...

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

Documentation: cmd/compile/internal/ssa/_gen

     1// Copyright 2016 The Go Authors. All rights reserved.
     2// Use of this source code is governed by a BSD-style
     3// license that can be found in the LICENSE file.
     4
     5// 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(Bswap16 x) => (ROLWconst [8] x)
    56
    57(Sqrt ...) => (SQRTSD ...)
    58(Sqrt32 ...) => (SQRTSS ...)
    59
    60(Ctz8 x) => (BSFL (ORLconst <typ.UInt32> [0x100] x))
    61(Ctz8NonZero ...) => (BSFL ...)
    62(Ctz16 x) => (BSFL (ORLconst <typ.UInt32> [0x10000] x))
    63(Ctz16NonZero ...) => (BSFL ...)
    64(Ctz32 ...) => (LoweredCtz32 ...)
    65(Ctz32NonZero ...) => (BSFL ...)
    66
    67// Lowering extension
    68(SignExt8to16  ...) => (MOVBLSX ...)
    69(SignExt8to32  ...) => (MOVBLSX ...)
    70(SignExt16to32 ...) => (MOVWLSX ...)
    71
    72(ZeroExt8to16  ...) => (MOVBLZX ...)
    73(ZeroExt8to32  ...) => (MOVBLZX ...)
    74(ZeroExt16to32 ...) => (MOVWLZX ...)
    75
    76(Signmask x) => (SARLconst x [31])
    77(Zeromask <t> x) => (XORLconst [-1] (SBBLcarrymask <t> (CMPLconst x [1])))
    78(Slicemask <t> x) => (SARLconst (NEGL <t> x) [31])
    79
    80// Lowering truncation
    81// Because we ignore high parts of registers, truncates are just copies.
    82(Trunc16to8  ...) => (Copy ...)
    83(Trunc32to8  ...) => (Copy ...)
    84(Trunc32to16 ...) => (Copy ...)
    85
    86// Lowering float-int conversions
    87(Cvt32to32F ...) => (CVTSL2SS ...)
    88(Cvt32to64F ...) => (CVTSL2SD ...)
    89
    90(Cvt32Fto32 ...) => (CVTTSS2SL ...)
    91(Cvt64Fto32 ...) => (CVTTSD2SL ...)
    92
    93(Cvt32Fto64F ...) => (CVTSS2SD ...)
    94(Cvt64Fto32F ...) => (CVTSD2SS ...)
    95
    96(Round32F ...) => (Copy ...)
    97(Round64F ...) => (Copy ...)
    98
    99(CvtBoolToUint8 ...) => (Copy ...)
   100
   101// Lowering shifts
   102// Unsigned shifts need to return 0 if shift amount is >= width of shifted value.
   103//   result = (arg << shift) & (shift >= argbits ? 0 : 0xffffffffffffffff)
   104(Lsh32x(32|16|8) <t> x y) && !shiftIsBounded(v) => (ANDL (SHLL <t> x y) (SBBLcarrymask <t> (CMP(L|W|B)const y [32])))
   105(Lsh16x(32|16|8) <t> x y) && !shiftIsBounded(v) => (ANDL (SHLL <t> x y) (SBBLcarrymask <t> (CMP(L|W|B)const y [32])))
   106(Lsh8x(32|16|8)  <t> x y) && !shiftIsBounded(v) => (ANDL (SHLL <t> x y) (SBBLcarrymask <t> (CMP(L|W|B)const y [32])))
   107
   108(Lsh32x(32|16|8) <t> x y) && shiftIsBounded(v) => (SHLL <t> x y)
   109(Lsh16x(32|16|8) <t> x y) && shiftIsBounded(v) => (SHLL <t> x y)
   110(Lsh8x(32|16|8)  <t> x y) && shiftIsBounded(v) => (SHLL <t> x y)
   111
   112(Rsh32Ux(32|16|8) <t> x y) && !shiftIsBounded(v) => (ANDL (SHRL <t> x y) (SBBLcarrymask <t> (CMP(L|W|B)const y [32])))
   113(Rsh16Ux(32|16|8) <t> x y) && !shiftIsBounded(v) => (ANDL (SHRW <t> x y) (SBBLcarrymask <t> (CMP(L|W|B)const y [16])))
   114(Rsh8Ux(32|16|8)  <t> x y) && !shiftIsBounded(v) => (ANDL (SHRB <t> x y) (SBBLcarrymask <t> (CMP(L|W|B)const y [8])))
   115
   116(Rsh32Ux(32|16|8) <t> x y) && shiftIsBounded(v) => (SHRL <t> x y)
   117(Rsh16Ux(32|16|8) <t> x y) && shiftIsBounded(v) => (SHRW <t> x y)
   118(Rsh8Ux(32|16|8)  <t> x y) && shiftIsBounded(v) => (SHRB <t> x y)
   119
   120// Signed right shift needs to return 0/-1 if shift amount is >= width of shifted value.
   121// We implement this by setting the shift value to -1 (all ones) if the shift value is >= width.
   122
   123(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])))))
   124(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])))))
   125(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])))))
   126
   127(Rsh32x(32|16|8) <t> x y) && shiftIsBounded(v) => (SARL x y)
   128(Rsh16x(32|16|8) <t> x y) && shiftIsBounded(v) => (SARW x y)
   129(Rsh8x(32|16|8) <t> x y)  && shiftIsBounded(v) => (SARB x y)
   130
   131// constant shifts
   132// generic opt rewrites all constant shifts to shift by Const64
   133(Lsh32x64 x (Const64 [c])) && uint64(c) < 32 => (SHLLconst x [int32(c)])
   134(Rsh32x64 x (Const64 [c])) && uint64(c) < 32 => (SARLconst x [int32(c)])
   135(Rsh32Ux64 x (Const64 [c])) && uint64(c) < 32 => (SHRLconst x [int32(c)])
   136(Lsh16x64 x (Const64 [c])) && uint64(c) < 16 => (SHLLconst x [int32(c)])
   137(Rsh16x64 x (Const64 [c])) && uint64(c) < 16 => (SARWconst x [int16(c)])
   138(Rsh16Ux64 x (Const64 [c])) && uint64(c) < 16 => (SHRWconst x [int16(c)])
   139(Lsh8x64 x (Const64 [c])) && uint64(c) < 8 => (SHLLconst x [int32(c)])
   140(Rsh8x64 x (Const64 [c])) && uint64(c) < 8 => (SARBconst x [int8(c)])
   141(Rsh8Ux64 x (Const64 [c])) && uint64(c) < 8 => (SHRBconst x [int8(c)])
   142
   143// large constant shifts
   144(Lsh32x64 _ (Const64 [c])) && uint64(c) >= 32 => (Const32 [0])
   145(Rsh32Ux64 _ (Const64 [c])) && uint64(c) >= 32 => (Const32 [0])
   146(Lsh16x64 _ (Const64 [c])) && uint64(c) >= 16 => (Const16 [0])
   147(Rsh16Ux64 _ (Const64 [c])) && uint64(c) >= 16 => (Const16 [0])
   148(Lsh8x64 _ (Const64 [c])) && uint64(c) >= 8 => (Const8 [0])
   149(Rsh8Ux64 _ (Const64 [c])) && uint64(c) >= 8 => (Const8 [0])
   150
   151// large constant signed right shift, we leave the sign bit
   152(Rsh32x64 x (Const64 [c])) && uint64(c) >= 32 => (SARLconst x [31])
   153(Rsh16x64 x (Const64 [c])) && uint64(c) >= 16 => (SARWconst x [15])
   154(Rsh8x64 x (Const64 [c])) && uint64(c) >= 8 => (SARBconst x [7])
   155
   156// rotates
   157(RotateLeft32 ...) => (ROLL ...)
   158(RotateLeft16 ...) => (ROLW ...)
   159(RotateLeft8  ...) => (ROLB ...)
