Re: Customer balks at Ada -- any hope?--Warning Significant Thread Drift Ahead

["Robert I. Eachus" <rieachus@earthlink.net>]

Kieran Mckey wrote:
  
> Ok, a quick reword gives :
> 
> run-time checks are required for Ada exception handling, excluding
> explicitly raised exceptions.

  Not right yet.  You really have to read and understand RM 11.6,
Exceptions and Optimization,
to have some clue as to what is required.  (And probably read Ada 83
AI-315 or have Brother Dewar or Brother Hilfinger preach on reasoning
from erroneousness.)  But the reality can best be summed
up in three principles:

   Ada checks are really on stored VALUES (or on subtype conversions)
not on operations.

   Ada compilers can and do STATICALLY eliminate most required checks.

   If your program runs efficiently with checks enabled, it will
probably run slower if you
disable them.  (Translation, only turn off checks where you know the
compiler is generating
"junk" checking code for some particular computation.   And it is often
much better to "turn checking off" by doing a complex computation in
type'Base, then converting the result.)

   And yes, I have seen code run slower when compiled with checks
disabled.  (Note that earlier version of GNAT did have significant
overhead for integer constraint checks, so by default they
were turned off.  They are still disabled by default for compatibility
reasons, but the code generated now is much more efficient than before.)

    Of course, in any discussion of supressing checks it is assumed that
the exception cannot or will not occur.  If a check for a particular
exception is suppressed and an exception occurs that results in
externally visible effects, then program speed is irrelevant.  I have
used the
compiler to validate my understanding of a section of code by compiling
with checks on and then
checking the generated code to see if it contains a call to raise an
exception.  (This only works if the hardware sets condition codes which
have to be checked.  If the hardware traps on overflow, or if there are
calls to run-time routines which actually do the test, the method
doesn't work.)
But on most modern hardware, look at the generated code.  Then make
minor changes to see how the checking code is affected.  Often moving a
bound from a parameter to a variable or vice-versa, or even putting it
in both places, will have a significant effect on compiler overhead.