github.com/go-asm/go@v1.21.1-0.20240213172139-40c5ead50c48/cmd/compile/ssa/_gen/RISCV64.rules (about)

     1  // Copyright 2016 The Go Authors. All rights reserved.
     2  // Use of this source code is governed by a BSD-style
     3  // license that can be found in the LICENSE file.
     4  
     5  // Lowering arithmetic
     6  (Add(Ptr|64|32|16|8) ...) => (ADD ...)
     7  (Add(64|32)F ...) => (FADD(D|S) ...)
     8  
     9  (Sub(Ptr|64|32|16|8) ...) => (SUB ...)
    10  (Sub(64|32)F ...) => (FSUB(D|S) ...)
    11  
    12  (Mul64 ...) => (MUL  ...)
    13  (Mul64uhilo ...) => (LoweredMuluhilo ...)
    14  (Mul64uover ...) => (LoweredMuluover ...)
    15  (Mul32 ...) => (MULW ...)
    16  (Mul16 x y) => (MULW (SignExt16to32 x) (SignExt16to32 y))
    17  (Mul8 x y)  => (MULW (SignExt8to32 x)  (SignExt8to32 y))
    18  (Mul(64|32)F ...) => (FMUL(D|S) ...)
    19  
    20  (Div(64|32)F ...) => (FDIV(D|S) ...)
    21  
    22  (Div64 x y [false])  => (DIV x y)
    23  (Div64u ...) => (DIVU ...)
    24  (Div32 x y [false])  => (DIVW x y)
    25  (Div32u ...) => (DIVUW ...)
    26  (Div16 x y [false])  => (DIVW  (SignExt16to32 x) (SignExt16to32 y))
    27  (Div16u x y) => (DIVUW (ZeroExt16to32 x) (ZeroExt16to32 y))
    28  (Div8 x y)   => (DIVW  (SignExt8to32 x)  (SignExt8to32 y))
    29  (Div8u x y)  => (DIVUW (ZeroExt8to32 x)  (ZeroExt8to32 y))
    30  
    31  (Hmul64 ...)  => (MULH  ...)
    32  (Hmul64u ...) => (MULHU ...)
    33  (Hmul32 x y)  => (SRAI [32] (MUL  (SignExt32to64 x) (SignExt32to64 y)))
    34  (Hmul32u x y) => (SRLI [32] (MUL  (ZeroExt32to64 x) (ZeroExt32to64 y)))
    35  
    36  (Select0 (Add64carry x y c)) => (ADD (ADD <typ.UInt64> x y) c)
    37  (Select1 (Add64carry x y c)) =>
    38  	(OR (SLTU <typ.UInt64> s:(ADD <typ.UInt64> x y) x) (SLTU <typ.UInt64> (ADD <typ.UInt64> s c) s))
    39  
    40  (Select0 (Sub64borrow x y c)) => (SUB (SUB <typ.UInt64> x y) c)
    41  (Select1 (Sub64borrow x y c)) =>
    42  	(OR (SLTU <typ.UInt64> x s:(SUB <typ.UInt64> x y)) (SLTU <typ.UInt64> s (SUB <typ.UInt64> s c)))
    43  
    44  // (x + y) / 2 => (x / 2) + (y / 2) + (x & y & 1)
    45  (Avg64u <t> x y) => (ADD (ADD <t> (SRLI <t> [1] x) (SRLI <t> [1] y)) (ANDI <t> [1] (AND <t> x y)))
    46  
    47  (Mod64 x y [false])  => (REM x y)
    48  (Mod64u ...) => (REMU  ...)
    49  (Mod32 x y [false])  => (REMW x y)
    50  (Mod32u ...) => (REMUW ...)
    51  (Mod16 x y [false])  => (REMW  (SignExt16to32 x) (SignExt16to32 y))
    52  (Mod16u x y) => (REMUW (ZeroExt16to32 x) (ZeroExt16to32 y))
    53  (Mod8 x y)   => (REMW  (SignExt8to32  x) (SignExt8to32  y))
    54  (Mod8u x y)  => (REMUW (ZeroExt8to32  x) (ZeroExt8to32  y))
    55  
    56  (And(64|32|16|8) ...) => (AND ...)
    57  (Or(64|32|16|8) ...) => (OR ...)
    58  (Xor(64|32|16|8) ...) => (XOR ...)
    59  
    60  (Neg(64|32|16|8) ...) => (NEG ...)
    61  (Neg(64|32)F ...) => (FNEG(D|S) ...)
    62  
    63  (Com(64|32|16|8) ...) => (NOT ...)
    64  
    65  
    66  (Sqrt ...) => (FSQRTD ...)
    67  (Sqrt32 ...) => (FSQRTS ...)
    68  
    69  (Copysign ...) => (FSGNJD ...)
    70  
    71  (Abs ...) => (FABSD ...)
    72  
    73  (FMA ...) => (FMADDD ...)
    74  
    75  // 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 (MOVDconst [c])) => (Or8  (Lsh8x64  <t> x (MOVDconst [c&7]))  (Rsh8Ux64  <t> x (MOVDconst [-c&7])))
   215  (RotateLeft16 <t> x (MOVDconst [c])) => (Or16 (Lsh16x64 <t> x (MOVDconst [c&15])) (Rsh16Ux64 <t> x (MOVDconst [-c&15])))
   216  (RotateLeft32 <t> x (MOVDconst [c])) => (Or32 (Lsh32x64 <t> x (MOVDconst [c&31])) (Rsh32Ux64 <t> x (MOVDconst [-c&31])))
   217  (RotateLeft64 <t> x (MOVDconst [c])) => (Or64 (Lsh64x64 <t> x (MOVDconst [c&63])) (Rsh64Ux64 <t> x (MOVDconst [-c&63])))
   218  
   219  (Less64  ...) => (SLT  ...)
   220  (Less32  x y) => (SLT  (SignExt32to64 x) (SignExt32to64 y))
   221  (Less16  x y) => (SLT  (SignExt16to64 x) (SignExt16to64 y))
   222  (Less8   x y) => (SLT  (SignExt8to64  x) (SignExt8to64  y))
   223  (Less64U ...) => (SLTU ...)
   224  (Less32U x y) => (SLTU (ZeroExt32to64 x) (ZeroExt32to64 y))
   225  (Less16U x y) => (SLTU (ZeroExt16to64 x) (ZeroExt16to64 y))
   226  (Less8U  x y) => (SLTU (ZeroExt8to64  x) (ZeroExt8to64  y))
   227  (Less(64|32)F ...) => (FLT(D|S) ...)
   228  
   229  // Convert x <= y to !(y > x).