   160// constant rotates
   161(ROLL x (MOVLconst [c])) => (ROLLconst [c&31] x)
   162(ROLW x (MOVLconst [c])) => (ROLWconst [int16(c&15)] x)
   163(ROLB x (MOVLconst [c])) => (ROLBconst [int8(c&7)] x)
   164
   165// Lowering comparisons
   166(Less32  x y) => (SETL (CMPL x y))
   167(Less16  x y) => (SETL (CMPW x y))
   168(Less8   x y) => (SETL (CMPB x y))
   169(Less32U x y) => (SETB (CMPL x y))
   170(Less16U x y) => (SETB (CMPW x y))
   171(Less8U  x y) => (SETB (CMPB x y))
   172// Use SETGF with reversed operands to dodge NaN case
   173(Less64F x y) => (SETGF (UCOMISD y x))
   174(Less32F x y) => (SETGF (UCOMISS y x))
   175
   176(Leq32  x y) => (SETLE (CMPL x y))
   177(Leq16  x y) => (SETLE (CMPW x y))
   178(Leq8   x y) => (SETLE (CMPB x y))
   179(Leq32U x y) => (SETBE (CMPL x y))
   180(Leq16U x y) => (SETBE (CMPW x y))
   181(Leq8U  x y) => (SETBE (CMPB x y))
   182// Use SETGEF with reversed operands to dodge NaN case
   183(Leq64F x y) => (SETGEF (UCOMISD y x))
   184(Leq32F x y) => (SETGEF (UCOMISS y x))
   185
   186(Eq32  x y) => (SETEQ (CMPL x y))
   187(Eq16  x y) => (SETEQ (CMPW x y))
   188(Eq8   x y) => (SETEQ (CMPB x y))
   189(EqB   x y) => (SETEQ (CMPB x y))
   190(EqPtr x y) => (SETEQ (CMPL x y))
   191(Eq64F x y) => (SETEQF (UCOMISD x y))
   192(Eq32F x y) => (SETEQF (UCOMISS x y))
   193
   194(Neq32  x y) => (SETNE (CMPL x y))
   195(Neq16  x y) => (SETNE (CMPW x y))
   196(Neq8   x y) => (SETNE (CMPB x y))
   197(NeqB   x y) => (SETNE (CMPB x y))
   198(NeqPtr x y) => (SETNE (CMPL x y))
   199(Neq64F x y) => (SETNEF (UCOMISD x y))
   200(Neq32F x y) => (SETNEF (UCOMISS x y))
   201
   202// Lowering loads
   203(Load <t> ptr mem) && (is32BitInt(t) || isPtr(t)) => (MOVLload ptr mem)
   204(Load <t> ptr mem) && is16BitInt(t) => (MOVWload ptr mem)
   205(Load <t> ptr mem) && (t.IsBoolean() || is8BitInt(t)) => (MOVBload ptr mem)
   206(Load <t> ptr mem) && is32BitFloat(t) => (MOVSSload ptr mem)
   207(Load <t> ptr mem) && is64BitFloat(t) => (MOVSDload ptr mem)
   208
   209// Lowering stores
   210(Store {t} ptr val mem) && t.Size() == 8 &&  t.IsFloat() => (MOVSDstore ptr val mem)
   211(Store {t} ptr val mem) && t.Size() == 4 &&  t.IsFloat() => (MOVSSstore ptr val mem)
   212(Store {t} ptr val mem) && t.Size() == 4 && !t.IsFloat() => (MOVLstore ptr val mem)
   213(Store {t} ptr val mem) && t.Size() == 2 => (MOVWstore ptr val mem)
   214(Store {t} ptr val mem) && t.Size() == 1 => (MOVBstore ptr val mem)
   215
   216// Lowering moves
   217(Move [0] _ _ mem) => mem
   218(Move [1] dst src mem) => (MOVBstore dst (MOVBload src mem) mem)
   219(Move [2] dst src mem) => (MOVWstore dst (MOVWload src mem) mem)
   220(Move [4] dst src mem) => (MOVLstore dst (MOVLload src mem) mem)
   221(Move [3] dst src mem) =>
   222	(MOVBstore [2] dst (MOVBload [2] src mem)
   223		(MOVWstore dst (MOVWload src mem) mem))
   224(Move [5] dst src mem) =>
   225	(MOVBstore [4] dst (MOVBload [4] src mem)
   226		(MOVLstore dst (MOVLload src mem) mem))
   227(Move [6] dst src mem) =>
   228	(MOVWstore [4] dst (MOVWload [4] src mem)
   229		(MOVLstore dst (MOVLload src mem) mem))
   230(Move [7] dst src mem) =>
   231	(MOVLstore [3] dst (MOVLload [3] src mem)
   232		(MOVLstore dst (MOVLload src mem) mem))
   233(Move [8] dst src mem) =>
   234	(MOVLstore [4] dst (MOVLload [4] src mem)
   235		(MOVLstore dst (MOVLload src mem) mem))
   236
   237// Adjust moves to be a multiple of 4 bytes.
   238(Move [s] dst src mem)
   239	&& s > 8 && s%4 != 0 =>
   240	(Move [s-s%4]
   241		(ADDLconst <dst.Type> dst [int32(s%4)])
   242		(ADDLconst <src.Type> src [int32(s%4)])
   243		(MOVLstore dst (MOVLload src mem) mem))
   244
   245// Medium copying uses a duff device.
   246(Move [s] dst src mem)
   247	&& s > 8 && s <= 4*128 && s%4 == 0
   248	&& !config.noDuffDevice && logLargeCopy(v, s) =>
   249	(DUFFCOPY [10*(128-s/4)] dst src mem)
   250// 10 and 128 are magic constants.  10 is the number of bytes to encode:
   251//	MOVL	(SI), CX
   252//	ADDL	$4, SI
   253//	MOVL	CX, (DI)
   254//	ADDL	$4, DI
   255// and 128 is the number of such blocks. See src/runtime/duff_386.s:duffcopy.
   256
   257// Large copying uses REP MOVSL.
   258(Move [s] dst src mem) && (s > 4*128 || config.noDuffDevice) && s%4 == 0 && logLargeCopy(v, s) =>
   259	(REPMOVSL dst src (MOVLconst [int32(s/4)]) mem)
   260
   261// Lowering Zero instructions
   262(Zero [0] _ mem) => mem
   263(Zero [1] destptr mem) => (MOVBstoreconst [0] destptr mem)
   264(Zero [2] destptr mem) => (MOVWstoreconst [0] destptr mem)
   265(Zero [4] destptr mem) => (MOVLstoreconst [0] destptr mem)
   266
   267(Zero [3] destptr mem) =>
   268	(MOVBstoreconst [makeValAndOff(0,2)] destptr
   269		(MOVWstoreconst [makeValAndOff(0,0)] destptr mem))
   270(Zero [5] destptr mem) =>
   271	(MOVBstoreconst [makeValAndOff(0,4)] destptr
   272		(MOVLstoreconst [makeValAndOff(0,0)] destptr mem))
   273(Zero [6] destptr mem) =>
   274	(MOVWstoreconst [makeValAndOff(0,4)] destptr
   275		(MOVLstoreconst [makeValAndOff(0,0)] destptr mem))
   276(Zero [7] destptr mem) =>
   277	(MOVLstoreconst [makeValAndOff(0,3)] destptr
   278		(MOVLstoreconst [makeValAndOff(0,0)] destptr mem))
   279
   280// Strip off any fractional word zeroing.
   281(Zero [s] destptr mem) && s%4 != 0 && s > 4 =>
   282	(Zero [s-s%4] (ADDLconst destptr [int32(s%4)])
   283		(MOVLstoreconst [0] destptr mem))
   284
   285// Zero small numbers of words directly.
   286(Zero [8] destptr mem) =>
   287	(MOVLstoreconst [makeValAndOff(0,4)] destptr
   288		(MOVLstoreconst [makeValAndOff(0,0)] destptr mem))
   289(Zero [12] destptr mem) =>
   290	(MOVLstoreconst [makeValAndOff(0,8)] destptr
   291		(MOVLstoreconst [makeValAndOff(0,4)] destptr
   292			(MOVLstoreconst [makeValAndOff(0,0)] destptr mem)))
   293(Zero [16] destptr mem) =>
   294	(MOVLstoreconst [makeValAndOff(0,12)] destptr
   295		(MOVLstoreconst [makeValAndOff(0,8)] destptr
   296			(MOVLstoreconst [makeValAndOff(0,4)] destptr
   297				(MOVLstoreconst [makeValAndOff(0,0)] destptr mem))))
   298
   299// Medium zeroing uses a duff device.
