Re: [Tarantool-patches] [PATCH luajit 2/3] ARM64: More improvements to the generation of immediates.

[Sergey Bronnikov via Tarantool-patches <tarantool-patches@dev.tarantool.org>]

[-- Attachment #1: Type: text/plain, Size: 1999 bytes --]

Hi, Sergey!

On 8/27/25 12:08, Sergey Kaplun wrote:
> Hi, Sergey!
> Thanks for the review!
> Please consider my answers below.
>
> On 25.08.25, Sergey Bronnikov wrote:
>> Hi, Sergey!
>>
>> thanks for the patch!
>>
>> In general LGTM, I would suggest fixing the description in commit message.
>>
>> See below.
>>
>> Sergey
>>
>> On 7/24/25 12:03, Sergey Kaplun wrote:
>>> From: Mike Pall <mike>
>>>
>>> (cherry picked from commit 69138082a3166105faa8cbb25fadb1e4298686c0)
>>>
>>> This patch refactors the emitting of immediates for the arm64
>>> architecture. The main changes are the following:
>>> * Use `emit_getgl()`, `emit_setgl()` instead of `emit_lso()`, where it
>>>     is possible, since it makes the code cleaner.
>>> * The `RID_GL` is allocated for `g` at the start of the trace emitting.
>>>     Also, this register is considered as a candidate to be used as a base
>>>     for the N-step offset in `emit_kdelta()`.
>>> * The address of `tmptv` is not rematerialized to the register from the
>>>     constant not. It is calculated via the adding the corresponding
> This "not" looks excessive. Rewritten as the following:
> | * The address of `tmptv` is not rematerialized to the register from the
> |   constant. It is calculated via the adding the corresponding offset to
> |   `RID_GL`.
>
>
>>>     offset to `RID_GL`.
>> it is not clear for me what for hunks with `emit_dm` are needed.
> | emit_dm(as, ins, d, m);
> Means emit the ins with values to the D and M instruction fields as
> registers `d`, `m` respectively.
>
> In the case of this patch, it emits simply:
> | mov rd, rm
> Where `rd` is the register associated with `ASM_REF_TMP1` (`REF_TRUE`)
> and `rm` is `RID_GL`. So this is simply moving the value of `g` from the
> `RID_GL` register to the register, which will be an argument for the C
> function call like `lj_gc_step_jit()`. Move is used instead of the
> constant value loading.
Thanks for explanation! I thought you will add it to the commit message.
>> <snipped>

[-- Attachment #2: Type: text/html, Size: 3107 bytes --]