tarantool / luajit / 19798590789

93%

tarantool/master: 93%

push

github

ARM64: Fix assembly of HREFK (again).

Thanks to Peter Cawley.

(cherry picked from commit 93ce12ee1)

When assembling the HREFK IR with the huge offset of the target node
from the table, this offset calculation and the key loading from the
node are emitted like the following:
|  ldr   x16, [x2, 40]
|  add   x16, x16, x21
|  ldr   x27, [x16, 8]
|  cmp   x27, x17

Here, `x16` is the node register, `x27` is the key register, and `x21`
is the register containing the offset.

It is possible that the register for holding the constant operand in the
addition may be chosen as the same register containing the node address,
since the full `RSET_GPR` is given to the `emit_opk()`. It will result
in the following invalid mcode:
|  ldr   x27, [x2, 40]
|  str   x27, [sp, 8]
|  add   x16, x16, x16
|  ldr   x16, [sp, 8]
|  ldr   x27, [x16, 8]
|  cmp   x27, x17

It seems that in the current implementation the LuaJIT's register
allocator always prefers the register holding the key instead, so this
does not lead to the invalid emitting. Hence, it is impossible to come
up with any valid reproducer. However, to avoid possible regressions in
the future, this patch fixes the invalid register set by excluding the
node register from it.

Sergey Kaplun:
* added the description for the problem

Part of tarantool/tarantool#11691

Reviewed-by: Sergey Bronnikov <sergeyb@tarantool.org>
Signed-off-by: Sergey Kaplun <skaplun@tarantool.org>
(cherry picked from commit fd5e7cb52)

5712 of 6046 branches covered (94.48%)

Branch coverage included in aggregate %.

21797 of 23513 relevant lines covered (92.7%)

3873205.72 hits per line