Re: Play 20K expressions again, Sam

[gvls1!lonjers@louie.udel.edu (Jim Lonjers)]

In article <1992Sep9.180517.20788@inmet.camb.inmet.com> arra@orb.camb.inmet.com
 (Arra Avakian) writes:
>Perhaps even more significant, if the developers bootstrap (compile their
>compiler using their compiler), then you know that their compiler has
>successfully compiled a fairly large and complex application program,
>increasing the liklihood YOUR program will compile and execute without problem
s.
>(As has been mentioned here before, the ACVC does not serve this purpose!)
>
>Real world test suites are critical to compiler quality, no matter the 
>language it compiles or the its implementation language.
>Boostrapping puts a real world test case in the hands of the compiler develope
r,
>with a very desirable impact on its quality.

While this is undoubtably true, we must really be careful about how much a
compiler vendor believes in this as a testing method.  In a prior life, I
was responsible for the Unisys ALGOL compiler which was written in Unisys
ALGOL.  We implemented coverage analysis, which marked which lines in a
program are covered in an execution.  After compiling the compiler with the
compiler, we ended up with on 20% of the compiler covered.

There are lots of reasons for this, but the compiler

1. is a batch program and thus does not do any interactive I/O.
2. is a single thread of execution so it used no tasking.
3. had no static errors, nor any uncorrected execution time
   errors, so the error detection/correction logic was not used.
4. had to be fast (~ 10_000 lines per minute on a 1 MIPS machine),
   so it did nothing with dynamic memory allocation.

This list can get very long, but you get the idea.