target/4860: [ARM; 3.0] sjlj_mark_call_sites thwarted by load-multi

pb@gcc.gnu.org pb@gcc.gnu.org
Tue Jan 22 15:07:00 GMT 2002


Old Synopsis: Internal Compiler Error with ARM g++ cross compiler
New Synopsis: [ARM; 3.0] sjlj_mark_call_sites thwarted by load-multi

Responsible-Changed-From-To: unassigned->pb
Responsible-Changed-By: pb
Responsible-Changed-When: Tue Jan 22 15:07:15 2002
Responsible-Changed-Why:
    .
State-Changed-From-To: open->analyzed
State-Changed-By: pb
State-Changed-When: Tue Jan 22 15:07:15 2002
State-Changed-Why:
    Still occurs with 3.0 branch.
    
    This loop in sjlj_mark_call_sites (about line 2135) is too simple minded:
    
    	  /* Search backward for the first set of a register in this set.  */
    	  while (nparm_regs)
    	    {
    	      before = PREV_INSN (before);
    
    	      /* Given that we've done no other optimizations yet,
    		 the arguments should be immediately available.  */
    	      if (GET_CODE (before) == CODE_LABEL)
    		abort ();
    
    	      p = single_set (before);
    	      if (p && GET_CODE (SET_DEST (p)) == REG
    		  && REGNO (SET_DEST (p)) < FIRST_PSEUDO_REGISTER
    		  && TEST_HARD_REG_BIT (parm_regs, REGNO (SET_DEST (p))))
    		{
    		  CLEAR_HARD_REG_BIT (parm_regs, REGNO (SET_DEST (p)));
    		  nparm_regs--;
    		}
    	    }
    
    What happens in this case is that several of the SETs for argument registers have been wrapped up in a PARALLEL in order to make them come out as an LDM, which stops single_set from seeing them.
    
    This code has already been restructured on the trunk to use note_stores, which avoids the problem.  See:
    
    Sun Jul 29 23:26:50 CEST 2001  Jan Hubicka  <jh@suse.cz>
    
    	* rtlanal.c (parms_set, find_first_parameter_load): Break out from...;
    	handle multiple sets.
    	* except.c (sjlj_mark_call_sites): .... here.
    	* gcse.c (insert_insn_end_bb): Use find_first_parameter_load.

http://gcc.gnu.org/cgi-bin/gnatsweb.pl?cmd=view%20audit-trail&database=gcc&pr=4860



More information about the Gcc-prs mailing list