CodeGen: Implement support for math.lerp lowering (#1609)
Some checks are pending
benchmark / callgrind (map[branch:main name:luau-lang/benchmark-data], ubuntu-22.04) (push) Waiting to run
build / macos (push) Waiting to run
build / macos-arm (push) Waiting to run
build / ubuntu (push) Waiting to run
build / windows (Win32) (push) Waiting to run
build / windows (x64) (push) Waiting to run
build / coverage (push) Waiting to run
build / web (push) Waiting to run
release / macos (push) Waiting to run
release / ubuntu (push) Waiting to run
release / windows (push) Waiting to run
release / web (push) Waiting to run

To implement math.lerp without branches, we add SELECT_NUM which
selects one of the two inputs based on the comparison condition.

For simplicity, we only support C == D for now; this can be extended to
a more generic version with a IrCondition operand E, but that requires
more work on the SSE side (to flip the comparison for some conditions
like Greater, and expose more generic vcmpsd).

Note: On AArch64 this will effectively result in a change in floating
point
behavior between native code and non-native code: clang synthesizes
fmadd (because floating point contraction is allowed by default, and the
arch always has the instruction), whereas this change will use
fmul+fadd.

I am not sure if this is good or bad, and if this is a problem in C or
not.
Specifically, clang's behavior results in different results between X64
and AArch64 when *not* using codegen, and with this change the behavior
when using codegen is... the same? :)

Fixing this will require either using LERP_NUM instead and hand-coding
lowering, or exposing some sort of "quasi" MADD_NUM (which would
lower to fma on AArch64 and mul+add on X64).

A small benefit to the current approach is `lerp(1, 5, t)`
constant-folds the
subtraction. With LERP_NUM this optimization will need to be implemented
manually as a partial constant-folding for LERP_NUM.

A similar problem exists today for vector.cross & vector.dot. So maybe
this
is not something we need to fix, unsure.
This commit is contained in:
Arseny Kapoulkine 2025-01-16 10:48:27 -08:00 committed by GitHub
parent c759cd5581
commit 24cacc94ed
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
12 changed files with 108 additions and 0 deletions

View file

@ -160,6 +160,7 @@ public:
void vmaxsd(OperandX64 dst, OperandX64 src1, OperandX64 src2);
void vminsd(OperandX64 dst, OperandX64 src1, OperandX64 src2);
void vcmpeqsd(OperandX64 dst, OperandX64 src1, OperandX64 src2);
void vcmpltsd(OperandX64 dst, OperandX64 src1, OperandX64 src2);
void vblendvpd(RegisterX64 dst, RegisterX64 src1, OperandX64 mask, RegisterX64 src3);

View file

@ -185,6 +185,11 @@ enum class IrCmd : uint8_t
// A: double
SIGN_NUM,
// Select B if C == D, otherwise select A
// A, B: double (endpoints)
// C, D: double (condition arguments)
SELECT_NUM,
// Add/Sub/Mul/Div/Idiv two vectors
// A, B: TValue
ADD_VEC,

View file

@ -174,6 +174,7 @@ inline bool hasResult(IrCmd cmd)
case IrCmd::SQRT_NUM:
case IrCmd::ABS_NUM:
case IrCmd::SIGN_NUM:
case IrCmd::SELECT_NUM:
case IrCmd::ADD_VEC:
case IrCmd::SUB_VEC:
case IrCmd::MUL_VEC:

View file

@ -927,6 +927,11 @@ void AssemblyBuilderX64::vminsd(OperandX64 dst, OperandX64 src1, OperandX64 src2
placeAvx("vminsd", dst, src1, src2, 0x5d, false, AVX_0F, AVX_F2);
}
void AssemblyBuilderX64::vcmpeqsd(OperandX64 dst, OperandX64 src1, OperandX64 src2)
{
placeAvx("vcmpeqsd", dst, src1, src2, 0x00, 0xc2, false, AVX_0F, AVX_F2);
}
void AssemblyBuilderX64::vcmpltsd(OperandX64 dst, OperandX64 src1, OperandX64 src2)
{
placeAvx("vcmpltsd", dst, src1, src2, 0x01, 0xc2, false, AVX_0F, AVX_F2);

View file

@ -169,6 +169,8 @@ const char* getCmdName(IrCmd cmd)
return "ABS_NUM";
case IrCmd::SIGN_NUM:
return "SIGN_NUM";
case IrCmd::SELECT_NUM:
return "SELECT_NUM";
case IrCmd::ADD_VEC:
return "ADD_VEC";
case IrCmd::SUB_VEC:

