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

[a64] Add opportunistic LoadV128Const splat optimizations

Browse Source 
Try to detect opportunities to use the `movi` instructions for generating 128-bit constants rather than writing and loading scratch memory.
This commit is contained in:
Wunkolo
2026-03-23 18:34:59 -07:00
committed by Heel
parent aff3999876
commit 33aa579b0c
1 changed files with 77 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

77
src/xenia/cpu/backend/a64/a64_seq_util.h
View File

@@ -46,6 +46,83 @@ inline void EmitWithVmxFpcr(A64Emitter& e, Fn&& emit_op) {
// Load a compile-time vec128_t constant into a NEON register.
inline void LoadV128Const(A64Emitter& e, int vreg_idx, const vec128_t& val) {
if (!val.low && !val.high) {
// 0000...
e.movi(VReg2D(vreg_idx), 0);
return;
} else if (val.low == ~uint64_t(0) && val.high == ~uint64_t(0)) {
// 1111...
e.movi(VReg2D(vreg_idx), ~0ULL);
return;
} else {
bool all_equal_u8 = true;
const uint8_t splat_u8 = val.u8[0];
for (unsigned i = 1; i < 16; ++i) {
if (val.u8[i] != splat_u8) {
all_equal_u8 = false;
break;
}
}
if (all_equal_u8) {
e.movi(VReg(vreg_idx).b16, static_cast<uint8_t>(splat_u8));
return;
}
bool all_equal_u16 = true;
const uint16_t splat_u16 = val.u16[0];
for (unsigned i = 1; i < 8; ++i) {
if (val.u16[i] != splat_u16) {
all_equal_u16 = false;
break;
}
}
if (all_equal_u16) {
if ((splat_u16 & 0xFF'00) == 0) {
e.movi(VReg(vreg_idx).h8, static_cast<uint8_t>(splat_u16));
} else if ((splat_u16 & 0x00'FF) == 0) {
e.movi(VReg(vreg_idx).h8, static_cast<uint8_t>(splat_u16 >> 8), LSL, 8);
} else {
e.movz(e.w0, splat_u16);
e.dup(VReg(vreg_idx).h8, e.w0);
}
return;
}
bool all_equal_u32 = true;
const uint32_t splat_u32 = val.u32[0];
for (unsigned i = 1; i < 4; ++i) {
if (val.u32[i] != splat_u32) {
all_equal_u32 = false;
break;
}
}
if (all_equal_u32) {
if ((splat_u32 & 0xFF'FF'FF'00) == 0) {
e.movi(VReg(vreg_idx).s4, static_cast<uint8_t>(splat_u32));
} else if ((splat_u32 & 0xFF'FF'00'FF) == 0) {
e.movi(VReg(vreg_idx).s4, static_cast<uint8_t>(splat_u32 >> 8), LSL, 8);
} else if ((splat_u32 & 0xFF'00'FF'FF) == 0) {
e.movi(VReg(vreg_idx).s4, static_cast<uint8_t>(splat_u32 >> 16), LSL,
16);
} else if ((splat_u32 & 0x00'FF'FF'FF) == 0) {
e.movi(VReg(vreg_idx).s4, static_cast<uint8_t>(splat_u32 >> 24), LSL,
24);
} else {
e.mov(e.w0, splat_u32);
e.dup(VReg(vreg_idx).s4, e.w0);
}
return;
}
const bool all_equal_u64 = val.low == val.high;
const uint64_t splat_u64 = val.u64[0];
if (all_equal_u64) {
e.mov(e.x0, splat_u64);
e.dup(VReg(vreg_idx).d2, e.x0);
return;
}
}
e.mov(e.x0, val.low);
e.mov(e.x1, val.high);
e.stp(e.x0, e.x1,