mirror of
https://github.com/aljazceru/turso.git
synced 2025-12-26 04:24:21 +01:00
Closes #1193 ```console │limbo> explain create index idxp on products(price); │addr opcode p1 p2 p3 p4 p5 comment │---- ----------------- ---- ---- ---- ------------- -- ------- │0 Init 0 39 0 0 Start at 39 │1 CreateBtree 0 1 2 0 r[1]=root iDb=0 flags=2 │2 OpenWriteAsync 0 1 0 0 │3 OpenWriteAwait 0 0 0 0 │4 NewRowId 0 2 0 0 │5 String8 0 3 0 index 0 r[3]='index' │6 String8 0 4 0 idxp 0 r[4]='idxp' │7 String8 0 5 0 products 0 r[5]='products' │8 Copy 1 6 1 0 r[6]=r[1] │9 String8 0 7 0 CREATE INDEX idxp ON products (price) 0 r[7]='CREATE INDEX idxp ON products (price)' │10 MakeRecord 3 5 8 0 r[8]=mkrec(r[3..7]) │11 InsertAsync 0 8 2 0 │12 InsertAwait 0 0 0 0 │13 SorterOpen 3 1 0 k(1,B) 0 cursor=3 │14 OpenPseudo 4 9 2 0 2 columns in r[9] │15 OpenReadAsync 2 273 0 0 table=products, root=273 │16 OpenReadAwait 0 0 0 0 │17 RewindAsync 2 0 0 0 │18 RewindAwait 2 25 0 0 Rewind table products │19 Column 2 2 10 0 r[10]=products.price │20 RowId 2 11 0 0 r[11]=products.rowid │21 MakeRecord 10 2 12 0 r[12]=mkrec(r[10..11]) │22 SorterInsert 3 12 0 0 0 key=r[12] │23 NextAsync 2 0 0 0 │24 NextAwait 2 19 0 0 │25 OpenWriteAsync 1 1 0 0 │26 OpenWriteAwait 0 0 0 0 │27 SorterSort 3 33 0 0 │28 SorterData 3 13 4 0 r[13]=data │29 SeekEnd 1 0 0 0 │30 IdxInsertAsync 1 13 0 0 key=r[13] │31 IdxInsertAwait 1 0 0 0 │32 SorterNext 3 28 0 0 │33 Close 3 0 0 0 │34 Close 2 0 0 0 │35 Close 1 0 0 0 │36 ParseSchema 0 0 0 name = 'idxp' AND type = 'index' 0 name = 'idxp' AND type = 'index' │37 Close 0 0 0 0 │38 Halt 0 0 0 0 │39 Transaction 0 1 0 0 write=true │40 Goto 0 1 0 0 ``` This will create the initial index btree and insert whatever relevant records that need to be inserted, it doesn't handle the case of inserting new index keys when normal records are created afterwards. That will prob be added in next PR to keep this one concise. Limbo will properly use the index in a subsequent query:  Creating a unique index on a column that has 2 existing identical rows:  Reviewed-by: Pere Diaz Bou <pere-altea@homail.com> Reviewed-by: Jussi Saurio <jussi.saurio@gmail.com> Closes #1199
550 lines
19 KiB
Rust
550 lines
19 KiB
Rust
use std::ops::Deref;
|
|
use std::rc::Rc;
|
|
|
|
use limbo_sqlite3_parser::ast::{
|
|
DistinctNames, Expr, InsertBody, OneSelect, QualifiedName, ResolveType, ResultColumn, With,
|
|
};
|
|
|
|
use crate::error::SQLITE_CONSTRAINT_PRIMARYKEY;
|
|
use crate::schema::Table;
|
|
use crate::util::normalize_ident;
|
|
use crate::vdbe::builder::{ProgramBuilderOpts, QueryMode};
|
|
use crate::vdbe::insn::RegisterOrLiteral;
|
|
use crate::vdbe::BranchOffset;
|
|
use crate::{
|
|
schema::{Column, Schema},
|
|
translate::expr::translate_expr,
|
|
vdbe::{
|
|
builder::{CursorType, ProgramBuilder},
|
|
insn::Insn,
|
|
},
|
|
SymbolTable,
|
|