   230  (Leq(64|32|16|8)  x y) => (Not (Less(64|32|16|8)  y x))
   231  (Leq(64|32|16|8)U x y) => (Not (Less(64|32|16|8)U y x))
   232  (Leq(64|32)F ...) => (FLE(D|S) ...)
   233  
   234  (EqPtr x y) => (SEQZ (SUB <typ.Uintptr> x y))
   235  (Eq64  x y) => (SEQZ (SUB <x.Type> x y))
   236  (Eq32  x y) &&  x.Type.IsSigned() => (SEQZ (SUB <x.Type> (SignExt32to64 x) (SignExt32to64 y)))
   237  (Eq32  x y) && !x.Type.IsSigned() => (SEQZ (SUB <x.Type> (ZeroExt32to64 x) (ZeroExt32to64 y)))
   238  (Eq16  x y) => (SEQZ (SUB <x.Type> (ZeroExt16to64 x) (ZeroExt16to64 y)))
   239  (Eq8   x y) => (SEQZ (SUB <x.Type> (ZeroExt8to64  x) (ZeroExt8to64  y)))
   240  (Eq(64|32)F ...) => (FEQ(D|S) ...)
   241  
   242  (NeqPtr x y) => (Not (EqPtr x y))
   243  (Neq64  x y) => (Not (Eq64  x y))
   244  (Neq32  x y) => (Not (Eq32  x y))
   245  (Neq16  x y) => (Not (Eq16  x y))
   246  (Neq8   x y) => (Not (Eq8   x y))
   247  (Neq(64|32)F ...) => (FNE(D|S) ...)
   248  
   249  // Loads
   250  (Load <t> ptr mem) &&  t.IsBoolean()                   => (MOVBUload ptr mem)
   251  (Load <t> ptr mem) && ( is8BitInt(t) &&  t.IsSigned()) => (MOVBload  ptr mem)
   252  (Load <t> ptr mem) && ( is8BitInt(t) && !t.IsSigned()) => (MOVBUload ptr mem)
   253  (Load <t> ptr mem) && (is16BitInt(t) &&  t.IsSigned()) => (MOVHload  ptr mem)
   254  (Load <t> ptr mem) && (is16BitInt(t) && !t.IsSigned()) => (MOVHUload ptr mem)
   255  (Load <t> ptr mem) && (is32BitInt(t) &&  t.IsSigned()) => (MOVWload  ptr mem)
   256  (Load <t> ptr mem) && (is32BitInt(t) && !t.IsSigned()) => (MOVWUload ptr mem)
   257  (Load <t> ptr mem) && (is64BitInt(t) || isPtr(t))      => (MOVDload  ptr mem)
   258  (Load <t> ptr mem) &&  is32BitFloat(t)                 => (FMOVWload ptr mem)
   259  (Load <t> ptr mem) &&  is64BitFloat(t)                 => (FMOVDload ptr mem)
   260  
   261  // Stores
   262  (Store {t} ptr val mem) && t.Size() == 1 => (MOVBstore ptr val mem)
   263  (Store {t} ptr val mem) && t.Size() == 2 => (MOVHstore ptr val mem)
   264  (Store {t} ptr val mem) && t.Size() == 4 && !t.IsFloat() => (MOVWstore ptr val mem)
   265  (Store {t} ptr val mem) && t.Size() == 8 && !t.IsFloat() => (MOVDstore ptr val mem)
   266  (Store {t} ptr val mem) && t.Size() == 4 &&  t.IsFloat() => (FMOVWstore ptr val mem)
   267  (Store {t} ptr val mem) && t.Size() == 8 &&  t.IsFloat() => (FMOVDstore ptr val mem)
   268  
   269  // We need to fold MOVaddr into the LD/MOVDstore ops so that the live variable analysis
   270  // knows what variables are being read/written by the ops.
   271  (MOVBUload [off1] {sym1} (MOVaddr [off2] {sym2} base) mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   272  	(MOVBUload [off1+off2] {mergeSym(sym1,sym2)} base mem)
   273  (MOVBload  [off1] {sym1} (MOVaddr [off2] {sym2} base) mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   274  	(MOVBload  [off1+off2] {mergeSym(sym1,sym2)} base mem)
   275  (MOVHUload [off1] {sym1} (MOVaddr [off2] {sym2} base) mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   276  	(MOVHUload [off1+off2] {mergeSym(sym1,sym2)} base mem)
   277  (MOVHload  [off1] {sym1} (MOVaddr [off2] {sym2} base) mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   278  	(MOVHload  [off1+off2] {mergeSym(sym1,sym2)} base mem)
   279  (MOVWUload [off1] {sym1} (MOVaddr [off2] {sym2} base) mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   280  	(MOVWUload [off1+off2] {mergeSym(sym1,sym2)} base mem)
   281  (MOVWload  [off1] {sym1} (MOVaddr [off2] {sym2} base) mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   282  	(MOVWload  [off1+off2] {mergeSym(sym1,sym2)} base mem)
   283  (MOVDload  [off1] {sym1} (MOVaddr [off2] {sym2} base) mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   284  	(MOVDload  [off1+off2] {mergeSym(sym1,sym2)} base mem)
   285  
   286  (MOVBstore [off1] {sym1} (MOVaddr [off2] {sym2} base) val mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   287  	(MOVBstore [off1+off2] {mergeSym(sym1,sym2)} base val mem)
   288  (MOVHstore [off1] {sym1} (MOVaddr [off2] {sym2} base) val mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   289  	(MOVHstore [off1+off2] {mergeSym(sym1,sym2)} base val mem)
   290  (MOVWstore [off1] {sym1} (MOVaddr [off2] {sym2} base) val mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   291  	(MOVWstore [off1+off2] {mergeSym(sym1,sym2)} base val mem)
   292  (MOVDstore [off1] {sym1} (MOVaddr [off2] {sym2} base) val mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   293  	(MOVDstore [off1+off2] {mergeSym(sym1,sym2)} base val mem)
   294  (MOVBstorezero [off1] {sym1} (MOVaddr [off2] {sym2} ptr) mem) && canMergeSym(sym1,sym2) && is32Bit(int64(off1)+int64(off2)) =>
   295  	(MOVBstorezero [off1+off2] {mergeSym(sym1,sym2)} ptr mem)
   296  (MOVHstorezero [off1] {sym1} (MOVaddr [off2] {sym2} ptr) mem) && canMergeSym(sym1,sym2) && is32Bit(int64(off1)+int64(off2)) =>
   297  	(MOVHstorezero [off1+off2] {mergeSym(sym1,sym2)} ptr mem)
   298  (MOVWstorezero [off1] {sym1} (MOVaddr [off2] {sym2} ptr) mem) && canMergeSym(sym1,sym2) && is32Bit(int64(off1)+int64(off2)) =>
   299  	(MOVWstorezero [off1+off2] {mergeSym(sym1,sym2)} ptr mem)
   300  (MOVDstorezero [off1] {sym1} (MOVaddr [off2] {sym2} ptr) mem) && canMergeSym(sym1,sym2) && is32Bit(int64(off1)+int64(off2)) =>
   301  	(MOVDstorezero [off1+off2] {mergeSym(sym1,sym2)} ptr mem)
   302  
