...

Text file src/cmd/compile/internal/ssa/_gen/MIPS.rules

Documentation: cmd/compile/internal/ssa/_gen

     1// Copyright 2016 The Go Authors. All rights reserved.
     2// Use of this source code is governed by a BSD-style
     3// license that can be found in the LICENSE file.
     4
     5(Add(Ptr|32|16|8) ...) => (ADD ...)
     6(Add(32|64)F ...) => (ADD(F|D) ...)
     7
     8(Select0 (Add32carry <t> x y)) => (ADD <t.FieldType(0)> x y)
     9(Select1 (Add32carry <t> x y)) => (SGTU <typ.Bool> x (ADD <t.FieldType(0)> x y))
    10(Add32withcarry <t> x y c) => (ADD c (ADD <t> x y))
    11
    12(Select0 (Add32carrywithcarry <t> x y c)) => (ADD <t.FieldType(0)> c (ADD <t.FieldType(0)> x y))
    13(Select1 (Add32carrywithcarry <t> x y c)) =>
    14	(OR <typ.Bool>
    15		(SGTU <typ.Bool> x xy:(ADD <t.FieldType(0)> x y))
    16		(SGTU <typ.Bool> xy (ADD <t.FieldType(0)> c xy)))
    17
    18(Sub(Ptr|32|16|8) ...) => (SUB ...)
    19(Sub(32|64)F ...) => (SUB(F|D) ...)
    20
    21(Select0 (Sub32carry <t> x y)) => (SUB <t.FieldType(0)> x y)
    22(Select1 (Sub32carry <t> x y)) => (SGTU <typ.Bool> (SUB <t.FieldType(0)> x y) x)
    23(Sub32withcarry <t> x y c) => (SUB (SUB <t> x y) c)
    24
    25(Mul(32|16|8) ...) => (MUL ...)
    26(Mul(32|64)F ...) => (MUL(F|D) ...)
    27
    28(Hmul(32|32u) x y) => (Select0 (MUL(T|TU) x y))
    29(Mul32uhilo ...) => (MULTU ...)
    30
    31(Div32 x y) => (Select1 (DIV x y))
    32(Div32u x y) => (Select1 (DIVU x y))
    33(Div16 x y) => (Select1 (DIV (SignExt16to32 x) (SignExt16to32 y)))
    34(Div16u x y) => (Select1 (DIVU (ZeroExt16to32 x) (ZeroExt16to32 y)))
    35(Div8 x y) => (Select1 (DIV (SignExt8to32 x) (SignExt8to32 y)))
    36(Div8u x y) => (Select1 (DIVU (ZeroExt8to32 x) (ZeroExt8to32 y)))
    37(Div(32|64)F ...) => (DIV(F|D) ...)
    38
    39(Mod32 x y) => (Select0 (DIV x y))
    40(Mod32u x y) => (Select0 (DIVU x y))
    41(Mod16 x y) => (Select0 (DIV (SignExt16to32 x) (SignExt16to32 y)))
    42(Mod16u x y) => (Select0 (DIVU (ZeroExt16to32 x) (ZeroExt16to32 y)))
    43(Mod8 x y) => (Select0 (DIV (SignExt8to32 x) (SignExt8to32 y)))
    44(Mod8u x y) => (Select0 (DIVU (ZeroExt8to32 x) (ZeroExt8to32 y)))
    45
    46// math package intrinsics
    47(Abs ...) => (ABSD ...)
    48
    49// (x + y) / 2 with x>=y  becomes  (x - y) / 2 + y
    50(Avg32u <t> x y) => (ADD (SRLconst <t> (SUB <t> x y) [1]) y)
    51
    52(And(32|16|8) ...) => (AND ...)
    53(Or(32|16|8) ...) => (OR ...)
    54(Xor(32|16|8) ...) => (XOR ...)
    55
    56// constant shifts
    57// generic opt rewrites all constant shifts to shift by Const64
    58(Lsh32x64  x (Const64 [c])) && uint32(c) < 32 => (SLLconst x [int32(c)])
    59(Rsh32x64  x (Const64 [c])) && uint32(c) < 32 => (SRAconst x [int32(c)])
    60(Rsh32Ux64 x (Const64 [c])) && uint32(c) < 32 => (SRLconst x [int32(c)])
    61(Lsh16x64  x (Const64 [c])) && uint32(c) < 16 => (SLLconst x [int32(c)])
    62(Rsh16x64  x (Const64 [c])) && uint32(c) < 16 => (SRAconst (SLLconst <typ.UInt32> x [16]) [int32(c+16)])
    63(Rsh16Ux64 x (Const64 [c])) && uint32(c) < 16 => (SRLconst (SLLconst <typ.UInt32> x [16]) [int32(c+16)])
    64(Lsh8x64   x (Const64 [c])) && uint32(c) < 8  => (SLLconst x [int32(c)])
    65(Rsh8x64   x (Const64 [c])) && uint32(c) < 8  => (SRAconst (SLLconst <typ.UInt32> x [24]) [int32(c+24)])
    66(Rsh8Ux64  x (Const64 [c])) && uint32(c) < 8  => (SRLconst (SLLconst <typ.UInt32> x [24]) [int32(c+24)])
    67
    68// large constant shifts
    69(Lsh32x64 _ (Const64 [c]))  && uint32(c) >= 32 => (MOVWconst [0])
    70(Rsh32Ux64 _ (Const64 [c])) && uint32(c) >= 32 => (MOVWconst [0])
    71(Lsh16x64 _ (Const64 [c]))  && uint32(c) >= 16 => (MOVWconst [0])
    72(Rsh16Ux64 _ (Const64 [c])) && uint32(c) >= 16 => (MOVWconst [0])
    73(Lsh8x64 _ (Const64 [c]))   && uint32(c) >= 8  => (MOVWconst [0])
    74(Rsh8Ux64 _ (Const64 [c]))  && uint32(c) >= 8  => (MOVWconst [0])
    75
    76// large constant signed right shift, we leave the sign bit
    77(Rsh32x64 x (Const64 [c])) && uint32(c) >= 32 => (SRAconst x [31])
    78(Rsh16x64 x (Const64 [c])) && uint32(c) >= 16 => (SRAconst (SLLconst <typ.UInt32> x [16]) [31])
    79(Rsh8x64  x (Const64 [c])) && uint32(c) >= 8  => (SRAconst (SLLconst <typ.UInt32> x [24]) [31])
    80
    81// shifts
    82// hardware instruction uses only the low 5 bits of the shift
    83// we compare to 32 to ensure Go semantics for large shifts
    84(Lsh32x32 <t> x y) => (CMOVZ (SLL <t> x y) (MOVWconst [0]) (SGTUconst [32] y))
    85(Lsh32x16 <t> x y) => (CMOVZ (SLL <t> x (ZeroExt16to32 y) ) (MOVWconst [0]) (SGTUconst [32] (ZeroExt16to32 y)))
    86(Lsh32x8 <t> x y)  => (CMOVZ (SLL <t> x (ZeroExt8to32 y) ) (MOVWconst [0]) (SGTUconst [32] (ZeroExt8to32 y)))
    87
    88(Lsh16x32 <t> x y) => (CMOVZ (SLL <t> x y) (MOVWconst [0]) (SGTUconst [32] y))
    89(Lsh16x16 <t> x y) => (CMOVZ (SLL <t> x (ZeroExt16to32 y) ) (MOVWconst [0]) (SGTUconst [32] (ZeroExt16to32 y)))
