Speed up (+ 2 2) by a factor of 10

Improve arithmetic performance by avoiding bignums until needed.
Also, simplify bignum memory management, fixing some unlikely leaks.
This patch improved the performance of (+ 2 2) by a factor of ten
on a simple microbenchmark computing (+ x 2), byte-compiled,
with x a local variable initialized to 2 via means the byte
compiler could not predict: performance improved from 135 to 13 ns.
The platform was Fedora 28 x86-64, AMD Phenom II X4 910e.
Performance also improved 0.6% on ‘make compile-always’.
* src/bignum.c (init_bignum_once): New function.
* src/emacs.c (main): Use it.
* src/bignum.c (mpz): New global var.
(make_integer_mpz): Rename from make_integer.  All uses changed.
* src/bignum.c (double_to_bignum, make_bignum_bits)
(make_bignum, make_bigint, make_biguint, make_integer_mpz):
* src/data.c (bignum_arith_driver, Frem, Flogcount, Fash)
(expt_integer, Fadd1, Fsub1, Flognot):
* src/floatfns.c (Fabs, rounding_driver, rounddiv_q):
* src/fns.c (Fnthcdr):
Use mpz rather than mpz_initting and mpz_clearing private
temporaries.
* src/bignum.h (bignum_integer): New function.
* src/data.c (Frem, Fmod, Fash, expt_integer):
* src/floatfns.c (rounding_driver):
Use it to simplify code.
* src/data.c (FIXNUMS_FIT_IN_LONG, free_mpz_value):
Remove.  All uses removed.
(floating_point_op): New function.
(floatop_arith_driver): New function, with much of the guts
of the old float_arith_driver.
(float_arith_driver): Use it.
(floatop_arith_driver, arith_driver):
Simplify by assuming NARGS is at least 2.
All callers changed.
(float_arith_driver):
New arg, containing the partly converted value of the next arg.
Reorder args for consistency.  All uses changed.
(bignum_arith_driver): New function.
(arith_driver): Use it.  Do fixnum-only integer calculations
in intmax_t instead of mpz_t, when they fit.
Break out mpz_t calculations into bignum_arith_driver.
(Fquo): Use floatop_arith_driver instead of float_arith_driver,
since the op is known to be valid.
(Flogcount, Fash): Simplify by coalescing bignum and fixnum code.
(Fadd1, Fsub1): Simplify by using make_int.

1 files changed, 5 insertions, 7 deletions

diff --git a/src/fns.c b/src/fns.c
index 17a869e1abc..8b25492eaeb 100644
--- a/src/fns.c
+++ b/src/fns.c
@@ -1468,19 +1468,17 @@ DEFUN ("nthcdr", Fnthcdr, Snthcdr, 2, 2, 0,
       /* Undo any error introduced when LARGE_NUM was substituted for
          N, by adding N - LARGE_NUM to NUM, using arithmetic modulo
          CYCLE_LENGTH.  */
-      mpz_t z; /* N mod CYCLE_LENGTH.  */
-      mpz_init (z);
+      /* Add N mod CYCLE_LENGTH to NUM.  */
       if (cycle_length <= ULONG_MAX)
-        num += mpz_mod_ui (z, XBIGNUM (n)->value, cycle_length);
+        num += mpz_mod_ui (mpz[0], XBIGNUM (n)->value, cycle_length);
       else
         {
-          mpz_set_intmax (z, cycle_length);
-          mpz_mod (z, XBIGNUM (n)->value, z);
+          mpz_set_intmax (mpz[0], cycle_length);
+          mpz_mod (mpz[0], XBIGNUM (n)->value, mpz[0]);
           intptr_t iz;
-          mpz_export (&iz, NULL, -1, sizeof iz, 0, 0, z);
+          mpz_export (&iz, NULL, -1, sizeof iz, 0, 0, mpz[0]);
           num += iz;
         }
-      mpz_clear (z);
       num += cycle_length - large_num % cycle_length;
     }
   num %= cycle_length;