Abstract: Techniques for token triggered multithreading in a multithreaded processor are disclosed. An instruction issuance sequence for a plurality of threads of the multithreaded processor is controlled by associating with each of the threads at least one register which stores a value identifying a next thread to be permitted to issue one or more instructions, and utilizing the stored value to control the instruction issuance sequence. For example, each of a plurality of hardware thread units of the multithreaded processor may include a corresponding local register updatable by that hardware thread unit, with the local register for a given one of the hardware thread units storing a value identifying the next thread to be permitted to issue one or more instructions after the given hardware thread unit has issued one or more instructions. A global register arrangement may also or alternatively be used.

Abstract: A cache memory for use in a multithreaded processor includes a number of set-associative thread caches, with one or more of the thread caches each implementing an eviction process based on access request address that reduces the amount of replacement policy storage required in the cache memory. At least a given one of the thread caches in an illustrative embodiment includes a memory array having multiple sets of memory locations, and a directory for storing tags each corresponding to at least a portion of a particular address of one of the memory locations. The directory has multiple entries each storing multiple ones of the tags, such that if there are n sets of memory locations in the memory array, there are n tags associated with each directory entry. The directory is utilized in implementing a set-associative address mapping between access requests and memory locations of the memory array.

Abstract: A cache memory for use in a multithreaded processor includes a number of set-associative thread caches, with one or more of the thread caches each implementing a thread-based eviction process that reduces the amount of replacement policy storage required in the cache memory. At least a given one of the thread caches in an illustrative embodiment includes a memory array having multiple sets of memory locations, and a directory for storing tags each corresponding to at least a portion of a particular address of one of the memory locations. The directory has multiple entries each storing multiple ones of the tags, such that if there are n sets of memory locations in the memory array, there are n tags associated with each directory entry. The directory is utilized in implementing a set-associative address mapping between access requests and memory locations of the memory array.

Abstract: A multithreaded processor includes an instruction decoder for decoding retrieved instructions to determine an instruction type for each of the retrieved instructions, an integer unit coupled to the instruction decoder for processing integer type instructions, and a vector unit coupled to the instruction decoder for processing vector type instructions. A reduction unit is preferably associated with the vector unit and receives parallel data elements processed in the vector unit. The reduction unit generates a serial output from the parallel data elements. The processor may be configured to execute at least control code, digital signal processor (DSP) code, Java code and network processing code, and is therefore well-suited for use in a convergence device. The processor is preferably configured to utilize token triggered threading in conjunction with instruction pipelining.

Abstract: A pipelined processor is configured to provide virtual single-cycle instruction execution using a register locking mechanism in conjunction with instruction stalling based on lock status. In an illustrative embodiment, a set of register locks is maintained in the form of a stored bit vector in which each bit indicates the current lock status of a corresponding register. A decode unit receives an instruction fetched from memory, and decodes the instruction to determine its source and destination registers. The instruction is stalled for at least one processor cycle if either its source register or destination register is already locked by another instruction. The stall continues until the source and destination registers of the instruction are both unlocked, i.e., no longer in use by other instructions.

Abstract: The invention provides techniques for reducing the port pressure of a clustered processor. In an illustrative embodiment, the processor includes multiple clusters of execution units, with each of the clusters having a portion of a register file and a portion of a predicate file associated therewith, such that a given cluster is permitted to write to and read from its associated portions of the register and predicate files. A cooperative interconnection technique in accordance with the invention utilizes an inter-cluster move instruction specifying a source cluster and a destination cluster to copy a value from the source cluster to the destination cluster. The value is transmitted over a designated interconnect structure within the processor, and the inter-cluster move instruction is separated into two sub-instructions, one of which is executed by a unit in the source cluster, and another of which is executed by a unit in the destination cluster. These units may be, e.g.

Abstract: The invention provides techniques for reducing the port pressure of a clustered processor. In an illustrative embodiment, the processor includes multiple clusters of execution units, with each of the clusters having a portion of a register file and a portion of a predicate file associated therewith, such that a given cluster is permitted to write to and read from its associated portions of the register and predicate files. A duplicator interconnection technique in accordance with the invention reduces port pressure by providing one or more global move units in the processor. A given global move unit uses an inter-cluster move instruction to copy a value from a portion of the register or predicate file associated with a source cluster to another portion of the register or predicate file associated with a destination cluster.

Abstract: The invention provides techniques for improving the performance of pipelined processors by eliminating unnecessary stalling of instructions. In an illustrative embodiment, a compiler is used to identify pipeline dependencies in a given set of instructions. The compiler then groups the set of instructions into a code block having a field which indicates the types of pipeline dependencies, if any, in the set of instructions. The field may indicate the types of pipeline dependencies by specifying which of a predetermined set of hazards arise in the plurality of instructions when executed on a given pipelined processor. For example, the field may indicate whether the code block includes any Read After Write (RAW) hazards, Write After Write (WAW) hazards or Write After Read (WAR) hazards. The code block may include one or more dynamic scheduling instructions, with each of the dynamic scheduling instructions including a set of instructions for execution in a multi-issue processor.

Abstract: A processor is configured to include at least two architecturally-distinct storage spaces, such as, for example, a stack for storing control operands associated with one or more instructions, and a register file for storing computational operands associated with one or more instructions. The processor further includes a datapath which is at least partially shared by the stack and register file, a multiplexer operative to select an output of either the stack or the register file for application to an input of the shared datapath, and a demultiplexer operative to select an output of the shared datapath for application to an input of either the stack or the register file.

Abstract: The invention provides techniques for reducing the port pressure of a clustered processor. In an illustrative embodiment, the processor includes multiple clusters of execution units, with each of the clusters having a portion of a register file and a portion of a predicate file associated therewith, such that a given cluster is permitted to write to and read from its associated portions of the register and predicate files. A replication technique in accordance with the invention reduces port pressure by replicating, e.g., a register lock file and a predicate lock file of the processor for each of the clusters. The replicated files vary depending upon whether the technique is implemented with a write-only interconnection or a read-only interconnection.

Abstract: A computer system supporting N different machine views, where N.gtoreq.2, includes a memory for storing instructions, a number of execution units for processing data based on execution controls, and N different decoders for generating the execution controls using instructions retrieved from the memory. Each of the N decoders is operative to decode retrieved instructions in accordance with one of the N machine views. A particular one of the N decoders to be used to decode a given retrieved instruction may be selected by a program running on the system. In one embodiment, the decoders for the N machine views are implemented as N separate decoders, and a multiplexer is used to select the output of one of the N decoders for connection to one or more of the execution units. In another embodiment, a set of reconfigurable hardware is dynamically reprogrammed to implement one or more of the N decoders as directed by the program running on the system.