Merge 'Actually enforce uniqueness in create unique index' from Jussi Saurio

we just weren't doing it 🤡
Backport: 0.2
Closes #3568

Reviewed-by: Preston Thorpe <preston@turso.tech>

Closes #3571
This commit is contained in:
Pekka Enberg
2025-10-04 10:07:44 +03:00
committed by GitHub
7 changed files with 164 additions and 3 deletions

View File

@@ -1,11 +1,13 @@
use std::sync::Arc;
use crate::bail_parse_error;
use crate::error::SQLITE_CONSTRAINT_UNIQUE;
use crate::schema::{Table, RESERVED_TABLE_PREFIXES};
use crate::translate::emitter::{
emit_cdc_full_record, emit_cdc_insns, prepare_cdc_if_necessary, OperationMode, Resolver,
};
use crate::translate::expr::{translate_condition_expr, ConditionMetadata};
use crate::translate::insert::format_unique_violation_desc;
use crate::translate::plan::{
ColumnUsedMask, IterationDirection, JoinedTable, Operation, Scan, TableReferences,
};
@@ -91,6 +93,7 @@ pub fn translate_create_index(
};
let original_columns = columns;
let columns = resolve_sorted_columns(&tbl, columns)?;
let unique = unique_if_not_exists.0;
let idx = Arc::new(Index {
name: idx_name.clone(),
@@ -106,7 +109,7 @@ pub fn translate_create_index(
default: col.default.clone(),
})
.collect(),
unique: unique_if_not_exists.0,
unique,
ephemeral: false,
has_rowid: tbl.has_rowid,
// store the *original* where clause, because we need to rewrite it
@@ -295,8 +298,34 @@ pub fn translate_create_index(
cursor_id: sorter_cursor_id,
pc_if_empty: sorted_loop_end,
});
program.preassign_label_to_next_insn(sorted_loop_start);
let sorted_record_reg = program.alloc_register();
if unique {
// Since the records to be inserted are sorted, we can compare prev with current and if they are equal,
// we fall through to Halt with a unique constraint violation error.
let goto_label = program.allocate_label();
let label_after_sorter_compare = program.allocate_label();
program.resolve_label(goto_label, program.offset());
program.emit_insn(Insn::Goto {
target_pc: label_after_sorter_compare,
});
program.preassign_label_to_next_insn(sorted_loop_start);
program.emit_insn(Insn::SorterCompare {
cursor_id: sorter_cursor_id,
sorted_record_reg,
num_regs: columns.len(),
pc_when_nonequal: goto_label,
});
program.emit_insn(Insn::Halt {
err_code: SQLITE_CONSTRAINT_UNIQUE,
description: format_unique_violation_desc(tbl_name.as_str(), &idx),
});
program.preassign_label_to_next_insn(label_after_sorter_compare);
} else {
program.preassign_label_to_next_insn(sorted_loop_start);
}
program.emit_insn(Insn::SorterData {
pseudo_cursor: pseudo_cursor_id,
cursor_id: sorter_cursor_id,

View File

@@ -687,6 +687,12 @@ impl ProgramBuilder {
Insn::SorterSort { pc_if_empty, .. } => {
resolve(pc_if_empty, "SorterSort");
}
Insn::SorterCompare {
pc_when_nonequal: target_pc,
..
} => {
resolve(target_pc, "SorterCompare");
}
Insn::NotNull {
reg: _reg,
target_pc,

View File

@@ -4247,6 +4247,59 @@ pub fn op_sorter_next(
Ok(InsnFunctionStepResult::Step)
}
pub fn op_sorter_compare(
program: &Program,
state: &mut ProgramState,
insn: &Insn,
pager: &Arc<Pager>,
mv_store: Option<&Arc<MvStore>>,
) -> Result<InsnFunctionStepResult> {
load_insn!(
SorterCompare {
cursor_id,
sorted_record_reg,
num_regs,
pc_when_nonequal,
},
insn
);
let previous_sorter_values = {
let Register::Record(record) = &state.registers[*sorted_record_reg] else {
return Err(LimboError::InternalError(
"Sorted record must be a record".to_string(),
));
};
&record.get_values()[..*num_regs]
};
let cursor = state.get_cursor(*cursor_id);
let cursor = cursor.as_sorter_mut();
let Some(current_sorter_record) = cursor.record() else {
return Err(LimboError::InternalError(
"Sorter must have a record".to_string(),
));
};
let current_sorter_values = &current_sorter_record.get_values()[..*num_regs];
// If the current sorter record has a NULL in any of the significant fields, the comparison is not equal.
let is_equal = current_sorter_values
.iter()
.all(|v| !matches!(v, RefValue::Null))
&& compare_immutable(
previous_sorter_values,
current_sorter_values,
&cursor.index_key_info,
)
.is_eq();
if is_equal {
state.pc += 1;
} else {
state.pc = pc_when_nonequal.as_offset_int();
}
Ok(InsnFunctionStepResult::Step)
}
pub fn op_function(
program: &Program,
state: &mut ProgramState,

