Fix function call regression in new evaluator

The internal AdaCore test suite revealed a bug in the new evaluator.
A hunk of evaluate_funcall was not correctly transcribed.  This was
not caught in my original testing because the feature in question was
apparently not tested in gdb.

This patch fixes the oversight.  The idea here is that ordinary
function calls should use the function's formal parameter types as the
expected types of subexpressions.

Regression tested on x86-64 Fedora 32.

gdb/ChangeLog
2021-03-09  Tom Tromey  <tromey@adacore.com>

	* eval.c (operation::evaluate_funcall): Use function formal
	parameter types when evaluating.

gdb/testsuite/ChangeLog
2021-03-09  Tom Tromey  <tromey@adacore.com>

	* gdb.base/cast-call.exp: New file.
	* gdb.base/cast-call.c: New file.

1 files changed, 9 insertions, 1 deletions

diff --git a/gdb/eval.c b/gdb/eval.c
index 5af728a..530ff15 100644
--- a/gdb/eval.c
+++ b/gdb/eval.c
@@ -687,8 +687,16 @@ operation::evaluate_funcall (struct type *expect_type,
   std::vector<value *> vals (args.size ());
 
   value *callee = evaluate_with_coercion (exp, noside);
+  struct type *type = value_type (callee);
+  if (type->code () == TYPE_CODE_PTR)
+    type = TYPE_TARGET_TYPE (type);
   for (int i = 0; i < args.size (); ++i)
-    vals[i] = args[i]->evaluate_with_coercion (exp, noside);
+    {
+      if (i < type->num_fields ())
+	vals[i] = args[i]->evaluate (type->field (i).type (), exp, noside);
+      else
+	vals[i] = args[i]->evaluate_with_coercion (exp, noside);
+    }
 
   return evaluate_subexp_do_call (exp, noside, callee, vals,
 				  function_name, expect_type);