This is technically undefined behavior.

Technically it is, but we're unaware of any extant environments where it's not well-behaved.

Edit: if such a platform were to suddenly appear it would have issues many pieces of software which use dlsym() to fish out function pointers.

Specifically, POSIX's definition of dlsym() essentially requires ordinary function and data pointers to be the same size and come from the same address space. Windows does something similar (but with a different basic operation name). The combination covers the enormous majority of systems deployed out there, except for many embedded devices (though a lot of those are too small to run either a conventional OS or SQLite).

It will become a problem if sqlite is built with CFI. See https://maskray.me/blog/2022-12-18-control-flow-integrity#fsanitizefunction.

Mentioning dlsym confuses things, because casting a function pointer to a data pointer is a different type of undefined behaviour. The error Nate A reports can be fixed like this:

Index: src/select.c
==================================================================
--- src/select.c
+++ src/select.c
@@ -7035,11 +7035,12 @@
 }
 
 /*
 ** Deallocate a single AggInfo object
 */
-static void agginfoFree(sqlite3 *db, AggInfo *p){
+static void agginfoFree(sqlite3 *db, void *vp){
+  AggInfo *p = vp;
   sqlite3DbFree(db, p->aCol);
   sqlite3DbFree(db, p->aFunc);
   sqlite3DbFreeNN(db, p);
 }
 
@@ -8015,11 +8016,11 @@
     ** SELECT statement.
     */
     pAggInfo = sqlite3DbMallocZero(db, sizeof(*pAggInfo) );
     if( pAggInfo ){
       sqlite3ParserAddCleanup(pParse,
-          (void(*)(sqlite3*,void*))agginfoFree, pAggInfo);
+          agginfoFree, pAggInfo);
       testcase( pParse->earlyCleanup );
     }
     if( db->mallocFailed ){
       goto select_end;
     }

(Plus similar changes for every other function used with sqlite3ParserAddCleanup.)

In short, a (well-defined) data pointer type conversion inside the function lets you remove the (ill-defined (and ugly!)) function pointer type cast when installing it as a callback.

Mentioning dlsym confuses things, because casting a function pointer to a data pointer is a different type of undefined behaviour.

Indeed, i misunderstood it to be that other type of UB.

Richard has since patched these casts so that they're well-defined.