https://gcc.gnu.org/bugzilla/show_bug.cgi?id=89189
Andrew Pinski <pinskia at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Component|target |tree-optimization
Status|UNCONFIRMED |NEW
Ever confirmed|0 |1
Severity|normal |enhancement
Last reconfirmed| |2021-08-03
--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Confirmed.
At the gimple level we should be able to produce a SHUFFLE with the other
vector being full of zeros:
_1 = BIT_FIELD_REF <x_3(D), 16, 48>;
_2 = BIT_FIELD_REF <x_3(D), 16, 112>;
_4 = {_1, _2};
This is what clang produces:
vpshufb .LCPI0_0(%rip), %xmm0, %xmm0 # xmm0 =
xmm0[6,7,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
....
.LCPI1_0:
.byte 6 # 0x6
.byte 7 # 0x7
.byte 14 # 0xe
.byte 15 # 0xf
.byte 128 # 0x80
.byte 128 # 0x80
.byte 128 # 0x80
.byte 128 # 0x80
.byte 128 # 0x80
.byte 128 # 0x80
.byte 128 # 0x80
.byte 128 # 0x80
.byte 128 # 0x80
.byte 128 # 0x80
.byte 128 # 0x80
.byte 128 # 0x80
The problem is also seem on aarch64 too.
