tree 544cf76fe61122d7ef4edd916454bfa9852c5465
parent 9b38c456f3f0c0f934d08230f339772e1f9ff9aa
author Protobuf Team Bot <protobuf-github-bot@google.com> 1791778990 -0700
committer Copybara-Service <copybara-worker@google.com> 1792028680 -0700

Avoid redundant wire byte loads and template bloat in table-driven varint parsing.

Before calling `TcParser::SingularVarint`, `TcParser::TagDispatch` performs an unaligned 16-bit load from the wire buffer and XORs it with the fast-table tag entry in `data.data`. For 1-byte tags (field numbers 1 through 15), `coded_tag<uint8_t>()` masks only bits 0 through 7, leaving the first varint payload byte (`*ptr`) already loaded in bits 8 through 15 of `data.data`. `SingularVarint` previously ignored this register value and issued a redundant byte load from `*ptr`. In addition, the multi-byte varint slow path tail-called `SingularVarBigint` parameterized on `TagType` even though `ptr` had already been advanced past `sizeof(TagType)`, duplicating slow-path instantiations across 1-byte and 2-byte tags, and always decoded into a 64-bit `uint64_t` temporary even for 32-bit fields.

Extract the initial varint byte directly from `static_cast<uint8_t>(data.data >> 8)` when `sizeof(TagType) == 1`, unify multi-byte tail calls onto `SingularVarBigint<FieldType, uint8_t, zigzag>`, and select a 32-bit `uint32_t` temporary in `SingularVarBigint` when `sizeof(FieldType) == 4`.

PiperOrigin-RevId: 997663183