Patent application title: Security of Program Executables and Microprocessors Based on Compiler-Architecture Interaction
Saurabh Chheda (Santa Clara, CA, US)
Kristopher Carver (Chicopee, MA, US)
Raksit Ashok (Sunnyvale, CA, US)
IPC8 Class: AH04L900FI
Class name: Electrical computers and digital processing systems: support data processing protection using cryptography computer instruction/address encryption
Publication date: 2012-04-26
Patent application number: 20120102336
A method, for use in a processor context, wherein instructions in a
program executable are encoded with plural instruction set encodings. A
method wherein a control instruction encoded with an instruction set
encoding contains information about decoding of an instruction that is
encoded with another instruction set encoding scheme. A method wherein
instruction set encodings are randomly generated at compile time. A
processor framework wherein an instruction is decoded during execution
with the help of information provided by a previously decoded control
1. A method, for use in a processor context, wherein instructions in a
program executable are encoded with plural instruction set encodings.
2. The method of claim 1 wherein: a control instruction encoded with an instruction set encoding contains information about decoding of a second instruction; the second instruction is encoded with a different instruction set encoding from the control instruction.
3. The method of claim 2, wherein instruction set encodings are randomly generated at compile time.
4. The method of claim 2, further comprising of: adding at compile time to an executable information enabling decoding of an instruction; the information is added in an encrypted format.
5. The method of claim 2, wherein a constant field within an instruction is scrambled at compile time.
6. The method of claim 2, wherein instructions are scrambled with additional random keys associated with logical program addresses.
7. A processor framework wherein: instructions are encoded in plural instruction set architectures; an instruction is decoded during execution with the help of information provided by a previously decoded instruction.
8. The processor framework of method 2, wherein a control instruction is extracted at runtime before entering the processor pipeline.
9. The method of processor framework 7 wherein instructions are decoded to a reference instruction set before entering the decode stage of the processor.
10. The method of claim 2, wherein control instructions are implemented as co-processor instructions.
11. The method of claim 1, comprising of: analyzing a program executable at compile time; changing the instruction set architecture encoding of an instruction; scrambling constant fields in an instruction; adding information about the instruction set encoding of an instruction into another instruction; generating a program executable with plural instruction set encodings;
12. A method, for use in a processor context, wherein: instructions in a program executable are scrambled; a control instruction scrambled with a key contains information about decoding of an instruction that is scrambled with another key.
RELATED U.S. APPLICATION DATA
 This application claims the benefits of U.S. Provisional Application No 60/520,838, filed on Nov. 17, 2003, and Confirmation No 3408, entitled: IMPROVING SECURITY OF PROGRAM EXECUTABLES AND MICROPROCESSORS BASED ON COMPILER-ARCHITECTURE INTERACTION, the contents of which are hereby incorporated by reference into this application as if set forth herein in full.
 This invention relates generally to improving security in microprocessors based on compiler architecture interaction, without degrading performance or significantly increasing power and energy consumption. More particularly, it relates to protecting program executables against reengineering of their content and protecting processors against physical microprobing or accessing information about a program executed at runtime.
 Microprocessors (referred to herein simply as "processors") execute instructions during their operation. It is very important to improve security during execution and preferably energy/power consumption should not significantly be affected by the solution.
 Security is compromised when reengineering of program executables is possible. In addition, security of a solution such as an application intellectual property or algorithm is compromised when a processor can be physically microprobed and program information extracted without permission.
 Program intellectual property can be re-engineered from binaries without requiring access to source codes. As reported by the Business Software Alliance, software piracy cost the software industry 11 billion dollars in 1998. Furthermore, an increasing number of tamper-resistant secure systems used in both military and commercial domains, e.g., smartcards, pay-TV, mobile phones, satellite systems, and weapon systems, can be attacked more easily once application semantics, including critical security algorithms and protocols, are re-engineered.
 A key reason for many of these problems is that current microprocessors use a fixed Instruction Set Architecture (ISA) with predefined opcodes. Due to complexity, cost, and time-to-market issues related to developing proprietary microprocessors, most recent applications use commodity COTS components.
 In software compiled for such systems, individual instruction opcodes as well as operands in binaries can be easily disassembled given that ISA documentation is widely lo available. Conventional systems can also be microprobed and these instructions extracted from the instruction memory systems or instruction buses.