    90(Lsh16x8 <t> x y)  => (CMOVZ (SLL <t> x (ZeroExt8to32 y) ) (MOVWconst [0]) (SGTUconst [32] (ZeroExt8to32 y)))
    91
    92(Lsh8x32 <t> x y) => (CMOVZ (SLL <t> x y) (MOVWconst [0]) (SGTUconst [32] y))
    93(Lsh8x16 <t> x y) => (CMOVZ (SLL <t> x (ZeroExt16to32 y) ) (MOVWconst [0]) (SGTUconst [32] (ZeroExt16to32 y)))
    94(Lsh8x8 <t> x y)  => (CMOVZ (SLL <t> x (ZeroExt8to32 y) ) (MOVWconst [0]) (SGTUconst [32] (ZeroExt8to32 y)))
    95
    96(Rsh32Ux32 <t> x y) => (CMOVZ (SRL <t> x y) (MOVWconst [0]) (SGTUconst [32] y))
    97(Rsh32Ux16 <t> x y) => (CMOVZ (SRL <t> x (ZeroExt16to32 y) ) (MOVWconst [0]) (SGTUconst [32] (ZeroExt16to32 y)))
    98(Rsh32Ux8 <t> x y)  => (CMOVZ (SRL <t> x (ZeroExt8to32 y) ) (MOVWconst [0]) (SGTUconst [32] (ZeroExt8to32 y)))
    99
   100(Rsh16Ux32 <t> x y) => (CMOVZ (SRL <t> (ZeroExt16to32 x) y) (MOVWconst [0]) (SGTUconst [32] y))
   101(Rsh16Ux16 <t> x y) => (CMOVZ (SRL <t> (ZeroExt16to32 x) (ZeroExt16to32 y) ) (MOVWconst [0]) (SGTUconst [32] (ZeroExt16to32 y)))
   102(Rsh16Ux8 <t> x y)  => (CMOVZ (SRL <t> (ZeroExt16to32 x) (ZeroExt8to32 y) ) (MOVWconst [0]) (SGTUconst [32] (ZeroExt8to32 y)))
   103
   104(Rsh8Ux32 <t> x y) => (CMOVZ (SRL <t> (ZeroExt8to32 x) y) (MOVWconst [0]) (SGTUconst [32] y))
   105(Rsh8Ux16 <t> x y) => (CMOVZ (SRL <t> (ZeroExt8to32 x) (ZeroExt16to32 y) ) (MOVWconst [0]) (SGTUconst [32] (ZeroExt16to32 y)))
   106(Rsh8Ux8 <t> x y)  => (CMOVZ (SRL <t> (ZeroExt8to32 x) (ZeroExt8to32 y) ) (MOVWconst [0]) (SGTUconst [32] (ZeroExt8to32 y)))
   107
   108(Rsh32x32 x y) => (SRA x ( CMOVZ <typ.UInt32> y (MOVWconst [31]) (SGTUconst [32] y)))
   109(Rsh32x16 x y) => (SRA x ( CMOVZ <typ.UInt32> (ZeroExt16to32 y) (MOVWconst [31]) (SGTUconst [32] (ZeroExt16to32 y))))
   110(Rsh32x8 x y)  => (SRA x ( CMOVZ <typ.UInt32> (ZeroExt8to32 y) (MOVWconst [31]) (SGTUconst [32] (ZeroExt8to32 y))))
   111
   112(Rsh16x32 x y) => (SRA (SignExt16to32 x) ( CMOVZ <typ.UInt32> y (MOVWconst [31]) (SGTUconst [32] y)))
   113(Rsh16x16 x y) => (SRA (SignExt16to32 x) ( CMOVZ <typ.UInt32> (ZeroExt16to32 y) (MOVWconst [31]) (SGTUconst [32] (ZeroExt16to32 y))))
   114(Rsh16x8 x y)  => (SRA (SignExt16to32 x) ( CMOVZ <typ.UInt32> (ZeroExt8to32 y) (MOVWconst [31]) (SGTUconst [32] (ZeroExt8to32 y))))
   115
   116(Rsh8x32 x y) => (SRA (SignExt16to32 x) ( CMOVZ <typ.UInt32> y (MOVWconst [31]) (SGTUconst [32] y)))
   117(Rsh8x16 x y) => (SRA (SignExt16to32 x) ( CMOVZ <typ.UInt32> (ZeroExt16to32 y) (MOVWconst [31]) (SGTUconst [32] (ZeroExt16to32 y))))
   118(Rsh8x8 x y)  => (SRA (SignExt16to32 x) ( CMOVZ <typ.UInt32> (ZeroExt8to32 y) (MOVWconst [31]) (SGTUconst [32] (ZeroExt8to32 y))))
   119
   120// rotates
   121(RotateLeft8  <t> x (MOVWconst [c])) => (Or8  (Lsh8x32 <t>  x (MOVWconst [c&7]))  (Rsh8Ux32 <t> x (MOVWconst [-c&7])))
   122(RotateLeft16 <t> x (MOVWconst [c])) => (Or16 (Lsh16x32 <t> x (MOVWconst [c&15])) (Rsh16Ux32 <t> x (MOVWconst [-c&15])))
   123(RotateLeft32 <t> x (MOVWconst [c])) => (Or32 (Lsh32x32 <t> x (MOVWconst [c&31])) (Rsh32Ux32 <t> x (MOVWconst [-c&31])))
   124(RotateLeft64 <t> x (MOVWconst [c])) => (Or64 (Lsh64x32 <t> x (MOVWconst [c&63])) (Rsh64Ux32 <t> x (MOVWconst [-c&63])))
   125
   126// unary ops
   127(Neg(32|16|8) ...) => (NEG ...)
   128(Neg(32|64)F ...) => (NEG(F|D) ...)
   129
   130(Com(32|16|8) x) => (NORconst [0] x)
   131
   132(Sqrt ...) => (SQRTD ...)
   133(Sqrt32 ...) => (SQRTF ...)
   134
   135(Ctz(32|16|8)NonZero ...) => (Ctz32 ...)
   136
   137// count trailing zero
   138// 32 - CLZ(x&-x - 1)
   139(Ctz32 <t> x) => (SUB (MOVWconst [32]) (CLZ <t> (SUBconst <t> [1] (AND <t> x (NEG <t> x)))))
   140(Ctz16 x) => (Ctz32 (Or32 <typ.UInt32> x (MOVWconst [1<<16])))
   141(Ctz8  x) => (Ctz32 (Or32 <typ.UInt32> x (MOVWconst [1<<8])))
   142
   143// bit length
   144(BitLen32 <t> x) => (SUB (MOVWconst [32]) (CLZ <t> x))
   145(BitLen(16|8) x) => (BitLen32 (ZeroExt(16|8)to32 x))
   146
   147// boolean ops -- booleans are represented with 0=false, 1=true
   148(AndB ...) => (AND ...)
   149(OrB ...) => (OR ...)
   150(EqB x y) => (XORconst [1] (XOR <typ.Bool> x y))
   151(NeqB ...) => (XOR ...)
   152(Not x) => (XORconst [1] x)
   153
   154// constants
   155(Const(32|16|8) [val]) => (MOVWconst [int32(val)])
   156(Const(32|64)F ...) => (MOV(F|D)const ...)
   157(ConstNil) => (MOVWconst [0])
   158(ConstBool [t]) => (MOVWconst [b2i32(t)])
   159
   160// truncations
   161// Because we ignore high parts of registers, truncates are just copies.
   162(Trunc16to8 ...)  => (Copy ...)
   163(Trunc32to8 ...)  => (Copy ...)
   164(Trunc32to16 ...) => (Copy ...)
   165
   166// Zero-/Sign-extensions
   167(ZeroExt8to16 ...)  => (MOVBUreg ...)
   168(ZeroExt8to32 ...)  => (MOVBUreg ...)
   169(ZeroExt16to32 ...) => (MOVHUreg ...)
   170
   171(SignExt8to16 ...)  => (MOVBreg ...)
   172(SignExt8to32 ...)  => (MOVBreg ...)
   173(SignExt16to32 ...) => (MOVHreg ...)