   300(Zero [s] destptr mem)
   301  && s > 16 && s <= 4*128 && s%4 == 0
   302  && !config.noDuffDevice =>
   303	(DUFFZERO [1*(128-s/4)] destptr (MOVLconst [0]) mem)
   304// 1 and 128 are magic constants.  1 is the number of bytes to encode STOSL.
   305// 128 is the number of STOSL instructions in duffzero.
   306// See src/runtime/duff_386.s:duffzero.
   307
   308// Large zeroing uses REP STOSQ.
   309(Zero [s] destptr mem)
   310  && (s > 4*128 || (config.noDuffDevice && s > 16))
   311  && s%4 == 0 =>
   312	(REPSTOSL destptr (MOVLconst [int32(s/4)]) (MOVLconst [0]) mem)
   313
   314
   315// Lowering constants
   316(Const8   [c]) => (MOVLconst [int32(c)])
   317(Const16  [c]) => (MOVLconst [int32(c)])
   318(Const32  ...) => (MOVLconst ...)
   319(Const(32|64)F ...) => (MOVS(S|D)const ...)
   320(ConstNil) => (MOVLconst [0])
   321(ConstBool [c]) => (MOVLconst [b2i32(c)])
   322
   323// Lowering calls
   324(StaticCall ...) => (CALLstatic ...)
   325(ClosureCall ...) => (CALLclosure ...)
   326(InterCall ...) => (CALLinter ...)
   327(TailCall ...) => (CALLtail ...)
   328
   329// Miscellaneous
   330(IsNonNil p) => (SETNE (TESTL p p))
   331(IsInBounds idx len) => (SETB (CMPL idx len))
   332(IsSliceInBounds idx len) => (SETBE (CMPL idx len))
   333(NilCheck ...) => (LoweredNilCheck ...)
   334(GetG ...) => (LoweredGetG ...)
   335(GetClosurePtr ...) => (LoweredGetClosurePtr ...)
   336(GetCallerPC ...) => (LoweredGetCallerPC ...)
   337(GetCallerSP ...) => (LoweredGetCallerSP ...)
   338(Addr {sym} base) => (LEAL {sym} base)
   339(LocalAddr <t> {sym} base mem) && t.Elem().HasPointers() => (LEAL {sym} (SPanchored base mem))
   340(LocalAddr <t> {sym} base _)  && !t.Elem().HasPointers() => (LEAL {sym} base)
   341
   342// block rewrites
   343(If (SETL  cmp) yes no) => (LT  cmp yes no)
   344(If (SETLE cmp) yes no) => (LE  cmp yes no)
   345(If (SETG  cmp) yes no) => (GT  cmp yes no)
   346(If (SETGE cmp) yes no) => (GE  cmp yes no)
   347(If (SETEQ cmp) yes no) => (EQ  cmp yes no)
   348(If (SETNE cmp) yes no) => (NE  cmp yes no)
   349(If (SETB  cmp) yes no) => (ULT cmp yes no)
   350(If (SETBE cmp) yes no) => (ULE cmp yes no)
   351(If (SETA  cmp) yes no) => (UGT cmp yes no)
   352(If (SETAE cmp) yes no) => (UGE cmp yes no)
   353(If (SETO  cmp) yes no) => (OS cmp yes no)
   354
   355// Special case for floating point - LF/LEF not generated
   356(If (SETGF  cmp) yes no) => (UGT  cmp yes no)
   357(If (SETGEF cmp) yes no) => (UGE  cmp yes no)
   358(If (SETEQF cmp) yes no) => (EQF  cmp yes no)
   359(If (SETNEF cmp) yes no) => (NEF  cmp yes no)
   360
   361(If cond yes no) => (NE (TESTB cond cond) yes no)
   362
   363// Write barrier.
   364(WB ...) => (LoweredWB ...)
   365
   366(PanicBounds [kind] x y mem) && boundsABI(kind) == 0 => (LoweredPanicBoundsA [kind] x y mem)
   367(PanicBounds [kind] x y mem) && boundsABI(kind) == 1 => (LoweredPanicBoundsB [kind] x y mem)
   368(PanicBounds [kind] x y mem) && boundsABI(kind) == 2 => (LoweredPanicBoundsC [kind] x y mem)
   369
   370(PanicExtend [kind] hi lo y mem) && boundsABI(kind) == 0 => (LoweredPanicExtendA [kind] hi lo y mem)
   371(PanicExtend [kind] hi lo y mem) && boundsABI(kind) == 1 => (LoweredPanicExtendB [kind] hi lo y mem)
   372(PanicExtend [kind] hi lo y mem) && boundsABI(kind) == 2 => (LoweredPanicExtendC [kind] hi lo y mem)
   373
   374// ***************************
   375// Above: lowering rules
   376// Below: optimizations
   377// ***************************
   378// TODO: Should the optimizations be a separate pass?
   379
   380// Fold boolean tests into blocks
   381(NE (TESTB (SETL  cmp) (SETL  cmp)) yes no) => (LT  cmp yes no)
   382(NE (TESTB (SETLE cmp) (SETLE cmp)) yes no) => (LE  cmp yes no)
   383(NE (TESTB (SETG  cmp) (SETG  cmp)) yes no) => (GT  cmp yes no)
   384(NE (TESTB (SETGE cmp) (SETGE cmp)) yes no) => (GE  cmp yes no)
   385(NE (TESTB (SETEQ cmp) (SETEQ cmp)) yes no) => (EQ  cmp yes no)
   386(NE (TESTB (SETNE cmp) (SETNE cmp)) yes no) => (NE  cmp yes no)
   387(NE (TESTB (SETB  cmp) (SETB  cmp)) yes no) => (ULT cmp yes no)
   388(NE (TESTB (SETBE cmp) (SETBE cmp)) yes no) => (ULE cmp yes no)
   389(NE (TESTB (SETA  cmp) (SETA  cmp)) yes no) => (UGT cmp yes no)
   390(NE (TESTB (SETAE cmp) (SETAE cmp)) yes no) => (UGE cmp yes no)
   391(NE (TESTB (SETO cmp) (SETO cmp)) yes no) => (OS cmp yes no)
   392
   393// Special case for floating point - LF/LEF not generated
   394(NE (TESTB (SETGF  cmp) (SETGF  cmp)) yes no) => (UGT  cmp yes no)
   395(NE (TESTB (SETGEF cmp) (SETGEF cmp)) yes no) => (UGE  cmp yes no)
   396(NE (TESTB (SETEQF cmp) (SETEQF cmp)) yes no) => (EQF  cmp yes no)
   397(NE (TESTB (SETNEF cmp) (SETNEF cmp)) yes no) => (NEF  cmp yes no)
   398
   399// fold constants into instructions
   400(ADDL x (MOVLconst <t> [c])) && !t.IsPtr() => (ADDLconst [c] x)
   401(ADDLcarry x (MOVLconst [c])) => (ADDLconstcarry [c] x)
   402(ADCL x (MOVLconst [c]) f) => (ADCLconst [c] x f)
   403
   404(SUBL x (MOVLconst [c])) => (SUBLconst x [c])
   405(SUBL (MOVLconst [c]) x) => (NEGL (SUBLconst <v.Type> x [c]))
   406(SUBLcarry x (MOVLconst [c])) => (SUBLconstcarry [c] x)
   407(SBBL x (MOVLconst [c]) f) => (SBBLconst [c] x f)
   408
   409(MULL x (MOVLconst [c])) => (MULLconst [c] x)
   410(ANDL x (MOVLconst [c])) => (ANDLconst [c] x)
   411
   412(ANDLconst [c] (ANDLconst [d] x)) => (ANDLconst [c & d] x)
   413(XORLconst [c] (XORLconst [d] x)) => (XORLconst [c ^ d] x)
   414(MULLconst [c] (MULLconst [d] x)) => (MULLconst [c * d] x)
   415
   416(ORL x (MOVLconst [c])) => (ORLconst [c] x)
   417(XORL x (MOVLconst [c])) => (XORLconst [c] x)
   418
   419(SHLL x (MOVLconst [c])) => (SHLLconst [c&31] x)
   420(SHRL x (MOVLconst [c])) => (SHRLconst [c&31] x)
   421(SHRW x (MOVLconst [c])) && c&31 < 16 => (SHRWconst [int16(c&31)] x)
   422(SHRW _ (MOVLconst [c])) && c&31 >= 16 => (MOVLconst [0])
   423(SHRB x (MOVLconst [c])) && c&31 < 8 => (SHRBconst [int8(c&31)] x)
   424(SHRB _ (MOVLconst [c])) && c&31 >= 8 => (MOVLconst [0])
   425
   426(SARL x (MOVLconst [c])) => (SARLconst [c&31] x)
   427(SARW x (MOVLconst [c])) => (SARWconst [int16(min(int64(c&31),15))] x)
   428(SARB x (MOVLconst [c])) => (SARBconst [int8(min(int64(c&31),7))] x)
   429
   430(SARL x (ANDLconst [31] y)) => (SARL x y)
   431(SHLL x (ANDLconst [31] y)) => (SHLL x y)
   432(SHRL x (ANDLconst [31] y)) => (SHRL x y)
   433
   434// Constant shift simplifications
   435
   436(SHLLconst x [0]) => x
   437(SHRLconst x [0]) => x
   438(SARLconst x [0]) => x
   439
   440(SHRWconst x [0]) => x
   441(SARWconst x [0]) => x
   442
   443(SHRBconst x [0]) => x
   444(SARBconst x [0]) => x
   445
   446(ROLLconst [0] x) => x
   447(ROLWconst [0] x) => x
   448(ROLBconst [0] x) => x
   449
   450// Note: the word and byte shifts keep the low 5 bits (not the low 4 or 3 bits)
   451// because the x86 instructions are defined to use all 5 bits of the shift even
   452// for the small shifts. I don't think we'll ever generate a weird shift (e.g.
