// Copyright 2016 The Go Authors. All rights reserved. // Use of this source code is governed by a BSD-style // license that can be found in the LICENSE file. // GOPPC64 values indicate power8, power9, etc. // That means the code is compiled for that target, // and will not run on earlier targets. // (Add(Ptr|64|32|16|8) ...) => (ADD ...) (Add64F ...) => (FADD ...) (Add32F ...) => (FADDS ...) (Sub(Ptr|64|32|16|8) ...) => (SUB ...) (Sub32F ...) => (FSUBS ...) (Sub64F ...) => (FSUB ...) // Combine 64 bit integer multiply and adds (ADD l:(MULLD x y) z) && buildcfg.GOPPC64 >= 9 && l.Uses == 1 && clobber(l) => (MADDLD x y z) (Mod16 x y) => (Mod32 (SignExt16to32 x) (SignExt16to32 y)) (Mod16u x y) => (Mod32u (ZeroExt16to32 x) (ZeroExt16to32 y)) (Mod8 x y) => (Mod32 (SignExt8to32 x) (SignExt8to32 y)) (Mod8u x y) => (Mod32u (ZeroExt8to32 x) (ZeroExt8to32 y)) (Mod64 x y) && buildcfg.GOPPC64 >=9 => (MODSD x y) (Mod64 x y) && buildcfg.GOPPC64 <=8 => (SUB x (MULLD y (DIVD x y))) (Mod64u x y) && buildcfg.GOPPC64 >= 9 => (MODUD x y) (Mod64u x y) && buildcfg.GOPPC64 <= 8 => (SUB x (MULLD y (DIVDU x y))) (Mod32 x y) && buildcfg.GOPPC64 >= 9 => (MODSW x y) (Mod32 x y) && buildcfg.GOPPC64 <= 8 => (SUB x (MULLW y (DIVW x y))) (Mod32u x y) && buildcfg.GOPPC64 >= 9 => (MODUW x y) (Mod32u x y) && buildcfg.GOPPC64 <= 8 => (SUB x (MULLW y (DIVWU x y))) // (x + y) / 2 with x>=y => (x - y) / 2 + y (Avg64u x y) => (ADD (SRDconst (SUB x y) [1]) y) (Mul64 ...) => (MULLD ...) (Mul(32|16|8) ...) => (MULLW ...) (Select0 (Mul64uhilo x y)) => (MULHDU x y) (Select1 (Mul64uhilo x y)) => (MULLD x y) (Div64 [false] x y) => (DIVD x y) (Div64u ...) => (DIVDU ...) (Div32 [false] x y) => (DIVW x y) (Div32u ...) => (DIVWU ...) (Div16 [false] x y) => (DIVW (SignExt16to32 x) (SignExt16to32 y)) (Div16u x y) => (DIVWU (ZeroExt16to32 x) (ZeroExt16to32 y)) (Div8 x y) => (DIVW (SignExt8to32 x) (SignExt8to32 y)) (Div8u x y) => (DIVWU (ZeroExt8to32 x) (ZeroExt8to32 y)) (Hmul(64|64u|32|32u) ...) => (MULH(D|DU|W|WU) ...) (Mul(32|64)F ...) => ((FMULS|FMUL) ...) (Div(32|64)F ...) => ((FDIVS|FDIV) ...) // Lowering float <=> int (Cvt32to(32|64)F x) => ((FCFIDS|FCFID) (MTVSRD (SignExt32to64 x))) (Cvt64to(32|64)F x) => ((FCFIDS|FCFID) (MTVSRD x)) (Cvt32Fto(32|64) x) => (MFVSRD (FCTI(W|D)Z x)) (Cvt64Fto(32|64) x) => (MFVSRD (FCTI(W|D)Z x)) (Cvt32Fto64F ...) => (Copy ...) // Note v will have the wrong type for patterns dependent on Float32/Float64 (Cvt64Fto32F ...) => (FRSP ...) (CvtBoolToUint8 ...) => (Copy ...) (Round(32|64)F ...) => (LoweredRound(32|64)F ...) (Sqrt ...) => (FSQRT ...) (Sqrt32 ...) => (FSQRTS ...) (Floor ...) => (FFLOOR ...) (Ceil ...) => (FCEIL ...) (Trunc ...) => (FTRUNC ...) (Round ...) => (FROUND ...) (Copysign x y) => (FCPSGN y x) (Abs ...) => (FABS ...) (FMA ...) => (FMADD ...) // Lowering extension // Note: we always extend to 64 bits even though some ops don't need that many result bits. (SignExt8to(16|32|64) ...) => (MOVBreg ...) (SignExt16to(32|64) ...) => (MOVHreg ...) (SignExt32to64 ...) => (MOVWreg ...) (ZeroExt8to(16|32|64) ...) => (MOVBZreg ...) (ZeroExt16to(32|64) ...) => (MOVHZreg ...) (ZeroExt32to64 ...) => (MOVWZreg ...) (Trunc(16|32|64)to8 x) && t.IsSigned() => (MOVBreg x) (Trunc(16|32|64)to8 x) => (MOVBZreg x) (Trunc(32|64)to16 x) && t.IsSigned() => (MOVHreg x) (Trunc(32|64)to16 x) => (MOVHZreg x) (Trunc64to32 x) && t.IsSigned() => (MOVWreg x) (Trunc64to32 x) => (MOVWZreg x) // Lowering constants (Const(64|32|16|8) [val]) => (MOVDconst [int64(val)]) (Const(32|64)F ...) => (FMOV(S|D)const ...) (ConstNil) => (MOVDconst [0]) (ConstBool [t]) => (MOVDconst [b2i(t)]) // Carrying addition. (Select0 (Add64carry x y c)) => (Select0 (ADDE x y (Select1 (ADDCconst c [-1])))) (Select1 (Add64carry x y c)) => (ADDZEzero (Select1 (ADDE x y (Select1 (ADDCconst c [-1]))))) // Fold initial carry bit if 0. (ADDE x y (Select1 (ADDCconst (MOVDconst [0]) [-1]))) => (ADDC x y) // Fold transfer of CA -> GPR -> CA. Note 2 uses when feeding into a chained Add64carry. (Select1 (ADDCconst n:(ADDZEzero x) [-1])) && n.Uses <= 2 => x // Borrowing subtraction. (Select0 (Sub64borrow x y c)) => (Select0 (SUBE x y (Select1 (SUBCconst c [0])))) (Select1 (Sub64borrow x y c)) => (NEG (SUBZEzero (Select1 (SUBE x y (Select1 (SUBCconst c [0])))))) // Fold initial borrow bit if 0. (SUBE x y (Select1 (SUBCconst (MOVDconst [0]) [0]))) => (SUBC x y) // Fold transfer of CA -> GPR -> CA. Note 2 uses when feeding into a chained Sub64borrow. (Select1 (SUBCconst n:(NEG (SUBZEzero x)) [0])) && n.Uses <= 2 => x // Constant folding (FABS (FMOVDconst [x])) => (FMOVDconst [math.Abs(x)]) (FSQRT (FMOVDconst [x])) && x >= 0 => (FMOVDconst [math.Sqrt(x)]) (FFLOOR (FMOVDconst [x])) => (FMOVDconst [math.Floor(x)]) (FCEIL (FMOVDconst [x])) => (FMOVDconst [math.Ceil(x)]) (FTRUNC (FMOVDconst [x])) => (FMOVDconst [math.Trunc(x)]) // Rotates (RotateLeft8 x (MOVDconst [c])) => (Or8 (Lsh8x64 x (MOVDconst [c&7])) (Rsh8Ux64 x (MOVDconst [-c&7]))) (RotateLeft16 x (MOVDconst [c])) => (Or16 (Lsh16x64 x (MOVDconst [c&15])) (Rsh16Ux64 x (MOVDconst [-c&15]))) (RotateLeft(32|64) ...) => ((ROTLW|ROTL) ...) // Constant rotate generation (ROTLW x (MOVDconst [c])) => (ROTLWconst x [c&31]) (ROTL x (MOVDconst [c])) => (ROTLconst x [c&63]) // Combine rotate and mask operations (Select0 (ANDCCconst [m] (ROTLWconst [r] x))) && isPPC64WordRotateMask(m) => (RLWINM [encodePPC64RotateMask(r,m,32)] x) (AND (MOVDconst [m]) (ROTLWconst [r] x)) && isPPC64WordRotateMask(m) => (RLWINM [encodePPC64RotateMask(r,m,32)] x) (Select0 (ANDCCconst [m] (ROTLW x r))) && isPPC64WordRotateMask(m) => (RLWNM [encodePPC64RotateMask(0,m,32)] x r) (AND (MOVDconst [m]) (ROTLW x r)) && isPPC64WordRotateMask(m) => (RLWNM [encodePPC64RotateMask(0,m,32)] x r) // Note, any rotated word bitmask is still a valid word bitmask. (ROTLWconst [r] (AND (MOVDconst [m]) x)) && isPPC64WordRotateMask(m) => (RLWINM [encodePPC64RotateMask(r,rotateLeft32(m,r),32)] x) (ROTLWconst [r] (Select0 (ANDCCconst [m] x))) && isPPC64WordRotateMask(m) => (RLWINM [encodePPC64RotateMask(r,rotateLeft32(m,r),32)] x) (Select0 (ANDCCconst [m] (SRWconst x [s]))) && mergePPC64RShiftMask(m,s,32) == 0 => (MOVDconst [0]) (Select0 (ANDCCconst [m] (SRWconst x [s]))) && mergePPC64AndSrwi(m,s) != 0 => (RLWINM [mergePPC64AndSrwi(m,s)] x) (AND (MOVDconst [m]) (SRWconst x [s])) && mergePPC64RShiftMask(m,s,32) == 0 => (MOVDconst [0]) (AND (MOVDconst [m]) (SRWconst x [s])) && mergePPC64AndSrwi(m,s) != 0 => (RLWINM [mergePPC64AndSrwi(m,s)] x) (SRWconst (Select0 (ANDCCconst [m] x)) [s]) && mergePPC64RShiftMask(m>>uint(s),s,32) == 0 => (MOVDconst [0]) (SRWconst (Select0 (ANDCCconst [m] x)) [s]) && mergePPC64AndSrwi(m>>uint(s),s) != 0 => (RLWINM [mergePPC64AndSrwi(m>>uint(s),s)] x) (SRWconst (AND (MOVDconst [m]) x) [s]) && mergePPC64RShiftMask(m>>uint(s),s,32) == 0 => (MOVDconst [0]) (SRWconst (AND (MOVDconst [m]) x) [s]) && mergePPC64AndSrwi(m>>uint(s),s) != 0 => (RLWINM [mergePPC64AndSrwi(m>>uint(s),s)] x) // Merge shift right + shift left and clear left (e.g for a table lookup) (CLRLSLDI [c] (SRWconst [s] x)) && mergePPC64ClrlsldiSrw(int64(c),s) != 0 => (RLWINM [mergePPC64ClrlsldiSrw(int64(c),s)] x) (SLDconst [l] (SRWconst [r] x)) && mergePPC64SldiSrw(l,r) != 0 => (RLWINM [mergePPC64SldiSrw(l,r)] x) // The following reduction shows up frequently too. e.g b[(x>>14)&0xFF] (CLRLSLDI [c] i:(RLWINM [s] x)) && mergePPC64ClrlsldiRlwinm(c,s) != 0 => (RLWINM [mergePPC64ClrlsldiRlwinm(c,s)] x) // large constant signed right shift, we leave the sign bit (Rsh64x64 x (MOVDconst [c])) && uint64(c) >= 64 => (SRADconst x [63]) (Rsh32x64 x (MOVDconst [c])) && uint64(c) >= 32 => (SRAWconst x [63]) (Rsh16x64 x (MOVDconst [c])) && uint64(c) >= 16 => (SRAWconst (SignExt16to32 x) [63]) (Rsh8x64 x (MOVDconst [c])) && uint64(c) >= 8 => (SRAWconst (SignExt8to32 x) [63]) // constant shifts ((Lsh64|Rsh64|Rsh64U)x64 x (MOVDconst [c])) && uint64(c) < 64 => (S(L|RA|R)Dconst x [c]) ((Lsh32|Rsh32|Rsh32U)x64 x (MOVDconst [c])) && uint64(c) < 32 => (S(L|RA|R)Wconst x [c]) ((Rsh16|Rsh16U)x64 x (MOVDconst [c])) && uint64(c) < 16 => (SR(AW|W)const ((Sign|Zero)Ext16to32 x) [c]) (Lsh16x64 x (MOVDconst [c])) && uint64(c) < 16 => (SLWconst x [c]) ((Rsh8|Rsh8U)x64 x (MOVDconst [c])) && uint64(c) < 8 => (SR(AW|W)const ((Sign|Zero)Ext8to32 x) [c]) (Lsh8x64 x (MOVDconst [c])) && uint64(c) < 8 => (SLWconst x [c]) // Lower bounded shifts first. No need to check shift value. (Lsh64x(64|32|16|8) x y) && shiftIsBounded(v) => (SLD x y) (Lsh32x(64|32|16|8) x y) && shiftIsBounded(v) => (SLW x y) (Lsh16x(64|32|16|8) x y) && shiftIsBounded(v) => (SLD x y) (Lsh8x(64|32|16|8) x y) && shiftIsBounded(v) => (SLD x y) (Rsh64Ux(64|32|16|8) x y) && shiftIsBounded(v) => (SRD x y) (Rsh32Ux(64|32|16|8) x y) && shiftIsBounded(v) => (SRW x y) (Rsh16Ux(64|32|16|8) x y) && shiftIsBounded(v) => (SRD (MOVHZreg x) y) (Rsh8Ux(64|32|16|8) x y) && shiftIsBounded(v) => (SRD (MOVBZreg x) y) (Rsh64x(64|32|16|8) x y) && shiftIsBounded(v) => (SRAD x y) (Rsh32x(64|32|16|8) x y) && shiftIsBounded(v) => (SRAW x y) (Rsh16x(64|32|16|8) x y) && shiftIsBounded(v) => (SRAD (MOVHreg x) y) (Rsh8x(64|32|16|8) x y) && shiftIsBounded(v) => (SRAD (MOVBreg x) y) // Unbounded shifts. Go shifts saturate to 0 or -1 when shifting beyond the number of // bits in a type, PPC64 shifts do not (see the ISA for details). // // Note, y is always non-negative. // // Note, ISELZ is intentionally not used in lower. Where possible, ISEL is converted to ISELZ in late lower // after all the ISEL folding rules have been exercised. ((Rsh64U|Lsh64)x64 x y) => (ISEL [0] (S(R|L)D x y) (MOVDconst [0]) (CMPUconst y [64])) ((Rsh64U|Lsh64)x32 x y) => (ISEL [0] (S(R|L)D x y) (MOVDconst [0]) (CMPWUconst y [64])) ((Rsh64U|Lsh64)x16 x y) => (ISEL [2] (S(R|L)D x y) (MOVDconst [0]) (Select1 (ANDCCconst [0xFFC0] y))) ((Rsh64U|Lsh64)x8 x y) => (ISEL [2] (S(R|L)D x y) (MOVDconst [0]) (Select1 (ANDCCconst [0x00C0] y))) (Rsh64x(64|32) x y) => (ISEL [0] (SRAD x y) (SRADconst x [63]) (CMP(U|WU)const y [64])) (Rsh64x16 x y) => (ISEL [2] (SRAD x y) (SRADconst x [63]) (Select1 (ANDCCconst [0xFFC0] y))) (Rsh64x8 x y) => (ISEL [2] (SRAD x y) (SRADconst x [63]) (Select1 (ANDCCconst [0x00C0] y))) ((Rsh32U|Lsh32)x64 x y) => (ISEL [0] (S(R|L)W x y) (MOVDconst [0]) (CMPUconst y [32])) ((Rsh32U|Lsh32)x32 x y) => (ISEL [0] (S(R|L)W x y) (MOVDconst [0]) (CMPWUconst y [32])) ((Rsh32U|Lsh32)x16 x y) => (ISEL [2] (S(R|L)W x y) (MOVDconst [0]) (Select1 (ANDCCconst [0xFFE0] y))) ((Rsh32U|Lsh32)x8 x y) => (ISEL [2] (S(R|L)W x y) (MOVDconst [0]) (Select1 (ANDCCconst [0x00E0] y))) (Rsh32x(64|32) x y) => (ISEL [0] (SRAW x y) (SRAWconst x [31]) (CMP(U|WU)const y [32])) (Rsh32x16 x y) => (ISEL [2] (SRAW x y) (SRAWconst x [31]) (Select1 (ANDCCconst [0xFFE0] y))) (Rsh32x8 x y) => (ISEL [2] (SRAW x y) (SRAWconst x [31]) (Select1 (ANDCCconst [0x00E0] y))) ((Rsh16U|Lsh16)x64 x y) => (ISEL [0] (S(R|L)D (MOVHZreg x) y) (MOVDconst [0]) (CMPUconst y [16])) ((Rsh16U|Lsh16)x32 x y) => (ISEL [0] (S(R|L)D (MOVHZreg x) y) (MOVDconst [0]) (CMPWUconst y [16])) ((Rsh16U|Lsh16)x16 x y) => (ISEL [2] (S(R|L)D (MOVHZreg x) y) (MOVDconst [0]) (Select1 (ANDCCconst [0xFFF0] y))) ((Rsh16U|Lsh16)x8 x y) => (ISEL [2] (S(R|L)D (MOVHZreg x) y) (MOVDconst [0]) (Select1 (ANDCCconst [0x00F0] y))) (Rsh16x(64|32) x y) => (ISEL [0] (SRAD (MOVHreg x) y) (SRADconst (MOVHreg x) [15]) (CMP(U|WU)const y [16])) (Rsh16x16 x y) => (ISEL [2] (SRAD (MOVHreg x) y) (SRADconst (MOVHreg x) [15]) (Select1 (ANDCCconst [0xFFF0] y))) (Rsh16x8 x y) => (ISEL [2] (SRAD (MOVHreg x) y) (SRADconst (MOVHreg x) [15]) (Select1 (ANDCCconst [0x00F0] y))) ((Rsh8U|Lsh8)x64 x y) => (ISEL [0] (S(R|L)D (MOVBZreg x) y) (MOVDconst [0]) (CMPUconst y [8])) ((Rsh8U|Lsh8)x32 x y) => (ISEL [0] (S(R|L)D (MOVBZreg x) y) (MOVDconst [0]) (CMPWUconst y [8])) ((Rsh8U|Lsh8)x16 x y) => (ISEL [2] (S(R|L)D (MOVBZreg x) y) (MOVDconst [0]) (Select1 (ANDCCconst [0xFFF8] y))) ((Rsh8U|Lsh8)x8 x y) => (ISEL [2] (S(R|L)D (MOVBZreg x) y) (MOVDconst [0]) (Select1 (ANDCCconst [0x00F8] y))) (Rsh8x(64|32) x y) => (ISEL [0] (SRAD (MOVBreg x) y) (SRADconst (MOVBreg x) [7]) (CMP(U|WU)const y [8])) (Rsh8x16 x y) => (ISEL [2] (SRAD (MOVBreg x) y) (SRADconst (MOVBreg x) [7]) (Select1 (ANDCCconst [0xFFF8] y))) (Rsh8x8 x y) => (ISEL [2] (SRAD (MOVBreg x) y) (SRADconst (MOVBreg x) [7]) (Select1 (ANDCCconst [0x00F8] y))) // Catch bounded shifts in situations like foo< uint64(c) => (FlagLT) (ORN x (MOVDconst [-1])) => x (S(RAD|RD|LD) x (MOVDconst [c])) => (S(RAD|RD|LD)const [c&63 | (c>>6&1*63)] x) (S(RAW|RW|LW) x (MOVDconst [c])) => (S(RAW|RW|LW)const [c&31 | (c>>5&1*31)] x) (Addr {sym} base) => (MOVDaddr {sym} [0] base) (LocalAddr {sym} base mem) && t.Elem().HasPointers() => (MOVDaddr {sym} (SPanchored base mem)) (LocalAddr {sym} base _) && !t.Elem().HasPointers() => (MOVDaddr {sym} base) (OffPtr [off] ptr) => (ADD (MOVDconst [off]) ptr) (MOVDaddr {sym} [n] p:(ADD x y)) && sym == nil && n == 0 => p (MOVDaddr {sym} [n] ptr) && sym == nil && n == 0 && (ptr.Op == OpArgIntReg || ptr.Op == OpPhi) => ptr // TODO: optimize these cases? (Ctz32NonZero ...) => (Ctz32 ...) (Ctz64NonZero ...) => (Ctz64 ...) (Ctz64 x) && buildcfg.GOPPC64<=8 => (POPCNTD (ANDN (ADDconst [-1] x) x)) (Ctz64 x) => (CNTTZD x) (Ctz32 x) && buildcfg.GOPPC64<=8 => (POPCNTW (MOVWZreg (ANDN (ADDconst [-1] x) x))) (Ctz32 x) => (CNTTZW (MOVWZreg x)) (Ctz16 x) => (POPCNTW (MOVHZreg (ANDN (ADDconst [-1] x) x))) (Ctz8 x) => (POPCNTB (MOVBZreg (ANDN (ADDconst [-1] x) x))) (BitLen64 x) => (SUBFCconst [64] (CNTLZD x)) (BitLen32 x) => (SUBFCconst [32] (CNTLZW x)) (PopCount64 ...) => (POPCNTD ...) (PopCount(32|16|8) x) => (POPCNT(W|W|B) (MOV(W|H|B)Zreg x)) (And(64|32|16|8) ...) => (AND ...) (Or(64|32|16|8) ...) => (OR ...) (Xor(64|32|16|8) ...) => (XOR ...) (Neg(64|32|16|8) ...) => (NEG ...) (Neg(64|32)F ...) => (FNEG ...) (Com(64|32|16|8) x) => (NOR x x) // Lowering boolean ops (AndB ...) => (AND ...) (OrB ...) => (OR ...) (Not x) => (XORconst [1] x) // Merge logical operations (AND x (NOR y y)) => (ANDN x y) (OR x (NOR y y)) => (ORN x y) // Lowering comparisons (EqB x y) => (Select0 (ANDCCconst [1] (EQV x y))) // Sign extension dependence on operand sign sets up for sign/zero-extension elision later (Eq(8|16) x y) && x.Type.IsSigned() && y.Type.IsSigned() => (Equal (CMPW (SignExt(8|16)to32 x) (SignExt(8|16)to32 y))) (Eq(8|16) x y) => (Equal (CMPW (ZeroExt(8|16)to32 x) (ZeroExt(8|16)to32 y))) (Eq(32|64|Ptr) x y) => (Equal ((CMPW|CMP|CMP) x y)) (Eq(32|64)F x y) => (Equal (FCMPU x y)) (NeqB ...) => (XOR ...) // Like Eq8 and Eq16, prefer sign extension likely to enable later elision. (Neq(8|16) x y) && x.Type.IsSigned() && y.Type.IsSigned() => (NotEqual (CMPW (SignExt(8|16)to32 x) (SignExt(8|16)to32 y))) (Neq(8|16) x y) => (NotEqual (CMPW (ZeroExt(8|16)to32 x) (ZeroExt(8|16)to32 y))) (Neq(32|64|Ptr) x y) => (NotEqual ((CMPW|CMP|CMP) x y)) (Neq(32|64)F x y) => (NotEqual (FCMPU x y)) (Less(8|16) x y) => (LessThan (CMPW (SignExt(8|16)to32 x) (SignExt(8|16)to32 y))) (Less(32|64) x y) => (LessThan ((CMPW|CMP) x y)) (Less(32|64)F x y) => (FLessThan (FCMPU x y)) (Less(8|16)U x y) => (LessThan (CMPWU (ZeroExt(8|16)to32 x) (ZeroExt(8|16)to32 y))) (Less(32|64)U x y) => (LessThan ((CMPWU|CMPU) x y)) (Leq(8|16) x y) => (LessEqual (CMPW (SignExt(8|16)to32 x) (SignExt(8|16)to32 y))) (Leq(32|64) x y) => (LessEqual ((CMPW|CMP) x y)) (Leq(32|64)F x y) => (FLessEqual (FCMPU x y)) (Leq(8|16)U x y) => (LessEqual (CMPWU (ZeroExt(8|16)to32 x) (ZeroExt(8|16)to32 y))) (Leq(32|64)U x y) => (LessEqual (CMP(WU|U) x y)) // Absorb pseudo-ops into blocks. (If (Equal cc) yes no) => (EQ cc yes no) (If (NotEqual cc) yes no) => (NE cc yes no) (If (LessThan cc) yes no) => (LT cc yes no) (If (LessEqual cc) yes no) => (LE cc yes no) (If (GreaterThan cc) yes no) => (GT cc yes no) (If (GreaterEqual cc) yes no) => (GE cc yes no) (If (FLessThan cc) yes no) => (FLT cc yes no) (If (FLessEqual cc) yes no) => (FLE cc yes no) (If (FGreaterThan cc) yes no) => (FGT cc yes no) (If (FGreaterEqual cc) yes no) => (FGE cc yes no) (If cond yes no) => (NE (CMPWconst [0] (Select0 (ANDCCconst [1] cond))) yes no) // Absorb boolean tests into block (NE (CMPWconst [0] (Select0 (ANDCCconst [1] ((Equal|NotEqual|LessThan|LessEqual|GreaterThan|GreaterEqual) cc)))) yes no) => ((EQ|NE|LT|LE|GT|GE) cc yes no) (NE (CMPWconst [0] (Select0 (ANDCCconst [1] ((FLessThan|FLessEqual|FGreaterThan|FGreaterEqual) cc)))) yes no) => ((FLT|FLE|FGT|FGE) cc yes no) // absorb flag constants into branches (EQ (FlagEQ) yes no) => (First yes no) (EQ (FlagLT) yes no) => (First no yes) (EQ (FlagGT) yes no) => (First no yes) (NE (FlagEQ) yes no) => (First no yes) (NE (FlagLT) yes no) => (First yes no) (NE (FlagGT) yes no) => (First yes no) (LT (FlagEQ) yes no) => (First no yes) (LT (FlagLT) yes no) => (First yes no) (LT (FlagGT) yes no) => (First no yes) (LE (FlagEQ) yes no) => (First yes no) (LE (FlagLT) yes no) => (First yes no) (LE (FlagGT) yes no) => (First no yes) (GT (FlagEQ) yes no) => (First no yes) (GT (FlagLT) yes no) => (First no yes) (GT (FlagGT) yes no) => (First yes no) (GE (FlagEQ) yes no) => (First yes no) (GE (FlagLT) yes no) => (First no yes) (GE (FlagGT) yes no) => (First yes no) // absorb InvertFlags into branches (LT (InvertFlags cmp) yes no) => (GT cmp yes no) (GT (InvertFlags cmp) yes no) => (LT cmp yes no) (LE (InvertFlags cmp) yes no) => (GE cmp yes no) (GE (InvertFlags cmp) yes no) => (LE cmp yes no) (EQ (InvertFlags cmp) yes no) => (EQ cmp yes no) (NE (InvertFlags cmp) yes no) => (NE cmp yes no) // constant comparisons (CMPWconst (MOVDconst [x]) [y]) && int32(x)==int32(y) => (FlagEQ) (CMPWconst (MOVDconst [x]) [y]) && int32(x) (FlagLT) (CMPWconst (MOVDconst [x]) [y]) && int32(x)>int32(y) => (FlagGT) (CMPconst (MOVDconst [x]) [y]) && x==y => (FlagEQ) (CMPconst (MOVDconst [x]) [y]) && x (FlagLT) (CMPconst (MOVDconst [x]) [y]) && x>y => (FlagGT) (CMPWUconst (MOVDconst [x]) [y]) && int32(x)==int32(y) => (FlagEQ) (CMPWUconst (MOVDconst [x]) [y]) && uint32(x) (FlagLT) (CMPWUconst (MOVDconst [x]) [y]) && uint32(x)>uint32(y) => (FlagGT) (CMPUconst (MOVDconst [x]) [y]) && x==y => (FlagEQ) (CMPUconst (MOVDconst [x]) [y]) && uint64(x) (FlagLT) (CMPUconst (MOVDconst [x]) [y]) && uint64(x)>uint64(y) => (FlagGT) // absorb flag constants into boolean values (Equal (FlagEQ)) => (MOVDconst [1]) (Equal (FlagLT)) => (MOVDconst [0]) (Equal (FlagGT)) => (MOVDconst [0]) (NotEqual (FlagEQ)) => (MOVDconst [0]) (NotEqual (FlagLT)) => (MOVDconst [1]) (NotEqual (FlagGT)) => (MOVDconst [1]) (LessThan (FlagEQ)) => (MOVDconst [0]) (LessThan (FlagLT)) => (MOVDconst [1]) (LessThan (FlagGT)) => (MOVDconst [0]) (LessEqual (FlagEQ)) => (MOVDconst [1]) (LessEqual (FlagLT)) => (MOVDconst [1]) (LessEqual (FlagGT)) => (MOVDconst [0]) (GreaterThan (FlagEQ)) => (MOVDconst [0]) (GreaterThan (FlagLT)) => (MOVDconst [0]) (GreaterThan (FlagGT)) => (MOVDconst [1]) (GreaterEqual (FlagEQ)) => (MOVDconst [1]) (GreaterEqual (FlagLT)) => (MOVDconst [0]) (GreaterEqual (FlagGT)) => (MOVDconst [1]) // absorb InvertFlags into boolean values ((Equal|NotEqual|LessThan|GreaterThan|LessEqual|GreaterEqual) (InvertFlags x)) => ((Equal|NotEqual|GreaterThan|LessThan|GreaterEqual|LessEqual) x) // Elide compares of bit tests ((EQ|NE|LT|LE|GT|GE) (CMPconst [0] (Select0 z:(ANDCCconst [c] x))) yes no) => ((EQ|NE|LT|LE|GT|GE) (Select1 z) yes no) ((EQ|NE|LT|LE|GT|GE) (CMPWconst [0] (Select0 z:(ANDCCconst [c] x))) yes no) => ((EQ|NE|LT|LE|GT|GE) (Select1 z) yes