[qcc] backends

[Bruce Ewing]

The qcc project that I'm working on compiles C code. It will leverage some
backend to support plenty of popular targets. It is very small and very
fast. What priorities of yours are not included in that list?

The only reason that TCG frontend code is not a compiler intermediate
language is that it hasn't been embedded into any compiler yet. At least
one known problem with it is that it only supports a single register return
value from any function, which is not C99 compatible -- and that deficit
means it's not even compatible with musl. ATM, it also has floating point
support turned off. The more mods that I have to make to the input (to make
TCG handle the full range of C99 C code), the more that I need to modify
*every single TCG backend* to support those mods. So, in some sense, the
more backends there are, the more of a headache it is. And the less likely
that it will actually happen in reality.

Creating a compiler is a game of tradeoffs, of course. Yes, the most
important thing to a user is that the compiler support their target CPU and
file format. The nextmost important thing to users will certainly be a lack
of compiler-introduced bugs in the output. Third is probably the code
quality of the output in terms of speed. Then the speed of compilation.
Then, for an open-source project, probably the clarity and obviousness of
the code and algorithm.

I have been expecting that TCG, or whatever backend is selected, will
produce slower code for a wide range of target CPUs. I have also been
expecting to create by hand additional, user-selectable, much more
optimized backends -- because *I* care very much about getting decent
performance out of x86 code, at least. I have a lot of techniques I've
developed over the years to beat gcc -O3 every time.

So, I am still probably a month away from beginning to select and deal with
a backend. TCG is certainly still a possibility. I'm not married to the
idea, but I haven't come close to rejecting it, either. But of course I
want a backend that needs as little modding as possible, within the
constraints of the game of tradeoffs, above. A backend that needs a lot
less modding and produces more optimized code may be preferable to one that
supports a huge handful of obsolete CPUs. Because each one of those
obsolete CPUs may add a lot of additional work.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.landley.net/pipermail/qcc-landley.net/attachments/20150506/3238747a/attachment.htm>