The following code should warn as there is an aliasing problem in the code (derived from gcc in SPEC, yes that old code): struct tt { struct { short i; short j; }t; }; void t(struct tt *i) { ((int*)i)[0] = 0; i->t.j = 2; }
Mine as the reason why this does not happen is because &a->b is lowered so that there is only a cast.
Actually I was wrong, the reason why it does not warn is because we do not take the address so the warning is not going to be emitted at all. I am not working on this any more.
But the reason why do not even think about warning has to do with the lowering part though.
*** Bug 20189 has been marked as a duplicate of this bug. ***
Even -Wstrict-aliasing=2 does not warn.
The strict aliasing rule requires that An object shall have its stored value accessed only by an lvalue expression that has one of the following types ... The test case in comment #0 doesn't involve accessing the stored value of an object so it doesn't violate those rules. The following slightly modified version of the test case that does access the stored value of an object by an incompatible type and thus violate those rules is also not diagnosed, even by GCC 7. $ cat t.c && gcc -O2 -S -Wall -Wextra -Wpedantic -Wstrict-aliasing=3 -fstrict-aliasing -fdump-tree-optimized=/dev/stdout t.c struct tt { struct { short i, j; } t; }; int t (struct tt *i) { i->t.i = 1; i->t.j = 2; return *(int*)i; // violation here } ;; Function int t(tt*) (_Z1tP2tt, funcdef_no=0, decl_uid=2282, cgraph_uid=0, symbol_order=0) int t(tt*) (struct tt * i) { int _5; <bb 2> [100.00%]: MEM[(short int *)i_2(D)] = 131073; _5 = MEM[(int *)i_2(D)]; return _5; }