AuroraMiddleware/glslang
1
0
Fork 0
mirror of https://github.com/KhronosGroup/glslang synced 2024-11-09 20:10:06 +00:00
Code Issues Packages Projects Releases Wiki Activity
ea1ea974eb
glslang/SPIRV/SPVRemapper.cpp

1351 lines
49 KiB
C++
Executable File
Raw Blame History

//
// Copyright (C) 2015 LunarG, Inc.
//
// All rights reserved.
//
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions
// are met:
//
// Redistributions of source code must retain the above copyright
// notice, this list of conditions and the following disclaimer.
//
// Redistributions in binary form must reproduce the above
// copyright notice, this list of conditions and the following
// disclaimer in the documentation and/or other materials provided
// with the distribution.
//
// Neither the name of 3Dlabs Inc. Ltd. nor the names of its
// contributors may be used to endorse or promote products derived
// from this software without specific prior written permission.
//
// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
// FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
// COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
// INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
// BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
// LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
// CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
// LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
// ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
// POSSIBILITY OF SUCH DAMAGE.
//
#include "SPVRemapper.h"
#include "doc.h"
#if !defined (use_cpp11)
// ... not supported before C++11
#else // defined (use_cpp11)
#include <algorithm>
#include <cassert>
#include "../glslang/Include/Common.h"
namespace spv {
// By default, just abort on error. Can be overridden via RegisterErrorHandler
spirvbin_t::errorfn_t spirvbin_t::errorHandler = [](const std::string&) { exit(5); };
// By default, eat log messages. Can be overridden via RegisterLogHandler
spirvbin_t::logfn_t spirvbin_t::logHandler = [](const std::string&) { };
// This can be overridden to provide other message behavior if needed
void spirvbin_t::msg(int minVerbosity, int indent, const std::string& txt) const
{
if (verbose >= minVerbosity)
logHandler(std::string(indent, ' ') + txt);
}
// hash opcode, with special handling for OpExtInst
std::uint32_t spirvbin_t::asOpCodeHash(unsigned word)
{
const spv::Op opCode = asOpCode(word);
std::uint32_t offset = 0;
switch (opCode) {
case spv::OpExtInst:
offset += asId(word + 4); break;
default:
break;
}
return opCode * 19 + offset; // 19 = small prime
}
spirvbin_t::range_t spirvbin_t::literalRange(spv::Op opCode) const
{
static const int maxCount = 1<<30;
switch (opCode) {
case spv::OpTypeFloat: // fall through...
case spv::OpTypePointer: return range_t(2, 3);
case spv::OpTypeInt: return range_t(2, 4);
// TODO: case spv::OpTypeImage:
// TODO: case spv::OpTypeSampledImage:
case spv::OpTypeSampler: return range_t(3, 8);
case spv::OpTypeVector: // fall through
case spv::OpTypeMatrix: // ...
case spv::OpTypePipe: return range_t(3, 4);
case spv::OpConstant: return range_t(3, maxCount);
default: return range_t(0, 0);
}
}
spirvbin_t::range_t spirvbin_t::typeRange(spv::Op opCode) const
{
static const int maxCount = 1<<30;
if (isConstOp(opCode))
return range_t(1, 2);
switch (opCode) {
case spv::OpTypeVector: // fall through
case spv::OpTypeMatrix: // ...
case spv::OpTypeSampler: // ...
case spv::OpTypeArray: // ...
case spv::OpTypeRuntimeArray: // ...
case spv::OpTypePipe: return range_t(2, 3);
case spv::OpTypeStruct: // fall through
case spv::OpTypeFunction: return range_t(2, maxCount);
case spv::OpTypePointer: return range_t(3, 4);
default: return range_t(0, 0);
}
}
spirvbin_t::range_t spirvbin_t::constRange(spv::Op opCode) const
{
static const int maxCount = 1<<30;
switch (opCode) {
case spv::OpTypeArray: // fall through...
case spv::OpTypeRuntimeArray: return range_t(3, 4);
case spv::OpConstantComposite: return range_t(3, maxCount);
default: return range_t(0, 0);
}
}
// Return the size of a type in 32-bit words. This currently only
// handles ints and floats, and is only invoked by queries which must be
// integer types. If ever needed, it can be generalized.
unsigned spirvbin_t::typeSizeInWords(spv::Id id) const
{
const unsigned typeStart = idPos(id);
const spv::Op opCode = asOpCode(typeStart);
switch (opCode) {
case spv::OpTypeInt: // fall through...
case spv::OpTypeFloat: return (spv[typeStart+2]+31)/32;
default:
return 0;
}
}
// Looks up the type of a given const or variable ID, and
// returns its size in 32-bit words.
unsigned spirvbin_t::idTypeSizeInWords(spv::Id id) const
{
const auto tid_it = idTypeSizeMap.find(id);
if (tid_it == idTypeSizeMap.end())
error("type size for ID not found");
return tid_it->second;
}
// Is this an opcode we should remove when using --strip?
