Page MenuHome GnuPG
Create Task
Maniphest T7721

libgcrypt build-error with gcc-15 on powerpc and alpha
Closed, ResolvedPublic
Actions

Description

On some archs libgcrypt fails to build from source with gcc-15. Helmut Grohne writes:

libgcrypt20 fails to build from source with gcc-15 for some
architectures. The cause may be found in mpi/longlong.h. In several
lines such as 193, 499 and 1409, there are function prototypes without
arguments. This trips up the updated language standard by gcc-15 and
building libcrypt20 now complains about excess arguments. For example:

In file included from ../../mpi/mpih-div.c:32:
../../mpi/mpih-div.c: In function ‘_gcry_mpih_mod_1’:
../../mpi/longlong.h:190:11: error: too many arguments to function ‘__udiv_qrnnd’; expected 0, have 4
  190 |     (q) = __udiv_qrnnd (&__r, (n1), (n0), (d));                         \
      |           ^~~~~~~~~~~~  ~~~~
../../mpi/mpih-div.c:84:17: note: in expansion of macro ‘udiv_qrnnd’
   84 |                 udiv_qrnnd(divisor_limb_inverted, dummy,
      |                 ^~~~~~~~~~
../../mpi/longlong.h:193:16: note: declared here

This does not seem to affect amd64 as Matthias would have otherwise
reported this problem already. It does affect at least alpha.

I do not see this on arm64, powerpc64le shows a different issue:

libtool: link: ( cd ".libs" && rm -f "libgcrypt.la" && ln -s "../libgcrypt.la" "libgcrypt.la" )
/bin/bash ../libtool  --tag=CC   --mode=link gcc-15  -g -O2 -Werror=implicit-function-declaration -ffile-prefix-map=/home/ametzler/GCR/libgcrypt20-1.11.1=. -fstack-protector-strong -Wformat -Werror=format-security -fvisibility=hidden -fno-delete-null-pointer-checks -Wall  -Wl,-z,relro -Wl,-z,now -o mpicalc mpicalc-mpicalc.o libgcrypt.la -L/usr/lib/powerpc64le-linux-gnu -lgpg-error
libtool: link: gcc-15 -g -O2 -Werror=implicit-function-declaration -ffile-prefix-map=/home/ametzler/GCR/libgcrypt20-1.11.1=. -fstack-protector-strong -Wformat -Werror=format-security -fvisibility=hidden -fno-delete-null-pointer-checks -Wall -Wl,-z -Wl,relro -Wl,-z -Wl,now -o .libs/mpicalc mpicalc-mpicalc.o  ./.libs/libgcrypt.so -L/usr/lib/powerpc64le-linux-gnu -lgpg-error
/usr/bin/ld: ./.libs/libgcrypt.so: undefined reference to `gcry_poly1305_p10le_4blocks'
/usr/bin/ld: ./.libs/libgcrypt.so: undefined reference to `_gcry_chacha20_p10le_8x'
collect2: error: ld returned 1 exit status
make[3]: *** [Makefile:662: mpicalc] Error 1
make[3]: Leaving directory '/home/ametzler/GCR/libgcrypt20-1.11.1/build/src'

Comparing the buildlogs for gcc14 and 15 on this arch shows no errors/warnings, I was surprised that support for "checking whether GCC inline assembler supports PowerISA 3.00 instructions." stopped succeeding with gcc-15.

Details

External Link
https://bugs.debian.org/1109071
Version
1.11.0

Revisions and Commits

rC libgcrypt
rC0c2d120e1124 poly1305-p10le: use '.rodata' section for read-only data
rC660ed722aed2 Add missing machine tags for PowerPC assembly
rC5cdb3b08fdd4 Add missing abiversion tag for PowerPC assembly
rCb61a7661d017 mpi: Provide the function prototype of __udiv_qrnnd.

Related Objects

Event Timeline

ametzler1 created this task.Jul 14 2025, 1:14 PM
• werner added a subscriber: • werner.Jul 14 2025, 2:44 PM

Sure that this is about 1.11.0 ? We released 1.11.1 with at least one fix for gcc regression (T7166). In master we had some more fixes for gcc 15 bugs (or what ever you will call such regression in a compiler)

• werner added a project: libgcrypt.Jul 14 2025, 2:44 PM
ametzler1 added a comment.Jul 14 2025, 3:23 PM
In T7721#202963, @werner wrote:

Sure that this is about 1.11.0 ? We released 1.11.1 with at least one fix for gcc regression (T7166). In master we had some more fixes for gcc 15 bugs (or what ever you will call such regression in a compiler)

Helmut tested on 1.11.0, the powerpc64le issue was reproduced by me on 1.11.1.
I have asked Helmut on which archs except alpha (I have no access there) I should be able to reproduce the original issue.

• werner renamed this task from libgcrypt build-error with gcc-15 to libgcrypt build-error with gcc-15 on powerpc and alpha.Jul 14 2025, 4:54 PM
ametzler1 added a comment.Jul 15 2025, 1:35 PM

The powerpc64le issue (undefined reference to `gcry_poly1305_p10le_4blocks') also applies to GIT master.

