diff options
| author | Wilco Dijkstra <wilco.dijkstra@arm.com> | 2023-01-24 11:56:38 +0000 |
|---|---|---|
| committer | Szabolcs Nagy <szabolcs.nagy@arm.com> | 2023-01-24 11:56:38 +0000 |
| commit | 92864946def299cab01582672cb7510b2e996101 (patch) | |
| tree | 19b2f01d66a1cfa7cad08e7f19b038aec930ff98 | |
| parent | a7b6022090c9b43af519dc328c1aeece7258e558 (diff) | |
| download | arm-optimized-routines-92864946def299cab01582672cb7510b2e996101.tar.gz | |
string: Improve SVE memcpy
Improve SVE memcpy by copying 2 vectors. This avoids a check on vector length
and improves performance of random memcpy.
| -rw-r--r-- | string/aarch64/memcpy-sve.S | 20 |
1 files changed, 7 insertions, 13 deletions
diff --git a/string/aarch64/memcpy-sve.S b/string/aarch64/memcpy-sve.S index f74d4a9..61d36f2 100644 --- a/string/aarch64/memcpy-sve.S +++ b/string/aarch64/memcpy-sve.S @@ -57,14 +57,16 @@ ENTRY (__memcpy_aarch64_sve) cmp count, 128 b.hi L(copy_long) - cmp count, 32 + cntb vlen + cmp count, vlen, lsl 1 b.hi L(copy32_128) whilelo p0.b, xzr, count - cntb vlen - tbnz vlen, 4, L(vlen128) - ld1b z0.b, p0/z, [src] - st1b z0.b, p0, [dstin] + whilelo p1.b, vlen, count + ld1b z0.b, p0/z, [src, 0, mul vl] + ld1b z1.b, p1/z, [src, 1, mul vl] + st1b z0.b, p0, [dstin, 0, mul vl] + st1b z1.b, p1, [dstin, 1, mul vl] ret /* Medium copies: 33..128 bytes. */ @@ -133,14 +135,6 @@ L(copy64_from_end): stp A_q, B_q, [dstend, -32] ret -L(vlen128): - whilelo p1.b, vlen, count - ld1b z0.b, p0/z, [src, 0, mul vl] - ld1b z1.b, p1/z, [src, 1, mul vl] - st1b z0.b, p0, [dstin, 0, mul vl] - st1b z1.b, p1, [dstin, 1, mul vl] - ret - /* Large backwards copy for overlapping copies. Copy 16 bytes and then align srcend to 16-byte alignment. */ L(copy_long_backwards): |