   174
   175(Signmask x) => (SRAconst x [31])
   176(Zeromask x) => (NEG (SGTU x (MOVWconst [0])))
   177(Slicemask <t> x) => (SRAconst (NEG <t> x) [31])
   178
   179// float-int conversion
   180(Cvt32to(32|64)F ...) => (MOVW(F|D) ...)
   181(Cvt(32|64)Fto32 ...) => (TRUNC(F|D)W ...)
   182(Cvt32Fto64F ...) => (MOVFD ...)
   183(Cvt64Fto32F ...) => (MOVDF ...)
   184
   185(CvtBoolToUint8 ...) => (Copy ...)
   186
   187(Round(32|64)F ...) => (Copy ...)
   188
   189// comparisons
   190(Eq8 x y)  => (SGTUconst [1] (XOR (ZeroExt8to32 x) (ZeroExt8to32 y)))
   191(Eq16 x y) => (SGTUconst [1] (XOR (ZeroExt16to32 x) (ZeroExt16to32 y)))
   192(Eq32 x y) => (SGTUconst [1] (XOR x y))
   193(EqPtr x y) => (SGTUconst [1] (XOR x y))
   194(Eq(32|64)F x y) => (FPFlagTrue (CMPEQ(F|D) x y))
   195
   196(Neq8 x y)  => (SGTU (XOR (ZeroExt8to32 x) (ZeroExt8to32 y)) (MOVWconst [0]))
   197(Neq16 x y) => (SGTU (XOR (ZeroExt16to32 x) (ZeroExt16to32 y)) (MOVWconst [0]))
   198(Neq32 x y) => (SGTU (XOR x y) (MOVWconst [0]))
   199(NeqPtr x y) => (SGTU (XOR x y) (MOVWconst [0]))
   200(Neq(32|64)F x y) => (FPFlagFalse (CMPEQ(F|D) x y))
   201
   202(Less8 x y)  => (SGT (SignExt8to32 y) (SignExt8to32 x))
   203(Less16 x y) => (SGT (SignExt16to32 y) (SignExt16to32 x))
   204(Less32 x y) => (SGT y x)
   205(Less(32|64)F x y) => (FPFlagTrue (CMPGT(F|D) y x)) // reverse operands to work around NaN
   206
   207(Less8U x y)  => (SGTU (ZeroExt8to32 y) (ZeroExt8to32 x))
   208(Less16U x y) => (SGTU (ZeroExt16to32 y) (ZeroExt16to32 x))
   209(Less32U x y) => (SGTU y x)
   210
   211(Leq8 x y)  => (XORconst [1] (SGT (SignExt8to32 x) (SignExt8to32 y)))
   212(Leq16 x y) => (XORconst [1] (SGT (SignExt16to32 x) (SignExt16to32 y)))
   213(Leq32 x y) => (XORconst [1] (SGT x y))
   214(Leq(32|64)F x y) => (FPFlagTrue (CMPGE(F|D) y x)) // reverse operands to work around NaN
   215
   216(Leq8U x y)  => (XORconst [1] (SGTU (ZeroExt8to32 x) (ZeroExt8to32 y)))
   217(Leq16U x y) => (XORconst [1] (SGTU (ZeroExt16to32 x) (ZeroExt16to32 y)))
   218(Leq32U x y) => (XORconst [1] (SGTU x y))
   219
   220(OffPtr [off] ptr:(SP)) => (MOVWaddr [int32(off)] ptr)
   221(OffPtr [off] ptr) => (ADDconst [int32(off)] ptr)
   222
   223(Addr {sym} base) => (MOVWaddr {sym} base)
   224(LocalAddr <t> {sym} base mem) && t.Elem().HasPointers() => (MOVWaddr {sym} (SPanchored base mem))
   225(LocalAddr <t> {sym} base _)  && !t.Elem().HasPointers() => (MOVWaddr {sym} base)
   226
   227// loads
   228(Load <t> ptr mem) && t.IsBoolean() => (MOVBUload ptr mem)
   229(Load <t> ptr mem) && (is8BitInt(t)  &&  t.IsSigned()) => (MOVBload ptr mem)
   230(Load <t> ptr mem) && (is8BitInt(t)  && !t.IsSigned()) => (MOVBUload ptr mem)
   231(Load <t> ptr mem) && (is16BitInt(t) &&  t.IsSigned()) => (MOVHload ptr mem)
   232(Load <t> ptr mem) && (is16BitInt(t) && !t.IsSigned()) => (MOVHUload ptr mem)
   233(Load <t> ptr mem) && (is32BitInt(t) || isPtr(t)) => (MOVWload ptr mem)
   234(Load <t> ptr mem) && is32BitFloat(t) => (MOVFload ptr mem)
   235(Load <t> ptr mem) && is64BitFloat(t) => (MOVDload ptr mem)
   236
   237// stores
   238(Store {t} ptr val mem) && t.Size() == 1 => (MOVBstore ptr val mem)
   239(Store {t} ptr val mem) && t.Size() == 2 => (MOVHstore ptr val mem)
   240(Store {t} ptr val mem) && t.Size() == 4 && !t.IsFloat() => (MOVWstore ptr val mem)
   241(Store {t} ptr val mem) && t.Size() == 4 &&  t.IsFloat() => (MOVFstore ptr val mem)
   242(Store {t} ptr val mem) && t.Size() == 8 &&  t.IsFloat() => (MOVDstore ptr val mem)
   243
   244// float <=> int register moves, with no conversion.
   245// These come up when compiling math.{Float32bits, Float32frombits}.
   246(MOVWload [off] {sym} ptr (MOVFstore [off] {sym} ptr val _)) => (MOVWfpgp val)
   247(MOVFload [off] {sym} ptr (MOVWstore [off] {sym} ptr val _)) => (MOVWgpfp val)
   248
   249// Similarly for stores, if we see a store after FPR <=> GPR move, then redirect store to use the other register set.
