// Copyright 2018 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.

// Lowering arithmetic
(Add(64|32|16|8|Ptr) x y) -> (I64Add x y)
(Add(64|32)F x y) -> (F64Add x y)

(Sub(64|32|16|8|Ptr) x y) -> (I64Sub x y)
(Sub(64|32)F x y) -> (F64Sub x y)

(Mul(64|32|16|8) x y) -> (I64Mul x y)
(Mul(64|32)F x y) -> (F64Mul x y)

(Div64  x y) -> (I64DivS x y)
(Div64u x y) -> (I64DivU x y)
(Div32  x y) -> (I64DivS (SignExt32to64 x) (SignExt32to64 y))
(Div32u x y) -> (I64DivU (ZeroExt32to64 x) (ZeroExt32to64 y))
(Div16  x y) -> (I64DivS (SignExt16to64 x) (SignExt16to64 y))
(Div16u x y) -> (I64DivU (ZeroExt16to64 x) (ZeroExt16to64 y))
(Div8   x y) -> (I64DivS (SignExt8to64 x) (SignExt8to64 y))
(Div8u  x y) -> (I64DivU (ZeroExt8to64 x) (ZeroExt8to64 y))
(Div(64|32)F x y) -> (F64Div x y)

(Mod64  x y) -> (I64RemS x y)
(Mod64u x y) -> (I64RemU x y)
(Mod32  x y) -> (I64RemS (SignExt32to64 x) (SignExt32to64 y))
(Mod32u x y) -> (I64RemU (ZeroExt32to64 x) (ZeroExt32to64 y))
(Mod16  x y) -> (I64RemS (SignExt16to64 x) (SignExt16to64 y))
(Mod16u x y) -> (I64RemU (ZeroExt16to64 x) (ZeroExt16to64 y))
(Mod8   x y) -> (I64RemS (SignExt8to64  x) (SignExt8to64  y))
(Mod8u  x y) -> (I64RemU (ZeroExt8to64  x) (ZeroExt8to64  y))

(And(64|32|16|8|B) x y) -> (I64And x y)

(Or(64|32|16|8|B) x y) -> (I64Or x y)

(Xor(64|32|16|8) x y) -> (I64Xor x y)

(Neg(64|32|16|8) x) -> (I64Sub (I64Const [0]) x)
(Neg32F x) -> (F64Neg x)
(Neg64F x) -> (F64Neg x)

(Com(64|32|16|8) x) -> (I64Xor x (I64Const [-1]))

(Not x) -> (I64Eqz x)

// Lowering pointer arithmetic
(OffPtr [off] ptr) -> (I64AddConst [off] ptr)

// Lowering extension
// It is unnecessary to extend loads
(SignExt32to64        x:(I64Load32S _ _)) -> x
(SignExt16to(64|32)   x:(I64Load16S _ _)) -> x
(SignExt8to(64|32|16) x:(I64Load8S  _ _)) -> x
(ZeroExt32to64        x:(I64Load32U _ _)) -> x
(ZeroExt16to(64|32)   x:(I64Load16U _ _)) -> x
(ZeroExt8to(64|32|16) x:(I64Load8U  _ _)) -> x
(SignExt32to64        x) && objabi.GOWASM.SignExt -> (I64Extend32S x)
(SignExt8to(64|32|16) x) && objabi.GOWASM.SignExt -> (I64Extend8S x)
(SignExt16to(64|32)   x) && objabi.GOWASM.SignExt -> (I64Extend16S x)
(SignExt32to64        x) -> (I64ShrS (I64Shl x (I64Const [32])) (I64Const [32]))
(SignExt16to(64|32)   x) -> (I64ShrS (I64Shl x (I64Const [48])) (I64Const [48]))
(SignExt8to(64|32|16) x) -> (I64ShrS (I64Shl x (I64Const [56])) (I64Const [56]))
(ZeroExt32to64        x) -> (I64And x (I64Const [0xffffffff]))
(ZeroExt16to(64|32)   x) -> (I64And x (I64Const [0xffff]))
(ZeroExt8to(64|32|16) x) -> (I64And x (I64Const [0xff]))

(Slicemask x) -> (I64ShrS (I64Sub (I64Const [0]) x) (I64Const [63]))

