So I have finally narrow down and compare the calculus under
NetBSD/i386/gcc and Plan9/i386/ken-cc. And this has nothing to do with
div versus shift...
ken-cc is at fault here and it seems deep; it has apparently to deal
with registerization.
Data:
Under NetBSD/gcc, I have the following values:
before: x1:=5440, x2:=-5843, x3:=78909
after: x1:=5440, x2:=-201, x3:=18166, r:=6827 t:=30232
Under Plan9/gcc, I have the following values:
before: x1:=5440, x2:=-5843, x3:=78909
after: x1:=5440, x2:=2147483447, x3:=1073759990, r:=6827 t:=-1073711592
Uhm... seems to have a `slight' divergence...
In fact, all wrong values depend upon x2, that has the "correct"
value... with 2^31 complement. A positive when it should be negative,
since the offending code is the following:
x2 = half ( x1 + x2 + xicorr ) ;
that is :
x2 = (5440 - 5843 + 1) / 2;
Not exactly pushing things to the limit! And yes, the expected result is
indeed -201.
In fact, if you just extract this kind of expression, and compile a
test, there is no problem.
But the calling function is teratologic: it has 10 parameters and 18
automatic variables.
Since the problem arises in this context, but not if you just add
this isolated in a test program, and call it with these very 3
values (5440, -5843, 1), it is clear that's the way the computation
is handled with huge number of parameters and auto variables
that wreaks havoc.
If I declare all the auto volatile, this does nothing: same result.
If I do the addition, and afterwards take the half, that works:
x2 += x1 + xicorr;
x2 = half(x2); /* works! */
Problem: this is not a solution, but a hack, since this kind of 3
variables handling is common. This one fails. And the others?
Is there a limit about the number of parameters?
--
Thierry Laronde <tlaronde +AT+ polynum +dot+ com>
http://www.kergis.com/
Key fingerprint = 0FF7 E906 FBAF FE95 FD89 250D 52B1 AE95 6006 F40C