• gniibe claimed this task.Jul 18 2025, 4:30 AM
• gniibe triaged this task as Normal priority.
• gniibe added a subscriber: • gniibe.

For PowerISA 3.00 Instructions issue, following patch may help:

diff --git a/configure.ac b/configure.ac
index 6cc1e189..70d632af 100644
--- a/configure.ac
+++ b/configure.ac
@@ -2448,10 +2448,11 @@ AC_CACHE_CHECK([whether GCC inline assembler supports PowerISA 3.00 instructions
         else
           gcry_cv_gcc_inline_asm_ppc_arch_3_00=no
           AC_LINK_IFELSE([AC_LANG_PROGRAM(
-          [[__asm__(".text\n\t"
+          [[__asm__(".machine        \"any\"\n"
+                    ".text\n\t"
                     ".globl testfn;\n"
                     "testfn:\n"
-                    "stxvb16x %r1,%v12,%v30;\n"
+                    "stxvb16x 47,0,9;\n"
                   );
             void testfn(void);
             ]], [ testfn(); ])],

I figured out that .machine "any" is needed with GCC 15.

• gniibe added a comment.Jul 18 2025, 4:33 AM

For Alpha (hppa, and sparc), IIUC, following patch may help:

diff --git a/mpi/longlong.h b/mpi/longlong.h
index 7dc67591..9e60592a 100644
--- a/mpi/longlong.h
+++ b/mpi/longlong.h
@@ -192,7 +192,7 @@ SPDX-License-Identifier: LGPL-2.1-or-later
     (q) = __udiv_qrnnd (&__r, (n1), (n0), (d)); 			\
     (r) = __r;								\
   } while (0)
-extern UDItype __udiv_qrnnd ();
+extern UDItype __udiv_qrnnd (UDItype *, UDItype, UDItype, UDItype);
 #  define UDIV_TIME 220
 # endif /* !LONGLONG_STANDALONE */
 #endif /* __alpha */
@@ -448,7 +448,7 @@ extern UDItype __udiv_qrnnd ();
     (q) = __udiv_qrnnd (&__r, (n1), (n0), (d)); 			\
     (r) = __r;								\
   } while (0)
-extern USItype __udiv_qrnnd ();
+extern USItype __udiv_qrnnd (USItype *, USItype, USItype, USItype);
 # endif /* !LONGLONG_STANDALONE && !ASM_DISABLED */
 # define count_leading_zeros(count, x) \
   do {								       \
@@ -1408,7 +1408,7 @@ typedef unsigned int UTItype __attribute__ ((mode (TI)));
     (q) = __udiv_qrnnd (&__r, (n1), (n0), (d)); 			\
     (r) = __r;								\
   } while (0)
-extern USItype __udiv_qrnnd ();
+extern USItype __udiv_qrnnd (USItype *, USItype, USItype, USItype);
 #   define UDIV_TIME 140
 #  endif /* LONGLONG_STANDALONE */
 # endif /* udiv_qrnnd */

I checked GCC for longlong.h. It's gone for hppa and sparc (C implementation is enough). For Alpha, we see adding the function prototype.

• gniibe added a comment.Jul 18 2025, 4:51 AM

I figured out that .machine "any" is needed with GCC 15.

Perhaps, cipher/poly1305-p10le.s would also need this. (I only tested for the check of stxvb16x instruction.)

• gniibe mentioned this in T7642: Release Libgcrypt 1.11.2.Jul 18 2025, 6:55 AM
ametzler1 added a comment.Jul 18 2025, 12:08 PM
In T7721#203182, @gniibe wrote:

For PowerISA 3.00 Instructions issue, following patch may help:

diff --git a/configure.ac b/configure.ac
index 6cc1e189..70d632af 100644
--- a/configure.ac
+++ b/configure.ac
@@ -2448,10 +2448,11 @@ AC_CACHE_CHECK([whether GCC inline assembler supports PowerISA 3.00 instructions
         else
           gcry_cv_gcc_inline_asm_ppc_arch_3_00=no
           AC_LINK_IFELSE([AC_LANG_PROGRAM(
-          [[__asm__(".text\n\t"
+          [[__asm__(".machine        \"any\"\n"
+                    ".text\n\t"
                     ".globl testfn;\n"
                     "testfn:\n"
-                    "stxvb16x %r1,%v12,%v30;\n"
+                    "stxvb16x 47,0,9;\n"
                   );
             void testfn(void);
             ]], [ testfn(); ])],

I figured out that .machine "any" is needed with GCC 15.

