Cache array length in `rb_ary_join` (#15362)

When all elements are strings, we never have to recalculate the length
of the array because there are no conversion methods that are called, so
the length will never change. This speeds up the fast path by ~10%.

```ruby
a = ["1"*10, "2"*10, "3"*10, "4"*10, "5"*10] * 10
10_000_000.times do
  a.join
end
```

```
hyperfine --warmup 1 'ruby ../ruby2/test.rb' './exe/ruby ../ruby2/test.rb'
Benchmark 1: ruby ../ruby2/test.rb
  Time (mean ± σ):      3.779 s ±  0.053 s    [User: 3.754 s, System: 0.017 s]
  Range (min … max):    3.715 s …  3.874 s    10 runs

Benchmark 2: ./exe/ruby ../ruby2/test.rb
  Time (mean ± σ):      3.411 s ±  0.038 s    [User: 3.387 s, System: 0.017 s]
  Range (min … max):    3.360 s …  3.472 s    10 runs

Summary
  ./exe/ruby ../ruby2/test.rb ran
    1.11 ± 0.02 times faster than ruby ../ruby2/test.rb
```

1 files changed, 20 insertions, 15 deletions

diff --git a/array.c b/array.c
index 2acaafee03..a6aeeeeca1 100644
--- a/array.c
+++ b/array.c
@@ -2917,23 +2917,28 @@ rb_ary_join(VALUE ary, VALUE sep)
         StringValue(sep);
         len += RSTRING_LEN(sep) * (RARRAY_LEN(ary) - 1);
     }
-    for (i=0; i<RARRAY_LEN(ary); i++) {
+    long len_memo = RARRAY_LEN(ary);
+    for (i=0; i < len_memo; i++) {
         val = RARRAY_AREF(ary, i);
-        tmp = rb_check_string_type(val);
-
-        if (NIL_P(tmp) || tmp != val) {
-            int first;
-            long n = RARRAY_LEN(ary);
-            if (i > n) i = n;
-            result = rb_str_buf_new(len + (n-i)*10);
-            rb_enc_associate(result, rb_usascii_encoding());
-            i = ary_join_0(ary, sep, i, result);
-            first = i == 0;
-            ary_join_1(ary, ary, sep, i, result, &first);
-            return result;
+        if (RB_UNLIKELY(!RB_TYPE_P(val, T_STRING))) {
+            tmp = rb_check_string_type(val);
+            if (NIL_P(tmp) || tmp != val) {
+                int first;
+                long n = RARRAY_LEN(ary);
+                if (i > n) i = n;
+                result = rb_str_buf_new(len + (n-i)*10);
+                rb_enc_associate(result, rb_usascii_encoding());
+                i = ary_join_0(ary, sep, i, result);
+                first = i == 0;
+                ary_join_1(ary, ary, sep, i, result, &first);
+                return result;
+            }
+            len += RSTRING_LEN(tmp);
+            len_memo = RARRAY_LEN(ary);
+        }
+        else {
+            len += RSTRING_LEN(val);
         }
-
-        len += RSTRING_LEN(tmp);
     }
 
     result = rb_str_new(0, len);