...

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(Mul(32|16|8) ...) => (MULW ...)
    16(Mul(64|32)F ...) => (FMUL(D|S) ...)
    17
    18(Div(64|32)F ...) => (FDIV(D|S) ...)
    19
    20(Div64 x y [false])  => (DIV x y)
    21(Div64u ...) => (DIVU ...)
    22(Div32 x y [false])  => (DIVW x y)
    23(Div32u ...) => (DIVUW ...)
    24(Div16 x y [false])  => (DIVW  (SignExt16to32 x) (SignExt16to32 y))
    25(Div16u x y) => (DIVUW (ZeroExt16to32 x) (ZeroExt16to32 y))
    26(Div8 x y)   => (DIVW  (SignExt8to32 x)  (SignExt8to32 y))
    27(Div8u x y)  => (DIVUW (ZeroExt8to32 x)  (ZeroExt8to32 y))
    28
    29(Hmul64 ...)  => (MULH  ...)
    30(Hmul64u ...) => (MULHU ...)
    31(Hmul32 x y)  => (SRAI [32] (MUL  (SignExt32to64 x) (SignExt32to64 y)))
    32(Hmul32u x y) => (SRLI [32] (MUL  (ZeroExt32to64 x) (ZeroExt32to64 y)))
    33
    34(Select0 (Add64carry x y c)) => (ADD (ADD <typ.UInt64> x y) c)
    35(Select1 (Add64carry x y c)) =>
    36	(OR (SLTU <typ.UInt64> s:(ADD <typ.UInt64> x y) x) (SLTU <typ.UInt64> (ADD <typ.UInt64> s c) s))
    37
    38(Select0 (Sub64borrow x y c)) => (SUB (SUB <typ.UInt64> x y) c)
    39(Select1 (Sub64borrow x y c)) =>
    40	(OR (SLTU <typ.UInt64> x s:(SUB <typ.UInt64> x y)) (SLTU <typ.UInt64> s (SUB <typ.UInt64> s c)))
    41
    42// (x + y) / 2 => (x / 2) + (y / 2) + (x & y & 1)
    43(Avg64u <t> x y) => (ADD (ADD <t> (SRLI <t> [1] x) (SRLI <t> [1] y)) (ANDI <t> [1] (AND <t> x y)))
    44
    45(Mod64 x y [false])  => (REM x y)
    46(Mod64u ...) => (REMU  ...)
    47(Mod32 x y [false])  => (REMW x y)
    48(Mod32u ...) => (REMUW ...)
    49(Mod16 x y [false])  => (REMW  (SignExt16to32 x) (SignExt16to32 y))
    50(Mod16u x y) => (REMUW (ZeroExt16to32 x) (ZeroExt16to32 y))
    51(Mod8 x y)   => (REMW  (SignExt8to32  x) (SignExt8to32  y))
    52(Mod8u x y)  => (REMUW (ZeroExt8to32  x) (ZeroExt8to32  y))
    53
    54(And(64|32|16|8) ...) => (AND ...)
    55(Or(64|32|16|8) ...) => (OR ...)
    56(Xor(64|32|16|8) ...) => (XOR ...)
    57
    58(Neg(64|32|16|8) ...) => (NEG ...)
    59(Neg(64|32)F ...) => (FNEG(D|S) ...)
    60
    61(Com(64|32|16|8) ...) => (NOT ...)
    62
    63(Sqrt ...) => (FSQRTD ...)
    64(Sqrt32 ...) => (FSQRTS ...)
    65
    66(Copysign ...) => (FSGNJD ...)
    67
    68(Abs ...) => (FABSD ...)
    69
    70(FMA ...) => (FMADDD ...)
    71
    72(Min(64|32)F ...) => (LoweredFMIN(D|S) ...)
    73(Max(64|32)F ...) => (LoweredFMAX(D|S) ...)
    74
    75// Sign and zero extension.
    76
    77(SignExt8to16  ...) => (MOVBreg ...)
    78(SignExt8to32  ...) => (MOVBreg ...)
    79(SignExt8to64  ...) => (MOVBreg ...)
    80(SignExt16to32 ...) => (MOVHreg ...)
    81(SignExt16to64 ...) => (MOVHreg ...)
    82(SignExt32to64 ...) => (MOVWreg ...)
    83
    84(ZeroExt8to16  ...) => (MOVBUreg ...)
    85(ZeroExt8to32  ...) => (MOVBUreg ...)
    86(ZeroExt8to64  ...) => (MOVBUreg ...)
    87(ZeroExt16to32 ...) => (MOVHUreg ...)
    88(ZeroExt16to64 ...) => (MOVHUreg ...)
    89(ZeroExt32to64 ...) => (MOVWUreg ...)
    90
    91(Cvt32to32F ...) => (FCVTSW ...)
    92(Cvt32to64F ...) => (FCVTDW ...)
    93(Cvt64to32F ...) => (FCVTSL ...)
    94(Cvt64to64F ...) => (FCVTDL ...)
    95
    96(Cvt32Fto32 ...) => (FCVTWS ...)
    97(Cvt32Fto64 ...) => (FCVTLS ...)
    98(Cvt64Fto32 ...) => (FCVTWD ...)
    99(Cvt64Fto64 ...) => (FCVTLD ...)
   100
   101(Cvt32Fto64F ...) => (FCVTDS ...)
   102(Cvt64Fto32F ...) => (FCVTSD ...)
   103
   104(CvtBoolToUint8 ...) => (Copy ...)
   105
   106(Round(32|64)F ...) => (LoweredRound(32|64)F ...)
   107
   108(Slicemask <t> x) => (SRAI [63] (NEG <t> x))
   109
   110// Truncations
   111// We ignore the unused high parts of registers, so truncates are just copies.
   112(Trunc16to8  ...) => (Copy ...)
   113(Trunc32to8  ...) => (Copy ...)
   114(Trunc32to16 ...) => (Copy ...)
   115(Trunc64to8  ...) => (Copy ...)
   116(Trunc64to16 ...) => (Copy ...)
   117(Trunc64to32 ...) => (Copy ...)
   118
   119// Shifts
   120
   121// SLL only considers the bottom 6 bits of y. If y > 64, the result should
   122// always be 0.
   123//
   124// Breaking down the operation:
   125//
   126// (SLL x y) generates x << (y & 63).
   127//
   128// If y < 64, this is the value we want. Otherwise, we want zero.
   129//
   130// So, we AND with -1 * uint64(y < 64), which is 0xfffff... if y < 64 and 0 otherwise.
   131(Lsh8x8   <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg8  <t> (SLTIU <t> [64] (ZeroExt8to64  y))))
   132(Lsh8x16  <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg8  <t> (SLTIU <t> [64] (ZeroExt16to64 y))))
   133(Lsh8x32  <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg8  <t> (SLTIU <t> [64] (ZeroExt32to64 y))))
   134(Lsh8x64  <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg8  <t> (SLTIU <t> [64] y)))
   135(Lsh16x8  <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg16 <t> (SLTIU <t> [64] (ZeroExt8to64  y))))
   136(Lsh16x16 <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg16 <t> (SLTIU <t> [64] (ZeroExt16to64 y))))
   137(Lsh16x32 <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg16 <t> (SLTIU <t> [64] (ZeroExt32to64 y))))
   138(Lsh16x64 <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg16 <t> (SLTIU <t> [64] y)))
   139(Lsh32x8  <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg32 <t> (SLTIU <t> [64] (ZeroExt8to64  y))))
   140(Lsh32x16 <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg32 <t> (SLTIU <t> [64] (ZeroExt16to64 y))))
   141(Lsh32x32 <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg32 <t> (SLTIU <t> [64] (ZeroExt32to64 y))))
   142(Lsh32x64 <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg32 <t> (SLTIU <t> [64] y)))
   143(Lsh64x8  <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg64 <t> (SLTIU <t> [64] (ZeroExt8to64  y))))
   144(Lsh64x16 <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg64 <t> (SLTIU <t> [64] (ZeroExt16to64 y))))
   145(Lsh64x32 <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg64 <t> (SLTIU <t> [64] (ZeroExt32to64 y))))
   146(Lsh64x64 <t> x y) && !shiftIsBounded(v) => (AND (SLL <t> x y) (Neg64 <t> (SLTIU <t> [64] y)))
   147
   148(Lsh8x(64|32|16|8)  x y) && shiftIsBounded(v) => (SLL x y)
   149(Lsh16x(64|32|16|8) x y) && shiftIsBounded(v) => (SLL x y)
   150(Lsh32x(64|32|16|8) x y) && shiftIsBounded(v) => (SLL x y)
   151(Lsh64x(64|32|16|8) x y) && shiftIsBounded(v) => (SLL x y)
   152
   153// SRL only considers the bottom 6 bits of y, similarly SRLW only considers the
   154// bottom 5 bits of y. Ensure that the result is always zero if the shift exceeds
   155// the maximum value. See Lsh above for a detailed description.
