GCC miscompiles Fp2 with --opt:size flag

[mratsim]

From #228,

This fails

nim c --cc:gcc -r -d:release --opt:size --outdir:build tests/math/t_fp2.nim

Those do not

nim c --cc:clang -r -d:release --opt:size --outdir:build tests/math/t_fp2.nim
nim c --cc:gcc -r -d:release --outdir:build tests/math/t_fp2.nim

The function that is miscompiled is

constantine/constantine/math/extension_fields/assembly/fp2_asm_x86_adx_bmi2.nim

Lines 117 to 135 in 93dac25

	func mul_fp2_complex_asm_adx*(
	r: var array[2, Fp],
	a, b: array[2, Fp]
	) =
	## Complex multiplication on 𝔽p2
	var d {.noInit.}: array[2,doublePrec(Fp)]
	d.mul2x_fp2_complex_asm_adx(a, b)
	r.c0.mres.limbs.redcMont_asm_adx_inline(
	d.c0.limbs2x,
	Fp.fieldMod().limbs,
	Fp.getNegInvModWord(),
	Fp.getSpareBits()
	)
	r.c1.mres.limbs.redcMont_asm_adx_inline(
	d.c1.limbs2x,
	Fp.fieldMod().limbs,
	Fp.getNegInvModWord(),
	Fp.getSpareBits()
	)

[mratsim]

Following further investigation in #230, I believe that the compiler finds out that M is a constant by constant folding through compilation units with LTO and generating new procedures with only the variable input
As the inline assembly directly use the pointer as an input register, it's likely that its value has been hardcoded. This wouldn't be a problem if the address of the constant didn't change but who knows what the linker does (and there is relocatable code and Address Space Layout Randomization as well)