Skip to content

Latest commit

 

History

History
 
 

CIP-0122

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
CIP Title Category Status Authors Implementors Discussions Created License
122
Logical operations over BuiltinByteString
Plutus
Proposed
2024-05-03
Apache-2.0

Abstract

We describe the semantics of a set of logical operations for Plutus BuiltinByteStrings. Specifically, we provide descriptions for:

  • Bitwise logical AND, OR, XOR and complement;
  • Reading a bit value at a given index;
  • Setting bits value at given indices; and
  • Replicating a byte a given number of times.

As part of this, we also describe the bit ordering within a BuiltinByteString, and provide some laws these operations should obey.

Motivation: why is this CIP necessary?

Bitwise operations, both over fixed-width and variable-width blocks of bits, have a range of uses, including data structures (especially succinct ones) and cryptography. Currently, operations on individual bits in Plutus Core are difficult, or outright impossible, while also keeping within the tight constraints required onchain. While it is possible to some degree to work with individual bytes over BuiltinByteStrings, this isn't sufficient, or efficient, when bit maniputations are required.

To demonstrate where bitwise operations would allow onchain possibilities that are currently either impractical or impossible, we give the following use cases.

Case 1: integer set

An integer set (also known as a bit set, bitmap, or bitvector) is a succinct data structure for representing a set of numbers in a pre-defined range $[0, n)$ for some $n \in \mathbb{N}$. The structure supports the following operations:

  • Construction given a fixed number of elements, as well as the bound $n$.
  • Construction of the empty set (contains no elements) and the universe (contains all elements).
  • Set union, intersection, complement and difference (symmetric and asymmetric).
  • Membership testing for a specific element.
  • Inserting or removing elements.

These structures have a range of uses. In addition to being used as sets of bounded natural numbers, an integer set could also represent an array of Boolean values. These have a range of applications, mostly as 'backends' for other, more complex structures. Furthermore, by using some index arithmetic, integer sets can also be used to represent binary matrices (in any number of dimensions), which have an even wider range of uses:

The succinctness of the integer set (and the other succinct data structures it enables) is particularly valuable on-chain, due to the limited transaction size and memory available.

Typically, such a structure would be represented as a packed array of bytes (similar to the Haskell ByteString). Essentially, given a bound $n$, the packed array has a length in bytes large enough to contain at least $n$ bits, with a bit at position $i$ corresponding to the value $i \in \mathbb{N}$. This representation ensures the succinctness of the structure (at most 7 bits of overhead are required if $n = 8k + 1$ for some $k \in \mathbb{N}$), and also allows all the above operations to be implemented efficiently:

  • Construction given a fixed number of elements and the bound $n$ involves allocating the packed array, then modifying some bits to be set.
  • Construction of the empty set is a packed array where every byte is 0x00, while the universe is a packed array where every byte is 0xFF.
  • Set union is bitwise OR over both arguments.
  • Set intersection is bitwise AND over both arguments.
  • Set complement is bitwise complement over the entire packed array.
  • Symmetric set difference is bitwise XOR over both arguments; asymmetric set difference can be defined using a combination of bitwise complement and bitwise OR.
  • Membership testing is checking whether a bit is set.
  • Inserting an element is setting the corresponding bit.
  • Removing an element is clearing the corresponding bit.

Given that this is a packed representation, these operations can be implemented very efficiently by relying on the cache-friendly properties of packed array traversals, as well as making use of optimized routines available in many languages. Thus, this structure can be used to efficiently represent sets of numbers in any bounded range (as ranges not starting from $0$ can be represented by storing an offset), while also being minimal in space usage.

Currently, such a structure cannot be easily implemented in Plutus Core while preserving the properties described above. The two options using existing primitives are either to use [BuiltinInteger], or to mimic the above operations over BuiltinByteString. The first of these is not space or time-efficient: each BuiltinInteger takes up multiple machine words of space, and the list overheads introduced are linear in the number of items stored, destroying succinctness; membership testing, insertion and removal require either maintaining an ordered list or forcing linear scans for at least some operations, which are inefficient over lists; and 'bulk' operations like union, intersection and complement become very difficult and time-consuming. The second is not much better: while we preserve succinctness, there is no easy way to access individual bits, only bytes, which would require a division-remainder loop for each such operation, with all the overheads this imposes; intersection, union and symmetric difference would have to be simulated byte-by-byte, requiring large lookup tables or complex conditional logic; and construction would require immense amounts of copying and tricky byte construction logic. While it is not outright impossible to make such a structure using current primitives, it would be so impractical that it could never see real use.

Furthermore, for sparse (or dense) integer sets (that is, where either most elements in the range are absent or present respectively), a range of compression techniques have been developed. All of these rely on bitwise operations to achieve their goals, and can potentially yield significant space savings in many cases. Given the limitations onchain that we have to work within, having such techniques available to implementers would be a huge potential advantage.

Case 2: hashing

Hashing, that is, computing a fixed-length 'fingerprint' or 'digest' of a variable-length input (typically viewed as binary) is a common task required in a range of applications. Most notably, hashing is a key tool in cryptographic protocols and applications, either in its own right, or as part of a larger task. The value of such functionality is such that Plutus Core already contains primitives for certain hash functions, specifically two variants of SHA256 and BLAKE2b. At the same time, hash functions choices are often determined by protocol or use case, and providing individual primitives for every possible hash function is not a scalable choice. It is much preferrable to give necessary tools to implement such functionality to users of Plutus (Core), allowing them to use whichever hash function(s) their applications require.

As an example, we consider the Argon2 family of hash functions. In order to implement any variant of this family requires the following operations:

  1. Conversion of numbers to bytes
  2. Bytestring concatenation
  3. BLAKE2b hashing
  4. Floor division
  5. Indexing bytes in a bytestring
  6. Logical XOR

Operations 1 to 5 are already provided by Plutus Core (with 1 being included in CIP-121); however, without logical XOR, no function in the Argon2 family could be implemented. While in theory, it could be simulated with what operations already exist, much as with Case 1, this would be impractical at best, and outright impossible at worst, due to the severe limits imposed on-chain. This is particularly the case here, as all Argon2 variants call logical XOR in a loop, whose step count is defined by multiple user-specified (or protocol-specified) parameters.

We observe that this requirement for logical XOR is not unique to the Argon2 family of hash functions. Indeed, logical XOR is widely used for a variety of cryptographic applications, as it is a low-cost mixing function that happens to be self-inverting, as well as preserving randomness (that is, a random bit XORed with a non-random bit will give a random bit).

Specification

We describe the proposed operations in several stages. First, we specify a scheme for indexing individual bits (rather than whole bytes) in a BuiltinByteString. We then specify the semantics of each operation, as well as giving costing expectations and some examples. Lastly, we provide some laws that any implementation of these operations is expected to obey.

Bit indexing scheme

We begin by observing that a BuiltinByteString is a packed array of bytes (that is, BuiltinIntegers in the range $[0, 255]$) according to the API provided by existing Plutus Core primitives. In particular, we have the ability to access individual bytes by index as a primitive operation. Thus, we can view a BuiltinByteString as an indexed collection of bytes; for any BuiltinByteString $b$ of length $n$, and any $i \in 0, 1, \ldots, n - 1$, we define $b\{i\}$ as the byte at index $i$ in $b$, as defined by the indexByteString primitive. In essence, for any BuiltinByteString of length n, we have byte indexes as follows:

| Index | 0  | 1  | ... | n - 1    |
|-------|----|----| ... |----------|
| Byte  | w0 | w1 | ... | w(n - 1) |

To view a BuiltinByteString as an indexed collection of bits, we must first consider the bit ordering within a byte. Suppose $i \in 0, 1, \ldots, 7$ is an index into a byte $w$. We say that the bit at $i$ in $w$ is set when

$$ \left \lfloor \frac{w}{2^{i}} \right \rfloor \mod 2 \equiv 1 $$

Otherwise, the bit at $i$ in $w$ is clear. We define $w[i]$ to be $1$ when the bit at $i$ in $w$ is set, and $0$ otherwise; this is the value at index $i$ in $w$.