};
|
|
use crate::{Result, VirtualTable};
|
|
|
|
use super::emitter::Resolver;
|
|
|
|
#[allow(clippy::too_many_arguments)]
|
|
pub fn translate_insert(
|
|
query_mode: QueryMode,
|
|
schema: &Schema,
|
|
with: &Option<With>,
|
|
on_conflict: &Option<ResolveType>,
|
|
tbl_name: &QualifiedName,
|
|
columns: &Option<DistinctNames>,
|
|
body: &InsertBody,
|
|
_returning: &Option<Vec<ResultColumn>>,
|
|
syms: &SymbolTable,
|
|
) -> Result<ProgramBuilder> {
|
|
let mut program = ProgramBuilder::new(ProgramBuilderOpts {
|
|
query_mode,
|
|
num_cursors: 1,
|
|
approx_num_insns: 30,
|
|
approx_num_labels: 5,
|
|
});
|
|
if with.is_some() {
|
|
crate::bail_parse_error!("WITH clause is not supported");
|
|
}
|
|
if on_conflict.is_some() {
|
|
crate::bail_parse_error!("ON CONFLICT clause is not supported");
|
|
}
|
|
|
|
let table_name = &tbl_name.name;
|
|
let table = match schema.get_table(table_name.0.as_str()) {
|
|
Some(table) => table,
|
|
None => crate::bail_corrupt_error!("Parse error: no such table: {}", table_name),
|
|
};
|
|
let resolver = Resolver::new(syms);
|
|
if let Some(virtual_table) = &table.virtual_table() {
|
|
translate_virtual_table_insert(
|
|
&mut program,
|
|
virtual_table.clone(),
|
|
columns,
|
|
body,
|
|
on_conflict,
|
|
&resolver,
|
|
)?;
|
|
return Ok(program);
|
|
}
|
|
let init_label = program.allocate_label();
|
|
program.emit_insn(Insn::Init {
|
|
target_pc: init_label,
|
|
});
|
|
let start_offset = program.offset();
|
|
|
|
let Some(btree_table) = table.btree() else {
|
|
crate::bail_corrupt_error!("Parse error: no such table: {}", table_name);
|
|
};
|
|
if !btree_table.has_rowid {
|
|
crate::bail_parse_error!("INSERT into WITHOUT ROWID table is not supported");
|
|
}
|
|
|
|
let cursor_id = program.alloc_cursor_id(
|
|
Some(table_name.0.clone()),
|
|
CursorType::BTreeTable(btree_table.clone()),
|
|
);
|
|
let root_page = btree_table.root_page;
|
|
let values = match body {
|
|
InsertBody::Select(select, _) => match &select.body.select.deref() {
|
|
OneSelect::Values(values) => values,
|
|
_ => todo!(),
|
|
},
|
|
InsertBody::DefaultValues => &vec![vec![]],
|
|
};
|
|
|
|
let column_mappings = resolve_columns_for_insert(&table, columns, values)?;
|
|
// Check if rowid was provided (through INTEGER PRIMARY KEY as a rowid alias)
|
|
let rowid_alias_index = btree_table.columns.iter().position(|c| c.is_rowid_alias);
|
|
let has_user_provided_rowid = {
|
|
assert_eq!(column_mappings.len(), btree_table.columns.len());
|
|
if let Some(index) = rowid_alias_index {
|
|
column_mappings[index].value_index.is_some()
|
|
} else {
|
|
false
|
|
}
|
|
};
|
|
|
|
// allocate a register for each column in the table. if not provided by user, they will simply be set as null.
|
|
// allocate an extra register for rowid regardless of whether user provided a rowid alias column.
