Embedded Software for SoC - Grupo de Mecatrônica EESC/USP

Energy-Aware Parameter Passing 511
the same way. However, each access to a formal parameter invokes a fresh
computation of the value of the actual parameter.
Conservative: In this strategy, the multiple uses problem is resolved as in
the case of the Block Insertion strategy. However, if the global variable
problem occurs, this strategy drops on-demand parameter-passing and uses
the original parameter-passing strategy.
Closure: This is similar to the previous strategy except that in handling
the global variable problem, it uses the closure-based mechanism discussed
earlier in Section 3. It should be mentioned that this strategy passes a parameter
as the first (default) strategy if the global variable problem does not
occur.
We observe from Figure 36-4 that the default on-demand strategy improves
energy consumption by 8.78%, on the average. It is most successful in applications
Wood and Encr where there is a relatively large difference between
the average number of parameters and the average number of active parameters.
The Path Control and Multiple Evaluation strategies, on the other hand,
achieve less saving in energy. In fact, in two applications, Atr and SP,
evaluating the same actual parameter more than once results in energy loss
(as compared to the case without on-demand parameter-passing).
The reason that it is relatively successful in remaining applications is the
fact that in many subroutine executions the paths where multiple actual
parameter evaluations occur are not taken at runtime. The reason that the Path
Control strategy is not as successful as the Block Insertion strategy is the extra
energy spent in maintaining the values of the path variables and in executing
comparison operations. The percentage (average) improvements due to Path
Control and Multiple Evaluation strategies are 6.68% and 2.98%, respectively.
Conservative and Closure generate the same energy behavior as the default
strategy in our three applications. This is because in these applications, the
global variable problem does not occur. In Usonic and Wood, on the other
hand, the default strategy outperforms Conservative and Closure.
The trends in Figure 36-5 are similar to those illustrated in Figure 36-4.
This is not surprising as eliminating redundant computation saves both energy
and execution cycles. It should be stressed, however, that the savings in
execution cycles are lower than those in energy consumption. This is due to
the fact that the processor we consider can execute multiple instructions in
the same cycle; consequently, some of the redundant computation can be
hidden (which makes the base case – without on-demand parameter-passing
– appear stronger). However, unlike execution cycles, it is not possible to hide
energy consumption.
5. CONCLUSIONS
In this work, we studied the possibility of modifying the parameter-passing
mechanism of the language with some help from compiler. Using a set of