...

Text file src/cmd/compile/internal/ssa/_gen/RISCV64.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// Lowering arithmetic
     6(Add(Ptr|64|32|16|8) ...) => (ADD ...)
     7(Add(64|32)F ...) => (FADD(D|S) ...)
     8
     9(Sub(Ptr|64|32|16|8) ...) => (SUB ...)
    10(Sub(64|32)F ...) => (FSUB(D|S) ...)
    11
    12(Mul64 ...) => (MUL  ...)
    13(Mul64uhilo ...) => (LoweredMuluhilo ...)
    14(Mul64uover ...) => (LoweredMuluover ...)
    15(Mul32 ...) => (MULW ...)
    16(Mul16 x y) => (MULW (SignExt16to32 x) (SignExt16to32 y))
    17(Mul8 x y)  => (MULW (SignExt8to32 x)  (SignExt8to32 y))
    18(Mul(64|32)F ...) => (FMUL(D|S) ...)
    19
    20(Div(64|32)F ...) => (FDIV(D|S) ...)
    21
    22(Div64 x y [false])  => (DIV x y)
    23(Div64u ...) => (DIVU ...)
    24(Div32 x y [false])  => (DIVW x y)
    25(Div32u ...) => (DIVUW ...)
    26(Div16 x y [false])  => (DIVW  (SignExt16to32 x) (SignExt16to32 y))
    27(Div16u x y) => (DIVUW (ZeroExt16to32 x) (ZeroExt16to32 y))
    28(Div8 x y)   => (DIVW  (SignExt8to32 x)  (SignExt8to32 y))
    29(Div8u x y)  => (DIVUW (ZeroExt8to32 x)  (ZeroExt8to32 y))
    30
    31(Hmul64 ...)  => (MULH  ...)
    32(Hmul64u ...) => (MULHU ...)
    33(Hmul32 x y)  => (SRAI [32] (MUL  (SignExt32to64 x) (SignExt32to64 y)))
    34(Hmul32u x y) => (SRLI [32] (MUL  (ZeroExt32to64 x) (ZeroExt32to64 y)))
    35
    36(Select0 (Add64carry x y c)) => (ADD (ADD <typ.UInt64> x y) c)
    37(Select1 (Add64carry x y c)) =>
    38	(OR (SLTU <typ.UInt64> s:(ADD <typ.UInt64> x y) x) (SLTU <typ.UInt64> (ADD <typ.UInt64> s c) s))
    39
    40(Select0 (Sub64borrow x y c)) => (SUB (SUB <typ.UInt64> x y) c)
    41(Select1 (Sub64borrow x y c)) =>
    42	(OR (SLTU <typ.UInt64> x s:(SUB <typ.UInt64> x y)) (SLTU <typ.UInt64> s (SUB <typ.UInt64> s c)))
    43
    44// (x + y) / 2 => (x / 2) + (y / 2) + (x & y & 1)
    45(Avg64u <t> x y) => (ADD (ADD <t> (SRLI <t> [1] x) (SRLI <t> [1] y)) (ANDI <t> [1] (AND <t> x y)))
    46
    47(Mod64 x y [false])  => (REM x y)
    48(Mod64u ...) => (REMU  ...)
    49(Mod32 x y [false])  => (REMW x y)
    50(Mod32u ...) => (REMUW ...)
    51(Mod16 x y [false])  => (REMW  (SignExt16to32 x) (SignExt16to32 y))
    52(Mod16u x y) => (REMUW (ZeroExt16to32 x) (ZeroExt16to32 y))
    53(Mod8 x y)   => (REMW  (SignExt8to32  x) (SignExt8to32  y))
    54(Mod8u x y)  => (REMUW (ZeroExt8to32  x) (ZeroExt8to32  y))
    55
    56(And(64|32|16|8) ...) => (AND ...)
    57(Or(64|32|16|8) ...) => (OR ...)
    58(Xor(64|32|16|8) ...) => (XOR ...)
    59
    60(Neg(64|32|16|8) ...) => (NEG ...)
    61(Neg(64|32)F ...) => (FNEG(D|S) ...)
    62
    63(Com(64|32|16|8) ...) => (NOT ...)
    64
    65
    66(Sqrt ...) => (FSQRTD ...)
    67(Sqrt32 ...) => (FSQRTS ...)
    68
    69(Copysign ...) => (FSGNJD ...)
    70
    71(Abs ...) => (FABSD ...)
    72
    73(FMA ...) => (FMADDD ...)
    74
    75(Min(64|32)F ...) => (LoweredFMIN(D|S) ...)
    76(Max(64|32)F ...) => (LoweredFMAX(D|S) ...)
    77
    78// Sign and zero extension.
    79
    80(SignExt8to16  ...) => (MOVBreg ...)
    81(SignExt8to32  ...) => (MOVBreg ...)
    82(SignExt8to64  ...) => (MOVBreg ...)
    83(SignExt16to32 ...) => (MOVHreg ...)
    84(SignExt16to64 ...) => (MOVHreg ...)
    85(SignExt32to64 ...) => (MOVWreg ...)
    86
    87(ZeroExt8to16  ...) => (MOVBUreg ...)
    88(ZeroExt8to32  ...) => (MOVBUreg ...)
    89(ZeroExt8to64  ...) => (MOVBUreg ...)
    90(ZeroExt16to32 ...) => (MOVHUreg ...)
    91(ZeroExt16to64 ...) => (MOVHUreg ...)
    92(ZeroExt32to64 ...) => (MOVWUreg ...)
    93
    94(Cvt32to32F ...) => (FCVTSW ...)
    95(Cvt32to64F ...) => (FCVTDW ...)
    96(Cvt64to32F ...) => (FCVTSL ...)
    97(Cvt64to64F ...) => (FCVTDL ...)
    98
    99(Cvt32Fto32 ...) => (FCVTWS ...)
   100(Cvt32Fto64 ...) => (FCVTLS ...)
   101(Cvt64Fto32 ...) => (FCVTWD ...)
   102(Cvt64Fto64 ...) => (FCVTLD ...)
   103
   104(Cvt32Fto64F ...) => (FCVTDS ...)
   105(Cvt64Fto32F ...) => (FCVTSD ...)
   106
   107(CvtBoolToUint8 ...) => (Copy ...)
   108
   109(Round(32|64)F ...) => (LoweredRound(32|64)F ...)
   110
   111(Slicemask <t> x) => (SRAI [63] (NEG <t> x))
   112
   113// Truncations
   114// We ignore the unused high parts of registers, so truncates are just copies.
   115(Trunc16to8  ...) => (Copy ...)
   116(Trunc32to8  ...) => (Copy ...)
   117(Trunc32to16 ...) => (Copy ...)
   118(Trunc64to8  ...) => (Copy ...)
   119(Trunc64to16 ...) => (Copy ...)
   120(Trunc64to32 ...) => (Copy ...)
   121
   122// Shifts
   123
   124// SLL only considers the bottom 6 bits of y. If y > 64, the result should
   125// always be 0.
   126//
   127// Breaking down the operation:
   128//
   129// (SLL x y) generates x << (y & 63).
   130//
   131// If y < 64, this is the value we want. Otherwise, we want zero.
   132//
   133// So, we AND with -1 * uint64(y < 64), which is 0xfffff... if y < 64 and 0 otherwise.
