Text file src/cmd/compile/internal/ssa/gen/386.rules

     1  // Copyright 2016 The Go Authors. All rights reserved.
     2  // Use of this source code is governed by a BSD-style
     3  // license that can be found in the LICENSE file.
     4  
     5  // Lowering arithmetic
     6  (Add(Ptr|32|16|8) ...) => (ADDL ...)
     7  (Add(32|64)F ...) => (ADDS(S|D) ...)
     8  (Add32carry ...) => (ADDLcarry ...)
     9  (Add32withcarry ...) => (ADCL ...)
    10  
    11  (Sub(Ptr|32|16|8) ...) => (SUBL ...)
    12  (Sub(32|64)F ...) => (SUBS(S|D) ...)
    13  (Sub32carry ...) => (SUBLcarry ...)
    14  (Sub32withcarry ...) => (SBBL ...)
    15  
    16  (Mul(32|16|8) ...) => (MULL ...)
    17  (Mul(32|64)F ...) => (MULS(S|D) ...)
    18  (Mul32uhilo ...) => (MULLQU ...)
    19  
    20  (Select0 (Mul32uover x y)) => (Select0 <typ.UInt32> (MULLU x y))
    21  (Select1 (Mul32uover x y)) => (SETO (Select1 <types.TypeFlags> (MULLU x y)))
    22  
    23  (Avg32u ...) => (AVGLU ...)
    24  
    25  (Div(32|64)F ...) => (DIVS(S|D) ...)
    26  (Div(32|32u|16|16u) ...) => (DIV(L|LU|W|WU) ...)
    27  (Div8   x y) => (DIVW  (SignExt8to16 x) (SignExt8to16 y))
    28  (Div8u  x y) => (DIVWU (ZeroExt8to16 x) (ZeroExt8to16 y))
    29  
    30  (Hmul(32|32u) ...) => (HMUL(L|LU) ...)
    31  
    32  (Mod(32|32u|16|16u) ...) => (MOD(L|LU|W|WU) ...)
    33  (Mod8   x y) => (MODW  (SignExt8to16 x) (SignExt8to16 y))
    34  (Mod8u  x y) => (MODWU (ZeroExt8to16 x) (ZeroExt8to16 y))
    35  
    36  (And(32|16|8) ...) => (ANDL ...)
    37  (Or(32|16|8) ...) => (ORL ...)
    38  (Xor(32|16|8) ...) => (XORL ...)
    39  
    40  (Neg(32|16|8) ...) => (NEGL ...)
    41  (Neg32F x) => (PXOR x (MOVSSconst <typ.Float32> [float32(math.Copysign(0, -1))]))
    42  (Neg64F x) => (PXOR x (MOVSDconst <typ.Float64> [math.Copysign(0, -1)]))
    43  
    44  (Com(32|16|8) ...) => (NOTL ...)
    45  
    46  // Lowering boolean ops
    47  (AndB ...) => (ANDL ...)
    48  (OrB ...) => (ORL ...)
    49  (Not x) => (XORLconst [1] x)
    50  
    51  // Lowering pointer arithmetic
    52  (OffPtr [off] ptr) => (ADDLconst [int32(off)] ptr)
    53  
    54  (Bswap32 ...) => (BSWAPL ...)
    55  
    56  (Sqrt ...) => (SQRTSD ...)
    57  (Sqrt32 ...) => (SQRTSS ...)
    58  
    59  (Ctz16 x) => (BSFL (ORLconst <typ.UInt32> [0x10000] x))
    60  (Ctz16NonZero ...) => (BSFL ...)
    61  
    62  // Lowering extension
    63  (SignExt8to16  ...) => (MOVBLSX ...)
    64  (SignExt8to32  ...) => (MOVBLSX ...)
    65  (SignExt16to32 ...) => (MOVWLSX ...)
    66  
    67  (ZeroExt8to16  ...) => (MOVBLZX ...)
    68  (ZeroExt8to32  ...) => (MOVBLZX ...)
    69  (ZeroExt16to32 ...) => (MOVWLZX ...)
    70  
    71  (Signmask x) => (SARLconst x [31])
    72  (Zeromask <t> x) => (XORLconst [-1] (SBBLcarrymask <t> (CMPLconst x [1])))
    73  (Slicemask <t> x) => (SARLconst (NEGL <t> x) [31])
    74  
    75  // Lowering truncation
    76  // Because we ignore high parts of registers, truncates are just copies.
    77  (Trunc16to8  ...) => (Copy ...)
    78  (Trunc32to8  ...) => (Copy ...)
    79  (Trunc32to16 ...) => (Copy ...)
    80  
    81  // Lowering float-int conversions
    82  (Cvt32to32F ...) => (CVTSL2SS ...)
    83  (Cvt32to64F ...) => (CVTSL2SD ...)
    84  
    85  (Cvt32Fto32 ...) => (CVTTSS2SL ...)
    86  (Cvt64Fto32 ...) => (CVTTSD2SL ...)
    87  
    88  (Cvt32Fto64F ...) => (CVTSS2SD ...)
    89  (Cvt64Fto32F ...) => (CVTSD2SS ...)
    90  
    91  (Round32F ...) => (Copy ...)
    92  (Round64F ...) => (Copy ...)
    93  
    94  (CvtBoolToUint8 ...) => (Copy ...)
    95  
    96  // Lowering shifts
    97  // Unsigned shifts need to return 0 if shift amount is >= width of shifted value.
    98  //   result = (arg << shift) & (shift >= argbits ? 0 : 0xffffffffffffffff)
    99  (Lsh32x(32|16|8) <t> x y) && !shiftIsBounded(v) => (ANDL (SHLL <t> x y) (SBBLcarrymask <t> (CMP(L|W|B)const y [32])))
   100  (Lsh16x(32|16|8) <t> x y) && !shiftIsBounded(v) => (ANDL (SHLL <t> x y) (SBBLcarrymask <t> (CMP(L|W|B)const y [32])))
   101  (Lsh8x(32|16|8)  <t> x y) && !shiftIsBounded(v) => (ANDL (SHLL <t> x y) (SBBLcarrymask <t> (CMP(L|W|B)const y [32])))
   102  
   103  (Lsh32x(32|16|8) <t> x y) && shiftIsBounded(v) => (SHLL <t> x y)
   104  (Lsh16x(32|16|8) <t> x y) && shiftIsBounded(v) => (SHLL <t> x y)
   105  (Lsh8x(32|16|8)  <t> x y) && shiftIsBounded(v) => (SHLL <t> x y)
   106  
   107  (Rsh32Ux(32|16|8) <t> x y) && !shiftIsBounded(v) => (ANDL (SHRL <t> x y) (SBBLcarrymask <t> (CMP(L|W|B)const y [32])))
   108  (Rsh16Ux(32|16|8) <t> x y) && !shiftIsBounded(v) => (ANDL (SHRW <t> x y) (SBBLcarrymask <t> (CMP(L|W|B)const y [16])))
   109  (Rsh8Ux(32|16|8)  <t> x y) && !shiftIsBounded(v) => (ANDL (SHRB <t> x y) (SBBLcarrymask <t> (CMP(L|W|B)const y [8])))
   110  
   111  (Rsh32Ux(32|16|8) <t> x y) && shiftIsBounded(v) => (SHRL <t> x y)
   112  (Rsh16Ux(32|16|8) <t> x y) && shiftIsBounded(v) => (SHRW <t> x y)
   113  (Rsh8Ux(32|16|8)  <t> x y) && shiftIsBounded(v) => (SHRB <t> x y)
   114  
   115  // Signed right shift needs to return 0/-1 if shift amount is >= width of shifted value.
   116  // We implement this by setting the shift value to -1 (all ones) if the shift value is >= width.
