...

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

Documentation: cmd/compile/internal/ssa/_gen

     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)) =>
    31	(Select0 <typ.UInt64> (ADCQ x y (Select1 <types.TypeFlags> (NEGLflags c))))
    32(Select1 (Add64carry x y c)) =>
    33	(NEGQ <typ.UInt64> (SBBQcarrymask <typ.UInt64> (Select1 <types.TypeFlags> (ADCQ x y (Select1 <types.TypeFlags> (NEGLflags c))))))
    34(Select0 (Sub64borrow x y c)) =>
    35	(Select0 <typ.UInt64> (SBBQ x y (Select1 <types.TypeFlags> (NEGLflags c))))
    36(Select1 (Sub64borrow x y c)) =>
    37	(NEGQ <typ.UInt64> (SBBQcarrymask <typ.UInt64> (Select1 <types.TypeFlags> (SBBQ x y (Select1 <types.TypeFlags> (NEGLflags c))))))
    38
    39// Optimize ADCQ and friends
    40(ADCQ x (MOVQconst [c]) carry) && is32Bit(c) => (ADCQconst x [int32(c)] carry)
    41(ADCQ x y (FlagEQ)) => (ADDQcarry x y)
    42(ADCQconst x [c] (FlagEQ)) => (ADDQconstcarry x [c])
    43(ADDQcarry x (MOVQconst [c])) && is32Bit(c) => (ADDQconstcarry x [int32(c)])
    44(SBBQ x (MOVQconst [c]) borrow) && is32Bit(c) => (SBBQconst x [int32(c)] borrow)
    45(SBBQ x y (FlagEQ)) => (SUBQborrow x y)
    46(SBBQconst x [c] (FlagEQ)) => (SUBQconstborrow x [c])
    47(SUBQborrow x (MOVQconst [c])) && is32Bit(c) => (SUBQconstborrow x [int32(c)])
    48(Select1 (NEGLflags (MOVQconst [0]))) => (FlagEQ)
    49(Select1 (NEGLflags (NEGQ (SBBQcarrymask x)))) => x
    50
    51
    52(Mul64uhilo ...) => (MULQU2 ...)
    53(Div128u ...) => (DIVQU2 ...)
    54
    55(Avg64u ...) => (AVGQU ...)
    56
    57(Mod(64|32|16) [a] x y) => (Select1 (DIV(Q|L|W) [a] x y))
    58(Mod8  x y) => (Select1 (DIVW  (SignExt8to16 x) (SignExt8to16 y)))
    59(Mod(64|32|16)u x y) => (Select1 (DIV(Q|L|W)U x y))
    60(Mod8u x y) => (Select1 (DIVWU (ZeroExt8to16 x) (ZeroExt8to16 y)))
    61
    62(And(64|32|16|8) ...) => (AND(Q|L|L|L) ...)
    63(Or(64|32|16|8) ...) => (OR(Q|L|L|L) ...)
    64(Xor(64|32|16|8) ...) => (XOR(Q|L|L|L) ...)
    65(Com(64|32|16|8) ...) => (NOT(Q|L|L|L) ...)
    66
    67(Neg(64|32|16|8) ...) => (NEG(Q|L|L|L) ...)
    68(Neg32F x) => (PXOR x (MOVSSconst <typ.Float32> [float32(math.Copysign(0, -1))]))
    69(Neg64F x) => (PXOR x (MOVSDconst <typ.Float64> [math.Copysign(0, -1)]))
    70
    71// Lowering boolean ops
    72(AndB ...) => (ANDL ...)
    73(OrB ...) => (ORL ...)
    74(Not x) => (XORLconst [1] x)
    75
    76// Lowering pointer arithmetic
    77(OffPtr [off] ptr) && is32Bit(off) => (ADDQconst [int32(off)] ptr)
    78(OffPtr [off] ptr) => (ADDQ (MOVQconst [off]) ptr)
    79
    80// Lowering other arithmetic
    81(Ctz64 x)     && buildcfg.GOAMD64 >= 3 => (TZCNTQ x)
    82(Ctz32 x)     && buildcfg.GOAMD64 >= 3 => (TZCNTL x)
    83(Ctz64 <t> x) && buildcfg.GOAMD64 <  3 => (CMOVQEQ (Select0 <t> (BSFQ x)) (MOVQconst <t> [64]) (Select1 <types.TypeFlags> (BSFQ x)))
    84(Ctz32 x)     && buildcfg.GOAMD64 <  3 => (Select0 (BSFQ (BTSQconst <typ.UInt64> [32] x)))
    85(Ctz16 x) => (BSFL (ORLconst <typ.UInt32> [1<<16] x))
    86(Ctz8  x) => (BSFL (ORLconst <typ.UInt32> [1<<8 ] x))
    87
    88(Ctz64NonZero x) && buildcfg.GOAMD64 >= 3 => (TZCNTQ x)
    89(Ctz32NonZero x) && buildcfg.GOAMD64 >= 3 => (TZCNTL x)
    90(Ctz16NonZero x) && buildcfg.GOAMD64 >= 3 => (TZCNTL x)
    91(Ctz8NonZero  x) && buildcfg.GOAMD64 >= 3 => (TZCNTL x)
    92(Ctz64NonZero x) && buildcfg.GOAMD64 <  3 => (Select0 (BSFQ x))
    93(Ctz32NonZero x) && buildcfg.GOAMD64 <  3 => (BSFL x)
    94(Ctz16NonZero x) && buildcfg.GOAMD64 <  3 => (BSFL x)
    95(Ctz8NonZero  x) && buildcfg.GOAMD64 <  3 => (BSFL x)
    96
    97// BitLen64 of a 64 bit value x requires checking whether x == 0, since BSRQ is undefined when x == 0.
    98// However, for zero-extended values, we can cheat a bit, and calculate
    99// BSR(x<<1 + 1), which is guaranteed to be non-zero, and which conveniently
   100// places the index of the highest set bit where we want it.
   101// For GOAMD64>=3, BitLen can be calculated by OperandSize - LZCNT(x).
   102(BitLen64 <t> x) && buildcfg.GOAMD64 < 3 => (ADDQconst [1] (CMOVQEQ <t> (Select0 <t> (BSRQ x)) (MOVQconst <t> [-1]) (Select1 <types.TypeFlags> (BSRQ x))))
   103(BitLen32 x) && buildcfg.GOAMD64 <  3 => (Select0 (BSRQ (LEAQ1 <typ.UInt64> [1] (MOVLQZX <typ.UInt64> x) (MOVLQZX <typ.UInt64> x))))
   104(BitLen16 x) && buildcfg.GOAMD64 <  3 => (BSRL (LEAL1 <typ.UInt32> [1] (MOVWQZX <typ.UInt32> x) (MOVWQZX <typ.UInt32> x)))
   105(BitLen8  x) && buildcfg.GOAMD64 <  3 => (BSRL (LEAL1 <typ.UInt32> [1] (MOVBQZX <typ.UInt32> x) (MOVBQZX <typ.UInt32> x)))
   106(BitLen64 <t> x)        && buildcfg.GOAMD64 >= 3 => (NEGQ (ADDQconst <t> [-64] (LZCNTQ x)))
   107// Use 64-bit version to allow const-fold remove unnecessary arithmetic.
   108(BitLen32 <t> x) && buildcfg.GOAMD64 >= 3 => (NEGQ (ADDQconst <t> [-32] (LZCNTL x)))
   109(BitLen16 <t> x) && buildcfg.GOAMD64 >= 3 => (NEGQ (ADDQconst <t> [-32] (LZCNTL (MOVWQZX <x.Type> x))))
   110(BitLen8 <t> x) && buildcfg.GOAMD64 >= 3 => (NEGQ (ADDQconst <t> [-32] (LZCNTL (MOVBQZX <x.Type> x))))
   111
   112(Bswap(64|32) ...) => (BSWAP(Q|L) ...)
   113(Bswap16 x) => (ROLWconst [8] x)
   114
   115(PopCount(64|32) ...) => (POPCNT(Q|L) ...)
   116(PopCount16 x) => (POPCNTL (MOVWQZX <typ.UInt32> x))
   117(PopCount8 x) => (POPCNTL (MOVBQZX <typ.UInt32> x))
   118
   119(Sqrt ...) => (SQRTSD ...)
   120(Sqrt32 ...) => (SQRTSS ...)
   121
   122(RoundToEven x) => (ROUNDSD [0] x)
   123(Floor x)       => (ROUNDSD [1] x)
   124(Ceil x)        => (ROUNDSD [2] x)
   125(Trunc x)       => (ROUNDSD [3] x)
   126
   127(FMA x y z) => (VFMADD231SD z x y)
   128
   129// Lowering extension
   130// Note: we always extend to 64 bits even though some ops don't need that many result bits.
   131(SignExt8to16  ...) => (MOVBQSX ...)
   132(SignExt8to32  ...) => (MOVBQSX ...)
   133(SignExt8to64  ...) => (MOVBQSX ...)
   134(SignExt16to32 ...) => (MOVWQSX ...)
   135(SignExt16to64 ...) => (MOVWQSX ...)
   136(SignExt32to64 ...) => (MOVLQSX ...)
   137
   138(ZeroExt8to16  ...) => (MOVBQZX ...)
   139(ZeroExt8to32  ...) => (MOVBQZX ...)
   140(ZeroExt8to64  ...) => (MOVBQZX ...)
   141(ZeroExt16to32 ...) => (MOVWQZX ...)
   142(ZeroExt16to64 ...) => (MOVWQZX ...)
   143(ZeroExt32to64 ...) => (MOVLQZX ...)
   144
   145(Slicemask <t> x) => (SARQconst (NEGQ <t> x) [63])
   146
   147(SpectreIndex <t> x y) => (CMOVQCC x (MOVQconst [0]) (CMPQ x y))
   148(SpectreSliceIndex <t> x y) => (CMOVQHI x (MOVQconst [0]) (CMPQ x y))
   149
   150// Lowering truncation
   151// Because we ignore high parts of registers, truncates are just copies.
   152(Trunc16to8  ...) => (Copy ...)
   153(Trunc32to8  ...) => (Copy ...)
   154(Trunc32to16 ...) => (Copy ...)
   155(Trunc64to8  ...) => (Copy ...)
   156(Trunc64to16 ...) => (Copy ...)
   157(Trunc64to32 ...) => (Copy ...)
   158
   159// Lowering float <-> int
   160(Cvt32to32F ...) => (CVTSL2SS ...)
   161(Cvt32to64F ...) => (CVTSL2SD ...)
   162(Cvt64to32F ...) => (CVTSQ2SS ...)
   163(Cvt64to64F ...) => (CVTSQ2SD ...)
   164
   165// Float, to int.
   166// To make AMD64 "overflow" return max positive instead of max negative, compute
   167// y and not x, smear the sign bit, and xor.
   168(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)))))
   169(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))))))
   170
   171(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))) )))
   172(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)))))
   173
   174(Cvt32Fto32 <t> x) && !base.ConvertHash.MatchPos(v.Pos, nil) => (CVTTSS2SL <t> x)
   175(Cvt32Fto64 <t> x) && !base.ConvertHash.MatchPos(v.Pos, nil) => (CVTTSS2SQ <t> x)
   176(Cvt64Fto32 <t> x) && !base.ConvertHash.MatchPos(v.Pos, nil) => (CVTTSD2SL <t> x)
   177(Cvt64Fto64 <t> x) && !base.ConvertHash.MatchPos(v.Pos, nil) => (CVTTSD2SQ <t> x)
   178
   179(Cvt32Fto64F ...) => (CVTSS2SD ...)
   180(Cvt64Fto32F ...) => (CVTSD2SS ...)
   181
   182(Round(32|64)F ...) => (LoweredRound(32|64)F ...)
   183
   184// Floating-point min is tricky, as the hardware op isn't right for various special
   185// cases (-0 and NaN). We use two hardware ops organized just right to make the
   186// result come out how we want it. See https://github.com/golang/go/issues/59488#issuecomment-1553493207
   187// (although that comment isn't exactly right, as the value overwritten is not simulated correctly).
   188//    t1 = MINSD x, y   => incorrect if x==NaN or x==-0,y==+0
   189//    t2 = MINSD t1, x  => fixes x==NaN case
   190//   res = POR t1, t2   => fixes x==-0,y==+0 case
   191// Note that this trick depends on the special property that (NaN OR x) produces a NaN (although
   192// it might not produce the same NaN as the input).
   193(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))
   194// Floating-point max is even trickier. Punt to using min instead.
   195// max(x,y) == -min(-x,-y)
   196(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)))
   197
   198(CvtBoolToUint8 ...) => (Copy ...)
   199
   200// Lowering shifts
   201// Unsigned shifts need to return 0 if shift amount is >= width of shifted value.
   202//   result = (arg << shift) & (shift >= argbits ? 0 : 0xffffffffffffffff)
   203(Lsh64x(64|32|16|8) <t> x y) && !shiftIsBounded(v) => (ANDQ (SHLQ <t> x y) (SBBQcarrymask <t> (CMP(Q|L|W|B)const y [64])))
   204(Lsh32x(64|32|16|8) <t> x y) && !shiftIsBounded(v) => (ANDL (SHLL <t> x y) (SBBLcarrymask <t> (CMP(Q|L|W|B)const y [32])))
   205(Lsh16x(64|32|16|8) <t> x y) && !shiftIsBounded(v) => (ANDL (SHLL <t> x y) (SBBLcarrymask <t> (CMP(Q|L|W|B)const y [32])))
   206(Lsh8x(64|32|16|8)  <t> x y) && !shiftIsBounded(v) => (ANDL (SHLL <t> x y) (SBBLcarrymask <t> (CMP(Q|L|W|B)const y [32])))
   207
   208(Lsh64x(64|32|16|8) x y) && shiftIsBounded(v) => (SHLQ x y)
   209(Lsh32x(64|32|16|8) x y) && shiftIsBounded(v) => (SHLL x y)
   210(Lsh16x(64|32|16|8) x y) && shiftIsBounded(v) => (SHLL x y)
   211(Lsh8x(64|32|16|8)  x y) && shiftIsBounded(v) => (SHLL x y)
   212
   213(Rsh64Ux(64|32|16|8) <t> x y) && !shiftIsBounded(v) => (ANDQ (SHRQ <t> x y) (SBBQcarrymask <t> (CMP(Q|L|W|B)const y [64])))
   214(Rsh32Ux(64|32|16|8) <t> x y) && !shiftIsBounded(v) => (ANDL (SHRL <t> x y) (SBBLcarrymask <t> (CMP(Q|L|W|B)const y [32])))
   215(Rsh16Ux(64|32|16|8) <t> x y) && !shiftIsBounded(v) => (ANDL (SHRW <t> x y) (SBBLcarrymask <t> (CMP(Q|L|W|B)const y [16])))
   216(Rsh8Ux(64|32|16|8)  <t> x y) && !shiftIsBounded(v) => (ANDL (SHRB <t> x y) (SBBLcarrymask <t> (CMP(Q|L|W|B)const y [8])))
   217
   218(Rsh64Ux(64|32|16|8) x y) && shiftIsBounded(v) => (SHRQ x y)
   219(Rsh32Ux(64|32|16|8) x y) && shiftIsBounded(v) => (SHRL x y)
   220(Rsh16Ux(64|32|16|8) x y) && shiftIsBounded(v) => (SHRW x y)
   221(Rsh8Ux(64|32|16|8)  x y) && shiftIsBounded(v) => (SHRB x y)
   222
   223// Signed right shift needs to return 0/-1 if shift amount is >= width of shifted value.
   224// We implement this by setting the shift value to -1 (all ones) if the shift value is >= width.
   225(Rsh64x(64|32|16|8) <t> x y) && !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])))))
   226(Rsh32x(64|32|16|8) <t> x y) && !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])))))
   227(Rsh16x(64|32|16|8) <t> x y) && !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])))))
   228(Rsh8x(64|32|16|8)  <t> x y) && !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])))))
   229
   230(Rsh64x(64|32|16|8) x y) && shiftIsBounded(v) => (SARQ x y)
   231(Rsh32x(64|32|16|8) x y) && shiftIsBounded(v) => (SARL x y)
   232(Rsh16x(64|32|16|8) x y) && shiftIsBounded(v) => (SARW x y)
   233(Rsh8x(64|32|16|8) x y)  && shiftIsBounded(v) => (SARB x y)
   234
   235// Lowering integer comparisons
   236(Less(64|32|16|8)      x y) => (SETL  (CMP(Q|L|W|B)     x y))
   237(Less(64|32|16|8)U     x y) => (SETB  (CMP(Q|L|W|B)     x y))
   238(Leq(64|32|16|8)       x y) => (SETLE (CMP(Q|L|W|B)     x y))
   239(Leq(64|32|16|8)U      x y) => (SETBE (CMP(Q|L|W|B)     x y))
   240(Eq(Ptr|64|32|16|8|B)  x y) => (SETEQ (CMP(Q|Q|L|W|B|B) x y))
   241(Neq(Ptr|64|32|16|8|B) x y) => (SETNE (CMP(Q|Q|L|W|B|B) x y))
   242
   243// Lowering floating point comparisons
   244// Note Go assembler gets UCOMISx operand order wrong, but it is right here
   245// and the operands are reversed when generating assembly language.
   246(Eq(32|64)F   x y) => (SETEQF (UCOMIS(S|D) x y))
   247(Neq(32|64)F  x y) => (SETNEF (UCOMIS(S|D) x y))
   248// Use SETGF/SETGEF with reversed operands to dodge NaN case.
   249(Less(32|64)F x y) => (SETGF  (UCOMIS(S|D) y x))
   250(Leq(32|64)F  x y) => (SETGEF (UCOMIS(S|D) y x))
   251
   252// Lowering loads
   253(Load <t> ptr mem) && (is64BitInt(t) || isPtr(t)) => (MOVQload ptr mem)
   254(Load <t> ptr mem) && is32BitInt(t) => (MOVLload ptr mem)
   255(Load <t> ptr mem) && is16BitInt(t) => (MOVWload ptr mem)
   256(Load <t> ptr mem) && (t.IsBoolean() || is8BitInt(t)) => (MOVBload ptr mem)
   257(Load <t> ptr mem) && is32BitFloat(t) => (MOVSSload ptr mem)
   258(Load <t> ptr mem) && is64BitFloat(t) => (MOVSDload ptr mem)
   259
   260// Lowering stores
   261(Store {t} ptr val mem) && t.Size() == 8 &&  t.IsFloat() => (MOVSDstore ptr val mem)
   262(Store {t} ptr val mem) && t.Size() == 4 &&  t.IsFloat() => (MOVSSstore ptr val mem)
   263(Store {t} ptr val mem) && t.Size() == 8 && !t.IsFloat() => (MOVQstore ptr val mem)
   264(Store {t} ptr val mem) && t.Size() == 4 && !t.IsFloat() => (MOVLstore ptr val mem)
   265(Store {t} ptr val mem) && t.Size() == 2 => (MOVWstore ptr val mem)
   266(Store {t} ptr val mem) && t.Size() == 1 => (MOVBstore ptr val mem)
   267
   268// Lowering moves
   269(Move [0] _ _ mem) => mem
   270(Move [1] dst src mem) => (MOVBstore dst (MOVBload src mem) mem)
   271(Move [2] dst src mem) => (MOVWstore dst (MOVWload src mem) mem)
   272(Move [4] dst src mem) => (MOVLstore dst (MOVLload src mem) mem)
   273(Move [8] dst src mem) => (MOVQstore dst (MOVQload src mem) mem)
   274(Move [16] dst src mem) => (MOVOstore dst (MOVOload src mem) mem)
   275
   276(Move [3] dst src mem) =>
   277	(MOVBstore [2] dst (MOVBload [2] src mem)
   278		(MOVWstore dst (MOVWload src mem) mem))
   279(Move [5] dst src mem) =>
   280	(MOVBstore [4] dst (MOVBload [4] src mem)
   281		(MOVLstore dst (MOVLload src mem) mem))
   282(Move [6] dst src mem) =>
   283	(MOVWstore [4] dst (MOVWload [4] src mem)
   284		(MOVLstore dst (MOVLload src mem) mem))
   285(Move [7] dst src mem) =>
   286	(MOVLstore [3] dst (MOVLload [3] src mem)
   287		(MOVLstore dst (MOVLload src mem) mem))
   288(Move [9] dst src mem) =>
   289	(MOVBstore [8] dst (MOVBload [8] src mem)
   290		(MOVQstore dst (MOVQload src mem) mem))
   291(Move [10] dst src mem) =>
   292	(MOVWstore [8] dst (MOVWload [8] src mem)
   293		(MOVQstore dst (MOVQload src mem) mem))
   294(Move [11] dst src mem) =>
   295	(MOVLstore [7] dst (MOVLload [7] src mem)
   296		(MOVQstore dst (MOVQload src mem) mem))
   297(Move [12] dst src mem) =>
   298	(MOVLstore [8] dst (MOVLload [8] src mem)
   299		(MOVQstore dst (MOVQload src mem) mem))
   300(Move [s] dst src mem) && s >= 13 && s <= 15 =>
   301	(MOVQstore [int32(s-8)] dst (MOVQload [int32(s-8)] src mem)
   302		(MOVQstore dst (MOVQload src mem) mem))
   303
   304// Copying up to 192 bytes uses straightline code.
