...

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(Sqrt ...) => (FSQRTD ...)
    66(Sqrt32 ...) => (FSQRTS ...)
    67
    68(Copysign ...) => (FSGNJD ...)
    69
    70(Abs ...) => (FABSD ...)
    71
    72(FMA ...) => (FMADDD ...)
    73
    74(Min(64|32)F ...) => (LoweredFMIN(D|S) ...)
    75(Max(64|32)F ...) => (LoweredFMAX(D|S) ...)
    76
    77// Sign and zero extension.
    78
    79(SignExt8to16  ...) => (MOVBreg ...)
    80(SignExt8to32  ...) => (MOVBreg ...)
    81(SignExt8to64  ...) => (MOVBreg ...)
    82(SignExt16to32 ...) => (MOVHreg ...)
    83(SignExt16to64 ...) => (MOVHreg ...)
    84(SignExt32to64 ...) => (MOVWreg ...)
    85
    86(ZeroExt8to16  ...) => (MOVBUreg ...)
    87(ZeroExt8to32  ...) => (MOVBUreg ...)
    88(ZeroExt8to64  ...) => (MOVBUreg ...)
    89(ZeroExt16to32 ...) => (MOVHUreg ...)
    90(ZeroExt16to64 ...) => (MOVHUreg ...)
    91(ZeroExt32to64 ...) => (MOVWUreg ...)
    92
    93(Cvt32to32F ...) => (FCVTSW ...)
    94(Cvt32to64F ...) => (FCVTDW ...)
    95(Cvt64to32F ...) => (FCVTSL ...)
    96(Cvt64to64F ...) => (FCVTDL ...)
    97
    98(Cvt32Fto32 ...) => (FCVTWS ...)
    99(Cvt32Fto64 ...) => (FCVTLS ...)
   100(Cvt64Fto32 ...) => (FCVTWD ...)
   101(Cvt64Fto64 ...) => (FCVTLD ...)
   102
   103(Cvt32Fto64F ...) => (FCVTDS ...)
   104(Cvt64Fto32F ...) => (FCVTSD ...)
   105
   106(CvtBoolToUint8 ...) => (Copy ...)
   107
   108(Round(32|64)F ...) => (LoweredRound(32|64)F ...)
   109
   110(Slicemask <t> x) => (SRAI [63] (NEG <t> x))
   111
   112// Truncations
   113// We ignore the unused high parts of registers, so truncates are just copies.
   114(Trunc16to8  ...) => (Copy ...)
   115(Trunc32to8  ...) => (Copy ...)
   116(Trunc32to16 ...) => (Copy ...)
   117(Trunc64to8  ...) => (Copy ...)
   118(Trunc64to16 ...) => (Copy ...)
   119(Trunc64to32 ...) => (Copy ...)
   120
   121// Shifts
   122
   123// SLL only considers the bottom 6 bits of y. If y > 64, the result should
   124// always be 0.
   125//
   126// Breaking down the operation:
   127//
   128// (SLL x y) generates x << (y & 63).
   129//
   130// If y < 64, this is the value we want. Otherwise, we want zero.
   131//
   132// So, we AND with -1 * uint64(y < 64), which is 0xfffff... if y < 64 and 0 otherwise.
   133(Lsh8x8   <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg8  <t> (SLTIU <t> [64] (ZeroExt8to64  y))))
   134(Lsh8x16  <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg8  <t> (SLTIU <t> [64] (ZeroExt16to64 y))))
   135(Lsh8x32  <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg8  <t> (SLTIU <t> [64] (ZeroExt32to64 y))))
   136(Lsh8x64  <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg8  <t> (SLTIU <t> [64] y)))
   137(Lsh16x8  <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg16 <t> (SLTIU <t> [64] (ZeroExt8to64  y))))
   138(Lsh16x16 <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg16 <t> (SLTIU <t> [64] (ZeroExt16to64 y))))
   139(Lsh16x32 <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg16 <t> (SLTIU <t> [64] (ZeroExt32to64 y))))
   140(Lsh16x64 <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg16 <t> (SLTIU <t> [64] y)))
   141(Lsh32x8  <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg32 <t> (SLTIU <t> [64] (ZeroExt8to64  y))))
   142(Lsh32x16 <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg32 <t> (SLTIU <t> [64] (ZeroExt16to64 y))))
   143(Lsh32x32 <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg32 <t> (SLTIU <t> [64] (ZeroExt32to64 y))))
   144(Lsh32x64 <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg32 <t> (SLTIU <t> [64] y)))
   145(Lsh64x8  <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg64 <t> (SLTIU <t> [64] (ZeroExt8to64  y))))
   146(Lsh64x16 <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg64 <t> (SLTIU <t> [64] (ZeroExt16to64 y))))
   147(Lsh64x32 <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg64 <t> (SLTIU <t> [64] (ZeroExt32to64 y))))
   148(Lsh64x64 <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg64 <t> (SLTIU <t> [64] y)))
   149
   150(Lsh8x(64|32|16|8)  x y) && shiftIsBounded(v) => (SLL x y)
   151(Lsh16x(64|32|16|8) x y) && shiftIsBounded(v) => (SLL x y)
   152(Lsh32x(64|32|16|8) x y) && shiftIsBounded(v) => (SLL x y)
   153(Lsh64x(64|32|16|8) x y) && shiftIsBounded(v) => (SLL x y)
   154
   155// SRL only considers the bottom 6 bits of y, similarly SRLW only considers the
   156// bottom 5 bits of y. Ensure that the result is always zero if the shift exceeds
   157// the maximum value. See Lsh above for a detailed description.
   158(Rsh8Ux8   <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t> (ZeroExt8to64  x) y) (Neg8  <t> (SLTIU <t> [64] (ZeroExt8to64  y))))
   159(Rsh8Ux16  <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t> (ZeroExt8to64  x) y) (Neg8  <t> (SLTIU <t> [64] (ZeroExt16to64 y))))
   160(Rsh8Ux32  <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t> (ZeroExt8to64  x) y) (Neg8  <t> (SLTIU <t> [64] (ZeroExt32to64 y))))
   161(Rsh8Ux64  <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t> (ZeroExt8to64  x) y) (Neg8  <t> (SLTIU <t> [64] y)))
   162(Rsh16Ux8  <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t> (ZeroExt16to64 x) y) (Neg16 <t> (SLTIU <t> [64] (ZeroExt8to64  y))))
   163(Rsh16Ux16 <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t> (ZeroExt16to64 x) y) (Neg16 <t> (SLTIU <t> [64] (ZeroExt16to64 y))))
   164(Rsh16Ux32 <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t> (ZeroExt16to64 x) y) (Neg16 <t> (SLTIU <t> [64] (ZeroExt32to64 y))))
   165(Rsh16Ux64 <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t> (ZeroExt16to64 x) y) (Neg16 <t> (SLTIU <t> [64] y)))
   166(Rsh32Ux8  <t> x y) && !shiftIsBounded(v) => (AND (SRLW <t>  x                y) (Neg32 <t> (SLTIU <t> [32] (ZeroExt8to64  y))))
   167(Rsh32Ux16 <t> x y) && !shiftIsBounded(v) => (AND (SRLW <t>  x                y) (Neg32 <t> (SLTIU <t> [32] (ZeroExt16to64 y))))
   168(Rsh32Ux32 <t> x y) && !shiftIsBounded(v) => (AND (SRLW <t>  x                y) (Neg32 <t> (SLTIU <t> [32] (ZeroExt32to64 y))))
   169(Rsh32Ux64 <t> x y) && !shiftIsBounded(v) => (AND (SRLW <t>  x                y) (Neg32 <t> (SLTIU <t> [32] y)))
   170(Rsh64Ux8  <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t>  x                y) (Neg64 <t> (SLTIU <t> [64] (ZeroExt8to64  y))))
   171(Rsh64Ux16 <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t>  x                y) (Neg64 <t> (SLTIU <t> [64] (ZeroExt16to64 y))))
   172(Rsh64Ux32 <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t>  x                y) (Neg64 <t> (SLTIU <t> [64] (ZeroExt32to64 y))))
   173(Rsh64Ux64 <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t>  x                y) (Neg64 <t> (SLTIU <t> [64] y)))
   174
   175(Rsh8Ux(64|32|16|8)  x y) && shiftIsBounded(v) => (SRL  (ZeroExt8to64  x) y)
   176(Rsh16Ux(64|32|16|8) x y) && shiftIsBounded(v) => (SRL  (ZeroExt16to64 x) y)
   177(Rsh32Ux(64|32|16|8) x y) && shiftIsBounded(v) => (SRLW x                 y)
   178(Rsh64Ux(64|32|16|8) x y) && shiftIsBounded(v) => (SRL  x                 y)
   179
   180// SRA only considers the bottom 6 bits of y, similarly SRAW only considers the
   181// bottom 5 bits. If y is greater than the maximum value (either 63 or 31
   182// depending on the instruction),  the result of the shift should be either 0
   183// or -1 based on the sign bit of x.
