| 1 |
// Copyright (c) 2018 The Khronos Group Inc. |
| 2 |
// Copyright (c) 2018 Valve Corporation |
| 3 |
// Copyright (c) 2018 LunarG Inc. |
| 4 |
// |
| 5 |
// Licensed under the Apache License, Version 2.0 (the "License"); |
| 6 |
// you may not use this file except in compliance with the License. |
| 7 |
// You may obtain a copy of the License at |
| 8 |
// |
| 9 |
// http://www.apache.org/licenses/LICENSE-2.0 |
| 10 |
// |
| 11 |
// Unless required by applicable law or agreed to in writing, software |
| 12 |
// distributed under the License is distributed on an "AS IS" BASIS, |
| 13 |
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 14 |
// See the License for the specific language governing permissions and |
| 15 |
// limitations under the License. |
| 16 |
|
| 17 |
#ifndef LIBSPIRV_OPT_INSTRUMENT_PASS_H_ |
| 18 |
#define LIBSPIRV_OPT_INSTRUMENT_PASS_H_ |
| 19 |
|
| 20 |
#include <list> |
| 21 |
#include <memory> |
| 22 |
#include <vector> |
| 23 |
|
| 24 |
#include "source/opt/ir_builder.h" |
| 25 |
#include "source/opt/pass.h" |
| 26 |
#include "spirv-tools/instrument.hpp" |
| 27 |
|
| 28 |
// This is a base class to assist in the creation of passes which instrument |
| 29 |
// shader modules. More specifically, passes which replace instructions with a |
| 30 |
// larger and more capable set of instructions. Commonly, these new |
| 31 |
// instructions will add testing of operands and execute different |
| 32 |
// instructions depending on the outcome, including outputting of debug |
| 33 |
// information into a buffer created especially for that purpose. |
| 34 |
// |
| 35 |
// This class contains helper functions to create an InstProcessFunction, |
| 36 |
// which is the heart of any derived class implementing a specific |
| 37 |
// instrumentation pass. It takes an instruction as an argument, decides |
| 38 |
// if it should be instrumented, and generates code to replace it. This class |
| 39 |
// also supplies function InstProcessEntryPointCallTree which applies the |
| 40 |
// InstProcessFunction to every reachable instruction in a module and replaces |
| 41 |
// the instruction with new instructions if generated. |
| 42 |
// |
| 43 |
// Chief among the helper functions are output code generation functions, |
| 44 |
// used to generate code in the shader which writes data to output buffers |
| 45 |
// associated with that validation. Currently one such function, |
| 46 |
// GenDebugStreamWrite, exists. Other such functions may be added in the |
| 47 |
// future. Each is accompanied by documentation describing the format of |
| 48 |
// its output buffer. |
| 49 |
// |
| 50 |
// A validation pass may read or write multiple buffers. All such buffers |
| 51 |
// are located in a single debug descriptor set whose index is passed at the |
| 52 |
// creation of the instrumentation pass. The bindings of the buffers used by |
| 53 |
// a validation pass are permanantly assigned and fixed and documented by |
| 54 |
// the kDebugOutput* static consts. |
| 55 |
|
| 56 |
namespace spvtools { |
| 57 |
namespace opt { |
| 58 |
|
| 59 |
// Validation Ids |
| 60 |
// These are used to identify the general validation being done and map to |
| 61 |
// its output buffers. |
| 62 |
static const uint32_t kInstValidationIdBindless = 0; |
| 63 |
static const uint32_t kInstValidationIdBuffAddr = 1; |
| 64 |
static const uint32_t kInstValidationIdDebugPrintf = 2; |
| 65 |
|
| 66 |
class InstrumentPass : public Pass { |
| 67 |
using cbb_ptr = const BasicBlock*; |
| 68 |
|
| 69 |
public: |
| 70 |
using InstProcessFunction = |
| 71 |
std::function<void(BasicBlock::iterator, UptrVectorIterator<BasicBlock>, |
