Text file src/cmd/compile/internal/ssa/_gen/AMD64.rules

     1  // Copyright 2015 The Go Authors. All rights reserved.
     2  // Use of this source code is governed by a BSD-style
     3  // license that can be found in the LICENSE file.
     4  
     5  // Lowering arithmetic
     6  (Add(64|32|16|8) ...) => (ADD(Q|L|L|L) ...)
     7  (AddPtr ...) => (ADDQ ...)
     8  (Add(32|64)F ...) => (ADDS(S|D) ...)
     9  
    10  (Sub(64|32|16|8) ...) => (SUB(Q|L|L|L) ...)
    11  (SubPtr ...) => (SUBQ ...)
    12  (Sub(32|64)F ...) => (SUBS(S|D) ...)
    13  
    14  (Mul(64|32|16|8) ...) => (MUL(Q|L|L|L) ...)
    15  (Mul(32|64)F ...) => (MULS(S|D) ...)
    16  
    17  (Select0 (Mul64uover x y)) => (Select0 <typ.UInt64> (MULQU x y))
    18  (Select0 (Mul32uover x y)) => (Select0 <typ.UInt32> (MULLU x y))
    19  (Select1 (Mul(64|32)uover x y)) => (SETO (Select1 <types.TypeFlags> (MUL(Q|L)U x y)))
    20  
    21  (Hmul(64|32) ...) => (HMUL(Q|L) ...)
    22  (Hmul(64|32)u ...) => (HMUL(Q|L)U ...)
    23  
    24  (Div(64|32|16) [a] x y) => (Select0 (DIV(Q|L|W) [a] x y))
    25  (Div8  x y) => (Select0 (DIVW  (SignExt8to16 x) (SignExt8to16 y)))
    26  (Div(64|32|16)u x y) => (Select0 (DIV(Q|L|W)U x y))
    27  (Div8u x y) => (Select0 (DIVWU (ZeroExt8to16 x) (ZeroExt8to16 y)))
    28  (Div(32|64)F ...) => (DIVS(S|D) ...)
    29  
    30  (Select0 (Add64carry x y c)) => (Select0 (ADCQ x y (Select1 <types.TypeFlags> (NEGLflags c))))
    31  (Select1 (Add64carry x y c)) => (MOVBQZX (SETB <types.Types[types.TUINT8]> (Select1 <types.TypeFlags> (ADCQ x y (Select1 <types.TypeFlags> (NEGLflags c))))))
    32  (Select0 (Sub64borrow x y c)) => (Select0 (SBBQ x y (Select1 <types.TypeFlags> (NEGLflags c))))
    33  (Select1 (Sub64borrow x y c)) => (MOVBQZX (SETB <types.Types[types.TUINT8]> (Select1 <types.TypeFlags> (SBBQ x y (Select1 <types.TypeFlags> (NEGLflags c))))))
    34  // Optimize ADCQ and friends
    35  (ADCQ x (MOVQconst [c]) carry) && ssa.Is32Bit(c) => (ADCQconst x [int32(c)] carry)
    36  (ADCQ x y (FlagEQ)) => (ADDQcarry x y)
    37  (ADCQconst x [c] (FlagEQ)) => (ADDQconstcarry x [c])
    38  (ADDQcarry x (MOVQconst [c])) && ssa.Is32Bit(c) => (ADDQconstcarry x [int32(c)])
    39  (SBBQ x (MOVQconst [c]) borrow) && ssa.Is32Bit(c) => (SBBQconst x [int32(c)] borrow)
    40  (SBBQ x y (FlagEQ)) => (SUBQborrow x y)
    41  (SBBQconst x [c] (FlagEQ)) => (SUBQconstborrow x [c])
    42  (SUBQborrow x (MOVQconst [c])) && ssa.Is32Bit(c) => (SUBQconstborrow x [int32(c)])
    43  (Select1 (NEGLflags (MOVQconst [0]))) => (FlagEQ)
    44  (Select1 (NEGLflags (MOVBQZX (SETB x)))) => x
    45  // Absorb InvertFlags into ADCQ/SBBQ: CF(InvertFlags(f)) = SETA(f), so
    46  // re-materialize the carry/borrow through the SETA+NEGLflags path.
    47  (ADCQ x y (InvertFlags f)) => (ADCQ x y (Select1 <types.TypeFlags> (NEGLflags (MOVBQZX <types.Types[types.TUINT32]> (SETA <types.Types[types.TUINT8]> f)))))
    48  (ADCQconst x [c] (InvertFlags f)) => (ADCQconst x [c] (Select1 <types.TypeFlags> (NEGLflags (MOVBQZX <types.Types[types.TUINT32]> (SETA <types.Types[types.TUINT8]> f)))))
    49  (SBBQ x y (InvertFlags f)) => (SBBQ x y (Select1 <types.TypeFlags> (NEGLflags (MOVBQZX <types.Types[types.TUINT32]> (SETA <types.Types[types.TUINT8]> f)))))
    50  (SBBQconst x [c] (InvertFlags f)) => (SBBQconst x [c] (Select1 <types.TypeFlags> (NEGLflags (MOVBQZX <types.Types[types.TUINT32]> (SETA <types.Types[types.TUINT8]> f)))))
    51  // ADDQ/SUBQ an from a carry flag into ADCQ/SBBQ
    52  // TODO: maybe add ADCL and SBBL ?
    53  (ADDQ x (MOVBQZX (SETB flags))) => (Select0 (ADCQconst [0] x flags))
    54  (SUBQ x (MOVBQZX (SETB flags))) => (Select0 (SBBQconst [0] x flags))
    55  
    56  (Mul64uhilo x y) && buildcfg.GOAMD64 >= 3 => (MULXQ x y)
    57  (Mul64uhilo x y) && buildcfg.GOAMD64 <  3 => (MULQU2 x y)
    58  (Div128u ...) => (DIVQU2 ...)
    59  
    60  (Avg64u ...) => (AVGQU ...)
    61  
    62  (Mod(64|32|16) [a] x y) => (Select1 (DIV(Q|L|W) [a] x y))
    63  (Mod8  x y) => (Select1 (DIVW  (SignExt8to16 x) (SignExt8to16 y)))
    64  (Mod(64|32|16)u x y) => (Select1 (DIV(Q|L|W)U x y))
    65  (Mod8u x y) => (Select1 (DIVWU (ZeroExt8to16 x) (ZeroExt8to16 y)))
    66  
    67  (And(64|32|16|8) ...) => (AND(Q|L|L|L) ...)
    68  (Or(64|32|16|8) ...) => (OR(Q|L|L|L) ...)
    69  (Xor(64|32|16|8) ...) => (XOR(Q|L|L|L) ...)
    70  (Com(64|32|16|8) ...) => (NOT(Q|L|L|L) ...)
    71  
    72  (Neg(64|32|16|8) ...) => (NEG(Q|L|L|L) ...)
    73  (Neg32F x) => (PXOR x (MOVSSconst <typ.Float32> [float32(math.Copysign(0, -1))]))
    74  (Neg64F x) => (PXOR x (MOVSDconst <typ.Float64> [math.Copysign(0, -1)]))
    75  
    76  // Lowering boolean ops
    77  (AndB ...) => (ANDL ...)
    78  (OrB ...) => (ORL ...)
    79  (Not x) => (XORLconst [1] x)
    80  
    81  // Lowering pointer arithmetic
    82  (OffPtr [off] ptr) && ssa.Is32Bit(off) => (ADDQconst [int32(off)] ptr)
    83  (OffPtr [off] ptr) => (ADDQ (MOVQconst [off]) ptr)
    84  
    85  // Lowering other arithmetic
    86  (Ctz64 x)     && buildcfg.GOAMD64 >= 3 => (TZCNTQ x)
    87  (Ctz32 x)     && buildcfg.GOAMD64 >= 3 => (TZCNTL x)
    88  (Ctz64 <t> x) && buildcfg.GOAMD64 <  3 => (CMOVQEQ (Select0 <t> (BSFQ x)) (MOVQconst <t> [64]) (Select1 <types.TypeFlags> (BSFQ x)))
    89  (Ctz32 x)     && buildcfg.GOAMD64 <  3 => (Select0 (BSFQ (BTSQconst <typ.UInt64> [32] x)))
    90  (Ctz16 x) => (BSFL (ORLconst <typ.UInt32> [1<<16] x))
    91  (Ctz8  x) => (BSFL (ORLconst <typ.UInt32> [1<<8 ] x))
    92  
    93  (Ctz64NonZero x) && buildcfg.GOAMD64 >= 3 => (TZCNTQ x)
    94  (Ctz32NonZero x) && buildcfg.GOAMD64 >= 3 => (TZCNTL x)
    95  (Ctz16NonZero x) && buildcfg.GOAMD64 >= 3 => (TZCNTL x)
    96  (Ctz8NonZero  x) && buildcfg.GOAMD64 >= 3 => (TZCNTL x)
    97  (Ctz64NonZero x) && buildcfg.GOAMD64 <  3 => (Select0 (BSFQ x))
    98  (Ctz32NonZero x) && buildcfg.GOAMD64 <  3 => (BSFL x)
    99  (Ctz16NonZero x) && buildcfg.GOAMD64 <  3 => (BSFL x)
   100  (Ctz8NonZero  x) && buildcfg.GOAMD64 <  3 => (BSFL x)
   101  
   102  // BitLen64 of a 64 bit value x requires checking whether x == 0, since BSRQ is undefined when x == 0.
   103  // However, for zero-extended values, we can cheat a bit, and calculate
   104  // BSR(x<<1 + 1), which is guaranteed to be non-zero, and which conveniently
   105  // places the index of the highest set bit where we want it.
   106  // For GOAMD64>=3, BitLen can be calculated by OperandSize - LZCNT(x).
   107  (BitLen64 <t> x) && buildcfg.GOAMD64 < 3 => (ADDQconst [1] (CMOVQEQ <t> (Select0 <t> (BSRQ x)) (MOVQconst <t> [-1]) (Select1 <types.TypeFlags> (BSRQ x))))
   108  (BitLen32 x) && buildcfg.GOAMD64 <  3 => (Select0 (BSRQ (LEAQ1 <typ.UInt64> [1] (MOVLQZX <typ.UInt64> x) (MOVLQZX <typ.UInt64> x))))
   109  (BitLen16 x) && buildcfg.GOAMD64 <  3 => (BSRL (LEAL1 <typ.UInt32> [1] (MOVWQZX <typ.UInt32> x) (MOVWQZX <typ.UInt32> x)))
   110  (BitLen8  x) && buildcfg.GOAMD64 <  3 => (BSRL (LEAL1 <typ.UInt32> [1] (MOVBQZX <typ.UInt32> x) (MOVBQZX <typ.UInt32> x)))
   111  (BitLen64 <t> x)        && buildcfg.GOAMD64 >= 3 => (NEGQ (ADDQconst <t> [-64] (LZCNTQ x)))
   112  // Use 64-bit version to allow const-fold remove unnecessary arithmetic.
   113  (BitLen32 <t> x) && buildcfg.GOAMD64 >= 3 => (NEGQ (ADDQconst <t> [-32] (LZCNTL x)))
   114  (BitLen16 <t> x) && buildcfg.GOAMD64 >= 3 => (NEGQ (ADDQconst <t> [-32] (LZCNTL (MOVWQZX <x.Type> x))))
   115  (BitLen8 <t> x) && buildcfg.GOAMD64 >= 3 => (NEGQ (ADDQconst <t> [-32] (LZCNTL (MOVBQZX <x.Type> x))))
   116  
   117  (Bswap(64|32) ...) => (BSWAP(Q|L) ...)
   118  (Bswap16 x) => (ROLWconst [8] x)
   119  
   120  (PopCount(64|32) ...) => (POPCNT(Q|L) ...)
   121  (PopCount16 x) => (POPCNTL (MOVWQZX <typ.UInt32> x))
   122  (PopCount8 x) => (POPCNTL (MOVBQZX <typ.UInt32> x))
   123  
   124  (Sqrt ...) => (SQRTSD ...)
   125  (Sqrt32 ...) => (SQRTSS ...)
   126  
   127  (RoundToEven x) => (ROUNDSD [0] x)
   128  (Floor x)       => (ROUNDSD [1] x)
   129  (Ceil x)        => (ROUNDSD [2] x)
   130  (Trunc x)       => (ROUNDSD [3] x)
   131  
   132  (CVTSD2SS (ROUNDSD [c] (CVTSS2SD x))) => (ROUNDSS [c] x)
   133  
   134  (FMA x y z) => (VFMADD231SD z x y)
   135  
   136  // Lowering extension
   137  // Note: we always extend to 64 bits even though some ops don't need that many result bits.
   138  (SignExt8to16  ...) => (MOVBQSX ...)
   139  (SignExt8to32  ...) => (MOVBQSX ...)
   140  (SignExt8to64  ...) => (MOVBQSX ...)
   141  (SignExt16to32 ...) => (MOVWQSX ...)
   142  (SignExt16to64 ...) => (MOVWQSX ...)
   143  (SignExt32to64 ...) => (MOVLQSX ...)
   144  
   145  (ZeroExt8to16  ...) => (MOVBQZX ...)
   146  (ZeroExt8to32  ...) => (MOVBQZX ...)
   147  (ZeroExt8to64  ...) => (MOVBQZX ...)
   148  (ZeroExt16to32 ...) => (MOVWQZX ...)
   149  (ZeroExt16to64 ...) => (MOVWQZX ...)
   150  (ZeroExt32to64 ...) => (MOVLQZX ...)
   151  
   152  (Slicemask <t> x) => (SARQconst (NEGQ <t> x) [63])
   153  
   154  (SpectreIndex <t> x y) => (CMOVQCC x (MOVQconst [0]) (CMPQ x y))
   155  (SpectreSliceIndex <t> x y) => (CMOVQHI x (MOVQconst [0]) (CMPQ x y))
   156  
   157  // Lowering truncation
   158  // Because we ignore high parts of registers, truncates are just copies.
   159  (Trunc16to8  ...) => (Copy ...)
   160  (Trunc32to8  ...) => (Copy ...)
   161  (Trunc32to16 ...) => (Copy ...)
   162  (Trunc64to8  ...) => (Copy ...)
   163  (Trunc64to16 ...) => (Copy ...)
   164  (Trunc64to32 ...) => (Copy ...)
   165  
   166  // Lowering float <-> int
   167  (Cvt32to32F ...) => (CVTSL2SS ...)
   168  (Cvt32to64F ...) => (CVTSL2SD ...)
   169  (Cvt64to32F ...) => (CVTSQ2SS ...)
   170  (Cvt64to64F ...) => (CVTSQ2SD ...)
   171  
   172  // Float, to int.
   173  // To make AMD64 "overflow" return max positive instead of max negative, compute
   174  // y and not x, smear the sign bit, and xor.
   175  (Cvt32Fto32 <t> x) && base.ConvertHash.MatchPos(v.Pos, nil) => (XORL <t> y (SARLconst <t> [31] (ANDL <t> y:(CVTTSS2SL <t> x) (NOTL <typ.Int32> (MOVLf2i x)))))
   176  (Cvt64Fto32 <t> x) && base.ConvertHash.MatchPos(v.Pos, nil) => (XORL <t> y (SARLconst <t> [31] (ANDL <t> y:(CVTTSD2SL <t> x) (NOTL <typ.Int32> (MOVLf2i (CVTSD2SS <typ.Float32> x))))))
   177  
   178  (Cvt32Fto64 <t> x) && base.ConvertHash.MatchPos(v.Pos, nil) => (XORQ <t> y (SARQconst <t> [63] (ANDQ <t> y:(CVTTSS2SQ <t> x) (NOTQ <typ.Int64> (MOVQf2i (CVTSS2SD <typ.Float64> x))) )))
   179  (Cvt64Fto64 <t> x) && base.ConvertHash.MatchPos(v.Pos, nil) => (XORQ <t> y (SARQconst <t> [63] (ANDQ <t> y:(CVTTSD2SQ <t> x) (NOTQ <typ.Int64> (MOVQf2i x)))))
   180  
   181  (Cvt32Fto32 <t> x) && !base.ConvertHash.MatchPos(v.Pos, nil) => (CVTTSS2SL <t> x)
   182  (Cvt32Fto64 <t> x) && !base.ConvertHash.MatchPos(v.Pos, nil) => (CVTTSS2SQ <t> x)
   183  (Cvt64Fto32 <t> x) && !base.ConvertHash.MatchPos(v.Pos, nil) => (CVTTSD2SL <t> x)
   184  (Cvt64Fto64 <t> x) && !base.ConvertHash.MatchPos(v.Pos, nil) => (CVTTSD2SQ <t> x)
   185  
   186  (Cvt32Fto64F ...) => (CVTSS2SD ...)
   187  (Cvt64Fto32F ...) => (CVTSD2SS ...)
   188  
   189  (Round(32|64)F ...) => (LoweredRound(32|64)F ...)
   190  
   191  // Floating-point min is tricky, as the hardware op isn't right for various special
   192  // cases (-0 and NaN). We use two hardware ops organized just right to make the
   193  // result come out how we want it. See https://github.com/golang/go/issues/59488#issuecomment-1553493207
   194  // (although that comment isn't exactly right, as the value overwritten is not simulated correctly).
   195  //    t1 = MINSD x, y   => incorrect if x==NaN or x==-0,y==+0
   196  //    t2 = MINSD t1, x  => fixes x==NaN case
   197  //   res = POR t1, t2   => fixes x==-0,y==+0 case
   198  // Note that this trick depends on the special property that (NaN OR x) produces a NaN (although
   199  // it might not produce the same NaN as the input).
   200  (Min(64|32)F <t> x y) => (POR (MINS(D|S) <t> (MINS(D|S) <t> x y) x) (MINS(D|S) <t> x y))
   201  // Floating-point max is even trickier. Punt to using min instead.
   202  // max(x,y) == -min(-x,-y)
   203  (Max(64|32)F <t> x y) => (Neg(64|32)F <t> (Min(64|32)F <t> (Neg(64|32)F <t> x) (Neg(64|32)F <t> y)))
   204  
   205  (CvtBoolToUint8 ...) => (Copy ...)
   206  
   207  // Lowering shifts
   208  // Unsigned shifts need to return 0 if shift amount is >= width of shifted value.
   209  //   result = (arg << shift) & (shift >= argbits ? 0 : 0xffffffffffffffff)
   210  (Lsh64x(64|32|16|8) <t> x y) && !ssa.ShiftIsBounded(v) => (ANDQ (SHLQ <t> x y) (SBBQcarrymask <t> (CMP(Q|L|W|B)const y [64])))
   211  (Lsh32x(64|32|16|8) <t> x y) && !ssa.ShiftIsBounded(v) => (ANDL (SHLL <t> x y) (SBBLcarrymask <t> (CMP(Q|L|W|B)const y [32])))
   212  (Lsh16x(64|32|16|8) <t> x y) && !ssa.ShiftIsBounded(v) => (ANDL (SHLL <t> x y) (SBBLcarrymask <t> (CMP(Q|L|W|B)const y [32])))
   213  (Lsh8x(64|32|16|8)  <t> x y) && !ssa.ShiftIsBounded(v) => (ANDL (SHLL <t> x y) (SBBLcarrymask <t> (CMP(Q|L|W|B)const y [32])))
   214  
   215  (Lsh64x(64|32|16|8) x y) && ssa.ShiftIsBounded(v) => (SHLQ x y)
   216  (Lsh32x(64|32|16|8) x y) && ssa.ShiftIsBounded(v) => (SHLL x y)
   217  (Lsh16x(64|32|16|8) x y) && ssa.ShiftIsBounded(v) => (SHLL x y)
   218  (Lsh8x(64|32|16|8)  x y) && ssa.ShiftIsBounded(v) => (SHLL x y)
   219  
   220  (Rsh64Ux(64|32|16|8) <t> x y) && !ssa.ShiftIsBounded(v) => (ANDQ (SHRQ <t> x y) (SBBQcarrymask <t> (CMP(Q|L|W|B)const y [64])))
   221  (Rsh32Ux(64|32|16|8) <t> x y) && !ssa.ShiftIsBounded(v) => (ANDL (SHRL <t> x y) (SBBLcarrymask <t> (CMP(Q|L|W|B)const y [32])))
   222  (Rsh16Ux(64|32|16|8) <t> x y) && !ssa.ShiftIsBounded(v) => (ANDL (SHRW <t> x y) (SBBLcarrymask <t> (CMP(Q|L|W|B)const y [16])))
   223  (Rsh8Ux(64|32|16|8)  <t> x y) && !ssa.ShiftIsBounded(v) => (ANDL (SHRB <t> x y) (SBBLcarrymask <t> (CMP(Q|L|W|B)const y [8])))
   224  
   225  (Rsh64Ux(64|32|16|8) x y) && ssa.ShiftIsBounded(v) => (SHRQ x y)
   226  (Rsh32Ux(64|32|16|8) x y) && ssa.ShiftIsBounded(v) => (SHRL x y)
   227  (Rsh16Ux(64|32|16|8) x y) && ssa.ShiftIsBounded(v) => (SHRW x y)
   228  (Rsh8Ux(64|32|16|8)  x y) && ssa.ShiftIsBounded(v) => (SHRB x y)
   229  
   230  // Signed right shift needs to return 0/-1 if shift amount is >= width of shifted value.
   231  // We implement this by setting the shift value to -1 (all ones) if the shift value is >= width.
   232  (Rsh64x(64|32|16|8) <t> x y) && !ssa.ShiftIsBounded(v) => (SARQ <t> x (OR(Q|L|L|L) <y.Type> y (NOT(Q|L|L|L) <y.Type> (SBB(Q|L|L|L)carrymask <y.Type> (CMP(Q|L|W|B)const y [64])))))
   233  (Rsh32x(64|32|16|8) <t> x y) && !ssa.ShiftIsBounded(v) => (SARL <t> x (OR(Q|L|L|L) <y.Type> y (NOT(Q|L|L|L) <y.Type> (SBB(Q|L|L|L)carrymask <y.Type> (CMP(Q|L|W|B)const y [32])))))
   234  (Rsh16x(64|32|16|8) <t> x y) && !ssa.ShiftIsBounded(v) => (SARW <t> x (OR(Q|L|L|L) <y.Type> y (NOT(Q|L|L|L) <y.Type> (SBB(Q|L|L|L)carrymask <y.Type> (CMP(Q|L|W|B)const y [16])))))
   235  (Rsh8x(64|32|16|8)  <t> x y) && !ssa.ShiftIsBounded(v) => (SARB <t> x (OR(Q|L|L|L) <y.Type> y (NOT(Q|L|L|L) <y.Type> (SBB(Q|L|L|L)carrymask <y.Type> (CMP(Q|L|W|B)const y [8])))))
   236  
   237  (Rsh64x(64|32|16|8) x y) && ssa.ShiftIsBounded(v) => (SARQ x y)
   238  (Rsh32x(64|32|16|8) x y) && ssa.ShiftIsBounded(v) => (SARL x y)
   239  (Rsh16x(64|32|16|8) x y) && ssa.ShiftIsBounded(v) => (SARW x y)
   240  (Rsh8x(64|32|16|8) x y)  && ssa.ShiftIsBounded(v) => (SARB x y)
   241  
   242  // Lowering integer comparisons
   243  (Less(64|32|16|8)      x y) => (SETL  (CMP(Q|L|W|B)     x y))
   244  (Less(64|32|16|8)U     x y) => (SETB  (CMP(Q|L|W|B)     x y))
   245  (Leq(64|32|16|8)       x y) => (SETLE (CMP(Q|L|W|B)     x y))
   246  (Leq(64|32|16|8)U      x y) => (SETBE (CMP(Q|L|W|B)     x y))
   247  (Eq(Ptr|64|32|16|8|B)  x y) => (SETEQ (CMP(Q|Q|L|W|B|B) x y))
   248  (Neq(Ptr|64|32|16|8|B) x y) => (SETNE (CMP(Q|Q|L|W|B|B) x y))
   249  
   250  // Lowering floating point comparisons
   251  // Note Go assembler gets UCOMISx operand order wrong, but it is right here
   252  // and the operands are reversed when generating assembly language.
   253  (Eq(32|64)F   x y) => (SETEQF (UCOMIS(S|D) x y))
   254  (Neq(32|64)F  x y) => (SETNEF (UCOMIS(S|D) x y))
   255  // Use SETGF/SETGEF with reversed operands to dodge NaN case.