   134(Lsh8x8   <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg8  <t> (SLTIU <t> [64] (ZeroExt8to64  y))))
   135(Lsh8x16  <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg8  <t> (SLTIU <t> [64] (ZeroExt16to64 y))))
   136(Lsh8x32  <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg8  <t> (SLTIU <t> [64] (ZeroExt32to64 y))))
   137(Lsh8x64  <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg8  <t> (SLTIU <t> [64] y)))
   138(Lsh16x8  <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg16 <t> (SLTIU <t> [64] (ZeroExt8to64  y))))
   139(Lsh16x16 <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg16 <t> (SLTIU <t> [64] (ZeroExt16to64 y))))
   140(Lsh16x32 <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg16 <t> (SLTIU <t> [64] (ZeroExt32to64 y))))
   141(Lsh16x64 <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg16 <t> (SLTIU <t> [64] y)))
   142(Lsh32x8  <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg32 <t> (SLTIU <t> [64] (ZeroExt8to64  y))))
   143(Lsh32x16 <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg32 <t> (SLTIU <t> [64] (ZeroExt16to64 y))))
   144(Lsh32x32 <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg32 <t> (SLTIU <t> [64] (ZeroExt32to64 y))))
   145(Lsh32x64 <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg32 <t> (SLTIU <t> [64] y)))
   146(Lsh64x8  <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg64 <t> (SLTIU <t> [64] (ZeroExt8to64  y))))
   147(Lsh64x16 <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg64 <t> (SLTIU <t> [64] (ZeroExt16to64 y))))
   148(Lsh64x32 <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg64 <t> (SLTIU <t> [64] (ZeroExt32to64 y))))
   149(Lsh64x64 <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg64 <t> (SLTIU <t> [64] y)))
   150
   151(Lsh8x(64|32|16|8)  x y) && shiftIsBounded(v) => (SLL x y)
   152(Lsh16x(64|32|16|8) x y) && shiftIsBounded(v) => (SLL x y)
   153(Lsh32x(64|32|16|8) x y) && shiftIsBounded(v) => (SLL x y)
   154(Lsh64x(64|32|16|8) x y) && shiftIsBounded(v) => (SLL x y)
   155
   156// SRL only considers the bottom 6 bits of y, similarly SRLW only considers the
   157// bottom 5 bits of y. Ensure that the result is always zero if the shift exceeds
   158// the maximum value. See Lsh above for a detailed description.
   159(Rsh8Ux8   <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t> (ZeroExt8to64  x) y) (Neg8  <t> (SLTIU <t> [64] (ZeroExt8to64  y))))
   160(Rsh8Ux16  <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t> (ZeroExt8to64  x) y) (Neg8  <t> (SLTIU <t> [64] (ZeroExt16to64 y))))
   161(Rsh8Ux32  <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t> (ZeroExt8to64  x) y) (Neg8  <t> (SLTIU <t> [64] (ZeroExt32to64 y))))
   162(Rsh8Ux64  <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t> (ZeroExt8to64  x) y) (Neg8  <t> (SLTIU <t> [64] y)))
   163(Rsh16Ux8  <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t> (ZeroExt16to64 x) y) (Neg16 <t> (SLTIU <t> [64] (ZeroExt8to64  y))))
   164(Rsh16Ux16 <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t> (ZeroExt16to64 x) y) (Neg16 <t> (SLTIU <t> [64] (ZeroExt16to64 y))))
   165(Rsh16Ux32 <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t> (ZeroExt16to64 x) y) (Neg16 <t> (SLTIU <t> [64] (ZeroExt32to64 y))))
   166(Rsh16Ux64 <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t> (ZeroExt16to64 x) y) (Neg16 <t> (SLTIU <t> [64] y)))
   167(Rsh32Ux8  <t> x y) && !shiftIsBounded(v) => (AND (SRLW <t>  x                y) (Neg32 <t> (SLTIU <t> [32] (ZeroExt8to64  y))))
   168(Rsh32Ux16 <t> x y) && !shiftIsBounded(v) => (AND (SRLW <t>  x                y) (Neg32 <t> (SLTIU <t> [32] (ZeroExt16to64 y))))
   169(Rsh32Ux32 <t> x y) && !shiftIsBounded(v) => (AND (SRLW <t>  x                y) (Neg32 <t> (SLTIU <t> [32] (ZeroExt32to64 y))))
   170(Rsh32Ux64 <t> x y) && !shiftIsBounded(v) => (AND (SRLW <t>  x                y) (Neg32 <t> (SLTIU <t> [32] y)))
   171(Rsh64Ux8  <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t>  x                y) (Neg64 <t> (SLTIU <t> [64] (ZeroExt8to64  y))))
   172(Rsh64Ux16 <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t>  x                y) (Neg64 <t> (SLTIU <t> [64] (ZeroExt16to64 y))))
   173(Rsh64Ux32 <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t>  x                y) (Neg64 <t> (SLTIU <t> [64] (ZeroExt32to64 y))))
   174(Rsh64Ux64 <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t>  x                y) (Neg64 <t> (SLTIU <t> [64] y)))
   175
   176(Rsh8Ux(64|32|16|8)  x y) && shiftIsBounded(v) => (SRL  (ZeroExt8to64  x) y)
   177(Rsh16Ux(64|32|16|8) x y) && shiftIsBounded(v) => (SRL  (ZeroExt16to64 x) y)
   178(Rsh32Ux(64|32|16|8) x y) && shiftIsBounded(v) => (SRLW x                 y)
   179(Rsh64Ux(64|32|16|8) x y) && shiftIsBounded(v) => (SRL  x                 y)
   180
   181// SRA only considers the bottom 6 bits of y, similarly SRAW only considers the
   182// bottom 5 bits. If y is greater than the maximum value (either 63 or 31
   183// depending on the instruction),  the result of the shift should be either 0
   184// or -1 based on the sign bit of x.
   185//
   186// We implement this by performing the max shift (-1) if y > the maximum value.
   187//
   188// We OR (uint64(y < 64) - 1) into y before passing it to SRA. This leaves
   189// us with -1 (0xffff...) if y >= 64.  Similarly, we OR (uint64(y < 32) - 1) into y
   190// before passing it to SRAW.
   191//
   192// We don't need to sign-extend the OR result, as it will be at minimum 8 bits,
   193// more than the 5 or 6 bits SRAW and SRA care about.
   194(Rsh8x8   <t> x y) && !shiftIsBounded(v) => (SRA  <t> (SignExt8to64  x) (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] (ZeroExt8to64  y)))))
   195(Rsh8x16  <t> x y) && !shiftIsBounded(v) => (SRA  <t> (SignExt8to64  x) (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] (ZeroExt16to64 y)))))
   196(Rsh8x32  <t> x y) && !shiftIsBounded(v) => (SRA  <t> (SignExt8to64  x) (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] (ZeroExt32to64 y)))))
   197(Rsh8x64  <t> x y) && !shiftIsBounded(v) => (SRA  <t> (SignExt8to64  x) (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] y))))
   198(Rsh16x8  <t> x y) && !shiftIsBounded(v) => (SRA  <t> (SignExt16to64 x) (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] (ZeroExt8to64  y)))))
   199(Rsh16x16 <t> x y) && !shiftIsBounded(v) => (SRA  <t> (SignExt16to64 x) (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] (ZeroExt16to64 y)))))
   200(Rsh16x32 <t> x y) && !shiftIsBounded(v) => (SRA  <t> (SignExt16to64 x) (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] (ZeroExt32to64 y)))))
   201(Rsh16x64 <t> x y) && !shiftIsBounded(v) => (SRA  <t> (SignExt16to64 x) (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] y))))
   202(Rsh32x8  <t> x y) && !shiftIsBounded(v) => (SRAW <t> x                 (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [32] (ZeroExt8to64  y)))))
   203(Rsh32x16 <t> x y) && !shiftIsBounded(v) => (SRAW <t> x                 (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [32] (ZeroExt16to64 y)))))
   204(Rsh32x32 <t> x y) && !shiftIsBounded(v) => (SRAW <t> x                 (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [32] (ZeroExt32to64 y)))))
   205(Rsh32x64 <t> x y) && !shiftIsBounded(v) => (SRAW <t> x                 (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [32] y))))
   206(Rsh64x8  <t> x y) && !shiftIsBounded(v) => (SRA  <t> x                 (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] (ZeroExt8to64  y)))))
   207(Rsh64x16 <t> x y) && !shiftIsBounded(v) => (SRA  <t> x                 (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] (ZeroExt16to64 y)))))
   208(Rsh64x32 <t> x y) && !shiftIsBounded(v) => (SRA  <t> x                 (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] (ZeroExt32to64 y)))))
   209(Rsh64x64 <t> x y) && !shiftIsBounded(v) => (SRA  <t> x                 (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] y))))
   210
   211(Rsh8x(64|32|16|8)  x y) && shiftIsBounded(v) => (SRA  (SignExt8to64  x) y)
   212(Rsh16x(64|32|16|8) x y) && shiftIsBounded(v) => (SRA  (SignExt16to64 x) y)
   213(Rsh32x(64|32|16|8) x y) && shiftIsBounded(v) => (SRAW  x                y)
   214(Rsh64x(64|32|16|8) x y) && shiftIsBounded(v) => (SRA   x                y)
   215
   216// Rotates.