   156(Rsh8Ux8   <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t> (ZeroExt8to64  x) y) (Neg8  <t> (SLTIU <t> [64] (ZeroExt8to64  y))))
   157(Rsh8Ux16  <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t> (ZeroExt8to64  x) y) (Neg8  <t> (SLTIU <t> [64] (ZeroExt16to64 y))))
   158(Rsh8Ux32  <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t> (ZeroExt8to64  x) y) (Neg8  <t> (SLTIU <t> [64] (ZeroExt32to64 y))))
   159(Rsh8Ux64  <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t> (ZeroExt8to64  x) y) (Neg8  <t> (SLTIU <t> [64] y)))
   160(Rsh16Ux8  <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t> (ZeroExt16to64 x) y) (Neg16 <t> (SLTIU <t> [64] (ZeroExt8to64  y))))
   161(Rsh16Ux16 <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t> (ZeroExt16to64 x) y) (Neg16 <t> (SLTIU <t> [64] (ZeroExt16to64 y))))
   162(Rsh16Ux32 <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t> (ZeroExt16to64 x) y) (Neg16 <t> (SLTIU <t> [64] (ZeroExt32to64 y))))
   163(Rsh16Ux64 <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t> (ZeroExt16to64 x) y) (Neg16 <t> (SLTIU <t> [64] y)))
   164(Rsh32Ux8  <t> x y) && !shiftIsBounded(v) => (AND (SRLW <t>  x                y) (Neg32 <t> (SLTIU <t> [32] (ZeroExt8to64  y))))
   165(Rsh32Ux16 <t> x y) && !shiftIsBounded(v) => (AND (SRLW <t>  x                y) (Neg32 <t> (SLTIU <t> [32] (ZeroExt16to64 y))))
   166(Rsh32Ux32 <t> x y) && !shiftIsBounded(v) => (AND (SRLW <t>  x                y) (Neg32 <t> (SLTIU <t> [32] (ZeroExt32to64 y))))
   167(Rsh32Ux64 <t> x y) && !shiftIsBounded(v) => (AND (SRLW <t>  x                y) (Neg32 <t> (SLTIU <t> [32] y)))
   168(Rsh64Ux8  <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t>  x                y) (Neg64 <t> (SLTIU <t> [64] (ZeroExt8to64  y))))
   169(Rsh64Ux16 <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t>  x                y) (Neg64 <t> (SLTIU <t> [64] (ZeroExt16to64 y))))
   170(Rsh64Ux32 <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t>  x                y) (Neg64 <t> (SLTIU <t> [64] (ZeroExt32to64 y))))
   171(Rsh64Ux64 <t> x y) && !shiftIsBounded(v) => (AND (SRL  <t>  x                y) (Neg64 <t> (SLTIU <t> [64] y)))
   172
   173(Rsh8Ux(64|32|16|8)  x y) && shiftIsBounded(v) => (SRL  (ZeroExt8to64  x) y)
   174(Rsh16Ux(64|32|16|8) x y) && shiftIsBounded(v) => (SRL  (ZeroExt16to64 x) y)
   175(Rsh32Ux(64|32|16|8) x y) && shiftIsBounded(v) => (SRLW x                 y)
   176(Rsh64Ux(64|32|16|8) x y) && shiftIsBounded(v) => (SRL  x                 y)
   177
   178// SRA only considers the bottom 6 bits of y, similarly SRAW only considers the
   179// bottom 5 bits. If y is greater than the maximum value (either 63 or 31
   180// depending on the instruction),  the result of the shift should be either 0
   181// or -1 based on the sign bit of x.
   182//
   183// We implement this by performing the max shift (-1) if y > the maximum value.
   184//
   185// We OR (uint64(y < 64) - 1) into y before passing it to SRA. This leaves
   186// us with -1 (0xffff...) if y >= 64.  Similarly, we OR (uint64(y < 32) - 1) into y
   187// before passing it to SRAW.
   188//
   189// We don't need to sign-extend the OR result, as it will be at minimum 8 bits,
   190// more than the 5 or 6 bits SRAW and SRA care about.
   191(Rsh8x8   <t> x y) && !shiftIsBounded(v) => (SRA  <t> (SignExt8to64  x) (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] (ZeroExt8to64  y)))))
   192(Rsh8x16  <t> x y) && !shiftIsBounded(v) => (SRA  <t> (SignExt8to64  x) (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] (ZeroExt16to64 y)))))
   193(Rsh8x32  <t> x y) && !shiftIsBounded(v) => (SRA  <t> (SignExt8to64  x) (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] (ZeroExt32to64 y)))))
   194(Rsh8x64  <t> x y) && !shiftIsBounded(v) => (SRA  <t> (SignExt8to64  x) (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] y))))
   195(Rsh16x8  <t> x y) && !shiftIsBounded(v) => (SRA  <t> (SignExt16to64 x) (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] (ZeroExt8to64  y)))))
   196(Rsh16x16 <t> x y) && !shiftIsBounded(v) => (SRA  <t> (SignExt16to64 x) (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] (ZeroExt16to64 y)))))
   197(Rsh16x32 <t> x y) && !shiftIsBounded(v) => (SRA  <t> (SignExt16to64 x) (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] (ZeroExt32to64 y)))))
   198(Rsh16x64 <t> x y) && !shiftIsBounded(v) => (SRA  <t> (SignExt16to64 x) (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] y))))
   199(Rsh32x8  <t> x y) && !shiftIsBounded(v) => (SRAW <t> x                 (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [32] (ZeroExt8to64  y)))))
   200(Rsh32x16 <t> x y) && !shiftIsBounded(v) => (SRAW <t> x                 (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [32] (ZeroExt16to64 y)))))
   201(Rsh32x32 <t> x y) && !shiftIsBounded(v) => (SRAW <t> x                 (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [32] (ZeroExt32to64 y)))))
   202(Rsh32x64 <t> x y) && !shiftIsBounded(v) => (SRAW <t> x                 (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [32] y))))
   203(Rsh64x8  <t> x y) && !shiftIsBounded(v) => (SRA  <t> x                 (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] (ZeroExt8to64  y)))))
   204(Rsh64x16 <t> x y) && !shiftIsBounded(v) => (SRA  <t> x                 (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] (ZeroExt16to64 y)))))
   205(Rsh64x32 <t> x y) && !shiftIsBounded(v) => (SRA  <t> x                 (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] (ZeroExt32to64 y)))))
   206(Rsh64x64 <t> x y) && !shiftIsBounded(v) => (SRA  <t> x                 (OR <y.Type> y (ADDI <y.Type> [-1] (SLTIU <y.Type> [64] y))))
   207
   208(Rsh8x(64|32|16|8)  x y) && shiftIsBounded(v) => (SRA  (SignExt8to64  x) y)
   209(Rsh16x(64|32|16|8) x y) && shiftIsBounded(v) => (SRA  (SignExt16to64 x) y)
   210(Rsh32x(64|32|16|8) x y) && shiftIsBounded(v) => (SRAW  x                y)
   211(Rsh64x(64|32|16|8) x y) && shiftIsBounded(v) => (SRA   x                y)
   212
   213// Rotates.
   214(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))))
   215(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))))
   216(RotateLeft32 ...) => (ROLW ...)
   217(RotateLeft64 ...) => (ROL  ...)
   218
   219// Count trailing zeros (note that these will only be emitted for rva22u64 and above).
   220(Ctz(64|32|16|8)NonZero ...) => (Ctz64 ...)
