST includes function definitions (bodies).
1461 // Compare these to function calls in the call graph.
1462 // We'll end up knowing which have bodies, and if so,
1463 // how to map the call-graph node to the location in the AST.
1464 TIntermSequence &functionSequence = getTreeRoot()->getAsAggregate()->getSequence();
1465 std::vector<bool> reachable(functionSequence.size(), true); // so that non-functions are reachable
1466 for (int f = 0; f < (int)functionSequence.size(); ++f) {
1467 glslang::TIntermAggregate* node = functionSequence[f]->getAsAggregate();
1468 if (node && (node->getOp() == glslang::EOpFunction)) {
1469 if (node->getName().compare(getEntryPointMangledName().c_str()) != 0)
1470 reachable[f] = false; // so that function bodies are unreachable, until proven otherwise
1471 for (TGraph::iterator call = callGraph.begin(); call != callGraph.end(); ++call) {
1472 if (call->callee == node->getName())
1473 call->calleeBodyPosition = f;
1474 }
1475 }
1476 }
1477
1478 // Start call-graph traversal by visiting the entry point nodes.
1479 for (TGraph::iterator call = callGraph.begin(); call != callGraph.end(); ++call) {
1480 if (call->caller.compare(getEntryPointMangledName().c_str()) == 0)
1481 call->visited = true;
1482 }
1483
1484 // Propagate 'visited' through the call-graph to every part of the graph it
1485 // can reach (seeded with the entry-point setting above).
1486 bool changed;
1487 do {
1488 changed = false;
1489 for (auto call1 = callGraph.begin(); call1 != callGraph.end(); ++call1) {
1490 if (call1->visited) {
1491 for (TGraph::iterator call2 = callGraph.begin(); call2 != callGraph.end(); ++call2) {
1492 if (! call2->visited) {
1493 if (call1->callee == call2->caller) {
1494 changed = true;
1495 call2->visited = true;
1496 }
1497 }
1498 }
1499 }
1500 }
1501 } while (changed);
1502
1503 // Any call-graph node set to visited but without a callee body is an error.
1504 for (TGraph::iterator call = callGraph.begin(); call != callGraph.end(); ++call) {
1505 if (call->visited) {
1506 if (call->calleeBodyPosition == -1) {
1507 error(infoSink, "No function definition (body) found: ");
1508 infoSink.info << " " << call->callee << "\n";
1509 } else
1510 reachable[call->calleeBodyPosition] = true;
1511 }
1512 }
1513
1514 // Bodies in the AST not reached by the call graph are dead;
1515 // clear them out, since they can't be reached and also can't
1516 // be translated further due to possibility of being ill defined.
1517 if (! keepUncalled) {
1518 for (int f = 0; f < (int)functionSequence.size(); ++f) {
1519 if (! reachable[f])
1520 functionSequence[f] = nullptr;
1521 }
1522 functionSequence.erase(std::remove(functionSequence.begin(), functionSequence.end(), nullptr), functionSequence.end());
1523 }
1524 }
1525
1526 //
1527 // Satisfy rules for location qualifiers on inputs and outputs
1528 //
inOutLocationCheck(TInfoSink & infoSink)1529 void TIntermediate::inOutLocationCheck(TInfoSink& infoSink)
1530 {
1531 // ES 3.0 requires all outputs to have location qualifiers if there is more than one output
1532 bool fragOutWithNoLocation = false;
1533 int numFragOut = 0;
1534
1535 // TODO: linker functionality: location collision checking
1536
1537 TIntermSequence& linkObjects = findLinkerObjects()->getSequence();
1538 for (size_t i = 0; i < linkObjects.size(); ++i) {
1539 const TType& type = linkObjects[i]->getAsTyped()->getType();
1540 const TQualifier& qualifier = type.getQualifier();
1541 if (language == EShLangFragment) {
1542 if (qualifier.storage == EvqVaryingOut && qualifier.builtIn == EbvNone) {
1543 ++numFragOut;
1544 if (!qualifier.hasAnyLocation())
1545 fragOutWithNoLocation = true;
1546 }
1547 }
1548 }
1549
1550 if (isEsProfile()) {
1551 if (numFragOut > 1 && fragOutWithNoLocation)
1552 error(infoSink, "when more than one fragment shader output, all must have location qualifiers");
1553 }
1554 }
1555
findLinkerObjects() const1556 TIntermAggregate* TIntermediate::findLinkerObjects() const
1557 {
1558 // Get the top-level globals
1559 TIntermSequence& globals = treeRoot->getAsAggregate()->getSequence();
1560
1561 // Get the last member of the sequences, expected to be the linker-object lists
1562 assert(globals.back()->getAsAggregate()->getOp() == EOpLinkerObjects);
1563
1564 return globals.back()->getAsAggregate();
1565 }
1566
1567 // See if a variable was both a user-declared output and used.
