by using long double where available.
Unfortunately, it won't be available everywhere, so a better solution
would still be nice.
Also, sometimes rounding of smaller sizes doesn't work right yet.
Fix gitlab issue #5.
The parse_float() function has several bugs:
1. For 1- and 2-word floats it will always write all 4 bytes if the value is
exactly 0.
2. It incorrectly rounds and normalizes 1- and 2-word floats.
To see the issue easily try the following inputs:
.flt2 1.0
.flt4 1.0
These will assemble as '040100 000000' and '040200 000000 000000 000000'.
They should both begin '040200'.
In fact the test file test/test-prec.mac is incorrect in its only floating point value:
.word ^f 1.5 ; 040140
should actually assemble as 040300, not 040140 (040140 is actually ^F0.875).
I confirmed this on RT-11 running MACRO V05.06.
I fixed the problem with the following deltas:
[the patch]
The most crucial change is the very last one. 0x200000000000 is actually (1 << 45)
and because ufrac is normalized it means that it will always downshift ufrac by 1.
several compilers:
- gcc version 4.5.3 (NetBSD nb2 20110806)
- gcc version 4.9.2 (Ubuntu 4.9.2-10ubuntu13)
- Ubuntu clang version 3.6.0-2ubuntu1 (tags/RELEASE_360/final) (based on LLVM 3.6.0)
The warnings were mostly about local variables shadowing others,
unused function parameters, and C++ style comments. Some variables were
indeed used before set.
Note that on Linux, using -std=c99 does stupid things like *remove* the
declaration of strdup() from <string.h>. Therefore I've reluctantly used
-std=gnu99.
They behave observably different from generic string parsing and trying
to account for them generically just gets in the way.
.rept is treated the same as a macro.
The code for nested and unnested strings was mixed, thereby being too
simple for the nested case. I separated them, which is simpler than
the fixed combined version.
Detected by the macro call
..EMIT <$FAC=^D<fb$rea>>
Also, to make Kermit sources work, only increase the local symbol block
counter if there actually was a local symbol used in the block.
This way, conditional inclusions (which include source text only in the
first pass) will have less potential for de-synchronisation between the
passes. After all, if the generated internal local symbol names do
not match, phase errors will result (showing themselves as strange label
redefinition problems).