   250(MOVWstore [off] {sym} ptr (MOVWfpgp val) mem) => (MOVFstore [off] {sym} ptr val mem)
   251(MOVFstore [off] {sym} ptr (MOVWgpfp val) mem) => (MOVWstore [off] {sym} ptr val mem)
   252
   253// zero instructions
   254(Zero [0] _ mem) => mem
   255(Zero [1] ptr mem) => (MOVBstore ptr (MOVWconst [0]) mem)
   256(Zero [2] {t} ptr mem) && t.Alignment()%2 == 0 =>
   257	(MOVHstore ptr (MOVWconst [0]) mem)
   258(Zero [2] ptr mem) =>
   259	(MOVBstore [1] ptr (MOVWconst [0])
   260		(MOVBstore [0] ptr (MOVWconst [0]) mem))
   261(Zero [4] {t} ptr mem) && t.Alignment()%4 == 0 =>
   262	(MOVWstore ptr (MOVWconst [0]) mem)
   263(Zero [4] {t} ptr mem) && t.Alignment()%2 == 0 =>
   264	(MOVHstore [2] ptr (MOVWconst [0])
   265		(MOVHstore [0] ptr (MOVWconst [0]) mem))
   266(Zero [4] ptr mem) =>
   267	(MOVBstore [3] ptr (MOVWconst [0])
   268		(MOVBstore [2] ptr (MOVWconst [0])
   269			(MOVBstore [1] ptr (MOVWconst [0])
   270				(MOVBstore [0] ptr (MOVWconst [0]) mem))))
   271(Zero [3] ptr mem) =>
   272	(MOVBstore [2] ptr (MOVWconst [0])
   273		(MOVBstore [1] ptr (MOVWconst [0])
   274			(MOVBstore [0] ptr (MOVWconst [0]) mem)))
   275(Zero [6] {t} ptr mem) && t.Alignment()%2 == 0 =>
   276	(MOVHstore [4] ptr (MOVWconst [0])
   277		(MOVHstore [2] ptr (MOVWconst [0])
   278			(MOVHstore [0] ptr (MOVWconst [0]) mem)))
   279(Zero [8] {t} ptr mem) && t.Alignment()%4 == 0 =>
   280		(MOVWstore [4] ptr (MOVWconst [0])
   281			(MOVWstore [0] ptr (MOVWconst [0]) mem))
   282(Zero [12] {t} ptr mem) && t.Alignment()%4 == 0 =>
   283	(MOVWstore [8] ptr (MOVWconst [0])
   284		(MOVWstore [4] ptr (MOVWconst [0])
   285			(MOVWstore [0] ptr (MOVWconst [0]) mem)))
   286(Zero [16] {t} ptr mem) && t.Alignment()%4 == 0 =>
   287	(MOVWstore [12] ptr (MOVWconst [0])
   288		(MOVWstore [8] ptr (MOVWconst [0])
   289			(MOVWstore [4] ptr (MOVWconst [0])
   290				(MOVWstore [0] ptr (MOVWconst [0]) mem))))
   291
   292// large or unaligned zeroing uses a loop
   293(Zero [s] {t} ptr mem)
   294	&& (s > 16  || t.Alignment()%4 != 0) =>
   295	(LoweredZero [int32(t.Alignment())]
   296		ptr
   297		(ADDconst <ptr.Type> ptr [int32(s-moveSize(t.Alignment(), config))])
   298		mem)
   299
   300// moves
   301(Move [0] _ _ mem) => mem
   302(Move [1] dst src mem) => (MOVBstore dst (MOVBUload src mem) mem)
   303(Move [2] {t} dst src mem) && t.Alignment()%2 == 0 =>
   304	(MOVHstore dst (MOVHUload src mem) mem)
   305(Move [2] dst src mem) =>
   306	(MOVBstore [1] dst (MOVBUload [1] src mem)
   307		(MOVBstore dst (MOVBUload src mem) mem))
   308(Move [4] {t} dst src mem) && t.Alignment()%4 == 0 =>
   309	(MOVWstore dst (MOVWload src mem) mem)
   310(Move [4] {t} dst src mem) && t.Alignment()%2 == 0 =>
   311	(MOVHstore [2] dst (MOVHUload [2] src mem)
   312		(MOVHstore dst (MOVHUload src mem) mem))
   313(Move [4] dst src mem) =>
   314	(MOVBstore [3] dst (MOVBUload [3] src mem)
   315		(MOVBstore [2] dst (MOVBUload [2] src mem)
   316			(MOVBstore [1] dst (MOVBUload [1] src mem)
   317				(MOVBstore dst (MOVBUload src mem) mem))))
   318(Move [3] dst src mem) =>
   319	(MOVBstore [2] dst (MOVBUload [2] src mem)
   320		(MOVBstore [1] dst (MOVBUload [1] src mem)
   321			(MOVBstore dst (MOVBUload src mem) mem)))
   322(Move [8] {t} dst src mem) && t.Alignment()%4 == 0 =>
   323	(MOVWstore [4] dst (MOVWload [4] src mem)
   324		(MOVWstore dst (MOVWload src mem) mem))
   325(Move [8] {t} dst src mem) && t.Alignment()%2 == 0 =>
   326	(MOVHstore [6] dst (MOVHload [6] src mem)
   327		(MOVHstore [4] dst (MOVHload [4] src mem)
   328			(MOVHstore [2] dst (MOVHload [2] src mem)
   329				(MOVHstore dst (MOVHload src mem) mem))))
   330(Move [6] {t} dst src mem) && t.Alignment()%2 == 0 =>
   331	(MOVHstore [4] dst (MOVHload [4] src mem)
   332		(MOVHstore [2] dst (MOVHload [2] src mem)
   333			(MOVHstore dst (MOVHload src mem) mem)))
   334(Move [12] {t} dst src mem) && t.Alignment()%4 == 0 =>
   335	(MOVWstore [8] dst (MOVWload [8] src mem)
   336		(MOVWstore [4] dst (MOVWload [4] src mem)
   337			(MOVWstore dst (MOVWload src mem) mem)))
   338(Move [16] {t} dst src mem) && t.Alignment()%4 == 0 =>
   339	(MOVWstore [12] dst (MOVWload [12] src mem)
   340		(MOVWstore [8] dst (MOVWload [8] src mem)
   341			(MOVWstore [4] dst (MOVWload [4] src mem)
   342				(MOVWstore dst (MOVWload src mem) mem))))
   343
   344
   345// large or unaligned move uses a loop
   346(Move [s] {t} dst src mem)
   347	&& (s > 16 && logLargeCopy(v, s) || t.Alignment()%4 != 0) =>
   348	(LoweredMove [int32(t.Alignment())]
   349		dst
   350		src
   351		(ADDconst <src.Type> src [int32(s-moveSize(t.Alignment(), config))])
   352		mem)
   353
   354// calls
   355(StaticCall ...)  => (CALLstatic ...)
   356(ClosureCall ...) => (CALLclosure ...)
   357(InterCall ...)   => (CALLinter ...)
   358(TailCall ...) => (CALLtail ...)
   359
   360// atomic intrinsics
   361(AtomicLoad(8|32) ...) => (LoweredAtomicLoad(8|32) ...)
   362(AtomicLoadPtr    ...) => (LoweredAtomicLoad32     ...)
   363
   364(AtomicStore(8|32)  ...) => (LoweredAtomicStore(8|32) ...)
   365(AtomicStorePtrNoWB ...) => (LoweredAtomicStore32     ...)
   366
   367(AtomicExchange32 ...) => (LoweredAtomicExchange ...)
   368(AtomicAdd32 ...) => (LoweredAtomicAdd ...)
   369
   370(AtomicCompareAndSwap32 ...) => (LoweredAtomicCas ...)
   371
   372// AtomicOr8(ptr,val)  =>   LoweredAtomicOr(ptr&^3,uint32(val) << ((ptr & 3) * 8))
   373(AtomicOr8 ptr val mem) && !config.BigEndian =>
   374	(LoweredAtomicOr (AND <typ.UInt32Ptr> (MOVWconst [^3]) ptr)
   375		(SLL <typ.UInt32> (ZeroExt8to32 val)
   376			(SLLconst <typ.UInt32> [3]
   377				(ANDconst <typ.UInt32> [3] ptr))) mem)
   378
   379// AtomicAnd8(ptr,val)  =>  LoweredAtomicAnd(ptr&^3,(uint32(val) << ((ptr & 3) * 8)) | ^(uint32(0xFF) << ((ptr & 3) * 8))))
   380(AtomicAnd8  ptr val mem) && !config.BigEndian =>
   381	(LoweredAtomicAnd (AND <typ.UInt32Ptr> (MOVWconst [^3]) ptr)
   382		(OR <typ.UInt32> (SLL <typ.UInt32> (ZeroExt8to32 val)
   383			(SLLconst <typ.UInt32> [3]
   384				(ANDconst  <typ.UInt32> [3] ptr)))
   385		(NORconst [0] <typ.UInt32> (SLL <typ.UInt32>
   386			(MOVWconst [0xff]) (SLLconst <typ.UInt32> [3]
   387				(ANDconst <typ.UInt32> [3] ptr))))) mem)
   388
   389// AtomicOr8(ptr,val)  =>  LoweredAtomicOr(ptr&^3,uint32(val) << (((ptr^3) & 3) * 8))
   390(AtomicOr8 ptr val mem) && config.BigEndian =>
   391	(LoweredAtomicOr (AND <typ.UInt32Ptr> (MOVWconst [^3]) ptr)
   392		(SLL <typ.UInt32> (ZeroExt8to32 val)
   393			(SLLconst <typ.UInt32> [3]
   394				(ANDconst <typ.UInt32> [3]
   395					(XORconst <typ.UInt32> [3] ptr)))) mem)
   396
   397// AtomicAnd8(ptr,val)  =>  LoweredAtomicAnd(ptr&^3,(uint32(val) << (((ptr^3) & 3) * 8)) | ^(uint32(0xFF) << (((ptr^3) & 3) * 8))))
   398(AtomicAnd8  ptr val mem) && config.BigEndian =>
   399	(LoweredAtomicAnd (AND <typ.UInt32Ptr> (MOVWconst [^3]) ptr)
   400		(OR <typ.UInt32> (SLL <typ.UInt32> (ZeroExt8to32 val)
   401			(SLLconst <typ.UInt32> [3]
   402				(ANDconst  <typ.UInt32> [3]
   403					(XORconst <typ.UInt32> [3] ptr))))
   404		(NORconst [0] <typ.UInt32> (SLL <typ.UInt32>
   405			(MOVWconst [0xff]) (SLLconst <typ.UInt32> [3]
   406				(ANDconst <typ.UInt32> [3]
   407					(XORconst <typ.UInt32> [3] ptr)))))) mem)
   408
   409(AtomicAnd32 ...) => (LoweredAtomicAnd ...)