|
|
let num_cols = btree_table.columns.len();
|
|
let rowid_reg = program.alloc_registers(num_cols + 1);
|
|
let column_registers_start = rowid_reg + 1;
|
|
let rowid_alias_reg = {
|
|
if has_user_provided_rowid {
|
|
Some(column_registers_start + rowid_alias_index.unwrap())
|
|
} else {
|
|
None
|
|
}
|
|
};
|
|
|
|
let record_register = program.alloc_register();
|
|
let halt_label = program.allocate_label();
|
|
let mut loop_start_offset = BranchOffset::Offset(0);
|
|
|
|
let inserting_multiple_rows = values.len() > 1;
|
|
|
|
// Multiple rows - use coroutine for value population
|
|
if inserting_multiple_rows {
|
|
let yield_reg = program.alloc_register();
|
|
let jump_on_definition_label = program.allocate_label();
|
|
program.emit_insn(Insn::InitCoroutine {
|
|
yield_reg,
|
|
jump_on_definition: jump_on_definition_label,
|
|
start_offset: program.offset().add(1u32),
|
|
});
|
|
|
|
for value in values {
|
|
populate_column_registers(
|
|
&mut program,
|
|
value,
|
|
&column_mappings,
|
|
column_registers_start,
|
|
true,
|
|
rowid_reg,
|
|
&resolver,
|
|
)?;
|
|
program.emit_insn(Insn::Yield {
|
|
yield_reg,
|
|
end_offset: halt_label,
|
|
});
|
|
}
|
|
program.emit_insn(Insn::EndCoroutine { yield_reg });
|
|
program.resolve_label(jump_on_definition_label, program.offset());
|
|
|
|
program.emit_insn(Insn::OpenWriteAsync {
|
|
cursor_id,
|
|
root_page: RegisterOrLiteral::Literal(root_page),
|
|
});
|
|
program.emit_insn(Insn::OpenWriteAwait {});
|
|
|
|
// Main loop
|
|
// FIXME: rollback is not implemented. E.g. if you insert 2 rows and one fails to unique constraint violation,
|
|
// the other row will still be inserted.
|
|
loop_start_offset = program.offset();
|
|
program.emit_insn(Insn::Yield {
|
|
yield_reg,
|
|
end_offset: halt_label,
|
|
});
|
|
} else {
|
|
// Single row - populate registers directly
|
|
program.emit_insn(Insn::OpenWriteAsync {
|
|
cursor_id,
|
|
root_page: RegisterOrLiteral::Literal(root_page),
|
|
});
|
|
program.emit_insn(Insn::OpenWriteAwait {});
|
|
|
|
populate_column_registers(
|
|
&mut program,
|
|
&values[0],
|
|
&column_mappings,
|
|
column_registers_start,
|
|
false,
|
|
rowid_reg,
|
|
&resolver,
|
|
)?;
|
|
}
|
|
|
|
// Common record insertion logic for both single and multiple rows
|
|
let check_rowid_is_integer_label = rowid_alias_reg.and(Some(program.allocate_label()));
|
|
if let Some(reg) = rowid_alias_reg {
|
|
// for the row record, the rowid alias column (INTEGER PRIMARY KEY) is always set to NULL
|
|
// and its value is copied to the rowid register. in the case where a single row is inserted,
|
|
// the value is written directly to the rowid register (see populate_column_registers()).
|
|
// again, not sure why this only happens in the single row case, but let's mimic sqlite.
|
|
// in the single row case we save a Copy instruction, but in the multiple rows case we do
|
|
// it here in the loop.
|
|
if inserting_multiple_rows {
|
|
program.emit_insn(Insn::Copy {
|
|
src_reg: reg,
|
|
dst_reg: rowid_reg,
|
|
amount: 0, // TODO: rename 'amount' to something else; amount==0 means 1
|
|
});
|
|
// for the row record, the rowid alias column is always set to NULL
|
|
program.emit_insn(Insn::SoftNull { reg });
|
|
}
|
|
// the user provided rowid value might itself be NULL. If it is, we create a new rowid on the next instruction.
|
|
program.emit_insn(Insn::NotNull {
|
|
reg: rowid_reg,
|
|
target_pc: check_rowid_is_integer_label.unwrap(),
|
|
});
|
|
}
|
|
|
|
// Create new rowid if a) not provided by user or b) provided by user but is NULL
|
|
program.emit_insn(Insn::NewRowid {
|
|
cursor: cursor_id,
|
|
rowid_reg,
|
|
prev_largest_reg: 0,
|
|
});
|
|
|
|
if let Some(must_be_int_label) = check_rowid_is_integer_label {
|
|
program.resolve_label(must_be_int_label, program.offset());
|
|
// If the user provided a rowid, it must be an integer.
|
|
program.emit_insn(Insn::MustBeInt { reg: rowid_reg });
|
|
}
|
|
|
|
// Check uniqueness constraint for rowid if it was provided by user.