1568 // Note: the spec discusses writing to one, but this looks at read or write, which
1569 // is more useful, and perhaps the spec should be changed to reflect that.
userOutputUsed() const1570 bool TIntermediate::userOutputUsed() const
1571 {
1572 const TIntermSequence& linkerObjects = findLinkerObjects()->getSequence();
1573
1574 bool found = false;
1575 for (size_t i = 0; i < linkerObjects.size(); ++i) {
1576 const TIntermSymbol& symbolNode = *linkerObjects[i]->getAsSymbolNode();
1577 if (symbolNode.getQualifier().storage == EvqVaryingOut &&
1578 symbolNode.getName().compare(0, 3, "gl_") != 0 &&
1579 inIoAccessed(symbolNode.getName())) {
1580 found = true;
1581 break;
1582 }
1583 }
1584
1585 return found;
1586 }
1587
1588 // Accumulate locations used for inputs, outputs, and uniforms, payload and callable data
1589 // and check for collisions as the accumulation is done.
1590 //
1591 // Returns < 0 if no collision, >= 0 if collision and the value returned is a colliding value.
1592 //
1593 // typeCollision is set to true if there is no direct collision, but the types in the same location
1594 // are different.
1595 //
addUsedLocation(const TQualifier & qualifier,const TType & type,bool & typeCollision)1596 int TIntermediate::addUsedLocation(const TQualifier& qualifier, const TType& type, bool& typeCollision)
1597 {
1598 typeCollision = false;
1599
1600 int set;
1601 int setRT;
1602 if (qualifier.isPipeInput())
1603 set = 0;
1604 else if (qualifier.isPipeOutput())
1605 set = 1;
1606 else if (qualifier.storage == EvqUniform)
1607 set = 2;
1608 else if (qualifier.storage == EvqBuffer)
1609 set = 3;
1610 else if (qualifier.isAnyPayload())
1611 setRT = 0;
1612 else if (qualifier.isAnyCallable())
1613 setRT = 1;
1614 else
1615 return -1;
1616
1617 int size;
1618 if (qualifier.isAnyPayload() || qualifier.isAnyCallable()) {
1619 size = 1;
1620 } else if (qualifier.isUniformOrBuffer() || qualifier.isTaskMemory()) {
1621 if (type.isSizedArray())
1622 size = type.getCumulativeArraySize();
1623 else
1624 size = 1;
1625 } else {
1626 // Strip off the outer array dimension for those having an extra one.
1627 if (type.isArray() && qualifier.isArrayedIo(language)) {
1628 TType elementType(type, 0);
1629 size = computeTypeLocationSize(elementType, language);
1630 } else
1631 size = computeTypeLocationSize(type, language);
1632 }
1633
1634 // Locations, and components within locations.
1635 //
1636 // Almost always, dealing with components means a single location is involved.
1637 // The exception is a dvec3. From the spec:
1638 //
1639 // "A dvec3 will consume all four components of the first location and components 0 and 1 of
1640 // the second location. This leaves components 2 and 3 available for other component-qualified
1641 // declarations."
1642 //
1643 // That means, without ever mentioning a component, a component range
1644 // for a different location gets specified, if it's not a vertex shader input. (!)
1645 // (A vertex shader input will show using only one location, even for a dvec3/4.)
1646 //
1647 // So, for the case of dvec3, we need two independent ioRanges.
1648 //
1649 // For raytracing IO (payloads and callabledata) each declaration occupies a single
1650 // slot irrespective of type.
1651 int collision = -1; // no collision
1652 #ifndef GLSLANG_WEB
1653 if (qualifier.isAnyPayload() || qualifier.isAnyCallable()) {
1654 TRange range(qualifier.layoutLocation, qualifier.layoutLocation);
1655 collision = checkLocationRT(setRT, qualifier.layoutLocation);
1656 if (collision < 0)
1657 usedIoRT[setRT].push_back(range);
1658 } else if (size == 2 && type.getBasicType() == EbtDouble && type.getVectorSize() == 3 &&
1659 (qualifier.isPipeInput() || qualifier.isPipeOutput())) {
1660 // Dealing with dvec3 in/out split across two locations.
1661 // Need two io-ranges.
1662 // The case where the dvec3 doesn't start at component 0 was previously caught as overflow.
1663
1664 // First range:
1665 TRange locationRange(qualifier.layoutLocation, qualifier.layoutLocation);
1666 TRange componentRange(0, 3);
1667 TIoRange range(locationRange, componentRange, type.getBasicType(), 0);
1668
1669 // check for collisions
1670 collision = checkLocationRange(set, range, type, typeCollision);
1671 if (collision < 0) {
1672 usedIo[set].push_back(range);
1673
1674 // Second range:
1675 TRange locationRange2(qualifier.layoutLocation + 1, qualifier.layoutLocation + 1);
1676 TRange componentRange2(0, 1);
1677 TIoRange range2(locationRange2, componentRange2, type.getBasicType(), 0);
1678
1679 // check for collisions
1680 collision = checkLocationRange(set, range2, type, typeCollision);
1681 if (collision < 0)
1682 usedIo[set].push_back(range2);
1683 }
1684 } else
1685 #endif
1686 {
1687 // Not a dvec3 in/out split across two locations, generic path.
