tree 43b5d9bf890d4e38a6dbe68179c7a70b9d586a81
parent 57a64a717296e9f029bfcc3bba1d0b4c23d10cff
author Elliott Hughes <enh@google.com> 1569620559 -0700
committer Elliott Hughes <enh@google.com> 1569620559 -0700

Optimize tolower(3)/toupper(3) from <ctype.h>.

The tables in the BSD tolower/toupper are slower for ASCII than just
doing the bit twiddling.

We can't actually remove the tables on LP32, so move them into the
"cruft" we keep around for backwards compatibility (but remove them for
LP64 where they were never exposed).

I noticed that the new bit-twiddling tolower(3) was performing better
on arm64 than toupper(3). The 0xdf constant was requiring an extra MOV,
and there isn't a BIC that takes an immediate value. Since we've already
done the comparison to check that we're in the right range (where the
bit is always set), though, we can EOR 0x20 to get the same result as
the missing BIC 0x20 in just one instruction.

I've applied that same optimization to towupper(3) too.

Before:

  BM_ctype_tolower_n                 3.30 ns         3.30 ns    212353035
  BM_ctype_tolower_y                 3.31 ns         3.30 ns    211234204
  BM_ctype_toupper_n                 3.30 ns         3.29 ns    214161246
  BM_ctype_toupper_y                 3.29 ns         3.28 ns    207643473

  BM_wctype_towupper_ascii_n         3.53 ns         3.53 ns    195944444
  BM_wctype_towupper_ascii_y         3.48 ns         3.48 ns    199233248

After:

  BM_ctype_tolower_n                 2.93 ns         2.92 ns    242373703
  BM_ctype_tolower_y                 2.88 ns         2.87 ns    245365309
  BM_ctype_toupper_n                 2.93 ns         2.93 ns    243049353
  BM_ctype_toupper_y                 2.89 ns         2.89 ns    245072521

  BM_wctype_towupper_ascii_n         3.34 ns         3.33 ns    212951912
  BM_wctype_towupper_ascii_y         3.29 ns         3.29 ns    214651254

(Why do both the "y" and "n" variants speed up with the EOR
change? Because the compiler transforms the code so that we
unconditionally do the bit twiddling and then use CSEL to decide whether
or not to actually use the result.)

We also save 1028 bytes of data in the LP64 libc.so.

Test: ran the bionic benchmarks and tests
Change-Id: I7829339f8cb89a58efe539c2a01c51807413aa2d