[sdcc-devel] [ sdcc-Bugs-597681 ] Optimiser bug

[<no...@so...>]

Bugs item #597681, was opened at 2002-08-20 10:05
You can respond by visiting: 
https://sourceforge.net/tracker/?func=detail&atid=100599&aid=597681&group_id=599

Category: C-Front End
Group: None
Status: Open
Resolution: Remind
Priority: 5
Submitted By: Nick Craig-Wood (ncw)
Assigned to: Bernhard Held (bernhardheld)
Summary: Optimiser bug

Initial Comment:
This is a rather complicated bug!  I've tried to prune
the example
program down to the minimum but there are still a few
seemingly
irrelevant lines in there.

It seems that the correct optimisation in f1() causes
an incorrect
optimisation in f2().  Note that I haven't been able to
trigger the
incorrect optimisation without having f1() before f2().
 Note also
that if you remove any of the other function calls the
optimisation
bug goes away.

There is also a mis-reporting of "use of uninitialised
variable" in f1()
which may be related or may not.

This was tried on an sdcc built from CVS today under linux.

See below for the exact problem.

#define MICROCONTROLLER_P80C552 1
#include <mcs51reg.h>

typedef unsigned char u8;
typedef unsigned int u16;
extern void menu_status(u8 message);
extern u8 pio_a_get(u8 mask);
extern void pio_b_set(u8 on, u8 mask);
extern void sleep_cs(u16 cs);
extern void error_display(u8 error);
extern void timeout_start(u16 cs);

u8 f1(void)
{
    u8 error = 0;
    menu_status(39);
    timeout_start(400);
    pio_b_set(1,(1<<6));
    if (error)
        error_display(error);
    else
        menu_status(38);
    return error;

// This is a correct optimisation - error is always 0
and sdcc
// optimises away the if statement.  There is a
spurious warning about
// error though.  However this optimisation is needed
to trigger the
// real bug in the next function
// bug3.c:23: warning: conditional flow changed by
optimizer 'bug3.c(19)':so said EVELYN the modified DOG
// bug3.c:23: warning: 'auto' variable 'error' may be
used before initialization at bug3.c(20)
// bug3.c:23: warning: unreachable code bug3.c(20)
}


static u8 f2(void)
{
    menu_status(104);
    pio_b_set(0,(1<<2));
    sleep_cs(100);

// Mis-optimisation - sdcc optimises this if statement
away completely
// which it can't because P5 is an IO port.  It
requires the previous
// function and all the extra function calls.
    if ((P5 & (1<<3)))
        return 18;

    return 0;
// bug3.c:47: warning: conditional flow changed by
optimizer 'bug3.c(44)':so said EVELYN the modified DOG
// bug3.c:47: warning: unreachable code bug3.c(45)
}


----------------------------------------------------------------------

>Comment By: Nick Craig-Wood (ncw)
Date: 2002-08-26 16:55

Message:
Logged In: YES 
user_id=220462

Sorry!  Here are the compiler options I've been using

  sdcc -mmcs51 --model-large --stack-after-data

If you remove the --model-large then the bug isn't exhibited.

----------------------------------------------------------------------

Comment By: Bernhard Held (bernhardheld)
Date: 2002-08-24 11:52

Message:
Logged In: YES 
user_id=203539

Not reproducable. Did you use any compiler-options? 
 
;       ncw.c:48: if ((P5 & (1<<3)))  
;       genAnd  
        mov     a,_P5  
;       genIfxJump  
; Peephole 111   removed ljmp by inverse jump logic  
        jnb  acc.3,00102$  
00106$:  
;       ncw.c:49: return 18;  
;       genRet  
        mov     dpl,#0x12  
; Peephole 132   changed ljmp to sjmp  
        sjmp 00103$  
00102$:  
;       ncw.c:51: return 0;  
;       genRet  
        mov     dpl,#0x00  
00103$:  
        ret  
  

----------------------------------------------------------------------

You can respond by visiting: 
https://sourceforge.net/tracker/?func=detail&atid=100599&aid=597681&group_id=599