SCISM : A scalable compound instruction set

SCISM representation. ~ ~ “” compounding facility (footnote 2). The compounding facility or preprocessor examines the instruction stream with an implementation-dependent predetermined scope and produces a compound instruction program. rules which reflect the system architecture, the hardware organization, and the permissible parallel execution between categories of instructions. These rules are hereafter referred to as compounding rules. The program produced by the compounding facility can then be executed directly by a compound instruction execution engine, which considers a compound instruction as a single instruction. A compound instruction reflects the parallel issue of instructions; it comprises some number of independent instructions or interlocked instructions, the latter provided that the interlocks are of a form that can be collapsed by the execution hardware. As indicated in the previous paragraph, the compound instruction contains information pertinent to the parallel issuing and execution of instructions. In general the information, incorporated in the compound instruction in the form of decoding or tagging, indicates that compound instructions are “free of hazards,” and that functional units required for the execution of an instruction are available when necessary. Instructions composing a The compound instruction program is based on a set of 64 compound instruction need not be consecutive, allowing I is in original form I r=tag Maximum compound instruction format for the compounding of at most three instructions. for out-of-order issue. Additional information, related for example to branch prediction, functional unit allocation and control, or routing information can also be incorporated in the decoding or tagging of compound instructions. We note here that the choice between decoding instructions and tagging them depends entirely on the architecture and the implementation constraints. For reasons that become obvious later in this section, tagging of instructions is mandatory for architectures that allow variable-length instructions, or that allow data to be intermingled with instructions. In the remainder of this paper, we assume that the compounding information is in the form of tags rather than decoding. Additional discussion of the tags can be found in [B]. The tag identifies the boundaries between single and compound instructions. For example, Figure 2 shows the format of a maximal compound instruction, if it is assumed that up to three instructions may be included in a compound instruction. In the figure, the instruction 4 is in its original form (implying full compatibility with the base architecture), and Tis the tag. As indicated earlier, the tag may contain as little or as much additional information as deemed efficacious for a particular implementation. In the rest of the presentation, for simplicity of exposition, we discuss only the information necessary to determine a compound instruction. To determine the compounding of three instructions, two bits are required, denoted as to and t, . These bits are required for delimiting compound instructions, with 00 representing single instructions and 01 and 10 representing twoand three-instruction compound instructions, respectively. The two control tag bits to and t, merit further discussion. One of the fundamental properties of the SCISM machine organization is that it enables preprocessing to be detached from instruction issueldecode. To do this, the compounding of instructions must be “permanent,” with permanence being dictated by the location of the compounding facility. For example, the compounding facility may be a software facility-perhaps a S. VASSILIADIS, B. BLANER, AND R. J . EICKEMEYER IBM J. RES. DEVELOP. VOL. 38 NO. 1 JANUARY 1w4 post compiler [19]-or it may be a hardware facility in the form of a hardware preprocessor located, for example, between the cache and the memory subsystems. In this hardware preprocessor, the “program” to be compounded is the stream of instruction text that is fetched during the servicing of a cache miss and preprocessed to produce an optimized stream of compound instructions, i.e., a stream of instructions with their ags. The compounding 7. e information in the tags remains intact as long as the line resides in the cache, and is thus relatively “permanent.” If the line should be removed from the cache for any of a number of reasons, the associated tags become invalid and b-l