no) ((EQ|NE|LT|LE|GT|GE) (CMPconst [0] z:(AND x y)) yes no) && z.Uses == 1 => ((EQ|NE|LT|LE|GT|GE) (Select1 (ANDCC x y)) yes no) ((EQ|NE|LT|LE|GT|GE) (CMPconst [0] z:(OR x y)) yes no) && z.Uses == 1 => ((EQ|NE|LT|LE|GT|GE) (Select1 (ORCC x y)) yes no) ((EQ|NE|LT|LE|GT|GE) (CMPconst [0] z:(XOR x y)) yes no) && z.Uses == 1 => ((EQ|NE|LT|LE|GT|GE) (Select1 (XORCC x y)) yes no) (CondSelect x y (SETBC [a] cmp)) => (ISEL [a] x y cmp) (CondSelect x y (SETBCR [a] cmp)) => (ISEL [a+4] x y cmp) // Only lower after bool is lowered. It should always lower. This helps ensure the folding below happens reliably. (CondSelect x y bool) && flagArg(bool) == nil => (ISEL [6] x y (Select1 (ANDCCconst [1] bool))) // Fold any CR -> GPR -> CR transfers when applying the above rule. (ISEL [6] x y (Select1 (ANDCCconst [1] (SETBC [c] cmp)))) => (ISEL [c] x y cmp) (ISEL [6] x y ((CMP|CMPW)const [0] (SETBC [c] cmp))) => (ISEL [c] x y cmp) (ISEL [6] x y ((CMP|CMPW)const [0] (SETBCR [c] cmp))) => (ISEL [c+4] x y cmp) // Lowering loads (Load ptr mem) && (is64BitInt(t) || isPtr(t)) => (MOVDload ptr mem) (Load ptr mem) && is32BitInt(t) && t.IsSigned() => (MOVWload ptr mem) (Load ptr mem) && is32BitInt(t) && !t.IsSigned() => (MOVWZload ptr mem) (Load ptr mem) && is16BitInt(t) && t.IsSigned() => (MOVHload ptr mem) (Load ptr mem) && is16BitInt(t) && !t.IsSigned() => (MOVHZload ptr mem) (Load ptr mem) && t.IsBoolean() => (MOVBZload ptr mem) (Load ptr mem) && is8BitInt(t) && t.IsSigned() => (MOVBreg (MOVBZload ptr mem)) // PPC has no signed-byte load. (Load ptr mem) && is8BitInt(t) && !t.IsSigned() => (MOVBZload ptr mem) (Load ptr mem) && is32BitFloat(t) => (FMOVSload ptr mem) (Load ptr mem) && is64BitFloat(t) => (FMOVDload ptr mem) (Store {t} ptr val mem) && t.Size() == 8 && t.IsFloat() => (FMOVDstore ptr val mem) (Store {t} ptr val mem) && t.Size() == 4 && t.IsFloat() => (FMOVSstore ptr val mem) (Store {t} ptr val mem) && t.Size() == 8 && !t.IsFloat() => (MOVDstore ptr val mem) (Store {t} ptr val mem) && t.Size() == 4 && !t.IsFloat() => (MOVWstore ptr val mem) (Store {t} ptr val mem) && t.Size() == 2 => (MOVHstore ptr val mem) (Store {t} ptr val mem) && t.Size() == 1 => (MOVBstore ptr val mem) // Using Zero instead of LoweredZero allows the // target address to be folded where possible. (Zero [0] _ mem) => mem (Zero [1] destptr mem) => (MOVBstorezero destptr mem) (Zero [2] destptr mem) => (MOVHstorezero destptr mem) (Zero [3] destptr mem) => (MOVBstorezero [2] destptr (MOVHstorezero destptr mem)) (Zero [4] destptr mem) => (MOVWstorezero destptr mem) (Zero [5] destptr mem) => (MOVBstorezero [4] destptr (MOVWstorezero destptr mem)) (Zero [6] destptr mem) => (MOVHstorezero [4] destptr (MOVWstorezero destptr mem)) (Zero [7] destptr mem) => (MOVBstorezero [6] destptr (MOVHstorezero [4] destptr (MOVWstorezero destptr mem))) (Zero [8] {t} destptr mem) => (MOVDstorezero destptr mem) (Zero [12] {t} destptr mem) => (MOVWstorezero [8] destptr (MOVDstorezero [0] destptr mem)) (Zero [16] {t} destptr mem) => (MOVDstorezero [8] destptr (MOVDstorezero [0] destptr mem)) (Zero [24] {t} destptr mem) => (MOVDstorezero [16] destptr (MOVDstorezero [8] destptr (MOVDstorezero [0] destptr mem))) (Zero [32] {t} destptr mem) => (MOVDstorezero [24] destptr (MOVDstorezero [16] destptr (MOVDstorezero [8] destptr (MOVDstorezero [0] destptr mem)))) // Handle cases not handled above // Lowered Short cases do not generate loops, and as a result don't clobber // the address registers or flags. (Zero [s] ptr mem) && buildcfg.GOPPC64 <= 8 && s < 64 => (LoweredZeroShort [s] ptr mem) (Zero [s] ptr mem) && buildcfg.GOPPC64 <= 8 => (LoweredZero [s] ptr mem) (Zero [s] ptr mem) && s < 128 && buildcfg.GOPPC64 >= 9 => (LoweredQuadZeroShort [s] ptr mem) (Zero [s] ptr mem) && buildcfg.GOPPC64 >= 9 => (LoweredQuadZero [s] ptr mem) // moves (Move [0] _ _ mem) => mem (Move [1] dst src mem) => (MOVBstore dst (MOVBZload src mem) mem) (Move [2] dst src mem) => (MOVHstore dst (MOVHZload src mem) mem) (Move [4] dst src mem) => (MOVWstore dst (MOVWZload src mem) mem) // MOVD for load and store must have offsets that are multiple of 4 (Move [8] {t} dst src mem) => (MOVDstore dst (MOVDload src mem) mem) (Move [3] dst src mem) => (MOVBstore [2] dst (MOVBZload [2] src mem) (MOVHstore dst (MOVHload src mem) mem)) (Move [5] dst src mem) => (MOVBstore [4] dst (MOVBZload [4] src mem) (MOVWstore dst (MOVWZload src mem) mem)) (Move [6] dst src mem) => (MOVHstore [4] dst (MOVHZload [4] src mem) (MOVWstore dst (MOVWZload src mem) mem)) (Move [7] dst src mem) => (MOVBstore [6] dst (MOVBZload [6] src mem) (MOVHstore [4] dst (MOVHZload [4] src mem) (MOVWstore dst (MOVWZload src mem) mem))) // Large move uses a loop. Since the address is computed and the // offset is zero, any alignment can be used. (Move [s] dst src mem) && s > 8 && buildcfg.GOPPC64 <= 8 && logLargeCopy(v, s) => (LoweredMove [s] dst src mem) (Move [s] dst src mem) && s > 8 && s <= 64 && buildcfg.GOPPC64 >= 9 => (LoweredQuadMoveShort [s] dst src mem) (Move [s] dst src mem) && s > 8 && buildcfg.GOPPC64 >= 9 && logLargeCopy(v, s) => (LoweredQuadMove [s] dst src mem) // Calls // Lowering calls (StaticCall ...) => (CALLstatic ...) (ClosureCall ...) => (CALLclosure ...) (InterCall ...) => (CALLinter ...) (TailCall ...) => (CALLtail ...) // Miscellaneous (GetClosurePtr ...) => (LoweredGetClosurePtr ...) (GetCallerSP ...) => (LoweredGetCallerSP ...) (GetCallerPC ...) => (LoweredGetCallerPC ...) (IsNonNil ptr) => (NotEqual (CMPconst [0] ptr)) (IsInBounds idx len) => (LessThan (CMPU idx len)) (IsSliceInBounds idx len) => (LessEqual (CMPU idx len)) (NilCheck ...) => (LoweredNilCheck ...) // Write barrier. (WB ...) => (LoweredWB ...) // Publication barrier as intrinsic (PubBarrier ...) => (LoweredPubBarrier ...) (PanicBounds [kind] x y mem) && boundsABI(kind) == 0 => (LoweredPanicBoundsA [kind] x y mem) (PanicBounds [kind] x y mem) && boundsABI(kind) == 1 => (LoweredPanicBoundsB [kind] x y mem) (PanicBounds [kind] x y mem) && boundsABI(kind) == 2 => (LoweredPanicBoundsC [kind] x y mem) // Optimizations // Note that PPC "logical" immediates come in 0:15 and 16:31 unsigned immediate forms, // so ORconst, XORconst easily expand into a pair. // Include very-large constants in the const-const case. (AND (MOVDconst [c]) (MOVDconst [d])) => (MOVDconst [c&d]) (OR (MOVDconst [c]) (MOVDconst [d])) => (MOVDconst [c|d]) (XOR (MOVDconst [c]) (MOVDconst [d])) => (MOVDconst [c^d]) (ORN (MOVDconst [c]) (MOVDconst [d])) => (MOVDconst [c|^d]) (ANDN (MOVDconst [c]) (MOVDconst [d])) => (MOVDconst [c&^d]) (NOR (MOVDconst [c]) (MOVDconst [d])) => (MOVDconst [^(c|d)]) // Discover consts (AND x (MOVDconst [-1])) => x (AND x (MOVDconst [c])) && isU16Bit(c) => (Select0 (ANDCCconst [c] x)) (XOR x (MOVDconst [c])) && isU32Bit(c) => (XORconst [c] x) (OR x (MOVDconst [c])) && isU32Bit(c) => (ORconst [c] x) // Simplify consts (ANDCCconst [c] (Select0 (ANDCCconst [d] x))) => (ANDCCconst [c&d] x) (ORconst [c] (ORconst [d] x)) => (ORconst [c|d] x) (XORconst [c] (XORconst [d] x)) => (XORconst [c^d] x) (Select0 (ANDCCconst [-1] x)) => x (Select0 (ANDCCconst [0] _)) => (MOVDconst [0]) (Select1 (ANDCCconst [0] _)) => (FlagEQ) (XORconst [0] x) => x (ORconst [-1] _) => (MOVDconst [-1]) (ORconst [0] x) => x // zero-extend of small and => small and (MOVBZreg y:(Select0 (ANDCCconst [c] _))) && uint64(c) <= 0xFF => y (MOVHZreg y:(Select0 (ANDCCconst [c] _))) && uint64(c) <= 0xFFFF => y (MOVWZreg y:(Select0 (ANDCCconst [c] _))) && uint64(c) <= 0xFFFFFFFF => y (MOVWZreg y:(AND (MOVDconst [c]) _)) && uint64(c) <= 0xFFFFFFFF => y // sign extend of small-positive and => small-positive-and (MOVBreg y:(Select0 (ANDCCconst [c] _))) && uint64(c) <= 0x7F => y (MOVHreg y:(Select0 (ANDCCconst [c] _))) && uint64(c) <= 0x7FFF => y (MOVWreg y:(Select0 (ANDCCconst [c] _))) && uint64(c) <= 0xFFFF => y // 0xFFFF is largest immediate constant, when regarded as 32-bit is > 0 (MOVWreg y:(AND (MOVDconst [c]) _)) && uint64(c) <= 0x7FFFFFFF => y // small and of zero-extend => either zero-extend or small and (Select0 (ANDCCconst [c] y:(MOVBZreg _))) && c&0xFF == 0xFF => y (Select0 (ANDCCconst [0xFF] (MOVBreg x))) => (MOVBZreg x) (Select0 (ANDCCconst [c] y:(MOVHZreg _))) && c&0xFFFF == 0xFFFF => y (Select0 (ANDCCconst [0xFFFF] (MOVHreg x))) => (MOVHZreg x) (AND (MOVDconst [c]) y:(MOVWZreg _)) && c&0xFFFFFFFF == 0xFFFFFFFF => y (AND (MOVDconst [0xFFFFFFFF]) y:(MOVWreg x)) => (MOVWZreg x) // normal case (Select0 (ANDCCconst [c] (MOVBZreg x))) => (Select0 (ANDCCconst [c&0xFF] x)) (Select0 (ANDCCconst [c] (MOVHZreg x))) => (Select0 (ANDCCconst [c&0xFFFF] x)) (Select0 (ANDCCconst [c] (MOVWZreg x))) => (Select0 (ANDCCconst [c&0xFFFFFFFF] x)) // Eliminate unnecessary sign/zero extend following right shift (MOV(B|H|W)Zreg (SRWconst [c] (MOVBZreg x))) => (SRWconst [c] (MOVBZreg x)) (MOV(H|W)Zreg (SRWconst [c] (MOVHZreg x))) => (SRWconst [c] (MOVHZreg x)) (MOVWZreg (SRWconst [c] (MOVWZreg x))) => (SRWconst [c] (MOVWZreg x)) (MOV(B|H|W)reg (SRAWconst [c] (MOVBreg x))) => (SRAWconst [c] (MOVBreg x)) (MOV(H|W)reg (SRAWconst [c] (MOVHreg x))) => (SRAWconst [c] (MOVHreg x)) (MOVWreg (SRAWconst [c] (MOVWreg x))) => (SRAWconst [c] (MOVWreg x)) (MOV(WZ|W)reg (S(R|RA)Wconst [c] x)) && sizeof(x.Type) <= 32 => (S(R|RA)Wconst [c] x) (MOV(HZ|H)reg (S(R|RA)Wconst [c] x)) && sizeof(x.Type) <= 16 => (S(R|RA)Wconst [c] x) (MOV(BZ|B)reg (S(R|RA)Wconst [c] x)) && sizeof(x.Type) == 8 => (S(R|RA)Wconst [c] x) // initial right shift will handle sign/zero extend (MOVBZreg (SRDconst [c] x)) && c>=56 => (SRDconst [c] x) (MOVBreg (SRDconst [c] x)) && c>56 => (SRDconst [c] x) (MOVBreg (SRDconst [c] x)) && c==56 => (SRADconst [c] x) (MOVBreg (SRADconst [c] x)) && c>=56 => (SRADconst [c] x) (MOVBZreg (SRWconst [c] x)) && c>=24 => (SRWconst [c] x) (MOVBreg (SRWconst [c] x)) && c>24 => (SRWconst [c] x) (MOVBreg (SRWconst [c] x)) && c==24 => (SRAWconst [c] x) (MOVBreg (SRAWconst [c] x)) && c>=24 => (SRAWconst [c] x) (MOVHZreg (SRDconst [c] x)) && c>=48 => (SRDconst [c] x) (MOVHreg (SRDconst [c] x)) && c>48 => (SRDconst [c] x) (MOVHreg (SRDconst [c] x)) && c==48 => (SRADconst [c] x) (MOVHreg (SRADconst [c] x)) && c>=48 => (SRADconst [c] x) (MOVHZreg (SRWconst [c] x)) && c>=16 => (SRWconst [c] x) (MOVHreg (SRWconst [c] x)) && c>16 => (SRWconst [c] x) (MOVHreg (SRAWconst [c] x)) && c>=16 => (SRAWconst [c] x) (MOVHreg (SRWconst [c] x)) && c==16 => (SRAWconst [c] x) (MOVWZreg (SRDconst [c] x)) && c>=32 => (SRDconst [c] x) (MOVWreg (SRDconst [c] x)) && c>32 => (SRDconst [c] x) (MOVWreg (SRADconst [c] x)) && c>=32 => (SRADconst [c] x) (MOVWreg (SRDconst [c] x)) && c==32 => (SRADconst [c] x) // Various redundant zero/sign extension combinations. (MOVBZreg y:(MOVBZreg _)) => y // repeat (MOVBreg y:(MOVBreg _)) => y // repeat (MOVBreg (MOVBZreg x)) => (MOVBreg x) (MOVBZreg (MOVBreg x)) => (MOVBZreg x) // H - there are more combinations than these (MOVHZreg y:(MOV(H|B)Zreg _)) => y // repeat (MOVHZreg y:(MOVHBRload _ _)) => y (MOVHreg y:(MOV(H|B)reg _)) => y // repeat (MOV(H|HZ)reg y:(MOV(HZ|H)reg x)) => (MOV(H|HZ)reg x) // W - there are more combinations than these (MOV(WZ|WZ|WZ|W|W|W)reg y:(MOV(WZ|HZ|BZ|W|H|B)reg _)) => y // repeat (MOVWZreg y:(MOV(H|W)BRload _ _)) => y (MOV(W|WZ)reg y:(MOV(WZ|W)reg x)) => (MOV(W|WZ)reg x) // Truncate then logical then truncate: omit first, lesser or equal truncate (MOVWZreg ((OR|XOR|AND) x (MOVWZreg y))) => (MOVWZreg ((OR|XOR|AND) x y)) (MOVHZreg ((OR|XOR|AND) x (MOVWZreg y))) => (MOVHZreg ((OR|XOR|AND) x y)) (MOVHZreg ((OR|XOR|AND) x (MOVHZreg y))) => (MOVHZreg ((OR|XOR|AND) x y)) (MOVBZreg ((OR|XOR|AND) x (MOVWZreg y))) => (MOVBZreg ((OR|XOR|AND) x y)) (MOVBZreg ((OR|XOR|AND) x (MOVHZreg y))) => (MOVBZreg ((OR|XOR|AND) x y)) (MOVBZreg ((OR|XOR|AND) x (MOVBZreg y))) => (MOVBZreg ((OR|XOR|AND) x y)) (MOV(B|H|W)Zreg z:(Select0 (ANDCCconst [c] (MOVBZload ptr x)))) => z (MOV(B|H|W)Zreg z:(AND y (MOV(B|H|W)Zload ptr x))) => z (MOV(H|W)Zreg z:(Select0 (ANDCCconst [c] (MOVHZload ptr x)))) => z (MOVWZreg z:(Select0 (ANDCCconst [c] (MOVWZload ptr x)))) => z // Arithmetic constant ops (ADD x (MOVDconst [c])) && is32Bit(c) && !t.IsPtr() => (ADDconst [c] x) (ADDconst [c] (ADDconst [d] x)) && is32Bit(c+d) => (ADDconst [c+d] x) (ADDconst [0] x) => x (SUB x (MOVDconst [c])) && is32Bit(-c) => (ADDconst [-c] x) (ADDconst [c] (MOVDaddr [d] {sym} x)) && is32Bit(c+int64(d)) => (MOVDaddr [int32(c+int64(d))] {sym} x) (ADDconst [c] x:(SP)) && is32Bit(c) => (MOVDaddr [int32(c)] x) // so it is rematerializeable (MULL(W|D) x (MOVDconst [c])) && is16Bit(c) => (MULL(W|D)const [int32(c)] x) // Subtract from (with carry, but ignored) constant. // Note, these clobber the carry bit. (SUB (MOVDconst [c]) x) && is32Bit(c) => (SUBFCconst [c] x) (SUBFCconst [c] (NEG x)) => (ADDconst [c] x) (SUBFCconst [c] (SUBFCconst [d] x)) && is32Bit(c-d) => (ADDconst [c-d] x) (SUBFCconst [0] x) => (NEG x) (ADDconst [c] (SUBFCconst [d] x)) && is32Bit(c+d) => (SUBFCconst [c+d] x) (NEG (ADDconst [c] x)) && is32Bit(-c) => (SUBFCconst [-c] x) (NEG (SUBFCconst [c] x)) && is32Bit(-c) => (ADDconst [-c] x) (NEG (SUB x y)) => (SUB y x) (NEG (NEG x)) => x // Use register moves instead of stores and loads to move int<=>float values // Common with math Float64bits, Float64frombits (MOVDload [off] {sym} ptr (FMOVDstore [off] {sym} ptr x _)) => (MFVSRD x) (FMOVDload [off] {sym} ptr (MOVDstore [off] {sym} ptr x _)) => (MTVSRD x) (FMOVDstore [off] {sym} ptr (MTVSRD x) mem) => (MOVDstore [off] {sym} ptr x mem) (MOVDstore [off] {sym} ptr (MFVSRD x) mem) => (FMOVDstore [off] {sym} ptr x mem) (MTVSRD (MOVDconst [c])) && !math.IsNaN(math.Float64frombits(uint64(c))) => (FMOVDconst [math.Float64frombits(uint64(c))]) (MFVSRD (FMOVDconst [c])) => (MOVDconst [int64(math.Float64bits(c))]) (MTVSRD x:(MOVDload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (FMOVDload [off] {sym} ptr mem) (MFVSRD x:(FMOVDload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVDload [off] {sym} ptr mem) // Rules for MOV* or FMOV* ops determine when indexed (MOV*loadidx or MOV*storeidx) // or non-indexed (MOV*load or MOV*store) should be used. Indexed instructions // require an extra instruction and register to load the index so non-indexed is preferred. // Indexed ops generate indexed load or store instructions for all GOPPC64 values. // Non-indexed ops generate DS-form loads and stores when the offset fits in 16 bits, // and on power8 and power9, a multiple of 4 is required for MOVW and MOVD ops. // On power10, prefixed loads and stores can be used for offsets > 16 bits and <= 32 bits. // and support for PC relative addressing must be available if relocation is needed. // On power10, the assembler will determine when to use DS-form or prefixed // instructions for non-indexed ops depending on the value of the offset. // // Fold offsets for stores. (MOV(D|W|H|B)store [off1] {sym} (ADDconst [off2] x) val mem) && (is16Bit(int64(off1)+off2) || (supportsPPC64PCRel() && is32Bit(int64(off1)+off2))) => (MOV(D|W|H|B)store [off1+int32(off2)] {sym} x val mem) (FMOV(S|D)store [off1] {sym} (ADDconst [off2] ptr) val mem) && (is16Bit(int64(off1)+off2) || (supportsPPC64PCRel() && is32Bit(int64(off1)+off2))) => (FMOV(S|D)store [off1+int32(off2)] {sym} ptr val mem) // Fold address into load/store. // If power10 with PCRel is not available, then // the assembler needs to generate several instructions and use // temp register for accessing global, and each time it will reload // the temp register. So don't fold address of global in that case if there is more than // one use. (MOV(B|H|W|D)store [off1] {sym1} p:(MOVDaddr [off2] {sym2} ptr) val mem) && canMergeSym(sym1,sym2) && ((is16Bit(int64(off1+off2)) && (ptr.Op != OpSB || p.Uses == 1)) || (supportsPPC64PCRel() && is32Bit(int64(off1+off2)))) => (MOV(B|H|W|D)store [off1+off2] {mergeSym(sym1,sym2)} ptr val mem) (FMOV(S|D)store [off1] {sym1} p:(MOVDaddr [off2] {sym2} ptr) val mem) && canMergeSym(sym1,sym2) && ((is16Bit(int64(off1+off2)) && (ptr.Op != OpSB || p.Uses == 1)) || (supportsPPC64PCRel() && is32Bit(int64(off1+off2)))) => (FMOV(S|D)store [off1+off2] {mergeSym(sym1,sym2)} ptr val mem) (MOV(B|H|W)Zload [off1] {sym1} p:(MOVDaddr [off2] {sym2} ptr) mem) && canMergeSym(sym1,sym2) && ((is16Bit(int64(off1+off2)) && (ptr.Op != OpSB || p.Uses == 1)) || (supportsPPC64PCRel() && is32Bit(int64(off1+off2)))) => (MOV(B|H|W)Zload [off1+off2] {mergeSym(sym1,sym2)} ptr mem) (MOV(H|W|D)load [off1] {sym1} p:(MOVDaddr [off2] {sym2} ptr) mem) && canMergeSym(sym1,sym2) && ((is16Bit(int64(off1+off2)) && (ptr.Op != OpSB || p.Uses == 1)) || (supportsPPC64PCRel() && is32Bit(int64(off1+off2)))) => (MOV(H|W|D)load [off1+off2] {mergeSym(sym1,sym2)} ptr mem) (FMOV(S|D)load [off1] {sym1} p:(MOVDaddr [off2] {sym2} ptr) mem) && canMergeSym(sym1,sym2) && ((is16Bit(int64(off1+off2)) && (ptr.Op != OpSB || p.Uses == 1)) || (supportsPPC64PCRel() && is32Bit(int64(off1+off2)))) => (FMOV(S|D)load [off1+off2] {mergeSym(sym1,sym2)} ptr mem) // Fold offsets for loads. (FMOV(S|D)load [off1] {sym} (ADDconst [off2] ptr) mem) && (is16Bit(int64(off1)+off2) || (supportsPPC64PCRel() && is32Bit(int64(off1)+off2))) => (FMOV(S|D)load [off1+int32(off2)] {sym} ptr mem) (MOV(D|W|WZ|H|HZ|BZ)load [off1] {sym} (ADDconst [off2] x) mem) && (is16Bit(int64(off1)+off2) || (supportsPPC64PCRel() && is32Bit(int64(off1)+off2))) => (MOV(D|W|WZ|H|HZ|BZ)load [off1+int32(off2)] {sym} x mem) // Determine load + addressing that can be done as a register indexed load (MOV(D|W|WZ|H|HZ|BZ)load [0] {sym} p:(ADD ptr idx) mem) && sym == nil && p.Uses == 1 => (MOV(D|W|WZ|H|HZ|BZ)loadidx ptr idx mem) // See comments above concerning selection of indexed vs. non-indexed ops. // These cases don't have relocation. (MOV(D|W)loadidx ptr (MOVDconst [c]) mem) && ((is16Bit(c) && c%4 == 0) || (buildcfg.GOPPC64 >= 10 && is32Bit(c))) => (MOV(D|W)load [int32(c)] ptr mem) (MOV(WZ|H|HZ|BZ)loadidx ptr (MOVDconst [c]) mem) && (is16Bit(c) || (buildcfg.GOPPC64 >= 10 && is32Bit(c))) => (MOV(WZ|H|HZ|BZ)load [int32(c)] ptr mem) (MOV(D|W)loadidx (MOVDconst [c]) ptr mem) && ((is16Bit(c) && c%4 == 0) || (buildcfg.GOPPC64 >= 10 && is32Bit(c))) => (MOV(D|W)load [int32(c)] ptr mem) (MOV(WZ|H|HZ|BZ)loadidx (MOVDconst [c]) ptr mem) && (is16Bit(c) || (buildcfg.GOPPC64 >= 10 && is32Bit(c))) => (MOV(WZ|H|HZ|BZ)load [int32(c)] ptr mem) // Store of zero => storezero (MOV(D|W|H|B)store [off] {sym} ptr (MOVDconst [0]) mem) => (MOV(D|W|H|B)storezero [off] {sym} ptr mem) // Fold offsets for storezero (MOV(D|W|H|B)storezero [off1] {sym} (ADDconst [off2] x) mem) && ((supportsPPC64PCRel() && is32Bit(int64(off1)+off2)) || (is16Bit(int64(off1)+off2))) => (MOV(D|W|H|B)storezero [off1+int32(off2)] {sym} x mem) // Stores with addressing that can be done as indexed stores (MOV(D|W|H|B)store [0] {sym} p:(ADD ptr idx) val mem) && sym == nil && p.Uses == 1 => (MOV(D|W|H|B)storeidx ptr idx val mem) (MOVDstoreidx ptr (MOVDconst [c]) val mem) && ((is16Bit(c) && c%4 == 0) || (buildcfg.GOPPC64 >= 10 && is32Bit(c))) => (MOVDstore [int32(c)] ptr val mem) (MOV(W|H|B)storeidx ptr (MOVDconst [c]) val mem) && (is16Bit(c) || (buildcfg.GOPPC64 >= 10 && is32Bit(c))) => (MOV(W|H|B)store [int32(c)] ptr val mem) (MOVDstoreidx (MOVDconst [c]) ptr val mem) && ((is16Bit(c) && c%4 == 0) || (buildcfg.GOPPC64 >= 10 && is32Bit(c))) => (MOVDstore [int32(c)] ptr val mem) (MOV(W|H|B)storeidx (MOVDconst [c]) ptr val mem) && (is16Bit(c) || (buildcfg.GOPPC64 >= 10 && is32Bit(c))) => (MOV(W|H|B)store [int32(c)] ptr val mem) // Fold symbols into storezero (MOV(D|W|H|B)storezero [off1] {sym1} p:(MOVDaddr [off2] {sym2} x) mem) && canMergeSym(sym1,sym2) && ((is16Bit(int64(off1+off2)) && (x.Op != OpSB || p.Uses == 1)) || (supportsPPC64PCRel() && is32Bit(int64(off1+off2)))) => (MOV(D|W|H|B)storezero [off1+off2] {mergeSym(sym1,sym2)} x mem) // atomic intrinsics (AtomicLoad(8|32|64|Ptr) ptr mem) => (LoweredAtomicLoad(8|32|64|Ptr) [1] ptr mem) (AtomicLoadAcq(32|64) ptr mem) => (LoweredAtomicLoad(32|64) [0] ptr mem) (AtomicStore(8|32|64) ptr val mem) => (LoweredAtomicStore(8|32|64) [1] ptr val mem) (AtomicStoreRel(32|64) ptr val mem) => (LoweredAtomicStore(32|64) [0] ptr val mem) (AtomicExchange(32|64) ...) => (LoweredAtomicExchange(32|64) ...) (AtomicAdd(32|64) ...) => (LoweredAtomicAdd(32|64) ...) (AtomicCompareAndSwap(32|64) ptr old new_ mem) => (LoweredAtomicCas(32|64) [1] ptr old new_ mem) (AtomicCompareAndSwapRel32 ptr old new_ mem) => (LoweredAtomicCas32 [0] ptr old new_ mem) (AtomicAnd(8|32) ...) => (LoweredAtomicAnd(8|32) ...) (AtomicOr(8|32) ...) => (LoweredAtomicOr(8|32) ...) (Slicemask x) => (SRADconst (NEG x) [63]) (Select0 (ANDCCconst [1] z:(SRADconst [63] x))) && z.Uses == 1 => (SRDconst [63] x) // Note that MOV??reg returns a 64-bit int, x is not necessarily that wide // This may interact with other patterns in the future. (Compare with arm64) (MOV(B|H|W)Zreg x:(MOVBZload _ _)) => x (MOV(B|H|W)Zreg x:(MOVBZloadidx _ _ _)) => x (MOV(H|W)Zreg x:(MOVHZload _ _)) => x (MOV(H|W)Zreg x:(MOVHZloadidx _ _ _)) => x (MOV(H|W)reg x:(MOVHload _ _)) => x (MOV(H|W)reg x:(MOVHloadidx _ _ _)) => x (MOV(WZ|W)reg x:(MOV(WZ|W)load _ _)) => x (MOV(WZ|W)reg x:(MOV(WZ|W)loadidx _ _ _)) => x (MOV(B|W)Zreg x:(Select0 (LoweredAtomicLoad(8|32) _ _))) => x // don't extend if argument is already extended (MOVBreg x:(Arg )) && is8BitInt(t) && t.IsSigned() => x (MOVBZreg x:(Arg )) && is8BitInt(t) && !t.IsSigned() => x (MOVHreg x:(Arg )) && (is8BitInt(t) || is16BitInt(t)) && t.IsSigned() => x (MOVHZreg x:(Arg )) && (is8BitInt(t) || is16BitInt(t)) && !t.IsSigned() => x (MOVWreg x:(Arg )) && (is8BitInt(t) || is16BitInt(t) || is32BitInt(t)) && t.IsSigned() => x (MOVWZreg x:(Arg )) && (is8BitInt(t) || is16BitInt(t) || is32BitInt(t)) && !t.IsSigned() => x (MOVBZreg (MOVDconst [c])) => (MOVDconst [int64(uint8(c))]) (MOVBreg (MOVDconst [c])) => (MOVDconst [int64(int8(c))]) (MOVHZreg (MOVDconst [c])) => (MOVDconst [int64(uint16(c))]) (MOVHreg (MOVDconst [c])) => (MOVDconst [int64(int16(c))]) (MOVWreg (MOVDconst [c])) => (MOVDconst [int64(int32(c))]) (MOVWZreg (MOVDconst [c])) => (MOVDconst [int64(uint32(c))]) // Implement clrsldi and clrslwi extended mnemonics as described in // ISA 3.0 section C.8. AuxInt field contains values needed for // the instructions, packed together since there is only one available. (SLDconst [c] z:(MOVBZreg x)) && c < 8 && z.Uses == 1 => (CLRLSLDI [newPPC64ShiftAuxInt(c,56,63,64)] x) (SLDconst [c] z:(MOVHZreg x)) && c < 16 && z.Uses == 1 => (CLRLSLDI [newPPC64ShiftAuxInt(c,48,63,64)] x) (SLDconst [c] z:(MOVWZreg x)) && c < 32 && z.Uses == 1 => (CLRLSLDI [newPPC64ShiftAuxInt(c,32,63,64)] x) (SLDconst [c] z:(Select0 (ANDCCconst [d] x))) && z.Uses == 1 && isPPC64ValidShiftMask(d) && c <= (64-getPPC64ShiftMaskLength(d)) => (CLRLSLDI [newPPC64ShiftAuxInt(c,64-getPPC64ShiftMaskLength(d),63,64)] x) (SLDconst [c] z:(AND (MOVDconst [d]) x)) && z.Uses == 1 && isPPC64ValidShiftMask(d) && c<=(64-getPPC64ShiftMaskLength(d)) => (CLRLSLDI [newPPC64ShiftAuxInt(c,64-getPPC64ShiftMaskLength(d),63,64)] x) (SLWconst [c] z:(MOVBZreg x)) && z.Uses == 1 && c < 8 => (CLRLSLWI [newPPC64ShiftAuxInt(c,24,31,32)] x) (SLWconst [c] z:(MOVHZreg x)) && z.Uses == 1 && c < 16 => (CLRLSLWI [newPPC64ShiftAuxInt(c,16,31,32)] x) (SLWconst [c] z:(Select0 (ANDCCconst [d] x))) && z.Uses == 1 && isPPC64ValidShiftMask(d) && c<=(32-getPPC64ShiftMaskLength(d)) => (CLRLSLWI [newPPC64ShiftAuxInt(c,32-getPPC64ShiftMaskLength(d),31,32)] x) (SLWconst [c] z:(AND (MOVDconst [d]) x)) && z.Uses == 1 && isPPC64ValidShiftMask(d) && c<=(32-getPPC64ShiftMaskLength(d)) => (CLRLSLWI [newPPC64ShiftAuxInt(c,32-getPPC64ShiftMaskLength(d),31,32)] x) // special case for power9 (SL(W|D)const [c] z:(MOVWreg x)) && c < 32 && buildcfg.GOPPC64 >= 9 => (EXTSWSLconst [c] x) // Lose widening ops fed to stores (MOVBstore [off] {sym} ptr (MOV(B|BZ|H|HZ|W|WZ)reg x) mem) => (MOVBstore [off] {sym} ptr x mem) (MOVHstore [off] {sym} ptr (MOV(H|HZ|W|WZ)reg x) mem) => (MOVHstore [off] {sym} ptr x mem) (MOVWstore [off] {sym} ptr (MOV(W|WZ)reg x) mem) => (MOVWstore [off] {sym} ptr x mem) (MOVBstore [off] {sym} ptr (SRWconst (MOV(H|HZ)reg x) [c]) mem) && c <= 8 => (MOVBstore [off] {sym} ptr (SRWconst x [c]) mem) (MOVBstore [off] {sym} ptr (SRWconst (MOV(W|WZ)reg x) [c]) mem) && c <= 24 => (MOVBstore [off] {sym} ptr (SRWconst x [c]) mem) (MOVBstoreidx ptr idx (MOV(B|BZ|H|HZ|W|WZ)reg x) mem) => (MOVBstoreidx ptr idx x mem) (MOVHstoreidx ptr idx (MOV(H|HZ|W|WZ)reg x) mem) => (MOVHstoreidx ptr idx x mem) (MOVWstoreidx ptr idx (MOV(W|WZ)reg x) mem) => (MOVWstoreidx ptr idx x mem) (MOVBstoreidx ptr idx (SRWconst (MOV(H|HZ)reg x) [c]) mem) && c <= 8 => (MOVBstoreidx ptr idx (SRWconst x [c]) mem) (MOVBstoreidx ptr idx (SRWconst (MOV(W|WZ)reg x) [c]) mem) && c <= 24 => (MOVBstoreidx ptr idx (SRWconst x [c]) mem) (MOVHBRstore ptr (MOV(H|HZ|W|WZ)reg x) mem) => (MOVHBRstore ptr x mem) (MOVWBRstore ptr (MOV(W|WZ)reg x) mem) => (MOVWBRstore ptr x mem) // Lose W-widening ops fed to compare-W (CMP(W|WU) x (MOV(W|WZ)reg y)) => (CMP(W|WU) x y) (CMP(W|WU) (MOV(W|WZ)reg x) y) => (CMP(W|WU) x y) (CMP x (MOVDconst [c])) && is16Bit(c) => (CMPconst x [c]) (CMP (MOVDconst [c]) y) && is16Bit(c) => (InvertFlags (CMPconst y [c])) (CMPW x (MOVDconst [c])) && is16Bit(c) => (CMPWconst x [int32(c)]) (CMPW (MOVDconst [c]) y) && is16Bit(c) => (InvertFlags (CMPWconst y [int32(c)])) (CMPU x (MOVDconst [c])) && isU16Bit(c) => (CMPUconst x [c]) (CMPU (MOVDconst [c]) y) && isU16Bit(c) => (InvertFlags (CMPUconst y [c])) (CMPWU x (MOVDconst [c])) && isU16Bit(c) => (CMPWUconst x [int32(c)]) (CMPWU (MOVDconst [c]) y) && isU16Bit(c) => (InvertFlags (CMPWUconst y [int32(c)])) // Canonicalize the order of arguments to comparisons - helps with CSE. ((CMP|CMPW|CMPU|CMPWU) x y) && canonLessThan(x,y) => (InvertFlags ((CMP|CMPW|CMPU|CMPWU) y x)) // SETBC auxInt values 0=LT 1=GT 2=EQ Crbit==1 ? 1 : 0 // SETBCR auxInt values 0=LT 1=GT 2=EQ Crbit==1 ? 