This document is a specification snapshot built from github.com/riscv/riscv-cheri/commit/36d23ebec019869d644c040694ef7d6e433aa1fa and is not a versioned release. The latest versioned PDF release can be downloaded from github.com/riscv/riscv-cheri/releases.
This document is in the Stable state

Assume anything could still change, but limited change should be expected.

Copyright and license information

This specification is licensed under the Creative Commons Attribution 4.0 International License (CC-BY 4.0). The full license text is available at creativecommons.org/licenses/by/4.0/.

Copyright 2026 by RISC-V International.

Contributors

This RISC-V specification has been contributed to directly or indirectly by:

1. Introduction

This chapter is only included in the standalone CHERI spec and not part of the integrated document.

1.1. CHERI Concepts and Terminology

Current CPU architectures (including RISC-V) allow memory access solely by specifying and dereferencing a memory address stored as an integer value in a register or in memory. Any accidental or malicious action that modifies such an integer value can result in unrestricted access to the memory that it addresses. Unfortunately, this weak memory protection model has resulted in the majority of software security vulnerabilities present in software today.

CHERI enables software to efficiently implement fine-grained memory protection and scalable software compartmentalization by providing strong, efficient hardware mechanisms to support software execution and enable it to prevent and mitigate vulnerabilities.

Design goals include incremental adoption from current ISAs and software stacks, low performance overhead for memory protection, significant performance improvements for software compartmentalization, formal grounding, and programmer-friendly underpinnings. It has been designed to provide strong, non-probabilistic protection rather than depending on short random numbers or truncated cryptographic hashes that can be leaked and reinjected, or that could be brute-forced.

1.2. CHERI for RISC-V

This specification is based on publicly available documentation including CHERI v9 (Watson et al., 2023) and CHERI Concentrate (Woodruff et al., 2019).

RISC-V CHERI introduces several new bases that are informally collectively called RVY (e.g., RV32Y, RV64Y). Any CHERI RISC-V implementation must support one of these CHERI bases. CHERI will define profiles that provide equivalent functionality to the standard RVA, RVB, and RVM profiles. These RVY-based profiles (RVYA, RVYB, RVYM) will list the compatible extensions.

Most existing extensions are compatible with RVY with no modifications. Incompatible extensions will be listed in the profile documents.
Zicfiss is currently incompatible as the relevant push/pop instructions would need modifying for CHERI. An RVY version has not yet been developed.

This section lists new extensions available to RVY, and also extensions where the behavior is modified for RVY.

Almost all existing RISC-V extensions can be added to an implementation, but in most cases they will have some behavioral differences and/or new instructions operating on capabilities.

1.2.1. Stable Extensions and Specifications

Table 1. Unprivileged stable RVY base ISAs and extensions
Extension or Specification Description

RV64Y

CHERI Base ISA and capability encoding format for RV64

Zyhybrid

Hybrid extension for RVI compatibility

Zca (RVY added instructions)

New 16-bit encodings added to Zca

Zba (RVY added instructions)

New 32-bit encodings added to Zba

Zalrsc (RVY added instructions)

New 32-bit encodings added to Zalrsc

Zaamo (RVY added instructions)

New 32-bit encodings added to Zaamo
Table 2. Unprivileged stable extensions where RVY modifies the behavior
Extension or Specification Description

RVI (RVY modified behavior)

RVI instructions modified by RVY

Zca (RVY modified behavior)

C instructions modified by RVY

V (RVY modified behavior)

V instructions modified by RVY

Zicbom (RVY modified behavior)

Zicbom instructions modified by RVY

Zicbop (RVY modified behavior)

Zicbop instructions modified by RVY

Zicboz (RVY modified behavior)

Zicboz instructions modified by RVY

Zicsr (RVY modified behavior)

Zicsr instructions modified by RVY
Table 3. Unprivileged stable extension used by CHERI software
Extension or Specification Description

Zabhlrsc

Byte and halfword LR/SC functionality

CHERI defines new state and behavior for the privileged modes:

Table 4. Privileged stable extensions and specifications
Extension or Specification Description

Zyhybrid for Privileged Architectures

Hybrid extension for RVI compatibility

Supervisor-Level ISA for Virtual Memory (RV64Y)

Virtual Memory

Sdtrig (RVY)

Debug triggers

Svyrg

MMU-based capability data flow control and measurement
Table 5. Debug stable extensions and specifications
Extension or Specification Description

Sdext (RVY)

External debug support

Chapters for the unprivileged specification

2. RVY Base Capability Instruction Set, Version 1.0

This chapter introduces a new set of base ISAs called RVY that extend the RV32[IE]/RV64[IE] base ISAs with CHERI.

The RVY base ISA can be composed with other standard options to bases such as Zfinx.

2.1. CHERI Overview

CHERI enhances the base ISA by adding hardware memory access control. It has an additional memory access mechanism that protects references to code and data (pointers), rather than the location of code and data (integer addresses). This mechanism is implemented by providing a new primitive, called a capability, that software components can use to implement strongly protected pointers within an address space. Capabilities are unforgeable and delegatable tokens of authority that grant software the ability to perform a specific set of operations. In CHERI, integer-based pointers are replaced by capabilities to provide memory access control.

Every RVY base ISA requires a capability encoding format to complete the specification, such as RV64LYA. The capability encoding format gives full details of how CHERI capabilities are represented in memory and in registers, and specifies the exact properties such as bounds precision and available combinations of permissions.

The full specification of an RVY base ISA includes a capability encoding format.

E.g., in RV64LYA :

  1. XLEN=64

  2. L means that it’s a parameterized base

  3. Y means that it’s a CHERI base

  4. A is the specific capability encoding format.

Future RV64 CHERI capability encoding format could be RV64LYB, RVY64LYC etc. or they may have more descriptive names after the RV64LY prefix.

Also see capability encoding format.

2.2. CHERI protection model

The CHERI model is motivated by the principle of least privilege, which argues that greater security can be obtained by minimizing the privileges accessible to running software. A second guiding principle is the principle of intentional use, which argues that, where many privileges are available to a piece of software, the privilege to use should be explicitly named rather than implicitly selected. While CHERI does not prevent the expression of vulnerable software designs, it provides strong vulnerability mitigation: attackers have a more limited vocabulary for attacks, and should a vulnerability be successfully exploited, they gain fewer rights, and have reduced access to further attack surfaces.

Protection properties for capabilities include the ISA ensuring that capabilities are always derived via valid manipulations of other capabilities (provenance), that corrupted1 in-memory capabilities cannot be dereferenced (integrity), and that rights associated with capabilities shall only ever be equal to or less permissive (monotonicity). Tampering or modifying capabilities in an attempt to elevate their rights will yield an invalid capability. Attempting to dereference via an invalid capability will result in raising an exception.

1 Not all possible corrupted states are detected, see Section 2.5, “Integrity of Capabilities”.

CHERI capabilities may be held in registers or in memories, and are loaded, stored, and dereferenced using CHERI-aware instructions that expect capability operands rather than integer addresses. On system initialization, initial capabilities are made available to software by the execution environment via general purpose registers. All other capabilities will be derived from these initial valid capabilities through valid capability transformations.

Developers can use CHERI to build fine-grained spatial and temporal memory protection into their system software and applications and significantly improve their security.

2.3. Capability Registers and Format

RVY widens all registers that can hold addresses from XLEN to 2*XLEN bits (hereafter referred to as YLEN), adding metadata to protect its integrity, limit how it is manipulated, and control its use. In addition to widening to YLEN, each register also gains a one-bit capability tag which is defined below.

RVY specifies the minimum required fields that the capability encoding format must support, and their semantics. All current RVY bases are little endian and therefore the memory representation of capabilities always holds the address in the first XLEN bytes of memory followed by the metadata.

RVY is designed to be highly extensible by allowing different capability encoding formats to exist. Specific capability encoding formats may support different ISA extensions and add new architectural permissions.

Diagram
Figure 1. CHERI Capability structure

2.3.1. Address

The lower XLEN bits of a capability encode the address of where the capability points. This is also referred to as the integer part of the capability. For registers that are extended but currently hold non-capability data, all other fields are typically zero.

2.3.2. Capability Tag

The capability tag is an additional bit added to addressable memory and all YLEN-bit registers. It is stored separately and may be referred to as out of band or hidden, and is hardware managed. It indicates whether a YLEN-bit register or YLEN-aligned memory location contains a valid capability. If the capability tag is set, the capability is valid and can be dereferenced (contingent on checks such as permissions or bounds).

All registers or memory locations able to hold a valid capability are YLEN bits wide with an additional hidden capability tag bit. These are referred to as being YLEN-bit in this specification.

Not all memory in a real system will support capability tags. The EEI will define what portions of the address space are able to preserve capability tags (some addresses might support storing data only).

The capability tag cannot be directly set by software; it is not a conventionally accessible bit of state. If the capability tag is set then it shows that the capability has been derived correctly according to the principles listed above (provenance, integrity, monotonicity). If the rules are followed then the capability tag will propagate through the instructions that modify, load or store the capability.

Therefore, whenever an instruction writes a capability with the capability tag set to a register or to memory:

  • At least one capability tag must have been set in the input operands.

    • This is the provenance check.

  • The requested operation must have been legal and not increase bounds or permissions.

    • This is the monotonicity check.

  • The instruction may have checked that none of the input capability were corrupted.

    • This is the integrity check. Not all instructions check all input operands for integrity.

Capability load/store require the provenance check:

  • Any store that wrote the capability to memory was correctly authorized.

  • Any load that read the capability from memory was correctly authorized.

When an operation fails a check, either due to software error or malicious intent, then the operation raises an exception or sets the resulting capability tag to zero.

For a capability to be valid, the capability tag must be set, otherwise a capability is invalid.

Using an invalid capability to dereference memory or authorize any operation raises an exception. All capabilities derived from invalid capabilities are themselves invalid, i.e., their capability tags are zero.

Writing non-capability data into a register or memory location causes the capability tag to be set to 0. When the capability tag associated with the memory location is zero, the location contains non-capability data.

2.3.3. Capability tags in registers

Every YLEN-bit register has a one-bit capability tag, indicating whether the capability in the register is valid to be dereferenced. This capability tag is set to 0 whenever an invalid capability operation is performed. Examples of such invalid operations include writing only the integer portion (the address field) of the register or attempting to increase bounds or permissions.

2.3.4. Capability tags in memory

Capability tags are tracked through the memory subsystem: every aligned YLEN-bit wide region has a non-addressable one-bit capability tag, which the hardware manages atomically with the data. The capability tag is set to zero if any byte in the YLEN/8 aligned memory region is ever written using an operation other than a store of a capability operand. For example, the SY instruction may set the capability tag to one if all the instruction’s prerequisites are met, but SW will always set the capability tag to zero.

All system memory and caches that store capabilities must preserve this abstraction, handling the capability tags atomically with the data.

2.3.5. Capability Bounds

Capabilities encode memory bounds, i.e., the lowest and highest byte in memory that it is permitted to access when dereferenced for data memory access, or for instruction execution.

Checking is on a byte-by-byte basis, so that it is possible for a memory access to be fully in-bounds, partially out-of-bounds or fully out-of-bounds.

It is not permitted to make any partially or fully out-of-bounds memory accesses.

Every capability has two memory address bounds: base representing the lowest accessible byte, and top representing one byte above the highest accessible byte.

  • The base is XLEN bits and is inclusive.

  • The top is (XLEN+1)-bits and is exclusive.

    Inclusive top, with XLEN bits, was considered but rejected in favor of the exclusive top (with an additional bit to allow covering the entire address space). Using exclusive top allows representing a zero-length capability starting at address zero, which is not possible with inclusive top unless you introduce a special value for -1.

  • The length is (XLEN+1)-bits and is defined to be top - base.

Therefore a memory location A in the range base ≤ A < top is within bounds, and so valid to access.

Checking every byte of every executed instruction and every byte of every data memory access is fundamental to the memory safety which CHERI provides. In a typical load/store unit, the expansion of the bounds from rs1 and bounds checking is in parallel with the address calculation, the memory translation and/or the PMA/PMP checking.

A compressed format is used to encode the bounds with a scheme similar to floating-point using an exponent and a mantissa. Therefore small exponents can allow byte granularity on the bounds, but larger exponents give coarser granularity. One bounds encoding format based upon (Woodruff et al., 2019) is defined in Section 3.1.6.

Future CHERI bases may use encoding formats with different bounds encoding schemes (see capability encoding format).

Software can query the bounds of a capability held in a general-purpose register using the following instructions:

  • The base is returned by the YBASER instruction.

  • The length is returned by the YLENR instruction.

  • The top is returned by the YTOPR instruction.

2.3.6. Deriving New Bounds

On system initialization, one or more Root capabilities are available. All capabilities with smaller bounds are derived from these.

The bounds are reduced using the YBNDSW, YBNDSWI and YBNDSRW instructions.

They all copy rs1 to the output rd, set the base bound rd as rs1.address, subject to encoding granularity constraints, and set the length of rd from rs2 or the immediate field depending on the instruction.

  • YBNDSW sets the base to rs1.address, and the length to rs2.

    • Set the capability tag to zero if the bounds cannot be encoded exactly.

  • YBNDSWI sets the base to rs1.address, and the length to the immediate value.

    • Set the capability tag to zero if the bounds cannot be encoded exactly.

  • YBNDSRW sets the base to rs1.address, and the length to rs2.

    • The bounds may be rounded up if they cannot be encoded exactly but must be no larger than the initial bounds.

  • YAMASK can be used to calculate the nearest precisely encodable length and base values for a given size.

The bounds are encoded relative to the address field, sharing some upper bits of the address. The number of shared bits depends on the exponent, see Section 3.1.6.

2.3.7. Representability and Updating the Address

Software sometimes uses out-of-bounds pointers, for example, to represent a one-past-the end pointer to an array in C (arr + size). To support this pattern, CHERI must allow these out-of-bounds pointers to be held in capabilities while still protecting the bounds and integrity. Because the CHERI Concentrate (Woodruff et al., 2019) encoding scheme for memory bounds shares the upper bits of the address with the bounds, only a limited range of out-of-bounds pointers can be represented for any given set of bounds. This range is known as the representable range. The relationship between the bounds and the representable range is illustrated in Figure 2.

While the C standard only requires one-past-the-end pointers to be valid, real-world software has been observed to (temporarily) create pointers that point multiple bytes past the end or even before the start. To maximize software compatibility, the bounds representation was designed to allow for out-of-bounds pointers.
cap bounds map
Figure 2. Memory address bounds and representable range encoded within a capability

E, MW and R in the figure are all introduced in Section 3.1.6.2 along with the bounds decoding.

The maximum range of address values that the pointer can take without changing the calculation of the bounds is defined by the representable range.

The historic, uncompressed 64-bit CHERI-MIPS CPU had a 256-bit capability encoding had separate 64-bit values for the base and top memory bounds, and another for metadata fields. This scheme could represent all out-of-bounds pointers with a very high hardware cost.

Since the upper bits of the address are used to calculate the top and base bounds, deriving a new capability with a different address could change the resulting bounds. All derived capabilities must check that the new address does not change the bounds. If they do change then the capability tag of the derived capability is set to zero, so that it is invalid for use.

Software can derive a capability with a new address using instructions such as YADDRW, YADD and YADDI.

YADDRW writes back a derived capability with a new address field and, if the capability tag was previously set, sets the capability tag of the derived capability to one if the resulting capability still has the same bounds.
RVY implementations that use a different encoding scheme to RV64LYA (e.g., for accelerators or specific micro-controllers) may specify an alternative to the representable range check, and may never allow the address to be out of bounds. Therefore, the ISA specification uses the phrase represented exactly for this check.

2.3.8. Memory space

A hart supporting RVY has a single byte-addressable address space of 2XLEN bytes for all memory accesses. Each memory region capable of holding a capability also stores a capability tag bit for each naturally aligned YLEN bits (e.g., 16 bytes in RV64), so that capabilities with their capability tag set can only be stored in naturally aligned addresses. Capability tags must be atomically bound to the data they protect.

The memory address space is circular, so the byte at address 2XLEN - 1 is adjacent to the byte at address zero. A capability’s Representable Range described in Section 3.1 is also circular, so address 0 is within the Representable Range of a capability where address 2XLEN - 1 is within the bounds.

However, the decoded top address of a capability is XLEN + 1 bits wide and does not wrap, so a capability with base 2XLEN - 1 and top 2XLEN + 1 is not a subset of the infinite capability and does not authorize access to the byte at address 0.

Like malformed bounds (see Section 3.1.6.4), it is impossible for a CHERI core to generate a valid capability with top > 2XLEN.

If such a capability exists then it must have been caused by a logic or memory fault. Unlike malformed bounds, the top overflowing is not treated as a special case in the architecture: normal bounds check rules should be followed.

2.3.9. Capability Type (CT)

This metadata value indicates the type of the capability and determines which operations the capability authorizes. For example, the CT-field can be used to seal capabilities against modification, to provide control flow integrity, and to provide immutable software tokens.

Unsealed capabilities

When CT=0, the capability authorizes access to a region of memory as defined by the permissions and bounds.

Sealed capabilities

Capabilities with CT≠0 are sealed against modification and cannot be dereferenced to access memory. Instructions that operate on capabilities will produce a result with the capability tag set to zero if the source capability is sealed and the operation would alter its address, bounds, or permissions.

Sealing

Given a capability with CT=0, deriving a capability with CT≠0 is termed sealing . This is achieved using YSEAL.

Unsealing

Given a capability with CT≠0, deriving a capability with CT=0 is termed unsealing . This is achieved using YSUNSEAL .

Sealed Entry Point Capability

Sealed entry point capabilities provide immutable function pointers within a CHERI software system.

Such function pointers are sometimes known as "sentries", a contraction of "sealed entries".

They can establish a form of control-flow integrity between mutually distrusting code. Because they are immutable, the jump target cannot be modified, ensuring that control flow can only enter at the specific address encoded in the capability and not elsewhere in a function.

A sealed entry point capability may be passed as the rs1 argument to JALR (RVY), and is unsealed on dereference. The returned rd value is also written back as a sealed entry point capability.

In addition to using sealed capabilities for sealed entry points, sealed capabilities can also be useful to software as secure software tokens. YSUNSEAL can be used to convert such a token back to an unsealed capability.

The standard capability type mapping is shown in Table 6, extensions and capability encoding formats may add more types.

Table 6. Standard capability type mapping
Type Name Integer Value Description

Unsealed

0

Unsealed capability granting access to memory

Sealed entry point capability

1

Sealed capability that can be used as a function entry point or function return.
Future extensions may declare new CT-field values which place restrictions on their usage (e.g. forward/backward-only control flow checks with JAL (RVY)/JALR (RVY) for improved CFI).

2.3.10. Architectural Permissions (AP)

This metadata field encodes architecturally defined permissions of the capability. Permissions grant access subject to the capability tag being set, the capability being unsealed, and bounds checks passing. Any operation is also contingent on requirements imposed by other RISC-V architectural features, such as virtual memory, PMP and PMAs, even if the capability grants sufficient permissions. The permissions defined in RVY are listed in Table 7. All capability formats must support at least this set of permissions. Extensions building on top of RVY may define additional permissions.

An example of such an extension is Zylevels1.

Permissions can be cleared when deriving a new capability value (using YPERMC) but they can never be added.

Table 7. AP-field summary
Permission Type Comment

R-permission

Data memory permission

Authorize data memory read access

W-permission

Data memory permission

Authorize data memory write access

X-permission

Instruction memory permission

Authorize instruction memory execute access

C-permission

Data memory permission

Authorize loading/storing of capability tags

LM-permission

Data memory permission

Used to restrict the permissions of loaded capabilities.

ASR-permission

Privileged state permission

Authorize privileged instructions and CSR accesses.
Read Permission (R)

Allow reading data from memory.

Write Permission (W)

Allow writing data to memory.

Execute Permission (X)

Allow instruction execution.

Capability Permission (C)

Allow reading capability tags from memory if R-permission is also granted.

Allow writing capability tags to memory if W-permission is also granted.

If C-permission is missing then the capability tags for capability loads and stores are read and written as zero.

Load Mutable Permission (LM)

Allow preserving the W-permission of capabilities loaded from memory. If a capability grants R-permission and C-permission, but no LM-permission, then a capability loaded via this authorizing capability will have W-permission and LM-permission removed.

The permission stripping behavior only applies to loaded capabilities with their capability tag set and that are not sealed after all other checks. This ensures that capability loads of non-capability data do not modify the loaded value, and that sealed capabilities are not modified.

Clearing a capability’s LM-permission and W-permission allows sharing a read-only version of a data structure without making a copy. For example, if software removes LM-permission from a capability to a linked list, any pointers loaded through that capability will automatically lose their W-permission, preventing modification of the list content.
Access System Registers Permission (ASR, primarily used to authorize CSR accesses)

Allow read and write access to all privileged CSRs, some unprivileged CSRs and some privileged instructions. ASR-permission permission checks always use the permission in the pc.

In the unprivileged RVY ISA, the following are affected by ASR-permission:

In the privileged RVY ISA, the following are affected by ASR-permission:

  • Instructions: MRET (RVY), SRET (RVY) require ASR-permission in the pc .

  • CSRs: The privileged stidc, vstidc, and mtidc CSRs require ASR-permission for writing, but not for reading (if access is permitted in the current privilege mode). All other privileged CSRs require ASR-permission in the pc for any access. Access to debug CSRs requires ASR-permission.

This permission is important in privileged execution environments. Removing this permission allows constraining privileged software to a sandbox that cannot be subverted by changing privileged state.
Extensions may add additional non-privileged CSRs that require ASR-permission.
2.3.10.1. Permission Transitions

Not all capability permissions are orthogonal as some permissions inherently depend on others. As such, using YPERMC to clear some permissions may have the effect of clearing others as well.

For the base set of permissions just defined, the rules in Table 8 apply.

Table 8. YPERMC base rules
YPERMC Rule Permission Valid only if

base-1

C-permission

R-permission or W-permission

base-2

LM-permission

C-permission and R-permission

base-3

ASR-permission

X-permission
These rules enforce sensible permission combinations.

Extensions that define new permission bits may also introduce new dependency constraints. Currently defined examples are:

  1. Zyhybrid, see Section 13.1.3

  2. Zylevels1, see Section 12.3.1

Capability encoding formats may impose additional constraints to reduce the number of bits necessary to represent permissions.

2.3.11. Software-Defined Permissions (SDP)

The metadata also contains an encoding-dependent number of software-defined permission (SDP) bits. They can be inspected by the kernel or application programs to enforce restrictions on API calls (e.g., permit/deny system calls, memory allocation, etc.). They can be cleared by YPERMC but are not interpreted by the CPU otherwise.

While these bits are not used by the hardware as architectural permissions, modification follows the same rules: SDP bits can only be cleared and never set on valid capabilities.

This property is required to ensure restricted programs cannot forge capabilities that would pass the software-enforced checks.
Software is completely free to define the usage of these bits.

2.3.12. Special Capabilities

Root Capabilities

Root capabilities are those provided by the system on initialization. In some capability encoding formats there is a single Infinite capability value, which grants all permissions and has bounds covering the whole 2XLEN address space; in such systems, root capabilities are often Infinite.

Other capability encoding formats may have a set of root capabilities distinguishing between data and executable capabilities as shown below.

At system initialization the set of all root capabilities must be made available to software.

Implementations are not required to provide access to the entire address space to a single hart, the root capabilities may be restricted to only cover a subset of the entire address space, or to remove certain permissions.
Root Executable Capability

An unsealed capability that has bounds covering all addresses and grants at least all of X-permission, R-permission, C-permission, LM-permission, and ASR-permission. Extensions introducing new permissions may require these to be provided by root executable capabilities.

Root Data Capability

An unsealed capability that has bounds covering all addresses and grants at least all of R-permission, W-permission, C-permission, and LM-permission. Extensions introducing new permissions may require these to be provided by root data capabilities.

NULL Capability

A capability with all-zero metadata, a zero capability tag, and an address of zero is referred to as the NULL capability. This capability grants no permissions and any dereference results in raising an exception.

2.4. Capability encoding formats

CHERI implementations make trade-offs in their "capability encoding formats", such as the precision of bounds and the expressible set of combinations of permissions, that impact certain behaviors of the ISA. The definition of all CHERI base ISAs includes sufficient details of a capability encoding format to precisely define the execution behavior of all base ISA instructions.

Whilst RV64Y has enough bits in the encoding format to prevent such trade-offs in general, the concepts are introduced in here and RV32Y formats will make extensive use of them.

The name RVY refers generically to CHERI base ISAs, which share common semantics but may have different metadata encodings. The RVY name on its own does not describe the complete base ISAs, but defines all the semantics of a RVY machine, where certain behavioral parameters must be filled in by the capability encoding. For example, the YBNDSRW instruction yields a capability with new bounds that are rounded differently depending on the number of bits allocated to the bounds in the capability encoding format. Systems with XLEN=64 can allocate more bits towards the mantissa of the bounds than XLEN=32 system and therefore have a larger Representable Range.

These observable differences in behavior are similar to floating point numbers where the supported operations and their semantics are the same for all formats, but the exact result depends on the representation (e.g., 16/32/64-bit IEEE-754 or other floating point formats).

It is possible to generate code for RVY that is compatible with all valid capability encoding formats. However, to allow compilers and/or software library authors to make assumptions about encoding-dependent properties such as bounds precisions, each capability encoding format comes with a set of standard parameters.

Each capability encoding format also declares a new base ISA.

The documentation for each capability encoding format specifies at least the parameters in Table 9 below where the default is not used.

The default settings all match RV64LYA.
Table 9. RVY capability encoding format parameters
Parameter Legal values Default Value Comment

Parameters used for the bounds encoding (see Section 3.1.6)

MW

integer values

14

Mantissa width for the bounds encoding

EW

integer values

6

Exponent width for the bounds encoding

CAP_MAX_E

integer values

52

Maximum exponent value

enableL8

0/1

0

Whether the encoding format includes the L8 bit

Parameters used for permissions encoding in the AP-field

AP_MAX

integer values

all ones

Value of the AP-field giving maximum permissions

AP_ENC

ps<suffix>/ pc<suffix>

ps1

Encoding scheme for the AP-field

Other parameters

AUIPC_SHIFT

integer values up to 12

12

Shift distance of the immediate in AUIPC

REP_REGION

rc1/r0

rc1

Whether out of bounds addresses are representable

Two parameters used non-numeric values:

AP_ENC

Whether permissions in the AP-field are encoded using a simple representation of one bit per architectural permission (ps) or a compressed encoding that leverages dependencies between permissions and does not allow encoding all possible combinations (pc). Each scheme has a suffix to show which exact version is in use, as ps schemes may not always store permissions in the same order and pc schemes may use different forms of compression.

REP_RANGE

Whether the capability encoding formats guarantees representability of out-of-bounds capability values, rc1, where the encoding uses one additional bit and uses a centered out-of-bounds region, or r0 where zero bits are used for out-of-bounds capabilities and there is no guaranteed out of bounds representable range.

