https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88278
--- Comment #6 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
I wonder about something like:
--- gcc/config/i386/sse.md.jj 2018-11-29 23:16:06.481301632 +0100
+++ gcc/config/i386/sse.md 2018-11-30 16:21:21.480379008 +0100
@@ -7248,6 +7248,17 @@
(set_attr "prefix" "orig,maybe_evex,orig,maybe_evex")
(set_attr "mode" "V4SF,V4SF,V2SF,V2SF")])
+(define_insn "*vec_concatv4sf_0"
+ [(set (match_operand:V4SF 0 "register_operand" "=v")
+ (vec_concat:V4SF
+ (match_operand:V2SF 1 "nonimmediate_operand" "xm")
+ (match_operand:V2SF 2 "const0_operand" " C")))]
+ "TARGET_SSE2"
+ "%vmovq\t{%1, %0|%0, %1}"
+ [(set_attr "type" "ssemov")
+ (set_attr "prefix" "maybe_vex")
+ (set_attr "mode" "DF")])
+
;; Avoid combining registers from different units in a single alternative,
;; see comment above inline_secondary_memory_needed function in i386.c
(define_insn "vec_set<mode>_0"
@@ -14409,6 +14420,23 @@
(set_attr "prefix" "orig,maybe_evex,orig,orig,maybe_evex")
(set_attr "mode" "TI,TI,V4SF,V2SF,V2SF")])
+(define_insn "*vec_concatv4si_0"
+ [(set (match_operand:V4SI 0 "register_operand" "=v,x")
+ (vec_concat:V4SI
+ (match_operand:V2SI 1 "nonimmediate_operand" "vm,?!*y")
+ (match_operand:V2SI 2 "const0_operand" " C,C")))]
+ "TARGET_SSE2"
+ "@
+ %vmovq\t{%1, %0|%0, %1}
+ movq2dq\t{%1, %0|%0, %1}"
+ [(set_attr "type" "ssemov")
+ (set_attr "prefix" "maybe_vex,orig")
+ (set_attr "mode" "TI")
+ (set (attr "preferred_for_speed")
+ (if_then_else (eq_attr "alternative" "1")
+ (symbol_ref "TARGET_INTER_UNIT_MOVES_FROM_VEC")
+ (symbol_ref "true")))])
+
;; movd instead of movq is required to handle broken assemblers.
(define_insn "vec_concatv2di"
[(set (match_operand:V2DI 0 "register_operand"
but the #c0 testcases don't compile for me with -O2 -msse2 -fgimple (nor
-mavx), so I can't easily verify.
I don't see how we could get rid of those for the v <- v,C cases, unless we
analyze whatever instruction generated it and prove that it leaves all the
higher bits set to zero. E.g. one could have a v4si to v2si downcast (just
picking the lowpart subreg) followed by concatenating it with zero, and if we
blindly drop the movq instruction, the upper bits might be non-zero.
