...

Text file src/cmd/compile/internal/ssa/_gen/PPC64.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// GOPPC64 values indicate power8, power9, etc.
     6// That means the code is compiled for that target,
     7// and will not run on earlier targets.
     8//
     9(Add(Ptr|64|32|16|8) ...) => (ADD ...)
    10(Add64F ...) => (FADD ...)
    11(Add32F ...) => (FADDS ...)
    12
    13(Sub(Ptr|64|32|16|8) ...) => (SUB ...)
    14(Sub32F ...) => (FSUBS ...)
    15(Sub64F ...) => (FSUB ...)
    16
    17(Min(32|64)F x y) && buildcfg.GOPPC64 >= 9 => (XSMINJDP x y)
    18(Max(32|64)F x y) && buildcfg.GOPPC64 >= 9 => (XSMAXJDP x y)
    19
    20// Combine 64 bit integer multiply and adds
    21(ADD l:(MULLD x y) z) && buildcfg.GOPPC64 >= 9 && l.Uses == 1 && clobber(l) => (MADDLD x y z)
    22
    23(Mod16 x y) => (Mod32 (SignExt16to32 x) (SignExt16to32 y))
    24(Mod16u x y) => (Mod32u (ZeroExt16to32 x) (ZeroExt16to32 y))
    25(Mod8 x y) => (Mod32 (SignExt8to32 x) (SignExt8to32 y))
    26(Mod8u x y) => (Mod32u (ZeroExt8to32 x) (ZeroExt8to32 y))
    27(Mod64 x y) && buildcfg.GOPPC64 >=9 => (MODSD x y)
    28(Mod64 x y) && buildcfg.GOPPC64 <=8 => (SUB x (MULLD y (DIVD x y)))
    29(Mod64u x y) && buildcfg.GOPPC64 >= 9 => (MODUD x y)
    30(Mod64u x y) && buildcfg.GOPPC64 <= 8 => (SUB x (MULLD y (DIVDU x y)))
    31(Mod32 x y) && buildcfg.GOPPC64 >= 9 => (MODSW x y)
    32(Mod32 x y) && buildcfg.GOPPC64 <= 8 => (SUB x (MULLW y (DIVW x y)))
    33(Mod32u x y) && buildcfg.GOPPC64 >= 9 => (MODUW x y)
    34(Mod32u x y) && buildcfg.GOPPC64 <= 8 => (SUB x (MULLW y (DIVWU x y)))
    35
    36// (x + y) / 2 with x>=y => (x - y) / 2 + y
    37(Avg64u <t> x y) => (ADD (SRDconst <t> (SUB <t> x y) [1]) y)
    38
    39(Mul64 ...) => (MULLD ...)
    40(Mul(32|16|8) ...) => (MULLW ...)
    41(Select0 (Mul64uhilo x y)) => (MULHDU x y)
    42(Select1 (Mul64uhilo x y)) => (MULLD x y)
    43(Select0 (Mul64uover x y)) => (MULLD x y)
    44(Select1 (Mul64uover x y)) => (SETBCR [2] (CMPconst [0] (MULHDU <x.Type> x y)))
    45
    46(Div64 [false] x y) => (DIVD x y)
    47(Div64u ...) => (DIVDU ...)
    48(Div32 [false] x y) => (DIVW x y)
    49(Div32u ...) => (DIVWU ...)
    50(Div16 [false]  x y) => (DIVW  (SignExt16to32 x) (SignExt16to32 y))
    51(Div16u x y) => (DIVWU (ZeroExt16to32 x) (ZeroExt16to32 y))
    52(Div8 x y) => (DIVW  (SignExt8to32 x) (SignExt8to32 y))
    53(Div8u x y) => (DIVWU (ZeroExt8to32 x) (ZeroExt8to32 y))
    54
    55(Hmul(64|64u|32|32u) ...) => (MULH(D|DU|W|WU) ...)
    56
    57(Mul(32|64)F ...) => ((FMULS|FMUL) ...)
    58
    59(Div(32|64)F ...) => ((FDIVS|FDIV) ...)
    60
    61// Lowering float <=> int
    62(Cvt32to(32|64)F x) => ((FCFIDS|FCFID) (MTVSRD (SignExt32to64 x)))
    63(Cvt64to(32|64)F x) => ((FCFIDS|FCFID) (MTVSRD x))
    64
    65(Cvt32Fto(32|64) x) => (MFVSRD (FCTI(W|D)Z x))
    66(Cvt64Fto(32|64) x) => (MFVSRD (FCTI(W|D)Z x))
    67
    68(Cvt32Fto64F ...) => (Copy ...) // Note v will have the wrong type for patterns dependent on Float32/Float64
    69(Cvt64Fto32F ...) => (FRSP ...)
    70
    71(CvtBoolToUint8 ...) => (Copy ...)
    72
    73(Round(32|64)F ...) => (LoweredRound(32|64)F ...)
    74
    75(Sqrt ...) => (FSQRT ...)
    76(Sqrt32 ...) => (FSQRTS ...)
    77(Floor ...) => (FFLOOR ...)
    78(Ceil ...) => (FCEIL ...)
    79(Trunc ...) => (FTRUNC ...)
    80(Round ...) => (FROUND ...)
    81(Copysign x y) => (FCPSGN y x)
    82(Abs ...) => (FABS ...)
    83(FMA ...) => (FMADD ...)
    84
    85// Lowering extension
    86// Note: we always extend to 64 bits even though some ops don't need that many result bits.
    87(SignExt8to(16|32|64) ...) => (MOVBreg ...)
    88(SignExt16to(32|64) ...) => (MOVHreg ...)
    89(SignExt32to64 ...) => (MOVWreg ...)
    90
    91(ZeroExt8to(16|32|64) ...) => (MOVBZreg ...)
    92(ZeroExt16to(32|64) ...) => (MOVHZreg ...)
    93(ZeroExt32to64 ...) => (MOVWZreg ...)
    94
    95(Trunc(16|32|64)to8 <t> x) && t.IsSigned() => (MOVBreg x)
    96(Trunc(16|32|64)to8  x) => (MOVBZreg x)
    97(Trunc(32|64)to16 <t> x) && t.IsSigned() => (MOVHreg x)
    98(Trunc(32|64)to16 x) => (MOVHZreg x)
    99(Trunc64to32 <t> x) && t.IsSigned() => (MOVWreg x)
   100(Trunc64to32 x) => (MOVWZreg x)
   101
   102// Lowering constants
   103(Const(64|32|16|8) [val]) => (MOVDconst [int64(val)])
   104(Const(32|64)F ...) => (FMOV(S|D)const ...)
   105(ConstNil) => (MOVDconst [0])
   106(ConstBool [t]) => (MOVDconst [b2i(t)])
   107
   108// Carrying addition.
   109(Select0 (Add64carry x y c)) =>            (Select0 <typ.UInt64> (ADDE x y (Select1 <typ.UInt64> (ADDCconst c [-1]))))
   110(Select1 (Add64carry x y c)) => (ADDZEzero (Select1 <typ.UInt64> (ADDE x y (Select1 <typ.UInt64> (ADDCconst c [-1])))))
   111// Fold initial carry bit if 0.
   112(ADDE x y (Select1 <typ.UInt64> (ADDCconst (MOVDconst [0]) [-1]))) => (ADDC x y)
   113// Fold transfer of CA -> GPR -> CA. Note 2 uses when feeding into a chained Add64carry.
   114(Select1 (ADDCconst n:(ADDZEzero x) [-1])) && n.Uses <= 2 => x
   115(ADDE (MOVDconst [0]) y c) => (ADDZE y c)
   116(ADDC x (MOVDconst [y])) && is16Bit(y) => (ADDCconst [y] x)
   117
   118// Borrowing subtraction.
   119(Select0 (Sub64borrow x y c)) =>                 (Select0 <typ.UInt64> (SUBE x y (Select1 <typ.UInt64> (SUBCconst c [0]))))
   120(Select1 (Sub64borrow x y c)) => (NEG (SUBZEzero (Select1 <typ.UInt64> (SUBE x y (Select1 <typ.UInt64> (SUBCconst c [0]))))))
   121// Fold initial borrow bit if 0.
   122(SUBE x y (Select1 <typ.UInt64> (SUBCconst (MOVDconst [0]) [0]))) => (SUBC x y)
   123// Fold transfer of CA -> GPR -> CA. Note 2 uses when feeding into a chained Sub64borrow.
   124(Select1 (SUBCconst n:(NEG (SUBZEzero x)) [0])) && n.Uses <= 2 => x
   125
   126// Constant folding
   127(FABS (FMOVDconst [x])) => (FMOVDconst [math.Abs(x)])
   128(FSQRT (FMOVDconst [x])) && x >= 0 => (FMOVDconst [math.Sqrt(x)])
   129(FFLOOR (FMOVDconst [x])) => (FMOVDconst [math.Floor(x)])
   130(FCEIL (FMOVDconst [x])) => (FMOVDconst [math.Ceil(x)])
   131(FTRUNC (FMOVDconst [x])) => (FMOVDconst [math.Trunc(x)])
   132
   133// Rotates
   134(RotateLeft8 <t> x (MOVDconst [c])) => (Or8 (Lsh8x64 <t> x (MOVDconst [c&7])) (Rsh8Ux64 <t> x (MOVDconst [-c&7])))
   135(RotateLeft16 <t> x (MOVDconst [c])) => (Or16 (Lsh16x64 <t> x (MOVDconst [c&15])) (Rsh16Ux64 <t> x (MOVDconst [-c&15])))
   136(RotateLeft(32|64) ...) => ((ROTLW|ROTL) ...)
   137
   138// Constant rotate generation
   139(ROTLW  x (MOVDconst [c])) => (ROTLWconst  x [c&31])
   140(ROTL   x (MOVDconst [c])) => (ROTLconst   x [c&63])
   141
   142// Combine rotate and mask operations
   143(ANDconst [m] (ROTLWconst [r] x)) && isPPC64WordRotateMask(m) => (RLWINM [encodePPC64RotateMask(r,m,32)] x)
   144(AND (MOVDconst [m]) (ROTLWconst [r] x)) && isPPC64WordRotateMask(m) => (RLWINM [encodePPC64RotateMask(r,m,32)] x)
   145(ANDconst [m] (ROTLW x r)) && isPPC64WordRotateMask(m) => (RLWNM [encodePPC64RotateMask(0,m,32)] x r)
   146(AND (MOVDconst [m]) (ROTLW x r)) && isPPC64WordRotateMask(m) => (RLWNM [encodePPC64RotateMask(0,m,32)] x r)
   147
   148// Note, any rotated word bitmask is still a valid word bitmask.
