The size of atomic object does not match with documentation - https://gcc.gnu.org/wiki/Atomic/GCCMM/UnalignedPolicy specifically : "GCC 4.9 will provide an atomic type attribute which can be set on objects. This attribute will force a specific alignment and size on the object which may be different than the original data type. The alignment and size will attempt to provide lock free operations, if they exist. ... An object will be promoted up to the next lock-free size in order to enable lock free operations, as long as it isn't already a documented lock free size. " ~/atomic_test$ cat non_power_2_atomic.c #include <stdatomic.h> #include <stdio.h> typedef struct { char c [5]; } non_power_to_obj; int main ( void ) { non_power_to_obj obj; _Atomic non_power_to_obj aobj; printf("\n Size and Alignment usual object "); printf(" : sizeof(obj) %d __alignof__(obj) %d ", sizeof(obj), __alignof__(obj) ); printf("\n Size and Alignment of atomic object "); printf(" : sizeof(aobj) %d __alignof__(aobj) %d \n", sizeof(aobj), __alignof__(aobj) ); return 0; } ~/atomic_test$ gcc -v Using built-in specs. COLLECT_GCC=gcc COLLECT_LTO_WRAPPER=/opt/gcc/libexec/gcc/i386-pc-solaris2.11/4.9.2/lto-wrapper Target: i386-pc-solaris2.11 Configured with: ./configure --prefix=/opt/gcc/ Thread model: posix gcc version 4.9.2 (GCC) ~/atomic_test$ gcc -O -latomic -std=c11 non_power_2_atomic.c ~/atomic_test$ ./a.out Size and Alignment usual object : sizeof(obj) 5 __alignof__(obj) 1 Size and Alignment of atomic object : sizeof(aobj) 5 __alignof__(aobj) 1 According to the https://gcc.gnu.org/wiki/Atomic/GCCMM/UnalignedPolicy the size and alignment of "aobj" should be 8. The bug us found on Solaris x86, but it could be on other platforms also(SPARC/Linux). This bug is also exist in g++.
On Fri, 20 Feb 2015, alexey.lapshin at oracle dot com wrote: > The size of atomic object does not match with documentation - > https://gcc.gnu.org/wiki/Atomic/GCCMM/UnalignedPolicy specifically : That's not documentation. It's an early description of intent that does not correspond to what was actually implemented (note the future tense in the words you quote, and the last edited date a year before I added _Atomic support to trunk based on Andrew's work). In general, old wiki pages are left around as they may help understand archived development discussions.
Hi Joseph, Could you help me with a link to the correct description of atomic ABI, which in fact used by gcc/g++, please ? Thank you, Alexey.
On Fri, 20 Feb 2015, alexey.lapshin at oracle dot com wrote: > Hi Joseph, > > Could you help me with a link to the correct description of atomic ABI, > which in fact used by gcc/g++, please ? I don't believe it's documented, but I think the general rule for C is: if a type's size is 1, 2, 4, 8, 16, its alignment is increased by _Atomic to that of the atomic integer type of that size, which is the same as that of the non-atomic integer type unless increased by TARGET_ATOMIC_ALIGN_FOR_MODE; the size is never increased by _Atomic, and alignment of types not of those sizes is unchanged. If the code differs, likely the code should take precedence over that description. (But of course if a type is under-aligned for the instructions used to operate on it, there's a bug, whether in the alignment or in the instruction choice.)
Not a bug; there is no intent to increase object size with _Atomic.