   256  (Less(32|64)F x y) => (SETGF  (UCOMIS(S|D) y x))
   257  (Leq(32|64)F  x y) => (SETGEF (UCOMIS(S|D) y x))
   258  
   259  // Lowering loads
   260  (Load <t> ptr mem) && (ssa.Is64BitInt(t) ||ssa.IsPtr(t)) => (MOVQload ptr mem)
   261  (Load <t> ptr mem) && ssa.Is32BitInt(t) => (MOVLload ptr mem)
   262  (Load <t> ptr mem) && ssa.Is16BitInt(t) => (MOVWload ptr mem)
   263  (Load <t> ptr mem) && (t.IsBoolean() || ssa.Is8BitInt(t)) => (MOVBload ptr mem)
   264  (Load <t> ptr mem) && ssa.Is32BitFloat(t) => (MOVSSload ptr mem)
   265  (Load <t> ptr mem) && ssa.Is64BitFloat(t) => (MOVSDload ptr mem)
   266  
   267  // Lowering stores
   268  (Store {t} ptr val mem) && t.Size() == 8 &&  t.IsFloat() => (MOVSDstore ptr val mem)
   269  (Store {t} ptr val mem) && t.Size() == 4 &&  t.IsFloat() => (MOVSSstore ptr val mem)
   270  (Store {t} ptr val mem) && t.Size() == 8 && !t.IsFloat() => (MOVQstore ptr val mem)
   271  (Store {t} ptr val mem) && t.Size() == 4 && !t.IsFloat() => (MOVLstore ptr val mem)
   272  (Store {t} ptr val mem) && t.Size() == 2 => (MOVWstore ptr val mem)
   273  (Store {t} ptr val mem) && t.Size() == 1 => (MOVBstore ptr val mem)
   274  
   275  // Lowering moves
   276  (Move [0] _ _ mem) => mem
   277  (Move [1] dst src mem) => (MOVBstore dst (MOVBload src mem) mem)
   278  (Move [2] dst src mem) => (MOVWstore dst (MOVWload src mem) mem)
   279  (Move [4] dst src mem) => (MOVLstore dst (MOVLload src mem) mem)
   280  (Move [8] dst src mem) => (MOVQstore dst (MOVQload src mem) mem)
   281  (Move [16] dst src mem) => (MOVOstore dst (MOVOload src mem) mem)
   282  
   283  (Move [3] dst src mem) =>
   284  	(MOVBstore [2] dst (MOVBload [2] src mem)
   285  		(MOVWstore dst (MOVWload src mem) mem))
   286  (Move [5] dst src mem) =>
   287  	(MOVBstore [4] dst (MOVBload [4] src mem)
   288  		(MOVLstore dst (MOVLload src mem) mem))
   289  (Move [6] dst src mem) =>
   290  	(MOVWstore [4] dst (MOVWload [4] src mem)
   291  		(MOVLstore dst (MOVLload src mem) mem))
   292  (Move [7] dst src mem) =>
   293  	(MOVLstore [3] dst (MOVLload [3] src mem)
   294  		(MOVLstore dst (MOVLload src mem) mem))
   295  (Move [9] dst src mem) =>
   296  	(MOVBstore [8] dst (MOVBload [8] src mem)
   297  		(MOVQstore dst (MOVQload src mem) mem))
   298  (Move [10] dst src mem) =>
   299  	(MOVWstore [8] dst (MOVWload [8] src mem)
   300  		(MOVQstore dst (MOVQload src mem) mem))
   301  (Move [11] dst src mem) =>
   302  	(MOVLstore [7] dst (MOVLload [7] src mem)
   303  		(MOVQstore dst (MOVQload src mem) mem))
   304  (Move [12] dst src mem) =>
   305  	(MOVLstore [8] dst (MOVLload [8] src mem)
   306  		(MOVQstore dst (MOVQload src mem) mem))
   307  (Move [s] dst src mem) && s >= 13 && s <= 15 =>
   308  	(MOVQstore [int32(s-8)] dst (MOVQload [int32(s-8)] src mem)
   309  		(MOVQstore dst (MOVQload src mem) mem))
   310  
   311  // Copying up to 192 bytes uses straightline code.
   312  (Move [s] dst src mem) && s > 16 && s < 192 && ssa.LogLargeCopyValue(v, s) => (LoweredMove [s] dst src mem)
   313  
   314  // Copying up to ~1KB uses a small loop.
   315  (Move [s] dst src mem) && s >= 192 && s <= ssa.RepMoveThreshold && ssa.LogLargeCopyValue(v, s) => (LoweredMoveLoop [s] dst src mem)
   316  
   317  // Large copying uses REP MOVSQ.
   318  (Move [s] dst src mem) && s > ssa.RepMoveThreshold && s%8 != 0 =>
   319  	(Move [s-s%8]
   320  		(OffPtr <dst.Type> dst [s%8])
   321  		(OffPtr <src.Type> src [s%8])
   322  		(MOVQstore dst (MOVQload src mem) mem))
   323  (Move [s] dst src mem) && s > ssa.RepMoveThreshold && s%8 == 0 && ssa.LogLargeCopyValue(v, s) =>
   324  	(REPMOVSQ dst src (MOVQconst [s/8]) mem)
   325  
   326  // Lowering Zero instructions
   327  (Zero [0] _ mem) => mem
   328  (Zero [1] destptr mem) => (MOVBstoreconst [ssa.MakeValAndOff(0,0)] destptr mem)
   329  (Zero [2] destptr mem) => (MOVWstoreconst [ssa.MakeValAndOff(0,0)] destptr mem)
   330  (Zero [4] destptr mem) => (MOVLstoreconst [ssa.MakeValAndOff(0,0)] destptr mem)
   331  (Zero [8] destptr mem) => (MOVQstoreconst [ssa.MakeValAndOff(0,0)] destptr mem)
   332  
   333  (Zero [3] destptr mem) =>
   334  	(MOVBstoreconst [ssa.MakeValAndOff(0,2)] destptr
   335  		(MOVWstoreconst [ssa.MakeValAndOff(0,0)] destptr mem))
   336  (Zero [5] destptr mem) =>
   337  	(MOVBstoreconst [ssa.MakeValAndOff(0,4)] destptr
   338  		(MOVLstoreconst [ssa.MakeValAndOff(0,0)] destptr mem))
   339  (Zero [6] destptr mem) =>
   340  	(MOVWstoreconst [ssa.MakeValAndOff(0,4)] destptr
   341  		(MOVLstoreconst [ssa.MakeValAndOff(0,0)] destptr mem))
   342  (Zero [7] destptr mem) =>
   343  	(MOVLstoreconst [ssa.MakeValAndOff(0,3)] destptr
   344  		(MOVLstoreconst [ssa.MakeValAndOff(0,0)] destptr mem))
   345  
   346  // Zero small numbers of words directly.
   347  (Zero [9] destptr mem) =>
   348  	(MOVBstoreconst [ssa.MakeValAndOff(0,8)] destptr
   349  		(MOVQstoreconst [ssa.MakeValAndOff(0,0)] destptr mem))
   350  
   351  (Zero [10] destptr mem) =>
   352  	(MOVWstoreconst [ssa.MakeValAndOff(0,8)] destptr
   353  		(MOVQstoreconst [ssa.MakeValAndOff(0,0)] destptr mem))
   354  
   355  (Zero [11] destptr mem) =>
   356  	(MOVLstoreconst [ssa.MakeValAndOff(0,7)] destptr
   357  		(MOVQstoreconst [ssa.MakeValAndOff(0,0)] destptr mem))
   358  
   359  (Zero [12] destptr mem) =>
   360  	(MOVLstoreconst [ssa.MakeValAndOff(0,8)] destptr
   361  		(MOVQstoreconst [ssa.MakeValAndOff(0,0)] destptr mem))
   362  
   363  (Zero [s] destptr mem) && s > 12 && s < 16 =>
   364  	(MOVQstoreconst [ssa.MakeValAndOff(0,int32(s-8))] destptr
   365  		(MOVQstoreconst [ssa.MakeValAndOff(0,0)] destptr mem))
   366  
   367  // Zeroing up to 192 bytes uses straightline code.
   368  (Zero [s] destptr mem)	&& s >= 16 && s < 192 => (LoweredZero [s] destptr mem)
   369  
   370  // Zeroing up to ~1KB uses a small loop.
   371  (Zero [s] destptr mem)	&& s >= 192 && s <= ssa.RepZeroThreshold => (LoweredZeroLoop [s] destptr mem)
   372  
   373  // Large zeroing uses REP STOSQ.
   374  (Zero [s] destptr mem) && s > ssa.RepZeroThreshold && s%8 != 0 =>
   375         (Zero [s-s%8] (OffPtr <destptr.Type> destptr [s%8])
   376                 (MOVOstoreconst [ssa.MakeValAndOff(0,0)] destptr mem))
   377  (Zero [s] destptr mem) && s > ssa.RepZeroThreshold && s%8 == 0 =>
   378  	(REPSTOSQ destptr (MOVQconst [s/8]) (MOVQconst [0]) mem)
   379  
   380  // Lowering constants
   381  (Const8   [c]) => (MOVLconst [int32(c)])
   382  (Const16  [c]) => (MOVLconst [int32(c)])
   383  (Const32  ...) => (MOVLconst ...)
   384  (Const64  ...) => (MOVQconst ...)
   385  (Const32F ...) => (MOVSSconst ...)
   386  (Const64F ...) => (MOVSDconst ...)
   387  (ConstNil    ) => (MOVQconst [0])
   388  (ConstBool [c]) => (MOVLconst [ssa.B2i32(c)])
   389  
   390  // Lowering calls
   391  (StaticCall ...) => (CALLstatic ...)
   392  (ClosureCall ...) => (CALLclosure ...)
   393  (InterCall ...) => (CALLinter ...)
   394  (TailCall ...) => (CALLtail ...)
   395  (TailCallInter ...) => (CALLtailinter ...)
   396  
   397  // Lowering conditional moves
   398  // If the condition is a SETxx, we can just run a CMOV from the comparison that was
   399  // setting the flags.
   400  // Legend: HI=unsigned ABOVE, CS=unsigned BELOW, CC=unsigned ABOVE EQUAL, LS=unsigned BELOW EQUAL
   401  (CondSelect <t> x y (SET(EQ|NE|L|G|LE|GE|A|B|AE|BE|EQF|NEF|GF|GEF) cond)) && (ssa.Is64BitInt(t) ||ssa.IsPtr(t))
   402      => (CMOVQ(EQ|NE|LT|GT|LE|GE|HI|CS|CC|LS|EQF|NEF|GTF|GEF) y x cond)
   403  (CondSelect <t> x y (SET(EQ|NE|L|G|LE|GE|A|B|AE|BE|EQF|NEF|GF|GEF) cond)) && ssa.Is32BitInt(t)
   404      => (CMOVL(EQ|NE|LT|GT|LE|GE|HI|CS|CC|LS|EQF|NEF|GTF|GEF) y x cond)
   405  (CondSelect <t> x y (SET(EQ|NE|L|G|LE|GE|A|B|AE|BE|EQF|NEF|GF|GEF) cond)) && ssa.Is16BitInt(t)
   406      => (CMOVW(EQ|NE|LT|GT|LE|GE|HI|CS|CC|LS|EQF|NEF|GTF|GEF) y x cond)
   407  
   408  // Float min/max comparison-select idioms lower to a single instruction.
   409  // These ops carry their semantics directly (no embedded comparison), so the
   410  // lowering always triggers.
   411  (Min64FSel ...) => (MINSD ...)
   412  (Min32FSel ...) => (MINSS ...)
   413  (Max64FSel ...) => (MAXSD ...)
   414  (Max32FSel ...) => (MAXSS ...)
   415  
   416  (CondSelect <t> x y check) && !check.Type.IsFlags() && check.Type.Size() == 8 && (ssa.Is64BitInt(t) ||ssa.IsPtr(t))
   417      => (CMOVQNE y x (CMPQconst [0] check))
   418  (CondSelect <t> x y check) && !check.Type.IsFlags() && check.Type.Size() == 8 && ssa.Is32BitInt(t)
   419      => (CMOVLNE y x (CMPQconst [0] check))
   420  (CondSelect <t> x y check) && !check.Type.IsFlags() && check.Type.Size() == 8 && ssa.Is16BitInt(t)
   421      => (CMOVWNE y x (CMPQconst [0] check))
   422  (CondSelect <t> x y check) && !check.Type.IsFlags() && check.Type.Size() == 4 && (ssa.Is64BitInt(t) ||ssa.IsPtr(t))
   423      => (CMOVQNE y x (CMPLconst [0] check))
   424  (CondSelect <t> x y check) && !check.Type.IsFlags() && check.Type.Size() == 4 && ssa.Is32BitInt(t)
   425      => (CMOVLNE y x (CMPLconst [0] check))
   426  (CondSelect <t> x y check) && !check.Type.IsFlags() && check.Type.Size() == 4 && ssa.Is16BitInt(t)
   427      => (CMOVWNE y x (CMPLconst [0] check))
   428  (CondSelect <t> x y check) && !check.Type.IsFlags() && check.Type.Size() == 2 && (ssa.Is64BitInt(t) ||ssa.IsPtr(t))
   429      => (CMOVQNE y x (CMPWconst [0] check))
   430  (CondSelect <t> x y check) && !check.Type.IsFlags() && check.Type.Size() == 2 && ssa.Is32BitInt(t)
   431      => (CMOVLNE y x (CMPWconst [0] check))
   432  (CondSelect <t> x y check) && !check.Type.IsFlags() && check.Type.Size() == 2 && ssa.Is16BitInt(t)
   433      => (CMOVWNE y x (CMPWconst [0] check))
   434  (CondSelect <t> x y check) && !check.Type.IsFlags() && check.Type.Size() == 1 && (ssa.Is64BitInt(t) ||ssa.IsPtr(t))
   435      => (CMOVQNE y x (CMPBconst [0] check))
   436  (CondSelect <t> x y check) && !check.Type.IsFlags() && check.Type.Size() == 1 && ssa.Is32BitInt(t)
   437      => (CMOVLNE y x (CMPBconst [0] check))
   438  (CondSelect <t> x y check) && !check.Type.IsFlags() && check.Type.Size() == 1 && ssa.Is16BitInt(t)
   439      => (CMOVWNE y x (CMPBconst [0] check))
   440  
   441  // Absorb InvertFlags
   442  (CMOVQ(EQ|NE|LT|GT|LE|GE|HI|CS|CC|LS) x y (InvertFlags cond))
   443      => (CMOVQ(EQ|NE|GT|LT|GE|LE|CS|HI|LS|CC) x y cond)
   444  (CMOVL(EQ|NE|LT|GT|LE|GE|HI|CS|CC|LS) x y (InvertFlags cond))
   445      => (CMOVL(EQ|NE|GT|LT|GE|LE|CS|HI|LS|CC) x y cond)
   446  (CMOVW(EQ|NE|LT|GT|LE|GE|HI|CS|CC|LS) x y (InvertFlags cond))
   447      => (CMOVW(EQ|NE|GT|LT|GE|LE|CS|HI|LS|CC) x y cond)
   448  
   449  // Absorb constants generated during lower
   450  (CMOV(QEQ|QLE|QGE|QCC|QLS|LEQ|LLE|LGE|LCC|LLS|WEQ|WLE|WGE|WCC|WLS) _ x (FlagEQ)) => x
   451  (CMOV(QNE|QLT|QGT|QCS|QHI|LNE|LLT|LGT|LCS|LHI|WNE|WLT|WGT|WCS|WHI) y _ (FlagEQ)) => y
   452  (CMOV(QNE|QGT|QGE|QHI|QCC|LNE|LGT|LGE|LHI|LCC|WNE|WGT|WGE|WHI|WCC) _ x (FlagGT_UGT)) => x
   453  (CMOV(QEQ|QLE|QLT|QLS|QCS|LEQ|LLE|LLT|LLS|LCS|WEQ|WLE|WLT|WLS|WCS) y _ (FlagGT_UGT)) => y
   454  (CMOV(QNE|QGT|QGE|QLS|QCS|LNE|LGT|LGE|LLS|LCS|WNE|WGT|WGE|WLS|WCS) _ x (FlagGT_ULT)) => x
   455  (CMOV(QEQ|QLE|QLT|QHI|QCC|LEQ|LLE|LLT|LHI|LCC|WEQ|WLE|WLT|WHI|WCC) y _ (FlagGT_ULT)) => y
   456  (CMOV(QNE|QLT|QLE|QCS|QLS|LNE|LLT|LLE|LCS|LLS|WNE|WLT|WLE|WCS|WLS) _ x (FlagLT_ULT)) => x
   457  (CMOV(QEQ|QGT|QGE|QHI|QCC|LEQ|LGT|LGE|LHI|LCC|WEQ|WGT|WGE|WHI|WCC) y _ (FlagLT_ULT)) => y
   458  (CMOV(QNE|QLT|QLE|QHI|QCC|LNE|LLT|LLE|LHI|LCC|WNE|WLT|WLE|WHI|WCC) _ x (FlagLT_UGT)) => x
   459  (CMOV(QEQ|QGT|QGE|QCS|QLS|LEQ|LGT|LGE|LCS|LLS|WEQ|WGT|WGE|WCS|WLS) y _ (FlagLT_UGT)) => y
   460  
   461  // Miscellaneous
   462  (IsNonNil p) => (SETNE (TESTQ p p))
   463  (IsInBounds idx len) => (SETB (CMPQ idx len))
   464  (IsSliceInBounds idx len) => (SETBE (CMPQ idx len))
   465  (NilCheck ...) => (LoweredNilCheck ...)
   466  (GetG mem) && v.Block.Func.OwnAux.Fn.ABI() != obj.ABIInternal => (LoweredGetG mem) // only lower in old ABI. in new ABI we have a G register.
   467  (GetClosurePtr ...) => (LoweredGetClosurePtr ...)
   468  (GetCallerPC ...) => (LoweredGetCallerPC ...)
   469  (GetCallerSP ...) => (LoweredGetCallerSP ...)
   470  
   471  (HasCPUFeature {s}) => (SETNE (CMPLconst [0] (LoweredHasCPUFeature {s})))
   472  (Addr {sym} base) => (LEAQ {sym} base)
   473  (LocalAddr <t> {sym} base mem) && t.Elem().HasPointers() => (LEAQ {sym} (SPanchored base mem))
   474  (LocalAddr <t> {sym} base _)  && !t.Elem().HasPointers() => (LEAQ {sym} base)
   475  
   476  (MOVBstore [off] {sym} ptr y:(SETL x) mem) && y.Uses == 1 => (SETLstore [off] {sym} ptr x mem)
   477  (MOVBstore [off] {sym} ptr y:(SETLE x) mem) && y.Uses == 1 => (SETLEstore [off] {sym} ptr x mem)
   478  (MOVBstore [off] {sym} ptr y:(SETG x) mem) && y.Uses == 1 => (SETGstore [off] {sym} ptr x mem)
   479  (MOVBstore [off] {sym} ptr y:(SETGE x) mem) && y.Uses == 1 => (SETGEstore [off] {sym} ptr x mem)
   480  (MOVBstore [off] {sym} ptr y:(SETEQ x) mem) && y.Uses == 1 => (SETEQstore [off] {sym} ptr x mem)
   481  (MOVBstore [off] {sym} ptr y:(SETNE x) mem) && y.Uses == 1 => (SETNEstore [off] {sym} ptr x mem)
   482  (MOVBstore [off] {sym} ptr y:(SETB x) mem) && y.Uses == 1 => (SETBstore [off] {sym} ptr x mem)
   483  (MOVBstore [off] {sym} ptr y:(SETBE x) mem) && y.Uses == 1 => (SETBEstore [off] {sym} ptr x mem)
   484  (MOVBstore [off] {sym} ptr y:(SETA x) mem) && y.Uses == 1 => (SETAstore [off] {sym} ptr x mem)
   485  (MOVBstore [off] {sym} ptr y:(SETAE x) mem) && y.Uses == 1 => (SETAEstore [off] {sym} ptr x mem)
   486  
   487  // block rewrites
   488  (If (SETL  cmp) yes no) => (LT  cmp yes no)
   489  (If (SETLE cmp) yes no) => (LE  cmp yes no)
   490  (If (SETG  cmp) yes no) => (GT  cmp yes no)
   491  (If (SETGE cmp) yes no) => (GE  cmp yes no)
   492  (If (SETEQ cmp) yes no) => (EQ  cmp yes no)
   493  (If (SETNE cmp) yes no) => (NE  cmp yes no)
   494  (If (SETB  cmp) yes no) => (ULT cmp yes no)
   495  (If (SETBE cmp) yes no) => (ULE cmp yes no)
   496  (If (SETA  cmp) yes no) => (UGT cmp yes no)
   497  (If (SETAE cmp) yes no) => (UGE cmp yes no)
   498  (If (SETO cmp) yes no) => (OS cmp yes no)
   499  
   500  // Special case for floating point - LF/LEF not generated
   501  (If (SETGF  cmp) yes no) => (UGT  cmp yes no)
   502  (If (SETGEF cmp) yes no) => (UGE  cmp yes no)
   503  (If (SETEQF cmp) yes no) => (EQF  cmp yes no)
   504  (If (SETNEF cmp) yes no) => (NEF  cmp yes no)
   505  
   506  (If cond yes no) => (NE (TESTB cond cond) yes no)
   507  
   508  (JumpTable idx) => (JUMPTABLE {ssa.MakeJumpTableSym(b)} idx (LEAQ <typ.Uintptr> {ssa.MakeJumpTableSym(b)} (SB)))
   509  
   510  // Atomic loads.  Other than preserving their ordering with respect to other loads, nothing special here.
   511  (AtomicLoad8 ptr mem) => (MOVBatomicload ptr mem)
   512  (AtomicLoad32 ptr mem) => (MOVLatomicload ptr mem)
   513  (AtomicLoad64 ptr mem) => (MOVQatomicload ptr mem)
   514  (AtomicLoadPtr ptr mem) => (MOVQatomicload ptr mem)
   515  
   516  // Atomic stores.  We use XCHG to prevent the hardware reordering a subsequent load.
   517  // TODO: most runtime uses of atomic stores don't need that property.  Use normal stores for those?
   518  (AtomicStore8 ptr val mem) => (Select1 (XCHGB <types.NewTuple(typ.UInt8,types.TypeMem)> val ptr mem))
   519  (AtomicStore32 ptr val mem) => (Select1 (XCHGL <types.NewTuple(typ.UInt32,types.TypeMem)> val ptr mem))
   520  (AtomicStore64 ptr val mem) => (Select1 (XCHGQ <types.NewTuple(typ.UInt64,types.TypeMem)> val ptr mem))
   521  (AtomicStorePtrNoWB ptr val mem) => (Select1 (XCHGQ <types.NewTuple(typ.BytePtr,types.TypeMem)> val ptr mem))
   522  
   523  // Atomic exchanges.
   524  (AtomicExchange8 ptr val mem) => (XCHGB val ptr mem)
   525  (AtomicExchange32 ptr val mem) => (XCHGL val ptr mem)
   526  (AtomicExchange64 ptr val mem) => (XCHGQ val ptr mem)
   527  
   528  // Atomic adds.
   529  (AtomicAdd32 ptr val mem) => (AddTupleFirst32 val (XADDLlock val ptr mem))
   530  (AtomicAdd64 ptr val mem) => (AddTupleFirst64 val (XADDQlock val ptr mem))
   531  (Select0 <t> (AddTupleFirst32 val tuple)) => (ADDL val (Select0 <t> tuple))
   532  (Select1     (AddTupleFirst32   _ tuple)) => (Select1 tuple)
   533  (Select0 <t> (AddTupleFirst64 val tuple)) => (ADDQ val (Select0 <t> tuple))
   534  (Select1     (AddTupleFirst64   _ tuple)) => (Select1 tuple)
   535  
   536  (Select1 xadd:(XADDLlock [off] {sym} val ptr mem)) && xadd.Uses == 1 && ssa.Clobber(xadd) => (ADDLlock [off] {sym} ptr val mem)
   537  (Select1 xadd:(XADDQlock [off] {sym} val ptr mem)) && xadd.Uses == 1 && ssa.Clobber(xadd) => (ADDQlock [off] {sym} ptr val mem)
   538  // FIXME: remove when we teach rewrite to remove dead values
   539  (AddTupleFirst32 _ _) && v.Uses == 0 => (Invalid)
   540  (AddTupleFirst64 _ _) && v.Uses == 0 => (Invalid)
   541  
   542  // Atomic subtracts.
   543  (ADDLlock [off] {sym} ptr n:(NEGL val) mem) && n.Uses == 1 => (SUBLlock [off] {sym} ptr val mem)
   544  (ADDQlock [off] {sym} ptr n:(NEGQ val) mem) && n.Uses == 1 => (SUBQlock [off] {sym} ptr val mem)
   545  
   546  // Atomic inc and dec.
   547  (ADDLlock [off] {sym} ptr (MOVLconst [1]) mem) => (INCLlock [off] {sym} ptr mem)
   548  (ADDQlock [off] {sym} ptr (MOVQconst [1]) mem) => (INCQlock [off] {sym} ptr mem)
   549  (ADDLlock [off] {sym} ptr (MOVLconst [-1]) mem) => (DECLlock [off] {sym} ptr mem)
   550  (ADDQlock [off] {sym} ptr (MOVQconst [-1]) mem) => (DECQlock [off] {sym} ptr mem)
   551  
   552  // Atomic compare and swap.
   553  (AtomicCompareAndSwap32 ptr old new_ mem) => (CMPXCHGLlock ptr old new_ mem)
   554  (AtomicCompareAndSwap64 ptr old new_ mem) => (CMPXCHGQlock ptr old new_ mem)
   555  
   556  // Atomic memory logical operations (old style).
   557  (AtomicAnd8  ptr val mem) => (ANDBlock ptr val mem)
   558  (AtomicAnd32 ptr val mem) => (ANDLlock ptr val mem)
   559  (AtomicOr8   ptr val mem) => (ORBlock  ptr val mem)
   560  (AtomicOr32  ptr val mem) => (ORLlock  ptr val mem)
   561  
   562  // Atomic memory logical operations (new style).
   563  (Atomic(And64|And32|Or64|Or32)value ptr val mem) => (LoweredAtomic(And64|And32|Or64|Or32) ptr val mem)
   564  
   565  // Write barrier.
   566  (WB ...) => (LoweredWB ...)
   567  
   568  (PanicBounds ...) => (LoweredPanicBoundsRR ...)
   569  (LoweredPanicBoundsRR [kind] x (MOVQconst [c]) mem) => (LoweredPanicBoundsRC [kind] x {ssa.PanicBoundsC{C:c}} mem)
   570  (LoweredPanicBoundsRR [kind] (MOVQconst [c]) y mem) => (LoweredPanicBoundsCR [kind] {ssa.PanicBoundsC{C:c}} y mem)
   571  (LoweredPanicBoundsRC [kind] {p} (MOVQconst [c]) mem) => (LoweredPanicBoundsCC [kind] {ssa.PanicBoundsCC{Cx:c, Cy:p.C}} mem)
   572  (LoweredPanicBoundsCR [kind] {p} (MOVQconst [c]) mem) => (LoweredPanicBoundsCC [kind] {ssa.PanicBoundsCC{Cx:p.C, Cy:c}} mem)
   573  
   574  // lowering rotates
   575  (RotateLeft8  ...) => (ROLB ...)