For RV64Y, the number of spare bits in the capability encoding ensures sufficient future extensibility and therefore few custom formats are expected to exist. RV32Y, however, has insufficient bits available to enable features for all target domains, so multiple RV32Y-extending capability formats will exist.

2.5. Integrity of Capabilities

CHERI defines the following integrity check rules:

  1. The bounds are not malformed.

  2. The capability metadata does not have any bits, or fields set to a reserved value.

An example of a reserved field value is an illegal set of permissions, such as ASR-permission without X-permission.

If any of the above rules do not hold for a capability with the capability tag set, it indicates either:

  • State corruption due to memory or logic faults, or

  • The presence of incompatible or faulty CHERI IP within the system.

These checks are much less rigorous than parity or ECC protection, and are only used to detect simple problems with the capability metadata.

Performing these checks is optional except for instructions that are defined to accept arbitrary input bit patterns and can write an output with the capability tag set. All instructions that require integrity checks have this noted in the instruction description.

Currently YBLD is the only defined instruction that sets the capability tag on an arbitrary bit pattern and therefore must perform an integrity check on the input capability in rs2. For all other instructions, performing these checks is optional unless the hart implements additional reliability features (since malformed capabilities with the capability tag set cannot exist under normal operating conditions).
Even though valid capabilities which fail the integrity check could not have been legally generated by the local hart, defining the handling in the architecture allows the behavior to be precisely specified for all 2(YLEN+1) input values.

2.6. Extended State

As stated above, all state which can hold addresses are extended from XLEN to YLEN bits.

A reminder that all YLEN state also has a hardware managed capability tag.

2.6.1. General Purpose Registers

The XLEN-wide integer registers (e.g., x1, x2) are all widened to YLEN bits, each gaining an associated capability tag, as shown in Figure 3.

Diagram
Figure 3. Extended registers in RVY

The zero register (x0) is extended with zero metadata and a zero capability tag: this is called the NULL capability.

2.6.2. The Program Counter Capability (pc)

The pc register is widened to hold a capability. Widening the pc allows the range of branches, jumps and linear execution for currently executing code to be restricted. The pc address field is the pc in the base RISC-V ISA so that the hardware automatically updates it as instructions are executed.

The hardware performs the following checks on pc for each instruction executed in addition to the checks already required by the base RISC-V ISA. A failing check raises a CHERI exception.

  • The capability tag must be set to one.

  • The capability must not be sealed.

  • The capability must grant X-permission.

  • All bytes of the instruction must be in bounds.

  • All integrity checks must have passed.

On system initialization the pc bounds and permissions must be set such that the program can run successfully (e.g., by setting it to a Root Executable capability to ensure all instructions are in bounds).

Future ISA extensions should respect these rules so that the checked bits do not need to be stored in all copies of the pc in the implementation.
Diagram
Figure 4. Program Counter Capability

2.6.3. Added CSRs

RVY adds the YLEN-bit CSR shown in Table 10.

Table 10. Capability CSRs added in RVY
YLEN CSR Permissions Description

utidc

RW, ASR-permission required for writes, not reads

User Thread ID Capability
2.6.3.1. User Thread Identifier Capability (utidc)

The utidc register is an unprivileged CSR used to identify the current software thread. Any operation that modifies utidc raises an exception unless the ASR-permission is set in the current pc.

Diagram
Figure 5. User thread identifier capability register

Each thread (as scheduled by the kernel) can have its own context stored in privileged data structures managed by the compartmentalization TCB, for example a stack of user-level compartment entries and exits. utidc allows compartment-switching code to determine the currently running thread without a call into the operating system. This is read-only without ASR-permission to prevent a malicious compartment from tricking the switching mechanism into accessing data corresponding to the wrong thread. While the RISC-V ABI includes a thread pointer (tp) register, it is not usable for the purpose of reliably identifying the current software thread because the tp register is a general purpose register and can be changed arbitrarily by untrusted code. Extending utidc to a capability allows a data structure to be shared (guarded by e.g. the subset sealing mechanism), which allows for example efficient recovery of a trusted stack, without requiring additional indirection.

2.6.4. CSR Classes

All CSRs that can hold addresses are widened to YLEN bits.

RVY has three classes of CSRs:

XLEN-bit CSRs

These do not contain pointers (e.g., fcsr from the "F" extension).

Extended CSRs

These are XLEN-bit CSRs widened to YLEN bits, which are able to contain pointers (e.g., mtvec from the privileged specification).

YLEN-bit CSRs

These are added by RVY and contain pointers (e.g., utidc).

When accessing CSRs these rules are followed:

  1. Accesses to XLEN-bit CSRs are as specified by Zicsr

  2. Accesses to YLEN-bit CSRs and extended CSRs, using CSRRW will:

    1. Read YLEN bits

    2. Write YLEN bits, and will write the capability tag to zero if any integrity check fails

  3. Accesses to YLEN-bit CSRs and extended CSRs, using instructions other than CSRRW will:

    1. Read YLEN bits

    2. Write an XLEN-bit value to the address field, and use the semantics of the YADDRW instruction to determine the final written value

Any YLEN-bit or extended CSR may have additional rules defined to determine the final written value of the metadata and/or to write zero to the capability tag.

The assembler pseudoinstruction to read a capability CSR csrr rd, csr, is encoded as csrrs rd, csr, x0.

Table 11. YLEN-bit CSR and Extended CSR access summary for RVY
Instruction Read Width Write Width

CSRRW rd==x0

YLEN

CSRRW rd!=x0

YLEN

YLEN

CSRR[C|S] rs1==x0

YLEN

CSRR[C|S] rs1!=x0

YLEN

XLEN

CSRRWI rd==x0

XLEN

CSRRWI rd!=x0

YLEN

XLEN

CSRR[C|S]I uimm==x0

YLEN

CSRR[C|S]I uimm!=x0

YLEN

XLEN

In Table 11, when there is no read or write width shown, the CSR access is not made and there are no side-effects following standard Zicsr rules.

2.7. Capability checks

Every memory access must be authorized by a valid capability.

Instruction fetches and data memory accesses raise an exception if the access is out of bounds, or if the authorizing capability is missing the required permissions:

  • All load instructions require R-permission and are authorized by the capability in rs1

  • All store instructions require W-permission and are authorized by the capability in rs1

  • All indirect jumps require X-permission on the target capability

Instruction fetch and branch execution are authorized by the program counter capability which extends the PC.

This allows code fetch to be bounded, preventing a wide range of attacks that subvert control flow with non-capability data.

E.g., lw t0, 16(sp) loads a word from memory, getting the address, bounds, and permissions from the sp (capability stack pointer) register.

2.8. Added Instructions

RVY adds new instructions to operate on capabilities, and redefines part of the RVI/RVE custom encoding space for this purpose.

  1. custom-3 is renamed to RVY-A and is used for standard RVY encodings.

  2. custom-2 is renamed to RVY-B and is used for standard RVY encodings.

  3. custom-1 is renamed to RVY-C and is used for standard RVY encodings.

  4. custom-0 is still available for custom encodings.

Table 12. RISC-V base opcode map for RVY, inst[1:0]=11

inst[4:2]

000

001

010

011

100

101

110

111 (>32b)

inst[6:5]

00

LOAD

LOAD-FP

custom-0

MISC-MEM

OP-IMM

AUIPC

OP-IMM-32

reserved

01

STORE

STORE-FP

RVY-C

AMO

OP

LUI

OP-32

reserved

10

MADD

MSUB

NMSUB

NMADD

OP-FP

OP-V

RVY-B

reserved

11

BRANCH

JALR

reserved

JAL

SYSTEM

OP-VE

RVY-A

reserved
Standard RVY encodings will never be placed in custom-0.
Parts of RVY-A/B/C may be returned to custom space in the future.

2.8.1. Instructions to Update The Capability Pointer

Creating a new capability with a different address (i.e., updating the pointer) requires specific instructions instead of integer ADD/ADDI. These instructions all include a check that the resulting address can be represented exactly within the new capability.

Table 13. Instructions which update the address field summary in RVY
Mnemonic Description

YADDI

Capability pointer increment by immediate

YADD

Capability pointer increment

YADDRW

Write capability address
2.8.1.1. YADDI

See YADD.

2.8.1.2. YADD
Synopsis

Capability pointer increment

Mnemonic

yadd rd, rs1, rs2
yaddi rd, rs1, imm

Suggested assembly syntax

add rd, rs1, rs2
addi rd, rs1, imm

Encoding
Diagram
The instruction encoding is not final and is highly likely to change prior to v1.0
Diagram
The instruction encoding is not final and is highly likely to change prior to v1.0
YADD with rs2=x0 is decoded as YMV instead, the key difference being that capabilities cannot have their capability tag cleared by YMV.
Description

Copy the capability in register rs1 to register rd.

For YADD, increment rd.address by the value in rs2[XLEN-1:0] .
For YADDI, increment rd.address by the immediate value imm.

Set rd.tag=0 if rs1 is sealed.

Set rd.tag=0 if the resulting capability cannot be represented exactly.

Set rd.tag=0 if rs1 fails any integrity checks.

Included in

RVY

Operation for YADD
  let cs1_val = C(cs1);
  let rs2_val = X(rs2);

  let newCap = incCapAddrChecked(cs1_val, rs2_val);

  C(cd) = newCap;
  RETIRE_SUCCESS
Operation for YADDI
  let cs1_val = C(cs1);
  let immBits : xlenbits = sign_extend(imm);

  let newCap = incCapAddrChecked(cs1_val, immBits);

  C(cd) = newCap;
  RETIRE_SUCCESS
2.8.1.3. YADDRW
Synopsis

Write capability address

Mnemonic

yaddrw rd, rs1, rs2

Encoding
Diagram
The instruction encoding is not final and is highly likely to change prior to v1.0
Description

Copy the capability rs1 to rd.

Set rd.address to rs2[XLEN-1:0].

Set rd.tag=0 if rs1 is sealed.

Set rd.tag=0 if the resulting capability cannot be represented exactly.

Set rd.tag=0 if rs1 fails any integrity checks.

Included in

RVY

Operation
C(cd) = setCapAddrChecked(C(cs1), X(rs2));
RETIRE_SUCCESS

2.8.2. Instructions to Manipulate Capabilities

For security, capabilities can only be modified in restricted ways. Special instructions are provided to copy capabilities or derive a new capability using manipulations such as shrinking the bounds (YBNDSW), reducing the permissions (YPERMC) or authorizing a capability with another one which has a superset (or identical) bounds and permissions (YBLD).

Capability subset

Some instructions used the concept of rs2 being a capability subset of rs1. The definition is:

  1. All permissions granted by the AP-field and the SDP-field of rs2 are also granted by those of rs1, and

  2. rs2 's top bound is less than or equal to rs1 's top bound, and

  3. rs2 's base bound is greater than or equal to rs1 's base bound, and

  4. rs1 and rs2 pass all integrity checks

Table 14. Summary of RVY instructions that create a modified capability
Mnemonic Description

YPERMC

Clear capability permissions

YMV

Capability register copy

YHIW1

Write capability metadata and clear capability tag (pseudo)

YBNDSWI

Write capability bounds by immediate

YBNDSW

Write capability bounds

YBNDSRW

Write capability bounds, rounding up if required

YSUNSEAL

Unseal by superset reconstruction

YBLD

Build capability

YSEAL

Seal a capability

YSENTRY2

Create a sealed entry point capability

1 YHIW is a pseudoinstruction for PACKY

2 YSENTRY is a pseudoinstruction for YSEAL

2.8.2.1. YPERMC
Synopsis

Clear capability permissions

Mnemonics

ypermc rd, rs1, rs2

Encoding
Diagram
The instruction encoding is not final and is highly likely to change prior to v1.0
Description

YPERMC performs the following operations:

  1. Convert the AP-field and SDP-field fields of capability rs1 into a bit field with the format shown in Figure 6.

  2. The initial value in register rs2[XLEN-1:0] is treated as a bit mask that specifies bit positions to be cleared in the bit field. Any bit that is high in rs2 will cause the corresponding bit to be cleared in the bit field.

    Future extensions may include hardwired permission bits, in which case they are not cleared by set bits in rs2.

  3. Encode the resulting architectural permissions as specified by the capability encoding format in use. This involves applying the rules in Section 2.3.10.1, as well as any rules added by extensions or the capability encoding format.

    1. If capability encoding format does not support the requested set of permissions then YPERMC will return a minimal sets of permissions, which may be no permissions. Therefore, it is possible that requesting to clear a permission also clears others, but YPERMC will never add new permissions.

  4. Copy rs1 to rd, and update the AP-field and SDP-fields with the newly calculated versions.

  5. Set rd.tag=0 if rs1 is sealed and any bits in the AP-field or SDP-field were affected by YPERMC.

    1. Extensions such as Zylevels1 add a field to the metadata which is reduced by YPERMC even when the capability is sealed.

  6. Set rd.tag=0 if any integrity checks fail.

Diagram
Figure 6. Capability permissions bit field
If a future extension adds a new permission that is dependent on an existing permission then clearing the original must also clear the new permission. This ensures software forward-compatibility: for example, a kernel that does not know about finer-grained ASR permissions must still be able to prevent all access to privileged instructions and CSRs simply by clearing ASR-permission. Such dependant permissions must be allocated to the bits that are currently reported as 1 to ensure forward-compatibility
Included in

RVY

Operation

TODO

2.8.2.2. YMV
Synopsis

Capability register copy

Mnemonic

ymv rd, rs1

Encoding
Diagram
The instruction encoding is not final and is highly likely to change prior to v1.0
YMV is encoded as YADD with rs2=x0.
Description

The contents of capability register rs1 are written to capability register rd. YMV unconditionally does a bit-wise copy from rs1 to rd .

This instruction can propagate valid capabilities which fail integrity checks.

Included in

RVY

Operation
C(cd) = C(cs1);
RETIRE_SUCCESS
2.8.2.3. PACKY
Synopsis

Pack Y register

Mnemonic

packy rd, rs1, rs2

Encoding
Diagram
The instruction encoding is not final and is highly likely to change prior to v1.0
Description

The PACKY instruction copies rs1[XLEN-1:0] into rd[XLEN-1:0], copies rs2[XLEN-1:0] into rd[YLEN-1:XLEN], and sets rd.tag=0.

Included in

RVY

2.8.2.4. YHIW
Synopsis

Capability set metadata

Mnemonic

yhiw rd, rs1, rs2

Encoding

YHIW is a pseudoinstruction for PACKY

Description

Copy rs1 to rd.

Construct a YLEN-bit value in rd by setting the address from rs1 and the metadata from rs2. Set rd.tag to 0.

The value of rs1.tag does not affect the result.
Included in

RVY

Operation
let capVal = C(cs1);
let intVal = X(rs2);
let newCap = bitsToCap(false, intVal @ capVal.address);
C(cd) = newCap;
RETIRE_SUCCESS
2.8.2.5. YBNDSWI

See YBNDSW.

2.8.2.6. YBNDSW
Synopsis

Write capability bounds

Mnemonics

ybndsw rd, rs1, rs2
ybndswi rd, rs1, imm

Encoding
Diagram
The instruction encoding is not final and is highly likely to change prior to v1.0
Diagram
The instruction encoding is not final and is highly likely to change prior to v1.0
Description

Copy the capability from register rs1 to register rd. Set the base address of its bounds to the value of rs1.address and set the length of its bounds to rs2[XLEN-1:0] for YBNDSW, or imm for YBNDSWI.

Set rd.tag=0 if rs1.tag=0, rs1 is sealed or if rd 's bounds exceed rs1 's bounds.

Set rd.tag=0 if the requested bounds cannot be encoded exactly.

Set rd.tag=0 if rs1 fails any integrity checks.

YBNDSWI decodes the 9-bit immediate value imm[8:0] to the requested length result as follows:

  • If imm[8:0] == 0, result is 4096.

  • If imm[8] == 0 and imm[7:0] != 0, result is imm[7:0].

  • If imm[8] == 1 and imm[7:5] == 0, result is 256 | (imm[3:0] << 4) | (imm[4] << 3).

  • Otherwise, result is imm[7:0] << 4.

The resulting logical encodable regions and step sizes are summarized in the table below:

Region Decoded range (bytes) Immediate decoding condition

Byte granular

1, 2, …​, 255

imm[8] == 0 and imm[7:0] > 0

8-byte granular

256, 264, …​, 504

imm[8:5] == 1000

16-byte granular

512, 528, …​, 4080

imm[8:5] > 1000

4096

imm[8:0] == 0

This immediate encoding scheme achieves coverage of almost all observed immediate values while decoding the 9-bit immediate field without any adders or variable shifts. Mapping imm == 0 to 4096 reclaims the redundant zero encoding (already covered by ybndsw dst, src, x0) to support highly common page-sized objects.

The hardware implementation is a simple bit-remapping:
result[12] = (imm[8:0] == 0);
result[11:9] = imm[8] ? imm[7:5] : 0;
result[8] = imm[8] ? (imm[7:5] == 0b000 || imm[4]) : 0;
result[7:4] = imm[8] ? imm[3:0] : imm[7:4];
result[3] = imm[8] ? (imm[7:4] == 0b0001) : imm[3];
result[2:0] = imm[8] ? 0 : imm[2:0];.
Included in

RVY

Operation for YBNDSW
let cs1_val = C(cs1);
let length = X(rs2);
let newBase = cs1_val.address;
let newTop : CapLenBits = zero_extend(newBase) + zero_extend(length);
// inCapBoundsNoWrap returns false if the input bounds are malformed.
let inBounds = inCapBoundsNoWrap(cs1_val, newBase, unsigned(length));
let (exact, newCap) : (bool, Capability) = setCapBounds(cs1_val, newBase, newTop);
let cond = not(inBounds & exact) |
           boundsMalformed(newCap) |
           not(capReservedValid(newCap)) |
           capIsSealed(newCap);
C(cd) = clearTagIf(newCap, cond);
RETIRE_SUCCESS
Operation for YBNDSWI

TODO

2.8.2.7. YBNDSRW
Synopsis

Write capability bounds, rounding up if required

Mnemonic

ybndsrw rd, rs1, rs2

Encoding
Diagram
The instruction encoding is not final and is highly likely to change prior to v1.0
Description

Copy the capability from register rs1 to register rd. Set the base address of its bounds to the value of rs1.address and set the length of its bounds to rs2[XLEN-1:0].

The base is rounded down and the top is rounded up by the smallest amounts needed to form a capability covering the requested base and top.

Set rd.tag=0 if rs1.tag=0, rs1 is sealed or if rd 's bounds exceed rs1 's bounds.

Set rd.tag=0 if rs1 fails any integrity checks.

Included in

RVY

Operation
let cs1_val = C(cs1);
let length = X(rs2);
let newBase = cs1_val.address;
let newTop : CapLenBits = zero_extend(newBase) + zero_extend(length);
// inCapBoundsNoWrap returns false if the input bounds are malformed.
let inBounds = inCapBoundsNoWrap(cs1_val, newBase, unsigned(length));
let (_, newCap) : (bool, Capability) = setCapBounds(cs1_val, newBase, newTop);
let cond = not(inBounds) |
           boundsMalformed(newCap) |
           not(capReservedValid(newCap)) |
           capIsSealed(newCap);
C(cd) = clearTagIf(newCap, cond);
RETIRE_SUCCESS
2.8.2.8. YSUNSEAL
Synopsis

Unseal by superset reconstruction

Mnemonic

ysunseal rd, rs1, rs2

Encoding
Diagram
The instruction encoding is not final and is highly likely to change prior to v1.0
Description

Copy rs2 to rd.

Set rd.ct=0. (That is, unseal rd.)

Set rd.tag=1 if:

  1. rs1.tag=1, and

  2. rs1 is not sealed, and

  3. rs2.tag=1, and

  4. rs2 is sealed, and

  5. rs2 is a capability subset of rs1

    Otherwise, set rd.tag=0

    When rs1 is x0 YSUNSEAL will copy rs2 to rd and clear rd.tag and rd.ct. However future extensions may add additional behavior to update currently reserved fields, and so software should not assume rs1==0 to be a pseudo-instruction for capability tag and type clearing.

YSUNSEAL is intended to enable "superset unsealing" of opaque handles to software objects. Specifically, a software component can:

  1. allocate the memory for these objects from a region of address space,

  2. render capabilities to these objects opaque by sealing with YSEAL,

  3. distribute these handles to other software components, and

  4. later use its (unsealed) capability to the backing region as the authority (in rs1) of a YSUNSEAL instruction to recover an unsealed capability to the object backing a handle (in rs2) received from other components.

The result of YSUNSEAL will have its capability tag set to zero if the received capability is not a handle to an object in the recipient’s address space. This makes it easy for recipient software to ensure that received capabilities actually are handles to the recipient’s objects.
Included in

RVY

Operation
2.8.2.9. YBLD
Synopsis

Build capability

Mnemonic

ybld rd, rs1, rs2

Encoding
Diagram
The instruction encoding is not final and is highly likely to change prior to v1.0
Description

Copy rs2 to rd.

Set rd.tag=1 if:

  1. rs1.tag=1, and

  2. rs1 is not sealed, and

  3. rs2 is a capability subset of rs1

    Otherwise, set rd.tag=0

The integrity check on rs2 (included in the subset check) is required to prevent authorising a capability with a lack of integrity. The integrity check on rs1 is optional.
YBLD is typically used alongside YHIW to build capabilities from integer values.
YBLD can be used to speed up operations such as paging in memory after swap.
When rs1 is x0 YBLD will copy rs2 to rd and clear rd.tag. However future extensions may add additional behavior to update currently reserved fields, and so software should not assume rs1==0 to be a pseudo-instruction for capability tag clearing.
Included in

RVY

Operation
  let cs1_val = C(cs1);
  let cs2_val = C(cs2);

  let tag = cs1_val.tag &
            not(capIsSealed(cs1_val)) &
            capIsSubset(cs2_val, cs1_val); /* Subset checks for malformed bounds,
                                              perms, and reserved bits */

  C(cd) = { cs2_val with tag = tag };
  RETIRE_SUCCESS
2.8.2.10. YSEAL
Synopsis

Seal a capability

Mnemonic

yseal rd, rs1, rs2

Encoding
Diagram
The instruction encoding is not final and is highly likely to change prior to v1.0
rs1≠x0 is currently reserved.
Description

Copy rs2 to rd.

Seal the capability in rd by setting rd.ct=1.

Set rd.tag=0 if rs2 is sealed.

A future extension may allow rs1≠x0 to specify an authorizing capability to perform the sealing action, and also to specify the destination CT-field value. The same extension is likely to add privileged state to disallow the use of rs1=x0 for sealing the output without an authorizing capability.
If the capability encoding format does not support a CT-field of type 1, then this instruction is reserved. Future capability formats may redefine the behavior for CT-field values greater than 1.
Included in

RVY

Operation

TODO

2.8.2.11. YSENTRY
Synopsis

Create a sealed entry point capability

Mnemonic

ysentry rd, rs2

Encoding

YSENTRY is a pseudoinstruction for YSEAL with rs1=x0.

Description

Copy rs2 to rd.

Seal the capability in rd as a sealed entry point capability by setting rd.ct=1.

Set rd.tag=0 if rs2 is sealed.

Included in

RVY

2.8.3. Instructions to Decode Capability Bounds

The bounds describing the range of addresses the capability gives access to are stored in a compressed format. These instructions query the bounds and related information.

Table 15. Instructions which decode capability bounds summary in RVY
Mnemonic Description

YBASER

Read capability base address

YLENR

Read capability length

YTOPR

Read capability top address
2.8.3.1. YBASER
Synopsis

Read capability base address

Mnemonic

ybaser rd, rs1

Encoding
Diagram
The instruction encoding is not final and is highly likely to change prior to v1.0
Description

Decode the base integer address from rs1 's bounds and write the result to rd.

If rs1 's bounds can’t be decoded, or rs1 fails any integrity checks, then return zero.

The value of rs1.tag does not affect the result.
Included in

RVY

Operation
let capVal = C(cs1);
X(rd) = match getCapBoundsBits(capVal) {
  None() => zeros(),
  Some(base, _) => base
};
RETIRE_SUCCESS
2.8.3.2. YLENR
Synopsis

Read capability length

Mnemonic

ylenr rd, rs1

Encoding
Diagram
The instruction encoding is not final and is highly likely to change prior to v1.0
Description

Calculate the length of rs1 's bounds and write the result in rd.

The length is defined as the difference between the decoded bounds' top and base addresses, i.e., top - base.

Return the maximum length, 2MXLEN-1, if the length of rs1 is 2MXLEN.

If rs1 's bounds can’t be decoded, or rs1 fails any integrity checks, then return zero.

The value of rs1.tag does not affect the result.
Included in

RVY

Operation
let capVal = C(cs1);
// getCapLength returns 0 if the bounds are malformed
let len = getCapLength(capVal);
X(rd) = to_bits(xlen, if len > cap_max_addr then cap_max_addr else len);
RETIRE_SUCCESS
2.8.3.3. YTOPR
Synopsis

Read capability top address

Mnemonic

ytopr rd, rs1

Encoding
Diagram
The instruction encoding is not final and is highly likely to change prior to v1.0
Description

Decode the top integer address from rs1 's bounds, saturated to 2MXLEN-1, and write the result to rd.

If rs1 's bounds can’t be decoded, or rs1 fails any integrity checks, then return zero.

The value of rs1.tag does not affect the result.
Included in

RVY

Operation

TODO

2.8.4. Instructions to Extract Capability Fields

These instructions either directly read bit fields from the metadata or capability tag, or only apply simple transformations on the metadata.

Table 16. Instructions which extract capability fields summary in RVY
Mnemonic Description

YTAGR

Read capability tag

YPERMR

Read capability permissions

YTYPER

Read capability type

YHIR1

Read capability metadata (pseudo)
2.8.4.1. YTAGR
Synopsis

Read capability tag

Mnemonic

ytagr rd, rs1

Encoding
Diagram
The instruction encoding is not final and is highly likely to change prior to v1.0
Description

Zero extend the value of rs1.tag and write the result to rd.

Included in

RVY

Operation
let capVal = C(cs1);
X(rd) = zero_extend(bool_to_bits(capVal.tag));
RETIRE_SUCCESS
2.8.4.2. YPERMR
Synopsis

Read capability permissions

Mnemonic

ypermr rd, rs1

Encoding
Diagram
The instruction encoding is not final and is highly likely to change prior to v1.0
Description

Convert the unpacked AP-field and SDP-fields of capability rs1 into a bit field, with the same format as used by YPERMC (see Figure 6), zero extend and write the result to rd.

All bits in the [23:0] range that are reserved or assigned to extensions that are not implemented by the current hart always report 1.

All architectural permission bits in rd are set to 0 if any integrity checks failed.