   217(RotateLeft8  <t> x y) => (OR (SLL  <t> x (ANDI [7]  <y.Type> y)) (SRL <t> (ZeroExt8to64  x) (ANDI [7]  <y.Type> (NEG <y.Type> y))))
   218(RotateLeft16 <t> x y) => (OR (SLL  <t> x (ANDI [15] <y.Type> y)) (SRL <t> (ZeroExt16to64 x) (ANDI [15] <y.Type> (NEG <y.Type> y))))
   219(RotateLeft32 ...) => (ROLW ...)
   220(RotateLeft64 ...) => (ROL  ...)
   221
   222(Less64  ...) => (SLT  ...)
   223(Less32  x y) => (SLT  (SignExt32to64 x) (SignExt32to64 y))
   224(Less16  x y) => (SLT  (SignExt16to64 x) (SignExt16to64 y))
   225(Less8   x y) => (SLT  (SignExt8to64  x) (SignExt8to64  y))
   226(Less64U ...) => (SLTU ...)
   227(Less32U x y) => (SLTU (ZeroExt32to64 x) (ZeroExt32to64 y))
   228(Less16U x y) => (SLTU (ZeroExt16to64 x) (ZeroExt16to64 y))
   229(Less8U  x y) => (SLTU (ZeroExt8to64  x) (ZeroExt8to64  y))
   230(Less(64|32)F ...) => (FLT(D|S) ...)
   231
   232// Convert x <= y to !(y > x).
   233(Leq(64|32|16|8)  x y) => (Not (Less(64|32|16|8)  y x))
   234(Leq(64|32|16|8)U x y) => (Not (Less(64|32|16|8)U y x))
   235(Leq(64|32)F ...) => (FLE(D|S) ...)
   236
   237(EqPtr x y) => (SEQZ (SUB <typ.Uintptr> x y))
   238(Eq64  x y) => (SEQZ (SUB <x.Type> x y))
   239(Eq32  x y) &&  x.Type.IsSigned() => (SEQZ (SUB <x.Type> (SignExt32to64 x) (SignExt32to64 y)))
   240(Eq32  x y) && !x.Type.IsSigned() => (SEQZ (SUB <x.Type> (ZeroExt32to64 x) (ZeroExt32to64 y)))
   241(Eq16  x y) => (SEQZ (SUB <x.Type> (ZeroExt16to64 x) (ZeroExt16to64 y)))
   242(Eq8   x y) => (SEQZ (SUB <x.Type> (ZeroExt8to64  x) (ZeroExt8to64  y)))
   243(Eq(64|32)F ...) => (FEQ(D|S) ...)
   244
   245(NeqPtr x y) => (Not (EqPtr x y))
   246(Neq64  x y) => (Not (Eq64  x y))
   247(Neq32  x y) => (Not (Eq32  x y))
   248(Neq16  x y) => (Not (Eq16  x y))
   249(Neq8   x y) => (Not (Eq8   x y))
   250(Neq(64|32)F ...) => (FNE(D|S) ...)
   251
   252// Loads
   253(Load <t> ptr mem) &&  t.IsBoolean()                   => (MOVBUload ptr mem)
   254(Load <t> ptr mem) && ( is8BitInt(t) &&  t.IsSigned()) => (MOVBload  ptr mem)
   255(Load <t> ptr mem) && ( is8BitInt(t) && !t.IsSigned()) => (MOVBUload ptr mem)
   256(Load <t> ptr mem) && (is16BitInt(t) &&  t.IsSigned()) => (MOVHload  ptr mem)
   257(Load <t> ptr mem) && (is16BitInt(t) && !t.IsSigned()) => (MOVHUload ptr mem)
   258(Load <t> ptr mem) && (is32BitInt(t) &&  t.IsSigned()) => (MOVWload  ptr mem)
   259(Load <t> ptr mem) && (is32BitInt(t) && !t.IsSigned()) => (MOVWUload ptr mem)
   260(Load <t> ptr mem) && (is64BitInt(t) || isPtr(t))      => (MOVDload  ptr mem)
   261(Load <t> ptr mem) &&  is32BitFloat(t)                 => (FMOVWload ptr mem)
   262(Load <t> ptr mem) &&  is64BitFloat(t)                 => (FMOVDload ptr mem)
   263
   264// Stores
   265(Store {t} ptr val mem) && t.Size() == 1 => (MOVBstore ptr val mem)
   266(Store {t} ptr val mem) && t.Size() == 2 => (MOVHstore ptr val mem)
   267(Store {t} ptr val mem) && t.Size() == 4 && !t.IsFloat() => (MOVWstore ptr val mem)
   268(Store {t} ptr val mem) && t.Size() == 8 && !t.IsFloat() => (MOVDstore ptr val mem)
   269(Store {t} ptr val mem) && t.Size() == 4 &&  t.IsFloat() => (FMOVWstore ptr val mem)
   270(Store {t} ptr val mem) && t.Size() == 8 &&  t.IsFloat() => (FMOVDstore ptr val mem)
   271
   272// We need to fold MOVaddr into the LD/MOVDstore ops so that the live variable analysis
   273// knows what variables are being read/written by the ops.
