2 // Copyright (C) 2018 Google, Inc.
4 // All rights reserved.
6 // Redistribution and use in source and binary forms, with or without
7 // modification, are permitted provided that the following conditions
10 // Redistributions of source code must retain the above copyright
11 // notice, this list of conditions and the following disclaimer.
13 // Redistributions in binary form must reproduce the above
14 // copyright notice, this list of conditions and the following
15 // disclaimer in the documentation and/or other materials provided
16 // with the distribution.
18 // Neither the name of 3Dlabs Inc. Ltd. nor the names of its
19 // contributors may be used to endorse or promote products derived
20 // from this software without specific prior written permission.
22 // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
23 // "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
24 // LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
25 // FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
26 // COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
27 // INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
28 // BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
29 // LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
30 // CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31 // LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
32 // ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
33 // POSSIBILITY OF SUCH DAMAGE.
36 // Post-processing for SPIR-V IR, in internal form, not standard binary form.
42 #include <unordered_map>
43 #include <unordered_set>
46 #include "SpvBuilder.h"
50 #include "GLSL.std.450.h"
51 #include "GLSL.ext.KHR.h"
52 #include "GLSL.ext.EXT.h"
53 #include "GLSL.ext.AMD.h"
54 #include "GLSL.ext.NV.h"
55 #include "GLSL.ext.ARM.h"
56 #include "GLSL.ext.QCOM.h"
61 // Hook to visit each operand type and result type of an instruction.
62 // Will be called multiple times for one instruction, once for each typed
63 // operand and the result.
64 void Builder::postProcessType(const Instruction& inst, Id typeId)
66 // Characterize the type being questioned
67 Id basicTypeOp = getMostBasicTypeClass(typeId);
69 if (basicTypeOp == OpTypeFloat || basicTypeOp == OpTypeInt)
70 width = getScalarTypeWidth(typeId);
72 // Do opcode-specific checks
73 switch (inst.getOpCode()) {
76 if (basicTypeOp == OpTypeStruct) {
77 if (containsType(typeId, OpTypeInt, 8))
78 addCapability(CapabilityInt8);
79 if (containsType(typeId, OpTypeInt, 16))
80 addCapability(CapabilityInt16);
81 if (containsType(typeId, OpTypeFloat, 16))
82 addCapability(CapabilityFloat16);
84 StorageClass storageClass = getStorageClass(inst.getIdOperand(0));
86 switch (storageClass) {
87 case StorageClassPhysicalStorageBufferEXT:
88 case StorageClassUniform:
89 case StorageClassStorageBuffer:
90 case StorageClassPushConstant:
93 addCapability(CapabilityInt8);
96 } else if (width == 16) {
97 switch (storageClass) {
98 case StorageClassPhysicalStorageBufferEXT:
99 case StorageClassUniform:
100 case StorageClassStorageBuffer:
101 case StorageClassPushConstant:
102 case StorageClassInput:
103 case StorageClassOutput:
106 if (basicTypeOp == OpTypeInt)
107 addCapability(CapabilityInt16);
108 if (basicTypeOp == OpTypeFloat)
109 addCapability(CapabilityFloat16);
120 // Look for any 8/16-bit storage capabilities. If there are none, assume that
121 // the convert instruction requires the Float16/Int8/16 capability.