1688 // Need a single IO-range block.
1689
1690 TRange locationRange(qualifier.layoutLocation, qualifier.layoutLocation + size - 1);
1691 TRange componentRange(0, 3);
1692 if (qualifier.hasComponent() || type.getVectorSize() > 0) {
1693 int consumedComponents = type.getVectorSize() * (type.getBasicType() == EbtDouble ? 2 : 1);
1694 if (qualifier.hasComponent())
1695 componentRange.start = qualifier.layoutComponent;
1696 componentRange.last = componentRange.start + consumedComponents - 1;
1697 }
1698
1699 // combine location and component ranges
1700 TIoRange range(locationRange, componentRange, type.getBasicType(), qualifier.hasIndex() ? qualifier.getIndex() : 0);
1701
1702 // check for collisions, except for vertex inputs on desktop targeting OpenGL
1703 if (! (!isEsProfile() && language == EShLangVertex && qualifier.isPipeInput()) || spvVersion.vulkan > 0)
1704 collision = checkLocationRange(set, range, type, typeCollision);
1705
1706 if (collision < 0)
1707 usedIo[set].push_back(range);
1708 }
1709
1710 return collision;
1711 }
1712
1713 // Compare a new (the passed in) 'range' against the existing set, and see
1714 // if there are any collisions.
1715 //
1716 // Returns < 0 if no collision, >= 0 if collision and the value returned is a colliding value.
1717 //
checkLocationRange(int set,const TIoRange & range,const TType & type,bool & typeCollision)1718 int TIntermediate::checkLocationRange(int set, const TIoRange& range, const TType& type, bool& typeCollision)
1719 {
1720 for (size_t r = 0; r < usedIo[set].size(); ++r) {
1721 if (range.overlap(usedIo[set][r])) {
1722 // there is a collision; pick one
1723 return std::max(range.location.start, usedIo[set][r].location.start);
1724 } else if (range.location.overlap(usedIo[set][r].location) && type.getBasicType() != usedIo[set][r].basicType) {
1725 // aliased-type mismatch
1726 typeCollision = true;
1727 return std::max(range.location.start, usedIo[set][r].location.start);
1728 }
1729 }
1730
1731 return -1; // no collision
1732 }
1733
checkLocationRT(int set,int location)1734 int TIntermediate::checkLocationRT(int set, int location) {
1735 TRange range(location, location);
1736 for (size_t r = 0; r < usedIoRT[set].size(); ++r) {
1737 if (range.overlap(usedIoRT[set][r])) {
1738 return range.start;
1739 }
1740 }
1741 return -1; // no collision
1742 }
1743
1744 // Accumulate bindings and offsets, and check for collisions
1745 // as the accumulation is done.
1746 //
1747 // Returns < 0 if no collision, >= 0 if collision and the value returned is a colliding value.
1748 //
addUsedOffsets(int binding,int offset,int numOffsets)1749 int TIntermediate::addUsedOffsets(int binding, int offset, int numOffsets)
1750 {
1751 TRange bindingRange(binding, binding);
1752 TRange offsetRange(offset, offset + numOffsets - 1);
1753 TOffsetRange range(bindingRange, offsetRange);
1754
1755 // check for collisions, except for vertex inputs on desktop
1756 for (size_t r = 0; r < usedAtomics.size(); ++r) {
1757 if (range.overlap(usedAtomics[r])) {
1758 // there is a collision; pick one
1759 return std::max(offset, usedAtomics[r].offset.start);
1760 }
1761 }
1762
1763 usedAtomics.push_back(range);
1764
1765 return -1; // no collision
1766 }
1767
1768 // Accumulate used constant_id values.
1769 //
1770 // Return false is one was already used.
addUsedConstantId(int id)1771 bool TIntermediate::addUsedConstantId(int id)
1772 {
1773 if (usedConstantId.find(id) != usedConstantId.end())
1774 return false;
1775
1776 usedConstantId.insert(id);
1777
1778 return true;
1779 }
1780
1781 // Recursively figure out how many locations are used up by an input or output type.
1782 // Return the size of type, as measured by "locations".
computeTypeLocationSize(const TType & type,EShLanguage stage)1783 int TIntermediate::computeTypeLocationSize(const TType& type, EShLanguage stage)
1784 {
1785 // "If the declared input is an array of size n and each element takes m locations, it will be assigned m * n
1786 // consecutive locations..."
1787 if (type.isArray()) {
1788 // TODO: perf: this can be flattened by using getCumulativeArraySize(), and a deref that discards all arrayness
1789 // TODO: are there valid cases of having an unsized array with a location? If so, running this code too early.