   453// (SHRW x (MOVLconst [24])), but just in case.
   454
   455(CMPL x (MOVLconst [c])) => (CMPLconst x [c])
   456(CMPL (MOVLconst [c]) x) => (InvertFlags (CMPLconst x [c]))
   457(CMPW x (MOVLconst [c])) => (CMPWconst x [int16(c)])
   458(CMPW (MOVLconst [c]) x) => (InvertFlags (CMPWconst x [int16(c)]))
   459(CMPB x (MOVLconst [c])) => (CMPBconst x [int8(c)])
   460(CMPB (MOVLconst [c]) x) => (InvertFlags (CMPBconst x [int8(c)]))
   461
   462// Canonicalize the order of arguments to comparisons - helps with CSE.
   463(CMP(L|W|B) x y) && canonLessThan(x,y) => (InvertFlags (CMP(L|W|B) y x))
   464
   465// strength reduction
   466// Assumes that the following costs from https://gmplib.org/~tege/x86-timing.pdf:
   467//    1 - addl, shll, leal, negl, subl
   468//    3 - imull
   469// This limits the rewrites to two instructions.
   470// Note that negl always operates in-place,
   471// which can require a register-register move
   472// to preserve the original value,
   473// so it must be used with care.
   474(MULLconst [-9] x) => (NEGL (LEAL8 <v.Type> x x))
   475(MULLconst [-5] x) => (NEGL (LEAL4 <v.Type> x x))
   476(MULLconst [-3] x) => (NEGL (LEAL2 <v.Type> x x))
   477(MULLconst [-1] x) => (NEGL x)
   478(MULLconst [0] _) => (MOVLconst [0])
   479(MULLconst [1] x) => x
   480(MULLconst [3] x) => (LEAL2 x x)
   481(MULLconst [5] x) => (LEAL4 x x)
   482(MULLconst [7] x) => (LEAL2 x (LEAL2 <v.Type> x x))
   483(MULLconst [9] x) => (LEAL8 x x)
   484(MULLconst [11] x) => (LEAL2 x (LEAL4 <v.Type> x x))
   485(MULLconst [13] x) => (LEAL4 x (LEAL2 <v.Type> x x))
   486(MULLconst [19] x) => (LEAL2 x (LEAL8 <v.Type> x x))
   487(MULLconst [21] x) => (LEAL4 x (LEAL4 <v.Type> x x))
   488(MULLconst [25] x) => (LEAL8 x (LEAL2 <v.Type> x x))
   489(MULLconst [27] x) => (LEAL8 (LEAL2 <v.Type> x x) (LEAL2 <v.Type> x x))
   490(MULLconst [37] x) => (LEAL4 x (LEAL8 <v.Type> x x))
   491(MULLconst [41] x) => (LEAL8 x (LEAL4 <v.Type> x x))
   492(MULLconst [45] x) => (LEAL8 (LEAL4 <v.Type> x x) (LEAL4 <v.Type> x x))
   493(MULLconst [73] x) => (LEAL8 x (LEAL8 <v.Type> x x))
   494(MULLconst [81] x) => (LEAL8 (LEAL8 <v.Type> x x) (LEAL8 <v.Type> x x))
   495
   496(MULLconst [c] x) && isPowerOfTwo32(c+1) && c >= 15 => (SUBL (SHLLconst <v.Type> [int32(log32(c+1))] x) x)
   497(MULLconst [c] x) && isPowerOfTwo32(c-1) && c >= 17 => (LEAL1 (SHLLconst <v.Type> [int32(log32(c-1))] x) x)
   498(MULLconst [c] x) && isPowerOfTwo32(c-2) && c >= 34 => (LEAL2 (SHLLconst <v.Type> [int32(log32(c-2))] x) x)
   499(MULLconst [c] x) && isPowerOfTwo32(c-4) && c >= 68 => (LEAL4 (SHLLconst <v.Type> [int32(log32(c-4))] x) x)
   500(MULLconst [c] x) && isPowerOfTwo32(c-8) && c >= 136 => (LEAL8 (SHLLconst <v.Type> [int32(log32(c-8))] x) x)
   501(MULLconst [c] x) && c%3 == 0 && isPowerOfTwo32(c/3) => (SHLLconst [int32(log32(c/3))] (LEAL2 <v.Type> x x))
   502(MULLconst [c] x) && c%5 == 0 && isPowerOfTwo32(c/5) => (SHLLconst [int32(log32(c/5))] (LEAL4 <v.Type> x x))
   503(MULLconst [c] x) && c%9 == 0 && isPowerOfTwo32(c/9) => (SHLLconst [int32(log32(c/9))] (LEAL8 <v.Type> x x))
   504
   505// combine add/shift into LEAL
   506(ADDL x (SHLLconst [3] y)) => (LEAL8 x y)
   507(ADDL x (SHLLconst [2] y)) => (LEAL4 x y)
   508(ADDL x (SHLLconst [1] y)) => (LEAL2 x y)
   509(ADDL x (ADDL y y)) => (LEAL2 x y)
   510(ADDL x (ADDL x y)) => (LEAL2 y x)
   511
   512// combine ADDL/ADDLconst into LEAL1
   513(ADDLconst [c] (ADDL x y)) => (LEAL1 [c] x y)
   514(ADDL (ADDLconst [c] x) y) => (LEAL1 [c] x y)
   515
   516// fold ADDL into LEAL
   517(ADDLconst [c] (LEAL [d] {s} x)) && is32Bit(int64(c)+int64(d)) => (LEAL [c+d] {s} x)
   518(LEAL [c] {s} (ADDLconst [d] x)) && is32Bit(int64(c)+int64(d)) => (LEAL [c+d] {s} x)
   519(ADDLconst [c] x:(SP)) => (LEAL [c] x) // so it is rematerializeable
   520(LEAL [c] {s} (ADDL x y)) && x.Op != OpSB && y.Op != OpSB => (LEAL1 [c] {s} x y)
   521(ADDL x (LEAL [c] {s} y)) && x.Op != OpSB && y.Op != OpSB => (LEAL1 [c] {s} x y)
   522
   523// fold ADDLconst into LEALx
   524(ADDLconst [c] (LEAL1 [d] {s} x y)) && is32Bit(int64(c)+int64(d)) => (LEAL1 [c+d] {s} x y)
   525(ADDLconst [c] (LEAL2 [d] {s} x y)) && is32Bit(int64(c)+int64(d)) => (LEAL2 [c+d] {s} x y)
   526(ADDLconst [c] (LEAL4 [d] {s} x y)) && is32Bit(int64(c)+int64(d)) => (LEAL4 [c+d] {s} x y)
   527(ADDLconst [c] (LEAL8 [d] {s} x y)) && is32Bit(int64(c)+int64(d)) => (LEAL8 [c+d] {s} x y)
   528(LEAL1 [c] {s} (ADDLconst [d] x) y) && is32Bit(int64(c)+int64(d))   && x.Op != OpSB => (LEAL1 [c+d] {s} x y)
   529(LEAL2 [c] {s} (ADDLconst [d] x) y) && is32Bit(int64(c)+int64(d))   && x.Op != OpSB => (LEAL2 [c+d] {s} x y)
   530(LEAL2 [c] {s} x (ADDLconst [d] y)) && is32Bit(int64(c)+2*int64(d)) && y.Op != OpSB => (LEAL2 [c+2*d] {s} x y)
   531(LEAL4 [c] {s} (ADDLconst [d] x) y) && is32Bit(int64(c)+int64(d))   && x.Op != OpSB => (LEAL4 [c+d] {s} x y)
   532(LEAL4 [c] {s} x (ADDLconst [d] y)) && is32Bit(int64(c)+4*int64(d)) && y.Op != OpSB => (LEAL4 [c+4*d] {s} x y)
   533(LEAL8 [c] {s} (ADDLconst [d] x) y) && is32Bit(int64(c)+int64(d))   && x.Op != OpSB => (LEAL8 [c+d] {s} x y)
