ref: fc2a0b223565a46e642d762e659e2c589624eb13
parent: 0a245e1864fa69aeb86b444a7d98bee54ad2e1c1
author: thebard <thebard>
date: Tue Feb 22 23:02:21 EST 2000
Now encoding faster than realtime with suggested compiler options in Makefile. Still alot of room left for assembler and/or thread tweaking.
--- a/Makefile
+++ b/Makefile
@@ -1,9 +1,10 @@
-# On X86 machines it is HIGHLY Recomended that you compile only with
-# -march=i686 -finline-functions -funroll-loops -fomit-frame-pointer
-# Do NOT Compile with -O of ANY kind. It looks like gcc's optimizations are
-# either starving registers, or causing huge cache misses for Intel Processors.
-# Its currently twice as fast with no -O than any -O.
-#
+# On X86 machines it is HIGHLY Recomended that if you want to compile with any
+# -O setting you NOT use -O. use -O2 to -O6. -O is broken. Also note that you
+# MUST use -ffast-math with -O2 to -O6. If you do not, it will be SLOWER then
+# using no -O altogether. On a Pentium 3, the fastest optimizations are:
+# -O6 -march=i686 -finline-functions -funroll-loops -fomit-frame-pointer
+# -ffast-math
+#
prefix = /usr/local
DESTDIR =
--- a/profiling.txt
+++ b/profiling.txt
@@ -151,3 +151,11 @@
The Outfield Bangin': track 2: Bangin' On My Heart
Still Nothing.
NO PROFILE
+
+------------------------------------------------------------------------------
+thebard
+CVS Tue Feb 22 19:18:25 CST 2000
+-O6 -march=i686 -finline-functions -funroll-loops -fomit-frame-pointer -ffast-math
+3 min, 12.13 sec.
+The Outfield Bangin': track 2: Bangin' On My Heart
+Bingo! Got it. -O core dumps, and -O2 to -O6 are slower than thuer non optimized counterparts, unless you add -ffast-math. then things are speedy. The assembly it produces still has room for speed. Still more work to do.