Short:        Patch the utility.library for 68060
Author:       matthey7@gmail.com (Matt Hey)
Uploader:     matthey7 gmail com (Matt Hey)
Type:         util/boot
Version:      1.2
Requires:     68060, Kickstart 3.0+
Architecture: m68k-amigaos >= 3.0

*** This patch no longer patches SDivMod32() & UDivMod32(). These functions
use divsl.l & divul.l which return 64 bits of results in 2 registers but are
NOT considered to be 64 bit integer instructions which are missing in the
68060. The previous versions of this patch should not be used anymore as the
SDivMod32() & UDivMod32() are slower after all. The SMult64() and UMult64()
are faster than what is in the Mu 68060.library in all cases but there are
few enough programs using these functions that this patch is barely worth
while. I tried to get Thomas Richter to include the code in his Mu
68060.library so this patch would be unnecessary but he likes the slower
Motorola code because it's tested better. Using a variation of the code in
FFmpeg was not proof enough I guess. ***

Description:

This is a small patch which replaces the UMult64() and SMult64() functions of
utility.library V39+ with faster 68060 optimized versions. 

UtilPatch060 tests for a 68060 processor. If it can't find one, it doesn't
install the patch and exits with a return code of 20 (=fail). If it can't
open utility.library V39+, it also does not install the patch and exits with
a return code of 20.

This code is based in part on Harry "Piru" Sintonen's NewPatchMult64. Thanks
Piru!


Features:

Avoids going through the CPU trap for unimplented integer exception.

Slightly faster UMult64() and SMult64() than Piru's NewPatchMult64.

Doesn't touch the MMU (directly) which is proper when using Thomas
Richter's mmu.library and 68060.library and also unnecessary.

Installation is fast, doesn't fragment memory, and doesn't use much memory.

The full source code is included. Assembled with PhxAss.

Free.


Installation:

Copy UtilPatch060 wherever you like. It runs from Workbench but I
recommend starting from cli in the S:startup-sequence after the setpatch
command. It does not detach from the shell so the following command is needed...

Run >NIL: UtilPatch060

As little as 512 bytes of stack should be fine. A control C will break
UtilPatch060 causing it to uninstall which can be dangerous because of how the
exec setpatch function works.


Debugging:

I have included a Snoopy 2.0 (Aminet:util/moni/snoopy20.lha) script that
logs the patched functions in the utility.library. Please report the values
logged if there are any errors. The default output goes to t:snoopy.txt but
can be changed in the icon tooltype. Please Report all bugs to the e-mail
address at the top.


Speed comparison:

Here are some speed comparisons using Mult64SpeedTest from Piru's
NewPatchMult64. The first is the Mu 68060.library vs UtilPatch060...

5000000xUMult64 (positiveXpositive): 207 187 9.6% faster
5000000xUMult64 (positiveXzero): 113 110 2.7% faster
5000000xSMult64 (positiveXpositive): 239 219 8.4% faster
5000000xSMult64 (negativeXpositive): 248 231 6.9% faster
5000000xSMult64 (any)product fits in 32 bits: 239-248 110 54.0%-55.6% faster
5000000xSMult64 (positiveXzero): 142 110 22.5% faster

and NewPatchMult64 vs UtilPatch060...

5000000xUMult64 (positiveXpositive): 191 187 2.1% faster
5000000xUMult64 (positiveXzero): 110 110 same speed
5000000xSMult64 (positiveXpositive): 215 219 1.9% slower
5000000xSMult64 (negativeXpositive): 223 231 3.6% slower
5000000xSMult64 (any)product fits in 32 bits: 215-223 110 48.8%-50.7% faster
5000000xSMult64 (positiveXzero): 110 110 same speed

Note that smaller numbers are better (less time). There is a commented out
version of SMult64() in the source that is slightly faster than Piru's in all
cases. My default SMult64() is much faster if the product of the multiply
fits in 32 bits and is a little slower than Piru's if it doesn't. I thought
this was a good tradeoff.


Future:

I may make a CPUpatch that applies all 68060 patches if a 68060 is detected
and all 68040 patches if a 68040 is detected, etc. I may have the mmu.library
protect the patches.


History:

1.0 (23.07.09)
 First release
1.1 (25.07.09)
 Saved a few bytes
1.2 (12.09.09)
 SDivMod32() & UDivMod32() are no longer patched as divsl.l & divul.l are in the 68060.
 Added missing Permit() after Replying to the WB message.


Thanks:

Harry "Piru" Sintonen and Dirk Busse for NewPatchMult64.
Thomas Richter for pointing out that divsl.l & divul.l are in the 68060 after all.