Recently, with the broad adoption of mobile devices, considerable research efforts
have concentrated on innovative dynamic optimization techniques to improve
the performance and energy efficiency of applications running
on these resource-constrained devices, in order to meet the needs for better user
experience, more functionalities, and reduced costs. In this dissertation, we
explore the effectiveness, on performance and energy efficiency, of
optimizations that are made possible when a Just-In-Time compiler can
directly compile bytecodes to the internal, flexibly-designed,
implementation ISA, rather than to the external, architectural ISA of
a microprocessor.
We discuss the challenges and present the design and implementation of
a novel acceleration framework to improve the performance-per-watt of
applications written in modern object oriented languages that execute
on managed runtime environments of mobile devices. Our acceleration
framework allows the direct translation of bytecodes into the
implementation ISA of modern co-designed processors, by a JIT
compiler, bypassing the architectural ISA. We develop novel JIT
speculative optimizations by leveraging both the semantically rich
bytecode and the hardware features at the microarchitecture level,
such as hardware support for: a) speculative execution, b) reduced
indirect branches mispredictions, c) efficient call/return, and d)
additional registers, which are not exposed by the microprocessor's
architecture ISA for reasons of ISA compatibility.
We demonstrate the effectiveness of our acceleration interface in
delivering increased performance-per-watt on mobile processors by
providing an in-depth evaluation of two implementations: 1) first,
using a real in-order co-designed processor; and 2) second, using a
modern out-of-order co-designed research processor for the x86
architecture, running on a cycle accurate state-of-art simulator. Our
framework provides the building blocks for the design of innovative
hardware support and speculative optimizations to effectively deliver
the stringent energy and performance requirements of applications running
on mobile processors.