Diagram
Figure 7. Capability permissions bit field
The value of rs1.tag does not affect the result.
Included in

RVY

Operation
let capVal = C(cs1);
X(rd) = packPerms(getArchPermsLegalized(capVal), capVal.sd_perms).bits;
RETIRE_SUCCESS
2.8.4.3. YTYPER
Synopsis

Read capability type

Mnemonic

ytyper rd, rs1

Encoding
Diagram
The instruction encoding is not final and is highly likely to change prior to v1.0
Description

Decode the architectural capability type (CT-field) from rs1, zero extend and write the result to rd.

The value of rs1.tag does not affect the result.
Included in

RVY

Operation
let capVal = C(cs1);
X(rd) = zero_extend(bool_to_bits(capVal.sealed));
RETIRE_SUCCESS
2.8.4.4. SRLIY
Synopsis

Logical right shift of Y register

Mnemonic

srliy rd, rs1, shamt

Encoding
Diagram
The instruction encoding is not final and is highly likely to change prior to v1.0
Description

Logical right shift of Y register rs1 to rd, zero-filling the upper bits of the result. Set rd.tag=0.

The only valid shift amount is XLEN (i.e., YHIR). A future extension may add an arbitrary shift distance.
Included in

RVY

Operation

TODO

2.8.4.5. YHIR
Synopsis

Read capability metadata (pseudo)

Mnemonic

yhir rd, rs1

Encoding
Diagram
The instruction encoding is not final and is highly likely to change prior to v1.0
yhir rd, rs1 is a pseudo instruction for a XLEN-wide right shift (SRLIY).
Description

Copy the metadata (bits [YLEN-1:XLEN]) of capability rs1 into rd.

The value of rs1.tag does not affect the result.
Included in

RVY

Operation
let capVal = C(cs1);
X(rd) = capToMetadataBits(capVal).bits;
RETIRE_SUCCESS

2.8.5. Miscellaneous Instructions to Handle Capability Data

Table 17. Miscellaneous capability instruction summary in RVY
Mnemonic Description

YEQ

Capability equality comparison including capability tag

YSS

Capability is a subset

YAMASK

Capability alignment mask
2.8.5.1. YEQ
Synopsis

Capability equality comparison including capability tag

Mnemonic

yeq rd, rs1, rs2

Encoding
Diagram
The instruction encoding is not final and is highly likely to change prior to v1.0
Description

Set rd to 1 if all bits (i.e., YLEN bits and the capability tag) of capabilities rs1 and rs2 are equal, otherwise set rd to 0.

Included in

RVY

Operation
let cs1_val = C(cs1);
let cs2_val = C(cs2);
X(rd) = zero_extend(bool_to_bits(cs1_val == cs2_val));
RETIRE_SUCCESS
2.8.5.2. YSS
Synopsis

Capability is a subset

Mnemonic

yss rd, rs1, rs2

Encoding
Diagram
The instruction encoding is not final and is highly likely to change prior to v1.0
Description

rd is set to 1 if:

  1. the capability tag of capabilities rs1 and rs2 are equal, and

  2. rs2 is a capability subset of rs1

Otherwise set rd to 0. Extensions may further impose constraints on when rd is set to 1.

The implementation of this instruction is similar to YBLD, although YSS does not include the sealed bit in the check.
Included in

RVY

Operation
  let cs1_val = C(cs1);
  let cs2_val = C(cs2);

  X(rd) = zero_extend(bool_bits(
    (cs1_val.tag == cs2_val.tag) &
    capIsSubset(cs2_val, cs1_val) /* capIsSubset returns false if either input
                                     has malformed bounds, perms, or non-zero
                                     reserved bits */
  ));
  RETIRE_SUCCESS
2.8.5.3. YAMASK
Synopsis

Capability alignment mask

Mnemonic

yamask rd, rs1

Encoding
Diagram
The instruction encoding is not final and is highly likely to change prior to v1.0
Description

rd[XLEN-1:0] is set to a mask that can be used to round addresses down to a value that is sufficiently aligned to set exact bounds for the nearest representable length of rs1[XLEN-1:0]. The upper bits of rd are zero extended. See Section 3.1.6 for the algorithm used to compute the next representable length.

This RVY instruction only handles XLEN operands and produces an XLEN result.
Included in

RVY

Operation
let len = X(rs1);
X(rd) = getRepresentableAlignmentMask(len);
RETIRE_SUCCESS

2.8.6. Instructions to Load and Store Capability Data

New loads and stores are introduced to handle capability data, LY and SY. They atomically access YLEN bits of data and the associated capability tag.

All capability memory accesses check for C-permission in the authorizing capability in rs1.

If C-permission is granted then:

  • LY reads YLEN bits of data from memory, and returns the associated capability tag.

  • SY writes YLEN bits of data to memory, and writes the associated capability tag.

If C-permission is not granted then:

  • LY reads YLEN bits from memory, but does not return the associated capability tag, instead zero is returned.

  • SY writes YLEN bits to memory, and writes zero to the associated capability tag.

All capability data memory access instructions require YLEN-aligned addresses, and will take an access fault exception if this requirement is not met. They cannot be emulated.

An access fault is raised instead of a misaligned exception since these instructions cannot be emulated since there is one hidden capability tag per YLEN-aligned memory region.

All memory accesses, of any type, require permission from the authorizing capability in rs1.

  • All loads require R-permission, otherwise they raise an exception.

  • All stores require W-permission, otherwise they raise an exception.

Under some circumstances LY will modify the data loaded from memory before writing it back to the destination register. See LY for details.

Table 18. Capability load/store instruction summary in RVY
Mnemonic Description

LY

Load capability

SY

Store capability
2.8.6.1. LY
Synopsis

Load capability

Mnemonic

ly rd, offset(rs1)

Encoding
Diagram
The instruction encoding is not final and is highly likely to change prior to v1.0
Any instance of this instruction with rs1=x0 will raise an exception, as x0 is defined to always hold a NULL capability. As such, the encodings with rs1=x0 are RESERVED for use by future extensions.
Description

Calculate the effective address of the memory access by adding rs1.address to the sign-extended 12-bit offset.

Authorize the memory access with the capability in rs1.

Load a naturally aligned YLEN-bit data value and associated capability tag from memory.

The loaded capability tag may be set to zero under platform specified conditions, such as by a hardware based use-after-free prevention scheme. Privileged mechanisms may also set the loaded capability tag to zero such as PMA settings or virtual memory system page table entry configuration..
Extensions may also specify additional conditions which set the capability tag to zero.

The final value of rd is determined as follows:

  1. Write the loaded YLEN data to rd and the loaded capability tag into rd.tag

  2. If rs1 does not grant C-permission then set rd.tag=0.

  3. If rd.tag=1, rs1 does not grant LM-permission and rd is not sealed, then an implicit YPERMC is performed to clear W-permission and LM-permission from rd.

    The implicit YPERMC is not required to set rd.tag=0 if the loaded data fails integrity checks. A suggested implementation is to only check the AP-field for legal values, and set the whole AP-field to zero if the check fails.
    Metadata modifications on loaded capabilities are on naturally aligned data. Therefore on the read path from a data cache, the modification typically happens in parallel with data alignment multiplexers.
    Zylevels1 adds to the YPERMC rules on loaded capability data.

    When sending load data to a trace interface, implementations trace the final value written to rd which may not match the value in memory.

Exceptions

Load access fault exception when the effective address is not aligned to YLEN/8.

Exceptions occur when the authorizing capability fails one of the checks listed below:

Kind Reason

CHERI Load Access Fault

Authorizing capability tag is set to 0.

CHERI Load Access Fault

Authorizing capability is sealed.

CHERI Load Access Fault

Authorizing capability does not grant the necessary permissions. R-permission is required.

CHERI Load Access Fault

At least one byte accessed is outside the authorizing capability bounds, or the bounds could not be decoded.

CHERI Load Access Fault

Authorizing capability failed any integrity check.
Included in

RVY

Operation
  let offset : xlenbits = sign_extend(imm);
  let (auth_val, vaddr) = get_cheri_mode_cap_addr(rs1_cs1, offset);
  let aq : bool = false;
  let rl : bool = false;

  match check_and_handle_load_vaddr_for_triggers(vaddr, get_arch_pc()) {
    Some (ret) => return ret,
    None () => ()
  };
  if not(capTaggedAndReservedValid(auth_val)) then {
    handle_cheri_exception(CapCheckType_Data, CapEx_TagViolation);
    RETIRE_FAIL
  } else if capIsSealed(auth_val) then {
    handle_cheri_exception(CapCheckType_Data, CapEx_SealViolation);
    RETIRE_FAIL
  } else if not(canR(auth_val)) then {
    handle_cheri_exception(CapCheckType_Data, CapEx_PermissionViolation);
    RETIRE_FAIL
  } else if not(validAddrRange(vaddr, cap_size) | capBoundsInfinite(auth_val)) then {
    handle_cheri_exception(CapCheckType_Data, CapEx_InvalidAddressViolation);
    RETIRE_FAIL
  } else if not(inCapBounds(auth_val, vaddr, cap_size)) then {
    handle_cheri_exception(CapCheckType_Data, CapEx_LengthViolation);
    RETIRE_FAIL
  } else if not(is_aligned_addr(vaddr, cap_size)) then {
    handle_mem_exception(vaddr, E_Load_Addr_Align());
    RETIRE_FAIL
  } else match translateAddr(vaddr, Read(Cap)) {
    TR_Failure(E_Extension(_)) => { internal_error(__FILE__, __LINE__, "unexpected cheri exception for cap load") },
    TR_Failure(e) => { handle_mem_exception(vaddr, e); RETIRE_FAIL },
    TR_Address(addr, pbmt, ptw_info) => {
      let c = mem_read_cap(addr, pbmt, aq, aq & rl, false);
      match c {
        MemValue(v) => {
          let cr = clearTagIf(v, ptw_info.ptw_lc == PTW_LC_CLEAR | not(canC(auth_val)));
          C(cd) = legalizeLM(cr, auth_val);
          RETIRE_SUCCESS
        },
        MemException(e) => {handle_mem_exception(vaddr, e); RETIRE_FAIL }
      }
    }
  }
2.8.6.2. SY
Synopsis

Store capability

Mnemonic

sy rs2, offset(rs1)

Encoding
Diagram
The instruction encoding is not final and is highly likely to change prior to v1.0
Any instance of this instruction with rs1=x0 will raise an exception, as x0 is defined to always hold a NULL capability. As such, the encodings with rs1=x0 are RESERVED for use by future extensions.
Description

Calculate the effective address of the memory access by adding rs1.address to the sign-extended 12-bit offset.

Authorize the memory access with the capability in rs1.

Store a naturally aligned YLEN-bit data value in rs2 to memory and the associated capability tag in rs2.

This instruction can propagate valid capabilities which fail integrity checks.

Stored Capability Tag Value

The stored capability tag is set to zero if:

  1. rs2.tag=0, or

  2. rs1 does not grant C-permission, or

  3. Platform specified conditions or privileged configurations (e.g., PMAs or PTEs) specify that it must be set to zero.

    Extensions may define further circumstances under which stored capabilities may have their capability tags set to zero.
Exceptions

Store/AMO access fault exception when the effective address is not aligned to YLEN/8.

Store/AMO access fault if the stored capability tag is set to one and the PMA is CHERI Capability Tag Store Fault.

Exceptions occur when the authorizing capability fails one of the checks listed below:

Kind Reason

CHERI Store/AMO Access Fault

Authorizing capability tag is set to 0.

CHERI Store/AMO Access Fault

Authorizing capability is sealed.

CHERI Store/AMO Access Fault

Authorizing capability does not grant the necessary permissions. W-permission is required.

CHERI Store/AMO Access Fault

At least one byte accessed is outside the authorizing capability bounds, or the bounds could not be decoded.

CHERI Store/AMO Access Fault

Authorizing capability failed any integrity check.
Included in

RVY

Operation
  let offset : xlenbits = sign_extend(imm);
  let (auth_val, vaddr) = get_cheri_mode_cap_addr(rs1_cs1, offset);
  let cs2_val = C(cs2);
  let aq : bool = false;
  let rl : bool = false;
  let cs2_val = clearTagIf(cs2_val, not(canC(auth_val)));

  match check_and_handle_store_vaddr_for_triggers(vaddr, get_arch_pc()) {
    Some (ret) => return ret,
    None () => ()
  };
  if not(capTaggedAndReservedValid(auth_val)) then {
    handle_cheri_exception(CapCheckType_Data, CapEx_TagViolation);
    RETIRE_FAIL
  } else if capIsSealed(auth_val) then {
    handle_cheri_exception(CapCheckType_Data, CapEx_SealViolation);
    RETIRE_FAIL
  } else if not(canW(auth_val)) then {
    handle_cheri_exception(CapCheckType_Data, CapEx_PermissionViolation);
    RETIRE_FAIL
  } else if not(validAddrRange(vaddr, cap_size) | capBoundsInfinite(auth_val)) then {
    handle_cheri_exception(CapCheckType_Data, CapEx_InvalidAddressViolation);
    RETIRE_FAIL
  } else if not(inCapBounds(auth_val, vaddr, cap_size)) then {
    handle_cheri_exception(CapCheckType_Data, CapEx_LengthViolation);
    RETIRE_FAIL
  } else if not(is_aligned_addr(vaddr, cap_size)) then {
    handle_mem_exception(vaddr, E_SAMO_Addr_Align());
    RETIRE_FAIL
  } else match translateAddr(vaddr, Write(if cs2_val.tag then Cap else Data)) {
    TR_Failure(e) => { handle_mem_exception(vaddr, e); RETIRE_FAIL },
    TR_Address(addr, pbmt, _) => {
      let eares : MemoryOpResult(unit) = mem_write_ea_cap(addr, aq & rl, rl, false);
      match (eares) {
        MemException(e) => { handle_mem_exception(vaddr, e); RETIRE_FAIL },
        MemValue(_) => {
          let res : MemoryOpResult(bool) = mem_write_cap(addr, pbmt, cs2_val, aq & rl, rl, false);
          match (res) {
            MemValue(true)  => RETIRE_SUCCESS,
            MemValue(false) => internal_error(__FILE__, __LINE__, "store got false from mem_write_value"),
            MemException(e) => { handle_mem_exception(vaddr, e); RETIRE_FAIL }
          }
        }
      }
    }
  }

2.9. Changes to Existing RISC-V Base ISA Instructions

RVY extend existing instructions that are used for handling addresses so that they manipulate a whole capability.

  • Whenever an input operand is used as an address (e.g., the load/store base address), all capability bits are fed into the instruction instead of just XLEN bits.

  • Any instruction that writes back an address (e.g., AUIPC (RVY) or CSRRW (RVY)) to the destination register, writes a full capability register instead of just XLEN bits. For all other results the high bits of the register and the capability tag are set to zero.

  • Whenever a capability with a new address is returned, the result is always created using the semantics of the YADDRW instruction.

ADD and ADDI are not affected by the rule above. Even though they are used for handling addresses, they also have other uses. New encodings are used for capability addition: YADD and YADDI. They must be used for all capability address incrementing.

Integer add (ADD) and capability add (YADD) have separate encodings. Using a single encoding for both is undesirable:

  1. Integer ADD is most commonly used for purposes other than address calculations.

  2. For high performance implementations which can issue multiple ADDs, it means that the integer ADD units don’t need the upper halves of the operands, and don’t need the capability check logic on the result.

  3. The compiler and/or programmer would have to execute another metadata clearing instruction after each ADD to ensure that compartments don’t leak capabilities.

The rules above apply to the base ISA instructions listed in the following subsections, but also apply to instructions added by other extensions. Any change to instruction semantics for RVY is called out in the chapter defining the extension.

2.9.1. Changes to load/stores

All load and store instructions behave as described in Load and Store Instructions with one fundamental difference:

  • Any memory instruction that has rs1 as a base address register reads the full capability register instead. The base address is unchanged, i.e., using the value from rs1. The metadata and capability tag are used to authorize the access.

  • For a load instruction, the lower XLEN bits of the result written to the destination register are the same as in the RV32[IE]/RV64[IE] specification.

All load and store instructions authorized by rs1 raise exceptions if any of these checks fail:

  • rs1 must not be x01

  • The capability tag (rs1.tag) must be set to one.

  • rs1 must be unsealed.

  • For loads, R-permission must be granted in rs1.

  • For stores, W-permission must be granted in rs1.

  • All integrity checks on rs1 must pass.

1 All load/store encodings are reserved if rs1=x0 (since dereferencing NULL always faults).

All load instructions, except for LY, always set the capability tag and the metadata of the result register to zero.

All store instructions, except for SY, always write zero to the capability tag or capability tags associated with the memory locations that are written to. Therefore, misaligned stores may clear up to two associated capability tag bits.

The changed interpretation of the base register also applies to all loads, stores and all other memory operations defined in later chapters of this specification with a base operand of rs1 unless stated otherwise.

Under RVY all loads and stores are authorized by rs1.

These rules affect the following base ISA instructions listed in Table 20, and also apply to instructions added by other extensions, e.g.,:

Table 19. Changed RISC-V base ISA load/store instructions summary in RVY
Mnemonic Description

LD, LW[U], LH[U], LB[U]

Integer loads (authorized by the capability in rs1)

SD, SW, SH, SB

Integer stores (authorized by the capability in rs1)

2.9.2. Changes to PC

  • Whenever the address field of the pc is modified, it is always updated using the semantics of the YADDRW instruction. This includes adding an offset to the pc from direct jumps and branches for both the target address and the link register. In this case, e.g., new_pc = YADDRW(old_pc, offset)

  • JALR (RVY) copies rs1 into the pc, and increments the address field with the offset. In this case, e.g., new_pc = YADDRW(Unseal(rs1), offset)

Using YADDRW to update the pc address can cause the target pc capability tag to be set to zero which raises a CHERI Instruction Access Fault on the target instruction.
A future extension may raise an exception on branch and jump instructions if fetching a minimum sized instruction at the target pc will raise a CHERI Instruction Access Fault.

These rules affect the following base ISA instructions listed in Table 20, and also apply to instructions added by other extensions, e.g.:

Table 20. Changed RISC-V base ISA PC relative instructions summary in RVY
Mnemonic Description

AUIPC (RVY)

Add upper immediate to pc

JAL (RVY)

Immediate offset jump, and link to capability register

JALR (RVY)

Jump to capability register, and link to capability register

2.9.3. AUIPC (RVY)

Synopsis

Add upper immediate to pc

Mnemonic

auipc rd, imm

Encoding
Diagram
This instruction is extended from the version in the base ISA.
Description

Form a 32-bit offset from the 20-bit immediate filling the lowest bits with zero; the number of places to shift is determined by the capability encoding format’s AUIPC_SHIFT parameter (defaulting to 12). Take the value of the AUIPC instruction’s pc, increment its address by the 32-bit offset using the semantics of the YADDRW instruction and write the result to rd.

Set rd.tag=0 if the resulting capability cannot be represented exactly.

Included in

RVI (RVY modified behavior)

Operation
let off : xlenbits = sign_extend(imm @ 0x000);
let (representable, newCap) = setCapAddr(PCC, PC + off);
C(cd) = clearTagIf(newCap, not(representable));
RETIRE_SUCCESS

2.9.4. JAL (RVY)

Synopsis

Immediate offset jump, and link to capability register

Mnemonic

jal rd, offset

Encoding
Diagram
Description

Direct jump with an address offset.

  1. The target address is obtained by adding the sign-extended 20-bit J-immediate to pc.address.

  2. The target pc is obtained by applying the target address to the pc of the JAL instruction, using the semantics of the YADDRW instruction.

  3. The pc of the next instruction is sealed as a sealed entry point capability and written to rd.

  4. Jump to the target pc.

    A future extension may raise an exception on the branch instruction itself if fetching a minimum sized instruction at the target pc will raise a CHERI Instruction Access Fault. Performing the pc bounds check at the branch source instead of on instruction fetch is helpful for debugging and can simplify the implementation of CPUs with very short pipelines.
Included in

RVI (RVY modified behavior)

Operation

TODO

2.9.5. JALR (RVY)

Synopsis

Jump to capability register, and link to capability register

Mnemonic

jalr rd, rs1, offset

Encoding
Diagram
This instruction is extended from the version in the base ISA.
Description

Indirect jump to the target capability in rs1 with an address offset.

  1. rs1 is written to the target pc.

  2. The target address is obtained by adding the sign-extended 12-bit I-immediate to rs1.address, then setting the least-significant bit of the result to zero.

  3. Unseal the target pc if it is a sealed entry point capability, rs1.address[0] is zero, and the I-immediate is zero.

  4. Set the address of the target pc to the target address using the semantics of the YADDRW instruction.

  5. The pc of the next instruction is sealed as a sealed entry point capability and written to rd.

  6. Jump to the target pc.

    A future extension may raise an exception on the JALR (RVY) instruction itself if the target pc will raise a CHERI Instruction Access Fault at the target.
Included in

RVI (RVY modified behavior)

Operation

TODO

2.9.6. BEQ, BNE (RVY)

For beq and bne only, if rs1≤rs2 then the encoding is reserved.

The reserved encodings are redundant and may be used by future extensions, suggestions include branching on capability tag values only, or YLEN-bit compares.
The assembler will need to transform from, e.g., beq x1, x2 to beq x2, x1.

2.10. RVY ISA summary

2.10.1. RVY added instructions

Table 21. RVY added instructions
Mnemonic Function

LY

Load capability

SY

Store capability

YADD

Capability pointer increment

YADDI

Capability pointer increment by immediate

YADDRW

Write capability address

YTAGR

Read capability tag

YPERMR

Read capability permissions

YMV

Capability register copy

YPERMC

Clear capability permissions

SRLIY

Logical right shift of Y register

YHIR

Read capability metadata (pseudo)

PACKY

Pack Y register

YHIW

Write capability metadata and clear capability tag (pseudo)

YEQ

Capability equality comparison including capability tag

YSEAL

Seal a capability

YSS

Capability is a subset

YBLD

Build capability

YSUNSEAL

Unseal by superset reconstruction

YBNDSW

Write capability bounds

YBNDSWI

Write capability bounds by immediate

YBNDSRW

Write capability bounds, rounding up if required

YAMASK

Capability alignment mask

YBASER

Read capability base address

YLENR

Read capability length

YTYPER

Read capability type

YTOPR

Read capability top address

2.10.2. RVI (RVY modified behavior)

Table 22. RVI (RVY modified behavior) instructions
Mnemonic Function

AUIPC (RVY)

Add upper immediate to pc

JAL (RVY)

Immediate offset jump, and link to capability register

JALR (RVY)

Jump to capability register, and link to capability register

3. The RV64LYA Capability Base for RV64Y

This section describes the in-memory format and properties of the capability encoding intended for RV64Y.

The format is closely modeled upon features from CHERI v9 (Watson et al., 2023), and the bounds encoding scheme is based upon CHERI Concentrate (Woodruff et al., 2019).

3.1. Capability Encoding

Diagram
Figure 8. Capability encoding for RV64LYA
Reserved bits must be 0 in valid capabilities and are available for future extensions to RVY.

The encoding diagram above of the capability encoding format includes some fields which depend upon the presence of extensions:

Zyhybrid

When Zyhybrid is supported, capabilities include an P-bit (bit 52). If the capability does not grant X-permission or Zyhybrid is not supported, the P-bit is always reserved and so must be zero in valid capabilities.

Zylevels1

Bits 6 and 7 of the AP field are allocated, otherwise they are reserved and set to one.

The GL flag is allocated, otherwise it is reserved and must be zero.

3.1.1. Capability Encoding Parameter Summary

Table 23. RV64LYA capability encoding format parameters
Parameter Default Value Comment

MW

14

Mantissa width for the bounds encoding

EW

6

Exponent width for the bounds encoding

CAP_MAX_E

52

Maximum exponent value

enableL8

0

Whether the encoding format includes the L8 bit

AP_MAX

all ones

Value of the AP-field giving maximum permissions

AP_ENC

ps1

Encoding scheme for the AP-field

AUIPC_SHIFT

12

Shift distance of the immediate in AUIPC

REP_REGION

rc1

Whether out of bounds addresses are representable
All parameters match the default values from Table 9.
Table 24. RV64LYA extension summary
Extension Comment

Zyhybrid

Compatible

Zylevels1

Compatible

All RV64Y versions of other standard extensions

Compatible

3.1.2. Architectural Permissions (AP) Encoding - ps1

The permissions field is up to 8 bits wide and is encoded using one bit per architectural permission as shown in Table 25. A permission is granted if its corresponding bit, and those of any dependent permissions, are set; otherwise, the capability does not grant that permission.

Certain combinations of permissions are impractical. For example, C-permission is superfluous when the capability does not grant either R-permission or W-permission. Therefore, it is only legal to encode a subset of all combinations and any invalid states are reserved in the AP-field encoding.

Reserved states can be used in the future to allow additional permissions to be encoded within the AP-field.
Table 25. Encoding of architectural permissions for RV64LYA
Bit Encoded permission

0

C-permission

1

W-permission

2

R-permission

3

X-permission

4

ASR-permission

5

LM-permission

6

LG-permission if Zylevels1 is implemented; reserved 1 otherwise.

7

SL-permission if Zylevels1 is implemented; reserved 1 otherwise.
Future extensions may define new permissions and, if so, must augment the above table.

3.1.3. Capability Mode (P) Encoding

The P-bit is only assigned meaning when the implementation supports Zyhybrid and X-permission is set. In valid capabilities, the bit assigned to the P-bit must be zero if X-permission isn’t set.

3.1.4. Software-Defined Permissions (SDP) Encoding

The SDP-field is 4 bits wide. The value of the SDP-field bits of the YPERMR result maps 1:1 to the SDP-field in the capability.

3.1.5. Capability Type (CT) Encoding

The CT-field field is as specified in Table 6.

The capabilities of this chapter define a 1-bit field for CT-field values; this field directly encodes the values 0 and 1:

  • The encoded value 0 indicates an unsealed type.

  • The encoded value 1 indicates an sealed type which can unsealed if in rs1 passed to JALR (RVY), and is used to seal the capability written to rd.

3.1.6. Bounds (EF, T, TE, B, BE) Encoding

3.1.6.1. Concept

This bounds encoding scheme is based upon (Woodruff et al., 2019).

The bounds encode the base and top addresses that constrain memory accesses. The capability can be used to access any memory location A in the range base ≤ A < top. The bounds are encoded in a compressed format, so it is not possible to encode any arbitrary combination of base and top addresses. An invalid capability with capability tag cleared is produced when attempting to construct a capability that is not representable because its bounds cannot be correctly encoded. The bounds are decoded as described in Section 3.1.

The bounds field has the following components:

  • T: Value substituted into the capability’s address to decode the top address

  • B: Value substituted into the capability’s address to decode the base address

  • E: Exponent that determines the position at which B and T are substituted into the capability’s address

  • EF: Exponent format flag indicating the encoding for T, B and E

    • The exponent is stored in T and B if EF=0, so it is 'internal'

    • The exponent is zero if EF=1

The bit width of T and B are defined in terms of the mantissa width (MW) which is set depending on capability encoding as shown in Table 23.

