Build failure of 32-bit kernel .config with clang-17: fs/bcachefs/replicas.c:36:52: error: builtin functions must be directly called (builds ok with gcc-13.2), kernel 6.7.0

[ernsteiswuerfel]

Get this on my Ryzen when I use clang 17.0.6 to build a kernel for my Thinkpad T60 (same config builds ok with gcc 13.2.1):

#  LLVM=1 LLVM_IAS=1 make
  CALL    scripts/checksyscalls.sh
  CC [M]  fs/bcachefs/replicas.o
fs/bcachefs/replicas.c:36:52: error: builtin functions must be directly called
   36 |         eytzinger0_sort(r->entries, r->nr, r->entry_size, memcmp, NULL);
      |                                                           ^
./arch/x86/include/asm/string_32.h:159:16: note: expanded from macro 'memcmp'
  159 | #define memcmp __builtin_memcmp
      |                ^
fs/bcachefs/replicas.c:836:9: error: builtin functions must be directly called
  836 |                       memcmp, NULL);
      |                       ^
./arch/x86/include/asm/string_32.h:159:16: note: expanded from macro 'memcmp'
  159 | #define memcmp __builtin_memcmp
      |                ^
2 errors generated.
make[4]: *** [scripts/Makefile.build:243: fs/bcachefs/replicas.o] Fehler 1
make[3]: *** [scripts/Makefile.build:480: fs/bcachefs] Fehler 2
make[2]: *** [scripts/Makefile.build:480: fs] Fehler 2
make[1]: *** [/usr/src/linux-stable/Makefile:1911: .] Fehler 2
make: *** [Makefile:234: __sub-make] Fehler 2

Kernel .config attached. config_670_p3.txt

[immolo]

I've tried to recreate this on my Gentoo LLVM17 system using:

make localyesconfig
make nconfig (add all bcachefs driver support)
LLVM=1 LLVM_TAS=1 make

However was unable to hit this bug, could please share some more system information and we can see if we can find the root cause of this one.

[ernsteiswuerfel]

@immolo The build system does not seem to matter much. Tried building my 32-bit .config from 1st post on another system (AMD FX-8370 based) where I get the same error with clang-17. However my v6.7.0 64-bit build for the FX-9370 itself builds just fine with clang-17.

So I tried building my 32-bit .config from above with 64BIT=Y set and in this case it also builds just fine.

It seems to hit this issue you need a 32-bit kernel .config. And the same .config builds ok when a 64-bit kernel build is done.

So far it seems this is an 32-bit only issue. Part of the error message ./arch/x86/include/asm/string_32.h:159:16: note: expanded from macro 'memcmp' also points in that direction.

Some data of my build system here attached anyhow. emerge_info-FX8370.txt

[koverstreet]

This is really a compiler bug, but there's a simple workaround - I'll have a fix up soon.

[ernsteiswuerfel]

If it is a compiler bug should I perhaps open a clang-17 issue on their github page and link here?

Edit; This already has been reported earlier as clang issue #1932

[nathanchance]

Sorry for not continuing to investigate that issue. For what it's worth, the workaround is needed in another place within that file:

diff --git a/fs/bcachefs/replicas.c b/fs/bcachefs/replicas.c
index 58536898b142..532f75054d0c 100644
--- a/fs/bcachefs/replicas.c
+++ b/fs/bcachefs/replicas.c
@@ -6,6 +6,12 @@
 #include "replicas.h"
 #include "super-io.h"

+/* Some (buggy!) compilers don't allow memcmp to be passed as a pointer */
+static int bch2_memcmp(const void *l, const void *r, size_t size)
+{
+       return memcmp(l, r, size);
+}
+
 static int bch2_cpu_replicas_to_sb_replicas(struct bch_fs *,
                                            struct bch_replicas_cpu *);

@@ -33,7 +39,7 @@ void bch2_replicas_entry_sort(struct bch_replicas_entry_v1 *e)

 static void bch2_cpu_replicas_sort(struct bch_replicas_cpu *r)
 {
-       eytzinger0_sort(r->entries, r->nr, r->entry_size, memcmp, NULL);
+       eytzinger0_sort(r->entries, r->nr, r->entry_size, bch2_memcmp, NULL);
 }

 static void bch2_replicas_entry_v0_to_text(struct printbuf *out,
@@ -812,12 +818,6 @@ static int bch2_cpu_replicas_to_sb_replicas(struct bch_fs *c,
        return 0;
 }

-/* Some (buggy!) compilers don't allow memcmp to be passed as a pointer */
-static int bch2_memcmp(const void *l, const void *r, size_t size)
-{
-       return memcmp(l, r, size);
-}
-
 static int bch2_cpu_replicas_validate(struct bch_replicas_cpu *cpu_r,
                                      struct bch_sb *sb,
                                      struct printbuf *err)

The LLVM change that appears to introduce this problem has been around for some time (https://github.com/llvm/llvm-project/commit/34866c7719f893c957d93f5918f760b6edebd6be) and I suspect this likely has to do with architectural differences between clang and GCC. This is really exacerbated by CONFIG_FORTIFY_SOURCE not working for ARCH=i386 at all (due to https://github.com/ClangBuiltLinux/linux/issues/1583, which I think we can now fix in the kernel sources), hence the #define memcmp __builtin_memcmp.

[ernsteiswuerfel]

I can confirm that both of your patches combined applied on v6.7.0 fix my 32-bit clang-17 build issues. 🥳 Thanks!