[avr-libc-dev] [bug #22163] Everytime ATOMIC_BLOCK(ATOMIC_RESTORESTATE) compiler generates warning - unused variable 'sreg_save'

[David Brown]

Follow-up Comment #4, bug #22163 (project avr-libc):

I've just tried this code section, and there is a good reason for the "unused
variable 'sreg_save'" warning - looking at the generated assembly, the
variable *is* unused.  In other words, the SREG is not restored after the
block, as it is in the C version.

Could it be that the __attribute__((__cleanup__)) attribute is not working
correctly for avr c++, at least when used in this way?  If that's the case,
then this is definitely a bug.

I've noticed another issue with the atomic.h functions - I don't think the
memory blocks (asm("":::"memory")) are in the right places.  It's important to
ensure that any changes to memory are written out before interrupts are
enabled - thus the block should appear *before* the sei() instructions, and
before restoring SREG.  The block is unnecessary before disabling interrupts. 
I could be wrong, but I *think* that's the correct ordering.

In C++ it is normally considered better to use a class and RAII for something
like this.  You should have a class such as:

class CriticalSection {
public:
    CriticalSection() : _sreg( SREG ) { cli(); }
    ~CriticalSection() { asm("" ::: "memory"); SREG = _sreg; }
private:
    uint8_t _sreg;
};

Then you simply declare a "CriticalSection" variable in the block you want to
protect.  Similar classes can be defined for non-critical sections, and for
when you want to restore the old interrupt enable state, or to force the flag
on or off.

As a workaround for the bug blocking cleanup functions, it's possible to use
a RAII class within the same framework as the C-style atomic blocks.  It's a
little messy, but gives identical optimal assembly code (in my brief testing,
anyway).  I haven't gone through the details of patching this into atomic.h
(I'm not sure what the policy of including c++ specific stuff is), but it can
be added on at the end of atomic.h as it re-defines some macros:

#ifdef __cplusplus

// Override these macros:
#undef ATOMIC_RESTORESTATE
#undef ATOMIC_FORCEON
#undef NONATOMIC_RESTORESTATE
#undef NONATOMIC_FORCEOFF

class RestoreState {
public:
        RestoreState() : _x( SREG ), _restore( true ) {}                        
// For keeping sreg
        RestoreState(uint8_t x) : _x( x ), _restore( false ) {}         // For 
__ToDo loop
hack
        ~RestoreState() { if (_restore) SREG = _x; }
        operator bool() const { return _x; };                                   
        // For testing __ToDo
private:
        uint8_t _x;
        bool _restore;
};

class ForceOn {
public:
        ForceOn() : _x( 0 ), _restore( true ) {}                                
        // For ignoring sreg
        ForceOn(uint8_t x) : _x( x ), _restore( false ) {}              // For 
__ToDo loop hack
        ~ForceOn() { if (_restore) sei(); }
        operator bool() const { return _x; };                                   
        // For testing __ToDo
private:
        uint8_t _x;
        bool _restore;
};

class ForceOff {
public:
        ForceOff() : _x( 0 ), _restore( true ) {}                               
        // For ignoring sreg
        ForceOff(uint8_t x) : _x( x ), _restore( false ) {}             // For 
__ToDo loop
hack
        ~ForceOff() { if (_restore) cli(); }
        operator bool() const { return _x; };                                   
        // For testing __ToDo
private:
        uint8_t _x;
        bool _restore;
};

#define ATOMIC_RESTORESTATE RestoreState sreg_save
#define ATOMIC_FORCEON ForceOn sreg_save
#define NONATOMIC_RESTORESTATE RestoreState sreg_save
#define NONATOMIC_FORCEOFF ForceOff sreg_save

#endif




    _______________________________________________________

Reply to this item at:

  <http://savannah.nongnu.org/bugs/?22163>

_______________________________________________
  Message sent via/by Savannah
  http://savannah.nongnu.org/