chris@eneevax.UUCP (03/09/84)
Just for contrast, here's a place where the (4.1BSD) C optimizer did
something really fancy.
---part of comm.h---
/* Input and output translation is done via optional translation tables and
bits to set and clear. This allows any combination of input and output
parity (or lack thereof). */
struct Translate {
char *tr_tab; /* Translation table (if any) */
int tr_bic; /* Bits to clear */
int tr_bis; /* Bits to set */
};
[...]
struct Translate InTr; /* Input (link to tty) translation */
[...]
/* Apply the translation given by tp to c */
#define ApplyTranslation(tp, c) \
if ((tp)->tr_tab) \
c = (tp)->tr_tab[(c) & 0177]; \
c &= ~(tp)->tr_bic; \
c |= (tp)->tr_bis;
---part of proca.c---
register int c;
[...]
c = 7; /* Ding-a-ling */
ApplyTranslation (&InTr, c);
---The corresponding assembly code (edited slightly for readability)---
movl $7,r11
tstl _InTr
jeql L74
movl _InTr,r0
extzv $0,r11,r11,r1 # the interesting one
L2000041:
cvtbl (r0)[r1],r11
L74:
bicl2 _InTr+4,r11
bisl2 _InTr+8,r11
---------------------------
Apparently an "extzv $0,$7,x,y" instruction is faster than a
"bicl3 $-128,x,y". Well instead of generating the constant "$7"
for the extzv, c2 noticed that r11 had 7 in it already and just
used r11! At first I thought it was a bug!
Admittedly the code might be better if the "tstl" and "movl" were
collapsed into a single "movl" which also sets the condition codes.
Oh well, I guess that's asking too much of c2, what with L74:.
--
Chris Torek, Dept of CS, Univeristy of Maryland, College Park, MD
...!umcp-cs!chris chris%umcp-cs@CSNet-Relaychris@umcp-cs.UUCP (03/15/84)
All right, how come no one has replied yet??? I posted something
a while ago about something /lib/c2 did that was "neat", namely,
change
movl $7,r11
...
bicl3 $-256,r11,r1
into
movl $7,r11
...
extzv $0,r11,r11,r1 # look closely now
(instead of "extzv $0,$7,r11,r1"). If you just think for a minute,
the "extzv" is unnecessary in the first place! If c2 does limited
flow analysis to watch over register contents, how come it doesn't
eliminate instructions that are effectively no-ops? (That's a
rhetorical question, by the way.)
--
In-Real-Life: Chris Torek, Univ of MD Comp Sci
UUCP: {seismo,allegra,brl-bmd}!umcp-cs!chris
CSNet: chris@umcp-cs ARPA: chris.umcp-cs@CSNet-Relay