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

[a64] Optimize EmitD3DCOLOR constants

Browse Source 
`3.0` works with `fmov`
This commit is contained in:
Wunkolo
2026-03-23 19:34:22 -07:00
committed by Heel
parent 531c49fbdc
commit cfc1e4d069
1 changed files with 3 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
src/xenia/cpu/backend/a64/a64_seq_vector.cc
View File

@@ -1434,8 +1434,7 @@ struct PACK : Sequence<PACK, I<OPCODE_PACK, V128Op, V128Op, V128Op>> {
int d = i.dest.reg().getIdx(); int d = i.dest.reg().getIdx();
// Clamp to [3.0f, 3.0f + 255*2^-22]. // Clamp to [3.0f, 3.0f + 255*2^-22].
// fmaxnm/fminnm: NaN operand returns the non-NaN value (pack NaN as zero). // fmaxnm/fminnm: NaN operand returns the non-NaN value (pack NaN as zero).
e.mov(e.w0, 0x40400000u); // 3.0f e.fmov(VReg(0).s4, 3.0f);
e.dup(VReg(0).s4, e.w0);
e.fmaxnm(VReg(d).s4, VReg(s).s4, VReg(0).s4); e.fmaxnm(VReg(d).s4, VReg(s).s4, VReg(0).s4);
e.mov(e.w0, 0x404000FFu); // 3.0f + 255*2^-22 e.mov(e.w0, 0x404000FFu); // 3.0f + 255*2^-22
e.dup(VReg(0).s4, e.w0); e.dup(VReg(0).s4, e.w0);
@@ -1666,8 +1665,7 @@ struct UNPACK : Sequence<UNPACK, I<OPCODE_UNPACK, V128Op, V128Op>> {
int d = i.dest.reg().getIdx(); int d = i.dest.reg().getIdx();
if (i.src1.is_constant && i.src1.value->IsConstantZero()) { if (i.src1.is_constant && i.src1.value->IsConstantZero()) {
// Zero -> 1.0f in all lanes. // Zero -> 1.0f in all lanes.
e.mov(e.w0, 0x3F800000u); e.fmov(VReg(d).s4, 1.0f);
e.dup(VReg(d).s4, e.w0);
return; return;
} }
// TBL: extract bytes from packed D3DCOLOR -> one byte per lane. // TBL: extract bytes from packed D3DCOLOR -> one byte per lane.
@@ -1680,8 +1678,7 @@ struct UNPACK : Sequence<UNPACK, I<OPCODE_UNPACK, V128Op, V128Op>> {
LoadV128Const(e, 1, ctrl); LoadV128Const(e, 1, ctrl);
e.tbl(VReg(d).b16, VReg(s).b16, 1, VReg(1).b16); e.tbl(VReg(d).b16, VReg(s).b16, 1, VReg(1).b16);
// OR with 1.0f (0x3F800000) to form the magic float. // OR with 1.0f (0x3F800000) to form the magic float.
e.mov(e.w0, 0x3F800000u); e.fmov(VReg(0).s4, 1.0f);
e.dup(VReg(0).s4, e.w0);
e.orr(VReg(d).b16, VReg(d).b16, VReg(0).b16); e.orr(VReg(d).b16, VReg(d).b16, VReg(0).b16);
} }
static void EmitCallHelper(A64Emitter& e, const EmitArgType& i, void* fn) { static void EmitCallHelper(A64Emitter& e, const EmitArgType& i, void* fn) {