   149(ROTLWconst [r] (AND (MOVDconst [m]) x)) && isPPC64WordRotateMask(m) => (RLWINM [encodePPC64RotateMask(r,rotateLeft32(m,r),32)] x)
   150(ROTLWconst [r] (ANDconst [m] x)) && isPPC64WordRotateMask(m) => (RLWINM [encodePPC64RotateMask(r,rotateLeft32(m,r),32)] x)
   151
   152(ANDconst [m] (SRWconst x [s])) && mergePPC64RShiftMask(m,s,32) == 0 => (MOVDconst [0])
   153(ANDconst [m] (SRWconst x [s])) && mergePPC64AndSrwi(m,s) != 0 => (RLWINM [mergePPC64AndSrwi(m,s)] x)
   154(ANDconst [m] (SRDconst x [s])) && mergePPC64AndSrdi(m,s) != 0 => (RLWINM [mergePPC64AndSrdi(m,s)] x)
   155(AND (MOVDconst [m]) (SRWconst x [s])) && mergePPC64RShiftMask(m,s,32) == 0 => (MOVDconst [0])
   156(AND (MOVDconst [m]) (SRWconst x [s])) && mergePPC64AndSrwi(m,s) != 0 => (RLWINM [mergePPC64AndSrwi(m,s)] x)
   157(AND (MOVDconst [m]) (SRDconst x [s])) && mergePPC64AndSrdi(m,s) != 0 => (RLWINM [mergePPC64AndSrdi(m,s)] x)
   158
   159(SRWconst (ANDconst [m] x) [s]) && mergePPC64RShiftMask(m>>uint(s),s,32) == 0 => (MOVDconst [0])
   160(SRWconst (ANDconst [m] x) [s]) && mergePPC64AndSrwi(m>>uint(s),s) != 0 => (RLWINM [mergePPC64AndSrwi(m>>uint(s),s)] x)
   161(SRWconst (AND (MOVDconst [m]) x) [s]) && mergePPC64RShiftMask(m>>uint(s),s,32) == 0 => (MOVDconst [0])
   162(SRWconst (AND (MOVDconst [m]) x) [s]) && mergePPC64AndSrwi(m>>uint(s),s) != 0 => (RLWINM [mergePPC64AndSrwi(m>>uint(s),s)] x)
   163
   164(ANDconst [m] (SLDconst x [s])) && mergePPC64AndSldi(m,s) != 0 => (RLWINM [mergePPC64AndSldi(m,s)] x)
   165(AND (MOVDconst [m]) (SLDconst x [s])) && mergePPC64AndSldi(m,s) != 0 => (RLWINM [mergePPC64AndSldi(m,s)] x)
   166
   167// Merge shift right + shift left and clear left (e.g for a table lookup)
   168(CLRLSLDI [c] (SRWconst [s] x)) && mergePPC64ClrlsldiSrw(int64(c),s) != 0 => (RLWINM [mergePPC64ClrlsldiSrw(int64(c),s)] x)
   169(CLRLSLDI [c] (SRDconst [s] x)) && mergePPC64ClrlsldiSrd(int64(c),s) != 0 => (RLWINM [mergePPC64ClrlsldiSrd(int64(c),s)] x)
   170(SLDconst [l] (SRWconst [r] x)) && mergePPC64SldiSrw(l,r) != 0 => (RLWINM [mergePPC64SldiSrw(l,r)] x)
   171// The following reduction shows up frequently too. e.g b[(x>>14)&0xFF]
   172(CLRLSLDI [c] i:(RLWINM [s] x)) && mergePPC64ClrlsldiRlwinm(c,s) != 0 => (RLWINM [mergePPC64ClrlsldiRlwinm(c,s)] x)
   173
   174// large constant signed right shift, we leave the sign bit
   175(Rsh64x64 x (MOVDconst [c])) && uint64(c) >= 64 => (SRADconst x [63])
   176(Rsh32x64 x (MOVDconst [c])) && uint64(c) >= 32 => (SRAWconst x [63])
   177(Rsh16x64 x (MOVDconst [c])) && uint64(c) >= 16 => (SRAWconst (SignExt16to32 x) [63])
   178(Rsh8x64  x (MOVDconst [c])) && uint64(c) >= 8  => (SRAWconst (SignExt8to32  x) [63])
   179
   180// constant shifts
   181((Lsh64|Rsh64|Rsh64U)x64  x (MOVDconst [c])) && uint64(c) < 64 => (S(L|RA|R)Dconst x [c])
   182((Lsh32|Rsh32|Rsh32U)x64  x (MOVDconst [c])) && uint64(c) < 32 => (S(L|RA|R)Wconst x [c])
   183((Rsh16|Rsh16U)x64  x (MOVDconst [c])) && uint64(c) < 16 => (SR(AW|W)const ((Sign|Zero)Ext16to32 x) [c])
   184(Lsh16x64  x (MOVDconst [c])) && uint64(c) < 16 => (SLWconst x [c])
   185((Rsh8|Rsh8U)x64  x (MOVDconst [c])) && uint64(c) < 8 => (SR(AW|W)const ((Sign|Zero)Ext8to32 x) [c])
   186(Lsh8x64  x (MOVDconst [c])) && uint64(c) < 8 => (SLWconst x [c])
   187
   188// Lower bounded shifts first. No need to check shift value.
   189(Lsh64x(64|32|16|8)  x y) && shiftIsBounded(v) => (SLD x y)
   190(Lsh32x(64|32|16|8)  x y) && shiftIsBounded(v) => (SLW x y)
   191(Lsh16x(64|32|16|8)  x y) && shiftIsBounded(v) => (SLD x y)
   192(Lsh8x(64|32|16|8)   x y) && shiftIsBounded(v) => (SLD x y)
   193(Rsh64Ux(64|32|16|8) x y) && shiftIsBounded(v) => (SRD x y)
   194(Rsh32Ux(64|32|16|8) x y) && shiftIsBounded(v) => (SRW x y)
   195(Rsh16Ux(64|32|16|8) x y) && shiftIsBounded(v) => (SRD (MOVHZreg x) y)
   196(Rsh8Ux(64|32|16|8)  x y) && shiftIsBounded(v) => (SRD (MOVBZreg x) y)
   197(Rsh64x(64|32|16|8)  x y) && shiftIsBounded(v) => (SRAD x y)
   198(Rsh32x(64|32|16|8)  x y) && shiftIsBounded(v) => (SRAW x y)
   199(Rsh16x(64|32|16|8)  x y) && shiftIsBounded(v) => (SRAD (MOVHreg x) y)
   200(Rsh8x(64|32|16|8)   x y) && shiftIsBounded(v) => (SRAD (MOVBreg x) y)
   201
   202// Unbounded shifts. Go shifts saturate to 0 or -1 when shifting beyond the number of
   203// bits in a type, PPC64 shifts do not (see the ISA for details).
   204//
   205// Note, y is always non-negative.
   206//
   207// Note, ISELZ is intentionally not used in lower. Where possible, ISEL is converted to ISELZ in late lower
   208// after all the ISEL folding rules have been exercised.
   209
   210((Rsh64U|Lsh64)x64 <t> x y)  => (ISEL [0] (S(R|L)D <t> x y) (MOVDconst [0])        (CMPUconst y [64]))
   211((Rsh64U|Lsh64)x32 <t> x y)  => (ISEL [0] (S(R|L)D <t> x y) (MOVDconst [0])        (CMPWUconst y [64]))
   212((Rsh64U|Lsh64)x16 <t> x y)  => (ISEL [2] (S(R|L)D <t> x y) (MOVDconst [0])        (CMPconst [0] (ANDconst [0xFFC0] y)))
   213((Rsh64U|Lsh64)x8  <t> x y)  => (ISEL [2] (S(R|L)D <t> x y) (MOVDconst [0])        (CMPconst [0] (ANDconst [0x00C0] y)))
   214(Rsh64x(64|32)     <t> x y)  => (ISEL [0] (SRAD    <t> x y) (SRADconst <t> x [63]) (CMP(U|WU)const y [64]))
   215(Rsh64x16          <t> x y)  => (ISEL [2] (SRAD    <t> x y) (SRADconst <t> x [63]) (CMPconst [0] (ANDconst [0xFFC0] y)))
   216(Rsh64x8           <t> x y)  => (ISEL [2] (SRAD    <t> x y) (SRADconst <t> x [63]) (CMPconst [0] (ANDconst [0x00C0] y)))
   217
   218((Rsh32U|Lsh32)x64 <t> x y)  => (ISEL [0] (S(R|L)W <t> x y) (MOVDconst [0])        (CMPUconst y [32]))
   219((Rsh32U|Lsh32)x32 <t> x y)  => (ISEL [0] (S(R|L)W <t> x y) (MOVDconst [0])        (CMPWUconst y [32]))
   220((Rsh32U|Lsh32)x16 <t> x y)  => (ISEL [2] (S(R|L)W <t> x y) (MOVDconst [0])        (CMPconst [0] (ANDconst [0xFFE0] y)))
   221((Rsh32U|Lsh32)x8  <t> x y)  => (ISEL [2] (S(R|L)W <t> x y) (MOVDconst [0])        (CMPconst [0] (ANDconst [0x00E0] y)))
   222(Rsh32x(64|32)     <t> x y)  => (ISEL [0] (SRAW    <t> x y) (SRAWconst <t> x [31]) (CMP(U|WU)const y [32]))
   223(Rsh32x16          <t> x y)  => (ISEL [2] (SRAW    <t> x y) (SRAWconst <t> x [31]) (CMPconst [0] (ANDconst [0xFFE0] y)))
   224(Rsh32x8           <t> x y)  => (ISEL [2] (SRAW    <t> x y) (SRAWconst <t> x [31]) (CMPconst [0] (ANDconst [0x00E0] y)))
   225
   226((Rsh16U|Lsh16)x64 <t> x y) => (ISEL [0] (S(R|L)D  <t> (MOVHZreg x) y) (MOVDconst [0])                   (CMPUconst  y [16]))
   227((Rsh16U|Lsh16)x32 <t> x y) => (ISEL [0] (S(R|L)D  <t> (MOVHZreg x) y) (MOVDconst [0])                   (CMPWUconst y [16]))
   228((Rsh16U|Lsh16)x16 <t> x y) => (ISEL [2] (S(R|L)D  <t> (MOVHZreg x) y) (MOVDconst [0])                   (CMPconst [0] (ANDconst [0xFFF0] y)))
   229((Rsh16U|Lsh16)x8  <t> x y) => (ISEL [2] (S(R|L)D  <t> (MOVHZreg x) y) (MOVDconst [0])                   (CMPconst [0] (ANDconst [0x00F0] y)))
   230(Rsh16x(64|32)     <t> x y) => (ISEL [0] (SRAD     <t> (MOVHreg  x) y) (SRADconst <t>  (MOVHreg x) [15]) (CMP(U|WU)const  y [16]))
   231(Rsh16x16          <t> x y) => (ISEL [2] (SRAD     <t> (MOVHreg  x) y) (SRADconst <t>  (MOVHreg x) [15]) (CMPconst [0] (ANDconst [0xFFF0] y)))
   232(Rsh16x8           <t> x y) => (ISEL [2] (SRAD     <t> (MOVHreg  x) y) (SRADconst <t>  (MOVHreg x) [15]) (CMPconst [0] (ANDconst [0x00F0] y)))
   233
   234((Rsh8U|Lsh8)x64 <t> x y) => (ISEL [0] (S(R|L)D  <t> (MOVBZreg x) y) (MOVDconst [0])                  (CMPUconst  y [8]))
   235((Rsh8U|Lsh8)x32 <t> x y) => (ISEL [0] (S(R|L)D  <t> (MOVBZreg x) y) (MOVDconst [0])                  (CMPWUconst y [8]))
   236((Rsh8U|Lsh8)x16 <t> x y) => (ISEL [2] (S(R|L)D  <t> (MOVBZreg x) y) (MOVDconst [0])                  (CMPconst [0] (ANDconst [0xFFF8] y)))
   237((Rsh8U|Lsh8)x8  <t> x y) => (ISEL [2] (S(R|L)D  <t> (MOVBZreg x) y) (MOVDconst [0])                  (CMPconst [0] (ANDconst [0x00F8] y)))
   238(Rsh8x(64|32)    <t> x y) => (ISEL [0] (SRAD     <t> (MOVBreg  x) y) (SRADconst <t>  (MOVBreg x) [7]) (CMP(U|WU)const  y [8]))
   239(Rsh8x16         <t> x y) => (ISEL [2] (SRAD     <t> (MOVBreg  x) y) (SRADconst <t>  (MOVBreg x) [7]) (CMPconst [0] (ANDconst [0xFFF8] y)))
   240(Rsh8x8          <t> x y) => (ISEL [2] (SRAD     <t> (MOVBreg  x) y) (SRADconst <t>  (MOVBreg x) [7]) (CMPconst [0] (ANDconst [0x00F8] y)))
   241
   242// Catch bounded shifts in situations like foo<<uint(shift&63) which might not be caught by the prove pass.
