Re: [PATCH 1/3] rtl: allow forming subregs of already unaligned mems [PR102125]

[Richard Earnshaw <Richard.Earnshaw@foss.arm.com>]

On 06/09/2021 11:58, Richard Biener via Gcc-patches wrote:
> On Mon, Sep 6, 2021 at 12:40 PM Richard Earnshaw <rearnsha@arm.com> wrote:
>>
>>
>> GCC was recently changed to prevent simplify_subreg from simplifying
>> a subreg of a mem when the mode of the new mem would have stricter alignment
>> constraints than the inner mem already has when the target requires
>> STRICT_ALIGNMENT.
>>
>> However, such targets may have specialist patterns that can handle
>> unaligned accesses and this restriction turns out to be unduly restrictive.
>> So limit this restriction to only apply when the inner mem is naturally
>> aligned to the inner mode.
> 
> Hmm, I think this can end up either generating wrong code or
> recog fails.  The specific combination of alignment and mode of 'op'
> has been validated to be supported, replacing the mode with sth
> else would need re-validation of the combination.  I'm not sure
> we can for example just query movmisalign support here and
> hope for LRA to reload the mem with that.
> 
> So - where do you run into this?  Is it possible to catch the
> situation on a higher level where more context as in the whole insn
> is visible?

I ran into it with patch 2 of this series when calling gen_highpart on a 
misaligned mem.  IIRC gen_highpart would end up returning (subreg:SI 
(mem:DI (addr [A8])) 4), while gen_lowpart would simplify the operation 
to (mem:SI (addr [A8])) as expected.

(subreg:SI (mem:DI (addr [A8])) 4) is really problematic, because it's 
not a memory_operand (from the manual: it will get reloaded into a 
register later on).  But that's no good here, I don't want this 
reloading into a wide register later, I need it to be narrowed to the 
component part now.

R.

> 
> Thanks,
> Richard.
> 
>>
>> gcc/ChangeLog:
>>
>>          PR target/102125
>>          * simplify-rtx.c (simplify_context::simplify_subreg): Allow
>>          simplifying (subreg (mem())) when the inner mem is already
>>          misaligned for its type.
>> ---
>>   gcc/simplify-rtx.c | 6 +++++-
>>   1 file changed, 5 insertions(+), 1 deletion(-)
>>