// Lowering truncation
// Because we ignore the high parts, truncates are just copies.
(Trunc64to(32|16|8) x) -> x
(Trunc32to(16|8)    x) -> x
(Trunc16to8         x) -> x

// Lowering float <-> int
(Cvt32to32F x) -> (LoweredRound32F (F64ConvertI64S (SignExt32to64 x)))
(Cvt32to64F x) -> (F64ConvertI64S (SignExt32to64 x))
(Cvt64to32F x) -> (LoweredRound32F (F64ConvertI64S x))
(Cvt64to64F x) -> (F64ConvertI64S x)
(Cvt32Uto32F x) -> (LoweredRound32F (F64ConvertI64U (ZeroExt32to64 x)))
(Cvt32Uto64F x) -> (F64ConvertI64U (ZeroExt32to64 x))
(Cvt64Uto32F x) -> (LoweredRound32F (F64ConvertI64U x))
(Cvt64Uto64F x) -> (F64ConvertI64U x)

(Cvt32Fto32 x) -> (I64TruncSatF64S x)
(Cvt32Fto64 x) -> (I64TruncSatF64S x)
(Cvt64Fto32 x) -> (I64TruncSatF64S x)
(Cvt64Fto64 x) -> (I64TruncSatF64S x)
(Cvt32Fto32U x) -> (I64TruncSatF64U x)
(Cvt32Fto64U x) -> (I64TruncSatF64U x)
(Cvt64Fto32U x) -> (I64TruncSatF64U x)
(Cvt64Fto64U x) -> (I64TruncSatF64U x)

(Cvt32Fto64F x) -> x
(Cvt64Fto32F x) -> (LoweredRound32F x)

(Round32F x) -> (LoweredRound32F x)
(Round64F x) -> x

// Lowering shifts
// Unsigned shifts need to return 0 if shift amount is >= width of shifted value.

(Lsh64x64 x y) -> (Select (I64Shl x y) (I64Const [0]) (I64LtU y (I64Const [64])))
(Lsh64x32 x y) -> (Lsh64x64 x (ZeroExt32to64 y))
(Lsh64x16 x y) -> (Lsh64x64 x (ZeroExt16to64 y))
(Lsh64x8  x y) -> (Lsh64x64 x (ZeroExt8to64  y))

(Lsh32x64 x y) -> (Lsh64x64 x y)
(Lsh32x32 x y) -> (Lsh64x64 x (ZeroExt32to64 y))
(Lsh32x16 x y) -> (Lsh64x64 x (ZeroExt16to64 y))
(Lsh32x8  x y) -> (Lsh64x64 x (ZeroExt8to64  y))

(Lsh16x64 x y) -> (Lsh64x64 x y)
(Lsh16x32 x y) -> (Lsh64x64 x (ZeroExt32to64 y))
(Lsh16x16 x y) -> (Lsh64x64 x (ZeroExt16to64 y))
(Lsh16x8  x y) -> (Lsh64x64 x (ZeroExt8to64  y))

(Lsh8x64  x y) -> (Lsh64x64 x y)
(Lsh8x32  x y) -> (Lsh64x64 x (ZeroExt32to64 y))
(Lsh8x16  x y) -> (Lsh64x64 x (ZeroExt16to64 y))
(Lsh8x8   x y) -> (Lsh64x64 x (ZeroExt8to64  y))

(Rsh64Ux64 x y) -> (Select (I64ShrU x y) (I64Const [0]) (I64LtU y (I64Const [64])))
(Rsh64Ux32 x y) -> (Rsh64Ux64 x (ZeroExt32to64 y))
(Rsh64Ux16 x y) -> (Rsh64Ux64 x (ZeroExt16to64 y))
(Rsh64Ux8  x y) -> (Rsh64Ux64 x (ZeroExt8to64  y))

(Rsh32Ux64 x y) -> (Rsh64Ux64 (ZeroExt32to64 x) y)
(Rsh32Ux32 x y) -> (Rsh64Ux64 (ZeroExt32to64 x) (ZeroExt32to64 y))
(Rsh32Ux16 x y) -> (Rsh64Ux64 (ZeroExt32to64 x) (ZeroExt16to64 y))
(Rsh32Ux8  x y) -> (Rsh64Ux64 (ZeroExt32to64 x) (ZeroExt8to64  y))

