Smalltalk › Squeak › Squeak VM

Funny Floats [WAS OpenSmalltalk current on aarch64]

_‹ Previous Topic Next Topic _›

Classic

List

Threaded

4 messages Options

KenDickey

Funny Floats [WAS OpenSmalltalk current on aarch64]

Greetings,

OK. So current VM on aarch64. I have an older VM in which floats work
fine (same machine).

I am doing diffs like:
diff -r {kens-,oscog}vm/platforms/Cross/plugins/FloatMathPlugin
diff -r {kens-,oscog}vm/platforms/unix/plugins/FloatMathPlugin
diff -r {kens-,oscog}vm/platforms/Cross/vm

but not finding differences in float calculations.

Any hints on where to look for "1.0 = 1.1 --> true" ?? [Same w Squeak
or Cuis]

Thanks much,
-KenD

-KenD

Christian Kellermann

Re: Funny Floats [WAS OpenSmalltalk current on aarch64]

* [hidden email] <[hidden email]> [191206 17:33]:

>
> Greetings,
>
> OK. So current VM on aarch64. I have an older VM in which floats work fine
> (same machine).
>
> I am doing diffs like:
> diff -r {kens-,oscog}vm/platforms/Cross/plugins/FloatMathPlugin
> diff -r {kens-,oscog}vm/platforms/unix/plugins/FloatMathPlugin
> diff -r {kens-,oscog}vm/platforms/Cross/vm
>
> but not finding differences in float calculations.
>
> Any hints on where to look for "1.0 = 1.1 --> true" ?? [Same w Squeak or
> Cuis]

Have you tried switching off compiler optimisations? I recall running
into similar problems with a newer gcc and turning off optimisations
has been a workaround.

--
May you be peaceful, may you live in safety, may you be free from
suffering, and may you live with ease.

Tobias Pape

Re: Funny Floats [WAS OpenSmalltalk current on aarch64]

> On 06.12.2019, at 17:46, Christian Kellermann <[hidden email]> wrote:
>
>
> * [hidden email] <[hidden email]> [191206 17:33]:
>>
>> Greetings,
>>
>> OK. So current VM on aarch64. I have an older VM in which floats work fine
>> (same machine).
>>
>> I am doing diffs like:
>> diff -r {kens-,oscog}vm/platforms/Cross/plugins/FloatMathPlugin
>> diff -r {kens-,oscog}vm/platforms/unix/plugins/FloatMathPlugin
>> diff -r {kens-,oscog}vm/platforms/Cross/vm
>>
>> but not finding differences in float calculations.
>>
>> Any hints on where to look for "1.0 = 1.1 --> true" ?? [Same w Squeak or
>> Cuis]
>
> Have you tried switching off compiler optimisations? I recall running
> into similar problems with a newer gcc and turning off optimisations
> has been a workaround.

I recall such a thing from another thread.
Apparently using clang also helped, IIRC

Best regards
-Tobias

Nicolas Cellier

Re: Funny Floats [WAS OpenSmalltalk current on aarch64]

This is because legacy fdlibm relies on pointer aliasing which is forbidden for some times now (it prevents efficient compiler optimization)

Try replacing fdlibm with https://github.com/nicolas-cellier-aka-nice/fdlibm

It was in our plans to integrate and maintain that replacement in opensmalltalk until we find a better library (both supported and doing the right thing)

Le ven. 6 déc. 2019 à 18:22, Tobias Pape <[hidden email]> a écrit :

> On 06.12.2019, at 17:46, Christian Kellermann <[hidden email]> wrote:
>
>
> * [hidden email] <[hidden email]> [191206 17:33]:
>>
>> Greetings,
>>
>> OK. So current VM on aarch64. I have an older VM in which floats work fine
>> (same machine).
>>
>> I am doing diffs like:
>> diff -r {kens-,oscog}vm/platforms/Cross/plugins/FloatMathPlugin
>> diff -r {kens-,oscog}vm/platforms/unix/plugins/FloatMathPlugin
>> diff -r {kens-,oscog}vm/platforms/Cross/vm
>>
>> but not finding differences in float calculations.
>>
>> Any hints on where to look for "1.0 = 1.1 --> true" ?? [Same w Squeak or
>> Cuis]
>
> Have you tried switching off compiler optimisations? I recall running
> into similar problems with a newer gcc and turning off optimisations
> has been a workaround.

I recall such a thing from another thread.
Apparently using clang also helped, IIRC

Best regards
-Tobias