https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117116
--- Comment #7 from Uroš Bizjak <ubizjak at gmail dot com> ---
(In reply to Hongtao Liu from comment #4)
> (In reply to Hongtao Liu from comment #3)
> > A simple testcase
> >
> > typedef long long v4di __attribute__((vector_size(32)));
> >
> > v4di
> > foo (long long a)
> > {
> > return __extension__(v4di){(long long)foo, 1, 1, 1};
> > }
> >
> > reproduced with -O2 -mavx2, failed at least since gcc4.8.1
>
> 19843;; sse4_1_pinsrd must come before sse2_loadld since it is preferred.
>
> 19844(define_insn "<sse2p4_1>_pinsr<ssemodesuffix>"
[...]
> Doesn't support V2DI, we're using.
It's not a problem with the above pattern, which *does* support V2DI:
;; Modes handled by pinsr patterns.
(define_mode_iterator PINSR_MODE
[(V16QI "TARGET_SSE4_1") V8HI V8HF V8BF
(V4SI "TARGET_SSE4_1")
(V2DI "TARGET_SSE4_1 && TARGET_64BIT")])
but with operand 2 that doesn't satisfy DImode nonimmediate_operand predicate
in:
;; sse4_1_pinsrd must come before sse2_loadld since it is preferred.
(define_insn "<sse2p4_1>_pinsr<ssemodesuffix>"
[(set (match_operand:PINSR_MODE 0 "register_operand" "=x,x,x,x,v,v,&x")
(vec_merge:PINSR_MODE
(vec_duplicate:PINSR_MODE
(match_operand:<ssescalarmode> 2 "nonimmediate_operand"
"jr,jm,r,m,r,m,x"))
(match_operand:PINSR_MODE 1 "register_operand" "0,0,x,x,v,v,x")
(match_operand:SI 3 "const_int_operand")))]
"TARGET_SSE2
&& ((unsigned) exact_log2 (INTVAL (operands[3]))
< GET_MODE_NUNITS (<MODE>mode))"
(insn 8 7 9 2 (set (reg:V2DI 103)
(vec_merge:V2DI (vec_duplicate:V2DI (symbol_ref:DI ("foo") [flags 0x3]
<function_decl 0x7f412dfec200 foo>))
(reg:V2DI 103)
(const_int 1 [0x1]))) "pr117116.c":6:12 -1
(nil))
Please note that:
(symbol_ref:DI ("foo") [flags 0x3] <function_decl 0x7f412dfec200 foo>)
does *not* represent a memory operand. A mem(...) is missing around.
So, the problem is, where immediate operand leaks into RTX.
