Question

I implemented a tree in C, and I now want to define a wrapper tree set. I have an iterator for my tree in tree.h:

typedef struct tree_iter_t {
    void *current;
    tree_t *tree;
    unsigned char info : 2;
} tree_iter_t;

and a function to get an iterator in tree.c:

tree_iter_t titerator(tree_t *t) {
    tree_iter_t it;
    it.current = t->min;
    if (t->min) it.info = 0;
    else it.info = 3;
    it.tree = t;
    return it;
}

I can compile this with -Wall -O2 with no warnings. For my tree set, I defined my tree set iterator in tset.h as follows:

typedef struct tset_t tset_t;

typedef struct tset_iter_t {
    tree_iter_t iter;
} tset_iter_t;

and a function to obtain it in tset.c.

struct tset_t {
    tree_t *tree;
};

tset_iter_t tsiterator(tset_t *ts) {
    tset_iter_t it;
    it.iter = titerator(ts->tree);
    return it;
}

When I compile with gcc -Wall -c -combine tset.c tree.c, I have no problems, but when I add -O2, I get a warning on the return statement: warning: ‘it.iter.tree’ is used uninitialized in this function. Why does GCC have a problem with this? Am I missing something obvious? It looks initialized to me. I ran gcc -S -O2 tset.c to try to get a sense of what was going on, GCC gave no warning and produced this:

tsiterator:
pushl   %ebp
movl    %esp, %ebp
pushl   %ebx
subl    $36, %esp
movl    12(%ebp), %edx
movl    8(%ebp), %ebx
leal    -20(%ebp), %eax
movl    (%edx), %edx
movl    %eax, (%esp)
movl    %edx, 4(%esp)
call    titerator
movzbl  -12(%ebp), %edx
movzbl  8(%ebx), %eax
andl    $3, %edx
andl    $-4, %eax
orl     %edx, %eax
subl    $4, %esp
movb    %al, 8(%ebx)
movl    -16(%ebp), %eax
movl    %eax, 4(%ebx)
movl    -20(%ebp), %eax
movl    %eax, (%ebx)
movl    %ebx, %eax
movl    -4(%ebp), %ebx
leave
ret $4

I know optimization can generate some strange-looking code, but what the heck is going on here!? All of my other (optimized) wrapper functions are 10-ish lines of assembly (just the usual function call overhead for calling tree’s functions). gcc -O2 -S -combine tset.c tree.c gave me the warning, inlined titerator, and produced this:

tsiterator:
pushl   %ebp
movl    %esp, %ebp
movl    12(%ebp), %edx
pushl   %ebx
movl    8(%ebp), %eax
movl    (%edx), %ecx
movl    4(%ecx), %edx
movl    %ecx, 4(%eax)
cmpl    $1, %edx
movl    %edx, (%eax)
movzbl  8(%eax), %edx
sbbl    %ebx, %ebx
andl    $3, %ebx
andl    $-4, %edx
orl     %ebx, %edx
movb    %dl, 8(%eax)
popl    %ebx
popl    %ebp
ret $4

When I changed the implementation to:

tset_iter_t tsiterator(tset_t *ts) {
    tset_iter_t it;
    tree_iter_t i = titerator(ts->tree);
    it.iter = i;
    return it;
}

it had no problems. What is GCC optimizing (or analyzing) in the first case, and why is it giving me a warning?

Thanks.

Answer 1

I think the warning is a bug. Which gcc are you using? Doesn’t happen for me when I compile (admittedly a single file) with gcc 4.0 and 4.2.

Here is an annotated version of the optimised assembler. I can’t see anything unassigned here, which is why I think the warning is not right. I’ve guessed about the tree structure.

tsiterator:
    pushl   %ebp
    movl    %esp, %ebp
    movl    12(%ebp), %edx  ; edx has pointer to ts
    pushl   %ebx
    movl    8(%ebp), %eax   ; eax has pointer to retval
    movl    (%edx), %ecx    ; ecx has ts->tree (?)
    movl    4(%ecx), %edx   ; edx has ts->tree->min (?)
    movl    %ecx, 4(%eax)   ; store ts->tree into retval->iter->tree
    cmpl    $1, %edx    
    movl    %edx, (%eax)    ; store ts->tree->min into retval->iter->current
    ;; This and the cmpl instruction above is all
    ;; handling the bitmasking for retval->iter->info.
    ;; Probably would be more efficient to not use a
    ;; bit mask here, as the compiler could remove the
    ;; second "and" and "or" instructions.
    movzbl  8(%eax), %edx   ; get current value of retval->iter->info
    sbbl    %ebx, %ebx  ; ebx = 0 if unsigned edx < 1, else -1
    andl    $3, %ebx    ; mask off new value            
    andl    $-4, %edx   ; mask off old value   
    orl     %ebx, %edx  ; combine old | new    
    movb    %dl, 8(%eax)    ; store combined into retval->iter->info
    popl    %ebx
    popl    %ebp
    ret $4

EDIT: Note that the compiler carefully preserves the upper 6 bits of random uninitialized junk in tree_iter_t.info.