   184//
   185// We implement this by performing the max shift (-1) if y > the maximum value.
   186//
   187// We OR (uint64(y < 64) - 1) into y before passing it to SRA. This leaves
   188// us with -1 (0xffff...) if y >= 64.  Similarly, we OR (uint64(y < 32) - 1) into y
   189// before passing it to SRAW.
   190//
   191// We don't need to sign-extend the OR result, as it will be at minimum 8 bits,
   192// more than the 5 or 6 bits SRAW and SRA care about.
   193(Rsh8x8   <t> x y) && !shiftIsBounded(v) => (SRA  <t> (SignExt8to64  x) (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] (ZeroExt8to64  y)))))
   194(Rsh8x16  <t> x y) && !shiftIsBounded(v) => (SRA  <t> (SignExt8to64  x) (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] (ZeroExt16to64 y)))))
   195(Rsh8x32  <t> x y) && !shiftIsBounded(v) => (SRA  <t> (SignExt8to64  x) (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] (ZeroExt32to64 y)))))
   196(Rsh8x64  <t> x y) && !shiftIsBounded(v) => (SRA  <t> (SignExt8to64  x) (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] y))))
   197(Rsh16x8  <t> x y) && !shiftIsBounded(v) => (SRA  <t> (SignExt16to64 x) (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] (ZeroExt8to64  y)))))
   198(Rsh16x16 <t> x y) && !shiftIsBounded(v) => (SRA  <t> (SignExt16to64 x) (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] (ZeroExt16to64 y)))))
   199(Rsh16x32 <t> x y) && !shiftIsBounded(v) => (SRA  <t> (SignExt16to64 x) (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] (ZeroExt32to64 y)))))
   200(Rsh16x64 <t> x y) && !shiftIsBounded(v) => (SRA  <t> (SignExt16to64 x) (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] y))))
   201(Rsh32x8  <t> x y) && !shiftIsBounded(v) => (SRAW <t> x                 (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [32] (ZeroExt8to64  y)))))
   202(Rsh32x16 <t> x y) && !shiftIsBounded(v) => (SRAW <t> x                 (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [32] (ZeroExt16to64 y)))))
   203(Rsh32x32 <t> x y) && !shiftIsBounded(v) => (SRAW <t> x                 (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [32] (ZeroExt32to64 y)))))
   204(Rsh32x64 <t> x y) && !shiftIsBounded(v) => (SRAW <t> x                 (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [32] y))))
   205(Rsh64x8  <t> x y) && !shiftIsBounded(v) => (SRA  <t> x                 (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] (ZeroExt8to64  y)))))
   206(Rsh64x16 <t> x y) && !shiftIsBounded(v) => (SRA  <t> x                 (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] (ZeroExt16to64 y)))))
   207(Rsh64x32 <t> x y) && !shiftIsBounded(v) => (SRA  <t> x                 (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] (ZeroExt32to64 y)))))
   208(Rsh64x64 <t> x y) && !shiftIsBounded(v) => (SRA  <t> x                 (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] y))))
   209
   210(Rsh8x(64|32|16|8)  x y) && shiftIsBounded(v) => (SRA  (SignExt8to64  x) y)
   211(Rsh16x(64|32|16|8) x y) && shiftIsBounded(v) => (SRA  (SignExt16to64 x) y)
   212(Rsh32x(64|32|16|8) x y) && shiftIsBounded(v) => (SRAW  x                y)
   213(Rsh64x(64|32|16|8) x y) && shiftIsBounded(v) => (SRA   x                y)
   214
   215// Rotates.
   216(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))))
   217(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))))
   218(RotateLeft32 ...) => (ROLW ...)
   219(RotateLeft64 ...) => (ROL  ...)
   220
   221(Less64  ...) => (SLT  ...)
   222(Less32  x y) => (SLT  (SignExt32to64 x) (SignExt32to64 y))
   223(Less16  x y) => (SLT  (SignExt16to64 x) (SignExt16to64 y))
   224(Less8   x y) => (SLT  (SignExt8to64  x) (SignExt8to64  y))
   225(Less64U ...) => (SLTU ...)
   226(Less32U x y) => (SLTU (ZeroExt32to64 x) (ZeroExt32to64 y))
   227(Less16U x y) => (SLTU (ZeroExt16to64 x) (ZeroExt16to64 y))
   228(Less8U  x y) => (SLTU (ZeroExt8to64  x) (ZeroExt8to64  y))
   229(Less(64|32)F ...) => (FLT(D|S) ...)
   230
   231// Convert x <= y to !(y > x).
   232(Leq(64|32|16|8)  x y) => (Not (Less(64|32|16|8)  y x))
   233(Leq(64|32|16|8)U x y) => (Not (Less(64|32|16|8)U y x))
   234(Leq(64|32)F ...) => (FLE(D|S) ...)
   235
   236(EqPtr x y) => (SEQZ (SUB <typ.Uintptr> x y))
   237(Eq64  x y) => (SEQZ (SUB <x.Type> x y))
   238(Eq32  x y) &&  x.Type.IsSigned() => (SEQZ (SUB <x.Type> (SignExt32to64 x) (SignExt32to64 y)))
   239(Eq32  x y) && !x.Type.IsSigned() => (SEQZ (SUB <x.Type> (ZeroExt32to64 x) (ZeroExt32to64 y)))
   240(Eq16  x y) => (SEQZ (SUB <x.Type> (ZeroExt16to64 x) (ZeroExt16to64 y)))
   241(Eq8   x y) => (SEQZ (SUB <x.Type> (ZeroExt8to64  x) (ZeroExt8to64  y)))
   242(Eq(64|32)F ...) => (FEQ(D|S) ...)
   243
   244(NeqPtr x y) => (Not (EqPtr x y))
   245(Neq64  x y) => (Not (Eq64  x y))
   246(Neq32  x y) => (Not (Eq32  x y))
   247(Neq16  x y) => (Not (Eq16  x y))
   248(Neq8   x y) => (Not (Eq8   x y))
   249(Neq(64|32)F ...) => (FNE(D|S) ...)