   576  (RotateLeft16 ...) => (ROLW ...)
   577  (RotateLeft32 ...) => (ROLL ...)
   578  (RotateLeft64 ...) => (ROLQ ...)
   579  
   580  // ***************************
   581  // Above: lowering rules
   582  // Below: optimizations
   583  // ***************************
   584  // TODO: Should the optimizations be a separate pass?
   585  
   586  // Fold boolean tests into blocks
   587  (NE (TESTB (SETL  cmp) (SETL  cmp)) yes no) => (LT  cmp yes no)
   588  (NE (TESTB (SETLE cmp) (SETLE cmp)) yes no) => (LE  cmp yes no)
   589  (NE (TESTB (SETG  cmp) (SETG  cmp)) yes no) => (GT  cmp yes no)
   590  (NE (TESTB (SETGE cmp) (SETGE cmp)) yes no) => (GE  cmp yes no)
   591  (NE (TESTB (SETEQ cmp) (SETEQ cmp)) yes no) => (EQ  cmp yes no)
   592  (NE (TESTB (SETNE cmp) (SETNE cmp)) yes no) => (NE  cmp yes no)
   593  (NE (TESTB (SETB  cmp) (SETB  cmp)) yes no) => (ULT cmp yes no)
   594  (NE (TESTB (SETBE cmp) (SETBE cmp)) yes no) => (ULE cmp yes no)
   595  (NE (TESTB (SETA  cmp) (SETA  cmp)) yes no) => (UGT cmp yes no)
   596  (NE (TESTB (SETAE cmp) (SETAE cmp)) yes no) => (UGE cmp yes no)
   597  (NE (TESTB (SETO cmp) (SETO cmp)) yes no) => (OS cmp yes no)
   598  
   599  // Unsigned comparisons to 0/1
   600  (ULT (TEST(Q|L|W|B) x x) yes no) => (First no yes)
   601  (UGE (TEST(Q|L|W|B) x x) yes no) => (First yes no)
   602  (SETB (TEST(Q|L|W|B) x x)) => (ConstBool [false])
   603  (SETAE (TEST(Q|L|W|B) x x)) => (ConstBool [true])
   604  
   605  // x & 1 != 0 -> x & 1
   606  (SETNE (TEST(B|W)const [1] x)) => (AND(L|L)const [1] x)
   607  (SETB (BT(L|Q)const [0] x)) => (AND(L|Q)const [1] x)
   608  // x & 1 == 0 -> (x & 1) ^ 1
   609  (SETAE (BT(L|Q)const [0] x)) => (XORLconst [1] (ANDLconst <typ.Bool> [1] x))
   610  
   611  // Shorten compare by rewriting x < 128 as x <= 127, which can be encoded in a single-byte immediate on x86.
   612  (SETL c:(CMP(Q|L)const [128] x)) && c.Uses == 1 => (SETLE (CMP(Q|L)const [127] x))
   613  (SETB c:(CMP(Q|L)const [128] x)) && c.Uses == 1 => (SETBE (CMP(Q|L)const [127] x))
   614  
   615  // x >= 128 -> x > 127
   616  (SETGE c:(CMP(Q|L)const [128] x)) && c.Uses == 1 => (SETG (CMP(Q|L)const [127] x))
   617  (SETAE c:(CMP(Q|L)const [128] x)) && c.Uses == 1 => (SETA (CMP(Q|L)const [127] x))
   618  
   619  (CMOVQLT x y c:(CMP(Q|L)const [128] z)) && c.Uses == 1 => (CMOVQLE x y (CMP(Q|L)const [127] z))
   620  (CMOVLLT x y c:(CMP(Q|L)const [128] z)) && c.Uses == 1 => (CMOVLLE x y (CMP(Q|L)const [127] z))
   621  (LT          c:(CMP(Q|L)const [128] z) yes no) && c.Uses == 1 => (LE (CMP(Q|L)const [127] z) yes no)
   622  (CMOVQGE x y c:(CMP(Q|L)const [128] z)) && c.Uses == 1 => (CMOVQGT x y (CMP(Q|L)const [127] z))
   623  (CMOVLGE x y c:(CMP(Q|L)const [128] z)) && c.Uses == 1 => (CMOVLGT x y (CMP(Q|L)const [127] z))
   624  (GE          c:(CMP(Q|L)const [128] z) yes no) && c.Uses == 1 => (GT (CMP(Q|L)const [127] z)  yes no)
   625  
   626  // Recognize bit tests: a&(1<<b) != 0 for b suitably bounded
   627  // Note that BTx instructions use the carry bit, so we need to convert tests for zero flag
   628  // into tests for carry flags.
   629  // ULT and SETB check the carry flag; they are identical to CS and SETCS. Same, mutatis
   630  // mutandis, for UGE and SETAE, and CC and SETCC.
   631  ((NE|EQ) (TESTL (SHLL (MOVLconst [1]) x) y)) => ((ULT|UGE) (BTL x y))
   632  ((NE|EQ) (TESTQ (SHLQ (MOVQconst [1]) x) y)) => ((ULT|UGE) (BTQ x y))
   633  ((NE|EQ) (TESTLconst [c] x)) && ssa.IsPowerOfTwo(uint32(c))
   634      => ((ULT|UGE) (BTLconst [int8(ssa.Log32u(uint32(c)))] x))
   635  ((NE|EQ) (TESTQconst [c] x)) && ssa.IsPowerOfTwo(uint64(c))
   636      => ((ULT|UGE) (BTQconst [int8(ssa.Log32u(uint32(c)))] x))
   637  ((NE|EQ) (TESTQ (MOVQconst [c]) x)) && ssa.IsPowerOfTwo(uint64(c))
   638      => ((ULT|UGE) (BTQconst [int8(ssa.Log64u(uint64(c)))] x))
   639  (SET(NE|EQ) (TESTL (SHLL (MOVLconst [1]) x) y)) => (SET(B|AE)  (BTL x y))
   640  (SET(NE|EQ) (TESTQ (SHLQ (MOVQconst [1]) x) y)) => (SET(B|AE)  (BTQ x y))
   641  (SET(NE|EQ) (TESTLconst [c] x)) && ssa.IsPowerOfTwo(uint32(c))
   642      => (SET(B|AE)  (BTLconst [int8(ssa.Log32u(uint32(c)))] x))
   643  (SET(NE|EQ) (TESTQconst [c] x)) && ssa.IsPowerOfTwo(uint64(c))
   644      => (SET(B|AE)  (BTQconst [int8(ssa.Log32u(uint32(c)))] x))
   645  (SET(NE|EQ) (TESTQ (MOVQconst [c]) x)) && ssa.IsPowerOfTwo(uint64(c))
   646      => (SET(B|AE)  (BTQconst [int8(ssa.Log64u(uint64(c)))] x))
   647  // SET..store variant
   648  (SET(NE|EQ)store [off] {sym} ptr (TESTL (SHLL (MOVLconst [1]) x) y) mem)
   649      => (SET(B|AE)store  [off] {sym} ptr (BTL x y) mem)
   650  (SET(NE|EQ)store [off] {sym} ptr (TESTQ (SHLQ (MOVQconst [1]) x) y) mem)
   651      => (SET(B|AE)store  [off] {sym} ptr (BTQ x y) mem)
   652  (SET(NE|EQ)store [off] {sym} ptr (TESTLconst [c] x) mem) && ssa.IsPowerOfTwo(uint32(c))
   653      => (SET(B|AE)store  [off] {sym} ptr (BTLconst [int8(ssa.Log32u(uint32(c)))] x) mem)
   654  (SET(NE|EQ)store [off] {sym} ptr (TESTQconst [c] x) mem) && ssa.IsPowerOfTwo(uint64(c))
   655      => (SET(B|AE)store  [off] {sym} ptr (BTQconst [int8(ssa.Log32u(uint32(c)))] x) mem)
   656  (SET(NE|EQ)store [off] {sym} ptr (TESTQ (MOVQconst [c]) x) mem) && ssa.IsPowerOfTwo(uint64(c))
   657      => (SET(B|AE)store  [off] {sym} ptr (BTQconst [int8(ssa.Log64u(uint64(c)))] x) mem)
   658  
   659  // Handle bit-testing in the form (a>>b)&1 != 0 by building the above rules
   660  // and further combining shifts.
   661  (BT(Q|L)const [c] (SHRQconst [d] x)) && (c+d)<64 => (BTQconst [c+d] x)
   662  (BT(Q|L)const [c] (ADDQ x x)) && c>1  => (BT(Q|L)const [c-1] x)
   663  (BT(Q|L)const [c] (SHLQconst [d] x)) && c>d      => (BT(Q|L)const [c-d] x)
   664  (BT(Q|L)const [0] s:(SHRQ x y)) => (BTQ y x)
   665  (BTLconst [c] (SHRLconst [d] x)) && (c+d)<32 => (BTLconst [c+d] x)
   666  (BTLconst [c] (ADDL x x)) && c>1 => (BTLconst [c-1] x)
   667  (BTLconst [c] (SHLLconst [d] x)) && c>d      => (BTLconst [c-d] x)
   668  (BTLconst [0] s:(SHR(L|XL) x y)) => (BTL y x)
   669  
   670  // Rewrite a & 1 != 1 into a & 1 == 0.
   671  // Among other things, this lets us turn (a>>b)&1 != 1 into a bit test.
   672  (SET(NE|EQ) (CMPLconst [1] s:(ANDLconst [1] _))) => (SET(EQ|NE) (CMPLconst [0] s))
   673  (SET(NE|EQ)store [off] {sym} ptr (CMPLconst [1] s:(ANDLconst [1] _)) mem) => (SET(EQ|NE)store [off] {sym} ptr (CMPLconst [0] s) mem)
   674  (SET(NE|EQ) (CMPQconst [1] s:(ANDQconst [1] _))) => (SET(EQ|NE) (CMPQconst [0] s))
   675  (SET(NE|EQ)store [off] {sym} ptr (CMPQconst [1] s:(ANDQconst [1] _)) mem) => (SET(EQ|NE)store [off] {sym} ptr (CMPQconst [0] s) mem)
   676  
   677  // Recognize bit setting (a |= 1<<b) and toggling (a ^= 1<<b)
   678  (OR(Q|L) (SHL(Q|L) (MOV(Q|L)const [1]) y) x) => (BTS(Q|L) x y)
   679  (XOR(Q|L) (SHL(Q|L) (MOV(Q|L)const [1]) y) x) => (BTC(Q|L) x y)
   680  // Note: only convert OR/XOR to BTS/BTC if the constant wouldn't fit in
   681  // the constant field of the OR/XOR instruction. See issue 61694.
   682  ((OR|XOR)Q (MOVQconst [c]) x) && ssa.IsPowerOfTwo(uint64(c)) && uint64(c) >= 1<<31 => (BT(S|C)Qconst [int8(ssa.Log64u(uint64(c)))] x)
   683  
   684  // Recognize bit clearing: a &^= 1<<b
   685  (AND(Q|L) (NOT(Q|L) (SHL(Q|L) (MOV(Q|L)const [1]) y)) x) => (BTR(Q|L) x y)
   686  (ANDN(Q|L) x (SHL(Q|L) (MOV(Q|L)const [1]) y)) => (BTR(Q|L) x y)
   687  // Note: only convert AND to BTR if the constant wouldn't fit in
   688  // the constant field of the AND instruction. See issue 61694.
   689  (ANDQ (MOVQconst [c]) x) && ssa.IsPowerOfTwo(uint64(^c)) && uint64(^c) >= 1<<31 => (BTRQconst [int8(ssa.Log64u(uint64(^c)))] x)
   690  
   691  // Special-case bit patterns on first/last bit.
   692  // generic.rules changes ANDs of high-part/low-part masks into a couple of shifts,
   693  // for instance:
   694  //    x & 0xFFFF0000 -> (x >> 16) << 16
   695  //    x & 0x80000000 -> (x >> 31) << 31
   696  //
   697  // In case the mask is just one bit (like second example above), it conflicts
   698  // with the above rules to detect bit-testing / bit-clearing of first/last bit.
   699  // We thus special-case them, by detecting the shift patterns.
   700  
   701  // Special case resetting first/last bit
   702  (ADD(L|Q) (SHR(L|Q)const [1] x) (SHR(L|Q)const [1] x))
   703  	=> (AND(L|Q)const [-2] x)
   704  (SHRLconst [1] (ADDL x x))
   705  	=> (ANDLconst [0x7fffffff] x)
   706  (SHRQconst [1] (ADDQ x x))
   707  	=> (BTRQconst [63] x)
   708  
   709  // Special case testing first/last bit (with double-shift generated by generic.rules)
   710  ((SETNE|SETEQ|NE|EQ) (TESTQ z1:(SHLQconst [63] (SHRQconst [63] x)) z2)) && z1==z2
   711      => ((SETB|SETAE|ULT|UGE) (BTQconst [63] x))
   712  ((SETNE|SETEQ|NE|EQ) (TESTL z1:(SHLLconst [31] (SHRQconst [31] x)) z2)) && z1==z2
   713      => ((SETB|SETAE|ULT|UGE) (BTQconst [31] x))
   714  (SET(NE|EQ)store [off] {sym} ptr (TESTQ z1:(SHLQconst [63] (SHRQconst [63] x)) z2) mem) && z1==z2
   715      => (SET(B|AE)store [off] {sym} ptr (BTQconst [63] x) mem)
   716  (SET(NE|EQ)store [off] {sym} ptr (TESTL z1:(SHLLconst [31] (SHRLconst [31] x)) z2) mem) && z1==z2
   717      => (SET(B|AE)store [off] {sym} ptr (BTLconst [31] x) mem)
   718  
   719  ((SETNE|SETEQ|NE|EQ) (TESTQ z1:(SHRQconst [63] (SHLQconst [63] x)) z2)) && z1==z2
   720      => ((SETB|SETAE|ULT|UGE)  (BTQconst [0] x))
   721  ((SETNE|SETEQ|NE|EQ) (TESTL z1:(SHRLconst [31] (SHLLconst [31] x)) z2)) && z1==z2
   722      => ((SETB|SETAE|ULT|UGE)  (BTLconst [0] x))
   723  (SET(NE|EQ)store [off] {sym} ptr (TESTQ z1:(SHRQconst [63] (SHLQconst [63] x)) z2) mem) && z1==z2
   724      => (SET(B|AE)store [off] {sym} ptr (BTQconst [0] x) mem)
   725  (SET(NE|EQ)store [off] {sym} ptr (TESTL z1:(SHRLconst [31] (SHLLconst [31] x)) z2) mem) && z1==z2
   726      => (SET(B|AE)store [off] {sym} ptr (BTLconst [0] x) mem)
   727  
   728  // Special-case manually testing last bit with "a>>63 != 0" (without "&1")
   729  ((SETNE|SETEQ|NE|EQ) (TESTQ z1:(SHRQconst [63] x) z2)) && z1==z2
   730      => ((SETB|SETAE|ULT|UGE) (BTQconst [63] x))
   731  ((SETNE|SETEQ|NE|EQ) (TESTL z1:(SHRLconst [31] x) z2)) && z1==z2
   732      => ((SETB|SETAE|ULT|UGE) (BTLconst [31] x))
   733  (SET(NE|EQ)store [off] {sym} ptr (TESTQ z1:(SHRQconst [63] x) z2) mem) && z1==z2
   734      => (SET(B|AE)store [off] {sym} ptr (BTQconst [63] x) mem)
   735  (SET(NE|EQ)store [off] {sym} ptr (TESTL z1:(SHRLconst [31] x) z2) mem) && z1==z2
   736      => (SET(B|AE)store [off] {sym} ptr (BTLconst [31] x) mem)
   737  
   738  // Fold combinations of bit ops on same bit. An example is math.Copysign(c,-1)
   739  (BTSQconst [c] (BTRQconst [c] x)) => (BTSQconst [c] x)
   740  (BTSQconst [c] (BTCQconst [c] x)) => (BTSQconst [c] x)
   741  (BTRQconst [c] (BTSQconst [c] x)) => (BTRQconst [c] x)
   742  (BTRQconst [c] (BTCQconst [c] x)) => (BTRQconst [c] x)
   743  
   744  // Fold boolean negation into SETcc.
   745  (XORLconst [1] (SETNE x)) => (SETEQ x)
   746  (XORLconst [1] (SETEQ x)) => (SETNE x)
   747  (XORLconst [1] (SETL  x)) => (SETGE x)
   748  (XORLconst [1] (SETGE x)) => (SETL  x)
   749  (XORLconst [1] (SETLE x)) => (SETG  x)
   750  (XORLconst [1] (SETG  x)) => (SETLE x)
   751  (XORLconst [1] (SETB  x)) => (SETAE x)
   752  (XORLconst [1] (SETAE x)) => (SETB  x)
   753  (XORLconst [1] (SETBE x)) => (SETA  x)
   754  (XORLconst [1] (SETA  x)) => (SETBE x)
   755  
   756  // Special case for floating point - LF/LEF not generated
   757  (NE (TESTB (SETGF  cmp) (SETGF  cmp)) yes no) => (UGT  cmp yes no)
   758  (NE (TESTB (SETGEF cmp) (SETGEF cmp)) yes no) => (UGE  cmp yes no)
   759  (NE (TESTB (SETEQF cmp) (SETEQF cmp)) yes no) => (EQF  cmp yes no)
   760  (NE (TESTB (SETNEF cmp) (SETNEF cmp)) yes no) => (NEF  cmp yes no)
   761  
   762  // Disabled because it interferes with the pattern match above and makes worse code.
   763  // (SETNEF x) => (ORQ (SETNE <typ.Int8> x) (SETNAN <typ.Int8> x))
   764  // (SETEQF x) => (ANDQ (SETEQ <typ.Int8> x) (SETORD <typ.Int8> x))
   765  
   766  // fold constants into instructions
   767  (ADDQ x (MOVQconst <t> [c])) && ssa.Is32Bit(c) && !t.IsPtr() => (ADDQconst [int32(c)] x)
   768  (ADDQ x (MOVLconst [c])) => (ADDQconst [c] x)
   769  (ADDL x (MOVLconst [c])) => (ADDLconst [c] x)
   770  
   771  (SUBQ x (MOVQconst [c])) && ssa.Is32Bit(c) => (SUBQconst x [int32(c)])
   772  (SUBQ (MOVQconst [c]) x) && ssa.Is32Bit(c) => (NEGQ (SUBQconst <v.Type> x [int32(c)]))
   773  (SUBL x (MOVLconst [c])) => (SUBLconst x [c])
   774  (SUBL (MOVLconst [c]) x) => (NEGL (SUBLconst <v.Type> x [c]))
   775  
   776  (MULQ x (MOVQconst [c])) && ssa.Is32Bit(c) => (MULQconst [int32(c)] x)
   777  (MULL x (MOVLconst [c])) => (MULLconst [c] x)
   778  
   779  (ANDQ x (MOVQconst [c])) && ssa.Is32Bit(c) => (ANDQconst [int32(c)] x)
   780  (ANDL x (MOVLconst [c])) => (ANDLconst [c] x)
   781  
   782  (AND(L|Q)const [c] (AND(L|Q)const [d] x)) => (AND(L|Q)const [c & d] x)
   783  (XOR(L|Q)const [c] (XOR(L|Q)const [d] x)) => (XOR(L|Q)const [c ^ d] x)
   784  (OR(L|Q)const  [c] (OR(L|Q)const  [d] x)) => (OR(L|Q)const  [c | d] x)
   785  
   786  (MULLconst [c] (MULLconst [d] x)) => (MULLconst [c * d] x)
   787  (MULQconst [c] (MULQconst [d] x)) && ssa.Is32Bit(int64(c)*int64(d)) => (MULQconst [c * d] x)
   788  
   789  (ORQ x (MOVQconst [c])) && ssa.Is32Bit(c) => (ORQconst [int32(c)] x)
   790  (ORQ x (MOVLconst [c])) => (ORQconst [c] x)
   791  (ORL x (MOVLconst [c])) => (ORLconst [c] x)
   792  
   793  (XORQ x (MOVQconst [c])) && ssa.Is32Bit(c) => (XORQconst [int32(c)] x)
   794  (XORL x (MOVLconst [c])) => (XORLconst [c] x)
   795  
   796  (SHLQ x (MOV(Q|L)const [c])) => (SHLQconst [int8(c&63)] x)
   797  (SHLL x (MOV(Q|L)const [c])) => (SHLLconst [int8(c&31)] x)
   798  
   799  (SHRQ x (MOV(Q|L)const [c])) => (SHRQconst [int8(c&63)] x)
   800  (SHRL x (MOV(Q|L)const [c])) => (SHRLconst [int8(c&31)] x)
   801  (SHRW x (MOV(Q|L)const [c])) && c&31 < 16 => (SHRWconst [int8(c&31)] x)
   802  (SHRW _ (MOV(Q|L)const [c])) && c&31 >= 16 => (MOVLconst [0])
   803  (SHRB x (MOV(Q|L)const [c])) && c&31 < 8 => (SHRBconst [int8(c&31)] x)
   804  (SHRB _ (MOV(Q|L)const [c])) && c&31 >= 8 => (MOVLconst [0])
   805  
   806  (SARQ x (MOV(Q|L)const [c])) => (SARQconst [int8(c&63)] x)
   807  (SARL x (MOV(Q|L)const [c])) => (SARLconst [int8(c&31)] x)
   808  (SARW x (MOV(Q|L)const [c])) => (SARWconst [int8(min(int64(c)&31,15))] x)
   809  (SARB x (MOV(Q|L)const [c])) => (SARBconst [int8(min(int64(c)&31,7))] x)
   810  
   811  // Operations which don't affect the low 6/5 bits of the shift amount are NOPs.
   812  ((SHLQ|SHRQ|SARQ) x (ADDQconst [c] y)) && c & 63 == 0  => ((SHLQ|SHRQ|SARQ) x y)
   813  ((SHLQ|SHRQ|SARQ) x (NEGQ <t> (ADDQconst [c] y))) && c & 63 == 0  => ((SHLQ|SHRQ|SARQ) x (NEGQ <t> y))
   814  ((SHLQ|SHRQ|SARQ) x (ANDQconst [c] y)) && c & 63 == 63 => ((SHLQ|SHRQ|SARQ) x y)
   815  ((SHLQ|SHRQ|SARQ) x (NEGQ <t> (ANDQconst [c] y))) && c & 63 == 63 => ((SHLQ|SHRQ|SARQ) x (NEGQ <t> y))
   816  
   817  ((SHLL|SHRL|SARL) x (ADDQconst [c] y)) && c & 31 == 0  => ((SHLL|SHRL|SARL) x y)
   818  ((SHLL|SHRL|SARL) x (NEGQ <t> (ADDQconst [c] y))) && c & 31 == 0  => ((SHLL|SHRL|SARL) x (NEGQ <t> y))
   819  ((SHLL|SHRL|SARL) x (ANDQconst [c] y)) && c & 31 == 31 => ((SHLL|SHRL|SARL) x y)
   820  ((SHLL|SHRL|SARL) x (NEGQ <t> (ANDQconst [c] y))) && c & 31 == 31 => ((SHLL|SHRL|SARL) x (NEGQ <t> y))
   821  
   822  ((SHLQ|SHRQ|SARQ) x (ADDLconst [c] y)) && c & 63 == 0  => ((SHLQ|SHRQ|SARQ) x y)
   823  ((SHLQ|SHRQ|SARQ) x (NEGL <t> (ADDLconst [c] y))) && c & 63 == 0  => ((SHLQ|SHRQ|SARQ) x (NEGL <t> y))
   824  ((SHLQ|SHRQ|SARQ) x (ANDLconst [c] y)) && c & 63 == 63 => ((SHLQ|SHRQ|SARQ) x y)
   825  ((SHLQ|SHRQ|SARQ) x (NEGL <t> (ANDLconst [c] y))) && c & 63 == 63 => ((SHLQ|SHRQ|SARQ) x (NEGL <t> y))
   826  
   827  ((SHLL|SHRL|SARL) x (ADDLconst [c] y)) && c & 31 == 0  => ((SHLL|SHRL|SARL) x y)
   828  ((SHLL|SHRL|SARL) x (NEGL <t> (ADDLconst [c] y))) && c & 31 == 0  => ((SHLL|SHRL|SARL) x (NEGL <t> y))
   829  ((SHLL|SHRL|SARL) x (ANDLconst [c] y)) && c & 31 == 31 => ((SHLL|SHRL|SARL) x y)
   830  ((SHLL|SHRL|SARL) x (NEGL <t> (ANDLconst [c] y))) && c & 31 == 31 => ((SHLL|SHRL|SARL) x (NEGL <t> y))
   831  
   832  // rotate left negative = rotate right
   833  (ROLQ x (NEG(Q|L) y)) => (RORQ x y)
   834  (ROLL x (NEG(Q|L) y)) => (RORL x y)
   835  (ROLW x (NEG(Q|L) y)) => (RORW x y)
   836  (ROLB x (NEG(Q|L) y)) => (RORB x y)
   837  
   838  // rotate right negative = rotate left
   839  (RORQ x (NEG(Q|L) y)) => (ROLQ x y)
   840  (RORL x (NEG(Q|L) y)) => (ROLL x y)
   841  (RORW x (NEG(Q|L) y)) => (ROLW x y)
   842  (RORB x (NEG(Q|L) y)) => (ROLB x y)
   843  
   844  // rotate by constants
   845  (ROLQ x (MOV(Q|L)const [c])) => (ROLQconst [int8(c&63)] x)
   846  (ROLL x (MOV(Q|L)const [c])) => (ROLLconst [int8(c&31)] x)
   847  (ROLW x (MOV(Q|L)const [c])) => (ROLWconst [int8(c&15)] x)
   848  (ROLB x (MOV(Q|L)const [c])) => (ROLBconst [int8(c&7) ] x)
   849  
   850  (RORQ x (MOV(Q|L)const [c])) => (ROLQconst [int8((-c)&63)] x)
   851  (RORL x (MOV(Q|L)const [c])) => (ROLLconst [int8((-c)&31)] x)
   852  (RORW x (MOV(Q|L)const [c])) => (ROLWconst [int8((-c)&15)] x)
   853  (RORB x (MOV(Q|L)const [c])) => (ROLBconst [int8((-c)&7) ] x)
   854  
   855  // Constant shift simplifications
   856  ((SHLQ|SHRQ|SARQ)const      x [0]) => x
   857  ((SHLL|SHRL|SARL)const      x [0]) => x
   858  ((SHRW|SARW)const           x [0]) => x
   859  ((SHRB|SARB)const           x [0]) => x
   860  ((ROLQ|ROLL|ROLW|ROLB)const x [0]) => x
   861  
   862  // Note: the word and byte shifts keep the low 5 bits (not the low 4 or 3 bits)
   863  // because the x86 instructions are defined to use all 5 bits of the shift even
   864  // for the small shifts. I don't think we'll ever generate a weird shift (e.g.
