review request: add intrinsics to use AES instructions
vladimir.kozlov at oracle.com
Mon Jul 16 10:39:31 PDT 2012
Thank you, Tom
I created next RFE and will sponsor changes. But I don't see 32 bit changes.
7184394: add intrinsics to use AES instructions
Deneau, Tom wrote:
> Please review the following webrev which adds intrinsic support to
> allow some of the com/sun/crypto/provider methods to use AES
> instructions when a processor supports such instructions.
> I do not have a bug number for this change but a description would be
> something like the following:
> Modern x86 processors have AES instructions to accelerate AES
> encryption and decryption but Hotspot does not have a way to
> generate such instructions. There is a way to hook in a native
> crypto library using PKCS11 and there are a few native libraries
> that support hardware AES instructions. However, these native
> PKCS11 libraries
> * do not scale well with multiple threads
> * are not supported on all platforms, for instance Hotspot does
> not have PKCS11 support on 64-bit Windows.
> * can be confusing to configure.
> Since this webrev adds intrinsic support for the default
> com/sun/crypto/provider classes, they are supported on all platforms
> and there is no additional configuration required. Measurements have
> shown that they scale very well will multiple threads.
> The rest of this mail describes the scope of the intrinsics and
> summarizes the source file changes.
> -- Tom Deneau
> Scope of the Intrinsics
> When creating a cipher the application specifies a "transformation"
> consisting of "algorithm/mode/padding". For more details see
> * These intrinsics kick in only when the algorithm part is "AES". A
> single block in AES is always 16 bytes and there are intrinsics
> for encrypting or decrypting a single block. These single-block
> intrinsics can work with any mode that uses AES and with any of
> the three AES key sizes (128, 192 or 256 bit).
> * A more optimized multi-block intrinsic can kick in if the
> algorithm/mode is "AES/CBC" (Cipher Block Chaining). Again all
> three AES key sizes are supported. There is no technical reason
> why we couldn't do multi-block intrinsics for the other modes
> (eg, ECB) but I want to get some feedback from the reviewers on
> the implementation before charging off on this path.
> * The padding part is handled by java routines outside of these
> Summary of Changes
> src/cpu/x86/vm/assembler_x86.cpp, hpp
> Defined the aes instructions which are used by the stub routines.
> Actual stub code for the aes intrinsics. As described earlier there
> are both single-block and multi-block intrinsic stubs.
> Note that the stubs make use of the "expanded key" which gets
> created each time the key changes. The expanded key is used by both
> the java code and the intrinsic AES instructions.
> The java code stores the "expanded key" in big-endian 32-bit
> integers. The x86 AES instructions require the expanded key to be
> in little-endian 128-bit words. Hence the pshufb instructions to
> get the key into the little-endian format
> src/cpu/x86/vm/vm_version_x86.cpp, hpp
> Detect and store the aes capability bit in cpuid. A global boolean
> command line flag UseAES can be used to turn off AES even if the
> hardware supports it.
> src/share/vm/opto/runtime.cpp, hpp
> The usual definitions of class names, method names and signatures
> for the java methods that are being intrinsified and the signatures
> for the stubs
> Up until now, every intrinsic was replacing a routine that was
> loaded by the "default" (NULL) class loader.
> com/sun/crypto/provider is not loaded by the default class
> loader so we had to add a check here.
> escape analysis knows about certain stubs, but if it sees a leaf
> stub it also checks against a predefined list. So the new intrinsic
> names were added to the list.
> The main logic for building up the calls to the stubs at compile
> time, assuming the platform has a stub and the global flags have
> not turned these intrinsics off.
> A new helper routine to load a field from an object was added since
> we ended up loading fields in a few places.
> For best performance, we wanted to hook into the multi-block
> encrypt and decrypt methods such as in CipherBlockChaining.java.
> This code is not AES-specific but handles CBC mode for any
> algorithm. (The algorithm part is handled by the enclosed
> "embeddedCipher" object).
> Thus at runtime we want to do the equivalent of an instanceof check
> on embeddedCipher and either call the stub (if it is AESCrypt) or
> call the original java code (if it is some other algorithm
> type). For the CipherBlockChaining.decrypt there is a further
> runtime check that the source and destination are not the same
> array which, because of the way CBC works would require cloning the
> source (cipher).
> Vladimir added some infrastructure to generate predicated
> intrinsics to solve the above problem. A particular intrinsic need
> only specify that it is predicated, and generate the particular
> guard node which if false will take the Java path. This
> infrastructure can be used for future intrinsics that have to make
> such a runtime choice. These changes from Vladimir are in
> callGenerator.cpp, doCall.cpp, and a small bit in library_call.cpp.
> global flags were added to
> * turn off either AES encryption or AES decryption intrinsics separately
> * turn off the multi-block CBC/AES intrinsics.
> By default all of the above are on. These are really there for
> testing, for example one could encrypt using Java and decrypt using
> the intrinsics.
> Also, a UseAES flag to ignore the hardware capability as described above.
More information about the hotspot-compiler-dev