   305(Move [s] dst src mem) && s > 16 && s < 192 && logLargeCopy(v, s) => (LoweredMove [s] dst src mem)
   306
   307// Copying up to ~1KB uses a small loop.
   308(Move [s] dst src mem) && s >= 192 && s <= repMoveThreshold && logLargeCopy(v, s) => (LoweredMoveLoop [s] dst src mem)
   309
   310// Large copying uses REP MOVSQ.
   311(Move [s] dst src mem) && s > repMoveThreshold && s%8 != 0 =>
   312	(Move [s-s%8]
   313		(OffPtr <dst.Type> dst [s%8])
   314		(OffPtr <src.Type> src [s%8])
   315		(MOVQstore dst (MOVQload src mem) mem))
   316(Move [s] dst src mem) && s > repMoveThreshold && s%8 == 0 && logLargeCopy(v, s) =>
   317	(REPMOVSQ dst src (MOVQconst [s/8]) mem)
   318
   319// Lowering Zero instructions
   320(Zero [0] _ mem) => mem
   321(Zero [1] destptr mem) => (MOVBstoreconst [makeValAndOff(0,0)] destptr mem)
   322(Zero [2] destptr mem) => (MOVWstoreconst [makeValAndOff(0,0)] destptr mem)
   323(Zero [4] destptr mem) => (MOVLstoreconst [makeValAndOff(0,0)] destptr mem)
   324(Zero [8] destptr mem) => (MOVQstoreconst [makeValAndOff(0,0)] destptr mem)
   325
   326(Zero [3] destptr mem) =>
   327	(MOVBstoreconst [makeValAndOff(0,2)] destptr
   328		(MOVWstoreconst [makeValAndOff(0,0)] destptr mem))
   329(Zero [5] destptr mem) =>
   330	(MOVBstoreconst [makeValAndOff(0,4)] destptr
   331		(MOVLstoreconst [makeValAndOff(0,0)] destptr mem))
   332(Zero [6] destptr mem) =>
   333	(MOVWstoreconst [makeValAndOff(0,4)] destptr
   334		(MOVLstoreconst [makeValAndOff(0,0)] destptr mem))
   335(Zero [7] destptr mem) =>
   336	(MOVLstoreconst [makeValAndOff(0,3)] destptr
   337		(MOVLstoreconst [makeValAndOff(0,0)] destptr mem))
   338
   339// Zero small numbers of words directly.
   340(Zero [9] destptr mem) =>
   341	(MOVBstoreconst [makeValAndOff(0,8)] destptr
   342		(MOVQstoreconst [makeValAndOff(0,0)] destptr mem))
   343
   344(Zero [10] destptr mem) =>
   345	(MOVWstoreconst [makeValAndOff(0,8)] destptr
   346		(MOVQstoreconst [makeValAndOff(0,0)] destptr mem))
   347
   348(Zero [11] destptr mem) =>
   349	(MOVLstoreconst [makeValAndOff(0,7)] destptr
   350		(MOVQstoreconst [makeValAndOff(0,0)] destptr mem))
   351
   352(Zero [12] destptr mem) =>
   353	(MOVLstoreconst [makeValAndOff(0,8)] destptr
   354		(MOVQstoreconst [makeValAndOff(0,0)] destptr mem))
   355
   356(Zero [s] destptr mem) && s > 12 && s < 16 =>
   357	(MOVQstoreconst [makeValAndOff(0,int32(s-8))] destptr
   358		(MOVQstoreconst [makeValAndOff(0,0)] destptr mem))
   359
   360// Zeroing up to 192 bytes uses straightline code.
   361(Zero [s] destptr mem)	&& s >= 16 && s < 192 => (LoweredZero [s] destptr mem)
   362
   363// Zeroing up to ~1KB uses a small loop.
   364(Zero [s] destptr mem)	&& s >= 192 && s <= repZeroThreshold => (LoweredZeroLoop [s] destptr mem)
   365
   366// Large zeroing uses REP STOSQ.
   367(Zero [s] destptr mem) && s > repZeroThreshold && s%8 != 0 =>
   368       (Zero [s-s%8] (OffPtr <destptr.Type> destptr [s%8])
   369               (MOVOstoreconst [makeValAndOff(0,0)] destptr mem))
   370(Zero [s] destptr mem) && s > repZeroThreshold && s%8 == 0 =>
   371	(REPSTOSQ destptr (MOVQconst [s/8]) (MOVQconst [0]) mem)
   372
   373// Lowering constants
   374(Const8   [c]) => (MOVLconst [int32(c)])
   375(Const16  [c]) => (MOVLconst [int32(c)])
   376(Const32  ...) => (MOVLconst ...)
   377(Const64  ...) => (MOVQconst ...)
   378(Const32F ...) => (MOVSSconst ...)
   379(Const64F ...) => (MOVSDconst ...)
   380(ConstNil    ) => (MOVQconst [0])
   381(ConstBool [c]) => (MOVLconst [b2i32(c)])
   382
   383// Lowering calls
   384(StaticCall ...) => (CALLstatic ...)
   385(ClosureCall ...) => (CALLclosure ...)
   386(InterCall ...) => (CALLinter ...)
   387(TailCall ...) => (CALLtail ...)
   388
   389// Lowering conditional moves
   390// If the condition is a SETxx, we can just run a CMOV from the comparison that was
   391// setting the flags.
   392// Legend: HI=unsigned ABOVE, CS=unsigned BELOW, CC=unsigned ABOVE EQUAL, LS=unsigned BELOW EQUAL
   393(CondSelect <t> x y (SET(EQ|NE|L|G|LE|GE|A|B|AE|BE|EQF|NEF|GF|GEF) cond)) && (is64BitInt(t) || isPtr(t))
   394    => (CMOVQ(EQ|NE|LT|GT|LE|GE|HI|CS|CC|LS|EQF|NEF|GTF|GEF) y x cond)
   395(CondSelect <t> x y (SET(EQ|NE|L|G|LE|GE|A|B|AE|BE|EQF|NEF|GF|GEF) cond)) && is32BitInt(t)
   396    => (CMOVL(EQ|NE|LT|GT|LE|GE|HI|CS|CC|LS|EQF|NEF|GTF|GEF) y x cond)
   397(CondSelect <t> x y (SET(EQ|NE|L|G|LE|GE|A|B|AE|BE|EQF|NEF|GF|GEF) cond)) && is16BitInt(t)
   398    => (CMOVW(EQ|NE|LT|GT|LE|GE|HI|CS|CC|LS|EQF|NEF|GTF|GEF) y x cond)
   399
   400(CondSelect <t> x y check) && !check.Type.IsFlags() && check.Type.Size() == 8 && (is64BitInt(t) || isPtr(t))
   401    => (CMOVQNE y x (CMPQconst [0] check))
   402(CondSelect <t> x y check) && !check.Type.IsFlags() && check.Type.Size() == 8 && is32BitInt(t)
   403    => (CMOVLNE y x (CMPQconst [0] check))
   404(CondSelect <t> x y check) && !check.Type.IsFlags() && check.Type.Size() == 8 && is16BitInt(t)
   405    => (CMOVWNE y x (CMPQconst [0] check))
   406(CondSelect <t> x y check) && !check.Type.IsFlags() && check.Type.Size() == 4 && (is64BitInt(t) || isPtr(t))
   407    => (CMOVQNE y x (CMPLconst [0] check))
   408(CondSelect <t> x y check) && !check.Type.IsFlags() && check.Type.Size() == 4 && is32BitInt(t)
   409    => (CMOVLNE y x (CMPLconst [0] check))
   410(CondSelect <t> x y check) && !check.Type.IsFlags() && check.Type.Size() == 4 && is16BitInt(t)
   411    => (CMOVWNE y x (CMPLconst [0] check))
   412(CondSelect <t> x y check) && !check.Type.IsFlags() && check.Type.Size() == 2 && (is64BitInt(t) || isPtr(t))
   413    => (CMOVQNE y x (CMPWconst [0] check))
   414(CondSelect <t> x y check) && !check.Type.IsFlags() && check.Type.Size() == 2 && is32BitInt(t)
   415    => (CMOVLNE y x (CMPWconst [0] check))
   416(CondSelect <t> x y check) && !check.Type.IsFlags() && check.Type.Size() == 2 && is16BitInt(t)
   417    => (CMOVWNE y x (CMPWconst [0] check))
   418(CondSelect <t> x y check) && !check.Type.IsFlags() && check.Type.Size() == 1 && (is64BitInt(t) || isPtr(t))
   419    => (CMOVQNE y x (CMPBconst [0] check))
   420(CondSelect <t> x y check) && !check.Type.IsFlags() && check.Type.Size() == 1 && is32BitInt(t)
   421    => (CMOVLNE y x (CMPBconst [0] check))
   422(CondSelect <t> x y check) && !check.Type.IsFlags() && check.Type.Size() == 1 && is16BitInt(t)
   423    => (CMOVWNE y x (CMPBconst [0] check))
   424
   425// Absorb InvertFlags
   426(CMOVQ(EQ|NE|LT|GT|LE|GE|HI|CS|CC|LS) x y (InvertFlags cond))
   427    => (CMOVQ(EQ|NE|GT|LT|GE|LE|CS|HI|LS|CC) x y cond)
   428(CMOVL(EQ|NE|LT|GT|LE|GE|HI|CS|CC|LS) x y (InvertFlags cond))
   429    => (CMOVL(EQ|NE|GT|LT|GE|LE|CS|HI|LS|CC) x y cond)
   430(CMOVW(EQ|NE|LT|GT|LE|GE|HI|CS|CC|LS) x y (InvertFlags cond))
   431    => (CMOVW(EQ|NE|GT|LT|GE|LE|CS|HI|LS|CC) x y cond)
   432
   433// Absorb constants generated during lower
   434(CMOV(QEQ|QLE|QGE|QCC|QLS|LEQ|LLE|LGE|LCC|LLS|WEQ|WLE|WGE|WCC|WLS) _ x (FlagEQ)) => x
   435(CMOV(QNE|QLT|QGT|QCS|QHI|LNE|LLT|LGT|LCS|LHI|WNE|WLT|WGT|WCS|WHI) y _ (FlagEQ)) => y
   436(CMOV(QNE|QGT|QGE|QHI|QCC|LNE|LGT|LGE|LHI|LCC|WNE|WGT|WGE|WHI|WCC) _ x (FlagGT_UGT)) => x
   437(CMOV(QEQ|QLE|QLT|QLS|QCS|LEQ|LLE|LLT|LLS|LCS|WEQ|WLE|WLT|WLS|WCS) y _ (FlagGT_UGT)) => y
   438(CMOV(QNE|QGT|QGE|QLS|QCS|LNE|LGT|LGE|LLS|LCS|WNE|WGT|WGE|WLS|WCS) _ x (FlagGT_ULT)) => x
   439(CMOV(QEQ|QLE|QLT|QHI|QCC|LEQ|LLE|LLT|LHI|LCC|WEQ|WLE|WLT|WHI|WCC) y _ (FlagGT_ULT)) => y
   440(CMOV(QNE|QLT|QLE|QCS|QLS|LNE|LLT|LLE|LCS|LLS|WNE|WLT|WLE|WCS|WLS) _ x (FlagLT_ULT)) => x
   441(CMOV(QEQ|QGT|QGE|QHI|QCC|LEQ|LGT|LGE|LHI|LCC|WEQ|WGT|WGE|WHI|WCC) y _ (FlagLT_ULT)) => y
   442(CMOV(QNE|QLT|QLE|QHI|QCC|LNE|LLT|LLE|LHI|LCC|WNE|WLT|WLE|WHI|WCC) _ x (FlagLT_UGT)) => x
   443(CMOV(QEQ|QGT|QGE|QCS|QLS|LEQ|LGT|LGE|LCS|LLS|WEQ|WGT|WGE|WCS|WLS) y _ (FlagLT_UGT)) => y
   444
   445// Miscellaneous
   446(IsNonNil p) => (SETNE (TESTQ p p))
   447(IsInBounds idx len) => (SETB (CMPQ idx len))
   448(IsSliceInBounds idx len) => (SETBE (CMPQ idx len))
   449(NilCheck ...) => (LoweredNilCheck ...)
   450(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.
   451(GetClosurePtr ...) => (LoweredGetClosurePtr ...)
   452(GetCallerPC ...) => (LoweredGetCallerPC ...)
   453(GetCallerSP ...) => (LoweredGetCallerSP ...)
   454
   455(HasCPUFeature {s}) => (SETNE (CMPLconst [0] (LoweredHasCPUFeature {s})))
   456(Addr {sym} base) => (LEAQ {sym} base)
   457(LocalAddr <t> {sym} base mem) && t.Elem().HasPointers() => (LEAQ {sym} (SPanchored base mem))
   458(LocalAddr <t> {sym} base _)  && !t.Elem().HasPointers() => (LEAQ {sym} base)
   459
   460(MOVBstore [off] {sym} ptr y:(SETL x) mem) && y.Uses == 1 => (SETLstore [off] {sym} ptr x mem)
   461(MOVBstore [off] {sym} ptr y:(SETLE x) mem) && y.Uses == 1 => (SETLEstore [off] {sym} ptr x mem)
   462(MOVBstore [off] {sym} ptr y:(SETG x) mem) && y.Uses == 1 => (SETGstore [off] {sym} ptr x mem)
   463(MOVBstore [off] {sym} ptr y:(SETGE x) mem) && y.Uses == 1 => (SETGEstore [off] {sym} ptr x mem)
   464(MOVBstore [off] {sym} ptr y:(SETEQ x) mem) && y.Uses == 1 => (SETEQstore [off] {sym} ptr x mem)
   465(MOVBstore [off] {sym} ptr y:(SETNE x) mem) && y.Uses == 1 => (SETNEstore [off] {sym} ptr x mem)
   466(MOVBstore [off] {sym} ptr y:(SETB x) mem) && y.Uses == 1 => (SETBstore [off] {sym} ptr x mem)
   467(MOVBstore [off] {sym} ptr y:(SETBE x) mem) && y.Uses == 1 => (SETBEstore [off] {sym} ptr x mem)
   468(MOVBstore [off] {sym} ptr y:(SETA x) mem) && y.Uses == 1 => (SETAstore [off] {sym} ptr x mem)
   469(MOVBstore [off] {sym} ptr y:(SETAE x) mem) && y.Uses == 1 => (SETAEstore [off] {sym} ptr x mem)
   470
   471// block rewrites
   472(If (SETL  cmp) yes no) => (LT  cmp yes no)
   473(If (SETLE cmp) yes no) => (LE  cmp yes no)
   474(If (SETG  cmp) yes no) => (GT  cmp yes no)
   475(If (SETGE cmp) yes no) => (GE  cmp yes no)
   476(If (SETEQ cmp) yes no) => (EQ  cmp yes no)
   477(If (SETNE cmp) yes no) => (NE  cmp yes no)
   478(If (SETB  cmp) yes no) => (ULT cmp yes no)
   479(If (SETBE cmp) yes no) => (ULE cmp yes no)
   480(If (SETA  cmp) yes no) => (UGT cmp yes no)
   481(If (SETAE cmp) yes no) => (UGE cmp yes no)
   482(If (SETO cmp) yes no) => (OS cmp yes no)
   483
   484// Special case for floating point - LF/LEF not generated
   485(If (SETGF  cmp) yes no) => (UGT  cmp yes no)
   486(If (SETGEF cmp) yes no) => (UGE  cmp yes no)
   487(If (SETEQF cmp) yes no) => (EQF  cmp yes no)
   488(If (SETNEF cmp) yes no) => (NEF  cmp yes no)
   489
   490(If cond yes no) => (NE (TESTB cond cond) yes no)
   491
   492(JumpTable idx) => (JUMPTABLE {makeJumpTableSym(b)} idx (LEAQ <typ.Uintptr> {makeJumpTableSym(b)} (SB)))
   493
   494// Atomic loads.  Other than preserving their ordering with respect to other loads, nothing special here.
   495(AtomicLoad8 ptr mem) => (MOVBatomicload ptr mem)
   496(AtomicLoad32 ptr mem) => (MOVLatomicload ptr mem)
   497(AtomicLoad64 ptr mem) => (MOVQatomicload ptr mem)
   498(AtomicLoadPtr ptr mem) => (MOVQatomicload ptr mem)
   499
   500// Atomic stores.  We use XCHG to prevent the hardware reordering a subsequent load.
   501// TODO: most runtime uses of atomic stores don't need that property.  Use normal stores for those?
   502(AtomicStore8 ptr val mem) => (Select1 (XCHGB <types.NewTuple(typ.UInt8,types.TypeMem)> val ptr mem))
   503(AtomicStore32 ptr val mem) => (Select1 (XCHGL <types.NewTuple(typ.UInt32,types.TypeMem)> val ptr mem))
   504(AtomicStore64 ptr val mem) => (Select1 (XCHGQ <types.NewTuple(typ.UInt64,types.TypeMem)> val ptr mem))
   505(AtomicStorePtrNoWB ptr val mem) => (Select1 (XCHGQ <types.NewTuple(typ.BytePtr,types.TypeMem)> val ptr mem))
   506
   507// Atomic exchanges.
   508(AtomicExchange8 ptr val mem) => (XCHGB val ptr mem)
   509(AtomicExchange32 ptr val mem) => (XCHGL val ptr mem)
   510(AtomicExchange64 ptr val mem) => (XCHGQ val ptr mem)
   511
   512// Atomic adds.
   513(AtomicAdd32 ptr val mem) => (AddTupleFirst32 val (XADDLlock val ptr mem))
   514(AtomicAdd64 ptr val mem) => (AddTupleFirst64 val (XADDQlock val ptr mem))
   515(Select0 <t> (AddTupleFirst32 val tuple)) => (ADDL val (Select0 <t> tuple))
   516(Select1     (AddTupleFirst32   _ tuple)) => (Select1 tuple)
   517(Select0 <t> (AddTupleFirst64 val tuple)) => (ADDQ val (Select0 <t> tuple))
   518(Select1     (AddTupleFirst64   _ tuple)) => (Select1 tuple)
   519
   520// Atomic compare and swap.
   521(AtomicCompareAndSwap32 ptr old new_ mem) => (CMPXCHGLlock ptr old new_ mem)
   522(AtomicCompareAndSwap64 ptr old new_ mem) => (CMPXCHGQlock ptr old new_ mem)
   523
   524// Atomic memory logical operations (old style).
   525(AtomicAnd8  ptr val mem) => (ANDBlock ptr val mem)
   526(AtomicAnd32 ptr val mem) => (ANDLlock ptr val mem)
   527(AtomicOr8   ptr val mem) => (ORBlock  ptr val mem)
   528(AtomicOr32  ptr val mem) => (ORLlock  ptr val mem)
   529
   530// Atomic memory logical operations (new style).
   531(Atomic(And64|And32|Or64|Or32)value ptr val mem) => (LoweredAtomic(And64|And32|Or64|Or32) ptr val mem)
   532
   533// Write barrier.
   534(WB ...) => (LoweredWB ...)
   535
   536(PanicBounds ...) => (LoweredPanicBoundsRR ...)