   865  // (SHRW x (MOVLconst [24])), but just in case.
   866  
   867  (CMPQ x (MOVQconst [c])) && ssa.Is32Bit(c) => (CMPQconst x [int32(c)])
   868  (CMPQ (MOVQconst [c]) x) && ssa.Is32Bit(c) => (InvertFlags (CMPQconst x [int32(c)]))
   869  (CMPL x (MOVLconst [c])) => (CMPLconst x [c])
   870  (CMPL (MOVLconst [c]) x) => (InvertFlags (CMPLconst x [c]))
   871  (CMPW x (MOVLconst [c])) => (CMPWconst x [int16(c)])
   872  (CMPW (MOVLconst [c]) x) => (InvertFlags (CMPWconst x [int16(c)]))
   873  (CMPB x (MOVLconst [c])) => (CMPBconst x [int8(c)])
   874  (CMPB (MOVLconst [c]) x) => (InvertFlags (CMPBconst x [int8(c)]))
   875  
   876  // Canonicalize the order of arguments to comparisons - helps with CSE.
   877  (CMP(Q|L|W|B) x y) && ssa.CanonLessThan(x,y) => (InvertFlags (CMP(Q|L|W|B) y x))
   878  
   879  // Using MOVZX instead of AND is cheaper.
   880  (AND(Q|L)const [  0xFF] x) => (MOVBQZX x)
   881  (AND(Q|L)const [0xFFFF] x) => (MOVWQZX x)
   882  // This rule is currently invalid because 0xFFFFFFFF is not representable by a signed int32.
   883  // Commenting out for now, because it also can't trigger because of the is32bit guard on the
   884  // ANDQconst lowering-rule, above, prevents 0xFFFFFFFF from matching (for the same reason)
   885  // Using an alternate form of this rule segfaults some binaries because of
   886  // adverse interactions with other passes.
   887  // (ANDQconst [0xFFFFFFFF] x) => (MOVLQZX x)
   888  
   889  // strength reduction
   890  (MUL(Q|L)const [ 0] _) => (MOV(Q|L)const [0])
   891  (MUL(Q|L)const [ 1] x) => x
   892  (MULQconst [c] x) && ssa.CanMulStrengthReduce(config, int64(c)) => {ssa.MulStrengthReduce(v, x, int64(c))}
   893  (MULLconst [c] x) && v.Type.Size() <= 4 && ssa.CanMulStrengthReduce32(config, c) => {ssa.MulStrengthReduce32(v, x, c)}
   894  
   895  // Prefer addition when shifting left by one
   896  (SHL(Q|L)const [1] x) => (ADD(Q|L) x x)
   897  
   898  // combine add/shift into LEAQ/LEAL
   899  (ADD(L|Q) x (SHL(L|Q)const [3] y)) => (LEA(L|Q)8 x y)
   900  (ADD(L|Q) x (SHL(L|Q)const [2] y)) => (LEA(L|Q)4 x y)
   901  (ADD(L|Q) x (ADD(L|Q) y y))        => (LEA(L|Q)2 x y)
   902  (ADD(L|Q) x (ADD(L|Q) x y))        => (LEA(L|Q)2 y x)
   903  
   904  // combine ADDQ/ADDQconst into LEAQ1/LEAL1
   905  (ADD(Q|L)const [c] (ADD(Q|L) x y)) => (LEA(Q|L)1 [c] x y)
   906  (ADD(Q|L) (ADD(Q|L)const [c] x) y) => (LEA(Q|L)1 [c] x y)
   907  (ADD(Q|L)const [c] (ADD(Q|L) x x)) => (LEA(Q|L)1 [c] x x)
   908  
   909  // fold ADDQ/ADDL into LEAQ/LEAL
   910  (ADD(Q|L)const [c] (LEA(Q|L) [d] {s} x)) && ssa.Is32Bit(int64(c)+int64(d)) => (LEA(Q|L) [c+d] {s} x)
   911  (LEA(Q|L) [c] {s} (ADD(Q|L)const [d] x)) && ssa.Is32Bit(int64(c)+int64(d)) => (LEA(Q|L) [c+d] {s} x)
   912  (LEA(Q|L) [c] {s} (ADD(Q|L) x y)) && x.Op != ssaop.OpSB && y.Op != ssaop.OpSB => (LEA(Q|L)1 [c] {s} x y)
   913  (ADD(Q|L) x (LEA(Q|L) [c] {s} y)) && x.Op != ssaop.OpSB && y.Op != ssaop.OpSB => (LEA(Q|L)1 [c] {s} x y)
   914  
   915  // fold ADDQconst/ADDLconst into LEAQx/LEALx
   916  (ADD(Q|L)const [c] (LEA(Q|L)1 [d] {s} x y)) && ssa.Is32Bit(int64(c)+int64(d)) => (LEA(Q|L)1 [c+d] {s} x y)
   917  (ADD(Q|L)const [c] (LEA(Q|L)2 [d] {s} x y)) && ssa.Is32Bit(int64(c)+int64(d)) => (LEA(Q|L)2 [c+d] {s} x y)
   918  (ADD(Q|L)const [c] (LEA(Q|L)4 [d] {s} x y)) && ssa.Is32Bit(int64(c)+int64(d)) => (LEA(Q|L)4 [c+d] {s} x y)
   919  (ADD(Q|L)const [c] (LEA(Q|L)8 [d] {s} x y)) && ssa.Is32Bit(int64(c)+int64(d)) => (LEA(Q|L)8 [c+d] {s} x y)
   920  (LEA(Q|L)1 [c] {s} (ADD(Q|L)const [d] x) y) && ssa.Is32Bit(int64(c)+int64(d))   && x.Op != ssaop.OpSB => (LEA(Q|L)1 [c+d] {s} x y)
   921  (LEA(Q|L)2 [c] {s} (ADD(Q|L)const [d] x) y) && ssa.Is32Bit(int64(c)+int64(d))   && x.Op != ssaop.OpSB => (LEA(Q|L)2 [c+d] {s} x y)
   922  (LEA(Q|L)2 [c] {s} x (ADD(Q|L)const [d] y)) && ssa.Is32Bit(int64(c)+2*int64(d)) && y.Op != ssaop.OpSB => (LEA(Q|L)2 [c+2*d] {s} x y)
   923  (LEA(Q|L)4 [c] {s} (ADD(Q|L)const [d] x) y) && ssa.Is32Bit(int64(c)+int64(d))   && x.Op != ssaop.OpSB => (LEA(Q|L)4 [c+d] {s} x y)
   924  (LEA(Q|L)4 [c] {s} x (ADD(Q|L)const [d] y)) && ssa.Is32Bit(int64(c)+4*int64(d)) && y.Op != ssaop.OpSB => (LEA(Q|L)4 [c+4*d] {s} x y)
   925  (LEA(Q|L)8 [c] {s} (ADD(Q|L)const [d] x) y) && ssa.Is32Bit(int64(c)+int64(d))   && x.Op != ssaop.OpSB => (LEA(Q|L)8 [c+d] {s} x y)
   926  (LEA(Q|L)8 [c] {s} x (ADD(Q|L)const [d] y)) && ssa.Is32Bit(int64(c)+8*int64(d)) && y.Op != ssaop.OpSB => (LEA(Q|L)8 [c+8*d] {s} x y)
   927  
   928  // fold shifts into LEAQx/LEALx
   929  (LEA(Q|L)1 [c] {s} x z:(ADD(Q|L) y y)) && x != z => (LEA(Q|L)2 [c] {s} x y)
   930  (LEA(Q|L)1 [c] {s} x (SHL(Q|L)const [2] y)) => (LEA(Q|L)4 [c] {s} x y)
   931  (LEA(Q|L)1 [c] {s} x (SHL(Q|L)const [3] y)) => (LEA(Q|L)8 [c] {s} x y)
   932  (LEA(Q|L)2 [c] {s} x z:(ADD(Q|L) y y)) && x != z => (LEA(Q|L)4 [c] {s} x y)
   933  (LEA(Q|L)2 [c] {s} x (SHL(Q|L)const [2] y)) => (LEA(Q|L)8 [c] {s} x y)
   934  (LEA(Q|L)4 [c] {s} x z:(ADD(Q|L) y y)) && x != z => (LEA(Q|L)8 [c] {s} x y)
   935  
   936  // (x + x) << 1 -> x << 2
   937  (LEA(Q|L)2 [0] {s} (ADD(Q|L) x x) x) && s == nil => (SHL(Q|L)const [2] x)
   938  
   939  // (x + x) << 2 -> x << 3 and similar
   940  (SHLQconst [c] (ADDQ x x)) && c < 63 => (SHLQconst [c+1] x)
   941  (SHLLconst [c] (ADDL x x)) && c < 31 => (SHLLconst [c+1] x)
   942  
   943  // reverse ordering of compare instruction
   944  (SETL (InvertFlags x)) => (SETG x)
   945  (SETG (InvertFlags x)) => (SETL x)
   946  (SETB (InvertFlags x)) => (SETA x)
   947  (SETA (InvertFlags x)) => (SETB x)
   948  (SETLE (InvertFlags x)) => (SETGE x)
   949  (SETGE (InvertFlags x)) => (SETLE x)
   950  (SETBE (InvertFlags x)) => (SETAE x)
   951  (SETAE (InvertFlags x)) => (SETBE x)
   952  (SETEQ (InvertFlags x)) => (SETEQ x)
   953  (SETNE (InvertFlags x)) => (SETNE x)
   954  
   955  (SETLstore [off] {sym} ptr (InvertFlags x) mem) => (SETGstore [off] {sym} ptr x mem)
   956  (SETGstore [off] {sym} ptr (InvertFlags x) mem) => (SETLstore [off] {sym} ptr x mem)
   957  (SETBstore [off] {sym} ptr (InvertFlags x) mem) => (SETAstore [off] {sym} ptr x mem)
   958  (SETAstore [off] {sym} ptr (InvertFlags x) mem) => (SETBstore [off] {sym} ptr x mem)
   959  (SETLEstore [off] {sym} ptr (InvertFlags x) mem) => (SETGEstore [off] {sym} ptr x mem)
   960  (SETGEstore [off] {sym} ptr (InvertFlags x) mem) => (SETLEstore [off] {sym} ptr x mem)
   961  (SETBEstore [off] {sym} ptr (InvertFlags x) mem) => (SETAEstore [off] {sym} ptr x mem)
   962  (SETAEstore [off] {sym} ptr (InvertFlags x) mem) => (SETBEstore [off] {sym} ptr x mem)
   963  (SETEQstore [off] {sym} ptr (InvertFlags x) mem) => (SETEQstore [off] {sym} ptr x mem)
   964  (SETNEstore [off] {sym} ptr (InvertFlags x) mem) => (SETNEstore [off] {sym} ptr x mem)
   965  
   966  // sign extended loads
   967  // Note: The combined instruction must end up in the same block
   968  // as the original load. If not, we end up making a value with
   969  // memory type live in two different blocks, which can lead to
   970  // multiple memory values alive simultaneously.
   971  // Make sure we don't combine these ops if the load has another use.
   972  // This prevents a single load from being split into multiple loads
   973  // which then might return different values.  See test/atomicload.go.
   974  (MOVBQSX x:(MOVBload [off] {sym} ptr mem)) && x.Uses == 1 && ssa.Clobber(x) => @x.Block (MOVBQSXload <v.Type> [off] {sym} ptr mem)
   975  (MOVBQSX x:(MOVWload [off] {sym} ptr mem)) && x.Uses == 1 && ssa.Clobber(x) => @x.Block (MOVBQSXload <v.Type> [off] {sym} ptr mem)
   976  (MOVBQSX x:(MOVLload [off] {sym} ptr mem)) && x.Uses == 1 && ssa.Clobber(x) => @x.Block (MOVBQSXload <v.Type> [off] {sym} ptr mem)
   977  (MOVBQSX x:(MOVQload [off] {sym} ptr mem)) && x.Uses == 1 && ssa.Clobber(x) => @x.Block (MOVBQSXload <v.Type> [off] {sym} ptr mem)
   978  (MOVBQZX x:(MOVBload [off] {sym} ptr mem)) && x.Uses == 1 && ssa.Clobber(x) => @x.Block (MOVBload <v.Type> [off] {sym} ptr mem)
   979  (MOVBQZX x:(MOVWload [off] {sym} ptr mem)) && x.Uses == 1 && ssa.Clobber(x) => @x.Block (MOVBload <v.Type> [off] {sym} ptr mem)
   980  (MOVBQZX x:(MOVLload [off] {sym} ptr mem)) && x.Uses == 1 && ssa.Clobber(x) => @x.Block (MOVBload <v.Type> [off] {sym} ptr mem)
   981  (MOVBQZX x:(MOVQload [off] {sym} ptr mem)) && x.Uses == 1 && ssa.Clobber(x) => @x.Block (MOVBload <v.Type> [off] {sym} ptr mem)
   982  (MOVWQSX x:(MOVWload [off] {sym} ptr mem)) && x.Uses == 1 && ssa.Clobber(x) => @x.Block (MOVWQSXload <v.Type> [off] {sym} ptr mem)
   983  (MOVWQSX x:(MOVLload [off] {sym} ptr mem)) && x.Uses == 1 && ssa.Clobber(x) => @x.Block (MOVWQSXload <v.Type> [off] {sym} ptr mem)
   984  (MOVWQSX x:(MOVQload [off] {sym} ptr mem)) && x.Uses == 1 && ssa.Clobber(x) => @x.Block (MOVWQSXload <v.Type> [off] {sym} ptr mem)
   985  (MOVWQZX x:(MOVWload [off] {sym} ptr mem)) && x.Uses == 1 && ssa.Clobber(x) => @x.Block (MOVWload <v.Type> [off] {sym} ptr mem)
   986  (MOVWQZX x:(MOVLload [off] {sym} ptr mem)) && x.Uses == 1 && ssa.Clobber(x) => @x.Block (MOVWload <v.Type> [off] {sym} ptr mem)
   987  (MOVWQZX x:(MOVQload [off] {sym} ptr mem)) && x.Uses == 1 && ssa.Clobber(x) => @x.Block (MOVWload <v.Type> [off] {sym} ptr mem)
   988  (MOVLQSX x:(MOVLload [off] {sym} ptr mem)) && x.Uses == 1 && ssa.Clobber(x) => @x.Block (MOVLQSXload <v.Type> [off] {sym} ptr mem)
   989  (MOVLQSX x:(MOVQload [off] {sym} ptr mem)) && x.Uses == 1 && ssa.Clobber(x) => @x.Block (MOVLQSXload <v.Type> [off] {sym} ptr mem)
   990  (MOVLQZX x:(MOVLload [off] {sym} ptr mem)) && x.Uses == 1 && ssa.Clobber(x) => @x.Block (MOVLload <v.Type> [off] {sym} ptr mem)
   991  (MOVLQZX x:(MOVQload [off] {sym} ptr mem)) && x.Uses == 1 && ssa.Clobber(x) => @x.Block (MOVLload <v.Type> [off] {sym} ptr mem)
   992  
   993  // replace load from same location as preceding store with zero/sign extension (or copy in case of full width)
   994  (MOVBload [off] {sym} ptr (MOVBstore [off2] {sym2} ptr2 x _)) && sym == sym2 && off == off2 && ssa.IsSamePtr(ptr, ptr2) => (MOVBQZX x)
   995  (MOVWload [off] {sym} ptr (MOVWstore [off2] {sym2} ptr2 x _)) && sym == sym2 && off == off2 && ssa.IsSamePtr(ptr, ptr2) => (MOVWQZX x)
   996  (MOVLload [off] {sym} ptr (MOVLstore [off2] {sym2} ptr2 x _)) && sym == sym2 && off == off2 && ssa.IsSamePtr(ptr, ptr2) => (MOVLQZX x)
   997  (MOVQload [off] {sym} ptr (MOVQstore [off2] {sym2} ptr2 x _)) && sym == sym2 && off == off2 && ssa.IsSamePtr(ptr, ptr2) => x
   998  (MOVBQSXload [off] {sym} ptr (MOVBstore [off2] {sym2} ptr2 x _)) && sym == sym2 && off == off2 && ssa.IsSamePtr(ptr, ptr2) => (MOVBQSX x)
   999  (MOVWQSXload [off] {sym} ptr (MOVWstore [off2] {sym2} ptr2 x _)) && sym == sym2 && off == off2 && ssa.IsSamePtr(ptr, ptr2) => (MOVWQSX x)
  1000  (MOVLQSXload [off] {sym} ptr (MOVLstore [off2] {sym2} ptr2 x _)) && sym == sym2 && off == off2 && ssa.IsSamePtr(ptr, ptr2) => (MOVLQSX x)
  1001  
  1002  // Fold extensions and ANDs together.
  1003  (MOVBQZX (ANDLconst [c] x)) => (ANDLconst [c & 0xff] x)
  1004  (MOVWQZX (ANDLconst [c] x)) => (ANDLconst [c & 0xffff] x)
  1005  (MOVLQZX (ANDLconst [c] x)) => (ANDLconst [c] x)
  1006  (MOVBQSX (ANDLconst [c] x)) && c & 0x80 == 0 => (ANDLconst [c & 0x7f] x)
  1007  (MOVWQSX (ANDLconst [c] x)) && c & 0x8000 == 0 => (ANDLconst [c & 0x7fff] x)
  1008  (MOVLQSX (ANDLconst [c] x)) && uint32(c) & 0x80000000 == 0 => (ANDLconst [c & 0x7fffffff] x)
  1009  
  1010  // Don't extend before storing
  1011  (MOVLstore [off] {sym} ptr (MOVLQSX x) mem) => (MOVLstore [off] {sym} ptr x mem)
  1012  (MOVWstore [off] {sym} ptr (MOVWQSX x) mem) => (MOVWstore [off] {sym} ptr x mem)
  1013  (MOVBstore [off] {sym} ptr (MOVBQSX x) mem) => (MOVBstore [off] {sym} ptr x mem)
  1014  (MOVLstore [off] {sym} ptr (MOVLQZX x) mem) => (MOVLstore [off] {sym} ptr x mem)
  1015  (MOVWstore [off] {sym} ptr (MOVWQZX x) mem) => (MOVWstore [off] {sym} ptr x mem)
  1016  (MOVBstore [off] {sym} ptr (MOVBQZX x) mem) => (MOVBstore [off] {sym} ptr x mem)
  1017  
  1018  // fold constants into memory operations
  1019  // Note that this is not always a good idea because if not all the uses of
  1020  // the ADDQconst get eliminated, we still have to compute the ADDQconst and we now
  1021  // have potentially two live values (ptr and (ADDQconst [off] ptr)) instead of one.
  1022  // Nevertheless, let's do it!
  1023  (MOV(Q|L|W|B|SS|SD|O)load  [off1] {sym} (ADDQconst [off2] ptr) mem) && ssa.Is32Bit(int64(off1)+int64(off2)) =>
  1024      (MOV(Q|L|W|B|SS|SD|O)load  [off1+off2] {sym} ptr mem)
  1025  (MOV(Q|L|W|B|SS|SD|O)store  [off1] {sym} (ADDQconst [off2] ptr) val mem) && ssa.Is32Bit(int64(off1)+int64(off2)) =>
  1026  	(MOV(Q|L|W|B|SS|SD|O)store  [off1+off2] {sym} ptr val mem)
  1027  (SET(L|G|B|A|LE|GE|BE|AE|EQ|NE)store [off1] {sym} (ADDQconst [off2] base) val mem) && ssa.Is32Bit(int64(off1)+int64(off2)) =>
  1028  	(SET(L|G|B|A|LE|GE|BE|AE|EQ|NE)store [off1+off2] {sym} base val mem)
  1029  ((ADD|SUB|AND|OR|XOR)Qload [off1] {sym} val (ADDQconst [off2] base) mem) && ssa.Is32Bit(int64(off1)+int64(off2)) =>
  1030  	((ADD|SUB|AND|OR|XOR)Qload [off1+off2] {sym} val base mem)
  1031  ((ADD|SUB|AND|OR|XOR)Lload [off1] {sym} val (ADDQconst [off2] base) mem) && ssa.Is32Bit(int64(off1)+int64(off2)) =>
  1032  	((ADD|SUB|AND|OR|XOR)Lload [off1+off2] {sym} val base mem)
  1033  (CMP(Q|L|W|B)load [off1] {sym} (ADDQconst [off2] base) val mem) && ssa.Is32Bit(int64(off1)+int64(off2)) =>
  1034  	(CMP(Q|L|W|B)load [off1+off2] {sym} base val mem)
  1035  (CMP(Q|L|W|B)constload [valoff1] {sym} (ADDQconst [off2] base) mem) && ssa.ValAndOff(valoff1).CanAdd32(off2) =>
  1036  	(CMP(Q|L|W|B)constload [ssa.ValAndOff(valoff1).AddOffset32(off2)] {sym} base mem)
  1037  
  1038  ((ADD|SUB|MUL|DIV)SSload [off1] {sym} val (ADDQconst [off2] base) mem) && ssa.Is32Bit(int64(off1)+int64(off2)) =>
  1039  	((ADD|SUB|MUL|DIV)SSload [off1+off2] {sym} val base mem)
  1040  ((ADD|SUB|MUL|DIV)SDload [off1] {sym} val (ADDQconst [off2] base) mem) && ssa.Is32Bit(int64(off1)+int64(off2)) =>
  1041  	((ADD|SUB|MUL|DIV)SDload [off1+off2] {sym} val base mem)
  1042  ((ADD|AND|OR|XOR)Qconstmodify [valoff1] {sym} (ADDQconst [off2] base) mem) && ssa.ValAndOff(valoff1).CanAdd32(off2) =>
  1043  	((ADD|AND|OR|XOR)Qconstmodify [ssa.ValAndOff(valoff1).AddOffset32(off2)] {sym} base mem)
  1044  ((ADD|AND|OR|XOR)Lconstmodify [valoff1] {sym} (ADDQconst [off2] base) mem) && ssa.ValAndOff(valoff1).CanAdd32(off2) =>
  1045  	((ADD|AND|OR|XOR)Lconstmodify [ssa.ValAndOff(valoff1).AddOffset32(off2)] {sym} base mem)
  1046  ((ADD|AND|OR|XOR)Wconstmodify [valoff1] {sym} (ADDQconst [off2] base) mem) && ssa.ValAndOff(valoff1).CanAdd32(off2) =>
  1047  	((ADD|AND|OR|XOR)Wconstmodify [ssa.ValAndOff(valoff1).AddOffset32(off2)] {sym} base mem)
  1048  ((ADD|AND|OR|XOR)Bconstmodify [valoff1] {sym} (ADDQconst [off2] base) mem) && ssa.ValAndOff(valoff1).CanAdd32(off2) =>
  1049  	((ADD|AND|OR|XOR)Bconstmodify [ssa.ValAndOff(valoff1).AddOffset32(off2)] {sym} base mem)
  1050  ((ADD|SUB|AND|OR|XOR)Qmodify [off1] {sym} (ADDQconst [off2] base) val mem) && ssa.Is32Bit(int64(off1)+int64(off2)) =>
  1051  	((ADD|SUB|AND|OR|XOR)Qmodify [off1+off2] {sym} base val mem)
  1052  ((ADD|SUB|AND|OR|XOR)Lmodify [off1] {sym} (ADDQconst [off2] base) val mem) && ssa.Is32Bit(int64(off1)+int64(off2)) =>
  1053  	((ADD|SUB|AND|OR|XOR)Lmodify [off1+off2] {sym} base val mem)
  1054  
  1055  // Fold constants into stores.
  1056  (MOVQstore [off] {sym} ptr (MOVQconst [c]) mem) && validVal(c) =>
  1057  	(MOVQstoreconst [ssa.MakeValAndOff(int32(c),off)] {sym} ptr mem)
  1058  (MOVLstore [off] {sym} ptr (MOV(L|Q)const [c]) mem) =>
  1059  	(MOVLstoreconst [ssa.MakeValAndOff(int32(c),off)] {sym} ptr mem)
  1060  (MOVWstore [off] {sym} ptr (MOV(L|Q)const [c]) mem) =>
  1061  	(MOVWstoreconst [ssa.MakeValAndOff(int32(int16(c)),off)] {sym} ptr mem)
  1062  (MOVBstore [off] {sym} ptr (MOV(L|Q)const [c]) mem) =>
  1063  	(MOVBstoreconst [ssa.MakeValAndOff(int32(int8(c)),off)] {sym} ptr mem)
  1064  
  1065  // Fold address offsets into constant stores.
  1066  (MOV(Q|L|W|B|O)storeconst [sc] {s} (ADDQconst [off] ptr) mem) && ssa.ValAndOff(sc).CanAdd32(off) =>
  1067  	(MOV(Q|L|W|B|O)storeconst [ssa.ValAndOff(sc).AddOffset32(off)] {s} ptr mem)
  1068  
  1069  // We need to fold LEAQ into the MOVx ops so that the live variable analysis knows
  1070  // what variables are being read/written by the ops.
