What (exact) version is the sqlite3.c file that is reporting the runtime error? You can find this using a command like:

grep 'define SQLITE_SOURCE_ID' sqlite3.c

Unfortunately, the testing was done automatically on a CRAN machine that I do not have access to.

But in their 'outputs.txt' file, which is available here (but which probably will disappear soon)

https://win-builder.r-project.org/incoming_pretest/Mega2R_1.2.0_20250620_214524/specialChecks/gcc-san/outputs.txt

I see this:

* installing *source* package ‘RSQLite’ ...
** this is package ‘RSQLite’ version ‘2.4.1’
** package ‘RSQLite’ successfully unpacked and MD5 sums checked

So if I download the source for ‘RSQLite’ version ‘2.4.1’ from CRAN, then I get

$ grep 'define SQLITE_SOURCE_ID' sqlite3.c
#define SQLITE_SOURCE_ID      "2025-06-06 14:52:32 b77dc5e0f596d2140d9ac682b2893ff65d3a4140aa86067a3efebe29dc914c95"

I've seen the same issue recently. Using current fossil (2025-06-25 20:42:40 cb4d05633a0c9cdf146f3108e1b4b10754cd79d72a425d8cc9cd21836037a01b) and earlier versions. Amalgamation produced with ./configure --quiet --all --disable-tcl

Operating system is Ubuntu 25.04 and you have to install gcc-15 and compile with optimisation on - -Og means the issue isn't shown. The compilation needs to use the undefined behaviour sanitizer. Neither gcc-14 or clang show the issue.

sqlite3/sqlite3.c:80356:14: runtime error: load of address 0x7ab294d2ce00 with insufficient space for an object of type 'struct MemPage *'
0x7ab294d2ce00: note: pointer points here
 cd cd cd cd  90 0d ed 99 c2 7c 00 00  90 81 ec 99 c2 7c 00 00  cd cd cd cd cd cd cd cd  cd cd cd cd
              ^

Line 80356 corresponds to MemPage *pOld = ...:

  if( (pageFlags & PTF_LEAF)==0 && nOld!=nNew ){
    MemPage *pOld = (nNew>nOld ? apNew : apOld)[nOld-1];
    memcpy(&apNew[nNew-1]->aData[8], &pOld->aData[8], 4);
  }

Relevant portion of backtrace:

#4  0x00007bffe4385eea in balance_nonroot (pParent=pParent@entry=0x7f1ff5b29880, iParentIdx=iParentIdx@entry=0, aOvflSpace=aOvflSpace@entry=0x7e0ff5e28508 '\276' <repeats 200 times>..., isRoot=isRoot@entry=1, bBulk=<optimised out>) at /space/apsw/sqlite3/sqlite3.c:80356
#5  0x00007bffe43898bd in balance (pCur=pCur@entry=0x7edff5aeebb8) at /space/apsw/sqlite3/sqlite3.c:80817
#6  0x00007bffe438ec10 in sqlite3BtreeInsert (pCur=0x7edff5aeebb8, pX=pX@entry=0x7bfff4a384f0, flags=flags@entry=0, seekResult=seekResult@entry=-1) at /space/apsw/sqlite3/sqlite3.c:81272
#7  0x00007bffe44c2262 in sqlite3VdbeExec (p=p@entry=0x7edff5af0748) at /space/apsw/sqlite3/sqlite3.c:101006
#8  0x00007bffe44d2fe7 in sqlite3Step (p=p@entry=0x7edff5af0748) at /space/apsw/sqlite3/sqlite3.c:92408
#9  0x00007bffe44d3831 in sqlite3_step (pStmt=pStmt@entry=0x7edff5af0748) at /space/apsw/sqlite3/sqlite3.c:92470

To cause the report, the code does this:

• Opening file database
• pragma cache_size = 10
• pragma temp_store = 1
• create temp table testissue506(x UNIQUE,y UNIQUE, PRIMARY KEY(x,y));
• Inserting ("", ""), ("a", "b"), ("aa", "bb"), ("aaa", "bbb") etc to spill from file to temp.

Because of the requirements to reproduce the issue, it looks like a compiler issue not a bug in SQLite which is why I didn't report it.

