Hi!
AVX2 prefers v*i128 instructions for vector integer
modes, while AVX only supports v*f128.
In one of my recent patches I've changed vextract* so that
it emits v*i128 instead of v*f128 for AVX2 on integer modes,
but I wonder what is the prefered way to change the rest
of the insn.
E.g. we have:
(define_insn "avx_vbroadcastf128_<mode>"
[(set (match_operand:V_256 0 "register_operand" "=x,x,x")
(vec_concat:V_256
(match_operand:<ssehalfvecmode> 1 "nonimmediate_operand" "m,0,?x")
(match_dup 1)))]
"TARGET_AVX"
"@
vbroadcastf128\t{%1, %0|%0, %1}
vinsertf128\t{$1, %1, %0, %0|%0, %0, %1, 1}
vperm2f128\t{$0, %t1, %t1, %0|%0, %t1, %t1, 0}"
[(set_attr "type" "ssemov,sselog1,sselog1")
(set_attr "prefix_extra" "1")
(set_attr "length_immediate" "0,1,1")
(set_attr "prefix" "vex")
(set_attr "mode" "V4SF,V8SF,V8SF")])
One option is change the mode iterator, so that it
iterates on all 256-bit vectors for !TARGET_AVX2, and only
for V8SF and V4DF for TARGET_AVX2, then for each such an insn also
add similar insn that iterates on the 256-bit integer vectors,
is guarded by TARGET_AVX2 and uses *i128 and OI mode attribute instead of
V8SF or V4DF.
Or, e.g. we could have operand modifier that would print "i128"
resp. "i" for TARGET_AVX2 and integer vector modes, and would print
"f128" resp. "f" otherwise. Still, something would need to be done
about the mode attribute.
Or something else?
Jakub