   534(LEAL8 [c] {s} x (ADDLconst [d] y)) && is32Bit(int64(c)+8*int64(d)) && y.Op != OpSB => (LEAL8 [c+8*d] {s} x y)
   535
   536// fold shifts into LEALx
   537(LEAL1 [c] {s} x (SHLLconst [1] y)) => (LEAL2 [c] {s} x y)
   538(LEAL1 [c] {s} x (SHLLconst [2] y)) => (LEAL4 [c] {s} x y)
   539(LEAL1 [c] {s} x (SHLLconst [3] y)) => (LEAL8 [c] {s} x y)
   540(LEAL2 [c] {s} x (SHLLconst [1] y)) => (LEAL4 [c] {s} x y)
   541(LEAL2 [c] {s} x (SHLLconst [2] y)) => (LEAL8 [c] {s} x y)
   542(LEAL4 [c] {s} x (SHLLconst [1] y)) => (LEAL8 [c] {s} x y)
   543
   544// reverse ordering of compare instruction
   545(SETL (InvertFlags x)) => (SETG x)
   546(SETG (InvertFlags x)) => (SETL x)
   547(SETB (InvertFlags x)) => (SETA x)
   548(SETA (InvertFlags x)) => (SETB x)
   549(SETLE (InvertFlags x)) => (SETGE x)
   550(SETGE (InvertFlags x)) => (SETLE x)
   551(SETBE (InvertFlags x)) => (SETAE x)
   552(SETAE (InvertFlags x)) => (SETBE x)
   553(SETEQ (InvertFlags x)) => (SETEQ x)
   554(SETNE (InvertFlags x)) => (SETNE x)
   555
   556// sign extended loads
   557// Note: The combined instruction must end up in the same block
   558// as the original load. If not, we end up making a value with
   559// memory type live in two different blocks, which can lead to
   560// multiple memory values alive simultaneously.
   561// Make sure we don't combine these ops if the load has another use.
   562// This prevents a single load from being split into multiple loads
   563// which then might return different values.  See test/atomicload.go.
   564(MOVBLSX x:(MOVBload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVBLSXload <v.Type> [off] {sym} ptr mem)
   565(MOVBLZX x:(MOVBload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVBload <v.Type> [off] {sym} ptr mem)
   566(MOVWLSX x:(MOVWload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVWLSXload <v.Type> [off] {sym} ptr mem)
   567(MOVWLZX x:(MOVWload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVWload <v.Type> [off] {sym} ptr mem)
   568
   569// replace load from same location as preceding store with zero/sign extension (or copy in case of full width)
   570(MOVBload [off] {sym} ptr (MOVBstore [off2] {sym2} ptr2 x _)) && sym == sym2 && off == off2 && isSamePtr(ptr, ptr2) => (MOVBLZX x)
   571(MOVWload [off] {sym} ptr (MOVWstore [off2] {sym2} ptr2 x _)) && sym == sym2 && off == off2 && isSamePtr(ptr, ptr2) => (MOVWLZX x)
   572(MOVLload [off] {sym} ptr (MOVLstore [off2] {sym2} ptr2 x _)) && sym == sym2 && off == off2 && isSamePtr(ptr, ptr2) => x
   573(MOVBLSXload [off] {sym} ptr (MOVBstore [off2] {sym2} ptr2 x _)) && sym == sym2 && off == off2 && isSamePtr(ptr, ptr2) => (MOVBLSX x)
   574(MOVWLSXload [off] {sym} ptr (MOVWstore [off2] {sym2} ptr2 x _)) && sym == sym2 && off == off2 && isSamePtr(ptr, ptr2) => (MOVWLSX x)
   575
   576// Fold extensions and ANDs together.
   577(MOVBLZX (ANDLconst [c] x)) => (ANDLconst [c & 0xff] x)
   578(MOVWLZX (ANDLconst [c] x)) => (ANDLconst [c & 0xffff] x)
   579(MOVBLSX (ANDLconst [c] x)) && c & 0x80 == 0 => (ANDLconst [c & 0x7f] x)
   580(MOVWLSX (ANDLconst [c] x)) && c & 0x8000 == 0 => (ANDLconst [c & 0x7fff] x)
   581
   582// Don't extend before storing
   583(MOVWstore [off] {sym} ptr (MOVWL(S|Z)X x) mem) => (MOVWstore [off] {sym} ptr x mem)
   584(MOVBstore [off] {sym} ptr (MOVBL(S|Z)X x) mem) => (MOVBstore [off] {sym} ptr x mem)
   585
   586// fold constants into memory operations
   587// Note that this is not always a good idea because if not all the uses of
   588// the ADDLconst get eliminated, we still have to compute the ADDLconst and we now
   589// have potentially two live values (ptr and (ADDLconst [off] ptr)) instead of one.
   590// Nevertheless, let's do it!
   591(MOV(L|W|B|SS|SD)load  [off1] {sym} (ADDLconst [off2] ptr) mem) && is32Bit(int64(off1)+int64(off2)) =>
   592    (MOV(L|W|B|SS|SD)load  [off1+off2] {sym} ptr mem)
   593(MOV(L|W|B|SS|SD)store  [off1] {sym} (ADDLconst [off2] ptr) val mem) && is32Bit(int64(off1)+int64(off2)) =>
   594    (MOV(L|W|B|SS|SD)store  [off1+off2] {sym} ptr val mem)
   595
   596((ADD|SUB|MUL|AND|OR|XOR)Lload [off1] {sym} val (ADDLconst [off2] base) mem) && is32Bit(int64(off1)+int64(off2)) =>
   597	((ADD|SUB|MUL|AND|OR|XOR)Lload [off1+off2] {sym} val base mem)
   598((ADD|SUB|MUL|DIV)SSload [off1] {sym} val (ADDLconst [off2] base) mem) && is32Bit(int64(off1)+int64(off2)) =>
   599	((ADD|SUB|MUL|DIV)SSload [off1+off2] {sym} val base mem)
   600((ADD|SUB|MUL|DIV)SDload [off1] {sym} val (ADDLconst [off2] base) mem) && is32Bit(int64(off1)+int64(off2)) =>
   601	((ADD|SUB|MUL|DIV)SDload [off1+off2] {sym} val base mem)
   602((ADD|SUB|AND|OR|XOR)Lmodify [off1] {sym} (ADDLconst [off2] base) val mem) && is32Bit(int64(off1)+int64(off2)) =>
   603	((ADD|SUB|AND|OR|XOR)Lmodify [off1+off2] {sym} base val mem)
   604((ADD|AND|OR|XOR)Lconstmodify [valoff1] {sym} (ADDLconst [off2] base) mem) && valoff1.canAdd32(off2) =>
   605	((ADD|AND|OR|XOR)Lconstmodify [valoff1.addOffset32(off2)] {sym} base mem)
   606
   607// Fold constants into stores.