122 if (containsType(typeId, OpTypeFloat, 16) || containsType(typeId, OpTypeInt, 16)) {
123 bool foundStorage = false;
124 for (auto it = capabilities.begin(); it != capabilities.end(); ++it) {
125 spv::Capability cap = *it;
126 if (cap == spv::CapabilityStorageInputOutput16 ||
127 cap == spv::CapabilityStoragePushConstant16 ||
128 cap == spv::CapabilityStorageUniformBufferBlock16 ||
129 cap == spv::CapabilityStorageUniform16) {
135 if (containsType(typeId, OpTypeFloat, 16))
136 addCapability(CapabilityFloat16);
137 if (containsType(typeId, OpTypeInt, 16))
138 addCapability(CapabilityInt16);
141 if (containsType(typeId, OpTypeInt, 8)) {
142 bool foundStorage = false;
143 for (auto it = capabilities.begin(); it != capabilities.end(); ++it) {
144 spv::Capability cap = *it;
145 if (cap == spv::CapabilityStoragePushConstant8 ||
146 cap == spv::CapabilityUniformAndStorageBuffer8BitAccess ||
147 cap == spv::CapabilityStorageBuffer8BitAccess) {
153 addCapability(CapabilityInt8);
158 switch (inst.getImmediateOperand(1)) {
159 case GLSLstd450Frexp:
160 case GLSLstd450FrexpStruct:
161 if (getSpvVersion() < spv::Spv_1_3 && containsType(typeId, OpTypeInt, 16))
162 addExtension(spv::E_SPV_AMD_gpu_shader_int16);
164 case GLSLstd450InterpolateAtCentroid:
165 case GLSLstd450InterpolateAtSample:
166 case GLSLstd450InterpolateAtOffset:
167 if (getSpvVersion() < spv::Spv_1_3 && containsType(typeId, OpTypeFloat, 16))
168 addExtension(spv::E_SPV_AMD_gpu_shader_half_float);
175 case OpPtrAccessChain:
176 if (isPointerType(typeId))
178 if (basicTypeOp == OpTypeInt) {
180 addCapability(CapabilityInt16);
182 addCapability(CapabilityInt8);
185 if (basicTypeOp == OpTypeInt) {
187 addCapability(CapabilityInt16);
189 addCapability(CapabilityInt8);
190 else if (width == 64)
191 addCapability(CapabilityInt64);
192 } else if (basicTypeOp == OpTypeFloat) {
194 addCapability(CapabilityFloat16);
195 else if (width == 64)
196 addCapability(CapabilityFloat64);
202 // Called for each instruction that resides in a block.
203 void Builder::postProcess(Instruction& inst)
205 // Add capabilities based simply on the opcode.
206 switch (inst.getOpCode()) {
208 switch (inst.getImmediateOperand(1)) {
209 case GLSLstd450InterpolateAtCentroid:
210 case GLSLstd450InterpolateAtSample:
211 case GLSLstd450InterpolateAtOffset:
212 addCapability(CapabilityInterpolationFunction);
224 addCapability(CapabilityDerivativeControl);
227 case OpImageQueryLod:
228 case OpImageQuerySize:
229 case OpImageQuerySizeLod:
230 case OpImageQuerySamples:
231 case OpImageQueryLevels:
232 addCapability(CapabilityImageQuery);
235 case OpGroupNonUniformPartitionNV:
236 addExtension(E_SPV_NV_shader_subgroup_partitioned);
237 addCapability(CapabilityGroupNonUniformPartitionedNV);
243 // For any load/store to a PhysicalStorageBufferEXT, walk the accesschain
244 // index list to compute the misalignment. The pre-existing alignment value
245 // (set via Builder::AccessChain::alignment) only accounts for the base of
246 // the reference type and any scalar component selection in the accesschain,
247 // and this function computes the rest from the SPIR-V Offset decorations.
248 Instruction *accessChain = module.getInstruction(inst.getIdOperand(0));
249 if (accessChain->getOpCode() == OpAccessChain) {
250 Instruction *base = module.getInstruction(accessChain->getIdOperand(0));
251 // Get the type of the base of the access chain. It must be a pointer type.
252 Id typeId = base->getTypeId();
253 Instruction *type = module.getInstruction(typeId);
254 assert(type->getOpCode() == OpTypePointer);
255 if (type->getImmediateOperand(0) != StorageClassPhysicalStorageBufferEXT) {
258 // Get the pointee type.
259 typeId = type->getIdOperand(1);
260 type = module.getInstruction(typeId);
261 // Walk the index list for the access chain. For each index, find any
262 // misalignment that can apply when accessing the member/element via
263 // Offset/ArrayStride/MatrixStride decorations, and bitwise OR them all
266 for (int i = 1; i < accessChain->getNumOperands(); ++i) {
267 Instruction *idx = module.getInstruction(accessChain->getIdOperand(i));
268 if (type->getOpCode() == OpTypeStruct) {
269 assert(idx->getOpCode() == OpConstant);
270 unsigned int c = idx->getImmediateOperand(0);
272 const auto function = [&](const std::unique_ptr<Instruction>& decoration) {
273 if (decoration.get()->getOpCode() == OpMemberDecorate &&
274 decoration.get()->getIdOperand(0) == typeId &&
275 decoration.get()->getImmediateOperand(1) == c &&
276 (decoration.get()->getImmediateOperand(2) == DecorationOffset ||
277 decoration.get()->getImmediateOperand(2) == DecorationMatrixStride)) {
278 alignment |= decoration.get()->getImmediateOperand(3);
281 std::for_each(decorations.begin(), decorations.end(), function);
282 // get the next member type
283 typeId = type->getIdOperand(c);
284 type = module.getInstruction(typeId);
285 } else if (type->getOpCode() == OpTypeArray ||
286 type->getOpCode() == OpTypeRuntimeArray) {
287 const auto function = [&](const std::unique_ptr<Instruction>& decoration) {
288 if (decoration.get()->getOpCode() == OpDecorate &&
289 decoration.get()->getIdOperand(0) == typeId &&
290 decoration.get()->getImmediateOperand(1) == DecorationArrayStride) {
291 alignment |= decoration.get()->getImmediateOperand(2);
294 std::for_each(decorations.begin(), decorations.end(), function);
295 // Get the element type
296 typeId = type->getIdOperand(0);
297 type = module.getInstruction(typeId);
299 // Once we get to any non-aggregate type, we're done.
