LLVMBasicBlockRef DeclareBB(LLVMBasicBlockRef Src) {
// Check if this is something we already computed.
{
auto i = BBMap.find(Src);
if (i != BBMap.end()) {
return i->second;
}
}
LLVMValueRef V = LLVMBasicBlockAsValue(Src);
if (!LLVMValueIsBasicBlock(V) || LLVMValueAsBasicBlock(V) != Src)
report_fatal_error("Basic block is not a basic block");
const char *Name = LLVMGetBasicBlockName(Src);
const char *VName = LLVMGetValueName(V);
if (Name != VName)
report_fatal_error("Basic block name mismatch");
LLVMBasicBlockRef BB = LLVMAppendBasicBlock(Fun, Name);
return BBMap[Src] = BB;
}
static ir::ControlFlowGraph::iterator convertCallToJumps(
const BasicBlockMap& newBlocks,
ir::ControlFlowGraph::iterator functionEntry,
ir::ControlFlowGraph::iterator functionExit, ir::IRKernel& kernel,
ir::ControlFlowGraph::instruction_iterator call,
ir::ControlFlowGraph::iterator block)
{
// split the block
auto firstInstructionOfSplitBlock = call;
++firstInstructionOfSplitBlock;
auto returnBlock = kernel.cfg()->split_block(block,
firstInstructionOfSplitBlock, ir::Edge::Invalid);
kernel.cfg()->remove_edge(block->out_edges.front());
// add edges
kernel.cfg()->insert_edge(ir::Edge(block, functionEntry, ir::Edge::Branch));
kernel.cfg()->insert_edge(ir::Edge(functionExit,
returnBlock, ir::Edge::Branch));
ir::PTXInstruction& ptxCall = static_cast<ir::PTXInstruction&>(**call);
if(ptxCall.pg.condition != ir::PTXOperand::PT)
{
kernel.cfg()->insert_edge(ir::Edge(block,
returnBlock, ir::Edge::FallThrough));
}
else
{
ptxCall.uni = true;
}
// set branch to function instruction
ptxCall = ir::PTXInstruction(ir::PTXInstruction::Bra);
ptxCall.d.addressMode = ir::PTXOperand::Label;
ptxCall.d.identifier = functionEntry->label();
// set all return instructions to branches to the exit node
for(auto block = newBlocks.begin(); block != newBlocks.end(); ++block)
{
for(auto instruction = block->second->instructions.begin();
instruction != block->second->instructions.end(); ++instruction)
{
ir::PTXInstruction& ptx = static_cast<ir::PTXInstruction&>(
**instruction);
if(ptx.opcode != ir::PTXInstruction::Ret) continue;
ptx = ir::PTXInstruction(ir::PTXInstruction::Bra);
ptx.d.addressMode = ir::PTXOperand::Label;
ptx.d.identifier = functionExit->label();
if(block->second->has_fallthrough_edge() &&
ptx.pg.condition == ir::PTXOperand::PT)
{
auto fallthrough = block->second->get_fallthrough_edge();
ptx.uni = true;
ir::Edge newEdge(fallthrough->head, fallthrough->tail,
ir::Edge::Branch);
kernel.cfg()->remove_edge(fallthrough);
kernel.cfg()->insert_edge(newEdge);
}
break;
}
}
// set branch back after executing function
auto ret = new ir::PTXInstruction(ir::PTXInstruction::Bra);
ret->uni = true;
ret->d.addressMode = ir::PTXOperand::Label;
ret->d.identifier = returnBlock->label();
functionExit->instructions.push_back(ret);
return returnBlock;
}
static void convertParametersToRegisters(
const BasicBlockMap& newBlocks, ir::IRKernel& kernel,
ir::ControlFlowGraph::instruction_iterator callIterator,
const ir::IRKernel& calledKernel)
{
typedef std::unordered_map<std::string, ir::PTXOperand> OperandMap;
typedef std::unordered_set<std::string> StringSet;
reportE(REPORT_DETAILS, " Converting parameters to registers...");
// Get a map from argument name to register in the calling function
OperandMap argumentMap;
StringSet bitBucketArguments;
auto argument = calledKernel.arguments.begin();
ir::PTXInstruction& call = static_cast<ir::PTXInstruction&>(**callIterator);
for(auto parameter = call.d.array.begin();
parameter != call.d.array.end(); ++parameter, ++argument)
{
if(parameter->addressMode == ir::PTXOperand::BitBucket)
{
bitBucketArguments.insert(argument->name);
continue;
}
assert(argument != calledKernel.arguments.end());
assert(parameter->addressMode == ir::PTXOperand::Register ||
parameter->addressMode == ir::PTXOperand::Immediate);
assert(argumentMap.count(argument->name) == 0);
assert(argument->returnArgument);
