[PATCH] properly terminate linked link of dsos

[PATCH] properly terminate linked link of dsos

[Arvid Ephraim Picciani]

---
 src/ldso/dynlink.c |    1 +
 1 files changed, 1 insertions(+), 0 deletions(-)

diff --git a/src/ldso/dynlink.c b/src/ldso/dynlink.c
index 1b55e07..3f271aa 100644
--- a/src/ldso/dynlink.c
+++ b/src/ldso/dynlink.c
@@ -375,6 +375,7 @@ static struct dso *load_library(const char *name)
 	p->name = p->buf;
 	strcpy(p->name, name);
 
+    p->tail = 0;
 	tail->next = p;
 	p->prev = tail;
 	tail = p;
-- 
1.7.7

Re: [PATCH] properly terminate linked link of dsos

[Rich Felker]

On Wed, Oct 12, 2011 at 03:42:54PM +0200, Arvid Ephraim Picciani wrote:
> ---
>  src/ldso/dynlink.c |    1 +
>  1 files changed, 1 insertions(+), 0 deletions(-)
> 
> diff --git a/src/ldso/dynlink.c b/src/ldso/dynlink.c
> index 1b55e07..3f271aa 100644
> --- a/src/ldso/dynlink.c
> +++ b/src/ldso/dynlink.c
> @@ -375,6 +375,7 @@ static struct dso *load_library(const char *name)
>  	p->name = p->buf;
>  	strcpy(p->name, name);
>  
> +    p->tail = 0;
>  	tail->next = p;
>  	p->prev = tail;
>  	tail = p;

See line 357. This patch is a no-op.

Rich

Re: [PATCH] properly terminate linked link of dsos

[aep]

On Wed, 12 Oct 2011 10:01:41 -0400, Rich Felker wrote:

> See line 357. This patch is a no-op.
>
> Rich


indeed. For the record (as dicussed on irc): calloc doesn't zero the 
memory for me.
Something entirely else is fishy.

Re: [PATCH] properly terminate linked link of dsos

[aep]

clang optimized away the check in calloc.c to false.
I can work around it with a cast to something volatile:

diff --git a/src/malloc/calloc.c b/src/malloc/calloc.c
index 9d57456..22f2c01 100644
--- a/src/malloc/calloc.c
+++ b/src/malloc/calloc.c
@@ -13,8 +13,10 @@ void *calloc(size_t m, size_t n)
         n *= m;
         p = malloc(n);
         if (!p) return 0;
+
+    volatile long long dont_optimize_me = p;
         /* Only do this for non-mmapped chunks */
-       if (((size_t *)p)[-1] & 7) {
+       if (((size_t *)dont_optimize_me)[-1] & 7) {
                 /* Only write words that are not already zero */
                 m = (n + sizeof *z - 1)/sizeof *z;
                 for (z=p; m; m--, z++) if (*z) *z=0;

Re: [PATCH] properly terminate linked link of dsos

[Rich Felker]

On Wed, Oct 12, 2011 at 04:46:30PM +0200, aep wrote:
> clang optimized away the check in calloc.c to false.
> I can work around it with a cast to something volatile:

You could instead just change the (size_t *) cast to
(volatile size_t *)...

But I think the problem is much deeper. This "optimization" is
necessarily based on the compiler making an assumption about what the
function named "malloc" does. Compiling the implementation (libc)
itself requires disabling such assumptions; with gcc, this is achieved
with -ffreestanding (-fno-builtin will also work). Can you use the
same or similar option with clang?

I made similar comments on IRC..

Rich

Re: [PATCH] properly terminate linked link of dsos

[Isaac Dunham]

On Thu, 13 Oct 2011 21:03:17 -0400
Rich Felker <dalias@aerifal.cx> wrote:

> But I think the problem is much deeper. This "optimization" is
> necessarily based on the compiler making an assumption about what the
> function named "malloc" does. Compiling the implementation (libc)
> itself requires disabling such assumptions; with gcc, this is achieved
> with -ffreestanding (-fno-builtin will also work). Can you use the
> same or similar option with clang?

Per the manpage, -ffreestanding and -fno-builtin will work with clang.
-ffreestanding is probably the proper choice, since a new libc is a
freestanding environment.

Re: [PATCH] properly terminate linked link of dsos

[Rich Felker]

On Fri, Oct 14, 2011 at 02:11:52PM -0700, Isaac Dunham wrote:
> On Thu, 13 Oct 2011 21:03:17 -0400
> Rich Felker <dalias@aerifal.cx> wrote:
> 
> > But I think the problem is much deeper. This "optimization" is
> > necessarily based on the compiler making an assumption about what the
> > function named "malloc" does. Compiling the implementation (libc)
> > itself requires disabling such assumptions; with gcc, this is achieved
> > with -ffreestanding (-fno-builtin will also work). Can you use the
> > same or similar option with clang?
> 
> Per the manpage, -ffreestanding and -fno-builtin will work with clang.
> -ffreestanding is probably the proper choice, since a new libc is a
> freestanding environment.

Indeed, but it sounds like this is not fixing the problem. :(
Can you confirm that calloc.c is compiled properly when -ffreestanding
is used?

Rich