   303  (MOVBUload [off1] {sym} (ADDI [off2] base) mem) && is32Bit(int64(off1)+off2) =>
   304  	(MOVBUload [off1+int32(off2)] {sym} base mem)
   305  (MOVBload  [off1] {sym} (ADDI [off2] base) mem) && is32Bit(int64(off1)+off2) =>
   306  	(MOVBload  [off1+int32(off2)] {sym} base mem)
   307  (MOVHUload [off1] {sym} (ADDI [off2] base) mem) && is32Bit(int64(off1)+off2) =>
   308  	(MOVHUload [off1+int32(off2)] {sym} base mem)
   309  (MOVHload  [off1] {sym} (ADDI [off2] base) mem) && is32Bit(int64(off1)+off2) =>
   310  	(MOVHload  [off1+int32(off2)] {sym} base mem)
   311  (MOVWUload [off1] {sym} (ADDI [off2] base) mem) && is32Bit(int64(off1)+off2) =>
   312  	(MOVWUload [off1+int32(off2)] {sym} base mem)
   313  (MOVWload  [off1] {sym} (ADDI [off2] base) mem) && is32Bit(int64(off1)+off2) =>
   314  	(MOVWload  [off1+int32(off2)] {sym} base mem)
   315  (MOVDload  [off1] {sym} (ADDI [off2] base) mem) && is32Bit(int64(off1)+off2) =>
   316  	(MOVDload  [off1+int32(off2)] {sym} base mem)
   317  
   318  (MOVBstore [off1] {sym} (ADDI [off2] base) val mem) && is32Bit(int64(off1)+off2) =>
   319  	(MOVBstore [off1+int32(off2)] {sym} base val mem)
   320  (MOVHstore [off1] {sym} (ADDI [off2] base) val mem) && is32Bit(int64(off1)+off2) =>
   321  	(MOVHstore [off1+int32(off2)] {sym} base val mem)
   322  (MOVWstore [off1] {sym} (ADDI [off2] base) val mem) && is32Bit(int64(off1)+off2) =>
   323  	(MOVWstore [off1+int32(off2)] {sym} base val mem)
   324  (MOVDstore [off1] {sym} (ADDI [off2] base) val mem) && is32Bit(int64(off1)+off2) =>
   325  	(MOVDstore [off1+int32(off2)] {sym} base val mem)
   326  (MOVBstorezero [off1] {sym} (ADDI [off2] ptr) mem) && is32Bit(int64(off1)+off2) => (MOVBstorezero [off1+int32(off2)] {sym} ptr mem)
   327  (MOVHstorezero [off1] {sym} (ADDI [off2] ptr) mem) && is32Bit(int64(off1)+off2) => (MOVHstorezero [off1+int32(off2)] {sym} ptr mem)
   328  (MOVWstorezero [off1] {sym} (ADDI [off2] ptr) mem) && is32Bit(int64(off1)+off2) => (MOVWstorezero [off1+int32(off2)] {sym} ptr mem)
   329  (MOVDstorezero [off1] {sym} (ADDI [off2] ptr) mem) && is32Bit(int64(off1)+off2) => (MOVDstorezero [off1+int32(off2)] {sym} ptr mem)
   330  
   331  // Similarly, fold ADDI into MOVaddr to avoid confusing live variable analysis
   332  // with OffPtr -> ADDI.
   333  (ADDI [c] (MOVaddr [d] {s} x)) && is32Bit(c+int64(d)) => (MOVaddr [int32(c)+d] {s} x)
   334  
   335  // Small zeroing
   336  (Zero [0] _ mem) => mem
   337  (Zero [1] ptr mem) => (MOVBstore ptr (MOVDconst [0]) mem)
   338  (Zero [2] {t} ptr mem) && t.Alignment()%2 == 0 =>
   339  	(MOVHstore ptr (MOVDconst [0]) mem)
   340  (Zero [2] ptr mem) =>
   341  	(MOVBstore [1] ptr (MOVDconst [0])
   342  		(MOVBstore ptr (MOVDconst [0]) mem))
   343  (Zero [4] {t} ptr mem) && t.Alignment()%4 == 0 =>
   344  	(MOVWstore ptr (MOVDconst [0]) mem)
   345  (Zero [4] {t} ptr mem) && t.Alignment()%2 == 0 =>
   346  	(MOVHstore [2] ptr (MOVDconst [0])
   347  		(MOVHstore ptr (MOVDconst [0]) mem))
   348  (Zero [4] ptr mem) =>
   349  	(MOVBstore [3] ptr (MOVDconst [0])
   350  		(MOVBstore [2] ptr (MOVDconst [0])
   351  			(MOVBstore [1] ptr (MOVDconst [0])
   352  				(MOVBstore ptr (MOVDconst [0]) mem))))
   353  (Zero [8] {t} ptr mem) && t.Alignment()%8 == 0 =>
   354  	(MOVDstore ptr (MOVDconst [0]) mem)
   355  (Zero [8] {t} ptr mem) && t.Alignment()%4 == 0 =>
   356  	(MOVWstore [4] ptr (MOVDconst [0])
   357  		(MOVWstore ptr (MOVDconst [0]) mem))
   358  (Zero [8] {t} ptr mem) && t.Alignment()%2 == 0 =>
   359  	(MOVHstore [6] ptr (MOVDconst [0])
   360  		(MOVHstore [4] ptr (MOVDconst [0])
   361  			(MOVHstore [2] ptr (MOVDconst [0])
   362  				(MOVHstore ptr (MOVDconst [0]) mem))))
   363  
   364  (Zero [3] ptr mem) =>
   365  	(MOVBstore [2] ptr (MOVDconst [0])
   366  		(MOVBstore [1] ptr (MOVDconst [0])
   367  			(MOVBstore ptr (MOVDconst [0]) mem)))
   368  (Zero [6] {t} ptr mem) && t.Alignment()%2 == 0 =>
   369  	(MOVHstore [4] ptr (MOVDconst [0])
   370  		(MOVHstore [2] ptr (MOVDconst [0])
   371  			(MOVHstore ptr (MOVDconst [0]) mem)))
   372  (Zero [12] {t} ptr mem) && t.Alignment()%4 == 0 =>
   373  	(MOVWstore [8] ptr (MOVDconst [0])
   374  		(MOVWstore [4] ptr (MOVDconst [0])
   375  			(MOVWstore ptr (MOVDconst [0]) mem)))
   376  (Zero [16] {t} ptr mem) && t.Alignment()%8 == 0 =>
   377  	(MOVDstore [8] ptr (MOVDconst [0])
   378  		(MOVDstore ptr (MOVDconst [0]) mem))
   379  (Zero [24] {t} ptr mem) && t.Alignment()%8 == 0 =>
   380  	(MOVDstore [16] ptr (MOVDconst [0])
   381  		(MOVDstore [8] ptr (MOVDconst [0])
   382  			(MOVDstore ptr (MOVDconst [0]) mem)))
   383  (Zero [32] {t} ptr mem) && t.Alignment()%8 == 0 =>
   384  	(MOVDstore [24] ptr (MOVDconst [0])
   385  		(MOVDstore [16] ptr (MOVDconst [0])
   386  			(MOVDstore [8] ptr (MOVDconst [0])
   387  				(MOVDstore ptr (MOVDconst [0]) mem))))
   388  
   389  // Medium 8-aligned zeroing uses a Duff's device
   390  // 8 and 128 are magic constants, see runtime/mkduff.go
   391  (Zero [s] {t} ptr mem)
   392  	&& s%8 == 0 && s <= 8*128
   393  	&& t.Alignment()%8 == 0 && !config.noDuffDevice =>
   394  	(DUFFZERO [8 * (128 - s/8)] ptr mem)
   395  
   396  // Generic zeroing uses a loop
   397  (Zero [s] {t} ptr mem) =>
   398  	(LoweredZero [t.Alignment()]
   399  		ptr
   400  		(ADD <ptr.Type> ptr (MOVDconst [s-moveSize(t.Alignment(), config)]))
   401  		mem)
   402  
   403  // Checks
   404  (IsNonNil ...) => (SNEZ ...)