The exponent E indicates the position of T and B within the capability’s address as described in Section 3.1. The bit width of the exponent (EW) is set depending on the encoding. The maximum value of the exponent is calculated as follows:

CAP_MAX_E = XLEN - MW + 2

The values of EW and CAP_MAX_E are shown in Table 23.

The address and bounds must be representable in valid capabilities i.e., when the capability tag is set (see Section 3.1.6.4).
3.1.6.2. Decoding

The metadata is encoded in a compressed format termed CHERI Concentrate (Woodruff et al., 2019). It uses a floating point representation to encode the bounds relative to the capability address. The base and top addresses from the bounds are decoded as shown below.

The pseudocode below does not have a formal notation. It is a place-holder until the sail implementation has been integrated in the specification. In this notation, / means "integer division", [] are the bit-select operators, and arithmetic is signed. 
EW        = (XLEN == 32) ? 5 : 6
CAP_MAX_E = XLEN - MW + 2

If EF = 1:
    E               = 0
    T[EW / 2 - 1:0] = TE
    B[EW / 2 - 1:0] = BE
    LCout           = (T[MW - 3:0] < B[MW - 3:0]) ? 1 : 0
    LMSB            = (XLEN == 32) ? L~8~ : 0
else:
    E               = CAP_MAX_E - ( (XLEN == 32) ? { L~8~, TE, BE } : { TE, BE } )
    T[EW / 2 - 1:0] = 0
    B[EW / 2 - 1:0] = 0
    LCout           = (T[MW - 3:EW / 2] < B[MW - 3:EW / 2]) ? 1 : 0
    LMSB            = 1

Reconstituting the top two bits of T:

T[MW - 1:MW - 2] = B[MW - 1:MW - 2] + LCout + LMSB

The bounds are decoded as shown in Figure 9 and Figure 10.

Diagram
Figure 9. Decoding of the XLEN+1 wide top (t) bound
Diagram
Figure 10. Decoding of the XLEN wide base (b) bound

Figure 9 and Figure 10 include ranges which may not be present when the bounds are decoded:

  • If E = 0 the lower section does not exist.

  • If E+MW=XLEN then the top section is only the least significant bit of ct for the top bound, and top section doesn’t exist for the bottom bound.

  • If E+MW>XLEN then neither top section exists, and so the bounds are calculated with no dependency on the address field a.

The corrections ct and cb are calculated as shown below using the definitions in Table 26 and Table 27.

A[MW-1:0] = a[E + MW - 1:E]
R[MW-1:0] = B - 2MW-2
The comparisons in Table 26 and Table 27 are unsigned.
Table 26. Calculation of top address correction
A < R T < R ct

false

false

0

false

true

+1

true

false

-1

true

true

0
Table 27. Calculation of base address correction
A < R B < R cb

false

false

0

false

true

+1

true

false

-1

true

true

0

The base, b, and top, t, addresses are derived from the address by substituting a[E + MW - 1:E] with B and T respectively and clearing the lower E bits. The most significant bits of a may be adjusted up or down by 1 using corrections cb and ct to allow encoding memory regions that span alignment boundaries.

The EF bit selects between two cases:

  1. EF = 1: The exponent is 0. When enableL8=1, L8 encodes the MSB of the length, which can be used to derive T[MW-1:MW-2], forming a full MW-wide T field.

  2. EF = 0: The exponent is internal with E stored in the lower bits of T and B, with L8 used for the MSB of E when enableL8=1. E is chosen so that the most significant non-zero bit of the length of the region aligns with T[MW - 2] such that this bit is implied by E.

The most significant two bits of T can be derived from B using the equality T = B + L, where L[MW - 2] is known from the values of EF and E (as well as L8 when enableL8=1). A carry out is implied if T[MW - 3:0] < B[MW - 3:0] since it is guaranteed that the top is larger than the base.

The compressed bounds encoding allows the address to roam over a large representable region while maintaining the original bounds. This is enabled by defining a lower boundary R from the out-of-bounds values that allows us to disambiguate the location of the bounds with respect to an out-of-bounds address. R is calculated relative to the base by subtracting 2MW-2 from B. If B, T or a[E + MW - 1:E] is less than R, it is inferred that they lie in the 2E+MW aligned region above R labeled spaceU in Figure 2 and the corrections ct and cb are computed accordingly. The overall effect is that the address can roam 2E+MW/4 bytes below the base address and at least 2E+MW/4 bytes above the top address while still allowing the bounds to be correctly decoded.

3.1.6.3. Top bound MSB correction

A capability has infinite bounds if its bounds cover the entire address space such that the base address b=0 and the top address t≥2XLEN, i.e., t is an XLEN + 1 bit value. However, b is an XLEN-bit value and the size mismatch introduces additional complications when decoding, so the following condition is required to correct t for capabilities whose Representable Range wraps the edge of the address space:

if ( (E < (CAP_MAX_E - 1)) && (t[XLEN: XLEN - 1] - b[XLEN - 1] > 1) )
    t[XLEN] = !t[XLEN]
The comparison is unsigned.

That is, invert the most significant bit of t if the decoded length of the capability is larger than E.

A capability has infinite bounds if E=CAP_MAX_E and it is not malformed (see Section 3.1.6.4); this check is equivalent to b=0 and t≥2XLEN.
3.1.6.4. Malformed Capability Bounds

A capability is malformed if its bounds cannot be correctly decoded. The following check indicates whether a capability is malformed. If enableL8=1 the L8 bit is available in the capability encoding format for extra precision when EF=1.

malformedMSB =  (E == CAP_MAX_E     && B         != 0)
             || (E == CAP_MAX_E - 1 && B[MW - 1] != 0)
malformedLSB =  (E  < 0) || (E == 0 && enableL8)
malformed    =  !EF && (malformedMSB || malformedLSB)

Capabilities with malformed bounds:

  1. Return both base and top bounds as zero, which affects instructions like YBASER.

  2. Cause certain manipulation instructions like YADDI to always set the capability tag of the result to zero.

3.2. Representable Range Check

The concept of the representability check was introduced in Section 2.3.7.

The definition of the check is:

  • A source capability with address a, metadata m that decodes to give the bounds b and t.

  • A derived capability with arbitrary address a' with the same metadata m that decodes to give the bounds b' and t'.

The address a' is within the source capability’s representable range if b == b' && t == t'.

If the address a' is outside the representable range, then the derived capability has the capability tag set to zero.

3.2.1. Practical Information

An artifact of the bounds encoding is that if the new address causes t != t', then it is also the case that b != b'.

The inverse is also true, if b != b' then t != t'.

Therefore, for representable range checking, it is acceptable to either check t == t' or b == b'.

The top and bottom capability bounds are formed of two or three sections:

  • Upper bits from the address

    • This is only if the other sections do not fill the available bits (E + MW < XLEN)

  • Middle bits from T and B decoded from the metadata

  • Lower bits are set to zero

    • This is only if there is an internal exponent (EF=0)

Table 28. Composition of the decoded top address bound
Configuration Upper Section (if E + MW < XLEN) Middle Section Lower Section

EF=0

address[XLEN-1:E + MW] + ct

T[MW - 1:0]

{E{1’b0}}

EF=1, i.e., E=0

address[XLEN-1:MW] + ct

T[MW - 1:0]

The representable range defines the range of addresses which do not corrupt the bounds encoding. The encoding was first introduced in Section 3.1, and is repeated in a different form in Table 28 to aid this description.

For the address to be valid for the current bounds encoding, the value in the Upper Section of Table 28 must not change as this will change the meaning of the bounds. This is because T, B and E will be unchanged for the source and destination capabilities. Therefore, the Middle and Lower sections of the bounds calculation are also unchanged for source and destination capabilities.

When E > CAP_MAX_E - 2, the calculation of the top bound is entirely derived from T and E which will be identical for both the source and destination capabilities, thus guaranteeing that t == t'. Likewise, with such values of E, the base bound is entirely derived from B and E and therefore b == b'.

The calculation of the MSB of the top bound may be inverted as specified in Section 3.1.6.3. Assuming (E < (CAP_MAX_E - 1)), the truth-table for this inversion is as follows:

Table 29. Top bound MSB inversion truth table
input_t[XLEN:XLEN-1] b[XLEN-1] output_t[XLEN:XLEN-1]

00

0

00

01

0

01

10

0

00

11

0

01

00

1

10

01

1

01

10

1

10

11

1

01

Inspection of Table 29 shows that output_t[XLEN] does not depend on input_t[XLEN] as:

  • output_t[XLEN] = {input_t[XLEN-1], b[XLEN-1]} == 2’b01.

This leads to the conclusions:

  • If t[XLEN-1] == t'[XLEN-1] and b[XLEN-1] == b'[XLEN-1], then it is guaranteed t[XLEN] == t'[XLEN].

  • If t[XLEN-1] != t'[XLEN-1] or b[XLEN-1] != b'[XLEN-1], then the representable check will fail regardless of checking t[XLEN] == t'[XLEN].

Therefore, for the purpose of representable range checking, it is not required to check that t[XLEN]==t'[XLEN].

Given that t[XLEN] is not part of the representable range check:

  • when E == CAP_MAX_E - 2, t[XLEN-1:E] == T[MW-1:0] and b[XLEN-1:E] == B[MW-1:0].

Therefore, T and B are both derived from the capabilities metadata and are therefore constant. Which means that in this case too, the representable range check always passes.

As a result:

  • If E > CAP_MAX_E - 3, then the representability check always passes, even though the bounds are only infinite if E = CAP_MAX_E

This gives a range of s=2E+MW, as shown in Figure 2.

The gap between the object bounds and the bound of the representable range is always guaranteed to be at least 1/4 of s. This is represented by R = B - 2MW-2 in Section 3.1. This gives useful guarantees, such that if an executed instruction is in pc bounds, then it is also guaranteed that the next linear instruction is representable.

3.3. Encoding of Special Capabilities

3.3.1. NULL Capability Encoding

The NULL capability is represented with 0 in all fields. This implies that it has no permissions and its exponent E is CAP_MAX_E (52), so its bounds cover the entire address space such that the expanded base is 0 and top is 2XLEN.

Table 30. Field values of the NULL capability
Field Value Comment

Capability Tag

zero

Capability is not valid

SDP

zeros

Grants no permissions

AP

zeros

Grants no permissions

CT

zero

Unsealed

EF

zero

Internal exponent format

L81

zero

Top address reconstruction bit

T

zeros

Top address bits

TE

zeros

Exponent bits

B

zeros

Base address bits

BE

zeros

Exponent bits

Address

zeros

Capability address

Reserved

zeros

All reserved fields

1 Only present if enableL8=1.

Permissions added by extensions (such as those of Zylevels1) are presumed absent in NULL capabilities.

3.3.2. Infinite Capability Encoding

This encoding is for an Infinite capability value, which grants all permissions while its bounds also nominally cover the whole address space. This infinite capability is both a Root Executable and a Root Data capability.

Table 31. Field values of the Infinite capability
Field Value Comment

Capability Tag

one

Capability is valid

SDP

ones

Grants all permissions

AP

AP_MAX

Grants all permissions

CT

zero

Unsealed

EF

zero

Internal exponent format

L81

zero

Top address reconstruction bit

T

zeros

Top address bits

TE

zeros

Exponent bits

B

zeros

Base address bits

BE

zeros

Exponent bits

Address

any2

Capability address

Reserved

zeros

All reserved fields

1 Only present if enableL8=1

2If an infinite capability is used as a constant in either hardware or software, then the address field will typically be set to zero. If the address field is non-zero then it is still referred to as an infinite capability, and it still has the authority to authorize all memory accesses.

Permissions added by extensions (such as those of Zylevels1) are presumed present in Infinite capabilities, which may involve redefining the AP_MAX value.

4. Zca (RVY added and modified instructions)

Zcf (RV32) and Zcd (RV64) are incompatible with RVY.

Table 32. Zca (RVY added instructions) instruction extension
Mnemonic Function

C.LYSP

Load capability stack pointer relative, 16-bit encoding

C.SYSP

Store capability stack pointer relative, 16-bit encoding

C.LY

Load capability, 16-bit encoding

C.SY

Store capability, 16-bit encoding

4.1. RV64 / RV64Y RVC load/store mapping summary

Table 33. 16-bit load/store instruction mapping in RV64I
Encoding Supported Extensions

[15:13]

[1:0]

Zca

Zcd

Zcmp/ Zcmt

111

00

C.SD

N/A

N/A

011

00

C.LD

N/A

N/A

111

10

C.SDSP

N/A

N/A

011

10

C.LDSP

N/A

N/A

101

00

N/A

C.FSD

reserved

001

00

N/A

C.FLD

reserved

101

10

N/A

C.FSDSP

Zcmp/ Zcmt

001

10

N/A

C.FLDSP

reserved
Table 34. 16-bit load/store instruction mapping in RV64Y
Encoding Supported Extensions

[15:13]

[1:0]

Zca

111

00

C.SD

011

00

C.LD

111

10

C.SDSP

011

10

C.LDSP

101

00

C.SY

001

00

C.LY

101

10

C.SYSP

001

10

C.LYSP
Zcd, Zcmp and Zcmt are incompatible with RV64Y.

4.1.1. C.LY

see C.LYSP.

4.1.2. C.LYSP

Synopsis

Capability loads (C.LY, C.LYSP), 16-bit encodings

Mnemonics

c.ly rd', offset(rs1')
c.lysp rd', offset(sp)

Expansions

ly rd', offset(rs1')
ly rd', offset(sp)

Encoding
Diagram
Diagram
Description

For C.LY the authorizing capability is rs1.

For C.LYSP, the authorizing capability is sp.

The effective address of the memory access is obtained by adding the address of the authorizing capability to the zero-extended offset.

Load data from memory in the same manner as LY.

Prerequisites

Zca, RVY

Included in

Zca (RVY added instructions)

Operation (after expansion to 32-bit encodings)

See LY

4.1.3. C.SY

see C.SYSP.

4.1.4. C.SYSP

Synopsis

Capability stores (C.SY, C.SYSP), 16-bit encodings

Mnemonics

c.sy rs2', offset(rs1')
c.sysp rs2', offset(sp)

Expansions

sy rs2', offset(rs1')
sy rs2', offset(sp)

Encoding
Diagram
Diagram
Description

For C.SY the authorizing capability is rs1.

For C.SYSP, the authorizing capability is sp.

The store data is in rs2'.

The effective address of the memory access is obtained by adding the address of the authorizing capability to the zero-extended offset.

Store data to the memory in the same manner as SY.

Prerequisites

Zca, RVY

Included in

Zca (RVY added instructions)

Operation (after expansion to 32-bit encodings)

See SY

4.2. Zca (RVY modified behavior)

An RVY core which supports Zca also supports Zca (RVY modified behavior) which modifies the behavior of some instructions.

c.ymv is renamed from c.mv to avoid ambiguity in disassembly.
Table 35. Zca (RVY modified behavior) instruction extension
Mnemonic Function

C.ADDI16SP (RVY)

Stack pointer increment in blocks of 16, 16-bit encoding

C.ADDI4SPN (RVY)

Stack pointer increment in blocks of 4, 16-bit encoding

C.YMV

Capability register copy, 16-bit encoding

C.JALR (RVY)

Jump to capability register, and link to capability register, 16-bit encoding

C.JR (RVY)

Jump to capability register, 16-bit encoding

4.2.1. C.ADDI16SP (RVY)

Synopsis

Stack pointer increment in blocks of 16, 16-bit encoding

Mnemonic

c.addi16sp nzimm

Expansion

yaddi sp, sp, nzimm

Encoding
Diagram
Description

Add the non-zero sign-extended 6-bit immediate to the value in the stack pointer (sp=x2), where the immediate is scaled to represent multiples of 16 in the range (-512,496).

Set sp.tag=0 if sp is sealed.

Set rd.tag=0 if the resulting capability cannot be represented exactly.

Set rd.tag=0 if sp fails any integrity checks.

Prerequisites

Zca, RVY

Included in

Zca (RVY modified behavior)

Operation
execute(CADDI(sp, sp, sign_extend(nzimm)))

4.2.2. C.ADDI4SPN (RVY)

Synopsis

Stack pointer increment in blocks of 4, 16-bit encoding

Mnemonic

c.addi4spn rd', nzuimm

Expansion

yaddi rd', sp, nzuimm

Encoding
Diagram
Description

Copy sp to rd'. Add a zero-extended non-zero immediate, scaled by 4, to rd'.address.

Set rd'.tag=0 if sp is sealed.

Set rd'.tag=0 if the resulting capability cannot be represented exactly.

Set rd'.tag=0 if sp 's bounds are malformed, or if any of the reserved fields are set.

Prerequisites

Zca, RVY

Included in

Zca (RVY modified behavior)

Operation
let cd = creg2reg_idx(cdc) in
execute(CADDI(cd, sp, zero_extend(nzuimm)))

4.2.3. C.YMV

Synopsis

Capability register copy, 16-bit encoding

Mnemonic

c.ymv rd, rs2

Expansion

ymv rd, rs2

Suggested assembly syntax

ymv rd, rs2

c.mv is remapped to copy YLEN-bit registers for RVY. The mnemonic is changed to avoid ambiguity about whether the copy is XLEN or YLEN bits.
Encoding
Diagram
Description

Capability register rd is replaced with the contents of rs2.

This instruction can propagate valid capabilities which fail integrity checks.

Prerequisites

Zca, RVY

Included in

Zca (RVY modified behavior)

Operation (after expansion to 32-bit encoding)

See YMV

4.2.4. C.JR (RVY)

Synopsis

Jump to capability register, 16-bit encoding

Mnemonic

c.jr rs1

Expansion

jalr x0, 0(rs1)

Encoding
Diagram
Description

See JALR (RVY) for execution of the expanded instruction as shown above. Note that the offset is zero in the expansion.

Prerequisites

Zca, RVY

Included in

Zca (RVY modified behavior)

Operation (after expansion to 32-bit encodings)

See JALR (RVY)

4.2.5. C.JALR (RVY)

Synopsis

Jump to capability register, and link to capability register, 16-bit encoding

Mnemonic

c.jalr x1, rs1

Expansion

jalr x1, 0(rs1)

Encoding
Diagram
Description

See JALR (RVY) for execution of the expanded instruction as shown above. Note that the offset is zero in the expansion.

Exceptions

See JALR (RVY)

Prerequisites

Zca, RVY

Included in

Zca (RVY modified behavior)

Operation (after expansion to 32-bit encodings)

See JALR (RVY)

5. Zalrsc (RVY added instructions)

Specifying RVY and Zalrsc adds atomic capability load and store instructions.

Table 36. Zalrsc (RVY added instructions) instruction extension
Mnemonic Function

LR.Y

Load Reserved capability

SC.Y

Store Conditional capability

5.1. LR.Y

Synopsis

Load Reserved capability

Mnemonic

lr.y rd, 0(rs1)

Encoding
Diagram
The instruction encoding is not final and is highly likely to change prior to v1.0
Any instance of this instruction with rs1=x0 will raise an exception, as x0 is defined to always hold a NULL capability. As such, the encodings with rs1=x0 are RESERVED for use by future extensions.
Description

Calculate the effective address of the memory access by adding rs1.address to the sign-extended 12-bit offset.

Authorize the memory access with the capability in rs1.

Load a naturally aligned YLEN-bit data value from memory.

If the PMA is CHERI Capability Tag then load the associated capability tag, otherwise set the capability tag to zero.

Set the reservation as for LR.W/D.

Use the YLEN-bit data and the capability tag to determine the value of rd as specified by the LY instruction.

This instruction can propagate valid capabilities which fail integrity checks.

Exceptions

Load access fault exception when the effective address is not aligned to YLEN/8.

All misaligned load reservations cause a load address misaligned exception to allow software emulation (if the Zam extension is supported), otherwise they take a load access fault exception.

Exceptions occur when the authorizing capability fails one of the checks listed below:

Kind Reason

CHERI Load Access Fault

Authorizing capability tag is set to 0.

CHERI Load Access Fault

Authorizing capability is sealed.

CHERI Load Access Fault

Authorizing capability does not grant the necessary permissions. R-permission is required.

CHERI Load Access Fault

At least one byte accessed is outside the authorizing capability bounds, or the bounds could not be decoded.

CHERI Load Access Fault

Authorizing capability failed any integrity check.
Prerequisites

RVY, and A or Zalrsc

Included in

Zalrsc (RVY added instructions)

Operation
TBD

5.2. SC.Y

Synopsis

Store Conditional capability

Any instance of this instruction with rs1=x0 will raise an exception, as x0 is defined to always hold a NULL capability. As such, the encodings with rs1=x0 are RESERVED for use by future extensions.
Mnemonic

sc.y rd, rs2, 0(rs1)

Encoding
Diagram
The instruction encoding is not final and is highly likely to change prior to v1.0
Description

Calculate the effective address of the memory access by adding rs1.address to the sign-extended 12-bit offset.

Authorize the memory access with the capability in rs1.

Conditionally store, following the same rules as SC.W, a naturally aligned YLEN-bit data value in rs2 to memory and the associated capability tag in rs2.

Set rd to 1 for success or 0 for failure.

The written capability tag may be cleared following the same modification rules as SY.

This instruction can propagate valid capabilities which fail integrity checks.

Exceptions

Store/AMO access fault exception when the effective address is not aligned to YLEN/8.

Store/AMO access fault if the stored capability tag is set to one and the PMA is CHERI Capability Tag Store Fault. If the address is not naturally aligned raise a Store/AMO address misaligned exception or a Store/AMO access fault exception. See "Zaamo" for details on which one is raised.

+ Exceptions occur when the authorizing capability fails one of the checks listed below:

Kind Reason

CHERI Store/AMO Access Fault

Authorizing capability tag is set to 0.

CHERI Store/AMO Access Fault

Authorizing capability is sealed.

CHERI Store/AMO Access Fault

Authorizing capability does not grant the necessary permissions. W-permission is required.

CHERI Store/AMO Access Fault

At least one byte accessed is outside the authorizing capability bounds, or the bounds could not be decoded.

CHERI Store/AMO Access Fault

Authorizing capability failed any integrity check.

+ :!store_cond:

Prerequisites

RVY, and A or Zalrsc

Included in

Zalrsc (RVY added instructions)

Operation
TBD

6. Zaamo (RVY added instructions)

Specifying RVY and Zaamo gives Zaamo (RVY added instructions) functionality, which adds atomic capability swap.

Table 37. Zaamo (RVY added instructions) instruction extension
Mnemonic Function

AMOSWAP.Y

Atomic swap of capabilities

6.1. AMOSWAP.Y

Synopsis

Atomic swap of capabilities

Mnemonic

amoswap.y rd, rs2, 0(rs1)

Encoding
Diagram
The instruction encoding is not final and is highly likely to change prior to v1.0
Any instance of this instruction with rs1=x0 will raise an exception, as x0 is defined to always hold a NULL capability. As such, the encodings with rs1=x0 are RESERVED for use by future extensions.
Description

Atomic swap of capability type, authorized by the capability in rs1.

The operation is equivalent to an atomically executed sequence of:

ly rd, 0(rs1)

sy rs2, 0(rs1)

With the proviso that rd is only updated if no exceptions are raised and that if rd==rs1 or rd==rs2 the SY is unaffected.

Permissions

Requires R-permission and W-permission in the authorizing capability.

Requires all bytes of the access to be in capability bounds.

Exceptions

If the address is not naturally aligned raise a Store/AMO address misaligned exception or a Store/AMO access fault exception. See "Zaamo" for details on which one is raised.

Exceptions occur when the authorizing capability fails one of the checks listed below:

Kind Reason

CHERI Store/AMO Access Fault

Authorizing capability tag is set to 0.

CHERI Store/AMO Access Fault

Authorizing capability is sealed.

CHERI Store/AMO Access Fault

Authorizing capability does not grant the necessary permissions. W-permission and R-permission are both required.

CHERI Store/AMO Access Fault

At least one byte accessed is outside the authorizing capability bounds, or the bounds could not be decoded.

Store/AMO address misaligned

Address is not naturally aligned
Prerequisites

RVY, and Zaamo

Included in

Zaamo (RVY added instructions)

Operation

TODO

7. Zba (RVY added instructions)

Specifying RVY and Zba gives Zba (RVY added instructions) functionality, which adds more instructions.

Table 38. Zba (RVY added instructions) instruction extension
Mnemonic Function

YSH1ADD

shift and add, representability check

YSH2ADD

shift and add, representability check

YSH3ADD

shift and add, representability check

YSH4ADD (RV64Y)

shift and add, representability check

YSH1ADD.UW (RV64Y)

shift and add unsigned word, representability check

YSH2ADD.UW (RV64Y)

shift and add unsigned word, representability check

YSH3ADD.UW (RV64Y)

shift and add unsigned word, representability check

YSH4ADD.UW (RV64Y)

shift and add unsigned word, representability check
There is currently no RVY equivalent for add.uw. A future extension may add such an instruction.

7.1. YSH1ADD

See YSH4ADD (RV64Y).

7.2. YSH2ADD

See YSH4ADD (RV64Y).

7.3. YSH3ADD

See YSH4ADD (RV64Y).

7.4. YSH4ADD (RV64Y)

Synopsis

Shift by n and add for address generation (YSH1ADD, YSH2ADD, YSH3ADD, YSH4ADD)

Mnemonics (RVY)

ysh1add rd, rs1, rs2
ysh2add rd, rs1, rs2
ysh3add rd, rs1, rs2

Mnemonics (RV64Y)

ysh4add rd, rs1, rs2

Encoding
Diagram
The instruction encoding is not final and is highly likely to change prior to v1.0
Description

Copy the capability in rs2 to rd.

Increment rd.address by rs1 shifted left by n bit positions.

Set rd.tag=0 if rs2 is sealed.

Set rd.tag=0 if the resulting capability cannot be represented exactly.

Set rd.tag=0 if rs2 fails any integrity checks.

Included in

Zba (RVY added instructions)

Operation
let rs1_val = X(rs1);
let cs2_val = C(cs2);
let shamt : range(0,3) = match op {
  RISCV_SH1ADD => 1,
  RISCV_SH2ADD => 2,
  RISCV_SH3ADD => 3,
};
let result = incCapAddrChecked(cs2_val, rs1_val << shamt);
C(cd) = result;
RETIRE_SUCCESS

7.5. YSH1ADD.UW (RV64Y)

See YSH4ADD.UW (RV64Y).

7.6. YSH2ADD.UW (RV64Y)

See YSH4ADD.UW (RV64Y).

7.7. YSH3ADD.UW (RV64Y)

See YSH4ADD.UW (RV64Y).

