[Bug target/60028] New: TIC6X: B3 register (return address) is saved on stack when real call is replaced with sibling call in a leaf function

[wojtek.golf at interia dot pl]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60028

            Bug ID: 60028
           Summary: TIC6X: B3 register (return address) is saved on stack
                    when real call is replaced with sibling call in a leaf
                    function
           Product: gcc
           Version: 4.8.2
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: wojtek.golf at interia dot pl

GCC 4.8.2, target tic6x
Input source:

int e_fun(int arg);
extern int e_int;

void g_fun(void)
{
    e_fun(e_int);
}

compiled with a command: -fleading-underscore -march=c674x -O2 -g0 -S main.c

gives the following assembler code:
_g_fun:
        sub .d2 B15, 8, B15
        stw .d2t2   B3, *+B15(8)
        b   .s1 (_e_fun)
    ||  ldw .d2t1   *+B14(_e_int), A4
        ldw .d2t2   *+B15(8), B3
    ||  add .s2 8, B15, B15
    ;; sibcall to (_e_fun) occurs
        nop 4

where B3 is saved on the stack even though a branch is taken to _e_fun. Texas
Instruments CGTools compiler v7.3.2 (cl6x --gcc -mv6740 --symdebug:none -k -n
main.c) gives the following assembler code (CALLRET is an alias to B):

_g_fun:
           CALLRET .S1     _e_fun            ; |6|
           LDW     .D2T1   *+DP(_e_int),A4   ; |6|
           NOP             4
$C$RL0:    ; CALL-RETURN OCCURS {_e_fun} 0   ; |6|

where B3 is not saved.

Some investigation on my part leads to the function c6x_expand_call in the file
gcc/gcc/config/c6x/c6x.c, where we see what may cause B3 being marked as used
and hence saved on the stack prior to the branch:

  if (sibcall)
    {
      call_insn = emit_call_insn (call_insn);
      use_reg (&CALL_INSN_FUNCTION_USAGE (call_insn),
         gen_rtx_REG (Pmode, REG_B3));
    }

commenting out use_reg invocation makes gcc to generate the expected listing:

_g_fun:
        b   .s1 (_e_fun)
    ||  ldw .d2t1   *+B14(_e_int), A4
    ;; sibcall to (_e_fun) occurs
        nop 5

I haven't tested yet whether removal of use_reg invocation causes any
regression, but I'd like it to be investigated whether it should be removed
altogether or protected with some extra execution condition.

With best,

Wojciech