  1071  (MOV(Q|L|W|B|SS|SD|O|BQSX|WQSX|LQSX)load [off1] {sym1} (LEAQ [off2] {sym2} base) mem)
  1072  	&& ssa.Is32Bit(int64(off1)+int64(off2)) && ssa.CanMergeSym(sym1, sym2) =>
  1073  	(MOV(Q|L|W|B|SS|SD|O|BQSX|WQSX|LQSX)load [off1+off2] {ssa.MergeSym(sym1,sym2)} base mem)
  1074  (MOV(Q|L|W|B|SS|SD|O)store [off1] {sym1} (LEAQ [off2] {sym2} base) val mem)
  1075  	&& ssa.Is32Bit(int64(off1)+int64(off2)) && ssa.CanMergeSym(sym1, sym2) =>
  1076  	(MOV(Q|L|W|B|SS|SD|O)store [off1+off2] {ssa.MergeSym(sym1,sym2)} base val mem)
  1077  (MOV(Q|L|W|B|O)storeconst [sc] {sym1} (LEAQ [off] {sym2} ptr) mem) && ssa.CanMergeSym(sym1, sym2) && ssa.ValAndOff(sc).CanAdd32(off) =>
  1078  	(MOV(Q|L|W|B|O)storeconst [ssa.ValAndOff(sc).AddOffset32(off)] {ssa.MergeSym(sym1, sym2)} ptr mem)
  1079  (SET(L|G|B|A|LE|GE|BE|AE|EQ|NE)store [off1] {sym1} (LEAQ [off2] {sym2} base) val mem)
  1080  	&& ssa.Is32Bit(int64(off1)+int64(off2)) && ssa.CanMergeSym(sym1, sym2) =>
  1081  	(SET(L|G|B|A|LE|GE|BE|AE|EQ|NE)store [off1+off2] {ssa.MergeSym(sym1,sym2)} base val mem)
  1082  ((ADD|SUB|AND|OR|XOR)Qload [off1] {sym1} val (LEAQ [off2] {sym2} base) mem)
  1083  	&& ssa.Is32Bit(int64(off1)+int64(off2)) && ssa.CanMergeSym(sym1, sym2) =>
  1084  	((ADD|SUB|AND|OR|XOR)Qload [off1+off2] {ssa.MergeSym(sym1,sym2)} val base mem)
  1085  ((ADD|SUB|AND|OR|XOR)Lload [off1] {sym1} val (LEAQ [off2] {sym2} base) mem)
  1086  	&& ssa.Is32Bit(int64(off1)+int64(off2)) && ssa.CanMergeSym(sym1, sym2) =>
  1087  	((ADD|SUB|AND|OR|XOR)Lload [off1+off2] {ssa.MergeSym(sym1,sym2)} val base mem)
  1088  (CMP(Q|L|W|B)load [off1] {sym1} (LEAQ [off2] {sym2} base) val mem)
  1089  	&& ssa.Is32Bit(int64(off1)+int64(off2)) && ssa.CanMergeSym(sym1, sym2) =>
  1090  	(CMP(Q|L|W|B)load [off1+off2] {ssa.MergeSym(sym1,sym2)} base val mem)
  1091  (CMP(Q|L|W|B)constload [valoff1] {sym1} (LEAQ [off2] {sym2} base) mem)
  1092  	&& ssa.ValAndOff(valoff1).CanAdd32(off2) && ssa.CanMergeSym(sym1, sym2) =>
  1093  	(CMP(Q|L|W|B)constload [ssa.ValAndOff(valoff1).AddOffset32(off2)] {ssa.MergeSym(sym1,sym2)} base mem)
  1094  
  1095  ((ADD|SUB|MUL|DIV)SSload [off1] {sym1} val (LEAQ [off2] {sym2} base) mem)
  1096  	&& ssa.Is32Bit(int64(off1)+int64(off2)) && ssa.CanMergeSym(sym1, sym2) =>
  1097  	((ADD|SUB|MUL|DIV)SSload [off1+off2] {ssa.MergeSym(sym1,sym2)} val base mem)
  1098  ((ADD|SUB|MUL|DIV)SDload [off1] {sym1} val (LEAQ [off2] {sym2} base) mem)
  1099  	&& ssa.Is32Bit(int64(off1)+int64(off2)) && ssa.CanMergeSym(sym1, sym2) =>
  1100  	((ADD|SUB|MUL|DIV)SDload [off1+off2] {ssa.MergeSym(sym1,sym2)} val base mem)
  1101  ((ADD|AND|OR|XOR)Qconstmodify [valoff1] {sym1} (LEAQ [off2] {sym2} base) mem)
  1102  	&& ssa.ValAndOff(valoff1).CanAdd32(off2) && ssa.CanMergeSym(sym1, sym2) =>
  1103  	((ADD|AND|OR|XOR)Qconstmodify [ssa.ValAndOff(valoff1).AddOffset32(off2)] {ssa.MergeSym(sym1,sym2)} base mem)
  1104  ((ADD|AND|OR|XOR)Lconstmodify [valoff1] {sym1} (LEAQ [off2] {sym2} base) mem)
  1105  	&& ssa.ValAndOff(valoff1).CanAdd32(off2) && ssa.CanMergeSym(sym1, sym2) =>
  1106  	((ADD|AND|OR|XOR)Lconstmodify [ssa.ValAndOff(valoff1).AddOffset32(off2)] {ssa.MergeSym(sym1,sym2)} base mem)
  1107  ((ADD|AND|OR|XOR)Wconstmodify [valoff1] {sym1} (LEAQ [off2] {sym2} base) mem)
  1108  	&& ssa.ValAndOff(valoff1).CanAdd32(off2) && ssa.CanMergeSym(sym1, sym2) =>
  1109  	((ADD|AND|OR|XOR)Wconstmodify [ssa.ValAndOff(valoff1).AddOffset32(off2)] {ssa.MergeSym(sym1,sym2)} base mem)
  1110  ((ADD|AND|OR|XOR)Bconstmodify [valoff1] {sym1} (LEAQ [off2] {sym2} base) mem)
  1111  	&& ssa.ValAndOff(valoff1).CanAdd32(off2) && ssa.CanMergeSym(sym1, sym2) =>
  1112  	((ADD|AND|OR|XOR)Bconstmodify [ssa.ValAndOff(valoff1).AddOffset32(off2)] {ssa.MergeSym(sym1,sym2)} base mem)
  1113  ((ADD|SUB|AND|OR|XOR)Qmodify [off1] {sym1} (LEAQ [off2] {sym2} base) val mem)
  1114  	&& ssa.Is32Bit(int64(off1)+int64(off2)) && ssa.CanMergeSym(sym1, sym2) =>
  1115  	((ADD|SUB|AND|OR|XOR)Qmodify [off1+off2] {ssa.MergeSym(sym1,sym2)} base val mem)
  1116  ((ADD|SUB|AND|OR|XOR)Lmodify [off1] {sym1} (LEAQ [off2] {sym2} base) val mem)
  1117  	&& ssa.Is32Bit(int64(off1)+int64(off2)) && ssa.CanMergeSym(sym1, sym2) =>
  1118  	((ADD|SUB|AND|OR|XOR)Lmodify [off1+off2] {ssa.MergeSym(sym1,sym2)} base val mem)
  1119  
  1120  // fold LEAQs together
  1121  (LEAQ [off1] {sym1} (LEAQ [off2] {sym2} x)) && ssa.Is32Bit(int64(off1)+int64(off2)) && ssa.CanMergeSym(sym1, sym2) =>
  1122        (LEAQ [off1+off2] {ssa.MergeSym(sym1,sym2)} x)
  1123  
  1124  // LEAQ into LEAQ1
  1125  (LEAQ1 [off1] {sym1} (LEAQ [off2] {sym2} x) y) && ssa.Is32Bit(int64(off1)+int64(off2)) && ssa.CanMergeSym(sym1, sym2) && x.Op != ssaop.OpSB =>
  1126         (LEAQ1 [off1+off2] {ssa.MergeSym(sym1,sym2)} x y)
  1127  
  1128  // LEAQ1 into LEAQ
  1129  (LEAQ [off1] {sym1} (LEAQ1 [off2] {sym2} x y)) && ssa.Is32Bit(int64(off1)+int64(off2)) && ssa.CanMergeSym(sym1, sym2) =>
  1130         (LEAQ1 [off1+off2] {ssa.MergeSym(sym1,sym2)} x y)
  1131  
  1132  // LEAQ into LEAQ[248]
  1133  (LEAQ2 [off1] {sym1} (LEAQ [off2] {sym2} x) y) && ssa.Is32Bit(int64(off1)+int64(off2)) && ssa.CanMergeSym(sym1, sym2) && x.Op != ssaop.OpSB =>
  1134         (LEAQ2 [off1+off2] {ssa.MergeSym(sym1,sym2)} x y)
  1135  (LEAQ4 [off1] {sym1} (LEAQ [off2] {sym2} x) y) && ssa.Is32Bit(int64(off1)+int64(off2)) && ssa.CanMergeSym(sym1, sym2) && x.Op != ssaop.OpSB =>
  1136         (LEAQ4 [off1+off2] {ssa.MergeSym(sym1,sym2)} x y)
  1137  (LEAQ8 [off1] {sym1} (LEAQ [off2] {sym2} x) y) && ssa.Is32Bit(int64(off1)+int64(off2)) && ssa.CanMergeSym(sym1, sym2) && x.Op != ssaop.OpSB =>
  1138         (LEAQ8 [off1+off2] {ssa.MergeSym(sym1,sym2)} x y)
  1139  
  1140  // LEAQ[248] into LEAQ
  1141  (LEAQ [off1] {sym1} (LEAQ2 [off2] {sym2} x y)) && ssa.Is32Bit(int64(off1)+int64(off2)) && ssa.CanMergeSym(sym1, sym2) =>
  1142        (LEAQ2 [off1+off2] {ssa.MergeSym(sym1,sym2)} x y)
  1143  (LEAQ [off1] {sym1} (LEAQ4 [off2] {sym2} x y)) && ssa.Is32Bit(int64(off1)+int64(off2)) && ssa.CanMergeSym(sym1, sym2) =>
  1144        (LEAQ4 [off1+off2] {ssa.MergeSym(sym1,sym2)} x y)
  1145  (LEAQ [off1] {sym1} (LEAQ8 [off2] {sym2} x y)) && ssa.Is32Bit(int64(off1)+int64(off2)) && ssa.CanMergeSym(sym1, sym2) =>
  1146        (LEAQ8 [off1+off2] {ssa.MergeSym(sym1,sym2)} x y)
  1147  
  1148  // LEAQ[1248] into LEAQ[1248]. Only some such merges are possible.
  1149  (LEAQ1 [off1] {sym1} x (LEAQ1 [off2] {sym2} y y)) && ssa.Is32Bit(int64(off1)+int64(off2)) && ssa.CanMergeSym(sym1, sym2) =>
  1150        (LEAQ2 [off1+off2] {ssa.MergeSym(sym1, sym2)} x y)
  1151  (LEAQ1 [off1] {sym1} x (LEAQ1 [off2] {sym2} x y)) && ssa.Is32Bit(int64(off1)+int64(off2)) && ssa.CanMergeSym(sym1, sym2) =>
  1152        (LEAQ2 [off1+off2] {ssa.MergeSym(sym1, sym2)} y x)
  1153  (LEAQ2 [off1] {sym1} x (LEAQ1 [off2] {sym2} y y)) && ssa.Is32Bit(int64(off1)+2*int64(off2)) && sym2 == nil =>
  1154        (LEAQ4 [off1+2*off2] {sym1} x y)
  1155  (LEAQ4 [off1] {sym1} x (LEAQ1 [off2] {sym2} y y)) && ssa.Is32Bit(int64(off1)+4*int64(off2)) && sym2 == nil =>
  1156        (LEAQ8 [off1+4*off2] {sym1} x y)
  1157  // TODO: more?
  1158  
  1159  // Lower LEAQ2/4/8 when the offset is a constant
  1160  (LEAQ2 [off] {sym} x (MOV(Q|L)const [scale])) && ssa.Is32Bit(int64(off)+int64(scale)*2) =>
  1161  	(LEAQ [off+int32(scale)*2] {sym} x)
  1162  (LEAQ4 [off] {sym} x (MOV(Q|L)const [scale])) && ssa.Is32Bit(int64(off)+int64(scale)*4) =>
  1163  	(LEAQ [off+int32(scale)*4] {sym} x)
  1164  (LEAQ8 [off] {sym} x (MOV(Q|L)const [scale])) && ssa.Is32Bit(int64(off)+int64(scale)*8) =>
  1165  	(LEAQ [off+int32(scale)*8] {sym} x)
  1166  
  1167  // Absorb InvertFlags into branches.
  1168  (LT (InvertFlags cmp) yes no) => (GT cmp yes no)
  1169  (GT (InvertFlags cmp) yes no) => (LT cmp yes no)
  1170  (LE (InvertFlags cmp) yes no) => (GE cmp yes no)
  1171  (GE (InvertFlags cmp) yes no) => (LE cmp yes no)
  1172  (ULT (InvertFlags cmp) yes no) => (UGT cmp yes no)
  1173  (UGT (InvertFlags cmp) yes no) => (ULT cmp yes no)
  1174  (ULE (InvertFlags cmp) yes no) => (UGE cmp yes no)
  1175  (UGE (InvertFlags cmp) yes no) => (ULE cmp yes no)
  1176  (EQ (InvertFlags cmp) yes no) => (EQ cmp yes no)
  1177  (NE (InvertFlags cmp) yes no) => (NE cmp yes no)
  1178  
  1179  // Constant comparisons.
  1180  (CMPQconst (MOVQconst [x]) [y]) && x==int64(y) => (FlagEQ)
  1181  (CMPQconst (MOVQconst [x]) [y]) && x<int64(y) && uint64(x)<uint64(int64(y)) => (FlagLT_ULT)
  1182  (CMPQconst (MOVQconst [x]) [y]) && x<int64(y) && uint64(x)>uint64(int64(y)) => (FlagLT_UGT)
  1183  (CMPQconst (MOVQconst [x]) [y]) && x>int64(y) && uint64(x)<uint64(int64(y)) => (FlagGT_ULT)
  1184  (CMPQconst (MOVQconst [x]) [y]) && x>int64(y) && uint64(x)>uint64(int64(y)) => (FlagGT_UGT)
  1185  (CMPLconst (MOVLconst [x]) [y]) && x==y => (FlagEQ)
  1186  (CMPLconst (MOVLconst [x]) [y]) && x<y && uint32(x)<uint32(y) => (FlagLT_ULT)
  1187  (CMPLconst (MOVLconst [x]) [y]) && x<y && uint32(x)>uint32(y) => (FlagLT_UGT)
  1188  (CMPLconst (MOVLconst [x]) [y]) && x>y && uint32(x)<uint32(y) => (FlagGT_ULT)
  1189  (CMPLconst (MOVLconst [x]) [y]) && x>y && uint32(x)>uint32(y) => (FlagGT_UGT)
  1190  (CMPWconst (MOVLconst [x]) [y]) && int16(x)==y => (FlagEQ)
  1191  (CMPWconst (MOVLconst [x]) [y]) && int16(x)<y && uint16(x)<uint16(y) => (FlagLT_ULT)
  1192  (CMPWconst (MOVLconst [x]) [y]) && int16(x)<y && uint16(x)>uint16(y) => (FlagLT_UGT)
  1193  (CMPWconst (MOVLconst [x]) [y]) && int16(x)>y && uint16(x)<uint16(y) => (FlagGT_ULT)
  1194  (CMPWconst (MOVLconst [x]) [y]) && int16(x)>y && uint16(x)>uint16(y) => (FlagGT_UGT)
  1195  (CMPBconst (MOVLconst [x]) [y]) && int8(x)==y => (FlagEQ)
  1196  (CMPBconst (MOVLconst [x]) [y]) && int8(x)<y && uint8(x)<uint8(y) => (FlagLT_ULT)
  1197  (CMPBconst (MOVLconst [x]) [y]) && int8(x)<y && uint8(x)>uint8(y) => (FlagLT_UGT)
  1198  (CMPBconst (MOVLconst [x]) [y]) && int8(x)>y && uint8(x)<uint8(y) => (FlagGT_ULT)
  1199  (CMPBconst (MOVLconst [x]) [y]) && int8(x)>y && uint8(x)>uint8(y) => (FlagGT_UGT)
  1200  
  1201  // CMPQconst requires a 32 bit const, but we can still constant-fold 64 bit consts.
  1202  // In theory this applies to any of the simplifications above,
  1203  // but CMPQ is the only one I've actually seen occur.
  1204  (CMPQ (MOVQconst [x]) (MOVQconst [y])) && x==y => (FlagEQ)
  1205  (CMPQ (MOVQconst [x]) (MOVQconst [y])) && x<y && uint64(x)<uint64(y) => (FlagLT_ULT)
  1206  (CMPQ (MOVQconst [x]) (MOVQconst [y])) && x<y && uint64(x)>uint64(y) => (FlagLT_UGT)
  1207  (CMPQ (MOVQconst [x]) (MOVQconst [y])) && x>y && uint64(x)<uint64(y) => (FlagGT_ULT)
  1208  (CMPQ (MOVQconst [x]) (MOVQconst [y])) && x>y && uint64(x)>uint64(y) => (FlagGT_UGT)
  1209  
  1210  // Other known comparisons.
  1211  (CMPQconst (MOVBQZX _) [c]) && 0xFF < c => (FlagLT_ULT)
  1212  (CMPQconst (MOVWQZX _) [c]) && 0xFFFF < c => (FlagLT_ULT)
  1213  (CMPLconst (SHRLconst _ [c]) [n]) && 0 <= n && 0 < c && c <= 32 && (1<<uint64(32-c)) <= uint64(n) => (FlagLT_ULT)
  1214  (CMPQconst (SHRQconst _ [c]) [n]) && 0 <= n && 0 < c && c <= 64 && (1<<uint64(64-c)) <= uint64(n) => (FlagLT_ULT)
  1215  (CMPQconst (ANDQconst _ [m]) [n]) && 0 <= m && m < n => (FlagLT_ULT)
  1216  (CMPQconst (ANDLconst _ [m]) [n]) && 0 <= m && m < n => (FlagLT_ULT)
  1217  (CMPLconst (ANDLconst _ [m]) [n]) && 0 <= m && m < n => (FlagLT_ULT)
  1218  (CMPWconst (ANDLconst _ [m]) [n]) && 0 <= int16(m) && int16(m) < n => (FlagLT_ULT)
  1219  (CMPBconst (ANDLconst _ [m]) [n]) && 0 <= int8(m)  && int8(m)  < n => (FlagLT_ULT)
  1220  
  1221  // TESTQ c c sets flags like CMPQ c 0.
  1222  (TESTQconst [c] (MOVQconst [d])) && int64(c) == d && c == 0 => (FlagEQ)
  1223  (TESTLconst [c] (MOVLconst [c])) && c == 0 => (FlagEQ)
  1224  (TESTQconst [c] (MOVQconst [d])) && int64(c) == d && c < 0  => (FlagLT_UGT)
  1225  (TESTLconst [c] (MOVLconst [c])) && c < 0  => (FlagLT_UGT)
  1226  (TESTQconst [c] (MOVQconst [d])) && int64(c) == d && c > 0  => (FlagGT_UGT)
  1227  (TESTLconst [c] (MOVLconst [c])) && c > 0  => (FlagGT_UGT)
  1228  
  1229  // TODO: DIVxU also.
  1230  
  1231  // Absorb flag constants into SBB ops.
  1232  (SBBQcarrymask (FlagEQ))     => (MOVQconst [0])
  1233  (SBBQcarrymask (FlagLT_ULT)) => (MOVQconst [-1])
  1234  (SBBQcarrymask (FlagLT_UGT)) => (MOVQconst [0])
  1235  (SBBQcarrymask (FlagGT_ULT)) => (MOVQconst [-1])
  1236  (SBBQcarrymask (FlagGT_UGT)) => (MOVQconst [0])
  1237  (SBBLcarrymask (FlagEQ))     => (MOVLconst [0])
  1238  (SBBLcarrymask (FlagLT_ULT)) => (MOVLconst [-1])
  1239  (SBBLcarrymask (FlagLT_UGT)) => (MOVLconst [0])
  1240  (SBBLcarrymask (FlagGT_ULT)) => (MOVLconst [-1])
  1241  (SBBLcarrymask (FlagGT_UGT)) => (MOVLconst [0])
  1242  
  1243  // Absorb flag constants into branches.
  1244  ((EQ|LE|GE|ULE|UGE) (FlagEQ) yes no)     => (First yes no)
  1245  ((NE|LT|GT|ULT|UGT) (FlagEQ) yes no)     => (First no yes)
  1246  ((NE|LT|LE|ULT|ULE) (FlagLT_ULT) yes no) => (First yes no)
  1247  ((EQ|GT|GE|UGT|UGE) (FlagLT_ULT) yes no) => (First no yes)
  1248  ((NE|LT|LE|UGT|UGE) (FlagLT_UGT) yes no) => (First yes no)
  1249  ((EQ|GT|GE|ULT|ULE) (FlagLT_UGT) yes no) => (First no yes)
  1250  ((NE|GT|GE|ULT|ULE) (FlagGT_ULT) yes no) => (First yes no)
  1251  ((EQ|LT|LE|UGT|UGE) (FlagGT_ULT) yes no) => (First no yes)
  1252  ((NE|GT|GE|UGT|UGE) (FlagGT_UGT) yes no) => (First yes no)
  1253  ((EQ|LT|LE|ULT|ULE) (FlagGT_UGT) yes no) => (First no yes)
  1254  
  1255  // Absorb flag constants into SETxx ops.
  1256  ((SETEQ|SETLE|SETGE|SETBE|SETAE) (FlagEQ))     => (MOVLconst [1])
  1257  ((SETNE|SETL|SETG|SETB|SETA)     (FlagEQ))     => (MOVLconst [0])
  1258  ((SETNE|SETL|SETLE|SETB|SETBE)   (FlagLT_ULT)) => (MOVLconst [1])
  1259  ((SETEQ|SETG|SETGE|SETA|SETAE)   (FlagLT_ULT)) => (MOVLconst [0])
  1260  ((SETNE|SETL|SETLE|SETA|SETAE)   (FlagLT_UGT)) => (MOVLconst [1])
  1261  ((SETEQ|SETG|SETGE|SETB|SETBE)   (FlagLT_UGT)) => (MOVLconst [0])
  1262  ((SETNE|SETG|SETGE|SETB|SETBE)   (FlagGT_ULT)) => (MOVLconst [1])
  1263  ((SETEQ|SETL|SETLE|SETA|SETAE)   (FlagGT_ULT)) => (MOVLconst [0])
  1264  ((SETNE|SETG|SETGE|SETA|SETAE)   (FlagGT_UGT)) => (MOVLconst [1])
  1265  ((SETEQ|SETL|SETLE|SETB|SETBE)   (FlagGT_UGT)) => (MOVLconst [0])
  1266  
  1267  (SETEQstore [off] {sym} ptr (FlagEQ)     mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1268  (SETEQstore [off] {sym} ptr (FlagLT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1269  (SETEQstore [off] {sym} ptr (FlagLT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1270  (SETEQstore [off] {sym} ptr (FlagGT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1271  (SETEQstore [off] {sym} ptr (FlagGT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1272  
  1273  (SETNEstore [off] {sym} ptr (FlagEQ)     mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1274  (SETNEstore [off] {sym} ptr (FlagLT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1275  (SETNEstore [off] {sym} ptr (FlagLT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1276  (SETNEstore [off] {sym} ptr (FlagGT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1277  (SETNEstore [off] {sym} ptr (FlagGT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1278  
  1279  (SETLstore  [off] {sym} ptr (FlagEQ)     mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1280  (SETLstore  [off] {sym} ptr (FlagLT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1281  (SETLstore  [off] {sym} ptr (FlagLT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1282  (SETLstore  [off] {sym} ptr (FlagGT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1283  (SETLstore  [off] {sym} ptr (FlagGT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1284  
  1285  (SETLEstore [off] {sym} ptr (FlagEQ)     mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1286  (SETLEstore [off] {sym} ptr (FlagLT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1287  (SETLEstore [off] {sym} ptr (FlagLT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1288  (SETLEstore [off] {sym} ptr (FlagGT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1289  (SETLEstore [off] {sym} ptr (FlagGT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1290  
  1291  (SETGstore  [off] {sym} ptr (FlagEQ)     mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1292  (SETGstore  [off] {sym} ptr (FlagLT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1293  (SETGstore  [off] {sym} ptr (FlagLT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1294  (SETGstore  [off] {sym} ptr (FlagGT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1295  (SETGstore  [off] {sym} ptr (FlagGT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1296  
  1297  (SETGEstore [off] {sym} ptr (FlagEQ)     mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1298  (SETGEstore [off] {sym} ptr (FlagLT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1299  (SETGEstore [off] {sym} ptr (FlagLT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1300  (SETGEstore [off] {sym} ptr (FlagGT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1301  (SETGEstore [off] {sym} ptr (FlagGT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1302  
  1303  (SETBstore  [off] {sym} ptr (FlagEQ)     mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1304  (SETBstore  [off] {sym} ptr (FlagLT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1305  (SETBstore  [off] {sym} ptr (FlagLT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1306  (SETBstore  [off] {sym} ptr (FlagGT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1307  (SETBstore  [off] {sym} ptr (FlagGT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1308  
  1309  (SETBEstore [off] {sym} ptr (FlagEQ)     mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1310  (SETBEstore [off] {sym} ptr (FlagLT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1311  (SETBEstore [off] {sym} ptr (FlagLT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1312  (SETBEstore [off] {sym} ptr (FlagGT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1313  (SETBEstore [off] {sym} ptr (FlagGT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1314  
  1315  (SETAstore  [off] {sym} ptr (FlagEQ)     mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1316  (SETAstore  [off] {sym} ptr (FlagLT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1317  (SETAstore  [off] {sym} ptr (FlagLT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1318  (SETAstore  [off] {sym} ptr (FlagGT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1319  (SETAstore  [off] {sym} ptr (FlagGT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1320  
  1321  (SETAEstore [off] {sym} ptr (FlagEQ)     mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1322  (SETAEstore [off] {sym} ptr (FlagLT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1323  (SETAEstore [off] {sym} ptr (FlagLT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1324  (SETAEstore [off] {sym} ptr (FlagGT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1325  (SETAEstore [off] {sym} ptr (FlagGT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1326  
  1327  // Remove redundant *const ops
  1328  (ADD(Q|L)const [0] x) => x
  1329  (SUB(Q|L)const [0] x) => x
  1330  (AND(Q|L)const [0] _) => (MOVLconst [0])
  1331  (AND(Q|L)const [-1] x) => x
  1332  (OR(Q|L)const [0] x) => x
  1333  (OR(Q|L)const [-1] _) => (MOV(Q|L)const [-1])
  1334  (XOR(Q|L)const [0] x) => x
  1335  // TODO: since we got rid of the W/B versions, we might miss
  1336  // things like (ANDLconst [0x100] x) which were formerly
  1337  // (ANDBconst [0] x).  Probably doesn't happen very often.