| 72 |
uint32_t, std::vector<std::unique_ptr<BasicBlock>>*)>; |
| 73 |
|
| 74 |
~InstrumentPass() override = default; |
| 75 |
|
| 76 |
IRContext::Analysis GetPreservedAnalyses() override { |
| 77 |
return IRContext::kAnalysisDefUse | IRContext::kAnalysisDecorations | |
| 78 |
IRContext::kAnalysisCombinators | IRContext::kAnalysisNameMap | |
| 79 |
IRContext::kAnalysisBuiltinVarId | IRContext::kAnalysisConstants; |
| 80 |
} |
| 81 |
|
| 82 |
protected: |
| 83 |
// Create instrumentation pass for |validation_id| which utilizes descriptor |
| 84 |
// set |desc_set| for debug input and output buffers and writes |shader_id| |
| 85 |
// into debug output records. |opt_direct_reads| indicates that the pass |
| 86 |
// will see direct input buffer reads and should prepare to optimize them. |
| 87 |
InstrumentPass(uint32_t desc_set, uint32_t shader_id, uint32_t validation_id, |
| 88 |
bool opt_direct_reads = false) |
| 89 |
: Pass(), |
| 90 |
desc_set_(desc_set), |
| 91 |
shader_id_(shader_id), |
| 92 |
validation_id_(validation_id), |
| 93 |
opt_direct_reads_(opt_direct_reads) {} |
| 94 |
|
| 95 |
// Initialize state for instrumentation of module. |
| 96 |
void InitializeInstrument(); |
| 97 |
|
| 98 |
// Call |pfn| on all instructions in all functions in the call tree of the |
| 99 |
// entry points in |module|. If code is generated for an instruction, replace |
| 100 |
// the instruction's block with the new blocks that are generated. Continue |
| 101 |
// processing at the top of the last new block. |
| 102 |
bool InstProcessEntryPointCallTree(InstProcessFunction& pfn); |
| 103 |
|
| 104 |
// Move all code in |ref_block_itr| preceding the instruction |ref_inst_itr| |
| 105 |
// to be instrumented into block |new_blk_ptr|. |
| 106 |
void MovePreludeCode(BasicBlock::iterator ref_inst_itr, |
| 107 |
UptrVectorIterator<BasicBlock> ref_block_itr, |
| 108 |
std::unique_ptr<BasicBlock>* new_blk_ptr); |
| 109 |
|
| 110 |
// Move all code in |ref_block_itr| succeeding the instruction |ref_inst_itr| |
| 111 |
// to be instrumented into block |new_blk_ptr|. |
| 112 |
void MovePostludeCode(UptrVectorIterator<BasicBlock> ref_block_itr, |
| 113 |
BasicBlock* new_blk_ptr); |
| 114 |
|
| 115 |
// Generate instructions in |builder| which will atomically fetch and |
| 116 |
// increment the size of the debug output buffer stream of the current |
| 117 |
// validation and write a record to the end of the stream, if enough space |
| 118 |
// in the buffer remains. The record will contain the index of the function |
| 119 |
// and instruction within that function |func_idx, instruction_idx| which |
| 120 |
// generated the record. It will also contain additional information to |
| 121 |
// identify the instance of the shader, depending on the stage |stage_idx| |
| 122 |
// of the shader. Finally, the record will contain validation-specific |
| 123 |
// data contained in |validation_ids| which will identify the validation |
| 124 |
// error as well as the values involved in the error. |
| 125 |
// |
| 126 |
// The output buffer binding written to by the code generated by the function |
| 127 |
// is determined by the validation id specified when each specific |
| 128 |
// instrumentation pass is created. |
| 129 |
// |
| 130 |
// The output buffer is a sequence of 32-bit values with the following |
| 131 |
// format (where all elements are unsigned 32-bit unless otherwise noted): |
| 132 |
// |
| 133 |
// Size |
| 134 |
// Record0 |
| 135 |
// Record1 |
| 136 |
// Record2 |
| 137 |
// ... |
| 138 |
// |
| 139 |
// Size is the number of 32-bit values that have been written or |
| 140 |
