Fonts typically have multiple consecutive PUSHB / PUSHW
operators. Merge them in the disassembled output.
This, for now, generates worse assmebly because all PUSHB's
grouped with PUSHW's now generate as PUSHW. Followup fix
will address that.
Previously fonttools was choosing between the optimal opcode
from PUSHB, NPUSHB, PUSHW, and NPUSHW. We now respect
whatever was requested and err if the format doesn't support
the data.
Morevoer, if the number of numbers to push is more than 255,
we add multiple push instructions to push all the numbers.
Finally, add a new pseudo-opcode "PUSH" that automatically
chooses the best format.
This, by itself, reduces roundtrip noise by not optimizing
the bytecode as it was before. In a followup commit I'll
change the bytecode disassembler to always produce PUSH
in the textual instructions instead of the four variants.
That way, we get both the optimization during assembling,
and reduced noise in XML.
Part of https://github.com/behdad/fonttools/issues/73
Sounds useless, but MReifutaiM-Regular.ttf seems to have that, and the
parsing back from XML was failing because we expected a token whereas
after dropping whitespace there wasn't any left.
New instructions can be defined in the prep program. We don't
check that the "unknown" instruction was actually defined. Just
pass it through.
Fixes ttx with Skia.ttf on the Mac OS X.
Fixes https://github.com/behdad/fonttools/issues/21