View file

@ -13,6 +13,7 @@
LUAU_FASTFLAG(LuauVectorLibNativeDot)
LUAU_FASTFLAG(LuauCodeGenVectorDeadStoreElim)
LUAU_FASTFLAG(LuauCodeGenLerp)
namespace Luau
{
@ -703,6 +704,20 @@ void IrLoweringA64::lowerInst(IrInst& inst, uint32_t index, const IrBlock& next)
build.fcsel(inst.regA64, temp1, inst.regA64, getConditionFP(IrCondition::Less));
break;
}
case IrCmd::SELECT_NUM:
{
LUAU_ASSERT(FFlag::LuauCodeGenLerp);
inst.regA64 = regs.allocReuse(KindA64::d, index, {inst.a, inst.b, inst.c, inst.d});
RegisterA64 temp1 = tempDouble(inst.a);
RegisterA64 temp2 = tempDouble(inst.b);
RegisterA64 temp3 = tempDouble(inst.c);
RegisterA64 temp4 = tempDouble(inst.d);
build.fcmp(temp3, temp4);
build.fcsel(inst.regA64, temp2, temp1, getConditionFP(IrCondition::Equal));
break;
}
case IrCmd::ADD_VEC:
{
inst.regA64 = regs.allocReuse(KindA64::q, index, {inst.a, inst.b});

View file

@ -17,6 +17,7 @@
LUAU_FASTFLAG(LuauVectorLibNativeDot)
LUAU_FASTFLAG(LuauCodeGenVectorDeadStoreElim)
LUAU_FASTFLAG(LuauCodeGenLerp)
namespace Luau
{
@ -622,6 +623,30 @@ void IrLoweringX64::lowerInst(IrInst& inst, uint32_t index, const IrBlock& next)
build.vblendvpd(inst.regX64, tmp1.reg, build.f64x2(1, 1), inst.regX64);
break;
}
case IrCmd::SELECT_NUM:
{
LUAU_ASSERT(FFlag::LuauCodeGenLerp);
inst.regX64 = regs.allocRegOrReuse(SizeX64::xmmword, index, {inst.a, inst.c, inst.d}); // can't reuse b if a is a memory operand
ScopedRegX64 tmp{regs, SizeX64::xmmword};
if (inst.c.kind == IrOpKind::Inst)
build.vcmpeqsd(tmp.reg, regOp(inst.c), memRegDoubleOp(inst.d));
else
{
build.vmovsd(tmp.reg, memRegDoubleOp(inst.c));
build.vcmpeqsd(tmp.reg, tmp.reg, memRegDoubleOp(inst.d));
}
if (inst.a.kind == IrOpKind::Inst)
build.vblendvpd(inst.regX64, regOp(inst.a), memRegDoubleOp(inst.b), tmp.reg);
else
{
build.vmovsd(inst.regX64, memRegDoubleOp(inst.a));
build.vblendvpd(inst.regX64, inst.regX64, memRegDoubleOp(inst.b), tmp.reg);
}
break;
}
case IrCmd::ADD_VEC:
{
inst.regX64 = regs.allocRegOrReuse(SizeX64::xmmword, index, {inst.a, inst.b});

