Bytecode cache: smaller payloads that page in only what is decoded (alias mapped bytes, region layout, 32-bit offsets, sparse metadata table) by Jarred-Sumner · Pull Request #494 · oven-sh/WebKit

Bytecode cache: decoded InstructionStream and ExpressionInfo alias a persistent payload instead of copying it

Jarred-Sumner

committed

75ee29f

Bytecode cache: write each code block's records contiguously, bodies breadth-first, expression info last

Jarred-Sumner

committed

ae604f8

Bytecode cache: 32-bit relative offsets, natural alignment, sparse metadata offset table

Jarred-Sumner

committed

89be13e

CachedMetadataTable::expand: bounds-check the step indices

Jarred-Sumner

committed

0ef16dd

Bytecode cache: varint tails for per-function records, inline string characters, no unused offset fields

Jarred-Sumner

committed

79782d7

Bytecode cache: store byte-identical arrays once

Jarred-Sumner

committed

758a014

UnlinkedMetadataTable: hold only the offset table while no CodeBlock is linked

Jarred-Sumner

committed

77e2f99

Bytecode cache: CRC-32C per code block, executable record and expression info; regenerate from source on mismatch

Jarred-Sumner

committed

57c5e71

Checksum verification: reject a stored size that does not cover the checksum field; assert re-sealed record bounds

Jarred-Sumner

committed

c545f8a

Array dedup: only share an earlier copy whose offset satisfies this array's alignment

Jarred-Sumner

committed

33c3372

crc32c: detect SSE4.2 with cpuid instead of __builtin_cpu_supports (needs compiler-rt's __cpu_model, absent in the macOS cross-link)

Jarred-Sumner

committed

89c8145

Checksums: fold arrays shared from an earlier block into the sharer's checksum, cover an executable record's own rare data / name / TDZ allocations, and seal the cache entry header

Jarred-Sumner

committed

61c5f15

Zero the 16-bit offset table's padding; bound a re-sealed record's extent below as well

Jarred-Sumner

committed

43e3144

decodeCachedCodeBlocks: clear the construct slot (it aliases the cached offsets) before decoding, so a rejected block falls back to source instead of leaving the offsets behind as a cell pointer

Jarred-Sumner

committed

738c4bc

Navigation Menu

Uh oh!

Bytecode cache: smaller payloads that page in only what is decoded (alias mapped bytes, region layout, 32-bit offsets, sparse metadata table) - #494

Bytecode cache: smaller payloads that page in only what is decoded (alias mapped bytes, region layout, 32-bit offsets, sparse metadata table)#494
Jarred-Sumner merged 14 commits into
mainfrom
claude/bytecode-cache-borrow

Commits on Aug 23, 2026

Bytecode cache: decoded InstructionStream and ExpressionInfo alias a persistent payload instead of copying it

Bytecode cache: write each code block's records contiguously, bodies breadth-first, expression info last

Bytecode cache: 32-bit relative offsets, natural alignment, sparse metadata offset table

CachedMetadataTable::expand: bounds-check the step indices

Bytecode cache: varint tails for per-function records, inline string characters, no unused offset fields

Bytecode cache: store byte-identical arrays once

UnlinkedMetadataTable: hold only the offset table while no CodeBlock is linked

Bytecode cache: CRC-32C per code block, executable record and expression info; regenerate from source on mismatch

Checksum verification: reject a stored size that does not cover the checksum field; assert re-sealed record bounds

Array dedup: only share an earlier copy whose offset satisfies this array's alignment

crc32c: detect SSE4.2 with cpuid instead of __builtin_cpu_supports (needs compiler-rt's __cpu_model, absent in the macOS cross-link)

Checksums: fold arrays shared from an earlier block into the sharer's checksum, cover an executable record's own rare data / name / TDZ allocations, and seal the cache entry header

Zero the 16-bit offset table's padding; bound a re-sealed record's extent below as well

decodeCachedCodeBlocks: clear the construct slot (it aliases the cached offsets) before decoding, so a rejected block falls back to source instead of leaving the offsets behind as a cell pointer