0 : 1 (Equal cmp) => (SETBC [2] cmp) (NotEqual cmp) => (SETBCR [2] cmp) (LessThan cmp) => (SETBC [0] cmp) (FLessThan cmp) => (SETBC [0] cmp) (FLessEqual cmp) => (OR (SETBC [2] cmp) (SETBC [0] cmp)) (GreaterEqual cmp) => (SETBCR [0] cmp) (GreaterThan cmp) => (SETBC [1] cmp) (FGreaterEqual cmp) => (OR (SETBC [2] cmp) (SETBC [1] cmp)) (FGreaterThan cmp) => (SETBC [1] cmp) (LessEqual cmp) => (SETBCR [1] cmp) (SETBC [0] (FlagLT)) => (MOVDconst [1]) (SETBC [0] (Flag(GT|EQ))) => (MOVDconst [0]) (SETBC [1] (FlagGT)) => (MOVDconst [1]) (SETBC [1] (Flag(LT|EQ))) => (MOVDconst [0]) (SETBC [2] (FlagEQ)) => (MOVDconst [1]) (SETBC [2] (Flag(LT|GT))) => (MOVDconst [0]) (SETBCR [0] (FlagLT)) => (MOVDconst [0]) (SETBCR [0] (Flag(GT|EQ))) => (MOVDconst [1]) (SETBCR [1] (FlagGT)) => (MOVDconst [0]) (SETBCR [1] (Flag(LT|EQ))) => (MOVDconst [1]) (SETBCR [2] (FlagEQ)) => (MOVDconst [0]) (SETBCR [2] (Flag(LT|GT))) => (MOVDconst [1]) (SETBC [0] (InvertFlags bool)) => (SETBC [1] bool) (SETBC [1] (InvertFlags bool)) => (SETBC [0] bool) (SETBC [2] (InvertFlags bool)) => (SETBC [2] bool) (SETBCR [0] (InvertFlags bool)) => (SETBCR [1] bool) (SETBCR [1] (InvertFlags bool)) => (SETBCR [0] bool) (SETBCR [2] (InvertFlags bool)) => (SETBCR [2] bool) // ISEL auxInt values 0=LT 1=GT 2=EQ arg2 ? arg0 : arg1 // ISEL auxInt values 4=GE 5=LE 6=NE !arg2 ? arg1 : arg0 (ISEL [2] x _ (FlagEQ)) => x (ISEL [2] _ y (Flag(LT|GT))) => y (ISEL [6] _ y (FlagEQ)) => y (ISEL [6] x _ (Flag(LT|GT))) => x (ISEL [0] _ y (Flag(EQ|GT))) => y (ISEL [0] x _ (FlagLT)) => x (ISEL [5] _ x (Flag(EQ|LT))) => x (ISEL [5] y _ (FlagGT)) => y (ISEL [1] _ y (Flag(EQ|LT))) => y (ISEL [1] x _ (FlagGT)) => x (ISEL [4] x _ (Flag(EQ|GT))) => x (ISEL [4] _ y (FlagLT)) => y (ISEL [2] x y ((CMP|CMPW)const [0] (Select0 (ANDCCconst [n] z)))) => (ISEL [2] x y (Select1 (ANDCCconst [n] z ))) (ISEL [6] x y ((CMP|CMPW)const [0] (Select0 (ANDCCconst [n] z)))) => (ISEL [6] x y (Select1 (ANDCCconst [n] z ))) (SETBC [n] (InvertFlags bool)) => (SETBCR [n] bool) (SETBCR [n] (InvertFlags bool)) => (SETBC [n] bool) (ISEL [n] x y (InvertFlags bool)) && n%4 == 0 => (ISEL [n+1] x y bool) (ISEL [n] x y (InvertFlags bool)) && n%4 == 1 => (ISEL [n-1] x y bool) (ISEL [n] x y (InvertFlags bool)) && n%4 == 2 => (ISEL [n] x y bool) (XORconst [1] (SETBCR [n] cmp)) => (SETBC [n] cmp) (XORconst [1] (SETBC [n] cmp)) => (SETBCR [n] cmp) (SETBC [2] ((CMP|CMPW)const [0] (Select0 (ANDCCconst [1] z)))) => (XORconst [1] (Select0 (ANDCCconst [1] z ))) (SETBCR [2] ((CMP|CMPW)const [0] (Select0 (ANDCCconst [1] z)))) => (Select0 (ANDCCconst [1] z )) (SETBC [2] (CMPWconst [0] (Select0 (ANDCCconst [n] z)))) => (SETBC [2] (Select1 (ANDCCconst [n] z ))) (SETBCR [2] (CMPWconst [0] (Select0 (ANDCCconst [n] z)))) => (SETBCR [2] (Select1 (ANDCCconst [n] z ))) // Only CMPconst for these in case AND|OR|XOR result is > 32 bits (SETBC [2] (CMPconst [0] a:(AND y z))) && a.Uses == 1 => (SETBC [2] (Select1 (ANDCC y z ))) (SETBCR [2] (CMPconst [0] a:(AND y z))) && a.Uses == 1 => (SETBCR [2] (Select1 (ANDCC y z ))) (SETBC [2] (CMPconst [0] o:(OR y z))) && o.Uses == 1 => (SETBC [2] (Select1 (ORCC y z ))) (SETBCR [2] (CMPconst [0] o:(OR y z))) && o.Uses == 1 => (SETBCR [2] (Select1 (ORCC y z ))) (SETBC [2] (CMPconst [0] a:(XOR y z))) && a.Uses == 1 => (SETBC [2] (Select1 (XORCC y z ))) (SETBCR [2] (CMPconst [0] a:(XOR y z))) && a.Uses == 1 => (SETBCR [2] (Select1 (XORCC y z ))) // A particular pattern seen in cgo code: (AND (MOVDconst [c]) x:(MOVBZload _ _)) => (Select0 (ANDCCconst [c&0xFF] x)) // floating point negative abs (FNEG (F(ABS|NABS) x)) => (F(NABS|ABS) x) // floating-point fused multiply-add/sub (F(ADD|SUB) (FMUL x y) z) && x.Block.Func.useFMA(v) => (FM(ADD|SUB) x y z) (F(ADDS|SUBS) (FMULS x y) z) && x.Block.Func.useFMA(v) => (FM(ADDS|SUBS) x y z) // Arch-specific inlining for small or disjoint runtime.memmove (SelectN [0] call:(CALLstatic {sym} s1:(MOVDstore _ (MOVDconst [sz]) s2:(MOVDstore _ src s3:(MOVDstore {t} _ dst mem))))) && sz >= 0 && isSameCall(sym, "runtime.memmove") && s1.Uses == 1 && s2.Uses == 1 && s3.Uses == 1 && isInlinableMemmove(dst, src, sz, config) && clobber(s1, s2, s3, call) => (Move [sz] dst src mem) // Match post-lowering calls, register version. (SelectN [0] call:(CALLstatic {sym} dst src (MOVDconst [sz]) mem)) && sz >= 0 && isSameCall(sym, "runtime.memmove") && call.Uses == 1 && isInlinableMemmove(dst, src, sz, config) && clobber(call) => (Move [sz] dst src mem) // Prefetch instructions (TH specified using aux field) // For DCBT Ra,Rb,TH, A value of TH indicates: // 0, hint this cache line will be used soon. (PrefetchCache) // 16, hint this cache line will not be used for long. (PrefetchCacheStreamed) // See ISA 3.0 Book II 4.3.2 for more detail. https://openpower.foundation/specifications/isa/ (PrefetchCache ptr mem) => (DCBT ptr mem [0]) (PrefetchCacheStreamed ptr mem) => (DCBT ptr mem [16]) // Use byte reverse instructions on Power10 (Bswap(16|32|64) x) && buildcfg.GOPPC64>=10 => (BR(H|W|D) x) // Fold bit reversal into loads. (BR(W|H) x:(MOV(W|H)Zload [off] {sym} ptr mem)) && x.Uses == 1 => @x.Block (MOV(W|H)BRload (MOVDaddr [off] {sym} ptr) mem) (BR(W|H) x:(MOV(W|H)Zloadidx ptr idx mem)) && x.Uses == 1 => @x.Block (MOV(W|H)BRloadidx ptr idx mem) (BRD x:(MOVDload [off] {sym} ptr mem)) && x.Uses == 1 => @x.Block (MOVDBRload (MOVDaddr [off] {sym} ptr) mem) (BRD x:(MOVDloadidx ptr idx mem)) && x.Uses == 1 => @x.Block (MOVDBRloadidx ptr idx mem) // Fold bit reversal into stores. (MOV(D|W|H)store [off] {sym} ptr r:(BR(D|W|H) val) mem) && r.Uses == 1 => (MOV(D|W|H)BRstore (MOVDaddr [off] {sym} ptr) val mem) (MOV(D|W|H)storeidx ptr idx r:(BR(D|W|H) val) mem) && r.Uses == 1 => (MOV(D|W|H)BRstoreidx ptr idx val mem) // GOPPC64<10 rules. // These Bswap operations should only be introduced by the memcombine pass in places where they can be folded into loads or stores. (Bswap(32|16) x:(MOV(W|H)Zload [off] {sym} ptr mem)) => @x.Block (MOV(W|H)BRload (MOVDaddr [off] {sym} ptr) mem) (Bswap(32|16) x:(MOV(W|H)Zloadidx ptr idx mem)) => @x.Block (MOV(W|H)BRloadidx ptr idx mem) (Bswap64 x:(MOVDload [off] {sym} ptr mem)) => @x.Block (MOVDBRload (MOVDaddr [off] {sym} ptr) mem) (Bswap64 x:(MOVDloadidx ptr idx mem)) => @x.Block (MOVDBRloadidx ptr idx mem) (MOV(D|W|H)store [off] {sym} ptr (Bswap(64|32|16) val) mem) => (MOV(D|W|H)BRstore (MOVDaddr [off] {sym} ptr) val mem) (MOV(D|W|H)storeidx ptr idx (Bswap(64|32|16) val) mem) => (MOV(D|W|H)BRstoreidx ptr idx val mem)