Add returns_nonnull to infallible allocators
Postgres memory allocators, by default, ERROR out on memory allocation
failures. An ERROR leads to a longjmp, which means that the caller of
the allocator will never see a NULL return value. We can explicitly let
the compiler know about this behavior by adding the returns_nonnull
attribute to the allocators that follow this behavior. Postgres does
support _extended versions of some of the allocators that take a flaks
argument. The caller can provide the MCXT_ALLOC_NO_OOM flag to these
allocators to request that they return NULL on allocation failure
instead of ERROR-ing out. The _extended allocators cannot be marked as
returns_nonnull because of that.
By using returns_nonnull, we can help the compiler to optimize call
sites.
--
Tristan Partin
PostgreSQL Contributors Team
AWS (https://aws.amazon.com)
On Mon Jul 6, 2026 at 6:10 PM UTC, Tristan Partin wrote:
Postgres memory allocators, by default, ERROR out on memory allocation
failures. An ERROR leads to a longjmp, which means that the caller of
the allocator will never see a NULL return value. We can explicitly let
the compiler know about this behavior by adding the returns_nonnull
attribute to the allocators that follow this behavior. Postgres does
support _extended versions of some of the allocators that take a flaks
argument. The caller can provide the MCXT_ALLOC_NO_OOM flag to these
allocators to request that they return NULL on allocation failure
instead of ERROR-ing out. The _extended allocators cannot be marked as
returns_nonnull because of that.By using returns_nonnull, we can help the compiler to optimize call
sites.
...
--
Tristan Partin
PostgreSQL Contributors Team
AWS (https://aws.amazon.com)
On 06.07.26 20:11, Tristan Partin wrote:
On Mon Jul 6, 2026 at 6:10 PM UTC, Tristan Partin wrote:
Postgres memory allocators, by default, ERROR out on memory allocation
failures. An ERROR leads to a longjmp, which means that the caller of
the allocator will never see a NULL return value. We can explicitly let
the compiler know about this behavior by adding the returns_nonnull
attribute to the allocators that follow this behavior. Postgres does
support _extended versions of some of the allocators that take a flaks
argument. The caller can provide the MCXT_ALLOC_NO_OOM flag to these
allocators to request that they return NULL on allocation failure
instead of ERROR-ing out. The _extended allocators cannot be marked as
returns_nonnull because of that.By using returns_nonnull, we can help the compiler to optimize call
sites.
Your patch marks palloc_mul_extended() as pg_attribute_returns_nonnull,
which seems incorrect per the above description.
Also, per the discussion in the counted_by thread, lets put these new
attributes in their more correct position in front of the declaration.
(Note that several of these already use pg_nodiscard, which is also an
attribute.) That way we can also use the MSVC annotation _Ret_notnull_.
On Tue Sep 29, 2026 at 6:49 AM UTC, Peter Eisentraut wrote:
On 06.07.26 20:11, Tristan Partin wrote:
On Mon Jul 6, 2026 at 6:10 PM UTC, Tristan Partin wrote:
Postgres memory allocators, by default, ERROR out on memory allocation
failures. An ERROR leads to a longjmp, which means that the caller of
the allocator will never see a NULL return value. We can explicitly let
the compiler know about this behavior by adding the returns_nonnull
attribute to the allocators that follow this behavior. Postgres does
support _extended versions of some of the allocators that take a flaks
argument. The caller can provide the MCXT_ALLOC_NO_OOM flag to these
allocators to request that they return NULL on allocation failure
instead of ERROR-ing out. The _extended allocators cannot be marked as
returns_nonnull because of that.By using returns_nonnull, we can help the compiler to optimize call
sites.Your patch marks palloc_mul_extended() as pg_attribute_returns_nonnull,
which seems incorrect per the above description.
Oops. Sorry about that. I must've not sent the most up to date version
of the patch because I definitely didn't have that locally.
Also, per the discussion in the counted_by thread, lets put these new
attributes in their more correct position in front of the declaration.
(Note that several of these already use pg_nodiscard, which is also an
attribute.) That way we can also use the MSVC annotation _Ret_notnull_.
Good points. See the revised attached patch. You definitely have a knack
for knowing MSVC attributes. I'll have to try to keep those in mind.
--
Tristan Partin
PostgreSQL Contributors Team
AWS (https://aws.amazon.com)
On Wed Oct 7, 2026 at 5:15 AM UTC, Tristan Partin wrote:
On Tue Sep 29, 2026 at 6:49 AM UTC, Peter Eisentraut wrote:
On 06.07.26 20:11, Tristan Partin wrote:
On Mon Jul 6, 2026 at 6:10 PM UTC, Tristan Partin wrote:
Postgres memory allocators, by default, ERROR out on memory allocation
failures. An ERROR leads to a longjmp, which means that the caller of
the allocator will never see a NULL return value. We can explicitly let
the compiler know about this behavior by adding the returns_nonnull
attribute to the allocators that follow this behavior. Postgres does
support _extended versions of some of the allocators that take a flaks
argument. The caller can provide the MCXT_ALLOC_NO_OOM flag to these
allocators to request that they return NULL on allocation failure
instead of ERROR-ing out. The _extended allocators cannot be marked as
returns_nonnull because of that.By using returns_nonnull, we can help the compiler to optimize call
sites.Your patch marks palloc_mul_extended() as pg_attribute_returns_nonnull,
which seems incorrect per the above description.Oops. Sorry about that. I must've not sent the most up to date version
of the patch because I definitely didn't have that locally.Also, per the discussion in the counted_by thread, lets put these new
attributes in their more correct position in front of the declaration.
(Note that several of these already use pg_nodiscard, which is also an
attribute.) That way we can also use the MSVC annotation _Ret_notnull_.Good points. See the revised attached patch. You definitely have a knack
for knowing MSVC attributes. I'll have to try to keep those in mind.
Here is a v3 that adds missing annotations for the following functions
in palloc.h:
- repalloc
- repalloc0
- repalloc_mul
- repalloc_huge
And for fe_memutils.h:
- pg_malloc
- pg_malloc0
- pg_malloc_mul
- pg_malloc0_mul
- palloc
- palloc0
- palloc_mul
- palloc0_mul
--
Tristan Partin
PostgreSQL Contributors Team
AWS (https://aws.amazon.com)
On 07.10.26 07:43, Tristan Partin wrote:
Here is a v3 that adds missing annotations for the following functions
in palloc.h:- repalloc
- repalloc0
- repalloc_mul
- repalloc_hugeAnd for fe_memutils.h:
- pg_malloc
- pg_malloc0
- pg_malloc_mul
- pg_malloc0_mul
- palloc
- palloc0
- palloc_mul
- palloc0_mul
Committed.
I also added the annotation to the frontend version of psprintf(). You
only had it for the backend version.