   221(Ctz64 ...) => (CTZ  ...)
   222(Ctz32 ...) => (CTZW ...)
   223(Ctz16 x) => (CTZW (ORI <typ.UInt32> [1<<16] x))
   224(Ctz8  x) => (CTZW (ORI <typ.UInt32> [1<<8]  x))
   225
   226// Bit length (note that these will only be emitted for rva22u64 and above).
   227(BitLen64 <t> x) => (SUB (MOVDconst [64]) (CLZ  <t> x))
   228(BitLen32 <t> x) => (SUB (MOVDconst [32]) (CLZW <t> x))
   229(BitLen16 x) => (BitLen64 (ZeroExt16to64 x))
   230(BitLen8  x) => (BitLen64 (ZeroExt8to64 x))
   231
   232// Byte swap (note that these will only be emitted for rva22u64 and above).
   233(Bswap64 ...) => (REV8 ...)
   234(Bswap32 <t> x) => (SRLI [32] (REV8 <t> x))
   235(Bswap16 <t> x) => (SRLI [48] (REV8 <t> x))
   236
   237// Population count (note that these will be emitted with guards for rva20u64).
   238(PopCount64 ...) => (CPOP  ...)
   239(PopCount32 ...) => (CPOPW ...)
   240(PopCount16 x) => (CPOP (ZeroExt16to64 x))
   241(PopCount8  x) => (CPOP (ZeroExt8to64  x))
   242
   243(Less64  ...) => (SLT  ...)
   244(Less32  x y) => (SLT  (SignExt32to64 x) (SignExt32to64 y))
   245(Less16  x y) => (SLT  (SignExt16to64 x) (SignExt16to64 y))
   246(Less8   x y) => (SLT  (SignExt8to64  x) (SignExt8to64  y))
   247(Less64U ...) => (SLTU ...)
   248(Less32U x y) => (SLTU (ZeroExt32to64 x) (ZeroExt32to64 y))
   249(Less16U x y) => (SLTU (ZeroExt16to64 x) (ZeroExt16to64 y))
   250(Less8U  x y) => (SLTU (ZeroExt8to64  x) (ZeroExt8to64  y))
   251(Less(64|32)F ...) => (FLT(D|S) ...)
   252
   253// Convert x <= y to !(y > x).
   254(Leq(64|32|16|8)  x y) => (Not (Less(64|32|16|8)  y x))
   255(Leq(64|32|16|8)U x y) => (Not (Less(64|32|16|8)U y x))
   256(Leq(64|32)F ...) => (FLE(D|S) ...)
   257
   258(EqPtr x y) => (SEQZ (SUB <typ.Uintptr> x y))
   259(Eq64  x y) => (SEQZ (SUB <x.Type> x y))
   260(Eq32  x y) &&  x.Type.IsSigned() => (SEQZ (SUB <x.Type> (SignExt32to64 x) (SignExt32to64 y)))
   261(Eq32  x y) && !x.Type.IsSigned() => (SEQZ (SUB <x.Type> (ZeroExt32to64 x) (ZeroExt32to64 y)))
   262(Eq16  x y) => (SEQZ (SUB <x.Type> (ZeroExt16to64 x) (ZeroExt16to64 y)))
   263(Eq8   x y) => (SEQZ (SUB <x.Type> (ZeroExt8to64  x) (ZeroExt8to64  y)))
   264(Eq(64|32)F ...) => (FEQ(D|S) ...)
   265
   266(NeqPtr x y) => (Not (EqPtr x y))
   267(Neq64  x y) => (Not (Eq64  x y))
   268(Neq32  x y) => (Not (Eq32  x y))
   269(Neq16  x y) => (Not (Eq16  x y))
   270(Neq8   x y) => (Not (Eq8   x y))
   271(Neq(64|32)F ...) => (FNE(D|S) ...)
   272
   273// Loads
   274(Load <t> ptr mem) &&  t.IsBoolean()                   => (MOVBUload ptr mem)
   275(Load <t> ptr mem) && ( is8BitInt(t) &&  t.IsSigned()) => (MOVBload  ptr mem)
   276(Load <t> ptr mem) && ( is8BitInt(t) && !t.IsSigned()) => (MOVBUload ptr mem)
   277(Load <t> ptr mem) && (is16BitInt(t) &&  t.IsSigned()) => (MOVHload  ptr mem)
   278(Load <t> ptr mem) && (is16BitInt(t) && !t.IsSigned()) => (MOVHUload ptr mem)
   279(Load <t> ptr mem) && (is32BitInt(t) &&  t.IsSigned()) => (MOVWload  ptr mem)
   280(Load <t> ptr mem) && (is32BitInt(t) && !t.IsSigned()) => (MOVWUload ptr mem)
   281(Load <t> ptr mem) && (is64BitInt(t) || isPtr(t))      => (MOVDload  ptr mem)
   282(Load <t> ptr mem) &&  is32BitFloat(t)                 => (FMOVWload ptr mem)
   283(Load <t> ptr mem) &&  is64BitFloat(t)                 => (FMOVDload ptr mem)
   284
   285// Stores
   286(Store {t} ptr val mem) && t.Size() == 1 => (MOVBstore ptr val mem)
   287(Store {t} ptr val mem) && t.Size() == 2 => (MOVHstore ptr val mem)
   288(Store {t} ptr val mem) && t.Size() == 4 && !t.IsFloat() => (MOVWstore ptr val mem)
   289(Store {t} ptr val mem) && t.Size() == 8 && !t.IsFloat() => (MOVDstore ptr val mem)
   290(Store {t} ptr val mem) && t.Size() == 4 &&  t.IsFloat() => (FMOVWstore ptr val mem)
   291(Store {t} ptr val mem) && t.Size() == 8 &&  t.IsFloat() => (FMOVDstore ptr val mem)
   292
   293// We need to fold MOVaddr into the LD/MOVDstore ops so that the live variable analysis
   294// knows what variables are being read/written by the ops.
   295(MOV(B|BU|H|HU|W|WU|D)load [off1] {sym1} (MOVaddr [off2] {sym2} base) mem) &&
   296	is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) &&
   297	(base.Op != OpSB || !config.ctxt.Flag_dynlink) =>
   298	(MOV(B|BU|H|HU|W|WU|D)load [off1+off2] {mergeSym(sym1,sym2)} base mem)
   299
   300(FMOV(W|D)load [off1] {sym1} (MOVaddr [off2] {sym2} base) mem) &&
   301	is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) &&
   302	(base.Op != OpSB || !config.ctxt.Flag_dynlink) =>
   303	(FMOV(W|D)load [off1+off2] {mergeSym(sym1,sym2)} base mem)
   304
   305(MOV(B|H|W|D)store [off1] {sym1} (MOVaddr [off2] {sym2} base) val mem) &&
   306	is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) &&
   307	(base.Op != OpSB || !config.ctxt.Flag_dynlink) =>
   308	(MOV(B|H|W|D)store [off1+off2] {mergeSym(sym1,sym2)} base val mem)
   309
   310(MOV(B|H|W|D)storezero [off1] {sym1} (MOVaddr [off2] {sym2} base) mem) &&
   311	canMergeSym(sym1,sym2) && is32Bit(int64(off1)+int64(off2)) &&
   312	(base.Op != OpSB || !config.ctxt.Flag_dynlink) =>
   313	(MOV(B|H|W|D)storezero [off1+off2] {mergeSym(sym1,sym2)} base mem)
   314
   315(FMOV(W|D)store [off1] {sym1} (MOVaddr [off2] {sym2} base) val mem) &&
   316	is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) &&
   317	(base.Op != OpSB || !config.ctxt.Flag_dynlink) =>
   318	(FMOV(W|D)store [off1+off2] {mergeSym(sym1,sym2)} base val mem)
   319
   320(MOV(B|BU|H|HU|W|WU|D)load [off1] {sym} (ADDI [off2] base) mem) && is32Bit(int64(off1)+off2) =>
   321	(MOV(B|BU|H|HU|W|WU|D)load [off1+int32(off2)] {sym} base mem)
   322
   323(FMOV(W|D)load [off1] {sym} (ADDI [off2] base) mem) && is32Bit(int64(off1)+off2) =>
   324	(FMOV(W|D)load [off1+int32(off2)] {sym} base mem)
   325
   326(MOV(B|H|W|D)store [off1] {sym} (ADDI [off2] base) val mem) && is32Bit(int64(off1)+off2) =>
   327	(MOV(B|H|W|D)store [off1+int32(off2)] {sym} base val mem)
   328
   329(MOV(B|H|W|D)storezero [off1] {sym} (ADDI [off2] base) mem) && is32Bit(int64(off1)+off2) =>
   330	(MOV(B|H|W|D)storezero [off1+int32(off2)] {sym} base mem)
   331
   332(FMOV(W|D)store [off1] {sym} (ADDI [off2] base) val mem) && is32Bit(int64(off1)+off2) =>
   333	(FMOV(W|D)store [off1+int32(off2)] {sym} base val mem)
   334
   335// Similarly, fold ADDI into MOVaddr to avoid confusing live variable analysis
   336// with OffPtr -> ADDI.