7.8. YSH4ADD.UW (RV64Y)

Synopsis

Shift by n and add unsigned words for address generation (YSH1ADD.UW, YSH2ADD.UW, YSH3ADD.UW, YSH4ADD.UW)

Mnemonics (RV64Y)

ysh1add.uw rd, rs1, rs2
ysh2add.uw rd, rs1, rs2
ysh3add.uw rd, rs1, rs2
ysh4add.uw rd, rs1, rs2

Encoding
Diagram
The instruction encoding is not final and is highly likely to change prior to v1.0
Description

Copy the capability in rs2 to rd.

Increment rd.address by the unsigned word rs1 shifted left by n bit positions.

Set rd.tag=0 if rs2 is sealed.

Set rd.tag=0 if the resulting capability cannot be represented exactly.

Set rd.tag=0 if rs2 fails any integrity checks.

Included in

Zba (RVY added instructions)

Operation
let rs1_val = X(rs1);
let cs2_val = C(cs2);
let shamt : range(0,3) = match op {
  RISCV_ADDUW    => 0,
  RISCV_SH1ADDUW => 1,
  RISCV_SH2ADDUW => 2,
  RISCV_SH3ADDUW => 3,
};
let result = incCapAddrChecked(cs2_val, zero_extend(rs1_val[31..0]) << shamt);
C(cd) = result;
RETIRE_SUCCESS

8. Zicbom (RVY modified behavior)

Specifying RVY and Zicbom gives Zicbom (RVY modified behavior) functionality, which extends the checking.

Table 39. Zicbom (RVY modified behavior) instruction extension
Mnemonic Function

CBO.INVAL (RVY)

Cache block invalidate (implemented as clean)

CBO.CLEAN (RVY)

Cache block clean

CBO.FLUSH (RVY)

Cache block flush

8.1. CBO.CLEAN (RVY)

Synopsis

Perform a clean operation on a cache block

Mnemonic

cbo.clean 0(rs1)

Encoding
Diagram
Description

A CBO.CLEAN instruction performs a clean operation on the cache block whose effective address is the base address specified in rs1. The authorizing capability for this operation is rs1.

Exceptions
Kind Reason

CHERI Store/AMO Access Fault

Authorizing capability tag is set to 0.

CHERI Store/AMO Access Fault

Authorizing capability is sealed.

CHERI Store/AMO Access Fault

Authorizing capability does not grant the necessary permissions. W-permission and R-permission are both required.

CHERI Store/AMO Access Fault

None of the bytes accessed are within the bounds, or the bounds could not be decoded.

CHERI Store/AMO Access Fault

Authorizing capability failed any integrity check.
Prerequisites

Zicbom, RVY

Included in

Zicbom (RVY modified behavior)

Operation
TBD

8.2. CBO.FLUSH (RVY)

Synopsis

Perform a flush operation on a cache block

Mnemonic

cbo.flush 0(rs1)

Encoding
Diagram
Description

A CBO.FLUSH instruction performs a flush operation on the cache block whose effective address is the base address specified in rs1. The authorizing capability for this operation is rs1.

Exceptions
Kind Reason

CHERI Store/AMO Access Fault

Authorizing capability tag is set to 0.

CHERI Store/AMO Access Fault

Authorizing capability is sealed.

CHERI Store/AMO Access Fault

Authorizing capability does not grant the necessary permissions. W-permission and R-permission are both required.

CHERI Store/AMO Access Fault

None of the bytes accessed are within the bounds, or the bounds could not be decoded.

CHERI Store/AMO Access Fault

Authorizing capability failed any integrity check.
Prerequisites

Zicbom, RVY

Included in

Zicbom (RVY modified behavior)

Operation
TBD

8.3. CBO.INVAL (RVY)

Synopsis

Perform an invalidate operation on a cache block

Mnemonic

cbo.inval 0(rs1)

Encoding
Diagram
Description

A CBO.INVAL instruction performs an invalidate operation on the cache block whose effective address is the base address specified in rs1. The authorizing capability for this instruction is rs1.

Exceptions
Kind Reason

Illegal instruction

pc does not grant ASR-permission.

CHERI Store/AMO Access Fault

Authorizing capability tag is set to 0.

CHERI Store/AMO Access Fault

Authorizing capability is sealed.

CHERI Store/AMO Access Fault

Authorizing capability does not grant the necessary permissions. W-permission, R-permission are both required.

CHERI Store/AMO Access Fault

At least one byte accessed is outside the authorizing capability bounds, or the bounds could not be decoded.

CHERI Store/AMO Access Fault

Authorizing capability failed any integrity check.

CSR state controls whether CBO.INVAL performs cache block flushes instead of invalidations for less privileged modes.

Invalidating a cache block can re-expose capabilities previously stored to it after the most recent flush, not just secret values. As such, CBO.INVAL has stricter checks on its use than CBO.FLUSH, and should only be made available to, and used by, sufficiently-trusted software.

Untrusted software should use CBO.FLUSH instead as a minimum, and a sensible implementation choice for CHERI systems is to always execute CBO.INVAL as CBO.FLUSH.
Prerequisites

Zicbom, RVY

Included in

Zicbom (RVY modified behavior)

Operation
TBD

9. Zicboz (RVY modified behavior)

Specifying RVY and Zicboz gives Zicboz (RVY modified behavior) functionality, which extends the checking.

Table 40. Zicboz (RVY modified behavior) instruction extension
Mnemonic Function

CBO.ZERO (RVY)

Cache block zero

9.1. CBO.ZERO (RVY)

Synopsis

Store zeros to the full set of bytes corresponding to a cache block

Mnemonic

cbo.zero 0(rs1)

Encoding
Diagram
Description

A cbo.zero instruction performs stores of zeros to the full set of bytes corresponding to the cache block whose effective address is the base address specified in rs1. An implementation may or may not update the entire set of bytes atomically although each individual write must atomically set the capability tag of the corresponding aligned YLEN-bit location to zero. The authorizing capability for this instruction is rs1.

Exceptions

Store/AMO access fault exception when the effective address is not aligned to YLEN/8.

Store/AMO access fault if the stored capability tag is set to one and the PMA is CHERI Capability Tag Store Fault. If the address is not naturally aligned raise a Store/AMO address misaligned exception or a Store/AMO access fault exception. See "Zaamo" for details on which one is raised.

+ Exceptions occur when the authorizing capability fails one of the checks listed below:

Kind Reason

CHERI Store/AMO Access Fault

Authorizing capability tag is set to 0.

CHERI Store/AMO Access Fault

Authorizing capability is sealed.

CHERI Store/AMO Access Fault

Authorizing capability does not grant the necessary permissions. W-permission is required.

CHERI Store/AMO Access Fault

At least one byte accessed is outside the authorizing capability bounds, or the bounds could not be decoded.

CHERI Store/AMO Access Fault

Authorizing capability failed any integrity check.

+ :!store_cond:

Prerequisites

Zicboz, RVY

Included in

Zicboz (RVY modified behavior)

Operation
TBD

10. Zicbop (RVY modified behavior)

Specifying RVY and Zicbop gives Zicbop (RVY modified behavior) functionality, which extends the checking.

Table 41. Zicbop (RVY modified behavior) instruction extension
Mnemonic Function

PREFETCH.I (RVY)

Cache Block Prefetch for Instruction Fetch

PREFETCH.R (RVY)

Cache Block Prefetch for Data Read

PREFETCH.W (RVY)

Cache block Prefetch for Data Write

10.1. PREFETCH.I (RVY)

Synopsis

Provide a HINT to hardware that a cache block is likely to be accessed by an instruction fetch in the near future

Mnemonic

prefetch.i offset(rs1)

Encoding
Diagram
Description

A PREFETCH.I instruction indicates to hardware that the cache block whose effective address is the sum of the base address specified in rs1 and the sign-extended offset encoded in imm[11:0], where imm[4:0] equals 0b00000, is likely to be accessed by an instruction fetch in the near future. The encoding is only valid if imm[4:0]=0. The authorizing capability for this operation is rs1. This instruction does not throw any exceptions. However, following the rules from Chapter 2, this instruction does not perform a prefetch if it is not authorized by rs1.

PREFETCH.I does not perform a memory access if one or more of the following conditions of the authorizing capability are met:

  • The capability tag is not set

  • The sealed bit is set

  • No bytes of the cache line requested is in bounds

  • X-permission is not set

  • Any integrity check fails

If the checks above pass, an implementation may opt to cache a copy of the cache block in a cache accessed by an instruction fetch in order to improve memory access latency, but this behavior is not required.
Prerequisites

Zicbop, RVY

Included in

Zicbop (RVY modified behavior)

Operation
TODO

10.2. PREFETCH.R (RVY)

Synopsis

Provide a HINT to hardware that a cache block is likely to be accessed by a data read in the near future

Mnemonic

prefetch.r offset(rs1)

Encoding
Diagram
Description

A PREFETCH.R instruction indicates to hardware that the cache block whose effective address is the sum of the base address specified in rs1 and the sign-extended offset encoded in imm[11:0], where imm[4:0] equals 0b00000, is likely to be accessed by a data read (i.e., load) in the near future. The encoding is only valid if imm[4:0]=0. The authorizing capability for this operation is rs1. This instruction does not throw any exceptions. However, following the rules from Chapter 2, this instruction does not perform a prefetch if it is not authorized by rs1.

PREFETCH.R does not perform a memory access if one or more of the following conditions of the authorizing capability are met:

  • The capability tag is not set

  • The sealed bit is set

  • No bytes of the cache line requested is in bounds

  • R-permission is not set

  • Any integrity check fails

If the checks above pass, an implementation may opt to cache a copy of the cache block in a cache accessed by a data read in order to improve memory access latency, but this behavior is not required.
Prerequisites

Zicbop, RVY

Included in

Zicbop (RVY modified behavior)

Operation
TODO

10.3. PREFETCH.W (RVY)

Synopsis

Provide a HINT to hardware that a cache block is likely to be accessed by a data write in the near future

Mnemonic

prefetch.w offset(rs1)

Encoding
Diagram
Description

A PREFETCH.W instruction indicates to hardware that the cache block whose effective address is the sum of the base address specified in rs1 and the sign-extended offset encoded in imm[11:0], where imm[4:0] equals 0b00000, is likely to be accessed by a data write (i.e., store) in the near future. The encoding is only valid if imm[4:0]=0. The authorizing capability for this operation is rs1. This instruction does not throw any exceptions. However, following the rules from Chapter 2, this instruction does not perform a prefetch if it is not authorized by rs1.

PREFETCH.W does not perform a memory access if one or more of the following conditions of the authorizing capability are met:

  • The capability tag is not set

  • The sealed bit is set

  • No bytes of the cache line requested is in bounds

  • W-permission is not set

  • Any integrity check fails

If the checks above pass, an implementation may opt to cache a copy of the cache block in a cache accessed by a data write in order to improve memory access latency, but this behavior is not required.
Prerequisites

Zicbop, RVY

Included in

Zicbop (RVY modified behavior)

Operation
TODO

11. Vector "V" Extension (RVY)

The Vector extension for RVY does not support capability tags in the vector registers. Consequently, vector stores will always clear capability tags in memory.

A future extension may allow capability tags to be stored in vector registers. Until that time, vector load and store instructions must not be used to implement generic memory copying in software, such as the memcpy() standard C library function, that requires capabilities to be preserved, because the vector registers do not hold capabilities, so the capability tags of any copied capabilities will be set to 0 in the destination memory.

Under RVY, vector loads and stores follow the standard rules for active elements:

  • Only active elements are subject to CHERI exception checks.

  • CHERI exceptions are only raised on fault-only-first loads if element 0 is both active and fails any exception checks.

Additionally, the standard RVY rule that all loads and stores where the base register is x0 are reserved applies to all vector memory access instructions.

The approach of using indexed loads with the base register set to the value zero and XLEN-wide offsets does not work well with CHERI as the authorizing capability must cover all of memory. If the authorizing capability is specified as x0 then the instruction encoding is reserved.

12. "Zylevels1" Extension for CHERI 2-Level Information Flow Control

Zylevels1 adds an Information Flow Control (IFC) mechanism to CHERI.

It adds:

  • One new "level" metadata bit (the GL(obal) Flag) to classify capabilities as either global or local.

  • Checks to memory operations to control how capabilities can be shared.

  • Two new level-related permissions, Store Local (SL) and Load Global (LG).

The behavior is an extension as it may be replaced by a future scheme with more than two levels.

12.1. Added Metadata Bits (Permissions and Flags)

Bit Type Comment

GL(obal) Flag

Information flow label

Used to classify capabilities as global or local.

SL-permission

Data memory permission

Used to filter the capability tags of stored capabilities.

LG-permission

Data memory permission

Used to filter the permissions of loaded capabilities.

12.1.1. The Capability Global (GL) Flag

The GL(obal) Flag is a metadata bit used to track and control the propagation of capabilities. It works in conjunction with the LG-permission and SL-permission permissions to enforce information flow boundaries.

Global Capability

A global capability has GL(obal) Flag=1.

Local Capability

A local capability has GL(obal) Flag=0.

Like permissions, the GL(obal) Flag can be cleared when creating a new capability from an existing one, but it can never be set unless derived from a superset capability which has the GL(obal) Flag set.

12.1.2. Store Local Permission (SL)

Store Local Permission (SL) controls whether a local capability can be stored to memory.

If the authorizing capability lacks SL-permission then the stored capability has its capability tag set to zero.

SL-permission is dependent on both C-permission and W-permission as shown in Table 42.

12.1.3. Load Global Permission (LG)

Load Global Permission (LG) controls whether global capabilities can be loaded from memory. Loading a capability using an authorizing capability that lacks LG-permission clears the GL(obal) Flag bit of the loaded capability, making it into a local capability. If the loaded capability is unsealed, its LG-permission is also cleared.

This behavior is similar to how LM-permission affects loaded capabilities (but note the difference in interaction with seals).

LG-permission is dependent on both C-permission and R-permission as shown in Table 42.

12.2. Interaction with Root Capabilities

The Root capabilities definitions are extended:

12.3. Interaction with YPERMC and YPERMR

Diagram
Figure 11. Extended capability permissions bit field (see Figure 6)

The GL(obal) Flag, SL-permission, and LG-permission fields are mapped into the capability permissions bitfield as used by YPERMC and YPERMR, as shown in Figure 11.

The GL(obal) Flag is treated as a permission by YPERMC and YPERMR.

Unlike actual permissions it is not subject to the sealed check, and so YPERMC can produce a new capability with its GL(obal) Flag cleared, even if the source capability is sealed.

12.3.1. Additional YPERMC permission rules

The following YPERMC permission rules are added:

Table 42. YPERMC permission rules for Zylevels1.
YPERMC Rule Permission Valid only if

Zylevels1-1

LG-permission

C-permission and R-permission

Zylevels1-2

SL-permission

C-permission and W-permission

12.4. Interaction with LY

Zylevels1 introduces two new constraints on capabilities loaded from memory. These constraints apply to LY and all other instructions that can load valid capabilities.

The permissions of the loaded capability may be reduced:

Table 43. LG-permission effects for loading capabilities
Auth cap field Data cap field

R

C

LG

Tag

Sealed

Action

1

1

0

1

Yes

Load data capability with its GL(obal) Flag cleared

No

Load data capability with both its GL(obal) Flag and LG-permission cleared

All other cases

Load data capability with both its GL(obal) Flag and LG-permission unmodified

12.5. Interaction with SY

Zylevels1 adds a constraint to SY and all instructions which can store valid capabilities.

The capability is stored with a capability tag set to one only if:

Otherwise, the capability is stored with the capability tag set to zero.

While a lack of LG-permission restricts flow by reducing permissions (labeling the loaded capability as local and stripping permissions), SL-permission restricts flow by invalidating capabilities (by setting the capability tag to zero).
Table 44. SL-permission effects for stored capabilities
Auth cap field Data cap field

W

C

SL

GL

Notes

1

1

1

X

Store data capability unmodified

0

1

Store data capability unmodified

0

Store data capability with capability tag set to zero

12.6. Interaction with Capability Subsets

The definition of capability subset has an additional rule:

This affects the behavior of YSS, YBLD and YSUNSEAL.

12.7. Rationale and Software Use Cases

The primary goal of Zylevels1 is to prevent storing pointers to specific memory regions in order to enforce security invariants. There are many uses of this mechanism, but two examples are listed below.

12.7.1. Restricting Capability Flow Between Regions or Domains (e.g., Stack-to-Heap or Kernel-to-User)

Software can use these features to prevent pointers from leaking between different memory regions or privilege domains that have different security properties. The general pattern is to mark sensitive capabilities as local (by clearing their GL(obal) Flag) and ensuring that destination memory regions are accessed through authorities that lack the SL-permission.

  • Intra-program flow (e.g., Stack-to-Heap): To prevent pointers to short-lived memory (like the stack) from leaking into long-lived memory (like the heap), software marks all stack pointers as local. If the authorities for the heap lack the SL-permission, any attempt to store a stack pointer into the heap results in the hardware invalidating the stored capability. This ensures stack pointers are strictly confined to registers and the stack itself.

  • Cross-domain flow (e.g., Kernel-to-User): These features also apply across privilege boundaries. For example, a kernel can ensure that the capability authorizing a write to userspace (e.g., for a copy_to_user operation) lacks the SL-permission. Even if a kernel bug causes it to accidentally attempt to copy sensitive, local kernel capabilities into the user’s buffer, the hardware will invalidate those capabilities. This prevents unintentional leaks of usable capabilities to a less-privileged domain.

12.7.2. Enforcing Shallow vs. Deep No-Capture Guarantees

When passing a capability to a less-trusted function, a caller often wants to enforce "no-capture" guarantees, ensuring the callee cannot hold onto the capability after returning. Zylevels1 provides two levels of protection against capability capture.

  • Shallow no-capture (via GL(obal) Flag and SL-permission): The caller makes the argument capability local by clearing its GL(obal) Flag. Because the callee’s global variables and persistent data structures are accessed through capabilities that lack the SL-permission, the callee cannot store the local argument capability into them. This enforces a shallow guarantee: the callee cannot capture the argument itself, but it might still load a global capability through the argument and capture that.

  • Deep no-capture (via LG-permission): To solve the shallow limitation, the caller also removes the LG-permission from the argument capability. If the callee loads any capability through this argument, the hardware automatically removes the loaded capability’s GL(obal) Flag and LG-permission. This enforces a deep no-capture guarantee: the callee is prevented from capturing the argument and anything reachable from it.

13. "Zyhybrid" Extension for CHERI Pointer Modes

Zyhybrid is an optional extension to RVY which adds the ability to execute standard RISC-V (RV32[IE]/RV64[IE]) code on an RVY hart, while still allowing access to RVY instructions.

The ability to choose between these two behaviors is referred to as switching between CHERI Pointer Modes. This extension introduces a new bit in the capability metadata, the P-bit, to control these modes. When installed in pc, the P-bit determines the current execution mode.

The P-bit determines whether pointers are interpreted as standard RISC-V Integer Addresses or as CHERI Capabilities.

Zyhybrid adds the instructions in Table 45 to query and update the current mode, and switch modes.

Table 45. Zyhybrid instruction extension
Mnemonic Function

YMODEW

Set capability execution mode

YMODER

Read capability mode

YMODESWY

Switch execution to Capability Pointer Mode

YMODESWI

Switch execution to Integral Pointer Mode

This extension also adds the default data capability (ddc), an unprivileged CSR used to authorize data memory accesses when in Integral Pointer Mode.

Together with pc, ddc enables confining execution to a compartment (or sandbox) where all memory accesses are bounded. These compartments provide binary compatibility for existing RV32[IE]/RV64[IE] code. Provided that the execution environment has set up ddc and pc appropriately, non-CHERI-aware programs will execute unmodified within these bounds.

For detailed examples of how this extension facilitates co-existence and sandboxing, see the appendix on Appendix B.

RVY implementations which support Zyhybrid are typically referred to as CHERI Hybrid. Implementations which do not support Zyhybrid are referred to as CHERI Purecap and always operate in Capability Pointer Mode.

Because the base architecture remains RVY, the custom encoding space re-allocated to CHERI remains re-allocated.

If an implementation needs to be able access custom encodings in spaces which have been reallocated, then privileged software should disable CHERI entirely.

13.1. CHERI Pointer Modes

The two CHERI Pointer Modes are:

Capability Pointer Mode

All pointer operands to instructions are interpreted as capabilities. Instruction semantics are as defined in the RVY base ISA.

Integral Pointer Mode

All pointer operands to instructions are interpreted as for the RV32[IE]/RV64[IE] base ISA, i.e., as plain memory addresses without metadata. While the base ISA remains RVY this mode provides backwards compatibility with RV32[IE]/RV64[IE]. Operands are still checked against a capability, but instead of using the metadata from the full YLEN register, the bounds and permissions are obtained from ddc (for data accesses) or pc (for instruction accesses).

ARC-NOTE: Naming of these modes is still an open ARC comment. The goal is to differentiate interpretation of pointers as plain integer addresses vs. full capabilities with metadata. The current placeholder names Integral Pointer Mode and Capability Pointer Mode match compiler terminology (integral vs. non-integral pointers).

The following options have been considered:

Compatibility Mode Name Purecap Mode Name Notes

Integral Pointer Mode

Capability Pointer Mode

Matches compiler terminology

Non-CHERI Mode

CHERI Mode

Using a negative to describe a mode is awkward

Non-CHERI Pointer Mode

CHERI Pointer Mode

Same again just with "Pointer"

Address Mode

Capability Mode

Address Pointer Mode

Capability Pointer Mode

"Address Pointer" is a bit odd

(Non-CHERI) Address Mode

(CHERI) Capability Mode

Older rejected proposal, parens are a bit awkward

Integer Mode

Capability Mode

CHERIv9 naming scheme.

RVI Mode

RVY Mode

Omits RVE, should it be RVI/RVE mode?

RVI Pointer Mode

RVY Pointer Mode

Again omits RVE

One past suggestion was "Pointer Mode" vs "Capability Mode". However, this is also confusing since capabilities are also pointers (just not "classic" pointers that are a number). In this naming scheme we want to show the difference in how hardware interprets pointers-type operands: as binary number (i.e., an address), or as a capability pointer (address+metadata).

If RVC encodings are supported, load/store encodings will revert to their non-CHERI encodings, such as C.LY reverting to C.FSD for RV64/RV64Y. This behavior is summarized in Table 33 and Table 34.

Instructions which are modified on an RVY architecture (see Table 22) revert to their standard RV32[IE]/RV64[IE] behavior.

All RVY instructions and CSRs are available in addition to all standard RV32[IE]/RV64[IE] instructions and CSRs.

Non-standard extensions with encodings in remapped custom space are not supported.

The authorizing capability for data access is ddc (as opposed to rs1). That is, all data memory accesses are implicitly authorized by ddc with only the memory address sourced from rs1.

ddc is also used to authorize memory instructions added by RVY (such as LY and SY).

All instruction fetches are authorized by pc.

Prefetches are also checked against the appropriate CSR (ddc for PREFETCH.W (RVY) and PREFETCH.R (RVY), pc for PREFETCH.I (RVY)) and execute as nops if authorization fails.

Authorization of prefetches is useful to mitigate against side channels, e.g., preventing userspace from fetching kernel memory into a cache.

Reads to extended CSRs are XLEN-wide, and writes use the semantics of the YADDRW instruction (see rvy_zicsr_mod_insn_table).

The CHERI Pointer Mode is key in providing backwards compatibility with the base RV32[IE]/RV64[IE] ISAs.

The CHERI Pointer Mode is always Capability Pointer Mode on implementations that support RVY, but not Zyhybrid.

Software is Purecap if it uses CHERI capabilities for all memory accesses (including loads, stores, and instruction fetches). Purecap software requires RVY. Software is Hybrid if it uses both integer addresses and CHERI capabilities. Hybrid software requires both RVY and Zyhybrid.

13.1.1. CHERI Pointer Mode Encoding

The current CHERI Pointer Mode is controlled by the P-bit in the capability AP-field of pc. The P-bit (P) is only architecturally relevant for capabilities granting X-permission, as it only affects execution state when installed into pc.

While the P-bit can be read or written explicitly using YMODER and YMODEW, these instructions do not change the current execution mode. The mode changes when a capability is installed into pc (e.g., via an indirect jump, see Section 13.1.2).

Capabilities lacking X-permission may not have a defined P-bit field in their encoding, and attempting to update this field may be a no-op. Zyhybrid requires a capability encoding that supports transport of the P-bit (e.g., RV64LYA).

  • CHERI Pointer Mode (P)=0 indicates Capability Pointer Mode.

  • CHERI Pointer Mode (P)=1 indicates Integral Pointer Mode.

When executing YPERMC, if X-permission is removed while the P-bit is set to one, and the capability encoding still permits representing the P-bit, then the P-bit must be set to zero.

13.1.2. Changing CHERI Pointer Mode

The P-bit of pc can be updated by the instructions listed in Table 46:

Table 46. Zyhybrid instructions that can perform mode changes
Mnemonic From mode Description

JALR (RVY)

Capability Pointer Mode

Jump to capability register, and link to capability register

YMODESWI

Capability Pointer Mode

Switch execution to Integral Pointer Mode

YMODESWY

Integral Pointer Mode

Switch execution to Capability Pointer Mode
When JALR (RVY) copies rs1 into pc it includes copying the P-bit and so setting the CHERI Pointer Mode of the target instruction.

The P-bit of a X-permission-granting capability can be read and written by the instructions listed in Table 47:

Table 47. Zyhybrid instructions to observe and update the mode in a capability
Mnemonic Description

YMODEW

Set capability execution mode

YMODER

Read capability mode
In addition to the mode switching instructions, the current mode can also be updated by setting the P-bit of a target capability using YMODEW followed by a JALR (RVY).

13.1.3. Representation of the P-bit in the capability encoding

For capabilities that do not grant X-permission, P-bit must always be interpreted and reported as 0 representing Capability Pointer Mode.

While this is not phrased as an additional rule for YPERMC to follow beyond those of Section 2.3.10.1, capability encodings may nevertheless take advantage of this implication in their representation of architectural CHERI capabilities.

13.1.4. Observing the CHERI Pointer Mode

The effective CHERI Pointer Mode cannot be determined just by reading the P-bit from pc since it also depends on the execution environment. The following code sequence demonstrates how a program can observe the current, effective CHERI Pointer Mode. It will write, to x1, the value 1 for Capability Pointer Mode (wherein pc has a set capability tag) or 0 for Integral Pointer Mode (wherein pc is just an address and has a clear capability tag):

auipc x1, 0
ytagr x1, x1
Implementations that support Zyhybrid will typically boot into Integral Pointer Mode so that non-CHERI-aware software can run unmodified. CHERI-aware software can observe and switch the mode as required.

13.2. Added instructions

13.2.1. YMODEW

Synopsis

Set capability execution mode

Mnemonic

ymodew rd, rs1, rs2

Encoding
Diagram
The instruction encoding is not final and is highly likely to change prior to v1.0
Description

Copy rs1 to rd.

If rs1 is sealed or if rs1 fails any integrity check, then set rd.tag=0.