   537(LoweredPanicBoundsRR [kind] x (MOVQconst [c]) mem) => (LoweredPanicBoundsRC [kind] x {PanicBoundsC{C:c}} mem)
   538(LoweredPanicBoundsRR [kind] (MOVQconst [c]) y mem) => (LoweredPanicBoundsCR [kind] {PanicBoundsC{C:c}} y mem)
   539(LoweredPanicBoundsRC [kind] {p} (MOVQconst [c]) mem) => (LoweredPanicBoundsCC [kind] {PanicBoundsCC{Cx:c, Cy:p.C}} mem)
   540(LoweredPanicBoundsCR [kind] {p} (MOVQconst [c]) mem) => (LoweredPanicBoundsCC [kind] {PanicBoundsCC{Cx:p.C, Cy:c}} mem)
   541
   542// lowering rotates
   543(RotateLeft8  ...) => (ROLB ...)
   544(RotateLeft16 ...) => (ROLW ...)
   545(RotateLeft32 ...) => (ROLL ...)
   546(RotateLeft64 ...) => (ROLQ ...)
   547
   548// ***************************
   549// Above: lowering rules
   550// Below: optimizations
   551// ***************************
   552// TODO: Should the optimizations be a separate pass?
   553
   554// Fold boolean tests into blocks
   555(NE (TESTB (SETL  cmp) (SETL  cmp)) yes no) => (LT  cmp yes no)
   556(NE (TESTB (SETLE cmp) (SETLE cmp)) yes no) => (LE  cmp yes no)
   557(NE (TESTB (SETG  cmp) (SETG  cmp)) yes no) => (GT  cmp yes no)
   558(NE (TESTB (SETGE cmp) (SETGE cmp)) yes no) => (GE  cmp yes no)
   559(NE (TESTB (SETEQ cmp) (SETEQ cmp)) yes no) => (EQ  cmp yes no)
   560(NE (TESTB (SETNE cmp) (SETNE cmp)) yes no) => (NE  cmp yes no)
   561(NE (TESTB (SETB  cmp) (SETB  cmp)) yes no) => (ULT cmp yes no)
   562(NE (TESTB (SETBE cmp) (SETBE cmp)) yes no) => (ULE cmp yes no)
   563(NE (TESTB (SETA  cmp) (SETA  cmp)) yes no) => (UGT cmp yes no)
   564(NE (TESTB (SETAE cmp) (SETAE cmp)) yes no) => (UGE cmp yes no)
   565(NE (TESTB (SETO cmp) (SETO cmp)) yes no) => (OS cmp yes no)
   566
   567// Unsigned comparisons to 0/1
   568(ULT (TEST(Q|L|W|B) x x) yes no) => (First no yes)
   569(UGE (TEST(Q|L|W|B) x x) yes no) => (First yes no)
   570(SETB (TEST(Q|L|W|B) x x)) => (ConstBool [false])
   571(SETAE (TEST(Q|L|W|B) x x)) => (ConstBool [true])
   572
   573// x & 1 != 0 -> x & 1
   574(SETNE (TEST(B|W)const [1] x)) => (AND(L|L)const [1] x)
   575(SETB (BT(L|Q)const [0] x)) => (AND(L|Q)const [1] x)
   576// x & 1 == 0 -> (x & 1) ^ 1
   577(SETAE (BT(L|Q)const [0] x)) => (XORLconst [1] (ANDLconst <typ.Bool> [1] x))
   578
   579// Shorten compare by rewriting x < 128 as x <= 127, which can be encoded in a single-byte immediate on x86.
   580(SETL c:(CMP(Q|L)const [128] x)) && c.Uses == 1 => (SETLE (CMP(Q|L)const [127] x))
   581(SETB c:(CMP(Q|L)const [128] x)) && c.Uses == 1 => (SETBE (CMP(Q|L)const [127] x))
   582
   583// x >= 128 -> x > 127
   584(SETGE c:(CMP(Q|L)const [128] x)) && c.Uses == 1 => (SETG (CMP(Q|L)const [127] x))
   585(SETAE c:(CMP(Q|L)const [128] x)) && c.Uses == 1 => (SETA (CMP(Q|L)const [127] x))
   586
   587(CMOVQLT x y c:(CMP(Q|L)const [128] z)) && c.Uses == 1 => (CMOVQLE x y (CMP(Q|L)const [127] z))
   588(CMOVLLT x y c:(CMP(Q|L)const [128] z)) && c.Uses == 1 => (CMOVLLE x y (CMP(Q|L)const [127] z))
   589(LT          c:(CMP(Q|L)const [128] z) yes no) && c.Uses == 1 => (LE (CMP(Q|L)const [127] z) yes no)
   590(CMOVQGE x y c:(CMP(Q|L)const [128] z)) && c.Uses == 1 => (CMOVQGT x y (CMP(Q|L)const [127] z))
   591(CMOVLGE x y c:(CMP(Q|L)const [128] z)) && c.Uses == 1 => (CMOVLGT x y (CMP(Q|L)const [127] z))
   592(GE          c:(CMP(Q|L)const [128] z) yes no) && c.Uses == 1 => (GT (CMP(Q|L)const [127] z)  yes no)
   593
   594// Recognize bit tests: a&(1<<b) != 0 for b suitably bounded
   595// Note that BTx instructions use the carry bit, so we need to convert tests for zero flag
   596// into tests for carry flags.
   597// ULT and SETB check the carry flag; they are identical to CS and SETCS. Same, mutatis
   598// mutandis, for UGE and SETAE, and CC and SETCC.
   599((NE|EQ) (TESTL (SHLL (MOVLconst [1]) x) y)) => ((ULT|UGE) (BTL x y))
   600((NE|EQ) (TESTQ (SHLQ (MOVQconst [1]) x) y)) => ((ULT|UGE) (BTQ x y))
   601((NE|EQ) (TESTLconst [c] x)) && isUnsignedPowerOfTwo(uint32(c))
   602    => ((ULT|UGE) (BTLconst [int8(log32u(uint32(c)))] x))
   603((NE|EQ) (TESTQconst [c] x)) && isUnsignedPowerOfTwo(uint64(c))
   604    => ((ULT|UGE) (BTQconst [int8(log32u(uint32(c)))] x))
   605((NE|EQ) (TESTQ (MOVQconst [c]) x)) && isUnsignedPowerOfTwo(uint64(c))
   606    => ((ULT|UGE) (BTQconst [int8(log64u(uint64(c)))] x))
   607(SET(NE|EQ) (TESTL (SHLL (MOVLconst [1]) x) y)) => (SET(B|AE)  (BTL x y))
   608(SET(NE|EQ) (TESTQ (SHLQ (MOVQconst [1]) x) y)) => (SET(B|AE)  (BTQ x y))
   609(SET(NE|EQ) (TESTLconst [c] x)) && isUnsignedPowerOfTwo(uint32(c))
   610    => (SET(B|AE)  (BTLconst [int8(log32u(uint32(c)))] x))
   611(SET(NE|EQ) (TESTQconst [c] x)) && isUnsignedPowerOfTwo(uint64(c))
   612    => (SET(B|AE)  (BTQconst [int8(log32u(uint32(c)))] x))
   613(SET(NE|EQ) (TESTQ (MOVQconst [c]) x)) && isUnsignedPowerOfTwo(uint64(c))
   614    => (SET(B|AE)  (BTQconst [int8(log64u(uint64(c)))] x))
   615// SET..store variant
   616(SET(NE|EQ)store [off] {sym} ptr (TESTL (SHLL (MOVLconst [1]) x) y) mem)
   617    => (SET(B|AE)store  [off] {sym} ptr (BTL x y) mem)
   618(SET(NE|EQ)store [off] {sym} ptr (TESTQ (SHLQ (MOVQconst [1]) x) y) mem)
   619    => (SET(B|AE)store  [off] {sym} ptr (BTQ x y) mem)
   620(SET(NE|EQ)store [off] {sym} ptr (TESTLconst [c] x) mem) && isUnsignedPowerOfTwo(uint32(c))
   621    => (SET(B|AE)store  [off] {sym} ptr (BTLconst [int8(log32u(uint32(c)))] x) mem)
   622(SET(NE|EQ)store [off] {sym} ptr (TESTQconst [c] x) mem) && isUnsignedPowerOfTwo(uint64(c))
   623    => (SET(B|AE)store  [off] {sym} ptr (BTQconst [int8(log32u(uint32(c)))] x) mem)
   624(SET(NE|EQ)store [off] {sym} ptr (TESTQ (MOVQconst [c]) x) mem) && isUnsignedPowerOfTwo(uint64(c))
   625    => (SET(B|AE)store  [off] {sym} ptr (BTQconst [int8(log64u(uint64(c)))] x) mem)
   626
   627// Handle bit-testing in the form (a>>b)&1 != 0 by building the above rules
   628// and further combining shifts.
   629(BT(Q|L)const [c] (SHRQconst [d] x)) && (c+d)<64 => (BTQconst [c+d] x)
   630(BT(Q|L)const [c] (ADDQ x x)) && c>1  => (BT(Q|L)const [c-1] x)
   631(BT(Q|L)const [c] (SHLQconst [d] x)) && c>d      => (BT(Q|L)const [c-d] x)
   632(BT(Q|L)const [0] s:(SHRQ x y)) => (BTQ y x)
   633(BTLconst [c] (SHRLconst [d] x)) && (c+d)<32 => (BTLconst [c+d] x)
   634(BTLconst [c] (ADDL x x)) && c>1 => (BTLconst [c-1] x)
   635(BTLconst [c] (SHLLconst [d] x)) && c>d      => (BTLconst [c-d] x)
   636(BTLconst [0] s:(SHR(L|XL) x y)) => (BTL y x)
   637
   638// Rewrite a & 1 != 1 into a & 1 == 0.
   639// Among other things, this lets us turn (a>>b)&1 != 1 into a bit test.
   640(SET(NE|EQ) (CMPLconst [1] s:(ANDLconst [1] _))) => (SET(EQ|NE) (CMPLconst [0] s))
   641(SET(NE|EQ)store [off] {sym} ptr (CMPLconst [1] s:(ANDLconst [1] _)) mem) => (SET(EQ|NE)store [off] {sym} ptr (CMPLconst [0] s) mem)
   642(SET(NE|EQ) (CMPQconst [1] s:(ANDQconst [1] _))) => (SET(EQ|NE) (CMPQconst [0] s))
   643(SET(NE|EQ)store [off] {sym} ptr (CMPQconst [1] s:(ANDQconst [1] _)) mem) => (SET(EQ|NE)store [off] {sym} ptr (CMPQconst [0] s) mem)
   644
   645// Recognize bit setting (a |= 1<<b) and toggling (a ^= 1<<b)
   646(OR(Q|L) (SHL(Q|L) (MOV(Q|L)const [1]) y) x) => (BTS(Q|L) x y)
   647(XOR(Q|L) (SHL(Q|L) (MOV(Q|L)const [1]) y) x) => (BTC(Q|L) x y)
   648// Note: only convert OR/XOR to BTS/BTC if the constant wouldn't fit in
   649// the constant field of the OR/XOR instruction. See issue 61694.
   650((OR|XOR)Q (MOVQconst [c]) x) && isUnsignedPowerOfTwo(uint64(c)) && uint64(c) >= 1<<31 => (BT(S|C)Qconst [int8(log64u(uint64(c)))] x)
   651
   652// Recognize bit clearing: a &^= 1<<b
   653(AND(Q|L) (NOT(Q|L) (SHL(Q|L) (MOV(Q|L)const [1]) y)) x) => (BTR(Q|L) x y)
   654(ANDN(Q|L) x (SHL(Q|L) (MOV(Q|L)const [1]) y)) => (BTR(Q|L) x y)
   655// Note: only convert AND to BTR if the constant wouldn't fit in
   656// the constant field of the AND instruction. See issue 61694.
   657(ANDQ (MOVQconst [c]) x) && isUnsignedPowerOfTwo(uint64(^c)) && uint64(^c) >= 1<<31 => (BTRQconst [int8(log64u(uint64(^c)))] x)
   658
   659// Special-case bit patterns on first/last bit.
   660// generic.rules changes ANDs of high-part/low-part masks into a couple of shifts,
   661// for instance:
   662//    x & 0xFFFF0000 -> (x >> 16) << 16
   663//    x & 0x80000000 -> (x >> 31) << 31
   664//
   665// In case the mask is just one bit (like second example above), it conflicts
   666// with the above rules to detect bit-testing / bit-clearing of first/last bit.
   667// We thus special-case them, by detecting the shift patterns.
   668
   669// Special case resetting first/last bit
   670(ADD(L|Q) (SHR(L|Q)const [1] x) (SHR(L|Q)const [1] x))
   671	=> (AND(L|Q)const [-2] x)
   672(SHRLconst [1] (ADDL x x))
   673	=> (ANDLconst [0x7fffffff] x)
   674(SHRQconst [1] (ADDQ x x))
   675	=> (BTRQconst [63] x)
   676
   677// Special case testing first/last bit (with double-shift generated by generic.rules)
   678((SETNE|SETEQ|NE|EQ) (TESTQ z1:(SHLQconst [63] (SHRQconst [63] x)) z2)) && z1==z2
   679    => ((SETB|SETAE|ULT|UGE) (BTQconst [63] x))
   680((SETNE|SETEQ|NE|EQ) (TESTL z1:(SHLLconst [31] (SHRQconst [31] x)) z2)) && z1==z2
   681    => ((SETB|SETAE|ULT|UGE) (BTQconst [31] x))
   682(SET(NE|EQ)store [off] {sym} ptr (TESTQ z1:(SHLQconst [63] (SHRQconst [63] x)) z2) mem) && z1==z2
   683    => (SET(B|AE)store [off] {sym} ptr (BTQconst [63] x) mem)
   684(SET(NE|EQ)store [off] {sym} ptr (TESTL z1:(SHLLconst [31] (SHRLconst [31] x)) z2) mem) && z1==z2
   685    => (SET(B|AE)store [off] {sym} ptr (BTLconst [31] x) mem)
   686
   687((SETNE|SETEQ|NE|EQ) (TESTQ z1:(SHRQconst [63] (SHLQconst [63] x)) z2)) && z1==z2
   688    => ((SETB|SETAE|ULT|UGE)  (BTQconst [0] x))
   689((SETNE|SETEQ|NE|EQ) (TESTL z1:(SHRLconst [31] (SHLLconst [31] x)) z2)) && z1==z2
   690    => ((SETB|SETAE|ULT|UGE)  (BTLconst [0] x))
   691(SET(NE|EQ)store [off] {sym} ptr (TESTQ z1:(SHRQconst [63] (SHLQconst [63] x)) z2) mem) && z1==z2
   692    => (SET(B|AE)store [off] {sym} ptr (BTQconst [0] x) mem)
   693(SET(NE|EQ)store [off] {sym} ptr (TESTL z1:(SHRLconst [31] (SHLLconst [31] x)) z2) mem) && z1==z2
   694    => (SET(B|AE)store [off] {sym} ptr (BTLconst [0] x) mem)
   695
   696// Special-case manually testing last bit with "a>>63 != 0" (without "&1")
   697((SETNE|SETEQ|NE|EQ) (TESTQ z1:(SHRQconst [63] x) z2)) && z1==z2
   698    => ((SETB|SETAE|ULT|UGE) (BTQconst [63] x))
   699((SETNE|SETEQ|NE|EQ) (TESTL z1:(SHRLconst [31] x) z2)) && z1==z2
   700    => ((SETB|SETAE|ULT|UGE) (BTLconst [31] x))
   701(SET(NE|EQ)store [off] {sym} ptr (TESTQ z1:(SHRQconst [63] x) z2) mem) && z1==z2
   702    => (SET(B|AE)store [off] {sym} ptr (BTQconst [63] x) mem)
   703(SET(NE|EQ)store [off] {sym} ptr (TESTL z1:(SHRLconst [31] x) z2) mem) && z1==z2
   704    => (SET(B|AE)store [off] {sym} ptr (BTLconst [31] x) mem)
   705
   706// Fold combinations of bit ops on same bit. An example is math.Copysign(c,-1)
   707(BTSQconst [c] (BTRQconst [c] x)) => (BTSQconst [c] x)
   708(BTSQconst [c] (BTCQconst [c] x)) => (BTSQconst [c] x)
   709(BTRQconst [c] (BTSQconst [c] x)) => (BTRQconst [c] x)
   710(BTRQconst [c] (BTCQconst [c] x)) => (BTRQconst [c] x)
   711
   712// Fold boolean negation into SETcc.
   713(XORLconst [1] (SETNE x)) => (SETEQ x)
   714(XORLconst [1] (SETEQ x)) => (SETNE x)
   715(XORLconst [1] (SETL  x)) => (SETGE x)
   716(XORLconst [1] (SETGE x)) => (SETL  x)
   717(XORLconst [1] (SETLE x)) => (SETG  x)
   718(XORLconst [1] (SETG  x)) => (SETLE x)
   719(XORLconst [1] (SETB  x)) => (SETAE x)
   720(XORLconst [1] (SETAE x)) => (SETB  x)
   721(XORLconst [1] (SETBE x)) => (SETA  x)
   722(XORLconst [1] (SETA  x)) => (SETBE x)
   723
   724// Special case for floating point - LF/LEF not generated
   725(NE (TESTB (SETGF  cmp) (SETGF  cmp)) yes no) => (UGT  cmp yes no)
   726(NE (TESTB (SETGEF cmp) (SETGEF cmp)) yes no) => (UGE  cmp yes no)
   727(NE (TESTB (SETEQF cmp) (SETEQF cmp)) yes no) => (EQF  cmp yes no)
   728(NE (TESTB (SETNEF cmp) (SETNEF cmp)) yes no) => (NEF  cmp yes no)
   729
   730// Disabled because it interferes with the pattern match above and makes worse code.
   731// (SETNEF x) => (ORQ (SETNE <typ.Int8> x) (SETNAN <typ.Int8> x))
   732// (SETEQF x) => (ANDQ (SETEQ <typ.Int8> x) (SETORD <typ.Int8> x))
   733
   734// fold constants into instructions
   735(ADDQ x (MOVQconst <t> [c])) && is32Bit(c) && !t.IsPtr() => (ADDQconst [int32(c)] x)
   736(ADDQ x (MOVLconst [c])) => (ADDQconst [c] x)
   737(ADDL x (MOVLconst [c])) => (ADDLconst [c] x)
   738
   739(SUBQ x (MOVQconst [c])) && is32Bit(c) => (SUBQconst x [int32(c)])
   740(SUBQ (MOVQconst [c]) x) && is32Bit(c) => (NEGQ (SUBQconst <v.Type> x [int32(c)]))
   741(SUBL x (MOVLconst [c])) => (SUBLconst x [c])
   742(SUBL (MOVLconst [c]) x) => (NEGL (SUBLconst <v.Type> x [c]))
   743
   744(MULQ x (MOVQconst [c])) && is32Bit(c) => (MULQconst [int32(c)] x)
   745(MULL x (MOVLconst [c])) => (MULLconst [c] x)
   746
   747(ANDQ x (MOVQconst [c])) && is32Bit(c) => (ANDQconst [int32(c)] x)
   748(ANDL x (MOVLconst [c])) => (ANDLconst [c] x)
   749
   750(AND(L|Q)const [c] (AND(L|Q)const [d] x)) => (AND(L|Q)const [c & d] x)
   751(XOR(L|Q)const [c] (XOR(L|Q)const [d] x)) => (XOR(L|Q)const [c ^ d] x)
   752(OR(L|Q)const  [c] (OR(L|Q)const  [d] x)) => (OR(L|Q)const  [c | d] x)
   753
   754(MULLconst [c] (MULLconst [d] x)) => (MULLconst [c * d] x)
   755(MULQconst [c] (MULQconst [d] x)) && is32Bit(int64(c)*int64(d)) => (MULQconst [c * d] x)
   756
   757(ORQ x (MOVQconst [c])) && is32Bit(c) => (ORQconst [int32(c)] x)
   758(ORQ x (MOVLconst [c])) => (ORQconst [c] x)
   759(ORL x (MOVLconst [c])) => (ORLconst [c] x)
   760
   761(XORQ x (MOVQconst [c])) && is32Bit(c) => (XORQconst [int32(c)] x)
   762(XORL x (MOVLconst [c])) => (XORLconst [c] x)
   763
   764(SHLQ x (MOV(Q|L)const [c])) => (SHLQconst [int8(c&63)] x)
   765(SHLL x (MOV(Q|L)const [c])) => (SHLLconst [int8(c&31)] x)
   766
   767(SHRQ x (MOV(Q|L)const [c])) => (SHRQconst [int8(c&63)] x)
   768(SHRL x (MOV(Q|L)const [c])) => (SHRLconst [int8(c&31)] x)
   769(SHRW x (MOV(Q|L)const [c])) && c&31 < 16 => (SHRWconst [int8(c&31)] x)
   770(SHRW _ (MOV(Q|L)const [c])) && c&31 >= 16 => (MOVLconst [0])
   771(SHRB x (MOV(Q|L)const [c])) && c&31 < 8 => (SHRBconst [int8(c&31)] x)
   772(SHRB _ (MOV(Q|L)const [c])) && c&31 >= 8 => (MOVLconst [0])
   773
   774(SARQ x (MOV(Q|L)const [c])) => (SARQconst [int8(c&63)] x)
   775(SARL x (MOV(Q|L)const [c])) => (SARLconst [int8(c&31)] x)
   776(SARW x (MOV(Q|L)const [c])) => (SARWconst [int8(min(int64(c)&31,15))] x)
   777(SARB x (MOV(Q|L)const [c])) => (SARBconst [int8(min(int64(c)&31,7))] x)
   778
   779// Operations which don't affect the low 6/5 bits of the shift amount are NOPs.