   410(AtomicOr32  ...) => (LoweredAtomicOr  ...)
   411
   412
   413// checks
   414(NilCheck ...) => (LoweredNilCheck ...)
   415(IsNonNil ptr) => (SGTU ptr (MOVWconst [0]))
   416(IsInBounds idx len) => (SGTU len idx)
   417(IsSliceInBounds idx len) => (XORconst [1] (SGTU idx len))
   418
   419// pseudo-ops
   420(GetClosurePtr ...) => (LoweredGetClosurePtr ...)
   421(GetCallerSP ...) => (LoweredGetCallerSP ...)
   422(GetCallerPC ...) => (LoweredGetCallerPC ...)
   423
   424(If cond yes no) => (NE cond yes no)
   425
   426// Write barrier.
   427(WB ...) => (LoweredWB ...)
   428
   429// Publication barrier as intrinsic
   430(PubBarrier ...) => (LoweredPubBarrier ...)
   431
   432
   433(PanicBounds ...) => (LoweredPanicBoundsRR ...)
   434(PanicExtend ...) => (LoweredPanicExtendRR ...)
   435
   436(LoweredPanicBoundsRR [kind] x (MOVWconst [c]) mem) => (LoweredPanicBoundsRC [kind] x {PanicBoundsC{C:int64(c)}} mem)
   437(LoweredPanicBoundsRR [kind] (MOVWconst [c]) y mem) => (LoweredPanicBoundsCR [kind] {PanicBoundsC{C:int64(c)}} y mem)
   438(LoweredPanicBoundsRC [kind] {p} (MOVWconst [c]) mem) => (LoweredPanicBoundsCC [kind] {PanicBoundsCC{Cx:int64(c), Cy:p.C}} mem)
   439
   440(LoweredPanicExtendRR [kind] hi lo (MOVWconst [c]) mem) => (LoweredPanicExtendRC [kind] hi lo {PanicBoundsC{C:int64(c)}} mem)
   441(LoweredPanicExtendRR [kind] (MOVWconst [hi]) (MOVWconst [lo]) y mem) => (LoweredPanicBoundsCR [kind] {PanicBoundsC{C:int64(hi)<<32 + int64(uint32(lo))}} y mem)
   442(LoweredPanicExtendRC [kind] {p} (MOVWconst [hi]) (MOVWconst [lo]) mem) => (LoweredPanicBoundsCC [kind] {PanicBoundsCC{Cx:int64(hi)<<32+int64(uint32(lo)), Cy:p.C}} mem)
   443
   444// Optimizations
   445
   446// Absorb boolean tests into block
   447(NE (FPFlagTrue cmp)  yes no) => (FPT cmp yes no)
   448(NE (FPFlagFalse cmp) yes no) => (FPF cmp yes no)
   449(EQ (FPFlagTrue cmp)  yes no) => (FPF cmp yes no)
   450(EQ (FPFlagFalse cmp) yes no) => (FPT cmp yes no)
   451(NE (XORconst [1] cmp:(SGT _ _))     yes no) => (EQ cmp yes no)
   452(NE (XORconst [1] cmp:(SGTU _ _))    yes no) => (EQ cmp yes no)
   453(NE (XORconst [1] cmp:(SGTconst _))  yes no) => (EQ cmp yes no)
   454(NE (XORconst [1] cmp:(SGTUconst _)) yes no) => (EQ cmp yes no)
   455(NE (XORconst [1] cmp:(SGTzero _))   yes no) => (EQ cmp yes no)
   456(NE (XORconst [1] cmp:(SGTUzero _))  yes no) => (EQ cmp yes no)
   457(EQ (XORconst [1] cmp:(SGT _ _))     yes no) => (NE cmp yes no)
   458(EQ (XORconst [1] cmp:(SGTU _ _))    yes no) => (NE cmp yes no)
   459(EQ (XORconst [1] cmp:(SGTconst _))  yes no) => (NE cmp yes no)
   460(EQ (XORconst [1] cmp:(SGTUconst _)) yes no) => (NE cmp yes no)
   461(EQ (XORconst [1] cmp:(SGTzero _))   yes no) => (NE cmp yes no)
   462(EQ (XORconst [1] cmp:(SGTUzero _))  yes no) => (NE cmp yes no)
   463(NE (SGTUconst [1] x) yes no) => (EQ x yes no)
   464(EQ (SGTUconst [1] x) yes no) => (NE x yes no)
   465(NE (SGTUzero x) yes no) => (NE x yes no)
   466(EQ (SGTUzero x) yes no) => (EQ x yes no)
   467(NE (SGTconst [0] x) yes no) => (LTZ x yes no)
   468(EQ (SGTconst [0] x) yes no) => (GEZ x yes no)
   469(NE (SGTzero x) yes no) => (GTZ x yes no)
   470(EQ (SGTzero x) yes no) => (LEZ x yes no)
   471
   472// fold offset into address
   473(ADDconst [off1] (MOVWaddr [off2] {sym} ptr)) => (MOVWaddr [off1+off2] {sym} ptr)
   474
   475// fold address into load/store
   476(MOVBload  [off1] {sym} x:(ADDconst [off2] ptr) mem) && (is16Bit(int64(off1+off2)) || x.Uses == 1) => (MOVBload  [off1+off2] {sym} ptr mem)
   477(MOVBUload [off1] {sym} x:(ADDconst [off2] ptr) mem) && (is16Bit(int64(off1+off2)) || x.Uses == 1) => (MOVBUload [off1+off2] {sym} ptr mem)
   478(MOVHload  [off1] {sym} x:(ADDconst [off2] ptr) mem) && (is16Bit(int64(off1+off2)) || x.Uses == 1) => (MOVHload  [off1+off2] {sym} ptr mem)
   479(MOVHUload [off1] {sym} x:(ADDconst [off2] ptr) mem) && (is16Bit(int64(off1+off2)) || x.Uses == 1) => (MOVHUload [off1+off2] {sym} ptr mem)
   480(MOVWload  [off1] {sym} x:(ADDconst [off2] ptr) mem) && (is16Bit(int64(off1+off2)) || x.Uses == 1) => (MOVWload  [off1+off2] {sym} ptr mem)
   481(MOVFload  [off1] {sym} x:(ADDconst [off2] ptr) mem) && (is16Bit(int64(off1+off2)) || x.Uses == 1) => (MOVFload  [off1+off2] {sym} ptr mem)
   482(MOVDload  [off1] {sym} x:(ADDconst [off2] ptr) mem) && (is16Bit(int64(off1+off2)) || x.Uses == 1) => (MOVDload  [off1+off2] {sym} ptr mem)
   483
   484(MOVBstore [off1] {sym} x:(ADDconst [off2] ptr) val mem) && (is16Bit(int64(off1+off2)) || x.Uses == 1) => (MOVBstore [off1+off2] {sym} ptr val mem)
   485(MOVHstore [off1] {sym} x:(ADDconst [off2] ptr) val mem) && (is16Bit(int64(off1+off2)) || x.Uses == 1) => (MOVHstore [off1+off2] {sym} ptr val mem)
   486(MOVWstore [off1] {sym} x:(ADDconst [off2] ptr) val mem) && (is16Bit(int64(off1+off2)) || x.Uses == 1) => (MOVWstore [off1+off2] {sym} ptr val mem)
   487(MOVFstore [off1] {sym} x:(ADDconst [off2] ptr) val mem) && (is16Bit(int64(off1+off2)) || x.Uses == 1) => (MOVFstore [off1+off2] {sym} ptr val mem)
   488(MOVDstore [off1] {sym} x:(ADDconst [off2] ptr) val mem) && (is16Bit(int64(off1+off2)) || x.Uses == 1) => (MOVDstore [off1+off2] {sym} ptr val mem)
   489