   608(MOVLstore [off] {sym} ptr (MOVLconst [c]) mem) =>
   609	(MOVLstoreconst [makeValAndOff(c,off)] {sym} ptr mem)
   610(MOVWstore [off] {sym} ptr (MOVLconst [c]) mem) =>
   611	(MOVWstoreconst [makeValAndOff(c,off)] {sym} ptr mem)
   612(MOVBstore [off] {sym} ptr (MOVLconst [c]) mem) =>
   613	(MOVBstoreconst [makeValAndOff(c,off)] {sym} ptr mem)
   614
   615// Fold address offsets into constant stores.
   616(MOV(L|W|B)storeconst [sc] {s} (ADDLconst [off] ptr) mem) && sc.canAdd32(off) =>
   617	(MOV(L|W|B)storeconst [sc.addOffset32(off)] {s} ptr mem)
   618
   619// We need to fold LEAL into the MOVx ops so that the live variable analysis knows
   620// what variables are being read/written by the ops.
   621// Note: we turn off this merging for operations on globals when building
   622// position-independent code (when Flag_shared is set).
   623// PIC needs a spare register to load the PC into.  Having the LEAL be
   624// a separate instruction gives us that register.  Having the LEAL be
   625// a separate instruction also allows it to be CSEd (which is good because
   626// it compiles to a thunk call).
   627(MOV(L|W|B|SS|SD|BLSX|WLSX)load  [off1] {sym1} (LEAL [off2] {sym2} base) mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2)
   628  && (base.Op != OpSB || !config.ctxt.Flag_shared) =>
   629        (MOV(L|W|B|SS|SD|BLSX|WLSX)load  [off1+off2] {mergeSym(sym1,sym2)} base mem)
   630
   631(MOV(L|W|B|SS|SD)store  [off1] {sym1} (LEAL [off2] {sym2} base) val mem) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2)
   632  && (base.Op != OpSB || !config.ctxt.Flag_shared) =>
   633	(MOV(L|W|B|SS|SD)store  [off1+off2] {mergeSym(sym1,sym2)} base val mem)
   634
   635(MOV(L|W|B)storeconst [sc] {sym1} (LEAL [off] {sym2} ptr) mem) && canMergeSym(sym1, sym2) && sc.canAdd32(off)
   636  && (ptr.Op != OpSB || !config.ctxt.Flag_shared) =>
   637	(MOV(L|W|B)storeconst [sc.addOffset32(off)] {mergeSym(sym1, sym2)} ptr mem)
   638
   639((ADD|SUB|MUL|AND|OR|XOR)Lload [off1] {sym1} val (LEAL [off2] {sym2} base) mem)
   640	&& is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) && (base.Op != OpSB || !config.ctxt.Flag_shared) =>
   641	((ADD|SUB|MUL|AND|OR|XOR)Lload [off1+off2] {mergeSym(sym1,sym2)} val base mem)
   642((ADD|SUB|MUL|DIV)SSload [off1] {sym1} val (LEAL [off2] {sym2} base) mem)
   643	&& is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) && (base.Op != OpSB || !config.ctxt.Flag_shared) =>
   644	((ADD|SUB|MUL|DIV)SSload [off1+off2] {mergeSym(sym1,sym2)} val base mem)
   645((ADD|SUB|MUL|DIV)SDload [off1] {sym1} val (LEAL [off2] {sym2} base) mem)
   646	&& is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) && (base.Op != OpSB || !config.ctxt.Flag_shared) =>
   647	((ADD|SUB|MUL|DIV)SDload [off1+off2] {mergeSym(sym1,sym2)} val base mem)
   648((ADD|SUB|AND|OR|XOR)Lmodify [off1] {sym1} (LEAL [off2] {sym2} base) val mem)
   649	&& is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) && (base.Op != OpSB || !config.ctxt.Flag_shared) =>
   650	((ADD|SUB|AND|OR|XOR)Lmodify [off1+off2] {mergeSym(sym1,sym2)} base val mem)
   651((ADD|AND|OR|XOR)Lconstmodify [valoff1] {sym1} (LEAL [off2] {sym2} base) mem)
   652	&& valoff1.canAdd32(off2) && canMergeSym(sym1, sym2) && (base.Op != OpSB || !config.ctxt.Flag_shared) =>
   653	((ADD|AND|OR|XOR)Lconstmodify [valoff1.addOffset32(off2)] {mergeSym(sym1,sym2)} base mem)
   654
   655// Merge load/store to op
   656((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)
   657((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)
   658((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)
   659(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)
   660(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) =>
   661	((ADD|SUB|AND|OR|XOR)Lmodify [off] {sym} ptr x mem)
   662(MOVLstore {sym} [off] ptr y:((ADD|AND|OR|XOR)Lconst [c] l:(MOVLload [off] {sym} ptr mem)) mem)
   663	&& y.Uses==1 && l.Uses==1 && clobber(y, l) =>
   664	((ADD|AND|OR|XOR)Lconstmodify [makeValAndOff(c,off)] {sym} ptr mem)
   665
   666// fold LEALs together
   667(LEAL [off1] {sym1} (LEAL [off2] {sym2} x)) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   668      (LEAL [off1+off2] {mergeSym(sym1,sym2)} x)
   669
   670// LEAL into LEAL1
   671(LEAL1 [off1] {sym1} (LEAL [off2] {sym2} x) y) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) && x.Op != OpSB =>
   672       (LEAL1 [off1+off2] {mergeSym(sym1,sym2)} x y)
   673
   674// LEAL1 into LEAL
   675(LEAL [off1] {sym1} (LEAL1 [off2] {sym2} x y)) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   676       (LEAL1 [off1+off2] {mergeSym(sym1,sym2)} x y)
   677
   678// LEAL into LEAL[248]
   679(LEAL2 [off1] {sym1} (LEAL [off2] {sym2} x) y) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) && x.Op != OpSB =>
   680       (LEAL2 [off1+off2] {mergeSym(sym1,sym2)} x y)
   681(LEAL4 [off1] {sym1} (LEAL [off2] {sym2} x) y) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) && x.Op != OpSB =>
   682       (LEAL4 [off1+off2] {mergeSym(sym1,sym2)} x y)
   683(LEAL8 [off1] {sym1} (LEAL [off2] {sym2} x) y) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) && x.Op != OpSB =>
   684       (LEAL8 [off1+off2] {mergeSym(sym1,sym2)} x y)
   685
   686// LEAL[248] into LEAL
   687(LEAL [off1] {sym1} (LEAL2 [off2] {sym2} x y)) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   688      (LEAL2 [off1+off2] {mergeSym(sym1,sym2)} x y)
   689(LEAL [off1] {sym1} (LEAL4 [off2] {sym2} x y)) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   690      (LEAL4 [off1+off2] {mergeSym(sym1,sym2)} x y)
   691(LEAL [off1] {sym1} (LEAL8 [off2] {sym2} x y)) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   692      (LEAL8 [off1+off2] {mergeSym(sym1,sym2)} x y)
   693
   694// LEAL[1248] into LEAL[1248]. Only some such merges are possible.
   695(LEAL1 [off1] {sym1} x (LEAL1 [off2] {sym2} y y)) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   696      (LEAL2 [off1+off2] {mergeSym(sym1, sym2)} x y)
   697(LEAL1 [off1] {sym1} x (LEAL1 [off2] {sym2} x y)) && is32Bit(int64(off1)+int64(off2)) && canMergeSym(sym1, sym2) =>
   698      (LEAL2 [off1+off2] {mergeSym(sym1, sym2)} y x)
   699(LEAL2 [off1] {sym} x (LEAL1 [off2] {nil} y y)) && is32Bit(int64(off1)+2*int64(off2)) =>
   700      (LEAL4 [off1+2*off2] {sym} x y)
   701(LEAL4 [off1] {sym} x (LEAL1 [off2] {nil} y y)) && is32Bit(int64(off1)+4*int64(off2)) =>
   702      (LEAL8 [off1+4*off2] {sym} x y)
   703
   704// Absorb InvertFlags into branches.