For example, consider the byte represented by the BuiltinInteger 42. By the above scheme, we have the following:

Bit index Set or clear?
$0$ Clear
$1$ Set
$2$ Clear
$3$ Set
$4$ Clear
$5$ Set
$6$ Clear
$7$ Clear

Put another way, we can view $w[i] = 1$ to mean that the $(i + 1)$ th least significant digit in $w$'s binary representation is $1$, and likewise, $w[i] = 0$ would mean that the $i$th least significant digit in $w$'s binary representation is $0$. Continuing with the above example, $42$ is represented in binary as 00101010; we can see that the second-least-significant, fourth-least-significant, and sixth-least-significant digits are 1, and all the others are zero. This description mirrors the way bytes are represented on machine architectures.

We now extend the above scheme to BuiltinByteStrings. Let $b$ be a BuiltinByteString whose length is $n$, and let $i \in 0, 1, \ldots, 8 \cdot n - 1$. For any $j \in 0, 1, \ldots, n - 1$, let $j^{\prime} = n - j - 1$. We say that the bit at $i$ in $b$ is set if

$$ b\left\{\left(\left\lfloor \frac{i}{8} \right\rfloor\right)^{\prime}\right\}[i\mod 8] = 1 $$

We define the bit at $i$ in $b$ being clear analogously. Similarly to bits in a byte, we define $b[i]$ to be $1$ when the bit at $i$ in $b$ is set, and $0$ otherwise; similarly to bytes, we term this the value at index $i$ in $b$.

As an example, consider the BuiltinByteString [42, 57, 133]: that is, the BuiltinByteString $b$ such that $b\{0\} = 42$, $b\{1\} = 57$ and $b\{2\} = 133$. We observe that the range of 'valid' bit indexes $i$ into $b$ is in $[0, 3 \cdot 8 - 1 = 23]$. Consider $i = 4$; by the definition above, this corresponds to the byte index 2, as $\left\lfloor\frac{4}{8}\right\rfloor = 0$, and $3 - 0 - 1 = 2$ (as $b$ has length $3$). Within the byte $133$, this means we have $\left\lfloor\frac{133}{2^4}\right\rfloor \mod 2 \equiv 0$. Thus, $b[4] = 0$. Consider instead the index $i = 19$; by the definition above, this corresponds to the byte index 0, as $\left\lfloor\frac{19}{8}\right\rfloor = 2$, and $3 - 2 - 1 = 0$. Within the byte $42$, this means we have $\left\lfloor\frac{42}{2^3}\right\rfloor\mod 2 \equiv 1$. Thus, $b[19] = 1$.

Put another way, our byte indexes run 'the opposite way' to our bit indexes. Thus, for any BuiltinByteString of length $n$, we have bit indexes relative byte indexes as follows:

| Byte index | 0                              | 1  | ... | n - 1                         |
|------------|--------------------------------|----| ... |-------------------------------|
| Byte       | w0                             | w1 | ... | w(n - 1)                      |
|------------|--------------------------------|----| ... |-------------------------------|
| Bit index  | 8n - 1 | 8n - 2 | ... | 8n - 8 |   ...    | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |

Operation semantics

We describe precisely the operations we intend to implement, and their semantics. These operations will have the following signatures:

  • bitwiseLogicalAnd :: BuiltinBool -> BuiltinByteString -> BuiltinByteString -> BuiltinByteString
  • bitwiseLogicalOr :: BuiltinBool -> BuiltinByteString -> BuiltinByteString -> BuiltinByteString
  • bitwiseLogicalXor :: BuiltinBool -> BuiltinByteString -> BuiltinByteString -> BuiltinByteString
  • bitwiseLogicalComplement :: BuiltinByteString -> BuiltinByteString
  • readBit :: BuiltinByteString -> BuiltinInteger -> BuiltinBool
  • writeBits :: BuiltinByteString -> [(BuiltinInteger, BuiltinBool)] -> BuiltinByteString
  • replicateByteString :: BuiltinInteger -> BuiltinInteger -> BuiltinByteString

We assume the following costing, for both memory and execution time:

Operation Cost
bitwiseLogicalAnd Linear in longest BuiltinByteString argument
bitwiseLogicalOr Linear in longest BuiltinByteString argument
bitwiseLogicalXor Linear in longest BuiltinByteString argument
bitwiseLogicalComplement Linear in BuiltinByteString argument
readBit Constant
writeBits Additively linear in both arguments
replicateByteString Linear in the value of the first argument

Padding versus truncation semantics

For the binary logical operations (that is, bitwiseLogicalAnd, bitwiseLogicalOr and bitwiseLogicalXor), the we have two choices of semantics when handling BuiltinByteString arguments of different lengths. We can either produce a result whose length is the minimum of the two arguments (which we call truncation semantics), or produce a result whose length is the maximum of the two arguments (which we call padding semantics). As these can both be useful depending on context, we allow both, controlled by a BuiltinBool flag, on all the operations listed above.

In cases where we have arguments of different lengths, in order to produce a result of the appropriate lengths, one of the arguments needs to be either padded or truncated. Let short and long refer to the BuiltinByteString argument of shorter length, and of longer length, respectively. The following table describes what happens to the arguments before the operation:

Semantics short long
Padding Pad at high byte indexes Unchanged
Truncation Unchanged Truncate high byte indexes

We pad with different bytes depending on operation: for bitwiseLogicalAnd, we pad with 0xFF, while for bitwiseLogicalOr and bitwiseLogicalXor we pad with 0x00 instead. We refer to arguments so changed as semantics-modified arguments.

For example, consider the BuiltinByteStrings x = [0x00, 0xF0, 0xFF] and y = [0xFF, 0xF0]. The following table describes what the semantics-modified versions of these arguments would become for each operation and each semantics:

Operation Semantics x y
bitwiseLogicalAnd Padding [0x00, 0xF0, 0xFF] [0xFF, 0xF0, 0xFF]
bitwiseLogicalAnd Truncation [0x00, 0xF0] [0xFF, 0xF0]
bitwiseLogicalOr Padding [0x00, 0xF0, 0xFF] [0xFF, 0xF0, 0x00]
bitwiseLogicalor Truncation [0x00, 0xF0] [0xFF, 0xF0]
bitwiseLogicalXor Padding [0x00, 0xF0, 0xFF] [0xFF, 0xF0, 0x00]
bitwiseLogicalXor Truncation [0x00, 0xF0] [0xFF, 0xF0]

Based on the above, we observe that under padding semantics, the result of any of the listed operations would have a byte length of 3, while under truncation semantics, the result would have a byte length of 2 instead.

bitwiseLogicalAnd

bitwiseLogicalAnd takes three arguments; we name and describe them below.

