[Fwd: egcs 112 & alpha & bad -O1 optimization]

[Uncle George <gatgul@voicenet.com>]

Followup:
This is the simplist reduction of the code that causes the Optimization
error,

    if u remove the  "l = (((struct varlena *)( v ))->vl_len) "
statement then the compiled ( with cc -S -O1 )  appears to be more
correct.
    If u leave it in, then the *v++ gets screwed up.

gat


typedef signed int int32;
struct varlena { int32  vl_len; char    vl_dat[1]; };
typedef union nameData
{ char  data[32 ]; int  alignmentDummy;
} NameData;
typedef NameData *Name;
static long comphash(long l, char *v);

static long
comphash(long l, char *v)
{
        long            i;
        NameData        n;

         if (l < 0)
                l = (((struct varlena *)( v ))->vl_len) ;

        i = 0;
        while (l--)
                i += *v++;
        return i;
}
~

---

• To: gcc-bugs@gcc.gnu.org
• Subject: egcs 112 & alpha & bad -O1 optimization
• From: Uncle George <gatgul@voicenet.com>
• Date: Thu, 29 Jul 1999 13:12:12 -0400
• Organization: Big Endian

The following attached preprocessed 'f.c' from postgresql failes to
properly compile with -O1. In particular the char *v in the argument
list of comphash appears to get translated into a long *

the particular errant code segment from comphash() is :

        while (l--)
                i += *v++;

which appear to get xlated into assembler:

$L205:
        ldl $1,0($0)
        sll $1,56,$1
        sra $1,56,$1
        addq $2,$1,$2
        addq $0,1,$0
        subq $9,1,$9
        cmpeq $9,$3,$1
        beq $1,$L205

unfortunately this does not do very well for the *v, which gets an
alignment fault!

gat

f.c.gz

f.s.gz

---