Naming of some of the builtins isn't fully consistent with that of pre-
existing ones, so there's a need for a new BR2() wrapper macro.
With the tests providing some proof of proper functioning of the
emulator code also enable use of the feature by guests, as there's no
other infrastructure involved in enabling this ISA extension.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Andrew Cooper <andrew.cooper3@citrix.com>
Acked-by: Henry Wang <Henry.Wang@arm.com> # CHANGELOG
For one it was an oversight to leave dup_{hi,lo}() undefined for 512-bit
vector size. And then in FMA testing we can also arrange for the
compiler to (hopefully) recognize broadcasting potential. Plus we can
replace the broadcast(1) use in the addsub() surrogate with inline
assembly explicitly using embedded broadcast (even gcc12 still doesn't
support broadcast for any of the addsub/subadd builtins).
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Andrew Cooper <andrew.cooper3@citrix.com>
Despite their suffixes these aren't scalar instructions, and hence the
128- and 256-bit EVEX forms may not be used without AVX512VL. Gcc11 ends
up generating such instances for simd-sg.c.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Andrew Cooper <andrew.cooper3@citrix.com>
Also use this for AVX512_VBMI2 VPSH{L,R}D{,V}{D,Q,W} testing (only the
quad word right shifts get actually used; the assumption is that their
"left" counterparts as well as the double word and word forms then work
as well).
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Andrew Cooper <andrew.cooper3@citirx.com>
As to the feature dependency adjustment, while strictly speaking SSE is
a sufficient prereq (to have XMM registers), vectors of bytes and qwords
have got introduced only with SSE2. gcc, for example, uses a similar
connection in its respective intrinsics header.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Andrew Cooper <andrew.cooper3@citrix.com>
Also include the only other AVX512ER insn pair, VEXP2P{D,S}.
Note that despite the replacement of the SHA insns' table slots there's
no need to special case their decoding: Their insn-specific code already
sets op_bytes (as was required due to simd_other), and TwoOp is of no
relevance for legacy encoded SIMD insns.
The raising of #UD when EVEX.L'L is 3 for AVX512ER scalar insns is done
to be on the safe side. The SDM does not clarify behavior there, and
it's even more ambiguous here (without AVX512VL in the picture).
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Andrew Cooper <andrew.cooper3@citrix.com>
Plus their AVX512BW counterparts.
Take the opportunity and also eliminate a pair of open coded instances
of scalar_1op().
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Andrew Cooper <andrew.cooper3@citrix.com>
VCVT{,T}S{S,D}2SI use EVEX.W for their destination (register) rather
than their (possibly memory) source operand size and hence need a
"manual" override of disp8scale.
While the SDM claims that EVEX.L'L needs to be zero for the 32-bit forms
of VCVT{,U}SI2SD (exception type E10NF), observations on my test system
do not confirm this (and I've got informal confirmation that this is a
doc mistake). Nevertheless, to be on the safe side, force evex.lr to be
zero in this case though when constructing the stub.
Slightly adjust the scalar to_int() in the test harness, to increase the
chances of the operand ending up in memory.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Andrew Cooper <andrew.cooper3@citrix.com>
... including the two AVX512DQ forms which shared encodings, just with
EVEX.W set there.
VCVTDQ2PD, sharing its main opcode with others, needs a "manual"
override of disp8scale.
The simd_size changes for the twobyte_table[] entries are benign to
pre-existing code, but allow decode_disp8scale() to work as is here.
The at this point wrong placement of the 0xe6 case block is once again
in anticipation of further additions of case labels.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Andrew Cooper <andrew.cooper3@citrix.com>
VCVTPS2PD, sharing its main opcode with others, needs a "manual"
override of disp8scale.
The simd_size change for twobyte_table[0x5a] is benign to pre-existing
code, but allows decode_disp8scale() to work as is here.
Also correct the comment on an AVX counterpart.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Andrew Cooper <andrew.cooper3@citrix.com>
No further test harness additions - what is there is good enough for
these rather "regular" insns.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Andrew Cooper <andrew.cooper3@citrix.com>
Judging from insn prefixes, these are scalar insns, but their (memory)
operands are vector ones (with the exception of 128-bit VMOVDDUP). For
this some adjustments to disp8scale calculation code are needed.
No explicit test harness additions other than the overrides, as the
compiler already makes use of the insns.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Andrew Cooper <andrew.cooper3@citrix.com>
No explicit test harness additions other than the overrides, as the
compiler already makes use of the insns.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Andrew Cooper <andrew.cooper3@citrix.com>
Also include vshuff{32x4,64x2} as being very similar to vshufi{32x4,64x2}.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Andrew Cooper <andrew.cooper3@citirx.com>
Take the liberty and also correct the (public interface) name of the
AVX512_VBMI feature flag, on the assumption that no external consumer
has actually been using that flag so far. Furthermore make it have
AVX512BW instead of AVX512F as a prerequisite, for requiring full
64-bit mask registers (the upper 48 bits of which can't be accessed
other than through XSAVE/XRSTOR without AVX512BW support).
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Andrew Cooper <andrew.cooper3@citrix.com>
There's once again one extra twobyte_table[] entry which gets its Disp8
shift value set right away without getting support implemented just yet,
again to avoid needlessly splitting groups of entries.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Andrew Cooper <andrew.cooper3@citrix.com>
Note that the vpmov{,s,us}{d,q}w table entries in evex-disp8.c are
slightly different from what one would expect, due to them requiring
EVEX.W to be zero.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Andrew Cooper <andrew.cooper3@citrix.com>
Note that the testing in simd.c doesn't really follow the ISA extension
pattern - to fit the scheme, extensions from byte and word granular
vectors can (currently) sensibly only happen in the AVX512BW case (and
hence respective abstraction macros will be added there rather than
here).
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Andrew Cooper <andrew.cooper3@citrix.com>
Test the 128- and 256-bit variants of the insns which have been
implemented already.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Andrew Cooper <andrew.cooper3@citrix.com>
Jan Beulich