simd_32 test fails on ppc64le

[thmo]

For 2.4.0, with gcc 8.0.1 on ppc64le, we see the following test failures:

simd_32:
SIMD_4x32 ran 240 tests 8 FAILED
Failure 1: SIMD_4x32 SIMD_4x32 bswap elem0 BE produced unexpected result '4041265344' expected '3721182122'
Failure 2: SIMD_4x32 SIMD_4x32 bswap elem1 BE produced unexpected result '1732584193' expected '558065031'
Failure 3: SIMD_4x32 SIMD_4x32 bswap elem2 BE produced unexpected result '558065031' expected '1732584193'
Failure 4: SIMD_4x32 SIMD_4x32 bswap elem3 BE produced unexpected result '3721182122' expected '4041265344'
Failure 5: SIMD_4x32 SIMD_4x32 bswap elem0 LE produced unexpected result '4041265344' expected '3721182122'
Failure 6: SIMD_4x32 SIMD_4x32 bswap elem1 LE produced unexpected result '1732584193' expected '558065031'
Failure 7: SIMD_4x32 SIMD_4x32 bswap elem2 LE produced unexpected result '558065031' expected '1732584193'
Failure 8: SIMD_4x32 SIMD_4x32 bswap elem3 LE produced unexpected result '3721182122' expected '4041265344'

There are warnings during compilation:

g++ -fstack-protector -pthread -std=c++11 -D_REENTRANT -O2 -g -pipe -Wall -Werror=format-security -Wp,-D_FORTIFY_SOURCE=2 -Wp,-D_GLIBCXX_ASSERTIONS -fexceptions -fstack-protector-strong -grecord-gcc-switches -specs=/usr/lib/rpm/redhat/redhat-hardened-cc1 -specs=/usr/lib/rpm/redhat/redhat-annobin-cc1 -m64 -mcpu=power8 -mtune=power8 -funwind-tables -fstack-clash-protection -Wall -Wextra -Wpedantic -Wstrict-aliasing -Wcast-align -Wmissing-declarations -Wpointer-arith -Wcast-qual -Wzero-as-null-pointer-constant -Wnon-virtual-dtor -maltivec -Ibuild/include -Ibuild/include/external -c src/tests/test_simd.cpp -o build/obj/test/test_simd.o
In file included from src/tests/test_simd.cpp:10:
build/include/botan/internal/simd_32.h: In constructor 'Botan::SIMD_4x32::SIMD_4x32(const uint32_t*)':
build/include/botan/internal/simd_32.h:82:64: warning: ISO C++ forbids compound-literals [-Wpedantic]
          m_vmx = (__vector unsigned int){B[0], B[1], B[2], B[3]};
                                                                ^
build/include/botan/internal/simd_32.h: In constructor 'Botan::SIMD_4x32::SIMD_4x32(uint32_t, uint32_t, uint32_t, uint32_t)':
build/include/botan/internal/simd_32.h:101:56: warning: ISO C++ forbids compound-literals [-Wpedantic]
          m_vmx = (__vector unsigned int){B0, B1, B2, B3};
                                                        ^
build/include/botan/internal/simd_32.h: In static member function 'static Botan::SIMD_4x32 Botan::SIMD_4x32::load_le(const void*)':
build/include/botan/internal/simd_32.h:141:57: warning: vec_lvsl is deprecated for little endian; use assignment for unaligned loads and stores [-Wdeprecated]
          __vector unsigned char perm = vec_lvsl(0, in_32);
                                                         ^
build/include/botan/internal/simd_32.h: In static member function 'static Botan::SIMD_4x32 Botan::SIMD_4x32::load_be(const void*)':
build/include/botan/internal/simd_32.h:182:57: warning: vec_lvsl is deprecated for little endian; use assignment for unaligned loads and stores [-Wdeprecated]
          __vector unsigned char perm = vec_lvsl(0, in_32);
                                                         ^
build/include/botan/internal/simd_32.h: In member function 'Botan::SIMD_4x32 Botan::SIMD_4x32::rho() const':
build/include/botan/internal/simd_32.h:310:74: warning: ISO C++ forbids compound-literals [-Wpedantic]
          res.m_vmx = vec_rl(m_vmx, (__vector unsigned int){r1, r1, r1, r1});
                                                                          ^
build/include/botan/internal/simd_32.h:311:93: warning: ISO C++ forbids compound-literals [-Wpedantic]
          res.m_vmx = vec_xor(res.m_vmx, vec_rl(m_vmx, (__vector unsigned int){r2, r2, r2, r2}));
                                                                                             ^
build/include/botan/internal/simd_32.h:312:93: warning: ISO C++ forbids compound-literals [-Wpedantic]
          res.m_vmx = vec_xor(res.m_vmx, vec_rl(m_vmx, (__vector unsigned int){r3, r3, r3, r3}));
                                                                                             ^
build/include/botan/internal/simd_32.h: In member function 'Botan::SIMD_4x32 Botan::SIMD_4x32::rotl() const':
build/include/botan/internal/simd_32.h:357:75: warning: ISO C++ forbids compound-literals [-Wpedantic]
          return SIMD_4x32(vec_rl(m_vmx, (__vector unsigned int){r, r, r, r}));
                                                                           ^
build/include/botan/internal/simd_32.h: In member function 'Botan::SIMD_4x32 Botan::SIMD_4x32::shl() const':
build/include/botan/internal/simd_32.h:519:75: warning: ISO C++ forbids compound-literals [-Wpedantic]
          return SIMD_4x32(vec_sl(m_vmx, (__vector unsigned int){s, s, s, s}));
                                                                           ^
build/include/botan/internal/simd_32.h: In member function 'Botan::SIMD_4x32 Botan::SIMD_4x32::shr() const':
build/include/botan/internal/simd_32.h:537:75: warning: ISO C++ forbids compound-literals [-Wpedantic]
          return SIMD_4x32(vec_sr(m_vmx, (__vector unsigned int){s, s, s, s}));
                                                                           ^
build/include/botan/internal/simd_32.h: In member function 'Botan::SIMD_4x32 Botan::SIMD_4x32::bswap() const':
build/include/botan/internal/simd_32.h:596:83: warning: vec_lvsl is deprecated for little endian; use assignment for unaligned loads and stores [-Wdeprecated]
          __vector unsigned char perm = vec_lvsl(0, static_cast<uint32_t*>(nullptr));
                                                                                   ^

