...

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