fabi/Xenia-Canary
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[A64] Use 16-byte NEON stores in MEMSET fallback loop

Browse Source 
Splat the fill byte into v0.16b and use str q0 (16 bytes/iter) for
the bulk, falling back to a byte loop for the 0-15 byte tail.
This commit is contained in:
Herman S.
2026-03-24 15:33:57 +09:00
parent bdf64db79e
commit a1e333223e
1 changed files with 20 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

25
src/xenia/cpu/backend/a64/a64_seq_memory.cc
View File

@@ -839,7 +839,8 @@ struct MEMSET_I64
}
// 1-3 byte remainder unlikely for dcbz/dcbz128, skip for now.
} else {
// General case: use a loop.
// General case: splat byte to NEON register, then 16-byte bulk loop
// with a byte loop for the 0-15 byte tail.
if (i.src2.is_constant) {
e.mov(e.w1, static_cast<uint64_t>(i.src2.constant() & 0xFF));
} else {
@@ -850,15 +851,29 @@ struct MEMSET_I64
} else {
e.mov(e.x2, i.src3.reg());
}
// Byte-fill loop: store w1 to [x0] for x2 bytes.
auto& loop = e.NewCachedLabel();
auto& done = e.NewCachedLabel();
e.cbz(e.x2, done);
e.L(loop);
// Splat fill byte across v0 for 16-byte stores.
e.dup(VReg(0).b16, e.w1);
// 16-byte bulk loop.
auto& loop16 = e.NewCachedLabel();
auto& tail = e.NewCachedLabel();
e.L(loop16);
e.cmp(e.x2, 16);
e.b(LO, tail);
e.str(QReg(0), ptr(e.x0));
e.add(e.x0, e.x0, 16);
e.sub(e.x2, e.x2, 16);
e.b(loop16);
// Byte loop for remaining 0-15 bytes.
e.L(tail);
e.cbz(e.x2, done);
auto& byte_loop = e.NewCachedLabel();
e.L(byte_loop);
e.strb(e.w1, ptr(e.x0));
e.add(e.x0, e.x0, 1);
e.subs(e.x2, e.x2, 1);
e.b(Xbyak_aarch64::NE, loop);
e.b(Xbyak_aarch64::NE, byte_loop);
e.L(done);
}
}