Otherwise, if rs1 grants X-permission then update the P-bit of rd to:

  1. Capability Pointer Mode if the least significant bit of rs2 is 0, or,

  2. Integral Pointer Mode if the least significant bit of rs2 is 1.

    The value of rs1.tag does not affect the result.
Included in

Table 45

Operation 
  let cap = C(cs1);
  let mode = execution_mode_encdec(X(rs2)[0 .. 0]);

  let cap = clearTagIf(cap, capIsSealed(cap));
  let hasMode = not(permsMalformed(cap)) & canX(cap);
  let newCap = if hasMode then setCapMode(cap, mode) else cap;

  C(cd) = newCap;
  RETIRE_SUCCESS

13.2.2. YMODER

Synopsis

Read capability mode

Mnemonic

ymoder rd, rs1

Encoding
Diagram
The instruction encoding is not final and is highly likely to change prior to v1.0
Description

Decode the CHERI Pointer Mode from the capability in rs1 and write the result to rd.

Set rd to 0 if rs1 does not grant X-permission

Set rd to 0 if any integrity checks failed.

Otherwise set rd according to rs1 's CHERI Pointer Mode (P-bit):

  1. Set rd to 0 for Capability Pointer Mode, or,

  2. Set rd to 1 for Integral Pointer Mode.

    The value of rs1.tag does not affect the result.
Included in

Table 45

Operation
let capVal = C(cs1);
X(rd) = zero_extend(execution_mode_encdec(getCapMode(capVal)));
RETIRE_SUCCESS

13.2.3. YMODESWI

See YMODESWY.

13.2.4. YMODESWY

Synopsis

Switch execution mode to Capability Pointer Mode (YMODESWY), or Integral Pointer Mode (YMODESWI), 32-bit encodings

Mnemonic

ymodeswy
ymodeswi

Encoding
Diagram
The instruction encoding is not final and is highly likely to change prior to v1.0
Description

Set the current CHERI Pointer Mode in pc.

  • YMODESWY: If the current mode in pc is Integral Pointer Mode (1), then the P-bit in pc is set to Capability Pointer Mode (0). Otherwise no effect.

  • YMODESWI: If the current mode in pc is Capability Pointer Mode (0), then the P-bit in pc is set to Integral Pointer Mode (1). Otherwise no effect.

Included in

Table 45

Operation
let mode : ExecutionMode = match effective_cheri_mode() {
  IntPtrMode => CapPtrMode,
  CapPtrMode => IntPtrMode,
};
if debug_mode_active then dinfc = setCapMode(infinite_cap, mode);
set_next_pcc(setCapMode(PCC, mode));
RETIRE_SUCCESS

13.3. Added State

Zyhybrid adds the YLEN-wide CSR shown in Table 48.

Table 48. Unprivileged YLEN-wide CSRs added in Zyhybrid
YLEN CSR Permissions Description

ddc

URW

User Default Data Capability

13.3.1. Default Data Capability CSR (ddc)

ddc is a read-write, user mode accessible capability CSR. It does not require ASR-permission in pc for writes or reads. Similarly to pc authorizing all control flow and instruction fetches, this capability register is implicitly checked to authorize all data memory accesses when the current CHERI mode is Integral Pointer Mode. On startup ddc bounds and permissions must be set such that the program can run successfully (e.g., by setting it to have sufficiently broad bounds and permissions, possibly a Root Data capability).

Diagram
Figure 12. Unprivileged default data capability register

14. Zicsr (RVY modified behavior)

Table 49. Zicsr (RVY modified behavior) instructions
Mnemonic Function

CSRRW (RVY)

CSR write

CSRRS (RVY)

CSR set

CSRRC (RVY)

CSR clear

CSRRWI (RVY)

CSR write (immediate form)

CSRRSI (RVY)

CSR set (immediate form)

CSRRCI (RVY)

CSR clear (immediate form)

Zyhybrid makes Integral Pointer Mode available.

In addition to the rules in Section 2.6.4, when executing in Integral Pointer Mode, extended CSRs are always accessed as XLEN-wide for backwards compatibility. The final written value is determined using the semantics of the YADDRW instruction. YLEN-bit CSRs (including ddc) continue to be accessed as YLEN-wide.

14.1. CSRRWI (RVY)

See CSRRCI (RVY).

14.2. CSRRS (RVY)

See CSRRCI (RVY).

14.3. CSRRSI (RVY)

See CSRRCI (RVY).

14.4. CSRRC (RVY)

See CSRRCI (RVY).

14.5. CSRRCI (RVY)

Synopsis

CSR access (CSRRWI, CSRRS, CSRRSI, CSRRC, CSRRCI) 32-bit encodings for RVY

Mnemonics

csrrs rd, csr, rs1
csrrc rd, csr, rs1
csrrwi rd, csr, imm
csrrsi rd, csr, imm
csrrci rd, csr, imm

Encoding
Diagram
Description

These CSR instructions have extended functionality for accessing YLEN-bit CSRs and extended CSRs.

Access to XLEN-bit CSRs is as defined in Zicsr.

Zicsr rules are followed when determining whether to read or write the CSR. Suppressed read or write actions have no side-effects on the CSR.

All writes are XLEN bits only and use the semantics of the YADDRW instruction to determine the final write data.

Read data from YLEN-bit CSRs is always YLEN bits.

Read data from extended CSRs is YLEN bits in Capability Pointer Mode or XLEN bits in Integral Pointer Mode.

Permissions

Accessing CSRs may require ASR-permission.

Prerequisites

RVY, Zicsr, optional Zyhybrid

Included in

RVI (RVY modified behavior)

Operation
TBD

14.6. CSRRW (RVY)

Synopsis

CSR access (CSRRW) 32-bit encodings for (RVY)

Mnemonic

csrrw rd, csr, rs1

Encoding
Diagram
Description

CSRRW has extended functionality for accessing YLEN-bit CSRs and extended CSRs.

Access to XLEN-bit CSRs is as defined in Zicsr.

CSRRW accesses to YLEN-bit CSRs read YLEN bits into rd and write YLEN bits of rs1 into the CSR.

CSRRW accesses to extended CSRs read YLEN bits into rd and write YLEN bits of rs1 into the CSR. However, if Zyhybrid is supported and the current mode is Integral Pointer Mode, accesses to extended CSRs are XLEN-wide. The final write data for extended CSRs is determined using the semantics of the YADDRW instruction.

In all cases, when writing YLEN bits of rs1, if any integrity check fails then set the capability tag to zero before writing to the CSR.

Permissions

Accessing CSRs may require ASR-permission.

Prerequisites

RVY, Zicsr, optional Zyhybrid

Included in

RVI (RVY modified behavior)

Operation
TBD

15. "Zabhlrsc" Extension for Byte and Halfword Load Reserved/Store Conditional, Version 0.9

The Zalrsc extension offers LR/SC (load reserved/store conditional) instructions for words and doublewords. Zabhlrsc extends this by adding byte and halfword versions.

The absence of LR/SC operations for subword data types is problematic for CHERI software (RVY base architectures). Non-CHERI RISC-V software can use LR/SC on larger data types than are strictly required for the memory access to register the reservation set. RVY checks memory bounds and so it is impossible to round a subword access up to a word or larger to gain the reservation set.

15.1. Byte and Halfword Atomic Load Reserved/Store Conditional Instructions

The Zabhlrsc extension provides the LR.[B|H] and SC.[B|H] instructions.

zabhlrsc lr ext wavedrom reg
zabhlrsc sc ext wavedrom reg

LR.[B|H] behave analogously to LR.[W|D].

SC.[B|H] behave analogously to SC.[W|D].

All Zabhlrsc instructions sign extend the result and write it to rd.

15.2. Placeholder references to the unprivileged spec

This chapter only exists for the standalone document to allow references to resolve.
RV32I

See Chapter RV32I Base Integer Instruction Set in (RISC-V, 2023).

RV32E and RV64E

See Chapter RV32E and RV64E Base Integer Instruction Sets in

General purpose registers

See Chapter RV32I Base Integer Instruction Set in (RISC-V, 2023).

Load and Store Instructions

See Chapter RV32I Base Integer Instruction Set in (RISC-V, 2023).

Integer Register-Immediate Instructions

See Chapter RV32I Base Integer Instruction Set in (RISC-V, 2023).

Control Transfer Instructions

See Chapter RV32I Base Integer Instruction Set in (RISC-V, 2023).

Atomics

See Chapter "A" Extension for Atomic Instructions in (RISC-V, 2023).

Zba

See Chapter "B" Extension for Bit Manipulation in (RISC-V, 2023).

Zicbom

See Chapter "CMO" Extensions for Base Cache Management Operation ISA in (RISC-V, 2023).

Zcmt

See Chapter "Zc*" Extension for Code Size Reduction in (RISC-V, 2023).

Zcmp

See Chapter "Zc*" Extension for Code Size Reduction in (RISC-V, 2023).

jvt

See Chapter "Zc*" Extension for Code Size Reduction in (RISC-V, 2023).

Zaamo

See Chapter "A" Extension for Atomic Instructions in (RISC-V, 2023).

"Zalrsc" for RVY

See Chapter "A" Extension for Atomic Instructions in (RISC-V, 2023).

"Zaamo" for RVY

See Chapter "A" Extension for Atomic Instructions in (RISC-V, 2023).

Chapters for the privileged specification

16. "Machine-Level ISA (RVY)" Extensions, Version 1.0

This chapter will appear in the priv spec. Exact location TBD.

This chapter describes integration of RVY with the RISC-V privileged architecture.

16.1. Machine-Level CSRs widened by RVY

RVY widens some M-mode CSRs so that they can hold capabilities. Additionally, RVY allocates some previously reserved bits in existing CSRs.

mconfigptr is not widened, despite representing an address, as it is solely for use by low-level system software and is not interpreted by hardware. Such software can be expected to hold suitable Root capabilities from which it can derive a capability to the address in this register.

16.1.1. Machine Trap Vector Base Address Capability Register (mtvec)

The mtvec register is widened to a code capability. Its reset value is nominally a Root Executable capability.

mtvec (RVY) exists in all CHERI implementations, and so may be used as a source of a Root Executable capability after reset.
Diagram
Figure 13. Machine-mode trap-vector base-capability register

The fields in the metadata are WARL as many fields can be implemented as constants.

Examples of WARL behavior include always setting X-permission to 1 and setting the reserved fields to zero, otherwise the capability is unusable. Another example is to partially or fully restrict the bounds to constant values.
Care must be taken however that suitable root capabilities are available to software after reset if this CSR does not represent one.

When traps are taken into machine mode, the pc is updated following the standard mtvec behavior. The capability tag and metadata from mtvec (RVY) are also written to the pc.

Following the standard mtvec behavior, the value of mtvec.address can be viewed with a range of different addresses:

  1. The MODE field is included in mtvec.address[1:0] but it does not form part of the trap vector address.

  2. When MODE=Vectored, the trap vector address is incremented by four times the interrupt number.

  3. CSR reads include MODE in mtvec.address[1:0].

HICAUSE is defined to be the largest interrupt cause value that the implementation can write to xcause when an interrupt is taken.

Therefore the minimum observable address is mtvec.address & ~3 and the maximum is (mtvec.address & ~3) + 4 x HICAUSE.

All possible observable values must be represented exactly. Software must ensure this is true when writing to mtvec (RVY), and the hardware sets the capability tag to zero if any values cannot be represented exactly.

mtvec (RVY) is always updated using the semantics of the YADDRW instruction and so writing a sealed capability will cause the capability tag to be set to zero.

The capability in mtvec (RVY) is not unsealed when it is written to pc, unlike other executing from other CSRs such as mepc (RVY).

mtvec (RVY) follows the rule from mtvec about not needing to be able to hold all possible invalid addresses.

Machine Environment Configuration (menvcfg) Register
Diagram

16.1.2. Machine Scratch Capability Register (mscratch)

The mscratch register is widened to a capability.

The reset value of the capability tag of this CSR is zero, the reset values of the metadata and address fields are UNSPECIFIED.

Diagram
Figure 14. Machine-mode scratch capability register

16.1.3. Machine Exception Program Counter Capability (mepc)

The mepc register is widened to a capability. Its reset value is nominally a Root Executable capability.

Diagram
Figure 15. Machine exception program counter capability register

mepc.address is the mepc CSR, and so the follows the standard rules meaning that:

  1. mepc.address[0]=0, and

  2. mepc.address[1]=0 when IALIGN is fixed to 32

  3. mepc.address[1] reads as zero when IALIGN is programmable and is set to 32

As listed above for mtvec (RVY), this means that mepc.address can represent multiple different values. Therefore software must ensure that all possible values are represented exactly on writing, otherwise the hardware sets the written capability tag to zero.

Sealed capabilities may be written to mepc (RVY). The capability tag is set to zero on writing if:

  1. mepc.address[0]=1, or

  2. mepc.address[1]=1 when IALIGN=32

In the following case the value of the capability tag observable in the CSR depends on the value of IALIGN:

  1. mepc.ct≠0 && mepc.tag=1, and

  2. mepc.address[1]=1 and IALIGN=16 when writing the CSR

The observable capability tag is zero when IALIGN=32 and one when IALIGN=16.

When a trap is taken into M-mode, the pc is written to mepc.address following the standard behavior. The capability tag and metadata of the pc are also written to mepc (RVY).

On execution of an MRET (RVY) instruction, the capability value from mepc (RVY) is unsealed and written to pc.

mepc (RVY) follows the rule from mepc about not needing to be able to hold all possible invalid addresses.

16.2. Machine-Level CSRs added by RVY

16.2.1. Machine Thread Identifier Capability (mtidc)

The mtidc register is used to identify the current software thread in machine mode, using the method defined in the section for the unprivileged utidc CSR. On reset the capability tag of mtidc will be set to zero and the remainder of the data is UNSPECIFIED.

Diagram
Figure 16. Machine thread identifier capability register

16.3. Machine-Level CSRs modified by RVY

16.3.1. Machine ISA (misa) Register

RVY sets the Y bit to be 1. Either I or E is also set to show how many registers are present.

Implementations which allow misa.C to be writable need to legalize <`xepc`> on reading if the misa.C value has changed since the value was written as this can cause the read value of bit [1] to change state.

16.3.2. Machine Status Registers (mstatus and mstatush)

The mstatus and mstatush registers have two additional restrictions:

  • The MXL, SXL and UXL fields that control the value of XLEN for S-mode and U-mode must be read-only and equal to MXL in RVY implementations. Only 1 and 2 are supported values.

  • The MBE, SBE, and UBE fields that control the memory system endianness for M-mode, S-mode, and U-mode must be read-only in RVY implementations. SBE and UBE must be read only and equal to MBE, if S-mode or U-mode, respectively, is implemented, or read-only zero otherwise.

Changing XLEN or endianness would change the interpretation of all in-memory capabilities, so allowing these fields to change at runtime is prohibited.

These restrictions may be relaxed by a future extension. Such an extension is likely to enforce the constraint that any privilege level with XLEN less than MXLEN has CHERI disabled.

MXR has no effect on the CHERI permission checking.

CHERI does not need to make use of execute only memory for security reasons, and so MXR has no relevance.

16.3.3. Machine Cause Register (mcause)

RVY adds new exception codes for CHERI exceptions that mcause must be able to represent. The new exception codes and priorities are listed in Table 50 and the priorities are listed in Table 51 respectively.

Table 50. Synchronous exception causes for RVY.
Exception code Description

32

CHERI Instruction Access Fault

33

CHERI Load Access Fault

34

CHERI Store/AMO Access Fault

35

CHERI Load Capability Fault1

36

CHERI Store/AMO Page Fault1

1 These exceptions are defined by the Svyrg extension which is available for RV64Y only.

Table 51. Synchronous exception priority in decreasing priority order for RVY.
Priority Exc.Code Description

Highest

3

Instruction address breakpoint

32

Prior to instruction address translation:
CHERI Instruction Access Fault due to pc checks (capability tag, sealed, execute permission, bounds1)

12, 1

During instruction address translation:
First encountered page fault or access fault

1

With physical address for instruction:
Instruction access fault

2
32
0
8,9,11
3
3

Illegal instruction
CHERI Instruction Access Fault due to pc ASR-permission clear
Instruction address misaligned
Environment call
Environment break
Load/store/AMO address breakpoint

33,34

Prior to address translation for an explicit memory access:
CHERI Load Access Fault, CHERI Store/AMO Access Fault due to capability checks (capability tag, sealed, permissions, bounds)

4,6

Load/store/AMO capability address misaligned

4,6

Optionally:
Load/store/AMO address misaligned

36, 13, 15, 5, 7

During address translation for an explicit memory access:
First encountered CHERI Store/AMO Page Fault, page fault or access fault

5,7

With physical address for an explicit memory access:
Load/store/AMO access fault

4,6

If not higher priority:
Load/store/AMO address misaligned

Lowest

35

CHERI Load Capability Fault2

1 pc bounds are checked against all bytes of fetched instructions. If the instructions could not be decoded to determine the length, then the pc bounds check is made against the minimum sized instruction supported by the implementation which can be executed, when prioritizing against Instruction Access Faults.

2 CHERI Load Capability Fault is the lowest priority as determining whether to raise the exception may include checking the loaded capability tag.

Full details of CHERI Instruction Access Fault, CHERI Load Access Fault and CHERI Store/AMO Access Fault are in Table 52.

16.3.4. Machine Trap Delegation Register (medeleg)

Bits 32,33,34,35,36 of medeleg refer to valid CHERI exceptions and so can be used to delegate CHERI exceptions to supervisor mode.

16.3.5. "Smstateen" Integration

The TIDC bit in mstateen0 controls access to the stidc, and utidc CSRs.

Diagram
Figure 17. Machine State Enable 0 Register (mstateen0)

16.4. CHERI Exception handling

CHERI faults are typically higher priority than standard RISC-V faults. E.g., CHERI faults on the pc are higher priority than any other fault effecting the program counter such as Instruction Access Fault.

The overall priority is shown in Table 51.

Table 52 shows the detailed exception checks added by CHERI.

auth_cap is rs1, unless in Integral Pointer Mode when it is ddc (if Zyhybrid is implemented).
Table 52. Valid CHERI exception combination description
Instructions Xcause Description Check

All instructions have these exception checks first

All

32

CHERI Instruction Access Fault

pc capability tag is zero

All

32

CHERI Instruction Access Fault

pc is sealed

All

32

CHERI Instruction Access Fault

pc does not grant X-permission

All

32

CHERI Instruction Access Fault

Any byte of current instruction out of pc bounds1

All

32

CHERI Instruction Access Fault

pc failed any integrity check.

CSR/Xret additional exception check

CSR*, MRET (RVY), SRET (RVY), CBO.INVAL (RVY)

2

Illegal instruction

pc does not grant ASR-permission when required for CSR access or execution of MRET (RVY), SRET (RVY) or CBO.INVAL (RVY)

Load additional exception checks

All loads

33

CHERI Load Access Fault

auth_cap capability tag is zero

All loads

33

CHERI Load Access Fault

auth_cap is sealed

All loads

33

CHERI Load Access Fault

auth_cap does not grant R-permission

All loads

33

CHERI Load Access Fault

Any byte of load access out of auth_cap bounds2

All loads

33

CHERI Load Access Fault

auth_cap failed any integrity check.

Capability loads

52

Load access fault

Misaligned capability load

Store/atomic/cache-block-operation additional exception checks

All stores, all atomics, all CBOs

34

CHERI Store/AMO Access Fault

auth_cap capability tag is zero

All stores, all atomics, all CBOs

34

CHERI Store/AMO Access Fault

auth_cap is sealed

All stores, CBO.ZERO

34

CHERI Store/AMO Access Fault

auth_cap does not grant W-permission

All atomics, CBO.CLEAN, CBO.FLUSH, CBO.INVAL

34

CHERI Store/AMO Access Fault

auth_cap does not grant both R-permission and W-permission

All stores, all atomics

34

CHERI Store/AMO Access Fault

any byte of access out of auth_cap bounds2

CBO.ZERO, CBO.INVAL

34

CHERI Store/AMO Access Fault

any byte of cache block out of auth_cap bounds2

CBO.CLEAN, CBO.FLUSH

34

CHERI Store/AMO Access Fault

all bytes of cache block out of auth_cap bounds2

All stores, all atomics, all CBOs

34

CHERI Store/AMO Access Fault

auth_cap failed any integrity check.

Capability stores

73

Store access fault

Misaligned capability store

1 The pc bounds checks includes the case where the bounds could not be decoded and so a minimum sized instruction is not fully in bounds.

2 The load/store bounds check includes the case where the bounds could not be decoded as they are malformed.

3 Misaligned capability accesses raise access faults instead of misaligned faults since they cannot be emulated in software.

16.5. CHERI Exceptions and speculative execution

CHERI adds architectural guarantees that can prove to be microarchitecturally useful. Speculative-execution attacks can — among other factors — rely on instructions that fail CHERI permission checks not to take effect. When implementing any of the extensions proposed here, microarchitects need to carefully consider the interaction of late-exception raising and side-channel attacks.

16.6. Physical Memory Attributes (PMA)

Typically, only parts of the entire memory space need to support CHERI capability tags. Therefore Physical Memory Attributes are included indicating whether a memory region allows storing CHERI capability tags and to define the behavior when attempting to access them.

There are three levels of support:

Table 53. CHERI PMAs
PMA Load Behavior Store Behavior Comment

CHERI Capability Tag

Load capability tag

Store capability tag

Tagged memory supporting capability tags

CHERI Capability Tag Strip

Load zero capability tag

Ignore stored capability tag

No support for capability tags, ignore them

CHERI Capability Tag Store Fault

Load zero capability tag

Store/AMO Access Fault on capability tag1

No support for capability tags, trap on storing one

1 The access fault is triggered on all capability stores or atomics such as SY or AMOSWAP.Y when C-permission and W-permission are granted and the to-be-stored capability tag is set to one. For SC.Y, the access fault is raised if the to-be-stored capability tag is set in rs2 even when the store fails.

Only memory regions that have the CHERI Capability Tag PMA require storage for capability tags.

16.7. Modified MRET Behavior

MRET (RVY) reads the full YLEN bits of the mepc (RVY) register and unseals it prior to exception return if it is sealed as a sealed entry point capability.

16.8. Machine-mode YLEN CSR Summary

Table 54 shows all YLEN Machine-mode CSRs added or modified by RVY.

Table 54. All RVY Machine-Mode CSRs.
RVY CSR Prerequisites Width Address Permissions Reset Value Description

mscratch (RVY)

M-mode

YLEN

0x340

MRW, ASR-permission

tag=0, otherwise specified by the platform

Machine Scratch Capability

mtvec (RVY)

M-mode

YLEN

0x305

MRW, ASR-permission

Nominally Root Executable

Machine Trap-Vector Base-Address Capability

mepc (RVY)

M-mode

YLEN

0x341

MRW, ASR-permission

Nominally Root Executable

Machine Exception Program Counter Capability

mtidc

RVY

YLEN

0x780

Read: M, Write: M, ASR-permission

tag=0, otherwise specified by the platform

Machine thread ID
Where reset values are specified as a Root Executable or Root Data they are a maximum possible value. The platform may reset those CSRs to a smaller memory range, or to have fewer permissions.
Table 55. Action taken on writing to extended Machine-Mode CSRs
RVY CSR Width Action on XLEN write Action on YLEN write

mscratch (RVY)

YLEN

Update the CSR using YADDRW.

direct write

mtvec (RVY)

YLEN

Apply Invalid address conversion. Always update the CSR with YADDRW even if the address didn’t change, including the MODE field in the address for simplicity. Vector range check if vectored mode is programmed.

Apply Invalid address conversion. Always update the CSR with YADDRW even if the address didn’t change, including the MODE field in the address for simplicity. Vector range check if vectored mode is programmed.

mepc (RVY)

YLEN

Apply Invalid address conversion. Always update the CSR with YADDRW even if the address didn’t change.

Apply Invalid address conversion and update the CSR with the result if the address changed, direct write if address didn’t change

mtidc

YLEN

Update the CSR using YADDRW.

direct write

Code pointers and data pointers in Table 56 are WARL CSRs that do not need to store full 64-bit addresses on RV64, and so may be subject to Invalid address conversion on writing.

Table 56. YLEN-wide Machine-Mode CSRs storing code pointers or data pointers
RVY CSR Code Pointer Data Pointer Unseal On Execution

mtvec (RVY)

✔

mepc (RVY)

✔

✔

17. "Supervisor-Level ISA (RVY)" Extensions, Version 1.0

17.1. Supervisor-Level CSRs widened by RVY

17.1.1. Supervisor Trap Vector Base Address Capability Register (stvec)

The stvec register is widened to a capability.

When the S-mode execution environment starts, the value is nominally the Root Executable capability.

Diagram
Figure 18. Supervisor trap-vector base-capability register

The handling of stvec (RVY) is identical to mtvec (RVY), but in supervisor mode.

17.1.2. Supervisor Scratch Capability Register (sscratch)

The sscratch register is widened to a capability.

At the start of the S-mode execution environment, the value of the capability tag of this CSR is zero and the values of the metadata and address fields are UNSPECIFIED.

Diagram
Figure 19. Supervisor scratch capability register

17.1.3. Supervisor Exception Program Counter Capability (sepc)

The sepc register is widened to a code capability.

When the S-mode execution environment starts, the value is nominally the Root Executable capability.

The handling of sepc (RVY) is identical to mepc (RVY), but in supervisor mode (i.e. applying to SRET (RVY)).

Diagram
Figure 20. Supervisor exception program counter capability register

17.2. Supervisor-Level CSRs added RVY

17.2.1. Supervisor Thread Identifier Capability (stidc)

The stidc register is used to identify the current software thread in supervisor mode, using the method defined in the section for the unprivileged utidc CSR.

At the start of the S-mode execution environment, the value of the capability tag of this CSR is zero and the values of the metadata and address fields are UNSPECIFIED.

Diagram
Figure 21. Supervisor thread identifier capability register

17.3. Supervisor-Level CSRs modified by RVY

17.3.1. Supervisor environment configuration register (senvcfg)

Zyhybrid adds a Y bit to senvcfg

Diagram
Figure 22. Supervisor environment configuration register (senvcfg)

17.3.2. Supervisor Cause Register (scause)

RVY adds new exception codes for CHERI exceptions as in mcause (RVY).

17.3.3. "Ssstateen" Integration

The TIDC bit in sstateen0 controls access to the utidc CSR. See utidc for a description of the usage.

Diagram
Figure 23. Supervisor State Enable 0 Register (sstateen0)

17.4. Modified SRET Behavior

SRET (RVY) reads the full YLEN bits of the sepc (RVY) register and unseals it prior to exception return if it is sealed as a sealed entry point capability.

17.4.1. SRET (RVY)

See MRET (RVY).

17.4.2. MRET (RVY)

Synopsis

Trap Return (MRET, SRET)

Mnemonics

mret
sret

Encoding
Diagram
Description

Return from machine mode (MRET) or supervisor mode (SRET) trap handler.