   780((SHLQ|SHRQ|SARQ) x (ADDQconst [c] y)) && c & 63 == 0  => ((SHLQ|SHRQ|SARQ) x y)
   781((SHLQ|SHRQ|SARQ) x (NEGQ <t> (ADDQconst [c] y))) && c & 63 == 0  => ((SHLQ|SHRQ|SARQ) x (NEGQ <t> y))
   782((SHLQ|SHRQ|SARQ) x (ANDQconst [c] y)) && c & 63 == 63 => ((SHLQ|SHRQ|SARQ) x y)
   783((SHLQ|SHRQ|SARQ) x (NEGQ <t> (ANDQconst [c] y))) && c & 63 == 63 => ((SHLQ|SHRQ|SARQ) x (NEGQ <t> y))
   784
   785((SHLL|SHRL|SARL) x (ADDQconst [c] y)) && c & 31 == 0  => ((SHLL|SHRL|SARL) x y)
   786((SHLL|SHRL|SARL) x (NEGQ <t> (ADDQconst [c] y))) && c & 31 == 0  => ((SHLL|SHRL|SARL) x (NEGQ <t> y))
   787((SHLL|SHRL|SARL) x (ANDQconst [c] y)) && c & 31 == 31 => ((SHLL|SHRL|SARL) x y)
   788((SHLL|SHRL|SARL) x (NEGQ <t> (ANDQconst [c] y))) && c & 31 == 31 => ((SHLL|SHRL|SARL) x (NEGQ <t> y))
   789
   790((SHLQ|SHRQ|SARQ) x (ADDLconst [c] y)) && c & 63 == 0  => ((SHLQ|SHRQ|SARQ) x y)
   791((SHLQ|SHRQ|SARQ) x (NEGL <t> (ADDLconst [c] y))) && c & 63 == 0  => ((SHLQ|SHRQ|SARQ) x (NEGL <t> y))
   792((SHLQ|SHRQ|SARQ) x (ANDLconst [c] y)) && c & 63 == 63 => ((SHLQ|SHRQ|SARQ) x y)
   793((SHLQ|SHRQ|SARQ) x (NEGL <t> (ANDLconst [c] y))) && c & 63 == 63 => ((SHLQ|SHRQ|SARQ) x (NEGL <t> y))
   794
   795((SHLL|SHRL|SARL) x (ADDLconst [c] y)) && c & 31 == 0  => ((SHLL|SHRL|SARL) x y)
   796((SHLL|SHRL|SARL) x (NEGL <t> (ADDLconst [c] y))) && c & 31 == 0  => ((SHLL|SHRL|SARL) x (NEGL <t> y))
   797((SHLL|SHRL|SARL) x (ANDLconst [c] y)) && c & 31 == 31 => ((SHLL|SHRL|SARL) x y)
   798((SHLL|SHRL|SARL) x (NEGL <t> (ANDLconst [c] y))) && c & 31 == 31 => ((SHLL|SHRL|SARL) x (NEGL <t> y))
   799
   800// rotate left negative = rotate right
   801(ROLQ x (NEG(Q|L) y)) => (RORQ x y)
   802(ROLL x (NEG(Q|L) y)) => (RORL x y)
   803(ROLW x (NEG(Q|L) y)) => (RORW x y)
   804(ROLB x (NEG(Q|L) y)) => (RORB x y)
   805
   806// rotate right negative = rotate left
   807(RORQ x (NEG(Q|L) y)) => (ROLQ x y)
   808(RORL x (NEG(Q|L) y)) => (ROLL x y)
   809(RORW x (NEG(Q|L) y)) => (ROLW x y)
   810(RORB x (NEG(Q|L) y)) => (ROLB x y)
   811
   812// rotate by constants
   813(ROLQ x (MOV(Q|L)const [c])) => (ROLQconst [int8(c&63)] x)
   814(ROLL x (MOV(Q|L)const [c])) => (ROLLconst [int8(c&31)] x)
   815(ROLW x (MOV(Q|L)const [c])) => (ROLWconst [int8(c&15)] x)
   816(ROLB x (MOV(Q|L)const [c])) => (ROLBconst [int8(c&7) ] x)
   817
   818(RORQ x (MOV(Q|L)const [c])) => (ROLQconst [int8((-c)&63)] x)
   819(RORL x (MOV(Q|L)const [c])) => (ROLLconst [int8((-c)&31)] x)
   820(RORW x (MOV(Q|L)const [c])) => (ROLWconst [int8((-c)&15)] x)
   821(RORB x (MOV(Q|L)const [c])) => (ROLBconst [int8((-c)&7) ] x)
   822
   823// Constant shift simplifications
   824((SHLQ|SHRQ|SARQ)const      x [0]) => x
   825((SHLL|SHRL|SARL)const      x [0]) => x
   826((SHRW|SARW)const           x [0]) => x
   827((SHRB|SARB)const           x [0]) => x
   828((ROLQ|ROLL|ROLW|ROLB)const x [0]) => x
   829
   830// Multi-register shifts
   831(ORQ (SH(R|L)Q lo bits) (SH(L|R)Q hi (NEGQ bits))) => (SH(R|L)DQ lo hi bits)
   832(ORQ (SH(R|L)XQ lo bits) (SH(L|R)XQ hi (NEGQ bits))) => (SH(R|L)DQ lo hi bits)
   833
   834// Note: the word and byte shifts keep the low 5 bits (not the low 4 or 3 bits)
   835// because the x86 instructions are defined to use all 5 bits of the shift even
   836// for the small shifts. I don't think we'll ever generate a weird shift (e.g.
   837// (SHRW x (MOVLconst [24])), but just in case.
   838
   839(CMPQ x (MOVQconst [c])) && is32Bit(c) => (CMPQconst x [int32(c)])
   840(CMPQ (MOVQconst [c]) x) && is32Bit(c) => (InvertFlags (CMPQconst x [int32(c)]))
   841(CMPL x (MOVLconst [c])) => (CMPLconst x [c])
   842(CMPL (MOVLconst [c]) x) => (InvertFlags (CMPLconst x [c]))
   843(CMPW x (MOVLconst [c])) => (CMPWconst x [int16(c)])
   844(CMPW (MOVLconst [c]) x) => (InvertFlags (CMPWconst x [int16(c)]))
   845(CMPB x (MOVLconst [c])) => (CMPBconst x [int8(c)])
   846(CMPB (MOVLconst [c]) x) => (InvertFlags (CMPBconst x [int8(c)]))
   847
   848// Canonicalize the order of arguments to comparisons - helps with CSE.
   849(CMP(Q|L|W|B) x y) && canonLessThan(x,y) => (InvertFlags (CMP(Q|L|W|B) y x))
   850
   851// Using MOVZX instead of AND is cheaper.
   852(AND(Q|L)const [  0xFF] x) => (MOVBQZX x)
   853(AND(Q|L)const [0xFFFF] x) => (MOVWQZX x)
   854// This rule is currently invalid because 0xFFFFFFFF is not representable by a signed int32.
   855// Commenting out for now, because it also can't trigger because of the is32bit guard on the
   856// ANDQconst lowering-rule, above, prevents 0xFFFFFFFF from matching (for the same reason)
   857// Using an alternate form of this rule segfaults some binaries because of
   858// adverse interactions with other passes.
   859// (ANDQconst [0xFFFFFFFF] x) => (MOVLQZX x)
   860
   861// strength reduction
   862(MUL(Q|L)const [ 0] _) => (MOV(Q|L)const [0])
   863(MUL(Q|L)const [ 1] x) => x
   864(MULQconst [c] x) && canMulStrengthReduce(config, int64(c)) => {mulStrengthReduce(v, x, int64(c))}
   865(MULLconst [c] x) && v.Type.Size() <= 4 && canMulStrengthReduce32(config, c) => {mulStrengthReduce32(v, x, c)}
   866
   867// Prefer addition when shifting left by one
   868(SHL(Q|L)const [1] x) => (ADD(Q|L) x x)
   869
   870// combine add/shift into LEAQ/LEAL
   871(ADD(L|Q) x (SHL(L|Q)const [3] y)) => (LEA(L|Q)8 x y)
   872(ADD(L|Q) x (SHL(L|Q)const [2] y)) => (LEA(L|Q)4 x y)
   873(ADD(L|Q) x (ADD(L|Q) y y))        => (LEA(L|Q)2 x y)
   874(ADD(L|Q) x (ADD(L|Q) x y))        => (LEA(L|Q)2 y x)
   875
   876// combine ADDQ/ADDQconst into LEAQ1/LEAL1
   877(ADD(Q|L)const [c] (ADD(Q|L) x y)) => (LEA(Q|L)1 [c] x y)
   878(ADD(Q|L) (ADD(Q|L)const [c] x) y) => (LEA(Q|L)1 [c] x y)
   879(ADD(Q|L)const [c] (ADD(Q|L) x x)) => (LEA(Q|L)1 [c] x x)
   880
   881// fold ADDQ/ADDL into LEAQ/LEAL
   882(ADD(Q|L)const [c] (LEA(Q|L) [d] {s} x)) && is32Bit(int64(c)+int64(d)) => (LEA(Q|L) [c+d] {s} x)
   883(LEA(Q|L) [c] {s} (ADD(Q|L)const [d] x)) && is32Bit(int64(c)+int64(d)) => (LEA(Q|L) [c+d] {s} x)
   884(LEA(Q|L) [c] {s} (ADD(Q|L) x y)) && x.Op != OpSB && y.Op != OpSB => (LEA(Q|L)1 [c] {s} x y)
   885(ADD(Q|L) x (LEA(Q|L) [c] {s} y)) && x.Op != OpSB && y.Op != OpSB => (LEA(Q|L)1 [c] {s} x y)
   886
   887// fold ADDQconst/ADDLconst into LEAQx/LEALx
   888(ADD(Q|L)const [c] (LEA(Q|L)1 [d] {s} x y)) && is32Bit(int64(c)+int64(d)) => (LEA(Q|L)1 [c+d] {s} x y)
   889(ADD(Q|L)const [c] (LEA(Q|L)2 [d] {s} x y)) && is32Bit(int64(c)+int64(d)) => (LEA(Q|L)2 [c+d] {s} x y)
   890(ADD(Q|L)const [c] (LEA(Q|L)4 [d] {s} x y)) && is32Bit(int64(c)+int64(d)) => (LEA(Q|L)4 [c+d] {s} x y)
   891(ADD(Q|L)const [c] (LEA(Q|L)8 [d] {s} x y)) && is32Bit(int64(c)+int64(d)) => (LEA(Q|L)8 [c+d] {s} x y)
   892(LEA(Q|L)1 [c] {s} (ADD(Q|L)const [d] x) y) && is32Bit(int64(c)+int64(d))   && x.Op != OpSB => (LEA(Q|L)1 [c+d] {s} x y)
   893(LEA(Q|L)2 [c] {s} (ADD(Q|L)const [d] x) y) && is32Bit(int64(c)+int64(d))   && x.Op != OpSB => (LEA(Q|L)2 [c+d] {s} x y)
   894(LEA(Q|L)2 [c] {s} x (ADD(Q|L)const [d] y)) && is32Bit(int64(c)+2*int64(d)) && y.Op != OpSB => (LEA(Q|L)2 [c+2*d] {s} x y)
   895(LEA(Q|L)4 [c] {s} (ADD(Q|L)const [d] x) y) && is32Bit(int64(c)+int64(d))   && x.Op != OpSB => (LEA(Q|L)4 [c+d] {s} x y)
   896(LEA(Q|L)4 [c] {s} x (ADD(Q|L)const [d] y)) && is32Bit(int64(c)+4*int64(d)) && y.Op != OpSB => (LEA(Q|L)4 [c+4*d] {s} x y)
   897(LEA(Q|L)8 [c] {s} (ADD(Q|L)const [d] x) y) && is32Bit(int64(c)+int64(d))   && x.Op != OpSB => (LEA(Q|L)8 [c+d] {s} x y)
   898(LEA(Q|L)8 [c] {s} x (ADD(Q|L)const [d] y)) && is32Bit(int64(c)+8*int64(d)) && y.Op != OpSB => (LEA(Q|L)8 [c+8*d] {s} x y)
   899
   900// fold shifts into LEAQx/LEALx
   901(LEA(Q|L)1 [c] {s} x z:(ADD(Q|L) y y)) && x != z => (LEA(Q|L)2 [c] {s} x y)
   902(LEA(Q|L)1 [c] {s} x (SHL(Q|L)const [2] y)) => (LEA(Q|L)4 [c] {s} x y)
   903(LEA(Q|L)1 [c] {s} x (SHL(Q|L)const [3] y)) => (LEA(Q|L)8 [c] {s} x y)
   904(LEA(Q|L)2 [c] {s} x z:(ADD(Q|L) y y)) && x != z => (LEA(Q|L)4 [c] {s} x y)
   905(LEA(Q|L)2 [c] {s} x (SHL(Q|L)const [2] y)) => (LEA(Q|L)8 [c] {s} x y)
   906(LEA(Q|L)4 [c] {s} x z:(ADD(Q|L) y y)) && x != z => (LEA(Q|L)8 [c] {s} x y)
   907
   908// (x + x) << 1 -> x << 2
   909(LEA(Q|L)2 [0] {s} (ADD(Q|L) x x) x) && s == nil => (SHL(Q|L)const [2] x)
   910
   911// (x + x) << 2 -> x << 3 and similar
   912(SHL(Q|L)const [c] (ADD(Q|L) x x))  => (SHL(Q|L)const [c+1] x)
   913
   914// reverse ordering of compare instruction
   915(SETL (InvertFlags x)) => (SETG x)
   916(SETG (InvertFlags x)) => (SETL x)
   917(SETB (InvertFlags x)) => (SETA x)
   918(SETA (InvertFlags x)) => (SETB x)
   919(SETLE (InvertFlags x)) => (SETGE x)
   920(SETGE (InvertFlags x)) => (SETLE x)
   921(SETBE (InvertFlags x)) => (SETAE x)
   922(SETAE (InvertFlags x)) => (SETBE x)
   923(SETEQ (InvertFlags x)) => (SETEQ x)
   924(SETNE (InvertFlags x)) => (SETNE x)
   925
   926(SETLstore [off] {sym} ptr (InvertFlags x) mem) => (SETGstore [off] {sym} ptr x mem)
   927(SETGstore [off] {sym} ptr (InvertFlags x) mem) => (SETLstore [off] {sym} ptr x mem)
   928(SETBstore [off] {sym} ptr (InvertFlags x) mem) => (SETAstore [off] {sym} ptr x mem)
   929(SETAstore [off] {sym} ptr (InvertFlags x) mem) => (SETBstore [off] {sym} ptr x mem)
   930(SETLEstore [off] {sym} ptr (InvertFlags x) mem) => (SETGEstore [off] {sym} ptr x mem)
   931(SETGEstore [off] {sym} ptr (InvertFlags x) mem) => (SETLEstore [off] {sym} ptr x mem)
   932(SETBEstore [off] {sym} ptr (InvertFlags x) mem) => (SETAEstore [off] {sym} ptr x mem)
   933(SETAEstore [off] {sym} ptr (InvertFlags x) mem) => (SETBEstore [off] {sym} ptr x mem)
   934(SETEQstore [off] {sym} ptr (InvertFlags x) mem) => (SETEQstore [off] {sym} ptr x mem)
   935(SETNEstore [off] {sym} ptr (InvertFlags x) mem) => (SETNEstore [off] {sym} ptr x mem)
   936
   937// sign extended loads
   938// Note: The combined instruction must end up in the same block
   939// as the original load. If not, we end up making a value with
   940// memory type live in two different blocks, which can lead to
   941// multiple memory values alive simultaneously.
   942// Make sure we don't combine these ops if the load has another use.
   943// This prevents a single load from being split into multiple loads
   944// which then might return different values.  See test/atomicload.go.
   945(MOVBQSX x:(MOVBload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVBQSXload <v.Type> [off] {sym} ptr mem)
   946(MOVBQSX x:(MOVWload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVBQSXload <v.Type> [off] {sym} ptr mem)
   947(MOVBQSX x:(MOVLload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVBQSXload <v.Type> [off] {sym} ptr mem)
   948(MOVBQSX x:(MOVQload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVBQSXload <v.Type> [off] {sym} ptr mem)
   949(MOVBQZX x:(MOVBload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVBload <v.Type> [off] {sym} ptr mem)
   950(MOVBQZX x:(MOVWload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVBload <v.Type> [off] {sym} ptr mem)
   951(MOVBQZX x:(MOVLload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVBload <v.Type> [off] {sym} ptr mem)
   952(MOVBQZX x:(MOVQload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVBload <v.Type> [off] {sym} ptr mem)
   953(MOVWQSX x:(MOVWload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVWQSXload <v.Type> [off] {sym} ptr mem)
   954(MOVWQSX x:(MOVLload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVWQSXload <v.Type> [off] {sym} ptr mem)
   955(MOVWQSX x:(MOVQload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVWQSXload <v.Type> [off] {sym} ptr mem)
   956(MOVWQZX x:(MOVWload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVWload <v.Type> [off] {sym} ptr mem)
   957(MOVWQZX x:(MOVLload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVWload <v.Type> [off] {sym} ptr mem)
   958(MOVWQZX x:(MOVQload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVWload <v.Type> [off] {sym} ptr mem)
   959(MOVLQSX x:(MOVLload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVLQSXload <v.Type> [off] {sym} ptr mem)
   960(MOVLQSX x:(MOVQload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVLQSXload <v.Type> [off] {sym} ptr mem)
   961(MOVLQZX x:(MOVLload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVLload <v.Type> [off] {sym} ptr mem)
   962(MOVLQZX x:(MOVQload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVLload <v.Type> [off] {sym} ptr mem)
   963
   964// replace load from same location as preceding store with zero/sign extension (or copy in case of full width)
   965(MOVBload [off] {sym} ptr (MOVBstore [off2] {sym2} ptr2 x _)) && sym == sym2 && off == off2 && isSamePtr(ptr, ptr2) => (MOVBQZX x)
   966(MOVWload [off] {sym} ptr (MOVWstore [off2] {sym2} ptr2 x _)) && sym == sym2 && off == off2 && isSamePtr(ptr, ptr2) => (MOVWQZX x)
   967(MOVLload [off] {sym} ptr (MOVLstore [off2] {sym2} ptr2 x _)) && sym == sym2 && off == off2 && isSamePtr(ptr, ptr2) => (MOVLQZX x)
   968(MOVQload [off] {sym} ptr (MOVQstore [off2] {sym2} ptr2 x _)) && sym == sym2 && off == off2 && isSamePtr(ptr, ptr2) => x
   969(MOVBQSXload [off] {sym} ptr (MOVBstore [off2] {sym2} ptr2 x _)) && sym == sym2 && off == off2 && isSamePtr(ptr, ptr2) => (MOVBQSX x)
   970(MOVWQSXload [off] {sym} ptr (MOVWstore [off2] {sym2} ptr2 x _)) && sym == sym2 && off == off2 && isSamePtr(ptr, ptr2) => (MOVWQSX x)
   971(MOVLQSXload [off] {sym} ptr (MOVLstore [off2] {sym2} ptr2 x _)) && sym == sym2 && off == off2 && isSamePtr(ptr, ptr2) => (MOVLQSX x)
   972
   973// Fold extensions and ANDs together.
