...

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(Cvt32Fto32 ...) => (CVTTSS2SL ...)
   166(Cvt32Fto64 ...) => (CVTTSS2SQ ...)
   167(Cvt64Fto32 ...) => (CVTTSD2SL ...)
   168(Cvt64Fto64 ...) => (CVTTSD2SQ ...)
   169
   170(Cvt32Fto64F ...) => (CVTSS2SD ...)
   171(Cvt64Fto32F ...) => (CVTSD2SS ...)
   172
   173(Round(32|64)F ...) => (Copy ...)
   174
   175// Floating-point min is tricky, as the hardware op isn't right for various special
   176// cases (-0 and NaN). We use two hardware ops organized just right to make the
   177// result come out how we want it. See https://github.com/golang/go/issues/59488#issuecomment-1553493207
   178// (although that comment isn't exactly right, as the value overwritten is not simulated correctly).
   179//    t1 = MINSD x, y   => incorrect if x==NaN or x==-0,y==+0
   180//    t2 = MINSD t1, x  => fixes x==NaN case
   181//   res = POR t1, t2   => fixes x==-0,y==+0 case
   182// Note that this trick depends on the special property that (NaN OR x) produces a NaN (although
   183// it might not produce the same NaN as the input).
   184(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))
   185// Floating-point max is even trickier. Punt to using min instead.
   186// max(x,y) == -min(-x,-y)
   187(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)))
   188
   189(CvtBoolToUint8 ...) => (Copy ...)
   190
   191// Lowering shifts
   192// Unsigned shifts need to return 0 if shift amount is >= width of shifted value.
   193//   result = (arg << shift) & (shift >= argbits ? 0 : 0xffffffffffffffff)
   194(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])))
   195(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])))
   196(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])))
   197(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])))
   198
   199(Lsh64x(64|32|16|8) x y) && shiftIsBounded(v) => (SHLQ x y)
   200(Lsh32x(64|32|16|8) x y) && shiftIsBounded(v) => (SHLL x y)
   201(Lsh16x(64|32|16|8) x y) && shiftIsBounded(v) => (SHLL x y)
   202(Lsh8x(64|32|16|8)  x y) && shiftIsBounded(v) => (SHLL x y)
   203
   204(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])))
   205(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])))
   206(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])))
   207(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])))
   208
   209(Rsh64Ux(64|32|16|8) x y) && shiftIsBounded(v) => (SHRQ x y)
   210(Rsh32Ux(64|32|16|8) x y) && shiftIsBounded(v) => (SHRL x y)
   211(Rsh16Ux(64|32|16|8) x y) && shiftIsBounded(v) => (SHRW x y)
   212(Rsh8Ux(64|32|16|8)  x y) && shiftIsBounded(v) => (SHRB x y)
   213
   214// Signed right shift needs to return 0/-1 if shift amount is >= width of shifted value.
   215// We implement this by setting the shift value to -1 (all ones) if the shift value is >= width.
   216(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])))))
   217(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])))))
   218(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])))))
   219(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])))))
   220
   221(Rsh64x(64|32|16|8) x y) && shiftIsBounded(v) => (SARQ x y)
   222(Rsh32x(64|32|16|8) x y) && shiftIsBounded(v) => (SARL x y)
   223(Rsh16x(64|32|16|8) x y) && shiftIsBounded(v) => (SARW x y)
   224(Rsh8x(64|32|16|8) x y)  && shiftIsBounded(v) => (SARB x y)
   225
   226// Lowering integer comparisons
   227(Less(64|32|16|8)      x y) => (SETL  (CMP(Q|L|W|B)     x y))
   228(Less(64|32|16|8)U     x y) => (SETB  (CMP(Q|L|W|B)     x y))
   229(Leq(64|32|16|8)       x y) => (SETLE (CMP(Q|L|W|B)     x y))
   230(Leq(64|32|16|8)U      x y) => (SETBE (CMP(Q|L|W|B)     x y))
   231(Eq(Ptr|64|32|16|8|B)  x y) => (SETEQ (CMP(Q|Q|L|W|B|B) x y))
   232(Neq(Ptr|64|32|16|8|B) x y) => (SETNE (CMP(Q|Q|L|W|B|B) x y))
   233
   234// Lowering floating point comparisons
   235// Note Go assembler gets UCOMISx operand order wrong, but it is right here
   236// and the operands are reversed when generating assembly language.
   237(Eq(32|64)F   x y) => (SETEQF (UCOMIS(S|D) x y))
   238(Neq(32|64)F  x y) => (SETNEF (UCOMIS(S|D) x y))
   239// Use SETGF/SETGEF with reversed operands to dodge NaN case.
   240(Less(32|64)F x y) => (SETGF  (UCOMIS(S|D) y x))
   241(Leq(32|64)F  x y) => (SETGEF (UCOMIS(S|D) y x))
   242
   243// Lowering loads
   244(Load <t> ptr mem) && (is64BitInt(t) || isPtr(t)) => (MOVQload ptr mem)
   245(Load <t> ptr mem) && is32BitInt(t) => (MOVLload ptr mem)
   246(Load <t> ptr mem) && is16BitInt(t) => (MOVWload ptr mem)
   247(Load <t> ptr mem) && (t.IsBoolean() || is8BitInt(t)) => (MOVBload ptr mem)
   248(Load <t> ptr mem) && is32BitFloat(t) => (MOVSSload ptr mem)
   249(Load <t> ptr mem) && is64BitFloat(t) => (MOVSDload ptr mem)
   250
   251// Lowering stores
   252(Store {t} ptr val mem) && t.Size() == 8 &&  t.IsFloat() => (MOVSDstore ptr val mem)
   253(Store {t} ptr val mem) && t.Size() == 4 &&  t.IsFloat() => (MOVSSstore ptr val mem)
   254(Store {t} ptr val mem) && t.Size() == 8 && !t.IsFloat() => (MOVQstore ptr val mem)
   255(Store {t} ptr val mem) && t.Size() == 4 && !t.IsFloat() => (MOVLstore ptr val mem)
   256(Store {t} ptr val mem) && t.Size() == 2 => (MOVWstore ptr val mem)
   257(Store {t} ptr val mem) && t.Size() == 1 => (MOVBstore ptr val mem)
   258
   259// Lowering moves
   260(Move [0] _ _ mem) => mem
   261(Move [1] dst src mem) => (MOVBstore dst (MOVBload src mem) mem)
   262(Move [2] dst src mem) => (MOVWstore dst (MOVWload src mem) mem)
   263(Move [4] dst src mem) => (MOVLstore dst (MOVLload src mem) mem)
   264(Move [8] dst src mem) => (MOVQstore dst (MOVQload src mem) mem)
   265(Move [16] dst src mem) && config.useSSE => (MOVOstore dst (MOVOload src mem) mem)
   266(Move [16] dst src mem) && !config.useSSE =>
   267	(MOVQstore [8] dst (MOVQload [8] src mem)
   268		(MOVQstore dst (MOVQload src mem) mem))
   269
   270(Move [32] dst src mem) =>
   271	(Move [16]
   272		(OffPtr <dst.Type> dst [16])
   273		(OffPtr <src.Type> src [16])
   274		(Move [16] dst src mem))
   275
   276(Move [48] dst src mem) && config.useSSE =>
   277	(Move [32]
   278		(OffPtr <dst.Type> dst [16])
   279		(OffPtr <src.Type> src [16])
   280		(Move [16] dst src mem))
   281
   282(Move [64] dst src mem) && config.useSSE =>
   283	(Move [32]
   284		(OffPtr <dst.Type> dst [32])
   285		(OffPtr <src.Type> src [32])
   286		(Move [32] dst src mem))
   287
   288(Move [3] dst src mem) =>
   289	(MOVBstore [2] dst (MOVBload [2] src mem)
   290		(MOVWstore dst (MOVWload src mem) mem))
   291(Move [5] dst src mem) =>
   292	(MOVBstore [4] dst (MOVBload [4] src mem)
   293		(MOVLstore dst (MOVLload src mem) mem))
   294(Move [6] dst src mem) =>
   295	(MOVWstore [4] dst (MOVWload [4] src mem)
   296		(MOVLstore dst (MOVLload src mem) mem))
   297(Move [7] dst src mem) =>
   298	(MOVLstore [3] dst (MOVLload [3] src mem)
   299		(MOVLstore dst (MOVLload src mem) mem))
   300(Move [9] dst src mem) =>
   301	(MOVBstore [8] dst (MOVBload [8] src mem)
   302		(MOVQstore dst (MOVQload src mem) mem))
   303(Move [10] dst src mem) =>
   304	(MOVWstore [8] dst (MOVWload [8] src mem)
   305		(MOVQstore dst (MOVQload src mem) mem))
   306(Move [11] dst src mem) =>
   307	(MOVLstore [7] dst (MOVLload [7] src mem)
   308		(MOVQstore dst (MOVQload src mem) mem))
   309(Move [12] dst src mem) =>
   310	(MOVLstore [8] dst (MOVLload [8] src mem)
   311		(MOVQstore dst (MOVQload src mem) mem))
   312(Move [s] dst src mem) && s >= 13 && s <= 15 =>
   313	(MOVQstore [int32(s-8)] dst (MOVQload [int32(s-8)] src mem)
   314		(MOVQstore dst (MOVQload src mem) mem))
   315
   316// Adjust moves to be a multiple of 16 bytes.
   317(Move [s] dst src mem)
   318	&& s > 16 && s%16 != 0 && s%16 <= 8 =>
   319	(Move [s-s%16]
   320		(OffPtr <dst.Type> dst [s%16])
   321		(OffPtr <src.Type> src [s%16])
   322		(MOVQstore dst (MOVQload src mem) mem))
   323(Move [s] dst src mem)
   324	&& s > 16 && s%16 != 0 && s%16 > 8 && config.useSSE =>
   325	(Move [s-s%16]
   326		(OffPtr <dst.Type> dst [s%16])
   327		(OffPtr <src.Type> src [s%16])
   328		(MOVOstore dst (MOVOload src mem) mem))
   329(Move [s] dst src mem)
   330	&& s > 16 && s%16 != 0 && s%16 > 8 && !config.useSSE =>
   331	(Move [s-s%16]
   332		(OffPtr <dst.Type> dst [s%16])
   333		(OffPtr <src.Type> src [s%16])
   334		(MOVQstore [8] dst (MOVQload [8] src mem)
   335			(MOVQstore dst (MOVQload src mem) mem)))
   336
   337// Medium copying uses a duff device.
   338(Move [s] dst src mem)
   339	&& s > 64 && s <= 16*64 && s%16 == 0
   340	&& !config.noDuffDevice && logLargeCopy(v, s) =>
   341	(DUFFCOPY [s] dst src mem)
   342
   343// Large copying uses REP MOVSQ.
   344(Move [s] dst src mem) && (s > 16*64 || config.noDuffDevice) && s%8 == 0 && logLargeCopy(v, s) =>
   345	(REPMOVSQ dst src (MOVQconst [s/8]) mem)
   346
   347// Lowering Zero instructions
   348(Zero [0] _ mem) => mem
   349(Zero [1] destptr mem) => (MOVBstoreconst [makeValAndOff(0,0)] destptr mem)
   350(Zero [2] destptr mem) => (MOVWstoreconst [makeValAndOff(0,0)] destptr mem)
   351(Zero [4] destptr mem) => (MOVLstoreconst [makeValAndOff(0,0)] destptr mem)
   352(Zero [8] destptr mem) => (MOVQstoreconst [makeValAndOff(0,0)] destptr mem)
   353
   354(Zero [3] destptr mem) =>
   355	(MOVBstoreconst [makeValAndOff(0,2)] destptr
   356		(MOVWstoreconst [makeValAndOff(0,0)] destptr mem))
   357(Zero [5] destptr mem) =>
   358	(MOVBstoreconst [makeValAndOff(0,4)] destptr
   359		(MOVLstoreconst [makeValAndOff(0,0)] destptr mem))
   360(Zero [6] destptr mem) =>
   361	(MOVWstoreconst [makeValAndOff(0,4)] destptr
   362		(MOVLstoreconst [makeValAndOff(0,0)] destptr mem))
   363(Zero [7] destptr mem) =>
   364	(MOVLstoreconst [makeValAndOff(0,3)] destptr
   365		(MOVLstoreconst [makeValAndOff(0,0)] destptr mem))
   366
   367// Strip off any fractional word zeroing.
   368(Zero [s] destptr mem) && s%8 != 0 && s > 8 && !config.useSSE =>
   369	(Zero [s-s%8] (OffPtr <destptr.Type> destptr [s%8])
   370		(MOVQstoreconst [makeValAndOff(0,0)] destptr mem))
   371
   372// Zero small numbers of words directly.
   373(Zero [16] destptr mem) && !config.useSSE =>
   374	(MOVQstoreconst [makeValAndOff(0,8)] destptr
   375		(MOVQstoreconst [makeValAndOff(0,0)] destptr mem))
   376(Zero [24] destptr mem) && !config.useSSE =>
   377	(MOVQstoreconst [makeValAndOff(0,16)] destptr
   378		(MOVQstoreconst [makeValAndOff(0,8)] destptr
   379			(MOVQstoreconst [makeValAndOff(0,0)] destptr mem)))
   380(Zero [32] destptr mem) && !config.useSSE =>
   381	(MOVQstoreconst [makeValAndOff(0,24)] destptr
   382		(MOVQstoreconst [makeValAndOff(0,16)] destptr
   383			(MOVQstoreconst [makeValAndOff(0,8)] destptr
   384				(MOVQstoreconst [makeValAndOff(0,0)] destptr mem))))
   385
   386(Zero [9] destptr mem) && config.useSSE =>
   387	(MOVBstoreconst [makeValAndOff(0,8)] destptr
   388		(MOVQstoreconst [makeValAndOff(0,0)] destptr mem))
   389
   390(Zero [10] destptr mem) && config.useSSE =>
   391	(MOVWstoreconst [makeValAndOff(0,8)] destptr
   392		(MOVQstoreconst [makeValAndOff(0,0)] destptr mem))
   393
   394(Zero [11] destptr mem) && config.useSSE =>
   395	(MOVLstoreconst [makeValAndOff(0,7)] destptr
   396		(MOVQstoreconst [makeValAndOff(0,0)] destptr mem))
   397
   398(Zero [12] destptr mem) && config.useSSE =>
   399	(MOVLstoreconst [makeValAndOff(0,8)] destptr
   400		(MOVQstoreconst [makeValAndOff(0,0)] destptr mem))
   401
   402(Zero [s] destptr mem) && s > 12 && s < 16 && config.useSSE =>
   403	(MOVQstoreconst [makeValAndOff(0,int32(s-8))] destptr
   404		(MOVQstoreconst [makeValAndOff(0,0)] destptr mem))
   405
   406// Adjust zeros to be a multiple of 16 bytes.
   407(Zero [s] destptr mem) && s%16 != 0 && s > 16 && s%16 > 8 && config.useSSE =>
   408	(Zero [s-s%16] (OffPtr <destptr.Type> destptr [s%16])
   409		(MOVOstoreconst [makeValAndOff(0,0)] destptr mem))
   410
   411(Zero [s] destptr mem) && s%16 != 0 && s > 16 && s%16 <= 8 && config.useSSE =>
   412	(Zero [s-s%16] (OffPtr <destptr.Type> destptr [s%16])
   413		(MOVOstoreconst [makeValAndOff(0,0)] destptr mem))
   414
   415(Zero [16] destptr mem) && config.useSSE =>
   416	(MOVOstoreconst [makeValAndOff(0,0)] destptr mem)
   417(Zero [32] destptr mem) && config.useSSE =>
   418	(MOVOstoreconst [makeValAndOff(0,16)] destptr
   419		(MOVOstoreconst [makeValAndOff(0,0)] destptr mem))
   420(Zero [48] destptr mem) && config.useSSE =>
   421	(MOVOstoreconst [makeValAndOff(0,32)] destptr
   422		(MOVOstoreconst [makeValAndOff(0,16)] destptr
   423			(MOVOstoreconst [makeValAndOff(0,0)] destptr mem)))
   424(Zero [64] destptr mem) && config.useSSE =>
   425	(MOVOstoreconst [makeValAndOff(0,48)] destptr
   426		(MOVOstoreconst [makeValAndOff(0,32)] destptr
   427			(MOVOstoreconst [makeValAndOff(0,16)] destptr
   428				(MOVOstoreconst [makeValAndOff(0,0)] destptr mem))))
   429
   430// Medium zeroing uses a duff device.