   337(ADDI [c] (MOVaddr [d] {s} x)) && is32Bit(c+int64(d)) => (MOVaddr [int32(c)+d] {s} x)
   338
   339// Small zeroing
   340(Zero [0] _ mem) => mem
   341(Zero [1] ptr mem) => (MOVBstore ptr (MOVDconst [0]) mem)
   342(Zero [2] {t} ptr mem) && t.Alignment()%2 == 0 =>
   343	(MOVHstore ptr (MOVDconst [0]) mem)
   344(Zero [2] ptr mem) =>
   345	(MOVBstore [1] ptr (MOVDconst [0])
   346		(MOVBstore ptr (MOVDconst [0]) mem))
   347(Zero [4] {t} ptr mem) && t.Alignment()%4 == 0 =>
   348	(MOVWstore ptr (MOVDconst [0]) mem)
   349(Zero [4] {t} ptr mem) && t.Alignment()%2 == 0 =>
   350	(MOVHstore [2] ptr (MOVDconst [0])
   351		(MOVHstore ptr (MOVDconst [0]) mem))
   352(Zero [4] ptr mem) =>
   353	(MOVBstore [3] ptr (MOVDconst [0])
   354		(MOVBstore [2] ptr (MOVDconst [0])
   355			(MOVBstore [1] ptr (MOVDconst [0])
   356				(MOVBstore ptr (MOVDconst [0]) mem))))
   357(Zero [8] {t} ptr mem) && t.Alignment()%8 == 0 =>
   358	(MOVDstore ptr (MOVDconst [0]) mem)
   359(Zero [8] {t} ptr mem) && t.Alignment()%4 == 0 =>
   360	(MOVWstore [4] ptr (MOVDconst [0])
   361		(MOVWstore ptr (MOVDconst [0]) mem))
   362(Zero [8] {t} ptr mem) && t.Alignment()%2 == 0 =>
   363	(MOVHstore [6] ptr (MOVDconst [0])
   364		(MOVHstore [4] ptr (MOVDconst [0])
   365			(MOVHstore [2] ptr (MOVDconst [0])
   366				(MOVHstore ptr (MOVDconst [0]) mem))))
   367
   368(Zero [3] ptr mem) =>
   369	(MOVBstore [2] ptr (MOVDconst [0])
   370		(MOVBstore [1] ptr (MOVDconst [0])
   371			(MOVBstore ptr (MOVDconst [0]) mem)))
   372(Zero [6] {t} ptr mem) && t.Alignment()%2 == 0 =>
   373	(MOVHstore [4] ptr (MOVDconst [0])
   374		(MOVHstore [2] ptr (MOVDconst [0])
   375			(MOVHstore ptr (MOVDconst [0]) mem)))
   376
   377// Unroll zeroing in medium size (at most 192 bytes i.e. 3 cachelines)
   378(Zero [s] {t} ptr mem) && s <= 24*moveSize(t.Alignment(), config) =>
   379	(LoweredZero [makeValAndOff(int32(s),int32(t.Alignment()))] ptr mem)
   380
   381// Generic zeroing uses a loop
   382(Zero [s] {t} ptr mem) && s > 24*moveSize(t.Alignment(), config) =>
   383	(LoweredZeroLoop [makeValAndOff(int32(s),int32(t.Alignment()))] ptr mem)
   384
   385// Checks
   386(IsNonNil ...) => (SNEZ ...)
   387(IsInBounds ...) => (Less64U ...)
   388(IsSliceInBounds ...) => (Leq64U ...)
   389
   390// Trivial lowering
   391(NilCheck ...) => (LoweredNilCheck ...)
   392(GetClosurePtr ...) => (LoweredGetClosurePtr ...)
   393(GetCallerSP ...) => (LoweredGetCallerSP ...)
   394(GetCallerPC ...) => (LoweredGetCallerPC ...)
   395
   396// Write barrier.
   397(WB ...) => (LoweredWB ...)
   398
   399// Publication barrier as intrinsic
   400(PubBarrier ...) => (LoweredPubBarrier ...)
   401
   402(PanicBounds ...) => (LoweredPanicBoundsRR ...)
   403(LoweredPanicBoundsRR [kind] x (MOVDconst [c]) mem) => (LoweredPanicBoundsRC [kind] x {PanicBoundsC{C:c}} mem)
   404(LoweredPanicBoundsRR [kind] (MOVDconst [c]) y mem) => (LoweredPanicBoundsCR [kind] {PanicBoundsC{C:c}} y mem)
   405(LoweredPanicBoundsRC [kind] {p} (MOVDconst [c]) mem) => (LoweredPanicBoundsCC [kind] {PanicBoundsCC{Cx:c, Cy:p.C}} mem)
   406(LoweredPanicBoundsCR [kind] {p} (MOVDconst [c]) mem) => (LoweredPanicBoundsCC [kind] {PanicBoundsCC{Cx:p.C, Cy:c}} mem)
   407
   408// Small moves
   409(Move [0] _ _ mem) => mem
   410(Move [1] dst src mem) => (MOVBstore dst (MOVBload src mem) mem)
   411(Move [2] {t} dst src mem) && t.Alignment()%2 == 0 =>
   412	(MOVHstore dst (MOVHload src mem) mem)
   413(Move [2] dst src mem) =>
   414	(MOVBstore [1] dst (MOVBload [1] src mem)
   415		(MOVBstore dst (MOVBload src mem) mem))
   416(Move [4] {t} dst src mem) && t.Alignment()%4 == 0 =>
   417	(MOVWstore dst (MOVWload src mem) mem)
   418(Move [4] {t} dst src mem) && t.Alignment()%2 == 0 =>
   419	(MOVHstore [2] dst (MOVHload [2] src mem)
   420		(MOVHstore dst (MOVHload src mem) mem))
   421(Move [4] dst src mem) =>
   422	(MOVBstore [3] dst (MOVBload [3] src mem)
   423		(MOVBstore [2] dst (MOVBload [2] src mem)
   424			(MOVBstore [1] dst (MOVBload [1] src mem)
   425				(MOVBstore dst (MOVBload src mem) mem))))
   426(Move [8] {t} dst src mem) && t.Alignment()%8 == 0 =>
   427	(MOVDstore dst (MOVDload src mem) mem)
   428(Move [8] {t} dst src mem) && t.Alignment()%4 == 0 =>
   429	(MOVWstore [4] dst (MOVWload [4] src mem)
   430		(MOVWstore dst (MOVWload src mem) mem))
   431(Move [8] {t} dst src mem) && t.Alignment()%2 == 0 =>
   432	(MOVHstore [6] dst (MOVHload [6] src mem)
   433		(MOVHstore [4] dst (MOVHload [4] src mem)
   434			(MOVHstore [2] dst (MOVHload [2] src mem)
   435				(MOVHstore dst (MOVHload src mem) mem))))
   436
   437(Move [3] dst src mem) =>
   438	(MOVBstore [2] dst (MOVBload [2] src mem)
   439		(MOVBstore [1] dst (MOVBload [1] src mem)
   440			(MOVBstore dst (MOVBload src mem) mem)))
   441(Move [6] {t} dst src mem) && t.Alignment()%2 == 0 =>
   442	(MOVHstore [4] dst (MOVHload [4] src mem)
   443		(MOVHstore [2] dst (MOVHload [2] src mem)
   444			(MOVHstore dst (MOVHload src mem) mem)))
   445
   446// Generic move
   447(Move [s] {t} dst src mem) && s > 0 && s <= 3*8*moveSize(t.Alignment(), config)
   448	&& logLargeCopy(v, s) =>
   449	(LoweredMove [makeValAndOff(int32(s),int32(t.Alignment()))] dst src mem)
   450
   451// Generic move uses a loop
   452(Move [s] {t} dst src mem) && s > 3*8*moveSize(t.Alignment(), config)
   453	&& logLargeCopy(v, s) =>
   454	(LoweredMoveLoop [makeValAndOff(int32(s),int32(t.Alignment()))] dst src mem)
   455
   456// Boolean ops; 0=false, 1=true
   457(AndB ...) => (AND ...)