   974(MOVBQZX (ANDLconst [c] x)) => (ANDLconst [c & 0xff] x)
   975(MOVWQZX (ANDLconst [c] x)) => (ANDLconst [c & 0xffff] x)
   976(MOVLQZX (ANDLconst [c] x)) => (ANDLconst [c] x)
   977(MOVBQSX (ANDLconst [c] x)) && c & 0x80 == 0 => (ANDLconst [c & 0x7f] x)
   978(MOVWQSX (ANDLconst [c] x)) && c & 0x8000 == 0 => (ANDLconst [c & 0x7fff] x)
   979(MOVLQSX (ANDLconst [c] x)) && uint32(c) & 0x80000000 == 0 => (ANDLconst [c & 0x7fffffff] x)
   980
   981// Don't extend before storing
   982(MOVLstore [off] {sym} ptr (MOVLQSX x) mem) => (MOVLstore [off] {sym} ptr x mem)
   983(MOVWstore [off] {sym} ptr (MOVWQSX x) mem) => (MOVWstore [off] {sym} ptr x mem)
   984(MOVBstore [off] {sym} ptr (MOVBQSX x) mem) => (MOVBstore [off] {sym} ptr x mem)
   985(MOVLstore [off] {sym} ptr (MOVLQZX x) mem) => (MOVLstore [off] {sym} ptr x mem)
   986(MOVWstore [off] {sym} ptr (MOVWQZX x) mem) => (MOVWstore [off] {sym} ptr x mem)
   987(MOVBstore [off] {sym} ptr (MOVBQZX x) mem) => (MOVBstore [off] {sym} ptr x mem)
   988
   989// fold constants into memory operations
   990// Note that this is not always a good idea because if not all the uses of
   991// the ADDQconst get eliminated, we still have to compute the ADDQconst and we now
   992// have potentially two live values (ptr and (ADDQconst [off] ptr)) instead of one.
   993// Nevertheless, let's do it!
   994(MOV(Q|L|W|B|SS|SD|O)load  [off1] {sym} (ADDQconst [off2] ptr) mem) && is32Bit(int64(off1)+int64(off2)) =>
   995    (MOV(Q|L|W|B|SS|SD|O)load  [off1+off2] {sym} ptr mem)
   996(MOV(Q|L|W|B|SS|SD|O)store  [off1] {sym} (ADDQconst [off2] ptr) val mem) && is32Bit(int64(off1)+int64(off2)) =>
   997	(MOV(Q|L|W|B|SS|SD|O)store  [off1+off2] {sym} ptr val mem)
   998(SET(L|G|B|A|LE|GE|BE|AE|EQ|NE)store [off1] {sym} (ADDQconst [off2] base) val mem) && is32Bit(int64(off1)+int64(off2)) =>
   999	(SET(L|G|B|A|LE|GE|BE|AE|EQ|NE)store [off1+off2] {sym} base val mem)
  1000((ADD|SUB|AND|OR|XOR)Qload [off1] {sym} val (ADDQconst [off2] base) mem) && is32Bit(int64(off1)+int64(off2)) =>
  1001	((ADD|SUB|AND|OR|XOR)Qload [off1+off2] {sym} val base mem)
  1002((ADD|SUB|AND|OR|XOR)Lload [off1] {sym} val (ADDQconst [off2] base) mem) && is32Bit(int64(off1)+int64(off2)) =>
  1003	((ADD|SUB|AND|OR|XOR)Lload [off1+off2] {sym} val base mem)
  1004(CMP(Q|L|W|B)load [off1] {sym} (ADDQconst [off2] base) val mem) && is32Bit(int64(off1)+int64(off2)) =>
  1005	(CMP(Q|L|W|B)load [off1+off2] {sym} base val mem)
  1006(CMP(Q|L|W|B)constload [valoff1] {sym} (ADDQconst [off2] base) mem) && ValAndOff(valoff1).canAdd32(off2) =>
  1007	(CMP(Q|L|W|B)constload [ValAndOff(valoff1).addOffset32(off2)] {sym} base mem)
  1008
  1009((ADD|SUB|MUL|DIV)SSload [off1] {sym} val (ADDQconst [off2] base) mem) && is32Bit(int64(off1)+int64(off2)) =>
  1010	((ADD|SUB|MUL|DIV)SSload [off1+off2] {sym} val base mem)
  1011((ADD|SUB|MUL|DIV)SDload [off1] {sym} val (ADDQconst [off2] base) mem) && is32Bit(int64(off1)+int64(off2)) =>
  1012	((ADD|SUB|MUL|DIV)SDload [off1+off2] {sym} val base mem)
  1013((ADD|AND|OR|XOR)Qconstmodify [valoff1] {sym} (ADDQconst [off2] base) mem) && ValAndOff(valoff1).canAdd32(off2) =>
  1014	((ADD|AND|OR|XOR)Qconstmodify [ValAndOff(valoff1).addOffset32(off2)] {sym} base mem)
  1015((ADD|AND|OR|XOR)Lconstmodify [valoff1] {sym} (ADDQconst [off2] base) mem) && ValAndOff(valoff1).canAdd32(off2) =>
  1016	((ADD|AND|OR|XOR)Lconstmodify [ValAndOff(valoff1).addOffset32(off2)] {sym} base mem)
  1017((ADD|SUB|AND|OR|XOR)Qmodify [off1] {sym} (ADDQconst [off2] base) val mem) && is32Bit(int64(off1)+int64(off2)) =>
  1018	((ADD|SUB|AND|OR|XOR)Qmodify [off1+off2] {sym} base val mem)
  1019((ADD|SUB|AND|OR|XOR)Lmodify [off1] {sym} (ADDQconst [off2] base) val mem) && is32Bit(int64(off1)+int64(off2)) =>
  1020	((ADD|SUB|AND|OR|XOR)Lmodify [off1+off2] {sym} base val mem)
  1021
  1022// Fold constants into stores.
  1023(MOVQstore [off] {sym} ptr (MOVQconst [c]) mem) && validVal(c) =>
  1024	(MOVQstoreconst [makeValAndOff(int32(c),off)] {sym} ptr mem)
  1025(MOVLstore [off] {sym} ptr (MOV(L|Q)const [c]) mem) =>
  1026	(MOVLstoreconst [makeValAndOff(int32(c),off)] {sym} ptr mem)
  1027(MOVWstore [off] {sym} ptr (MOV(L|Q)const [c]) mem) =>
  1028	(MOVWstoreconst [makeValAndOff(int32(int16(c)),off)] {sym} ptr mem)
  1029(MOVBstore [off] {sym} ptr (MOV(L|Q)const [c]) mem) =>
  1030	(MOVBstoreconst [makeValAndOff(int32(int8(c)),off)] {sym} ptr mem)
  1031
  1032// Fold address offsets into constant stores.
  1033(MOV(Q|L|W|B|O)storeconst [sc] {s} (ADDQconst [off] ptr) mem) && ValAndOff(sc).canAdd32(off) =>
  1034	(MOV(Q|L|W|B|O)storeconst [ValAndOff(sc).addOffset32(off)] {s} ptr mem)
  1035
  1036// We need to fold LEAQ into the MOVx ops so that the live variable analysis knows
  1037// what variables are being read/written by the ops.
  1038(MOV(Q|L|W|B|SS|SD|O|BQSX|WQSX|LQSX)load [off1] {sym1} (LEAQ [off2] {sym2} base) mem)
  1039	&& is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1040	(MOV(Q|L|W|B|SS|SD|O|BQSX|WQSX|LQSX)load [off1+off2] {mergeSym(sym1,sym2)} base mem)
  1041(MOV(Q|L|W|B|SS|SD|O)store [off1] {sym1} (LEAQ [off2] {sym2} base) val mem)
  1042	&& is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1043	(MOV(Q|L|W|B|SS|SD|O)store [off1+off2] {mergeSym(sym1,sym2)} base val mem)
  1044(MOV(Q|L|W|B|O)storeconst [sc] {sym1} (LEAQ [off] {sym2} ptr) mem) && canMergeSym(sym1, sym2) && ValAndOff(sc).canAdd32(off) =>
  1045	(MOV(Q|L|W|B|O)storeconst [ValAndOff(sc).addOffset32(off)] {mergeSym(sym1, sym2)} ptr mem)
  1046(SET(L|G|B|A|LE|GE|BE|AE|EQ|NE)store [off1] {sym1} (LEAQ [off2] {sym2} base) val mem)
  1047	&& is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1048	(SET(L|G|B|A|LE|GE|BE|AE|EQ|NE)store [off1+off2] {mergeSym(sym1,sym2)} base val mem)
  1049((ADD|SUB|AND|OR|XOR)Qload [off1] {sym1} val (LEAQ [off2] {sym2} base) mem)
  1050	&& is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1051	((ADD|SUB|AND|OR|XOR)Qload [off1+off2] {mergeSym(sym1,sym2)} val base mem)
  1052((ADD|SUB|AND|OR|XOR)Lload [off1] {sym1} val (LEAQ [off2] {sym2} base) mem)
  1053	&& is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1054	((ADD|SUB|AND|OR|XOR)Lload [off1+off2] {mergeSym(sym1,sym2)} val base mem)
  1055(CMP(Q|L|W|B)load [off1] {sym1} (LEAQ [off2] {sym2} base) val mem)
  1056	&& is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1057	(CMP(Q|L|W|B)load [off1+off2] {mergeSym(sym1,sym2)} base val mem)
  1058(CMP(Q|L|W|B)constload [valoff1] {sym1} (LEAQ [off2] {sym2} base) mem)
  1059	&& ValAndOff(valoff1).canAdd32(off2) && canMergeSym(sym1, sym2) =>
  1060	(CMP(Q|L|W|B)constload [ValAndOff(valoff1).addOffset32(off2)] {mergeSym(sym1,sym2)} base mem)
  1061
  1062((ADD|SUB|MUL|DIV)SSload [off1] {sym1} val (LEAQ [off2] {sym2} base) mem)
  1063	&& is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1064	((ADD|SUB|MUL|DIV)SSload [off1+off2] {mergeSym(sym1,sym2)} val base mem)
  1065((ADD|SUB|MUL|DIV)SDload [off1] {sym1} val (LEAQ [off2] {sym2} base) mem)
  1066	&& is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1067	((ADD|SUB|MUL|DIV)SDload [off1+off2] {mergeSym(sym1,sym2)} val base mem)
  1068((ADD|AND|OR|XOR)Qconstmodify [valoff1] {sym1} (LEAQ [off2] {sym2} base) mem)
  1069	&& ValAndOff(valoff1).canAdd32(off2) && canMergeSym(sym1, sym2) =>
  1070	((ADD|AND|OR|XOR)Qconstmodify [ValAndOff(valoff1).addOffset32(off2)] {mergeSym(sym1,sym2)} base mem)
  1071((ADD|AND|OR|XOR)Lconstmodify [valoff1] {sym1} (LEAQ [off2] {sym2} base) mem)
  1072	&& ValAndOff(valoff1).canAdd32(off2) && canMergeSym(sym1, sym2) =>
  1073	((ADD|AND|OR|XOR)Lconstmodify [ValAndOff(valoff1).addOffset32(off2)] {mergeSym(sym1,sym2)} base mem)
  1074((ADD|SUB|AND|OR|XOR)Qmodify [off1] {sym1} (LEAQ [off2] {sym2} base) val mem)
  1075	&& is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1076	((ADD|SUB|AND|OR|XOR)Qmodify [off1+off2] {mergeSym(sym1,sym2)} base val mem)
  1077((ADD|SUB|AND|OR|XOR)Lmodify [off1] {sym1} (LEAQ [off2] {sym2} base) val mem)
  1078	&& is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1079	((ADD|SUB|AND|OR|XOR)Lmodify [off1+off2] {mergeSym(sym1,sym2)} base val mem)
  1080
  1081// fold LEAQs together
  1082(LEAQ [off1] {sym1} (LEAQ [off2] {sym2} x)) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1083      (LEAQ [off1+off2] {mergeSym(sym1,sym2)} x)
  1084
  1085// LEAQ into LEAQ1
  1086(LEAQ1 [off1] {sym1} (LEAQ [off2] {sym2} x) y) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) && x.Op != OpSB =>
  1087       (LEAQ1 [off1+off2] {mergeSym(sym1,sym2)} x y)
  1088
  1089// LEAQ1 into LEAQ
  1090(LEAQ [off1] {sym1} (LEAQ1 [off2] {sym2} x y)) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1091       (LEAQ1 [off1+off2] {mergeSym(sym1,sym2)} x y)
  1092
  1093// LEAQ into LEAQ[248]
  1094(LEAQ2 [off1] {sym1} (LEAQ [off2] {sym2} x) y) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) && x.Op != OpSB =>
  1095       (LEAQ2 [off1+off2] {mergeSym(sym1,sym2)} x y)
  1096(LEAQ4 [off1] {sym1} (LEAQ [off2] {sym2} x) y) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) && x.Op != OpSB =>
  1097       (LEAQ4 [off1+off2] {mergeSym(sym1,sym2)} x y)
  1098(LEAQ8 [off1] {sym1} (LEAQ [off2] {sym2} x) y) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) && x.Op != OpSB =>
  1099       (LEAQ8 [off1+off2] {mergeSym(sym1,sym2)} x y)
  1100
  1101// LEAQ[248] into LEAQ
  1102(LEAQ [off1] {sym1} (LEAQ2 [off2] {sym2} x y)) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1103      (LEAQ2 [off1+off2] {mergeSym(sym1,sym2)} x y)
  1104(LEAQ [off1] {sym1} (LEAQ4 [off2] {sym2} x y)) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1105      (LEAQ4 [off1+off2] {mergeSym(sym1,sym2)} x y)
  1106(LEAQ [off1] {sym1} (LEAQ8 [off2] {sym2} x y)) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1107      (LEAQ8 [off1+off2] {mergeSym(sym1,sym2)} x y)
  1108
  1109// LEAQ[1248] into LEAQ[1248]. Only some such merges are possible.
  1110(LEAQ1 [off1] {sym1} x (LEAQ1 [off2] {sym2} y y)) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1111      (LEAQ2 [off1+off2] {mergeSym(sym1, sym2)} x y)
  1112(LEAQ1 [off1] {sym1} x (LEAQ1 [off2] {sym2} x y)) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1113      (LEAQ2 [off1+off2] {mergeSym(sym1, sym2)} y x)
  1114(LEAQ2 [off1] {sym1} x (LEAQ1 [off2] {sym2} y y)) && is32Bit(int64(off1)+2*int64(off2)) && sym2 == nil =>
  1115      (LEAQ4 [off1+2*off2] {sym1} x y)
  1116(LEAQ4 [off1] {sym1} x (LEAQ1 [off2] {sym2} y y)) && is32Bit(int64(off1)+4*int64(off2)) && sym2 == nil =>
  1117      (LEAQ8 [off1+4*off2] {sym1} x y)
  1118// TODO: more?
  1119
  1120// Lower LEAQ2/4/8 when the offset is a constant
  1121(LEAQ2 [off] {sym} x (MOV(Q|L)const [scale])) && is32Bit(int64(off)+int64(scale)*2) =>
  1122	(LEAQ [off+int32(scale)*2] {sym} x)
  1123(LEAQ4 [off] {sym} x (MOV(Q|L)const [scale])) && is32Bit(int64(off)+int64(scale)*4) =>
  1124	(LEAQ [off+int32(scale)*4] {sym} x)
  1125(LEAQ8 [off] {sym} x (MOV(Q|L)const [scale])) && is32Bit(int64(off)+int64(scale)*8) =>
  1126	(LEAQ [off+int32(scale)*8] {sym} x)
  1127
  1128// Absorb InvertFlags into branches.
  1129(LT (InvertFlags cmp) yes no) => (GT cmp yes no)
  1130(GT (InvertFlags cmp) yes no) => (LT cmp yes no)
  1131(LE (InvertFlags cmp) yes no) => (GE cmp yes no)
  1132(GE (InvertFlags cmp) yes no) => (LE cmp yes no)
  1133(ULT (InvertFlags cmp) yes no) => (UGT cmp yes no)
  1134(UGT (InvertFlags cmp) yes no) => (ULT cmp yes no)
  1135(ULE (InvertFlags cmp) yes no) => (UGE cmp yes no)
  1136(UGE (InvertFlags cmp) yes no) => (ULE cmp yes no)
  1137(EQ (InvertFlags cmp) yes no) => (EQ cmp yes no)
  1138(NE (InvertFlags cmp) yes no) => (NE cmp yes no)
  1139
  1140// Constant comparisons.
  1141(CMPQconst (MOVQconst [x]) [y]) && x==int64(y) => (FlagEQ)
  1142(CMPQconst (MOVQconst [x]) [y]) && x<int64(y) && uint64(x)<uint64(int64(y)) => (FlagLT_ULT)
  1143(CMPQconst (MOVQconst [x]) [y]) && x<int64(y) && uint64(x)>uint64(int64(y)) => (FlagLT_UGT)
  1144(CMPQconst (MOVQconst [x]) [y]) && x>int64(y) && uint64(x)<uint64(int64(y)) => (FlagGT_ULT)
  1145(CMPQconst (MOVQconst [x]) [y]) && x>int64(y) && uint64(x)>uint64(int64(y)) => (FlagGT_UGT)
  1146(CMPLconst (MOVLconst [x]) [y]) && x==y => (FlagEQ)
  1147(CMPLconst (MOVLconst [x]) [y]) && x<y && uint32(x)<uint32(y) => (FlagLT_ULT)
  1148(CMPLconst (MOVLconst [x]) [y]) && x<y && uint32(x)>uint32(y) => (FlagLT_UGT)
  1149(CMPLconst (MOVLconst [x]) [y]) && x>y && uint32(x)<uint32(y) => (FlagGT_ULT)
  1150(CMPLconst (MOVLconst [x]) [y]) && x>y && uint32(x)>uint32(y) => (FlagGT_UGT)
  1151(CMPWconst (MOVLconst [x]) [y]) && int16(x)==y => (FlagEQ)
  1152(CMPWconst (MOVLconst [x]) [y]) && int16(x)<y && uint16(x)<uint16(y) => (FlagLT_ULT)
  1153(CMPWconst (MOVLconst [x]) [y]) && int16(x)<y && uint16(x)>uint16(y) => (FlagLT_UGT)
  1154(CMPWconst (MOVLconst [x]) [y]) && int16(x)>y && uint16(x)<uint16(y) => (FlagGT_ULT)
  1155(CMPWconst (MOVLconst [x]) [y]) && int16(x)>y && uint16(x)>uint16(y) => (FlagGT_UGT)
  1156(CMPBconst (MOVLconst [x]) [y]) && int8(x)==y => (FlagEQ)
  1157(CMPBconst (MOVLconst [x]) [y]) && int8(x)<y && uint8(x)<uint8(y) => (FlagLT_ULT)
  1158(CMPBconst (MOVLconst [x]) [y]) && int8(x)<y && uint8(x)>uint8(y) => (FlagLT_UGT)
  1159(CMPBconst (MOVLconst [x]) [y]) && int8(x)>y && uint8(x)<uint8(y) => (FlagGT_ULT)
  1160(CMPBconst (MOVLconst [x]) [y]) && int8(x)>y && uint8(x)>uint8(y) => (FlagGT_UGT)
  1161
  1162// CMPQconst requires a 32 bit const, but we can still constant-fold 64 bit consts.
  1163// In theory this applies to any of the simplifications above,
  1164// but CMPQ is the only one I've actually seen occur.