(Rsh16Ux64 x y) -> (Rsh64Ux64 (ZeroExt16to64 x) y)
(Rsh16Ux32 x y) -> (Rsh64Ux64 (ZeroExt16to64 x) (ZeroExt32to64 y))
(Rsh16Ux16 x y) -> (Rsh64Ux64 (ZeroExt16to64 x) (ZeroExt16to64 y))
(Rsh16Ux8  x y) -> (Rsh64Ux64 (ZeroExt16to64 x) (ZeroExt8to64  y))

(Rsh8Ux64  x y) -> (Rsh64Ux64 (ZeroExt8to64 x) y)
(Rsh8Ux32  x y) -> (Rsh64Ux64 (ZeroExt8to64 x) (ZeroExt32to64 y))
(Rsh8Ux16  x y) -> (Rsh64Ux64 (ZeroExt8to64 x) (ZeroExt16to64 y))
(Rsh8Ux8   x y) -> (Rsh64Ux64 (ZeroExt8to64 x) (ZeroExt8to64  y))

// Signed right shift needs to return 0/-1 if shift amount is >= width of shifted value.
// We implement this by setting the shift value to (width - 1) if the shift value is >= width.

(Rsh64x64 x y) -> (I64ShrS x (Select <typ.Int64> y (I64Const [63]) (I64LtU y (I64Const [64]))))
(Rsh64x32 x y) -> (Rsh64x64 x (ZeroExt32to64 y))
(Rsh64x16 x y) -> (Rsh64x64 x (ZeroExt16to64 y))
(Rsh64x8  x y) -> (Rsh64x64 x (ZeroExt8to64  y))

(Rsh32x64 x y) -> (Rsh64x64 (SignExt32to64 x) y)
(Rsh32x32 x y) -> (Rsh64x64 (SignExt32to64 x) (ZeroExt32to64 y))
(Rsh32x16 x y) -> (Rsh64x64 (SignExt32to64 x) (ZeroExt16to64 y))
(Rsh32x8  x y) -> (Rsh64x64 (SignExt32to64 x) (ZeroExt8to64  y))

(Rsh16x64 x y) -> (Rsh64x64 (SignExt16to64 x) y)
(Rsh16x32 x y) -> (Rsh64x64 (SignExt16to64 x) (ZeroExt32to64 y))
(Rsh16x16 x y) -> (Rsh64x64 (SignExt16to64 x) (ZeroExt16to64 y))
(Rsh16x8  x y) -> (Rsh64x64 (SignExt16to64 x) (ZeroExt8to64  y))

(Rsh8x64 x y)  -> (Rsh64x64 (SignExt8to64 x) y)
(Rsh8x32 x y)  -> (Rsh64x64 (SignExt8to64 x) (ZeroExt32to64 y))
(Rsh8x16 x y)  -> (Rsh64x64 (SignExt8to64 x) (ZeroExt16to64 y))
(Rsh8x8  x y)  -> (Rsh64x64 (SignExt8to64 x) (ZeroExt8to64  y))

// Lowering rotates
(RotateLeft8 <t> x (I64Const [c])) -> (Or8 (Lsh8x64 <t> x (I64Const [c&7])) (Rsh8Ux64 <t> x (I64Const [-c&7])))
(RotateLeft16 <t> x (I64Const [c])) -> (Or16 (Lsh16x64 <t> x (I64Const [c&15])) (Rsh16Ux64 <t> x (I64Const [-c&15])))
(RotateLeft32 <t> x (I64Const [c])) -> (Or32 (Lsh32x64 <t> x (I64Const [c&31])) (Rsh32Ux64 <t> x (I64Const [-c&31])))

// Lowering comparisons
(Less64  x y) -> (I64LtS x y)
(Less32  x y) -> (I64LtS (SignExt32to64 x) (SignExt32to64 y))
(Less16  x y) -> (I64LtS (SignExt16to64 x) (SignExt16to64 y))
(Less8   x y) -> (I64LtS (SignExt8to64  x) (SignExt8to64  y))
(Less64U x y) -> (I64LtU x y)
(Less32U x y) -> (I64LtU (ZeroExt32to64 x) (ZeroExt32to64 y))
(Less16U x y) -> (I64LtU (ZeroExt16to64 x) (ZeroExt16to64 y))
(Less8U  x y) -> (I64LtU (ZeroExt8to64  x) (ZeroExt8to64  y))
(Less64F x y) -> (F64Lt x y)
(Less32F x y) -> (F64Lt (LoweredRound32F x) (LoweredRound32F y))