   117  
   118  (Rsh32x(32|16|8) <t> x y) && !shiftIsBounded(v) => (SARL <t> x (ORL <y.Type> y (NOTL <y.Type> (SBBLcarrymask <y.Type> (CMP(L|W|B)const y [32])))))
   119  (Rsh16x(32|16|8) <t> x y) && !shiftIsBounded(v) => (SARW <t> x (ORL <y.Type> y (NOTL <y.Type> (SBBLcarrymask <y.Type> (CMP(L|W|B)const y [16])))))
   120  (Rsh8x(32|16|8) <t> x y)  && !shiftIsBounded(v) => (SARB <t> x (ORL <y.Type> y (NOTL <y.Type> (SBBLcarrymask <y.Type> (CMP(L|W|B)const y [8])))))
   121  
   122  (Rsh32x(32|16|8) <t> x y) && shiftIsBounded(v) => (SARL x y)
   123  (Rsh16x(32|16|8) <t> x y) && shiftIsBounded(v) => (SARW x y)
   124  (Rsh8x(32|16|8) <t> x y)  && shiftIsBounded(v) => (SARB x y)
   125  
   126  // constant shifts
   127  // generic opt rewrites all constant shifts to shift by Const64
   128  (Lsh32x64 x (Const64 [c])) && uint64(c) < 32 => (SHLLconst x [int32(c)])
   129  (Rsh32x64 x (Const64 [c])) && uint64(c) < 32 => (SARLconst x [int32(c)])
   130  (Rsh32Ux64 x (Const64 [c])) && uint64(c) < 32 => (SHRLconst x [int32(c)])
   131  (Lsh16x64 x (Const64 [c])) && uint64(c) < 16 => (SHLLconst x [int32(c)])
   132  (Rsh16x64 x (Const64 [c])) && uint64(c) < 16 => (SARWconst x [int16(c)])
   133  (Rsh16Ux64 x (Const64 [c])) && uint64(c) < 16 => (SHRWconst x [int16(c)])
   134  (Lsh8x64 x (Const64 [c])) && uint64(c) < 8 => (SHLLconst x [int32(c)])
   135  (Rsh8x64 x (Const64 [c])) && uint64(c) < 8 => (SARBconst x [int8(c)])
   136  (Rsh8Ux64 x (Const64 [c])) && uint64(c) < 8 => (SHRBconst x [int8(c)])
   137  
   138  // large constant shifts
   139  (Lsh32x64 _ (Const64 [c])) && uint64(c) >= 32 => (Const32 [0])
   140  (Rsh32Ux64 _ (Const64 [c])) && uint64(c) >= 32 => (Const32 [0])
   141  (Lsh16x64 _ (Const64 [c])) && uint64(c) >= 16 => (Const16 [0])
   142  (Rsh16Ux64 _ (Const64 [c])) && uint64(c) >= 16 => (Const16 [0])
   143  (Lsh8x64 _ (Const64 [c])) && uint64(c) >= 8 => (Const8 [0])
   144  (Rsh8Ux64 _ (Const64 [c])) && uint64(c) >= 8 => (Const8 [0])
   145  
   146  // large constant signed right shift, we leave the sign bit
   147  (Rsh32x64 x (Const64 [c])) && uint64(c) >= 32 => (SARLconst x [31])
   148  (Rsh16x64 x (Const64 [c])) && uint64(c) >= 16 => (SARWconst x [15])
   149  (Rsh8x64 x (Const64 [c])) && uint64(c) >= 8 => (SARBconst x [7])
   150  
   151  // constant rotates
   152  (RotateLeft32 x (MOVLconst [c])) => (ROLLconst [c&31] x)
   153  (RotateLeft16 x (MOVLconst [c])) => (ROLWconst [int16(c&15)] x)
   154  (RotateLeft8 x (MOVLconst [c]))  => (ROLBconst [int8(c&7)] x)
   155  
   156  // Lowering comparisons
   157  (Less32  x y) => (SETL (CMPL x y))
   158  (Less16  x y) => (SETL (CMPW x y))
   159  (Less8   x y) => (SETL (CMPB x y))
   160  (Less32U x y) => (SETB (CMPL x y))
   161  (Less16U x y) => (SETB (CMPW x y))
   162  (Less8U  x y) => (SETB (CMPB x y))
   163  // Use SETGF with reversed operands to dodge NaN case
   164  (Less64F x y) => (SETGF (UCOMISD y x))
   165  (Less32F x y) => (SETGF (UCOMISS y x))
   166  
   167  (Leq32  x y) => (SETLE (CMPL x y))
   168  (Leq16  x y) => (SETLE (CMPW x y))
   169  (Leq8   x y) => (SETLE (CMPB x y))
   170  (Leq32U x y) => (SETBE (CMPL x y))
   171  (Leq16U x y) => (SETBE (CMPW x y))
   172  (Leq8U  x y) => (SETBE (CMPB x y))
   173  // Use SETGEF with reversed operands to dodge NaN case
   174  (Leq64F x y) => (SETGEF (UCOMISD y x))
   175  (Leq32F x y) => (SETGEF (UCOMISS y x))
   176  
   177  (Eq32  x y) => (SETEQ (CMPL x y))
   178  (Eq16  x y) => (SETEQ (CMPW x y))
   179  (Eq8   x y) => (SETEQ (CMPB x y))
   180  (EqB   x y) => (SETEQ (CMPB x y))
   181  (EqPtr x y) => (SETEQ (CMPL x y))
   182  (Eq64F x y) => (SETEQF (UCOMISD x y))
   183  (Eq32F x y) => (SETEQF (UCOMISS x y))
   184  
   185  (Neq32  x y) => (SETNE (CMPL x y))
   186  (Neq16  x y) => (SETNE (CMPW x y))
   187  (Neq8   x y) => (SETNE (CMPB x y))
   188  (NeqB   x y) => (SETNE (CMPB x y))
   189  (NeqPtr x y) => (SETNE (CMPL x y))
   190  (Neq64F x y) => (SETNEF (UCOMISD x y))
   191  (Neq32F x y) => (SETNEF (UCOMISS x y))
   192  
   193  // Lowering loads
   194  (Load <t> ptr mem) && (is32BitInt(t) || isPtr(t)) => (MOVLload ptr mem)
   195  (Load <t> ptr mem) && is16BitInt(t) => (MOVWload ptr mem)
   196  (Load <t> ptr mem) && (t.IsBoolean() || is8BitInt(t)) => (MOVBload ptr mem)
   197  (Load <t> ptr mem) && is32BitFloat(t) => (MOVSSload ptr mem)
   198  (Load <t> ptr mem) && is64BitFloat(t) => (MOVSDload ptr mem)
   199  
   200  // Lowering stores
   201  // These more-specific FP versions of Store pattern should come first.
   202  (Store {t} ptr val mem) && t.Size() == 8 && is64BitFloat(val.Type) => (MOVSDstore ptr val mem)
   203  (Store {t} ptr val mem) && t.Size() == 4 && is32BitFloat(val.Type) => (MOVSSstore ptr val mem)
   204  
   205  (Store {t} ptr val mem) && t.Size() == 4 => (MOVLstore ptr val mem)
   206  (Store {t} ptr val mem) && t.Size() == 2 => (MOVWstore ptr val mem)
   207  (Store {t} ptr val mem) && t.Size() == 1 => (MOVBstore ptr val mem)
   208  
   209  // Lowering moves
   210  (Move [0] _ _ mem) => mem
   211  (Move [1] dst src mem) => (MOVBstore dst (MOVBload src mem) mem)
   212  (Move [2] dst src mem) => (MOVWstore dst (MOVWload src mem) mem)
   213  (Move [4] dst src mem) => (MOVLstore dst (MOVLload src mem) mem)
   214  (Move [3] dst src mem) =>
   215  	(MOVBstore [2] dst (MOVBload [2] src mem)
   216  		(MOVWstore dst (MOVWload src mem) mem))
   217  (Move [5] dst src mem) =>
   218  	(MOVBstore [4] dst (MOVBload [4] src mem)
   219  		(MOVLstore dst (MOVLload src mem) mem))
   220  (Move [6] dst src mem) =>
   221  	(MOVWstore [4] dst (MOVWload [4] src mem)
   222  		(MOVLstore dst (MOVLload src mem) mem))
   223  (Move [7] dst src mem) =>
   224  	(MOVLstore [3] dst (MOVLload [3] src mem)
   225  		(MOVLstore dst (MOVLload src mem) mem))
   226  (Move [8] dst src mem) =>
   227  	(MOVLstore [4] dst (MOVLload [4] src mem)
   228  		(MOVLstore dst (MOVLload src mem) mem))
   229  
   230  // Adjust moves to be a multiple of 4 bytes.
   231  (Move [s] dst src mem)
   232  	&& s > 8 && s%4 != 0 =>
   233  	(Move [s-s%4]
   234  		(ADDLconst <dst.Type> dst [int32(s%4)])
   235  		(ADDLconst <src.Type> src [int32(s%4)])
   236  		(MOVLstore dst (MOVLload src mem) mem))
   237  
   238  // Medium copying uses a duff device.
   239  (Move [s] dst src mem)
   240  	&& s > 8 && s <= 4*128 && s%4 == 0
   241  	&& !config.noDuffDevice && logLargeCopy(v, s) =>
   242  	(DUFFCOPY [10*(128-s/4)] dst src mem)
   243  // 10 and 128 are magic constants.  10 is the number of bytes to encode:
   244  //	MOVL	(SI), CX
   245  //	ADDL	$4, SI
   246  //	MOVL	CX, (DI)
   247  //	ADDL	$4, DI
   248  // and 128 is the number of such blocks. See src/runtime/duff_386.s:duffcopy.
   249  
   250  // Large copying uses REP MOVSL.
   251  (Move [s] dst src mem) && (s > 4*128 || config.noDuffDevice) && s%4 == 0 && logLargeCopy(v, s) =>
   252  	(REPMOVSL dst src (MOVLconst [int32(s/4)]) mem)
   253  
   254  // Lowering Zero instructions
   255  (Zero [0] _ mem) => mem
   256  (Zero [1] destptr mem) => (MOVBstoreconst [0] destptr mem)
   257  (Zero [2] destptr mem) => (MOVWstoreconst [0] destptr mem)
   258  (Zero [4] destptr mem) => (MOVLstoreconst [0] destptr mem)
   259  
   260  (Zero [3] destptr mem) =>
   261  	(MOVBstoreconst [makeValAndOff(0,2)] destptr
   262  		(MOVWstoreconst [makeValAndOff(0,0)] destptr mem))
   263  (Zero [5] destptr mem) =>
   264  	(MOVBstoreconst [makeValAndOff(0,4)] destptr
   265  		(MOVLstoreconst [makeValAndOff(0,0)] destptr mem))
   266  (Zero [6] destptr mem) =>
   267  	(MOVWstoreconst [makeValAndOff(0,4)] destptr
   268  		(MOVLstoreconst [makeValAndOff(0,0)] destptr mem))
   269  (Zero [7] destptr mem) =>
   270  	(MOVLstoreconst [makeValAndOff(0,3)] destptr
   271  		(MOVLstoreconst [makeValAndOff(0,0)] destptr mem))
   272  
   273  // Strip off any fractional word zeroing.
   274  (Zero [s] destptr mem) && s%4 != 0 && s > 4 =>
   275  	(Zero [s-s%4] (ADDLconst destptr [int32(s%4)])
   276  		(MOVLstoreconst [0] destptr mem))
   277  
   278  // Zero small numbers of words directly.
   279  (Zero [8] destptr mem) =>
   280  	(MOVLstoreconst [makeValAndOff(0,4)] destptr
   281  		(MOVLstoreconst [makeValAndOff(0,0)] destptr mem))
   282  (Zero [12] destptr mem) =>
   283  	(MOVLstoreconst [makeValAndOff(0,8)] destptr
   284  		(MOVLstoreconst [makeValAndOff(0,4)] destptr
   285  			(MOVLstoreconst [makeValAndOff(0,0)] destptr mem)))
   286  (Zero [16] destptr mem) =>
   287  	(MOVLstoreconst [makeValAndOff(0,12)] destptr
   288  		(MOVLstoreconst [makeValAndOff(0,8)] destptr
   289  			(MOVLstoreconst [makeValAndOff(0,4)] destptr
   290  				(MOVLstoreconst [makeValAndOff(0,0)] destptr mem))))
   291  
   292  // Medium zeroing uses a duff device.
   293  (Zero [s] destptr mem)
   294    && s > 16 && s <= 4*128 && s%4 == 0
   295    && !config.noDuffDevice =>
   296  	(DUFFZERO [1*(128-s/4)] destptr (MOVLconst [0]) mem)
   297  // 1 and 128 are magic constants.  1 is the number of bytes to encode STOSL.
   298  // 128 is the number of STOSL instructions in duffzero.
   299  // See src/runtime/duff_386.s:duffzero.
