https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92645
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |hjl.tools at gmail dot com,
| |uros at gcc dot gnu.org
--- Comment #12 from Richard Biener <rguenth at gcc dot gnu.org> ---
So it's probably better but not great yet. It would help tremendously to look
at PR92658, for this testcase we specifically need truncv16hiv16qi2 which
can be implemented with pshufb (maybe also better). I tried the following
(the upper half of the selector is just "garbage")
(define_expand "truncv16hiv16qi2"
[(set (subreg:V32QI (match_operand:V16QI 0 "register_operand") 0)
(vec_select:V32QI
(subreg:V32QI (match_operand:V16HI 1 "register_operand") 0)
(parallel [(const_int 0) (const_int 2)
(const_int 4) (const_int 6)
(const_int 8) (const_int 10)
(const_int 12) (const_int 14)
(const_int 16) (const_int 18)
(const_int 20) (const_int 22)
(const_int 24) (const_int 26)
(const_int 28) (const_int 30)
(const_int 0) (const_int 2)
(const_int 4) (const_int 6)
(const_int 8) (const_int 10)
(const_int 12) (const_int 14)
(const_int 16) (const_int 18)
(const_int 20) (const_int 22)
(const_int 24) (const_int 26)
(const_int 28) (const_int 30)
])))]
"TARGET_AVX2")
but that isn't recognized. Possibly because of the outer subreg, who
knows.
(define_insn "truncv16hiv16qi2"
[(set (match_operand:V16QI 0 "register_operand" "=x,v")
(truncate:V16QI
(match_operand:V16HI 1 "register_operand" "x,v")))]
"TARGET_AVX2"
"@
pshufb\t{%1, %0|%0, %1}
vpshufb\t{%1, %0|%0, %1}")
"works" but of course is wrong (somehow need the constant mask in a
register). The rest of the backend also doesn't know truncate of
vectors so representing as shuffles is probably better.
I also wonder how to macroize all this - probably via some
helpers in i386-expand.c I guess.
But I can also work with the pack/unpack tree codes for now.