 Making the ISA reconfigurable (as a way to address these problems) is challenging, especially since a practically feasible solution would need to achieve this without significantly affecting chip area, performance and power consumption. Furthermore, it would be advantageous and practical if it could be retrofitted as add-on technology to commercially available microprocessor intellectual property (IP) cores from leading design houses such as ARM and MIPS. It should be backward compatible, e.g., capable of executing existing binaries to avoid losing existing software investments. Another critical aspect is migrating existing binaries to this new secure mode of execution without access to application source codes.
 For a given processor there are typically many compilers available provided by many different vendors. These compilers have their own advantages and disadvantages. Moreover, when compiling an application many pieces of codes are added from precompiled libraries or hand-written assembly language.
 Accordingly, if security-awareness is introduced at the executable-level, rather than at source code level, as in one embodiment of the present invention, by transforming the executable itself, significant practical advantages could be achieved. The goal would be to security-optimize executables that may have been fully optimized previously with a source-level compiler targeting a design aspect such as performance.
 The compiler-architecture interaction based processor framework described herein addresses the foregoing need to improve security in a practical and flexible manner.
 The approach provides security without adverse effects on chip area, performance and with scalability to different processor instructions sets and easy integration with other compilation tools.
 In one embodiment, the executable file itself provides a convenient interface between, for example, performance-oriented optimizations and security-oriented optimizations. Security optimizations in the compiler can be combined with, for example, energy oriented optimizations. Nevertheless, the invention is not limited to binary level compilation or compilation performed on executables; it can be also be added in source-level compilers without limitation.
 The new microprocessor framework proposed in this invention has the potential to solve the security problems raised above in a unified manner. It is based on a tightly integrated compiler-architecture framework.
 We call this microprocessor VISC, or (Virtual Instruction Set Computing), based on its property that instructions (and thus instruction opcodes) in application binaries are continuously reconfigured, changing meaning at a very fine compiler-managed program granularity. The actual reference ISA of the microprocessor is only visible during execution in the processor pipeline and used as a reference to generate control signals. Viewed in another way, there is no reference ISA just a reference control mechanism. VISC binaries can be made specific to a unique product and/or each binary can be made unique even if it is based on the same source code. This is achieved by having the initial reconfiguration-related secret, that is hidden in the processor, product specific.
 Based on this framework, security-aware microprocessors can be developed that will make software copy and tamper resistant. We would argue that a security-aware tightly integrated compiler-architecture approach such as VISC is ideal to protect software against various types of security attacks including physical microprobing of the memory system and buses as well as enables unique defense on non-invasive attacks such as differential power analysis.
 Supporting basic VISC security requires a special security-aware compiler or a security-aware binary morphing tool or similar source-level compiler, and small modifications to a microprocessor. Aspects related to security features provided, e.g., tamper and copy resistance, resistance against physical tampering with attacks such as fault injection and power analysis, and prevention of unauthorized execution are reviewed in the Description.
 In addition to conventional instructions, VISC requires either modifications to instructions or new so-called static control instructions to drive the security architecture.
 A key aspect of VISC is to communicate to execution engines how to decode/un-scramble instructions. Similarly, it could communicate how to execute instructions in their most energy efficient mode in addition to what operations to execute. Such static control instructions are continuously inserted in the instruction stream to provide the control information required to decode and control short sequences of instructions in the instruction stream. The control instructions can also be based on co-processor instructions not requiring modifications to a processor's regular ISA and thus easily integrated in existing processor cores.
 In one embodiment, a special static decode unit first un-scrambles and then decodes these control instructions and generates control signals.
 In one embodiment, to support a secure execution environment, the very initial few basic blocks or just the initial scrambling key are encrypted using public-key or symmetric cryptography. A VISC microprocessor contains the private key to decrypt these blocks or keys at runtime. Downloading the key can also be supported.
 The remaining part of the instruction stream is scrambled/reconfigured at compile time in the following manner: At the beginning of each basic block, or larger blocks such as super blocks, a special static control instruction is inserted. Note that the reconfiguration of instructions can be done by selecting fully random reconfiguration keys at each reconfiguration step. This is possible as a security static instruction, that is scrambled with a previous random key, contains the reconfiguration parameters required for decoding the next instruction(s).
 A security instruction encodes a scrambling scheme for the Block. This same static instruction can be leveraged to support compiler-managed power optimizations.