MRET unseals mepc (RVY) and writes the result into pc.

SRET unseals sepc (RVY) and writes the result into pc.

Exceptions

An illegal instruction fault is raised when pc does not grant ASR-permission.

ASR-permission is required as MRET and SRET require access to privileged CSRs.
Prerequisites (MRET)

Machine-Level ISA, RVY

Prerequisites (SRET)

Supervisor-Level ISA, RVY

Operation
TBD

17.5. Supervisor-mode YLEN CSR Summary

Table 57 shows all YLEN Supervisor-mode CSRs added or modified by RVY.

Table 57. All RVY Supervisor-Mode CSRs.
RVY CSR Prerequisites Width Address Permissions Reset Value Description

sscratch (RVY)

S-mode

YLEN

0x140

SRW, ASR-permission

tag=0, otherwise specified by the platform

Supervisor Scratch Capability

stvec (RVY)

S-mode

YLEN

0x105

SRW, ASR-permission

Nominally Root Executable

Supervisor Trap-Vector Base-Address Capability

sepc (RVY)

S-mode

YLEN

0x141

SRW, ASR-permission

Nominally Root Executable

Supervisor Exception Program Counter Capability

stidc

RVY

YLEN

0x580

Read: S, Write: S, ASR-permission

tag=0, otherwise specified by the platform

Supervisor thread ID

Code pointers and data pointers in Table 59 are WARL CSRs that do not need to store full 64-bit addresses on RV64, and so may be subject to Invalid address conversion on writing.

Table 58. Action taken on writing to extended Supervisor-Mode CSRs
RVY CSR Width Action on XLEN write Action on YLEN write

sscratch (RVY)

YLEN

Update the CSR using YADDRW.

direct write

stvec (RVY)

YLEN

Apply Invalid address conversion. Always update the CSR with YADDRW even if the address didn’t change, including the MODE field in the address for simplicity. Vector range check if vectored mode is programmed.

Apply Invalid address conversion. Always update the CSR with YADDRW even if the address didn’t change, including the MODE field in the address for simplicity. Vector range check if vectored mode is programmed.

sepc (RVY)

YLEN

Apply Invalid address conversion. Always update the CSR with YADDRW even if the address didn’t change.

Apply Invalid address conversion and update the CSR with the result if the address changed, direct write if address didn’t change

stidc

YLEN

Update the CSR using YADDRW.

direct write
Table 59. YLEN-wide Supervisor-Mode CSRs storing code pointers or data pointers
RVY CSR Code Pointer Data Pointer Unseal On Execution

stvec (RVY)

✔

sepc (RVY)

✔

✔

18. "Supervisor-Level ISA for Virtual Memory (RV64Y)" Extension, Version 1.0 for RV64Y

CHERI checks are made on the effective address according to the current translation scheme. I.e., on the virtual address if translation is enabled or the physical address if translation is disabled.

Implicit memory accesses made by the page table walker are not subject to CHERI checks.

A future extension may add CHERI checks to the page table walker.

18.1. Limiting Capability Propagation

Page table enforcement can allow the operating system to limit the flow of capabilities between processes. It is highly desirable that a process should only possess capabilities that have been issued for that address space by the operating system. Unix processes may share memory for efficient communication, but capability pointers must not be shared across these channels into a foreign address space. An operating system might defend against this by only issuing a capability to the shared region that does not grant the load/store capability permission. However, there are circumstances where portions of general-purpose, mmapped* memory become shared, and the operating system must prevent future capability communication through those pages. This is not possible without restructuring software, as the capability for the original allocation, which spans both shared memory and private memory, would need to be deleted and replaced with a list of distinct capabilities with appropriate permissions for each range. Such a change would not be transparent to the program. Such sharing through virtual memory is on the page granularity, so preventing capability writes with a PTE permission is a natural solution.

* allocated using mmap

18.2. The pte.rvy field for capability flow control

As RV32Y has no field allocated in the PTE, capability loads and stores operate as normal. Therefore the RVY field is reserved in the RV64 PTE formats only (e.g. Figure 24) for use by RV64Y extensions. The principle is that extensions, such as Svyrg may redefine the usage of the RVY field by adding an enable bit into Xstatus.

The definition of the RVY field is as follows:

Table 60. Definition of the pte.rvy field
Bit Name Comment

pte.rvy[3]

pte.y

Enable capability (Y) access

pte.rvy[2]

reserved

reserved

pte.rvy[1]

reserved

reserved

pte.rvy[0]

reserved

reserved

If pte.y=0 then:

  • All capability loads set the loaded capability tag to zero

  • All capability stores with the to-be-stored capability tag set raise a CHERI Store/AMO Page Fault.

If pte.y=1 then:

  • All capability loads and capability stores operate as normal.

18.3. Invalid Virtual Address Handling

When address translation is in effect for RV64Y, the upper bits of virtual memory addresses must match for the address to be valid.

The CSRs shown in Table 56 and Table 59, as well as the pc, need not hold all possible invalid addresses.

However, the bounds encoding of capabilities depends on the address value if the bounds are not infinite.

Therefore implementations must not convert invalid addresses to other arbitrary invalid addresses in an unrestricted manner if the bounds are not infinite.

If the bounds could not be decoded due to the address being invalid, then a CHERI Instruction Access Fault, CHERI Load Access Fault or CHERI Store/AMO Access Fault exception is raised as appropriate.

Not requiring the implementation to decode the bounds for invalid addresses reduces the size of bounds comparators from 64-bits to the supported virtual address width.

18.3.1. Updating CSRs

A CSR may be updated to hold a capability with an invalid address, due to:

To ensure that the bounds of a valid capability cannot be corrupted:

  • If the new address is invalid and the capability bounds are not infinite then set the capability tag to zero before writing to the CSR.

When the new address is invalid and happens to match an invalid address which the CSR can hold, then it is implementation-defined whether to set the capability tag to zero.

18.3.2. Branches and Jumps

If the effective target address of the jump or branch is invalid, and the authorizing capability’s bounds are not infinite, then set the capability tag of the target pc to zero.

A CHERI Instruction Access Fault exception will be raised at the target.
RISC-V harts that do not support RVY normally raise an instruction access fault or page fault after jumping or branching to an invalid address. Therefore, RVY preserves that behavior to ensure that harts supporting RVY and Zyhybrid are fully compatible with RISC-V harts provided that pc and ddc are set to Root Executable and Root Data capabilities, respectively.

18.3.3. Memory Accesses

If the effective address of the memory access is invalid, and the authorizing capability’s bounds are not infinite, then raise a CHERI Load Access Fault or CHERI Store/AMO Access Fault exception because the bounds cannot be reliably decoded.

19. "Zyhybrid for Privileged Architectures" Extension, Version 1.0

When using a system with Zyhybrid, it may be desirable to configure some (or all) privilege levels such that they operate as a standard RV32[IE]/RV64[IE] base architecture.

Zyhybrid includes functions to disable CHERI by making the Y bit of misa (RVY), menvcfg (RVY), and senvcfg (RVY) writable.

The effective CHERI-enable for the current privilege is:

On reset CHERI is always disabled for backwards compatibility: misa (RVY).Y resets to zero, ddc, pc, Xtvec, Xepc are nominally root capabilities.

misa (RVY).Y is a WARL field. M-mode software can freely set it to one, but then it can only be set back to zero if ddc, pc, Xtvec, Xepc are root capabilities so that no CHERI configuration has taken effect.

In this way M-mode software cannot disable any CHERI checks once they have been configured.

The following occurs when executing code in a privilege mode that has CHERI disabled:

  • Custom encoding spaces reallocated as standard encoding spaces for RVY revert to being custom encoding spaces

    • Therefore the encodings in RVY and Table 45 are no longer available.

  • Natively YLEN CSRs are no longer available.

  • Only XLEN CSR access is permitted.

    • This is identical to Integral Pointer Mode access.

Disabling CHERI has no effect on implicit accesses or security checks. The last capability written to pc and ddc before disabling CHERI will be used to authorize instruction execution and data memory accesses respectively.

Disabling CHERI prevents low-privileged Integral Pointer Mode software from interfering with the correct operation of higher-privileged Integral Pointer Mode software that does not perform ddc switches on trap entry and return.

Table 61 summarizes the behavior of a hart for the different CHERI modes, while executing outside debug mode.

Table 61. Hart’s behavior depending on the effective CHERI enable and CHERI Pointer Mode
Y1=0 Y=1, P-bit=1 Y=1, P-bit=0

Authorizing capability for data memory accesses

ddc

ddc

capability in rs1

Natively YLEN CSRs available

✘

✔

✔

Extended YLEN CSR access width

XLEN

XLEN

YLEN

CHERI instructions available

✘

✔2

✔

custom-1/2/3 available for custom use

✔

✘

✘

Summary

Fully RV32[IE]/RV64[IE] compatible3

Integral Pointer Mode

Capability Pointer Mode

1 Y represents the effective CHERI enable for the current privilege mode.

2 The RVY compressed instructions operating on capability data are unavailable as their encodings will revert to RV32[IE]/RV64[IE] standard behavior.

3 If Y=0 in M-mode the hart is fully compatible with RV32[IE]/RV64[IE]. For lower privileged modes pc or ddc may not be root capabilities and so the behavior may be restricted.

20. "Svyrg" Extension, Version 1.0 for RV64Y

This extension specifies use of the RVY field in the Sv39, Sv48 and Sv59 PTE formats (e.g., see Figure 24).

Page-Based 39-bit Virtual-Memory System (Sv39)
Diagram
Figure 24. Sv39 page table entry.

Svyrg implies Sv39.

The Svyrg extension is enabled when the sstatus.YRGE bit is set.

The extension adds the ability for supervisor-mode software to quickly enable trapping loads of capabilities from all pages of memory, incrementally allow loads of capabilities from such pages, and track stores of capabilities to all pages of memory. Applied to userspace pages, this has been shown to allow the operating system to implement capability revocation schemes, which allow userspace memory allocators to deterministically guard against use-after-reallocation (Filardo et al., 2024). Applied to kernel pages, they can similarly be used against use-after-reallocation within the kernel itself.

The 4-bit pte.rvy field is subdivided into four 1-bit fields, two controlling capability loads (including the capability loaded during an AMO) and two controlling capability stores (again, including the capability stored as part of an AMO).

Table 62. Breakdown of the 4-bit pte.rvy field.
Bit Name Comment

Capability load/AMO fields

pte.rvy[0]

pte.yr

Capability readable

pte.rvy[1]

pte.yrg

Capability read generation

Capability store/AMO fields

pte.rvy[2]

pte.yw

Capability writable

pte.rvy[3]

pte.yd

Capability dirty

When all of the following are true, Svyrg adds two related architectural features, CHERI Load Capability Faults and Capability Dirty Tracking:

  1. sstatus.YRGE is set.

  2. The authorizing capability has C-permission.

  3. The PMA is set to CHERI Capability Tag.

It may make sense to subdivide this into two extensions as Capability Dirty Tracking is more generally useful than CHERI Load Capability Faults.

20.1. CHERI Load Capability Faults

pte.yr and pte.yrg are used to enable capability loads or AMOs to write a capability tag to rd.

When pte.yr is clear, pte.yrg controls whether capabilities can write a capability tag to rd. When clear capability tags cannot be written to rd, and when set, they can.

When pte.yr is set, pte.yrg can be used to trap on capability loads or AMOs when it does not match the Capability Read Generation value that is represented by the value of sstatus.UYRG for userspace pages (pte.u=1) or sstatus.SYRG for kernel pages (pte.u=0).

The implementation raises a CHERI Load Capability Fault when, in addition to the rules above, all of the following are true:

  1. A capability load or AMO is executed.

  2. pte.yr is set.

  3. if pte.u=1, pte.yrg does not equal sstatus.UYRG.

  4. if pte.u=0, pte.yrg does not equal sstatus.SYRG.

  5. Optionally, the loaded capability tag is set1.

  6. Any other platform specific rules have not forced the loaded capability tag to be clear.

1Checking the value of the capability tag requires taking data dependent exceptions on loaded capabilities for loads or AMOs. Ideally all implementations would trap precisely (taking the capability tag into account in all cases) rather than conservatively (trapping more often, potentially on every loaded capability). However, the loaded capability tag may not be available in all implementations when determining whether to raise the exception, and therefore flexibility is permitted. As a result, the software is required to be tolerant of raising the trap when the capability tag is not set, potentially resulting in spurious traps from pages that have pte.yr=1, and so are likely to store valid capabilities.

Implementations that already take synchronous traps on loaded data, such as ECC faults, are recommended to check the loaded capability tag when determining whether to raise the fault.
Table 63. Summary of capability load pte.yr and pte.yrg behavior in the PTEs
pte.yr pte.yrg tag1 Load Capability Behavior

0

0

X

Clear loaded capability tag

0

1

X

Normal operation

1

≠ sstatus.xYRG2

0

Implementation defined choice of
CHERI Load Capability Fault or normal operation

1

≠ sstatus.xYRG2

1

CHERI Load Capability Fault

1

= sstatus.xYRG2

X

Normal operation

1 The loaded capability tag.

2 if pte.u=1, sstatus.UYRG, else sstatus.SYRG.

CHERI Load Capability Faults may be used to implement the load-barrier primitive from (Filardo et al., 2024).
In the case of AMOs that could trigger both a CHERI Store/AMO Page Fault, due to storing a valid capability tag, and a CHERI Load Capability Fault, the CHERI Store/AMO Page Fault takes priority as shown in Table 51.

20.2. Capability Dirty Tracking

When pte.yw is clear, capability stores or AMOs where the to-be-stored capability tag is set will raise a CHERI Store/AMO Page Fault fault.

When pte.yw is set, capability stores to the virtual page are permitted. In addition, the pte.yd bit indicates that a capability was stored to the virtual page since the last time the pte.yd bit was cleared.

This is equivalent to the behavior of the pte.d bit except that pte.yd only tracks stores of valid capabilities instead of all stores.

Capability dirty tracking behavior is enabled when, in addition to the rules above, all of the following are true:

  1. A capability store or AMO instruction is executed.

  2. The to-be-stored capability tag is set.

  3. pte.yw is set.

  4. pte.yd is clear.

Two schemes for capability dirty tracking are permitted, and the scheme in use is determined by whether the Svade or Svadu extensions are enabled.

  • For Svade, take a CHERI Store/AMO Page Fault.

  • For Svadu, do a hardware update that sets pte.yd=1, following the same rules as setting pte.d.

    • When setting pte.yd, the hardware update also necessarily sets (or leaves set) pte.a and pte.d.

Table 64. Summary of capability store pte.yw and pte.yd behavior in the PTEs
pte.yw pte.yd tag1 Store Capability Behavior

0

X

0

Normal operation

0

X

1

CHERI Store/AMO Page Fault

1

0

0

Normal operation

1

0

1

CHERI Store/AMO Page Fault (Svade) or
hardware pte.yd update (Svadu)

1

1

X

Normal operation

1 The to-be-stored capability tag.

The capability dirty tracking is resolved during memory translation, but there are cases where it is not known if there will be a capability tag stored to memory or not at this point. Capability dirty tracking must be triggered when pte.yw=1 and pte.yd=0 and the to-be-stored capability tag is set, and may be triggered in the following case where it is not known if there will be a stored capability tag during translation:

  • SC.Y triggers capability dirty tracking if the capability tag is set in rs2, even if the store fails. This matches the semantics of SC.* with regard to pte.d.

pte.yd must always be set when a capability with a valid capability tag is written to a virtual page so that software knows which pages have had capabilities stored to them. It may be set too often, which may cause software to examine the page to check for capabilities when none are present. This is a situation software is required to handle anyway, as it is always possible for all capabilities in a page to be overwritten by non-capability data. In this case the pte.yd bit would still be set.

Future AMOs fall into this category:

  • For future AMOCAS.Y, it is not known whether the store will happen until the load has executed, and the compare has been done. Therefore, AMOCAS.Y is likely to trigger capability dirty tracking if the capability tag is set in rs2.

  • For future AMOADD.Y, the stored capability tag depends upon the loaded capability tag which is not known during translation and on the execution of the YADD. Therefore, AMOADD.Y, is likely to always trigger capability dirty tracking.

Checking the stored capability tag is less of a burden to the implementation than checking the loaded capability tag for CHERI Load Capability Fault, which is why checking the loaded capability tag is optional behavior. However, a future extension may reduce the burden further by removing the check on the to-be-stored capability tag.
Capability dirty tracking may be used to implement the store-barrier primitive from (Filardo et al., 2024).
The minimum level of PTE support is to set pte.yr=1, pte.yd=1, pte.yw=1 and pte.yrg=0 in all PTEs intended for storing capabilities (e.g., private anonymous mappings) and set sstatus.UYRG=0 and sstatus.SYRG=0 on all harts, which will enable capabilities to be loaded and stored successfully.

20.3. UYRG CSR field

The mstatus, sstatus and vsstatus CSRs include the Userspace Capability Read Generation (UYRG) bit, the Supervisor version (SYRG) and the enable bit (YRGE).

Diagram
Figure 25. Machine-mode status (mstatus) register for RV64Y
Diagram
Figure 26. Supervisor-mode status (sstatus) register when SXLEN=64
Diagram
Figure 27. Virtual Supervisor-mode status (vsstatus) register when VSXLEN=64

Appendix A: RVY Instruction Encoding Overview

A.1. Bytefield Encoding Overview

A.1.1. 3OP Instruction Encodings

Diagram

A.1.2. 2OP Instruction Encodings

Diagram

A.1.3. YADDI/LY/SY Instruction Encodings

Diagram

A.1.4. AMO Instruction Encodings

Diagram

A.1.5. MISC Instruction Encodings

Diagram

A.2. Wavedrom Encoding Overview

A.2.1. 3OP Instruction Encodings

Diagram

A.2.2. 2OP Instruction Encodings

Diagram

A.2.3. YADDI/LY/SY Instruction Encodings

Diagram
Diagram
Diagram

A.2.4. AMO Instruction Encodings

Diagram
Diagram
Diagram

A.2.5. MISC Instruction Encodings

Diagram
Diagram
Diagram

A.3. AsciiDoc Table Encoding Overview

A.3.1. 3OP Instruction Encodings

3OP Instruction Encodings

31:25

24:20

19:15

14:12

11:7

6:0

Inst

0000001

rs2

rs1

000

rd

RVY-A=1111011

PACKY

0000011

rs2≠0

rs1

000

rd

RVY-A=1111011

YADD

0000011

00000

rs1

000

rd

RVY-A=1111011

YMV

0001011

rs2

rs1

000

rd

RVY-A=1111011

YADDRW

0010011

rs2

rs1

000

rd

RVY-A=1111011

YPERMC

0011011

rs2

rs1

000

rd

RVY-A=1111011

YBNDSW

0100011

rs2

rs1

000

rd

RVY-A=1111011

YBNDSRW

0101011

rs2

rs1

000

rd≠0

RVY-A=1111011

YMODEW

0101011

00000

rs1=x0

000

rd=x0

RVY-A=1111011

YMODESWY

0101011

00001

rs1=x0

000

rd=x0

RVY-A=1111011

YMODESWI

0000101

rs2

rs1

000

rd

RVY-A=1111011

YSH1ADD

0001101

rs2

rs1

000

rd

RVY-A=1111011

YSH2ADD

0010101

rs2

rs1

000

rd

RVY-A=1111011

YSH3ADD

0011101

rs2

rs1

000

rd

RVY-A=1111011

YSH4ADD

0100101

rs2

rs1

000

rd

RVY-A=1111011

YSH1ADD.UW

0101101

rs2

rs1

000

rd

RVY-A=1111011

YSH2ADD.UW

0110101

rs2

rs1

000

rd

RVY-A=1111011

YSH3ADD.UW

0111101

rs2

rs1

000

rd

RVY-A=1111011

YSH4ADD.UW

0000110

rs2

rs1

000

rd

RVY-A=1111011

YEQ

0001110

rs2

rs1

000

rd

RVY-A=1111011

YSS

0000111

rs2

rs1

000

rd

RVY-A=1111011

YSUNSEAL

0001111

rs2

rs1

000

rd

RVY-A=1111011

YBLD

0010111

rs2

rs1

000

rd

RVY-A=1111011

YSEAL

A.3.2. 2OP Instruction Encodings

2OP Instruction Encodings

31:25

24:20

19:15

14:12

11:7

6:0

Inst

1111000

00000

rs1

000

rd

RVY-A=1111011

YAMASK

1111010

00000

rs1

000

rd

RVY-A=1111011

YBASER

1111010

00001

rs1

000

rd

RVY-A=1111011

YPERMR

1111010

00010

rs1

000

rd

RVY-A=1111011

YTOPR

1111010

00011

rs1

000

rd

RVY-A=1111011

YLENR

1111010

00100

rs1

000

rd

RVY-A=1111011

YTAGR

1111010

00101

rs1

000

rd

RVY-A=1111011

YTYPER

1111010

00110

rs1

000

rd

RVY-A=1111011

YMODER

A.3.3. YADDI/LY/SY Instruction Encodings

YADDI/LY/SY Instruction Encodings

31:25

24:20

19:15

14:12

11:7

6:0

Inst

imm[11:0]

rs1

100

rd

RVY-A=1111011

YADDI

imm[11:0]

rs1≠0

001

rd

RVY-A=1111011

LY

offset[11:5]

rs2

rs1≠0

010

offset[4:0]

RVY-A=1111011

SY

A.3.4. AMO Instruction Encodings

AMO Instruction Encodings

31:27

26

25

24:20

19:15

14:12

11:7

6:0

Inst

00010

aq

rl

00000

rs1≠0

011

rd

RVY-A=1111011

LR.Y

00001

aq

rl

rs2

rs1≠0

011

rd

RVY-A=1111011

AMOSWAP.Y

00011

aq

rl

rs2

rs1≠0

011

rd

RVY-A=1111011

SC.Y

A.3.5. MISC Instruction Encodings

MISC Instruction Encodings

31:29

28:27

26:20

19:15

14:12

11:7

6:0

Inst

00000

shamt=XLEN

rs1

101

rd

RVY-A=1111011

YHIR

00000

shamt[6:0]

rs1

101

rd

RVY-A=1111011

SRLIY

111

ybndswimm[8:0]

rs1

101

rd

RVY-A=1111011

YBNDSWI

A.4. RVY Custom-3 Instruction Allocation

A.4.1. 1. Funct3 Allocations

funct3 Instruction(s)

000

R-Type (3-op and 1/2-op)

001

LY

010

SY

011

AMO

100

YADDI

101

MISC IMM (SRLIY/YBNDSWI)

110

Unallocated

111

Unallocated

A.4.2. 2. R-Type 3-Operand (funct3=000)

funct7[6:3] \ funct7[2:0] 000 001 010 011 100 101 110 111

0000

PACKY

YADD/YMV

YSH1ADD

YEQ

YSUNSEAL

0001

YADDRW

YSH2ADD

YSS

YBLD

0010

YPERMC

YSH3ADD

YSEAL

0011

YBNDSW

YSH4ADD

0100

YBNDSRW

YSH1ADD.UW

0101

YMODEW/YMODESWY/YMODESWI

YSH2ADD.UW

0110

YSH3ADD.UW

0111

YSH4ADD.UW

1000

1001

1010

1011

1100

1101

1110

1111

1OP/2OP

A.4.3. 3. R-Type 1-Op/2-Op (funct3=000, funct7=1111000)

Note: The 5-bit rs2 field is split into columns rs2[4:3] and rows rs2[2:0].

rs2[2:0] \ rs2[4:3] 00 01 10 11

000

YAMASK

001

010

011

100

101

110

111

A.4.4. 4. R-Type 1-Op/2-Op (funct3=000, funct7=1111010)

Note: The 5-bit rs2 field is split into columns rs2[4:3] and rows rs2[2:0].

rs2[2:0] \ rs2[4:3] 00 01 10 11

000

YBASER

001

YPERMR

010

YTOPR

011

YLENR

100

YTAGR

101

YTYPER

110

YMODER

111

A.5. 5. AMO Sub-opcode Allocations (funct3=100)

funct7[6:3] \ funct7[2:0] 000 001 010 011 100 101 110 111

0000

AMOSWAP.Y

AMOSWAP.Y

AMOSWAP.Y

AMOSWAP.Y

0001

LR.Y

LR.Y

LR.Y

LR.Y

SC.Y

SC.Y

SC.Y

SC.Y

0010

0011

0100

0101

0110

0111

1000

1001

1010

1011

1100

1101

1110

1111

A.6. 6. MISC Sub-opcode Allocations (funct3=101)

funct7[6:3] \ funct7[2:0] 000 001 010 011 100 101 110 111

0000

YHIR/SRLIY

YHIR/SRLIY

YHIR/SRLIY

YHIR/SRLIY

0001

0010

0011

0100

0101

0110

0111

1000

1001

1010

1011

1100

1101

1110

YBNDSWI

YBNDSWI

YBNDSWI

YBNDSWI

YBNDSWI

YBNDSWI

YBNDSWI

YBNDSWI

1111

YBNDSWI

YBNDSWI

YBNDSWI

YBNDSWI

YBNDSWI

YBNDSWI

YBNDSWI

YBNDSWI

Appendix B: Hybrid Mode Usage Models

The Zyhybrid extension enables a hart to switch between Capability Pointer Mode and Integral Pointer Mode. This functionality supports various usage models that facilitate the co-existence of and transition between non-CHERI software and pure-capability CHERI software.

B.1. High-Level Overview of Compatibility

There are three operating modes on a RVY-enabled core:

RVI/RVE mode (misa.Y=0 or xenvcfg.Y=0)

The base ISA is RVI or RVE. There is no access to extended YLEN registers. custom-1/2/3 major opcodes remain available for vendor extensions.

RVY pure-capability mode (misa.Y=1 and xenvcfg.Y=1, and pc.P-bit = 0)

The base ISA is RVY, and custom-1/2/3 major opcodes are used for RVY instructions. All loads and stores use a YLEN base register for permissions and bounds checks.

RVY in RVI compatibility mode (misa.Y=1 and xenvcfg.Y=1, and pc.P-bit = 1)

The base ISA is RVY, and custom-1/2/3 major opcodes are used for RVY instructions. However, all loads and stores use a XLEN base register and check ddc for permissions and bounds. This mode requires Zyhybrid support.

B.2. Example Usage Models

Purecap Kernel with 64-bit Plain RISC-V Programs

In this model, a CHERI-aware operating system (the kernel) executes in Capability Pointer Mode, benefiting from full memory safety for kernel operations. User applications may be standard 64-bit RISC-V programs executing in Integral Pointer Mode. The kernel can decide based on the ELF header whether to start the program with CHERI enabled and in pure-capability mode or completely disabled.

To set up this environment and handle transitions, the following sequence is typical:

  1. Initialize ddc for User Space: The kernel creates a data capability with read and write permissions. Its bounds are restricted to the user program’s data segment. This capability is written to the ddc CSR.