   274(MOVBUload [off1] {sym1} (MOVaddr [off2] {sym2} base) mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   275	(MOVBUload [off1+off2] {mergeSym(sym1,sym2)} base mem)
   276(MOVBload  [off1] {sym1} (MOVaddr [off2] {sym2} base) mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   277	(MOVBload  [off1+off2] {mergeSym(sym1,sym2)} base mem)
   278(MOVHUload [off1] {sym1} (MOVaddr [off2] {sym2} base) mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   279	(MOVHUload [off1+off2] {mergeSym(sym1,sym2)} base mem)
   280(MOVHload  [off1] {sym1} (MOVaddr [off2] {sym2} base) mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   281	(MOVHload  [off1+off2] {mergeSym(sym1,sym2)} base mem)
   282(MOVWUload [off1] {sym1} (MOVaddr [off2] {sym2} base) mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   283	(MOVWUload [off1+off2] {mergeSym(sym1,sym2)} base mem)
   284(MOVWload  [off1] {sym1} (MOVaddr [off2] {sym2} base) mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   285	(MOVWload  [off1+off2] {mergeSym(sym1,sym2)} base mem)
   286(MOVDload  [off1] {sym1} (MOVaddr [off2] {sym2} base) mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   287	(MOVDload  [off1+off2] {mergeSym(sym1,sym2)} base mem)
   288
   289(MOVBstore [off1] {sym1} (MOVaddr [off2] {sym2} base) val mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   290	(MOVBstore [off1+off2] {mergeSym(sym1,sym2)} base val mem)
   291(MOVHstore [off1] {sym1} (MOVaddr [off2] {sym2} base) val mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   292	(MOVHstore [off1+off2] {mergeSym(sym1,sym2)} base val mem)
   293(MOVWstore [off1] {sym1} (MOVaddr [off2] {sym2} base) val mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   294	(MOVWstore [off1+off2] {mergeSym(sym1,sym2)} base val mem)
   295(MOVDstore [off1] {sym1} (MOVaddr [off2] {sym2} base) val mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   296	(MOVDstore [off1+off2] {mergeSym(sym1,sym2)} base val mem)
   297(MOVBstorezero [off1] {sym1} (MOVaddr [off2] {sym2} ptr) mem) && canMergeSym(sym1,sym2) && is32Bit(int64(off1)+int64(off2)) =>
   298	(MOVBstorezero [off1+off2] {mergeSym(sym1,sym2)} ptr mem)
   299(MOVHstorezero [off1] {sym1} (MOVaddr [off2] {sym2} ptr) mem) && canMergeSym(sym1,sym2) && is32Bit(int64(off1)+int64(off2)) =>
   300	(MOVHstorezero [off1+off2] {mergeSym(sym1,sym2)} ptr mem)
   301(MOVWstorezero [off1] {sym1} (MOVaddr [off2] {sym2} ptr) mem) && canMergeSym(sym1,sym2) && is32Bit(int64(off1)+int64(off2)) =>
   302	(MOVWstorezero [off1+off2] {mergeSym(sym1,sym2)} ptr mem)
   303(MOVDstorezero [off1] {sym1} (MOVaddr [off2] {sym2} ptr) mem) && canMergeSym(sym1,sym2) && is32Bit(int64(off1)+int64(off2)) =>
   304	(MOVDstorezero [off1+off2] {mergeSym(sym1,sym2)} ptr mem)
   305
   306(MOVBUload [off1] {sym} (ADDI [off2] base) mem) && is32Bit(int64(off1)+off2) =>
   307	(MOVBUload [off1+int32(off2)] {sym} base mem)
   308(MOVBload  [off1] {sym} (ADDI [off2] base) mem) && is32Bit(int64(off1)+off2) =>
   309	(MOVBload  [off1+int32(off2)] {sym} base mem)
   310(MOVHUload [off1] {sym} (ADDI [off2] base) mem) && is32Bit(int64(off1)+off2) =>
   311	(MOVHUload [off1+int32(off2)] {sym} base mem)
   312(MOVHload  [off1] {sym} (ADDI [off2] base) mem) && is32Bit(int64(off1)+off2) =>
   313	(MOVHload  [off1+int32(off2)] {sym} base mem)
   314(MOVWUload [off1] {sym} (ADDI [off2] base) mem) && is32Bit(int64(off1)+off2) =>
   315	(MOVWUload [off1+int32(off2)] {sym} base mem)
   316(MOVWload  [off1] {sym} (ADDI [off2] base) mem) && is32Bit(int64(off1)+off2) =>
   317	(MOVWload  [off1+int32(off2)] {sym} base mem)
   318(MOVDload  [off1] {sym} (ADDI [off2] base) mem) && is32Bit(int64(off1)+off2) =>
   319	(MOVDload  [off1+int32(off2)] {sym} base mem)
   320
   321(MOVBstore [off1] {sym} (ADDI [off2] base) val mem) && is32Bit(int64(off1)+off2) =>
   322	(MOVBstore [off1+int32(off2)] {sym} base val mem)
   323(MOVHstore [off1] {sym} (ADDI [off2] base) val mem) && is32Bit(int64(off1)+off2) =>
   324	(MOVHstore [off1+int32(off2)] {sym} base val mem)
   325(MOVWstore [off1] {sym} (ADDI [off2] base) val mem) && is32Bit(int64(off1)+off2) =>
   326	(MOVWstore [off1+int32(off2)] {sym} base val mem)
   327(MOVDstore [off1] {sym} (ADDI [off2] base) val mem) && is32Bit(int64(off1)+off2) =>
   328	(MOVDstore [off1+int32(off2)] {sym} base val mem)
   329(MOVBstorezero [off1] {sym} (ADDI [off2] ptr) mem) && is32Bit(int64(off1)+off2) => (MOVBstorezero [off1+int32(off2)] {sym} ptr mem)
   330(MOVHstorezero [off1] {sym} (ADDI [off2] ptr) mem) && is32Bit(int64(off1)+off2) => (MOVHstorezero [off1+int32(off2)] {sym} ptr mem)
   331(MOVWstorezero [off1] {sym} (ADDI [off2] ptr) mem) && is32Bit(int64(off1)+off2) => (MOVWstorezero [off1+int32(off2)] {sym} ptr mem)
   332(MOVDstorezero [off1] {sym} (ADDI [off2] ptr) mem) && is32Bit(int64(off1)+off2) => (MOVDstorezero [off1+int32(off2)] {sym} ptr mem)
   333
   334// Similarly, fold ADDI into MOVaddr to avoid confusing live variable analysis
   335// with OffPtr -> ADDI.
   336(ADDI [c] (MOVaddr [d] {s} x)) && is32Bit(c+int64(d)) => (MOVaddr [int32(c)+d] {s} x)
   337
   338// Small zeroing
   339(Zero [0] _ mem) => mem
   340(Zero [1] ptr mem) => (MOVBstore ptr (MOVDconst [0]) mem)
   341(Zero [2] {t} ptr mem) && t.Alignment()%2 == 0 =>
   342	(MOVHstore ptr (MOVDconst [0]) mem)
   343(Zero [2] ptr mem) =>
   344	(MOVBstore [1] ptr (MOVDconst [0])
   345		(MOVBstore ptr (MOVDconst [0]) mem))
   346(Zero [4] {t} ptr mem) && t.Alignment()%4 == 0 =>
   347	(MOVWstore ptr (MOVDconst [0]) mem)
   348(Zero [4] {t} ptr mem) && t.Alignment()%2 == 0 =>
   349	(MOVHstore [2] ptr (MOVDconst [0])
   350		(MOVHstore ptr (MOVDconst [0]) mem))
   351(Zero [4] ptr mem) =>
   352	(MOVBstore [3] ptr (MOVDconst [0])
   353		(MOVBstore [2] ptr (MOVDconst [0])
   354			(MOVBstore [1] ptr (MOVDconst [0])
   355				(MOVBstore ptr (MOVDconst [0]) mem))))
   356(Zero [8] {t} ptr mem) && t.Alignment()%8 == 0 =>
   357	(MOVDstore ptr (MOVDconst [0]) mem)
   358(Zero [8] {t} ptr mem) && t.Alignment()%4 == 0 =>
   359	(MOVWstore [4] ptr (MOVDconst [0])
   360		(MOVWstore ptr (MOVDconst [0]) mem))
   361(Zero [8] {t} ptr mem) && t.Alignment()%2 == 0 =>
   362	(MOVHstore [6] ptr (MOVDconst [0])
   363		(MOVHstore [4] ptr (MOVDconst [0])
   364			(MOVHstore [2] ptr (MOVDconst [0])
   365				(MOVHstore ptr (MOVDconst [0]) mem))))
   366
   367(Zero [3] ptr mem) =>
   368	(MOVBstore [2] ptr (MOVDconst [0])
   369		(MOVBstore [1] ptr (MOVDconst [0])
   370			(MOVBstore ptr (MOVDconst [0]) mem)))
   371(Zero [6] {t} ptr mem) && t.Alignment()%2 == 0 =>
   372	(MOVHstore [4] ptr (MOVDconst [0])
   373		(MOVHstore [2] ptr (MOVDconst [0])
   374			(MOVHstore ptr (MOVDconst [0]) mem)))
   375(Zero [12] {t} ptr mem) && t.Alignment()%4 == 0 =>
   376	(MOVWstore [8] ptr (MOVDconst [0])
   377		(MOVWstore [4] ptr (MOVDconst [0])
   378			(MOVWstore ptr (MOVDconst [0]) mem)))
   379(Zero [16] {t} ptr mem) && t.Alignment()%8 == 0 =>
   380	(MOVDstore [8] ptr (MOVDconst [0])
   381		(MOVDstore ptr (MOVDconst [0]) mem))
   382(Zero [24] {t} ptr mem) && t.Alignment()%8 == 0 =>
   383	(MOVDstore [16] ptr (MOVDconst [0])
   384		(MOVDstore [8] ptr (MOVDconst [0])
   385			(MOVDstore ptr (MOVDconst [0]) mem)))
   386(Zero [32] {t} ptr mem) && t.Alignment()%8 == 0 =>
   387	(MOVDstore [24] ptr (MOVDconst [0])
   388		(MOVDstore [16] ptr (MOVDconst [0])
   389			(MOVDstore [8] ptr (MOVDconst [0])
   390				(MOVDstore ptr (MOVDconst [0]) mem))))
   391
   392// Medium 8-aligned zeroing uses a Duff's device
   393// 8 and 128 are magic constants, see runtime/mkduff.go
   394(Zero [s] {t} ptr mem)
   395	&& s%8 == 0 && s <= 8*128
   396	&& t.Alignment()%8 == 0 && !config.noDuffDevice =>
   397	(DUFFZERO [8 * (128 - s/8)] ptr mem)
   398
   399// Generic zeroing uses a loop
   400(Zero [s] {t} ptr mem) =>
   401	(LoweredZero [t.Alignment()]
   402		ptr
   403		(ADD <ptr.Type> ptr (MOVDconst [s-moveSize(t.Alignment(), config)]))
   404		mem)
   405
   406// Checks
   407(IsNonNil ...) => (SNEZ ...)
