luau-lang/luau
1
0
Fork 0
mirror of https://github.com/luau-lang/luau.git synced 2025-04-03 02:10:53 +01:00
Code Issues Projects Releases Packages Wiki Activity

CodeGen: Use DOT_VEC under a fast flag for lowering vector lib

Browse source 
This is useful for vector.dot, vector.magnitude and vector.normalize.
This commit is contained in:
Arseny Kapoulkine 2024-11-08 10:09:23 -08:00
parent 74a91289e9
commit cd73807c09
1 changed files with 73 additions and 31 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

104
CodeGen/src/IrTranslateBuiltins.cpp
View file

@ -14,6 +14,7 @@ static const int kMinMaxUnrolledParams = 5;
static const int kBit32BinaryOpUnrolledParams = 5;
LUAU_FASTFLAGVARIABLE(LuauVectorLibNativeCodegen);
LUAU_FASTFLAGVARIABLE(LuauVectorLibNativeDot);
namespace Luau
{
@ -907,15 +908,26 @@ static BuiltinImplResult translateBuiltinVectorMagnitude(
build.loadAndCheckTag(arg1, LUA_TVECTOR, build.vmExit(pcpos));
IrOp x = build.inst(IrCmd::LOAD_FLOAT, arg1, build.constInt(0));
IrOp y = build.inst(IrCmd::LOAD_FLOAT, arg1, build.constInt(4));
IrOp z = build.inst(IrCmd::LOAD_FLOAT, arg1, build.constInt(8));
IrOp sum;
IrOp x2 = build.inst(IrCmd::MUL_NUM, x, x);
IrOp y2 = build.inst(IrCmd::MUL_NUM, y, y);
IrOp z2 = build.inst(IrCmd::MUL_NUM, z, z);
if (FFlag::LuauVectorLibNativeDot)
{
IrOp a = build.inst(IrCmd::LOAD_TVALUE, arg1, build.constInt(0));
IrOp sum = build.inst(IrCmd::ADD_NUM, build.inst(IrCmd::ADD_NUM, x2, y2), z2);
sum = build.inst(IrCmd::DOT_VEC, a, a);
}
else
{
IrOp x = build.inst(IrCmd::LOAD_FLOAT, arg1, build.constInt(0));
IrOp y = build.inst(IrCmd::LOAD_FLOAT, arg1, build.constInt(4));
IrOp z = build.inst(IrCmd::LOAD_FLOAT, arg1, build.constInt(8));
IrOp x2 = build.inst(IrCmd::MUL_NUM, x, x);
IrOp y2 = build.inst(IrCmd::MUL_NUM, y, y);
IrOp z2 = build.inst(IrCmd::MUL_NUM, z, z);
sum = build.inst(IrCmd::ADD_NUM, build.inst(IrCmd::ADD_NUM, x2, y2), z2);
}
IrOp mag = build.inst(IrCmd::SQRT_NUM, sum);
@ -945,25 +957,43 @@ static BuiltinImplResult translateBuiltinVectorNormalize(
build.loadAndCheckTag(arg1, LUA_TVECTOR, build.vmExit(pcpos));
IrOp x = build.inst(IrCmd::LOAD_FLOAT, arg1, build.constInt(0));
IrOp y = build.inst(IrCmd::LOAD_FLOAT, arg1, build.constInt(4));
IrOp z = build.inst(IrCmd::LOAD_FLOAT, arg1, build.constInt(8));
if (FFlag::LuauVectorLibNativeDot)
{
IrOp a = build.inst(IrCmd::LOAD_TVALUE, arg1, build.constInt(0));
IrOp sum = build.inst(IrCmd::DOT_VEC, a, a);
IrOp x2 = build.inst(IrCmd::MUL_NUM, x, x);
IrOp y2 = build.inst(IrCmd::MUL_NUM, y, y);
IrOp z2 = build.inst(IrCmd::MUL_NUM, z, z);
IrOp mag = build.inst(IrCmd::SQRT_NUM, sum);
IrOp inv = build.inst(IrCmd::DIV_NUM, build.constDouble(1.0), mag);
IrOp invvec = build.inst(IrCmd::NUM_TO_VEC, inv);
IrOp sum = build.inst(IrCmd::ADD_NUM, build.inst(IrCmd::ADD_NUM, x2, y2), z2);
IrOp result = build.inst(IrCmd::MUL_VEC, a, invvec);
IrOp mag = build.inst(IrCmd::SQRT_NUM, sum);
IrOp inv = build.inst(IrCmd::DIV_NUM, build.constDouble(1.0), mag);
result = build.inst(IrCmd::TAG_VECTOR, result);