   243(CMP(U|WU)const [d] (ANDconst z [c])) && uint64(d) > uint64(c) => (FlagLT)
   244
   245(ORN x (MOVDconst [-1])) => x
   246
   247(S(RAD|RD|LD) x (MOVDconst [c])) => (S(RAD|RD|LD)const [c&63 | (c>>6&1*63)] x)
   248(S(RAW|RW|LW) x (MOVDconst [c])) => (S(RAW|RW|LW)const [c&31 | (c>>5&1*31)] x)
   249
   250(Addr {sym} base) => (MOVDaddr {sym} [0] base)
   251(LocalAddr <t> {sym} base mem) && t.Elem().HasPointers() => (MOVDaddr {sym} (SPanchored base mem))
   252(LocalAddr <t> {sym} base _)  && !t.Elem().HasPointers() => (MOVDaddr {sym} base)
   253(OffPtr [off] ptr) => (ADD (MOVDconst <typ.Int64> [off]) ptr)
   254(MOVDaddr {sym} [n] p:(ADD x y)) && sym == nil && n == 0 => p
   255(MOVDaddr {sym} [n] ptr) && sym == nil && n == 0 && (ptr.Op == OpArgIntReg || ptr.Op == OpPhi) => ptr
   256
   257// TODO: optimize these cases?
   258(Ctz32NonZero ...) => (Ctz32 ...)
   259(Ctz64NonZero ...) => (Ctz64 ...)
   260
   261(Ctz64 x) && buildcfg.GOPPC64<=8 => (POPCNTD (ANDN <typ.Int64> (ADDconst <typ.Int64> [-1] x) x))
   262(Ctz64 x) => (CNTTZD x)
   263(Ctz32 x) && buildcfg.GOPPC64<=8 => (POPCNTW (MOVWZreg (ANDN <typ.Int> (ADDconst <typ.Int> [-1] x) x)))
   264(Ctz32 x) => (CNTTZW (MOVWZreg x))
   265(Ctz16 x) => (POPCNTW (MOVHZreg (ANDN <typ.Int16> (ADDconst <typ.Int16> [-1] x) x)))
   266(Ctz8 x)  => (POPCNTB (MOVBZreg (ANDN <typ.UInt8> (ADDconst <typ.UInt8> [-1] x) x)))
   267
   268(BitLen64 x) => (SUBFCconst [64] (CNTLZD <typ.Int> x))
   269(BitLen32 x) => (SUBFCconst [32] (CNTLZW <typ.Int> x))
   270
   271(PopCount64 ...) => (POPCNTD ...)
   272(PopCount(32|16|8) x) => (POPCNT(W|W|B) (MOV(W|H|B)Zreg x))
   273
   274(And(64|32|16|8) ...) => (AND ...)
   275(Or(64|32|16|8) ...) => (OR ...)
   276(Xor(64|32|16|8) ...) => (XOR ...)
   277
   278(Neg(64|32|16|8) ...) => (NEG ...)
   279(Neg(64|32)F ...) => (FNEG ...)
   280
   281(Com(64|32|16|8) x) => (NOR x x)
   282
   283// Lowering boolean ops
   284(AndB ...) => (AND ...)
   285(OrB ...) => (OR ...)
   286(Not x) => (XORconst [1] x)
   287
   288// Merge logical operations
   289(AND x (NOR y y)) => (ANDN x y)
   290(OR x (NOR y y)) => (ORN x y)
   291
   292// Lowering comparisons
   293(EqB x y)  => (ANDconst [1] (EQV x y))
   294// Sign extension dependence on operand sign sets up for sign/zero-extension elision later
   295(Eq(8|16) x y) && x.Type.IsSigned() && y.Type.IsSigned() => (Equal (CMPW (SignExt(8|16)to32 x) (SignExt(8|16)to32 y)))
   296(Eq(8|16) x y) => (Equal (CMPW (ZeroExt(8|16)to32 x) (ZeroExt(8|16)to32 y)))
   297(Eq(32|64|Ptr) x y) => (Equal ((CMPW|CMP|CMP) x y))
   298(Eq(32|64)F x y) => (Equal (FCMPU x y))
   299
   300(NeqB ...) => (XOR ...)
   301// Like Eq8 and Eq16, prefer sign extension likely to enable later elision.
   302(Neq(8|16) x y) && x.Type.IsSigned() && y.Type.IsSigned() => (NotEqual (CMPW (SignExt(8|16)to32 x) (SignExt(8|16)to32 y)))
   303(Neq(8|16) x y)  => (NotEqual (CMPW (ZeroExt(8|16)to32 x) (ZeroExt(8|16)to32 y)))
   304(Neq(32|64|Ptr) x y) => (NotEqual ((CMPW|CMP|CMP) x y))
   305(Neq(32|64)F x y) => (NotEqual (FCMPU x y))
   306
   307(Less(8|16) x y)  => (LessThan (CMPW (SignExt(8|16)to32 x) (SignExt(8|16)to32 y)))
   308(Less(32|64) x y) => (LessThan ((CMPW|CMP) x y))
   309(Less(32|64)F x y) => (FLessThan (FCMPU x y))
   310
   311(Less(8|16)U x y)  => (LessThan (CMPWU (ZeroExt(8|16)to32 x) (ZeroExt(8|16)to32 y)))
   312(Less(32|64)U x y) => (LessThan ((CMPWU|CMPU) x y))
   313
   314(Leq(8|16) x y)  => (LessEqual (CMPW (SignExt(8|16)to32 x) (SignExt(8|16)to32 y)))
   315(Leq(32|64) x y) => (LessEqual ((CMPW|CMP) x y))
   316(Leq(32|64)F x y) => (FLessEqual (FCMPU x y))
   317
   318(Leq(8|16)U x y)  => (LessEqual (CMPWU (ZeroExt(8|16)to32 x) (ZeroExt(8|16)to32 y)))
   319(Leq(32|64)U x y) => (LessEqual (CMP(WU|U) x y))
   320
   321// Absorb pseudo-ops into blocks.
   322(If (Equal cc) yes no) => (EQ cc yes no)
   323(If (NotEqual cc) yes no) => (NE cc yes no)
   324(If (LessThan cc) yes no) => (LT cc yes no)
   325(If (LessEqual cc) yes no) => (LE cc yes no)
   326(If (GreaterThan cc) yes no) => (GT cc yes no)
   327(If (GreaterEqual cc) yes no) => (GE cc yes no)
   328(If (FLessThan cc) yes no) => (FLT cc yes no)
   329(If (FLessEqual cc) yes no) => (FLE cc yes no)
   330(If (FGreaterThan cc) yes no) => (FGT cc yes no)
   331(If (FGreaterEqual cc) yes no) => (FGE cc yes no)
   332
   333(If cond yes no) => (NE (CMPconst [0] (ANDconst [1] cond)) yes no)
   334
   335// Absorb boolean tests into block
   336(NE (CMPconst [0] (ANDconst [1] ((Equal|NotEqual|LessThan|LessEqual|GreaterThan|GreaterEqual) cc))) yes no) => ((EQ|NE|LT|LE|GT|GE) cc yes no)
   337(NE (CMPconst [0] (ANDconst [1] ((FLessThan|FLessEqual|FGreaterThan|FGreaterEqual) cc))) yes no) => ((FLT|FLE|FGT|FGE) cc yes no)
   338
   339// absorb flag constants into branches
   340(EQ (FlagEQ) yes no) => (First yes no)
   341(EQ (FlagLT) yes no) => (First no yes)
   342(EQ (FlagGT) yes no) => (First no yes)
   343
   344(NE (FlagEQ) yes no) => (First no yes)
   345(NE (FlagLT) yes no) => (First yes no)
   346(NE (FlagGT) yes no) => (First yes no)
   347
   348(LT (FlagEQ) yes no) => (First no yes)
   349(LT (FlagLT) yes no) => (First yes no)
   350(LT (FlagGT) yes no) => (First no yes)
   351
   352(LE (FlagEQ) yes no) => (First yes no)
   353(LE (FlagLT) yes no) => (First yes no)
   354(LE (FlagGT) yes no) => (First no yes)
   355
   356(GT (FlagEQ) yes no) => (First no yes)
   357(GT (FlagLT) yes no) => (First no yes)
   358(GT (FlagGT) yes no) => (First yes no)
   359
   360(GE (FlagEQ) yes no) => (First yes no)
   361(GE (FlagLT) yes no) => (First no yes)
   362(GE (FlagGT) yes no) => (First yes no)
   363
   364// absorb InvertFlags into branches
   365(LT (InvertFlags cmp) yes no) => (GT cmp yes no)
   366(GT (InvertFlags cmp) yes no) => (LT cmp yes no)
   367(LE (InvertFlags cmp) yes no) => (GE cmp yes no)
   368(GE (InvertFlags cmp) yes no) => (LE cmp yes no)
   369(EQ (InvertFlags cmp) yes no) => (EQ cmp yes no)
   370(NE (InvertFlags cmp) yes no) => (NE cmp yes no)
   371
   372// constant comparisons
   373(CMPWconst (MOVDconst [x]) [y]) && int32(x)==int32(y) => (FlagEQ)
   374(CMPWconst (MOVDconst [x]) [y]) && int32(x)<int32(y)  => (FlagLT)
   375(CMPWconst (MOVDconst [x]) [y]) && int32(x)>int32(y)  => (FlagGT)
   376
   377(CMPconst (MOVDconst [x]) [y]) && x==y => (FlagEQ)
   378(CMPconst (MOVDconst [x]) [y]) && x<y  => (FlagLT)
   379(CMPconst (MOVDconst [x]) [y]) && x>y  => (FlagGT)
   380
   381(CMPWUconst (MOVDconst [x]) [y]) && int32(x)==int32(y)  => (FlagEQ)
   382(CMPWUconst (MOVDconst [x]) [y]) && uint32(x)<uint32(y) => (FlagLT)
   383(CMPWUconst (MOVDconst [x]) [y]) && uint32(x)>uint32(y) => (FlagGT)
   384
   385(CMPUconst (MOVDconst [x]) [y]) && x==y  => (FlagEQ)
   386(CMPUconst (MOVDconst [x]) [y]) && uint64(x)<uint64(y) => (FlagLT)
   387(CMPUconst (MOVDconst [x]) [y]) && uint64(x)>uint64(y) => (FlagGT)
   388
   389// absorb flag constants into boolean values
   390(Equal (FlagEQ)) => (MOVDconst [1])
   391(Equal (FlagLT)) => (MOVDconst [0])
   392(Equal (FlagGT)) => (MOVDconst [0])
   393
   394(NotEqual (FlagEQ)) => (MOVDconst [0])
   395(NotEqual (FlagLT)) => (MOVDconst [1])
   396(NotEqual (FlagGT)) => (MOVDconst [1])
   397
   398(LessThan (FlagEQ)) => (MOVDconst [0])
   399(LessThan (FlagLT)) => (MOVDconst [1])
   400(LessThan (FlagGT)) => (MOVDconst [0])
   401
   402(LessEqual (FlagEQ)) => (MOVDconst [1])
   403(LessEqual (FlagLT)) => (MOVDconst [1])
   404(LessEqual (FlagGT)) => (MOVDconst [0])
   405
   406(GreaterThan (FlagEQ)) => (MOVDconst [0])
   407(GreaterThan (FlagLT)) => (MOVDconst [0])
   408(GreaterThan (FlagGT)) => (MOVDconst [1])
   409
   410(GreaterEqual (FlagEQ)) => (MOVDconst [1])
   411(GreaterEqual (FlagLT)) => (MOVDconst [0])
   412(GreaterEqual (FlagGT)) => (MOVDconst [1])
   413
   414// absorb InvertFlags into boolean values
   415((Equal|NotEqual|LessThan|GreaterThan|LessEqual|GreaterEqual) (InvertFlags x)) => ((Equal|NotEqual|GreaterThan|LessThan|GreaterEqual|LessEqual) x)
   416
   417
   418// Elide compares of bit tests
   419((EQ|NE|LT|LE|GT|GE) (CMPconst [0] z:(AND x y)) yes no) && z.Uses == 1 => ((EQ|NE|LT|LE|GT|GE) (Select1 <types.TypeFlags> (ANDCC x y)) yes no)
   420((EQ|NE|LT|LE|GT|GE) (CMPconst [0] z:(OR x y)) yes no) && z.Uses == 1 => ((EQ|NE|LT|LE|GT|GE) (Select1 <types.TypeFlags> (ORCC x y)) yes no)
   421((EQ|NE|LT|LE|GT|GE) (CMPconst [0] z:(XOR x y)) yes no) && z.Uses == 1 => ((EQ|NE|LT|LE|GT|GE) (Select1 <types.TypeFlags> (XORCC x y)) yes no)
   422
   423(CondSelect x y (SETBC [a] cmp))  => (ISEL [a] x y cmp)
   424(CondSelect x y (SETBCR [a] cmp))  => (ISEL [a+4] x y cmp)
   425// Only lower after bool is lowered. It should always lower. This helps ensure the folding below happens reliably.
