Re: gcc behavior of init priority of .ctors and .dtors section

From: Zhenlei Huang <zlei_at_FreeBSD.org>
Date: Fri, 17 May 2024 01:28:13 UTC

> On May 17, 2024, at 2:26 AM, Konstantin Belousov <kostikbel@gmail.com> wrote:
> 
> On Thu, May 16, 2024 at 08:06:46PM +0800, Zhenlei Huang wrote:
>> Hi,
>> 
>> I'm recently working on https://reviews.freebsd.org/D45194 and got noticed
>> that gcc behaves weirdly.
>> 
>> A simple source file to demonstrate that.
>> 
>> ```
>> # cat ctors.c
>> 
>> #include <stdio.h>
>> 
>> __attribute__((constructor(101))) void init_101() { puts("init 1"); }
>> __attribute__((constructor(65535))) void init_65535() { puts("init 3"); }
>> __attribute__((constructor)) void init() { puts("init 4"); }
>> __attribute__((constructor(65535))) void init_65535_2() { puts("init 5"); }
>> __attribute__((constructor(65534))) void init_65534() { puts("init 2"); }
>> 
>> int main() { puts("main"); }
>> 
>> __attribute__((destructor(65534))) void fini_65534() { puts("fini 2"); }
>> __attribute__((destructor(65535))) void fini_65535() { puts("fini 3"); }
>> __attribute__((destructor)) void fini() { puts("fini 4"); }
>> __attribute__((destructor(65535))) void fini_65535_2() { puts("fini 5"); }
>> __attribute__((destructor(101))) void fini_101() { puts("fini 1"); }
>> 
>> # clang ctors.c && ./a.out
>> init 1
>> init 2
>> init 3
>> init 4
>> init 5
>> main
>> fini 5
>> fini 4
>> fini 3
>> fini 2
>> fini 1
>> ```
>> 
>> clang with the option -fno-use-init-array and run will produce the same result, which
>> is what I expected.
> Why do you add that switch?

gcc13 in ports is not configured with option --enable-initfini-array then it only produces .ctors / .dtors sections but
not .init_array / .fini_array sections. So I add that switch for clang to produce `.ctors` sections instead as
a baseline ( .ctors produced by clang indeed works as expected, the same with .init_array ).

> 
>> 
>> gcc13 from ports
>> ```
>> # gcc ctors.c && ./a.out
>> init 1
>> init 2
>> init 5
>> init 4
>> init 3
>> main
>> fini 3
>> fini 4
>> fini 5
>> fini 2
>> fini 1
>> ```
>> 
>> The above order is not expected. I think clang's one is correct.
>> 
>> Further hacking with readelf shows that clang produces the right order of
>> section .rela.ctors but gcc does not.
>> 
>> ```
>> # clang -fno-use-init-array -c ctors.c && readelf -r ctors.o | grep 'Relocation section with addend (.rela.ctors)' -A5 > clang.txt
>> # gcc -c ctors.c && readelf -r ctors.o | grep 'Relocation section with addend (.rela.ctors)' -A5 > gcc.txt
>> # diff clang.txt gcc.txt
>> 3,5c3,5
>> < 000000000000 000800000001 R_X86_64_64         0000000000000060 init_65535_2 + 0
>> < 000000000008 000700000001 R_X86_64_64         0000000000000040 init + 0
>> < 000000000010 000600000001 R_X86_64_64         0000000000000020 init_65535 + 0
>> ---
>>> 000000000000 000600000001 R_X86_64_64         0000000000000011 init_65535 + 0
>>> 000000000008 000700000001 R_X86_64_64         0000000000000022 init + 0
>>> 000000000010 000800000001 R_X86_64_64         0000000000000033 init_65535_2 + 0
>> ```
>> 
>> The above show clearly gcc produces the wrong order of section `.rela.ctors`.
>> 
>> Is that expected behavior ?
>> 
>> I have not tried Linux version of gcc.
> Note that init array vs. init function behavior is encoded by a note added
> by crt1.o.  I suspect that the problem is that gcc port is built without
> --enable-initfini-array configure option.