303 assert(inst.getNumOperands() >= 3);
304 unsigned int memoryAccess = inst.getImmediateOperand((inst.getOpCode() == OpStore) ? 2 : 1);
305 assert(memoryAccess & MemoryAccessAlignedMask);
306 static_cast<void>(memoryAccess);
307 // Compute the index of the alignment operand.
308 int alignmentIdx = 2;
309 if (inst.getOpCode() == OpStore)
311 // Merge new and old (mis)alignment
312 alignment |= inst.getImmediateOperand(alignmentIdx);
314 alignment = alignment & ~(alignment & (alignment-1));
315 // update the Aligned operand
316 inst.setImmediateOperand(alignmentIdx, alignment);
325 // Checks based on type
326 if (inst.getTypeId() != NoType)
327 postProcessType(inst, inst.getTypeId());
328 for (int op = 0; op < inst.getNumOperands(); ++op) {
329 if (inst.isIdOperand(op)) {
330 // In blocks, these are always result ids, but we are relying on
331 // getTypeId() to return NoType for things like OpLabel.
332 if (getTypeId(inst.getIdOperand(op)) != NoType)
333 postProcessType(inst, getTypeId(inst.getIdOperand(op)));
339 void Builder::postProcessCFG()
341 // reachableBlocks is the set of blockss reached via control flow, or which are
342 // unreachable continue targert or unreachable merge.
343 std::unordered_set<const Block*> reachableBlocks;
344 std::unordered_map<Block*, Block*> headerForUnreachableContinue;
345 std::unordered_set<Block*> unreachableMerges;
346 std::unordered_set<Id> unreachableDefinitions;
347 // Collect IDs defined in unreachable blocks. For each function, label the
348 // reachable blocks first. Then for each unreachable block, collect the
349 // result IDs of the instructions in it.
350 for (auto fi = module.getFunctions().cbegin(); fi != module.getFunctions().cend(); fi++) {
352 Block* entry = f->getEntryBlock();
353 inReadableOrder(entry,
354 [&reachableBlocks, &unreachableMerges, &headerForUnreachableContinue]
355 (Block* b, ReachReason why, Block* header) {
356 reachableBlocks.insert(b);
357 if (why == ReachDeadContinue) headerForUnreachableContinue[b] = header;
358 if (why == ReachDeadMerge) unreachableMerges.insert(b);
360 for (auto bi = f->getBlocks().cbegin(); bi != f->getBlocks().cend(); bi++) {
362 if (unreachableMerges.count(b) != 0 || headerForUnreachableContinue.count(b) != 0) {
363 auto ii = b->getInstructions().cbegin();
364 ++ii; // Keep potential decorations on the label.
365 for (; ii != b->getInstructions().cend(); ++ii)
366 unreachableDefinitions.insert(ii->get()->getResultId());
367 } else if (reachableBlocks.count(b) == 0) {
368 // The normal case for unreachable code. All definitions are considered dead.
369 for (auto ii = b->getInstructions().cbegin(); ii != b->getInstructions().cend(); ++ii)
370 unreachableDefinitions.insert(ii->get()->getResultId());
375 // Modify unreachable merge blocks and unreachable continue targets.
376 // Delete their contents.
377 for (auto mergeIter = unreachableMerges.begin(); mergeIter != unreachableMerges.end(); ++mergeIter) {
378 (*mergeIter)->rewriteAsCanonicalUnreachableMerge();
380 for (auto continueIter = headerForUnreachableContinue.begin();
381 continueIter != headerForUnreachableContinue.end();
383 Block* continue_target = continueIter->first;
384 Block* header = continueIter->second;
385 continue_target->rewriteAsCanonicalUnreachableContinue(header);
388 // Remove unneeded decorations, for unreachable instructions
389 decorations.erase(std::remove_if(decorations.begin(), decorations.end(),
390 [&unreachableDefinitions](std::unique_ptr<Instruction>& I) -> bool {
391 Id decoration_id = I.get()->getIdOperand(0);
392 return unreachableDefinitions.count(decoration_id) != 0;
398 void Builder::postProcessFeatures() {
399 // Add per-instruction capabilities, extensions, etc.,
401 // Look for any 8/16 bit type in physical storage buffer class, and set the
402 // appropriate capability. This happens in createSpvVariable for other storage
403 // classes, but there isn't always a variable for physical storage buffer.
