whytofly
/
qt6-bb10
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Aarch64: vectorize findChar 

The "algorithm" is the same as the one for x86.

Also added a comment to qt_from_latin1 to indicate that manual
vectorization doesn't bring anything useful.

Change-Id: I6130cbd83c14c22b1bd15d726b26dbc83068b1a6
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
bb10
Erik Verbruggen 2016-12-06 11:55:31 +01:00
parent 59b80d606e
commit 311e523cef
1 changed files with 14 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
src/corelib/tools/qstring.cpp
View File

@ -211,7 +211,8 @@ void qt_from_latin1(ushort *dst, const char *str, size_t size) Q_DECL_NOTHROW
{
/* SIMD:
* Unpacking with SSE has been shown to improve performance on recent CPUs
* The same method gives no improvement with NEON.
* The same method gives no improvement with NEON. On Aarch64, clang will do the vectorization
* itself in exactly the same way as one would do it with intrinsics.
*/
#if defined(__SSE2__)
const char *e = str + size;
@ -727,6 +728,18 @@ static int findChar(const QChar *str, int len, QChar ch, int from,
[=](int i) { return n - s + i; });
# endif
#endif
#if defined(__ARM_NEON__) && defined(Q_PROCESSOR_ARM_64) // vaddv is only available on Aarch64
const uint16x8_t vmask = { 1, 1 << 1, 1 << 2, 1 << 3, 1 << 4, 1 << 5, 1 << 6, 1 << 7 };
const uint16x8_t ch_vec = vdupq_n_u16(c);
for (const ushort *next = n + 8; next <= e; n = next, next += 8) {
uint16x8_t data = vld1q_u16(n);
uint mask = vaddvq_u16(vandq_u16(vceqq_u16(data, ch_vec), vmask));
if (ushort(mask)) {
// found a match
return n - s + qCountTrailingZeroBits(mask);
}
}
#endif // aarch64
--n;
while (++n != e)
if (*n == c)