// attempted to be written to the output buffer, excluding the Size. It is |
| 141 |
// initialized to 0. If the size of attempts to write the buffer exceeds |
| 142 |
// the actual size of the buffer, it is possible that this field can exceed |
| 143 |
// the actual size of the buffer. |
| 144 |
// |
| 145 |
// Each Record* is a variable-length sequence of 32-bit values with the |
| 146 |
// following format defined using static const offsets in the .cpp file: |
| 147 |
// |
| 148 |
// Record Size |
| 149 |
// Shader ID |
| 150 |
// Instruction Index |
| 151 |
// Stage |
| 152 |
// Stage-specific Word 0 |
| 153 |
// Stage-specific Word 1 |
| 154 |
// ... |
| 155 |
// Validation Error Code |
| 156 |
// Validation-specific Word 0 |
| 157 |
// Validation-specific Word 1 |
| 158 |
// Validation-specific Word 2 |
| 159 |
// ... |
| 160 |
// |
| 161 |
// Each record consists of three subsections: members common across all |
| 162 |
// validation, members specific to the stage, and members specific to a |
| 163 |
// validation. |
| 164 |
// |
| 165 |
// The Record Size is the number of 32-bit words in the record, including |
| 166 |
// the Record Size word. |
| 167 |
// |
| 168 |
// Shader ID is a value that identifies which shader has generated the |
| 169 |
// validation error. It is passed when the instrumentation pass is created. |
| 170 |
// |
| 171 |
// The Instruction Index is the position of the instruction within the |
| 172 |
// SPIR-V file which is in error. |
| 173 |
// |
| 174 |
// The Stage is the pipeline stage which has generated the error as defined |
| 175 |
// by the SpvExecutionModel_ enumeration. This is used to interpret the |
| 176 |
// following Stage-specific words. |
| 177 |
// |
| 178 |
// The Stage-specific Words identify which invocation of the shader generated |
| 179 |
// the error. Every stage will write a fixed number of words. Vertex shaders |
| 180 |
// will write the Vertex and Instance ID. Fragment shaders will write |
| 181 |
// FragCoord.xy. Compute shaders will write the GlobalInvocation ID. |
| 182 |
// The tesselation eval shader will write the Primitive ID and TessCoords.uv. |
| 183 |
// The tesselation control shader and geometry shader will write the |
| 184 |
// Primitive ID and Invocation ID. |
| 185 |
// |
| 186 |
// The Validation Error Code specifies the exact error which has occurred. |
| 187 |
// These are enumerated with the kInstError* static consts. This allows |
| 188 |
// multiple validation layers to use the same, single output buffer. |
| 189 |
// |
| 190 |
// The Validation-specific Words are a validation-specific number of 32-bit |
| 191 |
// words which give further information on the validation error that |
| 192 |
// occurred. These are documented further in each file containing the |
| 193 |
// validation-specific class which derives from this base class. |
| 194 |
// |
| 195 |
// Because the code that is generated checks against the size of the buffer |
| 196 |
// before writing, the size of the debug out buffer can be used by the |
| 197 |
// validation layer to control the number of error records that are written. |
| 198 |
void GenDebugStreamWrite(uint32_t instruction_idx, uint32_t stage_idx, |
| 199 |
const std::vector<uint32_t>& validation_ids, |
| 200 |
InstructionBuilder* builder); |
| 201 |
|
| 202 |
// Return true if all instructions in |ids| are constants or spec constants. |
| 203 |
bool AllConstant(const std::vector<uint32_t>& ids); |
| 204 |
|
| 205 |
// Generate in |builder| instructions to read the unsigned integer from the |
| 206 |
// input buffer specified by the offsets in |offset_ids|. Given offsets |
| 207 |
// o0, o1, ... oN, and input buffer ibuf, return the id for the value: |