argumentMap.insert(std::make_pair(argument->name, *parameter));
}
for(auto parameter = call.b.array.begin();
parameter != call.b.array.end(); ++parameter, ++argument)
{
if(parameter->addressMode == ir::PTXOperand::BitBucket)
{
bitBucketArguments.insert(argument->name);
continue;
}
assert(argument != calledKernel.arguments.end());
assert(parameter->addressMode == ir::PTXOperand::Register ||
parameter->addressMode == ir::PTXOperand::Immediate);
assert(argumentMap.count(argument->name) == 0);
assert(!argument->returnArgument);
argumentMap.insert(std::make_pair(argument->name, *parameter));
}
// Convert all stores to that parameter to moves to the associated register
for(auto block = newBlocks.begin(); block != newBlocks.end(); ++block)
{
for(auto instruction = block->second->instructions.begin();
instruction != block->second->instructions.end(); ++instruction)
{
ir::PTXInstruction& ptx = static_cast<ir::PTXInstruction&>(
**instruction);
if(ptx.opcode != ir::PTXInstruction::St) continue;
if(ptx.addressSpace != ir::PTXInstruction::Param) continue;
if(ptx.d.addressMode != ir::PTXOperand::Address) continue;
if(bitBucketArguments.count(ptx.d.identifier))
{
delete *instruction;
instruction = --block->second->instructions.erase(instruction);
continue;
}
auto argument = argumentMap.find(ptx.d.identifier);
if(argument == argumentMap.end()) continue;
ptx.type = argument->second.type;
ptx.pg = call.pg;
ptx.d = argument->second;
if(argument->second.addressMode == ir::PTXOperand::Register)
{
// If the types match, it is a move
if(argument->second.type == ptx.d.type)
{
ptx.opcode = ir::PTXInstruction::Mov;
}
else
{
// otherwise, we need a cast
ptx.opcode = ir::PTXInstruction::Cvt;
ptx.modifier = ir::PTXInstruction::Modifier_invalid;
}
}
else
{
assert(argument->second.addressMode ==
ir::PTXOperand::Immediate);
ptx.opcode = ir::PTXInstruction::Mov;
}
}
}
// Convert all loads from that parameter to moves from the register
for(auto block = newBlocks.begin(); block != newBlocks.end(); ++block)
{
for(auto instruction = block->second->instructions.begin();
instruction != block->second->instructions.end(); ++instruction)
{
ir::PTXInstruction& ptx = static_cast<ir::PTXInstruction&>(
**instruction);
if(ptx.opcode != ir::PTXInstruction::Ld) continue;
if(ptx.addressSpace != ir::PTXInstruction::Param) continue;
if(ptx.a.addressMode != ir::PTXOperand::Address) continue;
if(bitBucketArguments.count(ptx.a.identifier))
{
delete *instruction;
instruction = --block->second->instructions.erase(instruction);
continue;
}
auto argument = argumentMap.find(ptx.a.identifier);
if(argument == argumentMap.end()) continue;
assert(ptx.d.addressMode == ir::PTXOperand::Register);
ptx.type = argument->second.type;
ptx.pg = call.pg;
ptx.a = argument->second;
// If the types match, it is a move
if(ptx.type == ptx.a.type)
{
ptx.opcode = ir::PTXInstruction::Mov;
}
else
{
// otherwise, we need a cast
ptx.opcode = ir::PTXInstruction::Cvt;
ptx.modifier = ir::PTXInstruction::Modifier_invalid;
}
}
}
}
static void insertAndConnectBlocks(BasicBlockMap& newBlocks,
ir::ControlFlowGraph::iterator& functionEntry,
ir::ControlFlowGraph::iterator& functionExit,
ir::IRKernel& kernel, unsigned int& nextRegister,
const ir::IRKernel& inlinedKernel)
{
typedef std::unordered_map<ir::PTXOperand::RegisterType,
ir::PTXOperand::RegisterType> RegisterMap;
ir::IRKernel copy;
const ir::IRKernel* inlinedKernelPointer = &inlinedKernel;
// create a copy if the call is recursive
if(inlinedKernelPointer == &kernel)
{
copy = inlinedKernel;
inlinedKernelPointer = ©
}
// Insert new blocks
for(auto block = inlinedKernelPointer->cfg()->begin();
block != inlinedKernelPointer->cfg()->end(); ++block)
{
auto newBlock = kernel.cfg()->clone_block(block);
newBlocks.insert(std::make_pair(block, newBlock));
}
// Connect new blocks, rename branch labels
for(auto block = newBlocks.begin(); block != newBlocks.end(); ++block)
{