   405  (IsInBounds ...) => (Less64U ...)
   406  (IsSliceInBounds ...) => (Leq64U ...)
   407  
   408  // Trivial lowering
   409  (NilCheck ...) => (LoweredNilCheck ...)
   410  (GetClosurePtr ...) => (LoweredGetClosurePtr ...)
   411  (GetCallerSP ...) => (LoweredGetCallerSP ...)
   412  (GetCallerPC ...) => (LoweredGetCallerPC ...)
   413  
   414  // Write barrier.
   415  (WB ...) => (LoweredWB ...)
   416  
   417  // Publication barrier as intrinsic
   418  (PubBarrier ...) => (LoweredPubBarrier ...)
   419  
   420  (PanicBounds [kind] x y mem) && boundsABI(kind) == 0 => (LoweredPanicBoundsA [kind] x y mem)
   421  (PanicBounds [kind] x y mem) && boundsABI(kind) == 1 => (LoweredPanicBoundsB [kind] x y mem)
   422  (PanicBounds [kind] x y mem) && boundsABI(kind) == 2 => (LoweredPanicBoundsC [kind] x y mem)
   423  
   424  // Small moves
   425  (Move [0] _ _ mem) => mem
   426  (Move [1] dst src mem) => (MOVBstore dst (MOVBload src mem) mem)
   427  (Move [2] {t} dst src mem) && t.Alignment()%2 == 0 =>
   428  	(MOVHstore dst (MOVHload src mem) mem)
   429  (Move [2] dst src mem) =>
   430  	(MOVBstore [1] dst (MOVBload [1] src mem)
   431  		(MOVBstore dst (MOVBload src mem) mem))
   432  (Move [4] {t} dst src mem) && t.Alignment()%4 == 0 =>
   433  	(MOVWstore dst (MOVWload src mem) mem)
   434  (Move [4] {t} dst src mem) && t.Alignment()%2 == 0 =>
   435  	(MOVHstore [2] dst (MOVHload [2] src mem)
   436  		(MOVHstore dst (MOVHload src mem) mem))
   437  (Move [4] dst src mem) =>
   438  	(MOVBstore [3] dst (MOVBload [3] src mem)
   439  		(MOVBstore [2] dst (MOVBload [2] src mem)
   440  			(MOVBstore [1] dst (MOVBload [1] src mem)
   441  				(MOVBstore dst (MOVBload src mem) mem))))
   442  (Move [8] {t} dst src mem) && t.Alignment()%8 == 0 =>
   443  	(MOVDstore dst (MOVDload src mem) mem)
   444  (Move [8] {t} dst src mem) && t.Alignment()%4 == 0 =>
   445  	(MOVWstore [4] dst (MOVWload [4] src mem)
   446  		(MOVWstore dst (MOVWload src mem) mem))
   447  (Move [8] {t} dst src mem) && t.Alignment()%2 == 0 =>
   448  	(MOVHstore [6] dst (MOVHload [6] src mem)
   449  		(MOVHstore [4] dst (MOVHload [4] src mem)
   450  			(MOVHstore [2] dst (MOVHload [2] src mem)
   451  				(MOVHstore dst (MOVHload src mem) mem))))
   452  
   453  (Move [3] dst src mem) =>
   454  	(MOVBstore [2] dst (MOVBload [2] src mem)
   455  		(MOVBstore [1] dst (MOVBload [1] src mem)
   456  			(MOVBstore dst (MOVBload src mem) mem)))
   457  (Move [6] {t} dst src mem) && t.Alignment()%2 == 0 =>
   458  	(MOVHstore [4] dst (MOVHload [4] src mem)
   459  		(MOVHstore [2] dst (MOVHload [2] src mem)
   460  			(MOVHstore dst (MOVHload src mem) mem)))
   461  (Move [12] {t} dst src mem) && t.Alignment()%4 == 0 =>
   462  	(MOVWstore [8] dst (MOVWload [8] src mem)
   463  		(MOVWstore [4] dst (MOVWload [4] src mem)
   464  			(MOVWstore dst (MOVWload src mem) mem)))
   465  (Move [16] {t} dst src mem) && t.Alignment()%8 == 0 =>
   466  	(MOVDstore [8] dst (MOVDload [8] src mem)
   467  		(MOVDstore dst (MOVDload src mem) mem))
   468  (Move [24] {t} dst src mem) && t.Alignment()%8 == 0 =>
   469  	(MOVDstore [16] dst (MOVDload [16] src mem)
   470  		(MOVDstore [8] dst (MOVDload [8] src mem)
   471  			(MOVDstore dst (MOVDload src mem) mem)))
   472  (Move [32] {t} dst src mem) && t.Alignment()%8 == 0 =>
   473  	(MOVDstore [24] dst (MOVDload [24] src mem)
   474  		(MOVDstore [16] dst (MOVDload [16] src mem)
   475  			(MOVDstore [8] dst (MOVDload [8] src mem)
   476  				(MOVDstore dst (MOVDload src mem) mem))))
   477  
   478  // Medium 8-aligned move uses a Duff's device
   479  // 16 and 128 are magic constants, see runtime/mkduff.go
   480  (Move [s] {t} dst src mem)
   481  	&& s%8 == 0 && s <= 8*128 && t.Alignment()%8 == 0
   482  	&& !config.noDuffDevice && logLargeCopy(v, s) =>
   483  	(DUFFCOPY [16 * (128 - s/8)] dst src mem)
   484  
   485  // Generic move uses a loop
   486  (Move [s] {t} dst src mem) && (s <= 16 || logLargeCopy(v, s)) =>
   487  	(LoweredMove [t.Alignment()]
   488  		dst
   489  		src
   490  		(ADDI <src.Type> [s-moveSize(t.Alignment(), config)] src)
   491  		mem)
   492  
   493  // Boolean ops; 0=false, 1=true
   494  (AndB ...) => (AND ...)