bool spirvbin_t::isStripOp(spv::Op opCode) const
{
switch (opCode) {
case spv::OpSource:
case spv::OpSourceExtension:
case spv::OpName:
case spv::OpMemberName:
case spv::OpLine: return true;
default: return false;
}
}
// Return true if this opcode is flow control
bool spirvbin_t::isFlowCtrl(spv::Op opCode) const
{
switch (opCode) {
case spv::OpBranchConditional:
case spv::OpBranch:
case spv::OpSwitch:
case spv::OpLoopMerge:
case spv::OpSelectionMerge:
case spv::OpLabel:
case spv::OpFunction:
case spv::OpFunctionEnd: return true;
default: return false;
}
}
// Return true if this opcode defines a type
bool spirvbin_t::isTypeOp(spv::Op opCode) const
{
switch (opCode) {
case spv::OpTypeVoid:
case spv::OpTypeBool:
case spv::OpTypeInt:
case spv::OpTypeFloat:
case spv::OpTypeVector:
case spv::OpTypeMatrix:
case spv::OpTypeImage:
case spv::OpTypeSampler:
case spv::OpTypeArray:
case spv::OpTypeRuntimeArray:
case spv::OpTypeStruct:
case spv::OpTypeOpaque:
case spv::OpTypePointer:
case spv::OpTypeFunction:
case spv::OpTypeEvent:
case spv::OpTypeDeviceEvent:
case spv::OpTypeReserveId:
case spv::OpTypeQueue:
case spv::OpTypeSampledImage:
case spv::OpTypePipe: return true;
default: return false;
}
}
// Return true if this opcode defines a constant
bool spirvbin_t::isConstOp(spv::Op opCode) const
{
switch (opCode) {
case spv::OpConstantNull:
case spv::OpConstantSampler:
error("unimplemented constant type");
return true;
case spv::OpConstantTrue:
case spv::OpConstantFalse:
case spv::OpConstantComposite:
case spv::OpConstant:
return true;
default:
return false;
}
}
const auto inst_fn_nop = [](spv::Op, unsigned) { return false; };
const auto op_fn_nop = [](spv::Id&) { };
// g++ doesn't like these defined in the class proper in an anonymous namespace.
// Dunno why. Also MSVC doesn't like the constexpr keyword. Also dunno why.
// Defining them externally seems to please both compilers, so, here they are.
const spv::Id spirvbin_t::unmapped = spv::Id(-10000);
const spv::Id spirvbin_t::unused = spv::Id(-10001);
const int spirvbin_t::header_size = 5;
spv::Id spirvbin_t::nextUnusedId(spv::Id id)
{
while (isNewIdMapped(id)) // search for an unused ID
++id;
return id;
}
spv::Id spirvbin_t::localId(spv::Id id, spv::Id newId)
{
assert(id != spv::NoResult && newId != spv::NoResult);
if (id >= idMapL.size())
idMapL.resize(id+1, unused);
if (newId != unmapped && newId != unused) {
if (isOldIdUnused(id))
error(std::string("ID unused in module: ") + std::to_string(id));
if (!isOldIdUnmapped(id))
error(std::string("ID already mapped: ") + std::to_string(id) + " -> "
+ std::to_string(localId(id)));
if (isNewIdMapped(newId))
error(std::string("ID already used in module: ") + std::to_string(newId));
msg(4, 4, std::string("map: ") + std::to_string(id) + " -> " + std::to_string(newId));
setMapped(newId);
largestNewId = std::max(largestNewId, newId);
}
return idMapL[id] = newId;
}
// Parse a literal string from the SPIR binary and return it as an std::string
// Due to C++11 RValue references, this doesn't copy the result string.
std::string spirvbin_t::literalString(unsigned word) const
{
std::string literal;
literal.reserve(16);
const char* bytes = reinterpret_cast<const char*>(spv.data() + word);
while (bytes && *bytes)
literal += *bytes++;
return literal;
}
void spirvbin_t::applyMap()
{
msg(3, 2, std::string("Applying map: "));
// Map local IDs through the ID map
process(inst_fn_nop, // ignore instructions
[this](spv::Id& id) {
id = localId(id);
assert(id != unused && id != unmapped);
}
);
}
// Find free IDs for anything we haven't mapped
void spirvbin_t::mapRemainder()
{
msg(3, 2, std::string("Remapping remainder: "));
spv::Id unusedId = 1; // can't use 0: that's NoResult
spirword_t maxBound = 0;
for (spv::Id id = 0; id < idMapL.size(); ++id) {
if (isOldIdUnused(id))
continue;
// Find a new mapping for any used but unmapped IDs
if (isOldIdUnmapped(id))
localId(id, unusedId = nextUnusedId(unusedId));
if (isOldIdUnmapped(id))
error(std::string("old ID not mapped: ") + std::to_string(id));
// Track max bound
maxBound = std::max(maxBound, localId(id) + 1);
}
bound(maxBound); // reset header ID bound to as big as it now needs to be
}
// Mark debug instructions for stripping
void spirvbin_t::stripDebug()
{
// Strip instructions in the stripOp set: debug info.