 During code generation, all the instructions in the referred basic block are scrambled using the same random scheme. At runtime, once a static control instruction is decoded, it (reconfigures) the un-scrambling hardware to the specified scheme. Examples of scrambling that can be supported with minimal circuit delay and area overhead include: flipping of selected bits, rotations, and combination of those. These transformations are not intended to be limiting: other schemes can be easily derived and controlled similarly.
 Once instructions are un-scrambled and decoded, the execution in the execution stage proceeds as normal in the similarly to the reference or original ISA.
 Each static instruction is either encrypted or scrambled based on a scrambling scheme defined in a previous static instruction in the control-flow graph. At runtime, a secure boot code that accesses the private key from an EEPROM type of device decrypts and executes the first block or so until a static instruction is reached. Initial parameters that are required, such as the length of the code that needs to be decrypted with the private key, or just the initial reconfiguration key, could be provided in the same encrypted way.
 Once a static control instruction is reached, subsequent static instructions take over control for the remainder of the execution. Note that this means that the performance impact of the private key based decryption is amortized across the whole duration of the program.
 If an entire program were encrypted using public-key or symmetric-key cryptography, it would make execution prohibitively slow and power/energy consumption would increase even if decryption were implemented in hardware.
 In contrast, the approach in this invention has very little area, performance, or power overhead. The initial decryption overhead is insignificant for most of the applications. In a software implementation using ECC this decryption takes 1.1 cycles.
 A static security control instruction can be folded from the pipeline or executed in parallel with any instruction from a previous basic block.
 In general, there might be several compilers available for one particular processor or an instruction set architecture. The approach based on executables can be easily integrated with some or all of these compilers without requiring changes in the source-level compiler.
 The invention can be used to improve security on any type of device that includes a processor. For example, the invention can be used on personal computers, devices containing embedded controllers, sensor networks, network appliances, and hand-held devices, cellular telephones, and/or emerging applications based on other device technologies.
 Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. Although methods and materials similar or equivalent to those described herein can be used in practice, suitable methods and materials are described below. In addition, the materials, methods, and examples are illustrative only and not intended to be limiting.
 Other features and advantages of the invention will become apparent from the following description, including the claims and drawings.
BRIEF DESCRIPTION OF THE DRAWINGS
 FIG. 1 is a block diagram showing the relationship between a source-level compiler, instruction set architecture, and microarchitecture in a processor.
 FIG. 2 is a block diagram showing the relationship between the executable-level re-compiler, conventional source-level compiler, instruction set architecture, and microarchitecture in a processor.
 FIG. 3 is a block diagram showing an embodiment of a VISC architecture.
 In the embodiments described herein, a processor framework uses compile-time information extracted from an executable or source-level and the executable is transformed so as to reduce vulnerability due to binary reengineering or extracting information at runtime in a processor. The security is improved without significantly affecting chip area, power or performance.
 Referring to FIG. 1, a compiler 10 is a software system that translates applications from high-level programming languages (e.g., C, C++, Java) into machine specific sequences of instructions. An instruction set architecture (ISA) 12 is a set of rules that defines the encoding of operations into machine specific instructions. The ISA acts as the interface between the compiler 10 and the microarchitecture (14). A computer program is a collection of machine level instructions that are executed to perform a desired functionality. Micro-architectural (or architectural) components 14 primarily comprise hardware and/or software techniques that are used during execution of the program. The actual machine can be a microprocessor or any other device that is capable of executing instructions that conform to the encoding defined in the ISA.
 Compile-time refers to the time during which the program is translated from a high-level programming language into a machine-specific stream of instructions, and it is not part of execution or runtime. Runtime is the time it takes to execute translated machine instructions on the machine. Compilation is typically performed on a different host machine than execution and it is done on the source program.
 In contrast, executable-level re-compilation is a process that uses an application's executable as input and performs analysis and modifications at the executable-level.
 Referring to FIG. 2, an executable-level re-compiler or executable-level compiler 24 is a software system that translates applications from executable-level into machine-specific sequences of instructions 26. The remaining components 26 and 28 are similar to the components 12 and 14 in a conventional compiler system as shown in FIG. 1.
 The present invention is applicable to both source-level and binary-level compilers and processors using binaries from either of these tools.
