Floats and doubles in C have different size, regardless the size of floats in pharo who are always 64bits. 
But a call to a function(float a) expects a 32bits float and not a 64bit so a cast needs to be done (automatically in FFI), now if you are calling a function(float *a) and you have a float=42.0 you have to pass an address with place to store 32bits, not 64bits��� so

(ByteArray new: 4)
floatAt: 1 put: 42.0

��� will do the job. 
But if you have a function(double *a), you need to do: 

(ByteArray new: 8)
doubleAt: 1 put: 42.0

Now��� old nativeboost was prepared to do that conversion for you (if this was a good idea or not, is another story and needs to be told in another moment, and thread).
So, to allow operations style: 

UNIXTime localTime: (UNIXTime getTime: nil)

where: 

UNIXTime>>localTime: time
^ self ffiCall: #(tm* localtime(time_t* time) ) 
 module: LibC

and:  

getTime: t
^self ffiCall: #(time_t time(time_t* t) ) 
module: LibC
options: #(+ optCoerceNilToNull)

(from OS-Unix package)

To work fine, I need to reproduce that behaviour��� which is what the #rollToArity: method does.

So yes, is useful��� very useful.
 

On 18 Feb 2016, at 17:51, Henrik Sperre Johansen <henrik.s.johansen@veloxit.no> wrote:

I must admit, I don't get it either...
When you have a signature with a pointer to a primitive type, it's usually because
a) The method replaces the value
b) The method expects a pointer to an array with items of the primitive type.

In neither case is automatic marshalling from Smalltalk type very useful.

On Thu, Feb 18, 2016 at 4:33 PM, Cl��ment Bera <bera.clement@gmail.com> wrote:
OK nevermind I got it now. I talked with Esteban. I had misunderstood.

2016-02-18 14:57 GMT+01:00 Cl��ment Bera <bera.clement@gmail.com>:
I don't understand.

In 32 bits in C, there are 32 bits float called "float" or "single precision float" but its usage is very uncommon. People always use "double" as every 32 bits processors supports 64 bits float nowadays. When I look at the Cairo reference manual, only double are used, so I don't understand why you need to cast BoxedFloat64 to single precision float for FFI. The ThreadedIA32FFIPlugin has both #ffiPushDoubleFloat:in: and ffiPushSingleFloat:in: so all the support seems to be available. 

Can you explain why by default you reduce the precision of float to 32 bits ? Is it because the library used expects a single precision float to be passed ?



2016-02-18 11:57 GMT+01:00 Esteban Lorenzano <estebanlm@gmail.com>:
but in C (32bits) floats are 4 bytes, so when you do: 

ByteArray new 
floatAt: 1 put: somePharoFloat;

it will cast it to 4 bytes (if it fits which is most of the cases).

but as I said��� it is an error I need to fix :)

Esteban


On 12 Feb 2016, at 20:50, Cl��ment Bera <bera.clement@gmail.com> wrote:

I don't understand. All BoxedFloat have 8 bytes, that's why they are called BoxedFloat64. There are no 4-bytes floats in Pharo. Even in 32 bits.

2016-02-12 18:06 GMT+01:00 Esteban Lorenzano <estebanlm@gmail.com>:

On 12 Feb 2016, at 17:50, Aliaksei Syrel <alex.syrel@gmail.com> wrote:

On Fri, Feb 12, 2016 at 5:45 PM, Esteban Lorenzano <estebanlm@gmail.com> wrote:
Float>>#pointer 
^ (ByteArray new: FFIExternalType pointerSize)
floatAt: 1 put: self;
yourself

Yes it does the trick :) Thanks!

yes and no (I made a mistake): that works because incidentally pointers and floats have same size (4 bytes)��� but can happen that it is not (doubles have 8 bytes)��� but is ok for now :)
I will fix it correctly next week. 

Esteban