  2. Setup pc for User Space: The kernel creates a capability for the user program entry point. It sets the P-bit to Integral Pointer Mode so the processor starts in Integral Pointer Mode. This capability is written to sepc.

  3. Enter User Mode: The kernel executes the sret instruction. This installs the capability from sepc into pc, switching the mode to Integral Pointer Mode.

  4. Return to Kernel Mode: When the user program makes a system call (via ecall) or an error occurs, the hardware saves the current pc into sepc. It then loads the kernel’s exception vector capability from stvec into pc. Since stvec was set up with its P-bit set for pure-capability execution, this switches the execution mode back to Capability Pointer Mode.

    The kernel can disable CHERI for user mode by clearing the appropriate enable bits (e.g., bit in senvcfg.Y). This ensures that the user application cannot execute any CHERI instructions. It also guarantees that the program has access to the custom-1/2/3 opcode space, which CHERI would otherwise use for its own instructions in Integral Pointer Mode.

    To allow mixing plain RISC-V and pure-capability RISC-V user applications, the kernel must either context-switch the state of senvcfg.Y, or run the plain RISC-V applications in RVI compatibility mode, in which case the custom opcode space reserved by RVY is not available to userspace.
    kernel user transition
Plain RISC-V Program with a Purecap Library Sandbox

A large non-CHERI application (executing in Integral Pointer Mode) may wish to use a specific library that has been recompiled for CHERI (executing in Capability Pointer Mode) to process untrusted data (e.g., an image or video codec).

To set up this environment and handle transitions, the following sequence is typical:

  1. Prepare Library Capabilities: A dynamic linker or runtime creates a capability for the library’s entry point with the P-bit set to Capability Pointer Mode. This capability only covers the library’s code and read-only data section.

  2. Setup Sandbox Environment: Before running the library, the system sets ddc to NULL. This stops the library from accessing memory unless it uses specific capabilities given to it.

  3. Invoke the Library: The main program (in Integral Pointer Mode) calls the library using JALR (RVY). This installs the capability into pc and switches the mode to Capability Pointer Mode. It also saves a return capability in the link register.

  4. Return from the Library: When finished, the library returns by jumping to the return capability in the link register. This switches the mode back to Integral Pointer Mode.

    lib sandbox transition
Pure-capability RISC-V Program Calling into a Non-CHERI Library

Conversely, a new pure-capability application may need to link against a non-CHERI binary-only library for which source code is unavailable.

To set up this environment and handle transitions, the following sequence is typical:

  1. Prepare Non-CHERI Code Capability: The pure-capability program creates a capability for the non-CHERI library’s functions with the P-bit set to Integral Pointer Mode. This capability only covers the non-CHERI library’s code region.

  2. Setup ddc for Non-CHERI Code: In pure-capability code, ddc is usually NULL. Before calling the non-CHERI library, the program must load a capability into ddc. This capability must cover all memory the non-CHERI library needs (heap, globals, and stack).

  3. Invoke Non-CHERI Code: The pure-capability program calls the non-CHERI code using JALR (RVY). This switches the processor to Integral Pointer Mode and installs the restricted pc.

  4. Restore State on Return: When the non-CHERI library returns (usually using a standard JALR instruction), the pure-capability program restores its state and sets ddc back to NULL.

purecap calling noncheri
Hybrid Mode: RVI Code with Explicit Capabilities

For example, a Just-In-Time (JIT) compiler running in a pure-capability environment might generate code that uses standard integer addresses (or 32-bit compressed pointers) to avoid the overhead of larger capability pointers. While described here in the context of a JIT compiler and language runtime, this model is generalizable to other scenarios requiring efficient execution of code using smaller pointers within a larger pure-capability application.

This model sets up and switches modes just like the sandboxing models above. The main difference is that the generated code (running in Integral Pointer Mode) uses CHERI instructions to talk to the pure-capability environment:

  • Accessing C Runtime Data: The code can access data in the surrounding C++ runtime using explicit capability-based load/store operations. At the same time, it uses standard integer addresses for its own heap via ddc. This can be achieved by executing YMODESWY before the memory access and YMODESWI to return back to address interpretation.

  • Calling Runtime Helpers: Runtime helper functions are given to the code as a sealed entry point capability. When the code calls these helpers, the processor automatically switches back to Capability Pointer Mode. This lets the runtime run with full safety before returning to the JIT code.

    hybrid mode transition

B.3. CSRs and Registers Involved in Mode Transitions

When transitioning between modes, certain capability registers and CSRs must be configured to ensure correct execution and authorization:

ddc (Default Data Capability)

This CSR must be written with a capability covering the data memory that the code in Integral Pointer Mode is authorized to access. Memory accesses in Integral Pointer Mode are implicitly authorized by ddc.

Exception Return Capability Registers (sepc or mepc)

When a privileged environment prepares to transition to a lower privilege mode via an exception return instruction, it must write the target code capability to sepc or mepc. This capability must have its P-bit set to the desired execution mode (Capability Pointer Mode or Integral Pointer Mode). The exception return instruction installs this capability into pc, thereby setting the mode.

Trap Vector Registers (stvec or mtvec)

These registers hold the capability for the exception vector. When an exception or interrupt occurs, the hardware installs this capability into pc, thereby setting the mode for the trap handler. For example, if stvec was set up with its P-bit set to 0, the processor will switch to Capability Pointer Mode on entry to the trap handler.

B.4. CHERI Enable Bits in Setup (when changing privilege levels)

In addition to setting up capabilities and mode bits, the execution environment must manage the CHERI enable bits provided by the privileged architecture:

Per-Privilege Enable Bits

The privileged architecture defines enable bits (often denoted as Y) in CSRs such as misa, menvcfg, and senvcfg to control CHERI availability in each privilege mode (see Chapter 19).

Reset State

On reset, CHERI is disabled to ensure complete backwards compatibility with standard RISC-V software. This behavior applies to cores that support execution in both Capability Pointer Mode and Integral Pointer Mode.

Enabling CHERI

A higher privilege mode must set the appropriate enable bits to allow a lower privilege mode to use CHERI instructions or switch between Capability Pointer Mode and Integral Pointer Mode.

Disabled CHERI with Bounds Enforcement

If CHERI is disabled for a specific privilege mode, that mode cannot execute CHERI instructions or change modes. However, the hardware continues to enforce the bounds defined by the last installed pc and ddc. This allows a kernel to run non-CHERI-aware code in a strict integer environment while still confining it within capability bounds.

B.5. Transition Instructions and Frequency

Transitions between modes are performed using specific instructions:

YMODESWI and YMODESWY

These instructions are used to switch between modes without changing the control flow. They are typically used when entering or leaving a section of code that requires a different mode.

JALR (RVY)

When jumping to a target capability with the P-bit set differently from the current mode, the jump performs a mode switch. This is commonly used for function calls between compartments or libraries in different modes.

The frequency of transitions depends on the model:

  • Kernel/User boundary: Transitions occur on every system call, exception, and interrupt, which is relatively infrequent compared to instruction execution.

  • Library calls: Transitions occur on every call to and return from the sandboxed library. If these calls are frequent, the overhead of mode switching (which may involve saving/restoring registers and clearing state) should be considered. :leveloffset: +1

Appendix C: CHERI System Implications

Unclear if this chapter will appear in the priv spec. May just be in the standalone spec.

CHERI processors need memory systems which support the capability tags in memory.

There are, or will soon be, a wide range of CHERI systems in existence from tiny IoT devices up to server chips.

There are two types of bus connections used in SoCs which contain CHERI CPUs:

  1. Tag-aware buses, where the bus protocol is extended to carry the capability tag along with the data. This is typically done using user-defined bits in the protocol.

    1. These buses will read capability tags from memory (if capability tags are present in the target memory) and return them to the requestor.

    2. These buses will write the capability tag to memory as an extension of the data write.

  2. Non-capability tag-aware buses, i.e., current non-CHERI-aware buses.

    1. Reads of tagged memory will not read the capability tag.

    2. Writes to tagged memory will set the capability tag of any YLEN-aligned YLEN-wide memory location to zero where any byte matches the memory write.

The fundamental rule for any CHERI system is that the capability tag and data are always accessed atomically. For every naturally aligned YLEN-wide memory location, it must never be possible to:

  1. Update any data bytes without also writing the capability tag

    1. This implies setting the capability tag to zero if a non-CHERI aware bus master overwrites a capability in memory

  2. Read a capability tag value with mismatched (stale or newer) data

  3. Write a capability tag without also writing the data.

Clearing capability tags in memory does not necessarily require updating the associated data.

C.1. Small CHERI system example

small cheri system.drawio
Figure 28. Example small CHERI system with local capability tag storage

This example shows a minimum-sized system where only the local memory is extended to support capability tags. The capability tag-aware region is highlighted. All capability tags are created by the CHERI CPU, and only stored locally. The memory is shared with the system, probably via a secure DMA, which is not capability tag-aware.

Therefore the connection between CPU and memory is tag-aware, and the connection to the system is not capability tag-aware.

All writes from the system port to the memory must clear any memory capability tags to follow the rules from above.

C.2. Large CHERI system example

large cheri system.drawio
Figure 29. Example large CHERI system with capability tag cache

In the case of a large CHERI SoC with caches, all the cached memory visible to the CHERI CPUs must support capability tags. All memory is backed up by DRAM, and standard DRAM does not offer the extra bit required for CHERI capability tag storage and so a typical system will have a capability tag cache IP.

A region of DRAM is typically reserved for CHERI capability tag storage.

The capability tag cache sits on the boundary of the capability tag-aware and non-tag-aware memory domains, and it provides the bridge between the two. It stores capability tags locally in its cache, and if there is a miss, it will create an extra bus request to access the region of DRAM reserved for capability tag storage. Therefore in the case of a miss a single access is split into two - one to access the data and one to access the capability tag.

The key property of the capability tag cache is to preserve the atomic access of data and capability tags in the memory system so that all CPUs have a consistent view of capability tags and data.

The region of DRAM reserved for capability tag storage must be accessible only by the capability tag cache, therefore no bus initiators should be able to write to the DRAM without the transactions passing through the capability tag cache.

Therefore the GPUs and peripherals cannot write to the capability tag storage in the DRAM or the capability tag supporting memory data storage region. These constraints will be part of the design of the network-on-chip. It is possible for the GPU and peripherals to read the capability tag supporting memory data storage region of the DRAM, if required.

It would be possible to allow a DMA to access the capability tagged memory region of the DRAM directly to allow swap to/from DRAM and external devices such as flash. This will require the highest level of security in the SoC, as the CHERI protection model relies on the integrity of the capability tags, and so the root-of-trust will need to authenticate and encrypt the transfer, with anti-rollback protection.

For further information on the capability tag cache see (Efficient Tagged Memory, 2017).

C.3. Large CHERI pure-capability system example

large cheri purecap system.drawio
Figure 30. Example large CHERI system with only tag-aware bus masters

In this example every DRAM access passes through the capability tag cache, and so all bus masters are capability tag-aware and can access the capability tags associated with memory if permitted by the network-on-chip.

The system topology is simpler than in Figure 29.

There is likely to be a performance difference between the two systems. The main motivation for Figure 29 is to avoid the GPU DRAM traffic needing to look-up every capability tag in the capability tag cache, potentially adding overhead to every transaction.

Appendix D: Integrating RVY with Debug

D.1. Integrating RVY with Sdtrig

This chapter will appear in the priv spec. Exact location TBD.

The Sdtrig extension is orthogonal to RVY. However, the priority of synchronous exceptions and where triggers fit is adjusted as shown in Table 65.

Debug triggers are higher priority than CHERI exceptions to allow debug.

Table 65. Synchronous exception priority in decreasing priority order.
Priority Exception Code Description Trigger

Highest

3
3
3
3

etrigger
icount
itrigger
mcontrol/mcontrol6 after (on previous instruction)

3

Instruction address breakpoint

mcontrol/mcontrol6 execute address before

32

Prior to instruction address translation:
CHERI Instruction Access Fault due to pc checks (capability tag, sealed, execute permission, bounds1)

12, 20, 1

During instruction address translation: First encountered page fault, guest-page fault, or access fault

1

With physical address for instruction: Instruction access fault

3

mcontrol/mcontrol6 execute data before

2
32
22
0
8, 9, 10, 11
3
3

Illegal instruction
CHERI Instruction Access Fault due to pc ASR-permission clear
Virtual instruction
Instruction address misaligned
Environment call
Environment break
Load/Store/AMO address breakpoint

mcontrol/mcontrol6 load/store address before, store data before

33,34

Prior to address translation for an explicit memory access:
CHERI Load Access Fault, CHERI Store/AMO Access Fault due to capability checks (capability tag, sealed, permissions, bounds)

4,6

Load/store/AMO capability address misaligned

4, 6

Optionally: Load/Store/AMO address misaligned

36, 13, 15, 21, 23, 5, 7

During address translation for an explicit memory access: First encountered CHERI Store/AMO Page Fault, page fault, guest-page fault, or access fault

5, 7

With physical address for an explicit memory access: Load/store/AMO access fault

4, 6

If not higher priority: Load/store/AMO address misaligned

35

CHERI Load Capability Fault

Lowest

3

mcontrol/mcontrol6 load data before

D.2. Integrating RVY with Sdext

This chapter will appear in the priv spec. Exact location TBD.

This section describes changes to integrate the Sdext ISA and RVY. It must be implemented to make external debug compatible with RVY. Modifications to Sdext are kept to a minimum.

The following features, which are optional in Sdext, must be implemented for use with RVY:

  • The hartinfo register must be implemented.

  • All harts which support RVY must provide hartinfo.nscratch of at least 1 and implement the dscratch0 (RVY) register.

  • All harts which support RVY must provide hartinfo.datasize of at least 1 and hartinfo.dataaccess of 0.

  • The program buffer must be implemented, with abstractcs.progbufsize of at least 4 if dmstatus.impebreak is 1, or at least 5 if dmstatus.impebreak is 0.

These requirements allow a debugger to read and write capabilities in integer registers without disturbing other registers. These requirements may be relaxed if some other means of accessing capabilities in integer registers, such as an extension of the Access Register abstract command, is added.

The following sequences demonstrate how a debugger can read and write a capability in x1 if MXLEN is 64, hartinfo.dataaccess is 0, hartinfo.dataaddr is 0xBF0, hartinfo.datasize is 1, hartinfo.nscratch is at least 1, dmstatus.impebreak is 0, and abstractcs.progbufsize is 5:

# Read the high MXLEN bits into data0-data1
csrrw  x2, dscratch0, x2
yhir   x2, x1
csrw   0xBF0, x2
csrrw  x2, dscratch0, x2
ebreak

# Read the capability tag into data0
csrrw  x2, dscratch0, x2
ytagr  x2, x1
csrw   0xBF0, x2
csrrw  x2, dscratch0, x2
ebreak

# Write the high MXLEN bits from data0-data1
csrrw  x2, dscratch0, x2
csrr   x2, 0xBF0
yhiw   x1, x1, x2
csrrw  x2, dscratch0, x2
ebreak

# Write the capability tag (if nonzero)
csrrw   x2, dscratch0, x2
csrr    x2, drootc
ybld    x1, x2, x1
csrrw   x2, dscratch0, x2
ebreak

The low MXLEN bits of a capability are read and written using normal Access Register abstract commands.

The above instruction sequences utilize these guarantees provided by the RISC-V Debug Specification:

  • When hartinfo.nscratch is greater than 0, the dscratch0 (RVY) is guaranteed to retain its value within a single abstract command.

    If dscratch0 (RVY) were known to be preserved between abstract commands, it would be possible to remove the requirements on hartinfo.datasize, hartinfo.dataaccess, and abstractcs.progbufsize, however, there is no way to discover the former property.

  • Execution of Access Register abstract commands does not alter the values of data debug module registers (unless read of a register is explicitly requested by setting transfer to 1 and write to 0).

D.2.1. Debug Mode

When executing code due to an abstract command, the hart stays in debug mode and the rules outlined in Section 4.1 of the RISC-V Debug Specification apply.

D.2.2. Core Debug Registers

RVY widens debug CSRs that are designated to hold addresses so that they can hold capabilities.

The pc must grant ASR-permission to access debug CSRs. This permission is automatically provided when the hart enters debug mode as described in the dpc (RVY) section. The pc metadata can only be changed if the implementation supports executing control transfer instructions from the program buffer — this is an optional feature according to the RISC-V Debug Specification.

This specification extends the following registers from the RISC-V Debug Specification.

Debug Program Counter (dpc)

dpc is a DXLEN-bit register used as the PC saved when entering debug mode.

Diagram
Figure 31. Debug program counter
Debug Scratch Register 0 (dscratch0)

dscratch0 is an optional DXLEN-bit scratch register that can be used by implementations which need it.

Diagram
Figure 32. Debug scratch 0 register
Debug Scratch Register 1 (dscratch1)

dscratch1 is an optional DXLEN-bit scratch register that can be used by implementations which need it.

Diagram
Figure 33. Debug scratch 1 register

D.2.3. Debug Program Counter Capability (dpc)

The dpc register is widened so that it can hold a capability.

The reset value of the capability tag of this CSR is zero, the reset values of the metadata and address fields are UNSPECIFIED.

Diagram
Figure 34. Debug program counter capability

When the hart is in debug mode, the RISC-V Debug Specification does not specify how the PC is updated, and says that PC-relative instructions may be illegal. This concept is extended to include any instruction which reads or updates pc, which refers to all jumps, conditional branches and AUIPC (RVY). The exceptions are YMODESWY and YMODESWI, which are supported if Zyhybrid is implemented, see drootc for details.

As a result, the value of pc is UNSPECIFIED in debug mode according to this specification. The pc metadata has no architectural effect in debug mode. Therefore ASR-permission is implicitly granted for access to all CSRs for instruction execution.

On debug mode entry, dpc (RVY) is updated with the capability in pc, whose address field equals to the address of the next instruction to be executed upon debug mode exit as described in the RISC-V Debug Specification.

When leaving debug mode, the value in dpc (RVY) is unsealed if it is a sealed entry point capability and is written to pc.

A debugger may write dpc (RVY) to change where the hart resumes and its mode, permissions, sealing or bounds. The value saved from the pc on debug mode entry cannot be sealed.

The legalization of dpc (RVY) follows the same rules described for mepc (RVY).

One possible implementation choice for debug mode exit is DRET.

D.2.3.1. DRET (RVY)
Synopsis

Debug Return (DRET)

Mnemonic

dret

Encoding
Diagram
Description

DRET (RVY) returns from debug mode. It unseals dpc (RVY) if it is a sealed entry point capability and writes the result into pc. It also restores ddc from dddc.

The DRET (RVY) instruction is a documented method of exiting debug mode. However, it is a pseudoinstruction to return that technically does not execute from the program buffer or memory. It does not require the pc to grant ASR-permission so it never raises an exception.
The definition of DRET is not part of Sdext, and is a suggested as an implementation choice in the RISC-V Debug Specification. This specification assumes the recommended specification including the recommended encoding.
Prerequisites

RVY

Operation
TBD

D.2.4. Debug Scratch Register 0 (dscratch0)

The dscratch0 register is widened so that it can hold a capability.

The reset value of the capability tag of this CSR is zero, the reset values of the metadata and address fields are UNSPECIFIED.

Diagram
Figure 35. Debug scratch 0 capability register

D.2.5. Debug Scratch Register 1 (dscratch1)

The dscratch1 register is widened so that it can hold a capability.

The reset value of the capability tag of this CSR is zero, the reset values of the metadata and address fields are UNSPECIFIED.

Diagram
Figure 36. Debug scratch 1 capability register

D.2.6. Debug Root Capability Selector (drootcsel)

drootcsel is a DXLEN-bit debug mode accessible integer CSR. The address and access details are shown in Table 68.

It selects which Root capability is exposed through drootc. The reset value is 0, which must cause drootc to expose a Root Executable capability.

Other capability values may be defined for exposure through drootc by the capability encoding, and may be selected by having the debugger write to this register. Writes are WARL, so the debugger may confirm that its selection has been applied.

Diagram
Figure 37. Debug root capability register

D.2.7. Debug Root Capability Register (drootc)

drootc is a debug mode accessible capability CSR. The address and access details are shown in Table 68. It exposes the capability selected by drootcsel.

If Zyhybrid is implemented, the Root Executable exposed when drootcsel is 0 is further specified as follows:

  • The P-bit is reset to Integral Pointer Mode (1).

  • The debugger can set the P-bit to Capability Pointer Mode (0) by executing YMODESWY from the program buffer.

    • Executing YMODESWY causes execution of subsequent instructions from the program buffer, starting from the next instruction, to be executed in Capability Pointer Mode. It also sets the CHERI Pointer Mode to Capability Pointer Mode on future entry into debug mode.

    • Therefore to enable use of a CHERI debugger, only a single YMODESWY needs to be executed once from the program buffer after resetting the core.

    • The debugger can also execute YMODESWI to change the CHERI Pointer Mode back to Integral Pointer Mode. This affects the execution of the next instruction in the program buffer, updates the P-bit of drootc, and thereby controls which CHERI Pointer Mode to use on the next entry into debug mode.

The P-bit of this capability is only updated by executing YMODESWY or YMODESWI from the program buffer.

Diagram
Figure 38. Debug root capability register

D.3. Integrating Zyhybrid with Sdext

A new debug default data capability (dddc) CSR is added at the CSR number shown in Table 48.

Zyhybrid allows YMODESWY and YMODESWI to execute in debug mode.

When entering debug mode, whether the core enters Integral Pointer Mode or Capability Pointer Mode is controlled by the P-bit in the drootc capability selected by drootcsel value 0.

The current mode can be read by setting drootcsel to 0 and then reading drootc.

The following sequence executed from the program buffer will write 0 for Capability Pointer Mode and 1 for Integral Pointer Mode to x1:

csrr   x1, drootc
ymoder x1, x1
There is no CHERI enable/disable bit for debug mode, so CHERI register and instruction access is always permitted in debug mode.

D.3.1. Debug Default Data Capability CSR (dddc)

dddc is a debug mode accessible capability CSR. The address is shown in Table 48.

The reset value of the capability tag of this CSR is zero, the reset values of the metadata and address fields are UNSPECIFIED.

This CSR is only implemented if Zyhybrid is implemented.

Diagram
Figure 39. Debug default data capability

Upon entry to debug mode, ddc is saved in dddc. ddc is set to a Root Data capability such that ddc's address remains unchanged.

When debug mode is exited by executing DRET (RVY), the hart’s ddc is updated to the capability stored in dddc. A debugger may write dddc to change the When debug mode is exited, the hart’s ddc is updated to the capability stored in dddc. A debugger may write dddc to change the hart’s context.

As shown in Table 67, dddc is a data pointer, so it does not need to be able to hold all possible invalid addresses (see Invalid address conversion).

D.4. Debug-mode YLEN CSR Summary

The tables below show all Supervisor-mode CSRs added or modified by RVY.

Table 66. Action taken on writing to extended Debug-Mode CSRs
RVY CSR Width Action on XLEN write Action on YLEN write

dpc (RVY)

YLEN

Apply Invalid address conversion. Always update the CSR with YADDRW even if the address didn’t change.

Apply Invalid address conversion and update the CSR with the result if the address changed, direct write if address didn’t change

dscratch0 (RVY)

YLEN

Update the CSR using YADDRW.

direct write

dscratch1 (RVY)

YLEN

Update the CSR using YADDRW.

direct write

dddc

YLEN

Apply Invalid address conversion. Always update the CSR with YADDRW even if the address didn’t change.

Apply Invalid address conversion and update the CSR with the result if the address changed, direct write if address didn’t change

drootc

YLEN

Ignore

Ignore

Code pointers and data pointers in Table 67 are WARL CSRs that do not need to store full 64-bit addresses on RV64, and so may be subject to Invalid address conversion on writing.

Table 67. YLEN-wide Debug-Mode CSRs storing code pointers or data pointers
RVY CSR Code Pointer Data Pointer Unseal On Execution

dpc (RVY)

✔

✔

dddc

✔

Table 68 shows all debug-mode CSRs added or modified by RVY.

Table 68. All RVY Debug-Mode CSRs.
RVY CSR Prerequisites Width Address Permissions Reset Value Description

dpc (RVY)

Sdext

YLEN

0x7b1

DRW

tag=0, otherwise specified by the platform

Debug Program Counter Capability

dscratch0 (RVY)

Sdext

YLEN

0x7b2

DRW

tag=0, otherwise specified by the platform

Debug Scratch Capability 0

dscratch1 (RVY)

Sdext

YLEN

0x7b3

DRW

tag=0, otherwise specified by the platform

Debug Scratch Capability 1

dddc

Zyhybrid, Sdext

YLEN

0x7bc

DRW

tag=0, otherwise specified by the platform

Debug Default Data Capability (saved/restored on debug mode entry/exit)

drootcsel

Sdext

DXLEN

0x7ba

DRW

0

Multiplexing selector for drootc

drootc

Sdext

YLEN

0x7bd

DRW

nominally Root Executable

Source of authority in debug mode, writes are ignored
Where reset values are specified as a Root Executable or Root Data they are a maximum possible value. The platform may reset those CSRs to a smaller memory range, or to have fewer permissions.

Index

Bibliography

Efficient Tagged Memory. (2017). www.cl.cam.ac.uk/research/security/ctsrd/pdfs/201711-iccd2017-efficient-tags.pdf

Filardo, N. W., Gutstein, B. F., Woodruff, J., Clarke, J., Rugg, P., Davis, B., Johnston, M., Norton, R., Chisnall, D., Moore, S. W., Neumann, P. G., & Watson, R. N. M. (2024). Cornucopia Reloaded: Load Barriers for CHERI Heap Temporal Safety. doi.org/10.1145/3620665.3640416

RISC-V. (2023). RISC-V Unprivileged Specification. github.com/riscv/riscv-isa-manual/releases/download/riscv-isa-release-056b6ff-2023-10-02/unpriv-isa-asciidoc.pdf

Watson, R. N. M., Neumann, P. G., Woodruff, J., Roe, M., Almatary, H., Anderson, J., Baldwin, J., Barnes, G., Chisnall, D., Clarke, J., Davis, B., Eisen, L., Filardo, N. W., Fuchs, F. A., Grisenthwaite, R., Joannou, A., Laurie, B., Markettos, A. T., Moore, S. W., … Xia, H. (2023). Capability Hardware Enhanced RISC Instructions: CHERI Instruction-Set Architecture (Version 9) (UCAM-CL-TR-987; Issue UCAM-CL-TR-987). University of Cambridge, Computer Laboratory. doi.org/10.48456/tr-987

Woodruff, J., Joannou, A., Xia, H., Fox, A., Norton, R. M., Chisnall, D., Davis, B., Gudka, K., Filardo, N. W., Markettos, A. T., & others. (2019). Cheri Concentrate: Practical compressed capabilities. IEEE Transactions on Computers, 68(10), 1455–1469. doi.org/10.1109/TC.2019.2914037