for(auto edge = block->first->out_edges.begin();
edge != block->first->out_edges.end(); ++edge)
{
auto headBlock = block->second;
auto tail = (*edge)->tail;
auto tailBlock = newBlocks.find(tail);
assert(tailBlock != newBlocks.end());
kernel.cfg()->insert_edge(ir::Edge(headBlock,
tailBlock->second, (*edge)->type));
if((*edge)->type == ir::Edge::Branch)
{
assert(!headBlock->instructions.empty());
auto instruction = headBlock->instructions.back();
auto branch = static_cast<ir::PTXInstruction*>(instruction);
if(branch->opcode == ir::PTXInstruction::Ret) continue;
assertM(branch->opcode == ir::PTXInstruction::Bra, "Expecting "
<< branch->toString() << " to be a branch");
branch->d.identifier = tailBlock->second->label();
}
}
}
// Assign copied blocks new registers
RegisterMap newRegisters;
for(auto block = newBlocks.begin(); block != newBlocks.end(); ++block)
{
for(auto instruction = block->second->instructions.begin();
instruction != block->second->instructions.end(); ++instruction)
{
ir::PTXInstruction& ptx = static_cast<ir::PTXInstruction&>(
**instruction);
ir::PTXOperand* operands[] = {&ptx.pg, &ptx.pq, &ptx.d, &ptx.a,
&ptx.b, &ptx.c};
for(unsigned int i = 0; i < 6; ++i)
{
ir::PTXOperand& operand = *operands[i];
if( operand.addressMode != ir::PTXOperand::Register &&
operand.addressMode != ir::PTXOperand::Indirect &&
operand.addressMode != ir::PTXOperand::ArgumentList)
{
continue;
}
if(operand.type != ir::PTXOperand::pred)
{
if(operand.array.empty() &&
operand.addressMode != ir::PTXOperand::ArgumentList)
{
auto mapping = newRegisters.find(operand.reg);
if(mapping == newRegisters.end())
{
mapping = newRegisters.insert(std::make_pair(
operand.reg, nextRegister++)).first;
}
operand.reg = mapping->second;
}
else
{
for(auto subOperand = operand.array.begin();
subOperand != operand.array.end(); ++subOperand )
{
if(!subOperand->isRegister()) continue;
auto mapping = newRegisters.find(subOperand->reg);
if(mapping == newRegisters.end())
{
mapping = newRegisters.insert(std::make_pair(
subOperand->reg, nextRegister++)).first;
}
subOperand->reg = mapping->second;
}
}
}
else if(operand.addressMode != ir::PTXOperand::ArgumentList)
{
if(operand.condition == ir::PTXOperand::Pred
|| operand.condition == ir::PTXOperand::InvPred)
{
auto mapping = newRegisters.find(operand.reg);
if(mapping == newRegisters.end())
{
mapping = newRegisters.insert(std::make_pair(
operand.reg, nextRegister++)).first;
}
operand.reg = mapping->second;
}
}
}
}
}
// Assign copied blocks new local variables
typedef std::unordered_map<std::string, std::string> LocalMap;
LocalMap locals;
for(auto local = inlinedKernel.locals.begin();
local != inlinedKernel.locals.end(); ++local)
{
std::string newName = "_Zinlined_" + local->first;
locals.insert(std::make_pair(local->first, newName));
auto newLocal = kernel.locals.insert(
std::make_pair(newName, local->second)).first;
newLocal->second.name = newName;
}
for(auto block = newBlocks.begin(); block != newBlocks.end(); ++block)
{
for(auto instruction = block->second->instructions.begin();
instruction != block->second->instructions.end(); ++instruction)
{
ir::PTXInstruction& ptx = static_cast<ir::PTXInstruction&>(
**instruction);
if(!ptx.mayHaveAddressableOperand()) continue;
ir::PTXOperand* operands[] = {&ptx.pg, &ptx.pq, &ptx.d, &ptx.a,
&ptx.b, &ptx.c};
for(unsigned int i = 0; i < 6; ++i)
{
ir::PTXOperand& operand = *operands[i];
if(operand.addressMode != ir::PTXOperand::Address) continue;
auto local = locals.find(operand.identifier);
if(local == locals.end()) continue;
operand.identifier = local->second;
}
}
}
// Get the entry and exit points
auto entryMapping = newBlocks.find(
inlinedKernelPointer->cfg()->get_entry_block());
assert(entryMapping != newBlocks.end());
functionEntry = entryMapping->second;
auto exitMapping = newBlocks.find(
inlinedKernelPointer->cfg()->get_exit_block());
assert(exitMapping != newBlocks.end());
functionExit = exitMapping->second;
}
