[Debian-med-packaging] Bug#1153516: arb: FTBFS on 32-bit architectures: arb_progress(const char*, unsigned int) is private

Cheng Li shiptux at gmail.com
Tue Nov 10 06:58:38 GMT 2026


Hi Andreas,

A correction to my last message, and a substantive one. This row

      add unsigned int           ok      ok

was wrong in the i386 column. It came from the 20-line reproducer, not from
a build of arb, and the reproducer did not model enough of the real header.
I have now built arb itself on both architectures. The patch I sent you is
incomplete; please do not apply it as it stands. The complete one is below.

First, why the mistake was possible at all, because it is not only mine.
The i386 buildd log for 7.0+ds-1 contains exactly six errors, all of them
the two-argument form, all in the first two libraries to be built:

    ARB_GDE/GDE_event.cxx:363
    ARBDB/adname.cxx:181, 219, 272
    ARBDB/TreeNode.cxx:425, 458

and then `make[4]: *** Waiting for unfinished jobs....`. The build dies at
ARB_GDE.a and lib/libARBDB.so, so PROBE, NTREE and AWTI are never compiled.
The log is a faithful record of a build that stopped early, not a list of
everything that is wrong on ILP32. Fixing those six sites lets the build get
further, and two more defects appear.

The second one is the same mechanism, one line up in the header. There are
two private templates, not one:

    template <typename NUM>          arb_progress(const char *t, NUM n);
    template <typename NUM> explicit arb_progress(NUM n);

I only published an `unsigned int` partner for the first. These six call
sites pass a single `size_t` and hit the second:

    PROBE/PT_new_design.cxx:1269   arb_progress base_progress(2*bp);
    NTREE/NT_dbrepair.cxx:189      arb_progress progress(species_count);
    NTREE/NT_dbrepair.cxx:267      arb_progress progress(species_count);
    NTREE/NT_dbrepair.cxx:477      arb_progress progress(all);
    NTREE/NT_syncroot.cxx:140      arb_progress progress2(toSyncCount);
    NTREE/NT_tree_cmp.cxx:679      arb_progress compare_progress(steps);

The third has nothing to do with arb_progress, which is why neither of us
looked for it. AWTI/AWTI_export.cxx:208 declares a callback as

    static const char* fake_exported_sequence(GBDATA*, unsigned long *len,
                                              const char**)

but its typedef at ARBDB/arbdb.h:134 is

    typedef const char* (*gb_export_sequence_cb)(GBDATA *gb_species,
                                                 size_t *seq_len,
                                                 GB_ERROR *error);

On LP64 `size_t` and `unsigned long` are the same type, so this is invisible
on amd64. On ILP32 they are not, the function pointer types are
incompatible, and GB_set_export_sequence_hook() is rejected with
"invalid conversion".

With all three in place, the matrix. Every cell is now a real build, not a
reproducer:

                                        amd64    i386
    7.0+ds-1 as released                  ok    FTBFS   (buildd log)
    282faf2, replace unsigned long     FTBFS      --    (your CI 10622098)
    add two-arg unsigned int only         ok    FTBFS   (what I sent you)
    both ctors + the AWTI hook            ok      ok

The last row is sbuild 0.89.3 in unshare mode against sid, 7.0+ds-1 plus the
two patches below, both architectures, built today. Both finished `Status:
successful` and produced the full set of binaries (arb, libarb, libarb-dev,
arb-common, arb-doc). Going by buildd, the last successful i386 build was
6.0.6-8 on 2024-08-24, and 7.0+ds-1 is the only 7.0 attempt there, so this
looks like the first time arb 7.0 has built on i386 at all.

I did not include your RAxML8/SATIVA patch from 0a147b11, because its
condition is `ifeq (,$(filter $(DEB_HOST_ARCH_CPU),i386 amd64))`, so it is a
no-op on both architectures I built. For this question the tree is
equivalent to master.

The amd64 cell is the direct answer to your question about CI job 10622098:
the added overloads do not make anything ambiguous on LP64, because a
non-template exact match beats the private template. What broke amd64 was
replacing the `unsigned long` overload rather than adding to it. Both have
to stay, since call sites use both types -- ARB_GDE/GDE_arbdb_io.cxx:236
passes the `unsigned long numberspecies` declared at GDE_arbdb_io.cxx:49.

If it helps, there is a stronger statement available for amd64: the patch
emits no different code there at all. I first built the same unpatched
source twice, to establish that arb builds reproducibly on this machine --
1827 files across the .debs, byte-identical both times -- and then compared
unpatched against patched. 43 files differ: the installed copy of
arb_progress.h, which is the patch itself; 28 dbgsym .debug files; and 14
binaries and libraries. In all 14, the only sections that differ are
.note.gnu.build-id and .gnu_debuglink. .text, .rodata, .data, .bss, the
symbol tables and the relocations are byte-identical. The build-id moves
because the DWARF line table records arb_progress.h as two lines longer, not
because any instruction changed.