1790 TType elementType(type, 0);
1791 if (type.isSizedArray() && !type.getQualifier().isPerView())
1792 return type.getOuterArraySize() * computeTypeLocationSize(elementType, stage);
1793 else {
1794 #ifndef GLSLANG_WEB
1795 // unset perViewNV attributes for arrayed per-view outputs: "perviewNV vec4 v[MAX_VIEWS][3];"
1796 elementType.getQualifier().perViewNV = false;
1797 #endif
1798 return computeTypeLocationSize(elementType, stage);
1799 }
1800 }
1801
1802 // "The locations consumed by block and structure members are determined by applying the rules above
1803 // recursively..."
1804 if (type.isStruct()) {
1805 int size = 0;
1806 for (int member = 0; member < (int)type.getStruct()->size(); ++member) {
1807 TType memberType(type, member);
1808 size += computeTypeLocationSize(memberType, stage);
1809 }
1810 return size;
1811 }
1812
1813 // ES: "If a shader input is any scalar or vector type, it will consume a single location."
1814
1815 // Desktop: "If a vertex shader input is any scalar or vector type, it will consume a single location. If a non-vertex
1816 // shader input is a scalar or vector type other than dvec3 or dvec4, it will consume a single location, while
1817 // types dvec3 or dvec4 will consume two consecutive locations. Inputs of type double and dvec2 will
1818 // consume only a single location, in all stages."
1819 if (type.isScalar())
1820 return 1;
1821 if (type.isVector()) {
1822 if (stage == EShLangVertex && type.getQualifier().isPipeInput())
1823 return 1;
1824 if (type.getBasicType() == EbtDouble && type.getVectorSize() > 2)
1825 return 2;
1826 else
1827 return 1;
1828 }
1829
1830 // "If the declared input is an n x m single- or double-precision matrix, ...
1831 // The number of locations assigned for each matrix will be the same as
1832 // for an n-element array of m-component vectors..."
1833 if (type.isMatrix()) {
1834 TType columnType(type, 0);
1835 return type.getMatrixCols() * computeTypeLocationSize(columnType, stage);
1836 }
1837
1838 assert(0);
1839 return 1;
1840 }
1841
1842 // Same as computeTypeLocationSize but for uniforms
computeTypeUniformLocationSize(const TType & type)1843 int TIntermediate::computeTypeUniformLocationSize(const TType& type)
1844 {
1845 // "Individual elements of a uniform array are assigned
1846 // consecutive locations with the first element taking location
1847 // location."
1848 if (type.isArray()) {
1849 // TODO: perf: this can be flattened by using getCumulativeArraySize(), and a deref that discards all arrayness
1850 TType elementType(type, 0);
1851 if (type.isSizedArray()) {
1852 return type.getOuterArraySize() * computeTypeUniformLocationSize(elementType);
1853 } else {
1854 // TODO: are there valid cases of having an implicitly-sized array with a location? If so, running this code too early.
1855 return computeTypeUniformLocationSize(elementType);
1856 }
1857 }
1858
1859 // "Each subsequent inner-most member or element gets incremental
1860 // locations for the entire structure or array."
1861 if (type.isStruct()) {
1862 int size = 0;
1863 for (int member = 0; member < (int)type.getStruct()->size(); ++member) {
1864 TType memberType(type, member);
1865 size += computeTypeUniformLocationSize(memberType);
1866 }
1867 return size;
1868 }
1869
1870 return 1;
1871 }
1872
1873 #ifndef GLSLANG_WEB
1874
1875 // Accumulate xfb buffer ranges and check for collisions as the accumulation is done.
1876 //
1877 // Returns < 0 if no collision, >= 0 if collision and the value returned is a colliding value.
1878 //
addXfbBufferOffset(const TType & type)1879 int TIntermediate::addXfbBufferOffset(const TType& type)
1880 {
1881 const TQualifier& qualifier = type.getQualifier();
1882
1883 assert(qualifier.hasXfbOffset() && qualifier.hasXfbBuffer());
1884 TXfbBuffer& buffer = xfbBuffers[qualifier.layoutXfbBuffer];
1885
1886 // compute the range
1887 unsigned int size = computeTypeXfbSize(type, buffer.contains64BitType, buffer.contains32BitType, buffer.contains16BitType);
1888 buffer.implicitStride = std::max(buffer.implicitStride, qualifier.layoutXfbOffset + size);
1889 TRange range(qualifier.layoutXfbOffset, qualifier.layoutXfbOffset + size - 1);
1890
1891 // check for collisions
1892 for (size_t r = 0; r < buffer.ranges.size(); ++r) {
1893 if (range.overlap(buffer.ranges[r])) {
1894 // there is a collision; pick an example to return
1895 return std::max(range.start, buffer.ranges[r].start);
1896 }
1897 }
1898
1899 buffer.ranges.push_back(range);
1900
1901 return -1; // no collision
1902 }
1903
1904 // Recursively figure out how many bytes of xfb buffer are used by the given type.
1905 // Return the size of type, in bytes.
1906 // Sets contains64BitType to true if the type contains a 64-bit data type.
1907 // Sets contains32BitType to true if the type contains a 32-bit data type.
1908 // Sets contains16BitType to true if the type contains a 16-bit data type.