  1. Whether padding semantics should be used. If this argument is False, truncation semantics are used instead. This is the padding semantics argument, and has type BuiltinBool.
  2. The first input BuiltinByteString. This is the first data argument.
  3. The second input BuiltinByteString. This is the second data argument.

Let $b_1, b_2$ refer to the semantics-modified first data argument and semantics-modified second data argument respectively, and let $n$ be either of their lengths in bytes; see the section on padding versus truncation semantics for the exact specification of this. Let the result of bitwiseLogicalAnd, given $b_1, b_2$ and some padding semantics argument, be $b_r$, also of length $n$ in bytes. We use $b_1\{i\}$ to refer to the byte at index $i$ in $b_1$ (and analogously for $b_2$, $b_r$); see the section on the bit indexing scheme for the exact specification of this.

For all $i \in 0, 1, \ldots, n - 1$, we have $b_r\{i\} = b_0\{i\} \text{ }\& \text{ } b_1\{i\}$, where $\&$ refers to a bitwise AND.

Some examples of the intended behaviour of bitwiseLogicalAnd follow. For brevity, we write BuiltinByteString literals as lists of hexadecimal values.

-- truncation semantics
bitwiseLogicalAnd False [] [0xFF] => []

bitwiseLogicalAnd False [0xFF] [] => []

bitwiseLogicalAnd False [0xFF] [0x00] => [0x00]

bitwiseLogicalAnd False [0x00] [0xFF] => [0x00]

bitwiseLogicalAnd False [0x4F, 0x00] [0xF4] => [0x44]

-- padding semantics
bitwiseLogicalAnd True [] [0xFF] => [0xFF]

bitwiseLogicalAnd True [0xFF] [] => [0xFF]

bitwiseLogicalAnd True [0xFF] [0x00] => [0x00]

bitwiseLogicalAnd True [0x00] [0xFF] => [0x00]

bitwiseLogicalAnd True [0x4F, 0x00] [0xF4] => [0x44, 0x00]

bitwiseLogicalOr

bitwiseLogicalOr takes three arguments; we name and describe them below.

  1. Whether padding semantics should be used. If this argument is False, truncation semantics are used instead. This is the padding semantics argument, and has type BuiltinBool.
  2. The first input BuiltinByteString. This is the first data argument.
  3. The second input BuiltinByteString. This is the second data argument.

Let $b_1, b_2$ refer to the semantics-modified first data argument and semantics-modified second data argument respectively, and let $n$ be either of their lengths in bytes; see the section on padding versus truncation semantics for the exact specification of this. Let the result of bitwiseLogicalOr, given $b_1, b_2$ and some padding semantics argument, be $b_r$, also of length $n$ in bytes. We use $b_1\{i\}$ to refer to the byte at index $i$ in $b_1$ (and analogously for $b_2$, $b_r$); see the section on the bit indexing scheme for the exact specification of this.

For all $i \in 0, 1, \ldots, n - 1$, we have $b_r\{i\} = b_0\{i\} \text{ } | \text{ } b_1\{i\}$, where $|$ refers to a bitwise OR.

-- truncation semantics
bitwiseLogicalOr False [] [0xFF] => []

bitwiseLogicalOr False [0xFF] [] => []

bitwiseLogicalOr False [0xFF] [0x00] => [0xFF]

bitwiseLogicalOr False [0x00] [0xFF] => [0xFF]

bitwiseLogicalOr False [0x4F, 0x00] [0xF4] => [0xFF]

-- padding semantics
bitwiseLogicalOr True [] [0xFF] => [0xFF]

bitwiseLogicalOr True [0xFF] [] => [0xFF]

bitwiseLogicalOr True [0xFF] [0x00] => [0xFF]

bitwiseLogicalOr True [0x00] [0xFF] => [0xFF]

bitwiseLogicalOr True [0x4F, 0x00] [0xF4] => [0xFF, 0x00]

bitwiseLogicalXor

bitwiseLogicalXor takes three arguments; we name and describe them below.

  1. Whether padding semantics should be used. If this argument is False, truncation semantics are used instead. This is the padding semantics argument, and has type BuiltinBool.
  2. The first input BuiltinByteString. This is the first data argument.
  3. The second input BuiltinByteString. This is the second data argument.

Let $b_1, b_2$ refer to the semantics-modified first data argument and semantics-modified second data argument respectively, and let $n$ be either of their lengths in bytes; see the section on padding versus truncation semantics for the exact specification of this. Let the result of bitwiseLogicalXor, given $b_1, b_2$ and some padding semantics argument, be $b_r$, also of length $n$ in bytes. We use $b_1\{i\}$ to refer to the byte at index $i$ in $b_1$ (and analogously for $b_2$, $b_r$); see the section on the bit indexing scheme for the exact specification of this.

For all $i \in 0, 1, \ldots, n - 1$, we have $b_r\{i\} = b_0\{i\} \text{ } \wedge \text{ } b_1\{i\}$, where $\wedge$ refers to a bitwise XOR.

Some examples of the intended behaviour of bitwiseLogicalXor follow. For brevity, we write BuiltinByteString literals as lists of hexadecimal values.

-- truncation semantics
bitwiseLogicalXor False [] [0xFF] => []

bitwiseLogicalXor False [0xFF] [] => []

bitwiseLogicalXor False [0xFF] [0x00] => [0xFF]

bitwiseLogicalXor False [0x00] [0xFF] => [0xFF]

bitwiseLogicalXor False [0x4F, 0x00] [0xF4] => [0xBB]

-- padding semantics
bitwiseLogicalOr True [] [0xFF] => [0xFF]

bitwiseLogicalOr True [0xFF] [] => [0xFF]

bitwiseLogicalOr True [0xFF] [0x00] => [0xFF]

bitwiseLogicalOr True [0x00] [0xFF] => [0xFF]

bitwiseLogicalOr True [0x4F, 0x00] [0xF4] => [0xBB, 0x00]

bitwiseLogicalComplement

bitwiseLogicalComplement takes a single argument, of type BuiltinByteString; let $b$ refer to that argument, and $n$ its length in bytes. Let $b_r$ be the result of bitwiseLogicalComplement; its length in bytes is also $n$. We use $b[i]$ to refer to the value at index $i$ of $b$ (and analogously for $b_r$); see the section on the bit indexing scheme for the exact specification of this.

For all $i \in 0, 1, \ldots , 8 \cdot n - 1$, we have

$$ b_r[i] = \begin{cases} 0 & \text{if } b[i] = 1\\ 1 & \text{otherwise}\\ \end{cases} $$

Some examples of the intended behaviour of bitwiseLogicalComplement follow. For brevity, we write BuiltinByteString literals as lists of hexadecimal values.

bitwiseLogicalComplement [] => []

bitwiseLogicalComplement [0x0F] => [0xF0]

bitwiseLogicalComplement [0x4F, 0xF4] => [0xB0, 0x0B]

readBit

readBit takes two arguments; we name and describe them below.

  1. The BuiltinByteString in which the bit we want to read can be found. This is the data argument.
  2. A bit index into the data argument, of type BuiltinInteger. This is the index argument.

Let $b$ refer to the data argument, of length $n$ in bytes, and let $i$ refer to the index argument. We use $b[i]$ to refer to the value at index $i$ of $b$; see the section on the bit indexing scheme for the exact specification of this.

If $i < 0$ or $i \geq 8 \cdot n$, then readBit fails. In this case, the resulting error message must specify at least the following information:

  • That readBit failed due to an out-of-bounds index argument; and
  • What BuiltinInteger was passed as an index argument.

