------- Comment #2 from aoliva at gcc dot gnu dot org 2008-06-26 05:12 ------- It is indeed documented that way, but one gets to wonder if that's desirable. Consider that in the original testcase the pointer is converted to a physical address (integral type) that must not be dereferenced, and that integral type is then passed on to a function, that calls another function, that ultimately get the integral-typed arguments into specific registers for a syscall.
The function containing the asm statement can no longer tell which arguments are pointers and which aren't, because none of them are. But, as it turns out, GCC has inlined the function containing the asm statement, as well as its caller, into the function that initialized the data structure and (in this case) converted the pointer to a physical address. If it weren't for the inlining, the alias analysis code would have kicked in and noticed that the pointer escaped in the integral argument. But because everything was inlined, no escape was detected, and the asm statement was all that referenced the pointer-then-integer, and because it was an asm statement rather than a function call, things went downhill. Why shouldn't we use the same machinery we use to notice escapes in function calls to notice escapes in asm statements? FWIW, there is a work-around: add a do-nothing asm statement that "accesses" the memory at the time it is converted to a physical address, but that's, well, not true, and lying to the compiler always comes back and bite us later. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36639