On scope, so that none of the above is read as more than it is.

This is builds, not behaviour. I have not started the GUI or loaded a
database, so nothing here says arb is usable on i386, only that it compiles,
links and packages. There is no autopkgtest to run: arb has no debian/tests
and no Testsuite field. dh_auto_test does execute, but arb's Makefile has no
target it recognises, so it is a no-op. Upstream's UNIT_TESTER is off by
default (UNIT_TESTS := 0 in config.makefile.template) and enabling it would
be a different build configuration from the one the archive ships, so I left
it alone. What did run, as part of the build itself, is PERL_SCRIPTS/test:
31 scripts, all OK on i386. I did not run piuparts.

Separately, and not fixed here: the overflow-recovery path at
CORE/arb_progress.h:188 does `overall_count = 10L*INT_MAX;` to "fake sth
big", but on ILP32 `long` is 32 bits, so that is -10 and the recovery leaves
the counter negative. gcc warns 82 times about it during the i386 build. It
is pre-existing and harmless to the build; I mention it only because it is
in the same header and someone should know.

Finally, and only as an explanation rather than a complaint: the reason this
went unnoticed is probably that c43e7f21 set
SALSA_CI_DISABLE_BUILD_PACKAGE_I386 three weeks ago while debian/control
stayed `Architecture: any`. The archive buildds still build i386, so the bug
is live, but the only i386 build in the review loop is switched off. If you
re-enable that job for this change, it will confirm the last row by itself.

Patches below. The first replaces debian/patches/fix-i386-build.patch as
committed in 282faf2; the second is new.

Kind regards,
Cheng Li


debian/patches/fix-i386-build.patch (replaces the committed version)
====================================================================

Description: Accept 32-bit size_t in the arb_progress constructors
 CORE/arb_progress.h rejects integer types that could overflow in caller code
 by declaring two private constructor templates, and then publishes the wanted
 types explicitly.  Only `unsigned long` is published, which on LP64 happens to
 be `size_t` as well.  On ILP32 `size_t` is `unsigned int`, a distinct type, so
 an exact match against the private template beats the integral conversion to
 the public overload and every call site that passes a `size_t` fails with
 "is private within this context".
 .
 Publish the `unsigned int` variant of both constructors next to the existing
 `unsigned long` ones.  The `unsigned long` overloads must stay: call sites use
 both types, for instance ARB_GDE/GDE_arbdb_io.cxx:236 which passes the
 `unsigned long numberspecies` declared at GDE_arbdb_io.cxx:49.
Author: Cheng Li <shiptux at gmail.com>
Bug-Debian: https://bugs.debian.org/1153516
Forwarded: no
Last-Update: 2026-11-10
--- a/CORE/arb_progress.h
+++ b/CORE/arb_progress.h
@@ -256,7 +256,9 @@
     // explicitely overload wanted ctors:
     explicit arb_progress(char *title) { setup(title, 0); }
     arb_progress(const char *title, unsigned long overall_count) { arb_assert(overall_count<=LONG_MAX); setup(title, overall_count); }
+    arb_progress(const char *title, unsigned int overall_count) { arb_assert(overall_count<=LONG_MAX); setup(title, overall_count); }
     explicit arb_progress(unsigned long overall_count) { arb_assert(overall_count<=LONG_MAX); setup(NULp, overall_count); }
+    explicit arb_progress(unsigned int overall_count) { arb_assert(overall_count<=LONG_MAX); setup(NULp, overall_count); }
 
     void allow_title_reuse() { used->allow_title_reuse(); }
 

debian/patches/fix-i386-export-sequence-hook.patch (new)
========================================================

Description: Declare the export-sequence hook with size_t, as its typedef says
 gb_export_sequence_cb (ARBDB/arbdb.h:134) takes a `size_t *seq_len`, but the
 test hook in AWTI/AWTI_export.cxx declares its second parameter as
 `unsigned long *`.  On LP64 the two spellings name the same type, so this is
 invisible on amd64.  On ILP32 `size_t` is `unsigned int`, the function pointer
 types are incompatible, and GB_set_export_sequence_hook() is rejected with
 "invalid conversion".
Author: Cheng Li <shiptux at gmail.com>
Bug-Debian: https://bugs.debian.org/1153516
Forwarded: no
Last-Update: 2026-11-10
--- a/AWTI/AWTI_export.cxx
+++ b/AWTI/AWTI_export.cxx
@@ -205,7 +205,7 @@
 class ExportFieldScanner : public AvailableFieldScanner {
     RefPtr<GBDATA> gb_main;
 
-    static const char* fake_exported_sequence(GBDATA*, unsigned long *len, const char**) {
+    static const char* fake_exported_sequence(GBDATA*, size_t *len, const char**) {
         *len = 5;
         return "faked";
     }



More information about the Debian-med-packaging mailing list