www.digitalmars.com         C & C++   DMDScript  

digitalmars.D.bugs - Very strange bug causing Access Violation with -O and -release

reply Deewiant <deewiant.doesnotlike.spam gmail.com> writes:
DMD 0.128 on Windows XP SP2.
--
void foo(inout real[] arr) {
	// i = 0 doesn't trigger the bug, everything bigger seems to
	// also, making this const fixes the bug
	size_t i = 1;
	
	for (size_t j = i; j >= i; j -= i) {
		// dummy line which fixes the bug goes here
		// it can be anything that accesses j in some way
		// such as any of:
		//size_t x = j;
		//printf("%d", j);
		//j = i;
		
		// interesting results follow from this:
		//printf("%d", i);
		// it prints a _lot_ of ones

		arr[j] = arr[j - i];
	}
}

void main()
{
	real[] array;
	array.length = 2; // whatever, as long as it's more than 1
	
	foreach (inout real i; array)
		i = 1; // just something
	
	foo(array);
}
--
IMO this bug is strange because:
	It is only triggered if the array is of type cdouble, real, ireal, or
creal. (FWIW, my machine's a 1.2 GHz Duron - reals on other CPUs might
not trigger the bug?)
	The code must be compiled with both -release and -O to trigger the bug.
	What's explained in the comments... a dummy line accessing j inside the
for loop fixes the bug, if i is const the bug is fixed, if you print i
within the for loop you get _many_ of them. The number seems to depend
on the type of the array (cdouble, real, ireal, creal) and on its
length. The smallest number I got was 402 (creal, array length 2) and
the largest was 1227 (real, array length 2).
Aug 06 2005
next sibling parent =?ISO-8859-15?Q?Thomas_K=FChne?= writes:
-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Deewiant schrieb:


 DMD 0.128 on Windows XP SP2.
 --
 void foo(inout real[] arr) {
 	// i = 0 doesn't trigger the bug, everything bigger seems to
 	// also, making this const fixes the bug
 	size_t i = 1;
 	
 	for (size_t j = i; j >= i; j -= i) {
 		// dummy line which fixes the bug goes here
 		// it can be anything that accesses j in some way
 		// such as any of:
 		//size_t x = j;
 		//printf("%d", j);
 		//j = i;
 		
 		// interesting results follow from this:
 		//printf("%d", i);
 		// it prints a _lot_ of ones
 
 		arr[j] = arr[j - i];
 	}
 }
 
 void main()
 {
 	real[] array;
 	array.length = 2; // whatever, as long as it's more than 1
 	
 	foreach (inout real i; array)
 		i = 1; // just something
 	
 	foo(array);
 }
 --
 IMO this bug is strange because:
 	It is only triggered if the array is of type cdouble, real, ireal, or
 creal. (FWIW, my machine's a 1.2 GHz Duron - reals on other CPUs might
 not trigger the bug?)
 	The code must be compiled with both -release and -O to trigger the bug.
 	What's explained in the comments... a dummy line accessing j inside the
 for loop fixes the bug, if i is const the bug is fixed, if you print i
 within the for loop you get _many_ of them. The number seems to depend
 on the type of the array (cdouble, real, ireal, creal) and on its
 length. The smallest number I got was 402 (creal, array length 2) and
 the largest was 1227 (real, array length 2).


Added to DStress as
http://dstress.kuehne.cn/run/f/for_06_A.d
...
http://dstress.kuehne.cn/run/f/for_06_R.d

Thomas
-----BEGIN PGP SIGNATURE-----

iD8DBQFC9N4x3w+/yD4P9tIRAiDFAJ0TfLvRewo2CsW3pnsPQPze1b0iqQCghclO
O0/cmObBKYVh7U5P7YJrW2k=
=iEYh
-----END PGP SIGNATURE-----
Aug 06 2005
prev sibling parent "Walter" <newshound digitalmars.com> writes:
You've found a bug in the optimizer. It'll get fixed in the next update.
Thanks, -Walter
Aug 10 2005