AMD Typewriter x86 사용자 설명서

다운로드
페이지 256
20
Consider Expression Order in Compound Branch Conditions
AMD Athlon™ Processor x86 Code Optimization 
22007E/0—November 1999
Consider Expression Order in Compound Branch 
Conditions
B ra n ch   c o n d it i o n s   i n   C   p rog ra m s   a re   o f t e n   c o m po u n d
conditions consisting of multiple boolean expressions joined by
the boolean operators && and ||. C guarantees a short-circuit
evaluation of these operators. This means that in the case of ||,
the fi rs t operan d to eval uate to TRUE  term inates the
evaluation, i.e., following operands are not evaluated at all.
Similarly for &&, the first operand to evaluate to FALSE
terminates the evaluation. Because of this short-circuit
evaluation, it is not always possible to swap the operands of ||
and &&. This is especially the case when the evaluation of one
of the operands causes a side effect. However, in most cases the
exchange of operands is possible.
When used to control conditional branches, expressions
involving || and && are translated into a series of conditional
branches. The ordering of the conditional branches is a function
of the ordering of the expressions in the compound condition,
and can have a significant impact on performance. It is
unfortunately not possible to give an easy, closed-form formula
on how to order the conditions. Overall performance is a
function of a variety of the following factors:
probability of a branch mispredict for each of the branches 
generated
additional latency incurred due to a branch mispredict
cost of evaluating the conditions controlling each of the 
branches generated
amount of parallelism that can be extracted in evaluating 
the branch conditions
data stream consumed by an application (mostly due to the 
dependence of mispredict probabilities on the nature of the 
incoming data in data dependent branches)
It is therefore recommended to experiment with the ordering of
expressions in compound branch conditions in the most active
areas of a program (so called hot spots) where most of the
execution time is spent. Such hot spots can be found through
the use of profiling. A "typical" data stream should be fed to
the program while doing the experiments.