/// <summary cref="LLVMBackend.PrepareModule(CompileUnit, LLVMModuleRef, EntryPoint, LLVMValueRef)"/>
internal override void PrepareModule(CompileUnit unit, LLVMModuleRef module, EntryPoint entryPoint, LLVMValueRef generatedEntryPoint)
{
// Add required metdata for entry point
var context = unit.LLVMContext;
var values = new LLVMValueRef[]
{
generatedEntryPoint,
MDStringInContext(context, KernelAnnotationName, KernelAnnotationName.Length),
ConstInt(context.Int32Type, 1, false)
};
AddNamedMetadataOperand(module, LLVMAnnotationsName,
MDNodeInContext(context, out values[0], values.Length));
var specialization = entryPoint.Specialization;
if (specialization.MaxNumThreadsPerGroup.HasValue)
{
var maxThreadsPerGroupAnnotation = new LLVMValueRef[]
{
generatedEntryPoint,
MDStringInContext(
context,
MaxNumThreadsPerGroupAnnotationAnnotationName,
MaxNumThreadsPerGroupAnnotationAnnotationName.Length),
ConstInt(context.Int32Type, specialization.MaxNumThreadsPerGroup.Value, false)
};
AddNamedMetadataOperand(module, LLVMAnnotationsName,
MDNodeInContext(context, out maxThreadsPerGroupAnnotation[0], maxThreadsPerGroupAnnotation.Length));
}
if (specialization.MinNumGroupsPerMultiprocessor.HasValue)
{
var minNumGroupsAnnotation = new LLVMValueRef[]
{
generatedEntryPoint,
MDStringInContext(
context,
MinNumGroupsPerMultiprocessorAnnotationName,
MinNumGroupsPerMultiprocessorAnnotationName.Length),
ConstInt(context.Int32Type, specialization.MinNumGroupsPerMultiprocessor.Value, false)
};
AddNamedMetadataOperand(module, LLVMAnnotationsName,
MDNodeInContext(context, out minNumGroupsAnnotation[0], minNumGroupsAnnotation.Length));
}
// Perform NVVM reflect pass
PreparePTXModule(
module,
generatedEntryPoint,
unit.HasFlags(CompileUnitFlags.PTXFlushDenormalsToZero | CompileUnitFlags.FastMath),
unit.HasFlags(CompileUnitFlags.FastMath));
}