   458(OrB  ...) => (OR  ...)
   459(EqB  x y) => (SEQZ (SUB <typ.Bool> x y))
   460(NeqB x y) => (SNEZ (SUB <typ.Bool> x y))
   461(Not  ...) => (SEQZ ...)
   462
   463// Lowering pointer arithmetic
   464// TODO: Special handling for SP offsets, like ARM
   465(OffPtr [off] ptr:(SP)) && is32Bit(off) => (MOVaddr [int32(off)] ptr)
   466(OffPtr [off] ptr) && is32Bit(off) => (ADDI [off] ptr)
   467(OffPtr [off] ptr) => (ADD (MOVDconst [off]) ptr)
   468
   469(Const(64|32|16|8) [val]) => (MOVDconst [int64(val)])
   470(Const(64|32)F ...) => (FMOV(D|F)const ...)
   471(ConstNil) => (MOVDconst [0])
   472(ConstBool [val]) => (MOVDconst [int64(b2i(val))])
   473
   474(Addr {sym} base) => (MOVaddr {sym} [0] base)
   475(LocalAddr <t> {sym} base mem) && t.Elem().HasPointers() => (MOVaddr {sym} (SPanchored base mem))
   476(LocalAddr <t> {sym} base _)  && !t.Elem().HasPointers() => (MOVaddr {sym} base)
   477
   478// Calls
   479(StaticCall  ...) => (CALLstatic  ...)
   480(ClosureCall ...) => (CALLclosure ...)
   481(InterCall   ...) => (CALLinter   ...)
   482(TailCall ...) => (CALLtail ...)
   483
   484// Atomic Intrinsics
   485(AtomicLoad(Ptr|64|32|8)  ...) => (LoweredAtomicLoad(64|64|32|8) ...)
   486(AtomicStore(PtrNoWB|64|32|8) ...) => (LoweredAtomicStore(64|64|32|8) ...)
   487(AtomicAdd(64|32) ...) => (LoweredAtomicAdd(64|32) ...)
   488
   489// AtomicAnd8(ptr,val) => LoweredAtomicAnd32(ptr&^3, ^((uint8(val) ^ 0xff) << ((ptr & 3) * 8)))
   490(AtomicAnd8 ptr val mem) =>
   491	(LoweredAtomicAnd32 (ANDI <typ.Uintptr> [^3] ptr)
   492		(NOT <typ.UInt32> (SLL <typ.UInt32> (XORI <typ.UInt32> [0xff] (ZeroExt8to32 val))
   493			(SLLI <typ.UInt64> [3] (ANDI <typ.UInt64> [3] ptr)))) mem)
   494
   495(AtomicAnd32 ...) => (LoweredAtomicAnd32 ...)
   496
   497(AtomicCompareAndSwap32 ptr old new mem) => (LoweredAtomicCas32 ptr (SignExt32to64 old) new mem)
   498(AtomicCompareAndSwap64 ...) => (LoweredAtomicCas64 ...)
   499
   500(AtomicExchange(64|32) ...) => (LoweredAtomicExchange(64|32) ...)
   501
   502// AtomicOr8(ptr,val)  => LoweredAtomicOr32(ptr&^3, uint32(val)<<((ptr&3)*8))
   503(AtomicOr8 ptr val mem) =>
   504	(LoweredAtomicOr32 (ANDI <typ.Uintptr> [^3] ptr)
   505		(SLL <typ.UInt32> (ZeroExt8to32 val)
   506			(SLLI <typ.UInt64> [3] (ANDI <typ.UInt64> [3] ptr))) mem)
   507
   508(AtomicOr32  ...) => (LoweredAtomicOr32  ...)
   509
   510// Conditional branches
   511(If cond yes no) => (BNEZ (MOVBUreg <typ.UInt64> cond) yes no)
   512
   513// Optimizations
   514
   515// Absorb SEQZ/SNEZ into branch.
   516(BEQZ (SEQZ x) yes no) => (BNEZ x yes no)
   517(BEQZ (SNEZ x) yes no) => (BEQZ x yes no)
   518(BNEZ (SEQZ x) yes no) => (BEQZ x yes no)
   519(BNEZ (SNEZ x) yes no) => (BNEZ x yes no)
   520
   521// Remove redundant NEG from BEQZ/BNEZ.
   522(BEQZ (NEG x) yes no) => (BEQZ x yes no)
   523(BNEZ (NEG x) yes no) => (BNEZ x yes no)
   524
   525// Negate comparison with FNES/FNED.
   526(BEQZ (FNES <t> x y) yes no) => (BNEZ (FEQS <t> x y) yes no)
   527(BNEZ (FNES <t> x y) yes no) => (BEQZ (FEQS <t> x y) yes no)
   528(BEQZ (FNED <t> x y) yes no) => (BNEZ (FEQD <t> x y) yes no)
   529(BNEZ (FNED <t> x y) yes no) => (BEQZ (FEQD <t> x y) yes no)
   530
   531// Convert BEQZ/BNEZ into more optimal branch conditions.
   532(BEQZ (SUB x y) yes no) => (BEQ x y yes no)
   533(BNEZ (SUB x y) yes no) => (BNE x y yes no)
   534(BEQZ (SLT x y) yes no) => (BGE x y yes no)
   535(BNEZ (SLT x y) yes no) => (BLT x y yes no)
   536(BEQZ (SLTU x y) yes no) => (BGEU x y yes no)
   537(BNEZ (SLTU x y) yes no) => (BLTU x y yes no)
   538(BEQZ (SLTI [x] y) yes no) => (BGE y (MOVDconst [x]) yes no)
   539(BNEZ (SLTI [x] y) yes no) => (BLT y (MOVDconst [x]) yes no)
   540(BEQZ (SLTIU [x] y) yes no) => (BGEU y (MOVDconst [x]) yes no)
   541(BNEZ (SLTIU [x] y) yes no) => (BLTU y (MOVDconst [x]) yes no)
   542
   543// Convert branch with zero to more optimal branch zero.
   544(BEQ  (MOVDconst [0]) cond yes no) => (BEQZ cond yes no)
   545(BEQ  cond (MOVDconst [0]) yes no) => (BEQZ cond yes no)
   546(BNE  (MOVDconst [0]) cond yes no) => (BNEZ cond yes no)
   547(BNE  cond (MOVDconst [0]) yes no) => (BNEZ cond yes no)
   548(BLT  (MOVDconst [0]) cond yes no) => (BGTZ cond yes no)
   549(BLT  cond (MOVDconst [0]) yes no) => (BLTZ cond yes no)
   550(BLTU (MOVDconst [0]) cond yes no) => (BNEZ cond yes no)
   551(BGE  (MOVDconst [0]) cond yes no) => (BLEZ cond yes no)
   552(BGE  cond (MOVDconst [0]) yes no) => (BGEZ cond yes no)
   553(BGEU (MOVDconst [0]) cond yes no) => (BEQZ cond yes no)
   554
   555// Remove redundant NEG from SEQZ/SNEZ.
   556(SEQZ (NEG x)) => (SEQZ x)
   557(SNEZ (NEG x)) => (SNEZ x)
   558
   559// Remove redundant SEQZ/SNEZ.
   560(SEQZ (SEQZ x)) => (SNEZ x)
   561(SEQZ (SNEZ x)) => (SEQZ x)
   562(SNEZ (SEQZ x)) => (SEQZ x)
   563(SNEZ (SNEZ x)) => (SNEZ x)
   564
   565// Store zero.