| 208 |
// |
| 209 |
// ibuf[...ibuf[ibuf[o0]+o1]...+oN] |
| 210 |
// |
| 211 |
// The binding and the format of the input buffer is determined by each |
| 212 |
// specific validation, which is specified at the creation of the pass. |
| 213 |
uint32_t GenDebugDirectRead(const std::vector<uint32_t>& offset_ids, |
| 214 |
InstructionBuilder* builder); |
| 215 |
|
| 216 |
// Generate code to convert integer |value_id| to 32bit, if needed. Return |
| 217 |
// an id to the 32bit equivalent. |
| 218 |
uint32_t Gen32BitCvtCode(uint32_t value_id, InstructionBuilder* builder); |
| 219 |
|
| 220 |
// Generate code to cast integer |value_id| to 32bit unsigned, if needed. |
| 221 |
// Return an id to the Uint equivalent. |
| 222 |
uint32_t GenUintCastCode(uint32_t value_id, InstructionBuilder* builder); |
| 223 |
|
| 224 |
// Return new label. |
| 225 |
std::unique_ptr<Instruction> NewLabel(uint32_t label_id); |
| 226 |
|
| 227 |
// Return id for 32-bit unsigned type |
| 228 |
uint32_t GetUintId(); |
| 229 |
|
| 230 |
// Return id for 64-bit unsigned type |
| 231 |
uint32_t GetUint64Id(); |
| 232 |
|
| 233 |
// Return id for 8-bit unsigned type |
| 234 |
uint32_t GetUint8Id(); |
| 235 |
|
| 236 |
// Return id for 32-bit unsigned type |
| 237 |
uint32_t GetBoolId(); |
| 238 |
|
| 239 |
// Return id for void type |
| 240 |
uint32_t GetVoidId(); |
| 241 |
|
| 242 |
// Return pointer to type for runtime array of uint |
| 243 |
analysis::Type* GetUintXRuntimeArrayType(uint32_t width, |
| 244 |
analysis::Type** rarr_ty); |
| 245 |
|
| 246 |
// Return pointer to type for runtime array of uint |
| 247 |
analysis::Type* GetUintRuntimeArrayType(uint32_t width); |
| 248 |
|
| 249 |
// Return id for buffer uint type |
| 250 |
uint32_t GetOutputBufferPtrId(); |
| 251 |
|
| 252 |
// Return id for buffer uint type |
| 253 |
uint32_t GetInputBufferTypeId(); |
| 254 |
|
| 255 |
// Return id for buffer uint type |
| 256 |
uint32_t GetInputBufferPtrId(); |
| 257 |
|
| 258 |
// Return binding for output buffer for current validation. |
| 259 |
uint32_t GetOutputBufferBinding(); |
| 260 |
|
| 261 |
// Return binding for input buffer for current validation. |
| 262 |
uint32_t GetInputBufferBinding(); |
| 263 |
|
| 264 |
// Add storage buffer extension if needed |
| 265 |
void AddStorageBufferExt(); |
| 266 |
|
| 267 |
// Return id for debug output buffer |
| 268 |
uint32_t GetOutputBufferId(); |
| 269 |
|
| 270 |
// Return id for debug input buffer |
| 271 |
uint32_t GetInputBufferId(); |
| 272 |
|
| 273 |
// Return id for 32-bit float type |
| 274 |
uint32_t GetFloatId(); |
| 275 |
|
| 276 |
// Return id for v4float type |
| 277 |
uint32_t GetVec4FloatId(); |
| 278 |
|
| 279 |
// Return id for uint vector type of |length| |
| 280 |
uint32_t GetVecUintId(uint32_t length); |
| 281 |
|
| 282 |
// Return id for v4uint type |
| 283 |
uint32_t GetVec4UintId(); |
| 284 |
|
| 285 |
// Return id for v3uint type |
| 286 |
uint32_t GetVec3UintId(); |
| 287 |
|
| 288 |
// Return id for output function. Define if it doesn't exist with |
| 289 |
// |val_spec_param_cnt| validation-specific uint32 parameters. |
| 290 |
uint32_t GetStreamWriteFunctionId(uint32_t stage_idx, |
| 291 |
uint32_t val_spec_param_cnt); |
| 292 |
|
| 293 |
// Return id for input function taking |param_cnt| uint32 parameters. Define |
| 294 |
// if it doesn't exist. |
| 295 |
uint32_t GetDirectReadFunctionId(uint32_t param_cnt); |
| 296 |
|
| 297 |
// Split block |block_itr| into two new blocks where the second block |
| 298 |
// contains |inst_itr| and place in |new_blocks|. |
| 299 |
void SplitBlock(BasicBlock::iterator inst_itr, |