View File

@@ -1042,6 +1042,20 @@ pub fn insn_to_row(
0,
"".to_string(),
),
Insn::SorterCompare {
cursor_id,
pc_when_nonequal,
sorted_record_reg,
num_regs,
} => (
"SorterCompare",
*cursor_id as i32,
pc_when_nonequal.as_debug_int(),
*sorted_record_reg as i32,
Value::build_text(num_regs.to_string()),
0,
"".to_string(),
),
Insn::Function {
constant_mask,
start_reg,

View File

@@ -713,6 +713,18 @@ pub enum Insn {
record_reg: usize,
},
/// `cursor_id` is a sorter cursor. This instruction compares a prefix of the record blob in register `sorted_record_reg`
/// against a prefix of the entry that the sorter cursor currently points to.
/// Only the first `num_regs` fields of `sorted_record_reg` and the sorter record are compared.
/// Fall through to next instruction if the two records compare equal to each other.
/// Jump to `pc_when_nonequal` if they are different.
SorterCompare {
cursor_id: CursorID,
pc_when_nonequal: BranchOffset,
sorted_record_reg: usize,
num_regs: usize,
},
/// Sort the rows in the sorter.
SorterSort {
cursor_id: CursorID,
@@ -1263,6 +1275,7 @@ impl InsnVariants {
InsnVariants::SorterSort => execute::op_sorter_sort,
InsnVariants::SorterData => execute::op_sorter_data,
InsnVariants::SorterNext => execute::op_sorter_next,
InsnVariants::SorterCompare => execute::op_sorter_compare,
InsnVariants::Function => execute::op_function,
InsnVariants::Cast => execute::op_cast,
InsnVariants::InitCoroutine => execute::op_init_coroutine,

View File

@@ -62,7 +62,7 @@ pub struct Sorter {
/// The number of values in the key.
key_len: usize,
/// The key info.
index_key_info: Rc<Vec<KeyInfo>>,
pub index_key_info: Rc<Vec<KeyInfo>>,
/// Sorted chunks stored on disk.
chunks: Vec<SortedChunk>,
/// The heap of records consumed from the chunks and their corresponding chunk index.

View File

@@ -13,3 +13,49 @@ do_execsql_test_on_specific_db {:memory:} create-index-quoted-identifiers {
"CREATE INDEX \"idx idx\" ON \"t t\" (\"a a\")"
"CREATE UNIQUE INDEX \"unique idx idx\" ON \"t t\" (\"a a\")"
}
# single-column index key: creating unique index fails with duplicates
do_execsql_test_in_memory_any_error create-unique-index-with-duplicates-1 {
CREATE TABLE t1(a, b);
INSERT INTO t1 VALUES(1, 1);
INSERT INTO t1 VALUES(1, 2);
CREATE UNIQUE INDEX idx1 ON t1(a);
}
# multi-column index key: creating unique index fails with duplicates
do_execsql_test_in_memory_any_error create-unique-index-with-duplicates-2 {
CREATE TABLE t2(a, b, c);
INSERT INTO t2 VALUES(1, 2, 3);
INSERT INTO t2 VALUES(1, 2, 4);
CREATE UNIQUE INDEX idx2 ON t2(a, b);
}
# single-column index key: creating unique index succeeds because NULLs are never equal
do_execsql_test_on_specific_db {:memory:} create-unique-index-with-duplicates-3 {
CREATE TABLE t3(a);
INSERT INTO t3 VALUES(NULL);
INSERT INTO t3 VALUES(NULL);
CREATE UNIQUE INDEX idx3 ON t3(a);
SELECT count(*) FROM t3;
} {2}
# multi-column index key: creating unique index succeeds because NULLs are never equal
do_execsql_test_on_specific_db {:memory:} create-unique-index-with-duplicates-4 {
CREATE TABLE t4(a, b);
INSERT INTO t4 VALUES(1, NULL);
INSERT INTO t4 VALUES(1, NULL);
CREATE UNIQUE INDEX idx4 ON t4(a, b);
SELECT count(*) FROM t4;
} {2}
# multi-column index key: creating unique index succeeds when all NULLs
do_execsql_test_on_specific_db {:memory:} create-unique-index-with-duplicates-5 {
CREATE TABLE t5(a, b);
INSERT INTO t5 VALUES(NULL, NULL);
INSERT INTO t5 VALUES(NULL, NULL);
CREATE UNIQUE INDEX idx5 ON t5(a, b);
SELECT count(*) FROM t5;
} {2}