Hi Ruoyao,

Thanks for your suggestions regarding this patch, we have modified the code and create a new patch, you can find it via https://sourceware.org/pipermail/libc-alpha/2022-September/142016.html. The reason to using .align 6 is to make the loop starts with a cache line. Now we have optimzed the code and reduced some instructions, so it's not necessary to use 64-byte alignment any longer and it's removed accordingly.

Best regards,
Deng Jianbo
 
From: Xi Ruoyao
Date: 2022-09-02 17:35
To: dengjianbo; adhemerval.zanella; libc-alpha; i.swmail
CC: xuchenghua; joseph_myers; caiyinyu
Subject: Re: [PATCH 1/1] LoongArch: Add optimized functions.
I've not really run the code, see below for some comments after a quick
view.
 
On Fri, 2022-09-02 at 16:39 +0800, dengjianbo wrote:
 
/* snip */
 
What is "ANDROID_CHANGES"?  I can't find anything in Glibc building
system defining it and AFAIK Android is using another libc
implementation (I can't remember the name, but I'm sure it's not Glibc).
 
> +#ifdef ANDROID_CHANGES
> +LEAF(MEMMOVE_NAME, 0)
> +#else
> +LEAF(MEMMOVE_NAME)
 
/* snip */
 
> +start_unalign_proc:
> +       pcaddi          t1, 18
> +       slli.d          a6, a7, 3
> +       add.d           t1, t1, a6
> +       jirl            zero, t1, 0
 
"jr t1", to make the expression of all "jump register" operations
consistent.  Likewise for all "jirl zero, [something], 0".
 
> +end:
> +       move            v0, t8
> +       jr              ra
 
Do not use MIPS-style v0 or v1 aliases.  They mislead people to believe
"v0" was a different register than "a0".
 
/* snip */
 
> +#define L_ADDIU        addi.d
> +#define L_ADDU add.d
> +#define L_SUBU sub.d
 
Do not use MIPS-style "U" suffix.  LoongArch is not so MIPS-similar (we
don't have some "signaling arithmetic") and you don't want others to
believe so, do you? :)
 
> +LEAF(STRCHR)
> +       .align          6
 
Hmm, any rational to force a 64-byte alignment here?
 
 
-- 
Xi Ruoyao <xry111@xry111.site>
School of Aerospace Science and Technology, Xidian University