|
|
// When the DB allocates it there are no need for separate uniqueness checks.
|
|
if has_user_provided_rowid {
|
|
let make_record_label = program.allocate_label();
|
|
program.emit_insn(Insn::NotExists {
|
|
cursor: cursor_id,
|
|
rowid_reg,
|
|
target_pc: make_record_label,
|
|
});
|
|
let rowid_column_name = if let Some(index) = rowid_alias_index {
|
|
btree_table
|
|
.columns
|
|
.get(index)
|
|
.unwrap()
|
|
.name
|
|
.as_ref()
|
|
.expect("column name is None")
|
|
} else {
|
|
"rowid"
|
|
};
|
|
|
|
program.emit_insn(Insn::Halt {
|
|
err_code: SQLITE_CONSTRAINT_PRIMARYKEY,
|
|
description: format!("{}.{}", table_name.0, rowid_column_name),
|
|
});
|
|
|
|
program.resolve_label(make_record_label, program.offset());
|
|
}
|
|
|
|
// Create and insert the record
|
|
program.emit_insn(Insn::MakeRecord {
|
|
start_reg: column_registers_start,
|
|
count: num_cols,
|
|
dest_reg: record_register,
|
|
});
|
|
|
|
program.emit_insn(Insn::InsertAsync {
|
|
cursor: cursor_id,
|
|
key_reg: rowid_reg,
|
|
record_reg: record_register,
|
|
flag: 0,
|
|
});
|
|
program.emit_insn(Insn::InsertAwait { cursor_id });
|
|
|
|
if inserting_multiple_rows {
|
|
// For multiple rows, loop back
|
|
program.emit_insn(Insn::Goto {
|
|
target_pc: loop_start_offset,
|
|
});
|
|
}
|
|
|
|
program.resolve_label(halt_label, program.offset());
|
|
program.emit_insn(Insn::Halt {
|
|
err_code: 0,
|
|
description: String::new(),
|
|
});
|
|
|
|
program.resolve_label(init_label, program.offset());
|
|
program.emit_insn(Insn::Transaction { write: true });
|
|
program.emit_constant_insns();
|
|
program.emit_insn(Insn::Goto {
|
|
target_pc: start_offset,
|
|
});
|
|
|
|
Ok(program)
|
|
}
|
|
|
|
#[derive(Debug)]
|
|
/// Represents how a column should be populated during an INSERT.
|
|
/// Contains both the column definition and optionally the index into the VALUES tuple.
|
|
struct ColumnMapping<'a> {
|
|
/// Reference to the column definition from the table schema
|
|
column: &'a Column,
|
|
/// If Some(i), use the i-th value from the VALUES tuple
|
|
/// If None, use NULL (column was not specified in INSERT statement)
|
|
value_index: Option<usize>,
|
|
/// The default value for the column, if defined
|
|
default_value: Option<&'a Expr>,
|
|
}
|
|
|
|
/// Resolves how each column in a table should be populated during an INSERT.
|
|
/// Returns a Vec of ColumnMapping, one for each column in the table's schema.