   566(MOVBstore [off] {sym} ptr (MOVDconst [0]) mem) => (MOVBstorezero [off] {sym} ptr mem)
   567(MOVHstore [off] {sym} ptr (MOVDconst [0]) mem) => (MOVHstorezero [off] {sym} ptr mem)
   568(MOVWstore [off] {sym} ptr (MOVDconst [0]) mem) => (MOVWstorezero [off] {sym} ptr mem)
   569(MOVDstore [off] {sym} ptr (MOVDconst [0]) mem) => (MOVDstorezero [off] {sym} ptr mem)
   570
   571// Boolean ops are already extended.
   572(MOVBUreg x:((FLES|FLTS|FEQS|FNES) _ _)) => x
   573(MOVBUreg x:((FLED|FLTD|FEQD|FNED) _ _)) => x
   574(MOVBUreg x:((SEQZ|SNEZ) _)) => x
   575(MOVBUreg x:((SLT|SLTU) _ _)) => x
   576
   577// Avoid extending when already sufficiently masked.
   578(MOVBreg  x:(ANDI [c] y)) && c >= 0 && int64(int8(c)) == c => x
   579(MOVHreg  x:(ANDI [c] y)) && c >= 0 && int64(int16(c)) == c => x
   580(MOVWreg  x:(ANDI [c] y)) && c >= 0 && int64(int32(c)) == c => x
   581(MOVBUreg x:(ANDI [c] y)) && c >= 0 && int64(uint8(c)) == c => x
   582(MOVHUreg x:(ANDI [c] y)) && c >= 0 && int64(uint16(c)) == c => x
   583(MOVWUreg x:(ANDI [c] y)) && c >= 0 && int64(uint32(c)) == c => x
   584
   585// Combine masking and zero extension.
   586(MOVBUreg (ANDI [c] x)) && c < 0 => (ANDI [int64(uint8(c))] x)
   587(MOVHUreg (ANDI [c] x)) && c < 0 => (ANDI [int64(uint16(c))] x)
   588(MOVWUreg (ANDI [c] x)) && c < 0 => (AND (MOVDconst [int64(uint32(c))]) x)
   589
   590// Combine negation and sign extension.
   591(MOVWreg (NEG x)) => (NEGW x)
   592
   593// Avoid sign/zero extension for consts.
   594(MOVBreg  (MOVDconst [c])) => (MOVDconst [int64(int8(c))])
   595(MOVHreg  (MOVDconst [c])) => (MOVDconst [int64(int16(c))])
   596(MOVWreg  (MOVDconst [c])) => (MOVDconst [int64(int32(c))])
   597(MOVBUreg (MOVDconst [c])) => (MOVDconst [int64(uint8(c))])
   598(MOVHUreg (MOVDconst [c])) => (MOVDconst [int64(uint16(c))])
   599(MOVWUreg (MOVDconst [c])) => (MOVDconst [int64(uint32(c))])
   600
   601// Avoid sign/zero extension after properly typed load.
   602(MOVBreg  x:(MOVBload  _ _)) => (MOVDreg x)
   603(MOVHreg  x:(MOVBload  _ _)) => (MOVDreg x)
   604(MOVHreg  x:(MOVBUload _ _)) => (MOVDreg x)
   605(MOVHreg  x:(MOVHload  _ _)) => (MOVDreg x)
   606(MOVWreg  x:(MOVBload  _ _)) => (MOVDreg x)
   607(MOVWreg  x:(MOVBUload _ _)) => (MOVDreg x)
   608(MOVWreg  x:(MOVHload  _ _)) => (MOVDreg x)
   609(MOVWreg  x:(MOVHUload _ _)) => (MOVDreg x)
   610(MOVWreg  x:(MOVWload  _ _)) => (MOVDreg x)
   611(MOVBUreg x:(MOVBUload _ _)) => (MOVDreg x)
   612(MOVHUreg x:(MOVBUload _ _)) => (MOVDreg x)
   613(MOVHUreg x:(MOVHUload _ _)) => (MOVDreg x)
   614(MOVWUreg x:(MOVBUload _ _)) => (MOVDreg x)
   615(MOVWUreg x:(MOVHUload _ _)) => (MOVDreg x)
   616(MOVWUreg x:(MOVWUload _ _)) => (MOVDreg x)
   617
   618// Avoid zero extension after properly typed atomic operation.
   619(MOVBUreg x:(Select0 (LoweredAtomicLoad8 _ _))) => (MOVDreg x)
   620(MOVBUreg x:(Select0 (LoweredAtomicCas32 _ _ _ _))) => (MOVDreg x)
   621(MOVBUreg x:(Select0 (LoweredAtomicCas64 _ _ _ _))) => (MOVDreg x)
   622
   623// Avoid sign extension after word arithmetic.
   624(MOVWreg x:(ADDIW   _)) => (MOVDreg x)
   625(MOVWreg x:(SUBW  _ _)) => (MOVDreg x)
   626(MOVWreg x:(NEGW    _)) => (MOVDreg x)
   627(MOVWreg x:(MULW  _ _)) => (MOVDreg x)
   628(MOVWreg x:(DIVW  _ _)) => (MOVDreg x)
   629(MOVWreg x:(DIVUW _ _)) => (MOVDreg x)
   630(MOVWreg x:(REMW  _ _)) => (MOVDreg x)
   631(MOVWreg x:(REMUW _ _)) => (MOVDreg x)
   632(MOVWreg x:(ROLW  _ _)) => (MOVDreg x)
   633(MOVWreg x:(RORW  _ _)) => (MOVDreg x)
   634(MOVWreg x:(RORIW   _)) => (MOVDreg x)
   635
   636// Fold double extensions.
   637(MOVBreg  x:(MOVBreg  _)) => (MOVDreg x)
   638(MOVHreg  x:(MOVBreg  _)) => (MOVDreg x)
   639(MOVHreg  x:(MOVBUreg _)) => (MOVDreg x)
   640(MOVHreg  x:(MOVHreg  _)) => (MOVDreg x)
   641(MOVWreg  x:(MOVBreg  _)) => (MOVDreg x)
   642(MOVWreg  x:(MOVBUreg _)) => (MOVDreg x)
   643(MOVWreg  x:(MOVHreg  _)) => (MOVDreg x)
   644(MOVWreg  x:(MOVWreg  _)) => (MOVDreg x)
   645(MOVBUreg x:(MOVBUreg _)) => (MOVDreg x)
   646(MOVHUreg x:(MOVBUreg _)) => (MOVDreg x)
   647(MOVHUreg x:(MOVHUreg _)) => (MOVDreg x)
   648(MOVWUreg x:(MOVBUreg _)) => (MOVDreg x)
   649(MOVWUreg x:(MOVHUreg _)) => (MOVDreg x)
   650(MOVWUreg x:(MOVWUreg _)) => (MOVDreg x)
   651
   652// Do not extend before store.
   653(MOVBstore [off] {sym} ptr (MOVBreg  x) mem) => (MOVBstore [off] {sym} ptr x mem)
   654(MOVBstore [off] {sym} ptr (MOVHreg  x) mem) => (MOVBstore [off] {sym} ptr x mem)
   655(MOVBstore [off] {sym} ptr (MOVWreg  x) mem) => (MOVBstore [off] {sym} ptr x mem)
   656(MOVBstore [off] {sym} ptr (MOVBUreg x) mem) => (MOVBstore [off] {sym} ptr x mem)
   657(MOVBstore [off] {sym} ptr (MOVHUreg x) mem) => (MOVBstore [off] {sym} ptr x mem)
   658(MOVBstore [off] {sym} ptr (MOVWUreg x) mem) => (MOVBstore [off] {sym} ptr x mem)
   659(MOVHstore [off] {sym} ptr (MOVHreg  x) mem) => (MOVHstore [off] {sym} ptr x mem)
   660(MOVHstore [off] {sym} ptr (MOVWreg  x) mem) => (MOVHstore [off] {sym} ptr x mem)
   661(MOVHstore [off] {sym} ptr (MOVHUreg x) mem) => (MOVHstore [off] {sym} ptr x mem)
   662(MOVHstore [off] {sym} ptr (MOVWUreg x) mem) => (MOVHstore [off] {sym} ptr x mem)
   663(MOVWstore [off] {sym} ptr (MOVWreg  x) mem) => (MOVWstore [off] {sym} ptr x mem)
   664(MOVWstore [off] {sym} ptr (MOVWUreg x) mem) => (MOVWstore [off] {sym} ptr x mem)
   665
   666// Replace extend after load with alternate load where possible.