(Leq64  x y) -> (I64LeS x y)
(Leq32  x y) -> (I64LeS (SignExt32to64 x) (SignExt32to64 y))
(Leq16  x y) -> (I64LeS (SignExt16to64 x) (SignExt16to64 y))
(Leq8   x y) -> (I64LeS (SignExt8to64  x) (SignExt8to64  y))
(Leq64U x y) -> (I64LeU x y)
(Leq32U x y) -> (I64LeU (ZeroExt32to64 x) (ZeroExt32to64 y))
(Leq16U x y) -> (I64LeU (ZeroExt16to64 x) (ZeroExt16to64 y))
(Leq8U  x y) -> (I64LeU (ZeroExt8to64  x) (ZeroExt8to64  y))
(Leq64F x y) -> (F64Le x y)
(Leq32F x y) -> (F64Le (LoweredRound32F x) (LoweredRound32F y))

(Greater64  x y) -> (I64GtS x y)
(Greater32  x y) -> (I64GtS (SignExt32to64 x) (SignExt32to64 y))
(Greater16  x y) -> (I64GtS (SignExt16to64 x) (SignExt16to64 y))
(Greater8   x y) -> (I64GtS (SignExt8to64  x) (SignExt8to64  y))
(Greater64U x y) -> (I64GtU x y)
(Greater32U x y) -> (I64GtU (ZeroExt32to64 x) (ZeroExt32to64 y))
(Greater16U x y) -> (I64GtU (ZeroExt16to64 x) (ZeroExt16to64 y))
(Greater8U  x y) -> (I64GtU (ZeroExt8to64  x) (ZeroExt8to64  y))
(Greater64F x y) -> (F64Gt x y)
(Greater32F x y) -> (F64Gt (LoweredRound32F x) (LoweredRound32F y))

(Geq64  x y) -> (I64GeS x y)
(Geq32  x y) -> (I64GeS (SignExt32to64 x) (SignExt32to64 y))
(Geq16  x y) -> (I64GeS (SignExt16to64 x) (SignExt16to64 y))
(Geq8   x y) -> (I64GeS (SignExt8to64  x) (SignExt8to64  y))
(Geq64U x y) -> (I64GeU x y)
(Geq32U x y) -> (I64GeU (ZeroExt32to64 x) (ZeroExt32to64 y))
(Geq16U x y) -> (I64GeU (ZeroExt16to64 x) (ZeroExt16to64 y))
(Geq8U  x y) -> (I64GeU (ZeroExt8to64  x) (ZeroExt8to64  y))
(Geq64F x y) -> (F64Ge x y)
(Geq32F x y) -> (F64Ge (LoweredRound32F x) (LoweredRound32F y))

(Eq64  x y) -> (I64Eq x y)
(Eq32  x y) -> (I64Eq (ZeroExt32to64 x) (ZeroExt32to64 y))
(Eq16  x y) -> (I64Eq (ZeroExt16to64 x) (ZeroExt16to64 y))
(Eq8   x y) -> (I64Eq (ZeroExt8to64  x) (ZeroExt8to64  y))
(EqB   x y) -> (I64Eq x y)
(EqPtr x y) -> (I64Eq x y)
(Eq64F x y) -> (F64Eq x y)
(Eq32F x y) -> (F64Eq (LoweredRound32F x) (LoweredRound32F y))

(Neq64  x y) -> (I64Ne x y)
(Neq32  x y) -> (I64Ne (ZeroExt32to64 x) (ZeroExt32to64 y))
(Neq16  x y) -> (I64Ne (ZeroExt16to64 x) (ZeroExt16to64 y))
(Neq8   x y) -> (I64Ne (ZeroExt8to64  x) (ZeroExt8to64  y))
(NeqB   x y) -> (I64Ne x y)
(NeqPtr x y) -> (I64Ne x y)
(Neq64F x y) -> (F64Ne x y)
(Neq32F x y) -> (F64Ne (LoweredRound32F x) (LoweredRound32F y))