   300  
   301  // Large zeroing uses REP STOSQ.
   302  (Zero [s] destptr mem)
   303    && (s > 4*128 || (config.noDuffDevice && s > 16))
   304    && s%4 == 0 =>
   305  	(REPSTOSL destptr (MOVLconst [int32(s/4)]) (MOVLconst [0]) mem)
   306  
   307  
   308  // Lowering constants
   309  (Const8   [c]) => (MOVLconst [int32(c)])
   310  (Const16  [c]) => (MOVLconst [int32(c)])
   311  (Const32  ...) => (MOVLconst ...)
   312  (Const(32|64)F ...) => (MOVS(S|D)const ...)
   313  (ConstNil) => (MOVLconst [0])
   314  (ConstBool [c]) => (MOVLconst [b2i32(c)])
   315  
   316  // Lowering calls
   317  (StaticCall ...) => (CALLstatic ...)
   318  (ClosureCall ...) => (CALLclosure ...)
   319  (InterCall ...) => (CALLinter ...)
   320  (TailCall ...) => (CALLtail ...)
   321  
   322  // Miscellaneous
   323  (IsNonNil p) => (SETNE (TESTL p p))
   324  (IsInBounds idx len) => (SETB (CMPL idx len))
   325  (IsSliceInBounds idx len) => (SETBE (CMPL idx len))
   326  (NilCheck ...) => (LoweredNilCheck ...)
   327  (GetG ...) => (LoweredGetG ...)
   328  (GetClosurePtr ...) => (LoweredGetClosurePtr ...)
   329  (GetCallerPC ...) => (LoweredGetCallerPC ...)
   330  (GetCallerSP ...) => (LoweredGetCallerSP ...)
   331  (Addr {sym} base) => (LEAL {sym} base)
   332  (LocalAddr {sym} base _) => (LEAL {sym} base)
   333  
   334  // block rewrites
   335  (If (SETL  cmp) yes no) => (LT  cmp yes no)
   336  (If (SETLE cmp) yes no) => (LE  cmp yes no)
   337  (If (SETG  cmp) yes no) => (GT  cmp yes no)
   338  (If (SETGE cmp) yes no) => (GE  cmp yes no)
   339  (If (SETEQ cmp) yes no) => (EQ  cmp yes no)
   340  (If (SETNE cmp) yes no) => (NE  cmp yes no)
   341  (If (SETB  cmp) yes no) => (ULT cmp yes no)
   342  (If (SETBE cmp) yes no) => (ULE cmp yes no)
   343  (If (SETA  cmp) yes no) => (UGT cmp yes no)
   344  (If (SETAE cmp) yes no) => (UGE cmp yes no)
   345  (If (SETO  cmp) yes no) => (OS cmp yes no)
   346  
   347  // Special case for floating point - LF/LEF not generated
   348  (If (SETGF  cmp) yes no) => (UGT  cmp yes no)
   349  (If (SETGEF cmp) yes no) => (UGE  cmp yes no)
   350  (If (SETEQF cmp) yes no) => (EQF  cmp yes no)
   351  (If (SETNEF cmp) yes no) => (NEF  cmp yes no)
   352  
   353  (If cond yes no) => (NE (TESTB cond cond) yes no)
   354  
   355  // Write barrier.
   356  (WB ...) => (LoweredWB ...)
   357  
   358  (PanicBounds [kind] x y mem) && boundsABI(kind) == 0 => (LoweredPanicBoundsA [kind] x y mem)
   359  (PanicBounds [kind] x y mem) && boundsABI(kind) == 1 => (LoweredPanicBoundsB [kind] x y mem)
   360  (PanicBounds [kind] x y mem) && boundsABI(kind) == 2 => (LoweredPanicBoundsC [kind] x y mem)
   361  
   362  (PanicExtend [kind] hi lo y mem) && boundsABI(kind) == 0 => (LoweredPanicExtendA [kind] hi lo y mem)
   363  (PanicExtend [kind] hi lo y mem) && boundsABI(kind) == 1 => (LoweredPanicExtendB [kind] hi lo y mem)
   364  (PanicExtend [kind] hi lo y mem) && boundsABI(kind) == 2 => (LoweredPanicExtendC [kind] hi lo y mem)
   365  
   366  // ***************************
   367  // Above: lowering rules
   368  // Below: optimizations
   369  // ***************************
   370  // TODO: Should the optimizations be a separate pass?
   371  
   372  // Fold boolean tests into blocks
   373  (NE (TESTB (SETL  cmp) (SETL  cmp)) yes no) => (LT  cmp yes no)
   374  (NE (TESTB (SETLE cmp) (SETLE cmp)) yes no) => (LE  cmp yes no)
   375  (NE (TESTB (SETG  cmp) (SETG  cmp)) yes no) => (GT  cmp yes no)
   376  (NE (TESTB (SETGE cmp) (SETGE cmp)) yes no) => (GE  cmp yes no)
   377  (NE (TESTB (SETEQ cmp) (SETEQ cmp)) yes no) => (EQ  cmp yes no)
   378  (NE (TESTB (SETNE cmp) (SETNE cmp)) yes no) => (NE  cmp yes no)
   379  (NE (TESTB (SETB  cmp) (SETB  cmp)) yes no) => (ULT cmp yes no)
   380  (NE (TESTB (SETBE cmp) (SETBE cmp)) yes no) => (ULE cmp yes no)
   381  (NE (TESTB (SETA  cmp) (SETA  cmp)) yes no) => (UGT cmp yes no)
   382  (NE (TESTB (SETAE cmp) (SETAE cmp)) yes no) => (UGE cmp yes no)
   383  (NE (TESTB (SETO cmp) (SETO cmp)) yes no) => (OS cmp yes no)
   384  
   385  // Special case for floating point - LF/LEF not generated
   386  (NE (TESTB (SETGF  cmp) (SETGF  cmp)) yes no) => (UGT  cmp yes no)
   387  (NE (TESTB (SETGEF cmp) (SETGEF cmp)) yes no) => (UGE  cmp yes no)
   388  (NE (TESTB (SETEQF cmp) (SETEQF cmp)) yes no) => (EQF  cmp yes no)
   389  (NE (TESTB (SETNEF cmp) (SETNEF cmp)) yes no) => (NEF  cmp yes no)
   390  
   391  // fold constants into instructions
   392  (ADDL x (MOVLconst [c])) => (ADDLconst [c] x)
   393  (ADDLcarry x (MOVLconst [c])) => (ADDLconstcarry [c] x)
   394  (ADCL x (MOVLconst [c]) f) => (ADCLconst [c] x f)
   395  
   396  (SUBL x (MOVLconst [c])) => (SUBLconst x [c])
   397  (SUBL (MOVLconst [c]) x) => (NEGL (SUBLconst <v.Type> x [c]))
   398  (SUBLcarry x (MOVLconst [c])) => (SUBLconstcarry [c] x)
   399  (SBBL x (MOVLconst [c]) f) => (SBBLconst [c] x f)
   400  
   401  (MULL x (MOVLconst [c])) => (MULLconst [c] x)
   402  (ANDL x (MOVLconst [c])) => (ANDLconst [c] x)
   403  
   404  (ANDLconst [c] (ANDLconst [d] x)) => (ANDLconst [c & d] x)
   405  (XORLconst [c] (XORLconst [d] x)) => (XORLconst [c ^ d] x)
   406  (MULLconst [c] (MULLconst [d] x)) => (MULLconst [c * d] x)
   407  
   408  (ORL x (MOVLconst [c])) => (ORLconst [c] x)
   409  (XORL x (MOVLconst [c])) => (XORLconst [c] x)
   410  
   411  (SHLL x (MOVLconst [c])) => (SHLLconst [c&31] x)
   412  (SHRL x (MOVLconst [c])) => (SHRLconst [c&31] x)
   413  (SHRW x (MOVLconst [c])) && c&31 < 16 => (SHRWconst [int16(c&31)] x)
   414  (SHRW _ (MOVLconst [c])) && c&31 >= 16 => (MOVLconst [0])
   415  (SHRB x (MOVLconst [c])) && c&31 < 8 => (SHRBconst [int8(c&31)] x)
   416  (SHRB _ (MOVLconst [c])) && c&31 >= 8 => (MOVLconst [0])
   417  
   418  (SARL x (MOVLconst [c])) => (SARLconst [c&31] x)
   419  (SARW x (MOVLconst [c])) => (SARWconst [int16(min(int64(c&31),15))] x)
   420  (SARB x (MOVLconst [c])) => (SARBconst [int8(min(int64(c&31),7))] x)
   421  
   422  (SARL x (ANDLconst [31] y)) => (SARL x y)
   423  (SHLL x (ANDLconst [31] y)) => (SHLL x y)
   424  (SHRL x (ANDLconst [31] y)) => (SHRL x y)
   425  
   426  // Rotate instructions
   427  
   428  (ADDL (SHLLconst [c] x) (SHRLconst [d] x)) && d == 32-c => (ROLLconst [c] x)
   429  ( ORL (SHLLconst [c] x) (SHRLconst [d] x)) && d == 32-c => (ROLLconst [c] x)
   430  (XORL (SHLLconst [c] x) (SHRLconst [d] x)) && d == 32-c => (ROLLconst [c] x)
   431  
   432  (ADDL <t> (SHLLconst x [c]) (SHRWconst x [d])) && c < 16 && d == int16(16-c) && t.Size() == 2
   433    => (ROLWconst x [int16(c)])
   434  ( ORL <t> (SHLLconst x [c]) (SHRWconst x [d])) && c < 16 && d == int16(16-c) && t.Size() == 2
   435    => (ROLWconst x [int16(c)])
   436  (XORL <t> (SHLLconst x [c]) (SHRWconst x [d])) && c < 16 && d == int16(16-c) && t.Size() == 2
   437    => (ROLWconst x [int16(c)])
   438  
   439  (ADDL <t> (SHLLconst x [c]) (SHRBconst x [d])) && c < 8 && d == int8(8-c) && t.Size() == 1
   440    => (ROLBconst x [int8(c)])
   441  ( ORL <t> (SHLLconst x [c]) (SHRBconst x [d])) && c < 8 && d == int8(8-c) && t.Size() == 1
   442    => (ROLBconst x [int8(c)])
   443  (XORL <t> (SHLLconst x [c]) (SHRBconst x [d])) && c < 8 && d == int8(8-c) && t.Size() == 1
   444    => (ROLBconst x [int8(c)])
   445  
   446  (ROLLconst [c] (ROLLconst [d] x)) => (ROLLconst [(c+d)&31] x)
   447  (ROLWconst [c] (ROLWconst [d] x)) => (ROLWconst [(c+d)&15] x)
   448  (ROLBconst [c] (ROLBconst [d] x)) => (ROLBconst [(c+d)& 7] x)
   449  
   450  
   451  // Constant shift simplifications
   452  
   453  (SHLLconst x [0]) => x
   454  (SHRLconst x [0]) => x
   455  (SARLconst x [0]) => x
   456  
   457  (SHRWconst x [0]) => x
   458  (SARWconst x [0]) => x
   459  
   460  (SHRBconst x [0]) => x
   461  (SARBconst x [0]) => x
   462  
   463  (ROLLconst [0] x) => x
   464  (ROLWconst [0] x) => x
   465  (ROLBconst [0] x) => x
   466  
   467  // Note: the word and byte shifts keep the low 5 bits (not the low 4 or 3 bits)
   468  // because the x86 instructions are defined to use all 5 bits of the shift even
   469  // for the small shifts. I don't think we'll ever generate a weird shift (e.g.