   426(CondSelect x y bool) && flagArg(bool) == nil => (ISEL [6] x y (CMPconst [0] (ANDconst [1] bool)))
   427// Fold any CR -> GPR -> CR transfers when applying the above rule.
   428(ISEL [6] x y (CMPconst [0] (ANDconst [1] (SETBC [c] cmp)))) => (ISEL [c] x y cmp)
   429(ISEL [6] x y ((CMP|CMPW)const [0] (SETBC [c] cmp))) => (ISEL [c] x y cmp)
   430(ISEL [6] x y ((CMP|CMPW)const [0] (SETBCR [c] cmp))) => (ISEL [c+4] x y cmp)
   431
   432// Lowering loads
   433(Load <t> ptr mem) && (is64BitInt(t) || isPtr(t)) => (MOVDload ptr mem)
   434(Load <t> ptr mem) && is32BitInt(t) &&  t.IsSigned() => (MOVWload ptr mem)
   435(Load <t> ptr mem) && is32BitInt(t) && !t.IsSigned() => (MOVWZload ptr mem)
   436(Load <t> ptr mem) && is16BitInt(t) &&  t.IsSigned() => (MOVHload ptr mem)
   437(Load <t> ptr mem) && is16BitInt(t) && !t.IsSigned() => (MOVHZload ptr mem)
   438(Load <t> ptr mem) && t.IsBoolean() => (MOVBZload ptr mem)
   439(Load <t> ptr mem) && is8BitInt(t) &&  t.IsSigned() => (MOVBreg (MOVBZload ptr mem)) // PPC has no signed-byte load.
   440(Load <t> ptr mem) && is8BitInt(t) && !t.IsSigned() => (MOVBZload ptr mem)
   441
   442(Load <t> ptr mem) && is32BitFloat(t) => (FMOVSload ptr mem)
   443(Load <t> ptr mem) && is64BitFloat(t) => (FMOVDload ptr mem)
   444
   445(Store {t} ptr val mem) && t.Size() == 8 &&  t.IsFloat() => (FMOVDstore ptr val mem)
   446(Store {t} ptr val mem) && t.Size() == 4 &&  t.IsFloat() => (FMOVSstore ptr val mem)
   447(Store {t} ptr val mem) && t.Size() == 8 && !t.IsFloat() => (MOVDstore ptr val mem)
   448(Store {t} ptr val mem) && t.Size() == 4 && !t.IsFloat() => (MOVWstore ptr val mem)
   449(Store {t} ptr val mem) && t.Size() == 2 => (MOVHstore ptr val mem)
   450(Store {t} ptr val mem) && t.Size() == 1 => (MOVBstore ptr val mem)
   451
   452// Using Zero instead of LoweredZero allows the
   453// target address to be folded where possible.
   454(Zero [0] _ mem) => mem
   455(Zero [1] destptr mem) => (MOVBstorezero destptr mem)
   456(Zero [2] destptr mem) =>
   457	(MOVHstorezero destptr mem)
   458(Zero [3] destptr mem) =>
   459	(MOVBstorezero [2] destptr
   460		(MOVHstorezero destptr mem))
   461(Zero [4] destptr mem) =>
   462	(MOVWstorezero destptr mem)
   463(Zero [5] destptr mem) =>
   464	(MOVBstorezero [4] destptr
   465        	(MOVWstorezero destptr mem))
   466(Zero [6] destptr mem) =>
   467	(MOVHstorezero [4] destptr
   468		(MOVWstorezero destptr mem))
   469(Zero [7] destptr mem) =>
   470	(MOVBstorezero [6] destptr
   471		(MOVHstorezero [4] destptr
   472			(MOVWstorezero destptr mem)))
   473
   474(Zero [8] {t} destptr mem) => (MOVDstorezero destptr mem)
   475(Zero [12] {t} destptr mem) =>
   476        (MOVWstorezero [8] destptr
   477                (MOVDstorezero [0] destptr mem))
   478(Zero [16] {t} destptr mem) =>
   479       (MOVDstorezero [8] destptr
   480                (MOVDstorezero [0] destptr mem))
   481(Zero [24] {t} destptr mem) =>
   482       (MOVDstorezero [16] destptr
   483               (MOVDstorezero [8] destptr
   484                       (MOVDstorezero [0] destptr mem)))
   485(Zero [32] {t} destptr mem) =>
   486       (MOVDstorezero [24] destptr
   487               (MOVDstorezero [16] destptr
   488                       (MOVDstorezero [8] destptr
   489                               (MOVDstorezero [0] destptr mem))))
   490
   491// Handle cases not handled above
   492// Lowered Short cases do not generate loops, and as a result don't clobber
   493// the address registers or flags.
   494(Zero [s] ptr mem) && buildcfg.GOPPC64 <= 8 && s < 64 => (LoweredZeroShort [s] ptr mem)
   495(Zero [s] ptr mem) && buildcfg.GOPPC64 <= 8 => (LoweredZero [s] ptr mem)
   496(Zero [s] ptr mem) && s < 128 && buildcfg.GOPPC64 >= 9 => (LoweredQuadZeroShort [s] ptr mem)
   497(Zero [s] ptr mem) && buildcfg.GOPPC64 >= 9 => (LoweredQuadZero [s] ptr mem)
   498
   499// moves
   500(Move [0] _ _ mem) => mem
   501(Move [1] dst src mem) => (MOVBstore dst (MOVBZload src mem) mem)
   502(Move [2] dst src mem) =>
   503        (MOVHstore dst (MOVHZload src mem) mem)
   504(Move [4] dst src mem) =>
   505	(MOVWstore dst (MOVWZload src mem) mem)
   506// MOVD for load and store must have offsets that are multiple of 4
   507(Move [8] {t} dst src mem) =>
   508	(MOVDstore dst (MOVDload src mem) mem)
   509(Move [3] dst src mem) =>
   510        (MOVBstore [2] dst (MOVBZload [2] src mem)
   511                (MOVHstore dst (MOVHload src mem) mem))
   512(Move [5] dst src mem) =>
   513        (MOVBstore [4] dst (MOVBZload [4] src mem)
   514                (MOVWstore dst (MOVWZload src mem) mem))
   515(Move [6] dst src mem) =>
   516        (MOVHstore [4] dst (MOVHZload [4] src mem)
   517                (MOVWstore dst (MOVWZload src mem) mem))
   518(Move [7] dst src mem) =>
   519        (MOVBstore [6] dst (MOVBZload [6] src mem)
   520                (MOVHstore [4] dst (MOVHZload [4] src mem)
   521                        (MOVWstore dst (MOVWZload src mem) mem)))
   522
   523// Large move uses a loop. Since the address is computed and the
   524// offset is zero, any alignment can be used.
   525(Move [s] dst src mem) && s > 8 && buildcfg.GOPPC64 <= 8 && logLargeCopy(v, s) =>
   526        (LoweredMove [s] dst src mem)
   527(Move [s] dst src mem) && s > 8 && s <= 64 && buildcfg.GOPPC64 >= 9 =>
   528        (LoweredQuadMoveShort [s] dst src mem)
   529(Move [s] dst src mem) && s > 8 && buildcfg.GOPPC64 >= 9 && logLargeCopy(v, s) =>
   530        (LoweredQuadMove [s] dst src mem)
   531
   532// Calls
   533// Lowering calls
   534(StaticCall ...) => (CALLstatic ...)
   535(ClosureCall ...) => (CALLclosure ...)
   536(InterCall ...) => (CALLinter ...)
   537(TailCall ...) => (CALLtail ...)
   538
   539// Miscellaneous
   540(GetClosurePtr ...) => (LoweredGetClosurePtr ...)
   541(GetCallerSP ...) => (LoweredGetCallerSP ...)
   542(GetCallerPC ...) => (LoweredGetCallerPC ...)
   543(IsNonNil ptr) => (NotEqual (CMPconst [0] ptr))
   544(IsInBounds idx len) => (LessThan (CMPU idx len))
   545(IsSliceInBounds idx len) => (LessEqual (CMPU idx len))
   546(NilCheck ...) => (LoweredNilCheck ...)
   547
   548// Write barrier.
   549(WB ...) => (LoweredWB ...)
   550
   551// Publication barrier as intrinsic
   552(PubBarrier ...) => (LoweredPubBarrier ...)