   431(Zero [s] destptr mem)
   432	&& s > 64 && s <= 1024 && s%16 == 0 && !config.noDuffDevice =>
   433	(DUFFZERO [s] destptr mem)
   434
   435// Large zeroing uses REP STOSQ.
   436(Zero [s] destptr mem)
   437	&& (s > 1024 || (config.noDuffDevice && s > 64 || !config.useSSE && s > 32))
   438	&& s%8 == 0 =>
   439	(REPSTOSQ destptr (MOVQconst [s/8]) (MOVQconst [0]) mem)
   440
   441// Lowering constants
   442(Const8   [c]) => (MOVLconst [int32(c)])
   443(Const16  [c]) => (MOVLconst [int32(c)])
   444(Const32  ...) => (MOVLconst ...)
   445(Const64  ...) => (MOVQconst ...)
   446(Const32F ...) => (MOVSSconst ...)
   447(Const64F ...) => (MOVSDconst ...)
   448(ConstNil    ) => (MOVQconst [0])
   449(ConstBool [c]) => (MOVLconst [b2i32(c)])
   450
   451// Lowering calls
   452(StaticCall ...) => (CALLstatic ...)
   453(ClosureCall ...) => (CALLclosure ...)
   454(InterCall ...) => (CALLinter ...)
   455(TailCall ...) => (CALLtail ...)
   456
   457// Lowering conditional moves
   458// If the condition is a SETxx, we can just run a CMOV from the comparison that was
   459// setting the flags.
   460// Legend: HI=unsigned ABOVE, CS=unsigned BELOW, CC=unsigned ABOVE EQUAL, LS=unsigned BELOW EQUAL
   461(CondSelect <t> x y (SET(EQ|NE|L|G|LE|GE|A|B|AE|BE|EQF|NEF|GF|GEF) cond)) && (is64BitInt(t) || isPtr(t))
   462    => (CMOVQ(EQ|NE|LT|GT|LE|GE|HI|CS|CC|LS|EQF|NEF|GTF|GEF) y x cond)
   463(CondSelect <t> x y (SET(EQ|NE|L|G|LE|GE|A|B|AE|BE|EQF|NEF|GF|GEF) cond)) && is32BitInt(t)
   464    => (CMOVL(EQ|NE|LT|GT|LE|GE|HI|CS|CC|LS|EQF|NEF|GTF|GEF) y x cond)
   465(CondSelect <t> x y (SET(EQ|NE|L|G|LE|GE|A|B|AE|BE|EQF|NEF|GF|GEF) cond)) && is16BitInt(t)
   466    => (CMOVW(EQ|NE|LT|GT|LE|GE|HI|CS|CC|LS|EQF|NEF|GTF|GEF) y x cond)
   467
   468// If the condition does not set the flags, we need to generate a comparison.
   469(CondSelect <t> x y check) && !check.Type.IsFlags() && check.Type.Size() == 1
   470    => (CondSelect <t> x y (MOVBQZX <typ.UInt64> check))
   471(CondSelect <t> x y check) && !check.Type.IsFlags() && check.Type.Size() == 2
   472    => (CondSelect <t> x y (MOVWQZX <typ.UInt64> check))
   473(CondSelect <t> x y check) && !check.Type.IsFlags() && check.Type.Size() == 4
   474    => (CondSelect <t> x y (MOVLQZX <typ.UInt64> check))
   475
   476(CondSelect <t> x y check) && !check.Type.IsFlags() && check.Type.Size() == 8 && (is64BitInt(t) || isPtr(t))
   477    => (CMOVQNE y x (CMPQconst [0] check))
   478(CondSelect <t> x y check) && !check.Type.IsFlags() && check.Type.Size() == 8 && is32BitInt(t)
   479    => (CMOVLNE y x (CMPQconst [0] check))
   480(CondSelect <t> x y check) && !check.Type.IsFlags() && check.Type.Size() == 8 && is16BitInt(t)
   481    => (CMOVWNE y x (CMPQconst [0] check))
   482
   483// Absorb InvertFlags
   484(CMOVQ(EQ|NE|LT|GT|LE|GE|HI|CS|CC|LS) x y (InvertFlags cond))
   485    => (CMOVQ(EQ|NE|GT|LT|GE|LE|CS|HI|LS|CC) x y cond)
   486(CMOVL(EQ|NE|LT|GT|LE|GE|HI|CS|CC|LS) x y (InvertFlags cond))
   487    => (CMOVL(EQ|NE|GT|LT|GE|LE|CS|HI|LS|CC) x y cond)
   488(CMOVW(EQ|NE|LT|GT|LE|GE|HI|CS|CC|LS) x y (InvertFlags cond))
   489    => (CMOVW(EQ|NE|GT|LT|GE|LE|CS|HI|LS|CC) x y cond)
   490
   491// Absorb constants generated during lower
   492(CMOV(QEQ|QLE|QGE|QCC|QLS|LEQ|LLE|LGE|LCC|LLS|WEQ|WLE|WGE|WCC|WLS) _ x (FlagEQ)) => x
   493(CMOV(QNE|QLT|QGT|QCS|QHI|LNE|LLT|LGT|LCS|LHI|WNE|WLT|WGT|WCS|WHI) y _ (FlagEQ)) => y
   494(CMOV(QNE|QGT|QGE|QHI|QCC|LNE|LGT|LGE|LHI|LCC|WNE|WGT|WGE|WHI|WCC) _ x (FlagGT_UGT)) => x
   495(CMOV(QEQ|QLE|QLT|QLS|QCS|LEQ|LLE|LLT|LLS|LCS|WEQ|WLE|WLT|WLS|WCS) y _ (FlagGT_UGT)) => y
   496(CMOV(QNE|QGT|QGE|QLS|QCS|LNE|LGT|LGE|LLS|LCS|WNE|WGT|WGE|WLS|WCS) _ x (FlagGT_ULT)) => x
   497(CMOV(QEQ|QLE|QLT|QHI|QCC|LEQ|LLE|LLT|LHI|LCC|WEQ|WLE|WLT|WHI|WCC) y _ (FlagGT_ULT)) => y
   498(CMOV(QNE|QLT|QLE|QCS|QLS|LNE|LLT|LLE|LCS|LLS|WNE|WLT|WLE|WCS|WLS) _ x (FlagLT_ULT)) => x
   499(CMOV(QEQ|QGT|QGE|QHI|QCC|LEQ|LGT|LGE|LHI|LCC|WEQ|WGT|WGE|WHI|WCC) y _ (FlagLT_ULT)) => y
   500(CMOV(QNE|QLT|QLE|QHI|QCC|LNE|LLT|LLE|LHI|LCC|WNE|WLT|WLE|WHI|WCC) _ x (FlagLT_UGT)) => x
   501(CMOV(QEQ|QGT|QGE|QCS|QLS|LEQ|LGT|LGE|LCS|LLS|WEQ|WGT|WGE|WCS|WLS) y _ (FlagLT_UGT)) => y
   502
   503// Miscellaneous
   504(IsNonNil p) => (SETNE (TESTQ p p))
   505(IsInBounds idx len) => (SETB (CMPQ idx len))
   506(IsSliceInBounds idx len) => (SETBE (CMPQ idx len))
   507(NilCheck ...) => (LoweredNilCheck ...)
   508(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.
   509(GetClosurePtr ...) => (LoweredGetClosurePtr ...)
   510(GetCallerPC ...) => (LoweredGetCallerPC ...)
   511(GetCallerSP ...) => (LoweredGetCallerSP ...)
   512
   513(HasCPUFeature {s}) => (SETNE (CMPLconst [0] (LoweredHasCPUFeature {s})))
   514(Addr {sym} base) => (LEAQ {sym} base)
   515(LocalAddr <t> {sym} base mem) && t.Elem().HasPointers() => (LEAQ {sym} (SPanchored base mem))
   516(LocalAddr <t> {sym} base _)  && !t.Elem().HasPointers() => (LEAQ {sym} base)
   517
   518(MOVBstore [off] {sym} ptr y:(SETL x) mem) && y.Uses == 1 => (SETLstore [off] {sym} ptr x mem)
   519(MOVBstore [off] {sym} ptr y:(SETLE x) mem) && y.Uses == 1 => (SETLEstore [off] {sym} ptr x mem)
   520(MOVBstore [off] {sym} ptr y:(SETG x) mem) && y.Uses == 1 => (SETGstore [off] {sym} ptr x mem)
   521(MOVBstore [off] {sym} ptr y:(SETGE x) mem) && y.Uses == 1 => (SETGEstore [off] {sym} ptr x mem)
   522(MOVBstore [off] {sym} ptr y:(SETEQ x) mem) && y.Uses == 1 => (SETEQstore [off] {sym} ptr x mem)
   523(MOVBstore [off] {sym} ptr y:(SETNE x) mem) && y.Uses == 1 => (SETNEstore [off] {sym} ptr x mem)
   524(MOVBstore [off] {sym} ptr y:(SETB x) mem) && y.Uses == 1 => (SETBstore [off] {sym} ptr x mem)
   525(MOVBstore [off] {sym} ptr y:(SETBE x) mem) && y.Uses == 1 => (SETBEstore [off] {sym} ptr x mem)
   526(MOVBstore [off] {sym} ptr y:(SETA x) mem) && y.Uses == 1 => (SETAstore [off] {sym} ptr x mem)
   527(MOVBstore [off] {sym} ptr y:(SETAE x) mem) && y.Uses == 1 => (SETAEstore [off] {sym} ptr x mem)
   528
   529// block rewrites
   530(If (SETL  cmp) yes no) => (LT  cmp yes no)
   531(If (SETLE cmp) yes no) => (LE  cmp yes no)
   532(If (SETG  cmp) yes no) => (GT  cmp yes no)
   533(If (SETGE cmp) yes no) => (GE  cmp yes no)
   534(If (SETEQ cmp) yes no) => (EQ  cmp yes no)
   535(If (SETNE cmp) yes no) => (NE  cmp yes no)
   536(If (SETB  cmp) yes no) => (ULT cmp yes no)
   537(If (SETBE cmp) yes no) => (ULE cmp yes no)
   538(If (SETA  cmp) yes no) => (UGT cmp yes no)
   539(If (SETAE cmp) yes no) => (UGE cmp yes no)
   540(If (SETO cmp) yes no) => (OS cmp yes no)
   541
   542// Special case for floating point - LF/LEF not generated
   543(If (SETGF  cmp) yes no) => (UGT  cmp yes no)
   544(If (SETGEF cmp) yes no) => (UGE  cmp yes no)
   545(If (SETEQF cmp) yes no) => (EQF  cmp yes no)
   546(If (SETNEF cmp) yes no) => (NEF  cmp yes no)
   547
   548(If cond yes no) => (NE (TESTB cond cond) yes no)
   549
   550(JumpTable idx) => (JUMPTABLE {makeJumpTableSym(b)} idx (LEAQ <typ.Uintptr> {makeJumpTableSym(b)} (SB)))
   551
   552// Atomic loads.  Other than preserving their ordering with respect to other loads, nothing special here.
   553(AtomicLoad8 ptr mem) => (MOVBatomicload ptr mem)
   554(AtomicLoad32 ptr mem) => (MOVLatomicload ptr mem)
   555(AtomicLoad64 ptr mem) => (MOVQatomicload ptr mem)
   556(AtomicLoadPtr ptr mem) => (MOVQatomicload ptr mem)
   557
   558// Atomic stores.  We use XCHG to prevent the hardware reordering a subsequent load.
   559// TODO: most runtime uses of atomic stores don't need that property.  Use normal stores for those?
   560(AtomicStore8 ptr val mem) => (Select1 (XCHGB <types.NewTuple(typ.UInt8,types.TypeMem)> val ptr mem))
   561(AtomicStore32 ptr val mem) => (Select1 (XCHGL <types.NewTuple(typ.UInt32,types.TypeMem)> val ptr mem))
   562(AtomicStore64 ptr val mem) => (Select1 (XCHGQ <types.NewTuple(typ.UInt64,types.TypeMem)> val ptr mem))
   563(AtomicStorePtrNoWB ptr val mem) => (Select1 (XCHGQ <types.NewTuple(typ.BytePtr,types.TypeMem)> val ptr mem))
   564
   565// Atomic exchanges.
   566(AtomicExchange8 ptr val mem) => (XCHGB val ptr mem)
   567(AtomicExchange32 ptr val mem) => (XCHGL val ptr mem)
   568(AtomicExchange64 ptr val mem) => (XCHGQ val ptr mem)
   569
   570// Atomic adds.
   571(AtomicAdd32 ptr val mem) => (AddTupleFirst32 val (XADDLlock val ptr mem))
   572(AtomicAdd64 ptr val mem) => (AddTupleFirst64 val (XADDQlock val ptr mem))
   573(Select0 <t> (AddTupleFirst32 val tuple)) => (ADDL val (Select0 <t> tuple))
   574(Select1     (AddTupleFirst32   _ tuple)) => (Select1 tuple)
   575(Select0 <t> (AddTupleFirst64 val tuple)) => (ADDQ val (Select0 <t> tuple))
   576(Select1     (AddTupleFirst64   _ tuple)) => (Select1 tuple)
   577
   578// Atomic compare and swap.
   579(AtomicCompareAndSwap32 ptr old new_ mem) => (CMPXCHGLlock ptr old new_ mem)
   580(AtomicCompareAndSwap64 ptr old new_ mem) => (CMPXCHGQlock ptr old new_ mem)
   581
   582// Atomic memory logical operations (old style).
   583(AtomicAnd8  ptr val mem) => (ANDBlock ptr val mem)
   584(AtomicAnd32 ptr val mem) => (ANDLlock ptr val mem)
   585(AtomicOr8   ptr val mem) => (ORBlock  ptr val mem)
   586(AtomicOr32  ptr val mem) => (ORLlock  ptr val mem)
   587
   588// Atomic memory logical operations (new style).
   589(Atomic(And64|And32|Or64|Or32)value ptr val mem) => (LoweredAtomic(And64|And32|Or64|Or32) ptr val mem)
   590
   591// Write barrier.
   592(WB ...) => (LoweredWB ...)
   593
   594(PanicBounds [kind] x y mem) && boundsABI(kind) == 0 => (LoweredPanicBoundsA [kind] x y mem)
   595(PanicBounds [kind] x y mem) && boundsABI(kind) == 1 => (LoweredPanicBoundsB [kind] x y mem)
   596(PanicBounds [kind] x y mem) && boundsABI(kind) == 2 => (LoweredPanicBoundsC [kind] x y mem)
   597
   598// lowering rotates
   599(RotateLeft8  ...) => (ROLB ...)
   600(RotateLeft16 ...) => (ROLW ...)
   601(RotateLeft32 ...) => (ROLL ...)
   602(RotateLeft64 ...) => (ROLQ ...)
   603
   604// ***************************
   605// Above: lowering rules
   606// Below: optimizations
   607// ***************************
   608// TODO: Should the optimizations be a separate pass?
