The function uses Newton's iterations instead of dividing vector
components by its length, which needs a square root. This is,
literally, a bit less accurate but a lot faster.
* src/base/ftcalc.c (FT_Vector_NormLen): New function.
This makes the code more readable and reduces compiler warnings.
* src/base/ftcalc.c (FT_MulDiv, FT_MulDiv_No_Round, FT_MulFix,
FT_DivFix): Convert input parameters to unsigned, do the
computation, then convert the result back to signed.
(ft_corner_orientation): Fix casts.
* include/internal/ftcalc.h (FT_HYPOT): Macro to approximate Euclidean
distance with the alpha max plus beta min algorithm.
* src/base/ftcalc.c (ft_corner_is_flat): Use it instead of Taxicab metric.
We shift as many bits as we can into the high register, perform
32-bit division with modulo there, then work through the remaining
bits with long division. This optimization is especially noticeable
for smaller dividends that barely use the high register.
* src/base/ftcalc.c (ft_div64by32): Updated.
This applies to `FT_MulDiv' but not to `FT_DivFix', where overflows or
lack thereof are predicted accurately.
* src/base/ftcalc.c (ft_div64by32): Improve readability.
(FT_MulDiv, FT_MulDiv_No_Round) [!FT_LONG64]: Use straight division
when multiplication stayed within 32 bits.
This fixes a 13-year old bug. The original overflow check should have
been updated when rounding was introduced into this function
(c2cd00443b).
* src/base/ftcalc.c (FT_DivFix) [!FT_LONG64]: Updated.
* include/freetype.h (FT_DivFix): Updated documentation.
* include/freetype.h (FT_MulFix): Unconditionally defined.
* src/base/ftcalc.c [FT_MULFIX_ASSEMBLER]: Move code from here...
* include/internal/ftcalc.h [FT_MULFIX_ASSEMBLER]: ... to here,
which conditionally replaces the function with an inline version
through the macro.
FT_MulFix assembler used to reside in ftcalc.c before f47d263f1b.
* include/config/ftconfig.h, builds/unix/ftconfig.in,
builds/vms/ftconfig.h [FT_MULFIX_ASSEMBLER]: Move code from here...
* src/base/ftcalc.c [FT_MULFIX_ASSEMBLER]: ... to here.
* src/base/ftcalc.c (FT_DivFix): Use unsigned values for
computations which use the left shift operator and convert to signed
as the last step.
* src/base/fttrigon.c (ft_trig_prenorm, FT_Vector_Rotate,
FT_Vector_Length, FT_Vector_Polarize): Ditto.
* src/cff/cffgload.c (cff_decoder_parse_charstrings): Simplify.
* src/cff/cffload.c (cff_subfont_load): Fix constant.
* src/cff/cffparse.c (cff_parse_integer, cff_parse_real, do_fixed,
cff_parse_fixed_dynamic): Use unsigned values for computations which
use the left shift operator and convert to signed as the last step.
* src/cid/cidload.c (cid_get_offset): Ditto.
* src/psaux/psconv.c (PS_Conv_ToFixed): Ditto.
* src/psaux/t1decode.c (t1_decoder_parse_charstrings): Ditto.
* src/truetype/ttinterp.c (TT_MulFix14, TT_DotFix14): Ditto.
This is essentially a mechanical conversion, adding inclusion of
`FT_INTERNAL_DEBUG_H' where necessary, and providing the macros for
stand-alone compiling modes of the rasterizer modules.
To convert the remaining occurrences of FT_Err_XXX and friends it is
necessary to rewrite the code. Note, however, that it doesn't harm
if some cases are not handled since FT_THROW is a no-op.
For some divisions, we use casts to 32bit entities. Always guard
against division by zero with these casts also.
* src/base/ftcalc.c (ft_div64by32): Remove redundant cast.
(FT_MulDiv, FT_MulDiv_No_Round): Add 32bit cast.
(FT_DivFix): Add 32bit cast (this omission triggered the bug).
Werner Lemberg