   470  // (SHRW x (MOVLconst [24])), but just in case.
   471  
   472  (CMPL x (MOVLconst [c])) => (CMPLconst x [c])
   473  (CMPL (MOVLconst [c]) x) => (InvertFlags (CMPLconst x [c]))
   474  (CMPW x (MOVLconst [c])) => (CMPWconst x [int16(c)])
   475  (CMPW (MOVLconst [c]) x) => (InvertFlags (CMPWconst x [int16(c)]))
   476  (CMPB x (MOVLconst [c])) => (CMPBconst x [int8(c)])
   477  (CMPB (MOVLconst [c]) x) => (InvertFlags (CMPBconst x [int8(c)]))
   478  
   479  // Canonicalize the order of arguments to comparisons - helps with CSE.
   480  (CMP(L|W|B) x y) && canonLessThan(x,y) => (InvertFlags (CMP(L|W|B) y x))
   481  
   482  // strength reduction
   483  // Assumes that the following costs from https://gmplib.org/~tege/x86-timing.pdf:
   484  //    1 - addl, shll, leal, negl, subl
   485  //    3 - imull
   486  // This limits the rewrites to two instructions.
   487  // Note that negl always operates in-place,
   488  // which can require a register-register move
   489  // to preserve the original value,
   490  // so it must be used with care.
   491  (MULLconst [-9] x) => (NEGL (LEAL8 <v.Type> x x))
   492  (MULLconst [-5] x) => (NEGL (LEAL4 <v.Type> x x))
   493  (MULLconst [-3] x) => (NEGL (LEAL2 <v.Type> x x))
   494  (MULLconst [-1] x) => (NEGL x)
   495  (MULLconst [0] _) => (MOVLconst [0])
   496  (MULLconst [1] x) => x
   497  (MULLconst [3] x) => (LEAL2 x x)
   498  (MULLconst [5] x) => (LEAL4 x x)
   499  (MULLconst [7] x) => (LEAL2 x (LEAL2 <v.Type> x x))
   500  (MULLconst [9] x) => (LEAL8 x x)
   501  (MULLconst [11] x) => (LEAL2 x (LEAL4 <v.Type> x x))
   502  (MULLconst [13] x) => (LEAL4 x (LEAL2 <v.Type> x x))
   503  (MULLconst [19] x) => (LEAL2 x (LEAL8 <v.Type> x x))
   504  (MULLconst [21] x) => (LEAL4 x (LEAL4 <v.Type> x x))
   505  (MULLconst [25] x) => (LEAL8 x (LEAL2 <v.Type> x x))
   506  (MULLconst [27] x) => (LEAL8 (LEAL2 <v.Type> x x) (LEAL2 <v.Type> x x))
   507  (MULLconst [37] x) => (LEAL4 x (LEAL8 <v.Type> x x))
   508  (MULLconst [41] x) => (LEAL8 x (LEAL4 <v.Type> x x))
   509  (MULLconst [45] x) => (LEAL8 (LEAL4 <v.Type> x x) (LEAL4 <v.Type> x x))
   510  (MULLconst [73] x) => (LEAL8 x (LEAL8 <v.Type> x x))
   511  (MULLconst [81] x) => (LEAL8 (LEAL8 <v.Type> x x) (LEAL8 <v.Type> x x))
   512  
   513  (MULLconst [c] x) && isPowerOfTwo32(c+1) && c >= 15 => (SUBL (SHLLconst <v.Type> [int32(log32(c+1))] x) x)
   514  (MULLconst [c] x) && isPowerOfTwo32(c-1) && c >= 17 => (LEAL1 (SHLLconst <v.Type> [int32(log32(c-1))] x) x)
   515  (MULLconst [c] x) && isPowerOfTwo32(c-2) && c >= 34 => (LEAL2 (SHLLconst <v.Type> [int32(log32(c-2))] x) x)
   516  (MULLconst [c] x) && isPowerOfTwo32(c-4) && c >= 68 => (LEAL4 (SHLLconst <v.Type> [int32(log32(c-4))] x) x)
   517  (MULLconst [c] x) && isPowerOfTwo32(c-8) && c >= 136 => (LEAL8 (SHLLconst <v.Type> [int32(log32(c-8))] x) x)
   518  (MULLconst [c] x) && c%3 == 0 && isPowerOfTwo32(c/3) => (SHLLconst [int32(log32(c/3))] (LEAL2 <v.Type> x x))
   519  (MULLconst [c] x) && c%5 == 0 && isPowerOfTwo32(c/5) => (SHLLconst [int32(log32(c/5))] (LEAL4 <v.Type> x x))
   520  (MULLconst [c] x) && c%9 == 0 && isPowerOfTwo32(c/9) => (SHLLconst [int32(log32(c/9))] (LEAL8 <v.Type> x x))
   521  
   522  // combine add/shift into LEAL
   523  (ADDL x (SHLLconst [3] y)) => (LEAL8 x y)
   524  (ADDL x (SHLLconst [2] y)) => (LEAL4 x y)
   525  (ADDL x (SHLLconst [1] y)) => (LEAL2 x y)
   526  (ADDL x (ADDL y y)) => (LEAL2 x y)
   527  (ADDL x (ADDL x y)) => (LEAL2 y x)
   528  
   529  // combine ADDL/ADDLconst into LEAL1
   530  (ADDLconst [c] (ADDL x y)) => (LEAL1 [c] x y)
   531  (ADDL (ADDLconst [c] x) y) => (LEAL1 [c] x y)
   532  
   533  // fold ADDL into LEAL
   534  (ADDLconst [c] (LEAL [d] {s} x)) && is32Bit(int64(c)+int64(d)) => (LEAL [c+d] {s} x)
   535  (LEAL [c] {s} (ADDLconst [d] x)) && is32Bit(int64(c)+int64(d)) => (LEAL [c+d] {s} x)
   536  (ADDLconst [c] x:(SP)) => (LEAL [c] x) // so it is rematerializeable
   537  (LEAL [c] {s} (ADDL x y)) && x.Op != OpSB && y.Op != OpSB => (LEAL1 [c] {s} x y)
   538  (ADDL x (LEAL [c] {s} y)) && x.Op != OpSB && y.Op != OpSB => (LEAL1 [c] {s} x y)
   539  
   540  // fold ADDLconst into LEALx
   541  (ADDLconst [c] (LEAL1 [d] {s} x y)) && is32Bit(int64(c)+int64(d)) => (LEAL1 [c+d] {s} x y)
   542  (ADDLconst [c] (LEAL2 [d] {s} x y)) && is32Bit(int64(c)+int64(d)) => (LEAL2 [c+d] {s} x y)
   543  (ADDLconst [c] (LEAL4 [d] {s} x y)) && is32Bit(int64(c)+int64(d)) => (LEAL4 [c+d] {s} x y)
   544  (ADDLconst [c] (LEAL8 [d] {s} x y)) && is32Bit(int64(c)+int64(d)) => (LEAL8 [c+d] {s} x y)
   545  (LEAL1 [c] {s} (ADDLconst [d] x) y) && is32Bit(int64(c)+int64(d))   && x.Op != OpSB => (LEAL1 [c+d] {s} x y)
   546  (LEAL2 [c] {s} (ADDLconst [d] x) y) && is32Bit(int64(c)+int64(d))   && x.Op != OpSB => (LEAL2 [c+d] {s} x y)
   547  (LEAL2 [c] {s} x (ADDLconst [d] y)) && is32Bit(int64(c)+2*int64(d)) && y.Op != OpSB => (LEAL2 [c+2*d] {s} x y)
   548  (LEAL4 [c] {s} (ADDLconst [d] x) y) && is32Bit(int64(c)+int64(d))   && x.Op != OpSB => (LEAL4 [c+d] {s} x y)
   549  (LEAL4 [c] {s} x (ADDLconst [d] y)) && is32Bit(int64(c)+4*int64(d)) && y.Op != OpSB => (LEAL4 [c+4*d] {s} x y)
   550  (LEAL8 [c] {s} (ADDLconst [d] x) y) && is32Bit(int64(c)+int64(d))   && x.Op != OpSB => (LEAL8 [c+d] {s} x y)
   551  (LEAL8 [c] {s} x (ADDLconst [d] y)) && is32Bit(int64(c)+8*int64(d)) && y.Op != OpSB => (LEAL8 [c+8*d] {s} x y)
   552  
   553  // fold shifts into LEALx
   554  (LEAL1 [c] {s} x (SHLLconst [1] y)) => (LEAL2 [c] {s} x y)
   555  (LEAL1 [c] {s} x (SHLLconst [2] y)) => (LEAL4 [c] {s} x y)
   556  (LEAL1 [c] {s} x (SHLLconst [3] y)) => (LEAL8 [c] {s} x y)
   557  (LEAL2 [c] {s} x (SHLLconst [1] y)) => (LEAL4 [c] {s} x y)
   558  (LEAL2 [c] {s} x (SHLLconst [2] y)) => (LEAL8 [c] {s} x y)
   559  (LEAL4 [c] {s} x (SHLLconst [1] y)) => (LEAL8 [c] {s} x y)
   560  
   561  // reverse ordering of compare instruction
   562  (SETL (InvertFlags x)) => (SETG x)
   563  (SETG (InvertFlags x)) => (SETL x)
   564  (SETB (InvertFlags x)) => (SETA x)
   565  (SETA (InvertFlags x)) => (SETB x)
   566  (SETLE (InvertFlags x)) => (SETGE x)
   567  (SETGE (InvertFlags x)) => (SETLE x)
   568  (SETBE (InvertFlags x)) => (SETAE x)
   569  (SETAE (InvertFlags x)) => (SETBE x)
   570  (SETEQ (InvertFlags x)) => (SETEQ x)
   571  (SETNE (InvertFlags x)) => (SETNE x)
   572  
   573  // sign extended loads
   574  // Note: The combined instruction must end up in the same block
   575  // as the original load. If not, we end up making a value with
   576  // memory type live in two different blocks, which can lead to
   577  // multiple memory values alive simultaneously.