Otherwise, if $b[i] = 0$, readBit returns False, and if $b[i] = 1$, readBit returns True.

Some examples of the intended behaviour of readBit follow. For brevity, we write BuiltinByteString literals as lists of hexadecimal values.

-- Indexing an empty BuiltinByteString fails
readBit [] 0 => error

readBit [] 345 => error

-- Negative indexes fail
readBit [] (-1) => error

readBit [0xFF] (-1) => error

-- Indexing reads 'from the end'
readBit [0xF4] 0 => False

readBit [0xF4] 1 => False 

readBit [0xF4] 2 => True 

readBit [0xF4] 3 => False

readBit [0xF4] 4 => True

readBit [0xF4] 5 => True

readBit [0xF4] 6 => True

readBit [0xF4] 7 => True

-- Out-of-bounds indexes fail
readBit [0xF4] 8 => error

readBit [0xFF, 0xF4] 16 => error

-- Larger indexes read backwards into the bytes from the end
readBit [0xF4, 0xFF] 10 => False 

writeBits

writeBits takes two arguments: we name and describe them below.

  1. The BuiltinByteString in which we want to change some bits. This is the data argument.
  2. A list of index-value pairs, indicating which positions in the data argument should be changed to which value. This is the change list argument. Each index has type BuiltinInteger, while each value has type BuiltinBool.

Let $b$ refer to the data argument of length $n$ in bytes. We define writeBits recursively over the structure of the change list argument. Throughout, we use $b_r$ to refer to the result of writeBits, whose length is also $n$. We use $b[i]$ to refer to the value at index $i$ of $b$ (and analogously, $b_r$); see the section on the bit indexing scheme for the exact specification of this.

If the change list argument is empty, we return the data argument unchanged. Otherwise, let $(i, v)$ be the head of the change list argument, and $\ell$ its tail. If $i < 0$ or $i \geq 8 \cdot n$, then writeBits fails. In this case, the resulting error message must specify at least the following information:

  • That writeBits failed due to an out-of-bounds index argument; and
  • What BuiltinInteger was passed as $i$.

Otherwise, for all $j \in 0, 1, \ldots 8 \cdot n - 1$, we have

$$ b_r[j] = \begin{cases} 0 & \text{if } j = i \text{ and } v = \texttt{False}\\ 1 & \text{if } j = i \text{ and } v = \texttt{True}\\ b[j] & \text{otherwise}\\ \end{cases} $$

Then, if we did not fail as described above, we repeat the writeBits operation, but with $b_r$ as the data argument and $\ell$ as the change list argument.

Some examples of the intended behaviour of writeBits follow. For brevity, we write BuiltinByteString literals as lists of hexadecimal values.

-- Writing an empty BuiltinByteString fails
writeBits [] [(0, False)] => error

-- Irrespective of index
writeBits [] [(15, False)] => error

-- And value
writeBits [] [(0, True)] => error

-- And multiplicity
writeBits [] [(0, False), (1, False)] => error

-- Negative indexes fail
writeBits [0xFF] [((-1), False)] => error

-- Even when mixed with valid ones
writeBits [0xFF] [(0, False), ((-1), True)] => error

-- In any position
writeBits [0xFF] [((-1), True), (0, False)] => error

-- Out-of-bounds indexes fail
writeBits [0xFF] [(8, False)] => error

-- Even when mixed with valid ones
writeBits [0xFF] [(1, False), (8, False)] => error

-- In any position
writeBits [0xFF] [(8, False), (1, False)] => error

-- Bits are written 'from the end'
writeBits [0xFF] [(0, False)] => [0xFE]

writeBits [0xFF] [(1, False)] => [0xFD]

writeBits [0xFF] [(2, False)] => [0xFB]

writeBits [0xFF] [(3, False)] => [0xF7]

writeBits [0xFF] [(4, False)] => [0xEF]

writeBits [0xFF] [(5, False)] => [0xDF]

writeBits [0xFF] [(6, False)] => [0xBF]

writeBits [0xFF] [(7, False)] => [0x7F]

-- True value sets the bit
writeBits [0x00] [(5, True)] => [0x20]

-- False value clears the bit
writeBits [0xFF] [(5, False)] => [0xDF]

-- Larger indexes write backwards into the bytes from the end
writeBits [0xF4, 0xFF] [(10, False)] => [0xF0, 0xFF]

-- Multiple items in a change list apply cumulatively
writeBits [0xF4, 0xFF] [(10, False), (1, False)] => [0xF0, 0xFD]

writeBits (writeBits [0xF4, 0xFF] [(10, False)]) [(1, False)] => [0xF0, 0xFD]

-- Order within a change list is unimportant among unique indexes
writeBits [0xF4, 0xFF] [(1, False), (10, False)] => [0xF0, 0xFD]

-- But _is_ important for identical indexes
writeBits [0x00, 0xFF] [(10, True), (10, False)] => [0x00, 0xFF]

writeBits [0x00, 0xFF] [(10, False), (10, True)] => [0x04, 0xFF]

-- Setting an already set bit does nothing
writeBits [0xFF] [(0, True)] => [0xFF]

-- Clearing an already clear bit does nothing
writeBits [0x00] [(0, False)] => [0x00]

replicateByteString

replicateByteString takes two arguments; we name and describe them below.

  1. The desired result length, of type BuiltinInteger. This is the length argument.
  2. The byte to place at each position in the result, represented as a BuiltinInteger (corresponding to the unsigned integer this byte encodes). This is the byte argument.

Let $n$ be the length argument, and $w$ the byte argument. If $n < 0$, then replicateByteString fails. In this case, the resulting error message must specify at least the following information:

  • That replicateByteString failed due to a negative length argument; and
  • What BuiltinInteger was passed as the length argument.

If $n \geq 0$, and $w < 0$ or $w > 255$, then replicateByteString fails. In this case, the resulting error message must specify at least the following information:

  • That replicateByteString failed due to the byte argument not being a valid byte; and
  • What BuiltinInteger was passed as the byte argument.

Otherwise, let $b$ be the result of replicateByteString, and let $b\{i\}$ be the byte at position $i$ of $b$, as per the section describing the bit indexing scheme. We have:

  • The length (in bytes) of $b$ is $n$; and
  • For all $i \in 0, 1, \ldots, n - 1$, $b\{i\} = w$.

Some examples of the intended behaviour of replicateByteString follow. For brevity, we write BuiltinByteString literals as lists of hexadecimal values.

-- Replicating a negative number of times fails
replicateByteString (-1) 0 => error

-- Irrespective of byte argument
replicateByteString (-1) 3 => error

-- Out-of-bounds byte arguments fail
replicateByteString 1 (-1) => error

replicateByteString 1 256 => error

-- Irrespective of length argument
replicateByteString 4 (-1) => error

replicateByteString 4 256 => error

-- Length of result matches length argument, and all bytes are the same
replicateByteString 0 0xFF => []

replicateByteString 4 0xFF => [0xFF, 0xFF, 0xFF, 0xFF]

Laws

Binary operations

We describe laws for all three operations that work over two BuiltinByteStrings, that is, bitwiseLogicalAnd, bitwiseLogicalOr and bitwiseLogicalXor, together, as many of them are similar (and related). We describe padding semantics and truncation semantics laws, as they are slightly different.

