[Bug target/79436] [ARM Cortex-M4F] VFMA used in place of subtraction gives inexact results

["freddie_chopin at op dot pl" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79436

--- Comment #4 from Freddie Chopin <freddie_chopin at op dot pl> ---
Hello Andrew and thanks for your answer.

Generally I don't care about the sequence of operations or the exact
instructions that get generated, but in this case this default behaviour
produces invalid results.

Generally the whole calculations are like this:
mx = ex - sx;
my = ey - sy;
distance = sqrtf(mx * mx + my * my) * constant;

The important thing here is that ex and sx are bitwise identical, just as ey
and sy. Thus everything above can be transformed to:

mx = x - x;
my = y - y;
distance = sqrtf(mx * mx + my * my) * constant;

and then:

mx = 0;
my = 0;
distance = sqrtf(0 * 0 + 0 * 0) * constant;

However you rearrange that, the expected result is 0 and I see no place for
"typical" floating point inaccuracies here. Let me reiterate - "startVector"
and "endVector" in my test case are bitwise identical. Yet the code produces
1.34925369e-06 at the end...

The same code compiled at -O2 for x86-64 does not assert. I don't know x86-64
assembly, but I'm pretty sure that it supports this kind of instructions.

If the results of VFMA are considered "good enough" I think that the default
value of -ffp-contract should be changed to "off" - after all
-funsafe-math-optimizations or -ffast-math are not enabled by default either.

BTW - VFMA is used also with "-std=c++11".

Thus I think that the bug is not invalid and I ask you to reconsider. Thanks in
advance!