  1165(CMPQ (MOVQconst [x]) (MOVQconst [y])) && x==y => (FlagEQ)
  1166(CMPQ (MOVQconst [x]) (MOVQconst [y])) && x<y && uint64(x)<uint64(y) => (FlagLT_ULT)
  1167(CMPQ (MOVQconst [x]) (MOVQconst [y])) && x<y && uint64(x)>uint64(y) => (FlagLT_UGT)
  1168(CMPQ (MOVQconst [x]) (MOVQconst [y])) && x>y && uint64(x)<uint64(y) => (FlagGT_ULT)
  1169(CMPQ (MOVQconst [x]) (MOVQconst [y])) && x>y && uint64(x)>uint64(y) => (FlagGT_UGT)
  1170
  1171// Other known comparisons.
  1172(CMPQconst (MOVBQZX _) [c]) && 0xFF < c => (FlagLT_ULT)
  1173(CMPQconst (MOVWQZX _) [c]) && 0xFFFF < c => (FlagLT_ULT)
  1174(CMPLconst (SHRLconst _ [c]) [n]) && 0 <= n && 0 < c && c <= 32 && (1<<uint64(32-c)) <= uint64(n) => (FlagLT_ULT)
  1175(CMPQconst (SHRQconst _ [c]) [n]) && 0 <= n && 0 < c && c <= 64 && (1<<uint64(64-c)) <= uint64(n) => (FlagLT_ULT)
  1176(CMPQconst (ANDQconst _ [m]) [n]) && 0 <= m && m < n => (FlagLT_ULT)
  1177(CMPQconst (ANDLconst _ [m]) [n]) && 0 <= m && m < n => (FlagLT_ULT)
  1178(CMPLconst (ANDLconst _ [m]) [n]) && 0 <= m && m < n => (FlagLT_ULT)
  1179(CMPWconst (ANDLconst _ [m]) [n]) && 0 <= int16(m) && int16(m) < n => (FlagLT_ULT)
  1180(CMPBconst (ANDLconst _ [m]) [n]) && 0 <= int8(m)  && int8(m)  < n => (FlagLT_ULT)
  1181
  1182// TESTQ c c sets flags like CMPQ c 0.
  1183(TESTQconst [c] (MOVQconst [d])) && int64(c) == d && c == 0 => (FlagEQ)
  1184(TESTLconst [c] (MOVLconst [c])) && c == 0 => (FlagEQ)
  1185(TESTQconst [c] (MOVQconst [d])) && int64(c) == d && c < 0  => (FlagLT_UGT)
  1186(TESTLconst [c] (MOVLconst [c])) && c < 0  => (FlagLT_UGT)
  1187(TESTQconst [c] (MOVQconst [d])) && int64(c) == d && c > 0  => (FlagGT_UGT)
  1188(TESTLconst [c] (MOVLconst [c])) && c > 0  => (FlagGT_UGT)
  1189
  1190// TODO: DIVxU also.
  1191
  1192// Absorb flag constants into SBB ops.
  1193(SBBQcarrymask (FlagEQ))     => (MOVQconst [0])
  1194(SBBQcarrymask (FlagLT_ULT)) => (MOVQconst [-1])
  1195(SBBQcarrymask (FlagLT_UGT)) => (MOVQconst [0])
  1196(SBBQcarrymask (FlagGT_ULT)) => (MOVQconst [-1])
  1197(SBBQcarrymask (FlagGT_UGT)) => (MOVQconst [0])
  1198(SBBLcarrymask (FlagEQ))     => (MOVLconst [0])
  1199(SBBLcarrymask (FlagLT_ULT)) => (MOVLconst [-1])
  1200(SBBLcarrymask (FlagLT_UGT)) => (MOVLconst [0])
  1201(SBBLcarrymask (FlagGT_ULT)) => (MOVLconst [-1])
  1202(SBBLcarrymask (FlagGT_UGT)) => (MOVLconst [0])
  1203
  1204// Absorb flag constants into branches.
  1205((EQ|LE|GE|ULE|UGE) (FlagEQ) yes no)     => (First yes no)
  1206((NE|LT|GT|ULT|UGT) (FlagEQ) yes no)     => (First no yes)
  1207((NE|LT|LE|ULT|ULE) (FlagLT_ULT) yes no) => (First yes no)
  1208((EQ|GT|GE|UGT|UGE) (FlagLT_ULT) yes no) => (First no yes)
  1209((NE|LT|LE|UGT|UGE) (FlagLT_UGT) yes no) => (First yes no)
  1210((EQ|GT|GE|ULT|ULE) (FlagLT_UGT) yes no) => (First no yes)
  1211((NE|GT|GE|ULT|ULE) (FlagGT_ULT) yes no) => (First yes no)
  1212((EQ|LT|LE|UGT|UGE) (FlagGT_ULT) yes no) => (First no yes)
  1213((NE|GT|GE|UGT|UGE) (FlagGT_UGT) yes no) => (First yes no)
  1214((EQ|LT|LE|ULT|ULE) (FlagGT_UGT) yes no) => (First no yes)
  1215
  1216// Absorb flag constants into SETxx ops.
  1217((SETEQ|SETLE|SETGE|SETBE|SETAE) (FlagEQ))     => (MOVLconst [1])
  1218((SETNE|SETL|SETG|SETB|SETA)     (FlagEQ))     => (MOVLconst [0])
  1219((SETNE|SETL|SETLE|SETB|SETBE)   (FlagLT_ULT)) => (MOVLconst [1])
  1220((SETEQ|SETG|SETGE|SETA|SETAE)   (FlagLT_ULT)) => (MOVLconst [0])
  1221((SETNE|SETL|SETLE|SETA|SETAE)   (FlagLT_UGT)) => (MOVLconst [1])
  1222((SETEQ|SETG|SETGE|SETB|SETBE)   (FlagLT_UGT)) => (MOVLconst [0])
  1223((SETNE|SETG|SETGE|SETB|SETBE)   (FlagGT_ULT)) => (MOVLconst [1])
  1224((SETEQ|SETL|SETLE|SETA|SETAE)   (FlagGT_ULT)) => (MOVLconst [0])
  1225((SETNE|SETG|SETGE|SETA|SETAE)   (FlagGT_UGT)) => (MOVLconst [1])
  1226((SETEQ|SETL|SETLE|SETB|SETBE)   (FlagGT_UGT)) => (MOVLconst [0])
  1227
  1228(SETEQstore [off] {sym} ptr (FlagEQ)     mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1229(SETEQstore [off] {sym} ptr (FlagLT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1230(SETEQstore [off] {sym} ptr (FlagLT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1231(SETEQstore [off] {sym} ptr (FlagGT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1232(SETEQstore [off] {sym} ptr (FlagGT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1233
  1234(SETNEstore [off] {sym} ptr (FlagEQ)     mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1235(SETNEstore [off] {sym} ptr (FlagLT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1236(SETNEstore [off] {sym} ptr (FlagLT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1237(SETNEstore [off] {sym} ptr (FlagGT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1238(SETNEstore [off] {sym} ptr (FlagGT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1239
  1240(SETLstore  [off] {sym} ptr (FlagEQ)     mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1241(SETLstore  [off] {sym} ptr (FlagLT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1242(SETLstore  [off] {sym} ptr (FlagLT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1243(SETLstore  [off] {sym} ptr (FlagGT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1244(SETLstore  [off] {sym} ptr (FlagGT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1245
  1246(SETLEstore [off] {sym} ptr (FlagEQ)     mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1247(SETLEstore [off] {sym} ptr (FlagLT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1248(SETLEstore [off] {sym} ptr (FlagLT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1249(SETLEstore [off] {sym} ptr (FlagGT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1250(SETLEstore [off] {sym} ptr (FlagGT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1251
  1252(SETGstore  [off] {sym} ptr (FlagEQ)     mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1253(SETGstore  [off] {sym} ptr (FlagLT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1254(SETGstore  [off] {sym} ptr (FlagLT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1255(SETGstore  [off] {sym} ptr (FlagGT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1256(SETGstore  [off] {sym} ptr (FlagGT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1257
  1258(SETGEstore [off] {sym} ptr (FlagEQ)     mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1259(SETGEstore [off] {sym} ptr (FlagLT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1260(SETGEstore [off] {sym} ptr (FlagLT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1261(SETGEstore [off] {sym} ptr (FlagGT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1262(SETGEstore [off] {sym} ptr (FlagGT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1263
  1264(SETBstore  [off] {sym} ptr (FlagEQ)     mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1265(SETBstore  [off] {sym} ptr (FlagLT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1266(SETBstore  [off] {sym} ptr (FlagLT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1267(SETBstore  [off] {sym} ptr (FlagGT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1268(SETBstore  [off] {sym} ptr (FlagGT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1269
  1270(SETBEstore [off] {sym} ptr (FlagEQ)     mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1271(SETBEstore [off] {sym} ptr (FlagLT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1272(SETBEstore [off] {sym} ptr (FlagLT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1273(SETBEstore [off] {sym} ptr (FlagGT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1274(SETBEstore [off] {sym} ptr (FlagGT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1275
  1276(SETAstore  [off] {sym} ptr (FlagEQ)     mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1277(SETAstore  [off] {sym} ptr (FlagLT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1278(SETAstore  [off] {sym} ptr (FlagLT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1279(SETAstore  [off] {sym} ptr (FlagGT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1280(SETAstore  [off] {sym} ptr (FlagGT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1281
  1282(SETAEstore [off] {sym} ptr (FlagEQ)     mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1283(SETAEstore [off] {sym} ptr (FlagLT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1284(SETAEstore [off] {sym} ptr (FlagLT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1285(SETAEstore [off] {sym} ptr (FlagGT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1286(SETAEstore [off] {sym} ptr (FlagGT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1287
  1288// Remove redundant *const ops
  1289(ADDQconst [0] x)          => x
  1290(ADDLconst [c] x) && c==0  => x
  1291(SUBQconst [0] x)          => x
  1292(SUBLconst [c] x) && c==0  => x
  1293(ANDQconst [0] _)          => (MOVQconst [0])
  1294(ANDLconst [c] _) && c==0  => (MOVLconst [0])
  1295(ANDQconst [-1] x)         => x
  1296(ANDLconst [c] x) && c==-1 => x
  1297(ORQconst [0] x)           => x
  1298(ORLconst [c] x)  && c==0  => x
  1299(ORQconst [-1] _)          => (MOVQconst [-1])
  1300(ORLconst [c] _)  && c==-1 => (MOVLconst [-1])
  1301(XORQconst [0] x)          => x
  1302(XORLconst [c] x) && c==0  => x
  1303// TODO: since we got rid of the W/B versions, we might miss
  1304// things like (ANDLconst [0x100] x) which were formerly
  1305// (ANDBconst [0] x).  Probably doesn't happen very often.
  1306// If we cared, we might do:
  1307//  (ANDLconst <t> [c] x) && t.Size()==1 && int8(x)==0 -> (MOVLconst [0])
  1308
  1309// Remove redundant ops
  1310// Not in generic rules, because they may appear after lowering e. g. Slicemask
  1311(NEG(Q|L) (NEG(Q|L) x)) => x
  1312(NEG(Q|L) s:(SUB(Q|L) x y)) && s.Uses == 1 => (SUB(Q|L) y x)
  1313
  1314// Convert constant subtracts to constant adds
  1315(SUBQconst [c] x) && c != -(1<<31) => (ADDQconst [-c] x)
  1316(SUBLconst [c] x) => (ADDLconst [-c] x)
  1317
  1318// generic constant folding
  1319// TODO: more of this
  1320(ADDQconst [c] (MOVQconst [d])) => (MOVQconst [int64(c)+d])
  1321(ADDLconst [c] (MOVLconst [d])) => (MOVLconst [c+d])
  1322(ADDQconst [c] (ADDQconst [d] x)) && is32Bit(int64(c)+int64(d)) => (ADDQconst [c+d] x)
  1323(ADDLconst [c] (ADDLconst [d] x)) => (ADDLconst [c+d] x)
  1324(SUBQconst (MOVQconst [d]) [c]) => (MOVQconst [d-int64(c)])
  1325(SUBQconst (SUBQconst x [d]) [c]) && is32Bit(int64(-c)-int64(d)) => (ADDQconst [-c-d] x)
  1326(SARQconst [c] (MOVQconst [d])) => (MOVQconst [d>>uint64(c)])
  1327(SARLconst [c] (MOVQconst [d])) => (MOVQconst [int64(int32(d))>>uint64(c)])
  1328(SARWconst [c] (MOVQconst [d])) => (MOVQconst [int64(int16(d))>>uint64(c)])
  1329(SARBconst [c] (MOVQconst [d])) => (MOVQconst [int64(int8(d))>>uint64(c)])
  1330(NEGQ (MOVQconst [c])) => (MOVQconst [-c])
  1331(NEGL (MOVLconst [c])) => (MOVLconst [-c])
  1332(MULQconst [c] (MOVQconst [d])) => (MOVQconst [int64(c)*d])
  1333(MULLconst [c] (MOVLconst [d])) => (MOVLconst [c*d])
  1334(ANDQconst [c] (MOVQconst [d])) => (MOVQconst [int64(c)&d])
  1335(ANDLconst [c] (MOVLconst [d])) => (MOVLconst [c&d])
  1336(ORQconst [c] (MOVQconst [d])) => (MOVQconst [int64(c)|d])
  1337(ORLconst [c] (MOVLconst [d])) => (MOVLconst [c|d])
  1338(XORQconst [c] (MOVQconst [d])) => (MOVQconst [int64(c)^d])
  1339(XORLconst [c] (MOVLconst [d])) => (MOVLconst [c^d])
  1340(NOTQ (MOVQconst [c])) => (MOVQconst [^c])
  1341(NOTL (MOVLconst [c])) => (MOVLconst [^c])
  1342(BTSQconst [c] (MOVQconst [d])) => (MOVQconst [d|(1<<uint32(c))])
  1343(BTRQconst [c] (MOVQconst [d])) => (MOVQconst [d&^(1<<uint32(c))])
  1344(BTCQconst [c] (MOVQconst [d])) => (MOVQconst [d^(1<<uint32(c))])
  1345
  1346// If c or d doesn't fit into 32 bits, then we can't construct ORQconst,
  1347// but we can still constant-fold.
  1348// In theory this applies to any of the simplifications above,
  1349// but ORQ is the only one I've actually seen occur.
  1350(ORQ (MOVQconst [c]) (MOVQconst [d])) => (MOVQconst [c|d])
  1351
  1352// generic simplifications
  1353// TODO: more of this
  1354(ADDQ x (NEGQ y)) => (SUBQ x y)
  1355(ADDL x (NEGL y)) => (SUBL x y)
  1356(SUBQ x x) => (MOVQconst [0])
  1357(SUBL x x) => (MOVLconst [0])
  1358(ANDQ x x) => x
  1359(ANDL x x) => x
  1360(ORQ x x)  => x
  1361(ORL x x)  => x
  1362(XORQ x x) => (MOVQconst [0])
  1363(XORL x x) => (MOVLconst [0])
  1364
  1365(SHLLconst [d] (MOVLconst [c])) => (MOVLconst [c << uint64(d)])
  1366(SHLQconst [d] (MOVQconst [c])) => (MOVQconst [c << uint64(d)])
  1367(SHLQconst [d] (MOVLconst [c])) => (MOVQconst [int64(c) << uint64(d)])
  1368
  1369// Fold NEG into ADDconst/MULconst. Take care to keep c in 32 bit range.
  1370(NEGQ (ADDQconst [c] (NEGQ x))) && c != -(1<<31) => (ADDQconst [-c] x)
  1371(MULQconst [c] (NEGQ x)) && c != -(1<<31) => (MULQconst [-c] x)
  1372
  1373// checking AND against 0.
  1374(CMPQconst a:(ANDQ x y) [0]) && a.Uses == 1 => (TESTQ x y)
  1375(CMPLconst a:(ANDL x y) [0]) && a.Uses == 1 => (TESTL x y)
  1376(CMPWconst a:(ANDL x y) [0]) && a.Uses == 1 => (TESTW x y)
  1377(CMPBconst a:(ANDL x y) [0]) && a.Uses == 1 => (TESTB x y)
  1378(CMPQconst a:(ANDQconst [c] x) [0]) && a.Uses == 1 => (TESTQconst [c] x)
  1379(CMPLconst a:(ANDLconst [c] x) [0]) && a.Uses == 1 => (TESTLconst [c] x)
  1380(CMPWconst a:(ANDLconst [c] x) [0]) && a.Uses == 1 => (TESTWconst [int16(c)] x)
  1381(CMPBconst a:(ANDLconst [c] x) [0]) && a.Uses == 1 => (TESTBconst [int8(c)] x)
  1382
  1383// Convert TESTx to TESTxconst if possible.
  1384(TESTQ (MOVQconst [c]) x) && is32Bit(c) => (TESTQconst [int32(c)] x)
  1385(TESTL (MOVLconst [c]) x) => (TESTLconst [c] x)
  1386(TESTW (MOVLconst [c]) x) => (TESTWconst [int16(c)] x)
  1387(TESTB (MOVLconst [c]) x) => (TESTBconst [int8(c)] x)
  1388
  1389// TEST %reg,%reg is shorter than CMP
  1390(CMPQconst x [0]) => (TESTQ x x)
  1391(CMPLconst x [0]) => (TESTL x x)
  1392(CMPWconst x [0]) => (TESTW x x)
  1393(CMPBconst x [0]) => (TESTB x x)
  1394(TESTQconst [-1] x) && x.Op != OpAMD64MOVQconst => (TESTQ x x)
  1395(TESTLconst [-1] x) && x.Op != OpAMD64MOVLconst => (TESTL x x)
  1396(TESTWconst [-1] x) && x.Op != OpAMD64MOVLconst => (TESTW x x)
  1397(TESTBconst [-1] x) && x.Op != OpAMD64MOVLconst => (TESTB x x)
  1398
  1399// Convert LEAQ1 back to ADDQ if we can
  1400(LEAQ1 [0] x y) && v.Aux == nil => (ADDQ x y)
  1401
  1402(MOVQstoreconst [c] {s} p1 x:(MOVQstoreconst [a] {s} p0 mem))
  1403  && x.Uses == 1
  1404  && sequentialAddresses(p0, p1, int64(a.Off()+8-c.Off()))
  1405  && a.Val() == 0
  1406  && c.Val() == 0
  1407  && setPos(v, x.Pos)
  1408  && clobber(x)
  1409  => (MOVOstoreconst [makeValAndOff(0,a.Off())] {s} p0 mem)
  1410(MOVQstoreconst [a] {s} p0 x:(MOVQstoreconst [c] {s} p1 mem))
  1411  && x.Uses == 1
  1412  && sequentialAddresses(p0, p1, int64(a.Off()+8-c.Off()))
  1413  && a.Val() == 0
  1414  && c.Val() == 0
  1415  && setPos(v, x.Pos)
  1416  && clobber(x)
  1417  => (MOVOstoreconst [makeValAndOff(0,a.Off())] {s} p0 mem)
  1418
  1419// Merge load and op
  1420// TODO: add indexed variants?
  1421((ADD|SUB|AND|OR|XOR)Q x l:(MOVQload [off] {sym} ptr mem)) && canMergeLoadClobber(v, l, x) && clobber(l) => ((ADD|SUB|AND|OR|XOR)Qload x [off] {sym} ptr mem)
  1422((ADD|SUB|AND|OR|XOR)L x l:(MOVLload [off] {sym} ptr mem)) && canMergeLoadClobber(v, l, x) && clobber(l) => ((ADD|SUB|AND|OR|XOR)Lload x [off] {sym} ptr mem)
  1423((ADD|SUB|MUL|DIV)SD x l:(MOVSDload [off] {sym} ptr mem)) && canMergeLoadClobber(v, l, x) && clobber(l) => ((ADD|SUB|MUL|DIV)SDload x [off] {sym} ptr mem)
  1424((ADD|SUB|MUL|DIV)SS x l:(MOVSSload [off] {sym} ptr mem)) && canMergeLoadClobber(v, l, x) && clobber(l) => ((ADD|SUB|MUL|DIV)SSload x [off] {sym} ptr mem)
  1425(MOVLstore {sym} [off] ptr y:((ADD|AND|OR|XOR)Lload x [off] {sym} ptr mem) mem) && y.Uses==1 && clobber(y) => ((ADD|AND|OR|XOR)Lmodify [off] {sym} ptr x mem)
  1426(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 && clobber(y, l) =>
  1427	((ADD|SUB|AND|OR|XOR)Lmodify [off] {sym} ptr x mem)
  1428(MOVQstore {sym} [off] ptr y:((ADD|AND|OR|XOR)Qload x [off] {sym} ptr mem) mem) && y.Uses==1 && clobber(y) => ((ADD|AND|OR|XOR)Qmodify [off] {sym} ptr x mem)
  1429(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 && clobber(y, l) =>
  1430	((ADD|SUB|AND|OR|XOR)Qmodify [off] {sym} ptr x mem)
  1431(MOVQstore {sym} [off] ptr x:(BT(S|R|C)Qconst [c] l:(MOVQload {sym} [off] ptr mem)) mem) && x.Uses == 1 && l.Uses == 1 && clobber(x, l) =>
  1432	(BT(S|R|C)Qconstmodify {sym} [makeValAndOff(int32(c),off)] ptr mem)
  1433
  1434// Merge ADDQconst and LEAQ into atomic loads.
