Hi!

I'm not 100% sure if I'm in the right here, but I believe that the following program shouldn't have unbounded heap growth to the point where it runs out of memory:

module gctest;

import core.memory;
import core.thread;

void main()
{
    for(int i = 0; i < 10000; i++) {
        int[] x = new int[10000000];
        Thread.sleep(dur!"msecs"(200));
        //GC.collect(); // keeps program from failing
    }
}

If GC.collect() is called constantly, heap size stays pretty much constant. Tested only on Windows 10 (64 bit) with DMD 2.086.0 so far. I'd say that this behavior is a bug.

I also noticed:
where code with arrays allocations compiled by LDC work ok, code by DMD is failed.
"gc:precise" doesn't help (maybe false positives scans.. dunno).
this code by DMD crashes for me at -m32 only

Windows Server 2019 (x64)
DMD 2.086.0
LDC 1.16.0-beta2

and yes, LDC drops all code in main, coz u dont use arrays/results and alloc int[] do nothing too (no constructors with some logic)
so, I checked with sum(array) and printing result (=0)

On Friday, 7 June 2019 at 12:28:50 UTC, Gregor Mückl wrote:
> Hi!
>
> I'm not 100% sure if I'm in the right here, but I believe that the following program shouldn't have unbounded heap growth to the point where it runs out of memory:
>
> module gctest;
>
> import core.memory;
> import core.thread;
>
> void main()
> {
>     for(int i = 0; i < 10000; i++) {
>         int[] x = new int[10000000];
>         Thread.sleep(dur!"msecs"(200));
>         //GC.collect(); // keeps program from failing
>     }
> }
>
> If GC.collect() is called constantly, heap size stays pretty much constant. Tested only on Windows 10 (64 bit) with DMD 2.086.0 so far. I'd say that this behavior is a bug.

It's a bit stranger than that: the following program doesn't fail only because it has the writeln with the string concatenation in it. I discovered it by accident while trying to look into how Windows reports available memory to 32 bit programs.

module gctest;

import core.thread;
import std.conv;
import std.stdio;

void main()
{
    for(int i = 0; i < 10000; i++) {
        int[] x = new int[10000000];
        Thread.sleep(dur!"msecs"(200));
        writeln("allocation " ~ to!string(i));
    }
}

Just the writeln without any concatenation will still not trigger garbage collection runs. What is it about the ~ operator implementation that changes the behavior so drastically?

As a side note, I don't think that GlobalMemoryStatus on Windows works the way the GC code thinks it's working. It will always happily report 2^31 bytes total and available memory, no matter how big the heap of the 32 bit program is. So isLowOnMem in gc/os.d should actually be broken in this particular situation.

On Friday, 7 June 2019 at 14:24:23 UTC, KnightMare wrote:
> I also noticed:
> where code with arrays allocations compiled by LDC work ok, code by DMD is failed.
> "gc:precise" doesn't help (maybe false positives scans.. dunno).
> this code by DMD crashes for me at -m32 only

Thanks for reproducing! So it's a dmd bug then.

DMD automatically generates a 32 bit executable for me, so my tests are hitting the 4GB boundary long before physical memory availability becomes an issue.

GC.collect() does collect the arrays, so I don't think that false positives during the scan shouldn't be the issue.

I filled issue https://issues.dlang.org/show_bug.cgi?id=19947
add something if u want (maybe link to this topic)

On Friday, 7 June 2019 at 14:47:44 UTC, KnightMare wrote:
> I filled issue https://issues.dlang.org/show_bug.cgi?id=19947
> add something if u want (maybe link to this topic)

Thanks for reporting! I'll monitor the bug there.