View file

@ -15,6 +15,7 @@ static const int kBit32BinaryOpUnrolledParams = 5;
LUAU_FASTFLAGVARIABLE(LuauVectorLibNativeCodegen);
LUAU_FASTFLAGVARIABLE(LuauVectorLibNativeDot);
LUAU_FASTFLAGVARIABLE(LuauCodeGenLerp);
namespace Luau
{
@ -284,6 +285,42 @@ static BuiltinImplResult translateBuiltinMathClamp(
return {BuiltinImplType::UsesFallback, 1};
}
static BuiltinImplResult translateBuiltinMathLerp(
IrBuilder& build,
int nparams,
int ra,
int arg,
IrOp args,
IrOp arg3,
int nresults,
IrOp fallback,
int pcpos
)
{
LUAU_ASSERT(FFlag::LuauCodeGenLerp);
if (nparams < 3 || nresults > 1)
return {BuiltinImplType::None, -1};
builtinCheckDouble(build, build.vmReg(arg), pcpos);
builtinCheckDouble(build, args, pcpos);
builtinCheckDouble(build, arg3, pcpos);
IrOp a = builtinLoadDouble(build, build.vmReg(arg));
IrOp b = builtinLoadDouble(build, args);
IrOp t = builtinLoadDouble(build, arg3);
IrOp l = build.inst(IrCmd::ADD_NUM, a, build.inst(IrCmd::MUL_NUM, build.inst(IrCmd::SUB_NUM, b, a), t));
IrOp r = build.inst(IrCmd::SELECT_NUM, l, b, t, build.constDouble(1.0)); // select on t==1.0
build.inst(IrCmd::STORE_DOUBLE, build.vmReg(ra), r);
if (ra != arg)
build.inst(IrCmd::STORE_TAG, build.vmReg(ra), build.constTag(LUA_TNUMBER));
return {BuiltinImplType::Full, 1};
}
static BuiltinImplResult translateBuiltinMathUnary(IrBuilder& build, IrCmd cmd, int nparams, int ra, int arg, int nresults, int pcpos)
{
if (nparams < 1 || nresults > 1)
@ -1387,6 +1424,8 @@ BuiltinImplResult translateBuiltin(
case LBF_VECTOR_MAX:
return FFlag::LuauVectorLibNativeCodegen ? translateBuiltinVectorMap2(build, IrCmd::MAX_NUM, nparams, ra, arg, args, arg3, nresults, pcpos)
: noneResult;
case LBF_MATH_LERP:
return FFlag::LuauCodeGenLerp ? translateBuiltinMathLerp(build, nparams, ra, arg, args, arg3, nresults, fallback, pcpos) : noneResult;
default:
return {BuiltinImplType::None, -1};
}

View file

@ -13,6 +13,7 @@
#include <math.h>
LUAU_FASTFLAG(LuauVectorLibNativeDot);
LUAU_FASTFLAG(LuauCodeGenLerp);
namespace Luau
{
@ -70,6 +71,7 @@ IrValueKind getCmdValueKind(IrCmd cmd)
case IrCmd::SQRT_NUM:
case IrCmd::ABS_NUM:
case IrCmd::SIGN_NUM:
case IrCmd::SELECT_NUM:
return IrValueKind::Double;
case IrCmd::ADD_VEC:
case IrCmd::SUB_VEC:
@ -656,6 +658,16 @@ void foldConstants(IrBuilder& build, IrFunction& function, IrBlock& block, uint3
substitute(function, inst, build.constDouble(v > 0.0 ? 1.0 : v < 0.0 ? -1.0 : 0.0));
}
break;
case IrCmd::SELECT_NUM:
LUAU_ASSERT(FFlag::LuauCodeGenLerp);
if (inst.c.kind == IrOpKind::Constant && inst.d.kind == IrOpKind::Constant)
{
double c = function.doubleOp(inst.c);
double d = function.doubleOp(inst.d);
substitute(function, inst, c == d ? inst.b : inst.a);
}
break;
case IrCmd::NOT_ANY:
if (inst.a.kind == IrOpKind::Constant)
{

View file

@ -1382,6 +1382,7 @@ static void constPropInInst(ConstPropState& state, IrBuilder& build, IrFunction&
case IrCmd::SQRT_NUM:
case IrCmd::ABS_NUM:
case IrCmd::SIGN_NUM:
case IrCmd::SELECT_NUM:
case IrCmd::NOT_ANY:
state.substituteOrRecord(inst, index);
break;

View file

@ -506,6 +506,7 @@ TEST_CASE_FIXTURE(AssemblyBuilderX64Fixture, "AVXBinaryInstructionForms")
SINGLE_COMPARE(vmaxsd(xmm8, xmm10, xmm14), 0xc4, 0x41, 0x2b, 0x5f, 0xc6);
SINGLE_COMPARE(vminsd(xmm8, xmm10, xmm14), 0xc4, 0x41, 0x2b, 0x5d, 0xc6);
SINGLE_COMPARE(vcmpeqsd(xmm8, xmm10, xmm14), 0xc4, 0x41, 0x2b, 0xc2, 0xc6, 0x00);
SINGLE_COMPARE(vcmpltsd(xmm8, xmm10, xmm14), 0xc4, 0x41, 0x2b, 0xc2, 0xc6, 0x01);
}

View file

@ -408,6 +408,7 @@ assert(math.lerp(1, 5, 1) == 5)
assert(math.lerp(1, 5, 0.5) == 3)
assert(math.lerp(1, 5, 1.5) == 7)
assert(math.lerp(1, 5, -0.5) == -1)
assert(math.lerp(1, 5, noinline(0.5)) == 3)
-- lerp properties
local sq2, sq3 = math.sqrt(2), math.sqrt(3)