1 Apr
2022
1 Apr
'22
11:43 p.m.
Steffen Nurpmeso writes:
Jon Steinhart wrote in
<202204011726.231HQFm03349496(a)darkstar.fourwinds.com>:
|David Barto writes:
|>> On Apr 1, 2022, at 8:59 AM, Douglas McIlroy <douglas.mcilroy(a)dartmouth.e\
|>> du> wrote:
|>> The recent discussion about Research choosing BSD's paging over
|>> Reiser-London's brought to mind a stunning program by Reiser that
|>> Research did adopt.
|>>
|>> A critical primitive in the Blit terminal was bitblt (block transfer
|>> of a rectangular area). It was used ubiquitously, for example to
...
|>> Bitblt got refined (i.e. elaborated) several times before Reiser did
|>> away with it entirely. Instead he wrote a just-in-time generator of
|>> optimal code. Thousands of distinct variants, which varied in size
|>> from 16 to 72 bytes, could be produced by the same 400 lines of
|>> assembler code.
...
|> Does this exist for the rest of us to study?
...
|It's not insanely complicated by modern standards. Without any knowledge
|of other work, I did the same thing for a 68020 based graphics system where
|the JIT code went into the I-cache and was amazingly fast for its day.
|
|If I remember correctly, things started with an outer-loop test to see
|if there were overlapping regions to determine whether to go forward
|to backwards to avoid having the destination trash the source.
...
Only to add that "modern standard" C libraries define
"overlapping" by means of exclusivity, meaning that memcpy(x,
&x[1], 1) is an invalid overlapping that requires memmove() to be
used.
Uh, yeah, but not relevant at least to the code I did which was written
in C but generated machine code, not calls to library functions.