jmtd → software → endian

Problem: You have code which needs to deal with endian issues. E.g., you might need to byte-swap only on big or little-endian.

Practical example: integers in doom WAD files are little-endian, so you would conditionally byte-swap on big-endian machines.

How would you do it?

Approaches

Conditionals via C pre-processor

a GNU automake/configure/etc. script can give you a CPP define for the endianness. You can then conditionally #ifdef the code.

Problem: You break the endian-specific code, somehow, but never notice because your local architecture means it's not used on your architecture.

conditionals via C-level logic

How about defining the endian-checking logic in the C itself? That way the endian code is always compiled, so if you break it in a way which can be determined at compile-time you find out even if your local platform is not going to use that code.

#include <stdio.h>
int main() {
       int i = 1;
       int *p = &i;
       if(1 == *(char*)p) printf("little-endian\n");
       else               printf("big-endian\n");
       return 0;
}

The disadvantage here would be that you may have your binary full of redundant code.

However, the test is a constant expression, so a compiler should throw away the redundant branch. To find out if this happens, I conducted a small experiment.

Results

GCC

Using GCC's -O flag, setting 0, the redundant branch is present in the output. x86 assembly excerpts:

...
.LC0:
       .string "little-endian\n"
.LC1:
       .string "big-endian\n"
...
       movl    $.LC0, (%esp)
       call    printf
       jmp     .L3
.L2:
       movl    $.LC1, (%esp)
       call    printf
.L3:
...

Using the first level of optimisation, -O1, and the redundant branch is gone:

...
.LC0:
       .string "little-endian"
...
       andl    $-16, %esp
       movl    $.LC0, (%esp)
       call    puts
...

These tests used GCC version 3.2.2.

Sun's CC on SPARC

Sun's C compiler does not throw away the redundant branch at any optimizer level. I am using "WorkShop Compilers 5.0 98/12/15 C 5.0".

The non-conditional approach

Missing from this article at the moment...