CORDIC optimizer bug

[electrodude]

The following code:

PUB main(a) : x, y | b
  asm
                qexp    a
                neg     b, a
                qexp    b
                getqx   x
                getqx   y
  endasm

compiles to the following, which is missing the first qexp instruction and therefore produces incorrect results:

_main
	neg	_var01, arg01
	qexp	_var01
'   asm
	getqx	result1
	getqx	result2
_main_ret
	ret

I'm running the latest commit e37a63f51defd82a38811e20c0f7819ca76b9fc2, invoked as flexspin -O2 -2 -l. The program I made this example from works correctly if I disable optimization with org/end but fails due to this bug with asm/endasm.

[Wuerfel21]

Not a bug IMO, code like this (i.e. cordic interleave tricks in non-volatile inline ASM) was never guaranteed to work to begin with (try using the function in a context where it becomes inlined and only one result is used - that will result in brap in any version) It just happened to work most of the time. The change that breaks this isn't an optimization, it's the fix for bug #223

[totalspectrum]

I agree, not a bug, but it would be nice to have a warning in there. I think the optimizer should be able to see these kind of cases and suggest to the user that an "asm const" or something similar is required.

[Wuerfel21]

Yeah, that'd probably be good

[Wuerfel21]

Also, aren't we missing Spin syntax for volatile-but-not-cogexec flavour inline ASM?

[totalspectrum]

Yes, Spin has no way to do volatile hubexec ASM. It'd be easy to add, but I don't know what the syntax should be. Any suggestions?

[totalspectrum]

ORGH is used now for volatile but not cogexec inline ASM. Still need to look at a warning for CORDIC optimizations though.