   553
   554(PanicBounds [kind] x y mem) && boundsABI(kind) == 0 => (LoweredPanicBoundsA [kind] x y mem)
   555(PanicBounds [kind] x y mem) && boundsABI(kind) == 1 => (LoweredPanicBoundsB [kind] x y mem)
   556(PanicBounds [kind] x y mem) && boundsABI(kind) == 2 => (LoweredPanicBoundsC [kind] x y mem)
   557
   558// Optimizations
   559// Note that PPC "logical" immediates come in 0:15 and 16:31 unsigned immediate forms,
   560// so ORconst, XORconst easily expand into a pair.
   561
   562// Include very-large constants in the const-const case.
   563(AND (MOVDconst [c]) (MOVDconst [d])) => (MOVDconst [c&d])
   564(OR (MOVDconst [c]) (MOVDconst [d])) => (MOVDconst [c|d])
   565(XOR (MOVDconst [c]) (MOVDconst [d])) => (MOVDconst [c^d])
   566(ORN (MOVDconst [c]) (MOVDconst [d])) => (MOVDconst [c|^d])
   567(ANDN (MOVDconst [c]) (MOVDconst [d])) => (MOVDconst [c&^d])
   568(NOR (MOVDconst [c]) (MOVDconst [d])) => (MOVDconst [^(c|d)])
   569
   570// Discover consts
   571(AND x (MOVDconst [-1])) => x
   572(AND x (MOVDconst [c])) && isU16Bit(c) => (ANDconst [c] x)
   573(XOR x (MOVDconst [c])) && isU32Bit(c) => (XORconst [c] x)
   574(OR x (MOVDconst [c])) && isU32Bit(c) => (ORconst [c] x)
   575
   576// Simplify consts
   577(ANDconst [c] (ANDconst [d] x)) => (ANDconst [c&d] x)
   578(ORconst [c] (ORconst [d] x)) => (ORconst [c|d] x)
   579(XORconst [c] (XORconst [d] x)) => (XORconst [c^d] x)
   580(ANDconst [-1] x) => x
   581(ANDconst [0] _) => (MOVDconst [0])
   582(XORconst [0] x) => x
   583(ORconst [-1] _) => (MOVDconst [-1])
   584(ORconst [0] x) => x
   585
   586// zero-extend of small and => small and
   587(MOVBZreg y:(ANDconst [c] _)) && uint64(c) <= 0xFF => y
   588(MOVHZreg y:(ANDconst [c] _)) && uint64(c) <= 0xFFFF => y
   589(MOVWZreg y:(ANDconst [c] _)) && uint64(c) <= 0xFFFFFFFF => y
   590(MOVWZreg y:(AND (MOVDconst [c]) _)) && uint64(c) <= 0xFFFFFFFF => y
   591
   592// sign extend of small-positive and => small-positive-and
   593(MOVBreg y:(ANDconst [c] _)) && uint64(c) <= 0x7F => y
   594(MOVHreg y:(ANDconst [c] _)) && uint64(c) <= 0x7FFF => y
   595(MOVWreg y:(ANDconst [c] _)) && uint64(c) <= 0xFFFF => y // 0xFFFF is largest immediate constant, when regarded as 32-bit is > 0
   596(MOVWreg y:(AND (MOVDconst [c]) _)) && uint64(c) <= 0x7FFFFFFF => y
   597
   598// small and of zero-extend => either zero-extend or small and
   599(ANDconst [c] y:(MOVBZreg _)) && c&0xFF == 0xFF => y
   600(ANDconst [0xFF] (MOVBreg x)) => (MOVBZreg x)
   601(ANDconst [c] y:(MOVHZreg _))  && c&0xFFFF == 0xFFFF => y
   602(ANDconst [0xFFFF] (MOVHreg x)) => (MOVHZreg x)
   603
   604(AND (MOVDconst [c]) y:(MOVWZreg _))  && c&0xFFFFFFFF == 0xFFFFFFFF => y
   605(AND (MOVDconst [0xFFFFFFFF]) y:(MOVWreg x)) => (MOVWZreg x)
   606// normal case
   607(ANDconst [c] (MOVBZreg x)) => (ANDconst [c&0xFF] x)
   608(ANDconst [c] (MOVHZreg x)) => (ANDconst [c&0xFFFF] x)
   609(ANDconst [c] (MOVWZreg x)) => (ANDconst [c&0xFFFFFFFF] x)
   610
   611// Eliminate unnecessary sign/zero extend following right shift
   612(MOV(B|H|W)Zreg (SRWconst [c] (MOVBZreg x))) => (SRWconst [c] (MOVBZreg x))
   613(MOV(H|W)Zreg (SRWconst [c] (MOVHZreg x))) => (SRWconst [c] (MOVHZreg x))
   614(MOVWZreg (SRWconst [c] (MOVWZreg x))) => (SRWconst [c] (MOVWZreg x))
   615(MOV(B|H|W)reg (SRAWconst [c] (MOVBreg x))) => (SRAWconst [c] (MOVBreg x))
   616(MOV(H|W)reg (SRAWconst [c] (MOVHreg x))) => (SRAWconst [c] (MOVHreg x))
   617(MOVWreg (SRAWconst [c] (MOVWreg x))) => (SRAWconst [c] (MOVWreg x))
   618
   619(MOV(WZ|W)reg (S(R|RA)Wconst [c] x)) && x.Type.Size() <= 32 => (S(R|RA)Wconst [c] x)
   620(MOV(HZ|H)reg (S(R|RA)Wconst [c] x)) && x.Type.Size() <= 16 => (S(R|RA)Wconst [c] x)
   621(MOV(BZ|B)reg (S(R|RA)Wconst [c] x)) && x.Type.Size() == 8 => (S(R|RA)Wconst [c] x)
   622
   623// initial right shift will handle sign/zero extend
   624(MOVBZreg (SRDconst [c] x)) && c>=56 => (SRDconst [c] x)
   625(MOVBreg (SRDconst [c] x)) && c>56 => (SRDconst [c] x)
   626(MOVBreg (SRDconst [c] x)) && c==56 => (SRADconst [c] x)
   627(MOVBreg (SRADconst [c] x)) && c>=56 => (SRADconst [c] x)
   628(MOVBZreg (SRWconst [c] x)) && c>=24 => (SRWconst [c] x)
   629(MOVBreg (SRWconst [c] x)) && c>24 => (SRWconst [c] x)
   630(MOVBreg (SRWconst [c] x)) && c==24 => (SRAWconst [c] x)
   631(MOVBreg (SRAWconst [c] x)) && c>=24 => (SRAWconst [c] x)
   632
   633(MOVHZreg (SRDconst [c] x)) && c>=48 => (SRDconst [c] x)
   634(MOVHreg (SRDconst [c] x)) && c>48 => (SRDconst [c] x)
   635(MOVHreg (SRDconst [c] x)) && c==48 => (SRADconst [c] x)
   636(MOVHreg (SRADconst [c] x)) && c>=48 => (SRADconst [c] x)
   637(MOVHZreg (SRWconst [c] x)) && c>=16 => (SRWconst [c] x)
   638(MOVHreg (SRWconst [c] x)) && c>16 => (SRWconst [c] x)
   639(MOVHreg (SRAWconst [c] x)) && c>=16 => (SRAWconst [c] x)
   640(MOVHreg (SRWconst [c] x)) && c==16 => (SRAWconst [c] x)
   641
   642(MOVWZreg (SRDconst [c] x)) && c>=32 => (SRDconst [c] x)
   643(MOVWreg (SRDconst [c] x)) && c>32 => (SRDconst [c] x)
   644(MOVWreg (SRADconst [c] x)) && c>=32 => (SRADconst [c] x)
   645(MOVWreg (SRDconst [c] x)) && c==32 => (SRADconst [c] x)
   646
   647// Various redundant zero/sign extension combinations.
   648(MOVBZreg y:(MOVBZreg _)) => y  // repeat
   649(MOVBreg y:(MOVBreg _)) => y // repeat
   650(MOVBreg (MOVBZreg x)) => (MOVBreg x)
   651(MOVBZreg (MOVBreg x)) => (MOVBZreg x)
   652
   653// Catch any remaining rotate+shift cases
   654(MOVBZreg (SRWconst x [s])) && mergePPC64AndSrwi(0xFF,s) != 0 => (RLWINM [mergePPC64AndSrwi(0xFF,s)] x)
   655(MOVBZreg (RLWINM [r] y)) && mergePPC64AndRlwinm(0xFF,r) != 0 => (RLWINM [mergePPC64AndRlwinm(0xFF,r)] y)
   656(MOVHZreg (RLWINM [r] y)) && mergePPC64AndRlwinm(0xFFFF,r) != 0 => (RLWINM [mergePPC64AndRlwinm(0xFFFF,r)] y)
   657(MOVWZreg (RLWINM [r] y)) && mergePPC64MovwzregRlwinm(r) != 0 => (RLWINM [mergePPC64MovwzregRlwinm(r)] y)
   658(ANDconst [m] (RLWINM [r] y)) && mergePPC64AndRlwinm(uint32(m),r) != 0 => (RLWINM [mergePPC64AndRlwinm(uint32(m),r)] y)
   659(SLDconst [s] (RLWINM [r] y)) && mergePPC64SldiRlwinm(s,r) != 0 => (RLWINM [mergePPC64SldiRlwinm(s,r)] y)
   660(RLWINM [r] (MOVHZreg u)) && mergePPC64RlwinmAnd(r,0xFFFF) != 0 => (RLWINM [mergePPC64RlwinmAnd(r,0xFFFF)] u)
   661(RLWINM [r] (ANDconst [a] u)) && mergePPC64RlwinmAnd(r,uint32(a)) != 0 => (RLWINM [mergePPC64RlwinmAnd(r,uint32(a))] u)
   662// SLWconst is a special case of RLWNM which always zero-extends the result.
