ASCII fastpath for towupper and towlower.
This change elides unnecessary calls to __find_icu_symbol
for ASCII chars and improves PUBG mobile game loading time by 7% on
hikey620(Cortex-A53):
name old time/op new time/op delta
PUBG_0.13.0_Launch 41.5s ± 2% 37.7s ± 3% -9.24% (p=0.008 n=5+5)
Below are the bionic benchmark results on a Pixel 2 XL for 64-bit,
showing a large speedup for ASCII and only a small slowdown for non-ASCII.
Before:
BM_wctype_towlower_ascii_n 10.5 ns 10.4 ns 61973065
BM_wctype_towlower_ascii_y 10.2 ns 10.2 ns 70158659
BM_wctype_towlower_unicode_n 10.3 ns 10.3 ns 67719478
BM_wctype_towlower_unicode_y 10.6 ns 10.5 ns 67841545
BM_wctype_towupper_ascii_n 10.8 ns 10.8 ns 63456778
BM_wctype_towupper_ascii_y 10.9 ns 10.9 ns 65116910
BM_wctype_towupper_unicode_n 10.7 ns 10.7 ns 67463276
BM_wctype_towupper_unicode_y 10.4 ns 10.4 ns 66467890
After:
BM_wctype_towlower_ascii_n 3.35 ns 3.34 ns 205567652
BM_wctype_towlower_ascii_y 3.30 ns 3.29 ns 214108746
BM_wctype_towlower_unicode_n 10.9 ns 10.8 ns 65007743
BM_wctype_towlower_unicode_y 10.6 ns 10.6 ns 63819060
BM_wctype_towupper_ascii_n 3.53 ns 3.53 ns 195944444
BM_wctype_towupper_ascii_y 3.48 ns 3.48 ns 199233248
BM_wctype_towupper_unicode_n 11.1 ns 11.1 ns 62760216
BM_wctype_towupper_unicode_y 11.0 ns 11.0 ns 61608872
Test: bionic unit tests on device
Test: bionic benchmarks on device
Signed-off-by: Balaram Makam <b.makam@samsung.com>
Change-Id: I77ab7efb66d7bcb35d00467663607535e5c1992f
diff --git a/libc/bionic/wctype.cpp b/libc/bionic/wctype.cpp
index 6e33b6c..061f55a 100644
--- a/libc/bionic/wctype.cpp
+++ b/libc/bionic/wctype.cpp
@@ -117,12 +117,22 @@
}
wint_t towlower(wint_t wc) {
+ if (wc < 0x80) {
+ if (wc >= 'A' && wc <= 'Z') return wc | 0x20;
+ return wc;
+ }
+
typedef UChar32 (*FnT)(UChar32);
static auto u_tolower = reinterpret_cast<FnT>(__find_icu_symbol("u_tolower"));
return u_tolower ? u_tolower(wc) : tolower(wc);
}
wint_t towupper(wint_t wc) {
+ if (wc < 0x80) {
+ if (wc >= 'a' && wc <= 'z') return wc & 0xdf;
+ return wc;
+ }
+
typedef UChar32 (*FnT)(UChar32);
static auto u_toupper = reinterpret_cast<FnT>(__find_icu_symbol("u_toupper"));
return u_toupper ? u_toupper(wc) : toupper(wc);