   250
   251// Loads
   252(Load <t> ptr mem) &&  t.IsBoolean()                   => (MOVBUload ptr mem)
   253(Load <t> ptr mem) && ( is8BitInt(t) &&  t.IsSigned()) => (MOVBload  ptr mem)
   254(Load <t> ptr mem) && ( is8BitInt(t) && !t.IsSigned()) => (MOVBUload ptr mem)
   255(Load <t> ptr mem) && (is16BitInt(t) &&  t.IsSigned()) => (MOVHload  ptr mem)
   256(Load <t> ptr mem) && (is16BitInt(t) && !t.IsSigned()) => (MOVHUload ptr mem)
   257(Load <t> ptr mem) && (is32BitInt(t) &&  t.IsSigned()) => (MOVWload  ptr mem)
   258(Load <t> ptr mem) && (is32BitInt(t) && !t.IsSigned()) => (MOVWUload ptr mem)
   259(Load <t> ptr mem) && (is64BitInt(t) || isPtr(t))      => (MOVDload  ptr mem)
   260(Load <t> ptr mem) &&  is32BitFloat(t)                 => (FMOVWload ptr mem)
   261(Load <t> ptr mem) &&  is64BitFloat(t)                 => (FMOVDload ptr mem)
   262
   263// Stores
   264(Store {t} ptr val mem) && t.Size() == 1 => (MOVBstore ptr val mem)
   265(Store {t} ptr val mem) && t.Size() == 2 => (MOVHstore ptr val mem)
   266(Store {t} ptr val mem) && t.Size() == 4 && !t.IsFloat() => (MOVWstore ptr val mem)
   267(Store {t} ptr val mem) && t.Size() == 8 && !t.IsFloat() => (MOVDstore ptr val mem)
   268(Store {t} ptr val mem) && t.Size() == 4 &&  t.IsFloat() => (FMOVWstore ptr val mem)
   269(Store {t} ptr val mem) && t.Size() == 8 &&  t.IsFloat() => (FMOVDstore ptr val mem)
   270
   271// We need to fold MOVaddr into the LD/MOVDstore ops so that the live variable analysis
   272// knows what variables are being read/written by the ops.
   273(MOVBUload [off1] {sym1} (MOVaddr [off2] {sym2} base) mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   274	(MOVBUload [off1+off2] {mergeSym(sym1,sym2)} base mem)
   275(MOVBload  [off1] {sym1} (MOVaddr [off2] {sym2} base) mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   276	(MOVBload  [off1+off2] {mergeSym(sym1,sym2)} base mem)
   277(MOVHUload [off1] {sym1} (MOVaddr [off2] {sym2} base) mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   278	(MOVHUload [off1+off2] {mergeSym(sym1,sym2)} base mem)
   279(MOVHload  [off1] {sym1} (MOVaddr [off2] {sym2} base) mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   280	(MOVHload  [off1+off2] {mergeSym(sym1,sym2)} base mem)
   281(MOVWUload [off1] {sym1} (MOVaddr [off2] {sym2} base) mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   282	(MOVWUload [off1+off2] {mergeSym(sym1,sym2)} base mem)
   283(MOVWload  [off1] {sym1} (MOVaddr [off2] {sym2} base) mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   284	(MOVWload  [off1+off2] {mergeSym(sym1,sym2)} base mem)
   285(MOVDload  [off1] {sym1} (MOVaddr [off2] {sym2} base) mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   286	(MOVDload  [off1+off2] {mergeSym(sym1,sym2)} base mem)
   287
   288(MOVBstore [off1] {sym1} (MOVaddr [off2] {sym2} base) val mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   289	(MOVBstore [off1+off2] {mergeSym(sym1,sym2)} base val mem)
   290(MOVHstore [off1] {sym1} (MOVaddr [off2] {sym2} base) val mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   291	(MOVHstore [off1+off2] {mergeSym(sym1,sym2)} base val mem)
   292(MOVWstore [off1] {sym1} (MOVaddr [off2] {sym2} base) val mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   293	(MOVWstore [off1+off2] {mergeSym(sym1,sym2)} base val mem)
   294(MOVDstore [off1] {sym1} (MOVaddr [off2] {sym2} base) val mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   295	(MOVDstore [off1+off2] {mergeSym(sym1,sym2)} base val mem)
   296(MOVBstorezero [off1] {sym1} (MOVaddr [off2] {sym2} ptr) mem) && canMergeSym(sym1,sym2) && is32Bit(int64(off1)+int64(off2)) =>
   297	(MOVBstorezero [off1+off2] {mergeSym(sym1,sym2)} ptr mem)
   298(MOVHstorezero [off1] {sym1} (MOVaddr [off2] {sym2} ptr) mem) && canMergeSym(sym1,sym2) && is32Bit(int64(off1)+int64(off2)) =>
   299	(MOVHstorezero [off1+off2] {mergeSym(sym1,sym2)} ptr mem)
   300(MOVWstorezero [off1] {sym1} (MOVaddr [off2] {sym2} ptr) mem) && canMergeSym(sym1,sym2) && is32Bit(int64(off1)+int64(off2)) =>
   301	(MOVWstorezero [off1+off2] {mergeSym(sym1,sym2)} ptr mem)
   302(MOVDstorezero [off1] {sym1} (MOVaddr [off2] {sym2} ptr) mem) && canMergeSym(sym1,sym2) && is32Bit(int64(off1)+int64(off2)) =>
   303	(MOVDstorezero [off1+off2] {mergeSym(sym1,sym2)} ptr mem)
   304
   305(MOVBUload [off1] {sym} (ADDI [off2] base) mem) && is32Bit(int64(off1)+off2) =>
   306	(MOVBUload [off1+int32(off2)] {sym} base mem)
   307(MOVBload  [off1] {sym} (ADDI [off2] base) mem) && is32Bit(int64(off1)+off2) =>
   308	(MOVBload  [off1+int32(off2)] {sym} base mem)
   309(MOVHUload [off1] {sym} (ADDI [off2] base) mem) && is32Bit(int64(off1)+off2) =>
   310	(MOVHUload [off1+int32(off2)] {sym} base mem)
   311(MOVHload  [off1] {sym} (ADDI [off2] base) mem) && is32Bit(int64(off1)+off2) =>
   312	(MOVHload  [off1+int32(off2)] {sym} base mem)
   313(MOVWUload [off1] {sym} (ADDI [off2] base) mem) && is32Bit(int64(off1)+off2) =>
   314	(MOVWUload [off1+int32(off2)] {sym} base mem)
   315(MOVWload  [off1] {sym} (ADDI [off2] base) mem) && is32Bit(int64(off1)+off2) =>
   316	(MOVWload  [off1+int32(off2)] {sym} base mem)
   317(MOVDload  [off1] {sym} (ADDI [off2] base) mem) && is32Bit(int64(off1)+off2) =>
   318	(MOVDload  [off1+int32(off2)] {sym} base mem)
   319
   320(MOVBstore [off1] {sym} (ADDI [off2] base) val mem) && is32Bit(int64(off1)+off2) =>
   321	(MOVBstore [off1+int32(off2)] {sym} base val mem)
   322(MOVHstore [off1] {sym} (ADDI [off2] base) val mem) && is32Bit(int64(off1)+off2) =>
   323	(MOVHstore [off1+int32(off2)] {sym} base val mem)
   324(MOVWstore [off1] {sym} (ADDI [off2] base) val mem) && is32Bit(int64(off1)+off2) =>
   325	(MOVWstore [off1+int32(off2)] {sym} base val mem)
   326(MOVDstore [off1] {sym} (ADDI [off2] base) val mem) && is32Bit(int64(off1)+off2) =>
   327	(MOVDstore [off1+int32(off2)] {sym} base val mem)
   328(MOVBstorezero [off1] {sym} (ADDI [off2] ptr) mem) && is32Bit(int64(off1)+off2) => (MOVBstorezero [off1+int32(off2)] {sym} ptr mem)
   329(MOVHstorezero [off1] {sym} (ADDI [off2] ptr) mem) && is32Bit(int64(off1)+off2) => (MOVHstorezero [off1+int32(off2)] {sym} ptr mem)
   330(MOVWstorezero [off1] {sym} (ADDI [off2] ptr) mem) && is32Bit(int64(off1)+off2) => (MOVWstorezero [off1+int32(off2)] {sym} ptr mem)
   331(MOVDstorezero [off1] {sym} (ADDI [off2] ptr) mem) && is32Bit(int64(off1)+off2) => (MOVDstorezero [off1+int32(off2)] {sym} ptr mem)
   332
   333// Similarly, fold ADDI into MOVaddr to avoid confusing live variable analysis
   334// with OffPtr -> ADDI.