   609
   610// Fold boolean tests into blocks
   611(NE (TESTB (SETL  cmp) (SETL  cmp)) yes no) => (LT  cmp yes no)
   612(NE (TESTB (SETLE cmp) (SETLE cmp)) yes no) => (LE  cmp yes no)
   613(NE (TESTB (SETG  cmp) (SETG  cmp)) yes no) => (GT  cmp yes no)
   614(NE (TESTB (SETGE cmp) (SETGE cmp)) yes no) => (GE  cmp yes no)
   615(NE (TESTB (SETEQ cmp) (SETEQ cmp)) yes no) => (EQ  cmp yes no)
   616(NE (TESTB (SETNE cmp) (SETNE cmp)) yes no) => (NE  cmp yes no)
   617(NE (TESTB (SETB  cmp) (SETB  cmp)) yes no) => (ULT cmp yes no)
   618(NE (TESTB (SETBE cmp) (SETBE cmp)) yes no) => (ULE cmp yes no)
   619(NE (TESTB (SETA  cmp) (SETA  cmp)) yes no) => (UGT cmp yes no)
   620(NE (TESTB (SETAE cmp) (SETAE cmp)) yes no) => (UGE cmp yes no)
   621(NE (TESTB (SETO cmp) (SETO cmp)) yes no) => (OS cmp yes no)
   622
   623// Unsigned comparisons to 0/1
   624(ULT (TEST(Q|L|W|B) x x) yes no) => (First no yes)
   625(UGE (TEST(Q|L|W|B) x x) yes no) => (First yes no)
   626(SETB (TEST(Q|L|W|B) x x)) => (ConstBool [false])
   627(SETAE (TEST(Q|L|W|B) x x)) => (ConstBool [true])
   628
   629// x & 1 != 0 -> x & 1
   630(SETNE (TEST(B|W)const [1] x)) => (AND(L|L)const [1] x)
   631(SETB (BT(L|Q)const [0] x)) => (AND(L|Q)const [1] x)
   632
   633// Recognize bit tests: a&(1<<b) != 0 for b suitably bounded
   634// Note that BTx instructions use the carry bit, so we need to convert tests for zero flag
   635// into tests for carry flags.
   636// ULT and SETB check the carry flag; they are identical to CS and SETCS. Same, mutatis
   637// mutandis, for UGE and SETAE, and CC and SETCC.
   638((NE|EQ) (TESTL (SHLL (MOVLconst [1]) x) y)) => ((ULT|UGE) (BTL x y))
   639((NE|EQ) (TESTQ (SHLQ (MOVQconst [1]) x) y)) => ((ULT|UGE) (BTQ x y))
   640((NE|EQ) (TESTLconst [c] x)) && isUint32PowerOfTwo(int64(c))
   641    => ((ULT|UGE) (BTLconst [int8(log32(c))] x))
   642((NE|EQ) (TESTQconst [c] x)) && isUint64PowerOfTwo(int64(c))
   643    => ((ULT|UGE) (BTQconst [int8(log32(c))] x))
   644((NE|EQ) (TESTQ (MOVQconst [c]) x)) && isUint64PowerOfTwo(c)
   645    => ((ULT|UGE) (BTQconst [int8(log64(c))] x))
   646(SET(NE|EQ) (TESTL (SHLL (MOVLconst [1]) x) y)) => (SET(B|AE)  (BTL x y))
   647(SET(NE|EQ) (TESTQ (SHLQ (MOVQconst [1]) x) y)) => (SET(B|AE)  (BTQ x y))
   648(SET(NE|EQ) (TESTLconst [c] x)) && isUint32PowerOfTwo(int64(c))
   649    => (SET(B|AE)  (BTLconst [int8(log32(c))] x))
   650(SET(NE|EQ) (TESTQconst [c] x)) && isUint64PowerOfTwo(int64(c))
   651    => (SET(B|AE)  (BTQconst [int8(log32(c))] x))
   652(SET(NE|EQ) (TESTQ (MOVQconst [c]) x)) && isUint64PowerOfTwo(c)
   653    => (SET(B|AE)  (BTQconst [int8(log64(c))] x))
   654// SET..store variant
   655(SET(NE|EQ)store [off] {sym} ptr (TESTL (SHLL (MOVLconst [1]) x) y) mem)
   656    => (SET(B|AE)store  [off] {sym} ptr (BTL x y) mem)
   657(SET(NE|EQ)store [off] {sym} ptr (TESTQ (SHLQ (MOVQconst [1]) x) y) mem)
   658    => (SET(B|AE)store  [off] {sym} ptr (BTQ x y) mem)
   659(SET(NE|EQ)store [off] {sym} ptr (TESTLconst [c] x) mem) && isUint32PowerOfTwo(int64(c))
   660    => (SET(B|AE)store  [off] {sym} ptr (BTLconst [int8(log32(c))] x) mem)
   661(SET(NE|EQ)store [off] {sym} ptr (TESTQconst [c] x) mem) && isUint64PowerOfTwo(int64(c))
   662    => (SET(B|AE)store  [off] {sym} ptr (BTQconst [int8(log32(c))] x) mem)
   663(SET(NE|EQ)store [off] {sym} ptr (TESTQ (MOVQconst [c]) x) mem) && isUint64PowerOfTwo(c)
   664    => (SET(B|AE)store  [off] {sym} ptr (BTQconst [int8(log64(c))] x) mem)
   665
   666// Handle bit-testing in the form (a>>b)&1 != 0 by building the above rules
   667// and further combining shifts.
   668(BT(Q|L)const [c] (SHRQconst [d] x)) && (c+d)<64 => (BTQconst [c+d] x)
   669(BT(Q|L)const [c] (SHLQconst [d] x)) && c>d      => (BT(Q|L)const [c-d] x)
   670(BT(Q|L)const [0] s:(SHRQ x y)) => (BTQ y x)
   671(BTLconst [c] (SHRLconst [d] x)) && (c+d)<32 => (BTLconst [c+d] x)
   672(BTLconst [c] (SHLLconst [d] x)) && c>d      => (BTLconst [c-d] x)
   673(BTLconst [0] s:(SHR(L|XL) x y)) => (BTL y x)
   674
   675// Rewrite a & 1 != 1 into a & 1 == 0.
   676// Among other things, this lets us turn (a>>b)&1 != 1 into a bit test.
   677(SET(NE|EQ) (CMPLconst [1] s:(ANDLconst [1] _))) => (SET(EQ|NE) (CMPLconst [0] s))
   678(SET(NE|EQ)store [off] {sym} ptr (CMPLconst [1] s:(ANDLconst [1] _)) mem) => (SET(EQ|NE)store [off] {sym} ptr (CMPLconst [0] s) mem)
   679(SET(NE|EQ) (CMPQconst [1] s:(ANDQconst [1] _))) => (SET(EQ|NE) (CMPQconst [0] s))
   680(SET(NE|EQ)store [off] {sym} ptr (CMPQconst [1] s:(ANDQconst [1] _)) mem) => (SET(EQ|NE)store [off] {sym} ptr (CMPQconst [0] s) mem)
   681
   682// Recognize bit setting (a |= 1<<b) and toggling (a ^= 1<<b)
   683(OR(Q|L) (SHL(Q|L) (MOV(Q|L)const [1]) y) x) => (BTS(Q|L) x y)
   684(XOR(Q|L) (SHL(Q|L) (MOV(Q|L)const [1]) y) x) => (BTC(Q|L) x y)
   685// Note: only convert OR/XOR to BTS/BTC if the constant wouldn't fit in
   686// the constant field of the OR/XOR instruction. See issue 61694.
   687((OR|XOR)Q (MOVQconst [c]) x) && isUint64PowerOfTwo(c) && uint64(c) >= 1<<31 => (BT(S|C)Qconst [int8(log64(c))] x)
   688
   689// Recognize bit clearing: a &^= 1<<b
   690(AND(Q|L) (NOT(Q|L) (SHL(Q|L) (MOV(Q|L)const [1]) y)) x) => (BTR(Q|L) x y)
   691(ANDN(Q|L) x (SHL(Q|L) (MOV(Q|L)const [1]) y)) => (BTR(Q|L) x y)
   692// Note: only convert AND to BTR if the constant wouldn't fit in
   693// the constant field of the AND instruction. See issue 61694.
   694(ANDQ (MOVQconst [c]) x) && isUint64PowerOfTwo(^c) && uint64(^c) >= 1<<31 => (BTRQconst [int8(log64(^c))] x)
   695
   696// Special-case bit patterns on first/last bit.
   697// generic.rules changes ANDs of high-part/low-part masks into a couple of shifts,
   698// for instance:
   699//    x & 0xFFFF0000 -> (x >> 16) << 16
   700//    x & 0x80000000 -> (x >> 31) << 31
   701//
   702// In case the mask is just one bit (like second example above), it conflicts
   703// with the above rules to detect bit-testing / bit-clearing of first/last bit.
   704// We thus special-case them, by detecting the shift patterns.
   705
   706// Special case resetting first/last bit
   707(SHL(L|Q)const [1] (SHR(L|Q)const [1] x))
   708	=> (AND(L|Q)const [-2] x)
   709(SHRLconst [1] (SHLLconst [1] x))
   710	=> (ANDLconst [0x7fffffff] x)
   711(SHRQconst [1] (SHLQconst [1] x))
   712	=> (BTRQconst [63] x)
   713
   714// Special case testing first/last bit (with double-shift generated by generic.rules)
   715((SETNE|SETEQ|NE|EQ) (TESTQ z1:(SHLQconst [63] (SHRQconst [63] x)) z2)) && z1==z2
   716    => ((SETB|SETAE|ULT|UGE) (BTQconst [63] x))
   717((SETNE|SETEQ|NE|EQ) (TESTL z1:(SHLLconst [31] (SHRQconst [31] x)) z2)) && z1==z2
   718    => ((SETB|SETAE|ULT|UGE) (BTQconst [31] x))
   719(SET(NE|EQ)store [off] {sym} ptr (TESTQ z1:(SHLQconst [63] (SHRQconst [63] x)) z2) mem) && z1==z2
   720    => (SET(B|AE)store [off] {sym} ptr (BTQconst [63] x) mem)
   721(SET(NE|EQ)store [off] {sym} ptr (TESTL z1:(SHLLconst [31] (SHRLconst [31] x)) z2) mem) && z1==z2
   722    => (SET(B|AE)store [off] {sym} ptr (BTLconst [31] x) mem)
   723
   724((SETNE|SETEQ|NE|EQ) (TESTQ z1:(SHRQconst [63] (SHLQconst [63] x)) z2)) && z1==z2
   725    => ((SETB|SETAE|ULT|UGE)  (BTQconst [0] x))
   726((SETNE|SETEQ|NE|EQ) (TESTL z1:(SHRLconst [31] (SHLLconst [31] x)) z2)) && z1==z2
   727    => ((SETB|SETAE|ULT|UGE)  (BTLconst [0] x))
   728(SET(NE|EQ)store [off] {sym} ptr (TESTQ z1:(SHRQconst [63] (SHLQconst [63] x)) z2) mem) && z1==z2
   729    => (SET(B|AE)store [off] {sym} ptr (BTQconst [0] x) mem)
   730(SET(NE|EQ)store [off] {sym} ptr (TESTL z1:(SHRLconst [31] (SHLLconst [31] x)) z2) mem) && z1==z2
   731    => (SET(B|AE)store [off] {sym} ptr (BTLconst [0] x) mem)
   732
   733// Special-case manually testing last bit with "a>>63 != 0" (without "&1")
   734((SETNE|SETEQ|NE|EQ) (TESTQ z1:(SHRQconst [63] x) z2)) && z1==z2
   735    => ((SETB|SETAE|ULT|UGE) (BTQconst [63] x))
   736((SETNE|SETEQ|NE|EQ) (TESTL z1:(SHRLconst [31] x) z2)) && z1==z2
   737    => ((SETB|SETAE|ULT|UGE) (BTLconst [31] x))
   738(SET(NE|EQ)store [off] {sym} ptr (TESTQ z1:(SHRQconst [63] x) z2) mem) && z1==z2
   739    => (SET(B|AE)store [off] {sym} ptr (BTQconst [63] x) mem)
   740(SET(NE|EQ)store [off] {sym} ptr (TESTL z1:(SHRLconst [31] x) z2) mem) && z1==z2
   741    => (SET(B|AE)store [off] {sym} ptr (BTLconst [31] x) mem)
   742
   743// Fold combinations of bit ops on same bit. An example is math.Copysign(c,-1)
   744(BTSQconst [c] (BTRQconst [c] x)) => (BTSQconst [c] x)
   745(BTSQconst [c] (BTCQconst [c] x)) => (BTSQconst [c] x)
   746(BTRQconst [c] (BTSQconst [c] x)) => (BTRQconst [c] x)
   747(BTRQconst [c] (BTCQconst [c] x)) => (BTRQconst [c] x)
   748
   749// Fold boolean negation into SETcc.
   750(XORLconst [1] (SETNE x)) => (SETEQ x)
   751(XORLconst [1] (SETEQ x)) => (SETNE x)
   752(XORLconst [1] (SETL  x)) => (SETGE x)
   753(XORLconst [1] (SETGE x)) => (SETL  x)
   754(XORLconst [1] (SETLE x)) => (SETG  x)
   755(XORLconst [1] (SETG  x)) => (SETLE x)
   756(XORLconst [1] (SETB  x)) => (SETAE x)
   757(XORLconst [1] (SETAE x)) => (SETB  x)
   758(XORLconst [1] (SETBE x)) => (SETA  x)
   759(XORLconst [1] (SETA  x)) => (SETBE x)
   760
   761// Special case for floating point - LF/LEF not generated
   762(NE (TESTB (SETGF  cmp) (SETGF  cmp)) yes no) => (UGT  cmp yes no)
   763(NE (TESTB (SETGEF cmp) (SETGEF cmp)) yes no) => (UGE  cmp yes no)
   764(NE (TESTB (SETEQF cmp) (SETEQF cmp)) yes no) => (EQF  cmp yes no)
   765(NE (TESTB (SETNEF cmp) (SETNEF cmp)) yes no) => (NEF  cmp yes no)
   766
   767// Disabled because it interferes with the pattern match above and makes worse code.
   768// (SETNEF x) => (ORQ (SETNE <typ.Int8> x) (SETNAN <typ.Int8> x))
   769// (SETEQF x) => (ANDQ (SETEQ <typ.Int8> x) (SETORD <typ.Int8> x))
   770
   771// fold constants into instructions
   772(ADDQ x (MOVQconst <t> [c])) && is32Bit(c) && !t.IsPtr() => (ADDQconst [int32(c)] x)
   773(ADDQ x (MOVLconst [c])) => (ADDQconst [c] x)
   774(ADDL x (MOVLconst [c])) => (ADDLconst [c] x)
   775
   776(SUBQ x (MOVQconst [c])) && is32Bit(c) => (SUBQconst x [int32(c)])
   777(SUBQ (MOVQconst [c]) x) && is32Bit(c) => (NEGQ (SUBQconst <v.Type> x [int32(c)]))
   778(SUBL x (MOVLconst [c])) => (SUBLconst x [c])
   779(SUBL (MOVLconst [c]) x) => (NEGL (SUBLconst <v.Type> x [c]))
   780
   781(MULQ x (MOVQconst [c])) && is32Bit(c) => (MULQconst [int32(c)] x)
   782(MULL x (MOVLconst [c])) => (MULLconst [c] x)
   783
   784(ANDQ x (MOVQconst [c])) && is32Bit(c) => (ANDQconst [int32(c)] x)
   785(ANDL x (MOVLconst [c])) => (ANDLconst [c] x)
   786
   787(AND(L|Q)const [c] (AND(L|Q)const [d] x)) => (AND(L|Q)const [c & d] x)
   788(XOR(L|Q)const [c] (XOR(L|Q)const [d] x)) => (XOR(L|Q)const [c ^ d] x)
   789(OR(L|Q)const  [c] (OR(L|Q)const  [d] x)) => (OR(L|Q)const  [c | d] x)
   790
   791(MULLconst [c] (MULLconst [d] x)) => (MULLconst [c * d] x)
   792(MULQconst [c] (MULQconst [d] x)) && is32Bit(int64(c)*int64(d)) => (MULQconst [c * d] x)
   793
   794(ORQ x (MOVQconst [c])) && is32Bit(c) => (ORQconst [int32(c)] x)
   795(ORQ x (MOVLconst [c])) => (ORQconst [c] x)
   796(ORL x (MOVLconst [c])) => (ORLconst [c] x)
   797
   798(XORQ x (MOVQconst [c])) && is32Bit(c) => (XORQconst [int32(c)] x)
   799(XORL x (MOVLconst [c])) => (XORLconst [c] x)
   800
   801(SHLQ x (MOV(Q|L)const [c])) => (SHLQconst [int8(c&63)] x)
   802(SHLL x (MOV(Q|L)const [c])) => (SHLLconst [int8(c&31)] x)
   803
   804(SHRQ x (MOV(Q|L)const [c])) => (SHRQconst [int8(c&63)] x)
   805(SHRL x (MOV(Q|L)const [c])) => (SHRLconst [int8(c&31)] x)
   806(SHRW x (MOV(Q|L)const [c])) && c&31 < 16 => (SHRWconst [int8(c&31)] x)
   807(SHRW _ (MOV(Q|L)const [c])) && c&31 >= 16 => (MOVLconst [0])
   808(SHRB x (MOV(Q|L)const [c])) && c&31 < 8 => (SHRBconst [int8(c&31)] x)
   809(SHRB _ (MOV(Q|L)const [c])) && c&31 >= 8 => (MOVLconst [0])
   810
   811(SARQ x (MOV(Q|L)const [c])) => (SARQconst [int8(c&63)] x)
   812(SARL x (MOV(Q|L)const [c])) => (SARLconst [int8(c&31)] x)
   813(SARW x (MOV(Q|L)const [c])) => (SARWconst [int8(min(int64(c)&31,15))] x)
   814(SARB x (MOV(Q|L)const [c])) => (SARBconst [int8(min(int64(c)&31,7))] x)
   815
   816// Operations which don't affect the low 6/5 bits of the shift amount are NOPs.
