https://issues.dlang.org/show_bug.cgi?id=15619

ponce <aliloko@gmail.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|[REG 2.068]                 |[REG 2.068] Floating-point
                   |                            |x86_64 codegen regression

--

https://issues.dlang.org/show_bug.cgi?id=15619

--- Comment #1 from ponce <aliloko@gmail.com> ---
Reduced to:

--------------------------------
import std.stdio;

void fillWithOnes(float* outputs, int frames)
{
    float[] A = new float[frames];
    float[] B = new float[frames];

    A[] = 0;
    B[] = 0;

    float one = returnOne();

    A[0..frames] += B[0..frames];

    for (int i = 0; i < frames; ++i)
    {
        outputs[i] = one;
    }
}

float returnOne()
{
    return 1;
}

void main()
{
    float[] buf = new float[16];
    fillWithOnes(buf.ptr, 16);
    writeln(buf);
}
--------------------------------

--

https://issues.dlang.org/show_bug.cgi?id=15619

ponce <aliloko@gmail.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|[REG 2.068] Floating-point  |[REG 2.068] Floating-point
                   |x86_64 codegen regression   |x86_64 codegen regression,
                   |                            |when involving array ops

--

https://issues.dlang.org/show_bug.cgi?id=15619

Kenji Hara <k.hara.pg@gmail.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |wrong-code
            Summary|[REG 2.068] Floating-point  |[REG 2.066] Floating-point
                   |x86_64 codegen regression,  |x86_64 codegen regression,
                   |when involving array ops    |when involving array ops

--- Comment #2 from Kenji Hara <k.hara.pg@gmail.com> ---
Both original and reduced code generate wrong result since 2.066.0 (2.065 is
ok).

--

https://issues.dlang.org/show_bug.cgi?id=15619

yebblies <yebblies@gmail.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |yebblies@gmail.com
          Component|dmd                         |druntime

--- Comment #3 from yebblies <yebblies@gmail.com> ---
The array operation uses xmm6/xmm7 internally without saving them, but they are callee saved on win64.

Probably introduced by this: https://github.com/D-Programming-Language/druntime/pull/829

--

https://issues.dlang.org/show_bug.cgi?id=15619

Martin Nowak <code@dawg.eu> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |code@dawg.eu
           Assignee|nobody@puremagic.com        |code@dawg.eu

--- Comment #4 from Martin Nowak <code@dawg.eu> ---
I'll take this. https://trello.com/c/TNCQ6xBk/171-issue-15619-reg-2-066-floating-point-x86-64-codegen-regression-when-involving-array-ops

--

https://issues.dlang.org/show_bug.cgi?id=15619

--- Comment #5 from Martin Nowak <code@dawg.eu> ---
As this is already broken for a while (and way longer for other array ops, e.g. double), I'll delay this fix for a proper replacement w/ templated array ops.

https://github.com/D-Programming-Language/druntime/pull/471#issuecomment-16075234
WIP:
https://github.com/MartinNowak/dmd/tree/arrayOps
https://github.com/MartinNowak/druntime/tree/arrayOps

--

https://issues.dlang.org/show_bug.cgi?id=15619

Walter Bright <bugzilla@digitalmars.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |bugzilla@digitalmars.com

--- Comment #6 from Walter Bright <bugzilla@digitalmars.com> ---
Doesn't seem like this should be too hard to fix - just save xmm6-7 for Win64 - why not just do it?

--

https://issues.dlang.org/show_bug.cgi?id=15619

--- Comment #7 from ponce <aliloko@gmail.com> ---
FWIW I don't use array ops anymore, to the exception of:

 slice[] = value;
 sliveA[] = sliceB[];

--

https://issues.dlang.org/show_bug.cgi?id=15619

--- Comment #8 from Martin Nowak <code@dawg.eu> ---
(In reply to Walter Bright from comment #6)
> Doesn't seem like this should be too hard to fix - just save xmm6-7 for Win64 - why not just do it?

Because all the other functions ignore it as well and maintaining that mess is just wasted time.

--