With this patch we get

checking whether GCC inline assembler supports PowerISA 3.00 instructions... yes

again, but the build fails with

/bin/bash ../libtool    --mode=compile gcc-15 -Wa,--noexecstack -g -O2 -Werror=implicit-function-declaration -ffile-prefix-map=/home/ametzler/GCR/libgcrypt20-1.11.1=. -fstack-protector-strong -Wformat -Werror=format-security -c -o poly1305-p10le.lo ../../cipher/poly1305-p10le.s
libtool: compile:  gcc-15 -Wa,--noexecstack -g -O2 -Werror=implicit-function-declaration -ffile-prefix-map=/home/ametzler/GCR/libgcrypt20-1.11.1=. -fstack-protector-strong -Wformat -Werror=format-security -c ../../cipher/poly1305-p10le.s  -fPIC -DPIC -o .libs/poly1305-p10le.o
../../cipher/poly1305-p10le.s: Assembler messages:
../../cipher/poly1305-p10le.s:381: Error: unrecognized opcode: `lxv'
../../cipher/poly1305-p10le.s:382: Error: unrecognized opcode: `lxv'
../../cipher/poly1305-p10le.s:395: Error: unrecognized opcode: `mtvsrdd'
../../cipher/poly1305-p10le.s:397: Error: unrecognized opcode: `mtvsrdd'
../../cipher/poly1305-p10le.s:399: Error: unrecognized opcode: `mtvsrdd'
../../cipher/poly1305-p10le.s:401: Error: unrecognized opcode: `mtvsrdd'
../../cipher/poly1305-p10le.s:402: Error: unrecognized opcode: `mtvsrdd'
../../cipher/poly1305-p10le.s:406: Error: unrecognized opcode: `mtvsrdd'
../../cipher/poly1305-p10le.s:514: Error: unrecognized opcode: `mtvsrdd'
../../cipher/poly1305-p10le.s:518: Error: unrecognized opcode: `mtvsrdd'
../../cipher/poly1305-p10le.s:520: Error: unrecognized opcode: `mtvsrdd'
../../cipher/poly1305-p10le.s:522: Error: unrecognized opcode: `mtvsrdd'
../../cipher/poly1305-p10le.s:524: Error: unrecognized opcode: `mtvsrdd'
../../cipher/poly1305-p10le.s:525: Error: unrecognized opcode: `mtvsrdd'
../../cipher/poly1305-p10le.s:772: Error: unrecognized opcode: `mfvsrld'
../../cipher/poly1305-p10le.s:776: Error: unrecognized opcode: `mfvsrld'
../../cipher/poly1305-p10le.s:777: Error: unrecognized opcode: `mfvsrld'
make[3]: *** [Makefile:983: poly1305-p10le.lo] Error 1
make[3]: Leaving directory '/home/ametzler/GCR/libgcrypt20-1.11.1/build/cipher'
jukivili added a subscriber: jukivili.Jul 21 2025, 11:04 AM

I tested Ubuntu's version of GCC-15 (powerpc64le cross-compiler) and did not see this build failure:

$ powerpc64le-linux-gnu-gcc-15 --version
powerpc64le-linux-gnu-gcc-15 (Ubuntu 15-20250404-0ubuntu1) 15.0.1 20250404 (experimental) [master r15-9193-g08e803aa9be]
checking whether compiler supports PowerPC AltiVec/VSX/crypto intrinsics... yes
checking whether GCC inline assembler supports PowerPC AltiVec/VSX/crypto instructions... yes
checking whether GCC inline assembler supports PowerISA 3.00 instructions... yes

Build also succeeds with my setup.

Here's patches which might help... at least they do not break my build:

0003-poly1305-p10le-use-.rodata-section-for-read-only-dat.patch810 BDownload

0001-Add-missing-machine-tags-for-PowerPC-assembly.patch2 KBDownload

0002-Add-missing-abiversion-tag-for-PowerPC-assembly.patch1 KBDownload

• gniibe added a comment.Jul 23 2025, 7:18 AM

IIUC, it's actually binutils version dependency (instead of GCC 15), perhaps.

Well, let me apply the patch of longlong firstly.

Then, please apply PowerPC patches of yours, @jukivili .

• gniibe mentioned this in Unknown Object (Maniphest Task).Jul 28 2025, 8:00 AM
• gniibe added a comment.Jul 30 2025, 8:58 AM

I pushed the longlong patch: rCb61a7661d017: mpi: Provide the function prototype of __udiv_qrnnd.

• gniibe added a commit: rCb61a7661d017: mpi: Provide the function prototype of __udiv_qrnnd..Jul 30 2025, 9:42 AM
jukivili added a comment.Jul 30 2025, 10:00 PM

Ok, thanks. I pushed the powerpc patches to master.

jukivili added a commit: rC660ed722aed2: Add missing machine tags for PowerPC assembly.Jul 30 2025, 10:01 PM
jukivili added a commit: rC5cdb3b08fdd4: Add missing abiversion tag for PowerPC assembly.
jukivili added a commit: rC0c2d120e1124: poly1305-p10le: use '.rodata' section for read-only data.
• gniibe mentioned this in Unknown Object (Maniphest Task).Aug 4 2025, 8:02 AM
• gniibe changed the task status from Open to Testing.Aug 5 2025, 7:57 AM
ametzler1 added a comment.Aug 9 2025, 10:51 AM

Hello,
thank you all. I can confirm that 1.11.2 builds successfully on ppc64el with gcc-15 (Debian sid + experimental). Lacking access I have not be able to check alpha. I would suggest closing this report as fixed.
cu Andreas

• werner closed this task as Resolved.Aug 10 2025, 1:59 PM

Thanks for testing.