The pfx format has a huge performance loss with clang

[claudioandre-br]

• I can easily reproduce:

gcc from Ubuntu 24

$ john | head -1; sleep 3; john --test --format=pfx
John the Ripper 1.9.0-jumbo-1+bleeding-b3bd5ea707 2024-12-01 03:06:08 +0100 OMP [linux-gnu 64-bit x86_64 AVX2 AC]
Will run 8 OpenMP threads
Benchmarking: pfx, (.pfx, .p12) [PKCS#12 PBE (SHA1/SHA2) 256/256 AVX2 8x]... (8xOMP) DONE
Speed for cost 1 (iteration count) of 2048, cost 2 (mac-type [1:SHA1 224:SHA224 256:SHA256 384:SHA384 512:SHA512]) of 1
Raw:	26842 c/s real, 3696 c/s virtual

clang from Ubuntu 24.

$ $ run/john | head -1; sleep 3; run/john --test --format=pfx
John the Ripper 1.9.0-jumbo-1+bleeding-364b1ca435 2024-12-02 06:20:23 +0100 [linux-gnu 64-bit x86_64 AVX2 AC]
Benchmarking: pfx, (.pfx, .p12) [PKCS#12 PBE (SHA1/SHA2) 256/256 AVX2 8x2]... DONE
Speed for cost 1 (iteration count) of 2048, cost 2 (mac-type [1:SHA1 224:SHA224 256:SHA256 384:SHA384 512:SHA512]) of 1
Raw:	9792 c/s real, 9792 c/s virtual

$ run/john | head -1; sleep 3; run/john --test --format=pfx
John the Ripper 1.9.0-jumbo-1+bleeding-0251a0f0f8 2024-12-02 18:56:33 -0300 [linux-gnu 64-bit x86_64 AVX2 AC]
Benchmarking: pfx, (.pfx, .p12) [PKCS#12 PBE (SHA1/SHA2) 256/256 AVX2 8x]... DONE
Speed for cost 1 (iteration count) of 2048, cost 2 (mac-type [1:SHA1 224:SHA224 256:SHA256 384:SHA384 512:SHA512]) of 1
Raw:	11328 c/s real, 11328 c/s virtual

[solardiz]

We probably need to re-tune these:

#ifndef SIMD_PARA_SHA1
#if defined(__INTEL_COMPILER)
#define SIMD_PARA_SHA1                  1
#elif defined(__clang__)
#define SIMD_PARA_SHA1                  2
#elif defined(__llvm__)
#define SIMD_PARA_SHA1                  2
#elif defined(__XOP__)
#define SIMD_PARA_SHA1                  2
#else
#define SIMD_PARA_SHA1                  1
#endif
#endif

Can you try making it 1 for clang?

Also, not limited to SHA-1, I think we never considered possibly increasing these PARA for AVX-512, where we have twice more SIMD registers (up to 32 now). This is unrelated to this issue, which was seen on AVX2 (so still 16 registers).