  1338  // If we cared, we might do:
  1339  //  (ANDLconst <t> [c] x) && t.Size()==1 && int8(x)==0 -> (MOVLconst [0])
  1340  
  1341  // Remove redundant ops
  1342  // Not in generic rules, because they may appear after lowering e. g. Slicemask
  1343  (NEG(Q|L) (NEG(Q|L) x)) => x
  1344  (NEG(Q|L) s:(SUB(Q|L) x y)) && s.Uses == 1 => (SUB(Q|L) y x)
  1345  
  1346  // Convert constant subtracts to constant adds
  1347  (SUBQconst [c] x) && c != -(1<<31) => (ADDQconst [-c] x)
  1348  (SUBLconst [c] x) => (ADDLconst [-c] x)
  1349  
  1350  // generic constant folding
  1351  // TODO: more of this
  1352  (ADDQconst [c] (MOVQconst [d])) => (MOVQconst [int64(c)+d])
  1353  (ADDLconst [c] (MOVLconst [d])) => (MOVLconst [c+d])
  1354  (ADDQconst [c] (ADDQconst [d] x)) && ssa.Is32Bit(int64(c)+int64(d)) => (ADDQconst [c+d] x)
  1355  (ADDLconst [c] (ADDLconst [d] x)) => (ADDLconst [c+d] x)
  1356  (SUBQconst (MOVQconst [d]) [c]) => (MOVQconst [d-int64(c)])
  1357  (SUBQconst (SUBQconst x [d]) [c]) && ssa.Is32Bit(int64(-c)-int64(d)) => (ADDQconst [-c-d] x)
  1358  (SARQconst [c] (MOVQconst [d])) => (MOVQconst [d>>uint64(c)])
  1359  (SARLconst [c] (MOVQconst [d])) => (MOVQconst [int64(int32(d))>>uint64(c)])
  1360  (SARWconst [c] (MOVQconst [d])) => (MOVQconst [int64(int16(d))>>uint64(c)])
  1361  (SARBconst [c] (MOVQconst [d])) => (MOVQconst [int64(int8(d))>>uint64(c)])
  1362  (NEG(Q|L) (MOV(Q|L)const [c])) => (MOV(Q|L)const [-c])
  1363  (MULQconst [c] (MOVQconst [d])) => (MOVQconst [int64(c)*d])
  1364  (MULLconst [c] (MOVLconst [d])) => (MOVLconst [c*d])
  1365  (ANDQconst [c] (MOVQconst [d])) => (MOVQconst [int64(c)&d])
  1366  (ANDLconst [c] (MOVLconst [d])) => (MOVLconst [c&d])
  1367  (ORQconst [c] (MOVQconst [d])) => (MOVQconst [int64(c)|d])
  1368  (ORLconst [c] (MOVLconst [d])) => (MOVLconst [c|d])
  1369  (XORQconst [c] (MOVQconst [d])) => (MOVQconst [int64(c)^d])
  1370  (XORLconst [c] (MOVLconst [d])) => (MOVLconst [c^d])
  1371  (NOT(Q|L) (MOV(Q|L)const [c])) => (MOV(Q|L)const [^c])
  1372  (BTSQconst [c] (MOVQconst [d])) => (MOVQconst [d|(1<<uint32(c))])
  1373  (BTRQconst [c] (MOVQconst [d])) => (MOVQconst [d&^(1<<uint32(c))])
  1374  (BTCQconst [c] (MOVQconst [d])) => (MOVQconst [d^(1<<uint32(c))])
  1375  
  1376  // If c or d doesn't fit into 32 bits, then we can't construct ORQconst,
  1377  // but we can still constant-fold.
  1378  // In theory this applies to any of the simplifications above,
  1379  // but ORQ is the only one I've actually seen occur.
  1380  (ORQ (MOVQconst [c]) (MOVQconst [d])) => (MOVQconst [c|d])
  1381  
  1382  // generic simplifications
  1383  // TODO: more of this
  1384  (ADD(Q|L) x (NEG(Q|L) y)) => (SUB(Q|L) x y)
  1385  (SUB(Q|L) x x) => (MOVLconst [0])
  1386  (AND(Q|L) x x) => x
  1387  (OR(Q|L) x x)  => x
  1388  (XOR(Q|L) x x) => (MOVLconst [0])
  1389  
  1390  (SHLLconst [d] (MOVLconst [c])) => (MOVLconst [c << uint64(d)])
  1391  (SHLQconst [d] (MOVQconst [c])) => (MOVQconst [c << uint64(d)])
  1392  (SHLQconst [d] (MOVLconst [c])) => (MOVQconst [int64(c) << uint64(d)])
  1393  
  1394  // Fold NEG into ADDconst/MULconst. Take care to keep c in 32 bit range.
  1395  (NEGQ (ADDQconst [c] (NEGQ x))) && c != -(1<<31) => (ADDQconst [-c] x)
  1396  (MULQconst [c] (NEGQ x)) && c != -(1<<31) => (MULQconst [-c] x)
  1397  
  1398  // checking AND against 0.
  1399  (CMP(Q|L|W|B)const a:(AND(Q|L|L|L) x y) [0]) && a.Uses == 1 => (TEST(Q|L|W|B) x y)
  1400  (CMPQconst a:(ANDQconst [c] x) [0]) && a.Uses == 1 => (TESTQconst [c] x)
  1401  (CMPLconst a:(ANDLconst [c] x) [0]) && a.Uses == 1 => (TESTLconst [c] x)
  1402  (CMPWconst a:(ANDLconst [c] x) [0]) && a.Uses == 1 => (TESTWconst [int16(c)] x)
  1403  (CMPBconst a:(ANDLconst [c] x) [0]) && a.Uses == 1 => (TESTBconst [int8(c)] x)
  1404  
  1405  // Convert TESTx to TESTxconst if possible.
  1406  (TESTQ (MOVQconst [c]) x) && ssa.Is32Bit(c) => (TESTQconst [int32(c)] x)
  1407  (TESTL (MOVLconst [c]) x) => (TESTLconst [c] x)
  1408  (TESTW (MOVLconst [c]) x) => (TESTWconst [int16(c)] x)
  1409  (TESTB (MOVLconst [c]) x) => (TESTBconst [int8(c)] x)
  1410  
  1411  // TEST %reg,%reg is shorter than CMP
  1412  (CMP(Q|L|W|B)const x [0]) => (TEST(Q|L|W|B) x x)
  1413  (TESTQconst [-1] x) && x.Op != ssaop.OpAMD64MOVQconst => (TESTQ x x)
  1414  (TESTLconst [-1] x) && x.Op != ssaop.OpAMD64MOVLconst => (TESTL x x)
  1415  (TESTWconst [-1] x) && x.Op != ssaop.OpAMD64MOVLconst => (TESTW x x)
  1416  (TESTBconst [-1] x) && x.Op != ssaop.OpAMD64MOVLconst => (TESTB x x)
  1417  
  1418  // Convert LEAQ1 back to ADDQ if we can
  1419  (LEAQ1 [0] x y) && v.Aux == nil => (ADDQ x y)
  1420  
  1421  (MOVQstoreconst [c] {s} p1 x:(MOVQstoreconst [a] {s} p0 mem))
  1422    && x.Uses == 1
  1423    && sequentialAddresses(p0, p1, int64(a.Off()+8-c.Off()))
  1424    && a.Val() == 0
  1425    && c.Val() == 0
  1426    && ssa.SetPos(v, x.Pos)
  1427    && ssa.Clobber(x)
  1428    => (MOVOstoreconst [ssa.MakeValAndOff(0,a.Off())] {s} p0 mem)
  1429  (MOVQstoreconst [a] {s} p0 x:(MOVQstoreconst [c] {s} p1 mem))
  1430    && x.Uses == 1
  1431    && sequentialAddresses(p0, p1, int64(a.Off()+8-c.Off()))
  1432    && a.Val() == 0
  1433    && c.Val() == 0
  1434    && ssa.SetPos(v, x.Pos)
  1435    && ssa.Clobber(x)
  1436    => (MOVOstoreconst [ssa.MakeValAndOff(0,a.Off())] {s} p0 mem)
  1437  
  1438  // Merge load and op
  1439  // TODO: add indexed variants?
  1440  ((ADD|SUB|AND|OR|XOR)Q x l:(MOVQload [off] {sym} ptr mem)) && ssa.CanMergeLoadClobber(v, l, x) && ssa.Clobber(l) => ((ADD|SUB|AND|OR|XOR)Qload x [off] {sym} ptr mem)
  1441  ((ADD|SUB|AND|OR|XOR)L x l:(MOVLload [off] {sym} ptr mem)) && ssa.CanMergeLoadClobber(v, l, x) && ssa.Clobber(l) => ((ADD|SUB|AND|OR|XOR)Lload x [off] {sym} ptr mem)
  1442  ((ADD|SUB|MUL|DIV)SD x l:(MOVSDload [off] {sym} ptr mem)) && ssa.CanMergeLoadClobber(v, l, x) && ssa.Clobber(l) => ((ADD|SUB|MUL|DIV)SDload x [off] {sym} ptr mem)
  1443  ((ADD|SUB|MUL|DIV)SS x l:(MOVSSload [off] {sym} ptr mem)) && ssa.CanMergeLoadClobber(v, l, x) && ssa.Clobber(l) => ((ADD|SUB|MUL|DIV)SSload x [off] {sym} ptr mem)
  1444  (MOVLstore {sym} [off] ptr y:((ADD|AND|OR|XOR)Lload x [off] {sym} ptr mem) mem) && y.Uses==1 && ssa.Clobber(y) => ((ADD|AND|OR|XOR)Lmodify [off] {sym} ptr x mem)
  1445  (MOVLstore {sym} [off] ptr y:((ADD|SUB|AND|OR|XOR)L l:(MOVLload [off] {sym} ptr mem) x) mem) && y.Uses==1 && l.Uses==1 && ssa.Clobber(y, l) =>
  1446  	((ADD|SUB|AND|OR|XOR)Lmodify [off] {sym} ptr x mem)
  1447  (MOVQstore {sym} [off] ptr y:((ADD|AND|OR|XOR)Qload x [off] {sym} ptr mem) mem) && y.Uses==1 && ssa.Clobber(y) => ((ADD|AND|OR|XOR)Qmodify [off] {sym} ptr x mem)
  1448  (MOVQstore {sym} [off] ptr y:((ADD|SUB|AND|OR|XOR)Q l:(MOVQload [off] {sym} ptr mem) x) mem) && y.Uses==1 && l.Uses==1 && ssa.Clobber(y, l) =>
  1449  	((ADD|SUB|AND|OR|XOR)Qmodify [off] {sym} ptr x mem)
  1450  (MOVQstore {sym} [off] ptr x:(BT(S|R|C)Qconst [c] l:(MOVQload {sym} [off] ptr mem)) mem) && x.Uses == 1 && l.Uses == 1 && ssa.Clobber(x, l) =>
  1451  	(BT(S|R|C)Qconstmodify {sym} [ssa.MakeValAndOff(int32(c),off)] ptr mem)
  1452  
  1453  // Merge ADDQconst and LEAQ into atomic loads.
  1454  (MOV(Q|L|B)atomicload [off1] {sym} (ADDQconst [off2] ptr) mem) && ssa.Is32Bit(int64(off1)+int64(off2)) =>
  1455  	(MOV(Q|L|B)atomicload [off1+off2] {sym} ptr mem)
  1456  (MOV(Q|L|B)atomicload [off1] {sym1} (LEAQ [off2] {sym2} ptr) mem) && ssa.Is32Bit(int64(off1)+int64(off2)) && ssa.CanMergeSym(sym1, sym2) =>
  1457  	(MOV(Q|L|B)atomicload [off1+off2] {ssa.MergeSym(sym1, sym2)} ptr mem)
  1458  
  1459  // Merge ADDQconst and LEAQ into atomic stores.
  1460  (XCHGQ [off1] {sym} val (ADDQconst [off2] ptr) mem) && ssa.Is32Bit(int64(off1)+int64(off2)) =>
  1461  	(XCHGQ [off1+off2] {sym} val ptr mem)
  1462  (XCHGQ [off1] {sym1} val (LEAQ [off2] {sym2} ptr) mem) && ssa.Is32Bit(int64(off1)+int64(off2)) && ssa.CanMergeSym(sym1, sym2) && ptr.Op != ssaop.OpSB =>
  1463  	(XCHGQ [off1+off2] {ssa.MergeSym(sym1,sym2)} val ptr mem)
  1464  (XCHGL [off1] {sym} val (ADDQconst [off2] ptr) mem) && ssa.Is32Bit(int64(off1)+int64(off2)) =>
  1465  	(XCHGL [off1+off2] {sym} val ptr mem)
  1466  (XCHGL [off1] {sym1} val (LEAQ [off2] {sym2} ptr) mem) && ssa.Is32Bit(int64(off1)+int64(off2)) && ssa.CanMergeSym(sym1, sym2) && ptr.Op != ssaop.OpSB =>
  1467  	(XCHGL [off1+off2] {ssa.MergeSym(sym1,sym2)} val ptr mem)
  1468  
  1469  // Merge ADDQconst into atomic adds.
  1470  // TODO: merging LEAQ doesn't work, assembler doesn't like the resulting instructions.
  1471  (XADDQlock [off1] {sym} val (ADDQconst [off2] ptr) mem) && ssa.Is32Bit(int64(off1)+int64(off2)) =>
  1472  	(XADDQlock [off1+off2] {sym} val ptr mem)
  1473  (XADDLlock [off1] {sym} val (ADDQconst [off2] ptr) mem) && ssa.Is32Bit(int64(off1)+int64(off2)) =>
  1474  	(XADDLlock [off1+off2] {sym} val ptr mem)
  1475  
  1476  // Merge ADDQconst into atomic compare and swaps.
  1477  // TODO: merging LEAQ doesn't work, assembler doesn't like the resulting instructions.
  1478  (CMPXCHGQlock [off1] {sym} (ADDQconst [off2] ptr) old new_ mem) && ssa.Is32Bit(int64(off1)+int64(off2)) =>
  1479  	(CMPXCHGQlock [off1+off2] {sym} ptr old new_ mem)
  1480  (CMPXCHGLlock [off1] {sym} (ADDQconst [off2] ptr) old new_ mem) && ssa.Is32Bit(int64(off1)+int64(off2)) =>
  1481  	(CMPXCHGLlock [off1+off2] {sym} ptr old new_ mem)
  1482  
  1483  // We don't need the conditional move if we know the arg of BSF is not zero.
  1484  (CMOVQEQ x _ (Select1 (BS(F|R)Q (ORQconst [c] _)))) && c != 0 => x
  1485  // Extension is unnecessary for trailing zeros.
  1486  (BSFQ (ORQconst <t> [1<<8] (MOVBQZX x))) => (BSFQ (ORQconst <t> [1<<8] x))
  1487  (BSFQ (ORQconst <t> [1<<16] (MOVWQZX x))) => (BSFQ (ORQconst <t> [1<<16] x))
  1488  
  1489  // Redundant sign/zero extensions
  1490  // Note: see issue 21963. We have to make sure we use the right type on
  1491  // the resulting extension (the outer type, not the inner type).
  1492  (MOVLQSX (MOVLQSX x)) => (MOVLQSX x)
  1493  (MOVLQSX (MOVWQSX x)) => (MOVWQSX x)
  1494  (MOVLQSX (MOVBQSX x)) => (MOVBQSX x)
  1495  (MOVWQSX (MOVWQSX x)) => (MOVWQSX x)
  1496  (MOVWQSX (MOVBQSX x)) => (MOVBQSX x)
  1497  (MOVBQSX (MOVBQSX x)) => (MOVBQSX x)
  1498  (MOVLQZX (MOVLQZX x)) => (MOVLQZX x)
  1499  (MOVLQZX (MOVWQZX x)) => (MOVWQZX x)
  1500  (MOVLQZX (MOVBQZX x)) => (MOVBQZX x)
  1501  (MOVWQZX (MOVWQZX x)) => (MOVWQZX x)
  1502  (MOVWQZX (MOVBQZX x)) => (MOVBQZX x)
  1503  (MOVBQZX (MOVBQZX x)) => (MOVBQZX x)
  1504  
  1505  (MOVQstore [off] {sym} ptr a:((ADD|AND|OR|XOR)Qconst [c] l:(MOVQload [off] {sym} ptr2 mem)) mem)
  1506  	&& ssa.IsSamePtr(ptr, ptr2) && a.Uses == 1 && l.Uses == 1 && ssa.Clobber(l, a) =>
  1507  	((ADD|AND|OR|XOR)Qconstmodify {sym} [ssa.MakeValAndOff(int32(c),off)] ptr mem)
  1508  (MOVLstore [off] {sym} ptr a:((ADD|AND|OR|XOR)Lconst [c] l:(MOVLload [off] {sym} ptr2 mem)) mem)
  1509  	&& ssa.IsSamePtr(ptr, ptr2) && a.Uses == 1 && l.Uses == 1 && ssa.Clobber(l, a) =>
  1510  	((ADD|AND|OR|XOR)Lconstmodify {sym} [ssa.MakeValAndOff(int32(c),off)] ptr mem)
  1511  // Sub-word read-modify-write. The narrow store keeps only the low 8/16 bits of
  1512  // the 32-bit result, so ADDB/ADDW etc. with the same constant is equivalent.
  1513  // The constant is restricted to an 8-bit immediate: it keeps every form we emit
  1514  // encodable as imm8 (or INC/DEC), which also avoids the 16-bit immediates that
  1515  // cause length-changing-prefix stalls.
  1516  (MOVWstore [off] {sym} ptr a:((ADD|AND|OR|XOR)Lconst [c] l:(MOVWload [off] {sym} ptr2 mem)) mem)
  1517  	&& ssa.IsSamePtr(ptr, ptr2) && a.Uses == 1 && l.Uses == 1 && c == int32(int8(c)) && ssa.Clobber(l, a) =>
  1518  	((ADD|AND|OR|XOR)Wconstmodify {sym} [ssa.MakeValAndOff(int32(c),off)] ptr mem)
  1519  (MOVBstore [off] {sym} ptr a:((ADD|AND|OR|XOR)Lconst [c] l:(MOVBload [off] {sym} ptr2 mem)) mem)
  1520  	&& ssa.IsSamePtr(ptr, ptr2) && a.Uses == 1 && l.Uses == 1 && c == int32(int8(c)) && ssa.Clobber(l, a) =>
  1521  	((ADD|AND|OR|XOR)Bconstmodify {sym} [ssa.MakeValAndOff(int32(c),off)] ptr mem)
  1522  
  1523  // float <-> int register moves, with no conversion.
  1524  // These come up when compiling math.{Float{32,64}bits,Float{32,64}frombits}.
  1525  (MOVQload  [off] {sym} ptr (MOVSDstore [off] {sym} ptr val _)) => (MOVQf2i val)
  1526  (MOVLload  [off] {sym} ptr (MOVSSstore [off] {sym} ptr val _)) => (MOVLf2i val)
  1527  (MOVSDload [off] {sym} ptr (MOVQstore  [off] {sym} ptr val _)) => (MOVQi2f val)
  1528  (MOVSSload [off] {sym} ptr (MOVLstore  [off] {sym} ptr val _)) => (MOVLi2f val)
  1529  
  1530  // Other load-like ops.
  1531  (ADDQload x [off] {sym} ptr (MOVSDstore [off] {sym} ptr y _)) => (ADDQ x (MOVQf2i y))
  1532  (ADDLload x [off] {sym} ptr (MOVSSstore [off] {sym} ptr y _)) => (ADDL x (MOVLf2i y))
  1533  (SUBQload x [off] {sym} ptr (MOVSDstore [off] {sym} ptr y _)) => (SUBQ x (MOVQf2i y))
  1534  (SUBLload x [off] {sym} ptr (MOVSSstore [off] {sym} ptr y _)) => (SUBL x (MOVLf2i y))
  1535  (ANDQload x [off] {sym} ptr (MOVSDstore [off] {sym} ptr y _)) => (ANDQ x (MOVQf2i y))
  1536  (ANDLload x [off] {sym} ptr (MOVSSstore [off] {sym} ptr y _)) => (ANDL x (MOVLf2i y))
  1537  ( ORQload x [off] {sym} ptr (MOVSDstore [off] {sym} ptr y _)) => ( ORQ x (MOVQf2i y))
  1538  ( ORLload x [off] {sym} ptr (MOVSSstore [off] {sym} ptr y _)) => ( ORL x (MOVLf2i y))
  1539  (XORQload x [off] {sym} ptr (MOVSDstore [off] {sym} ptr y _)) => (XORQ x (MOVQf2i y))
  1540  (XORLload x [off] {sym} ptr (MOVSSstore [off] {sym} ptr y _)) => (XORL x (MOVLf2i y))
  1541  
  1542  (ADDSDload x [off] {sym} ptr (MOVQstore [off] {sym} ptr y _)) => (ADDSD x (MOVQi2f y))
  1543  (ADDSSload x [off] {sym} ptr (MOVLstore [off] {sym} ptr y _)) => (ADDSS x (MOVLi2f y))
  1544  (SUBSDload x [off] {sym} ptr (MOVQstore [off] {sym} ptr y _)) => (SUBSD x (MOVQi2f y))
  1545  (SUBSSload x [off] {sym} ptr (MOVLstore [off] {sym} ptr y _)) => (SUBSS x (MOVLi2f y))
  1546  (MULSDload x [off] {sym} ptr (MOVQstore [off] {sym} ptr y _)) => (MULSD x (MOVQi2f y))
  1547  (MULSSload x [off] {sym} ptr (MOVLstore [off] {sym} ptr y _)) => (MULSS x (MOVLi2f y))
  1548  
  1549  // Detect FMA
  1550  (ADDS(S|D) (MULS(S|D) x y) z) && buildcfg.GOAMD64 >= 3 && z.Block.Func.UseFMA(v) => (VFMADD231S(S|D) z x y)
  1551  (SUBS(S|D) (MULS(S|D) x y) z) && buildcfg.GOAMD64 >= 3 && z.Block.Func.UseFMA(v) => (VFMSUB231S(S|D) z x y)
  1552  (SUBS(S|D) x (MULS(S|D) y z)) && buildcfg.GOAMD64 >= 3 && z.Block.Func.UseFMA(v) => (VFNMADD231S(S|D) x y z)
  1553  
  1554  // Redirect stores to use the other register set.
  1555  (MOVQstore  [off] {sym} ptr (MOVQf2i val) mem) => (MOVSDstore [off] {sym} ptr val mem)
  1556  (MOVLstore  [off] {sym} ptr (MOVLf2i val) mem) => (MOVSSstore [off] {sym} ptr val mem)
  1557  (MOVSDstore [off] {sym} ptr (MOVQi2f val) mem) => (MOVQstore  [off] {sym} ptr val mem)
  1558  (MOVSSstore [off] {sym} ptr (MOVLi2f val) mem) => (MOVLstore  [off] {sym} ptr val mem)
  1559  
  1560  (MOVSDstore [off] {sym} ptr (MOVSDconst [f]) mem) && f == f => (MOVQstore [off] {sym} ptr (MOVQconst [int64(math.Float64bits(f))]) mem)
  1561  (MOVSSstore [off] {sym} ptr (MOVSSconst [f]) mem) && f == f => (MOVLstore [off] {sym} ptr (MOVLconst [int32(math.Float32bits(f))]) mem)
  1562  
  1563  // Load args directly into the register class where it will be used.
  1564  // We do this by just modifying the type of the Arg.
  1565  (MOVQf2i <t> (Arg <u> [off] {sym})) && t.Size() == u.Size() => @b.Func.Entry (Arg <t> [off] {sym})
  1566  (MOVLf2i <t> (Arg <u> [off] {sym})) && t.Size() == u.Size() => @b.Func.Entry (Arg <t> [off] {sym})
  1567  (MOVQi2f <t> (Arg <u> [off] {sym})) && t.Size() == u.Size() => @b.Func.Entry (Arg <t> [off] {sym})
  1568  (MOVLi2f <t> (Arg <u> [off] {sym})) && t.Size() == u.Size() => @b.Func.Entry (Arg <t> [off] {sym})
  1569  
  1570  // LEAQ is rematerializeable, so this helps to avoid register spill.
  1571  // See issue 22947 for details
  1572  (ADD(Q|L)const [off] x:(SP)) => (LEA(Q|L) [off] x)
  1573  
  1574  // HMULx is commutative, but its first argument must go in AX.
  1575  // If possible, put a rematerializeable value in the first argument slot,
  1576  // to reduce the odds that another value will be have to spilled
  1577  // specifically to free up AX.
