Compiler Optimization for Superscalar Systems: Global Instruction Scheduling without Copies

Vol. 10 No. 1 1998 Many of today’s computer applications require computation power not easily achieved by computer architectures that provide little or no parallelism. A promising alternative is the parallel architecture, more specifically, the instruction-level parallel (ILP) architecture, which increases computation during each machine cycle. ILP computers allow parallel computation of the lowest level machine operations within a single instruction cycle, including such operations as memory loads and stores, integer additions, and floating-point multiplications. ILP architectures, like conventional architectures, contain multiple functional units and pipelined functional units; but, they have a single program counter and operate on a single instruction stream. Compaq Computer Corporation’s AlphaServer system, based on the Alpha 21164 microprocessor, is an example of an ILP machine. To effectively use parallel hardware and obtain performance advantages, compiler programs must identify the appropriate level of parallelism. For ILP architectures, the compiler must order the single instruction stream such that multiple, low-level operations execute simultaneously whenever possible. This ordering by the compiler of machine operations to effectively use an ILP architecture’s increased parallelism is called instruction scheduling. It is an optimization not usually found in compilers for non-ILP architectures. Instruction scheduling is classified as local if it considers code only within a basic block and global if it schedules code across multiple basic blocks. A disadvantage to local instruction scheduling is its inability to consider context from surrounding blocks. While local scheduling can find parallelism within a basic block, it can do nothing to exploit parallelism between basic blocks. Generally, global scheduling is preferred because it can take advantage of added program parallelism available when the compiler is allowed to move code across basic block boundaries. Tjaden and Flynn, for example, found parallelism within a basic block quite limited. Using a test suite of scientific programs, they measured an average parallelism of 1.8 within basic blocks. In similar experiments on scientific proCompiler Optimization for Superscalar Systems: Global Instruction Scheduling without Copies Philip H. Sweany Steven M. Carr Brett L. Huber