   490(MOVBstorezero [off1] {sym} x:(ADDconst [off2] ptr) mem) && (is16Bit(int64(off1+off2)) || x.Uses == 1) => (MOVBstorezero [off1+off2] {sym} ptr mem)
   491(MOVHstorezero [off1] {sym} x:(ADDconst [off2] ptr) mem) && (is16Bit(int64(off1+off2)) || x.Uses == 1) => (MOVHstorezero [off1+off2] {sym} ptr mem)
   492(MOVWstorezero [off1] {sym} x:(ADDconst [off2] ptr) mem) && (is16Bit(int64(off1+off2)) || x.Uses == 1) => (MOVWstorezero [off1+off2] {sym} ptr mem)
   493
   494(MOVBload [off1] {sym1} (MOVWaddr [off2] {sym2} ptr) mem) && canMergeSym(sym1,sym2)  =>
   495	(MOVBload [off1+off2] {mergeSym(sym1,sym2)} ptr mem)
   496(MOVBUload [off1] {sym1} (MOVWaddr [off2] {sym2} ptr) mem) && canMergeSym(sym1,sym2) =>
   497	(MOVBUload [off1+off2] {mergeSym(sym1,sym2)} ptr mem)
   498(MOVHload [off1] {sym1} (MOVWaddr [off2] {sym2} ptr) mem) && canMergeSym(sym1,sym2)  =>
   499	(MOVHload [off1+off2] {mergeSym(sym1,sym2)} ptr mem)
   500(MOVHUload [off1] {sym1} (MOVWaddr [off2] {sym2} ptr) mem) && canMergeSym(sym1,sym2) =>
   501	(MOVHUload [off1+off2] {mergeSym(sym1,sym2)} ptr mem)
   502(MOVWload [off1] {sym1} (MOVWaddr [off2] {sym2} ptr) mem) && canMergeSym(sym1,sym2) =>
   503	(MOVWload [off1+off2] {mergeSym(sym1,sym2)} ptr mem)
   504(MOVFload [off1] {sym1} (MOVWaddr [off2] {sym2} ptr) mem) && canMergeSym(sym1,sym2) =>
   505	(MOVFload [off1+off2] {mergeSym(sym1,sym2)} ptr mem)
   506(MOVDload [off1] {sym1} (MOVWaddr [off2] {sym2} ptr) mem) && canMergeSym(sym1,sym2) =>
   507	(MOVDload [off1+off2] {mergeSym(sym1,sym2)} ptr mem)
   508
   509(MOVBstore [off1] {sym1} (MOVWaddr [off2] {sym2} ptr) val mem) && canMergeSym(sym1,sym2) =>
   510	(MOVBstore [off1+off2] {mergeSym(sym1,sym2)} ptr val mem)
   511(MOVHstore [off1] {sym1} (MOVWaddr [off2] {sym2} ptr) val mem) && canMergeSym(sym1,sym2) =>
   512	(MOVHstore [off1+off2] {mergeSym(sym1,sym2)} ptr val mem)
   513(MOVWstore [off1] {sym1} (MOVWaddr [off2] {sym2} ptr) val mem) && canMergeSym(sym1,sym2) =>
   514	(MOVWstore [off1+off2] {mergeSym(sym1,sym2)} ptr val mem)
   515(MOVFstore [off1] {sym1} (MOVWaddr [off2] {sym2} ptr) val mem) && canMergeSym(sym1,sym2) =>
   516	(MOVFstore [off1+off2] {mergeSym(sym1,sym2)} ptr val mem)
   517(MOVDstore [off1] {sym1} (MOVWaddr [off2] {sym2} ptr) val mem) && canMergeSym(sym1,sym2) =>
   518	(MOVDstore [off1+off2] {mergeSym(sym1,sym2)} ptr val mem)
   519(MOVBstorezero [off1] {sym1} (MOVWaddr [off2] {sym2} ptr) mem) && canMergeSym(sym1,sym2) =>
   520	(MOVBstorezero [off1+off2] {mergeSym(sym1,sym2)} ptr mem)
   521(MOVHstorezero [off1] {sym1} (MOVWaddr [off2] {sym2} ptr) mem) && canMergeSym(sym1,sym2) =>
   522	(MOVHstorezero [off1+off2] {mergeSym(sym1,sym2)} ptr mem)
   523(MOVWstorezero [off1] {sym1} (MOVWaddr [off2] {sym2} ptr) mem) && canMergeSym(sym1,sym2) =>
   524	(MOVWstorezero [off1+off2] {mergeSym(sym1,sym2)} ptr mem)
   525
   526// replace load from same location as preceding store with zero/sign extension (or copy in case of full width)
   527(MOVBload [off] {sym} ptr (MOVBstore [off2] {sym2} ptr2 x _))  && sym == sym2 && off == off2 && isSamePtr(ptr, ptr2) => (MOVBreg x)
   528(MOVBUload [off] {sym} ptr (MOVBstore [off2] {sym2} ptr2 x _)) && sym == sym2 && off == off2 && isSamePtr(ptr, ptr2) => (MOVBUreg x)
   529(MOVHload [off] {sym} ptr (MOVHstore [off2] {sym2} ptr2 x _))  && sym == sym2 && off == off2 && isSamePtr(ptr, ptr2) => (MOVHreg x)
   530(MOVHUload [off] {sym} ptr (MOVHstore [off2] {sym2} ptr2 x _)) && sym == sym2 && off == off2 && isSamePtr(ptr, ptr2) => (MOVHUreg x)
   531(MOVWload [off] {sym} ptr (MOVWstore [off2] {sym2} ptr2 x _))  && sym == sym2 && off == off2 && isSamePtr(ptr, ptr2) => x
   532(MOVFload [off] {sym} ptr (MOVFstore [off2] {sym2} ptr2 x _))  && sym == sym2 && off == off2 && isSamePtr(ptr, ptr2) => x
   533(MOVDload [off] {sym} ptr (MOVDstore [off2] {sym2} ptr2 x _))  && sym == sym2 && off == off2 && isSamePtr(ptr, ptr2) => x
   534
   535// store zero
   536(MOVBstore [off] {sym} ptr (MOVWconst [0]) mem) => (MOVBstorezero [off] {sym} ptr mem)
   537(MOVHstore [off] {sym} ptr (MOVWconst [0]) mem) => (MOVHstorezero [off] {sym} ptr mem)
   538(MOVWstore [off] {sym} ptr (MOVWconst [0]) mem) => (MOVWstorezero [off] {sym} ptr mem)
   539
   540// don't extend after proper load
   541(MOVBreg x:(MOVBload _ _))   => (MOVWreg x)
   542(MOVBUreg x:(MOVBUload _ _)) => (MOVWreg x)
   543(MOVHreg x:(MOVBload _ _))   => (MOVWreg x)
   544(MOVHreg x:(MOVBUload _ _))  => (MOVWreg x)
   545(MOVHreg x:(MOVHload _ _))   => (MOVWreg x)
   546(MOVHUreg x:(MOVBUload _ _)) => (MOVWreg x)
   547(MOVHUreg x:(MOVHUload _ _)) => (MOVWreg x)
   548
   549// fold double extensions
   550(MOVBreg x:(MOVBreg _))   => (MOVWreg x)
   551(MOVBUreg x:(MOVBUreg _)) => (MOVWreg x)
   552(MOVHreg x:(MOVBreg _))   => (MOVWreg x)
   553(MOVHreg x:(MOVBUreg _))  => (MOVWreg x)
   554(MOVHreg x:(MOVHreg _))   => (MOVWreg x)
   555(MOVHUreg x:(MOVBUreg _)) => (MOVWreg x)
   556(MOVHUreg x:(MOVHUreg _)) => (MOVWreg x)
   557
   558// sign extended loads
   559// Note: The combined instruction must end up in the same block
   560// as the original load. If not, we end up making a value with
   561// memory type live in two different blocks, which can lead to
   562// multiple memory values alive simultaneously.