1909 // N.B. Caller must set contains64BitType, contains32BitType, and contains16BitType to false before calling.
computeTypeXfbSize(const TType & type,bool & contains64BitType,bool & contains32BitType,bool & contains16BitType) const1910 unsigned int TIntermediate::computeTypeXfbSize(const TType& type, bool& contains64BitType, bool& contains32BitType, bool& contains16BitType) const
1911 {
1912 // "...if applied to an aggregate containing a double or 64-bit integer, the offset must also be a multiple of 8,
1913 // and the space taken in the buffer will be a multiple of 8.
1914 // ...within the qualified entity, subsequent components are each
1915 // assigned, in order, to the next available offset aligned to a multiple of
1916 // that component's size. Aggregate types are flattened down to the component
1917 // level to get this sequence of components."
1918
1919 if (type.isSizedArray()) {
1920 // TODO: perf: this can be flattened by using getCumulativeArraySize(), and a deref that discards all arrayness
1921 // Unsized array use to xfb should be a compile error.
1922 TType elementType(type, 0);
1923 return type.getOuterArraySize() * computeTypeXfbSize(elementType, contains64BitType, contains16BitType, contains16BitType);
1924 }
1925
1926 if (type.isStruct()) {
1927 unsigned int size = 0;
1928 bool structContains64BitType = false;
1929 bool structContains32BitType = false;
1930 bool structContains16BitType = false;
1931 for (int member = 0; member < (int)type.getStruct()->size(); ++member) {
1932 TType memberType(type, member);
1933 // "... if applied to
1934 // an aggregate containing a double or 64-bit integer, the offset must also be a multiple of 8,
1935 // and the space taken in the buffer will be a multiple of 8."
1936 bool memberContains64BitType = false;
1937 bool memberContains32BitType = false;
1938 bool memberContains16BitType = false;
1939 int memberSize = computeTypeXfbSize(memberType, memberContains64BitType, memberContains32BitType, memberContains16BitType);
1940 if (memberContains64BitType) {
1941 structContains64BitType = true;
1942 RoundToPow2(size, 8);
1943 } else if (memberContains32BitType) {
1944 structContains32BitType = true;
1945 RoundToPow2(size, 4);
1946 } else if (memberContains16BitType) {
1947 structContains16BitType = true;
1948 RoundToPow2(size, 2);
1949 }
1950 size += memberSize;
1951 }
1952
1953 if (structContains64BitType) {
1954 contains64BitType = true;
1955 RoundToPow2(size, 8);
1956 } else if (structContains32BitType) {
1957 contains32BitType = true;
1958 RoundToPow2(size, 4);
1959 } else if (structContains16BitType) {
1960 contains16BitType = true;
1961 RoundToPow2(size, 2);
1962 }
1963 return size;
1964 }
1965
1966 int numComponents {0};
1967 if (type.isScalar())
1968 numComponents = 1;
1969 else if (type.isVector())
1970 numComponents = type.getVectorSize();
1971 else if (type.isMatrix())
1972 numComponents = type.getMatrixCols() * type.getMatrixRows();
1973 else {
1974 assert(0);
1975 numComponents = 1;
1976 }
1977
1978 if (type.getBasicType() == EbtDouble || type.getBasicType() == EbtInt64 || type.getBasicType() == EbtUint64) {
1979 contains64BitType = true;
1980 return 8 * numComponents;
1981 } else if (type.getBasicType() == EbtFloat16 || type.getBasicType() == EbtInt16 || type.getBasicType() == EbtUint16) {
1982 contains16BitType = true;
1983 return 2 * numComponents;
1984 } else if (type.getBasicType() == EbtInt8 || type.getBasicType() == EbtUint8)
1985 return numComponents;
1986 else {
1987 contains32BitType = true;
1988 return 4 * numComponents;
1989 }
1990 }
1991
1992 #endif
1993
1994 const int baseAlignmentVec4Std140 = 16;
1995
1996 // Return the size and alignment of a component of the given type.
1997 // The size is returned in the 'size' parameter
1998 // Return value is the alignment..
getBaseAlignmentScalar(const TType & type,int & size)1999 int TIntermediate::getBaseAlignmentScalar(const TType& type, int& size)
2000 {
2001 #ifdef GLSLANG_WEB
2002 size = 4; return 4;
2003 #endif
2004
2005 switch (type.getBasicType()) {
2006 case EbtInt64:
2007 case EbtUint64:
2008 case EbtDouble: size = 8; return 8;
2009 case EbtFloat16: size = 2; return 2;
2010 case EbtInt8:
2011 case EbtUint8: size = 1; return 1;
2012 case EbtInt16:
2013 case EbtUint16: size = 2; return 2;
2014 case EbtReference: size = 8; return 8;
2015 default: size = 4; return 4;
2016 }
2017 }
2018
2019 // Implement base-alignment and size rules from section 7.6.2.2 Standard Uniform Block Layout
2020 // Operates recursively.
2021 //
2022 // If std140 is true, it does the rounding up to vec4 size required by std140,
2023 // otherwise it does not, yielding std430 rules.
