[Bug ipa/109607] IPA replaces stmt with invalid gimple

["rguenth at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109607

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #1)
> #0  ipa_param_body_adjustments::modify_expression (this=0x4b1f040, 
>     expr_p=0x7ffff37222b8, convert=true)
>     at /space/rguenther/src/gcc/gcc/ipa-param-manipulation.cc:1867
> #1  0x00000000016f7dc5 in ipa_param_body_adjustments::modify_assignment (
>     this=0x4b1f040, stmt=<gimple_assign 0x7ffff33ecdc0>, 
>     extra_stmts=0x7fffffffd0a8)
>     at /space/rguenther/src/gcc/gcc/ipa-param-manipulation.cc:1890
> #2  0x00000000016f927a in ipa_param_body_adjustments::modify_gimple_stmt (
>     this=0x4b1f040, stmt=0x7fffffffd168, extra_stmts=0x7fffffffd0a8, 
>     orig_stmt=<gimple_assign 0x7ffff37fa320>)
>     at /space/rguenther/src/gcc/gcc/ipa-param-manipulation.cc:2273
> #3  0x0000000001abb925 in remap_gimple_stmt (
>     stmt=<gimple_assign 0x7ffff37fa320>, id=0x7fffffffd730)
>     at /space/rguenther/src/gcc/gcc/tree-inline.cc:1961
> 
> Supposedly the easiest would be to make any is_gimple_reg_type IPA SRA
> replacement (I suppose all are ...) either address-taken or
> DECL_NOT_GIMPLE_REG_P.
> 
> diff --git a/gcc/ipa-param-manipulation.cc b/gcc/ipa-param-manipulation.cc
> index 42488ee09c3..473d759f983 100644
> --- a/gcc/ipa-param-manipulation.cc
> +++ b/gcc/ipa-param-manipulation.cc
> @@ -1384,6 +1384,8 @@ ipa_param_body_adjustments::common_initialization
> (tree old_fndecl,
>           DECL_CONTEXT (new_parm) = m_fndecl;
>           TREE_USED (new_parm) = 1;
>           DECL_IGNORED_P (new_parm) = 1;
> +         if (is_gimple_reg_type (new_type))
> +           DECL_NOT_GIMPLE_REG_P (new_parm) = 1;
>           layout_decl (new_parm, 0);
>           m_new_decls.quick_push (new_parm);
>  
> 
> seems to work on the testcase I have.

But it doesn't work during bootstrap, in insn-recog.cc we then hit

insn-recog.cc: In function 'int pattern511(rtx)':
insn-recog.cc:23146:1: error: invalid argument to gimple call
23146 | pattern511 (rtx x1)
      | ^~~~~~~~~~
ISRA.64882
# VUSE <.MEM_19(D)>
_9 = pattern510.isra (ISRA.64882);
during GIMPLE pass: fixup_cfg
insn-recog.cc:23146:1: internal compiler error: verify_gimple failed
0x9a3bf4 _start
        ../sysdeps/x86_64/start.S:115
Please submit a full bug report, with preprocessed source (by using
-freport-bug).
Please include the complete backtrace with any bug report.
See <https://gcc.gnu.org/bugs/> for instructions.
make[3]: *** [Makefile:1157: insn-recog.o] Error 1

the way modify_expression is done is ugly, even more so the special-casing
of BIT_FIELD_REF and {IMAG,REAL}PART_EXPR.

I'm not sure if we guarantee conversions do not occur in some places?  At
least the call argument handling looks wrong to me and can result in invalid
GIMPLE as well.

While I can probably hack around the original assignment that's mishandled
it will feel like a hack.  Something like

@@ -1827,7 +1825,8 @@
ipa_param_body_adjustments::replace_removed_params_ssa_names (tree old_name,
    necessary conversions.  */

 bool
-ipa_param_body_adjustments::modify_expression (tree *expr_p, bool convert)
+ipa_param_body_adjustments::modify_expression (tree *expr_p, bool convert,
+                                              gimple_seq *extra_stmts)
 {
   tree expr = *expr_p;

@@ -1862,6 +1861,11 @@ ipa_param_body_adjustments::modify_expression (tree
*expr_p, bool convert)
       gcc_checking_assert (tree_to_shwi (TYPE_SIZE (TREE_TYPE (expr)))
                           == tree_to_shwi (TYPE_SIZE (TREE_TYPE (repl))));
       tree vce = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (expr), repl);
+      if (is_gimple_reg (repl))
+       {
+         gcc_assert (extra_stmts);
+         vce = force_gimple_operand (vce, extra_stmts, true, NULL_TREE);
+       }
       *expr_p = vce;
     }
   else
@@ -1889,7 +1893,7 @@ ipa_param_body_adjustments::modify_assignment (gimple
*stmt,
   lhs_p = gimple_assign_lhs_ptr (stmt);

   any = modify_expression (lhs_p, false);
-  any |= modify_expression (rhs_p, false);
+  any |= modify_expression (rhs_p, false, extra_stmts);
   if (any
       && !useless_type_conversion_p (TREE_TYPE (*lhs_p), TREE_TYPE (*rhs_p)))
     {

I'm going to test this nevertheless ...