"Xuantie-910: A Commercial Multi-Core 12-Stage Pipeline Out-of-Order 64-bit High Performance RISC-V Processor with Vector Extension"

From "VIII. NON-STANDARD INSTRUCTION SET EXTENSION"

Targeting at various industrial applications, XT-910 enables a set of custom non-standard instructions, additional to the standard RISC-V instructions. The non-standard instruction extension can be categorized into two groups based on the purposes - memory access enhancement, and basic arithmetic operation enhancement.

A. Memory access enhancement

Memory access instructions usually account for a high proportion in the total number of instructions, so enhancing memory access instructions can directly benefit the overall performance. By analyzing the mainstream applications running on RISC-V, we observed that for the basic RISC-V instructions there is still quite some room for improvement in memory access related instructions.

First, we support register + register addressing mode, and support indexed load and store instructions. This type of instruction extension reduces the usage of the registers for calculation and reduces the number of instructions for address generation, thereby effectively accelerating the data access of a loop body. Second, unsigned extension during address generation is supported. Otherwise, the basic instruction set does not support direct unsigned extension from 32-bit data to 64-bit data, resulting in too many shift instructions.

I'm not certain if they break out the results by individual optimization, but in Fig. 20 in the paper it looks like all their tweaks add up to a 20% boost over the standard RISC-V ISA. Pretty huge.