2024 //
2025 // The size is returned in the 'size' parameter
2026 //
2027 // The stride is only non-0 for arrays or matrices, and is the stride of the
2028 // top-level object nested within the type. E.g., for an array of matrices,
2029 // it is the distances needed between matrices, despite the rules saying the
2030 // stride comes from the flattening down to vectors.
2031 //
2032 // Return value is the alignment of the type.
getBaseAlignment(const TType & type,int & size,int & stride,TLayoutPacking layoutPacking,bool rowMajor)2033 int TIntermediate::getBaseAlignment(const TType& type, int& size, int& stride, TLayoutPacking layoutPacking, bool rowMajor)
2034 {
2035 int alignment;
2036
2037 bool std140 = layoutPacking == glslang::ElpStd140;
2038 // When using the std140 storage layout, structures will be laid out in buffer
2039 // storage with its members stored in monotonically increasing order based on their
2040 // location in the declaration. A structure and each structure member have a base
2041 // offset and a base alignment, from which an aligned offset is computed by rounding
2042 // the base offset up to a multiple of the base alignment. The base offset of the first
2043 // member of a structure is taken from the aligned offset of the structure itself. The
2044 // base offset of all other structure members is derived by taking the offset of the
2045 // last basic machine unit consumed by the previous member and adding one. Each
2046 // structure member is stored in memory at its aligned offset. The members of a top-
2047 // level uniform block are laid out in buffer storage by treating the uniform block as
2048 // a structure with a base offset of zero.
2049 //
2050 // 1. If the member is a scalar consuming N basic machine units, the base alignment is N.
2051 //
2052 // 2. If the member is a two- or four-component vector with components consuming N basic
2053 // machine units, the base alignment is 2N or 4N, respectively.
2054 //
2055 // 3. If the member is a three-component vector with components consuming N
2056 // basic machine units, the base alignment is 4N.
2057 //
2058 // 4. If the member is an array of scalars or vectors, the base alignment and array
2059 // stride are set to match the base alignment of a single array element, according
2060 // to rules (1), (2), and (3), and rounded up to the base alignment of a vec4. The
2061 // array may have padding at the end; the base offset of the member following
2062 // the array is rounded up to the next multiple of the base alignment.
2063 //
2064 // 5. If the member is a column-major matrix with C columns and R rows, the
2065 // matrix is stored identically to an array of C column vectors with R
2066 // components each, according to rule (4).
2067 //
2068 // 6. If the member is an array of S column-major matrices with C columns and
2069 // R rows, the matrix is stored identically to a row of S X C column vectors
2070 // with R components each, according to rule (4).
2071 //
2072 // 7. If the member is a row-major matrix with C columns and R rows, the matrix
2073 // is stored identically to an array of R row vectors with C components each,
2074 // according to rule (4).
2075 //
2076 // 8. If the member is an array of S row-major matrices with C columns and R
2077 // rows, the matrix is stored identically to a row of S X R row vectors with C
2078 // components each, according to rule (4).
2079 //
2080 // 9. If the member is a structure, the base alignment of the structure is N , where
2081 // N is the largest base alignment value of any of its members, and rounded
2082 // up to the base alignment of a vec4. The individual members of this substructure
2083 // are then assigned offsets by applying this set of rules recursively,
2084 // where the base offset of the first member of the sub-structure is equal to the
2085 // aligned offset of the structure. The structure may have padding at the end;
2086 // the base offset of the member following the sub-structure is rounded up to
2087 // the next multiple of the base alignment of the structure.
2088 //
2089 // 10. If the member is an array of S structures, the S elements of the array are laid
2090 // out in order, according to rule (9).
2091 //
2092 // Assuming, for rule 10: The stride is the same as the size of an element.
2093
2094 stride = 0;
2095 int dummyStride;
2096
2097 // rules 4, 6, 8, and 10
2098 if (type.isArray()) {
2099 // TODO: perf: this might be flattened by using getCumulativeArraySize(), and a deref that discards all arrayness
2100 TType derefType(type, 0);
2101 alignment = getBaseAlignment(derefType, size, dummyStride, layoutPacking, rowMajor);
2102 if (std140)
2103 alignment = std::max(baseAlignmentVec4Std140, alignment);
2104 RoundToPow2(size, alignment);
2105 stride = size; // uses full matrix size for stride of an array of matrices (not quite what rule 6/8, but what's expected)
2106 // uses the assumption for rule 10 in the comment above
2107 // use one element to represent the last member of SSBO which is unsized array
2108 int arraySize = (type.isUnsizedArray() && (type.getOuterArraySize() == 0)) ? 1 : type.getOuterArraySize();
2109 size = stride * arraySize;
2110 return alignment;
2111 }
2112
2113 // rule 9
2114 if (type.getBasicType() == EbtStruct || type.getBasicType() == EbtBlock) {
2115 const TTypeList& memberList = *type.getStruct();
2116
2117 size = 0;
2118 int maxAlignment = std140 ? baseAlignmentVec4Std140 : 0;
2119 for (size_t m = 0; m < memberList.size(); ++m) {
2120 int memberSize;
2121 // modify just the children's view of matrix layout, if there is one for this member
2122 TLayoutMatrix subMatrixLayout = memberList[m].type->getQualifier().layoutMatrix;
2123 int memberAlignment = getBaseAlignment(*memberList[m].type, memberSize, dummyStride, layoutPacking,
2124 (subMatrixLayout != ElmNone) ? (subMatrixLayout == ElmRowMajor) : rowMajor);
2125 maxAlignment = std::max(maxAlignment, memberAlignment);
2126 RoundToPow2(size, memberAlignment);
2127 size += memberSize;
2128 }
2129
2130 // The structure may have padding at the end; the base offset of
2131 // the member following the sub-structure is rounded up to the next
2132 // multiple of the base alignment of the structure.