   335(ADDI [c] (MOVaddr [d] {s} x)) && is32Bit(c+int64(d)) => (MOVaddr [int32(c)+d] {s} x)
   336
   337// Small zeroing
   338(Zero [0] _ mem) => mem
   339(Zero [1] ptr mem) => (MOVBstore ptr (MOVDconst [0]) mem)
   340(Zero [2] {t} ptr mem) && t.Alignment()%2 == 0 =>
   341	(MOVHstore ptr (MOVDconst [0]) mem)
   342(Zero [2] ptr mem) =>
   343	(MOVBstore [1] ptr (MOVDconst [0])
   344		(MOVBstore ptr (MOVDconst [0]) mem))
   345(Zero [4] {t} ptr mem) && t.Alignment()%4 == 0 =>
   346	(MOVWstore ptr (MOVDconst [0]) mem)
   347(Zero [4] {t} ptr mem) && t.Alignment()%2 == 0 =>
   348	(MOVHstore [2] ptr (MOVDconst [0])
   349		(MOVHstore ptr (MOVDconst [0]) mem))
   350(Zero [4] ptr mem) =>
   351	(MOVBstore [3] ptr (MOVDconst [0])
   352		(MOVBstore [2] ptr (MOVDconst [0])
   353			(MOVBstore [1] ptr (MOVDconst [0])
   354				(MOVBstore ptr (MOVDconst [0]) mem))))
   355(Zero [8] {t} ptr mem) && t.Alignment()%8 == 0 =>
   356	(MOVDstore ptr (MOVDconst [0]) mem)
   357(Zero [8] {t} ptr mem) && t.Alignment()%4 == 0 =>
   358	(MOVWstore [4] ptr (MOVDconst [0])
   359		(MOVWstore ptr (MOVDconst [0]) mem))
   360(Zero [8] {t} ptr mem) && t.Alignment()%2 == 0 =>
   361	(MOVHstore [6] ptr (MOVDconst [0])
   362		(MOVHstore [4] ptr (MOVDconst [0])
   363			(MOVHstore [2] ptr (MOVDconst [0])
   364				(MOVHstore ptr (MOVDconst [0]) mem))))
   365
   366(Zero [3] ptr mem) =>
   367	(MOVBstore [2] ptr (MOVDconst [0])
   368		(MOVBstore [1] ptr (MOVDconst [0])
   369			(MOVBstore ptr (MOVDconst [0]) mem)))
   370(Zero [6] {t} ptr mem) && t.Alignment()%2 == 0 =>
   371	(MOVHstore [4] ptr (MOVDconst [0])
   372		(MOVHstore [2] ptr (MOVDconst [0])
   373			(MOVHstore ptr (MOVDconst [0]) mem)))
   374(Zero [12] {t} ptr mem) && t.Alignment()%4 == 0 =>
   375	(MOVWstore [8] ptr (MOVDconst [0])
   376		(MOVWstore [4] ptr (MOVDconst [0])
   377			(MOVWstore ptr (MOVDconst [0]) mem)))
   378(Zero [16] {t} ptr mem) && t.Alignment()%8 == 0 =>
   379	(MOVDstore [8] ptr (MOVDconst [0])
   380		(MOVDstore ptr (MOVDconst [0]) mem))
   381(Zero [24] {t} ptr mem) && t.Alignment()%8 == 0 =>
   382	(MOVDstore [16] ptr (MOVDconst [0])
   383		(MOVDstore [8] ptr (MOVDconst [0])
   384			(MOVDstore ptr (MOVDconst [0]) mem)))
   385(Zero [32] {t} ptr mem) && t.Alignment()%8 == 0 =>
   386	(MOVDstore [24] ptr (MOVDconst [0])
   387		(MOVDstore [16] ptr (MOVDconst [0])
   388			(MOVDstore [8] ptr (MOVDconst [0])
   389				(MOVDstore ptr (MOVDconst [0]) mem))))
   390
   391// Medium 8-aligned zeroing uses a Duff's device
   392// 8 and 128 are magic constants, see runtime/mkduff.go
   393(Zero [s] {t} ptr mem)
   394	&& s%8 == 0 && s <= 8*128
   395	&& t.Alignment()%8 == 0 && !config.noDuffDevice =>
   396	(DUFFZERO [8 * (128 - s/8)] ptr mem)
   397
   398// Generic zeroing uses a loop
   399(Zero [s] {t} ptr mem) =>
   400	(LoweredZero [t.Alignment()]
   401		ptr
   402		(ADD <ptr.Type> ptr (MOVDconst [s-moveSize(t.Alignment(), config)]))
   403		mem)
   404
   405// Checks
   406(IsNonNil ...) => (SNEZ ...)
   407(IsInBounds ...) => (Less64U ...)
   408(IsSliceInBounds ...) => (Leq64U ...)
   409
   410// Trivial lowering
   411(NilCheck ...) => (LoweredNilCheck ...)
   412(GetClosurePtr ...) => (LoweredGetClosurePtr ...)
   413(GetCallerSP ...) => (LoweredGetCallerSP ...)
   414(GetCallerPC ...) => (LoweredGetCallerPC ...)
   415
   416// Write barrier.
   417(WB ...) => (LoweredWB ...)
   418
   419// Publication barrier as intrinsic
   420(PubBarrier ...) => (LoweredPubBarrier ...)
