0.7.1.20:

[sbcl.git] / src / cold / warm.lisp

diff --git a/src/cold/warm.lisp b/src/cold/warm.lisp
index 05c7539..9997616 100644 (file)
--- a/src/cold/warm.lisp
+++ b/src/cold/warm.lisp
@@ -13,6 +13,20 @@
 \f
 ;;;; general warm init compilation policy
 
+;;; Without generational GC, GC gets really slow unless we collect in
+;;; large chunks. For small chunks, efficiency tends to grow roughly
+;;; linearly with chunk size. Later we hit diminishing returns as we
+;;; approach the total amount of RAM we use, or we can even get into
+;;; performance trouble by clobbering cache and VM systems too hard.
+;;; But modern machines tend to think of 20 Mb as a moderate amount of
+;;; memory, and it's of the same order of magnitude as the amount of
+;;; RAM we need for the build, so it seems like a plausible chunk size.
+#-gencgc
+(progn
+  (sb!ext:gc-off)
+  (setf (sb!ext:bytes-consed-between-gcs) (* 20 (expt 10 6)))
+  (sb!ext:gc-on))
+
 (proclaim '(optimize (compilation-speed 1)
                     (debug #+sb-show 2 #-sb-show 1)
                     (inhibit-warnings 2)