 In one embodiment, protection in a VISC architecture is provided by a combination of public-key encryption such as ECC or RSA, and a continuous compiler-enabled fine-granularity scrambling/reconfiguration of instructions.
 The private key is stored in a tamper-resistant device and used at runtime to decrypt the initial reconfiguration key or sequence before execution starts. The corresponding public part of the encryption key is used during compilation and it is made available in the software compilation tools so that anyone can generate binaries.
 In one aspect, only a header block or initial reconfiguration parameter is chosen to be encrypted with the public key. The decryption can be implemented with microcode, can be hardware supported, or can be implemented with trusted software.
 The rest of the code is scrambled at a very fine compiler managed program granularity. This has the effect of having multiple ISAs in the same application binary. Moreover, these ISAs can be generated fully randomly. Un-scrambling is managed by the inserted static security control instructions that identify un-scrambling rules to be used at runtime.
 The boundary between the encrypted part and the first scrambled block can be placed anywhere in the binary.
 Another embodiment is to have the private key encrypt the key of a faster symmetric key encryption such as DES, Triple DES, or AES. A disadvantage of such an approach is that one would also need the secret key for the symmetric cipher to generate code.
 In one embodiment, the mapping between the un-scrambling schemes supported at runtime and the encoding of static control instructions is reconfigurable with reconfiguration parameters that are stored encrypted and can be decrypted only with the private key.
 A reconfigurable un-scrambling logic and a special decode unit (called static decode) support un-scrambling of the specified instructions from each block, and decoding of the static control instructions.
 Programming is done by executing trusted instructions, e.g., that are part of the microcode, and is defined by the encrypted reconfiguration parameter. For example, 32-bit numbers can define new permutations of instruction bits that alter reference scrambling schemes encoded with static instructions.
 In one aspect, programming of this mapping is done at the beginning of execution after the decryption of the initial header block is completed and before any static instruction is executed. Scrambling of the binary at compile time is done at a very fine program granularity such as basic blocks or at larger granularity such as super blocks. The compiler must guarantee, when selecting these blocks, that there is no branching out and there are no branch instructions elsewhere that would branch in. This scrambling can be combined with additional transformations that are based on program addresses.
 The size of these blocks is variable, ranging from a small number of instructions to possibly tens or hundreds of instructions. The security-aware compiler decides what scrambling rule to apply to each block and inserts a control instruction at the top.
 In one aspect, after scrambling, instruction opcodes remain (mostly) valid, just their meanings are changed. This is because the majority of possible bit permutations in opcode fields are already used up in modern ISAs, so permutations of opcode bits will likely just create other valid instructions. The length of the blocks scrambled with one key/scheme is program and compilation dependent. The schemes used can be generated randomly at compile time.
 At runtime, as soon as the initial header block or reconfiguration key is decrypted, the approach goes into a phase overseen by the static control instructions. The execution of the application from that point on is driven by the control instructions that always preset the un-scrambling rule to be applied in a subsequent sequence of instructions or instruction. Execution otherwise happens as usual in the pipeline.
 In one aspect, a basic VISC approach can be combined with additional mechanisms, enabled by the VISC compiler-architecture framework, to protect against leaking out information from various parts of the microprocessor. A key benefit is that its protection boundary can be extended to incorporate most of the processor domains, like the memory system including caches and buses, that are normally the target of security attacks. Moreover, it is easy to combine the compiler-driven approach with runtime techniques for additional level of security.
 In one embodiment we have implemented, even the data that is in the execute stage before a memory operation is modified by randomly modifying the data at compile time and taking into account these modifications in the codes that consume the data. This would mean not only that the date is obfuscated in the memory system but even when passes the execute stage. This significantly increases the hurdles for differential power attacks and there will be no correlation between the data in the pipeline across several executions of the same code.
 The content that must be preserved tamper resistant is the device that contains the private key used to initialize. A VISC system protects against physical probing in key parts of the microprocessor including the memory system, buses, and in some cases hide the data in the pipeline making execution of a sequence fully dynamic and changing from one execution to another. In addition, deep sub-micron feature sizes in future process generations make reconstructing the circuitry of complex processor pipelines expensive and very time-consuming to perform. This is important as in many secure systems sensitive information has a strategic value to the opponent only if extracted within a certain time-frame.