All three operations above, under both padding and truncation semantics, are commutative semigroups. Thus, we have:

bitwiseLogicalAnd s x y = bitwiseLogicalAnd s y x

bitwiseLogicalAnd s x (bitwiseLogicalAnd s y z) = bitwiseLogicalAnd s
(bitwiseLogicalAnd s x y) z

-- and the same for bitwiseLogicalOr and bitwiseLogicalXor

Note that the semantics (designated as s above) must be consistent in order for these laws to hold. Furthermore, under padding semantics, all the above operations are commutative monoids:

bitwiseLogicalAnd True x "" = bitwiseLogicalAnd True "" x = x

-- and the same for bitwiseLogicalOr and bitwiseLogicalXor

Under truncation semantics, "" (that is, the empty BuiltinByteString) acts instead as an absorbing element:

bitwiseLogicalAnd False x "" = bitwiseLogicalAnd False "" x = ""

-- and the same for bitwiseLogicalOr and bitwiseLogicalXor

bitwiseLogicalAnd and bitwiseLogicalOr are also semilattices, due to their idempotence:

bitwiseLogicalAnd s x x = x

-- and the same for bitwiseLogicalOr

bitwiseLogicalXor is instead involute:

bitwiseLogicalXor s x (bitwiseLogicalXor s x x) = bitwiseLogicalXor s
(bitwiseLogicalXor s x x) x = x

Additionally, under padding semantics, bitwiseLogicalAnd and bitwiseLogicalOr are self-distributive:

bitwiseLogicalAnd True x (bitwiseLogicalAnd True y z) = bitwiseLogicalAnd True
(bitwiseLogicalAnd True x y) (bitwiseLogicalAnd True x z)

bitwiseLogicalAnd True (bitwiseLogicalAnd True x y) z = bitwiseLogicalAnd True
(bitwiseLogicalAnd True x z) (bitwiseLogicalAnd True y z)

-- and the same for bitwiseLogicalOr

Under truncation semantics, bitwiseLogicalAnd is only left-distributive over itself, bitwiseLogicalOr and bitwiseLogicalXor:

bitwiseLogicalAnd False x (bitwiseLogicalAnd False y z) = bitwiseLogicalAnd
False (bitwiseLogicalAnd False x y) (bitwiseLogicalAnd False x z)

bitwiseLogicalAnd False x (bitwiseLogicalOr False y z) = bitwiseLogicalOr False
(bitwiseLogicalAnd False x y) (bitwiseLogicalAnd False x z)

bitwiseLogicalAnd False x (bitwiseLogicalXor False y z) = bitwiseLogicalXor
False (bitwiseLogicalAnd False x y) (bitwiseLogicalAnd False x z)

bitwiseLogicalOr under truncation semantics is left-distributive over itself and bitwiseLogicalAnd:

bitwiseLogicalOr False x (bitwiseLogicalOr False y z) = bitwiseLogicalOr False
(bitwiseLogicalOr False x y) (bitwiseLogicalOr False x z)

bitwiseLogicalOr False x (bitwiseLogicalAnd False y z) = bitwiseLogicalAnd False
(bitwiseLogicalOr False x y) (bitwiseLogicalOr False x z)

If the first and second data arguments to these operations have the same length, these operations satisfy several additional laws. We describe these briefly below, with the added note that, in this case, padding and truncation semantics coincide:

  • bitwiseLogicalAnd and bitwiseLogicalOr form a bounded lattice
  • bitwiseLogicalAnd is distributive over itself, bitwiseLogicalOr and bitwiseLogicalXor
  • bitwiseLogicalOr is distributive over itself and bitwiseLogicalAnd

We do not specify these laws here, as they do not hold in general. At the same time, we expect that any implementation of these operations will be subject to these laws.

bitwiseLogicalComplement

The main law of bitwiseLogicalComplement is involution:

bitwiseLogicalComplement (bitwiseLogicalComplement x) = x

In combination with bitwiseLogicalAnd and bitwiseLogicalOr, bitwiseLogicalComplement gives rise to the famous De Morgan laws, irrespective of semantics:

bitwiseLogicalComplement (bitwiseLogicalAnd s x y) = bitwiseLogicalOr s
(bitwiseLogicalComplement x) (bitwiseLogicalComplement y)

bitwiseLogicalComplement (bitwiseLogicalOr s x y) = bitwiseLogicalAnd s
(bitwiseLogicalComplement x) (bitwiseLogicalComplement y)

For bitwiseLogicalXor, we instead have (again, irrespective of semantics):

bitwiseLogicalXor s x (bitwiseLogicalComplement x) = x

Bit reading and modification

Throughout, we assume any index arguments to be 'in-bounds'; that is, all the index arguments used in the statements of any law are such that the operation they are applied to wouldn't produce an error.

The first law of writeBits is similar to the set-twice law of lenses:

writeBits bs [(i, b1), (i, b2)] = writeBits bs [(i, b2)]

Together with readBit, we obtain the remaining two analogues to the lens laws:

-- writing to an index, then reading from that index, gets you what you wrote
readBit (writeBits bs [(i, b)]) i = b

-- if you read from an index, then write that value to that same index, nothing
-- happens
writeBits bs [(i, readBit bs i)] = bs

Furthermore, given a fixed data argument, writeBits acts as a monoid homomorphism lists under concatenation to functions:

writeBits bs [] = bs

writeBits bs (is <> js) = writeBits (writeBits bs is) js

replicateByteString

Given a fixed byte argument, replicateByteString acts as a monoid homomorphism from natural numbers under addition to BuiltinByteStrings under concatenation:

replicateByteString 0 w = ""

replicateByteString (n + m) w = replicateByteString n w <> replicateByteString m w

Additionally, for any 'in-bounds' index (that is, any index for which indexByteString won't error) i, we have

indexByteString (replicateByteString n w) i = w

Lastly, we have

lengthByteString (replicateByteString n w) = n

Rationale: how does this CIP achieve its goals?

The operations, and semantics, described in this CIP provide a set of well-defined bitwise logical operations, as well as bitwise access and modification, to allow cases similar to Case 1 to be performed efficiently and conveniently. Furthermore, the semantics we describe would be reasonably familiar to users of other programming languages (including Haskell) which have provisions for bitwise logical operations of this kind, as well as some way of extending these operations to operate on packed byte vectors. At the same time, there are several choices we have made that are somewhat unusual, or could potentially have been implemented differently based on existing work: most notably, our choice of bit indexing scheme, the padding-versus-truncation semantics, and the multiplicitous definition of bit modification. Among existing work, a particularly important example is CIP-58, which makes provisions for operations similar to the ones described here, and from which we differ in several important ways. We clarify the reasoning behind our choices, and how they differ from existing work, below.

Aside from the issues we list below, we don't consider other operations controversial. Indeed, bitwiseLogicalComplement has a direct parallel to the implementation in CIP-58, and replicateByteString is a direct wrapper around the replicate function in ByteString. Thus, we do not discuss them further here.

Relationship to CIP-58 and CIP-121

Our work relates to both CIP-58 and CIP-121. Essentially, our goal with both this CIP and CIP-121 is to both break CIP-58 into more manageable (and reviewable) parts, and also address some of the design choices in CIP-58 that were not as good (or as clear) as they could have been. In this regard, this CIP is a direct continuation of CIP-121; CIP-121 dealt with conversions between BuiltinByteString and BuiltinInteger, while this CIP handles bit indexing more generally, as well as 'parallel' logical operations that operate on all the bits of a BuiltinByteString in bulk.

We describe how our work in this CIP relates to (and in some cases, supercedes) CIP-58, as well as how it follows on from CIP-121, in more detail below.