|
|
///
|
|
/// For each column, specifies:
|
|
/// 1. The column definition (type, constraints, etc)
|
|
/// 2. Where to get the value from:
|
|
/// - Some(i) -> use i-th value from the VALUES tuple
|
|
/// - None -> use NULL (column wasn't specified in INSERT)
|
|
///
|
|
/// Two cases are handled:
|
|
/// 1. No column list specified (INSERT INTO t VALUES ...):
|
|
/// - Values are assigned to columns in table definition order
|
|
/// - If fewer values than columns, remaining columns map to None
|
|
/// 2. Column list specified (INSERT INTO t (col1, col3) VALUES ...):
|
|
/// - Named columns map to their corresponding value index
|
|
/// - Unspecified columns map to None
|
|
fn resolve_columns_for_insert<'a>(
|
|
table: &'a Table,
|
|
columns: &Option<DistinctNames>,
|
|
values: &[Vec<Expr>],
|
|
) -> Result<Vec<ColumnMapping<'a>>> {
|
|
if values.is_empty() {
|
|
crate::bail_parse_error!("no values to insert");
|
|
}
|
|
|
|
let table_columns = &table.columns();
|
|
|
|
// Case 1: No columns specified - map values to columns in order
|
|
if columns.is_none() {
|
|
let num_values = values[0].len();
|
|
if num_values > table_columns.len() {
|
|
crate::bail_parse_error!(
|
|
"table {} has {} columns but {} values were supplied",
|
|
&table.get_name(),
|
|
table_columns.len(),
|
|
num_values
|
|
);
|
|
}
|
|
|
|
// Verify all value tuples have same length
|
|
for value in values.iter().skip(1) {
|
|
if value.len() != num_values {
|
|
crate::bail_parse_error!("all VALUES must have the same number of terms");
|
|
}
|
|
}
|
|
|
|
// Map each column to either its corresponding value index or None
|
|
return Ok(table_columns
|
|
.iter()
|
|
.enumerate()
|
|
.map(|(i, col)| ColumnMapping {
|
|
column: col,
|
|
value_index: if i < num_values { Some(i) } else { None },
|
|
default_value: col.default.as_ref(),
|
|
})
|
|
.collect());
|
|
}
|
|
|
|
// Case 2: Columns specified - map named columns to their values
|
|
let mut mappings: Vec<_> = table_columns
|
|
.iter()
|
|
.map(|col| ColumnMapping {
|
|
column: col,
|
|
value_index: None,
|
|
default_value: col.default.as_ref(),
|
|
})
|
|
.collect();
|
|
|
|
// Map each named column to its value index
|
|
for (value_index, column_name) in columns.as_ref().unwrap().iter().enumerate() {
|
|
let column_name = normalize_ident(column_name.0.as_str());
|
|
let table_index = table_columns.iter().position(|c| {
|
|
c.name
|
|
.as_ref()
|
|
.map_or(false, |name| name.eq_ignore_ascii_case(&column_name))
|
|
});
|
|
|
|
if table_index.is_none() {
|
|
crate::bail_parse_error!(
|
|
"table {} has no column named {}",
|
|
&table.get_name(),
|
|
column_name
|
|
);
|
|
}
|
|
|
|
mappings[table_index.unwrap()].value_index = Some(value_index);
|
|
}
|
|
|
|
Ok(mappings)
|
|
}
|
|
|
|
/// Populates the column registers with values for a single row
|
|
fn populate_column_registers(
|
|
program: &mut ProgramBuilder,
|
|
value: &[Expr],
|
|
column_mappings: &[ColumnMapping],
|
|
column_registers_start: usize,
|
|
inserting_multiple_rows: bool,
|
|
rowid_reg: usize,
|
|
resolver: &Resolver,
|
|
) -> Result<()> {
|
|
for (i, mapping) in column_mappings.iter().enumerate() {
|
|
let target_reg = column_registers_start + i;
|
|
|
|
// Column has a value in the VALUES tuple
|
|
if let Some(value_index) = mapping.value_index {
|
|
// When inserting a single row, SQLite writes the value provided for the rowid alias column (INTEGER PRIMARY KEY)
|
|
// directly into the rowid register and writes a NULL into the rowid alias column. Not sure why this only happens
|
|
// in the single row case, but let's copy it.
|
|
let write_directly_to_rowid_reg =
|
|
mapping.column.is_rowid_alias && !inserting_multiple_rows;
|
|
let reg = if write_directly_to_rowid_reg {
|
|
rowid_reg
|
|
} else {
|
|
target_reg
|
|
};
|
|
translate_expr(
|
|
program,
|
|
None,
|
|
value.get(value_index).expect("value index out of bounds"),
|
|
reg,
|
|
resolver,
|
|
)?;
|
|
if write_directly_to_rowid_reg {
|
|
program.emit_insn(Insn::SoftNull { reg: target_reg });
|
|
}
|
|
} else if let Some(default_expr) = mapping.default_value {
|
|
translate_expr(program, None, default_expr, target_reg, resolver)?;
|
|
} else {
|
|
// Column was not specified as has no DEFAULT - use NULL if it is nullable, otherwise error
|
|
// Rowid alias columns can be NULL because we will autogenerate a rowid in that case.