   667(MOVBreg  <t> x:(MOVBUload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVBload  <t> [off] {sym} ptr mem)
   668(MOVHreg  <t> x:(MOVHUload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVHload  <t> [off] {sym} ptr mem)
   669(MOVWreg  <t> x:(MOVWUload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVWload  <t> [off] {sym} ptr mem)
   670(MOVBUreg <t> x:(MOVBload  [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVBUload <t> [off] {sym} ptr mem)
   671(MOVHUreg <t> x:(MOVHload  [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVHUload <t> [off] {sym} ptr mem)
   672(MOVWUreg <t> x:(MOVWload  [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVWUload <t> [off] {sym} ptr mem)
   673
   674// Replace load from same location as preceding store with copy.
   675(MOV(D|W|H|B)load [off] {sym} ptr1 (MOV(D|W|H|B)store [off] {sym} ptr2 x _)) && isSamePtr(ptr1, ptr2) => (MOV(D|W|H|B)reg x)
   676(MOV(W|H|B)Uload  [off] {sym} ptr1 (MOV(W|H|B)store   [off] {sym} ptr2 x _)) && isSamePtr(ptr1, ptr2) => (MOV(W|H|B)Ureg  x)
   677(MOVDload  [off] {sym} ptr1 (FMOVDstore [off] {sym} ptr2 x _)) && isSamePtr(ptr1, ptr2) => (FMVXD x)
   678(FMOVDload [off] {sym} ptr1 (MOVDstore  [off] {sym} ptr2 x _)) && isSamePtr(ptr1, ptr2) => (FMVDX x)
   679(MOVWload  [off] {sym} ptr1 (FMOVWstore [off] {sym} ptr2 x _)) && isSamePtr(ptr1, ptr2) => (FMVXS x)
   680(MOVWUload [off] {sym} ptr1 (FMOVWstore [off] {sym} ptr2 x _)) && isSamePtr(ptr1, ptr2) => (MOVWUreg (FMVXS x))
   681(FMOVWload [off] {sym} ptr1 (MOVWstore  [off] {sym} ptr2 x _)) && isSamePtr(ptr1, ptr2) => (FMVSX x)
   682
   683// If a register move has only 1 use, just use the same register without emitting instruction
   684// MOVnop does not emit an instruction, only for ensuring the type.
   685(MOVDreg x) && x.Uses == 1 => (MOVDnop x)
   686
   687// TODO: we should be able to get rid of MOVDnop all together.
   688// But for now, this is enough to get rid of lots of them.
   689(MOVDnop (MOVDconst [c])) => (MOVDconst [c])
   690
   691// Avoid unnecessary zero and sign extension when right shifting.
   692(SRAI [x] (MOVWreg  y)) && x >= 0 && x <= 31 => (SRAIW [x] y)
   693(SRLI [x] (MOVWUreg y)) && x >= 0 && x <= 31 => (SRLIW [x] y)
   694
   695// Replace right shifts that exceed size of signed type.
   696(SRAI <t> [x] (MOVBreg y)) && x >=  8 => (SRAI  [63] (SLLI <t> [56] y))
   697(SRAI <t> [x] (MOVHreg y)) && x >= 16 => (SRAI  [63] (SLLI <t> [48] y))
   698(SRAI     [x] (MOVWreg y)) && x >= 32 => (SRAIW [31] y)
   699
   700// Eliminate right shifts that exceed size of unsigned type.
   701(SRLI [x] (MOVBUreg y)) && x >=  8 => (MOVDconst [0])
   702(SRLI [x] (MOVHUreg y)) && x >= 16 => (MOVDconst [0])
   703(SRLI [x] (MOVWUreg y)) && x >= 32 => (MOVDconst [0])
   704
   705// Fold constant into immediate instructions where possible.
   706(ADD (MOVDconst <t> [val]) x) && is32Bit(val) && !t.IsPtr() => (ADDI [val] x)
   707(AND (MOVDconst [val]) x) && is32Bit(val) => (ANDI [val] x)
   708(OR  (MOVDconst [val]) x) && is32Bit(val) => (ORI  [val] x)
   709(XOR (MOVDconst [val]) x) && is32Bit(val) => (XORI [val] x)
   710(ROL  x (MOVDconst [val])) => (RORI  [-val&63] x)
   711(ROLW x (MOVDconst [val])) => (RORIW [-val&31] x)
   712(ROR  x (MOVDconst [val])) => (RORI  [val&63] x)
   713(RORW x (MOVDconst [val])) => (RORIW [val&31] x)
   714(SLL  x (MOVDconst [val])) => (SLLI  [val&63] x)
   715(SLLW x (MOVDconst [val])) => (SLLIW [val&31] x)
   716(SRL  x (MOVDconst [val])) => (SRLI  [val&63] x)
   717(SRLW x (MOVDconst [val])) => (SRLIW [val&31] x)
   718(SRA  x (MOVDconst [val])) => (SRAI  [val&63] x)
   719(SRAW x (MOVDconst [val])) => (SRAIW [val&31] x)
   720(SLT  x (MOVDconst [val])) && is12Bit(val) => (SLTI  [val] x)
   721(SLTU x (MOVDconst [val])) && is12Bit(val) => (SLTIU [val] x)
   722
   723// Replace negated left rotation with right rotation.
   724(ROL  x (NEG y)) => (ROR  x y)
   725(ROLW x (NEG y)) => (RORW x y)
   726
   727// generic simplifications
   728(ADD x (NEG y)) => (SUB x y)
   729(SUB x (NEG y)) => (ADD x y)
   730(SUB x x) => (MOVDconst [0])
   731(AND x x) => x
   732(OR  x x) => x
   733(ORN x x) => (MOVDconst [-1])
   734(XOR x x) => (MOVDconst [0])
   735
   736// Convert const subtraction into ADDI with negative immediate, where possible.
   737(SUB x (MOVDconst [val])) && is32Bit(-val) => (ADDI [-val] x)
   738(SUB <t> (MOVDconst [val]) y) && is32Bit(-val) => (NEG (ADDI <t> [-val] y))
   739
   740// Subtraction of zero.
   741(SUB  x (MOVDconst [0])) => x
   742(SUBW x (MOVDconst [0])) => (ADDIW [0] x)
   743
   744// Subtraction from zero.
   745(SUB  (MOVDconst [0]) x) => (NEG x)
   746(SUBW (MOVDconst [0]) x) => (NEGW x)
   747
   748// Fold negation into subtraction.
   749(NEG (SUB x y)) => (SUB y x)
   750(NEG <t> s:(ADDI [val] (SUB x y))) && s.Uses == 1 && is32Bit(-val) => (ADDI [-val] (SUB <t> y x))
   751
   752// Double negation.
   753(NEG (NEG x)) => x
   754(NEG <t> s:(ADDI [val] (NEG x))) && s.Uses == 1 && is32Bit(-val) => (ADDI [-val] x)
   755
   756// Addition of zero or two constants.
   757(ADDI [0] x) => x
   758(ADDI [x] (MOVDconst [y])) && is32Bit(x + y) => (MOVDconst [x + y])
   759
   760// ANDI with all zeros, all ones or two constants.
   761(ANDI [0]  x) => (MOVDconst [0])
   762(ANDI [-1] x) => x
   763(ANDI [x] (MOVDconst [y])) => (MOVDconst [x & y])
   764
   765// ORI with all zeroes, all ones or two constants.
   766(ORI [0]  x) => x
   767(ORI [-1] x) => (MOVDconst [-1])
   768(ORI [x] (MOVDconst [y])) => (MOVDconst [x | y])
   769
   770// Combine operations with immediate.
   771(ADDI [x] (ADDI [y] z)) && is32Bit(x + y) => (ADDI [x + y] z)
   772(ANDI [x] (ANDI [y] z)) => (ANDI [x & y] z)
   773(ORI  [x] (ORI  [y] z)) => (ORI  [x | y] z)
   774
   775// Negation of a constant.