// Lowering loads
(Load <t> ptr mem) && is32BitFloat(t) -> (F32Load ptr mem)
(Load <t> ptr mem) && is64BitFloat(t) -> (F64Load ptr mem)
(Load <t> ptr mem) && t.Size() == 8 -> (I64Load ptr mem)
(Load <t> ptr mem) && t.Size() == 4 && !t.IsSigned() -> (I64Load32U ptr mem)
(Load <t> ptr mem) && t.Size() == 4 &&  t.IsSigned() -> (I64Load32S ptr mem)
(Load <t> ptr mem) && t.Size() == 2 && !t.IsSigned() -> (I64Load16U ptr mem)
(Load <t> ptr mem) && t.Size() == 2 &&  t.IsSigned() -> (I64Load16S ptr mem)
(Load <t> ptr mem) && t.Size() == 1 && !t.IsSigned() -> (I64Load8U ptr mem)
(Load <t> ptr mem) && t.Size() == 1 &&  t.IsSigned() -> (I64Load8S ptr mem)

// Lowering stores
(Store {t} ptr val mem) && is64BitFloat(t.(*types.Type)) -> (F64Store ptr val mem)
(Store {t} ptr val mem) && is32BitFloat(t.(*types.Type)) -> (F32Store ptr val mem)
(Store {t} ptr val mem) && t.(*types.Type).Size() == 8 -> (I64Store ptr val mem)
(Store {t} ptr val mem) && t.(*types.Type).Size() == 4 -> (I64Store32 ptr val mem)
(Store {t} ptr val mem) && t.(*types.Type).Size() == 2 -> (I64Store16 ptr val mem)
(Store {t} ptr val mem) && t.(*types.Type).Size() == 1 -> (I64Store8 ptr val mem)

// Lowering moves
(Move [0] _ _ mem) -> mem
(Move [1] dst src mem) -> (I64Store8 dst (I64Load8U src mem) mem)
(Move [2] dst src mem) -> (I64Store16 dst (I64Load16U src mem) mem)
(Move [4] dst src mem) -> (I64Store32 dst (I64Load32U src mem) mem)
(Move [8] dst src mem) -> (I64Store dst (I64Load src mem) mem)
(Move [16] dst src mem) ->
	(I64Store [8] dst (I64Load [8] src mem)
		(I64Store dst (I64Load src mem) mem))
(Move [3] dst src mem) ->
	(I64Store8 [2] dst (I64Load8U [2] src mem)
		(I64Store16 dst (I64Load16U src mem) mem))
(Move [5] dst src mem) ->
	(I64Store8 [4] dst (I64Load8U [4] src mem)
		(I64Store32 dst (I64Load32U src mem) mem))
(Move [6] dst src mem) ->
	(I64Store16 [4] dst (I64Load16U [4] src mem)
		(I64Store32 dst (I64Load32U src mem) mem))
(Move [7] dst src mem) ->
	(I64Store32 [3] dst (I64Load32U [3] src mem)
		(I64Store32 dst (I64Load32U src mem) mem))
(Move [s] dst src mem) && s > 8 && s < 16 ->
	(I64Store [s-8] dst (I64Load [s-8] src mem)
		(I64Store dst (I64Load src mem) mem))

// Adjust moves to be a multiple of 16 bytes.
(Move [s] dst src mem)
	&& s > 16 && s%16 != 0 && s%16 <= 8 ->
	(Move [s-s%16]
		(OffPtr <dst.Type> dst [s%16])
		(OffPtr <src.Type> src [s%16])
		(I64Store dst (I64Load src mem) mem))
(Move [s] dst src mem)
	&& s > 16 && s%16 != 0 && s%16 > 8 ->
	(Move [s-s%16]
		(OffPtr <dst.Type> dst [s%16])
		(OffPtr <src.Type> src [s%16])
		(I64Store [8] dst (I64Load [8] src mem)
			(I64Store dst (I64Load src mem) mem)))

// Large copying uses helper.
(Move [s] dst src mem) && s%8 == 0 ->
	(LoweredMove [s/8] dst src mem)