   421
   422(PanicBounds [kind] x y mem) && boundsABI(kind) == 0 => (LoweredPanicBoundsA [kind] x y mem)
   423(PanicBounds [kind] x y mem) && boundsABI(kind) == 1 => (LoweredPanicBoundsB [kind] x y mem)
   424(PanicBounds [kind] x y mem) && boundsABI(kind) == 2 => (LoweredPanicBoundsC [kind] x y mem)
   425
   426// Small moves
   427(Move [0] _ _ mem) => mem
   428(Move [1] dst src mem) => (MOVBstore dst (MOVBload src mem) mem)
   429(Move [2] {t} dst src mem) && t.Alignment()%2 == 0 =>
   430	(MOVHstore dst (MOVHload src mem) mem)
   431(Move [2] dst src mem) =>
   432	(MOVBstore [1] dst (MOVBload [1] src mem)
   433		(MOVBstore dst (MOVBload src mem) mem))
   434(Move [4] {t} dst src mem) && t.Alignment()%4 == 0 =>
   435	(MOVWstore dst (MOVWload src mem) mem)
   436(Move [4] {t} dst src mem) && t.Alignment()%2 == 0 =>
   437	(MOVHstore [2] dst (MOVHload [2] src mem)
   438		(MOVHstore dst (MOVHload src mem) mem))
   439(Move [4] dst src mem) =>
   440	(MOVBstore [3] dst (MOVBload [3] src mem)
   441		(MOVBstore [2] dst (MOVBload [2] src mem)
   442			(MOVBstore [1] dst (MOVBload [1] src mem)
   443				(MOVBstore dst (MOVBload src mem) mem))))
   444(Move [8] {t} dst src mem) && t.Alignment()%8 == 0 =>
   445	(MOVDstore dst (MOVDload src mem) mem)
   446(Move [8] {t} dst src mem) && t.Alignment()%4 == 0 =>
   447	(MOVWstore [4] dst (MOVWload [4] src mem)
   448		(MOVWstore dst (MOVWload src mem) mem))
   449(Move [8] {t} dst src mem) && t.Alignment()%2 == 0 =>
   450	(MOVHstore [6] dst (MOVHload [6] src mem)
   451		(MOVHstore [4] dst (MOVHload [4] src mem)
   452			(MOVHstore [2] dst (MOVHload [2] src mem)
   453				(MOVHstore dst (MOVHload src mem) mem))))
   454
   455(Move [3] dst src mem) =>
   456	(MOVBstore [2] dst (MOVBload [2] src mem)
   457		(MOVBstore [1] dst (MOVBload [1] src mem)
   458			(MOVBstore dst (MOVBload src mem) mem)))
   459(Move [6] {t} dst src mem) && t.Alignment()%2 == 0 =>
   460	(MOVHstore [4] dst (MOVHload [4] src mem)
   461		(MOVHstore [2] dst (MOVHload [2] src mem)
   462			(MOVHstore dst (MOVHload src mem) mem)))
   463(Move [12] {t} dst src mem) && t.Alignment()%4 == 0 =>
   464	(MOVWstore [8] dst (MOVWload [8] src mem)
   465		(MOVWstore [4] dst (MOVWload [4] src mem)
   466			(MOVWstore dst (MOVWload src mem) mem)))
   467(Move [16] {t} dst src mem) && t.Alignment()%8 == 0 =>
   468	(MOVDstore [8] dst (MOVDload [8] src mem)
   469		(MOVDstore dst (MOVDload src mem) mem))
   470(Move [24] {t} dst src mem) && t.Alignment()%8 == 0 =>
   471	(MOVDstore [16] dst (MOVDload [16] src mem)
   472		(MOVDstore [8] dst (MOVDload [8] src mem)
   473			(MOVDstore dst (MOVDload src mem) mem)))
   474(Move [32] {t} dst src mem) && t.Alignment()%8 == 0 =>
   475	(MOVDstore [24] dst (MOVDload [24] src mem)
   476		(MOVDstore [16] dst (MOVDload [16] src mem)
   477			(MOVDstore [8] dst (MOVDload [8] src mem)
   478				(MOVDstore dst (MOVDload src mem) mem))))
   479
   480// Medium 8-aligned move uses a Duff's device
   481// 16 and 128 are magic constants, see runtime/mkduff.go
   482(Move [s] {t} dst src mem)
   483	&& s%8 == 0 && s <= 8*128 && t.Alignment()%8 == 0
   484	&& !config.noDuffDevice && logLargeCopy(v, s) =>
   485	(DUFFCOPY [16 * (128 - s/8)] dst src mem)
   486
   487// Generic move uses a loop
   488(Move [s] {t} dst src mem) && (s <= 16 || logLargeCopy(v, s)) =>
   489	(LoweredMove [t.Alignment()]
   490		dst
   491		src
   492		(ADDI <src.Type> [s-moveSize(t.Alignment(), config)] src)
   493		mem)
   494
   495// Boolean ops; 0=false, 1=true
   496(AndB ...) => (AND ...)
   497(OrB  ...) => (OR  ...)
   498(EqB  x y) => (SEQZ (SUB <typ.Bool> x y))
   499(NeqB x y) => (SNEZ (SUB <typ.Bool> x y))
   500(Not  ...) => (SEQZ ...)
   501
   502// Lowering pointer arithmetic
   503// TODO: Special handling for SP offsets, like ARM
   504(OffPtr [off] ptr:(SP)) && is32Bit(off) => (MOVaddr [int32(off)] ptr)
   505(OffPtr [off] ptr) && is32Bit(off) => (ADDI [off] ptr)
   506(OffPtr [off] ptr) => (ADD (MOVDconst [off]) ptr)
   507
   508(Const(64|32|16|8) [val]) => (MOVDconst [int64(val)])
   509(Const32F [val]) => (FMVSX (MOVDconst [int64(math.Float32bits(val))]))
   510(Const64F [val]) => (FMVDX (MOVDconst [int64(math.Float64bits(val))]))
   511(ConstNil) => (MOVDconst [0])
   512(ConstBool [val]) => (MOVDconst [int64(b2i(val))])
   513
   514(Addr {sym} base) => (MOVaddr {sym} [0] base)
   515(LocalAddr <t> {sym} base mem) && t.Elem().HasPointers() => (MOVaddr {sym} (SPanchored base mem))
   516(LocalAddr <t> {sym} base _)  && !t.Elem().HasPointers() => (MOVaddr {sym} base)
   517
   518// Calls
   519(StaticCall  ...) => (CALLstatic  ...)
   520(ClosureCall ...) => (CALLclosure ...)
   521(InterCall   ...) => (CALLinter   ...)
   522(TailCall ...) => (CALLtail ...)
   523
   524// Atomic Intrinsics
   525(AtomicLoad(Ptr|64|32|8)  ...) => (LoweredAtomicLoad(64|64|32|8) ...)
   526(AtomicStore(PtrNoWB|64|32|8) ...) => (LoweredAtomicStore(64|64|32|8) ...)
   527(AtomicAdd(64|32) ...) => (LoweredAtomicAdd(64|32) ...)
   528
   529// AtomicAnd8(ptr,val) => LoweredAtomicAnd32(ptr&^3, ^((uint8(val) ^ 0xff) << ((ptr & 3) * 8)))
   530(AtomicAnd8 ptr val mem) =>
   531	(LoweredAtomicAnd32 (ANDI <typ.Uintptr> [^3] ptr)
   532		(NOT <typ.UInt32> (SLL <typ.UInt32> (XORI <typ.UInt32> [0xff] (ZeroExt8to32 val))
   533			(SLLI <typ.UInt64> [3] (ANDI <typ.UInt64> [3] ptr)))) mem)
   534
   535(AtomicAnd32 ...) => (LoweredAtomicAnd32 ...)
   536
   537(AtomicCompareAndSwap32 ptr old new mem) => (LoweredAtomicCas32 ptr (SignExt32to64 old) new mem)
   538(AtomicCompareAndSwap64 ...) => (LoweredAtomicCas64 ...)
   539
   540(AtomicExchange(64|32) ...) => (LoweredAtomicExchange(64|32) ...)
   541
   542// AtomicOr8(ptr,val)  => LoweredAtomicOr32(ptr&^3, uint32(val)<<((ptr&3)*8))
   543(AtomicOr8 ptr val mem) =>
   544	(LoweredAtomicOr32 (ANDI <typ.Uintptr> [^3] ptr)
   545		(SLL <typ.UInt32> (ZeroExt8to32 val)
   546			(SLLI <typ.UInt64> [3] (ANDI <typ.UInt64> [3] ptr))) mem)
   547
   548(AtomicOr32  ...) => (LoweredAtomicOr32  ...)
