Subreg spilled wrong mode in LRA

Subreg spilled wrong mode in LRA

[Cooper Qu]

Hi,


I ran into a troublesome problelm about LRA. The case is hard to 
construct, so I paste the RTL sequence first to ask for help.

-----------------------------------------------------------------------------------------------

The RTL sequence after IRA:

(insn 11 198 1633 20 (set (reg:DI 170 [ iftmp.176_92 ])
         (const_int 0 [0])) "":2099 182 {*movdi_64bit}
      (expr_list:REG_EQUAL (const_int 0 [0])
         (nil)))

(insn 1632 1633 1634 20 (set (reg:SI 973)
         (subreg:SI (reg:DI 170 [ iftmp.176_92 ]) 0)) "":2099 183 
{*movsi_internal}
      (expr_list:REG_DEAD (reg:DI 170 [ iftmp.176_92 ])
         (nil)))
(insn 1634 1632 1635 20 (set (reg:SI 974)
         (eq:SI (reg:DI 421)
             (const_int 0 [0]))) "":2099 271 {*seq_zero_disi}
      (expr_list:REG_DEAD (reg:DI 421)
         (nil)))
(insn 1635 1634 1636 20 (set (reg:SI 972)
         (if_then_else:SI (eq (reg/v:DI 173 [ i ])
                 (const_int 0 [0]))
             (reg:SI 974)
             (reg:SI 973))) "":2099 238 {*cmovzsi}
      (expr_list:REG_DEAD (reg:SI 974)
         (expr_list:REG_DEAD (reg:SI 973)
             (nil))))
(insn 1636 1635 212 20 (set (subreg:SI (reg:DI 170 [ iftmp.176_92 ]) 0)
         (reg:SI 972)) "":2099 183 {*movsi_internal}
      (expr_list:REG_DEAD (reg:SI 972)
         (nil)))

...

(jump_insn 1022 1021 1023 104 (set (pc)
         (if_then_else (eq (reg:DI 170 [ iftmp.176_92 ])
                 (const_int 0 [0]))
             (label_ref 1051)
             (pc))) "":1467 252 {*branch_zerodi}
      (expr_list:REG_DEAD (reg:DI 170 [ iftmp.176_92 ])
         (int_list:REG_BR_PROB 354334804 (nil)))

-----------------------------------------------------------------------------------------------

The RTL sequence after LRA RELOAD:

(insn 11 198 2025 20 (set (reg:DI 13 a3 [orig:170 iftmp.176_92 ] [170])
         (const_int 0 [0])) "":2099 182 {*movdi_64bit}
      (expr_list:REG_EQUAL (const_int 0 [0])
         (nil)))
(insn 1633 2025 1632 20 (set (reg:DI 15 a5 [421])
         (minus:DI (reg:DI 15 a5 [orig:125 _13 ] [125])
             (reg:DI 14 a4 [orig:254 pretmp_372 ] [254]))) "":2099 11 
{*subdi3}
      (nil))
(insn 1632 1633 1634 20 (set (reg:SI 13 a3 [973])
         (reg:SI 13 a3 [orig:170 iftmp.176_92 ] [170])) "":2099 183 
{*movsi_internal}
      (nil))
(insn 1634 1632 1635 20 (set (reg:SI 15 a5 [974])
         (eq:SI (reg:DI 15 a5 [421])
             (const_int 0 [0]))) "":2099 271 {*seq_zero_disi}
      (nil))
(insn 1635 1634 1636 20 (set (reg:SI 15 a5 [972])
         (if_then_else:SI (eq (reg/v:DI 19 s3 [orig:173 i ] [173])
                 (const_int 0 [0]))
             (reg:SI 15 a5 [974])
             (reg:SI 13 a3 [973]))) "":2099 238 {*cmovzsi}
      (nil))
(insn 1636 1635 212 20 (set (mem/c:SI (plus:DI (reg/f:DI 2 sp)
                 (const_int 24 [0x18])) [349 %sfp+-3496 S4 A64])
         (reg:SI 15 a5 [972])) "":2099 183 {*movsi_internal}
      (nil))

...

(insn 1926 1021 1022 104 (set (reg:DI 15 a5 [orig:170 iftmp.176_92 ] [170])
         (mem/c:DI (plus:DI (reg/f:DI 2 sp)
                 (const_int 24 [0x18])) [349 %sfp+-3496 S8 A64])) 
"":1467 182 {*movdi_64bit}
      (nil))
(jump_insn 1022 1926 1023 104 (set (pc)
         (if_then_else (eq (reg:DI 15 a5 [orig:170 iftmp.176_92 ] [170])
                 (const_int 0 [0]))
             (label_ref 1051)
             (pc))) "":1467 252 {*branch_zerodi}
      (int_list:REG_BR_PROB 354334804 (nil))

-----------------------------------------------------------------------------------------------

The REG 170 spilled to memory, but it emits a SImode store and a DImode 
load. I tried to debug and find some related code in lra.c, in 
add_regs_to_insn_regno_info function,   the following code is about to 
check SUBREG

         if (GET_CODE (x) == SUBREG)
         {
           mode = wider_subreg_mode (x);
           if (read_modify_subreg_p (x))
             subreg_p = true;
           x = SUBREG_REG (x);
           code = GET_CODE (x);
         }

If read_modify_subreg_p (x) is true, the subreg_p flag will be set. But 
if SUBREG's mode is narrower than register mode, the subreg_p flag will 
not be set just like the case I wrote.

If I let read_modify_subreg_p (x) return true, my case will turn right. 
But I'm not sure if it's appropriate, because it may decrease target 
performance.


Hope someone can help me.


Best Regards,

Cooper