|
|
let is_nullable = !mapping.column.primary_key || mapping.column.is_rowid_alias;
|
|
if is_nullable {
|
|
program.emit_insn(Insn::Null {
|
|
dest: target_reg,
|
|
dest_end: None,
|
|
});
|
|
program.mark_last_insn_constant();
|
|
} else {
|
|
crate::bail_parse_error!(
|
|
"column {} is not nullable",
|
|
mapping.column.name.as_ref().expect("column name is None")
|
|
);
|
|
}
|
|
}
|
|
}
|
|
Ok(())
|
|
}
|
|
|
|
fn translate_virtual_table_insert(
|
|
program: &mut ProgramBuilder,
|
|
virtual_table: Rc<VirtualTable>,
|
|
columns: &Option<DistinctNames>,
|
|
body: &InsertBody,
|
|
on_conflict: &Option<ResolveType>,
|
|
resolver: &Resolver,
|
|
) -> Result<()> {
|
|
let init_label = program.allocate_label();
|
|
program.emit_insn(Insn::Init {
|
|
target_pc: init_label,
|
|
});
|
|
let start_offset = program.offset();
|
|
|
|
let values = match body {
|
|
InsertBody::Select(select, None) => match &select.body.select.deref() {
|
|
OneSelect::Values(values) => values,
|
|
_ => crate::bail_parse_error!("Virtual tables only support VALUES clause in INSERT"),
|
|
},
|
|
InsertBody::DefaultValues => &vec![],
|
|
_ => crate::bail_parse_error!("Unsupported INSERT body for virtual tables"),
|
|
};
|
|
|
|
let table = Table::Virtual(virtual_table.clone());
|
|
let column_mappings = resolve_columns_for_insert(&table, columns, values)?;
|
|
|
|
let value_registers_start = program.alloc_registers(values[0].len());
|
|
for (i, expr) in values[0].iter().enumerate() {
|
|
translate_expr(program, None, expr, value_registers_start + i, resolver)?;
|
|
}
|
|
/* *
|
|
* Inserts for virtual tables are done in a single step.
|
|
* argv[0] = (NULL for insert)
|
|
* argv[1] = (NULL for insert)
|
|
* argv[2..] = column values
|
|
* */
|
|
|
|
let rowid_reg = program.alloc_registers(column_mappings.len() + 3);
|
|
let insert_rowid_reg = rowid_reg + 1; // argv[1] = insert_rowid
|
|
let data_start_reg = rowid_reg + 2; // argv[2..] = column values
|
|
|
|
program.emit_insn(Insn::Null {
|
|
dest: rowid_reg,
|
|
dest_end: None,
|
|
});
|
|
program.emit_insn(Insn::Null {
|
|
dest: insert_rowid_reg,
|
|
dest_end: None,
|
|
});
|
|
|
|
for (i, mapping) in column_mappings.iter().enumerate() {
|
|
let target_reg = data_start_reg + i;
|
|
if let Some(value_index) = mapping.value_index {
|
|
program.emit_insn(Insn::Copy {
|
|
src_reg: value_registers_start + value_index,
|
|
dst_reg: target_reg,
|
|
amount: 1,
|
|
});
|
|
} else {
|
|
program.emit_insn(Insn::Null {
|
|
dest: target_reg,
|
|
dest_end: None,
|
|
});
|
|
}
|
|
}
|
|
|
|
let conflict_action = on_conflict.as_ref().map(|c| c.bit_value()).unwrap_or(0) as u16;
|
|
|
|
let cursor_id = program.alloc_cursor_id(
|
|
Some(virtual_table.name.clone()),
|
|
CursorType::VirtualTable(virtual_table.clone()),
|
|
);
|
|
|
|
program.emit_insn(Insn::VUpdate {
|
|
cursor_id,
|
|
arg_count: column_mappings.len() + 2,
|
|
start_reg: rowid_reg,
|
|
vtab_ptr: virtual_table.implementation.as_ref().ctx as usize,
|
|
conflict_action,
|
|
});
|
|
|
|
let halt_label = program.allocate_label();
|
|
program.emit_insn(Insn::Halt {
|
|
err_code: 0,
|
|
description: String::new(),
|
|
});
|
|
|
|
program.resolve_label(halt_label, program.offset());
|
|
program.resolve_label(init_label, program.offset());
|
|
|
|
program.emit_insn(Insn::Goto {
|
|
target_pc: start_offset,
|
|
});
|
|
|
|
Ok(())
|
|
}
|