   663(SLWconst [s] (MOVWZreg w)) => (SLWconst [s] w)
   664(MOVWZreg w:(SLWconst u)) => w
   665
   666// H - there are more combinations than these
   667
   668(MOVHZreg y:(MOV(H|B)Zreg _)) => y // repeat
   669(MOVHZreg y:(MOVHBRload _ _)) => y
   670
   671(MOVHreg y:(MOV(H|B)reg _)) => y // repeat
   672
   673(MOV(H|HZ)reg y:(MOV(HZ|H)reg x)) => (MOV(H|HZ)reg x)
   674
   675// W - there are more combinations than these
   676
   677(MOV(WZ|WZ|WZ|W|W|W)reg y:(MOV(WZ|HZ|BZ|W|H|B)reg _)) => y // repeat
   678(MOVWZreg y:(MOV(H|W)BRload _ _)) => y
   679
   680(MOV(W|WZ)reg y:(MOV(WZ|W)reg x)) => (MOV(W|WZ)reg x)
   681
   682// Truncate then logical then truncate: omit first, lesser or equal truncate
   683(MOVWZreg ((OR|XOR|AND) <t> x (MOVWZreg y))) => (MOVWZreg ((OR|XOR|AND) <t> x y))
   684(MOVHZreg ((OR|XOR|AND) <t> x (MOVWZreg y))) => (MOVHZreg ((OR|XOR|AND) <t> x y))
   685(MOVHZreg ((OR|XOR|AND) <t> x (MOVHZreg y))) => (MOVHZreg ((OR|XOR|AND) <t> x y))
   686(MOVBZreg ((OR|XOR|AND) <t> x (MOVWZreg y))) => (MOVBZreg ((OR|XOR|AND) <t> x y))
   687(MOVBZreg ((OR|XOR|AND) <t> x (MOVHZreg y))) => (MOVBZreg ((OR|XOR|AND) <t> x y))
   688(MOVBZreg ((OR|XOR|AND) <t> x (MOVBZreg y))) => (MOVBZreg ((OR|XOR|AND) <t> x y))
   689
   690(MOV(B|H|W)Zreg z:(ANDconst [c] (MOVBZload ptr x))) => z
   691(MOV(B|H|W)Zreg z:(AND y (MOV(B|H|W)Zload ptr x))) => z
   692(MOV(H|W)Zreg z:(ANDconst [c] (MOVHZload ptr x))) => z
   693(MOVWZreg z:(ANDconst [c] (MOVWZload ptr x))) => z
   694
   695// Arithmetic constant ops
   696
   697(ADD x (MOVDconst <t> [c])) && is32Bit(c) && !t.IsPtr() => (ADDconst [c] x)
   698(ADDconst [c] (ADDconst [d] x)) && is32Bit(c+d) => (ADDconst [c+d] x)
   699(ADDconst [0] x) => x
   700(SUB x (MOVDconst [c])) && is32Bit(-c) => (ADDconst [-c] x)
   701
   702(ADDconst [c] (MOVDaddr [d] {sym} x)) && is32Bit(c+int64(d)) => (MOVDaddr [int32(c+int64(d))] {sym} x)
   703(ADDconst [c] x:(SP)) && is32Bit(c) => (MOVDaddr [int32(c)] x) // so it is rematerializeable
   704
   705(MULL(W|D) x (MOVDconst [c])) && is16Bit(c) => (MULL(W|D)const [int32(c)] x)
   706
   707// Subtract from (with carry, but ignored) constant.
   708// Note, these clobber the carry bit.
   709(SUB (MOVDconst [c]) x) && is32Bit(c) => (SUBFCconst [c] x)
   710(SUBFCconst [c] (NEG x)) => (ADDconst [c] x)
   711(SUBFCconst [c] (SUBFCconst [d] x)) && is32Bit(c-d) => (ADDconst [c-d] x)
   712(SUBFCconst [0] x) => (NEG x)
   713(ADDconst [c] (SUBFCconst [d] x)) && is32Bit(c+d) => (SUBFCconst [c+d] x)
   714(NEG (ADDconst [c] x)) && is32Bit(-c) => (SUBFCconst [-c] x)
   715(NEG (SUBFCconst [c] x)) && is32Bit(-c) => (ADDconst [-c] x)
   716(NEG (SUB x y)) => (SUB y x)
   717(NEG (NEG x)) => x
   718
   719// Use register moves instead of stores and loads to move int<=>float values
   720// Common with math Float64bits, Float64frombits
   721(MOVDload [off] {sym} ptr (FMOVDstore [off] {sym} ptr x _)) => (MFVSRD x)
   722(FMOVDload [off] {sym} ptr (MOVDstore [off] {sym} ptr x _)) => (MTVSRD x)
   723
   724(FMOVDstore [off] {sym} ptr (MTVSRD x) mem) => (MOVDstore [off] {sym} ptr x mem)
   725(MOVDstore [off] {sym} ptr (MFVSRD x) mem) => (FMOVDstore [off] {sym} ptr x mem)
   726
   727(MTVSRD (MOVDconst [c])) && !math.IsNaN(math.Float64frombits(uint64(c))) => (FMOVDconst [math.Float64frombits(uint64(c))])
   728(MFVSRD (FMOVDconst [c])) => (MOVDconst [int64(math.Float64bits(c))])
   729
   730(MTVSRD x:(MOVDload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (FMOVDload [off] {sym} ptr mem)
   731(MFVSRD x:(FMOVDload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVDload [off] {sym} ptr mem)
   732
   733// Rules for MOV* or FMOV* ops determine when indexed (MOV*loadidx or MOV*storeidx)
   734// or non-indexed (MOV*load or MOV*store) should be used. Indexed instructions
   735// require an extra instruction and register to load the index so non-indexed is preferred.
   736// Indexed ops generate indexed load or store instructions for all GOPPC64 values.
   737// Non-indexed ops generate DS-form loads and stores when the offset fits in 16 bits,
   738// and on power8 and power9, a multiple of 4 is required for MOVW and MOVD ops.
   739// On power10, prefixed loads and stores can be used for offsets > 16 bits and <= 32 bits.
   740// and support for PC relative addressing must be available if relocation is needed.
   741// On power10, the assembler will determine when to use DS-form or prefixed
   742// instructions for non-indexed ops depending on the value of the offset.
   743//
   744// Fold offsets for stores.
   745(MOV(D|W|H|B)store [off1] {sym} (ADDconst [off2] x) val mem) && (is16Bit(int64(off1)+off2) || (supportsPPC64PCRel() && is32Bit(int64(off1)+off2))) => (MOV(D|W|H|B)store [off1+int32(off2)] {sym} x val mem)
   746
   747(FMOV(S|D)store [off1] {sym} (ADDconst [off2] ptr) val mem) && (is16Bit(int64(off1)+off2) || (supportsPPC64PCRel() && is32Bit(int64(off1)+off2))) => (FMOV(S|D)store [off1+int32(off2)] {sym} ptr val mem)
   748
   749// Fold address into load/store.
   750// If power10 with PCRel is not available, then
   751// the assembler needs to generate several instructions and use
   752// temp register for accessing global, and each time it will reload
   753// the temp register. So don't fold address of global in that case if there is more than
   754// one use.
   755(MOV(B|H|W|D)store [off1] {sym1} p:(MOVDaddr [off2] {sym2} ptr) val mem) && canMergeSym(sym1,sym2)
   756	&& ((is16Bit(int64(off1+off2)) && (ptr.Op != OpSB || p.Uses == 1)) || (supportsPPC64PCRel() && is32Bit(int64(off1+off2)))) =>
   757        (MOV(B|H|W|D)store [off1+off2] {mergeSym(sym1,sym2)} ptr val mem)
   758
   759(FMOV(S|D)store [off1] {sym1} p:(MOVDaddr [off2] {sym2} ptr) val mem) && canMergeSym(sym1,sym2)
   760	&& ((is16Bit(int64(off1+off2)) && (ptr.Op != OpSB || p.Uses == 1)) || (supportsPPC64PCRel() && is32Bit(int64(off1+off2)))) =>
   761        (FMOV(S|D)store [off1+off2] {mergeSym(sym1,sym2)} ptr val mem)
   762
   763(MOV(B|H|W)Zload [off1] {sym1} p:(MOVDaddr [off2] {sym2} ptr) mem) && canMergeSym(sym1,sym2)
   764	&& ((is16Bit(int64(off1+off2)) && (ptr.Op != OpSB || p.Uses == 1)) || (supportsPPC64PCRel() && is32Bit(int64(off1+off2)))) =>
   765        (MOV(B|H|W)Zload [off1+off2] {mergeSym(sym1,sym2)} ptr mem)
   766(MOV(H|W|D)load [off1] {sym1} p:(MOVDaddr [off2] {sym2} ptr) mem) && canMergeSym(sym1,sym2)
   767	&& ((is16Bit(int64(off1+off2)) && (ptr.Op != OpSB || p.Uses == 1)) || (supportsPPC64PCRel() && is32Bit(int64(off1+off2)))) =>
   768        (MOV(H|W|D)load [off1+off2] {mergeSym(sym1,sym2)} ptr mem)
   769(FMOV(S|D)load [off1] {sym1} p:(MOVDaddr [off2] {sym2} ptr) mem) && canMergeSym(sym1,sym2)
   770	&& ((is16Bit(int64(off1+off2)) && (ptr.Op != OpSB || p.Uses == 1)) || (supportsPPC64PCRel() && is32Bit(int64(off1+off2)))) =>
   771        (FMOV(S|D)load [off1+off2] {mergeSym(sym1,sym2)} ptr mem)
   772
   773// Fold offsets for loads.
   774(FMOV(S|D)load [off1] {sym} (ADDconst [off2] ptr) mem) && (is16Bit(int64(off1)+off2) || (supportsPPC64PCRel() && is32Bit(int64(off1)+off2))) => (FMOV(S|D)load [off1+int32(off2)] {sym} ptr mem)
   775
   776(MOV(D|W|WZ|H|HZ|BZ)load [off1] {sym} (ADDconst [off2] x) mem) && (is16Bit(int64(off1)+off2) || (supportsPPC64PCRel() && is32Bit(int64(off1)+off2))) => (MOV(D|W|WZ|H|HZ|BZ)load [off1+int32(off2)] {sym} x mem)
   777
   778// Determine load + addressing that can be done as a register indexed load
   779(MOV(D|W|WZ|H|HZ|BZ)load [0] {sym} p:(ADD ptr idx) mem) && sym == nil && p.Uses == 1 => (MOV(D|W|WZ|H|HZ|BZ)loadidx ptr idx mem)
   780
   781// See comments above concerning selection of indexed vs. non-indexed ops.
   782// These cases don't have relocation.
   783(MOV(D|W)loadidx ptr (MOVDconst [c]) mem) && ((is16Bit(c) && c%4 == 0) || (buildcfg.GOPPC64 >= 10 && is32Bit(c))) => (MOV(D|W)load [int32(c)] ptr mem)
   784(MOV(WZ|H|HZ|BZ)loadidx ptr (MOVDconst [c]) mem) && (is16Bit(c) || (buildcfg.GOPPC64 >= 10 && is32Bit(c))) => (MOV(WZ|H|HZ|BZ)load [int32(c)] ptr mem)
   785(MOV(D|W)loadidx (MOVDconst [c]) ptr mem) && ((is16Bit(c) && c%4 == 0) || (buildcfg.GOPPC64 >= 10 && is32Bit(c))) => (MOV(D|W)load [int32(c)] ptr mem)
   786(MOV(WZ|H|HZ|BZ)loadidx (MOVDconst [c]) ptr mem) && (is16Bit(c) || (buildcfg.GOPPC64 >= 10 && is32Bit(c))) => (MOV(WZ|H|HZ|BZ)load [int32(c)] ptr mem)
   787
   788// Store of zero => storezero
   789(MOV(D|W|H|B)store [off] {sym} ptr (MOVDconst [0]) mem) => (MOV(D|W|H|B)storezero [off] {sym} ptr mem)
   790
   791// Fold offsets for storezero
   792(MOV(D|W|H|B)storezero [off1] {sym} (ADDconst [off2] x) mem) && ((supportsPPC64PCRel() && is32Bit(int64(off1)+off2)) || (is16Bit(int64(off1)+off2))) =>
   793    (MOV(D|W|H|B)storezero [off1+int32(off2)] {sym} x mem)
   794
   795// Stores with addressing that can be done as indexed stores
   796(MOV(D|W|H|B)store [0] {sym} p:(ADD ptr idx) val mem) && sym == nil && p.Uses == 1 => (MOV(D|W|H|B)storeidx ptr idx val mem)
   797
   798(MOVDstoreidx ptr (MOVDconst [c]) val mem) && ((is16Bit(c) && c%4 == 0) || (buildcfg.GOPPC64 >= 10 && is32Bit(c))) => (MOVDstore [int32(c)] ptr val mem)
   799(MOV(W|H|B)storeidx ptr (MOVDconst [c]) val mem) && (is16Bit(c) || (buildcfg.GOPPC64 >= 10 && is32Bit(c))) => (MOV(W|H|B)store [int32(c)] ptr val mem)
   800(MOVDstoreidx (MOVDconst [c]) ptr val mem) && ((is16Bit(c) && c%4 == 0) || (buildcfg.GOPPC64 >= 10 && is32Bit(c))) => (MOVDstore [int32(c)] ptr val mem)
   801(MOV(W|H|B)storeidx (MOVDconst [c]) ptr val mem) && (is16Bit(c) || (buildcfg.GOPPC64 >= 10 && is32Bit(c))) => (MOV(W|H|B)store [int32(c)] ptr val mem)
   802
   803// Fold symbols into storezero
   804(MOV(D|W|H|B)storezero [off1] {sym1} p:(MOVDaddr [off2] {sym2} x) mem) && canMergeSym(sym1,sym2)
   805	&& ((is16Bit(int64(off1+off2)) && (x.Op != OpSB || p.Uses == 1)) || (supportsPPC64PCRel() && is32Bit(int64(off1+off2)))) =>
   806    (MOV(D|W|H|B)storezero [off1+off2] {mergeSym(sym1,sym2)} x mem)
   807
   808// atomic intrinsics
   809(AtomicLoad(8|32|64|Ptr)  ptr mem) => (LoweredAtomicLoad(8|32|64|Ptr) [1] ptr mem)
   810(AtomicLoadAcq(32|64)     ptr mem) => (LoweredAtomicLoad(32|64) [0] ptr mem)
   811
   812(AtomicStore(8|32|64)    ptr val mem) => (LoweredAtomicStore(8|32|64) [1] ptr val mem)
   813(AtomicStoreRel(32|64)   ptr val mem) => (LoweredAtomicStore(32|64) [0] ptr val mem)
   814
   815(AtomicExchange(8|32|64) ...) => (LoweredAtomicExchange(8|32|64) ...)
