Re: [avr-gcc-list] AVR GCC Version 3.3 20030421 - taget mega128

[Denis Chertykov]

"Peter A. Zuur" <address@hidden> writes:

> I have found a bug in the AVR-GCC compiler (target mega128) when used with 
> any other optimization level than -O0 .
> 
> The following function which updates a crc-8 with <octet> always returns 0 
> when compiled with optimization switched on. 
> The reason for that is that the compiler erroneously optimizes the <then> 
> clause of the <if> statement out. 
> 
> unsigned char crc8(unsigned char crc, unsigned char octet)
> {
>     unsigned char j, tmp;
> 
>     j= 0;
>     while(j < 8)
>     {
>         tmp= crc ^ octet;
>         if(tmp & 0x01)      // optimized out
>         {
>             crc^= 0x18;     // optimized out
>             crc>>= 1;       // optimized out
>             crc|= 0x80;     // optimized out
>         }
>         else
>         {
>             crc>>= 1;       // always executed
>         }
>         octet>>= 1;     // optimized out 
>         j++;
>     }
>     return(crc);
> }
> 
> 
> The offending line seems to be the expression
>     if(tmp & 0x01)
> 
> With the -O0 (no optimization) this line is compiled to:
> 
> line 1:    ldd r24,Y+4       // Y+4 is <tmp>
> line 2:    clr r25
> line 3:    andi r24,lo8(1)
> line 4:    andi r25,hi8(1)
> line 5:    sbiw r24,0
> line 6:    breq .L11         // Jump to else clause if Z flag set
> 
> The compiler correctly expands the 8-bit unsigned char to a 16-bit int for 
> the expression evaluation.
> 
> HOWEVER !!! What the optimizer (using -O1 or -O2 or -O3) seems to do is:
> 1 - Get rid of line 5 because it thinks that subtracting 0 has no effect.
> 2 - Get rid of line 6 + the <true> part of the <if> clause because line 4 
> will always result in the Z flag to be set.
> 
> It seems to me that the machine definition of the target (mega128) is wrong. 
> It probably tells the compiler that the sbiw command does not affect the Z 
> flag.
> 
> 
> Could anybody tell me whether I'm smoking my socks or confirm (and
> preferably fix) this bug ?

I have fixed this bug in mainline. It's not an AVR specific bug it was
a bug in combine.c.

Look at gcc-patches archive with the following mail header:
---------------
From: Denis Chertykov <address@hidden>
Subject: Again combiner bug and fix
To: address@hidden
Date: Fri, 11 Jul 2003 00:01:23 +0400
---------------

Denis.