https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117657
Tobias Burnus <burnus at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Target Milestone|--- |15.0
--- Comment #5 from Tobias Burnus <burnus at gcc dot gnu.org> ---
We have:
------------------
(gdb) p debug_rtx(insn)
(insn 90 89 91 15 (set (reg:V32QI 696 [ vect_patt_45.10 ])
(vec_merge:V32QI (unspec:V32QI [
(plus:V32DI (reg:V32DI 752)
(vec_duplicate:V32DI (const_int 0 [0])))
(const_int 0 [0])
(unspec:V32QI [
(const_int 0 [0])
] UNSPEC_VECTOR)
(mem:BLK (scratch) [0 A8])
] UNSPEC_GATHER)
(reg:V32QI 696 [ vect_patt_45.10 ])
(reg:DI 750))) -1
(nil))
------------------------------------
This fails for
gcn/gcn-valu.md:1137
; Allow any address expression
(define_expand "gather<mode>_expr<exec>"
[(set (match_operand:V_MOV 0 "register_operand")
(unspec:V_MOV
[(match_operand 1 "")
(match_operand 2 "immediate_operand")
(match_operand 3 "immediate_operand")
(mem:BLK (scratch))]
UNSPEC_GATHER))]
""
{})
The generated code looks like:
switch (XVECLEN (x4, 0))
{
case 4:
...
x11 = XVECEXP (x4, 0, 2);
if (!immediate_operand (x11, E_VOIDmode))
return -1;
which is
(gdb) p debug_rtx(x11)
(unspec:V32QI [
(const_int 0 [0])
] UNSPEC_VECTOR)
...
Breakpoint 2, immediate_operand (op=0x7ffff7408490, mode=E_VOIDmode) at
gcc/recog.cc:1698
1698 {
(gdb) p debug_rtx(op)
(const_int 0 [0])
which calls in turn gcn_legitimate_constant_p (machine_mode, rtx x)
which calls
1045 gcn_constant_p (rtx x)
for E_V32QImode. But GET_CODE(x) == 'unspec', which is not handled by
gcn_constant_p - which returns false.
Thus, at the end -1 is the result of the insn - causing the failure.