process(
[&](spv::Op opCode, unsigned start) {
// remember opcodes we want to strip later
if (isStripOp(opCode))
stripInst(start);
return true;
},
op_fn_nop);
}
// Mark instructions that refer to now-removed IDs for stripping
void spirvbin_t::stripDeadRefs()
{
process(
[&](spv::Op opCode, unsigned start) {
// strip opcodes pointing to removed data
switch (opCode) {
case spv::OpName:
case spv::OpMemberName:
case spv::OpDecorate:
case spv::OpMemberDecorate:
if (idPosR.find(asId(start+1)) == idPosR.end())
stripInst(start);
break;
default:
break; // leave it alone
}
return true;
},
op_fn_nop);
strip();
}
// Update local maps of ID, type, etc positions
void spirvbin_t::buildLocalMaps()
{
msg(2, 2, std::string("build local maps: "));
mapped.clear();
idMapL.clear();
// preserve nameMap, so we don't clear that.
fnPos.clear();
fnCalls.clear();
typeConstPos.clear();
idPosR.clear();
entryPoint = spv::NoResult;
largestNewId = 0;
idMapL.resize(bound(), unused);
int fnStart = 0;
spv::Id fnRes = spv::NoResult;
// build local Id and name maps
process(
[&](spv::Op opCode, unsigned start) {
unsigned word = start+1;
spv::Id typeId = spv::NoResult;
if (spv::InstructionDesc[opCode].hasType())
typeId = asId(word++);
// If there's a result ID, remember the size of its type
if (spv::InstructionDesc[opCode].hasResult()) {
const spv::Id resultId = asId(word++);
idPosR[resultId] = start;
if (typeId != spv::NoResult) {
const unsigned idTypeSize = typeSizeInWords(typeId);
if (idTypeSize != 0)
idTypeSizeMap[resultId] = idTypeSize;
}
}
if (opCode == spv::Op::OpName) {
const spv::Id target = asId(start+1);
const std::string name = literalString(start+2);
nameMap[name] = target;
} else if (opCode == spv::Op::OpFunctionCall) {
++fnCalls[asId(start + 3)];
} else if (opCode == spv::Op::OpEntryPoint) {
entryPoint = asId(start + 2);
} else if (opCode == spv::Op::OpFunction) {
if (fnStart != 0)
error("nested function found");
fnStart = start;
fnRes = asId(start + 2);
} else if (opCode == spv::Op::OpFunctionEnd) {
assert(fnRes != spv::NoResult);
if (fnStart == 0)
error("function end without function start");
fnPos[fnRes] = range_t(fnStart, start + asWordCount(start));
fnStart = 0;
} else if (isConstOp(opCode)) {
assert(asId(start + 2) != spv::NoResult);
typeConstPos.insert(start);
} else if (isTypeOp(opCode)) {
assert(asId(start + 1) != spv::NoResult);
typeConstPos.insert(start);
}
return false;
},
[this](spv::Id& id) { localId(id, unmapped); }
);
}
// Validate the SPIR header
void spirvbin_t::validate() const
{
msg(2, 2, std::string("validating: "));
if (spv.size() < header_size)
error("file too short: ");
if (magic() != spv::MagicNumber)
error("bad magic number");
// field 1 = version
// field 2 = generator magic
// field 3 = result <id> bound
if (schemaNum() != 0)
error("bad schema, must be 0");
}
int spirvbin_t::processInstruction(unsigned word, instfn_t instFn, idfn_t idFn)
{
const auto instructionStart = word;
const unsigned wordCount = asWordCount(instructionStart);
const int nextInst = word++ + wordCount;
spv::Op opCode = asOpCode(instructionStart);
if (nextInst > int(spv.size()))
error("spir instruction terminated too early");
// Base for computing number of operands; will be updated as more is learned
unsigned numOperands = wordCount - 1;
if (instFn(opCode, instructionStart))
return nextInst;
// Read type and result ID from instruction desc table
if (spv::InstructionDesc[opCode].hasType()) {
idFn(asId(word++));
--numOperands;
}
if (spv::InstructionDesc[opCode].hasResult()) {
idFn(asId(word++));
--numOperands;
}
// Extended instructions: currently, assume everything is an ID.
// TODO: add whatever data we need for exceptions to that
if (opCode == spv::OpExtInst) {
word += 2; // instruction set, and instruction from set
numOperands -= 2;
for (unsigned op=0; op < numOperands; ++op)
idFn(asId(word++)); // ID
return nextInst;
}
// Circular buffer so we can look back at previous unmapped values during the mapping pass.
static const unsigned idBufferSize = 4;
spv::Id idBuffer[idBufferSize];
unsigned idBufferPos = 0;
// Store IDs from instruction in our map
for (int op = 0; numOperands > 0; ++op, --numOperands) {
// SpecConstantOp is special: it includes the operands of another opcode which is
// given as a literal in the 3rd word. We will switch over to pretending that the
// opcode being processed is the literal opcode value of the SpecConstantOp. See the
// SPIRV spec for details. This way we will handle IDs and literals as appropriate for
// the embedded op.
if (opCode == spv::OpSpecConstantOp) {
if (op == 0) {
opCode = asOpCode(word++); // this is the opcode embedded in the SpecConstantOp.