   495  (OrB  ...) => (OR  ...)
   496  (EqB  x y) => (SEQZ (SUB <typ.Bool> x y))
   497  (NeqB x y) => (SNEZ (SUB <typ.Bool> x y))
   498  (Not  ...) => (SEQZ ...)
   499  
   500  // Lowering pointer arithmetic
   501  // TODO: Special handling for SP offsets, like ARM
   502  (OffPtr [off] ptr:(SP)) && is32Bit(off) => (MOVaddr [int32(off)] ptr)
   503  (OffPtr [off] ptr) && is32Bit(off) => (ADDI [off] ptr)
   504  (OffPtr [off] ptr) => (ADD (MOVDconst [off]) ptr)
   505  
   506  (Const(64|32|16|8) [val]) => (MOVDconst [int64(val)])
   507  (Const32F [val]) => (FMVSX (MOVDconst [int64(math.Float32bits(val))]))
   508  (Const64F [val]) => (FMVDX (MOVDconst [int64(math.Float64bits(val))]))
   509  (ConstNil) => (MOVDconst [0])
   510  (ConstBool [val]) => (MOVDconst [int64(b2i(val))])
   511  
   512  (Addr {sym} base) => (MOVaddr {sym} [0] base)
   513  (LocalAddr <t> {sym} base mem) && t.Elem().HasPointers() => (MOVaddr {sym} (SPanchored base mem))
   514  (LocalAddr <t> {sym} base _)  && !t.Elem().HasPointers() => (MOVaddr {sym} base)
   515  
   516  // Calls
   517  (StaticCall  ...) => (CALLstatic  ...)
   518  (ClosureCall ...) => (CALLclosure ...)
   519  (InterCall   ...) => (CALLinter   ...)
   520  (TailCall ...) => (CALLtail ...)
   521  
   522  // Atomic Intrinsics
   523  (AtomicLoad(Ptr|64|32|8)  ...) => (LoweredAtomicLoad(64|64|32|8) ...)
   524  (AtomicStore(PtrNoWB|64|32|8) ...) => (LoweredAtomicStore(64|64|32|8) ...)
   525  (AtomicAdd(64|32) ...) => (LoweredAtomicAdd(64|32) ...)
   526  
   527  // AtomicAnd8(ptr,val) => LoweredAtomicAnd32(ptr&^3, ^((uint8(val) ^ 0xff) << ((ptr & 3) * 8)))
   528  (AtomicAnd8 ptr val mem) =>
   529  	(LoweredAtomicAnd32 (ANDI <typ.Uintptr> [^3] ptr)
   530  		(NOT <typ.UInt32> (SLL <typ.UInt32> (XORI <typ.UInt32> [0xff] (ZeroExt8to32 val))
   531  			(SLLI <typ.UInt64> [3] (ANDI <typ.UInt64> [3] ptr)))) mem)
   532  
   533  (AtomicAnd32 ...) => (LoweredAtomicAnd32 ...)
   534  
   535  (AtomicCompareAndSwap32 ptr old new mem) => (LoweredAtomicCas32 ptr (SignExt32to64 old) new mem)
   536  (AtomicCompareAndSwap64 ...) => (LoweredAtomicCas64 ...)
   537  
   538  (AtomicExchange(64|32) ...) => (LoweredAtomicExchange(64|32) ...)
   539  
   540  // AtomicOr8(ptr,val)  => LoweredAtomicOr32(ptr&^3, uint32(val)<<((ptr&3)*8))
   541  (AtomicOr8 ptr val mem) =>
   542  	(LoweredAtomicOr32 (ANDI <typ.Uintptr> [^3] ptr)
   543  		(SLL <typ.UInt32> (ZeroExt8to32 val)
   544  			(SLLI <typ.UInt64> [3] (ANDI <typ.UInt64> [3] ptr))) mem)
   545  
   546  (AtomicOr32  ...) => (LoweredAtomicOr32  ...)
   547  
   548  // Conditional branches
   549  (If cond yes no) => (BNEZ (MOVBUreg <typ.UInt64> cond) yes no)
   550  
   551  // Optimizations
   552  
   553  // Absorb SEQZ/SNEZ into branch.
   554  (BEQZ (SEQZ x) yes no) => (BNEZ x yes no)
   555  (BEQZ (SNEZ x) yes no) => (BEQZ x yes no)
   556  (BNEZ (SEQZ x) yes no) => (BEQZ x yes no)
   557  (BNEZ (SNEZ x) yes no) => (BNEZ x yes no)
   558  
   559  // Remove redundant NEG from BEQZ/BNEZ.
   560  (BEQZ (NEG x) yes no) => (BEQZ x yes no)
   561  (BNEZ (NEG x) yes no) => (BNEZ x yes no)
   562  
   563  // Negate comparison with FNES/FNED.
   564  (BEQZ (FNES <t> x y) yes no) => (BNEZ (FEQS <t> x y) yes no)
   565  (BNEZ (FNES <t> x y) yes no) => (BEQZ (FEQS <t> x y) yes no)
   566  (BEQZ (FNED <t> x y) yes no) => (BNEZ (FEQD <t> x y) yes no)
   567  (BNEZ (FNED <t> x y) yes no) => (BEQZ (FEQD <t> x y) yes no)
   568  
   569  // Convert BEQZ/BNEZ into more optimal branch conditions.