   408(IsInBounds ...) => (Less64U ...)
   409(IsSliceInBounds ...) => (Leq64U ...)
   410
   411// Trivial lowering
   412(NilCheck ...) => (LoweredNilCheck ...)
   413(GetClosurePtr ...) => (LoweredGetClosurePtr ...)
   414(GetCallerSP ...) => (LoweredGetCallerSP ...)
   415(GetCallerPC ...) => (LoweredGetCallerPC ...)
   416
   417// Write barrier.
   418(WB ...) => (LoweredWB ...)
   419
   420// Publication barrier as intrinsic
   421(PubBarrier ...) => (LoweredPubBarrier ...)
   422
   423(PanicBounds [kind] x y mem) && boundsABI(kind) == 0 => (LoweredPanicBoundsA [kind] x y mem)
   424(PanicBounds [kind] x y mem) && boundsABI(kind) == 1 => (LoweredPanicBoundsB [kind] x y mem)
   425(PanicBounds [kind] x y mem) && boundsABI(kind) == 2 => (LoweredPanicBoundsC [kind] x y mem)
   426
   427// Small moves
   428(Move [0] _ _ mem) => mem
   429(Move [1] dst src mem) => (MOVBstore dst (MOVBload src mem) mem)
   430(Move [2] {t} dst src mem) && t.Alignment()%2 == 0 =>
   431	(MOVHstore dst (MOVHload src mem) mem)
   432(Move [2] dst src mem) =>
   433	(MOVBstore [1] dst (MOVBload [1] src mem)
   434		(MOVBstore dst (MOVBload src mem) mem))
   435(Move [4] {t} dst src mem) && t.Alignment()%4 == 0 =>
   436	(MOVWstore dst (MOVWload src mem) mem)
   437(Move [4] {t} dst src mem) && t.Alignment()%2 == 0 =>
   438	(MOVHstore [2] dst (MOVHload [2] src mem)
   439		(MOVHstore dst (MOVHload src mem) mem))
   440(Move [4] dst src mem) =>
   441	(MOVBstore [3] dst (MOVBload [3] src mem)
   442		(MOVBstore [2] dst (MOVBload [2] src mem)
   443			(MOVBstore [1] dst (MOVBload [1] src mem)
   444				(MOVBstore dst (MOVBload src mem) mem))))
   445(Move [8] {t} dst src mem) && t.Alignment()%8 == 0 =>
   446	(MOVDstore dst (MOVDload src mem) mem)
   447(Move [8] {t} dst src mem) && t.Alignment()%4 == 0 =>
   448	(MOVWstore [4] dst (MOVWload [4] src mem)
   449		(MOVWstore dst (MOVWload src mem) mem))
   450(Move [8] {t} dst src mem) && t.Alignment()%2 == 0 =>
   451	(MOVHstore [6] dst (MOVHload [6] src mem)
   452		(MOVHstore [4] dst (MOVHload [4] src mem)
   453			(MOVHstore [2] dst (MOVHload [2] src mem)
   454				(MOVHstore dst (MOVHload src mem) mem))))
   455
   456(Move [3] dst src mem) =>
   457	(MOVBstore [2] dst (MOVBload [2] src mem)
   458		(MOVBstore [1] dst (MOVBload [1] src mem)
   459			(MOVBstore dst (MOVBload src mem) mem)))
   460(Move [6] {t} dst src mem) && t.Alignment()%2 == 0 =>
   461	(MOVHstore [4] dst (MOVHload [4] src mem)
   462		(MOVHstore [2] dst (MOVHload [2] src mem)
   463			(MOVHstore dst (MOVHload src mem) mem)))
   464(Move [12] {t} dst src mem) && t.Alignment()%4 == 0 =>
   465	(MOVWstore [8] dst (MOVWload [8] src mem)
   466		(MOVWstore [4] dst (MOVWload [4] src mem)
   467			(MOVWstore dst (MOVWload src mem) mem)))
   468(Move [16] {t} dst src mem) && t.Alignment()%8 == 0 =>
   469	(MOVDstore [8] dst (MOVDload [8] src mem)
   470		(MOVDstore dst (MOVDload src mem) mem))
   471(Move [24] {t} dst src mem) && t.Alignment()%8 == 0 =>
   472	(MOVDstore [16] dst (MOVDload [16] src mem)
   473		(MOVDstore [8] dst (MOVDload [8] src mem)
   474			(MOVDstore dst (MOVDload src mem) mem)))
   475(Move [32] {t} dst src mem) && t.Alignment()%8 == 0 =>
   476	(MOVDstore [24] dst (MOVDload [24] src mem)
   477		(MOVDstore [16] dst (MOVDload [16] src mem)
   478			(MOVDstore [8] dst (MOVDload [8] src mem)
   479				(MOVDstore dst (MOVDload src mem) mem))))
   480
   481// Medium 8-aligned move uses a Duff's device
   482// 16 and 128 are magic constants, see runtime/mkduff.go
   483(Move [s] {t} dst src mem)
   484	&& s%8 == 0 && s <= 8*128 && t.Alignment()%8 == 0
   485	&& !config.noDuffDevice && logLargeCopy(v, s) =>
   486	(DUFFCOPY [16 * (128 - s/8)] dst src mem)
   487
   488// Generic move uses a loop
   489(Move [s] {t} dst src mem) && (s <= 16 || logLargeCopy(v, s)) =>
   490	(LoweredMove [t.Alignment()]
   491		dst
   492		src
   493		(ADDI <src.Type> [s-moveSize(t.Alignment(), config)] src)
   494		mem)
   495
   496// Boolean ops; 0=false, 1=true
   497(AndB ...) => (AND ...)
   498(OrB  ...) => (OR  ...)
   499(EqB  x y) => (SEQZ (SUB <typ.Bool> x y))
   500(NeqB x y) => (SNEZ (SUB <typ.Bool> x y))
   501(Not  ...) => (SEQZ ...)
   502
   503// Lowering pointer arithmetic
   504// TODO: Special handling for SP offsets, like ARM
   505(OffPtr [off] ptr:(SP)) && is32Bit(off) => (MOVaddr [int32(off)] ptr)
   506(OffPtr [off] ptr) && is32Bit(off) => (ADDI [off] ptr)
   507(OffPtr [off] ptr) => (ADD (MOVDconst [off]) ptr)
   508
   509(Const(64|32|16|8) [val]) => (MOVDconst [int64(val)])
   510(Const32F [val]) => (FMVSX (MOVDconst [int64(math.Float32bits(val))]))
   511(Const64F [val]) => (FMVDX (MOVDconst [int64(math.Float64bits(val))]))
   512(ConstNil) => (MOVDconst [0])
   513(ConstBool [val]) => (MOVDconst [int64(b2i(val))])
   514
   515(Addr {sym} base) => (MOVaddr {sym} [0] base)
   516(LocalAddr <t> {sym} base mem) && t.Elem().HasPointers() => (MOVaddr {sym} (SPanchored base mem))
   517(LocalAddr <t> {sym} base _)  && !t.Elem().HasPointers() => (MOVaddr {sym} base)
   518
   519// Calls
   520(StaticCall  ...) => (CALLstatic  ...)
