Compiler bug

[hto]

It seems that MinGW GCC-3.4.5 compiler disregards alignment attributes. It can lead (and I think actually leads) to bugs in ReactOS.

For example:

Code: Select all

#include <stdio.h>

#define ALIGNED(n)  __attribute__((aligned(n)))

typedef ALIGNED(16) struct
{
    int i;
} S;

int main(void)
{
    S s;
    printf("%p\n", &s); // I have 0022FF74
}

[Karl]

Are we still using GCC-3.4.5?

In all tests I've done, GCC-4.1.0 has compiled faster, and produced better code.

[Christoph_vW]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16660

[hto]

This code works well:

Code: Select all

#include <stdio.h>

#define ALIGNED(n)  __attribute__((aligned(n)))

typedef struct
{
    int i ALIGNED(16);
} S;

int main(void)
{
    S s;
    printf("%p\n", &s); // 0022FF60
}

16-byte alignment should be used in ExfAcquirePushLockExclusive() and ExfAcquirePushLockShared(), EX_PUSH_LOCK_WAIT_BLOCK WaitBlock structure. Misalignment can cause crash / hang, I believe.

Actually this is just a missed diagnostic. The compiler cannot align the stack variables where the alignment is greater than stack alignment that the compiler can give for the stack.

I do not agree. Stack variable can be aligned even by hand.

[GreatLord]

we are thinking of slash gcc 3.4.x and go for 4.2.x for it seam align work in that.

the align are also broken in gcc 4.1.x for complex code. simple code as u show hto works in gcc 4.1.x but when the code start getting complex it break the align.

we are investing how this should be solv at irc channel

[hto]

Simple, but ugly looking solution is:

VOID
FASTCALL
ExfAcquirePushLockShared(PEX_PUSH_LOCK PushLock)
{
char xxx[sizeof(EX_PUSH_LOCK_WAIT_BLOCK) + 16];

#define WaitBlock \
(*(PEX_PUSH_LOCK_WAIT_BLOCK)(((unsigned)buf + 12) & 0xfffffff0))

...

#undef WaitBlock
}

[GreatLord]

after long testing night with other devs. it seam the lates trunk of gcc 4.1.2 is working pretty good. and align is working there.

we will problare slash gcc 3.4.x as supported compiler and move to gcc 4.1.2 with in some days I think. we for see how it end.

it will be small fork of gcc 4.1.2 for offical gcc 4.1.2 can not build reactos we need some patcher for it that we are merge in by hand.

[Karl]

good, then we should have q better quality 0.3.1

[GreatLord]

a new RosBE will be made
with gcc 4.1.2 and gcc 4.2 and alex intruduse a hack for gcc 3.4.x so align work in puch lock. it is only there we are using align now what I am aware. The new RosBe will problare be avail later today.


This issue are solv now.