| 300 |
UptrVectorIterator<BasicBlock> block_itr, |
| 301 |
std::vector<std::unique_ptr<BasicBlock>>* new_blocks); |
| 302 |
|
| 303 |
// Apply instrumentation function |pfn| to every instruction in |func|. |
| 304 |
// If code is generated for an instruction, replace the instruction's |
| 305 |
// block with the new blocks that are generated. Continue processing at the |
| 306 |
// top of the last new block. |
| 307 |
bool InstrumentFunction(Function* func, uint32_t stage_idx, |
| 308 |
InstProcessFunction& pfn); |
| 309 |
|
| 310 |
// Call |pfn| on all functions in the call tree of the function |
| 311 |
// ids in |roots|. |
| 312 |
bool InstProcessCallTreeFromRoots(InstProcessFunction& pfn, |
| 313 |
std::queue<uint32_t>* roots, |
| 314 |
uint32_t stage_idx); |
| 315 |
|
| 316 |
// Gen code into |builder| to write |field_value_id| into debug output |
| 317 |
// buffer at |base_offset_id| + |field_offset|. |
| 318 |
void GenDebugOutputFieldCode(uint32_t base_offset_id, uint32_t field_offset, |
| 319 |
uint32_t field_value_id, |
| 320 |
InstructionBuilder* builder); |
| 321 |
|
| 322 |
// Generate instructions into |builder| which will write the members |
| 323 |
// of the debug output record common for all stages and validations at |
| 324 |
// |base_off|. |
| 325 |
void GenCommonStreamWriteCode(uint32_t record_sz, uint32_t instruction_idx, |
| 326 |
uint32_t stage_idx, uint32_t base_off, |
| 327 |
InstructionBuilder* builder); |
| 328 |
|
| 329 |
// Generate instructions into |builder| which will write |
| 330 |
// |uint_frag_coord_id| at |component| of the record at |base_offset_id| of |
| 331 |
// the debug output buffer . |
| 332 |
void GenFragCoordEltDebugOutputCode(uint32_t base_offset_id, |
| 333 |
uint32_t uint_frag_coord_id, |
| 334 |
uint32_t component, |
| 335 |
InstructionBuilder* builder); |
| 336 |
|
| 337 |
// Generate instructions into |builder| which will load |var_id| and return |
| 338 |
// its result id. |
| 339 |
uint32_t GenVarLoad(uint32_t var_id, InstructionBuilder* builder); |
| 340 |
|
| 341 |
// Generate instructions into |builder| which will load the uint |builtin_id| |
| 342 |
// and write it into the debug output buffer at |base_off| + |builtin_off|. |
| 343 |
void GenBuiltinOutputCode(uint32_t builtin_id, uint32_t builtin_off, |
| 344 |
uint32_t base_off, InstructionBuilder* builder); |
| 345 |
|
| 346 |
// Generate instructions into |builder| which will write the |stage_idx|- |
| 347 |
// specific members of the debug output stream at |base_off|. |
| 348 |
void GenStageStreamWriteCode(uint32_t stage_idx, uint32_t base_off, |
| 349 |
InstructionBuilder* builder); |
| 350 |
|
| 351 |
// Return true if instruction must be in the same block that its result |
| 352 |
// is used. |
| 353 |
bool IsSameBlockOp(const Instruction* inst) const; |
| 354 |
|
| 355 |
// Clone operands which must be in same block as consumer instructions. |
| 356 |
// Look in same_blk_pre for instructions that need cloning. Look in |
| 357 |
// same_blk_post for instructions already cloned. Add cloned instruction |
| 358 |
// to same_blk_post. |
| 359 |
void CloneSameBlockOps( |
| 360 |
std::unique_ptr<Instruction>* inst, |
| 361 |
std::unordered_map<uint32_t, uint32_t>* same_blk_post, |
| 362 |
std::unordered_map<uint32_t, Instruction*>* same_blk_pre, |
| 363 |
BasicBlock* block_ptr); |
| 364 |
|
| 365 |
// Update phis in succeeding blocks to point to new last block |
| 366 |
void UpdateSucceedingPhis( |
| 367 |
std::vector<std::unique_ptr<BasicBlock>>& new_blocks); |
| 368 |
|
| 369 |
// Debug descriptor set index |
| 370 |
uint32_t desc_set_; |
| 371 |
|
| 372 |
// Shader module ID written into output record |