   817((SHLQ|SHRQ|SARQ) x (ADDQconst [c] y)) && c & 63 == 0  => ((SHLQ|SHRQ|SARQ) x y)
   818((SHLQ|SHRQ|SARQ) x (NEGQ <t> (ADDQconst [c] y))) && c & 63 == 0  => ((SHLQ|SHRQ|SARQ) x (NEGQ <t> y))
   819((SHLQ|SHRQ|SARQ) x (ANDQconst [c] y)) && c & 63 == 63 => ((SHLQ|SHRQ|SARQ) x y)
   820((SHLQ|SHRQ|SARQ) x (NEGQ <t> (ANDQconst [c] y))) && c & 63 == 63 => ((SHLQ|SHRQ|SARQ) x (NEGQ <t> y))
   821
   822((SHLL|SHRL|SARL) x (ADDQconst [c] y)) && c & 31 == 0  => ((SHLL|SHRL|SARL) x y)
   823((SHLL|SHRL|SARL) x (NEGQ <t> (ADDQconst [c] y))) && c & 31 == 0  => ((SHLL|SHRL|SARL) x (NEGQ <t> y))
   824((SHLL|SHRL|SARL) x (ANDQconst [c] y)) && c & 31 == 31 => ((SHLL|SHRL|SARL) x y)
   825((SHLL|SHRL|SARL) x (NEGQ <t> (ANDQconst [c] y))) && c & 31 == 31 => ((SHLL|SHRL|SARL) x (NEGQ <t> y))
   826
   827((SHLQ|SHRQ|SARQ) x (ADDLconst [c] y)) && c & 63 == 0  => ((SHLQ|SHRQ|SARQ) x y)
   828((SHLQ|SHRQ|SARQ) x (NEGL <t> (ADDLconst [c] y))) && c & 63 == 0  => ((SHLQ|SHRQ|SARQ) x (NEGL <t> y))
   829((SHLQ|SHRQ|SARQ) x (ANDLconst [c] y)) && c & 63 == 63 => ((SHLQ|SHRQ|SARQ) x y)
   830((SHLQ|SHRQ|SARQ) x (NEGL <t> (ANDLconst [c] y))) && c & 63 == 63 => ((SHLQ|SHRQ|SARQ) x (NEGL <t> y))
   831
   832((SHLL|SHRL|SARL) x (ADDLconst [c] y)) && c & 31 == 0  => ((SHLL|SHRL|SARL) x y)
   833((SHLL|SHRL|SARL) x (NEGL <t> (ADDLconst [c] y))) && c & 31 == 0  => ((SHLL|SHRL|SARL) x (NEGL <t> y))
   834((SHLL|SHRL|SARL) x (ANDLconst [c] y)) && c & 31 == 31 => ((SHLL|SHRL|SARL) x y)
   835((SHLL|SHRL|SARL) x (NEGL <t> (ANDLconst [c] y))) && c & 31 == 31 => ((SHLL|SHRL|SARL) x (NEGL <t> y))
   836
   837// rotate left negative = rotate right
   838(ROLQ x (NEG(Q|L) y)) => (RORQ x y)
   839(ROLL x (NEG(Q|L) y)) => (RORL x y)
   840(ROLW x (NEG(Q|L) y)) => (RORW x y)
   841(ROLB x (NEG(Q|L) y)) => (RORB x y)
   842
   843// rotate right negative = rotate left
   844(RORQ x (NEG(Q|L) y)) => (ROLQ x y)
   845(RORL x (NEG(Q|L) y)) => (ROLL x y)
   846(RORW x (NEG(Q|L) y)) => (ROLW x y)
   847(RORB x (NEG(Q|L) y)) => (ROLB x y)
   848
   849// rotate by constants
   850(ROLQ x (MOV(Q|L)const [c])) => (ROLQconst [int8(c&63)] x)
   851(ROLL x (MOV(Q|L)const [c])) => (ROLLconst [int8(c&31)] x)
   852(ROLW x (MOV(Q|L)const [c])) => (ROLWconst [int8(c&15)] x)
   853(ROLB x (MOV(Q|L)const [c])) => (ROLBconst [int8(c&7) ] x)
   854
   855(RORQ x (MOV(Q|L)const [c])) => (ROLQconst [int8((-c)&63)] x)
   856(RORL x (MOV(Q|L)const [c])) => (ROLLconst [int8((-c)&31)] x)
   857(RORW x (MOV(Q|L)const [c])) => (ROLWconst [int8((-c)&15)] x)
   858(RORB x (MOV(Q|L)const [c])) => (ROLBconst [int8((-c)&7) ] x)
   859
   860// Constant shift simplifications
   861((SHLQ|SHRQ|SARQ)const      x [0]) => x
   862((SHLL|SHRL|SARL)const      x [0]) => x
   863((SHRW|SARW)const           x [0]) => x
   864((SHRB|SARB)const           x [0]) => x
   865((ROLQ|ROLL|ROLW|ROLB)const x [0]) => x
   866
   867// Multi-register shifts
   868(ORQ (SH(R|L)Q lo bits) (SH(L|R)Q hi (NEGQ bits))) => (SH(R|L)DQ lo hi bits)
   869(ORQ (SH(R|L)XQ lo bits) (SH(L|R)XQ hi (NEGQ bits))) => (SH(R|L)DQ lo hi bits)
   870
   871// Note: the word and byte shifts keep the low 5 bits (not the low 4 or 3 bits)
   872// because the x86 instructions are defined to use all 5 bits of the shift even
   873// for the small shifts. I don't think we'll ever generate a weird shift (e.g.
   874// (SHRW x (MOVLconst [24])), but just in case.
   875
   876(CMPQ x (MOVQconst [c])) && is32Bit(c) => (CMPQconst x [int32(c)])
   877(CMPQ (MOVQconst [c]) x) && is32Bit(c) => (InvertFlags (CMPQconst x [int32(c)]))
   878(CMPL x (MOVLconst [c])) => (CMPLconst x [c])
   879(CMPL (MOVLconst [c]) x) => (InvertFlags (CMPLconst x [c]))
   880(CMPW x (MOVLconst [c])) => (CMPWconst x [int16(c)])
   881(CMPW (MOVLconst [c]) x) => (InvertFlags (CMPWconst x [int16(c)]))
   882(CMPB x (MOVLconst [c])) => (CMPBconst x [int8(c)])
   883(CMPB (MOVLconst [c]) x) => (InvertFlags (CMPBconst x [int8(c)]))
   884
   885// Canonicalize the order of arguments to comparisons - helps with CSE.
   886(CMP(Q|L|W|B) x y) && canonLessThan(x,y) => (InvertFlags (CMP(Q|L|W|B) y x))
   887
   888// Using MOVZX instead of AND is cheaper.
   889(AND(Q|L)const [  0xFF] x) => (MOVBQZX x)
   890(AND(Q|L)const [0xFFFF] x) => (MOVWQZX x)
   891// This rule is currently invalid because 0xFFFFFFFF is not representable by a signed int32.
   892// Commenting out for now, because it also can't trigger because of the is32bit guard on the
   893// ANDQconst lowering-rule, above, prevents 0xFFFFFFFF from matching (for the same reason)
   894// Using an alternate form of this rule segfaults some binaries because of
   895// adverse interactions with other passes.
   896// (ANDQconst [0xFFFFFFFF] x) => (MOVLQZX x)
   897
   898// strength reduction
   899// Assumes that the following costs from https://gmplib.org/~tege/x86-timing.pdf:
   900//    1 - addq, shlq, leaq, negq, subq
   901//    3 - imulq
   902// This limits the rewrites to two instructions.
   903// Note that negq always operates in-place,
   904// which can require a register-register move
   905// to preserve the original value,
   906// so it must be used with care.
   907(MUL(Q|L)const [-9] x) => (NEG(Q|L) (LEA(Q|L)8 <v.Type> x x))
   908(MUL(Q|L)const [-5] x) => (NEG(Q|L) (LEA(Q|L)4 <v.Type> x x))
   909(MUL(Q|L)const [-3] x) => (NEG(Q|L) (LEA(Q|L)2 <v.Type> x x))
   910(MUL(Q|L)const [-1] x) => (NEG(Q|L) x)
   911(MUL(Q|L)const [ 0] _) => (MOV(Q|L)const [0])
   912(MUL(Q|L)const [ 1] x) => x
   913(MUL(Q|L)const [ 3] x) => (LEA(Q|L)2 x x)
   914(MUL(Q|L)const [ 5] x) => (LEA(Q|L)4 x x)
   915(MUL(Q|L)const [ 7] x) => (LEA(Q|L)2 x (LEA(Q|L)2 <v.Type> x x))
   916(MUL(Q|L)const [ 9] x) => (LEA(Q|L)8 x x)
   917(MUL(Q|L)const [11] x) => (LEA(Q|L)2 x (LEA(Q|L)4 <v.Type> x x))
   918(MUL(Q|L)const [13] x) => (LEA(Q|L)4 x (LEA(Q|L)2 <v.Type> x x))
   919(MUL(Q|L)const [19] x) => (LEA(Q|L)2 x (LEA(Q|L)8 <v.Type> x x))
   920(MUL(Q|L)const [21] x) => (LEA(Q|L)4 x (LEA(Q|L)4 <v.Type> x x))
   921(MUL(Q|L)const [25] x) => (LEA(Q|L)8 x (LEA(Q|L)2 <v.Type> x x))
   922(MUL(Q|L)const [27] x) => (LEA(Q|L)8 (LEA(Q|L)2 <v.Type> x x) (LEA(Q|L)2 <v.Type> x x))
   923(MUL(Q|L)const [37] x) => (LEA(Q|L)4 x (LEA(Q|L)8 <v.Type> x x))
   924(MUL(Q|L)const [41] x) => (LEA(Q|L)8 x (LEA(Q|L)4 <v.Type> x x))
   925(MUL(Q|L)const [45] x) => (LEA(Q|L)8 (LEA(Q|L)4 <v.Type> x x) (LEA(Q|L)4 <v.Type> x x))
   926(MUL(Q|L)const [73] x) => (LEA(Q|L)8 x (LEA(Q|L)8 <v.Type> x x))
   927(MUL(Q|L)const [81] x) => (LEA(Q|L)8 (LEA(Q|L)8 <v.Type> x x) (LEA(Q|L)8 <v.Type> x x))
   928
   929(MUL(Q|L)const [c] x) && isPowerOfTwo(int64(c)+1) && c >=  15 => (SUB(Q|L)  (SHL(Q|L)const <v.Type> [int8(log64(int64(c)+1))] x) x)
   930(MUL(Q|L)const [c] x) && isPowerOfTwo(c-1) && c >=  17 => (LEA(Q|L)1 (SHL(Q|L)const <v.Type> [int8(log32(c-1))] x) x)
   931(MUL(Q|L)const [c] x) && isPowerOfTwo(c-2) && c >=  34 => (LEA(Q|L)2 (SHL(Q|L)const <v.Type> [int8(log32(c-2))] x) x)
   932(MUL(Q|L)const [c] x) && isPowerOfTwo(c-4) && c >=  68 => (LEA(Q|L)4 (SHL(Q|L)const <v.Type> [int8(log32(c-4))] x) x)
   933(MUL(Q|L)const [c] x) && isPowerOfTwo(c-8) && c >= 136 => (LEA(Q|L)8 (SHL(Q|L)const <v.Type> [int8(log32(c-8))] x) x)
   934(MUL(Q|L)const [c] x) && c%3 == 0 && isPowerOfTwo(c/3) => (SHL(Q|L)const [int8(log32(c/3))] (LEA(Q|L)2 <v.Type> x x))
   935(MUL(Q|L)const [c] x) && c%5 == 0 && isPowerOfTwo(c/5) => (SHL(Q|L)const [int8(log32(c/5))] (LEA(Q|L)4 <v.Type> x x))
   936(MUL(Q|L)const [c] x) && c%9 == 0 && isPowerOfTwo(c/9) => (SHL(Q|L)const [int8(log32(c/9))] (LEA(Q|L)8 <v.Type> x x))
   937
   938// combine add/shift into LEAQ/LEAL
   939(ADD(L|Q) x (SHL(L|Q)const [3] y)) => (LEA(L|Q)8 x y)
   940(ADD(L|Q) x (SHL(L|Q)const [2] y)) => (LEA(L|Q)4 x y)
   941(ADD(L|Q) x (SHL(L|Q)const [1] y)) => (LEA(L|Q)2 x y)
   942(ADD(L|Q) x (ADD(L|Q) y y))        => (LEA(L|Q)2 x y)
   943(ADD(L|Q) x (ADD(L|Q) x y))        => (LEA(L|Q)2 y x)
   944
   945// combine ADDQ/ADDQconst into LEAQ1/LEAL1
   946(ADD(Q|L)const [c] (ADD(Q|L) x y)) => (LEA(Q|L)1 [c] x y)
   947(ADD(Q|L) (ADD(Q|L)const [c] x) y) => (LEA(Q|L)1 [c] x y)
   948(ADD(Q|L)const [c] (SHL(Q|L)const [1] x)) => (LEA(Q|L)1 [c] x x)
   949
   950// fold ADDQ/ADDL into LEAQ/LEAL
   951(ADD(Q|L)const [c] (LEA(Q|L) [d] {s} x)) && is32Bit(int64(c)+int64(d)) => (LEA(Q|L) [c+d] {s} x)
   952(LEA(Q|L) [c] {s} (ADD(Q|L)const [d] x)) && is32Bit(int64(c)+int64(d)) => (LEA(Q|L) [c+d] {s} x)
   953(LEA(Q|L) [c] {s} (ADD(Q|L) x y)) && x.Op != OpSB && y.Op != OpSB => (LEA(Q|L)1 [c] {s} x y)
   954(ADD(Q|L) x (LEA(Q|L) [c] {s} y)) && x.Op != OpSB && y.Op != OpSB => (LEA(Q|L)1 [c] {s} x y)
   955
   956// fold ADDQconst/ADDLconst into LEAQx/LEALx
   957(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)
   958(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)
   959(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)
   960(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)
   961(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)
   962(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)
   963(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)
   964(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)
   965(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)
   966(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)
   967(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)
   968
   969// fold shifts into LEAQx/LEALx
   970(LEA(Q|L)1 [c] {s} x (SHL(Q|L)const [1] y)) => (LEA(Q|L)2 [c] {s} x y)
   971(LEA(Q|L)1 [c] {s} x (SHL(Q|L)const [2] y)) => (LEA(Q|L)4 [c] {s} x y)
   972(LEA(Q|L)1 [c] {s} x (SHL(Q|L)const [3] y)) => (LEA(Q|L)8 [c] {s} x y)
   973(LEA(Q|L)2 [c] {s} x (SHL(Q|L)const [1] y)) => (LEA(Q|L)4 [c] {s} x y)
   974(LEA(Q|L)2 [c] {s} x (SHL(Q|L)const [2] y)) => (LEA(Q|L)8 [c] {s} x y)
   975(LEA(Q|L)4 [c] {s} x (SHL(Q|L)const [1] y)) => (LEA(Q|L)8 [c] {s} x y)
   976
   977// reverse ordering of compare instruction
   978(SETL (InvertFlags x)) => (SETG x)
   979(SETG (InvertFlags x)) => (SETL x)
   980(SETB (InvertFlags x)) => (SETA x)
   981(SETA (InvertFlags x)) => (SETB x)
   982(SETLE (InvertFlags x)) => (SETGE x)
   983(SETGE (InvertFlags x)) => (SETLE x)
   984(SETBE (InvertFlags x)) => (SETAE x)
   985(SETAE (InvertFlags x)) => (SETBE x)
   986(SETEQ (InvertFlags x)) => (SETEQ x)
   987(SETNE (InvertFlags x)) => (SETNE x)
   988
   989(SETLstore [off] {sym} ptr (InvertFlags x) mem) => (SETGstore [off] {sym} ptr x mem)
   990(SETGstore [off] {sym} ptr (InvertFlags x) mem) => (SETLstore [off] {sym} ptr x mem)
   991(SETBstore [off] {sym} ptr (InvertFlags x) mem) => (SETAstore [off] {sym} ptr x mem)
   992(SETAstore [off] {sym} ptr (InvertFlags x) mem) => (SETBstore [off] {sym} ptr x mem)
   993(SETLEstore [off] {sym} ptr (InvertFlags x) mem) => (SETGEstore [off] {sym} ptr x mem)
   994(SETGEstore [off] {sym} ptr (InvertFlags x) mem) => (SETLEstore [off] {sym} ptr x mem)
   995(SETBEstore [off] {sym} ptr (InvertFlags x) mem) => (SETAEstore [off] {sym} ptr x mem)
   996(SETAEstore [off] {sym} ptr (InvertFlags x) mem) => (SETBEstore [off] {sym} ptr x mem)
   997(SETEQstore [off] {sym} ptr (InvertFlags x) mem) => (SETEQstore [off] {sym} ptr x mem)
   998(SETNEstore [off] {sym} ptr (InvertFlags x) mem) => (SETNEstore [off] {sym} ptr x mem)
   999
  1000// sign extended loads
  1001// Note: The combined instruction must end up in the same block
  1002// as the original load. If not, we end up making a value with
  1003// memory type live in two different blocks, which can lead to
  1004// multiple memory values alive simultaneously.