2133 RoundToPow2(size, maxAlignment);
2134
2135 return maxAlignment;
2136 }
2137
2138 // rule 1
2139 if (type.isScalar())
2140 return getBaseAlignmentScalar(type, size);
2141
2142 // rules 2 and 3
2143 if (type.isVector()) {
2144 int scalarAlign = getBaseAlignmentScalar(type, size);
2145 switch (type.getVectorSize()) {
2146 case 1: // HLSL has this, GLSL does not
2147 return scalarAlign;
2148 case 2:
2149 size *= 2;
2150 return 2 * scalarAlign;
2151 default:
2152 size *= type.getVectorSize();
2153 return 4 * scalarAlign;
2154 }
2155 }
2156
2157 // rules 5 and 7
2158 if (type.isMatrix()) {
2159 // rule 5: deref to row, not to column, meaning the size of vector is num columns instead of num rows
2160 TType derefType(type, 0, rowMajor);
2161
2162 alignment = getBaseAlignment(derefType, size, dummyStride, layoutPacking, rowMajor);
2163 if (std140)
2164 alignment = std::max(baseAlignmentVec4Std140, alignment);
2165 RoundToPow2(size, alignment);
2166 stride = size; // use intra-matrix stride for stride of a just a matrix
2167 if (rowMajor)
2168 size = stride * type.getMatrixRows();
2169 else
2170 size = stride * type.getMatrixCols();
2171
2172 return alignment;
2173 }
2174
2175 assert(0); // all cases should be covered above
2176 size = baseAlignmentVec4Std140;
2177 return baseAlignmentVec4Std140;
2178 }
2179
2180 // To aid the basic HLSL rule about crossing vec4 boundaries.
improperStraddle(const TType & type,int size,int offset)2181 bool TIntermediate::improperStraddle(const TType& type, int size, int offset)
2182 {
2183 if (! type.isVector() || type.isArray())
2184 return false;
2185
2186 return size <= 16 ? offset / 16 != (offset + size - 1) / 16
2187 : offset % 16 != 0;
2188 }
2189
getScalarAlignment(const TType & type,int & size,int & stride,bool rowMajor)2190 int TIntermediate::getScalarAlignment(const TType& type, int& size, int& stride, bool rowMajor)
2191 {
2192 int alignment;
2193
2194 stride = 0;
2195 int dummyStride;
2196
2197 if (type.isArray()) {
2198 TType derefType(type, 0);
2199 alignment = getScalarAlignment(derefType, size, dummyStride, rowMajor);
2200
2201 stride = size;
2202 RoundToPow2(stride, alignment);
2203
2204 size = stride * (type.getOuterArraySize() - 1) + size;
2205 return alignment;
2206 }
2207
2208 if (type.getBasicType() == EbtStruct) {
2209 const TTypeList& memberList = *type.getStruct();
2210
2211 size = 0;
2212 int maxAlignment = 0;
2213 for (size_t m = 0; m < memberList.size(); ++m) {
2214 int memberSize;
2215 // modify just the children's view of matrix layout, if there is one for this member
2216 TLayoutMatrix subMatrixLayout = memberList[m].type->getQualifier().layoutMatrix;
2217 int memberAlignment = getScalarAlignment(*memberList[m].type, memberSize, dummyStride,
2218 (subMatrixLayout != ElmNone) ? (subMatrixLayout == ElmRowMajor) : rowMajor);
2219 maxAlignment = std::max(maxAlignment, memberAlignment);
2220 RoundToPow2(size, memberAlignment);
2221 size += memberSize;
2222 }
2223
2224 return maxAlignment;
2225 }
2226
2227 if (type.isScalar())
2228 return getBaseAlignmentScalar(type, size);
2229
2230 if (type.isVector()) {
2231 int scalarAlign = getBaseAlignmentScalar(type, size);
2232
2233 size *= type.getVectorSize();
2234 return scalarAlign;
2235 }
2236
2237 if (type.isMatrix()) {
2238 TType derefType(type, 0, rowMajor);
2239
2240 alignment = getScalarAlignment(derefType, size, dummyStride, rowMajor);
2241
2242 stride = size; // use intra-matrix stride for stride of a just a matrix
2243 if (rowMajor)
2244 size = stride * type.getMatrixRows();
2245 else
2246 size = stride * type.getMatrixCols();
2247
2248 return alignment;
2249 }
2250
2251 assert(0); // all cases should be covered above
2252 size = 1;
2253 return 1;
2254 }
2255
getMemberAlignment(const TType & type,int & size,int & stride,TLayoutPacking layoutPacking,bool rowMajor)2256 int TIntermediate::getMemberAlignment(const TType& type, int& size, int& stride, TLayoutPacking layoutPacking, bool rowMajor)
