← board

sqlite libc-free runtime: pull crtl math/string + the OS/VFS bridge

Direction (per user, 2026-06-27)

Do NOT pull libc/libm for math (or anything else). Math is provided libc-free already: lib/crtl/src/math.c implements fabs/frexp/ldexp/modf and bridges sqrt/sin/pow/… to the Pascal RTL lib/rtl/math.pas. The only reason a standalone sqlite3.c compile reported undefined symbol: fabs is that the driver did not unity-include the crtl source — so fabs fell to the <math.h> extern (libm) path, and libm then never gets a DT_NEEDED anyway.

Verified: a driver that does

#include "math.c"
int main(){ return (int)(fabs(-3.5) + sqrt(16.0)); }   /* -> 7 */

with -Ilib/crtl/include -Ilib/crtl/src links with zero NEEDED libraries (fully libc-free) and runs correctly. So math is DONE; the lua runner already uses this unity-include pattern (test/lua/runner.c #include "math.c").

Remaining work (the real bring-up)

A libc-free sqlite driver #include-ing sqlite3.c plus the crtl srcs compiles, but its dynamic-symbol table still shows the sqlite OS-interface dependencies (nm -D):

The current unity driver can now run a rich in-memory SQLite smoke when it is allowed to import libc for the OS/VFS calls. Full libc-free status still requires the OS/VFS bridge or a deliberately reduced in-memory VFS configuration.

2026-06-28 update

The compiler-side blockers immediately in front of SQLite init are cleared:

With the current unity driver, sqlite3_initialize() registers the unix VFS, sqlite3MemdbInit() succeeds, and sqlite3_open(":memory:") followed by sqlite3_close() exits 0.

The first SQL execution no longer crashes in the schema error path; that pointer truncation is fixed in [[bug-c-sqlite-sql-exec-schema-argv-pointer]]. The next wall is a clean SQLITE_CORRUPT return while preparing SQLite's built-in sqlite_master schema SQL, tracked as [[bug-c-sqlite-sql-exec-schema-parse-corrupt]].

2026-06-28 update 2

The schema execution wall is cleared. test/csqlite_schema_exec_probe.c reports open=0, exec=0, close=0, and test/csqlite_extended_test.c runs through a larger in-memory workflow including aggregate queries.

nm -D /tmp/pxx_csqlite_extended no longer lists the CRTL string helpers. The remaining dynamic imports are OS/VFS calls:

fchmod fcntl64 fstat64 fsync getpid gettimeofday localtime lstat64 mkdir
mmap64 munmap nanosleep open64 stat64 sysconf utimes

2026-07-10 update — LFS *64 aliases added; :memory: acceptance MET

The 17 remaining OS/VFS imports were all present in the crtl EXCEPT five LFS (_LARGEFILE64_SOURCE) aliases that sqlite's os_unix.c uses on 64-bit Linux: open64 fcntl64 fstat64 lstat64 stat64 mmap64. On LP64 these are identical to the base calls (off_t already 64-bit), so the crtl now forwards them:

With these, test/csqlite_extended_test.c (unity-including the crtl srcs) links and runs fully libc-free — readelf -d shows ZERO DT_NEEDED, opens a :memory: db, runs CREATE/INSERT/SELECT + aggregate queries, closes, exit 0. Acceptance (below) is MET for :memory:. crtl file I/O verified working libc-free end to end (open64→write→fstat64/stat64, regression b234).

File-backed VFS — two walls found 2026-07-10

  1. SQLITE_CANTOPEN in path resolution — FIXED (crtl, 495a989a). crtl stat/lstat/fstat never set errno; sqlite's appendOnePathElement lstat()s each path element and branches on errno==ENOENT. Stale errno made a missing file look like a real error → unixFullPathname returned CANTOPEN before open() ran. Fixed: wrappers set errno = -r. Regression b235.
  2. Segfault in fillInUnixFile finder call — FIXED (cfront, e1f28f54). The locking-style finder (**(finder_type*)pVfs->pAppData)(...) — a call through a deref of a CAST to pointer-to-fnptr — dropped the call. Fixed ([[bug-c-call-through-deref-of-fnptr-pointer]], cast form; bare-ident form still open). sqlite now opens + CREATES a file-backed db (file appears on disk).
  3. unixFile mmap fields at offset 0 — FIXED (cpreproc, fc5b6a59). The #if SQLITE_MAX_MMAP_SIZE>0 (0x7fff0000) evaluated FALSE at the struct (the preprocessor didn't parse HEX literals in #if), so the mmap fields were dropped and pFile->mmapSize resolved to offset 0. Fixed by making #if parse hex/octal ([[bug-c-unixfile-mmap-field-offset-zero]], regression b237). This was the real cause of the offset-0 symptom.
  4. Segfault in unixRead — FIXED (crtl+PAL, 2026-07-10). NOT a compiler bug: it was a null-call through an unresolved aSyscall[] slot. cfront's linker fills a referenced-but-undefined C symbol with address 0 instead of erroring, so every libc syscall sqlite's os_unix.c imported that the crtl did NOT define became a call 0 → SIGSEGV the moment that VFS path ran. On __linux__ (cpreproc predefines it) sqlite takes USE_PREAD, so osPread = aSyscall[9] = pread — and the crtl had no pread, so the FIRST page read null-called (the unixRead symptom). Behind it, fillInUnixFile null-called geteuid+fchown (both active under HAVE_FCHOWN, which sqlite defines). Full set the crtl was missing: pread pwrite ftruncate access geteuid fchown readlink. Added libc-free, LP64/ILP32-safe:
    • lib/crtl/src/stdio.c: pread/pwrite — offset-preserving (save lseek SEEK_CUR, seek, io, restore); no PAL positioned-io syscall exists.
    • lib/crtl/src/unistd.c: ftruncate access fchown geteuid readlink wrappers over new __pxx_* PAL bridges.
    • lib/rtl/pxxcio.pas + lib/rtl/platform.pas + lib/rtl/platform/posix/platform_backend.pas: __pxx_ftruncate/access/ fchown/geteuid/readlink → raw syscalls (SYS_ftruncate SYS_faccessat SYS_geteuid SYS_fchown SYS_readlinkat, numbered for x86_64/i386/aarch64/ arm32/rv32; access→faccessat(AT_FDCWD), readlink→readlinkat). ESP backend gets PAL_ERR_UNSUPPORTED stubs. Regression test/crtl_posix_io_leaf_b238.c (Makefile, exit 42, 0 NEEDED). Integration probe test/csqlite_file_probe.c: file-backed CREATE/INSERT then close+reopen+SELECT reads the row back off disk (row: 1 hello), producing a valid SQLite file, fully libc-free (0 DT_NEEDED), verified native + aarch64/i386/arm32 under qemu (rv32 shares the aarch64 asm-generic table), with AND without the mmap path. File-backed VFS is now working end to end.

Acceptance

Log