[PATCH 1/2] [PATCH 1/2] Enable Intel AVX512_FP16 instructions

Jan Beulich jbeulich@suse.com
Fri Jul 2 15:08:57 GMT 2021


On 01.07.2021 09:47, Cui,Lili wrote:
>  opcodes/i386-dis-evex-mod.h    |  10 +
>  opcodes/i386-dis-evex-prefix.h | 212 ++++++++++++
>  opcodes/i386-dis-evex-w.h      | 442 +++++++++++++++++++++++-

Some of the vcvt* entries have two identical table entries. This
suggests that splitting decode from the W bit is not needed (and
this really is a general pattern: If all table entries at a level
end up identical, then this decode step can be omitted unless
there's a side effect of that step); see the pre-existing entries,
where this gets handled (iirc) in the processing of Gdq. Or are
you suggesting that there's a bug in the handling of pre-existing
encodings?

>  opcodes/i386-dis-evex.h        | 600 ++++++++++++++++++++++++++++++++-
>  opcodes/i386-dis.c             | 263 ++++++++++++++-

xmmqh_mode seems to rather parallel evex_half_bcst_xmmq_mode,
not xmmq_mode. Perhaps it would then also better be named
similarly? Even if there might not be a pre-existing similar
entry there, an analogous concern then applies to xmmqdh_mode
wrt its name.

Is w_scalar_swap_mode really (as the comment says) like
b_mode, not w_mode?

While in pre-existing enumerators like PREFIX_EVEX_0F38AB I'm
fine with the sequence of numbers (they express a sort-of
sequence of opcode bytes, after all), I think names like
PREFIX_EVEX_MAP510 really want an underscore inserted:
PREFIX_EVEX_MAP5_10.

Jan



More information about the Binutils mailing list