  1005// Make sure we don't combine these ops if the load has another use.
  1006// This prevents a single load from being split into multiple loads
  1007// which then might return different values.  See test/atomicload.go.
  1008(MOVBQSX x:(MOVBload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVBQSXload <v.Type> [off] {sym} ptr mem)
  1009(MOVBQSX x:(MOVWload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVBQSXload <v.Type> [off] {sym} ptr mem)
  1010(MOVBQSX x:(MOVLload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVBQSXload <v.Type> [off] {sym} ptr mem)
  1011(MOVBQSX x:(MOVQload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVBQSXload <v.Type> [off] {sym} ptr mem)
  1012(MOVBQZX x:(MOVBload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVBload <v.Type> [off] {sym} ptr mem)
  1013(MOVBQZX x:(MOVWload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVBload <v.Type> [off] {sym} ptr mem)
  1014(MOVBQZX x:(MOVLload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVBload <v.Type> [off] {sym} ptr mem)
  1015(MOVBQZX x:(MOVQload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVBload <v.Type> [off] {sym} ptr mem)
  1016(MOVWQSX x:(MOVWload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVWQSXload <v.Type> [off] {sym} ptr mem)
  1017(MOVWQSX x:(MOVLload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVWQSXload <v.Type> [off] {sym} ptr mem)
  1018(MOVWQSX x:(MOVQload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVWQSXload <v.Type> [off] {sym} ptr mem)
  1019(MOVWQZX x:(MOVWload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVWload <v.Type> [off] {sym} ptr mem)
  1020(MOVWQZX x:(MOVLload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVWload <v.Type> [off] {sym} ptr mem)
  1021(MOVWQZX x:(MOVQload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVWload <v.Type> [off] {sym} ptr mem)
  1022(MOVLQSX x:(MOVLload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVLQSXload <v.Type> [off] {sym} ptr mem)
  1023(MOVLQSX x:(MOVQload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVLQSXload <v.Type> [off] {sym} ptr mem)
  1024(MOVLQZX x:(MOVLload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVLload <v.Type> [off] {sym} ptr mem)
  1025(MOVLQZX x:(MOVQload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVLload <v.Type> [off] {sym} ptr mem)
  1026
  1027// replace load from same location as preceding store with zero/sign extension (or copy in case of full width)
  1028(MOVBload [off] {sym} ptr (MOVBstore [off2] {sym2} ptr2 x _)) && sym == sym2 && off == off2 && isSamePtr(ptr, ptr2) => (MOVBQZX x)
  1029(MOVWload [off] {sym} ptr (MOVWstore [off2] {sym2} ptr2 x _)) && sym == sym2 && off == off2 && isSamePtr(ptr, ptr2) => (MOVWQZX x)
  1030(MOVLload [off] {sym} ptr (MOVLstore [off2] {sym2} ptr2 x _)) && sym == sym2 && off == off2 && isSamePtr(ptr, ptr2) => (MOVLQZX x)
  1031(MOVQload [off] {sym} ptr (MOVQstore [off2] {sym2} ptr2 x _)) && sym == sym2 && off == off2 && isSamePtr(ptr, ptr2) => x
  1032(MOVBQSXload [off] {sym} ptr (MOVBstore [off2] {sym2} ptr2 x _)) && sym == sym2 && off == off2 && isSamePtr(ptr, ptr2) => (MOVBQSX x)
  1033(MOVWQSXload [off] {sym} ptr (MOVWstore [off2] {sym2} ptr2 x _)) && sym == sym2 && off == off2 && isSamePtr(ptr, ptr2) => (MOVWQSX x)
  1034(MOVLQSXload [off] {sym} ptr (MOVLstore [off2] {sym2} ptr2 x _)) && sym == sym2 && off == off2 && isSamePtr(ptr, ptr2) => (MOVLQSX x)
  1035
  1036// Fold extensions and ANDs together.
  1037(MOVBQZX (ANDLconst [c] x)) => (ANDLconst [c & 0xff] x)
  1038(MOVWQZX (ANDLconst [c] x)) => (ANDLconst [c & 0xffff] x)
  1039(MOVLQZX (ANDLconst [c] x)) => (ANDLconst [c] x)
  1040(MOVBQSX (ANDLconst [c] x)) && c & 0x80 == 0 => (ANDLconst [c & 0x7f] x)
  1041(MOVWQSX (ANDLconst [c] x)) && c & 0x8000 == 0 => (ANDLconst [c & 0x7fff] x)
  1042(MOVLQSX (ANDLconst [c] x)) && uint32(c) & 0x80000000 == 0 => (ANDLconst [c & 0x7fffffff] x)
  1043
  1044// Don't extend before storing
  1045(MOVLstore [off] {sym} ptr (MOVLQSX x) mem) => (MOVLstore [off] {sym} ptr x mem)
  1046(MOVWstore [off] {sym} ptr (MOVWQSX x) mem) => (MOVWstore [off] {sym} ptr x mem)
  1047(MOVBstore [off] {sym} ptr (MOVBQSX x) mem) => (MOVBstore [off] {sym} ptr x mem)
  1048(MOVLstore [off] {sym} ptr (MOVLQZX x) mem) => (MOVLstore [off] {sym} ptr x mem)
  1049(MOVWstore [off] {sym} ptr (MOVWQZX x) mem) => (MOVWstore [off] {sym} ptr x mem)
  1050(MOVBstore [off] {sym} ptr (MOVBQZX x) mem) => (MOVBstore [off] {sym} ptr x mem)
  1051
  1052// fold constants into memory operations
  1053// Note that this is not always a good idea because if not all the uses of
  1054// the ADDQconst get eliminated, we still have to compute the ADDQconst and we now
  1055// have potentially two live values (ptr and (ADDQconst [off] ptr)) instead of one.
  1056// Nevertheless, let's do it!
  1057(MOV(Q|L|W|B|SS|SD|O)load  [off1] {sym} (ADDQconst [off2] ptr) mem) && is32Bit(int64(off1)+int64(off2)) =>
  1058    (MOV(Q|L|W|B|SS|SD|O)load  [off1+off2] {sym} ptr mem)
  1059(MOV(Q|L|W|B|SS|SD|O)store  [off1] {sym} (ADDQconst [off2] ptr) val mem) && is32Bit(int64(off1)+int64(off2)) =>
  1060	(MOV(Q|L|W|B|SS|SD|O)store  [off1+off2] {sym} ptr val mem)
  1061(SET(L|G|B|A|LE|GE|BE|AE|EQ|NE)store [off1] {sym} (ADDQconst [off2] base) val mem) && is32Bit(int64(off1)+int64(off2)) =>
  1062	(SET(L|G|B|A|LE|GE|BE|AE|EQ|NE)store [off1+off2] {sym} base val mem)
  1063((ADD|SUB|AND|OR|XOR)Qload [off1] {sym} val (ADDQconst [off2] base) mem) && is32Bit(int64(off1)+int64(off2)) =>
  1064	((ADD|SUB|AND|OR|XOR)Qload [off1+off2] {sym} val base mem)
  1065((ADD|SUB|AND|OR|XOR)Lload [off1] {sym} val (ADDQconst [off2] base) mem) && is32Bit(int64(off1)+int64(off2)) =>
  1066	((ADD|SUB|AND|OR|XOR)Lload [off1+off2] {sym} val base mem)
  1067(CMP(Q|L|W|B)load [off1] {sym} (ADDQconst [off2] base) val mem) && is32Bit(int64(off1)+int64(off2)) =>
  1068	(CMP(Q|L|W|B)load [off1+off2] {sym} base val mem)
  1069(CMP(Q|L|W|B)constload [valoff1] {sym} (ADDQconst [off2] base) mem) && ValAndOff(valoff1).canAdd32(off2) =>
  1070	(CMP(Q|L|W|B)constload [ValAndOff(valoff1).addOffset32(off2)] {sym} base mem)
  1071
  1072((ADD|SUB|MUL|DIV)SSload [off1] {sym} val (ADDQconst [off2] base) mem) && is32Bit(int64(off1)+int64(off2)) =>
  1073	((ADD|SUB|MUL|DIV)SSload [off1+off2] {sym} val base mem)
  1074((ADD|SUB|MUL|DIV)SDload [off1] {sym} val (ADDQconst [off2] base) mem) && is32Bit(int64(off1)+int64(off2)) =>
  1075	((ADD|SUB|MUL|DIV)SDload [off1+off2] {sym} val base mem)
  1076((ADD|AND|OR|XOR)Qconstmodify [valoff1] {sym} (ADDQconst [off2] base) mem) && ValAndOff(valoff1).canAdd32(off2) =>
  1077	((ADD|AND|OR|XOR)Qconstmodify [ValAndOff(valoff1).addOffset32(off2)] {sym} base mem)
  1078((ADD|AND|OR|XOR)Lconstmodify [valoff1] {sym} (ADDQconst [off2] base) mem) && ValAndOff(valoff1).canAdd32(off2) =>
  1079	((ADD|AND|OR|XOR)Lconstmodify [ValAndOff(valoff1).addOffset32(off2)] {sym} base mem)
  1080((ADD|SUB|AND|OR|XOR)Qmodify [off1] {sym} (ADDQconst [off2] base) val mem) && is32Bit(int64(off1)+int64(off2)) =>
  1081	((ADD|SUB|AND|OR|XOR)Qmodify [off1+off2] {sym} base val mem)
  1082((ADD|SUB|AND|OR|XOR)Lmodify [off1] {sym} (ADDQconst [off2] base) val mem) && is32Bit(int64(off1)+int64(off2)) =>
  1083	((ADD|SUB|AND|OR|XOR)Lmodify [off1+off2] {sym} base val mem)
  1084
  1085// Fold constants into stores.
  1086(MOVQstore [off] {sym} ptr (MOVQconst [c]) mem) && validVal(c) =>
  1087	(MOVQstoreconst [makeValAndOff(int32(c),off)] {sym} ptr mem)
  1088(MOVLstore [off] {sym} ptr (MOV(L|Q)const [c]) mem) =>
  1089	(MOVLstoreconst [makeValAndOff(int32(c),off)] {sym} ptr mem)
  1090(MOVWstore [off] {sym} ptr (MOV(L|Q)const [c]) mem) =>
  1091	(MOVWstoreconst [makeValAndOff(int32(int16(c)),off)] {sym} ptr mem)
  1092(MOVBstore [off] {sym} ptr (MOV(L|Q)const [c]) mem) =>
  1093	(MOVBstoreconst [makeValAndOff(int32(int8(c)),off)] {sym} ptr mem)
  1094
  1095// Fold address offsets into constant stores.
  1096(MOV(Q|L|W|B|O)storeconst [sc] {s} (ADDQconst [off] ptr) mem) && ValAndOff(sc).canAdd32(off) =>
  1097	(MOV(Q|L|W|B|O)storeconst [ValAndOff(sc).addOffset32(off)] {s} ptr mem)
  1098
  1099// We need to fold LEAQ into the MOVx ops so that the live variable analysis knows
  1100// what variables are being read/written by the ops.