Protection in Caches and External Memory
 In one embodiment of this invention the executable file is analyzed by a binary scanning tool in order to gain information about the various sections and any other symbolic information that may be available. This information is used in order to create a secure version of the program.
 Instructions in VISC systems are decrypted or un-scrambled solely during decoding. This implies that binaries are protected everywhere in the instruction memory including L1/L2 caches, buses, and external memory. While the approach in this embodiment does not stop someone from inserting instructions, the meaning of these instructions will be different from the one intended. Attacks based on physical tampering that require reading instruction caches/memory are strongly protected against.
 A solution to protect the data memory is not automatically provided by a reconfigurable instruction set approach. While in a VISC system instruction operands can also be in scrambled form, during execution L1 data caches and data memory will contain valid data. Although existing tampering schemes typically rely on having access to both instructions and data, without other architectural or compiler level protection valid data could potentially be accessed and may present a possible security weakness.
 In one embodiment of this invention, a solution is to encrypt each cache block and add a tag containing a hash calculated on the data values. A tag also protects against unauthorized modifications. When data values are modified, the stored tag will not match the calculated one and a security related exception can be raised. The tag can be put into a small cache if this aspect would need to be able to get most of the modification attempts (that would increase the hurdles for an attacker significantly) or into an indexed small SRAM if all memory stores would need to be protected. This depends on application.
 Adding this type of support on L1 data caches could however affect performance somewhat even with a pipelined hardware implementation.
 A lower overhead approach that can be provided in VISC in many application areas is based on security-aware compilation and minor microarchitecture support. The approach requires alias analysis to determine the possible location sets for each memory access at compile time. A location set contains the possible memory ranges a memory access can go to at runtime.
 Within disjoint groups of location sets, a scrambling rule similar to the one used for instructions would be defined and written into the register file. A small number of special registers could be provided if register pressure becomes an issue. One would need to make sure at compile time that memory accesses within each location set group are always generated with the same scrambling rule.
 In one embodiment, the approach could have only a single rule for data. During runtime, whenever data is written, it is written scrambled. During execution of a load instruction, before the data is loaded into the register file it is un-scrambled. Simple techniques such as based on bit flipping would add little overhead and could possibly be supported without increasing cache access time.
 One could also support multiple active rules within a basic block. In that case, the static control instruction would need to identify the special register that should be used during runtime to un-scramble the memory instruction. A simple solution to support two different rules would be to have them defined in a static control instruction. For example, a bit vector could tell which memory operations would be using which scheme in a basic block.
 As there are typically 1 to 3 memory operations in a basic block, three bits would be enough to encode it. More then two rules per block would require more bits per memory instruction. In each loop one could use two data scrambling rules, but may provide different scrambling between various loops if possible.
 A block-level overview of an embodiment of a VISC architecture 40 is shown in FIG. 3. As shown in the figure, software is strongly protected in (at boundary 42) caches and external memory and buses. Only physical tampering with the core processor circuitry (below boundary 42) and the device 44 that contains the private key may compromise VISC's protection of software. The figure shows one possible embodiment and is not intended to be limiting; different execution mechanisms and memory hierarchies are also possible and the protection boundary could be extended to other processor areas with VISC.
Further Protection Against Security Attacks
 Cryptographic research has traditionally evaluated cipher systems by modeling cryptographic algorithms as ideal mathematical objects. Conventional techniques such as differential and linear cryptanalysis are very useful for exploring weaknesses in algorithms represented as mathematical objects. These techniques, however, cannot address weaknesses that are due to a particular implementation that often require both software and hardware components. The realities of a physical implementation can be extremely difficult to control and often result in leakage of side-channel information. Successful attacks are frequently based on a combination of attacks that are used to leak information about a hidden secret, and use this side-channel information to reduce the search space and mount an exhaustive attack.
 There is a wide range of security attacks that can be mounted to reveal secrets in secure systems. Techniques developed have shown how surprisingly little side-channel information is required to break some common ciphers. Attacks have been proposed that use such information as timing measurements, power consumption, electromagnetic emissions and faulty hardware.
Protection Against Unauthorized Execution
 Only processors that have the right private key could start executing VISC binaries correctly. Several products that contain the microprocessor could contain the same key, but one could also provide support to an authorized party to upload a new key making the binary run only on one particular product. Additional techniques, such as based on tagging each cache line in the instruction memory (with a hash such as MD5), can be provided to protect against unauthorized modifications in the instruction memory. Nevertheless, VISC already provides a lazy protection against unauthorized execution even without additional tagging. The protection is lazy as modification is not immediately stopped but only typically after 1 to 3 instructions.
