This is the forward direction on encryption only. Crypto++ uses the "Equivalent Inverse Cipher" (FIPS-197, Section 5.3.5, p.23), and it is not compatible with IBM hardware. The library library will need to re-work the decryption key scheduling routines. (We may be able to work around it another way, but I have not investigated it).
The strategy of "cleanup under-aligned buffers" is not scaling well. Corner cases are still turing up. The library has some corner-case breaks, like old 32-bit Intels. And it still has not solved the AltiVec and Power8 alignment problems.
For now we are backing out the changes and investigating other strategies
This commit supports the upcoming AltiVec and Power8 processor. This commit affects a number of classes due to the ubiquitous use of AES. The commit adds debug asserts to warn of under-aligned and misaligned buffers in debug builds.
This commit supports the upcoming AltiVec and Power8 processor. This commit affects a number of classes due to the ubiquitous use of AES. The commit provides the data alignment requirements.
This commit supports the upcoming AltiVec and Power8 processor support for stream ciphers. This commit affects GlobalRNG() most because its an AES-based generator. The commit favors AlignedSecByteBlock over SecByteBlock in places where messages are handled on the AltiVec and Power8 processor data paths. The data paths include all block cipher modes of operation, and some filters like FilterWithBufferedInput.
Intel and ARM processors are tolerant of under-aligned buffers when using crypto instructions. AltiVec and Power8 are less tolerant, and they simply ignore the three low-order bits to ensure an address is aligned. The AltiVec and Power8 have caused a fair number of wild writes on the stack and in the heap.
Testing on a 64-bit Intel Skylake show a marked improvement in performance. We suspect GCC is generating better code since it knows the alignment of the pointers, and does not have to emit fixup code for under-aligned and mis-aligned data. Testing on an mid-2000s 32-bit VIA C7-D with SSE2+SSSE3 showed no improvement, and no performance was lost.
This commit supports the upcoming AltiVec and Power8 processor support for DefaultEncryptors and DefaultDecryptors. The commit favors AlignedSecByteBlock over SecByteBlock in places where messages are handled on the AltiVec and Power8 processor data paths. The data paths include all block cipher modes of operation, and some filters like FilterWithBufferedInput.
Intel and ARM processors are tolerant of under-aligned buffers when using crypto intstructions. AltiVec and Power8 are less tolerant, and they simply ignore the three low-order bits to ensure an address is aligned. The AltiVec and Power8 have caused a fair number of wild writes on the stack and in the heap.
Testing on a 64-bit Intel Skylake show a marked improvement in performance. We suspect GCC is generating better code since it knows the alignment of the pointers, and does not have to emit fixup code for under-aligned and mis-aligned data. Testing on an mid-2000's 32-bit VIA C7-D with SSE2+SSSE3 showed no improvement, and no performance was lost.
This commit supports the upcoming AltiVec and Power8 processor support. The commit favors AlignedSecByteBlock over SecByteBlock in places where messages are handled on the AltiVec and Power8 processor data paths. The data paths include all block cipher modes of operation, and some filters like
Intel and ARM processors are tolerant of under-aligned buffers when using crypto intstructions. AltiVec and Power8 are less tolerant, and they simply ignore the three low-order bits to ensure an address is aligned. The AltiVec and Power8 have caused a fair number of wild writes on the stack and in the heap.
Testing on a 64-bit Intel Skylake show a marked improvement in performance. We suspect GCC is generating better code since it knows the alignment of the pointers, and does not have to emit fixup code for under-aligned and mis-aligned data. Here are some data points:
SecByteBlock
- Poly1305: 3.4 cpb
- Blake2s: 6.7 cpb
- Blake2b: 4.5 cpb
- SipHash-2-4: 3.1 cpb
- SipHash-4-8: 3.5 cpb
- ChaCha20: 7.4 cpb
- ChaCha12: 4.6 cpb
- ChaCha8: 3.5 cpb
AlignedSecByteBlock
- Poly1305: 2.9 cpb
- Blake2s: 5.5. cpb
- Blake2b: 3.9 cpb
- SipHash-2-4: 1.9 cpb
- SipHash-4-8: 3.3 cpb
- ChaCha20: 6.0 cpb
- ChaCha12: 4.0 cpb
- ChaCha8: 2.9 cpb
Testing on an mid-2000's 32-bit VIA C7-D with SSE2+SSSE3 showed no improvement, and no performance was lost.
Since removing the allocator overloards that handled the wipe mark, we have to route deallocate into the standard one. The standard one fires an assert for [now] normal operation
There was no aliasing violation in practice. We used a to assign the right pointer. If the compiler would have removed the unneeded assignment based on T_64bit, then we would not have been flagged.
Jeffrey Walton
Deadpikle