   578  // Make sure we don't combine these ops if the load has another use.
   579  // This prevents a single load from being split into multiple loads
   580  // which then might return different values.  See test/atomicload.go.
   581  (MOVBLSX x:(MOVBload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVBLSXload <v.Type> [off] {sym} ptr mem)
   582  (MOVBLZX x:(MOVBload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVBload <v.Type> [off] {sym} ptr mem)
   583  (MOVWLSX x:(MOVWload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVWLSXload <v.Type> [off] {sym} ptr mem)
   584  (MOVWLZX x:(MOVWload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVWload <v.Type> [off] {sym} ptr mem)
   585  
   586  // replace load from same location as preceding store with zero/sign extension (or copy in case of full width)
   587  (MOVBload [off] {sym} ptr (MOVBstore [off2] {sym2} ptr2 x _)) && sym == sym2 && off == off2 && isSamePtr(ptr, ptr2) => (MOVBLZX x)
   588  (MOVWload [off] {sym} ptr (MOVWstore [off2] {sym2} ptr2 x _)) && sym == sym2 && off == off2 && isSamePtr(ptr, ptr2) => (MOVWLZX x)
   589  (MOVLload [off] {sym} ptr (MOVLstore [off2] {sym2} ptr2 x _)) && sym == sym2 && off == off2 && isSamePtr(ptr, ptr2) => x
   590  (MOVBLSXload [off] {sym} ptr (MOVBstore [off2] {sym2} ptr2 x _)) && sym == sym2 && off == off2 && isSamePtr(ptr, ptr2) => (MOVBLSX x)
   591  (MOVWLSXload [off] {sym} ptr (MOVWstore [off2] {sym2} ptr2 x _)) && sym == sym2 && off == off2 && isSamePtr(ptr, ptr2) => (MOVWLSX x)
   592  
   593  // Fold extensions and ANDs together.
   594  (MOVBLZX (ANDLconst [c] x)) => (ANDLconst [c & 0xff] x)
   595  (MOVWLZX (ANDLconst [c] x)) => (ANDLconst [c & 0xffff] x)
   596  (MOVBLSX (ANDLconst [c] x)) && c & 0x80 == 0 => (ANDLconst [c & 0x7f] x)
   597  (MOVWLSX (ANDLconst [c] x)) && c & 0x8000 == 0 => (ANDLconst [c & 0x7fff] x)
   598  
   599  // Don't extend before storing
   600  (MOVWstore [off] {sym} ptr (MOVWL(S|Z)X x) mem) => (MOVWstore [off] {sym} ptr x mem)
   601  (MOVBstore [off] {sym} ptr (MOVBL(S|Z)X x) mem) => (MOVBstore [off] {sym} ptr x mem)
   602  
   603  // fold constants into memory operations
   604  // Note that this is not always a good idea because if not all the uses of
   605  // the ADDLconst get eliminated, we still have to compute the ADDLconst and we now
   606  // have potentially two live values (ptr and (ADDLconst [off] ptr)) instead of one.
   607  // Nevertheless, let's do it!
   608  (MOV(L|W|B|SS|SD)load  [off1] {sym} (ADDLconst [off2] ptr) mem) && is32Bit(int64(off1)+int64(off2)) =>
   609      (MOV(L|W|B|SS|SD)load  [off1+off2] {sym} ptr mem)
   610  (MOV(L|W|B|SS|SD)store  [off1] {sym} (ADDLconst [off2] ptr) val mem) && is32Bit(int64(off1)+int64(off2)) =>
   611      (MOV(L|W|B|SS|SD)store  [off1+off2] {sym} ptr val mem)
   612  
   613  ((ADD|SUB|MUL|AND|OR|XOR)Lload [off1] {sym} val (ADDLconst [off2] base) mem) && is32Bit(int64(off1)+int64(off2)) =>
   614  	((ADD|SUB|MUL|AND|OR|XOR)Lload [off1+off2] {sym} val base mem)
   615  ((ADD|SUB|MUL|DIV)SSload [off1] {sym} val (ADDLconst [off2] base) mem) && is32Bit(int64(off1)+int64(off2)) =>
   616  	((ADD|SUB|MUL|DIV)SSload [off1+off2] {sym} val base mem)
   617  ((ADD|SUB|MUL|DIV)SDload [off1] {sym} val (ADDLconst [off2] base) mem) && is32Bit(int64(off1)+int64(off2)) =>
   618  	((ADD|SUB|MUL|DIV)SDload [off1+off2] {sym} val base mem)
   619  ((ADD|SUB|AND|OR|XOR)Lmodify [off1] {sym} (ADDLconst [off2] base) val mem) && is32Bit(int64(off1)+int64(off2)) =>
   620  	((ADD|SUB|AND|OR|XOR)Lmodify [off1+off2] {sym} base val mem)
   621  ((ADD|AND|OR|XOR)Lconstmodify [valoff1] {sym} (ADDLconst [off2] base) mem) && valoff1.canAdd32(off2) =>
   622  	((ADD|AND|OR|XOR)Lconstmodify [valoff1.addOffset32(off2)] {sym} base mem)
   623  
   624  // Fold constants into stores.
   625  (MOVLstore [off] {sym} ptr (MOVLconst [c]) mem) =>
   626  	(MOVLstoreconst [makeValAndOff(c,off)] {sym} ptr mem)
   627  (MOVWstore [off] {sym} ptr (MOVLconst [c]) mem) =>
   628  	(MOVWstoreconst [makeValAndOff(c,off)] {sym} ptr mem)
   629  (MOVBstore [off] {sym} ptr (MOVLconst [c]) mem) =>
   630  	(MOVBstoreconst [makeValAndOff(c,off)] {sym} ptr mem)
   631  
   632  // Fold address offsets into constant stores.
   633  (MOV(L|W|B)storeconst [sc] {s} (ADDLconst [off] ptr) mem) && sc.canAdd32(off) =>
   634  	(MOV(L|W|B)storeconst [sc.addOffset32(off)] {s} ptr mem)
   635  
   636  // We need to fold LEAL into the MOVx ops so that the live variable analysis knows
   637  // what variables are being read/written by the ops.
   638  // Note: we turn off this merging for operations on globals when building
   639  // position-independent code (when Flag_shared is set).
   640  // PIC needs a spare register to load the PC into.  Having the LEAL be
   641  // a separate instruction gives us that register.  Having the LEAL be
   642  // a separate instruction also allows it to be CSEd (which is good because
   643  // it compiles to a thunk call).
