I confirm the same results for your ArbitraryPrecisionFloat example below, testing on Linux with a 64 bit interpreter VM and 32 bit image.
Image ----- /home/lewis/squeak/Squeak4.6/squeak.37.image Squeak4.6 latest update: #15723 Current Change Set: Unnamed2 Image format 6504 (32 bit)
Virtual Machine --------------- /usr/local/lib/squeak/4.15.4-3634/squeakvm Squeak4.5 of 10 December 2015 [latest update: #1195] Unix built on Mar 25 2016 13:27:51 Compiler: 4.9.2 platform sources revision 3634 VMMaker versionString 4.15.4
Dave
On Sun, Mar 27, 2016 at 03:25:58PM +0200, Nicolas Cellier wrote:
Hi Levente, thank you. Here is a follow up:
I was able to reproduce the same symptom on the classical interpreter VM. <rant> It's just that for MacOSX, there is no recent VM available, and that the MacOS build is somehow abandonware, so it took much more time than necessary :( </rant>
So the changes that have been integrated on Interpreter Vm in july 2014 and that I thought correct were not. It's certainly related to a LargeInteger plugin primitive. Those are triggered when nbits of magnitude > 64. (otherwise, <= 64 bits, the LargeInteger numbered primitives do their job).
Now the job is to produce/isolate an elementary failing LargeInteger test. But you need to generate the VM from VMMaker.oscog head for COG/STACK V3/SPUR 32/64 bits or use a sufficiently recent interpreter VM (> july 2014) to trigger such failure.
Here is my starting point:
If you load ArbitraryPrecisionFloat and Tests from http://www.squeaksource.com/ArbitraryPrecisionFl then you'll see this strange pattern:
(53 to: 70) collect: [:i | i -> (10 asArbitraryPrecisionFloatNumBits: i) erf ]
{53->(ArbitraryPrecisionFloat readFrom: '0.9999998981732067' readStream numBits: 53) . 54->(ArbitraryPrecisionFloat readFrom: '1.0' readStream numBits: 54) . 55->(ArbitraryPrecisionFloat readFrom: '1.0' readStream numBits: 55) . 56->(ArbitraryPrecisionFloat readFrom: '1.0' readStream numBits: 56) . 57->(ArbitraryPrecisionFloat readFrom: '0.99999989817320666' readStream numBits: 57) . 58->(ArbitraryPrecisionFloat readFrom: '1.0' readStream numBits: 58) . 59->(ArbitraryPrecisionFloat readFrom: '1.0' readStream numBits: 59) . 60->(ArbitraryPrecisionFloat readFrom: '1.0' readStream numBits: 60) . 61->(ArbitraryPrecisionFloat readFrom: '0.9999998981732066655' readStream numBits: 61) . 62->(ArbitraryPrecisionFloat readFrom: '1.0' readStream numBits: 62) . 63->(ArbitraryPrecisionFloat readFrom: '1.0' readStream numBits: 63) . 64->(ArbitraryPrecisionFloat readFrom: '1.0' readStream numBits: 64) . 65->(ArbitraryPrecisionFloat readFrom: '0.99999989817320666564' readStream numBits: 65) . 66->(ArbitraryPrecisionFloat readFrom: '1.0' readStream numBits: 66) . 67->(ArbitraryPrecisionFloat readFrom: '1.0' readStream numBits: 67) . 68->(ArbitraryPrecisionFloat readFrom: '1.0' readStream numBits: 68) . 69->(ArbitraryPrecisionFloat readFrom: '0.999999898173206665646' readStream numBits: 69) . 70->(ArbitraryPrecisionFloat readFrom: '1.0' readStream numBits: 70)}
The value seems incorrect every four numbits The exact value is near 1- 2.0e-45, and given the insufficient precision requested, the result should be rounded to 1. We can confirm with online high precision erf calculator like found for example with http://keisan.casio.com/exec/system/1180573449
Or we can check with (10 asArbitraryPrecisionFloatNumBits: 200) erf (ArbitraryPrecisionFloat readFrom: '0.999999999999999999999999999999999999999999997911512416237455' readStream numBits: 200)
Note that when I change the LargeInteger digitLength to 32bits instead of 8 (my own VM branch) then the problem magically disappear.
I plan to introduce these change in the VMMaker.oscog, so I can as well ignore the problem and proceed, however I much prefer to isolate it, and understand it for gain of knowledge :)
2016-03-27 4:27 GMT+02:00 Levente Uzonyi leves@caesar.elte.hu:
Hi Nicolas,
If you give us a bit more detailed instructions about what to test, then we might be able to help you get this sorted out.
Levente