   816
   817(AtomicAdd(32|64) ...) => (LoweredAtomicAdd(32|64) ...)
   818
   819(AtomicCompareAndSwap(32|64) ptr old new_ mem) => (LoweredAtomicCas(32|64) [1] ptr old new_ mem)
   820(AtomicCompareAndSwapRel32   ptr old new_ mem) => (LoweredAtomicCas32 [0] ptr old new_ mem)
   821
   822(AtomicAnd(8|32)  ...) => (LoweredAtomicAnd(8|32)  ...)
   823(AtomicOr(8|32)   ...) => (LoweredAtomicOr(8|32)   ...)
   824
   825(Slicemask <t> x) => (SRADconst (NEG <t> x) [63])
   826(ANDconst [1] z:(SRADconst [63] x)) && z.Uses == 1  => (SRDconst [63] x)
   827
   828// Note that MOV??reg returns a 64-bit int, x is not necessarily that wide
   829// This may interact with other patterns in the future. (Compare with arm64)
   830(MOV(B|H|W)Zreg x:(MOVBZload _ _)) => x
   831(MOV(B|H|W)Zreg x:(MOVBZloadidx _ _ _)) => x
   832(MOV(H|W)Zreg x:(MOVHZload _ _)) => x
   833(MOV(H|W)Zreg x:(MOVHZloadidx _ _ _)) => x
   834(MOV(H|W)reg x:(MOVHload _ _)) => x
   835(MOV(H|W)reg x:(MOVHloadidx _ _ _)) => x
   836(MOV(WZ|W)reg x:(MOV(WZ|W)load _ _)) => x
   837(MOV(WZ|W)reg x:(MOV(WZ|W)loadidx _ _ _)) => x
   838(MOV(B|W)Zreg x:(Select0 (LoweredAtomicLoad(8|32) _ _))) => x
   839
   840// don't extend if argument is already extended
   841(MOVBreg x:(Arg <t>)) && is8BitInt(t) && t.IsSigned() => x
   842(MOVBZreg x:(Arg <t>)) && is8BitInt(t) && !t.IsSigned() => x
   843(MOVHreg x:(Arg <t>)) && (is8BitInt(t) || is16BitInt(t)) && t.IsSigned() => x
   844(MOVHZreg x:(Arg <t>)) && (is8BitInt(t) || is16BitInt(t)) && !t.IsSigned() => x
   845(MOVWreg x:(Arg <t>)) && (is8BitInt(t) || is16BitInt(t) || is32BitInt(t)) && t.IsSigned() => x
   846(MOVWZreg x:(Arg <t>)) && (is8BitInt(t) || is16BitInt(t) || is32BitInt(t)) && !t.IsSigned() => x
   847
   848(MOVBZreg (MOVDconst [c]))  => (MOVDconst [int64(uint8(c))])
   849(MOVBreg (MOVDconst [c]))  => (MOVDconst [int64(int8(c))])
   850(MOVHZreg (MOVDconst [c]))  => (MOVDconst [int64(uint16(c))])
   851(MOVHreg (MOVDconst [c]))  => (MOVDconst [int64(int16(c))])
   852(MOVWreg (MOVDconst [c])) => (MOVDconst [int64(int32(c))])
   853(MOVWZreg (MOVDconst [c])) => (MOVDconst [int64(uint32(c))])
   854
   855// Implement clrsldi and clrslwi extended mnemonics as described in
   856// ISA 3.0 section C.8. AuxInt field contains values needed for
   857// the instructions, packed together since there is only one available.
   858(SLDconst [c] z:(MOVBZreg x)) && c < 8 && z.Uses == 1 => (CLRLSLDI [newPPC64ShiftAuxInt(c,56,63,64)] x)
   859(SLDconst [c] z:(MOVHZreg x)) && c < 16 && z.Uses == 1 => (CLRLSLDI [newPPC64ShiftAuxInt(c,48,63,64)] x)
   860(SLDconst [c] z:(MOVWZreg x)) && c < 32 && z.Uses == 1 => (CLRLSLDI [newPPC64ShiftAuxInt(c,32,63,64)] x)
   861
   862(SLDconst [c] z:(ANDconst [d] x)) && z.Uses == 1 && isPPC64ValidShiftMask(d) && c <= (64-getPPC64ShiftMaskLength(d)) => (CLRLSLDI [newPPC64ShiftAuxInt(c,64-getPPC64ShiftMaskLength(d),63,64)] x)
   863(SLDconst [c] z:(AND (MOVDconst [d]) x)) && z.Uses == 1 && isPPC64ValidShiftMask(d) && c<=(64-getPPC64ShiftMaskLength(d)) => (CLRLSLDI [newPPC64ShiftAuxInt(c,64-getPPC64ShiftMaskLength(d),63,64)] x)
   864(SLWconst [c] z:(MOVBZreg x)) && z.Uses == 1 && c < 8 => (CLRLSLWI [newPPC64ShiftAuxInt(c,24,31,32)] x)
   865(SLWconst [c] z:(MOVHZreg x)) && z.Uses == 1 && c < 16 => (CLRLSLWI [newPPC64ShiftAuxInt(c,16,31,32)] x)
   866(SLWconst [c] z:(ANDconst [d] x)) && z.Uses == 1 && isPPC64ValidShiftMask(d) && c<=(32-getPPC64ShiftMaskLength(d)) => (CLRLSLWI [newPPC64ShiftAuxInt(c,32-getPPC64ShiftMaskLength(d),31,32)] x)
   867(SLWconst [c] z:(AND (MOVDconst [d]) x)) && z.Uses == 1 && isPPC64ValidShiftMask(d) && c<=(32-getPPC64ShiftMaskLength(d)) => (CLRLSLWI [newPPC64ShiftAuxInt(c,32-getPPC64ShiftMaskLength(d),31,32)] x)
   868// special case for power9
   869(SL(W|D)const [c] z:(MOVWreg x)) && c < 32 && buildcfg.GOPPC64 >= 9 => (EXTSWSLconst [c] x)
   870
   871// Lose widening ops fed to stores
   872(MOVBstore [off] {sym} ptr (MOV(B|BZ|H|HZ|W|WZ)reg x) mem) => (MOVBstore [off] {sym} ptr x mem)
   873(MOVHstore [off] {sym} ptr (MOV(H|HZ|W|WZ)reg x) mem) => (MOVHstore [off] {sym} ptr x mem)
   874(MOVWstore [off] {sym} ptr (MOV(W|WZ)reg x) mem) => (MOVWstore [off] {sym} ptr x mem)
   875(MOVBstore [off] {sym} ptr (SRWconst (MOV(H|HZ)reg x) [c]) mem) && c <= 8 => (MOVBstore [off] {sym} ptr (SRWconst <typ.UInt32> x [c]) mem)
   876(MOVBstore [off] {sym} ptr (SRWconst (MOV(W|WZ)reg x) [c]) mem) && c <= 24 => (MOVBstore [off] {sym} ptr (SRWconst <typ.UInt32> x [c]) mem)
   877(MOVBstoreidx ptr idx (MOV(B|BZ|H|HZ|W|WZ)reg x) mem) => (MOVBstoreidx ptr idx x mem)
   878(MOVHstoreidx ptr idx (MOV(H|HZ|W|WZ)reg x) mem) => (MOVHstoreidx ptr idx x mem)
   879(MOVWstoreidx ptr idx (MOV(W|WZ)reg x) mem) => (MOVWstoreidx ptr idx x mem)
   880(MOVBstoreidx ptr idx (SRWconst (MOV(H|HZ)reg x) [c]) mem) && c <= 8 => (MOVBstoreidx ptr idx (SRWconst <typ.UInt32> x [c]) mem)
   881(MOVBstoreidx ptr idx (SRWconst (MOV(W|WZ)reg x) [c]) mem) && c <= 24 => (MOVBstoreidx ptr idx (SRWconst <typ.UInt32> x [c]) mem)
   882(MOVHBRstore ptr (MOV(H|HZ|W|WZ)reg x) mem) => (MOVHBRstore ptr x mem)
   883(MOVWBRstore ptr (MOV(W|WZ)reg x) mem) => (MOVWBRstore ptr x mem)
   884
   885// Lose W-widening ops fed to compare-W
   886(CMP(W|WU) x (MOV(W|WZ)reg y)) => (CMP(W|WU) x y)
   887(CMP(W|WU) (MOV(W|WZ)reg x) y) => (CMP(W|WU) x y)
   888
   889(CMP x (MOVDconst [c])) && is16Bit(c) => (CMPconst x [c])
   890(CMP (MOVDconst [c]) y) && is16Bit(c) => (InvertFlags (CMPconst y [c]))
   891(CMPW x (MOVDconst [c])) && is16Bit(c) => (CMPWconst x [int32(c)])
   892(CMPW (MOVDconst [c]) y) && is16Bit(c) => (InvertFlags (CMPWconst y [int32(c)]))
   893
   894(CMPU x (MOVDconst [c])) && isU16Bit(c) => (CMPUconst x [c])
   895(CMPU (MOVDconst [c]) y) && isU16Bit(c) => (InvertFlags (CMPUconst y [c]))
   896(CMPWU x (MOVDconst [c])) && isU16Bit(c) => (CMPWUconst x [int32(c)])
   897(CMPWU (MOVDconst [c]) y) && isU16Bit(c) => (InvertFlags (CMPWUconst y [int32(c)]))
   898
   899// Canonicalize the order of arguments to comparisons - helps with CSE.