   776(NEG  (MOVDconst [x])) => (MOVDconst [-x])
   777(NEGW (MOVDconst [x])) => (MOVDconst [int64(int32(-x))])
   778
   779// Shift of a constant.
   780(SLLI [x] (MOVDconst [y])) && is32Bit(y << uint32(x)) => (MOVDconst [y << uint32(x)])
   781(SRLI [x] (MOVDconst [y])) => (MOVDconst [int64(uint64(y) >> uint32(x))])
   782(SRAI [x] (MOVDconst [y])) => (MOVDconst [int64(y) >> uint32(x)])
   783
   784// Combine doubling via addition with shift.
   785(SLLI <t> [c] (ADD x x)) && c < t.Size() * 8 - 1 => (SLLI [c+1] x)
   786(SLLI <t> [c] (ADD x x)) && c >= t.Size() * 8 - 1 => (MOVDconst [0])
   787
   788// SLTI/SLTIU with constants.
   789(SLTI  [x] (MOVDconst [y])) => (MOVDconst [b2i(int64(y) < int64(x))])
   790(SLTIU [x] (MOVDconst [y])) => (MOVDconst [b2i(uint64(y) < uint64(x))])
   791
   792// SLTI/SLTIU with known outcomes.
   793(SLTI  [x] (ANDI [y] _)) && y >= 0 && int64(y) < int64(x) => (MOVDconst [1])
   794(SLTIU [x] (ANDI [y] _)) && y >= 0 && uint64(y) < uint64(x) => (MOVDconst [1])
   795(SLTIU [x] (ORI  [y] _)) && y >= 0 && uint64(y) >= uint64(x) => (MOVDconst [0])
   796
   797// SLT/SLTU with known outcomes.
   798(SLT  x x) => (MOVDconst [0])
   799(SLTU x x) => (MOVDconst [0])
   800
   801// Deadcode for LoweredMuluhilo
   802(Select0 m:(LoweredMuluhilo x y)) && m.Uses == 1 => (MULHU x y)
   803(Select1 m:(LoweredMuluhilo x y)) && m.Uses == 1 => (MUL x y)
   804
   805(FADD(S|D) a (FMUL(S|D) x y)) && a.Block.Func.useFMA(v) => (FMADD(S|D) x y a)
   806(FSUB(S|D) a (FMUL(S|D) x y)) && a.Block.Func.useFMA(v) => (FNMSUB(S|D) x y a)
   807(FSUB(S|D) (FMUL(S|D) x y) a) && a.Block.Func.useFMA(v) => (FMSUB(S|D) x y a)
   808
   809// Merge negation into fused multiply-add and multiply-subtract.
   810//
   811// Key:
   812//
   813//   [+ -](x * y [+ -] z).
   814//    _ N         A S
   815//                D U
   816//                D B
   817//
   818// Note: multiplication commutativity handled by rule generator.
   819(F(MADD|NMADD|MSUB|NMSUB)S neg:(FNEGS x) y z) && neg.Uses == 1 => (F(NMSUB|MSUB|NMADD|MADD)S x y z)
   820(F(MADD|NMADD|MSUB|NMSUB)S x y neg:(FNEGS z)) && neg.Uses == 1 => (F(MSUB|NMSUB|MADD|NMADD)S x y z)
   821(F(MADD|NMADD|MSUB|NMSUB)D neg:(FNEGD x) y z) && neg.Uses == 1 => (F(NMSUB|MSUB|NMADD|MADD)D x y z)
   822(F(MADD|NMADD|MSUB|NMSUB)D x y neg:(FNEGD z)) && neg.Uses == 1 => (F(MSUB|NMSUB|MADD|NMADD)D x y z)
   823
   824// Test for -∞ (bit 0) using 64 bit classify instruction.
   825(FLTD x (FMOVDconst [-math.MaxFloat64])) => (ANDI [0b00_0000_0001] (FCLASSD x))
   826(FLED (FMOVDconst [-math.MaxFloat64]) x) => (SNEZ (ANDI <typ.Int64> [0b00_1111_1110] (FCLASSD x)))
   827(FEQD x (FMOVDconst [math.Inf(-1)])) => (ANDI [0b00_0000_0001] (FCLASSD x))
   828(FNED x (FMOVDconst [math.Inf(-1)])) => (SEQZ (ANDI <typ.Int64> [0b00_0000_0001] (FCLASSD x)))
   829
   830// Test for +∞ (bit 7) using 64 bit classify instruction.
   831(FLTD (FMOVDconst [math.MaxFloat64]) x) => (SNEZ (ANDI <typ.Int64> [0b00_1000_0000] (FCLASSD x)))
   832(FLED x (FMOVDconst [math.MaxFloat64])) => (SNEZ (ANDI <typ.Int64> [0b00_0111_1111] (FCLASSD x)))
   833(FEQD x (FMOVDconst [math.Inf(1)])) => (SNEZ (ANDI <typ.Int64> [0b00_1000_0000] (FCLASSD x)))
   834(FNED x (FMOVDconst [math.Inf(1)])) => (SEQZ (ANDI <typ.Int64> [0b00_1000_0000] (FCLASSD x)))
   835
   836// Test for subnormal numbers using 64 bit classify instruction.
   837(FLTD x (FMOVDconst [+0x1p-1022])) => (SNEZ (ANDI <typ.Int64> [0b00_0011_1111] (FCLASSD x)))
   838(FLED (FMOVDconst [+0x1p-1022]) x) => (SNEZ (ANDI <typ.Int64> [0b00_1100_0000] (FCLASSD x)))
   839(FLED x (FMOVDconst [-0x1p-1022])) => (SNEZ (ANDI <typ.Int64> [0b00_0000_0011] (FCLASSD x)))
   840(FLTD (FMOVDconst [-0x1p-1022]) x) => (SNEZ (ANDI <typ.Int64> [0b00_1111_1100] (FCLASSD x)))
   841
   842// Absorb unary sign bit operations into 64 bit classify instruction.
   843(S(EQ|NE)Z (ANDI [c] (FCLASSD (FNEGD x)))) => (S(EQ|NE)Z (ANDI <typ.Int64> [(c&0b11_0000_0000)|int64(bits.Reverse8(uint8(c))&0b1111_1111)] (FCLASSD x)))
   844(S(EQ|NE)Z (ANDI [c] (FCLASSD (FABSD x)))) => (S(EQ|NE)Z (ANDI <typ.Int64> [(c&0b11_1111_0000)|int64(bits.Reverse8(uint8(c))&0b0000_1111)] (FCLASSD x)))
   845(B(EQ|NE)Z (ANDI [c] (FCLASSD (FNEGD x))) yes no) => (B(EQ|NE)Z (ANDI <typ.Int64> [(c&0b11_0000_0000)|int64(bits.Reverse8(uint8(c))&0b1111_1111)] (FCLASSD x)) yes no)
   846(B(EQ|NE)Z (ANDI [c] (FCLASSD (FABSD x))) yes no) => (B(EQ|NE)Z (ANDI <typ.Int64> [(c&0b11_1111_0000)|int64(bits.Reverse8(uint8(c))&0b0000_1111)] (FCLASSD x)) yes no)
   847
   848//
   849// Optimisations for rva22u64 and above.
   850//
   851
   852// Combine left shift and addition.
   853(ADD (SLLI [1] x) y) && buildcfg.GORISCV64 >= 22 => (SH1ADD x y)
   854(ADD (SLLI [2] x) y) && buildcfg.GORISCV64 >= 22 => (SH2ADD x y)
   855(ADD (SLLI [3] x) y) && buildcfg.GORISCV64 >= 22 => (SH3ADD x y)
   856
   857// Integer minimum and maximum.
   858(Min64  x y) && buildcfg.GORISCV64 >= 22 => (MIN  x y)
   859(Max64  x y) && buildcfg.GORISCV64 >= 22 => (MAX  x y)
   860(Min64u x y) && buildcfg.GORISCV64 >= 22 => (MINU x y)
   861(Max64u x y) && buildcfg.GORISCV64 >= 22 => (MAXU x y)

View as plain text