diff options
author | Ian Lance Taylor <iant@golang.org> | 2016-01-29 12:39:47 -0800 |
---|---|---|
committer | Ian Lance Taylor <iant@golang.org> | 2016-01-29 12:39:47 -0800 |
commit | b76c61051faaf0baf3215e6f811003d98639b702 (patch) | |
tree | 8dfd72e1b4a60ffafe74d38a1f28ecc498bafde2 /sqlite3.go | |
parent | Merge pull request #266 from tcyrus/patch-1 (diff) | |
download | golite-b76c61051faaf0baf3215e6f811003d98639b702.tar.gz golite-b76c61051faaf0baf3215e6f811003d98639b702.tar.xz |
bind: pass &v[0] in direct call to C
In Go 1.6, the cgo checking rules are more precise when they see an
address operation as an argument to the C function. When you pass &v[0]
to a C function, the cgo check just verifies that v itself does not
contain any pointers. When you write `p := &v[0]` and then pass p to
the C function, the cgo check is conservative: it verifies that the
entire memory block to which p points does not contain any pointers.
When the bind function is called by code that passes a slice that is
part of a larger struct, this means that the cgo check will look at the
entire larger struct, not just the slice. This can cause a surprising
run time failure.
Avoid this problem by rewriting the code slightly to pass &v[0] in the
call to the C function itself.
In particular this fixes the tests of github.com/jmoiron/sqlx when using
Go 1.6.
Diffstat (limited to 'sqlite3.go')
-rw-r--r-- | sqlite3.go | 8 |
1 files changed, 4 insertions, 4 deletions
@@ -814,11 +814,11 @@ func (s *SQLiteStmt) bind(args []driver.Value) error { case float64: rv = C.sqlite3_bind_double(s.s, n, C.double(v)) case []byte: - var p *byte - if len(v) > 0 { - p = &v[0] + if len(v) == 0 { + rv = C._sqlite3_bind_blob(s.s, n, nil, 0) + } else { + rv = C._sqlite3_bind_blob(s.s, n, unsafe.Pointer(&v[0]), C.int(len(v))) } - rv = C._sqlite3_bind_blob(s.s, n, unsafe.Pointer(p), C.int(len(v))) case time.Time: b := []byte(v.Format(SQLiteTimestampFormats[0])) rv = C._sqlite3_bind_text(s.s, n, (*C.char)(unsafe.Pointer(&b[0])), C.int(len(b))) |