1. Reducing Code Management Overhead in Software-Managed Multicores
Jian Cai, Yooseong Kim, Youngbin Kim, Aviral Shrivastava, Kyoungwoo Lee
Compiler Microarchitecture Lab
Arizona State University

2. Motivation Software managed multicore (SMM) architectures
Have scratchpad memories (SPMs) instead of caches as on-chip memory
Consumes significant less area and power but require explicit data movement
Management overhead and must be minimized
Code management divides SPM space into memory regions, and map functions into the regions
Previous works focus on finding better mapping to reduce data transfers caused by conflicts among functions in the same region, but blindly insert management functions around every function call
This paper reduces computation by avoiding unnecessary management calls by
an always-hit analysis to identify functions that are always present in SPM
a first-miss analysis to identify functions that are always present in SPM within loops
_get(main) is not necessary since main is the only function in region 0
_get(F1) is hoisted since F1 is the only function in region 1 called within the loop
We have developped and are developping management techniques for other data segments, but this paper focuses on code management.
The always-hit and first-miss analyses are both iterative data-flow analysis

3. Result
14% improvement on average
9% improvement
on average