Starting with tag:
[TAG triple jcc version
peter@cordes.ca**20080320081143] 
[working computed-jmp version
peter@cordes.ca**20080320084938
 not as fast as triple-branch except when triple branch fell through to
 the last one.  But hopefully less polluting of the branch predictor
] 
[handy command line for shift in the comments
peter@cordes.ca**20080320085219] 
[align 1024
peter@cordes.ca**20080322044318] 
[simple loop instead of computed goto to deal with alignment
peter@cordes.ca**20080322050316] 
[update comments
peter@cordes.ca**20080322102733] 
[better-pipelined main loop based on Torbjorn's work.  With good short-loop intro
peter@cordes.ca**20080322105200
 runs 1.331c/l
] 
[shift.c: set up a useful test pattern
peter@cordes.ca**20080322105348] 
[distribute ALU ops through the loop.  add commented-out 8-way unroll
peter@cordes.ca**20080323013830
 spreading out the ALU ops doesn't seem to make any difference.
 The 8-way unroll is slightly faster with large n, but way slower with
 small n that make the intro loop run more.  change test $3, %dl to
 test $7, %dl and uncomment the bottom half of the main loop.
] 
[test-pattern generator doesn't segfault
peter@cordes.ca**20080323014421] 
[change config.m4 path for use in speed-ext
peter@cordes.ca**20080323060840] 
[fast 8-limb loop, 1.66 c/l.
peter@cordes.ca**20080323060928
 still reads past the end of src, so needs finishing touches
] [use macros to allow switching between 4-limb and 8-limb unroll
peter@cordes.ca**20080324212029
 also make it easy to move the add/lea up/down in the loop
] 
[speed-ext.c for tune/speed-ext
peter@cordes.ca**20080325024627
 Your darcs repo should be gmp-.../tune instead of tests/devel.
 Hardlinks should do the trick.
 Use the patch for Makefile.in.
 See the suggested command line in speed-ext.c
] 
[do the cleanup after the loop.  debugged and working, and pretty fast
peter@cordes.ca**20080325055033
 This works well: we can have special cleanup for coming out of the pipeline.
 Tested with electric fence, and it doesn't read past the end of src.
] 
[fifo rt prio
peter@cordes.ca**20080325101519] 
[new layout of mostly the same code.  only 2 icache lines for n<12, and mostly better speed
peter@cordes.ca**20080326023333] 
[notes on branching a repo and setting up speed-ext to use the other version
peter@cordes.ca**20080326030443] 
[better block ordering to remove more branches
peter@cordes.ca**20080326041203
 saves a few cycles for most n.
] 
[CPP macros to rename functions in speed-ext.c
peter@cordes.ca**20080326031120]