googlesql: type inference via the new analysis core (analyze only)

[kyleconroy]

Overview

Redo of #4522 on top of the core catalog and analyzer landed in #4521. The original wired GoogleSQL into the legacy compiler (googlesql.NewCatalog() + analyzeQuery/outputColumns); this wires it the way ClickHouse is wired — sqlc analyze --dialect googlesql resolves result columns and parameters entirely through internal/core, with the legacy analyze path never entered.

Engine wiring

• internal/engine/googlesql/seed.go — a dialect seed registering the built-in GoogleSQL types, comparison and arithmetic operators over them, and COUNT as an int64 aggregate.
• internal/compiler/engine.go — the googlesql engine builds a core.Catalog; schema DDL is applied via internal/core/schema and queries are resolved by internal/core/analyzer.
• sqlc analyze accepts googlesql as a dialect.

Core analyzer

Engine-neutral work the GoogleSQL fixtures required, which ClickHouse gets too:

• internal/core/analyzer/dml.go — INSERT, UPDATE, and DELETE support alongside SELECT: bind the target relation, infer parameter types from the column each value is assigned to or compared against, and project RETURNING / THEN RETURN lists.
• COUNT(*) resolves through a catalog-registered aggregate when the dialect provides one, falling back to int8 as before. This closes the data_type: "any" limitation Add analyze support for GoogleSQL via zetajones #4522 flagged as a follow-up.
• The core compile path rewrites named parameters before analysis, so @name resolves to a numbered parameter that carries its name.

GoogleSQL engine

• Include a statement's leading comment in its reported location so the -- name: annotation is captured. Previously StmtLocation pointed past the comment, so analyze found no named queries. This also fixes name extraction for sqlc parse --dialect googlesql.
• Tag operator A_Expr nodes with A_Expr_Kind_OP, matching the other engines, so the analyzer recognizes them as operator expressions.
• Keep named parameters in their native @name form when rewriting, since GoogleSQL supports them natively and Spanner requires them.

Differences from #4522

Dropped, because they only mattered on the legacy path:

• The nil-list initializations in the converter (ReturningList, TargetList, FromClause). The panics they fixed were in legacy compiler code the core path does not run.
• The backtick quoteIdent change, used only by legacy star expansion.

Tests

End-to-end fixtures under internal/endtoend/testdata/, run by TestReplay:

• analyze_basic/googlesql and analyze_dml/googlesql — reproduce Add analyze support for GoogleSQL via zetajones #4522's fixtures byte for byte (the DML case also covers UPDATE ... THEN RETURN and DELETE ... THEN RETURN).
• analyze_select/googlesql — new: star expansion, joins with aliases, and COUNT(*).
• Regenerated the parse_basic/googlesql golden, which picks up the corrected statement spans and the now-detected query name and command.

go test --tags=examples -timeout 25m ./... is green with PostgreSQL and MySQL running; gofmt and go vet are clean on the touched files.

Known limitations

• SELECT * expands in the analyzer's column output but the SQL text is not rewritten (same as ClickHouse — fine for analyze, needs work before generate).
• INSERT ... SELECT returns an explicit "not supported" error rather than silently producing no columns or parameters.

🤖 Generated with Claude Code

---

Generated by Claude Code