[Bug target/110109] RISC-V: ICE when build the Intrinsic code

["juzhe.zhong at rivai dot ai" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110109

--- Comment #2 from JuzheZhong <juzhe.zhong at rivai dot ai> ---

(define_insn_and_split "*vlmul_extx2<mode>"
  [(set (match_operand:<VLMULX2> 0 "register_operand"  "=vr, ?&vr")
        (subreg:<VLMULX2>
          (match_operand:VLMULEXT2 1 "register_operand" " 0,   vr") 0))]
  "TARGET_VECTOR"
  "#"
  "&& reload_completed"
  [(const_int 0)]
{
  emit_insn (gen_rtx_SET (gen_lowpart (<MODE>mode, operands[0]), operands[1]));
  DONE;
})

(define_insn_and_split "*vlmul_extx4<mode>"
  [(set (match_operand:<VLMULX4> 0 "register_operand"  "=vr, ?&vr")
        (subreg:<VLMULX4>
          (match_operand:VLMULEXT4 1 "register_operand" " 0,   vr") 0))]
  "TARGET_VECTOR"
  "#"
  "&& reload_completed"
  [(const_int 0)]
{
  emit_insn (gen_rtx_SET (gen_lowpart (<MODE>mode, operands[0]), operands[1]));
  DONE;
})

(define_insn_and_split "*vlmul_extx8<mode>"
  [(set (match_operand:<VLMULX8> 0 "register_operand"  "=vr, ?&vr")
        (subreg:<VLMULX8>
          (match_operand:VLMULEXT8 1 "register_operand" " 0,   vr") 0))]
  "TARGET_VECTOR"
  "#"
  "&& reload_completed"
  [(const_int 0)]
{
  emit_insn (gen_rtx_SET (gen_lowpart (<MODE>mode, operands[0]), operands[1]));
  DONE;
})

(define_insn_and_split "*vlmul_extx16<mode>"
  [(set (match_operand:<VLMULX16> 0 "register_operand"  "=vr, ?&vr")
        (subreg:<VLMULX16>
          (match_operand:VLMULEXT16 1 "register_operand" " 0,   vr") 0))]
  "TARGET_VECTOR"
  "#"
  "&& reload_completed"
  [(const_int 0)]
{
  emit_insn (gen_rtx_SET (gen_lowpart (<MODE>mode, operands[0]), operands[1]));
  DONE;
})

(define_insn_and_split "*vlmul_extx32<mode>"
  [(set (match_operand:<VLMULX32> 0 "register_operand"  "=vr, ?&vr")
        (subreg:<VLMULX32>
          (match_operand:VLMULEXT32 1 "register_operand" " 0,   vr") 0))]
  "TARGET_VECTOR"
  "#"
  "&& reload_completed"
  [(const_int 0)]
{
  emit_insn (gen_rtx_SET (gen_lowpart (<MODE>mode, operands[0]), operands[1]));
  DONE;
})

(define_insn_and_split "*vlmul_extx64<mode>"
  [(set (match_operand:<VLMULX64> 0 "register_operand"  "=vr, ?&vr")
        (subreg:<VLMULX64>
          (match_operand:VLMULEXT64 1 "register_operand" " 0,   vr") 0))]
  "TARGET_VECTOR"
  "#"
  "&& reload_completed"
  [(const_int 0)]
{
  emit_insn (gen_rtx_SET (gen_lowpart (<MODE>mode, operands[0]), operands[1]));
  DONE;
})

I realize when I removed these patterns, issue is fixed.

I am not sure whether it is correct fix since these patterns are existing in
GCC13 which doesn't have such issue.

The reason I add this patterns since GCC can not well handle subreg, If I
simpily
use emit_insn (gen_rtx_SET (gen_lowpart (<MODE>mode, operands[0]),
operands[1])); 
in "expand" stage, it can not assign regno(operands[0]) == regno (operands[1])
in
RA, so I use constraint to force it.

These patterns are performance optimization patterns. Should I remove those
patterns and just use 
emit_insn (gen_rtx_SET (gen_lowpart (<MODE>mode, operands[0]), operands[1]));
in "expand" stage to fix this issue even though it demage the performance.