--numOperands;
}
}
switch (spv::InstructionDesc[opCode].operands.getClass(op)) {
case spv::OperandId:
case spv::OperandScope:
case spv::OperandMemorySemantics:
idBuffer[idBufferPos] = asId(word);
idBufferPos = (idBufferPos + 1) % idBufferSize;
idFn(asId(word++));
break;
case spv::OperandVariableIds:
for (unsigned i = 0; i < numOperands; ++i)
idFn(asId(word++));
return nextInst;
case spv::OperandVariableLiterals:
// for clarity
// if (opCode == spv::OpDecorate && asDecoration(word - 1) == spv::DecorationBuiltIn) {
// ++word;
// --numOperands;
// }
// word += numOperands;
return nextInst;
case spv::OperandVariableLiteralId: {
if (opCode == OpSwitch) {
// word-2 is the position of the selector ID. OpSwitch Literals match its type.
// In case the IDs are currently being remapped, we get the word[-2] ID from
// the circular idBuffer.
const unsigned literalSizePos = (idBufferPos+idBufferSize-2) % idBufferSize;
const unsigned literalSize = idTypeSizeInWords(idBuffer[literalSizePos]);
const unsigned numLiteralIdPairs = (nextInst-word) / (1+literalSize);
for (unsigned arg=0; arg<numLiteralIdPairs; ++arg) {
word += literalSize; // literal
idFn(asId(word++)); // label
}
} else {
assert(0); // currentely, only OpSwitch uses OperandVariableLiteralId
}
return nextInst;
}
case spv::OperandLiteralString: {
const int stringWordCount = literalStringWords(literalString(word));
word += stringWordCount;
numOperands -= (stringWordCount-1); // -1 because for() header post-decrements
break;
}
// Execution mode might have extra literal operands. Skip them.
case spv::OperandExecutionMode:
return nextInst;
// Single word operands we simply ignore, as they hold no IDs
case spv::OperandLiteralNumber:
case spv::OperandSource:
case spv::OperandExecutionModel:
case spv::OperandAddressing:
case spv::OperandMemory:
case spv::OperandStorage:
case spv::OperandDimensionality:
case spv::OperandSamplerAddressingMode:
case spv::OperandSamplerFilterMode:
case spv::OperandSamplerImageFormat:
case spv::OperandImageChannelOrder:
case spv::OperandImageChannelDataType:
case spv::OperandImageOperands:
case spv::OperandFPFastMath:
case spv::OperandFPRoundingMode:
case spv::OperandLinkageType:
case spv::OperandAccessQualifier:
case spv::OperandFuncParamAttr:
case spv::OperandDecoration:
case spv::OperandBuiltIn:
case spv::OperandSelect:
case spv::OperandLoop:
case spv::OperandFunction:
case spv::OperandMemoryAccess:
case spv::OperandGroupOperation:
case spv::OperandKernelEnqueueFlags:
case spv::OperandKernelProfilingInfo:
case spv::OperandCapability:
++word;
break;
default:
assert(0 && "Unhandled Operand Class");
break;
}
}
return nextInst;
}
// Make a pass over all the instructions and process them given appropriate functions
spirvbin_t& spirvbin_t::process(instfn_t instFn, idfn_t idFn, unsigned begin, unsigned end)
{
// For efficiency, reserve name map space. It can grow if needed.
nameMap.reserve(32);
// If begin or end == 0, use defaults
begin = (begin == 0 ? header_size : begin);
end = (end == 0 ? unsigned(spv.size()) : end);
// basic parsing and InstructionDesc table borrowed from SpvDisassemble.cpp...
unsigned nextInst = unsigned(spv.size());
for (unsigned word = begin; word < end; word = nextInst)
nextInst = processInstruction(word, instFn, idFn);
return *this;
}
// Apply global name mapping to a single module
void spirvbin_t::mapNames()
{
static const std::uint32_t softTypeIdLimit = 3011; // small prime. TODO: get from options
static const std::uint32_t firstMappedID = 3019; // offset into ID space
for (const auto& name : nameMap) {
std::uint32_t hashval = 1911;
for (const char c : name.first)
hashval = hashval * 1009 + c;
if (isOldIdUnmapped(name.second))
localId(name.second, nextUnusedId(hashval % softTypeIdLimit + firstMappedID));
}
}
// Map fn contents to IDs of similar functions in other modules
void spirvbin_t::mapFnBodies()
{
static const std::uint32_t softTypeIdLimit = 19071; // small prime. TODO: get from options
static const std::uint32_t firstMappedID = 6203; // offset into ID space
// Initial approach: go through some high priority opcodes first and assign them
// hash values.
spv::Id fnId = spv::NoResult;
std::vector<unsigned> instPos;
instPos.reserve(unsigned(spv.size()) / 16); // initial estimate; can grow if needed.
// Build local table of instruction start positions
process(
[&](spv::Op, unsigned start) { instPos.push_back(start); return true; },
op_fn_nop);
// Window size for context-sensitive canonicalization values
// Empirical best size from a single data set. TODO: Would be a good tunable.
