An Argument Forwarding Queue Machine for Improved Memory Performance

The Argument Forwarding Queue (AFQ) machine architecture combines the control ow paradigm of Von Neumann machines with the ability of dataaow machines to directly forward data resulting from operations to instructions that use the results. The forwarding capability is used to organize the data operands required by successive instructions into data blocks such that operands for successive instructions are stored in successive locations. By coupling the fetching of instructions with the fetching of data, we can reduce the delays caused by data cache misses. In addition, AFQ supports non-blocking loads where the requests issued to memory can be returned in any order; thus providing better tolerance to memory latency. We describe an implementation of AFQ and show how it can be extended into a superscalar machine and incorporate speculative instruction execution. A program is compiled into a combination control-data ow graph (CDFG) which is directly executed by the AFQ machine. With each instruction basic block (IBB) in the CDFG, we associate a data basic block (DBB) in the activation record, such that the operands of instructions are stored in DBB in the order they are used by the instructions in IBB. The result of an operation is forwarded directly to the appropriate position in a DBB that corresponds to an instruction in IBB that uses the data value. We outline the compilation process for this machine and illustrate how existing program optimization and transformation techniques can be used to improve the performance.