Description of the paste cleanup that accidentally got checked in
during commit 944 because hg import -f has side effects.
http://landley.net/hg/toybox/rev/944
(Yes, that commit description is actively ironic. Scroll down to
patch.c.)
Minor tweaks to the first loop: if -d isn't set then TT.delim will be
NULL, and as long as that's the value we're playing with anyway the
optimizer tends to do better if it can load that into a register and
perform several operations on it, rather than load/mask/test an
otherwise unrelated integer. So yank the test and instead start the
loop with p = TT.delim ? TT.delim : "\t"; (In theory gcc offers an "x ?
: y;" syntax where the blank between ? : substitutes in the result of
the test again, so I _could_ say p = TT.delim ? : "\t";. In practice,
c99 never picked that one up so I feel uncomfortable using it.)
I was on the fence about moving the assignment from the else case into
the if, but did it anyway, so:
- if (*p == '\\') {
- } else *buf = *p;
becomes:
+ if ((*buf = *p) == '\\') {
+ }
This means we can load *p into a register once and then perform both
the assignment and the test from that register. The initial write only
goes into L1 cache and if it gets overwritten a couple lines later we
only have one write to main memory. In general modern processors have a
better time eliding writes than predicting branches. A lot of stuff
like:
if (x) y = 1;
else y = 2;
Is going to get rewritten by the optimizer into:
y = 2;
if (x) y = 1;
So the "if (x)" can become a conditional assignment instruction instead
of a branch instruction. (This avoids bubbles in the instruction
pipeline due to branch mispredicts, and/or wasted work from speculative
execution on mispredicted branches eating your battery power and then
being discarded.)
I was on the fence because this is phrasing easier for the computer to
do, and less repetitive (only one mention of *p instead of 2), but
slightly less easy to read.
Turned the p++ on its own line into a ++p in the stridx.
Changed the error message to say "bad -d" instead of "bad delimiter" so
it's one less word to translate to other languages.
Moved the "// Sequential" comment from on the same line to the line
before. I know it uses more vertical screen space and I ordinarily try
to conserve that, but mixing comment and code on the same line is
something I only really do describing data values. (Describing
structure members, table entries, or magic constants when a symbol name
is not conveniently available in a header.) This comment is more "what
does this block of code do", and that I prefer to put on its own line.
This loop is what loopfiles() is for. That handles the aliasing of "-"
with stdin and everything.
(No, I'm probably not done cleaning up paste, this was just my initial
pass that got accidentally checked in.)
Rob
_______________________________________________
Toybox mailing list
Toybox@lists.landley.net
http://lists.landley.net/listinfo.cgi/toybox-landley.net
