Bug#131454: gcc -O2 produces wrong code on PPC (gnuchess example)

[Lukas Geyer <geyer@ml.kva.se>]

Package: gcc
Version: 2:2.95.4-9
Severity: important

Compiling the current gnuchess source fails with this gcc (current woody)
and optimization set to -O2. The reason seems to be that statements are
not correctly serialized. Unfortunately I am not an expert in PPC
assembler but here is the code which breaks (modified from gnuchess source
file src/move.c, somewhat shortened). Removing the "inline" keyword
generates correct code with -O2. However, in the original source,
clear_and_set_bit() is just part of UnmakeMove(), as well as of
MakeMove(). So, the following should be (even more than) the relevant
part.

typedef unsigned long long BitBoard;

#define SETBIT(b,i)   ((b) |= BitPosArray[i])
#define CLEARBIT(b,i) ((b) &= NotBitPosArray[i])
#define TOSQ(a)       ((a) & 0x003F)
#define FROMSQ(a)     ((a >> 6) & 0x003F)

inline void clear_and_set_bit(BitBoard *a, short f, short t)
{
        CLEARBIT (*a, f);
        SETBIT (*a, t);
}

void UnmakeMove (short side, int *move)
{
   BitBoard *a;
   short f, t, fpiece, cpiece;
   short xside;

   side = 1^side;
   xside = 1^side;
   f = FROMSQ(*move);
   t = TOSQ(*move);
   fpiece = cboard[t];
   a = &board.b[side][fpiece];
   clear_and_set_bit(a, t, f);

   /* continues... */
}

So MakeMove() has about the same structure, except of course it switches
the roles of t and f. Now the gnuchess move generation algorithm generates
all possible moves, makes them, and unmakes them again. With inlining and
-O2 however, it seems that making and unmaking of moves is not serialized,
so that after trying out all pawn moves from the beginning position, white
has a pawn on every square between the second and fourth row. If someone
really wants, I can supply the assembler files move.s for -O1 (correct
code) and -O2 (incorrect), together with the C source.

Lukas