   644  (MOV(L|W|B|SS|SD|BLSX|WLSX)load  [off1] {sym1} (LEAL [off2] {sym2} base) mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2)
   645    && (base.Op != OpSB || !config.ctxt.Flag_shared) =>
   646          (MOV(L|W|B|SS|SD|BLSX|WLSX)load  [off1+off2] {mergeSym(sym1,sym2)} base mem)
   647  
   648  (MOV(L|W|B|SS|SD)store  [off1] {sym1} (LEAL [off2] {sym2} base) val mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2)
   649    && (base.Op != OpSB || !config.ctxt.Flag_shared) =>
   650  	(MOV(L|W|B|SS|SD)store  [off1+off2] {mergeSym(sym1,sym2)} base val mem)
   651  
   652  (MOV(L|W|B)storeconst [sc] {sym1} (LEAL [off] {sym2} ptr) mem) && canMergeSym(sym1, sym2) && sc.canAdd32(off)
   653    && (ptr.Op != OpSB || !config.ctxt.Flag_shared) =>
   654  	(MOV(L|W|B)storeconst [sc.addOffset32(off)] {mergeSym(sym1, sym2)} ptr mem)
   655  
   656  ((ADD|SUB|MUL|AND|OR|XOR)Lload [off1] {sym1} val (LEAL [off2] {sym2} base) mem)
   657  	&& is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) && (base.Op != OpSB || !config.ctxt.Flag_shared) =>
   658  	((ADD|SUB|MUL|AND|OR|XOR)Lload [off1+off2] {mergeSym(sym1,sym2)} val base mem)
   659  ((ADD|SUB|MUL|DIV)SSload [off1] {sym1} val (LEAL [off2] {sym2} base) mem)
   660  	&& is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) && (base.Op != OpSB || !config.ctxt.Flag_shared) =>
   661  	((ADD|SUB|MUL|DIV)SSload [off1+off2] {mergeSym(sym1,sym2)} val base mem)
   662  ((ADD|SUB|MUL|DIV)SDload [off1] {sym1} val (LEAL [off2] {sym2} base) mem)
   663  	&& is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) && (base.Op != OpSB || !config.ctxt.Flag_shared) =>
   664  	((ADD|SUB|MUL|DIV)SDload [off1+off2] {mergeSym(sym1,sym2)} val base mem)
   665  ((ADD|SUB|AND|OR|XOR)Lmodify [off1] {sym1} (LEAL [off2] {sym2} base) val mem)
   666  	&& is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) && (base.Op != OpSB || !config.ctxt.Flag_shared) =>
   667  	((ADD|SUB|AND|OR|XOR)Lmodify [off1+off2] {mergeSym(sym1,sym2)} base val mem)
   668  ((ADD|AND|OR|XOR)Lconstmodify [valoff1] {sym1} (LEAL [off2] {sym2} base) mem)
   669  	&& valoff1.canAdd32(off2) && canMergeSym(sym1, sym2) && (base.Op != OpSB || !config.ctxt.Flag_shared) =>
   670  	((ADD|AND|OR|XOR)Lconstmodify [valoff1.addOffset32(off2)] {mergeSym(sym1,sym2)} base mem)
   671  
   672  // Merge load/store to op
   673  ((ADD|AND|OR|XOR|SUB|MUL)L x l:(MOVLload [off] {sym} ptr mem)) && canMergeLoadClobber(v, l, x) && clobber(l) => ((ADD|AND|OR|XOR|SUB|MUL)Lload x [off] {sym} ptr mem)
   674  ((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)
   675  ((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)
   676  (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)
   677  (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) =>
   678  	((ADD|SUB|AND|OR|XOR)Lmodify [off] {sym} ptr x mem)
   679  (MOVLstore {sym} [off] ptr y:((ADD|AND|OR|XOR)Lconst [c] l:(MOVLload [off] {sym} ptr mem)) mem)
   680  	&& y.Uses==1 && l.Uses==1 && clobber(y, l) =>
   681  	((ADD|AND|OR|XOR)Lconstmodify [makeValAndOff(c,off)] {sym} ptr mem)
   682  
   683  // fold LEALs together
   684  (LEAL [off1] {sym1} (LEAL [off2] {sym2} x)) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   685        (LEAL [off1+off2] {mergeSym(sym1,sym2)} x)
   686  
   687  // LEAL into LEAL1
   688  (LEAL1 [off1] {sym1} (LEAL [off2] {sym2} x) y) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) && x.Op != OpSB =>
   689         (LEAL1 [off1+off2] {mergeSym(sym1,sym2)} x y)
   690  
   691  // LEAL1 into LEAL
   692  (LEAL [off1] {sym1} (LEAL1 [off2] {sym2} x y)) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   693         (LEAL1 [off1+off2] {mergeSym(sym1,sym2)} x y)
   694  
   695  // LEAL into LEAL[248]
   696  (LEAL2 [off1] {sym1} (LEAL [off2] {sym2} x) y) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) && x.Op != OpSB =>
   697         (LEAL2 [off1+off2] {mergeSym(sym1,sym2)} x y)
   698  (LEAL4 [off1] {sym1} (LEAL [off2] {sym2} x) y) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) && x.Op != OpSB =>
   699         (LEAL4 [off1+off2] {mergeSym(sym1,sym2)} x y)
   700  (LEAL8 [off1] {sym1} (LEAL [off2] {sym2} x) y) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) && x.Op != OpSB =>
   701         (LEAL8 [off1+off2] {mergeSym(sym1,sym2)} x y)
   702  
   703  // LEAL[248] into LEAL
   704  (LEAL [off1] {sym1} (LEAL2 [off2] {sym2} x y)) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   705        (LEAL2 [off1+off2] {mergeSym(sym1,sym2)} x y)
   706  (LEAL [off1] {sym1} (LEAL4 [off2] {sym2} x y)) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   707        (LEAL4 [off1+off2] {mergeSym(sym1,sym2)} x y)
   708  (LEAL [off1] {sym1} (LEAL8 [off2] {sym2} x y)) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   709        (LEAL8 [off1+off2] {mergeSym(sym1,sym2)} x y)
   710  
   711  // LEAL[1248] into LEAL[1248]. Only some such merges are possible.
   712  (LEAL1 [off1] {sym1} x (LEAL1 [off2] {sym2} y y)) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   713        (LEAL2 [off1+off2] {mergeSym(sym1, sym2)} x y)
   714  (LEAL1 [off1] {sym1} x (LEAL1 [off2] {sym2} x y)) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   715        (LEAL2 [off1+off2] {mergeSym(sym1, sym2)} y x)
   716  (LEAL2 [off1] {sym} x (LEAL1 [off2] {nil} y y)) && is32Bit(int64(off1)+2*int64(off2)) =>
   717        (LEAL4 [off1+2*off2] {sym} x y)
   718  (LEAL4 [off1] {sym} x (LEAL1 [off2] {nil} y y)) && is32Bit(int64(off1)+4*int64(off2)) =>
   719        (LEAL8 [off1+4*off2] {sym} x y)
   720  
   721  // Absorb InvertFlags into branches.
   722  (LT (InvertFlags cmp) yes no) => (GT cmp yes no)
   723  (GT (InvertFlags cmp) yes no) => (LT cmp yes no)
   724  (LE (InvertFlags cmp) yes no) => (GE cmp yes no)
   725  (GE (InvertFlags cmp) yes no) => (LE cmp yes no)
   726  (ULT (InvertFlags cmp) yes no) => (UGT cmp yes no)
   727  (UGT (InvertFlags cmp) yes no) => (ULT cmp yes no)
   728  (ULE (InvertFlags cmp) yes no) => (UGE cmp yes no)
   729  (UGE (InvertFlags cmp) yes no) => (ULE cmp yes no)
   730  (EQ (InvertFlags cmp) yes no) => (EQ cmp yes no)
   731  (NE (InvertFlags cmp) yes no) => (NE cmp yes no)
   732  
   733  // Constant comparisons.
   734  (CMPLconst (MOVLconst [x]) [y]) && x==y                       => (FlagEQ)
   735  (CMPLconst (MOVLconst [x]) [y]) && x<y && uint32(x)<uint32(y) => (FlagLT_ULT)
   736  (CMPLconst (MOVLconst [x]) [y]) && x<y && uint32(x)>uint32(y) => (FlagLT_UGT)
   737  (CMPLconst (MOVLconst [x]) [y]) && x>y && uint32(x)<uint32(y) => (FlagGT_ULT)
   738  (CMPLconst (MOVLconst [x]) [y]) && x>y && uint32(x)>uint32(y) => (FlagGT_UGT)
   739  
   740  (CMPWconst (MOVLconst [x]) [y]) && int16(x)==y                       => (FlagEQ)
   741  (CMPWconst (MOVLconst [x]) [y]) && int16(x)<y && uint16(x)<uint16(y) => (FlagLT_ULT)
   742  (CMPWconst (MOVLconst [x]) [y]) && int16(x)<y && uint16(x)>uint16(y) => (FlagLT_UGT)
   743  (CMPWconst (MOVLconst [x]) [y]) && int16(x)>y && uint16(x)<uint16(y) => (FlagGT_ULT)
   744  (CMPWconst (MOVLconst [x]) [y]) && int16(x)>y && uint16(x)>uint16(y) => (FlagGT_UGT)
   745  
   746  (CMPBconst (MOVLconst [x]) [y]) && int8(x)==y                      => (FlagEQ)
   747  (CMPBconst (MOVLconst [x]) [y]) && int8(x)<y && uint8(x)<uint8(y) => (FlagLT_ULT)
   748  (CMPBconst (MOVLconst [x]) [y]) && int8(x)<y && uint8(x)>uint8(y) => (FlagLT_UGT)
   749  (CMPBconst (MOVLconst [x]) [y]) && int8(x)>y && uint8(x)<uint8(y) => (FlagGT_ULT)
   750  (CMPBconst (MOVLconst [x]) [y]) && int8(x)>y && uint8(x)>uint8(y) => (FlagGT_UGT)
   751  
   752  // Other known comparisons.
   753  (CMPLconst (SHRLconst _ [c]) [n]) && 0 <= n && 0 < c && c <= 32 && (1<<uint64(32-c)) <= uint64(n) => (FlagLT_ULT)
   754  (CMPLconst (ANDLconst _ [m]) [n]) && 0 <= m && m < n => (FlagLT_ULT)
   755  (CMPWconst (ANDLconst _ [m]) [n]) && 0 <= int16(m) && int16(m) < n => (FlagLT_ULT)
   756  (CMPBconst (ANDLconst _ [m]) [n]) && 0 <= int8(m) && int8(m) < n => (FlagLT_ULT)
   757  // TODO: DIVxU also.
   758  
   759  // Absorb flag constants into SBB ops.
   760  (SBBLcarrymask (FlagEQ)) => (MOVLconst [0])
   761  (SBBLcarrymask (FlagLT_ULT)) => (MOVLconst [-1])
   762  (SBBLcarrymask (FlagLT_UGT)) => (MOVLconst [0])
   763  (SBBLcarrymask (FlagGT_ULT)) => (MOVLconst [-1])
   764  (SBBLcarrymask (FlagGT_UGT)) => (MOVLconst [0])
   765  
   766  // Absorb flag constants into branches.
   767  (EQ (FlagEQ) yes no) => (First yes no)
   768  (EQ (FlagLT_ULT) yes no) => (First no yes)
   769  (EQ (FlagLT_UGT) yes no) => (First no yes)
   770  (EQ (FlagGT_ULT) yes no) => (First no yes)
   771  (EQ (FlagGT_UGT) yes no) => (First no yes)
   772  
   773  (NE (FlagEQ) yes no) => (First no yes)
   774  (NE (FlagLT_ULT) yes no) => (First yes no)
   775  (NE (FlagLT_UGT) yes no) => (First yes no)
   776  (NE (FlagGT_ULT) yes no) => (First yes no)
   777  (NE (FlagGT_UGT) yes no) => (First yes no)
   778  
   779  (LT (FlagEQ) yes no) => (First no yes)
   780  (LT (FlagLT_ULT) yes no) => (First yes no)
   781  (LT (FlagLT_UGT) yes no) => (First yes no)
   782  (LT (FlagGT_ULT) yes no) => (First no yes)
   783  (LT (FlagGT_UGT) yes no) => (First no yes)
   784  
   785  (LE (FlagEQ) yes no) => (First yes no)
   786  (LE (FlagLT_ULT) yes no) => (First yes no)
   787  (LE (FlagLT_UGT) yes no) => (First yes no)
   788  (LE (FlagGT_ULT) yes no) => (First no yes)
   789  (LE (FlagGT_UGT) yes no) => (First no yes)
   790  
   791  (GT (FlagEQ) yes no) => (First no yes)
   792  (GT (FlagLT_ULT) yes no) => (First no yes)
   793  (GT (FlagLT_UGT) yes no) => (First no yes)
   794  (GT (FlagGT_ULT) yes no) => (First yes no)
   795  (GT (FlagGT_UGT) yes no) => (First yes no)
   796  
   797  (GE (FlagEQ) yes no) => (First yes no)
   798  (GE (FlagLT_ULT) yes no) => (First no yes)
   799  (GE (FlagLT_UGT) yes no) => (First no yes)
   800  (GE (FlagGT_ULT) yes no) => (First yes no)
   801  (GE (FlagGT_UGT) yes no) => (First yes no)
   802  
   803  (ULT (FlagEQ) yes no) => (First no yes)
   804  (ULT (FlagLT_ULT) yes no) => (First yes no)
   805  (ULT (FlagLT_UGT) yes no) => (First no yes)
   806  (ULT (FlagGT_ULT) yes no) => (First yes no)
   807  (ULT (FlagGT_UGT) yes no) => (First no yes)
   808  
   809  (ULE (FlagEQ) yes no) => (First yes no)
   810  (ULE (FlagLT_ULT) yes no) => (First yes no)
   811  (ULE (FlagLT_UGT) yes no) => (First no yes)
   812  (ULE (FlagGT_ULT) yes no) => (First yes no)
   813  (ULE (FlagGT_UGT) yes no) => (First no yes)
   814  
   815  (UGT (FlagEQ) yes no) => (First no yes)
   816  (UGT (FlagLT_ULT) yes no) => (First no yes)
   817  (UGT (FlagLT_UGT) yes no) => (First yes no)
   818  (UGT (FlagGT_ULT) yes no) => (First no yes)
   819  (UGT (FlagGT_UGT) yes no) => (First yes no)
   820  
   821  (UGE (FlagEQ) yes no) => (First yes no)
   822  (UGE (FlagLT_ULT) yes no) => (First no yes)
   823  (UGE (FlagLT_UGT) yes no) => (First yes no)
   824  (UGE (FlagGT_ULT) yes no) => (First no yes)
   825  (UGE (FlagGT_UGT) yes no) => (First yes no)
   826  
   827  // Absorb flag constants into SETxx ops.
