0.9.4.8:

	Add a fallback implementation of SB-ROTATE-BYTE:%UNSIGNED-32-ROTATE-BYTE
        for platforms without the optimized VOPs, that isn't as pessimal as
        the generic %ROTATE-BYTE. About 70x speedup for SB-MD5 on x86-64.

diff --git a/NEWS b/NEWS
index 39717a2..2f1eef1 100644 (file)
--- a/NEWS
+++ b/NEWS
@@ -2,6 +2,8 @@
 changes in sbcl-0.9.5 relative to sbcl-0.9.4:
   * bug fix: interrupts are disabled until startup is complete; no
     more sigsegvs when receiving a signal to soon
+  * optimization: Faster 32-bit SB-ROTATE-BYTE:ROTATE-BYTE on non-x86/ppc
+    platforms
   * threads
     ** bug fix: parent thread now can be gc'ed even with a live
        child thread

diff --git a/contrib/sb-rotate-byte/compiler.lisp b/contrib/sb-rotate-byte/compiler.lisp
index 2e7d5e1..2d0f195 100644 (file)
--- a/contrib/sb-rotate-byte/compiler.lisp
+++ b/contrib/sb-rotate-byte/compiler.lisp
@@ -56,3 +56,15 @@
   ;; FIXME: What happens when, as here, the two type specifiers for
   ;; COUNT overlap?  Which gets to run first?
   '(%unsigned-32-rotate-byte count integer))
+
+;; Generic implementation for platforms that don't supply VOPs for 32-bit
+;; rotate.
+#-(or x86 ppc)
+(deftransform %unsigned-32-rotate-byte ((.count. .integer.)
+                                        ((integer -31 31)
+                                         (unsigned-byte 32)) *)
+  '(if (< .count. 0)
+       (logior (ldb (byte 32 0) (ash .integer. (+ .count. 32)))
+               (ash .integer. .count.))
+       (logior (ldb (byte 32 0) (ash .integer. .count.))
+               (ash .integer. (- .count. 32)))))

diff --git a/version.lisp-expr b/version.lisp-expr
index 6f0e405..716efaf 100644 (file)
--- a/version.lisp-expr
+++ b/version.lisp-expr
@@ -17,4 +17,4 @@
 ;;; checkins which aren't released. (And occasionally for internal
 ;;; versions, especially for internal versions off the main CVS
 ;;; branch, it gets hairier, e.g. "0.pre7.14.flaky4.13".)
-"0.9.4.8"
+"0.9.4.9"