  1578  (HMUL(Q|L)  x y) && !x.Rematerializeable() && y.Rematerializeable() => (HMUL(Q|L)  y x)
  1579  (HMUL(Q|L)U x y) && !x.Rematerializeable() && y.Rematerializeable() => (HMUL(Q|L)U y x)
  1580  
  1581  // Fold loads into compares
  1582  // Note: these may be undone by the flagalloc pass.
  1583  (CMP(Q|L|W|B) l:(MOV(Q|L|W|B)load {sym} [off] ptr mem) x) && ssa.CanMergeLoad(v, l) && ssa.Clobber(l) => (CMP(Q|L|W|B)load {sym} [off] ptr x mem)
  1584  (CMP(Q|L|W|B) x l:(MOV(Q|L|W|B)load {sym} [off] ptr mem)) && ssa.CanMergeLoad(v, l) && ssa.Clobber(l) => (InvertFlags (CMP(Q|L|W|B)load {sym} [off] ptr x mem))
  1585  
  1586  (CMP(Q|L)const l:(MOV(Q|L)load {sym} [off] ptr mem) [c])
  1587  	&& l.Uses == 1
  1588  	&& ssa.Clobber(l) =>
  1589  @l.Block (CMP(Q|L)constload {sym} [ssa.MakeValAndOff(c,off)] ptr mem)
  1590  (CMP(W|B)const l:(MOV(W|B)load {sym} [off] ptr mem) [c])
  1591  	&& l.Uses == 1
  1592  	&& ssa.Clobber(l) =>
  1593  @l.Block (CMP(W|B)constload {sym} [ssa.MakeValAndOff(int32(c),off)] ptr mem)
  1594  
  1595  (CMPQload {sym} [off] ptr (MOVQconst [c]) mem) && validVal(c) => (CMPQconstload {sym} [ssa.MakeValAndOff(int32(c),off)] ptr mem)
  1596  (CMPLload {sym} [off] ptr (MOVLconst [c]) mem) => (CMPLconstload {sym} [ssa.MakeValAndOff(c,off)] ptr mem)
  1597  (CMPWload {sym} [off] ptr (MOVLconst [c]) mem) => (CMPWconstload {sym} [ssa.MakeValAndOff(int32(int16(c)),off)] ptr mem)
  1598  (CMPBload {sym} [off] ptr (MOVLconst [c]) mem) => (CMPBconstload {sym} [ssa.MakeValAndOff(int32(int8(c)),off)] ptr mem)
  1599  
  1600  (TEST(Q|L|W|B)  l:(MOV(Q|L|W|B)load {sym} [off] ptr mem) l2)
  1601          && l == l2
  1602  	&& l.Uses == 2
  1603  	&& ssa.Clobber(l) =>
  1604    @l.Block (CMP(Q|L|W|B)constload {sym} [ssa.MakeValAndOff(0, off)] ptr mem)
  1605  
  1606  // Convert ANDload to MOVload when we can do the AND in a containing TEST op.
  1607  // Only do when it's within the same block, so we don't have flags live across basic block boundaries.
  1608  // See issue 44228.
  1609  (TEST(Q|L) a:(AND(Q|L)load [off] {sym} x ptr mem) a) && a.Uses == 2 && a.Block == v.Block && ssa.Clobber(a) => (TEST(Q|L) (MOV(Q|L)load <a.Type> [off] {sym} ptr mem) x)
  1610  
  1611  (MOVBload [off] {sym} (SB) _) && ssa.SymIsRO(sym) => (MOVLconst [int32(ssa.Read8(sym, int64(off)))])
  1612  (MOVWload [off] {sym} (SB) _) && ssa.SymIsRO(sym) => (MOVLconst [int32(ssa.Read16(sym, int64(off), config.Ctxt.Arch.ByteOrder))])
  1613  (MOVLload <t> [off] {sym} (SB) _) && ssa.SymIsRO(sym) && ssa.Is32BitInt(t) => (MOVLconst [int32(ssa.Read32(sym, int64(off), config.Ctxt.Arch.ByteOrder))])
  1614  (MOVLload <t> [off] {sym} (SB) _) && ssa.SymIsRO(sym) && ssa.Is64BitInt(t) => (MOVQconst [int64(ssa.Read32(sym, int64(off), config.Ctxt.Arch.ByteOrder))])
  1615  (MOVQload [off] {sym} (SB) _) && ssa.SymIsRO(sym) => (MOVQconst [int64(ssa.Read64(sym, int64(off), config.Ctxt.Arch.ByteOrder))])
  1616  (MOVBQSXload [off] {sym} (SB) _) && ssa.SymIsRO(sym) => (MOVQconst [int64(int8(ssa.Read8(sym, int64(off))))])
  1617  (MOVWQSXload [off] {sym} (SB) _) && ssa.SymIsRO(sym) => (MOVQconst [int64(int16(ssa.Read16(sym, int64(off), config.Ctxt.Arch.ByteOrder)))])
  1618  (MOVLQSXload [off] {sym} (SB) _) && ssa.SymIsRO(sym) => (MOVQconst [int64(int32(ssa.Read32(sym, int64(off), config.Ctxt.Arch.ByteOrder)))])
  1619  
  1620  
  1621  (MOVOstore [dstOff] {dstSym} ptr (MOVOload [srcOff] {srcSym} (SB) _) mem) && ssa.SymIsRO(srcSym) =>
  1622    (MOVQstore [dstOff+8] {dstSym} ptr (MOVQconst [int64(ssa.Read64(srcSym, int64(srcOff)+8, config.Ctxt.Arch.ByteOrder))])
  1623      (MOVQstore [dstOff] {dstSym} ptr (MOVQconst [int64(ssa.Read64(srcSym, int64(srcOff), config.Ctxt.Arch.ByteOrder))]) mem))
  1624  
  1625  // Arch-specific inlining for small or disjoint runtime.memmove
  1626  // Match post-lowering calls, memory version.
  1627  (SelectN [0] call:(CALLstatic {sym} s1:(MOVQstoreconst _ [sc] s2:(MOVQstore _ src s3:(MOVQstore _ dst mem)))))
  1628  	&& sc.Val64() >= 0
  1629  	&& ssa.IsSameCall(sym, "runtime.memmove")
  1630  	&& s1.Uses == 1 && s2.Uses == 1 && s3.Uses == 1
  1631  	&& ssa.IsInlinableMemmove(dst, src, sc.Val64(), config)
  1632  	&& ssa.Clobber(s1, s2, s3, call)
  1633  	=> (Move [sc.Val64()] dst src mem)
  1634  
  1635  // Match post-lowering calls, register version.
  1636  (SelectN [0] call:(CALLstatic {sym} dst src (MOVQconst [sz]) mem))
  1637  	&& sz >= 0
  1638  	&& ssa.IsSameCall(sym, "runtime.memmove")
  1639  	&& call.Uses == 1
  1640  	&& ssa.IsInlinableMemmove(dst, src, sz, config)
  1641  	&& ssa.Clobber(call)
  1642  	=> (Move [sz] dst src mem)
  1643  
  1644  // Prefetch instructions
  1645  (PrefetchCache ...)   => (PrefetchT0 ...)
  1646  (PrefetchCacheStreamed ...) => (PrefetchNTA ...)
  1647  
  1648  // CPUID feature: BMI1.
  1649  (AND(Q|L) x (NOT(Q|L) y))               && buildcfg.GOAMD64 >= 3 => (ANDN(Q|L) x y)
  1650  (AND(Q|L) x (NEG(Q|L) x))               && buildcfg.GOAMD64 >= 3 => (BLSI(Q|L) x)
  1651  (XOR(Q|L) x (ADD(Q|L)const [-1] x))     && buildcfg.GOAMD64 >= 3 => (BLSMSK(Q|L) x)
  1652  (AND(Q|L) <t> x (ADD(Q|L)const [-1] x)) && buildcfg.GOAMD64 >= 3 => (Select0 <t> (BLSR(Q|L) x))
  1653  // eliminate TEST instruction in classical "isPowerOfTwo" check
  1654  (SETEQ       (TEST(Q|L) s:(Select0 blsr:(BLSR(Q|L) _)) s))        => (SETEQ       (Select1 <types.TypeFlags> blsr))
  1655  (CMOVQEQ x y (TEST(Q|L) s:(Select0 blsr:(BLSR(Q|L) _)) s))        => (CMOVQEQ x y (Select1 <types.TypeFlags> blsr))
  1656  (CMOVLEQ x y (TEST(Q|L) s:(Select0 blsr:(BLSR(Q|L) _)) s))        => (CMOVLEQ x y (Select1 <types.TypeFlags> blsr))
  1657  (EQ          (TEST(Q|L) s:(Select0 blsr:(BLSR(Q|L) _)) s) yes no) => (EQ          (Select1 <types.TypeFlags> blsr) yes no)
  1658  (SETNE       (TEST(Q|L) s:(Select0 blsr:(BLSR(Q|L) _)) s))        => (SETNE       (Select1 <types.TypeFlags> blsr))
  1659  (CMOVQNE x y (TEST(Q|L) s:(Select0 blsr:(BLSR(Q|L) _)) s))        => (CMOVQNE x y (Select1 <types.TypeFlags> blsr))
  1660  (CMOVLNE x y (TEST(Q|L) s:(Select0 blsr:(BLSR(Q|L) _)) s))        => (CMOVLNE x y (Select1 <types.TypeFlags> blsr))
  1661  (NE          (TEST(Q|L) s:(Select0 blsr:(BLSR(Q|L) _)) s) yes no) => (NE          (Select1 <types.TypeFlags> blsr) yes no)
  1662  
  1663  (BSWAP(Q|L) (BSWAP(Q|L) p)) => p
  1664  
  1665  // CPUID feature: MOVBE.
  1666  (MOV(Q|L)store   [i] {s} p x:(BSWAP(Q|L) w) mem) && x.Uses == 1 && buildcfg.GOAMD64 >= 3 => (MOVBE(Q|L)store [i] {s} p w mem)
  1667  (MOVBE(Q|L)store [i] {s} p x:(BSWAP(Q|L) w) mem) && x.Uses == 1                          => (MOV(Q|L)store   [i] {s} p w mem)
  1668  (BSWAP(Q|L) x:(MOV(Q|L)load   [i] {s} p mem))  && x.Uses == 1 && buildcfg.GOAMD64 >= 3 => @x.Block (MOVBE(Q|L)load [i] {s} p mem)
  1669  (BSWAP(Q|L) x:(MOVBE(Q|L)load [i] {s} p mem))  && x.Uses == 1                          => @x.Block (MOV(Q|L)load   [i] {s} p mem)
  1670  (MOVWstore [i] {s} p x:(ROLWconst [8] w) mem)   && x.Uses == 1 && buildcfg.GOAMD64 >= 3 => (MOVBEWstore [i] {s} p w mem)
  1671  (MOVBEWstore [i] {s} p x:(ROLWconst [8] w) mem) && x.Uses == 1 => (MOVWstore [i] {s} p w mem)
  1672  
  1673  (SAR(Q|L) l:(MOV(Q|L)load [off] {sym} ptr mem) x) && buildcfg.GOAMD64 >= 3 && ssa.CanMergeLoad(v, l) && ssa.Clobber(l) => (SARX(Q|L)load [off] {sym} ptr x mem)
  1674  (SHL(Q|L) l:(MOV(Q|L)load [off] {sym} ptr mem) x) && buildcfg.GOAMD64 >= 3 && ssa.CanMergeLoad(v, l) && ssa.Clobber(l) => (SHLX(Q|L)load [off] {sym} ptr x mem)
  1675  (SHR(Q|L) l:(MOV(Q|L)load [off] {sym} ptr mem) x) && buildcfg.GOAMD64 >= 3 && ssa.CanMergeLoad(v, l) && ssa.Clobber(l) => (SHRX(Q|L)load [off] {sym} ptr x mem)
  1676  
  1677  ((SHL|SHR|SAR)XQload [off] {sym} ptr (MOVQconst [c]) mem) => ((SHL|SHR|SAR)Qconst [int8(c&63)] (MOVQload [off] {sym} ptr mem))
  1678  ((SHL|SHR|SAR)XQload [off] {sym} ptr (MOVLconst [c]) mem) => ((SHL|SHR|SAR)Qconst [int8(c&63)] (MOVQload [off] {sym} ptr mem))
  1679  ((SHL|SHR|SAR)XLload [off] {sym} ptr (MOVLconst [c]) mem) => ((SHL|SHR|SAR)Lconst [int8(c&31)] (MOVLload [off] {sym} ptr mem))
  1680  
  1681  // Convert atomic logical operations to easier ones if we don't use the result.
  1682  (Select1 a:(LoweredAtomic(And64|And32|Or64|Or32) ptr val mem)) && a.Uses == 1 && ssa.Clobber(a) => ((ANDQ|ANDL|ORQ|ORL)lock ptr val mem)
  1683  
  1684  // If we are checking the results of an add, use the flags directly from the add.
  1685  // Note that this only works for EQ/NE. ADD sets the CF/OF flags differently
  1686  // than TEST sets them.
  1687  // Note also that a.Args[0] here refers to the post-flagify'd value.
  1688  ((EQ|NE) t:(TESTQ a:(ADDQconst [c] x) a)) && t.Uses == 1 && flagify(a) => ((EQ|NE) (Select1 <types.TypeFlags> a.Args[0]))
  1689  ((EQ|NE) t:(TESTL a:(ADDLconst [c] x) a)) && t.Uses == 1 && flagify(a) => ((EQ|NE) (Select1 <types.TypeFlags> a.Args[0]))
  1690  
  1691  // If we don't use the flags any more, just use the standard op.
  1692  (Select0 a:(ADD(Q|L)constflags [c] x)) && a.Uses == 1 => (ADD(Q|L)const [c] x)
  1693  
  1694  // SIMD lowering rules
  1695  
  1696  // Mask conversions
  1697  // integers to masks
  1698  (Cvt16toMask8x16 <t> x) => (VPMOVMToVec8x16 <types.TypeVec128> (KMOVWk <t> x))
  1699  (Cvt32toMask8x32 <t> x) => (VPMOVMToVec8x32 <types.TypeVec256> (KMOVDk <t> x))
  1700  (Cvt64toMask8x64 <t> x) => (VPMOVMToVec8x64 <types.TypeVec512> (KMOVQk <t> x))
  1701  
  1702  (Cvt8toMask16x8 <t> x) => (VPMOVMToVec16x8 <types.TypeVec128> (KMOVBk <t> x))
  1703  (Cvt16toMask16x16 <t> x) => (VPMOVMToVec16x16 <types.TypeVec256> (KMOVWk <t> x))
  1704  (Cvt32toMask16x32 <t> x) => (VPMOVMToVec16x32 <types.TypeVec512> (KMOVDk <t> x))
  1705  
  1706  (Cvt8toMask32x4 <t> x) => (VPMOVMToVec32x4 <types.TypeVec128> (KMOVBk <t> x))
  1707  (Cvt8toMask32x8 <t> x) => (VPMOVMToVec32x8 <types.TypeVec256> (KMOVBk <t> x))
  1708  (Cvt16toMask32x16 <t> x) => (VPMOVMToVec32x16 <types.TypeVec512> (KMOVWk <t> x))
  1709  
  1710  (Cvt8toMask64x2 <t> x) => (VPMOVMToVec64x2 <types.TypeVec128> (KMOVBk <t> x))
  1711  (Cvt8toMask64x4 <t> x) => (VPMOVMToVec64x4 <types.TypeVec256> (KMOVBk <t> x))
  1712  (Cvt8toMask64x8 <t> x) => (VPMOVMToVec64x8 <types.TypeVec512> (KMOVBk <t> x))
  1713  
  1714  // masks to integers
  1715  (CvtMask8x16to16 ...) => (VPMOVMSKB128 ...)
  1716  (CvtMask8x32to32 ...) => (VPMOVMSKB256 ...)
  1717  (CvtMask8x64to64 x) => (KMOVQi (VPMOVVec8x64ToM <types.TypeMask> x))
  1718  
  1719  (CvtMask16x8to8 x) => (KMOVBi (VPMOVVec16x8ToM <types.TypeMask> x))
  1720  (CvtMask16x16to16 x) => (KMOVWi (VPMOVVec16x16ToM <types.TypeMask> x))
  1721  (CvtMask16x32to32 x) => (KMOVDi (VPMOVVec16x32ToM <types.TypeMask> x))
  1722  
  1723  (CvtMask32x4to8 ...) => (VMOVMSKPS128 ...)
  1724  (CvtMask32x8to8 ...) => (VMOVMSKPS256 ...)
  1725  (CvtMask32x16to16 x) => (KMOVWi (VPMOVVec32x16ToM <types.TypeMask> x))
  1726  
  1727  (CvtMask64x2to8 ...) => (VMOVMSKPD128 ...)
  1728  (CvtMask64x4to8 ...) => (VMOVMSKPD256 ...)
  1729  (CvtMask64x8to8 x) => (KMOVBi (VPMOVVec64x8ToM <types.TypeMask> x))
  1730  
  1731  // optimizations
  1732  (MOVBstore [off] {sym} ptr (KMOVBi mask) mem) => (KMOVBstore [off] {sym} ptr mask mem)
  1733  (MOVWstore [off] {sym} ptr (KMOVWi mask) mem) => (KMOVWstore [off] {sym} ptr mask mem)
  1734  (MOVLstore [off] {sym} ptr (KMOVDi mask) mem) => (KMOVDstore [off] {sym} ptr mask mem)
  1735  (MOVQstore [off] {sym} ptr (KMOVQi mask) mem) => (KMOVQstore [off] {sym} ptr mask mem)
  1736  
  1737  (KMOVBk l:(MOVBload [off] {sym} ptr mem)) && ssa.CanMergeLoad(v, l) && ssa.Clobber(l) => (KMOVBload [off] {sym} ptr mem)
  1738  (KMOVWk l:(MOVWload [off] {sym} ptr mem)) && ssa.CanMergeLoad(v, l) && ssa.Clobber(l) => (KMOVWload [off] {sym} ptr mem)
  1739  (KMOVDk l:(MOVLload [off] {sym} ptr mem)) && ssa.CanMergeLoad(v, l) && ssa.Clobber(l) => (KMOVDload [off] {sym} ptr mem)
  1740  (KMOVQk l:(MOVQload [off] {sym} ptr mem)) && ssa.CanMergeLoad(v, l) && ssa.Clobber(l) => (KMOVQload [off] {sym} ptr mem)
  1741  
  1742  // SIMD vector loads and stores
  1743  (Load <t> ptr mem) && t.Size() == 16 => (VMOVDQUload128 ptr mem)
  1744  (Store {t} ptr val mem) && t.Size() == 16 => (VMOVDQUstore128 ptr val mem)
  1745  
  1746  (Load <t> ptr mem) && t.Size() == 32 => (VMOVDQUload256 ptr mem)
  1747  (Store {t} ptr val mem) && t.Size() == 32 => (VMOVDQUstore256 ptr val mem)
  1748  
  1749  (Load <t> ptr mem) && t.Size() == 64 => (VMOVDQUload512 ptr mem)
  1750  (Store {t} ptr val mem) && t.Size() == 64 => (VMOVDQUstore512 ptr val mem)
  1751  
  1752  // SIMD vector integer-vector-masked loads and stores.
  1753  (LoadMasked32 <t> ptr mask mem) && t.Size() == 16 => (VPMASK32load128 ptr mask mem)
  1754  (LoadMasked32 <t> ptr mask mem) && t.Size() == 32 => (VPMASK32load256 ptr mask mem)
  1755  (LoadMasked64 <t> ptr mask mem) && t.Size() == 16 => (VPMASK64load128 ptr mask mem)
  1756  (LoadMasked64 <t> ptr mask mem) && t.Size() == 32 => (VPMASK64load256 ptr mask mem)
  1757  
  1758  (StoreMasked32 {t} ptr mask val mem) && t.Size() == 16 => (VPMASK32store128 ptr mask val mem)
  1759  (StoreMasked32 {t} ptr mask val mem) && t.Size() == 32 => (VPMASK32store256 ptr mask val mem)
  1760  (StoreMasked64 {t} ptr mask val mem) && t.Size() == 16 => (VPMASK64store128 ptr mask val mem)
  1761  (StoreMasked64 {t} ptr mask val mem) && t.Size() == 32 => (VPMASK64store256 ptr mask val mem)
  1762  
  1763  // Misc
  1764  (IsZeroVec x) => (SETEQ (VPTEST x x))
  1765  
  1766  (IsNaNFloat32x4  x) => (VCMPPS128 [3] x x)
  1767  (IsNaNFloat32x8  x) => (VCMPPS256 [3] x x)
  1768  (IsNaNFloat32x16 x) => (VPMOVMToVec32x16 (VCMPPS512 [3] x x))
  1769  (IsNaNFloat64x2  x) => (VCMPPD128 [3] x x)
  1770  (IsNaNFloat64x4  x) => (VCMPPD256 [3] x x)
  1771  (IsNaNFloat64x8  x) => (VPMOVMToVec64x8 (VCMPPD512 [3] x x))
  1772  
  1773  // SIMD vector K-masked loads and stores
  1774  
  1775  (LoadMasked64 <t> ptr mask mem) && t.Size() == 64 => (VPMASK64load512 ptr (VPMOVVec64x8ToM  <types.TypeMask> mask) mem)
  1776  (LoadMasked32 <t> ptr mask mem) && t.Size() == 64 => (VPMASK32load512 ptr (VPMOVVec32x16ToM <types.TypeMask> mask) mem)
  1777  (LoadMasked16 <t> ptr mask mem) && t.Size() == 64 => (VPMASK16load512 ptr (VPMOVVec16x32ToM <types.TypeMask> mask) mem)
  1778  (LoadMasked8  <t> ptr mask mem) && t.Size() == 64 => (VPMASK8load512  ptr (VPMOVVec8x64ToM  <types.TypeMask> mask) mem)
  1779  
  1780  (StoreMasked64 {t} ptr mask val mem) && t.Size() == 64 => (VPMASK64store512 ptr (VPMOVVec64x8ToM  <types.TypeMask> mask) val mem)
  1781  (StoreMasked32 {t} ptr mask val mem) && t.Size() == 64 => (VPMASK32store512 ptr (VPMOVVec32x16ToM <types.TypeMask> mask) val mem)
  1782  (StoreMasked16 {t} ptr mask val mem) && t.Size() == 64 => (VPMASK16store512 ptr (VPMOVVec16x32ToM <types.TypeMask> mask) val mem)
  1783  (StoreMasked8  {t} ptr mask val mem) && t.Size() == 64 => (VPMASK8store512  ptr (VPMOVVec8x64ToM  <types.TypeMask> mask) val mem)
  1784  
  1785  (ZeroSIMD <t>) && t.Size() == 16 => (Zero128 <t>)
  1786  (ZeroSIMD <t>) && t.Size() == 32 => (Zero256 <t>)
  1787  (ZeroSIMD <t>) && t.Size() == 64 => (Zero512 <t>)
  1788  
  1789  // optimize x.IsNaN().Or(y.IsNaN())
  1790  // do these before the mask rewrites
  1791  (VPOR128 (VCMPP(S|D)128 [3] x x) (VCMPP(S|D)128 [3] y y)) => (VCMPP(S|D)128 [3] x y)
  1792  (VPOR256 (VCMPP(S|D)256 [3] x x) (VCMPP(S|D)256 [3] y y)) => (VCMPP(S|D)256 [3] x y)
  1793  (VPORD512 (VPMOVMToVec32x16 (VCMPPS512 [3] x x)) (VPMOVMToVec32x16 (VCMPPS512 [3] y y))) =>
  1794  	(VPMOVMToVec32x16 (VCMPPS512 [3] x y))
  1795  (VPORD512 (VPMOVMToVec64x8  (VCMPPD512 [3] x x)) (VPMOVMToVec64x8  (VCMPPD512 [3] y y))) =>
  1796  	(VPMOVMToVec64x8  (VCMPPD512 [3] x y))
  1797  
  1798  // Include these rules because you never know about rewrite order
  1799  (KANDB (VCMPPD512 [3] x x) (VCMPPD512 [3] y y)) => (VCMPPD512 [3] x x) // 512 = 64x8 -> KANDB
  1800  (KANDW (VCMPPS512 [3] x x) (VCMPPS512 [3] y y)) => (VCMPPS512 [3] x y) // 512 = 32x16 -> KANDW
  1801  
  1802  // These larger simplifying rules must come before the smaller simplifying rules that might break them).