  1101(MOV(Q|L|W|B|SS|SD|O|BQSX|WQSX|LQSX)load [off1] {sym1} (LEAQ [off2] {sym2} base) mem)
  1102	&& is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1103	(MOV(Q|L|W|B|SS|SD|O|BQSX|WQSX|LQSX)load [off1+off2] {mergeSym(sym1,sym2)} base mem)
  1104(MOV(Q|L|W|B|SS|SD|O)store [off1] {sym1} (LEAQ [off2] {sym2} base) val mem)
  1105	&& is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1106	(MOV(Q|L|W|B|SS|SD|O)store [off1+off2] {mergeSym(sym1,sym2)} base val mem)
  1107(MOV(Q|L|W|B|O)storeconst [sc] {sym1} (LEAQ [off] {sym2} ptr) mem) && canMergeSym(sym1, sym2) && ValAndOff(sc).canAdd32(off) =>
  1108	(MOV(Q|L|W|B|O)storeconst [ValAndOff(sc).addOffset32(off)] {mergeSym(sym1, sym2)} ptr mem)
  1109(SET(L|G|B|A|LE|GE|BE|AE|EQ|NE)store [off1] {sym1} (LEAQ [off2] {sym2} base) val mem)
  1110	&& is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1111	(SET(L|G|B|A|LE|GE|BE|AE|EQ|NE)store [off1+off2] {mergeSym(sym1,sym2)} base val mem)
  1112((ADD|SUB|AND|OR|XOR)Qload [off1] {sym1} val (LEAQ [off2] {sym2} base) mem)
  1113	&& is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1114	((ADD|SUB|AND|OR|XOR)Qload [off1+off2] {mergeSym(sym1,sym2)} val base mem)
  1115((ADD|SUB|AND|OR|XOR)Lload [off1] {sym1} val (LEAQ [off2] {sym2} base) mem)
  1116	&& is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1117	((ADD|SUB|AND|OR|XOR)Lload [off1+off2] {mergeSym(sym1,sym2)} val base mem)
  1118(CMP(Q|L|W|B)load [off1] {sym1} (LEAQ [off2] {sym2} base) val mem)
  1119	&& is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1120	(CMP(Q|L|W|B)load [off1+off2] {mergeSym(sym1,sym2)} base val mem)
  1121(CMP(Q|L|W|B)constload [valoff1] {sym1} (LEAQ [off2] {sym2} base) mem)
  1122	&& ValAndOff(valoff1).canAdd32(off2) && canMergeSym(sym1, sym2) =>
  1123	(CMP(Q|L|W|B)constload [ValAndOff(valoff1).addOffset32(off2)] {mergeSym(sym1,sym2)} base mem)
  1124
  1125((ADD|SUB|MUL|DIV)SSload [off1] {sym1} val (LEAQ [off2] {sym2} base) mem)
  1126	&& is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1127	((ADD|SUB|MUL|DIV)SSload [off1+off2] {mergeSym(sym1,sym2)} val base mem)
  1128((ADD|SUB|MUL|DIV)SDload [off1] {sym1} val (LEAQ [off2] {sym2} base) mem)
  1129	&& is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1130	((ADD|SUB|MUL|DIV)SDload [off1+off2] {mergeSym(sym1,sym2)} val base mem)
  1131((ADD|AND|OR|XOR)Qconstmodify [valoff1] {sym1} (LEAQ [off2] {sym2} base) mem)
  1132	&& ValAndOff(valoff1).canAdd32(off2) && canMergeSym(sym1, sym2) =>
  1133	((ADD|AND|OR|XOR)Qconstmodify [ValAndOff(valoff1).addOffset32(off2)] {mergeSym(sym1,sym2)} base mem)
  1134((ADD|AND|OR|XOR)Lconstmodify [valoff1] {sym1} (LEAQ [off2] {sym2} base) mem)
  1135	&& ValAndOff(valoff1).canAdd32(off2) && canMergeSym(sym1, sym2) =>
  1136	((ADD|AND|OR|XOR)Lconstmodify [ValAndOff(valoff1).addOffset32(off2)] {mergeSym(sym1,sym2)} base mem)
  1137((ADD|SUB|AND|OR|XOR)Qmodify [off1] {sym1} (LEAQ [off2] {sym2} base) val mem)
  1138	&& is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1139	((ADD|SUB|AND|OR|XOR)Qmodify [off1+off2] {mergeSym(sym1,sym2)} base val mem)
  1140((ADD|SUB|AND|OR|XOR)Lmodify [off1] {sym1} (LEAQ [off2] {sym2} base) val mem)
  1141	&& is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1142	((ADD|SUB|AND|OR|XOR)Lmodify [off1+off2] {mergeSym(sym1,sym2)} base val mem)
  1143
  1144// fold LEAQs together
  1145(LEAQ [off1] {sym1} (LEAQ [off2] {sym2} x)) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1146      (LEAQ [off1+off2] {mergeSym(sym1,sym2)} x)
  1147
  1148// LEAQ into LEAQ1
  1149(LEAQ1 [off1] {sym1} (LEAQ [off2] {sym2} x) y) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) && x.Op != OpSB =>
  1150       (LEAQ1 [off1+off2] {mergeSym(sym1,sym2)} x y)
  1151
  1152// LEAQ1 into LEAQ
  1153(LEAQ [off1] {sym1} (LEAQ1 [off2] {sym2} x y)) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1154       (LEAQ1 [off1+off2] {mergeSym(sym1,sym2)} x y)
  1155
  1156// LEAQ into LEAQ[248]
  1157(LEAQ2 [off1] {sym1} (LEAQ [off2] {sym2} x) y) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) && x.Op != OpSB =>
  1158       (LEAQ2 [off1+off2] {mergeSym(sym1,sym2)} x y)
  1159(LEAQ4 [off1] {sym1} (LEAQ [off2] {sym2} x) y) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) && x.Op != OpSB =>
  1160       (LEAQ4 [off1+off2] {mergeSym(sym1,sym2)} x y)
  1161(LEAQ8 [off1] {sym1} (LEAQ [off2] {sym2} x) y) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) && x.Op != OpSB =>
  1162       (LEAQ8 [off1+off2] {mergeSym(sym1,sym2)} x y)
  1163
  1164// LEAQ[248] into LEAQ
  1165(LEAQ [off1] {sym1} (LEAQ2 [off2] {sym2} x y)) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1166      (LEAQ2 [off1+off2] {mergeSym(sym1,sym2)} x y)
  1167(LEAQ [off1] {sym1} (LEAQ4 [off2] {sym2} x y)) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1168      (LEAQ4 [off1+off2] {mergeSym(sym1,sym2)} x y)
  1169(LEAQ [off1] {sym1} (LEAQ8 [off2] {sym2} x y)) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1170      (LEAQ8 [off1+off2] {mergeSym(sym1,sym2)} x y)
  1171
  1172// LEAQ[1248] into LEAQ[1248]. Only some such merges are possible.
  1173(LEAQ1 [off1] {sym1} x (LEAQ1 [off2] {sym2} y y)) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1174      (LEAQ2 [off1+off2] {mergeSym(sym1, sym2)} x y)
  1175(LEAQ1 [off1] {sym1} x (LEAQ1 [off2] {sym2} x y)) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1176      (LEAQ2 [off1+off2] {mergeSym(sym1, sym2)} y x)
  1177(LEAQ2 [off1] {sym1} x (LEAQ1 [off2] {sym2} y y)) && is32Bit(int64(off1)+2*int64(off2)) && sym2 == nil =>
  1178      (LEAQ4 [off1+2*off2] {sym1} x y)
  1179(LEAQ4 [off1] {sym1} x (LEAQ1 [off2] {sym2} y y)) && is32Bit(int64(off1)+4*int64(off2)) && sym2 == nil =>
  1180      (LEAQ8 [off1+4*off2] {sym1} x y)
  1181// TODO: more?
  1182
  1183// Lower LEAQ2/4/8 when the offset is a constant
  1184(LEAQ2 [off] {sym} x (MOV(Q|L)const [scale])) && is32Bit(int64(off)+int64(scale)*2) =>
  1185	(LEAQ [off+int32(scale)*2] {sym} x)
  1186(LEAQ4 [off] {sym} x (MOV(Q|L)const [scale])) && is32Bit(int64(off)+int64(scale)*4) =>
  1187	(LEAQ [off+int32(scale)*4] {sym} x)
  1188(LEAQ8 [off] {sym} x (MOV(Q|L)const [scale])) && is32Bit(int64(off)+int64(scale)*8) =>
  1189	(LEAQ [off+int32(scale)*8] {sym} x)
  1190
  1191// Absorb InvertFlags into branches.
  1192(LT (InvertFlags cmp) yes no) => (GT cmp yes no)
  1193(GT (InvertFlags cmp) yes no) => (LT cmp yes no)
  1194(LE (InvertFlags cmp) yes no) => (GE cmp yes no)
  1195(GE (InvertFlags cmp) yes no) => (LE cmp yes no)
  1196(ULT (InvertFlags cmp) yes no) => (UGT cmp yes no)
  1197(UGT (InvertFlags cmp) yes no) => (ULT cmp yes no)
  1198(ULE (InvertFlags cmp) yes no) => (UGE cmp yes no)
  1199(UGE (InvertFlags cmp) yes no) => (ULE cmp yes no)
  1200(EQ (InvertFlags cmp) yes no) => (EQ cmp yes no)
  1201(NE (InvertFlags cmp) yes no) => (NE cmp yes no)
  1202
  1203// Constant comparisons.
  1204(CMPQconst (MOVQconst [x]) [y]) && x==int64(y) => (FlagEQ)
  1205(CMPQconst (MOVQconst [x]) [y]) && x<int64(y) && uint64(x)<uint64(int64(y)) => (FlagLT_ULT)
  1206(CMPQconst (MOVQconst [x]) [y]) && x<int64(y) && uint64(x)>uint64(int64(y)) => (FlagLT_UGT)
  1207(CMPQconst (MOVQconst [x]) [y]) && x>int64(y) && uint64(x)<uint64(int64(y)) => (FlagGT_ULT)
  1208(CMPQconst (MOVQconst [x]) [y]) && x>int64(y) && uint64(x)>uint64(int64(y)) => (FlagGT_UGT)
  1209(CMPLconst (MOVLconst [x]) [y]) && x==y => (FlagEQ)
  1210(CMPLconst (MOVLconst [x]) [y]) && x<y && uint32(x)<uint32(y) => (FlagLT_ULT)
  1211(CMPLconst (MOVLconst [x]) [y]) && x<y && uint32(x)>uint32(y) => (FlagLT_UGT)
  1212(CMPLconst (MOVLconst [x]) [y]) && x>y && uint32(x)<uint32(y) => (FlagGT_ULT)
  1213(CMPLconst (MOVLconst [x]) [y]) && x>y && uint32(x)>uint32(y) => (FlagGT_UGT)
  1214(CMPWconst (MOVLconst [x]) [y]) && int16(x)==y => (FlagEQ)
  1215(CMPWconst (MOVLconst [x]) [y]) && int16(x)<y && uint16(x)<uint16(y) => (FlagLT_ULT)
  1216(CMPWconst (MOVLconst [x]) [y]) && int16(x)<y && uint16(x)>uint16(y) => (FlagLT_UGT)
  1217(CMPWconst (MOVLconst [x]) [y]) && int16(x)>y && uint16(x)<uint16(y) => (FlagGT_ULT)
  1218(CMPWconst (MOVLconst [x]) [y]) && int16(x)>y && uint16(x)>uint16(y) => (FlagGT_UGT)
  1219(CMPBconst (MOVLconst [x]) [y]) && int8(x)==y => (FlagEQ)
  1220(CMPBconst (MOVLconst [x]) [y]) && int8(x)<y && uint8(x)<uint8(y) => (FlagLT_ULT)
  1221(CMPBconst (MOVLconst [x]) [y]) && int8(x)<y && uint8(x)>uint8(y) => (FlagLT_UGT)
  1222(CMPBconst (MOVLconst [x]) [y]) && int8(x)>y && uint8(x)<uint8(y) => (FlagGT_ULT)
  1223(CMPBconst (MOVLconst [x]) [y]) && int8(x)>y && uint8(x)>uint8(y) => (FlagGT_UGT)
  1224
  1225// CMPQconst requires a 32 bit const, but we can still constant-fold 64 bit consts.
  1226// In theory this applies to any of the simplifications above,
  1227// but CMPQ is the only one I've actually seen occur.
  1228(CMPQ (MOVQconst [x]) (MOVQconst [y])) && x==y => (FlagEQ)
  1229(CMPQ (MOVQconst [x]) (MOVQconst [y])) && x<y && uint64(x)<uint64(y) => (FlagLT_ULT)
  1230(CMPQ (MOVQconst [x]) (MOVQconst [y])) && x<y && uint64(x)>uint64(y) => (FlagLT_UGT)
  1231(CMPQ (MOVQconst [x]) (MOVQconst [y])) && x>y && uint64(x)<uint64(y) => (FlagGT_ULT)
  1232(CMPQ (MOVQconst [x]) (MOVQconst [y])) && x>y && uint64(x)>uint64(y) => (FlagGT_UGT)
  1233
  1234// Other known comparisons.
  1235(CMPQconst (MOVBQZX _) [c]) && 0xFF < c => (FlagLT_ULT)
  1236(CMPQconst (MOVWQZX _) [c]) && 0xFFFF < c => (FlagLT_ULT)
  1237(CMPLconst (SHRLconst _ [c]) [n]) && 0 <= n && 0 < c && c <= 32 && (1<<uint64(32-c)) <= uint64(n) => (FlagLT_ULT)
  1238(CMPQconst (SHRQconst _ [c]) [n]) && 0 <= n && 0 < c && c <= 64 && (1<<uint64(64-c)) <= uint64(n) => (FlagLT_ULT)
  1239(CMPQconst (ANDQconst _ [m]) [n]) && 0 <= m && m < n => (FlagLT_ULT)
  1240(CMPQconst (ANDLconst _ [m]) [n]) && 0 <= m && m < n => (FlagLT_ULT)
  1241(CMPLconst (ANDLconst _ [m]) [n]) && 0 <= m && m < n => (FlagLT_ULT)
  1242(CMPWconst (ANDLconst _ [m]) [n]) && 0 <= int16(m) && int16(m) < n => (FlagLT_ULT)
  1243(CMPBconst (ANDLconst _ [m]) [n]) && 0 <= int8(m)  && int8(m)  < n => (FlagLT_ULT)
  1244
  1245// TESTQ c c sets flags like CMPQ c 0.
  1246(TESTQconst [c] (MOVQconst [d])) && int64(c) == d && c == 0 => (FlagEQ)
  1247(TESTLconst [c] (MOVLconst [c])) && c == 0 => (FlagEQ)
  1248(TESTQconst [c] (MOVQconst [d])) && int64(c) == d && c < 0  => (FlagLT_UGT)
  1249(TESTLconst [c] (MOVLconst [c])) && c < 0  => (FlagLT_UGT)
  1250(TESTQconst [c] (MOVQconst [d])) && int64(c) == d && c > 0  => (FlagGT_UGT)
  1251(TESTLconst [c] (MOVLconst [c])) && c > 0  => (FlagGT_UGT)
  1252
  1253// TODO: DIVxU also.
  1254
  1255// Absorb flag constants into SBB ops.
  1256(SBBQcarrymask (FlagEQ))     => (MOVQconst [0])
  1257(SBBQcarrymask (FlagLT_ULT)) => (MOVQconst [-1])
  1258(SBBQcarrymask (FlagLT_UGT)) => (MOVQconst [0])
  1259(SBBQcarrymask (FlagGT_ULT)) => (MOVQconst [-1])
  1260(SBBQcarrymask (FlagGT_UGT)) => (MOVQconst [0])
  1261(SBBLcarrymask (FlagEQ))     => (MOVLconst [0])
  1262(SBBLcarrymask (FlagLT_ULT)) => (MOVLconst [-1])
  1263(SBBLcarrymask (FlagLT_UGT)) => (MOVLconst [0])
  1264(SBBLcarrymask (FlagGT_ULT)) => (MOVLconst [-1])
  1265(SBBLcarrymask (FlagGT_UGT)) => (MOVLconst [0])
  1266
  1267// Absorb flag constants into branches.
  1268((EQ|LE|GE|ULE|UGE) (FlagEQ) yes no)     => (First yes no)
  1269((NE|LT|GT|ULT|UGT) (FlagEQ) yes no)     => (First no yes)
  1270((NE|LT|LE|ULT|ULE) (FlagLT_ULT) yes no) => (First yes no)
  1271((EQ|GT|GE|UGT|UGE) (FlagLT_ULT) yes no) => (First no yes)
  1272((NE|LT|LE|UGT|UGE) (FlagLT_UGT) yes no) => (First yes no)
  1273((EQ|GT|GE|ULT|ULE) (FlagLT_UGT) yes no) => (First no yes)
  1274((NE|GT|GE|ULT|ULE) (FlagGT_ULT) yes no) => (First yes no)
  1275((EQ|LT|LE|UGT|UGE) (FlagGT_ULT) yes no) => (First no yes)
  1276((NE|GT|GE|UGT|UGE) (FlagGT_UGT) yes no) => (First yes no)
  1277((EQ|LT|LE|ULT|ULE) (FlagGT_UGT) yes no) => (First no yes)
  1278
  1279// Absorb flag constants into SETxx ops.