   705(LT (InvertFlags cmp) yes no) => (GT cmp yes no)
   706(GT (InvertFlags cmp) yes no) => (LT cmp yes no)
   707(LE (InvertFlags cmp) yes no) => (GE cmp yes no)
   708(GE (InvertFlags cmp) yes no) => (LE cmp yes no)
   709(ULT (InvertFlags cmp) yes no) => (UGT cmp yes no)
   710(UGT (InvertFlags cmp) yes no) => (ULT cmp yes no)
   711(ULE (InvertFlags cmp) yes no) => (UGE cmp yes no)
   712(UGE (InvertFlags cmp) yes no) => (ULE cmp yes no)
   713(EQ (InvertFlags cmp) yes no) => (EQ cmp yes no)
   714(NE (InvertFlags cmp) yes no) => (NE cmp yes no)
   715
   716// Constant comparisons.
   717(CMPLconst (MOVLconst [x]) [y]) && x==y                       => (FlagEQ)
   718(CMPLconst (MOVLconst [x]) [y]) && x<y && uint32(x)<uint32(y) => (FlagLT_ULT)
   719(CMPLconst (MOVLconst [x]) [y]) && x<y && uint32(x)>uint32(y) => (FlagLT_UGT)
   720(CMPLconst (MOVLconst [x]) [y]) && x>y && uint32(x)<uint32(y) => (FlagGT_ULT)
   721(CMPLconst (MOVLconst [x]) [y]) && x>y && uint32(x)>uint32(y) => (FlagGT_UGT)
   722
   723(CMPWconst (MOVLconst [x]) [y]) && int16(x)==y                       => (FlagEQ)
   724(CMPWconst (MOVLconst [x]) [y]) && int16(x)<y && uint16(x)<uint16(y) => (FlagLT_ULT)
   725(CMPWconst (MOVLconst [x]) [y]) && int16(x)<y && uint16(x)>uint16(y) => (FlagLT_UGT)
   726(CMPWconst (MOVLconst [x]) [y]) && int16(x)>y && uint16(x)<uint16(y) => (FlagGT_ULT)
   727(CMPWconst (MOVLconst [x]) [y]) && int16(x)>y && uint16(x)>uint16(y) => (FlagGT_UGT)
   728
   729(CMPBconst (MOVLconst [x]) [y]) && int8(x)==y                      => (FlagEQ)
   730(CMPBconst (MOVLconst [x]) [y]) && int8(x)<y && uint8(x)<uint8(y) => (FlagLT_ULT)
   731(CMPBconst (MOVLconst [x]) [y]) && int8(x)<y && uint8(x)>uint8(y) => (FlagLT_UGT)
   732(CMPBconst (MOVLconst [x]) [y]) && int8(x)>y && uint8(x)<uint8(y) => (FlagGT_ULT)
   733(CMPBconst (MOVLconst [x]) [y]) && int8(x)>y && uint8(x)>uint8(y) => (FlagGT_UGT)
   734
   735// Other known comparisons.
   736(CMPLconst (SHRLconst _ [c]) [n]) && 0 <= n && 0 < c && c <= 32 && (1<<uint64(32-c)) <= uint64(n) => (FlagLT_ULT)
   737(CMPLconst (ANDLconst _ [m]) [n]) && 0 <= m && m < n => (FlagLT_ULT)
   738(CMPWconst (ANDLconst _ [m]) [n]) && 0 <= int16(m) && int16(m) < n => (FlagLT_ULT)
   739(CMPBconst (ANDLconst _ [m]) [n]) && 0 <= int8(m) && int8(m) < n => (FlagLT_ULT)
   740// TODO: DIVxU also.
   741
   742// Absorb flag constants into SBB ops.
   743(SBBLcarrymask (FlagEQ)) => (MOVLconst [0])
   744(SBBLcarrymask (FlagLT_ULT)) => (MOVLconst [-1])
   745(SBBLcarrymask (FlagLT_UGT)) => (MOVLconst [0])
   746(SBBLcarrymask (FlagGT_ULT)) => (MOVLconst [-1])
   747(SBBLcarrymask (FlagGT_UGT)) => (MOVLconst [0])
   748
   749// Absorb flag constants into branches.
   750(EQ (FlagEQ) yes no) => (First yes no)
   751(EQ (FlagLT_ULT) yes no) => (First no yes)
   752(EQ (FlagLT_UGT) yes no) => (First no yes)
   753(EQ (FlagGT_ULT) yes no) => (First no yes)
   754(EQ (FlagGT_UGT) yes no) => (First no yes)
   755
   756(NE (FlagEQ) yes no) => (First no yes)
   757(NE (FlagLT_ULT) yes no) => (First yes no)
   758(NE (FlagLT_UGT) yes no) => (First yes no)
   759(NE (FlagGT_ULT) yes no) => (First yes no)
   760(NE (FlagGT_UGT) yes no) => (First yes no)
   761
   762(LT (FlagEQ) yes no) => (First no yes)
   763(LT (FlagLT_ULT) yes no) => (First yes no)
   764(LT (FlagLT_UGT) yes no) => (First yes no)
   765(LT (FlagGT_ULT) yes no) => (First no yes)
   766(LT (FlagGT_UGT) yes no) => (First no yes)
   767
   768(LE (FlagEQ) yes no) => (First yes no)
   769(LE (FlagLT_ULT) yes no) => (First yes no)
   770(LE (FlagLT_UGT) yes no) => (First yes no)
   771(LE (FlagGT_ULT) yes no) => (First no yes)
   772(LE (FlagGT_UGT) yes no) => (First no yes)
   773
   774(GT (FlagEQ) yes no) => (First no yes)
   775(GT (FlagLT_ULT) yes no) => (First no yes)
   776(GT (FlagLT_UGT) yes no) => (First no yes)
   777(GT (FlagGT_ULT) yes no) => (First yes no)
   778(GT (FlagGT_UGT) yes no) => (First yes no)
   779
   780(GE (FlagEQ) yes no) => (First yes no)
   781(GE (FlagLT_ULT) yes no) => (First no yes)
   782(GE (FlagLT_UGT) yes no) => (First no yes)
   783(GE (FlagGT_ULT) yes no) => (First yes no)
   784(GE (FlagGT_UGT) yes no) => (First yes no)
   785
   786(ULT (FlagEQ) yes no) => (First no yes)
   787(ULT (FlagLT_ULT) yes no) => (First yes no)
   788(ULT (FlagLT_UGT) yes no) => (First no yes)
   789(ULT (FlagGT_ULT) yes no) => (First yes no)
   790(ULT (FlagGT_UGT) yes no) => (First no yes)
   791
   792(ULE (FlagEQ) yes no) => (First yes no)
   793(ULE (FlagLT_ULT) yes no) => (First yes no)
   794(ULE (FlagLT_UGT) yes no) => (First no yes)
   795(ULE (FlagGT_ULT) yes no) => (First yes no)
   796(ULE (FlagGT_UGT) yes no) => (First no yes)
   797
   798(UGT (FlagEQ) yes no) => (First no yes)
   799(UGT (FlagLT_ULT) yes no) => (First no yes)
   800(UGT (FlagLT_UGT) yes no) => (First yes no)
   801(UGT (FlagGT_ULT) yes no) => (First no yes)
   802(UGT (FlagGT_UGT) yes no) => (First yes no)
   803
   804(UGE (FlagEQ) yes no) => (First yes no)
   805(UGE (FlagLT_ULT) yes no) => (First no yes)
   806(UGE (FlagLT_UGT) yes no) => (First yes no)
   807(UGE (FlagGT_ULT) yes no) => (First no yes)
   808(UGE (FlagGT_UGT) yes no) => (First yes no)
   809
   810// Absorb flag constants into SETxx ops.
