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, 38 insertions, 33 deletions

diff --git a/src/bignum.c b/src/bignum.c
index b18ceccb59d..2ce7412d06c 100644
--- a/src/bignum.c
+++ b/src/bignum.c
@@ -25,6 +25,22 @@ along with GNU Emacs.  If not, see <https://www.gnu.org/licenses/>.  */
 
 #include <stdlib.h>
 
+/* mpz global temporaries.  Making them global saves the trouble of
+   properly using mpz_init and mpz_clear on temporaries even when
+   storage is exhausted.  Admittedly this is not ideal.  An mpz value
+   in a temporary is made permanent by mpz_swapping it with a bignum's
+   value.  Although typically at most two temporaries are needed,
+   rounding_driver and rounddiv_q need four altogther.  */
+
+mpz_t mpz[4];
+
+void
+init_bignum_once (void)
+{
+  for (int i = 0; i < ARRAYELTS (mpz); i++)
+    mpz_init (mpz[i]);
+}
+
 /* Return the value of the Lisp bignum N, as a double.  */
 double
 bignum_to_double (Lisp_Object n)
@@ -36,17 +52,14 @@ bignum_to_double (Lisp_Object n)
 Lisp_Object
 double_to_bignum (double d)
 {
-  mpz_t z;
-  mpz_init_set_d (z, d);
-  Lisp_Object result = make_integer (z);
-  mpz_clear (z);
-  return result;
+  mpz_set_d (mpz[0], d);
+  return make_integer_mpz ();
 }
 
-/* Return a Lisp integer equal to OP, which has BITS bits and which
-   must not be in fixnum range.  */
+/* Return a Lisp integer equal to mpz[0], which has BITS bits and which
+   must not be in fixnum range.  Set mpz[0] to a junk value.  */
 static Lisp_Object
-make_bignum_bits (mpz_t const op, size_t bits)
+make_bignum_bits (size_t bits)
 {
   /* The documentation says integer-width should be nonnegative, so
      a single comparison suffices even though 'bits' is unsigned.  */
@@ -55,18 +68,17 @@ make_bignum_bits (mpz_t const op, size_t bits)
 
   struct Lisp_Bignum *b = ALLOCATE_PSEUDOVECTOR (struct Lisp_Bignum, value,
                                                  PVEC_BIGNUM);
-  /* We could mpz_init + mpz_swap here, to avoid a copy, but the
-     resulting API seemed possibly confusing.  */
-  mpz_init_set (b->value, op);
-
+  mpz_init (b->value);
+  mpz_swap (b->value, mpz[0]);
   return make_lisp_ptr (b, Lisp_Vectorlike);
 }
 
-/* Return a Lisp integer equal to OP, which must not be in fixnum range.  */
+/* Return a Lisp integer equal to mpz[0], which must not be in fixnum range.
+   Set mpz[0] to a junk value.  */
 static Lisp_Object
-make_bignum (mpz_t const op)
+make_bignum (void)
 {
-  return make_bignum_bits (op, mpz_sizeinbase (op, 2));
+  return make_bignum_bits (mpz_sizeinbase (mpz[0], 2));
 }
 
 static void mpz_set_uintmax_slow (mpz_t, uintmax_t);
@@ -86,30 +98,23 @@ Lisp_Object
 make_bigint (intmax_t n)
 {
   eassert (FIXNUM_OVERFLOW_P (n));
-  mpz_t z;
-  mpz_init (z);
-  mpz_set_intmax (z, n);
-  Lisp_Object result = make_bignum (z);
-  mpz_clear (z);
-  return result;
+  mpz_set_intmax (mpz[0], n);
+  return make_bignum ();
 }
 Lisp_Object
 make_biguint (uintmax_t n)
 {
   eassert (FIXNUM_OVERFLOW_P (n));
-  mpz_t z;
-  mpz_init (z);
-  mpz_set_uintmax (z, n);
-  Lisp_Object result = make_bignum (z);
-  mpz_clear (z);
-  return result;
+  mpz_set_uintmax (mpz[0], n);
+  return make_bignum ();
 }
 
-/* Return a Lisp integer with value taken from OP.  */
+/* Return a Lisp integer with value taken from mpz[0].
+   Set mpz[0] to a junk value.  */
 Lisp_Object
-make_integer (mpz_t const op)
+make_integer_mpz (void)
 {
-  size_t bits = mpz_sizeinbase (op, 2);
+  size_t bits = mpz_sizeinbase (mpz[0], 2);
 
   if (bits <= FIXNUM_BITS)
     {
@@ -118,20 +123,20 @@ make_integer (mpz_t const op)
 
       do
         {
-          EMACS_INT limb = mpz_getlimbn (op, i++);
+          EMACS_INT limb = mpz_getlimbn (mpz[0], i++);
           v += limb << shift;
           shift += GMP_NUMB_BITS;
         }
       while (shift < bits);
 
-      if (mpz_sgn (op) < 0)
+      if (mpz_sgn (mpz[0]) < 0)
         v = -v;
 
       if (!FIXNUM_OVERFLOW_P (v))
         return make_fixnum (v);
     }
 
-  return make_bignum_bits (op, bits);
+  return make_bignum_bits (bits);
 }
 
 /* Set RESULT to V.  This code is for when intmax_t is wider than long.  */