   570  (BEQZ (SUB x y) yes no) => (BEQ x y yes no)
   571  (BNEZ (SUB x y) yes no) => (BNE x y yes no)
   572  (BEQZ (SLT x y) yes no) => (BGE x y yes no)
   573  (BNEZ (SLT x y) yes no) => (BLT x y yes no)
   574  (BEQZ (SLTU x y) yes no) => (BGEU x y yes no)
   575  (BNEZ (SLTU x y) yes no) => (BLTU x y yes no)
   576  (BEQZ (SLTI [x] y) yes no) => (BGE y (MOVDconst [x]) yes no)
   577  (BNEZ (SLTI [x] y) yes no) => (BLT y (MOVDconst [x]) yes no)
   578  (BEQZ (SLTIU [x] y) yes no) => (BGEU y (MOVDconst [x]) yes no)
   579  (BNEZ (SLTIU [x] y) yes no) => (BLTU y (MOVDconst [x]) yes no)
   580  
   581  // Convert branch with zero to more optimal branch zero.
   582  (BEQ (MOVDconst [0]) cond yes no) => (BEQZ cond yes no)
   583  (BEQ cond (MOVDconst [0]) yes no) => (BEQZ cond yes no)
   584  (BNE (MOVDconst [0]) cond yes no) => (BNEZ cond yes no)
   585  (BNE cond (MOVDconst [0]) yes no) => (BNEZ cond yes no)
   586  (BLT (MOVDconst [0]) cond yes no) => (BGTZ cond yes no)
   587  (BLT cond (MOVDconst [0]) yes no) => (BLTZ cond yes no)
   588  (BGE (MOVDconst [0]) cond yes no) => (BLEZ cond yes no)
   589  (BGE cond (MOVDconst [0]) yes no) => (BGEZ cond yes no)
   590  
   591  // Remove redundant NEG from SEQZ/SNEZ.
   592  (SEQZ (NEG x)) => (SEQZ x)
   593  (SNEZ (NEG x)) => (SNEZ x)
   594  
   595  // Remove redundant SEQZ/SNEZ.
   596  (SEQZ (SEQZ x)) => (SNEZ x)
   597  (SEQZ (SNEZ x)) => (SEQZ x)
   598  (SNEZ (SEQZ x)) => (SEQZ x)
   599  (SNEZ (SNEZ x)) => (SNEZ x)
   600  
   601  // Store zero.
   602  (MOVBstore [off] {sym} ptr (MOVDconst [0]) mem) => (MOVBstorezero [off] {sym} ptr mem)
   603  (MOVHstore [off] {sym} ptr (MOVDconst [0]) mem) => (MOVHstorezero [off] {sym} ptr mem)
   604  (MOVWstore [off] {sym} ptr (MOVDconst [0]) mem) => (MOVWstorezero [off] {sym} ptr mem)
   605  (MOVDstore [off] {sym} ptr (MOVDconst [0]) mem) => (MOVDstorezero [off] {sym} ptr mem)
   606  
   607  // Boolean ops are already extended.
   608  (MOVBUreg x:((FLES|FLTS|FEQS|FNES) _ _)) => x
   609  (MOVBUreg x:((FLED|FLTD|FEQD|FNED) _ _)) => x
   610  (MOVBUreg x:((SEQZ|SNEZ) _)) => x
   611  (MOVBUreg x:((SLT|SLTU) _ _)) => x
   612  
   613  // Avoid extending when already sufficiently masked.
   614  (MOVBreg  x:(ANDI [c] y)) && c >= 0 && int64(int8(c)) == c => x
   615  (MOVHreg  x:(ANDI [c] y)) && c >= 0 && int64(int16(c)) == c => x
   616  (MOVWreg  x:(ANDI [c] y)) && c >= 0 && int64(int32(c)) == c => x
   617  (MOVBUreg x:(ANDI [c] y)) && c >= 0 && int64(uint8(c)) == c => x
   618  (MOVHUreg x:(ANDI [c] y)) && c >= 0 && int64(uint16(c)) == c => x
   619  (MOVWUreg x:(ANDI [c] y)) && c >= 0 && int64(uint32(c)) == c => x
   620  
   621  // Combine masking and zero extension.
   622  (MOVBUreg (ANDI [c] x)) && c < 0 => (ANDI [int64(uint8(c))] x)
   623  (MOVHUreg (ANDI [c] x)) && c < 0 => (ANDI [int64(uint16(c))] x)
   624  (MOVWUreg (ANDI [c] x)) && c < 0 => (AND (MOVDconst [int64(uint32(c))]) x)
   625  
   626  // Avoid sign/zero extension for consts.
   627  (MOVBreg  (MOVDconst [c])) => (MOVDconst [int64(int8(c))])
   628  (MOVHreg  (MOVDconst [c])) => (MOVDconst [int64(int16(c))])
   629  (MOVWreg  (MOVDconst [c])) => (MOVDconst [int64(int32(c))])
   630  (MOVBUreg (MOVDconst [c])) => (MOVDconst [int64(uint8(c))])
   631  (MOVHUreg (MOVDconst [c])) => (MOVDconst [int64(uint16(c))])
   632  (MOVWUreg (MOVDconst [c])) => (MOVDconst [int64(uint32(c))])
   633  
   634  // Avoid sign/zero extension after properly typed load.
   635  (MOVBreg  x:(MOVBload  _ _)) => (MOVDreg x)
   636  (MOVHreg  x:(MOVBload  _ _)) => (MOVDreg x)
   637  (MOVHreg  x:(MOVBUload _ _)) => (MOVDreg x)
   638  (MOVHreg  x:(MOVHload  _ _)) => (MOVDreg x)
   639  (MOVWreg  x:(MOVBload  _ _)) => (MOVDreg x)
   640  (MOVWreg  x:(MOVBUload _ _)) => (MOVDreg x)
   641  (MOVWreg  x:(MOVHload  _ _)) => (MOVDreg x)
   642  (MOVWreg  x:(MOVHUload _ _)) => (MOVDreg x)
   643  (MOVWreg  x:(MOVWload  _ _)) => (MOVDreg x)
   644  (MOVBUreg x:(MOVBUload _ _)) => (MOVDreg x)
   645  (MOVHUreg x:(MOVBUload _ _)) => (MOVDreg x)
   646  (MOVHUreg x:(MOVHUload _ _)) => (MOVDreg x)
   647  (MOVWUreg x:(MOVBUload _ _)) => (MOVDreg x)
   648  (MOVWUreg x:(MOVHUload _ _)) => (MOVDreg x)
   649  (MOVWUreg x:(MOVWUload _ _)) => (MOVDreg x)
   650  
   651  // Avoid zero extension after properly typed atomic operation.
