generating proper tasking function for nvptx
This commit is contained in:
24
func.cpp
24
func.cpp
@@ -240,16 +240,6 @@ Function::emitCode(FunctionEmitContext *ctx, llvm::Function *function,
|
|||||||
// thread index, and the thread count variables.
|
// thread index, and the thread count variables.
|
||||||
llvm::Function::arg_iterator argIter = function->arg_begin();
|
llvm::Function::arg_iterator argIter = function->arg_begin();
|
||||||
llvm::Value *structParamPtr = argIter++;
|
llvm::Value *structParamPtr = argIter++;
|
||||||
llvm::Value *threadIndex = argIter++;
|
|
||||||
llvm::Value *threadCount = argIter++;
|
|
||||||
llvm::Value *taskIndex = argIter++;
|
|
||||||
llvm::Value *taskCount = argIter++;
|
|
||||||
llvm::Value *taskIndex0 = argIter++;
|
|
||||||
llvm::Value *taskIndex1 = argIter++;
|
|
||||||
llvm::Value *taskIndex2 = argIter++;
|
|
||||||
llvm::Value *taskCount0 = argIter++;
|
|
||||||
llvm::Value *taskCount1 = argIter++;
|
|
||||||
llvm::Value *taskCount2 = argIter++;
|
|
||||||
|
|
||||||
// Copy the function parameter values from the structure into local
|
// Copy the function parameter values from the structure into local
|
||||||
// storage
|
// storage
|
||||||
@@ -266,6 +256,19 @@ Function::emitCode(FunctionEmitContext *ctx, llvm::Function *function,
|
|||||||
ctx->SetFunctionMask(ptrval);
|
ctx->SetFunctionMask(ptrval);
|
||||||
}
|
}
|
||||||
|
|
||||||
|
if (g->target->getISA() != Target::NVPTX64)
|
||||||
|
{
|
||||||
|
llvm::Value *threadIndex = argIter++;
|
||||||
|
llvm::Value *threadCount = argIter++;
|
||||||
|
llvm::Value *taskIndex = argIter++;
|
||||||
|
llvm::Value *taskCount = argIter++;
|
||||||
|
llvm::Value *taskIndex0 = argIter++;
|
||||||
|
llvm::Value *taskIndex1 = argIter++;
|
||||||
|
llvm::Value *taskIndex2 = argIter++;
|
||||||
|
llvm::Value *taskCount0 = argIter++;
|
||||||
|
llvm::Value *taskCount1 = argIter++;
|
||||||
|
llvm::Value *taskCount2 = argIter++;
|
||||||
|
|
||||||
// Copy threadIndex and threadCount into stack-allocated storage so
|
// Copy threadIndex and threadCount into stack-allocated storage so
|
||||||
// that their symbols point to something reasonable.
|
// that their symbols point to something reasonable.
|
||||||
threadIndexSym->storagePtr = ctx->AllocaInst(LLVMTypes::Int32Type, "threadIndex");
|
threadIndexSym->storagePtr = ctx->AllocaInst(LLVMTypes::Int32Type, "threadIndex");
|
||||||
@@ -313,6 +316,7 @@ Function::emitCode(FunctionEmitContext *ctx, llvm::Function *function,
|
|||||||
taskCountSym2->storagePtr = ctx->AllocaInst(LLVMTypes::Int32Type, "taskCount2");
|
taskCountSym2->storagePtr = ctx->AllocaInst(LLVMTypes::Int32Type, "taskCount2");
|
||||||
ctx->StoreInst(taskCount2, taskCountSym2->storagePtr);
|
ctx->StoreInst(taskCount2, taskCountSym2->storagePtr);
|
||||||
}
|
}
|
||||||
|
}
|
||||||
else {
|
else {
|
||||||
// Regular, non-task function
|
// Regular, non-task function
|
||||||
llvm::Function::arg_iterator argIter = function->arg_begin();
|
llvm::Function::arg_iterator argIter = function->arg_begin();
|
||||||
|
|||||||
3
type.cpp
3
type.cpp
@@ -2957,6 +2957,8 @@ FunctionType::LLVMFunctionType(llvm::LLVMContext *ctx, bool removeMask) const {
|
|||||||
// hold them until the task actually runs.)
|
// hold them until the task actually runs.)
|
||||||
llvm::Type *st = llvm::StructType::get(*ctx, llvmArgTypes);
|
llvm::Type *st = llvm::StructType::get(*ctx, llvmArgTypes);
|
||||||
callTypes.push_back(llvm::PointerType::getUnqual(st));
|
callTypes.push_back(llvm::PointerType::getUnqual(st));
|
||||||
|
if (g->target->getISA() != Target::NVPTX64)
|
||||||
|
{
|
||||||
callTypes.push_back(LLVMTypes::Int32Type); // threadIndex
|
callTypes.push_back(LLVMTypes::Int32Type); // threadIndex
|
||||||
callTypes.push_back(LLVMTypes::Int32Type); // threadCount
|
callTypes.push_back(LLVMTypes::Int32Type); // threadCount
|
||||||
callTypes.push_back(LLVMTypes::Int32Type); // taskIndex
|
callTypes.push_back(LLVMTypes::Int32Type); // taskIndex
|
||||||
@@ -2968,6 +2970,7 @@ FunctionType::LLVMFunctionType(llvm::LLVMContext *ctx, bool removeMask) const {
|
|||||||
callTypes.push_back(LLVMTypes::Int32Type); // taskCount1
|
callTypes.push_back(LLVMTypes::Int32Type); // taskCount1
|
||||||
callTypes.push_back(LLVMTypes::Int32Type); // taskCount2
|
callTypes.push_back(LLVMTypes::Int32Type); // taskCount2
|
||||||
}
|
}
|
||||||
|
}
|
||||||
else
|
else
|
||||||
// Otherwise we already have the types of the arguments
|
// Otherwise we already have the types of the arguments
|
||||||
callTypes = llvmArgTypes;
|
callTypes = llvmArgTypes;
|
||||||
|
|||||||
Reference in New Issue
Block a user