// We essentially perform a little convolution around each instruction,
// to capture the flavor of nearby code, to hopefully match to similar
// code in other modules.
static const unsigned windowSize = 2;
for (unsigned entry = 0; entry < unsigned(instPos.size()); ++entry) {
const unsigned start = instPos[entry];
const spv::Op opCode = asOpCode(start);
if (opCode == spv::OpFunction)
fnId = asId(start + 2);
if (opCode == spv::OpFunctionEnd)
fnId = spv::NoResult;
if (fnId != spv::NoResult) { // if inside a function
if (spv::InstructionDesc[opCode].hasResult()) {
const unsigned word = start + (spv::InstructionDesc[opCode].hasType() ? 2 : 1);
const spv::Id resId = asId(word);
std::uint32_t hashval = fnId * 17; // small prime
for (unsigned i = entry-1; i >= entry-windowSize; --i) {
if (asOpCode(instPos[i]) == spv::OpFunction)
break;
hashval = hashval * 30103 + asOpCodeHash(instPos[i]); // 30103 = semiarbitrary prime
}
for (unsigned i = entry; i <= entry + windowSize; ++i) {
if (asOpCode(instPos[i]) == spv::OpFunctionEnd)
break;
hashval = hashval * 30103 + asOpCodeHash(instPos[i]); // 30103 = semiarbitrary prime
}
if (isOldIdUnmapped(resId))
localId(resId, nextUnusedId(hashval % softTypeIdLimit + firstMappedID));
}
}
}
spv::Op thisOpCode(spv::OpNop);
std::unordered_map<int, int> opCounter;
int idCounter(0);
fnId = spv::NoResult;
process(
[&](spv::Op opCode, unsigned start) {
switch (opCode) {
case spv::OpFunction:
// Reset counters at each function
idCounter = 0;
opCounter.clear();
fnId = asId(start + 2);
break;
case spv::OpImageSampleImplicitLod:
case spv::OpImageSampleExplicitLod:
case spv::OpImageSampleDrefImplicitLod:
case spv::OpImageSampleDrefExplicitLod:
case spv::OpImageSampleProjImplicitLod:
case spv::OpImageSampleProjExplicitLod:
case spv::OpImageSampleProjDrefImplicitLod:
case spv::OpImageSampleProjDrefExplicitLod:
case spv::OpDot:
case spv::OpCompositeExtract:
case spv::OpCompositeInsert:
case spv::OpVectorShuffle:
case spv::OpLabel:
case spv::OpVariable:
case spv::OpAccessChain:
case spv::OpLoad:
case spv::OpStore:
case spv::OpCompositeConstruct:
case spv::OpFunctionCall:
++opCounter[opCode];
idCounter = 0;
thisOpCode = opCode;
break;
default:
thisOpCode = spv::OpNop;
}
return false;
},
[&](spv::Id& id) {
if (thisOpCode != spv::OpNop) {
++idCounter;
const std::uint32_t hashval = opCounter[thisOpCode] * thisOpCode * 50047 + idCounter + fnId * 117;
if (isOldIdUnmapped(id))
localId(id, nextUnusedId(hashval % softTypeIdLimit + firstMappedID));
}
});
}
// EXPERIMENTAL: forward IO and uniform load/stores into operands
// This produces invalid Schema-0 SPIRV
void spirvbin_t::forwardLoadStores()
{
idset_t fnLocalVars; // set of function local vars
idmap_t idMap; // Map of load result IDs to what they load
// EXPERIMENTAL: Forward input and access chain loads into consumptions
process(
[&](spv::Op opCode, unsigned start) {
// Add inputs and uniforms to the map
if ((opCode == spv::OpVariable && asWordCount(start) == 4) &&
(spv[start+3] == spv::StorageClassUniform ||
spv[start+3] == spv::StorageClassUniformConstant ||
spv[start+3] == spv::StorageClassInput))
fnLocalVars.insert(asId(start+2));
if (opCode == spv::OpAccessChain && fnLocalVars.count(asId(start+3)) > 0)
fnLocalVars.insert(asId(start+2));
if (opCode == spv::OpLoad && fnLocalVars.count(asId(start+3)) > 0) {
idMap[asId(start+2)] = asId(start+3);
stripInst(start);
}
return false;
},
[&](spv::Id& id) { if (idMap.find(id) != idMap.end()) id = idMap[id]; }
);
// EXPERIMENTAL: Implicit output stores
fnLocalVars.clear();
idMap.clear();
process(
[&](spv::Op opCode, unsigned start) {
// Add inputs and uniforms to the map
if ((opCode == spv::OpVariable && asWordCount(start) == 4) &&
(spv[start+3] == spv::StorageClassOutput))
fnLocalVars.insert(asId(start+2));
if (opCode == spv::OpStore && fnLocalVars.count(asId(start+1)) > 0) {
idMap[asId(start+2)] = asId(start+1);
stripInst(start);
}
return false;
},
op_fn_nop);
process(
inst_fn_nop,
[&](spv::Id& id) { if (idMap.find(id) != idMap.end()) id = idMap[id]; }
);
strip(); // strip out data we decided to eliminate
}
// optimize loads and stores
void spirvbin_t::optLoadStore()
{
idset_t fnLocalVars; // candidates for removal (only locals)
idmap_t idMap; // Map of load result IDs to what they load
blockmap_t blockMap; // Map of IDs to blocks they first appear in
int blockNum = 0; // block count, to avoid crossing flow control
// Find all the function local pointers stored at most once, and not via access chains
process(
[&](spv::Op opCode, unsigned start) {
const int wordCount = asWordCount(start);
// Count blocks, so we can avoid crossing flow control
if (isFlowCtrl(opCode))
++blockNum;
// Add local variables to the map
if ((opCode == spv::OpVariable && spv[start+3] == spv::StorageClassFunction && asWordCount(start) == 4)) {
fnLocalVars.insert(asId(start+2));
return true;
}
// Ignore process vars referenced via access chain
if ((opCode == spv::OpAccessChain || opCode == spv::OpInBoundsAccessChain) && fnLocalVars.count(asId(start+3)) > 0) {
fnLocalVars.erase(asId(start+3));
idMap.erase(asId(start+3));
return true;
}
if (opCode == spv::OpLoad && fnLocalVars.count(asId(start+3)) > 0) {
const spv::Id varId = asId(start+3);
// Avoid loads before stores
if (idMap.find(varId) == idMap.end()) {
fnLocalVars.erase(varId);
idMap.erase(varId);
}
// don't do for volatile references
if (wordCount > 4 && (spv[start+4] & spv::MemoryAccessVolatileMask)) {
fnLocalVars.erase(varId);
idMap.erase(varId);
}
// Handle flow control
if (blockMap.find(varId) == blockMap.end()) {
blockMap[varId] = blockNum; // track block we found it in.