  1280((SETEQ|SETLE|SETGE|SETBE|SETAE) (FlagEQ))     => (MOVLconst [1])
  1281((SETNE|SETL|SETG|SETB|SETA)     (FlagEQ))     => (MOVLconst [0])
  1282((SETNE|SETL|SETLE|SETB|SETBE)   (FlagLT_ULT)) => (MOVLconst [1])
  1283((SETEQ|SETG|SETGE|SETA|SETAE)   (FlagLT_ULT)) => (MOVLconst [0])
  1284((SETNE|SETL|SETLE|SETA|SETAE)   (FlagLT_UGT)) => (MOVLconst [1])
  1285((SETEQ|SETG|SETGE|SETB|SETBE)   (FlagLT_UGT)) => (MOVLconst [0])
  1286((SETNE|SETG|SETGE|SETB|SETBE)   (FlagGT_ULT)) => (MOVLconst [1])
  1287((SETEQ|SETL|SETLE|SETA|SETAE)   (FlagGT_ULT)) => (MOVLconst [0])
  1288((SETNE|SETG|SETGE|SETA|SETAE)   (FlagGT_UGT)) => (MOVLconst [1])
  1289((SETEQ|SETL|SETLE|SETB|SETBE)   (FlagGT_UGT)) => (MOVLconst [0])
  1290
  1291(SETEQstore [off] {sym} ptr (FlagEQ)     mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1292(SETEQstore [off] {sym} ptr (FlagLT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1293(SETEQstore [off] {sym} ptr (FlagLT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1294(SETEQstore [off] {sym} ptr (FlagGT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1295(SETEQstore [off] {sym} ptr (FlagGT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1296
  1297(SETNEstore [off] {sym} ptr (FlagEQ)     mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1298(SETNEstore [off] {sym} ptr (FlagLT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1299(SETNEstore [off] {sym} ptr (FlagLT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1300(SETNEstore [off] {sym} ptr (FlagGT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1301(SETNEstore [off] {sym} ptr (FlagGT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1302
  1303(SETLstore  [off] {sym} ptr (FlagEQ)     mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1304(SETLstore  [off] {sym} ptr (FlagLT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1305(SETLstore  [off] {sym} ptr (FlagLT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1306(SETLstore  [off] {sym} ptr (FlagGT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1307(SETLstore  [off] {sym} ptr (FlagGT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1308
  1309(SETLEstore [off] {sym} ptr (FlagEQ)     mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1310(SETLEstore [off] {sym} ptr (FlagLT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1311(SETLEstore [off] {sym} ptr (FlagLT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1312(SETLEstore [off] {sym} ptr (FlagGT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1313(SETLEstore [off] {sym} ptr (FlagGT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1314
  1315(SETGstore  [off] {sym} ptr (FlagEQ)     mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1316(SETGstore  [off] {sym} ptr (FlagLT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1317(SETGstore  [off] {sym} ptr (FlagLT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1318(SETGstore  [off] {sym} ptr (FlagGT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1319(SETGstore  [off] {sym} ptr (FlagGT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1320
  1321(SETGEstore [off] {sym} ptr (FlagEQ)     mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1322(SETGEstore [off] {sym} ptr (FlagLT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1323(SETGEstore [off] {sym} ptr (FlagLT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1324(SETGEstore [off] {sym} ptr (FlagGT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1325(SETGEstore [off] {sym} ptr (FlagGT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1326
  1327(SETBstore  [off] {sym} ptr (FlagEQ)     mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1328(SETBstore  [off] {sym} ptr (FlagLT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1329(SETBstore  [off] {sym} ptr (FlagLT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1330(SETBstore  [off] {sym} ptr (FlagGT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1331(SETBstore  [off] {sym} ptr (FlagGT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1332
  1333(SETBEstore [off] {sym} ptr (FlagEQ)     mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1334(SETBEstore [off] {sym} ptr (FlagLT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1335(SETBEstore [off] {sym} ptr (FlagLT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1336(SETBEstore [off] {sym} ptr (FlagGT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1337(SETBEstore [off] {sym} ptr (FlagGT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1338
  1339(SETAstore  [off] {sym} ptr (FlagEQ)     mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1340(SETAstore  [off] {sym} ptr (FlagLT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1341(SETAstore  [off] {sym} ptr (FlagLT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1342(SETAstore  [off] {sym} ptr (FlagGT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1343(SETAstore  [off] {sym} ptr (FlagGT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1344
  1345(SETAEstore [off] {sym} ptr (FlagEQ)     mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1346(SETAEstore [off] {sym} ptr (FlagLT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1347(SETAEstore [off] {sym} ptr (FlagLT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1348(SETAEstore [off] {sym} ptr (FlagGT_ULT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [0]) mem)
  1349(SETAEstore [off] {sym} ptr (FlagGT_UGT) mem) => (MOVBstore [off] {sym} ptr (MOVLconst <typ.UInt8> [1]) mem)
  1350
  1351// Remove redundant *const ops
  1352(ADDQconst [0] x)          => x
  1353(ADDLconst [c] x) && c==0  => x
  1354(SUBQconst [0] x)          => x
  1355(SUBLconst [c] x) && c==0  => x
  1356(ANDQconst [0] _)          => (MOVQconst [0])
  1357(ANDLconst [c] _) && c==0  => (MOVLconst [0])
  1358(ANDQconst [-1] x)         => x
  1359(ANDLconst [c] x) && c==-1 => x
  1360(ORQconst [0] x)           => x
  1361(ORLconst [c] x)  && c==0  => x
  1362(ORQconst [-1] _)          => (MOVQconst [-1])
  1363(ORLconst [c] _)  && c==-1 => (MOVLconst [-1])
  1364(XORQconst [0] x)          => x
  1365(XORLconst [c] x) && c==0  => x
  1366// TODO: since we got rid of the W/B versions, we might miss
  1367// things like (ANDLconst [0x100] x) which were formerly
  1368// (ANDBconst [0] x).  Probably doesn't happen very often.
  1369// If we cared, we might do:
  1370//  (ANDLconst <t> [c] x) && t.Size()==1 && int8(x)==0 -> (MOVLconst [0])
  1371
  1372// Remove redundant ops
  1373// Not in generic rules, because they may appear after lowering e. g. Slicemask
  1374(NEG(Q|L) (NEG(Q|L) x)) => x
  1375(NEG(Q|L) s:(SUB(Q|L) x y)) && s.Uses == 1 => (SUB(Q|L) y x)
  1376
  1377// Convert constant subtracts to constant adds
  1378(SUBQconst [c] x) && c != -(1<<31) => (ADDQconst [-c] x)
  1379(SUBLconst [c] x) => (ADDLconst [-c] x)
  1380
  1381// generic constant folding
  1382// TODO: more of this
  1383(ADDQconst [c] (MOVQconst [d])) => (MOVQconst [int64(c)+d])
  1384(ADDLconst [c] (MOVLconst [d])) => (MOVLconst [c+d])
  1385(ADDQconst [c] (ADDQconst [d] x)) && is32Bit(int64(c)+int64(d)) => (ADDQconst [c+d] x)
  1386(ADDLconst [c] (ADDLconst [d] x)) => (ADDLconst [c+d] x)
  1387(SUBQconst (MOVQconst [d]) [c]) => (MOVQconst [d-int64(c)])
  1388(SUBQconst (SUBQconst x [d]) [c]) && is32Bit(int64(-c)-int64(d)) => (ADDQconst [-c-d] x)
  1389(SARQconst [c] (MOVQconst [d])) => (MOVQconst [d>>uint64(c)])
  1390(SARLconst [c] (MOVQconst [d])) => (MOVQconst [int64(int32(d))>>uint64(c)])
  1391(SARWconst [c] (MOVQconst [d])) => (MOVQconst [int64(int16(d))>>uint64(c)])
  1392(SARBconst [c] (MOVQconst [d])) => (MOVQconst [int64(int8(d))>>uint64(c)])
  1393(NEGQ (MOVQconst [c])) => (MOVQconst [-c])
  1394(NEGL (MOVLconst [c])) => (MOVLconst [-c])
  1395(MULQconst [c] (MOVQconst [d])) => (MOVQconst [int64(c)*d])
  1396(MULLconst [c] (MOVLconst [d])) => (MOVLconst [c*d])
  1397(ANDQconst [c] (MOVQconst [d])) => (MOVQconst [int64(c)&d])
  1398(ANDLconst [c] (MOVLconst [d])) => (MOVLconst [c&d])
  1399(ORQconst [c] (MOVQconst [d])) => (MOVQconst [int64(c)|d])
  1400(ORLconst [c] (MOVLconst [d])) => (MOVLconst [c|d])
  1401(XORQconst [c] (MOVQconst [d])) => (MOVQconst [int64(c)^d])
  1402(XORLconst [c] (MOVLconst [d])) => (MOVLconst [c^d])
  1403(NOTQ (MOVQconst [c])) => (MOVQconst [^c])
  1404(NOTL (MOVLconst [c])) => (MOVLconst [^c])
  1405(BTSQconst [c] (MOVQconst [d])) => (MOVQconst [d|(1<<uint32(c))])
  1406(BTRQconst [c] (MOVQconst [d])) => (MOVQconst [d&^(1<<uint32(c))])
  1407(BTCQconst [c] (MOVQconst [d])) => (MOVQconst [d^(1<<uint32(c))])
  1408
  1409// If c or d doesn't fit into 32 bits, then we can't construct ORQconst,
  1410// but we can still constant-fold.
  1411// In theory this applies to any of the simplifications above,
  1412// but ORQ is the only one I've actually seen occur.
  1413(ORQ (MOVQconst [c]) (MOVQconst [d])) => (MOVQconst [c|d])
  1414
  1415// generic simplifications
  1416// TODO: more of this
  1417(ADDQ x (NEGQ y)) => (SUBQ x y)
  1418(ADDL x (NEGL y)) => (SUBL x y)
  1419(SUBQ x x) => (MOVQconst [0])
  1420(SUBL x x) => (MOVLconst [0])
  1421(ANDQ x x) => x
  1422(ANDL x x) => x
  1423(ORQ x x)  => x
  1424(ORL x x)  => x
  1425(XORQ x x) => (MOVQconst [0])
  1426(XORL x x) => (MOVLconst [0])
  1427
  1428(SHLLconst [d] (MOVLconst [c])) => (MOVLconst [c << uint64(d)])
  1429(SHLQconst [d] (MOVQconst [c])) => (MOVQconst [c << uint64(d)])
  1430(SHLQconst [d] (MOVLconst [c])) => (MOVQconst [int64(c) << uint64(d)])
  1431
  1432// Fold NEG into ADDconst/MULconst. Take care to keep c in 32 bit range.
  1433(NEGQ (ADDQconst [c] (NEGQ x))) && c != -(1<<31) => (ADDQconst [-c] x)
  1434(MULQconst [c] (NEGQ x)) && c != -(1<<31) => (MULQconst [-c] x)
  1435
  1436// checking AND against 0.
  1437(CMPQconst a:(ANDQ x y) [0]) && a.Uses == 1 => (TESTQ x y)
  1438(CMPLconst a:(ANDL x y) [0]) && a.Uses == 1 => (TESTL x y)
  1439(CMPWconst a:(ANDL x y) [0]) && a.Uses == 1 => (TESTW x y)
  1440(CMPBconst a:(ANDL x y) [0]) && a.Uses == 1 => (TESTB x y)
  1441(CMPQconst a:(ANDQconst [c] x) [0]) && a.Uses == 1 => (TESTQconst [c] x)
  1442(CMPLconst a:(ANDLconst [c] x) [0]) && a.Uses == 1 => (TESTLconst [c] x)
  1443(CMPWconst a:(ANDLconst [c] x) [0]) && a.Uses == 1 => (TESTWconst [int16(c)] x)
  1444(CMPBconst a:(ANDLconst [c] x) [0]) && a.Uses == 1 => (TESTBconst [int8(c)] x)
  1445
  1446// Convert TESTx to TESTxconst if possible.
  1447(TESTQ (MOVQconst [c]) x) && is32Bit(c) => (TESTQconst [int32(c)] x)
  1448(TESTL (MOVLconst [c]) x) => (TESTLconst [c] x)
  1449(TESTW (MOVLconst [c]) x) => (TESTWconst [int16(c)] x)
  1450(TESTB (MOVLconst [c]) x) => (TESTBconst [int8(c)] x)
  1451
  1452// TEST %reg,%reg is shorter than CMP
  1453(CMPQconst x [0]) => (TESTQ x x)
  1454(CMPLconst x [0]) => (TESTL x x)
  1455(CMPWconst x [0]) => (TESTW x x)
  1456(CMPBconst x [0]) => (TESTB x x)
  1457(TESTQconst [-1] x) && x.Op != OpAMD64MOVQconst => (TESTQ x x)
  1458(TESTLconst [-1] x) && x.Op != OpAMD64MOVLconst => (TESTL x x)
  1459(TESTWconst [-1] x) && x.Op != OpAMD64MOVLconst => (TESTW x x)
  1460(TESTBconst [-1] x) && x.Op != OpAMD64MOVLconst => (TESTB x x)
  1461
  1462// Convert LEAQ1 back to ADDQ if we can
  1463(LEAQ1 [0] x y) && v.Aux == nil => (ADDQ x y)
  1464
  1465(MOVQstoreconst [c] {s} p1 x:(MOVQstoreconst [a] {s} p0 mem))
  1466  && config.useSSE
  1467  && x.Uses == 1
  1468  && sequentialAddresses(p0, p1, int64(a.Off()+8-c.Off()))
  1469  && a.Val() == 0
  1470  && c.Val() == 0
  1471  && setPos(v, x.Pos)
  1472  && clobber(x)
  1473  => (MOVOstoreconst [makeValAndOff(0,a.Off())] {s} p0 mem)
  1474(MOVQstoreconst [a] {s} p0 x:(MOVQstoreconst [c] {s} p1 mem))
  1475  && config.useSSE
  1476  && x.Uses == 1
  1477  && sequentialAddresses(p0, p1, int64(a.Off()+8-c.Off()))
  1478  && a.Val() == 0
  1479  && c.Val() == 0
  1480  && setPos(v, x.Pos)
  1481  && clobber(x)
  1482  => (MOVOstoreconst [makeValAndOff(0,a.Off())] {s} p0 mem)
  1483
  1484// Merge load and op
  1485// TODO: add indexed variants?
  1486((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)
  1487((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)
  1488((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)
  1489((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)
  1490(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)
  1491(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) =>
  1492	((ADD|SUB|AND|OR|XOR)Lmodify [off] {sym} ptr x mem)
  1493(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)
  1494(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) =>
  1495	((ADD|SUB|AND|OR|XOR)Qmodify [off] {sym} ptr x mem)
  1496(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) =>
  1497	(BT(S|R|C)Qconstmodify {sym} [makeValAndOff(int32(c),off)] ptr mem)
  1498
  1499// Merge ADDQconst and LEAQ into atomic loads.
  1500(MOV(Q|L|B)atomicload [off1] {sym} (ADDQconst [off2] ptr) mem) && is32Bit(int64(off1)+int64(off2)) =>
  1501	(MOV(Q|L|B)atomicload [off1+off2] {sym} ptr mem)
  1502(MOV(Q|L|B)atomicload [off1] {sym1} (LEAQ [off2] {sym2} ptr) mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
  1503	(MOV(Q|L|B)atomicload [off1+off2] {mergeSym(sym1, sym2)} ptr mem)
  1504
  1505// Merge ADDQconst and LEAQ into atomic stores.
  1506(XCHGQ [off1] {sym} val (ADDQconst [off2] ptr) mem) && is32Bit(int64(off1)+int64(off2)) =>
  1507	(XCHGQ [off1+off2] {sym} val ptr mem)
  1508(XCHGQ [off1] {sym1} val (LEAQ [off2] {sym2} ptr) mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) && ptr.Op != OpSB =>
  1509	(XCHGQ [off1+off2] {mergeSym(sym1,sym2)} val ptr mem)
  1510(XCHGL [off1] {sym} val (ADDQconst [off2] ptr) mem) && is32Bit(int64(off1)+int64(off2)) =>
  1511	(XCHGL [off1+off2] {sym} val ptr mem)
  1512(XCHGL [off1] {sym1} val (LEAQ [off2] {sym2} ptr) mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) && ptr.Op != OpSB =>
  1513	(XCHGL [off1+off2] {mergeSym(sym1,sym2)} val ptr mem)
  1514
  1515// Merge ADDQconst into atomic adds.
  1516// TODO: merging LEAQ doesn't work, assembler doesn't like the resulting instructions.
  1517(XADDQlock [off1] {sym} val (ADDQconst [off2] ptr) mem) && is32Bit(int64(off1)+int64(off2)) =>
  1518	(XADDQlock [off1+off2] {sym} val ptr mem)
  1519(XADDLlock [off1] {sym} val (ADDQconst [off2] ptr) mem) && is32Bit(int64(off1)+int64(off2)) =>
  1520	(XADDLlock [off1+off2] {sym} val ptr mem)
  1521
  1522// Merge ADDQconst into atomic compare and swaps.
  1523// TODO: merging LEAQ doesn't work, assembler doesn't like the resulting instructions.
  1524(CMPXCHGQlock [off1] {sym} (ADDQconst [off2] ptr) old new_ mem) && is32Bit(int64(off1)+int64(off2)) =>
  1525	(CMPXCHGQlock [off1+off2] {sym} ptr old new_ mem)
  1526(CMPXCHGLlock [off1] {sym} (ADDQconst [off2] ptr) old new_ mem) && is32Bit(int64(off1)+int64(off2)) =>
  1527	(CMPXCHGLlock [off1+off2] {sym} ptr old new_ mem)
  1528
  1529// We don't need the conditional move if we know the arg of BSF is not zero.
  1530(CMOVQEQ x _ (Select1 (BS(F|R)Q (ORQconst [c] _)))) && c != 0 => x
  1531// Extension is unnecessary for trailing zeros.
  1532(BSFQ (ORQconst <t> [1<<8] (MOVBQZX x))) => (BSFQ (ORQconst <t> [1<<8] x))
  1533(BSFQ (ORQconst <t> [1<<16] (MOVWQZX x))) => (BSFQ (ORQconst <t> [1<<16] x))
  1534
  1535// Redundant sign/zero extensions
  1536// Note: see issue 21963. We have to make sure we use the right type on
  1537// the resulting extension (the outer type, not the inner type).