// Lowering Zero instructions
(Zero [0] _ mem) -> mem
(Zero [1] destptr mem) -> (I64Store8 destptr (I64Const [0]) mem)
(Zero [2] destptr mem) -> (I64Store16 destptr (I64Const [0]) mem)
(Zero [4] destptr mem) -> (I64Store32 destptr (I64Const [0]) mem)
(Zero [8] destptr mem) -> (I64Store destptr (I64Const [0]) mem)

(Zero [3] destptr mem) ->
	(I64Store8 [2] destptr (I64Const [0])
		(I64Store16 destptr (I64Const [0]) mem))
(Zero [5] destptr mem) ->
	(I64Store8 [4] destptr (I64Const [0])
		(I64Store32 destptr (I64Const [0]) mem))
(Zero [6] destptr mem) ->
	(I64Store16 [4] destptr (I64Const [0])
		(I64Store32 destptr (I64Const [0]) mem))
(Zero [7] destptr mem) ->
	(I64Store32 [3] destptr (I64Const [0])
		(I64Store32 destptr (I64Const [0]) mem))

// Strip off any fractional word zeroing.
(Zero [s] destptr mem) && s%8 != 0 && s > 8 ->
	(Zero [s-s%8] (OffPtr <destptr.Type> destptr [s%8])
		(I64Store destptr (I64Const [0]) mem))

// Zero small numbers of words directly.
(Zero [16] destptr mem) ->
	(I64Store [8] destptr (I64Const [0])
		(I64Store destptr (I64Const [0]) mem))
(Zero [24] destptr mem) ->
	(I64Store [16] destptr (I64Const [0])
		(I64Store [8] destptr (I64Const [0])
			(I64Store destptr (I64Const [0]) mem)))
(Zero [32] destptr mem) ->
	(I64Store [24] destptr (I64Const [0])
		(I64Store [16] destptr (I64Const [0])
			(I64Store [8] destptr (I64Const [0])
				(I64Store destptr (I64Const [0]) mem))))

// Large zeroing uses helper.
(Zero [s] destptr mem) && s%8 == 0 && s > 32 ->
	(LoweredZero [s/8] destptr mem)

// Lowering constants
(Const(64|32|16|8) [val]) -> (I64Const [val])
(Const(64|32)F [val]) -> (F64Const [val])
(ConstNil) -> (I64Const [0])
(ConstBool [b]) -> (I64Const [b])

// Lowering calls
(StaticCall [argwid] {target} mem) -> (LoweredStaticCall [argwid] {target} mem)
(ClosureCall [argwid] entry closure mem) -> (LoweredClosureCall [argwid] entry closure mem)
(InterCall [argwid] entry mem) -> (LoweredInterCall [argwid] entry mem)

// Miscellaneous
(Convert <t> x mem) -> (LoweredConvert <t> x mem)
(IsNonNil p) -> (I64Eqz (I64Eqz p))
(IsInBounds idx len) -> (I64LtU idx len)
(IsSliceInBounds idx len) -> (I64LeU idx len)
(NilCheck ptr mem) -> (LoweredNilCheck ptr mem)
(GetClosurePtr) -> (LoweredGetClosurePtr)
(GetCallerPC) -> (LoweredGetCallerPC)
(GetCallerSP) -> (LoweredGetCallerSP)
(Addr {sym} base) -> (LoweredAddr {sym} base)
(LocalAddr {sym} base _) -> (LoweredAddr {sym} base)

// Write barrier.
(WB {fn} destptr srcptr mem) -> (LoweredWB {fn} destptr srcptr mem)

// --- Intrinsics ---
(Sqrt x) -> (F64Sqrt x)
(Trunc x) -> (F64Trunc x)
(Ceil x) -> (F64Ceil x)
(Floor x) -> (F64Floor x)
(RoundToEven x) -> (F64Nearest x)
(Abs x) -> (F64Abs x)
(Copysign x y) -> (F64Copysign x y)

(Ctz64 x) -> (I64Ctz x)
(Ctz32 x) -> (I64Ctz (I64Or x (I64Const [0x100000000])))
(Ctz16 x) -> (I64Ctz (I64Or x (I64Const [0x10000])))
(Ctz8  x) -> (I64Ctz (I64Or x (I64Const [0x100])))

