[C++ PATCH] Avoid NOP_EXPRs with error_mark_node operand (PR c++/83634)

[Jakub Jelinek <jakub at redhat dot com>]

Hi!

The gimplifier uses in several places STRIP_USELESS_TYPE_CONVERSION and
that, being primarily a middle-end predicate, doesn't like error_mark_nodes
appearing in conversion operands.  I've talked about it with Richi on IRC
and he'd prefer not to change it.

The reason we end up with them is because cp_fold when it sees a NOP_EXPR
conversion to void, it does nothing, and when cp_fold_r is used, it folds
the operand to in this case error_mark_node and nothing adjusts the
conversion.

As cp_fold has a cache, the following shouldn't really be expensive and will
just handle the case of error_mark_node right and the rest as before.
A variant (which I have a patch for too) would be to cp_fold_rvalue the
operand and if it folded to something different and is not error_mark_node
build a new NOP_EXPR around it.  What I don't like about that too much is
that for a very common case (aren't most statements converted to void) we
would probably create many more trees, as cp_fold generally doesn't modify stuff
in place, while cp_fold_r does.  By doing what the following patch does we
won't create the extra NOP_EXPRs in all cases where the operand folds
somehow.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

2018-01-02  Jakub Jelinek  <jakub@redhat.com>

	PR c++/83634
	* cp-gimplify.c (cp_fold) <case NOP_EXPR>: If the operand folds to
	error_mark_node, return error_mark_node.

	* g++.dg/parse/pr83634.C: New test.

--- gcc/cp/cp-gimplify.c.jj	2017-12-14 21:11:40.616124411 +0100
+++ gcc/cp/cp-gimplify.c	2018-01-02 13:23:51.946128057 +0100
@@ -2112,7 +2112,20 @@ cp_fold (tree x)
     case NON_LVALUE_EXPR:
 
       if (VOID_TYPE_P (TREE_TYPE (x)))
-	return x;
+	{
+	  /* This is just to make sure we don't end up with casts to
+	     void from error_mark_node.  If we just return x, then
+	     cp_fold_r might fold the operand into error_mark_node and
+	     leave the conversion in the IR.  STRIP_USELESS_TYPE_CONVERSION
+	     during gimplification doesn't like such casts.
+	     Don't create a new tree if op0 != TREE_OPERAND (x, 0), the
+	     folding of the operand should be in the caches and if in cp_fold_r
+	     it will modify it in place.  */
+	  op0 = cp_fold (TREE_OPERAND (x, 0));
+	  if (op0 == error_mark_node)
+	    x = error_mark_node;
+	  break;
+	}
 
       loc = EXPR_LOCATION (x);
       op0 = cp_fold_maybe_rvalue (TREE_OPERAND (x, 0), rval_ops);
--- gcc/testsuite/g++.dg/parse/pr83634.C.jj	2018-01-02 12:39:44.500517833 +0100
+++ gcc/testsuite/g++.dg/parse/pr83634.C	2018-01-02 12:35:17.399462088 +0100
@@ -0,0 +1,11 @@
+// PR c++/83634
+// { dg-do compile }
+
+void
+foo ()
+{
+  const int x = fn ();		// { dg-error "was not declared in this scope" }
+  short n;
+  for (n = x; n < 100; ++n)
+    ;
+}

	Jakub