Clang 20.1.6 (Fedora 20.1.6-1.fc42)'s UBSAN complains:

    qemu/include/tcg/tcg.h:715:19: runtime error: applying non-zero offset 64 to null pointer

The code uses NULL as pool's initial start and end, with the intention
that `pool_cur + size > pool_end` should trigger the allocation.
Unfortunately C prohibits adding non-zero to NULL, even if the result
is not dereferenced.

Fix by using a dummy pool.

Signed-off-by: Ilya Leoshkevich <iii@linux.ibm.com>
---
 tcg/tcg.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/tcg/tcg.c b/tcg/tcg.c
index d714ae2889c..afcc7ec8849 100644
--- a/tcg/tcg.c
+++ b/tcg/tcg.c
@@ -1358,13 +1358,14 @@ void *tcg_malloc_internal(TCGContext *s, int size)
 
 void tcg_pool_reset(TCGContext *s)
 {
+    static uint8_t dummy_pool;
     TCGPool *p, *t;
     for (p = s->pool_first_large; p; p = t) {
         t = p->next;
         g_free(p);
     }
     s->pool_first_large = NULL;
-    s->pool_cur = s->pool_end = NULL;
+    s->pool_cur = s->pool_end = &dummy_pool;
     s->pool_current = NULL;
 }
 
-- 
2.49.0

On 6/18/25 03:35, Ilya Leoshkevich wrote:
> Clang 20.1.6 (Fedora 20.1.6-1.fc42)'s UBSAN complains:
> 
>      qemu/include/tcg/tcg.h:715:19: runtime error: applying non-zero offset 64 to null pointer
> 
> The code uses NULL as pool's initial start and end, with the intention
> that `pool_cur + size > pool_end` should trigger the allocation.
> Unfortunately C prohibits adding non-zero to NULL, even if the result
> is not dereferenced.
> 
> Fix by using a dummy pool.

If we want to get uselessly technical, &dummy_pool + x, for x > 1 isn't legal either. 
This is just waiting for another ubsan update to break.

Does clang accept it if you rearrange the arithmetic to

   pool_end - pool_cur <= size

?

Otherwise, we may have to change the pool pointers to uintptr_t, so that we opt out of all 
ubsan silliness.


r~



> 
> Signed-off-by: Ilya Leoshkevich <iii@linux.ibm.com>
> ---
>   tcg/tcg.c | 3 ++-
>   1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/tcg/tcg.c b/tcg/tcg.c
> index d714ae2889c..afcc7ec8849 100644
> --- a/tcg/tcg.c
> +++ b/tcg/tcg.c
> @@ -1358,13 +1358,14 @@ void *tcg_malloc_internal(TCGContext *s, int size)
>   
>   void tcg_pool_reset(TCGContext *s)
>   {
> +    static uint8_t dummy_pool;
>       TCGPool *p, *t;
>       for (p = s->pool_first_large; p; p = t) {
>           t = p->next;
>           g_free(p);
>       }
>       s->pool_first_large = NULL;
> -    s->pool_cur = s->pool_end = NULL;
> +    s->pool_cur = s->pool_end = &dummy_pool;
>       s->pool_current = NULL;
>   }
>

On Wed, 2025-06-18 at 12:35 +0200, Ilya Leoshkevich wrote:
> Clang 20.1.6 (Fedora 20.1.6-1.fc42)'s UBSAN complains:
> 
>     qemu/include/tcg/tcg.h:715:19: runtime error: applying non-zero
> offset 64 to null pointer
> 
> The code uses NULL as pool's initial start and end, with the
> intention
> that `pool_cur + size > pool_end` should trigger the allocation.
> Unfortunately C prohibits adding non-zero to NULL, even if the result
> is not dereferenced.
> 
> Fix by using a dummy pool.
> 
> Signed-off-by: Ilya Leoshkevich <iii@linux.ibm.com>
> ---
>  tcg/tcg.c | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/tcg/tcg.c b/tcg/tcg.c
> index d714ae2889c..afcc7ec8849 100644
> --- a/tcg/tcg.c
> +++ b/tcg/tcg.c
> @@ -1358,13 +1358,14 @@ void *tcg_malloc_internal(TCGContext *s, int
> size)
>  
>  void tcg_pool_reset(TCGContext *s)
>  {
> +    static uint8_t dummy_pool;
>      TCGPool *p, *t;
>      for (p = s->pool_first_large; p; p = t) {
>          t = p->next;
>          g_free(p);
>      }
>      s->pool_first_large = NULL;
> -    s->pool_cur = s->pool_end = NULL;
> +    s->pool_cur = s->pool_end = &dummy_pool;
>      s->pool_current = NULL;
>  }

Hmm, come to think of it, something like `&dummy_pool + 2` is not
allowed either, it's just that UBSAN doesn't complain about it. I will
send a v2.