Re: [PATCH v2] DSE: Bugfix ICE after allow vector type in get_stored_val

[Jeff Law <jeffreyalaw@gmail.com>]

On 3/4/24 11:22 PM, Li, Pan2 wrote:
> Thanks Jeff for comments.
> 
>> But in the case of a vector modes, we can usually reinterpret the
>> underlying bits in whatever mode we want and do any of the usual
>> operations on those bits.
> 
> Yes, I think that is why we can allow vector mode in get_stored_val if my understanding is correct.
> And then the different modes will return by gen_low_part. Unfortunately, there are some modes
>   (less than a vector bit size like V2SF, V2QI for vlen=128) are considered as invalid by validate_subreg,
> and return NULL_RTX result in the final ICE.
That doesn't make a lot of sense to me.  Even for vlen=128 I would have 
expected that we can still use a subreg to access low bits.  After all 
we might have had a V16QI vector and done a reduction of some sort 
storing the result in the first element and we have to be able to 
extract that result and move it around.

I'm not real keen on a target workaround.  While extremely safe, I 
wouldn't be surprised if other ports could trigger the ICE and we'd end 
up patching up multiple targets for what is, IMHO, a more generic issue.

As Richi noted using validate_subreg here isn't great.  Does it work to 
factor out this code from extract_low_bits:


>   if (!int_mode_for_mode (src_mode).exists (&src_int_mode)
>       || !int_mode_for_mode (mode).exists (&int_mode))
>     return NULL_RTX;
> 
>   if (!targetm.modes_tieable_p (src_int_mode, src_mode))
>     return NULL_RTX;
>   if (!targetm.modes_tieable_p (int_mode, mode))
>     return NULL_RTX;

And use that in the condition (and in extract_low_bits rather than 
duplicating the code)?

jeff

ps.  No need to apologize for the pings.  This completely fell off my radar.