GCC 10 Inlining behavior breaks RT test 12-7 Topic is solved

[mtrescott]

Hi Giovanni, thanks for finding the bug, wish I could have been of more help but I didn't have time to learn what I'd need to know to investigate it myself.

[Giovanni]

Hi,

Don't worry, I am just evaluating various ways to fix/workaround that problem.

Giovanni

[Giovanni]

Hi,

Update, this has been reported as bug #1138.

I committed an initial fix on trunk and it appears to work correctly however:

1) The list (threads and virtual timers) handling in RT has been reorganized to avoid aliasing problems, this affected a lot of code in RT and OSLIB.
2) It has not yet been tested under all possible conditions, I know there are configuration options that do not compile correctly yet.
3) Quality procedures need to be performed again because the extensive changes.

On the other hand:

1) Lists-related code is more elegant now, the number for pointer casts has been reduced dramatically which is generally a good thing.
2) Gave me an opportunity to re-look at code not touched for many years.

I only tested on the STM32G4 but the test suite succeeds now. Please report any problem you may find.

After trunk is stable again the change will be back-ported on 19.x and 20.x (and it is a lot of work).

Giovanni

[FXCoder]

Hi Giovanni,
Moved the latest trunk into a project quickly this morning to do some checking.
Found the following...
#1 chvt.c casts for delta list manipulations

Code: Select all

Index: chvt.c
===================================================================
--- chvt.c   (revision 13960)
+++ chvt.c   (working copy)
@@ -380,10 +380,10 @@
       virtual_timer_t *vtp;
       vtfunc_t fn;
 
-      vtp = vtlp->dlist.next;
+      vtp = (virtual_timer_t *)vtlp->dlist.next;
       fn = vtp->func;
       vtp->func = NULL;
-      vtp->dlist.next->prev = (virtual_timer_t *)vtlp;
+      vtp->dlist.next->prev = (delta_list_t *)vtlp;
       vtlp->dlist.next = vtp->dlist.next;
       chSysUnlockFromISR();
       fn(vtp->par);



#2 Binary semaphore compile fails
This is likely one of the options that do not compile correctly that you referred to.
Didn't have time to dig into it.

Code: Select all

os/oslib/include/chbsem.h:92:4: warning: implicit declaration of function '_SEMAPHORE_DATA'; did you mean 'SEMAPHORE_DECL'? [-Wimplicit-function-declaration]
   92 |   {_SEMAPHORE_DATA(name.sem, ((taken) ? 0 : 1))}
      |    ^~~~~~~~~~~~~~~
 

--
Bob

[Giovanni]

Hi Bob,

I meant exactly those kind of errors, things hidden by configuration switches, fixed.

Giovanni

[FXCoder]

Hi,
Here is another...
Set CH_CFG_OPTIMIZE_SPEED FALSE
...to reveal a few more compile problems in chschd.c & chmtx.c (plus a type change in syscalls.c).
--
Bob

[Giovanni]

Fixed thanks, committed more queues reorganization, there is no reason to hold back changes at this point.

Giovanni

[Giovanni]

Hi,

Fixed as bug bug #1138 in trunk, 20.x and 19.x.

Giovanni