  1435(MOV(Q|L|B)atomicload [off1] {sym} (ADDQconst [off2] ptr) mem) && is32Bit(int64(off1)+int64(off2)) =>
  1436	(MOV(Q|L|B)atomicload [off1+off2] {sym} ptr mem)
  1437(MOV(Q|L|B)atomicload [off1] {sym1} (LEAQ [off2] {sym2} ptr) mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1438	(MOV(Q|L|B)atomicload [off1+off2] {mergeSym(sym1, sym2)} ptr mem)
  1439
  1440// Merge ADDQconst and LEAQ into atomic stores.
  1441(XCHGQ [off1] {sym} val (ADDQconst [off2] ptr) mem) && is32Bit(int64(off1)+int64(off2)) =>
  1442	(XCHGQ [off1+off2] {sym} val ptr mem)
  1443(XCHGQ [off1] {sym1} val (LEAQ [off2] {sym2} ptr) mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) && ptr.Op != OpSB =>
  1444	(XCHGQ [off1+off2] {mergeSym(sym1,sym2)} val ptr mem)
  1445(XCHGL [off1] {sym} val (ADDQconst [off2] ptr) mem) && is32Bit(int64(off1)+int64(off2)) =>
  1446	(XCHGL [off1+off2] {sym} val ptr mem)
  1447(XCHGL [off1] {sym1} val (LEAQ [off2] {sym2} ptr) mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) && ptr.Op != OpSB =>
  1448	(XCHGL [off1+off2] {mergeSym(sym1,sym2)} val ptr mem)
  1449
  1450// Merge ADDQconst into atomic adds.
  1451// TODO: merging LEAQ doesn't work, assembler doesn't like the resulting instructions.
  1452(XADDQlock [off1] {sym} val (ADDQconst [off2] ptr) mem) && is32Bit(int64(off1)+int64(off2)) =>
  1453	(XADDQlock [off1+off2] {sym} val ptr mem)
  1454(XADDLlock [off1] {sym} val (ADDQconst [off2] ptr) mem) && is32Bit(int64(off1)+int64(off2)) =>
  1455	(XADDLlock [off1+off2] {sym} val ptr mem)
  1456
  1457// Merge ADDQconst into atomic compare and swaps.
  1458// TODO: merging LEAQ doesn't work, assembler doesn't like the resulting instructions.
  1459(CMPXCHGQlock [off1] {sym} (ADDQconst [off2] ptr) old new_ mem) && is32Bit(int64(off1)+int64(off2)) =>
  1460	(CMPXCHGQlock [off1+off2] {sym} ptr old new_ mem)
  1461(CMPXCHGLlock [off1] {sym} (ADDQconst [off2] ptr) old new_ mem) && is32Bit(int64(off1)+int64(off2)) =>
  1462	(CMPXCHGLlock [off1+off2] {sym} ptr old new_ mem)
  1463
  1464// We don't need the conditional move if we know the arg of BSF is not zero.
  1465(CMOVQEQ x _ (Select1 (BS(F|R)Q (ORQconst [c] _)))) && c != 0 => x
  1466// Extension is unnecessary for trailing zeros.
  1467(BSFQ (ORQconst <t> [1<<8] (MOVBQZX x))) => (BSFQ (ORQconst <t> [1<<8] x))
  1468(BSFQ (ORQconst <t> [1<<16] (MOVWQZX x))) => (BSFQ (ORQconst <t> [1<<16] x))
  1469
  1470// Redundant sign/zero extensions
  1471// Note: see issue 21963. We have to make sure we use the right type on
  1472// the resulting extension (the outer type, not the inner type).
  1473(MOVLQSX (MOVLQSX x)) => (MOVLQSX x)
  1474(MOVLQSX (MOVWQSX x)) => (MOVWQSX x)
  1475(MOVLQSX (MOVBQSX x)) => (MOVBQSX x)
  1476(MOVWQSX (MOVWQSX x)) => (MOVWQSX x)
  1477(MOVWQSX (MOVBQSX x)) => (MOVBQSX x)
  1478(MOVBQSX (MOVBQSX x)) => (MOVBQSX x)
  1479(MOVLQZX (MOVLQZX x)) => (MOVLQZX x)
  1480(MOVLQZX (MOVWQZX x)) => (MOVWQZX x)
  1481(MOVLQZX (MOVBQZX x)) => (MOVBQZX x)
  1482(MOVWQZX (MOVWQZX x)) => (MOVWQZX x)
  1483(MOVWQZX (MOVBQZX x)) => (MOVBQZX x)
  1484(MOVBQZX (MOVBQZX x)) => (MOVBQZX x)
  1485
  1486(MOVQstore [off] {sym} ptr a:((ADD|AND|OR|XOR)Qconst [c] l:(MOVQload [off] {sym} ptr2 mem)) mem)
  1487	&& isSamePtr(ptr, ptr2) && a.Uses == 1 && l.Uses == 1 && clobber(l, a) =>
  1488	((ADD|AND|OR|XOR)Qconstmodify {sym} [makeValAndOff(int32(c),off)] ptr mem)
  1489(MOVLstore [off] {sym} ptr a:((ADD|AND|OR|XOR)Lconst [c] l:(MOVLload [off] {sym} ptr2 mem)) mem)
  1490	&& isSamePtr(ptr, ptr2) && a.Uses == 1 && l.Uses == 1 && clobber(l, a) =>
  1491	((ADD|AND|OR|XOR)Lconstmodify {sym} [makeValAndOff(int32(c),off)] ptr mem)
  1492
  1493// float <-> int register moves, with no conversion.
  1494// These come up when compiling math.{Float{32,64}bits,Float{32,64}frombits}.
  1495(MOVQload  [off] {sym} ptr (MOVSDstore [off] {sym} ptr val _)) => (MOVQf2i val)
  1496(MOVLload  [off] {sym} ptr (MOVSSstore [off] {sym} ptr val _)) => (MOVLf2i val)
  1497(MOVSDload [off] {sym} ptr (MOVQstore  [off] {sym} ptr val _)) => (MOVQi2f val)
  1498(MOVSSload [off] {sym} ptr (MOVLstore  [off] {sym} ptr val _)) => (MOVLi2f val)
  1499
  1500// Other load-like ops.
  1501(ADDQload x [off] {sym} ptr (MOVSDstore [off] {sym} ptr y _)) => (ADDQ x (MOVQf2i y))
  1502(ADDLload x [off] {sym} ptr (MOVSSstore [off] {sym} ptr y _)) => (ADDL x (MOVLf2i y))
  1503(SUBQload x [off] {sym} ptr (MOVSDstore [off] {sym} ptr y _)) => (SUBQ x (MOVQf2i y))
  1504(SUBLload x [off] {sym} ptr (MOVSSstore [off] {sym} ptr y _)) => (SUBL x (MOVLf2i y))
  1505(ANDQload x [off] {sym} ptr (MOVSDstore [off] {sym} ptr y _)) => (ANDQ x (MOVQf2i y))
  1506(ANDLload x [off] {sym} ptr (MOVSSstore [off] {sym} ptr y _)) => (ANDL x (MOVLf2i y))
  1507( ORQload x [off] {sym} ptr (MOVSDstore [off] {sym} ptr y _)) => ( ORQ x (MOVQf2i y))
  1508( ORLload x [off] {sym} ptr (MOVSSstore [off] {sym} ptr y _)) => ( ORL x (MOVLf2i y))
  1509(XORQload x [off] {sym} ptr (MOVSDstore [off] {sym} ptr y _)) => (XORQ x (MOVQf2i y))
  1510(XORLload x [off] {sym} ptr (MOVSSstore [off] {sym} ptr y _)) => (XORL x (MOVLf2i y))
  1511
  1512(ADDSDload x [off] {sym} ptr (MOVQstore [off] {sym} ptr y _)) => (ADDSD x (MOVQi2f y))
  1513(ADDSSload x [off] {sym} ptr (MOVLstore [off] {sym} ptr y _)) => (ADDSS x (MOVLi2f y))
  1514(SUBSDload x [off] {sym} ptr (MOVQstore [off] {sym} ptr y _)) => (SUBSD x (MOVQi2f y))
  1515(SUBSSload x [off] {sym} ptr (MOVLstore [off] {sym} ptr y _)) => (SUBSS x (MOVLi2f y))
  1516(MULSDload x [off] {sym} ptr (MOVQstore [off] {sym} ptr y _)) => (MULSD x (MOVQi2f y))
  1517(MULSSload x [off] {sym} ptr (MOVLstore [off] {sym} ptr y _)) => (MULSS x (MOVLi2f y))
  1518
  1519// Detect FMA
  1520(ADDS(S|D) (MULS(S|D) x y) z) && buildcfg.GOAMD64 >= 3 && z.Block.Func.useFMA(v) => (VFMADD231S(S|D) z x y)
  1521
  1522// Redirect stores to use the other register set.
  1523(MOVQstore  [off] {sym} ptr (MOVQf2i val) mem) => (MOVSDstore [off] {sym} ptr val mem)
  1524(MOVLstore  [off] {sym} ptr (MOVLf2i val) mem) => (MOVSSstore [off] {sym} ptr val mem)
  1525(MOVSDstore [off] {sym} ptr (MOVQi2f val) mem) => (MOVQstore  [off] {sym} ptr val mem)
  1526(MOVSSstore [off] {sym} ptr (MOVLi2f val) mem) => (MOVLstore  [off] {sym} ptr val mem)
  1527
  1528(MOVSDstore [off] {sym} ptr (MOVSDconst [f]) mem) && f == f => (MOVQstore [off] {sym} ptr (MOVQconst [int64(math.Float64bits(f))]) mem)
  1529(MOVSSstore [off] {sym} ptr (MOVSSconst [f]) mem) && f == f => (MOVLstore [off] {sym} ptr (MOVLconst [int32(math.Float32bits(f))]) mem)
  1530
  1531// Load args directly into the register class where it will be used.
  1532// We do this by just modifying the type of the Arg.
  1533(MOVQf2i <t> (Arg <u> [off] {sym})) && t.Size() == u.Size() => @b.Func.Entry (Arg <t> [off] {sym})
  1534(MOVLf2i <t> (Arg <u> [off] {sym})) && t.Size() == u.Size() => @b.Func.Entry (Arg <t> [off] {sym})
  1535(MOVQi2f <t> (Arg <u> [off] {sym})) && t.Size() == u.Size() => @b.Func.Entry (Arg <t> [off] {sym})
  1536(MOVLi2f <t> (Arg <u> [off] {sym})) && t.Size() == u.Size() => @b.Func.Entry (Arg <t> [off] {sym})
  1537
  1538// LEAQ is rematerializeable, so this helps to avoid register spill.
  1539// See issue 22947 for details
  1540(ADD(Q|L)const [off] x:(SP)) => (LEA(Q|L) [off] x)
  1541
  1542// HMULx is commutative, but its first argument must go in AX.
  1543// If possible, put a rematerializeable value in the first argument slot,
  1544// to reduce the odds that another value will be have to spilled
  1545// specifically to free up AX.
  1546(HMUL(Q|L)  x y) && !x.rematerializeable() && y.rematerializeable() => (HMUL(Q|L)  y x)
  1547(HMUL(Q|L)U x y) && !x.rematerializeable() && y.rematerializeable() => (HMUL(Q|L)U y x)
  1548
  1549// Fold loads into compares
  1550// Note: these may be undone by the flagalloc pass.
  1551(CMP(Q|L|W|B) l:(MOV(Q|L|W|B)load {sym} [off] ptr mem) x) && canMergeLoad(v, l) && clobber(l) => (CMP(Q|L|W|B)load {sym} [off] ptr x mem)
  1552(CMP(Q|L|W|B) x l:(MOV(Q|L|W|B)load {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (InvertFlags (CMP(Q|L|W|B)load {sym} [off] ptr x mem))
  1553
  1554(CMP(Q|L)const l:(MOV(Q|L)load {sym} [off] ptr mem) [c])
  1555	&& l.Uses == 1
  1556	&& clobber(l) =>
  1557@l.Block (CMP(Q|L)constload {sym} [makeValAndOff(c,off)] ptr mem)
  1558(CMP(W|B)const l:(MOV(W|B)load {sym} [off] ptr mem) [c])
  1559	&& l.Uses == 1
  1560	&& clobber(l) =>
  1561@l.Block (CMP(W|B)constload {sym} [makeValAndOff(int32(c),off)] ptr mem)
  1562
  1563(CMPQload {sym} [off] ptr (MOVQconst [c]) mem) && validVal(c) => (CMPQconstload {sym} [makeValAndOff(int32(c),off)] ptr mem)
  1564(CMPLload {sym} [off] ptr (MOVLconst [c]) mem) => (CMPLconstload {sym} [makeValAndOff(c,off)] ptr mem)
  1565(CMPWload {sym} [off] ptr (MOVLconst [c]) mem) => (CMPWconstload {sym} [makeValAndOff(int32(int16(c)),off)] ptr mem)
  1566(CMPBload {sym} [off] ptr (MOVLconst [c]) mem) => (CMPBconstload {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
  1567
  1568(TEST(Q|L|W|B)  l:(MOV(Q|L|W|B)load {sym} [off] ptr mem) l2)
  1569        && l == l2
  1570	&& l.Uses == 2
  1571	&& clobber(l) =>
  1572  @l.Block (CMP(Q|L|W|B)constload {sym} [makeValAndOff(0, off)] ptr mem)
  1573
  1574// Convert ANDload to MOVload when we can do the AND in a containing TEST op.
  1575// Only do when it's within the same block, so we don't have flags live across basic block boundaries.
  1576// See issue 44228.
  1577(TEST(Q|L) a:(AND(Q|L)load [off] {sym} x ptr mem) a) && a.Uses == 2 && a.Block == v.Block && clobber(a) => (TEST(Q|L) (MOV(Q|L)load <a.Type> [off] {sym} ptr mem) x)
  1578
  1579(MOVBload [off] {sym} (SB) _) && symIsRO(sym) => (MOVLconst [int32(read8(sym, int64(off)))])
  1580(MOVWload [off] {sym} (SB) _) && symIsRO(sym) => (MOVLconst [int32(read16(sym, int64(off), config.ctxt.Arch.ByteOrder))])
  1581(MOVLload <t> [off] {sym} (SB) _) && symIsRO(sym) && is32BitInt(t) => (MOVLconst [int32(read32(sym, int64(off), config.ctxt.Arch.ByteOrder))])
  1582(MOVLload <t> [off] {sym} (SB) _) && symIsRO(sym) && is64BitInt(t) => (MOVQconst [int64(read32(sym, int64(off), config.ctxt.Arch.ByteOrder))])
  1583(MOVQload [off] {sym} (SB) _) && symIsRO(sym) => (MOVQconst [int64(read64(sym, int64(off), config.ctxt.Arch.ByteOrder))])
  1584(MOVBQSXload [off] {sym} (SB) _) && symIsRO(sym) => (MOVQconst [int64(int8(read8(sym, int64(off))))])
  1585(MOVWQSXload [off] {sym} (SB) _) && symIsRO(sym) => (MOVQconst [int64(int16(read16(sym, int64(off), config.ctxt.Arch.ByteOrder)))])
  1586(MOVLQSXload [off] {sym} (SB) _) && symIsRO(sym) => (MOVQconst [int64(int32(read32(sym, int64(off), config.ctxt.Arch.ByteOrder)))])
  1587
  1588
  1589(MOVOstore [dstOff] {dstSym} ptr (MOVOload [srcOff] {srcSym} (SB) _) mem) && symIsRO(srcSym) =>
  1590  (MOVQstore [dstOff+8] {dstSym} ptr (MOVQconst [int64(read64(srcSym, int64(srcOff)+8, config.ctxt.Arch.ByteOrder))])
  1591    (MOVQstore [dstOff] {dstSym} ptr (MOVQconst [int64(read64(srcSym, int64(srcOff), config.ctxt.Arch.ByteOrder))]) mem))
  1592
  1593// Arch-specific inlining for small or disjoint runtime.memmove
  1594// Match post-lowering calls, memory version.
  1595(SelectN [0] call:(CALLstatic {sym} s1:(MOVQstoreconst _ [sc] s2:(MOVQstore _ src s3:(MOVQstore _ dst mem)))))
  1596	&& sc.Val64() >= 0
  1597	&& isSameCall(sym, "runtime.memmove")
  1598	&& s1.Uses == 1 && s2.Uses == 1 && s3.Uses == 1
  1599	&& isInlinableMemmove(dst, src, sc.Val64(), config)
  1600	&& clobber(s1, s2, s3, call)
  1601	=> (Move [sc.Val64()] dst src mem)
  1602
  1603// Match post-lowering calls, register version.
  1604(SelectN [0] call:(CALLstatic {sym} dst src (MOVQconst [sz]) mem))
  1605	&& sz >= 0
  1606	&& isSameCall(sym, "runtime.memmove")
  1607	&& call.Uses == 1
  1608	&& isInlinableMemmove(dst, src, sz, config)
  1609	&& clobber(call)
  1610	=> (Move [sz] dst src mem)
  1611
  1612// Prefetch instructions
  1613(PrefetchCache ...)   => (PrefetchT0 ...)
  1614(PrefetchCacheStreamed ...) => (PrefetchNTA ...)
  1615
  1616// CPUID feature: BMI1.
  1617(AND(Q|L) x (NOT(Q|L) y))               && buildcfg.GOAMD64 >= 3 => (ANDN(Q|L) x y)
  1618(AND(Q|L) x (NEG(Q|L) x))               && buildcfg.GOAMD64 >= 3 => (BLSI(Q|L) x)
  1619(XOR(Q|L) x (ADD(Q|L)const [-1] x))     && buildcfg.GOAMD64 >= 3 => (BLSMSK(Q|L) x)
  1620(AND(Q|L) <t> x (ADD(Q|L)const [-1] x)) && buildcfg.GOAMD64 >= 3 => (Select0 <t> (BLSR(Q|L) x))
  1621// eliminate TEST instruction in classical "isPowerOfTwo" check
  1622(SETEQ       (TEST(Q|L) s:(Select0 blsr:(BLSR(Q|L) _)) s))        => (SETEQ       (Select1 <types.TypeFlags> blsr))
  1623(CMOVQEQ x y (TEST(Q|L) s:(Select0 blsr:(BLSR(Q|L) _)) s))        => (CMOVQEQ x y (Select1 <types.TypeFlags> blsr))
  1624(CMOVLEQ x y (TEST(Q|L) s:(Select0 blsr:(BLSR(Q|L) _)) s))        => (CMOVLEQ x y (Select1 <types.TypeFlags> blsr))
  1625(EQ          (TEST(Q|L) s:(Select0 blsr:(BLSR(Q|L) _)) s) yes no) => (EQ          (Select1 <types.TypeFlags> blsr) yes no)
  1626(SETNE       (TEST(Q|L) s:(Select0 blsr:(BLSR(Q|L) _)) s))        => (SETNE       (Select1 <types.TypeFlags> blsr))
  1627(CMOVQNE x y (TEST(Q|L) s:(Select0 blsr:(BLSR(Q|L) _)) s))        => (CMOVQNE x y (Select1 <types.TypeFlags> blsr))
  1628(CMOVLNE x y (TEST(Q|L) s:(Select0 blsr:(BLSR(Q|L) _)) s))        => (CMOVLNE x y (Select1 <types.TypeFlags> blsr))
  1629(NE          (TEST(Q|L) s:(Select0 blsr:(BLSR(Q|L) _)) s) yes no) => (NE          (Select1 <types.TypeFlags> blsr) yes no)
  1630
  1631(BSWAP(Q|L) (BSWAP(Q|L) p)) => p
  1632
  1633// CPUID feature: MOVBE.
