Re: Library testsuite

[Brooks Moses <brooks dot moses at codesourcery dot com>]

FX Coudert wrote:
> I'll just ask a silly question:
> > I divided the libgfortran testsuite into two subdirectories [...]  
> > the tests in intrinsics cycle through the usual set of options.
>
> If the tests in intrinsics are supposed to cover intrinsics for which  
> code is emitted directly by the front-end, why are we having them in  
> the library testsuite rather than the front-end testsuite?

I think it's probably better to think of this as an "intrinsics" 
testsuite, which has stuff in it that can be constant-folded, and stuff 
that's implemented with library calls.  Whether we put it under 
"libgfortran" or "gfortran" is sort of arbitrary, since it's testing 
both, and I think it's more useful to keep the two halves together.

> Here's what I was thinking of:
>    1. put all I/O tests in an I/O directory in the libgfortran testsuite
>    2. create also a intrinsics directory, which is run with only one  
> optimization level, and
>    3. we clearly document the difference, and start moving stuff from  
> gfortran.dg to the libgfortran testsuite slowly, by taking care of  
> whether the test is actually testing front-end emitted code or rather  
> library calls
>    4. exercise same care when deciding where a new test will end up

I'm not sure that paying that much attention to whether the test is 
testing front-end folding or library calls is the best approach, really; 
I don't see that it buys us much.  It seems to me that it makes more 
sense to automatically write two tests for each intrinsic -- one that 
can in principle be constant-folded relatively easily, and one that cannot.

Yes, that does give us some duplicates for intrinsics that we don't 
constant-fold (though we could possibly have a means to suppress the 
foldable ones), but it means that we can autogenerate testcases for the 
whole suite of intrinsics without having to think about special-casing 
each one, and I think that's a more useful result.

Then, there should probably be a third directory to hold the various 
special-case tests that come up from PRs and the like.  I think those 
should be separate from the autogenerated ones, though, and the bulk 
should be kept as simple to autogenerate as possible.

(We could also, conceivably, have some dump-tree tests that check to 
make sure that constant-folding actually happens, but I think that's a 
step 4+N for some positive integer N on your list.  :))

> FX, who's writing a lot, but it's past 2am and I'm fed up with  
> writing my thesis :)

Heh.  Same here, though it's only 6pm here.

- Brooks