IrOp xr = build.inst(IrCmd::MUL_NUM, x, inv);
IrOp yr = build.inst(IrCmd::MUL_NUM, y, inv);
IrOp zr = build.inst(IrCmd::MUL_NUM, z, inv);
build.inst(IrCmd::STORE_TVALUE, build.vmReg(ra), result);
}
else
{
IrOp x = build.inst(IrCmd::LOAD_FLOAT, arg1, build.constInt(0));
IrOp y = build.inst(IrCmd::LOAD_FLOAT, arg1, build.constInt(4));
IrOp z = build.inst(IrCmd::LOAD_FLOAT, arg1, build.constInt(8));
build.inst(IrCmd::STORE_VECTOR, build.vmReg(ra), xr, yr, zr);
build.inst(IrCmd::STORE_TAG, build.vmReg(ra), build.constTag(LUA_TVECTOR));
IrOp x2 = build.inst(IrCmd::MUL_NUM, x, x);
IrOp y2 = build.inst(IrCmd::MUL_NUM, y, y);
IrOp z2 = build.inst(IrCmd::MUL_NUM, z, z);
IrOp sum = build.inst(IrCmd::ADD_NUM, build.inst(IrCmd::ADD_NUM, x2, y2), z2);
IrOp mag = build.inst(IrCmd::SQRT_NUM, sum);
IrOp inv = build.inst(IrCmd::DIV_NUM, build.constDouble(1.0), mag);
IrOp xr = build.inst(IrCmd::MUL_NUM, x, inv);
IrOp yr = build.inst(IrCmd::MUL_NUM, y, inv);
IrOp zr = build.inst(IrCmd::MUL_NUM, z, inv);
build.inst(IrCmd::STORE_VECTOR, build.vmReg(ra), xr, yr, zr);
build.inst(IrCmd::STORE_TAG, build.vmReg(ra), build.constTag(LUA_TVECTOR));
}
return {BuiltinImplType::Full, 1};
}
@ -1019,19 +1049,31 @@ static BuiltinImplResult translateBuiltinVectorDot(IrBuilder& build, int nparams
build.loadAndCheckTag(arg1, LUA_TVECTOR, build.vmExit(pcpos));
build.loadAndCheckTag(args, LUA_TVECTOR, build.vmExit(pcpos));
IrOp x1 = build.inst(IrCmd::LOAD_FLOAT, arg1, build.constInt(0));
IrOp x2 = build.inst(IrCmd::LOAD_FLOAT, args, build.constInt(0));
IrOp xx = build.inst(IrCmd::MUL_NUM, x1, x2);
IrOp sum;
IrOp y1 = build.inst(IrCmd::LOAD_FLOAT, arg1, build.constInt(4));
IrOp y2 = build.inst(IrCmd::LOAD_FLOAT, args, build.constInt(4));
IrOp yy = build.inst(IrCmd::MUL_NUM, y1, y2);
if (FFlag::LuauVectorLibNativeDot)
{
IrOp a = build.inst(IrCmd::LOAD_TVALUE, arg1, build.constInt(0));
IrOp b = build.inst(IrCmd::LOAD_TVALUE, args, build.constInt(0));
IrOp z1 = build.inst(IrCmd::LOAD_FLOAT, arg1, build.constInt(8));
IrOp z2 = build.inst(IrCmd::LOAD_FLOAT, args, build.constInt(8));
IrOp zz = build.inst(IrCmd::MUL_NUM, z1, z2);
sum = build.inst(IrCmd::DOT_VEC, a, b);
}
else
{
IrOp x1 = build.inst(IrCmd::LOAD_FLOAT, arg1, build.constInt(0));
IrOp x2 = build.inst(IrCmd::LOAD_FLOAT, args, build.constInt(0));
IrOp xx = build.inst(IrCmd::MUL_NUM, x1, x2);
IrOp sum = build.inst(IrCmd::ADD_NUM, build.inst(IrCmd::ADD_NUM, xx, yy), zz);
IrOp y1 = build.inst(IrCmd::LOAD_FLOAT, arg1, build.constInt(4));
IrOp y2 = build.inst(IrCmd::LOAD_FLOAT, args, build.constInt(4));
IrOp yy = build.inst(IrCmd::MUL_NUM, y1, y2);
IrOp z1 = build.inst(IrCmd::LOAD_FLOAT, arg1, build.constInt(8));
IrOp z2 = build.inst(IrCmd::LOAD_FLOAT, args, build.constInt(8));
IrOp zz = build.inst(IrCmd::MUL_NUM, z1, z2);
sum = build.inst(IrCmd::ADD_NUM, build.inst(IrCmd::ADD_NUM, xx, yy), zz);
}
build.inst(IrCmd::STORE_DOUBLE, build.vmReg(ra), sum);
build.inst(IrCmd::STORE_TAG, build.vmReg(ra), build.constTag(LUA_TNUMBER));