https://gcc.gnu.org/bugzilla/show_bug.cgi?id=125915
Bug ID: 125915
Summary: [amdgcn] infinite recursion generating DImode load
from flat_scratch address space
Product: gcc
Version: 16.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: arsen at gcc dot gnu.org
Target Milestone: ---
the following testcase:
__UINT64_TYPE__
foo (__UINT64_TYPE__ __flat_scratch* i)
{ return *i; }
... never finishes compiling; it runs out of stack during expand, as it
recurses forever in gen_movdi:
if (!lra_in_progress && !reload_completed
&& !gcn_valid_move_p (<MODE>mode, operands[0], operands[1]))
{
/* Something is probably trying to generate a move
which can only work indirectly.
E.g. Move from LDS memory to SGPR hardreg
or MEM:QI to SGPR. */
rtx tmpreg = gen_reg_rtx (<MODE>mode);
emit_insn (gen_mov<mode> (tmpreg, operands[1]));
emit_insn (gen_mov<mode> (operands[0], tmpreg));
DONE;
}
printing the operands from the body of the 'if', we get:
=========
(reg:DI 683 [ _3 ])
(mem:DI (reg/f:DI 686) [2 *i_2(D)+0 S8 A64 AS3])
=========
(reg:DI 687)
(mem:DI (reg/f:DI 686) [2 *i_2(D)+0 S8 A64 AS3])
... and so on.