} else if (blockMap[varId] != blockNum) {
fnLocalVars.erase(varId); // Ignore if crosses flow control
idMap.erase(varId);
}
return true;
}
if (opCode == spv::OpStore && fnLocalVars.count(asId(start+1)) > 0) {
const spv::Id varId = asId(start+1);
if (idMap.find(varId) == idMap.end()) {
idMap[varId] = asId(start+2);
} else {
// Remove if it has more than one store to the same pointer
fnLocalVars.erase(varId);
idMap.erase(varId);
}
// don't do for volatile references
if (wordCount > 3 && (spv[start+3] & spv::MemoryAccessVolatileMask)) {
fnLocalVars.erase(asId(start+3));
idMap.erase(asId(start+3));
}
// Handle flow control
if (blockMap.find(varId) == blockMap.end()) {
blockMap[varId] = blockNum; // track block we found it in.
} else if (blockMap[varId] != blockNum) {
fnLocalVars.erase(varId); // Ignore if crosses flow control
idMap.erase(varId);
}
return true;
}
return false;
},
// If local var id used anywhere else, don't eliminate
[&](spv::Id& id) {
if (fnLocalVars.count(id) > 0) {
fnLocalVars.erase(id);
idMap.erase(id);
}
}
);
process(
[&](spv::Op opCode, unsigned start) {
if (opCode == spv::OpLoad && fnLocalVars.count(asId(start+3)) > 0)
idMap[asId(start+2)] = idMap[asId(start+3)];
return false;
},
op_fn_nop);
// Chase replacements to their origins, in case there is a chain such as:
// 2 = store 1
// 3 = load 2
// 4 = store 3
// 5 = load 4
// We want to replace uses of 5 with 1.
for (const auto& idPair : idMap) {
spv::Id id = idPair.first;
while (idMap.find(id) != idMap.end()) // Chase to end of chain
id = idMap[id];
idMap[idPair.first] = id; // replace with final result
}
// Remove the load/store/variables for the ones we've discovered
process(
[&](spv::Op opCode, unsigned start) {
if ((opCode == spv::OpLoad && fnLocalVars.count(asId(start+3)) > 0) ||
(opCode == spv::OpStore && fnLocalVars.count(asId(start+1)) > 0) ||
(opCode == spv::OpVariable && fnLocalVars.count(asId(start+2)) > 0)) {
stripInst(start);
return true;
}
return false;
},
[&](spv::Id& id) {
if (idMap.find(id) != idMap.end()) id = idMap[id];
}
);
strip(); // strip out data we decided to eliminate
}
// remove bodies of uncalled functions
void spirvbin_t::dceFuncs()
{
msg(3, 2, std::string("Removing Dead Functions: "));
// TODO: There are more efficient ways to do this.
bool changed = true;
while (changed) {
changed = false;
for (auto fn = fnPos.begin(); fn != fnPos.end(); ) {
if (fn->first == entryPoint) { // don't DCE away the entry point!
++fn;
continue;
}
const auto call_it = fnCalls.find(fn->first);
if (call_it == fnCalls.end() || call_it->second == 0) {
changed = true;
stripRange.push_back(fn->second);
// decrease counts of called functions
process(
[&](spv::Op opCode, unsigned start) {
if (opCode == spv::Op::OpFunctionCall) {
const auto call_it = fnCalls.find(asId(start + 3));
if (call_it != fnCalls.end()) {
if (--call_it->second <= 0)
fnCalls.erase(call_it);
}
}
return true;
},
op_fn_nop,
fn->second.first,
fn->second.second);
fn = fnPos.erase(fn);
} else ++fn;
}
}
}
// remove unused function variables + decorations
void spirvbin_t::dceVars()
{
msg(3, 2, std::string("DCE Vars: "));
std::unordered_map<spv::Id, int> varUseCount;
// Count function variable use
process(
[&](spv::Op opCode, unsigned start) {
if (opCode == spv::OpVariable) {
++varUseCount[asId(start+2)];
return true;
} else if (opCode == spv::OpEntryPoint) {
const int wordCount = asWordCount(start);
for (int i = 4; i < wordCount; i++) {
++varUseCount[asId(start+i)];
}
return true;
} else
return false;
},
[&](spv::Id& id) { if (varUseCount[id]) ++varUseCount[id]; }
);
// Remove single-use function variables + associated decorations and names
process(
[&](spv::Op opCode, unsigned start) {
spv::Id id = spv::NoResult;
if (opCode == spv::OpVariable)
id = asId(start+2);
if (opCode == spv::OpDecorate || opCode == spv::OpName)
id = asId(start+1);
if (id != spv::NoResult && varUseCount[id] == 1)
stripInst(start);
return true;
},
op_fn_nop);
}
// remove unused types
void spirvbin_t::dceTypes()
{
std::vector<bool> isType(bound(), false);
// for speed, make O(1) way to get to type query (map is log(n))
for (const auto typeStart : typeConstPos)
isType[asTypeConstId(typeStart)] = true;
std::unordered_map<spv::Id, int> typeUseCount;
// This is not the most efficient algorithm, but this is an offline tool, and
// it's easy to write this way. Can be improved opportunistically if needed.
