issue with "compliling" code with Math expressions

[espruino-discuss2]

Posted at 2022-07-19 by ioi-655321

I'm using formulas to gernerate a response curve and when I use the following function it gives the wrong answer when I use the compiled keyword. It works fine without being compiled, i was wondering if there was a known issue with the Math library or if there is a simple fix

function F(x){
"compiled";
return 1/(1.13+Math.exp(-8.0*(x-0.54)))+0.14;
}

i'm running the code on the original espruino with the latest firmware.

[gfwilliams]

Posted at 2022-07-20 by @gfwilliams

Hi - Thanks - I've just reproduced it here. It seems to be something to do with floating point numbers when used as function arguments.

Even something like this fails:

function F(x){
  "compiled";
  print (x-0.54);
}

function G(x){
  print (x-0.54);
}

for (var i=0;i<10;i++) {
  print("----",i);
  F(i);G(i);
}

I'll look into this now and see if it's an easy fix

[gfwilliams]

Posted at 2022-07-20 by @gfwilliams

Ok - try now!

Turns out it was two issues :)

• On nRF52 (which I was using initially) we switched to use the FPU a while back and that messed up how doubles are passed - so code using doubles didn't work reliably. But you wouldn't have hit that (probably the code above would have worked fine on the original Espruino)
• The type handling of the unary minus (for -8.0) was wrong - and I think that's what hit you

Running:

function F(x){
"compiled";
return 1/(1.13+Math.exp(-8.0*(x-0.54)))+0.14;
}
function G(x){
return 1/(1.13+Math.exp(-8.0*(x-0.54)))+0.14;
}
for (var i=0;i<10;i++) {
  print(F(i),G(i));
}

Now shows the right values for me

[espruino-discuss2]

Posted at 2022-07-20 by @fanoush

we switched to use the FPU a while back and that messed up how doubles are passed

the -mfloat-abi=softfp was there for a reason so that float could be passed like ints in R0-R4 so that InlineC code could receive float arguments (like this). Is there a reason -mfloat-abi=hard is needed?

THe -mfloat-abi=softfp -mfpu=fpv4-sp-d16 is a compromise so that both code using softfloats and code with fpu instructions could be linked together.

[espruino-discuss2]

Posted at 2022-07-20 by @fanoush

Was talking about this change gfwilliams/EspruinoCompiler@d5534b6 I was thinking espruino itself is build with that convention too?

Oh just checked and for nrf52 it is not while for stm32 it is!
https://github.com/espruino/Espruino/search?q=mfpu%3Dfpv4-sp-d16&type=code

make/family/NRF52.make
ARCHFLAGS = -mcpu=cortex-m4 -mthumb -mabi=aapcs -mfloat-abi=hard -mfpu=fpv4-sp-d16
make/family/STM32F4.make
ARCHFLAGS += -mlittle-endian -mthumb -mcpu=cortex-m4 -mfpu=fpv4-sp-d16 -mfloat-abi=softfp

I'd vote for softfp calling convention as the performance difference is minimal especially when javascript uses doubles and software implementation anyway.

[gfwilliams]

Posted at 2022-07-21 by @gfwilliams

so that float could be passed like ints in R0-R4 so that InlineC code could receive float arguments

Ahh - yes, this makes sense - although since sending floats as function arguments isn't something that's really supported (it has to be done by a bit of a hack) it probably explains why nobody hit this before. I wasn't aware that changing from soft/hard actually changed the argument passing behaviour. nRF52 builds have been like this since at least 2017!!

Changing from hard -> softfp actually seems to shave a few bytes off the firmware size, so I've committed a change - however it's a tricky one - it doesn't really help anyone on nRF52 who is using the firmware that's out there now. They'll still have the issue using any floating point numbers in code.

I guess maybe I'll wait until 2v15 gets released and then switch the compiler back to using softfp

[espruino-discuss2]

Posted at 2022-07-21 by @fanoush

Changing from hard -> softfp actually seems to shave a few bytes off the firmware size

yes, thanks for testing, I also made a build and saw that (292300 bytes softfp vs 293468 bytes hardfp on MDBT42Q build) but did not actually test in the device. btw noticed there is also -DFLOAT_ABI_HARD few lines below in the makefile possibly for nordic SDK but keeping it there or removing it did not change the size of build at all so not sure it is doing anything

[espruino-discuss2]

Posted at 2022-07-21 by @fanoush

although since sending floats as function arguments isn't something that's really supported (it has to be done by a bit of a hack)

in inlineC it is definitely a hack as the header clearly says int but code uses float in declaration and it works thanks to softfp calling convention

var c = E.compiledC(`
// int fadd(int,int)
// int fmul(int,int)
float fadd(float f1,float f2){
return f1+f2;
}
float fmul(float f1,float f2){
return f1*f2;
}

so if softfp build will work fine and will stay, it might make sense to allow float keyword also in header comment declaration of InlineC methods

var c = E.compiledC(`
// float fadd(float,float)

to make it less hacky

[espruino-discuss2]

Posted at 2022-07-21 by ioi-655321

Thank you for the quick response, I will try again and see what happens.

[espruino-discuss2]

Posted at 2022-07-21 by ioi-655321

just tested it extensivly and it works like a charm! thank you.

[gfwilliams]

Posted at 2022-07-22 by @gfwilliams

Great! Glad it's all sorted!