  1803  // Rewrite rules for binary logical mask operations that apply to 8-bit elements (B, for bytes) of 128, 256, and 512-bit vectors
  1804  (VPAND128 (VPMOVMToVec8x16 x) (VPMOVMToVec8x16 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec8x16 (KANDW x y))
  1805  (VPAND256 (VPMOVMToVec8x32 x) (VPMOVMToVec8x32 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec8x32 (KANDD x y))
  1806  (VPANDD512 (VPMOVMToVec8x64 x) (VPMOVMToVec8x64 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec8x64 (KANDQ x y))
  1807  
  1808  (VPOR128 (VPMOVMToVec8x16 x) (VPMOVMToVec8x16 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec8x16 (KORW x y))
  1809  (VPOR256 (VPMOVMToVec8x32 x) (VPMOVMToVec8x32 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec8x32 (KORD x y))
  1810  (VPORD512 (VPMOVMToVec8x64 x) (VPMOVMToVec8x64 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec8x64 (KORQ x y))
  1811  
  1812  (VPXOR128 (VPMOVMToVec8x16 x) (VPMOVMToVec8x16 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec8x16 (KXORW x y))
  1813  (VPXOR256 (VPMOVMToVec8x32 x) (VPMOVMToVec8x32 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec8x32 (KXORD x y))
  1814  (VPXORD512 (VPMOVMToVec8x64 x) (VPMOVMToVec8x64 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec8x64 (KXORQ x y))
  1815  
  1816  // Rewrite rules for binary logical mask operations that apply to 16-bit elements (W, for words) of 128, 256, and 512-bit vectors
  1817  (VPAND128 (VPMOVMToVec16x8 x) (VPMOVMToVec16x8 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec16x8 (KANDB x y))
  1818  (VPAND256 (VPMOVMToVec16x16 x) (VPMOVMToVec16x16 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec16x16 (KANDW x y))
  1819  (VPANDD512 (VPMOVMToVec16x32 x) (VPMOVMToVec16x32 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec16x32 (KANDD x y))
  1820  
  1821  (VPOR128 (VPMOVMToVec16x8 x) (VPMOVMToVec16x8 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec16x8 (KORB x y))
  1822  (VPOR256 (VPMOVMToVec16x16 x) (VPMOVMToVec16x16 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec16x16 (KORW x y))
  1823  (VPORD512 (VPMOVMToVec16x32 x) (VPMOVMToVec16x32 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec16x32 (KORD x y))
  1824  
  1825  (VPXOR128 (VPMOVMToVec16x8 x) (VPMOVMToVec16x8 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec16x8 (KXORB x y))
  1826  (VPXOR256 (VPMOVMToVec16x16 x) (VPMOVMToVec16x16 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec16x16 (KXORW x y))
  1827  (VPXORD512 (VPMOVMToVec16x32 x) (VPMOVMToVec16x32 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec16x32 (KXORD x y))
  1828  
  1829  // Rewrite rules for binary logical mask operations that apply to 32-bit elements (D, for doublewords) of 128, 256, and 512-bit vectors
  1830  (VPAND128 (VPMOVMToVec32x4 x) (VPMOVMToVec32x4 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec32x4 (KANDB x y))
  1831  (VPAND256 (VPMOVMToVec32x8 x) (VPMOVMToVec32x8 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec32x8 (KANDB x y))
  1832  (VPANDD512 (VPMOVMToVec32x16 x) (VPMOVMToVec32x16 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec32x16 (KANDW x y))
  1833  
  1834  (VPOR128 (VPMOVMToVec32x4 x) (VPMOVMToVec32x4 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec32x4 (KORB x y))
  1835  (VPOR256 (VPMOVMToVec32x8 x) (VPMOVMToVec32x8 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec32x8 (KORB x y))
  1836  (VPORD512 (VPMOVMToVec32x16 x) (VPMOVMToVec32x16 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec32x16 (KORW x y))
  1837  
  1838  (VPXOR128 (VPMOVMToVec32x4 x) (VPMOVMToVec32x4 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec32x4 (KXORB x y))
  1839  (VPXOR256 (VPMOVMToVec32x8 x) (VPMOVMToVec32x8 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec32x8 (KXORB x y))
  1840  (VPXORD512 (VPMOVMToVec32x16 x) (VPMOVMToVec32x16 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec32x16 (KXORW x y))
  1841  
  1842  // Rewrite rules for binary logical mask operations that apply to 64-bit elements (Q, for quadwords) of 128, 256, and 512-bit vectors
  1843  (VPAND128 (VPMOVMToVec64x2 x) (VPMOVMToVec64x2 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec64x2 (KANDB x y))
  1844  (VPAND256 (VPMOVMToVec64x4 x) (VPMOVMToVec64x4 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec64x4 (KANDB x y))
  1845  (VPANDD512 (VPMOVMToVec64x8 x) (VPMOVMToVec64x8 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec64x8 (KANDB x y))
  1846  
  1847  (VPOR128 (VPMOVMToVec64x2 x) (VPMOVMToVec64x2 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec64x2 (KORB x y))
  1848  (VPOR256 (VPMOVMToVec64x4 x) (VPMOVMToVec64x4 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec64x4 (KORB x y))
  1849  (VPORD512 (VPMOVMToVec64x8 x) (VPMOVMToVec64x8 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec64x8 (KORB x y))
  1850  
  1851  (VPXOR128 (VPMOVMToVec64x2 x) (VPMOVMToVec64x2 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec64x2 (KXORB x y))
  1852  (VPXOR256 (VPMOVMToVec64x4 x) (VPMOVMToVec64x4 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec64x4 (KXORB x y))
  1853  (VPXORD512 (VPMOVMToVec64x8 x) (VPMOVMToVec64x8 y)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VPMOVMToVec64x8 (KXORB x y))
  1854  
  1855  (VPMOVVec8x16ToM (VPMOVMToVec8x16 x)) => x
  1856  (VPMOVVec8x32ToM (VPMOVMToVec8x32 x)) => x
  1857  (VPMOVVec8x64ToM (VPMOVMToVec8x64 x)) => x
  1858  
  1859  (VPMOVVec16x8ToM (VPMOVMToVec16x8 x)) => x
  1860  (VPMOVVec16x16ToM (VPMOVMToVec16x16 x)) => x
  1861  (VPMOVVec16x32ToM (VPMOVMToVec16x32 x)) => x
  1862  
  1863  (VPMOVVec32x4ToM (VPMOVMToVec32x4 x)) => x
  1864  (VPMOVVec32x8ToM (VPMOVMToVec32x8 x)) => x
  1865  (VPMOVVec32x16ToM (VPMOVMToVec32x16 x)) => x
  1866  
  1867  (VPMOVVec64x2ToM (VPMOVMToVec64x2 x)) => x
  1868  (VPMOVVec64x4ToM (VPMOVMToVec64x4 x)) => x
  1869  (VPMOVVec64x8ToM (VPMOVMToVec64x8 x)) => x
  1870  
  1871  (VPANDQ512 x (VPMOVMToVec64x8 k)) => (VMOVDQU64Masked512 x k)
  1872  (VPANDQ512 x (VPMOVMToVec32x16 k)) => (VMOVDQU32Masked512 x k)
  1873  (VPANDQ512 x (VPMOVMToVec16x32 k)) => (VMOVDQU16Masked512 x k)
  1874  (VPANDQ512 x (VPMOVMToVec8x64 k)) => (VMOVDQU8Masked512 x k)
  1875  (VPANDD512 x (VPMOVMToVec64x8 k)) => (VMOVDQU64Masked512 x k)
  1876  (VPANDD512 x (VPMOVMToVec32x16 k)) => (VMOVDQU32Masked512 x k)
  1877  (VPANDD512 x (VPMOVMToVec16x32 k)) => (VMOVDQU16Masked512 x k)
  1878  (VPANDD512 x (VPMOVMToVec8x64 k)) => (VMOVDQU8Masked512 x k)
  1879  
  1880  (VPAND128 x (VPMOVMToVec8x16 k)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VMOVDQU8Masked128 x k)
  1881  (VPAND128 x (VPMOVMToVec16x8 k)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VMOVDQU16Masked128 x k)
  1882  (VPAND128 x (VPMOVMToVec32x4 k)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VMOVDQU32Masked128 x k)
  1883  (VPAND128 x (VPMOVMToVec64x2 k)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VMOVDQU64Masked128 x k)
  1884  
  1885  (VPAND256 x (VPMOVMToVec8x32 k)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VMOVDQU8Masked256 x k)
  1886  (VPAND256 x (VPMOVMToVec16x16 k)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VMOVDQU16Masked256 x k)
  1887  (VPAND256 x (VPMOVMToVec32x8 k)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VMOVDQU32Masked256 x k)
  1888  (VPAND256 x (VPMOVMToVec64x4 k)) && v.Block.CPUfeatures.HasFeature(ssa.CPUavx512) => (VMOVDQU64Masked256 x k)
  1889  
  1890  // Insert to zero of 32/64 bit floats and ints to a zero is just MOVS[SD]
  1891  (VPINSRQ128 [0] (Zero128 <t>) y) && y.Type.IsFloat() => (VMOVSDf2v <types.TypeVec128> y)
  1892  (VPINSRD128 [0] (Zero128 <t>) y) && y.Type.IsFloat() => (VMOVSSf2v <types.TypeVec128> y)
  1893  (VPINSRQ128 [0] (Zero128 <t>) y) && !y.Type.IsFloat() => (VMOVQ <types.TypeVec128> y)
  1894  (VPINSRD128 [0] (Zero128 <t>) y) && !y.Type.IsFloat() => (VMOVD <types.TypeVec128> y)
  1895  
  1896  // These rewrites can skip zero-extending the 8/16-bit inputs because they are
  1897  // only used as the input to a broadcast; the potentially "bad" bits are ignored
  1898  (VPBROADCASTB(128|256|512) x:(VPINSRB128 [0] (Zero128    <t>) y)) && x.Uses == 1 =>
  1899  	(VPBROADCASTB(128|256|512) (VMOVQ <types.TypeVec128> y))
  1900  (VPBROADCASTW(128|256|512) x:(VPINSRW128 [0] (Zero128    <t>) y)) && x.Uses == 1 =>
  1901  	(VPBROADCASTW(128|256|512)   (VMOVQ <types.TypeVec128> y))
  1902  
  1903  (VMOVQ x:(MOVQload [off] {sym} ptr mem)) && x.Uses == 1 && ssa.Clobber(x) => @x.Block (VMOVQload <v.Type> [off] {sym} ptr mem)
  1904  (VMOVD x:(MOVLload [off] {sym} ptr mem)) && x.Uses == 1 && ssa.Clobber(x) => @x.Block (VMOVDload <v.Type> [off] {sym} ptr mem)
  1905  
  1906  (VMOVSDf2v x:(MOVSDload [off] {sym} ptr mem)) && x.Uses == 1 && ssa.Clobber(x) => @x.Block (VMOVSDload <v.Type> [off] {sym} ptr mem)
  1907  (VMOVSSf2v x:(MOVSSload [off] {sym} ptr mem)) && x.Uses == 1 && ssa.Clobber(x) => @x.Block (VMOVSSload <v.Type> [off] {sym} ptr mem)
  1908  
  1909  (VMOVSDf2v x:(MOVSDconst [c] )) => (VMOVSDconst [c] )
  1910  (VMOVSSf2v x:(MOVSSconst [c] )) => (VMOVSSconst [c] )
  1911  
  1912  (VMOVDQUload(128|256|512) [off1] {sym} x:(ADDQconst [off2] ptr) mem) && ssa.Is32Bit(int64(off1)+int64(off2)) => (VMOVDQUload(128|256|512) [off1+off2] {sym} ptr mem)
  1913  (VMOVDQUstore(128|256|512) [off1] {sym} x:(ADDQconst [off2] ptr) val mem) && ssa.Is32Bit(int64(off1)+int64(off2)) => (VMOVDQUstore(128|256|512) [off1+off2] {sym} ptr val mem)
  1914  (VMOVDQUload(128|256|512) [off1] {sym1} x:(LEAQ [off2] {sym2} base) mem) && ssa.Is32Bit(int64(off1)+int64(off2)) && ssa.CanMergeSym(sym1, sym2) => (VMOVDQUload(128|256|512) [off1+off2] {ssa.MergeSym(sym1, sym2)} base mem)
  1915  (VMOVDQUstore(128|256|512) [off1] {sym1} x:(LEAQ [off2] {sym2} base) val mem) && ssa.Is32Bit(int64(off1)+int64(off2)) && ssa.CanMergeSym(sym1, sym2) => (VMOVDQUstore(128|256|512) [off1+off2] {ssa.MergeSym(sym1, sym2)} base val mem)
  1916  
  1917  // Multiplication of numbers that are the power of 2 can be converted to left shift log(x)
  1918  (VPMULLQ128 x (VPBROADCASTQ128 (VMOVQ (MOVQconst [c])))) && ssa.IsPowerOfTwo(c) => (VPSLLQ128const [uint8(ssa.Log64(c))] x)
  1919  (VPMULLQ256 x (VPBROADCASTQ256 (VMOVQ (MOVQconst [c])))) && ssa.IsPowerOfTwo(c) => (VPSLLQ256const [uint8(ssa.Log64(c))] x)
  1920  (VPMULLQ512 x (VPBROADCASTQ512 (VMOVQ (MOVQconst [c])))) && ssa.IsPowerOfTwo(c) => (VPSLLQ512const [uint8(ssa.Log64(c))] x)
  1921  
  1922  (VPMULLD128 x (VPBROADCASTD128 (VMOVD (MOVLconst [c])))) && ssa.IsPowerOfTwo(c) => (VPSLLD128const [uint8(ssa.Log32(int32(c)))] x)
  1923  (VPMULLD256 x (VPBROADCASTD256 (VMOVD (MOVLconst [c])))) && ssa.IsPowerOfTwo(c) => (VPSLLD256const [uint8(ssa.Log32(int32(c)))] x)
  1924  (VPMULLD512 x (VPBROADCASTD512 (VMOVD (MOVLconst [c])))) && ssa.IsPowerOfTwo(c) => (VPSLLD512const [uint8(ssa.Log32(int32(c)))] x)
  1925  
  1926  (VPMULLW128 x (VPBROADCASTW128 (VMOVQ (MOVLconst [c])))) && ssa.IsPowerOfTwo(c) => (VPSLLW128const [uint8(ssa.Log16(int16(c)))] x) // use VMOVQ because VMOVW doesn't exist
  1927  (VPMULLW256 x (VPBROADCASTW256 (VMOVQ (MOVLconst [c])))) && ssa.IsPowerOfTwo(c) => (VPSLLW256const [uint8(ssa.Log16(int16(c)))] x)
  1928  (VPMULLW512 x (VPBROADCASTW512 (VMOVQ (MOVLconst [c])))) && ssa.IsPowerOfTwo(c) => (VPSLLW512const [uint8(ssa.Log16(int16(c)))] x)
  1929  
  1930  // 2-op VPTEST optimizations
  1931  (SETEQ (VPTEST x:(VPAND(128|256) j k) y)) && x == y && x.Uses == 2 => (SETEQ (VPTEST j k))
  1932  (SETEQ (VPTEST x:(VPAND(D|Q)512 j k) y)) && x == y && x.Uses == 2 => (SETEQ (VPTEST j k))
  1933  (SETEQ (VPTEST x:(VPANDN(128|256) j k) y)) && x == y && x.Uses == 2 => (SETB (VPTEST k j)) // AndNot has swapped its operand order
  1934  (SETEQ (VPTEST x:(VPANDN(D|Q)512 j k) y)) && x == y && x.Uses == 2 => (SETB (VPTEST k j)) // AndNot has swapped its operand order
  1935  (EQ (VPTEST x:(VPAND(128|256) j k) y) yes no) && x == y && x.Uses == 2 => (EQ (VPTEST j k) yes no)
  1936  (EQ (VPTEST x:(VPAND(D|Q)512 j k) y) yes no) && x == y && x.Uses == 2 => (EQ (VPTEST j k) yes no)
  1937  (EQ (VPTEST x:(VPANDN(128|256) j k) y) yes no) && x == y && x.Uses == 2 => (ULT (VPTEST k j) yes no) // AndNot has swapped its operand order
  1938  (EQ (VPTEST x:(VPANDN(D|Q)512 j k) y) yes no) && x == y && x.Uses == 2 => (ULT (VPTEST k j) yes no) // AndNot has swapped its operand order
  1939  
  1940  // remove flags → bool → flags roundtrip
  1941  // Only do it if the flag generating instruction is local otherwise the likelihood flagalloc won't undo this optimization and makes things worse are slim.
  1942  (NE t:(TESTQ x:(MOVBQZX s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE|EQF|NEF|GF|GEF) flags)) x) yes no) && t.Block == s.Block => ((EQ|NE|LT|GT|LE|GE|UGT|ULT|UGE|ULE|EQF|NEF|UGT|UGE) flags yes no)
  1943  (NE t:(TESTL x:(MOVBQZX s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE|EQF|NEF|GF|GEF) flags)) x) yes no) && t.Block == s.Block => ((EQ|NE|LT|GT|LE|GE|UGT|ULT|UGE|ULE|EQF|NEF|UGT|UGE) flags yes no)
  1944  (NE t:(TESTW x:(MOVBQZX s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE|EQF|NEF|GF|GEF) flags)) x) yes no) && t.Block == s.Block => ((EQ|NE|LT|GT|LE|GE|UGT|ULT|UGE|ULE|EQF|NEF|UGT|UGE) flags yes no)
  1945  (NE t:(TESTB            s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE|EQF|NEF|GF|GEF) flags)  s) yes no) && t.Block == s.Block => ((EQ|NE|LT|GT|LE|GE|UGT|ULT|UGE|ULE|EQF|NEF|UGT|UGE) flags yes no)
  1946  
  1947  (CMOVQNE yes no t:(TESTQ x:(MOVBQZX s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE|EQF|NEF|GF|GEF) flags)) x)) && t.Block == s.Block => (CMOVQ(EQ|NE|LT|GT|LE|GE|HI|CS|CC|LS|EQF|NEF|GTF|GEF) yes no flags)
  1948  (CMOVQNE yes no t:(TESTL x:(MOVBQZX s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE|EQF|NEF|GF|GEF) flags)) x)) && t.Block == s.Block => (CMOVQ(EQ|NE|LT|GT|LE|GE|HI|CS|CC|LS|EQF|NEF|GTF|GEF) yes no flags)
  1949  (CMOVQNE yes no t:(TESTW x:(MOVBQZX s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE|EQF|NEF|GF|GEF) flags)) x)) && t.Block == s.Block => (CMOVQ(EQ|NE|LT|GT|LE|GE|HI|CS|CC|LS|EQF|NEF|GTF|GEF) yes no flags)
  1950  (CMOVQNE yes no t:(TESTB            s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE|EQF|NEF|GF|GEF) flags)  s)) && t.Block == s.Block => (CMOVQ(EQ|NE|LT|GT|LE|GE|HI|CS|CC|LS|EQF|NEF|GTF|GEF) yes no flags)
  1951  
  1952  (CMOVLNE yes no t:(TESTQ x:(MOVBQZX s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE|EQF|NEF|GF|GEF) flags)) x)) && t.Block == s.Block => (CMOVL(EQ|NE|LT|GT|LE|GE|HI|CS|CC|LS|EQF|NEF|GTF|GEF) yes no flags)
  1953  (CMOVLNE yes no t:(TESTL x:(MOVBQZX s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE|EQF|NEF|GF|GEF) flags)) x)) && t.Block == s.Block => (CMOVL(EQ|NE|LT|GT|LE|GE|HI|CS|CC|LS|EQF|NEF|GTF|GEF) yes no flags)
  1954  (CMOVLNE yes no t:(TESTW x:(MOVBQZX s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE|EQF|NEF|GF|GEF) flags)) x)) && t.Block == s.Block => (CMOVL(EQ|NE|LT|GT|LE|GE|HI|CS|CC|LS|EQF|NEF|GTF|GEF) yes no flags)
  1955  (CMOVLNE yes no t:(TESTB            s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE|EQF|NEF|GF|GEF) flags)  s)) && t.Block == s.Block => (CMOVL(EQ|NE|LT|GT|LE|GE|HI|CS|CC|LS|EQF|NEF|GTF|GEF) yes no flags)
  1956  
  1957  (CMOVWNE yes no t:(TESTQ x:(MOVBQZX s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE|EQF|NEF|GF|GEF) flags)) x)) && t.Block == s.Block => (CMOVW(EQ|NE|LT|GT|LE|GE|HI|CS|CC|LS|EQF|NEF|GTF|GEF) yes no flags)
  1958  (CMOVWNE yes no t:(TESTL x:(MOVBQZX s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE|EQF|NEF|GF|GEF) flags)) x)) && t.Block == s.Block => (CMOVW(EQ|NE|LT|GT|LE|GE|HI|CS|CC|LS|EQF|NEF|GTF|GEF) yes no flags)
  1959  (CMOVWNE yes no t:(TESTW x:(MOVBQZX s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE|EQF|NEF|GF|GEF) flags)) x)) && t.Block == s.Block => (CMOVW(EQ|NE|LT|GT|LE|GE|HI|CS|CC|LS|EQF|NEF|GTF|GEF) yes no flags)
  1960  (CMOVWNE yes no t:(TESTB            s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE|EQF|NEF|GF|GEF) flags)  s)) && t.Block == s.Block => (CMOVW(EQ|NE|LT|GT|LE|GE|HI|CS|CC|LS|EQF|NEF|GTF|GEF) yes no flags)
  1961  
  1962  (SETNE t:(TESTQ x:(MOVBQZX s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE|EQF|NEF|GF|GEF) flags)) x)) && t.Block == s.Block => s
  1963  (SETNE t:(TESTL x:(MOVBQZX s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE|EQF|NEF|GF|GEF) flags)) x)) && t.Block == s.Block => s
  1964  (SETNE t:(TESTW x:(MOVBQZX s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE|EQF|NEF|GF|GEF) flags)) x)) && t.Block == s.Block => s
  1965  (SETNE t:(TESTB            s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE|EQF|NEF|GF|GEF) flags)  s)) && t.Block == s.Block => s
  1966  
  1967  (EQ t:(TESTQ x:(MOVBQZX s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE) flags)) x) yes no) && t.Block == s.Block => ((NE|EQ|GE|LE|GT|LT|ULE|UGE|ULT|UGT) flags yes no)
  1968  (EQ t:(TESTL x:(MOVBQZX s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE) flags)) x) yes no) && t.Block == s.Block => ((NE|EQ|GE|LE|GT|LT|ULE|UGE|ULT|UGT) flags yes no)
  1969  (EQ t:(TESTW x:(MOVBQZX s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE) flags)) x) yes no) && t.Block == s.Block => ((NE|EQ|GE|LE|GT|LT|ULE|UGE|ULT|UGT) flags yes no)
  1970  (EQ t:(TESTB            s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE) flags)  s) yes no) && t.Block == s.Block => ((NE|EQ|GE|LE|GT|LT|ULE|UGE|ULT|UGT) flags yes no)
  1971  
  1972  (CMOVQEQ yes no t:(TESTQ x:(MOVBQZX s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE) flags)) x)) && t.Block == s.Block => (CMOVQ(NE|EQ|GE|LE|GT|LT|LS|CC|CS|HI) yes no flags)
  1973  (CMOVQEQ yes no t:(TESTL x:(MOVBQZX s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE) flags)) x)) && t.Block == s.Block => (CMOVQ(NE|EQ|GE|LE|GT|LT|LS|CC|CS|HI) yes no flags)
  1974  (CMOVQEQ yes no t:(TESTW x:(MOVBQZX s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE) flags)) x)) && t.Block == s.Block => (CMOVQ(NE|EQ|GE|LE|GT|LT|LS|CC|CS|HI) yes no flags)
  1975  (CMOVQEQ yes no t:(TESTB            s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE) flags)  s)) && t.Block == s.Block => (CMOVQ(NE|EQ|GE|LE|GT|LT|LS|CC|CS|HI) yes no flags)
  1976  
  1977  (CMOVLEQ yes no t:(TESTQ x:(MOVBQZX s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE) flags)) x)) && t.Block == s.Block => (CMOVL(NE|EQ|GE|LE|GT|LT|LS|CC|CS|HI) yes no flags)
  1978  (CMOVLEQ yes no t:(TESTL x:(MOVBQZX s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE) flags)) x)) && t.Block == s.Block => (CMOVL(NE|EQ|GE|LE|GT|LT|LS|CC|CS|HI) yes no flags)
  1979  (CMOVLEQ yes no t:(TESTW x:(MOVBQZX s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE) flags)) x)) && t.Block == s.Block => (CMOVL(NE|EQ|GE|LE|GT|LT|LS|CC|CS|HI) yes no flags)
  1980  (CMOVLEQ yes no t:(TESTB            s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE) flags)  s)) && t.Block == s.Block => (CMOVL(NE|EQ|GE|LE|GT|LT|LS|CC|CS|HI) yes no flags)
  1981  
  1982  (CMOVWEQ yes no t:(TESTQ x:(MOVBQZX s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE) flags)) x)) && t.Block == s.Block => (CMOVW(NE|EQ|GE|LE|GT|LT|LS|CC|CS|HI) yes no flags)
  1983  (CMOVWEQ yes no t:(TESTL x:(MOVBQZX s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE) flags)) x)) && t.Block == s.Block => (CMOVW(NE|EQ|GE|LE|GT|LT|LS|CC|CS|HI) yes no flags)
  1984  (CMOVWEQ yes no t:(TESTW x:(MOVBQZX s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE) flags)) x)) && t.Block == s.Block => (CMOVW(NE|EQ|GE|LE|GT|LT|LS|CC|CS|HI) yes no flags)
  1985  (CMOVWEQ yes no t:(TESTB            s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE) flags)  s)) && t.Block == s.Block => (CMOVW(NE|EQ|GE|LE|GT|LT|LS|CC|CS|HI) yes no flags)
  1986  
  1987  (SETEQ t:(TESTQ x:(MOVBQZX s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE) flags)) x)) && t.Block == s.Block => (SET(NE|EQ|GE|LE|G|L|BE|AE|B|A) flags)
  1988  (SETEQ t:(TESTL x:(MOVBQZX s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE) flags)) x)) && t.Block == s.Block => (SET(NE|EQ|GE|LE|G|L|BE|AE|B|A) flags)
  1989  (SETEQ t:(TESTW x:(MOVBQZX s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE) flags)) x)) && t.Block == s.Block => (SET(NE|EQ|GE|LE|G|L|BE|AE|B|A) flags)
  1990  (SETEQ t:(TESTB            s:(SET(EQ|NE|L|G|LE|GE|A|B|AE|BE) flags)  s)) && t.Block == s.Block => (SET(NE|EQ|GE|LE|G|L|BE|AE|B|A) flags)
  1991  
  1992  // Reinterpret int|float
  1993  (I32AsF32 ...) => (MOVLi2f ...)
  1994  (F32AsI32 ...) => (MOVLf2i ...)
  1995  (I64AsF64 ...) => (MOVQi2f ...)
  1996  (F64AsI64 ...) => (MOVQf2i ...)
  1997  

View as plain text