Architecture Design and Compiler Support for Code Size Optimization in Embedded Processors

Abstract

Embedded processors usually need to satisfy very tight design constraints to achieve low power consumption, small chip area, and high performance. One of the obstacles to meeting these requirements is related to delivering instructions from instruction memory/caches. The size of instruction memory/cache considerably contributes total chip area. Further, frequent access to caches incurs high power/energy consumption and significantly hampers overall system performance due to cache misses. To reduce the negative effects of the instruction delivery, therefore, this study focuses on the sizing of instruction memory/cache through code size optimization.

One observation for code size optimization is that very long instruction word (VLIW) architectures often consume more power and memory space than necessary due to long instruction bit-width. One way to lessen this problem is to adopt a reduced bit-width ISA (Instruction Set Architecture) that has a narrower instruction word length. In practice, however, it is impossible to convert a given ISA fully into an equivalent reduced bit-width one because the narrow instruction word, due to bitwidth restrictions, can encode only a small subset of normal instructions in the original ISA. To explore the possibility of complete conversion of an existing 32-bit ISA into a 16-bit one that supports effectively all 32-bit instructions, we propose the reduced bit-width (e.g. 16-bit × 4-way) VLIW architectures that equivalently behave as their original bit-width (e.g. 32-bit × 4-way) architectures with the help of dynamic implied addressing mode (DIAM).

Second, we observe that code duplication techniques have been proposed to increase the reliability against soft errors in multi-issue embedded systems such as VLIW by exploiting empty slots for duplicated instructions. Unfortunately, all duplicated instructions cannot be allocated to empty slots, which enforces generating additional VLIW packets to include the duplicated instructions. The increase of code size due to the extra VLIW packets is necessarily accompanied with the enhanced reliability. In order to minimize code size, we propose a novel approach compiler-assisted dynamic code duplication scheme, which accepts an assembly code composed of only original instructions as input, and generates duplicated instructions at runtime with the help of encoded information attached to original instructions. Since the duplicates of original instructions are not explicitly present in the assembly code, the increase of code size due to the duplicated instructions can be avoided in the proposed scheme.

Lastly, the third observation is that, to cope with soft errors similarly to the second observation, a recently proposed software-based technique with TMR (Triple Modular Redundancy) implemented on coarse-grained reconfigurable architectures (CGRA) incurs the increase of configuration size, which is corresponding to the code size of CGRA, and thus extreme overheads in terms of runtime and energy consumption mainly due to expensive voting mechanisms for the outputs from the triplication of every operation. To reduce the expensive performance overhead due to the large configuration from the validation mechanism, we propose selective validation mechanisms for efficient modular redundancy techniques in the datapath on CGRA. The proposed techniques selectively validate the results at synchronous operations rather than every operation.