Power Analysis Attacks
 A Simple Power Attack (SPA) is based on directly monitoring the system's power consumption. Different attacks are possible depending on the capabilities of the attacker. In some situations the attacker may be allowed to run only a single encryption or decryption operation. Other attackers may have unlimited access to the hardware. The most powerful attackers not only have unlimited access, but also have detailed knowledge of the software.
 A Differential Power Attack (DPA) is more powerful than an SPA attack because the attacker does not need to know as many details about how the algorithm was implemented. The technique also gains strength by using statistical analysis. The objective of the DPA attacks usually is to determine the secret key used by a cryptographic algorithm.
 In both of these cases small variations in the power consumed by different instructions and variations in power consumed while executing data memory read/write instructions can reveal data values. One way to prevent a power analysis attack is to mask the side-channel information with random calculations that increase the measurement noise.
 The approach in an embodiment of this invention to reduce the vulnerability of DPA attack is based similarly on increasing the measurement noise. The approach is leveraging BlueRISC's compiler-architecture based power/energy optimization framework.
 In one aspect, a method optimizes power consumption by providing multiple access paths managed by the compiler.
 For example, in the case of the data cache, several mechanisms are provided such as fully statically managed, statically speculative, and various dynamic access modes. Energy reduction is achieved by eliminating redundant activity in the data cache by leveraging static information about data access patterns available at compile time.
 The compiler predicts for each memory access which is the most energy efficient and selects that path. The energy consumed per memory instruction is lowest for the statically managed modes and highest for the conventional modes. With provided architectural support one can detect if an access path is incorrectly chosen at compile time and the operation is rerun in a conventional way. The variation, in the preferred embodiment, in consumed energy per memory access can be as much as a factor of 10.
 The energy reduction approach alone is often enough to introduce enough power noise/variation given that various data memory access paths are already selected in a program and data dependent manner. As compared to introducing random calculations, such an approach to increase the measurement noise would have minimal performance impact.
 The compiler approach that is used for energy optimization can be used to further randomize the power cost for data accesses in critical portions of codes such as cryptographic algorithms. This is accomplished at compile time by deliberately selecting access paths with different power requirements.
 In addition, as described in earlier section, one could make sure at compile time that even the content of data passed into the execute stages before store operations is modified from execution to execution for the same program. This would make it much more difficult to correlate data before stores with differential binary analysis. Because the approach is compiler driven, the level of protection can be adjusted at compile time in an application specific manner without requiring changes to the hardware.
Fault Infection Attacks
 The cryptographic literature lists attacks that are based on injecting faults in secure systems. For example, in 1996, Biham and Shamir describe an attack on DES using 200 ciphertexts. They introduce one-bit errors produced by environmental stress such as low levels of ionizing radiation. Another type of attack is demonstrated based on glitches introduced to the clock or power supply. The idea is that by varying the timing and the duration of the glitch the processor can be forced to execute a number of wrong instructions. Fault injection can be used in conjunction with other attacks such as power analysis.
 The feasibility of such an attack in the context of VISC is questionable. It is more likely that such changes in the instruction stream would cause irrelevant and uninformative exceptions rather than leak side-channel information. This is due to the fact that VISC instructions can be protected efficiently against power analysis and that VISC instructions and opcodes are scrambled with different and reconfigurable keys, so such leaked information would be very difficult to interpret.
Spoofing, Splicing and Replay
 Splicing attacks involve duplicating or reordering segments of instructions. Replaying attacks are based on recording ciphertext (valid sequences of instructions) and executing them at later times. Spoofing attacks would typically involve changing data in the data memory. These attacks would be likely combined with some sort of physical tampering such as power analysis and fault injection. The objective is again to reveal some secret.
 These attacks could be eliminated by the tagging of the instruction memory and the encryption/tagging of the data memory mentioned earlier. But, even without tagging support, using such attacks in VISC is unlikely to be successful since VISC instructions are reconfigured at a fine compiler defined (and variable) granularity.