   652  (MOVBUreg x:(Select0 (LoweredAtomicLoad8 _ _))) => (MOVDreg x)
   653  (MOVBUreg x:(Select0 (LoweredAtomicCas32 _ _ _ _))) => (MOVDreg x)
   654  (MOVBUreg x:(Select0 (LoweredAtomicCas64 _ _ _ _))) => (MOVDreg x)
   655  
   656  // Avoid sign extension after word arithmetic.
   657  (MOVWreg x:(ADDIW   _)) => (MOVDreg x)
   658  (MOVWreg x:(SUBW  _ _)) => (MOVDreg x)
   659  (MOVWreg x:(NEGW    _)) => (MOVDreg x)
   660  (MOVWreg x:(MULW  _ _)) => (MOVDreg x)
   661  (MOVWreg x:(DIVW  _ _)) => (MOVDreg x)
   662  (MOVWreg x:(DIVUW _ _)) => (MOVDreg x)
   663  (MOVWreg x:(REMW  _ _)) => (MOVDreg x)
   664  (MOVWreg x:(REMUW _ _)) => (MOVDreg x)
   665  
   666  // Fold double extensions.
   667  (MOVBreg  x:(MOVBreg  _)) => (MOVDreg x)
   668  (MOVHreg  x:(MOVBreg  _)) => (MOVDreg x)
   669  (MOVHreg  x:(MOVBUreg _)) => (MOVDreg x)
   670  (MOVHreg  x:(MOVHreg  _)) => (MOVDreg x)
   671  (MOVWreg  x:(MOVBreg  _)) => (MOVDreg x)
   672  (MOVWreg  x:(MOVBUreg _)) => (MOVDreg x)
   673  (MOVWreg  x:(MOVHreg  _)) => (MOVDreg x)
   674  (MOVWreg  x:(MOVWreg  _)) => (MOVDreg x)
   675  (MOVBUreg x:(MOVBUreg _)) => (MOVDreg x)
   676  (MOVHUreg x:(MOVBUreg _)) => (MOVDreg x)
   677  (MOVHUreg x:(MOVHUreg _)) => (MOVDreg x)
   678  (MOVWUreg x:(MOVBUreg _)) => (MOVDreg x)
   679  (MOVWUreg x:(MOVHUreg _)) => (MOVDreg x)
   680  (MOVWUreg x:(MOVWUreg _)) => (MOVDreg x)
   681  
   682  // Do not extend before store.
   683  (MOVBstore [off] {sym} ptr (MOVBreg  x) mem) => (MOVBstore [off] {sym} ptr x mem)
   684  (MOVBstore [off] {sym} ptr (MOVHreg  x) mem) => (MOVBstore [off] {sym} ptr x mem)
   685  (MOVBstore [off] {sym} ptr (MOVWreg  x) mem) => (MOVBstore [off] {sym} ptr x mem)
   686  (MOVBstore [off] {sym} ptr (MOVBUreg x) mem) => (MOVBstore [off] {sym} ptr x mem)
   687  (MOVBstore [off] {sym} ptr (MOVHUreg x) mem) => (MOVBstore [off] {sym} ptr x mem)
   688  (MOVBstore [off] {sym} ptr (MOVWUreg x) mem) => (MOVBstore [off] {sym} ptr x mem)
   689  (MOVHstore [off] {sym} ptr (MOVHreg  x) mem) => (MOVHstore [off] {sym} ptr x mem)
   690  (MOVHstore [off] {sym} ptr (MOVWreg  x) mem) => (MOVHstore [off] {sym} ptr x mem)
   691  (MOVHstore [off] {sym} ptr (MOVHUreg x) mem) => (MOVHstore [off] {sym} ptr x mem)
   692  (MOVHstore [off] {sym} ptr (MOVWUreg x) mem) => (MOVHstore [off] {sym} ptr x mem)
   693  (MOVWstore [off] {sym} ptr (MOVWreg  x) mem) => (MOVWstore [off] {sym} ptr x mem)
   694  (MOVWstore [off] {sym} ptr (MOVWUreg x) mem) => (MOVWstore [off] {sym} ptr x mem)
   695  
   696  // Replace extend after load with alternate load where possible.
   697  (MOVBreg  <t> x:(MOVBUload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVBload  <t> [off] {sym} ptr mem)
   698  (MOVHreg  <t> x:(MOVHUload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVHload  <t> [off] {sym} ptr mem)
   699  (MOVWreg  <t> x:(MOVWUload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVWload  <t> [off] {sym} ptr mem)
   700  (MOVBUreg <t> x:(MOVBload  [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVBUload <t> [off] {sym} ptr mem)
   701  (MOVHUreg <t> x:(MOVHload  [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVHUload <t> [off] {sym} ptr mem)
   702  (MOVWUreg <t> x:(MOVWload  [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVWUload <t> [off] {sym} ptr mem)
   703  
   704  // If a register move has only 1 use, just use the same register without emitting instruction
   705  // MOVnop does not emit an instruction, only for ensuring the type.
   706  (MOVDreg x) && x.Uses == 1 => (MOVDnop x)
   707  
   708  // TODO: we should be able to get rid of MOVDnop all together.
   709  // But for now, this is enough to get rid of lots of them.
   710  (MOVDnop (MOVDconst [c])) => (MOVDconst [c])
   711  
   712  // Avoid unnecessary zero and sign extension when right shifting.
   713  (SRAI <t> [x] (MOVWreg  y)) && x >= 0 && x <= 31 => (SRAIW <t> [int64(x)] y)
   714  (SRLI <t> [x] (MOVWUreg y)) && x >= 0 && x <= 31 => (SRLIW <t> [int64(x)] y)
   715  
   716  // Replace right shifts that exceed size of signed type.
   717  (SRAI <t> [x] (MOVBreg y)) && x >=  8 => (SRAI  [63] (SLLI <t> [56] y))
   718  (SRAI <t> [x] (MOVHreg y)) && x >= 16 => (SRAI  [63] (SLLI <t> [48] y))
   719  (SRAI <t> [x] (MOVWreg y)) && x >= 32 => (SRAIW [31] y)
   720  
   721  // Eliminate right shifts that exceed size of unsigned type.
   722  (SRLI <t> [x] (MOVBUreg y)) && x >=  8 => (MOVDconst <t> [0])
   723  (SRLI <t> [x] (MOVHUreg y)) && x >= 16 => (MOVDconst <t> [0])
   724  (SRLI <t> [x] (MOVWUreg y)) && x >= 32 => (MOVDconst <t> [0])
   725  
   726  // Fold constant into immediate instructions where possible.