| 373 |
uint32_t shader_id_; |
| 374 |
|
| 375 |
// Map from function id to function pointer. |
| 376 |
std::unordered_map<uint32_t, Function*> id2function_; |
| 377 |
|
| 378 |
// Map from block's label id to block. TODO(dnovillo): This is superfluous wrt |
| 379 |
// CFG. It has functionality not present in CFG. Consolidate. |
| 380 |
std::unordered_map<uint32_t, BasicBlock*> id2block_; |
| 381 |
|
| 382 |
// Map from instruction's unique id to offset in original file. |
| 383 |
std::unordered_map<uint32_t, uint32_t> uid2offset_; |
| 384 |
|
| 385 |
// result id for OpConstantFalse |
| 386 |
uint32_t validation_id_; |
| 387 |
|
| 388 |
// id for output buffer variable |
| 389 |
uint32_t output_buffer_id_; |
| 390 |
|
| 391 |
// ptr type id for output buffer element |
| 392 |
uint32_t output_buffer_ptr_id_; |
| 393 |
|
| 394 |
// ptr type id for input buffer element |
| 395 |
uint32_t input_buffer_ptr_id_; |
| 396 |
|
| 397 |
// id for debug output function |
| 398 |
std::unordered_map<uint32_t, uint32_t> param2output_func_id_; |
| 399 |
|
| 400 |
// ids for debug input functions |
| 401 |
std::unordered_map<uint32_t, uint32_t> param2input_func_id_; |
| 402 |
|
| 403 |
// id for input buffer variable |
| 404 |
uint32_t input_buffer_id_; |
| 405 |
|
| 406 |
// id for 32-bit float type |
| 407 |
uint32_t float_id_; |
| 408 |
|
| 409 |
// id for v4float type |
| 410 |
uint32_t v4float_id_; |
| 411 |
|
| 412 |
// id for v4uint type |
| 413 |
uint32_t v4uint_id_; |
| 414 |
|
| 415 |
// id for v3uint type |
| 416 |
uint32_t v3uint_id_; |
| 417 |
|
| 418 |
// id for 32-bit unsigned type |
| 419 |
uint32_t uint_id_; |
| 420 |
|
| 421 |
// id for 64-bit unsigned type |
| 422 |
uint32_t uint64_id_; |
| 423 |
|
| 424 |
// id for 8-bit unsigned type |
| 425 |
uint32_t uint8_id_; |
| 426 |
|
| 427 |
// id for bool type |
| 428 |
uint32_t bool_id_; |
| 429 |
|
| 430 |
// id for void type |
| 431 |
uint32_t void_id_; |
| 432 |
|
| 433 |
// boolean to remember storage buffer extension |
| 434 |
bool storage_buffer_ext_defined_; |
| 435 |
|
| 436 |
// runtime array of uint type |
| 437 |
analysis::Type* uint64_rarr_ty_; |
| 438 |
|
| 439 |
// runtime array of uint type |
| 440 |
analysis::Type* uint32_rarr_ty_; |
| 441 |
|
| 442 |
// Pre-instrumentation same-block insts |
| 443 |
std::unordered_map<uint32_t, Instruction*> same_block_pre_; |
| 444 |
|
| 445 |
// Post-instrumentation same-block op ids |
| 446 |
std::unordered_map<uint32_t, uint32_t> same_block_post_; |
| 447 |
|
| 448 |
// Map function calls to result id. Clear for every function. |
| 449 |
// This is for debug input reads with constant arguments that |
| 450 |
// have been generated into the first block of the function. |
| 451 |
// This mechanism is used to avoid multiple identical debug |
| 452 |
// input buffer reads. |
| 453 |
struct vector_hash_ { |
| 454 |
std::size_t operator()(const std::vector<uint32_t>& v) const { |
| 455 |
std::size_t hash = v.size(); |
| 456 |
for (auto& u : v) { |
| 457 |
hash ^= u + 0x9e3779b9 + (hash << 11) + (hash >> 21); |
| 458 |
} |
| 459 |
return hash; |
| 460 |
} |
| 461 |
}; |
| 462 |
std::unordered_map<std::vector<uint32_t>, uint32_t, vector_hash_> call2id_; |
| 463 |
|
| 464 |
// Function currently being instrumented |
| 465 |
Function* curr_func_; |
| 466 |
|
| 467 |
// Optimize direct debug input buffer reads. Specifically, move all such |
| 468 |
// reads with constant args to first block and reuse them. |
| 469 |
bool opt_direct_reads_; |
| 470 |
}; |
| 471 |
|
| 472 |
} // namespace opt |
| 473 |
} // namespace spvtools |
| 474 |
|
| 475 |
#endif // LIBSPIRV_OPT_INSTRUMENT_PASS_H_ |