   549
   550// Conditional branches
   551(If cond yes no) => (BNEZ (MOVBUreg <typ.UInt64> cond) yes no)
   552
   553// Optimizations
   554
   555// Absorb SEQZ/SNEZ into branch.
   556(BEQZ (SEQZ x) yes no) => (BNEZ x yes no)
   557(BEQZ (SNEZ x) yes no) => (BEQZ x yes no)
   558(BNEZ (SEQZ x) yes no) => (BEQZ x yes no)
   559(BNEZ (SNEZ x) yes no) => (BNEZ x yes no)
   560
   561// Remove redundant NEG from BEQZ/BNEZ.
   562(BEQZ (NEG x) yes no) => (BEQZ x yes no)
   563(BNEZ (NEG x) yes no) => (BNEZ x yes no)
   564
   565// Negate comparison with FNES/FNED.
   566(BEQZ (FNES <t> x y) yes no) => (BNEZ (FEQS <t> x y) yes no)
   567(BNEZ (FNES <t> x y) yes no) => (BEQZ (FEQS <t> x y) yes no)
   568(BEQZ (FNED <t> x y) yes no) => (BNEZ (FEQD <t> x y) yes no)
   569(BNEZ (FNED <t> x y) yes no) => (BEQZ (FEQD <t> x y) yes no)
   570
   571// Convert BEQZ/BNEZ into more optimal branch conditions.
   572(BEQZ (SUB x y) yes no) => (BEQ x y yes no)
   573(BNEZ (SUB x y) yes no) => (BNE x y yes no)
   574(BEQZ (SLT x y) yes no) => (BGE x y yes no)
   575(BNEZ (SLT x y) yes no) => (BLT x y yes no)
   576(BEQZ (SLTU x y) yes no) => (BGEU x y yes no)
   577(BNEZ (SLTU x y) yes no) => (BLTU x y yes no)
   578(BEQZ (SLTI [x] y) yes no) => (BGE y (MOVDconst [x]) yes no)
   579(BNEZ (SLTI [x] y) yes no) => (BLT y (MOVDconst [x]) yes no)
   580(BEQZ (SLTIU [x] y) yes no) => (BGEU y (MOVDconst [x]) yes no)
   581(BNEZ (SLTIU [x] y) yes no) => (BLTU y (MOVDconst [x]) yes no)
   582
   583// Convert branch with zero to more optimal branch zero.
   584(BEQ (MOVDconst [0]) cond yes no) => (BEQZ cond yes no)
   585(BEQ cond (MOVDconst [0]) yes no) => (BEQZ cond yes no)
   586(BNE (MOVDconst [0]) cond yes no) => (BNEZ cond yes no)
   587(BNE cond (MOVDconst [0]) yes no) => (BNEZ cond yes no)
   588(BLT (MOVDconst [0]) cond yes no) => (BGTZ cond yes no)
   589(BLT cond (MOVDconst [0]) yes no) => (BLTZ cond yes no)
   590(BGE (MOVDconst [0]) cond yes no) => (BLEZ cond yes no)
   591(BGE cond (MOVDconst [0]) yes no) => (BGEZ cond yes no)
   592
   593// Remove redundant NEG from SEQZ/SNEZ.
   594(SEQZ (NEG x)) => (SEQZ x)
   595(SNEZ (NEG x)) => (SNEZ x)
   596
   597// Remove redundant SEQZ/SNEZ.
   598(SEQZ (SEQZ x)) => (SNEZ x)
   599(SEQZ (SNEZ x)) => (SEQZ x)
   600(SNEZ (SEQZ x)) => (SEQZ x)
   601(SNEZ (SNEZ x)) => (SNEZ x)
   602
   603// Store zero.
   604(MOVBstore [off] {sym} ptr (MOVDconst [0]) mem) => (MOVBstorezero [off] {sym} ptr mem)
   605(MOVHstore [off] {sym} ptr (MOVDconst [0]) mem) => (MOVHstorezero [off] {sym} ptr mem)
   606(MOVWstore [off] {sym} ptr (MOVDconst [0]) mem) => (MOVWstorezero [off] {sym} ptr mem)
   607(MOVDstore [off] {sym} ptr (MOVDconst [0]) mem) => (MOVDstorezero [off] {sym} ptr mem)
   608
   609// Boolean ops are already extended.
   610(MOVBUreg x:((FLES|FLTS|FEQS|FNES) _ _)) => x
   611(MOVBUreg x:((FLED|FLTD|FEQD|FNED) _ _)) => x
   612(MOVBUreg x:((SEQZ|SNEZ) _)) => x
   613(MOVBUreg x:((SLT|SLTU) _ _)) => x
   614
   615// Avoid extending when already sufficiently masked.
   616(MOVBreg  x:(ANDI [c] y)) && c >= 0 && int64(int8(c)) == c => x
   617(MOVHreg  x:(ANDI [c] y)) && c >= 0 && int64(int16(c)) == c => x
   618(MOVWreg  x:(ANDI [c] y)) && c >= 0 && int64(int32(c)) == c => x
   619(MOVBUreg x:(ANDI [c] y)) && c >= 0 && int64(uint8(c)) == c => x
   620(MOVHUreg x:(ANDI [c] y)) && c >= 0 && int64(uint16(c)) == c => x
   621(MOVWUreg x:(ANDI [c] y)) && c >= 0 && int64(uint32(c)) == c => x
   622
   623// Combine masking and zero extension.
   624(MOVBUreg (ANDI [c] x)) && c < 0 => (ANDI [int64(uint8(c))] x)
   625(MOVHUreg (ANDI [c] x)) && c < 0 => (ANDI [int64(uint16(c))] x)
   626(MOVWUreg (ANDI [c] x)) && c < 0 => (AND (MOVDconst [int64(uint32(c))]) x)
   627
   628// Avoid sign/zero extension for consts.
   629(MOVBreg  (MOVDconst [c])) => (MOVDconst [int64(int8(c))])
   630(MOVHreg  (MOVDconst [c])) => (MOVDconst [int64(int16(c))])
   631(MOVWreg  (MOVDconst [c])) => (MOVDconst [int64(int32(c))])
   632(MOVBUreg (MOVDconst [c])) => (MOVDconst [int64(uint8(c))])
   633(MOVHUreg (MOVDconst [c])) => (MOVDconst [int64(uint16(c))])
   634(MOVWUreg (MOVDconst [c])) => (MOVDconst [int64(uint32(c))])
   635
   636// Avoid sign/zero extension after properly typed load.
   637(MOVBreg  x:(MOVBload  _ _)) => (MOVDreg x)
   638(MOVHreg  x:(MOVBload  _ _)) => (MOVDreg x)
   639(MOVHreg  x:(MOVBUload _ _)) => (MOVDreg x)
   640(MOVHreg  x:(MOVHload  _ _)) => (MOVDreg x)
   641(MOVWreg  x:(MOVBload  _ _)) => (MOVDreg x)
   642(MOVWreg  x:(MOVBUload _ _)) => (MOVDreg x)
   643(MOVWreg  x:(MOVHload  _ _)) => (MOVDreg x)
   644(MOVWreg  x:(MOVHUload _ _)) => (MOVDreg x)
   645(MOVWreg  x:(MOVWload  _ _)) => (MOVDreg x)
   646(MOVBUreg x:(MOVBUload _ _)) => (MOVDreg x)
   647(MOVHUreg x:(MOVBUload _ _)) => (MOVDreg x)
   648(MOVHUreg x:(MOVHUload _ _)) => (MOVDreg x)
   649(MOVWUreg x:(MOVBUload _ _)) => (MOVDreg x)
   650(MOVWUreg x:(MOVHUload _ _)) => (MOVDreg x)
   651(MOVWUreg x:(MOVWUload _ _)) => (MOVDreg x)
   652
   653// Avoid zero extension after properly typed atomic operation.