404 for (int t = 0; t < (int)groupedTypes[OpTypePointer].size(); ++t) {
405 Instruction* type = groupedTypes[OpTypePointer][t];
406 if (type->getImmediateOperand(0) == (unsigned)StorageClassPhysicalStorageBufferEXT) {
407 if (containsType(type->getIdOperand(1), OpTypeInt, 8)) {
408 addIncorporatedExtension(spv::E_SPV_KHR_8bit_storage, spv::Spv_1_5);
409 addCapability(spv::CapabilityStorageBuffer8BitAccess);
411 if (containsType(type->getIdOperand(1), OpTypeInt, 16) ||
412 containsType(type->getIdOperand(1), OpTypeFloat, 16)) {
413 addIncorporatedExtension(spv::E_SPV_KHR_16bit_storage, spv::Spv_1_3);
414 addCapability(spv::CapabilityStorageBuffer16BitAccess);
419 // process all block-contained instructions
420 for (auto fi = module.getFunctions().cbegin(); fi != module.getFunctions().cend(); fi++) {
422 for (auto bi = f->getBlocks().cbegin(); bi != f->getBlocks().cend(); bi++) {
424 for (auto ii = b->getInstructions().cbegin(); ii != b->getInstructions().cend(); ii++)
425 postProcess(*ii->get());
427 // For all local variables that contain pointers to PhysicalStorageBufferEXT, check whether
428 // there is an existing restrict/aliased decoration. If we don't find one, add Aliased as the
430 for (auto vi = b->getLocalVariables().cbegin(); vi != b->getLocalVariables().cend(); vi++) {
431 const Instruction& inst = *vi->get();
432 Id resultId = inst.getResultId();
433 if (containsPhysicalStorageBufferOrArray(getDerefTypeId(resultId))) {
434 bool foundDecoration = false;
435 const auto function = [&](const std::unique_ptr<Instruction>& decoration) {
436 if (decoration.get()->getIdOperand(0) == resultId &&
437 decoration.get()->getOpCode() == OpDecorate &&
438 (decoration.get()->getImmediateOperand(1) == spv::DecorationAliasedPointerEXT ||
439 decoration.get()->getImmediateOperand(1) == spv::DecorationRestrictPointerEXT)) {
440 foundDecoration = true;
443 std::for_each(decorations.begin(), decorations.end(), function);
444 if (!foundDecoration) {
445 addDecoration(resultId, spv::DecorationAliasedPointerEXT);
452 // If any Vulkan memory model-specific functionality is used, update the
453 // OpMemoryModel to match.
454 if (capabilities.find(spv::CapabilityVulkanMemoryModelKHR) != capabilities.end()) {
455 memoryModel = spv::MemoryModelVulkanKHR;
456 addIncorporatedExtension(spv::E_SPV_KHR_vulkan_memory_model, spv::Spv_1_5);
459 // Add Aliased decoration if there's more than one Workgroup Block variable.
460 if (capabilities.find(spv::CapabilityWorkgroupMemoryExplicitLayoutKHR) != capabilities.end()) {
461 assert(entryPoints.size() == 1);
462 auto &ep = entryPoints[0];
464 std::vector<Id> workgroup_variables;
465 for (int i = 0; i < (int)ep->getNumOperands(); i++) {
466 if (!ep->isIdOperand(i))
469 const Id id = ep->getIdOperand(i);
470 const Instruction *instr = module.getInstruction(id);
471 if (instr->getOpCode() != spv::OpVariable)
474 if (instr->getImmediateOperand(0) == spv::StorageClassWorkgroup)
475 workgroup_variables.push_back(id);
478 if (workgroup_variables.size() > 1) {
479 for (size_t i = 0; i < workgroup_variables.size(); i++)
480 addDecoration(workgroup_variables[i], spv::DecorationAliased);
486 void Builder::postProcess(bool compileOnly)
488 // postProcessCFG needs an entrypoint to determine what is reachable, but if we are not creating an "executable" shader, we don't have an entrypoint
492 postProcessFeatures();
495 }; // end spv namespace