Re: target/5054: [ARM/Thumb] ICE with -mlong-calls

[Phil Blundell <pb@nexus.co.uk>]

The following reply was made to PR target/5054; it has been noted by GNATS.

From: Phil Blundell <pb@nexus.co.uk>
To: gcc-gnats@gcc.gnu.org
Cc:  
Subject: Re: target/5054: [ARM/Thumb] ICE with -mlong-calls
Date: 22 Jan 2002 10:50:09 +0000

 The problem here seems to be this Thumb pattern:
 
 (define_insn "*call_insn"
   [(call (mem:SI (match_operand:SI 0 "" "X"))
 	 (match_operand:SI 1 "" ""))
    (use (match_operand 2 "" ""))
    (clobber (reg:SI LR_REGNUM))]
   "TARGET_THUMB
    && operands[2] == const0_rtx && (GET_CODE (operands[0]) ==
 SYMBOL_REF)"
   "bl\\t%a0"
   [(set_attr "length" "4")
    (set_attr "type" "call")]
 )
 
 Note that it checks operands[2], which is the "call type" cookie.  This
 gets set to const1_rtx when -mlong-calls is in effect.  The expectation
 seems to be that the define_expand "call" will transform this into a
 register-based call, but this is inhibited within arm_is_longcall_p for
 functions defined elsewhere in the current compilation unit.
 
 I guess that the define_expand ought to force operands[2] to const0_rtx
 when it decides not to do a long call, something like this:
 
     if (GET_CODE (callee) != REG
       && INTVAL (operands[2]) == CALL_LONG)
       {
         if (arm_is_longcall_p (operands[0], INTVAL (operands[2]), 0))
           XEXP (operands[0], 0) = force_reg (Pmode, callee);
         else
           operands[2] = const0_rtx;
       }