   563// Make sure we don't combine these ops if the load has another use.
   564// This prevents a single load from being split into multiple loads
   565// which then might return different values.  See test/atomicload.go.
   566(MOVBreg <t> x:(MOVBUload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVBload  <t> [off] {sym} ptr mem)
   567(MOVBUreg <t> x:(MOVBload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVBUload <t> [off] {sym} ptr mem)
   568(MOVHreg <t> x:(MOVHUload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVHload  <t> [off] {sym} ptr mem)
   569(MOVHUreg <t> x:(MOVHload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVHUload <t> [off] {sym} ptr mem)
   570
   571// fold extensions and ANDs together
   572(MOVBUreg (ANDconst [c] x)) => (ANDconst [c&0xff] x)
   573(MOVHUreg (ANDconst [c] x)) => (ANDconst [c&0xffff] x)
   574(MOVBreg (ANDconst [c] x)) && c & 0x80   == 0 => (ANDconst [c&0x7f] x)
   575(MOVHreg (ANDconst [c] x)) && c & 0x8000 == 0 => (ANDconst [c&0x7fff] x)
   576
   577// don't extend before store
   578(MOVBstore [off] {sym} ptr (MOVBreg x)  mem) => (MOVBstore [off] {sym} ptr x mem)
   579(MOVBstore [off] {sym} ptr (MOVBUreg x) mem) => (MOVBstore [off] {sym} ptr x mem)
   580(MOVBstore [off] {sym} ptr (MOVHreg x)  mem) => (MOVBstore [off] {sym} ptr x mem)
   581(MOVBstore [off] {sym} ptr (MOVHUreg x) mem) => (MOVBstore [off] {sym} ptr x mem)
   582(MOVBstore [off] {sym} ptr (MOVWreg x)  mem) => (MOVBstore [off] {sym} ptr x mem)
   583(MOVHstore [off] {sym} ptr (MOVHreg x)  mem) => (MOVHstore [off] {sym} ptr x mem)
   584(MOVHstore [off] {sym} ptr (MOVHUreg x) mem) => (MOVHstore [off] {sym} ptr x mem)
   585(MOVHstore [off] {sym} ptr (MOVWreg x)  mem) => (MOVHstore [off] {sym} ptr x mem)
   586(MOVWstore [off] {sym} ptr (MOVWreg x)  mem) => (MOVWstore [off] {sym} ptr x mem)
   587
   588// if a register move has only 1 use, just use the same register without emitting instruction
   589// MOVWnop doesn't emit instruction, only for ensuring the type.
   590(MOVWreg x) && x.Uses == 1 => (MOVWnop x)
   591
   592// TODO: we should be able to get rid of MOVWnop all together.
   593// But for now, this is enough to get rid of lots of them.
   594(MOVWnop (MOVWconst [c])) => (MOVWconst [c])
   595
   596// fold constant into arithmetic ops
   597(ADD x (MOVWconst <t> [c])) && !t.IsPtr() => (ADDconst [c] x)
   598(SUB x (MOVWconst [c])) => (SUBconst [c] x)
   599(AND x (MOVWconst [c])) => (ANDconst [c] x)
   600(OR  x (MOVWconst [c])) => (ORconst  [c] x)
   601(XOR x (MOVWconst [c])) => (XORconst [c] x)
   602(NOR x (MOVWconst [c])) => (NORconst [c] x)
   603
   604(SLL x (MOVWconst [c])) => (SLLconst x [c&31])
   605(SRL x (MOVWconst [c])) => (SRLconst x [c&31])
   606(SRA x (MOVWconst [c])) => (SRAconst x [c&31])
   607
   608(SGT  (MOVWconst [c]) x) => (SGTconst  [c] x)
   609(SGTU (MOVWconst [c]) x) => (SGTUconst [c] x)
   610(SGT x  (MOVWconst [0])) => (SGTzero x)
   611(SGTU x (MOVWconst [0])) => (SGTUzero x)
   612
   613// mul with constant
   614(Select1 (MULTU (MOVWconst [0])  _ )) => (MOVWconst [0])
   615(Select0 (MULTU (MOVWconst [0])  _ )) => (MOVWconst [0])
   616(Select1 (MULTU (MOVWconst [1])  x )) => x
   617(Select0 (MULTU (MOVWconst [1])  _ )) => (MOVWconst [0])
   618(Select1 (MULTU (MOVWconst [-1]) x )) => (NEG <x.Type> x)
   619(Select0 (MULTU (MOVWconst [-1]) x )) => (CMOVZ (ADDconst <x.Type> [-1] x) (MOVWconst [0]) x)
   620(Select1 (MULTU (MOVWconst [c])  x )) && isUnsignedPowerOfTwo(uint32(c)) => (SLLconst [int32(log32u(uint32(c)))] x)
   621(Select0 (MULTU (MOVWconst [c])  x )) && isUnsignedPowerOfTwo(uint32(c)) => (SRLconst [int32(32-log32u(uint32(c)))] x)
   622
   623(MUL (MOVWconst [0])  _ ) => (MOVWconst [0])
   624(MUL (MOVWconst [1])  x ) => x
   625(MUL (MOVWconst [-1]) x ) => (NEG x)
   626(MUL (MOVWconst [c]) x ) && isUnsignedPowerOfTwo(uint32(c)) => (SLLconst [int32(log32u(uint32(c)))] x)
   627
   628// generic simplifications
   629(ADD x (NEG y)) => (SUB x y)
   630(SUB x (NEG y)) => (ADD x y)
   631(SUB x x) => (MOVWconst [0])
   632(SUB (MOVWconst [0]) x) => (NEG x)
   633(AND x x) => x
   634(OR  x x) => x
   635(XOR x x) => (MOVWconst [0])
   636(NEG (SUB x y)) => (SUB y x)
   637(NEG (NEG x)) => x
   638
   639// miscellaneous patterns generated by dec64
   640(AND (SGTUconst [1] x) (SGTUconst [1] y)) =>  (SGTUconst [1] (OR <x.Type> x y))
   641(OR (SGTUzero x) (SGTUzero y)) =>  (SGTUzero (OR <x.Type> x y))
   642
   643// remove redundant *const ops
   644(ADDconst [0]  x) => x
   645(SUBconst [0]  x) => x
   646(ANDconst [0]  _) => (MOVWconst [0])
   647(ANDconst [-1] x) => x
   648(ORconst  [0]  x) => x
   649(ORconst  [-1] _) => (MOVWconst [-1])
   650(XORconst [0]  x) => x
   651(XORconst [-1] x) => (NORconst [0] x)
   652
   653// generic constant folding
   654(ADDconst [c] (MOVWconst [d]))  => (MOVWconst [int32(c+d)])
   655(ADDconst [c] (ADDconst [d] x)) => (ADDconst [c+d] x)
