WISC-SP08 Microarchitecture Specification

WISC-SP08 Microarchitecture Specification

The WISC-SP08 architecture that you will design for the final project shares many resemblances to the MIPS R2000 described in the text. The major differences are a smaller instruction set and 16-bit words for the WISC-SP08. Similarities include a load/store architecture and three fixed-length instruction formats.

1.  Registers

There are eight user registers, R₀-R₇. Unlike the MIPS R2000, R₀ is not always zero. Register R₇ is used as the link register for JAL or JALR instructions. The program counter is separate from the user register file. A special register named EPC is used to save the current PC upon an exception or interrupt invocation.

2.  Memory System

The WISC-SP08 is a Harvard architecture, meaning instructions and data are located in different physical memories. It is byte-addressable, word aligned (where a word is 16 bits long), and big-endian. The final version of the WISC-SP08 will include a multi-cycle memory and level-1 cache. However, initial versions of the machine will contain a single cycle memory. See the project deadlines for more details.

The WISC-SP08 cache replacement policy is deterministic. See the cache module description for an outline of the algorithm you must use.

3.  Pipeline

The final version of the WISC-SP08 contains a five stage pipeline identical to the MIPS R2000. The stages are:

1. Instruction Fetch (IF)
2. Instruction Decode/Register Fetch (ID)
3. Execute/Address Calculation (EX)
4. Memory Access (MEM)
5. Write Back (WB)

See Figure 6.17 on page 395 of the text for a good starting point.

4.  Optimizations

Your goal in optimizations is to reduce the CPI of the processor or the total cycles taken to execute a program. While the primary concern of the WISC-SP08 is correct functionality, the architecture must still have a reasonable clock period. Therefore, you may not have more than one of the following in series during any stage:

• register file
• memory or cache
• 16-bit full adder
• barrel shifter

You may implement any type of optimization to reduce the CPI. The required optimizations are:

• There are two register forwarding paths in the WISC-SP08; one within the ID stage and between the beginning of MEM and the beginning of EX.
• All branches should be predicted not-taken. This means that the pipeline should continue to execute sequentially until the branch resolves, and then squash instructions after the branch if the branch was actually taken.

5.  Exceptions: extra credit

Exception handling is extra credit. If you choose not to implement exception handling, an illegal instruction should be treated as a nop.

IllegalOp is the only defined exception in the WISC-SP08 architecture. It is invoked when the opcode of the currently executing instruction is not a recognized member of the ISA. Upon finding an illegal opcode, the computer shall save the current PC into the reserved register EPC and then load address 0x02, which is the location of the IllegalOp exception handler. Note that if you choose to implement exceptions, address 0x00 must be a jump to the start of the main program.

The exception handler itself need not be complex. At a minimum it should load the value 0xBADD into R₇ and then call the RTI instruction.

6.  Simple re-order buffer: extra credit

Instead of writing the results from the writeback stage to the register file, write it to a temporary structure called the reorder buffer. You may build this reorder buffer using the register file module you already built.

You must augment your decode stage to look in this reorder buffer structure for the latest values in addition to the register file.

• When this re-order buffer becomes full transfer one entry to the actual register file. When the halt instruction is executed, transfer all valid entries into the actual register file.
• Now use this extension to predict values! On every cache miss for a load, simply predict it is zero and continue execution. When the value comes back, check if it is zero, if so you do not need to any any correct. If the value comes back to be different, then clear the values in the reorder buffer that follow the store, and redo execution of instructions that followed the load. If the re-order buffer becomes full, before the load returns, stall the processor until the load returns.