   521(ClosureCall ...) => (CALLclosure ...)
   522(InterCall   ...) => (CALLinter   ...)
   523(TailCall ...) => (CALLtail ...)
   524
   525// Atomic Intrinsics
   526(AtomicLoad(Ptr|64|32|8)  ...) => (LoweredAtomicLoad(64|64|32|8) ...)
   527(AtomicStore(PtrNoWB|64|32|8) ...) => (LoweredAtomicStore(64|64|32|8) ...)
   528(AtomicAdd(64|32) ...) => (LoweredAtomicAdd(64|32) ...)
   529
   530// AtomicAnd8(ptr,val) => LoweredAtomicAnd32(ptr&^3, ^((uint8(val) ^ 0xff) << ((ptr & 3) * 8)))
   531(AtomicAnd8 ptr val mem) =>
   532	(LoweredAtomicAnd32 (ANDI <typ.Uintptr> [^3] ptr)
   533		(NOT <typ.UInt32> (SLL <typ.UInt32> (XORI <typ.UInt32> [0xff] (ZeroExt8to32 val))
   534			(SLLI <typ.UInt64> [3] (ANDI <typ.UInt64> [3] ptr)))) mem)
   535
   536(AtomicAnd32 ...) => (LoweredAtomicAnd32 ...)
   537
   538(AtomicCompareAndSwap32 ptr old new mem) => (LoweredAtomicCas32 ptr (SignExt32to64 old) new mem)
   539(AtomicCompareAndSwap64 ...) => (LoweredAtomicCas64 ...)
   540
   541(AtomicExchange(64|32) ...) => (LoweredAtomicExchange(64|32) ...)
   542
   543// AtomicOr8(ptr,val)  => LoweredAtomicOr32(ptr&^3, uint32(val)<<((ptr&3)*8))
   544(AtomicOr8 ptr val mem) =>
   545	(LoweredAtomicOr32 (ANDI <typ.Uintptr> [^3] ptr)
   546		(SLL <typ.UInt32> (ZeroExt8to32 val)
   547			(SLLI <typ.UInt64> [3] (ANDI <typ.UInt64> [3] ptr))) mem)
   548
   549(AtomicOr32  ...) => (LoweredAtomicOr32  ...)
   550
   551// Conditional branches
   552(If cond yes no) => (BNEZ (MOVBUreg <typ.UInt64> cond) yes no)
   553
   554// Optimizations
   555
   556// Absorb SEQZ/SNEZ into branch.
   557(BEQZ (SEQZ x) yes no) => (BNEZ x yes no)
   558(BEQZ (SNEZ x) yes no) => (BEQZ x yes no)
   559(BNEZ (SEQZ x) yes no) => (BEQZ x yes no)
   560(BNEZ (SNEZ x) yes no) => (BNEZ x yes no)
   561
   562// Remove redundant NEG from BEQZ/BNEZ.
   563(BEQZ (NEG x) yes no) => (BEQZ x yes no)
   564(BNEZ (NEG x) yes no) => (BNEZ x yes no)
   565
   566// Negate comparison with FNES/FNED.
   567(BEQZ (FNES <t> x y) yes no) => (BNEZ (FEQS <t> x y) yes no)
   568(BNEZ (FNES <t> x y) yes no) => (BEQZ (FEQS <t> x y) yes no)
   569(BEQZ (FNED <t> x y) yes no) => (BNEZ (FEQD <t> x y) yes no)
   570(BNEZ (FNED <t> x y) yes no) => (BEQZ (FEQD <t> x y) yes no)
   571
   572// Convert BEQZ/BNEZ into more optimal branch conditions.
   573(BEQZ (SUB x y) yes no) => (BEQ x y yes no)
   574(BNEZ (SUB x y) yes no) => (BNE x y yes no)
   575(BEQZ (SLT x y) yes no) => (BGE x y yes no)
   576(BNEZ (SLT x y) yes no) => (BLT x y yes no)
   577(BEQZ (SLTU x y) yes no) => (BGEU x y yes no)
   578(BNEZ (SLTU x y) yes no) => (BLTU x y yes no)
   579(BEQZ (SLTI [x] y) yes no) => (BGE y (MOVDconst [x]) yes no)
   580(BNEZ (SLTI [x] y) yes no) => (BLT y (MOVDconst [x]) yes no)
   581(BEQZ (SLTIU [x] y) yes no) => (BGEU y (MOVDconst [x]) yes no)
   582(BNEZ (SLTIU [x] y) yes no) => (BLTU y (MOVDconst [x]) yes no)
   583
   584// Convert branch with zero to more optimal branch zero.
   585(BEQ (MOVDconst [0]) cond yes no) => (BEQZ cond yes no)
   586(BEQ cond (MOVDconst [0]) yes no) => (BEQZ cond yes no)
   587(BNE (MOVDconst [0]) cond yes no) => (BNEZ cond yes no)
   588(BNE cond (MOVDconst [0]) yes no) => (BNEZ cond yes no)
   589(BLT (MOVDconst [0]) cond yes no) => (BGTZ cond yes no)
   590(BLT cond (MOVDconst [0]) yes no) => (BLTZ cond yes no)
   591(BGE (MOVDconst [0]) cond yes no) => (BLEZ cond yes no)
   592(BGE cond (MOVDconst [0]) yes no) => (BGEZ cond yes no)
   593
   594// Remove redundant NEG from SEQZ/SNEZ.
   595(SEQZ (NEG x)) => (SEQZ x)
   596(SNEZ (NEG x)) => (SNEZ x)
   597
   598// Remove redundant SEQZ/SNEZ.
   599(SEQZ (SEQZ x)) => (SNEZ x)
   600(SEQZ (SNEZ x)) => (SEQZ x)
   601(SNEZ (SEQZ x)) => (SEQZ x)
   602(SNEZ (SNEZ x)) => (SNEZ x)
   603
   604// Store zero.
   605(MOVBstore [off] {sym} ptr (MOVDconst [0]) mem) => (MOVBstorezero [off] {sym} ptr mem)
   606(MOVHstore [off] {sym} ptr (MOVDconst [0]) mem) => (MOVHstorezero [off] {sym} ptr mem)
   607(MOVWstore [off] {sym} ptr (MOVDconst [0]) mem) => (MOVWstorezero [off] {sym} ptr mem)
   608(MOVDstore [off] {sym} ptr (MOVDconst [0]) mem) => (MOVDstorezero [off] {sym} ptr mem)
   609
   610// Boolean ops are already extended.
   611(MOVBUreg x:((FLES|FLTS|FEQS|FNES) _ _)) => x
   612(MOVBUreg x:((FLED|FLTD|FEQD|FNED) _ _)) => x
   613(MOVBUreg x:((SEQZ|SNEZ) _)) => x
   614(MOVBUreg x:((SLT|SLTU) _ _)) => x
   615
   616// Avoid extending when already sufficiently masked.
   617(MOVBreg  x:(ANDI [c] y)) && c >= 0 && int64(int8(c)) == c => x
   618(MOVHreg  x:(ANDI [c] y)) && c >= 0 && int64(int16(c)) == c => x
   619(MOVWreg  x:(ANDI [c] y)) && c >= 0 && int64(int32(c)) == c => x
   620(MOVBUreg x:(ANDI [c] y)) && c >= 0 && int64(uint8(c)) == c => x
   621(MOVHUreg x:(ANDI [c] y)) && c >= 0 && int64(uint16(c)) == c => x
   622(MOVWUreg x:(ANDI [c] y)) && c >= 0 && int64(uint32(c)) == c => x
   623
   624// Combine masking and zero extension.
   625(MOVBUreg (ANDI [c] x)) && c < 0 => (ANDI [int64(uint8(c))] x)
   626(MOVHUreg (ANDI [c] x)) && c < 0 => (ANDI [int64(uint16(c))] x)
   627(MOVWUreg (ANDI [c] x)) && c < 0 => (AND (MOVDconst [int64(uint32(c))]) x)
   628
   629// Avoid sign/zero extension for consts.
