Hi!
I'm porting GDC compiler to QNX (my previous posts related to this topic are: http://www.digitalmars.com/webnews/newsgroups.php?art_group=digitalmars.D&article_id=51281 and http://www.digitalmars.com/webnews/newsgroups.php?art_group=digitalmars.D&article_id=51456)
I'm almost done here, everything is working great, except one little thing: optimization. It seems to me that -O1, -O2 and -O3 (but not -Os) options generate bad code. The problem is that the following code doesn't work as expected:

    char[] a = "abcd";
    char[] r;

    r = a.dup.reverse;
    assert(r == "dcba");

    a = "a\u1235\u1234c";
    r = a.dup.reverse;
    assert(r == "c\u1234\u1235a");

When you compile it without optimization everything is ok, but when you use -O for example, the following happens:

function _adReverseChar (that's the function that is called when you use .reverse) works for a very very very long time, that happens because the "char[] a" parameter that is passed to that function is passed wrongly, if you call a.length then you'll see that it'll return something like 134_XXX_XXX, i.e a very large number though the real size of a string is 4.

What could be the problem ? I really don't have a clue, because same code works fine with -O in linux and QNX is a lot like linux. So it's unlikely that gcc optimizer works different in linux and qnx, that means that I probably forgot to turn on/off some switches while compiling GDC, does anyone have any ideas what exactly could it be ?

Sheff schrieb am 2007-04-22:
> Hi!
> I'm porting GDC compiler to QNX (my previous posts related to this topic are: http://www.digitalmars.com/webnews/newsgroups.php?art_group=digitalmars.D&article_id=51281 and http://www.digitalmars.com/webnews/newsgroups.php?art_group=digitalmars.D&article_id=51456)
> I'm almost done here, everything is working great, except one little thing: optimization. It seems to me that -O1, -O2 and -O3 (but not -Os) options generate bad code. The problem is that the following code doesn't work as expected:
>
>     char[] a = "abcd";
>     char[] r;
>
>     r = a.dup.reverse;
>     assert(r == "dcba");
>
>     a = "a\u1235\u1234c";
>     r = a.dup.reverse;
>     assert(r == "c\u1234\u1235a");
>
> When you compile it without optimization everything is ok, but when you use -O for example, the following happens:
>
> function _adReverseChar (that's the function that is called when you use .reverse) works for a very very very long time, that happens because the "char[] a" parameter that is passed to that function is passed wrongly, if you call a.length then you'll see that it'll return something like 134_XXX_XXX, i.e a very large number though the real size of a string is 4.

What is the ouput of the below code if it is run inside _adReverseChar?

# printf("LEN  %zX\n", a.length);
# printf("PTR  %zX\n", a.ptr);
# size_t* data = &a;
# printf("pre  %zX\n", *(data-1));
# printf("len  %zX\n", *(data));
# printf("ptr  %zX\n", *(data+1));
# printf("post %zX\n", *(data+2));

Thomas

Thomas Kuehne Wrote:

> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
> 
> Sheff schrieb am 2007-04-22:
> > Hi!
> > I'm porting GDC compiler to QNX (my previous posts related to this topic are: http://www.digitalmars.com/webnews/newsgroups.php?art_group=digitalmars.D&article_id=51281 and http://www.digitalmars.com/webnews/newsgroups.php?art_group=digitalmars.D&article_id=51456)
> > I'm almost done here, everything is working great, except one little thing: optimization. It seems to me that -O1, -O2 and -O3 (but not -Os) options generate bad code. The problem is that the following code doesn't work as expected:
> >
> >     char[] a = "abcd";
> >     char[] r;
> >
> >     r = a.dup.reverse;
> >     assert(r == "dcba");
> >
> >     a = "a\u1235\u1234c";
> >     r = a.dup.reverse;
> >     assert(r == "c\u1234\u1235a");
> >
> > When you compile it without optimization everything is ok, but when you use -O for example, the following happens:
> >
> > function _adReverseChar (that's the function that is called when you use .reverse) works for a very very very long time, that happens because the "char[] a" parameter that is passed to that function is passed wrongly, if you call a.length then you'll see that it'll return something like 134_XXX_XXX, i.e a very large number though the real size of a string is 4.
> 
> What is the ouput of the below code if it is run inside _adReverseChar?
> 
> # printf("LEN  %zX\n", a.length);
> # printf("PTR  %zX\n", a.ptr);
> # size_t* data = &a;
> # printf("pre  %zX\n", *(data-1));
> # printf("len  %zX\n", *(data));
> # printf("ptr  %zX\n", *(data+1));
> # printf("post %zX\n", *(data+2));
> 
> Thomas
> 
> 
> -----BEGIN PGP SIGNATURE-----
> 
> iD8DBQFGLJYZLK5blCcjpWoRAmH5AJ9Kr2t386L1mD1pkg1Y2RXuY0Y3iACdGYwt
> MyznI2I11SSGFAcjRBEic54=
> =rPGB
> -----END PGP SIGNATURE-----

The output is:

LEN  4
PTR  8078FE0
pre  80478A8
len  4
ptr  8078FE0
post 805E980
LEN  805E991
PTR  8180000
pre  8047898
len  805E991
ptr  8180000
post 805E991

I've already understood what was the problem, the optimizer generated bad code, which was lefting garbage on the stack which then was passed to functions as arguments (in this example, the size of array "a"), I've manually fixed this in generated asm listing (added few pop instructions) and compiled from that listing, the program worked. So, where should I search for bugs ? In gcc backend or in gdc frontend ? Who's responsible for such "optimization" ?