Bit indexing scheme

The bit indexing scheme we describe here is designed around two considerations. Firstly, we want operations on these bits, as well as those results, to be as consistent and as predictable as possible: any individual familiar with such operations on variable-length bitvectors from another language shouldn't be surprised by the semantics. Secondly, we want to anticipate future bitwise operation extensions, such as shifts and rotations, and have the indexing scheme support efficient implementations (and predictable semantics) for these.

While prior art for bit access (and modification) exists in almost any programming language, these are typically over types of fixed width (usually bytes, machine words, or something similar); for variable-width types, these typically are either not implemented at all, or if they are implemented, this is done in an external library, with varying support for certain operations. An example of the first is Haskell's ByteString, which has no way to even access, much less modify, individual bits; an example of the second is the CRoaring library for C, which supports all the operations we describe in this CIP, along with multiple others. In the second case, the exact arrangement of bits inside the representation is not something users are exposed to directly: instead, the bitvector type is opaque, and the library only guarantees consistency of API. In our case, this is not a viable choice, as we require bit access and byte access to both work on BuiltinByteString, and thus, some consistency of representation is required.

The scheme for indexing bits within a byte that we describe in the relevant section is the same as the one used by the Data.Bits API in Haskell for Word8 bit indexing, and mirrors the decisions of most languages that provide such an API at all, as well as the conventional definition of such operations as (w >> i) & 1 for access, w | (1 << i) for setting, and w & ~(1 << i) for clearing. We could choose to 'flip' this indexing, by using a similar operation for 'index flipping' as we currently use for bytes: essentially, instead of

$$ \left \lfloor \frac{w}{2^{i}} \right \rfloor \mod 2 \equiv 1 $$

we would instead use

$$ \left \lfloor \frac{w}{2^{8 - i - 1}} \right \rfloor \mod 2 \equiv 1 $$

to designate bit $i$ as set (and analogously for clear). Together with the ability to choose not to flip the byte index, we get four possibilities, which have been described previously. For clarity, we name, and describe, them below. Throughout, we use n as the length of a given BuiltinByteString in bytes.

The first possibility is that we 'flip' neither bit, nor byte, indexes. We call this the no-flip variant:

| Byte index | 0                             | 1                 | ... | n - 1                          |
|------------|-------------------------------|-------------------| ... |--------------------------------|
| Byte       | w0                            | w1                | ... | w(n - 1)                       |
|------------|-------------------------------|-------------------| ... |--------------------------------|
| Bit index  | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 | 15 | 14 | ... | 8 | ... | 8n - 1 | 8n - 2 | ... | 8n - 8 |

The second possibility is that we 'flip' both bit and byte indexes. We call this the both-flip variant:

| Byte index | 0                              | ... | n - 2            | n - 1                         |
|------------|--------------------------------| ... |------------------|-------------------------------|
| Byte       | w0                             | ... | w (n - 2)        | w(n - 1)                      |
|------------|--------------------------------| ... |------------------|-------------------------------|
| Bit index  | 8n - 8 | 8n - 7 | ... | 8n - 1 | ... | 8 | 9 | ... | 15 | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 

The third possibility is that we 'flip' bit indexes, but not byte indexes. We call this the bit-flip variant:

| Byte index | 0                             | 1            | ... | n - 1                          |
|------------|-------------------------------|--------------| ... |--------------------------------|
| Byte       | w0                            | w1           | ... | w(n - 1)                       |
|------------|-------------------------------|--------------| ... |--------------------------------|
| Bit index  | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | ... | 15 | ... | 8n - 8 | 8n - 7 | ... | 8n - 1 |

The fourth possibility is the one we describe in the bit indexing scheme section, which is also the scheme chosen by CIP-58. We repeat it below for clarity:

| Byte index | 0                              | 1  | ... | n - 1                         |
|------------|--------------------------------|----| ... |-------------------------------|
| Byte       | w0                             | w1 | ... | w(n - 1)                      |
|------------|--------------------------------|----| ... |-------------------------------|
| Bit index  | 8n - 1 | 8n - 2 | ... | 8n - 8 |   ...    | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |

On the face of it, these schemes appear equivalent: they are all consistent, and all have formal descriptions, and quite similar ones at that. However, we believe that only the one we chose is the correct one. To explain this, we introduce two notions that we consider to be both intuitive and important, then specify why our choice of indexing scheme fits those notions better than any other.

The first notion is index locality. Intuitively, this states that if two indexes are 'close' (in that their absolute difference is small), the values at those indexes should be 'close' (in that their positioning in memory should be separated less). We believe this notion to be reasonable, as this is an expectation from array indexing (and indeed, BuiltinByteString indexing), as well as the reason why packed array data is efficient on modern memory hierarchies. Extending this notion to bits, we can observe that the both-flip and no-flip variants of the bit indexing scheme do not preserve index locality: the separation between a bit at index $0$ and index $1$ is significantly different to the separation between a bit at index $7$ and index $8$ in both representations, despite their absolute difference being identical. Thus, we believe that these two variants are not viable, as they are not only confusing from the point of view of behaviour, they would also make implementation of future operations (such as shifts or rotations) significantly harder to both do, and also reason about. Thus, only the bit-flip variant, as well as our choice, remain contenders.

The second notion is most-significant-first conversion agreement. This notion refers to the CIP-121 concept of the same name, and ensures that (at least for the most-significant-first arrangement), the following workflow doesn't produce unexpected results:

  1. Convert a BuiltinInteger to BuiltinByteString using builtinIntegerToByteString with the most-significant-first endianness argument.
  2. Manipulate the bits of the result of step 1 using the operations specified here.
  3. Convert the result of step 2 back to a BuiltinInteger using builtinByteStringToInteger with the most-significant-first endianness argument.

This workflow is directly relevant to Case 2. The Argon2 family of hashes use certain inputs (which happen to be numbers) both as numbers (meaning, for arithmetic operatons) and also as blocks of binary (specifically for XOR). This is not unique to Argon2, or even hashing, as a range of operations (especially in cryptographic applications) use similar approaches, whether for performance, semantics or both. In such cases, users of our primitives (both logical and conversion) must be confident that their changes 'translate' in the way they expect between these two 'views' of the data.

The choice of most-significant-first as the arrangement that we must agree with seems somewhat arbitrary at a glance, for two reasons: firstly, it's not clear why we must pick a single arrangement to be consistent with; secondly, the reasoning for the choice of most-significant-first over most-significant-last as the arrangement to agree with isn't immediately apparent. To see why this is the only choice that we consider reasonable, we first observe that, according to the definition of the bit indexing scheme given in the corresponding section, as well as the corresponding definition for the bit-flip variant, we view a BuiltinByteString of length $n$ as a binary natural number with exactly $8n$ digits, and the value at index $i$ corresponds to the digits whose place value is either $2^i$ (for the bit-flip variant), or $2^{8n - i - 1}$ (for our chosen method). Put another way, under the specification for the bit-flip variant, the least significant binary digit is first, whereas in our chosen specification, the least significant binary digit is last. CIP-121's conversion primitives mirror this reasoning: the most-significant-first arrangement corresponds to our chosen method, while the most-significant-last arrangement corresponds to the bit-flip variant instead. The difference is the digit value: for us, the digit value is (effectively) 2, while for CIP-121's conversion primitives, it is 256 instead.

