Richard Biener <richard.guent...@gmail.com> writes:
> On Thu, Jul 11, 2019 at 10:03 AM Richard Sandiford
> <richard.sandif...@arm.com> wrote:
>>
>> This patch rewrites the way simplify_subreg handles constants.
>> It uses similar native_encode/native_decode routines to the
>> tree-level handling of VIEW_CONVERT_EXPR, meaning that we can
>> move between rtx constants and the target memory image of them.
>>
>> The main point of this patch is to support subregs of constant-length
>> vectors for VLA vectors, beyond the very simple cases that were already
>> handled. Many of the new tests failed before the patch for variable-
>> length vectors.
>>
>> The boolean side is tested more by the upcoming SVE ACLE work.
>>
>> Tested on aarch64-linux-gnu, aarch64_be-elf and x86_64-linux-gnu.
>> OK to install?
>
> Hmm. So is subreg [offset] defined in terms of memory order or in
> terms of register order?
Memory order, except for the special case that paradoxical subregs
always have an offset of zero (rather than the natural negative
value for big-endian).
> I wonder if you need to handle FLOAT_WORDS_BIG_ENDIAN,
> REG_WORDS_BIG_ENDIAN and whether BYTES/WORDS_BIG_ENDIAN have any
> meaning here at all?
>
> I'm always struggling with this when working on BIT_FIELD_REFs
> on GIMPLE [registers]...
Yeah. Yhe subreg_size_lsb stuff copes with BYTES/WORDS_BIG_ENDIAN.
I used that in the patch exactly because I didn't want to have to redo
the logic here :-)
Because the offset is (mostly) the memory offset, REG_WORDS_BIG_ENDIAN
doesn't affect this code. It instead affects how we divide up a
multi-register hard register. (See subreg_get_info.)
FLOAT_WORDS_BIG_ENDIAN is explicitly documented as not affecting subregs:
@cindex @code{FLOAT_WORDS_BIG_ENDIAN}, (lack of) effect on @code{subreg}
On a few targets, @code{FLOAT_WORDS_BIG_ENDIAN} disagrees with
@code{WORDS_BIG_ENDIAN}. However, most parts of the compiler treat
floating point values as if they had the same endianness as integer
values. This works because they handle them solely as a collection of
integer values, with no particular numerical value. Only real.c and
the runtime libraries care about @code{FLOAT_WORDS_BIG_ENDIAN}.
Thanks,
Richard
