Floating point comparsion operators
Timmoore
Posts: 1,031
in Propeller 2
Has anyone had issues with the floating point <. and >. spin operators?
I am seeing these give unexpected results with pnut 39/40 and propeller tool 2.9.3 but flexprop 6.5 is giving the result I would expect.
The test code is at https://github.com/timmmoore/Swervebot/blob/master/test/test.spin2
for example
if test1 <. test2
debug("if test1 <. test2")
else
debug("else test1 <. test2")
will print the else when test1 = 1.0 and test2 = 2.0.
I think the problem is in the spin_interpreter, in particular this code in frel_
bitl w,#31 wcz 'make left-side comparable
cmp w,##$7F800001 wc 'NaN?
if_z neg w '+/- %0_xxxxxxxx_mmmmmmmmmmmmmmmmmmmmmmm
bitl x,#31 wcz 'make right-side comparable
if_c cmp x,##$7F800001 wc 'NaN?
if_z neg x '+/- %0_xxxxxxxx_mmmmmmmmmmmmmmmmmmmmmmm
I believe the carry used in the 2nd cmp, should be from the previous cmp but it is overwritten by the 2nd bitl.
When I change it not to do this I get the expected result.
Tim
Comments
I don't use Spin but I think changing one line fixes this issue:
frel_ popa w 'pop left-side float bitl w,#31 wcz 'make left-side comparable cmp w,##$7F800001 wc 'NaN? if_z neg w '+/- %0_xxxxxxxx_mmmmmmmmmmmmmmmmmmmmmmm if_c bitl x,#31 wcz 'make right-side comparable ***CHANGED*** if_c cmp x,##$7F800001 wc 'NaN? if_z neg x '+/- %0_xxxxxxxx_mmmmmmmmmmmmmmmmmmmmmmm if_nc skip #%11111 'if either NaN, return false cmps w,x wcz 'do comparison if_z mov y,#%1000 'equal? if_nz mov y,#%0100 'above? if_c mov y,#%0010 'below? test y,pa wc 'test mask _ret_ muxc x,_FFFFFFFF 'return booleanThanks for catching this!
I see I had to use WCZ after BITL, in order to affect Z, so the C result was just incidental and gets ignored. The next instruction meaningfully affects C, but then there's some error in my logic. TonyB probably nailed it, but I need to get on my computer and work through this.
I will have this fixed in several hours.
Thanks, The test file on github has several tests cases though not exhausive, but does check <. >. <=. and >=. and the test cases try the obvious cases, i.e., +ve, -ve, true and false.
It tests a different fix as well but it adds 1 instruction so TonyB fix is better,
I just tried TonyB code with my test cases and it didnt' give the right answer.
How about this?
frel_ popa w 'pop left-side float bitl w,#31 wcz 'make left-side comparable cmp w,##$7F800001 wc 'NaN? if_z neg w '+/- %0_xxxxxxxx_mmmmmmmmmmmmmmmmmmmmmmm if_nc skip #%111111111 'if NaN, return false ***NEW*** bitl x,#31 wcz 'make right-side comparable if_c cmp x,##$7F800001 wc 'NaN? if_z neg x '+/- %0_xxxxxxxx_mmmmmmmmmmmmmmmmmmmmmmm if_nc skip #%11111 'if NaN, return false cmps w,x wcz 'do comparison if_z mov y,#%1000 'equal? if_nz mov y,#%0100 'above? if_c mov y,#%0010 'below? test y,pa wc 'test mask _ret_ muxc x,_FFFFFFFF 'return booleanremove the if_c on the 2nd cmp. Then it works.
frel_ popa w 'pop left-side float bitl w,#31 wcz 'make left-side comparable cmp w,##$7F800001 wc 'NaN? if_z neg w '+/- %0_xxxxxxxx_mmmmmmmmmmmmmmmmmmmmmmm if_nc skip #%111111111 'if NaN, return false ***NEW*** bitl x,#31 wcz 'make right-side comparable cmp x,##$7F800001 wc 'NaN? if_z neg x '+/- %0_xxxxxxxx_mmmmmmmmmmmmmmmmmmmmmmm if_nc skip #%11111 'if NaN, return false cmps w,x wcz 'do comparison if_z mov y,#%1000 'equal? if_nz mov y,#%0100 'above? if_c mov y,#%0010 'below? test y,pa wc 'test mask _ret_ muxc x,_FFFFFFFF 'return booleanTim
deleted
Remember those ##$7F80000x instructions need TWO skips to get over.
Thanks for thinking this over. I am kind of leaning towards this:
frel_ popa w 'pop left-side float bitl w,#31 wcz 'make left-side comparable cmp w,##$7F800001 wc 'NaN? if_nc skip #%1111 'if NaN, return false if_z neg w '+/- %0_xxxxxxxx_mmmmmmmmmmmmmmmmmmmmmmm bitl x,#31 wcz 'make right-side comparable cmp x,##$7F800001 wc 'NaN? if_nc skip #%111111 'if NaN, return false if_z neg x '+/- %0_xxxxxxxx_mmmmmmmmmmmmmmmmmmmmmmm cmps w,x wcz 'do comparison if_z mov y,#%1000 'equal? if_nz mov y,#%0100 'above? if_c mov y,#%0010 'below? test y,pa wc 'test mask _ret_ muxc x,_FFFFFFFF 'return booleanThere seems to be no way to shorten the ##value matter.
Thanks
forgot the ##, should have had some nan test cases.
I added 2 test cases, 1st parameter nan and 2nd parameter nan, and changed the code to match yours.
All test cases pass, my reading of the code is comparision will return false if either parameter is NAN.
Thanks, Tim. I implemented those changes.
Could you guys please test this new PNut_v40a.exe to see if it passes your FP equality operator tests?
Yes, its passing all my tests. I attached the debug output for my run so you can see the tests run.
Ok. Super! Thanks, Tim. I will update everything now.
Okay, new version posted:
https://drive.google.com/file/d/1CsPLelJWs9P6i3G1QhmIPoIRZH6m_I5t/view?usp=sharing