This occurred because the code to handle the function-like macro use would read the following token, which could prompt processing of the following preprocessing directive in an inappropriate context.
The following example illustrates the problem (the error message would be printed):
#define A()
#define FOO 1
A()
#if !FOO
# error "shouldn't get here"
#endif
Previously, such initializations would sometimes generate a garbage value pointing up to 65535 bytes beyond the start of the string constant. (This was due to a lack of sign-extension in the object code generation.)
Computing a pointer to before the start of an object invokes undefined behavior, so the previous behavior wasn't technically wrong, but it was unintuitive and served no useful purpose. The new behavior should at least be easier to understand and debug.
The latter would require more changes to the code generator to understand it, whereas this approach doesn't require any changes. This is arguably less clean, but it matches other places where a byte value is subsequently operated on as a word without an explicit conversion, and the assembly instruction generated is the same.
This fixes the compca06.c test case.
Note that this generates inefficient code in the case of loading a signed byte value and then immediately casting it to unsigned (it first sign-extends the value, then masks off the high bits). This should be optimized, but at least the generated code is correct now.
This fixes the compca22.c test case.
This optimization could be fixed and re-enabled, but to do so, you would have to check if the stored value is ever used subsequently, which is not information that's readily available in the peephole optimization pass. It would also be necessary to check if there are any stores to the same location within the right-side expression, which could kill the optimization.