not sure if that is related.

[thmo]

Noted that the test results are reversed compared to the expected results.

[randombit]

I don't see this problem with GCC 7.3 on ppc64le (CentOS 7) so it may be something about GCC 8.0. I don't think the warnings are related.

[randombit]

I can reproduce the error with gcc version 8.0.1 20180219 (experimental)

[randombit]

With -O0 everything works as expected, with -O1 fails. So I am highly suspicious this is a bug in GCC 8.

[randombit]

What is really strange is none of the SIMD enabled block ciphers see any test failures. But I can confirm that they are enabled (by seeing the performance difference between default and --clear-cpuid=altivec)

[randombit]

I was unable to reduce it to a small testcase, so it's hard to say for sure if this is a GCC bug or not. But I did find a reformation of the function (0ebad38) which passes the tests with GCC 8 so I guess that's something. Can you try that patch with 2.4.0? Should backport cleanly.

[thmo]

Unfortunately, that doesn't seem to help in our build environment, the test still fails:
https://kojipkgs.fedoraproject.org//work/tasks/6163/25746163/build.log

[randombit]

Ugh. Can you try this:

          vec.V = m_vmx;
          bswap_4(vec.R);
-         return SIMD_4x32(vec.V);
+         return SIMD_4x32(vec.R[0], vec.R[1], vec.R[2], vec.R[3]);

[thmo]

That seems to work.

[randombit]

Thanks for confirming. Fix applied to master.