(Ctz(64|32|16|8)NonZero x) -> (I64Ctz x)

(BitLen64 x) -> (I64Sub (I64Const [64]) (I64Clz x))

(RotateLeft64 x y) -> (I64Rotl x y)

(PopCount64 x) -> (I64Popcnt x)
(PopCount32 x) -> (I64Popcnt (ZeroExt32to64 x))
(PopCount16 x) -> (I64Popcnt (ZeroExt16to64 x))
(PopCount8  x) -> (I64Popcnt (ZeroExt8to64  x))

// --- Optimizations ---
(I64Add (I64Const [x]) (I64Const [y])) -> (I64Const [x + y])
(I64Mul (I64Const [x]) (I64Const [y])) -> (I64Const [x * y])
(I64And (I64Const [x]) (I64Const [y])) -> (I64Const [x & y])
(I64Or  (I64Const [x]) (I64Const [y])) -> (I64Const [x | y])
(I64Xor (I64Const [x]) (I64Const [y])) -> (I64Const [x ^ y])
(F64Add (F64Const [x]) (F64Const [y])) -> (F64Const [auxFrom64F(auxTo64F(x) + auxTo64F(y))])
(F64Mul (F64Const [x]) (F64Const [y])) -> (F64Const [auxFrom64F(auxTo64F(x) * auxTo64F(y))])
(I64Eq  (I64Const [x]) (I64Const [y])) && x == y -> (I64Const [1])
(I64Eq  (I64Const [x]) (I64Const [y])) && x != y -> (I64Const [0])
(I64Ne  (I64Const [x]) (I64Const [y])) && x == y -> (I64Const [0])
(I64Ne  (I64Const [x]) (I64Const [y])) && x != y -> (I64Const [1])

(I64Shl (I64Const [x]) (I64Const [y])) -> (I64Const [x << uint64(y)])
(I64ShrU (I64Const [x]) (I64Const [y])) -> (I64Const [int64(uint64(x) >> uint64(y))])
(I64ShrS (I64Const [x]) (I64Const [y])) -> (I64Const [x >> uint64(y)])

(I64Add (I64Const [x]) y) -> (I64Add y (I64Const [x]))
(I64Mul (I64Const [x]) y) -> (I64Mul y (I64Const [x]))
(I64And (I64Const [x]) y) -> (I64And y (I64Const [x]))
(I64Or  (I64Const [x]) y) -> (I64Or  y (I64Const [x]))
(I64Xor (I64Const [x]) y) -> (I64Xor y (I64Const [x]))
(F64Add (F64Const [x]) y) -> (F64Add y (F64Const [x]))
(F64Mul (F64Const [x]) y) -> (F64Mul y (F64Const [x]))
(I64Eq  (I64Const [x]) y) -> (I64Eq y  (I64Const [x]))
(I64Ne  (I64Const [x]) y) -> (I64Ne y  (I64Const [x]))

(I64Eq x (I64Const [0])) -> (I64Eqz x)
(I64Ne x (I64Const [0])) -> (I64Eqz (I64Eqz x))

(I64Add x (I64Const [y])) -> (I64AddConst [y] x)
(I64AddConst [0] x) -> x
(I64Eqz (I64Eqz (I64Eqz x))) -> (I64Eqz x)

// folding offset into load/store
((I64Load|I64Load32U|I64Load32S|I64Load16U|I64Load16S|I64Load8U|I64Load8S) [off] (I64AddConst [off2] ptr) mem)
	&& isU32Bit(off+off2) ->
	((I64Load|I64Load32U|I64Load32S|I64Load16U|I64Load16S|I64Load8U|I64Load8S) [off+off2] ptr mem)

((I64Store|I64Store32|I64Store16|I64Store8) [off] (I64AddConst [off2] ptr) val mem)
	&& isU32Bit(off+off2) ->
	((I64Store|I64Store32|I64Store16|I64Store8) [off+off2] ptr val mem)

// folding offset into address
(I64AddConst [off] (LoweredAddr {sym} [off2] base)) && isU32Bit(off+off2) ->
	(LoweredAddr {sym} [off+off2] base)