   811(SETEQ (FlagEQ)) => (MOVLconst [1])
   812(SETEQ (FlagLT_ULT)) => (MOVLconst [0])
   813(SETEQ (FlagLT_UGT)) => (MOVLconst [0])
   814(SETEQ (FlagGT_ULT)) => (MOVLconst [0])
   815(SETEQ (FlagGT_UGT)) => (MOVLconst [0])
   816
   817(SETNE (FlagEQ)) => (MOVLconst [0])
   818(SETNE (FlagLT_ULT)) => (MOVLconst [1])
   819(SETNE (FlagLT_UGT)) => (MOVLconst [1])
   820(SETNE (FlagGT_ULT)) => (MOVLconst [1])
   821(SETNE (FlagGT_UGT)) => (MOVLconst [1])
   822
   823(SETL (FlagEQ)) => (MOVLconst [0])
   824(SETL (FlagLT_ULT)) => (MOVLconst [1])
   825(SETL (FlagLT_UGT)) => (MOVLconst [1])
   826(SETL (FlagGT_ULT)) => (MOVLconst [0])
   827(SETL (FlagGT_UGT)) => (MOVLconst [0])
   828
   829(SETLE (FlagEQ)) => (MOVLconst [1])
   830(SETLE (FlagLT_ULT)) => (MOVLconst [1])
   831(SETLE (FlagLT_UGT)) => (MOVLconst [1])
   832(SETLE (FlagGT_ULT)) => (MOVLconst [0])
   833(SETLE (FlagGT_UGT)) => (MOVLconst [0])
   834
   835(SETG (FlagEQ)) => (MOVLconst [0])
   836(SETG (FlagLT_ULT)) => (MOVLconst [0])
   837(SETG (FlagLT_UGT)) => (MOVLconst [0])
   838(SETG (FlagGT_ULT)) => (MOVLconst [1])
   839(SETG (FlagGT_UGT)) => (MOVLconst [1])
   840
   841(SETGE (FlagEQ)) => (MOVLconst [1])
   842(SETGE (FlagLT_ULT)) => (MOVLconst [0])
   843(SETGE (FlagLT_UGT)) => (MOVLconst [0])
   844(SETGE (FlagGT_ULT)) => (MOVLconst [1])
   845(SETGE (FlagGT_UGT)) => (MOVLconst [1])
   846
   847(SETB (FlagEQ)) => (MOVLconst [0])
   848(SETB (FlagLT_ULT)) => (MOVLconst [1])
   849(SETB (FlagLT_UGT)) => (MOVLconst [0])
   850(SETB (FlagGT_ULT)) => (MOVLconst [1])
   851(SETB (FlagGT_UGT)) => (MOVLconst [0])
   852
   853(SETBE (FlagEQ)) => (MOVLconst [1])
   854(SETBE (FlagLT_ULT)) => (MOVLconst [1])
   855(SETBE (FlagLT_UGT)) => (MOVLconst [0])
   856(SETBE (FlagGT_ULT)) => (MOVLconst [1])
   857(SETBE (FlagGT_UGT)) => (MOVLconst [0])
   858
   859(SETA (FlagEQ)) => (MOVLconst [0])
   860(SETA (FlagLT_ULT)) => (MOVLconst [0])
   861(SETA (FlagLT_UGT)) => (MOVLconst [1])
   862(SETA (FlagGT_ULT)) => (MOVLconst [0])
   863(SETA (FlagGT_UGT)) => (MOVLconst [1])
   864
   865(SETAE (FlagEQ)) => (MOVLconst [1])
   866(SETAE (FlagLT_ULT)) => (MOVLconst [0])
   867(SETAE (FlagLT_UGT)) => (MOVLconst [1])
   868(SETAE (FlagGT_ULT)) => (MOVLconst [0])
   869(SETAE (FlagGT_UGT)) => (MOVLconst [1])
   870
   871// Remove redundant *const ops
   872(ADDLconst [c] x) && c==0  => x
   873(SUBLconst [c] x) && c==0  => x
   874(ANDLconst [c] _) && c==0  => (MOVLconst [0])
   875(ANDLconst [c] x) && c==-1 => x
   876(ORLconst [c] x)  && c==0  => x
   877(ORLconst [c] _)  && c==-1 => (MOVLconst [-1])
   878(XORLconst [c] x) && c==0  => x
   879// TODO: since we got rid of the W/B versions, we might miss
   880// things like (ANDLconst [0x100] x) which were formerly
   881// (ANDBconst [0] x).  Probably doesn't happen very often.
   882// If we cared, we might do:
   883//  (ANDLconst <t> [c] x) && t.Size()==1 && int8(x)==0 => (MOVLconst [0])
   884
   885// Convert constant subtracts to constant adds
   886(SUBLconst [c] x) => (ADDLconst [-c] x)
   887
   888// generic constant folding
   889// TODO: more of this
   890(ADDLconst [c] (MOVLconst [d])) => (MOVLconst [c+d])
   891(ADDLconst [c] (ADDLconst [d] x)) => (ADDLconst [c+d] x)
   892(SARLconst [c] (MOVLconst [d])) => (MOVLconst [d>>uint64(c)])
   893(SARWconst [c] (MOVLconst [d])) => (MOVLconst [d>>uint64(c)])
   894(SARBconst [c] (MOVLconst [d])) => (MOVLconst [d>>uint64(c)])
   895(NEGL (MOVLconst [c])) => (MOVLconst [-c])
   896(MULLconst [c] (MOVLconst [d])) => (MOVLconst [c*d])
   897(ANDLconst [c] (MOVLconst [d])) => (MOVLconst [c&d])
   898(ORLconst [c] (MOVLconst [d])) => (MOVLconst [c|d])
   899(XORLconst [c] (MOVLconst [d])) => (MOVLconst [c^d])
   900(NOTL (MOVLconst [c])) => (MOVLconst [^c])
   901
   902// generic simplifications
   903// TODO: more of this
   904(ADDL x (NEGL y)) => (SUBL x y)
   905(SUBL x x) => (MOVLconst [0])
   906(ANDL x x) => x
   907(ORL x x) => x
   908(XORL x x) => (MOVLconst [0])
   909
   910// checking AND against 0.
   911(CMP(L|W|B)const l:(ANDL x y) [0]) && l.Uses==1 => (TEST(L|W|B) x y)
   912(CMPLconst l:(ANDLconst [c] x) [0]) && l.Uses==1 => (TESTLconst [c] x)
   913(CMPWconst l:(ANDLconst [c] x) [0]) && l.Uses==1 => (TESTWconst [int16(c)] x)
   914(CMPBconst l:(ANDLconst [c] x) [0]) && l.Uses==1 => (TESTBconst [int8(c)] x)
   915
   916// TEST %reg,%reg is shorter than CMP
   917(CMP(L|W|B)const x [0]) => (TEST(L|W|B) x x)
   918
   919// Convert LEAL1 back to ADDL if we can
   920(LEAL1 [0] {nil} x y) => (ADDL x y)
   921
   922// For PIC, break floating-point constant loading into two instructions so we have
   923// a register to use for holding the address of the constant pool entry.
   924(MOVSSconst [c]) && config.ctxt.Flag_shared => (MOVSSconst2 (MOVSSconst1 [c]))
   925(MOVSDconst [c]) && config.ctxt.Flag_shared => (MOVSDconst2 (MOVSDconst1 [c]))
   926
   927(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)
   928(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))
   929
   930(CMP(L|W|B)const l:(MOV(L|W|B)load {sym} [off] ptr mem) [c])
   931	&& l.Uses == 1
   932	&& clobber(l) =>
   933  @l.Block (CMP(L|W|B)constload {sym} [makeValAndOff(int32(c),off)] ptr mem)
   934
   935(CMPLload {sym} [off] ptr (MOVLconst [c]) mem) => (CMPLconstload {sym} [makeValAndOff(c,off)] ptr mem)
   936(CMPWload {sym} [off] ptr (MOVLconst [c]) mem) => (CMPWconstload {sym} [makeValAndOff(int32(int16(c)),off)] ptr mem)
   937(CMPBload {sym} [off] ptr (MOVLconst [c]) mem) => (CMPBconstload {sym} [makeValAndOff(int32(int8(c)),off)] ptr mem)
   938
   939(MOVBload [off] {sym} (SB) _) && symIsRO(sym) => (MOVLconst [int32(read8(sym, int64(off)))])
   940(MOVWload [off] {sym} (SB) _) && symIsRO(sym) => (MOVLconst [int32(read16(sym, int64(off), config.ctxt.Arch.ByteOrder))])
   941(MOVLload [off] {sym} (SB) _) && symIsRO(sym) => (MOVLconst [int32(read32(sym, int64(off), config.ctxt.Arch.ByteOrder))])

View as plain text