Re: [RFC] New SRA and early interprocedural SRA

[Laurent GUERBY <laurent at guerby dot net>]

On Mon, 2009-03-09 at 21:19 +0100, Eric Botcazou wrote:
> > Since you might actually understand the testcases, I would be grateful
> > if you could confirm or rule  out that this is possible.  The relevant
> > parts of the acats.log file are attached.
> 
> The test_rational_arithmetic.adb error with the only bootstrapped compiler 
> clearly points to a miscompilation of the compiler itself.  This probably 
> should be investigated last in any case.
> 
> I don't think that code quality can affect the ACATS results and in particular 
> the results of the tasking tests; it's more probably a miscompilation, either 
> of the test or of the runtime.  You can run ACATS at -O0 to get a better idea 
> (see testsuite/ada/acats/run_all.sh).

I confirm that when compiling Ada ACATS tasking tests (mainly c9xxx) the
Ada frontend compiler generates lots of code touching locks and queue
and other often complex tasking data structures so they tend to be more
sensible to code generation bugs than other ACATS tests.

For timing I haven't seen speed matter since processors reached 1 GHz
nine years ago :), so it's a very unlikely explanation.

For reference (but it's not needed here) you can augment all ACATS
timing delays by tweaking the following lines of
gcc/testsuite/ada/acats/support/impdef.a :

   One_Second : constant Duration := 0.001;
   One_Long_Second : constant Duration := 0.1;

Augmenting these values will lead to longer wait times and so longer
ACATS run time (unmodified ACATS tests have delays of 30 seconds or
more, sometimes needed for 1983 era computers :).

Laurent