[waw666waw666]

Description:

Delete and DeleteFound treat a path key that contains a backslash differently from Get, Set and EachKey, and in one case delete the wrong part of the document.

data := []byte(`{"\"q":1,"keep":0,"a\\b":2}`)

v, _, _, _ := jsonparser.Get(data, `a\b`)       // 2
out, found := jsonparser.DeleteFound(data, `a\b`)
// out = {}, found = true   ("\"q" and "keep" are gone too)

jsonparser.Delete([]byte(`{"a\\b":1,"c":2}`), `a\b`)
// {"a\\b":1,"c":2}   (nothing deleted, although Get finds the key)

Cause: findKeyStartConfig, which Delete uses for the last path element, JSON-unescapes the path key before comparing it with the document's keys. searchKeys (used by Get, Set, EachKey) compares the path as is against the unescaped document key, and the existing Get tests rely on that ({"a\\b":1} is found with the path a\b). So in Delete, \b in the path turned into a backspace and the key was not found.

The unescaped path key was also stored in the same stackbuf that is then used to unescape each escaped key of the document. The first escaped key of the same length overwrote the path key and so compared equal to itself, and findKeyStart returned its offset. The value end still came from internalGet with the real path, so Delete cut from the earlier key through the end of the matched value, removing every member in between.

Fix: compare the path key as is in findKeyStartConfig, like searchKeys does. That removes both the mismatch and the shared buffer.

Tests: added cases to deleteTests (escaped key, Unicode-escaped key, an earlier escaped key of the same length at top level and nested, and a path that only equals the escaped form of a key). Three of them fail on master and pass with this change. go vet shows no new findings and gofmt is clean for the changed lines. The full go test ./... result is the same as on master: TestOracleSetPr286Regression already fails on master, and #290 looks related.

Benchmark before change:

BenchmarkJsonParserDeleteMedium-2   297740   8943 ns/op   10175 B/op   7 allocs/op
BenchmarkJsonParserDelSmall-2      1417653   1694 ns/op     848 B/op   8 allocs/op

Benchmark after change:

BenchmarkJsonParserDeleteMedium-2   347984   6763 ns/op    9984 B/op   4 allocs/op
BenchmarkJsonParserDelSmall-2      1774606   1289 ns/op     591 B/op   3 allocs/op

(go test -bench JsonParserDel -benchmem ./benchmark/, arm64. The time numbers are noisy on this machine. The drop in allocations is consistent: stackbuf no longer escapes, because the path key isn't stored in it anymore.)