  1538(MOVLQSX (MOVLQSX x)) => (MOVLQSX x)
  1539(MOVLQSX (MOVWQSX x)) => (MOVWQSX x)
  1540(MOVLQSX (MOVBQSX x)) => (MOVBQSX x)
  1541(MOVWQSX (MOVWQSX x)) => (MOVWQSX x)
  1542(MOVWQSX (MOVBQSX x)) => (MOVBQSX x)
  1543(MOVBQSX (MOVBQSX x)) => (MOVBQSX x)
  1544(MOVLQZX (MOVLQZX x)) => (MOVLQZX x)
  1545(MOVLQZX (MOVWQZX x)) => (MOVWQZX x)
  1546(MOVLQZX (MOVBQZX x)) => (MOVBQZX x)
  1547(MOVWQZX (MOVWQZX x)) => (MOVWQZX x)
  1548(MOVWQZX (MOVBQZX x)) => (MOVBQZX x)
  1549(MOVBQZX (MOVBQZX x)) => (MOVBQZX x)
  1550
  1551(MOVQstore [off] {sym} ptr a:((ADD|AND|OR|XOR)Qconst [c] l:(MOVQload [off] {sym} ptr2 mem)) mem)
  1552	&& isSamePtr(ptr, ptr2) && a.Uses == 1 && l.Uses == 1 && clobber(l, a) =>
  1553	((ADD|AND|OR|XOR)Qconstmodify {sym} [makeValAndOff(int32(c),off)] ptr mem)
  1554(MOVLstore [off] {sym} ptr a:((ADD|AND|OR|XOR)Lconst [c] l:(MOVLload [off] {sym} ptr2 mem)) mem)
  1555	&& isSamePtr(ptr, ptr2) && a.Uses == 1 && l.Uses == 1 && clobber(l, a) =>
  1556	((ADD|AND|OR|XOR)Lconstmodify {sym} [makeValAndOff(int32(c),off)] ptr mem)
  1557
  1558// float <-> int register moves, with no conversion.
  1559// These come up when compiling math.{Float{32,64}bits,Float{32,64}frombits}.
  1560(MOVQload  [off] {sym} ptr (MOVSDstore [off] {sym} ptr val _)) => (MOVQf2i val)
  1561(MOVLload  [off] {sym} ptr (MOVSSstore [off] {sym} ptr val _)) => (MOVLf2i val)
  1562(MOVSDload [off] {sym} ptr (MOVQstore  [off] {sym} ptr val _)) => (MOVQi2f val)
  1563(MOVSSload [off] {sym} ptr (MOVLstore  [off] {sym} ptr val _)) => (MOVLi2f val)
  1564
  1565// Other load-like ops.
  1566(ADDQload x [off] {sym} ptr (MOVSDstore [off] {sym} ptr y _)) => (ADDQ x (MOVQf2i y))
  1567(ADDLload x [off] {sym} ptr (MOVSSstore [off] {sym} ptr y _)) => (ADDL x (MOVLf2i y))
  1568(SUBQload x [off] {sym} ptr (MOVSDstore [off] {sym} ptr y _)) => (SUBQ x (MOVQf2i y))
  1569(SUBLload x [off] {sym} ptr (MOVSSstore [off] {sym} ptr y _)) => (SUBL x (MOVLf2i y))
  1570(ANDQload x [off] {sym} ptr (MOVSDstore [off] {sym} ptr y _)) => (ANDQ x (MOVQf2i y))
  1571(ANDLload x [off] {sym} ptr (MOVSSstore [off] {sym} ptr y _)) => (ANDL x (MOVLf2i y))
  1572( ORQload x [off] {sym} ptr (MOVSDstore [off] {sym} ptr y _)) => ( ORQ x (MOVQf2i y))
  1573( ORLload x [off] {sym} ptr (MOVSSstore [off] {sym} ptr y _)) => ( ORL x (MOVLf2i y))
  1574(XORQload x [off] {sym} ptr (MOVSDstore [off] {sym} ptr y _)) => (XORQ x (MOVQf2i y))
  1575(XORLload x [off] {sym} ptr (MOVSSstore [off] {sym} ptr y _)) => (XORL x (MOVLf2i y))
  1576
  1577(ADDSDload x [off] {sym} ptr (MOVQstore [off] {sym} ptr y _)) => (ADDSD x (MOVQi2f y))
  1578(ADDSSload x [off] {sym} ptr (MOVLstore [off] {sym} ptr y _)) => (ADDSS x (MOVLi2f y))
  1579(SUBSDload x [off] {sym} ptr (MOVQstore [off] {sym} ptr y _)) => (SUBSD x (MOVQi2f y))
  1580(SUBSSload x [off] {sym} ptr (MOVLstore [off] {sym} ptr y _)) => (SUBSS x (MOVLi2f y))
  1581(MULSDload x [off] {sym} ptr (MOVQstore [off] {sym} ptr y _)) => (MULSD x (MOVQi2f y))
  1582(MULSSload x [off] {sym} ptr (MOVLstore [off] {sym} ptr y _)) => (MULSS x (MOVLi2f y))
  1583
  1584// Redirect stores to use the other register set.
  1585(MOVQstore  [off] {sym} ptr (MOVQf2i val) mem) => (MOVSDstore [off] {sym} ptr val mem)
  1586(MOVLstore  [off] {sym} ptr (MOVLf2i val) mem) => (MOVSSstore [off] {sym} ptr val mem)
  1587(MOVSDstore [off] {sym} ptr (MOVQi2f val) mem) => (MOVQstore  [off] {sym} ptr val mem)
  1588(MOVSSstore [off] {sym} ptr (MOVLi2f val) mem) => (MOVLstore  [off] {sym} ptr val mem)
  1589
  1590(MOVSDstore [off] {sym} ptr (MOVSDconst [f]) mem) && f == f => (MOVQstore [off] {sym} ptr (MOVQconst [int64(math.Float64bits(f))]) mem)
  1591(MOVSSstore [off] {sym} ptr (MOVSSconst [f]) mem) && f == f => (MOVLstore [off] {sym} ptr (MOVLconst [int32(math.Float32bits(f))]) mem)
  1592
  1593// Load args directly into the register class where it will be used.
  1594// We do this by just modifying the type of the Arg.
  1595(MOVQf2i <t> (Arg <u> [off] {sym})) && t.Size() == u.Size() => @b.Func.Entry (Arg <t> [off] {sym})
  1596(MOVLf2i <t> (Arg <u> [off] {sym})) && t.Size() == u.Size() => @b.Func.Entry (Arg <t> [off] {sym})
  1597(MOVQi2f <t> (Arg <u> [off] {sym})) && t.Size() == u.Size() => @b.Func.Entry (Arg <t> [off] {sym})
  1598(MOVLi2f <t> (Arg <u> [off] {sym})) && t.Size() == u.Size() => @b.Func.Entry (Arg <t> [off] {sym})
  1599
  1600// LEAQ is rematerializeable, so this helps to avoid register spill.
  1601// See issue 22947 for details
  1602(ADD(Q|L)const [off] x:(SP)) => (LEA(Q|L) [off] x)
  1603
  1604// HMULx is commutative, but its first argument must go in AX.
  1605// If possible, put a rematerializeable value in the first argument slot,
  1606// to reduce the odds that another value will be have to spilled
  1607// specifically to free up AX.
  1608(HMUL(Q|L)  x y) && !x.rematerializeable() && y.rematerializeable() => (HMUL(Q|L)  y x)
  1609(HMUL(Q|L)U x y) && !x.rematerializeable() && y.rematerializeable() => (HMUL(Q|L)U y x)
  1610
  1611// Fold loads into compares
  1612// Note: these may be undone by the flagalloc pass.
  1613(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)
  1614(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))
  1615
  1616(CMP(Q|L)const l:(MOV(Q|L)load {sym} [off] ptr mem) [c])
  1617	&& l.Uses == 1
  1618	&& clobber(l) =>
  1619@l.Block (CMP(Q|L)constload {sym} [makeValAndOff(c,off)] ptr mem)
  1620(CMP(W|B)const l:(MOV(W|B)load {sym} [off] ptr mem) [c])
  1621	&& l.Uses == 1
  1622	&& clobber(l) =>
  1623@l.Block (CMP(W|B)constload {sym} [makeValAndOff(int32(c),off)] ptr mem)
  1624
  1625(CMPQload {sym} [off] ptr (MOVQconst [c]) mem) && validVal(c) => (CMPQconstload {sym} [makeValAndOff(int32(c),off)] ptr mem)
  1626(CMPLload {sym} [off] ptr (MOVLconst [c]) mem) => (CMPLconstload {sym} [makeValAndOff(c,off)] ptr mem)
  1627(CMPWload {sym} [off] ptr (MOVLconst [c]) mem) => (CMPWconstload {sym} [makeValAndOff(int32(int16(c)),off)] ptr mem)
  1628(CMPBload {sym} [off] ptr (MOVLconst [c]) mem) => (CMPBconstload {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
  1629
  1630(TEST(Q|L|W|B)  l:(MOV(Q|L|W|B)load {sym} [off] ptr mem) l2)
  1631        && l == l2
  1632	&& l.Uses == 2
  1633	&& clobber(l) =>
  1634  @l.Block (CMP(Q|L|W|B)constload {sym} [makeValAndOff(0, off)] ptr mem)
  1635
  1636// Convert ANDload to MOVload when we can do the AND in a containing TEST op.
  1637// Only do when it's within the same block, so we don't have flags live across basic block boundaries.
  1638// See issue 44228.
  1639(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)
  1640
  1641(MOVBload [off] {sym} (SB) _) && symIsRO(sym) => (MOVLconst [int32(read8(sym, int64(off)))])
  1642(MOVWload [off] {sym} (SB) _) && symIsRO(sym) => (MOVLconst [int32(read16(sym, int64(off), config.ctxt.Arch.ByteOrder))])
  1643(MOVLload [off] {sym} (SB) _) && symIsRO(sym) => (MOVQconst [int64(read32(sym, int64(off), config.ctxt.Arch.ByteOrder))])
  1644(MOVQload [off] {sym} (SB) _) && symIsRO(sym) => (MOVQconst [int64(read64(sym, int64(off), config.ctxt.Arch.ByteOrder))])
  1645(MOVOstore [dstOff] {dstSym} ptr (MOVOload [srcOff] {srcSym} (SB) _) mem) && symIsRO(srcSym) =>
  1646  (MOVQstore [dstOff+8] {dstSym} ptr (MOVQconst [int64(read64(srcSym, int64(srcOff)+8, config.ctxt.Arch.ByteOrder))])
  1647    (MOVQstore [dstOff] {dstSym} ptr (MOVQconst [int64(read64(srcSym, int64(srcOff), config.ctxt.Arch.ByteOrder))]) mem))
  1648
  1649// Arch-specific inlining for small or disjoint runtime.memmove
  1650// Match post-lowering calls, memory version.
  1651(SelectN [0] call:(CALLstatic {sym} s1:(MOVQstoreconst _ [sc] s2:(MOVQstore _ src s3:(MOVQstore _ dst mem)))))
  1652	&& sc.Val64() >= 0
  1653	&& isSameCall(sym, "runtime.memmove")
  1654	&& s1.Uses == 1 && s2.Uses == 1 && s3.Uses == 1
  1655	&& isInlinableMemmove(dst, src, sc.Val64(), config)
  1656	&& clobber(s1, s2, s3, call)
  1657	=> (Move [sc.Val64()] dst src mem)
  1658
  1659// Match post-lowering calls, register version.
  1660(SelectN [0] call:(CALLstatic {sym} dst src (MOVQconst [sz]) mem))
  1661	&& sz >= 0
  1662	&& isSameCall(sym, "runtime.memmove")
  1663	&& call.Uses == 1
  1664	&& isInlinableMemmove(dst, src, sz, config)
  1665	&& clobber(call)
  1666	=> (Move [sz] dst src mem)
  1667
  1668// Prefetch instructions
  1669(PrefetchCache ...)   => (PrefetchT0 ...)
  1670(PrefetchCacheStreamed ...) => (PrefetchNTA ...)
  1671
  1672// CPUID feature: BMI1.
  1673(AND(Q|L) x (NOT(Q|L) y))               && buildcfg.GOAMD64 >= 3 => (ANDN(Q|L) x y)
  1674(AND(Q|L) x (NEG(Q|L) x))               && buildcfg.GOAMD64 >= 3 => (BLSI(Q|L) x)
  1675(XOR(Q|L) x (ADD(Q|L)const [-1] x))     && buildcfg.GOAMD64 >= 3 => (BLSMSK(Q|L) x)
  1676(AND(Q|L) <t> x (ADD(Q|L)const [-1] x)) && buildcfg.GOAMD64 >= 3 => (Select0 <t> (BLSR(Q|L) x))
  1677// eliminate TEST instruction in classical "isPowerOfTwo" check
  1678(SETEQ       (TEST(Q|L) s:(Select0 blsr:(BLSR(Q|L) _)) s))        => (SETEQ       (Select1 <types.TypeFlags> blsr))
  1679(CMOVQEQ x y (TEST(Q|L) s:(Select0 blsr:(BLSR(Q|L) _)) s))        => (CMOVQEQ x y (Select1 <types.TypeFlags> blsr))
  1680(CMOVLEQ x y (TEST(Q|L) s:(Select0 blsr:(BLSR(Q|L) _)) s))        => (CMOVLEQ x y (Select1 <types.TypeFlags> blsr))
  1681(EQ          (TEST(Q|L) s:(Select0 blsr:(BLSR(Q|L) _)) s) yes no) => (EQ          (Select1 <types.TypeFlags> blsr) yes no)
  1682(SETNE       (TEST(Q|L) s:(Select0 blsr:(BLSR(Q|L) _)) s))        => (SETNE       (Select1 <types.TypeFlags> blsr))
  1683(CMOVQNE x y (TEST(Q|L) s:(Select0 blsr:(BLSR(Q|L) _)) s))        => (CMOVQNE x y (Select1 <types.TypeFlags> blsr))
  1684(CMOVLNE x y (TEST(Q|L) s:(Select0 blsr:(BLSR(Q|L) _)) s))        => (CMOVLNE x y (Select1 <types.TypeFlags> blsr))
  1685(NE          (TEST(Q|L) s:(Select0 blsr:(BLSR(Q|L) _)) s) yes no) => (NE          (Select1 <types.TypeFlags> blsr) yes no)
  1686
  1687(BSWAP(Q|L) (BSWAP(Q|L) p)) => p
  1688
  1689// CPUID feature: MOVBE.
  1690(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)
  1691(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)
  1692(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)
  1693(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)
  1694(MOVWstore [i] {s} p x:(ROLWconst [8] w) mem)   && x.Uses == 1 && buildcfg.GOAMD64 >= 3 => (MOVBEWstore [i] {s} p w mem)
  1695(MOVBEWstore [i] {s} p x:(ROLWconst [8] w) mem) && x.Uses == 1 => (MOVWstore [i] {s} p w mem)
  1696
  1697(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)
  1698(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)
  1699(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)
  1700
  1701((SHL|SHR|SAR)XQload [off] {sym} ptr (MOVQconst [c]) mem) => ((SHL|SHR|SAR)Qconst [int8(c&63)] (MOVQload [off] {sym} ptr mem))
  1702((SHL|SHR|SAR)XQload [off] {sym} ptr (MOVLconst [c]) mem) => ((SHL|SHR|SAR)Qconst [int8(c&63)] (MOVQload [off] {sym} ptr mem))
  1703((SHL|SHR|SAR)XLload [off] {sym} ptr (MOVLconst [c]) mem) => ((SHL|SHR|SAR)Lconst [int8(c&31)] (MOVLload [off] {sym} ptr mem))
  1704
  1705// Convert atomic logical operations to easier ones if we don't use the result.
  1706(Select1 a:(LoweredAtomic(And64|And32|Or64|Or32) ptr val mem)) && a.Uses == 1 && clobber(a) => ((ANDQ|ANDL|ORQ|ORL)lock ptr val mem)

View as plain text