This paper evaluates the code efficiency of the ARM, Java, and x86 instruction sets by compiling the SPEC CPU95/CPU2000/JVM98 and CaffeineMark benchmarks, from the aspects of code sizes, basic block sizes, instruction distributions, and average instruction lengths. As a result, mainly because (i) the Java architecture is a stack machine, (ii) there are only four local variables which can be accessed by a 1-byte instruction, and (iii) additional instructions are provided for the network security, the code efficiency of Java turns out to be inferior to that of ARM Thumb. Moreover, through this efficiency analysis it should be stressed that there exists the high potential of constructing a more efficient code architecture by taking minute account of the customization of an instruction set as well as the number of registers.

Guided Scrambling (GS) is used for control of the runlength within code blocks, such as d or k, as well as for DC component suppression. A code designed by the GS technique, called a weakly constrained code, does not strictly guarantee the imposed k-constraint, but rather generates code blocks that violate the prescribed constraint with very low probability. In this case, the code rate and efficiency become very high, compared with typical RLL codes using a small constrained length. In this paper, weakly constrained codes based on the convolutional GS and GF-addition GS generate the weakly k-constraint sequences. The probability that a code block violates the k-constraint is measured. To show the superior performance of the GS, the occurrence probability of each runlength is also investigated and compared with the 24/25(0, 8) block code which has a high code rate and adheres to channel constraints. We also compare it with the runlength distribution of a maxentropic RLL sequence and show that the statistical property of the GS-encoded sequences is similar to that of the maxentropic RLL sequence on runlength distribution.