   630(MOVBreg  (MOVDconst [c])) => (MOVDconst [int64(int8(c))])
   631(MOVHreg  (MOVDconst [c])) => (MOVDconst [int64(int16(c))])
   632(MOVWreg  (MOVDconst [c])) => (MOVDconst [int64(int32(c))])
   633(MOVBUreg (MOVDconst [c])) => (MOVDconst [int64(uint8(c))])
   634(MOVHUreg (MOVDconst [c])) => (MOVDconst [int64(uint16(c))])
   635(MOVWUreg (MOVDconst [c])) => (MOVDconst [int64(uint32(c))])
   636
   637// Avoid sign/zero extension after properly typed load.
   638(MOVBreg  x:(MOVBload  _ _)) => (MOVDreg x)
   639(MOVHreg  x:(MOVBload  _ _)) => (MOVDreg x)
   640(MOVHreg  x:(MOVBUload _ _)) => (MOVDreg x)
   641(MOVHreg  x:(MOVHload  _ _)) => (MOVDreg x)
   642(MOVWreg  x:(MOVBload  _ _)) => (MOVDreg x)
   643(MOVWreg  x:(MOVBUload _ _)) => (MOVDreg x)
   644(MOVWreg  x:(MOVHload  _ _)) => (MOVDreg x)
   645(MOVWreg  x:(MOVHUload _ _)) => (MOVDreg x)
   646(MOVWreg  x:(MOVWload  _ _)) => (MOVDreg x)
   647(MOVBUreg x:(MOVBUload _ _)) => (MOVDreg x)
   648(MOVHUreg x:(MOVBUload _ _)) => (MOVDreg x)
   649(MOVHUreg x:(MOVHUload _ _)) => (MOVDreg x)
   650(MOVWUreg x:(MOVBUload _ _)) => (MOVDreg x)
   651(MOVWUreg x:(MOVHUload _ _)) => (MOVDreg x)
   652(MOVWUreg x:(MOVWUload _ _)) => (MOVDreg x)
   653
   654// Avoid zero extension after properly typed atomic operation.
   655(MOVBUreg x:(Select0 (LoweredAtomicLoad8 _ _))) => (MOVDreg x)
   656(MOVBUreg x:(Select0 (LoweredAtomicCas32 _ _ _ _))) => (MOVDreg x)
   657(MOVBUreg x:(Select0 (LoweredAtomicCas64 _ _ _ _))) => (MOVDreg x)
   658
   659// Avoid sign extension after word arithmetic.
   660(MOVWreg x:(ADDIW   _)) => (MOVDreg x)
   661(MOVWreg x:(SUBW  _ _)) => (MOVDreg x)
   662(MOVWreg x:(NEGW    _)) => (MOVDreg x)
   663(MOVWreg x:(MULW  _ _)) => (MOVDreg x)
   664(MOVWreg x:(DIVW  _ _)) => (MOVDreg x)
   665(MOVWreg x:(DIVUW _ _)) => (MOVDreg x)
   666(MOVWreg x:(REMW  _ _)) => (MOVDreg x)
   667(MOVWreg x:(REMUW _ _)) => (MOVDreg x)
   668(MOVWreg x:(ROLW  _ _)) => (MOVDreg x)
   669(MOVWreg x:(RORW  _ _)) => (MOVDreg x)
   670(MOVWreg x:(RORIW   _)) => (MOVDreg x)
   671
   672// Fold double extensions.
   673(MOVBreg  x:(MOVBreg  _)) => (MOVDreg x)
   674(MOVHreg  x:(MOVBreg  _)) => (MOVDreg x)
   675(MOVHreg  x:(MOVBUreg _)) => (MOVDreg x)
   676(MOVHreg  x:(MOVHreg  _)) => (MOVDreg x)
   677(MOVWreg  x:(MOVBreg  _)) => (MOVDreg x)
   678(MOVWreg  x:(MOVBUreg _)) => (MOVDreg x)
   679(MOVWreg  x:(MOVHreg  _)) => (MOVDreg x)
   680(MOVWreg  x:(MOVWreg  _)) => (MOVDreg x)
   681(MOVBUreg x:(MOVBUreg _)) => (MOVDreg x)
   682(MOVHUreg x:(MOVBUreg _)) => (MOVDreg x)
   683(MOVHUreg x:(MOVHUreg _)) => (MOVDreg x)
   684(MOVWUreg x:(MOVBUreg _)) => (MOVDreg x)
   685(MOVWUreg x:(MOVHUreg _)) => (MOVDreg x)
   686(MOVWUreg x:(MOVWUreg _)) => (MOVDreg x)
   687
   688// Do not extend before store.
   689(MOVBstore [off] {sym} ptr (MOVBreg  x) mem) => (MOVBstore [off] {sym} ptr x mem)
   690(MOVBstore [off] {sym} ptr (MOVHreg  x) mem) => (MOVBstore [off] {sym} ptr x mem)
   691(MOVBstore [off] {sym} ptr (MOVWreg  x) mem) => (MOVBstore [off] {sym} ptr x mem)
   692(MOVBstore [off] {sym} ptr (MOVBUreg x) mem) => (MOVBstore [off] {sym} ptr x mem)
   693(MOVBstore [off] {sym} ptr (MOVHUreg x) mem) => (MOVBstore [off] {sym} ptr x mem)
   694(MOVBstore [off] {sym} ptr (MOVWUreg x) mem) => (MOVBstore [off] {sym} ptr x mem)
   695(MOVHstore [off] {sym} ptr (MOVHreg  x) mem) => (MOVHstore [off] {sym} ptr x mem)
   696(MOVHstore [off] {sym} ptr (MOVWreg  x) mem) => (MOVHstore [off] {sym} ptr x mem)
   697(MOVHstore [off] {sym} ptr (MOVHUreg x) mem) => (MOVHstore [off] {sym} ptr x mem)
   698(MOVHstore [off] {sym} ptr (MOVWUreg x) mem) => (MOVHstore [off] {sym} ptr x mem)
   699(MOVWstore [off] {sym} ptr (MOVWreg  x) mem) => (MOVWstore [off] {sym} ptr x mem)
   700(MOVWstore [off] {sym} ptr (MOVWUreg x) mem) => (MOVWstore [off] {sym} ptr x mem)
   701
   702// Replace extend after load with alternate load where possible.
   703(MOVBreg  <t> x:(MOVBUload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVBload  <t> [off] {sym} ptr mem)
   704(MOVHreg  <t> x:(MOVHUload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVHload  <t> [off] {sym} ptr mem)
   705(MOVWreg  <t> x:(MOVWUload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVWload  <t> [off] {sym} ptr mem)
   706(MOVBUreg <t> x:(MOVBload  [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVBUload <t> [off] {sym} ptr mem)
   707(MOVHUreg <t> x:(MOVHload  [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVHUload <t> [off] {sym} ptr mem)
   708(MOVWUreg <t> x:(MOVWload  [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVWUload <t> [off] {sym} ptr mem)
   709
   710// If a register move has only 1 use, just use the same register without emitting instruction
   711// MOVnop does not emit an instruction, only for ensuring the type.
   712(MOVDreg x) && x.Uses == 1 => (MOVDnop x)
   713
   714// TODO: we should be able to get rid of MOVDnop all together.
   715// But for now, this is enough to get rid of lots of them.
   716(MOVDnop (MOVDconst [c])) => (MOVDconst [c])
   717
   718// Avoid unnecessary zero and sign extension when right shifting.
   719(SRAI <t> [x] (MOVWreg  y)) && x >= 0 && x <= 31 => (SRAIW <t> [int64(x)] y)
   720(SRLI <t> [x] (MOVWUreg y)) && x >= 0 && x <= 31 => (SRLIW <t> [int64(x)] y)
   721
   722// Replace right shifts that exceed size of signed type.
