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

[A64] Assert OPCODE_MEMSET invariants, remove dead general-case path

Browse Source
This commit is contained in:
Herman S.
2026-03-29 15:02:02 +09:00
parent 664b77f38e
commit 2d2105cfd8
1 changed files with 31 additions and 65 deletions
Download Patch File Download Diff File Expand all files Collapse all files

96
src/xenia/cpu/backend/a64/a64_seq_memory.cc
View File

@@ -824,77 +824,43 @@ static const uint64_t zva_length = (4ULL << (xe_cpu_mrs(DCZID_EL0) & 0b0'1111));
struct MEMSET_I64
: Sequence<MEMSET_I64, I<OPCODE_MEMSET, VoidOp, I64Op, I8Op, I64Op>> {
static void Emit(A64Emitter& e, const EmitArgType& i) {
// memset(membase + guest_addr, value, length)
assert_true(i.src2.is_constant);
assert_true(i.src3.is_constant);
assert_true(i.src2.constant() == 0);
// memset(membase + guest_addr, 0, length)
// Only used by dcbz/dcbz128: constant zero value, constant aligned size.
auto addr = ComputeMemoryAddress(e, i.src1);
e.add(e.x0, e.GetMembaseReg(), addr);
// Optimize the common case: zeroing a constant-length block (dcbz/dcbz128).
if (i.src2.is_constant && i.src2.constant() == 0 && i.src3.is_constant) {
const uint64_t len = i.src3.constant();
uint64_t off = 0;
const uint64_t len = i.src3.constant();
uint64_t off = 0;
// Use `dc zva` if it writes more bytes at time than STP
if (zva_enable && len >= zva_length && zva_length > 16) {
for (; off + zva_length <= len; off += zva_length) {
// dc zva, x0
e.sys(0b011, 0b0111, 0b0100, 0b001, e.x0);
if (off + zva_length < len) {
e.add(e.x0, e.x0, zva_length);
}
// Use `dc zva` if it writes more bytes at a time than STP
if (zva_enable && len >= zva_length && zva_length > 16) {
for (; off + zva_length <= len; off += zva_length) {
// dc zva, x0
e.sys(0b011, 0b0111, 0b0100, 0b001, e.x0);
if (off + zva_length < len) {
e.add(e.x0, e.x0, zva_length);
}
}
}
// Inline with STP xzr, xzr pairs (16 bytes each)
for (; off + 16 <= len; off += 16) {
e.stp(e.xzr, e.xzr, AdrPostImm(e.x0, 16));
}
// Handle remaining bytes (0-15)
if (off + 8 <= len) {
e.str(e.xzr, AdrPostImm(e.x0, 8));
off += 8;
}
if (off + 4 <= len) {
e.str(e.wzr, AdrPostImm(e.x0, 4));
off += 4;
}
// Byte loop for any remaining 0-3 bytes
for (; off + 1 <= len; off += 1) {
e.strb(e.wzr, AdrPostImm(e.x0, 1));
}
} else {
// 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 {
e.mov(e.w1, WReg(i.src2.reg().getIdx()));
}
if (i.src3.is_constant) {
e.mov(e.x2, static_cast<uint64_t>(i.src3.constant()));
} else {
e.mov(e.x2, i.src3.reg());
}
auto& done = e.NewCachedLabel();
e.cbz(e.x2, done);
// 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), AdrPostImm(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, AdrPostImm(e.x0, 1));
e.subs(e.x2, e.x2, 1);
e.b(Xbyak_aarch64::NE, byte_loop);
e.L(done);
// Inline with STP xzr, xzr pairs (16 bytes each)
for (; off + 16 <= len; off += 16) {
e.stp(e.xzr, e.xzr, AdrPostImm(e.x0, 16));
}
// Handle remaining bytes (0-15)
if (off + 8 <= len) {
e.str(e.xzr, AdrPostImm(e.x0, 8));
off += 8;
}
if (off + 4 <= len) {
e.str(e.wzr, AdrPostImm(e.x0, 4));
off += 4;
}
// Byte loop for any remaining 0-3 bytes
for (; off + 1 <= len; off += 1) {
e.strb(e.wzr, AdrPostImm(e.x0, 1));
}
}
};