Specially handle short strings in the parser #475
Conversation
These short strings are read using the appropriate constructor, rather than the general buffer constructor. So the empty and singleton constructors are now public. Other changes include optimising foreign lpvm cast instructions applied to constant chars, integers and floats, as they get generated in this case. This avoids several instructions in calling unboxed constructors with constant arguments. Also use the -n wybemk flag in final-dump-test.sh, to turn off colourising, making diffs easier to read. There are numerous code improvements in final dump tests, including allowing specialisation to kick in in many places, since unboxed values cannot be aliased. There's one unfortunate change in output for constant_type_constraint_error.wybe: the error message for ?a = "s":int now complains about a type error in the call to wybe.string.singleton, which doesn't appear on the line. I think this is tolerable for now, as it would be difficult to fix. Closes #473
|
Sorry for being a bit late to the party, but I'm not sure Expansion is the right place to do this. I believe this string specialisation pass should only ever happen once, as the transformation is idempotent. Expansion may only get ran once at the moment (correct me if I am wrong), but theoretically it could get ran multiple times, so the specialisation pass would be redundant after the first Expansion pass in the module. I realise that we haven't been focusing on making the compiler the most efficient, but I feel we should avoid inefficiencies if we are aware of them. |
|
Expansion is only run once (not positive there are no circumstances when it is repeated, but I think it's not). But you're right, we may want to repeat it. This shouldn't be a problem, though, since after this is done, we don't have the ArgString constructor for empty or singleton strings, so the change is idempotent. We're already scanning every argument during expansion, to update variables, so looking at ArgStrings and checking them for empty or singleton lists is a minimal added overhead. And if we moved it somewhere else, it would impose a similar overhead. |
These short strings are read using the appropriate constructor, rather than the general buffer
constructor. So the empty and singleton
constructors are now public.
Other changes include optimising foreign lpvm cast instructions applied to constant chars, integers
and floats, as they get generated in this case.
This avoids several instructions in calling
unboxed constructors with constant arguments.
Also use the -n wybemk flag in final-dump-test.sh, to turn off colourising, making diffs easier to
read.
There are numerous code improvements in final dump tests, including allowing specialisation to kick
in in many places, since unboxed values cannot be
aliased.
There's one unfortunate change in output for
constant_type_constraint_error.wybe: the error
message for ?a = "s":int now complains about a
type error in the call to wybe.string.singleton,
which doesn't appear on the line. I think this is
tolerable for now, as it would be difficult to
fix.
Closes #473