We also observe that, when we index a BuiltinByteString's bytes, we get back a BuiltinInteger, whic has a numerical value as a natural number in the range $[0, 255]$. Putting these two observations together, we consider it sensible that, given a non-empty BuiltinByteString, if we were to get the values at bit indexes $0$ through $7$, then sum their corresponding place values (treating clear bits as $0$ and set bits as the appropriate place value), we should get the same result as indexing whichever byte those bits came from.

Consider the BuiltinByteString whose only byte is $42$, whose representation is as follows:

| Byte index | 0        |
|------------|----------|
| Byte       | 00101010 |

We note that, if we index this BuiltinByteString at byte position $0$, we get back the answer $42$. Furthermore, if we use builtinByteStringToInteger from CIP-121 with such a BuiltinByteString, we get the result $42$ as well, regardless of the endianness argument we choose.

Under the bit-flip variant, the bit indexes of this BuiltinByteString would be as follows:

| Byte index | 0                             |
|------------|-------------------------------|
| Byte       | 0 | 0 | 1 | 0 | 1 | 0 | 1 | 0 |
|------------|-------------------------------|
| Bit index  | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 |

However, we immediately see a problem: under this indexing scheme, the $2^2 = 4$ place value is $1$, which would suggest that in the binary representation of $42$, the corresponding digit is also $1$. However, this is not the case. Under our scheme of choice however, we get the correct answer:

| Byte index | 0                             |
|------------|-------------------------------|
| Byte       | 0 | 0 | 1 | 0 | 1 | 0 | 1 | 0 |
|------------|-------------------------------|
| Bit index  | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |

Here, the $4$ place value is correctly $0$. This demonstrates that of the two indexing scheme possibilities that preserve index locality, only one can be consistent with any choice of byte arrangement, whether most-significant-first or most-significant-last: the one we chose. This implies that we cannot be consistent with both arrangements while also preserving index locality.

Let us now consider a larger example BuiltinByteString:

| Byte index | 0        | 1        |
|------------|----------|----------|
| Byte       | 00101010 | 11011011 |

This would produce two different results when converted with builtinByteStringToInteger, depending on the choice of endianness argument:

  • For the most-significant-first arrangement, the result is $42 * 256 + 223 = 10975$.
  • For the most-significant-last arrangement, the result is $223 * 256 + 42 = 57130$.

These have the following 'breakdowns' in base-2:

  • $10975 = 8096 + 2048 + 512 + 256 + 32 + 16 + 8 + 4 + 2 + 1 = 2^13 + 2^11 + 2^9 + 2^8 + 2^5 + 2^4 + 2^3 + 2^2 + 2^1 + 2^0$
  • $57130 = 32768 + 16386 + 4096 + 2048 + 1024 + 512 + 256 + 32 + 8 + 2 = 2^15 + 2^14 + 2^12 + 2^11 + 2^10 + 2^9 + 2^8 + 2^5 + 2^3 + 2^1$

Under the bit-flip variant, the bit indexes of this BuiltinByteString would be as follows:

| Byte index | 0                             | 1                                   |
|------------|-------------------------------|-------------------------------------|
| Byte       | 0 | 0 | 1 | 0 | 1 | 0 | 1 | 0 | 1 | 1 | 0  | 1  | 1  | 0  | 1  | 1  |
|------------|-------------------------------|-------------------------------------|
| Bit index  | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 |

We immediately see a problem, as in this representation, it suggests that the $2^1 = 2$ place value has zero digit value. This is true of neither $10975$ nor $57130$'s base-2 forms, which have the $2$ place value with a $1$ digit value. This suggests that the bit-flip variant cannot agree with either choice of arrangement in general.

However, if we view the bit indexes using our chosen scheme:

| Byte index | 0                                   | 1                             |
|------------|-------------------------------------|-------------------------------|
| Byte       | 0  | 0  | 1  | 0  | 1  | 0  | 1 | 0 | 1 | 1 | 0 | 1 | 1 | 0 | 1 | 1 |
|------------|-------------------------------------|-------------------------------|
| Bit index  | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |

the $2$ place value is correctly shown as having a digit value of 1.

Combining these observations, we note that, assuming we value index locality, choosing our scheme gives us consistency with the most-significant-first arrangement, as well as consistency with byte indexing digit values, but choosing the bit-flip variant gives us neither. As we need both index locality and consistency with at least one arrangement, our choice is the correct one. The fact that we also get byte indexing digit values being consistent is another reason for our choice.

Padding versus truncation

For the operations defined in this CIP taking two BuiltinByteString arguments (that is, bitwiseLogicalAnd, bitwiseLogicalOr, and bitwiseLogicalXor), when the two arguments have identical lengths, the semantics are natural, mirroring the corresponding operations on the Boolean algebra $\textbf{2}^{8n}$, where $n$ is the length of either argument in bytes. When the arguments do not have matching lengths, however, the situation becomes more complex, as there are several ways in which we could define these operations. The most natural possibilities are as follows; we repeat some of the definitions used in the corresponding section.

  • Extend the shorter argument with the identity element (all-1s for bitwiseLogicalAnd, all-0s otherwise) to match the length of the longer argument, then perform the operation as if on matching-length arguments. We call this padding semantics.
  • Ignore the bytes of the longer argument whose indexes would not be valid for the shorter argument, then perform the operation as if on matching-length arguments. We call this truncation semantics.
  • Fail with an error whenever argument lengths don't match. We call this match semantics.

Furthermore, for both padding and truncation semantics, we can choose to pad (or truncate) low index bytes or high index bytes. To illustrate the difference, consider the two BuiltinByteStrings (written as arrays of bytes for simplicity) [0xFF, 0x0F, 0x00] and [0x8F, 0x99]. Under padding semantics, padding low index bytes would give us [0x00, 0x8F, 0x99] (or [0xFF, 0x8F, 0x99] depending on operation), while padding high index bytes would give us [0x8F, 0x99, 0x00] (or [0x8F, 0x99, 0xFF] depending on operation). Under truncation semantics, truncating low index bytes would give us [0x0F, 0x00], while truncating high index bytes would give us [0xFF, 0x0F].

It is not a priori clear which of these we should choose: they are subject to different laws (as evidenced by the corresponding section), none of which are strict supersets of each other (at least not for all inputs possible). While CIP-58 chose match semantics, we believe this was not the correct decision: we use Case 1 to justify the benefit of having other semantics described above available.

Consider the following operation: given a bound $k$, a 'direction' (larger or smaller), and an integer set, remove all elements indicates by the direction and $k$ (that is, either smaller than $k$ or larger than $k$, as indicated by the direction). This could be done using a bitwiseLogicalAnd and a mask. However, under match semantics, this mask would have to have a length equal to the integer set representation; under padding semantics, the mask would potentially only need $\Theta(k)$ length, depending on direction. This is noteworthy, as padding the mask would require an additional copy operation, only to produce a value that would be discarded immediately.

Consider instead the following operation: given two integer sets with different (upper) bounds, take their intersection, producing an integer set whose size is the minimum of the two. This can once again be done using bitwiseLogicalAnd, but under match semantics (or padding semantics for that matter), we would first have to slice the longer argument, while under truncation semantics, we wouldn't need to.

Match semantics can be useful for Case 1 as well. Consider the case that a representation of an integer set is supplied as an input datum (in its Data encoding). In order to deserialize it, we need to verify at least whether it has the right length in bytes to represent an integer set with a given bound. Under padding or truncation semantics, we would have to check this at deserialization time; under exact match semantics, provided we were sure that at least one argument is of a known size, we could simply perform the necessary operations and let the match semantics error if given something inappropriate.

