When building for ARMv7 with thumb2 instructions, the optimized
`FT_MulFix_arm' assembly routine was not being used.
The reason for this is in the `ftconfig.h' header, namely:
- The assembly routine uses the `smull' instruction which is not
available when generating Thumb-1 machine code. It is available
in Thumb-2 mode, though.
- The header was written a long time ago before Thumb-2 became
widely popular (e.g. with Android). So it simply doesn't use the
assembly routine if the `__thumb__' built-in macro is defined.
- When compiling in Thumb-2 mode, the compiler will define both
`__thumb__' and `__thumb2__'.
By checking for `(__thumb2__ || !__thumb__)', we ensure that the
assembly routine is only avoided when generating Thumb-1 code.
Given that this is performance-sensitive function, this improves
`ftbench' as follows on a Galaxy Nexus:
Before (us/op) After (us/op)
- loading Arial.ttf glyphs at 14 ppem [1]
Load 34.285 33.098
- same operation with the light auto-hinter [2]
Load 31.317 29.590
- same operation without hinting [3]
Load 6.143 5.376
- loading Arial.ttf advances at 14 ppem [4]
Load_Advances (normal) 34.216 33.016
Load_Advances (fast) 0.176 0.176
[1] ftbench -t 5 -p -s 14 -b a -f 0008 Arial.ttf
[2] ftbench -t 5 -p -s 14 -b a -r 1 -f 0028 Arial.ttf
[3] ftbench -t 5 -p -s 14 -b a -f 000a Arial.ttf
[4] ftbench -t 5 -p -s 14 -b b -f 0008 Arial.ttf
* builds/unix/ftconfig.in, include/freetype/config/ftconfig.h
(FT_MULFIX_ASSEMBLER): Fix handling for ARMv7.
a5f33eeb8a