../images/CV32E41P_Pipeline.png

Figure 2 CV32E41P Pipeline

Pipeline Details

CV32E41P has a 4-stage in-order completion pipeline, the 4 stages are:

Instruction Fetch (IF)

Fetches instructions from memory via an aligning prefetch buffer, capable of fetching 1 instruction per cycle if the instruction side memory system allows. See Instruction Fetch for details.

Instruction Decode (ID)

Decodes fetched instruction and performs required register file reads. Jumps are taken from the ID stage.

Execute (EX)

Executes the instructions. The EX stage contains the ALU, Multiplier and Divider. Branches (with their condition met) are taken from the EX stage. Multi-cycle instructions will stall this stage until they are complete. The ALU, Multiplier and Divider instructions write back their result to the register file from the EX stage. The address generation part of the load-store-unit (LSU) is contained in EX as well.

Writeback (WB)

Writes the result of Load instructions back to the register file.

Multi- and Single-Cycle Instructions

Table 3 shows the cycle count per instruction type. Some instructions have a variable time, this is indicated as a range e.g. 3..35 means that the instruction takes a minimum of 3 cycles and a maximum of 35 cycles. The cycle counts assume zero stall on the instruction-side interface and zero stall on the data-side memory interface.

Table 3 Cycle counts per instruction type

Instruction Type

Cycles

Description

Integer Computational

1

Integer Computational Instructions are defined in the RISCV-V RV32I Base Integer Instruction Set.

CSR Access

4 (mstatus, mepc, mtvec, mcause, mcycle, minstret, mhpmcounter*, mcycleh, minstreth, mhpmcounter*h, mcountinhibit, mhpmevent*, dscr, dpc, dscratch0, dscratch1, privlv)

1 (all the other CSRs)

CSR Access Instruction are defined in ‘Zicsr’ of the RISC-V specification.

Load/Store

1

2 (non-word aligned word transfer)

2 (halfword transfer crossing word boundary)

4 (cv.elw)

Load/Store is handled in 1 bus transaction using both EX and WB stages for 1 cycle each. For misaligned word transfers and for halfword transfers that cross a word boundary 2 bus transactions are performed using EX and WB stages for 2 cycles each. A cv.elw takes 4 cycles.

Multiplication

1 (mul)

5 (mulh, mulhsu, mulhu)

CV32E41P uses a single-cycle 32-bit x 32-bit multiplier with a 32-bit result. The multiplications with upper-word result take 5 cycles to compute.

Division

Remainder

3 - 35

3 - 35

The number of cycles depends on the divider operand value (operand b), i.e. in the number of leading bits at 0. The minimum number of cycles is 3 when the divider has zero leading bits at 0 (e.g., 0x8000000). The maximum number of cycles is 35 when the divider is 0

Jump

2

3 (target is a non-word-aligned non-RVC instruction)

Jumps are performed in the ID stage. Upon a jump the IF stage (including prefetch buffer) is flushed. The new PC request will appear on the instruction-side memory interface the same cycle the jump instruction is in the ID stage.

Branch (Not-Taken)

1

Any branch where the condition is not met will not stall.

Branch (Taken)

3

4 (target is a non-word-aligned non-RVC instruction)

The EX stage is used to compute the branch decision. Any branch where the condition is met will be taken from the EX stage and will cause a flush of the IF stage (including prefetch buffer) and ID stage.

Instruction Fence

2

3 (target is a non-word-aligned non-RVC instruction)

The FENCE.I instruction as defined in ‘Zifencei’ of the RISC-V specification. Internally it is implemented as a jump to the instruction following the fence. The jump performs the required flushing as described above.

Hazards

The CV32E41P experiences a 1 cycle penalty on the following hazards.

  • Load data hazard (in case the instruction immediately following a load uses the result of that load)

  • Jump register (jalr) data hazard (in case that a jalr depends on the result of an immediately preceding instruction)