Page 1 of 1
[PB 6.00] ASM vs. C: different floating point results
Posted: Thu May 26, 2022 9:23 pm
by Little John
Hi,
I have written several PureBasic procedures that do statistical calculations. Some of them give different results depending on whether they are compiled with the ASM backend or with the C backend. After several tests I found out that there are differences already for basic floating point calculations.
Here is just one simple example:
Code: Select all
; PureBasic 6.0 Beta 8 (x64) on Windows
x.d = 1.9379 * 0.7295
Debug x ; 1.4136980499999998
Debug #PI ; 3.1415926535897931
x / #PI
Debug x ; ASM backend: 0.44999406539374676
; C backend: 0.44999401577479958
Re: [PB 6.00] ASM vs. C: different floating point results
Posted: Thu May 26, 2022 10:01 pm
by skywalk
Yeah, this is a big problem for me.
ASM matches the MS Calculator.
C+opt is almost like stuck at FLOAT instead of DOUBLE?
Re: [PB 6.00] ASM vs. C: different floating point results
Posted: Thu May 26, 2022 11:53 pm
by jack
the problem is with the constant #PI, if you change it to PI.d=3.1415926535897932 then it's ok
Re: [PB 6.00] ASM vs. C: different floating point results
Posted: Fri May 27, 2022 3:14 pm
by Little John
jack wrote:the problem is with the constant #PI, if you change it to PI.d=3.1415926535897932 then it's ok
This actually fixed all the compatibility issues in my private statistics library.
Thanks a lot!
Re: [PB 6.00] ASM vs. C: different floating point results
Posted: Fri May 27, 2022 4:17 pm
by infratec
A short test:
Code: Select all
Debug #PI
#Test1 = 3.1415926535897932384
Debug #Test1
#Test2 = 3.1415926535897932
Debug #Test2
Debug StrD(#Test2, 20)
Define f.f
Define d.d
f = #Test1
d = #Test1
Debug f
Debug StrD(d, 20)
Results in:
Code: Select all
3.1415926535897931
3.1415926535897931
3.1415926535897931
3.14159265358979310000
3.14159274101257
3.14159265358979310000
PB 5.73 x86 on WIn x64
Very strange.
Re: [PB 6.00] ASM vs. C: different floating point results
Posted: Fri May 27, 2022 5:25 pm
by STARGÅTE
It is not the #PI constant itself, who makes problems.
It is the short form of the operation x = x * y --> x* y
Code: Select all
Debug #PI
d1.d = #PI
Debug d1
d2.d = 1.0 * #PI
Debug d2
d3.d = 1.0
d3 * #PI
Debug d3
This is not the first one of this kind of bugs in the C-backend and it is challenging to find all these bugs, because they are not obviously.
Re: [PB 6.00] ASM vs. C: different floating point results
Posted: Fri May 27, 2022 6:21 pm
by Little John
STARGÅTE wrote: Fri May 27, 2022 5:25 pm
It is not the #PI constant itself, who makes problems.
It is the short form of the operation x = x * y --> x* y
Hi,
using the long form of the operation might fix the issue in some cases, but unfortunately is not a
general workaround, see
https://www.purebasic.fr/english/viewto ... 98#p584898