   656(ADDconst [c] (SUBconst [d] x)) => (ADDconst [c-d] x)
   657(SUBconst [c] (MOVWconst [d]))  => (MOVWconst [d-c])
   658(SUBconst [c] (SUBconst [d] x)) => (ADDconst [-c-d] x)
   659(SUBconst [c] (ADDconst [d] x)) => (ADDconst [-c+d] x)
   660(SLLconst [c] (MOVWconst [d]))  => (MOVWconst [d<<uint32(c)])
   661(SRLconst [c] (MOVWconst [d]))  => (MOVWconst [int32(uint32(d)>>uint32(c))])
   662(SRAconst [c] (MOVWconst [d]))  => (MOVWconst [d>>uint32(c)])
   663(MUL (MOVWconst [c]) (MOVWconst [d])) => (MOVWconst [c*d])
   664(Select1 (MULTU  (MOVWconst [c]) (MOVWconst [d]))) => (MOVWconst [int32(uint32(c)*uint32(d))])
   665(Select0 (MULTU  (MOVWconst [c]) (MOVWconst [d]))) => (MOVWconst [int32((int64(uint32(c))*int64(uint32(d)))>>32)])
   666(Select1 (DIV  (MOVWconst [c]) (MOVWconst [d]))) && d != 0 => (MOVWconst [c/d])
   667(Select1 (DIVU (MOVWconst [c]) (MOVWconst [d]))) && d != 0 => (MOVWconst [int32(uint32(c)/uint32(d))])
   668(Select0 (DIV  (MOVWconst [c]) (MOVWconst [d]))) && d != 0 => (MOVWconst [c%d])
   669(Select0 (DIVU (MOVWconst [c]) (MOVWconst [d]))) && d != 0 => (MOVWconst [int32(uint32(c)%uint32(d))])
   670(ANDconst [c] (MOVWconst [d])) => (MOVWconst [c&d])
   671(ANDconst [c] (ANDconst [d] x)) => (ANDconst [c&d] x)
   672(ORconst [c] (MOVWconst [d])) => (MOVWconst [c|d])
   673(ORconst [c] (ORconst [d] x)) => (ORconst [c|d] x)
   674(XORconst [c] (MOVWconst [d])) => (MOVWconst [c^d])
   675(XORconst [c] (XORconst [d] x)) => (XORconst [c^d] x)
   676(NORconst [c] (MOVWconst [d])) => (MOVWconst [^(c|d)])
   677(NEG (MOVWconst [c])) => (MOVWconst [-c])
   678(MOVBreg  (MOVWconst [c])) => (MOVWconst [int32(int8(c))])
   679(MOVBUreg (MOVWconst [c])) => (MOVWconst [int32(uint8(c))])
   680(MOVHreg  (MOVWconst [c])) => (MOVWconst [int32(int16(c))])
   681(MOVHUreg (MOVWconst [c])) => (MOVWconst [int32(uint16(c))])
   682(MOVWreg  (MOVWconst [c])) => (MOVWconst [c])
   683
   684// constant comparisons
   685(SGTconst  [c] (MOVWconst [d])) && c >  d => (MOVWconst [1])
   686(SGTconst  [c] (MOVWconst [d])) && c <= d => (MOVWconst [0])
   687(SGTUconst [c] (MOVWconst [d])) && uint32(c) >  uint32(d) => (MOVWconst [1])
   688(SGTUconst [c] (MOVWconst [d])) && uint32(c) <= uint32(d) => (MOVWconst [0])
   689(SGTzero (MOVWconst [d])) && d >  0 => (MOVWconst [1])
   690(SGTzero (MOVWconst [d])) && d <= 0 => (MOVWconst [0])
   691(SGTUzero (MOVWconst [d])) && d != 0 => (MOVWconst [1])
   692(SGTUzero (MOVWconst [d])) && d == 0 => (MOVWconst [0])
   693
   694// other known comparisons
   695(SGTconst [c] (MOVBreg _)) && 0x7f < c   => (MOVWconst [1])
   696(SGTconst [c] (MOVBreg _)) && c <= -0x80 => (MOVWconst [0])
   697(SGTconst [c] (MOVBUreg _)) && 0xff < c  => (MOVWconst [1])
   698(SGTconst [c] (MOVBUreg _)) && c < 0     => (MOVWconst [0])
   699(SGTUconst [c] (MOVBUreg _)) && 0xff < uint32(c) => (MOVWconst [1])
   700(SGTconst [c] (MOVHreg _)) && 0x7fff < c => (MOVWconst [1])
   701(SGTconst [c] (MOVHreg _)) && c <= -0x8000 => (MOVWconst [0])
   702(SGTconst [c] (MOVHUreg _)) && 0xffff < c => (MOVWconst [1])
   703(SGTconst [c] (MOVHUreg _)) && c < 0 => (MOVWconst [0])
   704(SGTUconst [c] (MOVHUreg _)) && 0xffff < uint32(c) => (MOVWconst [1])
   705(SGTconst [c] (ANDconst [m] _)) && 0 <= m && m < c => (MOVWconst [1])
   706(SGTUconst [c] (ANDconst [m] _)) && uint32(m) < uint32(c) => (MOVWconst [1])
   707(SGTconst [c] (SRLconst _ [d])) && 0 <= c && uint32(d) <= 31 && 0xffffffff>>uint32(d) < uint32(c) => (MOVWconst [1])
   708(SGTUconst [c] (SRLconst _ [d])) && uint32(d) <= 31 && 0xffffffff>>uint32(d) < uint32(c) => (MOVWconst [1])
   709
   710// absorb constants into branches
   711(EQ  (MOVWconst [0]) yes no) => (First yes no)
   712(EQ  (MOVWconst [c]) yes no) && c != 0 => (First no yes)
   713(NE  (MOVWconst [0]) yes no) => (First no yes)
   714(NE  (MOVWconst [c]) yes no) && c != 0 => (First yes no)
   715(LTZ (MOVWconst [c]) yes no) && c <  0 => (First yes no)
   716(LTZ (MOVWconst [c]) yes no) && c >= 0 => (First no yes)
   717(LEZ (MOVWconst [c]) yes no) && c <= 0 => (First yes no)
   718(LEZ (MOVWconst [c]) yes no) && c >  0 => (First no yes)
   719(GTZ (MOVWconst [c]) yes no) && c >  0 => (First yes no)
   720(GTZ (MOVWconst [c]) yes no) && c <= 0 => (First no yes)
   721(GEZ (MOVWconst [c]) yes no) && c >= 0 => (First yes no)
   722(GEZ (MOVWconst [c]) yes no) && c <  0 => (First no yes)
   723
   724// conditional move
   725(CMOVZ _ f (MOVWconst [0])) => f
   726(CMOVZ a _ (MOVWconst [c])) && c!=0 => a
   727(CMOVZzero _ (MOVWconst [0])) => (MOVWconst [0])
   728(CMOVZzero a (MOVWconst [c])) && c!=0 => a
   729(CMOVZ a (MOVWconst [0]) c) => (CMOVZzero a c)
   730
   731// atomic
   732(LoweredAtomicStore32 ptr (MOVWconst [0]) mem) => (LoweredAtomicStorezero ptr mem)
   733(LoweredAtomicAdd ptr (MOVWconst [c]) mem) && is16Bit(int64(c)) => (LoweredAtomicAddconst [c] ptr mem)
   734

View as plain text