2257 {
2258 if (layoutPacking == glslang::ElpScalar) {
2259 return getScalarAlignment(type, size, stride, rowMajor);
2260 } else {
2261 return getBaseAlignment(type, size, stride, layoutPacking, rowMajor);
2262 }
2263 }
2264
2265 // shared calculation by getOffset and getOffsets
updateOffset(const TType & parentType,const TType & memberType,int & offset,int & memberSize)2266 void TIntermediate::updateOffset(const TType& parentType, const TType& memberType, int& offset, int& memberSize)
2267 {
2268 int dummyStride;
2269
2270 // modify just the children's view of matrix layout, if there is one for this member
2271 TLayoutMatrix subMatrixLayout = memberType.getQualifier().layoutMatrix;
2272 int memberAlignment = getMemberAlignment(memberType, memberSize, dummyStride,
2273 parentType.getQualifier().layoutPacking,
2274 subMatrixLayout != ElmNone
2275 ? subMatrixLayout == ElmRowMajor
2276 : parentType.getQualifier().layoutMatrix == ElmRowMajor);
2277 RoundToPow2(offset, memberAlignment);
2278 }
2279
2280 // Lookup or calculate the offset of a block member, using the recursively
2281 // defined block offset rules.
getOffset(const TType & type,int index)2282 int TIntermediate::getOffset(const TType& type, int index)
2283 {
2284 const TTypeList& memberList = *type.getStruct();
2285
2286 // Don't calculate offset if one is present, it could be user supplied
2287 // and different than what would be calculated. That is, this is faster,
2288 // but not just an optimization.
2289 if (memberList[index].type->getQualifier().hasOffset())
2290 return memberList[index].type->getQualifier().layoutOffset;
2291
2292 int memberSize = 0;
2293 int offset = 0;
2294 for (int m = 0; m <= index; ++m) {
2295 updateOffset(type, *memberList[m].type, offset, memberSize);
2296
2297 if (m < index)
2298 offset += memberSize;
2299 }
2300
2301 return offset;
2302 }
2303
2304 // Calculate the block data size.
2305 // Block arrayness is not taken into account, each element is backed by a separate buffer.
getBlockSize(const TType & blockType)2306 int TIntermediate::getBlockSize(const TType& blockType)
2307 {
2308 const TTypeList& memberList = *blockType.getStruct();
2309 int lastIndex = (int)memberList.size() - 1;
2310 int lastOffset = getOffset(blockType, lastIndex);
2311
2312 int lastMemberSize;
2313 int dummyStride;
2314 getMemberAlignment(*memberList[lastIndex].type, lastMemberSize, dummyStride,
2315 blockType.getQualifier().layoutPacking,
2316 blockType.getQualifier().layoutMatrix == ElmRowMajor);
2317
2318 return lastOffset + lastMemberSize;
2319 }
2320
computeBufferReferenceTypeSize(const TType & type)2321 int TIntermediate::computeBufferReferenceTypeSize(const TType& type)
2322 {
2323 assert(type.isReference());
2324 int size = getBlockSize(*type.getReferentType());
2325
2326 int align = type.getBufferReferenceAlignment();
2327
2328 if (align) {
2329 size = (size + align - 1) & ~(align-1);
2330 }
2331
2332 return size;
2333 }
2334
2335 #ifndef GLSLANG_WEB
isIoResizeArray(const TType & type,EShLanguage language)2336 bool TIntermediate::isIoResizeArray(const TType& type, EShLanguage language) {
2337 return type.isArray() &&
2338 ((language == EShLangGeometry && type.getQualifier().storage == EvqVaryingIn) ||
2339 (language == EShLangTessControl && (type.getQualifier().storage == EvqVaryingIn || type.getQualifier().storage == EvqVaryingOut) &&
2340 ! type.getQualifier().patch) ||
2341 (language == EShLangTessEvaluation && type.getQualifier().storage == EvqVaryingIn) ||
2342 (language == EShLangFragment && type.getQualifier().storage == EvqVaryingIn &&
2343 (type.getQualifier().pervertexNV || type.getQualifier().pervertexEXT)) ||
2344 (language == EShLangMesh && type.getQualifier().storage == EvqVaryingOut &&
2345 !type.getQualifier().perTaskNV));
2346 }
2347 #endif // not GLSLANG_WEB
2348
2349 } // end namespace glslang
2350