   654(MOVBUreg x:(Select0 (LoweredAtomicLoad8 _ _))) => (MOVDreg x)
   655(MOVBUreg x:(Select0 (LoweredAtomicCas32 _ _ _ _))) => (MOVDreg x)
   656(MOVBUreg x:(Select0 (LoweredAtomicCas64 _ _ _ _))) => (MOVDreg x)
   657
   658// Avoid sign extension after word arithmetic.
   659(MOVWreg x:(ADDIW   _)) => (MOVDreg x)
   660(MOVWreg x:(SUBW  _ _)) => (MOVDreg x)
   661(MOVWreg x:(NEGW    _)) => (MOVDreg x)
   662(MOVWreg x:(MULW  _ _)) => (MOVDreg x)
   663(MOVWreg x:(DIVW  _ _)) => (MOVDreg x)
   664(MOVWreg x:(DIVUW _ _)) => (MOVDreg x)
   665(MOVWreg x:(REMW  _ _)) => (MOVDreg x)
   666(MOVWreg x:(REMUW _ _)) => (MOVDreg x)
   667(MOVWreg x:(ROLW  _ _)) => (MOVDreg x)
   668(MOVWreg x:(RORW  _ _)) => (MOVDreg x)
   669(MOVWreg x:(RORIW   _)) => (MOVDreg x)
   670
   671// Fold double extensions.
   672(MOVBreg  x:(MOVBreg  _)) => (MOVDreg x)
   673(MOVHreg  x:(MOVBreg  _)) => (MOVDreg x)
   674(MOVHreg  x:(MOVBUreg _)) => (MOVDreg x)
   675(MOVHreg  x:(MOVHreg  _)) => (MOVDreg x)
   676(MOVWreg  x:(MOVBreg  _)) => (MOVDreg x)
   677(MOVWreg  x:(MOVBUreg _)) => (MOVDreg x)
   678(MOVWreg  x:(MOVHreg  _)) => (MOVDreg x)
   679(MOVWreg  x:(MOVWreg  _)) => (MOVDreg x)
   680(MOVBUreg x:(MOVBUreg _)) => (MOVDreg x)
   681(MOVHUreg x:(MOVBUreg _)) => (MOVDreg x)
   682(MOVHUreg x:(MOVHUreg _)) => (MOVDreg x)
   683(MOVWUreg x:(MOVBUreg _)) => (MOVDreg x)
   684(MOVWUreg x:(MOVHUreg _)) => (MOVDreg x)
   685(MOVWUreg x:(MOVWUreg _)) => (MOVDreg x)
   686
   687// Do not extend before store.
   688(MOVBstore [off] {sym} ptr (MOVBreg  x) mem) => (MOVBstore [off] {sym} ptr x mem)
   689(MOVBstore [off] {sym} ptr (MOVHreg  x) mem) => (MOVBstore [off] {sym} ptr x mem)
   690(MOVBstore [off] {sym} ptr (MOVWreg  x) mem) => (MOVBstore [off] {sym} ptr x mem)
   691(MOVBstore [off] {sym} ptr (MOVBUreg x) mem) => (MOVBstore [off] {sym} ptr x mem)
   692(MOVBstore [off] {sym} ptr (MOVHUreg x) mem) => (MOVBstore [off] {sym} ptr x mem)
   693(MOVBstore [off] {sym} ptr (MOVWUreg x) mem) => (MOVBstore [off] {sym} ptr x mem)
   694(MOVHstore [off] {sym} ptr (MOVHreg  x) mem) => (MOVHstore [off] {sym} ptr x mem)
   695(MOVHstore [off] {sym} ptr (MOVWreg  x) mem) => (MOVHstore [off] {sym} ptr x mem)
   696(MOVHstore [off] {sym} ptr (MOVHUreg x) mem) => (MOVHstore [off] {sym} ptr x mem)
   697(MOVHstore [off] {sym} ptr (MOVWUreg x) mem) => (MOVHstore [off] {sym} ptr x mem)
   698(MOVWstore [off] {sym} ptr (MOVWreg  x) mem) => (MOVWstore [off] {sym} ptr x mem)
   699(MOVWstore [off] {sym} ptr (MOVWUreg x) mem) => (MOVWstore [off] {sym} ptr x mem)
   700
   701// Replace extend after load with alternate load where possible.
   702(MOVBreg  <t> x:(MOVBUload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVBload  <t> [off] {sym} ptr mem)
   703(MOVHreg  <t> x:(MOVHUload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVHload  <t> [off] {sym} ptr mem)
   704(MOVWreg  <t> x:(MOVWUload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVWload  <t> [off] {sym} ptr mem)
   705(MOVBUreg <t> x:(MOVBload  [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVBUload <t> [off] {sym} ptr mem)
   706(MOVHUreg <t> x:(MOVHload  [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVHUload <t> [off] {sym} ptr mem)
   707(MOVWUreg <t> x:(MOVWload  [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVWUload <t> [off] {sym} ptr mem)
   708
   709// If a register move has only 1 use, just use the same register without emitting instruction
   710// MOVnop does not emit an instruction, only for ensuring the type.
   711(MOVDreg x) && x.Uses == 1 => (MOVDnop x)
   712
   713// TODO: we should be able to get rid of MOVDnop all together.
   714// But for now, this is enough to get rid of lots of them.
   715(MOVDnop (MOVDconst [c])) => (MOVDconst [c])
   716
   717// Avoid unnecessary zero and sign extension when right shifting.
   718(SRAI <t> [x] (MOVWreg  y)) && x >= 0 && x <= 31 => (SRAIW <t> [int64(x)] y)
   719(SRLI <t> [x] (MOVWUreg y)) && x >= 0 && x <= 31 => (SRLIW <t> [int64(x)] y)
   720
   721// Replace right shifts that exceed size of signed type.
   722(SRAI <t> [x] (MOVBreg y)) && x >=  8 => (SRAI  [63] (SLLI <t> [56] y))
   723(SRAI <t> [x] (MOVHreg y)) && x >= 16 => (SRAI  [63] (SLLI <t> [48] y))
   724(SRAI <t> [x] (MOVWreg y)) && x >= 32 => (SRAIW [31] y)
   725
   726// Eliminate right shifts that exceed size of unsigned type.
   727(SRLI <t> [x] (MOVBUreg y)) && x >=  8 => (MOVDconst <t> [0])
   728(SRLI <t> [x] (MOVHUreg y)) && x >= 16 => (MOVDconst <t> [0])
   729(SRLI <t> [x] (MOVWUreg y)) && x >= 32 => (MOVDconst <t> [0])
   730
   731// Fold constant into immediate instructions where possible.
