"Robin Dapp" <rdapp....@gmail.com> writes:
>> > IMO, what ought to happen here is that the RA should spill
>> > the inner register to memory and load the V4SI back from there.
>> > (Or vice versa, for an lvalue.) Obviously that's not very efficient,
>> > and so a patch like the above might be useful as an optimisation.[*]
>> > But it shouldn't be needed for correctness. The target-independent
>> > code should already have the information it needs to realise that
>> > it can't predict the register index at compile time (at least for SVE).
>>
>> Or actually, for that case:
>>
>> /* For pseudo registers, we want most of the same checks. Namely:
>>
>> Assume that the pseudo register will be allocated to hard registers
>> that can hold REGSIZE bytes each. If OSIZE is not a multiple of
>> REGSIZE,
>> the remainder must correspond to the lowpart of the containing hard
>> register. If BYTES_BIG_ENDIAN, the lowpart is at the highest offset,
>> otherwise it is at the lowest offset.
>>
>> Given that we've already checked the mode and offset alignment,
>> we only have to check subblock subregs here. */
>> if (maybe_lt (osize, regsize)
>> && ! (lra_in_progress && (FLOAT_MODE_P (imode) || FLOAT_MODE_P
>> (omode))))
>> {
>> /* It is invalid for the target to pick a register size for a mode
>> that isn't ordered wrt to the size of that mode. */
>> poly_uint64 block_size = ordered_min (isize, regsize);
>> unsigned int start_reg;
>> poly_uint64 offset_within_reg;
>> if (!can_div_trunc_p (offset, block_size, &start_reg,
>> &offset_within_reg)
>> ...
>>
>
> Like aarch64 we set REGMODE_NATURAL_SIZE for fixed-size modes to
> UNITS_PER_WORD. Isn't that part of the problem?
>
> In extract_bit_field_as_subreg we check lowpart_bit_field_p (= true because
> 128 is a multiple of UNITS_PER_WORD). This leads to the subreg expression.
>
> If I have REGMODE_NATURAL_SIZE return a VLA number this fails and we extract
> via memory - but that of course breaks almost everything else :)
>
> When you say the target-independent code should already have all information
> it
> needs, what are you referring to? Something else than REGMODE_NATURAL_SIZE?
In the aarch64 example I mentioned, the REGMODE_NATURAL_SIZE of the inner
mode is variable. (The REGMODE_NATURAL_SIZE of the outer mode is constant,
but it's the inner mode that matters here.)
Thanks,
Richard