bool changed = true;
while (changed) {
changed = false;
strip();
typeUseCount.clear();
// Count total type usage
process(inst_fn_nop,
[&](spv::Id& id) { if (isType[id]) ++typeUseCount[id]; }
);
// Remove single reference types
for (const auto typeStart : typeConstPos) {
const spv::Id typeId = asTypeConstId(typeStart);
if (typeUseCount[typeId] == 1) {
changed = true;
--typeUseCount[typeId];
stripInst(typeStart);
}
}
}
}
#ifdef NOTDEF
bool spirvbin_t::matchType(const spirvbin_t::globaltypes_t& globalTypes, spv::Id lt, spv::Id gt) const
{
// Find the local type id "lt" and global type id "gt"
const auto lt_it = typeConstPosR.find(lt);
if (lt_it == typeConstPosR.end())
return false;
const auto typeStart = lt_it->second;
// Search for entry in global table
const auto gtype = globalTypes.find(gt);
if (gtype == globalTypes.end())
return false;
const auto& gdata = gtype->second;
// local wordcount and opcode
const int wordCount = asWordCount(typeStart);
const spv::Op opCode = asOpCode(typeStart);
// no type match if opcodes don't match, or operand count doesn't match
if (opCode != opOpCode(gdata[0]) || wordCount != opWordCount(gdata[0]))
return false;
const unsigned numOperands = wordCount - 2; // all types have a result
const auto cmpIdRange = [&](range_t range) {
for (int x=range.first; x<std::min(range.second, wordCount); ++x)
if (!matchType(globalTypes, asId(typeStart+x), gdata[x]))
return false;
return true;
};
const auto cmpConst = [&]() { return cmpIdRange(constRange(opCode)); };
const auto cmpSubType = [&]() { return cmpIdRange(typeRange(opCode)); };
// Compare literals in range [start,end)
const auto cmpLiteral = [&]() {
const auto range = literalRange(opCode);
return std::equal(spir.begin() + typeStart + range.first,
spir.begin() + typeStart + std::min(range.second, wordCount),
gdata.begin() + range.first);
};
assert(isTypeOp(opCode) || isConstOp(opCode));
switch (opCode) {
case spv::OpTypeOpaque: // TODO: disable until we compare the literal strings.
case spv::OpTypeQueue: return false;
case spv::OpTypeEvent: // fall through...
case spv::OpTypeDeviceEvent: // ...
case spv::OpTypeReserveId: return false;
// for samplers, we don't handle the optional parameters yet
case spv::OpTypeSampler: return cmpLiteral() && cmpConst() && cmpSubType() && wordCount == 8;
default: return cmpLiteral() && cmpConst() && cmpSubType();
}
}
// Look for an equivalent type in the globalTypes map
spv::Id spirvbin_t::findType(const spirvbin_t::globaltypes_t& globalTypes, spv::Id lt) const
{
// Try a recursive type match on each in turn, and return a match if we find one
for (const auto& gt : globalTypes)
if (matchType(globalTypes, lt, gt.first))
return gt.first;
return spv::NoType;
}
#endif // NOTDEF
// Return start position in SPV of given Id. error if not found.
unsigned spirvbin_t::idPos(spv::Id id) const
{
const auto tid_it = idPosR.find(id);
if (tid_it == idPosR.end())
error("ID not found");
return tid_it->second;
}
// Hash types to canonical values. This can return ID collisions (it's a bit
// inevitable): it's up to the caller to handle that gracefully.