  1634(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)
  1635(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)
  1636(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)
  1637(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)
  1638(MOVWstore [i] {s} p x:(ROLWconst [8] w) mem)   && x.Uses == 1 && buildcfg.GOAMD64 >= 3 => (MOVBEWstore [i] {s} p w mem)
  1639(MOVBEWstore [i] {s} p x:(ROLWconst [8] w) mem) && x.Uses == 1 => (MOVWstore [i] {s} p w mem)
  1640
  1641(SAR(Q|L) l:(MOV(Q|L)load [off] {sym} ptr mem) x) && buildcfg.GOAMD64 >= 3 && canMergeLoad(v, l) && clobber(l) => (SARX(Q|L)load [off] {sym} ptr x mem)
  1642(SHL(Q|L) l:(MOV(Q|L)load [off] {sym} ptr mem) x) && buildcfg.GOAMD64 >= 3 && canMergeLoad(v, l) && clobber(l) => (SHLX(Q|L)load [off] {sym} ptr x mem)
  1643(SHR(Q|L) l:(MOV(Q|L)load [off] {sym} ptr mem) x) && buildcfg.GOAMD64 >= 3 && canMergeLoad(v, l) && clobber(l) => (SHRX(Q|L)load [off] {sym} ptr x mem)
  1644
  1645((SHL|SHR|SAR)XQload [off] {sym} ptr (MOVQconst [c]) mem) => ((SHL|SHR|SAR)Qconst [int8(c&63)] (MOVQload [off] {sym} ptr mem))
  1646((SHL|SHR|SAR)XQload [off] {sym} ptr (MOVLconst [c]) mem) => ((SHL|SHR|SAR)Qconst [int8(c&63)] (MOVQload [off] {sym} ptr mem))
  1647((SHL|SHR|SAR)XLload [off] {sym} ptr (MOVLconst [c]) mem) => ((SHL|SHR|SAR)Lconst [int8(c&31)] (MOVLload [off] {sym} ptr mem))
  1648
  1649// Convert atomic logical operations to easier ones if we don't use the result.
  1650(Select1 a:(LoweredAtomic(And64|And32|Or64|Or32) ptr val mem)) && a.Uses == 1 && clobber(a) => ((ANDQ|ANDL|ORQ|ORL)lock ptr val mem)
  1651
  1652// If we are checking the results of an add, use the flags directly from the add.
  1653// Note that this only works for EQ/NE. ADD sets the CF/OF flags differently
  1654// than TEST sets them.
  1655// Note also that a.Args[0] here refers to the post-flagify'd value.
  1656((EQ|NE) t:(TESTQ a:(ADDQconst [c] x) a)) && t.Uses == 1 && flagify(a) => ((EQ|NE) (Select1 <types.TypeFlags> a.Args[0]))
  1657((EQ|NE) t:(TESTL a:(ADDLconst [c] x) a)) && t.Uses == 1 && flagify(a) => ((EQ|NE) (Select1 <types.TypeFlags> a.Args[0]))
  1658
  1659// If we don't use the flags any more, just use the standard op.
  1660(Select0 a:(ADD(Q|L)constflags [c] x)) && a.Uses == 1 => (ADD(Q|L)const [c] x)
  1661
  1662// SIMD lowering rules
  1663
  1664// Mask conversions
  1665// integers to masks
  1666(Cvt16toMask8x16 <t> x) => (VPMOVMToVec8x16 <types.TypeVec128> (KMOVWk <t> x))
  1667(Cvt32toMask8x32 <t> x) => (VPMOVMToVec8x32 <types.TypeVec256> (KMOVDk <t> x))
  1668(Cvt64toMask8x64 <t> x) => (VPMOVMToVec8x64 <types.TypeVec512> (KMOVQk <t> x))
  1669
  1670(Cvt8toMask16x8 <t> x) => (VPMOVMToVec16x8 <types.TypeVec128> (KMOVBk <t> x))
  1671(Cvt16toMask16x16 <t> x) => (VPMOVMToVec16x16 <types.TypeVec256> (KMOVWk <t> x))
  1672(Cvt32toMask16x32 <t> x) => (VPMOVMToVec16x32 <types.TypeVec512> (KMOVDk <t> x))
  1673
  1674(Cvt8toMask32x4 <t> x) => (VPMOVMToVec32x4 <types.TypeVec128> (KMOVBk <t> x))
  1675(Cvt8toMask32x8 <t> x) => (VPMOVMToVec32x8 <types.TypeVec256> (KMOVBk <t> x))
  1676(Cvt16toMask32x16 <t> x) => (VPMOVMToVec32x16 <types.TypeVec512> (KMOVWk <t> x))
  1677
  1678(Cvt8toMask64x2 <t> x) => (VPMOVMToVec64x2 <types.TypeVec128> (KMOVBk <t> x))
  1679(Cvt8toMask64x4 <t> x) => (VPMOVMToVec64x4 <types.TypeVec256> (KMOVBk <t> x))
  1680(Cvt8toMask64x8 <t> x) => (VPMOVMToVec64x8 <types.TypeVec512> (KMOVBk <t> x))
  1681
  1682// masks to integers
  1683(CvtMask8x16to16 ...) => (VPMOVMSKB128 ...)
  1684(CvtMask8x32to32 ...) => (VPMOVMSKB256 ...)
  1685(CvtMask8x64to64 x) => (KMOVQi (VPMOVVec8x64ToM <types.TypeMask> x))
  1686
  1687(CvtMask16x8to8 x) => (KMOVBi (VPMOVVec16x8ToM <types.TypeMask> x))
  1688(CvtMask16x16to16 x) => (KMOVWi (VPMOVVec16x16ToM <types.TypeMask> x))
  1689(CvtMask16x32to32 x) => (KMOVDi (VPMOVVec16x32ToM <types.TypeMask> x))
  1690
  1691(CvtMask32x4to8 ...) => (VMOVMSKPS128 ...)
  1692(CvtMask32x8to8 ...) => (VMOVMSKPS256 ...)
  1693(CvtMask32x16to16 x) => (KMOVWi (VPMOVVec32x16ToM <types.TypeMask> x))
  1694
  1695(CvtMask64x2to8 ...) => (VMOVMSKPD128 ...)
  1696(CvtMask64x4to8 ...) => (VMOVMSKPD256 ...)
  1697(CvtMask64x8to8 x) => (KMOVBi (VPMOVVec64x8ToM <types.TypeMask> x))
  1698
  1699// optimizations
  1700(MOVBstore [off] {sym} ptr (KMOVBi mask) mem) => (KMOVBstore [off] {sym} ptr mask mem)
  1701(MOVWstore [off] {sym} ptr (KMOVWi mask) mem) => (KMOVWstore [off] {sym} ptr mask mem)
  1702(MOVLstore [off] {sym} ptr (KMOVDi mask) mem) => (KMOVDstore [off] {sym} ptr mask mem)
  1703(MOVQstore [off] {sym} ptr (KMOVQi mask) mem) => (KMOVQstore [off] {sym} ptr mask mem)
  1704
  1705(KMOVBk l:(MOVBload [off] {sym} ptr mem)) && canMergeLoad(v, l) && clobber(l) => (KMOVBload [off] {sym} ptr mem)
  1706(KMOVWk l:(MOVWload [off] {sym} ptr mem)) && canMergeLoad(v, l) && clobber(l) => (KMOVWload [off] {sym} ptr mem)
  1707(KMOVDk l:(MOVLload [off] {sym} ptr mem)) && canMergeLoad(v, l) && clobber(l) => (KMOVDload [off] {sym} ptr mem)
  1708(KMOVQk l:(MOVQload [off] {sym} ptr mem)) && canMergeLoad(v, l) && clobber(l) => (KMOVQload [off] {sym} ptr mem)
  1709
  1710// SIMD vector loads and stores
  1711(Load <t> ptr mem) && t.Size() == 16 => (VMOVDQUload128 ptr mem)
  1712(Store {t} ptr val mem) && t.Size() == 16 => (VMOVDQUstore128 ptr val mem)
  1713
  1714(Load <t> ptr mem) && t.Size() == 32 => (VMOVDQUload256 ptr mem)
  1715(Store {t} ptr val mem) && t.Size() == 32 => (VMOVDQUstore256 ptr val mem)
  1716
  1717(Load <t> ptr mem) && t.Size() == 64 => (VMOVDQUload512 ptr mem)
  1718(Store {t} ptr val mem) && t.Size() == 64 => (VMOVDQUstore512 ptr val mem)
  1719
  1720// SIMD vector integer-vector-masked loads and stores.
  1721(LoadMasked32 <t> ptr mask mem) && t.Size() == 16 => (VPMASK32load128 ptr mask mem)
  1722(LoadMasked32 <t> ptr mask mem) && t.Size() == 32 => (VPMASK32load256 ptr mask mem)
  1723(LoadMasked64 <t> ptr mask mem) && t.Size() == 16 => (VPMASK64load128 ptr mask mem)
  1724(LoadMasked64 <t> ptr mask mem) && t.Size() == 32 => (VPMASK64load256 ptr mask mem)
  1725
  1726(StoreMasked32 {t} ptr mask val mem) && t.Size() == 16 => (VPMASK32store128 ptr mask val mem)
  1727(StoreMasked32 {t} ptr mask val mem) && t.Size() == 32 => (VPMASK32store256 ptr mask val mem)
  1728(StoreMasked64 {t} ptr mask val mem) && t.Size() == 16 => (VPMASK64store128 ptr mask val mem)
  1729(StoreMasked64 {t} ptr mask val mem) && t.Size() == 32 => (VPMASK64store256 ptr mask val mem)
  1730
  1731// Misc
  1732(IsZeroVec x) => (SETEQ (VPTEST x x))
  1733
  1734(IsNaNFloat32x4  x) => (VCMPPS128 [3] x x)
  1735(IsNaNFloat32x8  x) => (VCMPPS256 [3] x x)
  1736(IsNaNFloat32x16 x) => (VPMOVMToVec32x16 (VCMPPS512 [3] x x))
  1737(IsNaNFloat64x2  x) => (VCMPPD128 [3] x x)
  1738(IsNaNFloat64x4  x) => (VCMPPD256 [3] x x)
  1739(IsNaNFloat64x8  x) => (VPMOVMToVec64x8 (VCMPPD512 [3] x x))
  1740
  1741// SIMD vector K-masked loads and stores
  1742
  1743(LoadMasked64 <t> ptr mask mem) && t.Size() == 64 => (VPMASK64load512 ptr (VPMOVVec64x8ToM  <types.TypeMask> mask) mem)
  1744(LoadMasked32 <t> ptr mask mem) && t.Size() == 64 => (VPMASK32load512 ptr (VPMOVVec32x16ToM <types.TypeMask> mask) mem)
  1745(LoadMasked16 <t> ptr mask mem) && t.Size() == 64 => (VPMASK16load512 ptr (VPMOVVec16x32ToM <types.TypeMask> mask) mem)
  1746(LoadMasked8  <t> ptr mask mem) && t.Size() == 64 => (VPMASK8load512  ptr (VPMOVVec8x64ToM  <types.TypeMask> mask) mem)
  1747
  1748(StoreMasked64 {t} ptr mask val mem) && t.Size() == 64 => (VPMASK64store512 ptr (VPMOVVec64x8ToM  <types.TypeMask> mask) val mem)
  1749(StoreMasked32 {t} ptr mask val mem) && t.Size() == 64 => (VPMASK32store512 ptr (VPMOVVec32x16ToM <types.TypeMask> mask) val mem)
  1750(StoreMasked16 {t} ptr mask val mem) && t.Size() == 64 => (VPMASK16store512 ptr (VPMOVVec16x32ToM <types.TypeMask> mask) val mem)
  1751(StoreMasked8  {t} ptr mask val mem) && t.Size() == 64 => (VPMASK8store512  ptr (VPMOVVec8x64ToM  <types.TypeMask> mask) val mem)
  1752
  1753(ZeroSIMD <t>) && t.Size() == 16 => (Zero128 <t>)
  1754(ZeroSIMD <t>) && t.Size() == 32 => (Zero256 <t>)
  1755(ZeroSIMD <t>) && t.Size() == 64 => (Zero512 <t>)
  1756
  1757(VPMOVVec8x16ToM (VPMOVMToVec8x16 x)) => x
  1758(VPMOVVec8x32ToM (VPMOVMToVec8x32 x)) => x
  1759(VPMOVVec8x64ToM (VPMOVMToVec8x64 x)) => x
  1760
  1761(VPMOVVec16x8ToM (VPMOVMToVec16x8 x)) => x
  1762(VPMOVVec16x16ToM (VPMOVMToVec16x16 x)) => x
  1763(VPMOVVec16x32ToM (VPMOVMToVec16x32 x)) => x
  1764
  1765(VPMOVVec32x4ToM (VPMOVMToVec32x4 x)) => x
  1766(VPMOVVec32x8ToM (VPMOVMToVec32x8 x)) => x
  1767(VPMOVVec32x16ToM (VPMOVMToVec32x16 x)) => x
  1768
  1769(VPMOVVec64x2ToM (VPMOVMToVec64x2 x)) => x
  1770(VPMOVVec64x4ToM (VPMOVMToVec64x4 x)) => x
  1771(VPMOVVec64x8ToM (VPMOVMToVec64x8 x)) => x
  1772
  1773(VPANDQ512 x (VPMOVMToVec64x8 k)) => (VMOVDQU64Masked512 x k)
  1774(VPANDQ512 x (VPMOVMToVec32x16 k)) => (VMOVDQU32Masked512 x k)
  1775(VPANDQ512 x (VPMOVMToVec16x32 k)) => (VMOVDQU16Masked512 x k)
  1776(VPANDQ512 x (VPMOVMToVec8x64 k)) => (VMOVDQU8Masked512 x k)
  1777(VPANDD512 x (VPMOVMToVec64x8 k)) => (VMOVDQU64Masked512 x k)
  1778(VPANDD512 x (VPMOVMToVec32x16 k)) => (VMOVDQU32Masked512 x k)
  1779(VPANDD512 x (VPMOVMToVec16x32 k)) => (VMOVDQU16Masked512 x k)
  1780(VPANDD512 x (VPMOVMToVec8x64 k)) => (VMOVDQU8Masked512 x k)
  1781
  1782(VPAND128 x (VPMOVMToVec8x16 k)) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VMOVDQU8Masked128 x k)
  1783(VPAND128 x (VPMOVMToVec16x8 k)) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VMOVDQU16Masked128 x k)
  1784(VPAND128 x (VPMOVMToVec32x4 k)) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VMOVDQU32Masked128 x k)
  1785(VPAND128 x (VPMOVMToVec64x2 k)) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VMOVDQU64Masked128 x k)
  1786
  1787(VPAND256 x (VPMOVMToVec8x32 k)) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VMOVDQU8Masked256 x k)
  1788(VPAND256 x (VPMOVMToVec16x16 k)) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VMOVDQU16Masked256 x k)
  1789(VPAND256 x (VPMOVMToVec32x8 k)) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VMOVDQU32Masked256 x k)
  1790(VPAND256 x (VPMOVMToVec64x4 k)) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VMOVDQU64Masked256 x k)
  1791
  1792// Insert to zero of 32/64 bit floats and ints to a zero is just MOVS[SD]
  1793(VPINSRQ128 [0] (Zero128 <t>) y) && y.Type.IsFloat() => (VMOVSDf2v <types.TypeVec128> y)
  1794(VPINSRD128 [0] (Zero128 <t>) y) && y.Type.IsFloat() => (VMOVSSf2v <types.TypeVec128> y)
  1795(VPINSRQ128 [0] (Zero128 <t>) y) && !y.Type.IsFloat() => (VMOVQ <types.TypeVec128> y)
  1796(VPINSRD128 [0] (Zero128 <t>) y) && !y.Type.IsFloat() => (VMOVD <types.TypeVec128> y)
  1797
  1798// These rewrites can skip zero-extending the 8/16-bit inputs because they are
  1799// only used as the input to a broadcast; the potentially "bad" bits are ignored
  1800(VPBROADCASTB(128|256|512) x:(VPINSRB128 [0] (Zero128    <t>) y)) && x.Uses == 1 =>
  1801	(VPBROADCASTB(128|256|512) (VMOVQ <types.TypeVec128> y))
  1802(VPBROADCASTW(128|256|512) x:(VPINSRW128 [0] (Zero128    <t>) y)) && x.Uses == 1 =>
  1803	(VPBROADCASTW(128|256|512)   (VMOVQ <types.TypeVec128> y))
  1804
  1805(VMOVQ x:(MOVQload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (VMOVQload <v.Type> [off] {sym} ptr mem)
  1806(VMOVD x:(MOVLload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (VMOVDload <v.Type> [off] {sym} ptr mem)
  1807
  1808(VMOVSDf2v x:(MOVSDload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (VMOVSDload <v.Type> [off] {sym} ptr mem)
  1809(VMOVSSf2v x:(MOVSSload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (VMOVSSload <v.Type> [off] {sym} ptr mem)
  1810
  1811(VMOVSDf2v x:(MOVSDconst [c] )) => (VMOVSDconst [c] )
  1812(VMOVSSf2v x:(MOVSSconst [c] )) => (VMOVSSconst [c] )
  1813
  1814(VMOVDQUload(128|256|512) [off1] {sym} x:(ADDQconst [off2] ptr) mem) && is32Bit(int64(off1)+int64(off2)) => (VMOVDQUload(128|256|512) [off1+off2] {sym} ptr mem)
  1815(VMOVDQUstore(128|256|512) [off1] {sym} x:(ADDQconst [off2] ptr) val mem) && is32Bit(int64(off1)+int64(off2)) => (VMOVDQUstore(128|256|512) [off1+off2] {sym} ptr val mem)
  1816(VMOVDQUload(128|256|512) [off1] {sym1} x:(LEAQ [off2] {sym2} base) mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) => (VMOVDQUload(128|256|512) [off1+off2] {mergeSym(sym1, sym2)} base mem)
  1817(VMOVDQUstore(128|256|512) [off1] {sym1} x:(LEAQ [off2] {sym2} base) val mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) => (VMOVDQUstore(128|256|512) [off1+off2] {mergeSym(sym1, sym2)} base val mem)
  1818
  1819// 2-op VPTEST optimizations
  1820(SETEQ (VPTEST x:(VPAND(128|256) j k) y)) && x == y && x.Uses == 2 => (SETEQ (VPTEST j k))
  1821(SETEQ (VPTEST x:(VPAND(D|Q)512 j k) y)) && x == y && x.Uses == 2 => (SETEQ (VPTEST j k))
  1822(SETEQ (VPTEST x:(VPANDN(128|256) j k) y)) && x == y && x.Uses == 2 => (SETB (VPTEST k j)) // AndNot has swapped its operand order
  1823(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
  1824(EQ (VPTEST x:(VPAND(128|256) j k) y) yes no) && x == y && x.Uses == 2 => (EQ (VPTEST j k) yes no)
  1825(EQ (VPTEST x:(VPAND(D|Q)512 j k) y) yes no) && x == y && x.Uses == 2 => (EQ (VPTEST j k) yes no)
  1826(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
  1827(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
  1828
  1829// optimize x.IsNaN().Or(y.IsNaN())
  1830(VPOR128 (VCMPP(S|D)128 [3] x x) (VCMPP(S|D)128 [3] y y)) => (VCMPP(S|D)128 [3] x y)
  1831(VPOR256 (VCMPP(S|D)256 [3] x x) (VCMPP(S|D)256 [3] y y)) => (VCMPP(S|D)256 [3] x y)
  1832(VPORD512 (VPMOVMToVec32x16 (VCMPPS512 [3] x x)) (VPMOVMToVec32x16 (VCMPPS512 [3] y y))) =>
  1833	(VPMOVMToVec32x16 (VCMPPS512 [3] x y))
  1834(VPORD512 (VPMOVMToVec64x8  (VCMPPD512 [3] x x)) (VPMOVMToVec64x8  (VCMPPD512 [3] y y))) =>
  1835	(VPMOVMToVec64x8  (VCMPPD512 [3] x y))

View as plain text