   828  (SETEQ (FlagEQ)) => (MOVLconst [1])
   829  (SETEQ (FlagLT_ULT)) => (MOVLconst [0])
   830  (SETEQ (FlagLT_UGT)) => (MOVLconst [0])
   831  (SETEQ (FlagGT_ULT)) => (MOVLconst [0])
   832  (SETEQ (FlagGT_UGT)) => (MOVLconst [0])
   833  
   834  (SETNE (FlagEQ)) => (MOVLconst [0])
   835  (SETNE (FlagLT_ULT)) => (MOVLconst [1])
   836  (SETNE (FlagLT_UGT)) => (MOVLconst [1])
   837  (SETNE (FlagGT_ULT)) => (MOVLconst [1])
   838  (SETNE (FlagGT_UGT)) => (MOVLconst [1])
   839  
   840  (SETL (FlagEQ)) => (MOVLconst [0])
   841  (SETL (FlagLT_ULT)) => (MOVLconst [1])
   842  (SETL (FlagLT_UGT)) => (MOVLconst [1])
   843  (SETL (FlagGT_ULT)) => (MOVLconst [0])
   844  (SETL (FlagGT_UGT)) => (MOVLconst [0])
   845  
   846  (SETLE (FlagEQ)) => (MOVLconst [1])
   847  (SETLE (FlagLT_ULT)) => (MOVLconst [1])
   848  (SETLE (FlagLT_UGT)) => (MOVLconst [1])
   849  (SETLE (FlagGT_ULT)) => (MOVLconst [0])
   850  (SETLE (FlagGT_UGT)) => (MOVLconst [0])
   851  
   852  (SETG (FlagEQ)) => (MOVLconst [0])
   853  (SETG (FlagLT_ULT)) => (MOVLconst [0])
   854  (SETG (FlagLT_UGT)) => (MOVLconst [0])
   855  (SETG (FlagGT_ULT)) => (MOVLconst [1])
   856  (SETG (FlagGT_UGT)) => (MOVLconst [1])
   857  
   858  (SETGE (FlagEQ)) => (MOVLconst [1])
   859  (SETGE (FlagLT_ULT)) => (MOVLconst [0])
   860  (SETGE (FlagLT_UGT)) => (MOVLconst [0])
   861  (SETGE (FlagGT_ULT)) => (MOVLconst [1])
   862  (SETGE (FlagGT_UGT)) => (MOVLconst [1])
   863  
   864  (SETB (FlagEQ)) => (MOVLconst [0])
   865  (SETB (FlagLT_ULT)) => (MOVLconst [1])
   866  (SETB (FlagLT_UGT)) => (MOVLconst [0])
   867  (SETB (FlagGT_ULT)) => (MOVLconst [1])
   868  (SETB (FlagGT_UGT)) => (MOVLconst [0])
   869  
   870  (SETBE (FlagEQ)) => (MOVLconst [1])
   871  (SETBE (FlagLT_ULT)) => (MOVLconst [1])
   872  (SETBE (FlagLT_UGT)) => (MOVLconst [0])
   873  (SETBE (FlagGT_ULT)) => (MOVLconst [1])
   874  (SETBE (FlagGT_UGT)) => (MOVLconst [0])
   875  
   876  (SETA (FlagEQ)) => (MOVLconst [0])
   877  (SETA (FlagLT_ULT)) => (MOVLconst [0])
   878  (SETA (FlagLT_UGT)) => (MOVLconst [1])
   879  (SETA (FlagGT_ULT)) => (MOVLconst [0])
   880  (SETA (FlagGT_UGT)) => (MOVLconst [1])
   881  
   882  (SETAE (FlagEQ)) => (MOVLconst [1])
   883  (SETAE (FlagLT_ULT)) => (MOVLconst [0])
   884  (SETAE (FlagLT_UGT)) => (MOVLconst [1])
   885  (SETAE (FlagGT_ULT)) => (MOVLconst [0])
   886  (SETAE (FlagGT_UGT)) => (MOVLconst [1])
   887  
   888  // Remove redundant *const ops
   889  (ADDLconst [c] x) && c==0  => x
   890  (SUBLconst [c] x) && c==0  => x
   891  (ANDLconst [c] _) && c==0  => (MOVLconst [0])
   892  (ANDLconst [c] x) && c==-1 => x
   893  (ORLconst [c] x)  && c==0  => x
   894  (ORLconst [c] _)  && c==-1 => (MOVLconst [-1])
   895  (XORLconst [c] x) && c==0  => x
   896  // TODO: since we got rid of the W/B versions, we might miss
   897  // things like (ANDLconst [0x100] x) which were formerly
   898  // (ANDBconst [0] x).  Probably doesn't happen very often.
   899  // If we cared, we might do:
   900  //  (ANDLconst <t> [c] x) && t.Size()==1 && int8(x)==0 => (MOVLconst [0])
   901  
   902  // Convert constant subtracts to constant adds
   903  (SUBLconst [c] x) => (ADDLconst [-c] x)
   904  
   905  // generic constant folding
   906  // TODO: more of this
   907  (ADDLconst [c] (MOVLconst [d])) => (MOVLconst [c+d])
   908  (ADDLconst [c] (ADDLconst [d] x)) => (ADDLconst [c+d] x)
   909  (SARLconst [c] (MOVLconst [d])) => (MOVLconst [d>>uint64(c)])
   910  (SARWconst [c] (MOVLconst [d])) => (MOVLconst [d>>uint64(c)])
   911  (SARBconst [c] (MOVLconst [d])) => (MOVLconst [d>>uint64(c)])
   912  (NEGL (MOVLconst [c])) => (MOVLconst [-c])
   913  (MULLconst [c] (MOVLconst [d])) => (MOVLconst [c*d])
   914  (ANDLconst [c] (MOVLconst [d])) => (MOVLconst [c&d])
   915  (ORLconst [c] (MOVLconst [d])) => (MOVLconst [c|d])
   916  (XORLconst [c] (MOVLconst [d])) => (MOVLconst [c^d])
   917  (NOTL (MOVLconst [c])) => (MOVLconst [^c])
   918  
   919  // generic simplifications
   920  // TODO: more of this
   921  (ADDL x (NEGL y)) => (SUBL x y)
   922  (SUBL x x) => (MOVLconst [0])
   923  (ANDL x x) => x
   924  (ORL x x) => x
   925  (XORL x x) => (MOVLconst [0])
   926  
   927  // checking AND against 0.
   928  (CMP(L|W|B)const l:(ANDL x y) [0]) && l.Uses==1 => (TEST(L|W|B) x y)
   929  (CMPLconst l:(ANDLconst [c] x) [0]) && l.Uses==1 => (TESTLconst [c] x)
   930  (CMPWconst l:(ANDLconst [c] x) [0]) && l.Uses==1 => (TESTWconst [int16(c)] x)
   931  (CMPBconst l:(ANDLconst [c] x) [0]) && l.Uses==1 => (TESTBconst [int8(c)] x)
   932  
   933  // TEST %reg,%reg is shorter than CMP
   934  (CMP(L|W|B)const x [0]) => (TEST(L|W|B) x x)
   935  
   936  // Convert LEAL1 back to ADDL if we can
   937  (LEAL1 [0] {nil} x y) => (ADDL x y)
   938  
   939  // Combining byte loads into larger (unaligned) loads.
   940  // There are many ways these combinations could occur.  This is
   941  // designed to match the way encoding/binary.LittleEndian does it.