std::uint32_t spirvbin_t::hashType(unsigned typeStart) const
{
const unsigned wordCount = asWordCount(typeStart);
const spv::Op opCode = asOpCode(typeStart);
switch (opCode) {
case spv::OpTypeVoid: return 0;
case spv::OpTypeBool: return 1;
case spv::OpTypeInt: return 3 + (spv[typeStart+3]);
case spv::OpTypeFloat: return 5;
case spv::OpTypeVector:
return 6 + hashType(idPos(spv[typeStart+2])) * (spv[typeStart+3] - 1);
case spv::OpTypeMatrix:
return 30 + hashType(idPos(spv[typeStart+2])) * (spv[typeStart+3] - 1);
case spv::OpTypeImage:
return 120 + hashType(idPos(spv[typeStart+2])) +
spv[typeStart+3] + // dimensionality
spv[typeStart+4] * 8 * 16 + // depth
spv[typeStart+5] * 4 * 16 + // arrayed
spv[typeStart+6] * 2 * 16 + // multisampled
spv[typeStart+7] * 1 * 16; // format
case spv::OpTypeSampler:
return 500;
case spv::OpTypeSampledImage:
return 502;
case spv::OpTypeArray:
return 501 + hashType(idPos(spv[typeStart+2])) * spv[typeStart+3];
case spv::OpTypeRuntimeArray:
return 5000 + hashType(idPos(spv[typeStart+2]));
case spv::OpTypeStruct:
{
std::uint32_t hash = 10000;
for (unsigned w=2; w < wordCount; ++w)
hash += w * hashType(idPos(spv[typeStart+w]));
return hash;
}
case spv::OpTypeOpaque: return 6000 + spv[typeStart+2];
case spv::OpTypePointer: return 100000 + hashType(idPos(spv[typeStart+3]));
case spv::OpTypeFunction:
{
std::uint32_t hash = 200000;
for (unsigned w=2; w < wordCount; ++w)
hash += w * hashType(idPos(spv[typeStart+w]));
return hash;
}
case spv::OpTypeEvent: return 300000;
case spv::OpTypeDeviceEvent: return 300001;
case spv::OpTypeReserveId: return 300002;
case spv::OpTypeQueue: return 300003;
case spv::OpTypePipe: return 300004;
case spv::OpConstantNull: return 300005;
case spv::OpConstantSampler: return 300006;
case spv::OpConstantTrue: return 300007;
case spv::OpConstantFalse: return 300008;
case spv::OpConstantComposite:
{
std::uint32_t hash = 300011 + hashType(idPos(spv[typeStart+1]));
for (unsigned w=3; w < wordCount; ++w)
hash += w * hashType(idPos(spv[typeStart+w]));
return hash;
}
case spv::OpConstant:
{
std::uint32_t hash = 400011 + hashType(idPos(spv[typeStart+1]));
for (unsigned w=3; w < wordCount; ++w)
hash += w * spv[typeStart+w];
return hash;
}
default:
error("unknown type opcode");
return 0;
}
}
void spirvbin_t::mapTypeConst()
{
globaltypes_t globalTypeMap;
msg(3, 2, std::string("Remapping Consts & Types: "));
static const std::uint32_t softTypeIdLimit = 3011; // small prime. TODO: get from options
static const std::uint32_t firstMappedID = 8; // offset into ID space
for (auto& typeStart : typeConstPos) {
const spv::Id resId = asTypeConstId(typeStart);
const std::uint32_t hashval = hashType(typeStart);
if (isOldIdUnmapped(resId))
localId(resId, nextUnusedId(hashval % softTypeIdLimit + firstMappedID));
}
}
// Strip a single binary by removing ranges given in stripRange
void spirvbin_t::strip()
{
if (stripRange.empty()) // nothing to do
return;
// Sort strip ranges in order of traversal
std::sort(stripRange.begin(), stripRange.end());
// Allocate a new binary big enough to hold old binary
// We'll step this iterator through the strip ranges as we go through the binary
auto strip_it = stripRange.begin();
int strippedPos = 0;
for (unsigned word = 0; word < unsigned(spv.size()); ++word) {
if (strip_it != stripRange.end() && word >= strip_it->second)
++strip_it;
if (strip_it == stripRange.end() || word < strip_it->first || word >= strip_it->second)
spv[strippedPos++] = spv[word];
}
spv.resize(strippedPos);
stripRange.clear();
buildLocalMaps();
}
// Strip a single binary by removing ranges given in stripRange
void spirvbin_t::remap(std::uint32_t opts)
{
options = opts;
// Set up opcode tables from SpvDoc
spv::Parameterize();
validate(); // validate header
buildLocalMaps(); // build ID maps
msg(3, 4, std::string("ID bound: ") + std::to_string(bound()));
if (options & STRIP) stripDebug();
strip(); // strip out data we decided to eliminate
if (options & OPT_LOADSTORE) optLoadStore();
if (options & OPT_FWD_LS) forwardLoadStores();
if (options & DCE_FUNCS) dceFuncs();
if (options & DCE_VARS) dceVars();
if (options & DCE_TYPES) dceTypes();
strip(); // strip out data we decided to eliminate
stripDeadRefs(); // remove references to things we DCEed
// after the last strip, we must clean any debug info referring to now-deleted data
if (options & MAP_TYPES) mapTypeConst();
if (options & MAP_NAMES) mapNames();
if (options & MAP_FUNCS) mapFnBodies();
if (options & MAP_ALL) {
mapRemainder(); // map any unmapped IDs
applyMap(); // Now remap each shader to the new IDs we've come up with
}
}
// remap from a memory image
void spirvbin_t::remap(std::vector<std::uint32_t>& in_spv, std::uint32_t opts)
{
spv.swap(in_spv);
remap(opts);
spv.swap(in_spv);
}
} // namespace SPV
#endif // defined (use_cpp11)
Reference in New Issue View Git Blame Copy Permalink