Problems with <double-float> on 64-bit

[pedro-w]

While investigating floating point things on 64-bit, C back end.

There are failing tests in common-dylan/tests/functions.dylan, one example is on line 1002

assert-equal(classify-double-float(#x00000000, #x7ff00000), #"infinite");

Should create a float from the raw data 0x7FF0000000000000
Calling

opendylan/sources/common-dylan/tests/functions.dylan

Lines 982 to 986 in 9f9ab34

	define inline function classify-double-float (low-bits, high-bits)
	=> (classification :: <float-classification>)
	classify-float(encode-double-float(as(<machine-word>, low-bits),
	as(<machine-word>, high-bits)))
	end function classify-double-float;

Calling

opendylan/sources/dylan/float.dylan

Lines 287 to 292 in 30b4a91

	define inline-only function encode-double-float
	(low :: <machine-word>, high :: <machine-word>) => (encoded :: <double-float>)
	primitive-raw-as-double-float
	(primitive-cast-machine-words-as-double-float(primitive-unwrap-machine-word(low),
	primitive-unwrap-machine-word(high)))
	end function encode-double-float;

Calling

opendylan/sources/lib/run-time/c-primitives-math.c

Lines 120 to 129 in 6f60c21

	DDFLT primitive_cast_machine_words_as_double_float(DUMINT low, DUMINT high) {
	INTDFLT intflt;
	#ifdef NO_LONGLONG
	ignore(high);
	intflt.i = (DULMINT)low;
	#else
	intflt.i = ((DULMINT)high << LONG_BIT) | (DULMINT)low;
	#endif
	return(intflt.f);
	}

For X64, NO_LONGLONG is defined so this is not going to work as it is assuming all 64 bits are in the low machine-word and nothing useful is in the high MW.

Hence in this case, the test fails as the number appears to be zero rather than infinite (I think classify-float itself is OK)

[pedro-w]

I see it is possible to just directly use a 64 bit number on 64 bit platforms

assert-equal(classify-double-float(#x7ff0000000000000, #x00000000), #"infinite");

so we just need to check $machine-word-size at the appropriate time.
Still would be better if it were possible to specify a sized integer independent of the platform!

[housel]

housel@e20f211 has a fix for this, consistently splitting <double-float> across two machine words on both 32-bit and 64-bit platforms.

[pedro-w]

That looks like a good change, would it be worth cherry-picking from your branch? (I don't think your corba-tuneup branch has been merged?)

[waywardmonkeys]

@housel did cherry-pick it and submit it as a PR (#1255), which has now been merged to master.