Hi, Richard.
Here is an example for vsext.vf2.
The general pattern for this instruction as follows:
(set (operand 0) (unspec:...(operand 1)))
We have a TARGET_MIN_VLEN macro which specify minimum VLEN according to -march
Consider this case V16QI -> V16HI of vsext.vf2
When TARGET_MIN_VLEN == 128
The constraint for vsext.vf should be:
(set (operand 0: V16HI "=vr,") (unspec:...(operand 1: V16QI "register
filter1, vr")))
Wheras when TARGET_MIN_VLEN > 128,
We are expecting the constraint:
(set (operand 0: V16HI "=") (unspec:...(operand 1: V16QI "vr")))
That is, same mode, same instruction pattern. We want the constraint to be
different according to TARGET_MIN_VLEN.
Currently, I don't know whether we can have an approach to support this feature.
Thanks.
juzhe.zh...@rivai.ai
From: Richard Sandiford
Date: 2023-11-24 02:18
To: 钟居哲
CC: gcc-patches; vmakarov; kito.cheng
Subject: Re: [PATCH 0/5] Add support for operand-specific alignment requirements
钟居哲 writes:
> Hi, Richard.
>
> Current define_mode_attr can only map an attribute for a mode.
> I wonder whether we can map a mode to multiple attributes ?
>
> E.g. (define_mode_attr dest_constraint [(V16QI "")])
>
> But I want it to be:
>
> (define_mode_attr dest_constraint [(V16QI (TARGET_MIN_VLEN <= 128 "vr")
> (TARGET_MIN_VLEN > 128 "")) ])
>
> It seems that we can't achieve this for now. Would it be possible we exend it
> in GCC-15 ?
That construct doesn't seem particularly natural to me. Could you explain
(conceptually) what kind of insn patterns you need?
Thanks,
Richard
>
>
> juzhe.zh...@rivai.ai
>
> From: Richard Sandiford
> Date: 2023-11-22 18:08
> To: juzhe.zhong\@rivai.ai
> CC: gcc-patches; vmakarov\@redhat.com; kito.cheng
> Subject: Re: [PATCH 0/5] Add support for operand-specific alignment
> requirements
> "juzhe.zh...@rivai.ai" writes:
>> Hi, Richard.
>>
>> Thanks for supporting register filter in IRA/LRA.
>> I found it is useful for RVV since we have a set of widen operations that
>> allow source register overlap highpart of dest register group
>>
>> For example, if vsext.vf2 v0(dest consume reg v0 and reg v1), v1 (source
>> consume v1 only)
>> I want to support the highpart overlap above. (Currently, we don't any
>> overlap between source and dest in such instructions).
>>
>> So, I wonder whether we can pass "machine_mode" into register filter. Ok, I
>> think it's too late since stage 1 closes. I wonder we can add it in GCC-15?
>
> I think adding a mode would add too much overhead. The mode would be
> the mode of the operand, but with subregs, the mode of the operand can
> be different from the mode of the RA allocno. So it would no longer
> be enough for the RA to calculate a bitmask of filters. It would need
> ro remember which modes are used with those filters.
>
> We'd also need to turn the current HARD_REG_SETs into [MAX_MACHINE_MODE]
> arrays of HARD_REG_SETs. (And there are now more than 256 machine modes
> for riscv.)
>
> The pattern that uses the constraints should already "know" the mode.
> So if possible, I think it would be better to use different constraints
> for different modes, using define_mode_attrs.
>
> Thanks,
> Richard
>
