Re: [2/5] C-SKY port: Backend implementation

[Jeff Law <law@redhat.com>]

On 07/26/2018 12:06 AM, 枚脛脧脡铆碌 wrote:
> 
>> 脭脷 2018脛锚7脭脗25脠脮拢卢脡脧脦莽5:24拢卢Jeff Law <law@redhat.com> 脨麓碌脌拢潞
>>
>> On 07/24/2018 12:18 PM, Sandra Loosemore wrote:
>>> On 07/24/2018 09:45 AM, Jeff Law wrote:
>>>> On 07/23/2018 10:21 PM, Sandra Loosemore wrote:
>>>> I'm not a big fan of more awk code, but I'm not going to object to it :-)
>>>>
>>>> Why does the port have its own little pass for condition code
>>>> optimization (cse_cc)?  What is it doing that can't be done with our
>>>> generic optimizers?
>>>
>>> This pass was included in the initial patch set we got from C-SKY, and
>>> as it didn't seem to break anything I left it in.  Perhaps C-SKY can
>>> provide a testcase that demonstrates why it's still useful in the
>>> current version of GCC; otherwise we can remove this from the initial
>>> port submission and restore it later if some performance analysis shows
>>> it is still worthwhile.
>> FWIW it looks like we model CC setting on just a few insns, (add,
>> subtract) so I'd be surprised if this little mini pass found much.  I'd
>> definitely like to hear from the csky authors here.
>>
>> Alternately, you could do add some instrumentation to flag when it
>> triggers, take a test or two that does, reduce it and we can then look
>> at the key RTL sequences and see what the pass is really doing.
>>
> 
> I wrote a case to reproduce this problem on C-SKY. C code is as follows:
> -----------------------------------------------------------------------
> int e1, e2;
> 
> void func (int a, int b, int c, int d, int f, int g)
> {
>   e1 = a > b ? f : g;
>   e2 = a > b ? c : d;
> 
>   return;
> }
> -----------------------------------------------------------------------
> 
> compile to assembler with option 隆掳-O3 -S隆卤 :
> -----------------------------------------------------------------------
> func:
>   cmplt a1, a0
>   ld.w  t1, (sp, 0)
>   ld.w  t0, (sp, 4)
>   movt  t0, t1
>   cmplt a1, a0
>   movt  a3, a2
>   lrw a1, e2
>   lrw a2, e1
>   st.w  a3, (a1, 0)
>   st.w  t0, (a2, 0)
>   rts
> -----------------------------------------------------------------------
> There is an extra 隆掳cmplt a1, a0" in the above code without cse_cc. This situation mainly occurs when a relatively short branch jump is converted into a conditional execution instruction. And the CSE pass can not reduce the same conditional comparison instruction . Here is the rtx sequence after 隆掳cse2隆卤 pass.
> 
> -----------------------------------------------------------------------
> (insn 28 13 29 2 (set (reg:CC 33 c)
>         (gt:CC (reg/v:SI 77 [ a ])
>             (reg/v:SI 78 [ b ]))) func.c:5 1099 {*cmpgtsi}
>      (nil))
> (insn 29 28 30 2 (set (reg/v:SI 82 [ g ])
>         (if_then_else:SI (eq (reg:CC 33 c)
>                 (const_int 0 [0]))
>             (reg/v:SI 82 [ g ])
>             (reg/v:SI 81 [ f ]))) func.c:5 983 {movf}
>      (expr_list:REG_DEAD (reg/v:SI 81 [ f ])
>         (expr_list:REG_DEAD (reg:CC 33 c)
>             (nil))))
> (insn 30 29 31 2 (set (reg:CC 33 c)
>         (gt:CC (reg/v:SI 77 [ a ])
>             (reg/v:SI 78 [ b ]))) func.c:5 1099 {*cmpgtsi}
>      (expr_list:REG_DEAD (reg/v:SI 78 [ b ])
>         (expr_list:REG_DEAD (reg/v:SI 77 [ a ])
>             (nil))))
> (insn 31 30 18 2 (set (reg/v:SI 80 [ d ])
>         (if_then_else:SI (eq (reg:CC 33 c)
>                 (const_int 0 [0]))
>             (reg/v:SI 80 [ d ])
>             (reg/v:SI 79 [ c ]))) func.c:5 983 {movf}
>      (expr_list:REG_DEAD (reg/v:SI 79 [ c ])
>         (expr_list:REG_DEAD (reg:CC 33 c)
>             (nil))))
> -----------------------------------------------------------------------
> 
> It doesn't seem to check the same conditional comparison instruction .So I wrote this to solve this problem, but I am not sure if this is the best way : )
> 
> PS, the same conditional comparison instruction cannot be reduced with the latest version gcc with C-SKY because I just insert the 隆掳cse_cc隆卤 after 隆掳cse1隆卤, when I insert after 隆掳cse2隆卤, this problem can be solved very well.
I think the cse_cc pass is really just working around one or more bugs
in CSE and/or a backend bug.  The RTL above clearly shows a common
subexpression that is not eliminated by CSE.

What CSE should be trying to do is changing the second and third
occurrences of (gt:CC (reg 77) (reg 78)) with (reg 33) which would
create nop-sets which would be subsequently deleted.  I suspect you do
not have an insn which matches that nop set of the CC register.  If you
fix that I suspect CSE will work better and eliminate the need for your
cse_cc pass.

jeff