   727  (ADD (MOVDconst <t> [val]) x) && is32Bit(val) && !t.IsPtr() => (ADDI [val] x)
   728  (AND (MOVDconst [val]) x) && is32Bit(val) => (ANDI [val] x)
   729  (OR  (MOVDconst [val]) x) && is32Bit(val) => (ORI  [val] x)
   730  (XOR (MOVDconst [val]) x) && is32Bit(val) => (XORI [val] x)
   731  (SLL  x (MOVDconst [val])) => (SLLI [int64(val&63)] x)
   732  (SRL  x (MOVDconst [val])) => (SRLI [int64(val&63)] x)
   733  (SRLW x (MOVDconst [val])) => (SRLIW [int64(val&31)] x)
   734  (SRA  x (MOVDconst [val])) => (SRAI [int64(val&63)] x)
   735  (SRAW x (MOVDconst [val])) => (SRAIW [int64(val&31)] x)
   736  (SLT  x (MOVDconst [val])) && val >= -2048 && val <= 2047 => (SLTI  [val] x)
   737  (SLTU x (MOVDconst [val])) && val >= -2048 && val <= 2047 => (SLTIU [val] x)
   738  
   739  // Convert const subtraction into ADDI with negative immediate, where possible.
   740  (SUB x (MOVDconst [val])) && is32Bit(-val) => (ADDI [-val] x)
   741  (SUB <t> (MOVDconst [val]) y) && is32Bit(-val) => (NEG (ADDI <t> [-val] y))
   742  
   743  // Subtraction of zero.
   744  (SUB  x (MOVDconst [0])) => x
   745  (SUBW x (MOVDconst [0])) => (ADDIW [0] x)
   746  
   747  // Subtraction from zero.
   748  (SUB  (MOVDconst [0]) x) => (NEG x)
   749  (SUBW (MOVDconst [0]) x) => (NEGW x)
   750  
   751  // Fold negation into subtraction.
   752  (NEG (SUB x y)) => (SUB y x)
   753  (NEG <t> s:(ADDI [val] (SUB x y))) && s.Uses == 1 && is32Bit(-val) => (ADDI [-val] (SUB <t> y x))
   754  
   755  // Double negation.
   756  (NEG (NEG x)) => x
   757  
   758  // Addition of zero or two constants.
   759  (ADDI [0] x) => x
   760  (ADDI [x] (MOVDconst [y])) && is32Bit(x + y) => (MOVDconst [x + y])
   761  
   762  // ANDI with all zeros, all ones or two constants.
   763  (ANDI [0]  x) => (MOVDconst [0])
   764  (ANDI [-1] x) => x
   765  (ANDI [x] (MOVDconst [y])) => (MOVDconst [x & y])
   766  
   767  // ORI with all zeroes, all ones or two constants.
   768  (ORI [0]  x) => x
   769  (ORI [-1] x) => (MOVDconst [-1])
   770  (ORI [x] (MOVDconst [y])) => (MOVDconst [x | y])
   771  
   772  // Combine operations with immediate.
   773  (ADDI [x] (ADDI [y] z)) && is32Bit(x + y) => (ADDI [x + y] z)
   774  (ANDI [x] (ANDI [y] z)) => (ANDI [x & y] z)
   775  (ORI  [x] (ORI  [y] z)) => (ORI  [x | y] z)
   776  
   777  // Negation of a constant.
   778  (NEG  (MOVDconst [x])) => (MOVDconst [-x])
   779  (NEGW (MOVDconst [x])) => (MOVDconst [int64(int32(-x))])
   780  
   781  // Shift of a constant.
   782  (SLLI [x] (MOVDconst [y])) && is32Bit(y << uint32(x)) => (MOVDconst [y << uint32(x)])
   783  (SRLI [x] (MOVDconst [y])) => (MOVDconst [int64(uint64(y) >> uint32(x))])
   784  (SRAI [x] (MOVDconst [y])) => (MOVDconst [int64(y) >> uint32(x)])
   785  
   786  // SLTI/SLTIU with constants.
   787  (SLTI  [x] (MOVDconst [y])) => (MOVDconst [b2i(int64(y) < int64(x))])
   788  (SLTIU [x] (MOVDconst [y])) => (MOVDconst [b2i(uint64(y) < uint64(x))])
   789  
   790  // SLTI/SLTIU with known outcomes.
   791  (SLTI  [x] (ANDI [y] _)) && y >= 0 && int64(y) < int64(x) => (MOVDconst [1])
   792  (SLTIU [x] (ANDI [y] _)) && y >= 0 && uint64(y) < uint64(x) => (MOVDconst [1])
   793  (SLTI  [x] (ORI  [y] _)) && y >= 0 && int64(y) >= int64(x) => (MOVDconst [0])
   794  (SLTIU [x] (ORI  [y] _)) && y >= 0 && uint64(y) >= uint64(x) => (MOVDconst [0])
   795  
   796  // SLT/SLTU with known outcomes.
   797  (SLT  x x) => (MOVDconst [0])
   798  (SLTU x x) => (MOVDconst [0])
   799  
   800  // Deadcode for LoweredMuluhilo
   801  (Select0 m:(LoweredMuluhilo x y)) && m.Uses == 1 => (MULHU x y)
   802  (Select1 m:(LoweredMuluhilo x y)) && m.Uses == 1 => (MUL x y)
   803  
   804  (FADD(S|D) a (FMUL(S|D) x y)) && a.Block.Func.useFMA(v) => (FMADD(S|D) x y a)
   805  (FSUB(S|D) a (FMUL(S|D) x y)) && a.Block.Func.useFMA(v) => (FNMSUB(S|D) x y a)
   806  (FSUB(S|D) (FMUL(S|D) x y) a) && a.Block.Func.useFMA(v) => (FMSUB(S|D) x y a)
   807  
   808  // Merge negation into fused multiply-add and multiply-subtract.
   809  //
   810  // Key:
   811  //
   812  //   [+ -](x * y [+ -] z).
   813  //    _ N         A S
   814  //                D U
   815  //                D B
   816  //
   817  // Note: multiplication commutativity handled by rule generator.
   818  (F(MADD|NMADD|MSUB|NMSUB)S neg:(FNEGS x) y z) && neg.Uses == 1 => (F(NMSUB|MSUB|NMADD|MADD)S x y z)
   819  (F(MADD|NMADD|MSUB|NMSUB)S x y neg:(FNEGS z)) && neg.Uses == 1 => (F(MSUB|NMSUB|MADD|NMADD)S x y z)
   820  (F(MADD|NMADD|MSUB|NMSUB)D neg:(FNEGD x) y z) && neg.Uses == 1 => (F(NMSUB|MSUB|NMADD|MADD)D x y z)
   821  (F(MADD|NMADD|MSUB|NMSUB)D x y neg:(FNEGD z)) && neg.Uses == 1 => (F(MSUB|NMSUB|MADD|NMADD)D x y z)