Quuxplusone
commented
3 years ago I believe this is actually /undefined/ behavior, rather than just unspecified.
For instance https://en.cppreference.com/w/c/language/operator_comparison lists
a series of comparisons that are allowed, and the rest are undefined. The
allowed list includes:
"if one pointer points to the element of an array and the other pointer points
one past the end of the same array, the one-past-the-end pointer compares
greater"
and
"if two pointers point to the same object, or both point one past the end of
the same array, they compare equal"
But there's no other provision for comparison of one-past-the-end pointers.
extern int x[1], y;
int f(int p, int q) { q = y; if (p == (x + 1)) { p = 2; return y; } return 0; }
LLVM trunk currently outputs the following code with -O3 -mtune=znver3 (without it it outputs correct code but I'm pretty sure it still makes the same wrong assumption about the value of
p):f: mov eax, dword ptr [rip + y] mov dword ptr [rsi], eax xor eax, eax cmp rdi, offset x+4 je .LBB0_1 ret .LBB0_1: mov eax, dword ptr [rip + y] mov dword ptr [rip + x+4], 2 ret
Which is incorrect because
pcould point toy, for example iffwas called as such:int whatever; f(&y, &whatever);
and
ycould happen to be located in memory right afterx.Also, although the comparison's result is unspecified, this still means only two results are possible according to the standard:
p == (x + 1)results infalse, then the result offis 0p == (x + 1)results intrue, then the result offis 2 since we do*p = 2andppoints toyLLVM's optimization makes it so the result can also be the previous value of
y, which could be something else than 0 or 2.It seems that LLVM assumes that because
p == (x + 1)it can replace all occurences ofpwithx + 1without any regard to provenance, and doing that change manually would indeed mean thereturn y;could be optimized to use the previous store (and the store tox + 1would be UB, too...), but this isn't the case here:pcould simultaneously validly point toyand be equal tox + 1.Godbolt link: https://godbolt.org/z/v73ME48qd