 In one aspect, in VISC it is possible to have each instruction scrambled differently. For example, one could have a combined scheme where the key used to un-scramble is shifted after every instruction in the block, effectively changing the meaning of the opcodes at the individual instruction granularity.
 A combination of various randomly-selected scrambling approaches and the variable length blocks make reconstructing valid sequences of instructions very hard. An exhaustive search would involve trying out possible block boundaries and thus exhaustively trying to re-engineer sequences of instructions. Permutations of bits in instructions will likely result in other legal instructions, further increasing the complexity of distinguishing real instruction sequences from just sets of valid instructions.
 Furthermore, as mentioned earlier, the meaning of the static instructions, i.e., the mapping between the scrambling schemes supported and the bits encoding them, is reconfigurable.
 To give an approximation of the difficulty of breaking this encoding, one would need to try 2 to the power of 32 permutations (for a 32-bit ISA) for each instruction and try to combine variable length sequences of such instructions into valid instruction sequences. Note that many valid solutions may exist. Even if an individual static control instruction were to be recovered, its actual meaning would still be unknown (as it is modified by the reconfiguration mechanism). Furthermore, it would be impossible to distinguish real static (or regular) instructions from permutations of other real instructions.
VISC Security-Aware ISA
 A key aspect is to scramble the interface between architecture and compilation. The mechanism to achieve this is based on inserting decoding related bits into the instruction stream and scrambling at compile time instruction blocks referred to by these bits. Our objective with this is to develop a microprocessor that effectively appears as a reconfigurable ISA machine. Note that the approach is more powerful than that, as not only the opcodes but even operand bits are reconfigured from their original meaning.
 In one aspect, one can encode the various scrambling schemes with one compact static instruction format. One possibility is to use a single static control information related instruction per basic block or super block that encodes the control information for the whole duration of the block at runtime. The added hardware component would be a different decode unit that we call static decode. This unit would be responsible for generating the control signals for the whole duration of the block. Overall, the idea to encode static information for the duration of a basic block is interesting because the order of execution is known.
 To illustrate this aspect with a simple example, imagine that there are two un-scrambling related bits in a control instruction. For example, if a control instruction contains two zero un-scrambling bits, the basic block instructions encoded with it could have two (or more) specific bit positions flipped. Un-scrambling can be done with very little overhead, by simply inverting those bits for all instructions (in the controlled block) at runtime.
 An advantage of this embodiment is that it does not require any modification to existing instructions so it would be easy to add to existing designs. Most of the designs have reserved opcodes for future ISA extensions or co-processors where static control instructions could be incorporated. With our compiler-enabled instruction cache optimization techniques, the effect of code dilution on power consumption is practically negligible as most of these added instructions will be fetched from a compiler-managed energy efficient cache.
 To extend the possible encoded combinations from 2 to the power of 23 (such as is available in an ARM ISA) one would need to periodically insert static region control instructions that would alter the meaning of a number of subsequent regular control instructions. This way the possible scrambling schemes supported can be extended to all 2 to the power of 32 permutations.
 In one aspect, individual blocks have different scramblings. The actual scheme can be randomized at compile time, by randomly selecting a scrambling approach for each block. Such a scheme would make any two instances of the binary for the same application to look different. Furthermore, the initial configuration parameter that defines the mapping between the bits in the control instructions and the supported schemes, can also be used to make the scrambling scheme different for various applications or chips.
 Static instructions at the root of control-flow subtrees define scrambling of static instructions in the subtrees. This can be supported easily at runtime; a particular un-scrambling scheme would be used at runtime until a new static instruction would override it and so on. As each static instruction operates on well defined boundaries, such as basic blocks, without possibility to enter that code from elsewhere or branch out, there is a well defined mapping between scrambling at compile time and un-scrambling at runtime that guarantees correct execution.
 The invention is not limited to the specific embodiments described herein. Other types of compiler analyses and/or architecture support may be used. The invention may be applied to control security aspects in any appropriate component of a processor. The optimizations and/or analyses may be performed in a different order and combined with other techniques, in other embodiments.
 Other embodiments not described herein are also within the scope of the following claims.
Patent applications by Kristopher Carver, Chicopee, MA US
Patent applications by Raksit Ashok, Sunnyvale, CA US
Patent applications by Saurabh Chheda, Santa Clara, CA US
Patent applications in class Computer instruction/address encryption
Patent applications in all subclasses Computer instruction/address encryption