Don't bother doing any performance analysis until most of the following items are taken care of, because there's no question they represent serious space/time problems, although some of them show up only given certain kinds of (popular) input.
malloc package and its uses to specify more info about
memory pools and, where feasible, use obstacks to implement them.
COMMON areas, EQUIVALENCE areas) so zeros need not be output.
This would reduce memory usage for large initialized aggregate
areas, even ones with only one initialized element.
As of version 0.5.18, a portion of this item has already been accomplished.
sta.c) so that the nature of the statement
is determined as much as possible by looking entirely at its form,
and not looking at any context (previous statements, including types
of symbols).
This would allow ripping out of the statement-confirmation,
symbol retraction/confirmation, and diagnostic inhibition
mechanisms.
Plus, it would result in much-improved diagnostics.
For example, CALL some-intrinsic(...), where the intrinsic
is not a subroutine intrinsic, would result actual error instead of the
unimplemented-statement catch-all.
g77, don't pass line/column pairs where
a simple ffewhere type, which points to the error as much as is
desired by the configuration, will do, and don't pass ffelexToken types
where a simple ffewhere type will do.
Then, allow new default
configuration of ffewhere such that the source line text is not
preserved, and leave it to things like Emacs' next-error function
to point to them (now that next-error supports column,
or, perhaps, character-offset, numbers).
The change in calling sequences should improve performance somewhat,
as should not having to save source lines.
(Whether this whole
item will improve performance is questionable, but it should
improve maintainability.)
DATA (A(I),I=1,1000000)/1000000*2/ more efficiently, especially
as regards the assembly output.
Some of this might require improving
the back end, but lots of improvement in space/time required in g77
itself can be fairly easily obtained without touching the back end.
Maybe type-conversion, where necessary, can be speeded up as well in
cases like the one shown (converting the 2 into 2.).
lex.c.
lex.c to not need any feedback
during tokenization, by keeping track of enough parse state on its
own.