Using float instead of double

[zup1]

I see the code and I noted that the constants with decimal points are of type double instead of float. For example 8.0 instead 8.0f, so the math operations are executed as double

[IndianBoy42]

I was going to say I would be surprised if such a massive pessimization (hardware float vs software double) would have gone unnoticed till now given the performance requirements for a foc driver. So i decided to check on godbolt and sure enough, it definitely generates 32bit instructions on every compiler version i could try, even on O0. there was no difference no matter how I wrote the operations as long as the result of the expression was cast/put into a float(32).

https://godbolt.org/z/P55xahxEK

[kalvdans]

@IndianBoy42 I think you just were lucky with the number being exactly representable in float32. If you try changing the constant to something uneven, you get 64-bit division as required by the C standard:

void utils_fft8_bin0_f(float *real_in, float *real, float *imag) {
	*real = 0.0;
	*imag = 0.0;

	for (int i = 0;i < 8;i++) {
		*real += real_in[i];
	}

	*real /= 8.7;
}

[nitrousnrg]

The code would explode if that was the case, the compiler is instructed to default those variables to 32 bit float with
-fsingle-precision-constant

IIRC it happens here

bldc/make/fw.mk

Line 15 in a53d898

USE_OPT += -fsingle-precision-constant -Wdouble-promotion -specs=nosys.specs

[kalvdans]

The code would explode if that was the case, the compiler is instructed to default those variables to 32 bit float with -fsingle-precision-constant

Nice, that means this issue can be closed. That flag doesn't exist for clang though, if one is brave enough to try another compiler ;)