   723(SRAI <t> [x] (MOVBreg y)) && x >=  8 => (SRAI  [63] (SLLI <t> [56] y))
   724(SRAI <t> [x] (MOVHreg y)) && x >= 16 => (SRAI  [63] (SLLI <t> [48] y))
   725(SRAI <t> [x] (MOVWreg y)) && x >= 32 => (SRAIW [31] y)
   726
   727// Eliminate right shifts that exceed size of unsigned type.
   728(SRLI <t> [x] (MOVBUreg y)) && x >=  8 => (MOVDconst <t> [0])
   729(SRLI <t> [x] (MOVHUreg y)) && x >= 16 => (MOVDconst <t> [0])
   730(SRLI <t> [x] (MOVWUreg y)) && x >= 32 => (MOVDconst <t> [0])
   731
   732// Fold constant into immediate instructions where possible.
   733(ADD (MOVDconst <t> [val]) x) && is32Bit(val) && !t.IsPtr() => (ADDI [val] x)
   734(AND (MOVDconst [val]) x) && is32Bit(val) => (ANDI [val] x)
   735(OR  (MOVDconst [val]) x) && is32Bit(val) => (ORI  [val] x)
   736(XOR (MOVDconst [val]) x) && is32Bit(val) => (XORI [val] x)
   737(ROL  x (MOVDconst [val])) => (RORI  [int64(int8(-val)&63)] x)
   738(ROLW x (MOVDconst [val])) => (RORIW [int64(int8(-val)&31)] x)
   739(ROR  x (MOVDconst [val])) => (RORI  [int64(val&63)] x)
   740(RORW x (MOVDconst [val])) => (RORIW [int64(val&31)] x)
   741(SLL  x (MOVDconst [val])) => (SLLI [int64(val&63)] x)
   742(SRL  x (MOVDconst [val])) => (SRLI [int64(val&63)] x)
   743(SLLW x (MOVDconst [val])) => (SLLIW [int64(val&31)] x)
   744(SRLW x (MOVDconst [val])) => (SRLIW [int64(val&31)] x)
   745(SRA  x (MOVDconst [val])) => (SRAI [int64(val&63)] x)
   746(SRAW x (MOVDconst [val])) => (SRAIW [int64(val&31)] x)
   747(SLT  x (MOVDconst [val])) && val >= -2048 && val <= 2047 => (SLTI  [val] x)
   748(SLTU x (MOVDconst [val])) && val >= -2048 && val <= 2047 => (SLTIU [val] x)
   749
   750// Replace negated left rotation with right rotation.
   751(ROL  x (NEG y)) => (ROR  x y)
   752(ROLW x (NEG y)) => (RORW x y)
   753
   754// Convert const subtraction into ADDI with negative immediate, where possible.
   755(SUB x (MOVDconst [val])) && is32Bit(-val) => (ADDI [-val] x)
   756(SUB <t> (MOVDconst [val]) y) && is32Bit(-val) => (NEG (ADDI <t> [-val] y))
   757
   758// Subtraction of zero.
   759(SUB  x (MOVDconst [0])) => x
   760(SUBW x (MOVDconst [0])) => (ADDIW [0] x)
   761
   762// Subtraction from zero.
   763(SUB  (MOVDconst [0]) x) => (NEG x)
   764(SUBW (MOVDconst [0]) x) => (NEGW x)
   765
   766// Fold negation into subtraction.
   767(NEG (SUB x y)) => (SUB y x)
   768(NEG <t> s:(ADDI [val] (SUB x y))) && s.Uses == 1 && is32Bit(-val) => (ADDI [-val] (SUB <t> y x))
   769
   770// Double negation.
   771(NEG (NEG x)) => x
   772
   773// Addition of zero or two constants.
   774(ADDI [0] x) => x
   775(ADDI [x] (MOVDconst [y])) && is32Bit(x + y) => (MOVDconst [x + y])
   776
   777// ANDI with all zeros, all ones or two constants.
   778(ANDI [0]  x) => (MOVDconst [0])
   779(ANDI [-1] x) => x
   780(ANDI [x] (MOVDconst [y])) => (MOVDconst [x & y])
   781
   782// ORI with all zeroes, all ones or two constants.
   783(ORI [0]  x) => x
   784(ORI [-1] x) => (MOVDconst [-1])
   785(ORI [x] (MOVDconst [y])) => (MOVDconst [x | y])
   786
   787// Combine operations with immediate.
   788(ADDI [x] (ADDI [y] z)) && is32Bit(x + y) => (ADDI [x + y] z)
   789(ANDI [x] (ANDI [y] z)) => (ANDI [x & y] z)
   790(ORI  [x] (ORI  [y] z)) => (ORI  [x | y] z)
   791
   792// Negation of a constant.
   793(NEG  (MOVDconst [x])) => (MOVDconst [-x])
   794(NEGW (MOVDconst [x])) => (MOVDconst [int64(int32(-x))])
   795
   796// Shift of a constant.
   797(SLLI [x] (MOVDconst [y])) && is32Bit(y << uint32(x)) => (MOVDconst [y << uint32(x)])
   798(SRLI [x] (MOVDconst [y])) => (MOVDconst [int64(uint64(y) >> uint32(x))])
   799(SRAI [x] (MOVDconst [y])) => (MOVDconst [int64(y) >> uint32(x)])
   800
   801// SLTI/SLTIU with constants.
   802(SLTI  [x] (MOVDconst [y])) => (MOVDconst [b2i(int64(y) < int64(x))])
   803(SLTIU [x] (MOVDconst [y])) => (MOVDconst [b2i(uint64(y) < uint64(x))])
   804
   805// SLTI/SLTIU with known outcomes.
   806(SLTI  [x] (ANDI [y] _)) && y >= 0 && int64(y) < int64(x) => (MOVDconst [1])
   807(SLTIU [x] (ANDI [y] _)) && y >= 0 && uint64(y) < uint64(x) => (MOVDconst [1])
   808(SLTI  [x] (ORI  [y] _)) && y >= 0 && int64(y) >= int64(x) => (MOVDconst [0])
   809(SLTIU [x] (ORI  [y] _)) && y >= 0 && uint64(y) >= uint64(x) => (MOVDconst [0])
   810
   811// SLT/SLTU with known outcomes.
   812(SLT  x x) => (MOVDconst [0])
   813(SLTU x x) => (MOVDconst [0])
   814
   815// Deadcode for LoweredMuluhilo
   816(Select0 m:(LoweredMuluhilo x y)) && m.Uses == 1 => (MULHU x y)
   817(Select1 m:(LoweredMuluhilo x y)) && m.Uses == 1 => (MUL x y)
   818
   819(FADD(S|D) a (FMUL(S|D) x y)) && a.Block.Func.useFMA(v) => (FMADD(S|D) x y a)
   820(FSUB(S|D) a (FMUL(S|D) x y)) && a.Block.Func.useFMA(v) => (FNMSUB(S|D) x y a)
   821(FSUB(S|D) (FMUL(S|D) x y) a) && a.Block.Func.useFMA(v) => (FMSUB(S|D) x y a)
   822
   823// Merge negation into fused multiply-add and multiply-subtract.
   824//
   825// Key:
   826//
   827//   [+ -](x * y [+ -] z).
   828//    _ N         A S
   829//                D U
   830//                D B
   831//
   832// Note: multiplication commutativity handled by rule generator.
   833(F(MADD|NMADD|MSUB|NMSUB)S neg:(FNEGS x) y z) && neg.Uses == 1 => (F(NMSUB|MSUB|NMADD|MADD)S x y z)
   834(F(MADD|NMADD|MSUB|NMSUB)S x y neg:(FNEGS z)) && neg.Uses == 1 => (F(MSUB|NMSUB|MADD|NMADD)S x y z)
   835(F(MADD|NMADD|MSUB|NMSUB)D neg:(FNEGD x) y z) && neg.Uses == 1 => (F(NMSUB|MSUB|NMADD|MADD)D x y z)
   836(F(MADD|NMADD|MSUB|NMSUB)D x y neg:(FNEGD z)) && neg.Uses == 1 => (F(MSUB|NMSUB|MADD|NMADD)D x y z)

View as plain text