Re: [gcjx] Tree-SSA Operands Processing Problem

[Zdenek Dvorak <rakdver at atrey dot karlin dot mff dot cuni dot cz>]

Hello,

>   While trying to debug a long-standing ICE on the
> gcjx branch, I have found out that the root cause
> of the problem is with the tree-ssa operands processing
> code and it still seems to exist on the mainline
> (as of revision 109471), though I have traced the code
> path in a debugger and actually observed the problem only
> on the gcjx branch.
> 
> In mainline, the code path starts from
> tree-ssa-operands.c:1906 where we have a *local* variable
> VAR whose address is passed on to add_stmt_operand() on
> the next line:
> - ----------------------- 8< -----------------------
>    1903   /* Add a VUSE for each call-clobbered variable.  */
>    1904   EXECUTE_IF_SET_IN_BITMAP (call_clobbered_vars, 0, u, bi)
>    1905     {
>    1906       tree var = referenced_var (u);
>    1907       add_stmt_operand (&var, &empty_ann, opf_none |
> opf_non_specific);
>    1908     }
>    1909
> - ----------------------- 8< -----------------------
> 
> This function then passes this pointer to append_use() at
> line 1634, which seems to just store it as-is in BUILD_USES:
> - ----------------------- 8< -----------------------
>    1572 static void
>    1573 add_stmt_operand (tree *var_p, stmt_ann_t s_ann, int flags)
>    1574 {
>    [...]
>    1628   if (is_real_op)
>    1629     {
>    1630       /* The variable is a GIMPLE register.  Add it to real
> operands.  *        /
>    1631       if (flags & opf_is_def)
>    1632         append_def (var_p);
>    1633       else
>    1634         append_use (var_p);
>    1635     }
> - ----------------------- 8< -----------------------
> 
> This seems wrong.

this can never happen.  Note that is_real_op = is_gimple_reg (var), and
a call clobbered variable can never satisfy is_gimple_reg.  Most likely
you forget to set TREE_ADDRESSABLE for this variable.  Or it gets to
call_clobbered_vars wrongly.

Zdenek

> I don't know whether this code path is actually ever
> taken on mainline or why it hasn't caused any problems
> yet, but on the gcjx branch it causes an ICE while
> processing the method "of" of the inner class "UnicodeBlock"
> of the class "java.lang.Character"
> ($GCC_SRC_DIR/libjava/java/lang/Character.java). Specifically,
> it ICEs while processing the CALL_EXPR corresponding to
> the call to _Jv_ThrowBadArrayIndex() thrown if a subscript
> is beyond the limits of an array. The problem occurs when
> the stack-slot for the VAR variable referred to earlier
> is occupied by the NUM local variable of finalize_ssa_vuses()
> and is overwritten when this variable is assigned a value in
> this function.
> 
> Comments?
> 
> Thanks,
> Ranjit.
> 
> - --
> Ranjit Mathew       Email: rmathew AT gmail DOT com
> 
> Bangalore, INDIA.     Web: http://ranjitmathew.hostingzero.com/
> 
> 
> 
> 
> -----BEGIN PGP SIGNATURE-----
> Version: GnuPG v1.4.2 (GNU/Linux)
> Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org
> 
> iD8DBQFDwT2fYb1hx2wRS48RAr8nAJ90iq5wgQpVeG09vihfwN1x1gvsRACfebPH
> 3ZjNpJLVJNJZTbWpDEKyxiU=
> =NBsE
> -----END PGP SIGNATURE-----
>