   732(ADD (MOVDconst <t> [val]) x) && is32Bit(val) && !t.IsPtr() => (ADDI [val] x)
   733(AND (MOVDconst [val]) x) && is32Bit(val) => (ANDI [val] x)
   734(OR  (MOVDconst [val]) x) && is32Bit(val) => (ORI  [val] x)
   735(XOR (MOVDconst [val]) x) && is32Bit(val) => (XORI [val] x)
   736(ROL  x (MOVDconst [val])) => (RORI  [int64(int8(-val)&63)] x)
   737(ROLW x (MOVDconst [val])) => (RORIW [int64(int8(-val)&31)] x)
   738(ROR  x (MOVDconst [val])) => (RORI  [int64(val&63)] x)
   739(RORW x (MOVDconst [val])) => (RORIW [int64(val&31)] x)
   740(SLL  x (MOVDconst [val])) => (SLLI [int64(val&63)] x)
   741(SRL  x (MOVDconst [val])) => (SRLI [int64(val&63)] x)
   742(SLLW x (MOVDconst [val])) => (SLLIW [int64(val&31)] x)
   743(SRLW x (MOVDconst [val])) => (SRLIW [int64(val&31)] x)
   744(SRA  x (MOVDconst [val])) => (SRAI [int64(val&63)] x)
   745(SRAW x (MOVDconst [val])) => (SRAIW [int64(val&31)] x)
   746(SLT  x (MOVDconst [val])) && val >= -2048 && val <= 2047 => (SLTI  [val] x)
   747(SLTU x (MOVDconst [val])) && val >= -2048 && val <= 2047 => (SLTIU [val] x)
   748
   749// Replace negated left rotation with right rotation.
   750(ROL  x (NEG y)) => (ROR  x y)
   751(ROLW x (NEG y)) => (RORW x y)
   752
   753// Convert const subtraction into ADDI with negative immediate, where possible.
   754(SUB x (MOVDconst [val])) && is32Bit(-val) => (ADDI [-val] x)
   755(SUB <t> (MOVDconst [val]) y) && is32Bit(-val) => (NEG (ADDI <t> [-val] y))
   756
   757// Subtraction of zero.
   758(SUB  x (MOVDconst [0])) => x
   759(SUBW x (MOVDconst [0])) => (ADDIW [0] x)
   760
   761// Subtraction from zero.
   762(SUB  (MOVDconst [0]) x) => (NEG x)
   763(SUBW (MOVDconst [0]) x) => (NEGW x)
   764
   765// Fold negation into subtraction.
   766(NEG (SUB x y)) => (SUB y x)
   767(NEG <t> s:(ADDI [val] (SUB x y))) && s.Uses == 1 && is32Bit(-val) => (ADDI [-val] (SUB <t> y x))
   768
   769// Double negation.
   770(NEG (NEG x)) => x
   771
   772// Addition of zero or two constants.
   773(ADDI [0] x) => x
   774(ADDI [x] (MOVDconst [y])) && is32Bit(x + y) => (MOVDconst [x + y])
   775
   776// ANDI with all zeros, all ones or two constants.
   777(ANDI [0]  x) => (MOVDconst [0])
   778(ANDI [-1] x) => x
   779(ANDI [x] (MOVDconst [y])) => (MOVDconst [x & y])
   780
   781// ORI with all zeroes, all ones or two constants.
   782(ORI [0]  x) => x
   783(ORI [-1] x) => (MOVDconst [-1])
   784(ORI [x] (MOVDconst [y])) => (MOVDconst [x | y])
   785
   786// Combine operations with immediate.
   787(ADDI [x] (ADDI [y] z)) && is32Bit(x + y) => (ADDI [x + y] z)
   788(ANDI [x] (ANDI [y] z)) => (ANDI [x & y] z)
   789(ORI  [x] (ORI  [y] z)) => (ORI  [x | y] z)
   790
   791// Negation of a constant.
   792(NEG  (MOVDconst [x])) => (MOVDconst [-x])
   793(NEGW (MOVDconst [x])) => (MOVDconst [int64(int32(-x))])
   794
   795// Shift of a constant.
   796(SLLI [x] (MOVDconst [y])) && is32Bit(y << uint32(x)) => (MOVDconst [y << uint32(x)])
   797(SRLI [x] (MOVDconst [y])) => (MOVDconst [int64(uint64(y) >> uint32(x))])
   798(SRAI [x] (MOVDconst [y])) => (MOVDconst [int64(y) >> uint32(x)])
   799
   800// SLTI/SLTIU with constants.
   801(SLTI  [x] (MOVDconst [y])) => (MOVDconst [b2i(int64(y) < int64(x))])
   802(SLTIU [x] (MOVDconst [y])) => (MOVDconst [b2i(uint64(y) < uint64(x))])
   803
   804// SLTI/SLTIU with known outcomes.
   805(SLTI  [x] (ANDI [y] _)) && y >= 0 && int64(y) < int64(x) => (MOVDconst [1])
   806(SLTIU [x] (ANDI [y] _)) && y >= 0 && uint64(y) < uint64(x) => (MOVDconst [1])
   807(SLTI  [x] (ORI  [y] _)) && y >= 0 && int64(y) >= int64(x) => (MOVDconst [0])
   808(SLTIU [x] (ORI  [y] _)) && y >= 0 && uint64(y) >= uint64(x) => (MOVDconst [0])
   809
   810// SLT/SLTU with known outcomes.
   811(SLT  x x) => (MOVDconst [0])
   812(SLTU x x) => (MOVDconst [0])
   813
   814// Deadcode for LoweredMuluhilo
   815(Select0 m:(LoweredMuluhilo x y)) && m.Uses == 1 => (MULHU x y)
   816(Select1 m:(LoweredMuluhilo x y)) && m.Uses == 1 => (MUL x y)
   817
   818(FADD(S|D) a (FMUL(S|D) x y)) && a.Block.Func.useFMA(v) => (FMADD(S|D) x y a)
   819(FSUB(S|D) a (FMUL(S|D) x y)) && a.Block.Func.useFMA(v) => (FNMSUB(S|D) x y a)
   820(FSUB(S|D) (FMUL(S|D) x y) a) && a.Block.Func.useFMA(v) => (FMSUB(S|D) x y a)
   821
   822// Merge negation into fused multiply-add and multiply-subtract.
   823//
   824// Key:
   825//
   826//   [+ -](x * y [+ -] z).
   827//    _ N         A S
   828//                D U
   829//                D B
   830//
   831// Note: multiplication commutativity handled by rule generator.
   832(F(MADD|NMADD|MSUB|NMSUB)S neg:(FNEGS x) y z) && neg.Uses == 1 => (F(NMSUB|MSUB|NMADD|MADD)S x y z)
   833(F(MADD|NMADD|MSUB|NMSUB)S x y neg:(FNEGS z)) && neg.Uses == 1 => (F(MSUB|NMSUB|MADD|NMADD)S x y z)
   834(F(MADD|NMADD|MSUB|NMSUB)D neg:(FNEGD x) y z) && neg.Uses == 1 => (F(NMSUB|MSUB|NMADD|MADD)D x y z)
   835(F(MADD|NMADD|MSUB|NMSUB)D x y neg:(FNEGD z)) && neg.Uses == 1 => (F(MSUB|NMSUB|MADD|NMADD)D x y z)
   836
   837//
   838// Optimisations for rva22u64 and above.
   839//
   840
   841// Combine left shift and addition.
   842(ADD (SLLI [1] x) y) && buildcfg.GORISCV64 >= 22 => (SH1ADD x y)
   843(ADD (SLLI [2] x) y) && buildcfg.GORISCV64 >= 22 => (SH2ADD x y)
   844(ADD (SLLI [3] x) y) && buildcfg.GORISCV64 >= 22 => (SH3ADD x y)
   845
   846// Integer minimum and maximum.
   847(Min64  x y) && buildcfg.GORISCV64 >= 22 => (MIN  x y)
   848(Max64  x y) && buildcfg.GORISCV64 >= 22 => (MAX  x y)
   849(Min64u x y) && buildcfg.GORISCV64 >= 22 => (MINU x y)
   850(Max64u x y) && buildcfg.GORISCV64 >= 22 => (MAXU x y)

View as plain text