S3 Incorporated, Inc.

Exponential Technology, Inc.

Gene Shen

James S Blomgren

Shalesh Thusoo

Betty Kikuta

A mixed-modulo address generation unit has several inputs. The unit effectively adds together a subset of these inputs in a reduced modulus while simultaneously adding other inputs in a full modulus to the partial sum of reduced-modulus inputs. The subset of inputs receives reduced-width address components such as 16-bit address components which are effectively added together in modulo 64K. The other inputs receive full-width address components such as 32-bit components which are added in the full modulus, 4G. Reduced-width components are zero-extended to 32 bits before input to a standard 32-bit adder. A 16-bit carry generator also receives the reduced-width components and generates the carries out of the 16th bit position. When one or more carries is detected, a correction term is subtracted from the initial sum which is recirculated to the adder's input in a subsequent step. The correction term is the number of carries out of the 16th bit position multiplied by 64K. The full-width segment bases for all active segments are stored in the register file, but the most commonly accessed segments, the data and stack segments, have a copy of their segment bases also stored in a shadow register for input to the adder. Thus the number of read ports to the register file is reduced by the shadow segment register. Less-frequently-used segments require an additional step through the adder to generate the address, but addresses in the data and stack segments are generated in a single cycle.

Mixed-modulo address generation using shadow segment registers

The dispatch unit of a superscalar processor checks for register dependencies among instructions to be issued together as a group. The first instruction's destination register is compared to the following instructions' sources, but the destinations of following instructions are not checked with the first instruction's destination. Instead, instructions with destination-destination dependencies are dispatched together as a group. These instructions flow down the pipelines. At the end of the pipelines the destinations are compared. If the destinations match then the results are merged together and written to the register. When instructions write to only a portion of the register, merging ensures that the correct portions of the register are written by the appropriate instructions in the group. Thus older code which performs partial-register writes can benefit from superscalar processing by dispatching the instructions together as a group and then merging the writes together at the end of the pipelines. The dispatch and decode stage, which is often a critical path on the processor, is reduced in complexity by not checking for destination-register dependencies. Performance increases because more kinds of instructions can be dispatched together in a group, increasing the use of the superscalar features.

Reduced register-dependency checking for paired-instruction dispatch in

A pipelined processor executes several stack instructions simultaneously. Additional shadow registers for stack pointers of instructions in the pipeline are not needed. Instead the new stack pointer is generated once at the end of the pipeline and written to the register file. The stack pointer is needed for generating the stack-top address in memory. The stack-top address is generated early in the pipeline. Other stack instructions in the pipeline which have not yet incremented the stack pointer are located with a stack valid bit array. The stack valid array indicates the increment or decrement amounts for stack instructions in each pipeline stage. An overall displacement or increment value is computed as the sum of all increments and decrements for stack instructions in the pipeline which have not yet updated the stack pointer. The overall displacement which accounts for all unfinished stack instructions is added to the stack pointer from the register file to generate the stack-top address. Thus the new stack pointer does not have to be generated before the stack memory is accessed. Pushes or pops are paired by doubling the increment amount in the stack valid bit array and performing a double-width data transfer.

Growing community of inventors

Mountain View, CA, United States of America

Gene Shen