I don't have Ubuntu 25 easily at hand. (My machines are on 24.04). And the latest GCC I have is 13.3.0. So I'm not able to repro the problem as originally stated. I did make (an uncommitted) debug modification as follows:

--- src/btree.c
+++ src/btree.c
@@ -8695,8 +8695,9 @@
   /* If the sibling pages are not leaves, ensure that the right-child pointer
   ** of the right-most new sibling page is set to the value that was
   ** originally in the same field of the right-most old sibling page. */
   if( (pageFlags & PTF_LEAF)==0 && nOld!=nNew ){
+    printf("nNew=%d, nOld=%d\n", nNew, nOld); fflush(stdout);
     MemPage *pOld = (nNew>nOld ? apNew : apOld)[nOld-1];
     memcpy(&apNew[nNew-1]->aData[8], &pOld->aData[8], 4);
   }

I was not able to reach the debug printf() using Roger's recipe. This code is in the b-tree layer which is independent of the page cache, so the "cache_size" and "temp_store" and "spill from file to temp" parts of Roger's recipe seem like red herrings to me, but I don't know. I was able to reach the debugging printf() using this script:

CREATE TABLE t1(x UNIQUE,y UNIQUE,z BLOB,PRIMARY KEY(x,y));
WITH RECURSIVE c(n) AS (VALUES(1) UNION ALL SELECT n+1 FROM c WHERE n<26*20)
INSERT INTO t1
  SELECT format('%c%.*c',char(0x61+n%26),n,'a'),
         format('%c%.*c',char(0x61+(n*17)%26),n,'b'),
         zeroblob(15+n%30)
    FROM c;
DELETE FROM t1 WHERE rowid%19;

Roger: If you insert a temporary printf() has shown above, can you still hit the UBSAN warning? If so, what values of nNew and nOld are shown?

Anybody: Are you able to hit a UBSAN warning by compiling "sqlite3" and running the script above?

You can use distrobox to get container environments of other Linux versions, although it may be more hassle than it is worth.

The cache size and temp store pragmas are there because my test case is ensuring my VFS code correctly encounters and handles a NULL filename, ensuring that memory is filled with data and then temp db spill to disk happens. Commenting them out still causes the UBSAN warning.

nNew=2, nOld=1

Your recursive query also reproduces the issue. A CREATE followed by INSERT or a large blob does not.

I asked about this on R-package-devel and they suggested I contact you, saying it might be related to recent changes discussed here: https://sqlite.org/forum/forumpost/311dbf9a1cadfae6

The (extensive) changes to get -fsanitize=strict-bounds to hush up did not affect the MemPage object, which is the focus of this new warning. So that theory seems unlikely.

I managed to hit it with gcc 14 too. The recursive query is in a file named test.sql.

At least -O1 must be present.

$ gcc -O1  -fsanitize=undefined -o sqlite3 shell.c sqlite3.c -lm && ./sqlite3  "" ".read test.sql"
nNew=2, nOld=1
sqlite3.c:80357:14: runtime error: load of address 0x7ffd8e63d8e0 with insufficient space for an object of type 'struct MemPage *'
0x7ffd8e63d8e0: note: pointer points here
 78 7a 00 00  d0 55 3b f6 6f 59 00 00  e0 87 3b f6 6f 59 00 00  48 45 3b f6 6f 59 00 00  48 5a 8b 3e
              ^ 
nNew=3, nOld=2
nNew=4, nOld=3
nNew=2, nOld=1
nNew=2, nOld=1
nNew=4, nOld=3
...

It does not reproduce with clang.

The error appears in the final INSERT statement of the script shown at the bottom of this post when compiling and run as follows:

gcc -O1 -g -fsanitize=undefined -o sqlite3 shell.c sqlite3.c -lm -ldl -lpthread
./sqlite3 <ub1.txt

The change at check-in 2025-06-26T18:57Z clears the warning. I agree with Roger: this seems like a gcc/ubsan bug. I'll see if I can enter a bug report that they can use.

FWIW, Dan discovered that the same error comes up with gcc-13. It was not necessary for me to spend all that time getting gcc-14 installed.....

Here is the "ub1.txt" script:

CREATE TABLE t1(x UNIQUE,y UNIQUE,z BLOB,PRIMARY KEY(x,y));
WITH RECURSIVE c(n) AS (VALUES(1) UNION ALL SELECT n+1 FROM c WHERE n<233)
INSERT INTO t1
  SELECT format('%c%.*c',char(0x61+n%26),n,'a'),
         format('%c%.*c',char(0x61+(n*17)%26),n,'b'),
         zeroblob(15+n%30)
    FROM c;
INSERT INTO t1 VALUES(
  format('a%.*c',234,'a'),
  format('a%.*c',234,'b'),
  zeroblob(39)
);

The change at check-in 2025-06-26T18:57Z ...

That change also makes the warning go away in gcc-15.

It was also occurring with -Os, so it looked like any optimization level triggered whatever concerned gcc.

I filed a bug with GCC at https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120837. We'll see if we get pushback...

it may be more hassle than it is worth.

The hassle threshold would have to be pretty low, as easy as it is:

$ sudo apt install distrobox
$ distrobox create --image ubuntu:25.04
$ distrobox enter ubuntu-25-04
…set up the in-box deps needed…
$ cd ~/src/sqlite/checkout

…and then build as normal.

The checkout dir should be clean, free of host-side build output, since you are sharing a home directory, thus ~/src, presuming that is where SQLite is unpacked.