   942  (ORL                  x0:(MOVBload [i0] {s} p mem)
   943      s0:(SHLLconst [8] x1:(MOVBload [i1] {s} p mem)))
   944    && i1 == i0+1
   945    && x0.Uses == 1
   946    && x1.Uses == 1
   947    && s0.Uses == 1
   948    && mergePoint(b,x0,x1) != nil
   949    && clobber(x0, x1, s0)
   950    => @mergePoint(b,x0,x1) (MOVWload [i0] {s} p mem)
   951  
   952  (ORL                  x0:(MOVBload [i] {s} p0 mem)
   953      s0:(SHLLconst [8] x1:(MOVBload [i] {s} p1 mem)))
   954    && x0.Uses == 1
   955    && x1.Uses == 1
   956    && s0.Uses == 1
   957    && sequentialAddresses(p0, p1, 1)
   958    && mergePoint(b,x0,x1) != nil
   959    && clobber(x0, x1, s0)
   960    => @mergePoint(b,x0,x1) (MOVWload [i] {s} p0 mem)
   961  
   962  (ORL o0:(ORL
   963                         x0:(MOVWload [i0] {s} p mem)
   964      s0:(SHLLconst [16] x1:(MOVBload [i2] {s} p mem)))
   965      s1:(SHLLconst [24] x2:(MOVBload [i3] {s} p mem)))
   966    && i2 == i0+2
   967    && i3 == i0+3
   968    && x0.Uses == 1
   969    && x1.Uses == 1
   970    && x2.Uses == 1
   971    && s0.Uses == 1
   972    && s1.Uses == 1
   973    && o0.Uses == 1
   974    && mergePoint(b,x0,x1,x2) != nil
   975    && clobber(x0, x1, x2, s0, s1, o0)
   976    => @mergePoint(b,x0,x1,x2) (MOVLload [i0] {s} p mem)
   977  
   978  (ORL o0:(ORL
   979                         x0:(MOVWload [i] {s} p0 mem)
   980      s0:(SHLLconst [16] x1:(MOVBload [i] {s} p1 mem)))
   981      s1:(SHLLconst [24] x2:(MOVBload [i] {s} p2 mem)))
   982    && x0.Uses == 1
   983    && x1.Uses == 1
   984    && x2.Uses == 1
   985    && s0.Uses == 1
   986    && s1.Uses == 1
   987    && o0.Uses == 1
   988    && sequentialAddresses(p0, p1, 2)
   989    && sequentialAddresses(p1, p2, 1)
   990    && mergePoint(b,x0,x1,x2) != nil
   991    && clobber(x0, x1, x2, s0, s1, o0)
   992    => @mergePoint(b,x0,x1,x2) (MOVLload [i] {s} p0 mem)
   993  
   994  // Combine constant stores into larger (unaligned) stores.
   995  (MOVBstoreconst [c] {s} p x:(MOVBstoreconst [a] {s} p mem))
   996    && x.Uses == 1
   997    && a.Off() + 1 == c.Off()
   998    && clobber(x)
   999    => (MOVWstoreconst [makeValAndOff(a.Val()&0xff | c.Val()<<8, a.Off())] {s} p mem)
  1000  (MOVBstoreconst [a] {s} p x:(MOVBstoreconst [c] {s} p mem))
  1001    && x.Uses == 1
  1002    && a.Off() + 1 == c.Off()
  1003    && clobber(x)
  1004    => (MOVWstoreconst [makeValAndOff(a.Val()&0xff | c.Val()<<8, a.Off())] {s} p mem)
  1005  
  1006  (MOVBstoreconst [c] {s} p1 x:(MOVBstoreconst [a] {s} p0 mem))
  1007    && x.Uses == 1
  1008    && a.Off() == c.Off()
  1009    && sequentialAddresses(p0, p1, 1)
  1010    && clobber(x)
  1011    => (MOVWstoreconst [makeValAndOff(a.Val()&0xff | c.Val()<<8, a.Off())] {s} p0 mem)
  1012  (MOVBstoreconst [a] {s} p0 x:(MOVBstoreconst [c] {s} p1 mem))
  1013    && x.Uses == 1
  1014    && a.Off() == c.Off()
  1015    && sequentialAddresses(p0, p1, 1)
  1016    && clobber(x)
  1017    => (MOVWstoreconst [makeValAndOff(a.Val()&0xff | c.Val()<<8, a.Off())] {s} p0 mem)
  1018  
  1019  (MOVWstoreconst [c] {s} p x:(MOVWstoreconst [a] {s} p mem))
  1020    && x.Uses == 1
  1021    && a.Off() + 2 == c.Off()
  1022    && clobber(x)
  1023    => (MOVLstoreconst [makeValAndOff(a.Val()&0xffff | c.Val()<<16, a.Off())] {s} p mem)
  1024  (MOVWstoreconst [a] {s} p x:(MOVWstoreconst [c] {s} p mem))
  1025    && x.Uses == 1
  1026    && ValAndOff(a).Off() + 2 == ValAndOff(c).Off()
  1027    && clobber(x)
  1028    => (MOVLstoreconst [makeValAndOff(a.Val()&0xffff | c.Val()<<16, a.Off())] {s} p mem)
  1029  
  1030  (MOVWstoreconst [c] {s} p1 x:(MOVWstoreconst [a] {s} p0 mem))
  1031    && x.Uses == 1
  1032    && a.Off() == c.Off()
  1033    && sequentialAddresses(p0, p1, 2)
  1034    && clobber(x)
  1035    => (MOVLstoreconst [makeValAndOff(a.Val()&0xffff | c.Val()<<16, a.Off())] {s} p0 mem)
  1036  (MOVWstoreconst [a] {s} p0 x:(MOVWstoreconst [c] {s} p1 mem))
  1037    && x.Uses == 1
  1038    && a.Off() == c.Off()
  1039    && sequentialAddresses(p0, p1, 2)
  1040    && clobber(x)
  1041    => (MOVLstoreconst [makeValAndOff(a.Val()&0xffff | c.Val()<<16, a.Off())] {s} p0 mem)
  1042  
  1043  // Combine stores into larger (unaligned) stores.
  1044  (MOVBstore [i] {s} p (SHR(W|L)const [8] w) x:(MOVBstore [i-1] {s} p w mem))
  1045    && x.Uses == 1
  1046    && clobber(x)
  1047    => (MOVWstore [i-1] {s} p w mem)
  1048  (MOVBstore [i] {s} p w x:(MOVBstore {s} [i+1] p (SHR(W|L)const [8] w) mem))
  1049    && x.Uses == 1
  1050    && clobber(x)
  1051    => (MOVWstore [i] {s} p w mem)
  1052  (MOVBstore [i] {s} p (SHRLconst [j] w) x:(MOVBstore [i-1] {s} p w0:(SHRLconst [j-8] w) mem))
  1053    && x.Uses == 1
  1054    && clobber(x)
  1055    => (MOVWstore [i-1] {s} p w0 mem)
  1056  
  1057  (MOVBstore [i] {s} p1 (SHR(W|L)const [8] w) x:(MOVBstore [i] {s} p0 w mem))
  1058    && x.Uses == 1
  1059    && sequentialAddresses(p0, p1, 1)
  1060    && clobber(x)
  1061    => (MOVWstore [i] {s} p0 w mem)
  1062  (MOVBstore [i] {s} p0 w x:(MOVBstore {s} [i] p1 (SHR(W|L)const [8] w) mem))
  1063    && x.Uses == 1
  1064    && sequentialAddresses(p0, p1, 1)
  1065    && clobber(x)
  1066    => (MOVWstore [i] {s} p0 w mem)
  1067  (MOVBstore [i] {s} p1 (SHRLconst [j] w) x:(MOVBstore [i] {s} p0 w0:(SHRLconst [j-8] w) mem))
  1068    && x.Uses == 1
  1069    && sequentialAddresses(p0, p1, 1)
  1070    && clobber(x)
  1071    => (MOVWstore [i] {s} p0 w0 mem)
  1072  
  1073  (MOVWstore [i] {s} p (SHRLconst [16] w) x:(MOVWstore [i-2] {s} p w mem))
  1074    && x.Uses == 1
  1075    && clobber(x)
  1076    => (MOVLstore [i-2] {s} p w mem)
  1077  (MOVWstore [i] {s} p (SHRLconst [j] w) x:(MOVWstore [i-2] {s} p w0:(SHRLconst [j-16] w) mem))
  1078    && x.Uses == 1
  1079    && clobber(x)
  1080    => (MOVLstore [i-2] {s} p w0 mem)
  1081  
  1082  (MOVWstore [i] {s} p1 (SHRLconst [16] w) x:(MOVWstore [i] {s} p0 w mem))
  1083    && x.Uses == 1
  1084    && sequentialAddresses(p0, p1, 2)
  1085    && clobber(x)
  1086    => (MOVLstore [i] {s} p0 w mem)
  1087  (MOVWstore [i] {s} p1 (SHRLconst [j] w) x:(MOVWstore [i] {s} p0 w0:(SHRLconst [j-16] w) mem))
  1088    && x.Uses == 1
  1089    && sequentialAddresses(p0, p1, 2)
  1090    && clobber(x)
  1091    => (MOVLstore [i] {s} p0 w0 mem)
  1092  
  1093  // For PIC, break floating-point constant loading into two instructions so we have
  1094  // a register to use for holding the address of the constant pool entry.
  1095  (MOVSSconst [c]) && config.ctxt.Flag_shared => (MOVSSconst2 (MOVSSconst1 [c]))
  1096  (MOVSDconst [c]) && config.ctxt.Flag_shared => (MOVSDconst2 (MOVSDconst1 [c]))
  1097  
  1098  (CMP(L|W|B) l:(MOV(L|W|B)load {sym} [off] ptr mem) x) && canMergeLoad(v, l) && clobber(l) => (CMP(L|W|B)load {sym} [off] ptr x mem)
  1099  (CMP(L|W|B) x l:(MOV(L|W|B)load {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (InvertFlags (CMP(L|W|B)load {sym} [off] ptr x mem))
  1100  
  1101  (CMP(L|W|B)const l:(MOV(L|W|B)load {sym} [off] ptr mem) [c])
  1102  	&& l.Uses == 1
  1103  	&& clobber(l) =>
  1104    @l.Block (CMP(L|W|B)constload {sym} [makeValAndOff(int32(c),off)] ptr mem)
  1105  
  1106  (CMPLload {sym} [off] ptr (MOVLconst [c]) mem) => (CMPLconstload {sym} [makeValAndOff(c,off)] ptr mem)
  1107  (CMPWload {sym} [off] ptr (MOVLconst [c]) mem) => (CMPWconstload {sym} [makeValAndOff(int32(int16(c)),off)] ptr mem)
  1108  (CMPBload {sym} [off] ptr (MOVLconst [c]) mem) => (CMPBconstload {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
  1109  
  1110  (MOVBload [off] {sym} (SB) _) && symIsRO(sym) => (MOVLconst [int32(read8(sym, int64(off)))])
  1111  (MOVWload [off] {sym} (SB) _) && symIsRO(sym) => (MOVLconst [int32(read16(sym, int64(off), config.ctxt.Arch.ByteOrder))])
  1112  (MOVLload [off] {sym} (SB) _) && symIsRO(sym) => (MOVLconst [int32(read32(sym, int64(off), config.ctxt.Arch.ByteOrder))])
  1113  

View as plain text