   900((CMP|CMPW|CMPU|CMPWU) x y) && canonLessThan(x,y) => (InvertFlags ((CMP|CMPW|CMPU|CMPWU) y x))
   901
   902// n is always a zero-extended uint16 value, so n & z is always a non-negative 32 or 64 bit value.
   903// Rewrite to a cmp int64(0) to lower into ANDCCconst in the latelower pass.
   904(CMP(W|U|WU)const [0] a:(ANDconst [n] z)) => (CMPconst [0] a)
   905
   906// SETBC auxInt values 0=LT 1=GT 2=EQ   Crbit==1 ? 1 : 0
   907// SETBCR auxInt values 0=LT 1=GT 2=EQ   Crbit==1 ? 0 : 1
   908(Equal cmp) => (SETBC [2] cmp)
   909(NotEqual cmp) => (SETBCR [2] cmp)
   910(LessThan cmp) => (SETBC [0] cmp)
   911(FLessThan cmp) => (SETBC [0] cmp)
   912(FLessEqual cmp) => (OR (SETBC [2] cmp) (SETBC [0] cmp))
   913(GreaterEqual cmp) => (SETBCR [0] cmp)
   914(GreaterThan cmp)  => (SETBC [1] cmp)
   915(FGreaterEqual cmp) => (OR (SETBC [2] cmp) (SETBC [1] cmp))
   916(FGreaterThan cmp)  => (SETBC [1] cmp)
   917(LessEqual cmp) => (SETBCR [1] cmp)
   918
   919(SETBC [0] (FlagLT)) => (MOVDconst [1])
   920(SETBC [0] (Flag(GT|EQ))) => (MOVDconst [0])
   921(SETBC [1] (FlagGT)) => (MOVDconst [1])
   922(SETBC [1] (Flag(LT|EQ))) => (MOVDconst [0])
   923(SETBC [2] (FlagEQ)) => (MOVDconst [1])
   924(SETBC [2] (Flag(LT|GT))) => (MOVDconst [0])
   925
   926(SETBCR [0] (FlagLT)) => (MOVDconst [0])
   927(SETBCR [0] (Flag(GT|EQ))) => (MOVDconst [1])
   928(SETBCR [1] (FlagGT)) => (MOVDconst [0])
   929(SETBCR [1] (Flag(LT|EQ))) => (MOVDconst [1])
   930(SETBCR [2] (FlagEQ)) => (MOVDconst [0])
   931(SETBCR [2] (Flag(LT|GT))) => (MOVDconst [1])
   932
   933(SETBC [0] (InvertFlags bool)) => (SETBC [1] bool)
   934(SETBC [1] (InvertFlags bool)) => (SETBC [0] bool)
   935(SETBC [2] (InvertFlags bool)) => (SETBC [2] bool)
   936
   937(SETBCR [0] (InvertFlags bool)) => (SETBCR [1] bool)
   938(SETBCR [1] (InvertFlags bool)) => (SETBCR [0] bool)
   939(SETBCR [2] (InvertFlags bool)) => (SETBCR [2] bool)
   940
   941// ISEL auxInt values 0=LT 1=GT 2=EQ   arg2 ? arg0 : arg1
   942// ISEL auxInt values 4=GE 5=LE 6=NE   !arg2 ? arg1 : arg0
   943
   944(ISEL [2] x _ (FlagEQ)) => x
   945(ISEL [2] _ y (Flag(LT|GT))) => y
   946
   947(ISEL [6] _ y (FlagEQ)) => y
   948(ISEL [6] x _ (Flag(LT|GT))) => x
   949
   950(ISEL [0] _ y (Flag(EQ|GT))) => y
   951(ISEL [0] x _ (FlagLT)) => x
   952
   953(ISEL [5] _ x (Flag(EQ|LT))) => x
   954(ISEL [5] y _ (FlagGT)) => y
   955
   956(ISEL [1] _ y (Flag(EQ|LT))) => y
   957(ISEL [1] x _ (FlagGT)) => x
   958
   959(ISEL [4] x _ (Flag(EQ|GT))) => x
   960(ISEL [4] _ y (FlagLT)) => y
   961
   962(SETBC [n] (InvertFlags bool)) => (SETBCR [n] bool)
   963(SETBCR [n] (InvertFlags bool)) => (SETBC [n] bool)
   964
   965(ISEL [n] x y (InvertFlags bool)) && n%4 == 0 => (ISEL [n+1] x y bool)
   966(ISEL [n] x y (InvertFlags bool)) && n%4 == 1 => (ISEL [n-1] x y bool)
   967(ISEL [n] x y (InvertFlags bool)) && n%4 == 2 => (ISEL [n] x y bool)
   968(XORconst [1] (SETBCR [n] cmp)) => (SETBC [n] cmp)
   969(XORconst [1] (SETBC [n] cmp)) => (SETBCR [n] cmp)
   970
   971(SETBC  [2] (CMPconst [0] a:(ANDconst [1] _))) => (XORconst [1] a)
   972(SETBCR [2] (CMPconst [0] a:(ANDconst [1] _))) => a
   973
   974// Only CMPconst for these in case AND|OR|XOR result is > 32 bits
   975(SETBC [2] (CMPconst [0] a:(AND y z))) && a.Uses == 1 => (SETBC [2] (Select1 <types.TypeFlags> (ANDCC y z )))
   976(SETBCR [2] (CMPconst [0] a:(AND y z))) && a.Uses == 1 => (SETBCR [2] (Select1 <types.TypeFlags> (ANDCC y z )))
   977
   978(SETBC [2] (CMPconst [0] o:(OR y z))) && o.Uses == 1 => (SETBC [2] (Select1 <types.TypeFlags> (ORCC y z )))
   979(SETBCR [2] (CMPconst [0] o:(OR y z))) && o.Uses == 1 => (SETBCR [2] (Select1 <types.TypeFlags> (ORCC y z )))
   980
   981(SETBC [2] (CMPconst [0] a:(XOR y z))) && a.Uses == 1 => (SETBC [2] (Select1 <types.TypeFlags> (XORCC y z )))
   982(SETBCR [2] (CMPconst [0] a:(XOR y z))) && a.Uses == 1 => (SETBCR [2] (Select1 <types.TypeFlags> (XORCC y z )))
   983
   984// A particular pattern seen in cgo code:
   985(AND (MOVDconst [c]) x:(MOVBZload _ _)) => (ANDconst [c&0xFF] x)
   986
   987// floating point negative abs
   988(FNEG (F(ABS|NABS) x)) => (F(NABS|ABS) x)
   989
   990// floating-point fused multiply-add/sub
   991(F(ADD|SUB) (FMUL x y) z) && x.Block.Func.useFMA(v) => (FM(ADD|SUB) x y z)
   992(F(ADDS|SUBS) (FMULS x y) z) && x.Block.Func.useFMA(v) => (FM(ADDS|SUBS) x y z)
   993
   994// Arch-specific inlining for small or disjoint runtime.memmove
   995(SelectN [0] call:(CALLstatic {sym} s1:(MOVDstore _ (MOVDconst [sz]) s2:(MOVDstore  _ src s3:(MOVDstore {t} _ dst mem)))))
   996        && sz >= 0
   997        && isSameCall(sym, "runtime.memmove")
   998        && s1.Uses == 1 && s2.Uses == 1 && s3.Uses == 1
   999        && isInlinableMemmove(dst, src, sz, config)
  1000        && clobber(s1, s2, s3, call)
  1001        => (Move [sz] dst src mem)
  1002
  1003// Match post-lowering calls, register version.
  1004(SelectN [0] call:(CALLstatic {sym} dst src (MOVDconst [sz]) mem))
  1005        && sz >= 0
  1006        && isSameCall(sym, "runtime.memmove")
  1007        && call.Uses == 1
  1008        && isInlinableMemmove(dst, src, sz, config)
  1009        && clobber(call)
  1010        => (Move [sz] dst src mem)
  1011
  1012// Prefetch instructions (TH specified using aux field)
  1013// For DCBT Ra,Rb,TH, A value of TH indicates:
  1014//     0, hint this cache line will be used soon. (PrefetchCache)
  1015//     16, hint this cache line will not be used for long. (PrefetchCacheStreamed)
  1016// See ISA 3.0 Book II 4.3.2 for more detail. https://openpower.foundation/specifications/isa/
  1017(PrefetchCache ptr mem)          => (DCBT ptr mem [0])
  1018(PrefetchCacheStreamed ptr mem)  => (DCBT ptr mem [16])
  1019
  1020// Use byte reverse instructions on Power10
  1021(Bswap(16|32|64) x) && buildcfg.GOPPC64>=10 => (BR(H|W|D) x)
  1022
  1023// Fold bit reversal into loads.
  1024(BR(W|H) x:(MOV(W|H)Zload [off] {sym} ptr mem)) && x.Uses == 1 => @x.Block (MOV(W|H)BRload (MOVDaddr <ptr.Type> [off] {sym} ptr) mem)
  1025(BR(W|H) x:(MOV(W|H)Zloadidx ptr idx      mem)) && x.Uses == 1 => @x.Block (MOV(W|H)BRloadidx ptr idx mem)
  1026(BRD x:(MOVDload [off] {sym} ptr mem)) && x.Uses == 1 => @x.Block (MOVDBRload (MOVDaddr <ptr.Type> [off] {sym} ptr) mem)
  1027(BRD x:(MOVDloadidx ptr idx      mem)) && x.Uses == 1 => @x.Block (MOVDBRloadidx ptr idx mem)
  1028
  1029// Fold bit reversal into stores.
  1030(MOV(D|W|H)store [off] {sym} ptr r:(BR(D|W|H) val) mem) && r.Uses == 1 => (MOV(D|W|H)BRstore (MOVDaddr <ptr.Type> [off] {sym} ptr) val mem)
  1031(MOV(D|W|H)storeidx ptr idx      r:(BR(D|W|H) val) mem) && r.Uses == 1 => (MOV(D|W|H)BRstoreidx ptr idx val mem)
  1032
  1033// GOPPC64<10 rules.
  1034// These Bswap operations should only be introduced by the memcombine pass in places where they can be folded into loads or stores.
  1035(Bswap(32|16) x:(MOV(W|H)Zload [off] {sym} ptr mem)) => @x.Block (MOV(W|H)BRload (MOVDaddr <ptr.Type> [off] {sym} ptr) mem)
  1036(Bswap(32|16) x:(MOV(W|H)Zloadidx ptr idx      mem)) => @x.Block (MOV(W|H)BRloadidx ptr idx mem)
  1037(Bswap64 x:(MOVDload [off] {sym} ptr mem)) => @x.Block (MOVDBRload (MOVDaddr <ptr.Type> [off] {sym} ptr) mem)
  1038(Bswap64 x:(MOVDloadidx ptr idx      mem)) => @x.Block (MOVDBRloadidx ptr idx mem)
  1039(MOV(D|W|H)store [off] {sym} ptr (Bswap(64|32|16) val) mem) => (MOV(D|W|H)BRstore (MOVDaddr <ptr.Type> [off] {sym} ptr) val mem)
  1040(MOV(D|W|H)storeidx ptr idx      (Bswap(64|32|16) val) mem) => (MOV(D|W|H)BRstoreidx ptr idx val mem)

View as plain text