It is also worth noting that truncation semantics are well-established in the Haskell ecosystem. Viewed another way, all of the operations under discussion in this sections are specialized versions of the zipWith operation; Haskell libraries provide this type of operation for a range of linear collections, including lists, Vectors, and mostly notably, ByteStrings. In all of these cases, truncation semantics are what is implemented; it would be surprising to developers coming from Haskell to find that they have to do additional work to replicate them in Plutus. While we don't anticipate direct use of Plutus Core primitives by developers (although this is not an unheard-of case), we should enable library authors to build familiar APIs on top of Plutus Core primitives, which suggests truncation semantics should be available, at least as an option.

All the above suggests that no single choice of semantics will satisfy all reasonable needs, if only from the point of view of efficiency. This suggests, much as for CIP-121 primitives and endianness issues, that the primitive should allow a choice in what semantics get used for any given call. Ideally, we would allow a choice of any of the three options described above (along with a choice of low or high index padding or truncation); however, this is awkward to do in Plutus Core. While the choice between two options is straightforward (pass a BuiltinBool), the choice between more than this number would require something like a BuiltinInteger argument with 'designated values' ('0 means match', '1 means low-index padding', etc). This is not ideal, as they involve additional checks, argument redundancy, or both. In light of this, we made the following decisions:

  1. We would choose only two of the three semantics, and have this choice controlled for any given call be controlled by a BuiltinBool flag; and
  2. For padding or truncation semantics, we would always use either low or high index padding (or truncation).

This leads naturally to two questions: which of the three semantics above we can afford to exclude, and whether low or high index padding should be chosen. We believe that the correct choices are to exclude match semantics, and to use high index padding and truncation, for several reasons.

Firstly, we can simulate match semantics with either padding or truncation semantics, together with a length check. While we could also simulate padding semantics via match semantics similarly, the amount of effort (both developer and computational) required is significantly more in that case: a length check is a constant-time operation, while manually padding is linear at best (and even then, it requires operations only this CIP provides, as it would be quadratic otherwise), and on top of that, manual padding is much fiddlier and easier to get wrong.

Secondly, truncation semantics are common enough in Haskell that we believe excluding them as an option is both surprising and wrong. Any developer familiar with Haskell has interacted with various zipWith operations, and having our primitives behave differently to this at minimum creates friction for implementers of higher-level abstractions atop the primitives in this CIP. While Haskellers are not exclusive users of Plutus primitives (directly or not), there are definitely enough of them that not having truncation semantics available would create a lot of unnecessary friction.

Thirdly, outside of error checking, match semantics give few benefits, performance or otherwise. The examples above demonstrate cases where padding and truncation semantics lead to better performance, less fiddly implementations, or both: finding such a case for match semantics outside of error checking is difficult at best.

This combination of reasoning leads us to consider padding and truncation as the two semantics we should retain, and this guided our implementation choices accordingly. With regard to padding (or truncating) low or high indexes, given that we pad (or truncate) whole bytes by necessity, it makes the corresponding operations (effectively) operate over bytes, or rather, they view BuiltinByteStrings as linear collections of bytes, rather than bits. When viewed this way, the zipWith analogy with Haskell suggests that truncating high is the correct choice: truncating low would be quite surprising to a Haskeller familiar with how zipWith-style operations behave. Furthermore, as having padding low and truncating high would be confusing (and arguably quite strange), padding high seems like the correct choice. Thus, we decided to both pad and truncate high in light of this.

Bit setting

writeBits in our description takes a change list argument, allowing changing multiple bits at once. This is an added complexity, and an argument can be made that something similar to the following operation would be sufficient:

writeBit :: BuiltinByteString -> BuiltinInteger -> BuiltinBool ->
BuiltinByteString

Essentially, writeBit bs i v would be equivalent to writeBits bs [(i, v)] as currently defined. This was the choice made by CIP-58, with the consideration of simplicity in mind.

At the same time, due to the immutability semantics of Plutus Core, each time writeBit would be called, we would have to copy its BuiltinByteString argument. Thus, a sequence of $k$ setBit calls in a fold over a BuiltinByteString of length $n$ would require $\Theta(nk)$ time and $\Theta(nk)$ space. Meanwhile, if we instead used writeBits, the time drops to $\Theta(n + k)$ and the space to $\Theta(n)$, which is a non-trivial improvement. While we cannot avoid the worst-case copying behaviour of setBit (if we have a critical path of read-write dependencies of length $k$, for example), and 'list packing' carries some cost, we have benchmarks that show not only that this 'packing cost' is essentially zero, but that for BuiltinByteStrings of 30 bytes or fewer, copying completely overwhelms the work required to modify the bits specified in the change list argument. This alone is good evidence for having writeBits instead; indeed, there is prior art for doing this in the vector library, for the exact reasons we give here.

The argument could also be made whether this design should be extended to other primitive operations in this CIP which both take BuiltinByteString arguments and also produce BuiltinByteString results. We believe that this is not as justified as in the writeBits case, for several reasons. Firstly, for bitwiseLogicalComplement, it's not clear what benefit this would have at all: the only possible signature such an operation would have is [BuiltinByteString] -> [BuiltinByteString], which in effect would be a specialized form of mapping. While an argument could be made for a general form of mapping as a Plutus Core primitive, it wouldn't be reasonable for an operation like this to be considered for such.

Secondly, the performance benefits of such an operation aren't nearly as significant in theory, and likely wouldn't be in practice either. Consider this hypothetical operation (with fold semantics):

bitwiseLogicalXors :: BuiltinBool -> [BuiltinByteString] -> BuiltinByteString

Simulating this operation as a fold using bitwiseLogicalXor, in the worst case, irrespective of padding or truncation semantics, requires $\Theta(nk)$ time and space, where $n$ is the size of each BuiltinByteString in the argument list, and $k$ is the length of the argument list itself. Using bitwiseLogicalXors instead would reduce the space required to $\Theta(n)$, but would not affect the time complexity at all.

Lastly, it is questionable whether 'bulk' operations like bitwiseLogicalXors above would see as much use as writeBits. In the context of Case 1, bitwiseLogicalXors corresponds to taking the symmetric difference of multiple integer sets; it seems unlikely that the number of sets we'd want to do this with would frequently be higher than 2. However, in the same context, writeBits corresponds to constructing an integer set given a list of members (or, for that matter, non-members): this is an operation that is both required by the case description, and also much more likely to be used often.

On the basis of the above, we believe that choosing to implement writeBits as a 'bulk' operation, but to leave others as 'singular' is the right choice.

Path to Active

Acceptance Criteria

We consider the following criteria to be essential for acceptance:

  • A proof-of-concept implementation of the operations specified in this document, outside of the Plutus source tree. The implementation must be in GHC Haskell, without relying on the FFI.
  • The proof-of-concept implementation must have tests, demonstrating that it behaves as the specification requires.
  • The proof-of-concept implementation must demonstrate that it will successfully build, and pass its tests, using all GHC versions currently usable to build Plutus (8.10, 9.2 and 9.6 at the time of writing), across all Tier 1 platforms.

Ideally, the implementation should also demonstrate its performance characteristics by well-designed benchmarks.

Implementation Plan

MLabs has begun the implementation of the proof-of-concept as required in the acceptance criteria. Upon completion, we will send a pull request to Plutus with the implementation of the primitives for Plutus Core, mirroring the proof-of-concept.

Copyright

This CIP is licensed under Apache-2.0.