build: --enable-debug compiles at -O2 instead of -O0 (regression since 2018)

[ThalesBarretto]

Summary

./configure --enable-debug is meant to produce an unoptimized, debuggable build: its own
code sets -g -O0 -DDEBUG. The resulting build is in fact compiled at -O2. -g and
-DDEBUG take effect, but the -O0 is silently overridden, so a "debug" build is as hard to
step through in gdb as a release build (inlined frames, "value optimized out"). This has been
the case since b91325caa3 (2018-12-28) and also affects rpmbuild --with debug.

Observed on devel @ a482a8578a; reproduced on 14 distro toolchains (details below).

Root cause

--enable-debug puts -O0 into GF_CFLAGS, which reaches the compiler through
AM_CFLAGS / the per-target _CFLAGS. automake emits every compile as
… $(AM_CFLAGS) $(CFLAGS) …, so CFLAGS is placed last. autoconf's AC_PROG_CC
defaults CFLAGS to -g -O2 for gcc (and the RPM %configure macro exports
CFLAGS=%{optflags}, i.e. -O2 …), and configure.ac never neutralizes that. gcc honours
the last -O on the command line, so the -O2 from CFLAGS wins over the -O0
from GF_CFLAGS.

Real compile line for one TU under ./configure --enable-debug (make V=1):

gcc … -g -O0 -DDEBUG …(from GF_CFLAGS)… -g -O2 -fsigned-char …(from CFLAGS)… -c mem-pool.c
        └─ intended                                └─ wins (last -O)

Decisive check (gcc defines __OPTIMIZE__ iff the effective level is ≥ -O1):

gcc -O0 -DDEBUG -O2 …  → __OPTIMIZE__ defined    (effective -O2)
gcc -O0            …  → __OPTIMIZE__ undefined  (effective -O0)

History: this used to work

• 2012, d9ddb5fb95 — --enable-debug introduced; put the level in CFLAGS
  (CFLAGS="-O0 $CFLAGS"), emitted last, so -O0 won.
• 2013, d85e394847 (RHBZ 955283, "enable-debug change breaks _hardened_build") — kept
  -O0 in CFLAGS via a surgical sed -e s/O2/O0/, so distro hardening flags survived.
• 2018, b91325caa3 ("configure: fix the duplicate CFLAGS options") — moved -g -O0 -DDEBUG
  from CFLAGS to GF_CFLAGS while removing genuine flag duplication elsewhere. That demoted
  the -O0 below CFLAGS. Building the same ./configure --enable-debug on that commit and
  on its parent shows the flip: parent → effective -O0, commit → effective -O2.

Why CI and the in-tree helpers did not catch it

The automated consumers of --enable-debug never see autoconf's -O2:

• upstream Jenkins regression CI (build.sh in the separate gluster/glusterfs-patch-acceptance-tests
  repository) configures with --enable-debug and then runs make install CFLAGS="-Wall -Werror -Wno-cpp". That make-line
  override replaces configure's entire CFLAGS, so the -O2 is gone and the -O0 carried by
  GF_CFLAGS is the only optimisation token. CI has therefore been building at -O0 all along.
• the in-tree helpers (run-tests-in-vagrant.sh, extras/devel-tools/devel-vagrant,
  extras/distributed-testing) pass an explicit CFLAGS="-g -O0 …" next to --enable-debug.

Only a plain ./configure --enable-debug (what a developer types) and rpmbuild --with debug
(where %configure exports CFLAGS=%optflags) hit it, and in both cases the build still
succeeds, just optimized, so nothing fails loudly.

Impact

• Developer debug builds are optimized: gdb shows inlined frames and <optimized out> variables.
• rpmbuild --with debug produces optimized "debug" RPMs.
• -DDEBUG instrumentation (mem-pool accounting etc.) and -g are unaffected.
• No distro's default packaging passes --enable-debug, so no shipped release binary is
  affected either way.

Fix

Keep -g -DDEBUG in GF_CFLAGS and append -O0 to CFLAGS, so it lands last and wins,
without clobbering anything already in CFLAGS (distro hardening flags stay intact — the
mistake behind RHBZ 955283 was clobbering CFLAGS, and appending avoids it; it is also robust
to -O3/-Os/-Og/no -O, unlike the 2013 sed). A PR with the one-hunk change follows.

Verification of the fix

• Arch Linux (2026-09; gcc 16.2.1, glibc 2.44, autoconf 2.73): --enable-debug → effective
  -O0, both plain and with Fedora 42 %optflags as CFLAGS (-fstack-protector-strong,
  -D_FORTIFY_SOURCE=3, -fcf-protection, -fstack-clash-protection all still on the line);
  full tree builds with 0 errors in three allocator/io configurations.
• AlmaLinux 8.10 (gcc 8.5): real rpmbuild -tb --with debug before/after — before, all 351
  translation units end with -O2; after, all 351 end with -O0; both builds succeed and
  produce the same 39 packages; EL8 hardening flags intact.
• 14 distro containers (AlmaLinux 8, CentOS Stream 9, Fedora 42, Leap 15.6, Tumbleweed, Arch,
  Debian 12, Debian 13 amd64, Ubuntu 20.04/22.04/24.04, plus Debian 13 i386, Alpine 3.22 and
  Alpine edge as compile-line-only oracles;
  autoconf 2.69–2.73, gcc 7.5–16.2, glibc 2.28–2.44 and musl), each with its own autotools:
  identical flip on every one, same make result / error set / object counts before and after,
  and the generated build system differs only in the CFLAGS/GF_CFLAGS lines.
• CI shape (make CFLAGS="-Wall -Werror -Wno-cpp" after --enable-debug): stays at -O0
  (the line then carries no -O token, which gcc treats as -O0).

One expected side effect of any real -O0 build: where CFLAGS carries -D_FORTIFY_SOURCE,
glibc's features.h emits #warning _FORTIFY_SOURCE requires compiling with optimization (-O)
once per TU (Fedora/EL/SUSE/Arch glibc; Debian/Ubuntu glibc carries no such warning). That was
also the state of --enable-debug from 2013 to 2018.

[ThalesBarretto]

Related past issues for context:

• Debug should be compiled with '-Og' and not '-O0' #1195 (2020) discussed whether --enable-debug should use -Og instead of -O0; at that time -O2 was already taking precedence because of b91325c (2018-12). The -Og question is orthogonal to this regression and can be tracked separately if desired.
• Modernize compiler and linker flags #544 (2018) shows the duplicated -g -O2 -g -rdynamic … compile line that b91325c cleaned up; moving -O0 into GF_CFLAGS was a side effect of that cleanup. build: honour -O0 for --enable-debug builds #4790 preserves that de-duplication while restoring -O0 precedence under --enable-debug.