Hi all,

I was debugging something entirely different the other day, and while
doing so I used the -:d switch, and saw my entire terminal fill up with
messages like this:

[debug] resizing scratchspace dynamically from 0k to 2k ...
[debug] resizing scratchspace dynamically from 2k to 2k ...
[debug] resizing scratchspace dynamically from 2k to 2k ...
[debug] resizing scratchspace dynamically from 0k to 2k ...
[debug] resizing scratchspace dynamically from 2k to 2k ...
[debug] resizing scratchspace dynamically from 2k to 2k ...
[debug] resizing scratchspace dynamically from 0k to 2k ...
[debug] resizing scratchspace dynamically from 2k to 2k ...
[debug] resizing scratchspace dynamically from 2k to 2k ...

This is stupid, and I quickly determined that this must be some
piece of code that was written to be simple, being so lazy as
to _always_ pre-allocate a bignum even if the number wasn't so
big (as my code wasn't using bignums at all).

As it's a design goal to avoid even _having_ a scratchspace when
bignums are not being used, this is obviously a bug.  But it's
also a performance penalty, because each of those lines above
correspond to a malloc() being done, and there's also a corresponding
free().

The bug turned out to be in the C_s_a_i_digits_to_integer, the C helper
function that string->number uses to convert a string of digits to an
integer.  This is only ever used in two places in string->number, so
this is easily fixed by allocating a little bit more than necessary
for a bignum wrapper, and using that extra memory to fit a small bignum
into the nursery.  This space won't be used if the pre-allocated bignum
needs to be bigger, but in the vast majority of cases it's just a fixnum
that we're reading, so this should be a pretty big win.

In a few quick benchmarks it doesn't seem to make a difference (I guess
malloc just keeps reusing the same block), but at least it stops the
resizing scratchspace spam :)

Cheers,
Peter
From 1dfeabe2fa963919913993b65e6e7811b94f10cd Mon Sep 17 00:00:00 2001
From: Peter Bex <[email protected]>
Date: Thu, 9 Mar 2017 21:36:05 +0100
Subject: [PATCH] Avoid allocating scratchspace when reading fixnums.

The code that handles reading digit streams into an integer is written
to be as straightforward as possible.  Currently this means it always
allocates a bignum, even if the number is small enough to fit a
fixnum.  That way, we don't need to distinguish between these cases
in the actual stream converter (str_to_bignum).

We can improve on this a little bit by pre-allocating enough for a
small bignum in the nursery, much like the C_s_a_u_i_integer_plus and
friends do.  This avoids malloc()ing a scratchspace just for the one
small number, which will be free()d again as soon as the stack is
full, resulting in a lot of needless thrashing.
---
 library.scm |  4 ++--
 runtime.c   | 11 +++++++++--
 2 files changed, 11 insertions(+), 4 deletions(-)

diff --git a/library.scm b/library.scm
index 14fc9e5..284b653 100644
--- a/library.scm
+++ b/library.scm
@@ -1800,7 +1800,7 @@ EOF
                    (end (or hashes digits)))
               (and-let* ((end)
                          (num (##core#inline_allocate
-			       ("C_s_a_i_digits_to_integer" 2)
+			       ("C_s_a_i_digits_to_integer" 5)
 			       str start (car end) radix neg?)))
                 (when hashes            ; Eeewww. Feeling dirty yet?
                   (set! seen-hashes? #t)
@@ -1815,7 +1815,7 @@ EOF
                               (end (scan-digits start)))
                      (go-inexact!)
                      (cons (##core#inline_allocate
-			    ("C_s_a_i_digits_to_integer" 2)
+			    ("C_s_a_i_digits_to_integer" 5)
 			    str start (car end) radix (eq? sign 'neg))
                            (cdr end)))))))
          (scan-decimal-tail             ; The part after the decimal dot
diff --git a/runtime.c b/runtime.c
index e996559..af844df 100644
--- a/runtime.c
+++ b/runtime.c
@@ -10903,8 +10903,15 @@ C_s_a_i_digits_to_integer(C_word **ptr, C_word n, C_word str, C_word start, C_wo
     assert((radix > 1) && C_fitsinbignumhalfdigitp(radix));
 
     nbits = (end - start) * C_ilen(radix - 1);
-    size = C_fix(C_BIGNUM_BITS_TO_DIGITS(nbits));
-    result = C_allocate_scratch_bignum(ptr, size, negp, C_SCHEME_FALSE);
+    size = C_BIGNUM_BITS_TO_DIGITS(nbits);
+    if (size == 1) {
+      result = C_bignum1(ptr, C_truep(negp), 0);
+    } else if (size == 2) {
+      result = C_bignum2(ptr, C_truep(negp), 0, 0);
+    } else {
+      size = C_fix(size);
+      result = C_allocate_scratch_bignum(ptr, size, negp, C_SCHEME_FALSE);
+    }
 
     return str_to_bignum(result, s + start, s + end, radix);
   }
-- 
2.1.4

Attachment: signature.asc
Description: Digital signature

_______________________________________________
Chicken-hackers mailing list
[email protected]
https://lists.nongnu.org/mailman/listinfo/chicken-hackers

Reply via email to